4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
32 #include <linux/module.h>
33 #include <linux/nmi.h>
34 #include <linux/init.h>
35 #include <linux/uaccess.h>
36 #include <linux/highmem.h>
37 #include <linux/smp_lock.h>
38 #include <asm/mmu_context.h>
39 #include <linux/interrupt.h>
40 #include <linux/capability.h>
41 #include <linux/completion.h>
42 #include <linux/kernel_stat.h>
43 #include <linux/debug_locks.h>
44 #include <linux/perf_event.h>
45 #include <linux/security.h>
46 #include <linux/notifier.h>
47 #include <linux/profile.h>
48 #include <linux/freezer.h>
49 #include <linux/vmalloc.h>
50 #include <linux/blkdev.h>
51 #include <linux/delay.h>
52 #include <linux/pid_namespace.h>
53 #include <linux/smp.h>
54 #include <linux/threads.h>
55 #include <linux/timer.h>
56 #include <linux/rcupdate.h>
57 #include <linux/cpu.h>
58 #include <linux/cpuset.h>
59 #include <linux/percpu.h>
60 #include <linux/kthread.h>
61 #include <linux/proc_fs.h>
62 #include <linux/seq_file.h>
63 #include <linux/sysctl.h>
64 #include <linux/syscalls.h>
65 #include <linux/times.h>
66 #include <linux/tsacct_kern.h>
67 #include <linux/kprobes.h>
68 #include <linux/delayacct.h>
69 #include <linux/unistd.h>
70 #include <linux/pagemap.h>
71 #include <linux/hrtimer.h>
72 #include <linux/tick.h>
73 #include <linux/debugfs.h>
74 #include <linux/ctype.h>
75 #include <linux/ftrace.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/sched.h>
86 * Convert user-nice values [ -20 ... 0 ... 19 ]
87 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
90 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
91 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
92 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
95 * 'User priority' is the nice value converted to something we
96 * can work with better when scaling various scheduler parameters,
97 * it's a [ 0 ... 39 ] range.
99 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
100 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
101 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
104 * Helpers for converting nanosecond timing to jiffy resolution
106 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
108 #define NICE_0_LOAD SCHED_LOAD_SCALE
109 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
112 * These are the 'tuning knobs' of the scheduler:
114 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
115 * Timeslices get refilled after they expire.
117 #define DEF_TIMESLICE (100 * HZ / 1000)
120 * single value that denotes runtime == period, ie unlimited time.
122 #define RUNTIME_INF ((u64)~0ULL)
124 static inline int rt_policy(int policy)
126 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
131 static inline int task_has_rt_policy(struct task_struct *p)
133 return rt_policy(p->policy);
137 * This is the priority-queue data structure of the RT scheduling class:
139 struct rt_prio_array {
140 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
141 struct list_head queue[MAX_RT_PRIO];
144 struct rt_bandwidth {
145 /* nests inside the rq lock: */
146 raw_spinlock_t rt_runtime_lock;
149 struct hrtimer rt_period_timer;
152 static struct rt_bandwidth def_rt_bandwidth;
154 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
156 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
158 struct rt_bandwidth *rt_b =
159 container_of(timer, struct rt_bandwidth, rt_period_timer);
165 now = hrtimer_cb_get_time(timer);
166 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
171 idle = do_sched_rt_period_timer(rt_b, overrun);
174 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
178 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
180 rt_b->rt_period = ns_to_ktime(period);
181 rt_b->rt_runtime = runtime;
183 raw_spin_lock_init(&rt_b->rt_runtime_lock);
185 hrtimer_init(&rt_b->rt_period_timer,
186 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
187 rt_b->rt_period_timer.function = sched_rt_period_timer;
190 static inline int rt_bandwidth_enabled(void)
192 return sysctl_sched_rt_runtime >= 0;
195 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
199 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
202 if (hrtimer_active(&rt_b->rt_period_timer))
205 raw_spin_lock(&rt_b->rt_runtime_lock);
210 if (hrtimer_active(&rt_b->rt_period_timer))
213 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
214 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
216 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
217 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
218 delta = ktime_to_ns(ktime_sub(hard, soft));
219 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
220 HRTIMER_MODE_ABS_PINNED, 0);
222 raw_spin_unlock(&rt_b->rt_runtime_lock);
225 #ifdef CONFIG_RT_GROUP_SCHED
226 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
228 hrtimer_cancel(&rt_b->rt_period_timer);
233 * sched_domains_mutex serializes calls to arch_init_sched_domains,
234 * detach_destroy_domains and partition_sched_domains.
236 static DEFINE_MUTEX(sched_domains_mutex);
238 #ifdef CONFIG_GROUP_SCHED
240 #include <linux/cgroup.h>
244 static LIST_HEAD(task_groups);
246 /* task group related information */
248 #ifdef CONFIG_CGROUP_SCHED
249 struct cgroup_subsys_state css;
252 #ifdef CONFIG_USER_SCHED
256 #ifdef CONFIG_FAIR_GROUP_SCHED
257 /* schedulable entities of this group on each cpu */
258 struct sched_entity **se;
259 /* runqueue "owned" by this group on each cpu */
260 struct cfs_rq **cfs_rq;
261 unsigned long shares;
264 #ifdef CONFIG_RT_GROUP_SCHED
265 struct sched_rt_entity **rt_se;
266 struct rt_rq **rt_rq;
268 struct rt_bandwidth rt_bandwidth;
272 struct list_head list;
274 struct task_group *parent;
275 struct list_head siblings;
276 struct list_head children;
279 #ifdef CONFIG_USER_SCHED
281 /* Helper function to pass uid information to create_sched_user() */
282 void set_tg_uid(struct user_struct *user)
284 user->tg->uid = user->uid;
289 * Every UID task group (including init_task_group aka UID-0) will
290 * be a child to this group.
292 struct task_group root_task_group;
294 #ifdef CONFIG_FAIR_GROUP_SCHED
295 /* Default task group's sched entity on each cpu */
296 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
297 /* Default task group's cfs_rq on each cpu */
298 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq, init_tg_cfs_rq);
299 #endif /* CONFIG_FAIR_GROUP_SCHED */
301 #ifdef CONFIG_RT_GROUP_SCHED
302 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
303 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq, init_rt_rq_var);
304 #endif /* CONFIG_RT_GROUP_SCHED */
305 #else /* !CONFIG_USER_SCHED */
306 #define root_task_group init_task_group
307 #endif /* CONFIG_USER_SCHED */
309 /* task_group_lock serializes add/remove of task groups and also changes to
310 * a task group's cpu shares.
312 static DEFINE_SPINLOCK(task_group_lock);
314 #ifdef CONFIG_FAIR_GROUP_SCHED
317 static int root_task_group_empty(void)
319 return list_empty(&root_task_group.children);
323 #ifdef CONFIG_USER_SCHED
324 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
325 #else /* !CONFIG_USER_SCHED */
326 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
327 #endif /* CONFIG_USER_SCHED */
330 * A weight of 0 or 1 can cause arithmetics problems.
331 * A weight of a cfs_rq is the sum of weights of which entities
332 * are queued on this cfs_rq, so a weight of a entity should not be
333 * too large, so as the shares value of a task group.
334 * (The default weight is 1024 - so there's no practical
335 * limitation from this.)
338 #define MAX_SHARES (1UL << 18)
340 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
343 /* Default task group.
344 * Every task in system belong to this group at bootup.
346 struct task_group init_task_group;
348 /* return group to which a task belongs */
349 static inline struct task_group *task_group(struct task_struct *p)
351 struct task_group *tg;
353 #ifdef CONFIG_USER_SCHED
355 tg = __task_cred(p)->user->tg;
357 #elif defined(CONFIG_CGROUP_SCHED)
358 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
359 struct task_group, css);
361 tg = &init_task_group;
366 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
367 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
369 #ifdef CONFIG_FAIR_GROUP_SCHED
370 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
371 p->se.parent = task_group(p)->se[cpu];
374 #ifdef CONFIG_RT_GROUP_SCHED
375 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
376 p->rt.parent = task_group(p)->rt_se[cpu];
382 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
383 static inline struct task_group *task_group(struct task_struct *p)
388 #endif /* CONFIG_GROUP_SCHED */
390 /* CFS-related fields in a runqueue */
392 struct load_weight load;
393 unsigned long nr_running;
398 struct rb_root tasks_timeline;
399 struct rb_node *rb_leftmost;
401 struct list_head tasks;
402 struct list_head *balance_iterator;
405 * 'curr' points to currently running entity on this cfs_rq.
406 * It is set to NULL otherwise (i.e when none are currently running).
408 struct sched_entity *curr, *next, *last;
410 unsigned int nr_spread_over;
412 #ifdef CONFIG_FAIR_GROUP_SCHED
413 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
416 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
417 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
418 * (like users, containers etc.)
420 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
421 * list is used during load balance.
423 struct list_head leaf_cfs_rq_list;
424 struct task_group *tg; /* group that "owns" this runqueue */
428 * the part of load.weight contributed by tasks
430 unsigned long task_weight;
433 * h_load = weight * f(tg)
435 * Where f(tg) is the recursive weight fraction assigned to
438 unsigned long h_load;
441 * this cpu's part of tg->shares
443 unsigned long shares;
446 * load.weight at the time we set shares
448 unsigned long rq_weight;
453 /* Real-Time classes' related field in a runqueue: */
455 struct rt_prio_array active;
456 unsigned long rt_nr_running;
457 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
459 int curr; /* highest queued rt task prio */
461 int next; /* next highest */
466 unsigned long rt_nr_migratory;
467 unsigned long rt_nr_total;
469 struct plist_head pushable_tasks;
474 /* Nests inside the rq lock: */
475 raw_spinlock_t rt_runtime_lock;
477 #ifdef CONFIG_RT_GROUP_SCHED
478 unsigned long rt_nr_boosted;
481 struct list_head leaf_rt_rq_list;
482 struct task_group *tg;
483 struct sched_rt_entity *rt_se;
490 * We add the notion of a root-domain which will be used to define per-domain
491 * variables. Each exclusive cpuset essentially defines an island domain by
492 * fully partitioning the member cpus from any other cpuset. Whenever a new
493 * exclusive cpuset is created, we also create and attach a new root-domain
500 cpumask_var_t online;
503 * The "RT overload" flag: it gets set if a CPU has more than
504 * one runnable RT task.
506 cpumask_var_t rto_mask;
509 struct cpupri cpupri;
514 * By default the system creates a single root-domain with all cpus as
515 * members (mimicking the global state we have today).
517 static struct root_domain def_root_domain;
522 * This is the main, per-CPU runqueue data structure.
524 * Locking rule: those places that want to lock multiple runqueues
525 * (such as the load balancing or the thread migration code), lock
526 * acquire operations must be ordered by ascending &runqueue.
533 * nr_running and cpu_load should be in the same cacheline because
534 * remote CPUs use both these fields when doing load calculation.
536 unsigned long nr_running;
537 #define CPU_LOAD_IDX_MAX 5
538 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
540 unsigned char in_nohz_recently;
542 /* capture load from *all* tasks on this cpu: */
543 struct load_weight load;
544 unsigned long nr_load_updates;
550 #ifdef CONFIG_FAIR_GROUP_SCHED
551 /* list of leaf cfs_rq on this cpu: */
552 struct list_head leaf_cfs_rq_list;
554 #ifdef CONFIG_RT_GROUP_SCHED
555 struct list_head leaf_rt_rq_list;
559 * This is part of a global counter where only the total sum
560 * over all CPUs matters. A task can increase this counter on
561 * one CPU and if it got migrated afterwards it may decrease
562 * it on another CPU. Always updated under the runqueue lock:
564 unsigned long nr_uninterruptible;
566 struct task_struct *curr, *idle;
567 unsigned long next_balance;
568 struct mm_struct *prev_mm;
575 struct root_domain *rd;
576 struct sched_domain *sd;
578 unsigned char idle_at_tick;
579 /* For active balancing */
583 /* cpu of this runqueue: */
587 unsigned long avg_load_per_task;
589 struct task_struct *migration_thread;
590 struct list_head migration_queue;
598 /* calc_load related fields */
599 unsigned long calc_load_update;
600 long calc_load_active;
602 #ifdef CONFIG_SCHED_HRTICK
604 int hrtick_csd_pending;
605 struct call_single_data hrtick_csd;
607 struct hrtimer hrtick_timer;
610 #ifdef CONFIG_SCHEDSTATS
612 struct sched_info rq_sched_info;
613 unsigned long long rq_cpu_time;
614 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
616 /* sys_sched_yield() stats */
617 unsigned int yld_count;
619 /* schedule() stats */
620 unsigned int sched_switch;
621 unsigned int sched_count;
622 unsigned int sched_goidle;
624 /* try_to_wake_up() stats */
625 unsigned int ttwu_count;
626 unsigned int ttwu_local;
629 unsigned int bkl_count;
633 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
636 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
638 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
641 static inline int cpu_of(struct rq *rq)
651 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
652 * See detach_destroy_domains: synchronize_sched for details.
654 * The domain tree of any CPU may only be accessed from within
655 * preempt-disabled sections.
657 #define for_each_domain(cpu, __sd) \
658 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
660 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
661 #define this_rq() (&__get_cpu_var(runqueues))
662 #define task_rq(p) cpu_rq(task_cpu(p))
663 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
664 #define raw_rq() (&__raw_get_cpu_var(runqueues))
666 inline void update_rq_clock(struct rq *rq)
668 rq->clock = sched_clock_cpu(cpu_of(rq));
672 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
674 #ifdef CONFIG_SCHED_DEBUG
675 # define const_debug __read_mostly
677 # define const_debug static const
682 * @cpu: the processor in question.
684 * Returns true if the current cpu runqueue is locked.
685 * This interface allows printk to be called with the runqueue lock
686 * held and know whether or not it is OK to wake up the klogd.
688 int runqueue_is_locked(int cpu)
690 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
694 * Debugging: various feature bits
697 #define SCHED_FEAT(name, enabled) \
698 __SCHED_FEAT_##name ,
701 #include "sched_features.h"
706 #define SCHED_FEAT(name, enabled) \
707 (1UL << __SCHED_FEAT_##name) * enabled |
709 const_debug unsigned int sysctl_sched_features =
710 #include "sched_features.h"
715 #ifdef CONFIG_SCHED_DEBUG
716 #define SCHED_FEAT(name, enabled) \
719 static __read_mostly char *sched_feat_names[] = {
720 #include "sched_features.h"
726 static int sched_feat_show(struct seq_file *m, void *v)
730 for (i = 0; sched_feat_names[i]; i++) {
731 if (!(sysctl_sched_features & (1UL << i)))
733 seq_printf(m, "%s ", sched_feat_names[i]);
741 sched_feat_write(struct file *filp, const char __user *ubuf,
742 size_t cnt, loff_t *ppos)
752 if (copy_from_user(&buf, ubuf, cnt))
757 if (strncmp(buf, "NO_", 3) == 0) {
762 for (i = 0; sched_feat_names[i]; i++) {
763 int len = strlen(sched_feat_names[i]);
765 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
767 sysctl_sched_features &= ~(1UL << i);
769 sysctl_sched_features |= (1UL << i);
774 if (!sched_feat_names[i])
782 static int sched_feat_open(struct inode *inode, struct file *filp)
784 return single_open(filp, sched_feat_show, NULL);
787 static const struct file_operations sched_feat_fops = {
788 .open = sched_feat_open,
789 .write = sched_feat_write,
792 .release = single_release,
795 static __init int sched_init_debug(void)
797 debugfs_create_file("sched_features", 0644, NULL, NULL,
802 late_initcall(sched_init_debug);
806 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
809 * Number of tasks to iterate in a single balance run.
810 * Limited because this is done with IRQs disabled.
812 const_debug unsigned int sysctl_sched_nr_migrate = 32;
815 * ratelimit for updating the group shares.
818 unsigned int sysctl_sched_shares_ratelimit = 250000;
819 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
822 * Inject some fuzzyness into changing the per-cpu group shares
823 * this avoids remote rq-locks at the expense of fairness.
826 unsigned int sysctl_sched_shares_thresh = 4;
829 * period over which we average the RT time consumption, measured
834 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
837 * period over which we measure -rt task cpu usage in us.
840 unsigned int sysctl_sched_rt_period = 1000000;
842 static __read_mostly int scheduler_running;
845 * part of the period that we allow rt tasks to run in us.
848 int sysctl_sched_rt_runtime = 950000;
850 static inline u64 global_rt_period(void)
852 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
855 static inline u64 global_rt_runtime(void)
857 if (sysctl_sched_rt_runtime < 0)
860 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
863 #ifndef prepare_arch_switch
864 # define prepare_arch_switch(next) do { } while (0)
866 #ifndef finish_arch_switch
867 # define finish_arch_switch(prev) do { } while (0)
870 static inline int task_current(struct rq *rq, struct task_struct *p)
872 return rq->curr == p;
875 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
876 static inline int task_running(struct rq *rq, struct task_struct *p)
878 return task_current(rq, p);
881 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
885 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
887 #ifdef CONFIG_DEBUG_SPINLOCK
888 /* this is a valid case when another task releases the spinlock */
889 rq->lock.owner = current;
892 * If we are tracking spinlock dependencies then we have to
893 * fix up the runqueue lock - which gets 'carried over' from
896 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
898 raw_spin_unlock_irq(&rq->lock);
901 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
902 static inline int task_running(struct rq *rq, struct task_struct *p)
907 return task_current(rq, p);
911 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
915 * We can optimise this out completely for !SMP, because the
916 * SMP rebalancing from interrupt is the only thing that cares
921 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
922 raw_spin_unlock_irq(&rq->lock);
924 raw_spin_unlock(&rq->lock);
928 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
932 * After ->oncpu is cleared, the task can be moved to a different CPU.
933 * We must ensure this doesn't happen until the switch is completely
939 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
943 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
946 * __task_rq_lock - lock the runqueue a given task resides on.
947 * Must be called interrupts disabled.
949 static inline struct rq *__task_rq_lock(struct task_struct *p)
953 struct rq *rq = task_rq(p);
954 raw_spin_lock(&rq->lock);
955 if (likely(rq == task_rq(p)))
957 raw_spin_unlock(&rq->lock);
962 * task_rq_lock - lock the runqueue a given task resides on and disable
963 * interrupts. Note the ordering: we can safely lookup the task_rq without
964 * explicitly disabling preemption.
966 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
972 local_irq_save(*flags);
974 raw_spin_lock(&rq->lock);
975 if (likely(rq == task_rq(p)))
977 raw_spin_unlock_irqrestore(&rq->lock, *flags);
981 void task_rq_unlock_wait(struct task_struct *p)
983 struct rq *rq = task_rq(p);
985 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
986 raw_spin_unlock_wait(&rq->lock);
989 static void __task_rq_unlock(struct rq *rq)
992 raw_spin_unlock(&rq->lock);
995 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
998 raw_spin_unlock_irqrestore(&rq->lock, *flags);
1002 * this_rq_lock - lock this runqueue and disable interrupts.
1004 static struct rq *this_rq_lock(void)
1005 __acquires(rq->lock)
1009 local_irq_disable();
1011 raw_spin_lock(&rq->lock);
1016 #ifdef CONFIG_SCHED_HRTICK
1018 * Use HR-timers to deliver accurate preemption points.
1020 * Its all a bit involved since we cannot program an hrt while holding the
1021 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1024 * When we get rescheduled we reprogram the hrtick_timer outside of the
1030 * - enabled by features
1031 * - hrtimer is actually high res
1033 static inline int hrtick_enabled(struct rq *rq)
1035 if (!sched_feat(HRTICK))
1037 if (!cpu_active(cpu_of(rq)))
1039 return hrtimer_is_hres_active(&rq->hrtick_timer);
1042 static void hrtick_clear(struct rq *rq)
1044 if (hrtimer_active(&rq->hrtick_timer))
1045 hrtimer_cancel(&rq->hrtick_timer);
1049 * High-resolution timer tick.
1050 * Runs from hardirq context with interrupts disabled.
1052 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1054 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1056 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1058 raw_spin_lock(&rq->lock);
1059 update_rq_clock(rq);
1060 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1061 raw_spin_unlock(&rq->lock);
1063 return HRTIMER_NORESTART;
1068 * called from hardirq (IPI) context
1070 static void __hrtick_start(void *arg)
1072 struct rq *rq = arg;
1074 raw_spin_lock(&rq->lock);
1075 hrtimer_restart(&rq->hrtick_timer);
1076 rq->hrtick_csd_pending = 0;
1077 raw_spin_unlock(&rq->lock);
1081 * Called to set the hrtick timer state.
1083 * called with rq->lock held and irqs disabled
1085 static void hrtick_start(struct rq *rq, u64 delay)
1087 struct hrtimer *timer = &rq->hrtick_timer;
1088 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1090 hrtimer_set_expires(timer, time);
1092 if (rq == this_rq()) {
1093 hrtimer_restart(timer);
1094 } else if (!rq->hrtick_csd_pending) {
1095 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1096 rq->hrtick_csd_pending = 1;
1101 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1103 int cpu = (int)(long)hcpu;
1106 case CPU_UP_CANCELED:
1107 case CPU_UP_CANCELED_FROZEN:
1108 case CPU_DOWN_PREPARE:
1109 case CPU_DOWN_PREPARE_FROZEN:
1111 case CPU_DEAD_FROZEN:
1112 hrtick_clear(cpu_rq(cpu));
1119 static __init void init_hrtick(void)
1121 hotcpu_notifier(hotplug_hrtick, 0);
1125 * Called to set the hrtick timer state.
1127 * called with rq->lock held and irqs disabled
1129 static void hrtick_start(struct rq *rq, u64 delay)
1131 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1132 HRTIMER_MODE_REL_PINNED, 0);
1135 static inline void init_hrtick(void)
1138 #endif /* CONFIG_SMP */
1140 static void init_rq_hrtick(struct rq *rq)
1143 rq->hrtick_csd_pending = 0;
1145 rq->hrtick_csd.flags = 0;
1146 rq->hrtick_csd.func = __hrtick_start;
1147 rq->hrtick_csd.info = rq;
1150 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1151 rq->hrtick_timer.function = hrtick;
1153 #else /* CONFIG_SCHED_HRTICK */
1154 static inline void hrtick_clear(struct rq *rq)
1158 static inline void init_rq_hrtick(struct rq *rq)
1162 static inline void init_hrtick(void)
1165 #endif /* CONFIG_SCHED_HRTICK */
1168 * resched_task - mark a task 'to be rescheduled now'.
1170 * On UP this means the setting of the need_resched flag, on SMP it
1171 * might also involve a cross-CPU call to trigger the scheduler on
1176 #ifndef tsk_is_polling
1177 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1180 static void resched_task(struct task_struct *p)
1184 assert_raw_spin_locked(&task_rq(p)->lock);
1186 if (test_tsk_need_resched(p))
1189 set_tsk_need_resched(p);
1192 if (cpu == smp_processor_id())
1195 /* NEED_RESCHED must be visible before we test polling */
1197 if (!tsk_is_polling(p))
1198 smp_send_reschedule(cpu);
1201 static void resched_cpu(int cpu)
1203 struct rq *rq = cpu_rq(cpu);
1204 unsigned long flags;
1206 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1208 resched_task(cpu_curr(cpu));
1209 raw_spin_unlock_irqrestore(&rq->lock, flags);
1214 * When add_timer_on() enqueues a timer into the timer wheel of an
1215 * idle CPU then this timer might expire before the next timer event
1216 * which is scheduled to wake up that CPU. In case of a completely
1217 * idle system the next event might even be infinite time into the
1218 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1219 * leaves the inner idle loop so the newly added timer is taken into
1220 * account when the CPU goes back to idle and evaluates the timer
1221 * wheel for the next timer event.
1223 void wake_up_idle_cpu(int cpu)
1225 struct rq *rq = cpu_rq(cpu);
1227 if (cpu == smp_processor_id())
1231 * This is safe, as this function is called with the timer
1232 * wheel base lock of (cpu) held. When the CPU is on the way
1233 * to idle and has not yet set rq->curr to idle then it will
1234 * be serialized on the timer wheel base lock and take the new
1235 * timer into account automatically.
1237 if (rq->curr != rq->idle)
1241 * We can set TIF_RESCHED on the idle task of the other CPU
1242 * lockless. The worst case is that the other CPU runs the
1243 * idle task through an additional NOOP schedule()
1245 set_tsk_need_resched(rq->idle);
1247 /* NEED_RESCHED must be visible before we test polling */
1249 if (!tsk_is_polling(rq->idle))
1250 smp_send_reschedule(cpu);
1252 #endif /* CONFIG_NO_HZ */
1254 static u64 sched_avg_period(void)
1256 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1259 static void sched_avg_update(struct rq *rq)
1261 s64 period = sched_avg_period();
1263 while ((s64)(rq->clock - rq->age_stamp) > period) {
1264 rq->age_stamp += period;
1269 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1271 rq->rt_avg += rt_delta;
1272 sched_avg_update(rq);
1275 #else /* !CONFIG_SMP */
1276 static void resched_task(struct task_struct *p)
1278 assert_raw_spin_locked(&task_rq(p)->lock);
1279 set_tsk_need_resched(p);
1282 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1285 #endif /* CONFIG_SMP */
1287 #if BITS_PER_LONG == 32
1288 # define WMULT_CONST (~0UL)
1290 # define WMULT_CONST (1UL << 32)
1293 #define WMULT_SHIFT 32
1296 * Shift right and round:
1298 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1301 * delta *= weight / lw
1303 static unsigned long
1304 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1305 struct load_weight *lw)
1309 if (!lw->inv_weight) {
1310 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1313 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1317 tmp = (u64)delta_exec * weight;
1319 * Check whether we'd overflow the 64-bit multiplication:
1321 if (unlikely(tmp > WMULT_CONST))
1322 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1325 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1327 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1330 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1336 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1343 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1344 * of tasks with abnormal "nice" values across CPUs the contribution that
1345 * each task makes to its run queue's load is weighted according to its
1346 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1347 * scaled version of the new time slice allocation that they receive on time
1351 #define WEIGHT_IDLEPRIO 3
1352 #define WMULT_IDLEPRIO 1431655765
1355 * Nice levels are multiplicative, with a gentle 10% change for every
1356 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1357 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1358 * that remained on nice 0.
1360 * The "10% effect" is relative and cumulative: from _any_ nice level,
1361 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1362 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1363 * If a task goes up by ~10% and another task goes down by ~10% then
1364 * the relative distance between them is ~25%.)
1366 static const int prio_to_weight[40] = {
1367 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1368 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1369 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1370 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1371 /* 0 */ 1024, 820, 655, 526, 423,
1372 /* 5 */ 335, 272, 215, 172, 137,
1373 /* 10 */ 110, 87, 70, 56, 45,
1374 /* 15 */ 36, 29, 23, 18, 15,
1378 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1380 * In cases where the weight does not change often, we can use the
1381 * precalculated inverse to speed up arithmetics by turning divisions
1382 * into multiplications:
1384 static const u32 prio_to_wmult[40] = {
1385 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1386 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1387 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1388 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1389 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1390 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1391 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1392 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1395 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1398 * runqueue iterator, to support SMP load-balancing between different
1399 * scheduling classes, without having to expose their internal data
1400 * structures to the load-balancing proper:
1402 struct rq_iterator {
1404 struct task_struct *(*start)(void *);
1405 struct task_struct *(*next)(void *);
1409 static unsigned long
1410 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1411 unsigned long max_load_move, struct sched_domain *sd,
1412 enum cpu_idle_type idle, int *all_pinned,
1413 int *this_best_prio, struct rq_iterator *iterator);
1416 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1417 struct sched_domain *sd, enum cpu_idle_type idle,
1418 struct rq_iterator *iterator);
1421 /* Time spent by the tasks of the cpu accounting group executing in ... */
1422 enum cpuacct_stat_index {
1423 CPUACCT_STAT_USER, /* ... user mode */
1424 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1426 CPUACCT_STAT_NSTATS,
1429 #ifdef CONFIG_CGROUP_CPUACCT
1430 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1431 static void cpuacct_update_stats(struct task_struct *tsk,
1432 enum cpuacct_stat_index idx, cputime_t val);
1434 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1435 static inline void cpuacct_update_stats(struct task_struct *tsk,
1436 enum cpuacct_stat_index idx, cputime_t val) {}
1439 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1441 update_load_add(&rq->load, load);
1444 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1446 update_load_sub(&rq->load, load);
1449 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1450 typedef int (*tg_visitor)(struct task_group *, void *);
1453 * Iterate the full tree, calling @down when first entering a node and @up when
1454 * leaving it for the final time.
1456 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1458 struct task_group *parent, *child;
1462 parent = &root_task_group;
1464 ret = (*down)(parent, data);
1467 list_for_each_entry_rcu(child, &parent->children, siblings) {
1474 ret = (*up)(parent, data);
1479 parent = parent->parent;
1488 static int tg_nop(struct task_group *tg, void *data)
1495 /* Used instead of source_load when we know the type == 0 */
1496 static unsigned long weighted_cpuload(const int cpu)
1498 return cpu_rq(cpu)->load.weight;
1502 * Return a low guess at the load of a migration-source cpu weighted
1503 * according to the scheduling class and "nice" value.
1505 * We want to under-estimate the load of migration sources, to
1506 * balance conservatively.
1508 static unsigned long source_load(int cpu, int type)
1510 struct rq *rq = cpu_rq(cpu);
1511 unsigned long total = weighted_cpuload(cpu);
1513 if (type == 0 || !sched_feat(LB_BIAS))
1516 return min(rq->cpu_load[type-1], total);
1520 * Return a high guess at the load of a migration-target cpu weighted
1521 * according to the scheduling class and "nice" value.
