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))
3194 * find_busiest_group finds and returns the busiest CPU group within the
3195 * domain. It calculates and returns the amount of weighted load which
3196 * should be moved to restore balance via the imbalance parameter.
3198 static struct sched_group *
3199 find_busiest_group(struct sched_domain *sd, int this_cpu,
3200 unsigned long *imbalance, enum cpu_idle_type idle,
3201 int *sd_idle, const struct cpumask *cpus, int *balance)
3203 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3204 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3205 unsigned long max_pull;
3206 unsigned long busiest_load_per_task, busiest_nr_running;
3207 unsigned long this_load_per_task, this_nr_running;
3208 int load_idx, group_imb = 0;
3209 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3210 int power_savings_balance = 1;
3211 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3212 unsigned long min_nr_running = ULONG_MAX;
3213 struct sched_group *group_min = NULL, *group_leader = NULL;
3216 max_load = this_load = total_load = total_pwr = 0;
3217 busiest_load_per_task = busiest_nr_running = 0;
3218 this_load_per_task = this_nr_running = 0;
3220 if (idle == CPU_NOT_IDLE)
3221 load_idx = sd->busy_idx;
3222 else if (idle == CPU_NEWLY_IDLE)
3223 load_idx = sd->newidle_idx;
3225 load_idx = sd->idle_idx;
3228 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3231 int __group_imb = 0;
3232 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3233 unsigned long sum_nr_running, sum_weighted_load;
3234 unsigned long sum_avg_load_per_task;
3235 unsigned long avg_load_per_task;
3237 local_group = cpumask_test_cpu(this_cpu,
3238 sched_group_cpus(group));
3241 balance_cpu = cpumask_first(sched_group_cpus(group));
3243 /* Tally up the load of all CPUs in the group */
3244 sum_weighted_load = sum_nr_running = avg_load = 0;
3245 sum_avg_load_per_task = avg_load_per_task = 0;
3248 min_cpu_load = ~0UL;
3250 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3251 struct rq *rq = cpu_rq(i);
3253 if (*sd_idle && rq->nr_running)
3256 /* Bias balancing toward cpus of our domain */
3258 if (idle_cpu(i) && !first_idle_cpu) {
3263 load = target_load(i, load_idx);
3265 load = source_load(i, load_idx);
3266 if (load > max_cpu_load)
3267 max_cpu_load = load;
3268 if (min_cpu_load > load)
3269 min_cpu_load = load;
3273 sum_nr_running += rq->nr_running;
3274 sum_weighted_load += weighted_cpuload(i);
3276 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3280 * First idle cpu or the first cpu(busiest) in this sched group
3281 * is eligible for doing load balancing at this and above
3282 * domains. In the newly idle case, we will allow all the cpu's
3283 * to do the newly idle load balance.
3285 if (idle != CPU_NEWLY_IDLE && local_group &&
3286 balance_cpu != this_cpu && balance) {
3291 total_load += avg_load;
3292 total_pwr += group->__cpu_power;
3294 /* Adjust by relative CPU power of the group */
3295 avg_load = sg_div_cpu_power(group,
3296 avg_load * SCHED_LOAD_SCALE);
3300 * Consider the group unbalanced when the imbalance is larger
3301 * than the average weight of two tasks.
3303 * APZ: with cgroup the avg task weight can vary wildly and
3304 * might not be a suitable number - should we keep a
3305 * normalized nr_running number somewhere that negates
3308 avg_load_per_task = sg_div_cpu_power(group,
3309 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3311 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3314 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3317 this_load = avg_load;
3319 this_nr_running = sum_nr_running;
3320 this_load_per_task = sum_weighted_load;
3321 } else if (avg_load > max_load &&
3322 (sum_nr_running > group_capacity || __group_imb)) {
3323 max_load = avg_load;
3325 busiest_nr_running = sum_nr_running;
3326 busiest_load_per_task = sum_weighted_load;
3327 group_imb = __group_imb;
3330 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3332 * Busy processors will not participate in power savings
3335 if (idle == CPU_NOT_IDLE ||
3336 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3340 * If the local group is idle or completely loaded
3341 * no need to do power savings balance at this domain
3343 if (local_group && (this_nr_running >= group_capacity ||
3345 power_savings_balance = 0;
3348 * If a group is already running at full capacity or idle,
3349 * don't include that group in power savings calculations
3351 if (!power_savings_balance || sum_nr_running >= group_capacity
3356 * Calculate the group which has the least non-idle load.
3357 * This is the group from where we need to pick up the load
3360 if ((sum_nr_running < min_nr_running) ||
3361 (sum_nr_running == min_nr_running &&
3362 cpumask_first(sched_group_cpus(group)) >
3363 cpumask_first(sched_group_cpus(group_min)))) {
3365 min_nr_running = sum_nr_running;
3366 min_load_per_task = sum_weighted_load /
3371 * Calculate the group which is almost near its
3372 * capacity but still has some space to pick up some load
3373 * from other group and save more power
3375 if (sum_nr_running <= group_capacity - 1) {
3376 if (sum_nr_running > leader_nr_running ||
3377 (sum_nr_running == leader_nr_running &&
3378 cpumask_first(sched_group_cpus(group)) <
3379 cpumask_first(sched_group_cpus(group_leader)))) {
3380 group_leader = group;
3381 leader_nr_running = sum_nr_running;
3386 group = group->next;
3387 } while (group != sd->groups);
3389 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3392 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3394 if (this_load >= avg_load ||
3395 100*max_load <= sd->imbalance_pct*this_load)
3398 busiest_load_per_task /= busiest_nr_running;
3400 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3403 * We're trying to get all the cpus to the average_load, so we don't
3404 * want to push ourselves above the average load, nor do we wish to
3405 * reduce the max loaded cpu below the average load, as either of these
3406 * actions would just result in more rebalancing later, and ping-pong
3407 * tasks around. Thus we look for the minimum possible imbalance.
3408 * Negative imbalances (*we* are more loaded than anyone else) will
3409 * be counted as no imbalance for these purposes -- we can't fix that
3410 * by pulling tasks to us. Be careful of negative numbers as they'll
3411 * appear as very large values with unsigned longs.
3413 if (max_load <= busiest_load_per_task)
3417 * In the presence of smp nice balancing, certain scenarios can have
3418 * max load less than avg load(as we skip the groups at or below
3419 * its cpu_power, while calculating max_load..)
3421 if (max_load < avg_load) {
3423 goto small_imbalance;
3426 /* Don't want to pull so many tasks that a group would go idle */
3427 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3429 /* How much load to actually move to equalise the imbalance */
3430 *imbalance = min(max_pull * busiest->__cpu_power,
3431 (avg_load - this_load) * this->__cpu_power)
3435 * if *imbalance is less than the average load per runnable task
3436 * there is no gaurantee that any tasks will be moved so we'll have
3437 * a think about bumping its value to force at least one task to be
3440 if (*imbalance < busiest_load_per_task) {
3441 unsigned long tmp, pwr_now, pwr_move;
3445 pwr_move = pwr_now = 0;
3447 if (this_nr_running) {
3448 this_load_per_task /= this_nr_running;
3449 if (busiest_load_per_task > this_load_per_task)
3452 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3454 if (max_load - this_load + busiest_load_per_task >=
3455 busiest_load_per_task * imbn) {
3456 *imbalance = busiest_load_per_task;
3461 * OK, we don't have enough imbalance to justify moving tasks,
3462 * however we may be able to increase total CPU power used by
3466 pwr_now += busiest->__cpu_power *
3467 min(busiest_load_per_task, max_load);
3468 pwr_now += this->__cpu_power *
3469 min(this_load_per_task, this_load);
3470 pwr_now /= SCHED_LOAD_SCALE;
3472 /* Amount of load we'd subtract */
3473 tmp = sg_div_cpu_power(busiest,
3474 busiest_load_per_task * SCHED_LOAD_SCALE);
3476 pwr_move += busiest->__cpu_power *
3477 min(busiest_load_per_task, max_load - tmp);
3479 /* Amount of load we'd add */
3480 if (max_load * busiest->__cpu_power <
3481 busiest_load_per_task * SCHED_LOAD_SCALE)
3482 tmp = sg_div_cpu_power(this,
3483 max_load * busiest->__cpu_power);
3485 tmp = sg_div_cpu_power(this,
3486 busiest_load_per_task * SCHED_LOAD_SCALE);
3487 pwr_move += this->__cpu_power *
3488 min(this_load_per_task, this_load + tmp);
3489 pwr_move /= SCHED_LOAD_SCALE;
3491 /* Move if we gain throughput */
3492 if (pwr_move > pwr_now)
3493 *imbalance = busiest_load_per_task;
3499 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3500 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3503 if (this == group_leader && group_leader != group_min) {
3504 *imbalance = min_load_per_task;
3505 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3506 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3507 cpumask_first(sched_group_cpus(group_leader));
3518 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3521 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3522 unsigned long imbalance, const struct cpumask *cpus)
3524 struct rq *busiest = NULL, *rq;
3525 unsigned long max_load = 0;
3528 for_each_cpu(i, sched_group_cpus(group)) {
3531 if (!cpumask_test_cpu(i, cpus))
3535 wl = weighted_cpuload(i);
3537 if (rq->nr_running == 1 && wl > imbalance)
3540 if (wl > max_load) {
3550 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3551 * so long as it is large enough.
3553 #define MAX_PINNED_INTERVAL 512
3556 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3557 * tasks if there is an imbalance.
3559 static int load_balance(int this_cpu, struct rq *this_rq,
3560 struct sched_domain *sd, enum cpu_idle_type idle,
3561 int *balance, struct cpumask *cpus)
3563 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3564 struct sched_group *group;
3565 unsigned long imbalance;
3567 unsigned long flags;
3569 cpumask_setall(cpus);
3572 * When power savings policy is enabled for the parent domain, idle
3573 * sibling can pick up load irrespective of busy siblings. In this case,
3574 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3575 * portraying it as CPU_NOT_IDLE.
3577 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3578 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3581 schedstat_inc(sd, lb_count[idle]);
3585 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3592 schedstat_inc(sd, lb_nobusyg[idle]);
3596 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3598 schedstat_inc(sd, lb_nobusyq[idle]);
3602 BUG_ON(busiest == this_rq);
3604 schedstat_add(sd, lb_imbalance[idle], imbalance);
3607 if (busiest->nr_running > 1) {
3609 * Attempt to move tasks. If find_busiest_group has found
3610 * an imbalance but busiest->nr_running <= 1, the group is
3611 * still unbalanced. ld_moved simply stays zero, so it is
3612 * correctly treated as an imbalance.
3614 local_irq_save(flags);
3615 double_rq_lock(this_rq, busiest);
3616 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3617 imbalance, sd, idle, &all_pinned);
3618 double_rq_unlock(this_rq, busiest);
3619 local_irq_restore(flags);
3622 * some other cpu did the load balance for us.
3624 if (ld_moved && this_cpu != smp_processor_id())
3625 resched_cpu(this_cpu);
3627 /* All tasks on this runqueue were pinned by CPU affinity */
3628 if (unlikely(all_pinned)) {
3629 cpumask_clear_cpu(cpu_of(busiest), cpus);
3630 if (!cpumask_empty(cpus))
3637 schedstat_inc(sd, lb_failed[idle]);
3638 sd->nr_balance_failed++;
3640 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3642 spin_lock_irqsave(&busiest->lock, flags);
3644 /* don't kick the migration_thread, if the curr
3645 * task on busiest cpu can't be moved to this_cpu
3647 if (!cpumask_test_cpu(this_cpu,
3648 &busiest->curr->cpus_allowed)) {
3649 spin_unlock_irqrestore(&busiest->lock, flags);
3651 goto out_one_pinned;
3654 if (!busiest->active_balance) {
3655 busiest->active_balance = 1;
3656 busiest->push_cpu = this_cpu;
3659 spin_unlock_irqrestore(&busiest->lock, flags);
3661 wake_up_process(busiest->migration_thread);
3664 * We've kicked active balancing, reset the failure
3667 sd->nr_balance_failed = sd->cache_nice_tries+1;
3670 sd->nr_balance_failed = 0;
3672 if (likely(!active_balance)) {
3673 /* We were unbalanced, so reset the balancing interval */
3674 sd->balance_interval = sd->min_interval;
3677 * If we've begun active balancing, start to back off. This
3678 * case may not be covered by the all_pinned logic if there
3679 * is only 1 task on the busy runqueue (because we don't call
3682 if (sd->balance_interval < sd->max_interval)
3683 sd->balance_interval *= 2;
3686 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3687 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3693 schedstat_inc(sd, lb_balanced[idle]);
3695 sd->nr_balance_failed = 0;
3698 /* tune up the balancing interval */
3699 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3700 (sd->balance_interval < sd->max_interval))
3701 sd->balance_interval *= 2;
3703 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3704 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3715 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3716 * tasks if there is an imbalance.
3718 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3719 * this_rq is locked.
3722 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3723 struct cpumask *cpus)
3725 struct sched_group *group;
3726 struct rq *busiest = NULL;
3727 unsigned long imbalance;
3732 cpumask_setall(cpus);
3735 * When power savings policy is enabled for the parent domain, idle
3736 * sibling can pick up load irrespective of busy siblings. In this case,
3737 * let the state of idle sibling percolate up as IDLE, instead of
3738 * portraying it as CPU_NOT_IDLE.
3740 if (sd->flags & SD_SHARE_CPUPOWER &&
3741 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3744 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3746 update_shares_locked(this_rq, sd);
3747 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3748 &sd_idle, cpus, NULL);
3750 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3754 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3756 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3760 BUG_ON(busiest == this_rq);
3762 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3765 if (busiest->nr_running > 1) {
3766 /* Attempt to move tasks */
3767 double_lock_balance(this_rq, busiest);
3768 /* this_rq->clock is already updated */
3769 update_rq_clock(busiest);
3770 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3771 imbalance, sd, CPU_NEWLY_IDLE,
3773 double_unlock_balance(this_rq, busiest);
3775 if (unlikely(all_pinned)) {
3776 cpumask_clear_cpu(cpu_of(busiest), cpus);
3777 if (!cpumask_empty(cpus))
3783 int active_balance = 0;
3785 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3786 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3787 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3790 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3793 if (sd->nr_balance_failed++ < 2)
3797 * The only task running in a non-idle cpu can be moved to this
3798 * cpu in an attempt to completely freeup the other CPU
3799 * package. The same method used to move task in load_balance()
3800 * have been extended for load_balance_newidle() to speedup
3801 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
3803 * The package power saving logic comes from
3804 * find_busiest_group(). If there are no imbalance, then
3805 * f_b_g() will return NULL. However when sched_mc={1,2} then
3806 * f_b_g() will select a group from which a running task may be
3807 * pulled to this cpu in order to make the other package idle.
3808 * If there is no opportunity to make a package idle and if
3809 * there are no imbalance, then f_b_g() will return NULL and no
3810 * action will be taken in load_balance_newidle().
3812 * Under normal task pull operation due to imbalance, there
3813 * will be more than one task in the source run queue and
3814 * move_tasks() will succeed. ld_moved will be true and this
3815 * active balance code will not be triggered.
3818 /* Lock busiest in correct order while this_rq is held */
3819 double_lock_balance(this_rq, busiest);
3822 * don't kick the migration_thread, if the curr
3823 * task on busiest cpu can't be moved to this_cpu
3825 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
3826 double_unlock_balance(this_rq, busiest);
3831 if (!busiest->active_balance) {
3832 busiest->active_balance = 1;
3833 busiest->push_cpu = this_cpu;
3837 double_unlock_balance(this_rq, busiest);
3839 * Should not call ttwu while holding a rq->lock
3841 spin_unlock(&this_rq->lock);
3843 wake_up_process(busiest->migration_thread);
3844 spin_lock(&this_rq->lock);
3847 sd->nr_balance_failed = 0;
3849 update_shares_locked(this_rq, sd);
3853 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3854 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3855 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3857 sd->nr_balance_failed = 0;
3863 * idle_balance is called by schedule() if this_cpu is about to become
3864 * idle. Attempts to pull tasks from other CPUs.
3866 static void idle_balance(int this_cpu, struct rq *this_rq)
3868 struct sched_domain *sd;
3869 int pulled_task = 0;
3870 unsigned long next_balance = jiffies + HZ;
3871 cpumask_var_t tmpmask;
3873 if (!alloc_cpumask_var(&tmpmask, GFP_ATOMIC))
3876 for_each_domain(this_cpu, sd) {
3877 unsigned long interval;
3879 if (!(sd->flags & SD_LOAD_BALANCE))
3882 if (sd->flags & SD_BALANCE_NEWIDLE)
3883 /* If we've pulled tasks over stop searching: */
3884 pulled_task = load_balance_newidle(this_cpu, this_rq,
3887 interval = msecs_to_jiffies(sd->balance_interval);
3888 if (time_after(next_balance, sd->last_balance + interval))
3889 next_balance = sd->last_balance + interval;
3893 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3895 * We are going idle. next_balance may be set based on
3896 * a busy processor. So reset next_balance.
3898 this_rq->next_balance = next_balance;
3900 free_cpumask_var(tmpmask);
3904 * active_load_balance is run by migration threads. It pushes running tasks
3905 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3906 * running on each physical CPU where possible, and avoids physical /
3907 * logical imbalances.
3909 * Called with busiest_rq locked.
3911 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3913 int target_cpu = busiest_rq->push_cpu;
3914 struct sched_domain *sd;
3915 struct rq *target_rq;
3917 /* Is there any task to move? */
3918 if (busiest_rq->nr_running <= 1)
3921 target_rq = cpu_rq(target_cpu);
3924 * This condition is "impossible", if it occurs
3925 * we need to fix it. Originally reported by
3926 * Bjorn Helgaas on a 128-cpu setup.
