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/perf_counter.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/reciprocal_div.h>
68 #include <linux/unistd.h>
69 #include <linux/pagemap.h>
70 #include <linux/hrtimer.h>
71 #include <linux/tick.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/sched.h>
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * single value that denotes runtime == period, ie unlimited time.
121 #define RUNTIME_INF ((u64)~0ULL)
125 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
128 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
129 * Since cpu_power is a 'constant', we can use a reciprocal divide.
131 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
133 return reciprocal_divide(load, sg->reciprocal_cpu_power);
137 * Each time a sched group cpu_power is changed,
138 * we must compute its reciprocal value
140 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
142 sg->__cpu_power += val;
143 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
147 static inline int rt_policy(int policy)
149 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
154 static inline int task_has_rt_policy(struct task_struct *p)
156 return rt_policy(p->policy);
160 * This is the priority-queue data structure of the RT scheduling class:
162 struct rt_prio_array {
163 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
164 struct list_head queue[MAX_RT_PRIO];
167 struct rt_bandwidth {
168 /* nests inside the rq lock: */
169 spinlock_t rt_runtime_lock;
172 struct hrtimer rt_period_timer;
175 static struct rt_bandwidth def_rt_bandwidth;
177 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
179 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
181 struct rt_bandwidth *rt_b =
182 container_of(timer, struct rt_bandwidth, rt_period_timer);
188 now = hrtimer_cb_get_time(timer);
189 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
194 idle = do_sched_rt_period_timer(rt_b, overrun);
197 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
201 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
203 rt_b->rt_period = ns_to_ktime(period);
204 rt_b->rt_runtime = runtime;
206 spin_lock_init(&rt_b->rt_runtime_lock);
208 hrtimer_init(&rt_b->rt_period_timer,
209 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
210 rt_b->rt_period_timer.function = sched_rt_period_timer;
213 static inline int rt_bandwidth_enabled(void)
215 return sysctl_sched_rt_runtime >= 0;
218 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
222 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
225 if (hrtimer_active(&rt_b->rt_period_timer))
228 spin_lock(&rt_b->rt_runtime_lock);
233 if (hrtimer_active(&rt_b->rt_period_timer))
236 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
237 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
239 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
240 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
241 delta = ktime_to_ns(ktime_sub(hard, soft));
242 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
243 HRTIMER_MODE_ABS_PINNED, 0);
245 spin_unlock(&rt_b->rt_runtime_lock);
248 #ifdef CONFIG_RT_GROUP_SCHED
249 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
251 hrtimer_cancel(&rt_b->rt_period_timer);
256 * sched_domains_mutex serializes calls to arch_init_sched_domains,
257 * detach_destroy_domains and partition_sched_domains.
259 static DEFINE_MUTEX(sched_domains_mutex);
261 #ifdef CONFIG_GROUP_SCHED
263 #include <linux/cgroup.h>
267 static LIST_HEAD(task_groups);
269 /* task group related information */
271 #ifdef CONFIG_CGROUP_SCHED
272 struct cgroup_subsys_state css;
275 #ifdef CONFIG_USER_SCHED
279 #ifdef CONFIG_FAIR_GROUP_SCHED
280 /* schedulable entities of this group on each cpu */
281 struct sched_entity **se;
282 /* runqueue "owned" by this group on each cpu */
283 struct cfs_rq **cfs_rq;
284 unsigned long shares;
287 #ifdef CONFIG_RT_GROUP_SCHED
288 struct sched_rt_entity **rt_se;
289 struct rt_rq **rt_rq;
291 struct rt_bandwidth rt_bandwidth;
295 struct list_head list;
297 struct task_group *parent;
298 struct list_head siblings;
299 struct list_head children;
302 #ifdef CONFIG_USER_SCHED
304 /* Helper function to pass uid information to create_sched_user() */
305 void set_tg_uid(struct user_struct *user)
307 user->tg->uid = user->uid;
312 * Every UID task group (including init_task_group aka UID-0) will
313 * be a child to this group.
315 struct task_group root_task_group;
317 #ifdef CONFIG_FAIR_GROUP_SCHED
318 /* Default task group's sched entity on each cpu */
319 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
320 /* Default task group's cfs_rq on each cpu */
321 static DEFINE_PER_CPU(struct cfs_rq, init_tg_cfs_rq) ____cacheline_aligned_in_smp;
322 #endif /* CONFIG_FAIR_GROUP_SCHED */
324 #ifdef CONFIG_RT_GROUP_SCHED
325 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
326 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
327 #endif /* CONFIG_RT_GROUP_SCHED */
328 #else /* !CONFIG_USER_SCHED */
329 #define root_task_group init_task_group
330 #endif /* CONFIG_USER_SCHED */
332 /* task_group_lock serializes add/remove of task groups and also changes to
333 * a task group's cpu shares.
335 static DEFINE_SPINLOCK(task_group_lock);
338 static int root_task_group_empty(void)
340 return list_empty(&root_task_group.children);
344 #ifdef CONFIG_FAIR_GROUP_SCHED
345 #ifdef CONFIG_USER_SCHED
346 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
347 #else /* !CONFIG_USER_SCHED */
348 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
349 #endif /* CONFIG_USER_SCHED */
352 * A weight of 0 or 1 can cause arithmetics problems.
353 * A weight of a cfs_rq is the sum of weights of which entities
354 * are queued on this cfs_rq, so a weight of a entity should not be
355 * too large, so as the shares value of a task group.
356 * (The default weight is 1024 - so there's no practical
357 * limitation from this.)
360 #define MAX_SHARES (1UL << 18)
362 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
365 /* Default task group.
366 * Every task in system belong to this group at bootup.
368 struct task_group init_task_group;
370 /* return group to which a task belongs */
371 static inline struct task_group *task_group(struct task_struct *p)
373 struct task_group *tg;
375 #ifdef CONFIG_USER_SCHED
377 tg = __task_cred(p)->user->tg;
379 #elif defined(CONFIG_CGROUP_SCHED)
380 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
381 struct task_group, css);
383 tg = &init_task_group;
388 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
389 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
391 #ifdef CONFIG_FAIR_GROUP_SCHED
392 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
393 p->se.parent = task_group(p)->se[cpu];
396 #ifdef CONFIG_RT_GROUP_SCHED
397 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
398 p->rt.parent = task_group(p)->rt_se[cpu];
405 static int root_task_group_empty(void)
411 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
412 static inline struct task_group *task_group(struct task_struct *p)
417 #endif /* CONFIG_GROUP_SCHED */
419 /* CFS-related fields in a runqueue */
421 struct load_weight load;
422 unsigned long nr_running;
427 struct rb_root tasks_timeline;
428 struct rb_node *rb_leftmost;
430 struct list_head tasks;
431 struct list_head *balance_iterator;
434 * 'curr' points to currently running entity on this cfs_rq.
435 * It is set to NULL otherwise (i.e when none are currently running).
437 struct sched_entity *curr, *next, *last;
439 unsigned int nr_spread_over;
441 #ifdef CONFIG_FAIR_GROUP_SCHED
442 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
445 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
446 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
447 * (like users, containers etc.)
449 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
450 * list is used during load balance.
452 struct list_head leaf_cfs_rq_list;
453 struct task_group *tg; /* group that "owns" this runqueue */
457 * the part of load.weight contributed by tasks
459 unsigned long task_weight;
462 * h_load = weight * f(tg)
464 * Where f(tg) is the recursive weight fraction assigned to
467 unsigned long h_load;
470 * this cpu's part of tg->shares
472 unsigned long shares;
475 * load.weight at the time we set shares
477 unsigned long rq_weight;
482 /* Real-Time classes' related field in a runqueue: */
484 struct rt_prio_array active;
485 unsigned long rt_nr_running;
486 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
488 int curr; /* highest queued rt task prio */
490 int next; /* next highest */
495 unsigned long rt_nr_migratory;
496 unsigned long rt_nr_total;
498 struct plist_head pushable_tasks;
503 /* Nests inside the rq lock: */
504 spinlock_t rt_runtime_lock;
506 #ifdef CONFIG_RT_GROUP_SCHED
507 unsigned long rt_nr_boosted;
510 struct list_head leaf_rt_rq_list;
511 struct task_group *tg;
512 struct sched_rt_entity *rt_se;
519 * We add the notion of a root-domain which will be used to define per-domain
520 * variables. Each exclusive cpuset essentially defines an island domain by
521 * fully partitioning the member cpus from any other cpuset. Whenever a new
522 * exclusive cpuset is created, we also create and attach a new root-domain
529 cpumask_var_t online;
532 * The "RT overload" flag: it gets set if a CPU has more than
533 * one runnable RT task.
535 cpumask_var_t rto_mask;
538 struct cpupri cpupri;
540 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
542 * Preferred wake up cpu nominated by sched_mc balance that will be
543 * used when most cpus are idle in the system indicating overall very
544 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
546 unsigned int sched_mc_preferred_wakeup_cpu;
551 * By default the system creates a single root-domain with all cpus as
552 * members (mimicking the global state we have today).
554 static struct root_domain def_root_domain;
559 * This is the main, per-CPU runqueue data structure.
561 * Locking rule: those places that want to lock multiple runqueues
562 * (such as the load balancing or the thread migration code), lock
563 * acquire operations must be ordered by ascending &runqueue.
570 * nr_running and cpu_load should be in the same cacheline because
571 * remote CPUs use both these fields when doing load calculation.
573 unsigned long nr_running;
574 #define CPU_LOAD_IDX_MAX 5
575 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
577 unsigned long last_tick_seen;
578 unsigned char in_nohz_recently;
580 /* capture load from *all* tasks on this cpu: */
581 struct load_weight load;
582 unsigned long nr_load_updates;
584 u64 nr_migrations_in;
589 #ifdef CONFIG_FAIR_GROUP_SCHED
590 /* list of leaf cfs_rq on this cpu: */
591 struct list_head leaf_cfs_rq_list;
593 #ifdef CONFIG_RT_GROUP_SCHED
594 struct list_head leaf_rt_rq_list;
598 * This is part of a global counter where only the total sum
599 * over all CPUs matters. A task can increase this counter on
600 * one CPU and if it got migrated afterwards it may decrease
601 * it on another CPU. Always updated under the runqueue lock:
603 unsigned long nr_uninterruptible;
605 struct task_struct *curr, *idle;
606 unsigned long next_balance;
607 struct mm_struct *prev_mm;
614 struct root_domain *rd;
615 struct sched_domain *sd;
617 unsigned char idle_at_tick;
618 /* For active balancing */
622 /* cpu of this runqueue: */
626 unsigned long avg_load_per_task;
628 struct task_struct *migration_thread;
629 struct list_head migration_queue;
635 /* calc_load related fields */
636 unsigned long calc_load_update;
637 long calc_load_active;
639 #ifdef CONFIG_SCHED_HRTICK
641 int hrtick_csd_pending;
642 struct call_single_data hrtick_csd;
644 struct hrtimer hrtick_timer;
647 #ifdef CONFIG_SCHEDSTATS
649 struct sched_info rq_sched_info;
650 unsigned long long rq_cpu_time;
651 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
653 /* sys_sched_yield() stats */
654 unsigned int yld_count;
656 /* schedule() stats */
657 unsigned int sched_switch;
658 unsigned int sched_count;
659 unsigned int sched_goidle;
661 /* try_to_wake_up() stats */
662 unsigned int ttwu_count;
663 unsigned int ttwu_local;
666 unsigned int bkl_count;
670 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
672 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
674 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
677 static inline int cpu_of(struct rq *rq)
687 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
688 * See detach_destroy_domains: synchronize_sched for details.
690 * The domain tree of any CPU may only be accessed from within
691 * preempt-disabled sections.
693 #define for_each_domain(cpu, __sd) \
694 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
696 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
697 #define this_rq() (&__get_cpu_var(runqueues))
698 #define task_rq(p) cpu_rq(task_cpu(p))
699 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
700 #define raw_rq() (&__raw_get_cpu_var(runqueues))
702 inline void update_rq_clock(struct rq *rq)
704 rq->clock = sched_clock_cpu(cpu_of(rq));
708 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
710 #ifdef CONFIG_SCHED_DEBUG
711 # define const_debug __read_mostly
713 # define const_debug static const
719 * Returns true if the current cpu runqueue is locked.
720 * This interface allows printk to be called with the runqueue lock
721 * held and know whether or not it is OK to wake up the klogd.
723 int runqueue_is_locked(void)
726 struct rq *rq = cpu_rq(cpu);
729 ret = spin_is_locked(&rq->lock);
735 * Debugging: various feature bits
738 #define SCHED_FEAT(name, enabled) \
739 __SCHED_FEAT_##name ,
742 #include "sched_features.h"
747 #define SCHED_FEAT(name, enabled) \
748 (1UL << __SCHED_FEAT_##name) * enabled |
750 const_debug unsigned int sysctl_sched_features =
751 #include "sched_features.h"
756 #ifdef CONFIG_SCHED_DEBUG
757 #define SCHED_FEAT(name, enabled) \
760 static __read_mostly char *sched_feat_names[] = {
761 #include "sched_features.h"
767 static int sched_feat_show(struct seq_file *m, void *v)
771 for (i = 0; sched_feat_names[i]; i++) {
772 if (!(sysctl_sched_features & (1UL << i)))
774 seq_printf(m, "%s ", sched_feat_names[i]);
782 sched_feat_write(struct file *filp, const char __user *ubuf,
783 size_t cnt, loff_t *ppos)
793 if (copy_from_user(&buf, ubuf, cnt))
798 if (strncmp(buf, "NO_", 3) == 0) {
803 for (i = 0; sched_feat_names[i]; i++) {
804 int len = strlen(sched_feat_names[i]);
806 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
808 sysctl_sched_features &= ~(1UL << i);
810 sysctl_sched_features |= (1UL << i);
815 if (!sched_feat_names[i])
823 static int sched_feat_open(struct inode *inode, struct file *filp)
825 return single_open(filp, sched_feat_show, NULL);
828 static struct file_operations sched_feat_fops = {
829 .open = sched_feat_open,
830 .write = sched_feat_write,
833 .release = single_release,
836 static __init int sched_init_debug(void)
838 debugfs_create_file("sched_features", 0644, NULL, NULL,
843 late_initcall(sched_init_debug);
847 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
850 * Number of tasks to iterate in a single balance run.
851 * Limited because this is done with IRQs disabled.
853 const_debug unsigned int sysctl_sched_nr_migrate = 32;
856 * ratelimit for updating the group shares.
859 unsigned int sysctl_sched_shares_ratelimit = 250000;
862 * Inject some fuzzyness into changing the per-cpu group shares
863 * this avoids remote rq-locks at the expense of fairness.
866 unsigned int sysctl_sched_shares_thresh = 4;
869 * period over which we average the RT time consumption, measured
874 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
877 * period over which we measure -rt task cpu usage in us.
880 unsigned int sysctl_sched_rt_period = 1000000;
882 static __read_mostly int scheduler_running;
885 * part of the period that we allow rt tasks to run in us.
888 int sysctl_sched_rt_runtime = 950000;
890 static inline u64 global_rt_period(void)
892 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
895 static inline u64 global_rt_runtime(void)
897 if (sysctl_sched_rt_runtime < 0)
900 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
903 #ifndef prepare_arch_switch
904 # define prepare_arch_switch(next) do { } while (0)
906 #ifndef finish_arch_switch
907 # define finish_arch_switch(prev) do { } while (0)
910 static inline int task_current(struct rq *rq, struct task_struct *p)
912 return rq->curr == p;
915 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
916 static inline int task_running(struct rq *rq, struct task_struct *p)
918 return task_current(rq, p);
921 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
925 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
927 #ifdef CONFIG_DEBUG_SPINLOCK
928 /* this is a valid case when another task releases the spinlock */
929 rq->lock.owner = current;
932 * If we are tracking spinlock dependencies then we have to
933 * fix up the runqueue lock - which gets 'carried over' from
936 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
938 spin_unlock_irq(&rq->lock);
941 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
942 static inline int task_running(struct rq *rq, struct task_struct *p)
947 return task_current(rq, p);
951 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
955 * We can optimise this out completely for !SMP, because the
956 * SMP rebalancing from interrupt is the only thing that cares
961 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
962 spin_unlock_irq(&rq->lock);
964 spin_unlock(&rq->lock);
968 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
972 * After ->oncpu is cleared, the task can be moved to a different CPU.
973 * We must ensure this doesn't happen until the switch is completely
979 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
983 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
986 * __task_rq_lock - lock the runqueue a given task resides on.
987 * Must be called interrupts disabled.
989 static inline struct rq *__task_rq_lock(struct task_struct *p)
993 struct rq *rq = task_rq(p);
994 spin_lock(&rq->lock);
995 if (likely(rq == task_rq(p)))
997 spin_unlock(&rq->lock);
1002 * task_rq_lock - lock the runqueue a given task resides on and disable
1003 * interrupts. Note the ordering: we can safely lookup the task_rq without
1004 * explicitly disabling preemption.
1006 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1007 __acquires(rq->lock)
1012 local_irq_save(*flags);
1014 spin_lock(&rq->lock);
1015 if (likely(rq == task_rq(p)))
1017 spin_unlock_irqrestore(&rq->lock, *flags);
1021 void task_rq_unlock_wait(struct task_struct *p)
1023 struct rq *rq = task_rq(p);
1025 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1026 spin_unlock_wait(&rq->lock);
1029 static void __task_rq_unlock(struct rq *rq)
1030 __releases(rq->lock)
1032 spin_unlock(&rq->lock);
1035 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1036 __releases(rq->lock)
1038 spin_unlock_irqrestore(&rq->lock, *flags);
1042 * this_rq_lock - lock this runqueue and disable interrupts.
1044 static struct rq *this_rq_lock(void)
1045 __acquires(rq->lock)
1049 local_irq_disable();
1051 spin_lock(&rq->lock);
1056 #ifdef CONFIG_SCHED_HRTICK
1058 * Use HR-timers to deliver accurate preemption points.
1060 * Its all a bit involved since we cannot program an hrt while holding the
1061 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1064 * When we get rescheduled we reprogram the hrtick_timer outside of the
1070 * - enabled by features
1071 * - hrtimer is actually high res
1073 static inline int hrtick_enabled(struct rq *rq)
1075 if (!sched_feat(HRTICK))
1077 if (!cpu_active(cpu_of(rq)))
1079 return hrtimer_is_hres_active(&rq->hrtick_timer);
1082 static void hrtick_clear(struct rq *rq)
1084 if (hrtimer_active(&rq->hrtick_timer))
1085 hrtimer_cancel(&rq->hrtick_timer);
1089 * High-resolution timer tick.
1090 * Runs from hardirq context with interrupts disabled.
1092 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1094 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1096 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1098 spin_lock(&rq->lock);
1099 update_rq_clock(rq);
1100 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1101 spin_unlock(&rq->lock);
1103 return HRTIMER_NORESTART;
1108 * called from hardirq (IPI) context
1110 static void __hrtick_start(void *arg)
1112 struct rq *rq = arg;
1114 spin_lock(&rq->lock);
1115 hrtimer_restart(&rq->hrtick_timer);
1116 rq->hrtick_csd_pending = 0;
1117 spin_unlock(&rq->lock);
1121 * Called to set the hrtick timer state.
1123 * called with rq->lock held and irqs disabled
1125 static void hrtick_start(struct rq *rq, u64 delay)
1127 struct hrtimer *timer = &rq->hrtick_timer;
1128 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1130 hrtimer_set_expires(timer, time);
1132 if (rq == this_rq()) {
1133 hrtimer_restart(timer);
1134 } else if (!rq->hrtick_csd_pending) {
1135 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1136 rq->hrtick_csd_pending = 1;
1141 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1143 int cpu = (int)(long)hcpu;
1146 case CPU_UP_CANCELED:
1147 case CPU_UP_CANCELED_FROZEN:
1148 case CPU_DOWN_PREPARE:
1149 case CPU_DOWN_PREPARE_FROZEN:
1151 case CPU_DEAD_FROZEN:
1152 hrtick_clear(cpu_rq(cpu));
1159 static __init void init_hrtick(void)
1161 hotcpu_notifier(hotplug_hrtick, 0);
1165 * Called to set the hrtick timer state.
1167 * called with rq->lock held and irqs disabled
1169 static void hrtick_start(struct rq *rq, u64 delay)
1171 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1172 HRTIMER_MODE_REL_PINNED, 0);
1175 static inline void init_hrtick(void)
1178 #endif /* CONFIG_SMP */
1180 static void init_rq_hrtick(struct rq *rq)
1183 rq->hrtick_csd_pending = 0;
1185 rq->hrtick_csd.flags = 0;
1186 rq->hrtick_csd.func = __hrtick_start;
1187 rq->hrtick_csd.info = rq;
1190 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1191 rq->hrtick_timer.function = hrtick;
1193 #else /* CONFIG_SCHED_HRTICK */
1194 static inline void hrtick_clear(struct rq *rq)
1198 static inline void init_rq_hrtick(struct rq *rq)
1202 static inline void init_hrtick(void)
1205 #endif /* CONFIG_SCHED_HRTICK */
1208 * resched_task - mark a task 'to be rescheduled now'.
1210 * On UP this means the setting of the need_resched flag, on SMP it
1211 * might also involve a cross-CPU call to trigger the scheduler on
1216 #ifndef tsk_is_polling
1217 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1220 static void resched_task(struct task_struct *p)
1224 assert_spin_locked(&task_rq(p)->lock);
1226 if (test_tsk_need_resched(p))
1229 set_tsk_need_resched(p);
1232 if (cpu == smp_processor_id())
1235 /* NEED_RESCHED must be visible before we test polling */
1237 if (!tsk_is_polling(p))
1238 smp_send_reschedule(cpu);
1241 static void resched_cpu(int cpu)
1243 struct rq *rq = cpu_rq(cpu);
1244 unsigned long flags;
1246 if (!spin_trylock_irqsave(&rq->lock, flags))
1248 resched_task(cpu_curr(cpu));
1249 spin_unlock_irqrestore(&rq->lock, flags);
1254 * When add_timer_on() enqueues a timer into the timer wheel of an
1255 * idle CPU then this timer might expire before the next timer event
1256 * which is scheduled to wake up that CPU. In case of a completely
1257 * idle system the next event might even be infinite time into the
1258 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1259 * leaves the inner idle loop so the newly added timer is taken into
1260 * account when the CPU goes back to idle and evaluates the timer
1261 * wheel for the next timer event.
1263 void wake_up_idle_cpu(int cpu)
1265 struct rq *rq = cpu_rq(cpu);
1267 if (cpu == smp_processor_id())
1271 * This is safe, as this function is called with the timer
1272 * wheel base lock of (cpu) held. When the CPU is on the way
1273 * to idle and has not yet set rq->curr to idle then it will
1274 * be serialized on the timer wheel base lock and take the new
1275 * timer into account automatically.
1277 if (rq->curr != rq->idle)
1281 * We can set TIF_RESCHED on the idle task of the other CPU
1282 * lockless. The worst case is that the other CPU runs the
1283 * idle task through an additional NOOP schedule()
1285 set_tsk_need_resched(rq->idle);
1287 /* NEED_RESCHED must be visible before we test polling */
1289 if (!tsk_is_polling(rq->idle))
1290 smp_send_reschedule(cpu);
1292 #endif /* CONFIG_NO_HZ */
1294 static u64 sched_avg_period(void)
1296 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1299 static void sched_avg_update(struct rq *rq)
1301 s64 period = sched_avg_period();
1303 while ((s64)(rq->clock - rq->age_stamp) > period) {
1304 rq->age_stamp += period;
1309 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1311 rq->rt_avg += rt_delta;
1312 sched_avg_update(rq);
1315 #else /* !CONFIG_SMP */
1316 static void resched_task(struct task_struct *p)
1318 assert_spin_locked(&task_rq(p)->lock);
1319 set_tsk_need_resched(p);
1322 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1325 #endif /* CONFIG_SMP */
1327 #if BITS_PER_LONG == 32
1328 # define WMULT_CONST (~0UL)
1330 # define WMULT_CONST (1UL << 32)
1333 #define WMULT_SHIFT 32
1336 * Shift right and round:
1338 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1341 * delta *= weight / lw
1343 static unsigned long
1344 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1345 struct load_weight *lw)
1349 if (!lw->inv_weight) {
1350 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1353 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1357 tmp = (u64)delta_exec * weight;
1359 * Check whether we'd overflow the 64-bit multiplication:
1361 if (unlikely(tmp > WMULT_CONST))
1362 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1365 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1367 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1370 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1376 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1383 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1384 * of tasks with abnormal "nice" values across CPUs the contribution that
1385 * each task makes to its run queue's load is weighted according to its
1386 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1387 * scaled version of the new time slice allocation that they receive on time
1391 #define WEIGHT_IDLEPRIO 3
1392 #define WMULT_IDLEPRIO 1431655765
1395 * Nice levels are multiplicative, with a gentle 10% change for every
1396 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1397 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1398 * that remained on nice 0.
