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
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task);
122 DEFINE_TRACE(sched_wakeup);
123 DEFINE_TRACE(sched_wakeup_new);
124 DEFINE_TRACE(sched_switch);
125 DEFINE_TRACE(sched_migrate_task);
129 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
132 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
133 * Since cpu_power is a 'constant', we can use a reciprocal divide.
135 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
137 return reciprocal_divide(load, sg->reciprocal_cpu_power);
141 * Each time a sched group cpu_power is changed,
142 * we must compute its reciprocal value
144 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
146 sg->__cpu_power += val;
147 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
151 static inline int rt_policy(int policy)
153 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
158 static inline int task_has_rt_policy(struct task_struct *p)
160 return rt_policy(p->policy);
164 * This is the priority-queue data structure of the RT scheduling class:
166 struct rt_prio_array {
167 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
168 struct list_head queue[MAX_RT_PRIO];
171 struct rt_bandwidth {
172 /* nests inside the rq lock: */
173 spinlock_t rt_runtime_lock;
176 struct hrtimer rt_period_timer;
179 static struct rt_bandwidth def_rt_bandwidth;
181 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
183 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
185 struct rt_bandwidth *rt_b =
186 container_of(timer, struct rt_bandwidth, rt_period_timer);
192 now = hrtimer_cb_get_time(timer);
193 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
198 idle = do_sched_rt_period_timer(rt_b, overrun);
201 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
205 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
207 rt_b->rt_period = ns_to_ktime(period);
208 rt_b->rt_runtime = runtime;
210 spin_lock_init(&rt_b->rt_runtime_lock);
212 hrtimer_init(&rt_b->rt_period_timer,
213 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
214 rt_b->rt_period_timer.function = sched_rt_period_timer;
217 static inline int rt_bandwidth_enabled(void)
219 return sysctl_sched_rt_runtime >= 0;
222 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
226 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
229 if (hrtimer_active(&rt_b->rt_period_timer))
232 spin_lock(&rt_b->rt_runtime_lock);
237 if (hrtimer_active(&rt_b->rt_period_timer))
240 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
241 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
243 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
244 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
245 delta = ktime_to_ns(ktime_sub(hard, soft));
246 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
247 HRTIMER_MODE_ABS, 0);
249 spin_unlock(&rt_b->rt_runtime_lock);
252 #ifdef CONFIG_RT_GROUP_SCHED
253 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
255 hrtimer_cancel(&rt_b->rt_period_timer);
260 * sched_domains_mutex serializes calls to arch_init_sched_domains,
261 * detach_destroy_domains and partition_sched_domains.
263 static DEFINE_MUTEX(sched_domains_mutex);
265 #ifdef CONFIG_GROUP_SCHED
267 #include <linux/cgroup.h>
271 static LIST_HEAD(task_groups);
273 /* task group related information */
275 #ifdef CONFIG_CGROUP_SCHED
276 struct cgroup_subsys_state css;
279 #ifdef CONFIG_USER_SCHED
283 #ifdef CONFIG_FAIR_GROUP_SCHED
284 /* schedulable entities of this group on each cpu */
285 struct sched_entity **se;
286 /* runqueue "owned" by this group on each cpu */
287 struct cfs_rq **cfs_rq;
288 unsigned long shares;
291 #ifdef CONFIG_RT_GROUP_SCHED
292 struct sched_rt_entity **rt_se;
293 struct rt_rq **rt_rq;
295 struct rt_bandwidth rt_bandwidth;
299 struct list_head list;
301 struct task_group *parent;
302 struct list_head siblings;
303 struct list_head children;
306 #ifdef CONFIG_USER_SCHED
308 /* Helper function to pass uid information to create_sched_user() */
309 void set_tg_uid(struct user_struct *user)
311 user->tg->uid = user->uid;
316 * Every UID task group (including init_task_group aka UID-0) will
317 * be a child to this group.
319 struct task_group root_task_group;
321 #ifdef CONFIG_FAIR_GROUP_SCHED
322 /* Default task group's sched entity on each cpu */
323 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
324 /* Default task group's cfs_rq on each cpu */
325 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
326 #endif /* CONFIG_FAIR_GROUP_SCHED */
328 #ifdef CONFIG_RT_GROUP_SCHED
329 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
330 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
331 #endif /* CONFIG_RT_GROUP_SCHED */
332 #else /* !CONFIG_USER_SCHED */
333 #define root_task_group init_task_group
334 #endif /* CONFIG_USER_SCHED */
336 /* task_group_lock serializes add/remove of task groups and also changes to
337 * a task group's cpu shares.
339 static DEFINE_SPINLOCK(task_group_lock);
342 static int root_task_group_empty(void)
344 return list_empty(&root_task_group.children);
348 #ifdef CONFIG_FAIR_GROUP_SCHED
349 #ifdef CONFIG_USER_SCHED
350 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
351 #else /* !CONFIG_USER_SCHED */
352 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
353 #endif /* CONFIG_USER_SCHED */
356 * A weight of 0 or 1 can cause arithmetics problems.
357 * A weight of a cfs_rq is the sum of weights of which entities
358 * are queued on this cfs_rq, so a weight of a entity should not be
359 * too large, so as the shares value of a task group.
360 * (The default weight is 1024 - so there's no practical
361 * limitation from this.)
364 #define MAX_SHARES (1UL << 18)
366 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
369 /* Default task group.
370 * Every task in system belong to this group at bootup.
372 struct task_group init_task_group;
374 /* return group to which a task belongs */
375 static inline struct task_group *task_group(struct task_struct *p)
377 struct task_group *tg;
379 #ifdef CONFIG_USER_SCHED
381 tg = __task_cred(p)->user->tg;
383 #elif defined(CONFIG_CGROUP_SCHED)
384 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
385 struct task_group, css);
387 tg = &init_task_group;
392 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
393 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
395 #ifdef CONFIG_FAIR_GROUP_SCHED
396 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
397 p->se.parent = task_group(p)->se[cpu];
400 #ifdef CONFIG_RT_GROUP_SCHED
401 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
402 p->rt.parent = task_group(p)->rt_se[cpu];
409 static int root_task_group_empty(void)
415 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
416 static inline struct task_group *task_group(struct task_struct *p)
421 #endif /* CONFIG_GROUP_SCHED */
423 /* CFS-related fields in a runqueue */
425 struct load_weight load;
426 unsigned long nr_running;
431 struct rb_root tasks_timeline;
432 struct rb_node *rb_leftmost;
434 struct list_head tasks;
435 struct list_head *balance_iterator;
438 * 'curr' points to currently running entity on this cfs_rq.
439 * It is set to NULL otherwise (i.e when none are currently running).
441 struct sched_entity *curr, *next, *last;
443 unsigned int nr_spread_over;
445 #ifdef CONFIG_FAIR_GROUP_SCHED
446 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
449 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
450 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
451 * (like users, containers etc.)
453 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
454 * list is used during load balance.
456 struct list_head leaf_cfs_rq_list;
457 struct task_group *tg; /* group that "owns" this runqueue */
461 * the part of load.weight contributed by tasks
463 unsigned long task_weight;
466 * h_load = weight * f(tg)
468 * Where f(tg) is the recursive weight fraction assigned to
471 unsigned long h_load;
474 * this cpu's part of tg->shares
476 unsigned long shares;
479 * load.weight at the time we set shares
481 unsigned long rq_weight;
486 /* Real-Time classes' related field in a runqueue: */
488 struct rt_prio_array active;
489 unsigned long rt_nr_running;
490 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
492 int curr; /* highest queued rt task prio */
494 int next; /* next highest */
499 unsigned long rt_nr_migratory;
501 struct plist_head pushable_tasks;
506 /* Nests inside the rq lock: */
507 spinlock_t rt_runtime_lock;
509 #ifdef CONFIG_RT_GROUP_SCHED
510 unsigned long rt_nr_boosted;
513 struct list_head leaf_rt_rq_list;
514 struct task_group *tg;
515 struct sched_rt_entity *rt_se;
522 * We add the notion of a root-domain which will be used to define per-domain
523 * variables. Each exclusive cpuset essentially defines an island domain by
524 * fully partitioning the member cpus from any other cpuset. Whenever a new
525 * exclusive cpuset is created, we also create and attach a new root-domain
532 cpumask_var_t online;
535 * The "RT overload" flag: it gets set if a CPU has more than
536 * one runnable RT task.
538 cpumask_var_t rto_mask;
541 struct cpupri cpupri;
543 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
545 * Preferred wake up cpu nominated by sched_mc balance that will be
546 * used when most cpus are idle in the system indicating overall very
547 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
549 unsigned int sched_mc_preferred_wakeup_cpu;
554 * By default the system creates a single root-domain with all cpus as
555 * members (mimicking the global state we have today).
557 static struct root_domain def_root_domain;
562 * This is the main, per-CPU runqueue data structure.
564 * Locking rule: those places that want to lock multiple runqueues
565 * (such as the load balancing or the thread migration code), lock
566 * acquire operations must be ordered by ascending &runqueue.
573 * nr_running and cpu_load should be in the same cacheline because
574 * remote CPUs use both these fields when doing load calculation.
576 unsigned long nr_running;
577 #define CPU_LOAD_IDX_MAX 5
578 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
580 unsigned long last_tick_seen;
581 unsigned char in_nohz_recently;
583 /* capture load from *all* tasks on this cpu: */
584 struct load_weight load;
585 unsigned long nr_load_updates;
587 u64 nr_migrations_in;
592 #ifdef CONFIG_FAIR_GROUP_SCHED
593 /* list of leaf cfs_rq on this cpu: */
594 struct list_head leaf_cfs_rq_list;
596 #ifdef CONFIG_RT_GROUP_SCHED
597 struct list_head leaf_rt_rq_list;
601 * This is part of a global counter where only the total sum
602 * over all CPUs matters. A task can increase this counter on
603 * one CPU and if it got migrated afterwards it may decrease
604 * it on another CPU. Always updated under the runqueue lock:
606 unsigned long nr_uninterruptible;
608 struct task_struct *curr, *idle;
609 unsigned long next_balance;
610 struct mm_struct *prev_mm;
617 struct root_domain *rd;
618 struct sched_domain *sd;
620 unsigned char idle_at_tick;
621 /* For active balancing */
624 /* cpu of this runqueue: */
628 unsigned long avg_load_per_task;
630 struct task_struct *migration_thread;
631 struct list_head migration_queue;
634 #ifdef CONFIG_SCHED_HRTICK
636 int hrtick_csd_pending;
637 struct call_single_data hrtick_csd;
639 struct hrtimer hrtick_timer;
642 #ifdef CONFIG_SCHEDSTATS
644 struct sched_info rq_sched_info;
645 unsigned long long rq_cpu_time;
646 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
648 /* sys_sched_yield() stats */
649 unsigned int yld_count;
651 /* schedule() stats */
652 unsigned int sched_switch;
653 unsigned int sched_count;
654 unsigned int sched_goidle;
656 /* try_to_wake_up() stats */
657 unsigned int ttwu_count;
658 unsigned int ttwu_local;
661 unsigned int bkl_count;
665 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
667 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
669 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
672 static inline int cpu_of(struct rq *rq)
682 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
683 * See detach_destroy_domains: synchronize_sched for details.
685 * The domain tree of any CPU may only be accessed from within
686 * preempt-disabled sections.
688 #define for_each_domain(cpu, __sd) \
689 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
691 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
692 #define this_rq() (&__get_cpu_var(runqueues))
693 #define task_rq(p) cpu_rq(task_cpu(p))
694 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
696 inline void update_rq_clock(struct rq *rq)
698 rq->clock = sched_clock_cpu(cpu_of(rq));
702 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
704 #ifdef CONFIG_SCHED_DEBUG
705 # define const_debug __read_mostly
707 # define const_debug static const
713 * Returns true if the current cpu runqueue is locked.
714 * This interface allows printk to be called with the runqueue lock
715 * held and know whether or not it is OK to wake up the klogd.
717 int runqueue_is_locked(void)
720 struct rq *rq = cpu_rq(cpu);
723 ret = spin_is_locked(&rq->lock);
729 * Debugging: various feature bits
732 #define SCHED_FEAT(name, enabled) \
733 __SCHED_FEAT_##name ,
736 #include "sched_features.h"
741 #define SCHED_FEAT(name, enabled) \
742 (1UL << __SCHED_FEAT_##name) * enabled |
744 const_debug unsigned int sysctl_sched_features =
745 #include "sched_features.h"
750 #ifdef CONFIG_SCHED_DEBUG
751 #define SCHED_FEAT(name, enabled) \
754 static __read_mostly char *sched_feat_names[] = {
755 #include "sched_features.h"
761 static int sched_feat_show(struct seq_file *m, void *v)
765 for (i = 0; sched_feat_names[i]; i++) {
766 if (!(sysctl_sched_features & (1UL << i)))
768 seq_printf(m, "%s ", sched_feat_names[i]);
776 sched_feat_write(struct file *filp, const char __user *ubuf,
777 size_t cnt, loff_t *ppos)
787 if (copy_from_user(&buf, ubuf, cnt))
792 if (strncmp(buf, "NO_", 3) == 0) {
797 for (i = 0; sched_feat_names[i]; i++) {
798 int len = strlen(sched_feat_names[i]);
800 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
802 sysctl_sched_features &= ~(1UL << i);
804 sysctl_sched_features |= (1UL << i);
809 if (!sched_feat_names[i])
817 static int sched_feat_open(struct inode *inode, struct file *filp)
819 return single_open(filp, sched_feat_show, NULL);
822 static struct file_operations sched_feat_fops = {
823 .open = sched_feat_open,
824 .write = sched_feat_write,
827 .release = single_release,
830 static __init int sched_init_debug(void)
832 debugfs_create_file("sched_features", 0644, NULL, NULL,
837 late_initcall(sched_init_debug);
841 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
844 * Number of tasks to iterate in a single balance run.
845 * Limited because this is done with IRQs disabled.
847 const_debug unsigned int sysctl_sched_nr_migrate = 32;
850 * ratelimit for updating the group shares.
853 unsigned int sysctl_sched_shares_ratelimit = 250000;
856 * Inject some fuzzyness into changing the per-cpu group shares
857 * this avoids remote rq-locks at the expense of fairness.
860 unsigned int sysctl_sched_shares_thresh = 4;
863 * period over which we measure -rt task cpu usage in us.
866 unsigned int sysctl_sched_rt_period = 1000000;
868 static __read_mostly int scheduler_running;
871 * part of the period that we allow rt tasks to run in us.
874 int sysctl_sched_rt_runtime = 950000;
876 static inline u64 global_rt_period(void)
878 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
881 static inline u64 global_rt_runtime(void)
883 if (sysctl_sched_rt_runtime < 0)
886 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
889 #ifndef prepare_arch_switch
890 # define prepare_arch_switch(next) do { } while (0)
892 #ifndef finish_arch_switch
893 # define finish_arch_switch(prev) do { } while (0)
896 static inline int task_current(struct rq *rq, struct task_struct *p)
898 return rq->curr == p;
901 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
902 static inline int task_running(struct rq *rq, struct task_struct *p)
904 return task_current(rq, p);
907 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
911 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
913 #ifdef CONFIG_DEBUG_SPINLOCK
914 /* this is a valid case when another task releases the spinlock */
915 rq->lock.owner = current;
918 * If we are tracking spinlock dependencies then we have to
919 * fix up the runqueue lock - which gets 'carried over' from
922 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
924 spin_unlock_irq(&rq->lock);
927 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
928 static inline int task_running(struct rq *rq, struct task_struct *p)
933 return task_current(rq, p);
937 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
941 * We can optimise this out completely for !SMP, because the
942 * SMP rebalancing from interrupt is the only thing that cares
947 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
948 spin_unlock_irq(&rq->lock);
950 spin_unlock(&rq->lock);
954 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
958 * After ->oncpu is cleared, the task can be moved to a different CPU.
959 * We must ensure this doesn't happen until the switch is completely
965 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
969 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
972 * __task_rq_lock - lock the runqueue a given task resides on.
973 * Must be called interrupts disabled.
975 static inline struct rq *__task_rq_lock(struct task_struct *p)
979 struct rq *rq = task_rq(p);
980 spin_lock(&rq->lock);
981 if (likely(rq == task_rq(p)))
983 spin_unlock(&rq->lock);
988 * task_rq_lock - lock the runqueue a given task resides on and disable
989 * interrupts. Note the ordering: we can safely lookup the task_rq without
990 * explicitly disabling preemption.
992 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
998 local_irq_save(*flags);
1000 spin_lock(&rq->lock);
1001 if (likely(rq == task_rq(p)))
1003 spin_unlock_irqrestore(&rq->lock, *flags);
1007 void task_rq_unlock_wait(struct task_struct *p)
1009 struct rq *rq = task_rq(p);
1011 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1012 spin_unlock_wait(&rq->lock);
1015 static void __task_rq_unlock(struct rq *rq)
1016 __releases(rq->lock)
1018 spin_unlock(&rq->lock);
1021 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1022 __releases(rq->lock)
1024 spin_unlock_irqrestore(&rq->lock, *flags);
1028 * this_rq_lock - lock this runqueue and disable interrupts.
1030 static struct rq *this_rq_lock(void)
1031 __acquires(rq->lock)
1035 local_irq_disable();
1037 spin_lock(&rq->lock);
1042 #ifdef CONFIG_SCHED_HRTICK
1044 * Use HR-timers to deliver accurate preemption points.
1046 * Its all a bit involved since we cannot program an hrt while holding the
1047 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1050 * When we get rescheduled we reprogram the hrtick_timer outside of the
1056 * - enabled by features
1057 * - hrtimer is actually high res
1059 static inline int hrtick_enabled(struct rq *rq)
1061 if (!sched_feat(HRTICK))
1063 if (!cpu_active(cpu_of(rq)))
1065 return hrtimer_is_hres_active(&rq->hrtick_timer);
1068 static void hrtick_clear(struct rq *rq)
1070 if (hrtimer_active(&rq->hrtick_timer))
1071 hrtimer_cancel(&rq->hrtick_timer);
1075 * High-resolution timer tick.
1076 * Runs from hardirq context with interrupts disabled.
1078 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1080 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1082 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1084 spin_lock(&rq->lock);
1085 update_rq_clock(rq);
1086 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1087 spin_unlock(&rq->lock);
1089 return HRTIMER_NORESTART;
1094 * called from hardirq (IPI) context
1096 static void __hrtick_start(void *arg)
1098 struct rq *rq = arg;
1100 spin_lock(&rq->lock);
1101 hrtimer_restart(&rq->hrtick_timer);
1102 rq->hrtick_csd_pending = 0;
1103 spin_unlock(&rq->lock);
1107 * Called to set the hrtick timer state.
1109 * called with rq->lock held and irqs disabled
1111 static void hrtick_start(struct rq *rq, u64 delay)
1113 struct hrtimer *timer = &rq->hrtick_timer;
1114 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1116 hrtimer_set_expires(timer, time);
1118 if (rq == this_rq()) {
1119 hrtimer_restart(timer);
1120 } else if (!rq->hrtick_csd_pending) {
1121 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1122 rq->hrtick_csd_pending = 1;
1127 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1129 int cpu = (int)(long)hcpu;
1132 case CPU_UP_CANCELED:
1133 case CPU_UP_CANCELED_FROZEN:
1134 case CPU_DOWN_PREPARE:
1135 case CPU_DOWN_PREPARE_FROZEN:
1137 case CPU_DEAD_FROZEN:
1138 hrtick_clear(cpu_rq(cpu));
1145 static __init void init_hrtick(void)
1147 hotcpu_notifier(hotplug_hrtick, 0);
1151 * Called to set the hrtick timer state.
1153 * called with rq->lock held and irqs disabled
1155 static void hrtick_start(struct rq *rq, u64 delay)
1157 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1158 HRTIMER_MODE_REL, 0);
1161 static inline void init_hrtick(void)
1164 #endif /* CONFIG_SMP */
1166 static void init_rq_hrtick(struct rq *rq)
1169 rq->hrtick_csd_pending = 0;
1171 rq->hrtick_csd.flags = 0;
1172 rq->hrtick_csd.func = __hrtick_start;
1173 rq->hrtick_csd.info = rq;
1176 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1177 rq->hrtick_timer.function = hrtick;
1179 #else /* CONFIG_SCHED_HRTICK */
1180 static inline void hrtick_clear(struct rq *rq)
1184 static inline void init_rq_hrtick(struct rq *rq)
1188 static inline void init_hrtick(void)
1191 #endif /* CONFIG_SCHED_HRTICK */
1194 * resched_task - mark a task 'to be rescheduled now'.
1196 * On UP this means the setting of the need_resched flag, on SMP it
1197 * might also involve a cross-CPU call to trigger the scheduler on
1202 #ifndef tsk_is_polling
1203 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1206 static void resched_task(struct task_struct *p)
1210 assert_spin_locked(&task_rq(p)->lock);
1212 if (test_tsk_need_resched(p))
1215 set_tsk_need_resched(p);
1218 if (cpu == smp_processor_id())
1221 /* NEED_RESCHED must be visible before we test polling */
1223 if (!tsk_is_polling(p))
1224 smp_send_reschedule(cpu);
1227 static void resched_cpu(int cpu)
1229 struct rq *rq = cpu_rq(cpu);
1230 unsigned long flags;
1232 if (!spin_trylock_irqsave(&rq->lock, flags))
1234 resched_task(cpu_curr(cpu));
1235 spin_unlock_irqrestore(&rq->lock, flags);
1240 * When add_timer_on() enqueues a timer into the timer wheel of an
1241 * idle CPU then this timer might expire before the next timer event
1242 * which is scheduled to wake up that CPU. In case of a completely
1243 * idle system the next event might even be infinite time into the
1244 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1245 * leaves the inner idle loop so the newly added timer is taken into
1246 * account when the CPU goes back to idle and evaluates the timer
1247 * wheel for the next timer event.
1249 void wake_up_idle_cpu(int cpu)
1251 struct rq *rq = cpu_rq(cpu);
1253 if (cpu == smp_processor_id())
1257 * This is safe, as this function is called with the timer
1258 * wheel base lock of (cpu) held. When the CPU is on the way
1259 * to idle and has not yet set rq->curr to idle then it will
1260 * be serialized on the timer wheel base lock and take the new
1261 * timer into account automatically.
1263 if (rq->curr != rq->idle)
1267 * We can set TIF_RESCHED on the idle task of the other CPU
1268 * lockless. The worst case is that the other CPU runs the
1269 * idle task through an additional NOOP schedule()
1271 set_tsk_need_resched(rq->idle);
1273 /* NEED_RESCHED must be visible before we test polling */
1275 if (!tsk_is_polling(rq->idle))
1276 smp_send_reschedule(cpu);
1278 #endif /* CONFIG_NO_HZ */
1280 #else /* !CONFIG_SMP */
1281 static void resched_task(struct task_struct *p)
1283 assert_spin_locked(&task_rq(p)->lock);
1284 set_tsk_need_resched(p);
1286 #endif /* CONFIG_SMP */
1288 #if BITS_PER_LONG == 32
1289 # define WMULT_CONST (~0UL)
1291 # define WMULT_CONST (1UL << 32)
1294 #define WMULT_SHIFT 32
1297 * Shift right and round:
1299 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1302 * delta *= weight / lw
1304 static unsigned long
1305 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1306 struct load_weight *lw)
1310 if (!lw->inv_weight) {
1311 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1314 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1318 tmp = (u64)delta_exec * weight;
1320 * Check whether we'd overflow the 64-bit multiplication:
1322 if (unlikely(tmp > WMULT_CONST))
1323 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1326 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1328 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1331 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1337 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1344 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1345 * of tasks with abnormal "nice" values across CPUs the contribution that
1346 * each task makes to its run queue's load is weighted according to its
1347 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1348 * scaled version of the new time slice allocation that they receive on time
1352 #define WEIGHT_IDLEPRIO 3
1353 #define WMULT_IDLEPRIO 1431655765
1356 * Nice levels are multiplicative, with a gentle 10% change for every
1357 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1358 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1359 * that remained on nice 0.
1361 * The "10% effect" is relative and cumulative: from _any_ nice level,
1362 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1363 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1364 * If a task goes up by ~10% and another task goes down by ~10% then
1365 * the relative distance between them is ~25%.)
1367 static const int prio_to_weight[40] = {
1368 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1369 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1370 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1371 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1372 /* 0 */ 1024, 820, 655, 526, 423,
1373 /* 5 */ 335, 272, 215, 172, 137,
1374 /* 10 */ 110, 87, 70, 56, 45,
1375 /* 15 */ 36, 29, 23, 18, 15,
1379 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1381 * In cases where the weight does not change often, we can use the
1382 * precalculated inverse to speed up arithmetics by turning divisions
1383 * into multiplications:
1385 static const u32 prio_to_wmult[40] = {
1386 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1387 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1388 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1389 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1390 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1391 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1392 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1393 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1396 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1399 * runqueue iterator, to support SMP load-balancing between different
1400 * scheduling classes, without having to expose their internal data
1401 * structures to the load-balancing proper:
1403 struct rq_iterator {
1405 struct task_struct *(*start)(void *);
1406 struct task_struct *(*next)(void *);
1410 static unsigned long
1411 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1412 unsigned long max_load_move, struct sched_domain *sd,
1413 enum cpu_idle_type idle, int *all_pinned,
1414 int *this_best_prio, struct rq_iterator *iterator);
1417 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1418 struct sched_domain *sd, enum cpu_idle_type idle,
1419 struct rq_iterator *iterator);
1422 #ifdef CONFIG_CGROUP_CPUACCT
1423 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1425 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1428 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1430 update_load_add(&rq->load, load);
1433 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1435 update_load_sub(&rq->load, load);
1438 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1439 typedef int (*tg_visitor)(struct task_group *, void *);
1442 * Iterate the full tree, calling @down when first entering a node and @up when
1443 * leaving it for the final time.
1445 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1447 struct task_group *parent, *child;
1451 parent = &root_task_group;
1453 ret = (*down)(parent, data);
1456 list_for_each_entry_rcu(child, &parent->children, siblings) {
1463 ret = (*up)(parent, data);
1468 parent = parent->parent;
1477 static int tg_nop(struct task_group *tg, void *data)
1484 static unsigned long source_load(int cpu, int type);
1485 static unsigned long target_load(int cpu, int type);
1486 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1488 static unsigned long cpu_avg_load_per_task(int cpu)
1490 struct rq *rq = cpu_rq(cpu);
1491 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1494 rq->avg_load_per_task = rq->load.weight / nr_running;
1496 rq->avg_load_per_task = 0;
1498 return rq->avg_load_per_task;
1501 #ifdef CONFIG_FAIR_GROUP_SCHED
1503 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1506 * Calculate and set the cpu's group shares.
1509 update_group_shares_cpu(struct task_group *tg, int cpu,
1510 unsigned long sd_shares, unsigned long sd_rq_weight)
1512 unsigned long shares;
1513 unsigned long rq_weight;
1518 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1521 * \Sum shares * rq_weight
1522 * shares = -----------------------
1526 shares = (sd_shares * rq_weight) / sd_rq_weight;
1527 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1529 if (abs(shares - tg->se[cpu]->load.weight) >
1530 sysctl_sched_shares_thresh) {
1531 struct rq *rq = cpu_rq(cpu);
1532 unsigned long flags;
1534 spin_lock_irqsave(&rq->lock, flags);
1535 tg->cfs_rq[cpu]->shares = shares;
1537 __set_se_shares(tg->se[cpu], shares);
1538 spin_unlock_irqrestore(&rq->lock, flags);
1543 * Re-compute the task group their per cpu shares over the given domain.
