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 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
130 * Since cpu_power is a 'constant', we can use a reciprocal divide.
132 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
134 return reciprocal_divide(load, sg->reciprocal_cpu_power);
138 * Each time a sched group cpu_power is changed,
139 * we must compute its reciprocal value
141 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
143 sg->__cpu_power += val;
144 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
148 static inline int rt_policy(int policy)
150 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
155 static inline int task_has_rt_policy(struct task_struct *p)
157 return rt_policy(p->policy);
161 * This is the priority-queue data structure of the RT scheduling class:
163 struct rt_prio_array {
164 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
165 struct list_head queue[MAX_RT_PRIO];
168 struct rt_bandwidth {
169 /* nests inside the rq lock: */
170 spinlock_t rt_runtime_lock;
173 struct hrtimer rt_period_timer;
176 static struct rt_bandwidth def_rt_bandwidth;
178 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
180 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
182 struct rt_bandwidth *rt_b =
183 container_of(timer, struct rt_bandwidth, rt_period_timer);
189 now = hrtimer_cb_get_time(timer);
190 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
195 idle = do_sched_rt_period_timer(rt_b, overrun);
198 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
202 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
204 rt_b->rt_period = ns_to_ktime(period);
205 rt_b->rt_runtime = runtime;
207 spin_lock_init(&rt_b->rt_runtime_lock);
209 hrtimer_init(&rt_b->rt_period_timer,
210 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
211 rt_b->rt_period_timer.function = sched_rt_period_timer;
214 static inline int rt_bandwidth_enabled(void)
216 return sysctl_sched_rt_runtime >= 0;
219 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
223 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
226 if (hrtimer_active(&rt_b->rt_period_timer))
229 spin_lock(&rt_b->rt_runtime_lock);
231 if (hrtimer_active(&rt_b->rt_period_timer))
234 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
235 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
236 hrtimer_start_expires(&rt_b->rt_period_timer,
239 spin_unlock(&rt_b->rt_runtime_lock);
242 #ifdef CONFIG_RT_GROUP_SCHED
243 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
245 hrtimer_cancel(&rt_b->rt_period_timer);
250 * sched_domains_mutex serializes calls to arch_init_sched_domains,
251 * detach_destroy_domains and partition_sched_domains.
253 static DEFINE_MUTEX(sched_domains_mutex);
255 #ifdef CONFIG_GROUP_SCHED
257 #include <linux/cgroup.h>
261 static LIST_HEAD(task_groups);
263 /* task group related information */
265 #ifdef CONFIG_CGROUP_SCHED
266 struct cgroup_subsys_state css;
269 #ifdef CONFIG_USER_SCHED
273 #ifdef CONFIG_FAIR_GROUP_SCHED
274 /* schedulable entities of this group on each cpu */
275 struct sched_entity **se;
276 /* runqueue "owned" by this group on each cpu */
277 struct cfs_rq **cfs_rq;
278 unsigned long shares;
281 #ifdef CONFIG_RT_GROUP_SCHED
282 struct sched_rt_entity **rt_se;
283 struct rt_rq **rt_rq;
285 struct rt_bandwidth rt_bandwidth;
289 struct list_head list;
291 struct task_group *parent;
292 struct list_head siblings;
293 struct list_head children;
296 #ifdef CONFIG_USER_SCHED
298 /* Helper function to pass uid information to create_sched_user() */
299 void set_tg_uid(struct user_struct *user)
301 user->tg->uid = user->uid;
306 * Every UID task group (including init_task_group aka UID-0) will
307 * be a child to this group.
309 struct task_group root_task_group;
311 #ifdef CONFIG_FAIR_GROUP_SCHED
312 /* Default task group's sched entity on each cpu */
313 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
314 /* Default task group's cfs_rq on each cpu */
315 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
316 #endif /* CONFIG_FAIR_GROUP_SCHED */
318 #ifdef CONFIG_RT_GROUP_SCHED
319 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
320 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
321 #endif /* CONFIG_RT_GROUP_SCHED */
322 #else /* !CONFIG_USER_SCHED */
323 #define root_task_group init_task_group
324 #endif /* CONFIG_USER_SCHED */
326 /* task_group_lock serializes add/remove of task groups and also changes to
327 * a task group's cpu shares.
329 static DEFINE_SPINLOCK(task_group_lock);
331 #ifdef CONFIG_FAIR_GROUP_SCHED
332 #ifdef CONFIG_USER_SCHED
333 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
334 #else /* !CONFIG_USER_SCHED */
335 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
336 #endif /* CONFIG_USER_SCHED */
339 * A weight of 0 or 1 can cause arithmetics problems.
340 * A weight of a cfs_rq is the sum of weights of which entities
341 * are queued on this cfs_rq, so a weight of a entity should not be
342 * too large, so as the shares value of a task group.
343 * (The default weight is 1024 - so there's no practical
344 * limitation from this.)
347 #define MAX_SHARES (1UL << 18)
349 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
352 /* Default task group.
353 * Every task in system belong to this group at bootup.
355 struct task_group init_task_group;
357 /* return group to which a task belongs */
358 static inline struct task_group *task_group(struct task_struct *p)
360 struct task_group *tg;
362 #ifdef CONFIG_USER_SCHED
364 tg = __task_cred(p)->user->tg;
366 #elif defined(CONFIG_CGROUP_SCHED)
367 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
368 struct task_group, css);
370 tg = &init_task_group;
375 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
376 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
378 #ifdef CONFIG_FAIR_GROUP_SCHED
379 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
380 p->se.parent = task_group(p)->se[cpu];
383 #ifdef CONFIG_RT_GROUP_SCHED
384 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
385 p->rt.parent = task_group(p)->rt_se[cpu];
391 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
392 static inline struct task_group *task_group(struct task_struct *p)
397 #endif /* CONFIG_GROUP_SCHED */
399 /* CFS-related fields in a runqueue */
401 struct load_weight load;
402 unsigned long nr_running;
407 struct rb_root tasks_timeline;
408 struct rb_node *rb_leftmost;
410 struct list_head tasks;
411 struct list_head *balance_iterator;
414 * 'curr' points to currently running entity on this cfs_rq.
415 * It is set to NULL otherwise (i.e when none are currently running).
417 struct sched_entity *curr, *next, *last;
419 unsigned int nr_spread_over;
421 #ifdef CONFIG_FAIR_GROUP_SCHED
422 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
425 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
426 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
427 * (like users, containers etc.)
429 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
430 * list is used during load balance.
432 struct list_head leaf_cfs_rq_list;
433 struct task_group *tg; /* group that "owns" this runqueue */
437 * the part of load.weight contributed by tasks
439 unsigned long task_weight;
442 * h_load = weight * f(tg)
444 * Where f(tg) is the recursive weight fraction assigned to
447 unsigned long h_load;
450 * this cpu's part of tg->shares
452 unsigned long shares;
455 * load.weight at the time we set shares
457 unsigned long rq_weight;
462 /* Real-Time classes' related field in a runqueue: */
464 struct rt_prio_array active;
465 unsigned long rt_nr_running;
466 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
467 int highest_prio; /* highest queued rt task prio */
470 unsigned long rt_nr_migratory;
476 /* Nests inside the rq lock: */
477 spinlock_t rt_runtime_lock;
479 #ifdef CONFIG_RT_GROUP_SCHED
480 unsigned long rt_nr_boosted;
483 struct list_head leaf_rt_rq_list;
484 struct task_group *tg;
485 struct sched_rt_entity *rt_se;
492 * We add the notion of a root-domain which will be used to define per-domain
493 * variables. Each exclusive cpuset essentially defines an island domain by
494 * fully partitioning the member cpus from any other cpuset. Whenever a new
495 * exclusive cpuset is created, we also create and attach a new root-domain
502 cpumask_var_t online;
505 * The "RT overload" flag: it gets set if a CPU has more than
506 * one runnable RT task.
508 cpumask_var_t rto_mask;
511 struct cpupri cpupri;
513 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
515 * Preferred wake up cpu nominated by sched_mc balance that will be
516 * used when most cpus are idle in the system indicating overall very
517 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
519 unsigned int sched_mc_preferred_wakeup_cpu;
524 * By default the system creates a single root-domain with all cpus as
525 * members (mimicking the global state we have today).
527 static struct root_domain def_root_domain;
532 * This is the main, per-CPU runqueue data structure.
534 * Locking rule: those places that want to lock multiple runqueues
535 * (such as the load balancing or the thread migration code), lock
536 * acquire operations must be ordered by ascending &runqueue.
543 * nr_running and cpu_load should be in the same cacheline because
544 * remote CPUs use both these fields when doing load calculation.
546 unsigned long nr_running;
547 #define CPU_LOAD_IDX_MAX 5
548 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
549 unsigned char idle_at_tick;
551 unsigned long last_tick_seen;
552 unsigned char in_nohz_recently;
554 /* capture load from *all* tasks on this cpu: */
555 struct load_weight load;
556 unsigned long nr_load_updates;
562 #ifdef CONFIG_FAIR_GROUP_SCHED
563 /* list of leaf cfs_rq on this cpu: */
564 struct list_head leaf_cfs_rq_list;
566 #ifdef CONFIG_RT_GROUP_SCHED
567 struct list_head leaf_rt_rq_list;
571 * This is part of a global counter where only the total sum
572 * over all CPUs matters. A task can increase this counter on
573 * one CPU and if it got migrated afterwards it may decrease
574 * it on another CPU. Always updated under the runqueue lock:
576 unsigned long nr_uninterruptible;
578 struct task_struct *curr, *idle;
579 unsigned long next_balance;
580 struct mm_struct *prev_mm;
587 struct root_domain *rd;
588 struct sched_domain *sd;
590 /* For active balancing */
593 /* cpu of this runqueue: */
597 unsigned long avg_load_per_task;
599 struct task_struct *migration_thread;
600 struct list_head migration_queue;
603 #ifdef CONFIG_SCHED_HRTICK
605 int hrtick_csd_pending;
606 struct call_single_data hrtick_csd;
608 struct hrtimer hrtick_timer;
611 #ifdef CONFIG_SCHEDSTATS
613 struct sched_info rq_sched_info;
614 unsigned long long rq_cpu_time;
615 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
617 /* sys_sched_yield() stats */
618 unsigned int yld_exp_empty;
619 unsigned int yld_act_empty;
620 unsigned int yld_both_empty;
621 unsigned int yld_count;
623 /* schedule() stats */
624 unsigned int sched_switch;
625 unsigned int sched_count;
626 unsigned int sched_goidle;
628 /* try_to_wake_up() stats */
629 unsigned int ttwu_count;
630 unsigned int ttwu_local;
633 unsigned int bkl_count;
637 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
639 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
641 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
644 static inline int cpu_of(struct rq *rq)
654 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
655 * See detach_destroy_domains: synchronize_sched for details.
657 * The domain tree of any CPU may only be accessed from within
658 * preempt-disabled sections.
660 #define for_each_domain(cpu, __sd) \
661 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
663 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
664 #define this_rq() (&__get_cpu_var(runqueues))
665 #define task_rq(p) cpu_rq(task_cpu(p))
666 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
668 inline void update_rq_clock(struct rq *rq)
670 rq->clock = sched_clock_cpu(cpu_of(rq));
674 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
676 #ifdef CONFIG_SCHED_DEBUG
677 # define const_debug __read_mostly
679 # define const_debug static const
685 * Returns true if the current cpu runqueue is locked.
686 * This interface allows printk to be called with the runqueue lock
687 * held and know whether or not it is OK to wake up the klogd.
689 int runqueue_is_locked(void)
692 struct rq *rq = cpu_rq(cpu);
695 ret = spin_is_locked(&rq->lock);
701 * Debugging: various feature bits
704 #define SCHED_FEAT(name, enabled) \
705 __SCHED_FEAT_##name ,
708 #include "sched_features.h"
713 #define SCHED_FEAT(name, enabled) \
714 (1UL << __SCHED_FEAT_##name) * enabled |
716 const_debug unsigned int sysctl_sched_features =
717 #include "sched_features.h"
722 #ifdef CONFIG_SCHED_DEBUG
723 #define SCHED_FEAT(name, enabled) \
726 static __read_mostly char *sched_feat_names[] = {
727 #include "sched_features.h"
733 static int sched_feat_show(struct seq_file *m, void *v)
737 for (i = 0; sched_feat_names[i]; i++) {
738 if (!(sysctl_sched_features & (1UL << i)))
740 seq_printf(m, "%s ", sched_feat_names[i]);
748 sched_feat_write(struct file *filp, const char __user *ubuf,
749 size_t cnt, loff_t *ppos)
759 if (copy_from_user(&buf, ubuf, cnt))
764 if (strncmp(buf, "NO_", 3) == 0) {
769 for (i = 0; sched_feat_names[i]; i++) {
770 int len = strlen(sched_feat_names[i]);
772 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
774 sysctl_sched_features &= ~(1UL << i);
776 sysctl_sched_features |= (1UL << i);
781 if (!sched_feat_names[i])
789 static int sched_feat_open(struct inode *inode, struct file *filp)
791 return single_open(filp, sched_feat_show, NULL);
794 static struct file_operations sched_feat_fops = {
795 .open = sched_feat_open,
796 .write = sched_feat_write,
799 .release = single_release,
802 static __init int sched_init_debug(void)
804 debugfs_create_file("sched_features", 0644, NULL, NULL,
809 late_initcall(sched_init_debug);
813 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
816 * Number of tasks to iterate in a single balance run.
817 * Limited because this is done with IRQs disabled.
819 const_debug unsigned int sysctl_sched_nr_migrate = 32;
822 * ratelimit for updating the group shares.
825 unsigned int sysctl_sched_shares_ratelimit = 250000;
828 * Inject some fuzzyness into changing the per-cpu group shares
829 * this avoids remote rq-locks at the expense of fairness.
832 unsigned int sysctl_sched_shares_thresh = 4;
835 * period over which we measure -rt task cpu usage in us.
838 unsigned int sysctl_sched_rt_period = 1000000;
840 static __read_mostly int scheduler_running;
843 * part of the period that we allow rt tasks to run in us.
846 int sysctl_sched_rt_runtime = 950000;
848 static inline u64 global_rt_period(void)
850 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
853 static inline u64 global_rt_runtime(void)
855 if (sysctl_sched_rt_runtime < 0)
858 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
861 #ifndef prepare_arch_switch
862 # define prepare_arch_switch(next) do { } while (0)
864 #ifndef finish_arch_switch
865 # define finish_arch_switch(prev) do { } while (0)
868 static inline int task_current(struct rq *rq, struct task_struct *p)
870 return rq->curr == p;
873 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
874 static inline int task_running(struct rq *rq, struct task_struct *p)
876 return task_current(rq, p);
879 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
883 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
885 #ifdef CONFIG_DEBUG_SPINLOCK
886 /* this is a valid case when another task releases the spinlock */
887 rq->lock.owner = current;
890 * If we are tracking spinlock dependencies then we have to
891 * fix up the runqueue lock - which gets 'carried over' from
894 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
896 spin_unlock_irq(&rq->lock);
899 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
900 static inline int task_running(struct rq *rq, struct task_struct *p)
905 return task_current(rq, p);
909 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
913 * We can optimise this out completely for !SMP, because the
914 * SMP rebalancing from interrupt is the only thing that cares
919 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
920 spin_unlock_irq(&rq->lock);
922 spin_unlock(&rq->lock);
926 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
930 * After ->oncpu is cleared, the task can be moved to a different CPU.
931 * We must ensure this doesn't happen until the switch is completely
937 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
941 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
944 * __task_rq_lock - lock the runqueue a given task resides on.
945 * Must be called interrupts disabled.
947 static inline struct rq *__task_rq_lock(struct task_struct *p)
951 struct rq *rq = task_rq(p);
952 spin_lock(&rq->lock);
953 if (likely(rq == task_rq(p)))
955 spin_unlock(&rq->lock);
960 * task_rq_lock - lock the runqueue a given task resides on and disable
961 * interrupts. Note the ordering: we can safely lookup the task_rq without
962 * explicitly disabling preemption.
964 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
970 local_irq_save(*flags);
972 spin_lock(&rq->lock);
973 if (likely(rq == task_rq(p)))
975 spin_unlock_irqrestore(&rq->lock, *flags);
979 void curr_rq_lock_irq_save(unsigned long *flags)
984 local_irq_save(*flags);
985 rq = cpu_rq(smp_processor_id());
986 spin_lock(&rq->lock);
989 void curr_rq_unlock_irq_restore(unsigned long *flags)
994 rq = cpu_rq(smp_processor_id());
995 spin_unlock(&rq->lock);
996 local_irq_restore(*flags);
999 void task_rq_unlock_wait(struct task_struct *p)
1001 struct rq *rq = task_rq(p);
1003 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1004 spin_unlock_wait(&rq->lock);
1007 static void __task_rq_unlock(struct rq *rq)
1008 __releases(rq->lock)
1010 spin_unlock(&rq->lock);
1013 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1014 __releases(rq->lock)
1016 spin_unlock_irqrestore(&rq->lock, *flags);
1020 * this_rq_lock - lock this runqueue and disable interrupts.
1022 static struct rq *this_rq_lock(void)
1023 __acquires(rq->lock)
1027 local_irq_disable();
1029 spin_lock(&rq->lock);
1034 #ifdef CONFIG_SCHED_HRTICK
1036 * Use HR-timers to deliver accurate preemption points.
1038 * Its all a bit involved since we cannot program an hrt while holding the
1039 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1042 * When we get rescheduled we reprogram the hrtick_timer outside of the
1048 * - enabled by features
1049 * - hrtimer is actually high res
1051 static inline int hrtick_enabled(struct rq *rq)
1053 if (!sched_feat(HRTICK))
1055 if (!cpu_active(cpu_of(rq)))
1057 return hrtimer_is_hres_active(&rq->hrtick_timer);
1060 static void hrtick_clear(struct rq *rq)
1062 if (hrtimer_active(&rq->hrtick_timer))
1063 hrtimer_cancel(&rq->hrtick_timer);
1067 * High-resolution timer tick.
1068 * Runs from hardirq context with interrupts disabled.
1070 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1072 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1074 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1076 spin_lock(&rq->lock);
1077 update_rq_clock(rq);
1078 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1079 spin_unlock(&rq->lock);
1081 return HRTIMER_NORESTART;
1086 * called from hardirq (IPI) context
1088 static void __hrtick_start(void *arg)
1090 struct rq *rq = arg;
1092 spin_lock(&rq->lock);
1093 hrtimer_restart(&rq->hrtick_timer);
1094 rq->hrtick_csd_pending = 0;
1095 spin_unlock(&rq->lock);
1099 * Called to set the hrtick timer state.
1101 * called with rq->lock held and irqs disabled
1103 static void hrtick_start(struct rq *rq, u64 delay)
1105 struct hrtimer *timer = &rq->hrtick_timer;
1106 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1108 hrtimer_set_expires(timer, time);
1110 if (rq == this_rq()) {
1111 hrtimer_restart(timer);
1112 } else if (!rq->hrtick_csd_pending) {
1113 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1114 rq->hrtick_csd_pending = 1;
1119 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1121 int cpu = (int)(long)hcpu;
1124 case CPU_UP_CANCELED:
1125 case CPU_UP_CANCELED_FROZEN:
1126 case CPU_DOWN_PREPARE:
1127 case CPU_DOWN_PREPARE_FROZEN:
1129 case CPU_DEAD_FROZEN:
1130 hrtick_clear(cpu_rq(cpu));
1137 static __init void init_hrtick(void)
1139 hotcpu_notifier(hotplug_hrtick, 0);
1143 * Called to set the hrtick timer state.
1145 * called with rq->lock held and irqs disabled
1147 static void hrtick_start(struct rq *rq, u64 delay)
1149 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1152 static inline void init_hrtick(void)
1155 #endif /* CONFIG_SMP */
1157 static void init_rq_hrtick(struct rq *rq)
1160 rq->hrtick_csd_pending = 0;
1162 rq->hrtick_csd.flags = 0;
1163 rq->hrtick_csd.func = __hrtick_start;
1164 rq->hrtick_csd.info = rq;
1167 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1168 rq->hrtick_timer.function = hrtick;
1170 #else /* CONFIG_SCHED_HRTICK */
1171 static inline void hrtick_clear(struct rq *rq)
1175 static inline void init_rq_hrtick(struct rq *rq)
1179 static inline void init_hrtick(void)
1182 #endif /* CONFIG_SCHED_HRTICK */
1185 * resched_task - mark a task 'to be rescheduled now'.
1187 * On UP this means the setting of the need_resched flag, on SMP it
1188 * might also involve a cross-CPU call to trigger the scheduler on
1193 #ifndef tsk_is_polling
1194 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1197 static void resched_task(struct task_struct *p)
1201 assert_spin_locked(&task_rq(p)->lock);
1203 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1206 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1209 if (cpu == smp_processor_id())
1212 /* NEED_RESCHED must be visible before we test polling */
1214 if (!tsk_is_polling(p))
1215 smp_send_reschedule(cpu);
1218 static void resched_cpu(int cpu)
1220 struct rq *rq = cpu_rq(cpu);
1221 unsigned long flags;
1223 if (!spin_trylock_irqsave(&rq->lock, flags))
1225 resched_task(cpu_curr(cpu));
1226 spin_unlock_irqrestore(&rq->lock, flags);
1231 * When add_timer_on() enqueues a timer into the timer wheel of an
1232 * idle CPU then this timer might expire before the next timer event
1233 * which is scheduled to wake up that CPU. In case of a completely
1234 * idle system the next event might even be infinite time into the
1235 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1236 * leaves the inner idle loop so the newly added timer is taken into
1237 * account when the CPU goes back to idle and evaluates the timer
1238 * wheel for the next timer event.
1240 void wake_up_idle_cpu(int cpu)
1242 struct rq *rq = cpu_rq(cpu);
1244 if (cpu == smp_processor_id())
1248 * This is safe, as this function is called with the timer
1249 * wheel base lock of (cpu) held. When the CPU is on the way
1250 * to idle and has not yet set rq->curr to idle then it will
1251 * be serialized on the timer wheel base lock and take the new
1252 * timer into account automatically.
1254 if (rq->curr != rq->idle)
1258 * We can set TIF_RESCHED on the idle task of the other CPU
1259 * lockless. The worst case is that the other CPU runs the
1260 * idle task through an additional NOOP schedule()
1262 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1264 /* NEED_RESCHED must be visible before we test polling */
1266 if (!tsk_is_polling(rq->idle))
1267 smp_send_reschedule(cpu);
1269 #endif /* CONFIG_NO_HZ */
1271 #else /* !CONFIG_SMP */
1272 static void resched_task(struct task_struct *p)
1274 assert_spin_locked(&task_rq(p)->lock);
1275 set_tsk_need_resched(p);
1277 #endif /* CONFIG_SMP */
1279 #if BITS_PER_LONG == 32
1280 # define WMULT_CONST (~0UL)
1282 # define WMULT_CONST (1UL << 32)
1285 #define WMULT_SHIFT 32
1288 * Shift right and round:
1290 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1293 * delta *= weight / lw
1295 static unsigned long
1296 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1297 struct load_weight *lw)
1301 if (!lw->inv_weight) {
1302 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1305 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1309 tmp = (u64)delta_exec * weight;
1311 * Check whether we'd overflow the 64-bit multiplication:
1313 if (unlikely(tmp > WMULT_CONST))
1314 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1317 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1319 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1322 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1328 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1335 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1336 * of tasks with abnormal "nice" values across CPUs the contribution that
1337 * each task makes to its run queue's load is weighted according to its
1338 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1339 * scaled version of the new time slice allocation that they receive on time
1343 #define WEIGHT_IDLEPRIO 2
1344 #define WMULT_IDLEPRIO (1 << 31)
1347 * Nice levels are multiplicative, with a gentle 10% change for every
1348 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1349 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1350 * that remained on nice 0.
1352 * The "10% effect" is relative and cumulative: from _any_ nice level,
1353 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1354 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1355 * If a task goes up by ~10% and another task goes down by ~10% then
1356 * the relative distance between them is ~25%.)
1358 static const int prio_to_weight[40] = {
1359 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1360 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1361 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1362 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1363 /* 0 */ 1024, 820, 655, 526, 423,
1364 /* 5 */ 335, 272, 215, 172, 137,
1365 /* 10 */ 110, 87, 70, 56, 45,
1366 /* 15 */ 36, 29, 23, 18, 15,
1370 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1372 * In cases where the weight does not change often, we can use the
1373 * precalculated inverse to speed up arithmetics by turning divisions
1374 * into multiplications:
1376 static const u32 prio_to_wmult[40] = {
1377 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1378 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1379 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1380 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1381 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1382 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1383 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1384 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1387 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1390 * runqueue iterator, to support SMP load-balancing between different
1391 * scheduling classes, without having to expose their internal data
1392 * structures to the load-balancing proper:
1394 struct rq_iterator {
1396 struct task_struct *(*start)(void *);
1397 struct task_struct *(*next)(void *);
1401 static unsigned long
1402 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1403 unsigned long max_load_move, struct sched_domain *sd,
1404 enum cpu_idle_type idle, int *all_pinned,
1405 int *this_best_prio, struct rq_iterator *iterator);
1408 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1409 struct sched_domain *sd, enum cpu_idle_type idle,
1410 struct rq_iterator *iterator);
1413 #ifdef CONFIG_CGROUP_CPUACCT
1414 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1416 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1419 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1421 update_load_add(&rq->load, load);
1424 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1426 update_load_sub(&rq->load, load);
1429 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1430 typedef int (*tg_visitor)(struct task_group *, void *);
1433 * Iterate the full tree, calling @down when first entering a node and @up when
1434 * leaving it for the final time.