1400 * The "10% effect" is relative and cumulative: from _any_ nice level,
1401 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1402 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1403 * If a task goes up by ~10% and another task goes down by ~10% then
1404 * the relative distance between them is ~25%.)
1406 static const int prio_to_weight[40] = {
1407 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1408 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1409 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1410 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1411 /* 0 */ 1024, 820, 655, 526, 423,
1412 /* 5 */ 335, 272, 215, 172, 137,
1413 /* 10 */ 110, 87, 70, 56, 45,
1414 /* 15 */ 36, 29, 23, 18, 15,
1418 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1420 * In cases where the weight does not change often, we can use the
1421 * precalculated inverse to speed up arithmetics by turning divisions
1422 * into multiplications:
1424 static const u32 prio_to_wmult[40] = {
1425 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1426 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1427 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1428 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1429 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1430 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1431 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1432 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1435 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1438 * runqueue iterator, to support SMP load-balancing between different
1439 * scheduling classes, without having to expose their internal data
1440 * structures to the load-balancing proper:
1442 struct rq_iterator {
1444 struct task_struct *(*start)(void *);
1445 struct task_struct *(*next)(void *);
1449 static unsigned long
1450 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1451 unsigned long max_load_move, struct sched_domain *sd,
1452 enum cpu_idle_type idle, int *all_pinned,
1453 int *this_best_prio, struct rq_iterator *iterator);
1456 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1457 struct sched_domain *sd, enum cpu_idle_type idle,
1458 struct rq_iterator *iterator);
1461 /* Time spent by the tasks of the cpu accounting group executing in ... */
1462 enum cpuacct_stat_index {
1463 CPUACCT_STAT_USER, /* ... user mode */
1464 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1466 CPUACCT_STAT_NSTATS,
1469 #ifdef CONFIG_CGROUP_CPUACCT
1470 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1471 static void cpuacct_update_stats(struct task_struct *tsk,
1472 enum cpuacct_stat_index idx, cputime_t val);
1474 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1475 static inline void cpuacct_update_stats(struct task_struct *tsk,
1476 enum cpuacct_stat_index idx, cputime_t val) {}
1479 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1481 update_load_add(&rq->load, load);
1484 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1486 update_load_sub(&rq->load, load);
1489 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1490 typedef int (*tg_visitor)(struct task_group *, void *);
1493 * Iterate the full tree, calling @down when first entering a node and @up when
1494 * leaving it for the final time.
1496 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1498 struct task_group *parent, *child;
1502 parent = &root_task_group;
1504 ret = (*down)(parent, data);
1507 list_for_each_entry_rcu(child, &parent->children, siblings) {
1514 ret = (*up)(parent, data);
1519 parent = parent->parent;
1528 static int tg_nop(struct task_group *tg, void *data)
1535 static unsigned long source_load(int cpu, int type);
1536 static unsigned long target_load(int cpu, int type);
1537 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1539 static unsigned long cpu_avg_load_per_task(int cpu)
1541 struct rq *rq = cpu_rq(cpu);
1542 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1545 rq->avg_load_per_task = rq->load.weight / nr_running;
1547 rq->avg_load_per_task = 0;
1549 return rq->avg_load_per_task;
1552 #ifdef CONFIG_FAIR_GROUP_SCHED
1554 struct update_shares_data {
1555 unsigned long rq_weight[NR_CPUS];
1558 static DEFINE_PER_CPU(struct update_shares_data, update_shares_data);
1560 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1563 * Calculate and set the cpu's group shares.
1565 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1566 unsigned long sd_shares,
1567 unsigned long sd_rq_weight,
1568 struct update_shares_data *usd)
1570 unsigned long shares, rq_weight;
1573 rq_weight = usd->rq_weight[cpu];
1576 rq_weight = NICE_0_LOAD;
1580 * \Sum_j shares_j * rq_weight_i
1581 * shares_i = -----------------------------
1582 * \Sum_j rq_weight_j
1584 shares = (sd_shares * rq_weight) / sd_rq_weight;
1585 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1587 if (abs(shares - tg->se[cpu]->load.weight) >
1588 sysctl_sched_shares_thresh) {
1589 struct rq *rq = cpu_rq(cpu);
1590 unsigned long flags;
1592 spin_lock_irqsave(&rq->lock, flags);
1593 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1594 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1595 __set_se_shares(tg->se[cpu], shares);
1596 spin_unlock_irqrestore(&rq->lock, flags);
1601 * Re-compute the task group their per cpu shares over the given domain.
1602 * This needs to be done in a bottom-up fashion because the rq weight of a
1603 * parent group depends on the shares of its child groups.
1605 static int tg_shares_up(struct task_group *tg, void *data)
1607 unsigned long weight, rq_weight = 0, shares = 0;
1608 struct update_shares_data *usd;
1609 struct sched_domain *sd = data;
1610 unsigned long flags;
1616 local_irq_save(flags);
1617 usd = &__get_cpu_var(update_shares_data);
1619 for_each_cpu(i, sched_domain_span(sd)) {
1620 weight = tg->cfs_rq[i]->load.weight;
1621 usd->rq_weight[i] = weight;
1624 * If there are currently no tasks on the cpu pretend there
1625 * is one of average load so that when a new task gets to
1626 * run here it will not get delayed by group starvation.
1629 weight = NICE_0_LOAD;
1631 rq_weight += weight;
1632 shares += tg->cfs_rq[i]->shares;
1635 if ((!shares && rq_weight) || shares > tg->shares)
1636 shares = tg->shares;
1638 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1639 shares = tg->shares;
1641 for_each_cpu(i, sched_domain_span(sd))
1642 update_group_shares_cpu(tg, i, shares, rq_weight, usd);
1644 local_irq_restore(flags);
1650 * Compute the cpu's hierarchical load factor for each task group.
1651 * This needs to be done in a top-down fashion because the load of a child
1652 * group is a fraction of its parents load.
1654 static int tg_load_down(struct task_group *tg, void *data)
1657 long cpu = (long)data;
1660 load = cpu_rq(cpu)->load.weight;
1662 load = tg->parent->cfs_rq[cpu]->h_load;
1663 load *= tg->cfs_rq[cpu]->shares;
1664 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1667 tg->cfs_rq[cpu]->h_load = load;
1672 static void update_shares(struct sched_domain *sd)
1677 if (root_task_group_empty())
1680 now = cpu_clock(raw_smp_processor_id());
1681 elapsed = now - sd->last_update;
1683 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1684 sd->last_update = now;
1685 walk_tg_tree(tg_nop, tg_shares_up, sd);
1689 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1691 if (root_task_group_empty())
1694 spin_unlock(&rq->lock);
1696 spin_lock(&rq->lock);
1699 static void update_h_load(long cpu)
1701 if (root_task_group_empty())
1704 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1709 static inline void update_shares(struct sched_domain *sd)
1713 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1719 #ifdef CONFIG_PREEMPT
1722 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1723 * way at the expense of forcing extra atomic operations in all
1724 * invocations. This assures that the double_lock is acquired using the
1725 * same underlying policy as the spinlock_t on this architecture, which
1726 * reduces latency compared to the unfair variant below. However, it
1727 * also adds more overhead and therefore may reduce throughput.
1729 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1730 __releases(this_rq->lock)
1731 __acquires(busiest->lock)
1732 __acquires(this_rq->lock)
1734 spin_unlock(&this_rq->lock);
1735 double_rq_lock(this_rq, busiest);
1742 * Unfair double_lock_balance: Optimizes throughput at the expense of
1743 * latency by eliminating extra atomic operations when the locks are
1744 * already in proper order on entry. This favors lower cpu-ids and will
1745 * grant the double lock to lower cpus over higher ids under contention,
1746 * regardless of entry order into the function.
1748 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1749 __releases(this_rq->lock)
1750 __acquires(busiest->lock)
1751 __acquires(this_rq->lock)
1755 if (unlikely(!spin_trylock(&busiest->lock))) {
1756 if (busiest < this_rq) {
1757 spin_unlock(&this_rq->lock);
1758 spin_lock(&busiest->lock);
1759 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1762 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1767 #endif /* CONFIG_PREEMPT */
1770 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1772 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1774 if (unlikely(!irqs_disabled())) {
1775 /* printk() doesn't work good under rq->lock */
1776 spin_unlock(&this_rq->lock);
1780 return _double_lock_balance(this_rq, busiest);
1783 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1784 __releases(busiest->lock)
1786 spin_unlock(&busiest->lock);
1787 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1791 #ifdef CONFIG_FAIR_GROUP_SCHED
1792 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1795 cfs_rq->shares = shares;
1800 static void calc_load_account_active(struct rq *this_rq);
1802 #include "sched_stats.h"
1803 #include "sched_idletask.c"
1804 #include "sched_fair.c"
1805 #include "sched_rt.c"
1806 #ifdef CONFIG_SCHED_DEBUG
1807 # include "sched_debug.c"
1810 #define sched_class_highest (&rt_sched_class)
1811 #define for_each_class(class) \
1812 for (class = sched_class_highest; class; class = class->next)
1814 static void inc_nr_running(struct rq *rq)
1819 static void dec_nr_running(struct rq *rq)
1824 static void set_load_weight(struct task_struct *p)
1826 if (task_has_rt_policy(p)) {
1827 p->se.load.weight = prio_to_weight[0] * 2;
1828 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1833 * SCHED_IDLE tasks get minimal weight:
1835 if (p->policy == SCHED_IDLE) {
1836 p->se.load.weight = WEIGHT_IDLEPRIO;
1837 p->se.load.inv_weight = WMULT_IDLEPRIO;
1841 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1842 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1845 static void update_avg(u64 *avg, u64 sample)
1847 s64 diff = sample - *avg;
1851 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1854 p->se.start_runtime = p->se.sum_exec_runtime;
1856 sched_info_queued(p);
1857 p->sched_class->enqueue_task(rq, p, wakeup);
1861 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1864 if (p->se.last_wakeup) {
1865 update_avg(&p->se.avg_overlap,
1866 p->se.sum_exec_runtime - p->se.last_wakeup);
1867 p->se.last_wakeup = 0;
1869 update_avg(&p->se.avg_wakeup,
1870 sysctl_sched_wakeup_granularity);
1874 sched_info_dequeued(p);
1875 p->sched_class->dequeue_task(rq, p, sleep);
1880 * __normal_prio - return the priority that is based on the static prio
1882 static inline int __normal_prio(struct task_struct *p)
1884 return p->static_prio;
1888 * Calculate the expected normal priority: i.e. priority
1889 * without taking RT-inheritance into account. Might be
1890 * boosted by interactivity modifiers. Changes upon fork,
1891 * setprio syscalls, and whenever the interactivity
1892 * estimator recalculates.
1894 static inline int normal_prio(struct task_struct *p)
1898 if (task_has_rt_policy(p))
1899 prio = MAX_RT_PRIO-1 - p->rt_priority;
1901 prio = __normal_prio(p);
1906 * Calculate the current priority, i.e. the priority
1907 * taken into account by the scheduler. This value might
1908 * be boosted by RT tasks, or might be boosted by
1909 * interactivity modifiers. Will be RT if the task got
1910 * RT-boosted. If not then it returns p->normal_prio.
1912 static int effective_prio(struct task_struct *p)
1914 p->normal_prio = normal_prio(p);
1916 * If we are RT tasks or we were boosted to RT priority,
1917 * keep the priority unchanged. Otherwise, update priority
1918 * to the normal priority:
1920 if (!rt_prio(p->prio))
1921 return p->normal_prio;
1926 * activate_task - move a task to the runqueue.
1928 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1930 if (task_contributes_to_load(p))
1931 rq->nr_uninterruptible--;
1933 enqueue_task(rq, p, wakeup);
1938 * deactivate_task - remove a task from the runqueue.
1940 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1942 if (task_contributes_to_load(p))
1943 rq->nr_uninterruptible++;
1945 dequeue_task(rq, p, sleep);
1950 * task_curr - is this task currently executing on a CPU?
1951 * @p: the task in question.
1953 inline int task_curr(const struct task_struct *p)
1955 return cpu_curr(task_cpu(p)) == p;
1958 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1960 set_task_rq(p, cpu);
1963 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1964 * successfuly executed on another CPU. We must ensure that updates of
1965 * per-task data have been completed by this moment.
1968 task_thread_info(p)->cpu = cpu;
1972 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1973 const struct sched_class *prev_class,
1974 int oldprio, int running)
1976 if (prev_class != p->sched_class) {
1977 if (prev_class->switched_from)
1978 prev_class->switched_from(rq, p, running);
1979 p->sched_class->switched_to(rq, p, running);
1981 p->sched_class->prio_changed(rq, p, oldprio, running);
1986 /* Used instead of source_load when we know the type == 0 */
1987 static unsigned long weighted_cpuload(const int cpu)
1989 return cpu_rq(cpu)->load.weight;
1993 * Is this task likely cache-hot:
1996 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2001 * Buddy candidates are cache hot:
2003 if (sched_feat(CACHE_HOT_BUDDY) &&
2004 (&p->se == cfs_rq_of(&p->se)->next ||
2005 &p->se == cfs_rq_of(&p->se)->last))
2008 if (p->sched_class != &fair_sched_class)
2011 if (sysctl_sched_migration_cost == -1)
2013 if (sysctl_sched_migration_cost == 0)
2016 delta = now - p->se.exec_start;
2018 return delta < (s64)sysctl_sched_migration_cost;
2022 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2024 int old_cpu = task_cpu(p);
2025 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2026 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2027 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2030 clock_offset = old_rq->clock - new_rq->clock;
2032 trace_sched_migrate_task(p, new_cpu);
2034 #ifdef CONFIG_SCHEDSTATS
2035 if (p->se.wait_start)
2036 p->se.wait_start -= clock_offset;
2037 if (p->se.sleep_start)
2038 p->se.sleep_start -= clock_offset;
2039 if (p->se.block_start)
2040 p->se.block_start -= clock_offset;
2042 if (old_cpu != new_cpu) {
2043 p->se.nr_migrations++;
2044 new_rq->nr_migrations_in++;
2045 #ifdef CONFIG_SCHEDSTATS
2046 if (task_hot(p, old_rq->clock, NULL))
2047 schedstat_inc(p, se.nr_forced2_migrations);
2049 perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2052 p->se.vruntime -= old_cfsrq->min_vruntime -
2053 new_cfsrq->min_vruntime;
2055 __set_task_cpu(p, new_cpu);
2058 struct migration_req {
2059 struct list_head list;
2061 struct task_struct *task;
2064 struct completion done;
2068 * The task's runqueue lock must be held.
2069 * Returns true if you have to wait for migration thread.
2072 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2074 struct rq *rq = task_rq(p);
2077 * If the task is not on a runqueue (and not running), then
2078 * it is sufficient to simply update the task's cpu field.
2080 if (!p->se.on_rq && !task_running(rq, p)) {
2081 set_task_cpu(p, dest_cpu);
2085 init_completion(&req->done);
2087 req->dest_cpu = dest_cpu;
2088 list_add(&req->list, &rq->migration_queue);
2094 * wait_task_context_switch - wait for a thread to complete at least one
2097 * @p must not be current.
2099 void wait_task_context_switch(struct task_struct *p)
2101 unsigned long nvcsw, nivcsw, flags;
2109 * The runqueue is assigned before the actual context
2110 * switch. We need to take the runqueue lock.
2112 * We could check initially without the lock but it is
2113 * very likely that we need to take the lock in every
2116 rq = task_rq_lock(p, &flags);
2117 running = task_running(rq, p);
2118 task_rq_unlock(rq, &flags);
2120 if (likely(!running))
2123 * The switch count is incremented before the actual
2124 * context switch. We thus wait for two switches to be
2125 * sure at least one completed.
2127 if ((p->nvcsw - nvcsw) > 1)
2129 if ((p->nivcsw - nivcsw) > 1)
2137 * wait_task_inactive - wait for a thread to unschedule.
2139 * If @match_state is nonzero, it's the @p->state value just checked and
2140 * not expected to change. If it changes, i.e. @p might have woken up,
2141 * then return zero. When we succeed in waiting for @p to be off its CPU,
2142 * we return a positive number (its total switch count). If a second call
2143 * a short while later returns the same number, the caller can be sure that
2144 * @p has remained unscheduled the whole time.
2146 * The caller must ensure that the task *will* unschedule sometime soon,
2147 * else this function might spin for a *long* time. This function can't
2148 * be called with interrupts off, or it may introduce deadlock with
2149 * smp_call_function() if an IPI is sent by the same process we are
2150 * waiting to become inactive.
2152 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2154 unsigned long flags;
2161 * We do the initial early heuristics without holding
2162 * any task-queue locks at all. We'll only try to get
2163 * the runqueue lock when things look like they will
2169 * If the task is actively running on another CPU
2170 * still, just relax and busy-wait without holding
2173 * NOTE! Since we don't hold any locks, it's not
2174 * even sure that "rq" stays as the right runqueue!
2175 * But we don't care, since "task_running()" will
2176 * return false if the runqueue has changed and p
2177 * is actually now running somewhere else!
2179 while (task_running(rq, p)) {
2180 if (match_state && unlikely(p->state != match_state))
2186 * Ok, time to look more closely! We need the rq
2187 * lock now, to be *sure*. If we're wrong, we'll
2188 * just go back and repeat.
2190 rq = task_rq_lock(p, &flags);
2191 trace_sched_wait_task(rq, p);
2192 running = task_running(rq, p);
2193 on_rq = p->se.on_rq;
2195 if (!match_state || p->state == match_state)
2196 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2197 task_rq_unlock(rq, &flags);
2200 * If it changed from the expected state, bail out now.
2202 if (unlikely(!ncsw))
2206 * Was it really running after all now that we
2207 * checked with the proper locks actually held?
2209 * Oops. Go back and try again..
2211 if (unlikely(running)) {
2217 * It's not enough that it's not actively running,
2218 * it must be off the runqueue _entirely_, and not
2221 * So if it was still runnable (but just not actively
2222 * running right now), it's preempted, and we should
2223 * yield - it could be a while.
2225 if (unlikely(on_rq)) {
2226 schedule_timeout_uninterruptible(1);
2231 * Ahh, all good. It wasn't running, and it wasn't
2232 * runnable, which means that it will never become
2233 * running in the future either. We're all done!
2242 * kick_process - kick a running thread to enter/exit the kernel
2243 * @p: the to-be-kicked thread
2245 * Cause a process which is running on another CPU to enter
2246 * kernel-mode, without any delay. (to get signals handled.)
2248 * NOTE: this function doesnt have to take the runqueue lock,
2249 * because all it wants to ensure is that the remote task enters
2250 * the kernel. If the IPI races and the task has been migrated
2251 * to another CPU then no harm is done and the purpose has been
2254 void kick_process(struct task_struct *p)
2260 if ((cpu != smp_processor_id()) && task_curr(p))
2261 smp_send_reschedule(cpu);
2264 EXPORT_SYMBOL_GPL(kick_process);
2267 * Return a low guess at the load of a migration-source cpu weighted
2268 * according to the scheduling class and "nice" value.
2270 * We want to under-estimate the load of migration sources, to
2271 * balance conservatively.
2273 static unsigned long source_load(int cpu, int type)
2275 struct rq *rq = cpu_rq(cpu);
2276 unsigned long total = weighted_cpuload(cpu);
2278 if (type == 0 || !sched_feat(LB_BIAS))
2281 return min(rq->cpu_load[type-1], total);
2285 * Return a high guess at the load of a migration-target cpu weighted
2286 * according to the scheduling class and "nice" value.
2288 static unsigned long target_load(int cpu, int type)
2290 struct rq *rq = cpu_rq(cpu);
2291 unsigned long total = weighted_cpuload(cpu);
2293 if (type == 0 || !sched_feat(LB_BIAS))
2296 return max(rq->cpu_load[type-1], total);
2300 * find_idlest_group finds and returns the least busy CPU group within the
2303 static struct sched_group *
2304 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2306 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2307 unsigned long min_load = ULONG_MAX, this_load = 0;
2308 int load_idx = sd->forkexec_idx;
2309 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2312 unsigned long load, avg_load;
2316 /* Skip over this group if it has no CPUs allowed */
2317 if (!cpumask_intersects(sched_group_cpus(group),
2321 local_group = cpumask_test_cpu(this_cpu,
2322 sched_group_cpus(group));
2324 /* Tally up the load of all CPUs in the group */
2327 for_each_cpu(i, sched_group_cpus(group)) {
2328 /* Bias balancing toward cpus of our domain */
2330 load = source_load(i, load_idx);
2332 load = target_load(i, load_idx);
2337 /* Adjust by relative CPU power of the group */
2338 avg_load = sg_div_cpu_power(group,
2339 avg_load * SCHED_LOAD_SCALE);
2342 this_load = avg_load;
2344 } else if (avg_load < min_load) {
2345 min_load = avg_load;
2348 } while (group = group->next, group != sd->groups);
2350 if (!idlest || 100*this_load < imbalance*min_load)
2356 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2359 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2361 unsigned long load, min_load = ULONG_MAX;
2365 /* Traverse only the allowed CPUs */
2366 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2367 load = weighted_cpuload(i);
2369 if (load < min_load || (load == min_load && i == this_cpu)) {
2379 * sched_balance_self: balance the current task (running on cpu) in domains
2380 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2383 * Balance, ie. select the least loaded group.
2385 * Returns the target CPU number, or the same CPU if no balancing is needed.
2387 * preempt must be disabled.
2389 static int sched_balance_self(int cpu, int flag)
2391 struct task_struct *t = current;
2392 struct sched_domain *tmp, *sd = NULL;
2394 for_each_domain(cpu, tmp) {
2396 * If power savings logic is enabled for a domain, stop there.
2398 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2400 if (tmp->flags & flag)
2408 struct sched_group *group;
2409 int new_cpu, weight;
2411 if (!(sd->flags & flag)) {
2416 group = find_idlest_group(sd, t, cpu);
2422 new_cpu = find_idlest_cpu(group, t, cpu);
2423 if (new_cpu == -1 || new_cpu == cpu) {
2424 /* Now try balancing at a lower domain level of cpu */
2429 /* Now try balancing at a lower domain level of new_cpu */
2431 weight = cpumask_weight(sched_domain_span(sd));
2433 for_each_domain(cpu, tmp) {
2434 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2436 if (tmp->flags & flag)
2439 /* while loop will break here if sd == NULL */
2445 #endif /* CONFIG_SMP */
2448 * task_oncpu_function_call - call a function on the cpu on which a task runs
2449 * @p: the task to evaluate
2450 * @func: the function to be called
2451 * @info: the function call argument
2453 * Calls the function @func when the task is currently running. This might
2454 * be on the current CPU, which just calls the function directly
2456 void task_oncpu_function_call(struct task_struct *p,
2457 void (*func) (void *info), void *info)
2464 smp_call_function_single(cpu, func, info, 1);
2469 * try_to_wake_up - wake up a thread
2470 * @p: the to-be-woken-up thread
2471 * @state: the mask of task states that can be woken
2472 * @sync: do a synchronous wakeup?