3928 BUG_ON(busiest_rq == target_rq);
3930 /* move a task from busiest_rq to target_rq */
3931 double_lock_balance(busiest_rq, target_rq);
3932 update_rq_clock(busiest_rq);
3933 update_rq_clock(target_rq);
3935 /* Search for an sd spanning us and the target CPU. */
3936 for_each_domain(target_cpu, sd) {
3937 if ((sd->flags & SD_LOAD_BALANCE) &&
3938 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3943 schedstat_inc(sd, alb_count);
3945 if (move_one_task(target_rq, target_cpu, busiest_rq,
3947 schedstat_inc(sd, alb_pushed);
3949 schedstat_inc(sd, alb_failed);
3951 double_unlock_balance(busiest_rq, target_rq);
3956 atomic_t load_balancer;
3957 cpumask_var_t cpu_mask;
3958 } nohz ____cacheline_aligned = {
3959 .load_balancer = ATOMIC_INIT(-1),
3963 * This routine will try to nominate the ilb (idle load balancing)
3964 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3965 * load balancing on behalf of all those cpus. If all the cpus in the system
3966 * go into this tickless mode, then there will be no ilb owner (as there is
3967 * no need for one) and all the cpus will sleep till the next wakeup event
3970 * For the ilb owner, tick is not stopped. And this tick will be used
3971 * for idle load balancing. ilb owner will still be part of
3974 * While stopping the tick, this cpu will become the ilb owner if there
3975 * is no other owner. And will be the owner till that cpu becomes busy
3976 * or if all cpus in the system stop their ticks at which point
3977 * there is no need for ilb owner.
3979 * When the ilb owner becomes busy, it nominates another owner, during the
3980 * next busy scheduler_tick()
3982 int select_nohz_load_balancer(int stop_tick)
3984 int cpu = smp_processor_id();
3987 cpu_rq(cpu)->in_nohz_recently = 1;
3989 if (!cpu_active(cpu)) {
3990 if (atomic_read(&nohz.load_balancer) != cpu)
3994 * If we are going offline and still the leader,
3997 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4003 cpumask_set_cpu(cpu, nohz.cpu_mask);
4005 /* time for ilb owner also to sleep */
4006 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4007 if (atomic_read(&nohz.load_balancer) == cpu)
4008 atomic_set(&nohz.load_balancer, -1);
4012 if (atomic_read(&nohz.load_balancer) == -1) {
4013 /* make me the ilb owner */
4014 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4016 } else if (atomic_read(&nohz.load_balancer) == cpu)
4019 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4022 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4024 if (atomic_read(&nohz.load_balancer) == cpu)
4025 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4032 static DEFINE_SPINLOCK(balancing);
4035 * It checks each scheduling domain to see if it is due to be balanced,
4036 * and initiates a balancing operation if so.
4038 * Balancing parameters are set up in arch_init_sched_domains.
4040 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4043 struct rq *rq = cpu_rq(cpu);
4044 unsigned long interval;
4045 struct sched_domain *sd;
4046 /* Earliest time when we have to do rebalance again */
4047 unsigned long next_balance = jiffies + 60*HZ;
4048 int update_next_balance = 0;
4052 /* Fails alloc? Rebalancing probably not a priority right now. */
4053 if (!alloc_cpumask_var(&tmp, GFP_ATOMIC))
4056 for_each_domain(cpu, sd) {
4057 if (!(sd->flags & SD_LOAD_BALANCE))
4060 interval = sd->balance_interval;
4061 if (idle != CPU_IDLE)
4062 interval *= sd->busy_factor;
4064 /* scale ms to jiffies */
4065 interval = msecs_to_jiffies(interval);
4066 if (unlikely(!interval))
4068 if (interval > HZ*NR_CPUS/10)
4069 interval = HZ*NR_CPUS/10;
4071 need_serialize = sd->flags & SD_SERIALIZE;
4073 if (need_serialize) {
4074 if (!spin_trylock(&balancing))
4078 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4079 if (load_balance(cpu, rq, sd, idle, &balance, tmp)) {
4081 * We've pulled tasks over so either we're no
4082 * longer idle, or one of our SMT siblings is
4085 idle = CPU_NOT_IDLE;
4087 sd->last_balance = jiffies;
4090 spin_unlock(&balancing);
4092 if (time_after(next_balance, sd->last_balance + interval)) {
4093 next_balance = sd->last_balance + interval;
4094 update_next_balance = 1;
4098 * Stop the load balance at this level. There is another
4099 * CPU in our sched group which is doing load balancing more
4107 * next_balance will be updated only when there is a need.
4108 * When the cpu is attached to null domain for ex, it will not be
4111 if (likely(update_next_balance))
4112 rq->next_balance = next_balance;
4114 free_cpumask_var(tmp);
4118 * run_rebalance_domains is triggered when needed from the scheduler tick.
4119 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4120 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4122 static void run_rebalance_domains(struct softirq_action *h)
4124 int this_cpu = smp_processor_id();
4125 struct rq *this_rq = cpu_rq(this_cpu);
4126 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4127 CPU_IDLE : CPU_NOT_IDLE;
4129 rebalance_domains(this_cpu, idle);
4133 * If this cpu is the owner for idle load balancing, then do the
4134 * balancing on behalf of the other idle cpus whose ticks are
4137 if (this_rq->idle_at_tick &&
4138 atomic_read(&nohz.load_balancer) == this_cpu) {
4142 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4143 if (balance_cpu == this_cpu)
4147 * If this cpu gets work to do, stop the load balancing
4148 * work being done for other cpus. Next load
4149 * balancing owner will pick it up.
4154 rebalance_domains(balance_cpu, CPU_IDLE);
4156 rq = cpu_rq(balance_cpu);
4157 if (time_after(this_rq->next_balance, rq->next_balance))
4158 this_rq->next_balance = rq->next_balance;
4164 static inline int on_null_domain(int cpu)
4166 return !rcu_dereference(cpu_rq(cpu)->sd);
4170 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4172 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4173 * idle load balancing owner or decide to stop the periodic load balancing,
4174 * if the whole system is idle.
4176 static inline void trigger_load_balance(struct rq *rq, int cpu)
4180 * If we were in the nohz mode recently and busy at the current
4181 * scheduler tick, then check if we need to nominate new idle
4184 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4185 rq->in_nohz_recently = 0;
4187 if (atomic_read(&nohz.load_balancer) == cpu) {
4188 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4189 atomic_set(&nohz.load_balancer, -1);
4192 if (atomic_read(&nohz.load_balancer) == -1) {
4194 * simple selection for now: Nominate the
4195 * first cpu in the nohz list to be the next
4198 * TBD: Traverse the sched domains and nominate
4199 * the nearest cpu in the nohz.cpu_mask.
4201 int ilb = cpumask_first(nohz.cpu_mask);
4203 if (ilb < nr_cpu_ids)
4209 * If this cpu is idle and doing idle load balancing for all the
4210 * cpus with ticks stopped, is it time for that to stop?
4212 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4213 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4219 * If this cpu is idle and the idle load balancing is done by
4220 * someone else, then no need raise the SCHED_SOFTIRQ
4222 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4223 cpumask_test_cpu(cpu, nohz.cpu_mask))
4226 /* Don't need to rebalance while attached to NULL domain */
4227 if (time_after_eq(jiffies, rq->next_balance) &&
4228 likely(!on_null_domain(cpu)))
4229 raise_softirq(SCHED_SOFTIRQ);
4232 #else /* CONFIG_SMP */
4235 * on UP we do not need to balance between CPUs:
4237 static inline void idle_balance(int cpu, struct rq *rq)
4243 DEFINE_PER_CPU(struct kernel_stat, kstat);
4245 EXPORT_PER_CPU_SYMBOL(kstat);
4248 * Return any ns on the sched_clock that have not yet been banked in
4249 * @p in case that task is currently running.
4251 unsigned long long task_delta_exec(struct task_struct *p)
4253 unsigned long flags;
4257 rq = task_rq_lock(p, &flags);
4259 if (task_current(rq, p)) {
4262 update_rq_clock(rq);
4263 delta_exec = rq->clock - p->se.exec_start;
4264 if ((s64)delta_exec > 0)
4268 task_rq_unlock(rq, &flags);
4274 * Account user cpu time to a process.
4275 * @p: the process that the cpu time gets accounted to
4276 * @cputime: the cpu time spent in user space since the last update
4277 * @cputime_scaled: cputime scaled by cpu frequency
4279 void account_user_time(struct task_struct *p, cputime_t cputime,
4280 cputime_t cputime_scaled)
4282 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4285 /* Add user time to process. */
4286 p->utime = cputime_add(p->utime, cputime);
4287 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4288 account_group_user_time(p, cputime);
4290 /* Add user time to cpustat. */
4291 tmp = cputime_to_cputime64(cputime);
4292 if (TASK_NICE(p) > 0)
4293 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4295 cpustat->user = cputime64_add(cpustat->user, tmp);
4296 /* Account for user time used */
4297 acct_update_integrals(p);
4301 * Account guest cpu time to a process.
4302 * @p: the process that the cpu time gets accounted to
4303 * @cputime: the cpu time spent in virtual machine since the last update
4304 * @cputime_scaled: cputime scaled by cpu frequency
4306 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4307 cputime_t cputime_scaled)
4310 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4312 tmp = cputime_to_cputime64(cputime);
4314 /* Add guest time to process. */
4315 p->utime = cputime_add(p->utime, cputime);
4316 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4317 account_group_user_time(p, cputime);
4318 p->gtime = cputime_add(p->gtime, cputime);
4320 /* Add guest time to cpustat. */
4321 cpustat->user = cputime64_add(cpustat->user, tmp);
4322 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4326 * Account system cpu time to a process.
4327 * @p: the process that the cpu time gets accounted to
4328 * @hardirq_offset: the offset to subtract from hardirq_count()
4329 * @cputime: the cpu time spent in kernel space since the last update
4330 * @cputime_scaled: cputime scaled by cpu frequency
4332 void account_system_time(struct task_struct *p, int hardirq_offset,
4333 cputime_t cputime, cputime_t cputime_scaled)
4335 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4338 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4339 account_guest_time(p, cputime, cputime_scaled);
4343 /* Add system time to process. */
4344 p->stime = cputime_add(p->stime, cputime);
4345 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4346 account_group_system_time(p, cputime);
4348 /* Add system time to cpustat. */
4349 tmp = cputime_to_cputime64(cputime);
4350 if (hardirq_count() - hardirq_offset)
4351 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4352 else if (softirq_count())
4353 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4355 cpustat->system = cputime64_add(cpustat->system, tmp);
4357 /* Account for system time used */
4358 acct_update_integrals(p);
4362 * Account for involuntary wait time.
4363 * @steal: the cpu time spent in involuntary wait
4365 void account_steal_time(cputime_t cputime)
4367 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4368 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4370 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4374 * Account for idle time.
4375 * @cputime: the cpu time spent in idle wait
4377 void account_idle_time(cputime_t cputime)
4379 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4380 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4381 struct rq *rq = this_rq();
4383 if (atomic_read(&rq->nr_iowait) > 0)
4384 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4386 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4389 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4392 * Account a single tick of cpu time.
4393 * @p: the process that the cpu time gets accounted to
4394 * @user_tick: indicates if the tick is a user or a system tick
4396 void account_process_tick(struct task_struct *p, int user_tick)
4398 cputime_t one_jiffy = jiffies_to_cputime(1);
4399 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
4400 struct rq *rq = this_rq();
4403 account_user_time(p, one_jiffy, one_jiffy_scaled);
4404 else if (p != rq->idle)
4405 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
4408 account_idle_time(one_jiffy);
4412 * Account multiple ticks of steal time.
4413 * @p: the process from which the cpu time has been stolen
4414 * @ticks: number of stolen ticks
4416 void account_steal_ticks(unsigned long ticks)
4418 account_steal_time(jiffies_to_cputime(ticks));
4422 * Account multiple ticks of idle time.
4423 * @ticks: number of stolen ticks
4425 void account_idle_ticks(unsigned long ticks)
4427 account_idle_time(jiffies_to_cputime(ticks));
4433 * Use precise platform statistics if available:
4435 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4436 cputime_t task_utime(struct task_struct *p)
4441 cputime_t task_stime(struct task_struct *p)
4446 cputime_t task_utime(struct task_struct *p)
4448 clock_t utime = cputime_to_clock_t(p->utime),
4449 total = utime + cputime_to_clock_t(p->stime);
4453 * Use CFS's precise accounting:
4455 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4459 do_div(temp, total);
4461 utime = (clock_t)temp;
4463 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4464 return p->prev_utime;
4467 cputime_t task_stime(struct task_struct *p)
4472 * Use CFS's precise accounting. (we subtract utime from
4473 * the total, to make sure the total observed by userspace
4474 * grows monotonically - apps rely on that):
4476 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4477 cputime_to_clock_t(task_utime(p));
4480 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4482 return p->prev_stime;
4486 inline cputime_t task_gtime(struct task_struct *p)
4492 * This function gets called by the timer code, with HZ frequency.
4493 * We call it with interrupts disabled.
4495 * It also gets called by the fork code, when changing the parent's
4498 void scheduler_tick(void)
4500 int cpu = smp_processor_id();
4501 struct rq *rq = cpu_rq(cpu);
4502 struct task_struct *curr = rq->curr;
4506 spin_lock(&rq->lock);
4507 update_rq_clock(rq);
4508 update_cpu_load(rq);
4509 curr->sched_class->task_tick(rq, curr, 0);
4510 spin_unlock(&rq->lock);
4513 rq->idle_at_tick = idle_cpu(cpu);
4514 trigger_load_balance(rq, cpu);
4518 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4519 defined(CONFIG_PREEMPT_TRACER))
4521 static inline unsigned long get_parent_ip(unsigned long addr)
4523 if (in_lock_functions(addr)) {
4524 addr = CALLER_ADDR2;
4525 if (in_lock_functions(addr))
4526 addr = CALLER_ADDR3;
4531 void __kprobes add_preempt_count(int val)
4533 #ifdef CONFIG_DEBUG_PREEMPT
4537 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4540 preempt_count() += val;
4541 #ifdef CONFIG_DEBUG_PREEMPT
4543 * Spinlock count overflowing soon?
4545 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4548 if (preempt_count() == val)
4549 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4551 EXPORT_SYMBOL(add_preempt_count);
4553 void __kprobes sub_preempt_count(int val)
4555 #ifdef CONFIG_DEBUG_PREEMPT
4559 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4562 * Is the spinlock portion underflowing?
4564 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4565 !(preempt_count() & PREEMPT_MASK)))
4569 if (preempt_count() == val)
4570 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4571 preempt_count() -= val;
4573 EXPORT_SYMBOL(sub_preempt_count);
4578 * Print scheduling while atomic bug:
4580 static noinline void __schedule_bug(struct task_struct *prev)
4582 struct pt_regs *regs = get_irq_regs();
4584 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4585 prev->comm, prev->pid, preempt_count());
4587 debug_show_held_locks(prev);
4589 if (irqs_disabled())
4590 print_irqtrace_events(prev);
4599 * Various schedule()-time debugging checks and statistics:
4601 static inline void schedule_debug(struct task_struct *prev)
4604 * Test if we are atomic. Since do_exit() needs to call into
4605 * schedule() atomically, we ignore that path for now.
4606 * Otherwise, whine if we are scheduling when we should not be.
4608 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4609 __schedule_bug(prev);
4611 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4613 schedstat_inc(this_rq(), sched_count);
4614 #ifdef CONFIG_SCHEDSTATS
4615 if (unlikely(prev->lock_depth >= 0)) {
4616 schedstat_inc(this_rq(), bkl_count);
4617 schedstat_inc(prev, sched_info.bkl_count);
4622 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4624 if (prev->state == TASK_RUNNING) {
4625 u64 runtime = prev->se.sum_exec_runtime;
4627 runtime -= prev->se.prev_sum_exec_runtime;
4628 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
4631 * In order to avoid avg_overlap growing stale when we are
4632 * indeed overlapping and hence not getting put to sleep, grow
4633 * the avg_overlap on preemption.
4635 * We use the average preemption runtime because that
4636 * correlates to the amount of cache footprint a task can
4639 update_avg(&prev->se.avg_overlap, runtime);
4641 prev->sched_class->put_prev_task(rq, prev);
4645 * Pick up the highest-prio task:
4647 static inline struct task_struct *
4648 pick_next_task(struct rq *rq)
4650 const struct sched_class *class;
4651 struct task_struct *p;
4654 * Optimization: we know that if all tasks are in
4655 * the fair class we can call that function directly:
4657 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4658 p = fair_sched_class.pick_next_task(rq);
4663 class = sched_class_highest;
4665 p = class->pick_next_task(rq);
4669 * Will never be NULL as the idle class always
4670 * returns a non-NULL p:
4672 class = class->next;
4677 * schedule() is the main scheduler function.
4679 asmlinkage void __sched schedule(void)
4681 struct task_struct *prev, *next;
4682 unsigned long *switch_count;
4688 cpu = smp_processor_id();
4692 switch_count = &prev->nivcsw;
4694 release_kernel_lock(prev);
4695 need_resched_nonpreemptible:
4697 schedule_debug(prev);
4699 if (sched_feat(HRTICK))
4702 spin_lock_irq(&rq->lock);
4703 update_rq_clock(rq);
4704 clear_tsk_need_resched(prev);
4706 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4707 if (unlikely(signal_pending_state(prev->state, prev)))
4708 prev->state = TASK_RUNNING;
4710 deactivate_task(rq, prev, 1);
4711 switch_count = &prev->nvcsw;
4715 if (prev->sched_class->pre_schedule)
4716 prev->sched_class->pre_schedule(rq, prev);
4719 if (unlikely(!rq->nr_running))
4720 idle_balance(cpu, rq);
4722 put_prev_task(rq, prev);
4723 next = pick_next_task(rq);
4725 if (likely(prev != next)) {
4726 sched_info_switch(prev, next);
4732 context_switch(rq, prev, next); /* unlocks the rq */
4734 * the context switch might have flipped the stack from under
4735 * us, hence refresh the local variables.
