4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task);
122 DEFINE_TRACE(sched_wakeup);
123 DEFINE_TRACE(sched_wakeup_new);
124 DEFINE_TRACE(sched_switch);
125 DEFINE_TRACE(sched_migrate_task);
129 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
132 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
133 * Since cpu_power is a 'constant', we can use a reciprocal divide.
135 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
137 return reciprocal_divide(load, sg->reciprocal_cpu_power);
141 * Each time a sched group cpu_power is changed,
142 * we must compute its reciprocal value
144 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
146 sg->__cpu_power += val;
147 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
151 static inline int rt_policy(int policy)
153 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
158 static inline int task_has_rt_policy(struct task_struct *p)
160 return rt_policy(p->policy);
164 * This is the priority-queue data structure of the RT scheduling class:
166 struct rt_prio_array {
167 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
168 struct list_head queue[MAX_RT_PRIO];
171 struct rt_bandwidth {
172 /* nests inside the rq lock: */
173 spinlock_t rt_runtime_lock;
176 struct hrtimer rt_period_timer;
179 static struct rt_bandwidth def_rt_bandwidth;
181 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
183 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
185 struct rt_bandwidth *rt_b =
186 container_of(timer, struct rt_bandwidth, rt_period_timer);
192 now = hrtimer_cb_get_time(timer);
193 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
198 idle = do_sched_rt_period_timer(rt_b, overrun);
201 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
205 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
207 rt_b->rt_period = ns_to_ktime(period);
208 rt_b->rt_runtime = runtime;
210 spin_lock_init(&rt_b->rt_runtime_lock);
212 hrtimer_init(&rt_b->rt_period_timer,
213 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
214 rt_b->rt_period_timer.function = sched_rt_period_timer;
217 static inline int rt_bandwidth_enabled(void)
219 return sysctl_sched_rt_runtime >= 0;
222 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
226 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
229 if (hrtimer_active(&rt_b->rt_period_timer))
232 spin_lock(&rt_b->rt_runtime_lock);
234 if (hrtimer_active(&rt_b->rt_period_timer))
237 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
238 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
239 hrtimer_start_expires(&rt_b->rt_period_timer,
242 spin_unlock(&rt_b->rt_runtime_lock);
245 #ifdef CONFIG_RT_GROUP_SCHED
246 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
248 hrtimer_cancel(&rt_b->rt_period_timer);
253 * sched_domains_mutex serializes calls to arch_init_sched_domains,
254 * detach_destroy_domains and partition_sched_domains.
256 static DEFINE_MUTEX(sched_domains_mutex);
258 #ifdef CONFIG_GROUP_SCHED
260 #include <linux/cgroup.h>
264 static LIST_HEAD(task_groups);
266 /* task group related information */
268 #ifdef CONFIG_CGROUP_SCHED
269 struct cgroup_subsys_state css;
272 #ifdef CONFIG_USER_SCHED
276 #ifdef CONFIG_FAIR_GROUP_SCHED
277 /* schedulable entities of this group on each cpu */
278 struct sched_entity **se;
279 /* runqueue "owned" by this group on each cpu */
280 struct cfs_rq **cfs_rq;
281 unsigned long shares;
284 #ifdef CONFIG_RT_GROUP_SCHED
285 struct sched_rt_entity **rt_se;
286 struct rt_rq **rt_rq;
288 struct rt_bandwidth rt_bandwidth;
292 struct list_head list;
294 struct task_group *parent;
295 struct list_head siblings;
296 struct list_head children;
299 #ifdef CONFIG_USER_SCHED
301 /* Helper function to pass uid information to create_sched_user() */
302 void set_tg_uid(struct user_struct *user)
304 user->tg->uid = user->uid;
309 * Every UID task group (including init_task_group aka UID-0) will
310 * be a child to this group.
312 struct task_group root_task_group;
314 #ifdef CONFIG_FAIR_GROUP_SCHED
315 /* Default task group's sched entity on each cpu */
316 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
317 /* Default task group's cfs_rq on each cpu */
318 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
319 #endif /* CONFIG_FAIR_GROUP_SCHED */
321 #ifdef CONFIG_RT_GROUP_SCHED
322 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
323 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
324 #endif /* CONFIG_RT_GROUP_SCHED */
325 #else /* !CONFIG_USER_SCHED */
326 #define root_task_group init_task_group
327 #endif /* CONFIG_USER_SCHED */
329 /* task_group_lock serializes add/remove of task groups and also changes to
330 * a task group's cpu shares.
332 static DEFINE_SPINLOCK(task_group_lock);
335 static int root_task_group_empty(void)
337 return list_empty(&root_task_group.children);
341 #ifdef CONFIG_FAIR_GROUP_SCHED
342 #ifdef CONFIG_USER_SCHED
343 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
344 #else /* !CONFIG_USER_SCHED */
345 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
346 #endif /* CONFIG_USER_SCHED */
349 * A weight of 0 or 1 can cause arithmetics problems.
350 * A weight of a cfs_rq is the sum of weights of which entities
351 * are queued on this cfs_rq, so a weight of a entity should not be
352 * too large, so as the shares value of a task group.
353 * (The default weight is 1024 - so there's no practical
354 * limitation from this.)
357 #define MAX_SHARES (1UL << 18)
359 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
362 /* Default task group.
363 * Every task in system belong to this group at bootup.
365 struct task_group init_task_group;
367 /* return group to which a task belongs */
368 static inline struct task_group *task_group(struct task_struct *p)
370 struct task_group *tg;
372 #ifdef CONFIG_USER_SCHED
374 tg = __task_cred(p)->user->tg;
376 #elif defined(CONFIG_CGROUP_SCHED)
377 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
378 struct task_group, css);
380 tg = &init_task_group;
385 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
386 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
388 #ifdef CONFIG_FAIR_GROUP_SCHED
389 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
390 p->se.parent = task_group(p)->se[cpu];
393 #ifdef CONFIG_RT_GROUP_SCHED
394 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
395 p->rt.parent = task_group(p)->rt_se[cpu];
402 static int root_task_group_empty(void)
408 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
409 static inline struct task_group *task_group(struct task_struct *p)
414 #endif /* CONFIG_GROUP_SCHED */
416 /* CFS-related fields in a runqueue */
418 struct load_weight load;
419 unsigned long nr_running;
424 struct rb_root tasks_timeline;
425 struct rb_node *rb_leftmost;
427 struct list_head tasks;
428 struct list_head *balance_iterator;
431 * 'curr' points to currently running entity on this cfs_rq.
432 * It is set to NULL otherwise (i.e when none are currently running).
434 struct sched_entity *curr, *next, *last;
436 unsigned int nr_spread_over;
438 #ifdef CONFIG_FAIR_GROUP_SCHED
439 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
442 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
443 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
444 * (like users, containers etc.)
446 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
447 * list is used during load balance.
449 struct list_head leaf_cfs_rq_list;
450 struct task_group *tg; /* group that "owns" this runqueue */
454 * the part of load.weight contributed by tasks
456 unsigned long task_weight;
459 * h_load = weight * f(tg)
461 * Where f(tg) is the recursive weight fraction assigned to
464 unsigned long h_load;
467 * this cpu's part of tg->shares
469 unsigned long shares;
472 * load.weight at the time we set shares
474 unsigned long rq_weight;
479 /* Real-Time classes' related field in a runqueue: */
481 struct rt_prio_array active;
482 unsigned long rt_nr_running;
483 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
485 int curr; /* highest queued rt task prio */
487 int next; /* next highest */
492 unsigned long rt_nr_migratory;
494 struct plist_head pushable_tasks;
499 /* Nests inside the rq lock: */
500 spinlock_t rt_runtime_lock;
502 #ifdef CONFIG_RT_GROUP_SCHED
503 unsigned long rt_nr_boosted;
506 struct list_head leaf_rt_rq_list;
507 struct task_group *tg;
508 struct sched_rt_entity *rt_se;
515 * We add the notion of a root-domain which will be used to define per-domain
516 * variables. Each exclusive cpuset essentially defines an island domain by
517 * fully partitioning the member cpus from any other cpuset. Whenever a new
518 * exclusive cpuset is created, we also create and attach a new root-domain
525 cpumask_var_t online;
528 * The "RT overload" flag: it gets set if a CPU has more than
529 * one runnable RT task.
531 cpumask_var_t rto_mask;
534 struct cpupri cpupri;
536 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
538 * Preferred wake up cpu nominated by sched_mc balance that will be
539 * used when most cpus are idle in the system indicating overall very
540 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
542 unsigned int sched_mc_preferred_wakeup_cpu;
547 * By default the system creates a single root-domain with all cpus as
548 * members (mimicking the global state we have today).
550 static struct root_domain def_root_domain;
555 * This is the main, per-CPU runqueue data structure.
557 * Locking rule: those places that want to lock multiple runqueues
558 * (such as the load balancing or the thread migration code), lock
559 * acquire operations must be ordered by ascending &runqueue.
566 * nr_running and cpu_load should be in the same cacheline because
567 * remote CPUs use both these fields when doing load calculation.
569 unsigned long nr_running;
570 #define CPU_LOAD_IDX_MAX 5
571 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
573 unsigned long last_tick_seen;
574 unsigned char in_nohz_recently;
576 /* capture load from *all* tasks on this cpu: */
577 struct load_weight load;
578 unsigned long nr_load_updates;
584 #ifdef CONFIG_FAIR_GROUP_SCHED
585 /* list of leaf cfs_rq on this cpu: */
586 struct list_head leaf_cfs_rq_list;
588 #ifdef CONFIG_RT_GROUP_SCHED
589 struct list_head leaf_rt_rq_list;
593 * This is part of a global counter where only the total sum
594 * over all CPUs matters. A task can increase this counter on
595 * one CPU and if it got migrated afterwards it may decrease
596 * it on another CPU. Always updated under the runqueue lock:
598 unsigned long nr_uninterruptible;
600 struct task_struct *curr, *idle;
601 unsigned long next_balance;
602 struct mm_struct *prev_mm;
609 struct root_domain *rd;
610 struct sched_domain *sd;
612 unsigned char idle_at_tick;
613 /* For active balancing */
616 /* cpu of this runqueue: */
620 unsigned long avg_load_per_task;
622 struct task_struct *migration_thread;
623 struct list_head migration_queue;
626 #ifdef CONFIG_SCHED_HRTICK
628 int hrtick_csd_pending;
629 struct call_single_data hrtick_csd;
631 struct hrtimer hrtick_timer;
634 #ifdef CONFIG_SCHEDSTATS
636 struct sched_info rq_sched_info;
637 unsigned long long rq_cpu_time;
638 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
640 /* sys_sched_yield() stats */
641 unsigned int yld_count;
643 /* schedule() stats */
644 unsigned int sched_switch;
645 unsigned int sched_count;
646 unsigned int sched_goidle;
648 /* try_to_wake_up() stats */
649 unsigned int ttwu_count;
650 unsigned int ttwu_local;
653 unsigned int bkl_count;
657 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
659 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
661 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
664 static inline int cpu_of(struct rq *rq)
674 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
675 * See detach_destroy_domains: synchronize_sched for details.
677 * The domain tree of any CPU may only be accessed from within
678 * preempt-disabled sections.
680 #define for_each_domain(cpu, __sd) \
681 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
683 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
684 #define this_rq() (&__get_cpu_var(runqueues))
685 #define task_rq(p) cpu_rq(task_cpu(p))
686 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
688 static inline void update_rq_clock(struct rq *rq)
690 rq->clock = sched_clock_cpu(cpu_of(rq));
694 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
696 #ifdef CONFIG_SCHED_DEBUG
697 # define const_debug __read_mostly
699 # define const_debug static const
705 * Returns true if the current cpu runqueue is locked.
706 * This interface allows printk to be called with the runqueue lock
707 * held and know whether or not it is OK to wake up the klogd.
709 int runqueue_is_locked(void)
712 struct rq *rq = cpu_rq(cpu);
715 ret = spin_is_locked(&rq->lock);
721 * Debugging: various feature bits
724 #define SCHED_FEAT(name, enabled) \
725 __SCHED_FEAT_##name ,
728 #include "sched_features.h"
733 #define SCHED_FEAT(name, enabled) \
734 (1UL << __SCHED_FEAT_##name) * enabled |
736 const_debug unsigned int sysctl_sched_features =
737 #include "sched_features.h"
742 #ifdef CONFIG_SCHED_DEBUG
743 #define SCHED_FEAT(name, enabled) \
746 static __read_mostly char *sched_feat_names[] = {
747 #include "sched_features.h"
753 static int sched_feat_show(struct seq_file *m, void *v)
757 for (i = 0; sched_feat_names[i]; i++) {
758 if (!(sysctl_sched_features & (1UL << i)))
760 seq_printf(m, "%s ", sched_feat_names[i]);
768 sched_feat_write(struct file *filp, const char __user *ubuf,
769 size_t cnt, loff_t *ppos)
779 if (copy_from_user(&buf, ubuf, cnt))
784 if (strncmp(buf, "NO_", 3) == 0) {
789 for (i = 0; sched_feat_names[i]; i++) {
790 int len = strlen(sched_feat_names[i]);
792 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
794 sysctl_sched_features &= ~(1UL << i);
796 sysctl_sched_features |= (1UL << i);
801 if (!sched_feat_names[i])
809 static int sched_feat_open(struct inode *inode, struct file *filp)
811 return single_open(filp, sched_feat_show, NULL);
814 static struct file_operations sched_feat_fops = {
815 .open = sched_feat_open,
816 .write = sched_feat_write,
819 .release = single_release,
822 static __init int sched_init_debug(void)
824 debugfs_create_file("sched_features", 0644, NULL, NULL,
829 late_initcall(sched_init_debug);
833 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
836 * Number of tasks to iterate in a single balance run.
837 * Limited because this is done with IRQs disabled.
839 const_debug unsigned int sysctl_sched_nr_migrate = 32;
842 * ratelimit for updating the group shares.
845 unsigned int sysctl_sched_shares_ratelimit = 250000;
848 * Inject some fuzzyness into changing the per-cpu group shares
849 * this avoids remote rq-locks at the expense of fairness.
852 unsigned int sysctl_sched_shares_thresh = 4;
855 * period over which we measure -rt task cpu usage in us.
858 unsigned int sysctl_sched_rt_period = 1000000;
860 static __read_mostly int scheduler_running;
863 * part of the period that we allow rt tasks to run in us.
866 int sysctl_sched_rt_runtime = 950000;
868 static inline u64 global_rt_period(void)
870 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
873 static inline u64 global_rt_runtime(void)
875 if (sysctl_sched_rt_runtime < 0)
878 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
881 #ifndef prepare_arch_switch
882 # define prepare_arch_switch(next) do { } while (0)
884 #ifndef finish_arch_switch
885 # define finish_arch_switch(prev) do { } while (0)
888 static inline int task_current(struct rq *rq, struct task_struct *p)
890 return rq->curr == p;
893 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
894 static inline int task_running(struct rq *rq, struct task_struct *p)
896 return task_current(rq, p);
899 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
903 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
905 #ifdef CONFIG_DEBUG_SPINLOCK
906 /* this is a valid case when another task releases the spinlock */
907 rq->lock.owner = current;
910 * If we are tracking spinlock dependencies then we have to
911 * fix up the runqueue lock - which gets 'carried over' from
914 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
916 spin_unlock_irq(&rq->lock);
919 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
920 static inline int task_running(struct rq *rq, struct task_struct *p)
925 return task_current(rq, p);
929 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
933 * We can optimise this out completely for !SMP, because the
934 * SMP rebalancing from interrupt is the only thing that cares
939 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
940 spin_unlock_irq(&rq->lock);
942 spin_unlock(&rq->lock);
946 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
950 * After ->oncpu is cleared, the task can be moved to a different CPU.
951 * We must ensure this doesn't happen until the switch is completely
957 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
961 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
964 * __task_rq_lock - lock the runqueue a given task resides on.
965 * Must be called interrupts disabled.
967 static inline struct rq *__task_rq_lock(struct task_struct *p)
971 struct rq *rq = task_rq(p);
972 spin_lock(&rq->lock);
973 if (likely(rq == task_rq(p)))
975 spin_unlock(&rq->lock);
980 * task_rq_lock - lock the runqueue a given task resides on and disable
981 * interrupts. Note the ordering: we can safely lookup the task_rq without
982 * explicitly disabling preemption.
984 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
990 local_irq_save(*flags);
992 spin_lock(&rq->lock);
993 if (likely(rq == task_rq(p)))
995 spin_unlock_irqrestore(&rq->lock, *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 (test_tsk_need_resched(p))
1206 set_tsk_need_resched(p);
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_need_resched(rq->idle);
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 3
1344 #define WMULT_IDLEPRIO 1431655765
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)
1630 #ifdef CONFIG_PREEMPT
1633 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1634 * way at the expense of forcing extra atomic operations in all
1635 * invocations. This assures that the double_lock is acquired using the
1636 * same underlying policy as the spinlock_t on this architecture, which
1637 * reduces latency compared to the unfair variant below. However, it
1638 * also adds more overhead and therefore may reduce throughput.
1640 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1641 __releases(this_rq->lock)
1642 __acquires(busiest->lock)
1643 __acquires(this_rq->lock)
1645 spin_unlock(&this_rq->lock);
1646 double_rq_lock(this_rq, busiest);
1653 * Unfair double_lock_balance: Optimizes throughput at the expense of
1654 * latency by eliminating extra atomic operations when the locks are
1655 * already in proper order on entry. This favors lower cpu-ids and will
1656 * grant the double lock to lower cpus over higher ids under contention,
1657 * regardless of entry order into the function.
1659 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1660 __releases(this_rq->lock)
1661 __acquires(busiest->lock)
1662 __acquires(this_rq->lock)
1666 if (unlikely(!spin_trylock(&busiest->lock))) {
1667 if (busiest < this_rq) {
1668 spin_unlock(&this_rq->lock);
1669 spin_lock(&busiest->lock);
1670 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1673 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1678 #endif /* CONFIG_PREEMPT */
1681 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1683 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1685 if (unlikely(!irqs_disabled())) {
1686 /* printk() doesn't work good under rq->lock */
1687 spin_unlock(&this_rq->lock);
1691 return _double_lock_balance(this_rq, busiest);
1694 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1695 __releases(busiest->lock)
1697 spin_unlock(&busiest->lock);
1698 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1702 #ifdef CONFIG_FAIR_GROUP_SCHED
1703 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1706 cfs_rq->shares = shares;
1711 #include "sched_stats.h"
1712 #include "sched_idletask.c"
1713 #include "sched_fair.c"
1714 #include "sched_rt.c"
1715 #ifdef CONFIG_SCHED_DEBUG
1716 # include "sched_debug.c"
1719 #define sched_class_highest (&rt_sched_class)
1720 #define for_each_class(class) \
1721 for (class = sched_class_highest; class; class = class->next)
1723 static void inc_nr_running(struct rq *rq)
1728 static void dec_nr_running(struct rq *rq)
1733 static void set_load_weight(struct task_struct *p)
1735 if (task_has_rt_policy(p)) {
1736 p->se.load.weight = prio_to_weight[0] * 2;
1737 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1742 * SCHED_IDLE tasks get minimal weight:
1744 if (p->policy == SCHED_IDLE) {
1745 p->se.load.weight = WEIGHT_IDLEPRIO;
1746 p->se.load.inv_weight = WMULT_IDLEPRIO;
1750 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1751 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1754 static void update_avg(u64 *avg, u64 sample)
1756 s64 diff = sample - *avg;
1760 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1763 p->se.start_runtime = p->se.sum_exec_runtime;
1765 sched_info_queued(p);
1766 p->sched_class->enqueue_task(rq, p, wakeup);
1770 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1773 if (p->se.last_wakeup) {
1774 update_avg(&p->se.avg_overlap,
1775 p->se.sum_exec_runtime - p->se.last_wakeup);
1776 p->se.last_wakeup = 0;
1778 update_avg(&p->se.avg_wakeup,
1779 sysctl_sched_wakeup_granularity);
1783 sched_info_dequeued(p);
1784 p->sched_class->dequeue_task(rq, p, sleep);
1789 * __normal_prio - return the priority that is based on the static prio
1791 static inline int __normal_prio(struct task_struct *p)
1793 return p->static_prio;
1797 * Calculate the expected normal priority: i.e. priority
1798 * without taking RT-inheritance into account. Might be
1799 * boosted by interactivity modifiers. Changes upon fork,
1800 * setprio syscalls, and whenever the interactivity
1801 * estimator recalculates.
1803 static inline int normal_prio(struct task_struct *p)
1807 if (task_has_rt_policy(p))
1808 prio = MAX_RT_PRIO-1 - p->rt_priority;
1810 prio = __normal_prio(p);
1815 * Calculate the current priority, i.e. the priority
1816 * taken into account by the scheduler. This value might
1817 * be boosted by RT tasks, or might be boosted by
1818 * interactivity modifiers. Will be RT if the task got
1819 * RT-boosted. If not then it returns p->normal_prio.
1821 static int effective_prio(struct task_struct *p)
1823 p->normal_prio = normal_prio(p);
1825 * If we are RT tasks or we were boosted to RT priority,
1826 * keep the priority unchanged. Otherwise, update priority
1827 * to the normal priority:
1829 if (!rt_prio(p->prio))
1830 return p->normal_prio;
1835 * activate_task - move a task to the runqueue.
1837 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1839 if (task_contributes_to_load(p))
1840 rq->nr_uninterruptible--;
1842 enqueue_task(rq, p, wakeup);
1847 * deactivate_task - remove a task from the runqueue.
1849 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1851 if (task_contributes_to_load(p))
1852 rq->nr_uninterruptible++;
1854 dequeue_task(rq, p, sleep);
1859 * task_curr - is this task currently executing on a CPU?
1860 * @p: the task in question.
1862 inline int task_curr(const struct task_struct *p)
1864 return cpu_curr(task_cpu(p)) == p;
1867 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1869 set_task_rq(p, cpu);
1872 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1873 * successfuly executed on another CPU. We must ensure that updates of
1874 * per-task data have been completed by this moment.
1877 task_thread_info(p)->cpu = cpu;
1881 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1882 const struct sched_class *prev_class,
1883 int oldprio, int running)
1885 if (prev_class != p->sched_class) {
1886 if (prev_class->switched_from)
1887 prev_class->switched_from(rq, p, running);
1888 p->sched_class->switched_to(rq, p, running);
1890 p->sched_class->prio_changed(rq, p, oldprio, running);
1895 /* Used instead of source_load when we know the type == 0 */
1896 static unsigned long weighted_cpuload(const int cpu)
1898 return cpu_rq(cpu)->load.weight;
1902 * Is this task likely cache-hot:
1905 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1910 * Buddy candidates are cache hot:
1912 if (sched_feat(CACHE_HOT_BUDDY) &&
1913 (&p->se == cfs_rq_of(&p->se)->next ||
1914 &p->se == cfs_rq_of(&p->se)->last))
1917 if (p->sched_class != &fair_sched_class)
1920 if (sysctl_sched_migration_cost == -1)
1922 if (sysctl_sched_migration_cost == 0)
1925 delta = now - p->se.exec_start;
1927 return delta < (s64)sysctl_sched_migration_cost;
1931 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1933 int old_cpu = task_cpu(p);
1934 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1935 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1936 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1939 clock_offset = old_rq->clock - new_rq->clock;
1941 trace_sched_migrate_task(p, task_cpu(p), new_cpu);
1943 #ifdef CONFIG_SCHEDSTATS
1944 if (p->se.wait_start)
1945 p->se.wait_start -= clock_offset;
1946 if (p->se.sleep_start)
1947 p->se.sleep_start -= clock_offset;
1948 if (p->se.block_start)
1949 p->se.block_start -= clock_offset;
1950 if (old_cpu != new_cpu) {
1951 schedstat_inc(p, se.nr_migrations);
1952 if (task_hot(p, old_rq->clock, NULL))
1953 schedstat_inc(p, se.nr_forced2_migrations);
1956 p->se.vruntime -= old_cfsrq->min_vruntime -
1957 new_cfsrq->min_vruntime;
1959 __set_task_cpu(p, new_cpu);
1962 struct migration_req {
1963 struct list_head list;
1965 struct task_struct *task;
1968 struct completion done;
1972 * The task's runqueue lock must be held.
1973 * Returns true if you have to wait for migration thread.
1976 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1978 struct rq *rq = task_rq(p);
1981 * If the task is not on a runqueue (and not running), then
1982 * it is sufficient to simply update the task's cpu field.
1984 if (!p->se.on_rq && !task_running(rq, p)) {
1985 set_task_cpu(p, dest_cpu);
1989 init_completion(&req->done);
1991 req->dest_cpu = dest_cpu;
1992 list_add(&req->list, &rq->migration_queue);
1998 * wait_task_inactive - wait for a thread to unschedule.
2000 * If @match_state is nonzero, it's the @p->state value just checked and
2001 * not expected to change. If it changes, i.e. @p might have woken up,
2002 * then return zero. When we succeed in waiting for @p to be off its CPU,
2003 * we return a positive number (its total switch count). If a second call
2004 * a short while later returns the same number, the caller can be sure that
2005 * @p has remained unscheduled the whole time.