1544 * This needs to be done in a bottom-up fashion because the rq weight of a
1545 * parent group depends on the shares of its child groups.
1547 static int tg_shares_up(struct task_group *tg, void *data)
1549 unsigned long weight, rq_weight = 0;
1550 unsigned long shares = 0;
1551 struct sched_domain *sd = data;
1554 for_each_cpu(i, sched_domain_span(sd)) {
1556 * If there are currently no tasks on the cpu pretend there
1557 * is one of average load so that when a new task gets to
1558 * run here it will not get delayed by group starvation.
1560 weight = tg->cfs_rq[i]->load.weight;
1562 weight = NICE_0_LOAD;
1564 tg->cfs_rq[i]->rq_weight = weight;
1565 rq_weight += weight;
1566 shares += tg->cfs_rq[i]->shares;
1569 if ((!shares && rq_weight) || shares > tg->shares)
1570 shares = tg->shares;
1572 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1573 shares = tg->shares;
1575 for_each_cpu(i, sched_domain_span(sd))
1576 update_group_shares_cpu(tg, i, shares, rq_weight);
1582 * Compute the cpu's hierarchical load factor for each task group.
1583 * This needs to be done in a top-down fashion because the load of a child
1584 * group is a fraction of its parents load.
1586 static int tg_load_down(struct task_group *tg, void *data)
1589 long cpu = (long)data;
1592 load = cpu_rq(cpu)->load.weight;
1594 load = tg->parent->cfs_rq[cpu]->h_load;
1595 load *= tg->cfs_rq[cpu]->shares;
1596 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1599 tg->cfs_rq[cpu]->h_load = load;
1604 static void update_shares(struct sched_domain *sd)
1606 u64 now = cpu_clock(raw_smp_processor_id());
1607 s64 elapsed = now - sd->last_update;
1609 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1610 sd->last_update = now;
1611 walk_tg_tree(tg_nop, tg_shares_up, sd);
1615 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1617 spin_unlock(&rq->lock);
1619 spin_lock(&rq->lock);
1622 static void update_h_load(long cpu)
1624 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1629 static inline void update_shares(struct sched_domain *sd)
1633 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1639 #ifdef CONFIG_PREEMPT
1642 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1643 * way at the expense of forcing extra atomic operations in all
1644 * invocations. This assures that the double_lock is acquired using the
1645 * same underlying policy as the spinlock_t on this architecture, which
1646 * reduces latency compared to the unfair variant below. However, it
1647 * also adds more overhead and therefore may reduce throughput.
1649 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1650 __releases(this_rq->lock)
1651 __acquires(busiest->lock)
1652 __acquires(this_rq->lock)
1654 spin_unlock(&this_rq->lock);
1655 double_rq_lock(this_rq, busiest);
1662 * Unfair double_lock_balance: Optimizes throughput at the expense of
1663 * latency by eliminating extra atomic operations when the locks are
1664 * already in proper order on entry. This favors lower cpu-ids and will
1665 * grant the double lock to lower cpus over higher ids under contention,
1666 * regardless of entry order into the function.
1668 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1669 __releases(this_rq->lock)
1670 __acquires(busiest->lock)
1671 __acquires(this_rq->lock)
1675 if (unlikely(!spin_trylock(&busiest->lock))) {
1676 if (busiest < this_rq) {
1677 spin_unlock(&this_rq->lock);
1678 spin_lock(&busiest->lock);
1679 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1682 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1687 #endif /* CONFIG_PREEMPT */
1690 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1692 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1694 if (unlikely(!irqs_disabled())) {
1695 /* printk() doesn't work good under rq->lock */
1696 spin_unlock(&this_rq->lock);
1700 return _double_lock_balance(this_rq, busiest);
1703 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1704 __releases(busiest->lock)
1706 spin_unlock(&busiest->lock);
1707 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1711 #ifdef CONFIG_FAIR_GROUP_SCHED
1712 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1715 cfs_rq->shares = shares;
1720 #include "sched_stats.h"
1721 #include "sched_idletask.c"
1722 #include "sched_fair.c"
1723 #include "sched_rt.c"
1724 #ifdef CONFIG_SCHED_DEBUG
1725 # include "sched_debug.c"
1728 #define sched_class_highest (&rt_sched_class)
1729 #define for_each_class(class) \
1730 for (class = sched_class_highest; class; class = class->next)
1732 static void inc_nr_running(struct rq *rq)
1737 static void dec_nr_running(struct rq *rq)
1742 static void set_load_weight(struct task_struct *p)
1744 if (task_has_rt_policy(p)) {
1745 p->se.load.weight = prio_to_weight[0] * 2;
1746 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1751 * SCHED_IDLE tasks get minimal weight:
1753 if (p->policy == SCHED_IDLE) {
1754 p->se.load.weight = WEIGHT_IDLEPRIO;
1755 p->se.load.inv_weight = WMULT_IDLEPRIO;
1759 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1760 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1763 static void update_avg(u64 *avg, u64 sample)
1765 s64 diff = sample - *avg;
1769 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1772 p->se.start_runtime = p->se.sum_exec_runtime;
1774 sched_info_queued(p);
1775 p->sched_class->enqueue_task(rq, p, wakeup);
1779 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1782 if (p->se.last_wakeup) {
1783 update_avg(&p->se.avg_overlap,
1784 p->se.sum_exec_runtime - p->se.last_wakeup);
1785 p->se.last_wakeup = 0;
1787 update_avg(&p->se.avg_wakeup,
1788 sysctl_sched_wakeup_granularity);
1792 sched_info_dequeued(p);
1793 p->sched_class->dequeue_task(rq, p, sleep);
1798 * __normal_prio - return the priority that is based on the static prio
1800 static inline int __normal_prio(struct task_struct *p)
1802 return p->static_prio;
1806 * Calculate the expected normal priority: i.e. priority
1807 * without taking RT-inheritance into account. Might be
1808 * boosted by interactivity modifiers. Changes upon fork,
1809 * setprio syscalls, and whenever the interactivity
1810 * estimator recalculates.
1812 static inline int normal_prio(struct task_struct *p)
1816 if (task_has_rt_policy(p))
1817 prio = MAX_RT_PRIO-1 - p->rt_priority;
1819 prio = __normal_prio(p);
1824 * Calculate the current priority, i.e. the priority
1825 * taken into account by the scheduler. This value might
1826 * be boosted by RT tasks, or might be boosted by
1827 * interactivity modifiers. Will be RT if the task got
1828 * RT-boosted. If not then it returns p->normal_prio.
1830 static int effective_prio(struct task_struct *p)
1832 p->normal_prio = normal_prio(p);
1834 * If we are RT tasks or we were boosted to RT priority,
1835 * keep the priority unchanged. Otherwise, update priority
1836 * to the normal priority:
1838 if (!rt_prio(p->prio))
1839 return p->normal_prio;
1844 * activate_task - move a task to the runqueue.
1846 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1848 if (task_contributes_to_load(p))
1849 rq->nr_uninterruptible--;
1851 enqueue_task(rq, p, wakeup);
1856 * deactivate_task - remove a task from the runqueue.
1858 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1860 if (task_contributes_to_load(p))
1861 rq->nr_uninterruptible++;
1863 dequeue_task(rq, p, sleep);
1868 * task_curr - is this task currently executing on a CPU?
1869 * @p: the task in question.
1871 inline int task_curr(const struct task_struct *p)
1873 return cpu_curr(task_cpu(p)) == p;
1876 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1878 set_task_rq(p, cpu);
1881 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1882 * successfuly executed on another CPU. We must ensure that updates of
1883 * per-task data have been completed by this moment.
1886 task_thread_info(p)->cpu = cpu;
1890 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1891 const struct sched_class *prev_class,
1892 int oldprio, int running)
1894 if (prev_class != p->sched_class) {
1895 if (prev_class->switched_from)
1896 prev_class->switched_from(rq, p, running);
1897 p->sched_class->switched_to(rq, p, running);
1899 p->sched_class->prio_changed(rq, p, oldprio, running);
1904 /* Used instead of source_load when we know the type == 0 */
1905 static unsigned long weighted_cpuload(const int cpu)
1907 return cpu_rq(cpu)->load.weight;
1911 * Is this task likely cache-hot:
1914 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1919 * Buddy candidates are cache hot:
1921 if (sched_feat(CACHE_HOT_BUDDY) &&
1922 (&p->se == cfs_rq_of(&p->se)->next ||
1923 &p->se == cfs_rq_of(&p->se)->last))
1926 if (p->sched_class != &fair_sched_class)
1929 if (sysctl_sched_migration_cost == -1)
1931 if (sysctl_sched_migration_cost == 0)
1934 delta = now - p->se.exec_start;
1936 return delta < (s64)sysctl_sched_migration_cost;
1940 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1942 int old_cpu = task_cpu(p);
1943 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1944 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1945 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1948 clock_offset = old_rq->clock - new_rq->clock;
1950 trace_sched_migrate_task(p, task_cpu(p), new_cpu);
1952 #ifdef CONFIG_SCHEDSTATS
1953 if (p->se.wait_start)
1954 p->se.wait_start -= clock_offset;
1955 if (p->se.sleep_start)
1956 p->se.sleep_start -= clock_offset;
1957 if (p->se.block_start)
1958 p->se.block_start -= clock_offset;
1960 if (old_cpu != new_cpu) {
1961 p->se.nr_migrations++;
1962 new_rq->nr_migrations_in++;
1963 #ifdef CONFIG_SCHEDSTATS
1964 if (task_hot(p, old_rq->clock, NULL))
1965 schedstat_inc(p, se.nr_forced2_migrations);
1968 p->se.vruntime -= old_cfsrq->min_vruntime -
1969 new_cfsrq->min_vruntime;
1971 __set_task_cpu(p, new_cpu);
1974 struct migration_req {
1975 struct list_head list;
1977 struct task_struct *task;
1980 struct completion done;
1984 * The task's runqueue lock must be held.
1985 * Returns true if you have to wait for migration thread.
1988 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1990 struct rq *rq = task_rq(p);
1993 * If the task is not on a runqueue (and not running), then
1994 * it is sufficient to simply update the task's cpu field.
1996 if (!p->se.on_rq && !task_running(rq, p)) {
1997 set_task_cpu(p, dest_cpu);
2001 init_completion(&req->done);
2003 req->dest_cpu = dest_cpu;
2004 list_add(&req->list, &rq->migration_queue);
2010 * wait_task_inactive - wait for a thread to unschedule.
2012 * If @match_state is nonzero, it's the @p->state value just checked and
2013 * not expected to change. If it changes, i.e. @p might have woken up,
2014 * then return zero. When we succeed in waiting for @p to be off its CPU,
2015 * we return a positive number (its total switch count). If a second call
2016 * a short while later returns the same number, the caller can be sure that
2017 * @p has remained unscheduled the whole time.
2019 * The caller must ensure that the task *will* unschedule sometime soon,
2020 * else this function might spin for a *long* time. This function can't
2021 * be called with interrupts off, or it may introduce deadlock with
2022 * smp_call_function() if an IPI is sent by the same process we are
2023 * waiting to become inactive.
2025 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2027 unsigned long flags;
2034 * We do the initial early heuristics without holding
2035 * any task-queue locks at all. We'll only try to get
2036 * the runqueue lock when things look like they will
2042 * If the task is actively running on another CPU
2043 * still, just relax and busy-wait without holding
2046 * NOTE! Since we don't hold any locks, it's not
2047 * even sure that "rq" stays as the right runqueue!
2048 * But we don't care, since "task_running()" will
2049 * return false if the runqueue has changed and p
2050 * is actually now running somewhere else!
2052 while (task_running(rq, p)) {
2053 if (match_state && unlikely(p->state != match_state))
2059 * Ok, time to look more closely! We need the rq
2060 * lock now, to be *sure*. If we're wrong, we'll
2061 * just go back and repeat.
2063 rq = task_rq_lock(p, &flags);
2064 trace_sched_wait_task(rq, p);
2065 running = task_running(rq, p);
2066 on_rq = p->se.on_rq;
2068 if (!match_state || p->state == match_state)
2069 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2070 task_rq_unlock(rq, &flags);
2073 * If it changed from the expected state, bail out now.
2075 if (unlikely(!ncsw))
2079 * Was it really running after all now that we
2080 * checked with the proper locks actually held?
2082 * Oops. Go back and try again..
2084 if (unlikely(running)) {
2090 * It's not enough that it's not actively running,
2091 * it must be off the runqueue _entirely_, and not
2094 * So if it was still runnable (but just not actively
2095 * running right now), it's preempted, and we should
2096 * yield - it could be a while.
2098 if (unlikely(on_rq)) {
2099 schedule_timeout_uninterruptible(1);
2104 * Ahh, all good. It wasn't running, and it wasn't
2105 * runnable, which means that it will never become
2106 * running in the future either. We're all done!
2115 * kick_process - kick a running thread to enter/exit the kernel
2116 * @p: the to-be-kicked thread
2118 * Cause a process which is running on another CPU to enter
2119 * kernel-mode, without any delay. (to get signals handled.)
2121 * NOTE: this function doesnt have to take the runqueue lock,
2122 * because all it wants to ensure is that the remote task enters
2123 * the kernel. If the IPI races and the task has been migrated
2124 * to another CPU then no harm is done and the purpose has been
2127 void kick_process(struct task_struct *p)
2133 if ((cpu != smp_processor_id()) && task_curr(p))
2134 smp_send_reschedule(cpu);
2139 * Return a low guess at the load of a migration-source cpu weighted
2140 * according to the scheduling class and "nice" value.
2142 * We want to under-estimate the load of migration sources, to
2143 * balance conservatively.
2145 static unsigned long source_load(int cpu, int type)
2147 struct rq *rq = cpu_rq(cpu);
2148 unsigned long total = weighted_cpuload(cpu);
2150 if (type == 0 || !sched_feat(LB_BIAS))
2153 return min(rq->cpu_load[type-1], total);
2157 * Return a high guess at the load of a migration-target cpu weighted
2158 * according to the scheduling class and "nice" value.
2160 static unsigned long target_load(int cpu, int type)
2162 struct rq *rq = cpu_rq(cpu);
2163 unsigned long total = weighted_cpuload(cpu);
2165 if (type == 0 || !sched_feat(LB_BIAS))
2168 return max(rq->cpu_load[type-1], total);
2172 * find_idlest_group finds and returns the least busy CPU group within the
2175 static struct sched_group *
2176 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2178 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2179 unsigned long min_load = ULONG_MAX, this_load = 0;
2180 int load_idx = sd->forkexec_idx;
2181 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2184 unsigned long load, avg_load;
2188 /* Skip over this group if it has no CPUs allowed */
2189 if (!cpumask_intersects(sched_group_cpus(group),
2193 local_group = cpumask_test_cpu(this_cpu,
2194 sched_group_cpus(group));
2196 /* Tally up the load of all CPUs in the group */
2199 for_each_cpu(i, sched_group_cpus(group)) {
2200 /* Bias balancing toward cpus of our domain */
2202 load = source_load(i, load_idx);
2204 load = target_load(i, load_idx);
2209 /* Adjust by relative CPU power of the group */
2210 avg_load = sg_div_cpu_power(group,
2211 avg_load * SCHED_LOAD_SCALE);
2214 this_load = avg_load;
2216 } else if (avg_load < min_load) {
2217 min_load = avg_load;
2220 } while (group = group->next, group != sd->groups);
2222 if (!idlest || 100*this_load < imbalance*min_load)
2228 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2231 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2233 unsigned long load, min_load = ULONG_MAX;
2237 /* Traverse only the allowed CPUs */
2238 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2239 load = weighted_cpuload(i);
2241 if (load < min_load || (load == min_load && i == this_cpu)) {
2251 * sched_balance_self: balance the current task (running on cpu) in domains
2252 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2255 * Balance, ie. select the least loaded group.
2257 * Returns the target CPU number, or the same CPU if no balancing is needed.
2259 * preempt must be disabled.
2261 static int sched_balance_self(int cpu, int flag)
2263 struct task_struct *t = current;
2264 struct sched_domain *tmp, *sd = NULL;
2266 for_each_domain(cpu, tmp) {
2268 * If power savings logic is enabled for a domain, stop there.
2270 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2272 if (tmp->flags & flag)
2280 struct sched_group *group;
2281 int new_cpu, weight;
2283 if (!(sd->flags & flag)) {
2288 group = find_idlest_group(sd, t, cpu);
2294 new_cpu = find_idlest_cpu(group, t, cpu);
2295 if (new_cpu == -1 || new_cpu == cpu) {
2296 /* Now try balancing at a lower domain level of cpu */
2301 /* Now try balancing at a lower domain level of new_cpu */
2303 weight = cpumask_weight(sched_domain_span(sd));
2305 for_each_domain(cpu, tmp) {
2306 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2308 if (tmp->flags & flag)
2311 /* while loop will break here if sd == NULL */
2317 #endif /* CONFIG_SMP */
2320 * task_oncpu_function_call - call a function on the cpu on which a task runs
2321 * @p: the task to evaluate
2322 * @func: the function to be called
2323 * @info: the function call argument
2325 * Calls the function @func when the task is currently running. This might
2326 * be on the current CPU, which just calls the function directly
2328 void task_oncpu_function_call(struct task_struct *p,
2329 void (*func) (void *info), void *info)
2336 smp_call_function_single(cpu, func, info, 1);
2341 * try_to_wake_up - wake up a thread
2342 * @p: the to-be-woken-up thread
2343 * @state: the mask of task states that can be woken
2344 * @sync: do a synchronous wakeup?
2346 * Put it on the run-queue if it's not already there. The "current"
2347 * thread is always on the run-queue (except when the actual
2348 * re-schedule is in progress), and as such you're allowed to do
2349 * the simpler "current->state = TASK_RUNNING" to mark yourself
2350 * runnable without the overhead of this.
2352 * returns failure only if the task is already active.
2354 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2356 int cpu, orig_cpu, this_cpu, success = 0;
2357 unsigned long flags;
2361 if (!sched_feat(SYNC_WAKEUPS))
2365 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2366 struct sched_domain *sd;
2368 this_cpu = raw_smp_processor_id();
2371 for_each_domain(this_cpu, sd) {
2372 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2381 rq = task_rq_lock(p, &flags);
2382 update_rq_clock(rq);
2383 old_state = p->state;
2384 if (!(old_state & state))
2392 this_cpu = smp_processor_id();
2395 if (unlikely(task_running(rq, p)))
2398 cpu = p->sched_class->select_task_rq(p, sync);
2399 if (cpu != orig_cpu) {
2400 set_task_cpu(p, cpu);
2401 task_rq_unlock(rq, &flags);
2402 /* might preempt at this point */
2403 rq = task_rq_lock(p, &flags);
2404 old_state = p->state;
2405 if (!(old_state & state))
2410 this_cpu = smp_processor_id();
2414 #ifdef CONFIG_SCHEDSTATS
2415 schedstat_inc(rq, ttwu_count);
2416 if (cpu == this_cpu)
2417 schedstat_inc(rq, ttwu_local);
2419 struct sched_domain *sd;
2420 for_each_domain(this_cpu, sd) {
2421 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2422 schedstat_inc(sd, ttwu_wake_remote);
2427 #endif /* CONFIG_SCHEDSTATS */
2430 #endif /* CONFIG_SMP */
2431 schedstat_inc(p, se.nr_wakeups);
2433 schedstat_inc(p, se.nr_wakeups_sync);
2434 if (orig_cpu != cpu)
2435 schedstat_inc(p, se.nr_wakeups_migrate);
2436 if (cpu == this_cpu)
2437 schedstat_inc(p, se.nr_wakeups_local);
2439 schedstat_inc(p, se.nr_wakeups_remote);
2440 activate_task(rq, p, 1);
2444 * Only attribute actual wakeups done by this task.
2446 if (!in_interrupt()) {
2447 struct sched_entity *se = ¤t->se;
2448 u64 sample = se->sum_exec_runtime;
2450 if (se->last_wakeup)
2451 sample -= se->last_wakeup;
2453 sample -= se->start_runtime;
2454 update_avg(&se->avg_wakeup, sample);
2456 se->last_wakeup = se->sum_exec_runtime;
2460 trace_sched_wakeup(rq, p, success);
2461 check_preempt_curr(rq, p, sync);
2463 p->state = TASK_RUNNING;
2465 if (p->sched_class->task_wake_up)
2466 p->sched_class->task_wake_up(rq, p);
2469 task_rq_unlock(rq, &flags);
2474 int wake_up_process(struct task_struct *p)
2476 return try_to_wake_up(p, TASK_ALL, 0);
2478 EXPORT_SYMBOL(wake_up_process);
2480 int wake_up_state(struct task_struct *p, unsigned int state)
2482 return try_to_wake_up(p, state, 0);
2486 * Perform scheduler related setup for a newly forked process p.
2487 * p is forked by current.
2489 * __sched_fork() is basic setup used by init_idle() too:
2491 static void __sched_fork(struct task_struct *p)
2493 p->se.exec_start = 0;
2494 p->se.sum_exec_runtime = 0;
2495 p->se.prev_sum_exec_runtime = 0;
2496 p->se.nr_migrations = 0;
2497 p->se.last_wakeup = 0;
2498 p->se.avg_overlap = 0;
2499 p->se.start_runtime = 0;
2500 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2502 #ifdef CONFIG_SCHEDSTATS
2503 p->se.wait_start = 0;
2504 p->se.sum_sleep_runtime = 0;
2505 p->se.sleep_start = 0;
2506 p->se.block_start = 0;
2507 p->se.sleep_max = 0;
2508 p->se.block_max = 0;
2510 p->se.slice_max = 0;
2514 INIT_LIST_HEAD(&p->rt.run_list);
2516 INIT_LIST_HEAD(&p->se.group_node);
2518 #ifdef CONFIG_PREEMPT_NOTIFIERS
2519 INIT_HLIST_HEAD(&p->preempt_notifiers);
2523 * We mark the process as running here, but have not actually
2524 * inserted it onto the runqueue yet. This guarantees that
2525 * nobody will actually run it, and a signal or other external
2526 * event cannot wake it up and insert it on the runqueue either.
2528 p->state = TASK_RUNNING;
2532 * fork()/clone()-time setup:
2534 void sched_fork(struct task_struct *p, int clone_flags)
2536 int cpu = get_cpu();
2541 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2543 set_task_cpu(p, cpu);
2546 * Make sure we do not leak PI boosting priority to the child:
2548 p->prio = current->normal_prio;
2549 if (!rt_prio(p->prio))
2550 p->sched_class = &fair_sched_class;
2552 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2553 if (likely(sched_info_on()))
2554 memset(&p->sched_info, 0, sizeof(p->sched_info));
2556 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2559 #ifdef CONFIG_PREEMPT
2560 /* Want to start with kernel preemption disabled. */
2561 task_thread_info(p)->preempt_count = 1;
2563 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2569 * wake_up_new_task - wake up a newly created task for the first time.
2571 * This function will do some initial scheduler statistics housekeeping
2572 * that must be done for every newly created context, then puts the task
2573 * on the runqueue and wakes it.
2575 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2577 unsigned long flags;
2580 rq = task_rq_lock(p, &flags);
2581 BUG_ON(p->state != TASK_RUNNING);
2582 update_rq_clock(rq);
2584 p->prio = effective_prio(p);
2586 if (!p->sched_class->task_new || !current->se.on_rq) {
2587 activate_task(rq, p, 0);
2590 * Let the scheduling class do new task startup
2591 * management (if any):
2593 p->sched_class->task_new(rq, p);
2596 trace_sched_wakeup_new(rq, p, 1);
2597 check_preempt_curr(rq, p, 0);
2599 if (p->sched_class->task_wake_up)
2600 p->sched_class->task_wake_up(rq, p);
2602 task_rq_unlock(rq, &flags);
2605 #ifdef CONFIG_PREEMPT_NOTIFIERS
2608 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2609 * @notifier: notifier struct to register
2611 void preempt_notifier_register(struct preempt_notifier *notifier)
2613 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2615 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2618 * preempt_notifier_unregister - no longer interested in preemption notifications
2619 * @notifier: notifier struct to unregister
2621 * This is safe to call from within a preemption notifier.
2623 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2625 hlist_del(¬ifier->link);
2627 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2629 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2631 struct preempt_notifier *notifier;
2632 struct hlist_node *node;
2634 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2635 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2639 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2640 struct task_struct *next)
2642 struct preempt_notifier *notifier;
2643 struct hlist_node *node;
2645 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2646 notifier->ops->sched_out(notifier, next);
2649 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2651 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2656 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2657 struct task_struct *next)
2661 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2664 * prepare_task_switch - prepare to switch tasks
2665 * @rq: the runqueue preparing to switch
2666 * @prev: the current task that is being switched out
2667 * @next: the task we are going to switch to.
2669 * This is called with the rq lock held and interrupts off. It must
2670 * be paired with a subsequent finish_task_switch after the context
2673 * prepare_task_switch sets up locking and calls architecture specific
2677 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2678 struct task_struct *next)
2680 fire_sched_out_preempt_notifiers(prev, next);
2681 prepare_lock_switch(rq, next);
2682 prepare_arch_switch(next);
2686 * finish_task_switch - clean up after a task-switch
2687 * @rq: runqueue associated with task-switch
2688 * @prev: the thread we just switched away from.
2690 * finish_task_switch must be called after the context switch, paired
2691 * with a prepare_task_switch call before the context switch.
2692 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2693 * and do any other architecture-specific cleanup actions.
2695 * Note that we may have delayed dropping an mm in context_switch(). If
2696 * so, we finish that here outside of the runqueue lock. (Doing it
2697 * with the lock held can cause deadlocks; see schedule() for
2700 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2701 __releases(rq->lock)
2703 struct mm_struct *mm = rq->prev_mm;
2706 int post_schedule = 0;
2708 if (current->sched_class->needs_post_schedule)
2709 post_schedule = current->sched_class->needs_post_schedule(rq);
2715 * A task struct has one reference for the use as "current".
2716 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2717 * schedule one last time. The schedule call will never return, and
2718 * the scheduled task must drop that reference.
2719 * The test for TASK_DEAD must occur while the runqueue locks are
2720 * still held, otherwise prev could be scheduled on another cpu, die
2721 * there before we look at prev->state, and then the reference would
2723 * Manfred Spraul <manfred@colorfullife.com>
2725 prev_state = prev->state;
2726 finish_arch_switch(prev);
2727 perf_counter_task_sched_in(current, cpu_of(rq));
2728 finish_lock_switch(rq, prev);
2731 current->sched_class->post_schedule(rq);
2734 fire_sched_in_preempt_notifiers(current);
2737 if (unlikely(prev_state == TASK_DEAD)) {
2739 * Remove function-return probe instances associated with this
2740 * task and put them back on the free list.
2742 kprobe_flush_task(prev);
2743 put_task_struct(prev);
2748 * schedule_tail - first thing a freshly forked thread must call.
2749 * @prev: the thread we just switched away from.
2751 asmlinkage void schedule_tail(struct task_struct *prev)
2752 __releases(rq->lock)
2754 struct rq *rq = this_rq();
2756 finish_task_switch(rq, prev);
2757 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2758 /* In this case, finish_task_switch does not reenable preemption */
2761 if (current->set_child_tid)
2762 put_user(task_pid_vnr(current), current->set_child_tid);
2766 * context_switch - switch to the new MM and the new
2767 * thread's register state.
