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_exp_empty;
642 unsigned int yld_act_empty;
643 unsigned int yld_both_empty;
644 unsigned int yld_count;
646 /* schedule() stats */
647 unsigned int sched_switch;
648 unsigned int sched_count;
649 unsigned int sched_goidle;
651 /* try_to_wake_up() stats */
652 unsigned int ttwu_count;
653 unsigned int ttwu_local;
656 unsigned int bkl_count;
660 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
662 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
664 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
667 static inline int cpu_of(struct rq *rq)
677 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
678 * See detach_destroy_domains: synchronize_sched for details.
680 * The domain tree of any CPU may only be accessed from within
681 * preempt-disabled sections.
683 #define for_each_domain(cpu, __sd) \
684 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
686 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
687 #define this_rq() (&__get_cpu_var(runqueues))
688 #define task_rq(p) cpu_rq(task_cpu(p))
689 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
691 static inline void update_rq_clock(struct rq *rq)
693 rq->clock = sched_clock_cpu(cpu_of(rq));
697 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
699 #ifdef CONFIG_SCHED_DEBUG
700 # define const_debug __read_mostly
702 # define const_debug static const
708 * Returns true if the current cpu runqueue is locked.
709 * This interface allows printk to be called with the runqueue lock
710 * held and know whether or not it is OK to wake up the klogd.
712 int runqueue_is_locked(void)
715 struct rq *rq = cpu_rq(cpu);
718 ret = spin_is_locked(&rq->lock);
724 * Debugging: various feature bits
727 #define SCHED_FEAT(name, enabled) \
728 __SCHED_FEAT_##name ,
731 #include "sched_features.h"
736 #define SCHED_FEAT(name, enabled) \
737 (1UL << __SCHED_FEAT_##name) * enabled |
739 const_debug unsigned int sysctl_sched_features =
740 #include "sched_features.h"
745 #ifdef CONFIG_SCHED_DEBUG
746 #define SCHED_FEAT(name, enabled) \
749 static __read_mostly char *sched_feat_names[] = {
750 #include "sched_features.h"
756 static int sched_feat_show(struct seq_file *m, void *v)
760 for (i = 0; sched_feat_names[i]; i++) {
761 if (!(sysctl_sched_features & (1UL << i)))
763 seq_printf(m, "%s ", sched_feat_names[i]);
771 sched_feat_write(struct file *filp, const char __user *ubuf,
772 size_t cnt, loff_t *ppos)
782 if (copy_from_user(&buf, ubuf, cnt))
787 if (strncmp(buf, "NO_", 3) == 0) {
792 for (i = 0; sched_feat_names[i]; i++) {
793 int len = strlen(sched_feat_names[i]);
795 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
797 sysctl_sched_features &= ~(1UL << i);
799 sysctl_sched_features |= (1UL << i);
804 if (!sched_feat_names[i])
812 static int sched_feat_open(struct inode *inode, struct file *filp)
814 return single_open(filp, sched_feat_show, NULL);
817 static struct file_operations sched_feat_fops = {
818 .open = sched_feat_open,
819 .write = sched_feat_write,
822 .release = single_release,
825 static __init int sched_init_debug(void)
827 debugfs_create_file("sched_features", 0644, NULL, NULL,
832 late_initcall(sched_init_debug);
836 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
839 * Number of tasks to iterate in a single balance run.
840 * Limited because this is done with IRQs disabled.
842 const_debug unsigned int sysctl_sched_nr_migrate = 32;
845 * ratelimit for updating the group shares.
848 unsigned int sysctl_sched_shares_ratelimit = 250000;
851 * Inject some fuzzyness into changing the per-cpu group shares
852 * this avoids remote rq-locks at the expense of fairness.
855 unsigned int sysctl_sched_shares_thresh = 4;
858 * period over which we measure -rt task cpu usage in us.
861 unsigned int sysctl_sched_rt_period = 1000000;
863 static __read_mostly int scheduler_running;
866 * part of the period that we allow rt tasks to run in us.
869 int sysctl_sched_rt_runtime = 950000;
871 static inline u64 global_rt_period(void)
873 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
876 static inline u64 global_rt_runtime(void)
878 if (sysctl_sched_rt_runtime < 0)
881 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
884 #ifndef prepare_arch_switch
885 # define prepare_arch_switch(next) do { } while (0)
887 #ifndef finish_arch_switch
888 # define finish_arch_switch(prev) do { } while (0)
891 static inline int task_current(struct rq *rq, struct task_struct *p)
893 return rq->curr == p;
896 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
897 static inline int task_running(struct rq *rq, struct task_struct *p)
899 return task_current(rq, p);
902 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
906 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
908 #ifdef CONFIG_DEBUG_SPINLOCK
909 /* this is a valid case when another task releases the spinlock */
910 rq->lock.owner = current;
913 * If we are tracking spinlock dependencies then we have to
914 * fix up the runqueue lock - which gets 'carried over' from
917 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
919 spin_unlock_irq(&rq->lock);
922 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
923 static inline int task_running(struct rq *rq, struct task_struct *p)
928 return task_current(rq, p);
932 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
936 * We can optimise this out completely for !SMP, because the
937 * SMP rebalancing from interrupt is the only thing that cares
942 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
943 spin_unlock_irq(&rq->lock);
945 spin_unlock(&rq->lock);
949 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
953 * After ->oncpu is cleared, the task can be moved to a different CPU.
954 * We must ensure this doesn't happen until the switch is completely
960 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
964 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
967 * __task_rq_lock - lock the runqueue a given task resides on.
968 * Must be called interrupts disabled.
970 static inline struct rq *__task_rq_lock(struct task_struct *p)
974 struct rq *rq = task_rq(p);
975 spin_lock(&rq->lock);
976 if (likely(rq == task_rq(p)))
978 spin_unlock(&rq->lock);
983 * task_rq_lock - lock the runqueue a given task resides on and disable
984 * interrupts. Note the ordering: we can safely lookup the task_rq without
985 * explicitly disabling preemption.
987 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
993 local_irq_save(*flags);
995 spin_lock(&rq->lock);
996 if (likely(rq == task_rq(p)))
998 spin_unlock_irqrestore(&rq->lock, *flags);
1002 void task_rq_unlock_wait(struct task_struct *p)
1004 struct rq *rq = task_rq(p);
1006 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1007 spin_unlock_wait(&rq->lock);
1010 static void __task_rq_unlock(struct rq *rq)
1011 __releases(rq->lock)
1013 spin_unlock(&rq->lock);
1016 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1017 __releases(rq->lock)
1019 spin_unlock_irqrestore(&rq->lock, *flags);
1023 * this_rq_lock - lock this runqueue and disable interrupts.
1025 static struct rq *this_rq_lock(void)
1026 __acquires(rq->lock)
1030 local_irq_disable();
1032 spin_lock(&rq->lock);
1037 #ifdef CONFIG_SCHED_HRTICK
1039 * Use HR-timers to deliver accurate preemption points.
1041 * Its all a bit involved since we cannot program an hrt while holding the
1042 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1045 * When we get rescheduled we reprogram the hrtick_timer outside of the
1051 * - enabled by features
1052 * - hrtimer is actually high res
1054 static inline int hrtick_enabled(struct rq *rq)
1056 if (!sched_feat(HRTICK))
1058 if (!cpu_active(cpu_of(rq)))
1060 return hrtimer_is_hres_active(&rq->hrtick_timer);
1063 static void hrtick_clear(struct rq *rq)
1065 if (hrtimer_active(&rq->hrtick_timer))
1066 hrtimer_cancel(&rq->hrtick_timer);
1070 * High-resolution timer tick.
1071 * Runs from hardirq context with interrupts disabled.
1073 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1075 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1077 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1079 spin_lock(&rq->lock);
1080 update_rq_clock(rq);
1081 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1082 spin_unlock(&rq->lock);
1084 return HRTIMER_NORESTART;
1089 * called from hardirq (IPI) context
1091 static void __hrtick_start(void *arg)
1093 struct rq *rq = arg;
1095 spin_lock(&rq->lock);
1096 hrtimer_restart(&rq->hrtick_timer);
1097 rq->hrtick_csd_pending = 0;
1098 spin_unlock(&rq->lock);
1102 * Called to set the hrtick timer state.
1104 * called with rq->lock held and irqs disabled
1106 static void hrtick_start(struct rq *rq, u64 delay)
1108 struct hrtimer *timer = &rq->hrtick_timer;
1109 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1111 hrtimer_set_expires(timer, time);
1113 if (rq == this_rq()) {
1114 hrtimer_restart(timer);
1115 } else if (!rq->hrtick_csd_pending) {
1116 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1117 rq->hrtick_csd_pending = 1;
1122 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1124 int cpu = (int)(long)hcpu;
1127 case CPU_UP_CANCELED:
1128 case CPU_UP_CANCELED_FROZEN:
1129 case CPU_DOWN_PREPARE:
1130 case CPU_DOWN_PREPARE_FROZEN:
1132 case CPU_DEAD_FROZEN:
1133 hrtick_clear(cpu_rq(cpu));
1140 static __init void init_hrtick(void)
1142 hotcpu_notifier(hotplug_hrtick, 0);
1146 * Called to set the hrtick timer state.
1148 * called with rq->lock held and irqs disabled
1150 static void hrtick_start(struct rq *rq, u64 delay)
1152 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1155 static inline void init_hrtick(void)
1158 #endif /* CONFIG_SMP */
1160 static void init_rq_hrtick(struct rq *rq)
1163 rq->hrtick_csd_pending = 0;
1165 rq->hrtick_csd.flags = 0;
1166 rq->hrtick_csd.func = __hrtick_start;
1167 rq->hrtick_csd.info = rq;
1170 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1171 rq->hrtick_timer.function = hrtick;
1173 #else /* CONFIG_SCHED_HRTICK */
1174 static inline void hrtick_clear(struct rq *rq)
1178 static inline void init_rq_hrtick(struct rq *rq)
1182 static inline void init_hrtick(void)
1185 #endif /* CONFIG_SCHED_HRTICK */
1188 * resched_task - mark a task 'to be rescheduled now'.
1190 * On UP this means the setting of the need_resched flag, on SMP it
1191 * might also involve a cross-CPU call to trigger the scheduler on
1196 #ifndef tsk_is_polling
1197 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1200 static void resched_task(struct task_struct *p)
1204 assert_spin_locked(&task_rq(p)->lock);
1206 if (test_tsk_need_resched(p))
1209 set_tsk_need_resched(p);
1212 if (cpu == smp_processor_id())
1215 /* NEED_RESCHED must be visible before we test polling */
1217 if (!tsk_is_polling(p))
1218 smp_send_reschedule(cpu);
1221 static void resched_cpu(int cpu)
1223 struct rq *rq = cpu_rq(cpu);
1224 unsigned long flags;
1226 if (!spin_trylock_irqsave(&rq->lock, flags))
1228 resched_task(cpu_curr(cpu));
1229 spin_unlock_irqrestore(&rq->lock, flags);
1234 * When add_timer_on() enqueues a timer into the timer wheel of an
1235 * idle CPU then this timer might expire before the next timer event
1236 * which is scheduled to wake up that CPU. In case of a completely
1237 * idle system the next event might even be infinite time into the
1238 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1239 * leaves the inner idle loop so the newly added timer is taken into
1240 * account when the CPU goes back to idle and evaluates the timer
1241 * wheel for the next timer event.
1243 void wake_up_idle_cpu(int cpu)
1245 struct rq *rq = cpu_rq(cpu);
1247 if (cpu == smp_processor_id())
1251 * This is safe, as this function is called with the timer
1252 * wheel base lock of (cpu) held. When the CPU is on the way
1253 * to idle and has not yet set rq->curr to idle then it will
1254 * be serialized on the timer wheel base lock and take the new
1255 * timer into account automatically.
1257 if (rq->curr != rq->idle)
1261 * We can set TIF_RESCHED on the idle task of the other CPU
1262 * lockless. The worst case is that the other CPU runs the
1263 * idle task through an additional NOOP schedule()
1265 set_tsk_need_resched(rq->idle);
1267 /* NEED_RESCHED must be visible before we test polling */
1269 if (!tsk_is_polling(rq->idle))
1270 smp_send_reschedule(cpu);
1272 #endif /* CONFIG_NO_HZ */
1274 #else /* !CONFIG_SMP */
1275 static void resched_task(struct task_struct *p)
1277 assert_spin_locked(&task_rq(p)->lock);
1278 set_tsk_need_resched(p);
1280 #endif /* CONFIG_SMP */
1282 #if BITS_PER_LONG == 32
1283 # define WMULT_CONST (~0UL)
1285 # define WMULT_CONST (1UL << 32)
1288 #define WMULT_SHIFT 32
1291 * Shift right and round:
1293 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1296 * delta *= weight / lw
1298 static unsigned long
1299 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1300 struct load_weight *lw)
1304 if (!lw->inv_weight) {
1305 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1308 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1312 tmp = (u64)delta_exec * weight;
1314 * Check whether we'd overflow the 64-bit multiplication:
1316 if (unlikely(tmp > WMULT_CONST))
1317 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1320 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1322 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1325 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1331 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1338 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1339 * of tasks with abnormal "nice" values across CPUs the contribution that
1340 * each task makes to its run queue's load is weighted according to its
1341 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1342 * scaled version of the new time slice allocation that they receive on time
1346 #define WEIGHT_IDLEPRIO 3
1347 #define WMULT_IDLEPRIO 1431655765
1350 * Nice levels are multiplicative, with a gentle 10% change for every
1351 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1352 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1353 * that remained on nice 0.
1355 * The "10% effect" is relative and cumulative: from _any_ nice level,
1356 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1357 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1358 * If a task goes up by ~10% and another task goes down by ~10% then
1359 * the relative distance between them is ~25%.)
1361 static const int prio_to_weight[40] = {
1362 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1363 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1364 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1365 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1366 /* 0 */ 1024, 820, 655, 526, 423,
1367 /* 5 */ 335, 272, 215, 172, 137,
1368 /* 10 */ 110, 87, 70, 56, 45,
1369 /* 15 */ 36, 29, 23, 18, 15,
1373 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1375 * In cases where the weight does not change often, we can use the
1376 * precalculated inverse to speed up arithmetics by turning divisions
1377 * into multiplications:
1379 static const u32 prio_to_wmult[40] = {
1380 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1381 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1382 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1383 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1384 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1385 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1386 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1387 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1390 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1393 * runqueue iterator, to support SMP load-balancing between different
1394 * scheduling classes, without having to expose their internal data
1395 * structures to the load-balancing proper:
1397 struct rq_iterator {
1399 struct task_struct *(*start)(void *);
1400 struct task_struct *(*next)(void *);
1404 static unsigned long
1405 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1406 unsigned long max_load_move, struct sched_domain *sd,
1407 enum cpu_idle_type idle, int *all_pinned,
1408 int *this_best_prio, struct rq_iterator *iterator);
1411 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1412 struct sched_domain *sd, enum cpu_idle_type idle,
1413 struct rq_iterator *iterator);
1416 #ifdef CONFIG_CGROUP_CPUACCT
1417 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1419 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1422 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1424 update_load_add(&rq->load, load);
1427 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1429 update_load_sub(&rq->load, load);
1432 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1433 typedef int (*tg_visitor)(struct task_group *, void *);
1436 * Iterate the full tree, calling @down when first entering a node and @up when
1437 * leaving it for the final time.
1439 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1441 struct task_group *parent, *child;
1445 parent = &root_task_group;
1447 ret = (*down)(parent, data);
1450 list_for_each_entry_rcu(child, &parent->children, siblings) {
1457 ret = (*up)(parent, data);
1462 parent = parent->parent;
1471 static int tg_nop(struct task_group *tg, void *data)
1478 static unsigned long source_load(int cpu, int type);
1479 static unsigned long target_load(int cpu, int type);
1480 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1482 static unsigned long cpu_avg_load_per_task(int cpu)
1484 struct rq *rq = cpu_rq(cpu);
1485 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1488 rq->avg_load_per_task = rq->load.weight / nr_running;
1490 rq->avg_load_per_task = 0;
1492 return rq->avg_load_per_task;
1495 #ifdef CONFIG_FAIR_GROUP_SCHED
1497 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1500 * Calculate and set the cpu's group shares.
1503 update_group_shares_cpu(struct task_group *tg, int cpu,
1504 unsigned long sd_shares, unsigned long sd_rq_weight)
1506 unsigned long shares;
1507 unsigned long rq_weight;
1512 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1515 * \Sum shares * rq_weight
1516 * shares = -----------------------
1520 shares = (sd_shares * rq_weight) / sd_rq_weight;
1521 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1523 if (abs(shares - tg->se[cpu]->load.weight) >
1524 sysctl_sched_shares_thresh) {
1525 struct rq *rq = cpu_rq(cpu);
1526 unsigned long flags;
1528 spin_lock_irqsave(&rq->lock, flags);
1529 tg->cfs_rq[cpu]->shares = shares;
1531 __set_se_shares(tg->se[cpu], shares);
1532 spin_unlock_irqrestore(&rq->lock, flags);
1537 * Re-compute the task group their per cpu shares over the given domain.
1538 * This needs to be done in a bottom-up fashion because the rq weight of a
1539 * parent group depends on the shares of its child groups.
1541 static int tg_shares_up(struct task_group *tg, void *data)
1543 unsigned long weight, rq_weight = 0;
1544 unsigned long shares = 0;
1545 struct sched_domain *sd = data;
1548 for_each_cpu(i, sched_domain_span(sd)) {
1550 * If there are currently no tasks on the cpu pretend there
1551 * is one of average load so that when a new task gets to
1552 * run here it will not get delayed by group starvation.
1554 weight = tg->cfs_rq[i]->load.weight;
1556 weight = NICE_0_LOAD;
1558 tg->cfs_rq[i]->rq_weight = weight;
1559 rq_weight += weight;
1560 shares += tg->cfs_rq[i]->shares;
1563 if ((!shares && rq_weight) || shares > tg->shares)
1564 shares = tg->shares;
1566 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1567 shares = tg->shares;
1569 for_each_cpu(i, sched_domain_span(sd))
1570 update_group_shares_cpu(tg, i, shares, rq_weight);
1576 * Compute the cpu's hierarchical load factor for each task group.
1577 * This needs to be done in a top-down fashion because the load of a child
1578 * group is a fraction of its parents load.
1580 static int tg_load_down(struct task_group *tg, void *data)
1583 long cpu = (long)data;
1586 load = cpu_rq(cpu)->load.weight;
1588 load = tg->parent->cfs_rq[cpu]->h_load;
1589 load *= tg->cfs_rq[cpu]->shares;
1590 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1593 tg->cfs_rq[cpu]->h_load = load;
1598 static void update_shares(struct sched_domain *sd)
1600 u64 now = cpu_clock(raw_smp_processor_id());
1601 s64 elapsed = now - sd->last_update;
1603 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1604 sd->last_update = now;
1605 walk_tg_tree(tg_nop, tg_shares_up, sd);
1609 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1611 spin_unlock(&rq->lock);
1613 spin_lock(&rq->lock);
1616 static void update_h_load(long cpu)
1618 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1623 static inline void update_shares(struct sched_domain *sd)
1627 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1633 #ifdef CONFIG_PREEMPT
1636 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1637 * way at the expense of forcing extra atomic operations in all
1638 * invocations. This assures that the double_lock is acquired using the
1639 * same underlying policy as the spinlock_t on this architecture, which
1640 * reduces latency compared to the unfair variant below. However, it
1641 * also adds more overhead and therefore may reduce throughput.
1643 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1644 __releases(this_rq->lock)
1645 __acquires(busiest->lock)
1646 __acquires(this_rq->lock)
1648 spin_unlock(&this_rq->lock);
1649 double_rq_lock(this_rq, busiest);
1656 * Unfair double_lock_balance: Optimizes throughput at the expense of
1657 * latency by eliminating extra atomic operations when the locks are
1658 * already in proper order on entry. This favors lower cpu-ids and will
1659 * grant the double lock to lower cpus over higher ids under contention,
1660 * regardless of entry order into the function.
1662 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1663 __releases(this_rq->lock)
1664 __acquires(busiest->lock)
1665 __acquires(this_rq->lock)
1669 if (unlikely(!spin_trylock(&busiest->lock))) {
1670 if (busiest < this_rq) {
1671 spin_unlock(&this_rq->lock);
1672 spin_lock(&busiest->lock);
1673 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1676 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1681 #endif /* CONFIG_PREEMPT */
1684 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1686 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1688 if (unlikely(!irqs_disabled())) {
1689 /* printk() doesn't work good under rq->lock */
1690 spin_unlock(&this_rq->lock);
1694 return _double_lock_balance(this_rq, busiest);
1697 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1698 __releases(busiest->lock)
1700 spin_unlock(&busiest->lock);
1701 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1705 #ifdef CONFIG_FAIR_GROUP_SCHED
1706 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1709 cfs_rq->shares = shares;
1714 #include "sched_stats.h"
1715 #include "sched_idletask.c"
1716 #include "sched_fair.c"
1717 #include "sched_rt.c"
1718 #ifdef CONFIG_SCHED_DEBUG
1719 # include "sched_debug.c"
1722 #define sched_class_highest (&rt_sched_class)
1723 #define for_each_class(class) \
1724 for (class = sched_class_highest; class; class = class->next)
1726 static void inc_nr_running(struct rq *rq)
1731 static void dec_nr_running(struct rq *rq)
1736 static void set_load_weight(struct task_struct *p)
1738 if (task_has_rt_policy(p)) {
1739 p->se.load.weight = prio_to_weight[0] * 2;
1740 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1745 * SCHED_IDLE tasks get minimal weight:
1747 if (p->policy == SCHED_IDLE) {
1748 p->se.load.weight = WEIGHT_IDLEPRIO;
1749 p->se.load.inv_weight = WMULT_IDLEPRIO;
1753 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1754 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1757 static void update_avg(u64 *avg, u64 sample)
1759 s64 diff = sample - *avg;
1763 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1766 p->se.start_runtime = p->se.sum_exec_runtime;
1768 sched_info_queued(p);
1769 p->sched_class->enqueue_task(rq, p, wakeup);
1773 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1776 if (p->se.last_wakeup) {
1777 update_avg(&p->se.avg_overlap,
1778 p->se.sum_exec_runtime - p->se.last_wakeup);
1779 p->se.last_wakeup = 0;
1781 update_avg(&p->se.avg_wakeup,
1782 sysctl_sched_wakeup_granularity);
1786 sched_info_dequeued(p);
1787 p->sched_class->dequeue_task(rq, p, sleep);
1792 * __normal_prio - return the priority that is based on the static prio
1794 static inline int __normal_prio(struct task_struct *p)
1796 return p->static_prio;
1800 * Calculate the expected normal priority: i.e. priority
1801 * without taking RT-inheritance into account. Might be
1802 * boosted by interactivity modifiers. Changes upon fork,
1803 * setprio syscalls, and whenever the interactivity
1804 * estimator recalculates.
1806 static inline int normal_prio(struct task_struct *p)
1810 if (task_has_rt_policy(p))
1811 prio = MAX_RT_PRIO-1 - p->rt_priority;
1813 prio = __normal_prio(p);
1818 * Calculate the current priority, i.e. the priority
1819 * taken into account by the scheduler. This value might
1820 * be boosted by RT tasks, or might be boosted by
1821 * interactivity modifiers. Will be RT if the task got
1822 * RT-boosted. If not then it returns p->normal_prio.
1824 static int effective_prio(struct task_struct *p)
1826 p->normal_prio = normal_prio(p);
1828 * If we are RT tasks or we were boosted to RT priority,
1829 * keep the priority unchanged. Otherwise, update priority
1830 * to the normal priority:
1832 if (!rt_prio(p->prio))
1833 return p->normal_prio;
1838 * activate_task - move a task to the runqueue.
1840 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1842 if (task_contributes_to_load(p))
1843 rq->nr_uninterruptible--;
1845 enqueue_task(rq, p, wakeup);
1850 * deactivate_task - remove a task from the runqueue.
1852 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1854 if (task_contributes_to_load(p))
1855 rq->nr_uninterruptible++;
1857 dequeue_task(rq, p, sleep);
1862 * task_curr - is this task currently executing on a CPU?
1863 * @p: the task in question.
1865 inline int task_curr(const struct task_struct *p)
1867 return cpu_curr(task_cpu(p)) == p;
1870 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1872 set_task_rq(p, cpu);
1875 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1876 * successfuly executed on another CPU. We must ensure that updates of
1877 * per-task data have been completed by this moment.