1523 static unsigned long target_load(int cpu, int type)
1525 struct rq *rq = cpu_rq(cpu);
1526 unsigned long total = weighted_cpuload(cpu);
1528 if (type == 0 || !sched_feat(LB_BIAS))
1531 return max(rq->cpu_load[type-1], total);
1534 static struct sched_group *group_of(int cpu)
1536 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1544 static unsigned long power_of(int cpu)
1546 struct sched_group *group = group_of(cpu);
1549 return SCHED_LOAD_SCALE;
1551 return group->cpu_power;
1554 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1556 static unsigned long cpu_avg_load_per_task(int cpu)
1558 struct rq *rq = cpu_rq(cpu);
1559 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1562 rq->avg_load_per_task = rq->load.weight / nr_running;
1564 rq->avg_load_per_task = 0;
1566 return rq->avg_load_per_task;
1569 #ifdef CONFIG_FAIR_GROUP_SCHED
1571 static __read_mostly unsigned long *update_shares_data;
1573 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1576 * Calculate and set the cpu's group shares.
1578 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1579 unsigned long sd_shares,
1580 unsigned long sd_rq_weight,
1581 unsigned long *usd_rq_weight)
1583 unsigned long shares, rq_weight;
1586 rq_weight = usd_rq_weight[cpu];
1589 rq_weight = NICE_0_LOAD;
1593 * \Sum_j shares_j * rq_weight_i
1594 * shares_i = -----------------------------
1595 * \Sum_j rq_weight_j
1597 shares = (sd_shares * rq_weight) / sd_rq_weight;
1598 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1600 if (abs(shares - tg->se[cpu]->load.weight) >
1601 sysctl_sched_shares_thresh) {
1602 struct rq *rq = cpu_rq(cpu);
1603 unsigned long flags;
1605 raw_spin_lock_irqsave(&rq->lock, flags);
1606 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1607 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1608 __set_se_shares(tg->se[cpu], shares);
1609 raw_spin_unlock_irqrestore(&rq->lock, flags);
1614 * Re-compute the task group their per cpu shares over the given domain.
1615 * This needs to be done in a bottom-up fashion because the rq weight of a
1616 * parent group depends on the shares of its child groups.
1618 static int tg_shares_up(struct task_group *tg, void *data)
1620 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1621 unsigned long *usd_rq_weight;
1622 struct sched_domain *sd = data;
1623 unsigned long flags;
1629 local_irq_save(flags);
1630 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1632 for_each_cpu(i, sched_domain_span(sd)) {
1633 weight = tg->cfs_rq[i]->load.weight;
1634 usd_rq_weight[i] = weight;
1636 rq_weight += weight;
1638 * If there are currently no tasks on the cpu pretend there
1639 * is one of average load so that when a new task gets to
1640 * run here it will not get delayed by group starvation.
1643 weight = NICE_0_LOAD;
1645 sum_weight += weight;
1646 shares += tg->cfs_rq[i]->shares;
1650 rq_weight = sum_weight;
1652 if ((!shares && rq_weight) || shares > tg->shares)
1653 shares = tg->shares;
1655 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1656 shares = tg->shares;
1658 for_each_cpu(i, sched_domain_span(sd))
1659 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1661 local_irq_restore(flags);
1667 * Compute the cpu's hierarchical load factor for each task group.
1668 * This needs to be done in a top-down fashion because the load of a child
1669 * group is a fraction of its parents load.
1671 static int tg_load_down(struct task_group *tg, void *data)
1674 long cpu = (long)data;
1677 load = cpu_rq(cpu)->load.weight;
1679 load = tg->parent->cfs_rq[cpu]->h_load;
1680 load *= tg->cfs_rq[cpu]->shares;
1681 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1684 tg->cfs_rq[cpu]->h_load = load;
1689 static void update_shares(struct sched_domain *sd)
1694 if (root_task_group_empty())
1697 now = cpu_clock(raw_smp_processor_id());
1698 elapsed = now - sd->last_update;
1700 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1701 sd->last_update = now;
1702 walk_tg_tree(tg_nop, tg_shares_up, sd);
1706 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1708 if (root_task_group_empty())
1711 raw_spin_unlock(&rq->lock);
1713 raw_spin_lock(&rq->lock);
1716 static void update_h_load(long cpu)
1718 if (root_task_group_empty())
1721 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1726 static inline void update_shares(struct sched_domain *sd)
1730 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1736 #ifdef CONFIG_PREEMPT
1738 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1741 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1742 * way at the expense of forcing extra atomic operations in all
1743 * invocations. This assures that the double_lock is acquired using the
1744 * same underlying policy as the spinlock_t on this architecture, which
1745 * reduces latency compared to the unfair variant below. However, it
1746 * also adds more overhead and therefore may reduce throughput.
1748 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1749 __releases(this_rq->lock)
1750 __acquires(busiest->lock)
1751 __acquires(this_rq->lock)
1753 raw_spin_unlock(&this_rq->lock);
1754 double_rq_lock(this_rq, busiest);
1761 * Unfair double_lock_balance: Optimizes throughput at the expense of
1762 * latency by eliminating extra atomic operations when the locks are
1763 * already in proper order on entry. This favors lower cpu-ids and will
1764 * grant the double lock to lower cpus over higher ids under contention,
1765 * regardless of entry order into the function.
1767 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1768 __releases(this_rq->lock)
1769 __acquires(busiest->lock)
1770 __acquires(this_rq->lock)
1774 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1775 if (busiest < this_rq) {
1776 raw_spin_unlock(&this_rq->lock);
1777 raw_spin_lock(&busiest->lock);
1778 raw_spin_lock_nested(&this_rq->lock,
1779 SINGLE_DEPTH_NESTING);
1782 raw_spin_lock_nested(&busiest->lock,
1783 SINGLE_DEPTH_NESTING);
1788 #endif /* CONFIG_PREEMPT */
1791 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1793 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1795 if (unlikely(!irqs_disabled())) {
1796 /* printk() doesn't work good under rq->lock */
1797 raw_spin_unlock(&this_rq->lock);
1801 return _double_lock_balance(this_rq, busiest);
1804 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1805 __releases(busiest->lock)
1807 raw_spin_unlock(&busiest->lock);
1808 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1812 #ifdef CONFIG_FAIR_GROUP_SCHED
1813 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1816 cfs_rq->shares = shares;
1821 static void calc_load_account_active(struct rq *this_rq);
1822 static void update_sysctl(void);
1823 static int get_update_sysctl_factor(void);
1825 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1827 set_task_rq(p, cpu);
1830 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1831 * successfuly executed on another CPU. We must ensure that updates of
1832 * per-task data have been completed by this moment.
1835 task_thread_info(p)->cpu = cpu;
1839 #include "sched_stats.h"
1840 #include "sched_idletask.c"
1841 #include "sched_fair.c"
1842 #include "sched_rt.c"
1843 #ifdef CONFIG_SCHED_DEBUG
1844 # include "sched_debug.c"
1847 #define sched_class_highest (&rt_sched_class)
1848 #define for_each_class(class) \
1849 for (class = sched_class_highest; class; class = class->next)
1851 static void inc_nr_running(struct rq *rq)
1856 static void dec_nr_running(struct rq *rq)
1861 static void set_load_weight(struct task_struct *p)
1863 if (task_has_rt_policy(p)) {
1864 p->se.load.weight = prio_to_weight[0] * 2;
1865 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1870 * SCHED_IDLE tasks get minimal weight:
1872 if (p->policy == SCHED_IDLE) {
1873 p->se.load.weight = WEIGHT_IDLEPRIO;
1874 p->se.load.inv_weight = WMULT_IDLEPRIO;
1878 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1879 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1882 static void update_avg(u64 *avg, u64 sample)
1884 s64 diff = sample - *avg;
1888 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1891 p->se.start_runtime = p->se.sum_exec_runtime;
1893 sched_info_queued(p);
1894 p->sched_class->enqueue_task(rq, p, wakeup);
1898 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1901 if (p->se.last_wakeup) {
1902 update_avg(&p->se.avg_overlap,
1903 p->se.sum_exec_runtime - p->se.last_wakeup);
1904 p->se.last_wakeup = 0;
1906 update_avg(&p->se.avg_wakeup,
1907 sysctl_sched_wakeup_granularity);
1911 sched_info_dequeued(p);
1912 p->sched_class->dequeue_task(rq, p, sleep);
1917 * __normal_prio - return the priority that is based on the static prio
1919 static inline int __normal_prio(struct task_struct *p)
1921 return p->static_prio;
1925 * Calculate the expected normal priority: i.e. priority
1926 * without taking RT-inheritance into account. Might be
1927 * boosted by interactivity modifiers. Changes upon fork,
1928 * setprio syscalls, and whenever the interactivity
1929 * estimator recalculates.
1931 static inline int normal_prio(struct task_struct *p)
1935 if (task_has_rt_policy(p))
1936 prio = MAX_RT_PRIO-1 - p->rt_priority;
1938 prio = __normal_prio(p);
1943 * Calculate the current priority, i.e. the priority
1944 * taken into account by the scheduler. This value might
1945 * be boosted by RT tasks, or might be boosted by
1946 * interactivity modifiers. Will be RT if the task got
1947 * RT-boosted. If not then it returns p->normal_prio.
1949 static int effective_prio(struct task_struct *p)
1951 p->normal_prio = normal_prio(p);
1953 * If we are RT tasks or we were boosted to RT priority,
1954 * keep the priority unchanged. Otherwise, update priority
1955 * to the normal priority:
1957 if (!rt_prio(p->prio))
1958 return p->normal_prio;
1963 * activate_task - move a task to the runqueue.
1965 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1967 if (task_contributes_to_load(p))
1968 rq->nr_uninterruptible--;
1970 enqueue_task(rq, p, wakeup);
1975 * deactivate_task - remove a task from the runqueue.
1977 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1979 if (task_contributes_to_load(p))
1980 rq->nr_uninterruptible++;
1982 dequeue_task(rq, p, sleep);
1987 * task_curr - is this task currently executing on a CPU?
1988 * @p: the task in question.
1990 inline int task_curr(const struct task_struct *p)
1992 return cpu_curr(task_cpu(p)) == p;
1995 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1996 const struct sched_class *prev_class,
1997 int oldprio, int running)
1999 if (prev_class != p->sched_class) {
2000 if (prev_class->switched_from)
2001 prev_class->switched_from(rq, p, running);
2002 p->sched_class->switched_to(rq, p, running);
2004 p->sched_class->prio_changed(rq, p, oldprio, running);
2009 * Is this task likely cache-hot:
2012 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2016 if (p->sched_class != &fair_sched_class)
2020 * Buddy candidates are cache hot:
2022 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2023 (&p->se == cfs_rq_of(&p->se)->next ||
2024 &p->se == cfs_rq_of(&p->se)->last))
2027 if (sysctl_sched_migration_cost == -1)
2029 if (sysctl_sched_migration_cost == 0)
2032 delta = now - p->se.exec_start;
2034 return delta < (s64)sysctl_sched_migration_cost;
2038 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2040 int old_cpu = task_cpu(p);
2041 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2042 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2044 #ifdef CONFIG_SCHED_DEBUG
2046 * We should never call set_task_cpu() on a blocked task,
2047 * ttwu() will sort out the placement.
2049 WARN_ON(p->state != TASK_RUNNING && p->state != TASK_WAKING);
2052 trace_sched_migrate_task(p, new_cpu);
2054 if (old_cpu != new_cpu) {
2055 p->se.nr_migrations++;
2056 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2059 p->se.vruntime -= old_cfsrq->min_vruntime -
2060 new_cfsrq->min_vruntime;
2062 __set_task_cpu(p, new_cpu);
2065 struct migration_req {
2066 struct list_head list;
2068 struct task_struct *task;
2071 struct completion done;
2075 * The task's runqueue lock must be held.
2076 * Returns true if you have to wait for migration thread.
2079 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2081 struct rq *rq = task_rq(p);
2084 * If the task is not on a runqueue (and not running), then
2085 * the next wake-up will properly place the task.
2087 if (!p->se.on_rq && !task_running(rq, p))
2090 init_completion(&req->done);
2092 req->dest_cpu = dest_cpu;
2093 list_add(&req->list, &rq->migration_queue);
2099 * wait_task_context_switch - wait for a thread to complete at least one
2102 * @p must not be current.
2104 void wait_task_context_switch(struct task_struct *p)
2106 unsigned long nvcsw, nivcsw, flags;
2114 * The runqueue is assigned before the actual context
2115 * switch. We need to take the runqueue lock.
2117 * We could check initially without the lock but it is
2118 * very likely that we need to take the lock in every
2121 rq = task_rq_lock(p, &flags);
2122 running = task_running(rq, p);
2123 task_rq_unlock(rq, &flags);
2125 if (likely(!running))
2128 * The switch count is incremented before the actual
2129 * context switch. We thus wait for two switches to be
2130 * sure at least one completed.
2132 if ((p->nvcsw - nvcsw) > 1)
2134 if ((p->nivcsw - nivcsw) > 1)
2142 * wait_task_inactive - wait for a thread to unschedule.
2144 * If @match_state is nonzero, it's the @p->state value just checked and
2145 * not expected to change. If it changes, i.e. @p might have woken up,
2146 * then return zero. When we succeed in waiting for @p to be off its CPU,
2147 * we return a positive number (its total switch count). If a second call
2148 * a short while later returns the same number, the caller can be sure that
2149 * @p has remained unscheduled the whole time.
2151 * The caller must ensure that the task *will* unschedule sometime soon,
2152 * else this function might spin for a *long* time. This function can't
2153 * be called with interrupts off, or it may introduce deadlock with
2154 * smp_call_function() if an IPI is sent by the same process we are
2155 * waiting to become inactive.
2157 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2159 unsigned long flags;
2166 * We do the initial early heuristics without holding
2167 * any task-queue locks at all. We'll only try to get
2168 * the runqueue lock when things look like they will
2174 * If the task is actively running on another CPU
2175 * still, just relax and busy-wait without holding
2178 * NOTE! Since we don't hold any locks, it's not
2179 * even sure that "rq" stays as the right runqueue!
2180 * But we don't care, since "task_running()" will
2181 * return false if the runqueue has changed and p
2182 * is actually now running somewhere else!
2184 while (task_running(rq, p)) {
2185 if (match_state && unlikely(p->state != match_state))
2191 * Ok, time to look more closely! We need the rq
2192 * lock now, to be *sure*. If we're wrong, we'll
2193 * just go back and repeat.
2195 rq = task_rq_lock(p, &flags);
2196 trace_sched_wait_task(rq, p);
2197 running = task_running(rq, p);
2198 on_rq = p->se.on_rq;
2200 if (!match_state || p->state == match_state)
2201 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2202 task_rq_unlock(rq, &flags);
2205 * If it changed from the expected state, bail out now.
2207 if (unlikely(!ncsw))
2211 * Was it really running after all now that we
2212 * checked with the proper locks actually held?
2214 * Oops. Go back and try again..
2216 if (unlikely(running)) {
2222 * It's not enough that it's not actively running,
2223 * it must be off the runqueue _entirely_, and not
2226 * So if it was still runnable (but just not actively
2227 * running right now), it's preempted, and we should
2228 * yield - it could be a while.
2230 if (unlikely(on_rq)) {
2231 schedule_timeout_uninterruptible(1);
2236 * Ahh, all good. It wasn't running, and it wasn't
2237 * runnable, which means that it will never become
2238 * running in the future either. We're all done!
2247 * kick_process - kick a running thread to enter/exit the kernel
2248 * @p: the to-be-kicked thread
2250 * Cause a process which is running on another CPU to enter
2251 * kernel-mode, without any delay. (to get signals handled.)
2253 * NOTE: this function doesnt have to take the runqueue lock,
2254 * because all it wants to ensure is that the remote task enters
2255 * the kernel. If the IPI races and the task has been migrated
2256 * to another CPU then no harm is done and the purpose has been
2259 void kick_process(struct task_struct *p)
2265 if ((cpu != smp_processor_id()) && task_curr(p))
2266 smp_send_reschedule(cpu);
2269 EXPORT_SYMBOL_GPL(kick_process);
2270 #endif /* CONFIG_SMP */
2273 * task_oncpu_function_call - call a function on the cpu on which a task runs
2274 * @p: the task to evaluate
2275 * @func: the function to be called
2276 * @info: the function call argument
2278 * Calls the function @func when the task is currently running. This might
2279 * be on the current CPU, which just calls the function directly
2281 void task_oncpu_function_call(struct task_struct *p,
2282 void (*func) (void *info), void *info)
2289 smp_call_function_single(cpu, func, info, 1);
2294 static int select_fallback_rq(int cpu, struct task_struct *p)
2297 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2299 /* Look for allowed, online CPU in same node. */
2300 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2301 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2304 /* Any allowed, online CPU? */
2305 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2306 if (dest_cpu < nr_cpu_ids)
2309 /* No more Mr. Nice Guy. */
2310 if (dest_cpu >= nr_cpu_ids) {
2312 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2314 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2317 * Don't tell them about moving exiting tasks or
2318 * kernel threads (both mm NULL), since they never
2321 if (p->mm && printk_ratelimit()) {
2322 printk(KERN_INFO "process %d (%s) no "
2323 "longer affine to cpu%d\n",
2324 task_pid_nr(p), p->comm, cpu);
2334 * - fork, @p is stable because it isn't on the tasklist yet
2336 * - exec, @p is unstable, retry loop
2338 * - wake-up, we serialize ->cpus_allowed against TASK_WAKING so
2339 * we should be good.
2342 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2344 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2347 * In order not to call set_task_cpu() on a blocking task we need
2348 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2351 * Since this is common to all placement strategies, this lives here.
2353 * [ this allows ->select_task() to simply return task_cpu(p) and
2354 * not worry about this generic constraint ]
2356 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2358 cpu = select_fallback_rq(task_cpu(p), p);
2365 * try_to_wake_up - wake up a thread
2366 * @p: the to-be-woken-up thread
2367 * @state: the mask of task states that can be woken
2368 * @sync: do a synchronous wakeup?
2370 * Put it on the run-queue if it's not already there. The "current"
2371 * thread is always on the run-queue (except when the actual
2372 * re-schedule is in progress), and as such you're allowed to do
2373 * the simpler "current->state = TASK_RUNNING" to mark yourself
2374 * runnable without the overhead of this.
2376 * returns failure only if the task is already active.
2378 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2381 int cpu, orig_cpu, this_cpu, success = 0;
2382 unsigned long flags;
2383 struct rq *rq, *orig_rq;
2385 if (!sched_feat(SYNC_WAKEUPS))
2386 wake_flags &= ~WF_SYNC;
2388 this_cpu = get_cpu();
2391 rq = orig_rq = task_rq_lock(p, &flags);
2392 update_rq_clock(rq);
2393 if (!(p->state & state))
2403 if (unlikely(task_running(rq, p)))
2407 * In order to handle concurrent wakeups and release the rq->lock
2408 * we put the task in TASK_WAKING state.
2410 * First fix up the nr_uninterruptible count:
2412 if (task_contributes_to_load(p))
2413 rq->nr_uninterruptible--;
2414 p->state = TASK_WAKING;
2416 if (p->sched_class->task_waking)
2417 p->sched_class->task_waking(rq, p);
2419 __task_rq_unlock(rq);
2421 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2422 if (cpu != orig_cpu)
2423 set_task_cpu(p, cpu);
2425 rq = __task_rq_lock(p);
2426 update_rq_clock(rq);
2428 WARN_ON(p->state != TASK_WAKING);
2431 #ifdef CONFIG_SCHEDSTATS
2432 schedstat_inc(rq, ttwu_count);
2433 if (cpu == this_cpu)
2434 schedstat_inc(rq, ttwu_local);
2436 struct sched_domain *sd;
2437 for_each_domain(this_cpu, sd) {
2438 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2439 schedstat_inc(sd, ttwu_wake_remote);
2444 #endif /* CONFIG_SCHEDSTATS */
2447 #endif /* CONFIG_SMP */
2448 schedstat_inc(p, se.nr_wakeups);
2449 if (wake_flags & WF_SYNC)
2450 schedstat_inc(p, se.nr_wakeups_sync);
2451 if (orig_cpu != cpu)
2452 schedstat_inc(p, se.nr_wakeups_migrate);
2453 if (cpu == this_cpu)
2454 schedstat_inc(p, se.nr_wakeups_local);
2456 schedstat_inc(p, se.nr_wakeups_remote);
2457 activate_task(rq, p, 1);
2461 * Only attribute actual wakeups done by this task.
2463 if (!in_interrupt()) {
2464 struct sched_entity *se = ¤t->se;
2465 u64 sample = se->sum_exec_runtime;
2467 if (se->last_wakeup)
2468 sample -= se->last_wakeup;
2470 sample -= se->start_runtime;
2471 update_avg(&se->avg_wakeup, sample);
2473 se->last_wakeup = se->sum_exec_runtime;
2477 trace_sched_wakeup(rq, p, success);
2478 check_preempt_curr(rq, p, wake_flags);
2480 p->state = TASK_RUNNING;
2482 if (p->sched_class->task_woken)
2483 p->sched_class->task_woken(rq, p);
2485 if (unlikely(rq->idle_stamp)) {
2486 u64 delta = rq->clock - rq->idle_stamp;
2487 u64 max = 2*sysctl_sched_migration_cost;
2492 update_avg(&rq->avg_idle, delta);
2497 task_rq_unlock(rq, &flags);
2504 * wake_up_process - Wake up a specific process
2505 * @p: The process to be woken up.
2507 * Attempt to wake up the nominated process and move it to the set of runnable
2508 * processes. Returns 1 if the process was woken up, 0 if it was already
2511 * It may be assumed that this function implies a write memory barrier before
2512 * changing the task state if and only if any tasks are woken up.
2514 int wake_up_process(struct task_struct *p)
2516 return try_to_wake_up(p, TASK_ALL, 0);
2518 EXPORT_SYMBOL(wake_up_process);
2520 int wake_up_state(struct task_struct *p, unsigned int state)
2522 return try_to_wake_up(p, state, 0);
2526 * Perform scheduler related setup for a newly forked process p.
2527 * p is forked by current.
2529 * __sched_fork() is basic setup used by init_idle() too:
2531 static void __sched_fork(struct task_struct *p)
2533 p->se.exec_start = 0;
2534 p->se.sum_exec_runtime = 0;
2535 p->se.prev_sum_exec_runtime = 0;
2536 p->se.nr_migrations = 0;
2537 p->se.last_wakeup = 0;
2538 p->se.avg_overlap = 0;
2539 p->se.start_runtime = 0;
2540 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2542 #ifdef CONFIG_SCHEDSTATS
2543 p->se.wait_start = 0;
2545 p->se.wait_count = 0;
2548 p->se.sleep_start = 0;
2549 p->se.sleep_max = 0;
2550 p->se.sum_sleep_runtime = 0;
2552 p->se.block_start = 0;
2553 p->se.block_max = 0;
2555 p->se.slice_max = 0;
2557 p->se.nr_migrations_cold = 0;
2558 p->se.nr_failed_migrations_affine = 0;
2559 p->se.nr_failed_migrations_running = 0;
2560 p->se.nr_failed_migrations_hot = 0;
2561 p->se.nr_forced_migrations = 0;
2563 p->se.nr_wakeups = 0;
2564 p->se.nr_wakeups_sync = 0;
2565 p->se.nr_wakeups_migrate = 0;
2566 p->se.nr_wakeups_local = 0;
2567 p->se.nr_wakeups_remote = 0;
2568 p->se.nr_wakeups_affine = 0;
2569 p->se.nr_wakeups_affine_attempts = 0;
2570 p->se.nr_wakeups_passive = 0;
2571 p->se.nr_wakeups_idle = 0;
2575 INIT_LIST_HEAD(&p->rt.run_list);
2577 INIT_LIST_HEAD(&p->se.group_node);
2579 #ifdef CONFIG_PREEMPT_NOTIFIERS
2580 INIT_HLIST_HEAD(&p->preempt_notifiers);
2585 * fork()/clone()-time setup:
2587 void sched_fork(struct task_struct *p, int clone_flags)
2589 int cpu = get_cpu();
2593 * We mark the process as waking here. This guarantees that
2594 * nobody will actually run it, and a signal or other external
2595 * event cannot wake it up and insert it on the runqueue either.
2597 p->state = TASK_WAKING;
2600 * Revert to default priority/policy on fork if requested.
2602 if (unlikely(p->sched_reset_on_fork)) {
2603 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2604 p->policy = SCHED_NORMAL;
2605 p->normal_prio = p->static_prio;
2608 if (PRIO_TO_NICE(p->static_prio) < 0) {
2609 p->static_prio = NICE_TO_PRIO(0);
2610 p->normal_prio = p->static_prio;
2615 * We don't need the reset flag anymore after the fork. It has
2616 * fulfilled its duty:
2618 p->sched_reset_on_fork = 0;
2622 * Make sure we do not leak PI boosting priority to the child.
2624 p->prio = current->normal_prio;
2626 if (!rt_prio(p->prio))
2627 p->sched_class = &fair_sched_class;
2629 if (p->sched_class->task_fork)
2630 p->sched_class->task_fork(p);
2633 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2635 set_task_cpu(p, cpu);
2637 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2638 if (likely(sched_info_on()))
2639 memset(&p->sched_info, 0, sizeof(p->sched_info));
2641 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2644 #ifdef CONFIG_PREEMPT
2645 /* Want to start with kernel preemption disabled. */
2646 task_thread_info(p)->preempt_count = 1;
2648 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2654 * wake_up_new_task - wake up a newly created task for the first time.
2656 * This function will do some initial scheduler statistics housekeeping
2657 * that must be done for every newly created context, then puts the task
2658 * on the runqueue and wakes it.
2660 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2662 unsigned long flags;
2665 rq = task_rq_lock(p, &flags);
2666 BUG_ON(p->state != TASK_WAKING);
2667 p->state = TASK_RUNNING;
2668 update_rq_clock(rq);
2669 activate_task(rq, p, 0);
2670 trace_sched_wakeup_new(rq, p, 1);
2671 check_preempt_curr(rq, p, WF_FORK);
2673 if (p->sched_class->task_woken)
2674 p->sched_class->task_woken(rq, p);
2676 task_rq_unlock(rq, &flags);
2679 #ifdef CONFIG_PREEMPT_NOTIFIERS
2682 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2683 * @notifier: notifier struct to register
2685 void preempt_notifier_register(struct preempt_notifier *notifier)
2687 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2689 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2692 * preempt_notifier_unregister - no longer interested in preemption notifications
2693 * @notifier: notifier struct to unregister
2695 * This is safe to call from within a preemption notifier.
2697 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2699 hlist_del(¬ifier->link);
2701 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2703 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2705 struct preempt_notifier *notifier;
2706 struct hlist_node *node;
2708 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2709 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2713 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2714 struct task_struct *next)
2716 struct preempt_notifier *notifier;
2717 struct hlist_node *node;
2719 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2720 notifier->ops->sched_out(notifier, next);
2723 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2725 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2730 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2731 struct task_struct *next)
2735 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2738 * prepare_task_switch - prepare to switch tasks
2739 * @rq: the runqueue preparing to switch
2740 * @prev: the current task that is being switched out
2741 * @next: the task we are going to switch to.
2743 * This is called with the rq lock held and interrupts off. It must
2744 * be paired with a subsequent finish_task_switch after the context
2747 * prepare_task_switch sets up locking and calls architecture specific
2751 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2752 struct task_struct *next)
2754 fire_sched_out_preempt_notifiers(prev, next);
2755 prepare_lock_switch(rq, next);
2756 prepare_arch_switch(next);
2760 * finish_task_switch - clean up after a task-switch
2761 * @rq: runqueue associated with task-switch
2762 * @prev: the thread we just switched away from.
2764 * finish_task_switch must be called after the context switch, paired
2765 * with a prepare_task_switch call before the context switch.
2766 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2767 * and do any other architecture-specific cleanup actions.
2769 * Note that we may have delayed dropping an mm in context_switch(). If
2770 * so, we finish that here outside of the runqueue lock. (Doing it
2771 * with the lock held can cause deadlocks; see schedule() for
2774 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2775 __releases(rq->lock)
2777 struct mm_struct *mm = rq->prev_mm;
2783 * A task struct has one reference for the use as "current".
2784 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2785 * schedule one last time. The schedule call will never return, and
2786 * the scheduled task must drop that reference.
2787 * The test for TASK_DEAD must occur while the runqueue locks are
2788 * still held, otherwise prev could be scheduled on another cpu, die
2789 * there before we look at prev->state, and then the reference would
2791 * Manfred Spraul <manfred@colorfullife.com>
2793 prev_state = prev->state;
2794 finish_arch_switch(prev);
2795 perf_event_task_sched_in(current, cpu_of(rq));
2796 finish_lock_switch(rq, prev);
2798 fire_sched_in_preempt_notifiers(current);
2801 if (unlikely(prev_state == TASK_DEAD)) {
2803 * Remove function-return probe instances associated with this
2804 * task and put them back on the free list.
2806 kprobe_flush_task(prev);
2807 put_task_struct(prev);
2813 /* assumes rq->lock is held */
2814 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2816 if (prev->sched_class->pre_schedule)
2817 prev->sched_class->pre_schedule(rq, prev);
2820 /* rq->lock is NOT held, but preemption is disabled */
2821 static inline void post_schedule(struct rq *rq)
2823 if (rq->post_schedule) {
2824 unsigned long flags;
2826 raw_spin_lock_irqsave(&rq->lock, flags);
2827 if (rq->curr->sched_class->post_schedule)
2828 rq->curr->sched_class->post_schedule(rq);
2829 raw_spin_unlock_irqrestore(&rq->lock, flags);
2831 rq->post_schedule = 0;
2837 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2841 static inline void post_schedule(struct rq *rq)
2848 * schedule_tail - first thing a freshly forked thread must call.
2849 * @prev: the thread we just switched away from.