2770 context_switch(struct rq *rq, struct task_struct *prev,
2771 struct task_struct *next)
2773 struct mm_struct *mm, *oldmm;
2775 prepare_task_switch(rq, prev, next);
2776 trace_sched_switch(rq, prev, next);
2778 oldmm = prev->active_mm;
2780 * For paravirt, this is coupled with an exit in switch_to to
2781 * combine the page table reload and the switch backend into
2784 arch_enter_lazy_cpu_mode();
2786 if (unlikely(!mm)) {
2787 next->active_mm = oldmm;
2788 atomic_inc(&oldmm->mm_count);
2789 enter_lazy_tlb(oldmm, next);
2791 switch_mm(oldmm, mm, next);
2793 if (unlikely(!prev->mm)) {
2794 prev->active_mm = NULL;
2795 rq->prev_mm = oldmm;
2798 * Since the runqueue lock will be released by the next
2799 * task (which is an invalid locking op but in the case
2800 * of the scheduler it's an obvious special-case), so we
2801 * do an early lockdep release here:
2803 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2804 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2807 /* Here we just switch the register state and the stack. */
2808 switch_to(prev, next, prev);
2812 * this_rq must be evaluated again because prev may have moved
2813 * CPUs since it called schedule(), thus the 'rq' on its stack
2814 * frame will be invalid.
2816 finish_task_switch(this_rq(), prev);
2820 * nr_running, nr_uninterruptible and nr_context_switches:
2822 * externally visible scheduler statistics: current number of runnable
2823 * threads, current number of uninterruptible-sleeping threads, total
2824 * number of context switches performed since bootup.
2826 unsigned long nr_running(void)
2828 unsigned long i, sum = 0;
2830 for_each_online_cpu(i)
2831 sum += cpu_rq(i)->nr_running;
2836 unsigned long nr_uninterruptible(void)
2838 unsigned long i, sum = 0;
2840 for_each_possible_cpu(i)
2841 sum += cpu_rq(i)->nr_uninterruptible;
2844 * Since we read the counters lockless, it might be slightly
2845 * inaccurate. Do not allow it to go below zero though:
2847 if (unlikely((long)sum < 0))
2853 unsigned long long nr_context_switches(void)
2856 unsigned long long sum = 0;
2858 for_each_possible_cpu(i)
2859 sum += cpu_rq(i)->nr_switches;
2864 unsigned long nr_iowait(void)
2866 unsigned long i, sum = 0;
2868 for_each_possible_cpu(i)
2869 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2874 unsigned long nr_active(void)
2876 unsigned long i, running = 0, uninterruptible = 0;
2878 for_each_online_cpu(i) {
2879 running += cpu_rq(i)->nr_running;
2880 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2883 if (unlikely((long)uninterruptible < 0))
2884 uninterruptible = 0;
2886 return running + uninterruptible;
2890 * Externally visible per-cpu scheduler statistics:
2891 * cpu_nr_migrations(cpu) - number of migrations into that cpu
2893 u64 cpu_nr_migrations(int cpu)
2895 return cpu_rq(cpu)->nr_migrations_in;
2899 * Update rq->cpu_load[] statistics. This function is usually called every
2900 * scheduler tick (TICK_NSEC).
2902 static void update_cpu_load(struct rq *this_rq)
2904 unsigned long this_load = this_rq->load.weight;
2907 this_rq->nr_load_updates++;
2909 /* Update our load: */
2910 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2911 unsigned long old_load, new_load;
2913 /* scale is effectively 1 << i now, and >> i divides by scale */
2915 old_load = this_rq->cpu_load[i];
2916 new_load = this_load;
2918 * Round up the averaging division if load is increasing. This
2919 * prevents us from getting stuck on 9 if the load is 10, for
2922 if (new_load > old_load)
2923 new_load += scale-1;
2924 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2931 * double_rq_lock - safely lock two runqueues
2933 * Note this does not disable interrupts like task_rq_lock,
2934 * you need to do so manually before calling.
2936 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2937 __acquires(rq1->lock)
2938 __acquires(rq2->lock)
2940 BUG_ON(!irqs_disabled());
2942 spin_lock(&rq1->lock);
2943 __acquire(rq2->lock); /* Fake it out ;) */
2946 spin_lock(&rq1->lock);
2947 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2949 spin_lock(&rq2->lock);
2950 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2953 update_rq_clock(rq1);
2954 update_rq_clock(rq2);
2958 * double_rq_unlock - safely unlock two runqueues
2960 * Note this does not restore interrupts like task_rq_unlock,
2961 * you need to do so manually after calling.
2963 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2964 __releases(rq1->lock)
2965 __releases(rq2->lock)
2967 spin_unlock(&rq1->lock);
2969 spin_unlock(&rq2->lock);
2971 __release(rq2->lock);
2975 * If dest_cpu is allowed for this process, migrate the task to it.
2976 * This is accomplished by forcing the cpu_allowed mask to only
2977 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2978 * the cpu_allowed mask is restored.
2980 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2982 struct migration_req req;
2983 unsigned long flags;
2986 rq = task_rq_lock(p, &flags);
2987 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
2988 || unlikely(!cpu_active(dest_cpu)))
2991 /* force the process onto the specified CPU */
2992 if (migrate_task(p, dest_cpu, &req)) {
2993 /* Need to wait for migration thread (might exit: take ref). */
2994 struct task_struct *mt = rq->migration_thread;
2996 get_task_struct(mt);
2997 task_rq_unlock(rq, &flags);
2998 wake_up_process(mt);
2999 put_task_struct(mt);
3000 wait_for_completion(&req.done);
3005 task_rq_unlock(rq, &flags);
3009 * sched_exec - execve() is a valuable balancing opportunity, because at
3010 * this point the task has the smallest effective memory and cache footprint.
3012 void sched_exec(void)
3014 int new_cpu, this_cpu = get_cpu();
3015 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3017 if (new_cpu != this_cpu)
3018 sched_migrate_task(current, new_cpu);
3022 * pull_task - move a task from a remote runqueue to the local runqueue.
3023 * Both runqueues must be locked.
3025 static void pull_task(struct rq *src_rq, struct task_struct *p,
3026 struct rq *this_rq, int this_cpu)
3028 deactivate_task(src_rq, p, 0);
3029 set_task_cpu(p, this_cpu);
3030 activate_task(this_rq, p, 0);
3032 * Note that idle threads have a prio of MAX_PRIO, for this test
3033 * to be always true for them.
3035 check_preempt_curr(this_rq, p, 0);
3039 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3042 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3043 struct sched_domain *sd, enum cpu_idle_type idle,
3046 int tsk_cache_hot = 0;
3048 * We do not migrate tasks that are:
3049 * 1) running (obviously), or
3050 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3051 * 3) are cache-hot on their current CPU.
3053 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3054 schedstat_inc(p, se.nr_failed_migrations_affine);
3059 if (task_running(rq, p)) {
3060 schedstat_inc(p, se.nr_failed_migrations_running);
3065 * Aggressive migration if:
3066 * 1) task is cache cold, or
3067 * 2) too many balance attempts have failed.
3070 tsk_cache_hot = task_hot(p, rq->clock, sd);
3071 if (!tsk_cache_hot ||
3072 sd->nr_balance_failed > sd->cache_nice_tries) {
3073 #ifdef CONFIG_SCHEDSTATS
3074 if (tsk_cache_hot) {
3075 schedstat_inc(sd, lb_hot_gained[idle]);
3076 schedstat_inc(p, se.nr_forced_migrations);
3082 if (tsk_cache_hot) {
3083 schedstat_inc(p, se.nr_failed_migrations_hot);
3089 static unsigned long
3090 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3091 unsigned long max_load_move, struct sched_domain *sd,
3092 enum cpu_idle_type idle, int *all_pinned,
3093 int *this_best_prio, struct rq_iterator *iterator)
3095 int loops = 0, pulled = 0, pinned = 0;
3096 struct task_struct *p;
3097 long rem_load_move = max_load_move;
3099 if (max_load_move == 0)
3105 * Start the load-balancing iterator:
3107 p = iterator->start(iterator->arg);
3109 if (!p || loops++ > sysctl_sched_nr_migrate)
3112 if ((p->se.load.weight >> 1) > rem_load_move ||
3113 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3114 p = iterator->next(iterator->arg);
3118 pull_task(busiest, p, this_rq, this_cpu);
3120 rem_load_move -= p->se.load.weight;
3122 #ifdef CONFIG_PREEMPT
3124 * NEWIDLE balancing is a source of latency, so preemptible kernels
3125 * will stop after the first task is pulled to minimize the critical
3128 if (idle == CPU_NEWLY_IDLE)
3133 * We only want to steal up to the prescribed amount of weighted load.
3135 if (rem_load_move > 0) {
3136 if (p->prio < *this_best_prio)
3137 *this_best_prio = p->prio;
3138 p = iterator->next(iterator->arg);
3143 * Right now, this is one of only two places pull_task() is called,
3144 * so we can safely collect pull_task() stats here rather than
3145 * inside pull_task().
3147 schedstat_add(sd, lb_gained[idle], pulled);
3150 *all_pinned = pinned;
3152 return max_load_move - rem_load_move;
3156 * move_tasks tries to move up to max_load_move weighted load from busiest to
3157 * this_rq, as part of a balancing operation within domain "sd".
3158 * Returns 1 if successful and 0 otherwise.
3160 * Called with both runqueues locked.
3162 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3163 unsigned long max_load_move,
3164 struct sched_domain *sd, enum cpu_idle_type idle,
3167 const struct sched_class *class = sched_class_highest;
3168 unsigned long total_load_moved = 0;
3169 int this_best_prio = this_rq->curr->prio;
3173 class->load_balance(this_rq, this_cpu, busiest,
3174 max_load_move - total_load_moved,
3175 sd, idle, all_pinned, &this_best_prio);
3176 class = class->next;
3178 #ifdef CONFIG_PREEMPT
3180 * NEWIDLE balancing is a source of latency, so preemptible
3181 * kernels will stop after the first task is pulled to minimize
3182 * the critical section.
3184 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3187 } while (class && max_load_move > total_load_moved);
3189 return total_load_moved > 0;
3193 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3194 struct sched_domain *sd, enum cpu_idle_type idle,
3195 struct rq_iterator *iterator)
3197 struct task_struct *p = iterator->start(iterator->arg);
3201 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3202 pull_task(busiest, p, this_rq, this_cpu);
3204 * Right now, this is only the second place pull_task()
3205 * is called, so we can safely collect pull_task()
3206 * stats here rather than inside pull_task().
3208 schedstat_inc(sd, lb_gained[idle]);
3212 p = iterator->next(iterator->arg);
3219 * move_one_task tries to move exactly one task from busiest to this_rq, as
3220 * part of active balancing operations within "domain".
3221 * Returns 1 if successful and 0 otherwise.
3223 * Called with both runqueues locked.
3225 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3226 struct sched_domain *sd, enum cpu_idle_type idle)
3228 const struct sched_class *class;
3230 for (class = sched_class_highest; class; class = class->next)
3231 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3236 /********** Helpers for find_busiest_group ************************/
3238 * sd_lb_stats - Structure to store the statistics of a sched_domain
3239 * during load balancing.
3241 struct sd_lb_stats {
3242 struct sched_group *busiest; /* Busiest group in this sd */
3243 struct sched_group *this; /* Local group in this sd */
3244 unsigned long total_load; /* Total load of all groups in sd */
3245 unsigned long total_pwr; /* Total power of all groups in sd */
3246 unsigned long avg_load; /* Average load across all groups in sd */
3248 /** Statistics of this group */
3249 unsigned long this_load;
3250 unsigned long this_load_per_task;
3251 unsigned long this_nr_running;
3253 /* Statistics of the busiest group */
3254 unsigned long max_load;
3255 unsigned long busiest_load_per_task;
3256 unsigned long busiest_nr_running;
3258 int group_imb; /* Is there imbalance in this sd */
3259 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3260 int power_savings_balance; /* Is powersave balance needed for this sd */
3261 struct sched_group *group_min; /* Least loaded group in sd */
3262 struct sched_group *group_leader; /* Group which relieves group_min */
3263 unsigned long min_load_per_task; /* load_per_task in group_min */
3264 unsigned long leader_nr_running; /* Nr running of group_leader */
3265 unsigned long min_nr_running; /* Nr running of group_min */
3270 * sg_lb_stats - stats of a sched_group required for load_balancing
3272 struct sg_lb_stats {
3273 unsigned long avg_load; /*Avg load across the CPUs of the group */
3274 unsigned long group_load; /* Total load over the CPUs of the group */
3275 unsigned long sum_nr_running; /* Nr tasks running in the group */
3276 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3277 unsigned long group_capacity;
3278 int group_imb; /* Is there an imbalance in the group ? */
3282 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3283 * @group: The group whose first cpu is to be returned.
3285 static inline unsigned int group_first_cpu(struct sched_group *group)
3287 return cpumask_first(sched_group_cpus(group));
3291 * get_sd_load_idx - Obtain the load index for a given sched domain.
3292 * @sd: The sched_domain whose load_idx is to be obtained.
3293 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3295 static inline int get_sd_load_idx(struct sched_domain *sd,
3296 enum cpu_idle_type idle)
3302 load_idx = sd->busy_idx;
3305 case CPU_NEWLY_IDLE:
3306 load_idx = sd->newidle_idx;
3309 load_idx = sd->idle_idx;
3317 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3319 * init_sd_power_savings_stats - Initialize power savings statistics for
3320 * the given sched_domain, during load balancing.
3322 * @sd: Sched domain whose power-savings statistics are to be initialized.
3323 * @sds: Variable containing the statistics for sd.
3324 * @idle: Idle status of the CPU at which we're performing load-balancing.
3326 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3327 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3330 * Busy processors will not participate in power savings
3333 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3334 sds->power_savings_balance = 0;
3336 sds->power_savings_balance = 1;
3337 sds->min_nr_running = ULONG_MAX;
3338 sds->leader_nr_running = 0;
3343 * update_sd_power_savings_stats - Update the power saving stats for a
3344 * sched_domain while performing load balancing.
3346 * @group: sched_group belonging to the sched_domain under consideration.
3347 * @sds: Variable containing the statistics of the sched_domain
3348 * @local_group: Does group contain the CPU for which we're performing
3350 * @sgs: Variable containing the statistics of the group.
3352 static inline void update_sd_power_savings_stats(struct sched_group *group,
3353 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3356 if (!sds->power_savings_balance)
3360 * If the local group is idle or completely loaded
3361 * no need to do power savings balance at this domain
3363 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3364 !sds->this_nr_running))
3365 sds->power_savings_balance = 0;
3368 * If a group is already running at full capacity or idle,
3369 * don't include that group in power savings calculations
3371 if (!sds->power_savings_balance ||
3372 sgs->sum_nr_running >= sgs->group_capacity ||
3373 !sgs->sum_nr_running)
3377 * Calculate the group which has the least non-idle load.
3378 * This is the group from where we need to pick up the load
3381 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3382 (sgs->sum_nr_running == sds->min_nr_running &&
3383 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3384 sds->group_min = group;
3385 sds->min_nr_running = sgs->sum_nr_running;
3386 sds->min_load_per_task = sgs->sum_weighted_load /
3387 sgs->sum_nr_running;
3391 * Calculate the group which is almost near its
3392 * capacity but still has some space to pick up some load
3393 * from other group and save more power
3395 if (sgs->sum_nr_running > sgs->group_capacity - 1)
3398 if (sgs->sum_nr_running > sds->leader_nr_running ||
3399 (sgs->sum_nr_running == sds->leader_nr_running &&
3400 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3401 sds->group_leader = group;
3402 sds->leader_nr_running = sgs->sum_nr_running;
3407 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3408 * @sds: Variable containing the statistics of the sched_domain
3409 * under consideration.
3410 * @this_cpu: Cpu at which we're currently performing load-balancing.
3411 * @imbalance: Variable to store the imbalance.
3414 * Check if we have potential to perform some power-savings balance.
3415 * If yes, set the busiest group to be the least loaded group in the
3416 * sched_domain, so that it's CPUs can be put to idle.
3418 * Returns 1 if there is potential to perform power-savings balance.
3421 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3422 int this_cpu, unsigned long *imbalance)
3424 if (!sds->power_savings_balance)
3427 if (sds->this != sds->group_leader ||
3428 sds->group_leader == sds->group_min)
3431 *imbalance = sds->min_load_per_task;
3432 sds->busiest = sds->group_min;
3434 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3435 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3436 group_first_cpu(sds->group_leader);
3442 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3443 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3444 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3449 static inline void update_sd_power_savings_stats(struct sched_group *group,
3450 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3455 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3456 int this_cpu, unsigned long *imbalance)
3460 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3464 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3465 * @group: sched_group whose statistics are to be updated.
3466 * @this_cpu: Cpu for which load balance is currently performed.
3467 * @idle: Idle status of this_cpu
3468 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3469 * @sd_idle: Idle status of the sched_domain containing group.
3470 * @local_group: Does group contain this_cpu.
3471 * @cpus: Set of cpus considered for load balancing.
3472 * @balance: Should we balance.
3473 * @sgs: variable to hold the statistics for this group.
3475 static inline void update_sg_lb_stats(struct sched_group *group, int this_cpu,
3476 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3477 int local_group, const struct cpumask *cpus,
3478 int *balance, struct sg_lb_stats *sgs)
3480 unsigned long load, max_cpu_load, min_cpu_load;
3482 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3483 unsigned long sum_avg_load_per_task;
3484 unsigned long avg_load_per_task;
3487 balance_cpu = group_first_cpu(group);
3489 /* Tally up the load of all CPUs in the group */
3490 sum_avg_load_per_task = avg_load_per_task = 0;
3492 min_cpu_load = ~0UL;
3494 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3495 struct rq *rq = cpu_rq(i);
3497 if (*sd_idle && rq->nr_running)
3500 /* Bias balancing toward cpus of our domain */
3502 if (idle_cpu(i) && !first_idle_cpu) {
3507 load = target_load(i, load_idx);
3509 load = source_load(i, load_idx);
3510 if (load > max_cpu_load)
3511 max_cpu_load = load;
3512 if (min_cpu_load > load)
3513 min_cpu_load = load;
3516 sgs->group_load += load;
3517 sgs->sum_nr_running += rq->nr_running;
3518 sgs->sum_weighted_load += weighted_cpuload(i);
3520 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3524 * First idle cpu or the first cpu(busiest) in this sched group
3525 * is eligible for doing load balancing at this and above
3526 * domains. In the newly idle case, we will allow all the cpu's
3527 * to do the newly idle load balance.
3529 if (idle != CPU_NEWLY_IDLE && local_group &&
3530 balance_cpu != this_cpu && balance) {
3535 /* Adjust by relative CPU power of the group */
3536 sgs->avg_load = sg_div_cpu_power(group,
3537 sgs->group_load * SCHED_LOAD_SCALE);
3541 * Consider the group unbalanced when the imbalance is larger
3542 * than the average weight of two tasks.
3544 * APZ: with cgroup the avg task weight can vary wildly and
3545 * might not be a suitable number - should we keep a
3546 * normalized nr_running number somewhere that negates
3549 avg_load_per_task = sg_div_cpu_power(group,
3550 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3552 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3555 sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3560 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3561 * @sd: sched_domain whose statistics are to be updated.
3562 * @this_cpu: Cpu for which load balance is currently performed.
3563 * @idle: Idle status of this_cpu
3564 * @sd_idle: Idle status of the sched_domain containing group.
3565 * @cpus: Set of cpus considered for load balancing.
3566 * @balance: Should we balance.
3567 * @sds: variable to hold the statistics for this sched_domain.
3569 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3570 enum cpu_idle_type idle, int *sd_idle,
3571 const struct cpumask *cpus, int *balance,
3572 struct sd_lb_stats *sds)
3574 struct sched_group *group = sd->groups;
3575 struct sg_lb_stats sgs;
3578 init_sd_power_savings_stats(sd, sds, idle);
3579 load_idx = get_sd_load_idx(sd, idle);
3584 local_group = cpumask_test_cpu(this_cpu,
3585 sched_group_cpus(group));
3586 memset(&sgs, 0, sizeof(sgs));
3587 update_sg_lb_stats(group, this_cpu, idle, load_idx, sd_idle,
3588 local_group, cpus, balance, &sgs);
3590 if (local_group && balance && !(*balance))
3593 sds->total_load += sgs.group_load;
3594 sds->total_pwr += group->__cpu_power;
3597 sds->this_load = sgs.avg_load;
3599 sds->this_nr_running = sgs.sum_nr_running;
3600 sds->this_load_per_task = sgs.sum_weighted_load;
3601 } else if (sgs.avg_load > sds->max_load &&
3602 (sgs.sum_nr_running > sgs.group_capacity ||
3604 sds->max_load = sgs.avg_load;
3605 sds->busiest = group;
3606 sds->busiest_nr_running = sgs.sum_nr_running;
3607 sds->busiest_load_per_task = sgs.sum_weighted_load;
3608 sds->group_imb = sgs.group_imb;
3611 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3612 group = group->next;
3613 } while (group != sd->groups);
3618 * fix_small_imbalance - Calculate the minor imbalance that exists
3619 * amongst the groups of a sched_domain, during
3621 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3622 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3623 * @imbalance: Variable to store the imbalance.
3625 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3626 int this_cpu, unsigned long *imbalance)
3628 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3629 unsigned int imbn = 2;
3631 if (sds->this_nr_running) {
3632 sds->this_load_per_task /= sds->this_nr_running;
3633 if (sds->busiest_load_per_task >
3634 sds->this_load_per_task)
3637 sds->this_load_per_task =
3638 cpu_avg_load_per_task(this_cpu);
3640 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3641 sds->busiest_load_per_task * imbn) {
3642 *imbalance = sds->busiest_load_per_task;
3647 * OK, we don't have enough imbalance to justify moving tasks,
3648 * however we may be able to increase total CPU power used by
3652 pwr_now += sds->busiest->__cpu_power *
3653 min(sds->busiest_load_per_task, sds->max_load);
3654 pwr_now += sds->this->__cpu_power *
3655 min(sds->this_load_per_task, sds->this_load);
3656 pwr_now /= SCHED_LOAD_SCALE;
3658 /* Amount of load we'd subtract */
3659 tmp = sg_div_cpu_power(sds->busiest,
3660 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3661 if (sds->max_load > tmp)
3662 pwr_move += sds->busiest->__cpu_power *
3663 min(sds->busiest_load_per_task, sds->max_load - tmp);
3665 /* Amount of load we'd add */
3666 if (sds->max_load * sds->busiest->__cpu_power <
3667 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3668 tmp = sg_div_cpu_power(sds->this,
3669 sds->max_load * sds->busiest->__cpu_power);
3671 tmp = sg_div_cpu_power(sds->this,
3672 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3673 pwr_move += sds->this->__cpu_power *
3674 min(sds->this_load_per_task, sds->this_load + tmp);
3675 pwr_move /= SCHED_LOAD_SCALE;
3677 /* Move if we gain throughput */
3678 if (pwr_move > pwr_now)
3679 *imbalance = sds->busiest_load_per_task;
3683 * calculate_imbalance - Calculate the amount of imbalance present within the
3684 * groups of a given sched_domain during load balance.
3685 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3686 * @this_cpu: Cpu for which currently load balance is being performed.
3687 * @imbalance: The variable to store the imbalance.
3689 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3690 unsigned long *imbalance)
3692 unsigned long max_pull;
3694 * In the presence of smp nice balancing, certain scenarios can have
3695 * max load less than avg load(as we skip the groups at or below
3696 * its cpu_power, while calculating max_load..)
3698 if (sds->max_load < sds->avg_load) {
3700 return fix_small_imbalance(sds, this_cpu, imbalance);
3703 /* Don't want to pull so many tasks that a group would go idle */
3704 max_pull = min(sds->max_load - sds->avg_load,
3705 sds->max_load - sds->busiest_load_per_task);
3707 /* How much load to actually move to equalise the imbalance */
3708 *imbalance = min(max_pull * sds->busiest->__cpu_power,
3709 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
3713 * if *imbalance is less than the average load per runnable task
3714 * there is no gaurantee that any tasks will be moved so we'll have
3715 * a think about bumping its value to force at least one task to be
3718 if (*imbalance < sds->busiest_load_per_task)
3719 return fix_small_imbalance(sds, this_cpu, imbalance);
3722 /******* find_busiest_group() helpers end here *********************/
3725 * find_busiest_group - Returns the busiest group within the sched_domain
3726 * if there is an imbalance. If there isn't an imbalance, and
3727 * the user has opted for power-savings, it returns a group whose
3728 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3729 * such a group exists.
3731 * Also calculates the amount of weighted load which should be moved
3732 * to restore balance.
3734 * @sd: The sched_domain whose busiest group is to be returned.
3735 * @this_cpu: The cpu for which load balancing is currently being performed.
3736 * @imbalance: Variable which stores amount of weighted load which should
3737 * be moved to restore balance/put a group to idle.
3738 * @idle: The idle status of this_cpu.
3739 * @sd_idle: The idleness of sd
3740 * @cpus: The set of CPUs under consideration for load-balancing.
3741 * @balance: Pointer to a variable indicating if this_cpu
3742 * is the appropriate cpu to perform load balancing at this_level.
3744 * Returns: - the busiest group if imbalance exists.
3745 * - If no imbalance and user has opted for power-savings balance,
3746 * return the least loaded group whose CPUs can be
3747 * put to idle by rebalancing its tasks onto our group.
3749 static struct sched_group *
3750 find_busiest_group(struct sched_domain *sd, int this_cpu,
3751 unsigned long *imbalance, enum cpu_idle_type idle,
3752 int *sd_idle, const struct cpumask *cpus, int *balance)
3754 struct sd_lb_stats sds;
3756 memset(&sds, 0, sizeof(sds));
3759 * Compute the various statistics relavent for load balancing at
3762 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3765 /* Cases where imbalance does not exist from POV of this_cpu */
3766 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3768 * 2) There is no busy sibling group to pull from.
3769 * 3) This group is the busiest group.
3770 * 4) This group is more busy than the avg busieness at this
3772 * 5) The imbalance is within the specified limit.
3773 * 6) Any rebalance would lead to ping-pong
3775 if (balance && !(*balance))
3778 if (!sds.busiest || sds.busiest_nr_running == 0)
3781 if (sds.this_load >= sds.max_load)
3784 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3786 if (sds.this_load >= sds.avg_load)
3789 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3792 sds.busiest_load_per_task /= sds.busiest_nr_running;
3794 sds.busiest_load_per_task =
3795 min(sds.busiest_load_per_task, sds.avg_load);
3798 * We're trying to get all the cpus to the average_load, so we don't
3799 * want to push ourselves above the average load, nor do we wish to
3800 * reduce the max loaded cpu below the average load, as either of these
3801 * actions would just result in more rebalancing later, and ping-pong
3802 * tasks around. Thus we look for the minimum possible imbalance.
3803 * Negative imbalances (*we* are more loaded than anyone else) will
3804 * be counted as no imbalance for these purposes -- we can't fix that
3805 * by pulling tasks to us. Be careful of negative numbers as they'll
3806 * appear as very large values with unsigned longs.
