4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task);
122 DEFINE_TRACE(sched_wakeup);
123 DEFINE_TRACE(sched_wakeup_new);
124 DEFINE_TRACE(sched_switch);
125 DEFINE_TRACE(sched_migrate_task);
129 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
130 * Since cpu_power is a 'constant', we can use a reciprocal divide.
132 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
134 return reciprocal_divide(load, sg->reciprocal_cpu_power);
138 * Each time a sched group cpu_power is changed,
139 * we must compute its reciprocal value
141 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
143 sg->__cpu_power += val;
144 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
148 static inline int rt_policy(int policy)
150 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
155 static inline int task_has_rt_policy(struct task_struct *p)
157 return rt_policy(p->policy);
161 * This is the priority-queue data structure of the RT scheduling class:
163 struct rt_prio_array {
164 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
165 struct list_head queue[MAX_RT_PRIO];
168 struct rt_bandwidth {
169 /* nests inside the rq lock: */
170 spinlock_t rt_runtime_lock;
173 struct hrtimer rt_period_timer;
176 static struct rt_bandwidth def_rt_bandwidth;
178 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
180 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
182 struct rt_bandwidth *rt_b =
183 container_of(timer, struct rt_bandwidth, rt_period_timer);
189 now = hrtimer_cb_get_time(timer);
190 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
195 idle = do_sched_rt_period_timer(rt_b, overrun);
198 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
202 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
204 rt_b->rt_period = ns_to_ktime(period);
205 rt_b->rt_runtime = runtime;
207 spin_lock_init(&rt_b->rt_runtime_lock);
209 hrtimer_init(&rt_b->rt_period_timer,
210 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
211 rt_b->rt_period_timer.function = sched_rt_period_timer;
212 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_UNLOCKED;
215 static inline int rt_bandwidth_enabled(void)
217 return sysctl_sched_rt_runtime >= 0;
220 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
224 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
227 if (hrtimer_active(&rt_b->rt_period_timer))
230 spin_lock(&rt_b->rt_runtime_lock);
232 if (hrtimer_active(&rt_b->rt_period_timer))
235 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
236 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
237 hrtimer_start_expires(&rt_b->rt_period_timer,
240 spin_unlock(&rt_b->rt_runtime_lock);
243 #ifdef CONFIG_RT_GROUP_SCHED
244 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
246 hrtimer_cancel(&rt_b->rt_period_timer);
251 * sched_domains_mutex serializes calls to arch_init_sched_domains,
252 * detach_destroy_domains and partition_sched_domains.
254 static DEFINE_MUTEX(sched_domains_mutex);
256 #ifdef CONFIG_GROUP_SCHED
258 #include <linux/cgroup.h>
262 static LIST_HEAD(task_groups);
264 /* task group related information */
266 #ifdef CONFIG_CGROUP_SCHED
267 struct cgroup_subsys_state css;
270 #ifdef CONFIG_USER_SCHED
274 #ifdef CONFIG_FAIR_GROUP_SCHED
275 /* schedulable entities of this group on each cpu */
276 struct sched_entity **se;
277 /* runqueue "owned" by this group on each cpu */
278 struct cfs_rq **cfs_rq;
279 unsigned long shares;
282 #ifdef CONFIG_RT_GROUP_SCHED
283 struct sched_rt_entity **rt_se;
284 struct rt_rq **rt_rq;
286 struct rt_bandwidth rt_bandwidth;
290 struct list_head list;
292 struct task_group *parent;
293 struct list_head siblings;
294 struct list_head children;
297 #ifdef CONFIG_USER_SCHED
299 /* Helper function to pass uid information to create_sched_user() */
300 void set_tg_uid(struct user_struct *user)
302 user->tg->uid = user->uid;
307 * Every UID task group (including init_task_group aka UID-0) will
308 * be a child to this group.
310 struct task_group root_task_group;
312 #ifdef CONFIG_FAIR_GROUP_SCHED
313 /* Default task group's sched entity on each cpu */
314 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
315 /* Default task group's cfs_rq on each cpu */
316 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
317 #endif /* CONFIG_FAIR_GROUP_SCHED */
319 #ifdef CONFIG_RT_GROUP_SCHED
320 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
321 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
322 #endif /* CONFIG_RT_GROUP_SCHED */
323 #else /* !CONFIG_USER_SCHED */
324 #define root_task_group init_task_group
325 #endif /* CONFIG_USER_SCHED */
327 /* task_group_lock serializes add/remove of task groups and also changes to
328 * a task group's cpu shares.
330 static DEFINE_SPINLOCK(task_group_lock);
332 #ifdef CONFIG_FAIR_GROUP_SCHED
333 #ifdef CONFIG_USER_SCHED
334 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
335 #else /* !CONFIG_USER_SCHED */
336 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
337 #endif /* CONFIG_USER_SCHED */
340 * A weight of 0 or 1 can cause arithmetics problems.
341 * A weight of a cfs_rq is the sum of weights of which entities
342 * are queued on this cfs_rq, so a weight of a entity should not be
343 * too large, so as the shares value of a task group.
344 * (The default weight is 1024 - so there's no practical
345 * limitation from this.)
348 #define MAX_SHARES (1UL << 18)
350 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
353 /* Default task group.
354 * Every task in system belong to this group at bootup.
356 struct task_group init_task_group;
358 /* return group to which a task belongs */
359 static inline struct task_group *task_group(struct task_struct *p)
361 struct task_group *tg;
363 #ifdef CONFIG_USER_SCHED
365 #elif defined(CONFIG_CGROUP_SCHED)
366 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
367 struct task_group, css);
369 tg = &init_task_group;
374 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
375 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
377 #ifdef CONFIG_FAIR_GROUP_SCHED
378 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
379 p->se.parent = task_group(p)->se[cpu];
382 #ifdef CONFIG_RT_GROUP_SCHED
383 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
384 p->rt.parent = task_group(p)->rt_se[cpu];
390 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
391 static inline struct task_group *task_group(struct task_struct *p)
396 #endif /* CONFIG_GROUP_SCHED */
398 /* CFS-related fields in a runqueue */
400 struct load_weight load;
401 unsigned long nr_running;
406 struct rb_root tasks_timeline;
407 struct rb_node *rb_leftmost;
409 struct list_head tasks;
410 struct list_head *balance_iterator;
413 * 'curr' points to currently running entity on this cfs_rq.
414 * It is set to NULL otherwise (i.e when none are currently running).
416 struct sched_entity *curr, *next, *last;
418 unsigned int nr_spread_over;
420 #ifdef CONFIG_FAIR_GROUP_SCHED
421 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
424 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
425 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
426 * (like users, containers etc.)
428 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
429 * list is used during load balance.
431 struct list_head leaf_cfs_rq_list;
432 struct task_group *tg; /* group that "owns" this runqueue */
436 * the part of load.weight contributed by tasks
438 unsigned long task_weight;
441 * h_load = weight * f(tg)
443 * Where f(tg) is the recursive weight fraction assigned to
446 unsigned long h_load;
449 * this cpu's part of tg->shares
451 unsigned long shares;
454 * load.weight at the time we set shares
456 unsigned long rq_weight;
461 /* Real-Time classes' related field in a runqueue: */
463 struct rt_prio_array active;
464 unsigned long rt_nr_running;
465 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
467 int curr; /* highest queued rt task prio */
468 int next; /* next highest */
472 unsigned long rt_nr_migratory;
474 struct plist_head pushable_tasks;
479 /* Nests inside the rq lock: */
480 spinlock_t rt_runtime_lock;
482 #ifdef CONFIG_RT_GROUP_SCHED
483 unsigned long rt_nr_boosted;
486 struct list_head leaf_rt_rq_list;
487 struct task_group *tg;
488 struct sched_rt_entity *rt_se;
495 * We add the notion of a root-domain which will be used to define per-domain
496 * variables. Each exclusive cpuset essentially defines an island domain by
497 * fully partitioning the member cpus from any other cpuset. Whenever a new
498 * exclusive cpuset is created, we also create and attach a new root-domain
505 cpumask_var_t online;
508 * The "RT overload" flag: it gets set if a CPU has more than
509 * one runnable RT task.
511 cpumask_var_t rto_mask;
514 struct cpupri cpupri;
516 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
518 * Preferred wake up cpu nominated by sched_mc balance that will be
519 * used when most cpus are idle in the system indicating overall very
520 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
522 unsigned int sched_mc_preferred_wakeup_cpu;
527 * By default the system creates a single root-domain with all cpus as
528 * members (mimicking the global state we have today).
530 static struct root_domain def_root_domain;
535 * This is the main, per-CPU runqueue data structure.
537 * Locking rule: those places that want to lock multiple runqueues
538 * (such as the load balancing or the thread migration code), lock
539 * acquire operations must be ordered by ascending &runqueue.
546 * nr_running and cpu_load should be in the same cacheline because
547 * remote CPUs use both these fields when doing load calculation.
549 unsigned long nr_running;
550 #define CPU_LOAD_IDX_MAX 5
551 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
552 unsigned char idle_at_tick;
554 unsigned long last_tick_seen;
555 unsigned char in_nohz_recently;
557 /* capture load from *all* tasks on this cpu: */
558 struct load_weight load;
559 unsigned long nr_load_updates;
565 #ifdef CONFIG_FAIR_GROUP_SCHED
566 /* list of leaf cfs_rq on this cpu: */
567 struct list_head leaf_cfs_rq_list;
569 #ifdef CONFIG_RT_GROUP_SCHED
570 struct list_head leaf_rt_rq_list;
574 * This is part of a global counter where only the total sum
575 * over all CPUs matters. A task can increase this counter on
576 * one CPU and if it got migrated afterwards it may decrease
577 * it on another CPU. Always updated under the runqueue lock:
579 unsigned long nr_uninterruptible;
581 struct task_struct *curr, *idle;
582 unsigned long next_balance;
583 struct mm_struct *prev_mm;
590 struct root_domain *rd;
591 struct sched_domain *sd;
593 /* For active balancing */
596 /* cpu of this runqueue: */
600 unsigned long avg_load_per_task;
602 struct task_struct *migration_thread;
603 struct list_head migration_queue;
606 #ifdef CONFIG_SCHED_HRTICK
608 int hrtick_csd_pending;
609 struct call_single_data hrtick_csd;
611 struct hrtimer hrtick_timer;
614 #ifdef CONFIG_SCHEDSTATS
616 struct sched_info rq_sched_info;
618 /* sys_sched_yield() stats */
619 unsigned int yld_exp_empty;
620 unsigned int yld_act_empty;
621 unsigned int yld_both_empty;
622 unsigned int yld_count;
624 /* schedule() stats */
625 unsigned int sched_switch;
626 unsigned int sched_count;
627 unsigned int sched_goidle;
629 /* try_to_wake_up() stats */
630 unsigned int ttwu_count;
631 unsigned int ttwu_local;
634 unsigned int bkl_count;
638 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
640 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
642 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
645 static inline int cpu_of(struct rq *rq)
655 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
656 * See detach_destroy_domains: synchronize_sched for details.
658 * The domain tree of any CPU may only be accessed from within
659 * preempt-disabled sections.
661 #define for_each_domain(cpu, __sd) \
662 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
664 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
665 #define this_rq() (&__get_cpu_var(runqueues))
666 #define task_rq(p) cpu_rq(task_cpu(p))
667 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
669 static inline void update_rq_clock(struct rq *rq)
671 rq->clock = sched_clock_cpu(cpu_of(rq));
675 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
677 #ifdef CONFIG_SCHED_DEBUG
678 # define const_debug __read_mostly
680 # define const_debug static const
686 * Returns true if the current cpu runqueue is locked.
687 * This interface allows printk to be called with the runqueue lock
688 * held and know whether or not it is OK to wake up the klogd.
690 int runqueue_is_locked(void)
693 struct rq *rq = cpu_rq(cpu);
696 ret = spin_is_locked(&rq->lock);
702 * Debugging: various feature bits
705 #define SCHED_FEAT(name, enabled) \
706 __SCHED_FEAT_##name ,
709 #include "sched_features.h"
714 #define SCHED_FEAT(name, enabled) \
715 (1UL << __SCHED_FEAT_##name) * enabled |
717 const_debug unsigned int sysctl_sched_features =
718 #include "sched_features.h"
723 #ifdef CONFIG_SCHED_DEBUG
724 #define SCHED_FEAT(name, enabled) \
727 static __read_mostly char *sched_feat_names[] = {
728 #include "sched_features.h"
734 static int sched_feat_show(struct seq_file *m, void *v)
738 for (i = 0; sched_feat_names[i]; i++) {
739 if (!(sysctl_sched_features & (1UL << i)))
741 seq_printf(m, "%s ", sched_feat_names[i]);
749 sched_feat_write(struct file *filp, const char __user *ubuf,
750 size_t cnt, loff_t *ppos)
760 if (copy_from_user(&buf, ubuf, cnt))
765 if (strncmp(buf, "NO_", 3) == 0) {
770 for (i = 0; sched_feat_names[i]; i++) {
771 int len = strlen(sched_feat_names[i]);
773 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
775 sysctl_sched_features &= ~(1UL << i);
777 sysctl_sched_features |= (1UL << i);
782 if (!sched_feat_names[i])
790 static int sched_feat_open(struct inode *inode, struct file *filp)
792 return single_open(filp, sched_feat_show, NULL);
795 static struct file_operations sched_feat_fops = {
796 .open = sched_feat_open,
797 .write = sched_feat_write,
800 .release = single_release,
803 static __init int sched_init_debug(void)
805 debugfs_create_file("sched_features", 0644, NULL, NULL,
810 late_initcall(sched_init_debug);
814 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
817 * Number of tasks to iterate in a single balance run.
818 * Limited because this is done with IRQs disabled.
820 const_debug unsigned int sysctl_sched_nr_migrate = 32;
823 * ratelimit for updating the group shares.
826 unsigned int sysctl_sched_shares_ratelimit = 250000;
829 * Inject some fuzzyness into changing the per-cpu group shares
830 * this avoids remote rq-locks at the expense of fairness.
833 unsigned int sysctl_sched_shares_thresh = 4;
836 * period over which we measure -rt task cpu usage in us.
839 unsigned int sysctl_sched_rt_period = 1000000;
841 static __read_mostly int scheduler_running;
844 * part of the period that we allow rt tasks to run in us.
847 int sysctl_sched_rt_runtime = 950000;
849 static inline u64 global_rt_period(void)
851 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
854 static inline u64 global_rt_runtime(void)
856 if (sysctl_sched_rt_runtime < 0)
859 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
862 #ifndef prepare_arch_switch
863 # define prepare_arch_switch(next) do { } while (0)
865 #ifndef finish_arch_switch
866 # define finish_arch_switch(prev) do { } while (0)
869 static inline int task_current(struct rq *rq, struct task_struct *p)
871 return rq->curr == p;
874 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
875 static inline int task_running(struct rq *rq, struct task_struct *p)
877 return task_current(rq, p);
880 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
884 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
886 #ifdef CONFIG_DEBUG_SPINLOCK
887 /* this is a valid case when another task releases the spinlock */
888 rq->lock.owner = current;
891 * If we are tracking spinlock dependencies then we have to
892 * fix up the runqueue lock - which gets 'carried over' from
895 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
897 spin_unlock_irq(&rq->lock);
900 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
901 static inline int task_running(struct rq *rq, struct task_struct *p)
906 return task_current(rq, p);
910 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
914 * We can optimise this out completely for !SMP, because the
915 * SMP rebalancing from interrupt is the only thing that cares
920 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
921 spin_unlock_irq(&rq->lock);
923 spin_unlock(&rq->lock);
927 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
931 * After ->oncpu is cleared, the task can be moved to a different CPU.
932 * We must ensure this doesn't happen until the switch is completely
938 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
942 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
945 * __task_rq_lock - lock the runqueue a given task resides on.
946 * Must be called interrupts disabled.
948 static inline struct rq *__task_rq_lock(struct task_struct *p)
952 struct rq *rq = task_rq(p);
953 spin_lock(&rq->lock);
954 if (likely(rq == task_rq(p)))
956 spin_unlock(&rq->lock);
961 * task_rq_lock - lock the runqueue a given task resides on and disable
962 * interrupts. Note the ordering: we can safely lookup the task_rq without
963 * explicitly disabling preemption.
965 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
971 local_irq_save(*flags);
973 spin_lock(&rq->lock);
974 if (likely(rq == task_rq(p)))
976 spin_unlock_irqrestore(&rq->lock, *flags);
980 void task_rq_unlock_wait(struct task_struct *p)
982 struct rq *rq = task_rq(p);
984 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
985 spin_unlock_wait(&rq->lock);
988 static void __task_rq_unlock(struct rq *rq)
991 spin_unlock(&rq->lock);
994 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
997 spin_unlock_irqrestore(&rq->lock, *flags);
1001 * this_rq_lock - lock this runqueue and disable interrupts.
1003 static struct rq *this_rq_lock(void)
1004 __acquires(rq->lock)
1008 local_irq_disable();
1010 spin_lock(&rq->lock);
1015 #ifdef CONFIG_SCHED_HRTICK
1017 * Use HR-timers to deliver accurate preemption points.
1019 * Its all a bit involved since we cannot program an hrt while holding the
1020 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1023 * When we get rescheduled we reprogram the hrtick_timer outside of the
1029 * - enabled by features
1030 * - hrtimer is actually high res
1032 static inline int hrtick_enabled(struct rq *rq)
1034 if (!sched_feat(HRTICK))
1036 if (!cpu_active(cpu_of(rq)))
1038 return hrtimer_is_hres_active(&rq->hrtick_timer);
1041 static void hrtick_clear(struct rq *rq)
1043 if (hrtimer_active(&rq->hrtick_timer))
1044 hrtimer_cancel(&rq->hrtick_timer);
1048 * High-resolution timer tick.
1049 * Runs from hardirq context with interrupts disabled.
1051 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1053 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1055 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1057 spin_lock(&rq->lock);
1058 update_rq_clock(rq);
1059 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1060 spin_unlock(&rq->lock);
1062 return HRTIMER_NORESTART;
1067 * called from hardirq (IPI) context
1069 static void __hrtick_start(void *arg)
1071 struct rq *rq = arg;
1073 spin_lock(&rq->lock);
1074 hrtimer_restart(&rq->hrtick_timer);
1075 rq->hrtick_csd_pending = 0;
1076 spin_unlock(&rq->lock);
1080 * Called to set the hrtick timer state.
1082 * called with rq->lock held and irqs disabled
1084 static void hrtick_start(struct rq *rq, u64 delay)
1086 struct hrtimer *timer = &rq->hrtick_timer;
1087 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1089 hrtimer_set_expires(timer, time);
1091 if (rq == this_rq()) {
1092 hrtimer_restart(timer);
1093 } else if (!rq->hrtick_csd_pending) {
1094 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1095 rq->hrtick_csd_pending = 1;
1100 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1102 int cpu = (int)(long)hcpu;
1105 case CPU_UP_CANCELED:
1106 case CPU_UP_CANCELED_FROZEN:
1107 case CPU_DOWN_PREPARE:
1108 case CPU_DOWN_PREPARE_FROZEN:
1110 case CPU_DEAD_FROZEN:
1111 hrtick_clear(cpu_rq(cpu));
1118 static __init void init_hrtick(void)
1120 hotcpu_notifier(hotplug_hrtick, 0);
1124 * Called to set the hrtick timer state.
1126 * called with rq->lock held and irqs disabled
1128 static void hrtick_start(struct rq *rq, u64 delay)
1130 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1133 static inline void init_hrtick(void)
1136 #endif /* CONFIG_SMP */
1138 static void init_rq_hrtick(struct rq *rq)
1141 rq->hrtick_csd_pending = 0;
1143 rq->hrtick_csd.flags = 0;
1144 rq->hrtick_csd.func = __hrtick_start;
1145 rq->hrtick_csd.info = rq;
1148 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1149 rq->hrtick_timer.function = hrtick;
1150 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_PERCPU;
1152 #else /* CONFIG_SCHED_HRTICK */
1153 static inline void hrtick_clear(struct rq *rq)
1157 static inline void init_rq_hrtick(struct rq *rq)
1161 static inline void init_hrtick(void)
1164 #endif /* CONFIG_SCHED_HRTICK */
1167 * resched_task - mark a task 'to be rescheduled now'.
1169 * On UP this means the setting of the need_resched flag, on SMP it
1170 * might also involve a cross-CPU call to trigger the scheduler on
1175 #ifndef tsk_is_polling
1176 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1179 static void resched_task(struct task_struct *p)
1183 assert_spin_locked(&task_rq(p)->lock);
1185 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1188 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1191 if (cpu == smp_processor_id())
1194 /* NEED_RESCHED must be visible before we test polling */
1196 if (!tsk_is_polling(p))
1197 smp_send_reschedule(cpu);
1200 static void resched_cpu(int cpu)
1202 struct rq *rq = cpu_rq(cpu);
1203 unsigned long flags;
1205 if (!spin_trylock_irqsave(&rq->lock, flags))
1207 resched_task(cpu_curr(cpu));
1208 spin_unlock_irqrestore(&rq->lock, flags);
1213 * When add_timer_on() enqueues a timer into the timer wheel of an
1214 * idle CPU then this timer might expire before the next timer event
1215 * which is scheduled to wake up that CPU. In case of a completely
1216 * idle system the next event might even be infinite time into the
1217 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1218 * leaves the inner idle loop so the newly added timer is taken into
1219 * account when the CPU goes back to idle and evaluates the timer
1220 * wheel for the next timer event.
1222 void wake_up_idle_cpu(int cpu)
1224 struct rq *rq = cpu_rq(cpu);
1226 if (cpu == smp_processor_id())
1230 * This is safe, as this function is called with the timer
1231 * wheel base lock of (cpu) held. When the CPU is on the way
1232 * to idle and has not yet set rq->curr to idle then it will
1233 * be serialized on the timer wheel base lock and take the new
1234 * timer into account automatically.
1236 if (rq->curr != rq->idle)
1240 * We can set TIF_RESCHED on the idle task of the other CPU
1241 * lockless. The worst case is that the other CPU runs the
1242 * idle task through an additional NOOP schedule()
1244 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1246 /* NEED_RESCHED must be visible before we test polling */
1248 if (!tsk_is_polling(rq->idle))
1249 smp_send_reschedule(cpu);
1251 #endif /* CONFIG_NO_HZ */
1253 #else /* !CONFIG_SMP */
1254 static void resched_task(struct task_struct *p)
1256 assert_spin_locked(&task_rq(p)->lock);
1257 set_tsk_need_resched(p);
1259 #endif /* CONFIG_SMP */
1261 #if BITS_PER_LONG == 32
1262 # define WMULT_CONST (~0UL)
1264 # define WMULT_CONST (1UL << 32)
1267 #define WMULT_SHIFT 32
1270 * Shift right and round:
1272 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1275 * delta *= weight / lw
1277 static unsigned long
1278 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1279 struct load_weight *lw)
1283 if (!lw->inv_weight) {
1284 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1287 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1291 tmp = (u64)delta_exec * weight;
1293 * Check whether we'd overflow the 64-bit multiplication:
1295 if (unlikely(tmp > WMULT_CONST))
1296 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1299 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1301 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1304 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1310 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1317 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1318 * of tasks with abnormal "nice" values across CPUs the contribution that
1319 * each task makes to its run queue's load is weighted according to its
1320 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1321 * scaled version of the new time slice allocation that they receive on time
1325 #define WEIGHT_IDLEPRIO 2
1326 #define WMULT_IDLEPRIO (1 << 31)
1329 * Nice levels are multiplicative, with a gentle 10% change for every
1330 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1331 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1332 * that remained on nice 0.
1334 * The "10% effect" is relative and cumulative: from _any_ nice level,
1335 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1336 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1337 * If a task goes up by ~10% and another task goes down by ~10% then
1338 * the relative distance between them is ~25%.)
1340 static const int prio_to_weight[40] = {
1341 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1342 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1343 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1344 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1345 /* 0 */ 1024, 820, 655, 526, 423,
1346 /* 5 */ 335, 272, 215, 172, 137,
1347 /* 10 */ 110, 87, 70, 56, 45,
1348 /* 15 */ 36, 29, 23, 18, 15,
1352 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1354 * In cases where the weight does not change often, we can use the
1355 * precalculated inverse to speed up arithmetics by turning divisions
1356 * into multiplications:
1358 static const u32 prio_to_wmult[40] = {
1359 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1360 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1361 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1362 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1363 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1364 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1365 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1366 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1369 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1372 * runqueue iterator, to support SMP load-balancing between different
1373 * scheduling classes, without having to expose their internal data
1374 * structures to the load-balancing proper:
1376 struct rq_iterator {
1378 struct task_struct *(*start)(void *);
1379 struct task_struct *(*next)(void *);
1383 static unsigned long
1384 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1385 unsigned long max_load_move, struct sched_domain *sd,
1386 enum cpu_idle_type idle, int *all_pinned,
1387 int *this_best_prio, struct rq_iterator *iterator);
1390 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1391 struct sched_domain *sd, enum cpu_idle_type idle,
1392 struct rq_iterator *iterator);
1395 #ifdef CONFIG_CGROUP_CPUACCT
1396 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1398 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1401 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1403 update_load_add(&rq->load, load);
1406 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1408 update_load_sub(&rq->load, load);
1411 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1412 typedef int (*tg_visitor)(struct task_group *, void *);
1415 * Iterate the full tree, calling @down when first entering a node and @up when
1416 * leaving it for the final time.