1880 task_thread_info(p)->cpu = cpu;
1884 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1885 const struct sched_class *prev_class,
1886 int oldprio, int running)
1888 if (prev_class != p->sched_class) {
1889 if (prev_class->switched_from)
1890 prev_class->switched_from(rq, p, running);
1891 p->sched_class->switched_to(rq, p, running);
1893 p->sched_class->prio_changed(rq, p, oldprio, running);
1898 /* Used instead of source_load when we know the type == 0 */
1899 static unsigned long weighted_cpuload(const int cpu)
1901 return cpu_rq(cpu)->load.weight;
1905 * Is this task likely cache-hot:
1908 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1913 * Buddy candidates are cache hot:
1915 if (sched_feat(CACHE_HOT_BUDDY) &&
1916 (&p->se == cfs_rq_of(&p->se)->next ||
1917 &p->se == cfs_rq_of(&p->se)->last))
1920 if (p->sched_class != &fair_sched_class)
1923 if (sysctl_sched_migration_cost == -1)
1925 if (sysctl_sched_migration_cost == 0)
1928 delta = now - p->se.exec_start;
1930 return delta < (s64)sysctl_sched_migration_cost;
1934 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1936 int old_cpu = task_cpu(p);
1937 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1938 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1939 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1942 clock_offset = old_rq->clock - new_rq->clock;
1944 trace_sched_migrate_task(p, task_cpu(p), new_cpu);
1946 #ifdef CONFIG_SCHEDSTATS
1947 if (p->se.wait_start)
1948 p->se.wait_start -= clock_offset;
1949 if (p->se.sleep_start)
1950 p->se.sleep_start -= clock_offset;
1951 if (p->se.block_start)
1952 p->se.block_start -= clock_offset;
1953 if (old_cpu != new_cpu) {
1954 schedstat_inc(p, se.nr_migrations);
1955 if (task_hot(p, old_rq->clock, NULL))
1956 schedstat_inc(p, se.nr_forced2_migrations);
1959 p->se.vruntime -= old_cfsrq->min_vruntime -
1960 new_cfsrq->min_vruntime;
1962 __set_task_cpu(p, new_cpu);
1965 struct migration_req {
1966 struct list_head list;
1968 struct task_struct *task;
1971 struct completion done;
1975 * The task's runqueue lock must be held.
1976 * Returns true if you have to wait for migration thread.
1979 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1981 struct rq *rq = task_rq(p);
1984 * If the task is not on a runqueue (and not running), then
1985 * it is sufficient to simply update the task's cpu field.
1987 if (!p->se.on_rq && !task_running(rq, p)) {
1988 set_task_cpu(p, dest_cpu);
1992 init_completion(&req->done);
1994 req->dest_cpu = dest_cpu;
1995 list_add(&req->list, &rq->migration_queue);
2001 * wait_task_inactive - wait for a thread to unschedule.
2003 * If @match_state is nonzero, it's the @p->state value just checked and
2004 * not expected to change. If it changes, i.e. @p might have woken up,
2005 * then return zero. When we succeed in waiting for @p to be off its CPU,
2006 * we return a positive number (its total switch count). If a second call
2007 * a short while later returns the same number, the caller can be sure that
2008 * @p has remained unscheduled the whole time.
2010 * The caller must ensure that the task *will* unschedule sometime soon,
2011 * else this function might spin for a *long* time. This function can't
2012 * be called with interrupts off, or it may introduce deadlock with
2013 * smp_call_function() if an IPI is sent by the same process we are
2014 * waiting to become inactive.
2016 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2018 unsigned long flags;
2025 * We do the initial early heuristics without holding
2026 * any task-queue locks at all. We'll only try to get
2027 * the runqueue lock when things look like they will
2033 * If the task is actively running on another CPU
2034 * still, just relax and busy-wait without holding
2037 * NOTE! Since we don't hold any locks, it's not
2038 * even sure that "rq" stays as the right runqueue!
2039 * But we don't care, since "task_running()" will
2040 * return false if the runqueue has changed and p
2041 * is actually now running somewhere else!
2043 while (task_running(rq, p)) {
2044 if (match_state && unlikely(p->state != match_state))
2050 * Ok, time to look more closely! We need the rq
2051 * lock now, to be *sure*. If we're wrong, we'll
2052 * just go back and repeat.
2054 rq = task_rq_lock(p, &flags);
2055 trace_sched_wait_task(rq, p);
2056 running = task_running(rq, p);
2057 on_rq = p->se.on_rq;
2059 if (!match_state || p->state == match_state)
2060 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2061 task_rq_unlock(rq, &flags);
2064 * If it changed from the expected state, bail out now.
2066 if (unlikely(!ncsw))
2070 * Was it really running after all now that we
2071 * checked with the proper locks actually held?
2073 * Oops. Go back and try again..
2075 if (unlikely(running)) {
2081 * It's not enough that it's not actively running,
2082 * it must be off the runqueue _entirely_, and not
2085 * So if it was still runnable (but just not actively
2086 * running right now), it's preempted, and we should
2087 * yield - it could be a while.
2089 if (unlikely(on_rq)) {
2090 schedule_timeout_uninterruptible(1);
2095 * Ahh, all good. It wasn't running, and it wasn't
2096 * runnable, which means that it will never become
2097 * running in the future either. We're all done!
2106 * kick_process - kick a running thread to enter/exit the kernel
2107 * @p: the to-be-kicked thread
2109 * Cause a process which is running on another CPU to enter
2110 * kernel-mode, without any delay. (to get signals handled.)
2112 * NOTE: this function doesnt have to take the runqueue lock,
2113 * because all it wants to ensure is that the remote task enters
2114 * the kernel. If the IPI races and the task has been migrated
2115 * to another CPU then no harm is done and the purpose has been
2118 void kick_process(struct task_struct *p)
2124 if ((cpu != smp_processor_id()) && task_curr(p))
2125 smp_send_reschedule(cpu);
2130 * Return a low guess at the load of a migration-source cpu weighted
2131 * according to the scheduling class and "nice" value.
2133 * We want to under-estimate the load of migration sources, to
2134 * balance conservatively.
2136 static unsigned long source_load(int cpu, int type)
2138 struct rq *rq = cpu_rq(cpu);
2139 unsigned long total = weighted_cpuload(cpu);
2141 if (type == 0 || !sched_feat(LB_BIAS))
2144 return min(rq->cpu_load[type-1], total);
2148 * Return a high guess at the load of a migration-target cpu weighted
2149 * according to the scheduling class and "nice" value.
2151 static unsigned long target_load(int cpu, int type)
2153 struct rq *rq = cpu_rq(cpu);
2154 unsigned long total = weighted_cpuload(cpu);
2156 if (type == 0 || !sched_feat(LB_BIAS))
2159 return max(rq->cpu_load[type-1], total);
2163 * find_idlest_group finds and returns the least busy CPU group within the
2166 static struct sched_group *
2167 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2169 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2170 unsigned long min_load = ULONG_MAX, this_load = 0;
2171 int load_idx = sd->forkexec_idx;
2172 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2175 unsigned long load, avg_load;
2179 /* Skip over this group if it has no CPUs allowed */
2180 if (!cpumask_intersects(sched_group_cpus(group),
2184 local_group = cpumask_test_cpu(this_cpu,
2185 sched_group_cpus(group));
2187 /* Tally up the load of all CPUs in the group */
2190 for_each_cpu(i, sched_group_cpus(group)) {
2191 /* Bias balancing toward cpus of our domain */
2193 load = source_load(i, load_idx);
2195 load = target_load(i, load_idx);
2200 /* Adjust by relative CPU power of the group */
2201 avg_load = sg_div_cpu_power(group,
2202 avg_load * SCHED_LOAD_SCALE);
2205 this_load = avg_load;
2207 } else if (avg_load < min_load) {
2208 min_load = avg_load;
2211 } while (group = group->next, group != sd->groups);
2213 if (!idlest || 100*this_load < imbalance*min_load)
2219 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2222 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2224 unsigned long load, min_load = ULONG_MAX;
2228 /* Traverse only the allowed CPUs */
2229 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2230 load = weighted_cpuload(i);
2232 if (load < min_load || (load == min_load && i == this_cpu)) {
2242 * sched_balance_self: balance the current task (running on cpu) in domains
2243 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2246 * Balance, ie. select the least loaded group.
2248 * Returns the target CPU number, or the same CPU if no balancing is needed.
2250 * preempt must be disabled.
2252 static int sched_balance_self(int cpu, int flag)
2254 struct task_struct *t = current;
2255 struct sched_domain *tmp, *sd = NULL;
2257 for_each_domain(cpu, tmp) {
2259 * If power savings logic is enabled for a domain, stop there.
2261 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2263 if (tmp->flags & flag)
2271 struct sched_group *group;
2272 int new_cpu, weight;
2274 if (!(sd->flags & flag)) {
2279 group = find_idlest_group(sd, t, cpu);
2285 new_cpu = find_idlest_cpu(group, t, cpu);
2286 if (new_cpu == -1 || new_cpu == cpu) {
2287 /* Now try balancing at a lower domain level of cpu */
2292 /* Now try balancing at a lower domain level of new_cpu */
2294 weight = cpumask_weight(sched_domain_span(sd));
2296 for_each_domain(cpu, tmp) {
2297 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2299 if (tmp->flags & flag)
2302 /* while loop will break here if sd == NULL */
2308 #endif /* CONFIG_SMP */
2311 * try_to_wake_up - wake up a thread
2312 * @p: the to-be-woken-up thread
2313 * @state: the mask of task states that can be woken
2314 * @sync: do a synchronous wakeup?
2316 * Put it on the run-queue if it's not already there. The "current"
2317 * thread is always on the run-queue (except when the actual
2318 * re-schedule is in progress), and as such you're allowed to do
2319 * the simpler "current->state = TASK_RUNNING" to mark yourself
2320 * runnable without the overhead of this.
2322 * returns failure only if the task is already active.
2324 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2326 int cpu, orig_cpu, this_cpu, success = 0;
2327 unsigned long flags;
2331 if (!sched_feat(SYNC_WAKEUPS))
2335 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2336 struct sched_domain *sd;
2338 this_cpu = raw_smp_processor_id();
2341 for_each_domain(this_cpu, sd) {
2342 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2351 rq = task_rq_lock(p, &flags);
2352 update_rq_clock(rq);
2353 old_state = p->state;
2354 if (!(old_state & state))
2362 this_cpu = smp_processor_id();
2365 if (unlikely(task_running(rq, p)))
2368 cpu = p->sched_class->select_task_rq(p, sync);
2369 if (cpu != orig_cpu) {
2370 set_task_cpu(p, cpu);
2371 task_rq_unlock(rq, &flags);
2372 /* might preempt at this point */
2373 rq = task_rq_lock(p, &flags);
2374 old_state = p->state;
2375 if (!(old_state & state))
2380 this_cpu = smp_processor_id();
2384 #ifdef CONFIG_SCHEDSTATS
2385 schedstat_inc(rq, ttwu_count);
2386 if (cpu == this_cpu)
2387 schedstat_inc(rq, ttwu_local);
2389 struct sched_domain *sd;
2390 for_each_domain(this_cpu, sd) {
2391 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2392 schedstat_inc(sd, ttwu_wake_remote);
2397 #endif /* CONFIG_SCHEDSTATS */
2400 #endif /* CONFIG_SMP */
2401 schedstat_inc(p, se.nr_wakeups);
2403 schedstat_inc(p, se.nr_wakeups_sync);
2404 if (orig_cpu != cpu)
2405 schedstat_inc(p, se.nr_wakeups_migrate);
2406 if (cpu == this_cpu)
2407 schedstat_inc(p, se.nr_wakeups_local);
2409 schedstat_inc(p, se.nr_wakeups_remote);
2410 activate_task(rq, p, 1);
2414 * Only attribute actual wakeups done by this task.
2416 if (!in_interrupt()) {
2417 struct sched_entity *se = ¤t->se;
2418 u64 sample = se->sum_exec_runtime;
2420 if (se->last_wakeup)
2421 sample -= se->last_wakeup;
2423 sample -= se->start_runtime;
2424 update_avg(&se->avg_wakeup, sample);
2426 se->last_wakeup = se->sum_exec_runtime;
2430 trace_sched_wakeup(rq, p, success);
2431 check_preempt_curr(rq, p, sync);
2433 p->state = TASK_RUNNING;
2435 if (p->sched_class->task_wake_up)
2436 p->sched_class->task_wake_up(rq, p);
2439 task_rq_unlock(rq, &flags);
2444 int wake_up_process(struct task_struct *p)
2446 return try_to_wake_up(p, TASK_ALL, 0);
2448 EXPORT_SYMBOL(wake_up_process);
2450 int wake_up_state(struct task_struct *p, unsigned int state)
2452 return try_to_wake_up(p, state, 0);
2456 * Perform scheduler related setup for a newly forked process p.
2457 * p is forked by current.
2459 * __sched_fork() is basic setup used by init_idle() too:
2461 static void __sched_fork(struct task_struct *p)
2463 p->se.exec_start = 0;
2464 p->se.sum_exec_runtime = 0;
2465 p->se.prev_sum_exec_runtime = 0;
2466 p->se.last_wakeup = 0;
2467 p->se.avg_overlap = 0;
2468 p->se.start_runtime = 0;
2469 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2471 #ifdef CONFIG_SCHEDSTATS
2472 p->se.wait_start = 0;
2473 p->se.sum_sleep_runtime = 0;
2474 p->se.sleep_start = 0;
2475 p->se.block_start = 0;
2476 p->se.sleep_max = 0;
2477 p->se.block_max = 0;
2479 p->se.slice_max = 0;
2483 INIT_LIST_HEAD(&p->rt.run_list);
2485 INIT_LIST_HEAD(&p->se.group_node);
2487 #ifdef CONFIG_PREEMPT_NOTIFIERS
2488 INIT_HLIST_HEAD(&p->preempt_notifiers);
2492 * We mark the process as running here, but have not actually
2493 * inserted it onto the runqueue yet. This guarantees that
2494 * nobody will actually run it, and a signal or other external
2495 * event cannot wake it up and insert it on the runqueue either.
2497 p->state = TASK_RUNNING;
2501 * fork()/clone()-time setup:
2503 void sched_fork(struct task_struct *p, int clone_flags)
2505 int cpu = get_cpu();
2510 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2512 set_task_cpu(p, cpu);
2515 * Make sure we do not leak PI boosting priority to the child:
2517 p->prio = current->normal_prio;
2518 if (!rt_prio(p->prio))
2519 p->sched_class = &fair_sched_class;
2521 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2522 if (likely(sched_info_on()))
2523 memset(&p->sched_info, 0, sizeof(p->sched_info));
2525 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2528 #ifdef CONFIG_PREEMPT
2529 /* Want to start with kernel preemption disabled. */
2530 task_thread_info(p)->preempt_count = 1;
2532 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2538 * wake_up_new_task - wake up a newly created task for the first time.
2540 * This function will do some initial scheduler statistics housekeeping
2541 * that must be done for every newly created context, then puts the task
2542 * on the runqueue and wakes it.
2544 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2546 unsigned long flags;
2549 rq = task_rq_lock(p, &flags);
2550 BUG_ON(p->state != TASK_RUNNING);
2551 update_rq_clock(rq);
2553 p->prio = effective_prio(p);
2555 if (!p->sched_class->task_new || !current->se.on_rq) {
2556 activate_task(rq, p, 0);
2559 * Let the scheduling class do new task startup
2560 * management (if any):
2562 p->sched_class->task_new(rq, p);
2565 trace_sched_wakeup_new(rq, p, 1);
2566 check_preempt_curr(rq, p, 0);
2568 if (p->sched_class->task_wake_up)
2569 p->sched_class->task_wake_up(rq, p);
2571 task_rq_unlock(rq, &flags);
2574 #ifdef CONFIG_PREEMPT_NOTIFIERS
2577 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2578 * @notifier: notifier struct to register
2580 void preempt_notifier_register(struct preempt_notifier *notifier)
2582 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2584 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2587 * preempt_notifier_unregister - no longer interested in preemption notifications
2588 * @notifier: notifier struct to unregister
2590 * This is safe to call from within a preemption notifier.
2592 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2594 hlist_del(¬ifier->link);
2596 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2598 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2600 struct preempt_notifier *notifier;
2601 struct hlist_node *node;
2603 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2604 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2608 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2609 struct task_struct *next)
2611 struct preempt_notifier *notifier;
2612 struct hlist_node *node;
2614 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2615 notifier->ops->sched_out(notifier, next);
2618 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2620 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2625 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2626 struct task_struct *next)
2630 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2633 * prepare_task_switch - prepare to switch tasks
2634 * @rq: the runqueue preparing to switch
2635 * @prev: the current task that is being switched out
2636 * @next: the task we are going to switch to.
2638 * This is called with the rq lock held and interrupts off. It must
2639 * be paired with a subsequent finish_task_switch after the context
2642 * prepare_task_switch sets up locking and calls architecture specific
2646 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2647 struct task_struct *next)
2649 fire_sched_out_preempt_notifiers(prev, next);
2650 prepare_lock_switch(rq, next);
2651 prepare_arch_switch(next);
2655 * finish_task_switch - clean up after a task-switch
2656 * @rq: runqueue associated with task-switch
2657 * @prev: the thread we just switched away from.
2659 * finish_task_switch must be called after the context switch, paired
2660 * with a prepare_task_switch call before the context switch.
2661 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2662 * and do any other architecture-specific cleanup actions.
2664 * Note that we may have delayed dropping an mm in context_switch(). If
2665 * so, we finish that here outside of the runqueue lock. (Doing it
2666 * with the lock held can cause deadlocks; see schedule() for
2669 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2670 __releases(rq->lock)
2672 struct mm_struct *mm = rq->prev_mm;
2675 int post_schedule = 0;
2677 if (current->sched_class->needs_post_schedule)
2678 post_schedule = current->sched_class->needs_post_schedule(rq);
2684 * A task struct has one reference for the use as "current".
2685 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2686 * schedule one last time. The schedule call will never return, and
2687 * the scheduled task must drop that reference.
2688 * The test for TASK_DEAD must occur while the runqueue locks are
2689 * still held, otherwise prev could be scheduled on another cpu, die
2690 * there before we look at prev->state, and then the reference would
2692 * Manfred Spraul <manfred@colorfullife.com>
2694 prev_state = prev->state;
2695 finish_arch_switch(prev);
2696 finish_lock_switch(rq, prev);
2699 current->sched_class->post_schedule(rq);
2702 fire_sched_in_preempt_notifiers(current);
2705 if (unlikely(prev_state == TASK_DEAD)) {
2707 * Remove function-return probe instances associated with this
2708 * task and put them back on the free list.
2710 kprobe_flush_task(prev);
2711 put_task_struct(prev);
2716 * schedule_tail - first thing a freshly forked thread must call.
2717 * @prev: the thread we just switched away from.
2719 asmlinkage void schedule_tail(struct task_struct *prev)
2720 __releases(rq->lock)
2722 struct rq *rq = this_rq();
2724 finish_task_switch(rq, prev);
2725 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2726 /* In this case, finish_task_switch does not reenable preemption */
2729 if (current->set_child_tid)
2730 put_user(task_pid_vnr(current), current->set_child_tid);
2734 * context_switch - switch to the new MM and the new
2735 * thread's register state.
2738 context_switch(struct rq *rq, struct task_struct *prev,
2739 struct task_struct *next)
2741 struct mm_struct *mm, *oldmm;
2743 prepare_task_switch(rq, prev, next);
2744 trace_sched_switch(rq, prev, next);
2746 oldmm = prev->active_mm;
2748 * For paravirt, this is coupled with an exit in switch_to to
2749 * combine the page table reload and the switch backend into
2752 arch_enter_lazy_cpu_mode();
2754 if (unlikely(!mm)) {
2755 next->active_mm = oldmm;
2756 atomic_inc(&oldmm->mm_count);
2757 enter_lazy_tlb(oldmm, next);
2759 switch_mm(oldmm, mm, next);
2761 if (unlikely(!prev->mm)) {
2762 prev->active_mm = NULL;
2763 rq->prev_mm = oldmm;
2766 * Since the runqueue lock will be released by the next
2767 * task (which is an invalid locking op but in the case
2768 * of the scheduler it's an obvious special-case), so we
2769 * do an early lockdep release here:
2771 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2772 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2775 /* Here we just switch the register state and the stack. */
2776 switch_to(prev, next, prev);
2780 * this_rq must be evaluated again because prev may have moved
2781 * CPUs since it called schedule(), thus the 'rq' on its stack
2782 * frame will be invalid.
2784 finish_task_switch(this_rq(), prev);
2788 * nr_running, nr_uninterruptible and nr_context_switches:
2790 * externally visible scheduler statistics: current number of runnable
2791 * threads, current number of uninterruptible-sleeping threads, total
2792 * number of context switches performed since bootup.
2794 unsigned long nr_running(void)
2796 unsigned long i, sum = 0;
2798 for_each_online_cpu(i)
2799 sum += cpu_rq(i)->nr_running;
2804 unsigned long nr_uninterruptible(void)
2806 unsigned long i, sum = 0;
2808 for_each_possible_cpu(i)
2809 sum += cpu_rq(i)->nr_uninterruptible;
2812 * Since we read the counters lockless, it might be slightly
2813 * inaccurate. Do not allow it to go below zero though:
2815 if (unlikely((long)sum < 0))
2821 unsigned long long nr_context_switches(void)
2824 unsigned long long sum = 0;
2826 for_each_possible_cpu(i)
2827 sum += cpu_rq(i)->nr_switches;
2832 unsigned long nr_iowait(void)
2834 unsigned long i, sum = 0;
2836 for_each_possible_cpu(i)
2837 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2842 unsigned long nr_active(void)
2844 unsigned long i, running = 0, uninterruptible = 0;
2846 for_each_online_cpu(i) {
2847 running += cpu_rq(i)->nr_running;
2848 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2851 if (unlikely((long)uninterruptible < 0))
2852 uninterruptible = 0;
2854 return running + uninterruptible;
2858 * Update rq->cpu_load[] statistics. This function is usually called every
2859 * scheduler tick (TICK_NSEC).
2861 static void update_cpu_load(struct rq *this_rq)
2863 unsigned long this_load = this_rq->load.weight;
2866 this_rq->nr_load_updates++;
2868 /* Update our load: */
2869 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2870 unsigned long old_load, new_load;
2872 /* scale is effectively 1 << i now, and >> i divides by scale */
2874 old_load = this_rq->cpu_load[i];
2875 new_load = this_load;
2877 * Round up the averaging division if load is increasing. This
2878 * prevents us from getting stuck on 9 if the load is 10, for
2881 if (new_load > old_load)
2882 new_load += scale-1;
2883 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2890 * double_rq_lock - safely lock two runqueues
2892 * Note this does not disable interrupts like task_rq_lock,
2893 * you need to do so manually before calling.
2895 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2896 __acquires(rq1->lock)
2897 __acquires(rq2->lock)
2899 BUG_ON(!irqs_disabled());
2901 spin_lock(&rq1->lock);
2902 __acquire(rq2->lock); /* Fake it out ;) */
2905 spin_lock(&rq1->lock);
2906 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2908 spin_lock(&rq2->lock);
2909 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2912 update_rq_clock(rq1);
2913 update_rq_clock(rq2);
2917 * double_rq_unlock - safely unlock two runqueues
2919 * Note this does not restore interrupts like task_rq_unlock,
2920 * you need to do so manually after calling.
2922 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2923 __releases(rq1->lock)
2924 __releases(rq2->lock)
2926 spin_unlock(&rq1->lock);
2928 spin_unlock(&rq2->lock);
2930 __release(rq2->lock);
2934 * If dest_cpu is allowed for this process, migrate the task to it.
2935 * This is accomplished by forcing the cpu_allowed mask to only
2936 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2937 * the cpu_allowed mask is restored.
2939 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2941 struct migration_req req;
2942 unsigned long flags;
2945 rq = task_rq_lock(p, &flags);
2946 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
2947 || unlikely(!cpu_active(dest_cpu)))
2950 /* force the process onto the specified CPU */
2951 if (migrate_task(p, dest_cpu, &req)) {
2952 /* Need to wait for migration thread (might exit: take ref). */
2953 struct task_struct *mt = rq->migration_thread;
2955 get_task_struct(mt);
2956 task_rq_unlock(rq, &flags);
2957 wake_up_process(mt);
2958 put_task_struct(mt);
2959 wait_for_completion(&req.done);
2964 task_rq_unlock(rq, &flags);
2968 * sched_exec - execve() is a valuable balancing opportunity, because at
2969 * this point the task has the smallest effective memory and cache footprint.
2971 void sched_exec(void)
2973 int new_cpu, this_cpu = get_cpu();
2974 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2976 if (new_cpu != this_cpu)
2977 sched_migrate_task(current, new_cpu);
2981 * pull_task - move a task from a remote runqueue to the local runqueue.
2982 * Both runqueues must be locked.
