4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task);
122 DEFINE_TRACE(sched_wakeup);
123 DEFINE_TRACE(sched_wakeup_new);
124 DEFINE_TRACE(sched_switch);
125 DEFINE_TRACE(sched_migrate_task);
129 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
130 * Since cpu_power is a 'constant', we can use a reciprocal divide.
132 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
134 return reciprocal_divide(load, sg->reciprocal_cpu_power);
138 * Each time a sched group cpu_power is changed,
139 * we must compute its reciprocal value
141 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
143 sg->__cpu_power += val;
144 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
148 static inline int rt_policy(int policy)
150 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
155 static inline int task_has_rt_policy(struct task_struct *p)
157 return rt_policy(p->policy);
161 * This is the priority-queue data structure of the RT scheduling class:
163 struct rt_prio_array {
164 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
165 struct list_head queue[MAX_RT_PRIO];
168 struct rt_bandwidth {
169 /* nests inside the rq lock: */
170 spinlock_t rt_runtime_lock;
173 struct hrtimer rt_period_timer;
176 static struct rt_bandwidth def_rt_bandwidth;
178 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
180 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
182 struct rt_bandwidth *rt_b =
183 container_of(timer, struct rt_bandwidth, rt_period_timer);
189 now = hrtimer_cb_get_time(timer);
190 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
195 idle = do_sched_rt_period_timer(rt_b, overrun);
198 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
202 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
204 rt_b->rt_period = ns_to_ktime(period);
205 rt_b->rt_runtime = runtime;
207 spin_lock_init(&rt_b->rt_runtime_lock);
209 hrtimer_init(&rt_b->rt_period_timer,
210 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
211 rt_b->rt_period_timer.function = sched_rt_period_timer;
214 static inline int rt_bandwidth_enabled(void)
216 return sysctl_sched_rt_runtime >= 0;
219 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
223 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
226 if (hrtimer_active(&rt_b->rt_period_timer))
229 spin_lock(&rt_b->rt_runtime_lock);
231 if (hrtimer_active(&rt_b->rt_period_timer))
234 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
235 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
236 hrtimer_start_expires(&rt_b->rt_period_timer,
239 spin_unlock(&rt_b->rt_runtime_lock);
242 #ifdef CONFIG_RT_GROUP_SCHED
243 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
245 hrtimer_cancel(&rt_b->rt_period_timer);
250 * sched_domains_mutex serializes calls to arch_init_sched_domains,
251 * detach_destroy_domains and partition_sched_domains.
253 static DEFINE_MUTEX(sched_domains_mutex);
255 #ifdef CONFIG_GROUP_SCHED
257 #include <linux/cgroup.h>
261 static LIST_HEAD(task_groups);
263 /* task group related information */
265 #ifdef CONFIG_CGROUP_SCHED
266 struct cgroup_subsys_state css;
269 #ifdef CONFIG_USER_SCHED
273 #ifdef CONFIG_FAIR_GROUP_SCHED
274 /* schedulable entities of this group on each cpu */
275 struct sched_entity **se;
276 /* runqueue "owned" by this group on each cpu */
277 struct cfs_rq **cfs_rq;
278 unsigned long shares;
281 #ifdef CONFIG_RT_GROUP_SCHED
282 struct sched_rt_entity **rt_se;
283 struct rt_rq **rt_rq;
285 struct rt_bandwidth rt_bandwidth;
289 struct list_head list;
291 struct task_group *parent;
292 struct list_head siblings;
293 struct list_head children;
296 #ifdef CONFIG_USER_SCHED
298 /* Helper function to pass uid information to create_sched_user() */
299 void set_tg_uid(struct user_struct *user)
301 user->tg->uid = user->uid;
306 * Every UID task group (including init_task_group aka UID-0) will
307 * be a child to this group.
309 struct task_group root_task_group;
311 #ifdef CONFIG_FAIR_GROUP_SCHED
312 /* Default task group's sched entity on each cpu */
313 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
314 /* Default task group's cfs_rq on each cpu */
315 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
316 #endif /* CONFIG_FAIR_GROUP_SCHED */
318 #ifdef CONFIG_RT_GROUP_SCHED
319 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
320 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
321 #endif /* CONFIG_RT_GROUP_SCHED */
322 #else /* !CONFIG_USER_SCHED */
323 #define root_task_group init_task_group
324 #endif /* CONFIG_USER_SCHED */
326 /* task_group_lock serializes add/remove of task groups and also changes to
327 * a task group's cpu shares.
329 static DEFINE_SPINLOCK(task_group_lock);
331 #ifdef CONFIG_FAIR_GROUP_SCHED
332 #ifdef CONFIG_USER_SCHED
333 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
334 #else /* !CONFIG_USER_SCHED */
335 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
336 #endif /* CONFIG_USER_SCHED */
339 * A weight of 0 or 1 can cause arithmetics problems.
340 * A weight of a cfs_rq is the sum of weights of which entities
341 * are queued on this cfs_rq, so a weight of a entity should not be
342 * too large, so as the shares value of a task group.
343 * (The default weight is 1024 - so there's no practical
344 * limitation from this.)
347 #define MAX_SHARES (1UL << 18)
349 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
352 /* Default task group.
353 * Every task in system belong to this group at bootup.
355 struct task_group init_task_group;
357 /* return group to which a task belongs */
358 static inline struct task_group *task_group(struct task_struct *p)
360 struct task_group *tg;
362 #ifdef CONFIG_USER_SCHED
364 tg = __task_cred(p)->user->tg;
366 #elif defined(CONFIG_CGROUP_SCHED)
367 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
368 struct task_group, css);
370 tg = &init_task_group;
375 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
376 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
378 #ifdef CONFIG_FAIR_GROUP_SCHED
379 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
380 p->se.parent = task_group(p)->se[cpu];
383 #ifdef CONFIG_RT_GROUP_SCHED
384 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
385 p->rt.parent = task_group(p)->rt_se[cpu];
391 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
392 static inline struct task_group *task_group(struct task_struct *p)
397 #endif /* CONFIG_GROUP_SCHED */
399 /* CFS-related fields in a runqueue */
401 struct load_weight load;
402 unsigned long nr_running;
407 struct rb_root tasks_timeline;
408 struct rb_node *rb_leftmost;
410 struct list_head tasks;
411 struct list_head *balance_iterator;
414 * 'curr' points to currently running entity on this cfs_rq.
415 * It is set to NULL otherwise (i.e when none are currently running).
417 struct sched_entity *curr, *next, *last;
419 unsigned int nr_spread_over;
421 #ifdef CONFIG_FAIR_GROUP_SCHED
422 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
425 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
426 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
427 * (like users, containers etc.)
429 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
430 * list is used during load balance.
432 struct list_head leaf_cfs_rq_list;
433 struct task_group *tg; /* group that "owns" this runqueue */
437 * the part of load.weight contributed by tasks
439 unsigned long task_weight;
442 * h_load = weight * f(tg)
444 * Where f(tg) is the recursive weight fraction assigned to
447 unsigned long h_load;
450 * this cpu's part of tg->shares
452 unsigned long shares;
455 * load.weight at the time we set shares
457 unsigned long rq_weight;
462 /* Real-Time classes' related field in a runqueue: */
464 struct rt_prio_array active;
465 unsigned long rt_nr_running;
466 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
467 int highest_prio; /* highest queued rt task prio */
470 unsigned long rt_nr_migratory;
476 /* Nests inside the rq lock: */
477 spinlock_t rt_runtime_lock;
479 #ifdef CONFIG_RT_GROUP_SCHED
480 unsigned long rt_nr_boosted;
483 struct list_head leaf_rt_rq_list;
484 struct task_group *tg;
485 struct sched_rt_entity *rt_se;
492 * We add the notion of a root-domain which will be used to define per-domain
493 * variables. Each exclusive cpuset essentially defines an island domain by
494 * fully partitioning the member cpus from any other cpuset. Whenever a new
495 * exclusive cpuset is created, we also create and attach a new root-domain
505 * The "RT overload" flag: it gets set if a CPU has more than
506 * one runnable RT task.
511 struct cpupri cpupri;
516 * By default the system creates a single root-domain with all cpus as
517 * members (mimicking the global state we have today).
519 static struct root_domain def_root_domain;
524 * This is the main, per-CPU runqueue data structure.
526 * Locking rule: those places that want to lock multiple runqueues
527 * (such as the load balancing or the thread migration code), lock
528 * acquire operations must be ordered by ascending &runqueue.
535 * nr_running and cpu_load should be in the same cacheline because
536 * remote CPUs use both these fields when doing load calculation.
538 unsigned long nr_running;
539 #define CPU_LOAD_IDX_MAX 5
540 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
541 unsigned char idle_at_tick;
543 unsigned long last_tick_seen;
544 unsigned char in_nohz_recently;
546 /* capture load from *all* tasks on this cpu: */
547 struct load_weight load;
548 unsigned long nr_load_updates;
554 #ifdef CONFIG_FAIR_GROUP_SCHED
555 /* list of leaf cfs_rq on this cpu: */
556 struct list_head leaf_cfs_rq_list;
558 #ifdef CONFIG_RT_GROUP_SCHED
559 struct list_head leaf_rt_rq_list;
563 * This is part of a global counter where only the total sum
564 * over all CPUs matters. A task can increase this counter on
565 * one CPU and if it got migrated afterwards it may decrease
566 * it on another CPU. Always updated under the runqueue lock:
568 unsigned long nr_uninterruptible;
570 struct task_struct *curr, *idle;
571 unsigned long next_balance;
572 struct mm_struct *prev_mm;
579 struct root_domain *rd;
580 struct sched_domain *sd;
582 /* For active balancing */
585 /* cpu of this runqueue: */
589 unsigned long avg_load_per_task;
591 struct task_struct *migration_thread;
592 struct list_head migration_queue;
595 #ifdef CONFIG_SCHED_HRTICK
597 int hrtick_csd_pending;
598 struct call_single_data hrtick_csd;
600 struct hrtimer hrtick_timer;
603 #ifdef CONFIG_SCHEDSTATS
605 struct sched_info rq_sched_info;
606 unsigned long long rq_cpu_time;
607 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
609 /* sys_sched_yield() stats */
610 unsigned int yld_exp_empty;
611 unsigned int yld_act_empty;
612 unsigned int yld_both_empty;
613 unsigned int yld_count;
615 /* schedule() stats */
616 unsigned int sched_switch;
617 unsigned int sched_count;
618 unsigned int sched_goidle;
620 /* try_to_wake_up() stats */
621 unsigned int ttwu_count;
622 unsigned int ttwu_local;
625 unsigned int bkl_count;
629 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
631 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
633 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
636 static inline int cpu_of(struct rq *rq)
646 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
647 * See detach_destroy_domains: synchronize_sched for details.
649 * The domain tree of any CPU may only be accessed from within
650 * preempt-disabled sections.
652 #define for_each_domain(cpu, __sd) \
653 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
655 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
656 #define this_rq() (&__get_cpu_var(runqueues))
657 #define task_rq(p) cpu_rq(task_cpu(p))
658 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
660 static inline void update_rq_clock(struct rq *rq)
662 rq->clock = sched_clock_cpu(cpu_of(rq));
666 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
668 #ifdef CONFIG_SCHED_DEBUG
669 # define const_debug __read_mostly
671 # define const_debug static const
677 * Returns true if the current cpu runqueue is locked.
678 * This interface allows printk to be called with the runqueue lock
679 * held and know whether or not it is OK to wake up the klogd.
681 int runqueue_is_locked(void)
684 struct rq *rq = cpu_rq(cpu);
687 ret = spin_is_locked(&rq->lock);
693 * Debugging: various feature bits
696 #define SCHED_FEAT(name, enabled) \
697 __SCHED_FEAT_##name ,
700 #include "sched_features.h"
705 #define SCHED_FEAT(name, enabled) \
706 (1UL << __SCHED_FEAT_##name) * enabled |
708 const_debug unsigned int sysctl_sched_features =
709 #include "sched_features.h"
714 #ifdef CONFIG_SCHED_DEBUG
715 #define SCHED_FEAT(name, enabled) \
718 static __read_mostly char *sched_feat_names[] = {
719 #include "sched_features.h"
725 static int sched_feat_show(struct seq_file *m, void *v)
729 for (i = 0; sched_feat_names[i]; i++) {
730 if (!(sysctl_sched_features & (1UL << i)))
732 seq_printf(m, "%s ", sched_feat_names[i]);
740 sched_feat_write(struct file *filp, const char __user *ubuf,
741 size_t cnt, loff_t *ppos)
751 if (copy_from_user(&buf, ubuf, cnt))
756 if (strncmp(buf, "NO_", 3) == 0) {
761 for (i = 0; sched_feat_names[i]; i++) {
762 int len = strlen(sched_feat_names[i]);
764 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
766 sysctl_sched_features &= ~(1UL << i);
768 sysctl_sched_features |= (1UL << i);
773 if (!sched_feat_names[i])
781 static int sched_feat_open(struct inode *inode, struct file *filp)
783 return single_open(filp, sched_feat_show, NULL);
786 static struct file_operations sched_feat_fops = {
787 .open = sched_feat_open,
788 .write = sched_feat_write,
791 .release = single_release,
794 static __init int sched_init_debug(void)
796 debugfs_create_file("sched_features", 0644, NULL, NULL,
801 late_initcall(sched_init_debug);
805 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
808 * Number of tasks to iterate in a single balance run.
809 * Limited because this is done with IRQs disabled.
811 const_debug unsigned int sysctl_sched_nr_migrate = 32;
814 * ratelimit for updating the group shares.
817 unsigned int sysctl_sched_shares_ratelimit = 250000;
820 * Inject some fuzzyness into changing the per-cpu group shares
821 * this avoids remote rq-locks at the expense of fairness.
824 unsigned int sysctl_sched_shares_thresh = 4;
827 * period over which we measure -rt task cpu usage in us.
830 unsigned int sysctl_sched_rt_period = 1000000;
832 static __read_mostly int scheduler_running;
835 * part of the period that we allow rt tasks to run in us.
838 int sysctl_sched_rt_runtime = 950000;
840 static inline u64 global_rt_period(void)
842 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
845 static inline u64 global_rt_runtime(void)
847 if (sysctl_sched_rt_runtime < 0)
850 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
853 #ifndef prepare_arch_switch
854 # define prepare_arch_switch(next) do { } while (0)
856 #ifndef finish_arch_switch
857 # define finish_arch_switch(prev) do { } while (0)
860 static inline int task_current(struct rq *rq, struct task_struct *p)
862 return rq->curr == p;
865 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
866 static inline int task_running(struct rq *rq, struct task_struct *p)
868 return task_current(rq, p);
871 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
875 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
877 #ifdef CONFIG_DEBUG_SPINLOCK
878 /* this is a valid case when another task releases the spinlock */
879 rq->lock.owner = current;
882 * If we are tracking spinlock dependencies then we have to
883 * fix up the runqueue lock - which gets 'carried over' from
886 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
888 spin_unlock_irq(&rq->lock);
891 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
892 static inline int task_running(struct rq *rq, struct task_struct *p)
897 return task_current(rq, p);
901 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
905 * We can optimise this out completely for !SMP, because the
906 * SMP rebalancing from interrupt is the only thing that cares
911 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
912 spin_unlock_irq(&rq->lock);
914 spin_unlock(&rq->lock);
918 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
922 * After ->oncpu is cleared, the task can be moved to a different CPU.
923 * We must ensure this doesn't happen until the switch is completely
929 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
933 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
936 * __task_rq_lock - lock the runqueue a given task resides on.
937 * Must be called interrupts disabled.
939 static inline struct rq *__task_rq_lock(struct task_struct *p)
943 struct rq *rq = task_rq(p);
944 spin_lock(&rq->lock);
945 if (likely(rq == task_rq(p)))
947 spin_unlock(&rq->lock);
952 * task_rq_lock - lock the runqueue a given task resides on and disable
953 * interrupts. Note the ordering: we can safely lookup the task_rq without
954 * explicitly disabling preemption.
956 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
962 local_irq_save(*flags);
964 spin_lock(&rq->lock);
965 if (likely(rq == task_rq(p)))
967 spin_unlock_irqrestore(&rq->lock, *flags);
971 void task_rq_unlock_wait(struct task_struct *p)
973 struct rq *rq = task_rq(p);
975 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
976 spin_unlock_wait(&rq->lock);
979 static void __task_rq_unlock(struct rq *rq)
982 spin_unlock(&rq->lock);
985 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
988 spin_unlock_irqrestore(&rq->lock, *flags);
992 * this_rq_lock - lock this runqueue and disable interrupts.
994 static struct rq *this_rq_lock(void)
1001 spin_lock(&rq->lock);
1006 #ifdef CONFIG_SCHED_HRTICK
1008 * Use HR-timers to deliver accurate preemption points.
1010 * Its all a bit involved since we cannot program an hrt while holding the
1011 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1014 * When we get rescheduled we reprogram the hrtick_timer outside of the
1020 * - enabled by features
1021 * - hrtimer is actually high res
1023 static inline int hrtick_enabled(struct rq *rq)
1025 if (!sched_feat(HRTICK))
1027 if (!cpu_active(cpu_of(rq)))
1029 return hrtimer_is_hres_active(&rq->hrtick_timer);
1032 static void hrtick_clear(struct rq *rq)
1034 if (hrtimer_active(&rq->hrtick_timer))
1035 hrtimer_cancel(&rq->hrtick_timer);
1039 * High-resolution timer tick.
1040 * Runs from hardirq context with interrupts disabled.
1042 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1044 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1046 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1048 spin_lock(&rq->lock);
1049 update_rq_clock(rq);
1050 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1051 spin_unlock(&rq->lock);
1053 return HRTIMER_NORESTART;
1058 * called from hardirq (IPI) context
1060 static void __hrtick_start(void *arg)
1062 struct rq *rq = arg;
1064 spin_lock(&rq->lock);
1065 hrtimer_restart(&rq->hrtick_timer);
1066 rq->hrtick_csd_pending = 0;
1067 spin_unlock(&rq->lock);
1071 * Called to set the hrtick timer state.
1073 * called with rq->lock held and irqs disabled
1075 static void hrtick_start(struct rq *rq, u64 delay)
1077 struct hrtimer *timer = &rq->hrtick_timer;
1078 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1080 hrtimer_set_expires(timer, time);
1082 if (rq == this_rq()) {
1083 hrtimer_restart(timer);
1084 } else if (!rq->hrtick_csd_pending) {
1085 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1086 rq->hrtick_csd_pending = 1;
1091 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1093 int cpu = (int)(long)hcpu;
1096 case CPU_UP_CANCELED:
1097 case CPU_UP_CANCELED_FROZEN:
1098 case CPU_DOWN_PREPARE:
1099 case CPU_DOWN_PREPARE_FROZEN:
1101 case CPU_DEAD_FROZEN:
1102 hrtick_clear(cpu_rq(cpu));
1109 static __init void init_hrtick(void)
1111 hotcpu_notifier(hotplug_hrtick, 0);
1115 * Called to set the hrtick timer state.
1117 * called with rq->lock held and irqs disabled
1119 static void hrtick_start(struct rq *rq, u64 delay)
1121 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1124 static inline void init_hrtick(void)
1127 #endif /* CONFIG_SMP */
1129 static void init_rq_hrtick(struct rq *rq)
1132 rq->hrtick_csd_pending = 0;
1134 rq->hrtick_csd.flags = 0;
1135 rq->hrtick_csd.func = __hrtick_start;
1136 rq->hrtick_csd.info = rq;
1139 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1140 rq->hrtick_timer.function = hrtick;
1142 #else /* CONFIG_SCHED_HRTICK */
1143 static inline void hrtick_clear(struct rq *rq)
1147 static inline void init_rq_hrtick(struct rq *rq)
1151 static inline void init_hrtick(void)
1154 #endif /* CONFIG_SCHED_HRTICK */
1157 * resched_task - mark a task 'to be rescheduled now'.
1159 * On UP this means the setting of the need_resched flag, on SMP it
1160 * might also involve a cross-CPU call to trigger the scheduler on
1165 #ifndef tsk_is_polling
1166 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1169 static void resched_task(struct task_struct *p)
1173 assert_spin_locked(&task_rq(p)->lock);
1175 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1178 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1181 if (cpu == smp_processor_id())
1184 /* NEED_RESCHED must be visible before we test polling */
1186 if (!tsk_is_polling(p))
1187 smp_send_reschedule(cpu);
1190 static void resched_cpu(int cpu)
1192 struct rq *rq = cpu_rq(cpu);
1193 unsigned long flags;
1195 if (!spin_trylock_irqsave(&rq->lock, flags))
1197 resched_task(cpu_curr(cpu));
1198 spin_unlock_irqrestore(&rq->lock, flags);
1203 * When add_timer_on() enqueues a timer into the timer wheel of an
1204 * idle CPU then this timer might expire before the next timer event
1205 * which is scheduled to wake up that CPU. In case of a completely
1206 * idle system the next event might even be infinite time into the
1207 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1208 * leaves the inner idle loop so the newly added timer is taken into
1209 * account when the CPU goes back to idle and evaluates the timer
1210 * wheel for the next timer event.
1212 void wake_up_idle_cpu(int cpu)
1214 struct rq *rq = cpu_rq(cpu);
1216 if (cpu == smp_processor_id())
1220 * This is safe, as this function is called with the timer
1221 * wheel base lock of (cpu) held. When the CPU is on the way
1222 * to idle and has not yet set rq->curr to idle then it will
1223 * be serialized on the timer wheel base lock and take the new
1224 * timer into account automatically.
1226 if (rq->curr != rq->idle)
1230 * We can set TIF_RESCHED on the idle task of the other CPU
1231 * lockless. The worst case is that the other CPU runs the
1232 * idle task through an additional NOOP schedule()
1234 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1236 /* NEED_RESCHED must be visible before we test polling */
1238 if (!tsk_is_polling(rq->idle))
1239 smp_send_reschedule(cpu);
1241 #endif /* CONFIG_NO_HZ */
1243 #else /* !CONFIG_SMP */
1244 static void resched_task(struct task_struct *p)
1246 assert_spin_locked(&task_rq(p)->lock);
1247 set_tsk_need_resched(p);
1249 #endif /* CONFIG_SMP */
1251 #if BITS_PER_LONG == 32
1252 # define WMULT_CONST (~0UL)
1254 # define WMULT_CONST (1UL << 32)
1257 #define WMULT_SHIFT 32
1260 * Shift right and round:
1262 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1265 * delta *= weight / lw
1267 static unsigned long
1268 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1269 struct load_weight *lw)
1273 if (!lw->inv_weight) {
1274 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1277 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1281 tmp = (u64)delta_exec * weight;
1283 * Check whether we'd overflow the 64-bit multiplication:
1285 if (unlikely(tmp > WMULT_CONST))
1286 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1289 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1291 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1294 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1300 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1307 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1308 * of tasks with abnormal "nice" values across CPUs the contribution that
1309 * each task makes to its run queue's load is weighted according to its
1310 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1311 * scaled version of the new time slice allocation that they receive on time
1315 #define WEIGHT_IDLEPRIO 2
1316 #define WMULT_IDLEPRIO (1 << 31)
1319 * Nice levels are multiplicative, with a gentle 10% change for every
1320 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1321 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1322 * that remained on nice 0.
1324 * The "10% effect" is relative and cumulative: from _any_ nice level,
1325 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1326 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1327 * If a task goes up by ~10% and another task goes down by ~10% then
1328 * the relative distance between them is ~25%.)
1330 static const int prio_to_weight[40] = {
1331 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1332 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1333 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1334 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1335 /* 0 */ 1024, 820, 655, 526, 423,
1336 /* 5 */ 335, 272, 215, 172, 137,
1337 /* 10 */ 110, 87, 70, 56, 45,
1338 /* 15 */ 36, 29, 23, 18, 15,
1342 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1344 * In cases where the weight does not change often, we can use the
1345 * precalculated inverse to speed up arithmetics by turning divisions
1346 * into multiplications:
1348 static const u32 prio_to_wmult[40] = {
1349 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1350 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1351 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1352 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1353 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1354 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1355 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1356 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1359 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1362 * runqueue iterator, to support SMP load-balancing between different
1363 * scheduling classes, without having to expose their internal data
1364 * structures to the load-balancing proper:
1366 struct rq_iterator {
1368 struct task_struct *(*start)(void *);
1369 struct task_struct *(*next)(void *);
1373 static unsigned long
1374 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1375 unsigned long max_load_move, struct sched_domain *sd,
1376 enum cpu_idle_type idle, int *all_pinned,
1377 int *this_best_prio, struct rq_iterator *iterator);
1380 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1381 struct sched_domain *sd, enum cpu_idle_type idle,
1382 struct rq_iterator *iterator);
1385 #ifdef CONFIG_CGROUP_CPUACCT
1386 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1388 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1391 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1393 update_load_add(&rq->load, load);
1396 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1398 update_load_sub(&rq->load, load);
1401 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1402 typedef int (*tg_visitor)(struct task_group *, void *);
1405 * Iterate the full tree, calling @down when first entering a node and @up when
1406 * leaving it for the final time.