2007 * The caller must ensure that the task *will* unschedule sometime soon,
2008 * else this function might spin for a *long* time. This function can't
2009 * be called with interrupts off, or it may introduce deadlock with
2010 * smp_call_function() if an IPI is sent by the same process we are
2011 * waiting to become inactive.
2013 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2015 unsigned long flags;
2022 * We do the initial early heuristics without holding
2023 * any task-queue locks at all. We'll only try to get
2024 * the runqueue lock when things look like they will
2030 * If the task is actively running on another CPU
2031 * still, just relax and busy-wait without holding
2034 * NOTE! Since we don't hold any locks, it's not
2035 * even sure that "rq" stays as the right runqueue!
2036 * But we don't care, since "task_running()" will
2037 * return false if the runqueue has changed and p
2038 * is actually now running somewhere else!
2040 while (task_running(rq, p)) {
2041 if (match_state && unlikely(p->state != match_state))
2047 * Ok, time to look more closely! We need the rq
2048 * lock now, to be *sure*. If we're wrong, we'll
2049 * just go back and repeat.
2051 rq = task_rq_lock(p, &flags);
2052 trace_sched_wait_task(rq, p);
2053 running = task_running(rq, p);
2054 on_rq = p->se.on_rq;
2056 if (!match_state || p->state == match_state)
2057 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2058 task_rq_unlock(rq, &flags);
2061 * If it changed from the expected state, bail out now.
2063 if (unlikely(!ncsw))
2067 * Was it really running after all now that we
2068 * checked with the proper locks actually held?
2070 * Oops. Go back and try again..
2072 if (unlikely(running)) {
2078 * It's not enough that it's not actively running,
2079 * it must be off the runqueue _entirely_, and not
2082 * So if it was still runnable (but just not actively
2083 * running right now), it's preempted, and we should
2084 * yield - it could be a while.
2086 if (unlikely(on_rq)) {
2087 schedule_timeout_uninterruptible(1);
2092 * Ahh, all good. It wasn't running, and it wasn't
2093 * runnable, which means that it will never become
2094 * running in the future either. We're all done!
2103 * kick_process - kick a running thread to enter/exit the kernel
2104 * @p: the to-be-kicked thread
2106 * Cause a process which is running on another CPU to enter
2107 * kernel-mode, without any delay. (to get signals handled.)
2109 * NOTE: this function doesnt have to take the runqueue lock,
2110 * because all it wants to ensure is that the remote task enters
2111 * the kernel. If the IPI races and the task has been migrated
2112 * to another CPU then no harm is done and the purpose has been
2115 void kick_process(struct task_struct *p)
2121 if ((cpu != smp_processor_id()) && task_curr(p))
2122 smp_send_reschedule(cpu);
2127 * Return a low guess at the load of a migration-source cpu weighted
2128 * according to the scheduling class and "nice" value.
2130 * We want to under-estimate the load of migration sources, to
2131 * balance conservatively.
2133 static unsigned long source_load(int cpu, int type)
2135 struct rq *rq = cpu_rq(cpu);
2136 unsigned long total = weighted_cpuload(cpu);
2138 if (type == 0 || !sched_feat(LB_BIAS))
2141 return min(rq->cpu_load[type-1], total);
2145 * Return a high guess at the load of a migration-target cpu weighted
2146 * according to the scheduling class and "nice" value.
2148 static unsigned long target_load(int cpu, int type)
2150 struct rq *rq = cpu_rq(cpu);
2151 unsigned long total = weighted_cpuload(cpu);
2153 if (type == 0 || !sched_feat(LB_BIAS))
2156 return max(rq->cpu_load[type-1], total);
2160 * find_idlest_group finds and returns the least busy CPU group within the
2163 static struct sched_group *
2164 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2166 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2167 unsigned long min_load = ULONG_MAX, this_load = 0;
2168 int load_idx = sd->forkexec_idx;
2169 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2172 unsigned long load, avg_load;
2176 /* Skip over this group if it has no CPUs allowed */
2177 if (!cpumask_intersects(sched_group_cpus(group),
2181 local_group = cpumask_test_cpu(this_cpu,
2182 sched_group_cpus(group));
2184 /* Tally up the load of all CPUs in the group */
2187 for_each_cpu(i, sched_group_cpus(group)) {
2188 /* Bias balancing toward cpus of our domain */
2190 load = source_load(i, load_idx);
2192 load = target_load(i, load_idx);
2197 /* Adjust by relative CPU power of the group */
2198 avg_load = sg_div_cpu_power(group,
2199 avg_load * SCHED_LOAD_SCALE);
2202 this_load = avg_load;
2204 } else if (avg_load < min_load) {
2205 min_load = avg_load;
2208 } while (group = group->next, group != sd->groups);
2210 if (!idlest || 100*this_load < imbalance*min_load)
2216 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2219 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2221 unsigned long load, min_load = ULONG_MAX;
2225 /* Traverse only the allowed CPUs */
2226 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2227 load = weighted_cpuload(i);
2229 if (load < min_load || (load == min_load && i == this_cpu)) {
2239 * sched_balance_self: balance the current task (running on cpu) in domains
2240 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2243 * Balance, ie. select the least loaded group.
2245 * Returns the target CPU number, or the same CPU if no balancing is needed.
2247 * preempt must be disabled.
2249 static int sched_balance_self(int cpu, int flag)
2251 struct task_struct *t = current;
2252 struct sched_domain *tmp, *sd = NULL;
2254 for_each_domain(cpu, tmp) {
2256 * If power savings logic is enabled for a domain, stop there.
2258 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2260 if (tmp->flags & flag)
2268 struct sched_group *group;
2269 int new_cpu, weight;
2271 if (!(sd->flags & flag)) {
2276 group = find_idlest_group(sd, t, cpu);
2282 new_cpu = find_idlest_cpu(group, t, cpu);
2283 if (new_cpu == -1 || new_cpu == cpu) {
2284 /* Now try balancing at a lower domain level of cpu */
2289 /* Now try balancing at a lower domain level of new_cpu */
2291 weight = cpumask_weight(sched_domain_span(sd));
2293 for_each_domain(cpu, tmp) {
2294 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2296 if (tmp->flags & flag)
2299 /* while loop will break here if sd == NULL */
2305 #endif /* CONFIG_SMP */
2308 * try_to_wake_up - wake up a thread
2309 * @p: the to-be-woken-up thread
2310 * @state: the mask of task states that can be woken
2311 * @sync: do a synchronous wakeup?
2313 * Put it on the run-queue if it's not already there. The "current"
2314 * thread is always on the run-queue (except when the actual
2315 * re-schedule is in progress), and as such you're allowed to do
2316 * the simpler "current->state = TASK_RUNNING" to mark yourself
2317 * runnable without the overhead of this.
2319 * returns failure only if the task is already active.
2321 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2323 int cpu, orig_cpu, this_cpu, success = 0;
2324 unsigned long flags;
2328 if (!sched_feat(SYNC_WAKEUPS))
2332 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2333 struct sched_domain *sd;
2335 this_cpu = raw_smp_processor_id();
2338 for_each_domain(this_cpu, sd) {
2339 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2348 rq = task_rq_lock(p, &flags);
2349 update_rq_clock(rq);
2350 old_state = p->state;
2351 if (!(old_state & state))
2359 this_cpu = smp_processor_id();
2362 if (unlikely(task_running(rq, p)))
2365 cpu = p->sched_class->select_task_rq(p, sync);
2366 if (cpu != orig_cpu) {
2367 set_task_cpu(p, cpu);
2368 task_rq_unlock(rq, &flags);
2369 /* might preempt at this point */
2370 rq = task_rq_lock(p, &flags);
2371 old_state = p->state;
2372 if (!(old_state & state))
2377 this_cpu = smp_processor_id();
2381 #ifdef CONFIG_SCHEDSTATS
2382 schedstat_inc(rq, ttwu_count);
2383 if (cpu == this_cpu)
2384 schedstat_inc(rq, ttwu_local);
2386 struct sched_domain *sd;
2387 for_each_domain(this_cpu, sd) {
2388 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2389 schedstat_inc(sd, ttwu_wake_remote);
2394 #endif /* CONFIG_SCHEDSTATS */
2397 #endif /* CONFIG_SMP */
2398 schedstat_inc(p, se.nr_wakeups);
2400 schedstat_inc(p, se.nr_wakeups_sync);
2401 if (orig_cpu != cpu)
2402 schedstat_inc(p, se.nr_wakeups_migrate);
2403 if (cpu == this_cpu)
2404 schedstat_inc(p, se.nr_wakeups_local);
2406 schedstat_inc(p, se.nr_wakeups_remote);
2407 activate_task(rq, p, 1);
2411 * Only attribute actual wakeups done by this task.
2413 if (!in_interrupt()) {
2414 struct sched_entity *se = ¤t->se;
2415 u64 sample = se->sum_exec_runtime;
2417 if (se->last_wakeup)
2418 sample -= se->last_wakeup;
2420 sample -= se->start_runtime;
2421 update_avg(&se->avg_wakeup, sample);
2423 se->last_wakeup = se->sum_exec_runtime;
2427 trace_sched_wakeup(rq, p, success);
2428 check_preempt_curr(rq, p, sync);
2430 p->state = TASK_RUNNING;
2432 if (p->sched_class->task_wake_up)
2433 p->sched_class->task_wake_up(rq, p);
2436 task_rq_unlock(rq, &flags);
2441 int wake_up_process(struct task_struct *p)
2443 return try_to_wake_up(p, TASK_ALL, 0);
2445 EXPORT_SYMBOL(wake_up_process);
2447 int wake_up_state(struct task_struct *p, unsigned int state)
2449 return try_to_wake_up(p, state, 0);
2453 * Perform scheduler related setup for a newly forked process p.
2454 * p is forked by current.
2456 * __sched_fork() is basic setup used by init_idle() too:
2458 static void __sched_fork(struct task_struct *p)
2460 p->se.exec_start = 0;
2461 p->se.sum_exec_runtime = 0;
2462 p->se.prev_sum_exec_runtime = 0;
2463 p->se.last_wakeup = 0;
2464 p->se.avg_overlap = 0;
2465 p->se.start_runtime = 0;
2466 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2468 #ifdef CONFIG_SCHEDSTATS
2469 p->se.wait_start = 0;
2470 p->se.sum_sleep_runtime = 0;
2471 p->se.sleep_start = 0;
2472 p->se.block_start = 0;
2473 p->se.sleep_max = 0;
2474 p->se.block_max = 0;
2476 p->se.slice_max = 0;
2480 INIT_LIST_HEAD(&p->rt.run_list);
2482 INIT_LIST_HEAD(&p->se.group_node);
2484 #ifdef CONFIG_PREEMPT_NOTIFIERS
2485 INIT_HLIST_HEAD(&p->preempt_notifiers);
2489 * We mark the process as running here, but have not actually
2490 * inserted it onto the runqueue yet. This guarantees that
2491 * nobody will actually run it, and a signal or other external
2492 * event cannot wake it up and insert it on the runqueue either.
2494 p->state = TASK_RUNNING;
2498 * fork()/clone()-time setup:
2500 void sched_fork(struct task_struct *p, int clone_flags)
2502 int cpu = get_cpu();
2507 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2509 set_task_cpu(p, cpu);
2512 * Make sure we do not leak PI boosting priority to the child:
2514 p->prio = current->normal_prio;
2515 if (!rt_prio(p->prio))
2516 p->sched_class = &fair_sched_class;
2518 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2519 if (likely(sched_info_on()))
2520 memset(&p->sched_info, 0, sizeof(p->sched_info));
2522 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2525 #ifdef CONFIG_PREEMPT
2526 /* Want to start with kernel preemption disabled. */
2527 task_thread_info(p)->preempt_count = 1;
2529 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2535 * wake_up_new_task - wake up a newly created task for the first time.
2537 * This function will do some initial scheduler statistics housekeeping
2538 * that must be done for every newly created context, then puts the task
2539 * on the runqueue and wakes it.
2541 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2543 unsigned long flags;
2546 rq = task_rq_lock(p, &flags);
2547 BUG_ON(p->state != TASK_RUNNING);
2548 update_rq_clock(rq);
2550 p->prio = effective_prio(p);
2552 if (!p->sched_class->task_new || !current->se.on_rq) {
2553 activate_task(rq, p, 0);
2556 * Let the scheduling class do new task startup
2557 * management (if any):
2559 p->sched_class->task_new(rq, p);
2562 trace_sched_wakeup_new(rq, p, 1);
2563 check_preempt_curr(rq, p, 0);
2565 if (p->sched_class->task_wake_up)
2566 p->sched_class->task_wake_up(rq, p);
2568 task_rq_unlock(rq, &flags);
2571 #ifdef CONFIG_PREEMPT_NOTIFIERS
2574 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2575 * @notifier: notifier struct to register
2577 void preempt_notifier_register(struct preempt_notifier *notifier)
2579 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2581 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2584 * preempt_notifier_unregister - no longer interested in preemption notifications
2585 * @notifier: notifier struct to unregister
2587 * This is safe to call from within a preemption notifier.
2589 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2591 hlist_del(¬ifier->link);
2593 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2595 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2597 struct preempt_notifier *notifier;
2598 struct hlist_node *node;
2600 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2601 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2605 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2606 struct task_struct *next)
2608 struct preempt_notifier *notifier;
2609 struct hlist_node *node;
2611 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2612 notifier->ops->sched_out(notifier, next);
2615 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2617 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2622 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2623 struct task_struct *next)
2627 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2630 * prepare_task_switch - prepare to switch tasks
2631 * @rq: the runqueue preparing to switch
2632 * @prev: the current task that is being switched out
2633 * @next: the task we are going to switch to.
2635 * This is called with the rq lock held and interrupts off. It must
2636 * be paired with a subsequent finish_task_switch after the context
2639 * prepare_task_switch sets up locking and calls architecture specific
2643 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2644 struct task_struct *next)
2646 fire_sched_out_preempt_notifiers(prev, next);
2647 prepare_lock_switch(rq, next);
2648 prepare_arch_switch(next);
2652 * finish_task_switch - clean up after a task-switch
2653 * @rq: runqueue associated with task-switch
2654 * @prev: the thread we just switched away from.
2656 * finish_task_switch must be called after the context switch, paired
2657 * with a prepare_task_switch call before the context switch.
2658 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2659 * and do any other architecture-specific cleanup actions.
2661 * Note that we may have delayed dropping an mm in context_switch(). If
2662 * so, we finish that here outside of the runqueue lock. (Doing it
2663 * with the lock held can cause deadlocks; see schedule() for
2666 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2667 __releases(rq->lock)
2669 struct mm_struct *mm = rq->prev_mm;
2672 int post_schedule = 0;
2674 if (current->sched_class->needs_post_schedule)
2675 post_schedule = current->sched_class->needs_post_schedule(rq);
2681 * A task struct has one reference for the use as "current".
2682 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2683 * schedule one last time. The schedule call will never return, and
2684 * the scheduled task must drop that reference.
2685 * The test for TASK_DEAD must occur while the runqueue locks are
2686 * still held, otherwise prev could be scheduled on another cpu, die
2687 * there before we look at prev->state, and then the reference would
2689 * Manfred Spraul <manfred@colorfullife.com>
2691 prev_state = prev->state;
2692 finish_arch_switch(prev);
2693 finish_lock_switch(rq, prev);
2696 current->sched_class->post_schedule(rq);
2699 fire_sched_in_preempt_notifiers(current);
2702 if (unlikely(prev_state == TASK_DEAD)) {
2704 * Remove function-return probe instances associated with this
2705 * task and put them back on the free list.
2707 kprobe_flush_task(prev);
2708 put_task_struct(prev);
2713 * schedule_tail - first thing a freshly forked thread must call.
2714 * @prev: the thread we just switched away from.
2716 asmlinkage void schedule_tail(struct task_struct *prev)
2717 __releases(rq->lock)
2719 struct rq *rq = this_rq();
2721 finish_task_switch(rq, prev);
2722 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2723 /* In this case, finish_task_switch does not reenable preemption */
2726 if (current->set_child_tid)
2727 put_user(task_pid_vnr(current), current->set_child_tid);
2731 * context_switch - switch to the new MM and the new
2732 * thread's register state.
2735 context_switch(struct rq *rq, struct task_struct *prev,
2736 struct task_struct *next)
2738 struct mm_struct *mm, *oldmm;
2740 prepare_task_switch(rq, prev, next);
2741 trace_sched_switch(rq, prev, next);
2743 oldmm = prev->active_mm;
2745 * For paravirt, this is coupled with an exit in switch_to to
2746 * combine the page table reload and the switch backend into
2749 arch_enter_lazy_cpu_mode();
2751 if (unlikely(!mm)) {
2752 next->active_mm = oldmm;
2753 atomic_inc(&oldmm->mm_count);
2754 enter_lazy_tlb(oldmm, next);
2756 switch_mm(oldmm, mm, next);
2758 if (unlikely(!prev->mm)) {
2759 prev->active_mm = NULL;
2760 rq->prev_mm = oldmm;
2763 * Since the runqueue lock will be released by the next
2764 * task (which is an invalid locking op but in the case
2765 * of the scheduler it's an obvious special-case), so we
2766 * do an early lockdep release here:
2768 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2769 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2772 /* Here we just switch the register state and the stack. */
2773 switch_to(prev, next, prev);
2777 * this_rq must be evaluated again because prev may have moved
2778 * CPUs since it called schedule(), thus the 'rq' on its stack
2779 * frame will be invalid.
2781 finish_task_switch(this_rq(), prev);
2785 * nr_running, nr_uninterruptible and nr_context_switches:
2787 * externally visible scheduler statistics: current number of runnable
2788 * threads, current number of uninterruptible-sleeping threads, total
2789 * number of context switches performed since bootup.
2791 unsigned long nr_running(void)
2793 unsigned long i, sum = 0;
2795 for_each_online_cpu(i)
2796 sum += cpu_rq(i)->nr_running;
2801 unsigned long nr_uninterruptible(void)
2803 unsigned long i, sum = 0;
2805 for_each_possible_cpu(i)
2806 sum += cpu_rq(i)->nr_uninterruptible;
2809 * Since we read the counters lockless, it might be slightly
2810 * inaccurate. Do not allow it to go below zero though:
2812 if (unlikely((long)sum < 0))
2818 unsigned long long nr_context_switches(void)
2821 unsigned long long sum = 0;
2823 for_each_possible_cpu(i)
2824 sum += cpu_rq(i)->nr_switches;
2829 unsigned long nr_iowait(void)
2831 unsigned long i, sum = 0;
2833 for_each_possible_cpu(i)
2834 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2839 unsigned long nr_active(void)
2841 unsigned long i, running = 0, uninterruptible = 0;
2843 for_each_online_cpu(i) {
2844 running += cpu_rq(i)->nr_running;
2845 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2848 if (unlikely((long)uninterruptible < 0))
2849 uninterruptible = 0;
2851 return running + uninterruptible;
2855 * Update rq->cpu_load[] statistics. This function is usually called every
2856 * scheduler tick (TICK_NSEC).
2858 static void update_cpu_load(struct rq *this_rq)
2860 unsigned long this_load = this_rq->load.weight;
2863 this_rq->nr_load_updates++;
2865 /* Update our load: */
2866 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2867 unsigned long old_load, new_load;
2869 /* scale is effectively 1 << i now, and >> i divides by scale */
2871 old_load = this_rq->cpu_load[i];
2872 new_load = this_load;
2874 * Round up the averaging division if load is increasing. This
2875 * prevents us from getting stuck on 9 if the load is 10, for
2878 if (new_load > old_load)
2879 new_load += scale-1;
2880 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2887 * double_rq_lock - safely lock two runqueues
2889 * Note this does not disable interrupts like task_rq_lock,
2890 * you need to do so manually before calling.
2892 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2893 __acquires(rq1->lock)
2894 __acquires(rq2->lock)
2896 BUG_ON(!irqs_disabled());
2898 spin_lock(&rq1->lock);
2899 __acquire(rq2->lock); /* Fake it out ;) */
2902 spin_lock(&rq1->lock);
2903 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2905 spin_lock(&rq2->lock);
2906 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2909 update_rq_clock(rq1);
2910 update_rq_clock(rq2);
2914 * double_rq_unlock - safely unlock two runqueues
2916 * Note this does not restore interrupts like task_rq_unlock,
2917 * you need to do so manually after calling.
2919 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2920 __releases(rq1->lock)
2921 __releases(rq2->lock)
2923 spin_unlock(&rq1->lock);
2925 spin_unlock(&rq2->lock);
2927 __release(rq2->lock);
2931 * If dest_cpu is allowed for this process, migrate the task to it.
2932 * This is accomplished by forcing the cpu_allowed mask to only
2933 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2934 * the cpu_allowed mask is restored.
2936 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2938 struct migration_req req;
2939 unsigned long flags;
2942 rq = task_rq_lock(p, &flags);
2943 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
2944 || unlikely(!cpu_active(dest_cpu)))
2947 /* force the process onto the specified CPU */
2948 if (migrate_task(p, dest_cpu, &req)) {
2949 /* Need to wait for migration thread (might exit: take ref). */
2950 struct task_struct *mt = rq->migration_thread;
2952 get_task_struct(mt);
2953 task_rq_unlock(rq, &flags);
2954 wake_up_process(mt);
2955 put_task_struct(mt);
2956 wait_for_completion(&req.done);
2961 task_rq_unlock(rq, &flags);
2965 * sched_exec - execve() is a valuable balancing opportunity, because at
2966 * this point the task has the smallest effective memory and cache footprint.
2968 void sched_exec(void)
2970 int new_cpu, this_cpu = get_cpu();
2971 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2973 if (new_cpu != this_cpu)
2974 sched_migrate_task(current, new_cpu);
2978 * pull_task - move a task from a remote runqueue to the local runqueue.
2979 * Both runqueues must be locked.
2981 static void pull_task(struct rq *src_rq, struct task_struct *p,
2982 struct rq *this_rq, int this_cpu)
2984 deactivate_task(src_rq, p, 0);
2985 set_task_cpu(p, this_cpu);
2986 activate_task(this_rq, p, 0);
2988 * Note that idle threads have a prio of MAX_PRIO, for this test
2989 * to be always true for them.
2991 check_preempt_curr(this_rq, p, 0);
2995 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2998 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2999 struct sched_domain *sd, enum cpu_idle_type idle,
3002 int tsk_cache_hot = 0;
3004 * We do not migrate tasks that are:
3005 * 1) running (obviously), or
3006 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3007 * 3) are cache-hot on their current CPU.
3009 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3010 schedstat_inc(p, se.nr_failed_migrations_affine);
3015 if (task_running(rq, p)) {
3016 schedstat_inc(p, se.nr_failed_migrations_running);
3021 * Aggressive migration if:
3022 * 1) task is cache cold, or
3023 * 2) too many balance attempts have failed.
3026 tsk_cache_hot = task_hot(p, rq->clock, sd);
3027 if (!tsk_cache_hot ||
3028 sd->nr_balance_failed > sd->cache_nice_tries) {
3029 #ifdef CONFIG_SCHEDSTATS
3030 if (tsk_cache_hot) {
3031 schedstat_inc(sd, lb_hot_gained[idle]);
3032 schedstat_inc(p, se.nr_forced_migrations);
3038 if (tsk_cache_hot) {
3039 schedstat_inc(p, se.nr_failed_migrations_hot);
3045 static unsigned long
3046 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3047 unsigned long max_load_move, struct sched_domain *sd,
3048 enum cpu_idle_type idle, int *all_pinned,
3049 int *this_best_prio, struct rq_iterator *iterator)
3051 int loops = 0, pulled = 0, pinned = 0;
3052 struct task_struct *p;
3053 long rem_load_move = max_load_move;
3055 if (max_load_move == 0)
3061 * Start the load-balancing iterator:
3063 p = iterator->start(iterator->arg);
3065 if (!p || loops++ > sysctl_sched_nr_migrate)
3068 if ((p->se.load.weight >> 1) > rem_load_move ||
3069 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3070 p = iterator->next(iterator->arg);
3074 pull_task(busiest, p, this_rq, this_cpu);
3076 rem_load_move -= p->se.load.weight;
3078 #ifdef CONFIG_PREEMPT
3080 * NEWIDLE balancing is a source of latency, so preemptible kernels
3081 * will stop after the first task is pulled to minimize the critical
3084 if (idle == CPU_NEWLY_IDLE)
3089 * We only want to steal up to the prescribed amount of weighted load.
3091 if (rem_load_move > 0) {
3092 if (p->prio < *this_best_prio)
3093 *this_best_prio = p->prio;
3094 p = iterator->next(iterator->arg);
3099 * Right now, this is one of only two places pull_task() is called,
3100 * so we can safely collect pull_task() stats here rather than
3101 * inside pull_task().
3103 schedstat_add(sd, lb_gained[idle], pulled);
3106 *all_pinned = pinned;
3108 return max_load_move - rem_load_move;
3112 * move_tasks tries to move up to max_load_move weighted load from busiest to
3113 * this_rq, as part of a balancing operation within domain "sd".
3114 * Returns 1 if successful and 0 otherwise.
3116 * Called with both runqueues locked.