2984 static void pull_task(struct rq *src_rq, struct task_struct *p,
2985 struct rq *this_rq, int this_cpu)
2987 deactivate_task(src_rq, p, 0);
2988 set_task_cpu(p, this_cpu);
2989 activate_task(this_rq, p, 0);
2991 * Note that idle threads have a prio of MAX_PRIO, for this test
2992 * to be always true for them.
2994 check_preempt_curr(this_rq, p, 0);
2998 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3001 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3002 struct sched_domain *sd, enum cpu_idle_type idle,
3005 int tsk_cache_hot = 0;
3007 * We do not migrate tasks that are:
3008 * 1) running (obviously), or
3009 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3010 * 3) are cache-hot on their current CPU.
3012 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3013 schedstat_inc(p, se.nr_failed_migrations_affine);
3018 if (task_running(rq, p)) {
3019 schedstat_inc(p, se.nr_failed_migrations_running);
3024 * Aggressive migration if:
3025 * 1) task is cache cold, or
3026 * 2) too many balance attempts have failed.
3029 tsk_cache_hot = task_hot(p, rq->clock, sd);
3030 if (!tsk_cache_hot ||
3031 sd->nr_balance_failed > sd->cache_nice_tries) {
3032 #ifdef CONFIG_SCHEDSTATS
3033 if (tsk_cache_hot) {
3034 schedstat_inc(sd, lb_hot_gained[idle]);
3035 schedstat_inc(p, se.nr_forced_migrations);
3041 if (tsk_cache_hot) {
3042 schedstat_inc(p, se.nr_failed_migrations_hot);
3048 static unsigned long
3049 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3050 unsigned long max_load_move, struct sched_domain *sd,
3051 enum cpu_idle_type idle, int *all_pinned,
3052 int *this_best_prio, struct rq_iterator *iterator)
3054 int loops = 0, pulled = 0, pinned = 0;
3055 struct task_struct *p;
3056 long rem_load_move = max_load_move;
3058 if (max_load_move == 0)
3064 * Start the load-balancing iterator:
3066 p = iterator->start(iterator->arg);
3068 if (!p || loops++ > sysctl_sched_nr_migrate)
3071 if ((p->se.load.weight >> 1) > rem_load_move ||
3072 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3073 p = iterator->next(iterator->arg);
3077 pull_task(busiest, p, this_rq, this_cpu);
3079 rem_load_move -= p->se.load.weight;
3081 #ifdef CONFIG_PREEMPT
3083 * NEWIDLE balancing is a source of latency, so preemptible kernels
3084 * will stop after the first task is pulled to minimize the critical
3087 if (idle == CPU_NEWLY_IDLE)
3092 * We only want to steal up to the prescribed amount of weighted load.
3094 if (rem_load_move > 0) {
3095 if (p->prio < *this_best_prio)
3096 *this_best_prio = p->prio;
3097 p = iterator->next(iterator->arg);
3102 * Right now, this is one of only two places pull_task() is called,
3103 * so we can safely collect pull_task() stats here rather than
3104 * inside pull_task().
3106 schedstat_add(sd, lb_gained[idle], pulled);
3109 *all_pinned = pinned;
3111 return max_load_move - rem_load_move;
3115 * move_tasks tries to move up to max_load_move weighted load from busiest to
3116 * this_rq, as part of a balancing operation within domain "sd".
3117 * Returns 1 if successful and 0 otherwise.
3119 * Called with both runqueues locked.
3121 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3122 unsigned long max_load_move,
3123 struct sched_domain *sd, enum cpu_idle_type idle,
3126 const struct sched_class *class = sched_class_highest;
3127 unsigned long total_load_moved = 0;
3128 int this_best_prio = this_rq->curr->prio;
3132 class->load_balance(this_rq, this_cpu, busiest,
3133 max_load_move - total_load_moved,
3134 sd, idle, all_pinned, &this_best_prio);
3135 class = class->next;
3137 #ifdef CONFIG_PREEMPT
3139 * NEWIDLE balancing is a source of latency, so preemptible
3140 * kernels will stop after the first task is pulled to minimize
3141 * the critical section.
3143 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3146 } while (class && max_load_move > total_load_moved);
3148 return total_load_moved > 0;
3152 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3153 struct sched_domain *sd, enum cpu_idle_type idle,
3154 struct rq_iterator *iterator)
3156 struct task_struct *p = iterator->start(iterator->arg);
3160 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3161 pull_task(busiest, p, this_rq, this_cpu);
3163 * Right now, this is only the second place pull_task()
3164 * is called, so we can safely collect pull_task()
3165 * stats here rather than inside pull_task().
3167 schedstat_inc(sd, lb_gained[idle]);
3171 p = iterator->next(iterator->arg);
3178 * move_one_task tries to move exactly one task from busiest to this_rq, as
3179 * part of active balancing operations within "domain".
3180 * Returns 1 if successful and 0 otherwise.
3182 * Called with both runqueues locked.
3184 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3185 struct sched_domain *sd, enum cpu_idle_type idle)
3187 const struct sched_class *class;
3189 for (class = sched_class_highest; class; class = class->next)
3190 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3197 * find_busiest_group finds and returns the busiest CPU group within the
3198 * domain. It calculates and returns the amount of weighted load which
3199 * should be moved to restore balance via the imbalance parameter.
3201 static struct sched_group *
3202 find_busiest_group(struct sched_domain *sd, int this_cpu,
3203 unsigned long *imbalance, enum cpu_idle_type idle,
3204 int *sd_idle, const struct cpumask *cpus, int *balance)
3206 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3207 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3208 unsigned long max_pull;
3209 unsigned long busiest_load_per_task, busiest_nr_running;
3210 unsigned long this_load_per_task, this_nr_running;
3211 int load_idx, group_imb = 0;
3212 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3213 int power_savings_balance = 1;
3214 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3215 unsigned long min_nr_running = ULONG_MAX;
3216 struct sched_group *group_min = NULL, *group_leader = NULL;
3219 max_load = this_load = total_load = total_pwr = 0;
3220 busiest_load_per_task = busiest_nr_running = 0;
3221 this_load_per_task = this_nr_running = 0;
3223 if (idle == CPU_NOT_IDLE)
3224 load_idx = sd->busy_idx;
3225 else if (idle == CPU_NEWLY_IDLE)
3226 load_idx = sd->newidle_idx;
3228 load_idx = sd->idle_idx;
3231 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3234 int __group_imb = 0;
3235 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3236 unsigned long sum_nr_running, sum_weighted_load;
3237 unsigned long sum_avg_load_per_task;
3238 unsigned long avg_load_per_task;
3240 local_group = cpumask_test_cpu(this_cpu,
3241 sched_group_cpus(group));
3244 balance_cpu = cpumask_first(sched_group_cpus(group));
3246 /* Tally up the load of all CPUs in the group */
3247 sum_weighted_load = sum_nr_running = avg_load = 0;
3248 sum_avg_load_per_task = avg_load_per_task = 0;
3251 min_cpu_load = ~0UL;
3253 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3254 struct rq *rq = cpu_rq(i);
3256 if (*sd_idle && rq->nr_running)
3259 /* Bias balancing toward cpus of our domain */
3261 if (idle_cpu(i) && !first_idle_cpu) {
3266 load = target_load(i, load_idx);
3268 load = source_load(i, load_idx);
3269 if (load > max_cpu_load)
3270 max_cpu_load = load;
3271 if (min_cpu_load > load)
3272 min_cpu_load = load;
3276 sum_nr_running += rq->nr_running;
3277 sum_weighted_load += weighted_cpuload(i);
3279 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3283 * First idle cpu or the first cpu(busiest) in this sched group
3284 * is eligible for doing load balancing at this and above
3285 * domains. In the newly idle case, we will allow all the cpu's
3286 * to do the newly idle load balance.
3288 if (idle != CPU_NEWLY_IDLE && local_group &&
3289 balance_cpu != this_cpu && balance) {
3294 total_load += avg_load;
3295 total_pwr += group->__cpu_power;
3297 /* Adjust by relative CPU power of the group */
3298 avg_load = sg_div_cpu_power(group,
3299 avg_load * SCHED_LOAD_SCALE);
3303 * Consider the group unbalanced when the imbalance is larger
3304 * than the average weight of two tasks.
3306 * APZ: with cgroup the avg task weight can vary wildly and
3307 * might not be a suitable number - should we keep a
3308 * normalized nr_running number somewhere that negates
3311 avg_load_per_task = sg_div_cpu_power(group,
3312 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3314 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3317 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3320 this_load = avg_load;
3322 this_nr_running = sum_nr_running;
3323 this_load_per_task = sum_weighted_load;
3324 } else if (avg_load > max_load &&
3325 (sum_nr_running > group_capacity || __group_imb)) {
3326 max_load = avg_load;
3328 busiest_nr_running = sum_nr_running;
3329 busiest_load_per_task = sum_weighted_load;
3330 group_imb = __group_imb;
3333 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3335 * Busy processors will not participate in power savings
3338 if (idle == CPU_NOT_IDLE ||
3339 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3343 * If the local group is idle or completely loaded
3344 * no need to do power savings balance at this domain
3346 if (local_group && (this_nr_running >= group_capacity ||
3348 power_savings_balance = 0;
3351 * If a group is already running at full capacity or idle,
3352 * don't include that group in power savings calculations
3354 if (!power_savings_balance || sum_nr_running >= group_capacity
3359 * Calculate the group which has the least non-idle load.
3360 * This is the group from where we need to pick up the load
3363 if ((sum_nr_running < min_nr_running) ||
3364 (sum_nr_running == min_nr_running &&
3365 cpumask_first(sched_group_cpus(group)) >
3366 cpumask_first(sched_group_cpus(group_min)))) {
3368 min_nr_running = sum_nr_running;
3369 min_load_per_task = sum_weighted_load /
3374 * Calculate the group which is almost near its
3375 * capacity but still has some space to pick up some load
3376 * from other group and save more power
3378 if (sum_nr_running <= group_capacity - 1) {
3379 if (sum_nr_running > leader_nr_running ||
3380 (sum_nr_running == leader_nr_running &&
3381 cpumask_first(sched_group_cpus(group)) <
3382 cpumask_first(sched_group_cpus(group_leader)))) {
3383 group_leader = group;
3384 leader_nr_running = sum_nr_running;
3389 group = group->next;
3390 } while (group != sd->groups);
3392 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3395 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3397 if (this_load >= avg_load ||
3398 100*max_load <= sd->imbalance_pct*this_load)
3401 busiest_load_per_task /= busiest_nr_running;
3403 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3406 * We're trying to get all the cpus to the average_load, so we don't
3407 * want to push ourselves above the average load, nor do we wish to
3408 * reduce the max loaded cpu below the average load, as either of these
3409 * actions would just result in more rebalancing later, and ping-pong
3410 * tasks around. Thus we look for the minimum possible imbalance.
3411 * Negative imbalances (*we* are more loaded than anyone else) will
3412 * be counted as no imbalance for these purposes -- we can't fix that
3413 * by pulling tasks to us. Be careful of negative numbers as they'll
3414 * appear as very large values with unsigned longs.
3416 if (max_load <= busiest_load_per_task)
3420 * In the presence of smp nice balancing, certain scenarios can have
3421 * max load less than avg load(as we skip the groups at or below
3422 * its cpu_power, while calculating max_load..)
3424 if (max_load < avg_load) {
3426 goto small_imbalance;
3429 /* Don't want to pull so many tasks that a group would go idle */
3430 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3432 /* How much load to actually move to equalise the imbalance */
3433 *imbalance = min(max_pull * busiest->__cpu_power,
3434 (avg_load - this_load) * this->__cpu_power)
3438 * if *imbalance is less than the average load per runnable task
3439 * there is no gaurantee that any tasks will be moved so we'll have
3440 * a think about bumping its value to force at least one task to be
3443 if (*imbalance < busiest_load_per_task) {
3444 unsigned long tmp, pwr_now, pwr_move;
3448 pwr_move = pwr_now = 0;
3450 if (this_nr_running) {
3451 this_load_per_task /= this_nr_running;
3452 if (busiest_load_per_task > this_load_per_task)
3455 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3457 if (max_load - this_load + busiest_load_per_task >=
3458 busiest_load_per_task * imbn) {
3459 *imbalance = busiest_load_per_task;
3464 * OK, we don't have enough imbalance to justify moving tasks,
3465 * however we may be able to increase total CPU power used by
3469 pwr_now += busiest->__cpu_power *
3470 min(busiest_load_per_task, max_load);
3471 pwr_now += this->__cpu_power *
3472 min(this_load_per_task, this_load);
3473 pwr_now /= SCHED_LOAD_SCALE;
3475 /* Amount of load we'd subtract */
3476 tmp = sg_div_cpu_power(busiest,
3477 busiest_load_per_task * SCHED_LOAD_SCALE);
3479 pwr_move += busiest->__cpu_power *
3480 min(busiest_load_per_task, max_load - tmp);
3482 /* Amount of load we'd add */
3483 if (max_load * busiest->__cpu_power <
3484 busiest_load_per_task * SCHED_LOAD_SCALE)
3485 tmp = sg_div_cpu_power(this,
3486 max_load * busiest->__cpu_power);
3488 tmp = sg_div_cpu_power(this,
3489 busiest_load_per_task * SCHED_LOAD_SCALE);
3490 pwr_move += this->__cpu_power *
3491 min(this_load_per_task, this_load + tmp);
3492 pwr_move /= SCHED_LOAD_SCALE;
3494 /* Move if we gain throughput */
3495 if (pwr_move > pwr_now)
3496 *imbalance = busiest_load_per_task;
3502 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3503 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3506 if (this == group_leader && group_leader != group_min) {
3507 *imbalance = min_load_per_task;
3508 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3509 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3510 cpumask_first(sched_group_cpus(group_leader));
3521 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3524 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3525 unsigned long imbalance, const struct cpumask *cpus)
3527 struct rq *busiest = NULL, *rq;
3528 unsigned long max_load = 0;
3531 for_each_cpu(i, sched_group_cpus(group)) {
3534 if (!cpumask_test_cpu(i, cpus))
3538 wl = weighted_cpuload(i);
3540 if (rq->nr_running == 1 && wl > imbalance)
3543 if (wl > max_load) {
3553 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3554 * so long as it is large enough.
3556 #define MAX_PINNED_INTERVAL 512
3559 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3560 * tasks if there is an imbalance.
3562 static int load_balance(int this_cpu, struct rq *this_rq,
3563 struct sched_domain *sd, enum cpu_idle_type idle,
3564 int *balance, struct cpumask *cpus)
3566 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3567 struct sched_group *group;
3568 unsigned long imbalance;
3570 unsigned long flags;
3572 cpumask_setall(cpus);
3575 * When power savings policy is enabled for the parent domain, idle
3576 * sibling can pick up load irrespective of busy siblings. In this case,
3577 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3578 * portraying it as CPU_NOT_IDLE.
3580 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3581 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3584 schedstat_inc(sd, lb_count[idle]);
3588 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3595 schedstat_inc(sd, lb_nobusyg[idle]);
3599 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3601 schedstat_inc(sd, lb_nobusyq[idle]);
3605 BUG_ON(busiest == this_rq);
3607 schedstat_add(sd, lb_imbalance[idle], imbalance);
3610 if (busiest->nr_running > 1) {
3612 * Attempt to move tasks. If find_busiest_group has found
3613 * an imbalance but busiest->nr_running <= 1, the group is
3614 * still unbalanced. ld_moved simply stays zero, so it is
3615 * correctly treated as an imbalance.
3617 local_irq_save(flags);
3618 double_rq_lock(this_rq, busiest);
3619 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3620 imbalance, sd, idle, &all_pinned);
3621 double_rq_unlock(this_rq, busiest);
3622 local_irq_restore(flags);
3625 * some other cpu did the load balance for us.
3627 if (ld_moved && this_cpu != smp_processor_id())
3628 resched_cpu(this_cpu);
3630 /* All tasks on this runqueue were pinned by CPU affinity */
3631 if (unlikely(all_pinned)) {
3632 cpumask_clear_cpu(cpu_of(busiest), cpus);
3633 if (!cpumask_empty(cpus))
3640 schedstat_inc(sd, lb_failed[idle]);
3641 sd->nr_balance_failed++;
3643 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3645 spin_lock_irqsave(&busiest->lock, flags);
3647 /* don't kick the migration_thread, if the curr
3648 * task on busiest cpu can't be moved to this_cpu
3650 if (!cpumask_test_cpu(this_cpu,
3651 &busiest->curr->cpus_allowed)) {
3652 spin_unlock_irqrestore(&busiest->lock, flags);
3654 goto out_one_pinned;
3657 if (!busiest->active_balance) {
3658 busiest->active_balance = 1;
3659 busiest->push_cpu = this_cpu;
3662 spin_unlock_irqrestore(&busiest->lock, flags);
3664 wake_up_process(busiest->migration_thread);
3667 * We've kicked active balancing, reset the failure
3670 sd->nr_balance_failed = sd->cache_nice_tries+1;
3673 sd->nr_balance_failed = 0;
3675 if (likely(!active_balance)) {
3676 /* We were unbalanced, so reset the balancing interval */
3677 sd->balance_interval = sd->min_interval;
3680 * If we've begun active balancing, start to back off. This
3681 * case may not be covered by the all_pinned logic if there
3682 * is only 1 task on the busy runqueue (because we don't call
3685 if (sd->balance_interval < sd->max_interval)
3686 sd->balance_interval *= 2;
3689 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3690 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3696 schedstat_inc(sd, lb_balanced[idle]);
3698 sd->nr_balance_failed = 0;
3701 /* tune up the balancing interval */
3702 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3703 (sd->balance_interval < sd->max_interval))
3704 sd->balance_interval *= 2;
3706 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3707 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3718 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3719 * tasks if there is an imbalance.
3721 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3722 * this_rq is locked.
3725 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3726 struct cpumask *cpus)
3728 struct sched_group *group;
3729 struct rq *busiest = NULL;
3730 unsigned long imbalance;
3735 cpumask_setall(cpus);
3738 * When power savings policy is enabled for the parent domain, idle
3739 * sibling can pick up load irrespective of busy siblings. In this case,
3740 * let the state of idle sibling percolate up as IDLE, instead of
3741 * portraying it as CPU_NOT_IDLE.
3743 if (sd->flags & SD_SHARE_CPUPOWER &&
3744 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3747 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3749 update_shares_locked(this_rq, sd);
3750 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3751 &sd_idle, cpus, NULL);
3753 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3757 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3759 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3763 BUG_ON(busiest == this_rq);
3765 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3768 if (busiest->nr_running > 1) {
3769 /* Attempt to move tasks */
3770 double_lock_balance(this_rq, busiest);
3771 /* this_rq->clock is already updated */
3772 update_rq_clock(busiest);
3773 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3774 imbalance, sd, CPU_NEWLY_IDLE,
3776 double_unlock_balance(this_rq, busiest);
3778 if (unlikely(all_pinned)) {
3779 cpumask_clear_cpu(cpu_of(busiest), cpus);
3780 if (!cpumask_empty(cpus))
3786 int active_balance = 0;
3788 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3789 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3790 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3793 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3796 if (sd->nr_balance_failed++ < 2)
3800 * The only task running in a non-idle cpu can be moved to this
3801 * cpu in an attempt to completely freeup the other CPU
3802 * package. The same method used to move task in load_balance()
3803 * have been extended for load_balance_newidle() to speedup
3804 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
3806 * The package power saving logic comes from
3807 * find_busiest_group(). If there are no imbalance, then
3808 * f_b_g() will return NULL. However when sched_mc={1,2} then
3809 * f_b_g() will select a group from which a running task may be
3810 * pulled to this cpu in order to make the other package idle.
3811 * If there is no opportunity to make a package idle and if
3812 * there are no imbalance, then f_b_g() will return NULL and no
3813 * action will be taken in load_balance_newidle().
3815 * Under normal task pull operation due to imbalance, there
3816 * will be more than one task in the source run queue and
3817 * move_tasks() will succeed. ld_moved will be true and this
3818 * active balance code will not be triggered.
3821 /* Lock busiest in correct order while this_rq is held */
3822 double_lock_balance(this_rq, busiest);
3825 * don't kick the migration_thread, if the curr
3826 * task on busiest cpu can't be moved to this_cpu
3828 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
3829 double_unlock_balance(this_rq, busiest);
3834 if (!busiest->active_balance) {
3835 busiest->active_balance = 1;
3836 busiest->push_cpu = this_cpu;
3840 double_unlock_balance(this_rq, busiest);
3842 * Should not call ttwu while holding a rq->lock
3844 spin_unlock(&this_rq->lock);
3846 wake_up_process(busiest->migration_thread);
3847 spin_lock(&this_rq->lock);
3850 sd->nr_balance_failed = 0;
3852 update_shares_locked(this_rq, sd);
3856 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3857 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3858 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3860 sd->nr_balance_failed = 0;
3866 * idle_balance is called by schedule() if this_cpu is about to become
3867 * idle. Attempts to pull tasks from other CPUs.
3869 static void idle_balance(int this_cpu, struct rq *this_rq)
3871 struct sched_domain *sd;
3872 int pulled_task = 0;
3873 unsigned long next_balance = jiffies + HZ;
3874 cpumask_var_t tmpmask;
3876 if (!alloc_cpumask_var(&tmpmask, GFP_ATOMIC))
3879 for_each_domain(this_cpu, sd) {
3880 unsigned long interval;
3882 if (!(sd->flags & SD_LOAD_BALANCE))
3885 if (sd->flags & SD_BALANCE_NEWIDLE)
3886 /* If we've pulled tasks over stop searching: */
3887 pulled_task = load_balance_newidle(this_cpu, this_rq,
3890 interval = msecs_to_jiffies(sd->balance_interval);
3891 if (time_after(next_balance, sd->last_balance + interval))
3892 next_balance = sd->last_balance + interval;
3896 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3898 * We are going idle. next_balance may be set based on
3899 * a busy processor. So reset next_balance.
3901 this_rq->next_balance = next_balance;
3903 free_cpumask_var(tmpmask);
3907 * active_load_balance is run by migration threads. It pushes running tasks
3908 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3909 * running on each physical CPU where possible, and avoids physical /
3910 * logical imbalances.
3912 * Called with busiest_rq locked.
3914 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3916 int target_cpu = busiest_rq->push_cpu;
3917 struct sched_domain *sd;
3918 struct rq *target_rq;
3920 /* Is there any task to move? */
3921 if (busiest_rq->nr_running <= 1)
3924 target_rq = cpu_rq(target_cpu);
3927 * This condition is "impossible", if it occurs
3928 * we need to fix it. Originally reported by
3929 * Bjorn Helgaas on a 128-cpu setup.
3931 BUG_ON(busiest_rq == target_rq);
3933 /* move a task from busiest_rq to target_rq */
3934 double_lock_balance(busiest_rq, target_rq);
3935 update_rq_clock(busiest_rq);
3936 update_rq_clock(target_rq);
3938 /* Search for an sd spanning us and the target CPU. */
3939 for_each_domain(target_cpu, sd) {
3940 if ((sd->flags & SD_LOAD_BALANCE) &&
3941 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3946 schedstat_inc(sd, alb_count);
3948 if (move_one_task(target_rq, target_cpu, busiest_rq,
3950 schedstat_inc(sd, alb_pushed);
3952 schedstat_inc(sd, alb_failed);
3954 double_unlock_balance(busiest_rq, target_rq);
3959 atomic_t load_balancer;
3960 cpumask_var_t cpu_mask;
3961 } nohz ____cacheline_aligned = {
3962 .load_balancer = ATOMIC_INIT(-1),
3966 * This routine will try to nominate the ilb (idle load balancing)
3967 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3968 * load balancing on behalf of all those cpus. If all the cpus in the system
3969 * go into this tickless mode, then there will be no ilb owner (as there is
3970 * no need for one) and all the cpus will sleep till the next wakeup event
3973 * For the ilb owner, tick is not stopped. And this tick will be used
3974 * for idle load balancing. ilb owner will still be part of
3977 * While stopping the tick, this cpu will become the ilb owner if there
3978 * is no other owner. And will be the owner till that cpu becomes busy
3979 * or if all cpus in the system stop their ticks at which point
3980 * there is no need for ilb owner.
3982 * When the ilb owner becomes busy, it nominates another owner, during the
3983 * next busy scheduler_tick()
3985 int select_nohz_load_balancer(int stop_tick)
3987 int cpu = smp_processor_id();
3990 cpu_rq(cpu)->in_nohz_recently = 1;
3992 if (!cpu_active(cpu)) {
3993 if (atomic_read(&nohz.load_balancer) != cpu)
3997 * If we are going offline and still the leader,
4000 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4006 cpumask_set_cpu(cpu, nohz.cpu_mask);
4008 /* time for ilb owner also to sleep */
4009 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4010 if (atomic_read(&nohz.load_balancer) == cpu)
4011 atomic_set(&nohz.load_balancer, -1);
4015 if (atomic_read(&nohz.load_balancer) == -1) {
4016 /* make me the ilb owner */
4017 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4019 } else if (atomic_read(&nohz.load_balancer) == cpu)
4022 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4025 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4027 if (atomic_read(&nohz.load_balancer) == cpu)
4028 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4035 static DEFINE_SPINLOCK(balancing);
4038 * It checks each scheduling domain to see if it is due to be balanced,
4039 * and initiates a balancing operation if so.