1408 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1410 struct task_group *parent, *child;
1414 parent = &root_task_group;
1416 ret = (*down)(parent, data);
1419 list_for_each_entry_rcu(child, &parent->children, siblings) {
1426 ret = (*up)(parent, data);
1431 parent = parent->parent;
1440 static int tg_nop(struct task_group *tg, void *data)
1447 static unsigned long source_load(int cpu, int type);
1448 static unsigned long target_load(int cpu, int type);
1449 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1451 static unsigned long cpu_avg_load_per_task(int cpu)
1453 struct rq *rq = cpu_rq(cpu);
1454 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1457 rq->avg_load_per_task = rq->load.weight / nr_running;
1459 rq->avg_load_per_task = 0;
1461 return rq->avg_load_per_task;
1464 #ifdef CONFIG_FAIR_GROUP_SCHED
1466 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1469 * Calculate and set the cpu's group shares.
1472 update_group_shares_cpu(struct task_group *tg, int cpu,
1473 unsigned long sd_shares, unsigned long sd_rq_weight)
1475 unsigned long shares;
1476 unsigned long rq_weight;
1481 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1484 * \Sum shares * rq_weight
1485 * shares = -----------------------
1489 shares = (sd_shares * rq_weight) / sd_rq_weight;
1490 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1492 if (abs(shares - tg->se[cpu]->load.weight) >
1493 sysctl_sched_shares_thresh) {
1494 struct rq *rq = cpu_rq(cpu);
1495 unsigned long flags;
1497 spin_lock_irqsave(&rq->lock, flags);
1498 tg->cfs_rq[cpu]->shares = shares;
1500 __set_se_shares(tg->se[cpu], shares);
1501 spin_unlock_irqrestore(&rq->lock, flags);
1506 * Re-compute the task group their per cpu shares over the given domain.
1507 * This needs to be done in a bottom-up fashion because the rq weight of a
1508 * parent group depends on the shares of its child groups.
1510 static int tg_shares_up(struct task_group *tg, void *data)
1512 unsigned long weight, rq_weight = 0;
1513 unsigned long shares = 0;
1514 struct sched_domain *sd = data;
1517 for_each_cpu_mask(i, sd->span) {
1519 * If there are currently no tasks on the cpu pretend there
1520 * is one of average load so that when a new task gets to
1521 * run here it will not get delayed by group starvation.
1523 weight = tg->cfs_rq[i]->load.weight;
1525 weight = NICE_0_LOAD;
1527 tg->cfs_rq[i]->rq_weight = weight;
1528 rq_weight += weight;
1529 shares += tg->cfs_rq[i]->shares;
1532 if ((!shares && rq_weight) || shares > tg->shares)
1533 shares = tg->shares;
1535 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1536 shares = tg->shares;
1538 for_each_cpu_mask(i, sd->span)
1539 update_group_shares_cpu(tg, i, shares, rq_weight);
1545 * Compute the cpu's hierarchical load factor for each task group.
1546 * This needs to be done in a top-down fashion because the load of a child
1547 * group is a fraction of its parents load.
1549 static int tg_load_down(struct task_group *tg, void *data)
1552 long cpu = (long)data;
1555 load = cpu_rq(cpu)->load.weight;
1557 load = tg->parent->cfs_rq[cpu]->h_load;
1558 load *= tg->cfs_rq[cpu]->shares;
1559 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1562 tg->cfs_rq[cpu]->h_load = load;
1567 static void update_shares(struct sched_domain *sd)
1569 u64 now = cpu_clock(raw_smp_processor_id());
1570 s64 elapsed = now - sd->last_update;
1572 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1573 sd->last_update = now;
1574 walk_tg_tree(tg_nop, tg_shares_up, sd);
1578 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1580 spin_unlock(&rq->lock);
1582 spin_lock(&rq->lock);
1585 static void update_h_load(long cpu)
1587 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1592 static inline void update_shares(struct sched_domain *sd)
1596 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1603 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1605 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1606 __releases(this_rq->lock)
1607 __acquires(busiest->lock)
1608 __acquires(this_rq->lock)
1612 if (unlikely(!irqs_disabled())) {
1613 /* printk() doesn't work good under rq->lock */
1614 spin_unlock(&this_rq->lock);
1617 if (unlikely(!spin_trylock(&busiest->lock))) {
1618 if (busiest < this_rq) {
1619 spin_unlock(&this_rq->lock);
1620 spin_lock(&busiest->lock);
1621 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1624 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1629 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1630 __releases(busiest->lock)
1632 spin_unlock(&busiest->lock);
1633 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1637 #ifdef CONFIG_FAIR_GROUP_SCHED
1638 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1641 cfs_rq->shares = shares;
1646 #include "sched_stats.h"
1647 #include "sched_idletask.c"
1648 #include "sched_fair.c"
1649 #include "sched_rt.c"
1650 #ifdef CONFIG_SCHED_DEBUG
1651 # include "sched_debug.c"
1654 #define sched_class_highest (&rt_sched_class)
1655 #define for_each_class(class) \
1656 for (class = sched_class_highest; class; class = class->next)
1658 static void inc_nr_running(struct rq *rq)
1663 static void dec_nr_running(struct rq *rq)
1668 static void set_load_weight(struct task_struct *p)
1670 if (task_has_rt_policy(p)) {
1671 p->se.load.weight = prio_to_weight[0] * 2;
1672 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1677 * SCHED_IDLE tasks get minimal weight:
1679 if (p->policy == SCHED_IDLE) {
1680 p->se.load.weight = WEIGHT_IDLEPRIO;
1681 p->se.load.inv_weight = WMULT_IDLEPRIO;
1685 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1686 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1689 static void update_avg(u64 *avg, u64 sample)
1691 s64 diff = sample - *avg;
1695 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1697 sched_info_queued(p);
1698 p->sched_class->enqueue_task(rq, p, wakeup);
1702 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1704 if (sleep && p->se.last_wakeup) {
1705 update_avg(&p->se.avg_overlap,
1706 p->se.sum_exec_runtime - p->se.last_wakeup);
1707 p->se.last_wakeup = 0;
1710 sched_info_dequeued(p);
1711 p->sched_class->dequeue_task(rq, p, sleep);
1716 * __normal_prio - return the priority that is based on the static prio
1718 static inline int __normal_prio(struct task_struct *p)
1720 return p->static_prio;
1724 * Calculate the expected normal priority: i.e. priority
1725 * without taking RT-inheritance into account. Might be
1726 * boosted by interactivity modifiers. Changes upon fork,
1727 * setprio syscalls, and whenever the interactivity
1728 * estimator recalculates.
1730 static inline int normal_prio(struct task_struct *p)
1734 if (task_has_rt_policy(p))
1735 prio = MAX_RT_PRIO-1 - p->rt_priority;
1737 prio = __normal_prio(p);
1742 * Calculate the current priority, i.e. the priority
1743 * taken into account by the scheduler. This value might
1744 * be boosted by RT tasks, or might be boosted by
1745 * interactivity modifiers. Will be RT if the task got
1746 * RT-boosted. If not then it returns p->normal_prio.
1748 static int effective_prio(struct task_struct *p)
1750 p->normal_prio = normal_prio(p);
1752 * If we are RT tasks or we were boosted to RT priority,
1753 * keep the priority unchanged. Otherwise, update priority
1754 * to the normal priority:
1756 if (!rt_prio(p->prio))
1757 return p->normal_prio;
1762 * activate_task - move a task to the runqueue.
1764 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1766 if (task_contributes_to_load(p))
1767 rq->nr_uninterruptible--;
1769 enqueue_task(rq, p, wakeup);
1774 * deactivate_task - remove a task from the runqueue.
1776 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1778 if (task_contributes_to_load(p))
1779 rq->nr_uninterruptible++;
1781 dequeue_task(rq, p, sleep);
1786 * task_curr - is this task currently executing on a CPU?
1787 * @p: the task in question.
1789 inline int task_curr(const struct task_struct *p)
1791 return cpu_curr(task_cpu(p)) == p;
1794 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1796 set_task_rq(p, cpu);
1799 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1800 * successfuly executed on another CPU. We must ensure that updates of
1801 * per-task data have been completed by this moment.
1804 task_thread_info(p)->cpu = cpu;
1808 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1809 const struct sched_class *prev_class,
1810 int oldprio, int running)
1812 if (prev_class != p->sched_class) {
1813 if (prev_class->switched_from)
1814 prev_class->switched_from(rq, p, running);
1815 p->sched_class->switched_to(rq, p, running);
1817 p->sched_class->prio_changed(rq, p, oldprio, running);
1822 /* Used instead of source_load when we know the type == 0 */
1823 static unsigned long weighted_cpuload(const int cpu)
1825 return cpu_rq(cpu)->load.weight;
1829 * Is this task likely cache-hot:
1832 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1837 * Buddy candidates are cache hot:
1839 if (sched_feat(CACHE_HOT_BUDDY) &&
1840 (&p->se == cfs_rq_of(&p->se)->next ||
1841 &p->se == cfs_rq_of(&p->se)->last))
1844 if (p->sched_class != &fair_sched_class)
1847 if (sysctl_sched_migration_cost == -1)
1849 if (sysctl_sched_migration_cost == 0)
1852 delta = now - p->se.exec_start;
1854 return delta < (s64)sysctl_sched_migration_cost;
1858 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1860 int old_cpu = task_cpu(p);
1861 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1862 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1863 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1866 clock_offset = old_rq->clock - new_rq->clock;
1868 trace_sched_migrate_task(p, task_cpu(p), new_cpu);
1870 #ifdef CONFIG_SCHEDSTATS
1871 if (p->se.wait_start)
1872 p->se.wait_start -= clock_offset;
1873 if (p->se.sleep_start)
1874 p->se.sleep_start -= clock_offset;
1875 if (p->se.block_start)
1876 p->se.block_start -= clock_offset;
1877 if (old_cpu != new_cpu) {
1878 schedstat_inc(p, se.nr_migrations);
1879 if (task_hot(p, old_rq->clock, NULL))
1880 schedstat_inc(p, se.nr_forced2_migrations);
1883 p->se.vruntime -= old_cfsrq->min_vruntime -
1884 new_cfsrq->min_vruntime;
1886 __set_task_cpu(p, new_cpu);
1889 struct migration_req {
1890 struct list_head list;
1892 struct task_struct *task;
1895 struct completion done;
1899 * The task's runqueue lock must be held.
1900 * Returns true if you have to wait for migration thread.
1903 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1905 struct rq *rq = task_rq(p);
1908 * If the task is not on a runqueue (and not running), then
1909 * it is sufficient to simply update the task's cpu field.
1911 if (!p->se.on_rq && !task_running(rq, p)) {
1912 set_task_cpu(p, dest_cpu);
1916 init_completion(&req->done);
1918 req->dest_cpu = dest_cpu;
1919 list_add(&req->list, &rq->migration_queue);
1925 * wait_task_inactive - wait for a thread to unschedule.
1927 * If @match_state is nonzero, it's the @p->state value just checked and
1928 * not expected to change. If it changes, i.e. @p might have woken up,
1929 * then return zero. When we succeed in waiting for @p to be off its CPU,
1930 * we return a positive number (its total switch count). If a second call
1931 * a short while later returns the same number, the caller can be sure that
1932 * @p has remained unscheduled the whole time.
1934 * The caller must ensure that the task *will* unschedule sometime soon,
1935 * else this function might spin for a *long* time. This function can't
1936 * be called with interrupts off, or it may introduce deadlock with
1937 * smp_call_function() if an IPI is sent by the same process we are
1938 * waiting to become inactive.
1940 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1942 unsigned long flags;
1949 * We do the initial early heuristics without holding
1950 * any task-queue locks at all. We'll only try to get
1951 * the runqueue lock when things look like they will
1957 * If the task is actively running on another CPU
1958 * still, just relax and busy-wait without holding
1961 * NOTE! Since we don't hold any locks, it's not
1962 * even sure that "rq" stays as the right runqueue!
1963 * But we don't care, since "task_running()" will
1964 * return false if the runqueue has changed and p
1965 * is actually now running somewhere else!
1967 while (task_running(rq, p)) {
1968 if (match_state && unlikely(p->state != match_state))
1974 * Ok, time to look more closely! We need the rq
1975 * lock now, to be *sure*. If we're wrong, we'll
1976 * just go back and repeat.
1978 rq = task_rq_lock(p, &flags);
1979 trace_sched_wait_task(rq, p);
1980 running = task_running(rq, p);
1981 on_rq = p->se.on_rq;
1983 if (!match_state || p->state == match_state)
1984 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1985 task_rq_unlock(rq, &flags);
1988 * If it changed from the expected state, bail out now.
1990 if (unlikely(!ncsw))
1994 * Was it really running after all now that we
1995 * checked with the proper locks actually held?
1997 * Oops. Go back and try again..
1999 if (unlikely(running)) {
2005 * It's not enough that it's not actively running,
2006 * it must be off the runqueue _entirely_, and not
2009 * So if it wa still runnable (but just not actively
2010 * running right now), it's preempted, and we should
2011 * yield - it could be a while.
2013 if (unlikely(on_rq)) {
2014 schedule_timeout_uninterruptible(1);
2019 * Ahh, all good. It wasn't running, and it wasn't
2020 * runnable, which means that it will never become
2021 * running in the future either. We're all done!
2030 * kick_process - kick a running thread to enter/exit the kernel
2031 * @p: the to-be-kicked thread
2033 * Cause a process which is running on another CPU to enter
2034 * kernel-mode, without any delay. (to get signals handled.)
2036 * NOTE: this function doesnt have to take the runqueue lock,
2037 * because all it wants to ensure is that the remote task enters
2038 * the kernel. If the IPI races and the task has been migrated
2039 * to another CPU then no harm is done and the purpose has been
2042 void kick_process(struct task_struct *p)
2048 if ((cpu != smp_processor_id()) && task_curr(p))
2049 smp_send_reschedule(cpu);
2054 * Return a low guess at the load of a migration-source cpu weighted
2055 * according to the scheduling class and "nice" value.
2057 * We want to under-estimate the load of migration sources, to
2058 * balance conservatively.
2060 static unsigned long source_load(int cpu, int type)
2062 struct rq *rq = cpu_rq(cpu);
2063 unsigned long total = weighted_cpuload(cpu);
2065 if (type == 0 || !sched_feat(LB_BIAS))
2068 return min(rq->cpu_load[type-1], total);
2072 * Return a high guess at the load of a migration-target cpu weighted
2073 * according to the scheduling class and "nice" value.
2075 static unsigned long target_load(int cpu, int type)
2077 struct rq *rq = cpu_rq(cpu);
2078 unsigned long total = weighted_cpuload(cpu);
2080 if (type == 0 || !sched_feat(LB_BIAS))
2083 return max(rq->cpu_load[type-1], total);
2087 * find_idlest_group finds and returns the least busy CPU group within the
2090 static struct sched_group *
2091 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2093 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2094 unsigned long min_load = ULONG_MAX, this_load = 0;
2095 int load_idx = sd->forkexec_idx;
2096 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2099 unsigned long load, avg_load;
2103 /* Skip over this group if it has no CPUs allowed */
2104 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2107 local_group = cpu_isset(this_cpu, group->cpumask);
2109 /* Tally up the load of all CPUs in the group */
2112 for_each_cpu_mask_nr(i, group->cpumask) {
2113 /* Bias balancing toward cpus of our domain */
2115 load = source_load(i, load_idx);
2117 load = target_load(i, load_idx);
2122 /* Adjust by relative CPU power of the group */
2123 avg_load = sg_div_cpu_power(group,
2124 avg_load * SCHED_LOAD_SCALE);
2127 this_load = avg_load;
2129 } else if (avg_load < min_load) {
2130 min_load = avg_load;
2133 } while (group = group->next, group != sd->groups);
2135 if (!idlest || 100*this_load < imbalance*min_load)
2141 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2144 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2147 unsigned long load, min_load = ULONG_MAX;
2151 /* Traverse only the allowed CPUs */
2152 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2154 for_each_cpu_mask_nr(i, *tmp) {
2155 load = weighted_cpuload(i);
2157 if (load < min_load || (load == min_load && i == this_cpu)) {
2167 * sched_balance_self: balance the current task (running on cpu) in domains
2168 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2171 * Balance, ie. select the least loaded group.
2173 * Returns the target CPU number, or the same CPU if no balancing is needed.
2175 * preempt must be disabled.
2177 static int sched_balance_self(int cpu, int flag)
2179 struct task_struct *t = current;
2180 struct sched_domain *tmp, *sd = NULL;
2182 for_each_domain(cpu, tmp) {
2184 * If power savings logic is enabled for a domain, stop there.
2186 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2188 if (tmp->flags & flag)
2196 cpumask_t span, tmpmask;
2197 struct sched_group *group;
2198 int new_cpu, weight;
2200 if (!(sd->flags & flag)) {
2206 group = find_idlest_group(sd, t, cpu);
2212 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2213 if (new_cpu == -1 || new_cpu == cpu) {
2214 /* Now try balancing at a lower domain level of cpu */
2219 /* Now try balancing at a lower domain level of new_cpu */
2222 weight = cpus_weight(span);
2223 for_each_domain(cpu, tmp) {
2224 if (weight <= cpus_weight(tmp->span))
2226 if (tmp->flags & flag)
2229 /* while loop will break here if sd == NULL */
2235 #endif /* CONFIG_SMP */
2238 * try_to_wake_up - wake up a thread
2239 * @p: the to-be-woken-up thread
2240 * @state: the mask of task states that can be woken
2241 * @sync: do a synchronous wakeup?
2243 * Put it on the run-queue if it's not already there. The "current"
2244 * thread is always on the run-queue (except when the actual
2245 * re-schedule is in progress), and as such you're allowed to do
2246 * the simpler "current->state = TASK_RUNNING" to mark yourself
2247 * runnable without the overhead of this.
2249 * returns failure only if the task is already active.
2251 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2253 int cpu, orig_cpu, this_cpu, success = 0;
2254 unsigned long flags;
2258 if (!sched_feat(SYNC_WAKEUPS))
2262 if (sched_feat(LB_WAKEUP_UPDATE)) {
2263 struct sched_domain *sd;
2265 this_cpu = raw_smp_processor_id();
2268 for_each_domain(this_cpu, sd) {
2269 if (cpu_isset(cpu, sd->span)) {
2278 rq = task_rq_lock(p, &flags);
2279 update_rq_clock(rq);
2280 old_state = p->state;
2281 if (!(old_state & state))
2289 this_cpu = smp_processor_id();
2292 if (unlikely(task_running(rq, p)))
2295 cpu = p->sched_class->select_task_rq(p, sync);
2296 if (cpu != orig_cpu) {
2297 set_task_cpu(p, cpu);
2298 task_rq_unlock(rq, &flags);
2299 /* might preempt at this point */
2300 rq = task_rq_lock(p, &flags);
2301 old_state = p->state;
2302 if (!(old_state & state))
2307 this_cpu = smp_processor_id();
2311 #ifdef CONFIG_SCHEDSTATS
2312 schedstat_inc(rq, ttwu_count);
2313 if (cpu == this_cpu)
2314 schedstat_inc(rq, ttwu_local);
2316 struct sched_domain *sd;
2317 for_each_domain(this_cpu, sd) {
2318 if (cpu_isset(cpu, sd->span)) {
2319 schedstat_inc(sd, ttwu_wake_remote);
2324 #endif /* CONFIG_SCHEDSTATS */
2327 #endif /* CONFIG_SMP */
2328 schedstat_inc(p, se.nr_wakeups);
2330 schedstat_inc(p, se.nr_wakeups_sync);
2331 if (orig_cpu != cpu)
2332 schedstat_inc(p, se.nr_wakeups_migrate);
2333 if (cpu == this_cpu)
2334 schedstat_inc(p, se.nr_wakeups_local);
2336 schedstat_inc(p, se.nr_wakeups_remote);
2337 activate_task(rq, p, 1);
2341 trace_sched_wakeup(rq, p, success);
2342 check_preempt_curr(rq, p, sync);
2344 p->state = TASK_RUNNING;
2346 if (p->sched_class->task_wake_up)
2347 p->sched_class->task_wake_up(rq, p);
2350 current->se.last_wakeup = current->se.sum_exec_runtime;
2352 task_rq_unlock(rq, &flags);
2357 int wake_up_process(struct task_struct *p)
2359 return try_to_wake_up(p, TASK_ALL, 0);
2361 EXPORT_SYMBOL(wake_up_process);
2363 int wake_up_state(struct task_struct *p, unsigned int state)
2365 return try_to_wake_up(p, state, 0);
2369 * Perform scheduler related setup for a newly forked process p.
2370 * p is forked by current.
2372 * __sched_fork() is basic setup used by init_idle() too:
2374 static void __sched_fork(struct task_struct *p)
2376 p->se.exec_start = 0;
2377 p->se.sum_exec_runtime = 0;
2378 p->se.prev_sum_exec_runtime = 0;
2379 p->se.last_wakeup = 0;
2380 p->se.avg_overlap = 0;
2382 #ifdef CONFIG_SCHEDSTATS
2383 p->se.wait_start = 0;
2384 p->se.sum_sleep_runtime = 0;
2385 p->se.sleep_start = 0;
2386 p->se.block_start = 0;
2387 p->se.sleep_max = 0;
2388 p->se.block_max = 0;
2390 p->se.slice_max = 0;
2394 INIT_LIST_HEAD(&p->rt.run_list);
2396 INIT_LIST_HEAD(&p->se.group_node);
2398 #ifdef CONFIG_PREEMPT_NOTIFIERS
2399 INIT_HLIST_HEAD(&p->preempt_notifiers);
2403 * We mark the process as running here, but have not actually
2404 * inserted it onto the runqueue yet. This guarantees that
2405 * nobody will actually run it, and a signal or other external
2406 * event cannot wake it up and insert it on the runqueue either.
2408 p->state = TASK_RUNNING;
2412 * fork()/clone()-time setup:
2414 void sched_fork(struct task_struct *p, int clone_flags)
2416 int cpu = get_cpu();
2421 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2423 set_task_cpu(p, cpu);
2426 * Make sure we do not leak PI boosting priority to the child:
2428 p->prio = current->normal_prio;
2429 if (!rt_prio(p->prio))
2430 p->sched_class = &fair_sched_class;
2432 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2433 if (likely(sched_info_on()))
2434 memset(&p->sched_info, 0, sizeof(p->sched_info));
2436 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2439 #ifdef CONFIG_PREEMPT
2440 /* Want to start with kernel preemption disabled. */
2441 task_thread_info(p)->preempt_count = 1;
2447 * wake_up_new_task - wake up a newly created task for the first time.
2449 * This function will do some initial scheduler statistics housekeeping
2450 * that must be done for every newly created context, then puts the task
2451 * on the runqueue and wakes it.
2453 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2455 unsigned long flags;
2458 rq = task_rq_lock(p, &flags);
2459 BUG_ON(p->state != TASK_RUNNING);
2460 update_rq_clock(rq);
2462 p->prio = effective_prio(p);
2464 if (!p->sched_class->task_new || !current->se.on_rq) {
2465 activate_task(rq, p, 0);
2468 * Let the scheduling class do new task startup
2469 * management (if any):
2471 p->sched_class->task_new(rq, p);
2474 trace_sched_wakeup_new(rq, p, 1);
2475 check_preempt_curr(rq, p, 0);
2477 if (p->sched_class->task_wake_up)
2478 p->sched_class->task_wake_up(rq, p);
2480 task_rq_unlock(rq, &flags);
2483 #ifdef CONFIG_PREEMPT_NOTIFIERS
2486 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2487 * @notifier: notifier struct to register
2489 void preempt_notifier_register(struct preempt_notifier *notifier)
2491 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2493 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2496 * preempt_notifier_unregister - no longer interested in preemption notifications
2497 * @notifier: notifier struct to unregister
2499 * This is safe to call from within a preemption notifier.
2501 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2503 hlist_del(¬ifier->link);
2505 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2507 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2509 struct preempt_notifier *notifier;
2510 struct hlist_node *node;
2512 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2513 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2517 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2518 struct task_struct *next)
2520 struct preempt_notifier *notifier;
2521 struct hlist_node *node;
2523 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2524 notifier->ops->sched_out(notifier, next);
2527 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2529 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2534 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2535 struct task_struct *next)
2539 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2542 * prepare_task_switch - prepare to switch tasks
2543 * @rq: the runqueue preparing to switch
2544 * @prev: the current task that is being switched out
2545 * @next: the task we are going to switch to.
2547 * This is called with the rq lock held and interrupts off. It must
2548 * be paired with a subsequent finish_task_switch after the context
2551 * prepare_task_switch sets up locking and calls architecture specific
2555 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2556 struct task_struct *next)
2558 fire_sched_out_preempt_notifiers(prev, next);
2559 prepare_lock_switch(rq, next);
2560 prepare_arch_switch(next);
2564 * finish_task_switch - clean up after a task-switch
2565 * @rq: runqueue associated with task-switch
2566 * @prev: the thread we just switched away from.
2568 * finish_task_switch must be called after the context switch, paired
2569 * with a prepare_task_switch call before the context switch.
2570 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2571 * and do any other architecture-specific cleanup actions.
2573 * Note that we may have delayed dropping an mm in context_switch(). If
2574 * so, we finish that here outside of the runqueue lock. (Doing it
2575 * with the lock held can cause deadlocks; see schedule() for
2578 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2579 __releases(rq->lock)
2581 struct mm_struct *mm = rq->prev_mm;
2587 * A task struct has one reference for the use as "current".
2588 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2589 * schedule one last time. The schedule call will never return, and
2590 * the scheduled task must drop that reference.