2851 asmlinkage void schedule_tail(struct task_struct *prev)
2852 __releases(rq->lock)
2854 struct rq *rq = this_rq();
2856 finish_task_switch(rq, prev);
2859 * FIXME: do we need to worry about rq being invalidated by the
2864 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2865 /* In this case, finish_task_switch does not reenable preemption */
2868 if (current->set_child_tid)
2869 put_user(task_pid_vnr(current), current->set_child_tid);
2873 * context_switch - switch to the new MM and the new
2874 * thread's register state.
2877 context_switch(struct rq *rq, struct task_struct *prev,
2878 struct task_struct *next)
2880 struct mm_struct *mm, *oldmm;
2882 prepare_task_switch(rq, prev, next);
2883 trace_sched_switch(rq, prev, next);
2885 oldmm = prev->active_mm;
2887 * For paravirt, this is coupled with an exit in switch_to to
2888 * combine the page table reload and the switch backend into
2891 arch_start_context_switch(prev);
2894 next->active_mm = oldmm;
2895 atomic_inc(&oldmm->mm_count);
2896 enter_lazy_tlb(oldmm, next);
2898 switch_mm(oldmm, mm, next);
2900 if (likely(!prev->mm)) {
2901 prev->active_mm = NULL;
2902 rq->prev_mm = oldmm;
2905 * Since the runqueue lock will be released by the next
2906 * task (which is an invalid locking op but in the case
2907 * of the scheduler it's an obvious special-case), so we
2908 * do an early lockdep release here:
2910 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2911 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2914 /* Here we just switch the register state and the stack. */
2915 switch_to(prev, next, prev);
2919 * this_rq must be evaluated again because prev may have moved
2920 * CPUs since it called schedule(), thus the 'rq' on its stack
2921 * frame will be invalid.
2923 finish_task_switch(this_rq(), prev);
2927 * nr_running, nr_uninterruptible and nr_context_switches:
2929 * externally visible scheduler statistics: current number of runnable
2930 * threads, current number of uninterruptible-sleeping threads, total
2931 * number of context switches performed since bootup.
2933 unsigned long nr_running(void)
2935 unsigned long i, sum = 0;
2937 for_each_online_cpu(i)
2938 sum += cpu_rq(i)->nr_running;
2943 unsigned long nr_uninterruptible(void)
2945 unsigned long i, sum = 0;
2947 for_each_possible_cpu(i)
2948 sum += cpu_rq(i)->nr_uninterruptible;
2951 * Since we read the counters lockless, it might be slightly
2952 * inaccurate. Do not allow it to go below zero though:
2954 if (unlikely((long)sum < 0))
2960 unsigned long long nr_context_switches(void)
2963 unsigned long long sum = 0;
2965 for_each_possible_cpu(i)
2966 sum += cpu_rq(i)->nr_switches;
2971 unsigned long nr_iowait(void)
2973 unsigned long i, sum = 0;
2975 for_each_possible_cpu(i)
2976 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2981 unsigned long nr_iowait_cpu(void)
2983 struct rq *this = this_rq();
2984 return atomic_read(&this->nr_iowait);
2987 unsigned long this_cpu_load(void)
2989 struct rq *this = this_rq();
2990 return this->cpu_load[0];
2994 /* Variables and functions for calc_load */
2995 static atomic_long_t calc_load_tasks;
2996 static unsigned long calc_load_update;
2997 unsigned long avenrun[3];
2998 EXPORT_SYMBOL(avenrun);
3001 * get_avenrun - get the load average array
3002 * @loads: pointer to dest load array
3003 * @offset: offset to add
3004 * @shift: shift count to shift the result left
3006 * These values are estimates at best, so no need for locking.
3008 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3010 loads[0] = (avenrun[0] + offset) << shift;
3011 loads[1] = (avenrun[1] + offset) << shift;
3012 loads[2] = (avenrun[2] + offset) << shift;
3015 static unsigned long
3016 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3019 load += active * (FIXED_1 - exp);
3020 return load >> FSHIFT;
3024 * calc_load - update the avenrun load estimates 10 ticks after the
3025 * CPUs have updated calc_load_tasks.
3027 void calc_global_load(void)
3029 unsigned long upd = calc_load_update + 10;
3032 if (time_before(jiffies, upd))
3035 active = atomic_long_read(&calc_load_tasks);
3036 active = active > 0 ? active * FIXED_1 : 0;
3038 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3039 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3040 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3042 calc_load_update += LOAD_FREQ;
3046 * Either called from update_cpu_load() or from a cpu going idle
3048 static void calc_load_account_active(struct rq *this_rq)
3050 long nr_active, delta;
3052 nr_active = this_rq->nr_running;
3053 nr_active += (long) this_rq->nr_uninterruptible;
3055 if (nr_active != this_rq->calc_load_active) {
3056 delta = nr_active - this_rq->calc_load_active;
3057 this_rq->calc_load_active = nr_active;
3058 atomic_long_add(delta, &calc_load_tasks);
3063 * Update rq->cpu_load[] statistics. This function is usually called every
3064 * scheduler tick (TICK_NSEC).
3066 static void update_cpu_load(struct rq *this_rq)
3068 unsigned long this_load = this_rq->load.weight;
3071 this_rq->nr_load_updates++;
3073 /* Update our load: */
3074 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3075 unsigned long old_load, new_load;
3077 /* scale is effectively 1 << i now, and >> i divides by scale */
3079 old_load = this_rq->cpu_load[i];
3080 new_load = this_load;
3082 * Round up the averaging division if load is increasing. This
3083 * prevents us from getting stuck on 9 if the load is 10, for
3086 if (new_load > old_load)
3087 new_load += scale-1;
3088 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3091 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3092 this_rq->calc_load_update += LOAD_FREQ;
3093 calc_load_account_active(this_rq);
3100 * double_rq_lock - safely lock two runqueues
3102 * Note this does not disable interrupts like task_rq_lock,
3103 * you need to do so manually before calling.
3105 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3106 __acquires(rq1->lock)
3107 __acquires(rq2->lock)
3109 BUG_ON(!irqs_disabled());
3111 raw_spin_lock(&rq1->lock);
3112 __acquire(rq2->lock); /* Fake it out ;) */
3115 raw_spin_lock(&rq1->lock);
3116 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3118 raw_spin_lock(&rq2->lock);
3119 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3122 update_rq_clock(rq1);
3123 update_rq_clock(rq2);
3127 * double_rq_unlock - safely unlock two runqueues
3129 * Note this does not restore interrupts like task_rq_unlock,
3130 * you need to do so manually after calling.
3132 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3133 __releases(rq1->lock)
3134 __releases(rq2->lock)
3136 raw_spin_unlock(&rq1->lock);
3138 raw_spin_unlock(&rq2->lock);
3140 __release(rq2->lock);
3144 * sched_exec - execve() is a valuable balancing opportunity, because at
3145 * this point the task has the smallest effective memory and cache footprint.
3147 void sched_exec(void)
3149 struct task_struct *p = current;
3150 struct migration_req req;
3151 int dest_cpu, this_cpu;
3152 unsigned long flags;
3156 this_cpu = get_cpu();
3157 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3158 if (dest_cpu == this_cpu) {
3163 rq = task_rq_lock(p, &flags);
3167 * select_task_rq() can race against ->cpus_allowed
3169 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3170 || unlikely(!cpu_active(dest_cpu))) {
3171 task_rq_unlock(rq, &flags);
3175 /* force the process onto the specified CPU */
3176 if (migrate_task(p, dest_cpu, &req)) {
3177 /* Need to wait for migration thread (might exit: take ref). */
3178 struct task_struct *mt = rq->migration_thread;
3180 get_task_struct(mt);
3181 task_rq_unlock(rq, &flags);
3182 wake_up_process(mt);
3183 put_task_struct(mt);
3184 wait_for_completion(&req.done);
3188 task_rq_unlock(rq, &flags);
3192 * pull_task - move a task from a remote runqueue to the local runqueue.
3193 * Both runqueues must be locked.
3195 static void pull_task(struct rq *src_rq, struct task_struct *p,
3196 struct rq *this_rq, int this_cpu)
3198 deactivate_task(src_rq, p, 0);
3199 set_task_cpu(p, this_cpu);
3200 activate_task(this_rq, p, 0);
3201 check_preempt_curr(this_rq, p, 0);
3205 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3208 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3209 struct sched_domain *sd, enum cpu_idle_type idle,
3212 int tsk_cache_hot = 0;
3214 * We do not migrate tasks that are:
3215 * 1) running (obviously), or
3216 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3217 * 3) are cache-hot on their current CPU.
3219 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3220 schedstat_inc(p, se.nr_failed_migrations_affine);
3225 if (task_running(rq, p)) {
3226 schedstat_inc(p, se.nr_failed_migrations_running);
3231 * Aggressive migration if:
3232 * 1) task is cache cold, or
3233 * 2) too many balance attempts have failed.
3236 tsk_cache_hot = task_hot(p, rq->clock, sd);
3237 if (!tsk_cache_hot ||
3238 sd->nr_balance_failed > sd->cache_nice_tries) {
3239 #ifdef CONFIG_SCHEDSTATS
3240 if (tsk_cache_hot) {
3241 schedstat_inc(sd, lb_hot_gained[idle]);
3242 schedstat_inc(p, se.nr_forced_migrations);
3248 if (tsk_cache_hot) {
3249 schedstat_inc(p, se.nr_failed_migrations_hot);
3255 static unsigned long
3256 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3257 unsigned long max_load_move, struct sched_domain *sd,
3258 enum cpu_idle_type idle, int *all_pinned,
3259 int *this_best_prio, struct rq_iterator *iterator)
3261 int loops = 0, pulled = 0, pinned = 0;
3262 struct task_struct *p;
3263 long rem_load_move = max_load_move;
3265 if (max_load_move == 0)
3271 * Start the load-balancing iterator:
3273 p = iterator->start(iterator->arg);
3275 if (!p || loops++ > sysctl_sched_nr_migrate)
3278 if ((p->se.load.weight >> 1) > rem_load_move ||
3279 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3280 p = iterator->next(iterator->arg);
3284 pull_task(busiest, p, this_rq, this_cpu);
3286 rem_load_move -= p->se.load.weight;
3288 #ifdef CONFIG_PREEMPT
3290 * NEWIDLE balancing is a source of latency, so preemptible kernels
3291 * will stop after the first task is pulled to minimize the critical
3294 if (idle == CPU_NEWLY_IDLE)
3299 * We only want to steal up to the prescribed amount of weighted load.
3301 if (rem_load_move > 0) {
3302 if (p->prio < *this_best_prio)
3303 *this_best_prio = p->prio;
3304 p = iterator->next(iterator->arg);
3309 * Right now, this is one of only two places pull_task() is called,
3310 * so we can safely collect pull_task() stats here rather than
3311 * inside pull_task().
3313 schedstat_add(sd, lb_gained[idle], pulled);
3316 *all_pinned = pinned;
3318 return max_load_move - rem_load_move;
3322 * move_tasks tries to move up to max_load_move weighted load from busiest to
3323 * this_rq, as part of a balancing operation within domain "sd".
3324 * Returns 1 if successful and 0 otherwise.
3326 * Called with both runqueues locked.
3328 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3329 unsigned long max_load_move,
3330 struct sched_domain *sd, enum cpu_idle_type idle,
3333 const struct sched_class *class = sched_class_highest;
3334 unsigned long total_load_moved = 0;
3335 int this_best_prio = this_rq->curr->prio;
3339 class->load_balance(this_rq, this_cpu, busiest,
3340 max_load_move - total_load_moved,
3341 sd, idle, all_pinned, &this_best_prio);
3342 class = class->next;
3344 #ifdef CONFIG_PREEMPT
3346 * NEWIDLE balancing is a source of latency, so preemptible
3347 * kernels will stop after the first task is pulled to minimize
3348 * the critical section.
3350 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3353 } while (class && max_load_move > total_load_moved);
3355 return total_load_moved > 0;
3359 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3360 struct sched_domain *sd, enum cpu_idle_type idle,
3361 struct rq_iterator *iterator)
3363 struct task_struct *p = iterator->start(iterator->arg);
3367 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3368 pull_task(busiest, p, this_rq, this_cpu);
3370 * Right now, this is only the second place pull_task()
3371 * is called, so we can safely collect pull_task()
3372 * stats here rather than inside pull_task().
3374 schedstat_inc(sd, lb_gained[idle]);
3378 p = iterator->next(iterator->arg);
3385 * move_one_task tries to move exactly one task from busiest to this_rq, as
3386 * part of active balancing operations within "domain".
3387 * Returns 1 if successful and 0 otherwise.
3389 * Called with both runqueues locked.
3391 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3392 struct sched_domain *sd, enum cpu_idle_type idle)
3394 const struct sched_class *class;
3396 for_each_class(class) {
3397 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3403 /********** Helpers for find_busiest_group ************************/
3405 * sd_lb_stats - Structure to store the statistics of a sched_domain
3406 * during load balancing.
3408 struct sd_lb_stats {
3409 struct sched_group *busiest; /* Busiest group in this sd */
3410 struct sched_group *this; /* Local group in this sd */
3411 unsigned long total_load; /* Total load of all groups in sd */
3412 unsigned long total_pwr; /* Total power of all groups in sd */
3413 unsigned long avg_load; /* Average load across all groups in sd */
3415 /** Statistics of this group */
3416 unsigned long this_load;
3417 unsigned long this_load_per_task;
3418 unsigned long this_nr_running;
3420 /* Statistics of the busiest group */
3421 unsigned long max_load;
3422 unsigned long busiest_load_per_task;
3423 unsigned long busiest_nr_running;
3425 int group_imb; /* Is there imbalance in this sd */
3426 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3427 int power_savings_balance; /* Is powersave balance needed for this sd */
3428 struct sched_group *group_min; /* Least loaded group in sd */
3429 struct sched_group *group_leader; /* Group which relieves group_min */
3430 unsigned long min_load_per_task; /* load_per_task in group_min */
3431 unsigned long leader_nr_running; /* Nr running of group_leader */
3432 unsigned long min_nr_running; /* Nr running of group_min */
3437 * sg_lb_stats - stats of a sched_group required for load_balancing
3439 struct sg_lb_stats {
3440 unsigned long avg_load; /*Avg load across the CPUs of the group */
3441 unsigned long group_load; /* Total load over the CPUs of the group */
3442 unsigned long sum_nr_running; /* Nr tasks running in the group */
3443 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3444 unsigned long group_capacity;
3445 int group_imb; /* Is there an imbalance in the group ? */
3449 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3450 * @group: The group whose first cpu is to be returned.
3452 static inline unsigned int group_first_cpu(struct sched_group *group)
3454 return cpumask_first(sched_group_cpus(group));
3458 * get_sd_load_idx - Obtain the load index for a given sched domain.
3459 * @sd: The sched_domain whose load_idx is to be obtained.
3460 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3462 static inline int get_sd_load_idx(struct sched_domain *sd,
3463 enum cpu_idle_type idle)
3469 load_idx = sd->busy_idx;
3472 case CPU_NEWLY_IDLE:
3473 load_idx = sd->newidle_idx;
3476 load_idx = sd->idle_idx;
3484 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3486 * init_sd_power_savings_stats - Initialize power savings statistics for
3487 * the given sched_domain, during load balancing.
3489 * @sd: Sched domain whose power-savings statistics are to be initialized.
3490 * @sds: Variable containing the statistics for sd.
3491 * @idle: Idle status of the CPU at which we're performing load-balancing.
3493 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3494 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3497 * Busy processors will not participate in power savings
3500 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3501 sds->power_savings_balance = 0;
3503 sds->power_savings_balance = 1;
3504 sds->min_nr_running = ULONG_MAX;
3505 sds->leader_nr_running = 0;
3510 * update_sd_power_savings_stats - Update the power saving stats for a
3511 * sched_domain while performing load balancing.
3513 * @group: sched_group belonging to the sched_domain under consideration.
3514 * @sds: Variable containing the statistics of the sched_domain
3515 * @local_group: Does group contain the CPU for which we're performing
3517 * @sgs: Variable containing the statistics of the group.
3519 static inline void update_sd_power_savings_stats(struct sched_group *group,
3520 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3523 if (!sds->power_savings_balance)
3527 * If the local group is idle or completely loaded
3528 * no need to do power savings balance at this domain
3530 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3531 !sds->this_nr_running))
3532 sds->power_savings_balance = 0;
3535 * If a group is already running at full capacity or idle,
3536 * don't include that group in power savings calculations
3538 if (!sds->power_savings_balance ||
3539 sgs->sum_nr_running >= sgs->group_capacity ||
3540 !sgs->sum_nr_running)
3544 * Calculate the group which has the least non-idle load.
3545 * This is the group from where we need to pick up the load
3548 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3549 (sgs->sum_nr_running == sds->min_nr_running &&
3550 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3551 sds->group_min = group;
3552 sds->min_nr_running = sgs->sum_nr_running;
3553 sds->min_load_per_task = sgs->sum_weighted_load /
3554 sgs->sum_nr_running;
3558 * Calculate the group which is almost near its
3559 * capacity but still has some space to pick up some load
3560 * from other group and save more power
3562 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3565 if (sgs->sum_nr_running > sds->leader_nr_running ||
3566 (sgs->sum_nr_running == sds->leader_nr_running &&
3567 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3568 sds->group_leader = group;
3569 sds->leader_nr_running = sgs->sum_nr_running;
3574 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3575 * @sds: Variable containing the statistics of the sched_domain
3576 * under consideration.
3577 * @this_cpu: Cpu at which we're currently performing load-balancing.
3578 * @imbalance: Variable to store the imbalance.
3581 * Check if we have potential to perform some power-savings balance.
3582 * If yes, set the busiest group to be the least loaded group in the
3583 * sched_domain, so that it's CPUs can be put to idle.
3585 * Returns 1 if there is potential to perform power-savings balance.
3588 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3589 int this_cpu, unsigned long *imbalance)
3591 if (!sds->power_savings_balance)
3594 if (sds->this != sds->group_leader ||
3595 sds->group_leader == sds->group_min)
3598 *imbalance = sds->min_load_per_task;
3599 sds->busiest = sds->group_min;
3604 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3605 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3606 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3611 static inline void update_sd_power_savings_stats(struct sched_group *group,
3612 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3617 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3618 int this_cpu, unsigned long *imbalance)
3622 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3625 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3627 return SCHED_LOAD_SCALE;
3630 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3632 return default_scale_freq_power(sd, cpu);
3635 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3637 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3638 unsigned long smt_gain = sd->smt_gain;
3645 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3647 return default_scale_smt_power(sd, cpu);
3650 unsigned long scale_rt_power(int cpu)
3652 struct rq *rq = cpu_rq(cpu);
3653 u64 total, available;
3655 sched_avg_update(rq);
3657 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3658 available = total - rq->rt_avg;
3660 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3661 total = SCHED_LOAD_SCALE;
3663 total >>= SCHED_LOAD_SHIFT;
3665 return div_u64(available, total);
3668 static void update_cpu_power(struct sched_domain *sd, int cpu)
3670 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3671 unsigned long power = SCHED_LOAD_SCALE;
3672 struct sched_group *sdg = sd->groups;
3674 if (sched_feat(ARCH_POWER))
3675 power *= arch_scale_freq_power(sd, cpu);
3677 power *= default_scale_freq_power(sd, cpu);
3679 power >>= SCHED_LOAD_SHIFT;
3681 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3682 if (sched_feat(ARCH_POWER))
3683 power *= arch_scale_smt_power(sd, cpu);
3685 power *= default_scale_smt_power(sd, cpu);
3687 power >>= SCHED_LOAD_SHIFT;
3690 power *= scale_rt_power(cpu);
3691 power >>= SCHED_LOAD_SHIFT;
3696 sdg->cpu_power = power;
3699 static void update_group_power(struct sched_domain *sd, int cpu)
3701 struct sched_domain *child = sd->child;
3702 struct sched_group *group, *sdg = sd->groups;
3703 unsigned long power;
3706 update_cpu_power(sd, cpu);
3712 group = child->groups;
3714 power += group->cpu_power;
3715 group = group->next;
3716 } while (group != child->groups);
3718 sdg->cpu_power = power;
3722 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3723 * @sd: The sched_domain whose statistics are to be updated.
3724 * @group: sched_group whose statistics are to be updated.
3725 * @this_cpu: Cpu for which load balance is currently performed.
3726 * @idle: Idle status of this_cpu
3727 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3728 * @sd_idle: Idle status of the sched_domain containing group.
3729 * @local_group: Does group contain this_cpu.
3730 * @cpus: Set of cpus considered for load balancing.
3731 * @balance: Should we balance.
3732 * @sgs: variable to hold the statistics for this group.
3734 static inline void update_sg_lb_stats(struct sched_domain *sd,
3735 struct sched_group *group, int this_cpu,
3736 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3737 int local_group, const struct cpumask *cpus,
3738 int *balance, struct sg_lb_stats *sgs)
3740 unsigned long load, max_cpu_load, min_cpu_load;
3742 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3743 unsigned long sum_avg_load_per_task;
3744 unsigned long avg_load_per_task;
3747 balance_cpu = group_first_cpu(group);
3748 if (balance_cpu == this_cpu)
3749 update_group_power(sd, this_cpu);
3752 /* Tally up the load of all CPUs in the group */
3753 sum_avg_load_per_task = avg_load_per_task = 0;
3755 min_cpu_load = ~0UL;
3757 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3758 struct rq *rq = cpu_rq(i);
3760 if (*sd_idle && rq->nr_running)
3763 /* Bias balancing toward cpus of our domain */
3765 if (idle_cpu(i) && !first_idle_cpu) {
3770 load = target_load(i, load_idx);
3772 load = source_load(i, load_idx);
3773 if (load > max_cpu_load)
3774 max_cpu_load = load;
3775 if (min_cpu_load > load)
3776 min_cpu_load = load;
3779 sgs->group_load += load;
3780 sgs->sum_nr_running += rq->nr_running;
3781 sgs->sum_weighted_load += weighted_cpuload(i);
3783 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3787 * First idle cpu or the first cpu(busiest) in this sched group
3788 * is eligible for doing load balancing at this and above
3789 * domains. In the newly idle case, we will allow all the cpu's
3790 * to do the newly idle load balance.
3792 if (idle != CPU_NEWLY_IDLE && local_group &&
3793 balance_cpu != this_cpu && balance) {
3798 /* Adjust by relative CPU power of the group */
3799 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3803 * Consider the group unbalanced when the imbalance is larger
3804 * than the average weight of two tasks.
3806 * APZ: with cgroup the avg task weight can vary wildly and
3807 * might not be a suitable number - should we keep a
3808 * normalized nr_running number somewhere that negates
3811 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3814 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3817 sgs->group_capacity =
3818 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3822 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3823 * @sd: sched_domain whose statistics are to be updated.
3824 * @this_cpu: Cpu for which load balance is currently performed.
3825 * @idle: Idle status of this_cpu
3826 * @sd_idle: Idle status of the sched_domain containing group.
3827 * @cpus: Set of cpus considered for load balancing.
3828 * @balance: Should we balance.
3829 * @sds: variable to hold the statistics for this sched_domain.
3831 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3832 enum cpu_idle_type idle, int *sd_idle,
3833 const struct cpumask *cpus, int *balance,
3834 struct sd_lb_stats *sds)
3836 struct sched_domain *child = sd->child;
3837 struct sched_group *group = sd->groups;
3838 struct sg_lb_stats sgs;
3839 int load_idx, prefer_sibling = 0;
3841 if (child && child->flags & SD_PREFER_SIBLING)
3844 init_sd_power_savings_stats(sd, sds, idle);
3845 load_idx = get_sd_load_idx(sd, idle);
3850 local_group = cpumask_test_cpu(this_cpu,
3851 sched_group_cpus(group));
3852 memset(&sgs, 0, sizeof(sgs));
3853 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3854 local_group, cpus, balance, &sgs);
3856 if (local_group && balance && !(*balance))
3859 sds->total_load += sgs.group_load;
3860 sds->total_pwr += group->cpu_power;
3863 * In case the child domain prefers tasks go to siblings
3864 * first, lower the group capacity to one so that we'll try
3865 * and move all the excess tasks away.
3868 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3871 sds->this_load = sgs.avg_load;
3873 sds->this_nr_running = sgs.sum_nr_running;
3874 sds->this_load_per_task = sgs.sum_weighted_load;
3875 } else if (sgs.avg_load > sds->max_load &&
3876 (sgs.sum_nr_running > sgs.group_capacity ||
3878 sds->max_load = sgs.avg_load;
3879 sds->busiest = group;
3880 sds->busiest_nr_running = sgs.sum_nr_running;
3881 sds->busiest_load_per_task = sgs.sum_weighted_load;
3882 sds->group_imb = sgs.group_imb;
3885 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3886 group = group->next;
3887 } while (group != sd->groups);
3891 * fix_small_imbalance - Calculate the minor imbalance that exists
3892 * amongst the groups of a sched_domain, during
3894 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3895 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3896 * @imbalance: Variable to store the imbalance.
3898 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3899 int this_cpu, unsigned long *imbalance)
3901 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3902 unsigned int imbn = 2;
3904 if (sds->this_nr_running) {
3905 sds->this_load_per_task /= sds->this_nr_running;
3906 if (sds->busiest_load_per_task >
3907 sds->this_load_per_task)
3910 sds->this_load_per_task =
3911 cpu_avg_load_per_task(this_cpu);
3913 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3914 sds->busiest_load_per_task * imbn) {
3915 *imbalance = sds->busiest_load_per_task;
3920 * OK, we don't have enough imbalance to justify moving tasks,
3921 * however we may be able to increase total CPU power used by
3925 pwr_now += sds->busiest->cpu_power *
3926 min(sds->busiest_load_per_task, sds->max_load);
3927 pwr_now += sds->this->cpu_power *
3928 min(sds->this_load_per_task, sds->this_load);
3929 pwr_now /= SCHED_LOAD_SCALE;
3931 /* Amount of load we'd subtract */
3932 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3933 sds->busiest->cpu_power;
3934 if (sds->max_load > tmp)
3935 pwr_move += sds->busiest->cpu_power *
3936 min(sds->busiest_load_per_task, sds->max_load - tmp);
3938 /* Amount of load we'd add */
3939 if (sds->max_load * sds->busiest->cpu_power <
3940 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3941 tmp = (sds->max_load * sds->busiest->cpu_power) /
3942 sds->this->cpu_power;
3944 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3945 sds->this->cpu_power;
3946 pwr_move += sds->this->cpu_power *
3947 min(sds->this_load_per_task, sds->this_load + tmp);
3948 pwr_move /= SCHED_LOAD_SCALE;
3950 /* Move if we gain throughput */
3951 if (pwr_move > pwr_now)
3952 *imbalance = sds->busiest_load_per_task;
3956 * calculate_imbalance - Calculate the amount of imbalance present within the
3957 * groups of a given sched_domain during load balance.
3958 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3959 * @this_cpu: Cpu for which currently load balance is being performed.
3960 * @imbalance: The variable to store the imbalance.
3962 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3963 unsigned long *imbalance)
3965 unsigned long max_pull;
3967 * In the presence of smp nice balancing, certain scenarios can have
3968 * max load less than avg load(as we skip the groups at or below
3969 * its cpu_power, while calculating max_load..)
3971 if (sds->max_load < sds->avg_load) {
3973 return fix_small_imbalance(sds, this_cpu, imbalance);
3976 /* Don't want to pull so many tasks that a group would go idle */
3977 max_pull = min(sds->max_load - sds->avg_load,
3978 sds->max_load - sds->busiest_load_per_task);
3980 /* How much load to actually move to equalise the imbalance */
3981 *imbalance = min(max_pull * sds->busiest->cpu_power,
3982 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3986 * if *imbalance is less than the average load per runnable task
3987 * there is no gaurantee that any tasks will be moved so we'll have
3988 * a think about bumping its value to force at least one task to be
3991 if (*imbalance < sds->busiest_load_per_task)
3992 return fix_small_imbalance(sds, this_cpu, imbalance);
3995 /******* find_busiest_group() helpers end here *********************/
3998 * find_busiest_group - Returns the busiest group within the sched_domain
3999 * if there is an imbalance. If there isn't an imbalance, and
4000 * the user has opted for power-savings, it returns a group whose
4001 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4002 * such a group exists.
4004 * Also calculates the amount of weighted load which should be moved
4005 * to restore balance.
4007 * @sd: The sched_domain whose busiest group is to be returned.
4008 * @this_cpu: The cpu for which load balancing is currently being performed.
4009 * @imbalance: Variable which stores amount of weighted load which should
4010 * be moved to restore balance/put a group to idle.
4011 * @idle: The idle status of this_cpu.
4012 * @sd_idle: The idleness of sd
4013 * @cpus: The set of CPUs under consideration for load-balancing.
4014 * @balance: Pointer to a variable indicating if this_cpu
4015 * is the appropriate cpu to perform load balancing at this_level.
4017 * Returns: - the busiest group if imbalance exists.
4018 * - If no imbalance and user has opted for power-savings balance,
4019 * return the least loaded group whose CPUs can be
4020 * put to idle by rebalancing its tasks onto our group.
4022 static struct sched_group *
4023 find_busiest_group(struct sched_domain *sd, int this_cpu,
4024 unsigned long *imbalance, enum cpu_idle_type idle,
4025 int *sd_idle, const struct cpumask *cpus, int *balance)
4027 struct sd_lb_stats sds;
4029 memset(&sds, 0, sizeof(sds));
4032 * Compute the various statistics relavent for load balancing at
4035 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4038 /* Cases where imbalance does not exist from POV of this_cpu */
4039 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4041 * 2) There is no busy sibling group to pull from.
4042 * 3) This group is the busiest group.
4043 * 4) This group is more busy than the avg busieness at this
4045 * 5) The imbalance is within the specified limit.