1436 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1438 struct task_group *parent, *child;
1442 parent = &root_task_group;
1444 ret = (*down)(parent, data);
1447 list_for_each_entry_rcu(child, &parent->children, siblings) {
1454 ret = (*up)(parent, data);
1459 parent = parent->parent;
1468 static int tg_nop(struct task_group *tg, void *data)
1475 static unsigned long source_load(int cpu, int type);
1476 static unsigned long target_load(int cpu, int type);
1477 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1479 static unsigned long cpu_avg_load_per_task(int cpu)
1481 struct rq *rq = cpu_rq(cpu);
1482 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1485 rq->avg_load_per_task = rq->load.weight / nr_running;
1487 rq->avg_load_per_task = 0;
1489 return rq->avg_load_per_task;
1492 #ifdef CONFIG_FAIR_GROUP_SCHED
1494 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1497 * Calculate and set the cpu's group shares.
1500 update_group_shares_cpu(struct task_group *tg, int cpu,
1501 unsigned long sd_shares, unsigned long sd_rq_weight)
1503 unsigned long shares;
1504 unsigned long rq_weight;
1509 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1512 * \Sum shares * rq_weight
1513 * shares = -----------------------
1517 shares = (sd_shares * rq_weight) / sd_rq_weight;
1518 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1520 if (abs(shares - tg->se[cpu]->load.weight) >
1521 sysctl_sched_shares_thresh) {
1522 struct rq *rq = cpu_rq(cpu);
1523 unsigned long flags;
1525 spin_lock_irqsave(&rq->lock, flags);
1526 tg->cfs_rq[cpu]->shares = shares;
1528 __set_se_shares(tg->se[cpu], shares);
1529 spin_unlock_irqrestore(&rq->lock, flags);
1534 * Re-compute the task group their per cpu shares over the given domain.
1535 * This needs to be done in a bottom-up fashion because the rq weight of a
1536 * parent group depends on the shares of its child groups.
1538 static int tg_shares_up(struct task_group *tg, void *data)
1540 unsigned long weight, rq_weight = 0;
1541 unsigned long shares = 0;
1542 struct sched_domain *sd = data;
1545 for_each_cpu(i, sched_domain_span(sd)) {
1547 * If there are currently no tasks on the cpu pretend there
1548 * is one of average load so that when a new task gets to
1549 * run here it will not get delayed by group starvation.
1551 weight = tg->cfs_rq[i]->load.weight;
1553 weight = NICE_0_LOAD;
1555 tg->cfs_rq[i]->rq_weight = weight;
1556 rq_weight += weight;
1557 shares += tg->cfs_rq[i]->shares;
1560 if ((!shares && rq_weight) || shares > tg->shares)
1561 shares = tg->shares;
1563 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1564 shares = tg->shares;
1566 for_each_cpu(i, sched_domain_span(sd))
1567 update_group_shares_cpu(tg, i, shares, rq_weight);
1573 * Compute the cpu's hierarchical load factor for each task group.
1574 * This needs to be done in a top-down fashion because the load of a child
1575 * group is a fraction of its parents load.
1577 static int tg_load_down(struct task_group *tg, void *data)
1580 long cpu = (long)data;
1583 load = cpu_rq(cpu)->load.weight;
1585 load = tg->parent->cfs_rq[cpu]->h_load;
1586 load *= tg->cfs_rq[cpu]->shares;
1587 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1590 tg->cfs_rq[cpu]->h_load = load;
1595 static void update_shares(struct sched_domain *sd)
1597 u64 now = cpu_clock(raw_smp_processor_id());
1598 s64 elapsed = now - sd->last_update;
1600 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1601 sd->last_update = now;
1602 walk_tg_tree(tg_nop, tg_shares_up, sd);
1606 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1608 spin_unlock(&rq->lock);
1610 spin_lock(&rq->lock);
1613 static void update_h_load(long cpu)
1615 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1620 static inline void update_shares(struct sched_domain *sd)
1624 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1631 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1633 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1634 __releases(this_rq->lock)
1635 __acquires(busiest->lock)
1636 __acquires(this_rq->lock)
1640 if (unlikely(!irqs_disabled())) {
1641 /* printk() doesn't work good under rq->lock */
1642 spin_unlock(&this_rq->lock);
1645 if (unlikely(!spin_trylock(&busiest->lock))) {
1646 if (busiest < this_rq) {
1647 spin_unlock(&this_rq->lock);
1648 spin_lock(&busiest->lock);
1649 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1652 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1657 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1658 __releases(busiest->lock)
1660 spin_unlock(&busiest->lock);
1661 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1665 #ifdef CONFIG_FAIR_GROUP_SCHED
1666 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1669 cfs_rq->shares = shares;
1674 #include "sched_stats.h"
1675 #include "sched_idletask.c"
1676 #include "sched_fair.c"
1677 #include "sched_rt.c"
1678 #ifdef CONFIG_SCHED_DEBUG
1679 # include "sched_debug.c"
1682 #define sched_class_highest (&rt_sched_class)
1683 #define for_each_class(class) \
1684 for (class = sched_class_highest; class; class = class->next)
1686 static void inc_nr_running(struct rq *rq)
1691 static void dec_nr_running(struct rq *rq)
1696 static void set_load_weight(struct task_struct *p)
1698 if (task_has_rt_policy(p)) {
1699 p->se.load.weight = prio_to_weight[0] * 2;
1700 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1705 * SCHED_IDLE tasks get minimal weight:
1707 if (p->policy == SCHED_IDLE) {
1708 p->se.load.weight = WEIGHT_IDLEPRIO;
1709 p->se.load.inv_weight = WMULT_IDLEPRIO;
1713 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1714 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1717 static void update_avg(u64 *avg, u64 sample)
1719 s64 diff = sample - *avg;
1723 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1725 sched_info_queued(p);
1726 p->sched_class->enqueue_task(rq, p, wakeup);
1730 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1732 if (sleep && p->se.last_wakeup) {
1733 update_avg(&p->se.avg_overlap,
1734 p->se.sum_exec_runtime - p->se.last_wakeup);
1735 p->se.last_wakeup = 0;
1738 sched_info_dequeued(p);
1739 p->sched_class->dequeue_task(rq, p, sleep);
1744 * __normal_prio - return the priority that is based on the static prio
1746 static inline int __normal_prio(struct task_struct *p)
1748 return p->static_prio;
1752 * Calculate the expected normal priority: i.e. priority
1753 * without taking RT-inheritance into account. Might be
1754 * boosted by interactivity modifiers. Changes upon fork,
1755 * setprio syscalls, and whenever the interactivity
1756 * estimator recalculates.
1758 static inline int normal_prio(struct task_struct *p)
1762 if (task_has_rt_policy(p))
1763 prio = MAX_RT_PRIO-1 - p->rt_priority;
1765 prio = __normal_prio(p);
1770 * Calculate the current priority, i.e. the priority
1771 * taken into account by the scheduler. This value might
1772 * be boosted by RT tasks, or might be boosted by
1773 * interactivity modifiers. Will be RT if the task got
1774 * RT-boosted. If not then it returns p->normal_prio.
1776 static int effective_prio(struct task_struct *p)
1778 p->normal_prio = normal_prio(p);
1780 * If we are RT tasks or we were boosted to RT priority,
1781 * keep the priority unchanged. Otherwise, update priority
1782 * to the normal priority:
1784 if (!rt_prio(p->prio))
1785 return p->normal_prio;
1790 * activate_task - move a task to the runqueue.
1792 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1794 if (task_contributes_to_load(p))
1795 rq->nr_uninterruptible--;
1797 enqueue_task(rq, p, wakeup);
1802 * deactivate_task - remove a task from the runqueue.
1804 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1806 if (task_contributes_to_load(p))
1807 rq->nr_uninterruptible++;
1809 dequeue_task(rq, p, sleep);
1814 * task_curr - is this task currently executing on a CPU?
1815 * @p: the task in question.
1817 inline int task_curr(const struct task_struct *p)
1819 return cpu_curr(task_cpu(p)) == p;
1822 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1824 set_task_rq(p, cpu);
1827 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1828 * successfuly executed on another CPU. We must ensure that updates of
1829 * per-task data have been completed by this moment.
1832 task_thread_info(p)->cpu = cpu;
1836 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1837 const struct sched_class *prev_class,
1838 int oldprio, int running)
1840 if (prev_class != p->sched_class) {
1841 if (prev_class->switched_from)
1842 prev_class->switched_from(rq, p, running);
1843 p->sched_class->switched_to(rq, p, running);
1845 p->sched_class->prio_changed(rq, p, oldprio, running);
1850 /* Used instead of source_load when we know the type == 0 */
1851 static unsigned long weighted_cpuload(const int cpu)
1853 return cpu_rq(cpu)->load.weight;
1857 * Is this task likely cache-hot:
1860 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1865 * Buddy candidates are cache hot:
1867 if (sched_feat(CACHE_HOT_BUDDY) &&
1868 (&p->se == cfs_rq_of(&p->se)->next ||
1869 &p->se == cfs_rq_of(&p->se)->last))
1872 if (p->sched_class != &fair_sched_class)
1875 if (sysctl_sched_migration_cost == -1)
1877 if (sysctl_sched_migration_cost == 0)
1880 delta = now - p->se.exec_start;
1882 return delta < (s64)sysctl_sched_migration_cost;
1886 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1888 int old_cpu = task_cpu(p);
1889 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1890 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1891 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1894 clock_offset = old_rq->clock - new_rq->clock;
1896 trace_sched_migrate_task(p, task_cpu(p), new_cpu);
1898 #ifdef CONFIG_SCHEDSTATS
1899 if (p->se.wait_start)
1900 p->se.wait_start -= clock_offset;
1901 if (p->se.sleep_start)
1902 p->se.sleep_start -= clock_offset;
1903 if (p->se.block_start)
1904 p->se.block_start -= clock_offset;
1906 if (old_cpu != new_cpu) {
1907 p->se.nr_migrations++;
1908 #ifdef CONFIG_SCHEDSTATS
1909 if (task_hot(p, old_rq->clock, NULL))
1910 schedstat_inc(p, se.nr_forced2_migrations);
1913 p->se.vruntime -= old_cfsrq->min_vruntime -
1914 new_cfsrq->min_vruntime;
1916 __set_task_cpu(p, new_cpu);
1919 struct migration_req {
1920 struct list_head list;
1922 struct task_struct *task;
1925 struct completion done;
1929 * The task's runqueue lock must be held.
1930 * Returns true if you have to wait for migration thread.
1933 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1935 struct rq *rq = task_rq(p);
1938 * If the task is not on a runqueue (and not running), then
1939 * it is sufficient to simply update the task's cpu field.
1941 if (!p->se.on_rq && !task_running(rq, p)) {
1942 set_task_cpu(p, dest_cpu);
1946 init_completion(&req->done);
1948 req->dest_cpu = dest_cpu;
1949 list_add(&req->list, &rq->migration_queue);
1955 * wait_task_inactive - wait for a thread to unschedule.
1957 * If @match_state is nonzero, it's the @p->state value just checked and
1958 * not expected to change. If it changes, i.e. @p might have woken up,
1959 * then return zero. When we succeed in waiting for @p to be off its CPU,
1960 * we return a positive number (its total switch count). If a second call
1961 * a short while later returns the same number, the caller can be sure that
1962 * @p has remained unscheduled the whole time.
1964 * The caller must ensure that the task *will* unschedule sometime soon,
1965 * else this function might spin for a *long* time. This function can't
1966 * be called with interrupts off, or it may introduce deadlock with
1967 * smp_call_function() if an IPI is sent by the same process we are
1968 * waiting to become inactive.
1970 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1972 unsigned long flags;
1979 * We do the initial early heuristics without holding
1980 * any task-queue locks at all. We'll only try to get
1981 * the runqueue lock when things look like they will
1987 * If the task is actively running on another CPU
1988 * still, just relax and busy-wait without holding
1991 * NOTE! Since we don't hold any locks, it's not
1992 * even sure that "rq" stays as the right runqueue!
1993 * But we don't care, since "task_running()" will
1994 * return false if the runqueue has changed and p
1995 * is actually now running somewhere else!
1997 while (task_running(rq, p)) {
1998 if (match_state && unlikely(p->state != match_state))
2004 * Ok, time to look more closely! We need the rq
2005 * lock now, to be *sure*. If we're wrong, we'll
2006 * just go back and repeat.
2008 rq = task_rq_lock(p, &flags);
2009 trace_sched_wait_task(rq, p);
2010 running = task_running(rq, p);
2011 on_rq = p->se.on_rq;
2013 if (!match_state || p->state == match_state)
2014 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2015 task_rq_unlock(rq, &flags);
2018 * If it changed from the expected state, bail out now.
2020 if (unlikely(!ncsw))
2024 * Was it really running after all now that we
2025 * checked with the proper locks actually held?
2027 * Oops. Go back and try again..
2029 if (unlikely(running)) {
2035 * It's not enough that it's not actively running,
2036 * it must be off the runqueue _entirely_, and not
2039 * So if it wa still runnable (but just not actively
2040 * running right now), it's preempted, and we should
2041 * yield - it could be a while.
2043 if (unlikely(on_rq)) {
2044 schedule_timeout_uninterruptible(1);
2049 * Ahh, all good. It wasn't running, and it wasn't
2050 * runnable, which means that it will never become
2051 * running in the future either. We're all done!
2060 * kick_process - kick a running thread to enter/exit the kernel
2061 * @p: the to-be-kicked thread
2063 * Cause a process which is running on another CPU to enter
2064 * kernel-mode, without any delay. (to get signals handled.)
2066 * NOTE: this function doesnt have to take the runqueue lock,
2067 * because all it wants to ensure is that the remote task enters
2068 * the kernel. If the IPI races and the task has been migrated
2069 * to another CPU then no harm is done and the purpose has been
2072 void kick_process(struct task_struct *p)
2078 if ((cpu != smp_processor_id()) && task_curr(p))
2079 smp_send_reschedule(cpu);
2084 * Return a low guess at the load of a migration-source cpu weighted
2085 * according to the scheduling class and "nice" value.
2087 * We want to under-estimate the load of migration sources, to
2088 * balance conservatively.
2090 static unsigned long source_load(int cpu, int type)
2092 struct rq *rq = cpu_rq(cpu);
2093 unsigned long total = weighted_cpuload(cpu);
2095 if (type == 0 || !sched_feat(LB_BIAS))
2098 return min(rq->cpu_load[type-1], total);
2102 * Return a high guess at the load of a migration-target cpu weighted
2103 * according to the scheduling class and "nice" value.
2105 static unsigned long target_load(int cpu, int type)
2107 struct rq *rq = cpu_rq(cpu);
2108 unsigned long total = weighted_cpuload(cpu);
2110 if (type == 0 || !sched_feat(LB_BIAS))
2113 return max(rq->cpu_load[type-1], total);
2117 * find_idlest_group finds and returns the least busy CPU group within the
2120 static struct sched_group *
2121 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2123 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2124 unsigned long min_load = ULONG_MAX, this_load = 0;
2125 int load_idx = sd->forkexec_idx;
2126 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2129 unsigned long load, avg_load;
2133 /* Skip over this group if it has no CPUs allowed */
2134 if (!cpumask_intersects(sched_group_cpus(group),
2138 local_group = cpumask_test_cpu(this_cpu,
2139 sched_group_cpus(group));
2141 /* Tally up the load of all CPUs in the group */
2144 for_each_cpu(i, sched_group_cpus(group)) {
2145 /* Bias balancing toward cpus of our domain */
2147 load = source_load(i, load_idx);
2149 load = target_load(i, load_idx);
2154 /* Adjust by relative CPU power of the group */
2155 avg_load = sg_div_cpu_power(group,
2156 avg_load * SCHED_LOAD_SCALE);
2159 this_load = avg_load;
2161 } else if (avg_load < min_load) {
2162 min_load = avg_load;
2165 } while (group = group->next, group != sd->groups);
2167 if (!idlest || 100*this_load < imbalance*min_load)
2173 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2176 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2178 unsigned long load, min_load = ULONG_MAX;
2182 /* Traverse only the allowed CPUs */
2183 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2184 load = weighted_cpuload(i);
2186 if (load < min_load || (load == min_load && i == this_cpu)) {
2196 * sched_balance_self: balance the current task (running on cpu) in domains
2197 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2200 * Balance, ie. select the least loaded group.
2202 * Returns the target CPU number, or the same CPU if no balancing is needed.
2204 * preempt must be disabled.
2206 static int sched_balance_self(int cpu, int flag)
2208 struct task_struct *t = current;
2209 struct sched_domain *tmp, *sd = NULL;
2211 for_each_domain(cpu, tmp) {
2213 * If power savings logic is enabled for a domain, stop there.
2215 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2217 if (tmp->flags & flag)
2225 struct sched_group *group;
2226 int new_cpu, weight;
2228 if (!(sd->flags & flag)) {
2233 group = find_idlest_group(sd, t, cpu);
2239 new_cpu = find_idlest_cpu(group, t, cpu);
2240 if (new_cpu == -1 || new_cpu == cpu) {
2241 /* Now try balancing at a lower domain level of cpu */
2246 /* Now try balancing at a lower domain level of new_cpu */
2248 weight = cpumask_weight(sched_domain_span(sd));
2250 for_each_domain(cpu, tmp) {
2251 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2253 if (tmp->flags & flag)
2256 /* while loop will break here if sd == NULL */
2262 #endif /* CONFIG_SMP */
2265 * task_oncpu_function_call - call a function on the cpu on which a task runs
2266 * @p: the task to evaluate
2267 * @func: the function to be called
2268 * @info: the function call argument
2270 * Calls the function @func when the task is currently running. This might
2271 * be on the current CPU, which just calls the function directly
2273 void task_oncpu_function_call(struct task_struct *p,
2274 void (*func) (void *info), void *info)
2281 smp_call_function_single(cpu, func, info, 1);
2286 * try_to_wake_up - wake up a thread
2287 * @p: the to-be-woken-up thread
2288 * @state: the mask of task states that can be woken
2289 * @sync: do a synchronous wakeup?
2291 * Put it on the run-queue if it's not already there. The "current"
2292 * thread is always on the run-queue (except when the actual
2293 * re-schedule is in progress), and as such you're allowed to do
2294 * the simpler "current->state = TASK_RUNNING" to mark yourself
2295 * runnable without the overhead of this.
2297 * returns failure only if the task is already active.
2299 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2301 int cpu, orig_cpu, this_cpu, success = 0;
2302 unsigned long flags;
2306 if (!sched_feat(SYNC_WAKEUPS))
2310 if (sched_feat(LB_WAKEUP_UPDATE)) {
2311 struct sched_domain *sd;
2313 this_cpu = raw_smp_processor_id();
2316 for_each_domain(this_cpu, sd) {
2317 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2326 rq = task_rq_lock(p, &flags);
2327 update_rq_clock(rq);
2328 old_state = p->state;
2329 if (!(old_state & state))
2337 this_cpu = smp_processor_id();
2340 if (unlikely(task_running(rq, p)))
2343 cpu = p->sched_class->select_task_rq(p, sync);
2344 if (cpu != orig_cpu) {
2345 set_task_cpu(p, cpu);
2346 task_rq_unlock(rq, &flags);
2347 /* might preempt at this point */
2348 rq = task_rq_lock(p, &flags);
2349 old_state = p->state;
2350 if (!(old_state & state))
2355 this_cpu = smp_processor_id();
2359 #ifdef CONFIG_SCHEDSTATS
2360 schedstat_inc(rq, ttwu_count);
2361 if (cpu == this_cpu)
2362 schedstat_inc(rq, ttwu_local);
2364 struct sched_domain *sd;
2365 for_each_domain(this_cpu, sd) {
2366 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2367 schedstat_inc(sd, ttwu_wake_remote);
2372 #endif /* CONFIG_SCHEDSTATS */
2375 #endif /* CONFIG_SMP */
2376 schedstat_inc(p, se.nr_wakeups);
2378 schedstat_inc(p, se.nr_wakeups_sync);
2379 if (orig_cpu != cpu)
2380 schedstat_inc(p, se.nr_wakeups_migrate);
2381 if (cpu == this_cpu)
2382 schedstat_inc(p, se.nr_wakeups_local);
2384 schedstat_inc(p, se.nr_wakeups_remote);
2385 activate_task(rq, p, 1);
2389 trace_sched_wakeup(rq, p, success);
2390 check_preempt_curr(rq, p, sync);
2392 p->state = TASK_RUNNING;
2394 if (p->sched_class->task_wake_up)
2395 p->sched_class->task_wake_up(rq, p);
2398 current->se.last_wakeup = current->se.sum_exec_runtime;
2400 task_rq_unlock(rq, &flags);
2405 int wake_up_process(struct task_struct *p)
2407 return try_to_wake_up(p, TASK_ALL, 0);
2409 EXPORT_SYMBOL(wake_up_process);
2411 int wake_up_state(struct task_struct *p, unsigned int state)
2413 return try_to_wake_up(p, state, 0);
2417 * Perform scheduler related setup for a newly forked process p.
2418 * p is forked by current.
2420 * __sched_fork() is basic setup used by init_idle() too:
2422 static void __sched_fork(struct task_struct *p)
2424 p->se.exec_start = 0;
2425 p->se.sum_exec_runtime = 0;
2426 p->se.prev_sum_exec_runtime = 0;
2427 p->se.nr_migrations = 0;
2428 p->se.last_wakeup = 0;
2429 p->se.avg_overlap = 0;
2431 #ifdef CONFIG_SCHEDSTATS
2432 p->se.wait_start = 0;
2433 p->se.sum_sleep_runtime = 0;
2434 p->se.sleep_start = 0;
2435 p->se.block_start = 0;
2436 p->se.sleep_max = 0;
2437 p->se.block_max = 0;
2439 p->se.slice_max = 0;
2443 INIT_LIST_HEAD(&p->rt.run_list);
2445 INIT_LIST_HEAD(&p->se.group_node);
2447 #ifdef CONFIG_PREEMPT_NOTIFIERS
2448 INIT_HLIST_HEAD(&p->preempt_notifiers);
2452 * We mark the process as running here, but have not actually
2453 * inserted it onto the runqueue yet. This guarantees that
2454 * nobody will actually run it, and a signal or other external
2455 * event cannot wake it up and insert it on the runqueue either.
2457 p->state = TASK_RUNNING;
2461 * fork()/clone()-time setup:
2463 void sched_fork(struct task_struct *p, int clone_flags)
2465 int cpu = get_cpu();
2470 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2472 set_task_cpu(p, cpu);
2475 * Make sure we do not leak PI boosting priority to the child:
2477 p->prio = current->normal_prio;
2478 if (!rt_prio(p->prio))
2479 p->sched_class = &fair_sched_class;
2481 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2482 if (likely(sched_info_on()))
2483 memset(&p->sched_info, 0, sizeof(p->sched_info));
2485 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2488 #ifdef CONFIG_PREEMPT
2489 /* Want to start with kernel preemption disabled. */
2490 task_thread_info(p)->preempt_count = 1;
2496 * wake_up_new_task - wake up a newly created task for the first time.
2498 * This function will do some initial scheduler statistics housekeeping
2499 * that must be done for every newly created context, then puts the task
2500 * on the runqueue and wakes it.
2502 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2504 unsigned long flags;
2507 rq = task_rq_lock(p, &flags);
2508 BUG_ON(p->state != TASK_RUNNING);
2509 update_rq_clock(rq);
2511 p->prio = effective_prio(p);
2513 if (!p->sched_class->task_new || !current->se.on_rq) {
2514 activate_task(rq, p, 0);
2517 * Let the scheduling class do new task startup
2518 * management (if any):
2520 p->sched_class->task_new(rq, p);
2523 trace_sched_wakeup_new(rq, p, 1);
2524 check_preempt_curr(rq, p, 0);
2526 if (p->sched_class->task_wake_up)
2527 p->sched_class->task_wake_up(rq, p);
2529 task_rq_unlock(rq, &flags);
2532 #ifdef CONFIG_PREEMPT_NOTIFIERS
2535 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2536 * @notifier: notifier struct to register
2538 void preempt_notifier_register(struct preempt_notifier *notifier)
2540 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2542 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2545 * preempt_notifier_unregister - no longer interested in preemption notifications
2546 * @notifier: notifier struct to unregister
2548 * This is safe to call from within a preemption notifier.
2550 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2552 hlist_del(¬ifier->link);
2554 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2556 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2558 struct preempt_notifier *notifier;
2559 struct hlist_node *node;
2561 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2562 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2566 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2567 struct task_struct *next)
2569 struct preempt_notifier *notifier;
2570 struct hlist_node *node;
2572 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2573 notifier->ops->sched_out(notifier, next);
2576 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2578 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2583 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2584 struct task_struct *next)
2588 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2591 * prepare_task_switch - prepare to switch tasks
2592 * @rq: the runqueue preparing to switch
2593 * @prev: the current task that is being switched out
2594 * @next: the task we are going to switch to.
2596 * This is called with the rq lock held and interrupts off. It must
2597 * be paired with a subsequent finish_task_switch after the context
2600 * prepare_task_switch sets up locking and calls architecture specific
2604 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2605 struct task_struct *next)
2607 fire_sched_out_preempt_notifiers(prev, next);
2608 prepare_lock_switch(rq, next);
2609 prepare_arch_switch(next);
2613 * finish_task_switch - clean up after a task-switch
2614 * @rq: runqueue associated with task-switch
2615 * @prev: the thread we just switched away from.
2617 * finish_task_switch must be called after the context switch, paired
2618 * with a prepare_task_switch call before the context switch.
2619 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2620 * and do any other architecture-specific cleanup actions.
2622 * Note that we may have delayed dropping an mm in context_switch(). If
2623 * so, we finish that here outside of the runqueue lock. (Doing it
2624 * with the lock held can cause deadlocks; see schedule() for
2627 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2628 __releases(rq->lock)
2630 struct mm_struct *mm = rq->prev_mm;
2636 * A task struct has one reference for the use as "current".