1418 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1420 struct task_group *parent, *child;
1424 parent = &root_task_group;
1426 ret = (*down)(parent, data);
1429 list_for_each_entry_rcu(child, &parent->children, siblings) {
1436 ret = (*up)(parent, data);
1441 parent = parent->parent;
1450 static int tg_nop(struct task_group *tg, void *data)
1457 static unsigned long source_load(int cpu, int type);
1458 static unsigned long target_load(int cpu, int type);
1459 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1461 static unsigned long cpu_avg_load_per_task(int cpu)
1463 struct rq *rq = cpu_rq(cpu);
1464 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1467 rq->avg_load_per_task = rq->load.weight / nr_running;
1469 rq->avg_load_per_task = 0;
1471 return rq->avg_load_per_task;
1474 #ifdef CONFIG_FAIR_GROUP_SCHED
1476 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1479 * Calculate and set the cpu's group shares.
1482 update_group_shares_cpu(struct task_group *tg, int cpu,
1483 unsigned long sd_shares, unsigned long sd_rq_weight)
1485 unsigned long shares;
1486 unsigned long rq_weight;
1491 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1494 * \Sum shares * rq_weight
1495 * shares = -----------------------
1499 shares = (sd_shares * rq_weight) / sd_rq_weight;
1500 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1502 if (abs(shares - tg->se[cpu]->load.weight) >
1503 sysctl_sched_shares_thresh) {
1504 struct rq *rq = cpu_rq(cpu);
1505 unsigned long flags;
1507 spin_lock_irqsave(&rq->lock, flags);
1508 tg->cfs_rq[cpu]->shares = shares;
1510 __set_se_shares(tg->se[cpu], shares);
1511 spin_unlock_irqrestore(&rq->lock, flags);
1516 * Re-compute the task group their per cpu shares over the given domain.
1517 * This needs to be done in a bottom-up fashion because the rq weight of a
1518 * parent group depends on the shares of its child groups.
1520 static int tg_shares_up(struct task_group *tg, void *data)
1522 unsigned long weight, rq_weight = 0;
1523 unsigned long shares = 0;
1524 struct sched_domain *sd = data;
1527 for_each_cpu(i, sched_domain_span(sd)) {
1529 * If there are currently no tasks on the cpu pretend there
1530 * is one of average load so that when a new task gets to
1531 * run here it will not get delayed by group starvation.
1533 weight = tg->cfs_rq[i]->load.weight;
1535 weight = NICE_0_LOAD;
1537 tg->cfs_rq[i]->rq_weight = weight;
1538 rq_weight += weight;
1539 shares += tg->cfs_rq[i]->shares;
1542 if ((!shares && rq_weight) || shares > tg->shares)
1543 shares = tg->shares;
1545 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1546 shares = tg->shares;
1548 for_each_cpu(i, sched_domain_span(sd))
1549 update_group_shares_cpu(tg, i, shares, rq_weight);
1555 * Compute the cpu's hierarchical load factor for each task group.
1556 * This needs to be done in a top-down fashion because the load of a child
1557 * group is a fraction of its parents load.
1559 static int tg_load_down(struct task_group *tg, void *data)
1562 long cpu = (long)data;
1565 load = cpu_rq(cpu)->load.weight;
1567 load = tg->parent->cfs_rq[cpu]->h_load;
1568 load *= tg->cfs_rq[cpu]->shares;
1569 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1572 tg->cfs_rq[cpu]->h_load = load;
1577 static void update_shares(struct sched_domain *sd)
1579 u64 now = cpu_clock(raw_smp_processor_id());
1580 s64 elapsed = now - sd->last_update;
1582 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1583 sd->last_update = now;
1584 walk_tg_tree(tg_nop, tg_shares_up, sd);
1588 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1590 spin_unlock(&rq->lock);
1592 spin_lock(&rq->lock);
1595 static void update_h_load(long cpu)
1597 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1602 static inline void update_shares(struct sched_domain *sd)
1606 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1612 #ifdef CONFIG_PREEMPT
1615 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1616 * way at the expense of forcing extra atomic operations in all
1617 * invocations. This assures that the double_lock is acquired using the
1618 * same underlying policy as the spinlock_t on this architecture, which
1619 * reduces latency compared to the unfair variant below. However, it
1620 * also adds more overhead and therefore may reduce throughput.
1622 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1623 __releases(this_rq->lock)
1624 __acquires(busiest->lock)
1625 __acquires(this_rq->lock)
1627 spin_unlock(&this_rq->lock);
1628 double_rq_lock(this_rq, busiest);
1635 * Unfair double_lock_balance: Optimizes throughput at the expense of
1636 * latency by eliminating extra atomic operations when the locks are
1637 * already in proper order on entry. This favors lower cpu-ids and will
1638 * grant the double lock to lower cpus over higher ids under contention,
1639 * regardless of entry order into the function.
1641 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1642 __releases(this_rq->lock)
1643 __acquires(busiest->lock)
1644 __acquires(this_rq->lock)
1648 if (unlikely(!spin_trylock(&busiest->lock))) {
1649 if (busiest < this_rq) {
1650 spin_unlock(&this_rq->lock);
1651 spin_lock(&busiest->lock);
1652 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1655 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1660 #endif /* CONFIG_PREEMPT */
1663 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1665 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1667 if (unlikely(!irqs_disabled())) {
1668 /* printk() doesn't work good under rq->lock */
1669 spin_unlock(&this_rq->lock);
1673 return _double_lock_balance(this_rq, busiest);
1676 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1677 __releases(busiest->lock)
1679 spin_unlock(&busiest->lock);
1680 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1684 #ifdef CONFIG_FAIR_GROUP_SCHED
1685 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1688 cfs_rq->shares = shares;
1693 #include "sched_stats.h"
1694 #include "sched_idletask.c"
1695 #include "sched_fair.c"
1696 #include "sched_rt.c"
1697 #ifdef CONFIG_SCHED_DEBUG
1698 # include "sched_debug.c"
1701 #define sched_class_highest (&rt_sched_class)
1702 #define for_each_class(class) \
1703 for (class = sched_class_highest; class; class = class->next)
1705 static void inc_nr_running(struct rq *rq)
1710 static void dec_nr_running(struct rq *rq)
1715 static void set_load_weight(struct task_struct *p)
1717 if (task_has_rt_policy(p)) {
1718 p->se.load.weight = prio_to_weight[0] * 2;
1719 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1724 * SCHED_IDLE tasks get minimal weight:
1726 if (p->policy == SCHED_IDLE) {
1727 p->se.load.weight = WEIGHT_IDLEPRIO;
1728 p->se.load.inv_weight = WMULT_IDLEPRIO;
1732 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1733 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1736 static void update_avg(u64 *avg, u64 sample)
1738 s64 diff = sample - *avg;
1742 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1744 sched_info_queued(p);
1745 p->sched_class->enqueue_task(rq, p, wakeup);
1749 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1751 if (sleep && p->se.last_wakeup) {
1752 update_avg(&p->se.avg_overlap,
1753 p->se.sum_exec_runtime - p->se.last_wakeup);
1754 p->se.last_wakeup = 0;
1757 sched_info_dequeued(p);
1758 p->sched_class->dequeue_task(rq, p, sleep);
1763 * __normal_prio - return the priority that is based on the static prio
1765 static inline int __normal_prio(struct task_struct *p)
1767 return p->static_prio;
1771 * Calculate the expected normal priority: i.e. priority
1772 * without taking RT-inheritance into account. Might be
1773 * boosted by interactivity modifiers. Changes upon fork,
1774 * setprio syscalls, and whenever the interactivity
1775 * estimator recalculates.
1777 static inline int normal_prio(struct task_struct *p)
1781 if (task_has_rt_policy(p))
1782 prio = MAX_RT_PRIO-1 - p->rt_priority;
1784 prio = __normal_prio(p);
1789 * Calculate the current priority, i.e. the priority
1790 * taken into account by the scheduler. This value might
1791 * be boosted by RT tasks, or might be boosted by
1792 * interactivity modifiers. Will be RT if the task got
1793 * RT-boosted. If not then it returns p->normal_prio.
1795 static int effective_prio(struct task_struct *p)
1797 p->normal_prio = normal_prio(p);
1799 * If we are RT tasks or we were boosted to RT priority,
1800 * keep the priority unchanged. Otherwise, update priority
1801 * to the normal priority:
1803 if (!rt_prio(p->prio))
1804 return p->normal_prio;
1809 * activate_task - move a task to the runqueue.
1811 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1813 if (task_contributes_to_load(p))
1814 rq->nr_uninterruptible--;
1816 enqueue_task(rq, p, wakeup);
1821 * deactivate_task - remove a task from the runqueue.
1823 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1825 if (task_contributes_to_load(p))
1826 rq->nr_uninterruptible++;
1828 dequeue_task(rq, p, sleep);
1833 * task_curr - is this task currently executing on a CPU?
1834 * @p: the task in question.
1836 inline int task_curr(const struct task_struct *p)
1838 return cpu_curr(task_cpu(p)) == p;
1841 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1843 set_task_rq(p, cpu);
1846 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1847 * successfuly executed on another CPU. We must ensure that updates of
1848 * per-task data have been completed by this moment.
1851 task_thread_info(p)->cpu = cpu;
1855 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1856 const struct sched_class *prev_class,
1857 int oldprio, int running)
1859 if (prev_class != p->sched_class) {
1860 if (prev_class->switched_from)
1861 prev_class->switched_from(rq, p, running);
1862 p->sched_class->switched_to(rq, p, running);
1864 p->sched_class->prio_changed(rq, p, oldprio, running);
1869 /* Used instead of source_load when we know the type == 0 */
1870 static unsigned long weighted_cpuload(const int cpu)
1872 return cpu_rq(cpu)->load.weight;
1876 * Is this task likely cache-hot:
1879 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1884 * Buddy candidates are cache hot:
1886 if (sched_feat(CACHE_HOT_BUDDY) &&
1887 (&p->se == cfs_rq_of(&p->se)->next ||
1888 &p->se == cfs_rq_of(&p->se)->last))
1891 if (p->sched_class != &fair_sched_class)
1894 if (sysctl_sched_migration_cost == -1)
1896 if (sysctl_sched_migration_cost == 0)
1899 delta = now - p->se.exec_start;
1901 return delta < (s64)sysctl_sched_migration_cost;
1905 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1907 int old_cpu = task_cpu(p);
1908 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1909 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1910 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1913 clock_offset = old_rq->clock - new_rq->clock;
1915 #ifdef CONFIG_SCHEDSTATS
1916 if (p->se.wait_start)
1917 p->se.wait_start -= clock_offset;
1918 if (p->se.sleep_start)
1919 p->se.sleep_start -= clock_offset;
1920 if (p->se.block_start)
1921 p->se.block_start -= clock_offset;
1922 if (old_cpu != new_cpu) {
1923 schedstat_inc(p, se.nr_migrations);
1924 if (task_hot(p, old_rq->clock, NULL))
1925 schedstat_inc(p, se.nr_forced2_migrations);
1928 p->se.vruntime -= old_cfsrq->min_vruntime -
1929 new_cfsrq->min_vruntime;
1931 __set_task_cpu(p, new_cpu);
1934 struct migration_req {
1935 struct list_head list;
1937 struct task_struct *task;
1940 struct completion done;
1944 * The task's runqueue lock must be held.
1945 * Returns true if you have to wait for migration thread.
1948 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1950 struct rq *rq = task_rq(p);
1953 * If the task is not on a runqueue (and not running), then
1954 * it is sufficient to simply update the task's cpu field.
1956 if (!p->se.on_rq && !task_running(rq, p)) {
1957 set_task_cpu(p, dest_cpu);
1961 init_completion(&req->done);
1963 req->dest_cpu = dest_cpu;
1964 list_add(&req->list, &rq->migration_queue);
1970 * wait_task_inactive - wait for a thread to unschedule.
1972 * If @match_state is nonzero, it's the @p->state value just checked and
1973 * not expected to change. If it changes, i.e. @p might have woken up,
1974 * then return zero. When we succeed in waiting for @p to be off its CPU,
1975 * we return a positive number (its total switch count). If a second call
1976 * a short while later returns the same number, the caller can be sure that
1977 * @p has remained unscheduled the whole time.
1979 * The caller must ensure that the task *will* unschedule sometime soon,
1980 * else this function might spin for a *long* time. This function can't
1981 * be called with interrupts off, or it may introduce deadlock with
1982 * smp_call_function() if an IPI is sent by the same process we are
1983 * waiting to become inactive.
1985 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1987 unsigned long flags;
1994 * We do the initial early heuristics without holding
1995 * any task-queue locks at all. We'll only try to get
1996 * the runqueue lock when things look like they will
2002 * If the task is actively running on another CPU
2003 * still, just relax and busy-wait without holding
2006 * NOTE! Since we don't hold any locks, it's not
2007 * even sure that "rq" stays as the right runqueue!
2008 * But we don't care, since "task_running()" will
2009 * return false if the runqueue has changed and p
2010 * is actually now running somewhere else!
2012 while (task_running(rq, p)) {
2013 if (match_state && unlikely(p->state != match_state))
2019 * Ok, time to look more closely! We need the rq
2020 * lock now, to be *sure*. If we're wrong, we'll
2021 * just go back and repeat.
2023 rq = task_rq_lock(p, &flags);
2024 trace_sched_wait_task(rq, p);
2025 running = task_running(rq, p);
2026 on_rq = p->se.on_rq;
2028 if (!match_state || p->state == match_state)
2029 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2030 task_rq_unlock(rq, &flags);
2033 * If it changed from the expected state, bail out now.
2035 if (unlikely(!ncsw))
2039 * Was it really running after all now that we
2040 * checked with the proper locks actually held?
2042 * Oops. Go back and try again..
2044 if (unlikely(running)) {
2050 * It's not enough that it's not actively running,
2051 * it must be off the runqueue _entirely_, and not
2054 * So if it wa still runnable (but just not actively
2055 * running right now), it's preempted, and we should
2056 * yield - it could be a while.
2058 if (unlikely(on_rq)) {
2059 schedule_timeout_uninterruptible(1);
2064 * Ahh, all good. It wasn't running, and it wasn't
2065 * runnable, which means that it will never become
2066 * running in the future either. We're all done!
2075 * kick_process - kick a running thread to enter/exit the kernel
2076 * @p: the to-be-kicked thread
2078 * Cause a process which is running on another CPU to enter
2079 * kernel-mode, without any delay. (to get signals handled.)
2081 * NOTE: this function doesnt have to take the runqueue lock,
2082 * because all it wants to ensure is that the remote task enters
2083 * the kernel. If the IPI races and the task has been migrated
2084 * to another CPU then no harm is done and the purpose has been
2087 void kick_process(struct task_struct *p)
2093 if ((cpu != smp_processor_id()) && task_curr(p))
2094 smp_send_reschedule(cpu);
2099 * Return a low guess at the load of a migration-source cpu weighted
2100 * according to the scheduling class and "nice" value.
2102 * We want to under-estimate the load of migration sources, to
2103 * balance conservatively.
2105 static unsigned long source_load(int cpu, int type)
2107 struct rq *rq = cpu_rq(cpu);
2108 unsigned long total = weighted_cpuload(cpu);
2110 if (type == 0 || !sched_feat(LB_BIAS))
2113 return min(rq->cpu_load[type-1], total);
2117 * Return a high guess at the load of a migration-target cpu weighted
2118 * according to the scheduling class and "nice" value.
2120 static unsigned long target_load(int cpu, int type)
2122 struct rq *rq = cpu_rq(cpu);
2123 unsigned long total = weighted_cpuload(cpu);
2125 if (type == 0 || !sched_feat(LB_BIAS))
2128 return max(rq->cpu_load[type-1], total);
2132 * find_idlest_group finds and returns the least busy CPU group within the
2135 static struct sched_group *
2136 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2138 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2139 unsigned long min_load = ULONG_MAX, this_load = 0;
2140 int load_idx = sd->forkexec_idx;
2141 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2144 unsigned long load, avg_load;
2148 /* Skip over this group if it has no CPUs allowed */
2149 if (!cpumask_intersects(sched_group_cpus(group),
2153 local_group = cpumask_test_cpu(this_cpu,
2154 sched_group_cpus(group));
2156 /* Tally up the load of all CPUs in the group */
2159 for_each_cpu(i, sched_group_cpus(group)) {
2160 /* Bias balancing toward cpus of our domain */
2162 load = source_load(i, load_idx);
2164 load = target_load(i, load_idx);
2169 /* Adjust by relative CPU power of the group */
2170 avg_load = sg_div_cpu_power(group,
2171 avg_load * SCHED_LOAD_SCALE);
2174 this_load = avg_load;
2176 } else if (avg_load < min_load) {
2177 min_load = avg_load;
2180 } while (group = group->next, group != sd->groups);
2182 if (!idlest || 100*this_load < imbalance*min_load)
2188 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2191 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2193 unsigned long load, min_load = ULONG_MAX;
2197 /* Traverse only the allowed CPUs */
2198 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2199 load = weighted_cpuload(i);
2201 if (load < min_load || (load == min_load && i == this_cpu)) {
2211 * sched_balance_self: balance the current task (running on cpu) in domains
2212 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2215 * Balance, ie. select the least loaded group.
2217 * Returns the target CPU number, or the same CPU if no balancing is needed.
2219 * preempt must be disabled.
2221 static int sched_balance_self(int cpu, int flag)
2223 struct task_struct *t = current;
2224 struct sched_domain *tmp, *sd = NULL;
2226 for_each_domain(cpu, tmp) {
2228 * If power savings logic is enabled for a domain, stop there.
2230 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2232 if (tmp->flags & flag)
2240 struct sched_group *group;
2241 int new_cpu, weight;
2243 if (!(sd->flags & flag)) {
2248 group = find_idlest_group(sd, t, cpu);
2254 new_cpu = find_idlest_cpu(group, t, cpu);
2255 if (new_cpu == -1 || new_cpu == cpu) {
2256 /* Now try balancing at a lower domain level of cpu */
2261 /* Now try balancing at a lower domain level of new_cpu */
2263 weight = cpumask_weight(sched_domain_span(sd));
2265 for_each_domain(cpu, tmp) {
2266 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2268 if (tmp->flags & flag)
2271 /* while loop will break here if sd == NULL */
2277 #endif /* CONFIG_SMP */
2280 * try_to_wake_up - wake up a thread
2281 * @p: the to-be-woken-up thread
2282 * @state: the mask of task states that can be woken
2283 * @sync: do a synchronous wakeup?
2285 * Put it on the run-queue if it's not already there. The "current"
2286 * thread is always on the run-queue (except when the actual
2287 * re-schedule is in progress), and as such you're allowed to do
2288 * the simpler "current->state = TASK_RUNNING" to mark yourself
2289 * runnable without the overhead of this.
2291 * returns failure only if the task is already active.
2293 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2295 int cpu, orig_cpu, this_cpu, success = 0;
2296 unsigned long flags;
2300 if (!sched_feat(SYNC_WAKEUPS))
2304 if (sched_feat(LB_WAKEUP_UPDATE)) {
2305 struct sched_domain *sd;
2307 this_cpu = raw_smp_processor_id();
2310 for_each_domain(this_cpu, sd) {
2311 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2320 rq = task_rq_lock(p, &flags);
2321 old_state = p->state;
2322 if (!(old_state & state))
2330 this_cpu = smp_processor_id();
2333 if (unlikely(task_running(rq, p)))
2336 cpu = p->sched_class->select_task_rq(p, sync);
2337 if (cpu != orig_cpu) {
2338 set_task_cpu(p, cpu);
2339 task_rq_unlock(rq, &flags);
2340 /* might preempt at this point */
2341 rq = task_rq_lock(p, &flags);
2342 old_state = p->state;
2343 if (!(old_state & state))
2348 this_cpu = smp_processor_id();
2352 #ifdef CONFIG_SCHEDSTATS
2353 schedstat_inc(rq, ttwu_count);
2354 if (cpu == this_cpu)
2355 schedstat_inc(rq, ttwu_local);
2357 struct sched_domain *sd;
2358 for_each_domain(this_cpu, sd) {
2359 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2360 schedstat_inc(sd, ttwu_wake_remote);
2365 #endif /* CONFIG_SCHEDSTATS */
2368 #endif /* CONFIG_SMP */
2369 schedstat_inc(p, se.nr_wakeups);
2371 schedstat_inc(p, se.nr_wakeups_sync);
2372 if (orig_cpu != cpu)
2373 schedstat_inc(p, se.nr_wakeups_migrate);
2374 if (cpu == this_cpu)
2375 schedstat_inc(p, se.nr_wakeups_local);
2377 schedstat_inc(p, se.nr_wakeups_remote);
2378 update_rq_clock(rq);
2379 activate_task(rq, p, 1);
2383 trace_sched_wakeup(rq, p);
2384 check_preempt_curr(rq, p, sync);
2386 p->state = TASK_RUNNING;
2388 if (p->sched_class->task_wake_up)
2389 p->sched_class->task_wake_up(rq, p);
2392 current->se.last_wakeup = current->se.sum_exec_runtime;
2394 task_rq_unlock(rq, &flags);
2399 int wake_up_process(struct task_struct *p)
2401 return try_to_wake_up(p, TASK_ALL, 0);
2403 EXPORT_SYMBOL(wake_up_process);
2405 int wake_up_state(struct task_struct *p, unsigned int state)
2407 return try_to_wake_up(p, state, 0);
2411 * Perform scheduler related setup for a newly forked process p.
2412 * p is forked by current.
2414 * __sched_fork() is basic setup used by init_idle() too:
2416 static void __sched_fork(struct task_struct *p)
2418 p->se.exec_start = 0;
2419 p->se.sum_exec_runtime = 0;
2420 p->se.prev_sum_exec_runtime = 0;
2421 p->se.last_wakeup = 0;
2422 p->se.avg_overlap = 0;
2424 #ifdef CONFIG_SCHEDSTATS
2425 p->se.wait_start = 0;
2426 p->se.sum_sleep_runtime = 0;
2427 p->se.sleep_start = 0;
2428 p->se.block_start = 0;
2429 p->se.sleep_max = 0;
2430 p->se.block_max = 0;
2432 p->se.slice_max = 0;
2436 INIT_LIST_HEAD(&p->rt.run_list);
2438 INIT_LIST_HEAD(&p->se.group_node);
2440 #ifdef CONFIG_PREEMPT_NOTIFIERS
2441 INIT_HLIST_HEAD(&p->preempt_notifiers);
2445 * We mark the process as running here, but have not actually
2446 * inserted it onto the runqueue yet. This guarantees that
2447 * nobody will actually run it, and a signal or other external
2448 * event cannot wake it up and insert it on the runqueue either.
2450 p->state = TASK_RUNNING;
2454 * fork()/clone()-time setup:
2456 void sched_fork(struct task_struct *p, int clone_flags)
2458 int cpu = get_cpu();
2463 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2465 set_task_cpu(p, cpu);
2468 * Make sure we do not leak PI boosting priority to the child:
2470 p->prio = current->normal_prio;
2471 if (!rt_prio(p->prio))
2472 p->sched_class = &fair_sched_class;
2474 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2475 if (likely(sched_info_on()))
2476 memset(&p->sched_info, 0, sizeof(p->sched_info));
2478 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2481 #ifdef CONFIG_PREEMPT
2482 /* Want to start with kernel preemption disabled. */
2483 task_thread_info(p)->preempt_count = 1;
2485 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2491 * wake_up_new_task - wake up a newly created task for the first time.
2493 * This function will do some initial scheduler statistics housekeeping
2494 * that must be done for every newly created context, then puts the task
2495 * on the runqueue and wakes it.
2497 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2499 unsigned long flags;
2502 rq = task_rq_lock(p, &flags);
2503 BUG_ON(p->state != TASK_RUNNING);
2504 update_rq_clock(rq);
2506 p->prio = effective_prio(p);
2508 if (!p->sched_class->task_new || !current->se.on_rq) {
2509 activate_task(rq, p, 0);
2512 * Let the scheduling class do new task startup
2513 * management (if any):
2515 p->sched_class->task_new(rq, p);
2518 trace_sched_wakeup_new(rq, p);
2519 check_preempt_curr(rq, p, 0);
2521 if (p->sched_class->task_wake_up)
2522 p->sched_class->task_wake_up(rq, p);
2524 task_rq_unlock(rq, &flags);
2527 #ifdef CONFIG_PREEMPT_NOTIFIERS
2530 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2531 * @notifier: notifier struct to register
2533 void preempt_notifier_register(struct preempt_notifier *notifier)
2535 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2537 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2540 * preempt_notifier_unregister - no longer interested in preemption notifications
2541 * @notifier: notifier struct to unregister
2543 * This is safe to call from within a preemption notifier.
2545 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2547 hlist_del(¬ifier->link);
2549 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2551 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2553 struct preempt_notifier *notifier;
2554 struct hlist_node *node;
2556 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2557 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2561 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2562 struct task_struct *next)
2564 struct preempt_notifier *notifier;
2565 struct hlist_node *node;
2567 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2568 notifier->ops->sched_out(notifier, next);
2571 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2573 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2578 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2579 struct task_struct *next)
2583 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2586 * prepare_task_switch - prepare to switch tasks
2587 * @rq: the runqueue preparing to switch
2588 * @prev: the current task that is being switched out
2589 * @next: the task we are going to switch to.
2591 * This is called with the rq lock held and interrupts off. It must
2592 * be paired with a subsequent finish_task_switch after the context
2595 * prepare_task_switch sets up locking and calls architecture specific
2599 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2600 struct task_struct *next)
2602 fire_sched_out_preempt_notifiers(prev, next);
2603 prepare_lock_switch(rq, next);
2604 prepare_arch_switch(next);
2608 * finish_task_switch - clean up after a task-switch
2609 * @rq: runqueue associated with task-switch
2610 * @prev: the thread we just switched away from.
2612 * finish_task_switch must be called after the context switch, paired
2613 * with a prepare_task_switch call before the context switch.
2614 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2615 * and do any other architecture-specific cleanup actions.