4046 * 6) Any rebalance would lead to ping-pong
4048 if (balance && !(*balance))
4051 if (!sds.busiest || sds.busiest_nr_running == 0)
4054 if (sds.this_load >= sds.max_load)
4057 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4059 if (sds.this_load >= sds.avg_load)
4062 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4065 sds.busiest_load_per_task /= sds.busiest_nr_running;
4067 sds.busiest_load_per_task =
4068 min(sds.busiest_load_per_task, sds.avg_load);
4071 * We're trying to get all the cpus to the average_load, so we don't
4072 * want to push ourselves above the average load, nor do we wish to
4073 * reduce the max loaded cpu below the average load, as either of these
4074 * actions would just result in more rebalancing later, and ping-pong
4075 * tasks around. Thus we look for the minimum possible imbalance.
4076 * Negative imbalances (*we* are more loaded than anyone else) will
4077 * be counted as no imbalance for these purposes -- we can't fix that
4078 * by pulling tasks to us. Be careful of negative numbers as they'll
4079 * appear as very large values with unsigned longs.
4081 if (sds.max_load <= sds.busiest_load_per_task)
4084 /* Looks like there is an imbalance. Compute it */
4085 calculate_imbalance(&sds, this_cpu, imbalance);
4090 * There is no obvious imbalance. But check if we can do some balancing
4093 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4101 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4104 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4105 unsigned long imbalance, const struct cpumask *cpus)
4107 struct rq *busiest = NULL, *rq;
4108 unsigned long max_load = 0;
4111 for_each_cpu(i, sched_group_cpus(group)) {
4112 unsigned long power = power_of(i);
4113 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4116 if (!cpumask_test_cpu(i, cpus))
4120 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4123 if (capacity && rq->nr_running == 1 && wl > imbalance)
4126 if (wl > max_load) {
4136 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4137 * so long as it is large enough.
4139 #define MAX_PINNED_INTERVAL 512
4141 /* Working cpumask for load_balance and load_balance_newidle. */
4142 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4145 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4146 * tasks if there is an imbalance.
4148 static int load_balance(int this_cpu, struct rq *this_rq,
4149 struct sched_domain *sd, enum cpu_idle_type idle,
4152 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4153 struct sched_group *group;
4154 unsigned long imbalance;
4156 unsigned long flags;
4157 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4159 cpumask_copy(cpus, cpu_active_mask);
4162 * When power savings policy is enabled for the parent domain, idle
4163 * sibling can pick up load irrespective of busy siblings. In this case,
4164 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4165 * portraying it as CPU_NOT_IDLE.
4167 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4168 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4171 schedstat_inc(sd, lb_count[idle]);
4175 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4182 schedstat_inc(sd, lb_nobusyg[idle]);
4186 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4188 schedstat_inc(sd, lb_nobusyq[idle]);
4192 BUG_ON(busiest == this_rq);
4194 schedstat_add(sd, lb_imbalance[idle], imbalance);
4197 if (busiest->nr_running > 1) {
4199 * Attempt to move tasks. If find_busiest_group has found
4200 * an imbalance but busiest->nr_running <= 1, the group is
4201 * still unbalanced. ld_moved simply stays zero, so it is
4202 * correctly treated as an imbalance.
4204 local_irq_save(flags);
4205 double_rq_lock(this_rq, busiest);
4206 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4207 imbalance, sd, idle, &all_pinned);
4208 double_rq_unlock(this_rq, busiest);
4209 local_irq_restore(flags);
4212 * some other cpu did the load balance for us.
4214 if (ld_moved && this_cpu != smp_processor_id())
4215 resched_cpu(this_cpu);
4217 /* All tasks on this runqueue were pinned by CPU affinity */
4218 if (unlikely(all_pinned)) {
4219 cpumask_clear_cpu(cpu_of(busiest), cpus);
4220 if (!cpumask_empty(cpus))
4227 schedstat_inc(sd, lb_failed[idle]);
4228 sd->nr_balance_failed++;
4230 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4232 raw_spin_lock_irqsave(&busiest->lock, flags);
4234 /* don't kick the migration_thread, if the curr
4235 * task on busiest cpu can't be moved to this_cpu
4237 if (!cpumask_test_cpu(this_cpu,
4238 &busiest->curr->cpus_allowed)) {
4239 raw_spin_unlock_irqrestore(&busiest->lock,
4242 goto out_one_pinned;
4245 if (!busiest->active_balance) {
4246 busiest->active_balance = 1;
4247 busiest->push_cpu = this_cpu;
4250 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4252 wake_up_process(busiest->migration_thread);
4255 * We've kicked active balancing, reset the failure
4258 sd->nr_balance_failed = sd->cache_nice_tries+1;
4261 sd->nr_balance_failed = 0;
4263 if (likely(!active_balance)) {
4264 /* We were unbalanced, so reset the balancing interval */
4265 sd->balance_interval = sd->min_interval;
4268 * If we've begun active balancing, start to back off. This
4269 * case may not be covered by the all_pinned logic if there
4270 * is only 1 task on the busy runqueue (because we don't call
4273 if (sd->balance_interval < sd->max_interval)
4274 sd->balance_interval *= 2;
4277 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4278 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4284 schedstat_inc(sd, lb_balanced[idle]);
4286 sd->nr_balance_failed = 0;
4289 /* tune up the balancing interval */
4290 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4291 (sd->balance_interval < sd->max_interval))
4292 sd->balance_interval *= 2;
4294 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4295 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4306 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4307 * tasks if there is an imbalance.
4309 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4310 * this_rq is locked.
4313 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4315 struct sched_group *group;
4316 struct rq *busiest = NULL;
4317 unsigned long imbalance;
4321 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4323 cpumask_copy(cpus, cpu_active_mask);
4326 * When power savings policy is enabled for the parent domain, idle
4327 * sibling can pick up load irrespective of busy siblings. In this case,
4328 * let the state of idle sibling percolate up as IDLE, instead of
4329 * portraying it as CPU_NOT_IDLE.
4331 if (sd->flags & SD_SHARE_CPUPOWER &&
4332 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4335 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4337 update_shares_locked(this_rq, sd);
4338 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4339 &sd_idle, cpus, NULL);
4341 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4345 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4347 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4351 BUG_ON(busiest == this_rq);
4353 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4356 if (busiest->nr_running > 1) {
4357 /* Attempt to move tasks */
4358 double_lock_balance(this_rq, busiest);
4359 /* this_rq->clock is already updated */
4360 update_rq_clock(busiest);
4361 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4362 imbalance, sd, CPU_NEWLY_IDLE,
4364 double_unlock_balance(this_rq, busiest);
4366 if (unlikely(all_pinned)) {
4367 cpumask_clear_cpu(cpu_of(busiest), cpus);
4368 if (!cpumask_empty(cpus))
4374 int active_balance = 0;
4376 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4377 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4378 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4381 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4384 if (sd->nr_balance_failed++ < 2)
4388 * The only task running in a non-idle cpu can be moved to this
4389 * cpu in an attempt to completely freeup the other CPU
4390 * package. The same method used to move task in load_balance()
4391 * have been extended for load_balance_newidle() to speedup
4392 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4394 * The package power saving logic comes from
4395 * find_busiest_group(). If there are no imbalance, then
4396 * f_b_g() will return NULL. However when sched_mc={1,2} then
4397 * f_b_g() will select a group from which a running task may be
4398 * pulled to this cpu in order to make the other package idle.
4399 * If there is no opportunity to make a package idle and if
4400 * there are no imbalance, then f_b_g() will return NULL and no
4401 * action will be taken in load_balance_newidle().
4403 * Under normal task pull operation due to imbalance, there
4404 * will be more than one task in the source run queue and
4405 * move_tasks() will succeed. ld_moved will be true and this
4406 * active balance code will not be triggered.
4409 /* Lock busiest in correct order while this_rq is held */
4410 double_lock_balance(this_rq, busiest);
4413 * don't kick the migration_thread, if the curr
4414 * task on busiest cpu can't be moved to this_cpu
4416 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4417 double_unlock_balance(this_rq, busiest);
4422 if (!busiest->active_balance) {
4423 busiest->active_balance = 1;
4424 busiest->push_cpu = this_cpu;
4428 double_unlock_balance(this_rq, busiest);
4430 * Should not call ttwu while holding a rq->lock
4432 raw_spin_unlock(&this_rq->lock);
4434 wake_up_process(busiest->migration_thread);
4435 raw_spin_lock(&this_rq->lock);
4438 sd->nr_balance_failed = 0;
4440 update_shares_locked(this_rq, sd);
4444 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4445 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4446 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4448 sd->nr_balance_failed = 0;
4454 * idle_balance is called by schedule() if this_cpu is about to become
4455 * idle. Attempts to pull tasks from other CPUs.
4457 static void idle_balance(int this_cpu, struct rq *this_rq)
4459 struct sched_domain *sd;
4460 int pulled_task = 0;
4461 unsigned long next_balance = jiffies + HZ;
4463 this_rq->idle_stamp = this_rq->clock;
4465 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4468 for_each_domain(this_cpu, sd) {
4469 unsigned long interval;
4471 if (!(sd->flags & SD_LOAD_BALANCE))
4474 if (sd->flags & SD_BALANCE_NEWIDLE)
4475 /* If we've pulled tasks over stop searching: */
4476 pulled_task = load_balance_newidle(this_cpu, this_rq,
4479 interval = msecs_to_jiffies(sd->balance_interval);
4480 if (time_after(next_balance, sd->last_balance + interval))
4481 next_balance = sd->last_balance + interval;
4483 this_rq->idle_stamp = 0;
4487 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4489 * We are going idle. next_balance may be set based on
4490 * a busy processor. So reset next_balance.
4492 this_rq->next_balance = next_balance;
4497 * active_load_balance is run by migration threads. It pushes running tasks
4498 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4499 * running on each physical CPU where possible, and avoids physical /
4500 * logical imbalances.
4502 * Called with busiest_rq locked.
4504 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4506 int target_cpu = busiest_rq->push_cpu;
4507 struct sched_domain *sd;
4508 struct rq *target_rq;
4510 /* Is there any task to move? */
4511 if (busiest_rq->nr_running <= 1)
4514 target_rq = cpu_rq(target_cpu);
4517 * This condition is "impossible", if it occurs
4518 * we need to fix it. Originally reported by
4519 * Bjorn Helgaas on a 128-cpu setup.
4521 BUG_ON(busiest_rq == target_rq);
4523 /* move a task from busiest_rq to target_rq */
4524 double_lock_balance(busiest_rq, target_rq);
4525 update_rq_clock(busiest_rq);
4526 update_rq_clock(target_rq);
4528 /* Search for an sd spanning us and the target CPU. */
4529 for_each_domain(target_cpu, sd) {
4530 if ((sd->flags & SD_LOAD_BALANCE) &&
4531 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4536 schedstat_inc(sd, alb_count);
4538 if (move_one_task(target_rq, target_cpu, busiest_rq,
4540 schedstat_inc(sd, alb_pushed);
4542 schedstat_inc(sd, alb_failed);
4544 double_unlock_balance(busiest_rq, target_rq);
4549 atomic_t load_balancer;
4550 cpumask_var_t cpu_mask;
4551 cpumask_var_t ilb_grp_nohz_mask;
4552 } nohz ____cacheline_aligned = {
4553 .load_balancer = ATOMIC_INIT(-1),
4556 int get_nohz_load_balancer(void)
4558 return atomic_read(&nohz.load_balancer);
4561 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4563 * lowest_flag_domain - Return lowest sched_domain containing flag.
4564 * @cpu: The cpu whose lowest level of sched domain is to
4566 * @flag: The flag to check for the lowest sched_domain
4567 * for the given cpu.
4569 * Returns the lowest sched_domain of a cpu which contains the given flag.
4571 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4573 struct sched_domain *sd;
4575 for_each_domain(cpu, sd)
4576 if (sd && (sd->flags & flag))
4583 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4584 * @cpu: The cpu whose domains we're iterating over.
4585 * @sd: variable holding the value of the power_savings_sd
4587 * @flag: The flag to filter the sched_domains to be iterated.
4589 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4590 * set, starting from the lowest sched_domain to the highest.
4592 #define for_each_flag_domain(cpu, sd, flag) \
4593 for (sd = lowest_flag_domain(cpu, flag); \
4594 (sd && (sd->flags & flag)); sd = sd->parent)
4597 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4598 * @ilb_group: group to be checked for semi-idleness
4600 * Returns: 1 if the group is semi-idle. 0 otherwise.
4602 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4603 * and atleast one non-idle CPU. This helper function checks if the given
4604 * sched_group is semi-idle or not.
4606 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4608 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4609 sched_group_cpus(ilb_group));
4612 * A sched_group is semi-idle when it has atleast one busy cpu
4613 * and atleast one idle cpu.
4615 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4618 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4624 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4625 * @cpu: The cpu which is nominating a new idle_load_balancer.
4627 * Returns: Returns the id of the idle load balancer if it exists,
4628 * Else, returns >= nr_cpu_ids.
4630 * This algorithm picks the idle load balancer such that it belongs to a
4631 * semi-idle powersavings sched_domain. The idea is to try and avoid
4632 * completely idle packages/cores just for the purpose of idle load balancing
4633 * when there are other idle cpu's which are better suited for that job.
4635 static int find_new_ilb(int cpu)
4637 struct sched_domain *sd;
4638 struct sched_group *ilb_group;
4641 * Have idle load balancer selection from semi-idle packages only
4642 * when power-aware load balancing is enabled
4644 if (!(sched_smt_power_savings || sched_mc_power_savings))
4648 * Optimize for the case when we have no idle CPUs or only one
4649 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4651 if (cpumask_weight(nohz.cpu_mask) < 2)
4654 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4655 ilb_group = sd->groups;
4658 if (is_semi_idle_group(ilb_group))
4659 return cpumask_first(nohz.ilb_grp_nohz_mask);
4661 ilb_group = ilb_group->next;
4663 } while (ilb_group != sd->groups);
4667 return cpumask_first(nohz.cpu_mask);
4669 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4670 static inline int find_new_ilb(int call_cpu)
4672 return cpumask_first(nohz.cpu_mask);
4677 * This routine will try to nominate the ilb (idle load balancing)
4678 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4679 * load balancing on behalf of all those cpus. If all the cpus in the system
4680 * go into this tickless mode, then there will be no ilb owner (as there is
4681 * no need for one) and all the cpus will sleep till the next wakeup event
4684 * For the ilb owner, tick is not stopped. And this tick will be used
4685 * for idle load balancing. ilb owner will still be part of
4688 * While stopping the tick, this cpu will become the ilb owner if there
4689 * is no other owner. And will be the owner till that cpu becomes busy
4690 * or if all cpus in the system stop their ticks at which point
4691 * there is no need for ilb owner.
4693 * When the ilb owner becomes busy, it nominates another owner, during the
4694 * next busy scheduler_tick()
4696 int select_nohz_load_balancer(int stop_tick)
4698 int cpu = smp_processor_id();
4701 cpu_rq(cpu)->in_nohz_recently = 1;
4703 if (!cpu_active(cpu)) {
4704 if (atomic_read(&nohz.load_balancer) != cpu)
4708 * If we are going offline and still the leader,
4711 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4717 cpumask_set_cpu(cpu, nohz.cpu_mask);
4719 /* time for ilb owner also to sleep */
4720 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4721 if (atomic_read(&nohz.load_balancer) == cpu)
4722 atomic_set(&nohz.load_balancer, -1);
4726 if (atomic_read(&nohz.load_balancer) == -1) {
4727 /* make me the ilb owner */
4728 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4730 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4733 if (!(sched_smt_power_savings ||
4734 sched_mc_power_savings))
4737 * Check to see if there is a more power-efficient
4740 new_ilb = find_new_ilb(cpu);
4741 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4742 atomic_set(&nohz.load_balancer, -1);
4743 resched_cpu(new_ilb);
4749 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4752 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4754 if (atomic_read(&nohz.load_balancer) == cpu)
4755 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4762 static DEFINE_SPINLOCK(balancing);
4765 * It checks each scheduling domain to see if it is due to be balanced,
4766 * and initiates a balancing operation if so.
4768 * Balancing parameters are set up in arch_init_sched_domains.
4770 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4773 struct rq *rq = cpu_rq(cpu);
4774 unsigned long interval;
4775 struct sched_domain *sd;
4776 /* Earliest time when we have to do rebalance again */
4777 unsigned long next_balance = jiffies + 60*HZ;
4778 int update_next_balance = 0;
4781 for_each_domain(cpu, sd) {
4782 if (!(sd->flags & SD_LOAD_BALANCE))
4785 interval = sd->balance_interval;
4786 if (idle != CPU_IDLE)
4787 interval *= sd->busy_factor;
4789 /* scale ms to jiffies */
4790 interval = msecs_to_jiffies(interval);
4791 if (unlikely(!interval))
4793 if (interval > HZ*NR_CPUS/10)
4794 interval = HZ*NR_CPUS/10;
4796 need_serialize = sd->flags & SD_SERIALIZE;
4798 if (need_serialize) {
4799 if (!spin_trylock(&balancing))
4803 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4804 if (load_balance(cpu, rq, sd, idle, &balance)) {
4806 * We've pulled tasks over so either we're no
4807 * longer idle, or one of our SMT siblings is
4810 idle = CPU_NOT_IDLE;
4812 sd->last_balance = jiffies;
4815 spin_unlock(&balancing);
4817 if (time_after(next_balance, sd->last_balance + interval)) {
4818 next_balance = sd->last_balance + interval;
4819 update_next_balance = 1;
4823 * Stop the load balance at this level. There is another
4824 * CPU in our sched group which is doing load balancing more
4832 * next_balance will be updated only when there is a need.
4833 * When the cpu is attached to null domain for ex, it will not be
4836 if (likely(update_next_balance))
4837 rq->next_balance = next_balance;
4841 * run_rebalance_domains is triggered when needed from the scheduler tick.
4842 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4843 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4845 static void run_rebalance_domains(struct softirq_action *h)
4847 int this_cpu = smp_processor_id();
4848 struct rq *this_rq = cpu_rq(this_cpu);
4849 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4850 CPU_IDLE : CPU_NOT_IDLE;
4852 rebalance_domains(this_cpu, idle);
4856 * If this cpu is the owner for idle load balancing, then do the
4857 * balancing on behalf of the other idle cpus whose ticks are
4860 if (this_rq->idle_at_tick &&
4861 atomic_read(&nohz.load_balancer) == this_cpu) {
4865 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4866 if (balance_cpu == this_cpu)
4870 * If this cpu gets work to do, stop the load balancing
4871 * work being done for other cpus. Next load
4872 * balancing owner will pick it up.
4877 rebalance_domains(balance_cpu, CPU_IDLE);
4879 rq = cpu_rq(balance_cpu);
4880 if (time_after(this_rq->next_balance, rq->next_balance))
4881 this_rq->next_balance = rq->next_balance;
4887 static inline int on_null_domain(int cpu)
4889 return !rcu_dereference(cpu_rq(cpu)->sd);
4893 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4895 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4896 * idle load balancing owner or decide to stop the periodic load balancing,
4897 * if the whole system is idle.
4899 static inline void trigger_load_balance(struct rq *rq, int cpu)
4903 * If we were in the nohz mode recently and busy at the current
4904 * scheduler tick, then check if we need to nominate new idle
4907 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4908 rq->in_nohz_recently = 0;
4910 if (atomic_read(&nohz.load_balancer) == cpu) {
4911 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4912 atomic_set(&nohz.load_balancer, -1);
4915 if (atomic_read(&nohz.load_balancer) == -1) {
4916 int ilb = find_new_ilb(cpu);
4918 if (ilb < nr_cpu_ids)
4924 * If this cpu is idle and doing idle load balancing for all the
4925 * cpus with ticks stopped, is it time for that to stop?
4927 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4928 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4934 * If this cpu is idle and the idle load balancing is done by
4935 * someone else, then no need raise the SCHED_SOFTIRQ
4937 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4938 cpumask_test_cpu(cpu, nohz.cpu_mask))
4941 /* Don't need to rebalance while attached to NULL domain */
4942 if (time_after_eq(jiffies, rq->next_balance) &&
4943 likely(!on_null_domain(cpu)))
4944 raise_softirq(SCHED_SOFTIRQ);
4947 #else /* CONFIG_SMP */
4950 * on UP we do not need to balance between CPUs:
4952 static inline void idle_balance(int cpu, struct rq *rq)
4958 DEFINE_PER_CPU(struct kernel_stat, kstat);
4960 EXPORT_PER_CPU_SYMBOL(kstat);
4963 * Return any ns on the sched_clock that have not yet been accounted in
4964 * @p in case that task is currently running.
4966 * Called with task_rq_lock() held on @rq.
4968 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4972 if (task_current(rq, p)) {
4973 update_rq_clock(rq);
4974 ns = rq->clock - p->se.exec_start;
4982 unsigned long long task_delta_exec(struct task_struct *p)
4984 unsigned long flags;
4988 rq = task_rq_lock(p, &flags);
4989 ns = do_task_delta_exec(p, rq);
4990 task_rq_unlock(rq, &flags);
4996 * Return accounted runtime for the task.
4997 * In case the task is currently running, return the runtime plus current's
4998 * pending runtime that have not been accounted yet.
5000 unsigned long long task_sched_runtime(struct task_struct *p)
5002 unsigned long flags;
5006 rq = task_rq_lock(p, &flags);
5007 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
5008 task_rq_unlock(rq, &flags);
5014 * Return sum_exec_runtime for the thread group.
5015 * In case the task is currently running, return the sum plus current's
5016 * pending runtime that have not been accounted yet.
5018 * Note that the thread group might have other running tasks as well,
5019 * so the return value not includes other pending runtime that other
5020 * running tasks might have.
5022 unsigned long long thread_group_sched_runtime(struct task_struct *p)
5024 struct task_cputime totals;
5025 unsigned long flags;
5029 rq = task_rq_lock(p, &flags);
5030 thread_group_cputime(p, &totals);
5031 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
5032 task_rq_unlock(rq, &flags);
5038 * Account user cpu time to a process.
5039 * @p: the process that the cpu time gets accounted to
5040 * @cputime: the cpu time spent in user space since the last update
5041 * @cputime_scaled: cputime scaled by cpu frequency
5043 void account_user_time(struct task_struct *p, cputime_t cputime,
5044 cputime_t cputime_scaled)
5046 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5049 /* Add user time to process. */
5050 p->utime = cputime_add(p->utime, cputime);
5051 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5052 account_group_user_time(p, cputime);
5054 /* Add user time to cpustat. */
5055 tmp = cputime_to_cputime64(cputime);
5056 if (TASK_NICE(p) > 0)
5057 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5059 cpustat->user = cputime64_add(cpustat->user, tmp);
5061 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5062 /* Account for user time used */
5063 acct_update_integrals(p);
5067 * Account guest cpu time to a process.
5068 * @p: the process that the cpu time gets accounted to
5069 * @cputime: the cpu time spent in virtual machine since the last update
5070 * @cputime_scaled: cputime scaled by cpu frequency
5072 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5073 cputime_t cputime_scaled)
5076 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5078 tmp = cputime_to_cputime64(cputime);
5080 /* Add guest time to process. */
5081 p->utime = cputime_add(p->utime, cputime);
5082 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5083 account_group_user_time(p, cputime);
5084 p->gtime = cputime_add(p->gtime, cputime);
5086 /* Add guest time to cpustat. */
5087 if (TASK_NICE(p) > 0) {
5088 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5089 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
5091 cpustat->user = cputime64_add(cpustat->user, tmp);
5092 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5097 * Account system cpu time to a process.
5098 * @p: the process that the cpu time gets accounted to
5099 * @hardirq_offset: the offset to subtract from hardirq_count()
5100 * @cputime: the cpu time spent in kernel space since the last update
5101 * @cputime_scaled: cputime scaled by cpu frequency
5103 void account_system_time(struct task_struct *p, int hardirq_offset,
5104 cputime_t cputime, cputime_t cputime_scaled)
5106 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5109 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5110 account_guest_time(p, cputime, cputime_scaled);
5114 /* Add system time to process. */
5115 p->stime = cputime_add(p->stime, cputime);
5116 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5117 account_group_system_time(p, cputime);
5119 /* Add system time to cpustat. */
5120 tmp = cputime_to_cputime64(cputime);
5121 if (hardirq_count() - hardirq_offset)
5122 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5123 else if (softirq_count())
5124 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5126 cpustat->system = cputime64_add(cpustat->system, tmp);
5128 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5130 /* Account for system time used */
5131 acct_update_integrals(p);
5135 * Account for involuntary wait time.
5136 * @steal: the cpu time spent in involuntary wait
5138 void account_steal_time(cputime_t cputime)
5140 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5141 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5143 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5147 * Account for idle time.
5148 * @cputime: the cpu time spent in idle wait
5150 void account_idle_time(cputime_t cputime)
5152 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5153 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5154 struct rq *rq = this_rq();
5156 if (atomic_read(&rq->nr_iowait) > 0)
5157 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5159 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5162 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5165 * Account a single tick of cpu time.
5166 * @p: the process that the cpu time gets accounted to
5167 * @user_tick: indicates if the tick is a user or a system tick
5169 void account_process_tick(struct task_struct *p, int user_tick)
5171 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5172 struct rq *rq = this_rq();
5175 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5176 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5177 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5180 account_idle_time(cputime_one_jiffy);
5184 * Account multiple ticks of steal time.
5185 * @p: the process from which the cpu time has been stolen
5186 * @ticks: number of stolen ticks
5188 void account_steal_ticks(unsigned long ticks)
5190 account_steal_time(jiffies_to_cputime(ticks));
5194 * Account multiple ticks of idle time.
5195 * @ticks: number of stolen ticks
5197 void account_idle_ticks(unsigned long ticks)
5199 account_idle_time(jiffies_to_cputime(ticks));
5205 * Use precise platform statistics if available:
5207 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5208 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5214 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5216 struct task_cputime cputime;
5218 thread_group_cputime(p, &cputime);
5220 *ut = cputime.utime;
5221 *st = cputime.stime;
5225 #ifndef nsecs_to_cputime
5226 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
5229 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5231 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
5234 * Use CFS's precise accounting:
5236 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
5241 temp = (u64)(rtime * utime);
5242 do_div(temp, total);
5243 utime = (cputime_t)temp;
5248 * Compare with previous values, to keep monotonicity:
5250 p->prev_utime = max(p->prev_utime, utime);
5251 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
5253 *ut = p->prev_utime;
5254 *st = p->prev_stime;
5258 * Must be called with siglock held.
5260 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5262 struct signal_struct *sig = p->signal;
5263 struct task_cputime cputime;
5264 cputime_t rtime, utime, total;
5266 thread_group_cputime(p, &cputime);
5268 total = cputime_add(cputime.utime, cputime.stime);
5269 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5274 temp = (u64)(rtime * cputime.utime);
5275 do_div(temp, total);
5276 utime = (cputime_t)temp;
5280 sig->prev_utime = max(sig->prev_utime, utime);
5281 sig->prev_stime = max(sig->prev_stime,
5282 cputime_sub(rtime, sig->prev_utime));
5284 *ut = sig->prev_utime;
5285 *st = sig->prev_stime;
5290 * This function gets called by the timer code, with HZ frequency.
5291 * We call it with interrupts disabled.
5293 * It also gets called by the fork code, when changing the parent's
5296 void scheduler_tick(void)
5298 int cpu = smp_processor_id();
5299 struct rq *rq = cpu_rq(cpu);
5300 struct task_struct *curr = rq->curr;
5304 raw_spin_lock(&rq->lock);
5305 update_rq_clock(rq);
5306 update_cpu_load(rq);
5307 curr->sched_class->task_tick(rq, curr, 0);
5308 raw_spin_unlock(&rq->lock);
5310 perf_event_task_tick(curr, cpu);
5313 rq->idle_at_tick = idle_cpu(cpu);
5314 trigger_load_balance(rq, cpu);
5318 notrace unsigned long get_parent_ip(unsigned long addr)
5320 if (in_lock_functions(addr)) {
5321 addr = CALLER_ADDR2;
5322 if (in_lock_functions(addr))
5323 addr = CALLER_ADDR3;
5328 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5329 defined(CONFIG_PREEMPT_TRACER))
5331 void __kprobes add_preempt_count(int val)
5333 #ifdef CONFIG_DEBUG_PREEMPT
5337 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5340 preempt_count() += val;
5341 #ifdef CONFIG_DEBUG_PREEMPT
5343 * Spinlock count overflowing soon?
5345 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5348 if (preempt_count() == val)
5349 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5351 EXPORT_SYMBOL(add_preempt_count);
5353 void __kprobes sub_preempt_count(int val)
5355 #ifdef CONFIG_DEBUG_PREEMPT
5359 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5362 * Is the spinlock portion underflowing?
5364 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5365 !(preempt_count() & PREEMPT_MASK)))
5369 if (preempt_count() == val)
5370 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5371 preempt_count() -= val;
5373 EXPORT_SYMBOL(sub_preempt_count);
5378 * Print scheduling while atomic bug:
5380 static noinline void __schedule_bug(struct task_struct *prev)
5382 struct pt_regs *regs = get_irq_regs();
5384 pr_err("BUG: scheduling while atomic: %s/%d/0x%08x\n",
5385 prev->comm, prev->pid, preempt_count());
5387 debug_show_held_locks(prev);
5389 if (irqs_disabled())
5390 print_irqtrace_events(prev);
5399 * Various schedule()-time debugging checks and statistics:
5401 static inline void schedule_debug(struct task_struct *prev)
5404 * Test if we are atomic. Since do_exit() needs to call into
5405 * schedule() atomically, we ignore that path for now.
5406 * Otherwise, whine if we are scheduling when we should not be.