4041 * Balancing parameters are set up in arch_init_sched_domains.
4043 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4046 struct rq *rq = cpu_rq(cpu);
4047 unsigned long interval;
4048 struct sched_domain *sd;
4049 /* Earliest time when we have to do rebalance again */
4050 unsigned long next_balance = jiffies + 60*HZ;
4051 int update_next_balance = 0;
4055 /* Fails alloc? Rebalancing probably not a priority right now. */
4056 if (!alloc_cpumask_var(&tmp, GFP_ATOMIC))
4059 for_each_domain(cpu, sd) {
4060 if (!(sd->flags & SD_LOAD_BALANCE))
4063 interval = sd->balance_interval;
4064 if (idle != CPU_IDLE)
4065 interval *= sd->busy_factor;
4067 /* scale ms to jiffies */
4068 interval = msecs_to_jiffies(interval);
4069 if (unlikely(!interval))
4071 if (interval > HZ*NR_CPUS/10)
4072 interval = HZ*NR_CPUS/10;
4074 need_serialize = sd->flags & SD_SERIALIZE;
4076 if (need_serialize) {
4077 if (!spin_trylock(&balancing))
4081 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4082 if (load_balance(cpu, rq, sd, idle, &balance, tmp)) {
4084 * We've pulled tasks over so either we're no
4085 * longer idle, or one of our SMT siblings is
4088 idle = CPU_NOT_IDLE;
4090 sd->last_balance = jiffies;
4093 spin_unlock(&balancing);
4095 if (time_after(next_balance, sd->last_balance + interval)) {
4096 next_balance = sd->last_balance + interval;
4097 update_next_balance = 1;
4101 * Stop the load balance at this level. There is another
4102 * CPU in our sched group which is doing load balancing more
4110 * next_balance will be updated only when there is a need.
4111 * When the cpu is attached to null domain for ex, it will not be
4114 if (likely(update_next_balance))
4115 rq->next_balance = next_balance;
4117 free_cpumask_var(tmp);
4121 * run_rebalance_domains is triggered when needed from the scheduler tick.
4122 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4123 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4125 static void run_rebalance_domains(struct softirq_action *h)
4127 int this_cpu = smp_processor_id();
4128 struct rq *this_rq = cpu_rq(this_cpu);
4129 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4130 CPU_IDLE : CPU_NOT_IDLE;
4132 rebalance_domains(this_cpu, idle);
4136 * If this cpu is the owner for idle load balancing, then do the
4137 * balancing on behalf of the other idle cpus whose ticks are
4140 if (this_rq->idle_at_tick &&
4141 atomic_read(&nohz.load_balancer) == this_cpu) {
4145 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4146 if (balance_cpu == this_cpu)
4150 * If this cpu gets work to do, stop the load balancing
4151 * work being done for other cpus. Next load
4152 * balancing owner will pick it up.
4157 rebalance_domains(balance_cpu, CPU_IDLE);
4159 rq = cpu_rq(balance_cpu);
4160 if (time_after(this_rq->next_balance, rq->next_balance))
4161 this_rq->next_balance = rq->next_balance;
4167 static inline int on_null_domain(int cpu)
4169 return !rcu_dereference(cpu_rq(cpu)->sd);
4173 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4175 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4176 * idle load balancing owner or decide to stop the periodic load balancing,
4177 * if the whole system is idle.
4179 static inline void trigger_load_balance(struct rq *rq, int cpu)
4183 * If we were in the nohz mode recently and busy at the current
4184 * scheduler tick, then check if we need to nominate new idle
4187 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4188 rq->in_nohz_recently = 0;
4190 if (atomic_read(&nohz.load_balancer) == cpu) {
4191 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4192 atomic_set(&nohz.load_balancer, -1);
4195 if (atomic_read(&nohz.load_balancer) == -1) {
4197 * simple selection for now: Nominate the
4198 * first cpu in the nohz list to be the next
4201 * TBD: Traverse the sched domains and nominate
4202 * the nearest cpu in the nohz.cpu_mask.
4204 int ilb = cpumask_first(nohz.cpu_mask);
4206 if (ilb < nr_cpu_ids)
4212 * If this cpu is idle and doing idle load balancing for all the
4213 * cpus with ticks stopped, is it time for that to stop?
4215 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4216 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4222 * If this cpu is idle and the idle load balancing is done by
4223 * someone else, then no need raise the SCHED_SOFTIRQ
4225 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4226 cpumask_test_cpu(cpu, nohz.cpu_mask))
4229 /* Don't need to rebalance while attached to NULL domain */
4230 if (time_after_eq(jiffies, rq->next_balance) &&
4231 likely(!on_null_domain(cpu)))
4232 raise_softirq(SCHED_SOFTIRQ);
4235 #else /* CONFIG_SMP */
4238 * on UP we do not need to balance between CPUs:
4240 static inline void idle_balance(int cpu, struct rq *rq)
4246 DEFINE_PER_CPU(struct kernel_stat, kstat);
4248 EXPORT_PER_CPU_SYMBOL(kstat);
4251 * Return any ns on the sched_clock that have not yet been banked in
4252 * @p in case that task is currently running.
4254 unsigned long long task_delta_exec(struct task_struct *p)
4256 unsigned long flags;
4260 rq = task_rq_lock(p, &flags);
4262 if (task_current(rq, p)) {
4265 update_rq_clock(rq);
4266 delta_exec = rq->clock - p->se.exec_start;
4267 if ((s64)delta_exec > 0)
4271 task_rq_unlock(rq, &flags);
4277 * Account user cpu time to a process.
4278 * @p: the process that the cpu time gets accounted to
4279 * @cputime: the cpu time spent in user space since the last update
4280 * @cputime_scaled: cputime scaled by cpu frequency
4282 void account_user_time(struct task_struct *p, cputime_t cputime,
4283 cputime_t cputime_scaled)
4285 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4288 /* Add user time to process. */
4289 p->utime = cputime_add(p->utime, cputime);
4290 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4291 account_group_user_time(p, cputime);
4293 /* Add user time to cpustat. */
4294 tmp = cputime_to_cputime64(cputime);
4295 if (TASK_NICE(p) > 0)
4296 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4298 cpustat->user = cputime64_add(cpustat->user, tmp);
4299 /* Account for user time used */
4300 acct_update_integrals(p);
4304 * Account guest cpu time to a process.
4305 * @p: the process that the cpu time gets accounted to
4306 * @cputime: the cpu time spent in virtual machine since the last update
4307 * @cputime_scaled: cputime scaled by cpu frequency
4309 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4310 cputime_t cputime_scaled)
4313 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4315 tmp = cputime_to_cputime64(cputime);
4317 /* Add guest time to process. */
4318 p->utime = cputime_add(p->utime, cputime);
4319 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4320 account_group_user_time(p, cputime);
4321 p->gtime = cputime_add(p->gtime, cputime);
4323 /* Add guest time to cpustat. */
4324 cpustat->user = cputime64_add(cpustat->user, tmp);
4325 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4329 * Account system cpu time to a process.
4330 * @p: the process that the cpu time gets accounted to
4331 * @hardirq_offset: the offset to subtract from hardirq_count()
4332 * @cputime: the cpu time spent in kernel space since the last update
4333 * @cputime_scaled: cputime scaled by cpu frequency
4335 void account_system_time(struct task_struct *p, int hardirq_offset,
4336 cputime_t cputime, cputime_t cputime_scaled)
4338 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4341 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4342 account_guest_time(p, cputime, cputime_scaled);
4346 /* Add system time to process. */
4347 p->stime = cputime_add(p->stime, cputime);
4348 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4349 account_group_system_time(p, cputime);
4351 /* Add system time to cpustat. */
4352 tmp = cputime_to_cputime64(cputime);
4353 if (hardirq_count() - hardirq_offset)
4354 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4355 else if (softirq_count())
4356 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4358 cpustat->system = cputime64_add(cpustat->system, tmp);
4360 /* Account for system time used */
4361 acct_update_integrals(p);
4365 * Account for involuntary wait time.
4366 * @steal: the cpu time spent in involuntary wait
4368 void account_steal_time(cputime_t cputime)
4370 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4371 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4373 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4377 * Account for idle time.
4378 * @cputime: the cpu time spent in idle wait
4380 void account_idle_time(cputime_t cputime)
4382 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4383 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4384 struct rq *rq = this_rq();
4386 if (atomic_read(&rq->nr_iowait) > 0)
4387 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4389 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4392 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4395 * Account a single tick of cpu time.
4396 * @p: the process that the cpu time gets accounted to
4397 * @user_tick: indicates if the tick is a user or a system tick
4399 void account_process_tick(struct task_struct *p, int user_tick)
4401 cputime_t one_jiffy = jiffies_to_cputime(1);
4402 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
4403 struct rq *rq = this_rq();
4406 account_user_time(p, one_jiffy, one_jiffy_scaled);
4407 else if (p != rq->idle)
4408 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
4411 account_idle_time(one_jiffy);
4415 * Account multiple ticks of steal time.
4416 * @p: the process from which the cpu time has been stolen
4417 * @ticks: number of stolen ticks
4419 void account_steal_ticks(unsigned long ticks)
4421 account_steal_time(jiffies_to_cputime(ticks));
4425 * Account multiple ticks of idle time.
4426 * @ticks: number of stolen ticks
4428 void account_idle_ticks(unsigned long ticks)
4430 account_idle_time(jiffies_to_cputime(ticks));
4436 * Use precise platform statistics if available:
4438 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4439 cputime_t task_utime(struct task_struct *p)
4444 cputime_t task_stime(struct task_struct *p)
4449 cputime_t task_utime(struct task_struct *p)
4451 clock_t utime = cputime_to_clock_t(p->utime),
4452 total = utime + cputime_to_clock_t(p->stime);
4456 * Use CFS's precise accounting:
4458 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4462 do_div(temp, total);
4464 utime = (clock_t)temp;
4466 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4467 return p->prev_utime;
4470 cputime_t task_stime(struct task_struct *p)
4475 * Use CFS's precise accounting. (we subtract utime from
4476 * the total, to make sure the total observed by userspace
4477 * grows monotonically - apps rely on that):
4479 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4480 cputime_to_clock_t(task_utime(p));
4483 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4485 return p->prev_stime;
4489 inline cputime_t task_gtime(struct task_struct *p)
4495 * This function gets called by the timer code, with HZ frequency.
4496 * We call it with interrupts disabled.
4498 * It also gets called by the fork code, when changing the parent's
4501 void scheduler_tick(void)
4503 int cpu = smp_processor_id();
4504 struct rq *rq = cpu_rq(cpu);
4505 struct task_struct *curr = rq->curr;
4509 spin_lock(&rq->lock);
4510 update_rq_clock(rq);
4511 update_cpu_load(rq);
4512 curr->sched_class->task_tick(rq, curr, 0);
4513 spin_unlock(&rq->lock);
4516 rq->idle_at_tick = idle_cpu(cpu);
4517 trigger_load_balance(rq, cpu);
4521 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4522 defined(CONFIG_PREEMPT_TRACER))
4524 static inline unsigned long get_parent_ip(unsigned long addr)
4526 if (in_lock_functions(addr)) {
4527 addr = CALLER_ADDR2;
4528 if (in_lock_functions(addr))
4529 addr = CALLER_ADDR3;
4534 void __kprobes add_preempt_count(int val)
4536 #ifdef CONFIG_DEBUG_PREEMPT
4540 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4543 preempt_count() += val;
4544 #ifdef CONFIG_DEBUG_PREEMPT
4546 * Spinlock count overflowing soon?
4548 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4551 if (preempt_count() == val)
4552 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4554 EXPORT_SYMBOL(add_preempt_count);
4556 void __kprobes sub_preempt_count(int val)
4558 #ifdef CONFIG_DEBUG_PREEMPT
4562 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4565 * Is the spinlock portion underflowing?
4567 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4568 !(preempt_count() & PREEMPT_MASK)))
4572 if (preempt_count() == val)
4573 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4574 preempt_count() -= val;
4576 EXPORT_SYMBOL(sub_preempt_count);
4581 * Print scheduling while atomic bug:
4583 static noinline void __schedule_bug(struct task_struct *prev)
4585 struct pt_regs *regs = get_irq_regs();
4587 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4588 prev->comm, prev->pid, preempt_count());
4590 debug_show_held_locks(prev);
4592 if (irqs_disabled())
4593 print_irqtrace_events(prev);
4602 * Various schedule()-time debugging checks and statistics:
4604 static inline void schedule_debug(struct task_struct *prev)
4607 * Test if we are atomic. Since do_exit() needs to call into
4608 * schedule() atomically, we ignore that path for now.
4609 * Otherwise, whine if we are scheduling when we should not be.
4611 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4612 __schedule_bug(prev);
4614 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4616 schedstat_inc(this_rq(), sched_count);
4617 #ifdef CONFIG_SCHEDSTATS
4618 if (unlikely(prev->lock_depth >= 0)) {
4619 schedstat_inc(this_rq(), bkl_count);
4620 schedstat_inc(prev, sched_info.bkl_count);
4625 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4627 if (prev->state == TASK_RUNNING) {
4628 u64 runtime = prev->se.sum_exec_runtime;
4630 runtime -= prev->se.prev_sum_exec_runtime;
4631 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
4634 * In order to avoid avg_overlap growing stale when we are
4635 * indeed overlapping and hence not getting put to sleep, grow
4636 * the avg_overlap on preemption.
4638 * We use the average preemption runtime because that
4639 * correlates to the amount of cache footprint a task can
4642 update_avg(&prev->se.avg_overlap, runtime);
4644 prev->sched_class->put_prev_task(rq, prev);
4648 * Pick up the highest-prio task:
4650 static inline struct task_struct *
4651 pick_next_task(struct rq *rq)
4653 const struct sched_class *class;
4654 struct task_struct *p;
4657 * Optimization: we know that if all tasks are in
4658 * the fair class we can call that function directly:
4660 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4661 p = fair_sched_class.pick_next_task(rq);
4666 class = sched_class_highest;
4668 p = class->pick_next_task(rq);
4672 * Will never be NULL as the idle class always
4673 * returns a non-NULL p:
4675 class = class->next;
4680 * schedule() is the main scheduler function.
4682 asmlinkage void __sched schedule(void)
4684 struct task_struct *prev, *next;
4685 unsigned long *switch_count;
4691 cpu = smp_processor_id();
4695 switch_count = &prev->nivcsw;
4697 release_kernel_lock(prev);
4698 need_resched_nonpreemptible:
4700 schedule_debug(prev);
4702 if (sched_feat(HRTICK))
4705 spin_lock_irq(&rq->lock);
4706 update_rq_clock(rq);
4707 clear_tsk_need_resched(prev);
4709 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4710 if (unlikely(signal_pending_state(prev->state, prev)))
4711 prev->state = TASK_RUNNING;
4713 deactivate_task(rq, prev, 1);
4714 switch_count = &prev->nvcsw;
4718 if (prev->sched_class->pre_schedule)
4719 prev->sched_class->pre_schedule(rq, prev);
4722 if (unlikely(!rq->nr_running))
4723 idle_balance(cpu, rq);
4725 put_prev_task(rq, prev);
4726 next = pick_next_task(rq);
4728 if (likely(prev != next)) {
4729 sched_info_switch(prev, next);
4735 context_switch(rq, prev, next); /* unlocks the rq */
4737 * the context switch might have flipped the stack from under
4738 * us, hence refresh the local variables.
4740 cpu = smp_processor_id();
4743 spin_unlock_irq(&rq->lock);
4745 if (unlikely(reacquire_kernel_lock(current) < 0))
4746 goto need_resched_nonpreemptible;
4748 preempt_enable_no_resched();
4749 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4752 EXPORT_SYMBOL(schedule);
4754 #ifdef CONFIG_PREEMPT
4756 * this is the entry point to schedule() from in-kernel preemption
4757 * off of preempt_enable. Kernel preemptions off return from interrupt
4758 * occur there and call schedule directly.
4760 asmlinkage void __sched preempt_schedule(void)
4762 struct thread_info *ti = current_thread_info();
4765 * If there is a non-zero preempt_count or interrupts are disabled,
4766 * we do not want to preempt the current task. Just return..
4768 if (likely(ti->preempt_count || irqs_disabled()))
4772 add_preempt_count(PREEMPT_ACTIVE);
4774 sub_preempt_count(PREEMPT_ACTIVE);
4777 * Check again in case we missed a preemption opportunity
4778 * between schedule and now.
4781 } while (need_resched());
4783 EXPORT_SYMBOL(preempt_schedule);
4786 * this is the entry point to schedule() from kernel preemption
4787 * off of irq context.
4788 * Note, that this is called and return with irqs disabled. This will
4789 * protect us against recursive calling from irq.
4791 asmlinkage void __sched preempt_schedule_irq(void)
4793 struct thread_info *ti = current_thread_info();
4795 /* Catch callers which need to be fixed */
4796 BUG_ON(ti->preempt_count || !irqs_disabled());
4799 add_preempt_count(PREEMPT_ACTIVE);
4802 local_irq_disable();
4803 sub_preempt_count(PREEMPT_ACTIVE);
4806 * Check again in case we missed a preemption opportunity
4807 * between schedule and now.
4810 } while (need_resched());
4813 #endif /* CONFIG_PREEMPT */
4815 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4818 return try_to_wake_up(curr->private, mode, sync);
4820 EXPORT_SYMBOL(default_wake_function);
4823 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4824 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4825 * number) then we wake all the non-exclusive tasks and one exclusive task.
4827 * There are circumstances in which we can try to wake a task which has already
4828 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4829 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4831 void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4832 int nr_exclusive, int sync, void *key)
4834 wait_queue_t *curr, *next;
4836 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4837 unsigned flags = curr->flags;
4839 if (curr->func(curr, mode, sync, key) &&
4840 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4846 * __wake_up - wake up threads blocked on a waitqueue.
4848 * @mode: which threads
4849 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4850 * @key: is directly passed to the wakeup function
4852 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4853 int nr_exclusive, void *key)
4855 unsigned long flags;
4857 spin_lock_irqsave(&q->lock, flags);
4858 __wake_up_common(q, mode, nr_exclusive, 0, key);
4859 spin_unlock_irqrestore(&q->lock, flags);
4861 EXPORT_SYMBOL(__wake_up);
4864 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4866 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4868 __wake_up_common(q, mode, 1, 0, NULL);
4872 * __wake_up_sync - wake up threads blocked on a waitqueue.
4874 * @mode: which threads
4875 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4877 * The sync wakeup differs that the waker knows that it will schedule
4878 * away soon, so while the target thread will be woken up, it will not
4879 * be migrated to another CPU - ie. the two threads are 'synchronized'
4880 * with each other. This can prevent needless bouncing between CPUs.
4882 * On UP it can prevent extra preemption.
4885 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4887 unsigned long flags;
4893 if (unlikely(!nr_exclusive))
4896 spin_lock_irqsave(&q->lock, flags);
4897 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4898 spin_unlock_irqrestore(&q->lock, flags);
4900 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4903 * complete: - signals a single thread waiting on this completion
4904 * @x: holds the state of this particular completion
4906 * This will wake up a single thread waiting on this completion. Threads will be
4907 * awakened in the same order in which they were queued.
4909 * See also complete_all(), wait_for_completion() and related routines.
4911 void complete(struct completion *x)
4913 unsigned long flags;
4915 spin_lock_irqsave(&x->wait.lock, flags);
4917 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4918 spin_unlock_irqrestore(&x->wait.lock, flags);
4920 EXPORT_SYMBOL(complete);
4923 * complete_all: - signals all threads waiting on this completion
4924 * @x: holds the state of this particular completion
4926 * This will wake up all threads waiting on this particular completion event.
4928 void complete_all(struct completion *x)
4930 unsigned long flags;
4932 spin_lock_irqsave(&x->wait.lock, flags);
4933 x->done += UINT_MAX/2;
4934 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4935 spin_unlock_irqrestore(&x->wait.lock, flags);
4937 EXPORT_SYMBOL(complete_all);
4939 static inline long __sched
4940 do_wait_for_common(struct completion *x, long timeout, int state)
4943 DECLARE_WAITQUEUE(wait, current);
4945 wait.flags |= WQ_FLAG_EXCLUSIVE;
4946 __add_wait_queue_tail(&x->wait, &wait);
4948 if (signal_pending_state(state, current)) {
4949 timeout = -ERESTARTSYS;
4952 __set_current_state(state);
4953 spin_unlock_irq(&x->wait.lock);
4954 timeout = schedule_timeout(timeout);
4955 spin_lock_irq(&x->wait.lock);
4956 } while (!x->done && timeout);
4957 __remove_wait_queue(&x->wait, &wait);
4962 return timeout ?: 1;
4966 wait_for_common(struct completion *x, long timeout, int state)
4970 spin_lock_irq(&x->wait.lock);
4971 timeout = do_wait_for_common(x, timeout, state);
4972 spin_unlock_irq(&x->wait.lock);
4977 * wait_for_completion: - waits for completion of a task
4978 * @x: holds the state of this particular completion
4980 * This waits to be signaled for completion of a specific task. It is NOT
4981 * interruptible and there is no timeout.
4983 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4984 * and interrupt capability. Also see complete().
4986 void __sched wait_for_completion(struct completion *x)
4988 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4990 EXPORT_SYMBOL(wait_for_completion);
4993 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4994 * @x: holds the state of this particular completion
4995 * @timeout: timeout value in jiffies
4997 * This waits for either a completion of a specific task to be signaled or for a
4998 * specified timeout to expire. The timeout is in jiffies. It is not
5001 unsigned long __sched
5002 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5004 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5006 EXPORT_SYMBOL(wait_for_completion_timeout);
5009 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5010 * @x: holds the state of this particular completion
5012 * This waits for completion of a specific task to be signaled. It is
5015 int __sched wait_for_completion_interruptible(struct completion *x)
5017 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5018 if (t == -ERESTARTSYS)
5022 EXPORT_SYMBOL(wait_for_completion_interruptible);
5025 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5026 * @x: holds the state of this particular completion
5027 * @timeout: timeout value in jiffies
5029 * This waits for either a completion of a specific task to be signaled or for a
5030 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5032 unsigned long __sched
5033 wait_for_completion_interruptible_timeout(struct completion *x,
5034 unsigned long timeout)
5036 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5038 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5041 * wait_for_completion_killable: - waits for completion of a task (killable)
5042 * @x: holds the state of this particular completion
5044 * This waits to be signaled for completion of a specific task. It can be
5045 * interrupted by a kill signal.
5047 int __sched wait_for_completion_killable(struct completion *x)
5049 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5050 if (t == -ERESTARTSYS)
5054 EXPORT_SYMBOL(wait_for_completion_killable);
5057 * try_wait_for_completion - try to decrement a completion without blocking
5058 * @x: completion structure
5060 * Returns: 0 if a decrement cannot be done without blocking
5061 * 1 if a decrement succeeded.
5063 * If a completion is being used as a counting completion,
5064 * attempt to decrement the counter without blocking. This
5065 * enables us to avoid waiting if the resource the completion
5066 * is protecting is not available.
5068 bool try_wait_for_completion(struct completion *x)
5072 spin_lock_irq(&x->wait.lock);
5077 spin_unlock_irq(&x->wait.lock);
5080 EXPORT_SYMBOL(try_wait_for_completion);
5083 * completion_done - Test to see if a completion has any waiters
5084 * @x: completion structure
5086 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5087 * 1 if there are no waiters.
5090 bool completion_done(struct completion *x)
5094 spin_lock_irq(&x->wait.lock);
5097 spin_unlock_irq(&x->wait.lock);
5100 EXPORT_SYMBOL(completion_done);
5103 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5105 unsigned long flags;
5108 init_waitqueue_entry(&wait, current);
5110 __set_current_state(state);
5112 spin_lock_irqsave(&q->lock, flags);
5113 __add_wait_queue(q, &wait);
5114 spin_unlock(&q->lock);
5115 timeout = schedule_timeout(timeout);
5116 spin_lock_irq(&q->lock);
5117 __remove_wait_queue(q, &wait);
5118 spin_unlock_irqrestore(&q->lock, flags);
5123 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5125 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5127 EXPORT_SYMBOL(interruptible_sleep_on);
5130 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5132 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5134 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5136 void __sched sleep_on(wait_queue_head_t *q)
5138 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5140 EXPORT_SYMBOL(sleep_on);
5142 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5144 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5146 EXPORT_SYMBOL(sleep_on_timeout);
5148 #ifdef CONFIG_RT_MUTEXES
5151 * rt_mutex_setprio - set the current priority of a task
5153 * @prio: prio value (kernel-internal form)
5155 * This function changes the 'effective' priority of a task. It does
5156 * not touch ->normal_prio like __setscheduler().