3118 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3119 unsigned long max_load_move,
3120 struct sched_domain *sd, enum cpu_idle_type idle,
3123 const struct sched_class *class = sched_class_highest;
3124 unsigned long total_load_moved = 0;
3125 int this_best_prio = this_rq->curr->prio;
3129 class->load_balance(this_rq, this_cpu, busiest,
3130 max_load_move - total_load_moved,
3131 sd, idle, all_pinned, &this_best_prio);
3132 class = class->next;
3134 #ifdef CONFIG_PREEMPT
3136 * NEWIDLE balancing is a source of latency, so preemptible
3137 * kernels will stop after the first task is pulled to minimize
3138 * the critical section.
3140 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3143 } while (class && max_load_move > total_load_moved);
3145 return total_load_moved > 0;
3149 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3150 struct sched_domain *sd, enum cpu_idle_type idle,
3151 struct rq_iterator *iterator)
3153 struct task_struct *p = iterator->start(iterator->arg);
3157 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3158 pull_task(busiest, p, this_rq, this_cpu);
3160 * Right now, this is only the second place pull_task()
3161 * is called, so we can safely collect pull_task()
3162 * stats here rather than inside pull_task().
3164 schedstat_inc(sd, lb_gained[idle]);
3168 p = iterator->next(iterator->arg);
3175 * move_one_task tries to move exactly one task from busiest to this_rq, as
3176 * part of active balancing operations within "domain".
3177 * Returns 1 if successful and 0 otherwise.
3179 * Called with both runqueues locked.
3181 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3182 struct sched_domain *sd, enum cpu_idle_type idle)
3184 const struct sched_class *class;
3186 for (class = sched_class_highest; class; class = class->next)
3187 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3192 /********** Helpers for find_busiest_group ************************/
3194 * sd_lb_stats - Structure to store the statistics of a sched_domain
3195 * during load balancing.
3197 struct sd_lb_stats {
3198 struct sched_group *busiest; /* Busiest group in this sd */
3199 struct sched_group *this; /* Local group in this sd */
3200 unsigned long total_load; /* Total load of all groups in sd */
3201 unsigned long total_pwr; /* Total power of all groups in sd */
3202 unsigned long avg_load; /* Average load across all groups in sd */
3204 /** Statistics of this group */
3205 unsigned long this_load;
3206 unsigned long this_load_per_task;
3207 unsigned long this_nr_running;
3209 /* Statistics of the busiest group */
3210 unsigned long max_load;
3211 unsigned long busiest_load_per_task;
3212 unsigned long busiest_nr_running;
3214 int group_imb; /* Is there imbalance in this sd */
3215 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3216 int power_savings_balance; /* Is powersave balance needed for this sd */
3217 struct sched_group *group_min; /* Least loaded group in sd */
3218 struct sched_group *group_leader; /* Group which relieves group_min */
3219 unsigned long min_load_per_task; /* load_per_task in group_min */
3220 unsigned long leader_nr_running; /* Nr running of group_leader */
3221 unsigned long min_nr_running; /* Nr running of group_min */
3226 * sg_lb_stats - stats of a sched_group required for load_balancing
3228 struct sg_lb_stats {
3229 unsigned long avg_load; /*Avg load across the CPUs of the group */
3230 unsigned long group_load; /* Total load over the CPUs of the group */
3231 unsigned long sum_nr_running; /* Nr tasks running in the group */
3232 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3233 unsigned long group_capacity;
3234 int group_imb; /* Is there an imbalance in the group ? */
3238 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3239 * @group: The group whose first cpu is to be returned.
3241 static inline unsigned int group_first_cpu(struct sched_group *group)
3243 return cpumask_first(sched_group_cpus(group));
3247 * get_sd_load_idx - Obtain the load index for a given sched domain.
3248 * @sd: The sched_domain whose load_idx is to be obtained.
3249 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3251 static inline int get_sd_load_idx(struct sched_domain *sd,
3252 enum cpu_idle_type idle)
3258 load_idx = sd->busy_idx;
3261 case CPU_NEWLY_IDLE:
3262 load_idx = sd->newidle_idx;
3265 load_idx = sd->idle_idx;
3273 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3275 * init_sd_power_savings_stats - Initialize power savings statistics for
3276 * the given sched_domain, during load balancing.
3278 * @sd: Sched domain whose power-savings statistics are to be initialized.
3279 * @sds: Variable containing the statistics for sd.
3280 * @idle: Idle status of the CPU at which we're performing load-balancing.
3282 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3283 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3286 * Busy processors will not participate in power savings
3289 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3290 sds->power_savings_balance = 0;
3292 sds->power_savings_balance = 1;
3293 sds->min_nr_running = ULONG_MAX;
3294 sds->leader_nr_running = 0;
3299 * update_sd_power_savings_stats - Update the power saving stats for a
3300 * sched_domain while performing load balancing.
3302 * @group: sched_group belonging to the sched_domain under consideration.
3303 * @sds: Variable containing the statistics of the sched_domain
3304 * @local_group: Does group contain the CPU for which we're performing
3306 * @sgs: Variable containing the statistics of the group.
3308 static inline void update_sd_power_savings_stats(struct sched_group *group,
3309 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3312 if (!sds->power_savings_balance)
3316 * If the local group is idle or completely loaded
3317 * no need to do power savings balance at this domain
3319 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3320 !sds->this_nr_running))
3321 sds->power_savings_balance = 0;
3324 * If a group is already running at full capacity or idle,
3325 * don't include that group in power savings calculations
3327 if (!sds->power_savings_balance ||
3328 sgs->sum_nr_running >= sgs->group_capacity ||
3329 !sgs->sum_nr_running)
3333 * Calculate the group which has the least non-idle load.
3334 * This is the group from where we need to pick up the load
3337 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3338 (sgs->sum_nr_running == sds->min_nr_running &&
3339 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3340 sds->group_min = group;
3341 sds->min_nr_running = sgs->sum_nr_running;
3342 sds->min_load_per_task = sgs->sum_weighted_load /
3343 sgs->sum_nr_running;
3347 * Calculate the group which is almost near its
3348 * capacity but still has some space to pick up some load
3349 * from other group and save more power
3351 if (sgs->sum_nr_running > sgs->group_capacity - 1)
3354 if (sgs->sum_nr_running > sds->leader_nr_running ||
3355 (sgs->sum_nr_running == sds->leader_nr_running &&
3356 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3357 sds->group_leader = group;
3358 sds->leader_nr_running = sgs->sum_nr_running;
3363 * check_power_save_busiest_group - Check if we have potential to perform
3364 * some power-savings balance. If yes, set the busiest group to be
3365 * the least loaded group in the sched_domain, so that it's CPUs can
3368 * @sds: Variable containing the statistics of the sched_domain
3369 * under consideration.
3370 * @this_cpu: Cpu at which we're currently performing load-balancing.
3371 * @imbalance: Variable to store the imbalance.
3373 * Returns 1 if there is potential to perform power-savings balance.
3376 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3377 int this_cpu, unsigned long *imbalance)
3379 if (!sds->power_savings_balance)
3382 if (sds->this != sds->group_leader ||
3383 sds->group_leader == sds->group_min)
3386 *imbalance = sds->min_load_per_task;
3387 sds->busiest = sds->group_min;
3389 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3390 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3391 group_first_cpu(sds->group_leader);
3397 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3398 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3399 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3404 static inline void update_sd_power_savings_stats(struct sched_group *group,
3405 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3410 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3411 int this_cpu, unsigned long *imbalance)
3415 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3419 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3420 * @group: sched_group whose statistics are to be updated.
3421 * @this_cpu: Cpu for which load balance is currently performed.
3422 * @idle: Idle status of this_cpu
3423 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3424 * @sd_idle: Idle status of the sched_domain containing group.
3425 * @local_group: Does group contain this_cpu.
3426 * @cpus: Set of cpus considered for load balancing.
3427 * @balance: Should we balance.
3428 * @sgs: variable to hold the statistics for this group.
3430 static inline void update_sg_lb_stats(struct sched_group *group, int this_cpu,
3431 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3432 int local_group, const struct cpumask *cpus,
3433 int *balance, struct sg_lb_stats *sgs)
3435 unsigned long load, max_cpu_load, min_cpu_load;
3437 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3438 unsigned long sum_avg_load_per_task;
3439 unsigned long avg_load_per_task;
3442 balance_cpu = group_first_cpu(group);
3444 /* Tally up the load of all CPUs in the group */
3445 sum_avg_load_per_task = avg_load_per_task = 0;
3447 min_cpu_load = ~0UL;
3449 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3450 struct rq *rq = cpu_rq(i);
3452 if (*sd_idle && rq->nr_running)
3455 /* Bias balancing toward cpus of our domain */
3457 if (idle_cpu(i) && !first_idle_cpu) {
3462 load = target_load(i, load_idx);
3464 load = source_load(i, load_idx);
3465 if (load > max_cpu_load)
3466 max_cpu_load = load;
3467 if (min_cpu_load > load)
3468 min_cpu_load = load;
3471 sgs->group_load += load;
3472 sgs->sum_nr_running += rq->nr_running;
3473 sgs->sum_weighted_load += weighted_cpuload(i);
3475 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3479 * First idle cpu or the first cpu(busiest) in this sched group
3480 * is eligible for doing load balancing at this and above
3481 * domains. In the newly idle case, we will allow all the cpu's
3482 * to do the newly idle load balance.
3484 if (idle != CPU_NEWLY_IDLE && local_group &&
3485 balance_cpu != this_cpu && balance) {
3490 /* Adjust by relative CPU power of the group */
3491 sgs->avg_load = sg_div_cpu_power(group,
3492 sgs->group_load * SCHED_LOAD_SCALE);
3496 * Consider the group unbalanced when the imbalance is larger
3497 * than the average weight of two tasks.
3499 * APZ: with cgroup the avg task weight can vary wildly and
3500 * might not be a suitable number - should we keep a
3501 * normalized nr_running number somewhere that negates
3504 avg_load_per_task = sg_div_cpu_power(group,
3505 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3507 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3510 sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3515 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3516 * @sd: sched_domain whose statistics are to be updated.
3517 * @this_cpu: Cpu for which load balance is currently performed.
3518 * @idle: Idle status of this_cpu
3519 * @sd_idle: Idle status of the sched_domain containing group.
3520 * @cpus: Set of cpus considered for load balancing.
3521 * @balance: Should we balance.
3522 * @sds: variable to hold the statistics for this sched_domain.
3524 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3525 enum cpu_idle_type idle, int *sd_idle,
3526 const struct cpumask *cpus, int *balance,
3527 struct sd_lb_stats *sds)
3529 struct sched_group *group = sd->groups;
3530 struct sg_lb_stats sgs;
3533 init_sd_power_savings_stats(sd, sds, idle);
3534 load_idx = get_sd_load_idx(sd, idle);
3539 local_group = cpumask_test_cpu(this_cpu,
3540 sched_group_cpus(group));
3541 memset(&sgs, 0, sizeof(sgs));
3542 update_sg_lb_stats(group, this_cpu, idle, load_idx, sd_idle,
3543 local_group, cpus, balance, &sgs);
3545 if (local_group && balance && !(*balance))
3548 sds->total_load += sgs.group_load;
3549 sds->total_pwr += group->__cpu_power;
3552 sds->this_load = sgs.avg_load;
3554 sds->this_nr_running = sgs.sum_nr_running;
3555 sds->this_load_per_task = sgs.sum_weighted_load;
3556 } else if (sgs.avg_load > sds->max_load &&
3557 (sgs.sum_nr_running > sgs.group_capacity ||
3559 sds->max_load = sgs.avg_load;
3560 sds->busiest = group;
3561 sds->busiest_nr_running = sgs.sum_nr_running;
3562 sds->busiest_load_per_task = sgs.sum_weighted_load;
3563 sds->group_imb = sgs.group_imb;
3566 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3567 group = group->next;
3568 } while (group != sd->groups);
3573 * fix_small_imbalance - Calculate the minor imbalance that exists
3574 * amongst the groups of a sched_domain, during
3576 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3577 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3578 * @imbalance: Variable to store the imbalance.
3580 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3581 int this_cpu, unsigned long *imbalance)
3583 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3584 unsigned int imbn = 2;
3586 if (sds->this_nr_running) {
3587 sds->this_load_per_task /= sds->this_nr_running;
3588 if (sds->busiest_load_per_task >
3589 sds->this_load_per_task)
3592 sds->this_load_per_task =
3593 cpu_avg_load_per_task(this_cpu);
3595 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3596 sds->busiest_load_per_task * imbn) {
3597 *imbalance = sds->busiest_load_per_task;
3602 * OK, we don't have enough imbalance to justify moving tasks,
3603 * however we may be able to increase total CPU power used by
3607 pwr_now += sds->busiest->__cpu_power *
3608 min(sds->busiest_load_per_task, sds->max_load);
3609 pwr_now += sds->this->__cpu_power *
3610 min(sds->this_load_per_task, sds->this_load);
3611 pwr_now /= SCHED_LOAD_SCALE;
3613 /* Amount of load we'd subtract */
3614 tmp = sg_div_cpu_power(sds->busiest,
3615 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3616 if (sds->max_load > tmp)
3617 pwr_move += sds->busiest->__cpu_power *
3618 min(sds->busiest_load_per_task, sds->max_load - tmp);
3620 /* Amount of load we'd add */
3621 if (sds->max_load * sds->busiest->__cpu_power <
3622 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3623 tmp = sg_div_cpu_power(sds->this,
3624 sds->max_load * sds->busiest->__cpu_power);
3626 tmp = sg_div_cpu_power(sds->this,
3627 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3628 pwr_move += sds->this->__cpu_power *
3629 min(sds->this_load_per_task, sds->this_load + tmp);
3630 pwr_move /= SCHED_LOAD_SCALE;
3632 /* Move if we gain throughput */
3633 if (pwr_move > pwr_now)
3634 *imbalance = sds->busiest_load_per_task;
3638 * calculate_imbalance - Calculate the amount of imbalance present within the
3639 * groups of a given sched_domain during load balance.
3640 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3641 * @this_cpu: Cpu for which currently load balance is being performed.
3642 * @imbalance: The variable to store the imbalance.
3644 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3645 unsigned long *imbalance)
3647 unsigned long max_pull;
3649 * In the presence of smp nice balancing, certain scenarios can have
3650 * max load less than avg load(as we skip the groups at or below
3651 * its cpu_power, while calculating max_load..)
3653 if (sds->max_load < sds->avg_load) {
3655 return fix_small_imbalance(sds, this_cpu, imbalance);
3658 /* Don't want to pull so many tasks that a group would go idle */
3659 max_pull = min(sds->max_load - sds->avg_load,
3660 sds->max_load - sds->busiest_load_per_task);
3662 /* How much load to actually move to equalise the imbalance */
3663 *imbalance = min(max_pull * sds->busiest->__cpu_power,
3664 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
3668 * if *imbalance is less than the average load per runnable task
3669 * there is no gaurantee that any tasks will be moved so we'll have
3670 * a think about bumping its value to force at least one task to be
3673 if (*imbalance < sds->busiest_load_per_task)
3674 return fix_small_imbalance(sds, this_cpu, imbalance);
3677 /******* find_busiest_group() helpers end here *********************/
3680 * find_busiest_group finds and returns the busiest CPU group within the
3681 * domain. It calculates and returns the amount of weighted load which
3682 * should be moved to restore balance via the imbalance parameter.
3684 static struct sched_group *
3685 find_busiest_group(struct sched_domain *sd, int this_cpu,
3686 unsigned long *imbalance, enum cpu_idle_type idle,
3687 int *sd_idle, const struct cpumask *cpus, int *balance)
3689 struct sd_lb_stats sds;
3691 memset(&sds, 0, sizeof(sds));
3694 * Compute the various statistics relavent for load balancing at
3697 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3700 if (balance && !(*balance))
3703 if (!sds.busiest || sds.this_load >= sds.max_load
3704 || sds.busiest_nr_running == 0)
3707 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3709 if (sds.this_load >= sds.avg_load ||
3710 100*sds.max_load <= sd->imbalance_pct * sds.this_load)
3713 sds.busiest_load_per_task /= sds.busiest_nr_running;
3715 sds.busiest_load_per_task =
3716 min(sds.busiest_load_per_task, sds.avg_load);
3719 * We're trying to get all the cpus to the average_load, so we don't
3720 * want to push ourselves above the average load, nor do we wish to
3721 * reduce the max loaded cpu below the average load, as either of these
3722 * actions would just result in more rebalancing later, and ping-pong
3723 * tasks around. Thus we look for the minimum possible imbalance.
3724 * Negative imbalances (*we* are more loaded than anyone else) will
3725 * be counted as no imbalance for these purposes -- we can't fix that
3726 * by pulling tasks to us. Be careful of negative numbers as they'll
3727 * appear as very large values with unsigned longs.
3729 if (sds.max_load <= sds.busiest_load_per_task)
3732 /* Looks like there is an imbalance. Compute it */
3733 calculate_imbalance(&sds, this_cpu, imbalance);
3738 * There is no obvious imbalance. But check if we can do some balancing
3741 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3749 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3752 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3753 unsigned long imbalance, const struct cpumask *cpus)
3755 struct rq *busiest = NULL, *rq;
3756 unsigned long max_load = 0;
3759 for_each_cpu(i, sched_group_cpus(group)) {
3762 if (!cpumask_test_cpu(i, cpus))
3766 wl = weighted_cpuload(i);
3768 if (rq->nr_running == 1 && wl > imbalance)
3771 if (wl > max_load) {
3781 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3782 * so long as it is large enough.
3784 #define MAX_PINNED_INTERVAL 512
3787 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3788 * tasks if there is an imbalance.
3790 static int load_balance(int this_cpu, struct rq *this_rq,
3791 struct sched_domain *sd, enum cpu_idle_type idle,
3792 int *balance, struct cpumask *cpus)
3794 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3795 struct sched_group *group;
3796 unsigned long imbalance;
3798 unsigned long flags;
3800 cpumask_setall(cpus);
3803 * When power savings policy is enabled for the parent domain, idle
3804 * sibling can pick up load irrespective of busy siblings. In this case,
3805 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3806 * portraying it as CPU_NOT_IDLE.
3808 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3809 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3812 schedstat_inc(sd, lb_count[idle]);
3816 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3823 schedstat_inc(sd, lb_nobusyg[idle]);
3827 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3829 schedstat_inc(sd, lb_nobusyq[idle]);
3833 BUG_ON(busiest == this_rq);
3835 schedstat_add(sd, lb_imbalance[idle], imbalance);
3838 if (busiest->nr_running > 1) {
3840 * Attempt to move tasks. If find_busiest_group has found
3841 * an imbalance but busiest->nr_running <= 1, the group is
3842 * still unbalanced. ld_moved simply stays zero, so it is
3843 * correctly treated as an imbalance.
3845 local_irq_save(flags);
3846 double_rq_lock(this_rq, busiest);
3847 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3848 imbalance, sd, idle, &all_pinned);
3849 double_rq_unlock(this_rq, busiest);
3850 local_irq_restore(flags);
3853 * some other cpu did the load balance for us.
3855 if (ld_moved && this_cpu != smp_processor_id())
3856 resched_cpu(this_cpu);
3858 /* All tasks on this runqueue were pinned by CPU affinity */
3859 if (unlikely(all_pinned)) {
3860 cpumask_clear_cpu(cpu_of(busiest), cpus);
3861 if (!cpumask_empty(cpus))
3868 schedstat_inc(sd, lb_failed[idle]);
3869 sd->nr_balance_failed++;
3871 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3873 spin_lock_irqsave(&busiest->lock, flags);
3875 /* don't kick the migration_thread, if the curr
3876 * task on busiest cpu can't be moved to this_cpu
3878 if (!cpumask_test_cpu(this_cpu,
3879 &busiest->curr->cpus_allowed)) {
3880 spin_unlock_irqrestore(&busiest->lock, flags);
3882 goto out_one_pinned;
3885 if (!busiest->active_balance) {
3886 busiest->active_balance = 1;
3887 busiest->push_cpu = this_cpu;
3890 spin_unlock_irqrestore(&busiest->lock, flags);
3892 wake_up_process(busiest->migration_thread);
3895 * We've kicked active balancing, reset the failure
3898 sd->nr_balance_failed = sd->cache_nice_tries+1;
3901 sd->nr_balance_failed = 0;
3903 if (likely(!active_balance)) {
3904 /* We were unbalanced, so reset the balancing interval */
3905 sd->balance_interval = sd->min_interval;
3908 * If we've begun active balancing, start to back off. This
3909 * case may not be covered by the all_pinned logic if there
3910 * is only 1 task on the busy runqueue (because we don't call
3913 if (sd->balance_interval < sd->max_interval)
3914 sd->balance_interval *= 2;
3917 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3918 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3924 schedstat_inc(sd, lb_balanced[idle]);
3926 sd->nr_balance_failed = 0;
3929 /* tune up the balancing interval */
3930 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3931 (sd->balance_interval < sd->max_interval))
3932 sd->balance_interval *= 2;
3934 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3935 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3946 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3947 * tasks if there is an imbalance.
3949 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3950 * this_rq is locked.
3953 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3954 struct cpumask *cpus)
3956 struct sched_group *group;
3957 struct rq *busiest = NULL;
3958 unsigned long imbalance;
3963 cpumask_setall(cpus);
3966 * When power savings policy is enabled for the parent domain, idle
3967 * sibling can pick up load irrespective of busy siblings. In this case,
3968 * let the state of idle sibling percolate up as IDLE, instead of
3969 * portraying it as CPU_NOT_IDLE.
3971 if (sd->flags & SD_SHARE_CPUPOWER &&
3972 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3975 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3977 update_shares_locked(this_rq, sd);
3978 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3979 &sd_idle, cpus, NULL);
3981 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3985 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3987 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3991 BUG_ON(busiest == this_rq);
3993 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3996 if (busiest->nr_running > 1) {
3997 /* Attempt to move tasks */
3998 double_lock_balance(this_rq, busiest);
3999 /* this_rq->clock is already updated */
4000 update_rq_clock(busiest);
4001 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4002 imbalance, sd, CPU_NEWLY_IDLE,
4004 double_unlock_balance(this_rq, busiest);
4006 if (unlikely(all_pinned)) {
4007 cpumask_clear_cpu(cpu_of(busiest), cpus);
4008 if (!cpumask_empty(cpus))
4014 int active_balance = 0;
4016 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4017 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4018 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4021 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4024 if (sd->nr_balance_failed++ < 2)
4028 * The only task running in a non-idle cpu can be moved to this
4029 * cpu in an attempt to completely freeup the other CPU
4030 * package. The same method used to move task in load_balance()
4031 * have been extended for load_balance_newidle() to speedup
4032 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4034 * The package power saving logic comes from
4035 * find_busiest_group(). If there are no imbalance, then
4036 * f_b_g() will return NULL. However when sched_mc={1,2} then
4037 * f_b_g() will select a group from which a running task may be
4038 * pulled to this cpu in order to make the other package idle.
4039 * If there is no opportunity to make a package idle and if
4040 * there are no imbalance, then f_b_g() will return NULL and no
4041 * action will be taken in load_balance_newidle().
4043 * Under normal task pull operation due to imbalance, there
4044 * will be more than one task in the source run queue and
4045 * move_tasks() will succeed. ld_moved will be true and this
4046 * active balance code will not be triggered.
4049 /* Lock busiest in correct order while this_rq is held */
4050 double_lock_balance(this_rq, busiest);
4053 * don't kick the migration_thread, if the curr
4054 * task on busiest cpu can't be moved to this_cpu
4056 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4057 double_unlock_balance(this_rq, busiest);
4062 if (!busiest->active_balance) {
4063 busiest->active_balance = 1;
4064 busiest->push_cpu = this_cpu;
4068 double_unlock_balance(this_rq, busiest);
4070 * Should not call ttwu while holding a rq->lock
4072 spin_unlock(&this_rq->lock);
4074 wake_up_process(busiest->migration_thread);
4075 spin_lock(&this_rq->lock);
4078 sd->nr_balance_failed = 0;
4080 update_shares_locked(this_rq, sd);
4084 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4085 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4086 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4088 sd->nr_balance_failed = 0;
4094 * idle_balance is called by schedule() if this_cpu is about to become
4095 * idle. Attempts to pull tasks from other CPUs.
4097 static void idle_balance(int this_cpu, struct rq *this_rq)
4099 struct sched_domain *sd;
4100 int pulled_task = 0;
4101 unsigned long next_balance = jiffies + HZ;
4102 cpumask_var_t tmpmask;
4104 if (!alloc_cpumask_var(&tmpmask, GFP_ATOMIC))
4107 for_each_domain(this_cpu, sd) {
4108 unsigned long interval;
4110 if (!(sd->flags & SD_LOAD_BALANCE))
4113 if (sd->flags & SD_BALANCE_NEWIDLE)
4114 /* If we've pulled tasks over stop searching: */
4115 pulled_task = load_balance_newidle(this_cpu, this_rq,
4118 interval = msecs_to_jiffies(sd->balance_interval);
4119 if (time_after(next_balance, sd->last_balance + interval))
4120 next_balance = sd->last_balance + interval;
4124 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4126 * We are going idle. next_balance may be set based on
4127 * a busy processor. So reset next_balance.
4129 this_rq->next_balance = next_balance;
4131 free_cpumask_var(tmpmask);
4135 * active_load_balance is run by migration threads. It pushes running tasks
4136 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4137 * running on each physical CPU where possible, and avoids physical /
4138 * logical imbalances.
4140 * Called with busiest_rq locked.
