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;
2037 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2039 #ifdef CONFIG_SCHED_DEBUG
2041 * We should never call set_task_cpu() on a blocked task,
2042 * ttwu() will sort out the placement.
2044 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING);
2047 trace_sched_migrate_task(p, new_cpu);
2049 if (task_cpu(p) == new_cpu)
2052 p->se.nr_migrations++;
2053 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2055 __set_task_cpu(p, new_cpu);
2058 struct migration_req {
2059 struct list_head list;
2061 struct task_struct *task;
2064 struct completion done;
2068 * The task's runqueue lock must be held.
2069 * Returns true if you have to wait for migration thread.
2072 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2074 struct rq *rq = task_rq(p);
2077 * If the task is not on a runqueue (and not running), then
2078 * the next wake-up will properly place the task.
2080 if (!p->se.on_rq && !task_running(rq, p))
2083 init_completion(&req->done);
2085 req->dest_cpu = dest_cpu;
2086 list_add(&req->list, &rq->migration_queue);
2092 * wait_task_context_switch - wait for a thread to complete at least one
2095 * @p must not be current.
2097 void wait_task_context_switch(struct task_struct *p)
2099 unsigned long nvcsw, nivcsw, flags;
2107 * The runqueue is assigned before the actual context
2108 * switch. We need to take the runqueue lock.
2110 * We could check initially without the lock but it is
2111 * very likely that we need to take the lock in every
2114 rq = task_rq_lock(p, &flags);
2115 running = task_running(rq, p);
2116 task_rq_unlock(rq, &flags);
2118 if (likely(!running))
2121 * The switch count is incremented before the actual
2122 * context switch. We thus wait for two switches to be
2123 * sure at least one completed.
2125 if ((p->nvcsw - nvcsw) > 1)
2127 if ((p->nivcsw - nivcsw) > 1)
2135 * wait_task_inactive - wait for a thread to unschedule.
2137 * If @match_state is nonzero, it's the @p->state value just checked and
2138 * not expected to change. If it changes, i.e. @p might have woken up,
2139 * then return zero. When we succeed in waiting for @p to be off its CPU,
2140 * we return a positive number (its total switch count). If a second call
2141 * a short while later returns the same number, the caller can be sure that
2142 * @p has remained unscheduled the whole time.
2144 * The caller must ensure that the task *will* unschedule sometime soon,
2145 * else this function might spin for a *long* time. This function can't
2146 * be called with interrupts off, or it may introduce deadlock with
2147 * smp_call_function() if an IPI is sent by the same process we are
2148 * waiting to become inactive.
2150 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2152 unsigned long flags;
2159 * We do the initial early heuristics without holding
2160 * any task-queue locks at all. We'll only try to get
2161 * the runqueue lock when things look like they will
2167 * If the task is actively running on another CPU
2168 * still, just relax and busy-wait without holding
2171 * NOTE! Since we don't hold any locks, it's not
2172 * even sure that "rq" stays as the right runqueue!
2173 * But we don't care, since "task_running()" will
2174 * return false if the runqueue has changed and p
2175 * is actually now running somewhere else!
2177 while (task_running(rq, p)) {
2178 if (match_state && unlikely(p->state != match_state))
2184 * Ok, time to look more closely! We need the rq
2185 * lock now, to be *sure*. If we're wrong, we'll
2186 * just go back and repeat.
2188 rq = task_rq_lock(p, &flags);
2189 trace_sched_wait_task(rq, p);
2190 running = task_running(rq, p);
2191 on_rq = p->se.on_rq;
2193 if (!match_state || p->state == match_state)
2194 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2195 task_rq_unlock(rq, &flags);
2198 * If it changed from the expected state, bail out now.
2200 if (unlikely(!ncsw))
2204 * Was it really running after all now that we
2205 * checked with the proper locks actually held?
2207 * Oops. Go back and try again..
2209 if (unlikely(running)) {
2215 * It's not enough that it's not actively running,
2216 * it must be off the runqueue _entirely_, and not
2219 * So if it was still runnable (but just not actively
2220 * running right now), it's preempted, and we should
2221 * yield - it could be a while.
2223 if (unlikely(on_rq)) {
2224 schedule_timeout_uninterruptible(1);
2229 * Ahh, all good. It wasn't running, and it wasn't
2230 * runnable, which means that it will never become
2231 * running in the future either. We're all done!
2240 * kick_process - kick a running thread to enter/exit the kernel
2241 * @p: the to-be-kicked thread
2243 * Cause a process which is running on another CPU to enter
2244 * kernel-mode, without any delay. (to get signals handled.)
2246 * NOTE: this function doesnt have to take the runqueue lock,
2247 * because all it wants to ensure is that the remote task enters
2248 * the kernel. If the IPI races and the task has been migrated
2249 * to another CPU then no harm is done and the purpose has been
2252 void kick_process(struct task_struct *p)
2258 if ((cpu != smp_processor_id()) && task_curr(p))
2259 smp_send_reschedule(cpu);
2262 EXPORT_SYMBOL_GPL(kick_process);
2263 #endif /* CONFIG_SMP */
2266 * task_oncpu_function_call - call a function on the cpu on which a task runs
2267 * @p: the task to evaluate
2268 * @func: the function to be called
2269 * @info: the function call argument
2271 * Calls the function @func when the task is currently running. This might
2272 * be on the current CPU, which just calls the function directly
2274 void task_oncpu_function_call(struct task_struct *p,
2275 void (*func) (void *info), void *info)
2282 smp_call_function_single(cpu, func, info, 1);
2287 static int select_fallback_rq(int cpu, struct task_struct *p)
2290 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2292 /* Look for allowed, online CPU in same node. */
2293 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2294 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2297 /* Any allowed, online CPU? */
2298 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2299 if (dest_cpu < nr_cpu_ids)
2302 /* No more Mr. Nice Guy. */
2303 if (dest_cpu >= nr_cpu_ids) {
2305 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2307 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2310 * Don't tell them about moving exiting tasks or
2311 * kernel threads (both mm NULL), since they never
2314 if (p->mm && printk_ratelimit()) {
2315 printk(KERN_INFO "process %d (%s) no "
2316 "longer affine to cpu%d\n",
2317 task_pid_nr(p), p->comm, cpu);
2327 * - fork, @p is stable because it isn't on the tasklist yet
2329 * - exec, @p is unstable, retry loop
2331 * - wake-up, we serialize ->cpus_allowed against TASK_WAKING so
2332 * we should be good.
2335 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2337 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2340 * In order not to call set_task_cpu() on a blocking task we need
2341 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2344 * Since this is common to all placement strategies, this lives here.
2346 * [ this allows ->select_task() to simply return task_cpu(p) and
2347 * not worry about this generic constraint ]
2349 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2351 cpu = select_fallback_rq(task_cpu(p), p);
2358 * try_to_wake_up - wake up a thread
2359 * @p: the to-be-woken-up thread
2360 * @state: the mask of task states that can be woken
2361 * @sync: do a synchronous wakeup?
2363 * Put it on the run-queue if it's not already there. The "current"
2364 * thread is always on the run-queue (except when the actual
2365 * re-schedule is in progress), and as such you're allowed to do
2366 * the simpler "current->state = TASK_RUNNING" to mark yourself
2367 * runnable without the overhead of this.
2369 * returns failure only if the task is already active.
2371 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2374 int cpu, orig_cpu, this_cpu, success = 0;
2375 unsigned long flags;
2376 struct rq *rq, *orig_rq;
2378 if (!sched_feat(SYNC_WAKEUPS))
2379 wake_flags &= ~WF_SYNC;
2381 this_cpu = get_cpu();
2384 rq = orig_rq = task_rq_lock(p, &flags);
2385 update_rq_clock(rq);
2386 if (!(p->state & state))
2396 if (unlikely(task_running(rq, p)))
2400 * In order to handle concurrent wakeups and release the rq->lock
2401 * we put the task in TASK_WAKING state.
2403 * First fix up the nr_uninterruptible count:
2405 if (task_contributes_to_load(p))
2406 rq->nr_uninterruptible--;
2407 p->state = TASK_WAKING;
2409 if (p->sched_class->task_waking)
2410 p->sched_class->task_waking(rq, p);
2412 __task_rq_unlock(rq);
2414 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2415 if (cpu != orig_cpu)
2416 set_task_cpu(p, cpu);
2418 rq = __task_rq_lock(p);
2419 update_rq_clock(rq);
2421 WARN_ON(p->state != TASK_WAKING);
2424 #ifdef CONFIG_SCHEDSTATS
2425 schedstat_inc(rq, ttwu_count);
2426 if (cpu == this_cpu)
2427 schedstat_inc(rq, ttwu_local);
2429 struct sched_domain *sd;
2430 for_each_domain(this_cpu, sd) {
2431 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2432 schedstat_inc(sd, ttwu_wake_remote);
2437 #endif /* CONFIG_SCHEDSTATS */
2440 #endif /* CONFIG_SMP */
2441 schedstat_inc(p, se.nr_wakeups);
2442 if (wake_flags & WF_SYNC)
2443 schedstat_inc(p, se.nr_wakeups_sync);
2444 if (orig_cpu != cpu)
2445 schedstat_inc(p, se.nr_wakeups_migrate);
2446 if (cpu == this_cpu)
2447 schedstat_inc(p, se.nr_wakeups_local);
2449 schedstat_inc(p, se.nr_wakeups_remote);
2450 activate_task(rq, p, 1);
2454 * Only attribute actual wakeups done by this task.
2456 if (!in_interrupt()) {
2457 struct sched_entity *se = ¤t->se;
2458 u64 sample = se->sum_exec_runtime;
2460 if (se->last_wakeup)
2461 sample -= se->last_wakeup;
2463 sample -= se->start_runtime;
2464 update_avg(&se->avg_wakeup, sample);
2466 se->last_wakeup = se->sum_exec_runtime;
2470 trace_sched_wakeup(rq, p, success);
2471 check_preempt_curr(rq, p, wake_flags);
2473 p->state = TASK_RUNNING;
2475 if (p->sched_class->task_woken)
2476 p->sched_class->task_woken(rq, p);
2478 if (unlikely(rq->idle_stamp)) {
2479 u64 delta = rq->clock - rq->idle_stamp;
2480 u64 max = 2*sysctl_sched_migration_cost;
2485 update_avg(&rq->avg_idle, delta);
2490 task_rq_unlock(rq, &flags);
2497 * wake_up_process - Wake up a specific process
2498 * @p: The process to be woken up.
2500 * Attempt to wake up the nominated process and move it to the set of runnable
2501 * processes. Returns 1 if the process was woken up, 0 if it was already
2504 * It may be assumed that this function implies a write memory barrier before
2505 * changing the task state if and only if any tasks are woken up.
2507 int wake_up_process(struct task_struct *p)
2509 return try_to_wake_up(p, TASK_ALL, 0);
2511 EXPORT_SYMBOL(wake_up_process);
2513 int wake_up_state(struct task_struct *p, unsigned int state)
2515 return try_to_wake_up(p, state, 0);
2519 * Perform scheduler related setup for a newly forked process p.
2520 * p is forked by current.
2522 * __sched_fork() is basic setup used by init_idle() too:
2524 static void __sched_fork(struct task_struct *p)
2526 p->se.exec_start = 0;
2527 p->se.sum_exec_runtime = 0;
2528 p->se.prev_sum_exec_runtime = 0;
2529 p->se.nr_migrations = 0;
2530 p->se.last_wakeup = 0;
2531 p->se.avg_overlap = 0;
2532 p->se.start_runtime = 0;
2533 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2535 #ifdef CONFIG_SCHEDSTATS
2536 p->se.wait_start = 0;
2538 p->se.wait_count = 0;
2541 p->se.sleep_start = 0;
2542 p->se.sleep_max = 0;
2543 p->se.sum_sleep_runtime = 0;
2545 p->se.block_start = 0;
2546 p->se.block_max = 0;
2548 p->se.slice_max = 0;
2550 p->se.nr_migrations_cold = 0;
2551 p->se.nr_failed_migrations_affine = 0;
2552 p->se.nr_failed_migrations_running = 0;
2553 p->se.nr_failed_migrations_hot = 0;
2554 p->se.nr_forced_migrations = 0;
2556 p->se.nr_wakeups = 0;
2557 p->se.nr_wakeups_sync = 0;
2558 p->se.nr_wakeups_migrate = 0;
2559 p->se.nr_wakeups_local = 0;
2560 p->se.nr_wakeups_remote = 0;
2561 p->se.nr_wakeups_affine = 0;
2562 p->se.nr_wakeups_affine_attempts = 0;
2563 p->se.nr_wakeups_passive = 0;
2564 p->se.nr_wakeups_idle = 0;
2568 INIT_LIST_HEAD(&p->rt.run_list);
2570 INIT_LIST_HEAD(&p->se.group_node);
2572 #ifdef CONFIG_PREEMPT_NOTIFIERS
2573 INIT_HLIST_HEAD(&p->preempt_notifiers);
2578 * fork()/clone()-time setup:
2580 void sched_fork(struct task_struct *p, int clone_flags)
2582 int cpu = get_cpu();
2586 * We mark the process as waking here. This guarantees that
2587 * nobody will actually run it, and a signal or other external
2588 * event cannot wake it up and insert it on the runqueue either.
2590 p->state = TASK_WAKING;
2593 * Revert to default priority/policy on fork if requested.
2595 if (unlikely(p->sched_reset_on_fork)) {
2596 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2597 p->policy = SCHED_NORMAL;
2598 p->normal_prio = p->static_prio;
2601 if (PRIO_TO_NICE(p->static_prio) < 0) {
2602 p->static_prio = NICE_TO_PRIO(0);
2603 p->normal_prio = p->static_prio;
2608 * We don't need the reset flag anymore after the fork. It has
2609 * fulfilled its duty:
2611 p->sched_reset_on_fork = 0;
2615 * Make sure we do not leak PI boosting priority to the child.
2617 p->prio = current->normal_prio;
2619 if (!rt_prio(p->prio))
2620 p->sched_class = &fair_sched_class;
2622 if (p->sched_class->task_fork)
2623 p->sched_class->task_fork(p);
2626 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2628 set_task_cpu(p, cpu);
2630 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2631 if (likely(sched_info_on()))
2632 memset(&p->sched_info, 0, sizeof(p->sched_info));
2634 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2637 #ifdef CONFIG_PREEMPT
2638 /* Want to start with kernel preemption disabled. */
2639 task_thread_info(p)->preempt_count = 1;
2641 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2647 * wake_up_new_task - wake up a newly created task for the first time.
2649 * This function will do some initial scheduler statistics housekeeping
2650 * that must be done for every newly created context, then puts the task
2651 * on the runqueue and wakes it.
2653 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2655 unsigned long flags;
2658 rq = task_rq_lock(p, &flags);
2659 BUG_ON(p->state != TASK_WAKING);
2660 p->state = TASK_RUNNING;
2661 update_rq_clock(rq);
2662 activate_task(rq, p, 0);
2663 trace_sched_wakeup_new(rq, p, 1);
2664 check_preempt_curr(rq, p, WF_FORK);
2666 if (p->sched_class->task_woken)
2667 p->sched_class->task_woken(rq, p);
2669 task_rq_unlock(rq, &flags);
2672 #ifdef CONFIG_PREEMPT_NOTIFIERS
2675 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2676 * @notifier: notifier struct to register
2678 void preempt_notifier_register(struct preempt_notifier *notifier)
2680 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2682 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2685 * preempt_notifier_unregister - no longer interested in preemption notifications
2686 * @notifier: notifier struct to unregister
2688 * This is safe to call from within a preemption notifier.
2690 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2692 hlist_del(¬ifier->link);
2694 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2696 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2698 struct preempt_notifier *notifier;
2699 struct hlist_node *node;
2701 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2702 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2706 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2707 struct task_struct *next)
2709 struct preempt_notifier *notifier;
2710 struct hlist_node *node;
2712 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2713 notifier->ops->sched_out(notifier, next);
2716 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2718 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2723 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2724 struct task_struct *next)
2728 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2731 * prepare_task_switch - prepare to switch tasks
2732 * @rq: the runqueue preparing to switch
2733 * @prev: the current task that is being switched out
2734 * @next: the task we are going to switch to.
2736 * This is called with the rq lock held and interrupts off. It must
2737 * be paired with a subsequent finish_task_switch after the context
2740 * prepare_task_switch sets up locking and calls architecture specific
2744 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2745 struct task_struct *next)
2747 fire_sched_out_preempt_notifiers(prev, next);
2748 prepare_lock_switch(rq, next);
2749 prepare_arch_switch(next);
2753 * finish_task_switch - clean up after a task-switch
2754 * @rq: runqueue associated with task-switch
2755 * @prev: the thread we just switched away from.
2757 * finish_task_switch must be called after the context switch, paired
2758 * with a prepare_task_switch call before the context switch.
2759 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2760 * and do any other architecture-specific cleanup actions.
2762 * Note that we may have delayed dropping an mm in context_switch(). If
2763 * so, we finish that here outside of the runqueue lock. (Doing it
2764 * with the lock held can cause deadlocks; see schedule() for
2767 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2768 __releases(rq->lock)
2770 struct mm_struct *mm = rq->prev_mm;
2776 * A task struct has one reference for the use as "current".
2777 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2778 * schedule one last time. The schedule call will never return, and
2779 * the scheduled task must drop that reference.
2780 * The test for TASK_DEAD must occur while the runqueue locks are
2781 * still held, otherwise prev could be scheduled on another cpu, die
2782 * there before we look at prev->state, and then the reference would
2784 * Manfred Spraul <manfred@colorfullife.com>
2786 prev_state = prev->state;
2787 finish_arch_switch(prev);
2788 perf_event_task_sched_in(current, cpu_of(rq));
2789 finish_lock_switch(rq, prev);
2791 fire_sched_in_preempt_notifiers(current);
2794 if (unlikely(prev_state == TASK_DEAD)) {
2796 * Remove function-return probe instances associated with this
2797 * task and put them back on the free list.
2799 kprobe_flush_task(prev);
2800 put_task_struct(prev);
2806 /* assumes rq->lock is held */
2807 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2809 if (prev->sched_class->pre_schedule)
2810 prev->sched_class->pre_schedule(rq, prev);
2813 /* rq->lock is NOT held, but preemption is disabled */
2814 static inline void post_schedule(struct rq *rq)
2816 if (rq->post_schedule) {
2817 unsigned long flags;
2819 raw_spin_lock_irqsave(&rq->lock, flags);
2820 if (rq->curr->sched_class->post_schedule)
2821 rq->curr->sched_class->post_schedule(rq);
2822 raw_spin_unlock_irqrestore(&rq->lock, flags);
2824 rq->post_schedule = 0;
2830 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2834 static inline void post_schedule(struct rq *rq)
2841 * schedule_tail - first thing a freshly forked thread must call.
2842 * @prev: the thread we just switched away from.
2844 asmlinkage void schedule_tail(struct task_struct *prev)
2845 __releases(rq->lock)
2847 struct rq *rq = this_rq();
2849 finish_task_switch(rq, prev);
2852 * FIXME: do we need to worry about rq being invalidated by the
2857 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2858 /* In this case, finish_task_switch does not reenable preemption */
2861 if (current->set_child_tid)
2862 put_user(task_pid_vnr(current), current->set_child_tid);
2866 * context_switch - switch to the new MM and the new
2867 * thread's register state.
2870 context_switch(struct rq *rq, struct task_struct *prev,
2871 struct task_struct *next)
2873 struct mm_struct *mm, *oldmm;
2875 prepare_task_switch(rq, prev, next);
2876 trace_sched_switch(rq, prev, next);
2878 oldmm = prev->active_mm;
2880 * For paravirt, this is coupled with an exit in switch_to to
2881 * combine the page table reload and the switch backend into
2884 arch_start_context_switch(prev);
2887 next->active_mm = oldmm;
2888 atomic_inc(&oldmm->mm_count);
2889 enter_lazy_tlb(oldmm, next);
2891 switch_mm(oldmm, mm, next);
2893 if (likely(!prev->mm)) {
2894 prev->active_mm = NULL;
2895 rq->prev_mm = oldmm;
2898 * Since the runqueue lock will be released by the next
2899 * task (which is an invalid locking op but in the case
2900 * of the scheduler it's an obvious special-case), so we
2901 * do an early lockdep release here:
2903 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2904 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2907 /* Here we just switch the register state and the stack. */
2908 switch_to(prev, next, prev);
2912 * this_rq must be evaluated again because prev may have moved
2913 * CPUs since it called schedule(), thus the 'rq' on its stack
2914 * frame will be invalid.
2916 finish_task_switch(this_rq(), prev);
2920 * nr_running, nr_uninterruptible and nr_context_switches:
2922 * externally visible scheduler statistics: current number of runnable
2923 * threads, current number of uninterruptible-sleeping threads, total
2924 * number of context switches performed since bootup.
2926 unsigned long nr_running(void)
2928 unsigned long i, sum = 0;
2930 for_each_online_cpu(i)
2931 sum += cpu_rq(i)->nr_running;
2936 unsigned long nr_uninterruptible(void)
2938 unsigned long i, sum = 0;
2940 for_each_possible_cpu(i)
2941 sum += cpu_rq(i)->nr_uninterruptible;
2944 * Since we read the counters lockless, it might be slightly
2945 * inaccurate. Do not allow it to go below zero though:
2947 if (unlikely((long)sum < 0))
2953 unsigned long long nr_context_switches(void)
2956 unsigned long long sum = 0;
2958 for_each_possible_cpu(i)
2959 sum += cpu_rq(i)->nr_switches;
2964 unsigned long nr_iowait(void)
2966 unsigned long i, sum = 0;
2968 for_each_possible_cpu(i)
2969 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2974 unsigned long nr_iowait_cpu(void)
2976 struct rq *this = this_rq();
2977 return atomic_read(&this->nr_iowait);
2980 unsigned long this_cpu_load(void)
2982 struct rq *this = this_rq();
2983 return this->cpu_load[0];
2987 /* Variables and functions for calc_load */
2988 static atomic_long_t calc_load_tasks;
2989 static unsigned long calc_load_update;
2990 unsigned long avenrun[3];
2991 EXPORT_SYMBOL(avenrun);
2994 * get_avenrun - get the load average array
2995 * @loads: pointer to dest load array
2996 * @offset: offset to add
2997 * @shift: shift count to shift the result left
2999 * These values are estimates at best, so no need for locking.
3001 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3003 loads[0] = (avenrun[0] + offset) << shift;
3004 loads[1] = (avenrun[1] + offset) << shift;
3005 loads[2] = (avenrun[2] + offset) << shift;
3008 static unsigned long
3009 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3012 load += active * (FIXED_1 - exp);
3013 return load >> FSHIFT;
3017 * calc_load - update the avenrun load estimates 10 ticks after the
3018 * CPUs have updated calc_load_tasks.
3020 void calc_global_load(void)
3022 unsigned long upd = calc_load_update + 10;
3025 if (time_before(jiffies, upd))
3028 active = atomic_long_read(&calc_load_tasks);
3029 active = active > 0 ? active * FIXED_1 : 0;
3031 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3032 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3033 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3035 calc_load_update += LOAD_FREQ;
3039 * Either called from update_cpu_load() or from a cpu going idle
3041 static void calc_load_account_active(struct rq *this_rq)
3043 long nr_active, delta;
3045 nr_active = this_rq->nr_running;
3046 nr_active += (long) this_rq->nr_uninterruptible;
3048 if (nr_active != this_rq->calc_load_active) {
3049 delta = nr_active - this_rq->calc_load_active;
3050 this_rq->calc_load_active = nr_active;
3051 atomic_long_add(delta, &calc_load_tasks);
3056 * Update rq->cpu_load[] statistics. This function is usually called every
3057 * scheduler tick (TICK_NSEC).
3059 static void update_cpu_load(struct rq *this_rq)
3061 unsigned long this_load = this_rq->load.weight;
3064 this_rq->nr_load_updates++;
3066 /* Update our load: */
3067 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3068 unsigned long old_load, new_load;
3070 /* scale is effectively 1 << i now, and >> i divides by scale */
3072 old_load = this_rq->cpu_load[i];
3073 new_load = this_load;
3075 * Round up the averaging division if load is increasing. This
3076 * prevents us from getting stuck on 9 if the load is 10, for
3079 if (new_load > old_load)
3080 new_load += scale-1;
3081 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3084 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3085 this_rq->calc_load_update += LOAD_FREQ;
3086 calc_load_account_active(this_rq);
3093 * double_rq_lock - safely lock two runqueues
3095 * Note this does not disable interrupts like task_rq_lock,
3096 * you need to do so manually before calling.