3808 if (sds.max_load <= sds.busiest_load_per_task)
3811 /* Looks like there is an imbalance. Compute it */
3812 calculate_imbalance(&sds, this_cpu, imbalance);
3817 * There is no obvious imbalance. But check if we can do some balancing
3820 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3828 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3831 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3832 unsigned long imbalance, const struct cpumask *cpus)
3834 struct rq *busiest = NULL, *rq;
3835 unsigned long max_load = 0;
3838 for_each_cpu(i, sched_group_cpus(group)) {
3841 if (!cpumask_test_cpu(i, cpus))
3845 wl = weighted_cpuload(i);
3847 if (rq->nr_running == 1 && wl > imbalance)
3850 if (wl > max_load) {
3860 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3861 * so long as it is large enough.
3863 #define MAX_PINNED_INTERVAL 512
3865 /* Working cpumask for load_balance and load_balance_newidle. */
3866 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3869 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3870 * tasks if there is an imbalance.
3872 static int load_balance(int this_cpu, struct rq *this_rq,
3873 struct sched_domain *sd, enum cpu_idle_type idle,
3876 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3877 struct sched_group *group;
3878 unsigned long imbalance;
3880 unsigned long flags;
3881 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3883 cpumask_setall(cpus);
3886 * When power savings policy is enabled for the parent domain, idle
3887 * sibling can pick up load irrespective of busy siblings. In this case,
3888 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3889 * portraying it as CPU_NOT_IDLE.
3891 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3892 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3895 schedstat_inc(sd, lb_count[idle]);
3899 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3906 schedstat_inc(sd, lb_nobusyg[idle]);
3910 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3912 schedstat_inc(sd, lb_nobusyq[idle]);
3916 BUG_ON(busiest == this_rq);
3918 schedstat_add(sd, lb_imbalance[idle], imbalance);
3921 if (busiest->nr_running > 1) {
3923 * Attempt to move tasks. If find_busiest_group has found
3924 * an imbalance but busiest->nr_running <= 1, the group is
3925 * still unbalanced. ld_moved simply stays zero, so it is
3926 * correctly treated as an imbalance.
3928 local_irq_save(flags);
3929 double_rq_lock(this_rq, busiest);
3930 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3931 imbalance, sd, idle, &all_pinned);
3932 double_rq_unlock(this_rq, busiest);
3933 local_irq_restore(flags);
3936 * some other cpu did the load balance for us.
3938 if (ld_moved && this_cpu != smp_processor_id())
3939 resched_cpu(this_cpu);
3941 /* All tasks on this runqueue were pinned by CPU affinity */
3942 if (unlikely(all_pinned)) {
3943 cpumask_clear_cpu(cpu_of(busiest), cpus);
3944 if (!cpumask_empty(cpus))
3951 schedstat_inc(sd, lb_failed[idle]);
3952 sd->nr_balance_failed++;
3954 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3956 spin_lock_irqsave(&busiest->lock, flags);
3958 /* don't kick the migration_thread, if the curr
3959 * task on busiest cpu can't be moved to this_cpu
3961 if (!cpumask_test_cpu(this_cpu,
3962 &busiest->curr->cpus_allowed)) {
3963 spin_unlock_irqrestore(&busiest->lock, flags);
3965 goto out_one_pinned;
3968 if (!busiest->active_balance) {
3969 busiest->active_balance = 1;
3970 busiest->push_cpu = this_cpu;
3973 spin_unlock_irqrestore(&busiest->lock, flags);
3975 wake_up_process(busiest->migration_thread);
3978 * We've kicked active balancing, reset the failure
3981 sd->nr_balance_failed = sd->cache_nice_tries+1;
3984 sd->nr_balance_failed = 0;
3986 if (likely(!active_balance)) {
3987 /* We were unbalanced, so reset the balancing interval */
3988 sd->balance_interval = sd->min_interval;
3991 * If we've begun active balancing, start to back off. This
3992 * case may not be covered by the all_pinned logic if there
3993 * is only 1 task on the busy runqueue (because we don't call
3996 if (sd->balance_interval < sd->max_interval)
3997 sd->balance_interval *= 2;
4000 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4001 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4007 schedstat_inc(sd, lb_balanced[idle]);
4009 sd->nr_balance_failed = 0;
4012 /* tune up the balancing interval */
4013 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4014 (sd->balance_interval < sd->max_interval))
4015 sd->balance_interval *= 2;
4017 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4018 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4029 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4030 * tasks if there is an imbalance.
4032 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4033 * this_rq is locked.
4036 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4038 struct sched_group *group;
4039 struct rq *busiest = NULL;
4040 unsigned long imbalance;
4044 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4046 cpumask_setall(cpus);
4049 * When power savings policy is enabled for the parent domain, idle
4050 * sibling can pick up load irrespective of busy siblings. In this case,
4051 * let the state of idle sibling percolate up as IDLE, instead of
4052 * portraying it as CPU_NOT_IDLE.
4054 if (sd->flags & SD_SHARE_CPUPOWER &&
4055 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4058 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4060 update_shares_locked(this_rq, sd);
4061 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4062 &sd_idle, cpus, NULL);
4064 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4068 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4070 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4074 BUG_ON(busiest == this_rq);
4076 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4079 if (busiest->nr_running > 1) {
4080 /* Attempt to move tasks */
4081 double_lock_balance(this_rq, busiest);
4082 /* this_rq->clock is already updated */
4083 update_rq_clock(busiest);
4084 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4085 imbalance, sd, CPU_NEWLY_IDLE,
4087 double_unlock_balance(this_rq, busiest);
4089 if (unlikely(all_pinned)) {
4090 cpumask_clear_cpu(cpu_of(busiest), cpus);
4091 if (!cpumask_empty(cpus))
4097 int active_balance = 0;
4099 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4100 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4101 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4104 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4107 if (sd->nr_balance_failed++ < 2)
4111 * The only task running in a non-idle cpu can be moved to this
4112 * cpu in an attempt to completely freeup the other CPU
4113 * package. The same method used to move task in load_balance()
4114 * have been extended for load_balance_newidle() to speedup
4115 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4117 * The package power saving logic comes from
4118 * find_busiest_group(). If there are no imbalance, then
4119 * f_b_g() will return NULL. However when sched_mc={1,2} then
4120 * f_b_g() will select a group from which a running task may be
4121 * pulled to this cpu in order to make the other package idle.
4122 * If there is no opportunity to make a package idle and if
4123 * there are no imbalance, then f_b_g() will return NULL and no
4124 * action will be taken in load_balance_newidle().
4126 * Under normal task pull operation due to imbalance, there
4127 * will be more than one task in the source run queue and
4128 * move_tasks() will succeed. ld_moved will be true and this
4129 * active balance code will not be triggered.
4132 /* Lock busiest in correct order while this_rq is held */
4133 double_lock_balance(this_rq, busiest);
4136 * don't kick the migration_thread, if the curr
4137 * task on busiest cpu can't be moved to this_cpu
4139 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4140 double_unlock_balance(this_rq, busiest);
4145 if (!busiest->active_balance) {
4146 busiest->active_balance = 1;
4147 busiest->push_cpu = this_cpu;
4151 double_unlock_balance(this_rq, busiest);
4153 * Should not call ttwu while holding a rq->lock
4155 spin_unlock(&this_rq->lock);
4157 wake_up_process(busiest->migration_thread);
4158 spin_lock(&this_rq->lock);
4161 sd->nr_balance_failed = 0;
4163 update_shares_locked(this_rq, sd);
4167 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4168 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4169 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4171 sd->nr_balance_failed = 0;
4177 * idle_balance is called by schedule() if this_cpu is about to become
4178 * idle. Attempts to pull tasks from other CPUs.
4180 static void idle_balance(int this_cpu, struct rq *this_rq)
4182 struct sched_domain *sd;
4183 int pulled_task = 0;
4184 unsigned long next_balance = jiffies + HZ;
4186 for_each_domain(this_cpu, sd) {
4187 unsigned long interval;
4189 if (!(sd->flags & SD_LOAD_BALANCE))
4192 if (sd->flags & SD_BALANCE_NEWIDLE)
4193 /* If we've pulled tasks over stop searching: */
4194 pulled_task = load_balance_newidle(this_cpu, this_rq,
4197 interval = msecs_to_jiffies(sd->balance_interval);
4198 if (time_after(next_balance, sd->last_balance + interval))
4199 next_balance = sd->last_balance + interval;
4203 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4205 * We are going idle. next_balance may be set based on
4206 * a busy processor. So reset next_balance.
4208 this_rq->next_balance = next_balance;
4213 * active_load_balance is run by migration threads. It pushes running tasks
4214 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4215 * running on each physical CPU where possible, and avoids physical /
4216 * logical imbalances.
4218 * Called with busiest_rq locked.
4220 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4222 int target_cpu = busiest_rq->push_cpu;
4223 struct sched_domain *sd;
4224 struct rq *target_rq;
4226 /* Is there any task to move? */
4227 if (busiest_rq->nr_running <= 1)
4230 target_rq = cpu_rq(target_cpu);
4233 * This condition is "impossible", if it occurs
4234 * we need to fix it. Originally reported by
4235 * Bjorn Helgaas on a 128-cpu setup.
4237 BUG_ON(busiest_rq == target_rq);
4239 /* move a task from busiest_rq to target_rq */
4240 double_lock_balance(busiest_rq, target_rq);
4241 update_rq_clock(busiest_rq);
4242 update_rq_clock(target_rq);
4244 /* Search for an sd spanning us and the target CPU. */
4245 for_each_domain(target_cpu, sd) {
4246 if ((sd->flags & SD_LOAD_BALANCE) &&
4247 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4252 schedstat_inc(sd, alb_count);
4254 if (move_one_task(target_rq, target_cpu, busiest_rq,
4256 schedstat_inc(sd, alb_pushed);
4258 schedstat_inc(sd, alb_failed);
4260 double_unlock_balance(busiest_rq, target_rq);
4265 atomic_t load_balancer;
4266 cpumask_var_t cpu_mask;
4267 } nohz ____cacheline_aligned = {
4268 .load_balancer = ATOMIC_INIT(-1),
4272 * This routine will try to nominate the ilb (idle load balancing)
4273 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4274 * load balancing on behalf of all those cpus. If all the cpus in the system
4275 * go into this tickless mode, then there will be no ilb owner (as there is
4276 * no need for one) and all the cpus will sleep till the next wakeup event
4279 * For the ilb owner, tick is not stopped. And this tick will be used
4280 * for idle load balancing. ilb owner will still be part of
4283 * While stopping the tick, this cpu will become the ilb owner if there
4284 * is no other owner. And will be the owner till that cpu becomes busy
4285 * or if all cpus in the system stop their ticks at which point
4286 * there is no need for ilb owner.
4288 * When the ilb owner becomes busy, it nominates another owner, during the
4289 * next busy scheduler_tick()
4291 int select_nohz_load_balancer(int stop_tick)
4293 int cpu = smp_processor_id();
4296 cpu_rq(cpu)->in_nohz_recently = 1;
4298 if (!cpu_active(cpu)) {
4299 if (atomic_read(&nohz.load_balancer) != cpu)
4303 * If we are going offline and still the leader,
4306 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4312 cpumask_set_cpu(cpu, nohz.cpu_mask);
4314 /* time for ilb owner also to sleep */
4315 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4316 if (atomic_read(&nohz.load_balancer) == cpu)
4317 atomic_set(&nohz.load_balancer, -1);
4321 if (atomic_read(&nohz.load_balancer) == -1) {
4322 /* make me the ilb owner */
4323 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4325 } else if (atomic_read(&nohz.load_balancer) == cpu)
4328 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4331 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4333 if (atomic_read(&nohz.load_balancer) == cpu)
4334 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4341 static DEFINE_SPINLOCK(balancing);
4344 * It checks each scheduling domain to see if it is due to be balanced,
4345 * and initiates a balancing operation if so.
4347 * Balancing parameters are set up in arch_init_sched_domains.
4349 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4352 struct rq *rq = cpu_rq(cpu);
4353 unsigned long interval;
4354 struct sched_domain *sd;
4355 /* Earliest time when we have to do rebalance again */
4356 unsigned long next_balance = jiffies + 60*HZ;
4357 int update_next_balance = 0;
4360 for_each_domain(cpu, sd) {
4361 if (!(sd->flags & SD_LOAD_BALANCE))
4364 interval = sd->balance_interval;
4365 if (idle != CPU_IDLE)
4366 interval *= sd->busy_factor;
4368 /* scale ms to jiffies */
4369 interval = msecs_to_jiffies(interval);
4370 if (unlikely(!interval))
4372 if (interval > HZ*NR_CPUS/10)
4373 interval = HZ*NR_CPUS/10;
4375 need_serialize = sd->flags & SD_SERIALIZE;
4377 if (need_serialize) {
4378 if (!spin_trylock(&balancing))
4382 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4383 if (load_balance(cpu, rq, sd, idle, &balance)) {
4385 * We've pulled tasks over so either we're no
4386 * longer idle, or one of our SMT siblings is
4389 idle = CPU_NOT_IDLE;
4391 sd->last_balance = jiffies;
4394 spin_unlock(&balancing);
4396 if (time_after(next_balance, sd->last_balance + interval)) {
4397 next_balance = sd->last_balance + interval;
4398 update_next_balance = 1;
4402 * Stop the load balance at this level. There is another
4403 * CPU in our sched group which is doing load balancing more
4411 * next_balance will be updated only when there is a need.
4412 * When the cpu is attached to null domain for ex, it will not be
4415 if (likely(update_next_balance))
4416 rq->next_balance = next_balance;
4420 * run_rebalance_domains is triggered when needed from the scheduler tick.
4421 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4422 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4424 static void run_rebalance_domains(struct softirq_action *h)
4426 int this_cpu = smp_processor_id();
4427 struct rq *this_rq = cpu_rq(this_cpu);
4428 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4429 CPU_IDLE : CPU_NOT_IDLE;
4431 rebalance_domains(this_cpu, idle);
4435 * If this cpu is the owner for idle load balancing, then do the
4436 * balancing on behalf of the other idle cpus whose ticks are
4439 if (this_rq->idle_at_tick &&
4440 atomic_read(&nohz.load_balancer) == this_cpu) {
4444 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4445 if (balance_cpu == this_cpu)
4449 * If this cpu gets work to do, stop the load balancing
4450 * work being done for other cpus. Next load
4451 * balancing owner will pick it up.
4456 rebalance_domains(balance_cpu, CPU_IDLE);
4458 rq = cpu_rq(balance_cpu);
4459 if (time_after(this_rq->next_balance, rq->next_balance))
4460 this_rq->next_balance = rq->next_balance;
4466 static inline int on_null_domain(int cpu)
4468 return !rcu_dereference(cpu_rq(cpu)->sd);
4472 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4474 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4475 * idle load balancing owner or decide to stop the periodic load balancing,
4476 * if the whole system is idle.
4478 static inline void trigger_load_balance(struct rq *rq, int cpu)
4482 * If we were in the nohz mode recently and busy at the current
4483 * scheduler tick, then check if we need to nominate new idle
4486 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4487 rq->in_nohz_recently = 0;
4489 if (atomic_read(&nohz.load_balancer) == cpu) {
4490 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4491 atomic_set(&nohz.load_balancer, -1);
4494 if (atomic_read(&nohz.load_balancer) == -1) {
4496 * simple selection for now: Nominate the
4497 * first cpu in the nohz list to be the next
4500 * TBD: Traverse the sched domains and nominate
4501 * the nearest cpu in the nohz.cpu_mask.
4503 int ilb = cpumask_first(nohz.cpu_mask);
4505 if (ilb < nr_cpu_ids)
4511 * If this cpu is idle and doing idle load balancing for all the
4512 * cpus with ticks stopped, is it time for that to stop?
4514 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4515 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4521 * If this cpu is idle and the idle load balancing is done by
4522 * someone else, then no need raise the SCHED_SOFTIRQ
4524 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4525 cpumask_test_cpu(cpu, nohz.cpu_mask))
4528 /* Don't need to rebalance while attached to NULL domain */
4529 if (time_after_eq(jiffies, rq->next_balance) &&
4530 likely(!on_null_domain(cpu)))
4531 raise_softirq(SCHED_SOFTIRQ);
4534 #else /* CONFIG_SMP */
4537 * on UP we do not need to balance between CPUs:
4539 static inline void idle_balance(int cpu, struct rq *rq)
4545 DEFINE_PER_CPU(struct kernel_stat, kstat);
4547 EXPORT_PER_CPU_SYMBOL(kstat);
4550 * Return any ns on the sched_clock that have not yet been banked in
4551 * @p in case that task is currently running.
4553 unsigned long long __task_delta_exec(struct task_struct *p, int update)
4559 WARN_ON_ONCE(!runqueue_is_locked());
4560 WARN_ON_ONCE(!task_current(rq, p));
4563 update_rq_clock(rq);
4565 delta_exec = rq->clock - p->se.exec_start;
4567 WARN_ON_ONCE(delta_exec < 0);
4573 * Return any ns on the sched_clock that have not yet been banked in
4574 * @p in case that task is currently running.
4576 unsigned long long task_delta_exec(struct task_struct *p)
4578 unsigned long flags;
4582 rq = task_rq_lock(p, &flags);
4584 if (task_current(rq, p)) {
4587 update_rq_clock(rq);
4588 delta_exec = rq->clock - p->se.exec_start;
4589 if ((s64)delta_exec > 0)
4593 task_rq_unlock(rq, &flags);
4599 * Account user cpu time to a process.
4600 * @p: the process that the cpu time gets accounted to
4601 * @cputime: the cpu time spent in user space since the last update
4602 * @cputime_scaled: cputime scaled by cpu frequency
4604 void account_user_time(struct task_struct *p, cputime_t cputime,
4605 cputime_t cputime_scaled)
4607 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4610 /* Add user time to process. */
4611 p->utime = cputime_add(p->utime, cputime);
4612 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4613 account_group_user_time(p, cputime);
4615 /* Add user time to cpustat. */
4616 tmp = cputime_to_cputime64(cputime);
4617 if (TASK_NICE(p) > 0)
4618 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4620 cpustat->user = cputime64_add(cpustat->user, tmp);
4621 /* Account for user time used */
4622 acct_update_integrals(p);
4626 * Account guest cpu time to a process.
4627 * @p: the process that the cpu time gets accounted to
4628 * @cputime: the cpu time spent in virtual machine since the last update
4629 * @cputime_scaled: cputime scaled by cpu frequency
4631 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4632 cputime_t cputime_scaled)
4635 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4637 tmp = cputime_to_cputime64(cputime);
4639 /* Add guest time to process. */
4640 p->utime = cputime_add(p->utime, cputime);
4641 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4642 account_group_user_time(p, cputime);
4643 p->gtime = cputime_add(p->gtime, cputime);
4645 /* Add guest time to cpustat. */
4646 cpustat->user = cputime64_add(cpustat->user, tmp);
4647 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4651 * Account system cpu time to a process.
4652 * @p: the process that the cpu time gets accounted to
4653 * @hardirq_offset: the offset to subtract from hardirq_count()
4654 * @cputime: the cpu time spent in kernel space since the last update
4655 * @cputime_scaled: cputime scaled by cpu frequency
4657 void account_system_time(struct task_struct *p, int hardirq_offset,
4658 cputime_t cputime, cputime_t cputime_scaled)
4660 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4663 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4664 account_guest_time(p, cputime, cputime_scaled);
4668 /* Add system time to process. */
4669 p->stime = cputime_add(p->stime, cputime);
4670 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4671 account_group_system_time(p, cputime);
4673 /* Add system time to cpustat. */
4674 tmp = cputime_to_cputime64(cputime);
4675 if (hardirq_count() - hardirq_offset)
4676 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4677 else if (softirq_count())
4678 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4680 cpustat->system = cputime64_add(cpustat->system, tmp);
4682 /* Account for system time used */
4683 acct_update_integrals(p);
4687 * Account for involuntary wait time.
4688 * @steal: the cpu time spent in involuntary wait
4690 void account_steal_time(cputime_t cputime)
4692 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4693 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4695 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4699 * Account for idle time.
4700 * @cputime: the cpu time spent in idle wait
4702 void account_idle_time(cputime_t cputime)
4704 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4705 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4706 struct rq *rq = this_rq();
4708 if (atomic_read(&rq->nr_iowait) > 0)
4709 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4711 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4714 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4717 * Account a single tick of cpu time.
4718 * @p: the process that the cpu time gets accounted to
4719 * @user_tick: indicates if the tick is a user or a system tick
4721 void account_process_tick(struct task_struct *p, int user_tick)
4723 cputime_t one_jiffy = jiffies_to_cputime(1);
4724 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
4725 struct rq *rq = this_rq();
4728 account_user_time(p, one_jiffy, one_jiffy_scaled);
4729 else if (p != rq->idle)
4730 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
4733 account_idle_time(one_jiffy);
4737 * Account multiple ticks of steal time.
4738 * @p: the process from which the cpu time has been stolen
4739 * @ticks: number of stolen ticks
4741 void account_steal_ticks(unsigned long ticks)
4743 account_steal_time(jiffies_to_cputime(ticks));
4747 * Account multiple ticks of idle time.
4748 * @ticks: number of stolen ticks
4750 void account_idle_ticks(unsigned long ticks)
4752 account_idle_time(jiffies_to_cputime(ticks));
4758 * Use precise platform statistics if available:
4760 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4761 cputime_t task_utime(struct task_struct *p)
4766 cputime_t task_stime(struct task_struct *p)
4771 cputime_t task_utime(struct task_struct *p)
4773 clock_t utime = cputime_to_clock_t(p->utime),
4774 total = utime + cputime_to_clock_t(p->stime);
4778 * Use CFS's precise accounting:
4780 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4784 do_div(temp, total);
4786 utime = (clock_t)temp;
4788 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4789 return p->prev_utime;
4792 cputime_t task_stime(struct task_struct *p)
4797 * Use CFS's precise accounting. (we subtract utime from
4798 * the total, to make sure the total observed by userspace
4799 * grows monotonically - apps rely on that):
4801 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4802 cputime_to_clock_t(task_utime(p));
4805 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4807 return p->prev_stime;
4811 inline cputime_t task_gtime(struct task_struct *p)
4817 * This function gets called by the timer code, with HZ frequency.
4818 * We call it with interrupts disabled.
4820 * It also gets called by the fork code, when changing the parent's
4823 void scheduler_tick(void)
4825 int cpu = smp_processor_id();
4826 struct rq *rq = cpu_rq(cpu);
4827 struct task_struct *curr = rq->curr;
4831 spin_lock(&rq->lock);
4832 update_rq_clock(rq);
4833 update_cpu_load(rq);
4834 curr->sched_class->task_tick(rq, curr, 0);
4835 perf_counter_task_tick(curr, cpu);
4836 spin_unlock(&rq->lock);
4839 rq->idle_at_tick = idle_cpu(cpu);
4840 trigger_load_balance(rq, cpu);
4844 unsigned long get_parent_ip(unsigned long addr)
4846 if (in_lock_functions(addr)) {
4847 addr = CALLER_ADDR2;
4848 if (in_lock_functions(addr))
4849 addr = CALLER_ADDR3;
4854 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4855 defined(CONFIG_PREEMPT_TRACER))
4857 void __kprobes add_preempt_count(int val)
4859 #ifdef CONFIG_DEBUG_PREEMPT
4863 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4866 preempt_count() += val;
4867 #ifdef CONFIG_DEBUG_PREEMPT
4869 * Spinlock count overflowing soon?
4871 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4874 if (preempt_count() == val)
4875 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4877 EXPORT_SYMBOL(add_preempt_count);
4879 void __kprobes sub_preempt_count(int val)
4881 #ifdef CONFIG_DEBUG_PREEMPT
4885 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4888 * Is the spinlock portion underflowing?
4890 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4891 !(preempt_count() & PREEMPT_MASK)))
4895 if (preempt_count() == val)
4896 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4897 preempt_count() -= val;
4899 EXPORT_SYMBOL(sub_preempt_count);
4904 * Print scheduling while atomic bug:
4906 static noinline void __schedule_bug(struct task_struct *prev)
4908 struct pt_regs *regs = get_irq_regs();
4910 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4911 prev->comm, prev->pid, preempt_count());
4913 debug_show_held_locks(prev);
4915 if (irqs_disabled())
4916 print_irqtrace_events(prev);
4925 * Various schedule()-time debugging checks and statistics:
4927 static inline void schedule_debug(struct task_struct *prev)
4930 * Test if we are atomic. Since do_exit() needs to call into
4931 * schedule() atomically, we ignore that path for now.
4932 * Otherwise, whine if we are scheduling when we should not be.
4934 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4935 __schedule_bug(prev);
4937 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4939 schedstat_inc(this_rq(), sched_count);
4940 #ifdef CONFIG_SCHEDSTATS
4941 if (unlikely(prev->lock_depth >= 0)) {
4942 schedstat_inc(this_rq(), bkl_count);
4943 schedstat_inc(prev, sched_info.bkl_count);
4948 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4950 if (prev->state == TASK_RUNNING) {
4951 u64 runtime = prev->se.sum_exec_runtime;
4953 runtime -= prev->se.prev_sum_exec_runtime;
4954 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
4957 * In order to avoid avg_overlap growing stale when we are
4958 * indeed overlapping and hence not getting put to sleep, grow
4959 * the avg_overlap on preemption.
4961 * We use the average preemption runtime because that
4962 * correlates to the amount of cache footprint a task can
4965 update_avg(&prev->se.avg_overlap, runtime);
4967 prev->sched_class->put_prev_task(rq, prev);
4971 * Pick up the highest-prio task:
4973 static inline struct task_struct *
4974 pick_next_task(struct rq *rq)
4976 const struct sched_class *class;
4977 struct task_struct *p;
4980 * Optimization: we know that if all tasks are in
4981 * the fair class we can call that function directly:
4983 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4984 p = fair_sched_class.pick_next_task(rq);
4989 class = sched_class_highest;
4991 p = class->pick_next_task(rq);
4995 * Will never be NULL as the idle class always
4996 * returns a non-NULL p:
4998 class = class->next;
5003 * schedule() is the main scheduler function.
5005 asmlinkage void __sched __schedule(void)
5007 struct task_struct *prev, *next;
5008 unsigned long *switch_count;
5012 cpu = smp_processor_id();
5016 switch_count = &prev->nivcsw;
5018 release_kernel_lock(prev);
5019 need_resched_nonpreemptible:
5021 schedule_debug(prev);
5023 if (sched_feat(HRTICK))
5026 spin_lock_irq(&rq->lock);
5027 update_rq_clock(rq);
5028 clear_tsk_need_resched(prev);
5030 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5031 if (unlikely(signal_pending_state(prev->state, prev)))
5032 prev->state = TASK_RUNNING;
5034 deactivate_task(rq, prev, 1);
5035 switch_count = &prev->nvcsw;
5039 if (prev->sched_class->pre_schedule)
5040 prev->sched_class->pre_schedule(rq, prev);
5043 if (unlikely(!rq->nr_running))
5044 idle_balance(cpu, rq);
5046 put_prev_task(rq, prev);
5047 next = pick_next_task(rq);
5049 if (likely(prev != next)) {
5050 sched_info_switch(prev, next);
5051 perf_counter_task_sched_out(prev, cpu);
5057 context_switch(rq, prev, next); /* unlocks the rq */
5059 * the context switch might have flipped the stack from under
5060 * us, hence refresh the local variables.