2474 * Put it on the run-queue if it's not already there. The "current"
2475 * thread is always on the run-queue (except when the actual
2476 * re-schedule is in progress), and as such you're allowed to do
2477 * the simpler "current->state = TASK_RUNNING" to mark yourself
2478 * runnable without the overhead of this.
2480 * returns failure only if the task is already active.
2482 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2484 int cpu, orig_cpu, this_cpu, success = 0;
2485 unsigned long flags;
2489 if (!sched_feat(SYNC_WAKEUPS))
2493 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2494 struct sched_domain *sd;
2496 this_cpu = raw_smp_processor_id();
2499 for_each_domain(this_cpu, sd) {
2500 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2509 rq = task_rq_lock(p, &flags);
2510 update_rq_clock(rq);
2511 old_state = p->state;
2512 if (!(old_state & state))
2520 this_cpu = smp_processor_id();
2523 if (unlikely(task_running(rq, p)))
2526 cpu = p->sched_class->select_task_rq(p, sync);
2527 if (cpu != orig_cpu) {
2528 set_task_cpu(p, cpu);
2529 task_rq_unlock(rq, &flags);
2530 /* might preempt at this point */
2531 rq = task_rq_lock(p, &flags);
2532 old_state = p->state;
2533 if (!(old_state & state))
2538 this_cpu = smp_processor_id();
2542 #ifdef CONFIG_SCHEDSTATS
2543 schedstat_inc(rq, ttwu_count);
2544 if (cpu == this_cpu)
2545 schedstat_inc(rq, ttwu_local);
2547 struct sched_domain *sd;
2548 for_each_domain(this_cpu, sd) {
2549 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2550 schedstat_inc(sd, ttwu_wake_remote);
2555 #endif /* CONFIG_SCHEDSTATS */
2558 #endif /* CONFIG_SMP */
2559 schedstat_inc(p, se.nr_wakeups);
2561 schedstat_inc(p, se.nr_wakeups_sync);
2562 if (orig_cpu != cpu)
2563 schedstat_inc(p, se.nr_wakeups_migrate);
2564 if (cpu == this_cpu)
2565 schedstat_inc(p, se.nr_wakeups_local);
2567 schedstat_inc(p, se.nr_wakeups_remote);
2568 activate_task(rq, p, 1);
2572 * Only attribute actual wakeups done by this task.
2574 if (!in_interrupt()) {
2575 struct sched_entity *se = ¤t->se;
2576 u64 sample = se->sum_exec_runtime;
2578 if (se->last_wakeup)
2579 sample -= se->last_wakeup;
2581 sample -= se->start_runtime;
2582 update_avg(&se->avg_wakeup, sample);
2584 se->last_wakeup = se->sum_exec_runtime;
2588 trace_sched_wakeup(rq, p, success);
2589 check_preempt_curr(rq, p, sync);
2591 p->state = TASK_RUNNING;
2593 if (p->sched_class->task_wake_up)
2594 p->sched_class->task_wake_up(rq, p);
2597 task_rq_unlock(rq, &flags);
2603 * wake_up_process - Wake up a specific process
2604 * @p: The process to be woken up.
2606 * Attempt to wake up the nominated process and move it to the set of runnable
2607 * processes. Returns 1 if the process was woken up, 0 if it was already
2610 * It may be assumed that this function implies a write memory barrier before
2611 * changing the task state if and only if any tasks are woken up.
2613 int wake_up_process(struct task_struct *p)
2615 return try_to_wake_up(p, TASK_ALL, 0);
2617 EXPORT_SYMBOL(wake_up_process);
2619 int wake_up_state(struct task_struct *p, unsigned int state)
2621 return try_to_wake_up(p, state, 0);
2625 * Perform scheduler related setup for a newly forked process p.
2626 * p is forked by current.
2628 * __sched_fork() is basic setup used by init_idle() too:
2630 static void __sched_fork(struct task_struct *p)
2632 p->se.exec_start = 0;
2633 p->se.sum_exec_runtime = 0;
2634 p->se.prev_sum_exec_runtime = 0;
2635 p->se.nr_migrations = 0;
2636 p->se.last_wakeup = 0;
2637 p->se.avg_overlap = 0;
2638 p->se.start_runtime = 0;
2639 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2641 #ifdef CONFIG_SCHEDSTATS
2642 p->se.wait_start = 0;
2644 p->se.wait_count = 0;
2647 p->se.sleep_start = 0;
2648 p->se.sleep_max = 0;
2649 p->se.sum_sleep_runtime = 0;
2651 p->se.block_start = 0;
2652 p->se.block_max = 0;
2654 p->se.slice_max = 0;
2656 p->se.nr_migrations_cold = 0;
2657 p->se.nr_failed_migrations_affine = 0;
2658 p->se.nr_failed_migrations_running = 0;
2659 p->se.nr_failed_migrations_hot = 0;
2660 p->se.nr_forced_migrations = 0;
2661 p->se.nr_forced2_migrations = 0;
2663 p->se.nr_wakeups = 0;
2664 p->se.nr_wakeups_sync = 0;
2665 p->se.nr_wakeups_migrate = 0;
2666 p->se.nr_wakeups_local = 0;
2667 p->se.nr_wakeups_remote = 0;
2668 p->se.nr_wakeups_affine = 0;
2669 p->se.nr_wakeups_affine_attempts = 0;
2670 p->se.nr_wakeups_passive = 0;
2671 p->se.nr_wakeups_idle = 0;
2675 INIT_LIST_HEAD(&p->rt.run_list);
2677 INIT_LIST_HEAD(&p->se.group_node);
2679 #ifdef CONFIG_PREEMPT_NOTIFIERS
2680 INIT_HLIST_HEAD(&p->preempt_notifiers);
2684 * We mark the process as running here, but have not actually
2685 * inserted it onto the runqueue yet. This guarantees that
2686 * nobody will actually run it, and a signal or other external
2687 * event cannot wake it up and insert it on the runqueue either.
2689 p->state = TASK_RUNNING;
2693 * fork()/clone()-time setup:
2695 void sched_fork(struct task_struct *p, int clone_flags)
2697 int cpu = get_cpu();
2702 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2704 set_task_cpu(p, cpu);
2707 * Make sure we do not leak PI boosting priority to the child.
2709 p->prio = current->normal_prio;
2712 * Revert to default priority/policy on fork if requested.
2714 if (unlikely(p->sched_reset_on_fork)) {
2715 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR)
2716 p->policy = SCHED_NORMAL;
2718 if (p->normal_prio < DEFAULT_PRIO)
2719 p->prio = DEFAULT_PRIO;
2721 if (PRIO_TO_NICE(p->static_prio) < 0) {
2722 p->static_prio = NICE_TO_PRIO(0);
2727 * We don't need the reset flag anymore after the fork. It has
2728 * fulfilled its duty:
2730 p->sched_reset_on_fork = 0;
2733 if (!rt_prio(p->prio))
2734 p->sched_class = &fair_sched_class;
2736 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2737 if (likely(sched_info_on()))
2738 memset(&p->sched_info, 0, sizeof(p->sched_info));
2740 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2743 #ifdef CONFIG_PREEMPT
2744 /* Want to start with kernel preemption disabled. */
2745 task_thread_info(p)->preempt_count = 1;
2747 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2753 * wake_up_new_task - wake up a newly created task for the first time.
2755 * This function will do some initial scheduler statistics housekeeping
2756 * that must be done for every newly created context, then puts the task
2757 * on the runqueue and wakes it.
2759 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2761 unsigned long flags;
2764 rq = task_rq_lock(p, &flags);
2765 BUG_ON(p->state != TASK_RUNNING);
2766 update_rq_clock(rq);
2768 p->prio = effective_prio(p);
2770 if (!p->sched_class->task_new || !current->se.on_rq) {
2771 activate_task(rq, p, 0);
2774 * Let the scheduling class do new task startup
2775 * management (if any):
2777 p->sched_class->task_new(rq, p);
2780 trace_sched_wakeup_new(rq, p, 1);
2781 check_preempt_curr(rq, p, 0);
2783 if (p->sched_class->task_wake_up)
2784 p->sched_class->task_wake_up(rq, p);
2786 task_rq_unlock(rq, &flags);
2789 #ifdef CONFIG_PREEMPT_NOTIFIERS
2792 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2793 * @notifier: notifier struct to register
2795 void preempt_notifier_register(struct preempt_notifier *notifier)
2797 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2799 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2802 * preempt_notifier_unregister - no longer interested in preemption notifications
2803 * @notifier: notifier struct to unregister
2805 * This is safe to call from within a preemption notifier.
2807 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2809 hlist_del(¬ifier->link);
2811 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2813 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2815 struct preempt_notifier *notifier;
2816 struct hlist_node *node;
2818 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2819 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2823 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2824 struct task_struct *next)
2826 struct preempt_notifier *notifier;
2827 struct hlist_node *node;
2829 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2830 notifier->ops->sched_out(notifier, next);
2833 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2835 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2840 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2841 struct task_struct *next)
2845 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2848 * prepare_task_switch - prepare to switch tasks
2849 * @rq: the runqueue preparing to switch
2850 * @prev: the current task that is being switched out
2851 * @next: the task we are going to switch to.
2853 * This is called with the rq lock held and interrupts off. It must
2854 * be paired with a subsequent finish_task_switch after the context
2857 * prepare_task_switch sets up locking and calls architecture specific
2861 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2862 struct task_struct *next)
2864 fire_sched_out_preempt_notifiers(prev, next);
2865 prepare_lock_switch(rq, next);
2866 prepare_arch_switch(next);
2870 * finish_task_switch - clean up after a task-switch
2871 * @rq: runqueue associated with task-switch
2872 * @prev: the thread we just switched away from.
2874 * finish_task_switch must be called after the context switch, paired
2875 * with a prepare_task_switch call before the context switch.
2876 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2877 * and do any other architecture-specific cleanup actions.
2879 * Note that we may have delayed dropping an mm in context_switch(). If
2880 * so, we finish that here outside of the runqueue lock. (Doing it
2881 * with the lock held can cause deadlocks; see schedule() for
2884 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2885 __releases(rq->lock)
2887 struct mm_struct *mm = rq->prev_mm;
2893 * A task struct has one reference for the use as "current".
2894 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2895 * schedule one last time. The schedule call will never return, and
2896 * the scheduled task must drop that reference.
2897 * The test for TASK_DEAD must occur while the runqueue locks are
2898 * still held, otherwise prev could be scheduled on another cpu, die
2899 * there before we look at prev->state, and then the reference would
2901 * Manfred Spraul <manfred@colorfullife.com>
2903 prev_state = prev->state;
2904 finish_arch_switch(prev);
2905 perf_counter_task_sched_in(current, cpu_of(rq));
2906 finish_lock_switch(rq, prev);
2908 fire_sched_in_preempt_notifiers(current);
2911 if (unlikely(prev_state == TASK_DEAD)) {
2913 * Remove function-return probe instances associated with this
2914 * task and put them back on the free list.
2916 kprobe_flush_task(prev);
2917 put_task_struct(prev);
2923 /* assumes rq->lock is held */
2924 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2926 if (prev->sched_class->pre_schedule)
2927 prev->sched_class->pre_schedule(rq, prev);
2930 /* rq->lock is NOT held, but preemption is disabled */
2931 static inline void post_schedule(struct rq *rq)
2933 if (rq->post_schedule) {
2934 unsigned long flags;
2936 spin_lock_irqsave(&rq->lock, flags);
2937 if (rq->curr->sched_class->post_schedule)
2938 rq->curr->sched_class->post_schedule(rq);
2939 spin_unlock_irqrestore(&rq->lock, flags);
2941 rq->post_schedule = 0;
2947 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2951 static inline void post_schedule(struct rq *rq)
2958 * schedule_tail - first thing a freshly forked thread must call.
2959 * @prev: the thread we just switched away from.
2961 asmlinkage void schedule_tail(struct task_struct *prev)
2962 __releases(rq->lock)
2964 struct rq *rq = this_rq();
2966 finish_task_switch(rq, prev);
2969 * FIXME: do we need to worry about rq being invalidated by the
2974 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2975 /* In this case, finish_task_switch does not reenable preemption */
2978 if (current->set_child_tid)
2979 put_user(task_pid_vnr(current), current->set_child_tid);
2983 * context_switch - switch to the new MM and the new
2984 * thread's register state.
2987 context_switch(struct rq *rq, struct task_struct *prev,
2988 struct task_struct *next)
2990 struct mm_struct *mm, *oldmm;
2992 prepare_task_switch(rq, prev, next);
2993 trace_sched_switch(rq, prev, next);
2995 oldmm = prev->active_mm;
2997 * For paravirt, this is coupled with an exit in switch_to to
2998 * combine the page table reload and the switch backend into
3001 arch_start_context_switch(prev);
3003 if (unlikely(!mm)) {
3004 next->active_mm = oldmm;
3005 atomic_inc(&oldmm->mm_count);
3006 enter_lazy_tlb(oldmm, next);
3008 switch_mm(oldmm, mm, next);
3010 if (unlikely(!prev->mm)) {
3011 prev->active_mm = NULL;
3012 rq->prev_mm = oldmm;
3015 * Since the runqueue lock will be released by the next
3016 * task (which is an invalid locking op but in the case
3017 * of the scheduler it's an obvious special-case), so we
3018 * do an early lockdep release here:
3020 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3021 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3024 /* Here we just switch the register state and the stack. */
3025 switch_to(prev, next, prev);
3029 * this_rq must be evaluated again because prev may have moved
3030 * CPUs since it called schedule(), thus the 'rq' on its stack
3031 * frame will be invalid.
3033 finish_task_switch(this_rq(), prev);
3037 * nr_running, nr_uninterruptible and nr_context_switches:
3039 * externally visible scheduler statistics: current number of runnable
3040 * threads, current number of uninterruptible-sleeping threads, total
3041 * number of context switches performed since bootup.
3043 unsigned long nr_running(void)
3045 unsigned long i, sum = 0;
3047 for_each_online_cpu(i)
3048 sum += cpu_rq(i)->nr_running;
3053 unsigned long nr_uninterruptible(void)
3055 unsigned long i, sum = 0;
3057 for_each_possible_cpu(i)
3058 sum += cpu_rq(i)->nr_uninterruptible;
3061 * Since we read the counters lockless, it might be slightly
3062 * inaccurate. Do not allow it to go below zero though:
3064 if (unlikely((long)sum < 0))
3070 unsigned long long nr_context_switches(void)
3073 unsigned long long sum = 0;
3075 for_each_possible_cpu(i)
3076 sum += cpu_rq(i)->nr_switches;
3081 unsigned long nr_iowait(void)
3083 unsigned long i, sum = 0;
3085 for_each_possible_cpu(i)
3086 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3091 /* Variables and functions for calc_load */
3092 static atomic_long_t calc_load_tasks;
3093 static unsigned long calc_load_update;
3094 unsigned long avenrun[3];
3095 EXPORT_SYMBOL(avenrun);
3098 * get_avenrun - get the load average array
3099 * @loads: pointer to dest load array
3100 * @offset: offset to add
3101 * @shift: shift count to shift the result left
3103 * These values are estimates at best, so no need for locking.
3105 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3107 loads[0] = (avenrun[0] + offset) << shift;
3108 loads[1] = (avenrun[1] + offset) << shift;
3109 loads[2] = (avenrun[2] + offset) << shift;
3112 static unsigned long
3113 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3116 load += active * (FIXED_1 - exp);
3117 return load >> FSHIFT;
3121 * calc_load - update the avenrun load estimates 10 ticks after the
3122 * CPUs have updated calc_load_tasks.
3124 void calc_global_load(void)
3126 unsigned long upd = calc_load_update + 10;
3129 if (time_before(jiffies, upd))
3132 active = atomic_long_read(&calc_load_tasks);
3133 active = active > 0 ? active * FIXED_1 : 0;
3135 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3136 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3137 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3139 calc_load_update += LOAD_FREQ;
3143 * Either called from update_cpu_load() or from a cpu going idle
3145 static void calc_load_account_active(struct rq *this_rq)
3147 long nr_active, delta;
3149 nr_active = this_rq->nr_running;
3150 nr_active += (long) this_rq->nr_uninterruptible;
3152 if (nr_active != this_rq->calc_load_active) {
3153 delta = nr_active - this_rq->calc_load_active;
3154 this_rq->calc_load_active = nr_active;
3155 atomic_long_add(delta, &calc_load_tasks);
3160 * Externally visible per-cpu scheduler statistics:
3161 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3163 u64 cpu_nr_migrations(int cpu)
3165 return cpu_rq(cpu)->nr_migrations_in;
3169 * Update rq->cpu_load[] statistics. This function is usually called every
3170 * scheduler tick (TICK_NSEC).
3172 static void update_cpu_load(struct rq *this_rq)
3174 unsigned long this_load = this_rq->load.weight;
3177 this_rq->nr_load_updates++;
3179 /* Update our load: */
3180 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3181 unsigned long old_load, new_load;
3183 /* scale is effectively 1 << i now, and >> i divides by scale */
3185 old_load = this_rq->cpu_load[i];
3186 new_load = this_load;
3188 * Round up the averaging division if load is increasing. This
3189 * prevents us from getting stuck on 9 if the load is 10, for
3192 if (new_load > old_load)
3193 new_load += scale-1;
3194 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3197 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3198 this_rq->calc_load_update += LOAD_FREQ;
3199 calc_load_account_active(this_rq);
3206 * double_rq_lock - safely lock two runqueues
3208 * Note this does not disable interrupts like task_rq_lock,
3209 * you need to do so manually before calling.
3211 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3212 __acquires(rq1->lock)
3213 __acquires(rq2->lock)
3215 BUG_ON(!irqs_disabled());
3217 spin_lock(&rq1->lock);
3218 __acquire(rq2->lock); /* Fake it out ;) */
3221 spin_lock(&rq1->lock);
3222 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3224 spin_lock(&rq2->lock);
3225 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3228 update_rq_clock(rq1);
3229 update_rq_clock(rq2);
3233 * double_rq_unlock - safely unlock two runqueues
3235 * Note this does not restore interrupts like task_rq_unlock,
3236 * you need to do so manually after calling.
3238 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3239 __releases(rq1->lock)
3240 __releases(rq2->lock)
3242 spin_unlock(&rq1->lock);
3244 spin_unlock(&rq2->lock);
3246 __release(rq2->lock);
3250 * If dest_cpu is allowed for this process, migrate the task to it.
3251 * This is accomplished by forcing the cpu_allowed mask to only
3252 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3253 * the cpu_allowed mask is restored.
3255 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3257 struct migration_req req;
3258 unsigned long flags;
3261 rq = task_rq_lock(p, &flags);
3262 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3263 || unlikely(!cpu_active(dest_cpu)))
3266 /* force the process onto the specified CPU */
3267 if (migrate_task(p, dest_cpu, &req)) {
3268 /* Need to wait for migration thread (might exit: take ref). */
3269 struct task_struct *mt = rq->migration_thread;
3271 get_task_struct(mt);
3272 task_rq_unlock(rq, &flags);
3273 wake_up_process(mt);
3274 put_task_struct(mt);
3275 wait_for_completion(&req.done);
3280 task_rq_unlock(rq, &flags);
3284 * sched_exec - execve() is a valuable balancing opportunity, because at
3285 * this point the task has the smallest effective memory and cache footprint.
3287 void sched_exec(void)
3289 int new_cpu, this_cpu = get_cpu();
3290 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3292 if (new_cpu != this_cpu)
3293 sched_migrate_task(current, new_cpu);
3297 * pull_task - move a task from a remote runqueue to the local runqueue.
3298 * Both runqueues must be locked.
3300 static void pull_task(struct rq *src_rq, struct task_struct *p,
3301 struct rq *this_rq, int this_cpu)
3303 deactivate_task(src_rq, p, 0);
3304 set_task_cpu(p, this_cpu);
3305 activate_task(this_rq, p, 0);
3307 * Note that idle threads have a prio of MAX_PRIO, for this test
3308 * to be always true for them.
3310 check_preempt_curr(this_rq, p, 0);
3314 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3317 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3318 struct sched_domain *sd, enum cpu_idle_type idle,
3321 int tsk_cache_hot = 0;
3323 * We do not migrate tasks that are:
3324 * 1) running (obviously), or
3325 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3326 * 3) are cache-hot on their current CPU.
3328 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3329 schedstat_inc(p, se.nr_failed_migrations_affine);
3334 if (task_running(rq, p)) {
3335 schedstat_inc(p, se.nr_failed_migrations_running);
3340 * Aggressive migration if:
3341 * 1) task is cache cold, or
3342 * 2) too many balance attempts have failed.
3345 tsk_cache_hot = task_hot(p, rq->clock, sd);
3346 if (!tsk_cache_hot ||
3347 sd->nr_balance_failed > sd->cache_nice_tries) {
3348 #ifdef CONFIG_SCHEDSTATS
3349 if (tsk_cache_hot) {
3350 schedstat_inc(sd, lb_hot_gained[idle]);
3351 schedstat_inc(p, se.nr_forced_migrations);
3357 if (tsk_cache_hot) {
3358 schedstat_inc(p, se.nr_failed_migrations_hot);
3364 static unsigned long
3365 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3366 unsigned long max_load_move, struct sched_domain *sd,
3367 enum cpu_idle_type idle, int *all_pinned,
3368 int *this_best_prio, struct rq_iterator *iterator)
3370 int loops = 0, pulled = 0, pinned = 0;
3371 struct task_struct *p;
3372 long rem_load_move = max_load_move;
3374 if (max_load_move == 0)
3380 * Start the load-balancing iterator:
3382 p = iterator->start(iterator->arg);
3384 if (!p || loops++ > sysctl_sched_nr_migrate)
3387 if ((p->se.load.weight >> 1) > rem_load_move ||
3388 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3389 p = iterator->next(iterator->arg);
3393 pull_task(busiest, p, this_rq, this_cpu);
3395 rem_load_move -= p->se.load.weight;
3397 #ifdef CONFIG_PREEMPT
3399 * NEWIDLE balancing is a source of latency, so preemptible kernels
3400 * will stop after the first task is pulled to minimize the critical
3403 if (idle == CPU_NEWLY_IDLE)
3408 * We only want to steal up to the prescribed amount of weighted load.
3410 if (rem_load_move > 0) {
3411 if (p->prio < *this_best_prio)
3412 *this_best_prio = p->prio;
3413 p = iterator->next(iterator->arg);
3418 * Right now, this is one of only two places pull_task() is called,
3419 * so we can safely collect pull_task() stats here rather than
3420 * inside pull_task().
3422 schedstat_add(sd, lb_gained[idle], pulled);
3425 *all_pinned = pinned;
3427 return max_load_move - rem_load_move;
3431 * move_tasks tries to move up to max_load_move weighted load from busiest to
3432 * this_rq, as part of a balancing operation within domain "sd".
3433 * Returns 1 if successful and 0 otherwise.
3435 * Called with both runqueues locked.
3437 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3438 unsigned long max_load_move,
3439 struct sched_domain *sd, enum cpu_idle_type idle,
3442 const struct sched_class *class = sched_class_highest;
3443 unsigned long total_load_moved = 0;
3444 int this_best_prio = this_rq->curr->prio;
3448 class->load_balance(this_rq, this_cpu, busiest,
3449 max_load_move - total_load_moved,
3450 sd, idle, all_pinned, &this_best_prio);
3451 class = class->next;
3453 #ifdef CONFIG_PREEMPT
3455 * NEWIDLE balancing is a source of latency, so preemptible
3456 * kernels will stop after the first task is pulled to minimize
3457 * the critical section.