5158 * Used by the rt_mutex code to implement priority inheritance logic.
5160 void rt_mutex_setprio(struct task_struct *p, int prio)
5162 unsigned long flags;
5163 int oldprio, on_rq, running;
5165 const struct sched_class *prev_class = p->sched_class;
5167 BUG_ON(prio < 0 || prio > MAX_PRIO);
5169 rq = task_rq_lock(p, &flags);
5170 update_rq_clock(rq);
5173 on_rq = p->se.on_rq;
5174 running = task_current(rq, p);
5176 dequeue_task(rq, p, 0);
5178 p->sched_class->put_prev_task(rq, p);
5181 p->sched_class = &rt_sched_class;
5183 p->sched_class = &fair_sched_class;
5188 p->sched_class->set_curr_task(rq);
5190 enqueue_task(rq, p, 0);
5192 check_class_changed(rq, p, prev_class, oldprio, running);
5194 task_rq_unlock(rq, &flags);
5199 void set_user_nice(struct task_struct *p, long nice)
5201 int old_prio, delta, on_rq;
5202 unsigned long flags;
5205 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5208 * We have to be careful, if called from sys_setpriority(),
5209 * the task might be in the middle of scheduling on another CPU.
5211 rq = task_rq_lock(p, &flags);
5212 update_rq_clock(rq);
5214 * The RT priorities are set via sched_setscheduler(), but we still
5215 * allow the 'normal' nice value to be set - but as expected
5216 * it wont have any effect on scheduling until the task is
5217 * SCHED_FIFO/SCHED_RR:
5219 if (task_has_rt_policy(p)) {
5220 p->static_prio = NICE_TO_PRIO(nice);
5223 on_rq = p->se.on_rq;
5225 dequeue_task(rq, p, 0);
5227 p->static_prio = NICE_TO_PRIO(nice);
5230 p->prio = effective_prio(p);
5231 delta = p->prio - old_prio;
5234 enqueue_task(rq, p, 0);
5236 * If the task increased its priority or is running and
5237 * lowered its priority, then reschedule its CPU:
5239 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5240 resched_task(rq->curr);
5243 task_rq_unlock(rq, &flags);
5245 EXPORT_SYMBOL(set_user_nice);
5248 * can_nice - check if a task can reduce its nice value
5252 int can_nice(const struct task_struct *p, const int nice)
5254 /* convert nice value [19,-20] to rlimit style value [1,40] */
5255 int nice_rlim = 20 - nice;
5257 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5258 capable(CAP_SYS_NICE));
5261 #ifdef __ARCH_WANT_SYS_NICE
5264 * sys_nice - change the priority of the current process.
5265 * @increment: priority increment
5267 * sys_setpriority is a more generic, but much slower function that
5268 * does similar things.
5270 SYSCALL_DEFINE1(nice, int, increment)
5275 * Setpriority might change our priority at the same moment.
5276 * We don't have to worry. Conceptually one call occurs first
5277 * and we have a single winner.
5279 if (increment < -40)
5284 nice = TASK_NICE(current) + increment;
5290 if (increment < 0 && !can_nice(current, nice))
5293 retval = security_task_setnice(current, nice);
5297 set_user_nice(current, nice);
5304 * task_prio - return the priority value of a given task.
5305 * @p: the task in question.
5307 * This is the priority value as seen by users in /proc.
5308 * RT tasks are offset by -200. Normal tasks are centered
5309 * around 0, value goes from -16 to +15.
5311 int task_prio(const struct task_struct *p)
5313 return p->prio - MAX_RT_PRIO;
5317 * task_nice - return the nice value of a given task.
5318 * @p: the task in question.
5320 int task_nice(const struct task_struct *p)
5322 return TASK_NICE(p);
5324 EXPORT_SYMBOL(task_nice);
5327 * idle_cpu - is a given cpu idle currently?
5328 * @cpu: the processor in question.
5330 int idle_cpu(int cpu)
5332 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5336 * idle_task - return the idle task for a given cpu.
5337 * @cpu: the processor in question.
5339 struct task_struct *idle_task(int cpu)
5341 return cpu_rq(cpu)->idle;
5345 * find_process_by_pid - find a process with a matching PID value.
5346 * @pid: the pid in question.
5348 static struct task_struct *find_process_by_pid(pid_t pid)
5350 return pid ? find_task_by_vpid(pid) : current;
5353 /* Actually do priority change: must hold rq lock. */
5355 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5357 BUG_ON(p->se.on_rq);
5360 switch (p->policy) {
5364 p->sched_class = &fair_sched_class;
5368 p->sched_class = &rt_sched_class;
5372 p->rt_priority = prio;
5373 p->normal_prio = normal_prio(p);
5374 /* we are holding p->pi_lock already */
5375 p->prio = rt_mutex_getprio(p);
5380 * check the target process has a UID that matches the current process's
5382 static bool check_same_owner(struct task_struct *p)
5384 const struct cred *cred = current_cred(), *pcred;
5388 pcred = __task_cred(p);
5389 match = (cred->euid == pcred->euid ||
5390 cred->euid == pcred->uid);
5395 static int __sched_setscheduler(struct task_struct *p, int policy,
5396 struct sched_param *param, bool user)
5398 int retval, oldprio, oldpolicy = -1, on_rq, running;
5399 unsigned long flags;
5400 const struct sched_class *prev_class = p->sched_class;
5403 /* may grab non-irq protected spin_locks */
5404 BUG_ON(in_interrupt());
5406 /* double check policy once rq lock held */
5408 policy = oldpolicy = p->policy;
5409 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5410 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5411 policy != SCHED_IDLE)
5414 * Valid priorities for SCHED_FIFO and SCHED_RR are
5415 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5416 * SCHED_BATCH and SCHED_IDLE is 0.
5418 if (param->sched_priority < 0 ||
5419 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5420 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5422 if (rt_policy(policy) != (param->sched_priority != 0))
5426 * Allow unprivileged RT tasks to decrease priority:
5428 if (user && !capable(CAP_SYS_NICE)) {
5429 if (rt_policy(policy)) {
5430 unsigned long rlim_rtprio;
5432 if (!lock_task_sighand(p, &flags))
5434 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5435 unlock_task_sighand(p, &flags);
5437 /* can't set/change the rt policy */
5438 if (policy != p->policy && !rlim_rtprio)
5441 /* can't increase priority */
5442 if (param->sched_priority > p->rt_priority &&
5443 param->sched_priority > rlim_rtprio)
5447 * Like positive nice levels, dont allow tasks to
5448 * move out of SCHED_IDLE either:
5450 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5453 /* can't change other user's priorities */
5454 if (!check_same_owner(p))
5459 #ifdef CONFIG_RT_GROUP_SCHED
5461 * Do not allow realtime tasks into groups that have no runtime
5464 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5465 task_group(p)->rt_bandwidth.rt_runtime == 0)
5469 retval = security_task_setscheduler(p, policy, param);
5475 * make sure no PI-waiters arrive (or leave) while we are
5476 * changing the priority of the task:
5478 spin_lock_irqsave(&p->pi_lock, flags);
5480 * To be able to change p->policy safely, the apropriate
5481 * runqueue lock must be held.
5483 rq = __task_rq_lock(p);
5484 /* recheck policy now with rq lock held */
5485 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5486 policy = oldpolicy = -1;
5487 __task_rq_unlock(rq);
5488 spin_unlock_irqrestore(&p->pi_lock, flags);
5491 update_rq_clock(rq);
5492 on_rq = p->se.on_rq;
5493 running = task_current(rq, p);
5495 deactivate_task(rq, p, 0);
5497 p->sched_class->put_prev_task(rq, p);
5500 __setscheduler(rq, p, policy, param->sched_priority);
5503 p->sched_class->set_curr_task(rq);
5505 activate_task(rq, p, 0);
5507 check_class_changed(rq, p, prev_class, oldprio, running);
5509 __task_rq_unlock(rq);
5510 spin_unlock_irqrestore(&p->pi_lock, flags);
5512 rt_mutex_adjust_pi(p);
5518 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5519 * @p: the task in question.
5520 * @policy: new policy.
5521 * @param: structure containing the new RT priority.
5523 * NOTE that the task may be already dead.
5525 int sched_setscheduler(struct task_struct *p, int policy,
5526 struct sched_param *param)
5528 return __sched_setscheduler(p, policy, param, true);
5530 EXPORT_SYMBOL_GPL(sched_setscheduler);
5533 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5534 * @p: the task in question.
5535 * @policy: new policy.
5536 * @param: structure containing the new RT priority.
5538 * Just like sched_setscheduler, only don't bother checking if the
5539 * current context has permission. For example, this is needed in
5540 * stop_machine(): we create temporary high priority worker threads,
5541 * but our caller might not have that capability.
5543 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5544 struct sched_param *param)
5546 return __sched_setscheduler(p, policy, param, false);
5550 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5552 struct sched_param lparam;
5553 struct task_struct *p;
5556 if (!param || pid < 0)
5558 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5563 p = find_process_by_pid(pid);
5565 retval = sched_setscheduler(p, policy, &lparam);
5572 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5573 * @pid: the pid in question.
5574 * @policy: new policy.
5575 * @param: structure containing the new RT priority.
5577 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5578 struct sched_param __user *, param)
5580 /* negative values for policy are not valid */
5584 return do_sched_setscheduler(pid, policy, param);
5588 * sys_sched_setparam - set/change the RT priority of a thread
5589 * @pid: the pid in question.
5590 * @param: structure containing the new RT priority.
5592 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5594 return do_sched_setscheduler(pid, -1, param);
5598 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5599 * @pid: the pid in question.
5601 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5603 struct task_struct *p;
5610 read_lock(&tasklist_lock);
5611 p = find_process_by_pid(pid);
5613 retval = security_task_getscheduler(p);
5617 read_unlock(&tasklist_lock);
5622 * sys_sched_getscheduler - get the RT priority of a thread
5623 * @pid: the pid in question.
5624 * @param: structure containing the RT priority.
5626 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5628 struct sched_param lp;
5629 struct task_struct *p;
5632 if (!param || pid < 0)
5635 read_lock(&tasklist_lock);
5636 p = find_process_by_pid(pid);
5641 retval = security_task_getscheduler(p);
5645 lp.sched_priority = p->rt_priority;
5646 read_unlock(&tasklist_lock);
5649 * This one might sleep, we cannot do it with a spinlock held ...
5651 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5656 read_unlock(&tasklist_lock);
5660 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5662 cpumask_var_t cpus_allowed, new_mask;
5663 struct task_struct *p;
5667 read_lock(&tasklist_lock);
5669 p = find_process_by_pid(pid);
5671 read_unlock(&tasklist_lock);
5677 * It is not safe to call set_cpus_allowed with the
5678 * tasklist_lock held. We will bump the task_struct's
5679 * usage count and then drop tasklist_lock.
5682 read_unlock(&tasklist_lock);
5684 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5688 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5690 goto out_free_cpus_allowed;
5693 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5696 retval = security_task_setscheduler(p, 0, NULL);
5700 cpuset_cpus_allowed(p, cpus_allowed);
5701 cpumask_and(new_mask, in_mask, cpus_allowed);
5703 retval = set_cpus_allowed_ptr(p, new_mask);
5706 cpuset_cpus_allowed(p, cpus_allowed);
5707 if (!cpumask_subset(new_mask, cpus_allowed)) {
5709 * We must have raced with a concurrent cpuset
5710 * update. Just reset the cpus_allowed to the
5711 * cpuset's cpus_allowed
5713 cpumask_copy(new_mask, cpus_allowed);
5718 free_cpumask_var(new_mask);
5719 out_free_cpus_allowed:
5720 free_cpumask_var(cpus_allowed);
5727 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5728 struct cpumask *new_mask)
5730 if (len < cpumask_size())
5731 cpumask_clear(new_mask);
5732 else if (len > cpumask_size())
5733 len = cpumask_size();
5735 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5739 * sys_sched_setaffinity - set the cpu affinity of a process
5740 * @pid: pid of the process
5741 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5742 * @user_mask_ptr: user-space pointer to the new cpu mask
5744 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5745 unsigned long __user *, user_mask_ptr)
5747 cpumask_var_t new_mask;
5750 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5753 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5755 retval = sched_setaffinity(pid, new_mask);
5756 free_cpumask_var(new_mask);
5760 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5762 struct task_struct *p;
5766 read_lock(&tasklist_lock);
5769 p = find_process_by_pid(pid);
5773 retval = security_task_getscheduler(p);
5777 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5780 read_unlock(&tasklist_lock);
5787 * sys_sched_getaffinity - get the cpu affinity of a process
5788 * @pid: pid of the process
5789 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5790 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5792 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5793 unsigned long __user *, user_mask_ptr)
5798 if (len < cpumask_size())
5801 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5804 ret = sched_getaffinity(pid, mask);
5806 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
5809 ret = cpumask_size();
5811 free_cpumask_var(mask);
5817 * sys_sched_yield - yield the current processor to other threads.
5819 * This function yields the current CPU to other tasks. If there are no
5820 * other threads running on this CPU then this function will return.
5822 SYSCALL_DEFINE0(sched_yield)
5824 struct rq *rq = this_rq_lock();
5826 schedstat_inc(rq, yld_count);
5827 current->sched_class->yield_task(rq);
5830 * Since we are going to call schedule() anyway, there's
5831 * no need to preempt or enable interrupts:
5833 __release(rq->lock);
5834 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5835 _raw_spin_unlock(&rq->lock);
5836 preempt_enable_no_resched();
5843 static void __cond_resched(void)
5845 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5846 __might_sleep(__FILE__, __LINE__);
5849 * The BKS might be reacquired before we have dropped
5850 * PREEMPT_ACTIVE, which could trigger a second
5851 * cond_resched() call.
5854 add_preempt_count(PREEMPT_ACTIVE);
5856 sub_preempt_count(PREEMPT_ACTIVE);
5857 } while (need_resched());
5860 int __sched _cond_resched(void)
5862 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5863 system_state == SYSTEM_RUNNING) {
5869 EXPORT_SYMBOL(_cond_resched);
5872 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5873 * call schedule, and on return reacquire the lock.
5875 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5876 * operations here to prevent schedule() from being called twice (once via
5877 * spin_unlock(), once by hand).
5879 int cond_resched_lock(spinlock_t *lock)
5881 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5884 if (spin_needbreak(lock) || resched) {
5886 if (resched && need_resched())
5895 EXPORT_SYMBOL(cond_resched_lock);
5897 int __sched cond_resched_softirq(void)
5899 BUG_ON(!in_softirq());
5901 if (need_resched() && system_state == SYSTEM_RUNNING) {
5909 EXPORT_SYMBOL(cond_resched_softirq);
5912 * yield - yield the current processor to other threads.
5914 * This is a shortcut for kernel-space yielding - it marks the
5915 * thread runnable and calls sys_sched_yield().
5917 void __sched yield(void)
5919 set_current_state(TASK_RUNNING);
5922 EXPORT_SYMBOL(yield);
5925 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5926 * that process accounting knows that this is a task in IO wait state.
5928 * But don't do that if it is a deliberate, throttling IO wait (this task
5929 * has set its backing_dev_info: the queue against which it should throttle)
5931 void __sched io_schedule(void)
5933 struct rq *rq = &__raw_get_cpu_var(runqueues);
5935 delayacct_blkio_start();
5936 atomic_inc(&rq->nr_iowait);
5938 atomic_dec(&rq->nr_iowait);
5939 delayacct_blkio_end();
5941 EXPORT_SYMBOL(io_schedule);
5943 long __sched io_schedule_timeout(long timeout)
5945 struct rq *rq = &__raw_get_cpu_var(runqueues);
5948 delayacct_blkio_start();
5949 atomic_inc(&rq->nr_iowait);
5950 ret = schedule_timeout(timeout);
5951 atomic_dec(&rq->nr_iowait);
5952 delayacct_blkio_end();
5957 * sys_sched_get_priority_max - return maximum RT priority.
5958 * @policy: scheduling class.
5960 * this syscall returns the maximum rt_priority that can be used
5961 * by a given scheduling class.
5963 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5970 ret = MAX_USER_RT_PRIO-1;
5982 * sys_sched_get_priority_min - return minimum RT priority.
5983 * @policy: scheduling class.
5985 * this syscall returns the minimum rt_priority that can be used
5986 * by a given scheduling class.
5988 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6006 * sys_sched_rr_get_interval - return the default timeslice of a process.
6007 * @pid: pid of the process.
6008 * @interval: userspace pointer to the timeslice value.
6010 * this syscall writes the default timeslice value of a given process
6011 * into the user-space timespec buffer. A value of '0' means infinity.
6013 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6014 struct timespec __user *, interval)
6016 struct task_struct *p;
6017 unsigned int time_slice;
6025 read_lock(&tasklist_lock);
6026 p = find_process_by_pid(pid);
6030 retval = security_task_getscheduler(p);
6035 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6036 * tasks that are on an otherwise idle runqueue:
6039 if (p->policy == SCHED_RR) {
6040 time_slice = DEF_TIMESLICE;
6041 } else if (p->policy != SCHED_FIFO) {
6042 struct sched_entity *se = &p->se;
6043 unsigned long flags;
6046 rq = task_rq_lock(p, &flags);
6047 if (rq->cfs.load.weight)
6048 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6049 task_rq_unlock(rq, &flags);
6051 read_unlock(&tasklist_lock);
6052 jiffies_to_timespec(time_slice, &t);
6053 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6057 read_unlock(&tasklist_lock);
6061 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6063 void sched_show_task(struct task_struct *p)
6065 unsigned long free = 0;
6068 state = p->state ? __ffs(p->state) + 1 : 0;
6069 printk(KERN_INFO "%-13.13s %c", p->comm,
6070 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6071 #if BITS_PER_LONG == 32
6072 if (state == TASK_RUNNING)
6073 printk(KERN_CONT " running ");
6075 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6077 if (state == TASK_RUNNING)
6078 printk(KERN_CONT " running task ");
6080 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6082 #ifdef CONFIG_DEBUG_STACK_USAGE
6084 unsigned long *n = end_of_stack(p);
6087 free = (unsigned long)n - (unsigned long)end_of_stack(p);
6090 printk(KERN_CONT "%5lu %5d %6d\n", free,
6091 task_pid_nr(p), task_pid_nr(p->real_parent));
6093 show_stack(p, NULL);
6096 void show_state_filter(unsigned long state_filter)
6098 struct task_struct *g, *p;
6100 #if BITS_PER_LONG == 32
6102 " task PC stack pid father\n");
6105 " task PC stack pid father\n");
6107 read_lock(&tasklist_lock);
6108 do_each_thread(g, p) {
6110 * reset the NMI-timeout, listing all files on a slow
6111 * console might take alot of time:
6113 touch_nmi_watchdog();
6114 if (!state_filter || (p->state & state_filter))
6116 } while_each_thread(g, p);
6118 touch_all_softlockup_watchdogs();
6120 #ifdef CONFIG_SCHED_DEBUG
6121 sysrq_sched_debug_show();
6123 read_unlock(&tasklist_lock);
6125 * Only show locks if all tasks are dumped:
6127 if (state_filter == -1)
6128 debug_show_all_locks();
6131 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6133 idle->sched_class = &idle_sched_class;
6137 * init_idle - set up an idle thread for a given CPU
6138 * @idle: task in question
6139 * @cpu: cpu the idle task belongs to
6141 * NOTE: this function does not set the idle thread's NEED_RESCHED
6142 * flag, to make booting more robust.
6144 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6146 struct rq *rq = cpu_rq(cpu);
6147 unsigned long flags;
6149 spin_lock_irqsave(&rq->lock, flags);
6152 idle->se.exec_start = sched_clock();
6154 idle->prio = idle->normal_prio = MAX_PRIO;
6155 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6156 __set_task_cpu(idle, cpu);
6158 rq->curr = rq->idle = idle;
6159 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6162 spin_unlock_irqrestore(&rq->lock, flags);
6164 /* Set the preempt count _outside_ the spinlocks! */
6165 #if defined(CONFIG_PREEMPT)
6166 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6168 task_thread_info(idle)->preempt_count = 0;
6171 * The idle tasks have their own, simple scheduling class:
6173 idle->sched_class = &idle_sched_class;
6174 ftrace_graph_init_task(idle);
6178 * In a system that switches off the HZ timer nohz_cpu_mask
6179 * indicates which cpus entered this state. This is used
6180 * in the rcu update to wait only for active cpus. For system
6181 * which do not switch off the HZ timer nohz_cpu_mask should
6182 * always be CPU_BITS_NONE.
6184 cpumask_var_t nohz_cpu_mask;
6187 * Increase the granularity value when there are more CPUs,
6188 * because with more CPUs the 'effective latency' as visible
6189 * to users decreases. But the relationship is not linear,
6190 * so pick a second-best guess by going with the log2 of the
6193 * This idea comes from the SD scheduler of Con Kolivas:
6195 static inline void sched_init_granularity(void)
6197 unsigned int factor = 1 + ilog2(num_online_cpus());
6198 const unsigned long limit = 200000000;
6200 sysctl_sched_min_granularity *= factor;
6201 if (sysctl_sched_min_granularity > limit)
6202 sysctl_sched_min_granularity = limit;
6204 sysctl_sched_latency *= factor;
6205 if (sysctl_sched_latency > limit)
6206 sysctl_sched_latency = limit;
6208 sysctl_sched_wakeup_granularity *= factor;
6210 sysctl_sched_shares_ratelimit *= factor;
6215 * This is how migration works:
6217 * 1) we queue a struct migration_req structure in the source CPU's
6218 * runqueue and wake up that CPU's migration thread.
6219 * 2) we down() the locked semaphore => thread blocks.
6220 * 3) migration thread wakes up (implicitly it forces the migrated
6221 * thread off the CPU)
6222 * 4) it gets the migration request and checks whether the migrated
6223 * task is still in the wrong runqueue.
6224 * 5) if it's in the wrong runqueue then the migration thread removes
6225 * it and puts it into the right queue.
6226 * 6) migration thread up()s the semaphore.
6227 * 7) we wake up and the migration is done.
6231 * Change a given task's CPU affinity. Migrate the thread to a
6232 * proper CPU and schedule it away if the CPU it's executing on
6233 * is removed from the allowed bitmask.
6235 * NOTE: the caller must have a valid reference to the task, the
6236 * task must not exit() & deallocate itself prematurely. The
6237 * call is not atomic; no spinlocks may be held.
6239 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6241 struct migration_req req;
6242 unsigned long flags;
6246 rq = task_rq_lock(p, &flags);
6247 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6252 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6253 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6258 if (p->sched_class->set_cpus_allowed)
6259 p->sched_class->set_cpus_allowed(p, new_mask);
6261 cpumask_copy(&p->cpus_allowed, new_mask);
6262 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6265 /* Can the task run on the task's current CPU? If so, we're done */
6266 if (cpumask_test_cpu(task_cpu(p), new_mask))
6269 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6270 /* Need help from migration thread: drop lock and wait. */
6271 task_rq_unlock(rq, &flags);
6272 wake_up_process(rq->migration_thread);
6273 wait_for_completion(&req.done);
6274 tlb_migrate_finish(p->mm);
6278 task_rq_unlock(rq, &flags);
6282 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6285 * Move (not current) task off this cpu, onto dest cpu. We're doing
6286 * this because either it can't run here any more (set_cpus_allowed()
6287 * away from this CPU, or CPU going down), or because we're
6288 * attempting to rebalance this task on exec (sched_exec).
6290 * So we race with normal scheduler movements, but that's OK, as long
6291 * as the task is no longer on this CPU.
6293 * Returns non-zero if task was successfully migrated.