2617 * Note that we may have delayed dropping an mm in context_switch(). If
2618 * so, we finish that here outside of the runqueue lock. (Doing it
2619 * with the lock held can cause deadlocks; see schedule() for
2622 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2623 __releases(rq->lock)
2625 struct mm_struct *mm = rq->prev_mm;
2628 int post_schedule = 0;
2630 if (current->sched_class->needs_post_schedule)
2631 post_schedule = current->sched_class->needs_post_schedule(rq);
2637 * A task struct has one reference for the use as "current".
2638 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2639 * schedule one last time. The schedule call will never return, and
2640 * the scheduled task must drop that reference.
2641 * The test for TASK_DEAD must occur while the runqueue locks are
2642 * still held, otherwise prev could be scheduled on another cpu, die
2643 * there before we look at prev->state, and then the reference would
2645 * Manfred Spraul <manfred@colorfullife.com>
2647 prev_state = prev->state;
2648 finish_arch_switch(prev);
2649 finish_lock_switch(rq, prev);
2652 current->sched_class->post_schedule(rq);
2655 fire_sched_in_preempt_notifiers(current);
2658 if (unlikely(prev_state == TASK_DEAD)) {
2660 * Remove function-return probe instances associated with this
2661 * task and put them back on the free list.
2663 kprobe_flush_task(prev);
2664 put_task_struct(prev);
2669 * schedule_tail - first thing a freshly forked thread must call.
2670 * @prev: the thread we just switched away from.
2672 asmlinkage void schedule_tail(struct task_struct *prev)
2673 __releases(rq->lock)
2675 struct rq *rq = this_rq();
2677 finish_task_switch(rq, prev);
2678 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2679 /* In this case, finish_task_switch does not reenable preemption */
2682 if (current->set_child_tid)
2683 put_user(task_pid_vnr(current), current->set_child_tid);
2687 * context_switch - switch to the new MM and the new
2688 * thread's register state.
2691 context_switch(struct rq *rq, struct task_struct *prev,
2692 struct task_struct *next)
2694 struct mm_struct *mm, *oldmm;
2696 prepare_task_switch(rq, prev, next);
2697 trace_sched_switch(rq, prev, next);
2699 oldmm = prev->active_mm;
2701 * For paravirt, this is coupled with an exit in switch_to to
2702 * combine the page table reload and the switch backend into
2705 arch_enter_lazy_cpu_mode();
2707 if (unlikely(!mm)) {
2708 next->active_mm = oldmm;
2709 atomic_inc(&oldmm->mm_count);
2710 enter_lazy_tlb(oldmm, next);
2712 switch_mm(oldmm, mm, next);
2714 if (unlikely(!prev->mm)) {
2715 prev->active_mm = NULL;
2716 rq->prev_mm = oldmm;
2719 * Since the runqueue lock will be released by the next
2720 * task (which is an invalid locking op but in the case
2721 * of the scheduler it's an obvious special-case), so we
2722 * do an early lockdep release here:
2724 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2725 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2728 /* Here we just switch the register state and the stack. */
2729 switch_to(prev, next, prev);
2733 * this_rq must be evaluated again because prev may have moved
2734 * CPUs since it called schedule(), thus the 'rq' on its stack
2735 * frame will be invalid.
2737 finish_task_switch(this_rq(), prev);
2741 * nr_running, nr_uninterruptible and nr_context_switches:
2743 * externally visible scheduler statistics: current number of runnable
2744 * threads, current number of uninterruptible-sleeping threads, total
2745 * number of context switches performed since bootup.
2747 unsigned long nr_running(void)
2749 unsigned long i, sum = 0;
2751 for_each_online_cpu(i)
2752 sum += cpu_rq(i)->nr_running;
2757 unsigned long nr_uninterruptible(void)
2759 unsigned long i, sum = 0;
2761 for_each_possible_cpu(i)
2762 sum += cpu_rq(i)->nr_uninterruptible;
2765 * Since we read the counters lockless, it might be slightly
2766 * inaccurate. Do not allow it to go below zero though:
2768 if (unlikely((long)sum < 0))
2774 unsigned long long nr_context_switches(void)
2777 unsigned long long sum = 0;
2779 for_each_possible_cpu(i)
2780 sum += cpu_rq(i)->nr_switches;
2785 unsigned long nr_iowait(void)
2787 unsigned long i, sum = 0;
2789 for_each_possible_cpu(i)
2790 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2795 unsigned long nr_active(void)
2797 unsigned long i, running = 0, uninterruptible = 0;
2799 for_each_online_cpu(i) {
2800 running += cpu_rq(i)->nr_running;
2801 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2804 if (unlikely((long)uninterruptible < 0))
2805 uninterruptible = 0;
2807 return running + uninterruptible;
2811 * Update rq->cpu_load[] statistics. This function is usually called every
2812 * scheduler tick (TICK_NSEC).
2814 static void update_cpu_load(struct rq *this_rq)
2816 unsigned long this_load = this_rq->load.weight;
2819 this_rq->nr_load_updates++;
2821 /* Update our load: */
2822 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2823 unsigned long old_load, new_load;
2825 /* scale is effectively 1 << i now, and >> i divides by scale */
2827 old_load = this_rq->cpu_load[i];
2828 new_load = this_load;
2830 * Round up the averaging division if load is increasing. This
2831 * prevents us from getting stuck on 9 if the load is 10, for
2834 if (new_load > old_load)
2835 new_load += scale-1;
2836 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2843 * double_rq_lock - safely lock two runqueues
2845 * Note this does not disable interrupts like task_rq_lock,
2846 * you need to do so manually before calling.
2848 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2849 __acquires(rq1->lock)
2850 __acquires(rq2->lock)
2852 BUG_ON(!irqs_disabled());
2854 spin_lock(&rq1->lock);
2855 __acquire(rq2->lock); /* Fake it out ;) */
2858 spin_lock(&rq1->lock);
2859 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2861 spin_lock(&rq2->lock);
2862 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2865 update_rq_clock(rq1);
2866 update_rq_clock(rq2);
2870 * double_rq_unlock - safely unlock two runqueues
2872 * Note this does not restore interrupts like task_rq_unlock,
2873 * you need to do so manually after calling.
2875 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2876 __releases(rq1->lock)
2877 __releases(rq2->lock)
2879 spin_unlock(&rq1->lock);
2881 spin_unlock(&rq2->lock);
2883 __release(rq2->lock);
2887 * If dest_cpu is allowed for this process, migrate the task to it.
2888 * This is accomplished by forcing the cpu_allowed mask to only
2889 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2890 * the cpu_allowed mask is restored.
2892 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2894 struct migration_req req;
2895 unsigned long flags;
2898 rq = task_rq_lock(p, &flags);
2899 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
2900 || unlikely(!cpu_active(dest_cpu)))
2903 trace_sched_migrate_task(rq, p, dest_cpu);
2904 /* force the process onto the specified CPU */
2905 if (migrate_task(p, dest_cpu, &req)) {
2906 /* Need to wait for migration thread (might exit: take ref). */
2907 struct task_struct *mt = rq->migration_thread;
2909 get_task_struct(mt);
2910 task_rq_unlock(rq, &flags);
2911 wake_up_process(mt);
2912 put_task_struct(mt);
2913 wait_for_completion(&req.done);
2918 task_rq_unlock(rq, &flags);
2922 * sched_exec - execve() is a valuable balancing opportunity, because at
2923 * this point the task has the smallest effective memory and cache footprint.
2925 void sched_exec(void)
2927 int new_cpu, this_cpu = get_cpu();
2928 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2930 if (new_cpu != this_cpu)
2931 sched_migrate_task(current, new_cpu);
2935 * pull_task - move a task from a remote runqueue to the local runqueue.
2936 * Both runqueues must be locked.
2938 static void pull_task(struct rq *src_rq, struct task_struct *p,
2939 struct rq *this_rq, int this_cpu)
2941 deactivate_task(src_rq, p, 0);
2942 set_task_cpu(p, this_cpu);
2943 activate_task(this_rq, p, 0);
2945 * Note that idle threads have a prio of MAX_PRIO, for this test
2946 * to be always true for them.
2948 check_preempt_curr(this_rq, p, 0);
2952 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2955 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2956 struct sched_domain *sd, enum cpu_idle_type idle,
2960 * We do not migrate tasks that are:
2961 * 1) running (obviously), or
2962 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2963 * 3) are cache-hot on their current CPU.
2965 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
2966 schedstat_inc(p, se.nr_failed_migrations_affine);
2971 if (task_running(rq, p)) {
2972 schedstat_inc(p, se.nr_failed_migrations_running);
2977 * Aggressive migration if:
2978 * 1) task is cache cold, or
2979 * 2) too many balance attempts have failed.
2982 if (!task_hot(p, rq->clock, sd) ||
2983 sd->nr_balance_failed > sd->cache_nice_tries) {
2984 #ifdef CONFIG_SCHEDSTATS
2985 if (task_hot(p, rq->clock, sd)) {
2986 schedstat_inc(sd, lb_hot_gained[idle]);
2987 schedstat_inc(p, se.nr_forced_migrations);
2993 if (task_hot(p, rq->clock, sd)) {
2994 schedstat_inc(p, se.nr_failed_migrations_hot);
3000 static unsigned long
3001 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3002 unsigned long max_load_move, struct sched_domain *sd,
3003 enum cpu_idle_type idle, int *all_pinned,
3004 int *this_best_prio, struct rq_iterator *iterator)
3006 int loops = 0, pulled = 0, pinned = 0;
3007 struct task_struct *p;
3008 long rem_load_move = max_load_move;
3010 if (max_load_move == 0)
3016 * Start the load-balancing iterator:
3018 p = iterator->start(iterator->arg);
3020 if (!p || loops++ > sysctl_sched_nr_migrate)
3023 if ((p->se.load.weight >> 1) > rem_load_move ||
3024 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3025 p = iterator->next(iterator->arg);
3029 pull_task(busiest, p, this_rq, this_cpu);
3031 rem_load_move -= p->se.load.weight;
3033 #ifdef CONFIG_PREEMPT
3035 * NEWIDLE balancing is a source of latency, so preemptible kernels
3036 * will stop after the first task is pulled to minimize the critical
3039 if (idle == CPU_NEWLY_IDLE)
3044 * We only want to steal up to the prescribed amount of weighted load.
3046 if (rem_load_move > 0) {
3047 if (p->prio < *this_best_prio)
3048 *this_best_prio = p->prio;
3049 p = iterator->next(iterator->arg);
3054 * Right now, this is one of only two places pull_task() is called,
3055 * so we can safely collect pull_task() stats here rather than
3056 * inside pull_task().
3058 schedstat_add(sd, lb_gained[idle], pulled);
3061 *all_pinned = pinned;
3063 return max_load_move - rem_load_move;
3067 * move_tasks tries to move up to max_load_move weighted load from busiest to
3068 * this_rq, as part of a balancing operation within domain "sd".
3069 * Returns 1 if successful and 0 otherwise.
3071 * Called with both runqueues locked.
3073 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3074 unsigned long max_load_move,
3075 struct sched_domain *sd, enum cpu_idle_type idle,
3078 const struct sched_class *class = sched_class_highest;
3079 unsigned long total_load_moved = 0;
3080 int this_best_prio = this_rq->curr->prio;
3084 class->load_balance(this_rq, this_cpu, busiest,
3085 max_load_move - total_load_moved,
3086 sd, idle, all_pinned, &this_best_prio);
3087 class = class->next;
3089 #ifdef CONFIG_PREEMPT
3091 * NEWIDLE balancing is a source of latency, so preemptible
3092 * kernels will stop after the first task is pulled to minimize
3093 * the critical section.
3095 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3098 } while (class && max_load_move > total_load_moved);
3100 return total_load_moved > 0;
3104 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3105 struct sched_domain *sd, enum cpu_idle_type idle,
3106 struct rq_iterator *iterator)
3108 struct task_struct *p = iterator->start(iterator->arg);
3112 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3113 pull_task(busiest, p, this_rq, this_cpu);
3115 * Right now, this is only the second place pull_task()
3116 * is called, so we can safely collect pull_task()
3117 * stats here rather than inside pull_task().
3119 schedstat_inc(sd, lb_gained[idle]);
3123 p = iterator->next(iterator->arg);
3130 * move_one_task tries to move exactly one task from busiest to this_rq, as
3131 * part of active balancing operations within "domain".
3132 * Returns 1 if successful and 0 otherwise.
3134 * Called with both runqueues locked.
3136 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3137 struct sched_domain *sd, enum cpu_idle_type idle)
3139 const struct sched_class *class;
3141 for (class = sched_class_highest; class; class = class->next)
3142 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3149 * find_busiest_group finds and returns the busiest CPU group within the
3150 * domain. It calculates and returns the amount of weighted load which
3151 * should be moved to restore balance via the imbalance parameter.
3153 static struct sched_group *
3154 find_busiest_group(struct sched_domain *sd, int this_cpu,
3155 unsigned long *imbalance, enum cpu_idle_type idle,
3156 int *sd_idle, const struct cpumask *cpus, int *balance)
3158 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3159 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3160 unsigned long max_pull;
3161 unsigned long busiest_load_per_task, busiest_nr_running;
3162 unsigned long this_load_per_task, this_nr_running;
3163 int load_idx, group_imb = 0;
3164 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3165 int power_savings_balance = 1;
3166 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3167 unsigned long min_nr_running = ULONG_MAX;
3168 struct sched_group *group_min = NULL, *group_leader = NULL;
3171 max_load = this_load = total_load = total_pwr = 0;
3172 busiest_load_per_task = busiest_nr_running = 0;
3173 this_load_per_task = this_nr_running = 0;
3175 if (idle == CPU_NOT_IDLE)
3176 load_idx = sd->busy_idx;
3177 else if (idle == CPU_NEWLY_IDLE)
3178 load_idx = sd->newidle_idx;
3180 load_idx = sd->idle_idx;
3183 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3186 int __group_imb = 0;
3187 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3188 unsigned long sum_nr_running, sum_weighted_load;
3189 unsigned long sum_avg_load_per_task;
3190 unsigned long avg_load_per_task;
3192 local_group = cpumask_test_cpu(this_cpu,
3193 sched_group_cpus(group));
3196 balance_cpu = cpumask_first(sched_group_cpus(group));
3198 /* Tally up the load of all CPUs in the group */
3199 sum_weighted_load = sum_nr_running = avg_load = 0;
3200 sum_avg_load_per_task = avg_load_per_task = 0;
3203 min_cpu_load = ~0UL;
3205 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3206 struct rq *rq = cpu_rq(i);
3208 if (*sd_idle && rq->nr_running)
3211 /* Bias balancing toward cpus of our domain */
3213 if (idle_cpu(i) && !first_idle_cpu) {
3218 load = target_load(i, load_idx);
3220 load = source_load(i, load_idx);
3221 if (load > max_cpu_load)
3222 max_cpu_load = load;
3223 if (min_cpu_load > load)
3224 min_cpu_load = load;
3228 sum_nr_running += rq->nr_running;
3229 sum_weighted_load += weighted_cpuload(i);
3231 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3235 * First idle cpu or the first cpu(busiest) in this sched group
3236 * is eligible for doing load balancing at this and above
3237 * domains. In the newly idle case, we will allow all the cpu's
3238 * to do the newly idle load balance.
3240 if (idle != CPU_NEWLY_IDLE && local_group &&
3241 balance_cpu != this_cpu && balance) {
3246 total_load += avg_load;
3247 total_pwr += group->__cpu_power;
3249 /* Adjust by relative CPU power of the group */
3250 avg_load = sg_div_cpu_power(group,
3251 avg_load * SCHED_LOAD_SCALE);
3255 * Consider the group unbalanced when the imbalance is larger
3256 * than the average weight of two tasks.
3258 * APZ: with cgroup the avg task weight can vary wildly and
3259 * might not be a suitable number - should we keep a
3260 * normalized nr_running number somewhere that negates
3263 avg_load_per_task = sg_div_cpu_power(group,
3264 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3266 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3269 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3272 this_load = avg_load;
3274 this_nr_running = sum_nr_running;
3275 this_load_per_task = sum_weighted_load;
3276 } else if (avg_load > max_load &&
3277 (sum_nr_running > group_capacity || __group_imb)) {
3278 max_load = avg_load;
3280 busiest_nr_running = sum_nr_running;
3281 busiest_load_per_task = sum_weighted_load;
3282 group_imb = __group_imb;
3285 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3287 * Busy processors will not participate in power savings
3290 if (idle == CPU_NOT_IDLE ||
3291 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3295 * If the local group is idle or completely loaded
3296 * no need to do power savings balance at this domain
3298 if (local_group && (this_nr_running >= group_capacity ||
3300 power_savings_balance = 0;
3303 * If a group is already running at full capacity or idle,
3304 * don't include that group in power savings calculations
3306 if (!power_savings_balance || sum_nr_running >= group_capacity
3311 * Calculate the group which has the least non-idle load.
3312 * This is the group from where we need to pick up the load
3315 if ((sum_nr_running < min_nr_running) ||
3316 (sum_nr_running == min_nr_running &&
3317 cpumask_first(sched_group_cpus(group)) >
3318 cpumask_first(sched_group_cpus(group_min)))) {
3320 min_nr_running = sum_nr_running;
3321 min_load_per_task = sum_weighted_load /
3326 * Calculate the group which is almost near its
3327 * capacity but still has some space to pick up some load
3328 * from other group and save more power
3330 if (sum_nr_running <= group_capacity - 1) {
3331 if (sum_nr_running > leader_nr_running ||
3332 (sum_nr_running == leader_nr_running &&
3333 cpumask_first(sched_group_cpus(group)) <
3334 cpumask_first(sched_group_cpus(group_leader)))) {
3335 group_leader = group;
3336 leader_nr_running = sum_nr_running;
3341 group = group->next;
3342 } while (group != sd->groups);
3344 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3347 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3349 if (this_load >= avg_load ||
3350 100*max_load <= sd->imbalance_pct*this_load)
3353 busiest_load_per_task /= busiest_nr_running;
3355 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3358 * We're trying to get all the cpus to the average_load, so we don't
3359 * want to push ourselves above the average load, nor do we wish to
3360 * reduce the max loaded cpu below the average load, as either of these
3361 * actions would just result in more rebalancing later, and ping-pong
3362 * tasks around. Thus we look for the minimum possible imbalance.
3363 * Negative imbalances (*we* are more loaded than anyone else) will
3364 * be counted as no imbalance for these purposes -- we can't fix that
3365 * by pulling tasks to us. Be careful of negative numbers as they'll
3366 * appear as very large values with unsigned longs.
3368 if (max_load <= busiest_load_per_task)
3372 * In the presence of smp nice balancing, certain scenarios can have
3373 * max load less than avg load(as we skip the groups at or below
3374 * its cpu_power, while calculating max_load..)
3376 if (max_load < avg_load) {
3378 goto small_imbalance;
3381 /* Don't want to pull so many tasks that a group would go idle */
3382 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3384 /* How much load to actually move to equalise the imbalance */
3385 *imbalance = min(max_pull * busiest->__cpu_power,
3386 (avg_load - this_load) * this->__cpu_power)
3390 * if *imbalance is less than the average load per runnable task
3391 * there is no gaurantee that any tasks will be moved so we'll have
3392 * a think about bumping its value to force at least one task to be
3395 if (*imbalance < busiest_load_per_task) {
3396 unsigned long tmp, pwr_now, pwr_move;
3400 pwr_move = pwr_now = 0;
3402 if (this_nr_running) {
3403 this_load_per_task /= this_nr_running;
3404 if (busiest_load_per_task > this_load_per_task)
3407 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3409 if (max_load - this_load + busiest_load_per_task >=
3410 busiest_load_per_task * imbn) {
3411 *imbalance = busiest_load_per_task;
3416 * OK, we don't have enough imbalance to justify moving tasks,
3417 * however we may be able to increase total CPU power used by
3421 pwr_now += busiest->__cpu_power *
3422 min(busiest_load_per_task, max_load);
3423 pwr_now += this->__cpu_power *
3424 min(this_load_per_task, this_load);
3425 pwr_now /= SCHED_LOAD_SCALE;
3427 /* Amount of load we'd subtract */
3428 tmp = sg_div_cpu_power(busiest,
3429 busiest_load_per_task * SCHED_LOAD_SCALE);
3431 pwr_move += busiest->__cpu_power *
3432 min(busiest_load_per_task, max_load - tmp);
3434 /* Amount of load we'd add */
3435 if (max_load * busiest->__cpu_power <
3436 busiest_load_per_task * SCHED_LOAD_SCALE)
3437 tmp = sg_div_cpu_power(this,
3438 max_load * busiest->__cpu_power);
3440 tmp = sg_div_cpu_power(this,
3441 busiest_load_per_task * SCHED_LOAD_SCALE);
3442 pwr_move += this->__cpu_power *
3443 min(this_load_per_task, this_load + tmp);
3444 pwr_move /= SCHED_LOAD_SCALE;
3446 /* Move if we gain throughput */
3447 if (pwr_move > pwr_now)
3448 *imbalance = busiest_load_per_task;
3454 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3455 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3458 if (this == group_leader && group_leader != group_min) {
3459 *imbalance = min_load_per_task;
3460 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3461 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3462 cpumask_first(sched_group_cpus(group_leader));
3473 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3476 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3477 unsigned long imbalance, const struct cpumask *cpus)
3479 struct rq *busiest = NULL, *rq;
3480 unsigned long max_load = 0;
3483 for_each_cpu(i, sched_group_cpus(group)) {
3486 if (!cpumask_test_cpu(i, cpus))
3490 wl = weighted_cpuload(i);
3492 if (rq->nr_running == 1 && wl > imbalance)
3495 if (wl > max_load) {
3505 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3506 * so long as it is large enough.
3508 #define MAX_PINNED_INTERVAL 512
3511 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3512 * tasks if there is an imbalance.
3514 static int load_balance(int this_cpu, struct rq *this_rq,
3515 struct sched_domain *sd, enum cpu_idle_type idle,
3516 int *balance, struct cpumask *cpus)
3518 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3519 struct sched_group *group;
3520 unsigned long imbalance;
3522 unsigned long flags;
3524 cpumask_setall(cpus);
3527 * When power savings policy is enabled for the parent domain, idle
3528 * sibling can pick up load irrespective of busy siblings. In this case,
3529 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3530 * portraying it as CPU_NOT_IDLE.
3532 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3533 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3536 schedstat_inc(sd, lb_count[idle]);
3540 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3547 schedstat_inc(sd, lb_nobusyg[idle]);
3551 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3553 schedstat_inc(sd, lb_nobusyq[idle]);
3557 BUG_ON(busiest == this_rq);
3559 schedstat_add(sd, lb_imbalance[idle], imbalance);
3562 if (busiest->nr_running > 1) {
3564 * Attempt to move tasks. If find_busiest_group has found
3565 * an imbalance but busiest->nr_running <= 1, the group is
3566 * still unbalanced. ld_moved simply stays zero, so it is
3567 * correctly treated as an imbalance.
3569 local_irq_save(flags);
3570 double_rq_lock(this_rq, busiest);
3571 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3572 imbalance, sd, idle, &all_pinned);
3573 double_rq_unlock(this_rq, busiest);
3574 local_irq_restore(flags);
3577 * some other cpu did the load balance for us.
3579 if (ld_moved && this_cpu != smp_processor_id())
3580 resched_cpu(this_cpu);
3582 /* All tasks on this runqueue were pinned by CPU affinity */
3583 if (unlikely(all_pinned)) {
3584 cpumask_clear_cpu(cpu_of(busiest), cpus);
3585 if (!cpumask_empty(cpus))
3592 schedstat_inc(sd, lb_failed[idle]);
3593 sd->nr_balance_failed++;
3595 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3597 spin_lock_irqsave(&busiest->lock, flags);
3599 /* don't kick the migration_thread, if the curr
3600 * task on busiest cpu can't be moved to this_cpu
3602 if (!cpumask_test_cpu(this_cpu,
3603 &busiest->curr->cpus_allowed)) {
3604 spin_unlock_irqrestore(&busiest->lock, flags);
3606 goto out_one_pinned;
3609 if (!busiest->active_balance) {
3610 busiest->active_balance = 1;
3611 busiest->push_cpu = this_cpu;
3614 spin_unlock_irqrestore(&busiest->lock, flags);
3616 wake_up_process(busiest->migration_thread);
3619 * We've kicked active balancing, reset the failure
3622 sd->nr_balance_failed = sd->cache_nice_tries+1;
3625 sd->nr_balance_failed = 0;
3627 if (likely(!active_balance)) {
3628 /* We were unbalanced, so reset the balancing interval */
3629 sd->balance_interval = sd->min_interval;
3632 * If we've begun active balancing, start to back off. This
3633 * case may not be covered by the all_pinned logic if there
3634 * is only 1 task on the busy runqueue (because we don't call
3637 if (sd->balance_interval < sd->max_interval)
3638 sd->balance_interval *= 2;
3641 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3642 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3648 schedstat_inc(sd, lb_balanced[idle]);
3650 sd->nr_balance_failed = 0;
3653 /* tune up the balancing interval */
3654 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3655 (sd->balance_interval < sd->max_interval))
3656 sd->balance_interval *= 2;
3658 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3659 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3670 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3671 * tasks if there is an imbalance.
3673 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3674 * this_rq is locked.
3677 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3678 struct cpumask *cpus)
3680 struct sched_group *group;
3681 struct rq *busiest = NULL;
3682 unsigned long imbalance;
3687 cpumask_setall(cpus);
3690 * When power savings policy is enabled for the parent domain, idle
3691 * sibling can pick up load irrespective of busy siblings. In this case,
3692 * let the state of idle sibling percolate up as IDLE, instead of
3693 * portraying it as CPU_NOT_IDLE.