4142 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4144 int target_cpu = busiest_rq->push_cpu;
4145 struct sched_domain *sd;
4146 struct rq *target_rq;
4148 /* Is there any task to move? */
4149 if (busiest_rq->nr_running <= 1)
4152 target_rq = cpu_rq(target_cpu);
4155 * This condition is "impossible", if it occurs
4156 * we need to fix it. Originally reported by
4157 * Bjorn Helgaas on a 128-cpu setup.
4159 BUG_ON(busiest_rq == target_rq);
4161 /* move a task from busiest_rq to target_rq */
4162 double_lock_balance(busiest_rq, target_rq);
4163 update_rq_clock(busiest_rq);
4164 update_rq_clock(target_rq);
4166 /* Search for an sd spanning us and the target CPU. */
4167 for_each_domain(target_cpu, sd) {
4168 if ((sd->flags & SD_LOAD_BALANCE) &&
4169 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4174 schedstat_inc(sd, alb_count);
4176 if (move_one_task(target_rq, target_cpu, busiest_rq,
4178 schedstat_inc(sd, alb_pushed);
4180 schedstat_inc(sd, alb_failed);
4182 double_unlock_balance(busiest_rq, target_rq);
4187 atomic_t load_balancer;
4188 cpumask_var_t cpu_mask;
4189 } nohz ____cacheline_aligned = {
4190 .load_balancer = ATOMIC_INIT(-1),
4194 * This routine will try to nominate the ilb (idle load balancing)
4195 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4196 * load balancing on behalf of all those cpus. If all the cpus in the system
4197 * go into this tickless mode, then there will be no ilb owner (as there is
4198 * no need for one) and all the cpus will sleep till the next wakeup event
4201 * For the ilb owner, tick is not stopped. And this tick will be used
4202 * for idle load balancing. ilb owner will still be part of
4205 * While stopping the tick, this cpu will become the ilb owner if there
4206 * is no other owner. And will be the owner till that cpu becomes busy
4207 * or if all cpus in the system stop their ticks at which point
4208 * there is no need for ilb owner.
4210 * When the ilb owner becomes busy, it nominates another owner, during the
4211 * next busy scheduler_tick()
4213 int select_nohz_load_balancer(int stop_tick)
4215 int cpu = smp_processor_id();
4218 cpu_rq(cpu)->in_nohz_recently = 1;
4220 if (!cpu_active(cpu)) {
4221 if (atomic_read(&nohz.load_balancer) != cpu)
4225 * If we are going offline and still the leader,
4228 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4234 cpumask_set_cpu(cpu, nohz.cpu_mask);
4236 /* time for ilb owner also to sleep */
4237 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4238 if (atomic_read(&nohz.load_balancer) == cpu)
4239 atomic_set(&nohz.load_balancer, -1);
4243 if (atomic_read(&nohz.load_balancer) == -1) {
4244 /* make me the ilb owner */
4245 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4247 } else if (atomic_read(&nohz.load_balancer) == cpu)
4250 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4253 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4255 if (atomic_read(&nohz.load_balancer) == cpu)
4256 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4263 static DEFINE_SPINLOCK(balancing);
4266 * It checks each scheduling domain to see if it is due to be balanced,
4267 * and initiates a balancing operation if so.
4269 * Balancing parameters are set up in arch_init_sched_domains.
4271 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4274 struct rq *rq = cpu_rq(cpu);
4275 unsigned long interval;
4276 struct sched_domain *sd;
4277 /* Earliest time when we have to do rebalance again */
4278 unsigned long next_balance = jiffies + 60*HZ;
4279 int update_next_balance = 0;
4283 /* Fails alloc? Rebalancing probably not a priority right now. */
4284 if (!alloc_cpumask_var(&tmp, GFP_ATOMIC))
4287 for_each_domain(cpu, sd) {
4288 if (!(sd->flags & SD_LOAD_BALANCE))
4291 interval = sd->balance_interval;
4292 if (idle != CPU_IDLE)
4293 interval *= sd->busy_factor;
4295 /* scale ms to jiffies */
4296 interval = msecs_to_jiffies(interval);
4297 if (unlikely(!interval))
4299 if (interval > HZ*NR_CPUS/10)
4300 interval = HZ*NR_CPUS/10;
4302 need_serialize = sd->flags & SD_SERIALIZE;
4304 if (need_serialize) {
4305 if (!spin_trylock(&balancing))
4309 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4310 if (load_balance(cpu, rq, sd, idle, &balance, tmp)) {
4312 * We've pulled tasks over so either we're no
4313 * longer idle, or one of our SMT siblings is
4316 idle = CPU_NOT_IDLE;
4318 sd->last_balance = jiffies;
4321 spin_unlock(&balancing);
4323 if (time_after(next_balance, sd->last_balance + interval)) {
4324 next_balance = sd->last_balance + interval;
4325 update_next_balance = 1;
4329 * Stop the load balance at this level. There is another
4330 * CPU in our sched group which is doing load balancing more
4338 * next_balance will be updated only when there is a need.
4339 * When the cpu is attached to null domain for ex, it will not be
4342 if (likely(update_next_balance))
4343 rq->next_balance = next_balance;
4345 free_cpumask_var(tmp);
4349 * run_rebalance_domains is triggered when needed from the scheduler tick.
4350 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4351 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4353 static void run_rebalance_domains(struct softirq_action *h)
4355 int this_cpu = smp_processor_id();
4356 struct rq *this_rq = cpu_rq(this_cpu);
4357 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4358 CPU_IDLE : CPU_NOT_IDLE;
4360 rebalance_domains(this_cpu, idle);
4364 * If this cpu is the owner for idle load balancing, then do the
4365 * balancing on behalf of the other idle cpus whose ticks are
4368 if (this_rq->idle_at_tick &&
4369 atomic_read(&nohz.load_balancer) == this_cpu) {
4373 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4374 if (balance_cpu == this_cpu)
4378 * If this cpu gets work to do, stop the load balancing
4379 * work being done for other cpus. Next load
4380 * balancing owner will pick it up.
4385 rebalance_domains(balance_cpu, CPU_IDLE);
4387 rq = cpu_rq(balance_cpu);
4388 if (time_after(this_rq->next_balance, rq->next_balance))
4389 this_rq->next_balance = rq->next_balance;
4395 static inline int on_null_domain(int cpu)
4397 return !rcu_dereference(cpu_rq(cpu)->sd);
4401 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4403 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4404 * idle load balancing owner or decide to stop the periodic load balancing,
4405 * if the whole system is idle.
4407 static inline void trigger_load_balance(struct rq *rq, int cpu)
4411 * If we were in the nohz mode recently and busy at the current
4412 * scheduler tick, then check if we need to nominate new idle
4415 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4416 rq->in_nohz_recently = 0;
4418 if (atomic_read(&nohz.load_balancer) == cpu) {
4419 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4420 atomic_set(&nohz.load_balancer, -1);
4423 if (atomic_read(&nohz.load_balancer) == -1) {
4425 * simple selection for now: Nominate the
4426 * first cpu in the nohz list to be the next
4429 * TBD: Traverse the sched domains and nominate
4430 * the nearest cpu in the nohz.cpu_mask.
4432 int ilb = cpumask_first(nohz.cpu_mask);
4434 if (ilb < nr_cpu_ids)
4440 * If this cpu is idle and doing idle load balancing for all the
4441 * cpus with ticks stopped, is it time for that to stop?
4443 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4444 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4450 * If this cpu is idle and the idle load balancing is done by
4451 * someone else, then no need raise the SCHED_SOFTIRQ
4453 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4454 cpumask_test_cpu(cpu, nohz.cpu_mask))
4457 /* Don't need to rebalance while attached to NULL domain */
4458 if (time_after_eq(jiffies, rq->next_balance) &&
4459 likely(!on_null_domain(cpu)))
4460 raise_softirq(SCHED_SOFTIRQ);
4463 #else /* CONFIG_SMP */
4466 * on UP we do not need to balance between CPUs:
4468 static inline void idle_balance(int cpu, struct rq *rq)
4474 DEFINE_PER_CPU(struct kernel_stat, kstat);
4476 EXPORT_PER_CPU_SYMBOL(kstat);
4479 * Return any ns on the sched_clock that have not yet been banked in
4480 * @p in case that task is currently running.
4482 unsigned long long task_delta_exec(struct task_struct *p)
4484 unsigned long flags;
4488 rq = task_rq_lock(p, &flags);
4490 if (task_current(rq, p)) {
4493 update_rq_clock(rq);
4494 delta_exec = rq->clock - p->se.exec_start;
4495 if ((s64)delta_exec > 0)
4499 task_rq_unlock(rq, &flags);
4505 * Account user cpu time to a process.
4506 * @p: the process that the cpu time gets accounted to
4507 * @cputime: the cpu time spent in user space since the last update
4508 * @cputime_scaled: cputime scaled by cpu frequency
4510 void account_user_time(struct task_struct *p, cputime_t cputime,
4511 cputime_t cputime_scaled)
4513 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4516 /* Add user time to process. */
4517 p->utime = cputime_add(p->utime, cputime);
4518 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4519 account_group_user_time(p, cputime);
4521 /* Add user time to cpustat. */
4522 tmp = cputime_to_cputime64(cputime);
4523 if (TASK_NICE(p) > 0)
4524 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4526 cpustat->user = cputime64_add(cpustat->user, tmp);
4527 /* Account for user time used */
4528 acct_update_integrals(p);
4532 * Account guest cpu time to a process.
4533 * @p: the process that the cpu time gets accounted to
4534 * @cputime: the cpu time spent in virtual machine since the last update
4535 * @cputime_scaled: cputime scaled by cpu frequency
4537 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4538 cputime_t cputime_scaled)
4541 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4543 tmp = cputime_to_cputime64(cputime);
4545 /* Add guest time to process. */
4546 p->utime = cputime_add(p->utime, cputime);
4547 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4548 account_group_user_time(p, cputime);
4549 p->gtime = cputime_add(p->gtime, cputime);
4551 /* Add guest time to cpustat. */
4552 cpustat->user = cputime64_add(cpustat->user, tmp);
4553 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4557 * Account system cpu time to a process.
4558 * @p: the process that the cpu time gets accounted to
4559 * @hardirq_offset: the offset to subtract from hardirq_count()
4560 * @cputime: the cpu time spent in kernel space since the last update
4561 * @cputime_scaled: cputime scaled by cpu frequency
4563 void account_system_time(struct task_struct *p, int hardirq_offset,
4564 cputime_t cputime, cputime_t cputime_scaled)
4566 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4569 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4570 account_guest_time(p, cputime, cputime_scaled);
4574 /* Add system time to process. */
4575 p->stime = cputime_add(p->stime, cputime);
4576 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4577 account_group_system_time(p, cputime);
4579 /* Add system time to cpustat. */
4580 tmp = cputime_to_cputime64(cputime);
4581 if (hardirq_count() - hardirq_offset)
4582 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4583 else if (softirq_count())
4584 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4586 cpustat->system = cputime64_add(cpustat->system, tmp);
4588 /* Account for system time used */
4589 acct_update_integrals(p);
4593 * Account for involuntary wait time.
4594 * @steal: the cpu time spent in involuntary wait
4596 void account_steal_time(cputime_t cputime)
4598 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4599 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4601 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4605 * Account for idle time.
4606 * @cputime: the cpu time spent in idle wait
4608 void account_idle_time(cputime_t cputime)
4610 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4611 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4612 struct rq *rq = this_rq();
4614 if (atomic_read(&rq->nr_iowait) > 0)
4615 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4617 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4620 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4623 * Account a single tick of cpu time.
4624 * @p: the process that the cpu time gets accounted to
4625 * @user_tick: indicates if the tick is a user or a system tick
4627 void account_process_tick(struct task_struct *p, int user_tick)
4629 cputime_t one_jiffy = jiffies_to_cputime(1);
4630 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
4631 struct rq *rq = this_rq();
4634 account_user_time(p, one_jiffy, one_jiffy_scaled);
4635 else if (p != rq->idle)
4636 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
4639 account_idle_time(one_jiffy);
4643 * Account multiple ticks of steal time.
4644 * @p: the process from which the cpu time has been stolen
4645 * @ticks: number of stolen ticks
4647 void account_steal_ticks(unsigned long ticks)
4649 account_steal_time(jiffies_to_cputime(ticks));
4653 * Account multiple ticks of idle time.
4654 * @ticks: number of stolen ticks
4656 void account_idle_ticks(unsigned long ticks)
4658 account_idle_time(jiffies_to_cputime(ticks));
4664 * Use precise platform statistics if available:
4666 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4667 cputime_t task_utime(struct task_struct *p)
4672 cputime_t task_stime(struct task_struct *p)
4677 cputime_t task_utime(struct task_struct *p)
4679 clock_t utime = cputime_to_clock_t(p->utime),
4680 total = utime + cputime_to_clock_t(p->stime);
4684 * Use CFS's precise accounting:
4686 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4690 do_div(temp, total);
4692 utime = (clock_t)temp;
4694 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4695 return p->prev_utime;
4698 cputime_t task_stime(struct task_struct *p)
4703 * Use CFS's precise accounting. (we subtract utime from
4704 * the total, to make sure the total observed by userspace
4705 * grows monotonically - apps rely on that):
4707 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4708 cputime_to_clock_t(task_utime(p));
4711 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4713 return p->prev_stime;
4717 inline cputime_t task_gtime(struct task_struct *p)
4723 * This function gets called by the timer code, with HZ frequency.
4724 * We call it with interrupts disabled.
4726 * It also gets called by the fork code, when changing the parent's
4729 void scheduler_tick(void)
4731 int cpu = smp_processor_id();
4732 struct rq *rq = cpu_rq(cpu);
4733 struct task_struct *curr = rq->curr;
4737 spin_lock(&rq->lock);
4738 update_rq_clock(rq);
4739 update_cpu_load(rq);
4740 curr->sched_class->task_tick(rq, curr, 0);
4741 spin_unlock(&rq->lock);
4744 rq->idle_at_tick = idle_cpu(cpu);
4745 trigger_load_balance(rq, cpu);
4749 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4750 defined(CONFIG_PREEMPT_TRACER))
4752 static inline unsigned long get_parent_ip(unsigned long addr)
4754 if (in_lock_functions(addr)) {
4755 addr = CALLER_ADDR2;
4756 if (in_lock_functions(addr))
4757 addr = CALLER_ADDR3;
4762 void __kprobes add_preempt_count(int val)
4764 #ifdef CONFIG_DEBUG_PREEMPT
4768 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4771 preempt_count() += val;
4772 #ifdef CONFIG_DEBUG_PREEMPT
4774 * Spinlock count overflowing soon?
4776 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4779 if (preempt_count() == val)
4780 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4782 EXPORT_SYMBOL(add_preempt_count);
4784 void __kprobes sub_preempt_count(int val)
4786 #ifdef CONFIG_DEBUG_PREEMPT
4790 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4793 * Is the spinlock portion underflowing?
4795 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4796 !(preempt_count() & PREEMPT_MASK)))
4800 if (preempt_count() == val)
4801 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4802 preempt_count() -= val;
4804 EXPORT_SYMBOL(sub_preempt_count);
4809 * Print scheduling while atomic bug:
4811 static noinline void __schedule_bug(struct task_struct *prev)
4813 struct pt_regs *regs = get_irq_regs();
4815 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4816 prev->comm, prev->pid, preempt_count());
4818 debug_show_held_locks(prev);
4820 if (irqs_disabled())
4821 print_irqtrace_events(prev);
4830 * Various schedule()-time debugging checks and statistics:
4832 static inline void schedule_debug(struct task_struct *prev)
4835 * Test if we are atomic. Since do_exit() needs to call into
4836 * schedule() atomically, we ignore that path for now.
4837 * Otherwise, whine if we are scheduling when we should not be.
4839 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4840 __schedule_bug(prev);
4842 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4844 schedstat_inc(this_rq(), sched_count);
4845 #ifdef CONFIG_SCHEDSTATS
4846 if (unlikely(prev->lock_depth >= 0)) {
4847 schedstat_inc(this_rq(), bkl_count);
4848 schedstat_inc(prev, sched_info.bkl_count);
4853 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4855 if (prev->state == TASK_RUNNING) {
4856 u64 runtime = prev->se.sum_exec_runtime;
4858 runtime -= prev->se.prev_sum_exec_runtime;
4859 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
4862 * In order to avoid avg_overlap growing stale when we are
4863 * indeed overlapping and hence not getting put to sleep, grow
4864 * the avg_overlap on preemption.
4866 * We use the average preemption runtime because that
4867 * correlates to the amount of cache footprint a task can
4870 update_avg(&prev->se.avg_overlap, runtime);
4872 prev->sched_class->put_prev_task(rq, prev);
4876 * Pick up the highest-prio task:
4878 static inline struct task_struct *
4879 pick_next_task(struct rq *rq)
4881 const struct sched_class *class;
4882 struct task_struct *p;
4885 * Optimization: we know that if all tasks are in
4886 * the fair class we can call that function directly:
4888 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4889 p = fair_sched_class.pick_next_task(rq);
4894 class = sched_class_highest;
4896 p = class->pick_next_task(rq);
4900 * Will never be NULL as the idle class always
4901 * returns a non-NULL p:
4903 class = class->next;
4908 * schedule() is the main scheduler function.
4910 asmlinkage void __sched schedule(void)
4912 struct task_struct *prev, *next;
4913 unsigned long *switch_count;
4919 cpu = smp_processor_id();
4923 switch_count = &prev->nivcsw;
4925 release_kernel_lock(prev);
4926 need_resched_nonpreemptible:
4928 schedule_debug(prev);
4930 if (sched_feat(HRTICK))
4933 spin_lock_irq(&rq->lock);
4934 update_rq_clock(rq);
4935 clear_tsk_need_resched(prev);
4937 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4938 if (unlikely(signal_pending_state(prev->state, prev)))
4939 prev->state = TASK_RUNNING;
4941 deactivate_task(rq, prev, 1);
4942 switch_count = &prev->nvcsw;
4946 if (prev->sched_class->pre_schedule)
4947 prev->sched_class->pre_schedule(rq, prev);
4950 if (unlikely(!rq->nr_running))
4951 idle_balance(cpu, rq);
4953 put_prev_task(rq, prev);
4954 next = pick_next_task(rq);
4956 if (likely(prev != next)) {
4957 sched_info_switch(prev, next);
4963 context_switch(rq, prev, next); /* unlocks the rq */
4965 * the context switch might have flipped the stack from under
4966 * us, hence refresh the local variables.
4968 cpu = smp_processor_id();
4971 spin_unlock_irq(&rq->lock);
4973 if (unlikely(reacquire_kernel_lock(current) < 0))
4974 goto need_resched_nonpreemptible;
4976 preempt_enable_no_resched();
4977 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4980 EXPORT_SYMBOL(schedule);
4982 #ifdef CONFIG_PREEMPT
4984 * this is the entry point to schedule() from in-kernel preemption
4985 * off of preempt_enable. Kernel preemptions off return from interrupt
4986 * occur there and call schedule directly.
4988 asmlinkage void __sched preempt_schedule(void)
4990 struct thread_info *ti = current_thread_info();
4993 * If there is a non-zero preempt_count or interrupts are disabled,
4994 * we do not want to preempt the current task. Just return..
4996 if (likely(ti->preempt_count || irqs_disabled()))
5000 add_preempt_count(PREEMPT_ACTIVE);
5002 sub_preempt_count(PREEMPT_ACTIVE);
5005 * Check again in case we missed a preemption opportunity
5006 * between schedule and now.
5009 } while (need_resched());
5011 EXPORT_SYMBOL(preempt_schedule);
5014 * this is the entry point to schedule() from kernel preemption
5015 * off of irq context.
5016 * Note, that this is called and return with irqs disabled. This will
5017 * protect us against recursive calling from irq.
5019 asmlinkage void __sched preempt_schedule_irq(void)
5021 struct thread_info *ti = current_thread_info();
5023 /* Catch callers which need to be fixed */
5024 BUG_ON(ti->preempt_count || !irqs_disabled());
5027 add_preempt_count(PREEMPT_ACTIVE);
5030 local_irq_disable();
5031 sub_preempt_count(PREEMPT_ACTIVE);
5034 * Check again in case we missed a preemption opportunity
5035 * between schedule and now.
5038 } while (need_resched());
5041 #endif /* CONFIG_PREEMPT */
5043 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5046 return try_to_wake_up(curr->private, mode, sync);
5048 EXPORT_SYMBOL(default_wake_function);
5051 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5052 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5053 * number) then we wake all the non-exclusive tasks and one exclusive task.
5055 * There are circumstances in which we can try to wake a task which has already
5056 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5057 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5059 void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5060 int nr_exclusive, int sync, void *key)
5062 wait_queue_t *curr, *next;
5064 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5065 unsigned flags = curr->flags;
5067 if (curr->func(curr, mode, sync, key) &&
5068 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5074 * __wake_up - wake up threads blocked on a waitqueue.
5076 * @mode: which threads
5077 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5078 * @key: is directly passed to the wakeup function
5080 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5081 int nr_exclusive, void *key)
5083 unsigned long flags;
5085 spin_lock_irqsave(&q->lock, flags);
5086 __wake_up_common(q, mode, nr_exclusive, 0, key);
5087 spin_unlock_irqrestore(&q->lock, flags);
5089 EXPORT_SYMBOL(__wake_up);
5092 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5094 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5096 __wake_up_common(q, mode, 1, 0, NULL);
5100 * __wake_up_sync - wake up threads blocked on a waitqueue.
5102 * @mode: which threads
5103 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5105 * The sync wakeup differs that the waker knows that it will schedule
5106 * away soon, so while the target thread will be woken up, it will not
5107 * be migrated to another CPU - ie. the two threads are 'synchronized'
5108 * with each other. This can prevent needless bouncing between CPUs.
5110 * On UP it can prevent extra preemption.
5113 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5115 unsigned long flags;
5121 if (unlikely(!nr_exclusive))
5124 spin_lock_irqsave(&q->lock, flags);
5125 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
5126 spin_unlock_irqrestore(&q->lock, flags);
5128 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5131 * complete: - signals a single thread waiting on this completion
5132 * @x: holds the state of this particular completion
5134 * This will wake up a single thread waiting on this completion. Threads will be
5135 * awakened in the same order in which they were queued.
5137 * See also complete_all(), wait_for_completion() and related routines.
5139 void complete(struct completion *x)
5141 unsigned long flags;
5143 spin_lock_irqsave(&x->wait.lock, flags);
5145 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5146 spin_unlock_irqrestore(&x->wait.lock, flags);
5148 EXPORT_SYMBOL(complete);
5151 * complete_all: - signals all threads waiting on this completion
5152 * @x: holds the state of this particular completion
5154 * This will wake up all threads waiting on this particular completion event.
5156 void complete_all(struct completion *x)
5158 unsigned long flags;
5160 spin_lock_irqsave(&x->wait.lock, flags);
5161 x->done += UINT_MAX/2;
5162 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5163 spin_unlock_irqrestore(&x->wait.lock, flags);
5165 EXPORT_SYMBOL(complete_all);
5167 static inline long __sched
5168 do_wait_for_common(struct completion *x, long timeout, int state)
5171 DECLARE_WAITQUEUE(wait, current);
5173 wait.flags |= WQ_FLAG_EXCLUSIVE;
5174 __add_wait_queue_tail(&x->wait, &wait);
5176 if (signal_pending_state(state, current)) {
5177 timeout = -ERESTARTSYS;
5180 __set_current_state(state);
5181 spin_unlock_irq(&x->wait.lock);
5182 timeout = schedule_timeout(timeout);
5183 spin_lock_irq(&x->wait.lock);
5184 } while (!x->done && timeout);
5185 __remove_wait_queue(&x->wait, &wait);
5190 return timeout ?: 1;
5194 wait_for_common(struct completion *x, long timeout, int state)
5198 spin_lock_irq(&x->wait.lock);
5199 timeout = do_wait_for_common(x, timeout, state);
5200 spin_unlock_irq(&x->wait.lock);
5205 * wait_for_completion: - waits for completion of a task
5206 * @x: holds the state of this particular completion
5208 * This waits to be signaled for completion of a specific task. It is NOT
5209 * interruptible and there is no timeout.
5211 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5212 * and interrupt capability. Also see complete().
5214 void __sched wait_for_completion(struct completion *x)
5216 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5218 EXPORT_SYMBOL(wait_for_completion);
5221 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5222 * @x: holds the state of this particular completion
5223 * @timeout: timeout value in jiffies
5225 * This waits for either a completion of a specific task to be signaled or for a
5226 * specified timeout to expire. The timeout is in jiffies. It is not
5229 unsigned long __sched
5230 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5232 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5234 EXPORT_SYMBOL(wait_for_completion_timeout);
5237 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5238 * @x: holds the state of this particular completion
5240 * This waits for completion of a specific task to be signaled. It is
5243 int __sched wait_for_completion_interruptible(struct completion *x)
5245 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5246 if (t == -ERESTARTSYS)
5250 EXPORT_SYMBOL(wait_for_completion_interruptible);
5253 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5254 * @x: holds the state of this particular completion
5255 * @timeout: timeout value in jiffies
5257 * This waits for either a completion of a specific task to be signaled or for a
5258 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5260 unsigned long __sched
5261 wait_for_completion_interruptible_timeout(struct completion *x,
5262 unsigned long timeout)
5264 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5266 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5269 * wait_for_completion_killable: - waits for completion of a task (killable)
5270 * @x: holds the state of this particular completion
5272 * This waits to be signaled for completion of a specific task. It can be
5273 * interrupted by a kill signal.