2591 * The test for TASK_DEAD must occur while the runqueue locks are
2592 * still held, otherwise prev could be scheduled on another cpu, die
2593 * there before we look at prev->state, and then the reference would
2595 * Manfred Spraul <manfred@colorfullife.com>
2597 prev_state = prev->state;
2598 finish_arch_switch(prev);
2599 finish_lock_switch(rq, prev);
2601 if (current->sched_class->post_schedule)
2602 current->sched_class->post_schedule(rq);
2605 fire_sched_in_preempt_notifiers(current);
2608 if (unlikely(prev_state == TASK_DEAD)) {
2610 * Remove function-return probe instances associated with this
2611 * task and put them back on the free list.
2613 kprobe_flush_task(prev);
2614 put_task_struct(prev);
2619 * schedule_tail - first thing a freshly forked thread must call.
2620 * @prev: the thread we just switched away from.
2622 asmlinkage void schedule_tail(struct task_struct *prev)
2623 __releases(rq->lock)
2625 struct rq *rq = this_rq();
2627 finish_task_switch(rq, prev);
2628 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2629 /* In this case, finish_task_switch does not reenable preemption */
2632 if (current->set_child_tid)
2633 put_user(task_pid_vnr(current), current->set_child_tid);
2637 * context_switch - switch to the new MM and the new
2638 * thread's register state.
2641 context_switch(struct rq *rq, struct task_struct *prev,
2642 struct task_struct *next)
2644 struct mm_struct *mm, *oldmm;
2646 prepare_task_switch(rq, prev, next);
2647 trace_sched_switch(rq, prev, next);
2649 oldmm = prev->active_mm;
2651 * For paravirt, this is coupled with an exit in switch_to to
2652 * combine the page table reload and the switch backend into
2655 arch_enter_lazy_cpu_mode();
2657 if (unlikely(!mm)) {
2658 next->active_mm = oldmm;
2659 atomic_inc(&oldmm->mm_count);
2660 enter_lazy_tlb(oldmm, next);
2662 switch_mm(oldmm, mm, next);
2664 if (unlikely(!prev->mm)) {
2665 prev->active_mm = NULL;
2666 rq->prev_mm = oldmm;
2669 * Since the runqueue lock will be released by the next
2670 * task (which is an invalid locking op but in the case
2671 * of the scheduler it's an obvious special-case), so we
2672 * do an early lockdep release here:
2674 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2675 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2678 /* Here we just switch the register state and the stack. */
2679 switch_to(prev, next, prev);
2683 * this_rq must be evaluated again because prev may have moved
2684 * CPUs since it called schedule(), thus the 'rq' on its stack
2685 * frame will be invalid.
2687 finish_task_switch(this_rq(), prev);
2691 * nr_running, nr_uninterruptible and nr_context_switches:
2693 * externally visible scheduler statistics: current number of runnable
2694 * threads, current number of uninterruptible-sleeping threads, total
2695 * number of context switches performed since bootup.
2697 unsigned long nr_running(void)
2699 unsigned long i, sum = 0;
2701 for_each_online_cpu(i)
2702 sum += cpu_rq(i)->nr_running;
2707 unsigned long nr_uninterruptible(void)
2709 unsigned long i, sum = 0;
2711 for_each_possible_cpu(i)
2712 sum += cpu_rq(i)->nr_uninterruptible;
2715 * Since we read the counters lockless, it might be slightly
2716 * inaccurate. Do not allow it to go below zero though:
2718 if (unlikely((long)sum < 0))
2724 unsigned long long nr_context_switches(void)
2727 unsigned long long sum = 0;
2729 for_each_possible_cpu(i)
2730 sum += cpu_rq(i)->nr_switches;
2735 unsigned long nr_iowait(void)
2737 unsigned long i, sum = 0;
2739 for_each_possible_cpu(i)
2740 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2745 unsigned long nr_active(void)
2747 unsigned long i, running = 0, uninterruptible = 0;
2749 for_each_online_cpu(i) {
2750 running += cpu_rq(i)->nr_running;
2751 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2754 if (unlikely((long)uninterruptible < 0))
2755 uninterruptible = 0;
2757 return running + uninterruptible;
2761 * Update rq->cpu_load[] statistics. This function is usually called every
2762 * scheduler tick (TICK_NSEC).
2764 static void update_cpu_load(struct rq *this_rq)
2766 unsigned long this_load = this_rq->load.weight;
2769 this_rq->nr_load_updates++;
2771 /* Update our load: */
2772 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2773 unsigned long old_load, new_load;
2775 /* scale is effectively 1 << i now, and >> i divides by scale */
2777 old_load = this_rq->cpu_load[i];
2778 new_load = this_load;
2780 * Round up the averaging division if load is increasing. This
2781 * prevents us from getting stuck on 9 if the load is 10, for
2784 if (new_load > old_load)
2785 new_load += scale-1;
2786 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2793 * double_rq_lock - safely lock two runqueues
2795 * Note this does not disable interrupts like task_rq_lock,
2796 * you need to do so manually before calling.
2798 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2799 __acquires(rq1->lock)
2800 __acquires(rq2->lock)
2802 BUG_ON(!irqs_disabled());
2804 spin_lock(&rq1->lock);
2805 __acquire(rq2->lock); /* Fake it out ;) */
2808 spin_lock(&rq1->lock);
2809 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2811 spin_lock(&rq2->lock);
2812 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2815 update_rq_clock(rq1);
2816 update_rq_clock(rq2);
2820 * double_rq_unlock - safely unlock two runqueues
2822 * Note this does not restore interrupts like task_rq_unlock,
2823 * you need to do so manually after calling.
2825 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2826 __releases(rq1->lock)
2827 __releases(rq2->lock)
2829 spin_unlock(&rq1->lock);
2831 spin_unlock(&rq2->lock);
2833 __release(rq2->lock);
2837 * If dest_cpu is allowed for this process, migrate the task to it.
2838 * This is accomplished by forcing the cpu_allowed mask to only
2839 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2840 * the cpu_allowed mask is restored.
2842 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2844 struct migration_req req;
2845 unsigned long flags;
2848 rq = task_rq_lock(p, &flags);
2849 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2850 || unlikely(!cpu_active(dest_cpu)))
2853 /* force the process onto the specified CPU */
2854 if (migrate_task(p, dest_cpu, &req)) {
2855 /* Need to wait for migration thread (might exit: take ref). */
2856 struct task_struct *mt = rq->migration_thread;
2858 get_task_struct(mt);
2859 task_rq_unlock(rq, &flags);
2860 wake_up_process(mt);
2861 put_task_struct(mt);
2862 wait_for_completion(&req.done);
2867 task_rq_unlock(rq, &flags);
2871 * sched_exec - execve() is a valuable balancing opportunity, because at
2872 * this point the task has the smallest effective memory and cache footprint.
2874 void sched_exec(void)
2876 int new_cpu, this_cpu = get_cpu();
2877 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2879 if (new_cpu != this_cpu)
2880 sched_migrate_task(current, new_cpu);
2884 * pull_task - move a task from a remote runqueue to the local runqueue.
2885 * Both runqueues must be locked.
2887 static void pull_task(struct rq *src_rq, struct task_struct *p,
2888 struct rq *this_rq, int this_cpu)
2890 deactivate_task(src_rq, p, 0);
2891 set_task_cpu(p, this_cpu);
2892 activate_task(this_rq, p, 0);
2894 * Note that idle threads have a prio of MAX_PRIO, for this test
2895 * to be always true for them.
2897 check_preempt_curr(this_rq, p, 0);
2901 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2904 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2905 struct sched_domain *sd, enum cpu_idle_type idle,
2909 * We do not migrate tasks that are:
2910 * 1) running (obviously), or
2911 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2912 * 3) are cache-hot on their current CPU.
2914 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2915 schedstat_inc(p, se.nr_failed_migrations_affine);
2920 if (task_running(rq, p)) {
2921 schedstat_inc(p, se.nr_failed_migrations_running);
2926 * Aggressive migration if:
2927 * 1) task is cache cold, or
2928 * 2) too many balance attempts have failed.
2931 if (!task_hot(p, rq->clock, sd) ||
2932 sd->nr_balance_failed > sd->cache_nice_tries) {
2933 #ifdef CONFIG_SCHEDSTATS
2934 if (task_hot(p, rq->clock, sd)) {
2935 schedstat_inc(sd, lb_hot_gained[idle]);
2936 schedstat_inc(p, se.nr_forced_migrations);
2942 if (task_hot(p, rq->clock, sd)) {
2943 schedstat_inc(p, se.nr_failed_migrations_hot);
2949 static unsigned long
2950 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2951 unsigned long max_load_move, struct sched_domain *sd,
2952 enum cpu_idle_type idle, int *all_pinned,
2953 int *this_best_prio, struct rq_iterator *iterator)
2955 int loops = 0, pulled = 0, pinned = 0;
2956 struct task_struct *p;
2957 long rem_load_move = max_load_move;
2959 if (max_load_move == 0)
2965 * Start the load-balancing iterator:
2967 p = iterator->start(iterator->arg);
2969 if (!p || loops++ > sysctl_sched_nr_migrate)
2972 if ((p->se.load.weight >> 1) > rem_load_move ||
2973 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2974 p = iterator->next(iterator->arg);
2978 pull_task(busiest, p, this_rq, this_cpu);
2980 rem_load_move -= p->se.load.weight;
2983 * We only want to steal up to the prescribed amount of weighted load.
2985 if (rem_load_move > 0) {
2986 if (p->prio < *this_best_prio)
2987 *this_best_prio = p->prio;
2988 p = iterator->next(iterator->arg);
2993 * Right now, this is one of only two places pull_task() is called,
2994 * so we can safely collect pull_task() stats here rather than
2995 * inside pull_task().
2997 schedstat_add(sd, lb_gained[idle], pulled);
3000 *all_pinned = pinned;
3002 return max_load_move - rem_load_move;
3006 * move_tasks tries to move up to max_load_move weighted load from busiest to
3007 * this_rq, as part of a balancing operation within domain "sd".
3008 * Returns 1 if successful and 0 otherwise.
3010 * Called with both runqueues locked.
3012 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3013 unsigned long max_load_move,
3014 struct sched_domain *sd, enum cpu_idle_type idle,
3017 const struct sched_class *class = sched_class_highest;
3018 unsigned long total_load_moved = 0;
3019 int this_best_prio = this_rq->curr->prio;
3023 class->load_balance(this_rq, this_cpu, busiest,
3024 max_load_move - total_load_moved,
3025 sd, idle, all_pinned, &this_best_prio);
3026 class = class->next;
3028 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3031 } while (class && max_load_move > total_load_moved);
3033 return total_load_moved > 0;
3037 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3038 struct sched_domain *sd, enum cpu_idle_type idle,
3039 struct rq_iterator *iterator)
3041 struct task_struct *p = iterator->start(iterator->arg);
3045 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3046 pull_task(busiest, p, this_rq, this_cpu);
3048 * Right now, this is only the second place pull_task()
3049 * is called, so we can safely collect pull_task()
3050 * stats here rather than inside pull_task().
3052 schedstat_inc(sd, lb_gained[idle]);
3056 p = iterator->next(iterator->arg);
3063 * move_one_task tries to move exactly one task from busiest to this_rq, as
3064 * part of active balancing operations within "domain".
3065 * Returns 1 if successful and 0 otherwise.
3067 * Called with both runqueues locked.
3069 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3070 struct sched_domain *sd, enum cpu_idle_type idle)
3072 const struct sched_class *class;
3074 for (class = sched_class_highest; class; class = class->next)
3075 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3082 * find_busiest_group finds and returns the busiest CPU group within the
3083 * domain. It calculates and returns the amount of weighted load which
3084 * should be moved to restore balance via the imbalance parameter.
3086 static struct sched_group *
3087 find_busiest_group(struct sched_domain *sd, int this_cpu,
3088 unsigned long *imbalance, enum cpu_idle_type idle,
3089 int *sd_idle, const cpumask_t *cpus, int *balance)
3091 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3092 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3093 unsigned long max_pull;
3094 unsigned long busiest_load_per_task, busiest_nr_running;
3095 unsigned long this_load_per_task, this_nr_running;
3096 int load_idx, group_imb = 0;
3097 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3098 int power_savings_balance = 1;
3099 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3100 unsigned long min_nr_running = ULONG_MAX;
3101 struct sched_group *group_min = NULL, *group_leader = NULL;
3104 max_load = this_load = total_load = total_pwr = 0;
3105 busiest_load_per_task = busiest_nr_running = 0;
3106 this_load_per_task = this_nr_running = 0;
3108 if (idle == CPU_NOT_IDLE)
3109 load_idx = sd->busy_idx;
3110 else if (idle == CPU_NEWLY_IDLE)
3111 load_idx = sd->newidle_idx;
3113 load_idx = sd->idle_idx;
3116 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3119 int __group_imb = 0;
3120 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3121 unsigned long sum_nr_running, sum_weighted_load;
3122 unsigned long sum_avg_load_per_task;
3123 unsigned long avg_load_per_task;
3125 local_group = cpu_isset(this_cpu, group->cpumask);
3128 balance_cpu = first_cpu(group->cpumask);
3130 /* Tally up the load of all CPUs in the group */
3131 sum_weighted_load = sum_nr_running = avg_load = 0;
3132 sum_avg_load_per_task = avg_load_per_task = 0;
3135 min_cpu_load = ~0UL;
3137 for_each_cpu_mask_nr(i, group->cpumask) {
3140 if (!cpu_isset(i, *cpus))
3145 if (*sd_idle && rq->nr_running)
3148 /* Bias balancing toward cpus of our domain */
3150 if (idle_cpu(i) && !first_idle_cpu) {
3155 load = target_load(i, load_idx);
3157 load = source_load(i, load_idx);
3158 if (load > max_cpu_load)
3159 max_cpu_load = load;
3160 if (min_cpu_load > load)
3161 min_cpu_load = load;
3165 sum_nr_running += rq->nr_running;
3166 sum_weighted_load += weighted_cpuload(i);
3168 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3172 * First idle cpu or the first cpu(busiest) in this sched group
3173 * is eligible for doing load balancing at this and above
3174 * domains. In the newly idle case, we will allow all the cpu's
3175 * to do the newly idle load balance.
3177 if (idle != CPU_NEWLY_IDLE && local_group &&
3178 balance_cpu != this_cpu && balance) {
3183 total_load += avg_load;
3184 total_pwr += group->__cpu_power;
3186 /* Adjust by relative CPU power of the group */
3187 avg_load = sg_div_cpu_power(group,
3188 avg_load * SCHED_LOAD_SCALE);
3192 * Consider the group unbalanced when the imbalance is larger
3193 * than the average weight of two tasks.
3195 * APZ: with cgroup the avg task weight can vary wildly and
3196 * might not be a suitable number - should we keep a
3197 * normalized nr_running number somewhere that negates
3200 avg_load_per_task = sg_div_cpu_power(group,
3201 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3203 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3206 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3209 this_load = avg_load;
3211 this_nr_running = sum_nr_running;
3212 this_load_per_task = sum_weighted_load;
3213 } else if (avg_load > max_load &&
3214 (sum_nr_running > group_capacity || __group_imb)) {
3215 max_load = avg_load;
3217 busiest_nr_running = sum_nr_running;
3218 busiest_load_per_task = sum_weighted_load;
3219 group_imb = __group_imb;
3222 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3224 * Busy processors will not participate in power savings
3227 if (idle == CPU_NOT_IDLE ||
3228 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3232 * If the local group is idle or completely loaded
3233 * no need to do power savings balance at this domain
3235 if (local_group && (this_nr_running >= group_capacity ||
3237 power_savings_balance = 0;
3240 * If a group is already running at full capacity or idle,
3241 * don't include that group in power savings calculations
3243 if (!power_savings_balance || sum_nr_running >= group_capacity
3248 * Calculate the group which has the least non-idle load.
3249 * This is the group from where we need to pick up the load
3252 if ((sum_nr_running < min_nr_running) ||
3253 (sum_nr_running == min_nr_running &&
3254 first_cpu(group->cpumask) <
3255 first_cpu(group_min->cpumask))) {
3257 min_nr_running = sum_nr_running;
3258 min_load_per_task = sum_weighted_load /
3263 * Calculate the group which is almost near its
3264 * capacity but still has some space to pick up some load
3265 * from other group and save more power
3267 if (sum_nr_running <= group_capacity - 1) {
3268 if (sum_nr_running > leader_nr_running ||
3269 (sum_nr_running == leader_nr_running &&
3270 first_cpu(group->cpumask) >
3271 first_cpu(group_leader->cpumask))) {
3272 group_leader = group;
3273 leader_nr_running = sum_nr_running;
3278 group = group->next;
3279 } while (group != sd->groups);
3281 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3284 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3286 if (this_load >= avg_load ||
3287 100*max_load <= sd->imbalance_pct*this_load)
3290 busiest_load_per_task /= busiest_nr_running;
3292 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3295 * We're trying to get all the cpus to the average_load, so we don't
3296 * want to push ourselves above the average load, nor do we wish to
3297 * reduce the max loaded cpu below the average load, as either of these
3298 * actions would just result in more rebalancing later, and ping-pong
3299 * tasks around. Thus we look for the minimum possible imbalance.
3300 * Negative imbalances (*we* are more loaded than anyone else) will
3301 * be counted as no imbalance for these purposes -- we can't fix that
3302 * by pulling tasks to us. Be careful of negative numbers as they'll
3303 * appear as very large values with unsigned longs.
3305 if (max_load <= busiest_load_per_task)
3309 * In the presence of smp nice balancing, certain scenarios can have
3310 * max load less than avg load(as we skip the groups at or below
3311 * its cpu_power, while calculating max_load..)
3313 if (max_load < avg_load) {
3315 goto small_imbalance;
3318 /* Don't want to pull so many tasks that a group would go idle */
3319 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3321 /* How much load to actually move to equalise the imbalance */
3322 *imbalance = min(max_pull * busiest->__cpu_power,
3323 (avg_load - this_load) * this->__cpu_power)
3327 * if *imbalance is less than the average load per runnable task
3328 * there is no gaurantee that any tasks will be moved so we'll have
3329 * a think about bumping its value to force at least one task to be
3332 if (*imbalance < busiest_load_per_task) {
3333 unsigned long tmp, pwr_now, pwr_move;
3337 pwr_move = pwr_now = 0;
3339 if (this_nr_running) {
3340 this_load_per_task /= this_nr_running;
3341 if (busiest_load_per_task > this_load_per_task)
3344 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3346 if (max_load - this_load + busiest_load_per_task >=
3347 busiest_load_per_task * imbn) {
3348 *imbalance = busiest_load_per_task;
3353 * OK, we don't have enough imbalance to justify moving tasks,
3354 * however we may be able to increase total CPU power used by
3358 pwr_now += busiest->__cpu_power *
3359 min(busiest_load_per_task, max_load);
3360 pwr_now += this->__cpu_power *
3361 min(this_load_per_task, this_load);
3362 pwr_now /= SCHED_LOAD_SCALE;
3364 /* Amount of load we'd subtract */
3365 tmp = sg_div_cpu_power(busiest,
3366 busiest_load_per_task * SCHED_LOAD_SCALE);
3368 pwr_move += busiest->__cpu_power *
3369 min(busiest_load_per_task, max_load - tmp);
3371 /* Amount of load we'd add */
3372 if (max_load * busiest->__cpu_power <
3373 busiest_load_per_task * SCHED_LOAD_SCALE)
3374 tmp = sg_div_cpu_power(this,
3375 max_load * busiest->__cpu_power);
3377 tmp = sg_div_cpu_power(this,
3378 busiest_load_per_task * SCHED_LOAD_SCALE);
3379 pwr_move += this->__cpu_power *
3380 min(this_load_per_task, this_load + tmp);
3381 pwr_move /= SCHED_LOAD_SCALE;
3383 /* Move if we gain throughput */
3384 if (pwr_move > pwr_now)
3385 *imbalance = busiest_load_per_task;
3391 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3392 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3395 if (this == group_leader && group_leader != group_min) {
3396 *imbalance = min_load_per_task;
3406 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3409 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3410 unsigned long imbalance, const cpumask_t *cpus)
3412 struct rq *busiest = NULL, *rq;
3413 unsigned long max_load = 0;
3416 for_each_cpu_mask_nr(i, group->cpumask) {
3419 if (!cpu_isset(i, *cpus))
3423 wl = weighted_cpuload(i);
3425 if (rq->nr_running == 1 && wl > imbalance)
3428 if (wl > max_load) {
3438 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3439 * so long as it is large enough.
3441 #define MAX_PINNED_INTERVAL 512
3444 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3445 * tasks if there is an imbalance.
3447 static int load_balance(int this_cpu, struct rq *this_rq,
3448 struct sched_domain *sd, enum cpu_idle_type idle,
3449 int *balance, cpumask_t *cpus)
3451 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3452 struct sched_group *group;
3453 unsigned long imbalance;
3455 unsigned long flags;
3460 * When power savings policy is enabled for the parent domain, idle
3461 * sibling can pick up load irrespective of busy siblings. In this case,
3462 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3463 * portraying it as CPU_NOT_IDLE.
3465 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3466 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3469 schedstat_inc(sd, lb_count[idle]);
3473 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3480 schedstat_inc(sd, lb_nobusyg[idle]);
3484 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3486 schedstat_inc(sd, lb_nobusyq[idle]);
3490 BUG_ON(busiest == this_rq);
3492 schedstat_add(sd, lb_imbalance[idle], imbalance);
3495 if (busiest->nr_running > 1) {
3497 * Attempt to move tasks. If find_busiest_group has found
3498 * an imbalance but busiest->nr_running <= 1, the group is
3499 * still unbalanced. ld_moved simply stays zero, so it is
3500 * correctly treated as an imbalance.
3502 local_irq_save(flags);
3503 double_rq_lock(this_rq, busiest);
3504 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3505 imbalance, sd, idle, &all_pinned);
3506 double_rq_unlock(this_rq, busiest);
3507 local_irq_restore(flags);
3510 * some other cpu did the load balance for us.
3512 if (ld_moved && this_cpu != smp_processor_id())
3513 resched_cpu(this_cpu);
3515 /* All tasks on this runqueue were pinned by CPU affinity */
3516 if (unlikely(all_pinned)) {
3517 cpu_clear(cpu_of(busiest), *cpus);
3518 if (!cpus_empty(*cpus))
3525 schedstat_inc(sd, lb_failed[idle]);
3526 sd->nr_balance_failed++;
3528 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3530 spin_lock_irqsave(&busiest->lock, flags);
3532 /* don't kick the migration_thread, if the curr
3533 * task on busiest cpu can't be moved to this_cpu
3535 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3536 spin_unlock_irqrestore(&busiest->lock, flags);
3538 goto out_one_pinned;
3541 if (!busiest->active_balance) {
3542 busiest->active_balance = 1;
3543 busiest->push_cpu = this_cpu;
3546 spin_unlock_irqrestore(&busiest->lock, flags);
3548 wake_up_process(busiest->migration_thread);
3551 * We've kicked active balancing, reset the failure
3554 sd->nr_balance_failed = sd->cache_nice_tries+1;
3557 sd->nr_balance_failed = 0;
3559 if (likely(!active_balance)) {
3560 /* We were unbalanced, so reset the balancing interval */
3561 sd->balance_interval = sd->min_interval;
3564 * If we've begun active balancing, start to back off. This
3565 * case may not be covered by the all_pinned logic if there
3566 * is only 1 task on the busy runqueue (because we don't call
3569 if (sd->balance_interval < sd->max_interval)
3570 sd->balance_interval *= 2;
3573 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3574 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3580 schedstat_inc(sd, lb_balanced[idle]);
3582 sd->nr_balance_failed = 0;
3585 /* tune up the balancing interval */
3586 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3587 (sd->balance_interval < sd->max_interval))
3588 sd->balance_interval *= 2;
3590 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3591 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3602 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3603 * tasks if there is an imbalance.
3605 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3606 * this_rq is locked.
3609 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3612 struct sched_group *group;
3613 struct rq *busiest = NULL;
3614 unsigned long imbalance;
3622 * When power savings policy is enabled for the parent domain, idle
3623 * sibling can pick up load irrespective of busy siblings. In this case,
3624 * let the state of idle sibling percolate up as IDLE, instead of
3625 * portraying it as CPU_NOT_IDLE.
3627 if (sd->flags & SD_SHARE_CPUPOWER &&
3628 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3631 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3633 update_shares_locked(this_rq, sd);
3634 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3635 &sd_idle, cpus, NULL);
3637 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3641 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3643 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3647 BUG_ON(busiest == this_rq);
3649 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3652 if (busiest->nr_running > 1) {
3653 /* Attempt to move tasks */
3654 double_lock_balance(this_rq, busiest);
3655 /* this_rq->clock is already updated */
3656 update_rq_clock(busiest);
3657 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3658 imbalance, sd, CPU_NEWLY_IDLE,
3660 double_unlock_balance(this_rq, busiest);
3662 if (unlikely(all_pinned)) {
3663 cpu_clear(cpu_of(busiest), *cpus);
3664 if (!cpus_empty(*cpus))
3670 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3671 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3672 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3675 sd->nr_balance_failed = 0;
3677 update_shares_locked(this_rq, sd);
3681 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3682 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3683 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3685 sd->nr_balance_failed = 0;
3691 * idle_balance is called by schedule() if this_cpu is about to become
3692 * idle. Attempts to pull tasks from other CPUs.