2637 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2638 * schedule one last time. The schedule call will never return, and
2639 * the scheduled task must drop that reference.
2640 * The test for TASK_DEAD must occur while the runqueue locks are
2641 * still held, otherwise prev could be scheduled on another cpu, die
2642 * there before we look at prev->state, and then the reference would
2644 * Manfred Spraul <manfred@colorfullife.com>
2646 prev_state = prev->state;
2647 finish_arch_switch(prev);
2648 perf_counter_task_sched_in(current, cpu_of(rq));
2649 finish_lock_switch(rq, prev);
2651 if (current->sched_class->post_schedule)
2652 current->sched_class->post_schedule(rq);
2655 fire_sched_in_preempt_notifiers(current);
2658 if (unlikely(prev_state == TASK_DEAD)) {
2660 * Remove function-return probe instances associated with this
2661 * task and put them back on the free list.
2663 kprobe_flush_task(prev);
2664 put_task_struct(prev);
2669 * schedule_tail - first thing a freshly forked thread must call.
2670 * @prev: the thread we just switched away from.
2672 asmlinkage void schedule_tail(struct task_struct *prev)
2673 __releases(rq->lock)
2675 struct rq *rq = this_rq();
2677 finish_task_switch(rq, prev);
2678 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2679 /* In this case, finish_task_switch does not reenable preemption */
2682 if (current->set_child_tid)
2683 put_user(task_pid_vnr(current), current->set_child_tid);
2687 * context_switch - switch to the new MM and the new
2688 * thread's register state.
2691 context_switch(struct rq *rq, struct task_struct *prev,
2692 struct task_struct *next)
2694 struct mm_struct *mm, *oldmm;
2696 prepare_task_switch(rq, prev, next);
2697 trace_sched_switch(rq, prev, next);
2699 oldmm = prev->active_mm;
2701 * For paravirt, this is coupled with an exit in switch_to to
2702 * combine the page table reload and the switch backend into
2705 arch_enter_lazy_cpu_mode();
2707 if (unlikely(!mm)) {
2708 next->active_mm = oldmm;
2709 atomic_inc(&oldmm->mm_count);
2710 enter_lazy_tlb(oldmm, next);
2712 switch_mm(oldmm, mm, next);
2714 if (unlikely(!prev->mm)) {
2715 prev->active_mm = NULL;
2716 rq->prev_mm = oldmm;
2719 * Since the runqueue lock will be released by the next
2720 * task (which is an invalid locking op but in the case
2721 * of the scheduler it's an obvious special-case), so we
2722 * do an early lockdep release here:
2724 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2725 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2728 /* Here we just switch the register state and the stack. */
2729 switch_to(prev, next, prev);
2733 * this_rq must be evaluated again because prev may have moved
2734 * CPUs since it called schedule(), thus the 'rq' on its stack
2735 * frame will be invalid.
2737 finish_task_switch(this_rq(), prev);
2741 * nr_running, nr_uninterruptible and nr_context_switches:
2743 * externally visible scheduler statistics: current number of runnable
2744 * threads, current number of uninterruptible-sleeping threads, total
2745 * number of context switches performed since bootup.
2747 unsigned long nr_running(void)
2749 unsigned long i, sum = 0;
2751 for_each_online_cpu(i)
2752 sum += cpu_rq(i)->nr_running;
2757 unsigned long nr_uninterruptible(void)
2759 unsigned long i, sum = 0;
2761 for_each_possible_cpu(i)
2762 sum += cpu_rq(i)->nr_uninterruptible;
2765 * Since we read the counters lockless, it might be slightly
2766 * inaccurate. Do not allow it to go below zero though:
2768 if (unlikely((long)sum < 0))
2774 unsigned long long nr_context_switches(void)
2777 unsigned long long sum = 0;
2779 for_each_possible_cpu(i)
2780 sum += cpu_rq(i)->nr_switches;
2785 unsigned long nr_iowait(void)
2787 unsigned long i, sum = 0;
2789 for_each_possible_cpu(i)
2790 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2795 unsigned long nr_active(void)
2797 unsigned long i, running = 0, uninterruptible = 0;
2799 for_each_online_cpu(i) {
2800 running += cpu_rq(i)->nr_running;
2801 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2804 if (unlikely((long)uninterruptible < 0))
2805 uninterruptible = 0;
2807 return running + uninterruptible;
2811 * Update rq->cpu_load[] statistics. This function is usually called every
2812 * scheduler tick (TICK_NSEC).
2814 static void update_cpu_load(struct rq *this_rq)
2816 unsigned long this_load = this_rq->load.weight;
2819 this_rq->nr_load_updates++;
2821 /* Update our load: */
2822 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2823 unsigned long old_load, new_load;
2825 /* scale is effectively 1 << i now, and >> i divides by scale */
2827 old_load = this_rq->cpu_load[i];
2828 new_load = this_load;
2830 * Round up the averaging division if load is increasing. This
2831 * prevents us from getting stuck on 9 if the load is 10, for
2834 if (new_load > old_load)
2835 new_load += scale-1;
2836 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2843 * double_rq_lock - safely lock two runqueues
2845 * Note this does not disable interrupts like task_rq_lock,
2846 * you need to do so manually before calling.
2848 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2849 __acquires(rq1->lock)
2850 __acquires(rq2->lock)
2852 BUG_ON(!irqs_disabled());
2854 spin_lock(&rq1->lock);
2855 __acquire(rq2->lock); /* Fake it out ;) */
2858 spin_lock(&rq1->lock);
2859 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2861 spin_lock(&rq2->lock);
2862 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2865 update_rq_clock(rq1);
2866 update_rq_clock(rq2);
2870 * double_rq_unlock - safely unlock two runqueues
2872 * Note this does not restore interrupts like task_rq_unlock,
2873 * you need to do so manually after calling.
2875 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2876 __releases(rq1->lock)
2877 __releases(rq2->lock)
2879 spin_unlock(&rq1->lock);
2881 spin_unlock(&rq2->lock);
2883 __release(rq2->lock);
2887 * If dest_cpu is allowed for this process, migrate the task to it.
2888 * This is accomplished by forcing the cpu_allowed mask to only
2889 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2890 * the cpu_allowed mask is restored.
2892 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2894 struct migration_req req;
2895 unsigned long flags;
2898 rq = task_rq_lock(p, &flags);
2899 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
2900 || unlikely(!cpu_active(dest_cpu)))
2903 /* force the process onto the specified CPU */
2904 if (migrate_task(p, dest_cpu, &req)) {
2905 /* Need to wait for migration thread (might exit: take ref). */
2906 struct task_struct *mt = rq->migration_thread;
2908 get_task_struct(mt);
2909 task_rq_unlock(rq, &flags);
2910 wake_up_process(mt);
2911 put_task_struct(mt);
2912 wait_for_completion(&req.done);
2917 task_rq_unlock(rq, &flags);
2921 * sched_exec - execve() is a valuable balancing opportunity, because at
2922 * this point the task has the smallest effective memory and cache footprint.
2924 void sched_exec(void)
2926 int new_cpu, this_cpu = get_cpu();
2927 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2929 if (new_cpu != this_cpu)
2930 sched_migrate_task(current, new_cpu);
2934 * pull_task - move a task from a remote runqueue to the local runqueue.
2935 * Both runqueues must be locked.
2937 static void pull_task(struct rq *src_rq, struct task_struct *p,
2938 struct rq *this_rq, int this_cpu)
2940 deactivate_task(src_rq, p, 0);
2941 set_task_cpu(p, this_cpu);
2942 activate_task(this_rq, p, 0);
2944 * Note that idle threads have a prio of MAX_PRIO, for this test
2945 * to be always true for them.
2947 check_preempt_curr(this_rq, p, 0);
2951 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2954 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2955 struct sched_domain *sd, enum cpu_idle_type idle,
2959 * We do not migrate tasks that are:
2960 * 1) running (obviously), or
2961 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2962 * 3) are cache-hot on their current CPU.
2964 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
2965 schedstat_inc(p, se.nr_failed_migrations_affine);
2970 if (task_running(rq, p)) {
2971 schedstat_inc(p, se.nr_failed_migrations_running);
2976 * Aggressive migration if:
2977 * 1) task is cache cold, or
2978 * 2) too many balance attempts have failed.
2981 if (!task_hot(p, rq->clock, sd) ||
2982 sd->nr_balance_failed > sd->cache_nice_tries) {
2983 #ifdef CONFIG_SCHEDSTATS
2984 if (task_hot(p, rq->clock, sd)) {
2985 schedstat_inc(sd, lb_hot_gained[idle]);
2986 schedstat_inc(p, se.nr_forced_migrations);
2992 if (task_hot(p, rq->clock, sd)) {
2993 schedstat_inc(p, se.nr_failed_migrations_hot);
2999 static unsigned long
3000 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3001 unsigned long max_load_move, struct sched_domain *sd,
3002 enum cpu_idle_type idle, int *all_pinned,
3003 int *this_best_prio, struct rq_iterator *iterator)
3005 int loops = 0, pulled = 0, pinned = 0;
3006 struct task_struct *p;
3007 long rem_load_move = max_load_move;
3009 if (max_load_move == 0)
3015 * Start the load-balancing iterator:
3017 p = iterator->start(iterator->arg);
3019 if (!p || loops++ > sysctl_sched_nr_migrate)
3022 if ((p->se.load.weight >> 1) > rem_load_move ||
3023 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3024 p = iterator->next(iterator->arg);
3028 pull_task(busiest, p, this_rq, this_cpu);
3030 rem_load_move -= p->se.load.weight;
3033 * We only want to steal up to the prescribed amount of weighted load.
3035 if (rem_load_move > 0) {
3036 if (p->prio < *this_best_prio)
3037 *this_best_prio = p->prio;
3038 p = iterator->next(iterator->arg);
3043 * Right now, this is one of only two places pull_task() is called,
3044 * so we can safely collect pull_task() stats here rather than
3045 * inside pull_task().
3047 schedstat_add(sd, lb_gained[idle], pulled);
3050 *all_pinned = pinned;
3052 return max_load_move - rem_load_move;
3056 * move_tasks tries to move up to max_load_move weighted load from busiest to
3057 * this_rq, as part of a balancing operation within domain "sd".
3058 * Returns 1 if successful and 0 otherwise.
3060 * Called with both runqueues locked.
3062 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3063 unsigned long max_load_move,
3064 struct sched_domain *sd, enum cpu_idle_type idle,
3067 const struct sched_class *class = sched_class_highest;
3068 unsigned long total_load_moved = 0;
3069 int this_best_prio = this_rq->curr->prio;
3073 class->load_balance(this_rq, this_cpu, busiest,
3074 max_load_move - total_load_moved,
3075 sd, idle, all_pinned, &this_best_prio);
3076 class = class->next;
3078 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3081 } while (class && max_load_move > total_load_moved);
3083 return total_load_moved > 0;
3087 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3088 struct sched_domain *sd, enum cpu_idle_type idle,
3089 struct rq_iterator *iterator)
3091 struct task_struct *p = iterator->start(iterator->arg);
3095 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3096 pull_task(busiest, p, this_rq, this_cpu);
3098 * Right now, this is only the second place pull_task()
3099 * is called, so we can safely collect pull_task()
3100 * stats here rather than inside pull_task().
3102 schedstat_inc(sd, lb_gained[idle]);
3106 p = iterator->next(iterator->arg);
3113 * move_one_task tries to move exactly one task from busiest to this_rq, as
3114 * part of active balancing operations within "domain".
3115 * Returns 1 if successful and 0 otherwise.
3117 * Called with both runqueues locked.
3119 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3120 struct sched_domain *sd, enum cpu_idle_type idle)
3122 const struct sched_class *class;
3124 for (class = sched_class_highest; class; class = class->next)
3125 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3132 * find_busiest_group finds and returns the busiest CPU group within the
3133 * domain. It calculates and returns the amount of weighted load which
3134 * should be moved to restore balance via the imbalance parameter.
3136 static struct sched_group *
3137 find_busiest_group(struct sched_domain *sd, int this_cpu,
3138 unsigned long *imbalance, enum cpu_idle_type idle,
3139 int *sd_idle, const struct cpumask *cpus, int *balance)
3141 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3142 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3143 unsigned long max_pull;
3144 unsigned long busiest_load_per_task, busiest_nr_running;
3145 unsigned long this_load_per_task, this_nr_running;
3146 int load_idx, group_imb = 0;
3147 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3148 int power_savings_balance = 1;
3149 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3150 unsigned long min_nr_running = ULONG_MAX;
3151 struct sched_group *group_min = NULL, *group_leader = NULL;
3154 max_load = this_load = total_load = total_pwr = 0;
3155 busiest_load_per_task = busiest_nr_running = 0;
3156 this_load_per_task = this_nr_running = 0;
3158 if (idle == CPU_NOT_IDLE)
3159 load_idx = sd->busy_idx;
3160 else if (idle == CPU_NEWLY_IDLE)
3161 load_idx = sd->newidle_idx;
3163 load_idx = sd->idle_idx;
3166 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3169 int __group_imb = 0;
3170 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3171 unsigned long sum_nr_running, sum_weighted_load;
3172 unsigned long sum_avg_load_per_task;
3173 unsigned long avg_load_per_task;
3175 local_group = cpumask_test_cpu(this_cpu,
3176 sched_group_cpus(group));
3179 balance_cpu = cpumask_first(sched_group_cpus(group));
3181 /* Tally up the load of all CPUs in the group */
3182 sum_weighted_load = sum_nr_running = avg_load = 0;
3183 sum_avg_load_per_task = avg_load_per_task = 0;
3186 min_cpu_load = ~0UL;
3188 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3189 struct rq *rq = cpu_rq(i);
3191 if (*sd_idle && rq->nr_running)
3194 /* Bias balancing toward cpus of our domain */
3196 if (idle_cpu(i) && !first_idle_cpu) {
3201 load = target_load(i, load_idx);
3203 load = source_load(i, load_idx);
3204 if (load > max_cpu_load)
3205 max_cpu_load = load;
3206 if (min_cpu_load > load)
3207 min_cpu_load = load;
3211 sum_nr_running += rq->nr_running;
3212 sum_weighted_load += weighted_cpuload(i);
3214 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3218 * First idle cpu or the first cpu(busiest) in this sched group
3219 * is eligible for doing load balancing at this and above
3220 * domains. In the newly idle case, we will allow all the cpu's
3221 * to do the newly idle load balance.
3223 if (idle != CPU_NEWLY_IDLE && local_group &&
3224 balance_cpu != this_cpu && balance) {
3229 total_load += avg_load;
3230 total_pwr += group->__cpu_power;
3232 /* Adjust by relative CPU power of the group */
3233 avg_load = sg_div_cpu_power(group,
3234 avg_load * SCHED_LOAD_SCALE);
3238 * Consider the group unbalanced when the imbalance is larger
3239 * than the average weight of two tasks.
3241 * APZ: with cgroup the avg task weight can vary wildly and
3242 * might not be a suitable number - should we keep a
3243 * normalized nr_running number somewhere that negates
3246 avg_load_per_task = sg_div_cpu_power(group,
3247 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3249 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3252 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3255 this_load = avg_load;
3257 this_nr_running = sum_nr_running;
3258 this_load_per_task = sum_weighted_load;
3259 } else if (avg_load > max_load &&
3260 (sum_nr_running > group_capacity || __group_imb)) {
3261 max_load = avg_load;
3263 busiest_nr_running = sum_nr_running;
3264 busiest_load_per_task = sum_weighted_load;
3265 group_imb = __group_imb;
3268 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3270 * Busy processors will not participate in power savings
3273 if (idle == CPU_NOT_IDLE ||
3274 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3278 * If the local group is idle or completely loaded
3279 * no need to do power savings balance at this domain
3281 if (local_group && (this_nr_running >= group_capacity ||
3283 power_savings_balance = 0;
3286 * If a group is already running at full capacity or idle,
3287 * don't include that group in power savings calculations
3289 if (!power_savings_balance || sum_nr_running >= group_capacity
3294 * Calculate the group which has the least non-idle load.
3295 * This is the group from where we need to pick up the load
3298 if ((sum_nr_running < min_nr_running) ||
3299 (sum_nr_running == min_nr_running &&
3300 cpumask_first(sched_group_cpus(group)) >
3301 cpumask_first(sched_group_cpus(group_min)))) {
3303 min_nr_running = sum_nr_running;
3304 min_load_per_task = sum_weighted_load /
3309 * Calculate the group which is almost near its
3310 * capacity but still has some space to pick up some load
3311 * from other group and save more power
3313 if (sum_nr_running <= group_capacity - 1) {
3314 if (sum_nr_running > leader_nr_running ||
3315 (sum_nr_running == leader_nr_running &&
3316 cpumask_first(sched_group_cpus(group)) <
3317 cpumask_first(sched_group_cpus(group_leader)))) {
3318 group_leader = group;
3319 leader_nr_running = sum_nr_running;
3324 group = group->next;
3325 } while (group != sd->groups);
3327 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3330 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3332 if (this_load >= avg_load ||
3333 100*max_load <= sd->imbalance_pct*this_load)
3336 busiest_load_per_task /= busiest_nr_running;
3338 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3341 * We're trying to get all the cpus to the average_load, so we don't
3342 * want to push ourselves above the average load, nor do we wish to
3343 * reduce the max loaded cpu below the average load, as either of these
3344 * actions would just result in more rebalancing later, and ping-pong
3345 * tasks around. Thus we look for the minimum possible imbalance.
3346 * Negative imbalances (*we* are more loaded than anyone else) will
3347 * be counted as no imbalance for these purposes -- we can't fix that
3348 * by pulling tasks to us. Be careful of negative numbers as they'll
3349 * appear as very large values with unsigned longs.
3351 if (max_load <= busiest_load_per_task)
3355 * In the presence of smp nice balancing, certain scenarios can have
3356 * max load less than avg load(as we skip the groups at or below
3357 * its cpu_power, while calculating max_load..)
3359 if (max_load < avg_load) {
3361 goto small_imbalance;
3364 /* Don't want to pull so many tasks that a group would go idle */
3365 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3367 /* How much load to actually move to equalise the imbalance */
3368 *imbalance = min(max_pull * busiest->__cpu_power,
3369 (avg_load - this_load) * this->__cpu_power)
3373 * if *imbalance is less than the average load per runnable task
3374 * there is no gaurantee that any tasks will be moved so we'll have
3375 * a think about bumping its value to force at least one task to be
3378 if (*imbalance < busiest_load_per_task) {
3379 unsigned long tmp, pwr_now, pwr_move;
3383 pwr_move = pwr_now = 0;
3385 if (this_nr_running) {
3386 this_load_per_task /= this_nr_running;
3387 if (busiest_load_per_task > this_load_per_task)
3390 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3392 if (max_load - this_load + busiest_load_per_task >=
3393 busiest_load_per_task * imbn) {
3394 *imbalance = busiest_load_per_task;
3399 * OK, we don't have enough imbalance to justify moving tasks,
3400 * however we may be able to increase total CPU power used by
3404 pwr_now += busiest->__cpu_power *
3405 min(busiest_load_per_task, max_load);
3406 pwr_now += this->__cpu_power *
3407 min(this_load_per_task, this_load);
3408 pwr_now /= SCHED_LOAD_SCALE;
3410 /* Amount of load we'd subtract */
3411 tmp = sg_div_cpu_power(busiest,
3412 busiest_load_per_task * SCHED_LOAD_SCALE);
3414 pwr_move += busiest->__cpu_power *
3415 min(busiest_load_per_task, max_load - tmp);
3417 /* Amount of load we'd add */
3418 if (max_load * busiest->__cpu_power <
3419 busiest_load_per_task * SCHED_LOAD_SCALE)
3420 tmp = sg_div_cpu_power(this,
3421 max_load * busiest->__cpu_power);
3423 tmp = sg_div_cpu_power(this,
3424 busiest_load_per_task * SCHED_LOAD_SCALE);
3425 pwr_move += this->__cpu_power *
3426 min(this_load_per_task, this_load + tmp);
3427 pwr_move /= SCHED_LOAD_SCALE;
3429 /* Move if we gain throughput */
3430 if (pwr_move > pwr_now)
3431 *imbalance = busiest_load_per_task;
3437 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3438 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3441 if (this == group_leader && group_leader != group_min) {
3442 *imbalance = min_load_per_task;
3443 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3444 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3445 cpumask_first(sched_group_cpus(group_leader));
3456 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3459 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3460 unsigned long imbalance, const struct cpumask *cpus)
3462 struct rq *busiest = NULL, *rq;
3463 unsigned long max_load = 0;
3466 for_each_cpu(i, sched_group_cpus(group)) {
3469 if (!cpumask_test_cpu(i, cpus))
3473 wl = weighted_cpuload(i);
3475 if (rq->nr_running == 1 && wl > imbalance)
3478 if (wl > max_load) {
3488 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3489 * so long as it is large enough.
3491 #define MAX_PINNED_INTERVAL 512
3494 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3495 * tasks if there is an imbalance.
3497 static int load_balance(int this_cpu, struct rq *this_rq,
3498 struct sched_domain *sd, enum cpu_idle_type idle,
3499 int *balance, struct cpumask *cpus)
3501 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3502 struct sched_group *group;
3503 unsigned long imbalance;
3505 unsigned long flags;
3507 cpumask_setall(cpus);
3510 * When power savings policy is enabled for the parent domain, idle
3511 * sibling can pick up load irrespective of busy siblings. In this case,
3512 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3513 * portraying it as CPU_NOT_IDLE.
3515 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3516 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3519 schedstat_inc(sd, lb_count[idle]);
3523 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3530 schedstat_inc(sd, lb_nobusyg[idle]);
3534 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3536 schedstat_inc(sd, lb_nobusyq[idle]);
3540 BUG_ON(busiest == this_rq);
3542 schedstat_add(sd, lb_imbalance[idle], imbalance);
3545 if (busiest->nr_running > 1) {
3547 * Attempt to move tasks. If find_busiest_group has found
3548 * an imbalance but busiest->nr_running <= 1, the group is
3549 * still unbalanced. ld_moved simply stays zero, so it is
3550 * correctly treated as an imbalance.
3552 local_irq_save(flags);
3553 double_rq_lock(this_rq, busiest);
3554 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3555 imbalance, sd, idle, &all_pinned);
3556 double_rq_unlock(this_rq, busiest);
3557 local_irq_restore(flags);
3560 * some other cpu did the load balance for us.
3562 if (ld_moved && this_cpu != smp_processor_id())
3563 resched_cpu(this_cpu);
3565 /* All tasks on this runqueue were pinned by CPU affinity */
3566 if (unlikely(all_pinned)) {
3567 cpumask_clear_cpu(cpu_of(busiest), cpus);
3568 if (!cpumask_empty(cpus))
3575 schedstat_inc(sd, lb_failed[idle]);
3576 sd->nr_balance_failed++;
3578 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3580 spin_lock_irqsave(&busiest->lock, flags);
3582 /* don't kick the migration_thread, if the curr
3583 * task on busiest cpu can't be moved to this_cpu
3585 if (!cpumask_test_cpu(this_cpu,
3586 &busiest->curr->cpus_allowed)) {
3587 spin_unlock_irqrestore(&busiest->lock, flags);
3589 goto out_one_pinned;
3592 if (!busiest->active_balance) {
3593 busiest->active_balance = 1;
3594 busiest->push_cpu = this_cpu;
3597 spin_unlock_irqrestore(&busiest->lock, flags);
3599 wake_up_process(busiest->migration_thread);
3602 * We've kicked active balancing, reset the failure
3605 sd->nr_balance_failed = sd->cache_nice_tries+1;
3608 sd->nr_balance_failed = 0;
3610 if (likely(!active_balance)) {
3611 /* We were unbalanced, so reset the balancing interval */
3612 sd->balance_interval = sd->min_interval;
3615 * If we've begun active balancing, start to back off. This
3616 * case may not be covered by the all_pinned logic if there
3617 * is only 1 task on the busy runqueue (because we don't call
3620 if (sd->balance_interval < sd->max_interval)
3621 sd->balance_interval *= 2;
3624 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3625 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3631 schedstat_inc(sd, lb_balanced[idle]);
3633 sd->nr_balance_failed = 0;
3636 /* tune up the balancing interval */
3637 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3638 (sd->balance_interval < sd->max_interval))
3639 sd->balance_interval *= 2;
3641 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3642 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3653 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3654 * tasks if there is an imbalance.
3656 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3657 * this_rq is locked.
3660 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3661 struct cpumask *cpus)
3663 struct sched_group *group;
3664 struct rq *busiest = NULL;
3665 unsigned long imbalance;
3670 cpumask_setall(cpus);
3673 * When power savings policy is enabled for the parent domain, idle
3674 * sibling can pick up load irrespective of busy siblings. In this case,
3675 * let the state of idle sibling percolate up as IDLE, instead of
3676 * portraying it as CPU_NOT_IDLE.
3678 if (sd->flags & SD_SHARE_CPUPOWER &&
3679 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3682 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3684 update_shares_locked(this_rq, sd);
3685 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3686 &sd_idle, cpus, NULL);
3688 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3692 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3694 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3698 BUG_ON(busiest == this_rq);
3700 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3703 if (busiest->nr_running > 1) {
3704 /* Attempt to move tasks */
3705 double_lock_balance(this_rq, busiest);
3706 /* this_rq->clock is already updated */
3707 update_rq_clock(busiest);
3708 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3709 imbalance, sd, CPU_NEWLY_IDLE,
3711 double_unlock_balance(this_rq, busiest);
3713 if (unlikely(all_pinned)) {
3714 cpumask_clear_cpu(cpu_of(busiest), cpus);
3715 if (!cpumask_empty(cpus))
3721 int active_balance = 0;
3723 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3724 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3725 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3728 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3731 if (sd->nr_balance_failed++ < 2)
3735 * The only task running in a non-idle cpu can be moved to this
3736 * cpu in an attempt to completely freeup the other CPU
3737 * package. The same method used to move task in load_balance()
3738 * have been extended for load_balance_newidle() to speedup
3739 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
3741 * The package power saving logic comes from
3742 * find_busiest_group(). If there are no imbalance, then
3743 * f_b_g() will return NULL. However when sched_mc={1,2} then
3744 * f_b_g() will select a group from which a running task may be
3745 * pulled to this cpu in order to make the other package idle.