4737 cpu = smp_processor_id();
4740 spin_unlock_irq(&rq->lock);
4742 if (unlikely(reacquire_kernel_lock(current) < 0))
4743 goto need_resched_nonpreemptible;
4745 preempt_enable_no_resched();
4746 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4749 EXPORT_SYMBOL(schedule);
4751 #ifdef CONFIG_PREEMPT
4753 * this is the entry point to schedule() from in-kernel preemption
4754 * off of preempt_enable. Kernel preemptions off return from interrupt
4755 * occur there and call schedule directly.
4757 asmlinkage void __sched preempt_schedule(void)
4759 struct thread_info *ti = current_thread_info();
4762 * If there is a non-zero preempt_count or interrupts are disabled,
4763 * we do not want to preempt the current task. Just return..
4765 if (likely(ti->preempt_count || irqs_disabled()))
4769 add_preempt_count(PREEMPT_ACTIVE);
4771 sub_preempt_count(PREEMPT_ACTIVE);
4774 * Check again in case we missed a preemption opportunity
4775 * between schedule and now.
4778 } while (need_resched());
4780 EXPORT_SYMBOL(preempt_schedule);
4783 * this is the entry point to schedule() from kernel preemption
4784 * off of irq context.
4785 * Note, that this is called and return with irqs disabled. This will
4786 * protect us against recursive calling from irq.
4788 asmlinkage void __sched preempt_schedule_irq(void)
4790 struct thread_info *ti = current_thread_info();
4792 /* Catch callers which need to be fixed */
4793 BUG_ON(ti->preempt_count || !irqs_disabled());
4796 add_preempt_count(PREEMPT_ACTIVE);
4799 local_irq_disable();
4800 sub_preempt_count(PREEMPT_ACTIVE);
4803 * Check again in case we missed a preemption opportunity
4804 * between schedule and now.
4807 } while (need_resched());
4810 #endif /* CONFIG_PREEMPT */
4812 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4815 return try_to_wake_up(curr->private, mode, sync);
4817 EXPORT_SYMBOL(default_wake_function);
4820 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4821 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4822 * number) then we wake all the non-exclusive tasks and one exclusive task.
4824 * There are circumstances in which we can try to wake a task which has already
4825 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4826 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4828 void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4829 int nr_exclusive, int sync, void *key)
4831 wait_queue_t *curr, *next;
4833 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4834 unsigned flags = curr->flags;
4836 if (curr->func(curr, mode, sync, key) &&
4837 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4843 * __wake_up - wake up threads blocked on a waitqueue.
4845 * @mode: which threads
4846 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4847 * @key: is directly passed to the wakeup function
4849 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4850 int nr_exclusive, void *key)
4852 unsigned long flags;
4854 spin_lock_irqsave(&q->lock, flags);
4855 __wake_up_common(q, mode, nr_exclusive, 0, key);
4856 spin_unlock_irqrestore(&q->lock, flags);
4858 EXPORT_SYMBOL(__wake_up);
4861 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4863 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4865 __wake_up_common(q, mode, 1, 0, NULL);
4869 * __wake_up_sync - wake up threads blocked on a waitqueue.
4871 * @mode: which threads
4872 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4874 * The sync wakeup differs that the waker knows that it will schedule
4875 * away soon, so while the target thread will be woken up, it will not
4876 * be migrated to another CPU - ie. the two threads are 'synchronized'
4877 * with each other. This can prevent needless bouncing between CPUs.
4879 * On UP it can prevent extra preemption.
4882 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4884 unsigned long flags;
4890 if (unlikely(!nr_exclusive))
4893 spin_lock_irqsave(&q->lock, flags);
4894 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4895 spin_unlock_irqrestore(&q->lock, flags);
4897 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4900 * complete: - signals a single thread waiting on this completion
4901 * @x: holds the state of this particular completion
4903 * This will wake up a single thread waiting on this completion. Threads will be
4904 * awakened in the same order in which they were queued.
4906 * See also complete_all(), wait_for_completion() and related routines.
4908 void complete(struct completion *x)
4910 unsigned long flags;
4912 spin_lock_irqsave(&x->wait.lock, flags);
4914 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4915 spin_unlock_irqrestore(&x->wait.lock, flags);
4917 EXPORT_SYMBOL(complete);
4920 * complete_all: - signals all threads waiting on this completion
4921 * @x: holds the state of this particular completion
4923 * This will wake up all threads waiting on this particular completion event.
4925 void complete_all(struct completion *x)
4927 unsigned long flags;
4929 spin_lock_irqsave(&x->wait.lock, flags);
4930 x->done += UINT_MAX/2;
4931 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4932 spin_unlock_irqrestore(&x->wait.lock, flags);
4934 EXPORT_SYMBOL(complete_all);
4936 static inline long __sched
4937 do_wait_for_common(struct completion *x, long timeout, int state)
4940 DECLARE_WAITQUEUE(wait, current);
4942 wait.flags |= WQ_FLAG_EXCLUSIVE;
4943 __add_wait_queue_tail(&x->wait, &wait);
4945 if (signal_pending_state(state, current)) {
4946 timeout = -ERESTARTSYS;
4949 __set_current_state(state);
4950 spin_unlock_irq(&x->wait.lock);
4951 timeout = schedule_timeout(timeout);
4952 spin_lock_irq(&x->wait.lock);
4953 } while (!x->done && timeout);
4954 __remove_wait_queue(&x->wait, &wait);
4959 return timeout ?: 1;
4963 wait_for_common(struct completion *x, long timeout, int state)
4967 spin_lock_irq(&x->wait.lock);
4968 timeout = do_wait_for_common(x, timeout, state);
4969 spin_unlock_irq(&x->wait.lock);
4974 * wait_for_completion: - waits for completion of a task
4975 * @x: holds the state of this particular completion
4977 * This waits to be signaled for completion of a specific task. It is NOT
4978 * interruptible and there is no timeout.
4980 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4981 * and interrupt capability. Also see complete().
4983 void __sched wait_for_completion(struct completion *x)
4985 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4987 EXPORT_SYMBOL(wait_for_completion);
4990 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4991 * @x: holds the state of this particular completion
4992 * @timeout: timeout value in jiffies
4994 * This waits for either a completion of a specific task to be signaled or for a
4995 * specified timeout to expire. The timeout is in jiffies. It is not
4998 unsigned long __sched
4999 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5001 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5003 EXPORT_SYMBOL(wait_for_completion_timeout);
5006 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5007 * @x: holds the state of this particular completion
5009 * This waits for completion of a specific task to be signaled. It is
5012 int __sched wait_for_completion_interruptible(struct completion *x)
5014 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5015 if (t == -ERESTARTSYS)
5019 EXPORT_SYMBOL(wait_for_completion_interruptible);
5022 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5023 * @x: holds the state of this particular completion
5024 * @timeout: timeout value in jiffies
5026 * This waits for either a completion of a specific task to be signaled or for a
5027 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5029 unsigned long __sched
5030 wait_for_completion_interruptible_timeout(struct completion *x,
5031 unsigned long timeout)
5033 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5035 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5038 * wait_for_completion_killable: - waits for completion of a task (killable)
5039 * @x: holds the state of this particular completion
5041 * This waits to be signaled for completion of a specific task. It can be
5042 * interrupted by a kill signal.
5044 int __sched wait_for_completion_killable(struct completion *x)
5046 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5047 if (t == -ERESTARTSYS)
5051 EXPORT_SYMBOL(wait_for_completion_killable);
5054 * try_wait_for_completion - try to decrement a completion without blocking
5055 * @x: completion structure
5057 * Returns: 0 if a decrement cannot be done without blocking
5058 * 1 if a decrement succeeded.
5060 * If a completion is being used as a counting completion,
5061 * attempt to decrement the counter without blocking. This
5062 * enables us to avoid waiting if the resource the completion
5063 * is protecting is not available.
5065 bool try_wait_for_completion(struct completion *x)
5069 spin_lock_irq(&x->wait.lock);
5074 spin_unlock_irq(&x->wait.lock);
5077 EXPORT_SYMBOL(try_wait_for_completion);
5080 * completion_done - Test to see if a completion has any waiters
5081 * @x: completion structure
5083 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5084 * 1 if there are no waiters.
5087 bool completion_done(struct completion *x)
5091 spin_lock_irq(&x->wait.lock);
5094 spin_unlock_irq(&x->wait.lock);
5097 EXPORT_SYMBOL(completion_done);
5100 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5102 unsigned long flags;
5105 init_waitqueue_entry(&wait, current);
5107 __set_current_state(state);
5109 spin_lock_irqsave(&q->lock, flags);
5110 __add_wait_queue(q, &wait);
5111 spin_unlock(&q->lock);
5112 timeout = schedule_timeout(timeout);
5113 spin_lock_irq(&q->lock);
5114 __remove_wait_queue(q, &wait);
5115 spin_unlock_irqrestore(&q->lock, flags);
5120 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5122 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5124 EXPORT_SYMBOL(interruptible_sleep_on);
5127 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5129 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5131 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5133 void __sched sleep_on(wait_queue_head_t *q)
5135 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5137 EXPORT_SYMBOL(sleep_on);
5139 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5141 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5143 EXPORT_SYMBOL(sleep_on_timeout);
5145 #ifdef CONFIG_RT_MUTEXES
5148 * rt_mutex_setprio - set the current priority of a task
5150 * @prio: prio value (kernel-internal form)
5152 * This function changes the 'effective' priority of a task. It does
5153 * not touch ->normal_prio like __setscheduler().
5155 * Used by the rt_mutex code to implement priority inheritance logic.
5157 void rt_mutex_setprio(struct task_struct *p, int prio)
5159 unsigned long flags;
5160 int oldprio, on_rq, running;
5162 const struct sched_class *prev_class = p->sched_class;
5164 BUG_ON(prio < 0 || prio > MAX_PRIO);
5166 rq = task_rq_lock(p, &flags);
5167 update_rq_clock(rq);
5170 on_rq = p->se.on_rq;
5171 running = task_current(rq, p);
5173 dequeue_task(rq, p, 0);
5175 p->sched_class->put_prev_task(rq, p);
5178 p->sched_class = &rt_sched_class;
5180 p->sched_class = &fair_sched_class;
5185 p->sched_class->set_curr_task(rq);
5187 enqueue_task(rq, p, 0);
5189 check_class_changed(rq, p, prev_class, oldprio, running);
5191 task_rq_unlock(rq, &flags);
5196 void set_user_nice(struct task_struct *p, long nice)
5198 int old_prio, delta, on_rq;
5199 unsigned long flags;
5202 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5205 * We have to be careful, if called from sys_setpriority(),
5206 * the task might be in the middle of scheduling on another CPU.
5208 rq = task_rq_lock(p, &flags);
5209 update_rq_clock(rq);
5211 * The RT priorities are set via sched_setscheduler(), but we still
5212 * allow the 'normal' nice value to be set - but as expected
5213 * it wont have any effect on scheduling until the task is
5214 * SCHED_FIFO/SCHED_RR:
5216 if (task_has_rt_policy(p)) {
5217 p->static_prio = NICE_TO_PRIO(nice);
5220 on_rq = p->se.on_rq;
5222 dequeue_task(rq, p, 0);
5224 p->static_prio = NICE_TO_PRIO(nice);
5227 p->prio = effective_prio(p);
5228 delta = p->prio - old_prio;
5231 enqueue_task(rq, p, 0);
5233 * If the task increased its priority or is running and
5234 * lowered its priority, then reschedule its CPU:
5236 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5237 resched_task(rq->curr);
5240 task_rq_unlock(rq, &flags);
5242 EXPORT_SYMBOL(set_user_nice);
5245 * can_nice - check if a task can reduce its nice value
5249 int can_nice(const struct task_struct *p, const int nice)
5251 /* convert nice value [19,-20] to rlimit style value [1,40] */
5252 int nice_rlim = 20 - nice;
5254 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5255 capable(CAP_SYS_NICE));
5258 #ifdef __ARCH_WANT_SYS_NICE
5261 * sys_nice - change the priority of the current process.
5262 * @increment: priority increment
5264 * sys_setpriority is a more generic, but much slower function that
5265 * does similar things.
5267 SYSCALL_DEFINE1(nice, int, increment)
5272 * Setpriority might change our priority at the same moment.
5273 * We don't have to worry. Conceptually one call occurs first
5274 * and we have a single winner.
5276 if (increment < -40)
5281 nice = TASK_NICE(current) + increment;
5287 if (increment < 0 && !can_nice(current, nice))
5290 retval = security_task_setnice(current, nice);
5294 set_user_nice(current, nice);
5301 * task_prio - return the priority value of a given task.
5302 * @p: the task in question.
5304 * This is the priority value as seen by users in /proc.
5305 * RT tasks are offset by -200. Normal tasks are centered
5306 * around 0, value goes from -16 to +15.
5308 int task_prio(const struct task_struct *p)
5310 return p->prio - MAX_RT_PRIO;
5314 * task_nice - return the nice value of a given task.
5315 * @p: the task in question.
5317 int task_nice(const struct task_struct *p)
5319 return TASK_NICE(p);
5321 EXPORT_SYMBOL(task_nice);
5324 * idle_cpu - is a given cpu idle currently?
5325 * @cpu: the processor in question.
5327 int idle_cpu(int cpu)
5329 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5333 * idle_task - return the idle task for a given cpu.
5334 * @cpu: the processor in question.
5336 struct task_struct *idle_task(int cpu)
5338 return cpu_rq(cpu)->idle;
5342 * find_process_by_pid - find a process with a matching PID value.
5343 * @pid: the pid in question.
5345 static struct task_struct *find_process_by_pid(pid_t pid)
5347 return pid ? find_task_by_vpid(pid) : current;
5350 /* Actually do priority change: must hold rq lock. */
5352 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5354 BUG_ON(p->se.on_rq);
5357 switch (p->policy) {
5361 p->sched_class = &fair_sched_class;
5365 p->sched_class = &rt_sched_class;
5369 p->rt_priority = prio;
5370 p->normal_prio = normal_prio(p);
5371 /* we are holding p->pi_lock already */
5372 p->prio = rt_mutex_getprio(p);
5377 * check the target process has a UID that matches the current process's
5379 static bool check_same_owner(struct task_struct *p)
5381 const struct cred *cred = current_cred(), *pcred;
5385 pcred = __task_cred(p);
5386 match = (cred->euid == pcred->euid ||
5387 cred->euid == pcred->uid);
5392 static int __sched_setscheduler(struct task_struct *p, int policy,
5393 struct sched_param *param, bool user)
5395 int retval, oldprio, oldpolicy = -1, on_rq, running;
5396 unsigned long flags;
5397 const struct sched_class *prev_class = p->sched_class;
5400 /* may grab non-irq protected spin_locks */
5401 BUG_ON(in_interrupt());
5403 /* double check policy once rq lock held */
5405 policy = oldpolicy = p->policy;
5406 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5407 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5408 policy != SCHED_IDLE)
5411 * Valid priorities for SCHED_FIFO and SCHED_RR are
5412 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5413 * SCHED_BATCH and SCHED_IDLE is 0.
5415 if (param->sched_priority < 0 ||
5416 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5417 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5419 if (rt_policy(policy) != (param->sched_priority != 0))
5423 * Allow unprivileged RT tasks to decrease priority:
5425 if (user && !capable(CAP_SYS_NICE)) {
5426 if (rt_policy(policy)) {
5427 unsigned long rlim_rtprio;
5429 if (!lock_task_sighand(p, &flags))
5431 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5432 unlock_task_sighand(p, &flags);
5434 /* can't set/change the rt policy */
5435 if (policy != p->policy && !rlim_rtprio)
5438 /* can't increase priority */
5439 if (param->sched_priority > p->rt_priority &&
5440 param->sched_priority > rlim_rtprio)
5444 * Like positive nice levels, dont allow tasks to
5445 * move out of SCHED_IDLE either:
5447 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5450 /* can't change other user's priorities */
5451 if (!check_same_owner(p))
5456 #ifdef CONFIG_RT_GROUP_SCHED
5458 * Do not allow realtime tasks into groups that have no runtime
5461 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5462 task_group(p)->rt_bandwidth.rt_runtime == 0)
5466 retval = security_task_setscheduler(p, policy, param);
5472 * make sure no PI-waiters arrive (or leave) while we are
5473 * changing the priority of the task:
5475 spin_lock_irqsave(&p->pi_lock, flags);
5477 * To be able to change p->policy safely, the apropriate
5478 * runqueue lock must be held.
5480 rq = __task_rq_lock(p);
5481 /* recheck policy now with rq lock held */
5482 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5483 policy = oldpolicy = -1;
5484 __task_rq_unlock(rq);
5485 spin_unlock_irqrestore(&p->pi_lock, flags);
5488 update_rq_clock(rq);
5489 on_rq = p->se.on_rq;
5490 running = task_current(rq, p);
5492 deactivate_task(rq, p, 0);
5494 p->sched_class->put_prev_task(rq, p);
5497 __setscheduler(rq, p, policy, param->sched_priority);
5500 p->sched_class->set_curr_task(rq);
5502 activate_task(rq, p, 0);
5504 check_class_changed(rq, p, prev_class, oldprio, running);
5506 __task_rq_unlock(rq);
5507 spin_unlock_irqrestore(&p->pi_lock, flags);
5509 rt_mutex_adjust_pi(p);
5515 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5516 * @p: the task in question.
5517 * @policy: new policy.
5518 * @param: structure containing the new RT priority.
5520 * NOTE that the task may be already dead.
5522 int sched_setscheduler(struct task_struct *p, int policy,
5523 struct sched_param *param)
5525 return __sched_setscheduler(p, policy, param, true);
5527 EXPORT_SYMBOL_GPL(sched_setscheduler);
5530 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5531 * @p: the task in question.
5532 * @policy: new policy.
5533 * @param: structure containing the new RT priority.
5535 * Just like sched_setscheduler, only don't bother checking if the
5536 * current context has permission. For example, this is needed in
5537 * stop_machine(): we create temporary high priority worker threads,
5538 * but our caller might not have that capability.