5408 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5409 __schedule_bug(prev);
5411 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5413 schedstat_inc(this_rq(), sched_count);
5414 #ifdef CONFIG_SCHEDSTATS
5415 if (unlikely(prev->lock_depth >= 0)) {
5416 schedstat_inc(this_rq(), bkl_count);
5417 schedstat_inc(prev, sched_info.bkl_count);
5422 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5424 if (prev->state == TASK_RUNNING) {
5425 u64 runtime = prev->se.sum_exec_runtime;
5427 runtime -= prev->se.prev_sum_exec_runtime;
5428 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5431 * In order to avoid avg_overlap growing stale when we are
5432 * indeed overlapping and hence not getting put to sleep, grow
5433 * the avg_overlap on preemption.
5435 * We use the average preemption runtime because that
5436 * correlates to the amount of cache footprint a task can
5439 update_avg(&prev->se.avg_overlap, runtime);
5441 prev->sched_class->put_prev_task(rq, prev);
5445 * Pick up the highest-prio task:
5447 static inline struct task_struct *
5448 pick_next_task(struct rq *rq)
5450 const struct sched_class *class;
5451 struct task_struct *p;
5454 * Optimization: we know that if all tasks are in
5455 * the fair class we can call that function directly:
5457 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5458 p = fair_sched_class.pick_next_task(rq);
5463 class = sched_class_highest;
5465 p = class->pick_next_task(rq);
5469 * Will never be NULL as the idle class always
5470 * returns a non-NULL p:
5472 class = class->next;
5477 * schedule() is the main scheduler function.
5479 asmlinkage void __sched schedule(void)
5481 struct task_struct *prev, *next;
5482 unsigned long *switch_count;
5488 cpu = smp_processor_id();
5492 switch_count = &prev->nivcsw;
5494 release_kernel_lock(prev);
5495 need_resched_nonpreemptible:
5497 schedule_debug(prev);
5499 if (sched_feat(HRTICK))
5502 raw_spin_lock_irq(&rq->lock);
5503 update_rq_clock(rq);
5504 clear_tsk_need_resched(prev);
5506 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5507 if (unlikely(signal_pending_state(prev->state, prev)))
5508 prev->state = TASK_RUNNING;
5510 deactivate_task(rq, prev, 1);
5511 switch_count = &prev->nvcsw;
5514 pre_schedule(rq, prev);
5516 if (unlikely(!rq->nr_running))
5517 idle_balance(cpu, rq);
5519 put_prev_task(rq, prev);
5520 next = pick_next_task(rq);
5522 if (likely(prev != next)) {
5523 sched_info_switch(prev, next);
5524 perf_event_task_sched_out(prev, next, cpu);
5530 context_switch(rq, prev, next); /* unlocks the rq */
5532 * the context switch might have flipped the stack from under
5533 * us, hence refresh the local variables.
5535 cpu = smp_processor_id();
5538 raw_spin_unlock_irq(&rq->lock);
5542 if (unlikely(reacquire_kernel_lock(current) < 0))
5543 goto need_resched_nonpreemptible;
5545 preempt_enable_no_resched();
5549 EXPORT_SYMBOL(schedule);
5551 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5553 * Look out! "owner" is an entirely speculative pointer
5554 * access and not reliable.
5556 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5561 if (!sched_feat(OWNER_SPIN))
5564 #ifdef CONFIG_DEBUG_PAGEALLOC
5566 * Need to access the cpu field knowing that
5567 * DEBUG_PAGEALLOC could have unmapped it if
5568 * the mutex owner just released it and exited.
5570 if (probe_kernel_address(&owner->cpu, cpu))
5577 * Even if the access succeeded (likely case),
5578 * the cpu field may no longer be valid.
5580 if (cpu >= nr_cpumask_bits)
5584 * We need to validate that we can do a
5585 * get_cpu() and that we have the percpu area.
5587 if (!cpu_online(cpu))
5594 * Owner changed, break to re-assess state.
5596 if (lock->owner != owner)
5600 * Is that owner really running on that cpu?
5602 if (task_thread_info(rq->curr) != owner || need_resched())
5612 #ifdef CONFIG_PREEMPT
5614 * this is the entry point to schedule() from in-kernel preemption
5615 * off of preempt_enable. Kernel preemptions off return from interrupt
5616 * occur there and call schedule directly.
5618 asmlinkage void __sched preempt_schedule(void)
5620 struct thread_info *ti = current_thread_info();
5623 * If there is a non-zero preempt_count or interrupts are disabled,
5624 * we do not want to preempt the current task. Just return..
5626 if (likely(ti->preempt_count || irqs_disabled()))
5630 add_preempt_count(PREEMPT_ACTIVE);
5632 sub_preempt_count(PREEMPT_ACTIVE);
5635 * Check again in case we missed a preemption opportunity
5636 * between schedule and now.
5639 } while (need_resched());
5641 EXPORT_SYMBOL(preempt_schedule);
5644 * this is the entry point to schedule() from kernel preemption
5645 * off of irq context.
5646 * Note, that this is called and return with irqs disabled. This will
5647 * protect us against recursive calling from irq.
5649 asmlinkage void __sched preempt_schedule_irq(void)
5651 struct thread_info *ti = current_thread_info();
5653 /* Catch callers which need to be fixed */
5654 BUG_ON(ti->preempt_count || !irqs_disabled());
5657 add_preempt_count(PREEMPT_ACTIVE);
5660 local_irq_disable();
5661 sub_preempt_count(PREEMPT_ACTIVE);
5664 * Check again in case we missed a preemption opportunity
5665 * between schedule and now.
5668 } while (need_resched());
5671 #endif /* CONFIG_PREEMPT */
5673 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5676 return try_to_wake_up(curr->private, mode, wake_flags);
5678 EXPORT_SYMBOL(default_wake_function);
5681 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5682 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5683 * number) then we wake all the non-exclusive tasks and one exclusive task.
5685 * There are circumstances in which we can try to wake a task which has already
5686 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5687 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5689 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5690 int nr_exclusive, int wake_flags, void *key)
5692 wait_queue_t *curr, *next;
5694 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5695 unsigned flags = curr->flags;
5697 if (curr->func(curr, mode, wake_flags, key) &&
5698 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5704 * __wake_up - wake up threads blocked on a waitqueue.
5706 * @mode: which threads
5707 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5708 * @key: is directly passed to the wakeup function
5710 * It may be assumed that this function implies a write memory barrier before
5711 * changing the task state if and only if any tasks are woken up.
5713 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5714 int nr_exclusive, void *key)
5716 unsigned long flags;
5718 spin_lock_irqsave(&q->lock, flags);
5719 __wake_up_common(q, mode, nr_exclusive, 0, key);
5720 spin_unlock_irqrestore(&q->lock, flags);
5722 EXPORT_SYMBOL(__wake_up);
5725 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5727 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5729 __wake_up_common(q, mode, 1, 0, NULL);
5732 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5734 __wake_up_common(q, mode, 1, 0, key);
5738 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5740 * @mode: which threads
5741 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5742 * @key: opaque value to be passed to wakeup targets
5744 * The sync wakeup differs that the waker knows that it will schedule
5745 * away soon, so while the target thread will be woken up, it will not
5746 * be migrated to another CPU - ie. the two threads are 'synchronized'
5747 * with each other. This can prevent needless bouncing between CPUs.
5749 * On UP it can prevent extra preemption.
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 __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5755 int nr_exclusive, void *key)
5757 unsigned long flags;
5758 int wake_flags = WF_SYNC;
5763 if (unlikely(!nr_exclusive))
5766 spin_lock_irqsave(&q->lock, flags);
5767 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5768 spin_unlock_irqrestore(&q->lock, flags);
5770 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5773 * __wake_up_sync - see __wake_up_sync_key()
5775 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5777 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5779 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5782 * complete: - signals a single thread waiting on this completion
5783 * @x: holds the state of this particular completion
5785 * This will wake up a single thread waiting on this completion. Threads will be
5786 * awakened in the same order in which they were queued.
5788 * See also complete_all(), wait_for_completion() and related routines.
5790 * It may be assumed that this function implies a write memory barrier before
5791 * changing the task state if and only if any tasks are woken up.
5793 void complete(struct completion *x)
5795 unsigned long flags;
5797 spin_lock_irqsave(&x->wait.lock, flags);
5799 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5800 spin_unlock_irqrestore(&x->wait.lock, flags);
5802 EXPORT_SYMBOL(complete);
5805 * complete_all: - signals all threads waiting on this completion
5806 * @x: holds the state of this particular completion
5808 * This will wake up all threads waiting on this particular completion event.
5810 * It may be assumed that this function implies a write memory barrier before
5811 * changing the task state if and only if any tasks are woken up.
5813 void complete_all(struct completion *x)
5815 unsigned long flags;
5817 spin_lock_irqsave(&x->wait.lock, flags);
5818 x->done += UINT_MAX/2;
5819 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5820 spin_unlock_irqrestore(&x->wait.lock, flags);
5822 EXPORT_SYMBOL(complete_all);
5824 static inline long __sched
5825 do_wait_for_common(struct completion *x, long timeout, int state)
5828 DECLARE_WAITQUEUE(wait, current);
5830 wait.flags |= WQ_FLAG_EXCLUSIVE;
5831 __add_wait_queue_tail(&x->wait, &wait);
5833 if (signal_pending_state(state, current)) {
5834 timeout = -ERESTARTSYS;
5837 __set_current_state(state);
5838 spin_unlock_irq(&x->wait.lock);
5839 timeout = schedule_timeout(timeout);
5840 spin_lock_irq(&x->wait.lock);
5841 } while (!x->done && timeout);
5842 __remove_wait_queue(&x->wait, &wait);
5847 return timeout ?: 1;
5851 wait_for_common(struct completion *x, long timeout, int state)
5855 spin_lock_irq(&x->wait.lock);
5856 timeout = do_wait_for_common(x, timeout, state);
5857 spin_unlock_irq(&x->wait.lock);
5862 * wait_for_completion: - waits for completion of a task
5863 * @x: holds the state of this particular completion
5865 * This waits to be signaled for completion of a specific task. It is NOT
5866 * interruptible and there is no timeout.
5868 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5869 * and interrupt capability. Also see complete().
5871 void __sched wait_for_completion(struct completion *x)
5873 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5875 EXPORT_SYMBOL(wait_for_completion);
5878 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5879 * @x: holds the state of this particular completion
5880 * @timeout: timeout value in jiffies
5882 * This waits for either a completion of a specific task to be signaled or for a
5883 * specified timeout to expire. The timeout is in jiffies. It is not
5886 unsigned long __sched
5887 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5889 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5891 EXPORT_SYMBOL(wait_for_completion_timeout);
5894 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5895 * @x: holds the state of this particular completion
5897 * This waits for completion of a specific task to be signaled. It is
5900 int __sched wait_for_completion_interruptible(struct completion *x)
5902 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5903 if (t == -ERESTARTSYS)
5907 EXPORT_SYMBOL(wait_for_completion_interruptible);
5910 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5911 * @x: holds the state of this particular completion
5912 * @timeout: timeout value in jiffies
5914 * This waits for either a completion of a specific task to be signaled or for a
5915 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5917 unsigned long __sched
5918 wait_for_completion_interruptible_timeout(struct completion *x,
5919 unsigned long timeout)
5921 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5923 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5926 * wait_for_completion_killable: - waits for completion of a task (killable)
5927 * @x: holds the state of this particular completion
5929 * This waits to be signaled for completion of a specific task. It can be
5930 * interrupted by a kill signal.
5932 int __sched wait_for_completion_killable(struct completion *x)
5934 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5935 if (t == -ERESTARTSYS)
5939 EXPORT_SYMBOL(wait_for_completion_killable);
5942 * try_wait_for_completion - try to decrement a completion without blocking
5943 * @x: completion structure
5945 * Returns: 0 if a decrement cannot be done without blocking
5946 * 1 if a decrement succeeded.
5948 * If a completion is being used as a counting completion,
5949 * attempt to decrement the counter without blocking. This
5950 * enables us to avoid waiting if the resource the completion
5951 * is protecting is not available.
5953 bool try_wait_for_completion(struct completion *x)
5955 unsigned long flags;
5958 spin_lock_irqsave(&x->wait.lock, flags);
5963 spin_unlock_irqrestore(&x->wait.lock, flags);
5966 EXPORT_SYMBOL(try_wait_for_completion);
5969 * completion_done - Test to see if a completion has any waiters
5970 * @x: completion structure
5972 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5973 * 1 if there are no waiters.
5976 bool completion_done(struct completion *x)
5978 unsigned long flags;
5981 spin_lock_irqsave(&x->wait.lock, flags);
5984 spin_unlock_irqrestore(&x->wait.lock, flags);
5987 EXPORT_SYMBOL(completion_done);
5990 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5992 unsigned long flags;
5995 init_waitqueue_entry(&wait, current);
5997 __set_current_state(state);
5999 spin_lock_irqsave(&q->lock, flags);
6000 __add_wait_queue(q, &wait);
6001 spin_unlock(&q->lock);
6002 timeout = schedule_timeout(timeout);
6003 spin_lock_irq(&q->lock);
6004 __remove_wait_queue(q, &wait);
6005 spin_unlock_irqrestore(&q->lock, flags);
6010 void __sched interruptible_sleep_on(wait_queue_head_t *q)
6012 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6014 EXPORT_SYMBOL(interruptible_sleep_on);
6017 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
6019 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
6021 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
6023 void __sched sleep_on(wait_queue_head_t *q)
6025 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6027 EXPORT_SYMBOL(sleep_on);
6029 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
6031 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
6033 EXPORT_SYMBOL(sleep_on_timeout);
6035 #ifdef CONFIG_RT_MUTEXES
6038 * rt_mutex_setprio - set the current priority of a task
6040 * @prio: prio value (kernel-internal form)
6042 * This function changes the 'effective' priority of a task. It does
6043 * not touch ->normal_prio like __setscheduler().
6045 * Used by the rt_mutex code to implement priority inheritance logic.
6047 void rt_mutex_setprio(struct task_struct *p, int prio)
6049 unsigned long flags;
6050 int oldprio, on_rq, running;
6052 const struct sched_class *prev_class = p->sched_class;
6054 BUG_ON(prio < 0 || prio > MAX_PRIO);
6056 rq = task_rq_lock(p, &flags);
6057 update_rq_clock(rq);
6060 on_rq = p->se.on_rq;
6061 running = task_current(rq, p);
6063 dequeue_task(rq, p, 0);
6065 p->sched_class->put_prev_task(rq, p);
6068 p->sched_class = &rt_sched_class;
6070 p->sched_class = &fair_sched_class;
6075 p->sched_class->set_curr_task(rq);
6077 enqueue_task(rq, p, 0);
6079 check_class_changed(rq, p, prev_class, oldprio, running);
6081 task_rq_unlock(rq, &flags);
6086 void set_user_nice(struct task_struct *p, long nice)
6088 int old_prio, delta, on_rq;
6089 unsigned long flags;
6092 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6095 * We have to be careful, if called from sys_setpriority(),
6096 * the task might be in the middle of scheduling on another CPU.
6098 rq = task_rq_lock(p, &flags);
6099 update_rq_clock(rq);
6101 * The RT priorities are set via sched_setscheduler(), but we still
6102 * allow the 'normal' nice value to be set - but as expected
6103 * it wont have any effect on scheduling until the task is
6104 * SCHED_FIFO/SCHED_RR:
6106 if (task_has_rt_policy(p)) {
6107 p->static_prio = NICE_TO_PRIO(nice);
6110 on_rq = p->se.on_rq;
6112 dequeue_task(rq, p, 0);
6114 p->static_prio = NICE_TO_PRIO(nice);
6117 p->prio = effective_prio(p);
6118 delta = p->prio - old_prio;
6121 enqueue_task(rq, p, 0);
6123 * If the task increased its priority or is running and
6124 * lowered its priority, then reschedule its CPU:
6126 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6127 resched_task(rq->curr);
6130 task_rq_unlock(rq, &flags);
6132 EXPORT_SYMBOL(set_user_nice);
6135 * can_nice - check if a task can reduce its nice value
6139 int can_nice(const struct task_struct *p, const int nice)
6141 /* convert nice value [19,-20] to rlimit style value [1,40] */
6142 int nice_rlim = 20 - nice;
6144 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6145 capable(CAP_SYS_NICE));
6148 #ifdef __ARCH_WANT_SYS_NICE
6151 * sys_nice - change the priority of the current process.
6152 * @increment: priority increment
6154 * sys_setpriority is a more generic, but much slower function that
6155 * does similar things.
6157 SYSCALL_DEFINE1(nice, int, increment)
6162 * Setpriority might change our priority at the same moment.
6163 * We don't have to worry. Conceptually one call occurs first
6164 * and we have a single winner.
6166 if (increment < -40)
6171 nice = TASK_NICE(current) + increment;
6177 if (increment < 0 && !can_nice(current, nice))
6180 retval = security_task_setnice(current, nice);
6184 set_user_nice(current, nice);
6191 * task_prio - return the priority value of a given task.
6192 * @p: the task in question.
6194 * This is the priority value as seen by users in /proc.
6195 * RT tasks are offset by -200. Normal tasks are centered
6196 * around 0, value goes from -16 to +15.
6198 int task_prio(const struct task_struct *p)
6200 return p->prio - MAX_RT_PRIO;
6204 * task_nice - return the nice value of a given task.
6205 * @p: the task in question.
6207 int task_nice(const struct task_struct *p)
6209 return TASK_NICE(p);
6211 EXPORT_SYMBOL(task_nice);
6214 * idle_cpu - is a given cpu idle currently?
6215 * @cpu: the processor in question.
6217 int idle_cpu(int cpu)
6219 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6223 * idle_task - return the idle task for a given cpu.
6224 * @cpu: the processor in question.
6226 struct task_struct *idle_task(int cpu)
6228 return cpu_rq(cpu)->idle;
6232 * find_process_by_pid - find a process with a matching PID value.
6233 * @pid: the pid in question.
6235 static struct task_struct *find_process_by_pid(pid_t pid)
6237 return pid ? find_task_by_vpid(pid) : current;
6240 /* Actually do priority change: must hold rq lock. */
6242 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6244 BUG_ON(p->se.on_rq);
6247 p->rt_priority = prio;
6248 p->normal_prio = normal_prio(p);
6249 /* we are holding p->pi_lock already */
6250 p->prio = rt_mutex_getprio(p);
6251 if (rt_prio(p->prio))
6252 p->sched_class = &rt_sched_class;
6254 p->sched_class = &fair_sched_class;
6259 * check the target process has a UID that matches the current process's
6261 static bool check_same_owner(struct task_struct *p)
6263 const struct cred *cred = current_cred(), *pcred;
6267 pcred = __task_cred(p);
6268 match = (cred->euid == pcred->euid ||
6269 cred->euid == pcred->uid);
6274 static int __sched_setscheduler(struct task_struct *p, int policy,
6275 struct sched_param *param, bool user)
6277 int retval, oldprio, oldpolicy = -1, on_rq, running;
6278 unsigned long flags;
6279 const struct sched_class *prev_class = p->sched_class;
6283 /* may grab non-irq protected spin_locks */
6284 BUG_ON(in_interrupt());
6286 /* double check policy once rq lock held */
6288 reset_on_fork = p->sched_reset_on_fork;
6289 policy = oldpolicy = p->policy;
6291 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6292 policy &= ~SCHED_RESET_ON_FORK;
6294 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6295 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6296 policy != SCHED_IDLE)
6301 * Valid priorities for SCHED_FIFO and SCHED_RR are
6302 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6303 * SCHED_BATCH and SCHED_IDLE is 0.
6305 if (param->sched_priority < 0 ||
6306 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6307 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6309 if (rt_policy(policy) != (param->sched_priority != 0))
6313 * Allow unprivileged RT tasks to decrease priority:
6315 if (user && !capable(CAP_SYS_NICE)) {
6316 if (rt_policy(policy)) {
6317 unsigned long rlim_rtprio;
6319 if (!lock_task_sighand(p, &flags))
6321 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6322 unlock_task_sighand(p, &flags);
6324 /* can't set/change the rt policy */
6325 if (policy != p->policy && !rlim_rtprio)
6328 /* can't increase priority */
6329 if (param->sched_priority > p->rt_priority &&
6330 param->sched_priority > rlim_rtprio)
6334 * Like positive nice levels, dont allow tasks to
6335 * move out of SCHED_IDLE either:
6337 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6340 /* can't change other user's priorities */
6341 if (!check_same_owner(p))
6344 /* Normal users shall not reset the sched_reset_on_fork flag */
6345 if (p->sched_reset_on_fork && !reset_on_fork)
6350 #ifdef CONFIG_RT_GROUP_SCHED
6352 * Do not allow realtime tasks into groups that have no runtime
6355 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6356 task_group(p)->rt_bandwidth.rt_runtime == 0)
6360 retval = security_task_setscheduler(p, policy, param);
6366 * make sure no PI-waiters arrive (or leave) while we are
6367 * changing the priority of the task:
6369 raw_spin_lock_irqsave(&p->pi_lock, flags);
6371 * To be able to change p->policy safely, the apropriate
6372 * runqueue lock must be held.
6374 rq = __task_rq_lock(p);
6375 /* recheck policy now with rq lock held */
6376 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6377 policy = oldpolicy = -1;
6378 __task_rq_unlock(rq);
6379 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6382 update_rq_clock(rq);
6383 on_rq = p->se.on_rq;
6384 running = task_current(rq, p);
6386 deactivate_task(rq, p, 0);
6388 p->sched_class->put_prev_task(rq, p);
6390 p->sched_reset_on_fork = reset_on_fork;
6393 __setscheduler(rq, p, policy, param->sched_priority);
6396 p->sched_class->set_curr_task(rq);
6398 activate_task(rq, p, 0);
6400 check_class_changed(rq, p, prev_class, oldprio, running);
6402 __task_rq_unlock(rq);
6403 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6405 rt_mutex_adjust_pi(p);
6411 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6412 * @p: the task in question.
6413 * @policy: new policy.
6414 * @param: structure containing the new RT priority.
6416 * NOTE that the task may be already dead.
6418 int sched_setscheduler(struct task_struct *p, int policy,
6419 struct sched_param *param)
6421 return __sched_setscheduler(p, policy, param, true);
6423 EXPORT_SYMBOL_GPL(sched_setscheduler);
6426 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6427 * @p: the task in question.
6428 * @policy: new policy.
6429 * @param: structure containing the new RT priority.
6431 * Just like sched_setscheduler, only don't bother checking if the
6432 * current context has permission. For example, this is needed in
6433 * stop_machine(): we create temporary high priority worker threads,
6434 * but our caller might not have that capability.
6436 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6437 struct sched_param *param)
6439 return __sched_setscheduler(p, policy, param, false);
6443 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6445 struct sched_param lparam;
6446 struct task_struct *p;
6449 if (!param || pid < 0)
6451 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6456 p = find_process_by_pid(pid);
6458 retval = sched_setscheduler(p, policy, &lparam);
6465 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6466 * @pid: the pid in question.
6467 * @policy: new policy.
6468 * @param: structure containing the new RT priority.
6470 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6471 struct sched_param __user *, param)
6473 /* negative values for policy are not valid */
6477 return do_sched_setscheduler(pid, policy, param);
6481 * sys_sched_setparam - set/change the RT priority of a thread
6482 * @pid: the pid in question.
6483 * @param: structure containing the new RT priority.
6485 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6487 return do_sched_setscheduler(pid, -1, param);
6491 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6492 * @pid: the pid in question.
6494 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6496 struct task_struct *p;
6504 p = find_process_by_pid(pid);
6506 retval = security_task_getscheduler(p);
6509 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6516 * sys_sched_getparam - get the RT priority of a thread
6517 * @pid: the pid in question.
6518 * @param: structure containing the RT priority.
6520 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6522 struct sched_param lp;
6523 struct task_struct *p;
6526 if (!param || pid < 0)
6530 p = find_process_by_pid(pid);
6535 retval = security_task_getscheduler(p);
6539 lp.sched_priority = p->rt_priority;
6543 * This one might sleep, we cannot do it with a spinlock held ...
6545 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6554 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6556 cpumask_var_t cpus_allowed, new_mask;
6557 struct task_struct *p;
6563 p = find_process_by_pid(pid);
6570 /* Prevent p going away */
6574 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6578 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6580 goto out_free_cpus_allowed;
6583 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6586 retval = security_task_setscheduler(p, 0, NULL);
6590 cpuset_cpus_allowed(p, cpus_allowed);
6591 cpumask_and(new_mask, in_mask, cpus_allowed);
6593 retval = set_cpus_allowed_ptr(p, new_mask);
6596 cpuset_cpus_allowed(p, cpus_allowed);
6597 if (!cpumask_subset(new_mask, cpus_allowed)) {
6599 * We must have raced with a concurrent cpuset
6600 * update. Just reset the cpus_allowed to the
6601 * cpuset's cpus_allowed
6603 cpumask_copy(new_mask, cpus_allowed);
6608 free_cpumask_var(new_mask);
6609 out_free_cpus_allowed:
6610 free_cpumask_var(cpus_allowed);
6617 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6618 struct cpumask *new_mask)
6620 if (len < cpumask_size())
6621 cpumask_clear(new_mask);
6622 else if (len > cpumask_size())
6623 len = cpumask_size();
6625 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6629 * sys_sched_setaffinity - set the cpu affinity of a process
6630 * @pid: pid of the process
6631 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6632 * @user_mask_ptr: user-space pointer to the new cpu mask
6634 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6635 unsigned long __user *, user_mask_ptr)
6637 cpumask_var_t new_mask;
6640 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6643 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6645 retval = sched_setaffinity(pid, new_mask);
6646 free_cpumask_var(new_mask);
6650 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6652 struct task_struct *p;
6653 unsigned long flags;
6661 p = find_process_by_pid(pid);
6665 retval = security_task_getscheduler(p);
6669 rq = task_rq_lock(p, &flags);
6670 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6671 task_rq_unlock(rq, &flags);
6681 * sys_sched_getaffinity - get the cpu affinity of a process
6682 * @pid: pid of the process
6683 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6684 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6686 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6687 unsigned long __user *, user_mask_ptr)
6692 if (len < cpumask_size())
6695 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6698 ret = sched_getaffinity(pid, mask);
6700 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6703 ret = cpumask_size();
6705 free_cpumask_var(mask);
6711 * sys_sched_yield - yield the current processor to other threads.
6713 * This function yields the current CPU to other tasks. If there are no
6714 * other threads running on this CPU then this function will return.
6716 SYSCALL_DEFINE0(sched_yield)
6718 struct rq *rq = this_rq_lock();
6720 schedstat_inc(rq, yld_count);
6721 current->sched_class->yield_task(rq);
6724 * Since we are going to call schedule() anyway, there's
6725 * no need to preempt or enable interrupts:
6727 __release(rq->lock);
6728 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6729 do_raw_spin_unlock(&rq->lock);
6730 preempt_enable_no_resched();
6737 static inline int should_resched(void)
6739 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6742 static void __cond_resched(void)
6744 add_preempt_count(PREEMPT_ACTIVE);
6746 sub_preempt_count(PREEMPT_ACTIVE);
6749 int __sched _cond_resched(void)
6751 if (should_resched()) {
6757 EXPORT_SYMBOL(_cond_resched);
6760 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6761 * call schedule, and on return reacquire the lock.
6763 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6764 * operations here to prevent schedule() from being called twice (once via
6765 * spin_unlock(), once by hand).
6767 int __cond_resched_lock(spinlock_t *lock)
6769 int resched = should_resched();
6772 lockdep_assert_held(lock);
6774 if (spin_needbreak(lock) || resched) {
6785 EXPORT_SYMBOL(__cond_resched_lock);
6787 int __sched __cond_resched_softirq(void)
6789 BUG_ON(!in_softirq());
6791 if (should_resched()) {
6799 EXPORT_SYMBOL(__cond_resched_softirq);
6802 * yield - yield the current processor to other threads.
6804 * This is a shortcut for kernel-space yielding - it marks the
6805 * thread runnable and calls sys_sched_yield().
6807 void __sched yield(void)
6809 set_current_state(TASK_RUNNING);
6812 EXPORT_SYMBOL(yield);
6815 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6816 * that process accounting knows that this is a task in IO wait state.
6818 void __sched io_schedule(void)
6820 struct rq *rq = raw_rq();
6822 delayacct_blkio_start();
6823 atomic_inc(&rq->nr_iowait);
6824 current->in_iowait = 1;
6826 current->in_iowait = 0;
6827 atomic_dec(&rq->nr_iowait);
6828 delayacct_blkio_end();
6830 EXPORT_SYMBOL(io_schedule);
6832 long __sched io_schedule_timeout(long timeout)
6834 struct rq *rq = raw_rq();
6837 delayacct_blkio_start();
6838 atomic_inc(&rq->nr_iowait);
6839 current->in_iowait = 1;
6840 ret = schedule_timeout(timeout);
6841 current->in_iowait = 0;
6842 atomic_dec(&rq->nr_iowait);
6843 delayacct_blkio_end();
6848 * sys_sched_get_priority_max - return maximum RT priority.
6849 * @policy: scheduling class.
6851 * this syscall returns the maximum rt_priority that can be used
6852 * by a given scheduling class.
6854 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6861 ret = MAX_USER_RT_PRIO-1;
6873 * sys_sched_get_priority_min - return minimum RT priority.
6874 * @policy: scheduling class.
6876 * this syscall returns the minimum rt_priority that can be used
6877 * by a given scheduling class.
6879 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6897 * sys_sched_rr_get_interval - return the default timeslice of a process.
6898 * @pid: pid of the process.
6899 * @interval: userspace pointer to the timeslice value.
6901 * this syscall writes the default timeslice value of a given process
6902 * into the user-space timespec buffer. A value of '0' means infinity.