3694 static void idle_balance(int this_cpu, struct rq *this_rq)
3696 struct sched_domain *sd;
3697 int pulled_task = 0;
3698 unsigned long next_balance = jiffies + HZ;
3701 for_each_domain(this_cpu, sd) {
3702 unsigned long interval;
3704 if (!(sd->flags & SD_LOAD_BALANCE))
3707 if (sd->flags & SD_BALANCE_NEWIDLE)
3708 /* If we've pulled tasks over stop searching: */
3709 pulled_task = load_balance_newidle(this_cpu, this_rq,
3712 interval = msecs_to_jiffies(sd->balance_interval);
3713 if (time_after(next_balance, sd->last_balance + interval))
3714 next_balance = sd->last_balance + interval;
3718 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3720 * We are going idle. next_balance may be set based on
3721 * a busy processor. So reset next_balance.
3723 this_rq->next_balance = next_balance;
3728 * active_load_balance is run by migration threads. It pushes running tasks
3729 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3730 * running on each physical CPU where possible, and avoids physical /
3731 * logical imbalances.
3733 * Called with busiest_rq locked.
3735 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3737 int target_cpu = busiest_rq->push_cpu;
3738 struct sched_domain *sd;
3739 struct rq *target_rq;
3741 /* Is there any task to move? */
3742 if (busiest_rq->nr_running <= 1)
3745 target_rq = cpu_rq(target_cpu);
3748 * This condition is "impossible", if it occurs
3749 * we need to fix it. Originally reported by
3750 * Bjorn Helgaas on a 128-cpu setup.
3752 BUG_ON(busiest_rq == target_rq);
3754 /* move a task from busiest_rq to target_rq */
3755 double_lock_balance(busiest_rq, target_rq);
3756 update_rq_clock(busiest_rq);
3757 update_rq_clock(target_rq);
3759 /* Search for an sd spanning us and the target CPU. */
3760 for_each_domain(target_cpu, sd) {
3761 if ((sd->flags & SD_LOAD_BALANCE) &&
3762 cpu_isset(busiest_cpu, sd->span))
3767 schedstat_inc(sd, alb_count);
3769 if (move_one_task(target_rq, target_cpu, busiest_rq,
3771 schedstat_inc(sd, alb_pushed);
3773 schedstat_inc(sd, alb_failed);
3775 double_unlock_balance(busiest_rq, target_rq);
3780 atomic_t load_balancer;
3782 } nohz ____cacheline_aligned = {
3783 .load_balancer = ATOMIC_INIT(-1),
3784 .cpu_mask = CPU_MASK_NONE,
3788 * This routine will try to nominate the ilb (idle load balancing)
3789 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3790 * load balancing on behalf of all those cpus. If all the cpus in the system
3791 * go into this tickless mode, then there will be no ilb owner (as there is
3792 * no need for one) and all the cpus will sleep till the next wakeup event
3795 * For the ilb owner, tick is not stopped. And this tick will be used
3796 * for idle load balancing. ilb owner will still be part of
3799 * While stopping the tick, this cpu will become the ilb owner if there
3800 * is no other owner. And will be the owner till that cpu becomes busy
3801 * or if all cpus in the system stop their ticks at which point
3802 * there is no need for ilb owner.
3804 * When the ilb owner becomes busy, it nominates another owner, during the
3805 * next busy scheduler_tick()
3807 int select_nohz_load_balancer(int stop_tick)
3809 int cpu = smp_processor_id();
3812 cpu_set(cpu, nohz.cpu_mask);
3813 cpu_rq(cpu)->in_nohz_recently = 1;
3816 * If we are going offline and still the leader, give up!
3818 if (!cpu_active(cpu) &&
3819 atomic_read(&nohz.load_balancer) == cpu) {
3820 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3825 /* time for ilb owner also to sleep */
3826 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3827 if (atomic_read(&nohz.load_balancer) == cpu)
3828 atomic_set(&nohz.load_balancer, -1);
3832 if (atomic_read(&nohz.load_balancer) == -1) {
3833 /* make me the ilb owner */
3834 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3836 } else if (atomic_read(&nohz.load_balancer) == cpu)
3839 if (!cpu_isset(cpu, nohz.cpu_mask))
3842 cpu_clear(cpu, nohz.cpu_mask);
3844 if (atomic_read(&nohz.load_balancer) == cpu)
3845 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3852 static DEFINE_SPINLOCK(balancing);
3855 * It checks each scheduling domain to see if it is due to be balanced,
3856 * and initiates a balancing operation if so.
3858 * Balancing parameters are set up in arch_init_sched_domains.
3860 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3863 struct rq *rq = cpu_rq(cpu);
3864 unsigned long interval;
3865 struct sched_domain *sd;
3866 /* Earliest time when we have to do rebalance again */
3867 unsigned long next_balance = jiffies + 60*HZ;
3868 int update_next_balance = 0;
3872 for_each_domain(cpu, sd) {
3873 if (!(sd->flags & SD_LOAD_BALANCE))
3876 interval = sd->balance_interval;
3877 if (idle != CPU_IDLE)
3878 interval *= sd->busy_factor;
3880 /* scale ms to jiffies */
3881 interval = msecs_to_jiffies(interval);
3882 if (unlikely(!interval))
3884 if (interval > HZ*NR_CPUS/10)
3885 interval = HZ*NR_CPUS/10;
3887 need_serialize = sd->flags & SD_SERIALIZE;
3889 if (need_serialize) {
3890 if (!spin_trylock(&balancing))
3894 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3895 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3897 * We've pulled tasks over so either we're no
3898 * longer idle, or one of our SMT siblings is
3901 idle = CPU_NOT_IDLE;
3903 sd->last_balance = jiffies;
3906 spin_unlock(&balancing);
3908 if (time_after(next_balance, sd->last_balance + interval)) {
3909 next_balance = sd->last_balance + interval;
3910 update_next_balance = 1;
3914 * Stop the load balance at this level. There is another
3915 * CPU in our sched group which is doing load balancing more
3923 * next_balance will be updated only when there is a need.
3924 * When the cpu is attached to null domain for ex, it will not be
3927 if (likely(update_next_balance))
3928 rq->next_balance = next_balance;
3932 * run_rebalance_domains is triggered when needed from the scheduler tick.
3933 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3934 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3936 static void run_rebalance_domains(struct softirq_action *h)
3938 int this_cpu = smp_processor_id();
3939 struct rq *this_rq = cpu_rq(this_cpu);
3940 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3941 CPU_IDLE : CPU_NOT_IDLE;
3943 rebalance_domains(this_cpu, idle);
3947 * If this cpu is the owner for idle load balancing, then do the
3948 * balancing on behalf of the other idle cpus whose ticks are
3951 if (this_rq->idle_at_tick &&
3952 atomic_read(&nohz.load_balancer) == this_cpu) {
3953 cpumask_t cpus = nohz.cpu_mask;
3957 cpu_clear(this_cpu, cpus);
3958 for_each_cpu_mask_nr(balance_cpu, cpus) {
3960 * If this cpu gets work to do, stop the load balancing
3961 * work being done for other cpus. Next load
3962 * balancing owner will pick it up.
3967 rebalance_domains(balance_cpu, CPU_IDLE);
3969 rq = cpu_rq(balance_cpu);
3970 if (time_after(this_rq->next_balance, rq->next_balance))
3971 this_rq->next_balance = rq->next_balance;
3978 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3980 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3981 * idle load balancing owner or decide to stop the periodic load balancing,
3982 * if the whole system is idle.
3984 static inline void trigger_load_balance(struct rq *rq, int cpu)
3988 * If we were in the nohz mode recently and busy at the current
3989 * scheduler tick, then check if we need to nominate new idle
3992 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3993 rq->in_nohz_recently = 0;
3995 if (atomic_read(&nohz.load_balancer) == cpu) {
3996 cpu_clear(cpu, nohz.cpu_mask);
3997 atomic_set(&nohz.load_balancer, -1);
4000 if (atomic_read(&nohz.load_balancer) == -1) {
4002 * simple selection for now: Nominate the
4003 * first cpu in the nohz list to be the next
4006 * TBD: Traverse the sched domains and nominate
4007 * the nearest cpu in the nohz.cpu_mask.
4009 int ilb = first_cpu(nohz.cpu_mask);
4011 if (ilb < nr_cpu_ids)
4017 * If this cpu is idle and doing idle load balancing for all the
4018 * cpus with ticks stopped, is it time for that to stop?
4020 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4021 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4027 * If this cpu is idle and the idle load balancing is done by
4028 * someone else, then no need raise the SCHED_SOFTIRQ
4030 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4031 cpu_isset(cpu, nohz.cpu_mask))
4034 if (time_after_eq(jiffies, rq->next_balance))
4035 raise_softirq(SCHED_SOFTIRQ);
4038 #else /* CONFIG_SMP */
4041 * on UP we do not need to balance between CPUs:
4043 static inline void idle_balance(int cpu, struct rq *rq)
4049 DEFINE_PER_CPU(struct kernel_stat, kstat);
4051 EXPORT_PER_CPU_SYMBOL(kstat);
4054 * Return any ns on the sched_clock that have not yet been banked in
4055 * @p in case that task is currently running.
4057 unsigned long long task_delta_exec(struct task_struct *p)
4059 unsigned long flags;
4063 rq = task_rq_lock(p, &flags);
4065 if (task_current(rq, p)) {
4068 update_rq_clock(rq);
4069 delta_exec = rq->clock - p->se.exec_start;
4070 if ((s64)delta_exec > 0)
4074 task_rq_unlock(rq, &flags);
4080 * Account user cpu time to a process.
4081 * @p: the process that the cpu time gets accounted to
4082 * @cputime: the cpu time spent in user space since the last update
4083 * @cputime_scaled: cputime scaled by cpu frequency
4085 void account_user_time(struct task_struct *p, cputime_t cputime,
4086 cputime_t cputime_scaled)
4088 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4091 /* Add user time to process. */
4092 p->utime = cputime_add(p->utime, cputime);
4093 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4094 account_group_user_time(p, cputime);
4096 /* Add user time to cpustat. */
4097 tmp = cputime_to_cputime64(cputime);
4098 if (TASK_NICE(p) > 0)
4099 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4101 cpustat->user = cputime64_add(cpustat->user, tmp);
4102 /* Account for user time used */
4103 acct_update_integrals(p);
4107 * Account guest cpu time to a process.
4108 * @p: the process that the cpu time gets accounted to
4109 * @cputime: the cpu time spent in virtual machine since the last update
4110 * @cputime_scaled: cputime scaled by cpu frequency
4112 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4113 cputime_t cputime_scaled)
4116 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4118 tmp = cputime_to_cputime64(cputime);
4120 /* Add guest time to process. */
4121 p->utime = cputime_add(p->utime, cputime);
4122 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4123 account_group_user_time(p, cputime);
4124 p->gtime = cputime_add(p->gtime, cputime);
4126 /* Add guest time to cpustat. */
4127 cpustat->user = cputime64_add(cpustat->user, tmp);
4128 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4132 * Account system cpu time to a process.
4133 * @p: the process that the cpu time gets accounted to
4134 * @hardirq_offset: the offset to subtract from hardirq_count()
4135 * @cputime: the cpu time spent in kernel space since the last update
4136 * @cputime_scaled: cputime scaled by cpu frequency
4138 void account_system_time(struct task_struct *p, int hardirq_offset,
4139 cputime_t cputime, cputime_t cputime_scaled)
4141 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4142 struct rq *rq = this_rq();
4145 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4146 account_guest_time(p, cputime, cputime_scaled);
4150 /* Add system time to process. */
4151 p->stime = cputime_add(p->stime, cputime);
4152 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4153 account_group_system_time(p, cputime);
4155 /* Add system time to cpustat. */
4156 tmp = cputime_to_cputime64(cputime);
4157 if (hardirq_count() - hardirq_offset)
4158 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4159 else if (softirq_count())
4160 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4161 else if (p != rq->idle)
4162 cpustat->system = cputime64_add(cpustat->system, tmp);
4163 else if (atomic_read(&rq->nr_iowait) > 0)
4164 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4166 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4167 /* Account for system time used */
4168 acct_update_integrals(p);
4172 * Account for involuntary wait time.
4173 * @p: the process from which the cpu time has been stolen
4174 * @steal: the cpu time spent in involuntary wait
4176 void account_steal_time(struct task_struct *p, cputime_t steal)
4178 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4179 cputime64_t tmp = cputime_to_cputime64(steal);
4180 struct rq *rq = this_rq();
4182 if (p == rq->idle) {
4183 p->stime = cputime_add(p->stime, steal);
4184 if (atomic_read(&rq->nr_iowait) > 0)
4185 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4187 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4189 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4193 * Use precise platform statistics if available:
4195 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4196 cputime_t task_utime(struct task_struct *p)
4201 cputime_t task_stime(struct task_struct *p)
4206 cputime_t task_utime(struct task_struct *p)
4208 clock_t utime = cputime_to_clock_t(p->utime),
4209 total = utime + cputime_to_clock_t(p->stime);
4213 * Use CFS's precise accounting:
4215 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4219 do_div(temp, total);
4221 utime = (clock_t)temp;
4223 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4224 return p->prev_utime;
4227 cputime_t task_stime(struct task_struct *p)
4232 * Use CFS's precise accounting. (we subtract utime from
4233 * the total, to make sure the total observed by userspace
4234 * grows monotonically - apps rely on that):
4236 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4237 cputime_to_clock_t(task_utime(p));
4240 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4242 return p->prev_stime;
4246 inline cputime_t task_gtime(struct task_struct *p)
4252 * This function gets called by the timer code, with HZ frequency.
4253 * We call it with interrupts disabled.
4255 * It also gets called by the fork code, when changing the parent's
4258 void scheduler_tick(void)
4260 int cpu = smp_processor_id();
4261 struct rq *rq = cpu_rq(cpu);
4262 struct task_struct *curr = rq->curr;
4266 spin_lock(&rq->lock);
4267 update_rq_clock(rq);
4268 update_cpu_load(rq);
4269 curr->sched_class->task_tick(rq, curr, 0);
4270 spin_unlock(&rq->lock);
4273 rq->idle_at_tick = idle_cpu(cpu);
4274 trigger_load_balance(rq, cpu);
4278 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4279 defined(CONFIG_PREEMPT_TRACER))
4281 static inline unsigned long get_parent_ip(unsigned long addr)
4283 if (in_lock_functions(addr)) {
4284 addr = CALLER_ADDR2;
4285 if (in_lock_functions(addr))
4286 addr = CALLER_ADDR3;
4291 void __kprobes add_preempt_count(int val)
4293 #ifdef CONFIG_DEBUG_PREEMPT
4297 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4300 preempt_count() += val;
4301 #ifdef CONFIG_DEBUG_PREEMPT
4303 * Spinlock count overflowing soon?
4305 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4308 if (preempt_count() == val)
4309 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4311 EXPORT_SYMBOL(add_preempt_count);
4313 void __kprobes sub_preempt_count(int val)
4315 #ifdef CONFIG_DEBUG_PREEMPT
4319 if (DEBUG_LOCKS_WARN_ON(val > preempt_count() - (!!kernel_locked())))
4322 * Is the spinlock portion underflowing?
4324 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4325 !(preempt_count() & PREEMPT_MASK)))
4329 if (preempt_count() == val)
4330 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4331 preempt_count() -= val;
4333 EXPORT_SYMBOL(sub_preempt_count);
4338 * Print scheduling while atomic bug:
4340 static noinline void __schedule_bug(struct task_struct *prev)
4342 struct pt_regs *regs = get_irq_regs();
4344 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4345 prev->comm, prev->pid, preempt_count());
4347 debug_show_held_locks(prev);
4349 if (irqs_disabled())
4350 print_irqtrace_events(prev);
4359 * Various schedule()-time debugging checks and statistics:
4361 static inline void schedule_debug(struct task_struct *prev)
4364 * Test if we are atomic. Since do_exit() needs to call into
4365 * schedule() atomically, we ignore that path for now.
4366 * Otherwise, whine if we are scheduling when we should not be.
4368 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4369 __schedule_bug(prev);
4371 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4373 schedstat_inc(this_rq(), sched_count);
4374 #ifdef CONFIG_SCHEDSTATS
4375 if (unlikely(prev->lock_depth >= 0)) {
4376 schedstat_inc(this_rq(), bkl_count);
4377 schedstat_inc(prev, sched_info.bkl_count);
4383 * Pick up the highest-prio task:
4385 static inline struct task_struct *
4386 pick_next_task(struct rq *rq, struct task_struct *prev)
4388 const struct sched_class *class;
4389 struct task_struct *p;
4392 * Optimization: we know that if all tasks are in
4393 * the fair class we can call that function directly:
4395 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4396 p = fair_sched_class.pick_next_task(rq);
4401 class = sched_class_highest;
4403 p = class->pick_next_task(rq);
4407 * Will never be NULL as the idle class always
4408 * returns a non-NULL p:
4410 class = class->next;
4415 * schedule() is the main scheduler function.
4417 asmlinkage void __sched schedule(void)
4419 struct task_struct *prev, *next;
4420 unsigned long *switch_count;
4426 cpu = smp_processor_id();
4430 switch_count = &prev->nivcsw;
4432 release_kernel_lock(prev);
4433 need_resched_nonpreemptible:
4435 schedule_debug(prev);
4437 if (sched_feat(HRTICK))
4440 spin_lock_irq(&rq->lock);
4441 update_rq_clock(rq);
4442 clear_tsk_need_resched(prev);
4444 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4445 if (unlikely(signal_pending_state(prev->state, prev)))
4446 prev->state = TASK_RUNNING;
4448 deactivate_task(rq, prev, 1);
4449 switch_count = &prev->nvcsw;
4453 if (prev->sched_class->pre_schedule)
4454 prev->sched_class->pre_schedule(rq, prev);
4457 if (unlikely(!rq->nr_running))
4458 idle_balance(cpu, rq);
4460 prev->sched_class->put_prev_task(rq, prev);
4461 next = pick_next_task(rq, prev);
4463 if (likely(prev != next)) {
4464 sched_info_switch(prev, next);
4470 context_switch(rq, prev, next); /* unlocks the rq */
4472 * the context switch might have flipped the stack from under
4473 * us, hence refresh the local variables.
4475 cpu = smp_processor_id();
4478 spin_unlock_irq(&rq->lock);
4480 if (unlikely(reacquire_kernel_lock(current) < 0))
4481 goto need_resched_nonpreemptible;
4483 preempt_enable_no_resched();
4484 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4487 EXPORT_SYMBOL(schedule);
4489 #ifdef CONFIG_PREEMPT
4491 * this is the entry point to schedule() from in-kernel preemption
4492 * off of preempt_enable. Kernel preemptions off return from interrupt
4493 * occur there and call schedule directly.
4495 asmlinkage void __sched preempt_schedule(void)
4497 struct thread_info *ti = current_thread_info();
4500 * If there is a non-zero preempt_count or interrupts are disabled,
4501 * we do not want to preempt the current task. Just return..
4503 if (likely(ti->preempt_count || irqs_disabled()))
4507 add_preempt_count(PREEMPT_ACTIVE);
4509 sub_preempt_count(PREEMPT_ACTIVE);
4512 * Check again in case we missed a preemption opportunity
4513 * between schedule and now.
4516 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4518 EXPORT_SYMBOL(preempt_schedule);
4521 * this is the entry point to schedule() from kernel preemption
4522 * off of irq context.
4523 * Note, that this is called and return with irqs disabled. This will
4524 * protect us against recursive calling from irq.
4526 asmlinkage void __sched preempt_schedule_irq(void)
4528 struct thread_info *ti = current_thread_info();
4530 /* Catch callers which need to be fixed */
4531 BUG_ON(ti->preempt_count || !irqs_disabled());
4534 add_preempt_count(PREEMPT_ACTIVE);
4537 local_irq_disable();
4538 sub_preempt_count(PREEMPT_ACTIVE);
4541 * Check again in case we missed a preemption opportunity
4542 * between schedule and now.
4545 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4548 #endif /* CONFIG_PREEMPT */
4550 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4553 return try_to_wake_up(curr->private, mode, sync);
4555 EXPORT_SYMBOL(default_wake_function);
4558 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4559 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4560 * number) then we wake all the non-exclusive tasks and one exclusive task.
4562 * There are circumstances in which we can try to wake a task which has already
4563 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4564 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4566 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4567 int nr_exclusive, int sync, void *key)
4569 wait_queue_t *curr, *next;
4571 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4572 unsigned flags = curr->flags;
4574 if (curr->func(curr, mode, sync, key) &&
4575 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4581 * __wake_up - wake up threads blocked on a waitqueue.
4583 * @mode: which threads
4584 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4585 * @key: is directly passed to the wakeup function
4587 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4588 int nr_exclusive, void *key)
4590 unsigned long flags;
4592 spin_lock_irqsave(&q->lock, flags);
4593 __wake_up_common(q, mode, nr_exclusive, 0, key);
4594 spin_unlock_irqrestore(&q->lock, flags);
4596 EXPORT_SYMBOL(__wake_up);
4599 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4601 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4603 __wake_up_common(q, mode, 1, 0, NULL);
4607 * __wake_up_sync - wake up threads blocked on a waitqueue.
4609 * @mode: which threads
4610 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4612 * The sync wakeup differs that the waker knows that it will schedule
4613 * away soon, so while the target thread will be woken up, it will not
4614 * be migrated to another CPU - ie. the two threads are 'synchronized'
4615 * with each other. This can prevent needless bouncing between CPUs.
4617 * On UP it can prevent extra preemption.
4620 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4622 unsigned long flags;
4628 if (unlikely(!nr_exclusive))
4631 spin_lock_irqsave(&q->lock, flags);
4632 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4633 spin_unlock_irqrestore(&q->lock, flags);
4635 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4638 * complete: - signals a single thread waiting on this completion
4639 * @x: holds the state of this particular completion
4641 * This will wake up a single thread waiting on this completion. Threads will be
4642 * awakened in the same order in which they were queued.
4644 * See also complete_all(), wait_for_completion() and related routines.
4646 void complete(struct completion *x)
4648 unsigned long flags;
4650 spin_lock_irqsave(&x->wait.lock, flags);
4652 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4653 spin_unlock_irqrestore(&x->wait.lock, flags);
4655 EXPORT_SYMBOL(complete);
4658 * complete_all: - signals all threads waiting on this completion
4659 * @x: holds the state of this particular completion
4661 * This will wake up all threads waiting on this particular completion event.
4663 void complete_all(struct completion *x)
4665 unsigned long flags;
4667 spin_lock_irqsave(&x->wait.lock, flags);
4668 x->done += UINT_MAX/2;
4669 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4670 spin_unlock_irqrestore(&x->wait.lock, flags);
4672 EXPORT_SYMBOL(complete_all);
4674 static inline long __sched
4675 do_wait_for_common(struct completion *x, long timeout, int state)
4678 DECLARE_WAITQUEUE(wait, current);
4680 wait.flags |= WQ_FLAG_EXCLUSIVE;
4681 __add_wait_queue_tail(&x->wait, &wait);
4683 if (signal_pending_state(state, current)) {
4684 timeout = -ERESTARTSYS;
4687 __set_current_state(state);
4688 spin_unlock_irq(&x->wait.lock);
4689 timeout = schedule_timeout(timeout);
4690 spin_lock_irq(&x->wait.lock);
4691 } while (!x->done && timeout);
4692 __remove_wait_queue(&x->wait, &wait);
4697 return timeout ?: 1;
4701 wait_for_common(struct completion *x, long timeout, int state)
4705 spin_lock_irq(&x->wait.lock);
4706 timeout = do_wait_for_common(x, timeout, state);
4707 spin_unlock_irq(&x->wait.lock);
4712 * wait_for_completion: - waits for completion of a task
4713 * @x: holds the state of this particular completion
4715 * This waits to be signaled for completion of a specific task. It is NOT
4716 * interruptible and there is no timeout.
4718 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4719 * and interrupt capability. Also see complete().
4721 void __sched wait_for_completion(struct completion *x)
4723 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4725 EXPORT_SYMBOL(wait_for_completion);
4728 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4729 * @x: holds the state of this particular completion
4730 * @timeout: timeout value in jiffies
4732 * This waits for either a completion of a specific task to be signaled or for a
4733 * specified timeout to expire. The timeout is in jiffies. It is not
4736 unsigned long __sched
4737 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4739 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4741 EXPORT_SYMBOL(wait_for_completion_timeout);
4744 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4745 * @x: holds the state of this particular completion
4747 * This waits for completion of a specific task to be signaled. It is
4750 int __sched wait_for_completion_interruptible(struct completion *x)
4752 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4753 if (t == -ERESTARTSYS)
4757 EXPORT_SYMBOL(wait_for_completion_interruptible);
4760 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4761 * @x: holds the state of this particular completion
4762 * @timeout: timeout value in jiffies
4764 * This waits for either a completion of a specific task to be signaled or for a
4765 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4767 unsigned long __sched
4768 wait_for_completion_interruptible_timeout(struct completion *x,
4769 unsigned long timeout)
4771 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4773 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4776 * wait_for_completion_killable: - waits for completion of a task (killable)
4777 * @x: holds the state of this particular completion
4779 * This waits to be signaled for completion of a specific task. It can be
4780 * interrupted by a kill signal.
4782 int __sched wait_for_completion_killable(struct completion *x)
4784 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4785 if (t == -ERESTARTSYS)
4789 EXPORT_SYMBOL(wait_for_completion_killable);
4792 * try_wait_for_completion - try to decrement a completion without blocking
4793 * @x: completion structure
4795 * Returns: 0 if a decrement cannot be done without blocking
4796 * 1 if a decrement succeeded.