5540 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5541 struct sched_param *param)
5543 return __sched_setscheduler(p, policy, param, false);
5547 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5549 struct sched_param lparam;
5550 struct task_struct *p;
5553 if (!param || pid < 0)
5555 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5560 p = find_process_by_pid(pid);
5562 retval = sched_setscheduler(p, policy, &lparam);
5569 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5570 * @pid: the pid in question.
5571 * @policy: new policy.
5572 * @param: structure containing the new RT priority.
5574 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5575 struct sched_param __user *, param)
5577 /* negative values for policy are not valid */
5581 return do_sched_setscheduler(pid, policy, param);
5585 * sys_sched_setparam - set/change the RT priority of a thread
5586 * @pid: the pid in question.
5587 * @param: structure containing the new RT priority.
5589 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5591 return do_sched_setscheduler(pid, -1, param);
5595 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5596 * @pid: the pid in question.
5598 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5600 struct task_struct *p;
5607 read_lock(&tasklist_lock);
5608 p = find_process_by_pid(pid);
5610 retval = security_task_getscheduler(p);
5614 read_unlock(&tasklist_lock);
5619 * sys_sched_getscheduler - get the RT priority of a thread
5620 * @pid: the pid in question.
5621 * @param: structure containing the RT priority.
5623 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5625 struct sched_param lp;
5626 struct task_struct *p;
5629 if (!param || pid < 0)
5632 read_lock(&tasklist_lock);
5633 p = find_process_by_pid(pid);
5638 retval = security_task_getscheduler(p);
5642 lp.sched_priority = p->rt_priority;
5643 read_unlock(&tasklist_lock);
5646 * This one might sleep, we cannot do it with a spinlock held ...
5648 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5653 read_unlock(&tasklist_lock);
5657 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5659 cpumask_var_t cpus_allowed, new_mask;
5660 struct task_struct *p;
5664 read_lock(&tasklist_lock);
5666 p = find_process_by_pid(pid);
5668 read_unlock(&tasklist_lock);
5674 * It is not safe to call set_cpus_allowed with the
5675 * tasklist_lock held. We will bump the task_struct's
5676 * usage count and then drop tasklist_lock.
5679 read_unlock(&tasklist_lock);
5681 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5685 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5687 goto out_free_cpus_allowed;
5690 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5693 retval = security_task_setscheduler(p, 0, NULL);
5697 cpuset_cpus_allowed(p, cpus_allowed);
5698 cpumask_and(new_mask, in_mask, cpus_allowed);
5700 retval = set_cpus_allowed_ptr(p, new_mask);
5703 cpuset_cpus_allowed(p, cpus_allowed);
5704 if (!cpumask_subset(new_mask, cpus_allowed)) {
5706 * We must have raced with a concurrent cpuset
5707 * update. Just reset the cpus_allowed to the
5708 * cpuset's cpus_allowed
5710 cpumask_copy(new_mask, cpus_allowed);
5715 free_cpumask_var(new_mask);
5716 out_free_cpus_allowed:
5717 free_cpumask_var(cpus_allowed);
5724 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5725 struct cpumask *new_mask)
5727 if (len < cpumask_size())
5728 cpumask_clear(new_mask);
5729 else if (len > cpumask_size())
5730 len = cpumask_size();
5732 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5736 * sys_sched_setaffinity - set the cpu affinity of a process
5737 * @pid: pid of the process
5738 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5739 * @user_mask_ptr: user-space pointer to the new cpu mask
5741 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5742 unsigned long __user *, user_mask_ptr)
5744 cpumask_var_t new_mask;
5747 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5750 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5752 retval = sched_setaffinity(pid, new_mask);
5753 free_cpumask_var(new_mask);
5757 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5759 struct task_struct *p;
5763 read_lock(&tasklist_lock);
5766 p = find_process_by_pid(pid);
5770 retval = security_task_getscheduler(p);
5774 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5777 read_unlock(&tasklist_lock);
5784 * sys_sched_getaffinity - get the cpu affinity of a process
5785 * @pid: pid of the process
5786 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5787 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5789 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5790 unsigned long __user *, user_mask_ptr)
5795 if (len < cpumask_size())
5798 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5801 ret = sched_getaffinity(pid, mask);
5803 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
5806 ret = cpumask_size();
5808 free_cpumask_var(mask);
5814 * sys_sched_yield - yield the current processor to other threads.
5816 * This function yields the current CPU to other tasks. If there are no
5817 * other threads running on this CPU then this function will return.
5819 SYSCALL_DEFINE0(sched_yield)
5821 struct rq *rq = this_rq_lock();
5823 schedstat_inc(rq, yld_count);
5824 current->sched_class->yield_task(rq);
5827 * Since we are going to call schedule() anyway, there's
5828 * no need to preempt or enable interrupts:
5830 __release(rq->lock);
5831 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5832 _raw_spin_unlock(&rq->lock);
5833 preempt_enable_no_resched();
5840 static void __cond_resched(void)
5842 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5843 __might_sleep(__FILE__, __LINE__);
5846 * The BKS might be reacquired before we have dropped
5847 * PREEMPT_ACTIVE, which could trigger a second
5848 * cond_resched() call.
5851 add_preempt_count(PREEMPT_ACTIVE);
5853 sub_preempt_count(PREEMPT_ACTIVE);
5854 } while (need_resched());
5857 int __sched _cond_resched(void)
5859 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5860 system_state == SYSTEM_RUNNING) {
5866 EXPORT_SYMBOL(_cond_resched);
5869 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5870 * call schedule, and on return reacquire the lock.
5872 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5873 * operations here to prevent schedule() from being called twice (once via
5874 * spin_unlock(), once by hand).
5876 int cond_resched_lock(spinlock_t *lock)
5878 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5881 if (spin_needbreak(lock) || resched) {
5883 if (resched && need_resched())
5892 EXPORT_SYMBOL(cond_resched_lock);
5894 int __sched cond_resched_softirq(void)
5896 BUG_ON(!in_softirq());
5898 if (need_resched() && system_state == SYSTEM_RUNNING) {
5906 EXPORT_SYMBOL(cond_resched_softirq);
5909 * yield - yield the current processor to other threads.
5911 * This is a shortcut for kernel-space yielding - it marks the
5912 * thread runnable and calls sys_sched_yield().
5914 void __sched yield(void)
5916 set_current_state(TASK_RUNNING);
5919 EXPORT_SYMBOL(yield);
5922 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5923 * that process accounting knows that this is a task in IO wait state.
5925 * But don't do that if it is a deliberate, throttling IO wait (this task
5926 * has set its backing_dev_info: the queue against which it should throttle)
5928 void __sched io_schedule(void)
5930 struct rq *rq = &__raw_get_cpu_var(runqueues);
5932 delayacct_blkio_start();
5933 atomic_inc(&rq->nr_iowait);
5935 atomic_dec(&rq->nr_iowait);
5936 delayacct_blkio_end();
5938 EXPORT_SYMBOL(io_schedule);
5940 long __sched io_schedule_timeout(long timeout)
5942 struct rq *rq = &__raw_get_cpu_var(runqueues);
5945 delayacct_blkio_start();
5946 atomic_inc(&rq->nr_iowait);
5947 ret = schedule_timeout(timeout);
5948 atomic_dec(&rq->nr_iowait);
5949 delayacct_blkio_end();
5954 * sys_sched_get_priority_max - return maximum RT priority.
5955 * @policy: scheduling class.
5957 * this syscall returns the maximum rt_priority that can be used
5958 * by a given scheduling class.
5960 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5967 ret = MAX_USER_RT_PRIO-1;
5979 * sys_sched_get_priority_min - return minimum RT priority.
5980 * @policy: scheduling class.
5982 * this syscall returns the minimum rt_priority that can be used
5983 * by a given scheduling class.
5985 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6003 * sys_sched_rr_get_interval - return the default timeslice of a process.
6004 * @pid: pid of the process.
6005 * @interval: userspace pointer to the timeslice value.
6007 * this syscall writes the default timeslice value of a given process
6008 * into the user-space timespec buffer. A value of '0' means infinity.
6010 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6011 struct timespec __user *, interval)
6013 struct task_struct *p;
6014 unsigned int time_slice;
6022 read_lock(&tasklist_lock);
6023 p = find_process_by_pid(pid);
6027 retval = security_task_getscheduler(p);
6032 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6033 * tasks that are on an otherwise idle runqueue:
6036 if (p->policy == SCHED_RR) {
6037 time_slice = DEF_TIMESLICE;
6038 } else if (p->policy != SCHED_FIFO) {
6039 struct sched_entity *se = &p->se;
6040 unsigned long flags;
6043 rq = task_rq_lock(p, &flags);
6044 if (rq->cfs.load.weight)
6045 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6046 task_rq_unlock(rq, &flags);
6048 read_unlock(&tasklist_lock);
6049 jiffies_to_timespec(time_slice, &t);
6050 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6054 read_unlock(&tasklist_lock);
6058 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6060 void sched_show_task(struct task_struct *p)
6062 unsigned long free = 0;
6065 state = p->state ? __ffs(p->state) + 1 : 0;
6066 printk(KERN_INFO "%-13.13s %c", p->comm,
6067 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6068 #if BITS_PER_LONG == 32
6069 if (state == TASK_RUNNING)
6070 printk(KERN_CONT " running ");
6072 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6074 if (state == TASK_RUNNING)
6075 printk(KERN_CONT " running task ");
6077 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6079 #ifdef CONFIG_DEBUG_STACK_USAGE
6081 unsigned long *n = end_of_stack(p);
6084 free = (unsigned long)n - (unsigned long)end_of_stack(p);
6087 printk(KERN_CONT "%5lu %5d %6d\n", free,
6088 task_pid_nr(p), task_pid_nr(p->real_parent));
6090 show_stack(p, NULL);
6093 void show_state_filter(unsigned long state_filter)
6095 struct task_struct *g, *p;
6097 #if BITS_PER_LONG == 32
6099 " task PC stack pid father\n");
6102 " task PC stack pid father\n");
6104 read_lock(&tasklist_lock);
6105 do_each_thread(g, p) {
6107 * reset the NMI-timeout, listing all files on a slow
6108 * console might take alot of time:
6110 touch_nmi_watchdog();
6111 if (!state_filter || (p->state & state_filter))
6113 } while_each_thread(g, p);
6115 touch_all_softlockup_watchdogs();
6117 #ifdef CONFIG_SCHED_DEBUG
6118 sysrq_sched_debug_show();
6120 read_unlock(&tasklist_lock);
6122 * Only show locks if all tasks are dumped:
6124 if (state_filter == -1)
6125 debug_show_all_locks();
6128 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6130 idle->sched_class = &idle_sched_class;
6134 * init_idle - set up an idle thread for a given CPU
6135 * @idle: task in question
6136 * @cpu: cpu the idle task belongs to
6138 * NOTE: this function does not set the idle thread's NEED_RESCHED
6139 * flag, to make booting more robust.
6141 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6143 struct rq *rq = cpu_rq(cpu);
6144 unsigned long flags;
6146 spin_lock_irqsave(&rq->lock, flags);
6149 idle->se.exec_start = sched_clock();
6151 idle->prio = idle->normal_prio = MAX_PRIO;
6152 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6153 __set_task_cpu(idle, cpu);
6155 rq->curr = rq->idle = idle;
6156 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6159 spin_unlock_irqrestore(&rq->lock, flags);
6161 /* Set the preempt count _outside_ the spinlocks! */
6162 #if defined(CONFIG_PREEMPT)
6163 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6165 task_thread_info(idle)->preempt_count = 0;
6168 * The idle tasks have their own, simple scheduling class:
6170 idle->sched_class = &idle_sched_class;
6171 ftrace_graph_init_task(idle);
6175 * In a system that switches off the HZ timer nohz_cpu_mask
6176 * indicates which cpus entered this state. This is used
6177 * in the rcu update to wait only for active cpus. For system
6178 * which do not switch off the HZ timer nohz_cpu_mask should
6179 * always be CPU_BITS_NONE.
6181 cpumask_var_t nohz_cpu_mask;
6184 * Increase the granularity value when there are more CPUs,
6185 * because with more CPUs the 'effective latency' as visible
6186 * to users decreases. But the relationship is not linear,
6187 * so pick a second-best guess by going with the log2 of the
6190 * This idea comes from the SD scheduler of Con Kolivas:
6192 static inline void sched_init_granularity(void)
6194 unsigned int factor = 1 + ilog2(num_online_cpus());
6195 const unsigned long limit = 200000000;
6197 sysctl_sched_min_granularity *= factor;
6198 if (sysctl_sched_min_granularity > limit)
6199 sysctl_sched_min_granularity = limit;
6201 sysctl_sched_latency *= factor;
6202 if (sysctl_sched_latency > limit)
6203 sysctl_sched_latency = limit;
6205 sysctl_sched_wakeup_granularity *= factor;
6207 sysctl_sched_shares_ratelimit *= factor;
6212 * This is how migration works:
6214 * 1) we queue a struct migration_req structure in the source CPU's
6215 * runqueue and wake up that CPU's migration thread.
6216 * 2) we down() the locked semaphore => thread blocks.
6217 * 3) migration thread wakes up (implicitly it forces the migrated
6218 * thread off the CPU)
6219 * 4) it gets the migration request and checks whether the migrated
6220 * task is still in the wrong runqueue.
6221 * 5) if it's in the wrong runqueue then the migration thread removes
6222 * it and puts it into the right queue.
6223 * 6) migration thread up()s the semaphore.
6224 * 7) we wake up and the migration is done.
6228 * Change a given task's CPU affinity. Migrate the thread to a
6229 * proper CPU and schedule it away if the CPU it's executing on
6230 * is removed from the allowed bitmask.
6232 * NOTE: the caller must have a valid reference to the task, the
6233 * task must not exit() & deallocate itself prematurely. The
6234 * call is not atomic; no spinlocks may be held.
6236 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6238 struct migration_req req;
6239 unsigned long flags;
6243 rq = task_rq_lock(p, &flags);
6244 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6249 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6250 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6255 if (p->sched_class->set_cpus_allowed)
6256 p->sched_class->set_cpus_allowed(p, new_mask);
6258 cpumask_copy(&p->cpus_allowed, new_mask);
6259 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6262 /* Can the task run on the task's current CPU? If so, we're done */
6263 if (cpumask_test_cpu(task_cpu(p), new_mask))
6266 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6267 /* Need help from migration thread: drop lock and wait. */
6268 task_rq_unlock(rq, &flags);
6269 wake_up_process(rq->migration_thread);
6270 wait_for_completion(&req.done);
6271 tlb_migrate_finish(p->mm);
6275 task_rq_unlock(rq, &flags);
6279 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6282 * Move (not current) task off this cpu, onto dest cpu. We're doing
6283 * this because either it can't run here any more (set_cpus_allowed()
6284 * away from this CPU, or CPU going down), or because we're
6285 * attempting to rebalance this task on exec (sched_exec).
6287 * So we race with normal scheduler movements, but that's OK, as long
6288 * as the task is no longer on this CPU.
6290 * Returns non-zero if task was successfully migrated.
6292 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6294 struct rq *rq_dest, *rq_src;
6297 if (unlikely(!cpu_active(dest_cpu)))
6300 rq_src = cpu_rq(src_cpu);
6301 rq_dest = cpu_rq(dest_cpu);
6303 double_rq_lock(rq_src, rq_dest);
6304 /* Already moved. */
6305 if (task_cpu(p) != src_cpu)
6307 /* Affinity changed (again). */
6308 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6311 on_rq = p->se.on_rq;
6313 deactivate_task(rq_src, p, 0);
6315 set_task_cpu(p, dest_cpu);
6317 activate_task(rq_dest, p, 0);
6318 check_preempt_curr(rq_dest, p, 0);
6323 double_rq_unlock(rq_src, rq_dest);
6328 * migration_thread - this is a highprio system thread that performs
6329 * thread migration by bumping thread off CPU then 'pushing' onto
6332 static int migration_thread(void *data)
6334 int cpu = (long)data;
6338 BUG_ON(rq->migration_thread != current);
6340 set_current_state(TASK_INTERRUPTIBLE);
6341 while (!kthread_should_stop()) {
6342 struct migration_req *req;
6343 struct list_head *head;
6345 spin_lock_irq(&rq->lock);
6347 if (cpu_is_offline(cpu)) {
6348 spin_unlock_irq(&rq->lock);
6352 if (rq->active_balance) {
6353 active_load_balance(rq, cpu);
6354 rq->active_balance = 0;
6357 head = &rq->migration_queue;
6359 if (list_empty(head)) {
6360 spin_unlock_irq(&rq->lock);
6362 set_current_state(TASK_INTERRUPTIBLE);
6365 req = list_entry(head->next, struct migration_req, list);
6366 list_del_init(head->next);
6368 spin_unlock(&rq->lock);
6369 __migrate_task(req->task, cpu, req->dest_cpu);
6372 complete(&req->done);
6374 __set_current_state(TASK_RUNNING);
6378 /* Wait for kthread_stop */
6379 set_current_state(TASK_INTERRUPTIBLE);
6380 while (!kthread_should_stop()) {
6382 set_current_state(TASK_INTERRUPTIBLE);
6384 __set_current_state(TASK_RUNNING);
6388 #ifdef CONFIG_HOTPLUG_CPU
6390 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6394 local_irq_disable();
6395 ret = __migrate_task(p, src_cpu, dest_cpu);
6401 * Figure out where task on dead CPU should go, use force if necessary.