3098 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3099 __acquires(rq1->lock)
3100 __acquires(rq2->lock)
3102 BUG_ON(!irqs_disabled());
3104 raw_spin_lock(&rq1->lock);
3105 __acquire(rq2->lock); /* Fake it out ;) */
3108 raw_spin_lock(&rq1->lock);
3109 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3111 raw_spin_lock(&rq2->lock);
3112 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3115 update_rq_clock(rq1);
3116 update_rq_clock(rq2);
3120 * double_rq_unlock - safely unlock two runqueues
3122 * Note this does not restore interrupts like task_rq_unlock,
3123 * you need to do so manually after calling.
3125 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3126 __releases(rq1->lock)
3127 __releases(rq2->lock)
3129 raw_spin_unlock(&rq1->lock);
3131 raw_spin_unlock(&rq2->lock);
3133 __release(rq2->lock);
3137 * sched_exec - execve() is a valuable balancing opportunity, because at
3138 * this point the task has the smallest effective memory and cache footprint.
3140 void sched_exec(void)
3142 struct task_struct *p = current;
3143 struct migration_req req;
3144 int dest_cpu, this_cpu;
3145 unsigned long flags;
3149 this_cpu = get_cpu();
3150 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3151 if (dest_cpu == this_cpu) {
3156 rq = task_rq_lock(p, &flags);
3160 * select_task_rq() can race against ->cpus_allowed
3162 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3163 || unlikely(!cpu_active(dest_cpu))) {
3164 task_rq_unlock(rq, &flags);
3168 /* force the process onto the specified CPU */
3169 if (migrate_task(p, dest_cpu, &req)) {
3170 /* Need to wait for migration thread (might exit: take ref). */
3171 struct task_struct *mt = rq->migration_thread;
3173 get_task_struct(mt);
3174 task_rq_unlock(rq, &flags);
3175 wake_up_process(mt);
3176 put_task_struct(mt);
3177 wait_for_completion(&req.done);
3181 task_rq_unlock(rq, &flags);
3185 * pull_task - move a task from a remote runqueue to the local runqueue.
3186 * Both runqueues must be locked.
3188 static void pull_task(struct rq *src_rq, struct task_struct *p,
3189 struct rq *this_rq, int this_cpu)
3191 deactivate_task(src_rq, p, 0);
3192 set_task_cpu(p, this_cpu);
3193 activate_task(this_rq, p, 0);
3194 check_preempt_curr(this_rq, p, 0);
3198 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3201 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3202 struct sched_domain *sd, enum cpu_idle_type idle,
3205 int tsk_cache_hot = 0;
3207 * We do not migrate tasks that are:
3208 * 1) running (obviously), or
3209 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3210 * 3) are cache-hot on their current CPU.
3212 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3213 schedstat_inc(p, se.nr_failed_migrations_affine);
3218 if (task_running(rq, p)) {
3219 schedstat_inc(p, se.nr_failed_migrations_running);
3224 * Aggressive migration if:
3225 * 1) task is cache cold, or
3226 * 2) too many balance attempts have failed.
3229 tsk_cache_hot = task_hot(p, rq->clock, sd);
3230 if (!tsk_cache_hot ||
3231 sd->nr_balance_failed > sd->cache_nice_tries) {
3232 #ifdef CONFIG_SCHEDSTATS
3233 if (tsk_cache_hot) {
3234 schedstat_inc(sd, lb_hot_gained[idle]);
3235 schedstat_inc(p, se.nr_forced_migrations);
3241 if (tsk_cache_hot) {
3242 schedstat_inc(p, se.nr_failed_migrations_hot);
3248 static unsigned long
3249 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3250 unsigned long max_load_move, struct sched_domain *sd,
3251 enum cpu_idle_type idle, int *all_pinned,
3252 int *this_best_prio, struct rq_iterator *iterator)
3254 int loops = 0, pulled = 0, pinned = 0;
3255 struct task_struct *p;
3256 long rem_load_move = max_load_move;
3258 if (max_load_move == 0)
3264 * Start the load-balancing iterator:
3266 p = iterator->start(iterator->arg);
3268 if (!p || loops++ > sysctl_sched_nr_migrate)
3271 if ((p->se.load.weight >> 1) > rem_load_move ||
3272 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3273 p = iterator->next(iterator->arg);
3277 pull_task(busiest, p, this_rq, this_cpu);
3279 rem_load_move -= p->se.load.weight;
3281 #ifdef CONFIG_PREEMPT
3283 * NEWIDLE balancing is a source of latency, so preemptible kernels
3284 * will stop after the first task is pulled to minimize the critical
3287 if (idle == CPU_NEWLY_IDLE)
3292 * We only want to steal up to the prescribed amount of weighted load.
3294 if (rem_load_move > 0) {
3295 if (p->prio < *this_best_prio)
3296 *this_best_prio = p->prio;
3297 p = iterator->next(iterator->arg);
3302 * Right now, this is one of only two places pull_task() is called,
3303 * so we can safely collect pull_task() stats here rather than
3304 * inside pull_task().
3306 schedstat_add(sd, lb_gained[idle], pulled);
3309 *all_pinned = pinned;
3311 return max_load_move - rem_load_move;
3315 * move_tasks tries to move up to max_load_move weighted load from busiest to
3316 * this_rq, as part of a balancing operation within domain "sd".
3317 * Returns 1 if successful and 0 otherwise.
3319 * Called with both runqueues locked.
3321 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3322 unsigned long max_load_move,
3323 struct sched_domain *sd, enum cpu_idle_type idle,
3326 const struct sched_class *class = sched_class_highest;
3327 unsigned long total_load_moved = 0;
3328 int this_best_prio = this_rq->curr->prio;
3332 class->load_balance(this_rq, this_cpu, busiest,
3333 max_load_move - total_load_moved,
3334 sd, idle, all_pinned, &this_best_prio);
3335 class = class->next;
3337 #ifdef CONFIG_PREEMPT
3339 * NEWIDLE balancing is a source of latency, so preemptible
3340 * kernels will stop after the first task is pulled to minimize
3341 * the critical section.
3343 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3346 } while (class && max_load_move > total_load_moved);
3348 return total_load_moved > 0;
3352 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3353 struct sched_domain *sd, enum cpu_idle_type idle,
3354 struct rq_iterator *iterator)
3356 struct task_struct *p = iterator->start(iterator->arg);
3360 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3361 pull_task(busiest, p, this_rq, this_cpu);
3363 * Right now, this is only the second place pull_task()
3364 * is called, so we can safely collect pull_task()
3365 * stats here rather than inside pull_task().
3367 schedstat_inc(sd, lb_gained[idle]);
3371 p = iterator->next(iterator->arg);
3378 * move_one_task tries to move exactly one task from busiest to this_rq, as
3379 * part of active balancing operations within "domain".
3380 * Returns 1 if successful and 0 otherwise.
3382 * Called with both runqueues locked.
3384 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3385 struct sched_domain *sd, enum cpu_idle_type idle)
3387 const struct sched_class *class;
3389 for_each_class(class) {
3390 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3396 /********** Helpers for find_busiest_group ************************/
3398 * sd_lb_stats - Structure to store the statistics of a sched_domain
3399 * during load balancing.
3401 struct sd_lb_stats {
3402 struct sched_group *busiest; /* Busiest group in this sd */
3403 struct sched_group *this; /* Local group in this sd */
3404 unsigned long total_load; /* Total load of all groups in sd */
3405 unsigned long total_pwr; /* Total power of all groups in sd */
3406 unsigned long avg_load; /* Average load across all groups in sd */
3408 /** Statistics of this group */
3409 unsigned long this_load;
3410 unsigned long this_load_per_task;
3411 unsigned long this_nr_running;
3413 /* Statistics of the busiest group */
3414 unsigned long max_load;
3415 unsigned long busiest_load_per_task;
3416 unsigned long busiest_nr_running;
3418 int group_imb; /* Is there imbalance in this sd */
3419 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3420 int power_savings_balance; /* Is powersave balance needed for this sd */
3421 struct sched_group *group_min; /* Least loaded group in sd */
3422 struct sched_group *group_leader; /* Group which relieves group_min */
3423 unsigned long min_load_per_task; /* load_per_task in group_min */
3424 unsigned long leader_nr_running; /* Nr running of group_leader */
3425 unsigned long min_nr_running; /* Nr running of group_min */
3430 * sg_lb_stats - stats of a sched_group required for load_balancing
3432 struct sg_lb_stats {
3433 unsigned long avg_load; /*Avg load across the CPUs of the group */
3434 unsigned long group_load; /* Total load over the CPUs of the group */
3435 unsigned long sum_nr_running; /* Nr tasks running in the group */
3436 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3437 unsigned long group_capacity;
3438 int group_imb; /* Is there an imbalance in the group ? */
3442 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3443 * @group: The group whose first cpu is to be returned.
3445 static inline unsigned int group_first_cpu(struct sched_group *group)
3447 return cpumask_first(sched_group_cpus(group));
3451 * get_sd_load_idx - Obtain the load index for a given sched domain.
3452 * @sd: The sched_domain whose load_idx is to be obtained.
3453 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3455 static inline int get_sd_load_idx(struct sched_domain *sd,
3456 enum cpu_idle_type idle)
3462 load_idx = sd->busy_idx;
3465 case CPU_NEWLY_IDLE:
3466 load_idx = sd->newidle_idx;
3469 load_idx = sd->idle_idx;
3477 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3479 * init_sd_power_savings_stats - Initialize power savings statistics for
3480 * the given sched_domain, during load balancing.
3482 * @sd: Sched domain whose power-savings statistics are to be initialized.
3483 * @sds: Variable containing the statistics for sd.
3484 * @idle: Idle status of the CPU at which we're performing load-balancing.
3486 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3487 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3490 * Busy processors will not participate in power savings
3493 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3494 sds->power_savings_balance = 0;
3496 sds->power_savings_balance = 1;
3497 sds->min_nr_running = ULONG_MAX;
3498 sds->leader_nr_running = 0;
3503 * update_sd_power_savings_stats - Update the power saving stats for a
3504 * sched_domain while performing load balancing.
3506 * @group: sched_group belonging to the sched_domain under consideration.
3507 * @sds: Variable containing the statistics of the sched_domain
3508 * @local_group: Does group contain the CPU for which we're performing
3510 * @sgs: Variable containing the statistics of the group.
3512 static inline void update_sd_power_savings_stats(struct sched_group *group,
3513 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3516 if (!sds->power_savings_balance)
3520 * If the local group is idle or completely loaded
3521 * no need to do power savings balance at this domain
3523 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3524 !sds->this_nr_running))
3525 sds->power_savings_balance = 0;
3528 * If a group is already running at full capacity or idle,
3529 * don't include that group in power savings calculations
3531 if (!sds->power_savings_balance ||
3532 sgs->sum_nr_running >= sgs->group_capacity ||
3533 !sgs->sum_nr_running)
3537 * Calculate the group which has the least non-idle load.
3538 * This is the group from where we need to pick up the load
3541 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3542 (sgs->sum_nr_running == sds->min_nr_running &&
3543 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3544 sds->group_min = group;
3545 sds->min_nr_running = sgs->sum_nr_running;
3546 sds->min_load_per_task = sgs->sum_weighted_load /
3547 sgs->sum_nr_running;
3551 * Calculate the group which is almost near its
3552 * capacity but still has some space to pick up some load
3553 * from other group and save more power
3555 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3558 if (sgs->sum_nr_running > sds->leader_nr_running ||
3559 (sgs->sum_nr_running == sds->leader_nr_running &&
3560 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3561 sds->group_leader = group;
3562 sds->leader_nr_running = sgs->sum_nr_running;
3567 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3568 * @sds: Variable containing the statistics of the sched_domain
3569 * under consideration.
3570 * @this_cpu: Cpu at which we're currently performing load-balancing.
3571 * @imbalance: Variable to store the imbalance.
3574 * Check if we have potential to perform some power-savings balance.
3575 * If yes, set the busiest group to be the least loaded group in the
3576 * sched_domain, so that it's CPUs can be put to idle.
3578 * Returns 1 if there is potential to perform power-savings balance.
3581 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3582 int this_cpu, unsigned long *imbalance)
3584 if (!sds->power_savings_balance)
3587 if (sds->this != sds->group_leader ||
3588 sds->group_leader == sds->group_min)
3591 *imbalance = sds->min_load_per_task;
3592 sds->busiest = sds->group_min;
3597 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3598 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3599 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3604 static inline void update_sd_power_savings_stats(struct sched_group *group,
3605 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3610 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3611 int this_cpu, unsigned long *imbalance)
3615 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3618 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3620 return SCHED_LOAD_SCALE;
3623 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3625 return default_scale_freq_power(sd, cpu);
3628 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3630 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3631 unsigned long smt_gain = sd->smt_gain;
3638 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3640 return default_scale_smt_power(sd, cpu);
3643 unsigned long scale_rt_power(int cpu)
3645 struct rq *rq = cpu_rq(cpu);
3646 u64 total, available;
3648 sched_avg_update(rq);
3650 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3651 available = total - rq->rt_avg;
3653 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3654 total = SCHED_LOAD_SCALE;
3656 total >>= SCHED_LOAD_SHIFT;
3658 return div_u64(available, total);
3661 static void update_cpu_power(struct sched_domain *sd, int cpu)
3663 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3664 unsigned long power = SCHED_LOAD_SCALE;
3665 struct sched_group *sdg = sd->groups;
3667 if (sched_feat(ARCH_POWER))
3668 power *= arch_scale_freq_power(sd, cpu);
3670 power *= default_scale_freq_power(sd, cpu);
3672 power >>= SCHED_LOAD_SHIFT;
3674 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3675 if (sched_feat(ARCH_POWER))
3676 power *= arch_scale_smt_power(sd, cpu);
3678 power *= default_scale_smt_power(sd, cpu);
3680 power >>= SCHED_LOAD_SHIFT;
3683 power *= scale_rt_power(cpu);
3684 power >>= SCHED_LOAD_SHIFT;
3689 sdg->cpu_power = power;
3692 static void update_group_power(struct sched_domain *sd, int cpu)
3694 struct sched_domain *child = sd->child;
3695 struct sched_group *group, *sdg = sd->groups;
3696 unsigned long power;
3699 update_cpu_power(sd, cpu);
3705 group = child->groups;
3707 power += group->cpu_power;
3708 group = group->next;
3709 } while (group != child->groups);
3711 sdg->cpu_power = power;
3715 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3716 * @sd: The sched_domain whose statistics are to be updated.
3717 * @group: sched_group whose statistics are to be updated.
3718 * @this_cpu: Cpu for which load balance is currently performed.
3719 * @idle: Idle status of this_cpu
3720 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3721 * @sd_idle: Idle status of the sched_domain containing group.
3722 * @local_group: Does group contain this_cpu.
3723 * @cpus: Set of cpus considered for load balancing.
3724 * @balance: Should we balance.
3725 * @sgs: variable to hold the statistics for this group.
3727 static inline void update_sg_lb_stats(struct sched_domain *sd,
3728 struct sched_group *group, int this_cpu,
3729 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3730 int local_group, const struct cpumask *cpus,
3731 int *balance, struct sg_lb_stats *sgs)
3733 unsigned long load, max_cpu_load, min_cpu_load;
3735 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3736 unsigned long sum_avg_load_per_task;
3737 unsigned long avg_load_per_task;
3740 balance_cpu = group_first_cpu(group);
3741 if (balance_cpu == this_cpu)
3742 update_group_power(sd, this_cpu);
3745 /* Tally up the load of all CPUs in the group */
3746 sum_avg_load_per_task = avg_load_per_task = 0;
3748 min_cpu_load = ~0UL;
3750 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3751 struct rq *rq = cpu_rq(i);
3753 if (*sd_idle && rq->nr_running)
3756 /* Bias balancing toward cpus of our domain */
3758 if (idle_cpu(i) && !first_idle_cpu) {
3763 load = target_load(i, load_idx);
3765 load = source_load(i, load_idx);
3766 if (load > max_cpu_load)
3767 max_cpu_load = load;
3768 if (min_cpu_load > load)
3769 min_cpu_load = load;
3772 sgs->group_load += load;
3773 sgs->sum_nr_running += rq->nr_running;
3774 sgs->sum_weighted_load += weighted_cpuload(i);
3776 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3780 * First idle cpu or the first cpu(busiest) in this sched group
3781 * is eligible for doing load balancing at this and above
3782 * domains. In the newly idle case, we will allow all the cpu's
3783 * to do the newly idle load balance.
3785 if (idle != CPU_NEWLY_IDLE && local_group &&
3786 balance_cpu != this_cpu && balance) {
3791 /* Adjust by relative CPU power of the group */
3792 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3796 * Consider the group unbalanced when the imbalance is larger
3797 * than the average weight of two tasks.
3799 * APZ: with cgroup the avg task weight can vary wildly and
3800 * might not be a suitable number - should we keep a
3801 * normalized nr_running number somewhere that negates
3804 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3807 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3810 sgs->group_capacity =
3811 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3815 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3816 * @sd: sched_domain whose statistics are to be updated.
3817 * @this_cpu: Cpu for which load balance is currently performed.
3818 * @idle: Idle status of this_cpu
3819 * @sd_idle: Idle status of the sched_domain containing group.
3820 * @cpus: Set of cpus considered for load balancing.
3821 * @balance: Should we balance.
3822 * @sds: variable to hold the statistics for this sched_domain.
3824 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3825 enum cpu_idle_type idle, int *sd_idle,
3826 const struct cpumask *cpus, int *balance,
3827 struct sd_lb_stats *sds)
3829 struct sched_domain *child = sd->child;
3830 struct sched_group *group = sd->groups;
3831 struct sg_lb_stats sgs;
3832 int load_idx, prefer_sibling = 0;
3834 if (child && child->flags & SD_PREFER_SIBLING)
3837 init_sd_power_savings_stats(sd, sds, idle);
3838 load_idx = get_sd_load_idx(sd, idle);
3843 local_group = cpumask_test_cpu(this_cpu,
3844 sched_group_cpus(group));
3845 memset(&sgs, 0, sizeof(sgs));
3846 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3847 local_group, cpus, balance, &sgs);
3849 if (local_group && balance && !(*balance))
3852 sds->total_load += sgs.group_load;
3853 sds->total_pwr += group->cpu_power;
3856 * In case the child domain prefers tasks go to siblings
3857 * first, lower the group capacity to one so that we'll try
3858 * and move all the excess tasks away.
3861 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3864 sds->this_load = sgs.avg_load;
3866 sds->this_nr_running = sgs.sum_nr_running;
3867 sds->this_load_per_task = sgs.sum_weighted_load;
3868 } else if (sgs.avg_load > sds->max_load &&
3869 (sgs.sum_nr_running > sgs.group_capacity ||
3871 sds->max_load = sgs.avg_load;
3872 sds->busiest = group;
3873 sds->busiest_nr_running = sgs.sum_nr_running;
3874 sds->busiest_load_per_task = sgs.sum_weighted_load;
3875 sds->group_imb = sgs.group_imb;
3878 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3879 group = group->next;
3880 } while (group != sd->groups);
3884 * fix_small_imbalance - Calculate the minor imbalance that exists
3885 * amongst the groups of a sched_domain, during
3887 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3888 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3889 * @imbalance: Variable to store the imbalance.
3891 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3892 int this_cpu, unsigned long *imbalance)
3894 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3895 unsigned int imbn = 2;
3897 if (sds->this_nr_running) {
3898 sds->this_load_per_task /= sds->this_nr_running;
3899 if (sds->busiest_load_per_task >
3900 sds->this_load_per_task)
3903 sds->this_load_per_task =
3904 cpu_avg_load_per_task(this_cpu);
3906 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3907 sds->busiest_load_per_task * imbn) {
3908 *imbalance = sds->busiest_load_per_task;
3913 * OK, we don't have enough imbalance to justify moving tasks,
3914 * however we may be able to increase total CPU power used by
3918 pwr_now += sds->busiest->cpu_power *
3919 min(sds->busiest_load_per_task, sds->max_load);
3920 pwr_now += sds->this->cpu_power *
3921 min(sds->this_load_per_task, sds->this_load);
3922 pwr_now /= SCHED_LOAD_SCALE;
3924 /* Amount of load we'd subtract */
3925 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3926 sds->busiest->cpu_power;
3927 if (sds->max_load > tmp)
3928 pwr_move += sds->busiest->cpu_power *
3929 min(sds->busiest_load_per_task, sds->max_load - tmp);
3931 /* Amount of load we'd add */
3932 if (sds->max_load * sds->busiest->cpu_power <
3933 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3934 tmp = (sds->max_load * sds->busiest->cpu_power) /
3935 sds->this->cpu_power;
3937 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3938 sds->this->cpu_power;
3939 pwr_move += sds->this->cpu_power *
3940 min(sds->this_load_per_task, sds->this_load + tmp);
3941 pwr_move /= SCHED_LOAD_SCALE;
3943 /* Move if we gain throughput */
3944 if (pwr_move > pwr_now)
3945 *imbalance = sds->busiest_load_per_task;
3949 * calculate_imbalance - Calculate the amount of imbalance present within the
3950 * groups of a given sched_domain during load balance.
3951 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3952 * @this_cpu: Cpu for which currently load balance is being performed.
3953 * @imbalance: The variable to store the imbalance.
3955 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3956 unsigned long *imbalance)
3958 unsigned long max_pull;
3960 * In the presence of smp nice balancing, certain scenarios can have
3961 * max load less than avg load(as we skip the groups at or below
3962 * its cpu_power, while calculating max_load..)
3964 if (sds->max_load < sds->avg_load) {
3966 return fix_small_imbalance(sds, this_cpu, imbalance);
3969 /* Don't want to pull so many tasks that a group would go idle */
3970 max_pull = min(sds->max_load - sds->avg_load,
3971 sds->max_load - sds->busiest_load_per_task);
3973 /* How much load to actually move to equalise the imbalance */
3974 *imbalance = min(max_pull * sds->busiest->cpu_power,
3975 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3979 * if *imbalance is less than the average load per runnable task
3980 * there is no gaurantee that any tasks will be moved so we'll have
3981 * a think about bumping its value to force at least one task to be
3984 if (*imbalance < sds->busiest_load_per_task)
3985 return fix_small_imbalance(sds, this_cpu, imbalance);
3988 /******* find_busiest_group() helpers end here *********************/
3991 * find_busiest_group - Returns the busiest group within the sched_domain
3992 * if there is an imbalance. If there isn't an imbalance, and
3993 * the user has opted for power-savings, it returns a group whose
3994 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3995 * such a group exists.
3997 * Also calculates the amount of weighted load which should be moved
3998 * to restore balance.
4000 * @sd: The sched_domain whose busiest group is to be returned.
4001 * @this_cpu: The cpu for which load balancing is currently being performed.
4002 * @imbalance: Variable which stores amount of weighted load which should
4003 * be moved to restore balance/put a group to idle.
4004 * @idle: The idle status of this_cpu.
4005 * @sd_idle: The idleness of sd
4006 * @cpus: The set of CPUs under consideration for load-balancing.
4007 * @balance: Pointer to a variable indicating if this_cpu
4008 * is the appropriate cpu to perform load balancing at this_level.
4010 * Returns: - the busiest group if imbalance exists.
4011 * - If no imbalance and user has opted for power-savings balance,
4012 * return the least loaded group whose CPUs can be
4013 * put to idle by rebalancing its tasks onto our group.
4015 static struct sched_group *
4016 find_busiest_group(struct sched_domain *sd, int this_cpu,
4017 unsigned long *imbalance, enum cpu_idle_type idle,
4018 int *sd_idle, const struct cpumask *cpus, int *balance)
4020 struct sd_lb_stats sds;
4022 memset(&sds, 0, sizeof(sds));
4025 * Compute the various statistics relavent for load balancing at
4028 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4031 /* Cases where imbalance does not exist from POV of this_cpu */
4032 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4034 * 2) There is no busy sibling group to pull from.
4035 * 3) This group is the busiest group.
4036 * 4) This group is more busy than the avg busieness at this
4038 * 5) The imbalance is within the specified limit.
4039 * 6) Any rebalance would lead to ping-pong
4041 if (balance && !(*balance))
4044 if (!sds.busiest || sds.busiest_nr_running == 0)
4047 if (sds.this_load >= sds.max_load)
4050 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4052 if (sds.this_load >= sds.avg_load)
4055 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4058 sds.busiest_load_per_task /= sds.busiest_nr_running;
4060 sds.busiest_load_per_task =
4061 min(sds.busiest_load_per_task, sds.avg_load);
4064 * We're trying to get all the cpus to the average_load, so we don't
4065 * want to push ourselves above the average load, nor do we wish to
4066 * reduce the max loaded cpu below the average load, as either of these
4067 * actions would just result in more rebalancing later, and ping-pong
4068 * tasks around. Thus we look for the minimum possible imbalance.