5275 int __sched wait_for_completion_killable(struct completion *x)
5277 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5278 if (t == -ERESTARTSYS)
5282 EXPORT_SYMBOL(wait_for_completion_killable);
5285 * try_wait_for_completion - try to decrement a completion without blocking
5286 * @x: completion structure
5288 * Returns: 0 if a decrement cannot be done without blocking
5289 * 1 if a decrement succeeded.
5291 * If a completion is being used as a counting completion,
5292 * attempt to decrement the counter without blocking. This
5293 * enables us to avoid waiting if the resource the completion
5294 * is protecting is not available.
5296 bool try_wait_for_completion(struct completion *x)
5300 spin_lock_irq(&x->wait.lock);
5305 spin_unlock_irq(&x->wait.lock);
5308 EXPORT_SYMBOL(try_wait_for_completion);
5311 * completion_done - Test to see if a completion has any waiters
5312 * @x: completion structure
5314 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5315 * 1 if there are no waiters.
5318 bool completion_done(struct completion *x)
5322 spin_lock_irq(&x->wait.lock);
5325 spin_unlock_irq(&x->wait.lock);
5328 EXPORT_SYMBOL(completion_done);
5331 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5333 unsigned long flags;
5336 init_waitqueue_entry(&wait, current);
5338 __set_current_state(state);
5340 spin_lock_irqsave(&q->lock, flags);
5341 __add_wait_queue(q, &wait);
5342 spin_unlock(&q->lock);
5343 timeout = schedule_timeout(timeout);
5344 spin_lock_irq(&q->lock);
5345 __remove_wait_queue(q, &wait);
5346 spin_unlock_irqrestore(&q->lock, flags);
5351 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5353 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5355 EXPORT_SYMBOL(interruptible_sleep_on);
5358 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5360 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5362 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5364 void __sched sleep_on(wait_queue_head_t *q)
5366 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5368 EXPORT_SYMBOL(sleep_on);
5370 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5372 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5374 EXPORT_SYMBOL(sleep_on_timeout);
5376 #ifdef CONFIG_RT_MUTEXES
5379 * rt_mutex_setprio - set the current priority of a task
5381 * @prio: prio value (kernel-internal form)
5383 * This function changes the 'effective' priority of a task. It does
5384 * not touch ->normal_prio like __setscheduler().
5386 * Used by the rt_mutex code to implement priority inheritance logic.
5388 void rt_mutex_setprio(struct task_struct *p, int prio)
5390 unsigned long flags;
5391 int oldprio, on_rq, running;
5393 const struct sched_class *prev_class = p->sched_class;
5395 BUG_ON(prio < 0 || prio > MAX_PRIO);
5397 rq = task_rq_lock(p, &flags);
5398 update_rq_clock(rq);
5401 on_rq = p->se.on_rq;
5402 running = task_current(rq, p);
5404 dequeue_task(rq, p, 0);
5406 p->sched_class->put_prev_task(rq, p);
5409 p->sched_class = &rt_sched_class;
5411 p->sched_class = &fair_sched_class;
5416 p->sched_class->set_curr_task(rq);
5418 enqueue_task(rq, p, 0);
5420 check_class_changed(rq, p, prev_class, oldprio, running);
5422 task_rq_unlock(rq, &flags);
5427 void set_user_nice(struct task_struct *p, long nice)
5429 int old_prio, delta, on_rq;
5430 unsigned long flags;
5433 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5436 * We have to be careful, if called from sys_setpriority(),
5437 * the task might be in the middle of scheduling on another CPU.
5439 rq = task_rq_lock(p, &flags);
5440 update_rq_clock(rq);
5442 * The RT priorities are set via sched_setscheduler(), but we still
5443 * allow the 'normal' nice value to be set - but as expected
5444 * it wont have any effect on scheduling until the task is
5445 * SCHED_FIFO/SCHED_RR:
5447 if (task_has_rt_policy(p)) {
5448 p->static_prio = NICE_TO_PRIO(nice);
5451 on_rq = p->se.on_rq;
5453 dequeue_task(rq, p, 0);
5455 p->static_prio = NICE_TO_PRIO(nice);
5458 p->prio = effective_prio(p);
5459 delta = p->prio - old_prio;
5462 enqueue_task(rq, p, 0);
5464 * If the task increased its priority or is running and
5465 * lowered its priority, then reschedule its CPU:
5467 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5468 resched_task(rq->curr);
5471 task_rq_unlock(rq, &flags);
5473 EXPORT_SYMBOL(set_user_nice);
5476 * can_nice - check if a task can reduce its nice value
5480 int can_nice(const struct task_struct *p, const int nice)
5482 /* convert nice value [19,-20] to rlimit style value [1,40] */
5483 int nice_rlim = 20 - nice;
5485 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5486 capable(CAP_SYS_NICE));
5489 #ifdef __ARCH_WANT_SYS_NICE
5492 * sys_nice - change the priority of the current process.
5493 * @increment: priority increment
5495 * sys_setpriority is a more generic, but much slower function that
5496 * does similar things.
5498 SYSCALL_DEFINE1(nice, int, increment)
5503 * Setpriority might change our priority at the same moment.
5504 * We don't have to worry. Conceptually one call occurs first
5505 * and we have a single winner.
5507 if (increment < -40)
5512 nice = TASK_NICE(current) + increment;
5518 if (increment < 0 && !can_nice(current, nice))
5521 retval = security_task_setnice(current, nice);
5525 set_user_nice(current, nice);
5532 * task_prio - return the priority value of a given task.
5533 * @p: the task in question.
5535 * This is the priority value as seen by users in /proc.
5536 * RT tasks are offset by -200. Normal tasks are centered
5537 * around 0, value goes from -16 to +15.
5539 int task_prio(const struct task_struct *p)
5541 return p->prio - MAX_RT_PRIO;
5545 * task_nice - return the nice value of a given task.
5546 * @p: the task in question.
5548 int task_nice(const struct task_struct *p)
5550 return TASK_NICE(p);
5552 EXPORT_SYMBOL(task_nice);
5555 * idle_cpu - is a given cpu idle currently?
5556 * @cpu: the processor in question.
5558 int idle_cpu(int cpu)
5560 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5564 * idle_task - return the idle task for a given cpu.
5565 * @cpu: the processor in question.
5567 struct task_struct *idle_task(int cpu)
5569 return cpu_rq(cpu)->idle;
5573 * find_process_by_pid - find a process with a matching PID value.
5574 * @pid: the pid in question.
5576 static struct task_struct *find_process_by_pid(pid_t pid)
5578 return pid ? find_task_by_vpid(pid) : current;
5581 /* Actually do priority change: must hold rq lock. */
5583 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5585 BUG_ON(p->se.on_rq);
5588 switch (p->policy) {
5592 p->sched_class = &fair_sched_class;
5596 p->sched_class = &rt_sched_class;
5600 p->rt_priority = prio;
5601 p->normal_prio = normal_prio(p);
5602 /* we are holding p->pi_lock already */
5603 p->prio = rt_mutex_getprio(p);
5608 * check the target process has a UID that matches the current process's
5610 static bool check_same_owner(struct task_struct *p)
5612 const struct cred *cred = current_cred(), *pcred;
5616 pcred = __task_cred(p);
5617 match = (cred->euid == pcred->euid ||
5618 cred->euid == pcred->uid);
5623 static int __sched_setscheduler(struct task_struct *p, int policy,
5624 struct sched_param *param, bool user)
5626 int retval, oldprio, oldpolicy = -1, on_rq, running;
5627 unsigned long flags;
5628 const struct sched_class *prev_class = p->sched_class;
5631 /* may grab non-irq protected spin_locks */
5632 BUG_ON(in_interrupt());
5634 /* double check policy once rq lock held */
5636 policy = oldpolicy = p->policy;
5637 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5638 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5639 policy != SCHED_IDLE)
5642 * Valid priorities for SCHED_FIFO and SCHED_RR are
5643 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5644 * SCHED_BATCH and SCHED_IDLE is 0.
5646 if (param->sched_priority < 0 ||
5647 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5648 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5650 if (rt_policy(policy) != (param->sched_priority != 0))
5654 * Allow unprivileged RT tasks to decrease priority:
5656 if (user && !capable(CAP_SYS_NICE)) {
5657 if (rt_policy(policy)) {
5658 unsigned long rlim_rtprio;
5660 if (!lock_task_sighand(p, &flags))
5662 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5663 unlock_task_sighand(p, &flags);
5665 /* can't set/change the rt policy */
5666 if (policy != p->policy && !rlim_rtprio)
5669 /* can't increase priority */
5670 if (param->sched_priority > p->rt_priority &&
5671 param->sched_priority > rlim_rtprio)
5675 * Like positive nice levels, dont allow tasks to
5676 * move out of SCHED_IDLE either:
5678 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5681 /* can't change other user's priorities */
5682 if (!check_same_owner(p))
5687 #ifdef CONFIG_RT_GROUP_SCHED
5689 * Do not allow realtime tasks into groups that have no runtime
5692 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5693 task_group(p)->rt_bandwidth.rt_runtime == 0)
5697 retval = security_task_setscheduler(p, policy, param);
5703 * make sure no PI-waiters arrive (or leave) while we are
5704 * changing the priority of the task:
5706 spin_lock_irqsave(&p->pi_lock, flags);
5708 * To be able to change p->policy safely, the apropriate
5709 * runqueue lock must be held.
5711 rq = __task_rq_lock(p);
5712 /* recheck policy now with rq lock held */
5713 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5714 policy = oldpolicy = -1;
5715 __task_rq_unlock(rq);
5716 spin_unlock_irqrestore(&p->pi_lock, flags);
5719 update_rq_clock(rq);
5720 on_rq = p->se.on_rq;
5721 running = task_current(rq, p);
5723 deactivate_task(rq, p, 0);
5725 p->sched_class->put_prev_task(rq, p);
5728 __setscheduler(rq, p, policy, param->sched_priority);
5731 p->sched_class->set_curr_task(rq);
5733 activate_task(rq, p, 0);
5735 check_class_changed(rq, p, prev_class, oldprio, running);
5737 __task_rq_unlock(rq);
5738 spin_unlock_irqrestore(&p->pi_lock, flags);
5740 rt_mutex_adjust_pi(p);
5746 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5747 * @p: the task in question.
5748 * @policy: new policy.
5749 * @param: structure containing the new RT priority.
5751 * NOTE that the task may be already dead.
5753 int sched_setscheduler(struct task_struct *p, int policy,
5754 struct sched_param *param)
5756 return __sched_setscheduler(p, policy, param, true);
5758 EXPORT_SYMBOL_GPL(sched_setscheduler);
5761 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5762 * @p: the task in question.
5763 * @policy: new policy.
5764 * @param: structure containing the new RT priority.
5766 * Just like sched_setscheduler, only don't bother checking if the
5767 * current context has permission. For example, this is needed in
5768 * stop_machine(): we create temporary high priority worker threads,
5769 * but our caller might not have that capability.
5771 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5772 struct sched_param *param)
5774 return __sched_setscheduler(p, policy, param, false);
5778 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5780 struct sched_param lparam;
5781 struct task_struct *p;
5784 if (!param || pid < 0)
5786 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5791 p = find_process_by_pid(pid);
5793 retval = sched_setscheduler(p, policy, &lparam);
5800 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5801 * @pid: the pid in question.
5802 * @policy: new policy.
5803 * @param: structure containing the new RT priority.
5805 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5806 struct sched_param __user *, param)
5808 /* negative values for policy are not valid */
5812 return do_sched_setscheduler(pid, policy, param);
5816 * sys_sched_setparam - set/change the RT priority of a thread
5817 * @pid: the pid in question.
5818 * @param: structure containing the new RT priority.
5820 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5822 return do_sched_setscheduler(pid, -1, param);
5826 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5827 * @pid: the pid in question.
5829 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5831 struct task_struct *p;
5838 read_lock(&tasklist_lock);
5839 p = find_process_by_pid(pid);
5841 retval = security_task_getscheduler(p);
5845 read_unlock(&tasklist_lock);
5850 * sys_sched_getscheduler - get the RT priority of a thread
5851 * @pid: the pid in question.
5852 * @param: structure containing the RT priority.
5854 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5856 struct sched_param lp;
5857 struct task_struct *p;
5860 if (!param || pid < 0)
5863 read_lock(&tasklist_lock);
5864 p = find_process_by_pid(pid);
5869 retval = security_task_getscheduler(p);
5873 lp.sched_priority = p->rt_priority;
5874 read_unlock(&tasklist_lock);
5877 * This one might sleep, we cannot do it with a spinlock held ...
5879 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5884 read_unlock(&tasklist_lock);
5888 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5890 cpumask_var_t cpus_allowed, new_mask;
5891 struct task_struct *p;
5895 read_lock(&tasklist_lock);
5897 p = find_process_by_pid(pid);
5899 read_unlock(&tasklist_lock);
5905 * It is not safe to call set_cpus_allowed with the
5906 * tasklist_lock held. We will bump the task_struct's
5907 * usage count and then drop tasklist_lock.
5910 read_unlock(&tasklist_lock);
5912 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5916 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5918 goto out_free_cpus_allowed;
5921 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5924 retval = security_task_setscheduler(p, 0, NULL);
5928 cpuset_cpus_allowed(p, cpus_allowed);
5929 cpumask_and(new_mask, in_mask, cpus_allowed);
5931 retval = set_cpus_allowed_ptr(p, new_mask);
5934 cpuset_cpus_allowed(p, cpus_allowed);
5935 if (!cpumask_subset(new_mask, cpus_allowed)) {
5937 * We must have raced with a concurrent cpuset
5938 * update. Just reset the cpus_allowed to the
5939 * cpuset's cpus_allowed
5941 cpumask_copy(new_mask, cpus_allowed);
5946 free_cpumask_var(new_mask);
5947 out_free_cpus_allowed:
5948 free_cpumask_var(cpus_allowed);
5955 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5956 struct cpumask *new_mask)
5958 if (len < cpumask_size())
5959 cpumask_clear(new_mask);
5960 else if (len > cpumask_size())
5961 len = cpumask_size();
5963 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5967 * sys_sched_setaffinity - set the cpu affinity of a process
5968 * @pid: pid of the process
5969 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5970 * @user_mask_ptr: user-space pointer to the new cpu mask
5972 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5973 unsigned long __user *, user_mask_ptr)
5975 cpumask_var_t new_mask;
5978 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5981 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5983 retval = sched_setaffinity(pid, new_mask);
5984 free_cpumask_var(new_mask);
5988 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5990 struct task_struct *p;
5994 read_lock(&tasklist_lock);
5997 p = find_process_by_pid(pid);
6001 retval = security_task_getscheduler(p);
6005 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6008 read_unlock(&tasklist_lock);
6015 * sys_sched_getaffinity - get the cpu affinity of a process
6016 * @pid: pid of the process
6017 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6018 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6020 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6021 unsigned long __user *, user_mask_ptr)
6026 if (len < cpumask_size())
6029 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6032 ret = sched_getaffinity(pid, mask);
6034 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6037 ret = cpumask_size();
6039 free_cpumask_var(mask);
6045 * sys_sched_yield - yield the current processor to other threads.
6047 * This function yields the current CPU to other tasks. If there are no
6048 * other threads running on this CPU then this function will return.
6050 SYSCALL_DEFINE0(sched_yield)
6052 struct rq *rq = this_rq_lock();
6054 schedstat_inc(rq, yld_count);
6055 current->sched_class->yield_task(rq);
6058 * Since we are going to call schedule() anyway, there's
6059 * no need to preempt or enable interrupts:
6061 __release(rq->lock);
6062 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6063 _raw_spin_unlock(&rq->lock);
6064 preempt_enable_no_resched();
6071 static void __cond_resched(void)
6073 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6074 __might_sleep(__FILE__, __LINE__);
6077 * The BKS might be reacquired before we have dropped
6078 * PREEMPT_ACTIVE, which could trigger a second
6079 * cond_resched() call.
6082 add_preempt_count(PREEMPT_ACTIVE);
6084 sub_preempt_count(PREEMPT_ACTIVE);
6085 } while (need_resched());
6088 int __sched _cond_resched(void)
6090 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
6091 system_state == SYSTEM_RUNNING) {
6097 EXPORT_SYMBOL(_cond_resched);
6100 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6101 * call schedule, and on return reacquire the lock.
6103 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6104 * operations here to prevent schedule() from being called twice (once via
6105 * spin_unlock(), once by hand).
6107 int cond_resched_lock(spinlock_t *lock)
6109 int resched = need_resched() && system_state == SYSTEM_RUNNING;
6112 if (spin_needbreak(lock) || resched) {
6114 if (resched && need_resched())
6123 EXPORT_SYMBOL(cond_resched_lock);
6125 int __sched cond_resched_softirq(void)
6127 BUG_ON(!in_softirq());
6129 if (need_resched() && system_state == SYSTEM_RUNNING) {
6137 EXPORT_SYMBOL(cond_resched_softirq);
6140 * yield - yield the current processor to other threads.
6142 * This is a shortcut for kernel-space yielding - it marks the
6143 * thread runnable and calls sys_sched_yield().
6145 void __sched yield(void)
6147 set_current_state(TASK_RUNNING);
6150 EXPORT_SYMBOL(yield);
6153 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6154 * that process accounting knows that this is a task in IO wait state.
6156 * But don't do that if it is a deliberate, throttling IO wait (this task
6157 * has set its backing_dev_info: the queue against which it should throttle)
6159 void __sched io_schedule(void)
6161 struct rq *rq = &__raw_get_cpu_var(runqueues);
6163 delayacct_blkio_start();
6164 atomic_inc(&rq->nr_iowait);
6166 atomic_dec(&rq->nr_iowait);
6167 delayacct_blkio_end();
6169 EXPORT_SYMBOL(io_schedule);
6171 long __sched io_schedule_timeout(long timeout)
6173 struct rq *rq = &__raw_get_cpu_var(runqueues);
6176 delayacct_blkio_start();
6177 atomic_inc(&rq->nr_iowait);
6178 ret = schedule_timeout(timeout);
6179 atomic_dec(&rq->nr_iowait);
6180 delayacct_blkio_end();
6185 * sys_sched_get_priority_max - return maximum RT priority.
6186 * @policy: scheduling class.
6188 * this syscall returns the maximum rt_priority that can be used
6189 * by a given scheduling class.
6191 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6198 ret = MAX_USER_RT_PRIO-1;
6210 * sys_sched_get_priority_min - return minimum RT priority.
6211 * @policy: scheduling class.
6213 * this syscall returns the minimum rt_priority that can be used
6214 * by a given scheduling class.
6216 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6234 * sys_sched_rr_get_interval - return the default timeslice of a process.
6235 * @pid: pid of the process.
6236 * @interval: userspace pointer to the timeslice value.
6238 * this syscall writes the default timeslice value of a given process
6239 * into the user-space timespec buffer. A value of '0' means infinity.
6241 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6242 struct timespec __user *, interval)
6244 struct task_struct *p;
6245 unsigned int time_slice;
6253 read_lock(&tasklist_lock);
6254 p = find_process_by_pid(pid);
6258 retval = security_task_getscheduler(p);
6263 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6264 * tasks that are on an otherwise idle runqueue:
6267 if (p->policy == SCHED_RR) {
6268 time_slice = DEF_TIMESLICE;
6269 } else if (p->policy != SCHED_FIFO) {
6270 struct sched_entity *se = &p->se;
6271 unsigned long flags;
6274 rq = task_rq_lock(p, &flags);
6275 if (rq->cfs.load.weight)
6276 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6277 task_rq_unlock(rq, &flags);
6279 read_unlock(&tasklist_lock);
6280 jiffies_to_timespec(time_slice, &t);
6281 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6285 read_unlock(&tasklist_lock);
6289 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6291 void sched_show_task(struct task_struct *p)
6293 unsigned long free = 0;
6296 state = p->state ? __ffs(p->state) + 1 : 0;
6297 printk(KERN_INFO "%-13.13s %c", p->comm,
6298 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6299 #if BITS_PER_LONG == 32
6300 if (state == TASK_RUNNING)
6301 printk(KERN_CONT " running ");
6303 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6305 if (state == TASK_RUNNING)
6306 printk(KERN_CONT " running task ");
6308 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6310 #ifdef CONFIG_DEBUG_STACK_USAGE
6312 unsigned long *n = end_of_stack(p);
6315 free = (unsigned long)n - (unsigned long)end_of_stack(p);
6318 printk(KERN_CONT "%5lu %5d %6d\n", free,
6319 task_pid_nr(p), task_pid_nr(p->real_parent));
6321 show_stack(p, NULL);
6324 void show_state_filter(unsigned long state_filter)
6326 struct task_struct *g, *p;
6328 #if BITS_PER_LONG == 32
6330 " task PC stack pid father\n");
6333 " task PC stack pid father\n");
6335 read_lock(&tasklist_lock);
6336 do_each_thread(g, p) {
6338 * reset the NMI-timeout, listing all files on a slow
6339 * console might take alot of time:
6341 touch_nmi_watchdog();
6342 if (!state_filter || (p->state & state_filter))
6344 } while_each_thread(g, p);
6346 touch_all_softlockup_watchdogs();
6348 #ifdef CONFIG_SCHED_DEBUG
6349 sysrq_sched_debug_show();
6351 read_unlock(&tasklist_lock);
6353 * Only show locks if all tasks are dumped:
6355 if (state_filter == -1)
6356 debug_show_all_locks();
6359 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6361 idle->sched_class = &idle_sched_class;
6365 * init_idle - set up an idle thread for a given CPU
6366 * @idle: task in question
6367 * @cpu: cpu the idle task belongs to
6369 * NOTE: this function does not set the idle thread's NEED_RESCHED
6370 * flag, to make booting more robust.
6372 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6374 struct rq *rq = cpu_rq(cpu);
6375 unsigned long flags;
6377 spin_lock_irqsave(&rq->lock, flags);
6380 idle->se.exec_start = sched_clock();
6382 idle->prio = idle->normal_prio = MAX_PRIO;
6383 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6384 __set_task_cpu(idle, cpu);
6386 rq->curr = rq->idle = idle;
6387 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6390 spin_unlock_irqrestore(&rq->lock, flags);
6392 /* Set the preempt count _outside_ the spinlocks! */
6393 #if defined(CONFIG_PREEMPT)
6394 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6396 task_thread_info(idle)->preempt_count = 0;
6399 * The idle tasks have their own, simple scheduling class:
6401 idle->sched_class = &idle_sched_class;
6402 ftrace_graph_init_task(idle);
6406 * In a system that switches off the HZ timer nohz_cpu_mask
6407 * indicates which cpus entered this state. This is used
6408 * in the rcu update to wait only for active cpus. For system
6409 * which do not switch off the HZ timer nohz_cpu_mask should
6410 * always be CPU_BITS_NONE.
6412 cpumask_var_t nohz_cpu_mask;
6415 * Increase the granularity value when there are more CPUs,
6416 * because with more CPUs the 'effective latency' as visible
6417 * to users decreases. But the relationship is not linear,
6418 * so pick a second-best guess by going with the log2 of the
6421 * This idea comes from the SD scheduler of Con Kolivas:
6423 static inline void sched_init_granularity(void)
6425 unsigned int factor = 1 + ilog2(num_online_cpus());
6426 const unsigned long limit = 200000000;
6428 sysctl_sched_min_granularity *= factor;
6429 if (sysctl_sched_min_granularity > limit)
6430 sysctl_sched_min_granularity = limit;
6432 sysctl_sched_latency *= factor;
6433 if (sysctl_sched_latency > limit)
6434 sysctl_sched_latency = limit;
6436 sysctl_sched_wakeup_granularity *= factor;
6438 sysctl_sched_shares_ratelimit *= factor;
6443 * This is how migration works:
6445 * 1) we queue a struct migration_req structure in the source CPU's
6446 * runqueue and wake up that CPU's migration thread.
6447 * 2) we down() the locked semaphore => thread blocks.
6448 * 3) migration thread wakes up (implicitly it forces the migrated
6449 * thread off the CPU)
6450 * 4) it gets the migration request and checks whether the migrated
6451 * task is still in the wrong runqueue.
6452 * 5) if it's in the wrong runqueue then the migration thread removes
6453 * it and puts it into the right queue.
6454 * 6) migration thread up()s the semaphore.
6455 * 7) we wake up and the migration is done.
6459 * Change a given task's CPU affinity. Migrate the thread to a
6460 * proper CPU and schedule it away if the CPU it's executing on
6461 * is removed from the allowed bitmask.
6463 * NOTE: the caller must have a valid reference to the task, the
6464 * task must not exit() & deallocate itself prematurely. The
6465 * call is not atomic; no spinlocks may be held.