3746 * If there is no opportunity to make a package idle and if
3747 * there are no imbalance, then f_b_g() will return NULL and no
3748 * action will be taken in load_balance_newidle().
3750 * Under normal task pull operation due to imbalance, there
3751 * will be more than one task in the source run queue and
3752 * move_tasks() will succeed. ld_moved will be true and this
3753 * active balance code will not be triggered.
3756 /* Lock busiest in correct order while this_rq is held */
3757 double_lock_balance(this_rq, busiest);
3760 * don't kick the migration_thread, if the curr
3761 * task on busiest cpu can't be moved to this_cpu
3763 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
3764 double_unlock_balance(this_rq, busiest);
3769 if (!busiest->active_balance) {
3770 busiest->active_balance = 1;
3771 busiest->push_cpu = this_cpu;
3775 double_unlock_balance(this_rq, busiest);
3777 * Should not call ttwu while holding a rq->lock
3779 spin_unlock(&this_rq->lock);
3781 wake_up_process(busiest->migration_thread);
3782 spin_lock(&this_rq->lock);
3785 sd->nr_balance_failed = 0;
3787 update_shares_locked(this_rq, sd);
3791 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3792 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3793 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3795 sd->nr_balance_failed = 0;
3801 * idle_balance is called by schedule() if this_cpu is about to become
3802 * idle. Attempts to pull tasks from other CPUs.
3804 static void idle_balance(int this_cpu, struct rq *this_rq)
3806 struct sched_domain *sd;
3807 int pulled_task = 0;
3808 unsigned long next_balance = jiffies + HZ;
3809 cpumask_var_t tmpmask;
3811 if (!alloc_cpumask_var(&tmpmask, GFP_ATOMIC))
3814 for_each_domain(this_cpu, sd) {
3815 unsigned long interval;
3817 if (!(sd->flags & SD_LOAD_BALANCE))
3820 if (sd->flags & SD_BALANCE_NEWIDLE)
3821 /* If we've pulled tasks over stop searching: */
3822 pulled_task = load_balance_newidle(this_cpu, this_rq,
3825 interval = msecs_to_jiffies(sd->balance_interval);
3826 if (time_after(next_balance, sd->last_balance + interval))
3827 next_balance = sd->last_balance + interval;
3831 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3833 * We are going idle. next_balance may be set based on
3834 * a busy processor. So reset next_balance.
3836 this_rq->next_balance = next_balance;
3838 free_cpumask_var(tmpmask);
3842 * active_load_balance is run by migration threads. It pushes running tasks
3843 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3844 * running on each physical CPU where possible, and avoids physical /
3845 * logical imbalances.
3847 * Called with busiest_rq locked.
3849 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3851 int target_cpu = busiest_rq->push_cpu;
3852 struct sched_domain *sd;
3853 struct rq *target_rq;
3855 /* Is there any task to move? */
3856 if (busiest_rq->nr_running <= 1)
3859 target_rq = cpu_rq(target_cpu);
3862 * This condition is "impossible", if it occurs
3863 * we need to fix it. Originally reported by
3864 * Bjorn Helgaas on a 128-cpu setup.
3866 BUG_ON(busiest_rq == target_rq);
3868 /* move a task from busiest_rq to target_rq */
3869 double_lock_balance(busiest_rq, target_rq);
3870 update_rq_clock(busiest_rq);
3871 update_rq_clock(target_rq);
3873 /* Search for an sd spanning us and the target CPU. */
3874 for_each_domain(target_cpu, sd) {
3875 if ((sd->flags & SD_LOAD_BALANCE) &&
3876 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3881 schedstat_inc(sd, alb_count);
3883 if (move_one_task(target_rq, target_cpu, busiest_rq,
3885 schedstat_inc(sd, alb_pushed);
3887 schedstat_inc(sd, alb_failed);
3889 double_unlock_balance(busiest_rq, target_rq);
3894 atomic_t load_balancer;
3895 cpumask_var_t cpu_mask;
3896 } nohz ____cacheline_aligned = {
3897 .load_balancer = ATOMIC_INIT(-1),
3901 * This routine will try to nominate the ilb (idle load balancing)
3902 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3903 * load balancing on behalf of all those cpus. If all the cpus in the system
3904 * go into this tickless mode, then there will be no ilb owner (as there is
3905 * no need for one) and all the cpus will sleep till the next wakeup event
3908 * For the ilb owner, tick is not stopped. And this tick will be used
3909 * for idle load balancing. ilb owner will still be part of
3912 * While stopping the tick, this cpu will become the ilb owner if there
3913 * is no other owner. And will be the owner till that cpu becomes busy
3914 * or if all cpus in the system stop their ticks at which point
3915 * there is no need for ilb owner.
3917 * When the ilb owner becomes busy, it nominates another owner, during the
3918 * next busy scheduler_tick()
3920 int select_nohz_load_balancer(int stop_tick)
3922 int cpu = smp_processor_id();
3925 cpumask_set_cpu(cpu, nohz.cpu_mask);
3926 cpu_rq(cpu)->in_nohz_recently = 1;
3929 * If we are going offline and still the leader, give up!
3931 if (!cpu_active(cpu) &&
3932 atomic_read(&nohz.load_balancer) == cpu) {
3933 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3938 /* time for ilb owner also to sleep */
3939 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
3940 if (atomic_read(&nohz.load_balancer) == cpu)
3941 atomic_set(&nohz.load_balancer, -1);
3945 if (atomic_read(&nohz.load_balancer) == -1) {
3946 /* make me the ilb owner */
3947 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3949 } else if (atomic_read(&nohz.load_balancer) == cpu)
3952 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
3955 cpumask_clear_cpu(cpu, nohz.cpu_mask);
3957 if (atomic_read(&nohz.load_balancer) == cpu)
3958 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3965 static DEFINE_SPINLOCK(balancing);
3968 * It checks each scheduling domain to see if it is due to be balanced,
3969 * and initiates a balancing operation if so.
3971 * Balancing parameters are set up in arch_init_sched_domains.
3973 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3976 struct rq *rq = cpu_rq(cpu);
3977 unsigned long interval;
3978 struct sched_domain *sd;
3979 /* Earliest time when we have to do rebalance again */
3980 unsigned long next_balance = jiffies + 60*HZ;
3981 int update_next_balance = 0;
3985 /* Fails alloc? Rebalancing probably not a priority right now. */
3986 if (!alloc_cpumask_var(&tmp, GFP_ATOMIC))
3989 for_each_domain(cpu, sd) {
3990 if (!(sd->flags & SD_LOAD_BALANCE))
3993 interval = sd->balance_interval;
3994 if (idle != CPU_IDLE)
3995 interval *= sd->busy_factor;
3997 /* scale ms to jiffies */
3998 interval = msecs_to_jiffies(interval);
3999 if (unlikely(!interval))
4001 if (interval > HZ*NR_CPUS/10)
4002 interval = HZ*NR_CPUS/10;
4004 need_serialize = sd->flags & SD_SERIALIZE;
4006 if (need_serialize) {
4007 if (!spin_trylock(&balancing))
4011 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4012 if (load_balance(cpu, rq, sd, idle, &balance, tmp)) {
4014 * We've pulled tasks over so either we're no
4015 * longer idle, or one of our SMT siblings is
4018 idle = CPU_NOT_IDLE;
4020 sd->last_balance = jiffies;
4023 spin_unlock(&balancing);
4025 if (time_after(next_balance, sd->last_balance + interval)) {
4026 next_balance = sd->last_balance + interval;
4027 update_next_balance = 1;
4031 * Stop the load balance at this level. There is another
4032 * CPU in our sched group which is doing load balancing more
4040 * next_balance will be updated only when there is a need.
4041 * When the cpu is attached to null domain for ex, it will not be
4044 if (likely(update_next_balance))
4045 rq->next_balance = next_balance;
4047 free_cpumask_var(tmp);
4051 * run_rebalance_domains is triggered when needed from the scheduler tick.
4052 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4053 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4055 static void run_rebalance_domains(struct softirq_action *h)
4057 int this_cpu = smp_processor_id();
4058 struct rq *this_rq = cpu_rq(this_cpu);
4059 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4060 CPU_IDLE : CPU_NOT_IDLE;
4062 rebalance_domains(this_cpu, idle);
4066 * If this cpu is the owner for idle load balancing, then do the
4067 * balancing on behalf of the other idle cpus whose ticks are
4070 if (this_rq->idle_at_tick &&
4071 atomic_read(&nohz.load_balancer) == this_cpu) {
4075 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4076 if (balance_cpu == this_cpu)
4080 * If this cpu gets work to do, stop the load balancing
4081 * work being done for other cpus. Next load
4082 * balancing owner will pick it up.
4087 rebalance_domains(balance_cpu, CPU_IDLE);
4089 rq = cpu_rq(balance_cpu);
4090 if (time_after(this_rq->next_balance, rq->next_balance))
4091 this_rq->next_balance = rq->next_balance;
4098 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4100 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4101 * idle load balancing owner or decide to stop the periodic load balancing,
4102 * if the whole system is idle.
4104 static inline void trigger_load_balance(struct rq *rq, int cpu)
4108 * If we were in the nohz mode recently and busy at the current
4109 * scheduler tick, then check if we need to nominate new idle
4112 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4113 rq->in_nohz_recently = 0;
4115 if (atomic_read(&nohz.load_balancer) == cpu) {
4116 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4117 atomic_set(&nohz.load_balancer, -1);
4120 if (atomic_read(&nohz.load_balancer) == -1) {
4122 * simple selection for now: Nominate the
4123 * first cpu in the nohz list to be the next
4126 * TBD: Traverse the sched domains and nominate
4127 * the nearest cpu in the nohz.cpu_mask.
4129 int ilb = cpumask_first(nohz.cpu_mask);
4131 if (ilb < nr_cpu_ids)
4137 * If this cpu is idle and doing idle load balancing for all the
4138 * cpus with ticks stopped, is it time for that to stop?
4140 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4141 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4147 * If this cpu is idle and the idle load balancing is done by
4148 * someone else, then no need raise the SCHED_SOFTIRQ
4150 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4151 cpumask_test_cpu(cpu, nohz.cpu_mask))
4154 if (time_after_eq(jiffies, rq->next_balance))
4155 raise_softirq(SCHED_SOFTIRQ);
4158 #else /* CONFIG_SMP */
4161 * on UP we do not need to balance between CPUs:
4163 static inline void idle_balance(int cpu, struct rq *rq)
4169 DEFINE_PER_CPU(struct kernel_stat, kstat);
4171 EXPORT_PER_CPU_SYMBOL(kstat);
4174 * Return any ns on the sched_clock that have not yet been banked in
4175 * @p in case that task is currently running.
4177 unsigned long long __task_delta_exec(struct task_struct *p, int update)
4183 WARN_ON_ONCE(!runqueue_is_locked());
4184 WARN_ON_ONCE(!task_current(rq, p));
4187 update_rq_clock(rq);
4189 delta_exec = rq->clock - p->se.exec_start;
4191 WARN_ON_ONCE(delta_exec < 0);
4197 * Return any ns on the sched_clock that have not yet been banked in
4198 * @p in case that task is currently running.
4200 unsigned long long task_delta_exec(struct task_struct *p)
4202 unsigned long flags;
4206 rq = task_rq_lock(p, &flags);
4208 if (task_current(rq, p)) {
4211 update_rq_clock(rq);
4212 delta_exec = rq->clock - p->se.exec_start;
4213 if ((s64)delta_exec > 0)
4217 task_rq_unlock(rq, &flags);
4223 * Account user cpu time to a process.
4224 * @p: the process that the cpu time gets accounted to
4225 * @cputime: the cpu time spent in user space since the last update
4226 * @cputime_scaled: cputime scaled by cpu frequency
4228 void account_user_time(struct task_struct *p, cputime_t cputime,
4229 cputime_t cputime_scaled)
4231 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4234 /* Add user time to process. */
4235 p->utime = cputime_add(p->utime, cputime);
4236 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4237 account_group_user_time(p, cputime);
4239 /* Add user time to cpustat. */
4240 tmp = cputime_to_cputime64(cputime);
4241 if (TASK_NICE(p) > 0)
4242 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4244 cpustat->user = cputime64_add(cpustat->user, tmp);
4245 /* Account for user time used */
4246 acct_update_integrals(p);
4250 * Account guest cpu time to a process.
4251 * @p: the process that the cpu time gets accounted to
4252 * @cputime: the cpu time spent in virtual machine since the last update
4253 * @cputime_scaled: cputime scaled by cpu frequency
4255 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4256 cputime_t cputime_scaled)
4259 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4261 tmp = cputime_to_cputime64(cputime);
4263 /* Add guest time to process. */
4264 p->utime = cputime_add(p->utime, cputime);
4265 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4266 account_group_user_time(p, cputime);
4267 p->gtime = cputime_add(p->gtime, cputime);
4269 /* Add guest time to cpustat. */
4270 cpustat->user = cputime64_add(cpustat->user, tmp);
4271 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4275 * Account system cpu time to a process.
4276 * @p: the process that the cpu time gets accounted to
4277 * @hardirq_offset: the offset to subtract from hardirq_count()
4278 * @cputime: the cpu time spent in kernel space since the last update
4279 * @cputime_scaled: cputime scaled by cpu frequency
4281 void account_system_time(struct task_struct *p, int hardirq_offset,
4282 cputime_t cputime, cputime_t cputime_scaled)
4284 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4287 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4288 account_guest_time(p, cputime, cputime_scaled);
4292 /* Add system time to process. */
4293 p->stime = cputime_add(p->stime, cputime);
4294 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4295 account_group_system_time(p, cputime);
4297 /* Add system time to cpustat. */
4298 tmp = cputime_to_cputime64(cputime);
4299 if (hardirq_count() - hardirq_offset)
4300 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4301 else if (softirq_count())
4302 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4304 cpustat->system = cputime64_add(cpustat->system, tmp);
4306 /* Account for system time used */
4307 acct_update_integrals(p);
4311 * Account for involuntary wait time.
4312 * @steal: the cpu time spent in involuntary wait
4314 void account_steal_time(cputime_t cputime)
4316 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4317 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4319 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4323 * Account for idle time.
4324 * @cputime: the cpu time spent in idle wait
4326 void account_idle_time(cputime_t cputime)
4328 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4329 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4330 struct rq *rq = this_rq();
4332 if (atomic_read(&rq->nr_iowait) > 0)
4333 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4335 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4338 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4341 * Account a single tick of cpu time.
4342 * @p: the process that the cpu time gets accounted to
4343 * @user_tick: indicates if the tick is a user or a system tick
4345 void account_process_tick(struct task_struct *p, int user_tick)
4347 cputime_t one_jiffy = jiffies_to_cputime(1);
4348 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
4349 struct rq *rq = this_rq();
4352 account_user_time(p, one_jiffy, one_jiffy_scaled);
4353 else if (p != rq->idle)
4354 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
4357 account_idle_time(one_jiffy);
4361 * Account multiple ticks of steal time.
4362 * @p: the process from which the cpu time has been stolen
4363 * @ticks: number of stolen ticks
4365 void account_steal_ticks(unsigned long ticks)
4367 account_steal_time(jiffies_to_cputime(ticks));
4371 * Account multiple ticks of idle time.
4372 * @ticks: number of stolen ticks
4374 void account_idle_ticks(unsigned long ticks)
4376 account_idle_time(jiffies_to_cputime(ticks));
4382 * Use precise platform statistics if available:
4384 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4385 cputime_t task_utime(struct task_struct *p)
4390 cputime_t task_stime(struct task_struct *p)
4395 cputime_t task_utime(struct task_struct *p)
4397 clock_t utime = cputime_to_clock_t(p->utime),
4398 total = utime + cputime_to_clock_t(p->stime);
4402 * Use CFS's precise accounting:
4404 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4408 do_div(temp, total);
4410 utime = (clock_t)temp;
4412 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4413 return p->prev_utime;
4416 cputime_t task_stime(struct task_struct *p)
4421 * Use CFS's precise accounting. (we subtract utime from
4422 * the total, to make sure the total observed by userspace
4423 * grows monotonically - apps rely on that):
4425 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4426 cputime_to_clock_t(task_utime(p));
4429 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4431 return p->prev_stime;
4435 inline cputime_t task_gtime(struct task_struct *p)
4441 * This function gets called by the timer code, with HZ frequency.
4442 * We call it with interrupts disabled.
4444 * It also gets called by the fork code, when changing the parent's
4447 void scheduler_tick(void)
4449 int cpu = smp_processor_id();
4450 struct rq *rq = cpu_rq(cpu);
4451 struct task_struct *curr = rq->curr;
4455 spin_lock(&rq->lock);
4456 update_rq_clock(rq);
4457 update_cpu_load(rq);
4458 curr->sched_class->task_tick(rq, curr, 0);
4459 perf_counter_task_tick(curr, cpu);
4460 spin_unlock(&rq->lock);
4463 rq->idle_at_tick = idle_cpu(cpu);
4464 trigger_load_balance(rq, cpu);
4468 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4469 defined(CONFIG_PREEMPT_TRACER))
4471 static inline unsigned long get_parent_ip(unsigned long addr)
4473 if (in_lock_functions(addr)) {
4474 addr = CALLER_ADDR2;
4475 if (in_lock_functions(addr))
4476 addr = CALLER_ADDR3;
4481 void __kprobes add_preempt_count(int val)
4483 #ifdef CONFIG_DEBUG_PREEMPT
4487 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4490 preempt_count() += val;
4491 #ifdef CONFIG_DEBUG_PREEMPT
4493 * Spinlock count overflowing soon?
4495 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4498 if (preempt_count() == val)
4499 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4501 EXPORT_SYMBOL(add_preempt_count);
4503 void __kprobes sub_preempt_count(int val)
4505 #ifdef CONFIG_DEBUG_PREEMPT
4509 if (DEBUG_LOCKS_WARN_ON(val > preempt_count() - (!!kernel_locked())))
4512 * Is the spinlock portion underflowing?
4514 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4515 !(preempt_count() & PREEMPT_MASK)))
4519 if (preempt_count() == val)
4520 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4521 preempt_count() -= val;
4523 EXPORT_SYMBOL(sub_preempt_count);
4528 * Print scheduling while atomic bug:
4530 static noinline void __schedule_bug(struct task_struct *prev)
4532 struct pt_regs *regs = get_irq_regs();
4534 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4535 prev->comm, prev->pid, preempt_count());
4537 debug_show_held_locks(prev);
4539 if (irqs_disabled())
4540 print_irqtrace_events(prev);
4549 * Various schedule()-time debugging checks and statistics:
4551 static inline void schedule_debug(struct task_struct *prev)
4554 * Test if we are atomic. Since do_exit() needs to call into
4555 * schedule() atomically, we ignore that path for now.
4556 * Otherwise, whine if we are scheduling when we should not be.
4558 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4559 __schedule_bug(prev);
4561 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4563 schedstat_inc(this_rq(), sched_count);
4564 #ifdef CONFIG_SCHEDSTATS
4565 if (unlikely(prev->lock_depth >= 0)) {
4566 schedstat_inc(this_rq(), bkl_count);
4567 schedstat_inc(prev, sched_info.bkl_count);
4573 * Pick up the highest-prio task:
4575 static inline struct task_struct *
4576 pick_next_task(struct rq *rq, struct task_struct *prev)
4578 const struct sched_class *class;
4579 struct task_struct *p;
4582 * Optimization: we know that if all tasks are in
4583 * the fair class we can call that function directly:
4585 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4586 p = fair_sched_class.pick_next_task(rq);
4591 class = sched_class_highest;
4593 p = class->pick_next_task(rq);
4597 * Will never be NULL as the idle class always
4598 * returns a non-NULL p:
4600 class = class->next;
4605 * schedule() is the main scheduler function.
4607 asmlinkage void __sched schedule(void)
4609 struct task_struct *prev, *next;
4610 unsigned long *switch_count;
4616 cpu = smp_processor_id();
4620 switch_count = &prev->nivcsw;
4622 release_kernel_lock(prev);
4623 need_resched_nonpreemptible:
4625 schedule_debug(prev);
4627 if (sched_feat(HRTICK))
4630 spin_lock_irq(&rq->lock);
4631 update_rq_clock(rq);
4632 clear_tsk_need_resched(prev);
4634 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4635 if (unlikely(signal_pending_state(prev->state, prev)))
4636 prev->state = TASK_RUNNING;
4638 deactivate_task(rq, prev, 1);
4639 switch_count = &prev->nvcsw;
4643 if (prev->sched_class->pre_schedule)
4644 prev->sched_class->pre_schedule(rq, prev);
4647 if (unlikely(!rq->nr_running))
4648 idle_balance(cpu, rq);
4650 prev->sched_class->put_prev_task(rq, prev);
4651 next = pick_next_task(rq, prev);
4653 if (likely(prev != next)) {
4654 sched_info_switch(prev, next);
4655 perf_counter_task_sched_out(prev, cpu);
4661 context_switch(rq, prev, next); /* unlocks the rq */
4663 * the context switch might have flipped the stack from under
4664 * us, hence refresh the local variables.
4666 cpu = smp_processor_id();
4669 spin_unlock_irq(&rq->lock);
4671 if (unlikely(reacquire_kernel_lock(current) < 0))
4672 goto need_resched_nonpreemptible;
4674 preempt_enable_no_resched();
4675 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4678 EXPORT_SYMBOL(schedule);
4680 #ifdef CONFIG_PREEMPT
4682 * this is the entry point to schedule() from in-kernel preemption
4683 * off of preempt_enable. Kernel preemptions off return from interrupt
4684 * occur there and call schedule directly.
4686 asmlinkage void __sched preempt_schedule(void)
4688 struct thread_info *ti = current_thread_info();
4691 * If there is a non-zero preempt_count or interrupts are disabled,
4692 * we do not want to preempt the current task. Just return..
4694 if (likely(ti->preempt_count || irqs_disabled()))
4698 add_preempt_count(PREEMPT_ACTIVE);
4700 sub_preempt_count(PREEMPT_ACTIVE);
4703 * Check again in case we missed a preemption opportunity
4704 * between schedule and now.
4707 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4709 EXPORT_SYMBOL(preempt_schedule);
4712 * this is the entry point to schedule() from kernel preemption
4713 * off of irq context.
4714 * Note, that this is called and return with irqs disabled. This will
4715 * protect us against recursive calling from irq.
4717 asmlinkage void __sched preempt_schedule_irq(void)
4719 struct thread_info *ti = current_thread_info();
4721 /* Catch callers which need to be fixed */
4722 BUG_ON(ti->preempt_count || !irqs_disabled());
4725 add_preempt_count(PREEMPT_ACTIVE);
4728 local_irq_disable();
4729 sub_preempt_count(PREEMPT_ACTIVE);
4732 * Check again in case we missed a preemption opportunity
4733 * between schedule and now.
4736 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4739 #endif /* CONFIG_PREEMPT */
4741 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4744 return try_to_wake_up(curr->private, mode, sync);
4746 EXPORT_SYMBOL(default_wake_function);
4749 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4750 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4751 * number) then we wake all the non-exclusive tasks and one exclusive task.
4753 * There are circumstances in which we can try to wake a task which has already
4754 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4755 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4757 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4758 int nr_exclusive, int sync, void *key)
4760 wait_queue_t *curr, *next;
4762 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4763 unsigned flags = curr->flags;
4765 if (curr->func(curr, mode, sync, key) &&
4766 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4772 * __wake_up - wake up threads blocked on a waitqueue.
4774 * @mode: which threads
4775 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4776 * @key: is directly passed to the wakeup function
4778 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4779 int nr_exclusive, void *key)
4781 unsigned long flags;
4783 spin_lock_irqsave(&q->lock, flags);
4784 __wake_up_common(q, mode, nr_exclusive, 0, key);
4785 spin_unlock_irqrestore(&q->lock, flags);
4787 EXPORT_SYMBOL(__wake_up);
4790 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4792 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4794 __wake_up_common(q, mode, 1, 0, NULL);
4798 * __wake_up_sync - wake up threads blocked on a waitqueue.
4800 * @mode: which threads
4801 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4803 * The sync wakeup differs that the waker knows that it will schedule
4804 * away soon, so while the target thread will be woken up, it will not
4805 * be migrated to another CPU - ie. the two threads are 'synchronized'
4806 * with each other. This can prevent needless bouncing between CPUs.
4808 * On UP it can prevent extra preemption.
4811 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4813 unsigned long flags;
4819 if (unlikely(!nr_exclusive))
4822 spin_lock_irqsave(&q->lock, flags);
4823 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4824 spin_unlock_irqrestore(&q->lock, flags);
4826 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4829 * complete: - signals a single thread waiting on this completion
4830 * @x: holds the state of this particular completion
4832 * This will wake up a single thread waiting on this completion. Threads will be
4833 * awakened in the same order in which they were queued.
4835 * See also complete_all(), wait_for_completion() and related routines.
4837 void complete(struct completion *x)
4839 unsigned long flags;
4841 spin_lock_irqsave(&x->wait.lock, flags);
4843 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4844 spin_unlock_irqrestore(&x->wait.lock, flags);
4846 EXPORT_SYMBOL(complete);
4849 * complete_all: - signals all threads waiting on this completion
4850 * @x: holds the state of this particular completion
4852 * This will wake up all threads waiting on this particular completion event.