4069 * Negative imbalances (*we* are more loaded than anyone else) will
4070 * be counted as no imbalance for these purposes -- we can't fix that
4071 * by pulling tasks to us. Be careful of negative numbers as they'll
4072 * appear as very large values with unsigned longs.
4074 if (sds.max_load <= sds.busiest_load_per_task)
4077 /* Looks like there is an imbalance. Compute it */
4078 calculate_imbalance(&sds, this_cpu, imbalance);
4083 * There is no obvious imbalance. But check if we can do some balancing
4086 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4094 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4097 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4098 unsigned long imbalance, const struct cpumask *cpus)
4100 struct rq *busiest = NULL, *rq;
4101 unsigned long max_load = 0;
4104 for_each_cpu(i, sched_group_cpus(group)) {
4105 unsigned long power = power_of(i);
4106 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4109 if (!cpumask_test_cpu(i, cpus))
4113 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4116 if (capacity && rq->nr_running == 1 && wl > imbalance)
4119 if (wl > max_load) {
4129 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4130 * so long as it is large enough.
4132 #define MAX_PINNED_INTERVAL 512
4134 /* Working cpumask for load_balance and load_balance_newidle. */
4135 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4138 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4139 * tasks if there is an imbalance.
4141 static int load_balance(int this_cpu, struct rq *this_rq,
4142 struct sched_domain *sd, enum cpu_idle_type idle,
4145 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4146 struct sched_group *group;
4147 unsigned long imbalance;
4149 unsigned long flags;
4150 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4152 cpumask_copy(cpus, cpu_active_mask);
4155 * When power savings policy is enabled for the parent domain, idle
4156 * sibling can pick up load irrespective of busy siblings. In this case,
4157 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4158 * portraying it as CPU_NOT_IDLE.
4160 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4161 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4164 schedstat_inc(sd, lb_count[idle]);
4168 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4175 schedstat_inc(sd, lb_nobusyg[idle]);
4179 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4181 schedstat_inc(sd, lb_nobusyq[idle]);
4185 BUG_ON(busiest == this_rq);
4187 schedstat_add(sd, lb_imbalance[idle], imbalance);
4190 if (busiest->nr_running > 1) {
4192 * Attempt to move tasks. If find_busiest_group has found
4193 * an imbalance but busiest->nr_running <= 1, the group is
4194 * still unbalanced. ld_moved simply stays zero, so it is
4195 * correctly treated as an imbalance.
4197 local_irq_save(flags);
4198 double_rq_lock(this_rq, busiest);
4199 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4200 imbalance, sd, idle, &all_pinned);
4201 double_rq_unlock(this_rq, busiest);
4202 local_irq_restore(flags);
4205 * some other cpu did the load balance for us.
4207 if (ld_moved && this_cpu != smp_processor_id())
4208 resched_cpu(this_cpu);
4210 /* All tasks on this runqueue were pinned by CPU affinity */
4211 if (unlikely(all_pinned)) {
4212 cpumask_clear_cpu(cpu_of(busiest), cpus);
4213 if (!cpumask_empty(cpus))
4220 schedstat_inc(sd, lb_failed[idle]);
4221 sd->nr_balance_failed++;
4223 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4225 raw_spin_lock_irqsave(&busiest->lock, flags);
4227 /* don't kick the migration_thread, if the curr
4228 * task on busiest cpu can't be moved to this_cpu
4230 if (!cpumask_test_cpu(this_cpu,
4231 &busiest->curr->cpus_allowed)) {
4232 raw_spin_unlock_irqrestore(&busiest->lock,
4235 goto out_one_pinned;
4238 if (!busiest->active_balance) {
4239 busiest->active_balance = 1;
4240 busiest->push_cpu = this_cpu;
4243 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4245 wake_up_process(busiest->migration_thread);
4248 * We've kicked active balancing, reset the failure
4251 sd->nr_balance_failed = sd->cache_nice_tries+1;
4254 sd->nr_balance_failed = 0;
4256 if (likely(!active_balance)) {
4257 /* We were unbalanced, so reset the balancing interval */
4258 sd->balance_interval = sd->min_interval;
4261 * If we've begun active balancing, start to back off. This
4262 * case may not be covered by the all_pinned logic if there
4263 * is only 1 task on the busy runqueue (because we don't call
4266 if (sd->balance_interval < sd->max_interval)
4267 sd->balance_interval *= 2;
4270 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4271 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4277 schedstat_inc(sd, lb_balanced[idle]);
4279 sd->nr_balance_failed = 0;
4282 /* tune up the balancing interval */
4283 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4284 (sd->balance_interval < sd->max_interval))
4285 sd->balance_interval *= 2;
4287 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4288 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4299 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4300 * tasks if there is an imbalance.
4302 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4303 * this_rq is locked.
4306 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4308 struct sched_group *group;
4309 struct rq *busiest = NULL;
4310 unsigned long imbalance;
4314 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4316 cpumask_copy(cpus, cpu_active_mask);
4319 * When power savings policy is enabled for the parent domain, idle
4320 * sibling can pick up load irrespective of busy siblings. In this case,
4321 * let the state of idle sibling percolate up as IDLE, instead of
4322 * portraying it as CPU_NOT_IDLE.
4324 if (sd->flags & SD_SHARE_CPUPOWER &&
4325 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4328 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4330 update_shares_locked(this_rq, sd);
4331 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4332 &sd_idle, cpus, NULL);
4334 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4338 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4340 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4344 BUG_ON(busiest == this_rq);
4346 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4349 if (busiest->nr_running > 1) {
4350 /* Attempt to move tasks */
4351 double_lock_balance(this_rq, busiest);
4352 /* this_rq->clock is already updated */
4353 update_rq_clock(busiest);
4354 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4355 imbalance, sd, CPU_NEWLY_IDLE,
4357 double_unlock_balance(this_rq, busiest);
4359 if (unlikely(all_pinned)) {
4360 cpumask_clear_cpu(cpu_of(busiest), cpus);
4361 if (!cpumask_empty(cpus))
4367 int active_balance = 0;
4369 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4370 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4371 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4374 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4377 if (sd->nr_balance_failed++ < 2)
4381 * The only task running in a non-idle cpu can be moved to this
4382 * cpu in an attempt to completely freeup the other CPU
4383 * package. The same method used to move task in load_balance()
4384 * have been extended for load_balance_newidle() to speedup
4385 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4387 * The package power saving logic comes from
4388 * find_busiest_group(). If there are no imbalance, then
4389 * f_b_g() will return NULL. However when sched_mc={1,2} then
4390 * f_b_g() will select a group from which a running task may be
4391 * pulled to this cpu in order to make the other package idle.
4392 * If there is no opportunity to make a package idle and if
4393 * there are no imbalance, then f_b_g() will return NULL and no
4394 * action will be taken in load_balance_newidle().
4396 * Under normal task pull operation due to imbalance, there
4397 * will be more than one task in the source run queue and
4398 * move_tasks() will succeed. ld_moved will be true and this
4399 * active balance code will not be triggered.
4402 /* Lock busiest in correct order while this_rq is held */
4403 double_lock_balance(this_rq, busiest);
4406 * don't kick the migration_thread, if the curr
4407 * task on busiest cpu can't be moved to this_cpu
4409 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4410 double_unlock_balance(this_rq, busiest);
4415 if (!busiest->active_balance) {
4416 busiest->active_balance = 1;
4417 busiest->push_cpu = this_cpu;
4421 double_unlock_balance(this_rq, busiest);
4423 * Should not call ttwu while holding a rq->lock
4425 raw_spin_unlock(&this_rq->lock);
4427 wake_up_process(busiest->migration_thread);
4428 raw_spin_lock(&this_rq->lock);
4431 sd->nr_balance_failed = 0;
4433 update_shares_locked(this_rq, sd);
4437 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4438 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4439 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4441 sd->nr_balance_failed = 0;
4447 * idle_balance is called by schedule() if this_cpu is about to become
4448 * idle. Attempts to pull tasks from other CPUs.
4450 static void idle_balance(int this_cpu, struct rq *this_rq)
4452 struct sched_domain *sd;
4453 int pulled_task = 0;
4454 unsigned long next_balance = jiffies + HZ;
4456 this_rq->idle_stamp = this_rq->clock;
4458 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4461 for_each_domain(this_cpu, sd) {
4462 unsigned long interval;
4464 if (!(sd->flags & SD_LOAD_BALANCE))
4467 if (sd->flags & SD_BALANCE_NEWIDLE)
4468 /* If we've pulled tasks over stop searching: */
4469 pulled_task = load_balance_newidle(this_cpu, this_rq,
4472 interval = msecs_to_jiffies(sd->balance_interval);
4473 if (time_after(next_balance, sd->last_balance + interval))
4474 next_balance = sd->last_balance + interval;
4476 this_rq->idle_stamp = 0;
4480 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4482 * We are going idle. next_balance may be set based on
4483 * a busy processor. So reset next_balance.
4485 this_rq->next_balance = next_balance;
4490 * active_load_balance is run by migration threads. It pushes running tasks
4491 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4492 * running on each physical CPU where possible, and avoids physical /
4493 * logical imbalances.
4495 * Called with busiest_rq locked.
4497 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4499 int target_cpu = busiest_rq->push_cpu;
4500 struct sched_domain *sd;
4501 struct rq *target_rq;
4503 /* Is there any task to move? */
4504 if (busiest_rq->nr_running <= 1)
4507 target_rq = cpu_rq(target_cpu);
4510 * This condition is "impossible", if it occurs
4511 * we need to fix it. Originally reported by
4512 * Bjorn Helgaas on a 128-cpu setup.
4514 BUG_ON(busiest_rq == target_rq);
4516 /* move a task from busiest_rq to target_rq */
4517 double_lock_balance(busiest_rq, target_rq);
4518 update_rq_clock(busiest_rq);
4519 update_rq_clock(target_rq);
4521 /* Search for an sd spanning us and the target CPU. */
4522 for_each_domain(target_cpu, sd) {
4523 if ((sd->flags & SD_LOAD_BALANCE) &&
4524 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4529 schedstat_inc(sd, alb_count);
4531 if (move_one_task(target_rq, target_cpu, busiest_rq,
4533 schedstat_inc(sd, alb_pushed);
4535 schedstat_inc(sd, alb_failed);
4537 double_unlock_balance(busiest_rq, target_rq);
4542 atomic_t load_balancer;
4543 cpumask_var_t cpu_mask;
4544 cpumask_var_t ilb_grp_nohz_mask;
4545 } nohz ____cacheline_aligned = {
4546 .load_balancer = ATOMIC_INIT(-1),
4549 int get_nohz_load_balancer(void)
4551 return atomic_read(&nohz.load_balancer);
4554 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4556 * lowest_flag_domain - Return lowest sched_domain containing flag.
4557 * @cpu: The cpu whose lowest level of sched domain is to
4559 * @flag: The flag to check for the lowest sched_domain
4560 * for the given cpu.
4562 * Returns the lowest sched_domain of a cpu which contains the given flag.
4564 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4566 struct sched_domain *sd;
4568 for_each_domain(cpu, sd)
4569 if (sd && (sd->flags & flag))
4576 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4577 * @cpu: The cpu whose domains we're iterating over.
4578 * @sd: variable holding the value of the power_savings_sd
4580 * @flag: The flag to filter the sched_domains to be iterated.
4582 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4583 * set, starting from the lowest sched_domain to the highest.
4585 #define for_each_flag_domain(cpu, sd, flag) \
4586 for (sd = lowest_flag_domain(cpu, flag); \
4587 (sd && (sd->flags & flag)); sd = sd->parent)
4590 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4591 * @ilb_group: group to be checked for semi-idleness
4593 * Returns: 1 if the group is semi-idle. 0 otherwise.
4595 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4596 * and atleast one non-idle CPU. This helper function checks if the given
4597 * sched_group is semi-idle or not.
4599 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4601 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4602 sched_group_cpus(ilb_group));
4605 * A sched_group is semi-idle when it has atleast one busy cpu
4606 * and atleast one idle cpu.
4608 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4611 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4617 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4618 * @cpu: The cpu which is nominating a new idle_load_balancer.
4620 * Returns: Returns the id of the idle load balancer if it exists,
4621 * Else, returns >= nr_cpu_ids.
4623 * This algorithm picks the idle load balancer such that it belongs to a
4624 * semi-idle powersavings sched_domain. The idea is to try and avoid
4625 * completely idle packages/cores just for the purpose of idle load balancing
4626 * when there are other idle cpu's which are better suited for that job.
4628 static int find_new_ilb(int cpu)
4630 struct sched_domain *sd;
4631 struct sched_group *ilb_group;
4634 * Have idle load balancer selection from semi-idle packages only
4635 * when power-aware load balancing is enabled
4637 if (!(sched_smt_power_savings || sched_mc_power_savings))
4641 * Optimize for the case when we have no idle CPUs or only one
4642 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4644 if (cpumask_weight(nohz.cpu_mask) < 2)
4647 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4648 ilb_group = sd->groups;
4651 if (is_semi_idle_group(ilb_group))
4652 return cpumask_first(nohz.ilb_grp_nohz_mask);
4654 ilb_group = ilb_group->next;
4656 } while (ilb_group != sd->groups);
4660 return cpumask_first(nohz.cpu_mask);
4662 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4663 static inline int find_new_ilb(int call_cpu)
4665 return cpumask_first(nohz.cpu_mask);
4670 * This routine will try to nominate the ilb (idle load balancing)
4671 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4672 * load balancing on behalf of all those cpus. If all the cpus in the system
4673 * go into this tickless mode, then there will be no ilb owner (as there is
4674 * no need for one) and all the cpus will sleep till the next wakeup event
4677 * For the ilb owner, tick is not stopped. And this tick will be used
4678 * for idle load balancing. ilb owner will still be part of
4681 * While stopping the tick, this cpu will become the ilb owner if there
4682 * is no other owner. And will be the owner till that cpu becomes busy
4683 * or if all cpus in the system stop their ticks at which point
4684 * there is no need for ilb owner.
4686 * When the ilb owner becomes busy, it nominates another owner, during the
4687 * next busy scheduler_tick()
4689 int select_nohz_load_balancer(int stop_tick)
4691 int cpu = smp_processor_id();
4694 cpu_rq(cpu)->in_nohz_recently = 1;
4696 if (!cpu_active(cpu)) {
4697 if (atomic_read(&nohz.load_balancer) != cpu)
4701 * If we are going offline and still the leader,
4704 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4710 cpumask_set_cpu(cpu, nohz.cpu_mask);
4712 /* time for ilb owner also to sleep */
4713 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4714 if (atomic_read(&nohz.load_balancer) == cpu)
4715 atomic_set(&nohz.load_balancer, -1);
4719 if (atomic_read(&nohz.load_balancer) == -1) {
4720 /* make me the ilb owner */
4721 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4723 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4726 if (!(sched_smt_power_savings ||
4727 sched_mc_power_savings))
4730 * Check to see if there is a more power-efficient
4733 new_ilb = find_new_ilb(cpu);
4734 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4735 atomic_set(&nohz.load_balancer, -1);
4736 resched_cpu(new_ilb);
4742 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4745 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4747 if (atomic_read(&nohz.load_balancer) == cpu)
4748 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4755 static DEFINE_SPINLOCK(balancing);
4758 * It checks each scheduling domain to see if it is due to be balanced,
4759 * and initiates a balancing operation if so.
4761 * Balancing parameters are set up in arch_init_sched_domains.
4763 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4766 struct rq *rq = cpu_rq(cpu);
4767 unsigned long interval;
4768 struct sched_domain *sd;
4769 /* Earliest time when we have to do rebalance again */
4770 unsigned long next_balance = jiffies + 60*HZ;
4771 int update_next_balance = 0;
4774 for_each_domain(cpu, sd) {
4775 if (!(sd->flags & SD_LOAD_BALANCE))
4778 interval = sd->balance_interval;
4779 if (idle != CPU_IDLE)
4780 interval *= sd->busy_factor;
4782 /* scale ms to jiffies */
4783 interval = msecs_to_jiffies(interval);
4784 if (unlikely(!interval))
4786 if (interval > HZ*NR_CPUS/10)
4787 interval = HZ*NR_CPUS/10;
4789 need_serialize = sd->flags & SD_SERIALIZE;
4791 if (need_serialize) {
4792 if (!spin_trylock(&balancing))
4796 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4797 if (load_balance(cpu, rq, sd, idle, &balance)) {
4799 * We've pulled tasks over so either we're no
4800 * longer idle, or one of our SMT siblings is
4803 idle = CPU_NOT_IDLE;
4805 sd->last_balance = jiffies;
4808 spin_unlock(&balancing);
4810 if (time_after(next_balance, sd->last_balance + interval)) {
4811 next_balance = sd->last_balance + interval;
4812 update_next_balance = 1;
4816 * Stop the load balance at this level. There is another
4817 * CPU in our sched group which is doing load balancing more
4825 * next_balance will be updated only when there is a need.
4826 * When the cpu is attached to null domain for ex, it will not be
4829 if (likely(update_next_balance))
4830 rq->next_balance = next_balance;
4834 * run_rebalance_domains is triggered when needed from the scheduler tick.
4835 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4836 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4838 static void run_rebalance_domains(struct softirq_action *h)
4840 int this_cpu = smp_processor_id();
4841 struct rq *this_rq = cpu_rq(this_cpu);
4842 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4843 CPU_IDLE : CPU_NOT_IDLE;
4845 rebalance_domains(this_cpu, idle);
4849 * If this cpu is the owner for idle load balancing, then do the
4850 * balancing on behalf of the other idle cpus whose ticks are
4853 if (this_rq->idle_at_tick &&
4854 atomic_read(&nohz.load_balancer) == this_cpu) {
4858 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4859 if (balance_cpu == this_cpu)
4863 * If this cpu gets work to do, stop the load balancing
4864 * work being done for other cpus. Next load
4865 * balancing owner will pick it up.
4870 rebalance_domains(balance_cpu, CPU_IDLE);
4872 rq = cpu_rq(balance_cpu);
4873 if (time_after(this_rq->next_balance, rq->next_balance))
4874 this_rq->next_balance = rq->next_balance;
4880 static inline int on_null_domain(int cpu)
4882 return !rcu_dereference(cpu_rq(cpu)->sd);
4886 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4888 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4889 * idle load balancing owner or decide to stop the periodic load balancing,
4890 * if the whole system is idle.
4892 static inline void trigger_load_balance(struct rq *rq, int cpu)
4896 * If we were in the nohz mode recently and busy at the current
4897 * scheduler tick, then check if we need to nominate new idle
4900 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4901 rq->in_nohz_recently = 0;
4903 if (atomic_read(&nohz.load_balancer) == cpu) {
4904 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4905 atomic_set(&nohz.load_balancer, -1);
4908 if (atomic_read(&nohz.load_balancer) == -1) {
4909 int ilb = find_new_ilb(cpu);
4911 if (ilb < nr_cpu_ids)
4917 * If this cpu is idle and doing idle load balancing for all the
4918 * cpus with ticks stopped, is it time for that to stop?
4920 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4921 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4927 * If this cpu is idle and the idle load balancing is done by
4928 * someone else, then no need raise the SCHED_SOFTIRQ
4930 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4931 cpumask_test_cpu(cpu, nohz.cpu_mask))
4934 /* Don't need to rebalance while attached to NULL domain */
4935 if (time_after_eq(jiffies, rq->next_balance) &&
4936 likely(!on_null_domain(cpu)))
4937 raise_softirq(SCHED_SOFTIRQ);
4940 #else /* CONFIG_SMP */
4943 * on UP we do not need to balance between CPUs:
4945 static inline void idle_balance(int cpu, struct rq *rq)
4951 DEFINE_PER_CPU(struct kernel_stat, kstat);
4953 EXPORT_PER_CPU_SYMBOL(kstat);
4956 * Return any ns on the sched_clock that have not yet been accounted in
4957 * @p in case that task is currently running.
4959 * Called with task_rq_lock() held on @rq.
4961 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4965 if (task_current(rq, p)) {
4966 update_rq_clock(rq);
4967 ns = rq->clock - p->se.exec_start;
4975 unsigned long long task_delta_exec(struct task_struct *p)
4977 unsigned long flags;
4981 rq = task_rq_lock(p, &flags);
4982 ns = do_task_delta_exec(p, rq);
4983 task_rq_unlock(rq, &flags);
4989 * Return accounted runtime for the task.
4990 * In case the task is currently running, return the runtime plus current's
4991 * pending runtime that have not been accounted yet.
4993 unsigned long long task_sched_runtime(struct task_struct *p)
4995 unsigned long flags;
4999 rq = task_rq_lock(p, &flags);
5000 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
5001 task_rq_unlock(rq, &flags);
5007 * Return sum_exec_runtime for the thread group.
5008 * In case the task is currently running, return the sum plus current's
5009 * pending runtime that have not been accounted yet.
5011 * Note that the thread group might have other running tasks as well,
5012 * so the return value not includes other pending runtime that other
5013 * running tasks might have.
5015 unsigned long long thread_group_sched_runtime(struct task_struct *p)
5017 struct task_cputime totals;
5018 unsigned long flags;
5022 rq = task_rq_lock(p, &flags);
5023 thread_group_cputime(p, &totals);
5024 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
5025 task_rq_unlock(rq, &flags);
5031 * Account user cpu time to a process.
5032 * @p: the process that the cpu time gets accounted to
5033 * @cputime: the cpu time spent in user space since the last update
5034 * @cputime_scaled: cputime scaled by cpu frequency
5036 void account_user_time(struct task_struct *p, cputime_t cputime,
5037 cputime_t cputime_scaled)
5039 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5042 /* Add user time to process. */
5043 p->utime = cputime_add(p->utime, cputime);
5044 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5045 account_group_user_time(p, cputime);
5047 /* Add user time to cpustat. */
5048 tmp = cputime_to_cputime64(cputime);
5049 if (TASK_NICE(p) > 0)
5050 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5052 cpustat->user = cputime64_add(cpustat->user, tmp);
5054 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5055 /* Account for user time used */
5056 acct_update_integrals(p);
5060 * Account guest cpu time to a process.
5061 * @p: the process that the cpu time gets accounted to
5062 * @cputime: the cpu time spent in virtual machine since the last update
5063 * @cputime_scaled: cputime scaled by cpu frequency
5065 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5066 cputime_t cputime_scaled)
5069 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5071 tmp = cputime_to_cputime64(cputime);
5073 /* Add guest time to process. */
5074 p->utime = cputime_add(p->utime, cputime);
5075 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5076 account_group_user_time(p, cputime);
5077 p->gtime = cputime_add(p->gtime, cputime);
5079 /* Add guest time to cpustat. */
5080 if (TASK_NICE(p) > 0) {
5081 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5082 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
5084 cpustat->user = cputime64_add(cpustat->user, tmp);
5085 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5090 * Account system cpu time to a process.
5091 * @p: the process that the cpu time gets accounted to
5092 * @hardirq_offset: the offset to subtract from hardirq_count()
5093 * @cputime: the cpu time spent in kernel space since the last update
5094 * @cputime_scaled: cputime scaled by cpu frequency
5096 void account_system_time(struct task_struct *p, int hardirq_offset,
5097 cputime_t cputime, cputime_t cputime_scaled)
5099 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5102 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5103 account_guest_time(p, cputime, cputime_scaled);
5107 /* Add system time to process. */
5108 p->stime = cputime_add(p->stime, cputime);
5109 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5110 account_group_system_time(p, cputime);
5112 /* Add system time to cpustat. */
5113 tmp = cputime_to_cputime64(cputime);
5114 if (hardirq_count() - hardirq_offset)
5115 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5116 else if (softirq_count())
5117 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5119 cpustat->system = cputime64_add(cpustat->system, tmp);
5121 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5123 /* Account for system time used */
5124 acct_update_integrals(p);
5128 * Account for involuntary wait time.
5129 * @steal: the cpu time spent in involuntary wait
5131 void account_steal_time(cputime_t cputime)
5133 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5134 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5136 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5140 * Account for idle time.
5141 * @cputime: the cpu time spent in idle wait
5143 void account_idle_time(cputime_t cputime)
5145 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5146 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5147 struct rq *rq = this_rq();
5149 if (atomic_read(&rq->nr_iowait) > 0)
5150 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5152 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5155 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5158 * Account a single tick of cpu time.
5159 * @p: the process that the cpu time gets accounted to
5160 * @user_tick: indicates if the tick is a user or a system tick
5162 void account_process_tick(struct task_struct *p, int user_tick)
5164 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5165 struct rq *rq = this_rq();
5168 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5169 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5170 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5173 account_idle_time(cputime_one_jiffy);
5177 * Account multiple ticks of steal time.
5178 * @p: the process from which the cpu time has been stolen
5179 * @ticks: number of stolen ticks
5181 void account_steal_ticks(unsigned long ticks)
5183 account_steal_time(jiffies_to_cputime(ticks));
5187 * Account multiple ticks of idle time.