3459 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3462 } while (class && max_load_move > total_load_moved);
3464 return total_load_moved > 0;
3468 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3469 struct sched_domain *sd, enum cpu_idle_type idle,
3470 struct rq_iterator *iterator)
3472 struct task_struct *p = iterator->start(iterator->arg);
3476 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3477 pull_task(busiest, p, this_rq, this_cpu);
3479 * Right now, this is only the second place pull_task()
3480 * is called, so we can safely collect pull_task()
3481 * stats here rather than inside pull_task().
3483 schedstat_inc(sd, lb_gained[idle]);
3487 p = iterator->next(iterator->arg);
3494 * move_one_task tries to move exactly one task from busiest to this_rq, as
3495 * part of active balancing operations within "domain".
3496 * Returns 1 if successful and 0 otherwise.
3498 * Called with both runqueues locked.
3500 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3501 struct sched_domain *sd, enum cpu_idle_type idle)
3503 const struct sched_class *class;
3505 for_each_class(class) {
3506 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3512 /********** Helpers for find_busiest_group ************************/
3514 * sd_lb_stats - Structure to store the statistics of a sched_domain
3515 * during load balancing.
3517 struct sd_lb_stats {
3518 struct sched_group *busiest; /* Busiest group in this sd */
3519 struct sched_group *this; /* Local group in this sd */
3520 unsigned long total_load; /* Total load of all groups in sd */
3521 unsigned long total_pwr; /* Total power of all groups in sd */
3522 unsigned long avg_load; /* Average load across all groups in sd */
3524 /** Statistics of this group */
3525 unsigned long this_load;
3526 unsigned long this_load_per_task;
3527 unsigned long this_nr_running;
3529 /* Statistics of the busiest group */
3530 unsigned long max_load;
3531 unsigned long busiest_load_per_task;
3532 unsigned long busiest_nr_running;
3534 int group_imb; /* Is there imbalance in this sd */
3535 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3536 int power_savings_balance; /* Is powersave balance needed for this sd */
3537 struct sched_group *group_min; /* Least loaded group in sd */
3538 struct sched_group *group_leader; /* Group which relieves group_min */
3539 unsigned long min_load_per_task; /* load_per_task in group_min */
3540 unsigned long leader_nr_running; /* Nr running of group_leader */
3541 unsigned long min_nr_running; /* Nr running of group_min */
3546 * sg_lb_stats - stats of a sched_group required for load_balancing
3548 struct sg_lb_stats {
3549 unsigned long avg_load; /*Avg load across the CPUs of the group */
3550 unsigned long group_load; /* Total load over the CPUs of the group */
3551 unsigned long sum_nr_running; /* Nr tasks running in the group */
3552 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3553 unsigned long group_capacity;
3554 int group_imb; /* Is there an imbalance in the group ? */
3558 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3559 * @group: The group whose first cpu is to be returned.
3561 static inline unsigned int group_first_cpu(struct sched_group *group)
3563 return cpumask_first(sched_group_cpus(group));
3567 * get_sd_load_idx - Obtain the load index for a given sched domain.
3568 * @sd: The sched_domain whose load_idx is to be obtained.
3569 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3571 static inline int get_sd_load_idx(struct sched_domain *sd,
3572 enum cpu_idle_type idle)
3578 load_idx = sd->busy_idx;
3581 case CPU_NEWLY_IDLE:
3582 load_idx = sd->newidle_idx;
3585 load_idx = sd->idle_idx;
3593 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3595 * init_sd_power_savings_stats - Initialize power savings statistics for
3596 * the given sched_domain, during load balancing.
3598 * @sd: Sched domain whose power-savings statistics are to be initialized.
3599 * @sds: Variable containing the statistics for sd.
3600 * @idle: Idle status of the CPU at which we're performing load-balancing.
3602 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3603 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3606 * Busy processors will not participate in power savings
3609 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3610 sds->power_savings_balance = 0;
3612 sds->power_savings_balance = 1;
3613 sds->min_nr_running = ULONG_MAX;
3614 sds->leader_nr_running = 0;
3619 * update_sd_power_savings_stats - Update the power saving stats for a
3620 * sched_domain while performing load balancing.
3622 * @group: sched_group belonging to the sched_domain under consideration.
3623 * @sds: Variable containing the statistics of the sched_domain
3624 * @local_group: Does group contain the CPU for which we're performing
3626 * @sgs: Variable containing the statistics of the group.
3628 static inline void update_sd_power_savings_stats(struct sched_group *group,
3629 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3632 if (!sds->power_savings_balance)
3636 * If the local group is idle or completely loaded
3637 * no need to do power savings balance at this domain
3639 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3640 !sds->this_nr_running))
3641 sds->power_savings_balance = 0;
3644 * If a group is already running at full capacity or idle,
3645 * don't include that group in power savings calculations
3647 if (!sds->power_savings_balance ||
3648 sgs->sum_nr_running >= sgs->group_capacity ||
3649 !sgs->sum_nr_running)
3653 * Calculate the group which has the least non-idle load.
3654 * This is the group from where we need to pick up the load
3657 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3658 (sgs->sum_nr_running == sds->min_nr_running &&
3659 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3660 sds->group_min = group;
3661 sds->min_nr_running = sgs->sum_nr_running;
3662 sds->min_load_per_task = sgs->sum_weighted_load /
3663 sgs->sum_nr_running;
3667 * Calculate the group which is almost near its
3668 * capacity but still has some space to pick up some load
3669 * from other group and save more power
3671 if (sgs->sum_nr_running > sgs->group_capacity - 1)
3674 if (sgs->sum_nr_running > sds->leader_nr_running ||
3675 (sgs->sum_nr_running == sds->leader_nr_running &&
3676 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3677 sds->group_leader = group;
3678 sds->leader_nr_running = sgs->sum_nr_running;
3683 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3684 * @sds: Variable containing the statistics of the sched_domain
3685 * under consideration.
3686 * @this_cpu: Cpu at which we're currently performing load-balancing.
3687 * @imbalance: Variable to store the imbalance.
3690 * Check if we have potential to perform some power-savings balance.
3691 * If yes, set the busiest group to be the least loaded group in the
3692 * sched_domain, so that it's CPUs can be put to idle.
3694 * Returns 1 if there is potential to perform power-savings balance.
3697 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3698 int this_cpu, unsigned long *imbalance)
3700 if (!sds->power_savings_balance)
3703 if (sds->this != sds->group_leader ||
3704 sds->group_leader == sds->group_min)
3707 *imbalance = sds->min_load_per_task;
3708 sds->busiest = sds->group_min;
3710 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3711 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3712 group_first_cpu(sds->group_leader);
3718 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3719 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3720 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3725 static inline void update_sd_power_savings_stats(struct sched_group *group,
3726 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3731 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3732 int this_cpu, unsigned long *imbalance)
3736 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3738 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3740 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3741 unsigned long smt_gain = sd->smt_gain;
3748 unsigned long scale_rt_power(int cpu)
3750 struct rq *rq = cpu_rq(cpu);
3751 u64 total, available;
3753 sched_avg_update(rq);
3755 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3756 available = total - rq->rt_avg;
3758 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3759 total = SCHED_LOAD_SCALE;
3761 total >>= SCHED_LOAD_SHIFT;
3763 return div_u64(available, total);
3766 static void update_cpu_power(struct sched_domain *sd, int cpu)
3768 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3769 unsigned long power = SCHED_LOAD_SCALE;
3770 struct sched_group *sdg = sd->groups;
3771 unsigned long old = sdg->__cpu_power;
3773 /* here we could scale based on cpufreq */
3775 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3776 power *= arch_scale_smt_power(sd, cpu);
3777 power >>= SCHED_LOAD_SHIFT;
3780 power *= scale_rt_power(cpu);
3781 power >>= SCHED_LOAD_SHIFT;
3787 sdg->__cpu_power = power;
3788 sdg->reciprocal_cpu_power = reciprocal_value(power);
3792 static void update_group_power(struct sched_domain *sd, int cpu)
3794 struct sched_domain *child = sd->child;
3795 struct sched_group *group, *sdg = sd->groups;
3796 unsigned long power = sdg->__cpu_power;
3799 update_cpu_power(sd, cpu);
3803 sdg->__cpu_power = 0;
3805 group = child->groups;
3807 sdg->__cpu_power += group->__cpu_power;
3808 group = group->next;
3809 } while (group != child->groups);
3811 if (power != sdg->__cpu_power)
3812 sdg->reciprocal_cpu_power = reciprocal_value(sdg->__cpu_power);
3816 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3817 * @group: sched_group whose statistics are to be updated.
3818 * @this_cpu: Cpu for which load balance is currently performed.
3819 * @idle: Idle status of this_cpu
3820 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3821 * @sd_idle: Idle status of the sched_domain containing group.
3822 * @local_group: Does group contain this_cpu.
3823 * @cpus: Set of cpus considered for load balancing.
3824 * @balance: Should we balance.
3825 * @sgs: variable to hold the statistics for this group.
3827 static inline void update_sg_lb_stats(struct sched_domain *sd,
3828 struct sched_group *group, int this_cpu,
3829 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3830 int local_group, const struct cpumask *cpus,
3831 int *balance, struct sg_lb_stats *sgs)
3833 unsigned long load, max_cpu_load, min_cpu_load;
3835 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3836 unsigned long sum_avg_load_per_task;
3837 unsigned long avg_load_per_task;
3840 balance_cpu = group_first_cpu(group);
3841 if (balance_cpu == this_cpu)
3842 update_group_power(sd, this_cpu);
3845 /* Tally up the load of all CPUs in the group */
3846 sum_avg_load_per_task = avg_load_per_task = 0;
3848 min_cpu_load = ~0UL;
3850 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3851 struct rq *rq = cpu_rq(i);
3853 if (*sd_idle && rq->nr_running)
3856 /* Bias balancing toward cpus of our domain */
3858 if (idle_cpu(i) && !first_idle_cpu) {
3863 load = target_load(i, load_idx);
3865 load = source_load(i, load_idx);
3866 if (load > max_cpu_load)
3867 max_cpu_load = load;
3868 if (min_cpu_load > load)
3869 min_cpu_load = load;
3872 sgs->group_load += load;
3873 sgs->sum_nr_running += rq->nr_running;
3874 sgs->sum_weighted_load += weighted_cpuload(i);
3876 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3880 * First idle cpu or the first cpu(busiest) in this sched group
3881 * is eligible for doing load balancing at this and above
3882 * domains. In the newly idle case, we will allow all the cpu's
3883 * to do the newly idle load balance.
3885 if (idle != CPU_NEWLY_IDLE && local_group &&
3886 balance_cpu != this_cpu && balance) {
3891 /* Adjust by relative CPU power of the group */
3892 sgs->avg_load = sg_div_cpu_power(group,
3893 sgs->group_load * SCHED_LOAD_SCALE);
3897 * Consider the group unbalanced when the imbalance is larger
3898 * than the average weight of two tasks.
3900 * APZ: with cgroup the avg task weight can vary wildly and
3901 * might not be a suitable number - should we keep a
3902 * normalized nr_running number somewhere that negates
3905 avg_load_per_task = sg_div_cpu_power(group,
3906 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3908 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3911 sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3916 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3917 * @sd: sched_domain whose statistics are to be updated.
3918 * @this_cpu: Cpu for which load balance is currently performed.
3919 * @idle: Idle status of this_cpu
3920 * @sd_idle: Idle status of the sched_domain containing group.
3921 * @cpus: Set of cpus considered for load balancing.
3922 * @balance: Should we balance.
3923 * @sds: variable to hold the statistics for this sched_domain.
3925 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3926 enum cpu_idle_type idle, int *sd_idle,
3927 const struct cpumask *cpus, int *balance,
3928 struct sd_lb_stats *sds)
3930 struct sched_domain *child = sd->child;
3931 struct sched_group *group = sd->groups;
3932 struct sg_lb_stats sgs;
3933 int load_idx, prefer_sibling = 0;
3935 if (child && child->flags & SD_PREFER_SIBLING)
3938 init_sd_power_savings_stats(sd, sds, idle);
3939 load_idx = get_sd_load_idx(sd, idle);
3944 local_group = cpumask_test_cpu(this_cpu,
3945 sched_group_cpus(group));
3946 memset(&sgs, 0, sizeof(sgs));
3947 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3948 local_group, cpus, balance, &sgs);
3950 if (local_group && balance && !(*balance))
3953 sds->total_load += sgs.group_load;
3954 sds->total_pwr += group->__cpu_power;
3957 * In case the child domain prefers tasks go to siblings
3958 * first, lower the group capacity to one so that we'll try
3959 * and move all the excess tasks away.
3962 sgs.group_capacity = 1;
3965 sds->this_load = sgs.avg_load;
3967 sds->this_nr_running = sgs.sum_nr_running;
3968 sds->this_load_per_task = sgs.sum_weighted_load;
3969 } else if (sgs.avg_load > sds->max_load &&
3970 (sgs.sum_nr_running > sgs.group_capacity ||
3972 sds->max_load = sgs.avg_load;
3973 sds->busiest = group;
3974 sds->busiest_nr_running = sgs.sum_nr_running;
3975 sds->busiest_load_per_task = sgs.sum_weighted_load;
3976 sds->group_imb = sgs.group_imb;
3979 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3980 group = group->next;
3981 } while (group != sd->groups);
3985 * fix_small_imbalance - Calculate the minor imbalance that exists
3986 * amongst the groups of a sched_domain, during
3988 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3989 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3990 * @imbalance: Variable to store the imbalance.
3992 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3993 int this_cpu, unsigned long *imbalance)
3995 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3996 unsigned int imbn = 2;
3998 if (sds->this_nr_running) {
3999 sds->this_load_per_task /= sds->this_nr_running;
4000 if (sds->busiest_load_per_task >
4001 sds->this_load_per_task)
4004 sds->this_load_per_task =
4005 cpu_avg_load_per_task(this_cpu);
4007 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
4008 sds->busiest_load_per_task * imbn) {
4009 *imbalance = sds->busiest_load_per_task;
4014 * OK, we don't have enough imbalance to justify moving tasks,
4015 * however we may be able to increase total CPU power used by
4019 pwr_now += sds->busiest->__cpu_power *
4020 min(sds->busiest_load_per_task, sds->max_load);
4021 pwr_now += sds->this->__cpu_power *
4022 min(sds->this_load_per_task, sds->this_load);
4023 pwr_now /= SCHED_LOAD_SCALE;
4025 /* Amount of load we'd subtract */
4026 tmp = sg_div_cpu_power(sds->busiest,
4027 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
4028 if (sds->max_load > tmp)
4029 pwr_move += sds->busiest->__cpu_power *
4030 min(sds->busiest_load_per_task, sds->max_load - tmp);
4032 /* Amount of load we'd add */
4033 if (sds->max_load * sds->busiest->__cpu_power <
4034 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
4035 tmp = sg_div_cpu_power(sds->this,
4036 sds->max_load * sds->busiest->__cpu_power);
4038 tmp = sg_div_cpu_power(sds->this,
4039 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
4040 pwr_move += sds->this->__cpu_power *
4041 min(sds->this_load_per_task, sds->this_load + tmp);
4042 pwr_move /= SCHED_LOAD_SCALE;
4044 /* Move if we gain throughput */
4045 if (pwr_move > pwr_now)
4046 *imbalance = sds->busiest_load_per_task;
4050 * calculate_imbalance - Calculate the amount of imbalance present within the
4051 * groups of a given sched_domain during load balance.
4052 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4053 * @this_cpu: Cpu for which currently load balance is being performed.
4054 * @imbalance: The variable to store the imbalance.
4056 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
4057 unsigned long *imbalance)
4059 unsigned long max_pull;
4061 * In the presence of smp nice balancing, certain scenarios can have
4062 * max load less than avg load(as we skip the groups at or below
4063 * its cpu_power, while calculating max_load..)
4065 if (sds->max_load < sds->avg_load) {
4067 return fix_small_imbalance(sds, this_cpu, imbalance);
4070 /* Don't want to pull so many tasks that a group would go idle */
4071 max_pull = min(sds->max_load - sds->avg_load,
4072 sds->max_load - sds->busiest_load_per_task);
4074 /* How much load to actually move to equalise the imbalance */
4075 *imbalance = min(max_pull * sds->busiest->__cpu_power,
4076 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
4080 * if *imbalance is less than the average load per runnable task
4081 * there is no gaurantee that any tasks will be moved so we'll have
4082 * a think about bumping its value to force at least one task to be
4085 if (*imbalance < sds->busiest_load_per_task)
4086 return fix_small_imbalance(sds, this_cpu, imbalance);
4089 /******* find_busiest_group() helpers end here *********************/
4092 * find_busiest_group - Returns the busiest group within the sched_domain
4093 * if there is an imbalance. If there isn't an imbalance, and
4094 * the user has opted for power-savings, it returns a group whose
4095 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4096 * such a group exists.
4098 * Also calculates the amount of weighted load which should be moved
4099 * to restore balance.
4101 * @sd: The sched_domain whose busiest group is to be returned.
4102 * @this_cpu: The cpu for which load balancing is currently being performed.
4103 * @imbalance: Variable which stores amount of weighted load which should
4104 * be moved to restore balance/put a group to idle.
4105 * @idle: The idle status of this_cpu.
4106 * @sd_idle: The idleness of sd
4107 * @cpus: The set of CPUs under consideration for load-balancing.
4108 * @balance: Pointer to a variable indicating if this_cpu
4109 * is the appropriate cpu to perform load balancing at this_level.
4111 * Returns: - the busiest group if imbalance exists.
4112 * - If no imbalance and user has opted for power-savings balance,
4113 * return the least loaded group whose CPUs can be
4114 * put to idle by rebalancing its tasks onto our group.
4116 static struct sched_group *
4117 find_busiest_group(struct sched_domain *sd, int this_cpu,
4118 unsigned long *imbalance, enum cpu_idle_type idle,
4119 int *sd_idle, const struct cpumask *cpus, int *balance)
4121 struct sd_lb_stats sds;
4123 memset(&sds, 0, sizeof(sds));
4126 * Compute the various statistics relavent for load balancing at
4129 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4132 /* Cases where imbalance does not exist from POV of this_cpu */
4133 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4135 * 2) There is no busy sibling group to pull from.
4136 * 3) This group is the busiest group.
4137 * 4) This group is more busy than the avg busieness at this
4139 * 5) The imbalance is within the specified limit.
4140 * 6) Any rebalance would lead to ping-pong
4142 if (balance && !(*balance))
4145 if (!sds.busiest || sds.busiest_nr_running == 0)
4148 if (sds.this_load >= sds.max_load)
4151 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4153 if (sds.this_load >= sds.avg_load)
4156 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4159 sds.busiest_load_per_task /= sds.busiest_nr_running;
4161 sds.busiest_load_per_task =
4162 min(sds.busiest_load_per_task, sds.avg_load);
4165 * We're trying to get all the cpus to the average_load, so we don't
4166 * want to push ourselves above the average load, nor do we wish to
4167 * reduce the max loaded cpu below the average load, as either of these
4168 * actions would just result in more rebalancing later, and ping-pong
4169 * tasks around. Thus we look for the minimum possible imbalance.
4170 * Negative imbalances (*we* are more loaded than anyone else) will
4171 * be counted as no imbalance for these purposes -- we can't fix that
4172 * by pulling tasks to us. Be careful of negative numbers as they'll
4173 * appear as very large values with unsigned longs.
4175 if (sds.max_load <= sds.busiest_load_per_task)
4178 /* Looks like there is an imbalance. Compute it */
4179 calculate_imbalance(&sds, this_cpu, imbalance);
4184 * There is no obvious imbalance. But check if we can do some balancing
4187 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4195 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4198 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4199 unsigned long imbalance, const struct cpumask *cpus)
4201 struct rq *busiest = NULL, *rq;
4202 unsigned long max_load = 0;
4205 for_each_cpu(i, sched_group_cpus(group)) {
4208 if (!cpumask_test_cpu(i, cpus))
4212 wl = weighted_cpuload(i);
4214 if (rq->nr_running == 1 && wl > imbalance)
4217 if (wl > max_load) {
4227 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4228 * so long as it is large enough.
4230 #define MAX_PINNED_INTERVAL 512
4232 /* Working cpumask for load_balance and load_balance_newidle. */
4233 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4236 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4237 * tasks if there is an imbalance.
4239 static int load_balance(int this_cpu, struct rq *this_rq,
4240 struct sched_domain *sd, enum cpu_idle_type idle,
4243 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4244 struct sched_group *group;
4245 unsigned long imbalance;
4247 unsigned long flags;
4248 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4250 cpumask_setall(cpus);
4253 * When power savings policy is enabled for the parent domain, idle
4254 * sibling can pick up load irrespective of busy siblings. In this case,
4255 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4256 * portraying it as CPU_NOT_IDLE.
4258 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4259 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4262 schedstat_inc(sd, lb_count[idle]);
4266 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4273 schedstat_inc(sd, lb_nobusyg[idle]);
4277 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4279 schedstat_inc(sd, lb_nobusyq[idle]);
4283 BUG_ON(busiest == this_rq);
4285 schedstat_add(sd, lb_imbalance[idle], imbalance);
4288 if (busiest->nr_running > 1) {
4290 * Attempt to move tasks. If find_busiest_group has found
4291 * an imbalance but busiest->nr_running <= 1, the group is
4292 * still unbalanced. ld_moved simply stays zero, so it is
4293 * correctly treated as an imbalance.
4295 local_irq_save(flags);
4296 double_rq_lock(this_rq, busiest);
4297 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4298 imbalance, sd, idle, &all_pinned);
4299 double_rq_unlock(this_rq, busiest);
4300 local_irq_restore(flags);
4303 * some other cpu did the load balance for us.
4305 if (ld_moved && this_cpu != smp_processor_id())
4306 resched_cpu(this_cpu);
4308 /* All tasks on this runqueue were pinned by CPU affinity */
4309 if (unlikely(all_pinned)) {
4310 cpumask_clear_cpu(cpu_of(busiest), cpus);
4311 if (!cpumask_empty(cpus))
4318 schedstat_inc(sd, lb_failed[idle]);
4319 sd->nr_balance_failed++;
4321 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4323 spin_lock_irqsave(&busiest->lock, flags);
4325 /* don't kick the migration_thread, if the curr
4326 * task on busiest cpu can't be moved to this_cpu
4328 if (!cpumask_test_cpu(this_cpu,
4329 &busiest->curr->cpus_allowed)) {
4330 spin_unlock_irqrestore(&busiest->lock, flags);
4332 goto out_one_pinned;
4335 if (!busiest->active_balance) {
4336 busiest->active_balance = 1;
4337 busiest->push_cpu = this_cpu;
4340 spin_unlock_irqrestore(&busiest->lock, flags);
4342 wake_up_process(busiest->migration_thread);
4345 * We've kicked active balancing, reset the failure
4348 sd->nr_balance_failed = sd->cache_nice_tries+1;
4351 sd->nr_balance_failed = 0;
4353 if (likely(!active_balance)) {
4354 /* We were unbalanced, so reset the balancing interval */
4355 sd->balance_interval = sd->min_interval;
4358 * If we've begun active balancing, start to back off. This
4359 * case may not be covered by the all_pinned logic if there
4360 * is only 1 task on the busy runqueue (because we don't call
4363 if (sd->balance_interval < sd->max_interval)
4364 sd->balance_interval *= 2;
4367 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4368 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4374 schedstat_inc(sd, lb_balanced[idle]);
4376 sd->nr_balance_failed = 0;
4379 /* tune up the balancing interval */
4380 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4381 (sd->balance_interval < sd->max_interval))
4382 sd->balance_interval *= 2;
4384 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4385 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4396 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4397 * tasks if there is an imbalance.
4399 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4400 * this_rq is locked.
4403 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4405 struct sched_group *group;
4406 struct rq *busiest = NULL;
4407 unsigned long imbalance;
4411 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4413 cpumask_setall(cpus);
4416 * When power savings policy is enabled for the parent domain, idle
4417 * sibling can pick up load irrespective of busy siblings. In this case,
4418 * let the state of idle sibling percolate up as IDLE, instead of
4419 * portraying it as CPU_NOT_IDLE.