6467 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6469 struct migration_req req;
6470 unsigned long flags;
6474 rq = task_rq_lock(p, &flags);
6475 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6480 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6481 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6486 if (p->sched_class->set_cpus_allowed)
6487 p->sched_class->set_cpus_allowed(p, new_mask);
6489 cpumask_copy(&p->cpus_allowed, new_mask);
6490 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6493 /* Can the task run on the task's current CPU? If so, we're done */
6494 if (cpumask_test_cpu(task_cpu(p), new_mask))
6497 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6498 /* Need help from migration thread: drop lock and wait. */
6499 task_rq_unlock(rq, &flags);
6500 wake_up_process(rq->migration_thread);
6501 wait_for_completion(&req.done);
6502 tlb_migrate_finish(p->mm);
6506 task_rq_unlock(rq, &flags);
6510 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6513 * Move (not current) task off this cpu, onto dest cpu. We're doing
6514 * this because either it can't run here any more (set_cpus_allowed()
6515 * away from this CPU, or CPU going down), or because we're
6516 * attempting to rebalance this task on exec (sched_exec).
6518 * So we race with normal scheduler movements, but that's OK, as long
6519 * as the task is no longer on this CPU.
6521 * Returns non-zero if task was successfully migrated.
6523 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6525 struct rq *rq_dest, *rq_src;
6528 if (unlikely(!cpu_active(dest_cpu)))
6531 rq_src = cpu_rq(src_cpu);
6532 rq_dest = cpu_rq(dest_cpu);
6534 double_rq_lock(rq_src, rq_dest);
6535 /* Already moved. */
6536 if (task_cpu(p) != src_cpu)
6538 /* Affinity changed (again). */
6539 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6542 on_rq = p->se.on_rq;
6544 deactivate_task(rq_src, p, 0);
6546 set_task_cpu(p, dest_cpu);
6548 activate_task(rq_dest, p, 0);
6549 check_preempt_curr(rq_dest, p, 0);
6554 double_rq_unlock(rq_src, rq_dest);
6559 * migration_thread - this is a highprio system thread that performs
6560 * thread migration by bumping thread off CPU then 'pushing' onto
6563 static int migration_thread(void *data)
6565 int cpu = (long)data;
6569 BUG_ON(rq->migration_thread != current);
6571 set_current_state(TASK_INTERRUPTIBLE);
6572 while (!kthread_should_stop()) {
6573 struct migration_req *req;
6574 struct list_head *head;
6576 spin_lock_irq(&rq->lock);
6578 if (cpu_is_offline(cpu)) {
6579 spin_unlock_irq(&rq->lock);
6583 if (rq->active_balance) {
6584 active_load_balance(rq, cpu);
6585 rq->active_balance = 0;
6588 head = &rq->migration_queue;
6590 if (list_empty(head)) {
6591 spin_unlock_irq(&rq->lock);
6593 set_current_state(TASK_INTERRUPTIBLE);
6596 req = list_entry(head->next, struct migration_req, list);
6597 list_del_init(head->next);
6599 spin_unlock(&rq->lock);
6600 __migrate_task(req->task, cpu, req->dest_cpu);
6603 complete(&req->done);
6605 __set_current_state(TASK_RUNNING);
6609 /* Wait for kthread_stop */
6610 set_current_state(TASK_INTERRUPTIBLE);
6611 while (!kthread_should_stop()) {
6613 set_current_state(TASK_INTERRUPTIBLE);
6615 __set_current_state(TASK_RUNNING);
6619 #ifdef CONFIG_HOTPLUG_CPU
6621 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6625 local_irq_disable();
6626 ret = __migrate_task(p, src_cpu, dest_cpu);
6632 * Figure out where task on dead CPU should go, use force if necessary.
6634 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6637 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
6640 /* Look for allowed, online CPU in same node. */
6641 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
6642 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6645 /* Any allowed, online CPU? */
6646 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
6647 if (dest_cpu < nr_cpu_ids)
6650 /* No more Mr. Nice Guy. */
6651 if (dest_cpu >= nr_cpu_ids) {
6652 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
6653 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
6656 * Don't tell them about moving exiting tasks or
6657 * kernel threads (both mm NULL), since they never
6660 if (p->mm && printk_ratelimit()) {
6661 printk(KERN_INFO "process %d (%s) no "
6662 "longer affine to cpu%d\n",
6663 task_pid_nr(p), p->comm, dead_cpu);
6668 /* It can have affinity changed while we were choosing. */
6669 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
6674 * While a dead CPU has no uninterruptible tasks queued at this point,
6675 * it might still have a nonzero ->nr_uninterruptible counter, because
6676 * for performance reasons the counter is not stricly tracking tasks to
6677 * their home CPUs. So we just add the counter to another CPU's counter,
6678 * to keep the global sum constant after CPU-down:
6680 static void migrate_nr_uninterruptible(struct rq *rq_src)
6682 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
6683 unsigned long flags;
6685 local_irq_save(flags);
6686 double_rq_lock(rq_src, rq_dest);
6687 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6688 rq_src->nr_uninterruptible = 0;
6689 double_rq_unlock(rq_src, rq_dest);
6690 local_irq_restore(flags);
6693 /* Run through task list and migrate tasks from the dead cpu. */
6694 static void migrate_live_tasks(int src_cpu)
6696 struct task_struct *p, *t;
6698 read_lock(&tasklist_lock);
6700 do_each_thread(t, p) {
6704 if (task_cpu(p) == src_cpu)
6705 move_task_off_dead_cpu(src_cpu, p);
6706 } while_each_thread(t, p);
6708 read_unlock(&tasklist_lock);
6712 * Schedules idle task to be the next runnable task on current CPU.
6713 * It does so by boosting its priority to highest possible.
6714 * Used by CPU offline code.
6716 void sched_idle_next(void)
6718 int this_cpu = smp_processor_id();
6719 struct rq *rq = cpu_rq(this_cpu);
6720 struct task_struct *p = rq->idle;
6721 unsigned long flags;
6723 /* cpu has to be offline */
6724 BUG_ON(cpu_online(this_cpu));
6727 * Strictly not necessary since rest of the CPUs are stopped by now
6728 * and interrupts disabled on the current cpu.
6730 spin_lock_irqsave(&rq->lock, flags);
6732 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6734 update_rq_clock(rq);
6735 activate_task(rq, p, 0);
6737 spin_unlock_irqrestore(&rq->lock, flags);
6741 * Ensures that the idle task is using init_mm right before its cpu goes
6744 void idle_task_exit(void)
6746 struct mm_struct *mm = current->active_mm;
6748 BUG_ON(cpu_online(smp_processor_id()));
6751 switch_mm(mm, &init_mm, current);
6755 /* called under rq->lock with disabled interrupts */
6756 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6758 struct rq *rq = cpu_rq(dead_cpu);
6760 /* Must be exiting, otherwise would be on tasklist. */
6761 BUG_ON(!p->exit_state);
6763 /* Cannot have done final schedule yet: would have vanished. */
6764 BUG_ON(p->state == TASK_DEAD);
6769 * Drop lock around migration; if someone else moves it,
6770 * that's OK. No task can be added to this CPU, so iteration is
6773 spin_unlock_irq(&rq->lock);
6774 move_task_off_dead_cpu(dead_cpu, p);
6775 spin_lock_irq(&rq->lock);
6780 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6781 static void migrate_dead_tasks(unsigned int dead_cpu)
6783 struct rq *rq = cpu_rq(dead_cpu);
6784 struct task_struct *next;
6787 if (!rq->nr_running)
6789 update_rq_clock(rq);
6790 next = pick_next_task(rq);
6793 next->sched_class->put_prev_task(rq, next);
6794 migrate_dead(dead_cpu, next);
6798 #endif /* CONFIG_HOTPLUG_CPU */
6800 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6802 static struct ctl_table sd_ctl_dir[] = {
6804 .procname = "sched_domain",
6810 static struct ctl_table sd_ctl_root[] = {
6812 .ctl_name = CTL_KERN,
6813 .procname = "kernel",
6815 .child = sd_ctl_dir,
6820 static struct ctl_table *sd_alloc_ctl_entry(int n)
6822 struct ctl_table *entry =
6823 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6828 static void sd_free_ctl_entry(struct ctl_table **tablep)
6830 struct ctl_table *entry;
6833 * In the intermediate directories, both the child directory and
6834 * procname are dynamically allocated and could fail but the mode
6835 * will always be set. In the lowest directory the names are
6836 * static strings and all have proc handlers.
6838 for (entry = *tablep; entry->mode; entry++) {
6840 sd_free_ctl_entry(&entry->child);
6841 if (entry->proc_handler == NULL)
6842 kfree(entry->procname);
6850 set_table_entry(struct ctl_table *entry,
6851 const char *procname, void *data, int maxlen,
6852 mode_t mode, proc_handler *proc_handler)
6854 entry->procname = procname;
6856 entry->maxlen = maxlen;
6858 entry->proc_handler = proc_handler;
6861 static struct ctl_table *
6862 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6864 struct ctl_table *table = sd_alloc_ctl_entry(13);
6869 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6870 sizeof(long), 0644, proc_doulongvec_minmax);
6871 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6872 sizeof(long), 0644, proc_doulongvec_minmax);
6873 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6874 sizeof(int), 0644, proc_dointvec_minmax);
6875 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6876 sizeof(int), 0644, proc_dointvec_minmax);
6877 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6878 sizeof(int), 0644, proc_dointvec_minmax);
6879 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6880 sizeof(int), 0644, proc_dointvec_minmax);
6881 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6882 sizeof(int), 0644, proc_dointvec_minmax);
6883 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6884 sizeof(int), 0644, proc_dointvec_minmax);
6885 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6886 sizeof(int), 0644, proc_dointvec_minmax);
6887 set_table_entry(&table[9], "cache_nice_tries",
6888 &sd->cache_nice_tries,
6889 sizeof(int), 0644, proc_dointvec_minmax);
6890 set_table_entry(&table[10], "flags", &sd->flags,
6891 sizeof(int), 0644, proc_dointvec_minmax);
6892 set_table_entry(&table[11], "name", sd->name,
6893 CORENAME_MAX_SIZE, 0444, proc_dostring);
6894 /* &table[12] is terminator */
6899 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6901 struct ctl_table *entry, *table;
6902 struct sched_domain *sd;
6903 int domain_num = 0, i;
6906 for_each_domain(cpu, sd)
6908 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6913 for_each_domain(cpu, sd) {
6914 snprintf(buf, 32, "domain%d", i);
6915 entry->procname = kstrdup(buf, GFP_KERNEL);
6917 entry->child = sd_alloc_ctl_domain_table(sd);
6924 static struct ctl_table_header *sd_sysctl_header;
6925 static void register_sched_domain_sysctl(void)
6927 int i, cpu_num = num_online_cpus();
6928 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6931 WARN_ON(sd_ctl_dir[0].child);
6932 sd_ctl_dir[0].child = entry;
6937 for_each_online_cpu(i) {
6938 snprintf(buf, 32, "cpu%d", i);
6939 entry->procname = kstrdup(buf, GFP_KERNEL);
6941 entry->child = sd_alloc_ctl_cpu_table(i);
6945 WARN_ON(sd_sysctl_header);
6946 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6949 /* may be called multiple times per register */
6950 static void unregister_sched_domain_sysctl(void)
6952 if (sd_sysctl_header)
6953 unregister_sysctl_table(sd_sysctl_header);
6954 sd_sysctl_header = NULL;
6955 if (sd_ctl_dir[0].child)
6956 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6959 static void register_sched_domain_sysctl(void)
6962 static void unregister_sched_domain_sysctl(void)
6967 static void set_rq_online(struct rq *rq)
6970 const struct sched_class *class;
6972 cpumask_set_cpu(rq->cpu, rq->rd->online);
6975 for_each_class(class) {
6976 if (class->rq_online)
6977 class->rq_online(rq);
6982 static void set_rq_offline(struct rq *rq)
6985 const struct sched_class *class;
6987 for_each_class(class) {
6988 if (class->rq_offline)
6989 class->rq_offline(rq);
6992 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6998 * migration_call - callback that gets triggered when a CPU is added.
6999 * Here we can start up the necessary migration thread for the new CPU.
7001 static int __cpuinit
7002 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7004 struct task_struct *p;
7005 int cpu = (long)hcpu;
7006 unsigned long flags;
7011 case CPU_UP_PREPARE:
7012 case CPU_UP_PREPARE_FROZEN:
7013 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7016 kthread_bind(p, cpu);
7017 /* Must be high prio: stop_machine expects to yield to it. */
7018 rq = task_rq_lock(p, &flags);
7019 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7020 task_rq_unlock(rq, &flags);
7021 cpu_rq(cpu)->migration_thread = p;
7025 case CPU_ONLINE_FROZEN:
7026 /* Strictly unnecessary, as first user will wake it. */
7027 wake_up_process(cpu_rq(cpu)->migration_thread);
7029 /* Update our root-domain */
7031 spin_lock_irqsave(&rq->lock, flags);
7033 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7037 spin_unlock_irqrestore(&rq->lock, flags);
7040 #ifdef CONFIG_HOTPLUG_CPU
7041 case CPU_UP_CANCELED:
7042 case CPU_UP_CANCELED_FROZEN:
7043 if (!cpu_rq(cpu)->migration_thread)
7045 /* Unbind it from offline cpu so it can run. Fall thru. */
7046 kthread_bind(cpu_rq(cpu)->migration_thread,
7047 cpumask_any(cpu_online_mask));
7048 kthread_stop(cpu_rq(cpu)->migration_thread);
7049 cpu_rq(cpu)->migration_thread = NULL;
7053 case CPU_DEAD_FROZEN:
7054 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7055 migrate_live_tasks(cpu);
7057 kthread_stop(rq->migration_thread);
7058 rq->migration_thread = NULL;
7059 /* Idle task back to normal (off runqueue, low prio) */
7060 spin_lock_irq(&rq->lock);
7061 update_rq_clock(rq);
7062 deactivate_task(rq, rq->idle, 0);
7063 rq->idle->static_prio = MAX_PRIO;
7064 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7065 rq->idle->sched_class = &idle_sched_class;
7066 migrate_dead_tasks(cpu);
7067 spin_unlock_irq(&rq->lock);
7069 migrate_nr_uninterruptible(rq);
7070 BUG_ON(rq->nr_running != 0);
7073 * No need to migrate the tasks: it was best-effort if
7074 * they didn't take sched_hotcpu_mutex. Just wake up
7077 spin_lock_irq(&rq->lock);
7078 while (!list_empty(&rq->migration_queue)) {
7079 struct migration_req *req;
7081 req = list_entry(rq->migration_queue.next,
7082 struct migration_req, list);
7083 list_del_init(&req->list);
7084 spin_unlock_irq(&rq->lock);
7085 complete(&req->done);
7086 spin_lock_irq(&rq->lock);
7088 spin_unlock_irq(&rq->lock);
7092 case CPU_DYING_FROZEN:
7093 /* Update our root-domain */
7095 spin_lock_irqsave(&rq->lock, flags);
7097 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7100 spin_unlock_irqrestore(&rq->lock, flags);
7107 /* Register at highest priority so that task migration (migrate_all_tasks)
7108 * happens before everything else.
7110 static struct notifier_block __cpuinitdata migration_notifier = {
7111 .notifier_call = migration_call,
7115 static int __init migration_init(void)
7117 void *cpu = (void *)(long)smp_processor_id();
7120 /* Start one for the boot CPU: */
7121 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7122 BUG_ON(err == NOTIFY_BAD);
7123 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7124 register_cpu_notifier(&migration_notifier);
7128 early_initcall(migration_init);
7133 #ifdef CONFIG_SCHED_DEBUG
7135 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7136 struct cpumask *groupmask)
7138 struct sched_group *group = sd->groups;
7141 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7142 cpumask_clear(groupmask);
7144 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7146 if (!(sd->flags & SD_LOAD_BALANCE)) {
7147 printk("does not load-balance\n");
7149 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7154 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7156 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7157 printk(KERN_ERR "ERROR: domain->span does not contain "
7160 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7161 printk(KERN_ERR "ERROR: domain->groups does not contain"
7165 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7169 printk(KERN_ERR "ERROR: group is NULL\n");
7173 if (!group->__cpu_power) {
7174 printk(KERN_CONT "\n");
7175 printk(KERN_ERR "ERROR: domain->cpu_power not "
7180 if (!cpumask_weight(sched_group_cpus(group))) {
7181 printk(KERN_CONT "\n");
7182 printk(KERN_ERR "ERROR: empty group\n");
7186 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7187 printk(KERN_CONT "\n");
7188 printk(KERN_ERR "ERROR: repeated CPUs\n");
7192 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7194 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7195 printk(KERN_CONT " %s", str);
7197 group = group->next;
7198 } while (group != sd->groups);
7199 printk(KERN_CONT "\n");
7201 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7202 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7205 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7206 printk(KERN_ERR "ERROR: parent span is not a superset "
7207 "of domain->span\n");
7211 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7213 cpumask_var_t groupmask;
7217 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7221 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7223 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7224 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7229 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7236 free_cpumask_var(groupmask);
7238 #else /* !CONFIG_SCHED_DEBUG */
7239 # define sched_domain_debug(sd, cpu) do { } while (0)
7240 #endif /* CONFIG_SCHED_DEBUG */
7242 static int sd_degenerate(struct sched_domain *sd)
7244 if (cpumask_weight(sched_domain_span(sd)) == 1)
7247 /* Following flags need at least 2 groups */
7248 if (sd->flags & (SD_LOAD_BALANCE |
7249 SD_BALANCE_NEWIDLE |
7253 SD_SHARE_PKG_RESOURCES)) {
7254 if (sd->groups != sd->groups->next)
7258 /* Following flags don't use groups */
7259 if (sd->flags & (SD_WAKE_IDLE |
7268 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7270 unsigned long cflags = sd->flags, pflags = parent->flags;
7272 if (sd_degenerate(parent))
7275 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7278 /* Does parent contain flags not in child? */
7279 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7280 if (cflags & SD_WAKE_AFFINE)
7281 pflags &= ~SD_WAKE_BALANCE;
7282 /* Flags needing groups don't count if only 1 group in parent */
7283 if (parent->groups == parent->groups->next) {
7284 pflags &= ~(SD_LOAD_BALANCE |
7285 SD_BALANCE_NEWIDLE |
7289 SD_SHARE_PKG_RESOURCES);
7290 if (nr_node_ids == 1)
7291 pflags &= ~SD_SERIALIZE;
7293 if (~cflags & pflags)
7299 static void free_rootdomain(struct root_domain *rd)
7301 cpupri_cleanup(&rd->cpupri);
7303 free_cpumask_var(rd->rto_mask);
7304 free_cpumask_var(rd->online);
7305 free_cpumask_var(rd->span);
7309 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7311 struct root_domain *old_rd = NULL;
7312 unsigned long flags;
7314 spin_lock_irqsave(&rq->lock, flags);
7319 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7322 cpumask_clear_cpu(rq->cpu, old_rd->span);
7325 * If we dont want to free the old_rt yet then
7326 * set old_rd to NULL to skip the freeing later
7329 if (!atomic_dec_and_test(&old_rd->refcount))
7333 atomic_inc(&rd->refcount);
7336 cpumask_set_cpu(rq->cpu, rd->span);
7337 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7340 spin_unlock_irqrestore(&rq->lock, flags);
7343 free_rootdomain(old_rd);
7346 static int __init_refok init_rootdomain(struct root_domain *rd, bool bootmem)
7348 memset(rd, 0, sizeof(*rd));
7351 alloc_bootmem_cpumask_var(&def_root_domain.span);
7352 alloc_bootmem_cpumask_var(&def_root_domain.online);
7353 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
7354 cpupri_init(&rd->cpupri, true);
7358 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7360 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7362 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7365 if (cpupri_init(&rd->cpupri, false) != 0)
7370 free_cpumask_var(rd->rto_mask);
7372 free_cpumask_var(rd->online);
7374 free_cpumask_var(rd->span);
7379 static void init_defrootdomain(void)
7381 init_rootdomain(&def_root_domain, true);
7383 atomic_set(&def_root_domain.refcount, 1);
7386 static struct root_domain *alloc_rootdomain(void)
7388 struct root_domain *rd;
7390 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7394 if (init_rootdomain(rd, false) != 0) {
7403 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7404 * hold the hotplug lock.
7407 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7409 struct rq *rq = cpu_rq(cpu);
7410 struct sched_domain *tmp;
7412 /* Remove the sched domains which do not contribute to scheduling. */
7413 for (tmp = sd; tmp; ) {
7414 struct sched_domain *parent = tmp->parent;
7418 if (sd_parent_degenerate(tmp, parent)) {
7419 tmp->parent = parent->parent;
7421 parent->parent->child = tmp;
7426 if (sd && sd_degenerate(sd)) {
7432 sched_domain_debug(sd, cpu);
7434 rq_attach_root(rq, rd);
7435 rcu_assign_pointer(rq->sd, sd);
7438 /* cpus with isolated domains */
7439 static cpumask_var_t cpu_isolated_map;
7441 /* Setup the mask of cpus configured for isolated domains */
7442 static int __init isolated_cpu_setup(char *str)
7444 cpulist_parse(str, cpu_isolated_map);
7448 __setup("isolcpus=", isolated_cpu_setup);
7451 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7452 * to a function which identifies what group(along with sched group) a CPU
7453 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7454 * (due to the fact that we keep track of groups covered with a struct cpumask).
7456 * init_sched_build_groups will build a circular linked list of the groups
7457 * covered by the given span, and will set each group's ->cpumask correctly,
7458 * and ->cpu_power to 0.
7461 init_sched_build_groups(const struct cpumask *span,
7462 const struct cpumask *cpu_map,
7463 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7464 struct sched_group **sg,
7465 struct cpumask *tmpmask),
7466 struct cpumask *covered, struct cpumask *tmpmask)
7468 struct sched_group *first = NULL, *last = NULL;
7471 cpumask_clear(covered);
7473 for_each_cpu(i, span) {
7474 struct sched_group *sg;
7475 int group = group_fn(i, cpu_map, &sg, tmpmask);
7478 if (cpumask_test_cpu(i, covered))
7481 cpumask_clear(sched_group_cpus(sg));
7482 sg->__cpu_power = 0;
7484 for_each_cpu(j, span) {
7485 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7488 cpumask_set_cpu(j, covered);
7489 cpumask_set_cpu(j, sched_group_cpus(sg));
7500 #define SD_NODES_PER_DOMAIN 16
7505 * find_next_best_node - find the next node to include in a sched_domain
7506 * @node: node whose sched_domain we're building
7507 * @used_nodes: nodes already in the sched_domain
7509 * Find the next node to include in a given scheduling domain. Simply
7510 * finds the closest node not already in the @used_nodes map.
7512 * Should use nodemask_t.
7514 static int find_next_best_node(int node, nodemask_t *used_nodes)
7516 int i, n, val, min_val, best_node = 0;
7520 for (i = 0; i < nr_node_ids; i++) {
7521 /* Start at @node */
7522 n = (node + i) % nr_node_ids;
7524 if (!nr_cpus_node(n))
7527 /* Skip already used nodes */
7528 if (node_isset(n, *used_nodes))
7531 /* Simple min distance search */
7532 val = node_distance(node, n);
7534 if (val < min_val) {
7540 node_set(best_node, *used_nodes);
7545 * sched_domain_node_span - get a cpumask for a node's sched_domain
7546 * @node: node whose cpumask we're constructing
7547 * @span: resulting cpumask
7549 * Given a node, construct a good cpumask for its sched_domain to span. It
7550 * should be one that prevents unnecessary balancing, but also spreads tasks
7553 static void sched_domain_node_span(int node, struct cpumask *span)
7555 nodemask_t used_nodes;
7558 cpumask_clear(span);
7559 nodes_clear(used_nodes);
7561 cpumask_or(span, span, cpumask_of_node(node));
7562 node_set(node, used_nodes);
7564 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7565 int next_node = find_next_best_node(node, &used_nodes);
7567 cpumask_or(span, span, cpumask_of_node(next_node));
7570 #endif /* CONFIG_NUMA */
7572 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7575 * The cpus mask in sched_group and sched_domain hangs off the end.
7576 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7577 * for nr_cpu_ids < CONFIG_NR_CPUS.
7579 struct static_sched_group {
7580 struct sched_group sg;
7581 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7584 struct static_sched_domain {
7585 struct sched_domain sd;
7586 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7590 * SMT sched-domains:
7592 #ifdef CONFIG_SCHED_SMT
7593 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7594 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7597 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
7598 struct sched_group **sg, struct cpumask *unused)
7601 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7604 #endif /* CONFIG_SCHED_SMT */
7607 * multi-core sched-domains:
7609 #ifdef CONFIG_SCHED_MC
7610 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7611 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7612 #endif /* CONFIG_SCHED_MC */
7614 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7616 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7617 struct sched_group **sg, struct cpumask *mask)
7621 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7622 group = cpumask_first(mask);
7624 *sg = &per_cpu(sched_group_core, group).sg;
7627 #elif defined(CONFIG_SCHED_MC)
7629 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7630 struct sched_group **sg, struct cpumask *unused)
7633 *sg = &per_cpu(sched_group_core, cpu).sg;
7638 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
7639 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
7642 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
7643 struct sched_group **sg, struct cpumask *mask)
7646 #ifdef CONFIG_SCHED_MC
7647 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
7648 group = cpumask_first(mask);
7649 #elif defined(CONFIG_SCHED_SMT)
7650 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7651 group = cpumask_first(mask);
7656 *sg = &per_cpu(sched_group_phys, group).sg;
7662 * The init_sched_build_groups can't handle what we want to do with node
7663 * groups, so roll our own. Now each node has its own list of groups which
7664 * gets dynamically allocated.