5062 cpu = smp_processor_id();
5065 spin_unlock_irq(&rq->lock);
5067 if (unlikely(reacquire_kernel_lock(current) < 0))
5068 goto need_resched_nonpreemptible;
5071 asmlinkage void __sched schedule(void)
5076 preempt_enable_no_resched();
5077 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
5080 EXPORT_SYMBOL(schedule);
5084 * Look out! "owner" is an entirely speculative pointer
5085 * access and not reliable.
5087 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5092 if (!sched_feat(OWNER_SPIN))
5095 #ifdef CONFIG_DEBUG_PAGEALLOC
5097 * Need to access the cpu field knowing that
5098 * DEBUG_PAGEALLOC could have unmapped it if
5099 * the mutex owner just released it and exited.
5101 if (probe_kernel_address(&owner->cpu, cpu))
5108 * Even if the access succeeded (likely case),
5109 * the cpu field may no longer be valid.
5111 if (cpu >= nr_cpumask_bits)
5115 * We need to validate that we can do a
5116 * get_cpu() and that we have the percpu area.
5118 if (!cpu_online(cpu))
5125 * Owner changed, break to re-assess state.
5127 if (lock->owner != owner)
5131 * Is that owner really running on that cpu?
5133 if (task_thread_info(rq->curr) != owner || need_resched())
5143 #ifdef CONFIG_PREEMPT
5145 * this is the entry point to schedule() from in-kernel preemption
5146 * off of preempt_enable. Kernel preemptions off return from interrupt
5147 * occur there and call schedule directly.
5149 asmlinkage void __sched preempt_schedule(void)
5151 struct thread_info *ti = current_thread_info();
5154 * If there is a non-zero preempt_count or interrupts are disabled,
5155 * we do not want to preempt the current task. Just return..
5157 if (likely(ti->preempt_count || irqs_disabled()))
5161 add_preempt_count(PREEMPT_ACTIVE);
5163 sub_preempt_count(PREEMPT_ACTIVE);
5166 * Check again in case we missed a preemption opportunity
5167 * between schedule and now.
5170 } while (need_resched());
5172 EXPORT_SYMBOL(preempt_schedule);
5175 * this is the entry point to schedule() from kernel preemption
5176 * off of irq context.
5177 * Note, that this is called and return with irqs disabled. This will
5178 * protect us against recursive calling from irq.
5180 asmlinkage void __sched preempt_schedule_irq(void)
5182 struct thread_info *ti = current_thread_info();
5184 /* Catch callers which need to be fixed */
5185 BUG_ON(ti->preempt_count || !irqs_disabled());
5188 add_preempt_count(PREEMPT_ACTIVE);
5191 local_irq_disable();
5192 sub_preempt_count(PREEMPT_ACTIVE);
5195 * Check again in case we missed a preemption opportunity
5196 * between schedule and now.
5199 } while (need_resched());
5202 #endif /* CONFIG_PREEMPT */
5204 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5207 return try_to_wake_up(curr->private, mode, sync);
5209 EXPORT_SYMBOL(default_wake_function);
5212 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5213 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5214 * number) then we wake all the non-exclusive tasks and one exclusive task.
5216 * There are circumstances in which we can try to wake a task which has already
5217 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5218 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5220 void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5221 int nr_exclusive, int sync, void *key)
5223 wait_queue_t *curr, *next;
5225 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5226 unsigned flags = curr->flags;
5228 if (curr->func(curr, mode, sync, key) &&
5229 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5235 * __wake_up - wake up threads blocked on a waitqueue.
5237 * @mode: which threads
5238 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5239 * @key: is directly passed to the wakeup function
5241 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5242 int nr_exclusive, void *key)
5244 unsigned long flags;
5246 spin_lock_irqsave(&q->lock, flags);
5247 __wake_up_common(q, mode, nr_exclusive, 0, key);
5248 spin_unlock_irqrestore(&q->lock, flags);
5250 EXPORT_SYMBOL(__wake_up);
5253 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5255 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5257 __wake_up_common(q, mode, 1, 0, NULL);
5260 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5262 __wake_up_common(q, mode, 1, 0, key);
5266 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5268 * @mode: which threads
5269 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5270 * @key: opaque value to be passed to wakeup targets
5272 * The sync wakeup differs that the waker knows that it will schedule
5273 * away soon, so while the target thread will be woken up, it will not
5274 * be migrated to another CPU - ie. the two threads are 'synchronized'
5275 * with each other. This can prevent needless bouncing between CPUs.
5277 * On UP it can prevent extra preemption.
5279 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5280 int nr_exclusive, void *key)
5282 unsigned long flags;
5288 if (unlikely(!nr_exclusive))
5291 spin_lock_irqsave(&q->lock, flags);
5292 __wake_up_common(q, mode, nr_exclusive, sync, key);
5293 spin_unlock_irqrestore(&q->lock, flags);
5295 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5298 * __wake_up_sync - see __wake_up_sync_key()
5300 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5302 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5304 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5307 * complete: - signals a single thread waiting on this completion
5308 * @x: holds the state of this particular completion
5310 * This will wake up a single thread waiting on this completion. Threads will be
5311 * awakened in the same order in which they were queued.
5313 * See also complete_all(), wait_for_completion() and related routines.
5315 void complete(struct completion *x)
5317 unsigned long flags;
5319 spin_lock_irqsave(&x->wait.lock, flags);
5321 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5322 spin_unlock_irqrestore(&x->wait.lock, flags);
5324 EXPORT_SYMBOL(complete);
5327 * complete_all: - signals all threads waiting on this completion
5328 * @x: holds the state of this particular completion
5330 * This will wake up all threads waiting on this particular completion event.
5332 void complete_all(struct completion *x)
5334 unsigned long flags;
5336 spin_lock_irqsave(&x->wait.lock, flags);
5337 x->done += UINT_MAX/2;
5338 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5339 spin_unlock_irqrestore(&x->wait.lock, flags);
5341 EXPORT_SYMBOL(complete_all);
5343 static inline long __sched
5344 do_wait_for_common(struct completion *x, long timeout, int state)
5347 DECLARE_WAITQUEUE(wait, current);
5349 wait.flags |= WQ_FLAG_EXCLUSIVE;
5350 __add_wait_queue_tail(&x->wait, &wait);
5352 if (signal_pending_state(state, current)) {
5353 timeout = -ERESTARTSYS;
5356 __set_current_state(state);
5357 spin_unlock_irq(&x->wait.lock);
5358 timeout = schedule_timeout(timeout);
5359 spin_lock_irq(&x->wait.lock);
5360 } while (!x->done && timeout);
5361 __remove_wait_queue(&x->wait, &wait);
5366 return timeout ?: 1;
5370 wait_for_common(struct completion *x, long timeout, int state)
5374 spin_lock_irq(&x->wait.lock);
5375 timeout = do_wait_for_common(x, timeout, state);
5376 spin_unlock_irq(&x->wait.lock);
5381 * wait_for_completion: - waits for completion of a task
5382 * @x: holds the state of this particular completion
5384 * This waits to be signaled for completion of a specific task. It is NOT
5385 * interruptible and there is no timeout.
5387 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5388 * and interrupt capability. Also see complete().
5390 void __sched wait_for_completion(struct completion *x)
5392 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5394 EXPORT_SYMBOL(wait_for_completion);
5397 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5398 * @x: holds the state of this particular completion
5399 * @timeout: timeout value in jiffies
5401 * This waits for either a completion of a specific task to be signaled or for a
5402 * specified timeout to expire. The timeout is in jiffies. It is not
5405 unsigned long __sched
5406 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5408 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5410 EXPORT_SYMBOL(wait_for_completion_timeout);
5413 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5414 * @x: holds the state of this particular completion
5416 * This waits for completion of a specific task to be signaled. It is
5419 int __sched wait_for_completion_interruptible(struct completion *x)
5421 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5422 if (t == -ERESTARTSYS)
5426 EXPORT_SYMBOL(wait_for_completion_interruptible);
5429 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5430 * @x: holds the state of this particular completion
5431 * @timeout: timeout value in jiffies
5433 * This waits for either a completion of a specific task to be signaled or for a
5434 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5436 unsigned long __sched
5437 wait_for_completion_interruptible_timeout(struct completion *x,
5438 unsigned long timeout)
5440 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5442 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5445 * wait_for_completion_killable: - waits for completion of a task (killable)
5446 * @x: holds the state of this particular completion
5448 * This waits to be signaled for completion of a specific task. It can be
5449 * interrupted by a kill signal.
5451 int __sched wait_for_completion_killable(struct completion *x)
5453 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5454 if (t == -ERESTARTSYS)
5458 EXPORT_SYMBOL(wait_for_completion_killable);
5461 * try_wait_for_completion - try to decrement a completion without blocking
5462 * @x: completion structure
5464 * Returns: 0 if a decrement cannot be done without blocking
5465 * 1 if a decrement succeeded.
5467 * If a completion is being used as a counting completion,
5468 * attempt to decrement the counter without blocking. This
5469 * enables us to avoid waiting if the resource the completion
5470 * is protecting is not available.
5472 bool try_wait_for_completion(struct completion *x)
5476 spin_lock_irq(&x->wait.lock);
5481 spin_unlock_irq(&x->wait.lock);
5484 EXPORT_SYMBOL(try_wait_for_completion);
5487 * completion_done - Test to see if a completion has any waiters
5488 * @x: completion structure
5490 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5491 * 1 if there are no waiters.
5494 bool completion_done(struct completion *x)
5498 spin_lock_irq(&x->wait.lock);
5501 spin_unlock_irq(&x->wait.lock);
5504 EXPORT_SYMBOL(completion_done);
5507 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5509 unsigned long flags;
5512 init_waitqueue_entry(&wait, current);
5514 __set_current_state(state);
5516 spin_lock_irqsave(&q->lock, flags);
5517 __add_wait_queue(q, &wait);
5518 spin_unlock(&q->lock);
5519 timeout = schedule_timeout(timeout);
5520 spin_lock_irq(&q->lock);
5521 __remove_wait_queue(q, &wait);
5522 spin_unlock_irqrestore(&q->lock, flags);
5527 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5529 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5531 EXPORT_SYMBOL(interruptible_sleep_on);
5534 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5536 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5538 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5540 void __sched sleep_on(wait_queue_head_t *q)
5542 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5544 EXPORT_SYMBOL(sleep_on);
5546 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5548 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5550 EXPORT_SYMBOL(sleep_on_timeout);
5552 #ifdef CONFIG_RT_MUTEXES
5555 * rt_mutex_setprio - set the current priority of a task
5557 * @prio: prio value (kernel-internal form)
5559 * This function changes the 'effective' priority of a task. It does
5560 * not touch ->normal_prio like __setscheduler().
5562 * Used by the rt_mutex code to implement priority inheritance logic.
5564 void rt_mutex_setprio(struct task_struct *p, int prio)
5566 unsigned long flags;
5567 int oldprio, on_rq, running;
5569 const struct sched_class *prev_class = p->sched_class;
5571 BUG_ON(prio < 0 || prio > MAX_PRIO);
5573 rq = task_rq_lock(p, &flags);
5574 update_rq_clock(rq);
5577 on_rq = p->se.on_rq;
5578 running = task_current(rq, p);
5580 dequeue_task(rq, p, 0);
5582 p->sched_class->put_prev_task(rq, p);
5585 p->sched_class = &rt_sched_class;
5587 p->sched_class = &fair_sched_class;
5592 p->sched_class->set_curr_task(rq);
5594 enqueue_task(rq, p, 0);
5596 check_class_changed(rq, p, prev_class, oldprio, running);
5598 task_rq_unlock(rq, &flags);
5603 void set_user_nice(struct task_struct *p, long nice)
5605 int old_prio, delta, on_rq;
5606 unsigned long flags;
5609 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5612 * We have to be careful, if called from sys_setpriority(),
5613 * the task might be in the middle of scheduling on another CPU.
5615 rq = task_rq_lock(p, &flags);
5616 update_rq_clock(rq);
5618 * The RT priorities are set via sched_setscheduler(), but we still
5619 * allow the 'normal' nice value to be set - but as expected
5620 * it wont have any effect on scheduling until the task is
5621 * SCHED_FIFO/SCHED_RR:
5623 if (task_has_rt_policy(p)) {
5624 p->static_prio = NICE_TO_PRIO(nice);
5627 on_rq = p->se.on_rq;
5629 dequeue_task(rq, p, 0);
5631 p->static_prio = NICE_TO_PRIO(nice);
5634 p->prio = effective_prio(p);
5635 delta = p->prio - old_prio;
5638 enqueue_task(rq, p, 0);
5640 * If the task increased its priority or is running and
5641 * lowered its priority, then reschedule its CPU:
5643 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5644 resched_task(rq->curr);
5647 task_rq_unlock(rq, &flags);
5649 EXPORT_SYMBOL(set_user_nice);
5652 * can_nice - check if a task can reduce its nice value
5656 int can_nice(const struct task_struct *p, const int nice)
5658 /* convert nice value [19,-20] to rlimit style value [1,40] */
5659 int nice_rlim = 20 - nice;
5661 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5662 capable(CAP_SYS_NICE));
5665 #ifdef __ARCH_WANT_SYS_NICE
5668 * sys_nice - change the priority of the current process.
5669 * @increment: priority increment
5671 * sys_setpriority is a more generic, but much slower function that
5672 * does similar things.
5674 SYSCALL_DEFINE1(nice, int, increment)
5679 * Setpriority might change our priority at the same moment.
5680 * We don't have to worry. Conceptually one call occurs first
5681 * and we have a single winner.
5683 if (increment < -40)
5688 nice = TASK_NICE(current) + increment;
5694 if (increment < 0 && !can_nice(current, nice))
5697 retval = security_task_setnice(current, nice);
5701 set_user_nice(current, nice);
5708 * task_prio - return the priority value of a given task.
5709 * @p: the task in question.
5711 * This is the priority value as seen by users in /proc.
5712 * RT tasks are offset by -200. Normal tasks are centered
5713 * around 0, value goes from -16 to +15.
5715 int task_prio(const struct task_struct *p)
5717 return p->prio - MAX_RT_PRIO;
5721 * task_nice - return the nice value of a given task.
5722 * @p: the task in question.
5724 int task_nice(const struct task_struct *p)
5726 return TASK_NICE(p);
5728 EXPORT_SYMBOL(task_nice);
5731 * idle_cpu - is a given cpu idle currently?
5732 * @cpu: the processor in question.
5734 int idle_cpu(int cpu)
5736 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5740 * idle_task - return the idle task for a given cpu.
5741 * @cpu: the processor in question.
5743 struct task_struct *idle_task(int cpu)
5745 return cpu_rq(cpu)->idle;
5749 * find_process_by_pid - find a process with a matching PID value.
5750 * @pid: the pid in question.
5752 static struct task_struct *find_process_by_pid(pid_t pid)
5754 return pid ? find_task_by_vpid(pid) : current;
5757 /* Actually do priority change: must hold rq lock. */
5759 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5761 BUG_ON(p->se.on_rq);
5764 switch (p->policy) {
5768 p->sched_class = &fair_sched_class;
5772 p->sched_class = &rt_sched_class;
5776 p->rt_priority = prio;
5777 p->normal_prio = normal_prio(p);
5778 /* we are holding p->pi_lock already */
5779 p->prio = rt_mutex_getprio(p);
5784 * check the target process has a UID that matches the current process's
5786 static bool check_same_owner(struct task_struct *p)
5788 const struct cred *cred = current_cred(), *pcred;
5792 pcred = __task_cred(p);
5793 match = (cred->euid == pcred->euid ||
5794 cred->euid == pcred->uid);
5799 static int __sched_setscheduler(struct task_struct *p, int policy,
5800 struct sched_param *param, bool user)
5802 int retval, oldprio, oldpolicy = -1, on_rq, running;
5803 unsigned long flags;
5804 const struct sched_class *prev_class = p->sched_class;
5807 /* may grab non-irq protected spin_locks */
5808 BUG_ON(in_interrupt());
5810 /* double check policy once rq lock held */
5812 policy = oldpolicy = p->policy;
5813 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5814 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5815 policy != SCHED_IDLE)
5818 * Valid priorities for SCHED_FIFO and SCHED_RR are
5819 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5820 * SCHED_BATCH and SCHED_IDLE is 0.
5822 if (param->sched_priority < 0 ||
5823 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5824 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5826 if (rt_policy(policy) != (param->sched_priority != 0))
5830 * Allow unprivileged RT tasks to decrease priority:
5832 if (user && !capable(CAP_SYS_NICE)) {
5833 if (rt_policy(policy)) {
5834 unsigned long rlim_rtprio;
5836 if (!lock_task_sighand(p, &flags))
5838 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5839 unlock_task_sighand(p, &flags);
5841 /* can't set/change the rt policy */
5842 if (policy != p->policy && !rlim_rtprio)
5845 /* can't increase priority */
5846 if (param->sched_priority > p->rt_priority &&
5847 param->sched_priority > rlim_rtprio)
5851 * Like positive nice levels, dont allow tasks to
5852 * move out of SCHED_IDLE either:
5854 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5857 /* can't change other user's priorities */
5858 if (!check_same_owner(p))
5863 #ifdef CONFIG_RT_GROUP_SCHED
5865 * Do not allow realtime tasks into groups that have no runtime
5868 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5869 task_group(p)->rt_bandwidth.rt_runtime == 0)
5873 retval = security_task_setscheduler(p, policy, param);
5879 * make sure no PI-waiters arrive (or leave) while we are
5880 * changing the priority of the task:
5882 spin_lock_irqsave(&p->pi_lock, flags);
5884 * To be able to change p->policy safely, the apropriate
5885 * runqueue lock must be held.
5887 rq = __task_rq_lock(p);
5888 /* recheck policy now with rq lock held */
5889 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5890 policy = oldpolicy = -1;
5891 __task_rq_unlock(rq);
5892 spin_unlock_irqrestore(&p->pi_lock, flags);
5895 update_rq_clock(rq);
5896 on_rq = p->se.on_rq;
5897 running = task_current(rq, p);
5899 deactivate_task(rq, p, 0);
5901 p->sched_class->put_prev_task(rq, p);
5904 __setscheduler(rq, p, policy, param->sched_priority);
5907 p->sched_class->set_curr_task(rq);
5909 activate_task(rq, p, 0);
5911 check_class_changed(rq, p, prev_class, oldprio, running);
5913 __task_rq_unlock(rq);
5914 spin_unlock_irqrestore(&p->pi_lock, flags);
5916 rt_mutex_adjust_pi(p);
5922 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5923 * @p: the task in question.
5924 * @policy: new policy.
5925 * @param: structure containing the new RT priority.
5927 * NOTE that the task may be already dead.
5929 int sched_setscheduler(struct task_struct *p, int policy,
5930 struct sched_param *param)
5932 return __sched_setscheduler(p, policy, param, true);
5934 EXPORT_SYMBOL_GPL(sched_setscheduler);
5937 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5938 * @p: the task in question.
5939 * @policy: new policy.
5940 * @param: structure containing the new RT priority.
5942 * Just like sched_setscheduler, only don't bother checking if the
5943 * current context has permission. For example, this is needed in
5944 * stop_machine(): we create temporary high priority worker threads,
5945 * but our caller might not have that capability.
5947 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5948 struct sched_param *param)
5950 return __sched_setscheduler(p, policy, param, false);
5954 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5956 struct sched_param lparam;
5957 struct task_struct *p;
5960 if (!param || pid < 0)
5962 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5967 p = find_process_by_pid(pid);
5969 retval = sched_setscheduler(p, policy, &lparam);
5976 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5977 * @pid: the pid in question.
5978 * @policy: new policy.
5979 * @param: structure containing the new RT priority.
5981 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5982 struct sched_param __user *, param)
5984 /* negative values for policy are not valid */
5988 return do_sched_setscheduler(pid, policy, param);
5992 * sys_sched_setparam - set/change the RT priority of a thread
5993 * @pid: the pid in question.
5994 * @param: structure containing the new RT priority.
5996 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5998 return do_sched_setscheduler(pid, -1, param);
6002 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6003 * @pid: the pid in question.
6005 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6007 struct task_struct *p;
6014 read_lock(&tasklist_lock);
6015 p = find_process_by_pid(pid);
6017 retval = security_task_getscheduler(p);
6021 read_unlock(&tasklist_lock);
6026 * sys_sched_getscheduler - get the RT priority of a thread
6027 * @pid: the pid in question.
6028 * @param: structure containing the RT priority.
6030 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6032 struct sched_param lp;
6033 struct task_struct *p;
6036 if (!param || pid < 0)
6039 read_lock(&tasklist_lock);
6040 p = find_process_by_pid(pid);
6045 retval = security_task_getscheduler(p);
6049 lp.sched_priority = p->rt_priority;
6050 read_unlock(&tasklist_lock);
6053 * This one might sleep, we cannot do it with a spinlock held ...
6055 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6060 read_unlock(&tasklist_lock);
6064 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6066 cpumask_var_t cpus_allowed, new_mask;
6067 struct task_struct *p;
6071 read_lock(&tasklist_lock);
6073 p = find_process_by_pid(pid);
6075 read_unlock(&tasklist_lock);
6081 * It is not safe to call set_cpus_allowed with the
6082 * tasklist_lock held. We will bump the task_struct's
6083 * usage count and then drop tasklist_lock.
6086 read_unlock(&tasklist_lock);
6088 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6092 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6094 goto out_free_cpus_allowed;
6097 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6100 retval = security_task_setscheduler(p, 0, NULL);
6104 cpuset_cpus_allowed(p, cpus_allowed);
6105 cpumask_and(new_mask, in_mask, cpus_allowed);
6107 retval = set_cpus_allowed_ptr(p, new_mask);
6110 cpuset_cpus_allowed(p, cpus_allowed);
6111 if (!cpumask_subset(new_mask, cpus_allowed)) {
6113 * We must have raced with a concurrent cpuset
6114 * update. Just reset the cpus_allowed to the
6115 * cpuset's cpus_allowed
6117 cpumask_copy(new_mask, cpus_allowed);
6122 free_cpumask_var(new_mask);
6123 out_free_cpus_allowed:
6124 free_cpumask_var(cpus_allowed);
6131 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6132 struct cpumask *new_mask)
6134 if (len < cpumask_size())
6135 cpumask_clear(new_mask);
6136 else if (len > cpumask_size())
6137 len = cpumask_size();
6139 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6143 * sys_sched_setaffinity - set the cpu affinity of a process
6144 * @pid: pid of the process
6145 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6146 * @user_mask_ptr: user-space pointer to the new cpu mask
6148 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6149 unsigned long __user *, user_mask_ptr)
6151 cpumask_var_t new_mask;
6154 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6157 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6159 retval = sched_setaffinity(pid, new_mask);
6160 free_cpumask_var(new_mask);
6164 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6166 struct task_struct *p;
6170 read_lock(&tasklist_lock);
6173 p = find_process_by_pid(pid);
6177 retval = security_task_getscheduler(p);
6181 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6184 read_unlock(&tasklist_lock);
6191 * sys_sched_getaffinity - get the cpu affinity of a process
6192 * @pid: pid of the process
6193 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6194 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6196 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6197 unsigned long __user *, user_mask_ptr)
6202 if (len < cpumask_size())
6205 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6208 ret = sched_getaffinity(pid, mask);
6210 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6213 ret = cpumask_size();
6215 free_cpumask_var(mask);
6221 * sys_sched_yield - yield the current processor to other threads.
6223 * This function yields the current CPU to other tasks. If there are no
6224 * other threads running on this CPU then this function will return.
6226 SYSCALL_DEFINE0(sched_yield)
6228 struct rq *rq = this_rq_lock();
6230 schedstat_inc(rq, yld_count);
6231 current->sched_class->yield_task(rq);
6234 * Since we are going to call schedule() anyway, there's
6235 * no need to preempt or enable interrupts:
6237 __release(rq->lock);
6238 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6239 _raw_spin_unlock(&rq->lock);
6240 preempt_enable_no_resched();
6247 static void __cond_resched(void)
6249 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6250 __might_sleep(__FILE__, __LINE__);
6253 * The BKS might be reacquired before we have dropped
6254 * PREEMPT_ACTIVE, which could trigger a second
6255 * cond_resched() call.
6258 add_preempt_count(PREEMPT_ACTIVE);
6260 sub_preempt_count(PREEMPT_ACTIVE);
6261 } while (need_resched());
6264 int __sched _cond_resched(void)
6266 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
6267 system_state == SYSTEM_RUNNING) {
6273 EXPORT_SYMBOL(_cond_resched);
6276 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6277 * call schedule, and on return reacquire the lock.
6279 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6280 * operations here to prevent schedule() from being called twice (once via
6281 * spin_unlock(), once by hand).
6283 int cond_resched_lock(spinlock_t *lock)
6285 int resched = need_resched() && system_state == SYSTEM_RUNNING;
6288 if (spin_needbreak(lock) || resched) {
6290 if (resched && need_resched())
6299 EXPORT_SYMBOL(cond_resched_lock);
6301 int __sched cond_resched_softirq(void)
6303 BUG_ON(!in_softirq());
6305 if (need_resched() && system_state == SYSTEM_RUNNING) {
6313 EXPORT_SYMBOL(cond_resched_softirq);
6316 * yield - yield the current processor to other threads.
6318 * This is a shortcut for kernel-space yielding - it marks the
6319 * thread runnable and calls sys_sched_yield().
6321 void __sched yield(void)
6323 set_current_state(TASK_RUNNING);
6326 EXPORT_SYMBOL(yield);
6329 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6330 * that process accounting knows that this is a task in IO wait state.
6332 * But don't do that if it is a deliberate, throttling IO wait (this task
6333 * has set its backing_dev_info: the queue against which it should throttle)
6335 void __sched io_schedule(void)
6337 struct rq *rq = &__raw_get_cpu_var(runqueues);
6339 delayacct_blkio_start();
6340 atomic_inc(&rq->nr_iowait);
6342 atomic_dec(&rq->nr_iowait);
6343 delayacct_blkio_end();
6345 EXPORT_SYMBOL(io_schedule);
6347 long __sched io_schedule_timeout(long timeout)
6349 struct rq *rq = &__raw_get_cpu_var(runqueues);
6352 delayacct_blkio_start();
6353 atomic_inc(&rq->nr_iowait);
6354 ret = schedule_timeout(timeout);
6355 atomic_dec(&rq->nr_iowait);
6356 delayacct_blkio_end();
6361 * sys_sched_get_priority_max - return maximum RT priority.
6362 * @policy: scheduling class.
6364 * this syscall returns the maximum rt_priority that can be used
6365 * by a given scheduling class.
6367 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6374 ret = MAX_USER_RT_PRIO-1;
6386 * sys_sched_get_priority_min - return minimum RT priority.
6387 * @policy: scheduling class.
6389 * this syscall returns the minimum rt_priority that can be used
6390 * by a given scheduling class.
6392 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6410 * sys_sched_rr_get_interval - return the default timeslice of a process.
6411 * @pid: pid of the process.
6412 * @interval: userspace pointer to the timeslice value.
6414 * this syscall writes the default timeslice value of a given process
6415 * into the user-space timespec buffer. A value of '0' means infinity.