6403 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6406 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
6409 /* Look for allowed, online CPU in same node. */
6410 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
6411 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6414 /* Any allowed, online CPU? */
6415 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
6416 if (dest_cpu < nr_cpu_ids)
6419 /* No more Mr. Nice Guy. */
6420 if (dest_cpu >= nr_cpu_ids) {
6421 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
6422 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
6425 * Don't tell them about moving exiting tasks or
6426 * kernel threads (both mm NULL), since they never
6429 if (p->mm && printk_ratelimit()) {
6430 printk(KERN_INFO "process %d (%s) no "
6431 "longer affine to cpu%d\n",
6432 task_pid_nr(p), p->comm, dead_cpu);
6437 /* It can have affinity changed while we were choosing. */
6438 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
6443 * While a dead CPU has no uninterruptible tasks queued at this point,
6444 * it might still have a nonzero ->nr_uninterruptible counter, because
6445 * for performance reasons the counter is not stricly tracking tasks to
6446 * their home CPUs. So we just add the counter to another CPU's counter,
6447 * to keep the global sum constant after CPU-down:
6449 static void migrate_nr_uninterruptible(struct rq *rq_src)
6451 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
6452 unsigned long flags;
6454 local_irq_save(flags);
6455 double_rq_lock(rq_src, rq_dest);
6456 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6457 rq_src->nr_uninterruptible = 0;
6458 double_rq_unlock(rq_src, rq_dest);
6459 local_irq_restore(flags);
6462 /* Run through task list and migrate tasks from the dead cpu. */
6463 static void migrate_live_tasks(int src_cpu)
6465 struct task_struct *p, *t;
6467 read_lock(&tasklist_lock);
6469 do_each_thread(t, p) {
6473 if (task_cpu(p) == src_cpu)
6474 move_task_off_dead_cpu(src_cpu, p);
6475 } while_each_thread(t, p);
6477 read_unlock(&tasklist_lock);
6481 * Schedules idle task to be the next runnable task on current CPU.
6482 * It does so by boosting its priority to highest possible.
6483 * Used by CPU offline code.
6485 void sched_idle_next(void)
6487 int this_cpu = smp_processor_id();
6488 struct rq *rq = cpu_rq(this_cpu);
6489 struct task_struct *p = rq->idle;
6490 unsigned long flags;
6492 /* cpu has to be offline */
6493 BUG_ON(cpu_online(this_cpu));
6496 * Strictly not necessary since rest of the CPUs are stopped by now
6497 * and interrupts disabled on the current cpu.
6499 spin_lock_irqsave(&rq->lock, flags);
6501 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6503 update_rq_clock(rq);
6504 activate_task(rq, p, 0);
6506 spin_unlock_irqrestore(&rq->lock, flags);
6510 * Ensures that the idle task is using init_mm right before its cpu goes
6513 void idle_task_exit(void)
6515 struct mm_struct *mm = current->active_mm;
6517 BUG_ON(cpu_online(smp_processor_id()));
6520 switch_mm(mm, &init_mm, current);
6524 /* called under rq->lock with disabled interrupts */
6525 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6527 struct rq *rq = cpu_rq(dead_cpu);
6529 /* Must be exiting, otherwise would be on tasklist. */
6530 BUG_ON(!p->exit_state);
6532 /* Cannot have done final schedule yet: would have vanished. */
6533 BUG_ON(p->state == TASK_DEAD);
6538 * Drop lock around migration; if someone else moves it,
6539 * that's OK. No task can be added to this CPU, so iteration is
6542 spin_unlock_irq(&rq->lock);
6543 move_task_off_dead_cpu(dead_cpu, p);
6544 spin_lock_irq(&rq->lock);
6549 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6550 static void migrate_dead_tasks(unsigned int dead_cpu)
6552 struct rq *rq = cpu_rq(dead_cpu);
6553 struct task_struct *next;
6556 if (!rq->nr_running)
6558 update_rq_clock(rq);
6559 next = pick_next_task(rq);
6562 next->sched_class->put_prev_task(rq, next);
6563 migrate_dead(dead_cpu, next);
6567 #endif /* CONFIG_HOTPLUG_CPU */
6569 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6571 static struct ctl_table sd_ctl_dir[] = {
6573 .procname = "sched_domain",
6579 static struct ctl_table sd_ctl_root[] = {
6581 .ctl_name = CTL_KERN,
6582 .procname = "kernel",
6584 .child = sd_ctl_dir,
6589 static struct ctl_table *sd_alloc_ctl_entry(int n)
6591 struct ctl_table *entry =
6592 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6597 static void sd_free_ctl_entry(struct ctl_table **tablep)
6599 struct ctl_table *entry;
6602 * In the intermediate directories, both the child directory and
6603 * procname are dynamically allocated and could fail but the mode
6604 * will always be set. In the lowest directory the names are
6605 * static strings and all have proc handlers.
6607 for (entry = *tablep; entry->mode; entry++) {
6609 sd_free_ctl_entry(&entry->child);
6610 if (entry->proc_handler == NULL)
6611 kfree(entry->procname);
6619 set_table_entry(struct ctl_table *entry,
6620 const char *procname, void *data, int maxlen,
6621 mode_t mode, proc_handler *proc_handler)
6623 entry->procname = procname;
6625 entry->maxlen = maxlen;
6627 entry->proc_handler = proc_handler;
6630 static struct ctl_table *
6631 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6633 struct ctl_table *table = sd_alloc_ctl_entry(13);
6638 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6639 sizeof(long), 0644, proc_doulongvec_minmax);
6640 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6641 sizeof(long), 0644, proc_doulongvec_minmax);
6642 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6643 sizeof(int), 0644, proc_dointvec_minmax);
6644 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6645 sizeof(int), 0644, proc_dointvec_minmax);
6646 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6647 sizeof(int), 0644, proc_dointvec_minmax);
6648 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6649 sizeof(int), 0644, proc_dointvec_minmax);
6650 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6651 sizeof(int), 0644, proc_dointvec_minmax);
6652 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6653 sizeof(int), 0644, proc_dointvec_minmax);
6654 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6655 sizeof(int), 0644, proc_dointvec_minmax);
6656 set_table_entry(&table[9], "cache_nice_tries",
6657 &sd->cache_nice_tries,
6658 sizeof(int), 0644, proc_dointvec_minmax);
6659 set_table_entry(&table[10], "flags", &sd->flags,
6660 sizeof(int), 0644, proc_dointvec_minmax);
6661 set_table_entry(&table[11], "name", sd->name,
6662 CORENAME_MAX_SIZE, 0444, proc_dostring);
6663 /* &table[12] is terminator */
6668 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6670 struct ctl_table *entry, *table;
6671 struct sched_domain *sd;
6672 int domain_num = 0, i;
6675 for_each_domain(cpu, sd)
6677 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6682 for_each_domain(cpu, sd) {
6683 snprintf(buf, 32, "domain%d", i);
6684 entry->procname = kstrdup(buf, GFP_KERNEL);
6686 entry->child = sd_alloc_ctl_domain_table(sd);
6693 static struct ctl_table_header *sd_sysctl_header;
6694 static void register_sched_domain_sysctl(void)
6696 int i, cpu_num = num_online_cpus();
6697 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6700 WARN_ON(sd_ctl_dir[0].child);
6701 sd_ctl_dir[0].child = entry;
6706 for_each_online_cpu(i) {
6707 snprintf(buf, 32, "cpu%d", i);
6708 entry->procname = kstrdup(buf, GFP_KERNEL);
6710 entry->child = sd_alloc_ctl_cpu_table(i);
6714 WARN_ON(sd_sysctl_header);
6715 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6718 /* may be called multiple times per register */
6719 static void unregister_sched_domain_sysctl(void)
6721 if (sd_sysctl_header)
6722 unregister_sysctl_table(sd_sysctl_header);
6723 sd_sysctl_header = NULL;
6724 if (sd_ctl_dir[0].child)
6725 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6728 static void register_sched_domain_sysctl(void)
6731 static void unregister_sched_domain_sysctl(void)
6736 static void set_rq_online(struct rq *rq)
6739 const struct sched_class *class;
6741 cpumask_set_cpu(rq->cpu, rq->rd->online);
6744 for_each_class(class) {
6745 if (class->rq_online)
6746 class->rq_online(rq);
6751 static void set_rq_offline(struct rq *rq)
6754 const struct sched_class *class;
6756 for_each_class(class) {
6757 if (class->rq_offline)
6758 class->rq_offline(rq);
6761 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6767 * migration_call - callback that gets triggered when a CPU is added.
6768 * Here we can start up the necessary migration thread for the new CPU.
6770 static int __cpuinit
6771 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6773 struct task_struct *p;
6774 int cpu = (long)hcpu;
6775 unsigned long flags;
6780 case CPU_UP_PREPARE:
6781 case CPU_UP_PREPARE_FROZEN:
6782 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6785 kthread_bind(p, cpu);
6786 /* Must be high prio: stop_machine expects to yield to it. */
6787 rq = task_rq_lock(p, &flags);
6788 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6789 task_rq_unlock(rq, &flags);
6790 cpu_rq(cpu)->migration_thread = p;
6794 case CPU_ONLINE_FROZEN:
6795 /* Strictly unnecessary, as first user will wake it. */
6796 wake_up_process(cpu_rq(cpu)->migration_thread);
6798 /* Update our root-domain */
6800 spin_lock_irqsave(&rq->lock, flags);
6802 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6806 spin_unlock_irqrestore(&rq->lock, flags);
6809 #ifdef CONFIG_HOTPLUG_CPU
6810 case CPU_UP_CANCELED:
6811 case CPU_UP_CANCELED_FROZEN:
6812 if (!cpu_rq(cpu)->migration_thread)
6814 /* Unbind it from offline cpu so it can run. Fall thru. */
6815 kthread_bind(cpu_rq(cpu)->migration_thread,
6816 cpumask_any(cpu_online_mask));
6817 kthread_stop(cpu_rq(cpu)->migration_thread);
6818 cpu_rq(cpu)->migration_thread = NULL;
6822 case CPU_DEAD_FROZEN:
6823 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6824 migrate_live_tasks(cpu);
6826 kthread_stop(rq->migration_thread);
6827 rq->migration_thread = NULL;
6828 /* Idle task back to normal (off runqueue, low prio) */
6829 spin_lock_irq(&rq->lock);
6830 update_rq_clock(rq);
6831 deactivate_task(rq, rq->idle, 0);
6832 rq->idle->static_prio = MAX_PRIO;
6833 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6834 rq->idle->sched_class = &idle_sched_class;
6835 migrate_dead_tasks(cpu);
6836 spin_unlock_irq(&rq->lock);
6838 migrate_nr_uninterruptible(rq);
6839 BUG_ON(rq->nr_running != 0);
6842 * No need to migrate the tasks: it was best-effort if
6843 * they didn't take sched_hotcpu_mutex. Just wake up
6846 spin_lock_irq(&rq->lock);
6847 while (!list_empty(&rq->migration_queue)) {
6848 struct migration_req *req;
6850 req = list_entry(rq->migration_queue.next,
6851 struct migration_req, list);
6852 list_del_init(&req->list);
6853 spin_unlock_irq(&rq->lock);
6854 complete(&req->done);
6855 spin_lock_irq(&rq->lock);
6857 spin_unlock_irq(&rq->lock);
6861 case CPU_DYING_FROZEN:
6862 /* Update our root-domain */
6864 spin_lock_irqsave(&rq->lock, flags);
6866 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6869 spin_unlock_irqrestore(&rq->lock, flags);
6876 /* Register at highest priority so that task migration (migrate_all_tasks)
6877 * happens before everything else.
6879 static struct notifier_block __cpuinitdata migration_notifier = {
6880 .notifier_call = migration_call,
6884 static int __init migration_init(void)
6886 void *cpu = (void *)(long)smp_processor_id();
6889 /* Start one for the boot CPU: */
6890 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6891 BUG_ON(err == NOTIFY_BAD);
6892 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6893 register_cpu_notifier(&migration_notifier);
6897 early_initcall(migration_init);
6902 #ifdef CONFIG_SCHED_DEBUG
6904 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6905 struct cpumask *groupmask)
6907 struct sched_group *group = sd->groups;
6910 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6911 cpumask_clear(groupmask);
6913 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6915 if (!(sd->flags & SD_LOAD_BALANCE)) {
6916 printk("does not load-balance\n");
6918 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6923 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6925 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6926 printk(KERN_ERR "ERROR: domain->span does not contain "
6929 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6930 printk(KERN_ERR "ERROR: domain->groups does not contain"
6934 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6938 printk(KERN_ERR "ERROR: group is NULL\n");
6942 if (!group->__cpu_power) {
6943 printk(KERN_CONT "\n");
6944 printk(KERN_ERR "ERROR: domain->cpu_power not "
6949 if (!cpumask_weight(sched_group_cpus(group))) {
6950 printk(KERN_CONT "\n");
6951 printk(KERN_ERR "ERROR: empty group\n");
6955 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6956 printk(KERN_CONT "\n");
6957 printk(KERN_ERR "ERROR: repeated CPUs\n");
6961 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6963 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6964 printk(KERN_CONT " %s", str);
6966 group = group->next;
6967 } while (group != sd->groups);
6968 printk(KERN_CONT "\n");
6970 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6971 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6974 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6975 printk(KERN_ERR "ERROR: parent span is not a superset "
6976 "of domain->span\n");
6980 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6982 cpumask_var_t groupmask;
6986 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6990 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6992 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6993 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6998 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7005 free_cpumask_var(groupmask);
7007 #else /* !CONFIG_SCHED_DEBUG */
7008 # define sched_domain_debug(sd, cpu) do { } while (0)
7009 #endif /* CONFIG_SCHED_DEBUG */
7011 static int sd_degenerate(struct sched_domain *sd)
7013 if (cpumask_weight(sched_domain_span(sd)) == 1)
7016 /* Following flags need at least 2 groups */
7017 if (sd->flags & (SD_LOAD_BALANCE |
7018 SD_BALANCE_NEWIDLE |
7022 SD_SHARE_PKG_RESOURCES)) {
7023 if (sd->groups != sd->groups->next)
7027 /* Following flags don't use groups */
7028 if (sd->flags & (SD_WAKE_IDLE |
7037 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7039 unsigned long cflags = sd->flags, pflags = parent->flags;
7041 if (sd_degenerate(parent))
7044 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7047 /* Does parent contain flags not in child? */
7048 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7049 if (cflags & SD_WAKE_AFFINE)
7050 pflags &= ~SD_WAKE_BALANCE;
7051 /* Flags needing groups don't count if only 1 group in parent */
7052 if (parent->groups == parent->groups->next) {
7053 pflags &= ~(SD_LOAD_BALANCE |
7054 SD_BALANCE_NEWIDLE |
7058 SD_SHARE_PKG_RESOURCES);
7059 if (nr_node_ids == 1)
7060 pflags &= ~SD_SERIALIZE;
7062 if (~cflags & pflags)
7068 static void free_rootdomain(struct root_domain *rd)
7070 cpupri_cleanup(&rd->cpupri);
7072 free_cpumask_var(rd->rto_mask);
7073 free_cpumask_var(rd->online);
7074 free_cpumask_var(rd->span);
7078 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7080 struct root_domain *old_rd = NULL;
7081 unsigned long flags;
7083 spin_lock_irqsave(&rq->lock, flags);
7088 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7091 cpumask_clear_cpu(rq->cpu, old_rd->span);
7094 * If we dont want to free the old_rt yet then
7095 * set old_rd to NULL to skip the freeing later
7098 if (!atomic_dec_and_test(&old_rd->refcount))
7102 atomic_inc(&rd->refcount);
7105 cpumask_set_cpu(rq->cpu, rd->span);
7106 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7109 spin_unlock_irqrestore(&rq->lock, flags);
7112 free_rootdomain(old_rd);
7115 static int __init_refok init_rootdomain(struct root_domain *rd, bool bootmem)
7117 memset(rd, 0, sizeof(*rd));
7120 alloc_bootmem_cpumask_var(&def_root_domain.span);
7121 alloc_bootmem_cpumask_var(&def_root_domain.online);
7122 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
7123 cpupri_init(&rd->cpupri, true);
7127 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7129 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7131 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7134 if (cpupri_init(&rd->cpupri, false) != 0)
7139 free_cpumask_var(rd->rto_mask);
7141 free_cpumask_var(rd->online);
7143 free_cpumask_var(rd->span);
7148 static void init_defrootdomain(void)
7150 init_rootdomain(&def_root_domain, true);
7152 atomic_set(&def_root_domain.refcount, 1);
7155 static struct root_domain *alloc_rootdomain(void)
7157 struct root_domain *rd;
7159 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7163 if (init_rootdomain(rd, false) != 0) {
7172 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7173 * hold the hotplug lock.
7176 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7178 struct rq *rq = cpu_rq(cpu);
7179 struct sched_domain *tmp;
7181 /* Remove the sched domains which do not contribute to scheduling. */
7182 for (tmp = sd; tmp; ) {
7183 struct sched_domain *parent = tmp->parent;
7187 if (sd_parent_degenerate(tmp, parent)) {
7188 tmp->parent = parent->parent;
7190 parent->parent->child = tmp;
7195 if (sd && sd_degenerate(sd)) {
7201 sched_domain_debug(sd, cpu);
7203 rq_attach_root(rq, rd);
7204 rcu_assign_pointer(rq->sd, sd);
7207 /* cpus with isolated domains */
7208 static cpumask_var_t cpu_isolated_map;
7210 /* Setup the mask of cpus configured for isolated domains */
7211 static int __init isolated_cpu_setup(char *str)
7213 cpulist_parse(str, cpu_isolated_map);
7217 __setup("isolcpus=", isolated_cpu_setup);
7220 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7221 * to a function which identifies what group(along with sched group) a CPU
7222 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7223 * (due to the fact that we keep track of groups covered with a struct cpumask).
7225 * init_sched_build_groups will build a circular linked list of the groups
7226 * covered by the given span, and will set each group's ->cpumask correctly,
7227 * and ->cpu_power to 0.
7230 init_sched_build_groups(const struct cpumask *span,
7231 const struct cpumask *cpu_map,
7232 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7233 struct sched_group **sg,
7234 struct cpumask *tmpmask),
7235 struct cpumask *covered, struct cpumask *tmpmask)
7237 struct sched_group *first = NULL, *last = NULL;
7240 cpumask_clear(covered);
7242 for_each_cpu(i, span) {
7243 struct sched_group *sg;
7244 int group = group_fn(i, cpu_map, &sg, tmpmask);
7247 if (cpumask_test_cpu(i, covered))
7250 cpumask_clear(sched_group_cpus(sg));
7251 sg->__cpu_power = 0;
7253 for_each_cpu(j, span) {
7254 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7257 cpumask_set_cpu(j, covered);
7258 cpumask_set_cpu(j, sched_group_cpus(sg));
7269 #define SD_NODES_PER_DOMAIN 16
7274 * find_next_best_node - find the next node to include in a sched_domain
7275 * @node: node whose sched_domain we're building
7276 * @used_nodes: nodes already in the sched_domain
7278 * Find the next node to include in a given scheduling domain. Simply
7279 * finds the closest node not already in the @used_nodes map.