4421 if (sd->flags & SD_SHARE_CPUPOWER &&
4422 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4425 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4427 update_shares_locked(this_rq, sd);
4428 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4429 &sd_idle, cpus, NULL);
4431 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4435 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4437 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4441 BUG_ON(busiest == this_rq);
4443 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4446 if (busiest->nr_running > 1) {
4447 /* Attempt to move tasks */
4448 double_lock_balance(this_rq, busiest);
4449 /* this_rq->clock is already updated */
4450 update_rq_clock(busiest);
4451 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4452 imbalance, sd, CPU_NEWLY_IDLE,
4454 double_unlock_balance(this_rq, busiest);
4456 if (unlikely(all_pinned)) {
4457 cpumask_clear_cpu(cpu_of(busiest), cpus);
4458 if (!cpumask_empty(cpus))
4464 int active_balance = 0;
4466 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4467 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4468 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4471 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4474 if (sd->nr_balance_failed++ < 2)
4478 * The only task running in a non-idle cpu can be moved to this
4479 * cpu in an attempt to completely freeup the other CPU
4480 * package. The same method used to move task in load_balance()
4481 * have been extended for load_balance_newidle() to speedup
4482 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4484 * The package power saving logic comes from
4485 * find_busiest_group(). If there are no imbalance, then
4486 * f_b_g() will return NULL. However when sched_mc={1,2} then
4487 * f_b_g() will select a group from which a running task may be
4488 * pulled to this cpu in order to make the other package idle.
4489 * If there is no opportunity to make a package idle and if
4490 * there are no imbalance, then f_b_g() will return NULL and no
4491 * action will be taken in load_balance_newidle().
4493 * Under normal task pull operation due to imbalance, there
4494 * will be more than one task in the source run queue and
4495 * move_tasks() will succeed. ld_moved will be true and this
4496 * active balance code will not be triggered.
4499 /* Lock busiest in correct order while this_rq is held */
4500 double_lock_balance(this_rq, busiest);
4503 * don't kick the migration_thread, if the curr
4504 * task on busiest cpu can't be moved to this_cpu
4506 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4507 double_unlock_balance(this_rq, busiest);
4512 if (!busiest->active_balance) {
4513 busiest->active_balance = 1;
4514 busiest->push_cpu = this_cpu;
4518 double_unlock_balance(this_rq, busiest);
4520 * Should not call ttwu while holding a rq->lock
4522 spin_unlock(&this_rq->lock);
4524 wake_up_process(busiest->migration_thread);
4525 spin_lock(&this_rq->lock);
4528 sd->nr_balance_failed = 0;
4530 update_shares_locked(this_rq, sd);
4534 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4535 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4536 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4538 sd->nr_balance_failed = 0;
4544 * idle_balance is called by schedule() if this_cpu is about to become
4545 * idle. Attempts to pull tasks from other CPUs.
4547 static void idle_balance(int this_cpu, struct rq *this_rq)
4549 struct sched_domain *sd;
4550 int pulled_task = 0;
4551 unsigned long next_balance = jiffies + HZ;
4553 for_each_domain(this_cpu, sd) {
4554 unsigned long interval;
4556 if (!(sd->flags & SD_LOAD_BALANCE))
4559 if (sd->flags & SD_BALANCE_NEWIDLE)
4560 /* If we've pulled tasks over stop searching: */
4561 pulled_task = load_balance_newidle(this_cpu, this_rq,
4564 interval = msecs_to_jiffies(sd->balance_interval);
4565 if (time_after(next_balance, sd->last_balance + interval))
4566 next_balance = sd->last_balance + interval;
4570 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4572 * We are going idle. next_balance may be set based on
4573 * a busy processor. So reset next_balance.
4575 this_rq->next_balance = next_balance;
4580 * active_load_balance is run by migration threads. It pushes running tasks
4581 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4582 * running on each physical CPU where possible, and avoids physical /
4583 * logical imbalances.
4585 * Called with busiest_rq locked.
4587 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4589 int target_cpu = busiest_rq->push_cpu;
4590 struct sched_domain *sd;
4591 struct rq *target_rq;
4593 /* Is there any task to move? */
4594 if (busiest_rq->nr_running <= 1)
4597 target_rq = cpu_rq(target_cpu);
4600 * This condition is "impossible", if it occurs
4601 * we need to fix it. Originally reported by
4602 * Bjorn Helgaas on a 128-cpu setup.
4604 BUG_ON(busiest_rq == target_rq);
4606 /* move a task from busiest_rq to target_rq */
4607 double_lock_balance(busiest_rq, target_rq);
4608 update_rq_clock(busiest_rq);
4609 update_rq_clock(target_rq);
4611 /* Search for an sd spanning us and the target CPU. */
4612 for_each_domain(target_cpu, sd) {
4613 if ((sd->flags & SD_LOAD_BALANCE) &&
4614 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4619 schedstat_inc(sd, alb_count);
4621 if (move_one_task(target_rq, target_cpu, busiest_rq,
4623 schedstat_inc(sd, alb_pushed);
4625 schedstat_inc(sd, alb_failed);
4627 double_unlock_balance(busiest_rq, target_rq);
4632 atomic_t load_balancer;
4633 cpumask_var_t cpu_mask;
4634 cpumask_var_t ilb_grp_nohz_mask;
4635 } nohz ____cacheline_aligned = {
4636 .load_balancer = ATOMIC_INIT(-1),
4639 int get_nohz_load_balancer(void)
4641 return atomic_read(&nohz.load_balancer);
4644 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4646 * lowest_flag_domain - Return lowest sched_domain containing flag.
4647 * @cpu: The cpu whose lowest level of sched domain is to
4649 * @flag: The flag to check for the lowest sched_domain
4650 * for the given cpu.
4652 * Returns the lowest sched_domain of a cpu which contains the given flag.
4654 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4656 struct sched_domain *sd;
4658 for_each_domain(cpu, sd)
4659 if (sd && (sd->flags & flag))
4666 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4667 * @cpu: The cpu whose domains we're iterating over.
4668 * @sd: variable holding the value of the power_savings_sd
4670 * @flag: The flag to filter the sched_domains to be iterated.
4672 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4673 * set, starting from the lowest sched_domain to the highest.
4675 #define for_each_flag_domain(cpu, sd, flag) \
4676 for (sd = lowest_flag_domain(cpu, flag); \
4677 (sd && (sd->flags & flag)); sd = sd->parent)
4680 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4681 * @ilb_group: group to be checked for semi-idleness
4683 * Returns: 1 if the group is semi-idle. 0 otherwise.
4685 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4686 * and atleast one non-idle CPU. This helper function checks if the given
4687 * sched_group is semi-idle or not.
4689 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4691 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4692 sched_group_cpus(ilb_group));
4695 * A sched_group is semi-idle when it has atleast one busy cpu
4696 * and atleast one idle cpu.
4698 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4701 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4707 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4708 * @cpu: The cpu which is nominating a new idle_load_balancer.
4710 * Returns: Returns the id of the idle load balancer if it exists,
4711 * Else, returns >= nr_cpu_ids.
4713 * This algorithm picks the idle load balancer such that it belongs to a
4714 * semi-idle powersavings sched_domain. The idea is to try and avoid
4715 * completely idle packages/cores just for the purpose of idle load balancing
4716 * when there are other idle cpu's which are better suited for that job.
4718 static int find_new_ilb(int cpu)
4720 struct sched_domain *sd;
4721 struct sched_group *ilb_group;
4724 * Have idle load balancer selection from semi-idle packages only
4725 * when power-aware load balancing is enabled
4727 if (!(sched_smt_power_savings || sched_mc_power_savings))
4731 * Optimize for the case when we have no idle CPUs or only one
4732 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4734 if (cpumask_weight(nohz.cpu_mask) < 2)
4737 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4738 ilb_group = sd->groups;
4741 if (is_semi_idle_group(ilb_group))
4742 return cpumask_first(nohz.ilb_grp_nohz_mask);
4744 ilb_group = ilb_group->next;
4746 } while (ilb_group != sd->groups);
4750 return cpumask_first(nohz.cpu_mask);
4752 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4753 static inline int find_new_ilb(int call_cpu)
4755 return cpumask_first(nohz.cpu_mask);
4760 * This routine will try to nominate the ilb (idle load balancing)
4761 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4762 * load balancing on behalf of all those cpus. If all the cpus in the system
4763 * go into this tickless mode, then there will be no ilb owner (as there is
4764 * no need for one) and all the cpus will sleep till the next wakeup event
4767 * For the ilb owner, tick is not stopped. And this tick will be used
4768 * for idle load balancing. ilb owner will still be part of
4771 * While stopping the tick, this cpu will become the ilb owner if there
4772 * is no other owner. And will be the owner till that cpu becomes busy
4773 * or if all cpus in the system stop their ticks at which point
4774 * there is no need for ilb owner.
4776 * When the ilb owner becomes busy, it nominates another owner, during the
4777 * next busy scheduler_tick()
4779 int select_nohz_load_balancer(int stop_tick)
4781 int cpu = smp_processor_id();
4784 cpu_rq(cpu)->in_nohz_recently = 1;
4786 if (!cpu_active(cpu)) {
4787 if (atomic_read(&nohz.load_balancer) != cpu)
4791 * If we are going offline and still the leader,
4794 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4800 cpumask_set_cpu(cpu, nohz.cpu_mask);
4802 /* time for ilb owner also to sleep */
4803 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4804 if (atomic_read(&nohz.load_balancer) == cpu)
4805 atomic_set(&nohz.load_balancer, -1);
4809 if (atomic_read(&nohz.load_balancer) == -1) {
4810 /* make me the ilb owner */
4811 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4813 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4816 if (!(sched_smt_power_savings ||
4817 sched_mc_power_savings))
4820 * Check to see if there is a more power-efficient
4823 new_ilb = find_new_ilb(cpu);
4824 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4825 atomic_set(&nohz.load_balancer, -1);
4826 resched_cpu(new_ilb);
4832 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4835 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4837 if (atomic_read(&nohz.load_balancer) == cpu)
4838 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4845 static DEFINE_SPINLOCK(balancing);
4848 * It checks each scheduling domain to see if it is due to be balanced,
4849 * and initiates a balancing operation if so.
4851 * Balancing parameters are set up in arch_init_sched_domains.
4853 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4856 struct rq *rq = cpu_rq(cpu);
4857 unsigned long interval;
4858 struct sched_domain *sd;
4859 /* Earliest time when we have to do rebalance again */
4860 unsigned long next_balance = jiffies + 60*HZ;
4861 int update_next_balance = 0;
4864 for_each_domain(cpu, sd) {
4865 if (!(sd->flags & SD_LOAD_BALANCE))
4868 interval = sd->balance_interval;
4869 if (idle != CPU_IDLE)
4870 interval *= sd->busy_factor;
4872 /* scale ms to jiffies */
4873 interval = msecs_to_jiffies(interval);
4874 if (unlikely(!interval))
4876 if (interval > HZ*NR_CPUS/10)
4877 interval = HZ*NR_CPUS/10;
4879 need_serialize = sd->flags & SD_SERIALIZE;
4881 if (need_serialize) {
4882 if (!spin_trylock(&balancing))
4886 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4887 if (load_balance(cpu, rq, sd, idle, &balance)) {
4889 * We've pulled tasks over so either we're no
4890 * longer idle, or one of our SMT siblings is
4893 idle = CPU_NOT_IDLE;
4895 sd->last_balance = jiffies;
4898 spin_unlock(&balancing);
4900 if (time_after(next_balance, sd->last_balance + interval)) {
4901 next_balance = sd->last_balance + interval;
4902 update_next_balance = 1;
4906 * Stop the load balance at this level. There is another
4907 * CPU in our sched group which is doing load balancing more
4915 * next_balance will be updated only when there is a need.
4916 * When the cpu is attached to null domain for ex, it will not be
4919 if (likely(update_next_balance))
4920 rq->next_balance = next_balance;
4924 * run_rebalance_domains is triggered when needed from the scheduler tick.
4925 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4926 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4928 static void run_rebalance_domains(struct softirq_action *h)
4930 int this_cpu = smp_processor_id();
4931 struct rq *this_rq = cpu_rq(this_cpu);
4932 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4933 CPU_IDLE : CPU_NOT_IDLE;
4935 rebalance_domains(this_cpu, idle);
4939 * If this cpu is the owner for idle load balancing, then do the
4940 * balancing on behalf of the other idle cpus whose ticks are
4943 if (this_rq->idle_at_tick &&
4944 atomic_read(&nohz.load_balancer) == this_cpu) {
4948 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4949 if (balance_cpu == this_cpu)
4953 * If this cpu gets work to do, stop the load balancing
4954 * work being done for other cpus. Next load
4955 * balancing owner will pick it up.
4960 rebalance_domains(balance_cpu, CPU_IDLE);
4962 rq = cpu_rq(balance_cpu);
4963 if (time_after(this_rq->next_balance, rq->next_balance))
4964 this_rq->next_balance = rq->next_balance;
4970 static inline int on_null_domain(int cpu)
4972 return !rcu_dereference(cpu_rq(cpu)->sd);
4976 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4978 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4979 * idle load balancing owner or decide to stop the periodic load balancing,
4980 * if the whole system is idle.
4982 static inline void trigger_load_balance(struct rq *rq, int cpu)
4986 * If we were in the nohz mode recently and busy at the current
4987 * scheduler tick, then check if we need to nominate new idle
4990 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4991 rq->in_nohz_recently = 0;
4993 if (atomic_read(&nohz.load_balancer) == cpu) {
4994 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4995 atomic_set(&nohz.load_balancer, -1);
4998 if (atomic_read(&nohz.load_balancer) == -1) {
4999 int ilb = find_new_ilb(cpu);
5001 if (ilb < nr_cpu_ids)
5007 * If this cpu is idle and doing idle load balancing for all the
5008 * cpus with ticks stopped, is it time for that to stop?
5010 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
5011 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
5017 * If this cpu is idle and the idle load balancing is done by
5018 * someone else, then no need raise the SCHED_SOFTIRQ
5020 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
5021 cpumask_test_cpu(cpu, nohz.cpu_mask))
5024 /* Don't need to rebalance while attached to NULL domain */
5025 if (time_after_eq(jiffies, rq->next_balance) &&
5026 likely(!on_null_domain(cpu)))
5027 raise_softirq(SCHED_SOFTIRQ);
5030 #else /* CONFIG_SMP */
5033 * on UP we do not need to balance between CPUs:
5035 static inline void idle_balance(int cpu, struct rq *rq)
5041 DEFINE_PER_CPU(struct kernel_stat, kstat);
5043 EXPORT_PER_CPU_SYMBOL(kstat);
5046 * Return any ns on the sched_clock that have not yet been accounted in
5047 * @p in case that task is currently running.
5049 * Called with task_rq_lock() held on @rq.
5051 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
5055 if (task_current(rq, p)) {
5056 update_rq_clock(rq);
5057 ns = rq->clock - p->se.exec_start;
5065 unsigned long long task_delta_exec(struct task_struct *p)
5067 unsigned long flags;
5071 rq = task_rq_lock(p, &flags);
5072 ns = do_task_delta_exec(p, rq);
5073 task_rq_unlock(rq, &flags);
5079 * Return accounted runtime for the task.
5080 * In case the task is currently running, return the runtime plus current's
5081 * pending runtime that have not been accounted yet.
5083 unsigned long long task_sched_runtime(struct task_struct *p)
5085 unsigned long flags;
5089 rq = task_rq_lock(p, &flags);
5090 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
5091 task_rq_unlock(rq, &flags);
5097 * Return sum_exec_runtime for the thread group.
5098 * In case the task is currently running, return the sum plus current's
5099 * pending runtime that have not been accounted yet.
5101 * Note that the thread group might have other running tasks as well,
5102 * so the return value not includes other pending runtime that other
5103 * running tasks might have.
5105 unsigned long long thread_group_sched_runtime(struct task_struct *p)
5107 struct task_cputime totals;
5108 unsigned long flags;
5112 rq = task_rq_lock(p, &flags);
5113 thread_group_cputime(p, &totals);
5114 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
5115 task_rq_unlock(rq, &flags);
5121 * Account user cpu time to a process.
5122 * @p: the process that the cpu time gets accounted to
5123 * @cputime: the cpu time spent in user space since the last update
5124 * @cputime_scaled: cputime scaled by cpu frequency
5126 void account_user_time(struct task_struct *p, cputime_t cputime,
5127 cputime_t cputime_scaled)
5129 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5132 /* Add user time to process. */
5133 p->utime = cputime_add(p->utime, cputime);
5134 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5135 account_group_user_time(p, cputime);
5137 /* Add user time to cpustat. */
5138 tmp = cputime_to_cputime64(cputime);
5139 if (TASK_NICE(p) > 0)
5140 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5142 cpustat->user = cputime64_add(cpustat->user, tmp);
5144 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5145 /* Account for user time used */
5146 acct_update_integrals(p);
5150 * Account guest cpu time to a process.
5151 * @p: the process that the cpu time gets accounted to
5152 * @cputime: the cpu time spent in virtual machine since the last update
5153 * @cputime_scaled: cputime scaled by cpu frequency
5155 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5156 cputime_t cputime_scaled)
5159 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5161 tmp = cputime_to_cputime64(cputime);
5163 /* Add guest time to process. */
5164 p->utime = cputime_add(p->utime, cputime);
5165 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5166 account_group_user_time(p, cputime);
5167 p->gtime = cputime_add(p->gtime, cputime);
5169 /* Add guest time to cpustat. */
5170 cpustat->user = cputime64_add(cpustat->user, tmp);
5171 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5175 * Account system cpu time to a process.
5176 * @p: the process that the cpu time gets accounted to
5177 * @hardirq_offset: the offset to subtract from hardirq_count()
5178 * @cputime: the cpu time spent in kernel space since the last update
5179 * @cputime_scaled: cputime scaled by cpu frequency
5181 void account_system_time(struct task_struct *p, int hardirq_offset,
5182 cputime_t cputime, cputime_t cputime_scaled)
5184 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5187 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5188 account_guest_time(p, cputime, cputime_scaled);
5192 /* Add system time to process. */
5193 p->stime = cputime_add(p->stime, cputime);
5194 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5195 account_group_system_time(p, cputime);
5197 /* Add system time to cpustat. */
5198 tmp = cputime_to_cputime64(cputime);
5199 if (hardirq_count() - hardirq_offset)
5200 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5201 else if (softirq_count())
5202 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5204 cpustat->system = cputime64_add(cpustat->system, tmp);
5206 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5208 /* Account for system time used */
5209 acct_update_integrals(p);
5213 * Account for involuntary wait time.
5214 * @steal: the cpu time spent in involuntary wait
5216 void account_steal_time(cputime_t cputime)
5218 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5219 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5221 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5225 * Account for idle time.
5226 * @cputime: the cpu time spent in idle wait
5228 void account_idle_time(cputime_t cputime)
5230 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5231 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5232 struct rq *rq = this_rq();
5234 if (atomic_read(&rq->nr_iowait) > 0)
5235 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5237 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5240 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5243 * Account a single tick of cpu time.
5244 * @p: the process that the cpu time gets accounted to
5245 * @user_tick: indicates if the tick is a user or a system tick
5247 void account_process_tick(struct task_struct *p, int user_tick)
5249 cputime_t one_jiffy = jiffies_to_cputime(1);
5250 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
5251 struct rq *rq = this_rq();
5254 account_user_time(p, one_jiffy, one_jiffy_scaled);
5255 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5256 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
5259 account_idle_time(one_jiffy);
5263 * Account multiple ticks of steal time.
5264 * @p: the process from which the cpu time has been stolen
5265 * @ticks: number of stolen ticks
5267 void account_steal_ticks(unsigned long ticks)
5269 account_steal_time(jiffies_to_cputime(ticks));
5273 * Account multiple ticks of idle time.
5274 * @ticks: number of stolen ticks
5276 void account_idle_ticks(unsigned long ticks)
5278 account_idle_time(jiffies_to_cputime(ticks));
5284 * Use precise platform statistics if available:
5286 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5287 cputime_t task_utime(struct task_struct *p)
5292 cputime_t task_stime(struct task_struct *p)
5297 cputime_t task_utime(struct task_struct *p)
5299 clock_t utime = cputime_to_clock_t(p->utime),
5300 total = utime + cputime_to_clock_t(p->stime);
5304 * Use CFS's precise accounting:
5306 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5310 do_div(temp, total);
5312 utime = (clock_t)temp;
5314 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5315 return p->prev_utime;
5318 cputime_t task_stime(struct task_struct *p)
5323 * Use CFS's precise accounting. (we subtract utime from
5324 * the total, to make sure the total observed by userspace
5325 * grows monotonically - apps rely on that):
5327 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5328 cputime_to_clock_t(task_utime(p));
5331 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5333 return p->prev_stime;
5337 inline cputime_t task_gtime(struct task_struct *p)
5343 * This function gets called by the timer code, with HZ frequency.
5344 * We call it with interrupts disabled.
5346 * It also gets called by the fork code, when changing the parent's
5349 void scheduler_tick(void)
5351 int cpu = smp_processor_id();
5352 struct rq *rq = cpu_rq(cpu);
5353 struct task_struct *curr = rq->curr;
5357 spin_lock(&rq->lock);
5358 update_rq_clock(rq);
5359 update_cpu_load(rq);
5360 curr->sched_class->task_tick(rq, curr, 0);
5361 spin_unlock(&rq->lock);
5363 perf_counter_task_tick(curr, cpu);
5366 rq->idle_at_tick = idle_cpu(cpu);
5367 trigger_load_balance(rq, cpu);
5371 notrace unsigned long get_parent_ip(unsigned long addr)
5373 if (in_lock_functions(addr)) {
5374 addr = CALLER_ADDR2;
5375 if (in_lock_functions(addr))
5376 addr = CALLER_ADDR3;
5381 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5382 defined(CONFIG_PREEMPT_TRACER))
5384 void __kprobes add_preempt_count(int val)
5386 #ifdef CONFIG_DEBUG_PREEMPT
5390 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5393 preempt_count() += val;
5394 #ifdef CONFIG_DEBUG_PREEMPT
5396 * Spinlock count overflowing soon?
5398 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5401 if (preempt_count() == val)
5402 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5404 EXPORT_SYMBOL(add_preempt_count);
5406 void __kprobes sub_preempt_count(int val)
5408 #ifdef CONFIG_DEBUG_PREEMPT
5412 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5415 * Is the spinlock portion underflowing?
5417 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5418 !(preempt_count() & PREEMPT_MASK)))
5422 if (preempt_count() == val)
5423 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5424 preempt_count() -= val;
5426 EXPORT_SYMBOL(sub_preempt_count);
5431 * Print scheduling while atomic bug:
5433 static noinline void __schedule_bug(struct task_struct *prev)
5435 struct pt_regs *regs = get_irq_regs();
5437 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5438 prev->comm, prev->pid, preempt_count());
5440 debug_show_held_locks(prev);
5442 if (irqs_disabled())
5443 print_irqtrace_events(prev);
5452 * Various schedule()-time debugging checks and statistics:
5454 static inline void schedule_debug(struct task_struct *prev)
5457 * Test if we are atomic. Since do_exit() needs to call into
5458 * schedule() atomically, we ignore that path for now.
5459 * Otherwise, whine if we are scheduling when we should not be.
5461 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5462 __schedule_bug(prev);
5464 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5466 schedstat_inc(this_rq(), sched_count);
5467 #ifdef CONFIG_SCHEDSTATS
5468 if (unlikely(prev->lock_depth >= 0)) {
5469 schedstat_inc(this_rq(), bkl_count);
5470 schedstat_inc(prev, sched_info.bkl_count);
5475 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5477 if (prev->state == TASK_RUNNING) {
5478 u64 runtime = prev->se.sum_exec_runtime;
5480 runtime -= prev->se.prev_sum_exec_runtime;
5481 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5484 * In order to avoid avg_overlap growing stale when we are
5485 * indeed overlapping and hence not getting put to sleep, grow
5486 * the avg_overlap on preemption.