6295 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6297 struct rq *rq_dest, *rq_src;
6300 if (unlikely(!cpu_active(dest_cpu)))
6303 rq_src = cpu_rq(src_cpu);
6304 rq_dest = cpu_rq(dest_cpu);
6306 double_rq_lock(rq_src, rq_dest);
6307 /* Already moved. */
6308 if (task_cpu(p) != src_cpu)
6310 /* Affinity changed (again). */
6311 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6314 on_rq = p->se.on_rq;
6316 deactivate_task(rq_src, p, 0);
6318 set_task_cpu(p, dest_cpu);
6320 activate_task(rq_dest, p, 0);
6321 check_preempt_curr(rq_dest, p, 0);
6326 double_rq_unlock(rq_src, rq_dest);
6331 * migration_thread - this is a highprio system thread that performs
6332 * thread migration by bumping thread off CPU then 'pushing' onto
6335 static int migration_thread(void *data)
6337 int cpu = (long)data;
6341 BUG_ON(rq->migration_thread != current);
6343 set_current_state(TASK_INTERRUPTIBLE);
6344 while (!kthread_should_stop()) {
6345 struct migration_req *req;
6346 struct list_head *head;
6348 spin_lock_irq(&rq->lock);
6350 if (cpu_is_offline(cpu)) {
6351 spin_unlock_irq(&rq->lock);
6355 if (rq->active_balance) {
6356 active_load_balance(rq, cpu);
6357 rq->active_balance = 0;
6360 head = &rq->migration_queue;
6362 if (list_empty(head)) {
6363 spin_unlock_irq(&rq->lock);
6365 set_current_state(TASK_INTERRUPTIBLE);
6368 req = list_entry(head->next, struct migration_req, list);
6369 list_del_init(head->next);
6371 spin_unlock(&rq->lock);
6372 __migrate_task(req->task, cpu, req->dest_cpu);
6375 complete(&req->done);
6377 __set_current_state(TASK_RUNNING);
6381 /* Wait for kthread_stop */
6382 set_current_state(TASK_INTERRUPTIBLE);
6383 while (!kthread_should_stop()) {
6385 set_current_state(TASK_INTERRUPTIBLE);
6387 __set_current_state(TASK_RUNNING);
6391 #ifdef CONFIG_HOTPLUG_CPU
6393 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6397 local_irq_disable();
6398 ret = __migrate_task(p, src_cpu, dest_cpu);
6404 * Figure out where task on dead CPU should go, use force if necessary.
6406 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6409 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
6412 /* Look for allowed, online CPU in same node. */
6413 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
6414 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6417 /* Any allowed, online CPU? */
6418 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
6419 if (dest_cpu < nr_cpu_ids)
6422 /* No more Mr. Nice Guy. */
6423 if (dest_cpu >= nr_cpu_ids) {
6424 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
6425 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
6428 * Don't tell them about moving exiting tasks or
6429 * kernel threads (both mm NULL), since they never
6432 if (p->mm && printk_ratelimit()) {
6433 printk(KERN_INFO "process %d (%s) no "
6434 "longer affine to cpu%d\n",
6435 task_pid_nr(p), p->comm, dead_cpu);
6440 /* It can have affinity changed while we were choosing. */
6441 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
6446 * While a dead CPU has no uninterruptible tasks queued at this point,
6447 * it might still have a nonzero ->nr_uninterruptible counter, because
6448 * for performance reasons the counter is not stricly tracking tasks to
6449 * their home CPUs. So we just add the counter to another CPU's counter,
6450 * to keep the global sum constant after CPU-down:
6452 static void migrate_nr_uninterruptible(struct rq *rq_src)
6454 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
6455 unsigned long flags;
6457 local_irq_save(flags);
6458 double_rq_lock(rq_src, rq_dest);
6459 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6460 rq_src->nr_uninterruptible = 0;
6461 double_rq_unlock(rq_src, rq_dest);
6462 local_irq_restore(flags);
6465 /* Run through task list and migrate tasks from the dead cpu. */
6466 static void migrate_live_tasks(int src_cpu)
6468 struct task_struct *p, *t;
6470 read_lock(&tasklist_lock);
6472 do_each_thread(t, p) {
6476 if (task_cpu(p) == src_cpu)
6477 move_task_off_dead_cpu(src_cpu, p);
6478 } while_each_thread(t, p);
6480 read_unlock(&tasklist_lock);
6484 * Schedules idle task to be the next runnable task on current CPU.
6485 * It does so by boosting its priority to highest possible.
6486 * Used by CPU offline code.
6488 void sched_idle_next(void)
6490 int this_cpu = smp_processor_id();
6491 struct rq *rq = cpu_rq(this_cpu);
6492 struct task_struct *p = rq->idle;
6493 unsigned long flags;
6495 /* cpu has to be offline */
6496 BUG_ON(cpu_online(this_cpu));
6499 * Strictly not necessary since rest of the CPUs are stopped by now
6500 * and interrupts disabled on the current cpu.
6502 spin_lock_irqsave(&rq->lock, flags);
6504 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6506 update_rq_clock(rq);
6507 activate_task(rq, p, 0);
6509 spin_unlock_irqrestore(&rq->lock, flags);
6513 * Ensures that the idle task is using init_mm right before its cpu goes
6516 void idle_task_exit(void)
6518 struct mm_struct *mm = current->active_mm;
6520 BUG_ON(cpu_online(smp_processor_id()));
6523 switch_mm(mm, &init_mm, current);
6527 /* called under rq->lock with disabled interrupts */
6528 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6530 struct rq *rq = cpu_rq(dead_cpu);
6532 /* Must be exiting, otherwise would be on tasklist. */
6533 BUG_ON(!p->exit_state);
6535 /* Cannot have done final schedule yet: would have vanished. */
6536 BUG_ON(p->state == TASK_DEAD);
6541 * Drop lock around migration; if someone else moves it,
6542 * that's OK. No task can be added to this CPU, so iteration is
6545 spin_unlock_irq(&rq->lock);
6546 move_task_off_dead_cpu(dead_cpu, p);
6547 spin_lock_irq(&rq->lock);
6552 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6553 static void migrate_dead_tasks(unsigned int dead_cpu)
6555 struct rq *rq = cpu_rq(dead_cpu);
6556 struct task_struct *next;
6559 if (!rq->nr_running)
6561 update_rq_clock(rq);
6562 next = pick_next_task(rq);
6565 next->sched_class->put_prev_task(rq, next);
6566 migrate_dead(dead_cpu, next);
6570 #endif /* CONFIG_HOTPLUG_CPU */
6572 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6574 static struct ctl_table sd_ctl_dir[] = {
6576 .procname = "sched_domain",
6582 static struct ctl_table sd_ctl_root[] = {
6584 .ctl_name = CTL_KERN,
6585 .procname = "kernel",
6587 .child = sd_ctl_dir,
6592 static struct ctl_table *sd_alloc_ctl_entry(int n)
6594 struct ctl_table *entry =
6595 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6600 static void sd_free_ctl_entry(struct ctl_table **tablep)
6602 struct ctl_table *entry;
6605 * In the intermediate directories, both the child directory and
6606 * procname are dynamically allocated and could fail but the mode
6607 * will always be set. In the lowest directory the names are
6608 * static strings and all have proc handlers.
6610 for (entry = *tablep; entry->mode; entry++) {
6612 sd_free_ctl_entry(&entry->child);
6613 if (entry->proc_handler == NULL)
6614 kfree(entry->procname);
6622 set_table_entry(struct ctl_table *entry,
6623 const char *procname, void *data, int maxlen,
6624 mode_t mode, proc_handler *proc_handler)
6626 entry->procname = procname;
6628 entry->maxlen = maxlen;
6630 entry->proc_handler = proc_handler;
6633 static struct ctl_table *
6634 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6636 struct ctl_table *table = sd_alloc_ctl_entry(13);
6641 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6642 sizeof(long), 0644, proc_doulongvec_minmax);
6643 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6644 sizeof(long), 0644, proc_doulongvec_minmax);
6645 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6646 sizeof(int), 0644, proc_dointvec_minmax);
6647 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6648 sizeof(int), 0644, proc_dointvec_minmax);
6649 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6650 sizeof(int), 0644, proc_dointvec_minmax);
6651 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6652 sizeof(int), 0644, proc_dointvec_minmax);
6653 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6654 sizeof(int), 0644, proc_dointvec_minmax);
6655 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6656 sizeof(int), 0644, proc_dointvec_minmax);
6657 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6658 sizeof(int), 0644, proc_dointvec_minmax);
6659 set_table_entry(&table[9], "cache_nice_tries",
6660 &sd->cache_nice_tries,
6661 sizeof(int), 0644, proc_dointvec_minmax);
6662 set_table_entry(&table[10], "flags", &sd->flags,
6663 sizeof(int), 0644, proc_dointvec_minmax);
6664 set_table_entry(&table[11], "name", sd->name,
6665 CORENAME_MAX_SIZE, 0444, proc_dostring);
6666 /* &table[12] is terminator */
6671 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6673 struct ctl_table *entry, *table;
6674 struct sched_domain *sd;
6675 int domain_num = 0, i;
6678 for_each_domain(cpu, sd)
6680 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6685 for_each_domain(cpu, sd) {
6686 snprintf(buf, 32, "domain%d", i);
6687 entry->procname = kstrdup(buf, GFP_KERNEL);
6689 entry->child = sd_alloc_ctl_domain_table(sd);
6696 static struct ctl_table_header *sd_sysctl_header;
6697 static void register_sched_domain_sysctl(void)
6699 int i, cpu_num = num_online_cpus();
6700 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6703 WARN_ON(sd_ctl_dir[0].child);
6704 sd_ctl_dir[0].child = entry;
6709 for_each_online_cpu(i) {
6710 snprintf(buf, 32, "cpu%d", i);
6711 entry->procname = kstrdup(buf, GFP_KERNEL);
6713 entry->child = sd_alloc_ctl_cpu_table(i);
6717 WARN_ON(sd_sysctl_header);
6718 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6721 /* may be called multiple times per register */
6722 static void unregister_sched_domain_sysctl(void)
6724 if (sd_sysctl_header)
6725 unregister_sysctl_table(sd_sysctl_header);
6726 sd_sysctl_header = NULL;
6727 if (sd_ctl_dir[0].child)
6728 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6731 static void register_sched_domain_sysctl(void)
6734 static void unregister_sched_domain_sysctl(void)
6739 static void set_rq_online(struct rq *rq)
6742 const struct sched_class *class;
6744 cpumask_set_cpu(rq->cpu, rq->rd->online);
6747 for_each_class(class) {
6748 if (class->rq_online)
6749 class->rq_online(rq);
6754 static void set_rq_offline(struct rq *rq)
6757 const struct sched_class *class;
6759 for_each_class(class) {
6760 if (class->rq_offline)
6761 class->rq_offline(rq);
6764 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6770 * migration_call - callback that gets triggered when a CPU is added.
6771 * Here we can start up the necessary migration thread for the new CPU.
6773 static int __cpuinit
6774 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6776 struct task_struct *p;
6777 int cpu = (long)hcpu;
6778 unsigned long flags;
6783 case CPU_UP_PREPARE:
6784 case CPU_UP_PREPARE_FROZEN:
6785 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6788 kthread_bind(p, cpu);
6789 /* Must be high prio: stop_machine expects to yield to it. */
6790 rq = task_rq_lock(p, &flags);
6791 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6792 task_rq_unlock(rq, &flags);
6793 cpu_rq(cpu)->migration_thread = p;
6797 case CPU_ONLINE_FROZEN:
6798 /* Strictly unnecessary, as first user will wake it. */
6799 wake_up_process(cpu_rq(cpu)->migration_thread);
6801 /* Update our root-domain */
6803 spin_lock_irqsave(&rq->lock, flags);
6805 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6809 spin_unlock_irqrestore(&rq->lock, flags);
6812 #ifdef CONFIG_HOTPLUG_CPU
6813 case CPU_UP_CANCELED:
6814 case CPU_UP_CANCELED_FROZEN:
6815 if (!cpu_rq(cpu)->migration_thread)
6817 /* Unbind it from offline cpu so it can run. Fall thru. */
6818 kthread_bind(cpu_rq(cpu)->migration_thread,
6819 cpumask_any(cpu_online_mask));
6820 kthread_stop(cpu_rq(cpu)->migration_thread);
6821 cpu_rq(cpu)->migration_thread = NULL;
6825 case CPU_DEAD_FROZEN:
6826 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6827 migrate_live_tasks(cpu);
6829 kthread_stop(rq->migration_thread);
6830 rq->migration_thread = NULL;
6831 /* Idle task back to normal (off runqueue, low prio) */
6832 spin_lock_irq(&rq->lock);
6833 update_rq_clock(rq);
6834 deactivate_task(rq, rq->idle, 0);
6835 rq->idle->static_prio = MAX_PRIO;
6836 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6837 rq->idle->sched_class = &idle_sched_class;
6838 migrate_dead_tasks(cpu);
6839 spin_unlock_irq(&rq->lock);
6841 migrate_nr_uninterruptible(rq);
6842 BUG_ON(rq->nr_running != 0);
6845 * No need to migrate the tasks: it was best-effort if
6846 * they didn't take sched_hotcpu_mutex. Just wake up
6849 spin_lock_irq(&rq->lock);
6850 while (!list_empty(&rq->migration_queue)) {
6851 struct migration_req *req;
6853 req = list_entry(rq->migration_queue.next,
6854 struct migration_req, list);
6855 list_del_init(&req->list);
6856 spin_unlock_irq(&rq->lock);
6857 complete(&req->done);
6858 spin_lock_irq(&rq->lock);
6860 spin_unlock_irq(&rq->lock);
6864 case CPU_DYING_FROZEN:
6865 /* Update our root-domain */
6867 spin_lock_irqsave(&rq->lock, flags);
6869 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6872 spin_unlock_irqrestore(&rq->lock, flags);
6879 /* Register at highest priority so that task migration (migrate_all_tasks)
6880 * happens before everything else.
6882 static struct notifier_block __cpuinitdata migration_notifier = {
6883 .notifier_call = migration_call,
6887 static int __init migration_init(void)
6889 void *cpu = (void *)(long)smp_processor_id();
6892 /* Start one for the boot CPU: */
6893 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6894 BUG_ON(err == NOTIFY_BAD);
6895 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6896 register_cpu_notifier(&migration_notifier);
6900 early_initcall(migration_init);
6905 #ifdef CONFIG_SCHED_DEBUG
6907 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6908 struct cpumask *groupmask)
6910 struct sched_group *group = sd->groups;
6913 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6914 cpumask_clear(groupmask);
6916 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6918 if (!(sd->flags & SD_LOAD_BALANCE)) {
6919 printk("does not load-balance\n");
6921 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6926 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6928 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6929 printk(KERN_ERR "ERROR: domain->span does not contain "
6932 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6933 printk(KERN_ERR "ERROR: domain->groups does not contain"
6937 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6941 printk(KERN_ERR "ERROR: group is NULL\n");
6945 if (!group->__cpu_power) {
6946 printk(KERN_CONT "\n");
6947 printk(KERN_ERR "ERROR: domain->cpu_power not "
6952 if (!cpumask_weight(sched_group_cpus(group))) {
6953 printk(KERN_CONT "\n");
6954 printk(KERN_ERR "ERROR: empty group\n");
6958 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6959 printk(KERN_CONT "\n");
6960 printk(KERN_ERR "ERROR: repeated CPUs\n");
6964 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6966 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6967 printk(KERN_CONT " %s", str);
6969 group = group->next;
6970 } while (group != sd->groups);
6971 printk(KERN_CONT "\n");
6973 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6974 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6977 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6978 printk(KERN_ERR "ERROR: parent span is not a superset "
6979 "of domain->span\n");
6983 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6985 cpumask_var_t groupmask;
6989 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6993 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6995 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6996 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7001 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7008 free_cpumask_var(groupmask);
7010 #else /* !CONFIG_SCHED_DEBUG */
7011 # define sched_domain_debug(sd, cpu) do { } while (0)
7012 #endif /* CONFIG_SCHED_DEBUG */
7014 static int sd_degenerate(struct sched_domain *sd)
7016 if (cpumask_weight(sched_domain_span(sd)) == 1)
7019 /* Following flags need at least 2 groups */
7020 if (sd->flags & (SD_LOAD_BALANCE |
7021 SD_BALANCE_NEWIDLE |
7025 SD_SHARE_PKG_RESOURCES)) {
7026 if (sd->groups != sd->groups->next)
7030 /* Following flags don't use groups */
7031 if (sd->flags & (SD_WAKE_IDLE |
7040 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7042 unsigned long cflags = sd->flags, pflags = parent->flags;
7044 if (sd_degenerate(parent))
7047 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7050 /* Does parent contain flags not in child? */
7051 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7052 if (cflags & SD_WAKE_AFFINE)
7053 pflags &= ~SD_WAKE_BALANCE;
7054 /* Flags needing groups don't count if only 1 group in parent */
7055 if (parent->groups == parent->groups->next) {
7056 pflags &= ~(SD_LOAD_BALANCE |
7057 SD_BALANCE_NEWIDLE |
7061 SD_SHARE_PKG_RESOURCES);
7062 if (nr_node_ids == 1)
7063 pflags &= ~SD_SERIALIZE;
7065 if (~cflags & pflags)
7071 static void free_rootdomain(struct root_domain *rd)
7073 cpupri_cleanup(&rd->cpupri);
7075 free_cpumask_var(rd->rto_mask);
7076 free_cpumask_var(rd->online);
7077 free_cpumask_var(rd->span);
7081 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7083 struct root_domain *old_rd = NULL;
7084 unsigned long flags;
7086 spin_lock_irqsave(&rq->lock, flags);
7091 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7094 cpumask_clear_cpu(rq->cpu, old_rd->span);
7097 * If we dont want to free the old_rt yet then
7098 * set old_rd to NULL to skip the freeing later
7101 if (!atomic_dec_and_test(&old_rd->refcount))
7105 atomic_inc(&rd->refcount);
7108 cpumask_set_cpu(rq->cpu, rd->span);
7109 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7112 spin_unlock_irqrestore(&rq->lock, flags);
7115 free_rootdomain(old_rd);
7118 static int __init_refok init_rootdomain(struct root_domain *rd, bool bootmem)
7120 memset(rd, 0, sizeof(*rd));
7123 alloc_bootmem_cpumask_var(&def_root_domain.span);
7124 alloc_bootmem_cpumask_var(&def_root_domain.online);
7125 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
7126 cpupri_init(&rd->cpupri, true);
7130 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7132 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7134 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7137 if (cpupri_init(&rd->cpupri, false) != 0)
7142 free_cpumask_var(rd->rto_mask);
7144 free_cpumask_var(rd->online);
7146 free_cpumask_var(rd->span);
7151 static void init_defrootdomain(void)
7153 init_rootdomain(&def_root_domain, true);
7155 atomic_set(&def_root_domain.refcount, 1);
7158 static struct root_domain *alloc_rootdomain(void)
7160 struct root_domain *rd;
7162 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7166 if (init_rootdomain(rd, false) != 0) {
7175 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7176 * hold the hotplug lock.
7179 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7181 struct rq *rq = cpu_rq(cpu);
7182 struct sched_domain *tmp;
7184 /* Remove the sched domains which do not contribute to scheduling. */
7185 for (tmp = sd; tmp; ) {
7186 struct sched_domain *parent = tmp->parent;
7190 if (sd_parent_degenerate(tmp, parent)) {
7191 tmp->parent = parent->parent;
7193 parent->parent->child = tmp;
7198 if (sd && sd_degenerate(sd)) {
7204 sched_domain_debug(sd, cpu);
7206 rq_attach_root(rq, rd);
7207 rcu_assign_pointer(rq->sd, sd);
7210 /* cpus with isolated domains */
7211 static cpumask_var_t cpu_isolated_map;
7213 /* Setup the mask of cpus configured for isolated domains */
7214 static int __init isolated_cpu_setup(char *str)
7216 cpulist_parse(str, cpu_isolated_map);
7220 __setup("isolcpus=", isolated_cpu_setup);
7223 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7224 * to a function which identifies what group(along with sched group) a CPU
7225 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7226 * (due to the fact that we keep track of groups covered with a struct cpumask).
7228 * init_sched_build_groups will build a circular linked list of the groups
7229 * covered by the given span, and will set each group's ->cpumask correctly,
7230 * and ->cpu_power to 0.
7233 init_sched_build_groups(const struct cpumask *span,
7234 const struct cpumask *cpu_map,
7235 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7236 struct sched_group **sg,
7237 struct cpumask *tmpmask),
7238 struct cpumask *covered, struct cpumask *tmpmask)
7240 struct sched_group *first = NULL, *last = NULL;
7243 cpumask_clear(covered);
7245 for_each_cpu(i, span) {
7246 struct sched_group *sg;
7247 int group = group_fn(i, cpu_map, &sg, tmpmask);
7250 if (cpumask_test_cpu(i, covered))
7253 cpumask_clear(sched_group_cpus(sg));
7254 sg->__cpu_power = 0;
7256 for_each_cpu(j, span) {
7257 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7260 cpumask_set_cpu(j, covered);
7261 cpumask_set_cpu(j, sched_group_cpus(sg));
7272 #define SD_NODES_PER_DOMAIN 16
7277 * find_next_best_node - find the next node to include in a sched_domain
7278 * @node: node whose sched_domain we're building
7279 * @used_nodes: nodes already in the sched_domain
7281 * Find the next node to include in a given scheduling domain. Simply
7282 * finds the closest node not already in the @used_nodes map.
7284 * Should use nodemask_t.
7286 static int find_next_best_node(int node, nodemask_t *used_nodes)
7288 int i, n, val, min_val, best_node = 0;
7292 for (i = 0; i < nr_node_ids; i++) {
7293 /* Start at @node */
7294 n = (node + i) % nr_node_ids;
7296 if (!nr_cpus_node(n))
7299 /* Skip already used nodes */
7300 if (node_isset(n, *used_nodes))
7303 /* Simple min distance search */
7304 val = node_distance(node, n);
7306 if (val < min_val) {
7312 node_set(best_node, *used_nodes);
7317 * sched_domain_node_span - get a cpumask for a node's sched_domain
7318 * @node: node whose cpumask we're constructing
7319 * @span: resulting cpumask
7321 * Given a node, construct a good cpumask for its sched_domain to span. It
7322 * should be one that prevents unnecessary balancing, but also spreads tasks
7325 static void sched_domain_node_span(int node, struct cpumask *span)
7327 nodemask_t used_nodes;
7330 cpumask_clear(span);
7331 nodes_clear(used_nodes);
7333 cpumask_or(span, span, cpumask_of_node(node));
7334 node_set(node, used_nodes);
7336 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7337 int next_node = find_next_best_node(node, &used_nodes);
7339 cpumask_or(span, span, cpumask_of_node(next_node));
7342 #endif /* CONFIG_NUMA */
7344 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7347 * The cpus mask in sched_group and sched_domain hangs off the end.
7348 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7349 * for nr_cpu_ids < CONFIG_NR_CPUS.
7351 struct static_sched_group {
7352 struct sched_group sg;
7353 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7356 struct static_sched_domain {
7357 struct sched_domain sd;
7358 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7362 * SMT sched-domains:
7364 #ifdef CONFIG_SCHED_SMT
7365 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7366 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7369 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
7370 struct sched_group **sg, struct cpumask *unused)
7373 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7376 #endif /* CONFIG_SCHED_SMT */
7379 * multi-core sched-domains:
7381 #ifdef CONFIG_SCHED_MC
7382 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7383 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7384 #endif /* CONFIG_SCHED_MC */
7386 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7388 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7389 struct sched_group **sg, struct cpumask *mask)
7393 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7394 group = cpumask_first(mask);
7396 *sg = &per_cpu(sched_group_core, group).sg;
7399 #elif defined(CONFIG_SCHED_MC)
7401 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7402 struct sched_group **sg, struct cpumask *unused)
7405 *sg = &per_cpu(sched_group_core, cpu).sg;
7410 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
7411 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
7414 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
7415 struct sched_group **sg, struct cpumask *mask)
7418 #ifdef CONFIG_SCHED_MC
7419 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
7420 group = cpumask_first(mask);
7421 #elif defined(CONFIG_SCHED_SMT)
7422 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7423 group = cpumask_first(mask);
7428 *sg = &per_cpu(sched_group_phys, group).sg;
7434 * The init_sched_build_groups can't handle what we want to do with node
7435 * groups, so roll our own. Now each node has its own list of groups which
7436 * gets dynamically allocated.
7438 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
7439 static struct sched_group ***sched_group_nodes_bycpu;
7441 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
7442 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7444 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
7445 struct sched_group **sg,
7446 struct cpumask *nodemask)
7450 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7451 group = cpumask_first(nodemask);
7454 *sg = &per_cpu(sched_group_allnodes, group).sg;
7458 static void init_numa_sched_groups_power(struct sched_group *group_head)
7460 struct sched_group *sg = group_head;
7466 for_each_cpu(j, sched_group_cpus(sg)) {
7467 struct sched_domain *sd;
7469 sd = &per_cpu(phys_domains, j).sd;
7470 if (j != cpumask_first(sched_group_cpus(sd->groups))) {
7472 * Only add "power" once for each
7478 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7481 } while (sg != group_head);
7483 #endif /* CONFIG_NUMA */
7486 /* Free memory allocated for various sched_group structures */
7487 static void free_sched_groups(const struct cpumask *cpu_map,
7488 struct cpumask *nodemask)
7492 for_each_cpu(cpu, cpu_map) {
7493 struct sched_group **sched_group_nodes
7494 = sched_group_nodes_bycpu[cpu];
7496 if (!sched_group_nodes)
7499 for (i = 0; i < nr_node_ids; i++) {
7500 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7502 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7503 if (cpumask_empty(nodemask))
7513 if (oldsg != sched_group_nodes[i])
7516 kfree(sched_group_nodes);
7517 sched_group_nodes_bycpu[cpu] = NULL;
7520 #else /* !CONFIG_NUMA */
7521 static void free_sched_groups(const struct cpumask *cpu_map,
7522 struct cpumask *nodemask)
7525 #endif /* CONFIG_NUMA */
7528 * Initialize sched groups cpu_power.