4854 void complete_all(struct completion *x)
4856 unsigned long flags;
4858 spin_lock_irqsave(&x->wait.lock, flags);
4859 x->done += UINT_MAX/2;
4860 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4861 spin_unlock_irqrestore(&x->wait.lock, flags);
4863 EXPORT_SYMBOL(complete_all);
4865 static inline long __sched
4866 do_wait_for_common(struct completion *x, long timeout, int state)
4869 DECLARE_WAITQUEUE(wait, current);
4871 wait.flags |= WQ_FLAG_EXCLUSIVE;
4872 __add_wait_queue_tail(&x->wait, &wait);
4874 if (signal_pending_state(state, current)) {
4875 timeout = -ERESTARTSYS;
4878 __set_current_state(state);
4879 spin_unlock_irq(&x->wait.lock);
4880 timeout = schedule_timeout(timeout);
4881 spin_lock_irq(&x->wait.lock);
4882 } while (!x->done && timeout);
4883 __remove_wait_queue(&x->wait, &wait);
4888 return timeout ?: 1;
4892 wait_for_common(struct completion *x, long timeout, int state)
4896 spin_lock_irq(&x->wait.lock);
4897 timeout = do_wait_for_common(x, timeout, state);
4898 spin_unlock_irq(&x->wait.lock);
4903 * wait_for_completion: - waits for completion of a task
4904 * @x: holds the state of this particular completion
4906 * This waits to be signaled for completion of a specific task. It is NOT
4907 * interruptible and there is no timeout.
4909 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4910 * and interrupt capability. Also see complete().
4912 void __sched wait_for_completion(struct completion *x)
4914 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4916 EXPORT_SYMBOL(wait_for_completion);
4919 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4920 * @x: holds the state of this particular completion
4921 * @timeout: timeout value in jiffies
4923 * This waits for either a completion of a specific task to be signaled or for a
4924 * specified timeout to expire. The timeout is in jiffies. It is not
4927 unsigned long __sched
4928 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4930 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4932 EXPORT_SYMBOL(wait_for_completion_timeout);
4935 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4936 * @x: holds the state of this particular completion
4938 * This waits for completion of a specific task to be signaled. It is
4941 int __sched wait_for_completion_interruptible(struct completion *x)
4943 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4944 if (t == -ERESTARTSYS)
4948 EXPORT_SYMBOL(wait_for_completion_interruptible);
4951 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4952 * @x: holds the state of this particular completion
4953 * @timeout: timeout value in jiffies
4955 * This waits for either a completion of a specific task to be signaled or for a
4956 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4958 unsigned long __sched
4959 wait_for_completion_interruptible_timeout(struct completion *x,
4960 unsigned long timeout)
4962 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4964 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4967 * wait_for_completion_killable: - waits for completion of a task (killable)
4968 * @x: holds the state of this particular completion
4970 * This waits to be signaled for completion of a specific task. It can be
4971 * interrupted by a kill signal.
4973 int __sched wait_for_completion_killable(struct completion *x)
4975 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4976 if (t == -ERESTARTSYS)
4980 EXPORT_SYMBOL(wait_for_completion_killable);
4983 * try_wait_for_completion - try to decrement a completion without blocking
4984 * @x: completion structure
4986 * Returns: 0 if a decrement cannot be done without blocking
4987 * 1 if a decrement succeeded.
4989 * If a completion is being used as a counting completion,
4990 * attempt to decrement the counter without blocking. This
4991 * enables us to avoid waiting if the resource the completion
4992 * is protecting is not available.
4994 bool try_wait_for_completion(struct completion *x)
4998 spin_lock_irq(&x->wait.lock);
5003 spin_unlock_irq(&x->wait.lock);
5006 EXPORT_SYMBOL(try_wait_for_completion);
5009 * completion_done - Test to see if a completion has any waiters
5010 * @x: completion structure
5012 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5013 * 1 if there are no waiters.
5016 bool completion_done(struct completion *x)
5020 spin_lock_irq(&x->wait.lock);
5023 spin_unlock_irq(&x->wait.lock);
5026 EXPORT_SYMBOL(completion_done);
5029 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5031 unsigned long flags;
5034 init_waitqueue_entry(&wait, current);
5036 __set_current_state(state);
5038 spin_lock_irqsave(&q->lock, flags);
5039 __add_wait_queue(q, &wait);
5040 spin_unlock(&q->lock);
5041 timeout = schedule_timeout(timeout);
5042 spin_lock_irq(&q->lock);
5043 __remove_wait_queue(q, &wait);
5044 spin_unlock_irqrestore(&q->lock, flags);
5049 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5051 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5053 EXPORT_SYMBOL(interruptible_sleep_on);
5056 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5058 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5060 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5062 void __sched sleep_on(wait_queue_head_t *q)
5064 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5066 EXPORT_SYMBOL(sleep_on);
5068 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5070 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5072 EXPORT_SYMBOL(sleep_on_timeout);
5074 #ifdef CONFIG_RT_MUTEXES
5077 * rt_mutex_setprio - set the current priority of a task
5079 * @prio: prio value (kernel-internal form)
5081 * This function changes the 'effective' priority of a task. It does
5082 * not touch ->normal_prio like __setscheduler().
5084 * Used by the rt_mutex code to implement priority inheritance logic.
5086 void rt_mutex_setprio(struct task_struct *p, int prio)
5088 unsigned long flags;
5089 int oldprio, on_rq, running;
5091 const struct sched_class *prev_class = p->sched_class;
5093 BUG_ON(prio < 0 || prio > MAX_PRIO);
5095 rq = task_rq_lock(p, &flags);
5096 update_rq_clock(rq);
5099 on_rq = p->se.on_rq;
5100 running = task_current(rq, p);
5102 dequeue_task(rq, p, 0);
5104 p->sched_class->put_prev_task(rq, p);
5107 p->sched_class = &rt_sched_class;
5109 p->sched_class = &fair_sched_class;
5114 p->sched_class->set_curr_task(rq);
5116 enqueue_task(rq, p, 0);
5118 check_class_changed(rq, p, prev_class, oldprio, running);
5120 task_rq_unlock(rq, &flags);
5125 void set_user_nice(struct task_struct *p, long nice)
5127 int old_prio, delta, on_rq;
5128 unsigned long flags;
5131 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5134 * We have to be careful, if called from sys_setpriority(),
5135 * the task might be in the middle of scheduling on another CPU.
5137 rq = task_rq_lock(p, &flags);
5138 update_rq_clock(rq);
5140 * The RT priorities are set via sched_setscheduler(), but we still
5141 * allow the 'normal' nice value to be set - but as expected
5142 * it wont have any effect on scheduling until the task is
5143 * SCHED_FIFO/SCHED_RR:
5145 if (task_has_rt_policy(p)) {
5146 p->static_prio = NICE_TO_PRIO(nice);
5149 on_rq = p->se.on_rq;
5151 dequeue_task(rq, p, 0);
5153 p->static_prio = NICE_TO_PRIO(nice);
5156 p->prio = effective_prio(p);
5157 delta = p->prio - old_prio;
5160 enqueue_task(rq, p, 0);
5162 * If the task increased its priority or is running and
5163 * lowered its priority, then reschedule its CPU:
5165 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5166 resched_task(rq->curr);
5169 task_rq_unlock(rq, &flags);
5171 EXPORT_SYMBOL(set_user_nice);
5174 * can_nice - check if a task can reduce its nice value
5178 int can_nice(const struct task_struct *p, const int nice)
5180 /* convert nice value [19,-20] to rlimit style value [1,40] */
5181 int nice_rlim = 20 - nice;
5183 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5184 capable(CAP_SYS_NICE));
5187 #ifdef __ARCH_WANT_SYS_NICE
5190 * sys_nice - change the priority of the current process.
5191 * @increment: priority increment
5193 * sys_setpriority is a more generic, but much slower function that
5194 * does similar things.
5196 asmlinkage long sys_nice(int increment)
5201 * Setpriority might change our priority at the same moment.
5202 * We don't have to worry. Conceptually one call occurs first
5203 * and we have a single winner.
5205 if (increment < -40)
5210 nice = PRIO_TO_NICE(current->static_prio) + increment;
5216 if (increment < 0 && !can_nice(current, nice))
5219 retval = security_task_setnice(current, nice);
5223 set_user_nice(current, nice);
5230 * task_prio - return the priority value of a given task.
5231 * @p: the task in question.
5233 * This is the priority value as seen by users in /proc.
5234 * RT tasks are offset by -200. Normal tasks are centered
5235 * around 0, value goes from -16 to +15.
5237 int task_prio(const struct task_struct *p)
5239 return p->prio - MAX_RT_PRIO;
5243 * task_nice - return the nice value of a given task.
5244 * @p: the task in question.
5246 int task_nice(const struct task_struct *p)
5248 return TASK_NICE(p);
5250 EXPORT_SYMBOL(task_nice);
5253 * idle_cpu - is a given cpu idle currently?
5254 * @cpu: the processor in question.
5256 int idle_cpu(int cpu)
5258 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5262 * idle_task - return the idle task for a given cpu.
5263 * @cpu: the processor in question.
5265 struct task_struct *idle_task(int cpu)
5267 return cpu_rq(cpu)->idle;
5271 * find_process_by_pid - find a process with a matching PID value.
5272 * @pid: the pid in question.
5274 static struct task_struct *find_process_by_pid(pid_t pid)
5276 return pid ? find_task_by_vpid(pid) : current;
5279 /* Actually do priority change: must hold rq lock. */
5281 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5283 BUG_ON(p->se.on_rq);
5286 switch (p->policy) {
5290 p->sched_class = &fair_sched_class;
5294 p->sched_class = &rt_sched_class;
5298 p->rt_priority = prio;
5299 p->normal_prio = normal_prio(p);
5300 /* we are holding p->pi_lock already */
5301 p->prio = rt_mutex_getprio(p);
5306 * check the target process has a UID that matches the current process's
5308 static bool check_same_owner(struct task_struct *p)
5310 const struct cred *cred = current_cred(), *pcred;
5314 pcred = __task_cred(p);
5315 match = (cred->euid == pcred->euid ||
5316 cred->euid == pcred->uid);
5321 static int __sched_setscheduler(struct task_struct *p, int policy,
5322 struct sched_param *param, bool user)
5324 int retval, oldprio, oldpolicy = -1, on_rq, running;
5325 unsigned long flags;
5326 const struct sched_class *prev_class = p->sched_class;
5329 /* may grab non-irq protected spin_locks */
5330 BUG_ON(in_interrupt());
5332 /* double check policy once rq lock held */
5334 policy = oldpolicy = p->policy;
5335 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5336 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5337 policy != SCHED_IDLE)
5340 * Valid priorities for SCHED_FIFO and SCHED_RR are
5341 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5342 * SCHED_BATCH and SCHED_IDLE is 0.
5344 if (param->sched_priority < 0 ||
5345 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5346 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5348 if (rt_policy(policy) != (param->sched_priority != 0))
5352 * Allow unprivileged RT tasks to decrease priority:
5354 if (user && !capable(CAP_SYS_NICE)) {
5355 if (rt_policy(policy)) {
5356 unsigned long rlim_rtprio;
5358 if (!lock_task_sighand(p, &flags))
5360 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5361 unlock_task_sighand(p, &flags);
5363 /* can't set/change the rt policy */
5364 if (policy != p->policy && !rlim_rtprio)
5367 /* can't increase priority */
5368 if (param->sched_priority > p->rt_priority &&
5369 param->sched_priority > rlim_rtprio)
5373 * Like positive nice levels, dont allow tasks to
5374 * move out of SCHED_IDLE either:
5376 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5379 /* can't change other user's priorities */
5380 if (!check_same_owner(p))
5385 #ifdef CONFIG_RT_GROUP_SCHED
5387 * Do not allow realtime tasks into groups that have no runtime
5390 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5391 task_group(p)->rt_bandwidth.rt_runtime == 0)
5395 retval = security_task_setscheduler(p, policy, param);
5401 * make sure no PI-waiters arrive (or leave) while we are
5402 * changing the priority of the task:
5404 spin_lock_irqsave(&p->pi_lock, flags);
5406 * To be able to change p->policy safely, the apropriate
5407 * runqueue lock must be held.
5409 rq = __task_rq_lock(p);
5410 /* recheck policy now with rq lock held */
5411 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5412 policy = oldpolicy = -1;
5413 __task_rq_unlock(rq);
5414 spin_unlock_irqrestore(&p->pi_lock, flags);
5417 update_rq_clock(rq);
5418 on_rq = p->se.on_rq;
5419 running = task_current(rq, p);
5421 deactivate_task(rq, p, 0);
5423 p->sched_class->put_prev_task(rq, p);
5426 __setscheduler(rq, p, policy, param->sched_priority);
5429 p->sched_class->set_curr_task(rq);
5431 activate_task(rq, p, 0);
5433 check_class_changed(rq, p, prev_class, oldprio, running);
5435 __task_rq_unlock(rq);
5436 spin_unlock_irqrestore(&p->pi_lock, flags);
5438 rt_mutex_adjust_pi(p);
5444 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5445 * @p: the task in question.
5446 * @policy: new policy.
5447 * @param: structure containing the new RT priority.
5449 * NOTE that the task may be already dead.
5451 int sched_setscheduler(struct task_struct *p, int policy,
5452 struct sched_param *param)
5454 return __sched_setscheduler(p, policy, param, true);
5456 EXPORT_SYMBOL_GPL(sched_setscheduler);
5459 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5460 * @p: the task in question.
5461 * @policy: new policy.
5462 * @param: structure containing the new RT priority.
5464 * Just like sched_setscheduler, only don't bother checking if the
5465 * current context has permission. For example, this is needed in
5466 * stop_machine(): we create temporary high priority worker threads,
5467 * but our caller might not have that capability.
5469 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5470 struct sched_param *param)
5472 return __sched_setscheduler(p, policy, param, false);
5476 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5478 struct sched_param lparam;
5479 struct task_struct *p;
5482 if (!param || pid < 0)
5484 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5489 p = find_process_by_pid(pid);
5491 retval = sched_setscheduler(p, policy, &lparam);
5498 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5499 * @pid: the pid in question.
5500 * @policy: new policy.
5501 * @param: structure containing the new RT priority.
5504 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5506 /* negative values for policy are not valid */
5510 return do_sched_setscheduler(pid, policy, param);
5514 * sys_sched_setparam - set/change the RT priority of a thread
5515 * @pid: the pid in question.
5516 * @param: structure containing the new RT priority.
5518 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5520 return do_sched_setscheduler(pid, -1, param);
5524 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5525 * @pid: the pid in question.
5527 asmlinkage long sys_sched_getscheduler(pid_t pid)
5529 struct task_struct *p;
5536 read_lock(&tasklist_lock);
5537 p = find_process_by_pid(pid);
5539 retval = security_task_getscheduler(p);
5543 read_unlock(&tasklist_lock);
5548 * sys_sched_getscheduler - get the RT priority of a thread
5549 * @pid: the pid in question.
5550 * @param: structure containing the RT priority.
5552 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5554 struct sched_param lp;
5555 struct task_struct *p;
5558 if (!param || pid < 0)
5561 read_lock(&tasklist_lock);
5562 p = find_process_by_pid(pid);
5567 retval = security_task_getscheduler(p);
5571 lp.sched_priority = p->rt_priority;
5572 read_unlock(&tasklist_lock);
5575 * This one might sleep, we cannot do it with a spinlock held ...
5577 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5582 read_unlock(&tasklist_lock);
5586 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5588 cpumask_var_t cpus_allowed, new_mask;
5589 struct task_struct *p;
5593 read_lock(&tasklist_lock);
5595 p = find_process_by_pid(pid);
5597 read_unlock(&tasklist_lock);
5603 * It is not safe to call set_cpus_allowed with the
5604 * tasklist_lock held. We will bump the task_struct's
5605 * usage count and then drop tasklist_lock.
5608 read_unlock(&tasklist_lock);
5610 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5614 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5616 goto out_free_cpus_allowed;
5619 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5622 retval = security_task_setscheduler(p, 0, NULL);
5626 cpuset_cpus_allowed(p, cpus_allowed);
5627 cpumask_and(new_mask, in_mask, cpus_allowed);
5629 retval = set_cpus_allowed_ptr(p, new_mask);
5632 cpuset_cpus_allowed(p, cpus_allowed);
5633 if (!cpumask_subset(new_mask, cpus_allowed)) {
5635 * We must have raced with a concurrent cpuset
5636 * update. Just reset the cpus_allowed to the
5637 * cpuset's cpus_allowed
5639 cpumask_copy(new_mask, cpus_allowed);
5644 free_cpumask_var(new_mask);
5645 out_free_cpus_allowed:
5646 free_cpumask_var(cpus_allowed);
5653 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5654 struct cpumask *new_mask)
5656 if (len < cpumask_size())
5657 cpumask_clear(new_mask);
5658 else if (len > cpumask_size())
5659 len = cpumask_size();
5661 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5665 * sys_sched_setaffinity - set the cpu affinity of a process
5666 * @pid: pid of the process
5667 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5668 * @user_mask_ptr: user-space pointer to the new cpu mask
5670 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5671 unsigned long __user *user_mask_ptr)
5673 cpumask_var_t new_mask;
5676 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5679 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5681 retval = sched_setaffinity(pid, new_mask);
5682 free_cpumask_var(new_mask);
5686 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5688 struct task_struct *p;
5692 read_lock(&tasklist_lock);
5695 p = find_process_by_pid(pid);
5699 retval = security_task_getscheduler(p);
5703 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5706 read_unlock(&tasklist_lock);
5713 * sys_sched_getaffinity - get the cpu affinity of a process
5714 * @pid: pid of the process
5715 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5716 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5718 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5719 unsigned long __user *user_mask_ptr)
5724 if (len < cpumask_size())
5727 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5730 ret = sched_getaffinity(pid, mask);
5732 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
5735 ret = cpumask_size();
5737 free_cpumask_var(mask);
5743 * sys_sched_yield - yield the current processor to other threads.
5745 * This function yields the current CPU to other tasks. If there are no
5746 * other threads running on this CPU then this function will return.
5748 asmlinkage long sys_sched_yield(void)
5750 struct rq *rq = this_rq_lock();
5752 schedstat_inc(rq, yld_count);
5753 current->sched_class->yield_task(rq);
5756 * Since we are going to call schedule() anyway, there's
5757 * no need to preempt or enable interrupts:
5759 __release(rq->lock);
5760 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5761 _raw_spin_unlock(&rq->lock);
5762 preempt_enable_no_resched();
5769 static void __cond_resched(void)
5771 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5772 __might_sleep(__FILE__, __LINE__);
5775 * The BKS might be reacquired before we have dropped
5776 * PREEMPT_ACTIVE, which could trigger a second
5777 * cond_resched() call.
5780 add_preempt_count(PREEMPT_ACTIVE);
5782 sub_preempt_count(PREEMPT_ACTIVE);
5783 } while (need_resched());
5786 int __sched _cond_resched(void)
5788 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5789 system_state == SYSTEM_RUNNING) {
5795 EXPORT_SYMBOL(_cond_resched);
5798 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5799 * call schedule, and on return reacquire the lock.
5801 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5802 * operations here to prevent schedule() from being called twice (once via
5803 * spin_unlock(), once by hand).
5805 int cond_resched_lock(spinlock_t *lock)
5807 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5810 if (spin_needbreak(lock) || resched) {
5812 if (resched && need_resched())
5821 EXPORT_SYMBOL(cond_resched_lock);
5823 int __sched cond_resched_softirq(void)
5825 BUG_ON(!in_softirq());
5827 if (need_resched() && system_state == SYSTEM_RUNNING) {
5835 EXPORT_SYMBOL(cond_resched_softirq);
5838 * yield - yield the current processor to other threads.
5840 * This is a shortcut for kernel-space yielding - it marks the
5841 * thread runnable and calls sys_sched_yield().
5843 void __sched yield(void)
5845 set_current_state(TASK_RUNNING);
5848 EXPORT_SYMBOL(yield);
5851 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5852 * that process accounting knows that this is a task in IO wait state.
5854 * But don't do that if it is a deliberate, throttling IO wait (this task
5855 * has set its backing_dev_info: the queue against which it should throttle)
5857 void __sched io_schedule(void)
5859 struct rq *rq = &__raw_get_cpu_var(runqueues);
5861 delayacct_blkio_start();
5862 atomic_inc(&rq->nr_iowait);
5864 atomic_dec(&rq->nr_iowait);
5865 delayacct_blkio_end();
5867 EXPORT_SYMBOL(io_schedule);
5869 long __sched io_schedule_timeout(long timeout)
5871 struct rq *rq = &__raw_get_cpu_var(runqueues);
5874 delayacct_blkio_start();
5875 atomic_inc(&rq->nr_iowait);
5876 ret = schedule_timeout(timeout);
5877 atomic_dec(&rq->nr_iowait);
5878 delayacct_blkio_end();
5883 * sys_sched_get_priority_max - return maximum RT priority.
5884 * @policy: scheduling class.
5886 * this syscall returns the maximum rt_priority that can be used
5887 * by a given scheduling class.
5889 asmlinkage long sys_sched_get_priority_max(int policy)
5896 ret = MAX_USER_RT_PRIO-1;
5908 * sys_sched_get_priority_min - return minimum RT priority.
5909 * @policy: scheduling class.
5911 * this syscall returns the minimum rt_priority that can be used
5912 * by a given scheduling class.
5914 asmlinkage long sys_sched_get_priority_min(int policy)
5932 * sys_sched_rr_get_interval - return the default timeslice of a process.
5933 * @pid: pid of the process.
5934 * @interval: userspace pointer to the timeslice value.
5936 * this syscall writes the default timeslice value of a given process
5937 * into the user-space timespec buffer. A value of '0' means infinity.
5940 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5942 struct task_struct *p;
5943 unsigned int time_slice;
5951 read_lock(&tasklist_lock);
5952 p = find_process_by_pid(pid);
5956 retval = security_task_getscheduler(p);
5961 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5962 * tasks that are on an otherwise idle runqueue:
5965 if (p->policy == SCHED_RR) {
5966 time_slice = DEF_TIMESLICE;
5967 } else if (p->policy != SCHED_FIFO) {
5968 struct sched_entity *se = &p->se;
5969 unsigned long flags;
5972 rq = task_rq_lock(p, &flags);
5973 if (rq->cfs.load.weight)
5974 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5975 task_rq_unlock(rq, &flags);
5977 read_unlock(&tasklist_lock);
5978 jiffies_to_timespec(time_slice, &t);
5979 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5983 read_unlock(&tasklist_lock);
5987 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5989 void sched_show_task(struct task_struct *p)
5991 unsigned long free = 0;
5994 state = p->state ? __ffs(p->state) + 1 : 0;
5995 printk(KERN_INFO "%-13.13s %c", p->comm,
5996 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5997 #if BITS_PER_LONG == 32
5998 if (state == TASK_RUNNING)
5999 printk(KERN_CONT " running ");
6001 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6003 if (state == TASK_RUNNING)
6004 printk(KERN_CONT " running task ");
6006 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6008 #ifdef CONFIG_DEBUG_STACK_USAGE
6010 unsigned long *n = end_of_stack(p);
6013 free = (unsigned long)n - (unsigned long)end_of_stack(p);
6016 printk(KERN_CONT "%5lu %5d %6d\n", free,
6017 task_pid_nr(p), task_pid_nr(p->real_parent));
6019 show_stack(p, NULL);
6022 void show_state_filter(unsigned long state_filter)
6024 struct task_struct *g, *p;
6026 #if BITS_PER_LONG == 32
6028 " task PC stack pid father\n");
6031 " task PC stack pid father\n");
6033 read_lock(&tasklist_lock);
6034 do_each_thread(g, p) {
6036 * reset the NMI-timeout, listing all files on a slow
6037 * console might take alot of time:
6039 touch_nmi_watchdog();
6040 if (!state_filter || (p->state & state_filter))
6042 } while_each_thread(g, p);
6044 touch_all_softlockup_watchdogs();
6046 #ifdef CONFIG_SCHED_DEBUG
6047 sysrq_sched_debug_show();
6049 read_unlock(&tasklist_lock);
6051 * Only show locks if all tasks are dumped:
6053 if (state_filter == -1)
6054 debug_show_all_locks();
6057 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6059 idle->sched_class = &idle_sched_class;
6063 * init_idle - set up an idle thread for a given CPU
6064 * @idle: task in question
6065 * @cpu: cpu the idle task belongs to
6067 * NOTE: this function does not set the idle thread's NEED_RESCHED
6068 * flag, to make booting more robust.
6070 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6072 struct rq *rq = cpu_rq(cpu);
6073 unsigned long flags;
6075 spin_lock_irqsave(&rq->lock, flags);
6078 idle->se.exec_start = sched_clock();
6080 idle->prio = idle->normal_prio = MAX_PRIO;
6081 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6082 __set_task_cpu(idle, cpu);
6084 rq->curr = rq->idle = idle;
6085 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6088 spin_unlock_irqrestore(&rq->lock, flags);
6090 /* Set the preempt count _outside_ the spinlocks! */
6091 #if defined(CONFIG_PREEMPT)
6092 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6094 task_thread_info(idle)->preempt_count = 0;
6097 * The idle tasks have their own, simple scheduling class:
6099 idle->sched_class = &idle_sched_class;
6100 ftrace_graph_init_task(idle);
6104 * In a system that switches off the HZ timer nohz_cpu_mask
6105 * indicates which cpus entered this state. This is used
6106 * in the rcu update to wait only for active cpus. For system
6107 * which do not switch off the HZ timer nohz_cpu_mask should
6108 * always be CPU_BITS_NONE.