7281 * Should use nodemask_t.
7283 static int find_next_best_node(int node, nodemask_t *used_nodes)
7285 int i, n, val, min_val, best_node = 0;
7289 for (i = 0; i < nr_node_ids; i++) {
7290 /* Start at @node */
7291 n = (node + i) % nr_node_ids;
7293 if (!nr_cpus_node(n))
7296 /* Skip already used nodes */
7297 if (node_isset(n, *used_nodes))
7300 /* Simple min distance search */
7301 val = node_distance(node, n);
7303 if (val < min_val) {
7309 node_set(best_node, *used_nodes);
7314 * sched_domain_node_span - get a cpumask for a node's sched_domain
7315 * @node: node whose cpumask we're constructing
7316 * @span: resulting cpumask
7318 * Given a node, construct a good cpumask for its sched_domain to span. It
7319 * should be one that prevents unnecessary balancing, but also spreads tasks
7322 static void sched_domain_node_span(int node, struct cpumask *span)
7324 nodemask_t used_nodes;
7327 cpumask_clear(span);
7328 nodes_clear(used_nodes);
7330 cpumask_or(span, span, cpumask_of_node(node));
7331 node_set(node, used_nodes);
7333 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7334 int next_node = find_next_best_node(node, &used_nodes);
7336 cpumask_or(span, span, cpumask_of_node(next_node));
7339 #endif /* CONFIG_NUMA */
7341 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7344 * The cpus mask in sched_group and sched_domain hangs off the end.
7345 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7346 * for nr_cpu_ids < CONFIG_NR_CPUS.
7348 struct static_sched_group {
7349 struct sched_group sg;
7350 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7353 struct static_sched_domain {
7354 struct sched_domain sd;
7355 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7359 * SMT sched-domains:
7361 #ifdef CONFIG_SCHED_SMT
7362 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7363 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7366 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
7367 struct sched_group **sg, struct cpumask *unused)
7370 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7373 #endif /* CONFIG_SCHED_SMT */
7376 * multi-core sched-domains:
7378 #ifdef CONFIG_SCHED_MC
7379 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7380 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7381 #endif /* CONFIG_SCHED_MC */
7383 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7385 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7386 struct sched_group **sg, struct cpumask *mask)
7390 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7391 group = cpumask_first(mask);
7393 *sg = &per_cpu(sched_group_core, group).sg;
7396 #elif defined(CONFIG_SCHED_MC)
7398 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7399 struct sched_group **sg, struct cpumask *unused)
7402 *sg = &per_cpu(sched_group_core, cpu).sg;
7407 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
7408 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
7411 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
7412 struct sched_group **sg, struct cpumask *mask)
7415 #ifdef CONFIG_SCHED_MC
7416 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
7417 group = cpumask_first(mask);
7418 #elif defined(CONFIG_SCHED_SMT)
7419 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7420 group = cpumask_first(mask);
7425 *sg = &per_cpu(sched_group_phys, group).sg;
7431 * The init_sched_build_groups can't handle what we want to do with node
7432 * groups, so roll our own. Now each node has its own list of groups which
7433 * gets dynamically allocated.
7435 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
7436 static struct sched_group ***sched_group_nodes_bycpu;
7438 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
7439 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7441 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
7442 struct sched_group **sg,
7443 struct cpumask *nodemask)
7447 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7448 group = cpumask_first(nodemask);
7451 *sg = &per_cpu(sched_group_allnodes, group).sg;
7455 static void init_numa_sched_groups_power(struct sched_group *group_head)
7457 struct sched_group *sg = group_head;
7463 for_each_cpu(j, sched_group_cpus(sg)) {
7464 struct sched_domain *sd;
7466 sd = &per_cpu(phys_domains, j).sd;
7467 if (j != cpumask_first(sched_group_cpus(sd->groups))) {
7469 * Only add "power" once for each
7475 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7478 } while (sg != group_head);
7480 #endif /* CONFIG_NUMA */
7483 /* Free memory allocated for various sched_group structures */
7484 static void free_sched_groups(const struct cpumask *cpu_map,
7485 struct cpumask *nodemask)
7489 for_each_cpu(cpu, cpu_map) {
7490 struct sched_group **sched_group_nodes
7491 = sched_group_nodes_bycpu[cpu];
7493 if (!sched_group_nodes)
7496 for (i = 0; i < nr_node_ids; i++) {
7497 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7499 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7500 if (cpumask_empty(nodemask))
7510 if (oldsg != sched_group_nodes[i])
7513 kfree(sched_group_nodes);
7514 sched_group_nodes_bycpu[cpu] = NULL;
7517 #else /* !CONFIG_NUMA */
7518 static void free_sched_groups(const struct cpumask *cpu_map,
7519 struct cpumask *nodemask)
7522 #endif /* CONFIG_NUMA */
7525 * Initialize sched groups cpu_power.
7527 * cpu_power indicates the capacity of sched group, which is used while
7528 * distributing the load between different sched groups in a sched domain.
7529 * Typically cpu_power for all the groups in a sched domain will be same unless
7530 * there are asymmetries in the topology. If there are asymmetries, group
7531 * having more cpu_power will pickup more load compared to the group having
7534 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7535 * the maximum number of tasks a group can handle in the presence of other idle
7536 * or lightly loaded groups in the same sched domain.
7538 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7540 struct sched_domain *child;
7541 struct sched_group *group;
7543 WARN_ON(!sd || !sd->groups);
7545 if (cpu != cpumask_first(sched_group_cpus(sd->groups)))
7550 sd->groups->__cpu_power = 0;
7553 * For perf policy, if the groups in child domain share resources
7554 * (for example cores sharing some portions of the cache hierarchy
7555 * or SMT), then set this domain groups cpu_power such that each group
7556 * can handle only one task, when there are other idle groups in the
7557 * same sched domain.
7559 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7561 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7562 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7567 * add cpu_power of each child group to this groups cpu_power
7569 group = child->groups;
7571 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7572 group = group->next;
7573 } while (group != child->groups);
7577 * Initializers for schedule domains
7578 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7581 #ifdef CONFIG_SCHED_DEBUG
7582 # define SD_INIT_NAME(sd, type) sd->name = #type
7584 # define SD_INIT_NAME(sd, type) do { } while (0)
7587 #define SD_INIT(sd, type) sd_init_##type(sd)
7589 #define SD_INIT_FUNC(type) \
7590 static noinline void sd_init_##type(struct sched_domain *sd) \
7592 memset(sd, 0, sizeof(*sd)); \
7593 *sd = SD_##type##_INIT; \
7594 sd->level = SD_LV_##type; \
7595 SD_INIT_NAME(sd, type); \
7600 SD_INIT_FUNC(ALLNODES)
7603 #ifdef CONFIG_SCHED_SMT
7604 SD_INIT_FUNC(SIBLING)
7606 #ifdef CONFIG_SCHED_MC
7610 static int default_relax_domain_level = -1;
7612 static int __init setup_relax_domain_level(char *str)
7616 val = simple_strtoul(str, NULL, 0);
7617 if (val < SD_LV_MAX)
7618 default_relax_domain_level = val;
7622 __setup("relax_domain_level=", setup_relax_domain_level);
7624 static void set_domain_attribute(struct sched_domain *sd,
7625 struct sched_domain_attr *attr)
7629 if (!attr || attr->relax_domain_level < 0) {
7630 if (default_relax_domain_level < 0)
7633 request = default_relax_domain_level;
7635 request = attr->relax_domain_level;
7636 if (request < sd->level) {
7637 /* turn off idle balance on this domain */
7638 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7640 /* turn on idle balance on this domain */
7641 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7646 * Build sched domains for a given set of cpus and attach the sched domains
7647 * to the individual cpus
7649 static int __build_sched_domains(const struct cpumask *cpu_map,
7650 struct sched_domain_attr *attr)
7652 int i, err = -ENOMEM;
7653 struct root_domain *rd;
7654 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
7657 cpumask_var_t domainspan, covered, notcovered;
7658 struct sched_group **sched_group_nodes = NULL;
7659 int sd_allnodes = 0;
7661 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
7663 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
7664 goto free_domainspan;
7665 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
7669 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
7670 goto free_notcovered;
7671 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
7673 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
7674 goto free_this_sibling_map;
7675 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
7676 goto free_this_core_map;
7677 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
7678 goto free_send_covered;
7682 * Allocate the per-node list of sched groups
7684 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7686 if (!sched_group_nodes) {
7687 printk(KERN_WARNING "Can not alloc sched group node list\n");
7692 rd = alloc_rootdomain();
7694 printk(KERN_WARNING "Cannot alloc root domain\n");
7695 goto free_sched_groups;
7699 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
7703 * Set up domains for cpus specified by the cpu_map.
7705 for_each_cpu(i, cpu_map) {
7706 struct sched_domain *sd = NULL, *p;
7708 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
7711 if (cpumask_weight(cpu_map) >
7712 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
7713 sd = &per_cpu(allnodes_domains, i).sd;
7714 SD_INIT(sd, ALLNODES);
7715 set_domain_attribute(sd, attr);
7716 cpumask_copy(sched_domain_span(sd), cpu_map);
7717 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7723 sd = &per_cpu(node_domains, i).sd;
7725 set_domain_attribute(sd, attr);
7726 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7730 cpumask_and(sched_domain_span(sd),
7731 sched_domain_span(sd), cpu_map);
7735 sd = &per_cpu(phys_domains, i).sd;
7737 set_domain_attribute(sd, attr);
7738 cpumask_copy(sched_domain_span(sd), nodemask);
7742 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7744 #ifdef CONFIG_SCHED_MC
7746 sd = &per_cpu(core_domains, i).sd;
7748 set_domain_attribute(sd, attr);
7749 cpumask_and(sched_domain_span(sd), cpu_map,
7750 cpu_coregroup_mask(i));
7753 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7756 #ifdef CONFIG_SCHED_SMT
7758 sd = &per_cpu(cpu_domains, i).sd;
7759 SD_INIT(sd, SIBLING);
7760 set_domain_attribute(sd, attr);
7761 cpumask_and(sched_domain_span(sd),
7762 &per_cpu(cpu_sibling_map, i), cpu_map);
7765 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7769 #ifdef CONFIG_SCHED_SMT
7770 /* Set up CPU (sibling) groups */
7771 for_each_cpu(i, cpu_map) {
7772 cpumask_and(this_sibling_map,
7773 &per_cpu(cpu_sibling_map, i), cpu_map);
7774 if (i != cpumask_first(this_sibling_map))
7777 init_sched_build_groups(this_sibling_map, cpu_map,
7779 send_covered, tmpmask);
7783 #ifdef CONFIG_SCHED_MC
7784 /* Set up multi-core groups */
7785 for_each_cpu(i, cpu_map) {
7786 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
7787 if (i != cpumask_first(this_core_map))
7790 init_sched_build_groups(this_core_map, cpu_map,
7792 send_covered, tmpmask);
7796 /* Set up physical groups */
7797 for (i = 0; i < nr_node_ids; i++) {
7798 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7799 if (cpumask_empty(nodemask))
7802 init_sched_build_groups(nodemask, cpu_map,
7804 send_covered, tmpmask);
7808 /* Set up node groups */
7810 init_sched_build_groups(cpu_map, cpu_map,
7811 &cpu_to_allnodes_group,
7812 send_covered, tmpmask);
7815 for (i = 0; i < nr_node_ids; i++) {
7816 /* Set up node groups */
7817 struct sched_group *sg, *prev;
7820 cpumask_clear(covered);
7821 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7822 if (cpumask_empty(nodemask)) {
7823 sched_group_nodes[i] = NULL;
7827 sched_domain_node_span(i, domainspan);
7828 cpumask_and(domainspan, domainspan, cpu_map);
7830 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7833 printk(KERN_WARNING "Can not alloc domain group for "
7837 sched_group_nodes[i] = sg;
7838 for_each_cpu(j, nodemask) {
7839 struct sched_domain *sd;
7841 sd = &per_cpu(node_domains, j).sd;
7844 sg->__cpu_power = 0;
7845 cpumask_copy(sched_group_cpus(sg), nodemask);
7847 cpumask_or(covered, covered, nodemask);
7850 for (j = 0; j < nr_node_ids; j++) {
7851 int n = (i + j) % nr_node_ids;
7853 cpumask_complement(notcovered, covered);
7854 cpumask_and(tmpmask, notcovered, cpu_map);
7855 cpumask_and(tmpmask, tmpmask, domainspan);
7856 if (cpumask_empty(tmpmask))
7859 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
7860 if (cpumask_empty(tmpmask))
7863 sg = kmalloc_node(sizeof(struct sched_group) +
7868 "Can not alloc domain group for node %d\n", j);
7871 sg->__cpu_power = 0;
7872 cpumask_copy(sched_group_cpus(sg), tmpmask);
7873 sg->next = prev->next;
7874 cpumask_or(covered, covered, tmpmask);
7881 /* Calculate CPU power for physical packages and nodes */
7882 #ifdef CONFIG_SCHED_SMT
7883 for_each_cpu(i, cpu_map) {
7884 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
7886 init_sched_groups_power(i, sd);
7889 #ifdef CONFIG_SCHED_MC
7890 for_each_cpu(i, cpu_map) {
7891 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
7893 init_sched_groups_power(i, sd);
7897 for_each_cpu(i, cpu_map) {
7898 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
7900 init_sched_groups_power(i, sd);
7904 for (i = 0; i < nr_node_ids; i++)
7905 init_numa_sched_groups_power(sched_group_nodes[i]);
7908 struct sched_group *sg;
7910 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7912 init_numa_sched_groups_power(sg);
7916 /* Attach the domains */
7917 for_each_cpu(i, cpu_map) {
7918 struct sched_domain *sd;
7919 #ifdef CONFIG_SCHED_SMT
7920 sd = &per_cpu(cpu_domains, i).sd;
7921 #elif defined(CONFIG_SCHED_MC)
7922 sd = &per_cpu(core_domains, i).sd;
7924 sd = &per_cpu(phys_domains, i).sd;
7926 cpu_attach_domain(sd, rd, i);
7932 free_cpumask_var(tmpmask);
7934 free_cpumask_var(send_covered);
7936 free_cpumask_var(this_core_map);
7937 free_this_sibling_map:
7938 free_cpumask_var(this_sibling_map);
7940 free_cpumask_var(nodemask);
7943 free_cpumask_var(notcovered);
7945 free_cpumask_var(covered);
7947 free_cpumask_var(domainspan);
7954 kfree(sched_group_nodes);
7960 free_sched_groups(cpu_map, tmpmask);
7961 free_rootdomain(rd);
7966 static int build_sched_domains(const struct cpumask *cpu_map)
7968 return __build_sched_domains(cpu_map, NULL);
7971 static struct cpumask *doms_cur; /* current sched domains */
7972 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7973 static struct sched_domain_attr *dattr_cur;
7974 /* attribues of custom domains in 'doms_cur' */
7977 * Special case: If a kmalloc of a doms_cur partition (array of
7978 * cpumask) fails, then fallback to a single sched domain,
7979 * as determined by the single cpumask fallback_doms.
7981 static cpumask_var_t fallback_doms;
7984 * arch_update_cpu_topology lets virtualized architectures update the
7985 * cpu core maps. It is supposed to return 1 if the topology changed
7986 * or 0 if it stayed the same.
7988 int __attribute__((weak)) arch_update_cpu_topology(void)
7994 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7995 * For now this just excludes isolated cpus, but could be used to
7996 * exclude other special cases in the future.
7998 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8002 arch_update_cpu_topology();
8004 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8006 doms_cur = fallback_doms;
8007 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8009 err = build_sched_domains(doms_cur);
8010 register_sched_domain_sysctl();
8015 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8016 struct cpumask *tmpmask)
8018 free_sched_groups(cpu_map, tmpmask);
8022 * Detach sched domains from a group of cpus specified in cpu_map
8023 * These cpus will now be attached to the NULL domain
8025 static void detach_destroy_domains(const struct cpumask *cpu_map)
8027 /* Save because hotplug lock held. */
8028 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8031 for_each_cpu(i, cpu_map)
8032 cpu_attach_domain(NULL, &def_root_domain, i);
8033 synchronize_sched();
8034 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8037 /* handle null as "default" */
8038 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8039 struct sched_domain_attr *new, int idx_new)
8041 struct sched_domain_attr tmp;
8048 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8049 new ? (new + idx_new) : &tmp,
8050 sizeof(struct sched_domain_attr));
8054 * Partition sched domains as specified by the 'ndoms_new'
8055 * cpumasks in the array doms_new[] of cpumasks. This compares
8056 * doms_new[] to the current sched domain partitioning, doms_cur[].
8057 * It destroys each deleted domain and builds each new domain.
8059 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8060 * The masks don't intersect (don't overlap.) We should setup one
8061 * sched domain for each mask. CPUs not in any of the cpumasks will
8062 * not be load balanced. If the same cpumask appears both in the
8063 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8066 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8067 * ownership of it and will kfree it when done with it. If the caller
8068 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8069 * ndoms_new == 1, and partition_sched_domains() will fallback to
8070 * the single partition 'fallback_doms', it also forces the domains
8073 * If doms_new == NULL it will be replaced with cpu_online_mask.
8074 * ndoms_new == 0 is a special case for destroying existing domains,
8075 * and it will not create the default domain.