5188 * @ticks: number of stolen ticks
5190 void account_idle_ticks(unsigned long ticks)
5192 account_idle_time(jiffies_to_cputime(ticks));
5198 * Use precise platform statistics if available:
5200 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5201 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5207 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5209 struct task_cputime cputime;
5211 thread_group_cputime(p, &cputime);
5213 *ut = cputime.utime;
5214 *st = cputime.stime;
5218 #ifndef nsecs_to_cputime
5219 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
5222 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5224 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
5227 * Use CFS's precise accounting:
5229 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
5234 temp = (u64)(rtime * utime);
5235 do_div(temp, total);
5236 utime = (cputime_t)temp;
5241 * Compare with previous values, to keep monotonicity:
5243 p->prev_utime = max(p->prev_utime, utime);
5244 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
5246 *ut = p->prev_utime;
5247 *st = p->prev_stime;
5251 * Must be called with siglock held.
5253 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5255 struct signal_struct *sig = p->signal;
5256 struct task_cputime cputime;
5257 cputime_t rtime, utime, total;
5259 thread_group_cputime(p, &cputime);
5261 total = cputime_add(cputime.utime, cputime.stime);
5262 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5267 temp = (u64)(rtime * cputime.utime);
5268 do_div(temp, total);
5269 utime = (cputime_t)temp;
5273 sig->prev_utime = max(sig->prev_utime, utime);
5274 sig->prev_stime = max(sig->prev_stime,
5275 cputime_sub(rtime, sig->prev_utime));
5277 *ut = sig->prev_utime;
5278 *st = sig->prev_stime;
5283 * This function gets called by the timer code, with HZ frequency.
5284 * We call it with interrupts disabled.
5286 * It also gets called by the fork code, when changing the parent's
5289 void scheduler_tick(void)
5291 int cpu = smp_processor_id();
5292 struct rq *rq = cpu_rq(cpu);
5293 struct task_struct *curr = rq->curr;
5297 raw_spin_lock(&rq->lock);
5298 update_rq_clock(rq);
5299 update_cpu_load(rq);
5300 curr->sched_class->task_tick(rq, curr, 0);
5301 raw_spin_unlock(&rq->lock);
5303 perf_event_task_tick(curr, cpu);
5306 rq->idle_at_tick = idle_cpu(cpu);
5307 trigger_load_balance(rq, cpu);
5311 notrace unsigned long get_parent_ip(unsigned long addr)
5313 if (in_lock_functions(addr)) {
5314 addr = CALLER_ADDR2;
5315 if (in_lock_functions(addr))
5316 addr = CALLER_ADDR3;
5321 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5322 defined(CONFIG_PREEMPT_TRACER))
5324 void __kprobes add_preempt_count(int val)
5326 #ifdef CONFIG_DEBUG_PREEMPT
5330 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5333 preempt_count() += val;
5334 #ifdef CONFIG_DEBUG_PREEMPT
5336 * Spinlock count overflowing soon?
5338 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5341 if (preempt_count() == val)
5342 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5344 EXPORT_SYMBOL(add_preempt_count);
5346 void __kprobes sub_preempt_count(int val)
5348 #ifdef CONFIG_DEBUG_PREEMPT
5352 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5355 * Is the spinlock portion underflowing?
5357 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5358 !(preempt_count() & PREEMPT_MASK)))
5362 if (preempt_count() == val)
5363 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5364 preempt_count() -= val;
5366 EXPORT_SYMBOL(sub_preempt_count);
5371 * Print scheduling while atomic bug:
5373 static noinline void __schedule_bug(struct task_struct *prev)
5375 struct pt_regs *regs = get_irq_regs();
5377 pr_err("BUG: scheduling while atomic: %s/%d/0x%08x\n",
5378 prev->comm, prev->pid, preempt_count());
5380 debug_show_held_locks(prev);
5382 if (irqs_disabled())
5383 print_irqtrace_events(prev);
5392 * Various schedule()-time debugging checks and statistics:
5394 static inline void schedule_debug(struct task_struct *prev)
5397 * Test if we are atomic. Since do_exit() needs to call into
5398 * schedule() atomically, we ignore that path for now.
5399 * Otherwise, whine if we are scheduling when we should not be.
5401 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5402 __schedule_bug(prev);
5404 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5406 schedstat_inc(this_rq(), sched_count);
5407 #ifdef CONFIG_SCHEDSTATS
5408 if (unlikely(prev->lock_depth >= 0)) {
5409 schedstat_inc(this_rq(), bkl_count);
5410 schedstat_inc(prev, sched_info.bkl_count);
5415 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5417 if (prev->state == TASK_RUNNING) {
5418 u64 runtime = prev->se.sum_exec_runtime;
5420 runtime -= prev->se.prev_sum_exec_runtime;
5421 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5424 * In order to avoid avg_overlap growing stale when we are
5425 * indeed overlapping and hence not getting put to sleep, grow
5426 * the avg_overlap on preemption.
5428 * We use the average preemption runtime because that
5429 * correlates to the amount of cache footprint a task can
5432 update_avg(&prev->se.avg_overlap, runtime);
5434 prev->sched_class->put_prev_task(rq, prev);
5438 * Pick up the highest-prio task:
5440 static inline struct task_struct *
5441 pick_next_task(struct rq *rq)
5443 const struct sched_class *class;
5444 struct task_struct *p;
5447 * Optimization: we know that if all tasks are in
5448 * the fair class we can call that function directly:
5450 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5451 p = fair_sched_class.pick_next_task(rq);
5456 class = sched_class_highest;
5458 p = class->pick_next_task(rq);
5462 * Will never be NULL as the idle class always
5463 * returns a non-NULL p:
5465 class = class->next;
5470 * schedule() is the main scheduler function.
5472 asmlinkage void __sched schedule(void)
5474 struct task_struct *prev, *next;
5475 unsigned long *switch_count;
5481 cpu = smp_processor_id();
5485 switch_count = &prev->nivcsw;
5487 release_kernel_lock(prev);
5488 need_resched_nonpreemptible:
5490 schedule_debug(prev);
5492 if (sched_feat(HRTICK))
5495 raw_spin_lock_irq(&rq->lock);
5496 update_rq_clock(rq);
5497 clear_tsk_need_resched(prev);
5499 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5500 if (unlikely(signal_pending_state(prev->state, prev)))
5501 prev->state = TASK_RUNNING;
5503 deactivate_task(rq, prev, 1);
5504 switch_count = &prev->nvcsw;
5507 pre_schedule(rq, prev);
5509 if (unlikely(!rq->nr_running))
5510 idle_balance(cpu, rq);
5512 put_prev_task(rq, prev);
5513 next = pick_next_task(rq);
5515 if (likely(prev != next)) {
5516 sched_info_switch(prev, next);
5517 perf_event_task_sched_out(prev, next, cpu);
5523 context_switch(rq, prev, next); /* unlocks the rq */
5525 * the context switch might have flipped the stack from under
5526 * us, hence refresh the local variables.
5528 cpu = smp_processor_id();
5531 raw_spin_unlock_irq(&rq->lock);
5535 if (unlikely(reacquire_kernel_lock(current) < 0))
5536 goto need_resched_nonpreemptible;
5538 preempt_enable_no_resched();
5542 EXPORT_SYMBOL(schedule);
5544 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5546 * Look out! "owner" is an entirely speculative pointer
5547 * access and not reliable.
5549 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5554 if (!sched_feat(OWNER_SPIN))
5557 #ifdef CONFIG_DEBUG_PAGEALLOC
5559 * Need to access the cpu field knowing that
5560 * DEBUG_PAGEALLOC could have unmapped it if
5561 * the mutex owner just released it and exited.
5563 if (probe_kernel_address(&owner->cpu, cpu))
5570 * Even if the access succeeded (likely case),
5571 * the cpu field may no longer be valid.
5573 if (cpu >= nr_cpumask_bits)
5577 * We need to validate that we can do a
5578 * get_cpu() and that we have the percpu area.
5580 if (!cpu_online(cpu))
5587 * Owner changed, break to re-assess state.
5589 if (lock->owner != owner)
5593 * Is that owner really running on that cpu?
5595 if (task_thread_info(rq->curr) != owner || need_resched())
5605 #ifdef CONFIG_PREEMPT
5607 * this is the entry point to schedule() from in-kernel preemption
5608 * off of preempt_enable. Kernel preemptions off return from interrupt
5609 * occur there and call schedule directly.
5611 asmlinkage void __sched preempt_schedule(void)
5613 struct thread_info *ti = current_thread_info();
5616 * If there is a non-zero preempt_count or interrupts are disabled,
5617 * we do not want to preempt the current task. Just return..
5619 if (likely(ti->preempt_count || irqs_disabled()))
5623 add_preempt_count(PREEMPT_ACTIVE);
5625 sub_preempt_count(PREEMPT_ACTIVE);
5628 * Check again in case we missed a preemption opportunity
5629 * between schedule and now.
5632 } while (need_resched());
5634 EXPORT_SYMBOL(preempt_schedule);
5637 * this is the entry point to schedule() from kernel preemption
5638 * off of irq context.
5639 * Note, that this is called and return with irqs disabled. This will
5640 * protect us against recursive calling from irq.
5642 asmlinkage void __sched preempt_schedule_irq(void)
5644 struct thread_info *ti = current_thread_info();
5646 /* Catch callers which need to be fixed */
5647 BUG_ON(ti->preempt_count || !irqs_disabled());
5650 add_preempt_count(PREEMPT_ACTIVE);
5653 local_irq_disable();
5654 sub_preempt_count(PREEMPT_ACTIVE);
5657 * Check again in case we missed a preemption opportunity
5658 * between schedule and now.
5661 } while (need_resched());
5664 #endif /* CONFIG_PREEMPT */
5666 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5669 return try_to_wake_up(curr->private, mode, wake_flags);
5671 EXPORT_SYMBOL(default_wake_function);
5674 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5675 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5676 * number) then we wake all the non-exclusive tasks and one exclusive task.
5678 * There are circumstances in which we can try to wake a task which has already
5679 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5680 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5682 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5683 int nr_exclusive, int wake_flags, void *key)
5685 wait_queue_t *curr, *next;
5687 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5688 unsigned flags = curr->flags;
5690 if (curr->func(curr, mode, wake_flags, key) &&
5691 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5697 * __wake_up - wake up threads blocked on a waitqueue.
5699 * @mode: which threads
5700 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5701 * @key: is directly passed to the wakeup function
5703 * It may be assumed that this function implies a write memory barrier before
5704 * changing the task state if and only if any tasks are woken up.
5706 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5707 int nr_exclusive, void *key)
5709 unsigned long flags;
5711 spin_lock_irqsave(&q->lock, flags);
5712 __wake_up_common(q, mode, nr_exclusive, 0, key);
5713 spin_unlock_irqrestore(&q->lock, flags);
5715 EXPORT_SYMBOL(__wake_up);
5718 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5720 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5722 __wake_up_common(q, mode, 1, 0, NULL);
5725 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5727 __wake_up_common(q, mode, 1, 0, key);
5731 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5733 * @mode: which threads
5734 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5735 * @key: opaque value to be passed to wakeup targets
5737 * The sync wakeup differs that the waker knows that it will schedule
5738 * away soon, so while the target thread will be woken up, it will not
5739 * be migrated to another CPU - ie. the two threads are 'synchronized'
5740 * with each other. This can prevent needless bouncing between CPUs.
5742 * On UP it can prevent extra preemption.
5744 * It may be assumed that this function implies a write memory barrier before
5745 * changing the task state if and only if any tasks are woken up.
5747 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5748 int nr_exclusive, void *key)
5750 unsigned long flags;
5751 int wake_flags = WF_SYNC;
5756 if (unlikely(!nr_exclusive))
5759 spin_lock_irqsave(&q->lock, flags);
5760 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5761 spin_unlock_irqrestore(&q->lock, flags);
5763 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5766 * __wake_up_sync - see __wake_up_sync_key()
5768 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5770 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5772 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5775 * complete: - signals a single thread waiting on this completion
5776 * @x: holds the state of this particular completion
5778 * This will wake up a single thread waiting on this completion. Threads will be
5779 * awakened in the same order in which they were queued.
5781 * See also complete_all(), wait_for_completion() and related routines.
5783 * It may be assumed that this function implies a write memory barrier before
5784 * changing the task state if and only if any tasks are woken up.
5786 void complete(struct completion *x)
5788 unsigned long flags;
5790 spin_lock_irqsave(&x->wait.lock, flags);
5792 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5793 spin_unlock_irqrestore(&x->wait.lock, flags);
5795 EXPORT_SYMBOL(complete);
5798 * complete_all: - signals all threads waiting on this completion
5799 * @x: holds the state of this particular completion
5801 * This will wake up all threads waiting on this particular completion event.
5803 * It may be assumed that this function implies a write memory barrier before
5804 * changing the task state if and only if any tasks are woken up.
5806 void complete_all(struct completion *x)
5808 unsigned long flags;
5810 spin_lock_irqsave(&x->wait.lock, flags);
5811 x->done += UINT_MAX/2;
5812 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5813 spin_unlock_irqrestore(&x->wait.lock, flags);
5815 EXPORT_SYMBOL(complete_all);
5817 static inline long __sched
5818 do_wait_for_common(struct completion *x, long timeout, int state)
5821 DECLARE_WAITQUEUE(wait, current);
5823 wait.flags |= WQ_FLAG_EXCLUSIVE;
5824 __add_wait_queue_tail(&x->wait, &wait);
5826 if (signal_pending_state(state, current)) {
5827 timeout = -ERESTARTSYS;
5830 __set_current_state(state);
5831 spin_unlock_irq(&x->wait.lock);
5832 timeout = schedule_timeout(timeout);
5833 spin_lock_irq(&x->wait.lock);
5834 } while (!x->done && timeout);
5835 __remove_wait_queue(&x->wait, &wait);
5840 return timeout ?: 1;
5844 wait_for_common(struct completion *x, long timeout, int state)
5848 spin_lock_irq(&x->wait.lock);
5849 timeout = do_wait_for_common(x, timeout, state);
5850 spin_unlock_irq(&x->wait.lock);
5855 * wait_for_completion: - waits for completion of a task
5856 * @x: holds the state of this particular completion
5858 * This waits to be signaled for completion of a specific task. It is NOT
5859 * interruptible and there is no timeout.
5861 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5862 * and interrupt capability. Also see complete().
5864 void __sched wait_for_completion(struct completion *x)
5866 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5868 EXPORT_SYMBOL(wait_for_completion);
5871 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5872 * @x: holds the state of this particular completion
5873 * @timeout: timeout value in jiffies
5875 * This waits for either a completion of a specific task to be signaled or for a
5876 * specified timeout to expire. The timeout is in jiffies. It is not
5879 unsigned long __sched
5880 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5882 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5884 EXPORT_SYMBOL(wait_for_completion_timeout);
5887 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5888 * @x: holds the state of this particular completion
5890 * This waits for completion of a specific task to be signaled. It is
5893 int __sched wait_for_completion_interruptible(struct completion *x)
5895 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5896 if (t == -ERESTARTSYS)
5900 EXPORT_SYMBOL(wait_for_completion_interruptible);
5903 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5904 * @x: holds the state of this particular completion
5905 * @timeout: timeout value in jiffies
5907 * This waits for either a completion of a specific task to be signaled or for a
5908 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5910 unsigned long __sched
5911 wait_for_completion_interruptible_timeout(struct completion *x,
5912 unsigned long timeout)
5914 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5916 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5919 * wait_for_completion_killable: - waits for completion of a task (killable)
5920 * @x: holds the state of this particular completion
5922 * This waits to be signaled for completion of a specific task. It can be
5923 * interrupted by a kill signal.
5925 int __sched wait_for_completion_killable(struct completion *x)
5927 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5928 if (t == -ERESTARTSYS)
5932 EXPORT_SYMBOL(wait_for_completion_killable);
5935 * try_wait_for_completion - try to decrement a completion without blocking
5936 * @x: completion structure
5938 * Returns: 0 if a decrement cannot be done without blocking
5939 * 1 if a decrement succeeded.
5941 * If a completion is being used as a counting completion,
5942 * attempt to decrement the counter without blocking. This
5943 * enables us to avoid waiting if the resource the completion
5944 * is protecting is not available.
5946 bool try_wait_for_completion(struct completion *x)
5948 unsigned long flags;
5951 spin_lock_irqsave(&x->wait.lock, flags);
5956 spin_unlock_irqrestore(&x->wait.lock, flags);
5959 EXPORT_SYMBOL(try_wait_for_completion);
5962 * completion_done - Test to see if a completion has any waiters
5963 * @x: completion structure
5965 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5966 * 1 if there are no waiters.
5969 bool completion_done(struct completion *x)
5971 unsigned long flags;
5974 spin_lock_irqsave(&x->wait.lock, flags);
5977 spin_unlock_irqrestore(&x->wait.lock, flags);
5980 EXPORT_SYMBOL(completion_done);
5983 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5985 unsigned long flags;
5988 init_waitqueue_entry(&wait, current);
5990 __set_current_state(state);
5992 spin_lock_irqsave(&q->lock, flags);
5993 __add_wait_queue(q, &wait);
5994 spin_unlock(&q->lock);
5995 timeout = schedule_timeout(timeout);
5996 spin_lock_irq(&q->lock);
5997 __remove_wait_queue(q, &wait);
5998 spin_unlock_irqrestore(&q->lock, flags);
6003 void __sched interruptible_sleep_on(wait_queue_head_t *q)
6005 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6007 EXPORT_SYMBOL(interruptible_sleep_on);
6010 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
6012 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
6014 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
6016 void __sched sleep_on(wait_queue_head_t *q)
6018 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6020 EXPORT_SYMBOL(sleep_on);
6022 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
6024 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
6026 EXPORT_SYMBOL(sleep_on_timeout);
6028 #ifdef CONFIG_RT_MUTEXES
6031 * rt_mutex_setprio - set the current priority of a task
6033 * @prio: prio value (kernel-internal form)
6035 * This function changes the 'effective' priority of a task. It does
6036 * not touch ->normal_prio like __setscheduler().
6038 * Used by the rt_mutex code to implement priority inheritance logic.
6040 void rt_mutex_setprio(struct task_struct *p, int prio)
6042 unsigned long flags;
6043 int oldprio, on_rq, running;
6045 const struct sched_class *prev_class = p->sched_class;
6047 BUG_ON(prio < 0 || prio > MAX_PRIO);
6049 rq = task_rq_lock(p, &flags);
6050 update_rq_clock(rq);
6053 on_rq = p->se.on_rq;
6054 running = task_current(rq, p);
6056 dequeue_task(rq, p, 0);
6058 p->sched_class->put_prev_task(rq, p);
6061 p->sched_class = &rt_sched_class;
6063 p->sched_class = &fair_sched_class;
6068 p->sched_class->set_curr_task(rq);
6070 enqueue_task(rq, p, 0);
6072 check_class_changed(rq, p, prev_class, oldprio, running);
6074 task_rq_unlock(rq, &flags);
6079 void set_user_nice(struct task_struct *p, long nice)
6081 int old_prio, delta, on_rq;
6082 unsigned long flags;
6085 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6088 * We have to be careful, if called from sys_setpriority(),
6089 * the task might be in the middle of scheduling on another CPU.
6091 rq = task_rq_lock(p, &flags);
6092 update_rq_clock(rq);
6094 * The RT priorities are set via sched_setscheduler(), but we still
6095 * allow the 'normal' nice value to be set - but as expected
6096 * it wont have any effect on scheduling until the task is
6097 * SCHED_FIFO/SCHED_RR:
6099 if (task_has_rt_policy(p)) {
6100 p->static_prio = NICE_TO_PRIO(nice);
6103 on_rq = p->se.on_rq;
6105 dequeue_task(rq, p, 0);
6107 p->static_prio = NICE_TO_PRIO(nice);
6110 p->prio = effective_prio(p);
6111 delta = p->prio - old_prio;
6114 enqueue_task(rq, p, 0);
6116 * If the task increased its priority or is running and
6117 * lowered its priority, then reschedule its CPU:
6119 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6120 resched_task(rq->curr);
6123 task_rq_unlock(rq, &flags);
6125 EXPORT_SYMBOL(set_user_nice);
6128 * can_nice - check if a task can reduce its nice value
6132 int can_nice(const struct task_struct *p, const int nice)
6134 /* convert nice value [19,-20] to rlimit style value [1,40] */
6135 int nice_rlim = 20 - nice;
6137 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6138 capable(CAP_SYS_NICE));
6141 #ifdef __ARCH_WANT_SYS_NICE
6144 * sys_nice - change the priority of the current process.
6145 * @increment: priority increment
6147 * sys_setpriority is a more generic, but much slower function that
6148 * does similar things.
6150 SYSCALL_DEFINE1(nice, int, increment)
6155 * Setpriority might change our priority at the same moment.
6156 * We don't have to worry. Conceptually one call occurs first
6157 * and we have a single winner.
6159 if (increment < -40)
6164 nice = TASK_NICE(current) + increment;
6170 if (increment < 0 && !can_nice(current, nice))
6173 retval = security_task_setnice(current, nice);
6177 set_user_nice(current, nice);
6184 * task_prio - return the priority value of a given task.
6185 * @p: the task in question.
6187 * This is the priority value as seen by users in /proc.
6188 * RT tasks are offset by -200. Normal tasks are centered
6189 * around 0, value goes from -16 to +15.
6191 int task_prio(const struct task_struct *p)
6193 return p->prio - MAX_RT_PRIO;
6197 * task_nice - return the nice value of a given task.
6198 * @p: the task in question.
6200 int task_nice(const struct task_struct *p)
6202 return TASK_NICE(p);
6204 EXPORT_SYMBOL(task_nice);
6207 * idle_cpu - is a given cpu idle currently?
6208 * @cpu: the processor in question.
6210 int idle_cpu(int cpu)
6212 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6216 * idle_task - return the idle task for a given cpu.
6217 * @cpu: the processor in question.
6219 struct task_struct *idle_task(int cpu)
6221 return cpu_rq(cpu)->idle;
6225 * find_process_by_pid - find a process with a matching PID value.
6226 * @pid: the pid in question.
6228 static struct task_struct *find_process_by_pid(pid_t pid)
6230 return pid ? find_task_by_vpid(pid) : current;
6233 /* Actually do priority change: must hold rq lock. */
6235 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6237 BUG_ON(p->se.on_rq);
6240 p->rt_priority = prio;
6241 p->normal_prio = normal_prio(p);
6242 /* we are holding p->pi_lock already */
6243 p->prio = rt_mutex_getprio(p);
6244 if (rt_prio(p->prio))
6245 p->sched_class = &rt_sched_class;
6247 p->sched_class = &fair_sched_class;
6252 * check the target process has a UID that matches the current process's
6254 static bool check_same_owner(struct task_struct *p)
6256 const struct cred *cred = current_cred(), *pcred;
6260 pcred = __task_cred(p);
6261 match = (cred->euid == pcred->euid ||
6262 cred->euid == pcred->uid);
6267 static int __sched_setscheduler(struct task_struct *p, int policy,
6268 struct sched_param *param, bool user)
6270 int retval, oldprio, oldpolicy = -1, on_rq, running;
6271 unsigned long flags;
6272 const struct sched_class *prev_class = p->sched_class;
6276 /* may grab non-irq protected spin_locks */
6277 BUG_ON(in_interrupt());
6279 /* double check policy once rq lock held */
6281 reset_on_fork = p->sched_reset_on_fork;
6282 policy = oldpolicy = p->policy;
6284 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6285 policy &= ~SCHED_RESET_ON_FORK;
6287 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6288 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6289 policy != SCHED_IDLE)
6294 * Valid priorities for SCHED_FIFO and SCHED_RR are
6295 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6296 * SCHED_BATCH and SCHED_IDLE is 0.
6298 if (param->sched_priority < 0 ||
6299 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6300 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6302 if (rt_policy(policy) != (param->sched_priority != 0))
6306 * Allow unprivileged RT tasks to decrease priority:
6308 if (user && !capable(CAP_SYS_NICE)) {
6309 if (rt_policy(policy)) {
6310 unsigned long rlim_rtprio;
6312 if (!lock_task_sighand(p, &flags))
6314 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6315 unlock_task_sighand(p, &flags);
6317 /* can't set/change the rt policy */
6318 if (policy != p->policy && !rlim_rtprio)
6321 /* can't increase priority */
6322 if (param->sched_priority > p->rt_priority &&
6323 param->sched_priority > rlim_rtprio)
6327 * Like positive nice levels, dont allow tasks to
6328 * move out of SCHED_IDLE either:
6330 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6333 /* can't change other user's priorities */
6334 if (!check_same_owner(p))
6337 /* Normal users shall not reset the sched_reset_on_fork flag */
6338 if (p->sched_reset_on_fork && !reset_on_fork)
6343 #ifdef CONFIG_RT_GROUP_SCHED
6345 * Do not allow realtime tasks into groups that have no runtime
6348 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6349 task_group(p)->rt_bandwidth.rt_runtime == 0)
6353 retval = security_task_setscheduler(p, policy, param);
6359 * make sure no PI-waiters arrive (or leave) while we are
6360 * changing the priority of the task:
6362 raw_spin_lock_irqsave(&p->pi_lock, flags);
6364 * To be able to change p->policy safely, the apropriate
6365 * runqueue lock must be held.