7666 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
7667 static struct sched_group ***sched_group_nodes_bycpu;
7669 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
7670 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7672 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
7673 struct sched_group **sg,
7674 struct cpumask *nodemask)
7678 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7679 group = cpumask_first(nodemask);
7682 *sg = &per_cpu(sched_group_allnodes, group).sg;
7686 static void init_numa_sched_groups_power(struct sched_group *group_head)
7688 struct sched_group *sg = group_head;
7694 for_each_cpu(j, sched_group_cpus(sg)) {
7695 struct sched_domain *sd;
7697 sd = &per_cpu(phys_domains, j).sd;
7698 if (j != cpumask_first(sched_group_cpus(sd->groups))) {
7700 * Only add "power" once for each
7706 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7709 } while (sg != group_head);
7711 #endif /* CONFIG_NUMA */
7714 /* Free memory allocated for various sched_group structures */
7715 static void free_sched_groups(const struct cpumask *cpu_map,
7716 struct cpumask *nodemask)
7720 for_each_cpu(cpu, cpu_map) {
7721 struct sched_group **sched_group_nodes
7722 = sched_group_nodes_bycpu[cpu];
7724 if (!sched_group_nodes)
7727 for (i = 0; i < nr_node_ids; i++) {
7728 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7730 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7731 if (cpumask_empty(nodemask))
7741 if (oldsg != sched_group_nodes[i])
7744 kfree(sched_group_nodes);
7745 sched_group_nodes_bycpu[cpu] = NULL;
7748 #else /* !CONFIG_NUMA */
7749 static void free_sched_groups(const struct cpumask *cpu_map,
7750 struct cpumask *nodemask)
7753 #endif /* CONFIG_NUMA */
7756 * Initialize sched groups cpu_power.
7758 * cpu_power indicates the capacity of sched group, which is used while
7759 * distributing the load between different sched groups in a sched domain.
7760 * Typically cpu_power for all the groups in a sched domain will be same unless
7761 * there are asymmetries in the topology. If there are asymmetries, group
7762 * having more cpu_power will pickup more load compared to the group having
7765 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7766 * the maximum number of tasks a group can handle in the presence of other idle
7767 * or lightly loaded groups in the same sched domain.
7769 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7771 struct sched_domain *child;
7772 struct sched_group *group;
7774 WARN_ON(!sd || !sd->groups);
7776 if (cpu != cpumask_first(sched_group_cpus(sd->groups)))
7781 sd->groups->__cpu_power = 0;
7784 * For perf policy, if the groups in child domain share resources
7785 * (for example cores sharing some portions of the cache hierarchy
7786 * or SMT), then set this domain groups cpu_power such that each group
7787 * can handle only one task, when there are other idle groups in the
7788 * same sched domain.
7790 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7792 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7793 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7798 * add cpu_power of each child group to this groups cpu_power
7800 group = child->groups;
7802 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7803 group = group->next;
7804 } while (group != child->groups);
7808 * Initializers for schedule domains
7809 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7812 #ifdef CONFIG_SCHED_DEBUG
7813 # define SD_INIT_NAME(sd, type) sd->name = #type
7815 # define SD_INIT_NAME(sd, type) do { } while (0)
7818 #define SD_INIT(sd, type) sd_init_##type(sd)
7820 #define SD_INIT_FUNC(type) \
7821 static noinline void sd_init_##type(struct sched_domain *sd) \
7823 memset(sd, 0, sizeof(*sd)); \
7824 *sd = SD_##type##_INIT; \
7825 sd->level = SD_LV_##type; \
7826 SD_INIT_NAME(sd, type); \
7831 SD_INIT_FUNC(ALLNODES)
7834 #ifdef CONFIG_SCHED_SMT
7835 SD_INIT_FUNC(SIBLING)
7837 #ifdef CONFIG_SCHED_MC
7841 static int default_relax_domain_level = -1;
7843 static int __init setup_relax_domain_level(char *str)
7847 val = simple_strtoul(str, NULL, 0);
7848 if (val < SD_LV_MAX)
7849 default_relax_domain_level = val;
7853 __setup("relax_domain_level=", setup_relax_domain_level);
7855 static void set_domain_attribute(struct sched_domain *sd,
7856 struct sched_domain_attr *attr)
7860 if (!attr || attr->relax_domain_level < 0) {
7861 if (default_relax_domain_level < 0)
7864 request = default_relax_domain_level;
7866 request = attr->relax_domain_level;
7867 if (request < sd->level) {
7868 /* turn off idle balance on this domain */
7869 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7871 /* turn on idle balance on this domain */
7872 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7877 * Build sched domains for a given set of cpus and attach the sched domains
7878 * to the individual cpus
7880 static int __build_sched_domains(const struct cpumask *cpu_map,
7881 struct sched_domain_attr *attr)
7883 int i, err = -ENOMEM;
7884 struct root_domain *rd;
7885 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
7888 cpumask_var_t domainspan, covered, notcovered;
7889 struct sched_group **sched_group_nodes = NULL;
7890 int sd_allnodes = 0;
7892 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
7894 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
7895 goto free_domainspan;
7896 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
7900 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
7901 goto free_notcovered;
7902 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
7904 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
7905 goto free_this_sibling_map;
7906 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
7907 goto free_this_core_map;
7908 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
7909 goto free_send_covered;
7913 * Allocate the per-node list of sched groups
7915 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7917 if (!sched_group_nodes) {
7918 printk(KERN_WARNING "Can not alloc sched group node list\n");
7923 rd = alloc_rootdomain();
7925 printk(KERN_WARNING "Cannot alloc root domain\n");
7926 goto free_sched_groups;
7930 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
7934 * Set up domains for cpus specified by the cpu_map.
7936 for_each_cpu(i, cpu_map) {
7937 struct sched_domain *sd = NULL, *p;
7939 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
7942 if (cpumask_weight(cpu_map) >
7943 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
7944 sd = &per_cpu(allnodes_domains, i).sd;
7945 SD_INIT(sd, ALLNODES);
7946 set_domain_attribute(sd, attr);
7947 cpumask_copy(sched_domain_span(sd), cpu_map);
7948 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7954 sd = &per_cpu(node_domains, i).sd;
7956 set_domain_attribute(sd, attr);
7957 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7961 cpumask_and(sched_domain_span(sd),
7962 sched_domain_span(sd), cpu_map);
7966 sd = &per_cpu(phys_domains, i).sd;
7968 set_domain_attribute(sd, attr);
7969 cpumask_copy(sched_domain_span(sd), nodemask);
7973 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7975 #ifdef CONFIG_SCHED_MC
7977 sd = &per_cpu(core_domains, i).sd;
7979 set_domain_attribute(sd, attr);
7980 cpumask_and(sched_domain_span(sd), cpu_map,
7981 cpu_coregroup_mask(i));
7984 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7987 #ifdef CONFIG_SCHED_SMT
7989 sd = &per_cpu(cpu_domains, i).sd;
7990 SD_INIT(sd, SIBLING);
7991 set_domain_attribute(sd, attr);
7992 cpumask_and(sched_domain_span(sd),
7993 &per_cpu(cpu_sibling_map, i), cpu_map);
7996 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
8000 #ifdef CONFIG_SCHED_SMT
8001 /* Set up CPU (sibling) groups */
8002 for_each_cpu(i, cpu_map) {
8003 cpumask_and(this_sibling_map,
8004 &per_cpu(cpu_sibling_map, i), cpu_map);
8005 if (i != cpumask_first(this_sibling_map))
8008 init_sched_build_groups(this_sibling_map, cpu_map,
8010 send_covered, tmpmask);
8014 #ifdef CONFIG_SCHED_MC
8015 /* Set up multi-core groups */
8016 for_each_cpu(i, cpu_map) {
8017 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
8018 if (i != cpumask_first(this_core_map))
8021 init_sched_build_groups(this_core_map, cpu_map,
8023 send_covered, tmpmask);
8027 /* Set up physical groups */
8028 for (i = 0; i < nr_node_ids; i++) {
8029 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8030 if (cpumask_empty(nodemask))
8033 init_sched_build_groups(nodemask, cpu_map,
8035 send_covered, tmpmask);
8039 /* Set up node groups */
8041 init_sched_build_groups(cpu_map, cpu_map,
8042 &cpu_to_allnodes_group,
8043 send_covered, tmpmask);
8046 for (i = 0; i < nr_node_ids; i++) {
8047 /* Set up node groups */
8048 struct sched_group *sg, *prev;
8051 cpumask_clear(covered);
8052 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8053 if (cpumask_empty(nodemask)) {
8054 sched_group_nodes[i] = NULL;
8058 sched_domain_node_span(i, domainspan);
8059 cpumask_and(domainspan, domainspan, cpu_map);
8061 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8064 printk(KERN_WARNING "Can not alloc domain group for "
8068 sched_group_nodes[i] = sg;
8069 for_each_cpu(j, nodemask) {
8070 struct sched_domain *sd;
8072 sd = &per_cpu(node_domains, j).sd;
8075 sg->__cpu_power = 0;
8076 cpumask_copy(sched_group_cpus(sg), nodemask);
8078 cpumask_or(covered, covered, nodemask);
8081 for (j = 0; j < nr_node_ids; j++) {
8082 int n = (i + j) % nr_node_ids;
8084 cpumask_complement(notcovered, covered);
8085 cpumask_and(tmpmask, notcovered, cpu_map);
8086 cpumask_and(tmpmask, tmpmask, domainspan);
8087 if (cpumask_empty(tmpmask))
8090 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
8091 if (cpumask_empty(tmpmask))
8094 sg = kmalloc_node(sizeof(struct sched_group) +
8099 "Can not alloc domain group for node %d\n", j);
8102 sg->__cpu_power = 0;
8103 cpumask_copy(sched_group_cpus(sg), tmpmask);
8104 sg->next = prev->next;
8105 cpumask_or(covered, covered, tmpmask);
8112 /* Calculate CPU power for physical packages and nodes */
8113 #ifdef CONFIG_SCHED_SMT
8114 for_each_cpu(i, cpu_map) {
8115 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
8117 init_sched_groups_power(i, sd);
8120 #ifdef CONFIG_SCHED_MC
8121 for_each_cpu(i, cpu_map) {
8122 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
8124 init_sched_groups_power(i, sd);
8128 for_each_cpu(i, cpu_map) {
8129 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
8131 init_sched_groups_power(i, sd);
8135 for (i = 0; i < nr_node_ids; i++)
8136 init_numa_sched_groups_power(sched_group_nodes[i]);
8139 struct sched_group *sg;
8141 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8143 init_numa_sched_groups_power(sg);
8147 /* Attach the domains */
8148 for_each_cpu(i, cpu_map) {
8149 struct sched_domain *sd;
8150 #ifdef CONFIG_SCHED_SMT
8151 sd = &per_cpu(cpu_domains, i).sd;
8152 #elif defined(CONFIG_SCHED_MC)
8153 sd = &per_cpu(core_domains, i).sd;
8155 sd = &per_cpu(phys_domains, i).sd;
8157 cpu_attach_domain(sd, rd, i);
8163 free_cpumask_var(tmpmask);
8165 free_cpumask_var(send_covered);
8167 free_cpumask_var(this_core_map);
8168 free_this_sibling_map:
8169 free_cpumask_var(this_sibling_map);
8171 free_cpumask_var(nodemask);
8174 free_cpumask_var(notcovered);
8176 free_cpumask_var(covered);
8178 free_cpumask_var(domainspan);
8185 kfree(sched_group_nodes);
8191 free_sched_groups(cpu_map, tmpmask);
8192 free_rootdomain(rd);
8197 static int build_sched_domains(const struct cpumask *cpu_map)
8199 return __build_sched_domains(cpu_map, NULL);
8202 static struct cpumask *doms_cur; /* current sched domains */
8203 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8204 static struct sched_domain_attr *dattr_cur;
8205 /* attribues of custom domains in 'doms_cur' */
8208 * Special case: If a kmalloc of a doms_cur partition (array of
8209 * cpumask) fails, then fallback to a single sched domain,
8210 * as determined by the single cpumask fallback_doms.
8212 static cpumask_var_t fallback_doms;
8215 * arch_update_cpu_topology lets virtualized architectures update the
8216 * cpu core maps. It is supposed to return 1 if the topology changed
8217 * or 0 if it stayed the same.
8219 int __attribute__((weak)) arch_update_cpu_topology(void)
8225 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8226 * For now this just excludes isolated cpus, but could be used to
8227 * exclude other special cases in the future.
8229 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8233 arch_update_cpu_topology();
8235 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8237 doms_cur = fallback_doms;
8238 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8240 err = build_sched_domains(doms_cur);
8241 register_sched_domain_sysctl();
8246 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8247 struct cpumask *tmpmask)
8249 free_sched_groups(cpu_map, tmpmask);
8253 * Detach sched domains from a group of cpus specified in cpu_map
8254 * These cpus will now be attached to the NULL domain
8256 static void detach_destroy_domains(const struct cpumask *cpu_map)
8258 /* Save because hotplug lock held. */
8259 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8262 for_each_cpu(i, cpu_map)
8263 cpu_attach_domain(NULL, &def_root_domain, i);
8264 synchronize_sched();
8265 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8268 /* handle null as "default" */
8269 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8270 struct sched_domain_attr *new, int idx_new)
8272 struct sched_domain_attr tmp;
8279 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8280 new ? (new + idx_new) : &tmp,
8281 sizeof(struct sched_domain_attr));
8285 * Partition sched domains as specified by the 'ndoms_new'
8286 * cpumasks in the array doms_new[] of cpumasks. This compares
8287 * doms_new[] to the current sched domain partitioning, doms_cur[].
8288 * It destroys each deleted domain and builds each new domain.
8290 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8291 * The masks don't intersect (don't overlap.) We should setup one
8292 * sched domain for each mask. CPUs not in any of the cpumasks will
8293 * not be load balanced. If the same cpumask appears both in the
8294 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8297 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8298 * ownership of it and will kfree it when done with it. If the caller
8299 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8300 * ndoms_new == 1, and partition_sched_domains() will fallback to
8301 * the single partition 'fallback_doms', it also forces the domains
8304 * If doms_new == NULL it will be replaced with cpu_online_mask.
8305 * ndoms_new == 0 is a special case for destroying existing domains,
8306 * and it will not create the default domain.
8308 * Call with hotplug lock held
8310 /* FIXME: Change to struct cpumask *doms_new[] */
8311 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8312 struct sched_domain_attr *dattr_new)
8317 mutex_lock(&sched_domains_mutex);
8319 /* always unregister in case we don't destroy any domains */
8320 unregister_sched_domain_sysctl();
8322 /* Let architecture update cpu core mappings. */
8323 new_topology = arch_update_cpu_topology();
8325 n = doms_new ? ndoms_new : 0;
8327 /* Destroy deleted domains */
8328 for (i = 0; i < ndoms_cur; i++) {
8329 for (j = 0; j < n && !new_topology; j++) {
8330 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8331 && dattrs_equal(dattr_cur, i, dattr_new, j))
8334 /* no match - a current sched domain not in new doms_new[] */
8335 detach_destroy_domains(doms_cur + i);
8340 if (doms_new == NULL) {
8342 doms_new = fallback_doms;
8343 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8344 WARN_ON_ONCE(dattr_new);
8347 /* Build new domains */
8348 for (i = 0; i < ndoms_new; i++) {
8349 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8350 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8351 && dattrs_equal(dattr_new, i, dattr_cur, j))
8354 /* no match - add a new doms_new */
8355 __build_sched_domains(doms_new + i,
8356 dattr_new ? dattr_new + i : NULL);
8361 /* Remember the new sched domains */
8362 if (doms_cur != fallback_doms)
8364 kfree(dattr_cur); /* kfree(NULL) is safe */
8365 doms_cur = doms_new;
8366 dattr_cur = dattr_new;
8367 ndoms_cur = ndoms_new;
8369 register_sched_domain_sysctl();
8371 mutex_unlock(&sched_domains_mutex);
8374 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8375 static void arch_reinit_sched_domains(void)
8379 /* Destroy domains first to force the rebuild */
8380 partition_sched_domains(0, NULL, NULL);
8382 rebuild_sched_domains();
8386 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8388 unsigned int level = 0;
8390 if (sscanf(buf, "%u", &level) != 1)
8394 * level is always be positive so don't check for
8395 * level < POWERSAVINGS_BALANCE_NONE which is 0
8396 * What happens on 0 or 1 byte write,
8397 * need to check for count as well?
8400 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8404 sched_smt_power_savings = level;
8406 sched_mc_power_savings = level;
8408 arch_reinit_sched_domains();
8413 #ifdef CONFIG_SCHED_MC
8414 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8417 return sprintf(page, "%u\n", sched_mc_power_savings);
8419 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8420 const char *buf, size_t count)
8422 return sched_power_savings_store(buf, count, 0);
8424 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8425 sched_mc_power_savings_show,
8426 sched_mc_power_savings_store);
8429 #ifdef CONFIG_SCHED_SMT
8430 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8433 return sprintf(page, "%u\n", sched_smt_power_savings);
8435 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8436 const char *buf, size_t count)
8438 return sched_power_savings_store(buf, count, 1);
8440 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8441 sched_smt_power_savings_show,
8442 sched_smt_power_savings_store);
8445 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8449 #ifdef CONFIG_SCHED_SMT
8451 err = sysfs_create_file(&cls->kset.kobj,
8452 &attr_sched_smt_power_savings.attr);
8454 #ifdef CONFIG_SCHED_MC
8455 if (!err && mc_capable())
8456 err = sysfs_create_file(&cls->kset.kobj,
8457 &attr_sched_mc_power_savings.attr);
8461 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8463 #ifndef CONFIG_CPUSETS
8465 * Add online and remove offline CPUs from the scheduler domains.
8466 * When cpusets are enabled they take over this function.
8468 static int update_sched_domains(struct notifier_block *nfb,
8469 unsigned long action, void *hcpu)
8473 case CPU_ONLINE_FROZEN:
8475 case CPU_DEAD_FROZEN:
8476 partition_sched_domains(1, NULL, NULL);
8485 static int update_runtime(struct notifier_block *nfb,
8486 unsigned long action, void *hcpu)
8488 int cpu = (int)(long)hcpu;
8491 case CPU_DOWN_PREPARE:
8492 case CPU_DOWN_PREPARE_FROZEN:
8493 disable_runtime(cpu_rq(cpu));
8496 case CPU_DOWN_FAILED:
8497 case CPU_DOWN_FAILED_FROZEN:
8499 case CPU_ONLINE_FROZEN:
8500 enable_runtime(cpu_rq(cpu));
8508 void __init sched_init_smp(void)
8510 cpumask_var_t non_isolated_cpus;
8512 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8514 #if defined(CONFIG_NUMA)
8515 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8517 BUG_ON(sched_group_nodes_bycpu == NULL);
8520 mutex_lock(&sched_domains_mutex);
8521 arch_init_sched_domains(cpu_online_mask);
8522 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8523 if (cpumask_empty(non_isolated_cpus))
8524 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8525 mutex_unlock(&sched_domains_mutex);
8528 #ifndef CONFIG_CPUSETS
8529 /* XXX: Theoretical race here - CPU may be hotplugged now */
8530 hotcpu_notifier(update_sched_domains, 0);
8533 /* RT runtime code needs to handle some hotplug events */
8534 hotcpu_notifier(update_runtime, 0);
8538 /* Move init over to a non-isolated CPU */
8539 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8541 sched_init_granularity();
8542 free_cpumask_var(non_isolated_cpus);
8544 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8545 init_sched_rt_class();
8548 void __init sched_init_smp(void)
8550 sched_init_granularity();
8552 #endif /* CONFIG_SMP */
8554 int in_sched_functions(unsigned long addr)
8556 return in_lock_functions(addr) ||
8557 (addr >= (unsigned long)__sched_text_start
8558 && addr < (unsigned long)__sched_text_end);
8561 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8563 cfs_rq->tasks_timeline = RB_ROOT;
8564 INIT_LIST_HEAD(&cfs_rq->tasks);
8565 #ifdef CONFIG_FAIR_GROUP_SCHED
8568 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8571 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8573 struct rt_prio_array *array;
8576 array = &rt_rq->active;
8577 for (i = 0; i < MAX_RT_PRIO; i++) {
8578 INIT_LIST_HEAD(array->queue + i);
8579 __clear_bit(i, array->bitmap);
8581 /* delimiter for bitsearch: */
8582 __set_bit(MAX_RT_PRIO, array->bitmap);
8584 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8585 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8587 rt_rq->highest_prio.next = MAX_RT_PRIO;
8591 rt_rq->rt_nr_migratory = 0;
8592 rt_rq->overloaded = 0;
8593 plist_head_init(&rq->rt.pushable_tasks, &rq->lock);
8597 rt_rq->rt_throttled = 0;
8598 rt_rq->rt_runtime = 0;
8599 spin_lock_init(&rt_rq->rt_runtime_lock);
8601 #ifdef CONFIG_RT_GROUP_SCHED
8602 rt_rq->rt_nr_boosted = 0;
8607 #ifdef CONFIG_FAIR_GROUP_SCHED
8608 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8609 struct sched_entity *se, int cpu, int add,
8610 struct sched_entity *parent)
8612 struct rq *rq = cpu_rq(cpu);
8613 tg->cfs_rq[cpu] = cfs_rq;
8614 init_cfs_rq(cfs_rq, rq);
8617 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8620 /* se could be NULL for init_task_group */
8625 se->cfs_rq = &rq->cfs;
8627 se->cfs_rq = parent->my_q;
8630 se->load.weight = tg->shares;
8631 se->load.inv_weight = 0;
8632 se->parent = parent;
8636 #ifdef CONFIG_RT_GROUP_SCHED
8637 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8638 struct sched_rt_entity *rt_se, int cpu, int add,
8639 struct sched_rt_entity *parent)
8641 struct rq *rq = cpu_rq(cpu);
8643 tg->rt_rq[cpu] = rt_rq;
8644 init_rt_rq(rt_rq, rq);
8646 rt_rq->rt_se = rt_se;
8647 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8649 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8651 tg->rt_se[cpu] = rt_se;
8656 rt_se->rt_rq = &rq->rt;
8658 rt_se->rt_rq = parent->my_q;
8660 rt_se->my_q = rt_rq;
8661 rt_se->parent = parent;
8662 INIT_LIST_HEAD(&rt_se->run_list);
8666 void __init sched_init(void)
8669 unsigned long alloc_size = 0, ptr;
8671 #ifdef CONFIG_FAIR_GROUP_SCHED
8672 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8674 #ifdef CONFIG_RT_GROUP_SCHED
8675 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8677 #ifdef CONFIG_USER_SCHED
8681 * As sched_init() is called before page_alloc is setup,
8682 * we use alloc_bootmem().
8685 ptr = (unsigned long)alloc_bootmem(alloc_size);
8687 #ifdef CONFIG_FAIR_GROUP_SCHED
8688 init_task_group.se = (struct sched_entity **)ptr;
8689 ptr += nr_cpu_ids * sizeof(void **);
8691 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8692 ptr += nr_cpu_ids * sizeof(void **);
8694 #ifdef CONFIG_USER_SCHED
8695 root_task_group.se = (struct sched_entity **)ptr;
8696 ptr += nr_cpu_ids * sizeof(void **);
8698 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8699 ptr += nr_cpu_ids * sizeof(void **);
8700 #endif /* CONFIG_USER_SCHED */
8701 #endif /* CONFIG_FAIR_GROUP_SCHED */
8702 #ifdef CONFIG_RT_GROUP_SCHED
8703 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8704 ptr += nr_cpu_ids * sizeof(void **);
8706 init_task_group.rt_rq = (struct rt_rq **)ptr;
8707 ptr += nr_cpu_ids * sizeof(void **);
8709 #ifdef CONFIG_USER_SCHED
8710 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8711 ptr += nr_cpu_ids * sizeof(void **);
8713 root_task_group.rt_rq = (struct rt_rq **)ptr;
8714 ptr += nr_cpu_ids * sizeof(void **);
8715 #endif /* CONFIG_USER_SCHED */
8716 #endif /* CONFIG_RT_GROUP_SCHED */
8720 init_defrootdomain();
8723 init_rt_bandwidth(&def_rt_bandwidth,
8724 global_rt_period(), global_rt_runtime());
8726 #ifdef CONFIG_RT_GROUP_SCHED
8727 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8728 global_rt_period(), global_rt_runtime());
8729 #ifdef CONFIG_USER_SCHED
8730 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8731 global_rt_period(), RUNTIME_INF);
8732 #endif /* CONFIG_USER_SCHED */
8733 #endif /* CONFIG_RT_GROUP_SCHED */
8735 #ifdef CONFIG_GROUP_SCHED
8736 list_add(&init_task_group.list, &task_groups);
8737 INIT_LIST_HEAD(&init_task_group.children);
8739 #ifdef CONFIG_USER_SCHED
8740 INIT_LIST_HEAD(&root_task_group.children);
8741 init_task_group.parent = &root_task_group;
8742 list_add(&init_task_group.siblings, &root_task_group.children);
8743 #endif /* CONFIG_USER_SCHED */
8744 #endif /* CONFIG_GROUP_SCHED */
8746 for_each_possible_cpu(i) {
8750 spin_lock_init(&rq->lock);
8752 init_cfs_rq(&rq->cfs, rq);
8753 init_rt_rq(&rq->rt, rq);
8754 #ifdef CONFIG_FAIR_GROUP_SCHED
8755 init_task_group.shares = init_task_group_load;
8756 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8757 #ifdef CONFIG_CGROUP_SCHED
8759 * How much cpu bandwidth does init_task_group get?