8077 * Call with hotplug lock held
8079 /* FIXME: Change to struct cpumask *doms_new[] */
8080 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8081 struct sched_domain_attr *dattr_new)
8086 mutex_lock(&sched_domains_mutex);
8088 /* always unregister in case we don't destroy any domains */
8089 unregister_sched_domain_sysctl();
8091 /* Let architecture update cpu core mappings. */
8092 new_topology = arch_update_cpu_topology();
8094 n = doms_new ? ndoms_new : 0;
8096 /* Destroy deleted domains */
8097 for (i = 0; i < ndoms_cur; i++) {
8098 for (j = 0; j < n && !new_topology; j++) {
8099 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8100 && dattrs_equal(dattr_cur, i, dattr_new, j))
8103 /* no match - a current sched domain not in new doms_new[] */
8104 detach_destroy_domains(doms_cur + i);
8109 if (doms_new == NULL) {
8111 doms_new = fallback_doms;
8112 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8113 WARN_ON_ONCE(dattr_new);
8116 /* Build new domains */
8117 for (i = 0; i < ndoms_new; i++) {
8118 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8119 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8120 && dattrs_equal(dattr_new, i, dattr_cur, j))
8123 /* no match - add a new doms_new */
8124 __build_sched_domains(doms_new + i,
8125 dattr_new ? dattr_new + i : NULL);
8130 /* Remember the new sched domains */
8131 if (doms_cur != fallback_doms)
8133 kfree(dattr_cur); /* kfree(NULL) is safe */
8134 doms_cur = doms_new;
8135 dattr_cur = dattr_new;
8136 ndoms_cur = ndoms_new;
8138 register_sched_domain_sysctl();
8140 mutex_unlock(&sched_domains_mutex);
8143 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8144 static void arch_reinit_sched_domains(void)
8148 /* Destroy domains first to force the rebuild */
8149 partition_sched_domains(0, NULL, NULL);
8151 rebuild_sched_domains();
8155 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8157 unsigned int level = 0;
8159 if (sscanf(buf, "%u", &level) != 1)
8163 * level is always be positive so don't check for
8164 * level < POWERSAVINGS_BALANCE_NONE which is 0
8165 * What happens on 0 or 1 byte write,
8166 * need to check for count as well?
8169 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8173 sched_smt_power_savings = level;
8175 sched_mc_power_savings = level;
8177 arch_reinit_sched_domains();
8182 #ifdef CONFIG_SCHED_MC
8183 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8186 return sprintf(page, "%u\n", sched_mc_power_savings);
8188 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8189 const char *buf, size_t count)
8191 return sched_power_savings_store(buf, count, 0);
8193 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8194 sched_mc_power_savings_show,
8195 sched_mc_power_savings_store);
8198 #ifdef CONFIG_SCHED_SMT
8199 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8202 return sprintf(page, "%u\n", sched_smt_power_savings);
8204 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8205 const char *buf, size_t count)
8207 return sched_power_savings_store(buf, count, 1);
8209 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8210 sched_smt_power_savings_show,
8211 sched_smt_power_savings_store);
8214 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8218 #ifdef CONFIG_SCHED_SMT
8220 err = sysfs_create_file(&cls->kset.kobj,
8221 &attr_sched_smt_power_savings.attr);
8223 #ifdef CONFIG_SCHED_MC
8224 if (!err && mc_capable())
8225 err = sysfs_create_file(&cls->kset.kobj,
8226 &attr_sched_mc_power_savings.attr);
8230 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8232 #ifndef CONFIG_CPUSETS
8234 * Add online and remove offline CPUs from the scheduler domains.
8235 * When cpusets are enabled they take over this function.
8237 static int update_sched_domains(struct notifier_block *nfb,
8238 unsigned long action, void *hcpu)
8242 case CPU_ONLINE_FROZEN:
8244 case CPU_DEAD_FROZEN:
8245 partition_sched_domains(1, NULL, NULL);
8254 static int update_runtime(struct notifier_block *nfb,
8255 unsigned long action, void *hcpu)
8257 int cpu = (int)(long)hcpu;
8260 case CPU_DOWN_PREPARE:
8261 case CPU_DOWN_PREPARE_FROZEN:
8262 disable_runtime(cpu_rq(cpu));
8265 case CPU_DOWN_FAILED:
8266 case CPU_DOWN_FAILED_FROZEN:
8268 case CPU_ONLINE_FROZEN:
8269 enable_runtime(cpu_rq(cpu));
8277 void __init sched_init_smp(void)
8279 cpumask_var_t non_isolated_cpus;
8281 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8283 #if defined(CONFIG_NUMA)
8284 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8286 BUG_ON(sched_group_nodes_bycpu == NULL);
8289 mutex_lock(&sched_domains_mutex);
8290 arch_init_sched_domains(cpu_online_mask);
8291 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8292 if (cpumask_empty(non_isolated_cpus))
8293 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8294 mutex_unlock(&sched_domains_mutex);
8297 #ifndef CONFIG_CPUSETS
8298 /* XXX: Theoretical race here - CPU may be hotplugged now */
8299 hotcpu_notifier(update_sched_domains, 0);
8302 /* RT runtime code needs to handle some hotplug events */
8303 hotcpu_notifier(update_runtime, 0);
8307 /* Move init over to a non-isolated CPU */
8308 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8310 sched_init_granularity();
8311 free_cpumask_var(non_isolated_cpus);
8313 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8314 init_sched_rt_class();
8317 void __init sched_init_smp(void)
8319 sched_init_granularity();
8321 #endif /* CONFIG_SMP */
8323 int in_sched_functions(unsigned long addr)
8325 return in_lock_functions(addr) ||
8326 (addr >= (unsigned long)__sched_text_start
8327 && addr < (unsigned long)__sched_text_end);
8330 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8332 cfs_rq->tasks_timeline = RB_ROOT;
8333 INIT_LIST_HEAD(&cfs_rq->tasks);
8334 #ifdef CONFIG_FAIR_GROUP_SCHED
8337 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8340 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8342 struct rt_prio_array *array;
8345 array = &rt_rq->active;
8346 for (i = 0; i < MAX_RT_PRIO; i++) {
8347 INIT_LIST_HEAD(array->queue + i);
8348 __clear_bit(i, array->bitmap);
8350 /* delimiter for bitsearch: */
8351 __set_bit(MAX_RT_PRIO, array->bitmap);
8353 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8354 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8356 rt_rq->highest_prio.next = MAX_RT_PRIO;
8360 rt_rq->rt_nr_migratory = 0;
8361 rt_rq->overloaded = 0;
8362 plist_head_init(&rq->rt.pushable_tasks, &rq->lock);
8366 rt_rq->rt_throttled = 0;
8367 rt_rq->rt_runtime = 0;
8368 spin_lock_init(&rt_rq->rt_runtime_lock);
8370 #ifdef CONFIG_RT_GROUP_SCHED
8371 rt_rq->rt_nr_boosted = 0;
8376 #ifdef CONFIG_FAIR_GROUP_SCHED
8377 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8378 struct sched_entity *se, int cpu, int add,
8379 struct sched_entity *parent)
8381 struct rq *rq = cpu_rq(cpu);
8382 tg->cfs_rq[cpu] = cfs_rq;
8383 init_cfs_rq(cfs_rq, rq);
8386 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8389 /* se could be NULL for init_task_group */
8394 se->cfs_rq = &rq->cfs;
8396 se->cfs_rq = parent->my_q;
8399 se->load.weight = tg->shares;
8400 se->load.inv_weight = 0;
8401 se->parent = parent;
8405 #ifdef CONFIG_RT_GROUP_SCHED
8406 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8407 struct sched_rt_entity *rt_se, int cpu, int add,
8408 struct sched_rt_entity *parent)
8410 struct rq *rq = cpu_rq(cpu);
8412 tg->rt_rq[cpu] = rt_rq;
8413 init_rt_rq(rt_rq, rq);
8415 rt_rq->rt_se = rt_se;
8416 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8418 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8420 tg->rt_se[cpu] = rt_se;
8425 rt_se->rt_rq = &rq->rt;
8427 rt_se->rt_rq = parent->my_q;
8429 rt_se->my_q = rt_rq;
8430 rt_se->parent = parent;
8431 INIT_LIST_HEAD(&rt_se->run_list);
8435 void __init sched_init(void)
8438 unsigned long alloc_size = 0, ptr;
8440 #ifdef CONFIG_FAIR_GROUP_SCHED
8441 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8443 #ifdef CONFIG_RT_GROUP_SCHED
8444 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8446 #ifdef CONFIG_USER_SCHED
8450 * As sched_init() is called before page_alloc is setup,
8451 * we use alloc_bootmem().
8454 ptr = (unsigned long)alloc_bootmem(alloc_size);
8456 #ifdef CONFIG_FAIR_GROUP_SCHED
8457 init_task_group.se = (struct sched_entity **)ptr;
8458 ptr += nr_cpu_ids * sizeof(void **);
8460 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8461 ptr += nr_cpu_ids * sizeof(void **);
8463 #ifdef CONFIG_USER_SCHED
8464 root_task_group.se = (struct sched_entity **)ptr;
8465 ptr += nr_cpu_ids * sizeof(void **);
8467 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8468 ptr += nr_cpu_ids * sizeof(void **);
8469 #endif /* CONFIG_USER_SCHED */
8470 #endif /* CONFIG_FAIR_GROUP_SCHED */
8471 #ifdef CONFIG_RT_GROUP_SCHED
8472 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8473 ptr += nr_cpu_ids * sizeof(void **);
8475 init_task_group.rt_rq = (struct rt_rq **)ptr;
8476 ptr += nr_cpu_ids * sizeof(void **);
8478 #ifdef CONFIG_USER_SCHED
8479 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8480 ptr += nr_cpu_ids * sizeof(void **);
8482 root_task_group.rt_rq = (struct rt_rq **)ptr;
8483 ptr += nr_cpu_ids * sizeof(void **);
8484 #endif /* CONFIG_USER_SCHED */
8485 #endif /* CONFIG_RT_GROUP_SCHED */
8489 init_defrootdomain();
8492 init_rt_bandwidth(&def_rt_bandwidth,
8493 global_rt_period(), global_rt_runtime());
8495 #ifdef CONFIG_RT_GROUP_SCHED
8496 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8497 global_rt_period(), global_rt_runtime());
8498 #ifdef CONFIG_USER_SCHED
8499 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8500 global_rt_period(), RUNTIME_INF);
8501 #endif /* CONFIG_USER_SCHED */
8502 #endif /* CONFIG_RT_GROUP_SCHED */
8504 #ifdef CONFIG_GROUP_SCHED
8505 list_add(&init_task_group.list, &task_groups);
8506 INIT_LIST_HEAD(&init_task_group.children);
8508 #ifdef CONFIG_USER_SCHED
8509 INIT_LIST_HEAD(&root_task_group.children);
8510 init_task_group.parent = &root_task_group;
8511 list_add(&init_task_group.siblings, &root_task_group.children);
8512 #endif /* CONFIG_USER_SCHED */
8513 #endif /* CONFIG_GROUP_SCHED */
8515 for_each_possible_cpu(i) {
8519 spin_lock_init(&rq->lock);
8521 init_cfs_rq(&rq->cfs, rq);
8522 init_rt_rq(&rq->rt, rq);
8523 #ifdef CONFIG_FAIR_GROUP_SCHED
8524 init_task_group.shares = init_task_group_load;
8525 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8526 #ifdef CONFIG_CGROUP_SCHED
8528 * How much cpu bandwidth does init_task_group get?
8530 * In case of task-groups formed thr' the cgroup filesystem, it
8531 * gets 100% of the cpu resources in the system. This overall
8532 * system cpu resource is divided among the tasks of
8533 * init_task_group and its child task-groups in a fair manner,
8534 * based on each entity's (task or task-group's) weight
8535 * (se->load.weight).
8537 * In other words, if init_task_group has 10 tasks of weight
8538 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8539 * then A0's share of the cpu resource is:
8541 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8543 * We achieve this by letting init_task_group's tasks sit
8544 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8546 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8547 #elif defined CONFIG_USER_SCHED
8548 root_task_group.shares = NICE_0_LOAD;
8549 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8551 * In case of task-groups formed thr' the user id of tasks,
8552 * init_task_group represents tasks belonging to root user.
8553 * Hence it forms a sibling of all subsequent groups formed.
8554 * In this case, init_task_group gets only a fraction of overall
8555 * system cpu resource, based on the weight assigned to root
8556 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8557 * by letting tasks of init_task_group sit in a separate cfs_rq
8558 * (init_cfs_rq) and having one entity represent this group of
8559 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8561 init_tg_cfs_entry(&init_task_group,
8562 &per_cpu(init_cfs_rq, i),
8563 &per_cpu(init_sched_entity, i), i, 1,
8564 root_task_group.se[i]);
8567 #endif /* CONFIG_FAIR_GROUP_SCHED */
8569 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8570 #ifdef CONFIG_RT_GROUP_SCHED
8571 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8572 #ifdef CONFIG_CGROUP_SCHED
8573 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8574 #elif defined CONFIG_USER_SCHED
8575 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8576 init_tg_rt_entry(&init_task_group,
8577 &per_cpu(init_rt_rq, i),
8578 &per_cpu(init_sched_rt_entity, i), i, 1,
8579 root_task_group.rt_se[i]);
8583 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8584 rq->cpu_load[j] = 0;
8588 rq->active_balance = 0;
8589 rq->next_balance = jiffies;
8593 rq->migration_thread = NULL;
8594 INIT_LIST_HEAD(&rq->migration_queue);
8595 rq_attach_root(rq, &def_root_domain);
8598 atomic_set(&rq->nr_iowait, 0);
8601 set_load_weight(&init_task);
8603 #ifdef CONFIG_PREEMPT_NOTIFIERS
8604 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8608 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8611 #ifdef CONFIG_RT_MUTEXES
8612 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8616 * The boot idle thread does lazy MMU switching as well:
8618 atomic_inc(&init_mm.mm_count);
8619 enter_lazy_tlb(&init_mm, current);
8622 * Make us the idle thread. Technically, schedule() should not be
8623 * called from this thread, however somewhere below it might be,
8624 * but because we are the idle thread, we just pick up running again
8625 * when this runqueue becomes "idle".
8627 init_idle(current, smp_processor_id());
8629 * During early bootup we pretend to be a normal task:
8631 current->sched_class = &fair_sched_class;
8633 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8634 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
8637 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
8639 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8642 scheduler_running = 1;
8645 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8646 void __might_sleep(char *file, int line)
8649 static unsigned long prev_jiffy; /* ratelimiting */
8651 if ((!in_atomic() && !irqs_disabled()) ||
8652 system_state != SYSTEM_RUNNING || oops_in_progress)
8654 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8656 prev_jiffy = jiffies;
8659 "BUG: sleeping function called from invalid context at %s:%d\n",
8662 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8663 in_atomic(), irqs_disabled(),
8664 current->pid, current->comm);
8666 debug_show_held_locks(current);
8667 if (irqs_disabled())
8668 print_irqtrace_events(current);
8672 EXPORT_SYMBOL(__might_sleep);
8675 #ifdef CONFIG_MAGIC_SYSRQ
8676 static void normalize_task(struct rq *rq, struct task_struct *p)
8680 update_rq_clock(rq);
8681 on_rq = p->se.on_rq;
8683 deactivate_task(rq, p, 0);
8684 __setscheduler(rq, p, SCHED_NORMAL, 0);
8686 activate_task(rq, p, 0);
8687 resched_task(rq->curr);
8691 void normalize_rt_tasks(void)
8693 struct task_struct *g, *p;
8694 unsigned long flags;
8697 read_lock_irqsave(&tasklist_lock, flags);
8698 do_each_thread(g, p) {
8700 * Only normalize user tasks:
8705 p->se.exec_start = 0;
8706 #ifdef CONFIG_SCHEDSTATS
8707 p->se.wait_start = 0;
8708 p->se.sleep_start = 0;
8709 p->se.block_start = 0;
8714 * Renice negative nice level userspace
8717 if (TASK_NICE(p) < 0 && p->mm)
8718 set_user_nice(p, 0);
8722 spin_lock(&p->pi_lock);
8723 rq = __task_rq_lock(p);
8725 normalize_task(rq, p);
8727 __task_rq_unlock(rq);
8728 spin_unlock(&p->pi_lock);
8729 } while_each_thread(g, p);
8731 read_unlock_irqrestore(&tasklist_lock, flags);
8734 #endif /* CONFIG_MAGIC_SYSRQ */
8738 * These functions are only useful for the IA64 MCA handling.
8740 * They can only be called when the whole system has been
8741 * stopped - every CPU needs to be quiescent, and no scheduling
8742 * activity can take place. Using them for anything else would
8743 * be a serious bug, and as a result, they aren't even visible
8744 * under any other configuration.
8748 * curr_task - return the current task for a given cpu.
8749 * @cpu: the processor in question.
8751 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8753 struct task_struct *curr_task(int cpu)
8755 return cpu_curr(cpu);
8759 * set_curr_task - set the current task for a given cpu.
8760 * @cpu: the processor in question.
8761 * @p: the task pointer to set.
8763 * Description: This function must only be used when non-maskable interrupts
8764 * are serviced on a separate stack. It allows the architecture to switch the
8765 * notion of the current task on a cpu in a non-blocking manner. This function
8766 * must be called with all CPU's synchronized, and interrupts disabled, the
8767 * and caller must save the original value of the current task (see
8768 * curr_task() above) and restore that value before reenabling interrupts and
8769 * re-starting the system.