6367 rq = __task_rq_lock(p);
6368 /* recheck policy now with rq lock held */
6369 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6370 policy = oldpolicy = -1;
6371 __task_rq_unlock(rq);
6372 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6375 update_rq_clock(rq);
6376 on_rq = p->se.on_rq;
6377 running = task_current(rq, p);
6379 deactivate_task(rq, p, 0);
6381 p->sched_class->put_prev_task(rq, p);
6383 p->sched_reset_on_fork = reset_on_fork;
6386 __setscheduler(rq, p, policy, param->sched_priority);
6389 p->sched_class->set_curr_task(rq);
6391 activate_task(rq, p, 0);
6393 check_class_changed(rq, p, prev_class, oldprio, running);
6395 __task_rq_unlock(rq);
6396 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6398 rt_mutex_adjust_pi(p);
6404 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6405 * @p: the task in question.
6406 * @policy: new policy.
6407 * @param: structure containing the new RT priority.
6409 * NOTE that the task may be already dead.
6411 int sched_setscheduler(struct task_struct *p, int policy,
6412 struct sched_param *param)
6414 return __sched_setscheduler(p, policy, param, true);
6416 EXPORT_SYMBOL_GPL(sched_setscheduler);
6419 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6420 * @p: the task in question.
6421 * @policy: new policy.
6422 * @param: structure containing the new RT priority.
6424 * Just like sched_setscheduler, only don't bother checking if the
6425 * current context has permission. For example, this is needed in
6426 * stop_machine(): we create temporary high priority worker threads,
6427 * but our caller might not have that capability.
6429 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6430 struct sched_param *param)
6432 return __sched_setscheduler(p, policy, param, false);
6436 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6438 struct sched_param lparam;
6439 struct task_struct *p;
6442 if (!param || pid < 0)
6444 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6449 p = find_process_by_pid(pid);
6451 retval = sched_setscheduler(p, policy, &lparam);
6458 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6459 * @pid: the pid in question.
6460 * @policy: new policy.
6461 * @param: structure containing the new RT priority.
6463 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6464 struct sched_param __user *, param)
6466 /* negative values for policy are not valid */
6470 return do_sched_setscheduler(pid, policy, param);
6474 * sys_sched_setparam - set/change the RT priority of a thread
6475 * @pid: the pid in question.
6476 * @param: structure containing the new RT priority.
6478 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6480 return do_sched_setscheduler(pid, -1, param);
6484 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6485 * @pid: the pid in question.
6487 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6489 struct task_struct *p;
6497 p = find_process_by_pid(pid);
6499 retval = security_task_getscheduler(p);
6502 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6509 * sys_sched_getparam - get the RT priority of a thread
6510 * @pid: the pid in question.
6511 * @param: structure containing the RT priority.
6513 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6515 struct sched_param lp;
6516 struct task_struct *p;
6519 if (!param || pid < 0)
6523 p = find_process_by_pid(pid);
6528 retval = security_task_getscheduler(p);
6532 lp.sched_priority = p->rt_priority;
6536 * This one might sleep, we cannot do it with a spinlock held ...
6538 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6547 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6549 cpumask_var_t cpus_allowed, new_mask;
6550 struct task_struct *p;
6556 p = find_process_by_pid(pid);
6563 /* Prevent p going away */
6567 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6571 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6573 goto out_free_cpus_allowed;
6576 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6579 retval = security_task_setscheduler(p, 0, NULL);
6583 cpuset_cpus_allowed(p, cpus_allowed);
6584 cpumask_and(new_mask, in_mask, cpus_allowed);
6586 retval = set_cpus_allowed_ptr(p, new_mask);
6589 cpuset_cpus_allowed(p, cpus_allowed);
6590 if (!cpumask_subset(new_mask, cpus_allowed)) {
6592 * We must have raced with a concurrent cpuset
6593 * update. Just reset the cpus_allowed to the
6594 * cpuset's cpus_allowed
6596 cpumask_copy(new_mask, cpus_allowed);
6601 free_cpumask_var(new_mask);
6602 out_free_cpus_allowed:
6603 free_cpumask_var(cpus_allowed);
6610 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6611 struct cpumask *new_mask)
6613 if (len < cpumask_size())
6614 cpumask_clear(new_mask);
6615 else if (len > cpumask_size())
6616 len = cpumask_size();
6618 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6622 * sys_sched_setaffinity - set the cpu affinity of a process
6623 * @pid: pid of the process
6624 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6625 * @user_mask_ptr: user-space pointer to the new cpu mask
6627 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6628 unsigned long __user *, user_mask_ptr)
6630 cpumask_var_t new_mask;
6633 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6636 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6638 retval = sched_setaffinity(pid, new_mask);
6639 free_cpumask_var(new_mask);
6643 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6645 struct task_struct *p;
6646 unsigned long flags;
6654 p = find_process_by_pid(pid);
6658 retval = security_task_getscheduler(p);
6662 rq = task_rq_lock(p, &flags);
6663 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6664 task_rq_unlock(rq, &flags);
6674 * sys_sched_getaffinity - get the cpu affinity of a process
6675 * @pid: pid of the process
6676 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6677 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6679 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6680 unsigned long __user *, user_mask_ptr)
6685 if (len < cpumask_size())
6688 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6691 ret = sched_getaffinity(pid, mask);
6693 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6696 ret = cpumask_size();
6698 free_cpumask_var(mask);
6704 * sys_sched_yield - yield the current processor to other threads.
6706 * This function yields the current CPU to other tasks. If there are no
6707 * other threads running on this CPU then this function will return.
6709 SYSCALL_DEFINE0(sched_yield)
6711 struct rq *rq = this_rq_lock();
6713 schedstat_inc(rq, yld_count);
6714 current->sched_class->yield_task(rq);
6717 * Since we are going to call schedule() anyway, there's
6718 * no need to preempt or enable interrupts:
6720 __release(rq->lock);
6721 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6722 do_raw_spin_unlock(&rq->lock);
6723 preempt_enable_no_resched();
6730 static inline int should_resched(void)
6732 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6735 static void __cond_resched(void)
6737 add_preempt_count(PREEMPT_ACTIVE);
6739 sub_preempt_count(PREEMPT_ACTIVE);
6742 int __sched _cond_resched(void)
6744 if (should_resched()) {
6750 EXPORT_SYMBOL(_cond_resched);
6753 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6754 * call schedule, and on return reacquire the lock.
6756 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6757 * operations here to prevent schedule() from being called twice (once via
6758 * spin_unlock(), once by hand).
6760 int __cond_resched_lock(spinlock_t *lock)
6762 int resched = should_resched();
6765 lockdep_assert_held(lock);
6767 if (spin_needbreak(lock) || resched) {
6778 EXPORT_SYMBOL(__cond_resched_lock);
6780 int __sched __cond_resched_softirq(void)
6782 BUG_ON(!in_softirq());
6784 if (should_resched()) {
6792 EXPORT_SYMBOL(__cond_resched_softirq);
6795 * yield - yield the current processor to other threads.
6797 * This is a shortcut for kernel-space yielding - it marks the
6798 * thread runnable and calls sys_sched_yield().
6800 void __sched yield(void)
6802 set_current_state(TASK_RUNNING);
6805 EXPORT_SYMBOL(yield);
6808 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6809 * that process accounting knows that this is a task in IO wait state.
6811 void __sched io_schedule(void)
6813 struct rq *rq = raw_rq();
6815 delayacct_blkio_start();
6816 atomic_inc(&rq->nr_iowait);
6817 current->in_iowait = 1;
6819 current->in_iowait = 0;
6820 atomic_dec(&rq->nr_iowait);
6821 delayacct_blkio_end();
6823 EXPORT_SYMBOL(io_schedule);
6825 long __sched io_schedule_timeout(long timeout)
6827 struct rq *rq = raw_rq();
6830 delayacct_blkio_start();
6831 atomic_inc(&rq->nr_iowait);
6832 current->in_iowait = 1;
6833 ret = schedule_timeout(timeout);
6834 current->in_iowait = 0;
6835 atomic_dec(&rq->nr_iowait);
6836 delayacct_blkio_end();
6841 * sys_sched_get_priority_max - return maximum RT priority.
6842 * @policy: scheduling class.
6844 * this syscall returns the maximum rt_priority that can be used
6845 * by a given scheduling class.
6847 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6854 ret = MAX_USER_RT_PRIO-1;
6866 * sys_sched_get_priority_min - return minimum RT priority.
6867 * @policy: scheduling class.
6869 * this syscall returns the minimum rt_priority that can be used
6870 * by a given scheduling class.
6872 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6890 * sys_sched_rr_get_interval - return the default timeslice of a process.
6891 * @pid: pid of the process.
6892 * @interval: userspace pointer to the timeslice value.
6894 * this syscall writes the default timeslice value of a given process
6895 * into the user-space timespec buffer. A value of '0' means infinity.
6897 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6898 struct timespec __user *, interval)
6900 struct task_struct *p;
6901 unsigned int time_slice;
6902 unsigned long flags;
6912 p = find_process_by_pid(pid);
6916 retval = security_task_getscheduler(p);
6920 rq = task_rq_lock(p, &flags);
6921 time_slice = p->sched_class->get_rr_interval(rq, p);
6922 task_rq_unlock(rq, &flags);
6925 jiffies_to_timespec(time_slice, &t);
6926 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6934 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6936 void sched_show_task(struct task_struct *p)
6938 unsigned long free = 0;
6941 state = p->state ? __ffs(p->state) + 1 : 0;
6942 pr_info("%-13.13s %c", p->comm,
6943 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6944 #if BITS_PER_LONG == 32
6945 if (state == TASK_RUNNING)
6946 pr_cont(" running ");
6948 pr_cont(" %08lx ", thread_saved_pc(p));
6950 if (state == TASK_RUNNING)
6951 pr_cont(" running task ");
6953 pr_cont(" %016lx ", thread_saved_pc(p));
6955 #ifdef CONFIG_DEBUG_STACK_USAGE
6956 free = stack_not_used(p);
6958 pr_cont("%5lu %5d %6d 0x%08lx\n", free,
6959 task_pid_nr(p), task_pid_nr(p->real_parent),
6960 (unsigned long)task_thread_info(p)->flags);
6962 show_stack(p, NULL);
6965 void show_state_filter(unsigned long state_filter)
6967 struct task_struct *g, *p;
6969 #if BITS_PER_LONG == 32
6970 pr_info(" task PC stack pid father\n");
6972 pr_info(" task PC stack pid father\n");
6974 read_lock(&tasklist_lock);
6975 do_each_thread(g, p) {
6977 * reset the NMI-timeout, listing all files on a slow
6978 * console might take alot of time:
6980 touch_nmi_watchdog();
6981 if (!state_filter || (p->state & state_filter))
6983 } while_each_thread(g, p);
6985 touch_all_softlockup_watchdogs();
6987 #ifdef CONFIG_SCHED_DEBUG
6988 sysrq_sched_debug_show();
6990 read_unlock(&tasklist_lock);
6992 * Only show locks if all tasks are dumped:
6995 debug_show_all_locks();
6998 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
7000 idle->sched_class = &idle_sched_class;
7004 * init_idle - set up an idle thread for a given CPU
7005 * @idle: task in question
7006 * @cpu: cpu the idle task belongs to
7008 * NOTE: this function does not set the idle thread's NEED_RESCHED
7009 * flag, to make booting more robust.
7011 void __cpuinit init_idle(struct task_struct *idle, int cpu)
7013 struct rq *rq = cpu_rq(cpu);
7014 unsigned long flags;
7016 raw_spin_lock_irqsave(&rq->lock, flags);
7019 idle->state = TASK_RUNNING;
7020 idle->se.exec_start = sched_clock();
7022 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
7023 __set_task_cpu(idle, cpu);
7025 rq->curr = rq->idle = idle;
7026 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7029 raw_spin_unlock_irqrestore(&rq->lock, flags);
7031 /* Set the preempt count _outside_ the spinlocks! */
7032 #if defined(CONFIG_PREEMPT)
7033 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7035 task_thread_info(idle)->preempt_count = 0;
7038 * The idle tasks have their own, simple scheduling class:
7040 idle->sched_class = &idle_sched_class;
7041 ftrace_graph_init_task(idle);
7045 * In a system that switches off the HZ timer nohz_cpu_mask
7046 * indicates which cpus entered this state. This is used
7047 * in the rcu update to wait only for active cpus. For system
7048 * which do not switch off the HZ timer nohz_cpu_mask should
7049 * always be CPU_BITS_NONE.
7051 cpumask_var_t nohz_cpu_mask;
7054 * Increase the granularity value when there are more CPUs,
7055 * because with more CPUs the 'effective latency' as visible
7056 * to users decreases. But the relationship is not linear,
7057 * so pick a second-best guess by going with the log2 of the
7060 * This idea comes from the SD scheduler of Con Kolivas:
7062 static int get_update_sysctl_factor(void)
7064 unsigned int cpus = min_t(int, num_online_cpus(), 8);
7065 unsigned int factor;
7067 switch (sysctl_sched_tunable_scaling) {
7068 case SCHED_TUNABLESCALING_NONE:
7071 case SCHED_TUNABLESCALING_LINEAR:
7074 case SCHED_TUNABLESCALING_LOG:
7076 factor = 1 + ilog2(cpus);
7083 static void update_sysctl(void)
7085 unsigned int factor = get_update_sysctl_factor();
7087 #define SET_SYSCTL(name) \
7088 (sysctl_##name = (factor) * normalized_sysctl_##name)
7089 SET_SYSCTL(sched_min_granularity);
7090 SET_SYSCTL(sched_latency);
7091 SET_SYSCTL(sched_wakeup_granularity);
7092 SET_SYSCTL(sched_shares_ratelimit);
7096 static inline void sched_init_granularity(void)
7103 * This is how migration works:
7105 * 1) we queue a struct migration_req structure in the source CPU's
7106 * runqueue and wake up that CPU's migration thread.
7107 * 2) we down() the locked semaphore => thread blocks.
7108 * 3) migration thread wakes up (implicitly it forces the migrated
7109 * thread off the CPU)
7110 * 4) it gets the migration request and checks whether the migrated
7111 * task is still in the wrong runqueue.
7112 * 5) if it's in the wrong runqueue then the migration thread removes
7113 * it and puts it into the right queue.
7114 * 6) migration thread up()s the semaphore.
7115 * 7) we wake up and the migration is done.
7119 * Change a given task's CPU affinity. Migrate the thread to a
7120 * proper CPU and schedule it away if the CPU it's executing on
7121 * is removed from the allowed bitmask.
7123 * NOTE: the caller must have a valid reference to the task, the
7124 * task must not exit() & deallocate itself prematurely. The
7125 * call is not atomic; no spinlocks may be held.
7127 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7129 struct migration_req req;
7130 unsigned long flags;
7135 * Since we rely on wake-ups to migrate sleeping tasks, don't change
7136 * the ->cpus_allowed mask from under waking tasks, which would be
7137 * possible when we change rq->lock in ttwu(), so synchronize against
7138 * TASK_WAKING to avoid that.
7141 while (p->state == TASK_WAKING)
7144 rq = task_rq_lock(p, &flags);
7146 if (p->state == TASK_WAKING) {
7147 task_rq_unlock(rq, &flags);
7151 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7156 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7157 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7162 if (p->sched_class->set_cpus_allowed)
7163 p->sched_class->set_cpus_allowed(p, new_mask);
7165 cpumask_copy(&p->cpus_allowed, new_mask);
7166 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7169 /* Can the task run on the task's current CPU? If so, we're done */
7170 if (cpumask_test_cpu(task_cpu(p), new_mask))
7173 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7174 /* Need help from migration thread: drop lock and wait. */
7175 struct task_struct *mt = rq->migration_thread;
7177 get_task_struct(mt);
7178 task_rq_unlock(rq, &flags);
7179 wake_up_process(rq->migration_thread);
7180 put_task_struct(mt);
7181 wait_for_completion(&req.done);
7182 tlb_migrate_finish(p->mm);
7186 task_rq_unlock(rq, &flags);
7190 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7193 * Move (not current) task off this cpu, onto dest cpu. We're doing
7194 * this because either it can't run here any more (set_cpus_allowed()
7195 * away from this CPU, or CPU going down), or because we're
7196 * attempting to rebalance this task on exec (sched_exec).
7198 * So we race with normal scheduler movements, but that's OK, as long
7199 * as the task is no longer on this CPU.
7201 * Returns non-zero if task was successfully migrated.
7203 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7205 struct rq *rq_dest, *rq_src;
7208 if (unlikely(!cpu_active(dest_cpu)))
7211 rq_src = cpu_rq(src_cpu);
7212 rq_dest = cpu_rq(dest_cpu);
7214 double_rq_lock(rq_src, rq_dest);
7215 /* Already moved. */
7216 if (task_cpu(p) != src_cpu)
7218 /* Affinity changed (again). */
7219 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7223 * If we're not on a rq, the next wake-up will ensure we're
7227 deactivate_task(rq_src, p, 0);
7228 set_task_cpu(p, dest_cpu);
7229 activate_task(rq_dest, p, 0);
7230 check_preempt_curr(rq_dest, p, 0);
7235 double_rq_unlock(rq_src, rq_dest);
7239 #define RCU_MIGRATION_IDLE 0
7240 #define RCU_MIGRATION_NEED_QS 1
7241 #define RCU_MIGRATION_GOT_QS 2
7242 #define RCU_MIGRATION_MUST_SYNC 3
7245 * migration_thread - this is a highprio system thread that performs
7246 * thread migration by bumping thread off CPU then 'pushing' onto
7249 static int migration_thread(void *data)
7252 int cpu = (long)data;
7256 BUG_ON(rq->migration_thread != current);
7258 set_current_state(TASK_INTERRUPTIBLE);
7259 while (!kthread_should_stop()) {
7260 struct migration_req *req;
7261 struct list_head *head;
7263 raw_spin_lock_irq(&rq->lock);
7265 if (cpu_is_offline(cpu)) {
7266 raw_spin_unlock_irq(&rq->lock);
7270 if (rq->active_balance) {
7271 active_load_balance(rq, cpu);
7272 rq->active_balance = 0;
7275 head = &rq->migration_queue;
7277 if (list_empty(head)) {
7278 raw_spin_unlock_irq(&rq->lock);
7280 set_current_state(TASK_INTERRUPTIBLE);
7283 req = list_entry(head->next, struct migration_req, list);
7284 list_del_init(head->next);
7286 if (req->task != NULL) {
7287 raw_spin_unlock(&rq->lock);
7288 __migrate_task(req->task, cpu, req->dest_cpu);
7289 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7290 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7291 raw_spin_unlock(&rq->lock);
7293 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7294 raw_spin_unlock(&rq->lock);
7295 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7299 complete(&req->done);
7301 __set_current_state(TASK_RUNNING);
7306 #ifdef CONFIG_HOTPLUG_CPU
7308 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7312 local_irq_disable();
7313 ret = __migrate_task(p, src_cpu, dest_cpu);
7319 * Figure out where task on dead CPU should go, use force if necessary.
7321 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7326 dest_cpu = select_fallback_rq(dead_cpu, p);
7328 /* It can have affinity changed while we were choosing. */
7329 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7334 * While a dead CPU has no uninterruptible tasks queued at this point,
7335 * it might still have a nonzero ->nr_uninterruptible counter, because
7336 * for performance reasons the counter is not stricly tracking tasks to
7337 * their home CPUs. So we just add the counter to another CPU's counter,
7338 * to keep the global sum constant after CPU-down:
7340 static void migrate_nr_uninterruptible(struct rq *rq_src)
7342 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7343 unsigned long flags;
7345 local_irq_save(flags);
7346 double_rq_lock(rq_src, rq_dest);
7347 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7348 rq_src->nr_uninterruptible = 0;
7349 double_rq_unlock(rq_src, rq_dest);
7350 local_irq_restore(flags);
7353 /* Run through task list and migrate tasks from the dead cpu. */
7354 static void migrate_live_tasks(int src_cpu)
7356 struct task_struct *p, *t;
7358 read_lock(&tasklist_lock);
7360 do_each_thread(t, p) {
7364 if (task_cpu(p) == src_cpu)
7365 move_task_off_dead_cpu(src_cpu, p);
7366 } while_each_thread(t, p);
7368 read_unlock(&tasklist_lock);
7372 * Schedules idle task to be the next runnable task on current CPU.
7373 * It does so by boosting its priority to highest possible.
7374 * Used by CPU offline code.
7376 void sched_idle_next(void)
7378 int this_cpu = smp_processor_id();
7379 struct rq *rq = cpu_rq(this_cpu);
7380 struct task_struct *p = rq->idle;
7381 unsigned long flags;
7383 /* cpu has to be offline */
7384 BUG_ON(cpu_online(this_cpu));
7387 * Strictly not necessary since rest of the CPUs are stopped by now
7388 * and interrupts disabled on the current cpu.
7390 raw_spin_lock_irqsave(&rq->lock, flags);
7392 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7394 update_rq_clock(rq);
7395 activate_task(rq, p, 0);
7397 raw_spin_unlock_irqrestore(&rq->lock, flags);
7401 * Ensures that the idle task is using init_mm right before its cpu goes
7404 void idle_task_exit(void)
7406 struct mm_struct *mm = current->active_mm;
7408 BUG_ON(cpu_online(smp_processor_id()));
7411 switch_mm(mm, &init_mm, current);
7415 /* called under rq->lock with disabled interrupts */
7416 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7418 struct rq *rq = cpu_rq(dead_cpu);
7420 /* Must be exiting, otherwise would be on tasklist. */
7421 BUG_ON(!p->exit_state);
7423 /* Cannot have done final schedule yet: would have vanished. */
7424 BUG_ON(p->state == TASK_DEAD);
7429 * Drop lock around migration; if someone else moves it,
7430 * that's OK. No task can be added to this CPU, so iteration is
7433 raw_spin_unlock_irq(&rq->lock);
7434 move_task_off_dead_cpu(dead_cpu, p);
7435 raw_spin_lock_irq(&rq->lock);
7440 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7441 static void migrate_dead_tasks(unsigned int dead_cpu)
7443 struct rq *rq = cpu_rq(dead_cpu);
7444 struct task_struct *next;
7447 if (!rq->nr_running)
7449 update_rq_clock(rq);
7450 next = pick_next_task(rq);
7453 next->sched_class->put_prev_task(rq, next);
7454 migrate_dead(dead_cpu, next);
7460 * remove the tasks which were accounted by rq from calc_load_tasks.
7462 static void calc_global_load_remove(struct rq *rq)
7464 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7465 rq->calc_load_active = 0;
7467 #endif /* CONFIG_HOTPLUG_CPU */
7469 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7471 static struct ctl_table sd_ctl_dir[] = {
7473 .procname = "sched_domain",
7479 static struct ctl_table sd_ctl_root[] = {
7481 .procname = "kernel",
7483 .child = sd_ctl_dir,
7488 static struct ctl_table *sd_alloc_ctl_entry(int n)
7490 struct ctl_table *entry =
7491 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7496 static void sd_free_ctl_entry(struct ctl_table **tablep)
7498 struct ctl_table *entry;
7501 * In the intermediate directories, both the child directory and
7502 * procname are dynamically allocated and could fail but the mode
7503 * will always be set. In the lowest directory the names are
7504 * static strings and all have proc handlers.