3695 if (sd->flags & SD_SHARE_CPUPOWER &&
3696 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3699 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3701 update_shares_locked(this_rq, sd);
3702 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3703 &sd_idle, cpus, NULL);
3705 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3709 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3711 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3715 BUG_ON(busiest == this_rq);
3717 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3720 if (busiest->nr_running > 1) {
3721 /* Attempt to move tasks */
3722 double_lock_balance(this_rq, busiest);
3723 /* this_rq->clock is already updated */
3724 update_rq_clock(busiest);
3725 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3726 imbalance, sd, CPU_NEWLY_IDLE,
3728 double_unlock_balance(this_rq, busiest);
3730 if (unlikely(all_pinned)) {
3731 cpumask_clear_cpu(cpu_of(busiest), cpus);
3732 if (!cpumask_empty(cpus))
3738 int active_balance = 0;
3740 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3741 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3742 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3745 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3748 if (sd->nr_balance_failed++ < 2)
3752 * The only task running in a non-idle cpu can be moved to this
3753 * cpu in an attempt to completely freeup the other CPU
3754 * package. The same method used to move task in load_balance()
3755 * have been extended for load_balance_newidle() to speedup
3756 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
3758 * The package power saving logic comes from
3759 * find_busiest_group(). If there are no imbalance, then
3760 * f_b_g() will return NULL. However when sched_mc={1,2} then
3761 * f_b_g() will select a group from which a running task may be
3762 * pulled to this cpu in order to make the other package idle.
3763 * If there is no opportunity to make a package idle and if
3764 * there are no imbalance, then f_b_g() will return NULL and no
3765 * action will be taken in load_balance_newidle().
3767 * Under normal task pull operation due to imbalance, there
3768 * will be more than one task in the source run queue and
3769 * move_tasks() will succeed. ld_moved will be true and this
3770 * active balance code will not be triggered.
3773 /* Lock busiest in correct order while this_rq is held */
3774 double_lock_balance(this_rq, busiest);
3777 * don't kick the migration_thread, if the curr
3778 * task on busiest cpu can't be moved to this_cpu
3780 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3781 double_unlock_balance(this_rq, busiest);
3786 if (!busiest->active_balance) {
3787 busiest->active_balance = 1;
3788 busiest->push_cpu = this_cpu;
3792 double_unlock_balance(this_rq, busiest);
3794 wake_up_process(busiest->migration_thread);
3797 sd->nr_balance_failed = 0;
3799 update_shares_locked(this_rq, sd);
3803 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3804 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3805 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3807 sd->nr_balance_failed = 0;
3813 * idle_balance is called by schedule() if this_cpu is about to become
3814 * idle. Attempts to pull tasks from other CPUs.
3816 static void idle_balance(int this_cpu, struct rq *this_rq)
3818 struct sched_domain *sd;
3819 int pulled_task = 0;
3820 unsigned long next_balance = jiffies + HZ;
3821 cpumask_var_t tmpmask;
3823 if (!alloc_cpumask_var(&tmpmask, GFP_ATOMIC))
3826 for_each_domain(this_cpu, sd) {
3827 unsigned long interval;
3829 if (!(sd->flags & SD_LOAD_BALANCE))
3832 if (sd->flags & SD_BALANCE_NEWIDLE)
3833 /* If we've pulled tasks over stop searching: */
3834 pulled_task = load_balance_newidle(this_cpu, this_rq,
3837 interval = msecs_to_jiffies(sd->balance_interval);
3838 if (time_after(next_balance, sd->last_balance + interval))
3839 next_balance = sd->last_balance + interval;
3843 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3845 * We are going idle. next_balance may be set based on
3846 * a busy processor. So reset next_balance.
3848 this_rq->next_balance = next_balance;
3850 free_cpumask_var(tmpmask);
3854 * active_load_balance is run by migration threads. It pushes running tasks
3855 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3856 * running on each physical CPU where possible, and avoids physical /
3857 * logical imbalances.
3859 * Called with busiest_rq locked.
3861 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3863 int target_cpu = busiest_rq->push_cpu;
3864 struct sched_domain *sd;
3865 struct rq *target_rq;
3867 /* Is there any task to move? */
3868 if (busiest_rq->nr_running <= 1)
3871 target_rq = cpu_rq(target_cpu);
3874 * This condition is "impossible", if it occurs
3875 * we need to fix it. Originally reported by
3876 * Bjorn Helgaas on a 128-cpu setup.
3878 BUG_ON(busiest_rq == target_rq);
3880 /* move a task from busiest_rq to target_rq */
3881 double_lock_balance(busiest_rq, target_rq);
3882 update_rq_clock(busiest_rq);
3883 update_rq_clock(target_rq);
3885 /* Search for an sd spanning us and the target CPU. */
3886 for_each_domain(target_cpu, sd) {
3887 if ((sd->flags & SD_LOAD_BALANCE) &&
3888 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3893 schedstat_inc(sd, alb_count);
3895 if (move_one_task(target_rq, target_cpu, busiest_rq,
3897 schedstat_inc(sd, alb_pushed);
3899 schedstat_inc(sd, alb_failed);
3901 double_unlock_balance(busiest_rq, target_rq);
3906 atomic_t load_balancer;
3907 cpumask_var_t cpu_mask;
3908 } nohz ____cacheline_aligned = {
3909 .load_balancer = ATOMIC_INIT(-1),
3913 * This routine will try to nominate the ilb (idle load balancing)
3914 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3915 * load balancing on behalf of all those cpus. If all the cpus in the system
3916 * go into this tickless mode, then there will be no ilb owner (as there is
3917 * no need for one) and all the cpus will sleep till the next wakeup event
3920 * For the ilb owner, tick is not stopped. And this tick will be used
3921 * for idle load balancing. ilb owner will still be part of
3924 * While stopping the tick, this cpu will become the ilb owner if there
3925 * is no other owner. And will be the owner till that cpu becomes busy
3926 * or if all cpus in the system stop their ticks at which point
3927 * there is no need for ilb owner.
3929 * When the ilb owner becomes busy, it nominates another owner, during the
3930 * next busy scheduler_tick()
3932 int select_nohz_load_balancer(int stop_tick)
3934 int cpu = smp_processor_id();
3937 cpumask_set_cpu(cpu, nohz.cpu_mask);
3938 cpu_rq(cpu)->in_nohz_recently = 1;
3941 * If we are going offline and still the leader, give up!
3943 if (!cpu_active(cpu) &&
3944 atomic_read(&nohz.load_balancer) == cpu) {
3945 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3950 /* time for ilb owner also to sleep */
3951 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
3952 if (atomic_read(&nohz.load_balancer) == cpu)
3953 atomic_set(&nohz.load_balancer, -1);
3957 if (atomic_read(&nohz.load_balancer) == -1) {
3958 /* make me the ilb owner */
3959 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3961 } else if (atomic_read(&nohz.load_balancer) == cpu)
3964 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
3967 cpumask_clear_cpu(cpu, nohz.cpu_mask);
3969 if (atomic_read(&nohz.load_balancer) == cpu)
3970 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3977 static DEFINE_SPINLOCK(balancing);
3980 * It checks each scheduling domain to see if it is due to be balanced,
3981 * and initiates a balancing operation if so.
3983 * Balancing parameters are set up in arch_init_sched_domains.
3985 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3988 struct rq *rq = cpu_rq(cpu);
3989 unsigned long interval;
3990 struct sched_domain *sd;
3991 /* Earliest time when we have to do rebalance again */
3992 unsigned long next_balance = jiffies + 60*HZ;
3993 int update_next_balance = 0;
3997 /* Fails alloc? Rebalancing probably not a priority right now. */
3998 if (!alloc_cpumask_var(&tmp, GFP_ATOMIC))
4001 for_each_domain(cpu, sd) {
4002 if (!(sd->flags & SD_LOAD_BALANCE))
4005 interval = sd->balance_interval;
4006 if (idle != CPU_IDLE)
4007 interval *= sd->busy_factor;
4009 /* scale ms to jiffies */
4010 interval = msecs_to_jiffies(interval);
4011 if (unlikely(!interval))
4013 if (interval > HZ*NR_CPUS/10)
4014 interval = HZ*NR_CPUS/10;
4016 need_serialize = sd->flags & SD_SERIALIZE;
4018 if (need_serialize) {
4019 if (!spin_trylock(&balancing))
4023 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4024 if (load_balance(cpu, rq, sd, idle, &balance, tmp)) {
4026 * We've pulled tasks over so either we're no
4027 * longer idle, or one of our SMT siblings is
4030 idle = CPU_NOT_IDLE;
4032 sd->last_balance = jiffies;
4035 spin_unlock(&balancing);
4037 if (time_after(next_balance, sd->last_balance + interval)) {
4038 next_balance = sd->last_balance + interval;
4039 update_next_balance = 1;
4043 * Stop the load balance at this level. There is another
4044 * CPU in our sched group which is doing load balancing more
4052 * next_balance will be updated only when there is a need.
4053 * When the cpu is attached to null domain for ex, it will not be
4056 if (likely(update_next_balance))
4057 rq->next_balance = next_balance;
4059 free_cpumask_var(tmp);
4063 * run_rebalance_domains is triggered when needed from the scheduler tick.
4064 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4065 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4067 static void run_rebalance_domains(struct softirq_action *h)
4069 int this_cpu = smp_processor_id();
4070 struct rq *this_rq = cpu_rq(this_cpu);
4071 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4072 CPU_IDLE : CPU_NOT_IDLE;
4074 rebalance_domains(this_cpu, idle);
4078 * If this cpu is the owner for idle load balancing, then do the
4079 * balancing on behalf of the other idle cpus whose ticks are
4082 if (this_rq->idle_at_tick &&
4083 atomic_read(&nohz.load_balancer) == this_cpu) {
4087 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4088 if (balance_cpu == this_cpu)
4092 * If this cpu gets work to do, stop the load balancing
4093 * work being done for other cpus. Next load
4094 * balancing owner will pick it up.
4099 rebalance_domains(balance_cpu, CPU_IDLE);
4101 rq = cpu_rq(balance_cpu);
4102 if (time_after(this_rq->next_balance, rq->next_balance))
4103 this_rq->next_balance = rq->next_balance;
4110 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4112 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4113 * idle load balancing owner or decide to stop the periodic load balancing,
4114 * if the whole system is idle.
4116 static inline void trigger_load_balance(struct rq *rq, int cpu)
4120 * If we were in the nohz mode recently and busy at the current
4121 * scheduler tick, then check if we need to nominate new idle
4124 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4125 rq->in_nohz_recently = 0;
4127 if (atomic_read(&nohz.load_balancer) == cpu) {
4128 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4129 atomic_set(&nohz.load_balancer, -1);
4132 if (atomic_read(&nohz.load_balancer) == -1) {
4134 * simple selection for now: Nominate the
4135 * first cpu in the nohz list to be the next
4138 * TBD: Traverse the sched domains and nominate
4139 * the nearest cpu in the nohz.cpu_mask.
4141 int ilb = cpumask_first(nohz.cpu_mask);
4143 if (ilb < nr_cpu_ids)
4149 * If this cpu is idle and doing idle load balancing for all the
4150 * cpus with ticks stopped, is it time for that to stop?
4152 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4153 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4159 * If this cpu is idle and the idle load balancing is done by
4160 * someone else, then no need raise the SCHED_SOFTIRQ
4162 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4163 cpumask_test_cpu(cpu, nohz.cpu_mask))
4166 if (time_after_eq(jiffies, rq->next_balance))
4167 raise_softirq(SCHED_SOFTIRQ);
4170 #else /* CONFIG_SMP */
4173 * on UP we do not need to balance between CPUs:
4175 static inline void idle_balance(int cpu, struct rq *rq)
4181 DEFINE_PER_CPU(struct kernel_stat, kstat);
4183 EXPORT_PER_CPU_SYMBOL(kstat);
4186 * Return any ns on the sched_clock that have not yet been banked in
4187 * @p in case that task is currently running.
4189 unsigned long long task_delta_exec(struct task_struct *p)
4191 unsigned long flags;
4195 rq = task_rq_lock(p, &flags);
4197 if (task_current(rq, p)) {
4200 update_rq_clock(rq);
4201 delta_exec = rq->clock - p->se.exec_start;
4202 if ((s64)delta_exec > 0)
4206 task_rq_unlock(rq, &flags);
4212 * Account user cpu time to a process.
4213 * @p: the process that the cpu time gets accounted to
4214 * @cputime: the cpu time spent in user space since the last update
4216 void account_user_time(struct task_struct *p, cputime_t cputime)
4218 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4221 p->utime = cputime_add(p->utime, cputime);
4222 account_group_user_time(p, cputime);
4224 /* Add user time to cpustat. */
4225 tmp = cputime_to_cputime64(cputime);
4226 if (TASK_NICE(p) > 0)
4227 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4229 cpustat->user = cputime64_add(cpustat->user, tmp);
4230 /* Account for user time used */
4231 acct_update_integrals(p);
4235 * Account guest cpu time to a process.
4236 * @p: the process that the cpu time gets accounted to
4237 * @cputime: the cpu time spent in virtual machine since the last update
4239 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4242 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4244 tmp = cputime_to_cputime64(cputime);
4246 p->utime = cputime_add(p->utime, cputime);
4247 account_group_user_time(p, cputime);
4248 p->gtime = cputime_add(p->gtime, cputime);
4250 cpustat->user = cputime64_add(cpustat->user, tmp);
4251 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4255 * Account scaled user cpu time to a process.
4256 * @p: the process that the cpu time gets accounted to
4257 * @cputime: the cpu time spent in user space since the last update
4259 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4261 p->utimescaled = cputime_add(p->utimescaled, cputime);
4265 * Account system cpu time to a process.
4266 * @p: the process that the cpu time gets accounted to
4267 * @hardirq_offset: the offset to subtract from hardirq_count()
4268 * @cputime: the cpu time spent in kernel space since the last update
4270 void account_system_time(struct task_struct *p, int hardirq_offset,
4273 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4274 struct rq *rq = this_rq();
4277 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4278 account_guest_time(p, cputime);
4282 p->stime = cputime_add(p->stime, cputime);
4283 account_group_system_time(p, cputime);
4285 /* Add system time to cpustat. */
4286 tmp = cputime_to_cputime64(cputime);
4287 if (hardirq_count() - hardirq_offset)
4288 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4289 else if (softirq_count())
4290 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4291 else if (p != rq->idle)
4292 cpustat->system = cputime64_add(cpustat->system, tmp);
4293 else if (atomic_read(&rq->nr_iowait) > 0)
4294 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4296 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4297 /* Account for system time used */
4298 acct_update_integrals(p);
4302 * Account scaled system cpu time to a process.
4303 * @p: the process that the cpu time gets accounted to
4304 * @hardirq_offset: the offset to subtract from hardirq_count()
4305 * @cputime: the cpu time spent in kernel space since the last update
4307 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4309 p->stimescaled = cputime_add(p->stimescaled, cputime);
4313 * Account for involuntary wait time.
4314 * @p: the process from which the cpu time has been stolen
4315 * @steal: the cpu time spent in involuntary wait
4317 void account_steal_time(struct task_struct *p, cputime_t steal)
4319 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4320 cputime64_t tmp = cputime_to_cputime64(steal);
4321 struct rq *rq = this_rq();
4323 if (p == rq->idle) {
4324 p->stime = cputime_add(p->stime, steal);
4325 if (atomic_read(&rq->nr_iowait) > 0)
4326 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4328 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4330 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4334 * Use precise platform statistics if available:
4336 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4337 cputime_t task_utime(struct task_struct *p)
4342 cputime_t task_stime(struct task_struct *p)
4347 cputime_t task_utime(struct task_struct *p)
4349 clock_t utime = cputime_to_clock_t(p->utime),
4350 total = utime + cputime_to_clock_t(p->stime);
4354 * Use CFS's precise accounting:
4356 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4360 do_div(temp, total);
4362 utime = (clock_t)temp;
4364 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4365 return p->prev_utime;
4368 cputime_t task_stime(struct task_struct *p)
4373 * Use CFS's precise accounting. (we subtract utime from
4374 * the total, to make sure the total observed by userspace
4375 * grows monotonically - apps rely on that):
4377 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4378 cputime_to_clock_t(task_utime(p));
4381 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4383 return p->prev_stime;
4387 inline cputime_t task_gtime(struct task_struct *p)
4393 * This function gets called by the timer code, with HZ frequency.
4394 * We call it with interrupts disabled.
4396 * It also gets called by the fork code, when changing the parent's
4399 void scheduler_tick(void)
4401 int cpu = smp_processor_id();
4402 struct rq *rq = cpu_rq(cpu);
4403 struct task_struct *curr = rq->curr;
4407 spin_lock(&rq->lock);
4408 update_rq_clock(rq);
4409 update_cpu_load(rq);
4410 curr->sched_class->task_tick(rq, curr, 0);
4411 spin_unlock(&rq->lock);
4414 rq->idle_at_tick = idle_cpu(cpu);
4415 trigger_load_balance(rq, cpu);
4419 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4420 defined(CONFIG_PREEMPT_TRACER))
4422 static inline unsigned long get_parent_ip(unsigned long addr)
4424 if (in_lock_functions(addr)) {
4425 addr = CALLER_ADDR2;
4426 if (in_lock_functions(addr))
4427 addr = CALLER_ADDR3;
4432 void __kprobes add_preempt_count(int val)
4434 #ifdef CONFIG_DEBUG_PREEMPT
4438 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4441 preempt_count() += val;
4442 #ifdef CONFIG_DEBUG_PREEMPT
4444 * Spinlock count overflowing soon?
4446 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4449 if (preempt_count() == val)
4450 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4452 EXPORT_SYMBOL(add_preempt_count);
4454 void __kprobes sub_preempt_count(int val)
4456 #ifdef CONFIG_DEBUG_PREEMPT
4460 if (DEBUG_LOCKS_WARN_ON(val > preempt_count() - (!!kernel_locked())))
4463 * Is the spinlock portion underflowing?
4465 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4466 !(preempt_count() & PREEMPT_MASK)))
4470 if (preempt_count() == val)
4471 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4472 preempt_count() -= val;
4474 EXPORT_SYMBOL(sub_preempt_count);
4479 * Print scheduling while atomic bug:
4481 static noinline void __schedule_bug(struct task_struct *prev)
4483 struct pt_regs *regs = get_irq_regs();
4485 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4486 prev->comm, prev->pid, preempt_count());
4488 debug_show_held_locks(prev);
4490 if (irqs_disabled())
4491 print_irqtrace_events(prev);
4500 * Various schedule()-time debugging checks and statistics:
4502 static inline void schedule_debug(struct task_struct *prev)
4505 * Test if we are atomic. Since do_exit() needs to call into
4506 * schedule() atomically, we ignore that path for now.
4507 * Otherwise, whine if we are scheduling when we should not be.
4509 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4510 __schedule_bug(prev);
4512 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4514 schedstat_inc(this_rq(), sched_count);
4515 #ifdef CONFIG_SCHEDSTATS
4516 if (unlikely(prev->lock_depth >= 0)) {
4517 schedstat_inc(this_rq(), bkl_count);
4518 schedstat_inc(prev, sched_info.bkl_count);
4524 * Pick up the highest-prio task:
4526 static inline struct task_struct *
4527 pick_next_task(struct rq *rq, struct task_struct *prev)
4529 const struct sched_class *class;
4530 struct task_struct *p;
4533 * Optimization: we know that if all tasks are in
4534 * the fair class we can call that function directly:
4536 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4537 p = fair_sched_class.pick_next_task(rq);
4542 class = sched_class_highest;
4544 p = class->pick_next_task(rq);
4548 * Will never be NULL as the idle class always
4549 * returns a non-NULL p:
4551 class = class->next;
4556 * schedule() is the main scheduler function.
4558 asmlinkage void __sched schedule(void)
4560 struct task_struct *prev, *next;
4561 unsigned long *switch_count;
4567 cpu = smp_processor_id();
4571 switch_count = &prev->nivcsw;
4573 release_kernel_lock(prev);
4574 need_resched_nonpreemptible:
4576 schedule_debug(prev);
4578 if (sched_feat(HRTICK))
4581 spin_lock_irq(&rq->lock);
4582 update_rq_clock(rq);
4583 clear_tsk_need_resched(prev);
4585 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4586 if (unlikely(signal_pending_state(prev->state, prev)))
4587 prev->state = TASK_RUNNING;
4589 deactivate_task(rq, prev, 1);
4590 switch_count = &prev->nvcsw;
4594 if (prev->sched_class->pre_schedule)
4595 prev->sched_class->pre_schedule(rq, prev);
4598 if (unlikely(!rq->nr_running))
4599 idle_balance(cpu, rq);
4601 prev->sched_class->put_prev_task(rq, prev);
4602 next = pick_next_task(rq, prev);
4604 if (likely(prev != next)) {
4605 sched_info_switch(prev, next);
4611 context_switch(rq, prev, next); /* unlocks the rq */
4613 * the context switch might have flipped the stack from under
4614 * us, hence refresh the local variables.
4616 cpu = smp_processor_id();
4619 spin_unlock_irq(&rq->lock);
4621 if (unlikely(reacquire_kernel_lock(current) < 0))
4622 goto need_resched_nonpreemptible;
4624 preempt_enable_no_resched();
4625 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4628 EXPORT_SYMBOL(schedule);
4630 #ifdef CONFIG_PREEMPT
4632 * this is the entry point to schedule() from in-kernel preemption
4633 * off of preempt_enable. Kernel preemptions off return from interrupt
4634 * occur there and call schedule directly.
4636 asmlinkage void __sched preempt_schedule(void)
4638 struct thread_info *ti = current_thread_info();
4641 * If there is a non-zero preempt_count or interrupts are disabled,
4642 * we do not want to preempt the current task. Just return..
4644 if (likely(ti->preempt_count || irqs_disabled()))
4648 add_preempt_count(PREEMPT_ACTIVE);
4650 sub_preempt_count(PREEMPT_ACTIVE);
4653 * Check again in case we missed a preemption opportunity
4654 * between schedule and now.
4657 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4659 EXPORT_SYMBOL(preempt_schedule);
4662 * this is the entry point to schedule() from kernel preemption
4663 * off of irq context.
4664 * Note, that this is called and return with irqs disabled. This will
4665 * protect us against recursive calling from irq.
4667 asmlinkage void __sched preempt_schedule_irq(void)
4669 struct thread_info *ti = current_thread_info();
4671 /* Catch callers which need to be fixed */
4672 BUG_ON(ti->preempt_count || !irqs_disabled());
4675 add_preempt_count(PREEMPT_ACTIVE);
4678 local_irq_disable();
4679 sub_preempt_count(PREEMPT_ACTIVE);
4682 * Check again in case we missed a preemption opportunity
4683 * between schedule and now.
4686 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4689 #endif /* CONFIG_PREEMPT */
4691 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4694 return try_to_wake_up(curr->private, mode, sync);
4696 EXPORT_SYMBOL(default_wake_function);
4699 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4700 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4701 * number) then we wake all the non-exclusive tasks and one exclusive task.
4703 * There are circumstances in which we can try to wake a task which has already
4704 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4705 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4707 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4708 int nr_exclusive, int sync, void *key)
4710 wait_queue_t *curr, *next;
4712 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4713 unsigned flags = curr->flags;
4715 if (curr->func(curr, mode, sync, key) &&
4716 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4722 * __wake_up - wake up threads blocked on a waitqueue.
4724 * @mode: which threads
4725 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4726 * @key: is directly passed to the wakeup function
4728 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4729 int nr_exclusive, void *key)
4731 unsigned long flags;
4733 spin_lock_irqsave(&q->lock, flags);
4734 __wake_up_common(q, mode, nr_exclusive, 0, key);
4735 spin_unlock_irqrestore(&q->lock, flags);
4737 EXPORT_SYMBOL(__wake_up);
4740 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4742 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4744 __wake_up_common(q, mode, 1, 0, NULL);
4748 * __wake_up_sync - wake up threads blocked on a waitqueue.
4750 * @mode: which threads
4751 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4753 * The sync wakeup differs that the waker knows that it will schedule
4754 * away soon, so while the target thread will be woken up, it will not
4755 * be migrated to another CPU - ie. the two threads are 'synchronized'
4756 * with each other. This can prevent needless bouncing between CPUs.
4758 * On UP it can prevent extra preemption.
4761 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4763 unsigned long flags;
4769 if (unlikely(!nr_exclusive))
4772 spin_lock_irqsave(&q->lock, flags);
4773 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4774 spin_unlock_irqrestore(&q->lock, flags);
4776 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4779 * complete: - signals a single thread waiting on this completion
4780 * @x: holds the state of this particular completion
4782 * This will wake up a single thread waiting on this completion. Threads will be
4783 * awakened in the same order in which they were queued.
4785 * See also complete_all(), wait_for_completion() and related routines.
4787 void complete(struct completion *x)
4789 unsigned long flags;
4791 spin_lock_irqsave(&x->wait.lock, flags);
4793 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4794 spin_unlock_irqrestore(&x->wait.lock, flags);
4796 EXPORT_SYMBOL(complete);
4799 * complete_all: - signals all threads waiting on this completion
4800 * @x: holds the state of this particular completion
4802 * This will wake up all threads waiting on this particular completion event.
4804 void complete_all(struct completion *x)
4806 unsigned long flags;
4808 spin_lock_irqsave(&x->wait.lock, flags);
4809 x->done += UINT_MAX/2;
4810 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4811 spin_unlock_irqrestore(&x->wait.lock, flags);
4813 EXPORT_SYMBOL(complete_all);
4815 static inline long __sched
4816 do_wait_for_common(struct completion *x, long timeout, int state)
4819 DECLARE_WAITQUEUE(wait, current);
4821 wait.flags |= WQ_FLAG_EXCLUSIVE;
4822 __add_wait_queue_tail(&x->wait, &wait);
4824 if (signal_pending_state(state, current)) {
4825 timeout = -ERESTARTSYS;
4828 __set_current_state(state);
4829 spin_unlock_irq(&x->wait.lock);
4830 timeout = schedule_timeout(timeout);
4831 spin_lock_irq(&x->wait.lock);
4832 } while (!x->done && timeout);
4833 __remove_wait_queue(&x->wait, &wait);
4838 return timeout ?: 1;
4842 wait_for_common(struct completion *x, long timeout, int state)
4846 spin_lock_irq(&x->wait.lock);
4847 timeout = do_wait_for_common(x, timeout, state);
4848 spin_unlock_irq(&x->wait.lock);
4853 * wait_for_completion: - waits for completion of a task
4854 * @x: holds the state of this particular completion
4856 * This waits to be signaled for completion of a specific task. It is NOT
4857 * interruptible and there is no timeout.