8771 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8773 void set_curr_task(int cpu, struct task_struct *p)
8780 #ifdef CONFIG_FAIR_GROUP_SCHED
8781 static void free_fair_sched_group(struct task_group *tg)
8785 for_each_possible_cpu(i) {
8787 kfree(tg->cfs_rq[i]);
8797 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8799 struct cfs_rq *cfs_rq;
8800 struct sched_entity *se;
8804 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8807 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8811 tg->shares = NICE_0_LOAD;
8813 for_each_possible_cpu(i) {
8816 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8817 GFP_KERNEL, cpu_to_node(i));
8821 se = kzalloc_node(sizeof(struct sched_entity),
8822 GFP_KERNEL, cpu_to_node(i));
8826 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8835 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8837 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8838 &cpu_rq(cpu)->leaf_cfs_rq_list);
8841 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8843 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8845 #else /* !CONFG_FAIR_GROUP_SCHED */
8846 static inline void free_fair_sched_group(struct task_group *tg)
8851 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8856 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8860 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8863 #endif /* CONFIG_FAIR_GROUP_SCHED */
8865 #ifdef CONFIG_RT_GROUP_SCHED
8866 static void free_rt_sched_group(struct task_group *tg)
8870 destroy_rt_bandwidth(&tg->rt_bandwidth);
8872 for_each_possible_cpu(i) {
8874 kfree(tg->rt_rq[i]);
8876 kfree(tg->rt_se[i]);
8884 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8886 struct rt_rq *rt_rq;
8887 struct sched_rt_entity *rt_se;
8891 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8894 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8898 init_rt_bandwidth(&tg->rt_bandwidth,
8899 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8901 for_each_possible_cpu(i) {
8904 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8905 GFP_KERNEL, cpu_to_node(i));
8909 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8910 GFP_KERNEL, cpu_to_node(i));
8914 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8923 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8925 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8926 &cpu_rq(cpu)->leaf_rt_rq_list);
8929 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8931 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8933 #else /* !CONFIG_RT_GROUP_SCHED */
8934 static inline void free_rt_sched_group(struct task_group *tg)
8939 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8944 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8948 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8951 #endif /* CONFIG_RT_GROUP_SCHED */
8953 #ifdef CONFIG_GROUP_SCHED
8954 static void free_sched_group(struct task_group *tg)
8956 free_fair_sched_group(tg);
8957 free_rt_sched_group(tg);
8961 /* allocate runqueue etc for a new task group */
8962 struct task_group *sched_create_group(struct task_group *parent)
8964 struct task_group *tg;
8965 unsigned long flags;
8968 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8970 return ERR_PTR(-ENOMEM);
8972 if (!alloc_fair_sched_group(tg, parent))
8975 if (!alloc_rt_sched_group(tg, parent))
8978 spin_lock_irqsave(&task_group_lock, flags);
8979 for_each_possible_cpu(i) {
8980 register_fair_sched_group(tg, i);
8981 register_rt_sched_group(tg, i);
8983 list_add_rcu(&tg->list, &task_groups);
8985 WARN_ON(!parent); /* root should already exist */
8987 tg->parent = parent;
8988 INIT_LIST_HEAD(&tg->children);
8989 list_add_rcu(&tg->siblings, &parent->children);
8990 spin_unlock_irqrestore(&task_group_lock, flags);
8995 free_sched_group(tg);
8996 return ERR_PTR(-ENOMEM);
8999 /* rcu callback to free various structures associated with a task group */
9000 static void free_sched_group_rcu(struct rcu_head *rhp)
9002 /* now it should be safe to free those cfs_rqs */
9003 free_sched_group(container_of(rhp, struct task_group, rcu));
9006 /* Destroy runqueue etc associated with a task group */
9007 void sched_destroy_group(struct task_group *tg)
9009 unsigned long flags;
9012 spin_lock_irqsave(&task_group_lock, flags);
9013 for_each_possible_cpu(i) {
9014 unregister_fair_sched_group(tg, i);
9015 unregister_rt_sched_group(tg, i);
9017 list_del_rcu(&tg->list);
9018 list_del_rcu(&tg->siblings);
9019 spin_unlock_irqrestore(&task_group_lock, flags);
9021 /* wait for possible concurrent references to cfs_rqs complete */
9022 call_rcu(&tg->rcu, free_sched_group_rcu);
9025 /* change task's runqueue when it moves between groups.
9026 * The caller of this function should have put the task in its new group
9027 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9028 * reflect its new group.
9030 void sched_move_task(struct task_struct *tsk)
9033 unsigned long flags;
9036 rq = task_rq_lock(tsk, &flags);
9038 update_rq_clock(rq);
9040 running = task_current(rq, tsk);
9041 on_rq = tsk->se.on_rq;
9044 dequeue_task(rq, tsk, 0);
9045 if (unlikely(running))
9046 tsk->sched_class->put_prev_task(rq, tsk);
9048 set_task_rq(tsk, task_cpu(tsk));
9050 #ifdef CONFIG_FAIR_GROUP_SCHED
9051 if (tsk->sched_class->moved_group)
9052 tsk->sched_class->moved_group(tsk);
9055 if (unlikely(running))
9056 tsk->sched_class->set_curr_task(rq);
9058 enqueue_task(rq, tsk, 0);
9060 task_rq_unlock(rq, &flags);
9062 #endif /* CONFIG_GROUP_SCHED */
9064 #ifdef CONFIG_FAIR_GROUP_SCHED
9065 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9067 struct cfs_rq *cfs_rq = se->cfs_rq;
9072 dequeue_entity(cfs_rq, se, 0);
9074 se->load.weight = shares;
9075 se->load.inv_weight = 0;
9078 enqueue_entity(cfs_rq, se, 0);
9081 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9083 struct cfs_rq *cfs_rq = se->cfs_rq;
9084 struct rq *rq = cfs_rq->rq;
9085 unsigned long flags;
9087 spin_lock_irqsave(&rq->lock, flags);
9088 __set_se_shares(se, shares);
9089 spin_unlock_irqrestore(&rq->lock, flags);
9092 static DEFINE_MUTEX(shares_mutex);
9094 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9097 unsigned long flags;
9100 * We can't change the weight of the root cgroup.
9105 if (shares < MIN_SHARES)
9106 shares = MIN_SHARES;
9107 else if (shares > MAX_SHARES)
9108 shares = MAX_SHARES;
9110 mutex_lock(&shares_mutex);
9111 if (tg->shares == shares)
9114 spin_lock_irqsave(&task_group_lock, flags);
9115 for_each_possible_cpu(i)
9116 unregister_fair_sched_group(tg, i);
9117 list_del_rcu(&tg->siblings);
9118 spin_unlock_irqrestore(&task_group_lock, flags);
9120 /* wait for any ongoing reference to this group to finish */
9121 synchronize_sched();
9124 * Now we are free to modify the group's share on each cpu
9125 * w/o tripping rebalance_share or load_balance_fair.
9127 tg->shares = shares;
9128 for_each_possible_cpu(i) {
9132 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9133 set_se_shares(tg->se[i], shares);
9137 * Enable load balance activity on this group, by inserting it back on
9138 * each cpu's rq->leaf_cfs_rq_list.
9140 spin_lock_irqsave(&task_group_lock, flags);
9141 for_each_possible_cpu(i)
9142 register_fair_sched_group(tg, i);
9143 list_add_rcu(&tg->siblings, &tg->parent->children);
9144 spin_unlock_irqrestore(&task_group_lock, flags);
9146 mutex_unlock(&shares_mutex);
9150 unsigned long sched_group_shares(struct task_group *tg)
9156 #ifdef CONFIG_RT_GROUP_SCHED
9158 * Ensure that the real time constraints are schedulable.
9160 static DEFINE_MUTEX(rt_constraints_mutex);
9162 static unsigned long to_ratio(u64 period, u64 runtime)
9164 if (runtime == RUNTIME_INF)
9167 return div64_u64(runtime << 20, period);
9170 /* Must be called with tasklist_lock held */
9171 static inline int tg_has_rt_tasks(struct task_group *tg)
9173 struct task_struct *g, *p;
9175 do_each_thread(g, p) {
9176 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9178 } while_each_thread(g, p);
9183 struct rt_schedulable_data {
9184 struct task_group *tg;
9189 static int tg_schedulable(struct task_group *tg, void *data)
9191 struct rt_schedulable_data *d = data;
9192 struct task_group *child;
9193 unsigned long total, sum = 0;
9194 u64 period, runtime;
9196 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9197 runtime = tg->rt_bandwidth.rt_runtime;
9200 period = d->rt_period;
9201 runtime = d->rt_runtime;
9204 #ifdef CONFIG_USER_SCHED
9205 if (tg == &root_task_group) {
9206 period = global_rt_period();
9207 runtime = global_rt_runtime();
9212 * Cannot have more runtime than the period.
9214 if (runtime > period && runtime != RUNTIME_INF)
9218 * Ensure we don't starve existing RT tasks.
9220 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9223 total = to_ratio(period, runtime);
9226 * Nobody can have more than the global setting allows.
9228 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9232 * The sum of our children's runtime should not exceed our own.
9234 list_for_each_entry_rcu(child, &tg->children, siblings) {
9235 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9236 runtime = child->rt_bandwidth.rt_runtime;
9238 if (child == d->tg) {
9239 period = d->rt_period;
9240 runtime = d->rt_runtime;
9243 sum += to_ratio(period, runtime);
9252 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9254 struct rt_schedulable_data data = {
9256 .rt_period = period,
9257 .rt_runtime = runtime,
9260 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9263 static int tg_set_bandwidth(struct task_group *tg,
9264 u64 rt_period, u64 rt_runtime)
9268 mutex_lock(&rt_constraints_mutex);
9269 read_lock(&tasklist_lock);
9270 err = __rt_schedulable(tg, rt_period, rt_runtime);
9274 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9275 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9276 tg->rt_bandwidth.rt_runtime = rt_runtime;
9278 for_each_possible_cpu(i) {
9279 struct rt_rq *rt_rq = tg->rt_rq[i];
9281 spin_lock(&rt_rq->rt_runtime_lock);
9282 rt_rq->rt_runtime = rt_runtime;
9283 spin_unlock(&rt_rq->rt_runtime_lock);
9285 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9287 read_unlock(&tasklist_lock);
9288 mutex_unlock(&rt_constraints_mutex);
9293 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9295 u64 rt_runtime, rt_period;
9297 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9298 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9299 if (rt_runtime_us < 0)
9300 rt_runtime = RUNTIME_INF;
9302 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9305 long sched_group_rt_runtime(struct task_group *tg)
9309 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9312 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9313 do_div(rt_runtime_us, NSEC_PER_USEC);
9314 return rt_runtime_us;
9317 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9319 u64 rt_runtime, rt_period;
9321 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9322 rt_runtime = tg->rt_bandwidth.rt_runtime;
9327 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9330 long sched_group_rt_period(struct task_group *tg)
9334 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9335 do_div(rt_period_us, NSEC_PER_USEC);
9336 return rt_period_us;
9339 static int sched_rt_global_constraints(void)
9341 u64 runtime, period;
9344 if (sysctl_sched_rt_period <= 0)
9347 runtime = global_rt_runtime();
9348 period = global_rt_period();
9351 * Sanity check on the sysctl variables.
9353 if (runtime > period && runtime != RUNTIME_INF)
9356 mutex_lock(&rt_constraints_mutex);
9357 read_lock(&tasklist_lock);
9358 ret = __rt_schedulable(NULL, 0, 0);
9359 read_unlock(&tasklist_lock);
9360 mutex_unlock(&rt_constraints_mutex);
9365 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9367 /* Don't accept realtime tasks when there is no way for them to run */
9368 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9374 #else /* !CONFIG_RT_GROUP_SCHED */
9375 static int sched_rt_global_constraints(void)
9377 unsigned long flags;
9380 if (sysctl_sched_rt_period <= 0)
9383 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9384 for_each_possible_cpu(i) {
9385 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9387 spin_lock(&rt_rq->rt_runtime_lock);
9388 rt_rq->rt_runtime = global_rt_runtime();
9389 spin_unlock(&rt_rq->rt_runtime_lock);
9391 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9395 #endif /* CONFIG_RT_GROUP_SCHED */
9397 int sched_rt_handler(struct ctl_table *table, int write,
9398 struct file *filp, void __user *buffer, size_t *lenp,
9402 int old_period, old_runtime;
9403 static DEFINE_MUTEX(mutex);
9406 old_period = sysctl_sched_rt_period;
9407 old_runtime = sysctl_sched_rt_runtime;
9409 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9411 if (!ret && write) {
9412 ret = sched_rt_global_constraints();
9414 sysctl_sched_rt_period = old_period;
9415 sysctl_sched_rt_runtime = old_runtime;
9417 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9418 def_rt_bandwidth.rt_period =
9419 ns_to_ktime(global_rt_period());
9422 mutex_unlock(&mutex);
9427 #ifdef CONFIG_CGROUP_SCHED
9429 /* return corresponding task_group object of a cgroup */
9430 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9432 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9433 struct task_group, css);
9436 static struct cgroup_subsys_state *
9437 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9439 struct task_group *tg, *parent;
9441 if (!cgrp->parent) {
9442 /* This is early initialization for the top cgroup */
9443 return &init_task_group.css;
9446 parent = cgroup_tg(cgrp->parent);
9447 tg = sched_create_group(parent);
9449 return ERR_PTR(-ENOMEM);
9455 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9457 struct task_group *tg = cgroup_tg(cgrp);
9459 sched_destroy_group(tg);
9463 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9464 struct task_struct *tsk)
9466 #ifdef CONFIG_RT_GROUP_SCHED
9467 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9470 /* We don't support RT-tasks being in separate groups */
9471 if (tsk->sched_class != &fair_sched_class)
9479 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9480 struct cgroup *old_cont, struct task_struct *tsk)
9482 sched_move_task(tsk);
9485 #ifdef CONFIG_FAIR_GROUP_SCHED
9486 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9489 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9492 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9494 struct task_group *tg = cgroup_tg(cgrp);
9496 return (u64) tg->shares;
9498 #endif /* CONFIG_FAIR_GROUP_SCHED */
9500 #ifdef CONFIG_RT_GROUP_SCHED
9501 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9504 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9507 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9509 return sched_group_rt_runtime(cgroup_tg(cgrp));
9512 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9515 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9518 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9520 return sched_group_rt_period(cgroup_tg(cgrp));
9522 #endif /* CONFIG_RT_GROUP_SCHED */
9524 static struct cftype cpu_files[] = {
9525 #ifdef CONFIG_FAIR_GROUP_SCHED
9528 .read_u64 = cpu_shares_read_u64,
9529 .write_u64 = cpu_shares_write_u64,
9532 #ifdef CONFIG_RT_GROUP_SCHED
9534 .name = "rt_runtime_us",
9535 .read_s64 = cpu_rt_runtime_read,
9536 .write_s64 = cpu_rt_runtime_write,
9539 .name = "rt_period_us",
9540 .read_u64 = cpu_rt_period_read_uint,
9541 .write_u64 = cpu_rt_period_write_uint,
9546 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9548 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9551 struct cgroup_subsys cpu_cgroup_subsys = {
9553 .create = cpu_cgroup_create,
9554 .destroy = cpu_cgroup_destroy,
9555 .can_attach = cpu_cgroup_can_attach,
9556 .attach = cpu_cgroup_attach,
9557 .populate = cpu_cgroup_populate,
9558 .subsys_id = cpu_cgroup_subsys_id,
9562 #endif /* CONFIG_CGROUP_SCHED */
9564 #ifdef CONFIG_CGROUP_CPUACCT
9567 * CPU accounting code for task groups.
9569 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9570 * (balbir@in.ibm.com).
9573 /* track cpu usage of a group of tasks and its child groups */
9575 struct cgroup_subsys_state css;
9576 /* cpuusage holds pointer to a u64-type object on every cpu */
9578 struct cpuacct *parent;
9581 struct cgroup_subsys cpuacct_subsys;
9583 /* return cpu accounting group corresponding to this container */
9584 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9586 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9587 struct cpuacct, css);
9590 /* return cpu accounting group to which this task belongs */
9591 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9593 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9594 struct cpuacct, css);
9597 /* create a new cpu accounting group */
9598 static struct cgroup_subsys_state *cpuacct_create(
9599 struct cgroup_subsys *ss, struct cgroup *cgrp)
9601 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9604 return ERR_PTR(-ENOMEM);
9606 ca->cpuusage = alloc_percpu(u64);
9607 if (!ca->cpuusage) {
9609 return ERR_PTR(-ENOMEM);
9613 ca->parent = cgroup_ca(cgrp->parent);
9618 /* destroy an existing cpu accounting group */
9620 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9622 struct cpuacct *ca = cgroup_ca(cgrp);
9624 free_percpu(ca->cpuusage);
9628 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9630 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9633 #ifndef CONFIG_64BIT
9635 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9637 spin_lock_irq(&cpu_rq(cpu)->lock);
9639 spin_unlock_irq(&cpu_rq(cpu)->lock);
9647 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9649 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9651 #ifndef CONFIG_64BIT
9653 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9655 spin_lock_irq(&cpu_rq(cpu)->lock);
9657 spin_unlock_irq(&cpu_rq(cpu)->lock);
9663 /* return total cpu usage (in nanoseconds) of a group */
9664 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9666 struct cpuacct *ca = cgroup_ca(cgrp);
9667 u64 totalcpuusage = 0;
9670 for_each_present_cpu(i)
9671 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9673 return totalcpuusage;
9676 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9679 struct cpuacct *ca = cgroup_ca(cgrp);
9688 for_each_present_cpu(i)
9689 cpuacct_cpuusage_write(ca, i, 0);
9695 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9698 struct cpuacct *ca = cgroup_ca(cgroup);
9702 for_each_present_cpu(i) {
9703 percpu = cpuacct_cpuusage_read(ca, i);
9704 seq_printf(m, "%llu ", (unsigned long long) percpu);
9706 seq_printf(m, "\n");
9710 static struct cftype files[] = {
9713 .read_u64 = cpuusage_read,
9714 .write_u64 = cpuusage_write,
9717 .name = "usage_percpu",
9718 .read_seq_string = cpuacct_percpu_seq_read,
9723 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9725 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9729 * charge this task's execution time to its accounting group.
9731 * called with rq->lock held.
9733 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9738 if (unlikely(!cpuacct_subsys.active))
9741 cpu = task_cpu(tsk);
9744 for (; ca; ca = ca->parent) {
9745 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9746 *cpuusage += cputime;
9750 struct cgroup_subsys cpuacct_subsys = {
9752 .create = cpuacct_create,
9753 .destroy = cpuacct_destroy,
9754 .populate = cpuacct_populate,
9755 .subsys_id = cpuacct_subsys_id,
9757 #endif /* CONFIG_CGROUP_CPUACCT */