5488 * We use the average preemption runtime because that
5489 * correlates to the amount of cache footprint a task can
5492 update_avg(&prev->se.avg_overlap, runtime);
5494 prev->sched_class->put_prev_task(rq, prev);
5498 * Pick up the highest-prio task:
5500 static inline struct task_struct *
5501 pick_next_task(struct rq *rq)
5503 const struct sched_class *class;
5504 struct task_struct *p;
5507 * Optimization: we know that if all tasks are in
5508 * the fair class we can call that function directly:
5510 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5511 p = fair_sched_class.pick_next_task(rq);
5516 class = sched_class_highest;
5518 p = class->pick_next_task(rq);
5522 * Will never be NULL as the idle class always
5523 * returns a non-NULL p:
5525 class = class->next;
5530 * schedule() is the main scheduler function.
5532 asmlinkage void __sched schedule(void)
5534 struct task_struct *prev, *next;
5535 unsigned long *switch_count;
5541 cpu = smp_processor_id();
5545 switch_count = &prev->nivcsw;
5547 release_kernel_lock(prev);
5548 need_resched_nonpreemptible:
5550 schedule_debug(prev);
5552 if (sched_feat(HRTICK))
5555 spin_lock_irq(&rq->lock);
5556 update_rq_clock(rq);
5557 clear_tsk_need_resched(prev);
5559 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5560 if (unlikely(signal_pending_state(prev->state, prev)))
5561 prev->state = TASK_RUNNING;
5563 deactivate_task(rq, prev, 1);
5564 switch_count = &prev->nvcsw;
5567 pre_schedule(rq, prev);
5569 if (unlikely(!rq->nr_running))
5570 idle_balance(cpu, rq);
5572 put_prev_task(rq, prev);
5573 next = pick_next_task(rq);
5575 if (likely(prev != next)) {
5576 sched_info_switch(prev, next);
5577 perf_counter_task_sched_out(prev, next, cpu);
5583 context_switch(rq, prev, next); /* unlocks the rq */
5585 * the context switch might have flipped the stack from under
5586 * us, hence refresh the local variables.
5588 cpu = smp_processor_id();
5591 spin_unlock_irq(&rq->lock);
5595 if (unlikely(reacquire_kernel_lock(current) < 0))
5596 goto need_resched_nonpreemptible;
5598 preempt_enable_no_resched();
5602 EXPORT_SYMBOL(schedule);
5606 * Look out! "owner" is an entirely speculative pointer
5607 * access and not reliable.
5609 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5614 if (!sched_feat(OWNER_SPIN))
5617 #ifdef CONFIG_DEBUG_PAGEALLOC
5619 * Need to access the cpu field knowing that
5620 * DEBUG_PAGEALLOC could have unmapped it if
5621 * the mutex owner just released it and exited.
5623 if (probe_kernel_address(&owner->cpu, cpu))
5630 * Even if the access succeeded (likely case),
5631 * the cpu field may no longer be valid.
5633 if (cpu >= nr_cpumask_bits)
5637 * We need to validate that we can do a
5638 * get_cpu() and that we have the percpu area.
5640 if (!cpu_online(cpu))
5647 * Owner changed, break to re-assess state.
5649 if (lock->owner != owner)
5653 * Is that owner really running on that cpu?
5655 if (task_thread_info(rq->curr) != owner || need_resched())
5665 #ifdef CONFIG_PREEMPT
5667 * this is the entry point to schedule() from in-kernel preemption
5668 * off of preempt_enable. Kernel preemptions off return from interrupt
5669 * occur there and call schedule directly.
5671 asmlinkage void __sched preempt_schedule(void)
5673 struct thread_info *ti = current_thread_info();
5676 * If there is a non-zero preempt_count or interrupts are disabled,
5677 * we do not want to preempt the current task. Just return..
5679 if (likely(ti->preempt_count || irqs_disabled()))
5683 add_preempt_count(PREEMPT_ACTIVE);
5685 sub_preempt_count(PREEMPT_ACTIVE);
5688 * Check again in case we missed a preemption opportunity
5689 * between schedule and now.
5692 } while (need_resched());
5694 EXPORT_SYMBOL(preempt_schedule);
5697 * this is the entry point to schedule() from kernel preemption
5698 * off of irq context.
5699 * Note, that this is called and return with irqs disabled. This will
5700 * protect us against recursive calling from irq.
5702 asmlinkage void __sched preempt_schedule_irq(void)
5704 struct thread_info *ti = current_thread_info();
5706 /* Catch callers which need to be fixed */
5707 BUG_ON(ti->preempt_count || !irqs_disabled());
5710 add_preempt_count(PREEMPT_ACTIVE);
5713 local_irq_disable();
5714 sub_preempt_count(PREEMPT_ACTIVE);
5717 * Check again in case we missed a preemption opportunity
5718 * between schedule and now.
5721 } while (need_resched());
5724 #endif /* CONFIG_PREEMPT */
5726 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5729 return try_to_wake_up(curr->private, mode, sync);
5731 EXPORT_SYMBOL(default_wake_function);
5734 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5735 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5736 * number) then we wake all the non-exclusive tasks and one exclusive task.
5738 * There are circumstances in which we can try to wake a task which has already
5739 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5740 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5742 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5743 int nr_exclusive, int sync, void *key)
5745 wait_queue_t *curr, *next;
5747 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5748 unsigned flags = curr->flags;
5750 if (curr->func(curr, mode, sync, key) &&
5751 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5757 * __wake_up - wake up threads blocked on a waitqueue.
5759 * @mode: which threads
5760 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5761 * @key: is directly passed to the wakeup function
5763 * It may be assumed that this function implies a write memory barrier before
5764 * changing the task state if and only if any tasks are woken up.
5766 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5767 int nr_exclusive, void *key)
5769 unsigned long flags;
5771 spin_lock_irqsave(&q->lock, flags);
5772 __wake_up_common(q, mode, nr_exclusive, 0, key);
5773 spin_unlock_irqrestore(&q->lock, flags);
5775 EXPORT_SYMBOL(__wake_up);
5778 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5780 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5782 __wake_up_common(q, mode, 1, 0, NULL);
5785 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5787 __wake_up_common(q, mode, 1, 0, key);
5791 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5793 * @mode: which threads
5794 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5795 * @key: opaque value to be passed to wakeup targets
5797 * The sync wakeup differs that the waker knows that it will schedule
5798 * away soon, so while the target thread will be woken up, it will not
5799 * be migrated to another CPU - ie. the two threads are 'synchronized'
5800 * with each other. This can prevent needless bouncing between CPUs.
5802 * On UP it can prevent extra preemption.
5804 * It may be assumed that this function implies a write memory barrier before
5805 * changing the task state if and only if any tasks are woken up.
5807 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5808 int nr_exclusive, void *key)
5810 unsigned long flags;
5816 if (unlikely(!nr_exclusive))
5819 spin_lock_irqsave(&q->lock, flags);
5820 __wake_up_common(q, mode, nr_exclusive, sync, key);
5821 spin_unlock_irqrestore(&q->lock, flags);
5823 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5826 * __wake_up_sync - see __wake_up_sync_key()
5828 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5830 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5832 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5835 * complete: - signals a single thread waiting on this completion
5836 * @x: holds the state of this particular completion
5838 * This will wake up a single thread waiting on this completion. Threads will be
5839 * awakened in the same order in which they were queued.
5841 * See also complete_all(), wait_for_completion() and related routines.
5843 * It may be assumed that this function implies a write memory barrier before
5844 * changing the task state if and only if any tasks are woken up.
5846 void complete(struct completion *x)
5848 unsigned long flags;
5850 spin_lock_irqsave(&x->wait.lock, flags);
5852 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5853 spin_unlock_irqrestore(&x->wait.lock, flags);
5855 EXPORT_SYMBOL(complete);
5858 * complete_all: - signals all threads waiting on this completion
5859 * @x: holds the state of this particular completion
5861 * This will wake up all threads waiting on this particular completion event.
5863 * It may be assumed that this function implies a write memory barrier before
5864 * changing the task state if and only if any tasks are woken up.
5866 void complete_all(struct completion *x)
5868 unsigned long flags;
5870 spin_lock_irqsave(&x->wait.lock, flags);
5871 x->done += UINT_MAX/2;
5872 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5873 spin_unlock_irqrestore(&x->wait.lock, flags);
5875 EXPORT_SYMBOL(complete_all);
5877 static inline long __sched
5878 do_wait_for_common(struct completion *x, long timeout, int state)
5881 DECLARE_WAITQUEUE(wait, current);
5883 wait.flags |= WQ_FLAG_EXCLUSIVE;
5884 __add_wait_queue_tail(&x->wait, &wait);
5886 if (signal_pending_state(state, current)) {
5887 timeout = -ERESTARTSYS;
5890 __set_current_state(state);
5891 spin_unlock_irq(&x->wait.lock);
5892 timeout = schedule_timeout(timeout);
5893 spin_lock_irq(&x->wait.lock);
5894 } while (!x->done && timeout);
5895 __remove_wait_queue(&x->wait, &wait);
5900 return timeout ?: 1;
5904 wait_for_common(struct completion *x, long timeout, int state)
5908 spin_lock_irq(&x->wait.lock);
5909 timeout = do_wait_for_common(x, timeout, state);
5910 spin_unlock_irq(&x->wait.lock);
5915 * wait_for_completion: - waits for completion of a task
5916 * @x: holds the state of this particular completion
5918 * This waits to be signaled for completion of a specific task. It is NOT
5919 * interruptible and there is no timeout.
5921 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5922 * and interrupt capability. Also see complete().
5924 void __sched wait_for_completion(struct completion *x)
5926 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5928 EXPORT_SYMBOL(wait_for_completion);
5931 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5932 * @x: holds the state of this particular completion
5933 * @timeout: timeout value in jiffies
5935 * This waits for either a completion of a specific task to be signaled or for a
5936 * specified timeout to expire. The timeout is in jiffies. It is not
5939 unsigned long __sched
5940 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5942 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5944 EXPORT_SYMBOL(wait_for_completion_timeout);
5947 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5948 * @x: holds the state of this particular completion
5950 * This waits for completion of a specific task to be signaled. It is
5953 int __sched wait_for_completion_interruptible(struct completion *x)
5955 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5956 if (t == -ERESTARTSYS)
5960 EXPORT_SYMBOL(wait_for_completion_interruptible);
5963 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5964 * @x: holds the state of this particular completion
5965 * @timeout: timeout value in jiffies
5967 * This waits for either a completion of a specific task to be signaled or for a
5968 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5970 unsigned long __sched
5971 wait_for_completion_interruptible_timeout(struct completion *x,
5972 unsigned long timeout)
5974 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5976 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5979 * wait_for_completion_killable: - waits for completion of a task (killable)
5980 * @x: holds the state of this particular completion
5982 * This waits to be signaled for completion of a specific task. It can be
5983 * interrupted by a kill signal.
5985 int __sched wait_for_completion_killable(struct completion *x)
5987 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5988 if (t == -ERESTARTSYS)
5992 EXPORT_SYMBOL(wait_for_completion_killable);
5995 * try_wait_for_completion - try to decrement a completion without blocking
5996 * @x: completion structure
5998 * Returns: 0 if a decrement cannot be done without blocking
5999 * 1 if a decrement succeeded.
6001 * If a completion is being used as a counting completion,
6002 * attempt to decrement the counter without blocking. This
6003 * enables us to avoid waiting if the resource the completion
6004 * is protecting is not available.
6006 bool try_wait_for_completion(struct completion *x)
6010 spin_lock_irq(&x->wait.lock);
6015 spin_unlock_irq(&x->wait.lock);
6018 EXPORT_SYMBOL(try_wait_for_completion);
6021 * completion_done - Test to see if a completion has any waiters
6022 * @x: completion structure
6024 * Returns: 0 if there are waiters (wait_for_completion() in progress)
6025 * 1 if there are no waiters.
6028 bool completion_done(struct completion *x)
6032 spin_lock_irq(&x->wait.lock);
6035 spin_unlock_irq(&x->wait.lock);
6038 EXPORT_SYMBOL(completion_done);
6041 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
6043 unsigned long flags;
6046 init_waitqueue_entry(&wait, current);
6048 __set_current_state(state);
6050 spin_lock_irqsave(&q->lock, flags);
6051 __add_wait_queue(q, &wait);
6052 spin_unlock(&q->lock);
6053 timeout = schedule_timeout(timeout);
6054 spin_lock_irq(&q->lock);
6055 __remove_wait_queue(q, &wait);
6056 spin_unlock_irqrestore(&q->lock, flags);
6061 void __sched interruptible_sleep_on(wait_queue_head_t *q)
6063 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6065 EXPORT_SYMBOL(interruptible_sleep_on);
6068 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
6070 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
6072 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
6074 void __sched sleep_on(wait_queue_head_t *q)
6076 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6078 EXPORT_SYMBOL(sleep_on);
6080 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
6082 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
6084 EXPORT_SYMBOL(sleep_on_timeout);
6086 #ifdef CONFIG_RT_MUTEXES
6089 * rt_mutex_setprio - set the current priority of a task
6091 * @prio: prio value (kernel-internal form)
6093 * This function changes the 'effective' priority of a task. It does
6094 * not touch ->normal_prio like __setscheduler().
6096 * Used by the rt_mutex code to implement priority inheritance logic.
6098 void rt_mutex_setprio(struct task_struct *p, int prio)
6100 unsigned long flags;
6101 int oldprio, on_rq, running;
6103 const struct sched_class *prev_class = p->sched_class;
6105 BUG_ON(prio < 0 || prio > MAX_PRIO);
6107 rq = task_rq_lock(p, &flags);
6108 update_rq_clock(rq);
6111 on_rq = p->se.on_rq;
6112 running = task_current(rq, p);
6114 dequeue_task(rq, p, 0);
6116 p->sched_class->put_prev_task(rq, p);
6119 p->sched_class = &rt_sched_class;
6121 p->sched_class = &fair_sched_class;
6126 p->sched_class->set_curr_task(rq);
6128 enqueue_task(rq, p, 0);
6130 check_class_changed(rq, p, prev_class, oldprio, running);
6132 task_rq_unlock(rq, &flags);
6137 void set_user_nice(struct task_struct *p, long nice)
6139 int old_prio, delta, on_rq;
6140 unsigned long flags;
6143 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6146 * We have to be careful, if called from sys_setpriority(),
6147 * the task might be in the middle of scheduling on another CPU.
6149 rq = task_rq_lock(p, &flags);
6150 update_rq_clock(rq);
6152 * The RT priorities are set via sched_setscheduler(), but we still
6153 * allow the 'normal' nice value to be set - but as expected
6154 * it wont have any effect on scheduling until the task is
6155 * SCHED_FIFO/SCHED_RR:
6157 if (task_has_rt_policy(p)) {
6158 p->static_prio = NICE_TO_PRIO(nice);
6161 on_rq = p->se.on_rq;
6163 dequeue_task(rq, p, 0);
6165 p->static_prio = NICE_TO_PRIO(nice);
6168 p->prio = effective_prio(p);
6169 delta = p->prio - old_prio;
6172 enqueue_task(rq, p, 0);
6174 * If the task increased its priority or is running and
6175 * lowered its priority, then reschedule its CPU:
6177 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6178 resched_task(rq->curr);
6181 task_rq_unlock(rq, &flags);
6183 EXPORT_SYMBOL(set_user_nice);
6186 * can_nice - check if a task can reduce its nice value
6190 int can_nice(const struct task_struct *p, const int nice)
6192 /* convert nice value [19,-20] to rlimit style value [1,40] */
6193 int nice_rlim = 20 - nice;
6195 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6196 capable(CAP_SYS_NICE));
6199 #ifdef __ARCH_WANT_SYS_NICE
6202 * sys_nice - change the priority of the current process.
6203 * @increment: priority increment
6205 * sys_setpriority is a more generic, but much slower function that
6206 * does similar things.
6208 SYSCALL_DEFINE1(nice, int, increment)
6213 * Setpriority might change our priority at the same moment.
6214 * We don't have to worry. Conceptually one call occurs first
6215 * and we have a single winner.
6217 if (increment < -40)
6222 nice = TASK_NICE(current) + increment;
6228 if (increment < 0 && !can_nice(current, nice))
6231 retval = security_task_setnice(current, nice);
6235 set_user_nice(current, nice);
6242 * task_prio - return the priority value of a given task.
6243 * @p: the task in question.
6245 * This is the priority value as seen by users in /proc.
6246 * RT tasks are offset by -200. Normal tasks are centered
6247 * around 0, value goes from -16 to +15.
6249 int task_prio(const struct task_struct *p)
6251 return p->prio - MAX_RT_PRIO;
6255 * task_nice - return the nice value of a given task.
6256 * @p: the task in question.
6258 int task_nice(const struct task_struct *p)
6260 return TASK_NICE(p);
6262 EXPORT_SYMBOL(task_nice);
6265 * idle_cpu - is a given cpu idle currently?
6266 * @cpu: the processor in question.
6268 int idle_cpu(int cpu)
6270 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6274 * idle_task - return the idle task for a given cpu.
6275 * @cpu: the processor in question.
6277 struct task_struct *idle_task(int cpu)
6279 return cpu_rq(cpu)->idle;
6283 * find_process_by_pid - find a process with a matching PID value.
6284 * @pid: the pid in question.
6286 static struct task_struct *find_process_by_pid(pid_t pid)
6288 return pid ? find_task_by_vpid(pid) : current;
6291 /* Actually do priority change: must hold rq lock. */
6293 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6295 BUG_ON(p->se.on_rq);
6298 switch (p->policy) {
6302 p->sched_class = &fair_sched_class;
6306 p->sched_class = &rt_sched_class;
6310 p->rt_priority = prio;
6311 p->normal_prio = normal_prio(p);
6312 /* we are holding p->pi_lock already */
6313 p->prio = rt_mutex_getprio(p);
6318 * check the target process has a UID that matches the current process's
6320 static bool check_same_owner(struct task_struct *p)
6322 const struct cred *cred = current_cred(), *pcred;
6326 pcred = __task_cred(p);
6327 match = (cred->euid == pcred->euid ||
6328 cred->euid == pcred->uid);
6333 static int __sched_setscheduler(struct task_struct *p, int policy,
6334 struct sched_param *param, bool user)
6336 int retval, oldprio, oldpolicy = -1, on_rq, running;
6337 unsigned long flags;
6338 const struct sched_class *prev_class = p->sched_class;
6342 /* may grab non-irq protected spin_locks */
6343 BUG_ON(in_interrupt());
6345 /* double check policy once rq lock held */
6347 reset_on_fork = p->sched_reset_on_fork;
6348 policy = oldpolicy = p->policy;
6350 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6351 policy &= ~SCHED_RESET_ON_FORK;
6353 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6354 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6355 policy != SCHED_IDLE)
6360 * Valid priorities for SCHED_FIFO and SCHED_RR are
6361 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6362 * SCHED_BATCH and SCHED_IDLE is 0.
6364 if (param->sched_priority < 0 ||
6365 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6366 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6368 if (rt_policy(policy) != (param->sched_priority != 0))
6372 * Allow unprivileged RT tasks to decrease priority:
6374 if (user && !capable(CAP_SYS_NICE)) {
6375 if (rt_policy(policy)) {
6376 unsigned long rlim_rtprio;
6378 if (!lock_task_sighand(p, &flags))
6380 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6381 unlock_task_sighand(p, &flags);
6383 /* can't set/change the rt policy */
6384 if (policy != p->policy && !rlim_rtprio)
6387 /* can't increase priority */
6388 if (param->sched_priority > p->rt_priority &&
6389 param->sched_priority > rlim_rtprio)
6393 * Like positive nice levels, dont allow tasks to
6394 * move out of SCHED_IDLE either:
6396 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6399 /* can't change other user's priorities */
6400 if (!check_same_owner(p))
6403 /* Normal users shall not reset the sched_reset_on_fork flag */
6404 if (p->sched_reset_on_fork && !reset_on_fork)
6409 #ifdef CONFIG_RT_GROUP_SCHED
6411 * Do not allow realtime tasks into groups that have no runtime
6414 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6415 task_group(p)->rt_bandwidth.rt_runtime == 0)
6419 retval = security_task_setscheduler(p, policy, param);
6425 * make sure no PI-waiters arrive (or leave) while we are
6426 * changing the priority of the task:
6428 spin_lock_irqsave(&p->pi_lock, flags);
6430 * To be able to change p->policy safely, the apropriate
6431 * runqueue lock must be held.
6433 rq = __task_rq_lock(p);
6434 /* recheck policy now with rq lock held */
6435 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6436 policy = oldpolicy = -1;
6437 __task_rq_unlock(rq);
6438 spin_unlock_irqrestore(&p->pi_lock, flags);
6441 update_rq_clock(rq);
6442 on_rq = p->se.on_rq;
6443 running = task_current(rq, p);
6445 deactivate_task(rq, p, 0);
6447 p->sched_class->put_prev_task(rq, p);
6449 p->sched_reset_on_fork = reset_on_fork;
6452 __setscheduler(rq, p, policy, param->sched_priority);
6455 p->sched_class->set_curr_task(rq);
6457 activate_task(rq, p, 0);
6459 check_class_changed(rq, p, prev_class, oldprio, running);
6461 __task_rq_unlock(rq);
6462 spin_unlock_irqrestore(&p->pi_lock, flags);
6464 rt_mutex_adjust_pi(p);
6470 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6471 * @p: the task in question.
6472 * @policy: new policy.
6473 * @param: structure containing the new RT priority.
6475 * NOTE that the task may be already dead.
6477 int sched_setscheduler(struct task_struct *p, int policy,
6478 struct sched_param *param)
6480 return __sched_setscheduler(p, policy, param, true);
6482 EXPORT_SYMBOL_GPL(sched_setscheduler);
6485 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6486 * @p: the task in question.
6487 * @policy: new policy.
6488 * @param: structure containing the new RT priority.
6490 * Just like sched_setscheduler, only don't bother checking if the
6491 * current context has permission. For example, this is needed in
6492 * stop_machine(): we create temporary high priority worker threads,
6493 * but our caller might not have that capability.
6495 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6496 struct sched_param *param)
6498 return __sched_setscheduler(p, policy, param, false);
6502 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6504 struct sched_param lparam;
6505 struct task_struct *p;
6508 if (!param || pid < 0)
6510 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6515 p = find_process_by_pid(pid);
6517 retval = sched_setscheduler(p, policy, &lparam);
6524 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6525 * @pid: the pid in question.
6526 * @policy: new policy.
6527 * @param: structure containing the new RT priority.
6529 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6530 struct sched_param __user *, param)
6532 /* negative values for policy are not valid */
6536 return do_sched_setscheduler(pid, policy, param);
6540 * sys_sched_setparam - set/change the RT priority of a thread
6541 * @pid: the pid in question.
6542 * @param: structure containing the new RT priority.
6544 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6546 return do_sched_setscheduler(pid, -1, param);
6550 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6551 * @pid: the pid in question.
6553 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6555 struct task_struct *p;
6562 read_lock(&tasklist_lock);
6563 p = find_process_by_pid(pid);
6565 retval = security_task_getscheduler(p);
6568 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6570 read_unlock(&tasklist_lock);
6575 * sys_sched_getparam - get the RT priority of a thread
6576 * @pid: the pid in question.
6577 * @param: structure containing the RT priority.
6579 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6581 struct sched_param lp;
6582 struct task_struct *p;
6585 if (!param || pid < 0)
6588 read_lock(&tasklist_lock);
6589 p = find_process_by_pid(pid);
6594 retval = security_task_getscheduler(p);
6598 lp.sched_priority = p->rt_priority;
6599 read_unlock(&tasklist_lock);
6602 * This one might sleep, we cannot do it with a spinlock held ...