4798 * If a completion is being used as a counting completion,
4799 * attempt to decrement the counter without blocking. This
4800 * enables us to avoid waiting if the resource the completion
4801 * is protecting is not available.
4803 bool try_wait_for_completion(struct completion *x)
4807 spin_lock_irq(&x->wait.lock);
4812 spin_unlock_irq(&x->wait.lock);
4815 EXPORT_SYMBOL(try_wait_for_completion);
4818 * completion_done - Test to see if a completion has any waiters
4819 * @x: completion structure
4821 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4822 * 1 if there are no waiters.
4825 bool completion_done(struct completion *x)
4829 spin_lock_irq(&x->wait.lock);
4832 spin_unlock_irq(&x->wait.lock);
4835 EXPORT_SYMBOL(completion_done);
4838 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4840 unsigned long flags;
4843 init_waitqueue_entry(&wait, current);
4845 __set_current_state(state);
4847 spin_lock_irqsave(&q->lock, flags);
4848 __add_wait_queue(q, &wait);
4849 spin_unlock(&q->lock);
4850 timeout = schedule_timeout(timeout);
4851 spin_lock_irq(&q->lock);
4852 __remove_wait_queue(q, &wait);
4853 spin_unlock_irqrestore(&q->lock, flags);
4858 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4860 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4862 EXPORT_SYMBOL(interruptible_sleep_on);
4865 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4867 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4869 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4871 void __sched sleep_on(wait_queue_head_t *q)
4873 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4875 EXPORT_SYMBOL(sleep_on);
4877 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4879 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4881 EXPORT_SYMBOL(sleep_on_timeout);
4883 #ifdef CONFIG_RT_MUTEXES
4886 * rt_mutex_setprio - set the current priority of a task
4888 * @prio: prio value (kernel-internal form)
4890 * This function changes the 'effective' priority of a task. It does
4891 * not touch ->normal_prio like __setscheduler().
4893 * Used by the rt_mutex code to implement priority inheritance logic.
4895 void rt_mutex_setprio(struct task_struct *p, int prio)
4897 unsigned long flags;
4898 int oldprio, on_rq, running;
4900 const struct sched_class *prev_class = p->sched_class;
4902 BUG_ON(prio < 0 || prio > MAX_PRIO);
4904 rq = task_rq_lock(p, &flags);
4905 update_rq_clock(rq);
4908 on_rq = p->se.on_rq;
4909 running = task_current(rq, p);
4911 dequeue_task(rq, p, 0);
4913 p->sched_class->put_prev_task(rq, p);
4916 p->sched_class = &rt_sched_class;
4918 p->sched_class = &fair_sched_class;
4923 p->sched_class->set_curr_task(rq);
4925 enqueue_task(rq, p, 0);
4927 check_class_changed(rq, p, prev_class, oldprio, running);
4929 task_rq_unlock(rq, &flags);
4934 void set_user_nice(struct task_struct *p, long nice)
4936 int old_prio, delta, on_rq;
4937 unsigned long flags;
4940 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4943 * We have to be careful, if called from sys_setpriority(),
4944 * the task might be in the middle of scheduling on another CPU.
4946 rq = task_rq_lock(p, &flags);
4947 update_rq_clock(rq);
4949 * The RT priorities are set via sched_setscheduler(), but we still
4950 * allow the 'normal' nice value to be set - but as expected
4951 * it wont have any effect on scheduling until the task is
4952 * SCHED_FIFO/SCHED_RR:
4954 if (task_has_rt_policy(p)) {
4955 p->static_prio = NICE_TO_PRIO(nice);
4958 on_rq = p->se.on_rq;
4960 dequeue_task(rq, p, 0);
4962 p->static_prio = NICE_TO_PRIO(nice);
4965 p->prio = effective_prio(p);
4966 delta = p->prio - old_prio;
4969 enqueue_task(rq, p, 0);
4971 * If the task increased its priority or is running and
4972 * lowered its priority, then reschedule its CPU:
4974 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4975 resched_task(rq->curr);
4978 task_rq_unlock(rq, &flags);
4980 EXPORT_SYMBOL(set_user_nice);
4983 * can_nice - check if a task can reduce its nice value
4987 int can_nice(const struct task_struct *p, const int nice)
4989 /* convert nice value [19,-20] to rlimit style value [1,40] */
4990 int nice_rlim = 20 - nice;
4992 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4993 capable(CAP_SYS_NICE));
4996 #ifdef __ARCH_WANT_SYS_NICE
4999 * sys_nice - change the priority of the current process.
5000 * @increment: priority increment
5002 * sys_setpriority is a more generic, but much slower function that
5003 * does similar things.
5005 asmlinkage long sys_nice(int increment)
5010 * Setpriority might change our priority at the same moment.
5011 * We don't have to worry. Conceptually one call occurs first
5012 * and we have a single winner.
5014 if (increment < -40)
5019 nice = PRIO_TO_NICE(current->static_prio) + increment;
5025 if (increment < 0 && !can_nice(current, nice))
5028 retval = security_task_setnice(current, nice);
5032 set_user_nice(current, nice);
5039 * task_prio - return the priority value of a given task.
5040 * @p: the task in question.
5042 * This is the priority value as seen by users in /proc.
5043 * RT tasks are offset by -200. Normal tasks are centered
5044 * around 0, value goes from -16 to +15.
5046 int task_prio(const struct task_struct *p)
5048 return p->prio - MAX_RT_PRIO;
5052 * task_nice - return the nice value of a given task.
5053 * @p: the task in question.
5055 int task_nice(const struct task_struct *p)
5057 return TASK_NICE(p);
5059 EXPORT_SYMBOL(task_nice);
5062 * idle_cpu - is a given cpu idle currently?
5063 * @cpu: the processor in question.
5065 int idle_cpu(int cpu)
5067 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5071 * idle_task - return the idle task for a given cpu.
5072 * @cpu: the processor in question.
5074 struct task_struct *idle_task(int cpu)
5076 return cpu_rq(cpu)->idle;
5080 * find_process_by_pid - find a process with a matching PID value.
5081 * @pid: the pid in question.
5083 static struct task_struct *find_process_by_pid(pid_t pid)
5085 return pid ? find_task_by_vpid(pid) : current;
5088 /* Actually do priority change: must hold rq lock. */
5090 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5092 BUG_ON(p->se.on_rq);
5095 switch (p->policy) {
5099 p->sched_class = &fair_sched_class;
5103 p->sched_class = &rt_sched_class;
5107 p->rt_priority = prio;
5108 p->normal_prio = normal_prio(p);
5109 /* we are holding p->pi_lock already */
5110 p->prio = rt_mutex_getprio(p);
5115 * check the target process has a UID that matches the current process's
5117 static bool check_same_owner(struct task_struct *p)
5119 const struct cred *cred = current_cred(), *pcred;
5123 pcred = __task_cred(p);
5124 match = (cred->euid == pcred->euid ||
5125 cred->euid == pcred->uid);
5130 static int __sched_setscheduler(struct task_struct *p, int policy,
5131 struct sched_param *param, bool user)
5133 int retval, oldprio, oldpolicy = -1, on_rq, running;
5134 unsigned long flags;
5135 const struct sched_class *prev_class = p->sched_class;
5138 /* may grab non-irq protected spin_locks */
5139 BUG_ON(in_interrupt());
5141 /* double check policy once rq lock held */
5143 policy = oldpolicy = p->policy;
5144 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5145 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5146 policy != SCHED_IDLE)
5149 * Valid priorities for SCHED_FIFO and SCHED_RR are
5150 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5151 * SCHED_BATCH and SCHED_IDLE is 0.
5153 if (param->sched_priority < 0 ||
5154 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5155 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5157 if (rt_policy(policy) != (param->sched_priority != 0))
5161 * Allow unprivileged RT tasks to decrease priority:
5163 if (user && !capable(CAP_SYS_NICE)) {
5164 if (rt_policy(policy)) {
5165 unsigned long rlim_rtprio;
5167 if (!lock_task_sighand(p, &flags))
5169 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5170 unlock_task_sighand(p, &flags);
5172 /* can't set/change the rt policy */
5173 if (policy != p->policy && !rlim_rtprio)
5176 /* can't increase priority */
5177 if (param->sched_priority > p->rt_priority &&
5178 param->sched_priority > rlim_rtprio)
5182 * Like positive nice levels, dont allow tasks to
5183 * move out of SCHED_IDLE either:
5185 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5188 /* can't change other user's priorities */
5189 if (!check_same_owner(p))
5194 #ifdef CONFIG_RT_GROUP_SCHED
5196 * Do not allow realtime tasks into groups that have no runtime
5199 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5200 task_group(p)->rt_bandwidth.rt_runtime == 0)
5204 retval = security_task_setscheduler(p, policy, param);
5210 * make sure no PI-waiters arrive (or leave) while we are
5211 * changing the priority of the task:
5213 spin_lock_irqsave(&p->pi_lock, flags);
5215 * To be able to change p->policy safely, the apropriate
5216 * runqueue lock must be held.
5218 rq = __task_rq_lock(p);
5219 /* recheck policy now with rq lock held */
5220 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5221 policy = oldpolicy = -1;
5222 __task_rq_unlock(rq);
5223 spin_unlock_irqrestore(&p->pi_lock, flags);
5226 update_rq_clock(rq);
5227 on_rq = p->se.on_rq;
5228 running = task_current(rq, p);
5230 deactivate_task(rq, p, 0);
5232 p->sched_class->put_prev_task(rq, p);
5235 __setscheduler(rq, p, policy, param->sched_priority);
5238 p->sched_class->set_curr_task(rq);
5240 activate_task(rq, p, 0);
5242 check_class_changed(rq, p, prev_class, oldprio, running);
5244 __task_rq_unlock(rq);
5245 spin_unlock_irqrestore(&p->pi_lock, flags);
5247 rt_mutex_adjust_pi(p);
5253 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5254 * @p: the task in question.
5255 * @policy: new policy.
5256 * @param: structure containing the new RT priority.
5258 * NOTE that the task may be already dead.
5260 int sched_setscheduler(struct task_struct *p, int policy,
5261 struct sched_param *param)
5263 return __sched_setscheduler(p, policy, param, true);
5265 EXPORT_SYMBOL_GPL(sched_setscheduler);
5268 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5269 * @p: the task in question.
5270 * @policy: new policy.
5271 * @param: structure containing the new RT priority.
5273 * Just like sched_setscheduler, only don't bother checking if the
5274 * current context has permission. For example, this is needed in
5275 * stop_machine(): we create temporary high priority worker threads,
5276 * but our caller might not have that capability.
5278 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5279 struct sched_param *param)
5281 return __sched_setscheduler(p, policy, param, false);
5285 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5287 struct sched_param lparam;
5288 struct task_struct *p;
5291 if (!param || pid < 0)
5293 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5298 p = find_process_by_pid(pid);
5300 retval = sched_setscheduler(p, policy, &lparam);
5307 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5308 * @pid: the pid in question.
5309 * @policy: new policy.
5310 * @param: structure containing the new RT priority.
5313 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5315 /* negative values for policy are not valid */
5319 return do_sched_setscheduler(pid, policy, param);
5323 * sys_sched_setparam - set/change the RT priority of a thread
5324 * @pid: the pid in question.
5325 * @param: structure containing the new RT priority.
5327 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5329 return do_sched_setscheduler(pid, -1, param);
5333 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5334 * @pid: the pid in question.
5336 asmlinkage long sys_sched_getscheduler(pid_t pid)
5338 struct task_struct *p;
5345 read_lock(&tasklist_lock);
5346 p = find_process_by_pid(pid);
5348 retval = security_task_getscheduler(p);
5352 read_unlock(&tasklist_lock);
5357 * sys_sched_getscheduler - get the RT priority of a thread
5358 * @pid: the pid in question.
5359 * @param: structure containing the RT priority.
5361 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5363 struct sched_param lp;
5364 struct task_struct *p;
5367 if (!param || pid < 0)
5370 read_lock(&tasklist_lock);
5371 p = find_process_by_pid(pid);
5376 retval = security_task_getscheduler(p);
5380 lp.sched_priority = p->rt_priority;
5381 read_unlock(&tasklist_lock);
5384 * This one might sleep, we cannot do it with a spinlock held ...
5386 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5391 read_unlock(&tasklist_lock);
5395 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5397 cpumask_t cpus_allowed;
5398 cpumask_t new_mask = *in_mask;
5399 struct task_struct *p;
5403 read_lock(&tasklist_lock);
5405 p = find_process_by_pid(pid);
5407 read_unlock(&tasklist_lock);
5413 * It is not safe to call set_cpus_allowed with the
5414 * tasklist_lock held. We will bump the task_struct's
5415 * usage count and then drop tasklist_lock.
5418 read_unlock(&tasklist_lock);
5421 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5424 retval = security_task_setscheduler(p, 0, NULL);
5428 cpuset_cpus_allowed(p, &cpus_allowed);
5429 cpus_and(new_mask, new_mask, cpus_allowed);
5431 retval = set_cpus_allowed_ptr(p, &new_mask);
5434 cpuset_cpus_allowed(p, &cpus_allowed);
5435 if (!cpus_subset(new_mask, cpus_allowed)) {
5437 * We must have raced with a concurrent cpuset
5438 * update. Just reset the cpus_allowed to the
5439 * cpuset's cpus_allowed
5441 new_mask = cpus_allowed;
5451 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5452 cpumask_t *new_mask)
5454 if (len < sizeof(cpumask_t)) {
5455 memset(new_mask, 0, sizeof(cpumask_t));
5456 } else if (len > sizeof(cpumask_t)) {
5457 len = sizeof(cpumask_t);
5459 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5463 * sys_sched_setaffinity - set the cpu affinity of a process
5464 * @pid: pid of the process
5465 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5466 * @user_mask_ptr: user-space pointer to the new cpu mask
5468 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5469 unsigned long __user *user_mask_ptr)
5474 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5478 return sched_setaffinity(pid, &new_mask);
5481 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5483 struct task_struct *p;
5487 read_lock(&tasklist_lock);
5490 p = find_process_by_pid(pid);
5494 retval = security_task_getscheduler(p);
5498 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5501 read_unlock(&tasklist_lock);
5508 * sys_sched_getaffinity - get the cpu affinity of a process
5509 * @pid: pid of the process
5510 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5511 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5513 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5514 unsigned long __user *user_mask_ptr)
5519 if (len < sizeof(cpumask_t))
5522 ret = sched_getaffinity(pid, &mask);
5526 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5529 return sizeof(cpumask_t);
5533 * sys_sched_yield - yield the current processor to other threads.
5535 * This function yields the current CPU to other tasks. If there are no
5536 * other threads running on this CPU then this function will return.
5538 asmlinkage long sys_sched_yield(void)
5540 struct rq *rq = this_rq_lock();
5542 schedstat_inc(rq, yld_count);
5543 current->sched_class->yield_task(rq);
5546 * Since we are going to call schedule() anyway, there's
5547 * no need to preempt or enable interrupts:
5549 __release(rq->lock);
5550 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5551 _raw_spin_unlock(&rq->lock);
5552 preempt_enable_no_resched();
5559 static void __cond_resched(void)
5561 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5562 __might_sleep(__FILE__, __LINE__);
5565 * The BKS might be reacquired before we have dropped
5566 * PREEMPT_ACTIVE, which could trigger a second
5567 * cond_resched() call.
5570 add_preempt_count(PREEMPT_ACTIVE);
5572 sub_preempt_count(PREEMPT_ACTIVE);
5573 } while (need_resched());
5576 int __sched _cond_resched(void)
5578 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5579 system_state == SYSTEM_RUNNING) {
5585 EXPORT_SYMBOL(_cond_resched);
5588 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5589 * call schedule, and on return reacquire the lock.
5591 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5592 * operations here to prevent schedule() from being called twice (once via
5593 * spin_unlock(), once by hand).
5595 int cond_resched_lock(spinlock_t *lock)
5597 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5600 if (spin_needbreak(lock) || resched) {
5602 if (resched && need_resched())
5611 EXPORT_SYMBOL(cond_resched_lock);
5613 int __sched cond_resched_softirq(void)
5615 BUG_ON(!in_softirq());
5617 if (need_resched() && system_state == SYSTEM_RUNNING) {
5625 EXPORT_SYMBOL(cond_resched_softirq);
5628 * yield - yield the current processor to other threads.
5630 * This is a shortcut for kernel-space yielding - it marks the
5631 * thread runnable and calls sys_sched_yield().
5633 void __sched yield(void)
5635 set_current_state(TASK_RUNNING);
5638 EXPORT_SYMBOL(yield);
5641 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5642 * that process accounting knows that this is a task in IO wait state.
5644 * But don't do that if it is a deliberate, throttling IO wait (this task
5645 * has set its backing_dev_info: the queue against which it should throttle)
5647 void __sched io_schedule(void)
5649 struct rq *rq = &__raw_get_cpu_var(runqueues);
5651 delayacct_blkio_start();
5652 atomic_inc(&rq->nr_iowait);
5654 atomic_dec(&rq->nr_iowait);
5655 delayacct_blkio_end();
5657 EXPORT_SYMBOL(io_schedule);
5659 long __sched io_schedule_timeout(long timeout)
5661 struct rq *rq = &__raw_get_cpu_var(runqueues);
5664 delayacct_blkio_start();
5665 atomic_inc(&rq->nr_iowait);
5666 ret = schedule_timeout(timeout);
5667 atomic_dec(&rq->nr_iowait);
5668 delayacct_blkio_end();
5673 * sys_sched_get_priority_max - return maximum RT priority.
5674 * @policy: scheduling class.
5676 * this syscall returns the maximum rt_priority that can be used
5677 * by a given scheduling class.
5679 asmlinkage long sys_sched_get_priority_max(int policy)
5686 ret = MAX_USER_RT_PRIO-1;
5698 * sys_sched_get_priority_min - return minimum RT priority.
5699 * @policy: scheduling class.
5701 * this syscall returns the minimum rt_priority that can be used
5702 * by a given scheduling class.
5704 asmlinkage long sys_sched_get_priority_min(int policy)
5722 * sys_sched_rr_get_interval - return the default timeslice of a process.
5723 * @pid: pid of the process.
5724 * @interval: userspace pointer to the timeslice value.
5726 * this syscall writes the default timeslice value of a given process
5727 * into the user-space timespec buffer. A value of '0' means infinity.
5730 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5732 struct task_struct *p;
5733 unsigned int time_slice;
5741 read_lock(&tasklist_lock);
5742 p = find_process_by_pid(pid);
5746 retval = security_task_getscheduler(p);
5751 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5752 * tasks that are on an otherwise idle runqueue:
5755 if (p->policy == SCHED_RR) {
5756 time_slice = DEF_TIMESLICE;
5757 } else if (p->policy != SCHED_FIFO) {
5758 struct sched_entity *se = &p->se;
5759 unsigned long flags;
5762 rq = task_rq_lock(p, &flags);
5763 if (rq->cfs.load.weight)
5764 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5765 task_rq_unlock(rq, &flags);
5767 read_unlock(&tasklist_lock);
5768 jiffies_to_timespec(time_slice, &t);
5769 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5773 read_unlock(&tasklist_lock);
5777 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5779 void sched_show_task(struct task_struct *p)
5781 unsigned long free = 0;
5784 state = p->state ? __ffs(p->state) + 1 : 0;
5785 printk(KERN_INFO "%-13.13s %c", p->comm,
5786 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5787 #if BITS_PER_LONG == 32
5788 if (state == TASK_RUNNING)
5789 printk(KERN_CONT " running ");
5791 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5793 if (state == TASK_RUNNING)
5794 printk(KERN_CONT " running task ");
5796 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5798 #ifdef CONFIG_DEBUG_STACK_USAGE
5800 unsigned long *n = end_of_stack(p);
5803 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5806 printk(KERN_CONT "%5lu %5d %6d\n", free,
5807 task_pid_nr(p), task_pid_nr(p->real_parent));
5809 show_stack(p, NULL);
5812 void show_state_filter(unsigned long state_filter)
5814 struct task_struct *g, *p;
5816 #if BITS_PER_LONG == 32
5818 " task PC stack pid father\n");
5821 " task PC stack pid father\n");
5823 read_lock(&tasklist_lock);
5824 do_each_thread(g, p) {
5826 * reset the NMI-timeout, listing all files on a slow
5827 * console might take alot of time:
5829 touch_nmi_watchdog();
5830 if (!state_filter || (p->state & state_filter))
5832 } while_each_thread(g, p);
5834 touch_all_softlockup_watchdogs();
5836 #ifdef CONFIG_SCHED_DEBUG
5837 sysrq_sched_debug_show();
5839 read_unlock(&tasklist_lock);
5841 * Only show locks if all tasks are dumped:
5843 if (state_filter == -1)
5844 debug_show_all_locks();
5847 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5849 idle->sched_class = &idle_sched_class;
5853 * init_idle - set up an idle thread for a given CPU
5854 * @idle: task in question
5855 * @cpu: cpu the idle task belongs to
5857 * NOTE: this function does not set the idle thread's NEED_RESCHED
5858 * flag, to make booting more robust.
5860 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5862 struct rq *rq = cpu_rq(cpu);
5863 unsigned long flags;
5865 spin_lock_irqsave(&rq->lock, flags);
5868 idle->se.exec_start = sched_clock();
5870 idle->prio = idle->normal_prio = MAX_PRIO;
5871 idle->cpus_allowed = cpumask_of_cpu(cpu);
5872 __set_task_cpu(idle, cpu);
5874 rq->curr = rq->idle = idle;
5875 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5878 spin_unlock_irqrestore(&rq->lock, flags);
5880 /* Set the preempt count _outside_ the spinlocks! */
5881 #if defined(CONFIG_PREEMPT)
5882 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5884 task_thread_info(idle)->preempt_count = 0;
5887 * The idle tasks have their own, simple scheduling class:
5889 idle->sched_class = &idle_sched_class;
5890 ftrace_graph_init_task(idle);
5894 * In a system that switches off the HZ timer nohz_cpu_mask
5895 * indicates which cpus entered this state. This is used
5896 * in the rcu update to wait only for active cpus. For system
5897 * which do not switch off the HZ timer nohz_cpu_mask should
5898 * always be CPU_MASK_NONE.
5900 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5903 * Increase the granularity value when there are more CPUs,
5904 * because with more CPUs the 'effective latency' as visible
5905 * to users decreases. But the relationship is not linear,
5906 * so pick a second-best guess by going with the log2 of the
5909 * This idea comes from the SD scheduler of Con Kolivas:
5911 static inline void sched_init_granularity(void)
5913 unsigned int factor = 1 + ilog2(num_online_cpus());
5914 const unsigned long limit = 200000000;
5916 sysctl_sched_min_granularity *= factor;
5917 if (sysctl_sched_min_granularity > limit)
5918 sysctl_sched_min_granularity = limit;
5920 sysctl_sched_latency *= factor;
5921 if (sysctl_sched_latency > limit)
5922 sysctl_sched_latency = limit;
5924 sysctl_sched_wakeup_granularity *= factor;
5926 sysctl_sched_shares_ratelimit *= factor;
5931 * This is how migration works:
5933 * 1) we queue a struct migration_req structure in the source CPU's
5934 * runqueue and wake up that CPU's migration thread.
5935 * 2) we down() the locked semaphore => thread blocks.
5936 * 3) migration thread wakes up (implicitly it forces the migrated
5937 * thread off the CPU)
5938 * 4) it gets the migration request and checks whether the migrated
5939 * task is still in the wrong runqueue.
5940 * 5) if it's in the wrong runqueue then the migration thread removes
5941 * it and puts it into the right queue.
5942 * 6) migration thread up()s the semaphore.
5943 * 7) we wake up and the migration is done.
5947 * Change a given task's CPU affinity. Migrate the thread to a
5948 * proper CPU and schedule it away if the CPU it's executing on
5949 * is removed from the allowed bitmask.
5951 * NOTE: the caller must have a valid reference to the task, the
5952 * task must not exit() & deallocate itself prematurely. The
5953 * call is not atomic; no spinlocks may be held.
5955 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5957 struct migration_req req;
5958 unsigned long flags;
5962 rq = task_rq_lock(p, &flags);
5963 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5968 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5969 !cpus_equal(p->cpus_allowed, *new_mask))) {
5974 if (p->sched_class->set_cpus_allowed)
5975 p->sched_class->set_cpus_allowed(p, new_mask);
5977 p->cpus_allowed = *new_mask;
5978 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5981 /* Can the task run on the task's current CPU? If so, we're done */
5982 if (cpu_isset(task_cpu(p), *new_mask))
5985 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5986 /* Need help from migration thread: drop lock and wait. */
5987 task_rq_unlock(rq, &flags);
5988 wake_up_process(rq->migration_thread);
5989 wait_for_completion(&req.done);
5990 tlb_migrate_finish(p->mm);
5994 task_rq_unlock(rq, &flags);
5998 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6001 * Move (not current) task off this cpu, onto dest cpu. We're doing
6002 * this because either it can't run here any more (set_cpus_allowed()
6003 * away from this CPU, or CPU going down), or because we're
6004 * attempting to rebalance this task on exec (sched_exec).
6006 * So we race with normal scheduler movements, but that's OK, as long
6007 * as the task is no longer on this CPU.
6009 * Returns non-zero if task was successfully migrated.