6904 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6905 struct timespec __user *, interval)
6907 struct task_struct *p;
6908 unsigned int time_slice;
6909 unsigned long flags;
6919 p = find_process_by_pid(pid);
6923 retval = security_task_getscheduler(p);
6927 rq = task_rq_lock(p, &flags);
6928 time_slice = p->sched_class->get_rr_interval(rq, p);
6929 task_rq_unlock(rq, &flags);
6932 jiffies_to_timespec(time_slice, &t);
6933 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6941 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6943 void sched_show_task(struct task_struct *p)
6945 unsigned long free = 0;
6948 state = p->state ? __ffs(p->state) + 1 : 0;
6949 pr_info("%-13.13s %c", p->comm,
6950 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6951 #if BITS_PER_LONG == 32
6952 if (state == TASK_RUNNING)
6953 pr_cont(" running ");
6955 pr_cont(" %08lx ", thread_saved_pc(p));
6957 if (state == TASK_RUNNING)
6958 pr_cont(" running task ");
6960 pr_cont(" %016lx ", thread_saved_pc(p));
6962 #ifdef CONFIG_DEBUG_STACK_USAGE
6963 free = stack_not_used(p);
6965 pr_cont("%5lu %5d %6d 0x%08lx\n", free,
6966 task_pid_nr(p), task_pid_nr(p->real_parent),
6967 (unsigned long)task_thread_info(p)->flags);
6969 show_stack(p, NULL);
6972 void show_state_filter(unsigned long state_filter)
6974 struct task_struct *g, *p;
6976 #if BITS_PER_LONG == 32
6977 pr_info(" task PC stack pid father\n");
6979 pr_info(" task PC stack pid father\n");
6981 read_lock(&tasklist_lock);
6982 do_each_thread(g, p) {
6984 * reset the NMI-timeout, listing all files on a slow
6985 * console might take alot of time:
6987 touch_nmi_watchdog();
6988 if (!state_filter || (p->state & state_filter))
6990 } while_each_thread(g, p);
6992 touch_all_softlockup_watchdogs();
6994 #ifdef CONFIG_SCHED_DEBUG
6995 sysrq_sched_debug_show();
6997 read_unlock(&tasklist_lock);
6999 * Only show locks if all tasks are dumped:
7002 debug_show_all_locks();
7005 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
7007 idle->sched_class = &idle_sched_class;
7011 * init_idle - set up an idle thread for a given CPU
7012 * @idle: task in question
7013 * @cpu: cpu the idle task belongs to
7015 * NOTE: this function does not set the idle thread's NEED_RESCHED
7016 * flag, to make booting more robust.
7018 void __cpuinit init_idle(struct task_struct *idle, int cpu)
7020 struct rq *rq = cpu_rq(cpu);
7021 unsigned long flags;
7023 raw_spin_lock_irqsave(&rq->lock, flags);
7026 idle->state = TASK_RUNNING;
7027 idle->se.exec_start = sched_clock();
7029 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
7030 __set_task_cpu(idle, cpu);
7032 rq->curr = rq->idle = idle;
7033 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7036 raw_spin_unlock_irqrestore(&rq->lock, flags);
7038 /* Set the preempt count _outside_ the spinlocks! */
7039 #if defined(CONFIG_PREEMPT)
7040 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7042 task_thread_info(idle)->preempt_count = 0;
7045 * The idle tasks have their own, simple scheduling class:
7047 idle->sched_class = &idle_sched_class;
7048 ftrace_graph_init_task(idle);
7052 * In a system that switches off the HZ timer nohz_cpu_mask
7053 * indicates which cpus entered this state. This is used
7054 * in the rcu update to wait only for active cpus. For system
7055 * which do not switch off the HZ timer nohz_cpu_mask should
7056 * always be CPU_BITS_NONE.
7058 cpumask_var_t nohz_cpu_mask;
7061 * Increase the granularity value when there are more CPUs,
7062 * because with more CPUs the 'effective latency' as visible
7063 * to users decreases. But the relationship is not linear,
7064 * so pick a second-best guess by going with the log2 of the
7067 * This idea comes from the SD scheduler of Con Kolivas:
7069 static int get_update_sysctl_factor(void)
7071 unsigned int cpus = min_t(int, num_online_cpus(), 8);
7072 unsigned int factor;
7074 switch (sysctl_sched_tunable_scaling) {
7075 case SCHED_TUNABLESCALING_NONE:
7078 case SCHED_TUNABLESCALING_LINEAR:
7081 case SCHED_TUNABLESCALING_LOG:
7083 factor = 1 + ilog2(cpus);
7090 static void update_sysctl(void)
7092 unsigned int factor = get_update_sysctl_factor();
7094 #define SET_SYSCTL(name) \
7095 (sysctl_##name = (factor) * normalized_sysctl_##name)
7096 SET_SYSCTL(sched_min_granularity);
7097 SET_SYSCTL(sched_latency);
7098 SET_SYSCTL(sched_wakeup_granularity);
7099 SET_SYSCTL(sched_shares_ratelimit);
7103 static inline void sched_init_granularity(void)
7110 * This is how migration works:
7112 * 1) we queue a struct migration_req structure in the source CPU's
7113 * runqueue and wake up that CPU's migration thread.
7114 * 2) we down() the locked semaphore => thread blocks.
7115 * 3) migration thread wakes up (implicitly it forces the migrated
7116 * thread off the CPU)
7117 * 4) it gets the migration request and checks whether the migrated
7118 * task is still in the wrong runqueue.
7119 * 5) if it's in the wrong runqueue then the migration thread removes
7120 * it and puts it into the right queue.
7121 * 6) migration thread up()s the semaphore.
7122 * 7) we wake up and the migration is done.
7126 * Change a given task's CPU affinity. Migrate the thread to a
7127 * proper CPU and schedule it away if the CPU it's executing on
7128 * is removed from the allowed bitmask.
7130 * NOTE: the caller must have a valid reference to the task, the
7131 * task must not exit() & deallocate itself prematurely. The
7132 * call is not atomic; no spinlocks may be held.
7134 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7136 struct migration_req req;
7137 unsigned long flags;
7142 * Since we rely on wake-ups to migrate sleeping tasks, don't change
7143 * the ->cpus_allowed mask from under waking tasks, which would be
7144 * possible when we change rq->lock in ttwu(), so synchronize against
7145 * TASK_WAKING to avoid that.
7148 while (p->state == TASK_WAKING)
7151 rq = task_rq_lock(p, &flags);
7153 if (p->state == TASK_WAKING) {
7154 task_rq_unlock(rq, &flags);
7158 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7163 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7164 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7169 if (p->sched_class->set_cpus_allowed)
7170 p->sched_class->set_cpus_allowed(p, new_mask);
7172 cpumask_copy(&p->cpus_allowed, new_mask);
7173 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7176 /* Can the task run on the task's current CPU? If so, we're done */
7177 if (cpumask_test_cpu(task_cpu(p), new_mask))
7180 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7181 /* Need help from migration thread: drop lock and wait. */
7182 struct task_struct *mt = rq->migration_thread;
7184 get_task_struct(mt);
7185 task_rq_unlock(rq, &flags);
7186 wake_up_process(rq->migration_thread);
7187 put_task_struct(mt);
7188 wait_for_completion(&req.done);
7189 tlb_migrate_finish(p->mm);
7193 task_rq_unlock(rq, &flags);
7197 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7200 * Move (not current) task off this cpu, onto dest cpu. We're doing
7201 * this because either it can't run here any more (set_cpus_allowed()
7202 * away from this CPU, or CPU going down), or because we're
7203 * attempting to rebalance this task on exec (sched_exec).
7205 * So we race with normal scheduler movements, but that's OK, as long
7206 * as the task is no longer on this CPU.
7208 * Returns non-zero if task was successfully migrated.
7210 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7212 struct rq *rq_dest, *rq_src;
7215 if (unlikely(!cpu_active(dest_cpu)))
7218 rq_src = cpu_rq(src_cpu);
7219 rq_dest = cpu_rq(dest_cpu);
7221 double_rq_lock(rq_src, rq_dest);
7222 /* Already moved. */
7223 if (task_cpu(p) != src_cpu)
7225 /* Affinity changed (again). */
7226 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7230 * If we're not on a rq, the next wake-up will ensure we're
7234 deactivate_task(rq_src, p, 0);
7235 set_task_cpu(p, dest_cpu);
7236 activate_task(rq_dest, p, 0);
7237 check_preempt_curr(rq_dest, p, 0);
7242 double_rq_unlock(rq_src, rq_dest);
7246 #define RCU_MIGRATION_IDLE 0
7247 #define RCU_MIGRATION_NEED_QS 1
7248 #define RCU_MIGRATION_GOT_QS 2
7249 #define RCU_MIGRATION_MUST_SYNC 3
7252 * migration_thread - this is a highprio system thread that performs
7253 * thread migration by bumping thread off CPU then 'pushing' onto
7256 static int migration_thread(void *data)
7259 int cpu = (long)data;
7263 BUG_ON(rq->migration_thread != current);
7265 set_current_state(TASK_INTERRUPTIBLE);
7266 while (!kthread_should_stop()) {
7267 struct migration_req *req;
7268 struct list_head *head;
7270 raw_spin_lock_irq(&rq->lock);
7272 if (cpu_is_offline(cpu)) {
7273 raw_spin_unlock_irq(&rq->lock);
7277 if (rq->active_balance) {
7278 active_load_balance(rq, cpu);
7279 rq->active_balance = 0;
7282 head = &rq->migration_queue;
7284 if (list_empty(head)) {
7285 raw_spin_unlock_irq(&rq->lock);
7287 set_current_state(TASK_INTERRUPTIBLE);
7290 req = list_entry(head->next, struct migration_req, list);
7291 list_del_init(head->next);
7293 if (req->task != NULL) {
7294 raw_spin_unlock(&rq->lock);
7295 __migrate_task(req->task, cpu, req->dest_cpu);
7296 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7297 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7298 raw_spin_unlock(&rq->lock);
7300 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7301 raw_spin_unlock(&rq->lock);
7302 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7306 complete(&req->done);
7308 __set_current_state(TASK_RUNNING);
7313 #ifdef CONFIG_HOTPLUG_CPU
7315 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7319 local_irq_disable();
7320 ret = __migrate_task(p, src_cpu, dest_cpu);
7326 * Figure out where task on dead CPU should go, use force if necessary.
7328 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7333 dest_cpu = select_fallback_rq(dead_cpu, p);
7335 /* It can have affinity changed while we were choosing. */
7336 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7341 * While a dead CPU has no uninterruptible tasks queued at this point,
7342 * it might still have a nonzero ->nr_uninterruptible counter, because
7343 * for performance reasons the counter is not stricly tracking tasks to
7344 * their home CPUs. So we just add the counter to another CPU's counter,
7345 * to keep the global sum constant after CPU-down:
7347 static void migrate_nr_uninterruptible(struct rq *rq_src)
7349 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7350 unsigned long flags;
7352 local_irq_save(flags);
7353 double_rq_lock(rq_src, rq_dest);
7354 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7355 rq_src->nr_uninterruptible = 0;
7356 double_rq_unlock(rq_src, rq_dest);
7357 local_irq_restore(flags);
7360 /* Run through task list and migrate tasks from the dead cpu. */
7361 static void migrate_live_tasks(int src_cpu)
7363 struct task_struct *p, *t;
7365 read_lock(&tasklist_lock);
7367 do_each_thread(t, p) {
7371 if (task_cpu(p) == src_cpu)
7372 move_task_off_dead_cpu(src_cpu, p);
7373 } while_each_thread(t, p);
7375 read_unlock(&tasklist_lock);
7379 * Schedules idle task to be the next runnable task on current CPU.
7380 * It does so by boosting its priority to highest possible.
7381 * Used by CPU offline code.
7383 void sched_idle_next(void)
7385 int this_cpu = smp_processor_id();
7386 struct rq *rq = cpu_rq(this_cpu);
7387 struct task_struct *p = rq->idle;
7388 unsigned long flags;
7390 /* cpu has to be offline */
7391 BUG_ON(cpu_online(this_cpu));
7394 * Strictly not necessary since rest of the CPUs are stopped by now
7395 * and interrupts disabled on the current cpu.
7397 raw_spin_lock_irqsave(&rq->lock, flags);
7399 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7401 update_rq_clock(rq);
7402 activate_task(rq, p, 0);
7404 raw_spin_unlock_irqrestore(&rq->lock, flags);
7408 * Ensures that the idle task is using init_mm right before its cpu goes
7411 void idle_task_exit(void)
7413 struct mm_struct *mm = current->active_mm;
7415 BUG_ON(cpu_online(smp_processor_id()));
7418 switch_mm(mm, &init_mm, current);
7422 /* called under rq->lock with disabled interrupts */
7423 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7425 struct rq *rq = cpu_rq(dead_cpu);
7427 /* Must be exiting, otherwise would be on tasklist. */
7428 BUG_ON(!p->exit_state);
7430 /* Cannot have done final schedule yet: would have vanished. */
7431 BUG_ON(p->state == TASK_DEAD);
7436 * Drop lock around migration; if someone else moves it,
7437 * that's OK. No task can be added to this CPU, so iteration is
7440 raw_spin_unlock_irq(&rq->lock);
7441 move_task_off_dead_cpu(dead_cpu, p);
7442 raw_spin_lock_irq(&rq->lock);
7447 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7448 static void migrate_dead_tasks(unsigned int dead_cpu)
7450 struct rq *rq = cpu_rq(dead_cpu);
7451 struct task_struct *next;
7454 if (!rq->nr_running)
7456 update_rq_clock(rq);
7457 next = pick_next_task(rq);
7460 next->sched_class->put_prev_task(rq, next);
7461 migrate_dead(dead_cpu, next);
7467 * remove the tasks which were accounted by rq from calc_load_tasks.
7469 static void calc_global_load_remove(struct rq *rq)
7471 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7472 rq->calc_load_active = 0;
7474 #endif /* CONFIG_HOTPLUG_CPU */
7476 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7478 static struct ctl_table sd_ctl_dir[] = {
7480 .procname = "sched_domain",
7486 static struct ctl_table sd_ctl_root[] = {
7488 .procname = "kernel",
7490 .child = sd_ctl_dir,
7495 static struct ctl_table *sd_alloc_ctl_entry(int n)
7497 struct ctl_table *entry =
7498 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7503 static void sd_free_ctl_entry(struct ctl_table **tablep)
7505 struct ctl_table *entry;
7508 * In the intermediate directories, both the child directory and
7509 * procname are dynamically allocated and could fail but the mode
7510 * will always be set. In the lowest directory the names are
7511 * static strings and all have proc handlers.
7513 for (entry = *tablep; entry->mode; entry++) {
7515 sd_free_ctl_entry(&entry->child);
7516 if (entry->proc_handler == NULL)
7517 kfree(entry->procname);
7525 set_table_entry(struct ctl_table *entry,
7526 const char *procname, void *data, int maxlen,
7527 mode_t mode, proc_handler *proc_handler)
7529 entry->procname = procname;
7531 entry->maxlen = maxlen;
7533 entry->proc_handler = proc_handler;
7536 static struct ctl_table *
7537 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7539 struct ctl_table *table = sd_alloc_ctl_entry(13);
7544 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7545 sizeof(long), 0644, proc_doulongvec_minmax);
7546 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7547 sizeof(long), 0644, proc_doulongvec_minmax);
7548 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7549 sizeof(int), 0644, proc_dointvec_minmax);
7550 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7551 sizeof(int), 0644, proc_dointvec_minmax);
7552 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7553 sizeof(int), 0644, proc_dointvec_minmax);
7554 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7555 sizeof(int), 0644, proc_dointvec_minmax);
7556 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7557 sizeof(int), 0644, proc_dointvec_minmax);
7558 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7559 sizeof(int), 0644, proc_dointvec_minmax);
7560 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7561 sizeof(int), 0644, proc_dointvec_minmax);
7562 set_table_entry(&table[9], "cache_nice_tries",
7563 &sd->cache_nice_tries,
7564 sizeof(int), 0644, proc_dointvec_minmax);
7565 set_table_entry(&table[10], "flags", &sd->flags,
7566 sizeof(int), 0644, proc_dointvec_minmax);
7567 set_table_entry(&table[11], "name", sd->name,
7568 CORENAME_MAX_SIZE, 0444, proc_dostring);
7569 /* &table[12] is terminator */
7574 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7576 struct ctl_table *entry, *table;
7577 struct sched_domain *sd;
7578 int domain_num = 0, i;
7581 for_each_domain(cpu, sd)
7583 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7588 for_each_domain(cpu, sd) {
7589 snprintf(buf, 32, "domain%d", i);
7590 entry->procname = kstrdup(buf, GFP_KERNEL);
7592 entry->child = sd_alloc_ctl_domain_table(sd);
7599 static struct ctl_table_header *sd_sysctl_header;
7600 static void register_sched_domain_sysctl(void)
7602 int i, cpu_num = num_possible_cpus();
7603 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7606 WARN_ON(sd_ctl_dir[0].child);
7607 sd_ctl_dir[0].child = entry;
7612 for_each_possible_cpu(i) {
7613 snprintf(buf, 32, "cpu%d", i);
7614 entry->procname = kstrdup(buf, GFP_KERNEL);
7616 entry->child = sd_alloc_ctl_cpu_table(i);
7620 WARN_ON(sd_sysctl_header);
7621 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7624 /* may be called multiple times per register */
7625 static void unregister_sched_domain_sysctl(void)
7627 if (sd_sysctl_header)
7628 unregister_sysctl_table(sd_sysctl_header);
7629 sd_sysctl_header = NULL;
7630 if (sd_ctl_dir[0].child)
7631 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7634 static void register_sched_domain_sysctl(void)
7637 static void unregister_sched_domain_sysctl(void)
7642 static void set_rq_online(struct rq *rq)
7645 const struct sched_class *class;
7647 cpumask_set_cpu(rq->cpu, rq->rd->online);
7650 for_each_class(class) {
7651 if (class->rq_online)
7652 class->rq_online(rq);
7657 static void set_rq_offline(struct rq *rq)
7660 const struct sched_class *class;
7662 for_each_class(class) {
7663 if (class->rq_offline)
7664 class->rq_offline(rq);
7667 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7673 * migration_call - callback that gets triggered when a CPU is added.
7674 * Here we can start up the necessary migration thread for the new CPU.
7676 static int __cpuinit
7677 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7679 struct task_struct *p;
7680 int cpu = (long)hcpu;
7681 unsigned long flags;
7686 case CPU_UP_PREPARE:
7687 case CPU_UP_PREPARE_FROZEN:
7688 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7691 kthread_bind(p, cpu);
7692 /* Must be high prio: stop_machine expects to yield to it. */
7693 rq = task_rq_lock(p, &flags);
7694 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7695 task_rq_unlock(rq, &flags);
7697 cpu_rq(cpu)->migration_thread = p;
7698 rq->calc_load_update = calc_load_update;
7702 case CPU_ONLINE_FROZEN:
7703 /* Strictly unnecessary, as first user will wake it. */
7704 wake_up_process(cpu_rq(cpu)->migration_thread);
7706 /* Update our root-domain */
7708 raw_spin_lock_irqsave(&rq->lock, flags);
7710 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7714 raw_spin_unlock_irqrestore(&rq->lock, flags);
7717 #ifdef CONFIG_HOTPLUG_CPU
7718 case CPU_UP_CANCELED:
7719 case CPU_UP_CANCELED_FROZEN:
7720 if (!cpu_rq(cpu)->migration_thread)
7722 /* Unbind it from offline cpu so it can run. Fall thru. */
7723 kthread_bind(cpu_rq(cpu)->migration_thread,
7724 cpumask_any(cpu_online_mask));
7725 kthread_stop(cpu_rq(cpu)->migration_thread);
7726 put_task_struct(cpu_rq(cpu)->migration_thread);
7727 cpu_rq(cpu)->migration_thread = NULL;
7731 case CPU_DEAD_FROZEN:
7732 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7733 migrate_live_tasks(cpu);
7735 kthread_stop(rq->migration_thread);
7736 put_task_struct(rq->migration_thread);
7737 rq->migration_thread = NULL;
7738 /* Idle task back to normal (off runqueue, low prio) */
7739 raw_spin_lock_irq(&rq->lock);
7740 update_rq_clock(rq);
7741 deactivate_task(rq, rq->idle, 0);
7742 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7743 rq->idle->sched_class = &idle_sched_class;
7744 migrate_dead_tasks(cpu);
7745 raw_spin_unlock_irq(&rq->lock);
7747 migrate_nr_uninterruptible(rq);
7748 BUG_ON(rq->nr_running != 0);
7749 calc_global_load_remove(rq);
7751 * No need to migrate the tasks: it was best-effort if
7752 * they didn't take sched_hotcpu_mutex. Just wake up
7755 raw_spin_lock_irq(&rq->lock);
7756 while (!list_empty(&rq->migration_queue)) {
7757 struct migration_req *req;
7759 req = list_entry(rq->migration_queue.next,
7760 struct migration_req, list);
7761 list_del_init(&req->list);
7762 raw_spin_unlock_irq(&rq->lock);
7763 complete(&req->done);
7764 raw_spin_lock_irq(&rq->lock);
7766 raw_spin_unlock_irq(&rq->lock);
7770 case CPU_DYING_FROZEN:
7771 /* Update our root-domain */
7773 raw_spin_lock_irqsave(&rq->lock, flags);
7775 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7778 raw_spin_unlock_irqrestore(&rq->lock, flags);
7786 * Register at high priority so that task migration (migrate_all_tasks)
7787 * happens before everything else. This has to be lower priority than
7788 * the notifier in the perf_event subsystem, though.
7790 static struct notifier_block __cpuinitdata migration_notifier = {
7791 .notifier_call = migration_call,
7795 static int __init migration_init(void)
7797 void *cpu = (void *)(long)smp_processor_id();
7800 /* Start one for the boot CPU: */
7801 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7802 BUG_ON(err == NOTIFY_BAD);
7803 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7804 register_cpu_notifier(&migration_notifier);
7808 early_initcall(migration_init);
7813 #ifdef CONFIG_SCHED_DEBUG
7815 static __read_mostly int sched_domain_debug_enabled;
7817 static int __init sched_domain_debug_setup(char *str)
7819 sched_domain_debug_enabled = 1;
7823 early_param("sched_debug", sched_domain_debug_setup);
7825 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7826 struct cpumask *groupmask)
7828 struct sched_group *group = sd->groups;
7831 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7832 cpumask_clear(groupmask);
7834 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7836 if (!(sd->flags & SD_LOAD_BALANCE)) {
7837 pr_cont("does not load-balance\n");
7839 pr_err("ERROR: !SD_LOAD_BALANCE domain has parent\n");
7843 pr_cont("span %s level %s\n", str, sd->name);
7845 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7846 pr_err("ERROR: domain->span does not contain CPU%d\n", cpu);
7848 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7849 pr_err("ERROR: domain->groups does not contain CPU%d\n", cpu);
7852 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7856 pr_err("ERROR: group is NULL\n");
7860 if (!group->cpu_power) {
7862 pr_err("ERROR: domain->cpu_power not set\n");
7866 if (!cpumask_weight(sched_group_cpus(group))) {
7868 pr_err("ERROR: empty group\n");
7872 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7874 pr_err("ERROR: repeated CPUs\n");
7878 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7880 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7882 pr_cont(" %s", str);
7883 if (group->cpu_power != SCHED_LOAD_SCALE) {
7884 pr_cont(" (cpu_power = %d)", group->cpu_power);
7887 group = group->next;
7888 } while (group != sd->groups);
7891 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7892 pr_err("ERROR: groups don't span domain->span\n");
7895 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7896 pr_err("ERROR: parent span is not a superset of domain->span\n");
7900 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7902 cpumask_var_t groupmask;
7905 if (!sched_domain_debug_enabled)
7909 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7913 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7915 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7916 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7921 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7928 free_cpumask_var(groupmask);
7930 #else /* !CONFIG_SCHED_DEBUG */
7931 # define sched_domain_debug(sd, cpu) do { } while (0)
7932 #endif /* CONFIG_SCHED_DEBUG */
7934 static int sd_degenerate(struct sched_domain *sd)
7936 if (cpumask_weight(sched_domain_span(sd)) == 1)
7939 /* Following flags need at least 2 groups */
7940 if (sd->flags & (SD_LOAD_BALANCE |
7941 SD_BALANCE_NEWIDLE |
7945 SD_SHARE_PKG_RESOURCES)) {
7946 if (sd->groups != sd->groups->next)
7950 /* Following flags don't use groups */
7951 if (sd->flags & (SD_WAKE_AFFINE))
7958 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7960 unsigned long cflags = sd->flags, pflags = parent->flags;
7962 if (sd_degenerate(parent))
7965 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7968 /* Flags needing groups don't count if only 1 group in parent */
7969 if (parent->groups == parent->groups->next) {
7970 pflags &= ~(SD_LOAD_BALANCE |
7971 SD_BALANCE_NEWIDLE |
7975 SD_SHARE_PKG_RESOURCES);
7976 if (nr_node_ids == 1)
7977 pflags &= ~SD_SERIALIZE;
7979 if (~cflags & pflags)
7985 static void free_rootdomain(struct root_domain *rd)
7987 synchronize_sched();
7989 cpupri_cleanup(&rd->cpupri);
7991 free_cpumask_var(rd->rto_mask);
7992 free_cpumask_var(rd->online);
7993 free_cpumask_var(rd->span);
7997 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7999 struct root_domain *old_rd = NULL;
8000 unsigned long flags;
8002 raw_spin_lock_irqsave(&rq->lock, flags);
8007 if (cpumask_test_cpu(rq->cpu, old_rd->online))
8010 cpumask_clear_cpu(rq->cpu, old_rd->span);
8013 * If we dont want to free the old_rt yet then
8014 * set old_rd to NULL to skip the freeing later
8017 if (!atomic_dec_and_test(&old_rd->refcount))
8021 atomic_inc(&rd->refcount);
8024 cpumask_set_cpu(rq->cpu, rd->span);
8025 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
8028 raw_spin_unlock_irqrestore(&rq->lock, flags);
8031 free_rootdomain(old_rd);
8034 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8036 gfp_t gfp = GFP_KERNEL;
8038 memset(rd, 0, sizeof(*rd));
8043 if (!alloc_cpumask_var(&rd->span, gfp))
8045 if (!alloc_cpumask_var(&rd->online, gfp))
8047 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8050 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8055 free_cpumask_var(rd->rto_mask);
8057 free_cpumask_var(rd->online);
8059 free_cpumask_var(rd->span);
8064 static void init_defrootdomain(void)
8066 init_rootdomain(&def_root_domain, true);
8068 atomic_set(&def_root_domain.refcount, 1);
8071 static struct root_domain *alloc_rootdomain(void)
8073 struct root_domain *rd;
8075 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8079 if (init_rootdomain(rd, false) != 0) {
8088 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8089 * hold the hotplug lock.
8092 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8094 struct rq *rq = cpu_rq(cpu);
8095 struct sched_domain *tmp;
8097 /* Remove the sched domains which do not contribute to scheduling. */
8098 for (tmp = sd; tmp; ) {
8099 struct sched_domain *parent = tmp->parent;
8103 if (sd_parent_degenerate(tmp, parent)) {
8104 tmp->parent = parent->parent;
8106 parent->parent->child = tmp;
8111 if (sd && sd_degenerate(sd)) {
8117 sched_domain_debug(sd, cpu);
8119 rq_attach_root(rq, rd);
8120 rcu_assign_pointer(rq->sd, sd);
8123 /* cpus with isolated domains */
8124 static cpumask_var_t cpu_isolated_map;
8126 /* Setup the mask of cpus configured for isolated domains */
8127 static int __init isolated_cpu_setup(char *str)
8129 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8130 cpulist_parse(str, cpu_isolated_map);
8134 __setup("isolcpus=", isolated_cpu_setup);
8137 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8138 * to a function which identifies what group(along with sched group) a CPU
8139 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8140 * (due to the fact that we keep track of groups covered with a struct cpumask).
8142 * init_sched_build_groups will build a circular linked list of the groups
8143 * covered by the given span, and will set each group's ->cpumask correctly,
8144 * and ->cpu_power to 0.
8147 init_sched_build_groups(const struct cpumask *span,
8148 const struct cpumask *cpu_map,
8149 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8150 struct sched_group **sg,
8151 struct cpumask *tmpmask),
8152 struct cpumask *covered, struct cpumask *tmpmask)
8154 struct sched_group *first = NULL, *last = NULL;
8157 cpumask_clear(covered);
8159 for_each_cpu(i, span) {
8160 struct sched_group *sg;
8161 int group = group_fn(i, cpu_map, &sg, tmpmask);
8164 if (cpumask_test_cpu(i, covered))
8167 cpumask_clear(sched_group_cpus(sg));
8170 for_each_cpu(j, span) {
8171 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8174 cpumask_set_cpu(j, covered);
8175 cpumask_set_cpu(j, sched_group_cpus(sg));
8186 #define SD_NODES_PER_DOMAIN 16
8191 * find_next_best_node - find the next node to include in a sched_domain
8192 * @node: node whose sched_domain we're building
8193 * @used_nodes: nodes already in the sched_domain
8195 * Find the next node to include in a given scheduling domain. Simply
8196 * finds the closest node not already in the @used_nodes map.
8198 * Should use nodemask_t.