4859 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4860 * and interrupt capability. Also see complete().
4862 void __sched wait_for_completion(struct completion *x)
4864 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4866 EXPORT_SYMBOL(wait_for_completion);
4869 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4870 * @x: holds the state of this particular completion
4871 * @timeout: timeout value in jiffies
4873 * This waits for either a completion of a specific task to be signaled or for a
4874 * specified timeout to expire. The timeout is in jiffies. It is not
4877 unsigned long __sched
4878 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4880 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4882 EXPORT_SYMBOL(wait_for_completion_timeout);
4885 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4886 * @x: holds the state of this particular completion
4888 * This waits for completion of a specific task to be signaled. It is
4891 int __sched wait_for_completion_interruptible(struct completion *x)
4893 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4894 if (t == -ERESTARTSYS)
4898 EXPORT_SYMBOL(wait_for_completion_interruptible);
4901 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4902 * @x: holds the state of this particular completion
4903 * @timeout: timeout value in jiffies
4905 * This waits for either a completion of a specific task to be signaled or for a
4906 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4908 unsigned long __sched
4909 wait_for_completion_interruptible_timeout(struct completion *x,
4910 unsigned long timeout)
4912 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4914 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4917 * wait_for_completion_killable: - waits for completion of a task (killable)
4918 * @x: holds the state of this particular completion
4920 * This waits to be signaled for completion of a specific task. It can be
4921 * interrupted by a kill signal.
4923 int __sched wait_for_completion_killable(struct completion *x)
4925 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4926 if (t == -ERESTARTSYS)
4930 EXPORT_SYMBOL(wait_for_completion_killable);
4933 * try_wait_for_completion - try to decrement a completion without blocking
4934 * @x: completion structure
4936 * Returns: 0 if a decrement cannot be done without blocking
4937 * 1 if a decrement succeeded.
4939 * If a completion is being used as a counting completion,
4940 * attempt to decrement the counter without blocking. This
4941 * enables us to avoid waiting if the resource the completion
4942 * is protecting is not available.
4944 bool try_wait_for_completion(struct completion *x)
4948 spin_lock_irq(&x->wait.lock);
4953 spin_unlock_irq(&x->wait.lock);
4956 EXPORT_SYMBOL(try_wait_for_completion);
4959 * completion_done - Test to see if a completion has any waiters
4960 * @x: completion structure
4962 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4963 * 1 if there are no waiters.
4966 bool completion_done(struct completion *x)
4970 spin_lock_irq(&x->wait.lock);
4973 spin_unlock_irq(&x->wait.lock);
4976 EXPORT_SYMBOL(completion_done);
4979 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4981 unsigned long flags;
4984 init_waitqueue_entry(&wait, current);
4986 __set_current_state(state);
4988 spin_lock_irqsave(&q->lock, flags);
4989 __add_wait_queue(q, &wait);
4990 spin_unlock(&q->lock);
4991 timeout = schedule_timeout(timeout);
4992 spin_lock_irq(&q->lock);
4993 __remove_wait_queue(q, &wait);
4994 spin_unlock_irqrestore(&q->lock, flags);
4999 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5001 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5003 EXPORT_SYMBOL(interruptible_sleep_on);
5006 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5008 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5010 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5012 void __sched sleep_on(wait_queue_head_t *q)
5014 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5016 EXPORT_SYMBOL(sleep_on);
5018 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5020 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5022 EXPORT_SYMBOL(sleep_on_timeout);
5024 #ifdef CONFIG_RT_MUTEXES
5027 * rt_mutex_setprio - set the current priority of a task
5029 * @prio: prio value (kernel-internal form)
5031 * This function changes the 'effective' priority of a task. It does
5032 * not touch ->normal_prio like __setscheduler().
5034 * Used by the rt_mutex code to implement priority inheritance logic.
5036 void rt_mutex_setprio(struct task_struct *p, int prio)
5038 unsigned long flags;
5039 int oldprio, on_rq, running;
5041 const struct sched_class *prev_class = p->sched_class;
5043 BUG_ON(prio < 0 || prio > MAX_PRIO);
5045 rq = task_rq_lock(p, &flags);
5046 update_rq_clock(rq);
5049 on_rq = p->se.on_rq;
5050 running = task_current(rq, p);
5052 dequeue_task(rq, p, 0);
5054 p->sched_class->put_prev_task(rq, p);
5057 p->sched_class = &rt_sched_class;
5059 p->sched_class = &fair_sched_class;
5064 p->sched_class->set_curr_task(rq);
5066 enqueue_task(rq, p, 0);
5068 check_class_changed(rq, p, prev_class, oldprio, running);
5070 task_rq_unlock(rq, &flags);
5075 void set_user_nice(struct task_struct *p, long nice)
5077 int old_prio, delta, on_rq;
5078 unsigned long flags;
5081 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5084 * We have to be careful, if called from sys_setpriority(),
5085 * the task might be in the middle of scheduling on another CPU.
5087 rq = task_rq_lock(p, &flags);
5088 update_rq_clock(rq);
5090 * The RT priorities are set via sched_setscheduler(), but we still
5091 * allow the 'normal' nice value to be set - but as expected
5092 * it wont have any effect on scheduling until the task is
5093 * SCHED_FIFO/SCHED_RR:
5095 if (task_has_rt_policy(p)) {
5096 p->static_prio = NICE_TO_PRIO(nice);
5099 on_rq = p->se.on_rq;
5101 dequeue_task(rq, p, 0);
5103 p->static_prio = NICE_TO_PRIO(nice);
5106 p->prio = effective_prio(p);
5107 delta = p->prio - old_prio;
5110 enqueue_task(rq, p, 0);
5112 * If the task increased its priority or is running and
5113 * lowered its priority, then reschedule its CPU:
5115 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5116 resched_task(rq->curr);
5119 task_rq_unlock(rq, &flags);
5121 EXPORT_SYMBOL(set_user_nice);
5124 * can_nice - check if a task can reduce its nice value
5128 int can_nice(const struct task_struct *p, const int nice)
5130 /* convert nice value [19,-20] to rlimit style value [1,40] */
5131 int nice_rlim = 20 - nice;
5133 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5134 capable(CAP_SYS_NICE));
5137 #ifdef __ARCH_WANT_SYS_NICE
5140 * sys_nice - change the priority of the current process.
5141 * @increment: priority increment
5143 * sys_setpriority is a more generic, but much slower function that
5144 * does similar things.
5146 asmlinkage long sys_nice(int increment)
5151 * Setpriority might change our priority at the same moment.
5152 * We don't have to worry. Conceptually one call occurs first
5153 * and we have a single winner.
5155 if (increment < -40)
5160 nice = PRIO_TO_NICE(current->static_prio) + increment;
5166 if (increment < 0 && !can_nice(current, nice))
5169 retval = security_task_setnice(current, nice);
5173 set_user_nice(current, nice);
5180 * task_prio - return the priority value of a given task.
5181 * @p: the task in question.
5183 * This is the priority value as seen by users in /proc.
5184 * RT tasks are offset by -200. Normal tasks are centered
5185 * around 0, value goes from -16 to +15.
5187 int task_prio(const struct task_struct *p)
5189 return p->prio - MAX_RT_PRIO;
5193 * task_nice - return the nice value of a given task.
5194 * @p: the task in question.
5196 int task_nice(const struct task_struct *p)
5198 return TASK_NICE(p);
5200 EXPORT_SYMBOL(task_nice);
5203 * idle_cpu - is a given cpu idle currently?
5204 * @cpu: the processor in question.
5206 int idle_cpu(int cpu)
5208 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5212 * idle_task - return the idle task for a given cpu.
5213 * @cpu: the processor in question.
5215 struct task_struct *idle_task(int cpu)
5217 return cpu_rq(cpu)->idle;
5221 * find_process_by_pid - find a process with a matching PID value.
5222 * @pid: the pid in question.
5224 static struct task_struct *find_process_by_pid(pid_t pid)
5226 return pid ? find_task_by_vpid(pid) : current;
5229 /* Actually do priority change: must hold rq lock. */
5231 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5233 BUG_ON(p->se.on_rq);
5236 switch (p->policy) {
5240 p->sched_class = &fair_sched_class;
5244 p->sched_class = &rt_sched_class;
5248 p->rt_priority = prio;
5249 p->normal_prio = normal_prio(p);
5250 /* we are holding p->pi_lock already */
5251 p->prio = rt_mutex_getprio(p);
5255 static int __sched_setscheduler(struct task_struct *p, int policy,
5256 struct sched_param *param, bool user)
5258 int retval, oldprio, oldpolicy = -1, on_rq, running;
5259 unsigned long flags;
5260 const struct sched_class *prev_class = p->sched_class;
5263 /* may grab non-irq protected spin_locks */
5264 BUG_ON(in_interrupt());
5266 /* double check policy once rq lock held */
5268 policy = oldpolicy = p->policy;
5269 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5270 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5271 policy != SCHED_IDLE)
5274 * Valid priorities for SCHED_FIFO and SCHED_RR are
5275 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5276 * SCHED_BATCH and SCHED_IDLE is 0.
5278 if (param->sched_priority < 0 ||
5279 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5280 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5282 if (rt_policy(policy) != (param->sched_priority != 0))
5286 * Allow unprivileged RT tasks to decrease priority:
5288 if (user && !capable(CAP_SYS_NICE)) {
5289 if (rt_policy(policy)) {
5290 unsigned long rlim_rtprio;
5292 if (!lock_task_sighand(p, &flags))
5294 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5295 unlock_task_sighand(p, &flags);
5297 /* can't set/change the rt policy */
5298 if (policy != p->policy && !rlim_rtprio)
5301 /* can't increase priority */
5302 if (param->sched_priority > p->rt_priority &&
5303 param->sched_priority > rlim_rtprio)
5307 * Like positive nice levels, dont allow tasks to
5308 * move out of SCHED_IDLE either:
5310 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5313 /* can't change other user's priorities */
5314 if ((current->euid != p->euid) &&
5315 (current->euid != p->uid))
5320 #ifdef CONFIG_RT_GROUP_SCHED
5322 * Do not allow realtime tasks into groups that have no runtime
5325 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5326 task_group(p)->rt_bandwidth.rt_runtime == 0)
5330 retval = security_task_setscheduler(p, policy, param);
5336 * make sure no PI-waiters arrive (or leave) while we are
5337 * changing the priority of the task:
5339 spin_lock_irqsave(&p->pi_lock, flags);
5341 * To be able to change p->policy safely, the apropriate
5342 * runqueue lock must be held.
5344 rq = __task_rq_lock(p);
5345 /* recheck policy now with rq lock held */
5346 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5347 policy = oldpolicy = -1;
5348 __task_rq_unlock(rq);
5349 spin_unlock_irqrestore(&p->pi_lock, flags);
5352 update_rq_clock(rq);
5353 on_rq = p->se.on_rq;
5354 running = task_current(rq, p);
5356 deactivate_task(rq, p, 0);
5358 p->sched_class->put_prev_task(rq, p);
5361 __setscheduler(rq, p, policy, param->sched_priority);
5364 p->sched_class->set_curr_task(rq);
5366 activate_task(rq, p, 0);
5368 check_class_changed(rq, p, prev_class, oldprio, running);
5370 __task_rq_unlock(rq);
5371 spin_unlock_irqrestore(&p->pi_lock, flags);
5373 rt_mutex_adjust_pi(p);
5379 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5380 * @p: the task in question.
5381 * @policy: new policy.
5382 * @param: structure containing the new RT priority.
5384 * NOTE that the task may be already dead.
5386 int sched_setscheduler(struct task_struct *p, int policy,
5387 struct sched_param *param)
5389 return __sched_setscheduler(p, policy, param, true);
5391 EXPORT_SYMBOL_GPL(sched_setscheduler);
5394 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5395 * @p: the task in question.
5396 * @policy: new policy.
5397 * @param: structure containing the new RT priority.
5399 * Just like sched_setscheduler, only don't bother checking if the
5400 * current context has permission. For example, this is needed in
5401 * stop_machine(): we create temporary high priority worker threads,
5402 * but our caller might not have that capability.
5404 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5405 struct sched_param *param)
5407 return __sched_setscheduler(p, policy, param, false);
5411 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5413 struct sched_param lparam;
5414 struct task_struct *p;
5417 if (!param || pid < 0)
5419 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5424 p = find_process_by_pid(pid);
5426 retval = sched_setscheduler(p, policy, &lparam);
5433 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5434 * @pid: the pid in question.
5435 * @policy: new policy.
5436 * @param: structure containing the new RT priority.
5439 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5441 /* negative values for policy are not valid */
5445 return do_sched_setscheduler(pid, policy, param);
5449 * sys_sched_setparam - set/change the RT priority of a thread
5450 * @pid: the pid in question.
5451 * @param: structure containing the new RT priority.
5453 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5455 return do_sched_setscheduler(pid, -1, param);
5459 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5460 * @pid: the pid in question.
5462 asmlinkage long sys_sched_getscheduler(pid_t pid)
5464 struct task_struct *p;
5471 read_lock(&tasklist_lock);
5472 p = find_process_by_pid(pid);
5474 retval = security_task_getscheduler(p);
5478 read_unlock(&tasklist_lock);
5483 * sys_sched_getscheduler - get the RT priority of a thread
5484 * @pid: the pid in question.
5485 * @param: structure containing the RT priority.
5487 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5489 struct sched_param lp;
5490 struct task_struct *p;
5493 if (!param || pid < 0)
5496 read_lock(&tasklist_lock);
5497 p = find_process_by_pid(pid);
5502 retval = security_task_getscheduler(p);
5506 lp.sched_priority = p->rt_priority;
5507 read_unlock(&tasklist_lock);
5510 * This one might sleep, we cannot do it with a spinlock held ...
5512 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5517 read_unlock(&tasklist_lock);
5521 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5523 cpumask_var_t cpus_allowed, new_mask;
5524 struct task_struct *p;
5528 read_lock(&tasklist_lock);
5530 p = find_process_by_pid(pid);
5532 read_unlock(&tasklist_lock);
5538 * It is not safe to call set_cpus_allowed with the
5539 * tasklist_lock held. We will bump the task_struct's
5540 * usage count and then drop tasklist_lock.
5543 read_unlock(&tasklist_lock);
5545 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5549 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5551 goto out_free_cpus_allowed;
5554 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5555 !capable(CAP_SYS_NICE))
5558 retval = security_task_setscheduler(p, 0, NULL);
5562 cpuset_cpus_allowed(p, cpus_allowed);
5563 cpumask_and(new_mask, in_mask, cpus_allowed);
5565 retval = set_cpus_allowed_ptr(p, new_mask);
5568 cpuset_cpus_allowed(p, cpus_allowed);
5569 if (!cpumask_subset(new_mask, cpus_allowed)) {
5571 * We must have raced with a concurrent cpuset
5572 * update. Just reset the cpus_allowed to the
5573 * cpuset's cpus_allowed
5575 cpumask_copy(new_mask, cpus_allowed);
5580 free_cpumask_var(new_mask);
5581 out_free_cpus_allowed:
5582 free_cpumask_var(cpus_allowed);
5589 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5590 struct cpumask *new_mask)
5592 if (len < cpumask_size())
5593 cpumask_clear(new_mask);
5594 else if (len > cpumask_size())
5595 len = cpumask_size();
5597 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5601 * sys_sched_setaffinity - set the cpu affinity of a process
5602 * @pid: pid of the process
5603 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5604 * @user_mask_ptr: user-space pointer to the new cpu mask
5606 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5607 unsigned long __user *user_mask_ptr)
5609 cpumask_var_t new_mask;
5612 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5615 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5617 retval = sched_setaffinity(pid, new_mask);
5618 free_cpumask_var(new_mask);
5622 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5624 struct task_struct *p;
5628 read_lock(&tasklist_lock);
5631 p = find_process_by_pid(pid);
5635 retval = security_task_getscheduler(p);
5639 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5642 read_unlock(&tasklist_lock);
5649 * sys_sched_getaffinity - get the cpu affinity of a process
5650 * @pid: pid of the process
5651 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5652 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5654 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5655 unsigned long __user *user_mask_ptr)
5660 if (len < cpumask_size())
5663 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5666 ret = sched_getaffinity(pid, mask);
5668 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
5671 ret = cpumask_size();
5673 free_cpumask_var(mask);
5679 * sys_sched_yield - yield the current processor to other threads.
5681 * This function yields the current CPU to other tasks. If there are no
5682 * other threads running on this CPU then this function will return.
5684 asmlinkage long sys_sched_yield(void)
5686 struct rq *rq = this_rq_lock();
5688 schedstat_inc(rq, yld_count);
5689 current->sched_class->yield_task(rq);
5692 * Since we are going to call schedule() anyway, there's
5693 * no need to preempt or enable interrupts:
5695 __release(rq->lock);
5696 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5697 _raw_spin_unlock(&rq->lock);
5698 preempt_enable_no_resched();
5705 static void __cond_resched(void)
5707 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5708 __might_sleep(__FILE__, __LINE__);
5711 * The BKS might be reacquired before we have dropped
5712 * PREEMPT_ACTIVE, which could trigger a second
5713 * cond_resched() call.
5716 add_preempt_count(PREEMPT_ACTIVE);
5718 sub_preempt_count(PREEMPT_ACTIVE);
5719 } while (need_resched());
5722 int __sched _cond_resched(void)
5724 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5725 system_state == SYSTEM_RUNNING) {
5731 EXPORT_SYMBOL(_cond_resched);
5734 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5735 * call schedule, and on return reacquire the lock.
5737 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5738 * operations here to prevent schedule() from being called twice (once via
5739 * spin_unlock(), once by hand).
5741 int cond_resched_lock(spinlock_t *lock)
5743 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5746 if (spin_needbreak(lock) || resched) {
5748 if (resched && need_resched())
5757 EXPORT_SYMBOL(cond_resched_lock);
5759 int __sched cond_resched_softirq(void)
5761 BUG_ON(!in_softirq());
5763 if (need_resched() && system_state == SYSTEM_RUNNING) {
5771 EXPORT_SYMBOL(cond_resched_softirq);
5774 * yield - yield the current processor to other threads.
5776 * This is a shortcut for kernel-space yielding - it marks the
5777 * thread runnable and calls sys_sched_yield().
5779 void __sched yield(void)
5781 set_current_state(TASK_RUNNING);
5784 EXPORT_SYMBOL(yield);
5787 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5788 * that process accounting knows that this is a task in IO wait state.
5790 * But don't do that if it is a deliberate, throttling IO wait (this task
5791 * has set its backing_dev_info: the queue against which it should throttle)
5793 void __sched io_schedule(void)
5795 struct rq *rq = &__raw_get_cpu_var(runqueues);
5797 delayacct_blkio_start();
5798 atomic_inc(&rq->nr_iowait);
5800 atomic_dec(&rq->nr_iowait);
5801 delayacct_blkio_end();
5803 EXPORT_SYMBOL(io_schedule);
5805 long __sched io_schedule_timeout(long timeout)
5807 struct rq *rq = &__raw_get_cpu_var(runqueues);
5810 delayacct_blkio_start();
5811 atomic_inc(&rq->nr_iowait);
5812 ret = schedule_timeout(timeout);
5813 atomic_dec(&rq->nr_iowait);
5814 delayacct_blkio_end();
5819 * sys_sched_get_priority_max - return maximum RT priority.
5820 * @policy: scheduling class.
5822 * this syscall returns the maximum rt_priority that can be used
5823 * by a given scheduling class.
5825 asmlinkage long sys_sched_get_priority_max(int policy)
5832 ret = MAX_USER_RT_PRIO-1;
5844 * sys_sched_get_priority_min - return minimum RT priority.
5845 * @policy: scheduling class.
5847 * this syscall returns the minimum rt_priority that can be used
5848 * by a given scheduling class.
5850 asmlinkage long sys_sched_get_priority_min(int policy)
5868 * sys_sched_rr_get_interval - return the default timeslice of a process.
5869 * @pid: pid of the process.
5870 * @interval: userspace pointer to the timeslice value.
5872 * this syscall writes the default timeslice value of a given process
5873 * into the user-space timespec buffer. A value of '0' means infinity.
5876 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5878 struct task_struct *p;
5879 unsigned int time_slice;
5887 read_lock(&tasklist_lock);
5888 p = find_process_by_pid(pid);
5892 retval = security_task_getscheduler(p);
5897 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5898 * tasks that are on an otherwise idle runqueue:
5901 if (p->policy == SCHED_RR) {
5902 time_slice = DEF_TIMESLICE;
5903 } else if (p->policy != SCHED_FIFO) {
5904 struct sched_entity *se = &p->se;
5905 unsigned long flags;
5908 rq = task_rq_lock(p, &flags);
5909 if (rq->cfs.load.weight)
5910 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5911 task_rq_unlock(rq, &flags);
5913 read_unlock(&tasklist_lock);
5914 jiffies_to_timespec(time_slice, &t);
5915 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5919 read_unlock(&tasklist_lock);
5923 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5925 void sched_show_task(struct task_struct *p)
5927 unsigned long free = 0;
5930 state = p->state ? __ffs(p->state) + 1 : 0;
5931 printk(KERN_INFO "%-13.13s %c", p->comm,
5932 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5933 #if BITS_PER_LONG == 32
5934 if (state == TASK_RUNNING)
5935 printk(KERN_CONT " running ");
5937 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5939 if (state == TASK_RUNNING)
5940 printk(KERN_CONT " running task ");
5942 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5944 #ifdef CONFIG_DEBUG_STACK_USAGE
5946 unsigned long *n = end_of_stack(p);
5949 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5952 printk(KERN_CONT "%5lu %5d %6d\n", free,
5953 task_pid_nr(p), task_pid_nr(p->real_parent));
5955 show_stack(p, NULL);
5958 void show_state_filter(unsigned long state_filter)
5960 struct task_struct *g, *p;
5962 #if BITS_PER_LONG == 32
5964 " task PC stack pid father\n");
5967 " task PC stack pid father\n");
5969 read_lock(&tasklist_lock);
5970 do_each_thread(g, p) {
5972 * reset the NMI-timeout, listing all files on a slow
5973 * console might take alot of time:
5975 touch_nmi_watchdog();
5976 if (!state_filter || (p->state & state_filter))
5978 } while_each_thread(g, p);
5980 touch_all_softlockup_watchdogs();
5982 #ifdef CONFIG_SCHED_DEBUG
5983 sysrq_sched_debug_show();
5985 read_unlock(&tasklist_lock);
5987 * Only show locks if all tasks are dumped:
5989 if (state_filter == -1)
5990 debug_show_all_locks();
5993 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5995 idle->sched_class = &idle_sched_class;
5999 * init_idle - set up an idle thread for a given CPU
6000 * @idle: task in question
6001 * @cpu: cpu the idle task belongs to
6003 * NOTE: this function does not set the idle thread's NEED_RESCHED
6004 * flag, to make booting more robust.
6006 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6008 struct rq *rq = cpu_rq(cpu);
6009 unsigned long flags;
6011 spin_lock_irqsave(&rq->lock, flags);
6014 idle->se.exec_start = sched_clock();
6016 idle->prio = idle->normal_prio = MAX_PRIO;
6017 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6018 __set_task_cpu(idle, cpu);
6020 rq->curr = rq->idle = idle;
6021 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6024 spin_unlock_irqrestore(&rq->lock, flags);
6026 /* Set the preempt count _outside_ the spinlocks! */
6027 #if defined(CONFIG_PREEMPT)
6028 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6030 task_thread_info(idle)->preempt_count = 0;
6033 * The idle tasks have their own, simple scheduling class:
6035 idle->sched_class = &idle_sched_class;
6036 ftrace_graph_init_task(idle);
6040 * In a system that switches off the HZ timer nohz_cpu_mask
6041 * indicates which cpus entered this state. This is used
6042 * in the rcu update to wait only for active cpus. For system
6043 * which do not switch off the HZ timer nohz_cpu_mask should
6044 * always be CPU_BITS_NONE.
6046 cpumask_var_t nohz_cpu_mask;
6049 * Increase the granularity value when there are more CPUs,
6050 * because with more CPUs the 'effective latency' as visible
6051 * to users decreases. But the relationship is not linear,
6052 * so pick a second-best guess by going with the log2 of the
6055 * This idea comes from the SD scheduler of Con Kolivas:
6057 static inline void sched_init_granularity(void)
6059 unsigned int factor = 1 + ilog2(num_online_cpus());
6060 const unsigned long limit = 200000000;
6062 sysctl_sched_min_granularity *= factor;
6063 if (sysctl_sched_min_granularity > limit)
6064 sysctl_sched_min_granularity = limit;
6066 sysctl_sched_latency *= factor;
6067 if (sysctl_sched_latency > limit)
6068 sysctl_sched_latency = limit;
6070 sysctl_sched_wakeup_granularity *= factor;
6072 sysctl_sched_shares_ratelimit *= factor;
6077 * This is how migration works:
6079 * 1) we queue a struct migration_req structure in the source CPU's
6080 * runqueue and wake up that CPU's migration thread.
6081 * 2) we down() the locked semaphore => thread blocks.
6082 * 3) migration thread wakes up (implicitly it forces the migrated
6083 * thread off the CPU)
6084 * 4) it gets the migration request and checks whether the migrated
6085 * task is still in the wrong runqueue.
6086 * 5) if it's in the wrong runqueue then the migration thread removes
6087 * it and puts it into the right queue.