6011 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6013 struct rq *rq_dest, *rq_src;
6016 if (unlikely(!cpu_active(dest_cpu)))
6019 rq_src = cpu_rq(src_cpu);
6020 rq_dest = cpu_rq(dest_cpu);
6022 double_rq_lock(rq_src, rq_dest);
6023 /* Already moved. */
6024 if (task_cpu(p) != src_cpu)
6026 /* Affinity changed (again). */
6027 if (!cpu_isset(dest_cpu, p->cpus_allowed))
6030 on_rq = p->se.on_rq;
6032 deactivate_task(rq_src, p, 0);
6034 set_task_cpu(p, dest_cpu);
6036 activate_task(rq_dest, p, 0);
6037 check_preempt_curr(rq_dest, p, 0);
6042 double_rq_unlock(rq_src, rq_dest);
6047 * migration_thread - this is a highprio system thread that performs
6048 * thread migration by bumping thread off CPU then 'pushing' onto
6051 static int migration_thread(void *data)
6053 int cpu = (long)data;
6057 BUG_ON(rq->migration_thread != current);
6059 set_current_state(TASK_INTERRUPTIBLE);
6060 while (!kthread_should_stop()) {
6061 struct migration_req *req;
6062 struct list_head *head;
6064 spin_lock_irq(&rq->lock);
6066 if (cpu_is_offline(cpu)) {
6067 spin_unlock_irq(&rq->lock);
6071 if (rq->active_balance) {
6072 active_load_balance(rq, cpu);
6073 rq->active_balance = 0;
6076 head = &rq->migration_queue;
6078 if (list_empty(head)) {
6079 spin_unlock_irq(&rq->lock);
6081 set_current_state(TASK_INTERRUPTIBLE);
6084 req = list_entry(head->next, struct migration_req, list);
6085 list_del_init(head->next);
6087 spin_unlock(&rq->lock);
6088 __migrate_task(req->task, cpu, req->dest_cpu);
6091 complete(&req->done);
6093 __set_current_state(TASK_RUNNING);
6097 /* Wait for kthread_stop */
6098 set_current_state(TASK_INTERRUPTIBLE);
6099 while (!kthread_should_stop()) {
6101 set_current_state(TASK_INTERRUPTIBLE);
6103 __set_current_state(TASK_RUNNING);
6107 #ifdef CONFIG_HOTPLUG_CPU
6109 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6113 local_irq_disable();
6114 ret = __migrate_task(p, src_cpu, dest_cpu);
6120 * Figure out where task on dead CPU should go, use force if necessary.
6122 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6124 unsigned long flags;
6131 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6132 cpus_and(mask, mask, p->cpus_allowed);
6133 dest_cpu = any_online_cpu(mask);
6135 /* On any allowed CPU? */
6136 if (dest_cpu >= nr_cpu_ids)
6137 dest_cpu = any_online_cpu(p->cpus_allowed);
6139 /* No more Mr. Nice Guy. */
6140 if (dest_cpu >= nr_cpu_ids) {
6141 cpumask_t cpus_allowed;
6143 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6145 * Try to stay on the same cpuset, where the
6146 * current cpuset may be a subset of all cpus.
6147 * The cpuset_cpus_allowed_locked() variant of
6148 * cpuset_cpus_allowed() will not block. It must be
6149 * called within calls to cpuset_lock/cpuset_unlock.
6151 rq = task_rq_lock(p, &flags);
6152 p->cpus_allowed = cpus_allowed;
6153 dest_cpu = any_online_cpu(p->cpus_allowed);
6154 task_rq_unlock(rq, &flags);
6157 * Don't tell them about moving exiting tasks or
6158 * kernel threads (both mm NULL), since they never
6161 if (p->mm && printk_ratelimit()) {
6162 printk(KERN_INFO "process %d (%s) no "
6163 "longer affine to cpu%d\n",
6164 task_pid_nr(p), p->comm, dead_cpu);
6167 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6171 * While a dead CPU has no uninterruptible tasks queued at this point,
6172 * it might still have a nonzero ->nr_uninterruptible counter, because
6173 * for performance reasons the counter is not stricly tracking tasks to
6174 * their home CPUs. So we just add the counter to another CPU's counter,
6175 * to keep the global sum constant after CPU-down:
6177 static void migrate_nr_uninterruptible(struct rq *rq_src)
6179 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6180 unsigned long flags;
6182 local_irq_save(flags);
6183 double_rq_lock(rq_src, rq_dest);
6184 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6185 rq_src->nr_uninterruptible = 0;
6186 double_rq_unlock(rq_src, rq_dest);
6187 local_irq_restore(flags);
6190 /* Run through task list and migrate tasks from the dead cpu. */
6191 static void migrate_live_tasks(int src_cpu)
6193 struct task_struct *p, *t;
6195 read_lock(&tasklist_lock);
6197 do_each_thread(t, p) {
6201 if (task_cpu(p) == src_cpu)
6202 move_task_off_dead_cpu(src_cpu, p);
6203 } while_each_thread(t, p);
6205 read_unlock(&tasklist_lock);
6209 * Schedules idle task to be the next runnable task on current CPU.
6210 * It does so by boosting its priority to highest possible.
6211 * Used by CPU offline code.
6213 void sched_idle_next(void)
6215 int this_cpu = smp_processor_id();
6216 struct rq *rq = cpu_rq(this_cpu);
6217 struct task_struct *p = rq->idle;
6218 unsigned long flags;
6220 /* cpu has to be offline */
6221 BUG_ON(cpu_online(this_cpu));
6224 * Strictly not necessary since rest of the CPUs are stopped by now
6225 * and interrupts disabled on the current cpu.
6227 spin_lock_irqsave(&rq->lock, flags);
6229 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6231 update_rq_clock(rq);
6232 activate_task(rq, p, 0);
6234 spin_unlock_irqrestore(&rq->lock, flags);
6238 * Ensures that the idle task is using init_mm right before its cpu goes
6241 void idle_task_exit(void)
6243 struct mm_struct *mm = current->active_mm;
6245 BUG_ON(cpu_online(smp_processor_id()));
6248 switch_mm(mm, &init_mm, current);
6252 /* called under rq->lock with disabled interrupts */
6253 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6255 struct rq *rq = cpu_rq(dead_cpu);
6257 /* Must be exiting, otherwise would be on tasklist. */
6258 BUG_ON(!p->exit_state);
6260 /* Cannot have done final schedule yet: would have vanished. */
6261 BUG_ON(p->state == TASK_DEAD);
6266 * Drop lock around migration; if someone else moves it,
6267 * that's OK. No task can be added to this CPU, so iteration is
6270 spin_unlock_irq(&rq->lock);
6271 move_task_off_dead_cpu(dead_cpu, p);
6272 spin_lock_irq(&rq->lock);
6277 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6278 static void migrate_dead_tasks(unsigned int dead_cpu)
6280 struct rq *rq = cpu_rq(dead_cpu);
6281 struct task_struct *next;
6284 if (!rq->nr_running)
6286 update_rq_clock(rq);
6287 next = pick_next_task(rq, rq->curr);
6290 next->sched_class->put_prev_task(rq, next);
6291 migrate_dead(dead_cpu, next);
6295 #endif /* CONFIG_HOTPLUG_CPU */
6297 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6299 static struct ctl_table sd_ctl_dir[] = {
6301 .procname = "sched_domain",
6307 static struct ctl_table sd_ctl_root[] = {
6309 .ctl_name = CTL_KERN,
6310 .procname = "kernel",
6312 .child = sd_ctl_dir,
6317 static struct ctl_table *sd_alloc_ctl_entry(int n)
6319 struct ctl_table *entry =
6320 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6325 static void sd_free_ctl_entry(struct ctl_table **tablep)
6327 struct ctl_table *entry;
6330 * In the intermediate directories, both the child directory and
6331 * procname are dynamically allocated and could fail but the mode
6332 * will always be set. In the lowest directory the names are
6333 * static strings and all have proc handlers.
6335 for (entry = *tablep; entry->mode; entry++) {
6337 sd_free_ctl_entry(&entry->child);
6338 if (entry->proc_handler == NULL)
6339 kfree(entry->procname);
6347 set_table_entry(struct ctl_table *entry,
6348 const char *procname, void *data, int maxlen,
6349 mode_t mode, proc_handler *proc_handler)
6351 entry->procname = procname;
6353 entry->maxlen = maxlen;
6355 entry->proc_handler = proc_handler;
6358 static struct ctl_table *
6359 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6361 struct ctl_table *table = sd_alloc_ctl_entry(13);
6366 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6367 sizeof(long), 0644, proc_doulongvec_minmax);
6368 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6369 sizeof(long), 0644, proc_doulongvec_minmax);
6370 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6371 sizeof(int), 0644, proc_dointvec_minmax);
6372 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6373 sizeof(int), 0644, proc_dointvec_minmax);
6374 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6375 sizeof(int), 0644, proc_dointvec_minmax);
6376 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6377 sizeof(int), 0644, proc_dointvec_minmax);
6378 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6379 sizeof(int), 0644, proc_dointvec_minmax);
6380 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6381 sizeof(int), 0644, proc_dointvec_minmax);
6382 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6383 sizeof(int), 0644, proc_dointvec_minmax);
6384 set_table_entry(&table[9], "cache_nice_tries",
6385 &sd->cache_nice_tries,
6386 sizeof(int), 0644, proc_dointvec_minmax);
6387 set_table_entry(&table[10], "flags", &sd->flags,
6388 sizeof(int), 0644, proc_dointvec_minmax);
6389 set_table_entry(&table[11], "name", sd->name,
6390 CORENAME_MAX_SIZE, 0444, proc_dostring);
6391 /* &table[12] is terminator */
6396 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6398 struct ctl_table *entry, *table;
6399 struct sched_domain *sd;
6400 int domain_num = 0, i;
6403 for_each_domain(cpu, sd)
6405 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6410 for_each_domain(cpu, sd) {
6411 snprintf(buf, 32, "domain%d", i);
6412 entry->procname = kstrdup(buf, GFP_KERNEL);
6414 entry->child = sd_alloc_ctl_domain_table(sd);
6421 static struct ctl_table_header *sd_sysctl_header;
6422 static void register_sched_domain_sysctl(void)
6424 int i, cpu_num = num_online_cpus();
6425 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6428 WARN_ON(sd_ctl_dir[0].child);
6429 sd_ctl_dir[0].child = entry;
6434 for_each_online_cpu(i) {
6435 snprintf(buf, 32, "cpu%d", i);
6436 entry->procname = kstrdup(buf, GFP_KERNEL);
6438 entry->child = sd_alloc_ctl_cpu_table(i);
6442 WARN_ON(sd_sysctl_header);
6443 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6446 /* may be called multiple times per register */
6447 static void unregister_sched_domain_sysctl(void)
6449 if (sd_sysctl_header)
6450 unregister_sysctl_table(sd_sysctl_header);
6451 sd_sysctl_header = NULL;
6452 if (sd_ctl_dir[0].child)
6453 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6456 static void register_sched_domain_sysctl(void)
6459 static void unregister_sched_domain_sysctl(void)
6464 static void set_rq_online(struct rq *rq)
6467 const struct sched_class *class;
6469 cpu_set(rq->cpu, rq->rd->online);
6472 for_each_class(class) {
6473 if (class->rq_online)
6474 class->rq_online(rq);
6479 static void set_rq_offline(struct rq *rq)
6482 const struct sched_class *class;
6484 for_each_class(class) {
6485 if (class->rq_offline)
6486 class->rq_offline(rq);
6489 cpu_clear(rq->cpu, rq->rd->online);
6495 * migration_call - callback that gets triggered when a CPU is added.
6496 * Here we can start up the necessary migration thread for the new CPU.
6498 static int __cpuinit
6499 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6501 struct task_struct *p;
6502 int cpu = (long)hcpu;
6503 unsigned long flags;
6508 case CPU_UP_PREPARE:
6509 case CPU_UP_PREPARE_FROZEN:
6510 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6513 kthread_bind(p, cpu);
6514 /* Must be high prio: stop_machine expects to yield to it. */
6515 rq = task_rq_lock(p, &flags);
6516 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6517 task_rq_unlock(rq, &flags);
6518 cpu_rq(cpu)->migration_thread = p;
6522 case CPU_ONLINE_FROZEN:
6523 /* Strictly unnecessary, as first user will wake it. */
6524 wake_up_process(cpu_rq(cpu)->migration_thread);
6526 /* Update our root-domain */
6528 spin_lock_irqsave(&rq->lock, flags);
6530 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6534 spin_unlock_irqrestore(&rq->lock, flags);
6537 #ifdef CONFIG_HOTPLUG_CPU
6538 case CPU_UP_CANCELED:
6539 case CPU_UP_CANCELED_FROZEN:
6540 if (!cpu_rq(cpu)->migration_thread)
6542 /* Unbind it from offline cpu so it can run. Fall thru. */
6543 kthread_bind(cpu_rq(cpu)->migration_thread,
6544 any_online_cpu(cpu_online_map));
6545 kthread_stop(cpu_rq(cpu)->migration_thread);
6546 cpu_rq(cpu)->migration_thread = NULL;
6550 case CPU_DEAD_FROZEN:
6551 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6552 migrate_live_tasks(cpu);
6554 kthread_stop(rq->migration_thread);
6555 rq->migration_thread = NULL;
6556 /* Idle task back to normal (off runqueue, low prio) */
6557 spin_lock_irq(&rq->lock);
6558 update_rq_clock(rq);
6559 deactivate_task(rq, rq->idle, 0);
6560 rq->idle->static_prio = MAX_PRIO;
6561 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6562 rq->idle->sched_class = &idle_sched_class;
6563 migrate_dead_tasks(cpu);
6564 spin_unlock_irq(&rq->lock);
6566 migrate_nr_uninterruptible(rq);
6567 BUG_ON(rq->nr_running != 0);
6570 * No need to migrate the tasks: it was best-effort if
6571 * they didn't take sched_hotcpu_mutex. Just wake up
6574 spin_lock_irq(&rq->lock);
6575 while (!list_empty(&rq->migration_queue)) {
6576 struct migration_req *req;
6578 req = list_entry(rq->migration_queue.next,
6579 struct migration_req, list);
6580 list_del_init(&req->list);
6581 spin_unlock_irq(&rq->lock);
6582 complete(&req->done);
6583 spin_lock_irq(&rq->lock);
6585 spin_unlock_irq(&rq->lock);
6589 case CPU_DYING_FROZEN:
6590 /* Update our root-domain */
6592 spin_lock_irqsave(&rq->lock, flags);
6594 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6597 spin_unlock_irqrestore(&rq->lock, flags);
6604 /* Register at highest priority so that task migration (migrate_all_tasks)
6605 * happens before everything else.
6607 static struct notifier_block __cpuinitdata migration_notifier = {
6608 .notifier_call = migration_call,
6612 static int __init migration_init(void)
6614 void *cpu = (void *)(long)smp_processor_id();
6617 /* Start one for the boot CPU: */
6618 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6619 BUG_ON(err == NOTIFY_BAD);
6620 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6621 register_cpu_notifier(&migration_notifier);
6625 early_initcall(migration_init);
6630 #ifdef CONFIG_SCHED_DEBUG
6632 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6633 cpumask_t *groupmask)
6635 struct sched_group *group = sd->groups;
6638 cpulist_scnprintf(str, sizeof(str), sd->span);
6639 cpus_clear(*groupmask);
6641 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6643 if (!(sd->flags & SD_LOAD_BALANCE)) {
6644 printk("does not load-balance\n");
6646 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6651 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6653 if (!cpu_isset(cpu, sd->span)) {
6654 printk(KERN_ERR "ERROR: domain->span does not contain "
6657 if (!cpu_isset(cpu, group->cpumask)) {
6658 printk(KERN_ERR "ERROR: domain->groups does not contain"
6662 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6666 printk(KERN_ERR "ERROR: group is NULL\n");
6670 if (!group->__cpu_power) {
6671 printk(KERN_CONT "\n");
6672 printk(KERN_ERR "ERROR: domain->cpu_power not "
6677 if (!cpus_weight(group->cpumask)) {
6678 printk(KERN_CONT "\n");
6679 printk(KERN_ERR "ERROR: empty group\n");
6683 if (cpus_intersects(*groupmask, group->cpumask)) {
6684 printk(KERN_CONT "\n");
6685 printk(KERN_ERR "ERROR: repeated CPUs\n");
6689 cpus_or(*groupmask, *groupmask, group->cpumask);
6691 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6692 printk(KERN_CONT " %s", str);
6694 group = group->next;
6695 } while (group != sd->groups);
6696 printk(KERN_CONT "\n");
6698 if (!cpus_equal(sd->span, *groupmask))
6699 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6701 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6702 printk(KERN_ERR "ERROR: parent span is not a superset "
6703 "of domain->span\n");
6707 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6709 cpumask_t *groupmask;
6713 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6717 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6719 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6721 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6726 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6735 #else /* !CONFIG_SCHED_DEBUG */
6736 # define sched_domain_debug(sd, cpu) do { } while (0)
6737 #endif /* CONFIG_SCHED_DEBUG */
6739 static int sd_degenerate(struct sched_domain *sd)
6741 if (cpus_weight(sd->span) == 1)
6744 /* Following flags need at least 2 groups */
6745 if (sd->flags & (SD_LOAD_BALANCE |
6746 SD_BALANCE_NEWIDLE |
6750 SD_SHARE_PKG_RESOURCES)) {
6751 if (sd->groups != sd->groups->next)
6755 /* Following flags don't use groups */
6756 if (sd->flags & (SD_WAKE_IDLE |
6765 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6767 unsigned long cflags = sd->flags, pflags = parent->flags;
6769 if (sd_degenerate(parent))
6772 if (!cpus_equal(sd->span, parent->span))
6775 /* Does parent contain flags not in child? */
6776 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6777 if (cflags & SD_WAKE_AFFINE)
6778 pflags &= ~SD_WAKE_BALANCE;
6779 /* Flags needing groups don't count if only 1 group in parent */
6780 if (parent->groups == parent->groups->next) {
6781 pflags &= ~(SD_LOAD_BALANCE |
6782 SD_BALANCE_NEWIDLE |
6786 SD_SHARE_PKG_RESOURCES);
6787 if (nr_node_ids == 1)
6788 pflags &= ~SD_SERIALIZE;
6790 if (~cflags & pflags)
6796 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6798 unsigned long flags;
6800 spin_lock_irqsave(&rq->lock, flags);
6803 struct root_domain *old_rd = rq->rd;
6805 if (cpu_isset(rq->cpu, old_rd->online))
6808 cpu_clear(rq->cpu, old_rd->span);
6810 if (atomic_dec_and_test(&old_rd->refcount))
6814 atomic_inc(&rd->refcount);
6817 cpu_set(rq->cpu, rd->span);
6818 if (cpu_isset(rq->cpu, cpu_online_map))
6821 spin_unlock_irqrestore(&rq->lock, flags);
6824 static void init_rootdomain(struct root_domain *rd)
6826 memset(rd, 0, sizeof(*rd));
6828 cpus_clear(rd->span);
6829 cpus_clear(rd->online);
6831 cpupri_init(&rd->cpupri);
6834 static void init_defrootdomain(void)
6836 init_rootdomain(&def_root_domain);
6837 atomic_set(&def_root_domain.refcount, 1);
6840 static struct root_domain *alloc_rootdomain(void)
6842 struct root_domain *rd;
6844 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6848 init_rootdomain(rd);
6854 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6855 * hold the hotplug lock.
6858 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6860 struct rq *rq = cpu_rq(cpu);
6861 struct sched_domain *tmp;
6863 /* Remove the sched domains which do not contribute to scheduling. */
6864 for (tmp = sd; tmp; ) {
6865 struct sched_domain *parent = tmp->parent;
6869 if (sd_parent_degenerate(tmp, parent)) {
6870 tmp->parent = parent->parent;
6872 parent->parent->child = tmp;
6877 if (sd && sd_degenerate(sd)) {
6883 sched_domain_debug(sd, cpu);
6885 rq_attach_root(rq, rd);
6886 rcu_assign_pointer(rq->sd, sd);
6889 /* cpus with isolated domains */
6890 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6892 /* Setup the mask of cpus configured for isolated domains */
6893 static int __init isolated_cpu_setup(char *str)
6895 static int __initdata ints[NR_CPUS];
6898 str = get_options(str, ARRAY_SIZE(ints), ints);
6899 cpus_clear(cpu_isolated_map);
6900 for (i = 1; i <= ints[0]; i++)
6901 if (ints[i] < NR_CPUS)
6902 cpu_set(ints[i], cpu_isolated_map);
6906 __setup("isolcpus=", isolated_cpu_setup);
6909 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6910 * to a function which identifies what group(along with sched group) a CPU
6911 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6912 * (due to the fact that we keep track of groups covered with a cpumask_t).
6914 * init_sched_build_groups will build a circular linked list of the groups
6915 * covered by the given span, and will set each group's ->cpumask correctly,
6916 * and ->cpu_power to 0.
6919 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6920 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6921 struct sched_group **sg,
6922 cpumask_t *tmpmask),
6923 cpumask_t *covered, cpumask_t *tmpmask)
6925 struct sched_group *first = NULL, *last = NULL;
6928 cpus_clear(*covered);
6930 for_each_cpu_mask_nr(i, *span) {
6931 struct sched_group *sg;
6932 int group = group_fn(i, cpu_map, &sg, tmpmask);
6935 if (cpu_isset(i, *covered))
6938 cpus_clear(sg->cpumask);
6939 sg->__cpu_power = 0;
6941 for_each_cpu_mask_nr(j, *span) {
6942 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6945 cpu_set(j, *covered);
6946 cpu_set(j, sg->cpumask);
6957 #define SD_NODES_PER_DOMAIN 16
6962 * find_next_best_node - find the next node to include in a sched_domain
6963 * @node: node whose sched_domain we're building
6964 * @used_nodes: nodes already in the sched_domain
6966 * Find the next node to include in a given scheduling domain. Simply
6967 * finds the closest node not already in the @used_nodes map.
6969 * Should use nodemask_t.
6971 static int find_next_best_node(int node, nodemask_t *used_nodes)
6973 int i, n, val, min_val, best_node = 0;
6977 for (i = 0; i < nr_node_ids; i++) {
6978 /* Start at @node */
6979 n = (node + i) % nr_node_ids;
6981 if (!nr_cpus_node(n))
6984 /* Skip already used nodes */
6985 if (node_isset(n, *used_nodes))
6988 /* Simple min distance search */
6989 val = node_distance(node, n);
6991 if (val < min_val) {
6997 node_set(best_node, *used_nodes);
7002 * sched_domain_node_span - get a cpumask for a node's sched_domain
7003 * @node: node whose cpumask we're constructing
7004 * @span: resulting cpumask
7006 * Given a node, construct a good cpumask for its sched_domain to span. It
7007 * should be one that prevents unnecessary balancing, but also spreads tasks
7010 static void sched_domain_node_span(int node, cpumask_t *span)
7012 nodemask_t used_nodes;
7013 node_to_cpumask_ptr(nodemask, node);
7017 nodes_clear(used_nodes);
7019 cpus_or(*span, *span, *nodemask);
7020 node_set(node, used_nodes);
7022 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7023 int next_node = find_next_best_node(node, &used_nodes);
7025 node_to_cpumask_ptr_next(nodemask, next_node);
7026 cpus_or(*span, *span, *nodemask);
7029 #endif /* CONFIG_NUMA */
7031 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7034 * SMT sched-domains:
7036 #ifdef CONFIG_SCHED_SMT
7037 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
7038 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
7041 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7045 *sg = &per_cpu(sched_group_cpus, cpu);
7048 #endif /* CONFIG_SCHED_SMT */
7051 * multi-core sched-domains:
7053 #ifdef CONFIG_SCHED_MC
7054 static DEFINE_PER_CPU(struct sched_domain, core_domains);
7055 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
7056 #endif /* CONFIG_SCHED_MC */
7058 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7060 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7065 *mask = per_cpu(cpu_sibling_map, cpu);
7066 cpus_and(*mask, *mask, *cpu_map);
7067 group = first_cpu(*mask);
7069 *sg = &per_cpu(sched_group_core, group);
7072 #elif defined(CONFIG_SCHED_MC)
7074 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7078 *sg = &per_cpu(sched_group_core, cpu);
7083 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
7084 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
7087 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7091 #ifdef CONFIG_SCHED_MC
7092 *mask = cpu_coregroup_map(cpu);
7093 cpus_and(*mask, *mask, *cpu_map);
7094 group = first_cpu(*mask);
7095 #elif defined(CONFIG_SCHED_SMT)
7096 *mask = per_cpu(cpu_sibling_map, cpu);
7097 cpus_and(*mask, *mask, *cpu_map);
7098 group = first_cpu(*mask);
7103 *sg = &per_cpu(sched_group_phys, group);
7109 * The init_sched_build_groups can't handle what we want to do with node
7110 * groups, so roll our own. Now each node has its own list of groups which
7111 * gets dynamically allocated.