6604 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6609 read_unlock(&tasklist_lock);
6613 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6615 cpumask_var_t cpus_allowed, new_mask;
6616 struct task_struct *p;
6620 read_lock(&tasklist_lock);
6622 p = find_process_by_pid(pid);
6624 read_unlock(&tasklist_lock);
6630 * It is not safe to call set_cpus_allowed with the
6631 * tasklist_lock held. We will bump the task_struct's
6632 * usage count and then drop tasklist_lock.
6635 read_unlock(&tasklist_lock);
6637 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6641 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6643 goto out_free_cpus_allowed;
6646 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6649 retval = security_task_setscheduler(p, 0, NULL);
6653 cpuset_cpus_allowed(p, cpus_allowed);
6654 cpumask_and(new_mask, in_mask, cpus_allowed);
6656 retval = set_cpus_allowed_ptr(p, new_mask);
6659 cpuset_cpus_allowed(p, cpus_allowed);
6660 if (!cpumask_subset(new_mask, cpus_allowed)) {
6662 * We must have raced with a concurrent cpuset
6663 * update. Just reset the cpus_allowed to the
6664 * cpuset's cpus_allowed
6666 cpumask_copy(new_mask, cpus_allowed);
6671 free_cpumask_var(new_mask);
6672 out_free_cpus_allowed:
6673 free_cpumask_var(cpus_allowed);
6680 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6681 struct cpumask *new_mask)
6683 if (len < cpumask_size())
6684 cpumask_clear(new_mask);
6685 else if (len > cpumask_size())
6686 len = cpumask_size();
6688 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6692 * sys_sched_setaffinity - set the cpu affinity of a process
6693 * @pid: pid of the process
6694 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6695 * @user_mask_ptr: user-space pointer to the new cpu mask
6697 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6698 unsigned long __user *, user_mask_ptr)
6700 cpumask_var_t new_mask;
6703 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6706 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6708 retval = sched_setaffinity(pid, new_mask);
6709 free_cpumask_var(new_mask);
6713 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6715 struct task_struct *p;
6719 read_lock(&tasklist_lock);
6722 p = find_process_by_pid(pid);
6726 retval = security_task_getscheduler(p);
6730 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6733 read_unlock(&tasklist_lock);
6740 * sys_sched_getaffinity - get the cpu affinity of a process
6741 * @pid: pid of the process
6742 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6743 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6745 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6746 unsigned long __user *, user_mask_ptr)
6751 if (len < cpumask_size())
6754 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6757 ret = sched_getaffinity(pid, mask);
6759 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6762 ret = cpumask_size();
6764 free_cpumask_var(mask);
6770 * sys_sched_yield - yield the current processor to other threads.
6772 * This function yields the current CPU to other tasks. If there are no
6773 * other threads running on this CPU then this function will return.
6775 SYSCALL_DEFINE0(sched_yield)
6777 struct rq *rq = this_rq_lock();
6779 schedstat_inc(rq, yld_count);
6780 current->sched_class->yield_task(rq);
6783 * Since we are going to call schedule() anyway, there's
6784 * no need to preempt or enable interrupts:
6786 __release(rq->lock);
6787 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6788 _raw_spin_unlock(&rq->lock);
6789 preempt_enable_no_resched();
6796 static inline int should_resched(void)
6798 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6801 static void __cond_resched(void)
6803 add_preempt_count(PREEMPT_ACTIVE);
6805 sub_preempt_count(PREEMPT_ACTIVE);
6808 int __sched _cond_resched(void)
6810 if (should_resched()) {
6816 EXPORT_SYMBOL(_cond_resched);
6819 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6820 * call schedule, and on return reacquire the lock.
6822 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6823 * operations here to prevent schedule() from being called twice (once via
6824 * spin_unlock(), once by hand).
6826 int __cond_resched_lock(spinlock_t *lock)
6828 int resched = should_resched();
6831 if (spin_needbreak(lock) || resched) {
6842 EXPORT_SYMBOL(__cond_resched_lock);
6844 int __sched __cond_resched_softirq(void)
6846 BUG_ON(!in_softirq());
6848 if (should_resched()) {
6856 EXPORT_SYMBOL(__cond_resched_softirq);
6859 * yield - yield the current processor to other threads.
6861 * This is a shortcut for kernel-space yielding - it marks the
6862 * thread runnable and calls sys_sched_yield().
6864 void __sched yield(void)
6866 set_current_state(TASK_RUNNING);
6869 EXPORT_SYMBOL(yield);
6872 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6873 * that process accounting knows that this is a task in IO wait state.
6875 * But don't do that if it is a deliberate, throttling IO wait (this task
6876 * has set its backing_dev_info: the queue against which it should throttle)
6878 void __sched io_schedule(void)
6880 struct rq *rq = raw_rq();
6882 delayacct_blkio_start();
6883 atomic_inc(&rq->nr_iowait);
6884 current->in_iowait = 1;
6886 current->in_iowait = 0;
6887 atomic_dec(&rq->nr_iowait);
6888 delayacct_blkio_end();
6890 EXPORT_SYMBOL(io_schedule);
6892 long __sched io_schedule_timeout(long timeout)
6894 struct rq *rq = raw_rq();
6897 delayacct_blkio_start();
6898 atomic_inc(&rq->nr_iowait);
6899 current->in_iowait = 1;
6900 ret = schedule_timeout(timeout);
6901 current->in_iowait = 0;
6902 atomic_dec(&rq->nr_iowait);
6903 delayacct_blkio_end();
6908 * sys_sched_get_priority_max - return maximum RT priority.
6909 * @policy: scheduling class.
6911 * this syscall returns the maximum rt_priority that can be used
6912 * by a given scheduling class.
6914 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6921 ret = MAX_USER_RT_PRIO-1;
6933 * sys_sched_get_priority_min - return minimum RT priority.
6934 * @policy: scheduling class.
6936 * this syscall returns the minimum rt_priority that can be used
6937 * by a given scheduling class.
6939 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6957 * sys_sched_rr_get_interval - return the default timeslice of a process.
6958 * @pid: pid of the process.
6959 * @interval: userspace pointer to the timeslice value.
6961 * this syscall writes the default timeslice value of a given process
6962 * into the user-space timespec buffer. A value of '0' means infinity.
6964 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6965 struct timespec __user *, interval)
6967 struct task_struct *p;
6968 unsigned int time_slice;
6976 read_lock(&tasklist_lock);
6977 p = find_process_by_pid(pid);
6981 retval = security_task_getscheduler(p);
6986 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6987 * tasks that are on an otherwise idle runqueue:
6990 if (p->policy == SCHED_RR) {
6991 time_slice = DEF_TIMESLICE;
6992 } else if (p->policy != SCHED_FIFO) {
6993 struct sched_entity *se = &p->se;
6994 unsigned long flags;
6997 rq = task_rq_lock(p, &flags);
6998 if (rq->cfs.load.weight)
6999 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
7000 task_rq_unlock(rq, &flags);
7002 read_unlock(&tasklist_lock);
7003 jiffies_to_timespec(time_slice, &t);
7004 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
7008 read_unlock(&tasklist_lock);
7012 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
7014 void sched_show_task(struct task_struct *p)
7016 unsigned long free = 0;
7019 state = p->state ? __ffs(p->state) + 1 : 0;
7020 printk(KERN_INFO "%-13.13s %c", p->comm,
7021 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
7022 #if BITS_PER_LONG == 32
7023 if (state == TASK_RUNNING)
7024 printk(KERN_CONT " running ");
7026 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
7028 if (state == TASK_RUNNING)
7029 printk(KERN_CONT " running task ");
7031 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
7033 #ifdef CONFIG_DEBUG_STACK_USAGE
7034 free = stack_not_used(p);
7036 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
7037 task_pid_nr(p), task_pid_nr(p->real_parent),
7038 (unsigned long)task_thread_info(p)->flags);
7040 show_stack(p, NULL);
7043 void show_state_filter(unsigned long state_filter)
7045 struct task_struct *g, *p;
7047 #if BITS_PER_LONG == 32
7049 " task PC stack pid father\n");
7052 " task PC stack pid father\n");
7054 read_lock(&tasklist_lock);
7055 do_each_thread(g, p) {
7057 * reset the NMI-timeout, listing all files on a slow
7058 * console might take alot of time:
7060 touch_nmi_watchdog();
7061 if (!state_filter || (p->state & state_filter))
7063 } while_each_thread(g, p);
7065 touch_all_softlockup_watchdogs();
7067 #ifdef CONFIG_SCHED_DEBUG
7068 sysrq_sched_debug_show();
7070 read_unlock(&tasklist_lock);
7072 * Only show locks if all tasks are dumped:
7074 if (state_filter == -1)
7075 debug_show_all_locks();
7078 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
7080 idle->sched_class = &idle_sched_class;
7084 * init_idle - set up an idle thread for a given CPU
7085 * @idle: task in question
7086 * @cpu: cpu the idle task belongs to
7088 * NOTE: this function does not set the idle thread's NEED_RESCHED
7089 * flag, to make booting more robust.
7091 void __cpuinit init_idle(struct task_struct *idle, int cpu)
7093 struct rq *rq = cpu_rq(cpu);
7094 unsigned long flags;
7096 spin_lock_irqsave(&rq->lock, flags);
7099 idle->se.exec_start = sched_clock();
7101 idle->prio = idle->normal_prio = MAX_PRIO;
7102 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
7103 __set_task_cpu(idle, cpu);
7105 rq->curr = rq->idle = idle;
7106 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7109 spin_unlock_irqrestore(&rq->lock, flags);
7111 /* Set the preempt count _outside_ the spinlocks! */
7112 #if defined(CONFIG_PREEMPT)
7113 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7115 task_thread_info(idle)->preempt_count = 0;
7118 * The idle tasks have their own, simple scheduling class:
7120 idle->sched_class = &idle_sched_class;
7121 ftrace_graph_init_task(idle);
7125 * In a system that switches off the HZ timer nohz_cpu_mask
7126 * indicates which cpus entered this state. This is used
7127 * in the rcu update to wait only for active cpus. For system
7128 * which do not switch off the HZ timer nohz_cpu_mask should
7129 * always be CPU_BITS_NONE.
7131 cpumask_var_t nohz_cpu_mask;
7134 * Increase the granularity value when there are more CPUs,
7135 * because with more CPUs the 'effective latency' as visible
7136 * to users decreases. But the relationship is not linear,
7137 * so pick a second-best guess by going with the log2 of the
7140 * This idea comes from the SD scheduler of Con Kolivas:
7142 static inline void sched_init_granularity(void)
7144 unsigned int factor = 1 + ilog2(num_online_cpus());
7145 const unsigned long limit = 200000000;
7147 sysctl_sched_min_granularity *= factor;
7148 if (sysctl_sched_min_granularity > limit)
7149 sysctl_sched_min_granularity = limit;
7151 sysctl_sched_latency *= factor;
7152 if (sysctl_sched_latency > limit)
7153 sysctl_sched_latency = limit;
7155 sysctl_sched_wakeup_granularity *= factor;
7157 sysctl_sched_shares_ratelimit *= factor;
7162 * This is how migration works:
7164 * 1) we queue a struct migration_req structure in the source CPU's
7165 * runqueue and wake up that CPU's migration thread.
7166 * 2) we down() the locked semaphore => thread blocks.
7167 * 3) migration thread wakes up (implicitly it forces the migrated
7168 * thread off the CPU)
7169 * 4) it gets the migration request and checks whether the migrated
7170 * task is still in the wrong runqueue.
7171 * 5) if it's in the wrong runqueue then the migration thread removes
7172 * it and puts it into the right queue.
7173 * 6) migration thread up()s the semaphore.
7174 * 7) we wake up and the migration is done.
7178 * Change a given task's CPU affinity. Migrate the thread to a
7179 * proper CPU and schedule it away if the CPU it's executing on
7180 * is removed from the allowed bitmask.
7182 * NOTE: the caller must have a valid reference to the task, the
7183 * task must not exit() & deallocate itself prematurely. The
7184 * call is not atomic; no spinlocks may be held.
7186 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7188 struct migration_req req;
7189 unsigned long flags;
7193 rq = task_rq_lock(p, &flags);
7194 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
7199 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7200 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7205 if (p->sched_class->set_cpus_allowed)
7206 p->sched_class->set_cpus_allowed(p, new_mask);
7208 cpumask_copy(&p->cpus_allowed, new_mask);
7209 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7212 /* Can the task run on the task's current CPU? If so, we're done */
7213 if (cpumask_test_cpu(task_cpu(p), new_mask))
7216 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
7217 /* Need help from migration thread: drop lock and wait. */
7218 struct task_struct *mt = rq->migration_thread;
7220 get_task_struct(mt);
7221 task_rq_unlock(rq, &flags);
7222 wake_up_process(rq->migration_thread);
7223 put_task_struct(mt);
7224 wait_for_completion(&req.done);
7225 tlb_migrate_finish(p->mm);
7229 task_rq_unlock(rq, &flags);
7233 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7236 * Move (not current) task off this cpu, onto dest cpu. We're doing
7237 * this because either it can't run here any more (set_cpus_allowed()
7238 * away from this CPU, or CPU going down), or because we're
7239 * attempting to rebalance this task on exec (sched_exec).
7241 * So we race with normal scheduler movements, but that's OK, as long
7242 * as the task is no longer on this CPU.
7244 * Returns non-zero if task was successfully migrated.
7246 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7248 struct rq *rq_dest, *rq_src;
7251 if (unlikely(!cpu_active(dest_cpu)))
7254 rq_src = cpu_rq(src_cpu);
7255 rq_dest = cpu_rq(dest_cpu);
7257 double_rq_lock(rq_src, rq_dest);
7258 /* Already moved. */
7259 if (task_cpu(p) != src_cpu)
7261 /* Affinity changed (again). */
7262 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7265 on_rq = p->se.on_rq;
7267 deactivate_task(rq_src, p, 0);
7269 set_task_cpu(p, dest_cpu);
7271 activate_task(rq_dest, p, 0);
7272 check_preempt_curr(rq_dest, p, 0);
7277 double_rq_unlock(rq_src, rq_dest);
7282 * migration_thread - this is a highprio system thread that performs
7283 * thread migration by bumping thread off CPU then 'pushing' onto
7286 static int migration_thread(void *data)
7288 int cpu = (long)data;
7292 BUG_ON(rq->migration_thread != current);
7294 set_current_state(TASK_INTERRUPTIBLE);
7295 while (!kthread_should_stop()) {
7296 struct migration_req *req;
7297 struct list_head *head;
7299 spin_lock_irq(&rq->lock);
7301 if (cpu_is_offline(cpu)) {
7302 spin_unlock_irq(&rq->lock);
7306 if (rq->active_balance) {
7307 active_load_balance(rq, cpu);
7308 rq->active_balance = 0;
7311 head = &rq->migration_queue;
7313 if (list_empty(head)) {
7314 spin_unlock_irq(&rq->lock);
7316 set_current_state(TASK_INTERRUPTIBLE);
7319 req = list_entry(head->next, struct migration_req, list);
7320 list_del_init(head->next);
7322 spin_unlock(&rq->lock);
7323 __migrate_task(req->task, cpu, req->dest_cpu);
7326 complete(&req->done);
7328 __set_current_state(TASK_RUNNING);
7333 #ifdef CONFIG_HOTPLUG_CPU
7335 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7339 local_irq_disable();
7340 ret = __migrate_task(p, src_cpu, dest_cpu);
7346 * Figure out where task on dead CPU should go, use force if necessary.
7348 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7351 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7354 /* Look for allowed, online CPU in same node. */
7355 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7356 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7359 /* Any allowed, online CPU? */
7360 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7361 if (dest_cpu < nr_cpu_ids)
7364 /* No more Mr. Nice Guy. */
7365 if (dest_cpu >= nr_cpu_ids) {
7366 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7367 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7370 * Don't tell them about moving exiting tasks or
7371 * kernel threads (both mm NULL), since they never
7374 if (p->mm && printk_ratelimit()) {
7375 printk(KERN_INFO "process %d (%s) no "
7376 "longer affine to cpu%d\n",
7377 task_pid_nr(p), p->comm, dead_cpu);
7382 /* It can have affinity changed while we were choosing. */
7383 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7388 * While a dead CPU has no uninterruptible tasks queued at this point,
7389 * it might still have a nonzero ->nr_uninterruptible counter, because
7390 * for performance reasons the counter is not stricly tracking tasks to
7391 * their home CPUs. So we just add the counter to another CPU's counter,
7392 * to keep the global sum constant after CPU-down:
7394 static void migrate_nr_uninterruptible(struct rq *rq_src)
7396 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7397 unsigned long flags;
7399 local_irq_save(flags);
7400 double_rq_lock(rq_src, rq_dest);
7401 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7402 rq_src->nr_uninterruptible = 0;
7403 double_rq_unlock(rq_src, rq_dest);
7404 local_irq_restore(flags);
7407 /* Run through task list and migrate tasks from the dead cpu. */
7408 static void migrate_live_tasks(int src_cpu)
7410 struct task_struct *p, *t;
7412 read_lock(&tasklist_lock);
7414 do_each_thread(t, p) {
7418 if (task_cpu(p) == src_cpu)
7419 move_task_off_dead_cpu(src_cpu, p);
7420 } while_each_thread(t, p);
7422 read_unlock(&tasklist_lock);
7426 * Schedules idle task to be the next runnable task on current CPU.
7427 * It does so by boosting its priority to highest possible.
7428 * Used by CPU offline code.
7430 void sched_idle_next(void)
7432 int this_cpu = smp_processor_id();
7433 struct rq *rq = cpu_rq(this_cpu);
7434 struct task_struct *p = rq->idle;
7435 unsigned long flags;
7437 /* cpu has to be offline */
7438 BUG_ON(cpu_online(this_cpu));
7441 * Strictly not necessary since rest of the CPUs are stopped by now
7442 * and interrupts disabled on the current cpu.
7444 spin_lock_irqsave(&rq->lock, flags);
7446 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7448 update_rq_clock(rq);
7449 activate_task(rq, p, 0);
7451 spin_unlock_irqrestore(&rq->lock, flags);
7455 * Ensures that the idle task is using init_mm right before its cpu goes
7458 void idle_task_exit(void)
7460 struct mm_struct *mm = current->active_mm;
7462 BUG_ON(cpu_online(smp_processor_id()));
7465 switch_mm(mm, &init_mm, current);
7469 /* called under rq->lock with disabled interrupts */
7470 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7472 struct rq *rq = cpu_rq(dead_cpu);
7474 /* Must be exiting, otherwise would be on tasklist. */
7475 BUG_ON(!p->exit_state);
7477 /* Cannot have done final schedule yet: would have vanished. */
7478 BUG_ON(p->state == TASK_DEAD);
7483 * Drop lock around migration; if someone else moves it,
7484 * that's OK. No task can be added to this CPU, so iteration is
7487 spin_unlock_irq(&rq->lock);
7488 move_task_off_dead_cpu(dead_cpu, p);
7489 spin_lock_irq(&rq->lock);
7494 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7495 static void migrate_dead_tasks(unsigned int dead_cpu)
7497 struct rq *rq = cpu_rq(dead_cpu);
7498 struct task_struct *next;
7501 if (!rq->nr_running)
7503 update_rq_clock(rq);
7504 next = pick_next_task(rq);
7507 next->sched_class->put_prev_task(rq, next);
7508 migrate_dead(dead_cpu, next);
7514 * remove the tasks which were accounted by rq from calc_load_tasks.
7516 static void calc_global_load_remove(struct rq *rq)
7518 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7519 rq->calc_load_active = 0;
7521 #endif /* CONFIG_HOTPLUG_CPU */
7523 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7525 static struct ctl_table sd_ctl_dir[] = {
7527 .procname = "sched_domain",
7533 static struct ctl_table sd_ctl_root[] = {
7535 .ctl_name = CTL_KERN,
7536 .procname = "kernel",
7538 .child = sd_ctl_dir,
7543 static struct ctl_table *sd_alloc_ctl_entry(int n)
7545 struct ctl_table *entry =
7546 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7551 static void sd_free_ctl_entry(struct ctl_table **tablep)
7553 struct ctl_table *entry;
7556 * In the intermediate directories, both the child directory and
7557 * procname are dynamically allocated and could fail but the mode
7558 * will always be set. In the lowest directory the names are
7559 * static strings and all have proc handlers.
7561 for (entry = *tablep; entry->mode; entry++) {
7563 sd_free_ctl_entry(&entry->child);
7564 if (entry->proc_handler == NULL)
7565 kfree(entry->procname);
7573 set_table_entry(struct ctl_table *entry,
7574 const char *procname, void *data, int maxlen,
7575 mode_t mode, proc_handler *proc_handler)
7577 entry->procname = procname;
7579 entry->maxlen = maxlen;
7581 entry->proc_handler = proc_handler;
7584 static struct ctl_table *
7585 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7587 struct ctl_table *table = sd_alloc_ctl_entry(13);
7592 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7593 sizeof(long), 0644, proc_doulongvec_minmax);
7594 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7595 sizeof(long), 0644, proc_doulongvec_minmax);
7596 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7597 sizeof(int), 0644, proc_dointvec_minmax);
7598 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7599 sizeof(int), 0644, proc_dointvec_minmax);
7600 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7601 sizeof(int), 0644, proc_dointvec_minmax);
7602 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7603 sizeof(int), 0644, proc_dointvec_minmax);
7604 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7605 sizeof(int), 0644, proc_dointvec_minmax);
7606 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7607 sizeof(int), 0644, proc_dointvec_minmax);
7608 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7609 sizeof(int), 0644, proc_dointvec_minmax);
7610 set_table_entry(&table[9], "cache_nice_tries",
7611 &sd->cache_nice_tries,
7612 sizeof(int), 0644, proc_dointvec_minmax);
7613 set_table_entry(&table[10], "flags", &sd->flags,
7614 sizeof(int), 0644, proc_dointvec_minmax);
7615 set_table_entry(&table[11], "name", sd->name,
7616 CORENAME_MAX_SIZE, 0444, proc_dostring);
7617 /* &table[12] is terminator */
7622 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7624 struct ctl_table *entry, *table;
7625 struct sched_domain *sd;
7626 int domain_num = 0, i;
7629 for_each_domain(cpu, sd)
7631 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7636 for_each_domain(cpu, sd) {
7637 snprintf(buf, 32, "domain%d", i);
7638 entry->procname = kstrdup(buf, GFP_KERNEL);
7640 entry->child = sd_alloc_ctl_domain_table(sd);
7647 static struct ctl_table_header *sd_sysctl_header;
7648 static void register_sched_domain_sysctl(void)
7650 int i, cpu_num = num_online_cpus();
7651 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7654 WARN_ON(sd_ctl_dir[0].child);
7655 sd_ctl_dir[0].child = entry;
7660 for_each_online_cpu(i) {
7661 snprintf(buf, 32, "cpu%d", i);
7662 entry->procname = kstrdup(buf, GFP_KERNEL);
7664 entry->child = sd_alloc_ctl_cpu_table(i);
7668 WARN_ON(sd_sysctl_header);
7669 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7672 /* may be called multiple times per register */
7673 static void unregister_sched_domain_sysctl(void)
7675 if (sd_sysctl_header)
7676 unregister_sysctl_table(sd_sysctl_header);
7677 sd_sysctl_header = NULL;
7678 if (sd_ctl_dir[0].child)
7679 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7682 static void register_sched_domain_sysctl(void)
7685 static void unregister_sched_domain_sysctl(void)
7690 static void set_rq_online(struct rq *rq)
7693 const struct sched_class *class;
7695 cpumask_set_cpu(rq->cpu, rq->rd->online);
7698 for_each_class(class) {
7699 if (class->rq_online)
7700 class->rq_online(rq);
7705 static void set_rq_offline(struct rq *rq)
7708 const struct sched_class *class;
7710 for_each_class(class) {
7711 if (class->rq_offline)
7712 class->rq_offline(rq);
7715 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7721 * migration_call - callback that gets triggered when a CPU is added.