7506 for (entry = *tablep; entry->mode; entry++) {
7508 sd_free_ctl_entry(&entry->child);
7509 if (entry->proc_handler == NULL)
7510 kfree(entry->procname);
7518 set_table_entry(struct ctl_table *entry,
7519 const char *procname, void *data, int maxlen,
7520 mode_t mode, proc_handler *proc_handler)
7522 entry->procname = procname;
7524 entry->maxlen = maxlen;
7526 entry->proc_handler = proc_handler;
7529 static struct ctl_table *
7530 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7532 struct ctl_table *table = sd_alloc_ctl_entry(13);
7537 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7538 sizeof(long), 0644, proc_doulongvec_minmax);
7539 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7540 sizeof(long), 0644, proc_doulongvec_minmax);
7541 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7542 sizeof(int), 0644, proc_dointvec_minmax);
7543 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7544 sizeof(int), 0644, proc_dointvec_minmax);
7545 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7546 sizeof(int), 0644, proc_dointvec_minmax);
7547 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7548 sizeof(int), 0644, proc_dointvec_minmax);
7549 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7550 sizeof(int), 0644, proc_dointvec_minmax);
7551 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7552 sizeof(int), 0644, proc_dointvec_minmax);
7553 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7554 sizeof(int), 0644, proc_dointvec_minmax);
7555 set_table_entry(&table[9], "cache_nice_tries",
7556 &sd->cache_nice_tries,
7557 sizeof(int), 0644, proc_dointvec_minmax);
7558 set_table_entry(&table[10], "flags", &sd->flags,
7559 sizeof(int), 0644, proc_dointvec_minmax);
7560 set_table_entry(&table[11], "name", sd->name,
7561 CORENAME_MAX_SIZE, 0444, proc_dostring);
7562 /* &table[12] is terminator */
7567 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7569 struct ctl_table *entry, *table;
7570 struct sched_domain *sd;
7571 int domain_num = 0, i;
7574 for_each_domain(cpu, sd)
7576 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7581 for_each_domain(cpu, sd) {
7582 snprintf(buf, 32, "domain%d", i);
7583 entry->procname = kstrdup(buf, GFP_KERNEL);
7585 entry->child = sd_alloc_ctl_domain_table(sd);
7592 static struct ctl_table_header *sd_sysctl_header;
7593 static void register_sched_domain_sysctl(void)
7595 int i, cpu_num = num_possible_cpus();
7596 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7599 WARN_ON(sd_ctl_dir[0].child);
7600 sd_ctl_dir[0].child = entry;
7605 for_each_possible_cpu(i) {
7606 snprintf(buf, 32, "cpu%d", i);
7607 entry->procname = kstrdup(buf, GFP_KERNEL);
7609 entry->child = sd_alloc_ctl_cpu_table(i);
7613 WARN_ON(sd_sysctl_header);
7614 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7617 /* may be called multiple times per register */
7618 static void unregister_sched_domain_sysctl(void)
7620 if (sd_sysctl_header)
7621 unregister_sysctl_table(sd_sysctl_header);
7622 sd_sysctl_header = NULL;
7623 if (sd_ctl_dir[0].child)
7624 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7627 static void register_sched_domain_sysctl(void)
7630 static void unregister_sched_domain_sysctl(void)
7635 static void set_rq_online(struct rq *rq)
7638 const struct sched_class *class;
7640 cpumask_set_cpu(rq->cpu, rq->rd->online);
7643 for_each_class(class) {
7644 if (class->rq_online)
7645 class->rq_online(rq);
7650 static void set_rq_offline(struct rq *rq)
7653 const struct sched_class *class;
7655 for_each_class(class) {
7656 if (class->rq_offline)
7657 class->rq_offline(rq);
7660 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7666 * migration_call - callback that gets triggered when a CPU is added.
7667 * Here we can start up the necessary migration thread for the new CPU.
7669 static int __cpuinit
7670 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7672 struct task_struct *p;
7673 int cpu = (long)hcpu;
7674 unsigned long flags;
7679 case CPU_UP_PREPARE:
7680 case CPU_UP_PREPARE_FROZEN:
7681 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7684 kthread_bind(p, cpu);
7685 /* Must be high prio: stop_machine expects to yield to it. */
7686 rq = task_rq_lock(p, &flags);
7687 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7688 task_rq_unlock(rq, &flags);
7690 cpu_rq(cpu)->migration_thread = p;
7691 rq->calc_load_update = calc_load_update;
7695 case CPU_ONLINE_FROZEN:
7696 /* Strictly unnecessary, as first user will wake it. */
7697 wake_up_process(cpu_rq(cpu)->migration_thread);
7699 /* Update our root-domain */
7701 raw_spin_lock_irqsave(&rq->lock, flags);
7703 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7707 raw_spin_unlock_irqrestore(&rq->lock, flags);
7710 #ifdef CONFIG_HOTPLUG_CPU
7711 case CPU_UP_CANCELED:
7712 case CPU_UP_CANCELED_FROZEN:
7713 if (!cpu_rq(cpu)->migration_thread)
7715 /* Unbind it from offline cpu so it can run. Fall thru. */
7716 kthread_bind(cpu_rq(cpu)->migration_thread,
7717 cpumask_any(cpu_online_mask));
7718 kthread_stop(cpu_rq(cpu)->migration_thread);
7719 put_task_struct(cpu_rq(cpu)->migration_thread);
7720 cpu_rq(cpu)->migration_thread = NULL;
7724 case CPU_DEAD_FROZEN:
7725 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7726 migrate_live_tasks(cpu);
7728 kthread_stop(rq->migration_thread);
7729 put_task_struct(rq->migration_thread);
7730 rq->migration_thread = NULL;
7731 /* Idle task back to normal (off runqueue, low prio) */
7732 raw_spin_lock_irq(&rq->lock);
7733 update_rq_clock(rq);
7734 deactivate_task(rq, rq->idle, 0);
7735 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7736 rq->idle->sched_class = &idle_sched_class;
7737 migrate_dead_tasks(cpu);
7738 raw_spin_unlock_irq(&rq->lock);
7740 migrate_nr_uninterruptible(rq);
7741 BUG_ON(rq->nr_running != 0);
7742 calc_global_load_remove(rq);
7744 * No need to migrate the tasks: it was best-effort if
7745 * they didn't take sched_hotcpu_mutex. Just wake up
7748 raw_spin_lock_irq(&rq->lock);
7749 while (!list_empty(&rq->migration_queue)) {
7750 struct migration_req *req;
7752 req = list_entry(rq->migration_queue.next,
7753 struct migration_req, list);
7754 list_del_init(&req->list);
7755 raw_spin_unlock_irq(&rq->lock);
7756 complete(&req->done);
7757 raw_spin_lock_irq(&rq->lock);
7759 raw_spin_unlock_irq(&rq->lock);
7763 case CPU_DYING_FROZEN:
7764 /* Update our root-domain */
7766 raw_spin_lock_irqsave(&rq->lock, flags);
7768 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7771 raw_spin_unlock_irqrestore(&rq->lock, flags);
7779 * Register at high priority so that task migration (migrate_all_tasks)
7780 * happens before everything else. This has to be lower priority than
7781 * the notifier in the perf_event subsystem, though.
7783 static struct notifier_block __cpuinitdata migration_notifier = {
7784 .notifier_call = migration_call,
7788 static int __init migration_init(void)
7790 void *cpu = (void *)(long)smp_processor_id();
7793 /* Start one for the boot CPU: */
7794 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7795 BUG_ON(err == NOTIFY_BAD);
7796 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7797 register_cpu_notifier(&migration_notifier);
7801 early_initcall(migration_init);
7806 #ifdef CONFIG_SCHED_DEBUG
7808 static __read_mostly int sched_domain_debug_enabled;
7810 static int __init sched_domain_debug_setup(char *str)
7812 sched_domain_debug_enabled = 1;
7816 early_param("sched_debug", sched_domain_debug_setup);
7818 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7819 struct cpumask *groupmask)
7821 struct sched_group *group = sd->groups;
7824 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7825 cpumask_clear(groupmask);
7827 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7829 if (!(sd->flags & SD_LOAD_BALANCE)) {
7830 pr_cont("does not load-balance\n");
7832 pr_err("ERROR: !SD_LOAD_BALANCE domain has parent\n");
7836 pr_cont("span %s level %s\n", str, sd->name);
7838 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7839 pr_err("ERROR: domain->span does not contain CPU%d\n", cpu);
7841 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7842 pr_err("ERROR: domain->groups does not contain CPU%d\n", cpu);
7845 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7849 pr_err("ERROR: group is NULL\n");
7853 if (!group->cpu_power) {
7855 pr_err("ERROR: domain->cpu_power not set\n");
7859 if (!cpumask_weight(sched_group_cpus(group))) {
7861 pr_err("ERROR: empty group\n");
7865 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7867 pr_err("ERROR: repeated CPUs\n");
7871 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7873 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7875 pr_cont(" %s", str);
7876 if (group->cpu_power != SCHED_LOAD_SCALE) {
7877 pr_cont(" (cpu_power = %d)", group->cpu_power);
7880 group = group->next;
7881 } while (group != sd->groups);
7884 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7885 pr_err("ERROR: groups don't span domain->span\n");
7888 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7889 pr_err("ERROR: parent span is not a superset of domain->span\n");
7893 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7895 cpumask_var_t groupmask;
7898 if (!sched_domain_debug_enabled)
7902 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7906 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7908 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7909 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7914 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7921 free_cpumask_var(groupmask);
7923 #else /* !CONFIG_SCHED_DEBUG */
7924 # define sched_domain_debug(sd, cpu) do { } while (0)
7925 #endif /* CONFIG_SCHED_DEBUG */
7927 static int sd_degenerate(struct sched_domain *sd)
7929 if (cpumask_weight(sched_domain_span(sd)) == 1)
7932 /* Following flags need at least 2 groups */
7933 if (sd->flags & (SD_LOAD_BALANCE |
7934 SD_BALANCE_NEWIDLE |
7938 SD_SHARE_PKG_RESOURCES)) {
7939 if (sd->groups != sd->groups->next)
7943 /* Following flags don't use groups */
7944 if (sd->flags & (SD_WAKE_AFFINE))
7951 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7953 unsigned long cflags = sd->flags, pflags = parent->flags;
7955 if (sd_degenerate(parent))
7958 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7961 /* Flags needing groups don't count if only 1 group in parent */
7962 if (parent->groups == parent->groups->next) {
7963 pflags &= ~(SD_LOAD_BALANCE |
7964 SD_BALANCE_NEWIDLE |
7968 SD_SHARE_PKG_RESOURCES);
7969 if (nr_node_ids == 1)
7970 pflags &= ~SD_SERIALIZE;
7972 if (~cflags & pflags)
7978 static void free_rootdomain(struct root_domain *rd)
7980 synchronize_sched();
7982 cpupri_cleanup(&rd->cpupri);
7984 free_cpumask_var(rd->rto_mask);
7985 free_cpumask_var(rd->online);
7986 free_cpumask_var(rd->span);
7990 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7992 struct root_domain *old_rd = NULL;
7993 unsigned long flags;
7995 raw_spin_lock_irqsave(&rq->lock, flags);
8000 if (cpumask_test_cpu(rq->cpu, old_rd->online))
8003 cpumask_clear_cpu(rq->cpu, old_rd->span);
8006 * If we dont want to free the old_rt yet then
8007 * set old_rd to NULL to skip the freeing later
8010 if (!atomic_dec_and_test(&old_rd->refcount))
8014 atomic_inc(&rd->refcount);
8017 cpumask_set_cpu(rq->cpu, rd->span);
8018 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
8021 raw_spin_unlock_irqrestore(&rq->lock, flags);
8024 free_rootdomain(old_rd);
8027 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8029 gfp_t gfp = GFP_KERNEL;
8031 memset(rd, 0, sizeof(*rd));
8036 if (!alloc_cpumask_var(&rd->span, gfp))
8038 if (!alloc_cpumask_var(&rd->online, gfp))
8040 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8043 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8048 free_cpumask_var(rd->rto_mask);
8050 free_cpumask_var(rd->online);
8052 free_cpumask_var(rd->span);
8057 static void init_defrootdomain(void)
8059 init_rootdomain(&def_root_domain, true);
8061 atomic_set(&def_root_domain.refcount, 1);
8064 static struct root_domain *alloc_rootdomain(void)
8066 struct root_domain *rd;
8068 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8072 if (init_rootdomain(rd, false) != 0) {
8081 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8082 * hold the hotplug lock.
8085 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8087 struct rq *rq = cpu_rq(cpu);
8088 struct sched_domain *tmp;
8090 /* Remove the sched domains which do not contribute to scheduling. */
8091 for (tmp = sd; tmp; ) {
8092 struct sched_domain *parent = tmp->parent;
8096 if (sd_parent_degenerate(tmp, parent)) {
8097 tmp->parent = parent->parent;
8099 parent->parent->child = tmp;
8104 if (sd && sd_degenerate(sd)) {
8110 sched_domain_debug(sd, cpu);
8112 rq_attach_root(rq, rd);
8113 rcu_assign_pointer(rq->sd, sd);
8116 /* cpus with isolated domains */
8117 static cpumask_var_t cpu_isolated_map;
8119 /* Setup the mask of cpus configured for isolated domains */
8120 static int __init isolated_cpu_setup(char *str)
8122 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8123 cpulist_parse(str, cpu_isolated_map);
8127 __setup("isolcpus=", isolated_cpu_setup);
8130 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8131 * to a function which identifies what group(along with sched group) a CPU
8132 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8133 * (due to the fact that we keep track of groups covered with a struct cpumask).
8135 * init_sched_build_groups will build a circular linked list of the groups
8136 * covered by the given span, and will set each group's ->cpumask correctly,
8137 * and ->cpu_power to 0.
8140 init_sched_build_groups(const struct cpumask *span,
8141 const struct cpumask *cpu_map,
8142 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8143 struct sched_group **sg,
8144 struct cpumask *tmpmask),
8145 struct cpumask *covered, struct cpumask *tmpmask)
8147 struct sched_group *first = NULL, *last = NULL;
8150 cpumask_clear(covered);
8152 for_each_cpu(i, span) {
8153 struct sched_group *sg;
8154 int group = group_fn(i, cpu_map, &sg, tmpmask);
8157 if (cpumask_test_cpu(i, covered))
8160 cpumask_clear(sched_group_cpus(sg));
8163 for_each_cpu(j, span) {
8164 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8167 cpumask_set_cpu(j, covered);
8168 cpumask_set_cpu(j, sched_group_cpus(sg));
8179 #define SD_NODES_PER_DOMAIN 16
8184 * find_next_best_node - find the next node to include in a sched_domain
8185 * @node: node whose sched_domain we're building
8186 * @used_nodes: nodes already in the sched_domain
8188 * Find the next node to include in a given scheduling domain. Simply
8189 * finds the closest node not already in the @used_nodes map.
8191 * Should use nodemask_t.
8193 static int find_next_best_node(int node, nodemask_t *used_nodes)
8195 int i, n, val, min_val, best_node = 0;
8199 for (i = 0; i < nr_node_ids; i++) {
8200 /* Start at @node */
8201 n = (node + i) % nr_node_ids;
8203 if (!nr_cpus_node(n))
8206 /* Skip already used nodes */
8207 if (node_isset(n, *used_nodes))
8210 /* Simple min distance search */
8211 val = node_distance(node, n);
8213 if (val < min_val) {
8219 node_set(best_node, *used_nodes);
8224 * sched_domain_node_span - get a cpumask for a node's sched_domain
8225 * @node: node whose cpumask we're constructing
8226 * @span: resulting cpumask
8228 * Given a node, construct a good cpumask for its sched_domain to span. It
8229 * should be one that prevents unnecessary balancing, but also spreads tasks
8232 static void sched_domain_node_span(int node, struct cpumask *span)
8234 nodemask_t used_nodes;
8237 cpumask_clear(span);
8238 nodes_clear(used_nodes);
8240 cpumask_or(span, span, cpumask_of_node(node));
8241 node_set(node, used_nodes);
8243 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8244 int next_node = find_next_best_node(node, &used_nodes);
8246 cpumask_or(span, span, cpumask_of_node(next_node));
8249 #endif /* CONFIG_NUMA */
8251 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8254 * The cpus mask in sched_group and sched_domain hangs off the end.
8256 * ( See the the comments in include/linux/sched.h:struct sched_group
8257 * and struct sched_domain. )
8259 struct static_sched_group {
8260 struct sched_group sg;
8261 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8264 struct static_sched_domain {
8265 struct sched_domain sd;
8266 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8272 cpumask_var_t domainspan;
8273 cpumask_var_t covered;
8274 cpumask_var_t notcovered;
8276 cpumask_var_t nodemask;
8277 cpumask_var_t this_sibling_map;
8278 cpumask_var_t this_core_map;
8279 cpumask_var_t send_covered;
8280 cpumask_var_t tmpmask;
8281 struct sched_group **sched_group_nodes;
8282 struct root_domain *rd;
8286 sa_sched_groups = 0,
8291 sa_this_sibling_map,
8293 sa_sched_group_nodes,
8303 * SMT sched-domains:
8305 #ifdef CONFIG_SCHED_SMT
8306 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8307 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
8310 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8311 struct sched_group **sg, struct cpumask *unused)
8314 *sg = &per_cpu(sched_groups, cpu).sg;
8317 #endif /* CONFIG_SCHED_SMT */
8320 * multi-core sched-domains:
8322 #ifdef CONFIG_SCHED_MC
8323 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8324 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8325 #endif /* CONFIG_SCHED_MC */
8327 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8329 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8330 struct sched_group **sg, struct cpumask *mask)
8334 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8335 group = cpumask_first(mask);
8337 *sg = &per_cpu(sched_group_core, group).sg;
8340 #elif defined(CONFIG_SCHED_MC)
8342 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8343 struct sched_group **sg, struct cpumask *unused)
8346 *sg = &per_cpu(sched_group_core, cpu).sg;
8351 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8352 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8355 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8356 struct sched_group **sg, struct cpumask *mask)
8359 #ifdef CONFIG_SCHED_MC
8360 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8361 group = cpumask_first(mask);
8362 #elif defined(CONFIG_SCHED_SMT)
8363 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8364 group = cpumask_first(mask);
8369 *sg = &per_cpu(sched_group_phys, group).sg;
8375 * The init_sched_build_groups can't handle what we want to do with node
8376 * groups, so roll our own. Now each node has its own list of groups which
8377 * gets dynamically allocated.
8379 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8380 static struct sched_group ***sched_group_nodes_bycpu;
8382 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8383 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8385 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8386 struct sched_group **sg,
8387 struct cpumask *nodemask)
8391 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8392 group = cpumask_first(nodemask);
8395 *sg = &per_cpu(sched_group_allnodes, group).sg;
8399 static void init_numa_sched_groups_power(struct sched_group *group_head)
8401 struct sched_group *sg = group_head;
8407 for_each_cpu(j, sched_group_cpus(sg)) {
8408 struct sched_domain *sd;
8410 sd = &per_cpu(phys_domains, j).sd;
8411 if (j != group_first_cpu(sd->groups)) {
8413 * Only add "power" once for each
8419 sg->cpu_power += sd->groups->cpu_power;
8422 } while (sg != group_head);
8425 static int build_numa_sched_groups(struct s_data *d,
8426 const struct cpumask *cpu_map, int num)
8428 struct sched_domain *sd;
8429 struct sched_group *sg, *prev;
8432 cpumask_clear(d->covered);
8433 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8434 if (cpumask_empty(d->nodemask)) {
8435 d->sched_group_nodes[num] = NULL;
8439 sched_domain_node_span(num, d->domainspan);
8440 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8442 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8445 pr_warning("Can not alloc domain group for node %d\n", num);
8448 d->sched_group_nodes[num] = sg;
8450 for_each_cpu(j, d->nodemask) {
8451 sd = &per_cpu(node_domains, j).sd;
8456 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8458 cpumask_or(d->covered, d->covered, d->nodemask);
8461 for (j = 0; j < nr_node_ids; j++) {
8462 n = (num + j) % nr_node_ids;
8463 cpumask_complement(d->notcovered, d->covered);
8464 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8465 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8466 if (cpumask_empty(d->tmpmask))
8468 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8469 if (cpumask_empty(d->tmpmask))
8471 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8474 pr_warning("Can not alloc domain group for node %d\n",
8479 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8480 sg->next = prev->next;
8481 cpumask_or(d->covered, d->covered, d->tmpmask);
8488 #endif /* CONFIG_NUMA */
8491 /* Free memory allocated for various sched_group structures */
8492 static void free_sched_groups(const struct cpumask *cpu_map,
8493 struct cpumask *nodemask)
8497 for_each_cpu(cpu, cpu_map) {
8498 struct sched_group **sched_group_nodes
8499 = sched_group_nodes_bycpu[cpu];
8501 if (!sched_group_nodes)
8504 for (i = 0; i < nr_node_ids; i++) {
8505 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8507 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8508 if (cpumask_empty(nodemask))
8518 if (oldsg != sched_group_nodes[i])
8521 kfree(sched_group_nodes);
8522 sched_group_nodes_bycpu[cpu] = NULL;
8525 #else /* !CONFIG_NUMA */
8526 static void free_sched_groups(const struct cpumask *cpu_map,
8527 struct cpumask *nodemask)
8530 #endif /* CONFIG_NUMA */
8533 * Initialize sched groups cpu_power.
8535 * cpu_power indicates the capacity of sched group, which is used while
8536 * distributing the load between different sched groups in a sched domain.
8537 * Typically cpu_power for all the groups in a sched domain will be same unless
8538 * there are asymmetries in the topology. If there are asymmetries, group
8539 * having more cpu_power will pickup more load compared to the group having
8542 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8544 struct sched_domain *child;
8545 struct sched_group *group;
8549 WARN_ON(!sd || !sd->groups);
8551 if (cpu != group_first_cpu(sd->groups))
8556 sd->groups->cpu_power = 0;
8559 power = SCHED_LOAD_SCALE;
8560 weight = cpumask_weight(sched_domain_span(sd));
8562 * SMT siblings share the power of a single core.
8563 * Usually multiple threads get a better yield out of
8564 * that one core than a single thread would have,
8565 * reflect that in sd->smt_gain.
8567 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8568 power *= sd->smt_gain;
8570 power >>= SCHED_LOAD_SHIFT;
8572 sd->groups->cpu_power += power;
8577 * Add cpu_power of each child group to this groups cpu_power.