7113 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7114 static struct sched_group ***sched_group_nodes_bycpu;
7116 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7117 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7119 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7120 struct sched_group **sg, cpumask_t *nodemask)
7124 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7125 cpus_and(*nodemask, *nodemask, *cpu_map);
7126 group = first_cpu(*nodemask);
7129 *sg = &per_cpu(sched_group_allnodes, group);
7133 static void init_numa_sched_groups_power(struct sched_group *group_head)
7135 struct sched_group *sg = group_head;
7141 for_each_cpu_mask_nr(j, sg->cpumask) {
7142 struct sched_domain *sd;
7144 sd = &per_cpu(phys_domains, j);
7145 if (j != first_cpu(sd->groups->cpumask)) {
7147 * Only add "power" once for each
7153 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7156 } while (sg != group_head);
7158 #endif /* CONFIG_NUMA */
7161 /* Free memory allocated for various sched_group structures */
7162 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7166 for_each_cpu_mask_nr(cpu, *cpu_map) {
7167 struct sched_group **sched_group_nodes
7168 = sched_group_nodes_bycpu[cpu];
7170 if (!sched_group_nodes)
7173 for (i = 0; i < nr_node_ids; i++) {
7174 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7176 *nodemask = node_to_cpumask(i);
7177 cpus_and(*nodemask, *nodemask, *cpu_map);
7178 if (cpus_empty(*nodemask))
7188 if (oldsg != sched_group_nodes[i])
7191 kfree(sched_group_nodes);
7192 sched_group_nodes_bycpu[cpu] = NULL;
7195 #else /* !CONFIG_NUMA */
7196 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7199 #endif /* CONFIG_NUMA */
7202 * Initialize sched groups cpu_power.
7204 * cpu_power indicates the capacity of sched group, which is used while
7205 * distributing the load between different sched groups in a sched domain.
7206 * Typically cpu_power for all the groups in a sched domain will be same unless
7207 * there are asymmetries in the topology. If there are asymmetries, group
7208 * having more cpu_power will pickup more load compared to the group having
7211 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7212 * the maximum number of tasks a group can handle in the presence of other idle
7213 * or lightly loaded groups in the same sched domain.
7215 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7217 struct sched_domain *child;
7218 struct sched_group *group;
7220 WARN_ON(!sd || !sd->groups);
7222 if (cpu != first_cpu(sd->groups->cpumask))
7227 sd->groups->__cpu_power = 0;
7230 * For perf policy, if the groups in child domain share resources
7231 * (for example cores sharing some portions of the cache hierarchy
7232 * or SMT), then set this domain groups cpu_power such that each group
7233 * can handle only one task, when there are other idle groups in the
7234 * same sched domain.
7236 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7238 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7239 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7244 * add cpu_power of each child group to this groups cpu_power
7246 group = child->groups;
7248 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7249 group = group->next;
7250 } while (group != child->groups);
7254 * Initializers for schedule domains
7255 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7258 #ifdef CONFIG_SCHED_DEBUG
7259 # define SD_INIT_NAME(sd, type) sd->name = #type
7261 # define SD_INIT_NAME(sd, type) do { } while (0)
7264 #define SD_INIT(sd, type) sd_init_##type(sd)
7266 #define SD_INIT_FUNC(type) \
7267 static noinline void sd_init_##type(struct sched_domain *sd) \
7269 memset(sd, 0, sizeof(*sd)); \
7270 *sd = SD_##type##_INIT; \
7271 sd->level = SD_LV_##type; \
7272 SD_INIT_NAME(sd, type); \
7277 SD_INIT_FUNC(ALLNODES)
7280 #ifdef CONFIG_SCHED_SMT
7281 SD_INIT_FUNC(SIBLING)
7283 #ifdef CONFIG_SCHED_MC
7288 * To minimize stack usage kmalloc room for cpumasks and share the
7289 * space as the usage in build_sched_domains() dictates. Used only
7290 * if the amount of space is significant.
7293 cpumask_t tmpmask; /* make this one first */
7296 cpumask_t this_sibling_map;
7297 cpumask_t this_core_map;
7299 cpumask_t send_covered;
7302 cpumask_t domainspan;
7304 cpumask_t notcovered;
7309 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7310 static inline void sched_cpumask_alloc(struct allmasks **masks)
7312 *masks = kmalloc(sizeof(**masks), GFP_KERNEL);
7314 static inline void sched_cpumask_free(struct allmasks *masks)
7319 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7320 static inline void sched_cpumask_alloc(struct allmasks **masks)
7322 static inline void sched_cpumask_free(struct allmasks *masks)
7326 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7327 ((unsigned long)(a) + offsetof(struct allmasks, v))
7329 static int default_relax_domain_level = -1;
7331 static int __init setup_relax_domain_level(char *str)
7335 val = simple_strtoul(str, NULL, 0);
7336 if (val < SD_LV_MAX)
7337 default_relax_domain_level = val;
7341 __setup("relax_domain_level=", setup_relax_domain_level);
7343 static void set_domain_attribute(struct sched_domain *sd,
7344 struct sched_domain_attr *attr)
7348 if (!attr || attr->relax_domain_level < 0) {
7349 if (default_relax_domain_level < 0)
7352 request = default_relax_domain_level;
7354 request = attr->relax_domain_level;
7355 if (request < sd->level) {
7356 /* turn off idle balance on this domain */
7357 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7359 /* turn on idle balance on this domain */
7360 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7365 * Build sched domains for a given set of cpus and attach the sched domains
7366 * to the individual cpus
7368 static int __build_sched_domains(const cpumask_t *cpu_map,
7369 struct sched_domain_attr *attr)
7372 struct root_domain *rd;
7373 SCHED_CPUMASK_DECLARE(allmasks);
7376 struct sched_group **sched_group_nodes = NULL;
7377 int sd_allnodes = 0;
7380 * Allocate the per-node list of sched groups
7382 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7384 if (!sched_group_nodes) {
7385 printk(KERN_WARNING "Can not alloc sched group node list\n");
7390 rd = alloc_rootdomain();
7392 printk(KERN_WARNING "Cannot alloc root domain\n");
7394 kfree(sched_group_nodes);
7399 /* get space for all scratch cpumask variables */
7400 sched_cpumask_alloc(&allmasks);
7402 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7405 kfree(sched_group_nodes);
7410 tmpmask = (cpumask_t *)allmasks;
7414 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7418 * Set up domains for cpus specified by the cpu_map.
7420 for_each_cpu_mask_nr(i, *cpu_map) {
7421 struct sched_domain *sd = NULL, *p;
7422 SCHED_CPUMASK_VAR(nodemask, allmasks);
7424 *nodemask = node_to_cpumask(cpu_to_node(i));
7425 cpus_and(*nodemask, *nodemask, *cpu_map);
7428 if (cpus_weight(*cpu_map) >
7429 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7430 sd = &per_cpu(allnodes_domains, i);
7431 SD_INIT(sd, ALLNODES);
7432 set_domain_attribute(sd, attr);
7433 sd->span = *cpu_map;
7434 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7440 sd = &per_cpu(node_domains, i);
7442 set_domain_attribute(sd, attr);
7443 sched_domain_node_span(cpu_to_node(i), &sd->span);
7447 cpus_and(sd->span, sd->span, *cpu_map);
7451 sd = &per_cpu(phys_domains, i);
7453 set_domain_attribute(sd, attr);
7454 sd->span = *nodemask;
7458 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7460 #ifdef CONFIG_SCHED_MC
7462 sd = &per_cpu(core_domains, i);
7464 set_domain_attribute(sd, attr);
7465 sd->span = cpu_coregroup_map(i);
7466 cpus_and(sd->span, sd->span, *cpu_map);
7469 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7472 #ifdef CONFIG_SCHED_SMT
7474 sd = &per_cpu(cpu_domains, i);
7475 SD_INIT(sd, SIBLING);
7476 set_domain_attribute(sd, attr);
7477 sd->span = per_cpu(cpu_sibling_map, i);
7478 cpus_and(sd->span, sd->span, *cpu_map);
7481 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7485 #ifdef CONFIG_SCHED_SMT
7486 /* Set up CPU (sibling) groups */
7487 for_each_cpu_mask_nr(i, *cpu_map) {
7488 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7489 SCHED_CPUMASK_VAR(send_covered, allmasks);
7491 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7492 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7493 if (i != first_cpu(*this_sibling_map))
7496 init_sched_build_groups(this_sibling_map, cpu_map,
7498 send_covered, tmpmask);
7502 #ifdef CONFIG_SCHED_MC
7503 /* Set up multi-core groups */
7504 for_each_cpu_mask_nr(i, *cpu_map) {
7505 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7506 SCHED_CPUMASK_VAR(send_covered, allmasks);
7508 *this_core_map = cpu_coregroup_map(i);
7509 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7510 if (i != first_cpu(*this_core_map))
7513 init_sched_build_groups(this_core_map, cpu_map,
7515 send_covered, tmpmask);
7519 /* Set up physical groups */
7520 for (i = 0; i < nr_node_ids; i++) {
7521 SCHED_CPUMASK_VAR(nodemask, allmasks);
7522 SCHED_CPUMASK_VAR(send_covered, allmasks);
7524 *nodemask = node_to_cpumask(i);
7525 cpus_and(*nodemask, *nodemask, *cpu_map);
7526 if (cpus_empty(*nodemask))
7529 init_sched_build_groups(nodemask, cpu_map,
7531 send_covered, tmpmask);
7535 /* Set up node groups */
7537 SCHED_CPUMASK_VAR(send_covered, allmasks);
7539 init_sched_build_groups(cpu_map, cpu_map,
7540 &cpu_to_allnodes_group,
7541 send_covered, tmpmask);
7544 for (i = 0; i < nr_node_ids; i++) {
7545 /* Set up node groups */
7546 struct sched_group *sg, *prev;
7547 SCHED_CPUMASK_VAR(nodemask, allmasks);
7548 SCHED_CPUMASK_VAR(domainspan, allmasks);
7549 SCHED_CPUMASK_VAR(covered, allmasks);
7552 *nodemask = node_to_cpumask(i);
7553 cpus_clear(*covered);
7555 cpus_and(*nodemask, *nodemask, *cpu_map);
7556 if (cpus_empty(*nodemask)) {
7557 sched_group_nodes[i] = NULL;
7561 sched_domain_node_span(i, domainspan);
7562 cpus_and(*domainspan, *domainspan, *cpu_map);
7564 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7566 printk(KERN_WARNING "Can not alloc domain group for "
7570 sched_group_nodes[i] = sg;
7571 for_each_cpu_mask_nr(j, *nodemask) {
7572 struct sched_domain *sd;
7574 sd = &per_cpu(node_domains, j);
7577 sg->__cpu_power = 0;
7578 sg->cpumask = *nodemask;
7580 cpus_or(*covered, *covered, *nodemask);
7583 for (j = 0; j < nr_node_ids; j++) {
7584 SCHED_CPUMASK_VAR(notcovered, allmasks);
7585 int n = (i + j) % nr_node_ids;
7586 node_to_cpumask_ptr(pnodemask, n);
7588 cpus_complement(*notcovered, *covered);
7589 cpus_and(*tmpmask, *notcovered, *cpu_map);
7590 cpus_and(*tmpmask, *tmpmask, *domainspan);
7591 if (cpus_empty(*tmpmask))
7594 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7595 if (cpus_empty(*tmpmask))
7598 sg = kmalloc_node(sizeof(struct sched_group),
7602 "Can not alloc domain group for node %d\n", j);
7605 sg->__cpu_power = 0;
7606 sg->cpumask = *tmpmask;
7607 sg->next = prev->next;
7608 cpus_or(*covered, *covered, *tmpmask);
7615 /* Calculate CPU power for physical packages and nodes */
7616 #ifdef CONFIG_SCHED_SMT
7617 for_each_cpu_mask_nr(i, *cpu_map) {
7618 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7620 init_sched_groups_power(i, sd);
7623 #ifdef CONFIG_SCHED_MC
7624 for_each_cpu_mask_nr(i, *cpu_map) {
7625 struct sched_domain *sd = &per_cpu(core_domains, i);
7627 init_sched_groups_power(i, sd);
7631 for_each_cpu_mask_nr(i, *cpu_map) {
7632 struct sched_domain *sd = &per_cpu(phys_domains, i);
7634 init_sched_groups_power(i, sd);
7638 for (i = 0; i < nr_node_ids; i++)
7639 init_numa_sched_groups_power(sched_group_nodes[i]);
7642 struct sched_group *sg;
7644 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7646 init_numa_sched_groups_power(sg);
7650 /* Attach the domains */
7651 for_each_cpu_mask_nr(i, *cpu_map) {
7652 struct sched_domain *sd;
7653 #ifdef CONFIG_SCHED_SMT
7654 sd = &per_cpu(cpu_domains, i);
7655 #elif defined(CONFIG_SCHED_MC)
7656 sd = &per_cpu(core_domains, i);
7658 sd = &per_cpu(phys_domains, i);
7660 cpu_attach_domain(sd, rd, i);
7663 sched_cpumask_free(allmasks);
7668 free_sched_groups(cpu_map, tmpmask);
7669 sched_cpumask_free(allmasks);
7675 static int build_sched_domains(const cpumask_t *cpu_map)
7677 return __build_sched_domains(cpu_map, NULL);
7680 static cpumask_t *doms_cur; /* current sched domains */
7681 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7682 static struct sched_domain_attr *dattr_cur;
7683 /* attribues of custom domains in 'doms_cur' */
7686 * Special case: If a kmalloc of a doms_cur partition (array of
7687 * cpumask_t) fails, then fallback to a single sched domain,
7688 * as determined by the single cpumask_t fallback_doms.
7690 static cpumask_t fallback_doms;
7693 * arch_update_cpu_topology lets virtualized architectures update the
7694 * cpu core maps. It is supposed to return 1 if the topology changed
7695 * or 0 if it stayed the same.
7697 int __attribute__((weak)) arch_update_cpu_topology(void)
7703 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7704 * For now this just excludes isolated cpus, but could be used to
7705 * exclude other special cases in the future.
7707 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7711 arch_update_cpu_topology();
7713 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7715 doms_cur = &fallback_doms;
7716 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7718 err = build_sched_domains(doms_cur);
7719 register_sched_domain_sysctl();
7724 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7727 free_sched_groups(cpu_map, tmpmask);
7731 * Detach sched domains from a group of cpus specified in cpu_map
7732 * These cpus will now be attached to the NULL domain
7734 static void detach_destroy_domains(const cpumask_t *cpu_map)
7739 for_each_cpu_mask_nr(i, *cpu_map)
7740 cpu_attach_domain(NULL, &def_root_domain, i);
7741 synchronize_sched();
7742 arch_destroy_sched_domains(cpu_map, &tmpmask);
7745 /* handle null as "default" */
7746 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7747 struct sched_domain_attr *new, int idx_new)
7749 struct sched_domain_attr tmp;
7756 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7757 new ? (new + idx_new) : &tmp,
7758 sizeof(struct sched_domain_attr));
7762 * Partition sched domains as specified by the 'ndoms_new'
7763 * cpumasks in the array doms_new[] of cpumasks. This compares
7764 * doms_new[] to the current sched domain partitioning, doms_cur[].
7765 * It destroys each deleted domain and builds each new domain.
7767 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7768 * The masks don't intersect (don't overlap.) We should setup one
7769 * sched domain for each mask. CPUs not in any of the cpumasks will
7770 * not be load balanced. If the same cpumask appears both in the
7771 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7774 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7775 * ownership of it and will kfree it when done with it. If the caller
7776 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7777 * ndoms_new == 1, and partition_sched_domains() will fallback to
7778 * the single partition 'fallback_doms', it also forces the domains
7781 * If doms_new == NULL it will be replaced with cpu_online_map.
7782 * ndoms_new == 0 is a special case for destroying existing domains,
7783 * and it will not create the default domain.
7785 * Call with hotplug lock held
7787 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7788 struct sched_domain_attr *dattr_new)
7793 mutex_lock(&sched_domains_mutex);
7795 /* always unregister in case we don't destroy any domains */
7796 unregister_sched_domain_sysctl();
7798 /* Let architecture update cpu core mappings. */
7799 new_topology = arch_update_cpu_topology();
7801 n = doms_new ? ndoms_new : 0;
7803 /* Destroy deleted domains */
7804 for (i = 0; i < ndoms_cur; i++) {
7805 for (j = 0; j < n && !new_topology; j++) {
7806 if (cpus_equal(doms_cur[i], doms_new[j])
7807 && dattrs_equal(dattr_cur, i, dattr_new, j))
7810 /* no match - a current sched domain not in new doms_new[] */
7811 detach_destroy_domains(doms_cur + i);
7816 if (doms_new == NULL) {
7818 doms_new = &fallback_doms;
7819 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7820 WARN_ON_ONCE(dattr_new);
7823 /* Build new domains */
7824 for (i = 0; i < ndoms_new; i++) {
7825 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7826 if (cpus_equal(doms_new[i], doms_cur[j])
7827 && dattrs_equal(dattr_new, i, dattr_cur, j))
7830 /* no match - add a new doms_new */
7831 __build_sched_domains(doms_new + i,
7832 dattr_new ? dattr_new + i : NULL);
7837 /* Remember the new sched domains */
7838 if (doms_cur != &fallback_doms)
7840 kfree(dattr_cur); /* kfree(NULL) is safe */
7841 doms_cur = doms_new;
7842 dattr_cur = dattr_new;
7843 ndoms_cur = ndoms_new;
7845 register_sched_domain_sysctl();
7847 mutex_unlock(&sched_domains_mutex);
7850 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7851 int arch_reinit_sched_domains(void)
7855 /* Destroy domains first to force the rebuild */
7856 partition_sched_domains(0, NULL, NULL);
7858 rebuild_sched_domains();
7864 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7868 if (buf[0] != '0' && buf[0] != '1')
7872 sched_smt_power_savings = (buf[0] == '1');
7874 sched_mc_power_savings = (buf[0] == '1');
7876 ret = arch_reinit_sched_domains();
7878 return ret ? ret : count;
7881 #ifdef CONFIG_SCHED_MC
7882 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7885 return sprintf(page, "%u\n", sched_mc_power_savings);
7887 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7888 const char *buf, size_t count)
7890 return sched_power_savings_store(buf, count, 0);
7892 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7893 sched_mc_power_savings_show,
7894 sched_mc_power_savings_store);
7897 #ifdef CONFIG_SCHED_SMT
7898 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7901 return sprintf(page, "%u\n", sched_smt_power_savings);
7903 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7904 const char *buf, size_t count)
7906 return sched_power_savings_store(buf, count, 1);
7908 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7909 sched_smt_power_savings_show,
7910 sched_smt_power_savings_store);
7913 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7917 #ifdef CONFIG_SCHED_SMT
7919 err = sysfs_create_file(&cls->kset.kobj,
7920 &attr_sched_smt_power_savings.attr);
7922 #ifdef CONFIG_SCHED_MC
7923 if (!err && mc_capable())
7924 err = sysfs_create_file(&cls->kset.kobj,
7925 &attr_sched_mc_power_savings.attr);
7929 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7931 #ifndef CONFIG_CPUSETS
7933 * Add online and remove offline CPUs from the scheduler domains.
7934 * When cpusets are enabled they take over this function.
7936 static int update_sched_domains(struct notifier_block *nfb,
7937 unsigned long action, void *hcpu)
7941 case CPU_ONLINE_FROZEN:
7943 case CPU_DEAD_FROZEN:
7944 partition_sched_domains(1, NULL, NULL);
7953 static int update_runtime(struct notifier_block *nfb,
7954 unsigned long action, void *hcpu)
7956 int cpu = (int)(long)hcpu;
7959 case CPU_DOWN_PREPARE:
7960 case CPU_DOWN_PREPARE_FROZEN:
7961 disable_runtime(cpu_rq(cpu));
7964 case CPU_DOWN_FAILED:
7965 case CPU_DOWN_FAILED_FROZEN:
7967 case CPU_ONLINE_FROZEN:
7968 enable_runtime(cpu_rq(cpu));
7976 void __init sched_init_smp(void)
7978 cpumask_t non_isolated_cpus;
7980 #if defined(CONFIG_NUMA)
7981 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7983 BUG_ON(sched_group_nodes_bycpu == NULL);
7986 mutex_lock(&sched_domains_mutex);
7987 arch_init_sched_domains(&cpu_online_map);
7988 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7989 if (cpus_empty(non_isolated_cpus))
7990 cpu_set(smp_processor_id(), non_isolated_cpus);
7991 mutex_unlock(&sched_domains_mutex);
7994 #ifndef CONFIG_CPUSETS
7995 /* XXX: Theoretical race here - CPU may be hotplugged now */
7996 hotcpu_notifier(update_sched_domains, 0);
7999 /* RT runtime code needs to handle some hotplug events */
8000 hotcpu_notifier(update_runtime, 0);
8004 /* Move init over to a non-isolated CPU */
8005 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
8007 sched_init_granularity();
8010 void __init sched_init_smp(void)
8012 sched_init_granularity();
8014 #endif /* CONFIG_SMP */
8016 int in_sched_functions(unsigned long addr)
8018 return in_lock_functions(addr) ||
8019 (addr >= (unsigned long)__sched_text_start
8020 && addr < (unsigned long)__sched_text_end);
8023 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8025 cfs_rq->tasks_timeline = RB_ROOT;
8026 INIT_LIST_HEAD(&cfs_rq->tasks);
8027 #ifdef CONFIG_FAIR_GROUP_SCHED
8030 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8033 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8035 struct rt_prio_array *array;
8038 array = &rt_rq->active;
8039 for (i = 0; i < MAX_RT_PRIO; i++) {
8040 INIT_LIST_HEAD(array->queue + i);
8041 __clear_bit(i, array->bitmap);
8043 /* delimiter for bitsearch: */
8044 __set_bit(MAX_RT_PRIO, array->bitmap);
8046 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8047 rt_rq->highest_prio = MAX_RT_PRIO;
8050 rt_rq->rt_nr_migratory = 0;
8051 rt_rq->overloaded = 0;
8055 rt_rq->rt_throttled = 0;
8056 rt_rq->rt_runtime = 0;
8057 spin_lock_init(&rt_rq->rt_runtime_lock);
8059 #ifdef CONFIG_RT_GROUP_SCHED
8060 rt_rq->rt_nr_boosted = 0;
8065 #ifdef CONFIG_FAIR_GROUP_SCHED
8066 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8067 struct sched_entity *se, int cpu, int add,
8068 struct sched_entity *parent)
8070 struct rq *rq = cpu_rq(cpu);
8071 tg->cfs_rq[cpu] = cfs_rq;
8072 init_cfs_rq(cfs_rq, rq);
8075 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8078 /* se could be NULL for init_task_group */
8083 se->cfs_rq = &rq->cfs;
8085 se->cfs_rq = parent->my_q;
8088 se->load.weight = tg->shares;
8089 se->load.inv_weight = 0;
8090 se->parent = parent;
8094 #ifdef CONFIG_RT_GROUP_SCHED
8095 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8096 struct sched_rt_entity *rt_se, int cpu, int add,
8097 struct sched_rt_entity *parent)
8099 struct rq *rq = cpu_rq(cpu);
8101 tg->rt_rq[cpu] = rt_rq;
8102 init_rt_rq(rt_rq, rq);
8104 rt_rq->rt_se = rt_se;
8105 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8107 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8109 tg->rt_se[cpu] = rt_se;
8114 rt_se->rt_rq = &rq->rt;
8116 rt_se->rt_rq = parent->my_q;
8118 rt_se->my_q = rt_rq;
8119 rt_se->parent = parent;
8120 INIT_LIST_HEAD(&rt_se->run_list);
8124 void __init sched_init(void)
8127 unsigned long alloc_size = 0, ptr;
8129 #ifdef CONFIG_FAIR_GROUP_SCHED
8130 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8132 #ifdef CONFIG_RT_GROUP_SCHED
8133 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8135 #ifdef CONFIG_USER_SCHED
8139 * As sched_init() is called before page_alloc is setup,
8140 * we use alloc_bootmem().
8143 ptr = (unsigned long)alloc_bootmem(alloc_size);
8145 #ifdef CONFIG_FAIR_GROUP_SCHED
8146 init_task_group.se = (struct sched_entity **)ptr;
8147 ptr += nr_cpu_ids * sizeof(void **);
8149 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8150 ptr += nr_cpu_ids * sizeof(void **);
8152 #ifdef CONFIG_USER_SCHED
8153 root_task_group.se = (struct sched_entity **)ptr;
8154 ptr += nr_cpu_ids * sizeof(void **);
8156 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8157 ptr += nr_cpu_ids * sizeof(void **);
8158 #endif /* CONFIG_USER_SCHED */
8159 #endif /* CONFIG_FAIR_GROUP_SCHED */
8160 #ifdef CONFIG_RT_GROUP_SCHED
8161 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8162 ptr += nr_cpu_ids * sizeof(void **);
8164 init_task_group.rt_rq = (struct rt_rq **)ptr;
8165 ptr += nr_cpu_ids * sizeof(void **);
8167 #ifdef CONFIG_USER_SCHED
8168 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8169 ptr += nr_cpu_ids * sizeof(void **);
8171 root_task_group.rt_rq = (struct rt_rq **)ptr;
8172 ptr += nr_cpu_ids * sizeof(void **);
8173 #endif /* CONFIG_USER_SCHED */
8174 #endif /* CONFIG_RT_GROUP_SCHED */
8178 init_defrootdomain();
8181 init_rt_bandwidth(&def_rt_bandwidth,
8182 global_rt_period(), global_rt_runtime());
8184 #ifdef CONFIG_RT_GROUP_SCHED
8185 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8186 global_rt_period(), global_rt_runtime());
8187 #ifdef CONFIG_USER_SCHED
8188 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8189 global_rt_period(), RUNTIME_INF);
8190 #endif /* CONFIG_USER_SCHED */
8191 #endif /* CONFIG_RT_GROUP_SCHED */
8193 #ifdef CONFIG_GROUP_SCHED
8194 list_add(&init_task_group.list, &task_groups);
8195 INIT_LIST_HEAD(&init_task_group.children);
8197 #ifdef CONFIG_USER_SCHED
8198 INIT_LIST_HEAD(&root_task_group.children);
8199 init_task_group.parent = &root_task_group;
8200 list_add(&init_task_group.siblings, &root_task_group.children);
8201 #endif /* CONFIG_USER_SCHED */
8202 #endif /* CONFIG_GROUP_SCHED */
8204 for_each_possible_cpu(i) {
8208 spin_lock_init(&rq->lock);
8210 init_cfs_rq(&rq->cfs, rq);
8211 init_rt_rq(&rq->rt, rq);
8212 #ifdef CONFIG_FAIR_GROUP_SCHED
8213 init_task_group.shares = init_task_group_load;
8214 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8215 #ifdef CONFIG_CGROUP_SCHED
8217 * How much cpu bandwidth does init_task_group get?