7530 * cpu_power indicates the capacity of sched group, which is used while
7531 * distributing the load between different sched groups in a sched domain.
7532 * Typically cpu_power for all the groups in a sched domain will be same unless
7533 * there are asymmetries in the topology. If there are asymmetries, group
7534 * having more cpu_power will pickup more load compared to the group having
7537 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7538 * the maximum number of tasks a group can handle in the presence of other idle
7539 * or lightly loaded groups in the same sched domain.
7541 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7543 struct sched_domain *child;
7544 struct sched_group *group;
7546 WARN_ON(!sd || !sd->groups);
7548 if (cpu != cpumask_first(sched_group_cpus(sd->groups)))
7553 sd->groups->__cpu_power = 0;
7556 * For perf policy, if the groups in child domain share resources
7557 * (for example cores sharing some portions of the cache hierarchy
7558 * or SMT), then set this domain groups cpu_power such that each group
7559 * can handle only one task, when there are other idle groups in the
7560 * same sched domain.
7562 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7564 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7565 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7570 * add cpu_power of each child group to this groups cpu_power
7572 group = child->groups;
7574 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7575 group = group->next;
7576 } while (group != child->groups);
7580 * Initializers for schedule domains
7581 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7584 #ifdef CONFIG_SCHED_DEBUG
7585 # define SD_INIT_NAME(sd, type) sd->name = #type
7587 # define SD_INIT_NAME(sd, type) do { } while (0)
7590 #define SD_INIT(sd, type) sd_init_##type(sd)
7592 #define SD_INIT_FUNC(type) \
7593 static noinline void sd_init_##type(struct sched_domain *sd) \
7595 memset(sd, 0, sizeof(*sd)); \
7596 *sd = SD_##type##_INIT; \
7597 sd->level = SD_LV_##type; \
7598 SD_INIT_NAME(sd, type); \
7603 SD_INIT_FUNC(ALLNODES)
7606 #ifdef CONFIG_SCHED_SMT
7607 SD_INIT_FUNC(SIBLING)
7609 #ifdef CONFIG_SCHED_MC
7613 static int default_relax_domain_level = -1;
7615 static int __init setup_relax_domain_level(char *str)
7619 val = simple_strtoul(str, NULL, 0);
7620 if (val < SD_LV_MAX)
7621 default_relax_domain_level = val;
7625 __setup("relax_domain_level=", setup_relax_domain_level);
7627 static void set_domain_attribute(struct sched_domain *sd,
7628 struct sched_domain_attr *attr)
7632 if (!attr || attr->relax_domain_level < 0) {
7633 if (default_relax_domain_level < 0)
7636 request = default_relax_domain_level;
7638 request = attr->relax_domain_level;
7639 if (request < sd->level) {
7640 /* turn off idle balance on this domain */
7641 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7643 /* turn on idle balance on this domain */
7644 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7649 * Build sched domains for a given set of cpus and attach the sched domains
7650 * to the individual cpus
7652 static int __build_sched_domains(const struct cpumask *cpu_map,
7653 struct sched_domain_attr *attr)
7655 int i, err = -ENOMEM;
7656 struct root_domain *rd;
7657 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
7660 cpumask_var_t domainspan, covered, notcovered;
7661 struct sched_group **sched_group_nodes = NULL;
7662 int sd_allnodes = 0;
7664 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
7666 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
7667 goto free_domainspan;
7668 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
7672 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
7673 goto free_notcovered;
7674 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
7676 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
7677 goto free_this_sibling_map;
7678 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
7679 goto free_this_core_map;
7680 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
7681 goto free_send_covered;
7685 * Allocate the per-node list of sched groups
7687 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7689 if (!sched_group_nodes) {
7690 printk(KERN_WARNING "Can not alloc sched group node list\n");
7695 rd = alloc_rootdomain();
7697 printk(KERN_WARNING "Cannot alloc root domain\n");
7698 goto free_sched_groups;
7702 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
7706 * Set up domains for cpus specified by the cpu_map.
7708 for_each_cpu(i, cpu_map) {
7709 struct sched_domain *sd = NULL, *p;
7711 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
7714 if (cpumask_weight(cpu_map) >
7715 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
7716 sd = &per_cpu(allnodes_domains, i).sd;
7717 SD_INIT(sd, ALLNODES);
7718 set_domain_attribute(sd, attr);
7719 cpumask_copy(sched_domain_span(sd), cpu_map);
7720 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7726 sd = &per_cpu(node_domains, i).sd;
7728 set_domain_attribute(sd, attr);
7729 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7733 cpumask_and(sched_domain_span(sd),
7734 sched_domain_span(sd), cpu_map);
7738 sd = &per_cpu(phys_domains, i).sd;
7740 set_domain_attribute(sd, attr);
7741 cpumask_copy(sched_domain_span(sd), nodemask);
7745 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7747 #ifdef CONFIG_SCHED_MC
7749 sd = &per_cpu(core_domains, i).sd;
7751 set_domain_attribute(sd, attr);
7752 cpumask_and(sched_domain_span(sd), cpu_map,
7753 cpu_coregroup_mask(i));
7756 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7759 #ifdef CONFIG_SCHED_SMT
7761 sd = &per_cpu(cpu_domains, i).sd;
7762 SD_INIT(sd, SIBLING);
7763 set_domain_attribute(sd, attr);
7764 cpumask_and(sched_domain_span(sd),
7765 &per_cpu(cpu_sibling_map, i), cpu_map);
7768 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7772 #ifdef CONFIG_SCHED_SMT
7773 /* Set up CPU (sibling) groups */
7774 for_each_cpu(i, cpu_map) {
7775 cpumask_and(this_sibling_map,
7776 &per_cpu(cpu_sibling_map, i), cpu_map);
7777 if (i != cpumask_first(this_sibling_map))
7780 init_sched_build_groups(this_sibling_map, cpu_map,
7782 send_covered, tmpmask);
7786 #ifdef CONFIG_SCHED_MC
7787 /* Set up multi-core groups */
7788 for_each_cpu(i, cpu_map) {
7789 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
7790 if (i != cpumask_first(this_core_map))
7793 init_sched_build_groups(this_core_map, cpu_map,
7795 send_covered, tmpmask);
7799 /* Set up physical groups */
7800 for (i = 0; i < nr_node_ids; i++) {
7801 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7802 if (cpumask_empty(nodemask))
7805 init_sched_build_groups(nodemask, cpu_map,
7807 send_covered, tmpmask);
7811 /* Set up node groups */
7813 init_sched_build_groups(cpu_map, cpu_map,
7814 &cpu_to_allnodes_group,
7815 send_covered, tmpmask);
7818 for (i = 0; i < nr_node_ids; i++) {
7819 /* Set up node groups */
7820 struct sched_group *sg, *prev;
7823 cpumask_clear(covered);
7824 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7825 if (cpumask_empty(nodemask)) {
7826 sched_group_nodes[i] = NULL;
7830 sched_domain_node_span(i, domainspan);
7831 cpumask_and(domainspan, domainspan, cpu_map);
7833 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7836 printk(KERN_WARNING "Can not alloc domain group for "
7840 sched_group_nodes[i] = sg;
7841 for_each_cpu(j, nodemask) {
7842 struct sched_domain *sd;
7844 sd = &per_cpu(node_domains, j).sd;
7847 sg->__cpu_power = 0;
7848 cpumask_copy(sched_group_cpus(sg), nodemask);
7850 cpumask_or(covered, covered, nodemask);
7853 for (j = 0; j < nr_node_ids; j++) {
7854 int n = (i + j) % nr_node_ids;
7856 cpumask_complement(notcovered, covered);
7857 cpumask_and(tmpmask, notcovered, cpu_map);
7858 cpumask_and(tmpmask, tmpmask, domainspan);
7859 if (cpumask_empty(tmpmask))
7862 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
7863 if (cpumask_empty(tmpmask))
7866 sg = kmalloc_node(sizeof(struct sched_group) +
7871 "Can not alloc domain group for node %d\n", j);
7874 sg->__cpu_power = 0;
7875 cpumask_copy(sched_group_cpus(sg), tmpmask);
7876 sg->next = prev->next;
7877 cpumask_or(covered, covered, tmpmask);
7884 /* Calculate CPU power for physical packages and nodes */
7885 #ifdef CONFIG_SCHED_SMT
7886 for_each_cpu(i, cpu_map) {
7887 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
7889 init_sched_groups_power(i, sd);
7892 #ifdef CONFIG_SCHED_MC
7893 for_each_cpu(i, cpu_map) {
7894 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
7896 init_sched_groups_power(i, sd);
7900 for_each_cpu(i, cpu_map) {
7901 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
7903 init_sched_groups_power(i, sd);
7907 for (i = 0; i < nr_node_ids; i++)
7908 init_numa_sched_groups_power(sched_group_nodes[i]);
7911 struct sched_group *sg;
7913 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7915 init_numa_sched_groups_power(sg);
7919 /* Attach the domains */
7920 for_each_cpu(i, cpu_map) {
7921 struct sched_domain *sd;
7922 #ifdef CONFIG_SCHED_SMT
7923 sd = &per_cpu(cpu_domains, i).sd;
7924 #elif defined(CONFIG_SCHED_MC)
7925 sd = &per_cpu(core_domains, i).sd;
7927 sd = &per_cpu(phys_domains, i).sd;
7929 cpu_attach_domain(sd, rd, i);
7935 free_cpumask_var(tmpmask);
7937 free_cpumask_var(send_covered);
7939 free_cpumask_var(this_core_map);
7940 free_this_sibling_map:
7941 free_cpumask_var(this_sibling_map);
7943 free_cpumask_var(nodemask);
7946 free_cpumask_var(notcovered);
7948 free_cpumask_var(covered);
7950 free_cpumask_var(domainspan);
7957 kfree(sched_group_nodes);
7963 free_sched_groups(cpu_map, tmpmask);
7964 free_rootdomain(rd);
7969 static int build_sched_domains(const struct cpumask *cpu_map)
7971 return __build_sched_domains(cpu_map, NULL);
7974 static struct cpumask *doms_cur; /* current sched domains */
7975 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7976 static struct sched_domain_attr *dattr_cur;
7977 /* attribues of custom domains in 'doms_cur' */
7980 * Special case: If a kmalloc of a doms_cur partition (array of
7981 * cpumask) fails, then fallback to a single sched domain,
7982 * as determined by the single cpumask fallback_doms.
7984 static cpumask_var_t fallback_doms;
7987 * arch_update_cpu_topology lets virtualized architectures update the
7988 * cpu core maps. It is supposed to return 1 if the topology changed
7989 * or 0 if it stayed the same.
7991 int __attribute__((weak)) arch_update_cpu_topology(void)
7997 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7998 * For now this just excludes isolated cpus, but could be used to
7999 * exclude other special cases in the future.
8001 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8005 arch_update_cpu_topology();
8007 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8009 doms_cur = fallback_doms;
8010 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8012 err = build_sched_domains(doms_cur);
8013 register_sched_domain_sysctl();
8018 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8019 struct cpumask *tmpmask)
8021 free_sched_groups(cpu_map, tmpmask);
8025 * Detach sched domains from a group of cpus specified in cpu_map
8026 * These cpus will now be attached to the NULL domain
8028 static void detach_destroy_domains(const struct cpumask *cpu_map)
8030 /* Save because hotplug lock held. */
8031 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8034 for_each_cpu(i, cpu_map)
8035 cpu_attach_domain(NULL, &def_root_domain, i);
8036 synchronize_sched();
8037 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8040 /* handle null as "default" */
8041 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8042 struct sched_domain_attr *new, int idx_new)
8044 struct sched_domain_attr tmp;
8051 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8052 new ? (new + idx_new) : &tmp,
8053 sizeof(struct sched_domain_attr));
8057 * Partition sched domains as specified by the 'ndoms_new'
8058 * cpumasks in the array doms_new[] of cpumasks. This compares
8059 * doms_new[] to the current sched domain partitioning, doms_cur[].
8060 * It destroys each deleted domain and builds each new domain.
8062 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8063 * The masks don't intersect (don't overlap.) We should setup one
8064 * sched domain for each mask. CPUs not in any of the cpumasks will
8065 * not be load balanced. If the same cpumask appears both in the
8066 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8069 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8070 * ownership of it and will kfree it when done with it. If the caller
8071 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8072 * ndoms_new == 1, and partition_sched_domains() will fallback to
8073 * the single partition 'fallback_doms', it also forces the domains
8076 * If doms_new == NULL it will be replaced with cpu_online_mask.
8077 * ndoms_new == 0 is a special case for destroying existing domains,
8078 * and it will not create the default domain.
8080 * Call with hotplug lock held
8082 /* FIXME: Change to struct cpumask *doms_new[] */
8083 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8084 struct sched_domain_attr *dattr_new)
8089 mutex_lock(&sched_domains_mutex);
8091 /* always unregister in case we don't destroy any domains */
8092 unregister_sched_domain_sysctl();
8094 /* Let architecture update cpu core mappings. */
8095 new_topology = arch_update_cpu_topology();
8097 n = doms_new ? ndoms_new : 0;
8099 /* Destroy deleted domains */
8100 for (i = 0; i < ndoms_cur; i++) {
8101 for (j = 0; j < n && !new_topology; j++) {
8102 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8103 && dattrs_equal(dattr_cur, i, dattr_new, j))
8106 /* no match - a current sched domain not in new doms_new[] */
8107 detach_destroy_domains(doms_cur + i);
8112 if (doms_new == NULL) {
8114 doms_new = fallback_doms;
8115 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8116 WARN_ON_ONCE(dattr_new);
8119 /* Build new domains */
8120 for (i = 0; i < ndoms_new; i++) {
8121 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8122 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8123 && dattrs_equal(dattr_new, i, dattr_cur, j))
8126 /* no match - add a new doms_new */
8127 __build_sched_domains(doms_new + i,
8128 dattr_new ? dattr_new + i : NULL);
8133 /* Remember the new sched domains */
8134 if (doms_cur != fallback_doms)
8136 kfree(dattr_cur); /* kfree(NULL) is safe */
8137 doms_cur = doms_new;
8138 dattr_cur = dattr_new;
8139 ndoms_cur = ndoms_new;
8141 register_sched_domain_sysctl();
8143 mutex_unlock(&sched_domains_mutex);
8146 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8147 static void arch_reinit_sched_domains(void)
8151 /* Destroy domains first to force the rebuild */
8152 partition_sched_domains(0, NULL, NULL);
8154 rebuild_sched_domains();
8158 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8160 unsigned int level = 0;
8162 if (sscanf(buf, "%u", &level) != 1)
8166 * level is always be positive so don't check for
8167 * level < POWERSAVINGS_BALANCE_NONE which is 0
8168 * What happens on 0 or 1 byte write,
8169 * need to check for count as well?
8172 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8176 sched_smt_power_savings = level;
8178 sched_mc_power_savings = level;
8180 arch_reinit_sched_domains();
8185 #ifdef CONFIG_SCHED_MC
8186 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8189 return sprintf(page, "%u\n", sched_mc_power_savings);
8191 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8192 const char *buf, size_t count)
8194 return sched_power_savings_store(buf, count, 0);
8196 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8197 sched_mc_power_savings_show,
8198 sched_mc_power_savings_store);
8201 #ifdef CONFIG_SCHED_SMT
8202 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8205 return sprintf(page, "%u\n", sched_smt_power_savings);
8207 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8208 const char *buf, size_t count)
8210 return sched_power_savings_store(buf, count, 1);
8212 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8213 sched_smt_power_savings_show,
8214 sched_smt_power_savings_store);
8217 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8221 #ifdef CONFIG_SCHED_SMT
8223 err = sysfs_create_file(&cls->kset.kobj,
8224 &attr_sched_smt_power_savings.attr);
8226 #ifdef CONFIG_SCHED_MC
8227 if (!err && mc_capable())
8228 err = sysfs_create_file(&cls->kset.kobj,
8229 &attr_sched_mc_power_savings.attr);
8233 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8235 #ifndef CONFIG_CPUSETS
8237 * Add online and remove offline CPUs from the scheduler domains.
8238 * When cpusets are enabled they take over this function.
8240 static int update_sched_domains(struct notifier_block *nfb,
8241 unsigned long action, void *hcpu)
8245 case CPU_ONLINE_FROZEN:
8247 case CPU_DEAD_FROZEN:
8248 partition_sched_domains(1, NULL, NULL);
8257 static int update_runtime(struct notifier_block *nfb,
8258 unsigned long action, void *hcpu)
8260 int cpu = (int)(long)hcpu;
8263 case CPU_DOWN_PREPARE:
8264 case CPU_DOWN_PREPARE_FROZEN:
8265 disable_runtime(cpu_rq(cpu));
8268 case CPU_DOWN_FAILED:
8269 case CPU_DOWN_FAILED_FROZEN:
8271 case CPU_ONLINE_FROZEN:
8272 enable_runtime(cpu_rq(cpu));
8280 void __init sched_init_smp(void)
8282 cpumask_var_t non_isolated_cpus;
8284 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8286 #if defined(CONFIG_NUMA)
8287 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8289 BUG_ON(sched_group_nodes_bycpu == NULL);
8292 mutex_lock(&sched_domains_mutex);
8293 arch_init_sched_domains(cpu_online_mask);
8294 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8295 if (cpumask_empty(non_isolated_cpus))
8296 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8297 mutex_unlock(&sched_domains_mutex);
8300 #ifndef CONFIG_CPUSETS
8301 /* XXX: Theoretical race here - CPU may be hotplugged now */
8302 hotcpu_notifier(update_sched_domains, 0);
8305 /* RT runtime code needs to handle some hotplug events */
8306 hotcpu_notifier(update_runtime, 0);
8310 /* Move init over to a non-isolated CPU */
8311 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8313 sched_init_granularity();
8314 free_cpumask_var(non_isolated_cpus);
8316 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8317 init_sched_rt_class();
8320 void __init sched_init_smp(void)
8322 sched_init_granularity();
8324 #endif /* CONFIG_SMP */
8326 int in_sched_functions(unsigned long addr)
8328 return in_lock_functions(addr) ||
8329 (addr >= (unsigned long)__sched_text_start
8330 && addr < (unsigned long)__sched_text_end);
8333 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8335 cfs_rq->tasks_timeline = RB_ROOT;
8336 INIT_LIST_HEAD(&cfs_rq->tasks);
8337 #ifdef CONFIG_FAIR_GROUP_SCHED
8340 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8343 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8345 struct rt_prio_array *array;
8348 array = &rt_rq->active;
8349 for (i = 0; i < MAX_RT_PRIO; i++) {
8350 INIT_LIST_HEAD(array->queue + i);
8351 __clear_bit(i, array->bitmap);
8353 /* delimiter for bitsearch: */
8354 __set_bit(MAX_RT_PRIO, array->bitmap);
8356 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8357 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8359 rt_rq->highest_prio.next = MAX_RT_PRIO;
8363 rt_rq->rt_nr_migratory = 0;
8364 rt_rq->overloaded = 0;
8365 plist_head_init(&rq->rt.pushable_tasks, &rq->lock);
8369 rt_rq->rt_throttled = 0;
8370 rt_rq->rt_runtime = 0;
8371 spin_lock_init(&rt_rq->rt_runtime_lock);
8373 #ifdef CONFIG_RT_GROUP_SCHED
8374 rt_rq->rt_nr_boosted = 0;
8379 #ifdef CONFIG_FAIR_GROUP_SCHED
8380 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8381 struct sched_entity *se, int cpu, int add,
8382 struct sched_entity *parent)
8384 struct rq *rq = cpu_rq(cpu);
8385 tg->cfs_rq[cpu] = cfs_rq;
8386 init_cfs_rq(cfs_rq, rq);
8389 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8392 /* se could be NULL for init_task_group */
8397 se->cfs_rq = &rq->cfs;
8399 se->cfs_rq = parent->my_q;
8402 se->load.weight = tg->shares;
8403 se->load.inv_weight = 0;
8404 se->parent = parent;
8408 #ifdef CONFIG_RT_GROUP_SCHED
8409 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8410 struct sched_rt_entity *rt_se, int cpu, int add,
8411 struct sched_rt_entity *parent)
8413 struct rq *rq = cpu_rq(cpu);
8415 tg->rt_rq[cpu] = rt_rq;
8416 init_rt_rq(rt_rq, rq);
8418 rt_rq->rt_se = rt_se;
8419 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8421 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8423 tg->rt_se[cpu] = rt_se;
8428 rt_se->rt_rq = &rq->rt;
8430 rt_se->rt_rq = parent->my_q;
8432 rt_se->my_q = rt_rq;
8433 rt_se->parent = parent;
8434 INIT_LIST_HEAD(&rt_se->run_list);
8438 void __init sched_init(void)
8441 unsigned long alloc_size = 0, ptr;
8443 #ifdef CONFIG_FAIR_GROUP_SCHED
8444 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8446 #ifdef CONFIG_RT_GROUP_SCHED
8447 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8449 #ifdef CONFIG_USER_SCHED
8453 * As sched_init() is called before page_alloc is setup,
8454 * we use alloc_bootmem().
8457 ptr = (unsigned long)alloc_bootmem(alloc_size);
8459 #ifdef CONFIG_FAIR_GROUP_SCHED
8460 init_task_group.se = (struct sched_entity **)ptr;
8461 ptr += nr_cpu_ids * sizeof(void **);
8463 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8464 ptr += nr_cpu_ids * sizeof(void **);
8466 #ifdef CONFIG_USER_SCHED
8467 root_task_group.se = (struct sched_entity **)ptr;
8468 ptr += nr_cpu_ids * sizeof(void **);
8470 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8471 ptr += nr_cpu_ids * sizeof(void **);
8472 #endif /* CONFIG_USER_SCHED */
8473 #endif /* CONFIG_FAIR_GROUP_SCHED */
8474 #ifdef CONFIG_RT_GROUP_SCHED
8475 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8476 ptr += nr_cpu_ids * sizeof(void **);
8478 init_task_group.rt_rq = (struct rt_rq **)ptr;
8479 ptr += nr_cpu_ids * sizeof(void **);
8481 #ifdef CONFIG_USER_SCHED
8482 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8483 ptr += nr_cpu_ids * sizeof(void **);
8485 root_task_group.rt_rq = (struct rt_rq **)ptr;
8486 ptr += nr_cpu_ids * sizeof(void **);
8487 #endif /* CONFIG_USER_SCHED */
8488 #endif /* CONFIG_RT_GROUP_SCHED */
8492 init_defrootdomain();
8495 init_rt_bandwidth(&def_rt_bandwidth,
8496 global_rt_period(), global_rt_runtime());
8498 #ifdef CONFIG_RT_GROUP_SCHED
8499 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8500 global_rt_period(), global_rt_runtime());
8501 #ifdef CONFIG_USER_SCHED
8502 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8503 global_rt_period(), RUNTIME_INF);
8504 #endif /* CONFIG_USER_SCHED */
8505 #endif /* CONFIG_RT_GROUP_SCHED */
8507 #ifdef CONFIG_GROUP_SCHED
8508 list_add(&init_task_group.list, &task_groups);
8509 INIT_LIST_HEAD(&init_task_group.children);
8511 #ifdef CONFIG_USER_SCHED
8512 INIT_LIST_HEAD(&root_task_group.children);
8513 init_task_group.parent = &root_task_group;
8514 list_add(&init_task_group.siblings, &root_task_group.children);
8515 #endif /* CONFIG_USER_SCHED */
8516 #endif /* CONFIG_GROUP_SCHED */
8518 for_each_possible_cpu(i) {
8522 spin_lock_init(&rq->lock);
8524 init_cfs_rq(&rq->cfs, rq);
8525 init_rt_rq(&rq->rt, rq);
8526 #ifdef CONFIG_FAIR_GROUP_SCHED
8527 init_task_group.shares = init_task_group_load;
8528 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8529 #ifdef CONFIG_CGROUP_SCHED
8531 * How much cpu bandwidth does init_task_group get?