6417 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6418 struct timespec __user *, interval)
6420 struct task_struct *p;
6421 unsigned int time_slice;
6429 read_lock(&tasklist_lock);
6430 p = find_process_by_pid(pid);
6434 retval = security_task_getscheduler(p);
6439 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6440 * tasks that are on an otherwise idle runqueue:
6443 if (p->policy == SCHED_RR) {
6444 time_slice = DEF_TIMESLICE;
6445 } else if (p->policy != SCHED_FIFO) {
6446 struct sched_entity *se = &p->se;
6447 unsigned long flags;
6450 rq = task_rq_lock(p, &flags);
6451 if (rq->cfs.load.weight)
6452 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6453 task_rq_unlock(rq, &flags);
6455 read_unlock(&tasklist_lock);
6456 jiffies_to_timespec(time_slice, &t);
6457 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6461 read_unlock(&tasklist_lock);
6465 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6467 void sched_show_task(struct task_struct *p)
6469 unsigned long free = 0;
6472 state = p->state ? __ffs(p->state) + 1 : 0;
6473 printk(KERN_INFO "%-13.13s %c", p->comm,
6474 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6475 #if BITS_PER_LONG == 32
6476 if (state == TASK_RUNNING)
6477 printk(KERN_CONT " running ");
6479 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6481 if (state == TASK_RUNNING)
6482 printk(KERN_CONT " running task ");
6484 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6486 #ifdef CONFIG_DEBUG_STACK_USAGE
6487 free = stack_not_used(p);
6489 printk(KERN_CONT "%5lu %5d %6d\n", free,
6490 task_pid_nr(p), task_pid_nr(p->real_parent));
6492 show_stack(p, NULL);
6495 void show_state_filter(unsigned long state_filter)
6497 struct task_struct *g, *p;
6499 #if BITS_PER_LONG == 32
6501 " task PC stack pid father\n");
6504 " task PC stack pid father\n");
6506 read_lock(&tasklist_lock);
6507 do_each_thread(g, p) {
6509 * reset the NMI-timeout, listing all files on a slow
6510 * console might take alot of time:
6512 touch_nmi_watchdog();
6513 if (!state_filter || (p->state & state_filter))
6515 } while_each_thread(g, p);
6517 touch_all_softlockup_watchdogs();
6519 #ifdef CONFIG_SCHED_DEBUG
6520 sysrq_sched_debug_show();
6522 read_unlock(&tasklist_lock);
6524 * Only show locks if all tasks are dumped:
6526 if (state_filter == -1)
6527 debug_show_all_locks();
6530 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6532 idle->sched_class = &idle_sched_class;
6536 * init_idle - set up an idle thread for a given CPU
6537 * @idle: task in question
6538 * @cpu: cpu the idle task belongs to
6540 * NOTE: this function does not set the idle thread's NEED_RESCHED
6541 * flag, to make booting more robust.
6543 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6545 struct rq *rq = cpu_rq(cpu);
6546 unsigned long flags;
6548 spin_lock_irqsave(&rq->lock, flags);
6551 idle->se.exec_start = sched_clock();
6553 idle->prio = idle->normal_prio = MAX_PRIO;
6554 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6555 __set_task_cpu(idle, cpu);
6557 rq->curr = rq->idle = idle;
6558 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6561 spin_unlock_irqrestore(&rq->lock, flags);
6563 /* Set the preempt count _outside_ the spinlocks! */
6564 #if defined(CONFIG_PREEMPT)
6565 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6567 task_thread_info(idle)->preempt_count = 0;
6570 * The idle tasks have their own, simple scheduling class:
6572 idle->sched_class = &idle_sched_class;
6573 ftrace_graph_init_task(idle);
6577 * In a system that switches off the HZ timer nohz_cpu_mask
6578 * indicates which cpus entered this state. This is used
6579 * in the rcu update to wait only for active cpus. For system
6580 * which do not switch off the HZ timer nohz_cpu_mask should
6581 * always be CPU_BITS_NONE.
6583 cpumask_var_t nohz_cpu_mask;
6586 * Increase the granularity value when there are more CPUs,
6587 * because with more CPUs the 'effective latency' as visible
6588 * to users decreases. But the relationship is not linear,
6589 * so pick a second-best guess by going with the log2 of the
6592 * This idea comes from the SD scheduler of Con Kolivas:
6594 static inline void sched_init_granularity(void)
6596 unsigned int factor = 1 + ilog2(num_online_cpus());
6597 const unsigned long limit = 200000000;
6599 sysctl_sched_min_granularity *= factor;
6600 if (sysctl_sched_min_granularity > limit)
6601 sysctl_sched_min_granularity = limit;
6603 sysctl_sched_latency *= factor;
6604 if (sysctl_sched_latency > limit)
6605 sysctl_sched_latency = limit;
6607 sysctl_sched_wakeup_granularity *= factor;
6609 sysctl_sched_shares_ratelimit *= factor;
6614 * This is how migration works:
6616 * 1) we queue a struct migration_req structure in the source CPU's
6617 * runqueue and wake up that CPU's migration thread.
6618 * 2) we down() the locked semaphore => thread blocks.
6619 * 3) migration thread wakes up (implicitly it forces the migrated
6620 * thread off the CPU)
6621 * 4) it gets the migration request and checks whether the migrated
6622 * task is still in the wrong runqueue.
6623 * 5) if it's in the wrong runqueue then the migration thread removes
6624 * it and puts it into the right queue.
6625 * 6) migration thread up()s the semaphore.
6626 * 7) we wake up and the migration is done.
6630 * Change a given task's CPU affinity. Migrate the thread to a
6631 * proper CPU and schedule it away if the CPU it's executing on
6632 * is removed from the allowed bitmask.
6634 * NOTE: the caller must have a valid reference to the task, the
6635 * task must not exit() & deallocate itself prematurely. The
6636 * call is not atomic; no spinlocks may be held.
6638 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6640 struct migration_req req;
6641 unsigned long flags;
6645 rq = task_rq_lock(p, &flags);
6646 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6651 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6652 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6657 if (p->sched_class->set_cpus_allowed)
6658 p->sched_class->set_cpus_allowed(p, new_mask);
6660 cpumask_copy(&p->cpus_allowed, new_mask);
6661 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6664 /* Can the task run on the task's current CPU? If so, we're done */
6665 if (cpumask_test_cpu(task_cpu(p), new_mask))
6668 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6669 /* Need help from migration thread: drop lock and wait. */
6670 task_rq_unlock(rq, &flags);
6671 wake_up_process(rq->migration_thread);
6672 wait_for_completion(&req.done);
6673 tlb_migrate_finish(p->mm);
6677 task_rq_unlock(rq, &flags);
6681 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6684 * Move (not current) task off this cpu, onto dest cpu. We're doing
6685 * this because either it can't run here any more (set_cpus_allowed()
6686 * away from this CPU, or CPU going down), or because we're
6687 * attempting to rebalance this task on exec (sched_exec).
6689 * So we race with normal scheduler movements, but that's OK, as long
6690 * as the task is no longer on this CPU.
6692 * Returns non-zero if task was successfully migrated.
6694 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6696 struct rq *rq_dest, *rq_src;
6699 if (unlikely(!cpu_active(dest_cpu)))
6702 rq_src = cpu_rq(src_cpu);
6703 rq_dest = cpu_rq(dest_cpu);
6705 double_rq_lock(rq_src, rq_dest);
6706 /* Already moved. */
6707 if (task_cpu(p) != src_cpu)
6709 /* Affinity changed (again). */
6710 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6713 on_rq = p->se.on_rq;
6715 deactivate_task(rq_src, p, 0);
6717 set_task_cpu(p, dest_cpu);
6719 activate_task(rq_dest, p, 0);
6720 check_preempt_curr(rq_dest, p, 0);
6725 double_rq_unlock(rq_src, rq_dest);
6730 * migration_thread - this is a highprio system thread that performs
6731 * thread migration by bumping thread off CPU then 'pushing' onto
6734 static int migration_thread(void *data)
6736 int cpu = (long)data;
6740 BUG_ON(rq->migration_thread != current);
6742 set_current_state(TASK_INTERRUPTIBLE);
6743 while (!kthread_should_stop()) {
6744 struct migration_req *req;
6745 struct list_head *head;
6747 spin_lock_irq(&rq->lock);
6749 if (cpu_is_offline(cpu)) {
6750 spin_unlock_irq(&rq->lock);
6754 if (rq->active_balance) {
6755 active_load_balance(rq, cpu);
6756 rq->active_balance = 0;
6759 head = &rq->migration_queue;
6761 if (list_empty(head)) {
6762 spin_unlock_irq(&rq->lock);
6764 set_current_state(TASK_INTERRUPTIBLE);
6767 req = list_entry(head->next, struct migration_req, list);
6768 list_del_init(head->next);
6770 spin_unlock(&rq->lock);
6771 __migrate_task(req->task, cpu, req->dest_cpu);
6774 complete(&req->done);
6776 __set_current_state(TASK_RUNNING);
6780 /* Wait for kthread_stop */
6781 set_current_state(TASK_INTERRUPTIBLE);
6782 while (!kthread_should_stop()) {
6784 set_current_state(TASK_INTERRUPTIBLE);
6786 __set_current_state(TASK_RUNNING);
6790 #ifdef CONFIG_HOTPLUG_CPU
6792 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6796 local_irq_disable();
6797 ret = __migrate_task(p, src_cpu, dest_cpu);
6803 * Figure out where task on dead CPU should go, use force if necessary.
6805 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6808 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
6811 /* Look for allowed, online CPU in same node. */
6812 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
6813 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6816 /* Any allowed, online CPU? */
6817 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
6818 if (dest_cpu < nr_cpu_ids)
6821 /* No more Mr. Nice Guy. */
6822 if (dest_cpu >= nr_cpu_ids) {
6823 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
6824 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
6827 * Don't tell them about moving exiting tasks or
6828 * kernel threads (both mm NULL), since they never
6831 if (p->mm && printk_ratelimit()) {
6832 printk(KERN_INFO "process %d (%s) no "
6833 "longer affine to cpu%d\n",
6834 task_pid_nr(p), p->comm, dead_cpu);
6839 /* It can have affinity changed while we were choosing. */
6840 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
6845 * While a dead CPU has no uninterruptible tasks queued at this point,
6846 * it might still have a nonzero ->nr_uninterruptible counter, because
6847 * for performance reasons the counter is not stricly tracking tasks to
6848 * their home CPUs. So we just add the counter to another CPU's counter,
6849 * to keep the global sum constant after CPU-down:
6851 static void migrate_nr_uninterruptible(struct rq *rq_src)
6853 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
6854 unsigned long flags;
6856 local_irq_save(flags);
6857 double_rq_lock(rq_src, rq_dest);
6858 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6859 rq_src->nr_uninterruptible = 0;
6860 double_rq_unlock(rq_src, rq_dest);
6861 local_irq_restore(flags);
6864 /* Run through task list and migrate tasks from the dead cpu. */
6865 static void migrate_live_tasks(int src_cpu)
6867 struct task_struct *p, *t;
6869 read_lock(&tasklist_lock);
6871 do_each_thread(t, p) {
6875 if (task_cpu(p) == src_cpu)
6876 move_task_off_dead_cpu(src_cpu, p);
6877 } while_each_thread(t, p);
6879 read_unlock(&tasklist_lock);
6883 * Schedules idle task to be the next runnable task on current CPU.
6884 * It does so by boosting its priority to highest possible.
6885 * Used by CPU offline code.
6887 void sched_idle_next(void)
6889 int this_cpu = smp_processor_id();
6890 struct rq *rq = cpu_rq(this_cpu);
6891 struct task_struct *p = rq->idle;
6892 unsigned long flags;
6894 /* cpu has to be offline */
6895 BUG_ON(cpu_online(this_cpu));
6898 * Strictly not necessary since rest of the CPUs are stopped by now
6899 * and interrupts disabled on the current cpu.
6901 spin_lock_irqsave(&rq->lock, flags);
6903 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6905 update_rq_clock(rq);
6906 activate_task(rq, p, 0);
6908 spin_unlock_irqrestore(&rq->lock, flags);
6912 * Ensures that the idle task is using init_mm right before its cpu goes
6915 void idle_task_exit(void)
6917 struct mm_struct *mm = current->active_mm;
6919 BUG_ON(cpu_online(smp_processor_id()));
6922 switch_mm(mm, &init_mm, current);
6926 /* called under rq->lock with disabled interrupts */
6927 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6929 struct rq *rq = cpu_rq(dead_cpu);
6931 /* Must be exiting, otherwise would be on tasklist. */
6932 BUG_ON(!p->exit_state);
6934 /* Cannot have done final schedule yet: would have vanished. */
6935 BUG_ON(p->state == TASK_DEAD);
6940 * Drop lock around migration; if someone else moves it,
6941 * that's OK. No task can be added to this CPU, so iteration is
6944 spin_unlock_irq(&rq->lock);
6945 move_task_off_dead_cpu(dead_cpu, p);
6946 spin_lock_irq(&rq->lock);
6951 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6952 static void migrate_dead_tasks(unsigned int dead_cpu)
6954 struct rq *rq = cpu_rq(dead_cpu);
6955 struct task_struct *next;
6958 if (!rq->nr_running)
6960 update_rq_clock(rq);
6961 next = pick_next_task(rq);
6964 next->sched_class->put_prev_task(rq, next);
6965 migrate_dead(dead_cpu, next);
6969 #endif /* CONFIG_HOTPLUG_CPU */
6971 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6973 static struct ctl_table sd_ctl_dir[] = {
6975 .procname = "sched_domain",
6981 static struct ctl_table sd_ctl_root[] = {
6983 .ctl_name = CTL_KERN,
6984 .procname = "kernel",
6986 .child = sd_ctl_dir,
6991 static struct ctl_table *sd_alloc_ctl_entry(int n)
6993 struct ctl_table *entry =
6994 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6999 static void sd_free_ctl_entry(struct ctl_table **tablep)
7001 struct ctl_table *entry;
7004 * In the intermediate directories, both the child directory and
7005 * procname are dynamically allocated and could fail but the mode
7006 * will always be set. In the lowest directory the names are
7007 * static strings and all have proc handlers.
7009 for (entry = *tablep; entry->mode; entry++) {
7011 sd_free_ctl_entry(&entry->child);
7012 if (entry->proc_handler == NULL)
7013 kfree(entry->procname);
7021 set_table_entry(struct ctl_table *entry,
7022 const char *procname, void *data, int maxlen,
7023 mode_t mode, proc_handler *proc_handler)
7025 entry->procname = procname;
7027 entry->maxlen = maxlen;
7029 entry->proc_handler = proc_handler;
7032 static struct ctl_table *
7033 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7035 struct ctl_table *table = sd_alloc_ctl_entry(13);
7040 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7041 sizeof(long), 0644, proc_doulongvec_minmax);
7042 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7043 sizeof(long), 0644, proc_doulongvec_minmax);
7044 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7045 sizeof(int), 0644, proc_dointvec_minmax);
7046 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7047 sizeof(int), 0644, proc_dointvec_minmax);
7048 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7049 sizeof(int), 0644, proc_dointvec_minmax);
7050 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7051 sizeof(int), 0644, proc_dointvec_minmax);
7052 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7053 sizeof(int), 0644, proc_dointvec_minmax);
7054 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7055 sizeof(int), 0644, proc_dointvec_minmax);
7056 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7057 sizeof(int), 0644, proc_dointvec_minmax);
7058 set_table_entry(&table[9], "cache_nice_tries",
7059 &sd->cache_nice_tries,
7060 sizeof(int), 0644, proc_dointvec_minmax);
7061 set_table_entry(&table[10], "flags", &sd->flags,
7062 sizeof(int), 0644, proc_dointvec_minmax);
7063 set_table_entry(&table[11], "name", sd->name,
7064 CORENAME_MAX_SIZE, 0444, proc_dostring);
7065 /* &table[12] is terminator */
7070 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7072 struct ctl_table *entry, *table;
7073 struct sched_domain *sd;
7074 int domain_num = 0, i;
7077 for_each_domain(cpu, sd)
7079 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7084 for_each_domain(cpu, sd) {
7085 snprintf(buf, 32, "domain%d", i);
7086 entry->procname = kstrdup(buf, GFP_KERNEL);
7088 entry->child = sd_alloc_ctl_domain_table(sd);
7095 static struct ctl_table_header *sd_sysctl_header;
7096 static void register_sched_domain_sysctl(void)
7098 int i, cpu_num = num_online_cpus();
7099 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7102 WARN_ON(sd_ctl_dir[0].child);
7103 sd_ctl_dir[0].child = entry;
7108 for_each_online_cpu(i) {
7109 snprintf(buf, 32, "cpu%d", i);
7110 entry->procname = kstrdup(buf, GFP_KERNEL);
7112 entry->child = sd_alloc_ctl_cpu_table(i);
7116 WARN_ON(sd_sysctl_header);
7117 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7120 /* may be called multiple times per register */
7121 static void unregister_sched_domain_sysctl(void)
7123 if (sd_sysctl_header)
7124 unregister_sysctl_table(sd_sysctl_header);
7125 sd_sysctl_header = NULL;
7126 if (sd_ctl_dir[0].child)
7127 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7130 static void register_sched_domain_sysctl(void)
7133 static void unregister_sched_domain_sysctl(void)
7138 static void set_rq_online(struct rq *rq)
7141 const struct sched_class *class;
7143 cpumask_set_cpu(rq->cpu, rq->rd->online);
7146 for_each_class(class) {
7147 if (class->rq_online)
7148 class->rq_online(rq);
7153 static void set_rq_offline(struct rq *rq)
7156 const struct sched_class *class;
7158 for_each_class(class) {
7159 if (class->rq_offline)
7160 class->rq_offline(rq);
7163 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7169 * migration_call - callback that gets triggered when a CPU is added.
7170 * Here we can start up the necessary migration thread for the new CPU.
7172 static int __cpuinit
7173 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7175 struct task_struct *p;
7176 int cpu = (long)hcpu;
7177 unsigned long flags;
7182 case CPU_UP_PREPARE:
7183 case CPU_UP_PREPARE_FROZEN:
7184 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7187 kthread_bind(p, cpu);
7188 /* Must be high prio: stop_machine expects to yield to it. */
7189 rq = task_rq_lock(p, &flags);
7190 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7191 task_rq_unlock(rq, &flags);
7192 cpu_rq(cpu)->migration_thread = p;
7196 case CPU_ONLINE_FROZEN:
7197 /* Strictly unnecessary, as first user will wake it. */
7198 wake_up_process(cpu_rq(cpu)->migration_thread);
7200 /* Update our root-domain */
7202 spin_lock_irqsave(&rq->lock, flags);
7204 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7208 spin_unlock_irqrestore(&rq->lock, flags);
7211 #ifdef CONFIG_HOTPLUG_CPU
7212 case CPU_UP_CANCELED:
7213 case CPU_UP_CANCELED_FROZEN:
7214 if (!cpu_rq(cpu)->migration_thread)
7216 /* Unbind it from offline cpu so it can run. Fall thru. */
7217 kthread_bind(cpu_rq(cpu)->migration_thread,
7218 cpumask_any(cpu_online_mask));
7219 kthread_stop(cpu_rq(cpu)->migration_thread);
7220 cpu_rq(cpu)->migration_thread = NULL;
7224 case CPU_DEAD_FROZEN:
7225 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7226 migrate_live_tasks(cpu);
7228 kthread_stop(rq->migration_thread);
7229 rq->migration_thread = NULL;
7230 /* Idle task back to normal (off runqueue, low prio) */
7231 spin_lock_irq(&rq->lock);
7232 update_rq_clock(rq);
7233 deactivate_task(rq, rq->idle, 0);
7234 rq->idle->static_prio = MAX_PRIO;
7235 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7236 rq->idle->sched_class = &idle_sched_class;
7237 migrate_dead_tasks(cpu);
7238 spin_unlock_irq(&rq->lock);
7240 migrate_nr_uninterruptible(rq);
7241 BUG_ON(rq->nr_running != 0);
7244 * No need to migrate the tasks: it was best-effort if
7245 * they didn't take sched_hotcpu_mutex. Just wake up
7248 spin_lock_irq(&rq->lock);
7249 while (!list_empty(&rq->migration_queue)) {
7250 struct migration_req *req;
7252 req = list_entry(rq->migration_queue.next,
7253 struct migration_req, list);
7254 list_del_init(&req->list);
7255 spin_unlock_irq(&rq->lock);
7256 complete(&req->done);
7257 spin_lock_irq(&rq->lock);
7259 spin_unlock_irq(&rq->lock);
7263 case CPU_DYING_FROZEN:
7264 /* Update our root-domain */
7266 spin_lock_irqsave(&rq->lock, flags);
7268 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7271 spin_unlock_irqrestore(&rq->lock, flags);
7278 /* Register at highest priority so that task migration (migrate_all_tasks)
7279 * happens before everything else.
7281 static struct notifier_block __cpuinitdata migration_notifier = {
7282 .notifier_call = migration_call,
7286 static int __init migration_init(void)
7288 void *cpu = (void *)(long)smp_processor_id();
7291 /* Start one for the boot CPU: */
7292 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7293 BUG_ON(err == NOTIFY_BAD);
7294 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7295 register_cpu_notifier(&migration_notifier);
7299 early_initcall(migration_init);
7304 #ifdef CONFIG_SCHED_DEBUG
7306 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7307 struct cpumask *groupmask)
7309 struct sched_group *group = sd->groups;
7312 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7313 cpumask_clear(groupmask);
7315 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7317 if (!(sd->flags & SD_LOAD_BALANCE)) {
7318 printk("does not load-balance\n");
7320 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7325 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7327 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7328 printk(KERN_ERR "ERROR: domain->span does not contain "
7331 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7332 printk(KERN_ERR "ERROR: domain->groups does not contain"
7336 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7340 printk(KERN_ERR "ERROR: group is NULL\n");
7344 if (!group->__cpu_power) {
7345 printk(KERN_CONT "\n");
7346 printk(KERN_ERR "ERROR: domain->cpu_power not "
7351 if (!cpumask_weight(sched_group_cpus(group))) {
7352 printk(KERN_CONT "\n");
7353 printk(KERN_ERR "ERROR: empty group\n");
7357 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7358 printk(KERN_CONT "\n");
7359 printk(KERN_ERR "ERROR: repeated CPUs\n");
7363 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7365 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7366 printk(KERN_CONT " %s", str);
7368 group = group->next;
7369 } while (group != sd->groups);
7370 printk(KERN_CONT "\n");
7372 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7373 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7376 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7377 printk(KERN_ERR "ERROR: parent span is not a superset "
7378 "of domain->span\n");
7382 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7384 cpumask_var_t groupmask;
7388 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7392 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7394 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7395 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7400 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7407 free_cpumask_var(groupmask);
7409 #else /* !CONFIG_SCHED_DEBUG */
7410 # define sched_domain_debug(sd, cpu) do { } while (0)
7411 #endif /* CONFIG_SCHED_DEBUG */
7413 static int sd_degenerate(struct sched_domain *sd)
7415 if (cpumask_weight(sched_domain_span(sd)) == 1)
7418 /* Following flags need at least 2 groups */
7419 if (sd->flags & (SD_LOAD_BALANCE |
7420 SD_BALANCE_NEWIDLE |
7424 SD_SHARE_PKG_RESOURCES)) {
7425 if (sd->groups != sd->groups->next)
7429 /* Following flags don't use groups */
7430 if (sd->flags & (SD_WAKE_IDLE |
7439 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7441 unsigned long cflags = sd->flags, pflags = parent->flags;
7443 if (sd_degenerate(parent))
7446 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7449 /* Does parent contain flags not in child? */
7450 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7451 if (cflags & SD_WAKE_AFFINE)
7452 pflags &= ~SD_WAKE_BALANCE;
7453 /* Flags needing groups don't count if only 1 group in parent */
7454 if (parent->groups == parent->groups->next) {
7455 pflags &= ~(SD_LOAD_BALANCE |
7456 SD_BALANCE_NEWIDLE |
7460 SD_SHARE_PKG_RESOURCES);
7461 if (nr_node_ids == 1)
7462 pflags &= ~SD_SERIALIZE;
7464 if (~cflags & pflags)
7470 static void free_rootdomain(struct root_domain *rd)
7472 cpupri_cleanup(&rd->cpupri);
7474 free_cpumask_var(rd->rto_mask);
7475 free_cpumask_var(rd->online);
7476 free_cpumask_var(rd->span);
7480 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7482 struct root_domain *old_rd = NULL;
7483 unsigned long flags;
7485 spin_lock_irqsave(&rq->lock, flags);
7490 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7493 cpumask_clear_cpu(rq->cpu, old_rd->span);
7496 * If we dont want to free the old_rt yet then
7497 * set old_rd to NULL to skip the freeing later
7500 if (!atomic_dec_and_test(&old_rd->refcount))
7504 atomic_inc(&rd->refcount);
7507 cpumask_set_cpu(rq->cpu, rd->span);
7508 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7511 spin_unlock_irqrestore(&rq->lock, flags);
7514 free_rootdomain(old_rd);
7517 static int __init_refok init_rootdomain(struct root_domain *rd, bool bootmem)
7519 memset(rd, 0, sizeof(*rd));
7522 alloc_bootmem_cpumask_var(&def_root_domain.span);
7523 alloc_bootmem_cpumask_var(&def_root_domain.online);
7524 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
7525 cpupri_init(&rd->cpupri, true);
7529 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7531 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7533 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7536 if (cpupri_init(&rd->cpupri, false) != 0)
7541 free_cpumask_var(rd->rto_mask);
7543 free_cpumask_var(rd->online);
7545 free_cpumask_var(rd->span);
7550 static void init_defrootdomain(void)
7552 init_rootdomain(&def_root_domain, true);
7554 atomic_set(&def_root_domain.refcount, 1);
7557 static struct root_domain *alloc_rootdomain(void)
7559 struct root_domain *rd;
7561 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7565 if (init_rootdomain(rd, false) != 0) {
7574 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7575 * hold the hotplug lock.
7578 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7580 struct rq *rq = cpu_rq(cpu);
7581 struct sched_domain *tmp;
7583 /* Remove the sched domains which do not contribute to scheduling. */
7584 for (tmp = sd; tmp; ) {
7585 struct sched_domain *parent = tmp->parent;
7589 if (sd_parent_degenerate(tmp, parent)) {
7590 tmp->parent = parent->parent;
7592 parent->parent->child = tmp;
7597 if (sd && sd_degenerate(sd)) {
7603 sched_domain_debug(sd, cpu);
7605 rq_attach_root(rq, rd);
7606 rcu_assign_pointer(rq->sd, sd);
7609 /* cpus with isolated domains */
7610 static cpumask_var_t cpu_isolated_map;
7612 /* Setup the mask of cpus configured for isolated domains */
7613 static int __init isolated_cpu_setup(char *str)
7615 cpulist_parse(str, cpu_isolated_map);
7619 __setup("isolcpus=", isolated_cpu_setup);
7622 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7623 * to a function which identifies what group(along with sched group) a CPU
7624 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7625 * (due to the fact that we keep track of groups covered with a struct cpumask).
7627 * init_sched_build_groups will build a circular linked list of the groups
7628 * covered by the given span, and will set each group's ->cpumask correctly,
7629 * and ->cpu_power to 0.