8219 * In case of task-groups formed thr' the cgroup filesystem, it
8220 * gets 100% of the cpu resources in the system. This overall
8221 * system cpu resource is divided among the tasks of
8222 * init_task_group and its child task-groups in a fair manner,
8223 * based on each entity's (task or task-group's) weight
8224 * (se->load.weight).
8226 * In other words, if init_task_group has 10 tasks of weight
8227 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8228 * then A0's share of the cpu resource is:
8230 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8232 * We achieve this by letting init_task_group's tasks sit
8233 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8235 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8236 #elif defined CONFIG_USER_SCHED
8237 root_task_group.shares = NICE_0_LOAD;
8238 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8240 * In case of task-groups formed thr' the user id of tasks,
8241 * init_task_group represents tasks belonging to root user.
8242 * Hence it forms a sibling of all subsequent groups formed.
8243 * In this case, init_task_group gets only a fraction of overall
8244 * system cpu resource, based on the weight assigned to root
8245 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8246 * by letting tasks of init_task_group sit in a separate cfs_rq
8247 * (init_cfs_rq) and having one entity represent this group of
8248 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8250 init_tg_cfs_entry(&init_task_group,
8251 &per_cpu(init_cfs_rq, i),
8252 &per_cpu(init_sched_entity, i), i, 1,
8253 root_task_group.se[i]);
8256 #endif /* CONFIG_FAIR_GROUP_SCHED */
8258 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8259 #ifdef CONFIG_RT_GROUP_SCHED
8260 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8261 #ifdef CONFIG_CGROUP_SCHED
8262 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8263 #elif defined CONFIG_USER_SCHED
8264 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8265 init_tg_rt_entry(&init_task_group,
8266 &per_cpu(init_rt_rq, i),
8267 &per_cpu(init_sched_rt_entity, i), i, 1,
8268 root_task_group.rt_se[i]);
8272 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8273 rq->cpu_load[j] = 0;
8277 rq->active_balance = 0;
8278 rq->next_balance = jiffies;
8282 rq->migration_thread = NULL;
8283 INIT_LIST_HEAD(&rq->migration_queue);
8284 rq_attach_root(rq, &def_root_domain);
8287 atomic_set(&rq->nr_iowait, 0);
8290 set_load_weight(&init_task);
8292 #ifdef CONFIG_PREEMPT_NOTIFIERS
8293 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8297 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8300 #ifdef CONFIG_RT_MUTEXES
8301 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8305 * The boot idle thread does lazy MMU switching as well:
8307 atomic_inc(&init_mm.mm_count);
8308 enter_lazy_tlb(&init_mm, current);
8311 * Make us the idle thread. Technically, schedule() should not be
8312 * called from this thread, however somewhere below it might be,
8313 * but because we are the idle thread, we just pick up running again
8314 * when this runqueue becomes "idle".
8316 init_idle(current, smp_processor_id());
8318 * During early bootup we pretend to be a normal task:
8320 current->sched_class = &fair_sched_class;
8322 scheduler_running = 1;
8325 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8326 void __might_sleep(char *file, int line)
8329 static unsigned long prev_jiffy; /* ratelimiting */
8331 if ((!in_atomic() && !irqs_disabled()) ||
8332 system_state != SYSTEM_RUNNING || oops_in_progress)
8334 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8336 prev_jiffy = jiffies;
8339 "BUG: sleeping function called from invalid context at %s:%d\n",
8342 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8343 in_atomic(), irqs_disabled(),
8344 current->pid, current->comm);
8346 debug_show_held_locks(current);
8347 if (irqs_disabled())
8348 print_irqtrace_events(current);
8352 EXPORT_SYMBOL(__might_sleep);
8355 #ifdef CONFIG_MAGIC_SYSRQ
8356 static void normalize_task(struct rq *rq, struct task_struct *p)
8360 update_rq_clock(rq);
8361 on_rq = p->se.on_rq;
8363 deactivate_task(rq, p, 0);
8364 __setscheduler(rq, p, SCHED_NORMAL, 0);
8366 activate_task(rq, p, 0);
8367 resched_task(rq->curr);
8371 void normalize_rt_tasks(void)
8373 struct task_struct *g, *p;
8374 unsigned long flags;
8377 read_lock_irqsave(&tasklist_lock, flags);
8378 do_each_thread(g, p) {
8380 * Only normalize user tasks:
8385 p->se.exec_start = 0;
8386 #ifdef CONFIG_SCHEDSTATS
8387 p->se.wait_start = 0;
8388 p->se.sleep_start = 0;
8389 p->se.block_start = 0;
8394 * Renice negative nice level userspace
8397 if (TASK_NICE(p) < 0 && p->mm)
8398 set_user_nice(p, 0);
8402 spin_lock(&p->pi_lock);
8403 rq = __task_rq_lock(p);
8405 normalize_task(rq, p);
8407 __task_rq_unlock(rq);
8408 spin_unlock(&p->pi_lock);
8409 } while_each_thread(g, p);
8411 read_unlock_irqrestore(&tasklist_lock, flags);
8414 #endif /* CONFIG_MAGIC_SYSRQ */
8418 * These functions are only useful for the IA64 MCA handling.
8420 * They can only be called when the whole system has been
8421 * stopped - every CPU needs to be quiescent, and no scheduling
8422 * activity can take place. Using them for anything else would
8423 * be a serious bug, and as a result, they aren't even visible
8424 * under any other configuration.
8428 * curr_task - return the current task for a given cpu.
8429 * @cpu: the processor in question.
8431 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8433 struct task_struct *curr_task(int cpu)
8435 return cpu_curr(cpu);
8439 * set_curr_task - set the current task for a given cpu.
8440 * @cpu: the processor in question.
8441 * @p: the task pointer to set.
8443 * Description: This function must only be used when non-maskable interrupts
8444 * are serviced on a separate stack. It allows the architecture to switch the
8445 * notion of the current task on a cpu in a non-blocking manner. This function
8446 * must be called with all CPU's synchronized, and interrupts disabled, the
8447 * and caller must save the original value of the current task (see
8448 * curr_task() above) and restore that value before reenabling interrupts and
8449 * re-starting the system.
8451 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8453 void set_curr_task(int cpu, struct task_struct *p)
8460 #ifdef CONFIG_FAIR_GROUP_SCHED
8461 static void free_fair_sched_group(struct task_group *tg)
8465 for_each_possible_cpu(i) {
8467 kfree(tg->cfs_rq[i]);
8477 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8479 struct cfs_rq *cfs_rq;
8480 struct sched_entity *se;
8484 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8487 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8491 tg->shares = NICE_0_LOAD;
8493 for_each_possible_cpu(i) {
8496 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8497 GFP_KERNEL, cpu_to_node(i));
8501 se = kzalloc_node(sizeof(struct sched_entity),
8502 GFP_KERNEL, cpu_to_node(i));
8506 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8515 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8517 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8518 &cpu_rq(cpu)->leaf_cfs_rq_list);
8521 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8523 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8525 #else /* !CONFG_FAIR_GROUP_SCHED */
8526 static inline void free_fair_sched_group(struct task_group *tg)
8531 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8536 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8540 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8543 #endif /* CONFIG_FAIR_GROUP_SCHED */
8545 #ifdef CONFIG_RT_GROUP_SCHED
8546 static void free_rt_sched_group(struct task_group *tg)
8550 destroy_rt_bandwidth(&tg->rt_bandwidth);
8552 for_each_possible_cpu(i) {
8554 kfree(tg->rt_rq[i]);
8556 kfree(tg->rt_se[i]);
8564 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8566 struct rt_rq *rt_rq;
8567 struct sched_rt_entity *rt_se;
8571 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8574 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8578 init_rt_bandwidth(&tg->rt_bandwidth,
8579 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8581 for_each_possible_cpu(i) {
8584 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8585 GFP_KERNEL, cpu_to_node(i));
8589 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8590 GFP_KERNEL, cpu_to_node(i));
8594 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8603 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8605 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8606 &cpu_rq(cpu)->leaf_rt_rq_list);
8609 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8611 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8613 #else /* !CONFIG_RT_GROUP_SCHED */
8614 static inline void free_rt_sched_group(struct task_group *tg)
8619 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8624 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8628 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8631 #endif /* CONFIG_RT_GROUP_SCHED */
8633 #ifdef CONFIG_GROUP_SCHED
8634 static void free_sched_group(struct task_group *tg)
8636 free_fair_sched_group(tg);
8637 free_rt_sched_group(tg);
8641 /* allocate runqueue etc for a new task group */
8642 struct task_group *sched_create_group(struct task_group *parent)
8644 struct task_group *tg;
8645 unsigned long flags;
8648 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8650 return ERR_PTR(-ENOMEM);
8652 if (!alloc_fair_sched_group(tg, parent))
8655 if (!alloc_rt_sched_group(tg, parent))
8658 spin_lock_irqsave(&task_group_lock, flags);
8659 for_each_possible_cpu(i) {
8660 register_fair_sched_group(tg, i);
8661 register_rt_sched_group(tg, i);
8663 list_add_rcu(&tg->list, &task_groups);
8665 WARN_ON(!parent); /* root should already exist */
8667 tg->parent = parent;
8668 INIT_LIST_HEAD(&tg->children);
8669 list_add_rcu(&tg->siblings, &parent->children);
8670 spin_unlock_irqrestore(&task_group_lock, flags);
8675 free_sched_group(tg);
8676 return ERR_PTR(-ENOMEM);
8679 /* rcu callback to free various structures associated with a task group */
8680 static void free_sched_group_rcu(struct rcu_head *rhp)
8682 /* now it should be safe to free those cfs_rqs */
8683 free_sched_group(container_of(rhp, struct task_group, rcu));
8686 /* Destroy runqueue etc associated with a task group */
8687 void sched_destroy_group(struct task_group *tg)
8689 unsigned long flags;
8692 spin_lock_irqsave(&task_group_lock, flags);
8693 for_each_possible_cpu(i) {
8694 unregister_fair_sched_group(tg, i);
8695 unregister_rt_sched_group(tg, i);
8697 list_del_rcu(&tg->list);
8698 list_del_rcu(&tg->siblings);
8699 spin_unlock_irqrestore(&task_group_lock, flags);
8701 /* wait for possible concurrent references to cfs_rqs complete */
8702 call_rcu(&tg->rcu, free_sched_group_rcu);
8705 /* change task's runqueue when it moves between groups.
8706 * The caller of this function should have put the task in its new group
8707 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8708 * reflect its new group.
8710 void sched_move_task(struct task_struct *tsk)
8713 unsigned long flags;
8716 rq = task_rq_lock(tsk, &flags);
8718 update_rq_clock(rq);
8720 running = task_current(rq, tsk);
8721 on_rq = tsk->se.on_rq;
8724 dequeue_task(rq, tsk, 0);
8725 if (unlikely(running))
8726 tsk->sched_class->put_prev_task(rq, tsk);
8728 set_task_rq(tsk, task_cpu(tsk));
8730 #ifdef CONFIG_FAIR_GROUP_SCHED
8731 if (tsk->sched_class->moved_group)
8732 tsk->sched_class->moved_group(tsk);
8735 if (unlikely(running))
8736 tsk->sched_class->set_curr_task(rq);
8738 enqueue_task(rq, tsk, 0);
8740 task_rq_unlock(rq, &flags);
8742 #endif /* CONFIG_GROUP_SCHED */
8744 #ifdef CONFIG_FAIR_GROUP_SCHED
8745 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8747 struct cfs_rq *cfs_rq = se->cfs_rq;
8752 dequeue_entity(cfs_rq, se, 0);
8754 se->load.weight = shares;
8755 se->load.inv_weight = 0;
8758 enqueue_entity(cfs_rq, se, 0);
8761 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8763 struct cfs_rq *cfs_rq = se->cfs_rq;
8764 struct rq *rq = cfs_rq->rq;
8765 unsigned long flags;
8767 spin_lock_irqsave(&rq->lock, flags);
8768 __set_se_shares(se, shares);
8769 spin_unlock_irqrestore(&rq->lock, flags);
8772 static DEFINE_MUTEX(shares_mutex);
8774 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8777 unsigned long flags;
8780 * We can't change the weight of the root cgroup.
8785 if (shares < MIN_SHARES)
8786 shares = MIN_SHARES;
8787 else if (shares > MAX_SHARES)
8788 shares = MAX_SHARES;
8790 mutex_lock(&shares_mutex);
8791 if (tg->shares == shares)
8794 spin_lock_irqsave(&task_group_lock, flags);
8795 for_each_possible_cpu(i)
8796 unregister_fair_sched_group(tg, i);
8797 list_del_rcu(&tg->siblings);
8798 spin_unlock_irqrestore(&task_group_lock, flags);
8800 /* wait for any ongoing reference to this group to finish */
8801 synchronize_sched();
8804 * Now we are free to modify the group's share on each cpu
8805 * w/o tripping rebalance_share or load_balance_fair.
8807 tg->shares = shares;
8808 for_each_possible_cpu(i) {
8812 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8813 set_se_shares(tg->se[i], shares);
8817 * Enable load balance activity on this group, by inserting it back on
8818 * each cpu's rq->leaf_cfs_rq_list.
8820 spin_lock_irqsave(&task_group_lock, flags);
8821 for_each_possible_cpu(i)
8822 register_fair_sched_group(tg, i);
8823 list_add_rcu(&tg->siblings, &tg->parent->children);
8824 spin_unlock_irqrestore(&task_group_lock, flags);
8826 mutex_unlock(&shares_mutex);
8830 unsigned long sched_group_shares(struct task_group *tg)
8836 #ifdef CONFIG_RT_GROUP_SCHED
8838 * Ensure that the real time constraints are schedulable.
8840 static DEFINE_MUTEX(rt_constraints_mutex);
8842 static unsigned long to_ratio(u64 period, u64 runtime)
8844 if (runtime == RUNTIME_INF)
8847 return div64_u64(runtime << 20, period);
8850 /* Must be called with tasklist_lock held */
8851 static inline int tg_has_rt_tasks(struct task_group *tg)
8853 struct task_struct *g, *p;
8855 do_each_thread(g, p) {
8856 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8858 } while_each_thread(g, p);
8863 struct rt_schedulable_data {
8864 struct task_group *tg;
8869 static int tg_schedulable(struct task_group *tg, void *data)
8871 struct rt_schedulable_data *d = data;
8872 struct task_group *child;
8873 unsigned long total, sum = 0;
8874 u64 period, runtime;
8876 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8877 runtime = tg->rt_bandwidth.rt_runtime;
8880 period = d->rt_period;
8881 runtime = d->rt_runtime;
8885 * Cannot have more runtime than the period.
8887 if (runtime > period && runtime != RUNTIME_INF)
8891 * Ensure we don't starve existing RT tasks.
8893 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8896 total = to_ratio(period, runtime);
8899 * Nobody can have more than the global setting allows.
8901 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8905 * The sum of our children's runtime should not exceed our own.
8907 list_for_each_entry_rcu(child, &tg->children, siblings) {
8908 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8909 runtime = child->rt_bandwidth.rt_runtime;
8911 if (child == d->tg) {
8912 period = d->rt_period;
8913 runtime = d->rt_runtime;
8916 sum += to_ratio(period, runtime);
8925 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8927 struct rt_schedulable_data data = {
8929 .rt_period = period,
8930 .rt_runtime = runtime,
8933 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8936 static int tg_set_bandwidth(struct task_group *tg,
8937 u64 rt_period, u64 rt_runtime)
8941 mutex_lock(&rt_constraints_mutex);
8942 read_lock(&tasklist_lock);
8943 err = __rt_schedulable(tg, rt_period, rt_runtime);
8947 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8948 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8949 tg->rt_bandwidth.rt_runtime = rt_runtime;
8951 for_each_possible_cpu(i) {
8952 struct rt_rq *rt_rq = tg->rt_rq[i];
8954 spin_lock(&rt_rq->rt_runtime_lock);
8955 rt_rq->rt_runtime = rt_runtime;
8956 spin_unlock(&rt_rq->rt_runtime_lock);
8958 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8960 read_unlock(&tasklist_lock);
8961 mutex_unlock(&rt_constraints_mutex);
8966 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8968 u64 rt_runtime, rt_period;
8970 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8971 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8972 if (rt_runtime_us < 0)
8973 rt_runtime = RUNTIME_INF;
8975 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8978 long sched_group_rt_runtime(struct task_group *tg)
8982 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8985 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8986 do_div(rt_runtime_us, NSEC_PER_USEC);
8987 return rt_runtime_us;
8990 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8992 u64 rt_runtime, rt_period;
8994 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8995 rt_runtime = tg->rt_bandwidth.rt_runtime;
9000 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9003 long sched_group_rt_period(struct task_group *tg)
9007 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9008 do_div(rt_period_us, NSEC_PER_USEC);
9009 return rt_period_us;
9012 static int sched_rt_global_constraints(void)
9014 u64 runtime, period;
9017 if (sysctl_sched_rt_period <= 0)
9020 runtime = global_rt_runtime();
9021 period = global_rt_period();
9024 * Sanity check on the sysctl variables.
9026 if (runtime > period && runtime != RUNTIME_INF)
9029 mutex_lock(&rt_constraints_mutex);
9030 read_lock(&tasklist_lock);
9031 ret = __rt_schedulable(NULL, 0, 0);
9032 read_unlock(&tasklist_lock);
9033 mutex_unlock(&rt_constraints_mutex);
9037 #else /* !CONFIG_RT_GROUP_SCHED */
9038 static int sched_rt_global_constraints(void)
9040 unsigned long flags;
9043 if (sysctl_sched_rt_period <= 0)
9046 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9047 for_each_possible_cpu(i) {
9048 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9050 spin_lock(&rt_rq->rt_runtime_lock);
9051 rt_rq->rt_runtime = global_rt_runtime();
9052 spin_unlock(&rt_rq->rt_runtime_lock);
9054 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9058 #endif /* CONFIG_RT_GROUP_SCHED */
9060 int sched_rt_handler(struct ctl_table *table, int write,
9061 struct file *filp, void __user *buffer, size_t *lenp,
9065 int old_period, old_runtime;
9066 static DEFINE_MUTEX(mutex);
9069 old_period = sysctl_sched_rt_period;
9070 old_runtime = sysctl_sched_rt_runtime;
9072 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9074 if (!ret && write) {
9075 ret = sched_rt_global_constraints();
9077 sysctl_sched_rt_period = old_period;
9078 sysctl_sched_rt_runtime = old_runtime;
9080 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9081 def_rt_bandwidth.rt_period =
9082 ns_to_ktime(global_rt_period());
9085 mutex_unlock(&mutex);
9090 #ifdef CONFIG_CGROUP_SCHED
9092 /* return corresponding task_group object of a cgroup */
9093 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9095 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9096 struct task_group, css);
9099 static struct cgroup_subsys_state *
9100 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9102 struct task_group *tg, *parent;
9104 if (!cgrp->parent) {
9105 /* This is early initialization for the top cgroup */
9106 return &init_task_group.css;
9109 parent = cgroup_tg(cgrp->parent);
9110 tg = sched_create_group(parent);
9112 return ERR_PTR(-ENOMEM);
9118 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9120 struct task_group *tg = cgroup_tg(cgrp);
9122 sched_destroy_group(tg);
9126 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9127 struct task_struct *tsk)
9129 #ifdef CONFIG_RT_GROUP_SCHED
9130 /* Don't accept realtime tasks when there is no way for them to run */
9131 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9134 /* We don't support RT-tasks being in separate groups */
9135 if (tsk->sched_class != &fair_sched_class)
9143 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9144 struct cgroup *old_cont, struct task_struct *tsk)
9146 sched_move_task(tsk);
9149 #ifdef CONFIG_FAIR_GROUP_SCHED
9150 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9153 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9156 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9158 struct task_group *tg = cgroup_tg(cgrp);
9160 return (u64) tg->shares;
9162 #endif /* CONFIG_FAIR_GROUP_SCHED */
9164 #ifdef CONFIG_RT_GROUP_SCHED
9165 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9168 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9171 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9173 return sched_group_rt_runtime(cgroup_tg(cgrp));
9176 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9179 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9182 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9184 return sched_group_rt_period(cgroup_tg(cgrp));
9186 #endif /* CONFIG_RT_GROUP_SCHED */
9188 static struct cftype cpu_files[] = {
9189 #ifdef CONFIG_FAIR_GROUP_SCHED
9192 .read_u64 = cpu_shares_read_u64,
9193 .write_u64 = cpu_shares_write_u64,
9196 #ifdef CONFIG_RT_GROUP_SCHED
9198 .name = "rt_runtime_us",
9199 .read_s64 = cpu_rt_runtime_read,
9200 .write_s64 = cpu_rt_runtime_write,
9203 .name = "rt_period_us",
9204 .read_u64 = cpu_rt_period_read_uint,
9205 .write_u64 = cpu_rt_period_write_uint,
9210 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9212 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9215 struct cgroup_subsys cpu_cgroup_subsys = {
9217 .create = cpu_cgroup_create,
9218 .destroy = cpu_cgroup_destroy,
9219 .can_attach = cpu_cgroup_can_attach,
9220 .attach = cpu_cgroup_attach,
9221 .populate = cpu_cgroup_populate,
9222 .subsys_id = cpu_cgroup_subsys_id,
9226 #endif /* CONFIG_CGROUP_SCHED */
9228 #ifdef CONFIG_CGROUP_CPUACCT
9231 * CPU accounting code for task groups.
9233 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9234 * (balbir@in.ibm.com).
9237 /* track cpu usage of a group of tasks and its child groups */
9239 struct cgroup_subsys_state css;
9240 /* cpuusage holds pointer to a u64-type object on every cpu */
9242 struct cpuacct *parent;
9245 struct cgroup_subsys cpuacct_subsys;
9247 /* return cpu accounting group corresponding to this container */
9248 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9250 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9251 struct cpuacct, css);
9254 /* return cpu accounting group to which this task belongs */
9255 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9257 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9258 struct cpuacct, css);
9261 /* create a new cpu accounting group */
9262 static struct cgroup_subsys_state *cpuacct_create(
9263 struct cgroup_subsys *ss, struct cgroup *cgrp)
9265 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9268 return ERR_PTR(-ENOMEM);
9270 ca->cpuusage = alloc_percpu(u64);
9271 if (!ca->cpuusage) {
9273 return ERR_PTR(-ENOMEM);
9277 ca->parent = cgroup_ca(cgrp->parent);
9282 /* destroy an existing cpu accounting group */
9284 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9286 struct cpuacct *ca = cgroup_ca(cgrp);
9288 free_percpu(ca->cpuusage);
9292 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9294 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9297 #ifndef CONFIG_64BIT
9299 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9301 spin_lock_irq(&cpu_rq(cpu)->lock);
9303 spin_unlock_irq(&cpu_rq(cpu)->lock);
9311 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9313 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9315 #ifndef CONFIG_64BIT
9317 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9319 spin_lock_irq(&cpu_rq(cpu)->lock);
9321 spin_unlock_irq(&cpu_rq(cpu)->lock);
9327 /* return total cpu usage (in nanoseconds) of a group */
9328 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9330 struct cpuacct *ca = cgroup_ca(cgrp);
9331 u64 totalcpuusage = 0;
9334 for_each_present_cpu(i)
9335 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9337 return totalcpuusage;
9340 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9343 struct cpuacct *ca = cgroup_ca(cgrp);
9352 for_each_present_cpu(i)
9353 cpuacct_cpuusage_write(ca, i, 0);
9359 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9362 struct cpuacct *ca = cgroup_ca(cgroup);
9366 for_each_present_cpu(i) {
9367 percpu = cpuacct_cpuusage_read(ca, i);
9368 seq_printf(m, "%llu ", (unsigned long long) percpu);
9370 seq_printf(m, "\n");
9374 static struct cftype files[] = {
9377 .read_u64 = cpuusage_read,
9378 .write_u64 = cpuusage_write,
9381 .name = "usage_percpu",
9382 .read_seq_string = cpuacct_percpu_seq_read,
9387 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9389 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9393 * charge this task's execution time to its accounting group.
9395 * called with rq->lock held.
9397 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9402 if (!cpuacct_subsys.active)
9405 cpu = task_cpu(tsk);
9408 for (; ca; ca = ca->parent) {
9409 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9410 *cpuusage += cputime;
9414 struct cgroup_subsys cpuacct_subsys = {
9416 .create = cpuacct_create,
9417 .destroy = cpuacct_destroy,
9418 .populate = cpuacct_populate,
9419 .subsys_id = cpuacct_subsys_id,
9421 #endif /* CONFIG_CGROUP_CPUACCT */