8200 static int find_next_best_node(int node, nodemask_t *used_nodes)
8202 int i, n, val, min_val, best_node = 0;
8206 for (i = 0; i < nr_node_ids; i++) {
8207 /* Start at @node */
8208 n = (node + i) % nr_node_ids;
8210 if (!nr_cpus_node(n))
8213 /* Skip already used nodes */
8214 if (node_isset(n, *used_nodes))
8217 /* Simple min distance search */
8218 val = node_distance(node, n);
8220 if (val < min_val) {
8226 node_set(best_node, *used_nodes);
8231 * sched_domain_node_span - get a cpumask for a node's sched_domain
8232 * @node: node whose cpumask we're constructing
8233 * @span: resulting cpumask
8235 * Given a node, construct a good cpumask for its sched_domain to span. It
8236 * should be one that prevents unnecessary balancing, but also spreads tasks
8239 static void sched_domain_node_span(int node, struct cpumask *span)
8241 nodemask_t used_nodes;
8244 cpumask_clear(span);
8245 nodes_clear(used_nodes);
8247 cpumask_or(span, span, cpumask_of_node(node));
8248 node_set(node, used_nodes);
8250 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8251 int next_node = find_next_best_node(node, &used_nodes);
8253 cpumask_or(span, span, cpumask_of_node(next_node));
8256 #endif /* CONFIG_NUMA */
8258 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8261 * The cpus mask in sched_group and sched_domain hangs off the end.
8263 * ( See the the comments in include/linux/sched.h:struct sched_group
8264 * and struct sched_domain. )
8266 struct static_sched_group {
8267 struct sched_group sg;
8268 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8271 struct static_sched_domain {
8272 struct sched_domain sd;
8273 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8279 cpumask_var_t domainspan;
8280 cpumask_var_t covered;
8281 cpumask_var_t notcovered;
8283 cpumask_var_t nodemask;
8284 cpumask_var_t this_sibling_map;
8285 cpumask_var_t this_core_map;
8286 cpumask_var_t send_covered;
8287 cpumask_var_t tmpmask;
8288 struct sched_group **sched_group_nodes;
8289 struct root_domain *rd;
8293 sa_sched_groups = 0,
8298 sa_this_sibling_map,
8300 sa_sched_group_nodes,
8310 * SMT sched-domains:
8312 #ifdef CONFIG_SCHED_SMT
8313 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8314 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
8317 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8318 struct sched_group **sg, struct cpumask *unused)
8321 *sg = &per_cpu(sched_groups, cpu).sg;
8324 #endif /* CONFIG_SCHED_SMT */
8327 * multi-core sched-domains:
8329 #ifdef CONFIG_SCHED_MC
8330 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8331 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8332 #endif /* CONFIG_SCHED_MC */
8334 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8336 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8337 struct sched_group **sg, struct cpumask *mask)
8341 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8342 group = cpumask_first(mask);
8344 *sg = &per_cpu(sched_group_core, group).sg;
8347 #elif defined(CONFIG_SCHED_MC)
8349 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8350 struct sched_group **sg, struct cpumask *unused)
8353 *sg = &per_cpu(sched_group_core, cpu).sg;
8358 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8359 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8362 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8363 struct sched_group **sg, struct cpumask *mask)
8366 #ifdef CONFIG_SCHED_MC
8367 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8368 group = cpumask_first(mask);
8369 #elif defined(CONFIG_SCHED_SMT)
8370 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8371 group = cpumask_first(mask);
8376 *sg = &per_cpu(sched_group_phys, group).sg;
8382 * The init_sched_build_groups can't handle what we want to do with node
8383 * groups, so roll our own. Now each node has its own list of groups which
8384 * gets dynamically allocated.
8386 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8387 static struct sched_group ***sched_group_nodes_bycpu;
8389 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8390 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8392 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8393 struct sched_group **sg,
8394 struct cpumask *nodemask)
8398 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8399 group = cpumask_first(nodemask);
8402 *sg = &per_cpu(sched_group_allnodes, group).sg;
8406 static void init_numa_sched_groups_power(struct sched_group *group_head)
8408 struct sched_group *sg = group_head;
8414 for_each_cpu(j, sched_group_cpus(sg)) {
8415 struct sched_domain *sd;
8417 sd = &per_cpu(phys_domains, j).sd;
8418 if (j != group_first_cpu(sd->groups)) {
8420 * Only add "power" once for each
8426 sg->cpu_power += sd->groups->cpu_power;
8429 } while (sg != group_head);
8432 static int build_numa_sched_groups(struct s_data *d,
8433 const struct cpumask *cpu_map, int num)
8435 struct sched_domain *sd;
8436 struct sched_group *sg, *prev;
8439 cpumask_clear(d->covered);
8440 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8441 if (cpumask_empty(d->nodemask)) {
8442 d->sched_group_nodes[num] = NULL;
8446 sched_domain_node_span(num, d->domainspan);
8447 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8449 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8452 pr_warning("Can not alloc domain group for node %d\n", num);
8455 d->sched_group_nodes[num] = sg;
8457 for_each_cpu(j, d->nodemask) {
8458 sd = &per_cpu(node_domains, j).sd;
8463 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8465 cpumask_or(d->covered, d->covered, d->nodemask);
8468 for (j = 0; j < nr_node_ids; j++) {
8469 n = (num + j) % nr_node_ids;
8470 cpumask_complement(d->notcovered, d->covered);
8471 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8472 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8473 if (cpumask_empty(d->tmpmask))
8475 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8476 if (cpumask_empty(d->tmpmask))
8478 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8481 pr_warning("Can not alloc domain group for node %d\n",
8486 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8487 sg->next = prev->next;
8488 cpumask_or(d->covered, d->covered, d->tmpmask);
8495 #endif /* CONFIG_NUMA */
8498 /* Free memory allocated for various sched_group structures */
8499 static void free_sched_groups(const struct cpumask *cpu_map,
8500 struct cpumask *nodemask)
8504 for_each_cpu(cpu, cpu_map) {
8505 struct sched_group **sched_group_nodes
8506 = sched_group_nodes_bycpu[cpu];
8508 if (!sched_group_nodes)
8511 for (i = 0; i < nr_node_ids; i++) {
8512 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8514 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8515 if (cpumask_empty(nodemask))
8525 if (oldsg != sched_group_nodes[i])
8528 kfree(sched_group_nodes);
8529 sched_group_nodes_bycpu[cpu] = NULL;
8532 #else /* !CONFIG_NUMA */
8533 static void free_sched_groups(const struct cpumask *cpu_map,
8534 struct cpumask *nodemask)
8537 #endif /* CONFIG_NUMA */
8540 * Initialize sched groups cpu_power.
8542 * cpu_power indicates the capacity of sched group, which is used while
8543 * distributing the load between different sched groups in a sched domain.
8544 * Typically cpu_power for all the groups in a sched domain will be same unless
8545 * there are asymmetries in the topology. If there are asymmetries, group
8546 * having more cpu_power will pickup more load compared to the group having
8549 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8551 struct sched_domain *child;
8552 struct sched_group *group;
8556 WARN_ON(!sd || !sd->groups);
8558 if (cpu != group_first_cpu(sd->groups))
8563 sd->groups->cpu_power = 0;
8566 power = SCHED_LOAD_SCALE;
8567 weight = cpumask_weight(sched_domain_span(sd));
8569 * SMT siblings share the power of a single core.
8570 * Usually multiple threads get a better yield out of
8571 * that one core than a single thread would have,
8572 * reflect that in sd->smt_gain.
8574 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8575 power *= sd->smt_gain;
8577 power >>= SCHED_LOAD_SHIFT;
8579 sd->groups->cpu_power += power;
8584 * Add cpu_power of each child group to this groups cpu_power.
8586 group = child->groups;
8588 sd->groups->cpu_power += group->cpu_power;
8589 group = group->next;
8590 } while (group != child->groups);
8594 * Initializers for schedule domains
8595 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8598 #ifdef CONFIG_SCHED_DEBUG
8599 # define SD_INIT_NAME(sd, type) sd->name = #type
8601 # define SD_INIT_NAME(sd, type) do { } while (0)
8604 #define SD_INIT(sd, type) sd_init_##type(sd)
8606 #define SD_INIT_FUNC(type) \
8607 static noinline void sd_init_##type(struct sched_domain *sd) \
8609 memset(sd, 0, sizeof(*sd)); \
8610 *sd = SD_##type##_INIT; \
8611 sd->level = SD_LV_##type; \
8612 SD_INIT_NAME(sd, type); \
8617 SD_INIT_FUNC(ALLNODES)
8620 #ifdef CONFIG_SCHED_SMT
8621 SD_INIT_FUNC(SIBLING)
8623 #ifdef CONFIG_SCHED_MC
8627 static int default_relax_domain_level = -1;
8629 static int __init setup_relax_domain_level(char *str)
8633 val = simple_strtoul(str, NULL, 0);
8634 if (val < SD_LV_MAX)
8635 default_relax_domain_level = val;
8639 __setup("relax_domain_level=", setup_relax_domain_level);
8641 static void set_domain_attribute(struct sched_domain *sd,
8642 struct sched_domain_attr *attr)
8646 if (!attr || attr->relax_domain_level < 0) {
8647 if (default_relax_domain_level < 0)
8650 request = default_relax_domain_level;
8652 request = attr->relax_domain_level;
8653 if (request < sd->level) {
8654 /* turn off idle balance on this domain */
8655 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8657 /* turn on idle balance on this domain */
8658 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8662 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8663 const struct cpumask *cpu_map)
8666 case sa_sched_groups:
8667 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8668 d->sched_group_nodes = NULL;
8670 free_rootdomain(d->rd); /* fall through */
8672 free_cpumask_var(d->tmpmask); /* fall through */
8673 case sa_send_covered:
8674 free_cpumask_var(d->send_covered); /* fall through */
8675 case sa_this_core_map:
8676 free_cpumask_var(d->this_core_map); /* fall through */
8677 case sa_this_sibling_map:
8678 free_cpumask_var(d->this_sibling_map); /* fall through */
8680 free_cpumask_var(d->nodemask); /* fall through */
8681 case sa_sched_group_nodes:
8683 kfree(d->sched_group_nodes); /* fall through */
8685 free_cpumask_var(d->notcovered); /* fall through */
8687 free_cpumask_var(d->covered); /* fall through */
8689 free_cpumask_var(d->domainspan); /* fall through */
8696 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8697 const struct cpumask *cpu_map)
8700 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8702 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8703 return sa_domainspan;
8704 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8706 /* Allocate the per-node list of sched groups */
8707 d->sched_group_nodes = kcalloc(nr_node_ids,
8708 sizeof(struct sched_group *), GFP_KERNEL);
8709 if (!d->sched_group_nodes) {
8710 pr_warning("Can not alloc sched group node list\n");
8711 return sa_notcovered;
8713 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8715 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8716 return sa_sched_group_nodes;
8717 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8719 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8720 return sa_this_sibling_map;
8721 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8722 return sa_this_core_map;
8723 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8724 return sa_send_covered;
8725 d->rd = alloc_rootdomain();
8727 pr_warning("Cannot alloc root domain\n");
8730 return sa_rootdomain;
8733 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8734 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8736 struct sched_domain *sd = NULL;
8738 struct sched_domain *parent;
8741 if (cpumask_weight(cpu_map) >
8742 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8743 sd = &per_cpu(allnodes_domains, i).sd;
8744 SD_INIT(sd, ALLNODES);
8745 set_domain_attribute(sd, attr);
8746 cpumask_copy(sched_domain_span(sd), cpu_map);
8747 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8752 sd = &per_cpu(node_domains, i).sd;
8754 set_domain_attribute(sd, attr);
8755 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8756 sd->parent = parent;
8759 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8764 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8765 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8766 struct sched_domain *parent, int i)
8768 struct sched_domain *sd;
8769 sd = &per_cpu(phys_domains, i).sd;
8771 set_domain_attribute(sd, attr);
8772 cpumask_copy(sched_domain_span(sd), d->nodemask);
8773 sd->parent = parent;
8776 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8780 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8781 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8782 struct sched_domain *parent, int i)
8784 struct sched_domain *sd = parent;
8785 #ifdef CONFIG_SCHED_MC
8786 sd = &per_cpu(core_domains, i).sd;
8788 set_domain_attribute(sd, attr);
8789 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8790 sd->parent = parent;
8792 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8797 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8798 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8799 struct sched_domain *parent, int i)
8801 struct sched_domain *sd = parent;
8802 #ifdef CONFIG_SCHED_SMT
8803 sd = &per_cpu(cpu_domains, i).sd;
8804 SD_INIT(sd, SIBLING);
8805 set_domain_attribute(sd, attr);
8806 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8807 sd->parent = parent;
8809 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8814 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8815 const struct cpumask *cpu_map, int cpu)
8818 #ifdef CONFIG_SCHED_SMT
8819 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8820 cpumask_and(d->this_sibling_map, cpu_map,
8821 topology_thread_cpumask(cpu));
8822 if (cpu == cpumask_first(d->this_sibling_map))
8823 init_sched_build_groups(d->this_sibling_map, cpu_map,
8825 d->send_covered, d->tmpmask);
8828 #ifdef CONFIG_SCHED_MC
8829 case SD_LV_MC: /* set up multi-core groups */
8830 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8831 if (cpu == cpumask_first(d->this_core_map))
8832 init_sched_build_groups(d->this_core_map, cpu_map,
8834 d->send_covered, d->tmpmask);
8837 case SD_LV_CPU: /* set up physical groups */
8838 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8839 if (!cpumask_empty(d->nodemask))
8840 init_sched_build_groups(d->nodemask, cpu_map,
8842 d->send_covered, d->tmpmask);
8845 case SD_LV_ALLNODES:
8846 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8847 d->send_covered, d->tmpmask);
8856 * Build sched domains for a given set of cpus and attach the sched domains
8857 * to the individual cpus
8859 static int __build_sched_domains(const struct cpumask *cpu_map,
8860 struct sched_domain_attr *attr)
8862 enum s_alloc alloc_state = sa_none;
8864 struct sched_domain *sd;
8870 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8871 if (alloc_state != sa_rootdomain)
8873 alloc_state = sa_sched_groups;
8876 * Set up domains for cpus specified by the cpu_map.
8878 for_each_cpu(i, cpu_map) {
8879 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8882 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8883 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8884 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8885 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8888 for_each_cpu(i, cpu_map) {
8889 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8890 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8893 /* Set up physical groups */
8894 for (i = 0; i < nr_node_ids; i++)
8895 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8898 /* Set up node groups */
8900 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8902 for (i = 0; i < nr_node_ids; i++)
8903 if (build_numa_sched_groups(&d, cpu_map, i))
8907 /* Calculate CPU power for physical packages and nodes */
8908 #ifdef CONFIG_SCHED_SMT
8909 for_each_cpu(i, cpu_map) {
8910 sd = &per_cpu(cpu_domains, i).sd;
8911 init_sched_groups_power(i, sd);
8914 #ifdef CONFIG_SCHED_MC
8915 for_each_cpu(i, cpu_map) {
8916 sd = &per_cpu(core_domains, i).sd;
8917 init_sched_groups_power(i, sd);
8921 for_each_cpu(i, cpu_map) {
8922 sd = &per_cpu(phys_domains, i).sd;
8923 init_sched_groups_power(i, sd);
8927 for (i = 0; i < nr_node_ids; i++)
8928 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8930 if (d.sd_allnodes) {
8931 struct sched_group *sg;
8933 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8935 init_numa_sched_groups_power(sg);
8939 /* Attach the domains */
8940 for_each_cpu(i, cpu_map) {
8941 #ifdef CONFIG_SCHED_SMT
8942 sd = &per_cpu(cpu_domains, i).sd;
8943 #elif defined(CONFIG_SCHED_MC)
8944 sd = &per_cpu(core_domains, i).sd;
8946 sd = &per_cpu(phys_domains, i).sd;
8948 cpu_attach_domain(sd, d.rd, i);
8951 d.sched_group_nodes = NULL; /* don't free this we still need it */
8952 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8956 __free_domain_allocs(&d, alloc_state, cpu_map);
8960 static int build_sched_domains(const struct cpumask *cpu_map)
8962 return __build_sched_domains(cpu_map, NULL);
8965 static cpumask_var_t *doms_cur; /* current sched domains */
8966 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8967 static struct sched_domain_attr *dattr_cur;
8968 /* attribues of custom domains in 'doms_cur' */
8971 * Special case: If a kmalloc of a doms_cur partition (array of
8972 * cpumask) fails, then fallback to a single sched domain,
8973 * as determined by the single cpumask fallback_doms.
8975 static cpumask_var_t fallback_doms;
8978 * arch_update_cpu_topology lets virtualized architectures update the
8979 * cpu core maps. It is supposed to return 1 if the topology changed
8980 * or 0 if it stayed the same.
8982 int __attribute__((weak)) arch_update_cpu_topology(void)
8987 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
8990 cpumask_var_t *doms;
8992 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
8995 for (i = 0; i < ndoms; i++) {
8996 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
8997 free_sched_domains(doms, i);
9004 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
9007 for (i = 0; i < ndoms; i++)
9008 free_cpumask_var(doms[i]);
9013 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9014 * For now this just excludes isolated cpus, but could be used to
9015 * exclude other special cases in the future.
9017 static int arch_init_sched_domains(const struct cpumask *cpu_map)
9021 arch_update_cpu_topology();
9023 doms_cur = alloc_sched_domains(ndoms_cur);
9025 doms_cur = &fallback_doms;
9026 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
9028 err = build_sched_domains(doms_cur[0]);
9029 register_sched_domain_sysctl();
9034 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9035 struct cpumask *tmpmask)
9037 free_sched_groups(cpu_map, tmpmask);
9041 * Detach sched domains from a group of cpus specified in cpu_map
9042 * These cpus will now be attached to the NULL domain
9044 static void detach_destroy_domains(const struct cpumask *cpu_map)
9046 /* Save because hotplug lock held. */
9047 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9050 for_each_cpu(i, cpu_map)
9051 cpu_attach_domain(NULL, &def_root_domain, i);
9052 synchronize_sched();
9053 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9056 /* handle null as "default" */
9057 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9058 struct sched_domain_attr *new, int idx_new)
9060 struct sched_domain_attr tmp;
9067 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9068 new ? (new + idx_new) : &tmp,
9069 sizeof(struct sched_domain_attr));
9073 * Partition sched domains as specified by the 'ndoms_new'
9074 * cpumasks in the array doms_new[] of cpumasks. This compares
9075 * doms_new[] to the current sched domain partitioning, doms_cur[].
9076 * It destroys each deleted domain and builds each new domain.
9078 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
9079 * The masks don't intersect (don't overlap.) We should setup one
9080 * sched domain for each mask. CPUs not in any of the cpumasks will
9081 * not be load balanced. If the same cpumask appears both in the
9082 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9085 * The passed in 'doms_new' should be allocated using
9086 * alloc_sched_domains. This routine takes ownership of it and will
9087 * free_sched_domains it when done with it. If the caller failed the
9088 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9089 * and partition_sched_domains() will fallback to the single partition
9090 * 'fallback_doms', it also forces the domains to be rebuilt.
9092 * If doms_new == NULL it will be replaced with cpu_online_mask.
9093 * ndoms_new == 0 is a special case for destroying existing domains,
9094 * and it will not create the default domain.
9096 * Call with hotplug lock held
9098 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
9099 struct sched_domain_attr *dattr_new)
9104 mutex_lock(&sched_domains_mutex);
9106 /* always unregister in case we don't destroy any domains */
9107 unregister_sched_domain_sysctl();
9109 /* Let architecture update cpu core mappings. */
9110 new_topology = arch_update_cpu_topology();
9112 n = doms_new ? ndoms_new : 0;
9114 /* Destroy deleted domains */
9115 for (i = 0; i < ndoms_cur; i++) {
9116 for (j = 0; j < n && !new_topology; j++) {
9117 if (cpumask_equal(doms_cur[i], doms_new[j])
9118 && dattrs_equal(dattr_cur, i, dattr_new, j))
9121 /* no match - a current sched domain not in new doms_new[] */
9122 detach_destroy_domains(doms_cur[i]);
9127 if (doms_new == NULL) {
9129 doms_new = &fallback_doms;
9130 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
9131 WARN_ON_ONCE(dattr_new);
9134 /* Build new domains */
9135 for (i = 0; i < ndoms_new; i++) {
9136 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9137 if (cpumask_equal(doms_new[i], doms_cur[j])
9138 && dattrs_equal(dattr_new, i, dattr_cur, j))
9141 /* no match - add a new doms_new */
9142 __build_sched_domains(doms_new[i],
9143 dattr_new ? dattr_new + i : NULL);
9148 /* Remember the new sched domains */
9149 if (doms_cur != &fallback_doms)
9150 free_sched_domains(doms_cur, ndoms_cur);
9151 kfree(dattr_cur); /* kfree(NULL) is safe */
9152 doms_cur = doms_new;
9153 dattr_cur = dattr_new;
9154 ndoms_cur = ndoms_new;
9156 register_sched_domain_sysctl();
9158 mutex_unlock(&sched_domains_mutex);
9161 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9162 static void arch_reinit_sched_domains(void)
9166 /* Destroy domains first to force the rebuild */
9167 partition_sched_domains(0, NULL, NULL);
9169 rebuild_sched_domains();
9173 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9175 unsigned int level = 0;
9177 if (sscanf(buf, "%u", &level) != 1)
9181 * level is always be positive so don't check for
9182 * level < POWERSAVINGS_BALANCE_NONE which is 0
9183 * What happens on 0 or 1 byte write,
9184 * need to check for count as well?
9187 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9191 sched_smt_power_savings = level;
9193 sched_mc_power_savings = level;
9195 arch_reinit_sched_domains();
9200 #ifdef CONFIG_SCHED_MC
9201 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9204 return sprintf(page, "%u\n", sched_mc_power_savings);
9206 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9207 const char *buf, size_t count)
9209 return sched_power_savings_store(buf, count, 0);
9211 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9212 sched_mc_power_savings_show,
9213 sched_mc_power_savings_store);
9216 #ifdef CONFIG_SCHED_SMT
9217 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9220 return sprintf(page, "%u\n", sched_smt_power_savings);
9222 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9223 const char *buf, size_t count)
9225 return sched_power_savings_store(buf, count, 1);
9227 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9228 sched_smt_power_savings_show,
9229 sched_smt_power_savings_store);
9232 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9236 #ifdef CONFIG_SCHED_SMT
9238 err = sysfs_create_file(&cls->kset.kobj,
9239 &attr_sched_smt_power_savings.attr);
9241 #ifdef CONFIG_SCHED_MC
9242 if (!err && mc_capable())
9243 err = sysfs_create_file(&cls->kset.kobj,
9244 &attr_sched_mc_power_savings.attr);
9248 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9250 #ifndef CONFIG_CPUSETS
9252 * Add online and remove offline CPUs from the scheduler domains.
9253 * When cpusets are enabled they take over this function.
9255 static int update_sched_domains(struct notifier_block *nfb,
9256 unsigned long action, void *hcpu)
9260 case CPU_ONLINE_FROZEN:
9261 case CPU_DOWN_PREPARE:
9262 case CPU_DOWN_PREPARE_FROZEN:
9263 case CPU_DOWN_FAILED:
9264 case CPU_DOWN_FAILED_FROZEN:
9265 partition_sched_domains(1, NULL, NULL);
9274 static int update_runtime(struct notifier_block *nfb,
9275 unsigned long action, void *hcpu)
9277 int cpu = (int)(long)hcpu;
9280 case CPU_DOWN_PREPARE:
9281 case CPU_DOWN_PREPARE_FROZEN:
9282 disable_runtime(cpu_rq(cpu));
9285 case CPU_DOWN_FAILED:
9286 case CPU_DOWN_FAILED_FROZEN:
9288 case CPU_ONLINE_FROZEN:
9289 enable_runtime(cpu_rq(cpu));
9297 void __init sched_init_smp(void)
9299 cpumask_var_t non_isolated_cpus;
9301 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9302 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9304 #if defined(CONFIG_NUMA)
9305 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9307 BUG_ON(sched_group_nodes_bycpu == NULL);
9310 mutex_lock(&sched_domains_mutex);
9311 arch_init_sched_domains(cpu_active_mask);
9312 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9313 if (cpumask_empty(non_isolated_cpus))
9314 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9315 mutex_unlock(&sched_domains_mutex);
9318 #ifndef CONFIG_CPUSETS
9319 /* XXX: Theoretical race here - CPU may be hotplugged now */
9320 hotcpu_notifier(update_sched_domains, 0);
9323 /* RT runtime code needs to handle some hotplug events */
9324 hotcpu_notifier(update_runtime, 0);
9328 /* Move init over to a non-isolated CPU */
9329 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9331 sched_init_granularity();
9332 free_cpumask_var(non_isolated_cpus);
9334 init_sched_rt_class();
9337 void __init sched_init_smp(void)
9339 sched_init_granularity();
9341 #endif /* CONFIG_SMP */
9343 const_debug unsigned int sysctl_timer_migration = 1;
9345 int in_sched_functions(unsigned long addr)
9347 return in_lock_functions(addr) ||
9348 (addr >= (unsigned long)__sched_text_start
9349 && addr < (unsigned long)__sched_text_end);
9352 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9354 cfs_rq->tasks_timeline = RB_ROOT;
9355 INIT_LIST_HEAD(&cfs_rq->tasks);
9356 #ifdef CONFIG_FAIR_GROUP_SCHED
9359 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9362 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9364 struct rt_prio_array *array;
9367 array = &rt_rq->active;
9368 for (i = 0; i < MAX_RT_PRIO; i++) {
9369 INIT_LIST_HEAD(array->queue + i);
9370 __clear_bit(i, array->bitmap);
9372 /* delimiter for bitsearch: */
9373 __set_bit(MAX_RT_PRIO, array->bitmap);
9375 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9376 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9378 rt_rq->highest_prio.next = MAX_RT_PRIO;
9382 rt_rq->rt_nr_migratory = 0;
9383 rt_rq->overloaded = 0;
9384 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
9388 rt_rq->rt_throttled = 0;
9389 rt_rq->rt_runtime = 0;
9390 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
9392 #ifdef CONFIG_RT_GROUP_SCHED
9393 rt_rq->rt_nr_boosted = 0;
9398 #ifdef CONFIG_FAIR_GROUP_SCHED
9399 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9400 struct sched_entity *se, int cpu, int add,
9401 struct sched_entity *parent)
9403 struct rq *rq = cpu_rq(cpu);
9404 tg->cfs_rq[cpu] = cfs_rq;
9405 init_cfs_rq(cfs_rq, rq);
9408 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9411 /* se could be NULL for init_task_group */
9416 se->cfs_rq = &rq->cfs;
9418 se->cfs_rq = parent->my_q;
9421 se->load.weight = tg->shares;
9422 se->load.inv_weight = 0;
9423 se->parent = parent;
9427 #ifdef CONFIG_RT_GROUP_SCHED
9428 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9429 struct sched_rt_entity *rt_se, int cpu, int add,
9430 struct sched_rt_entity *parent)
9432 struct rq *rq = cpu_rq(cpu);
9434 tg->rt_rq[cpu] = rt_rq;
9435 init_rt_rq(rt_rq, rq);
9437 rt_rq->rt_se = rt_se;
9438 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9440 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9442 tg->rt_se[cpu] = rt_se;
9447 rt_se->rt_rq = &rq->rt;
9449 rt_se->rt_rq = parent->my_q;
9451 rt_se->my_q = rt_rq;
9452 rt_se->parent = parent;
9453 INIT_LIST_HEAD(&rt_se->run_list);
9457 void __init sched_init(void)
9460 unsigned long alloc_size = 0, ptr;
9462 #ifdef CONFIG_FAIR_GROUP_SCHED
9463 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9465 #ifdef CONFIG_RT_GROUP_SCHED
9466 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9468 #ifdef CONFIG_USER_SCHED
9471 #ifdef CONFIG_CPUMASK_OFFSTACK
9472 alloc_size += num_possible_cpus() * cpumask_size();
9475 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9477 #ifdef CONFIG_FAIR_GROUP_SCHED
9478 init_task_group.se = (struct sched_entity **)ptr;
9479 ptr += nr_cpu_ids * sizeof(void **);
9481 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9482 ptr += nr_cpu_ids * sizeof(void **);
9484 #ifdef CONFIG_USER_SCHED
9485 root_task_group.se = (struct sched_entity **)ptr;
9486 ptr += nr_cpu_ids * sizeof(void **);
9488 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9489 ptr += nr_cpu_ids * sizeof(void **);
9490 #endif /* CONFIG_USER_SCHED */
9491 #endif /* CONFIG_FAIR_GROUP_SCHED */
9492 #ifdef CONFIG_RT_GROUP_SCHED
9493 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9494 ptr += nr_cpu_ids * sizeof(void **);
9496 init_task_group.rt_rq = (struct rt_rq **)ptr;
9497 ptr += nr_cpu_ids * sizeof(void **);
9499 #ifdef CONFIG_USER_SCHED
9500 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9501 ptr += nr_cpu_ids * sizeof(void **);
9503 root_task_group.rt_rq = (struct rt_rq **)ptr;
9504 ptr += nr_cpu_ids * sizeof(void **);
9505 #endif /* CONFIG_USER_SCHED */
9506 #endif /* CONFIG_RT_GROUP_SCHED */
9507 #ifdef CONFIG_CPUMASK_OFFSTACK
9508 for_each_possible_cpu(i) {
9509 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9510 ptr += cpumask_size();
9512 #endif /* CONFIG_CPUMASK_OFFSTACK */
9516 init_defrootdomain();
9519 init_rt_bandwidth(&def_rt_bandwidth,
9520 global_rt_period(), global_rt_runtime());
9522 #ifdef CONFIG_RT_GROUP_SCHED
9523 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9524 global_rt_period(), global_rt_runtime());
9525 #ifdef CONFIG_USER_SCHED
9526 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9527 global_rt_period(), RUNTIME_INF);
9528 #endif /* CONFIG_USER_SCHED */
9529 #endif /* CONFIG_RT_GROUP_SCHED */
9531 #ifdef CONFIG_GROUP_SCHED
9532 list_add(&init_task_group.list, &task_groups);
9533 INIT_LIST_HEAD(&init_task_group.children);
9535 #ifdef CONFIG_USER_SCHED
9536 INIT_LIST_HEAD(&root_task_group.children);
9537 init_task_group.parent = &root_task_group;
9538 list_add(&init_task_group.siblings, &root_task_group.children);
9539 #endif /* CONFIG_USER_SCHED */
9540 #endif /* CONFIG_GROUP_SCHED */
9542 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9543 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9544 __alignof__(unsigned long));
9546 for_each_possible_cpu(i) {
9550 raw_spin_lock_init(&rq->lock);
9552 rq->calc_load_active = 0;
9553 rq->calc_load_update = jiffies + LOAD_FREQ;
9554 init_cfs_rq(&rq->cfs, rq);
9555 init_rt_rq(&rq->rt, rq);
9556 #ifdef CONFIG_FAIR_GROUP_SCHED
9557 init_task_group.shares = init_task_group_load;
9558 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9559 #ifdef CONFIG_CGROUP_SCHED
9561 * How much cpu bandwidth does init_task_group get?