6088 * 6) migration thread up()s the semaphore.
6089 * 7) we wake up and the migration is done.
6093 * Change a given task's CPU affinity. Migrate the thread to a
6094 * proper CPU and schedule it away if the CPU it's executing on
6095 * is removed from the allowed bitmask.
6097 * NOTE: the caller must have a valid reference to the task, the
6098 * task must not exit() & deallocate itself prematurely. The
6099 * call is not atomic; no spinlocks may be held.
6101 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6103 struct migration_req req;
6104 unsigned long flags;
6108 rq = task_rq_lock(p, &flags);
6109 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6114 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6115 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6120 if (p->sched_class->set_cpus_allowed)
6121 p->sched_class->set_cpus_allowed(p, new_mask);
6123 cpumask_copy(&p->cpus_allowed, new_mask);
6124 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6127 /* Can the task run on the task's current CPU? If so, we're done */
6128 if (cpumask_test_cpu(task_cpu(p), new_mask))
6131 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6132 /* Need help from migration thread: drop lock and wait. */
6133 task_rq_unlock(rq, &flags);
6134 wake_up_process(rq->migration_thread);
6135 wait_for_completion(&req.done);
6136 tlb_migrate_finish(p->mm);
6140 task_rq_unlock(rq, &flags);
6144 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6147 * Move (not current) task off this cpu, onto dest cpu. We're doing
6148 * this because either it can't run here any more (set_cpus_allowed()
6149 * away from this CPU, or CPU going down), or because we're
6150 * attempting to rebalance this task on exec (sched_exec).
6152 * So we race with normal scheduler movements, but that's OK, as long
6153 * as the task is no longer on this CPU.
6155 * Returns non-zero if task was successfully migrated.
6157 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6159 struct rq *rq_dest, *rq_src;
6162 if (unlikely(!cpu_active(dest_cpu)))
6165 rq_src = cpu_rq(src_cpu);
6166 rq_dest = cpu_rq(dest_cpu);
6168 double_rq_lock(rq_src, rq_dest);
6169 /* Already moved. */
6170 if (task_cpu(p) != src_cpu)
6172 /* Affinity changed (again). */
6173 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6176 on_rq = p->se.on_rq;
6178 deactivate_task(rq_src, p, 0);
6180 set_task_cpu(p, dest_cpu);
6182 activate_task(rq_dest, p, 0);
6183 check_preempt_curr(rq_dest, p, 0);
6188 double_rq_unlock(rq_src, rq_dest);
6193 * migration_thread - this is a highprio system thread that performs
6194 * thread migration by bumping thread off CPU then 'pushing' onto
6197 static int migration_thread(void *data)
6199 int cpu = (long)data;
6203 BUG_ON(rq->migration_thread != current);
6205 set_current_state(TASK_INTERRUPTIBLE);
6206 while (!kthread_should_stop()) {
6207 struct migration_req *req;
6208 struct list_head *head;
6210 spin_lock_irq(&rq->lock);
6212 if (cpu_is_offline(cpu)) {
6213 spin_unlock_irq(&rq->lock);
6217 if (rq->active_balance) {
6218 active_load_balance(rq, cpu);
6219 rq->active_balance = 0;
6222 head = &rq->migration_queue;
6224 if (list_empty(head)) {
6225 spin_unlock_irq(&rq->lock);
6227 set_current_state(TASK_INTERRUPTIBLE);
6230 req = list_entry(head->next, struct migration_req, list);
6231 list_del_init(head->next);
6233 spin_unlock(&rq->lock);
6234 __migrate_task(req->task, cpu, req->dest_cpu);
6237 complete(&req->done);
6239 __set_current_state(TASK_RUNNING);
6243 /* Wait for kthread_stop */
6244 set_current_state(TASK_INTERRUPTIBLE);
6245 while (!kthread_should_stop()) {
6247 set_current_state(TASK_INTERRUPTIBLE);
6249 __set_current_state(TASK_RUNNING);
6253 #ifdef CONFIG_HOTPLUG_CPU
6255 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6259 local_irq_disable();
6260 ret = __migrate_task(p, src_cpu, dest_cpu);
6266 * Figure out where task on dead CPU should go, use force if necessary.
6268 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6271 /* FIXME: Use cpumask_of_node here. */
6272 cpumask_t _nodemask = node_to_cpumask(cpu_to_node(dead_cpu));
6273 const struct cpumask *nodemask = &_nodemask;
6276 /* Look for allowed, online CPU in same node. */
6277 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
6278 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6281 /* Any allowed, online CPU? */
6282 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
6283 if (dest_cpu < nr_cpu_ids)
6286 /* No more Mr. Nice Guy. */
6287 if (dest_cpu >= nr_cpu_ids) {
6288 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
6289 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
6292 * Don't tell them about moving exiting tasks or
6293 * kernel threads (both mm NULL), since they never
6296 if (p->mm && printk_ratelimit()) {
6297 printk(KERN_INFO "process %d (%s) no "
6298 "longer affine to cpu%d\n",
6299 task_pid_nr(p), p->comm, dead_cpu);
6304 /* It can have affinity changed while we were choosing. */
6305 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
6310 * While a dead CPU has no uninterruptible tasks queued at this point,
6311 * it might still have a nonzero ->nr_uninterruptible counter, because
6312 * for performance reasons the counter is not stricly tracking tasks to
6313 * their home CPUs. So we just add the counter to another CPU's counter,
6314 * to keep the global sum constant after CPU-down:
6316 static void migrate_nr_uninterruptible(struct rq *rq_src)
6318 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
6319 unsigned long flags;
6321 local_irq_save(flags);
6322 double_rq_lock(rq_src, rq_dest);
6323 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6324 rq_src->nr_uninterruptible = 0;
6325 double_rq_unlock(rq_src, rq_dest);
6326 local_irq_restore(flags);
6329 /* Run through task list and migrate tasks from the dead cpu. */
6330 static void migrate_live_tasks(int src_cpu)
6332 struct task_struct *p, *t;
6334 read_lock(&tasklist_lock);
6336 do_each_thread(t, p) {
6340 if (task_cpu(p) == src_cpu)
6341 move_task_off_dead_cpu(src_cpu, p);
6342 } while_each_thread(t, p);
6344 read_unlock(&tasklist_lock);
6348 * Schedules idle task to be the next runnable task on current CPU.
6349 * It does so by boosting its priority to highest possible.
6350 * Used by CPU offline code.
6352 void sched_idle_next(void)
6354 int this_cpu = smp_processor_id();
6355 struct rq *rq = cpu_rq(this_cpu);
6356 struct task_struct *p = rq->idle;
6357 unsigned long flags;
6359 /* cpu has to be offline */
6360 BUG_ON(cpu_online(this_cpu));
6363 * Strictly not necessary since rest of the CPUs are stopped by now
6364 * and interrupts disabled on the current cpu.
6366 spin_lock_irqsave(&rq->lock, flags);
6368 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6370 update_rq_clock(rq);
6371 activate_task(rq, p, 0);
6373 spin_unlock_irqrestore(&rq->lock, flags);
6377 * Ensures that the idle task is using init_mm right before its cpu goes
6380 void idle_task_exit(void)
6382 struct mm_struct *mm = current->active_mm;
6384 BUG_ON(cpu_online(smp_processor_id()));
6387 switch_mm(mm, &init_mm, current);
6391 /* called under rq->lock with disabled interrupts */
6392 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6394 struct rq *rq = cpu_rq(dead_cpu);
6396 /* Must be exiting, otherwise would be on tasklist. */
6397 BUG_ON(!p->exit_state);
6399 /* Cannot have done final schedule yet: would have vanished. */
6400 BUG_ON(p->state == TASK_DEAD);
6405 * Drop lock around migration; if someone else moves it,
6406 * that's OK. No task can be added to this CPU, so iteration is
6409 spin_unlock_irq(&rq->lock);
6410 move_task_off_dead_cpu(dead_cpu, p);
6411 spin_lock_irq(&rq->lock);
6416 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6417 static void migrate_dead_tasks(unsigned int dead_cpu)
6419 struct rq *rq = cpu_rq(dead_cpu);
6420 struct task_struct *next;
6423 if (!rq->nr_running)
6425 update_rq_clock(rq);
6426 next = pick_next_task(rq, rq->curr);
6429 next->sched_class->put_prev_task(rq, next);
6430 migrate_dead(dead_cpu, next);
6434 #endif /* CONFIG_HOTPLUG_CPU */
6436 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6438 static struct ctl_table sd_ctl_dir[] = {
6440 .procname = "sched_domain",
6446 static struct ctl_table sd_ctl_root[] = {
6448 .ctl_name = CTL_KERN,
6449 .procname = "kernel",
6451 .child = sd_ctl_dir,
6456 static struct ctl_table *sd_alloc_ctl_entry(int n)
6458 struct ctl_table *entry =
6459 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6464 static void sd_free_ctl_entry(struct ctl_table **tablep)
6466 struct ctl_table *entry;
6469 * In the intermediate directories, both the child directory and
6470 * procname are dynamically allocated and could fail but the mode
6471 * will always be set. In the lowest directory the names are
6472 * static strings and all have proc handlers.
6474 for (entry = *tablep; entry->mode; entry++) {
6476 sd_free_ctl_entry(&entry->child);
6477 if (entry->proc_handler == NULL)
6478 kfree(entry->procname);
6486 set_table_entry(struct ctl_table *entry,
6487 const char *procname, void *data, int maxlen,
6488 mode_t mode, proc_handler *proc_handler)
6490 entry->procname = procname;
6492 entry->maxlen = maxlen;
6494 entry->proc_handler = proc_handler;
6497 static struct ctl_table *
6498 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6500 struct ctl_table *table = sd_alloc_ctl_entry(13);
6505 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6506 sizeof(long), 0644, proc_doulongvec_minmax);
6507 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6508 sizeof(long), 0644, proc_doulongvec_minmax);
6509 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6510 sizeof(int), 0644, proc_dointvec_minmax);
6511 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6512 sizeof(int), 0644, proc_dointvec_minmax);
6513 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6514 sizeof(int), 0644, proc_dointvec_minmax);
6515 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6516 sizeof(int), 0644, proc_dointvec_minmax);
6517 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6518 sizeof(int), 0644, proc_dointvec_minmax);
6519 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6520 sizeof(int), 0644, proc_dointvec_minmax);
6521 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6522 sizeof(int), 0644, proc_dointvec_minmax);
6523 set_table_entry(&table[9], "cache_nice_tries",
6524 &sd->cache_nice_tries,
6525 sizeof(int), 0644, proc_dointvec_minmax);
6526 set_table_entry(&table[10], "flags", &sd->flags,
6527 sizeof(int), 0644, proc_dointvec_minmax);
6528 set_table_entry(&table[11], "name", sd->name,
6529 CORENAME_MAX_SIZE, 0444, proc_dostring);
6530 /* &table[12] is terminator */
6535 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6537 struct ctl_table *entry, *table;
6538 struct sched_domain *sd;
6539 int domain_num = 0, i;
6542 for_each_domain(cpu, sd)
6544 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6549 for_each_domain(cpu, sd) {
6550 snprintf(buf, 32, "domain%d", i);
6551 entry->procname = kstrdup(buf, GFP_KERNEL);
6553 entry->child = sd_alloc_ctl_domain_table(sd);
6560 static struct ctl_table_header *sd_sysctl_header;
6561 static void register_sched_domain_sysctl(void)
6563 int i, cpu_num = num_online_cpus();
6564 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6567 WARN_ON(sd_ctl_dir[0].child);
6568 sd_ctl_dir[0].child = entry;
6573 for_each_online_cpu(i) {
6574 snprintf(buf, 32, "cpu%d", i);
6575 entry->procname = kstrdup(buf, GFP_KERNEL);
6577 entry->child = sd_alloc_ctl_cpu_table(i);
6581 WARN_ON(sd_sysctl_header);
6582 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6585 /* may be called multiple times per register */
6586 static void unregister_sched_domain_sysctl(void)
6588 if (sd_sysctl_header)
6589 unregister_sysctl_table(sd_sysctl_header);
6590 sd_sysctl_header = NULL;
6591 if (sd_ctl_dir[0].child)
6592 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6595 static void register_sched_domain_sysctl(void)
6598 static void unregister_sched_domain_sysctl(void)
6603 static void set_rq_online(struct rq *rq)
6606 const struct sched_class *class;
6608 cpumask_set_cpu(rq->cpu, rq->rd->online);
6611 for_each_class(class) {
6612 if (class->rq_online)
6613 class->rq_online(rq);
6618 static void set_rq_offline(struct rq *rq)
6621 const struct sched_class *class;
6623 for_each_class(class) {
6624 if (class->rq_offline)
6625 class->rq_offline(rq);
6628 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6634 * migration_call - callback that gets triggered when a CPU is added.
6635 * Here we can start up the necessary migration thread for the new CPU.
6637 static int __cpuinit
6638 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6640 struct task_struct *p;
6641 int cpu = (long)hcpu;
6642 unsigned long flags;
6647 case CPU_UP_PREPARE:
6648 case CPU_UP_PREPARE_FROZEN:
6649 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6652 kthread_bind(p, cpu);
6653 /* Must be high prio: stop_machine expects to yield to it. */
6654 rq = task_rq_lock(p, &flags);
6655 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6656 task_rq_unlock(rq, &flags);
6657 cpu_rq(cpu)->migration_thread = p;
6661 case CPU_ONLINE_FROZEN:
6662 /* Strictly unnecessary, as first user will wake it. */
6663 wake_up_process(cpu_rq(cpu)->migration_thread);
6665 /* Update our root-domain */
6667 spin_lock_irqsave(&rq->lock, flags);
6669 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6673 spin_unlock_irqrestore(&rq->lock, flags);
6676 #ifdef CONFIG_HOTPLUG_CPU
6677 case CPU_UP_CANCELED:
6678 case CPU_UP_CANCELED_FROZEN:
6679 if (!cpu_rq(cpu)->migration_thread)
6681 /* Unbind it from offline cpu so it can run. Fall thru. */
6682 kthread_bind(cpu_rq(cpu)->migration_thread,
6683 cpumask_any(cpu_online_mask));
6684 kthread_stop(cpu_rq(cpu)->migration_thread);
6685 cpu_rq(cpu)->migration_thread = NULL;
6689 case CPU_DEAD_FROZEN:
6690 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6691 migrate_live_tasks(cpu);
6693 kthread_stop(rq->migration_thread);
6694 rq->migration_thread = NULL;
6695 /* Idle task back to normal (off runqueue, low prio) */
6696 spin_lock_irq(&rq->lock);
6697 update_rq_clock(rq);
6698 deactivate_task(rq, rq->idle, 0);
6699 rq->idle->static_prio = MAX_PRIO;
6700 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6701 rq->idle->sched_class = &idle_sched_class;
6702 migrate_dead_tasks(cpu);
6703 spin_unlock_irq(&rq->lock);
6705 migrate_nr_uninterruptible(rq);
6706 BUG_ON(rq->nr_running != 0);
6709 * No need to migrate the tasks: it was best-effort if
6710 * they didn't take sched_hotcpu_mutex. Just wake up
6713 spin_lock_irq(&rq->lock);
6714 while (!list_empty(&rq->migration_queue)) {
6715 struct migration_req *req;
6717 req = list_entry(rq->migration_queue.next,
6718 struct migration_req, list);
6719 list_del_init(&req->list);
6720 spin_unlock_irq(&rq->lock);
6721 complete(&req->done);
6722 spin_lock_irq(&rq->lock);
6724 spin_unlock_irq(&rq->lock);
6728 case CPU_DYING_FROZEN:
6729 /* Update our root-domain */
6731 spin_lock_irqsave(&rq->lock, flags);
6733 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6736 spin_unlock_irqrestore(&rq->lock, flags);
6743 /* Register at highest priority so that task migration (migrate_all_tasks)
6744 * happens before everything else.
6746 static struct notifier_block __cpuinitdata migration_notifier = {
6747 .notifier_call = migration_call,
6751 static int __init migration_init(void)
6753 void *cpu = (void *)(long)smp_processor_id();
6756 /* Start one for the boot CPU: */
6757 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6758 BUG_ON(err == NOTIFY_BAD);
6759 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6760 register_cpu_notifier(&migration_notifier);
6764 early_initcall(migration_init);
6769 #ifdef CONFIG_SCHED_DEBUG
6771 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6772 struct cpumask *groupmask)
6774 struct sched_group *group = sd->groups;
6777 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6778 cpumask_clear(groupmask);
6780 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6782 if (!(sd->flags & SD_LOAD_BALANCE)) {
6783 printk("does not load-balance\n");
6785 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6790 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6792 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6793 printk(KERN_ERR "ERROR: domain->span does not contain "
6796 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6797 printk(KERN_ERR "ERROR: domain->groups does not contain"
6801 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6805 printk(KERN_ERR "ERROR: group is NULL\n");
6809 if (!group->__cpu_power) {
6810 printk(KERN_CONT "\n");
6811 printk(KERN_ERR "ERROR: domain->cpu_power not "
6816 if (!cpumask_weight(sched_group_cpus(group))) {
6817 printk(KERN_CONT "\n");
6818 printk(KERN_ERR "ERROR: empty group\n");
6822 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6823 printk(KERN_CONT "\n");
6824 printk(KERN_ERR "ERROR: repeated CPUs\n");
6828 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6830 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6831 printk(KERN_CONT " %s", str);
6833 group = group->next;
6834 } while (group != sd->groups);
6835 printk(KERN_CONT "\n");
6837 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6838 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6841 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6842 printk(KERN_ERR "ERROR: parent span is not a superset "
6843 "of domain->span\n");
6847 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6849 cpumask_var_t groupmask;
6853 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6857 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6859 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6860 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6865 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6872 free_cpumask_var(groupmask);
6874 #else /* !CONFIG_SCHED_DEBUG */
6875 # define sched_domain_debug(sd, cpu) do { } while (0)
6876 #endif /* CONFIG_SCHED_DEBUG */
6878 static int sd_degenerate(struct sched_domain *sd)
6880 if (cpumask_weight(sched_domain_span(sd)) == 1)
6883 /* Following flags need at least 2 groups */
6884 if (sd->flags & (SD_LOAD_BALANCE |
6885 SD_BALANCE_NEWIDLE |
6889 SD_SHARE_PKG_RESOURCES)) {
6890 if (sd->groups != sd->groups->next)
6894 /* Following flags don't use groups */
6895 if (sd->flags & (SD_WAKE_IDLE |
6904 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6906 unsigned long cflags = sd->flags, pflags = parent->flags;
6908 if (sd_degenerate(parent))
6911 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6914 /* Does parent contain flags not in child? */
6915 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6916 if (cflags & SD_WAKE_AFFINE)
6917 pflags &= ~SD_WAKE_BALANCE;
6918 /* Flags needing groups don't count if only 1 group in parent */
6919 if (parent->groups == parent->groups->next) {
6920 pflags &= ~(SD_LOAD_BALANCE |
6921 SD_BALANCE_NEWIDLE |
6925 SD_SHARE_PKG_RESOURCES);
6926 if (nr_node_ids == 1)
6927 pflags &= ~SD_SERIALIZE;
6929 if (~cflags & pflags)
6935 static void free_rootdomain(struct root_domain *rd)
6937 cpupri_cleanup(&rd->cpupri);
6939 free_cpumask_var(rd->rto_mask);
6940 free_cpumask_var(rd->online);
6941 free_cpumask_var(rd->span);
6945 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6947 unsigned long flags;
6949 spin_lock_irqsave(&rq->lock, flags);
6952 struct root_domain *old_rd = rq->rd;
6954 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6957 cpumask_clear_cpu(rq->cpu, old_rd->span);
6959 if (atomic_dec_and_test(&old_rd->refcount))
6960 free_rootdomain(old_rd);
6963 atomic_inc(&rd->refcount);
6966 cpumask_set_cpu(rq->cpu, rd->span);
6967 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
6970 spin_unlock_irqrestore(&rq->lock, flags);
6973 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6975 memset(rd, 0, sizeof(*rd));
6978 alloc_bootmem_cpumask_var(&def_root_domain.span);
6979 alloc_bootmem_cpumask_var(&def_root_domain.online);
6980 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
6981 cpupri_init(&rd->cpupri, true);
6985 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6987 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6989 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6992 if (cpupri_init(&rd->cpupri, false) != 0)
6997 free_cpumask_var(rd->rto_mask);
6999 free_cpumask_var(rd->online);
7001 free_cpumask_var(rd->span);
7007 static void init_defrootdomain(void)
7009 init_rootdomain(&def_root_domain, true);
7011 atomic_set(&def_root_domain.refcount, 1);
7014 static struct root_domain *alloc_rootdomain(void)
7016 struct root_domain *rd;
7018 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7022 if (init_rootdomain(rd, false) != 0) {
7031 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7032 * hold the hotplug lock.
7035 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7037 struct rq *rq = cpu_rq(cpu);
7038 struct sched_domain *tmp;
7040 /* Remove the sched domains which do not contribute to scheduling. */
7041 for (tmp = sd; tmp; ) {
7042 struct sched_domain *parent = tmp->parent;
7046 if (sd_parent_degenerate(tmp, parent)) {
7047 tmp->parent = parent->parent;
7049 parent->parent->child = tmp;
7054 if (sd && sd_degenerate(sd)) {
7060 sched_domain_debug(sd, cpu);
7062 rq_attach_root(rq, rd);
7063 rcu_assign_pointer(rq->sd, sd);
7066 /* cpus with isolated domains */
7067 static cpumask_var_t cpu_isolated_map;
7069 /* Setup the mask of cpus configured for isolated domains */
7070 static int __init isolated_cpu_setup(char *str)
7072 cpulist_parse(str, cpu_isolated_map);
7076 __setup("isolcpus=", isolated_cpu_setup);
7079 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7080 * to a function which identifies what group(along with sched group) a CPU
7081 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7082 * (due to the fact that we keep track of groups covered with a struct cpumask).
7084 * init_sched_build_groups will build a circular linked list of the groups
7085 * covered by the given span, and will set each group's ->cpumask correctly,
7086 * and ->cpu_power to 0.
7089 init_sched_build_groups(const struct cpumask *span,
7090 const struct cpumask *cpu_map,
7091 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7092 struct sched_group **sg,
7093 struct cpumask *tmpmask),
7094 struct cpumask *covered, struct cpumask *tmpmask)
7096 struct sched_group *first = NULL, *last = NULL;
7099 cpumask_clear(covered);
7101 for_each_cpu(i, span) {
7102 struct sched_group *sg;
7103 int group = group_fn(i, cpu_map, &sg, tmpmask);
7106 if (cpumask_test_cpu(i, covered))
7109 cpumask_clear(sched_group_cpus(sg));
7110 sg->__cpu_power = 0;
7112 for_each_cpu(j, span) {
7113 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7116 cpumask_set_cpu(j, covered);
7117 cpumask_set_cpu(j, sched_group_cpus(sg));
7128 #define SD_NODES_PER_DOMAIN 16
7133 * find_next_best_node - find the next node to include in a sched_domain
7134 * @node: node whose sched_domain we're building
7135 * @used_nodes: nodes already in the sched_domain
7137 * Find the next node to include in a given scheduling domain. Simply
7138 * finds the closest node not already in the @used_nodes map.
7140 * Should use nodemask_t.
7142 static int find_next_best_node(int node, nodemask_t *used_nodes)
7144 int i, n, val, min_val, best_node = 0;
7148 for (i = 0; i < nr_node_ids; i++) {
7149 /* Start at @node */
7150 n = (node + i) % nr_node_ids;
7152 if (!nr_cpus_node(n))
7155 /* Skip already used nodes */
7156 if (node_isset(n, *used_nodes))
7159 /* Simple min distance search */
7160 val = node_distance(node, n);
7162 if (val < min_val) {
7168 node_set(best_node, *used_nodes);
7173 * sched_domain_node_span - get a cpumask for a node's sched_domain
7174 * @node: node whose cpumask we're constructing
7175 * @span: resulting cpumask
7177 * Given a node, construct a good cpumask for its sched_domain to span. It
7178 * should be one that prevents unnecessary balancing, but also spreads tasks
7181 static void sched_domain_node_span(int node, struct cpumask *span)
7183 nodemask_t used_nodes;
7184 /* FIXME: use cpumask_of_node() */
7185 node_to_cpumask_ptr(nodemask, node);
7189 nodes_clear(used_nodes);
7191 cpus_or(*span, *span, *nodemask);
7192 node_set(node, used_nodes);
7194 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7195 int next_node = find_next_best_node(node, &used_nodes);
7197 node_to_cpumask_ptr_next(nodemask, next_node);
7198 cpus_or(*span, *span, *nodemask);
7201 #endif /* CONFIG_NUMA */
7203 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7206 * The cpus mask in sched_group and sched_domain hangs off the end.
7207 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7208 * for nr_cpu_ids < CONFIG_NR_CPUS.
7210 struct static_sched_group {
7211 struct sched_group sg;
7212 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7215 struct static_sched_domain {
7216 struct sched_domain sd;
7217 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7221 * SMT sched-domains:
7223 #ifdef CONFIG_SCHED_SMT
7224 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7225 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7228 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
7229 struct sched_group **sg, struct cpumask *unused)
7232 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7235 #endif /* CONFIG_SCHED_SMT */
7238 * multi-core sched-domains:
7240 #ifdef CONFIG_SCHED_MC
7241 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7242 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7243 #endif /* CONFIG_SCHED_MC */
7245 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7247 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7248 struct sched_group **sg, struct cpumask *mask)
7252 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7253 group = cpumask_first(mask);
7255 *sg = &per_cpu(sched_group_core, group).sg;
7258 #elif defined(CONFIG_SCHED_MC)
7260 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7261 struct sched_group **sg, struct cpumask *unused)
7264 *sg = &per_cpu(sched_group_core, cpu).sg;
7269 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
7270 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
7273 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
7274 struct sched_group **sg, struct cpumask *mask)
7277 #ifdef CONFIG_SCHED_MC
7278 /* FIXME: Use cpu_coregroup_mask. */
7279 *mask = cpu_coregroup_map(cpu);
7280 cpus_and(*mask, *mask, *cpu_map);
7281 group = cpumask_first(mask);
7282 #elif defined(CONFIG_SCHED_SMT)
7283 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7284 group = cpumask_first(mask);
7289 *sg = &per_cpu(sched_group_phys, group).sg;
7295 * The init_sched_build_groups can't handle what we want to do with node
7296 * groups, so roll our own. Now each node has its own list of groups which
7297 * gets dynamically allocated.