6110 cpumask_var_t nohz_cpu_mask;
6113 * Increase the granularity value when there are more CPUs,
6114 * because with more CPUs the 'effective latency' as visible
6115 * to users decreases. But the relationship is not linear,
6116 * so pick a second-best guess by going with the log2 of the
6119 * This idea comes from the SD scheduler of Con Kolivas:
6121 static inline void sched_init_granularity(void)
6123 unsigned int factor = 1 + ilog2(num_online_cpus());
6124 const unsigned long limit = 200000000;
6126 sysctl_sched_min_granularity *= factor;
6127 if (sysctl_sched_min_granularity > limit)
6128 sysctl_sched_min_granularity = limit;
6130 sysctl_sched_latency *= factor;
6131 if (sysctl_sched_latency > limit)
6132 sysctl_sched_latency = limit;
6134 sysctl_sched_wakeup_granularity *= factor;
6136 sysctl_sched_shares_ratelimit *= factor;
6141 * This is how migration works:
6143 * 1) we queue a struct migration_req structure in the source CPU's
6144 * runqueue and wake up that CPU's migration thread.
6145 * 2) we down() the locked semaphore => thread blocks.
6146 * 3) migration thread wakes up (implicitly it forces the migrated
6147 * thread off the CPU)
6148 * 4) it gets the migration request and checks whether the migrated
6149 * task is still in the wrong runqueue.
6150 * 5) if it's in the wrong runqueue then the migration thread removes
6151 * it and puts it into the right queue.
6152 * 6) migration thread up()s the semaphore.
6153 * 7) we wake up and the migration is done.
6157 * Change a given task's CPU affinity. Migrate the thread to a
6158 * proper CPU and schedule it away if the CPU it's executing on
6159 * is removed from the allowed bitmask.
6161 * NOTE: the caller must have a valid reference to the task, the
6162 * task must not exit() & deallocate itself prematurely. The
6163 * call is not atomic; no spinlocks may be held.
6165 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6167 struct migration_req req;
6168 unsigned long flags;
6172 rq = task_rq_lock(p, &flags);
6173 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6178 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6179 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6184 if (p->sched_class->set_cpus_allowed)
6185 p->sched_class->set_cpus_allowed(p, new_mask);
6187 cpumask_copy(&p->cpus_allowed, new_mask);
6188 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6191 /* Can the task run on the task's current CPU? If so, we're done */
6192 if (cpumask_test_cpu(task_cpu(p), new_mask))
6195 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6196 /* Need help from migration thread: drop lock and wait. */
6197 task_rq_unlock(rq, &flags);
6198 wake_up_process(rq->migration_thread);
6199 wait_for_completion(&req.done);
6200 tlb_migrate_finish(p->mm);
6204 task_rq_unlock(rq, &flags);
6208 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6211 * Move (not current) task off this cpu, onto dest cpu. We're doing
6212 * this because either it can't run here any more (set_cpus_allowed()
6213 * away from this CPU, or CPU going down), or because we're
6214 * attempting to rebalance this task on exec (sched_exec).
6216 * So we race with normal scheduler movements, but that's OK, as long
6217 * as the task is no longer on this CPU.
6219 * Returns non-zero if task was successfully migrated.
6221 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6223 struct rq *rq_dest, *rq_src;
6226 if (unlikely(!cpu_active(dest_cpu)))
6229 rq_src = cpu_rq(src_cpu);
6230 rq_dest = cpu_rq(dest_cpu);
6232 double_rq_lock(rq_src, rq_dest);
6233 /* Already moved. */
6234 if (task_cpu(p) != src_cpu)
6236 /* Affinity changed (again). */
6237 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6240 on_rq = p->se.on_rq;
6242 deactivate_task(rq_src, p, 0);
6244 set_task_cpu(p, dest_cpu);
6246 activate_task(rq_dest, p, 0);
6247 check_preempt_curr(rq_dest, p, 0);
6252 double_rq_unlock(rq_src, rq_dest);
6257 * migration_thread - this is a highprio system thread that performs
6258 * thread migration by bumping thread off CPU then 'pushing' onto
6261 static int migration_thread(void *data)
6263 int cpu = (long)data;
6267 BUG_ON(rq->migration_thread != current);
6269 set_current_state(TASK_INTERRUPTIBLE);
6270 while (!kthread_should_stop()) {
6271 struct migration_req *req;
6272 struct list_head *head;
6274 spin_lock_irq(&rq->lock);
6276 if (cpu_is_offline(cpu)) {
6277 spin_unlock_irq(&rq->lock);
6281 if (rq->active_balance) {
6282 active_load_balance(rq, cpu);
6283 rq->active_balance = 0;
6286 head = &rq->migration_queue;
6288 if (list_empty(head)) {
6289 spin_unlock_irq(&rq->lock);
6291 set_current_state(TASK_INTERRUPTIBLE);
6294 req = list_entry(head->next, struct migration_req, list);
6295 list_del_init(head->next);
6297 spin_unlock(&rq->lock);
6298 __migrate_task(req->task, cpu, req->dest_cpu);
6301 complete(&req->done);
6303 __set_current_state(TASK_RUNNING);
6307 /* Wait for kthread_stop */
6308 set_current_state(TASK_INTERRUPTIBLE);
6309 while (!kthread_should_stop()) {
6311 set_current_state(TASK_INTERRUPTIBLE);
6313 __set_current_state(TASK_RUNNING);
6317 #ifdef CONFIG_HOTPLUG_CPU
6319 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6323 local_irq_disable();
6324 ret = __migrate_task(p, src_cpu, dest_cpu);
6330 * Figure out where task on dead CPU should go, use force if necessary.
6332 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6335 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
6338 /* Look for allowed, online CPU in same node. */
6339 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
6340 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6343 /* Any allowed, online CPU? */
6344 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
6345 if (dest_cpu < nr_cpu_ids)
6348 /* No more Mr. Nice Guy. */
6349 if (dest_cpu >= nr_cpu_ids) {
6350 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
6351 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
6354 * Don't tell them about moving exiting tasks or
6355 * kernel threads (both mm NULL), since they never
6358 if (p->mm && printk_ratelimit()) {
6359 printk(KERN_INFO "process %d (%s) no "
6360 "longer affine to cpu%d\n",
6361 task_pid_nr(p), p->comm, dead_cpu);
6366 /* It can have affinity changed while we were choosing. */
6367 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
6372 * While a dead CPU has no uninterruptible tasks queued at this point,
6373 * it might still have a nonzero ->nr_uninterruptible counter, because
6374 * for performance reasons the counter is not stricly tracking tasks to
6375 * their home CPUs. So we just add the counter to another CPU's counter,
6376 * to keep the global sum constant after CPU-down:
6378 static void migrate_nr_uninterruptible(struct rq *rq_src)
6380 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
6381 unsigned long flags;
6383 local_irq_save(flags);
6384 double_rq_lock(rq_src, rq_dest);
6385 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6386 rq_src->nr_uninterruptible = 0;
6387 double_rq_unlock(rq_src, rq_dest);
6388 local_irq_restore(flags);
6391 /* Run through task list and migrate tasks from the dead cpu. */
6392 static void migrate_live_tasks(int src_cpu)
6394 struct task_struct *p, *t;
6396 read_lock(&tasklist_lock);
6398 do_each_thread(t, p) {
6402 if (task_cpu(p) == src_cpu)
6403 move_task_off_dead_cpu(src_cpu, p);
6404 } while_each_thread(t, p);
6406 read_unlock(&tasklist_lock);
6410 * Schedules idle task to be the next runnable task on current CPU.
6411 * It does so by boosting its priority to highest possible.
6412 * Used by CPU offline code.
6414 void sched_idle_next(void)
6416 int this_cpu = smp_processor_id();
6417 struct rq *rq = cpu_rq(this_cpu);
6418 struct task_struct *p = rq->idle;
6419 unsigned long flags;
6421 /* cpu has to be offline */
6422 BUG_ON(cpu_online(this_cpu));
6425 * Strictly not necessary since rest of the CPUs are stopped by now
6426 * and interrupts disabled on the current cpu.
6428 spin_lock_irqsave(&rq->lock, flags);
6430 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6432 update_rq_clock(rq);
6433 activate_task(rq, p, 0);
6435 spin_unlock_irqrestore(&rq->lock, flags);
6439 * Ensures that the idle task is using init_mm right before its cpu goes
6442 void idle_task_exit(void)
6444 struct mm_struct *mm = current->active_mm;
6446 BUG_ON(cpu_online(smp_processor_id()));
6449 switch_mm(mm, &init_mm, current);
6453 /* called under rq->lock with disabled interrupts */
6454 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6456 struct rq *rq = cpu_rq(dead_cpu);
6458 /* Must be exiting, otherwise would be on tasklist. */
6459 BUG_ON(!p->exit_state);
6461 /* Cannot have done final schedule yet: would have vanished. */
6462 BUG_ON(p->state == TASK_DEAD);
6467 * Drop lock around migration; if someone else moves it,
6468 * that's OK. No task can be added to this CPU, so iteration is
6471 spin_unlock_irq(&rq->lock);
6472 move_task_off_dead_cpu(dead_cpu, p);
6473 spin_lock_irq(&rq->lock);
6478 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6479 static void migrate_dead_tasks(unsigned int dead_cpu)
6481 struct rq *rq = cpu_rq(dead_cpu);
6482 struct task_struct *next;
6485 if (!rq->nr_running)
6487 update_rq_clock(rq);
6488 next = pick_next_task(rq, rq->curr);
6491 next->sched_class->put_prev_task(rq, next);
6492 migrate_dead(dead_cpu, next);
6496 #endif /* CONFIG_HOTPLUG_CPU */
6498 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6500 static struct ctl_table sd_ctl_dir[] = {
6502 .procname = "sched_domain",
6508 static struct ctl_table sd_ctl_root[] = {
6510 .ctl_name = CTL_KERN,
6511 .procname = "kernel",
6513 .child = sd_ctl_dir,
6518 static struct ctl_table *sd_alloc_ctl_entry(int n)
6520 struct ctl_table *entry =
6521 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6526 static void sd_free_ctl_entry(struct ctl_table **tablep)
6528 struct ctl_table *entry;
6531 * In the intermediate directories, both the child directory and
6532 * procname are dynamically allocated and could fail but the mode
6533 * will always be set. In the lowest directory the names are
6534 * static strings and all have proc handlers.
6536 for (entry = *tablep; entry->mode; entry++) {
6538 sd_free_ctl_entry(&entry->child);
6539 if (entry->proc_handler == NULL)
6540 kfree(entry->procname);
6548 set_table_entry(struct ctl_table *entry,
6549 const char *procname, void *data, int maxlen,
6550 mode_t mode, proc_handler *proc_handler)
6552 entry->procname = procname;
6554 entry->maxlen = maxlen;
6556 entry->proc_handler = proc_handler;
6559 static struct ctl_table *
6560 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6562 struct ctl_table *table = sd_alloc_ctl_entry(13);
6567 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6568 sizeof(long), 0644, proc_doulongvec_minmax);
6569 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6570 sizeof(long), 0644, proc_doulongvec_minmax);
6571 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6572 sizeof(int), 0644, proc_dointvec_minmax);
6573 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6574 sizeof(int), 0644, proc_dointvec_minmax);
6575 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6576 sizeof(int), 0644, proc_dointvec_minmax);
6577 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6578 sizeof(int), 0644, proc_dointvec_minmax);
6579 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6580 sizeof(int), 0644, proc_dointvec_minmax);
6581 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6582 sizeof(int), 0644, proc_dointvec_minmax);
6583 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6584 sizeof(int), 0644, proc_dointvec_minmax);
6585 set_table_entry(&table[9], "cache_nice_tries",
6586 &sd->cache_nice_tries,
6587 sizeof(int), 0644, proc_dointvec_minmax);
6588 set_table_entry(&table[10], "flags", &sd->flags,
6589 sizeof(int), 0644, proc_dointvec_minmax);
6590 set_table_entry(&table[11], "name", sd->name,
6591 CORENAME_MAX_SIZE, 0444, proc_dostring);
6592 /* &table[12] is terminator */
6597 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6599 struct ctl_table *entry, *table;
6600 struct sched_domain *sd;
6601 int domain_num = 0, i;
6604 for_each_domain(cpu, sd)
6606 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6611 for_each_domain(cpu, sd) {
6612 snprintf(buf, 32, "domain%d", i);
6613 entry->procname = kstrdup(buf, GFP_KERNEL);
6615 entry->child = sd_alloc_ctl_domain_table(sd);
6622 static struct ctl_table_header *sd_sysctl_header;
6623 static void register_sched_domain_sysctl(void)
6625 int i, cpu_num = num_online_cpus();
6626 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6629 WARN_ON(sd_ctl_dir[0].child);
6630 sd_ctl_dir[0].child = entry;
6635 for_each_online_cpu(i) {
6636 snprintf(buf, 32, "cpu%d", i);
6637 entry->procname = kstrdup(buf, GFP_KERNEL);
6639 entry->child = sd_alloc_ctl_cpu_table(i);
6643 WARN_ON(sd_sysctl_header);
6644 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6647 /* may be called multiple times per register */
6648 static void unregister_sched_domain_sysctl(void)
6650 if (sd_sysctl_header)
6651 unregister_sysctl_table(sd_sysctl_header);
6652 sd_sysctl_header = NULL;
6653 if (sd_ctl_dir[0].child)
6654 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6657 static void register_sched_domain_sysctl(void)
6660 static void unregister_sched_domain_sysctl(void)
6665 static void set_rq_online(struct rq *rq)
6668 const struct sched_class *class;
6670 cpumask_set_cpu(rq->cpu, rq->rd->online);
6673 for_each_class(class) {
6674 if (class->rq_online)
6675 class->rq_online(rq);
6680 static void set_rq_offline(struct rq *rq)
6683 const struct sched_class *class;
6685 for_each_class(class) {
6686 if (class->rq_offline)
6687 class->rq_offline(rq);
6690 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6696 * migration_call - callback that gets triggered when a CPU is added.
6697 * Here we can start up the necessary migration thread for the new CPU.
6699 static int __cpuinit
6700 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6702 struct task_struct *p;
6703 int cpu = (long)hcpu;
6704 unsigned long flags;
6709 case CPU_UP_PREPARE:
6710 case CPU_UP_PREPARE_FROZEN:
6711 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6714 kthread_bind(p, cpu);
6715 /* Must be high prio: stop_machine expects to yield to it. */
6716 rq = task_rq_lock(p, &flags);
6717 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6718 task_rq_unlock(rq, &flags);
6719 cpu_rq(cpu)->migration_thread = p;
6723 case CPU_ONLINE_FROZEN:
6724 /* Strictly unnecessary, as first user will wake it. */
6725 wake_up_process(cpu_rq(cpu)->migration_thread);
6727 /* Update our root-domain */
6729 spin_lock_irqsave(&rq->lock, flags);
6731 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6735 spin_unlock_irqrestore(&rq->lock, flags);
6738 #ifdef CONFIG_HOTPLUG_CPU
6739 case CPU_UP_CANCELED:
6740 case CPU_UP_CANCELED_FROZEN:
6741 if (!cpu_rq(cpu)->migration_thread)
6743 /* Unbind it from offline cpu so it can run. Fall thru. */
6744 kthread_bind(cpu_rq(cpu)->migration_thread,
6745 cpumask_any(cpu_online_mask));
6746 kthread_stop(cpu_rq(cpu)->migration_thread);
6747 cpu_rq(cpu)->migration_thread = NULL;
6751 case CPU_DEAD_FROZEN:
6752 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6753 migrate_live_tasks(cpu);
6755 kthread_stop(rq->migration_thread);
6756 rq->migration_thread = NULL;
6757 /* Idle task back to normal (off runqueue, low prio) */
6758 spin_lock_irq(&rq->lock);
6759 update_rq_clock(rq);
6760 deactivate_task(rq, rq->idle, 0);
6761 rq->idle->static_prio = MAX_PRIO;
6762 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6763 rq->idle->sched_class = &idle_sched_class;
6764 migrate_dead_tasks(cpu);
6765 spin_unlock_irq(&rq->lock);
6767 migrate_nr_uninterruptible(rq);
6768 BUG_ON(rq->nr_running != 0);
6771 * No need to migrate the tasks: it was best-effort if
6772 * they didn't take sched_hotcpu_mutex. Just wake up
6775 spin_lock_irq(&rq->lock);
6776 while (!list_empty(&rq->migration_queue)) {
6777 struct migration_req *req;
6779 req = list_entry(rq->migration_queue.next,
6780 struct migration_req, list);
6781 list_del_init(&req->list);
6782 spin_unlock_irq(&rq->lock);
6783 complete(&req->done);
6784 spin_lock_irq(&rq->lock);
6786 spin_unlock_irq(&rq->lock);
6790 case CPU_DYING_FROZEN:
6791 /* Update our root-domain */
6793 spin_lock_irqsave(&rq->lock, flags);
6795 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6798 spin_unlock_irqrestore(&rq->lock, flags);
6805 /* Register at highest priority so that task migration (migrate_all_tasks)
6806 * happens before everything else.
6808 static struct notifier_block __cpuinitdata migration_notifier = {
6809 .notifier_call = migration_call,
6813 static int __init migration_init(void)
6815 void *cpu = (void *)(long)smp_processor_id();
6818 /* Start one for the boot CPU: */
6819 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6820 BUG_ON(err == NOTIFY_BAD);
6821 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6822 register_cpu_notifier(&migration_notifier);
6826 early_initcall(migration_init);
6831 #ifdef CONFIG_SCHED_DEBUG
6833 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6834 struct cpumask *groupmask)
6836 struct sched_group *group = sd->groups;
6839 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6840 cpumask_clear(groupmask);
6842 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6844 if (!(sd->flags & SD_LOAD_BALANCE)) {
6845 printk("does not load-balance\n");
6847 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6852 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6854 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6855 printk(KERN_ERR "ERROR: domain->span does not contain "
6858 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6859 printk(KERN_ERR "ERROR: domain->groups does not contain"
6863 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6867 printk(KERN_ERR "ERROR: group is NULL\n");
6871 if (!group->__cpu_power) {
6872 printk(KERN_CONT "\n");
6873 printk(KERN_ERR "ERROR: domain->cpu_power not "
6878 if (!cpumask_weight(sched_group_cpus(group))) {
6879 printk(KERN_CONT "\n");
6880 printk(KERN_ERR "ERROR: empty group\n");
6884 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6885 printk(KERN_CONT "\n");
6886 printk(KERN_ERR "ERROR: repeated CPUs\n");
6890 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6892 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6893 printk(KERN_CONT " %s", str);
6895 group = group->next;
6896 } while (group != sd->groups);
6897 printk(KERN_CONT "\n");
6899 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6900 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6903 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6904 printk(KERN_ERR "ERROR: parent span is not a superset "
6905 "of domain->span\n");
6909 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6911 cpumask_var_t groupmask;
6915 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6919 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6921 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6922 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6927 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6934 free_cpumask_var(groupmask);
6936 #else /* !CONFIG_SCHED_DEBUG */
6937 # define sched_domain_debug(sd, cpu) do { } while (0)
6938 #endif /* CONFIG_SCHED_DEBUG */
6940 static int sd_degenerate(struct sched_domain *sd)
6942 if (cpumask_weight(sched_domain_span(sd)) == 1)
6945 /* Following flags need at least 2 groups */
6946 if (sd->flags & (SD_LOAD_BALANCE |
6947 SD_BALANCE_NEWIDLE |
6951 SD_SHARE_PKG_RESOURCES)) {
6952 if (sd->groups != sd->groups->next)
6956 /* Following flags don't use groups */
6957 if (sd->flags & (SD_WAKE_IDLE |
6966 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6968 unsigned long cflags = sd->flags, pflags = parent->flags;
6970 if (sd_degenerate(parent))
6973 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6976 /* Does parent contain flags not in child? */
6977 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6978 if (cflags & SD_WAKE_AFFINE)
6979 pflags &= ~SD_WAKE_BALANCE;
6980 /* Flags needing groups don't count if only 1 group in parent */
6981 if (parent->groups == parent->groups->next) {
6982 pflags &= ~(SD_LOAD_BALANCE |
6983 SD_BALANCE_NEWIDLE |
6987 SD_SHARE_PKG_RESOURCES);
6988 if (nr_node_ids == 1)
6989 pflags &= ~SD_SERIALIZE;
6991 if (~cflags & pflags)
6997 static void free_rootdomain(struct root_domain *rd)
6999 cpupri_cleanup(&rd->cpupri);
7001 free_cpumask_var(rd->rto_mask);
7002 free_cpumask_var(rd->online);
7003 free_cpumask_var(rd->span);
7007 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7009 unsigned long flags;
7011 spin_lock_irqsave(&rq->lock, flags);
7014 struct root_domain *old_rd = rq->rd;
7016 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7019 cpumask_clear_cpu(rq->cpu, old_rd->span);
7021 if (atomic_dec_and_test(&old_rd->refcount))
7022 free_rootdomain(old_rd);
7025 atomic_inc(&rd->refcount);
7028 cpumask_set_cpu(rq->cpu, rd->span);
7029 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7032 spin_unlock_irqrestore(&rq->lock, flags);
7035 static int __init_refok init_rootdomain(struct root_domain *rd, bool bootmem)
7037 memset(rd, 0, sizeof(*rd));
7040 alloc_bootmem_cpumask_var(&def_root_domain.span);
7041 alloc_bootmem_cpumask_var(&def_root_domain.online);
7042 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
7043 cpupri_init(&rd->cpupri, true);
7047 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7049 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7051 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7054 if (cpupri_init(&rd->cpupri, false) != 0)
7059 free_cpumask_var(rd->rto_mask);
7061 free_cpumask_var(rd->online);
7063 free_cpumask_var(rd->span);
7068 static void init_defrootdomain(void)
7070 init_rootdomain(&def_root_domain, true);
7072 atomic_set(&def_root_domain.refcount, 1);
7075 static struct root_domain *alloc_rootdomain(void)
7077 struct root_domain *rd;
7079 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7083 if (init_rootdomain(rd, false) != 0) {
7092 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7093 * hold the hotplug lock.
7096 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7098 struct rq *rq = cpu_rq(cpu);
7099 struct sched_domain *tmp;
7101 /* Remove the sched domains which do not contribute to scheduling. */
7102 for (tmp = sd; tmp; ) {
7103 struct sched_domain *parent = tmp->parent;
7107 if (sd_parent_degenerate(tmp, parent)) {
7108 tmp->parent = parent->parent;
7110 parent->parent->child = tmp;
7115 if (sd && sd_degenerate(sd)) {
7121 sched_domain_debug(sd, cpu);
7123 rq_attach_root(rq, rd);
7124 rcu_assign_pointer(rq->sd, sd);
7127 /* cpus with isolated domains */
7128 static cpumask_var_t cpu_isolated_map;
7130 /* Setup the mask of cpus configured for isolated domains */
7131 static int __init isolated_cpu_setup(char *str)
7133 cpulist_parse(str, cpu_isolated_map);
7137 __setup("isolcpus=", isolated_cpu_setup);
7140 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7141 * to a function which identifies what group(along with sched group) a CPU
7142 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7143 * (due to the fact that we keep track of groups covered with a struct cpumask).
7145 * init_sched_build_groups will build a circular linked list of the groups
7146 * covered by the given span, and will set each group's ->cpumask correctly,
7147 * and ->cpu_power to 0.
7150 init_sched_build_groups(const struct cpumask *span,
7151 const struct cpumask *cpu_map,
7152 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7153 struct sched_group **sg,
7154 struct cpumask *tmpmask),
7155 struct cpumask *covered, struct cpumask *tmpmask)
7157 struct sched_group *first = NULL, *last = NULL;
7160 cpumask_clear(covered);
7162 for_each_cpu(i, span) {
7163 struct sched_group *sg;
7164 int group = group_fn(i, cpu_map, &sg, tmpmask);
7167 if (cpumask_test_cpu(i, covered))
7170 cpumask_clear(sched_group_cpus(sg));
7171 sg->__cpu_power = 0;
7173 for_each_cpu(j, span) {
7174 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7177 cpumask_set_cpu(j, covered);
7178 cpumask_set_cpu(j, sched_group_cpus(sg));
7189 #define SD_NODES_PER_DOMAIN 16
7194 * find_next_best_node - find the next node to include in a sched_domain
7195 * @node: node whose sched_domain we're building
7196 * @used_nodes: nodes already in the sched_domain
7198 * Find the next node to include in a given scheduling domain. Simply
7199 * finds the closest node not already in the @used_nodes map.
7201 * Should use nodemask_t.
7203 static int find_next_best_node(int node, nodemask_t *used_nodes)
7205 int i, n, val, min_val, best_node = 0;
7209 for (i = 0; i < nr_node_ids; i++) {
7210 /* Start at @node */
7211 n = (node + i) % nr_node_ids;
7213 if (!nr_cpus_node(n))
7216 /* Skip already used nodes */
7217 if (node_isset(n, *used_nodes))
7220 /* Simple min distance search */
7221 val = node_distance(node, n);
7223 if (val < min_val) {
7229 node_set(best_node, *used_nodes);
7234 * sched_domain_node_span - get a cpumask for a node's sched_domain
7235 * @node: node whose cpumask we're constructing
7236 * @span: resulting cpumask
7238 * Given a node, construct a good cpumask for its sched_domain to span. It
7239 * should be one that prevents unnecessary balancing, but also spreads tasks
7242 static void sched_domain_node_span(int node, struct cpumask *span)
7244 nodemask_t used_nodes;
7247 cpumask_clear(span);
7248 nodes_clear(used_nodes);
7250 cpumask_or(span, span, cpumask_of_node(node));
7251 node_set(node, used_nodes);
7253 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7254 int next_node = find_next_best_node(node, &used_nodes);
7256 cpumask_or(span, span, cpumask_of_node(next_node));
7259 #endif /* CONFIG_NUMA */
7261 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7264 * The cpus mask in sched_group and sched_domain hangs off the end.