8579 group = child->groups;
8581 sd->groups->cpu_power += group->cpu_power;
8582 group = group->next;
8583 } while (group != child->groups);
8587 * Initializers for schedule domains
8588 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8591 #ifdef CONFIG_SCHED_DEBUG
8592 # define SD_INIT_NAME(sd, type) sd->name = #type
8594 # define SD_INIT_NAME(sd, type) do { } while (0)
8597 #define SD_INIT(sd, type) sd_init_##type(sd)
8599 #define SD_INIT_FUNC(type) \
8600 static noinline void sd_init_##type(struct sched_domain *sd) \
8602 memset(sd, 0, sizeof(*sd)); \
8603 *sd = SD_##type##_INIT; \
8604 sd->level = SD_LV_##type; \
8605 SD_INIT_NAME(sd, type); \
8610 SD_INIT_FUNC(ALLNODES)
8613 #ifdef CONFIG_SCHED_SMT
8614 SD_INIT_FUNC(SIBLING)
8616 #ifdef CONFIG_SCHED_MC
8620 static int default_relax_domain_level = -1;
8622 static int __init setup_relax_domain_level(char *str)
8626 val = simple_strtoul(str, NULL, 0);
8627 if (val < SD_LV_MAX)
8628 default_relax_domain_level = val;
8632 __setup("relax_domain_level=", setup_relax_domain_level);
8634 static void set_domain_attribute(struct sched_domain *sd,
8635 struct sched_domain_attr *attr)
8639 if (!attr || attr->relax_domain_level < 0) {
8640 if (default_relax_domain_level < 0)
8643 request = default_relax_domain_level;
8645 request = attr->relax_domain_level;
8646 if (request < sd->level) {
8647 /* turn off idle balance on this domain */
8648 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8650 /* turn on idle balance on this domain */
8651 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8655 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8656 const struct cpumask *cpu_map)
8659 case sa_sched_groups:
8660 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8661 d->sched_group_nodes = NULL;
8663 free_rootdomain(d->rd); /* fall through */
8665 free_cpumask_var(d->tmpmask); /* fall through */
8666 case sa_send_covered:
8667 free_cpumask_var(d->send_covered); /* fall through */
8668 case sa_this_core_map:
8669 free_cpumask_var(d->this_core_map); /* fall through */
8670 case sa_this_sibling_map:
8671 free_cpumask_var(d->this_sibling_map); /* fall through */
8673 free_cpumask_var(d->nodemask); /* fall through */
8674 case sa_sched_group_nodes:
8676 kfree(d->sched_group_nodes); /* fall through */
8678 free_cpumask_var(d->notcovered); /* fall through */
8680 free_cpumask_var(d->covered); /* fall through */
8682 free_cpumask_var(d->domainspan); /* fall through */
8689 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8690 const struct cpumask *cpu_map)
8693 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8695 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8696 return sa_domainspan;
8697 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8699 /* Allocate the per-node list of sched groups */
8700 d->sched_group_nodes = kcalloc(nr_node_ids,
8701 sizeof(struct sched_group *), GFP_KERNEL);
8702 if (!d->sched_group_nodes) {
8703 pr_warning("Can not alloc sched group node list\n");
8704 return sa_notcovered;
8706 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8708 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8709 return sa_sched_group_nodes;
8710 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8712 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8713 return sa_this_sibling_map;
8714 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8715 return sa_this_core_map;
8716 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8717 return sa_send_covered;
8718 d->rd = alloc_rootdomain();
8720 pr_warning("Cannot alloc root domain\n");
8723 return sa_rootdomain;
8726 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8727 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8729 struct sched_domain *sd = NULL;
8731 struct sched_domain *parent;
8734 if (cpumask_weight(cpu_map) >
8735 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8736 sd = &per_cpu(allnodes_domains, i).sd;
8737 SD_INIT(sd, ALLNODES);
8738 set_domain_attribute(sd, attr);
8739 cpumask_copy(sched_domain_span(sd), cpu_map);
8740 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8745 sd = &per_cpu(node_domains, i).sd;
8747 set_domain_attribute(sd, attr);
8748 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8749 sd->parent = parent;
8752 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8757 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8758 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8759 struct sched_domain *parent, int i)
8761 struct sched_domain *sd;
8762 sd = &per_cpu(phys_domains, i).sd;
8764 set_domain_attribute(sd, attr);
8765 cpumask_copy(sched_domain_span(sd), d->nodemask);
8766 sd->parent = parent;
8769 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8773 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8774 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8775 struct sched_domain *parent, int i)
8777 struct sched_domain *sd = parent;
8778 #ifdef CONFIG_SCHED_MC
8779 sd = &per_cpu(core_domains, i).sd;
8781 set_domain_attribute(sd, attr);
8782 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8783 sd->parent = parent;
8785 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8790 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8791 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8792 struct sched_domain *parent, int i)
8794 struct sched_domain *sd = parent;
8795 #ifdef CONFIG_SCHED_SMT
8796 sd = &per_cpu(cpu_domains, i).sd;
8797 SD_INIT(sd, SIBLING);
8798 set_domain_attribute(sd, attr);
8799 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8800 sd->parent = parent;
8802 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8807 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8808 const struct cpumask *cpu_map, int cpu)
8811 #ifdef CONFIG_SCHED_SMT
8812 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8813 cpumask_and(d->this_sibling_map, cpu_map,
8814 topology_thread_cpumask(cpu));
8815 if (cpu == cpumask_first(d->this_sibling_map))
8816 init_sched_build_groups(d->this_sibling_map, cpu_map,
8818 d->send_covered, d->tmpmask);
8821 #ifdef CONFIG_SCHED_MC
8822 case SD_LV_MC: /* set up multi-core groups */
8823 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8824 if (cpu == cpumask_first(d->this_core_map))
8825 init_sched_build_groups(d->this_core_map, cpu_map,
8827 d->send_covered, d->tmpmask);
8830 case SD_LV_CPU: /* set up physical groups */
8831 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8832 if (!cpumask_empty(d->nodemask))
8833 init_sched_build_groups(d->nodemask, cpu_map,
8835 d->send_covered, d->tmpmask);
8838 case SD_LV_ALLNODES:
8839 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8840 d->send_covered, d->tmpmask);
8849 * Build sched domains for a given set of cpus and attach the sched domains
8850 * to the individual cpus
8852 static int __build_sched_domains(const struct cpumask *cpu_map,
8853 struct sched_domain_attr *attr)
8855 enum s_alloc alloc_state = sa_none;
8857 struct sched_domain *sd;
8863 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8864 if (alloc_state != sa_rootdomain)
8866 alloc_state = sa_sched_groups;
8869 * Set up domains for cpus specified by the cpu_map.
8871 for_each_cpu(i, cpu_map) {
8872 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8875 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8876 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8877 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8878 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8881 for_each_cpu(i, cpu_map) {
8882 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8883 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8886 /* Set up physical groups */
8887 for (i = 0; i < nr_node_ids; i++)
8888 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8891 /* Set up node groups */
8893 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8895 for (i = 0; i < nr_node_ids; i++)
8896 if (build_numa_sched_groups(&d, cpu_map, i))
8900 /* Calculate CPU power for physical packages and nodes */
8901 #ifdef CONFIG_SCHED_SMT
8902 for_each_cpu(i, cpu_map) {
8903 sd = &per_cpu(cpu_domains, i).sd;
8904 init_sched_groups_power(i, sd);
8907 #ifdef CONFIG_SCHED_MC
8908 for_each_cpu(i, cpu_map) {
8909 sd = &per_cpu(core_domains, i).sd;
8910 init_sched_groups_power(i, sd);
8914 for_each_cpu(i, cpu_map) {
8915 sd = &per_cpu(phys_domains, i).sd;
8916 init_sched_groups_power(i, sd);
8920 for (i = 0; i < nr_node_ids; i++)
8921 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8923 if (d.sd_allnodes) {
8924 struct sched_group *sg;
8926 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8928 init_numa_sched_groups_power(sg);
8932 /* Attach the domains */
8933 for_each_cpu(i, cpu_map) {
8934 #ifdef CONFIG_SCHED_SMT
8935 sd = &per_cpu(cpu_domains, i).sd;
8936 #elif defined(CONFIG_SCHED_MC)
8937 sd = &per_cpu(core_domains, i).sd;
8939 sd = &per_cpu(phys_domains, i).sd;
8941 cpu_attach_domain(sd, d.rd, i);
8944 d.sched_group_nodes = NULL; /* don't free this we still need it */
8945 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8949 __free_domain_allocs(&d, alloc_state, cpu_map);
8953 static int build_sched_domains(const struct cpumask *cpu_map)
8955 return __build_sched_domains(cpu_map, NULL);
8958 static cpumask_var_t *doms_cur; /* current sched domains */
8959 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8960 static struct sched_domain_attr *dattr_cur;
8961 /* attribues of custom domains in 'doms_cur' */
8964 * Special case: If a kmalloc of a doms_cur partition (array of
8965 * cpumask) fails, then fallback to a single sched domain,
8966 * as determined by the single cpumask fallback_doms.
8968 static cpumask_var_t fallback_doms;
8971 * arch_update_cpu_topology lets virtualized architectures update the
8972 * cpu core maps. It is supposed to return 1 if the topology changed
8973 * or 0 if it stayed the same.
8975 int __attribute__((weak)) arch_update_cpu_topology(void)
8980 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
8983 cpumask_var_t *doms;
8985 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
8988 for (i = 0; i < ndoms; i++) {
8989 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
8990 free_sched_domains(doms, i);
8997 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
9000 for (i = 0; i < ndoms; i++)
9001 free_cpumask_var(doms[i]);
9006 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9007 * For now this just excludes isolated cpus, but could be used to
9008 * exclude other special cases in the future.
9010 static int arch_init_sched_domains(const struct cpumask *cpu_map)
9014 arch_update_cpu_topology();
9016 doms_cur = alloc_sched_domains(ndoms_cur);
9018 doms_cur = &fallback_doms;
9019 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
9021 err = build_sched_domains(doms_cur[0]);
9022 register_sched_domain_sysctl();
9027 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9028 struct cpumask *tmpmask)
9030 free_sched_groups(cpu_map, tmpmask);
9034 * Detach sched domains from a group of cpus specified in cpu_map
9035 * These cpus will now be attached to the NULL domain
9037 static void detach_destroy_domains(const struct cpumask *cpu_map)
9039 /* Save because hotplug lock held. */
9040 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9043 for_each_cpu(i, cpu_map)
9044 cpu_attach_domain(NULL, &def_root_domain, i);
9045 synchronize_sched();
9046 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9049 /* handle null as "default" */
9050 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9051 struct sched_domain_attr *new, int idx_new)
9053 struct sched_domain_attr tmp;
9060 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9061 new ? (new + idx_new) : &tmp,
9062 sizeof(struct sched_domain_attr));
9066 * Partition sched domains as specified by the 'ndoms_new'
9067 * cpumasks in the array doms_new[] of cpumasks. This compares
9068 * doms_new[] to the current sched domain partitioning, doms_cur[].
9069 * It destroys each deleted domain and builds each new domain.
9071 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
9072 * The masks don't intersect (don't overlap.) We should setup one
9073 * sched domain for each mask. CPUs not in any of the cpumasks will
9074 * not be load balanced. If the same cpumask appears both in the
9075 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9078 * The passed in 'doms_new' should be allocated using
9079 * alloc_sched_domains. This routine takes ownership of it and will
9080 * free_sched_domains it when done with it. If the caller failed the
9081 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9082 * and partition_sched_domains() will fallback to the single partition
9083 * 'fallback_doms', it also forces the domains to be rebuilt.
9085 * If doms_new == NULL it will be replaced with cpu_online_mask.
9086 * ndoms_new == 0 is a special case for destroying existing domains,
9087 * and it will not create the default domain.
9089 * Call with hotplug lock held
9091 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
9092 struct sched_domain_attr *dattr_new)
9097 mutex_lock(&sched_domains_mutex);
9099 /* always unregister in case we don't destroy any domains */
9100 unregister_sched_domain_sysctl();
9102 /* Let architecture update cpu core mappings. */
9103 new_topology = arch_update_cpu_topology();
9105 n = doms_new ? ndoms_new : 0;
9107 /* Destroy deleted domains */
9108 for (i = 0; i < ndoms_cur; i++) {
9109 for (j = 0; j < n && !new_topology; j++) {
9110 if (cpumask_equal(doms_cur[i], doms_new[j])
9111 && dattrs_equal(dattr_cur, i, dattr_new, j))
9114 /* no match - a current sched domain not in new doms_new[] */
9115 detach_destroy_domains(doms_cur[i]);
9120 if (doms_new == NULL) {
9122 doms_new = &fallback_doms;
9123 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
9124 WARN_ON_ONCE(dattr_new);
9127 /* Build new domains */
9128 for (i = 0; i < ndoms_new; i++) {
9129 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9130 if (cpumask_equal(doms_new[i], doms_cur[j])
9131 && dattrs_equal(dattr_new, i, dattr_cur, j))
9134 /* no match - add a new doms_new */
9135 __build_sched_domains(doms_new[i],
9136 dattr_new ? dattr_new + i : NULL);
9141 /* Remember the new sched domains */
9142 if (doms_cur != &fallback_doms)
9143 free_sched_domains(doms_cur, ndoms_cur);
9144 kfree(dattr_cur); /* kfree(NULL) is safe */
9145 doms_cur = doms_new;
9146 dattr_cur = dattr_new;
9147 ndoms_cur = ndoms_new;
9149 register_sched_domain_sysctl();
9151 mutex_unlock(&sched_domains_mutex);
9154 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9155 static void arch_reinit_sched_domains(void)
9159 /* Destroy domains first to force the rebuild */
9160 partition_sched_domains(0, NULL, NULL);
9162 rebuild_sched_domains();
9166 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9168 unsigned int level = 0;
9170 if (sscanf(buf, "%u", &level) != 1)
9174 * level is always be positive so don't check for
9175 * level < POWERSAVINGS_BALANCE_NONE which is 0
9176 * What happens on 0 or 1 byte write,
9177 * need to check for count as well?
9180 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9184 sched_smt_power_savings = level;
9186 sched_mc_power_savings = level;
9188 arch_reinit_sched_domains();
9193 #ifdef CONFIG_SCHED_MC
9194 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9197 return sprintf(page, "%u\n", sched_mc_power_savings);
9199 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9200 const char *buf, size_t count)
9202 return sched_power_savings_store(buf, count, 0);
9204 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9205 sched_mc_power_savings_show,
9206 sched_mc_power_savings_store);
9209 #ifdef CONFIG_SCHED_SMT
9210 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9213 return sprintf(page, "%u\n", sched_smt_power_savings);
9215 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9216 const char *buf, size_t count)
9218 return sched_power_savings_store(buf, count, 1);
9220 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9221 sched_smt_power_savings_show,
9222 sched_smt_power_savings_store);
9225 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9229 #ifdef CONFIG_SCHED_SMT
9231 err = sysfs_create_file(&cls->kset.kobj,
9232 &attr_sched_smt_power_savings.attr);
9234 #ifdef CONFIG_SCHED_MC
9235 if (!err && mc_capable())
9236 err = sysfs_create_file(&cls->kset.kobj,
9237 &attr_sched_mc_power_savings.attr);
9241 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9243 #ifndef CONFIG_CPUSETS
9245 * Add online and remove offline CPUs from the scheduler domains.
9246 * When cpusets are enabled they take over this function.
9248 static int update_sched_domains(struct notifier_block *nfb,
9249 unsigned long action, void *hcpu)
9253 case CPU_ONLINE_FROZEN:
9254 case CPU_DOWN_PREPARE:
9255 case CPU_DOWN_PREPARE_FROZEN:
9256 case CPU_DOWN_FAILED:
9257 case CPU_DOWN_FAILED_FROZEN:
9258 partition_sched_domains(1, NULL, NULL);
9267 static int update_runtime(struct notifier_block *nfb,
9268 unsigned long action, void *hcpu)
9270 int cpu = (int)(long)hcpu;
9273 case CPU_DOWN_PREPARE:
9274 case CPU_DOWN_PREPARE_FROZEN:
9275 disable_runtime(cpu_rq(cpu));
9278 case CPU_DOWN_FAILED:
9279 case CPU_DOWN_FAILED_FROZEN:
9281 case CPU_ONLINE_FROZEN:
9282 enable_runtime(cpu_rq(cpu));
9290 void __init sched_init_smp(void)
9292 cpumask_var_t non_isolated_cpus;
9294 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9295 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9297 #if defined(CONFIG_NUMA)
9298 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9300 BUG_ON(sched_group_nodes_bycpu == NULL);
9303 mutex_lock(&sched_domains_mutex);
9304 arch_init_sched_domains(cpu_active_mask);
9305 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9306 if (cpumask_empty(non_isolated_cpus))
9307 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9308 mutex_unlock(&sched_domains_mutex);
9311 #ifndef CONFIG_CPUSETS
9312 /* XXX: Theoretical race here - CPU may be hotplugged now */
9313 hotcpu_notifier(update_sched_domains, 0);
9316 /* RT runtime code needs to handle some hotplug events */
9317 hotcpu_notifier(update_runtime, 0);
9321 /* Move init over to a non-isolated CPU */
9322 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9324 sched_init_granularity();
9325 free_cpumask_var(non_isolated_cpus);
9327 init_sched_rt_class();
9330 void __init sched_init_smp(void)
9332 sched_init_granularity();
9334 #endif /* CONFIG_SMP */
9336 const_debug unsigned int sysctl_timer_migration = 1;
9338 int in_sched_functions(unsigned long addr)
9340 return in_lock_functions(addr) ||
9341 (addr >= (unsigned long)__sched_text_start
9342 && addr < (unsigned long)__sched_text_end);
9345 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9347 cfs_rq->tasks_timeline = RB_ROOT;
9348 INIT_LIST_HEAD(&cfs_rq->tasks);
9349 #ifdef CONFIG_FAIR_GROUP_SCHED
9352 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9355 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9357 struct rt_prio_array *array;
9360 array = &rt_rq->active;
9361 for (i = 0; i < MAX_RT_PRIO; i++) {
9362 INIT_LIST_HEAD(array->queue + i);
9363 __clear_bit(i, array->bitmap);
9365 /* delimiter for bitsearch: */
9366 __set_bit(MAX_RT_PRIO, array->bitmap);
9368 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9369 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9371 rt_rq->highest_prio.next = MAX_RT_PRIO;
9375 rt_rq->rt_nr_migratory = 0;
9376 rt_rq->overloaded = 0;
9377 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
9381 rt_rq->rt_throttled = 0;
9382 rt_rq->rt_runtime = 0;
9383 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
9385 #ifdef CONFIG_RT_GROUP_SCHED
9386 rt_rq->rt_nr_boosted = 0;
9391 #ifdef CONFIG_FAIR_GROUP_SCHED
9392 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9393 struct sched_entity *se, int cpu, int add,
9394 struct sched_entity *parent)
9396 struct rq *rq = cpu_rq(cpu);
9397 tg->cfs_rq[cpu] = cfs_rq;
9398 init_cfs_rq(cfs_rq, rq);
9401 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9404 /* se could be NULL for init_task_group */
9409 se->cfs_rq = &rq->cfs;
9411 se->cfs_rq = parent->my_q;
9414 se->load.weight = tg->shares;
9415 se->load.inv_weight = 0;
9416 se->parent = parent;
9420 #ifdef CONFIG_RT_GROUP_SCHED
9421 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9422 struct sched_rt_entity *rt_se, int cpu, int add,
9423 struct sched_rt_entity *parent)
9425 struct rq *rq = cpu_rq(cpu);
9427 tg->rt_rq[cpu] = rt_rq;
9428 init_rt_rq(rt_rq, rq);
9430 rt_rq->rt_se = rt_se;
9431 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9433 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9435 tg->rt_se[cpu] = rt_se;
9440 rt_se->rt_rq = &rq->rt;
9442 rt_se->rt_rq = parent->my_q;
9444 rt_se->my_q = rt_rq;
9445 rt_se->parent = parent;
9446 INIT_LIST_HEAD(&rt_se->run_list);
9450 void __init sched_init(void)
9453 unsigned long alloc_size = 0, ptr;
9455 #ifdef CONFIG_FAIR_GROUP_SCHED
9456 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9458 #ifdef CONFIG_RT_GROUP_SCHED
9459 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9461 #ifdef CONFIG_USER_SCHED
9464 #ifdef CONFIG_CPUMASK_OFFSTACK
9465 alloc_size += num_possible_cpus() * cpumask_size();
9468 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9470 #ifdef CONFIG_FAIR_GROUP_SCHED
9471 init_task_group.se = (struct sched_entity **)ptr;
9472 ptr += nr_cpu_ids * sizeof(void **);
9474 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9475 ptr += nr_cpu_ids * sizeof(void **);
9477 #ifdef CONFIG_USER_SCHED
9478 root_task_group.se = (struct sched_entity **)ptr;
9479 ptr += nr_cpu_ids * sizeof(void **);
9481 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9482 ptr += nr_cpu_ids * sizeof(void **);
9483 #endif /* CONFIG_USER_SCHED */
9484 #endif /* CONFIG_FAIR_GROUP_SCHED */
9485 #ifdef CONFIG_RT_GROUP_SCHED
9486 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9487 ptr += nr_cpu_ids * sizeof(void **);
9489 init_task_group.rt_rq = (struct rt_rq **)ptr;
9490 ptr += nr_cpu_ids * sizeof(void **);
9492 #ifdef CONFIG_USER_SCHED
9493 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9494 ptr += nr_cpu_ids * sizeof(void **);
9496 root_task_group.rt_rq = (struct rt_rq **)ptr;
9497 ptr += nr_cpu_ids * sizeof(void **);
9498 #endif /* CONFIG_USER_SCHED */
9499 #endif /* CONFIG_RT_GROUP_SCHED */
9500 #ifdef CONFIG_CPUMASK_OFFSTACK
9501 for_each_possible_cpu(i) {
9502 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9503 ptr += cpumask_size();
9505 #endif /* CONFIG_CPUMASK_OFFSTACK */
9509 init_defrootdomain();
9512 init_rt_bandwidth(&def_rt_bandwidth,
9513 global_rt_period(), global_rt_runtime());
9515 #ifdef CONFIG_RT_GROUP_SCHED
9516 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9517 global_rt_period(), global_rt_runtime());
9518 #ifdef CONFIG_USER_SCHED
9519 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9520 global_rt_period(), RUNTIME_INF);
9521 #endif /* CONFIG_USER_SCHED */
9522 #endif /* CONFIG_RT_GROUP_SCHED */
9524 #ifdef CONFIG_GROUP_SCHED
9525 list_add(&init_task_group.list, &task_groups);
9526 INIT_LIST_HEAD(&init_task_group.children);
9528 #ifdef CONFIG_USER_SCHED
9529 INIT_LIST_HEAD(&root_task_group.children);
9530 init_task_group.parent = &root_task_group;
9531 list_add(&init_task_group.siblings, &root_task_group.children);
9532 #endif /* CONFIG_USER_SCHED */
9533 #endif /* CONFIG_GROUP_SCHED */
9535 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9536 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9537 __alignof__(unsigned long));
9539 for_each_possible_cpu(i) {
9543 raw_spin_lock_init(&rq->lock);
9545 rq->calc_load_active = 0;
9546 rq->calc_load_update = jiffies + LOAD_FREQ;
9547 init_cfs_rq(&rq->cfs, rq);
9548 init_rt_rq(&rq->rt, rq);
9549 #ifdef CONFIG_FAIR_GROUP_SCHED
9550 init_task_group.shares = init_task_group_load;
9551 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9552 #ifdef CONFIG_CGROUP_SCHED
9554 * How much cpu bandwidth does init_task_group get?
9556 * In case of task-groups formed thr' the cgroup filesystem, it
9557 * gets 100% of the cpu resources in the system. This overall
9558 * system cpu resource is divided among the tasks of
9559 * init_task_group and its child task-groups in a fair manner,
9560 * based on each entity's (task or task-group's) weight
9561 * (se->load.weight).
9563 * In other words, if init_task_group has 10 tasks of weight
9564 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9565 * then A0's share of the cpu resource is:
9567 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9569 * We achieve this by letting init_task_group's tasks sit
9570 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9572 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9573 #elif defined CONFIG_USER_SCHED
9574 root_task_group.shares = NICE_0_LOAD;
9575 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9577 * In case of task-groups formed thr' the user id of tasks,
9578 * init_task_group represents tasks belonging to root user.
9579 * Hence it forms a sibling of all subsequent groups formed.
9580 * In this case, init_task_group gets only a fraction of overall
9581 * system cpu resource, based on the weight assigned to root
9582 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9583 * by letting tasks of init_task_group sit in a separate cfs_rq
9584 * (init_tg_cfs_rq) and having one entity represent this group of
9585 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9587 init_tg_cfs_entry(&init_task_group,
9588 &per_cpu(init_tg_cfs_rq, i),
9589 &per_cpu(init_sched_entity, i), i, 1,
9590 root_task_group.se[i]);
9593 #endif /* CONFIG_FAIR_GROUP_SCHED */
9595 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9596 #ifdef CONFIG_RT_GROUP_SCHED
9597 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9598 #ifdef CONFIG_CGROUP_SCHED
9599 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9600 #elif defined CONFIG_USER_SCHED
9601 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9602 init_tg_rt_entry(&init_task_group,
9603 &per_cpu(init_rt_rq_var, i),
9604 &per_cpu(init_sched_rt_entity, i), i, 1,
9605 root_task_group.rt_se[i]);
9609 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9610 rq->cpu_load[j] = 0;
9614 rq->post_schedule = 0;
9615 rq->active_balance = 0;
9616 rq->next_balance = jiffies;
9620 rq->migration_thread = NULL;
9622 rq->avg_idle = 2*sysctl_sched_migration_cost;
9623 INIT_LIST_HEAD(&rq->migration_queue);
9624 rq_attach_root(rq, &def_root_domain);
9627 atomic_set(&rq->nr_iowait, 0);
9630 set_load_weight(&init_task);
9632 #ifdef CONFIG_PREEMPT_NOTIFIERS
9633 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9637 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9640 #ifdef CONFIG_RT_MUTEXES
9641 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
9645 * The boot idle thread does lazy MMU switching as well:
9647 atomic_inc(&init_mm.mm_count);
9648 enter_lazy_tlb(&init_mm, current);
9651 * Make us the idle thread. Technically, schedule() should not be
9652 * called from this thread, however somewhere below it might be,
9653 * but because we are the idle thread, we just pick up running again
9654 * when this runqueue becomes "idle".
9656 init_idle(current, smp_processor_id());
9658 calc_load_update = jiffies + LOAD_FREQ;
9661 * During early bootup we pretend to be a normal task:
9663 current->sched_class = &fair_sched_class;
9665 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9666 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9669 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9670 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9672 /* May be allocated at isolcpus cmdline parse time */
9673 if (cpu_isolated_map == NULL)
9674 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9679 scheduler_running = 1;
9682 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9683 static inline int preempt_count_equals(int preempt_offset)
9685 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
9687 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9690 void __might_sleep(char *file, int line, int preempt_offset)
9693 static unsigned long prev_jiffy; /* ratelimiting */
9695 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9696 system_state != SYSTEM_RUNNING || oops_in_progress)
9698 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9700 prev_jiffy = jiffies;
9702 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9704 pr_err("in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9705 in_atomic(), irqs_disabled(),
9706 current->pid, current->comm);
9708 debug_show_held_locks(current);
9709 if (irqs_disabled())
9710 print_irqtrace_events(current);
9714 EXPORT_SYMBOL(__might_sleep);
9717 #ifdef CONFIG_MAGIC_SYSRQ
9718 static void normalize_task(struct rq *rq, struct task_struct *p)
9722 update_rq_clock(rq);
9723 on_rq = p->se.on_rq;
9725 deactivate_task(rq, p, 0);
9726 __setscheduler(rq, p, SCHED_NORMAL, 0);
9728 activate_task(rq, p, 0);
9729 resched_task(rq->curr);
9733 void normalize_rt_tasks(void)
9735 struct task_struct *g, *p;
9736 unsigned long flags;
9739 read_lock_irqsave(&tasklist_lock, flags);
9740 do_each_thread(g, p) {
9742 * Only normalize user tasks:
9747 p->se.exec_start = 0;
9748 #ifdef CONFIG_SCHEDSTATS
9749 p->se.wait_start = 0;
9750 p->se.sleep_start = 0;
9751 p->se.block_start = 0;
9756 * Renice negative nice level userspace
9759 if (TASK_NICE(p) < 0 && p->mm)
9760 set_user_nice(p, 0);
9764 raw_spin_lock(&p->pi_lock);
9765 rq = __task_rq_lock(p);
9767 normalize_task(rq, p);
9769 __task_rq_unlock(rq);
9770 raw_spin_unlock(&p->pi_lock);
9771 } while_each_thread(g, p);
9773 read_unlock_irqrestore(&tasklist_lock, flags);
9776 #endif /* CONFIG_MAGIC_SYSRQ */
9780 * These functions are only useful for the IA64 MCA handling.