7299 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7300 static struct sched_group ***sched_group_nodes_bycpu;
7302 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7303 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7305 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
7306 struct sched_group **sg,
7307 struct cpumask *nodemask)
7310 /* FIXME: use cpumask_of_node */
7311 node_to_cpumask_ptr(pnodemask, cpu_to_node(cpu));
7313 cpumask_and(nodemask, pnodemask, cpu_map);
7314 group = cpumask_first(nodemask);
7317 *sg = &per_cpu(sched_group_allnodes, group).sg;
7321 static void init_numa_sched_groups_power(struct sched_group *group_head)
7323 struct sched_group *sg = group_head;
7329 for_each_cpu(j, sched_group_cpus(sg)) {
7330 struct sched_domain *sd;
7332 sd = &per_cpu(phys_domains, j).sd;
7333 if (j != cpumask_first(sched_group_cpus(sd->groups))) {
7335 * Only add "power" once for each
7341 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7344 } while (sg != group_head);
7346 #endif /* CONFIG_NUMA */
7349 /* Free memory allocated for various sched_group structures */
7350 static void free_sched_groups(const struct cpumask *cpu_map,
7351 struct cpumask *nodemask)
7355 for_each_cpu(cpu, cpu_map) {
7356 struct sched_group **sched_group_nodes
7357 = sched_group_nodes_bycpu[cpu];
7359 if (!sched_group_nodes)
7362 for (i = 0; i < nr_node_ids; i++) {
7363 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7364 /* FIXME: Use cpumask_of_node */
7365 node_to_cpumask_ptr(pnodemask, i);
7367 cpus_and(*nodemask, *pnodemask, *cpu_map);
7368 if (cpumask_empty(nodemask))
7378 if (oldsg != sched_group_nodes[i])
7381 kfree(sched_group_nodes);
7382 sched_group_nodes_bycpu[cpu] = NULL;
7385 #else /* !CONFIG_NUMA */
7386 static void free_sched_groups(const struct cpumask *cpu_map,
7387 struct cpumask *nodemask)
7390 #endif /* CONFIG_NUMA */
7393 * Initialize sched groups cpu_power.
7395 * cpu_power indicates the capacity of sched group, which is used while
7396 * distributing the load between different sched groups in a sched domain.
7397 * Typically cpu_power for all the groups in a sched domain will be same unless
7398 * there are asymmetries in the topology. If there are asymmetries, group
7399 * having more cpu_power will pickup more load compared to the group having
7402 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7403 * the maximum number of tasks a group can handle in the presence of other idle
7404 * or lightly loaded groups in the same sched domain.
7406 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7408 struct sched_domain *child;
7409 struct sched_group *group;
7411 WARN_ON(!sd || !sd->groups);
7413 if (cpu != cpumask_first(sched_group_cpus(sd->groups)))
7418 sd->groups->__cpu_power = 0;
7421 * For perf policy, if the groups in child domain share resources
7422 * (for example cores sharing some portions of the cache hierarchy
7423 * or SMT), then set this domain groups cpu_power such that each group
7424 * can handle only one task, when there are other idle groups in the
7425 * same sched domain.
7427 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7429 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7430 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7435 * add cpu_power of each child group to this groups cpu_power
7437 group = child->groups;
7439 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7440 group = group->next;
7441 } while (group != child->groups);
7445 * Initializers for schedule domains
7446 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7449 #ifdef CONFIG_SCHED_DEBUG
7450 # define SD_INIT_NAME(sd, type) sd->name = #type
7452 # define SD_INIT_NAME(sd, type) do { } while (0)
7455 #define SD_INIT(sd, type) sd_init_##type(sd)
7457 #define SD_INIT_FUNC(type) \
7458 static noinline void sd_init_##type(struct sched_domain *sd) \
7460 memset(sd, 0, sizeof(*sd)); \
7461 *sd = SD_##type##_INIT; \
7462 sd->level = SD_LV_##type; \
7463 SD_INIT_NAME(sd, type); \
7468 SD_INIT_FUNC(ALLNODES)
7471 #ifdef CONFIG_SCHED_SMT
7472 SD_INIT_FUNC(SIBLING)
7474 #ifdef CONFIG_SCHED_MC
7478 static int default_relax_domain_level = -1;
7480 static int __init setup_relax_domain_level(char *str)
7484 val = simple_strtoul(str, NULL, 0);
7485 if (val < SD_LV_MAX)
7486 default_relax_domain_level = val;
7490 __setup("relax_domain_level=", setup_relax_domain_level);
7492 static void set_domain_attribute(struct sched_domain *sd,
7493 struct sched_domain_attr *attr)
7497 if (!attr || attr->relax_domain_level < 0) {
7498 if (default_relax_domain_level < 0)
7501 request = default_relax_domain_level;
7503 request = attr->relax_domain_level;
7504 if (request < sd->level) {
7505 /* turn off idle balance on this domain */
7506 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7508 /* turn on idle balance on this domain */
7509 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7514 * Build sched domains for a given set of cpus and attach the sched domains
7515 * to the individual cpus
7517 static int __build_sched_domains(const struct cpumask *cpu_map,
7518 struct sched_domain_attr *attr)
7520 int i, err = -ENOMEM;
7521 struct root_domain *rd;
7522 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
7525 cpumask_var_t domainspan, covered, notcovered;
7526 struct sched_group **sched_group_nodes = NULL;
7527 int sd_allnodes = 0;
7529 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
7531 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
7532 goto free_domainspan;
7533 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
7537 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
7538 goto free_notcovered;
7539 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
7541 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
7542 goto free_this_sibling_map;
7543 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
7544 goto free_this_core_map;
7545 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
7546 goto free_send_covered;
7550 * Allocate the per-node list of sched groups
7552 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7554 if (!sched_group_nodes) {
7555 printk(KERN_WARNING "Can not alloc sched group node list\n");
7560 rd = alloc_rootdomain();
7562 printk(KERN_WARNING "Cannot alloc root domain\n");
7563 goto free_sched_groups;
7567 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
7571 * Set up domains for cpus specified by the cpu_map.
7573 for_each_cpu(i, cpu_map) {
7574 struct sched_domain *sd = NULL, *p;
7576 /* FIXME: use cpumask_of_node */
7577 *nodemask = node_to_cpumask(cpu_to_node(i));
7578 cpus_and(*nodemask, *nodemask, *cpu_map);
7581 if (cpumask_weight(cpu_map) >
7582 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
7583 sd = &per_cpu(allnodes_domains, i);
7584 SD_INIT(sd, ALLNODES);
7585 set_domain_attribute(sd, attr);
7586 cpumask_copy(sched_domain_span(sd), cpu_map);
7587 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7593 sd = &per_cpu(node_domains, i);
7595 set_domain_attribute(sd, attr);
7596 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7600 cpumask_and(sched_domain_span(sd),
7601 sched_domain_span(sd), cpu_map);
7605 sd = &per_cpu(phys_domains, i).sd;
7607 set_domain_attribute(sd, attr);
7608 cpumask_copy(sched_domain_span(sd), nodemask);
7612 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7614 #ifdef CONFIG_SCHED_MC
7616 sd = &per_cpu(core_domains, i).sd;
7618 set_domain_attribute(sd, attr);
7619 *sched_domain_span(sd) = cpu_coregroup_map(i);
7620 cpumask_and(sched_domain_span(sd),
7621 sched_domain_span(sd), cpu_map);
7624 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7627 #ifdef CONFIG_SCHED_SMT
7629 sd = &per_cpu(cpu_domains, i).sd;
7630 SD_INIT(sd, SIBLING);
7631 set_domain_attribute(sd, attr);
7632 cpumask_and(sched_domain_span(sd),
7633 &per_cpu(cpu_sibling_map, i), cpu_map);
7636 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7640 #ifdef CONFIG_SCHED_SMT
7641 /* Set up CPU (sibling) groups */
7642 for_each_cpu(i, cpu_map) {
7643 cpumask_and(this_sibling_map,
7644 &per_cpu(cpu_sibling_map, i), cpu_map);
7645 if (i != cpumask_first(this_sibling_map))
7648 init_sched_build_groups(this_sibling_map, cpu_map,
7650 send_covered, tmpmask);
7654 #ifdef CONFIG_SCHED_MC
7655 /* Set up multi-core groups */
7656 for_each_cpu(i, cpu_map) {
7657 /* FIXME: Use cpu_coregroup_mask */
7658 *this_core_map = cpu_coregroup_map(i);
7659 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7660 if (i != cpumask_first(this_core_map))
7663 init_sched_build_groups(this_core_map, cpu_map,
7665 send_covered, tmpmask);
7669 /* Set up physical groups */
7670 for (i = 0; i < nr_node_ids; i++) {
7671 /* FIXME: Use cpumask_of_node */
7672 *nodemask = node_to_cpumask(i);
7673 cpus_and(*nodemask, *nodemask, *cpu_map);
7674 if (cpumask_empty(nodemask))
7677 init_sched_build_groups(nodemask, cpu_map,
7679 send_covered, tmpmask);
7683 /* Set up node groups */
7685 init_sched_build_groups(cpu_map, cpu_map,
7686 &cpu_to_allnodes_group,
7687 send_covered, tmpmask);
7690 for (i = 0; i < nr_node_ids; i++) {
7691 /* Set up node groups */
7692 struct sched_group *sg, *prev;
7695 /* FIXME: Use cpumask_of_node */
7696 *nodemask = node_to_cpumask(i);
7697 cpumask_clear(covered);
7699 cpus_and(*nodemask, *nodemask, *cpu_map);
7700 if (cpumask_empty(nodemask)) {
7701 sched_group_nodes[i] = NULL;
7705 sched_domain_node_span(i, domainspan);
7706 cpumask_and(domainspan, domainspan, cpu_map);
7708 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7711 printk(KERN_WARNING "Can not alloc domain group for "
7715 sched_group_nodes[i] = sg;
7716 for_each_cpu(j, nodemask) {
7717 struct sched_domain *sd;
7719 sd = &per_cpu(node_domains, j);
7722 sg->__cpu_power = 0;
7723 cpumask_copy(sched_group_cpus(sg), nodemask);
7725 cpumask_or(covered, covered, nodemask);
7728 for (j = 0; j < nr_node_ids; j++) {
7729 int n = (i + j) % nr_node_ids;
7730 /* FIXME: Use cpumask_of_node */
7731 node_to_cpumask_ptr(pnodemask, n);
7733 cpumask_complement(notcovered, covered);
7734 cpumask_and(tmpmask, notcovered, cpu_map);
7735 cpumask_and(tmpmask, tmpmask, domainspan);
7736 if (cpumask_empty(tmpmask))
7739 cpumask_and(tmpmask, tmpmask, pnodemask);
7740 if (cpumask_empty(tmpmask))
7743 sg = kmalloc_node(sizeof(struct sched_group) +
7748 "Can not alloc domain group for node %d\n", j);
7751 sg->__cpu_power = 0;
7752 cpumask_copy(sched_group_cpus(sg), tmpmask);
7753 sg->next = prev->next;
7754 cpumask_or(covered, covered, tmpmask);
7761 /* Calculate CPU power for physical packages and nodes */
7762 #ifdef CONFIG_SCHED_SMT
7763 for_each_cpu(i, cpu_map) {
7764 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
7766 init_sched_groups_power(i, sd);
7769 #ifdef CONFIG_SCHED_MC
7770 for_each_cpu(i, cpu_map) {
7771 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
7773 init_sched_groups_power(i, sd);
7777 for_each_cpu(i, cpu_map) {
7778 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
7780 init_sched_groups_power(i, sd);
7784 for (i = 0; i < nr_node_ids; i++)
7785 init_numa_sched_groups_power(sched_group_nodes[i]);
7788 struct sched_group *sg;
7790 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7792 init_numa_sched_groups_power(sg);
7796 /* Attach the domains */
7797 for_each_cpu(i, cpu_map) {
7798 struct sched_domain *sd;
7799 #ifdef CONFIG_SCHED_SMT
7800 sd = &per_cpu(cpu_domains, i).sd;
7801 #elif defined(CONFIG_SCHED_MC)
7802 sd = &per_cpu(core_domains, i).sd;
7804 sd = &per_cpu(phys_domains, i).sd;
7806 cpu_attach_domain(sd, rd, i);
7812 free_cpumask_var(tmpmask);
7814 free_cpumask_var(send_covered);
7816 free_cpumask_var(this_core_map);
7817 free_this_sibling_map:
7818 free_cpumask_var(this_sibling_map);
7820 free_cpumask_var(nodemask);
7823 free_cpumask_var(notcovered);
7825 free_cpumask_var(covered);
7827 free_cpumask_var(domainspan);
7834 kfree(sched_group_nodes);
7840 free_sched_groups(cpu_map, tmpmask);
7841 free_rootdomain(rd);
7846 static int build_sched_domains(const struct cpumask *cpu_map)
7848 return __build_sched_domains(cpu_map, NULL);
7851 static struct cpumask *doms_cur; /* current sched domains */
7852 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7853 static struct sched_domain_attr *dattr_cur;
7854 /* attribues of custom domains in 'doms_cur' */
7857 * Special case: If a kmalloc of a doms_cur partition (array of
7858 * cpumask) fails, then fallback to a single sched domain,
7859 * as determined by the single cpumask fallback_doms.
7861 static cpumask_var_t fallback_doms;
7864 * arch_update_cpu_topology lets virtualized architectures update the
7865 * cpu core maps. It is supposed to return 1 if the topology changed
7866 * or 0 if it stayed the same.
7868 int __attribute__((weak)) arch_update_cpu_topology(void)
7874 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7875 * For now this just excludes isolated cpus, but could be used to
7876 * exclude other special cases in the future.
7878 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7882 arch_update_cpu_topology();
7884 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
7886 doms_cur = fallback_doms;
7887 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
7889 err = build_sched_domains(doms_cur);
7890 register_sched_domain_sysctl();
7895 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7896 struct cpumask *tmpmask)
7898 free_sched_groups(cpu_map, tmpmask);
7902 * Detach sched domains from a group of cpus specified in cpu_map
7903 * These cpus will now be attached to the NULL domain
7905 static void detach_destroy_domains(const struct cpumask *cpu_map)
7907 /* Save because hotplug lock held. */
7908 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7911 for_each_cpu(i, cpu_map)
7912 cpu_attach_domain(NULL, &def_root_domain, i);
7913 synchronize_sched();
7914 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7917 /* handle null as "default" */
7918 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7919 struct sched_domain_attr *new, int idx_new)
7921 struct sched_domain_attr tmp;
7928 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7929 new ? (new + idx_new) : &tmp,
7930 sizeof(struct sched_domain_attr));
7934 * Partition sched domains as specified by the 'ndoms_new'
7935 * cpumasks in the array doms_new[] of cpumasks. This compares
7936 * doms_new[] to the current sched domain partitioning, doms_cur[].
7937 * It destroys each deleted domain and builds each new domain.
7939 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
7940 * The masks don't intersect (don't overlap.) We should setup one
7941 * sched domain for each mask. CPUs not in any of the cpumasks will
7942 * not be load balanced. If the same cpumask appears both in the
7943 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7946 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7947 * ownership of it and will kfree it when done with it. If the caller
7948 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7949 * ndoms_new == 1, and partition_sched_domains() will fallback to
7950 * the single partition 'fallback_doms', it also forces the domains
7953 * If doms_new == NULL it will be replaced with cpu_online_mask.
7954 * ndoms_new == 0 is a special case for destroying existing domains,
7955 * and it will not create the default domain.
7957 * Call with hotplug lock held
7959 /* FIXME: Change to struct cpumask *doms_new[] */
7960 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
7961 struct sched_domain_attr *dattr_new)
7966 mutex_lock(&sched_domains_mutex);
7968 /* always unregister in case we don't destroy any domains */
7969 unregister_sched_domain_sysctl();
7971 /* Let architecture update cpu core mappings. */
7972 new_topology = arch_update_cpu_topology();
7974 n = doms_new ? ndoms_new : 0;
7976 /* Destroy deleted domains */
7977 for (i = 0; i < ndoms_cur; i++) {
7978 for (j = 0; j < n && !new_topology; j++) {
7979 if (cpumask_equal(&doms_cur[i], &doms_new[j])
7980 && dattrs_equal(dattr_cur, i, dattr_new, j))
7983 /* no match - a current sched domain not in new doms_new[] */
7984 detach_destroy_domains(doms_cur + i);
7989 if (doms_new == NULL) {
7991 doms_new = fallback_doms;
7992 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
7993 WARN_ON_ONCE(dattr_new);
7996 /* Build new domains */
7997 for (i = 0; i < ndoms_new; i++) {
7998 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7999 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8000 && dattrs_equal(dattr_new, i, dattr_cur, j))
8003 /* no match - add a new doms_new */
8004 __build_sched_domains(doms_new + i,
8005 dattr_new ? dattr_new + i : NULL);
8010 /* Remember the new sched domains */
8011 if (doms_cur != fallback_doms)
8013 kfree(dattr_cur); /* kfree(NULL) is safe */
8014 doms_cur = doms_new;
8015 dattr_cur = dattr_new;
8016 ndoms_cur = ndoms_new;
8018 register_sched_domain_sysctl();
8020 mutex_unlock(&sched_domains_mutex);
8023 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8024 int arch_reinit_sched_domains(void)
8028 /* Destroy domains first to force the rebuild */
8029 partition_sched_domains(0, NULL, NULL);
8031 rebuild_sched_domains();
8037 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8040 unsigned int level = 0;
8042 if (sscanf(buf, "%u", &level) != 1)
8046 * level is always be positive so don't check for
8047 * level < POWERSAVINGS_BALANCE_NONE which is 0
8048 * What happens on 0 or 1 byte write,
8049 * need to check for count as well?
8052 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8056 sched_smt_power_savings = level;
8058 sched_mc_power_savings = level;
8060 ret = arch_reinit_sched_domains();
8062 return ret ? ret : count;
8065 #ifdef CONFIG_SCHED_MC
8066 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8069 return sprintf(page, "%u\n", sched_mc_power_savings);
8071 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8072 const char *buf, size_t count)
8074 return sched_power_savings_store(buf, count, 0);
8076 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8077 sched_mc_power_savings_show,
8078 sched_mc_power_savings_store);
8081 #ifdef CONFIG_SCHED_SMT
8082 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8085 return sprintf(page, "%u\n", sched_smt_power_savings);
8087 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8088 const char *buf, size_t count)
8090 return sched_power_savings_store(buf, count, 1);
8092 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8093 sched_smt_power_savings_show,
8094 sched_smt_power_savings_store);
8097 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8101 #ifdef CONFIG_SCHED_SMT
8103 err = sysfs_create_file(&cls->kset.kobj,
8104 &attr_sched_smt_power_savings.attr);
8106 #ifdef CONFIG_SCHED_MC
8107 if (!err && mc_capable())
8108 err = sysfs_create_file(&cls->kset.kobj,
8109 &attr_sched_mc_power_savings.attr);
8113 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8115 #ifndef CONFIG_CPUSETS
8117 * Add online and remove offline CPUs from the scheduler domains.
8118 * When cpusets are enabled they take over this function.
8120 static int update_sched_domains(struct notifier_block *nfb,
8121 unsigned long action, void *hcpu)
8125 case CPU_ONLINE_FROZEN:
8127 case CPU_DEAD_FROZEN:
8128 partition_sched_domains(1, NULL, NULL);
8137 static int update_runtime(struct notifier_block *nfb,
8138 unsigned long action, void *hcpu)
8140 int cpu = (int)(long)hcpu;
8143 case CPU_DOWN_PREPARE:
8144 case CPU_DOWN_PREPARE_FROZEN:
8145 disable_runtime(cpu_rq(cpu));
8148 case CPU_DOWN_FAILED:
8149 case CPU_DOWN_FAILED_FROZEN:
8151 case CPU_ONLINE_FROZEN:
8152 enable_runtime(cpu_rq(cpu));
8160 void __init sched_init_smp(void)
8162 cpumask_var_t non_isolated_cpus;
8164 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8166 #if defined(CONFIG_NUMA)
8167 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8169 BUG_ON(sched_group_nodes_bycpu == NULL);
8172 mutex_lock(&sched_domains_mutex);
8173 arch_init_sched_domains(cpu_online_mask);
8174 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8175 if (cpumask_empty(non_isolated_cpus))
8176 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8177 mutex_unlock(&sched_domains_mutex);
8180 #ifndef CONFIG_CPUSETS
8181 /* XXX: Theoretical race here - CPU may be hotplugged now */
8182 hotcpu_notifier(update_sched_domains, 0);
8185 /* RT runtime code needs to handle some hotplug events */
8186 hotcpu_notifier(update_runtime, 0);
8190 /* Move init over to a non-isolated CPU */
8191 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8193 sched_init_granularity();
8194 free_cpumask_var(non_isolated_cpus);
8196 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8197 init_sched_rt_class();
8200 void __init sched_init_smp(void)
8202 sched_init_granularity();
8204 #endif /* CONFIG_SMP */
8206 int in_sched_functions(unsigned long addr)
8208 return in_lock_functions(addr) ||
8209 (addr >= (unsigned long)__sched_text_start
8210 && addr < (unsigned long)__sched_text_end);
8213 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8215 cfs_rq->tasks_timeline = RB_ROOT;
8216 INIT_LIST_HEAD(&cfs_rq->tasks);
8217 #ifdef CONFIG_FAIR_GROUP_SCHED
8220 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8223 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8225 struct rt_prio_array *array;
8228 array = &rt_rq->active;
8229 for (i = 0; i < MAX_RT_PRIO; i++) {
8230 INIT_LIST_HEAD(array->queue + i);
8231 __clear_bit(i, array->bitmap);
8233 /* delimiter for bitsearch: */
8234 __set_bit(MAX_RT_PRIO, array->bitmap);
8236 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8237 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8238 rt_rq->highest_prio.next = MAX_RT_PRIO;
8241 rt_rq->rt_nr_migratory = 0;
8242 rt_rq->overloaded = 0;
8243 plist_head_init(&rq->rt.pushable_tasks, &rq->lock);
8247 rt_rq->rt_throttled = 0;
8248 rt_rq->rt_runtime = 0;
8249 spin_lock_init(&rt_rq->rt_runtime_lock);
8251 #ifdef CONFIG_RT_GROUP_SCHED
8252 rt_rq->rt_nr_boosted = 0;
8257 #ifdef CONFIG_FAIR_GROUP_SCHED
8258 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8259 struct sched_entity *se, int cpu, int add,
8260 struct sched_entity *parent)
8262 struct rq *rq = cpu_rq(cpu);
8263 tg->cfs_rq[cpu] = cfs_rq;
8264 init_cfs_rq(cfs_rq, rq);
8267 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8270 /* se could be NULL for init_task_group */
8275 se->cfs_rq = &rq->cfs;
8277 se->cfs_rq = parent->my_q;
8280 se->load.weight = tg->shares;
8281 se->load.inv_weight = 0;
8282 se->parent = parent;
8286 #ifdef CONFIG_RT_GROUP_SCHED
8287 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8288 struct sched_rt_entity *rt_se, int cpu, int add,
8289 struct sched_rt_entity *parent)
8291 struct rq *rq = cpu_rq(cpu);
8293 tg->rt_rq[cpu] = rt_rq;
8294 init_rt_rq(rt_rq, rq);
8296 rt_rq->rt_se = rt_se;
8297 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8299 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8301 tg->rt_se[cpu] = rt_se;
8306 rt_se->rt_rq = &rq->rt;
8308 rt_se->rt_rq = parent->my_q;
8310 rt_se->my_q = rt_rq;
8311 rt_se->parent = parent;
8312 INIT_LIST_HEAD(&rt_se->run_list);
8316 void __init sched_init(void)
8319 unsigned long alloc_size = 0, ptr;
8321 #ifdef CONFIG_FAIR_GROUP_SCHED
8322 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8324 #ifdef CONFIG_RT_GROUP_SCHED
8325 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8327 #ifdef CONFIG_USER_SCHED
8331 * As sched_init() is called before page_alloc is setup,
8332 * we use alloc_bootmem().