9563 * In case of task-groups formed thr' the cgroup filesystem, it
9564 * gets 100% of the cpu resources in the system. This overall
9565 * system cpu resource is divided among the tasks of
9566 * init_task_group and its child task-groups in a fair manner,
9567 * based on each entity's (task or task-group's) weight
9568 * (se->load.weight).
9570 * In other words, if init_task_group has 10 tasks of weight
9571 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9572 * then A0's share of the cpu resource is:
9574 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9576 * We achieve this by letting init_task_group's tasks sit
9577 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9579 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9580 #elif defined CONFIG_USER_SCHED
9581 root_task_group.shares = NICE_0_LOAD;
9582 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9584 * In case of task-groups formed thr' the user id of tasks,
9585 * init_task_group represents tasks belonging to root user.
9586 * Hence it forms a sibling of all subsequent groups formed.
9587 * In this case, init_task_group gets only a fraction of overall
9588 * system cpu resource, based on the weight assigned to root
9589 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9590 * by letting tasks of init_task_group sit in a separate cfs_rq
9591 * (init_tg_cfs_rq) and having one entity represent this group of
9592 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9594 init_tg_cfs_entry(&init_task_group,
9595 &per_cpu(init_tg_cfs_rq, i),
9596 &per_cpu(init_sched_entity, i), i, 1,
9597 root_task_group.se[i]);
9600 #endif /* CONFIG_FAIR_GROUP_SCHED */
9602 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9603 #ifdef CONFIG_RT_GROUP_SCHED
9604 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9605 #ifdef CONFIG_CGROUP_SCHED
9606 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9607 #elif defined CONFIG_USER_SCHED
9608 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9609 init_tg_rt_entry(&init_task_group,
9610 &per_cpu(init_rt_rq_var, i),
9611 &per_cpu(init_sched_rt_entity, i), i, 1,
9612 root_task_group.rt_se[i]);
9616 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9617 rq->cpu_load[j] = 0;
9621 rq->post_schedule = 0;
9622 rq->active_balance = 0;
9623 rq->next_balance = jiffies;
9627 rq->migration_thread = NULL;
9629 rq->avg_idle = 2*sysctl_sched_migration_cost;
9630 INIT_LIST_HEAD(&rq->migration_queue);
9631 rq_attach_root(rq, &def_root_domain);
9634 atomic_set(&rq->nr_iowait, 0);
9637 set_load_weight(&init_task);
9639 #ifdef CONFIG_PREEMPT_NOTIFIERS
9640 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9644 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9647 #ifdef CONFIG_RT_MUTEXES
9648 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
9652 * The boot idle thread does lazy MMU switching as well:
9654 atomic_inc(&init_mm.mm_count);
9655 enter_lazy_tlb(&init_mm, current);
9658 * Make us the idle thread. Technically, schedule() should not be
9659 * called from this thread, however somewhere below it might be,
9660 * but because we are the idle thread, we just pick up running again
9661 * when this runqueue becomes "idle".
9663 init_idle(current, smp_processor_id());
9665 calc_load_update = jiffies + LOAD_FREQ;
9668 * During early bootup we pretend to be a normal task:
9670 current->sched_class = &fair_sched_class;
9672 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9673 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9676 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9677 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9679 /* May be allocated at isolcpus cmdline parse time */
9680 if (cpu_isolated_map == NULL)
9681 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9686 scheduler_running = 1;
9689 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9690 static inline int preempt_count_equals(int preempt_offset)
9692 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9694 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9697 void __might_sleep(char *file, int line, int preempt_offset)
9700 static unsigned long prev_jiffy; /* ratelimiting */
9702 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9703 system_state != SYSTEM_RUNNING || oops_in_progress)
9705 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9707 prev_jiffy = jiffies;
9709 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9711 pr_err("in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9712 in_atomic(), irqs_disabled(),
9713 current->pid, current->comm);
9715 debug_show_held_locks(current);
9716 if (irqs_disabled())
9717 print_irqtrace_events(current);
9721 EXPORT_SYMBOL(__might_sleep);
9724 #ifdef CONFIG_MAGIC_SYSRQ
9725 static void normalize_task(struct rq *rq, struct task_struct *p)
9729 update_rq_clock(rq);
9730 on_rq = p->se.on_rq;
9732 deactivate_task(rq, p, 0);
9733 __setscheduler(rq, p, SCHED_NORMAL, 0);
9735 activate_task(rq, p, 0);
9736 resched_task(rq->curr);
9740 void normalize_rt_tasks(void)
9742 struct task_struct *g, *p;
9743 unsigned long flags;
9746 read_lock_irqsave(&tasklist_lock, flags);
9747 do_each_thread(g, p) {
9749 * Only normalize user tasks:
9754 p->se.exec_start = 0;
9755 #ifdef CONFIG_SCHEDSTATS
9756 p->se.wait_start = 0;
9757 p->se.sleep_start = 0;
9758 p->se.block_start = 0;
9763 * Renice negative nice level userspace
9766 if (TASK_NICE(p) < 0 && p->mm)
9767 set_user_nice(p, 0);
9771 raw_spin_lock(&p->pi_lock);
9772 rq = __task_rq_lock(p);
9774 normalize_task(rq, p);
9776 __task_rq_unlock(rq);
9777 raw_spin_unlock(&p->pi_lock);
9778 } while_each_thread(g, p);
9780 read_unlock_irqrestore(&tasklist_lock, flags);
9783 #endif /* CONFIG_MAGIC_SYSRQ */
9787 * These functions are only useful for the IA64 MCA handling.
9789 * They can only be called when the whole system has been
9790 * stopped - every CPU needs to be quiescent, and no scheduling
9791 * activity can take place. Using them for anything else would
9792 * be a serious bug, and as a result, they aren't even visible
9793 * under any other configuration.
9797 * curr_task - return the current task for a given cpu.
9798 * @cpu: the processor in question.
9800 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9802 struct task_struct *curr_task(int cpu)
9804 return cpu_curr(cpu);
9808 * set_curr_task - set the current task for a given cpu.
9809 * @cpu: the processor in question.
9810 * @p: the task pointer to set.
9812 * Description: This function must only be used when non-maskable interrupts
9813 * are serviced on a separate stack. It allows the architecture to switch the
9814 * notion of the current task on a cpu in a non-blocking manner. This function
9815 * must be called with all CPU's synchronized, and interrupts disabled, the
9816 * and caller must save the original value of the current task (see
9817 * curr_task() above) and restore that value before reenabling interrupts and
9818 * re-starting the system.
9820 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9822 void set_curr_task(int cpu, struct task_struct *p)
9829 #ifdef CONFIG_FAIR_GROUP_SCHED
9830 static void free_fair_sched_group(struct task_group *tg)
9834 for_each_possible_cpu(i) {
9836 kfree(tg->cfs_rq[i]);
9846 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9848 struct cfs_rq *cfs_rq;
9849 struct sched_entity *se;
9853 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9856 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9860 tg->shares = NICE_0_LOAD;
9862 for_each_possible_cpu(i) {
9865 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9866 GFP_KERNEL, cpu_to_node(i));
9870 se = kzalloc_node(sizeof(struct sched_entity),
9871 GFP_KERNEL, cpu_to_node(i));
9875 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9886 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9888 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9889 &cpu_rq(cpu)->leaf_cfs_rq_list);
9892 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9894 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9896 #else /* !CONFG_FAIR_GROUP_SCHED */
9897 static inline void free_fair_sched_group(struct task_group *tg)
9902 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9907 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9911 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9914 #endif /* CONFIG_FAIR_GROUP_SCHED */
9916 #ifdef CONFIG_RT_GROUP_SCHED
9917 static void free_rt_sched_group(struct task_group *tg)
9921 destroy_rt_bandwidth(&tg->rt_bandwidth);
9923 for_each_possible_cpu(i) {
9925 kfree(tg->rt_rq[i]);
9927 kfree(tg->rt_se[i]);
9935 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9937 struct rt_rq *rt_rq;
9938 struct sched_rt_entity *rt_se;
9942 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9945 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9949 init_rt_bandwidth(&tg->rt_bandwidth,
9950 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9952 for_each_possible_cpu(i) {
9955 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9956 GFP_KERNEL, cpu_to_node(i));
9960 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9961 GFP_KERNEL, cpu_to_node(i));
9965 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9976 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9978 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9979 &cpu_rq(cpu)->leaf_rt_rq_list);
9982 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9984 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9986 #else /* !CONFIG_RT_GROUP_SCHED */
9987 static inline void free_rt_sched_group(struct task_group *tg)
9992 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9997 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10001 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10004 #endif /* CONFIG_RT_GROUP_SCHED */
10006 #ifdef CONFIG_GROUP_SCHED
10007 static void free_sched_group(struct task_group *tg)
10009 free_fair_sched_group(tg);
10010 free_rt_sched_group(tg);
10014 /* allocate runqueue etc for a new task group */
10015 struct task_group *sched_create_group(struct task_group *parent)
10017 struct task_group *tg;
10018 unsigned long flags;
10021 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
10023 return ERR_PTR(-ENOMEM);
10025 if (!alloc_fair_sched_group(tg, parent))
10028 if (!alloc_rt_sched_group(tg, parent))
10031 spin_lock_irqsave(&task_group_lock, flags);
10032 for_each_possible_cpu(i) {
10033 register_fair_sched_group(tg, i);
10034 register_rt_sched_group(tg, i);
10036 list_add_rcu(&tg->list, &task_groups);
10038 WARN_ON(!parent); /* root should already exist */
10040 tg->parent = parent;
10041 INIT_LIST_HEAD(&tg->children);
10042 list_add_rcu(&tg->siblings, &parent->children);
10043 spin_unlock_irqrestore(&task_group_lock, flags);
10048 free_sched_group(tg);
10049 return ERR_PTR(-ENOMEM);
10052 /* rcu callback to free various structures associated with a task group */
10053 static void free_sched_group_rcu(struct rcu_head *rhp)
10055 /* now it should be safe to free those cfs_rqs */
10056 free_sched_group(container_of(rhp, struct task_group, rcu));
10059 /* Destroy runqueue etc associated with a task group */
10060 void sched_destroy_group(struct task_group *tg)
10062 unsigned long flags;
10065 spin_lock_irqsave(&task_group_lock, flags);
10066 for_each_possible_cpu(i) {
10067 unregister_fair_sched_group(tg, i);
10068 unregister_rt_sched_group(tg, i);
10070 list_del_rcu(&tg->list);
10071 list_del_rcu(&tg->siblings);
10072 spin_unlock_irqrestore(&task_group_lock, flags);
10074 /* wait for possible concurrent references to cfs_rqs complete */
10075 call_rcu(&tg->rcu, free_sched_group_rcu);
10078 /* change task's runqueue when it moves between groups.
10079 * The caller of this function should have put the task in its new group
10080 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10081 * reflect its new group.
10083 void sched_move_task(struct task_struct *tsk)
10085 int on_rq, running;
10086 unsigned long flags;
10089 rq = task_rq_lock(tsk, &flags);
10091 update_rq_clock(rq);
10093 running = task_current(rq, tsk);
10094 on_rq = tsk->se.on_rq;
10097 dequeue_task(rq, tsk, 0);
10098 if (unlikely(running))
10099 tsk->sched_class->put_prev_task(rq, tsk);
10101 set_task_rq(tsk, task_cpu(tsk));
10103 #ifdef CONFIG_FAIR_GROUP_SCHED
10104 if (tsk->sched_class->moved_group)
10105 tsk->sched_class->moved_group(tsk);
10108 if (unlikely(running))
10109 tsk->sched_class->set_curr_task(rq);
10111 enqueue_task(rq, tsk, 0);
10113 task_rq_unlock(rq, &flags);
10115 #endif /* CONFIG_GROUP_SCHED */
10117 #ifdef CONFIG_FAIR_GROUP_SCHED
10118 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10120 struct cfs_rq *cfs_rq = se->cfs_rq;
10125 dequeue_entity(cfs_rq, se, 0);
10127 se->load.weight = shares;
10128 se->load.inv_weight = 0;
10131 enqueue_entity(cfs_rq, se, 0);
10134 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10136 struct cfs_rq *cfs_rq = se->cfs_rq;
10137 struct rq *rq = cfs_rq->rq;
10138 unsigned long flags;
10140 raw_spin_lock_irqsave(&rq->lock, flags);
10141 __set_se_shares(se, shares);
10142 raw_spin_unlock_irqrestore(&rq->lock, flags);
10145 static DEFINE_MUTEX(shares_mutex);
10147 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10150 unsigned long flags;
10153 * We can't change the weight of the root cgroup.
10158 if (shares < MIN_SHARES)
10159 shares = MIN_SHARES;
10160 else if (shares > MAX_SHARES)
10161 shares = MAX_SHARES;
10163 mutex_lock(&shares_mutex);
10164 if (tg->shares == shares)
10167 spin_lock_irqsave(&task_group_lock, flags);
10168 for_each_possible_cpu(i)
10169 unregister_fair_sched_group(tg, i);
10170 list_del_rcu(&tg->siblings);
10171 spin_unlock_irqrestore(&task_group_lock, flags);
10173 /* wait for any ongoing reference to this group to finish */
10174 synchronize_sched();
10177 * Now we are free to modify the group's share on each cpu
10178 * w/o tripping rebalance_share or load_balance_fair.
10180 tg->shares = shares;
10181 for_each_possible_cpu(i) {
10183 * force a rebalance
10185 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10186 set_se_shares(tg->se[i], shares);
10190 * Enable load balance activity on this group, by inserting it back on
10191 * each cpu's rq->leaf_cfs_rq_list.
10193 spin_lock_irqsave(&task_group_lock, flags);
10194 for_each_possible_cpu(i)
10195 register_fair_sched_group(tg, i);
10196 list_add_rcu(&tg->siblings, &tg->parent->children);
10197 spin_unlock_irqrestore(&task_group_lock, flags);
10199 mutex_unlock(&shares_mutex);
10203 unsigned long sched_group_shares(struct task_group *tg)
10209 #ifdef CONFIG_RT_GROUP_SCHED
10211 * Ensure that the real time constraints are schedulable.
10213 static DEFINE_MUTEX(rt_constraints_mutex);
10215 static unsigned long to_ratio(u64 period, u64 runtime)
10217 if (runtime == RUNTIME_INF)
10220 return div64_u64(runtime << 20, period);
10223 /* Must be called with tasklist_lock held */
10224 static inline int tg_has_rt_tasks(struct task_group *tg)
10226 struct task_struct *g, *p;
10228 do_each_thread(g, p) {
10229 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10231 } while_each_thread(g, p);
10236 struct rt_schedulable_data {
10237 struct task_group *tg;
10242 static int tg_schedulable(struct task_group *tg, void *data)
10244 struct rt_schedulable_data *d = data;
10245 struct task_group *child;
10246 unsigned long total, sum = 0;
10247 u64 period, runtime;
10249 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10250 runtime = tg->rt_bandwidth.rt_runtime;
10253 period = d->rt_period;
10254 runtime = d->rt_runtime;
10257 #ifdef CONFIG_USER_SCHED
10258 if (tg == &root_task_group) {
10259 period = global_rt_period();
10260 runtime = global_rt_runtime();
10265 * Cannot have more runtime than the period.
10267 if (runtime > period && runtime != RUNTIME_INF)
10271 * Ensure we don't starve existing RT tasks.
10273 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10276 total = to_ratio(period, runtime);
10279 * Nobody can have more than the global setting allows.
10281 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10285 * The sum of our children's runtime should not exceed our own.
10287 list_for_each_entry_rcu(child, &tg->children, siblings) {
10288 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10289 runtime = child->rt_bandwidth.rt_runtime;
10291 if (child == d->tg) {
10292 period = d->rt_period;
10293 runtime = d->rt_runtime;
10296 sum += to_ratio(period, runtime);
10305 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10307 struct rt_schedulable_data data = {
10309 .rt_period = period,
10310 .rt_runtime = runtime,
10313 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10316 static int tg_set_bandwidth(struct task_group *tg,
10317 u64 rt_period, u64 rt_runtime)
10321 mutex_lock(&rt_constraints_mutex);
10322 read_lock(&tasklist_lock);
10323 err = __rt_schedulable(tg, rt_period, rt_runtime);
10327 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10328 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10329 tg->rt_bandwidth.rt_runtime = rt_runtime;
10331 for_each_possible_cpu(i) {
10332 struct rt_rq *rt_rq = tg->rt_rq[i];
10334 raw_spin_lock(&rt_rq->rt_runtime_lock);
10335 rt_rq->rt_runtime = rt_runtime;
10336 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10338 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10340 read_unlock(&tasklist_lock);
10341 mutex_unlock(&rt_constraints_mutex);
10346 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10348 u64 rt_runtime, rt_period;
10350 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10351 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10352 if (rt_runtime_us < 0)
10353 rt_runtime = RUNTIME_INF;
10355 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10358 long sched_group_rt_runtime(struct task_group *tg)
10362 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10365 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10366 do_div(rt_runtime_us, NSEC_PER_USEC);
10367 return rt_runtime_us;
10370 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10372 u64 rt_runtime, rt_period;
10374 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10375 rt_runtime = tg->rt_bandwidth.rt_runtime;
10377 if (rt_period == 0)
10380 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10383 long sched_group_rt_period(struct task_group *tg)
10387 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10388 do_div(rt_period_us, NSEC_PER_USEC);
10389 return rt_period_us;
10392 static int sched_rt_global_constraints(void)
10394 u64 runtime, period;
10397 if (sysctl_sched_rt_period <= 0)
10400 runtime = global_rt_runtime();
10401 period = global_rt_period();
10404 * Sanity check on the sysctl variables.
10406 if (runtime > period && runtime != RUNTIME_INF)
10409 mutex_lock(&rt_constraints_mutex);
10410 read_lock(&tasklist_lock);
10411 ret = __rt_schedulable(NULL, 0, 0);
10412 read_unlock(&tasklist_lock);
10413 mutex_unlock(&rt_constraints_mutex);
10418 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10420 /* Don't accept realtime tasks when there is no way for them to run */
10421 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10427 #else /* !CONFIG_RT_GROUP_SCHED */
10428 static int sched_rt_global_constraints(void)
10430 unsigned long flags;
10433 if (sysctl_sched_rt_period <= 0)
10437 * There's always some RT tasks in the root group
10438 * -- migration, kstopmachine etc..
10440 if (sysctl_sched_rt_runtime == 0)
10443 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10444 for_each_possible_cpu(i) {
10445 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10447 raw_spin_lock(&rt_rq->rt_runtime_lock);
10448 rt_rq->rt_runtime = global_rt_runtime();
10449 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10451 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10455 #endif /* CONFIG_RT_GROUP_SCHED */
10457 int sched_rt_handler(struct ctl_table *table, int write,
10458 void __user *buffer, size_t *lenp,
10462 int old_period, old_runtime;
10463 static DEFINE_MUTEX(mutex);
10465 mutex_lock(&mutex);
10466 old_period = sysctl_sched_rt_period;
10467 old_runtime = sysctl_sched_rt_runtime;
10469 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10471 if (!ret && write) {
10472 ret = sched_rt_global_constraints();
10474 sysctl_sched_rt_period = old_period;
10475 sysctl_sched_rt_runtime = old_runtime;
10477 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10478 def_rt_bandwidth.rt_period =
10479 ns_to_ktime(global_rt_period());
10482 mutex_unlock(&mutex);
10487 #ifdef CONFIG_CGROUP_SCHED
10489 /* return corresponding task_group object of a cgroup */
10490 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10492 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10493 struct task_group, css);
10496 static struct cgroup_subsys_state *
10497 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10499 struct task_group *tg, *parent;
10501 if (!cgrp->parent) {
10502 /* This is early initialization for the top cgroup */
10503 return &init_task_group.css;
10506 parent = cgroup_tg(cgrp->parent);
10507 tg = sched_create_group(parent);
10509 return ERR_PTR(-ENOMEM);
10515 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10517 struct task_group *tg = cgroup_tg(cgrp);
10519 sched_destroy_group(tg);
10523 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10525 #ifdef CONFIG_RT_GROUP_SCHED
10526 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10529 /* We don't support RT-tasks being in separate groups */
10530 if (tsk->sched_class != &fair_sched_class)
10537 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10538 struct task_struct *tsk, bool threadgroup)
10540 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10544 struct task_struct *c;
10546 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10547 retval = cpu_cgroup_can_attach_task(cgrp, c);
10559 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10560 struct cgroup *old_cont, struct task_struct *tsk,
10563 sched_move_task(tsk);
10565 struct task_struct *c;
10567 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10568 sched_move_task(c);
10574 #ifdef CONFIG_FAIR_GROUP_SCHED
10575 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10578 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10581 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10583 struct task_group *tg = cgroup_tg(cgrp);
10585 return (u64) tg->shares;
10587 #endif /* CONFIG_FAIR_GROUP_SCHED */
10589 #ifdef CONFIG_RT_GROUP_SCHED
10590 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10593 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10596 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10598 return sched_group_rt_runtime(cgroup_tg(cgrp));
10601 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10604 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10607 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10609 return sched_group_rt_period(cgroup_tg(cgrp));
10611 #endif /* CONFIG_RT_GROUP_SCHED */
10613 static struct cftype cpu_files[] = {
10614 #ifdef CONFIG_FAIR_GROUP_SCHED
10617 .read_u64 = cpu_shares_read_u64,
10618 .write_u64 = cpu_shares_write_u64,
10621 #ifdef CONFIG_RT_GROUP_SCHED
10623 .name = "rt_runtime_us",
10624 .read_s64 = cpu_rt_runtime_read,
10625 .write_s64 = cpu_rt_runtime_write,
10628 .name = "rt_period_us",
10629 .read_u64 = cpu_rt_period_read_uint,
10630 .write_u64 = cpu_rt_period_write_uint,
10635 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10637 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10640 struct cgroup_subsys cpu_cgroup_subsys = {
10642 .create = cpu_cgroup_create,
10643 .destroy = cpu_cgroup_destroy,
10644 .can_attach = cpu_cgroup_can_attach,
10645 .attach = cpu_cgroup_attach,
10646 .populate = cpu_cgroup_populate,
10647 .subsys_id = cpu_cgroup_subsys_id,
10651 #endif /* CONFIG_CGROUP_SCHED */
10653 #ifdef CONFIG_CGROUP_CPUACCT
10656 * CPU accounting code for task groups.
10658 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10659 * (balbir@in.ibm.com).
10662 /* track cpu usage of a group of tasks and its child groups */
10664 struct cgroup_subsys_state css;
10665 /* cpuusage holds pointer to a u64-type object on every cpu */
10667 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10668 struct cpuacct *parent;
10671 struct cgroup_subsys cpuacct_subsys;
10673 /* return cpu accounting group corresponding to this container */
10674 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10676 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10677 struct cpuacct, css);
10680 /* return cpu accounting group to which this task belongs */
10681 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10683 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10684 struct cpuacct, css);
10687 /* create a new cpu accounting group */
10688 static struct cgroup_subsys_state *cpuacct_create(
10689 struct cgroup_subsys *ss, struct cgroup *cgrp)
10691 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10697 ca->cpuusage = alloc_percpu(u64);
10701 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10702 if (percpu_counter_init(&ca->cpustat[i], 0))
10703 goto out_free_counters;
10706 ca->parent = cgroup_ca(cgrp->parent);
10712 percpu_counter_destroy(&ca->cpustat[i]);
10713 free_percpu(ca->cpuusage);
10717 return ERR_PTR(-ENOMEM);
10720 /* destroy an existing cpu accounting group */
10722 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10724 struct cpuacct *ca = cgroup_ca(cgrp);
10727 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10728 percpu_counter_destroy(&ca->cpustat[i]);
10729 free_percpu(ca->cpuusage);
10733 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10735 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10738 #ifndef CONFIG_64BIT
10740 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10742 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10744 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10752 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10754 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10756 #ifndef CONFIG_64BIT
10758 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10760 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10762 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10768 /* return total cpu usage (in nanoseconds) of a group */
10769 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10771 struct cpuacct *ca = cgroup_ca(cgrp);
10772 u64 totalcpuusage = 0;
10775 for_each_present_cpu(i)
10776 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10778 return totalcpuusage;
10781 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10784 struct cpuacct *ca = cgroup_ca(cgrp);
10793 for_each_present_cpu(i)
10794 cpuacct_cpuusage_write(ca, i, 0);
10800 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10801 struct seq_file *m)
10803 struct cpuacct *ca = cgroup_ca(cgroup);
10807 for_each_present_cpu(i) {
10808 percpu = cpuacct_cpuusage_read(ca, i);
10809 seq_printf(m, "%llu ", (unsigned long long) percpu);
10811 seq_printf(m, "\n");
10815 static const char *cpuacct_stat_desc[] = {
10816 [CPUACCT_STAT_USER] = "user",
10817 [CPUACCT_STAT_SYSTEM] = "system",
10820 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10821 struct cgroup_map_cb *cb)
10823 struct cpuacct *ca = cgroup_ca(cgrp);
10826 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10827 s64 val = percpu_counter_read(&ca->cpustat[i]);
10828 val = cputime64_to_clock_t(val);
10829 cb->fill(cb, cpuacct_stat_desc[i], val);
10834 static struct cftype files[] = {
10837 .read_u64 = cpuusage_read,
10838 .write_u64 = cpuusage_write,
10841 .name = "usage_percpu",
10842 .read_seq_string = cpuacct_percpu_seq_read,
10846 .read_map = cpuacct_stats_show,
10850 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10852 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10856 * charge this task's execution time to its accounting group.
10858 * called with rq->lock held.
10860 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10862 struct cpuacct *ca;
10865 if (unlikely(!cpuacct_subsys.active))
10868 cpu = task_cpu(tsk);
10874 for (; ca; ca = ca->parent) {
10875 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10876 *cpuusage += cputime;
10883 * Charge the system/user time to the task's accounting group.
10885 static void cpuacct_update_stats(struct task_struct *tsk,
10886 enum cpuacct_stat_index idx, cputime_t val)
10888 struct cpuacct *ca;
10890 if (unlikely(!cpuacct_subsys.active))
10897 percpu_counter_add(&ca->cpustat[idx], val);
10903 struct cgroup_subsys cpuacct_subsys = {
10905 .create = cpuacct_create,
10906 .destroy = cpuacct_destroy,
10907 .populate = cpuacct_populate,
10908 .subsys_id = cpuacct_subsys_id,
10910 #endif /* CONFIG_CGROUP_CPUACCT */
10914 int rcu_expedited_torture_stats(char *page)
10918 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10920 void synchronize_sched_expedited(void)
10923 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10925 #else /* #ifndef CONFIG_SMP */
10927 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10928 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10930 #define RCU_EXPEDITED_STATE_POST -2
10931 #define RCU_EXPEDITED_STATE_IDLE -1
10933 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10935 int rcu_expedited_torture_stats(char *page)
10940 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10941 for_each_online_cpu(cpu) {
10942 cnt += sprintf(&page[cnt], " %d:%d",
10943 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10945 cnt += sprintf(&page[cnt], "\n");
10948 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10950 static long synchronize_sched_expedited_count;
10953 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10954 * approach to force grace period to end quickly. This consumes
10955 * significant time on all CPUs, and is thus not recommended for
10956 * any sort of common-case code.
10958 * Note that it is illegal to call this function while holding any
10959 * lock that is acquired by a CPU-hotplug notifier. Failing to
10960 * observe this restriction will result in deadlock.
10962 void synchronize_sched_expedited(void)
10965 unsigned long flags;
10966 bool need_full_sync = 0;
10968 struct migration_req *req;
10972 smp_mb(); /* ensure prior mod happens before capturing snap. */
10973 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10975 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10977 if (trycount++ < 10)
10978 udelay(trycount * num_online_cpus());
10980 synchronize_sched();
10983 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10984 smp_mb(); /* ensure test happens before caller kfree */
10989 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10990 for_each_online_cpu(cpu) {
10992 req = &per_cpu(rcu_migration_req, cpu);
10993 init_completion(&req->done);
10995 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10996 raw_spin_lock_irqsave(&rq->lock, flags);
10997 list_add(&req->list, &rq->migration_queue);
10998 raw_spin_unlock_irqrestore(&rq->lock, flags);
10999 wake_up_process(rq->migration_thread);
11001 for_each_online_cpu(cpu) {
11002 rcu_expedited_state = cpu;
11003 req = &per_cpu(rcu_migration_req, cpu);
11005 wait_for_completion(&req->done);
11006 raw_spin_lock_irqsave(&rq->lock, flags);
11007 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
11008 need_full_sync = 1;
11009 req->dest_cpu = RCU_MIGRATION_IDLE;
11010 raw_spin_unlock_irqrestore(&rq->lock, flags);
11012 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
11013 synchronize_sched_expedited_count++;
11014 mutex_unlock(&rcu_sched_expedited_mutex);
11016 if (need_full_sync)
11017 synchronize_sched();
11019 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
11021 #endif /* #else #ifndef CONFIG_SMP */