7265 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7266 * for nr_cpu_ids < CONFIG_NR_CPUS.
7268 struct static_sched_group {
7269 struct sched_group sg;
7270 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7273 struct static_sched_domain {
7274 struct sched_domain sd;
7275 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7279 * SMT sched-domains:
7281 #ifdef CONFIG_SCHED_SMT
7282 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7283 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7286 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
7287 struct sched_group **sg, struct cpumask *unused)
7290 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7293 #endif /* CONFIG_SCHED_SMT */
7296 * multi-core sched-domains:
7298 #ifdef CONFIG_SCHED_MC
7299 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7300 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7301 #endif /* CONFIG_SCHED_MC */
7303 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7305 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7306 struct sched_group **sg, struct cpumask *mask)
7310 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7311 group = cpumask_first(mask);
7313 *sg = &per_cpu(sched_group_core, group).sg;
7316 #elif defined(CONFIG_SCHED_MC)
7318 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7319 struct sched_group **sg, struct cpumask *unused)
7322 *sg = &per_cpu(sched_group_core, cpu).sg;
7327 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
7328 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
7331 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
7332 struct sched_group **sg, struct cpumask *mask)
7335 #ifdef CONFIG_SCHED_MC
7336 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
7337 group = cpumask_first(mask);
7338 #elif defined(CONFIG_SCHED_SMT)
7339 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7340 group = cpumask_first(mask);
7345 *sg = &per_cpu(sched_group_phys, group).sg;
7351 * The init_sched_build_groups can't handle what we want to do with node
7352 * groups, so roll our own. Now each node has its own list of groups which
7353 * gets dynamically allocated.
7355 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7356 static struct sched_group ***sched_group_nodes_bycpu;
7358 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7359 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7361 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
7362 struct sched_group **sg,
7363 struct cpumask *nodemask)
7367 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7368 group = cpumask_first(nodemask);
7371 *sg = &per_cpu(sched_group_allnodes, group).sg;
7375 static void init_numa_sched_groups_power(struct sched_group *group_head)
7377 struct sched_group *sg = group_head;
7383 for_each_cpu(j, sched_group_cpus(sg)) {
7384 struct sched_domain *sd;
7386 sd = &per_cpu(phys_domains, j).sd;
7387 if (j != cpumask_first(sched_group_cpus(sd->groups))) {
7389 * Only add "power" once for each
7395 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7398 } while (sg != group_head);
7400 #endif /* CONFIG_NUMA */
7403 /* Free memory allocated for various sched_group structures */
7404 static void free_sched_groups(const struct cpumask *cpu_map,
7405 struct cpumask *nodemask)
7409 for_each_cpu(cpu, cpu_map) {
7410 struct sched_group **sched_group_nodes
7411 = sched_group_nodes_bycpu[cpu];
7413 if (!sched_group_nodes)
7416 for (i = 0; i < nr_node_ids; i++) {
7417 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7419 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7420 if (cpumask_empty(nodemask))
7430 if (oldsg != sched_group_nodes[i])
7433 kfree(sched_group_nodes);
7434 sched_group_nodes_bycpu[cpu] = NULL;
7437 #else /* !CONFIG_NUMA */
7438 static void free_sched_groups(const struct cpumask *cpu_map,
7439 struct cpumask *nodemask)
7442 #endif /* CONFIG_NUMA */
7445 * Initialize sched groups cpu_power.
7447 * cpu_power indicates the capacity of sched group, which is used while
7448 * distributing the load between different sched groups in a sched domain.
7449 * Typically cpu_power for all the groups in a sched domain will be same unless
7450 * there are asymmetries in the topology. If there are asymmetries, group
7451 * having more cpu_power will pickup more load compared to the group having
7454 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7455 * the maximum number of tasks a group can handle in the presence of other idle
7456 * or lightly loaded groups in the same sched domain.
7458 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7460 struct sched_domain *child;
7461 struct sched_group *group;
7463 WARN_ON(!sd || !sd->groups);
7465 if (cpu != cpumask_first(sched_group_cpus(sd->groups)))
7470 sd->groups->__cpu_power = 0;
7473 * For perf policy, if the groups in child domain share resources
7474 * (for example cores sharing some portions of the cache hierarchy
7475 * or SMT), then set this domain groups cpu_power such that each group
7476 * can handle only one task, when there are other idle groups in the
7477 * same sched domain.
7479 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7481 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7482 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7487 * add cpu_power of each child group to this groups cpu_power
7489 group = child->groups;
7491 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7492 group = group->next;
7493 } while (group != child->groups);
7497 * Initializers for schedule domains
7498 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7501 #ifdef CONFIG_SCHED_DEBUG
7502 # define SD_INIT_NAME(sd, type) sd->name = #type
7504 # define SD_INIT_NAME(sd, type) do { } while (0)
7507 #define SD_INIT(sd, type) sd_init_##type(sd)
7509 #define SD_INIT_FUNC(type) \
7510 static noinline void sd_init_##type(struct sched_domain *sd) \
7512 memset(sd, 0, sizeof(*sd)); \
7513 *sd = SD_##type##_INIT; \
7514 sd->level = SD_LV_##type; \
7515 SD_INIT_NAME(sd, type); \
7520 SD_INIT_FUNC(ALLNODES)
7523 #ifdef CONFIG_SCHED_SMT
7524 SD_INIT_FUNC(SIBLING)
7526 #ifdef CONFIG_SCHED_MC
7530 static int default_relax_domain_level = -1;
7532 static int __init setup_relax_domain_level(char *str)
7536 val = simple_strtoul(str, NULL, 0);
7537 if (val < SD_LV_MAX)
7538 default_relax_domain_level = val;
7542 __setup("relax_domain_level=", setup_relax_domain_level);
7544 static void set_domain_attribute(struct sched_domain *sd,
7545 struct sched_domain_attr *attr)
7549 if (!attr || attr->relax_domain_level < 0) {
7550 if (default_relax_domain_level < 0)
7553 request = default_relax_domain_level;
7555 request = attr->relax_domain_level;
7556 if (request < sd->level) {
7557 /* turn off idle balance on this domain */
7558 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7560 /* turn on idle balance on this domain */
7561 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7566 * Build sched domains for a given set of cpus and attach the sched domains
7567 * to the individual cpus
7569 static int __build_sched_domains(const struct cpumask *cpu_map,
7570 struct sched_domain_attr *attr)
7572 int i, err = -ENOMEM;
7573 struct root_domain *rd;
7574 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
7577 cpumask_var_t domainspan, covered, notcovered;
7578 struct sched_group **sched_group_nodes = NULL;
7579 int sd_allnodes = 0;
7581 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
7583 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
7584 goto free_domainspan;
7585 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
7589 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
7590 goto free_notcovered;
7591 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
7593 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
7594 goto free_this_sibling_map;
7595 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
7596 goto free_this_core_map;
7597 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
7598 goto free_send_covered;
7602 * Allocate the per-node list of sched groups
7604 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7606 if (!sched_group_nodes) {
7607 printk(KERN_WARNING "Can not alloc sched group node list\n");
7612 rd = alloc_rootdomain();
7614 printk(KERN_WARNING "Cannot alloc root domain\n");
7615 goto free_sched_groups;
7619 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
7623 * Set up domains for cpus specified by the cpu_map.
7625 for_each_cpu(i, cpu_map) {
7626 struct sched_domain *sd = NULL, *p;
7628 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
7631 if (cpumask_weight(cpu_map) >
7632 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
7633 sd = &per_cpu(allnodes_domains, i);
7634 SD_INIT(sd, ALLNODES);
7635 set_domain_attribute(sd, attr);
7636 cpumask_copy(sched_domain_span(sd), cpu_map);
7637 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7643 sd = &per_cpu(node_domains, i);
7645 set_domain_attribute(sd, attr);
7646 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7650 cpumask_and(sched_domain_span(sd),
7651 sched_domain_span(sd), cpu_map);
7655 sd = &per_cpu(phys_domains, i).sd;
7657 set_domain_attribute(sd, attr);
7658 cpumask_copy(sched_domain_span(sd), nodemask);
7662 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7664 #ifdef CONFIG_SCHED_MC
7666 sd = &per_cpu(core_domains, i).sd;
7668 set_domain_attribute(sd, attr);
7669 cpumask_and(sched_domain_span(sd), cpu_map,
7670 cpu_coregroup_mask(i));
7673 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7676 #ifdef CONFIG_SCHED_SMT
7678 sd = &per_cpu(cpu_domains, i).sd;
7679 SD_INIT(sd, SIBLING);
7680 set_domain_attribute(sd, attr);
7681 cpumask_and(sched_domain_span(sd),
7682 &per_cpu(cpu_sibling_map, i), cpu_map);
7685 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7689 #ifdef CONFIG_SCHED_SMT
7690 /* Set up CPU (sibling) groups */
7691 for_each_cpu(i, cpu_map) {
7692 cpumask_and(this_sibling_map,
7693 &per_cpu(cpu_sibling_map, i), cpu_map);
7694 if (i != cpumask_first(this_sibling_map))
7697 init_sched_build_groups(this_sibling_map, cpu_map,
7699 send_covered, tmpmask);
7703 #ifdef CONFIG_SCHED_MC
7704 /* Set up multi-core groups */
7705 for_each_cpu(i, cpu_map) {
7706 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
7707 if (i != cpumask_first(this_core_map))
7710 init_sched_build_groups(this_core_map, cpu_map,
7712 send_covered, tmpmask);
7716 /* Set up physical groups */
7717 for (i = 0; i < nr_node_ids; i++) {
7718 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7719 if (cpumask_empty(nodemask))
7722 init_sched_build_groups(nodemask, cpu_map,
7724 send_covered, tmpmask);
7728 /* Set up node groups */
7730 init_sched_build_groups(cpu_map, cpu_map,
7731 &cpu_to_allnodes_group,
7732 send_covered, tmpmask);
7735 for (i = 0; i < nr_node_ids; i++) {
7736 /* Set up node groups */
7737 struct sched_group *sg, *prev;
7740 cpumask_clear(covered);
7741 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7742 if (cpumask_empty(nodemask)) {
7743 sched_group_nodes[i] = NULL;
7747 sched_domain_node_span(i, domainspan);
7748 cpumask_and(domainspan, domainspan, cpu_map);
7750 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7753 printk(KERN_WARNING "Can not alloc domain group for "
7757 sched_group_nodes[i] = sg;
7758 for_each_cpu(j, nodemask) {
7759 struct sched_domain *sd;
7761 sd = &per_cpu(node_domains, j);
7764 sg->__cpu_power = 0;
7765 cpumask_copy(sched_group_cpus(sg), nodemask);
7767 cpumask_or(covered, covered, nodemask);
7770 for (j = 0; j < nr_node_ids; j++) {
7771 int n = (i + j) % nr_node_ids;
7773 cpumask_complement(notcovered, covered);
7774 cpumask_and(tmpmask, notcovered, cpu_map);
7775 cpumask_and(tmpmask, tmpmask, domainspan);
7776 if (cpumask_empty(tmpmask))
7779 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
7780 if (cpumask_empty(tmpmask))
7783 sg = kmalloc_node(sizeof(struct sched_group) +
7788 "Can not alloc domain group for node %d\n", j);
7791 sg->__cpu_power = 0;
7792 cpumask_copy(sched_group_cpus(sg), tmpmask);
7793 sg->next = prev->next;
7794 cpumask_or(covered, covered, tmpmask);
7801 /* Calculate CPU power for physical packages and nodes */
7802 #ifdef CONFIG_SCHED_SMT
7803 for_each_cpu(i, cpu_map) {
7804 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
7806 init_sched_groups_power(i, sd);
7809 #ifdef CONFIG_SCHED_MC
7810 for_each_cpu(i, cpu_map) {
7811 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
7813 init_sched_groups_power(i, sd);
7817 for_each_cpu(i, cpu_map) {
7818 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
7820 init_sched_groups_power(i, sd);
7824 for (i = 0; i < nr_node_ids; i++)
7825 init_numa_sched_groups_power(sched_group_nodes[i]);
7828 struct sched_group *sg;
7830 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7832 init_numa_sched_groups_power(sg);
7836 /* Attach the domains */
7837 for_each_cpu(i, cpu_map) {
7838 struct sched_domain *sd;
7839 #ifdef CONFIG_SCHED_SMT
7840 sd = &per_cpu(cpu_domains, i).sd;
7841 #elif defined(CONFIG_SCHED_MC)
7842 sd = &per_cpu(core_domains, i).sd;
7844 sd = &per_cpu(phys_domains, i).sd;
7846 cpu_attach_domain(sd, rd, i);
7852 free_cpumask_var(tmpmask);
7854 free_cpumask_var(send_covered);
7856 free_cpumask_var(this_core_map);
7857 free_this_sibling_map:
7858 free_cpumask_var(this_sibling_map);
7860 free_cpumask_var(nodemask);
7863 free_cpumask_var(notcovered);
7865 free_cpumask_var(covered);
7867 free_cpumask_var(domainspan);
7874 kfree(sched_group_nodes);
7880 free_sched_groups(cpu_map, tmpmask);
7881 free_rootdomain(rd);
7886 static int build_sched_domains(const struct cpumask *cpu_map)
7888 return __build_sched_domains(cpu_map, NULL);
7891 static struct cpumask *doms_cur; /* current sched domains */
7892 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7893 static struct sched_domain_attr *dattr_cur;
7894 /* attribues of custom domains in 'doms_cur' */
7897 * Special case: If a kmalloc of a doms_cur partition (array of
7898 * cpumask) fails, then fallback to a single sched domain,
7899 * as determined by the single cpumask fallback_doms.
7901 static cpumask_var_t fallback_doms;
7904 * arch_update_cpu_topology lets virtualized architectures update the
7905 * cpu core maps. It is supposed to return 1 if the topology changed
7906 * or 0 if it stayed the same.
7908 int __attribute__((weak)) arch_update_cpu_topology(void)
7914 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7915 * For now this just excludes isolated cpus, but could be used to
7916 * exclude other special cases in the future.
7918 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7922 arch_update_cpu_topology();
7924 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
7926 doms_cur = fallback_doms;
7927 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
7929 err = build_sched_domains(doms_cur);
7930 register_sched_domain_sysctl();
7935 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7936 struct cpumask *tmpmask)
7938 free_sched_groups(cpu_map, tmpmask);
7942 * Detach sched domains from a group of cpus specified in cpu_map
7943 * These cpus will now be attached to the NULL domain
7945 static void detach_destroy_domains(const struct cpumask *cpu_map)
7947 /* Save because hotplug lock held. */
7948 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7951 for_each_cpu(i, cpu_map)
7952 cpu_attach_domain(NULL, &def_root_domain, i);
7953 synchronize_sched();
7954 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7957 /* handle null as "default" */
7958 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7959 struct sched_domain_attr *new, int idx_new)
7961 struct sched_domain_attr tmp;
7968 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7969 new ? (new + idx_new) : &tmp,
7970 sizeof(struct sched_domain_attr));
7974 * Partition sched domains as specified by the 'ndoms_new'
7975 * cpumasks in the array doms_new[] of cpumasks. This compares
7976 * doms_new[] to the current sched domain partitioning, doms_cur[].
7977 * It destroys each deleted domain and builds each new domain.
7979 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
7980 * The masks don't intersect (don't overlap.) We should setup one
7981 * sched domain for each mask. CPUs not in any of the cpumasks will
7982 * not be load balanced. If the same cpumask appears both in the
7983 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7986 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7987 * ownership of it and will kfree it when done with it. If the caller
7988 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7989 * ndoms_new == 1, and partition_sched_domains() will fallback to
7990 * the single partition 'fallback_doms', it also forces the domains
7993 * If doms_new == NULL it will be replaced with cpu_online_mask.
7994 * ndoms_new == 0 is a special case for destroying existing domains,
7995 * and it will not create the default domain.
7997 * Call with hotplug lock held
7999 /* FIXME: Change to struct cpumask *doms_new[] */
8000 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8001 struct sched_domain_attr *dattr_new)
8006 mutex_lock(&sched_domains_mutex);
8008 /* always unregister in case we don't destroy any domains */
8009 unregister_sched_domain_sysctl();
8011 /* Let architecture update cpu core mappings. */
8012 new_topology = arch_update_cpu_topology();
8014 n = doms_new ? ndoms_new : 0;
8016 /* Destroy deleted domains */
8017 for (i = 0; i < ndoms_cur; i++) {
8018 for (j = 0; j < n && !new_topology; j++) {
8019 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8020 && dattrs_equal(dattr_cur, i, dattr_new, j))
8023 /* no match - a current sched domain not in new doms_new[] */
8024 detach_destroy_domains(doms_cur + i);
8029 if (doms_new == NULL) {
8031 doms_new = fallback_doms;
8032 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8033 WARN_ON_ONCE(dattr_new);
8036 /* Build new domains */
8037 for (i = 0; i < ndoms_new; i++) {
8038 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8039 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8040 && dattrs_equal(dattr_new, i, dattr_cur, j))
8043 /* no match - add a new doms_new */
8044 __build_sched_domains(doms_new + i,
8045 dattr_new ? dattr_new + i : NULL);
8050 /* Remember the new sched domains */
8051 if (doms_cur != fallback_doms)
8053 kfree(dattr_cur); /* kfree(NULL) is safe */
8054 doms_cur = doms_new;
8055 dattr_cur = dattr_new;
8056 ndoms_cur = ndoms_new;
8058 register_sched_domain_sysctl();
8060 mutex_unlock(&sched_domains_mutex);
8063 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8064 static void arch_reinit_sched_domains(void)
8068 /* Destroy domains first to force the rebuild */
8069 partition_sched_domains(0, NULL, NULL);
8071 rebuild_sched_domains();
8075 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8077 unsigned int level = 0;
8079 if (sscanf(buf, "%u", &level) != 1)
8083 * level is always be positive so don't check for
8084 * level < POWERSAVINGS_BALANCE_NONE which is 0
8085 * What happens on 0 or 1 byte write,
8086 * need to check for count as well?
8089 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8093 sched_smt_power_savings = level;
8095 sched_mc_power_savings = level;
8097 arch_reinit_sched_domains();
8102 #ifdef CONFIG_SCHED_MC
8103 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8106 return sprintf(page, "%u\n", sched_mc_power_savings);
8108 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8109 const char *buf, size_t count)
8111 return sched_power_savings_store(buf, count, 0);
8113 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8114 sched_mc_power_savings_show,
8115 sched_mc_power_savings_store);
8118 #ifdef CONFIG_SCHED_SMT
8119 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8122 return sprintf(page, "%u\n", sched_smt_power_savings);
8124 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8125 const char *buf, size_t count)
8127 return sched_power_savings_store(buf, count, 1);
8129 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8130 sched_smt_power_savings_show,
8131 sched_smt_power_savings_store);
8134 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8138 #ifdef CONFIG_SCHED_SMT
8140 err = sysfs_create_file(&cls->kset.kobj,
8141 &attr_sched_smt_power_savings.attr);
8143 #ifdef CONFIG_SCHED_MC
8144 if (!err && mc_capable())
8145 err = sysfs_create_file(&cls->kset.kobj,
8146 &attr_sched_mc_power_savings.attr);
8150 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8152 #ifndef CONFIG_CPUSETS
8154 * Add online and remove offline CPUs from the scheduler domains.
8155 * When cpusets are enabled they take over this function.
8157 static int update_sched_domains(struct notifier_block *nfb,
8158 unsigned long action, void *hcpu)
8162 case CPU_ONLINE_FROZEN:
8164 case CPU_DEAD_FROZEN:
8165 partition_sched_domains(1, NULL, NULL);
8174 static int update_runtime(struct notifier_block *nfb,
8175 unsigned long action, void *hcpu)
8177 int cpu = (int)(long)hcpu;
8180 case CPU_DOWN_PREPARE:
8181 case CPU_DOWN_PREPARE_FROZEN:
8182 disable_runtime(cpu_rq(cpu));
8185 case CPU_DOWN_FAILED:
8186 case CPU_DOWN_FAILED_FROZEN:
8188 case CPU_ONLINE_FROZEN:
8189 enable_runtime(cpu_rq(cpu));
8197 void __init sched_init_smp(void)
8199 cpumask_var_t non_isolated_cpus;
8201 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8203 #if defined(CONFIG_NUMA)
8204 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8206 BUG_ON(sched_group_nodes_bycpu == NULL);
8209 mutex_lock(&sched_domains_mutex);
8210 arch_init_sched_domains(cpu_online_mask);
8211 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8212 if (cpumask_empty(non_isolated_cpus))
8213 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8214 mutex_unlock(&sched_domains_mutex);
8217 #ifndef CONFIG_CPUSETS
8218 /* XXX: Theoretical race here - CPU may be hotplugged now */
8219 hotcpu_notifier(update_sched_domains, 0);
8222 /* RT runtime code needs to handle some hotplug events */
8223 hotcpu_notifier(update_runtime, 0);
8227 /* Move init over to a non-isolated CPU */
8228 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8230 sched_init_granularity();
8231 free_cpumask_var(non_isolated_cpus);
8233 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8234 init_sched_rt_class();
8237 void __init sched_init_smp(void)
8239 sched_init_granularity();
8241 #endif /* CONFIG_SMP */
8243 int in_sched_functions(unsigned long addr)
8245 return in_lock_functions(addr) ||
8246 (addr >= (unsigned long)__sched_text_start
8247 && addr < (unsigned long)__sched_text_end);
8250 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8252 cfs_rq->tasks_timeline = RB_ROOT;
8253 INIT_LIST_HEAD(&cfs_rq->tasks);
8254 #ifdef CONFIG_FAIR_GROUP_SCHED
8257 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8260 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8262 struct rt_prio_array *array;
8265 array = &rt_rq->active;
8266 for (i = 0; i < MAX_RT_PRIO; i++) {
8267 INIT_LIST_HEAD(array->queue + i);
8268 __clear_bit(i, array->bitmap);
8270 /* delimiter for bitsearch: */
8271 __set_bit(MAX_RT_PRIO, array->bitmap);
8273 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8274 rt_rq->highest_prio = MAX_RT_PRIO;
8277 rt_rq->rt_nr_migratory = 0;
8278 rt_rq->overloaded = 0;
8282 rt_rq->rt_throttled = 0;
8283 rt_rq->rt_runtime = 0;
8284 spin_lock_init(&rt_rq->rt_runtime_lock);
8286 #ifdef CONFIG_RT_GROUP_SCHED
8287 rt_rq->rt_nr_boosted = 0;
8292 #ifdef CONFIG_FAIR_GROUP_SCHED
8293 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8294 struct sched_entity *se, int cpu, int add,
8295 struct sched_entity *parent)
8297 struct rq *rq = cpu_rq(cpu);
8298 tg->cfs_rq[cpu] = cfs_rq;
8299 init_cfs_rq(cfs_rq, rq);
8302 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8305 /* se could be NULL for init_task_group */
8310 se->cfs_rq = &rq->cfs;
8312 se->cfs_rq = parent->my_q;
8315 se->load.weight = tg->shares;
8316 se->load.inv_weight = 0;
8317 se->parent = parent;
8321 #ifdef CONFIG_RT_GROUP_SCHED
8322 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8323 struct sched_rt_entity *rt_se, int cpu, int add,
8324 struct sched_rt_entity *parent)
8326 struct rq *rq = cpu_rq(cpu);
8328 tg->rt_rq[cpu] = rt_rq;
8329 init_rt_rq(rt_rq, rq);
8331 rt_rq->rt_se = rt_se;
8332 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8334 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8336 tg->rt_se[cpu] = rt_se;
8341 rt_se->rt_rq = &rq->rt;
8343 rt_se->rt_rq = parent->my_q;
8345 rt_se->my_q = rt_rq;
8346 rt_se->parent = parent;
8347 INIT_LIST_HEAD(&rt_se->run_list);
8351 void __init sched_init(void)
8354 unsigned long alloc_size = 0, ptr;
8356 #ifdef CONFIG_FAIR_GROUP_SCHED
8357 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8359 #ifdef CONFIG_RT_GROUP_SCHED
8360 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8362 #ifdef CONFIG_USER_SCHED
8366 * As sched_init() is called before page_alloc is setup,
8367 * we use alloc_bootmem().
8370 ptr = (unsigned long)alloc_bootmem(alloc_size);
8372 #ifdef CONFIG_FAIR_GROUP_SCHED
8373 init_task_group.se = (struct sched_entity **)ptr;
8374 ptr += nr_cpu_ids * sizeof(void **);
8376 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8377 ptr += nr_cpu_ids * sizeof(void **);
8379 #ifdef CONFIG_USER_SCHED
8380 root_task_group.se = (struct sched_entity **)ptr;
8381 ptr += nr_cpu_ids * sizeof(void **);
8383 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8384 ptr += nr_cpu_ids * sizeof(void **);
8385 #endif /* CONFIG_USER_SCHED */
8386 #endif /* CONFIG_FAIR_GROUP_SCHED */
8387 #ifdef CONFIG_RT_GROUP_SCHED
8388 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8389 ptr += nr_cpu_ids * sizeof(void **);
8391 init_task_group.rt_rq = (struct rt_rq **)ptr;
8392 ptr += nr_cpu_ids * sizeof(void **);
8394 #ifdef CONFIG_USER_SCHED
8395 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8396 ptr += nr_cpu_ids * sizeof(void **);
8398 root_task_group.rt_rq = (struct rt_rq **)ptr;
8399 ptr += nr_cpu_ids * sizeof(void **);
8400 #endif /* CONFIG_USER_SCHED */
8401 #endif /* CONFIG_RT_GROUP_SCHED */
8405 init_defrootdomain();
8408 init_rt_bandwidth(&def_rt_bandwidth,
8409 global_rt_period(), global_rt_runtime());
8411 #ifdef CONFIG_RT_GROUP_SCHED
8412 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8413 global_rt_period(), global_rt_runtime());
8414 #ifdef CONFIG_USER_SCHED
8415 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8416 global_rt_period(), RUNTIME_INF);
8417 #endif /* CONFIG_USER_SCHED */
8418 #endif /* CONFIG_RT_GROUP_SCHED */
8420 #ifdef CONFIG_GROUP_SCHED
8421 list_add(&init_task_group.list, &task_groups);
8422 INIT_LIST_HEAD(&init_task_group.children);
8424 #ifdef CONFIG_USER_SCHED
8425 INIT_LIST_HEAD(&root_task_group.children);
8426 init_task_group.parent = &root_task_group;
8427 list_add(&init_task_group.siblings, &root_task_group.children);
8428 #endif /* CONFIG_USER_SCHED */
8429 #endif /* CONFIG_GROUP_SCHED */
8431 for_each_possible_cpu(i) {
8435 spin_lock_init(&rq->lock);
8437 init_cfs_rq(&rq->cfs, rq);
8438 init_rt_rq(&rq->rt, rq);
8439 #ifdef CONFIG_FAIR_GROUP_SCHED
8440 init_task_group.shares = init_task_group_load;
8441 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8442 #ifdef CONFIG_CGROUP_SCHED
8444 * How much cpu bandwidth does init_task_group get?