9782 * They can only be called when the whole system has been
9783 * stopped - every CPU needs to be quiescent, and no scheduling
9784 * activity can take place. Using them for anything else would
9785 * be a serious bug, and as a result, they aren't even visible
9786 * under any other configuration.
9790 * curr_task - return the current task for a given cpu.
9791 * @cpu: the processor in question.
9793 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9795 struct task_struct *curr_task(int cpu)
9797 return cpu_curr(cpu);
9801 * set_curr_task - set the current task for a given cpu.
9802 * @cpu: the processor in question.
9803 * @p: the task pointer to set.
9805 * Description: This function must only be used when non-maskable interrupts
9806 * are serviced on a separate stack. It allows the architecture to switch the
9807 * notion of the current task on a cpu in a non-blocking manner. This function
9808 * must be called with all CPU's synchronized, and interrupts disabled, the
9809 * and caller must save the original value of the current task (see
9810 * curr_task() above) and restore that value before reenabling interrupts and
9811 * re-starting the system.
9813 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9815 void set_curr_task(int cpu, struct task_struct *p)
9822 #ifdef CONFIG_FAIR_GROUP_SCHED
9823 static void free_fair_sched_group(struct task_group *tg)
9827 for_each_possible_cpu(i) {
9829 kfree(tg->cfs_rq[i]);
9839 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9841 struct cfs_rq *cfs_rq;
9842 struct sched_entity *se;
9846 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9849 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9853 tg->shares = NICE_0_LOAD;
9855 for_each_possible_cpu(i) {
9858 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9859 GFP_KERNEL, cpu_to_node(i));
9863 se = kzalloc_node(sizeof(struct sched_entity),
9864 GFP_KERNEL, cpu_to_node(i));
9868 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9879 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9881 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9882 &cpu_rq(cpu)->leaf_cfs_rq_list);
9885 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9887 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9889 #else /* !CONFG_FAIR_GROUP_SCHED */
9890 static inline void free_fair_sched_group(struct task_group *tg)
9895 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9900 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9904 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9907 #endif /* CONFIG_FAIR_GROUP_SCHED */
9909 #ifdef CONFIG_RT_GROUP_SCHED
9910 static void free_rt_sched_group(struct task_group *tg)
9914 destroy_rt_bandwidth(&tg->rt_bandwidth);
9916 for_each_possible_cpu(i) {
9918 kfree(tg->rt_rq[i]);
9920 kfree(tg->rt_se[i]);
9928 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9930 struct rt_rq *rt_rq;
9931 struct sched_rt_entity *rt_se;
9935 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9938 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9942 init_rt_bandwidth(&tg->rt_bandwidth,
9943 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9945 for_each_possible_cpu(i) {
9948 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9949 GFP_KERNEL, cpu_to_node(i));
9953 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9954 GFP_KERNEL, cpu_to_node(i));
9958 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9969 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9971 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9972 &cpu_rq(cpu)->leaf_rt_rq_list);
9975 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9977 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9979 #else /* !CONFIG_RT_GROUP_SCHED */
9980 static inline void free_rt_sched_group(struct task_group *tg)
9985 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9990 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9994 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9997 #endif /* CONFIG_RT_GROUP_SCHED */
9999 #ifdef CONFIG_GROUP_SCHED
10000 static void free_sched_group(struct task_group *tg)
10002 free_fair_sched_group(tg);
10003 free_rt_sched_group(tg);
10007 /* allocate runqueue etc for a new task group */
10008 struct task_group *sched_create_group(struct task_group *parent)
10010 struct task_group *tg;
10011 unsigned long flags;
10014 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
10016 return ERR_PTR(-ENOMEM);
10018 if (!alloc_fair_sched_group(tg, parent))
10021 if (!alloc_rt_sched_group(tg, parent))
10024 spin_lock_irqsave(&task_group_lock, flags);
10025 for_each_possible_cpu(i) {
10026 register_fair_sched_group(tg, i);
10027 register_rt_sched_group(tg, i);
10029 list_add_rcu(&tg->list, &task_groups);
10031 WARN_ON(!parent); /* root should already exist */
10033 tg->parent = parent;
10034 INIT_LIST_HEAD(&tg->children);
10035 list_add_rcu(&tg->siblings, &parent->children);
10036 spin_unlock_irqrestore(&task_group_lock, flags);
10041 free_sched_group(tg);
10042 return ERR_PTR(-ENOMEM);
10045 /* rcu callback to free various structures associated with a task group */
10046 static void free_sched_group_rcu(struct rcu_head *rhp)
10048 /* now it should be safe to free those cfs_rqs */
10049 free_sched_group(container_of(rhp, struct task_group, rcu));
10052 /* Destroy runqueue etc associated with a task group */
10053 void sched_destroy_group(struct task_group *tg)
10055 unsigned long flags;
10058 spin_lock_irqsave(&task_group_lock, flags);
10059 for_each_possible_cpu(i) {
10060 unregister_fair_sched_group(tg, i);
10061 unregister_rt_sched_group(tg, i);
10063 list_del_rcu(&tg->list);
10064 list_del_rcu(&tg->siblings);
10065 spin_unlock_irqrestore(&task_group_lock, flags);
10067 /* wait for possible concurrent references to cfs_rqs complete */
10068 call_rcu(&tg->rcu, free_sched_group_rcu);
10071 /* change task's runqueue when it moves between groups.
10072 * The caller of this function should have put the task in its new group
10073 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10074 * reflect its new group.
10076 void sched_move_task(struct task_struct *tsk)
10078 int on_rq, running;
10079 unsigned long flags;
10082 rq = task_rq_lock(tsk, &flags);
10084 update_rq_clock(rq);
10086 running = task_current(rq, tsk);
10087 on_rq = tsk->se.on_rq;
10090 dequeue_task(rq, tsk, 0);
10091 if (unlikely(running))
10092 tsk->sched_class->put_prev_task(rq, tsk);
10094 set_task_rq(tsk, task_cpu(tsk));
10096 #ifdef CONFIG_FAIR_GROUP_SCHED
10097 if (tsk->sched_class->moved_group)
10098 tsk->sched_class->moved_group(tsk, on_rq);
10101 if (unlikely(running))
10102 tsk->sched_class->set_curr_task(rq);
10104 enqueue_task(rq, tsk, 0);
10106 task_rq_unlock(rq, &flags);
10108 #endif /* CONFIG_GROUP_SCHED */
10110 #ifdef CONFIG_FAIR_GROUP_SCHED
10111 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10113 struct cfs_rq *cfs_rq = se->cfs_rq;
10118 dequeue_entity(cfs_rq, se, 0);
10120 se->load.weight = shares;
10121 se->load.inv_weight = 0;
10124 enqueue_entity(cfs_rq, se, 0);
10127 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10129 struct cfs_rq *cfs_rq = se->cfs_rq;
10130 struct rq *rq = cfs_rq->rq;
10131 unsigned long flags;
10133 raw_spin_lock_irqsave(&rq->lock, flags);
10134 __set_se_shares(se, shares);
10135 raw_spin_unlock_irqrestore(&rq->lock, flags);
10138 static DEFINE_MUTEX(shares_mutex);
10140 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10143 unsigned long flags;
10146 * We can't change the weight of the root cgroup.
10151 if (shares < MIN_SHARES)
10152 shares = MIN_SHARES;
10153 else if (shares > MAX_SHARES)
10154 shares = MAX_SHARES;
10156 mutex_lock(&shares_mutex);
10157 if (tg->shares == shares)
10160 spin_lock_irqsave(&task_group_lock, flags);
10161 for_each_possible_cpu(i)
10162 unregister_fair_sched_group(tg, i);
10163 list_del_rcu(&tg->siblings);
10164 spin_unlock_irqrestore(&task_group_lock, flags);
10166 /* wait for any ongoing reference to this group to finish */
10167 synchronize_sched();
10170 * Now we are free to modify the group's share on each cpu
10171 * w/o tripping rebalance_share or load_balance_fair.
10173 tg->shares = shares;
10174 for_each_possible_cpu(i) {
10176 * force a rebalance
10178 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10179 set_se_shares(tg->se[i], shares);
10183 * Enable load balance activity on this group, by inserting it back on
10184 * each cpu's rq->leaf_cfs_rq_list.
10186 spin_lock_irqsave(&task_group_lock, flags);
10187 for_each_possible_cpu(i)
10188 register_fair_sched_group(tg, i);
10189 list_add_rcu(&tg->siblings, &tg->parent->children);
10190 spin_unlock_irqrestore(&task_group_lock, flags);
10192 mutex_unlock(&shares_mutex);
10196 unsigned long sched_group_shares(struct task_group *tg)
10202 #ifdef CONFIG_RT_GROUP_SCHED
10204 * Ensure that the real time constraints are schedulable.
10206 static DEFINE_MUTEX(rt_constraints_mutex);
10208 static unsigned long to_ratio(u64 period, u64 runtime)
10210 if (runtime == RUNTIME_INF)
10213 return div64_u64(runtime << 20, period);
10216 /* Must be called with tasklist_lock held */
10217 static inline int tg_has_rt_tasks(struct task_group *tg)
10219 struct task_struct *g, *p;
10221 do_each_thread(g, p) {
10222 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10224 } while_each_thread(g, p);
10229 struct rt_schedulable_data {
10230 struct task_group *tg;
10235 static int tg_schedulable(struct task_group *tg, void *data)
10237 struct rt_schedulable_data *d = data;
10238 struct task_group *child;
10239 unsigned long total, sum = 0;
10240 u64 period, runtime;
10242 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10243 runtime = tg->rt_bandwidth.rt_runtime;
10246 period = d->rt_period;
10247 runtime = d->rt_runtime;
10250 #ifdef CONFIG_USER_SCHED
10251 if (tg == &root_task_group) {
10252 period = global_rt_period();
10253 runtime = global_rt_runtime();
10258 * Cannot have more runtime than the period.
10260 if (runtime > period && runtime != RUNTIME_INF)
10264 * Ensure we don't starve existing RT tasks.
10266 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10269 total = to_ratio(period, runtime);
10272 * Nobody can have more than the global setting allows.
10274 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10278 * The sum of our children's runtime should not exceed our own.
10280 list_for_each_entry_rcu(child, &tg->children, siblings) {
10281 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10282 runtime = child->rt_bandwidth.rt_runtime;
10284 if (child == d->tg) {
10285 period = d->rt_period;
10286 runtime = d->rt_runtime;
10289 sum += to_ratio(period, runtime);
10298 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10300 struct rt_schedulable_data data = {
10302 .rt_period = period,
10303 .rt_runtime = runtime,
10306 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10309 static int tg_set_bandwidth(struct task_group *tg,
10310 u64 rt_period, u64 rt_runtime)
10314 mutex_lock(&rt_constraints_mutex);
10315 read_lock(&tasklist_lock);
10316 err = __rt_schedulable(tg, rt_period, rt_runtime);
10320 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10321 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10322 tg->rt_bandwidth.rt_runtime = rt_runtime;
10324 for_each_possible_cpu(i) {
10325 struct rt_rq *rt_rq = tg->rt_rq[i];
10327 raw_spin_lock(&rt_rq->rt_runtime_lock);
10328 rt_rq->rt_runtime = rt_runtime;
10329 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10331 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10333 read_unlock(&tasklist_lock);
10334 mutex_unlock(&rt_constraints_mutex);
10339 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10341 u64 rt_runtime, rt_period;
10343 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10344 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10345 if (rt_runtime_us < 0)
10346 rt_runtime = RUNTIME_INF;
10348 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10351 long sched_group_rt_runtime(struct task_group *tg)
10355 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10358 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10359 do_div(rt_runtime_us, NSEC_PER_USEC);
10360 return rt_runtime_us;
10363 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10365 u64 rt_runtime, rt_period;
10367 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10368 rt_runtime = tg->rt_bandwidth.rt_runtime;
10370 if (rt_period == 0)
10373 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10376 long sched_group_rt_period(struct task_group *tg)
10380 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10381 do_div(rt_period_us, NSEC_PER_USEC);
10382 return rt_period_us;
10385 static int sched_rt_global_constraints(void)
10387 u64 runtime, period;
10390 if (sysctl_sched_rt_period <= 0)
10393 runtime = global_rt_runtime();
10394 period = global_rt_period();
10397 * Sanity check on the sysctl variables.
10399 if (runtime > period && runtime != RUNTIME_INF)
10402 mutex_lock(&rt_constraints_mutex);
10403 read_lock(&tasklist_lock);
10404 ret = __rt_schedulable(NULL, 0, 0);
10405 read_unlock(&tasklist_lock);
10406 mutex_unlock(&rt_constraints_mutex);
10411 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10413 /* Don't accept realtime tasks when there is no way for them to run */
10414 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10420 #else /* !CONFIG_RT_GROUP_SCHED */
10421 static int sched_rt_global_constraints(void)
10423 unsigned long flags;
10426 if (sysctl_sched_rt_period <= 0)
10430 * There's always some RT tasks in the root group
10431 * -- migration, kstopmachine etc..
10433 if (sysctl_sched_rt_runtime == 0)
10436 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10437 for_each_possible_cpu(i) {
10438 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10440 raw_spin_lock(&rt_rq->rt_runtime_lock);
10441 rt_rq->rt_runtime = global_rt_runtime();
10442 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10444 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10448 #endif /* CONFIG_RT_GROUP_SCHED */
10450 int sched_rt_handler(struct ctl_table *table, int write,
10451 void __user *buffer, size_t *lenp,
10455 int old_period, old_runtime;
10456 static DEFINE_MUTEX(mutex);
10458 mutex_lock(&mutex);
10459 old_period = sysctl_sched_rt_period;
10460 old_runtime = sysctl_sched_rt_runtime;
10462 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10464 if (!ret && write) {
10465 ret = sched_rt_global_constraints();
10467 sysctl_sched_rt_period = old_period;
10468 sysctl_sched_rt_runtime = old_runtime;
10470 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10471 def_rt_bandwidth.rt_period =
10472 ns_to_ktime(global_rt_period());
10475 mutex_unlock(&mutex);
10480 #ifdef CONFIG_CGROUP_SCHED
10482 /* return corresponding task_group object of a cgroup */
10483 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10485 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10486 struct task_group, css);
10489 static struct cgroup_subsys_state *
10490 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10492 struct task_group *tg, *parent;
10494 if (!cgrp->parent) {
10495 /* This is early initialization for the top cgroup */
10496 return &init_task_group.css;
10499 parent = cgroup_tg(cgrp->parent);
10500 tg = sched_create_group(parent);
10502 return ERR_PTR(-ENOMEM);
10508 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10510 struct task_group *tg = cgroup_tg(cgrp);
10512 sched_destroy_group(tg);
10516 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10518 #ifdef CONFIG_RT_GROUP_SCHED
10519 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10522 /* We don't support RT-tasks being in separate groups */
10523 if (tsk->sched_class != &fair_sched_class)
10530 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10531 struct task_struct *tsk, bool threadgroup)
10533 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10537 struct task_struct *c;
10539 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10540 retval = cpu_cgroup_can_attach_task(cgrp, c);
10552 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10553 struct cgroup *old_cont, struct task_struct *tsk,
10556 sched_move_task(tsk);
10558 struct task_struct *c;
10560 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10561 sched_move_task(c);
10567 #ifdef CONFIG_FAIR_GROUP_SCHED
10568 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10571 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10574 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10576 struct task_group *tg = cgroup_tg(cgrp);
10578 return (u64) tg->shares;
10580 #endif /* CONFIG_FAIR_GROUP_SCHED */
10582 #ifdef CONFIG_RT_GROUP_SCHED
10583 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10586 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10589 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10591 return sched_group_rt_runtime(cgroup_tg(cgrp));
10594 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10597 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10600 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10602 return sched_group_rt_period(cgroup_tg(cgrp));
10604 #endif /* CONFIG_RT_GROUP_SCHED */
10606 static struct cftype cpu_files[] = {
10607 #ifdef CONFIG_FAIR_GROUP_SCHED
10610 .read_u64 = cpu_shares_read_u64,
10611 .write_u64 = cpu_shares_write_u64,
10614 #ifdef CONFIG_RT_GROUP_SCHED
10616 .name = "rt_runtime_us",
10617 .read_s64 = cpu_rt_runtime_read,
10618 .write_s64 = cpu_rt_runtime_write,
10621 .name = "rt_period_us",
10622 .read_u64 = cpu_rt_period_read_uint,
10623 .write_u64 = cpu_rt_period_write_uint,
10628 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10630 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10633 struct cgroup_subsys cpu_cgroup_subsys = {
10635 .create = cpu_cgroup_create,
10636 .destroy = cpu_cgroup_destroy,
10637 .can_attach = cpu_cgroup_can_attach,
10638 .attach = cpu_cgroup_attach,
10639 .populate = cpu_cgroup_populate,
10640 .subsys_id = cpu_cgroup_subsys_id,
10644 #endif /* CONFIG_CGROUP_SCHED */
10646 #ifdef CONFIG_CGROUP_CPUACCT
10649 * CPU accounting code for task groups.
10651 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10652 * (balbir@in.ibm.com).
10655 /* track cpu usage of a group of tasks and its child groups */
10657 struct cgroup_subsys_state css;
10658 /* cpuusage holds pointer to a u64-type object on every cpu */
10660 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10661 struct cpuacct *parent;
10664 struct cgroup_subsys cpuacct_subsys;
10666 /* return cpu accounting group corresponding to this container */
10667 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10669 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10670 struct cpuacct, css);
10673 /* return cpu accounting group to which this task belongs */
10674 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10676 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10677 struct cpuacct, css);
10680 /* create a new cpu accounting group */
10681 static struct cgroup_subsys_state *cpuacct_create(
10682 struct cgroup_subsys *ss, struct cgroup *cgrp)
10684 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10690 ca->cpuusage = alloc_percpu(u64);
10694 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10695 if (percpu_counter_init(&ca->cpustat[i], 0))
10696 goto out_free_counters;
10699 ca->parent = cgroup_ca(cgrp->parent);
10705 percpu_counter_destroy(&ca->cpustat[i]);
10706 free_percpu(ca->cpuusage);
10710 return ERR_PTR(-ENOMEM);
10713 /* destroy an existing cpu accounting group */
10715 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10717 struct cpuacct *ca = cgroup_ca(cgrp);
10720 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10721 percpu_counter_destroy(&ca->cpustat[i]);
10722 free_percpu(ca->cpuusage);
10726 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10728 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10731 #ifndef CONFIG_64BIT
10733 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10735 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10737 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10745 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10747 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10749 #ifndef CONFIG_64BIT
10751 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10753 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10755 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10761 /* return total cpu usage (in nanoseconds) of a group */
10762 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10764 struct cpuacct *ca = cgroup_ca(cgrp);
10765 u64 totalcpuusage = 0;
10768 for_each_present_cpu(i)
10769 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10771 return totalcpuusage;
10774 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10777 struct cpuacct *ca = cgroup_ca(cgrp);
10786 for_each_present_cpu(i)
10787 cpuacct_cpuusage_write(ca, i, 0);
10793 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10794 struct seq_file *m)
10796 struct cpuacct *ca = cgroup_ca(cgroup);
10800 for_each_present_cpu(i) {
10801 percpu = cpuacct_cpuusage_read(ca, i);
10802 seq_printf(m, "%llu ", (unsigned long long) percpu);
10804 seq_printf(m, "\n");
10808 static const char *cpuacct_stat_desc[] = {
10809 [CPUACCT_STAT_USER] = "user",
10810 [CPUACCT_STAT_SYSTEM] = "system",
10813 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10814 struct cgroup_map_cb *cb)
10816 struct cpuacct *ca = cgroup_ca(cgrp);
10819 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10820 s64 val = percpu_counter_read(&ca->cpustat[i]);
10821 val = cputime64_to_clock_t(val);
10822 cb->fill(cb, cpuacct_stat_desc[i], val);
10827 static struct cftype files[] = {
10830 .read_u64 = cpuusage_read,
10831 .write_u64 = cpuusage_write,
10834 .name = "usage_percpu",
10835 .read_seq_string = cpuacct_percpu_seq_read,
10839 .read_map = cpuacct_stats_show,
10843 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10845 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10849 * charge this task's execution time to its accounting group.
10851 * called with rq->lock held.
10853 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10855 struct cpuacct *ca;
10858 if (unlikely(!cpuacct_subsys.active))
10861 cpu = task_cpu(tsk);
10867 for (; ca; ca = ca->parent) {
10868 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10869 *cpuusage += cputime;
10876 * Charge the system/user time to the task's accounting group.
10878 static void cpuacct_update_stats(struct task_struct *tsk,
10879 enum cpuacct_stat_index idx, cputime_t val)
10881 struct cpuacct *ca;
10883 if (unlikely(!cpuacct_subsys.active))
10890 percpu_counter_add(&ca->cpustat[idx], val);
10896 struct cgroup_subsys cpuacct_subsys = {
10898 .create = cpuacct_create,
10899 .destroy = cpuacct_destroy,
10900 .populate = cpuacct_populate,
10901 .subsys_id = cpuacct_subsys_id,
10903 #endif /* CONFIG_CGROUP_CPUACCT */
10907 int rcu_expedited_torture_stats(char *page)
10911 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10913 void synchronize_sched_expedited(void)
10916 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10918 #else /* #ifndef CONFIG_SMP */
10920 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10921 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10923 #define RCU_EXPEDITED_STATE_POST -2
10924 #define RCU_EXPEDITED_STATE_IDLE -1
10926 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10928 int rcu_expedited_torture_stats(char *page)
10933 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10934 for_each_online_cpu(cpu) {
10935 cnt += sprintf(&page[cnt], " %d:%d",
10936 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10938 cnt += sprintf(&page[cnt], "\n");
10941 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10943 static long synchronize_sched_expedited_count;
10946 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10947 * approach to force grace period to end quickly. This consumes
10948 * significant time on all CPUs, and is thus not recommended for
10949 * any sort of common-case code.
10951 * Note that it is illegal to call this function while holding any
10952 * lock that is acquired by a CPU-hotplug notifier. Failing to
10953 * observe this restriction will result in deadlock.
10955 void synchronize_sched_expedited(void)
10958 unsigned long flags;
10959 bool need_full_sync = 0;
10961 struct migration_req *req;
10965 smp_mb(); /* ensure prior mod happens before capturing snap. */
10966 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10968 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10970 if (trycount++ < 10)
10971 udelay(trycount * num_online_cpus());
10973 synchronize_sched();
10976 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10977 smp_mb(); /* ensure test happens before caller kfree */
10982 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10983 for_each_online_cpu(cpu) {
10985 req = &per_cpu(rcu_migration_req, cpu);
10986 init_completion(&req->done);
10988 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10989 raw_spin_lock_irqsave(&rq->lock, flags);
10990 list_add(&req->list, &rq->migration_queue);
10991 raw_spin_unlock_irqrestore(&rq->lock, flags);
10992 wake_up_process(rq->migration_thread);
10994 for_each_online_cpu(cpu) {
10995 rcu_expedited_state = cpu;
10996 req = &per_cpu(rcu_migration_req, cpu);
10998 wait_for_completion(&req->done);
10999 raw_spin_lock_irqsave(&rq->lock, flags);
11000 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
11001 need_full_sync = 1;
11002 req->dest_cpu = RCU_MIGRATION_IDLE;
11003 raw_spin_unlock_irqrestore(&rq->lock, flags);
11005 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
11006 synchronize_sched_expedited_count++;
11007 mutex_unlock(&rcu_sched_expedited_mutex);
11009 if (need_full_sync)
11010 synchronize_sched();
11012 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
11014 #endif /* #else #ifndef CONFIG_SMP */