7632 init_sched_build_groups(const struct cpumask *span,
7633 const struct cpumask *cpu_map,
7634 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7635 struct sched_group **sg,
7636 struct cpumask *tmpmask),
7637 struct cpumask *covered, struct cpumask *tmpmask)
7639 struct sched_group *first = NULL, *last = NULL;
7642 cpumask_clear(covered);
7644 for_each_cpu(i, span) {
7645 struct sched_group *sg;
7646 int group = group_fn(i, cpu_map, &sg, tmpmask);
7649 if (cpumask_test_cpu(i, covered))
7652 cpumask_clear(sched_group_cpus(sg));
7653 sg->__cpu_power = 0;
7655 for_each_cpu(j, span) {
7656 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7659 cpumask_set_cpu(j, covered);
7660 cpumask_set_cpu(j, sched_group_cpus(sg));
7671 #define SD_NODES_PER_DOMAIN 16
7676 * find_next_best_node - find the next node to include in a sched_domain
7677 * @node: node whose sched_domain we're building
7678 * @used_nodes: nodes already in the sched_domain
7680 * Find the next node to include in a given scheduling domain. Simply
7681 * finds the closest node not already in the @used_nodes map.
7683 * Should use nodemask_t.
7685 static int find_next_best_node(int node, nodemask_t *used_nodes)
7687 int i, n, val, min_val, best_node = 0;
7691 for (i = 0; i < nr_node_ids; i++) {
7692 /* Start at @node */
7693 n = (node + i) % nr_node_ids;
7695 if (!nr_cpus_node(n))
7698 /* Skip already used nodes */
7699 if (node_isset(n, *used_nodes))
7702 /* Simple min distance search */
7703 val = node_distance(node, n);
7705 if (val < min_val) {
7711 node_set(best_node, *used_nodes);
7716 * sched_domain_node_span - get a cpumask for a node's sched_domain
7717 * @node: node whose cpumask we're constructing
7718 * @span: resulting cpumask
7720 * Given a node, construct a good cpumask for its sched_domain to span. It
7721 * should be one that prevents unnecessary balancing, but also spreads tasks
7724 static void sched_domain_node_span(int node, struct cpumask *span)
7726 nodemask_t used_nodes;
7729 cpumask_clear(span);
7730 nodes_clear(used_nodes);
7732 cpumask_or(span, span, cpumask_of_node(node));
7733 node_set(node, used_nodes);
7735 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7736 int next_node = find_next_best_node(node, &used_nodes);
7738 cpumask_or(span, span, cpumask_of_node(next_node));
7741 #endif /* CONFIG_NUMA */
7743 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7746 * The cpus mask in sched_group and sched_domain hangs off the end.
7747 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7748 * for nr_cpu_ids < CONFIG_NR_CPUS.
7750 struct static_sched_group {
7751 struct sched_group sg;
7752 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7755 struct static_sched_domain {
7756 struct sched_domain sd;
7757 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7761 * SMT sched-domains:
7763 #ifdef CONFIG_SCHED_SMT
7764 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7765 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7768 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
7769 struct sched_group **sg, struct cpumask *unused)
7772 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7775 #endif /* CONFIG_SCHED_SMT */
7778 * multi-core sched-domains:
7780 #ifdef CONFIG_SCHED_MC
7781 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7782 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7783 #endif /* CONFIG_SCHED_MC */
7785 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7787 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7788 struct sched_group **sg, struct cpumask *mask)
7792 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
7793 group = cpumask_first(mask);
7795 *sg = &per_cpu(sched_group_core, group).sg;
7798 #elif defined(CONFIG_SCHED_MC)
7800 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7801 struct sched_group **sg, struct cpumask *unused)
7804 *sg = &per_cpu(sched_group_core, cpu).sg;
7809 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
7810 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
7813 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
7814 struct sched_group **sg, struct cpumask *mask)
7817 #ifdef CONFIG_SCHED_MC
7818 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
7819 group = cpumask_first(mask);
7820 #elif defined(CONFIG_SCHED_SMT)
7821 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
7822 group = cpumask_first(mask);
7827 *sg = &per_cpu(sched_group_phys, group).sg;
7833 * The init_sched_build_groups can't handle what we want to do with node
7834 * groups, so roll our own. Now each node has its own list of groups which
7835 * gets dynamically allocated.
7837 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
7838 static struct sched_group ***sched_group_nodes_bycpu;
7840 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
7841 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7843 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
7844 struct sched_group **sg,
7845 struct cpumask *nodemask)
7849 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7850 group = cpumask_first(nodemask);
7853 *sg = &per_cpu(sched_group_allnodes, group).sg;
7857 static void init_numa_sched_groups_power(struct sched_group *group_head)
7859 struct sched_group *sg = group_head;
7865 for_each_cpu(j, sched_group_cpus(sg)) {
7866 struct sched_domain *sd;
7868 sd = &per_cpu(phys_domains, j).sd;
7869 if (j != cpumask_first(sched_group_cpus(sd->groups))) {
7871 * Only add "power" once for each
7877 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7880 } while (sg != group_head);
7882 #endif /* CONFIG_NUMA */
7885 /* Free memory allocated for various sched_group structures */
7886 static void free_sched_groups(const struct cpumask *cpu_map,
7887 struct cpumask *nodemask)
7891 for_each_cpu(cpu, cpu_map) {
7892 struct sched_group **sched_group_nodes
7893 = sched_group_nodes_bycpu[cpu];
7895 if (!sched_group_nodes)
7898 for (i = 0; i < nr_node_ids; i++) {
7899 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7901 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7902 if (cpumask_empty(nodemask))
7912 if (oldsg != sched_group_nodes[i])
7915 kfree(sched_group_nodes);
7916 sched_group_nodes_bycpu[cpu] = NULL;
7919 #else /* !CONFIG_NUMA */
7920 static void free_sched_groups(const struct cpumask *cpu_map,
7921 struct cpumask *nodemask)
7924 #endif /* CONFIG_NUMA */
7927 * Initialize sched groups cpu_power.
7929 * cpu_power indicates the capacity of sched group, which is used while
7930 * distributing the load between different sched groups in a sched domain.
7931 * Typically cpu_power for all the groups in a sched domain will be same unless
7932 * there are asymmetries in the topology. If there are asymmetries, group
7933 * having more cpu_power will pickup more load compared to the group having
7936 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7937 * the maximum number of tasks a group can handle in the presence of other idle
7938 * or lightly loaded groups in the same sched domain.
7940 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7942 struct sched_domain *child;
7943 struct sched_group *group;
7945 WARN_ON(!sd || !sd->groups);
7947 if (cpu != cpumask_first(sched_group_cpus(sd->groups)))
7952 sd->groups->__cpu_power = 0;
7955 * For perf policy, if the groups in child domain share resources
7956 * (for example cores sharing some portions of the cache hierarchy
7957 * or SMT), then set this domain groups cpu_power such that each group
7958 * can handle only one task, when there are other idle groups in the
7959 * same sched domain.
7961 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7963 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7964 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7969 * add cpu_power of each child group to this groups cpu_power
7971 group = child->groups;
7973 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7974 group = group->next;
7975 } while (group != child->groups);
7979 * Initializers for schedule domains
7980 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7983 #ifdef CONFIG_SCHED_DEBUG
7984 # define SD_INIT_NAME(sd, type) sd->name = #type
7986 # define SD_INIT_NAME(sd, type) do { } while (0)
7989 #define SD_INIT(sd, type) sd_init_##type(sd)
7991 #define SD_INIT_FUNC(type) \
7992 static noinline void sd_init_##type(struct sched_domain *sd) \
7994 memset(sd, 0, sizeof(*sd)); \
7995 *sd = SD_##type##_INIT; \
7996 sd->level = SD_LV_##type; \
7997 SD_INIT_NAME(sd, type); \
8002 SD_INIT_FUNC(ALLNODES)
8005 #ifdef CONFIG_SCHED_SMT
8006 SD_INIT_FUNC(SIBLING)
8008 #ifdef CONFIG_SCHED_MC
8012 static int default_relax_domain_level = -1;
8014 static int __init setup_relax_domain_level(char *str)
8018 val = simple_strtoul(str, NULL, 0);
8019 if (val < SD_LV_MAX)
8020 default_relax_domain_level = val;
8024 __setup("relax_domain_level=", setup_relax_domain_level);
8026 static void set_domain_attribute(struct sched_domain *sd,
8027 struct sched_domain_attr *attr)
8031 if (!attr || attr->relax_domain_level < 0) {
8032 if (default_relax_domain_level < 0)
8035 request = default_relax_domain_level;
8037 request = attr->relax_domain_level;
8038 if (request < sd->level) {
8039 /* turn off idle balance on this domain */
8040 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
8042 /* turn on idle balance on this domain */
8043 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
8048 * Build sched domains for a given set of cpus and attach the sched domains
8049 * to the individual cpus
8051 static int __build_sched_domains(const struct cpumask *cpu_map,
8052 struct sched_domain_attr *attr)
8054 int i, err = -ENOMEM;
8055 struct root_domain *rd;
8056 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
8059 cpumask_var_t domainspan, covered, notcovered;
8060 struct sched_group **sched_group_nodes = NULL;
8061 int sd_allnodes = 0;
8063 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
8065 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
8066 goto free_domainspan;
8067 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
8071 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
8072 goto free_notcovered;
8073 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
8075 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
8076 goto free_this_sibling_map;
8077 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
8078 goto free_this_core_map;
8079 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
8080 goto free_send_covered;
8084 * Allocate the per-node list of sched groups
8086 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
8088 if (!sched_group_nodes) {
8089 printk(KERN_WARNING "Can not alloc sched group node list\n");
8094 rd = alloc_rootdomain();
8096 printk(KERN_WARNING "Cannot alloc root domain\n");
8097 goto free_sched_groups;
8101 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
8105 * Set up domains for cpus specified by the cpu_map.
8107 for_each_cpu(i, cpu_map) {
8108 struct sched_domain *sd = NULL, *p;
8110 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
8113 if (cpumask_weight(cpu_map) >
8114 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
8115 sd = &per_cpu(allnodes_domains, i).sd;
8116 SD_INIT(sd, ALLNODES);
8117 set_domain_attribute(sd, attr);
8118 cpumask_copy(sched_domain_span(sd), cpu_map);
8119 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
8125 sd = &per_cpu(node_domains, i).sd;
8127 set_domain_attribute(sd, attr);
8128 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8132 cpumask_and(sched_domain_span(sd),
8133 sched_domain_span(sd), cpu_map);
8137 sd = &per_cpu(phys_domains, i).sd;
8139 set_domain_attribute(sd, attr);
8140 cpumask_copy(sched_domain_span(sd), nodemask);
8144 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
8146 #ifdef CONFIG_SCHED_MC
8148 sd = &per_cpu(core_domains, i).sd;
8150 set_domain_attribute(sd, attr);
8151 cpumask_and(sched_domain_span(sd), cpu_map,
8152 cpu_coregroup_mask(i));
8155 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
8158 #ifdef CONFIG_SCHED_SMT
8160 sd = &per_cpu(cpu_domains, i).sd;
8161 SD_INIT(sd, SIBLING);
8162 set_domain_attribute(sd, attr);
8163 cpumask_and(sched_domain_span(sd),
8164 topology_thread_cpumask(i), cpu_map);
8167 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
8171 #ifdef CONFIG_SCHED_SMT
8172 /* Set up CPU (sibling) groups */
8173 for_each_cpu(i, cpu_map) {
8174 cpumask_and(this_sibling_map,
8175 topology_thread_cpumask(i), cpu_map);
8176 if (i != cpumask_first(this_sibling_map))
8179 init_sched_build_groups(this_sibling_map, cpu_map,
8181 send_covered, tmpmask);
8185 #ifdef CONFIG_SCHED_MC
8186 /* Set up multi-core groups */
8187 for_each_cpu(i, cpu_map) {
8188 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
8189 if (i != cpumask_first(this_core_map))
8192 init_sched_build_groups(this_core_map, cpu_map,
8194 send_covered, tmpmask);
8198 /* Set up physical groups */
8199 for (i = 0; i < nr_node_ids; i++) {
8200 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8201 if (cpumask_empty(nodemask))
8204 init_sched_build_groups(nodemask, cpu_map,
8206 send_covered, tmpmask);
8210 /* Set up node groups */
8212 init_sched_build_groups(cpu_map, cpu_map,
8213 &cpu_to_allnodes_group,
8214 send_covered, tmpmask);
8217 for (i = 0; i < nr_node_ids; i++) {
8218 /* Set up node groups */
8219 struct sched_group *sg, *prev;
8222 cpumask_clear(covered);
8223 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8224 if (cpumask_empty(nodemask)) {
8225 sched_group_nodes[i] = NULL;
8229 sched_domain_node_span(i, domainspan);
8230 cpumask_and(domainspan, domainspan, cpu_map);
8232 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8235 printk(KERN_WARNING "Can not alloc domain group for "
8239 sched_group_nodes[i] = sg;
8240 for_each_cpu(j, nodemask) {
8241 struct sched_domain *sd;
8243 sd = &per_cpu(node_domains, j).sd;
8246 sg->__cpu_power = 0;
8247 cpumask_copy(sched_group_cpus(sg), nodemask);
8249 cpumask_or(covered, covered, nodemask);
8252 for (j = 0; j < nr_node_ids; j++) {
8253 int n = (i + j) % nr_node_ids;
8255 cpumask_complement(notcovered, covered);
8256 cpumask_and(tmpmask, notcovered, cpu_map);
8257 cpumask_and(tmpmask, tmpmask, domainspan);
8258 if (cpumask_empty(tmpmask))
8261 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
8262 if (cpumask_empty(tmpmask))
8265 sg = kmalloc_node(sizeof(struct sched_group) +
8270 "Can not alloc domain group for node %d\n", j);
8273 sg->__cpu_power = 0;
8274 cpumask_copy(sched_group_cpus(sg), tmpmask);
8275 sg->next = prev->next;
8276 cpumask_or(covered, covered, tmpmask);
8283 /* Calculate CPU power for physical packages and nodes */
8284 #ifdef CONFIG_SCHED_SMT
8285 for_each_cpu(i, cpu_map) {
8286 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
8288 init_sched_groups_power(i, sd);
8291 #ifdef CONFIG_SCHED_MC
8292 for_each_cpu(i, cpu_map) {
8293 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
8295 init_sched_groups_power(i, sd);
8299 for_each_cpu(i, cpu_map) {
8300 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
8302 init_sched_groups_power(i, sd);
8306 for (i = 0; i < nr_node_ids; i++)
8307 init_numa_sched_groups_power(sched_group_nodes[i]);
8310 struct sched_group *sg;
8312 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8314 init_numa_sched_groups_power(sg);
8318 /* Attach the domains */
8319 for_each_cpu(i, cpu_map) {
8320 struct sched_domain *sd;
8321 #ifdef CONFIG_SCHED_SMT
8322 sd = &per_cpu(cpu_domains, i).sd;
8323 #elif defined(CONFIG_SCHED_MC)
8324 sd = &per_cpu(core_domains, i).sd;
8326 sd = &per_cpu(phys_domains, i).sd;
8328 cpu_attach_domain(sd, rd, i);
8334 free_cpumask_var(tmpmask);
8336 free_cpumask_var(send_covered);
8338 free_cpumask_var(this_core_map);
8339 free_this_sibling_map:
8340 free_cpumask_var(this_sibling_map);
8342 free_cpumask_var(nodemask);
8345 free_cpumask_var(notcovered);
8347 free_cpumask_var(covered);
8349 free_cpumask_var(domainspan);
8356 kfree(sched_group_nodes);
8362 free_sched_groups(cpu_map, tmpmask);
8363 free_rootdomain(rd);
8368 static int build_sched_domains(const struct cpumask *cpu_map)
8370 return __build_sched_domains(cpu_map, NULL);
8373 static struct cpumask *doms_cur; /* current sched domains */
8374 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8375 static struct sched_domain_attr *dattr_cur;
8376 /* attribues of custom domains in 'doms_cur' */
8379 * Special case: If a kmalloc of a doms_cur partition (array of
8380 * cpumask) fails, then fallback to a single sched domain,
8381 * as determined by the single cpumask fallback_doms.
8383 static cpumask_var_t fallback_doms;
8386 * arch_update_cpu_topology lets virtualized architectures update the
8387 * cpu core maps. It is supposed to return 1 if the topology changed
8388 * or 0 if it stayed the same.
8390 int __attribute__((weak)) arch_update_cpu_topology(void)
8396 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8397 * For now this just excludes isolated cpus, but could be used to
8398 * exclude other special cases in the future.
8400 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8404 arch_update_cpu_topology();
8406 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8408 doms_cur = fallback_doms;
8409 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8411 err = build_sched_domains(doms_cur);
8412 register_sched_domain_sysctl();
8417 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8418 struct cpumask *tmpmask)
8420 free_sched_groups(cpu_map, tmpmask);
8424 * Detach sched domains from a group of cpus specified in cpu_map
8425 * These cpus will now be attached to the NULL domain
8427 static void detach_destroy_domains(const struct cpumask *cpu_map)
8429 /* Save because hotplug lock held. */
8430 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8433 for_each_cpu(i, cpu_map)
8434 cpu_attach_domain(NULL, &def_root_domain, i);
8435 synchronize_sched();
8436 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8439 /* handle null as "default" */
8440 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8441 struct sched_domain_attr *new, int idx_new)
8443 struct sched_domain_attr tmp;
8450 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8451 new ? (new + idx_new) : &tmp,
8452 sizeof(struct sched_domain_attr));
8456 * Partition sched domains as specified by the 'ndoms_new'
8457 * cpumasks in the array doms_new[] of cpumasks. This compares
8458 * doms_new[] to the current sched domain partitioning, doms_cur[].
8459 * It destroys each deleted domain and builds each new domain.
8461 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8462 * The masks don't intersect (don't overlap.) We should setup one
8463 * sched domain for each mask. CPUs not in any of the cpumasks will
8464 * not be load balanced. If the same cpumask appears both in the
8465 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8468 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8469 * ownership of it and will kfree it when done with it. If the caller
8470 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8471 * ndoms_new == 1, and partition_sched_domains() will fallback to
8472 * the single partition 'fallback_doms', it also forces the domains
8475 * If doms_new == NULL it will be replaced with cpu_online_mask.
8476 * ndoms_new == 0 is a special case for destroying existing domains,
8477 * and it will not create the default domain.
8479 * Call with hotplug lock held
8481 /* FIXME: Change to struct cpumask *doms_new[] */
8482 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8483 struct sched_domain_attr *dattr_new)
8488 mutex_lock(&sched_domains_mutex);
8490 /* always unregister in case we don't destroy any domains */
8491 unregister_sched_domain_sysctl();
8493 /* Let architecture update cpu core mappings. */
8494 new_topology = arch_update_cpu_topology();
8496 n = doms_new ? ndoms_new : 0;
8498 /* Destroy deleted domains */
8499 for (i = 0; i < ndoms_cur; i++) {
8500 for (j = 0; j < n && !new_topology; j++) {
8501 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8502 && dattrs_equal(dattr_cur, i, dattr_new, j))
8505 /* no match - a current sched domain not in new doms_new[] */
8506 detach_destroy_domains(doms_cur + i);
8511 if (doms_new == NULL) {
8513 doms_new = fallback_doms;
8514 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8515 WARN_ON_ONCE(dattr_new);
8518 /* Build new domains */
8519 for (i = 0; i < ndoms_new; i++) {
8520 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8521 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8522 && dattrs_equal(dattr_new, i, dattr_cur, j))
8525 /* no match - add a new doms_new */
8526 __build_sched_domains(doms_new + i,
8527 dattr_new ? dattr_new + i : NULL);
8532 /* Remember the new sched domains */
8533 if (doms_cur != fallback_doms)
8535 kfree(dattr_cur); /* kfree(NULL) is safe */
8536 doms_cur = doms_new;
8537 dattr_cur = dattr_new;
8538 ndoms_cur = ndoms_new;
8540 register_sched_domain_sysctl();
8542 mutex_unlock(&sched_domains_mutex);
8545 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8546 static void arch_reinit_sched_domains(void)
8550 /* Destroy domains first to force the rebuild */
8551 partition_sched_domains(0, NULL, NULL);
8553 rebuild_sched_domains();
8557 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8559 unsigned int level = 0;
8561 if (sscanf(buf, "%u", &level) != 1)
8565 * level is always be positive so don't check for
8566 * level < POWERSAVINGS_BALANCE_NONE which is 0
8567 * What happens on 0 or 1 byte write,
8568 * need to check for count as well?
8571 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8575 sched_smt_power_savings = level;
8577 sched_mc_power_savings = level;
8579 arch_reinit_sched_domains();
8584 #ifdef CONFIG_SCHED_MC
8585 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8588 return sprintf(page, "%u\n", sched_mc_power_savings);
8590 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8591 const char *buf, size_t count)
8593 return sched_power_savings_store(buf, count, 0);
8595 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8596 sched_mc_power_savings_show,
8597 sched_mc_power_savings_store);
8600 #ifdef CONFIG_SCHED_SMT
8601 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8604 return sprintf(page, "%u\n", sched_smt_power_savings);
8606 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8607 const char *buf, size_t count)
8609 return sched_power_savings_store(buf, count, 1);
8611 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8612 sched_smt_power_savings_show,
8613 sched_smt_power_savings_store);
8616 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8620 #ifdef CONFIG_SCHED_SMT
8622 err = sysfs_create_file(&cls->kset.kobj,
8623 &attr_sched_smt_power_savings.attr);
8625 #ifdef CONFIG_SCHED_MC
8626 if (!err && mc_capable())
8627 err = sysfs_create_file(&cls->kset.kobj,
8628 &attr_sched_mc_power_savings.attr);
8632 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8634 #ifndef CONFIG_CPUSETS
8636 * Add online and remove offline CPUs from the scheduler domains.
8637 * When cpusets are enabled they take over this function.
8639 static int update_sched_domains(struct notifier_block *nfb,
8640 unsigned long action, void *hcpu)
8644 case CPU_ONLINE_FROZEN:
8646 case CPU_DEAD_FROZEN:
8647 partition_sched_domains(1, NULL, NULL);
8656 static int update_runtime(struct notifier_block *nfb,
8657 unsigned long action, void *hcpu)
8659 int cpu = (int)(long)hcpu;
8662 case CPU_DOWN_PREPARE:
8663 case CPU_DOWN_PREPARE_FROZEN:
8664 disable_runtime(cpu_rq(cpu));
8667 case CPU_DOWN_FAILED:
8668 case CPU_DOWN_FAILED_FROZEN:
8670 case CPU_ONLINE_FROZEN:
8671 enable_runtime(cpu_rq(cpu));
8679 void __init sched_init_smp(void)
8681 cpumask_var_t non_isolated_cpus;
8683 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8685 #if defined(CONFIG_NUMA)
8686 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8688 BUG_ON(sched_group_nodes_bycpu == NULL);
8691 mutex_lock(&sched_domains_mutex);
8692 arch_init_sched_domains(cpu_online_mask);
8693 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8694 if (cpumask_empty(non_isolated_cpus))
8695 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8696 mutex_unlock(&sched_domains_mutex);
8699 #ifndef CONFIG_CPUSETS
8700 /* XXX: Theoretical race here - CPU may be hotplugged now */
8701 hotcpu_notifier(update_sched_domains, 0);
8704 /* RT runtime code needs to handle some hotplug events */
8705 hotcpu_notifier(update_runtime, 0);
8709 /* Move init over to a non-isolated CPU */
8710 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8712 sched_init_granularity();
8713 free_cpumask_var(non_isolated_cpus);
8715 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8716 init_sched_rt_class();
8719 void __init sched_init_smp(void)
8721 sched_init_granularity();
8723 #endif /* CONFIG_SMP */
8725 int in_sched_functions(unsigned long addr)
8727 return in_lock_functions(addr) ||
8728 (addr >= (unsigned long)__sched_text_start
8729 && addr < (unsigned long)__sched_text_end);
8732 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8734 cfs_rq->tasks_timeline = RB_ROOT;
8735 INIT_LIST_HEAD(&cfs_rq->tasks);
8736 #ifdef CONFIG_FAIR_GROUP_SCHED
8739 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8742 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8744 struct rt_prio_array *array;
8747 array = &rt_rq->active;
8748 for (i = 0; i < MAX_RT_PRIO; i++) {
8749 INIT_LIST_HEAD(array->queue + i);
8750 __clear_bit(i, array->bitmap);
8752 /* delimiter for bitsearch: */
8753 __set_bit(MAX_RT_PRIO, array->bitmap);
8755 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8756 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8758 rt_rq->highest_prio.next = MAX_RT_PRIO;
8762 rt_rq->rt_nr_migratory = 0;
8763 rt_rq->overloaded = 0;
8764 plist_head_init(&rq->rt.pushable_tasks, &rq->lock);
8768 rt_rq->rt_throttled = 0;
8769 rt_rq->rt_runtime = 0;
8770 spin_lock_init(&rt_rq->rt_runtime_lock);
8772 #ifdef CONFIG_RT_GROUP_SCHED
8773 rt_rq->rt_nr_boosted = 0;
8778 #ifdef CONFIG_FAIR_GROUP_SCHED
8779 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8780 struct sched_entity *se, int cpu, int add,
8781 struct sched_entity *parent)
8783 struct rq *rq = cpu_rq(cpu);
8784 tg->cfs_rq[cpu] = cfs_rq;
8785 init_cfs_rq(cfs_rq, rq);
8788 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8791 /* se could be NULL for init_task_group */
8796 se->cfs_rq = &rq->cfs;
8798 se->cfs_rq = parent->my_q;
8801 se->load.weight = tg->shares;
8802 se->load.inv_weight = 0;
8803 se->parent = parent;
8807 #ifdef CONFIG_RT_GROUP_SCHED
8808 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8809 struct sched_rt_entity *rt_se, int cpu, int add,
8810 struct sched_rt_entity *parent)
8812 struct rq *rq = cpu_rq(cpu);
8814 tg->rt_rq[cpu] = rt_rq;
8815 init_rt_rq(rt_rq, rq);
8817 rt_rq->rt_se = rt_se;
8818 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8820 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8822 tg->rt_se[cpu] = rt_se;
8827 rt_se->rt_rq = &rq->rt;
8829 rt_se->rt_rq = parent->my_q;
8831 rt_se->my_q = rt_rq;
8832 rt_se->parent = parent;
8833 INIT_LIST_HEAD(&rt_se->run_list);
8837 void __init sched_init(void)
8840 unsigned long alloc_size = 0, ptr;
8842 #ifdef CONFIG_FAIR_GROUP_SCHED
8843 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8845 #ifdef CONFIG_RT_GROUP_SCHED
8846 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8848 #ifdef CONFIG_USER_SCHED
8851 #ifdef CONFIG_CPUMASK_OFFSTACK
8852 alloc_size += num_possible_cpus() * cpumask_size();
8855 * As sched_init() is called before page_alloc is setup,
8856 * we use alloc_bootmem().