8335 ptr = (unsigned long)alloc_bootmem(alloc_size);
8337 #ifdef CONFIG_FAIR_GROUP_SCHED
8338 init_task_group.se = (struct sched_entity **)ptr;
8339 ptr += nr_cpu_ids * sizeof(void **);
8341 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8342 ptr += nr_cpu_ids * sizeof(void **);
8344 #ifdef CONFIG_USER_SCHED
8345 root_task_group.se = (struct sched_entity **)ptr;
8346 ptr += nr_cpu_ids * sizeof(void **);
8348 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8349 ptr += nr_cpu_ids * sizeof(void **);
8350 #endif /* CONFIG_USER_SCHED */
8351 #endif /* CONFIG_FAIR_GROUP_SCHED */
8352 #ifdef CONFIG_RT_GROUP_SCHED
8353 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8354 ptr += nr_cpu_ids * sizeof(void **);
8356 init_task_group.rt_rq = (struct rt_rq **)ptr;
8357 ptr += nr_cpu_ids * sizeof(void **);
8359 #ifdef CONFIG_USER_SCHED
8360 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8361 ptr += nr_cpu_ids * sizeof(void **);
8363 root_task_group.rt_rq = (struct rt_rq **)ptr;
8364 ptr += nr_cpu_ids * sizeof(void **);
8365 #endif /* CONFIG_USER_SCHED */
8366 #endif /* CONFIG_RT_GROUP_SCHED */
8370 init_defrootdomain();
8373 init_rt_bandwidth(&def_rt_bandwidth,
8374 global_rt_period(), global_rt_runtime());
8376 #ifdef CONFIG_RT_GROUP_SCHED
8377 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8378 global_rt_period(), global_rt_runtime());
8379 #ifdef CONFIG_USER_SCHED
8380 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8381 global_rt_period(), RUNTIME_INF);
8382 #endif /* CONFIG_USER_SCHED */
8383 #endif /* CONFIG_RT_GROUP_SCHED */
8385 #ifdef CONFIG_GROUP_SCHED
8386 list_add(&init_task_group.list, &task_groups);
8387 INIT_LIST_HEAD(&init_task_group.children);
8389 #ifdef CONFIG_USER_SCHED
8390 INIT_LIST_HEAD(&root_task_group.children);
8391 init_task_group.parent = &root_task_group;
8392 list_add(&init_task_group.siblings, &root_task_group.children);
8393 #endif /* CONFIG_USER_SCHED */
8394 #endif /* CONFIG_GROUP_SCHED */
8396 for_each_possible_cpu(i) {
8400 spin_lock_init(&rq->lock);
8402 init_cfs_rq(&rq->cfs, rq);
8403 init_rt_rq(&rq->rt, rq);
8404 #ifdef CONFIG_FAIR_GROUP_SCHED
8405 init_task_group.shares = init_task_group_load;
8406 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8407 #ifdef CONFIG_CGROUP_SCHED
8409 * How much cpu bandwidth does init_task_group get?
8411 * In case of task-groups formed thr' the cgroup filesystem, it
8412 * gets 100% of the cpu resources in the system. This overall
8413 * system cpu resource is divided among the tasks of
8414 * init_task_group and its child task-groups in a fair manner,
8415 * based on each entity's (task or task-group's) weight
8416 * (se->load.weight).
8418 * In other words, if init_task_group has 10 tasks of weight
8419 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8420 * then A0's share of the cpu resource is:
8422 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8424 * We achieve this by letting init_task_group's tasks sit
8425 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8427 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8428 #elif defined CONFIG_USER_SCHED
8429 root_task_group.shares = NICE_0_LOAD;
8430 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8432 * In case of task-groups formed thr' the user id of tasks,
8433 * init_task_group represents tasks belonging to root user.
8434 * Hence it forms a sibling of all subsequent groups formed.
8435 * In this case, init_task_group gets only a fraction of overall
8436 * system cpu resource, based on the weight assigned to root
8437 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8438 * by letting tasks of init_task_group sit in a separate cfs_rq
8439 * (init_cfs_rq) and having one entity represent this group of
8440 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8442 init_tg_cfs_entry(&init_task_group,
8443 &per_cpu(init_cfs_rq, i),
8444 &per_cpu(init_sched_entity, i), i, 1,
8445 root_task_group.se[i]);
8448 #endif /* CONFIG_FAIR_GROUP_SCHED */
8450 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8451 #ifdef CONFIG_RT_GROUP_SCHED
8452 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8453 #ifdef CONFIG_CGROUP_SCHED
8454 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8455 #elif defined CONFIG_USER_SCHED
8456 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8457 init_tg_rt_entry(&init_task_group,
8458 &per_cpu(init_rt_rq, i),
8459 &per_cpu(init_sched_rt_entity, i), i, 1,
8460 root_task_group.rt_se[i]);
8464 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8465 rq->cpu_load[j] = 0;
8469 rq->active_balance = 0;
8470 rq->next_balance = jiffies;
8474 rq->migration_thread = NULL;
8475 INIT_LIST_HEAD(&rq->migration_queue);
8476 rq_attach_root(rq, &def_root_domain);
8479 atomic_set(&rq->nr_iowait, 0);
8482 set_load_weight(&init_task);
8484 #ifdef CONFIG_PREEMPT_NOTIFIERS
8485 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8489 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8492 #ifdef CONFIG_RT_MUTEXES
8493 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8497 * The boot idle thread does lazy MMU switching as well:
8499 atomic_inc(&init_mm.mm_count);
8500 enter_lazy_tlb(&init_mm, current);
8503 * Make us the idle thread. Technically, schedule() should not be
8504 * called from this thread, however somewhere below it might be,
8505 * but because we are the idle thread, we just pick up running again
8506 * when this runqueue becomes "idle".
8508 init_idle(current, smp_processor_id());
8510 * During early bootup we pretend to be a normal task:
8512 current->sched_class = &fair_sched_class;
8514 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8515 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
8518 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
8520 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8523 scheduler_running = 1;
8526 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8527 void __might_sleep(char *file, int line)
8530 static unsigned long prev_jiffy; /* ratelimiting */
8532 if ((!in_atomic() && !irqs_disabled()) ||
8533 system_state != SYSTEM_RUNNING || oops_in_progress)
8535 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8537 prev_jiffy = jiffies;
8540 "BUG: sleeping function called from invalid context at %s:%d\n",
8543 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8544 in_atomic(), irqs_disabled(),
8545 current->pid, current->comm);
8547 debug_show_held_locks(current);
8548 if (irqs_disabled())
8549 print_irqtrace_events(current);
8553 EXPORT_SYMBOL(__might_sleep);
8556 #ifdef CONFIG_MAGIC_SYSRQ
8557 static void normalize_task(struct rq *rq, struct task_struct *p)
8561 update_rq_clock(rq);
8562 on_rq = p->se.on_rq;
8564 deactivate_task(rq, p, 0);
8565 __setscheduler(rq, p, SCHED_NORMAL, 0);
8567 activate_task(rq, p, 0);
8568 resched_task(rq->curr);
8572 void normalize_rt_tasks(void)
8574 struct task_struct *g, *p;
8575 unsigned long flags;
8578 read_lock_irqsave(&tasklist_lock, flags);
8579 do_each_thread(g, p) {
8581 * Only normalize user tasks:
8586 p->se.exec_start = 0;
8587 #ifdef CONFIG_SCHEDSTATS
8588 p->se.wait_start = 0;
8589 p->se.sleep_start = 0;
8590 p->se.block_start = 0;
8595 * Renice negative nice level userspace
8598 if (TASK_NICE(p) < 0 && p->mm)
8599 set_user_nice(p, 0);
8603 spin_lock(&p->pi_lock);
8604 rq = __task_rq_lock(p);
8606 normalize_task(rq, p);
8608 __task_rq_unlock(rq);
8609 spin_unlock(&p->pi_lock);
8610 } while_each_thread(g, p);
8612 read_unlock_irqrestore(&tasklist_lock, flags);
8615 #endif /* CONFIG_MAGIC_SYSRQ */
8619 * These functions are only useful for the IA64 MCA handling.
8621 * They can only be called when the whole system has been
8622 * stopped - every CPU needs to be quiescent, and no scheduling
8623 * activity can take place. Using them for anything else would
8624 * be a serious bug, and as a result, they aren't even visible
8625 * under any other configuration.
8629 * curr_task - return the current task for a given cpu.
8630 * @cpu: the processor in question.
8632 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8634 struct task_struct *curr_task(int cpu)
8636 return cpu_curr(cpu);
8640 * set_curr_task - set the current task for a given cpu.
8641 * @cpu: the processor in question.
8642 * @p: the task pointer to set.
8644 * Description: This function must only be used when non-maskable interrupts
8645 * are serviced on a separate stack. It allows the architecture to switch the
8646 * notion of the current task on a cpu in a non-blocking manner. This function
8647 * must be called with all CPU's synchronized, and interrupts disabled, the
8648 * and caller must save the original value of the current task (see
8649 * curr_task() above) and restore that value before reenabling interrupts and
8650 * re-starting the system.
8652 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8654 void set_curr_task(int cpu, struct task_struct *p)
8661 #ifdef CONFIG_FAIR_GROUP_SCHED
8662 static void free_fair_sched_group(struct task_group *tg)
8666 for_each_possible_cpu(i) {
8668 kfree(tg->cfs_rq[i]);
8678 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8680 struct cfs_rq *cfs_rq;
8681 struct sched_entity *se;
8685 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8688 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8692 tg->shares = NICE_0_LOAD;
8694 for_each_possible_cpu(i) {
8697 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8698 GFP_KERNEL, cpu_to_node(i));
8702 se = kzalloc_node(sizeof(struct sched_entity),
8703 GFP_KERNEL, cpu_to_node(i));
8707 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8716 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8718 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8719 &cpu_rq(cpu)->leaf_cfs_rq_list);
8722 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8724 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8726 #else /* !CONFG_FAIR_GROUP_SCHED */
8727 static inline void free_fair_sched_group(struct task_group *tg)
8732 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8737 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8741 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8744 #endif /* CONFIG_FAIR_GROUP_SCHED */
8746 #ifdef CONFIG_RT_GROUP_SCHED
8747 static void free_rt_sched_group(struct task_group *tg)
8751 destroy_rt_bandwidth(&tg->rt_bandwidth);
8753 for_each_possible_cpu(i) {
8755 kfree(tg->rt_rq[i]);
8757 kfree(tg->rt_se[i]);
8765 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8767 struct rt_rq *rt_rq;
8768 struct sched_rt_entity *rt_se;
8772 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8775 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8779 init_rt_bandwidth(&tg->rt_bandwidth,
8780 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8782 for_each_possible_cpu(i) {
8785 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8786 GFP_KERNEL, cpu_to_node(i));
8790 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8791 GFP_KERNEL, cpu_to_node(i));
8795 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8804 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8806 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8807 &cpu_rq(cpu)->leaf_rt_rq_list);
8810 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8812 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8814 #else /* !CONFIG_RT_GROUP_SCHED */
8815 static inline void free_rt_sched_group(struct task_group *tg)
8820 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8825 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8829 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8832 #endif /* CONFIG_RT_GROUP_SCHED */
8834 #ifdef CONFIG_GROUP_SCHED
8835 static void free_sched_group(struct task_group *tg)
8837 free_fair_sched_group(tg);
8838 free_rt_sched_group(tg);
8842 /* allocate runqueue etc for a new task group */
8843 struct task_group *sched_create_group(struct task_group *parent)
8845 struct task_group *tg;
8846 unsigned long flags;
8849 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8851 return ERR_PTR(-ENOMEM);
8853 if (!alloc_fair_sched_group(tg, parent))
8856 if (!alloc_rt_sched_group(tg, parent))
8859 spin_lock_irqsave(&task_group_lock, flags);
8860 for_each_possible_cpu(i) {
8861 register_fair_sched_group(tg, i);
8862 register_rt_sched_group(tg, i);
8864 list_add_rcu(&tg->list, &task_groups);
8866 WARN_ON(!parent); /* root should already exist */
8868 tg->parent = parent;
8869 INIT_LIST_HEAD(&tg->children);
8870 list_add_rcu(&tg->siblings, &parent->children);
8871 spin_unlock_irqrestore(&task_group_lock, flags);
8876 free_sched_group(tg);
8877 return ERR_PTR(-ENOMEM);
8880 /* rcu callback to free various structures associated with a task group */
8881 static void free_sched_group_rcu(struct rcu_head *rhp)
8883 /* now it should be safe to free those cfs_rqs */
8884 free_sched_group(container_of(rhp, struct task_group, rcu));
8887 /* Destroy runqueue etc associated with a task group */
8888 void sched_destroy_group(struct task_group *tg)
8890 unsigned long flags;
8893 spin_lock_irqsave(&task_group_lock, flags);
8894 for_each_possible_cpu(i) {
8895 unregister_fair_sched_group(tg, i);
8896 unregister_rt_sched_group(tg, i);
8898 list_del_rcu(&tg->list);
8899 list_del_rcu(&tg->siblings);
8900 spin_unlock_irqrestore(&task_group_lock, flags);
8902 /* wait for possible concurrent references to cfs_rqs complete */
8903 call_rcu(&tg->rcu, free_sched_group_rcu);
8906 /* change task's runqueue when it moves between groups.
8907 * The caller of this function should have put the task in its new group
8908 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8909 * reflect its new group.
8911 void sched_move_task(struct task_struct *tsk)
8914 unsigned long flags;
8917 rq = task_rq_lock(tsk, &flags);
8919 update_rq_clock(rq);
8921 running = task_current(rq, tsk);
8922 on_rq = tsk->se.on_rq;
8925 dequeue_task(rq, tsk, 0);
8926 if (unlikely(running))
8927 tsk->sched_class->put_prev_task(rq, tsk);
8929 set_task_rq(tsk, task_cpu(tsk));
8931 #ifdef CONFIG_FAIR_GROUP_SCHED
8932 if (tsk->sched_class->moved_group)
8933 tsk->sched_class->moved_group(tsk);
8936 if (unlikely(running))
8937 tsk->sched_class->set_curr_task(rq);
8939 enqueue_task(rq, tsk, 0);
8941 task_rq_unlock(rq, &flags);
8943 #endif /* CONFIG_GROUP_SCHED */
8945 #ifdef CONFIG_FAIR_GROUP_SCHED
8946 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8948 struct cfs_rq *cfs_rq = se->cfs_rq;
8953 dequeue_entity(cfs_rq, se, 0);
8955 se->load.weight = shares;
8956 se->load.inv_weight = 0;
8959 enqueue_entity(cfs_rq, se, 0);
8962 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8964 struct cfs_rq *cfs_rq = se->cfs_rq;
8965 struct rq *rq = cfs_rq->rq;
8966 unsigned long flags;
8968 spin_lock_irqsave(&rq->lock, flags);
8969 __set_se_shares(se, shares);
8970 spin_unlock_irqrestore(&rq->lock, flags);
8973 static DEFINE_MUTEX(shares_mutex);
8975 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8978 unsigned long flags;
8981 * We can't change the weight of the root cgroup.
8986 if (shares < MIN_SHARES)
8987 shares = MIN_SHARES;
8988 else if (shares > MAX_SHARES)
8989 shares = MAX_SHARES;
8991 mutex_lock(&shares_mutex);
8992 if (tg->shares == shares)
8995 spin_lock_irqsave(&task_group_lock, flags);
8996 for_each_possible_cpu(i)
8997 unregister_fair_sched_group(tg, i);
8998 list_del_rcu(&tg->siblings);
8999 spin_unlock_irqrestore(&task_group_lock, flags);
9001 /* wait for any ongoing reference to this group to finish */
9002 synchronize_sched();
9005 * Now we are free to modify the group's share on each cpu
9006 * w/o tripping rebalance_share or load_balance_fair.
9008 tg->shares = shares;
9009 for_each_possible_cpu(i) {
9013 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9014 set_se_shares(tg->se[i], shares);
9018 * Enable load balance activity on this group, by inserting it back on
9019 * each cpu's rq->leaf_cfs_rq_list.
9021 spin_lock_irqsave(&task_group_lock, flags);
9022 for_each_possible_cpu(i)
9023 register_fair_sched_group(tg, i);
9024 list_add_rcu(&tg->siblings, &tg->parent->children);
9025 spin_unlock_irqrestore(&task_group_lock, flags);
9027 mutex_unlock(&shares_mutex);
9031 unsigned long sched_group_shares(struct task_group *tg)
9037 #ifdef CONFIG_RT_GROUP_SCHED
9039 * Ensure that the real time constraints are schedulable.
9041 static DEFINE_MUTEX(rt_constraints_mutex);
9043 static unsigned long to_ratio(u64 period, u64 runtime)
9045 if (runtime == RUNTIME_INF)
9048 return div64_u64(runtime << 20, period);
9051 /* Must be called with tasklist_lock held */
9052 static inline int tg_has_rt_tasks(struct task_group *tg)
9054 struct task_struct *g, *p;
9056 do_each_thread(g, p) {
9057 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9059 } while_each_thread(g, p);
9064 struct rt_schedulable_data {
9065 struct task_group *tg;
9070 static int tg_schedulable(struct task_group *tg, void *data)
9072 struct rt_schedulable_data *d = data;
9073 struct task_group *child;
9074 unsigned long total, sum = 0;
9075 u64 period, runtime;
9077 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9078 runtime = tg->rt_bandwidth.rt_runtime;
9081 period = d->rt_period;
9082 runtime = d->rt_runtime;
9086 * Cannot have more runtime than the period.
9088 if (runtime > period && runtime != RUNTIME_INF)
9092 * Ensure we don't starve existing RT tasks.
9094 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9097 total = to_ratio(period, runtime);
9100 * Nobody can have more than the global setting allows.
9102 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9106 * The sum of our children's runtime should not exceed our own.
9108 list_for_each_entry_rcu(child, &tg->children, siblings) {
9109 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9110 runtime = child->rt_bandwidth.rt_runtime;
9112 if (child == d->tg) {
9113 period = d->rt_period;
9114 runtime = d->rt_runtime;
9117 sum += to_ratio(period, runtime);
9126 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9128 struct rt_schedulable_data data = {
9130 .rt_period = period,
9131 .rt_runtime = runtime,
9134 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9137 static int tg_set_bandwidth(struct task_group *tg,
9138 u64 rt_period, u64 rt_runtime)
9142 mutex_lock(&rt_constraints_mutex);
9143 read_lock(&tasklist_lock);
9144 err = __rt_schedulable(tg, rt_period, rt_runtime);
9148 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9149 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9150 tg->rt_bandwidth.rt_runtime = rt_runtime;
9152 for_each_possible_cpu(i) {
9153 struct rt_rq *rt_rq = tg->rt_rq[i];
9155 spin_lock(&rt_rq->rt_runtime_lock);
9156 rt_rq->rt_runtime = rt_runtime;
9157 spin_unlock(&rt_rq->rt_runtime_lock);
9159 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9161 read_unlock(&tasklist_lock);
9162 mutex_unlock(&rt_constraints_mutex);
9167 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9169 u64 rt_runtime, rt_period;
9171 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9172 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9173 if (rt_runtime_us < 0)
9174 rt_runtime = RUNTIME_INF;
9176 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9179 long sched_group_rt_runtime(struct task_group *tg)
9183 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9186 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9187 do_div(rt_runtime_us, NSEC_PER_USEC);
9188 return rt_runtime_us;
9191 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9193 u64 rt_runtime, rt_period;
9195 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9196 rt_runtime = tg->rt_bandwidth.rt_runtime;
9201 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9204 long sched_group_rt_period(struct task_group *tg)
9208 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9209 do_div(rt_period_us, NSEC_PER_USEC);
9210 return rt_period_us;
9213 static int sched_rt_global_constraints(void)
9215 u64 runtime, period;
9218 if (sysctl_sched_rt_period <= 0)
9221 runtime = global_rt_runtime();
9222 period = global_rt_period();
9225 * Sanity check on the sysctl variables.
9227 if (runtime > period && runtime != RUNTIME_INF)
9230 mutex_lock(&rt_constraints_mutex);
9231 read_lock(&tasklist_lock);
9232 ret = __rt_schedulable(NULL, 0, 0);
9233 read_unlock(&tasklist_lock);
9234 mutex_unlock(&rt_constraints_mutex);
9238 #else /* !CONFIG_RT_GROUP_SCHED */
9239 static int sched_rt_global_constraints(void)
9241 unsigned long flags;
9244 if (sysctl_sched_rt_period <= 0)
9247 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9248 for_each_possible_cpu(i) {
9249 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9251 spin_lock(&rt_rq->rt_runtime_lock);
9252 rt_rq->rt_runtime = global_rt_runtime();
9253 spin_unlock(&rt_rq->rt_runtime_lock);
9255 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9259 #endif /* CONFIG_RT_GROUP_SCHED */
9261 int sched_rt_handler(struct ctl_table *table, int write,
9262 struct file *filp, void __user *buffer, size_t *lenp,
9266 int old_period, old_runtime;
9267 static DEFINE_MUTEX(mutex);
9270 old_period = sysctl_sched_rt_period;
9271 old_runtime = sysctl_sched_rt_runtime;
9273 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9275 if (!ret && write) {
9276 ret = sched_rt_global_constraints();
9278 sysctl_sched_rt_period = old_period;
9279 sysctl_sched_rt_runtime = old_runtime;
9281 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9282 def_rt_bandwidth.rt_period =
9283 ns_to_ktime(global_rt_period());
9286 mutex_unlock(&mutex);
9291 #ifdef CONFIG_CGROUP_SCHED
9293 /* return corresponding task_group object of a cgroup */
9294 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9296 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9297 struct task_group, css);
9300 static struct cgroup_subsys_state *
9301 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9303 struct task_group *tg, *parent;
9305 if (!cgrp->parent) {
9306 /* This is early initialization for the top cgroup */
9307 return &init_task_group.css;
9310 parent = cgroup_tg(cgrp->parent);
9311 tg = sched_create_group(parent);
9313 return ERR_PTR(-ENOMEM);
9319 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9321 struct task_group *tg = cgroup_tg(cgrp);
9323 sched_destroy_group(tg);
9327 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9328 struct task_struct *tsk)
9330 #ifdef CONFIG_RT_GROUP_SCHED
9331 /* Don't accept realtime tasks when there is no way for them to run */
9332 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9335 /* We don't support RT-tasks being in separate groups */
9336 if (tsk->sched_class != &fair_sched_class)
9344 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9345 struct cgroup *old_cont, struct task_struct *tsk)
9347 sched_move_task(tsk);
9350 #ifdef CONFIG_FAIR_GROUP_SCHED
9351 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9354 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9357 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9359 struct task_group *tg = cgroup_tg(cgrp);
9361 return (u64) tg->shares;
9363 #endif /* CONFIG_FAIR_GROUP_SCHED */
9365 #ifdef CONFIG_RT_GROUP_SCHED
9366 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9369 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9372 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9374 return sched_group_rt_runtime(cgroup_tg(cgrp));
9377 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9380 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9383 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9385 return sched_group_rt_period(cgroup_tg(cgrp));
9387 #endif /* CONFIG_RT_GROUP_SCHED */
9389 static struct cftype cpu_files[] = {
9390 #ifdef CONFIG_FAIR_GROUP_SCHED
9393 .read_u64 = cpu_shares_read_u64,
9394 .write_u64 = cpu_shares_write_u64,
9397 #ifdef CONFIG_RT_GROUP_SCHED
9399 .name = "rt_runtime_us",
9400 .read_s64 = cpu_rt_runtime_read,
9401 .write_s64 = cpu_rt_runtime_write,
9404 .name = "rt_period_us",
9405 .read_u64 = cpu_rt_period_read_uint,
9406 .write_u64 = cpu_rt_period_write_uint,
9411 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9413 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9416 struct cgroup_subsys cpu_cgroup_subsys = {
9418 .create = cpu_cgroup_create,
9419 .destroy = cpu_cgroup_destroy,
9420 .can_attach = cpu_cgroup_can_attach,
9421 .attach = cpu_cgroup_attach,
9422 .populate = cpu_cgroup_populate,
9423 .subsys_id = cpu_cgroup_subsys_id,
9427 #endif /* CONFIG_CGROUP_SCHED */
9429 #ifdef CONFIG_CGROUP_CPUACCT
9432 * CPU accounting code for task groups.
9434 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9435 * (balbir@in.ibm.com).
9438 /* track cpu usage of a group of tasks and its child groups */
9440 struct cgroup_subsys_state css;
9441 /* cpuusage holds pointer to a u64-type object on every cpu */
9443 struct cpuacct *parent;
9446 struct cgroup_subsys cpuacct_subsys;
9448 /* return cpu accounting group corresponding to this container */
9449 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9451 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9452 struct cpuacct, css);
9455 /* return cpu accounting group to which this task belongs */
9456 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9458 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9459 struct cpuacct, css);
9462 /* create a new cpu accounting group */
9463 static struct cgroup_subsys_state *cpuacct_create(
9464 struct cgroup_subsys *ss, struct cgroup *cgrp)
9466 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9469 return ERR_PTR(-ENOMEM);
9471 ca->cpuusage = alloc_percpu(u64);
9472 if (!ca->cpuusage) {
9474 return ERR_PTR(-ENOMEM);
9478 ca->parent = cgroup_ca(cgrp->parent);
9483 /* destroy an existing cpu accounting group */
9485 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9487 struct cpuacct *ca = cgroup_ca(cgrp);
9489 free_percpu(ca->cpuusage);
9493 /* return total cpu usage (in nanoseconds) of a group */
9494 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9496 struct cpuacct *ca = cgroup_ca(cgrp);
9497 u64 totalcpuusage = 0;
9500 for_each_possible_cpu(i) {
9501 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9504 * Take rq->lock to make 64-bit addition safe on 32-bit
9507 spin_lock_irq(&cpu_rq(i)->lock);
9508 totalcpuusage += *cpuusage;
9509 spin_unlock_irq(&cpu_rq(i)->lock);
9512 return totalcpuusage;
9515 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9518 struct cpuacct *ca = cgroup_ca(cgrp);
9527 for_each_possible_cpu(i) {
9528 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9530 spin_lock_irq(&cpu_rq(i)->lock);
9532 spin_unlock_irq(&cpu_rq(i)->lock);
9538 static struct cftype files[] = {
9541 .read_u64 = cpuusage_read,
9542 .write_u64 = cpuusage_write,
9546 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9548 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9552 * charge this task's execution time to its accounting group.
9554 * called with rq->lock held.
9556 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9561 if (!cpuacct_subsys.active)
9564 cpu = task_cpu(tsk);
9567 for (; ca; ca = ca->parent) {
9568 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9569 *cpuusage += cputime;
9573 struct cgroup_subsys cpuacct_subsys = {
9575 .create = cpuacct_create,
9576 .destroy = cpuacct_destroy,
9577 .populate = cpuacct_populate,
9578 .subsys_id = cpuacct_subsys_id,
9580 #endif /* CONFIG_CGROUP_CPUACCT */