8533 * In case of task-groups formed thr' the cgroup filesystem, it
8534 * gets 100% of the cpu resources in the system. This overall
8535 * system cpu resource is divided among the tasks of
8536 * init_task_group and its child task-groups in a fair manner,
8537 * based on each entity's (task or task-group's) weight
8538 * (se->load.weight).
8540 * In other words, if init_task_group has 10 tasks of weight
8541 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8542 * then A0's share of the cpu resource is:
8544 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8546 * We achieve this by letting init_task_group's tasks sit
8547 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8549 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8550 #elif defined CONFIG_USER_SCHED
8551 root_task_group.shares = NICE_0_LOAD;
8552 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8554 * In case of task-groups formed thr' the user id of tasks,
8555 * init_task_group represents tasks belonging to root user.
8556 * Hence it forms a sibling of all subsequent groups formed.
8557 * In this case, init_task_group gets only a fraction of overall
8558 * system cpu resource, based on the weight assigned to root
8559 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8560 * by letting tasks of init_task_group sit in a separate cfs_rq
8561 * (init_cfs_rq) and having one entity represent this group of
8562 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8564 init_tg_cfs_entry(&init_task_group,
8565 &per_cpu(init_cfs_rq, i),
8566 &per_cpu(init_sched_entity, i), i, 1,
8567 root_task_group.se[i]);
8570 #endif /* CONFIG_FAIR_GROUP_SCHED */
8572 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8573 #ifdef CONFIG_RT_GROUP_SCHED
8574 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8575 #ifdef CONFIG_CGROUP_SCHED
8576 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8577 #elif defined CONFIG_USER_SCHED
8578 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8579 init_tg_rt_entry(&init_task_group,
8580 &per_cpu(init_rt_rq, i),
8581 &per_cpu(init_sched_rt_entity, i), i, 1,
8582 root_task_group.rt_se[i]);
8586 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8587 rq->cpu_load[j] = 0;
8591 rq->active_balance = 0;
8592 rq->next_balance = jiffies;
8596 rq->migration_thread = NULL;
8597 INIT_LIST_HEAD(&rq->migration_queue);
8598 rq_attach_root(rq, &def_root_domain);
8601 atomic_set(&rq->nr_iowait, 0);
8604 set_load_weight(&init_task);
8606 #ifdef CONFIG_PREEMPT_NOTIFIERS
8607 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8611 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8614 #ifdef CONFIG_RT_MUTEXES
8615 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8619 * The boot idle thread does lazy MMU switching as well:
8621 atomic_inc(&init_mm.mm_count);
8622 enter_lazy_tlb(&init_mm, current);
8625 * Make us the idle thread. Technically, schedule() should not be
8626 * called from this thread, however somewhere below it might be,
8627 * but because we are the idle thread, we just pick up running again
8628 * when this runqueue becomes "idle".
8630 init_idle(current, smp_processor_id());
8632 * During early bootup we pretend to be a normal task:
8634 current->sched_class = &fair_sched_class;
8636 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8637 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
8640 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
8642 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8645 scheduler_running = 1;
8648 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8649 void __might_sleep(char *file, int line)
8652 static unsigned long prev_jiffy; /* ratelimiting */
8654 if ((!in_atomic() && !irqs_disabled()) ||
8655 system_state != SYSTEM_RUNNING || oops_in_progress)
8657 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8659 prev_jiffy = jiffies;
8662 "BUG: sleeping function called from invalid context at %s:%d\n",
8665 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8666 in_atomic(), irqs_disabled(),
8667 current->pid, current->comm);
8669 debug_show_held_locks(current);
8670 if (irqs_disabled())
8671 print_irqtrace_events(current);
8675 EXPORT_SYMBOL(__might_sleep);
8678 #ifdef CONFIG_MAGIC_SYSRQ
8679 static void normalize_task(struct rq *rq, struct task_struct *p)
8683 update_rq_clock(rq);
8684 on_rq = p->se.on_rq;
8686 deactivate_task(rq, p, 0);
8687 __setscheduler(rq, p, SCHED_NORMAL, 0);
8689 activate_task(rq, p, 0);
8690 resched_task(rq->curr);
8694 void normalize_rt_tasks(void)
8696 struct task_struct *g, *p;
8697 unsigned long flags;
8700 read_lock_irqsave(&tasklist_lock, flags);
8701 do_each_thread(g, p) {
8703 * Only normalize user tasks:
8708 p->se.exec_start = 0;
8709 #ifdef CONFIG_SCHEDSTATS
8710 p->se.wait_start = 0;
8711 p->se.sleep_start = 0;
8712 p->se.block_start = 0;
8717 * Renice negative nice level userspace
8720 if (TASK_NICE(p) < 0 && p->mm)
8721 set_user_nice(p, 0);
8725 spin_lock(&p->pi_lock);
8726 rq = __task_rq_lock(p);
8728 normalize_task(rq, p);
8730 __task_rq_unlock(rq);
8731 spin_unlock(&p->pi_lock);
8732 } while_each_thread(g, p);
8734 read_unlock_irqrestore(&tasklist_lock, flags);
8737 #endif /* CONFIG_MAGIC_SYSRQ */
8741 * These functions are only useful for the IA64 MCA handling.
8743 * They can only be called when the whole system has been
8744 * stopped - every CPU needs to be quiescent, and no scheduling
8745 * activity can take place. Using them for anything else would
8746 * be a serious bug, and as a result, they aren't even visible
8747 * under any other configuration.
8751 * curr_task - return the current task for a given cpu.
8752 * @cpu: the processor in question.
8754 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8756 struct task_struct *curr_task(int cpu)
8758 return cpu_curr(cpu);
8762 * set_curr_task - set the current task for a given cpu.
8763 * @cpu: the processor in question.
8764 * @p: the task pointer to set.
8766 * Description: This function must only be used when non-maskable interrupts
8767 * are serviced on a separate stack. It allows the architecture to switch the
8768 * notion of the current task on a cpu in a non-blocking manner. This function
8769 * must be called with all CPU's synchronized, and interrupts disabled, the
8770 * and caller must save the original value of the current task (see
8771 * curr_task() above) and restore that value before reenabling interrupts and
8772 * re-starting the system.
8774 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8776 void set_curr_task(int cpu, struct task_struct *p)
8783 #ifdef CONFIG_FAIR_GROUP_SCHED
8784 static void free_fair_sched_group(struct task_group *tg)
8788 for_each_possible_cpu(i) {
8790 kfree(tg->cfs_rq[i]);
8800 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8802 struct cfs_rq *cfs_rq;
8803 struct sched_entity *se;
8807 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8810 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8814 tg->shares = NICE_0_LOAD;
8816 for_each_possible_cpu(i) {
8819 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8820 GFP_KERNEL, cpu_to_node(i));
8824 se = kzalloc_node(sizeof(struct sched_entity),
8825 GFP_KERNEL, cpu_to_node(i));
8829 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8838 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8840 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8841 &cpu_rq(cpu)->leaf_cfs_rq_list);
8844 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8846 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8848 #else /* !CONFG_FAIR_GROUP_SCHED */
8849 static inline void free_fair_sched_group(struct task_group *tg)
8854 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8859 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8863 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8866 #endif /* CONFIG_FAIR_GROUP_SCHED */
8868 #ifdef CONFIG_RT_GROUP_SCHED
8869 static void free_rt_sched_group(struct task_group *tg)
8873 destroy_rt_bandwidth(&tg->rt_bandwidth);
8875 for_each_possible_cpu(i) {
8877 kfree(tg->rt_rq[i]);
8879 kfree(tg->rt_se[i]);
8887 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8889 struct rt_rq *rt_rq;
8890 struct sched_rt_entity *rt_se;
8894 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8897 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8901 init_rt_bandwidth(&tg->rt_bandwidth,
8902 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8904 for_each_possible_cpu(i) {
8907 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8908 GFP_KERNEL, cpu_to_node(i));
8912 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8913 GFP_KERNEL, cpu_to_node(i));
8917 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8926 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8928 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8929 &cpu_rq(cpu)->leaf_rt_rq_list);
8932 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8934 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8936 #else /* !CONFIG_RT_GROUP_SCHED */
8937 static inline void free_rt_sched_group(struct task_group *tg)
8942 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8947 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8951 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8954 #endif /* CONFIG_RT_GROUP_SCHED */
8956 #ifdef CONFIG_GROUP_SCHED
8957 static void free_sched_group(struct task_group *tg)
8959 free_fair_sched_group(tg);
8960 free_rt_sched_group(tg);
8964 /* allocate runqueue etc for a new task group */
8965 struct task_group *sched_create_group(struct task_group *parent)
8967 struct task_group *tg;
8968 unsigned long flags;
8971 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8973 return ERR_PTR(-ENOMEM);
8975 if (!alloc_fair_sched_group(tg, parent))
8978 if (!alloc_rt_sched_group(tg, parent))
8981 spin_lock_irqsave(&task_group_lock, flags);
8982 for_each_possible_cpu(i) {
8983 register_fair_sched_group(tg, i);
8984 register_rt_sched_group(tg, i);
8986 list_add_rcu(&tg->list, &task_groups);
8988 WARN_ON(!parent); /* root should already exist */
8990 tg->parent = parent;
8991 INIT_LIST_HEAD(&tg->children);
8992 list_add_rcu(&tg->siblings, &parent->children);
8993 spin_unlock_irqrestore(&task_group_lock, flags);
8998 free_sched_group(tg);
8999 return ERR_PTR(-ENOMEM);
9002 /* rcu callback to free various structures associated with a task group */
9003 static void free_sched_group_rcu(struct rcu_head *rhp)
9005 /* now it should be safe to free those cfs_rqs */
9006 free_sched_group(container_of(rhp, struct task_group, rcu));
9009 /* Destroy runqueue etc associated with a task group */
9010 void sched_destroy_group(struct task_group *tg)
9012 unsigned long flags;
9015 spin_lock_irqsave(&task_group_lock, flags);
9016 for_each_possible_cpu(i) {
9017 unregister_fair_sched_group(tg, i);
9018 unregister_rt_sched_group(tg, i);
9020 list_del_rcu(&tg->list);
9021 list_del_rcu(&tg->siblings);
9022 spin_unlock_irqrestore(&task_group_lock, flags);
9024 /* wait for possible concurrent references to cfs_rqs complete */
9025 call_rcu(&tg->rcu, free_sched_group_rcu);
9028 /* change task's runqueue when it moves between groups.
9029 * The caller of this function should have put the task in its new group
9030 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9031 * reflect its new group.
9033 void sched_move_task(struct task_struct *tsk)
9036 unsigned long flags;
9039 rq = task_rq_lock(tsk, &flags);
9041 update_rq_clock(rq);
9043 running = task_current(rq, tsk);
9044 on_rq = tsk->se.on_rq;
9047 dequeue_task(rq, tsk, 0);
9048 if (unlikely(running))
9049 tsk->sched_class->put_prev_task(rq, tsk);
9051 set_task_rq(tsk, task_cpu(tsk));
9053 #ifdef CONFIG_FAIR_GROUP_SCHED
9054 if (tsk->sched_class->moved_group)
9055 tsk->sched_class->moved_group(tsk);
9058 if (unlikely(running))
9059 tsk->sched_class->set_curr_task(rq);
9061 enqueue_task(rq, tsk, 0);
9063 task_rq_unlock(rq, &flags);
9065 #endif /* CONFIG_GROUP_SCHED */
9067 #ifdef CONFIG_FAIR_GROUP_SCHED
9068 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9070 struct cfs_rq *cfs_rq = se->cfs_rq;
9075 dequeue_entity(cfs_rq, se, 0);
9077 se->load.weight = shares;
9078 se->load.inv_weight = 0;
9081 enqueue_entity(cfs_rq, se, 0);
9084 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9086 struct cfs_rq *cfs_rq = se->cfs_rq;
9087 struct rq *rq = cfs_rq->rq;
9088 unsigned long flags;
9090 spin_lock_irqsave(&rq->lock, flags);
9091 __set_se_shares(se, shares);
9092 spin_unlock_irqrestore(&rq->lock, flags);
9095 static DEFINE_MUTEX(shares_mutex);
9097 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9100 unsigned long flags;
9103 * We can't change the weight of the root cgroup.
9108 if (shares < MIN_SHARES)
9109 shares = MIN_SHARES;
9110 else if (shares > MAX_SHARES)
9111 shares = MAX_SHARES;
9113 mutex_lock(&shares_mutex);
9114 if (tg->shares == shares)
9117 spin_lock_irqsave(&task_group_lock, flags);
9118 for_each_possible_cpu(i)
9119 unregister_fair_sched_group(tg, i);
9120 list_del_rcu(&tg->siblings);
9121 spin_unlock_irqrestore(&task_group_lock, flags);
9123 /* wait for any ongoing reference to this group to finish */
9124 synchronize_sched();
9127 * Now we are free to modify the group's share on each cpu
9128 * w/o tripping rebalance_share or load_balance_fair.
9130 tg->shares = shares;
9131 for_each_possible_cpu(i) {
9135 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9136 set_se_shares(tg->se[i], shares);
9140 * Enable load balance activity on this group, by inserting it back on
9141 * each cpu's rq->leaf_cfs_rq_list.
9143 spin_lock_irqsave(&task_group_lock, flags);
9144 for_each_possible_cpu(i)
9145 register_fair_sched_group(tg, i);
9146 list_add_rcu(&tg->siblings, &tg->parent->children);
9147 spin_unlock_irqrestore(&task_group_lock, flags);
9149 mutex_unlock(&shares_mutex);
9153 unsigned long sched_group_shares(struct task_group *tg)
9159 #ifdef CONFIG_RT_GROUP_SCHED
9161 * Ensure that the real time constraints are schedulable.
9163 static DEFINE_MUTEX(rt_constraints_mutex);
9165 static unsigned long to_ratio(u64 period, u64 runtime)
9167 if (runtime == RUNTIME_INF)
9170 return div64_u64(runtime << 20, period);
9173 /* Must be called with tasklist_lock held */
9174 static inline int tg_has_rt_tasks(struct task_group *tg)
9176 struct task_struct *g, *p;
9178 do_each_thread(g, p) {
9179 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9181 } while_each_thread(g, p);
9186 struct rt_schedulable_data {
9187 struct task_group *tg;
9192 static int tg_schedulable(struct task_group *tg, void *data)
9194 struct rt_schedulable_data *d = data;
9195 struct task_group *child;
9196 unsigned long total, sum = 0;
9197 u64 period, runtime;
9199 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9200 runtime = tg->rt_bandwidth.rt_runtime;
9203 period = d->rt_period;
9204 runtime = d->rt_runtime;
9207 #ifdef CONFIG_USER_SCHED
9208 if (tg == &root_task_group) {
9209 period = global_rt_period();
9210 runtime = global_rt_runtime();
9215 * Cannot have more runtime than the period.
9217 if (runtime > period && runtime != RUNTIME_INF)
9221 * Ensure we don't starve existing RT tasks.
9223 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9226 total = to_ratio(period, runtime);
9229 * Nobody can have more than the global setting allows.
9231 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9235 * The sum of our children's runtime should not exceed our own.
9237 list_for_each_entry_rcu(child, &tg->children, siblings) {
9238 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9239 runtime = child->rt_bandwidth.rt_runtime;
9241 if (child == d->tg) {
9242 period = d->rt_period;
9243 runtime = d->rt_runtime;
9246 sum += to_ratio(period, runtime);
9255 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9257 struct rt_schedulable_data data = {
9259 .rt_period = period,
9260 .rt_runtime = runtime,
9263 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9266 static int tg_set_bandwidth(struct task_group *tg,
9267 u64 rt_period, u64 rt_runtime)
9271 mutex_lock(&rt_constraints_mutex);
9272 read_lock(&tasklist_lock);
9273 err = __rt_schedulable(tg, rt_period, rt_runtime);
9277 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9278 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9279 tg->rt_bandwidth.rt_runtime = rt_runtime;
9281 for_each_possible_cpu(i) {
9282 struct rt_rq *rt_rq = tg->rt_rq[i];
9284 spin_lock(&rt_rq->rt_runtime_lock);
9285 rt_rq->rt_runtime = rt_runtime;
9286 spin_unlock(&rt_rq->rt_runtime_lock);
9288 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9290 read_unlock(&tasklist_lock);
9291 mutex_unlock(&rt_constraints_mutex);
9296 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9298 u64 rt_runtime, rt_period;
9300 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9301 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9302 if (rt_runtime_us < 0)
9303 rt_runtime = RUNTIME_INF;
9305 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9308 long sched_group_rt_runtime(struct task_group *tg)
9312 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9315 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9316 do_div(rt_runtime_us, NSEC_PER_USEC);
9317 return rt_runtime_us;
9320 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9322 u64 rt_runtime, rt_period;
9324 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9325 rt_runtime = tg->rt_bandwidth.rt_runtime;
9330 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9333 long sched_group_rt_period(struct task_group *tg)
9337 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9338 do_div(rt_period_us, NSEC_PER_USEC);
9339 return rt_period_us;
9342 static int sched_rt_global_constraints(void)
9344 u64 runtime, period;
9347 if (sysctl_sched_rt_period <= 0)
9350 runtime = global_rt_runtime();
9351 period = global_rt_period();
9354 * Sanity check on the sysctl variables.
9356 if (runtime > period && runtime != RUNTIME_INF)
9359 mutex_lock(&rt_constraints_mutex);
9360 read_lock(&tasklist_lock);
9361 ret = __rt_schedulable(NULL, 0, 0);
9362 read_unlock(&tasklist_lock);
9363 mutex_unlock(&rt_constraints_mutex);
9368 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9370 /* Don't accept realtime tasks when there is no way for them to run */
9371 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9377 #else /* !CONFIG_RT_GROUP_SCHED */
9378 static int sched_rt_global_constraints(void)
9380 unsigned long flags;
9383 if (sysctl_sched_rt_period <= 0)
9386 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9387 for_each_possible_cpu(i) {
9388 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9390 spin_lock(&rt_rq->rt_runtime_lock);
9391 rt_rq->rt_runtime = global_rt_runtime();
9392 spin_unlock(&rt_rq->rt_runtime_lock);
9394 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9398 #endif /* CONFIG_RT_GROUP_SCHED */
9400 int sched_rt_handler(struct ctl_table *table, int write,
9401 struct file *filp, void __user *buffer, size_t *lenp,
9405 int old_period, old_runtime;
9406 static DEFINE_MUTEX(mutex);
9409 old_period = sysctl_sched_rt_period;
9410 old_runtime = sysctl_sched_rt_runtime;
9412 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9414 if (!ret && write) {
9415 ret = sched_rt_global_constraints();
9417 sysctl_sched_rt_period = old_period;
9418 sysctl_sched_rt_runtime = old_runtime;
9420 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9421 def_rt_bandwidth.rt_period =
9422 ns_to_ktime(global_rt_period());
9425 mutex_unlock(&mutex);
9430 #ifdef CONFIG_CGROUP_SCHED
9432 /* return corresponding task_group object of a cgroup */
9433 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9435 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9436 struct task_group, css);
9439 static struct cgroup_subsys_state *
9440 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9442 struct task_group *tg, *parent;
9444 if (!cgrp->parent) {
9445 /* This is early initialization for the top cgroup */
9446 return &init_task_group.css;
9449 parent = cgroup_tg(cgrp->parent);
9450 tg = sched_create_group(parent);
9452 return ERR_PTR(-ENOMEM);
9458 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9460 struct task_group *tg = cgroup_tg(cgrp);
9462 sched_destroy_group(tg);
9466 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9467 struct task_struct *tsk)
9469 #ifdef CONFIG_RT_GROUP_SCHED
9470 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9473 /* We don't support RT-tasks being in separate groups */
9474 if (tsk->sched_class != &fair_sched_class)
9482 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9483 struct cgroup *old_cont, struct task_struct *tsk)
9485 sched_move_task(tsk);
9488 #ifdef CONFIG_FAIR_GROUP_SCHED
9489 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9492 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9495 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9497 struct task_group *tg = cgroup_tg(cgrp);
9499 return (u64) tg->shares;
9501 #endif /* CONFIG_FAIR_GROUP_SCHED */
9503 #ifdef CONFIG_RT_GROUP_SCHED
9504 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9507 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9510 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9512 return sched_group_rt_runtime(cgroup_tg(cgrp));
9515 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9518 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9521 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9523 return sched_group_rt_period(cgroup_tg(cgrp));
9525 #endif /* CONFIG_RT_GROUP_SCHED */
9527 static struct cftype cpu_files[] = {
9528 #ifdef CONFIG_FAIR_GROUP_SCHED
9531 .read_u64 = cpu_shares_read_u64,
9532 .write_u64 = cpu_shares_write_u64,
9535 #ifdef CONFIG_RT_GROUP_SCHED
9537 .name = "rt_runtime_us",
9538 .read_s64 = cpu_rt_runtime_read,
9539 .write_s64 = cpu_rt_runtime_write,
9542 .name = "rt_period_us",
9543 .read_u64 = cpu_rt_period_read_uint,
9544 .write_u64 = cpu_rt_period_write_uint,
9549 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9551 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9554 struct cgroup_subsys cpu_cgroup_subsys = {
9556 .create = cpu_cgroup_create,
9557 .destroy = cpu_cgroup_destroy,
9558 .can_attach = cpu_cgroup_can_attach,
9559 .attach = cpu_cgroup_attach,
9560 .populate = cpu_cgroup_populate,
9561 .subsys_id = cpu_cgroup_subsys_id,
9565 #endif /* CONFIG_CGROUP_SCHED */
9567 #ifdef CONFIG_CGROUP_CPUACCT
9570 * CPU accounting code for task groups.
9572 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9573 * (balbir@in.ibm.com).
9576 /* track cpu usage of a group of tasks and its child groups */
9578 struct cgroup_subsys_state css;
9579 /* cpuusage holds pointer to a u64-type object on every cpu */
9581 struct cpuacct *parent;
9584 struct cgroup_subsys cpuacct_subsys;
9586 /* return cpu accounting group corresponding to this container */
9587 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9589 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9590 struct cpuacct, css);
9593 /* return cpu accounting group to which this task belongs */
9594 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9596 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9597 struct cpuacct, css);
9600 /* create a new cpu accounting group */
9601 static struct cgroup_subsys_state *cpuacct_create(
9602 struct cgroup_subsys *ss, struct cgroup *cgrp)
9604 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9607 return ERR_PTR(-ENOMEM);
9609 ca->cpuusage = alloc_percpu(u64);
9610 if (!ca->cpuusage) {
9612 return ERR_PTR(-ENOMEM);
9616 ca->parent = cgroup_ca(cgrp->parent);
9621 /* destroy an existing cpu accounting group */
9623 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9625 struct cpuacct *ca = cgroup_ca(cgrp);
9627 free_percpu(ca->cpuusage);
9631 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9633 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9636 #ifndef CONFIG_64BIT
9638 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9640 spin_lock_irq(&cpu_rq(cpu)->lock);
9642 spin_unlock_irq(&cpu_rq(cpu)->lock);
9650 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9652 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9654 #ifndef CONFIG_64BIT
9656 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9658 spin_lock_irq(&cpu_rq(cpu)->lock);
9660 spin_unlock_irq(&cpu_rq(cpu)->lock);
9666 /* return total cpu usage (in nanoseconds) of a group */
9667 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9669 struct cpuacct *ca = cgroup_ca(cgrp);
9670 u64 totalcpuusage = 0;
9673 for_each_present_cpu(i)
9674 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9676 return totalcpuusage;
9679 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9682 struct cpuacct *ca = cgroup_ca(cgrp);
9691 for_each_present_cpu(i)
9692 cpuacct_cpuusage_write(ca, i, 0);
9698 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9701 struct cpuacct *ca = cgroup_ca(cgroup);
9705 for_each_present_cpu(i) {
9706 percpu = cpuacct_cpuusage_read(ca, i);
9707 seq_printf(m, "%llu ", (unsigned long long) percpu);
9709 seq_printf(m, "\n");
9713 static struct cftype files[] = {
9716 .read_u64 = cpuusage_read,
9717 .write_u64 = cpuusage_write,
9720 .name = "usage_percpu",
9721 .read_seq_string = cpuacct_percpu_seq_read,
9726 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9728 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9732 * charge this task's execution time to its accounting group.
9734 * called with rq->lock held.
9736 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9741 if (unlikely(!cpuacct_subsys.active))
9744 cpu = task_cpu(tsk);
9747 for (; ca; ca = ca->parent) {
9748 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9749 *cpuusage += cputime;
9753 struct cgroup_subsys cpuacct_subsys = {
9755 .create = cpuacct_create,
9756 .destroy = cpuacct_destroy,
9757 .populate = cpuacct_populate,
9758 .subsys_id = cpuacct_subsys_id,
9760 #endif /* CONFIG_CGROUP_CPUACCT */