8446 * In case of task-groups formed thr' the cgroup filesystem, it
8447 * gets 100% of the cpu resources in the system. This overall
8448 * system cpu resource is divided among the tasks of
8449 * init_task_group and its child task-groups in a fair manner,
8450 * based on each entity's (task or task-group's) weight
8451 * (se->load.weight).
8453 * In other words, if init_task_group has 10 tasks of weight
8454 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8455 * then A0's share of the cpu resource is:
8457 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8459 * We achieve this by letting init_task_group's tasks sit
8460 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8462 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8463 #elif defined CONFIG_USER_SCHED
8464 root_task_group.shares = NICE_0_LOAD;
8465 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8467 * In case of task-groups formed thr' the user id of tasks,
8468 * init_task_group represents tasks belonging to root user.
8469 * Hence it forms a sibling of all subsequent groups formed.
8470 * In this case, init_task_group gets only a fraction of overall
8471 * system cpu resource, based on the weight assigned to root
8472 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8473 * by letting tasks of init_task_group sit in a separate cfs_rq
8474 * (init_cfs_rq) and having one entity represent this group of
8475 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8477 init_tg_cfs_entry(&init_task_group,
8478 &per_cpu(init_cfs_rq, i),
8479 &per_cpu(init_sched_entity, i), i, 1,
8480 root_task_group.se[i]);
8483 #endif /* CONFIG_FAIR_GROUP_SCHED */
8485 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8486 #ifdef CONFIG_RT_GROUP_SCHED
8487 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8488 #ifdef CONFIG_CGROUP_SCHED
8489 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8490 #elif defined CONFIG_USER_SCHED
8491 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8492 init_tg_rt_entry(&init_task_group,
8493 &per_cpu(init_rt_rq, i),
8494 &per_cpu(init_sched_rt_entity, i), i, 1,
8495 root_task_group.rt_se[i]);
8499 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8500 rq->cpu_load[j] = 0;
8504 rq->active_balance = 0;
8505 rq->next_balance = jiffies;
8509 rq->migration_thread = NULL;
8510 INIT_LIST_HEAD(&rq->migration_queue);
8511 rq_attach_root(rq, &def_root_domain);
8514 atomic_set(&rq->nr_iowait, 0);
8517 set_load_weight(&init_task);
8519 #ifdef CONFIG_PREEMPT_NOTIFIERS
8520 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8524 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8527 #ifdef CONFIG_RT_MUTEXES
8528 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8532 * The boot idle thread does lazy MMU switching as well:
8534 atomic_inc(&init_mm.mm_count);
8535 enter_lazy_tlb(&init_mm, current);
8538 * Make us the idle thread. Technically, schedule() should not be
8539 * called from this thread, however somewhere below it might be,
8540 * but because we are the idle thread, we just pick up running again
8541 * when this runqueue becomes "idle".
8543 init_idle(current, smp_processor_id());
8545 * During early bootup we pretend to be a normal task:
8547 current->sched_class = &fair_sched_class;
8549 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8550 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
8553 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
8555 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8558 scheduler_running = 1;
8561 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8562 void __might_sleep(char *file, int line)
8565 static unsigned long prev_jiffy; /* ratelimiting */
8567 if ((!in_atomic() && !irqs_disabled()) ||
8568 system_state != SYSTEM_RUNNING || oops_in_progress)
8570 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8572 prev_jiffy = jiffies;
8575 "BUG: sleeping function called from invalid context at %s:%d\n",
8578 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8579 in_atomic(), irqs_disabled(),
8580 current->pid, current->comm);
8582 debug_show_held_locks(current);
8583 if (irqs_disabled())
8584 print_irqtrace_events(current);
8588 EXPORT_SYMBOL(__might_sleep);
8591 #ifdef CONFIG_MAGIC_SYSRQ
8592 static void normalize_task(struct rq *rq, struct task_struct *p)
8596 update_rq_clock(rq);
8597 on_rq = p->se.on_rq;
8599 deactivate_task(rq, p, 0);
8600 __setscheduler(rq, p, SCHED_NORMAL, 0);
8602 activate_task(rq, p, 0);
8603 resched_task(rq->curr);
8607 void normalize_rt_tasks(void)
8609 struct task_struct *g, *p;
8610 unsigned long flags;
8613 read_lock_irqsave(&tasklist_lock, flags);
8614 do_each_thread(g, p) {
8616 * Only normalize user tasks:
8621 p->se.exec_start = 0;
8622 #ifdef CONFIG_SCHEDSTATS
8623 p->se.wait_start = 0;
8624 p->se.sleep_start = 0;
8625 p->se.block_start = 0;
8630 * Renice negative nice level userspace
8633 if (TASK_NICE(p) < 0 && p->mm)
8634 set_user_nice(p, 0);
8638 spin_lock(&p->pi_lock);
8639 rq = __task_rq_lock(p);
8641 normalize_task(rq, p);
8643 __task_rq_unlock(rq);
8644 spin_unlock(&p->pi_lock);
8645 } while_each_thread(g, p);
8647 read_unlock_irqrestore(&tasklist_lock, flags);
8650 #endif /* CONFIG_MAGIC_SYSRQ */
8654 * These functions are only useful for the IA64 MCA handling.
8656 * They can only be called when the whole system has been
8657 * stopped - every CPU needs to be quiescent, and no scheduling
8658 * activity can take place. Using them for anything else would
8659 * be a serious bug, and as a result, they aren't even visible
8660 * under any other configuration.
8664 * curr_task - return the current task for a given cpu.
8665 * @cpu: the processor in question.
8667 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8669 struct task_struct *curr_task(int cpu)
8671 return cpu_curr(cpu);
8675 * set_curr_task - set the current task for a given cpu.
8676 * @cpu: the processor in question.
8677 * @p: the task pointer to set.
8679 * Description: This function must only be used when non-maskable interrupts
8680 * are serviced on a separate stack. It allows the architecture to switch the
8681 * notion of the current task on a cpu in a non-blocking manner. This function
8682 * must be called with all CPU's synchronized, and interrupts disabled, the
8683 * and caller must save the original value of the current task (see
8684 * curr_task() above) and restore that value before reenabling interrupts and
8685 * re-starting the system.
8687 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8689 void set_curr_task(int cpu, struct task_struct *p)
8696 #ifdef CONFIG_FAIR_GROUP_SCHED
8697 static void free_fair_sched_group(struct task_group *tg)
8701 for_each_possible_cpu(i) {
8703 kfree(tg->cfs_rq[i]);
8713 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8715 struct cfs_rq *cfs_rq;
8716 struct sched_entity *se;
8720 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8723 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8727 tg->shares = NICE_0_LOAD;
8729 for_each_possible_cpu(i) {
8732 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8733 GFP_KERNEL, cpu_to_node(i));
8737 se = kzalloc_node(sizeof(struct sched_entity),
8738 GFP_KERNEL, cpu_to_node(i));
8742 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8751 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8753 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8754 &cpu_rq(cpu)->leaf_cfs_rq_list);
8757 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8759 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8761 #else /* !CONFG_FAIR_GROUP_SCHED */
8762 static inline void free_fair_sched_group(struct task_group *tg)
8767 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8772 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8776 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8779 #endif /* CONFIG_FAIR_GROUP_SCHED */
8781 #ifdef CONFIG_RT_GROUP_SCHED
8782 static void free_rt_sched_group(struct task_group *tg)
8786 destroy_rt_bandwidth(&tg->rt_bandwidth);
8788 for_each_possible_cpu(i) {
8790 kfree(tg->rt_rq[i]);
8792 kfree(tg->rt_se[i]);
8800 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8802 struct rt_rq *rt_rq;
8803 struct sched_rt_entity *rt_se;
8807 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8810 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8814 init_rt_bandwidth(&tg->rt_bandwidth,
8815 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8817 for_each_possible_cpu(i) {
8820 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8821 GFP_KERNEL, cpu_to_node(i));
8825 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8826 GFP_KERNEL, cpu_to_node(i));
8830 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8839 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8841 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8842 &cpu_rq(cpu)->leaf_rt_rq_list);
8845 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8847 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8849 #else /* !CONFIG_RT_GROUP_SCHED */
8850 static inline void free_rt_sched_group(struct task_group *tg)
8855 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8860 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8864 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8867 #endif /* CONFIG_RT_GROUP_SCHED */
8869 #ifdef CONFIG_GROUP_SCHED
8870 static void free_sched_group(struct task_group *tg)
8872 free_fair_sched_group(tg);
8873 free_rt_sched_group(tg);
8877 /* allocate runqueue etc for a new task group */
8878 struct task_group *sched_create_group(struct task_group *parent)
8880 struct task_group *tg;
8881 unsigned long flags;
8884 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8886 return ERR_PTR(-ENOMEM);
8888 if (!alloc_fair_sched_group(tg, parent))
8891 if (!alloc_rt_sched_group(tg, parent))
8894 spin_lock_irqsave(&task_group_lock, flags);
8895 for_each_possible_cpu(i) {
8896 register_fair_sched_group(tg, i);
8897 register_rt_sched_group(tg, i);
8899 list_add_rcu(&tg->list, &task_groups);
8901 WARN_ON(!parent); /* root should already exist */
8903 tg->parent = parent;
8904 INIT_LIST_HEAD(&tg->children);
8905 list_add_rcu(&tg->siblings, &parent->children);
8906 spin_unlock_irqrestore(&task_group_lock, flags);
8911 free_sched_group(tg);
8912 return ERR_PTR(-ENOMEM);
8915 /* rcu callback to free various structures associated with a task group */
8916 static void free_sched_group_rcu(struct rcu_head *rhp)
8918 /* now it should be safe to free those cfs_rqs */
8919 free_sched_group(container_of(rhp, struct task_group, rcu));
8922 /* Destroy runqueue etc associated with a task group */
8923 void sched_destroy_group(struct task_group *tg)
8925 unsigned long flags;
8928 spin_lock_irqsave(&task_group_lock, flags);
8929 for_each_possible_cpu(i) {
8930 unregister_fair_sched_group(tg, i);
8931 unregister_rt_sched_group(tg, i);
8933 list_del_rcu(&tg->list);
8934 list_del_rcu(&tg->siblings);
8935 spin_unlock_irqrestore(&task_group_lock, flags);
8937 /* wait for possible concurrent references to cfs_rqs complete */
8938 call_rcu(&tg->rcu, free_sched_group_rcu);
8941 /* change task's runqueue when it moves between groups.
8942 * The caller of this function should have put the task in its new group
8943 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8944 * reflect its new group.
8946 void sched_move_task(struct task_struct *tsk)
8949 unsigned long flags;
8952 rq = task_rq_lock(tsk, &flags);
8954 update_rq_clock(rq);
8956 running = task_current(rq, tsk);
8957 on_rq = tsk->se.on_rq;
8960 dequeue_task(rq, tsk, 0);
8961 if (unlikely(running))
8962 tsk->sched_class->put_prev_task(rq, tsk);
8964 set_task_rq(tsk, task_cpu(tsk));
8966 #ifdef CONFIG_FAIR_GROUP_SCHED
8967 if (tsk->sched_class->moved_group)
8968 tsk->sched_class->moved_group(tsk);
8971 if (unlikely(running))
8972 tsk->sched_class->set_curr_task(rq);
8974 enqueue_task(rq, tsk, 0);
8976 task_rq_unlock(rq, &flags);
8978 #endif /* CONFIG_GROUP_SCHED */
8980 #ifdef CONFIG_FAIR_GROUP_SCHED
8981 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8983 struct cfs_rq *cfs_rq = se->cfs_rq;
8988 dequeue_entity(cfs_rq, se, 0);
8990 se->load.weight = shares;
8991 se->load.inv_weight = 0;
8994 enqueue_entity(cfs_rq, se, 0);
8997 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8999 struct cfs_rq *cfs_rq = se->cfs_rq;
9000 struct rq *rq = cfs_rq->rq;
9001 unsigned long flags;
9003 spin_lock_irqsave(&rq->lock, flags);
9004 __set_se_shares(se, shares);
9005 spin_unlock_irqrestore(&rq->lock, flags);
9008 static DEFINE_MUTEX(shares_mutex);
9010 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9013 unsigned long flags;
9016 * We can't change the weight of the root cgroup.
9021 if (shares < MIN_SHARES)
9022 shares = MIN_SHARES;
9023 else if (shares > MAX_SHARES)
9024 shares = MAX_SHARES;
9026 mutex_lock(&shares_mutex);
9027 if (tg->shares == shares)
9030 spin_lock_irqsave(&task_group_lock, flags);
9031 for_each_possible_cpu(i)
9032 unregister_fair_sched_group(tg, i);
9033 list_del_rcu(&tg->siblings);
9034 spin_unlock_irqrestore(&task_group_lock, flags);
9036 /* wait for any ongoing reference to this group to finish */
9037 synchronize_sched();
9040 * Now we are free to modify the group's share on each cpu
9041 * w/o tripping rebalance_share or load_balance_fair.
9043 tg->shares = shares;
9044 for_each_possible_cpu(i) {
9048 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9049 set_se_shares(tg->se[i], shares);
9053 * Enable load balance activity on this group, by inserting it back on
9054 * each cpu's rq->leaf_cfs_rq_list.
9056 spin_lock_irqsave(&task_group_lock, flags);
9057 for_each_possible_cpu(i)
9058 register_fair_sched_group(tg, i);
9059 list_add_rcu(&tg->siblings, &tg->parent->children);
9060 spin_unlock_irqrestore(&task_group_lock, flags);
9062 mutex_unlock(&shares_mutex);
9066 unsigned long sched_group_shares(struct task_group *tg)
9072 #ifdef CONFIG_RT_GROUP_SCHED
9074 * Ensure that the real time constraints are schedulable.
9076 static DEFINE_MUTEX(rt_constraints_mutex);
9078 static unsigned long to_ratio(u64 period, u64 runtime)
9080 if (runtime == RUNTIME_INF)
9083 return div64_u64(runtime << 20, period);
9086 /* Must be called with tasklist_lock held */
9087 static inline int tg_has_rt_tasks(struct task_group *tg)
9089 struct task_struct *g, *p;
9091 do_each_thread(g, p) {
9092 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9094 } while_each_thread(g, p);
9099 struct rt_schedulable_data {
9100 struct task_group *tg;
9105 static int tg_schedulable(struct task_group *tg, void *data)
9107 struct rt_schedulable_data *d = data;
9108 struct task_group *child;
9109 unsigned long total, sum = 0;
9110 u64 period, runtime;
9112 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9113 runtime = tg->rt_bandwidth.rt_runtime;
9116 period = d->rt_period;
9117 runtime = d->rt_runtime;
9121 * Cannot have more runtime than the period.
9123 if (runtime > period && runtime != RUNTIME_INF)
9127 * Ensure we don't starve existing RT tasks.
9129 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9132 total = to_ratio(period, runtime);
9135 * Nobody can have more than the global setting allows.
9137 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9141 * The sum of our children's runtime should not exceed our own.
9143 list_for_each_entry_rcu(child, &tg->children, siblings) {
9144 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9145 runtime = child->rt_bandwidth.rt_runtime;
9147 if (child == d->tg) {
9148 period = d->rt_period;
9149 runtime = d->rt_runtime;
9152 sum += to_ratio(period, runtime);
9161 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9163 struct rt_schedulable_data data = {
9165 .rt_period = period,
9166 .rt_runtime = runtime,
9169 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9172 static int tg_set_bandwidth(struct task_group *tg,
9173 u64 rt_period, u64 rt_runtime)
9177 mutex_lock(&rt_constraints_mutex);
9178 read_lock(&tasklist_lock);
9179 err = __rt_schedulable(tg, rt_period, rt_runtime);
9183 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9184 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9185 tg->rt_bandwidth.rt_runtime = rt_runtime;
9187 for_each_possible_cpu(i) {
9188 struct rt_rq *rt_rq = tg->rt_rq[i];
9190 spin_lock(&rt_rq->rt_runtime_lock);
9191 rt_rq->rt_runtime = rt_runtime;
9192 spin_unlock(&rt_rq->rt_runtime_lock);
9194 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9196 read_unlock(&tasklist_lock);
9197 mutex_unlock(&rt_constraints_mutex);
9202 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9204 u64 rt_runtime, rt_period;
9206 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9207 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9208 if (rt_runtime_us < 0)
9209 rt_runtime = RUNTIME_INF;
9211 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9214 long sched_group_rt_runtime(struct task_group *tg)
9218 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9221 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9222 do_div(rt_runtime_us, NSEC_PER_USEC);
9223 return rt_runtime_us;
9226 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9228 u64 rt_runtime, rt_period;
9230 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9231 rt_runtime = tg->rt_bandwidth.rt_runtime;
9236 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9239 long sched_group_rt_period(struct task_group *tg)
9243 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9244 do_div(rt_period_us, NSEC_PER_USEC);
9245 return rt_period_us;
9248 static int sched_rt_global_constraints(void)
9250 u64 runtime, period;
9253 if (sysctl_sched_rt_period <= 0)
9256 runtime = global_rt_runtime();
9257 period = global_rt_period();
9260 * Sanity check on the sysctl variables.
9262 if (runtime > period && runtime != RUNTIME_INF)
9265 mutex_lock(&rt_constraints_mutex);
9266 read_lock(&tasklist_lock);
9267 ret = __rt_schedulable(NULL, 0, 0);
9268 read_unlock(&tasklist_lock);
9269 mutex_unlock(&rt_constraints_mutex);
9273 #else /* !CONFIG_RT_GROUP_SCHED */
9274 static int sched_rt_global_constraints(void)
9276 unsigned long flags;
9279 if (sysctl_sched_rt_period <= 0)
9282 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9283 for_each_possible_cpu(i) {
9284 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9286 spin_lock(&rt_rq->rt_runtime_lock);
9287 rt_rq->rt_runtime = global_rt_runtime();
9288 spin_unlock(&rt_rq->rt_runtime_lock);
9290 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9294 #endif /* CONFIG_RT_GROUP_SCHED */
9296 int sched_rt_handler(struct ctl_table *table, int write,
9297 struct file *filp, void __user *buffer, size_t *lenp,
9301 int old_period, old_runtime;
9302 static DEFINE_MUTEX(mutex);
9305 old_period = sysctl_sched_rt_period;
9306 old_runtime = sysctl_sched_rt_runtime;
9308 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9310 if (!ret && write) {
9311 ret = sched_rt_global_constraints();
9313 sysctl_sched_rt_period = old_period;
9314 sysctl_sched_rt_runtime = old_runtime;
9316 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9317 def_rt_bandwidth.rt_period =
9318 ns_to_ktime(global_rt_period());
9321 mutex_unlock(&mutex);
9326 #ifdef CONFIG_CGROUP_SCHED
9328 /* return corresponding task_group object of a cgroup */
9329 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9331 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9332 struct task_group, css);
9335 static struct cgroup_subsys_state *
9336 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9338 struct task_group *tg, *parent;
9340 if (!cgrp->parent) {
9341 /* This is early initialization for the top cgroup */
9342 return &init_task_group.css;
9345 parent = cgroup_tg(cgrp->parent);
9346 tg = sched_create_group(parent);
9348 return ERR_PTR(-ENOMEM);
9354 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9356 struct task_group *tg = cgroup_tg(cgrp);
9358 sched_destroy_group(tg);
9362 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9363 struct task_struct *tsk)
9365 #ifdef CONFIG_RT_GROUP_SCHED
9366 /* Don't accept realtime tasks when there is no way for them to run */
9367 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9370 /* We don't support RT-tasks being in separate groups */
9371 if (tsk->sched_class != &fair_sched_class)
9379 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9380 struct cgroup *old_cont, struct task_struct *tsk)
9382 sched_move_task(tsk);
9385 #ifdef CONFIG_FAIR_GROUP_SCHED
9386 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9389 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9392 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9394 struct task_group *tg = cgroup_tg(cgrp);
9396 return (u64) tg->shares;
9398 #endif /* CONFIG_FAIR_GROUP_SCHED */
9400 #ifdef CONFIG_RT_GROUP_SCHED
9401 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9404 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9407 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9409 return sched_group_rt_runtime(cgroup_tg(cgrp));
9412 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9415 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9418 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9420 return sched_group_rt_period(cgroup_tg(cgrp));
9422 #endif /* CONFIG_RT_GROUP_SCHED */
9424 static struct cftype cpu_files[] = {
9425 #ifdef CONFIG_FAIR_GROUP_SCHED
9428 .read_u64 = cpu_shares_read_u64,
9429 .write_u64 = cpu_shares_write_u64,
9432 #ifdef CONFIG_RT_GROUP_SCHED
9434 .name = "rt_runtime_us",
9435 .read_s64 = cpu_rt_runtime_read,
9436 .write_s64 = cpu_rt_runtime_write,
9439 .name = "rt_period_us",
9440 .read_u64 = cpu_rt_period_read_uint,
9441 .write_u64 = cpu_rt_period_write_uint,
9446 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9448 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9451 struct cgroup_subsys cpu_cgroup_subsys = {
9453 .create = cpu_cgroup_create,
9454 .destroy = cpu_cgroup_destroy,
9455 .can_attach = cpu_cgroup_can_attach,
9456 .attach = cpu_cgroup_attach,
9457 .populate = cpu_cgroup_populate,
9458 .subsys_id = cpu_cgroup_subsys_id,
9462 #endif /* CONFIG_CGROUP_SCHED */
9464 #ifdef CONFIG_CGROUP_CPUACCT
9467 * CPU accounting code for task groups.
9469 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9470 * (balbir@in.ibm.com).
9473 /* track cpu usage of a group of tasks and its child groups */
9475 struct cgroup_subsys_state css;
9476 /* cpuusage holds pointer to a u64-type object on every cpu */
9478 struct cpuacct *parent;
9481 struct cgroup_subsys cpuacct_subsys;
9483 /* return cpu accounting group corresponding to this container */
9484 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9486 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9487 struct cpuacct, css);
9490 /* return cpu accounting group to which this task belongs */
9491 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9493 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9494 struct cpuacct, css);
9497 /* create a new cpu accounting group */
9498 static struct cgroup_subsys_state *cpuacct_create(
9499 struct cgroup_subsys *ss, struct cgroup *cgrp)
9501 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9504 return ERR_PTR(-ENOMEM);
9506 ca->cpuusage = alloc_percpu(u64);
9507 if (!ca->cpuusage) {
9509 return ERR_PTR(-ENOMEM);
9513 ca->parent = cgroup_ca(cgrp->parent);
9518 /* destroy an existing cpu accounting group */
9520 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9522 struct cpuacct *ca = cgroup_ca(cgrp);
9524 free_percpu(ca->cpuusage);
9528 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9530 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9533 #ifndef CONFIG_64BIT
9535 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9537 spin_lock_irq(&cpu_rq(cpu)->lock);
9539 spin_unlock_irq(&cpu_rq(cpu)->lock);
9547 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9549 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9551 #ifndef CONFIG_64BIT
9553 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9555 spin_lock_irq(&cpu_rq(cpu)->lock);
9557 spin_unlock_irq(&cpu_rq(cpu)->lock);
9563 /* return total cpu usage (in nanoseconds) of a group */
9564 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9566 struct cpuacct *ca = cgroup_ca(cgrp);
9567 u64 totalcpuusage = 0;
9570 for_each_present_cpu(i)
9571 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9573 return totalcpuusage;
9576 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9579 struct cpuacct *ca = cgroup_ca(cgrp);
9588 for_each_present_cpu(i)
9589 cpuacct_cpuusage_write(ca, i, 0);
9595 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9598 struct cpuacct *ca = cgroup_ca(cgroup);
9602 for_each_present_cpu(i) {
9603 percpu = cpuacct_cpuusage_read(ca, i);
9604 seq_printf(m, "%llu ", (unsigned long long) percpu);
9606 seq_printf(m, "\n");
9610 static struct cftype files[] = {
9613 .read_u64 = cpuusage_read,
9614 .write_u64 = cpuusage_write,
9617 .name = "usage_percpu",
9618 .read_seq_string = cpuacct_percpu_seq_read,
9623 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9625 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9629 * charge this task's execution time to its accounting group.
9631 * called with rq->lock held.
9633 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9638 if (!cpuacct_subsys.active)
9641 cpu = task_cpu(tsk);
9644 for (; ca; ca = ca->parent) {
9645 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9646 *cpuusage += cputime;
9650 struct cgroup_subsys cpuacct_subsys = {
9652 .create = cpuacct_create,
9653 .destroy = cpuacct_destroy,
9654 .populate = cpuacct_populate,
9655 .subsys_id = cpuacct_subsys_id,
9657 #endif /* CONFIG_CGROUP_CPUACCT */