8761 * In case of task-groups formed thr' the cgroup filesystem, it
8762 * gets 100% of the cpu resources in the system. This overall
8763 * system cpu resource is divided among the tasks of
8764 * init_task_group and its child task-groups in a fair manner,
8765 * based on each entity's (task or task-group's) weight
8766 * (se->load.weight).
8768 * In other words, if init_task_group has 10 tasks of weight
8769 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8770 * then A0's share of the cpu resource is:
8772 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8774 * We achieve this by letting init_task_group's tasks sit
8775 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8777 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8778 #elif defined CONFIG_USER_SCHED
8779 root_task_group.shares = NICE_0_LOAD;
8780 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8782 * In case of task-groups formed thr' the user id of tasks,
8783 * init_task_group represents tasks belonging to root user.
8784 * Hence it forms a sibling of all subsequent groups formed.
8785 * In this case, init_task_group gets only a fraction of overall
8786 * system cpu resource, based on the weight assigned to root
8787 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8788 * by letting tasks of init_task_group sit in a separate cfs_rq
8789 * (init_cfs_rq) and having one entity represent this group of
8790 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8792 init_tg_cfs_entry(&init_task_group,
8793 &per_cpu(init_cfs_rq, i),
8794 &per_cpu(init_sched_entity, i), i, 1,
8795 root_task_group.se[i]);
8798 #endif /* CONFIG_FAIR_GROUP_SCHED */
8800 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8801 #ifdef CONFIG_RT_GROUP_SCHED
8802 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8803 #ifdef CONFIG_CGROUP_SCHED
8804 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8805 #elif defined CONFIG_USER_SCHED
8806 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8807 init_tg_rt_entry(&init_task_group,
8808 &per_cpu(init_rt_rq, i),
8809 &per_cpu(init_sched_rt_entity, i), i, 1,
8810 root_task_group.rt_se[i]);
8814 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8815 rq->cpu_load[j] = 0;
8819 rq->active_balance = 0;
8820 rq->next_balance = jiffies;
8824 rq->migration_thread = NULL;
8825 INIT_LIST_HEAD(&rq->migration_queue);
8826 rq_attach_root(rq, &def_root_domain);
8829 atomic_set(&rq->nr_iowait, 0);
8832 set_load_weight(&init_task);
8834 #ifdef CONFIG_PREEMPT_NOTIFIERS
8835 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8839 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8842 #ifdef CONFIG_RT_MUTEXES
8843 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8847 * The boot idle thread does lazy MMU switching as well:
8849 atomic_inc(&init_mm.mm_count);
8850 enter_lazy_tlb(&init_mm, current);
8853 * Make us the idle thread. Technically, schedule() should not be
8854 * called from this thread, however somewhere below it might be,
8855 * but because we are the idle thread, we just pick up running again
8856 * when this runqueue becomes "idle".
8858 init_idle(current, smp_processor_id());
8860 * During early bootup we pretend to be a normal task:
8862 current->sched_class = &fair_sched_class;
8864 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8865 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
8868 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
8870 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8873 scheduler_running = 1;
8876 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8877 void __might_sleep(char *file, int line)
8880 static unsigned long prev_jiffy; /* ratelimiting */
8882 if ((!in_atomic() && !irqs_disabled()) ||
8883 system_state != SYSTEM_RUNNING || oops_in_progress)
8885 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8887 prev_jiffy = jiffies;
8890 "BUG: sleeping function called from invalid context at %s:%d\n",
8893 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8894 in_atomic(), irqs_disabled(),
8895 current->pid, current->comm);
8897 debug_show_held_locks(current);
8898 if (irqs_disabled())
8899 print_irqtrace_events(current);
8903 EXPORT_SYMBOL(__might_sleep);
8906 #ifdef CONFIG_MAGIC_SYSRQ
8907 static void normalize_task(struct rq *rq, struct task_struct *p)
8911 update_rq_clock(rq);
8912 on_rq = p->se.on_rq;
8914 deactivate_task(rq, p, 0);
8915 __setscheduler(rq, p, SCHED_NORMAL, 0);
8917 activate_task(rq, p, 0);
8918 resched_task(rq->curr);
8922 void normalize_rt_tasks(void)
8924 struct task_struct *g, *p;
8925 unsigned long flags;
8928 read_lock_irqsave(&tasklist_lock, flags);
8929 do_each_thread(g, p) {
8931 * Only normalize user tasks:
8936 p->se.exec_start = 0;
8937 #ifdef CONFIG_SCHEDSTATS
8938 p->se.wait_start = 0;
8939 p->se.sleep_start = 0;
8940 p->se.block_start = 0;
8945 * Renice negative nice level userspace
8948 if (TASK_NICE(p) < 0 && p->mm)
8949 set_user_nice(p, 0);
8953 spin_lock(&p->pi_lock);
8954 rq = __task_rq_lock(p);
8956 normalize_task(rq, p);
8958 __task_rq_unlock(rq);
8959 spin_unlock(&p->pi_lock);
8960 } while_each_thread(g, p);
8962 read_unlock_irqrestore(&tasklist_lock, flags);
8965 #endif /* CONFIG_MAGIC_SYSRQ */
8969 * These functions are only useful for the IA64 MCA handling.
8971 * They can only be called when the whole system has been
8972 * stopped - every CPU needs to be quiescent, and no scheduling
8973 * activity can take place. Using them for anything else would
8974 * be a serious bug, and as a result, they aren't even visible
8975 * under any other configuration.
8979 * curr_task - return the current task for a given cpu.
8980 * @cpu: the processor in question.
8982 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8984 struct task_struct *curr_task(int cpu)
8986 return cpu_curr(cpu);
8990 * set_curr_task - set the current task for a given cpu.
8991 * @cpu: the processor in question.
8992 * @p: the task pointer to set.
8994 * Description: This function must only be used when non-maskable interrupts
8995 * are serviced on a separate stack. It allows the architecture to switch the
8996 * notion of the current task on a cpu in a non-blocking manner. This function
8997 * must be called with all CPU's synchronized, and interrupts disabled, the
8998 * and caller must save the original value of the current task (see
8999 * curr_task() above) and restore that value before reenabling interrupts and
9000 * re-starting the system.
9002 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9004 void set_curr_task(int cpu, struct task_struct *p)
9011 #ifdef CONFIG_FAIR_GROUP_SCHED
9012 static void free_fair_sched_group(struct task_group *tg)
9016 for_each_possible_cpu(i) {
9018 kfree(tg->cfs_rq[i]);
9028 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9030 struct cfs_rq *cfs_rq;
9031 struct sched_entity *se;
9035 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9038 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9042 tg->shares = NICE_0_LOAD;
9044 for_each_possible_cpu(i) {
9047 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9048 GFP_KERNEL, cpu_to_node(i));
9052 se = kzalloc_node(sizeof(struct sched_entity),
9053 GFP_KERNEL, cpu_to_node(i));
9057 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9066 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9068 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9069 &cpu_rq(cpu)->leaf_cfs_rq_list);
9072 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9074 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9076 #else /* !CONFG_FAIR_GROUP_SCHED */
9077 static inline void free_fair_sched_group(struct task_group *tg)
9082 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9087 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9091 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9094 #endif /* CONFIG_FAIR_GROUP_SCHED */
9096 #ifdef CONFIG_RT_GROUP_SCHED
9097 static void free_rt_sched_group(struct task_group *tg)
9101 destroy_rt_bandwidth(&tg->rt_bandwidth);
9103 for_each_possible_cpu(i) {
9105 kfree(tg->rt_rq[i]);
9107 kfree(tg->rt_se[i]);
9115 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9117 struct rt_rq *rt_rq;
9118 struct sched_rt_entity *rt_se;
9122 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9125 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9129 init_rt_bandwidth(&tg->rt_bandwidth,
9130 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9132 for_each_possible_cpu(i) {
9135 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9136 GFP_KERNEL, cpu_to_node(i));
9140 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9141 GFP_KERNEL, cpu_to_node(i));
9145 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9154 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9156 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9157 &cpu_rq(cpu)->leaf_rt_rq_list);
9160 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9162 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9164 #else /* !CONFIG_RT_GROUP_SCHED */
9165 static inline void free_rt_sched_group(struct task_group *tg)
9170 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9175 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9179 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9182 #endif /* CONFIG_RT_GROUP_SCHED */
9184 #ifdef CONFIG_GROUP_SCHED
9185 static void free_sched_group(struct task_group *tg)
9187 free_fair_sched_group(tg);
9188 free_rt_sched_group(tg);
9192 /* allocate runqueue etc for a new task group */
9193 struct task_group *sched_create_group(struct task_group *parent)
9195 struct task_group *tg;
9196 unsigned long flags;
9199 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9201 return ERR_PTR(-ENOMEM);
9203 if (!alloc_fair_sched_group(tg, parent))
9206 if (!alloc_rt_sched_group(tg, parent))
9209 spin_lock_irqsave(&task_group_lock, flags);
9210 for_each_possible_cpu(i) {
9211 register_fair_sched_group(tg, i);
9212 register_rt_sched_group(tg, i);
9214 list_add_rcu(&tg->list, &task_groups);
9216 WARN_ON(!parent); /* root should already exist */
9218 tg->parent = parent;
9219 INIT_LIST_HEAD(&tg->children);
9220 list_add_rcu(&tg->siblings, &parent->children);
9221 spin_unlock_irqrestore(&task_group_lock, flags);
9226 free_sched_group(tg);
9227 return ERR_PTR(-ENOMEM);
9230 /* rcu callback to free various structures associated with a task group */
9231 static void free_sched_group_rcu(struct rcu_head *rhp)
9233 /* now it should be safe to free those cfs_rqs */
9234 free_sched_group(container_of(rhp, struct task_group, rcu));
9237 /* Destroy runqueue etc associated with a task group */
9238 void sched_destroy_group(struct task_group *tg)
9240 unsigned long flags;
9243 spin_lock_irqsave(&task_group_lock, flags);
9244 for_each_possible_cpu(i) {
9245 unregister_fair_sched_group(tg, i);
9246 unregister_rt_sched_group(tg, i);
9248 list_del_rcu(&tg->list);
9249 list_del_rcu(&tg->siblings);
9250 spin_unlock_irqrestore(&task_group_lock, flags);
9252 /* wait for possible concurrent references to cfs_rqs complete */
9253 call_rcu(&tg->rcu, free_sched_group_rcu);
9256 /* change task's runqueue when it moves between groups.
9257 * The caller of this function should have put the task in its new group
9258 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9259 * reflect its new group.
9261 void sched_move_task(struct task_struct *tsk)
9264 unsigned long flags;
9267 rq = task_rq_lock(tsk, &flags);
9269 update_rq_clock(rq);
9271 running = task_current(rq, tsk);
9272 on_rq = tsk->se.on_rq;
9275 dequeue_task(rq, tsk, 0);
9276 if (unlikely(running))
9277 tsk->sched_class->put_prev_task(rq, tsk);
9279 set_task_rq(tsk, task_cpu(tsk));
9281 #ifdef CONFIG_FAIR_GROUP_SCHED
9282 if (tsk->sched_class->moved_group)
9283 tsk->sched_class->moved_group(tsk);
9286 if (unlikely(running))
9287 tsk->sched_class->set_curr_task(rq);
9289 enqueue_task(rq, tsk, 0);
9291 task_rq_unlock(rq, &flags);
9293 #endif /* CONFIG_GROUP_SCHED */
9295 #ifdef CONFIG_FAIR_GROUP_SCHED
9296 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9298 struct cfs_rq *cfs_rq = se->cfs_rq;
9303 dequeue_entity(cfs_rq, se, 0);
9305 se->load.weight = shares;
9306 se->load.inv_weight = 0;
9309 enqueue_entity(cfs_rq, se, 0);
9312 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9314 struct cfs_rq *cfs_rq = se->cfs_rq;
9315 struct rq *rq = cfs_rq->rq;
9316 unsigned long flags;
9318 spin_lock_irqsave(&rq->lock, flags);
9319 __set_se_shares(se, shares);
9320 spin_unlock_irqrestore(&rq->lock, flags);
9323 static DEFINE_MUTEX(shares_mutex);
9325 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9328 unsigned long flags;
9331 * We can't change the weight of the root cgroup.
9336 if (shares < MIN_SHARES)
9337 shares = MIN_SHARES;
9338 else if (shares > MAX_SHARES)
9339 shares = MAX_SHARES;
9341 mutex_lock(&shares_mutex);
9342 if (tg->shares == shares)
9345 spin_lock_irqsave(&task_group_lock, flags);
9346 for_each_possible_cpu(i)
9347 unregister_fair_sched_group(tg, i);
9348 list_del_rcu(&tg->siblings);
9349 spin_unlock_irqrestore(&task_group_lock, flags);
9351 /* wait for any ongoing reference to this group to finish */
9352 synchronize_sched();
9355 * Now we are free to modify the group's share on each cpu
9356 * w/o tripping rebalance_share or load_balance_fair.
9358 tg->shares = shares;
9359 for_each_possible_cpu(i) {
9363 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9364 set_se_shares(tg->se[i], shares);
9368 * Enable load balance activity on this group, by inserting it back on
9369 * each cpu's rq->leaf_cfs_rq_list.
9371 spin_lock_irqsave(&task_group_lock, flags);
9372 for_each_possible_cpu(i)
9373 register_fair_sched_group(tg, i);
9374 list_add_rcu(&tg->siblings, &tg->parent->children);
9375 spin_unlock_irqrestore(&task_group_lock, flags);
9377 mutex_unlock(&shares_mutex);
9381 unsigned long sched_group_shares(struct task_group *tg)
9387 #ifdef CONFIG_RT_GROUP_SCHED
9389 * Ensure that the real time constraints are schedulable.
9391 static DEFINE_MUTEX(rt_constraints_mutex);
9393 static unsigned long to_ratio(u64 period, u64 runtime)
9395 if (runtime == RUNTIME_INF)
9398 return div64_u64(runtime << 20, period);
9401 /* Must be called with tasklist_lock held */
9402 static inline int tg_has_rt_tasks(struct task_group *tg)
9404 struct task_struct *g, *p;
9406 do_each_thread(g, p) {
9407 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9409 } while_each_thread(g, p);
9414 struct rt_schedulable_data {
9415 struct task_group *tg;
9420 static int tg_schedulable(struct task_group *tg, void *data)
9422 struct rt_schedulable_data *d = data;
9423 struct task_group *child;
9424 unsigned long total, sum = 0;
9425 u64 period, runtime;
9427 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9428 runtime = tg->rt_bandwidth.rt_runtime;
9431 period = d->rt_period;
9432 runtime = d->rt_runtime;
9435 #ifdef CONFIG_USER_SCHED
9436 if (tg == &root_task_group) {
9437 period = global_rt_period();
9438 runtime = global_rt_runtime();
9443 * Cannot have more runtime than the period.
9445 if (runtime > period && runtime != RUNTIME_INF)
9449 * Ensure we don't starve existing RT tasks.
9451 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9454 total = to_ratio(period, runtime);
9457 * Nobody can have more than the global setting allows.
9459 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9463 * The sum of our children's runtime should not exceed our own.
9465 list_for_each_entry_rcu(child, &tg->children, siblings) {
9466 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9467 runtime = child->rt_bandwidth.rt_runtime;
9469 if (child == d->tg) {
9470 period = d->rt_period;
9471 runtime = d->rt_runtime;
9474 sum += to_ratio(period, runtime);
9483 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9485 struct rt_schedulable_data data = {
9487 .rt_period = period,
9488 .rt_runtime = runtime,
9491 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9494 static int tg_set_bandwidth(struct task_group *tg,
9495 u64 rt_period, u64 rt_runtime)
9499 mutex_lock(&rt_constraints_mutex);
9500 read_lock(&tasklist_lock);
9501 err = __rt_schedulable(tg, rt_period, rt_runtime);
9505 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9506 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9507 tg->rt_bandwidth.rt_runtime = rt_runtime;
9509 for_each_possible_cpu(i) {
9510 struct rt_rq *rt_rq = tg->rt_rq[i];
9512 spin_lock(&rt_rq->rt_runtime_lock);
9513 rt_rq->rt_runtime = rt_runtime;
9514 spin_unlock(&rt_rq->rt_runtime_lock);
9516 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9518 read_unlock(&tasklist_lock);
9519 mutex_unlock(&rt_constraints_mutex);
9524 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9526 u64 rt_runtime, rt_period;
9528 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9529 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9530 if (rt_runtime_us < 0)
9531 rt_runtime = RUNTIME_INF;
9533 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9536 long sched_group_rt_runtime(struct task_group *tg)
9540 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9543 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9544 do_div(rt_runtime_us, NSEC_PER_USEC);
9545 return rt_runtime_us;
9548 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9550 u64 rt_runtime, rt_period;
9552 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9553 rt_runtime = tg->rt_bandwidth.rt_runtime;
9558 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9561 long sched_group_rt_period(struct task_group *tg)
9565 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9566 do_div(rt_period_us, NSEC_PER_USEC);
9567 return rt_period_us;
9570 static int sched_rt_global_constraints(void)
9572 u64 runtime, period;
9575 if (sysctl_sched_rt_period <= 0)
9578 runtime = global_rt_runtime();
9579 period = global_rt_period();
9582 * Sanity check on the sysctl variables.
9584 if (runtime > period && runtime != RUNTIME_INF)
9587 mutex_lock(&rt_constraints_mutex);
9588 read_lock(&tasklist_lock);
9589 ret = __rt_schedulable(NULL, 0, 0);
9590 read_unlock(&tasklist_lock);
9591 mutex_unlock(&rt_constraints_mutex);
9596 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9598 /* Don't accept realtime tasks when there is no way for them to run */
9599 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9605 #else /* !CONFIG_RT_GROUP_SCHED */
9606 static int sched_rt_global_constraints(void)
9608 unsigned long flags;
9611 if (sysctl_sched_rt_period <= 0)
9614 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9615 for_each_possible_cpu(i) {
9616 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9618 spin_lock(&rt_rq->rt_runtime_lock);
9619 rt_rq->rt_runtime = global_rt_runtime();
9620 spin_unlock(&rt_rq->rt_runtime_lock);
9622 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9626 #endif /* CONFIG_RT_GROUP_SCHED */
9628 int sched_rt_handler(struct ctl_table *table, int write,
9629 struct file *filp, void __user *buffer, size_t *lenp,
9633 int old_period, old_runtime;
9634 static DEFINE_MUTEX(mutex);
9637 old_period = sysctl_sched_rt_period;
9638 old_runtime = sysctl_sched_rt_runtime;
9640 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9642 if (!ret && write) {
9643 ret = sched_rt_global_constraints();
9645 sysctl_sched_rt_period = old_period;
9646 sysctl_sched_rt_runtime = old_runtime;
9648 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9649 def_rt_bandwidth.rt_period =
9650 ns_to_ktime(global_rt_period());
9653 mutex_unlock(&mutex);
9658 #ifdef CONFIG_CGROUP_SCHED
9660 /* return corresponding task_group object of a cgroup */
9661 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9663 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9664 struct task_group, css);
9667 static struct cgroup_subsys_state *
9668 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9670 struct task_group *tg, *parent;
9672 if (!cgrp->parent) {
9673 /* This is early initialization for the top cgroup */
9674 return &init_task_group.css;
9677 parent = cgroup_tg(cgrp->parent);
9678 tg = sched_create_group(parent);
9680 return ERR_PTR(-ENOMEM);
9686 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9688 struct task_group *tg = cgroup_tg(cgrp);
9690 sched_destroy_group(tg);
9694 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9695 struct task_struct *tsk)
9697 #ifdef CONFIG_RT_GROUP_SCHED
9698 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9701 /* We don't support RT-tasks being in separate groups */
9702 if (tsk->sched_class != &fair_sched_class)
9710 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9711 struct cgroup *old_cont, struct task_struct *tsk)
9713 sched_move_task(tsk);
9716 #ifdef CONFIG_FAIR_GROUP_SCHED
9717 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9720 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9723 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9725 struct task_group *tg = cgroup_tg(cgrp);
9727 return (u64) tg->shares;
9729 #endif /* CONFIG_FAIR_GROUP_SCHED */
9731 #ifdef CONFIG_RT_GROUP_SCHED
9732 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9735 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9738 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9740 return sched_group_rt_runtime(cgroup_tg(cgrp));
9743 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9746 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9749 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9751 return sched_group_rt_period(cgroup_tg(cgrp));
9753 #endif /* CONFIG_RT_GROUP_SCHED */
9755 static struct cftype cpu_files[] = {
9756 #ifdef CONFIG_FAIR_GROUP_SCHED
9759 .read_u64 = cpu_shares_read_u64,
9760 .write_u64 = cpu_shares_write_u64,
9763 #ifdef CONFIG_RT_GROUP_SCHED
9765 .name = "rt_runtime_us",
9766 .read_s64 = cpu_rt_runtime_read,
9767 .write_s64 = cpu_rt_runtime_write,
9770 .name = "rt_period_us",
9771 .read_u64 = cpu_rt_period_read_uint,
9772 .write_u64 = cpu_rt_period_write_uint,
9777 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9779 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9782 struct cgroup_subsys cpu_cgroup_subsys = {
9784 .create = cpu_cgroup_create,
9785 .destroy = cpu_cgroup_destroy,
9786 .can_attach = cpu_cgroup_can_attach,
9787 .attach = cpu_cgroup_attach,
9788 .populate = cpu_cgroup_populate,
9789 .subsys_id = cpu_cgroup_subsys_id,
9793 #endif /* CONFIG_CGROUP_SCHED */
9795 #ifdef CONFIG_CGROUP_CPUACCT
9798 * CPU accounting code for task groups.
9800 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9801 * (balbir@in.ibm.com).
9804 /* track cpu usage of a group of tasks and its child groups */
9806 struct cgroup_subsys_state css;
9807 /* cpuusage holds pointer to a u64-type object on every cpu */
9809 struct cpuacct *parent;
9812 struct cgroup_subsys cpuacct_subsys;
9814 /* return cpu accounting group corresponding to this container */
9815 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9817 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9818 struct cpuacct, css);
9821 /* return cpu accounting group to which this task belongs */
9822 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9824 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9825 struct cpuacct, css);
9828 /* create a new cpu accounting group */
9829 static struct cgroup_subsys_state *cpuacct_create(
9830 struct cgroup_subsys *ss, struct cgroup *cgrp)
9832 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9835 return ERR_PTR(-ENOMEM);
9837 ca->cpuusage = alloc_percpu(u64);
9838 if (!ca->cpuusage) {
9840 return ERR_PTR(-ENOMEM);
9844 ca->parent = cgroup_ca(cgrp->parent);
9849 /* destroy an existing cpu accounting group */
9851 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9853 struct cpuacct *ca = cgroup_ca(cgrp);
9855 free_percpu(ca->cpuusage);
9859 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9861 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9864 #ifndef CONFIG_64BIT
9866 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9868 spin_lock_irq(&cpu_rq(cpu)->lock);
9870 spin_unlock_irq(&cpu_rq(cpu)->lock);
9878 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9880 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9882 #ifndef CONFIG_64BIT
9884 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9886 spin_lock_irq(&cpu_rq(cpu)->lock);
9888 spin_unlock_irq(&cpu_rq(cpu)->lock);
9894 /* return total cpu usage (in nanoseconds) of a group */
9895 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9897 struct cpuacct *ca = cgroup_ca(cgrp);
9898 u64 totalcpuusage = 0;
9901 for_each_present_cpu(i)
9902 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9904 return totalcpuusage;
9907 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9910 struct cpuacct *ca = cgroup_ca(cgrp);
9919 for_each_present_cpu(i)
9920 cpuacct_cpuusage_write(ca, i, 0);
9926 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9929 struct cpuacct *ca = cgroup_ca(cgroup);
9933 for_each_present_cpu(i) {
9934 percpu = cpuacct_cpuusage_read(ca, i);
9935 seq_printf(m, "%llu ", (unsigned long long) percpu);
9937 seq_printf(m, "\n");
9941 static struct cftype files[] = {
9944 .read_u64 = cpuusage_read,
9945 .write_u64 = cpuusage_write,
9948 .name = "usage_percpu",
9949 .read_seq_string = cpuacct_percpu_seq_read,
9954 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9956 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9960 * charge this task's execution time to its accounting group.
9962 * called with rq->lock held.
9964 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9969 if (unlikely(!cpuacct_subsys.active))
9972 cpu = task_cpu(tsk);
9975 for (; ca; ca = ca->parent) {
9976 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9977 *cpuusage += cputime;
9981 struct cgroup_subsys cpuacct_subsys = {
9983 .create = cpuacct_create,
9984 .destroy = cpuacct_destroy,
9985 .populate = cpuacct_populate,
9986 .subsys_id = cpuacct_subsys_id,
9988 #endif /* CONFIG_CGROUP_CPUACCT */