8859 ptr = (unsigned long)alloc_bootmem(alloc_size);
8861 #ifdef CONFIG_FAIR_GROUP_SCHED
8862 init_task_group.se = (struct sched_entity **)ptr;
8863 ptr += nr_cpu_ids * sizeof(void **);
8865 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8866 ptr += nr_cpu_ids * sizeof(void **);
8868 #ifdef CONFIG_USER_SCHED
8869 root_task_group.se = (struct sched_entity **)ptr;
8870 ptr += nr_cpu_ids * sizeof(void **);
8872 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8873 ptr += nr_cpu_ids * sizeof(void **);
8874 #endif /* CONFIG_USER_SCHED */
8875 #endif /* CONFIG_FAIR_GROUP_SCHED */
8876 #ifdef CONFIG_RT_GROUP_SCHED
8877 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8878 ptr += nr_cpu_ids * sizeof(void **);
8880 init_task_group.rt_rq = (struct rt_rq **)ptr;
8881 ptr += nr_cpu_ids * sizeof(void **);
8883 #ifdef CONFIG_USER_SCHED
8884 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8885 ptr += nr_cpu_ids * sizeof(void **);
8887 root_task_group.rt_rq = (struct rt_rq **)ptr;
8888 ptr += nr_cpu_ids * sizeof(void **);
8889 #endif /* CONFIG_USER_SCHED */
8890 #endif /* CONFIG_RT_GROUP_SCHED */
8891 #ifdef CONFIG_CPUMASK_OFFSTACK
8892 for_each_possible_cpu(i) {
8893 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8894 ptr += cpumask_size();
8896 #endif /* CONFIG_CPUMASK_OFFSTACK */
8900 init_defrootdomain();
8903 init_rt_bandwidth(&def_rt_bandwidth,
8904 global_rt_period(), global_rt_runtime());
8906 #ifdef CONFIG_RT_GROUP_SCHED
8907 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8908 global_rt_period(), global_rt_runtime());
8909 #ifdef CONFIG_USER_SCHED
8910 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8911 global_rt_period(), RUNTIME_INF);
8912 #endif /* CONFIG_USER_SCHED */
8913 #endif /* CONFIG_RT_GROUP_SCHED */
8915 #ifdef CONFIG_GROUP_SCHED
8916 list_add(&init_task_group.list, &task_groups);
8917 INIT_LIST_HEAD(&init_task_group.children);
8919 #ifdef CONFIG_USER_SCHED
8920 INIT_LIST_HEAD(&root_task_group.children);
8921 init_task_group.parent = &root_task_group;
8922 list_add(&init_task_group.siblings, &root_task_group.children);
8923 #endif /* CONFIG_USER_SCHED */
8924 #endif /* CONFIG_GROUP_SCHED */
8926 for_each_possible_cpu(i) {
8930 spin_lock_init(&rq->lock);
8932 init_cfs_rq(&rq->cfs, rq);
8933 init_rt_rq(&rq->rt, rq);
8934 #ifdef CONFIG_FAIR_GROUP_SCHED
8935 init_task_group.shares = init_task_group_load;
8936 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8937 #ifdef CONFIG_CGROUP_SCHED
8939 * How much cpu bandwidth does init_task_group get?
8941 * In case of task-groups formed thr' the cgroup filesystem, it
8942 * gets 100% of the cpu resources in the system. This overall
8943 * system cpu resource is divided among the tasks of
8944 * init_task_group and its child task-groups in a fair manner,
8945 * based on each entity's (task or task-group's) weight
8946 * (se->load.weight).
8948 * In other words, if init_task_group has 10 tasks of weight
8949 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8950 * then A0's share of the cpu resource is:
8952 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8954 * We achieve this by letting init_task_group's tasks sit
8955 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8957 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8958 #elif defined CONFIG_USER_SCHED
8959 root_task_group.shares = NICE_0_LOAD;
8960 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8962 * In case of task-groups formed thr' the user id of tasks,
8963 * init_task_group represents tasks belonging to root user.
8964 * Hence it forms a sibling of all subsequent groups formed.
8965 * In this case, init_task_group gets only a fraction of overall
8966 * system cpu resource, based on the weight assigned to root
8967 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8968 * by letting tasks of init_task_group sit in a separate cfs_rq
8969 * (init_cfs_rq) and having one entity represent this group of
8970 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8972 init_tg_cfs_entry(&init_task_group,
8973 &per_cpu(init_cfs_rq, i),
8974 &per_cpu(init_sched_entity, i), i, 1,
8975 root_task_group.se[i]);
8978 #endif /* CONFIG_FAIR_GROUP_SCHED */
8980 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8981 #ifdef CONFIG_RT_GROUP_SCHED
8982 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8983 #ifdef CONFIG_CGROUP_SCHED
8984 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8985 #elif defined CONFIG_USER_SCHED
8986 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8987 init_tg_rt_entry(&init_task_group,
8988 &per_cpu(init_rt_rq, i),
8989 &per_cpu(init_sched_rt_entity, i), i, 1,
8990 root_task_group.rt_se[i]);
8994 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8995 rq->cpu_load[j] = 0;
8999 rq->active_balance = 0;
9000 rq->next_balance = jiffies;
9004 rq->migration_thread = NULL;
9005 INIT_LIST_HEAD(&rq->migration_queue);
9006 rq_attach_root(rq, &def_root_domain);
9009 atomic_set(&rq->nr_iowait, 0);
9012 set_load_weight(&init_task);
9014 #ifdef CONFIG_PREEMPT_NOTIFIERS
9015 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9019 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9022 #ifdef CONFIG_RT_MUTEXES
9023 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9027 * The boot idle thread does lazy MMU switching as well:
9029 atomic_inc(&init_mm.mm_count);
9030 enter_lazy_tlb(&init_mm, current);
9033 * Make us the idle thread. Technically, schedule() should not be
9034 * called from this thread, however somewhere below it might be,
9035 * but because we are the idle thread, we just pick up running again
9036 * when this runqueue becomes "idle".
9038 init_idle(current, smp_processor_id());
9040 * During early bootup we pretend to be a normal task:
9042 current->sched_class = &fair_sched_class;
9044 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9045 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
9048 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
9050 alloc_bootmem_cpumask_var(&cpu_isolated_map);
9053 scheduler_running = 1;
9056 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9057 void __might_sleep(char *file, int line)
9060 static unsigned long prev_jiffy; /* ratelimiting */
9062 if ((!in_atomic() && !irqs_disabled()) ||
9063 system_state != SYSTEM_RUNNING || oops_in_progress)
9065 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9067 prev_jiffy = jiffies;
9070 "BUG: sleeping function called from invalid context at %s:%d\n",
9073 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9074 in_atomic(), irqs_disabled(),
9075 current->pid, current->comm);
9077 debug_show_held_locks(current);
9078 if (irqs_disabled())
9079 print_irqtrace_events(current);
9083 EXPORT_SYMBOL(__might_sleep);
9086 #ifdef CONFIG_MAGIC_SYSRQ
9087 static void normalize_task(struct rq *rq, struct task_struct *p)
9091 update_rq_clock(rq);
9092 on_rq = p->se.on_rq;
9094 deactivate_task(rq, p, 0);
9095 __setscheduler(rq, p, SCHED_NORMAL, 0);
9097 activate_task(rq, p, 0);
9098 resched_task(rq->curr);
9102 void normalize_rt_tasks(void)
9104 struct task_struct *g, *p;
9105 unsigned long flags;
9108 read_lock_irqsave(&tasklist_lock, flags);
9109 do_each_thread(g, p) {
9111 * Only normalize user tasks:
9116 p->se.exec_start = 0;
9117 #ifdef CONFIG_SCHEDSTATS
9118 p->se.wait_start = 0;
9119 p->se.sleep_start = 0;
9120 p->se.block_start = 0;
9125 * Renice negative nice level userspace
9128 if (TASK_NICE(p) < 0 && p->mm)
9129 set_user_nice(p, 0);
9133 spin_lock(&p->pi_lock);
9134 rq = __task_rq_lock(p);
9136 normalize_task(rq, p);
9138 __task_rq_unlock(rq);
9139 spin_unlock(&p->pi_lock);
9140 } while_each_thread(g, p);
9142 read_unlock_irqrestore(&tasklist_lock, flags);
9145 #endif /* CONFIG_MAGIC_SYSRQ */
9149 * These functions are only useful for the IA64 MCA handling.
9151 * They can only be called when the whole system has been
9152 * stopped - every CPU needs to be quiescent, and no scheduling
9153 * activity can take place. Using them for anything else would
9154 * be a serious bug, and as a result, they aren't even visible
9155 * under any other configuration.
9159 * curr_task - return the current task for a given cpu.
9160 * @cpu: the processor in question.
9162 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9164 struct task_struct *curr_task(int cpu)
9166 return cpu_curr(cpu);
9170 * set_curr_task - set the current task for a given cpu.
9171 * @cpu: the processor in question.
9172 * @p: the task pointer to set.
9174 * Description: This function must only be used when non-maskable interrupts
9175 * are serviced on a separate stack. It allows the architecture to switch the
9176 * notion of the current task on a cpu in a non-blocking manner. This function
9177 * must be called with all CPU's synchronized, and interrupts disabled, the
9178 * and caller must save the original value of the current task (see
9179 * curr_task() above) and restore that value before reenabling interrupts and
9180 * re-starting the system.
9182 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9184 void set_curr_task(int cpu, struct task_struct *p)
9191 #ifdef CONFIG_FAIR_GROUP_SCHED
9192 static void free_fair_sched_group(struct task_group *tg)
9196 for_each_possible_cpu(i) {
9198 kfree(tg->cfs_rq[i]);
9208 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9210 struct cfs_rq *cfs_rq;
9211 struct sched_entity *se;
9215 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9218 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9222 tg->shares = NICE_0_LOAD;
9224 for_each_possible_cpu(i) {
9227 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9228 GFP_KERNEL, cpu_to_node(i));
9232 se = kzalloc_node(sizeof(struct sched_entity),
9233 GFP_KERNEL, cpu_to_node(i));
9237 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9246 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9248 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9249 &cpu_rq(cpu)->leaf_cfs_rq_list);
9252 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9254 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9256 #else /* !CONFG_FAIR_GROUP_SCHED */
9257 static inline void free_fair_sched_group(struct task_group *tg)
9262 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9267 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9271 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9274 #endif /* CONFIG_FAIR_GROUP_SCHED */
9276 #ifdef CONFIG_RT_GROUP_SCHED
9277 static void free_rt_sched_group(struct task_group *tg)
9281 destroy_rt_bandwidth(&tg->rt_bandwidth);
9283 for_each_possible_cpu(i) {
9285 kfree(tg->rt_rq[i]);
9287 kfree(tg->rt_se[i]);
9295 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9297 struct rt_rq *rt_rq;
9298 struct sched_rt_entity *rt_se;
9302 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9305 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9309 init_rt_bandwidth(&tg->rt_bandwidth,
9310 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9312 for_each_possible_cpu(i) {
9315 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9316 GFP_KERNEL, cpu_to_node(i));
9320 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9321 GFP_KERNEL, cpu_to_node(i));
9325 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9334 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9336 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9337 &cpu_rq(cpu)->leaf_rt_rq_list);
9340 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9342 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9344 #else /* !CONFIG_RT_GROUP_SCHED */
9345 static inline void free_rt_sched_group(struct task_group *tg)
9350 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9355 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9359 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9362 #endif /* CONFIG_RT_GROUP_SCHED */
9364 #ifdef CONFIG_GROUP_SCHED
9365 static void free_sched_group(struct task_group *tg)
9367 free_fair_sched_group(tg);
9368 free_rt_sched_group(tg);
9372 /* allocate runqueue etc for a new task group */
9373 struct task_group *sched_create_group(struct task_group *parent)
9375 struct task_group *tg;
9376 unsigned long flags;
9379 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9381 return ERR_PTR(-ENOMEM);
9383 if (!alloc_fair_sched_group(tg, parent))
9386 if (!alloc_rt_sched_group(tg, parent))
9389 spin_lock_irqsave(&task_group_lock, flags);
9390 for_each_possible_cpu(i) {
9391 register_fair_sched_group(tg, i);
9392 register_rt_sched_group(tg, i);
9394 list_add_rcu(&tg->list, &task_groups);
9396 WARN_ON(!parent); /* root should already exist */
9398 tg->parent = parent;
9399 INIT_LIST_HEAD(&tg->children);
9400 list_add_rcu(&tg->siblings, &parent->children);
9401 spin_unlock_irqrestore(&task_group_lock, flags);
9406 free_sched_group(tg);
9407 return ERR_PTR(-ENOMEM);
9410 /* rcu callback to free various structures associated with a task group */
9411 static void free_sched_group_rcu(struct rcu_head *rhp)
9413 /* now it should be safe to free those cfs_rqs */
9414 free_sched_group(container_of(rhp, struct task_group, rcu));
9417 /* Destroy runqueue etc associated with a task group */
9418 void sched_destroy_group(struct task_group *tg)
9420 unsigned long flags;
9423 spin_lock_irqsave(&task_group_lock, flags);
9424 for_each_possible_cpu(i) {
9425 unregister_fair_sched_group(tg, i);
9426 unregister_rt_sched_group(tg, i);
9428 list_del_rcu(&tg->list);
9429 list_del_rcu(&tg->siblings);
9430 spin_unlock_irqrestore(&task_group_lock, flags);
9432 /* wait for possible concurrent references to cfs_rqs complete */
9433 call_rcu(&tg->rcu, free_sched_group_rcu);
9436 /* change task's runqueue when it moves between groups.
9437 * The caller of this function should have put the task in its new group
9438 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9439 * reflect its new group.
9441 void sched_move_task(struct task_struct *tsk)
9444 unsigned long flags;
9447 rq = task_rq_lock(tsk, &flags);
9449 update_rq_clock(rq);
9451 running = task_current(rq, tsk);
9452 on_rq = tsk->se.on_rq;
9455 dequeue_task(rq, tsk, 0);
9456 if (unlikely(running))
9457 tsk->sched_class->put_prev_task(rq, tsk);
9459 set_task_rq(tsk, task_cpu(tsk));
9461 #ifdef CONFIG_FAIR_GROUP_SCHED
9462 if (tsk->sched_class->moved_group)
9463 tsk->sched_class->moved_group(tsk);
9466 if (unlikely(running))
9467 tsk->sched_class->set_curr_task(rq);
9469 enqueue_task(rq, tsk, 0);
9471 task_rq_unlock(rq, &flags);
9473 #endif /* CONFIG_GROUP_SCHED */
9475 #ifdef CONFIG_FAIR_GROUP_SCHED
9476 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9478 struct cfs_rq *cfs_rq = se->cfs_rq;
9483 dequeue_entity(cfs_rq, se, 0);
9485 se->load.weight = shares;
9486 se->load.inv_weight = 0;
9489 enqueue_entity(cfs_rq, se, 0);
9492 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9494 struct cfs_rq *cfs_rq = se->cfs_rq;
9495 struct rq *rq = cfs_rq->rq;
9496 unsigned long flags;
9498 spin_lock_irqsave(&rq->lock, flags);
9499 __set_se_shares(se, shares);
9500 spin_unlock_irqrestore(&rq->lock, flags);
9503 static DEFINE_MUTEX(shares_mutex);
9505 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9508 unsigned long flags;
9511 * We can't change the weight of the root cgroup.
9516 if (shares < MIN_SHARES)
9517 shares = MIN_SHARES;
9518 else if (shares > MAX_SHARES)
9519 shares = MAX_SHARES;
9521 mutex_lock(&shares_mutex);
9522 if (tg->shares == shares)
9525 spin_lock_irqsave(&task_group_lock, flags);
9526 for_each_possible_cpu(i)
9527 unregister_fair_sched_group(tg, i);
9528 list_del_rcu(&tg->siblings);
9529 spin_unlock_irqrestore(&task_group_lock, flags);
9531 /* wait for any ongoing reference to this group to finish */
9532 synchronize_sched();
9535 * Now we are free to modify the group's share on each cpu
9536 * w/o tripping rebalance_share or load_balance_fair.
9538 tg->shares = shares;
9539 for_each_possible_cpu(i) {
9543 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9544 set_se_shares(tg->se[i], shares);
9548 * Enable load balance activity on this group, by inserting it back on
9549 * each cpu's rq->leaf_cfs_rq_list.
9551 spin_lock_irqsave(&task_group_lock, flags);
9552 for_each_possible_cpu(i)
9553 register_fair_sched_group(tg, i);
9554 list_add_rcu(&tg->siblings, &tg->parent->children);
9555 spin_unlock_irqrestore(&task_group_lock, flags);
9557 mutex_unlock(&shares_mutex);
9561 unsigned long sched_group_shares(struct task_group *tg)
9567 #ifdef CONFIG_RT_GROUP_SCHED
9569 * Ensure that the real time constraints are schedulable.
9571 static DEFINE_MUTEX(rt_constraints_mutex);
9573 static unsigned long to_ratio(u64 period, u64 runtime)
9575 if (runtime == RUNTIME_INF)
9578 return div64_u64(runtime << 20, period);
9581 /* Must be called with tasklist_lock held */
9582 static inline int tg_has_rt_tasks(struct task_group *tg)
9584 struct task_struct *g, *p;
9586 do_each_thread(g, p) {
9587 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9589 } while_each_thread(g, p);
9594 struct rt_schedulable_data {
9595 struct task_group *tg;
9600 static int tg_schedulable(struct task_group *tg, void *data)
9602 struct rt_schedulable_data *d = data;
9603 struct task_group *child;
9604 unsigned long total, sum = 0;
9605 u64 period, runtime;
9607 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9608 runtime = tg->rt_bandwidth.rt_runtime;
9611 period = d->rt_period;
9612 runtime = d->rt_runtime;
9615 #ifdef CONFIG_USER_SCHED
9616 if (tg == &root_task_group) {
9617 period = global_rt_period();
9618 runtime = global_rt_runtime();
9623 * Cannot have more runtime than the period.
9625 if (runtime > period && runtime != RUNTIME_INF)
9629 * Ensure we don't starve existing RT tasks.
9631 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9634 total = to_ratio(period, runtime);
9637 * Nobody can have more than the global setting allows.
9639 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9643 * The sum of our children's runtime should not exceed our own.
9645 list_for_each_entry_rcu(child, &tg->children, siblings) {
9646 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9647 runtime = child->rt_bandwidth.rt_runtime;
9649 if (child == d->tg) {
9650 period = d->rt_period;
9651 runtime = d->rt_runtime;
9654 sum += to_ratio(period, runtime);
9663 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9665 struct rt_schedulable_data data = {
9667 .rt_period = period,
9668 .rt_runtime = runtime,
9671 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9674 static int tg_set_bandwidth(struct task_group *tg,
9675 u64 rt_period, u64 rt_runtime)
9679 mutex_lock(&rt_constraints_mutex);
9680 read_lock(&tasklist_lock);
9681 err = __rt_schedulable(tg, rt_period, rt_runtime);
9685 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9686 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9687 tg->rt_bandwidth.rt_runtime = rt_runtime;
9689 for_each_possible_cpu(i) {
9690 struct rt_rq *rt_rq = tg->rt_rq[i];
9692 spin_lock(&rt_rq->rt_runtime_lock);
9693 rt_rq->rt_runtime = rt_runtime;
9694 spin_unlock(&rt_rq->rt_runtime_lock);
9696 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9698 read_unlock(&tasklist_lock);
9699 mutex_unlock(&rt_constraints_mutex);
9704 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9706 u64 rt_runtime, rt_period;
9708 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9709 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9710 if (rt_runtime_us < 0)
9711 rt_runtime = RUNTIME_INF;
9713 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9716 long sched_group_rt_runtime(struct task_group *tg)
9720 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9723 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9724 do_div(rt_runtime_us, NSEC_PER_USEC);
9725 return rt_runtime_us;
9728 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9730 u64 rt_runtime, rt_period;
9732 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9733 rt_runtime = tg->rt_bandwidth.rt_runtime;
9738 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9741 long sched_group_rt_period(struct task_group *tg)
9745 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9746 do_div(rt_period_us, NSEC_PER_USEC);
9747 return rt_period_us;
9750 static int sched_rt_global_constraints(void)
9752 u64 runtime, period;
9755 if (sysctl_sched_rt_period <= 0)
9758 runtime = global_rt_runtime();
9759 period = global_rt_period();
9762 * Sanity check on the sysctl variables.
9764 if (runtime > period && runtime != RUNTIME_INF)
9767 mutex_lock(&rt_constraints_mutex);
9768 read_lock(&tasklist_lock);
9769 ret = __rt_schedulable(NULL, 0, 0);
9770 read_unlock(&tasklist_lock);
9771 mutex_unlock(&rt_constraints_mutex);
9776 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9778 /* Don't accept realtime tasks when there is no way for them to run */
9779 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9785 #else /* !CONFIG_RT_GROUP_SCHED */
9786 static int sched_rt_global_constraints(void)
9788 unsigned long flags;
9791 if (sysctl_sched_rt_period <= 0)
9794 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9795 for_each_possible_cpu(i) {
9796 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9798 spin_lock(&rt_rq->rt_runtime_lock);
9799 rt_rq->rt_runtime = global_rt_runtime();
9800 spin_unlock(&rt_rq->rt_runtime_lock);
9802 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9806 #endif /* CONFIG_RT_GROUP_SCHED */
9808 int sched_rt_handler(struct ctl_table *table, int write,
9809 struct file *filp, void __user *buffer, size_t *lenp,
9813 int old_period, old_runtime;
9814 static DEFINE_MUTEX(mutex);
9817 old_period = sysctl_sched_rt_period;
9818 old_runtime = sysctl_sched_rt_runtime;
9820 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9822 if (!ret && write) {
9823 ret = sched_rt_global_constraints();
9825 sysctl_sched_rt_period = old_period;
9826 sysctl_sched_rt_runtime = old_runtime;
9828 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9829 def_rt_bandwidth.rt_period =
9830 ns_to_ktime(global_rt_period());
9833 mutex_unlock(&mutex);
9838 #ifdef CONFIG_CGROUP_SCHED
9840 /* return corresponding task_group object of a cgroup */
9841 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9843 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9844 struct task_group, css);
9847 static struct cgroup_subsys_state *
9848 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9850 struct task_group *tg, *parent;
9852 if (!cgrp->parent) {
9853 /* This is early initialization for the top cgroup */
9854 return &init_task_group.css;
9857 parent = cgroup_tg(cgrp->parent);
9858 tg = sched_create_group(parent);
9860 return ERR_PTR(-ENOMEM);
9866 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9868 struct task_group *tg = cgroup_tg(cgrp);
9870 sched_destroy_group(tg);
9874 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9875 struct task_struct *tsk)
9877 #ifdef CONFIG_RT_GROUP_SCHED
9878 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9881 /* We don't support RT-tasks being in separate groups */
9882 if (tsk->sched_class != &fair_sched_class)
9890 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9891 struct cgroup *old_cont, struct task_struct *tsk)
9893 sched_move_task(tsk);
9896 #ifdef CONFIG_FAIR_GROUP_SCHED
9897 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9900 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9903 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9905 struct task_group *tg = cgroup_tg(cgrp);
9907 return (u64) tg->shares;
9909 #endif /* CONFIG_FAIR_GROUP_SCHED */
9911 #ifdef CONFIG_RT_GROUP_SCHED
9912 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9915 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9918 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9920 return sched_group_rt_runtime(cgroup_tg(cgrp));
9923 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9926 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9929 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9931 return sched_group_rt_period(cgroup_tg(cgrp));
9933 #endif /* CONFIG_RT_GROUP_SCHED */
9935 static struct cftype cpu_files[] = {
9936 #ifdef CONFIG_FAIR_GROUP_SCHED
9939 .read_u64 = cpu_shares_read_u64,
9940 .write_u64 = cpu_shares_write_u64,
9943 #ifdef CONFIG_RT_GROUP_SCHED
9945 .name = "rt_runtime_us",
9946 .read_s64 = cpu_rt_runtime_read,
9947 .write_s64 = cpu_rt_runtime_write,
9950 .name = "rt_period_us",
9951 .read_u64 = cpu_rt_period_read_uint,
9952 .write_u64 = cpu_rt_period_write_uint,
9957 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9959 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9962 struct cgroup_subsys cpu_cgroup_subsys = {
9964 .create = cpu_cgroup_create,
9965 .destroy = cpu_cgroup_destroy,
9966 .can_attach = cpu_cgroup_can_attach,
9967 .attach = cpu_cgroup_attach,
9968 .populate = cpu_cgroup_populate,
9969 .subsys_id = cpu_cgroup_subsys_id,
9973 #endif /* CONFIG_CGROUP_SCHED */
9975 #ifdef CONFIG_CGROUP_CPUACCT
9978 * CPU accounting code for task groups.
9980 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9981 * (balbir@in.ibm.com).
9984 /* track cpu usage of a group of tasks and its child groups */
9986 struct cgroup_subsys_state css;
9987 /* cpuusage holds pointer to a u64-type object on every cpu */
9989 struct cpuacct *parent;
9992 struct cgroup_subsys cpuacct_subsys;
9994 /* return cpu accounting group corresponding to this container */
9995 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9997 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9998 struct cpuacct, css);
10001 /* return cpu accounting group to which this task belongs */
10002 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10004 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10005 struct cpuacct, css);
10008 /* create a new cpu accounting group */
10009 static struct cgroup_subsys_state *cpuacct_create(
10010 struct cgroup_subsys *ss, struct cgroup *cgrp)
10012 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10015 return ERR_PTR(-ENOMEM);
10017 ca->cpuusage = alloc_percpu(u64);
10018 if (!ca->cpuusage) {
10020 return ERR_PTR(-ENOMEM);
10024 ca->parent = cgroup_ca(cgrp->parent);
10029 /* destroy an existing cpu accounting group */
10031 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10033 struct cpuacct *ca = cgroup_ca(cgrp);
10035 free_percpu(ca->cpuusage);
10039 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10041 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10044 #ifndef CONFIG_64BIT
10046 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10048 spin_lock_irq(&cpu_rq(cpu)->lock);
10050 spin_unlock_irq(&cpu_rq(cpu)->lock);
10058 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10060 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10062 #ifndef CONFIG_64BIT
10064 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10066 spin_lock_irq(&cpu_rq(cpu)->lock);
10068 spin_unlock_irq(&cpu_rq(cpu)->lock);
10074 /* return total cpu usage (in nanoseconds) of a group */
10075 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10077 struct cpuacct *ca = cgroup_ca(cgrp);
10078 u64 totalcpuusage = 0;
10081 for_each_present_cpu(i)
10082 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10084 return totalcpuusage;
10087 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10090 struct cpuacct *ca = cgroup_ca(cgrp);
10099 for_each_present_cpu(i)
10100 cpuacct_cpuusage_write(ca, i, 0);
10106 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10107 struct seq_file *m)
10109 struct cpuacct *ca = cgroup_ca(cgroup);
10113 for_each_present_cpu(i) {
10114 percpu = cpuacct_cpuusage_read(ca, i);
10115 seq_printf(m, "%llu ", (unsigned long long) percpu);
10117 seq_printf(m, "\n");
10121 static struct cftype files[] = {
10124 .read_u64 = cpuusage_read,
10125 .write_u64 = cpuusage_write,
10128 .name = "usage_percpu",
10129 .read_seq_string = cpuacct_percpu_seq_read,
10134 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10136 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10140 * charge this task's execution time to its accounting group.
10142 * called with rq->lock held.
10144 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10146 struct cpuacct *ca;
10149 if (unlikely(!cpuacct_subsys.active))
10152 cpu = task_cpu(tsk);
10155 for (; ca; ca = ca->parent) {
10156 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10157 *cpuusage += cputime;
10161 struct cgroup_subsys cpuacct_subsys = {
10163 .create = cpuacct_create,
10164 .destroy = cpuacct_destroy,
10165 .populate = cpuacct_populate,
10166 .subsys_id = cpuacct_subsys_id,
10168 #endif /* CONFIG_CGROUP_CPUACCT */