7722 * Here we can start up the necessary migration thread for the new CPU.
7724 static int __cpuinit
7725 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7727 struct task_struct *p;
7728 int cpu = (long)hcpu;
7729 unsigned long flags;
7734 case CPU_UP_PREPARE:
7735 case CPU_UP_PREPARE_FROZEN:
7736 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7739 kthread_bind(p, cpu);
7740 /* Must be high prio: stop_machine expects to yield to it. */
7741 rq = task_rq_lock(p, &flags);
7742 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7743 task_rq_unlock(rq, &flags);
7745 cpu_rq(cpu)->migration_thread = p;
7746 rq->calc_load_update = calc_load_update;
7750 case CPU_ONLINE_FROZEN:
7751 /* Strictly unnecessary, as first user will wake it. */
7752 wake_up_process(cpu_rq(cpu)->migration_thread);
7754 /* Update our root-domain */
7756 spin_lock_irqsave(&rq->lock, flags);
7758 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7762 spin_unlock_irqrestore(&rq->lock, flags);
7765 #ifdef CONFIG_HOTPLUG_CPU
7766 case CPU_UP_CANCELED:
7767 case CPU_UP_CANCELED_FROZEN:
7768 if (!cpu_rq(cpu)->migration_thread)
7770 /* Unbind it from offline cpu so it can run. Fall thru. */
7771 kthread_bind(cpu_rq(cpu)->migration_thread,
7772 cpumask_any(cpu_online_mask));
7773 kthread_stop(cpu_rq(cpu)->migration_thread);
7774 put_task_struct(cpu_rq(cpu)->migration_thread);
7775 cpu_rq(cpu)->migration_thread = NULL;
7779 case CPU_DEAD_FROZEN:
7780 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7781 migrate_live_tasks(cpu);
7783 kthread_stop(rq->migration_thread);
7784 put_task_struct(rq->migration_thread);
7785 rq->migration_thread = NULL;
7786 /* Idle task back to normal (off runqueue, low prio) */
7787 spin_lock_irq(&rq->lock);
7788 update_rq_clock(rq);
7789 deactivate_task(rq, rq->idle, 0);
7790 rq->idle->static_prio = MAX_PRIO;
7791 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7792 rq->idle->sched_class = &idle_sched_class;
7793 migrate_dead_tasks(cpu);
7794 spin_unlock_irq(&rq->lock);
7796 migrate_nr_uninterruptible(rq);
7797 BUG_ON(rq->nr_running != 0);
7798 calc_global_load_remove(rq);
7800 * No need to migrate the tasks: it was best-effort if
7801 * they didn't take sched_hotcpu_mutex. Just wake up
7804 spin_lock_irq(&rq->lock);
7805 while (!list_empty(&rq->migration_queue)) {
7806 struct migration_req *req;
7808 req = list_entry(rq->migration_queue.next,
7809 struct migration_req, list);
7810 list_del_init(&req->list);
7811 spin_unlock_irq(&rq->lock);
7812 complete(&req->done);
7813 spin_lock_irq(&rq->lock);
7815 spin_unlock_irq(&rq->lock);
7819 case CPU_DYING_FROZEN:
7820 /* Update our root-domain */
7822 spin_lock_irqsave(&rq->lock, flags);
7824 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7827 spin_unlock_irqrestore(&rq->lock, flags);
7835 * Register at high priority so that task migration (migrate_all_tasks)
7836 * happens before everything else. This has to be lower priority than
7837 * the notifier in the perf_counter subsystem, though.
7839 static struct notifier_block __cpuinitdata migration_notifier = {
7840 .notifier_call = migration_call,
7844 static int __init migration_init(void)
7846 void *cpu = (void *)(long)smp_processor_id();
7849 /* Start one for the boot CPU: */
7850 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7851 BUG_ON(err == NOTIFY_BAD);
7852 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7853 register_cpu_notifier(&migration_notifier);
7857 early_initcall(migration_init);
7862 #ifdef CONFIG_SCHED_DEBUG
7864 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7865 struct cpumask *groupmask)
7867 struct sched_group *group = sd->groups;
7870 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7871 cpumask_clear(groupmask);
7873 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7875 if (!(sd->flags & SD_LOAD_BALANCE)) {
7876 printk("does not load-balance\n");
7878 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7883 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7885 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7886 printk(KERN_ERR "ERROR: domain->span does not contain "
7889 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7890 printk(KERN_ERR "ERROR: domain->groups does not contain"
7894 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7898 printk(KERN_ERR "ERROR: group is NULL\n");
7902 if (!group->__cpu_power) {
7903 printk(KERN_CONT "\n");
7904 printk(KERN_ERR "ERROR: domain->cpu_power not "
7909 if (!cpumask_weight(sched_group_cpus(group))) {
7910 printk(KERN_CONT "\n");
7911 printk(KERN_ERR "ERROR: empty group\n");
7915 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7916 printk(KERN_CONT "\n");
7917 printk(KERN_ERR "ERROR: repeated CPUs\n");
7921 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7923 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7925 printk(KERN_CONT " %s", str);
7926 if (group->__cpu_power != SCHED_LOAD_SCALE) {
7927 printk(KERN_CONT " (__cpu_power = %d)",
7928 group->__cpu_power);
7931 group = group->next;
7932 } while (group != sd->groups);
7933 printk(KERN_CONT "\n");
7935 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7936 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7939 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7940 printk(KERN_ERR "ERROR: parent span is not a superset "
7941 "of domain->span\n");
7945 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7947 cpumask_var_t groupmask;
7951 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7955 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7957 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7958 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7963 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7970 free_cpumask_var(groupmask);
7972 #else /* !CONFIG_SCHED_DEBUG */
7973 # define sched_domain_debug(sd, cpu) do { } while (0)
7974 #endif /* CONFIG_SCHED_DEBUG */
7976 static int sd_degenerate(struct sched_domain *sd)
7978 if (cpumask_weight(sched_domain_span(sd)) == 1)
7981 /* Following flags need at least 2 groups */
7982 if (sd->flags & (SD_LOAD_BALANCE |
7983 SD_BALANCE_NEWIDLE |
7987 SD_SHARE_PKG_RESOURCES)) {
7988 if (sd->groups != sd->groups->next)
7992 /* Following flags don't use groups */
7993 if (sd->flags & (SD_WAKE_IDLE |
8002 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
8004 unsigned long cflags = sd->flags, pflags = parent->flags;
8006 if (sd_degenerate(parent))
8009 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
8012 /* Does parent contain flags not in child? */
8013 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
8014 if (cflags & SD_WAKE_AFFINE)
8015 pflags &= ~SD_WAKE_BALANCE;
8016 /* Flags needing groups don't count if only 1 group in parent */
8017 if (parent->groups == parent->groups->next) {
8018 pflags &= ~(SD_LOAD_BALANCE |
8019 SD_BALANCE_NEWIDLE |
8023 SD_SHARE_PKG_RESOURCES);
8024 if (nr_node_ids == 1)
8025 pflags &= ~SD_SERIALIZE;
8027 if (~cflags & pflags)
8033 static void free_rootdomain(struct root_domain *rd)
8035 cpupri_cleanup(&rd->cpupri);
8037 free_cpumask_var(rd->rto_mask);
8038 free_cpumask_var(rd->online);
8039 free_cpumask_var(rd->span);
8043 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
8045 struct root_domain *old_rd = NULL;
8046 unsigned long flags;
8048 spin_lock_irqsave(&rq->lock, flags);
8053 if (cpumask_test_cpu(rq->cpu, old_rd->online))
8056 cpumask_clear_cpu(rq->cpu, old_rd->span);
8059 * If we dont want to free the old_rt yet then
8060 * set old_rd to NULL to skip the freeing later
8063 if (!atomic_dec_and_test(&old_rd->refcount))
8067 atomic_inc(&rd->refcount);
8070 cpumask_set_cpu(rq->cpu, rd->span);
8071 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
8074 spin_unlock_irqrestore(&rq->lock, flags);
8077 free_rootdomain(old_rd);
8080 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8082 gfp_t gfp = GFP_KERNEL;
8084 memset(rd, 0, sizeof(*rd));
8089 if (!alloc_cpumask_var(&rd->span, gfp))
8091 if (!alloc_cpumask_var(&rd->online, gfp))
8093 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8096 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8101 free_cpumask_var(rd->rto_mask);
8103 free_cpumask_var(rd->online);
8105 free_cpumask_var(rd->span);
8110 static void init_defrootdomain(void)
8112 init_rootdomain(&def_root_domain, true);
8114 atomic_set(&def_root_domain.refcount, 1);
8117 static struct root_domain *alloc_rootdomain(void)
8119 struct root_domain *rd;
8121 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8125 if (init_rootdomain(rd, false) != 0) {
8134 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8135 * hold the hotplug lock.
8138 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8140 struct rq *rq = cpu_rq(cpu);
8141 struct sched_domain *tmp;
8143 /* Remove the sched domains which do not contribute to scheduling. */
8144 for (tmp = sd; tmp; ) {
8145 struct sched_domain *parent = tmp->parent;
8149 if (sd_parent_degenerate(tmp, parent)) {
8150 tmp->parent = parent->parent;
8152 parent->parent->child = tmp;
8157 if (sd && sd_degenerate(sd)) {
8163 sched_domain_debug(sd, cpu);
8165 rq_attach_root(rq, rd);
8166 rcu_assign_pointer(rq->sd, sd);
8169 /* cpus with isolated domains */
8170 static cpumask_var_t cpu_isolated_map;
8172 /* Setup the mask of cpus configured for isolated domains */
8173 static int __init isolated_cpu_setup(char *str)
8175 cpulist_parse(str, cpu_isolated_map);
8179 __setup("isolcpus=", isolated_cpu_setup);
8182 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8183 * to a function which identifies what group(along with sched group) a CPU
8184 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8185 * (due to the fact that we keep track of groups covered with a struct cpumask).
8187 * init_sched_build_groups will build a circular linked list of the groups
8188 * covered by the given span, and will set each group's ->cpumask correctly,
8189 * and ->cpu_power to 0.
8192 init_sched_build_groups(const struct cpumask *span,
8193 const struct cpumask *cpu_map,
8194 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8195 struct sched_group **sg,
8196 struct cpumask *tmpmask),
8197 struct cpumask *covered, struct cpumask *tmpmask)
8199 struct sched_group *first = NULL, *last = NULL;
8202 cpumask_clear(covered);
8204 for_each_cpu(i, span) {
8205 struct sched_group *sg;
8206 int group = group_fn(i, cpu_map, &sg, tmpmask);
8209 if (cpumask_test_cpu(i, covered))
8212 cpumask_clear(sched_group_cpus(sg));
8213 sg->__cpu_power = 0;
8215 for_each_cpu(j, span) {
8216 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8219 cpumask_set_cpu(j, covered);
8220 cpumask_set_cpu(j, sched_group_cpus(sg));
8231 #define SD_NODES_PER_DOMAIN 16
8236 * find_next_best_node - find the next node to include in a sched_domain
8237 * @node: node whose sched_domain we're building
8238 * @used_nodes: nodes already in the sched_domain
8240 * Find the next node to include in a given scheduling domain. Simply
8241 * finds the closest node not already in the @used_nodes map.
8243 * Should use nodemask_t.
8245 static int find_next_best_node(int node, nodemask_t *used_nodes)
8247 int i, n, val, min_val, best_node = 0;
8251 for (i = 0; i < nr_node_ids; i++) {
8252 /* Start at @node */
8253 n = (node + i) % nr_node_ids;
8255 if (!nr_cpus_node(n))
8258 /* Skip already used nodes */
8259 if (node_isset(n, *used_nodes))
8262 /* Simple min distance search */
8263 val = node_distance(node, n);
8265 if (val < min_val) {
8271 node_set(best_node, *used_nodes);
8276 * sched_domain_node_span - get a cpumask for a node's sched_domain
8277 * @node: node whose cpumask we're constructing
8278 * @span: resulting cpumask
8280 * Given a node, construct a good cpumask for its sched_domain to span. It
8281 * should be one that prevents unnecessary balancing, but also spreads tasks
8284 static void sched_domain_node_span(int node, struct cpumask *span)
8286 nodemask_t used_nodes;
8289 cpumask_clear(span);
8290 nodes_clear(used_nodes);
8292 cpumask_or(span, span, cpumask_of_node(node));
8293 node_set(node, used_nodes);
8295 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8296 int next_node = find_next_best_node(node, &used_nodes);
8298 cpumask_or(span, span, cpumask_of_node(next_node));
8301 #endif /* CONFIG_NUMA */
8303 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8306 * The cpus mask in sched_group and sched_domain hangs off the end.
8308 * ( See the the comments in include/linux/sched.h:struct sched_group
8309 * and struct sched_domain. )
8311 struct static_sched_group {
8312 struct sched_group sg;
8313 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8316 struct static_sched_domain {
8317 struct sched_domain sd;
8318 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8324 cpumask_var_t domainspan;
8325 cpumask_var_t covered;
8326 cpumask_var_t notcovered;
8328 cpumask_var_t nodemask;
8329 cpumask_var_t this_sibling_map;
8330 cpumask_var_t this_core_map;
8331 cpumask_var_t send_covered;
8332 cpumask_var_t tmpmask;
8333 struct sched_group **sched_group_nodes;
8334 struct root_domain *rd;
8338 sa_sched_groups = 0,
8343 sa_this_sibling_map,
8345 sa_sched_group_nodes,
8355 * SMT sched-domains:
8357 #ifdef CONFIG_SCHED_SMT
8358 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8359 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8362 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8363 struct sched_group **sg, struct cpumask *unused)
8366 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8369 #endif /* CONFIG_SCHED_SMT */
8372 * multi-core sched-domains:
8374 #ifdef CONFIG_SCHED_MC
8375 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8376 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8377 #endif /* CONFIG_SCHED_MC */
8379 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8381 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8382 struct sched_group **sg, struct cpumask *mask)
8386 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8387 group = cpumask_first(mask);
8389 *sg = &per_cpu(sched_group_core, group).sg;
8392 #elif defined(CONFIG_SCHED_MC)
8394 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8395 struct sched_group **sg, struct cpumask *unused)
8398 *sg = &per_cpu(sched_group_core, cpu).sg;
8403 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8404 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8407 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8408 struct sched_group **sg, struct cpumask *mask)
8411 #ifdef CONFIG_SCHED_MC
8412 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8413 group = cpumask_first(mask);
8414 #elif defined(CONFIG_SCHED_SMT)
8415 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8416 group = cpumask_first(mask);
8421 *sg = &per_cpu(sched_group_phys, group).sg;
8427 * The init_sched_build_groups can't handle what we want to do with node
8428 * groups, so roll our own. Now each node has its own list of groups which
8429 * gets dynamically allocated.
8431 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8432 static struct sched_group ***sched_group_nodes_bycpu;
8434 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8435 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8437 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8438 struct sched_group **sg,
8439 struct cpumask *nodemask)
8443 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8444 group = cpumask_first(nodemask);
8447 *sg = &per_cpu(sched_group_allnodes, group).sg;
8451 static void init_numa_sched_groups_power(struct sched_group *group_head)
8453 struct sched_group *sg = group_head;
8459 for_each_cpu(j, sched_group_cpus(sg)) {
8460 struct sched_domain *sd;
8462 sd = &per_cpu(phys_domains, j).sd;
8463 if (j != group_first_cpu(sd->groups)) {
8465 * Only add "power" once for each
8471 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
8474 } while (sg != group_head);
8477 static int build_numa_sched_groups(struct s_data *d,
8478 const struct cpumask *cpu_map, int num)
8480 struct sched_domain *sd;
8481 struct sched_group *sg, *prev;
8484 cpumask_clear(d->covered);
8485 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8486 if (cpumask_empty(d->nodemask)) {
8487 d->sched_group_nodes[num] = NULL;
8491 sched_domain_node_span(num, d->domainspan);
8492 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8494 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8497 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8501 d->sched_group_nodes[num] = sg;
8503 for_each_cpu(j, d->nodemask) {
8504 sd = &per_cpu(node_domains, j).sd;
8508 sg->__cpu_power = 0;
8509 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8511 cpumask_or(d->covered, d->covered, d->nodemask);
8514 for (j = 0; j < nr_node_ids; j++) {
8515 n = (num + j) % nr_node_ids;
8516 cpumask_complement(d->notcovered, d->covered);
8517 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8518 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8519 if (cpumask_empty(d->tmpmask))
8521 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8522 if (cpumask_empty(d->tmpmask))
8524 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8528 "Can not alloc domain group for node %d\n", j);
8531 sg->__cpu_power = 0;
8532 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8533 sg->next = prev->next;
8534 cpumask_or(d->covered, d->covered, d->tmpmask);
8541 #endif /* CONFIG_NUMA */
8544 /* Free memory allocated for various sched_group structures */
8545 static void free_sched_groups(const struct cpumask *cpu_map,
8546 struct cpumask *nodemask)
8550 for_each_cpu(cpu, cpu_map) {
8551 struct sched_group **sched_group_nodes
8552 = sched_group_nodes_bycpu[cpu];
8554 if (!sched_group_nodes)
8557 for (i = 0; i < nr_node_ids; i++) {
8558 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8560 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8561 if (cpumask_empty(nodemask))
8571 if (oldsg != sched_group_nodes[i])
8574 kfree(sched_group_nodes);
8575 sched_group_nodes_bycpu[cpu] = NULL;
8578 #else /* !CONFIG_NUMA */
8579 static void free_sched_groups(const struct cpumask *cpu_map,
8580 struct cpumask *nodemask)
8583 #endif /* CONFIG_NUMA */
8586 * Initialize sched groups cpu_power.
8588 * cpu_power indicates the capacity of sched group, which is used while
8589 * distributing the load between different sched groups in a sched domain.
8590 * Typically cpu_power for all the groups in a sched domain will be same unless
8591 * there are asymmetries in the topology. If there are asymmetries, group
8592 * having more cpu_power will pickup more load compared to the group having
8595 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8597 struct sched_domain *child;
8598 struct sched_group *group;
8602 WARN_ON(!sd || !sd->groups);
8604 if (cpu != group_first_cpu(sd->groups))
8609 sd->groups->__cpu_power = 0;
8612 power = SCHED_LOAD_SCALE;
8613 weight = cpumask_weight(sched_domain_span(sd));
8615 * SMT siblings share the power of a single core.
8616 * Usually multiple threads get a better yield out of
8617 * that one core than a single thread would have,
8618 * reflect that in sd->smt_gain.
8620 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8621 power *= sd->smt_gain;
8623 power >>= SCHED_LOAD_SHIFT;
8625 sg_inc_cpu_power(sd->groups, power);
8630 * Add cpu_power of each child group to this groups cpu_power.
8632 group = child->groups;
8634 sg_inc_cpu_power(sd->groups, group->__cpu_power);
8635 group = group->next;
8636 } while (group != child->groups);
8640 * Initializers for schedule domains
8641 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8644 #ifdef CONFIG_SCHED_DEBUG
8645 # define SD_INIT_NAME(sd, type) sd->name = #type
8647 # define SD_INIT_NAME(sd, type) do { } while (0)
8650 #define SD_INIT(sd, type) sd_init_##type(sd)
8652 #define SD_INIT_FUNC(type) \
8653 static noinline void sd_init_##type(struct sched_domain *sd) \
8655 memset(sd, 0, sizeof(*sd)); \
8656 *sd = SD_##type##_INIT; \
8657 sd->level = SD_LV_##type; \
8658 SD_INIT_NAME(sd, type); \
8663 SD_INIT_FUNC(ALLNODES)
8666 #ifdef CONFIG_SCHED_SMT
8667 SD_INIT_FUNC(SIBLING)
8669 #ifdef CONFIG_SCHED_MC
8673 static int default_relax_domain_level = -1;
8675 static int __init setup_relax_domain_level(char *str)
8679 val = simple_strtoul(str, NULL, 0);
8680 if (val < SD_LV_MAX)
8681 default_relax_domain_level = val;
8685 __setup("relax_domain_level=", setup_relax_domain_level);
8687 static void set_domain_attribute(struct sched_domain *sd,
8688 struct sched_domain_attr *attr)
8692 if (!attr || attr->relax_domain_level < 0) {
8693 if (default_relax_domain_level < 0)
8696 request = default_relax_domain_level;
8698 request = attr->relax_domain_level;
8699 if (request < sd->level) {
8700 /* turn off idle balance on this domain */
8701 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
8703 /* turn on idle balance on this domain */
8704 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
8708 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8709 const struct cpumask *cpu_map)
8712 case sa_sched_groups:
8713 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8714 d->sched_group_nodes = NULL;
8716 free_rootdomain(d->rd); /* fall through */
8718 free_cpumask_var(d->tmpmask); /* fall through */
8719 case sa_send_covered:
8720 free_cpumask_var(d->send_covered); /* fall through */
8721 case sa_this_core_map:
8722 free_cpumask_var(d->this_core_map); /* fall through */
8723 case sa_this_sibling_map:
8724 free_cpumask_var(d->this_sibling_map); /* fall through */
8726 free_cpumask_var(d->nodemask); /* fall through */
8727 case sa_sched_group_nodes:
8729 kfree(d->sched_group_nodes); /* fall through */
8731 free_cpumask_var(d->notcovered); /* fall through */
8733 free_cpumask_var(d->covered); /* fall through */
8735 free_cpumask_var(d->domainspan); /* fall through */
8742 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8743 const struct cpumask *cpu_map)
8746 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8748 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8749 return sa_domainspan;
8750 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8752 /* Allocate the per-node list of sched groups */
8753 d->sched_group_nodes = kcalloc(nr_node_ids,
8754 sizeof(struct sched_group *), GFP_KERNEL);
8755 if (!d->sched_group_nodes) {
8756 printk(KERN_WARNING "Can not alloc sched group node list\n");
8757 return sa_notcovered;
8759 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8761 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8762 return sa_sched_group_nodes;
8763 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8765 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8766 return sa_this_sibling_map;
8767 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8768 return sa_this_core_map;
8769 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8770 return sa_send_covered;
8771 d->rd = alloc_rootdomain();
8773 printk(KERN_WARNING "Cannot alloc root domain\n");
8776 return sa_rootdomain;
8779 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8780 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8782 struct sched_domain *sd = NULL;
8784 struct sched_domain *parent;
8787 if (cpumask_weight(cpu_map) >
8788 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8789 sd = &per_cpu(allnodes_domains, i).sd;
8790 SD_INIT(sd, ALLNODES);
8791 set_domain_attribute(sd, attr);
8792 cpumask_copy(sched_domain_span(sd), cpu_map);
8793 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8798 sd = &per_cpu(node_domains, i).sd;
8800 set_domain_attribute(sd, attr);
8801 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8802 sd->parent = parent;
8805 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8810 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8811 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8812 struct sched_domain *parent, int i)
8814 struct sched_domain *sd;
8815 sd = &per_cpu(phys_domains, i).sd;
8817 set_domain_attribute(sd, attr);
8818 cpumask_copy(sched_domain_span(sd), d->nodemask);
8819 sd->parent = parent;
8822 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8826 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8827 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8828 struct sched_domain *parent, int i)
8830 struct sched_domain *sd = parent;
8831 #ifdef CONFIG_SCHED_MC
8832 sd = &per_cpu(core_domains, i).sd;
8834 set_domain_attribute(sd, attr);
8835 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8836 sd->parent = parent;
8838 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8843 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8844 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8845 struct sched_domain *parent, int i)
8847 struct sched_domain *sd = parent;
8848 #ifdef CONFIG_SCHED_SMT
8849 sd = &per_cpu(cpu_domains, i).sd;
8850 SD_INIT(sd, SIBLING);
8851 set_domain_attribute(sd, attr);
8852 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8853 sd->parent = parent;
8855 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8860 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8861 const struct cpumask *cpu_map, int cpu)
8864 #ifdef CONFIG_SCHED_SMT
8865 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8866 cpumask_and(d->this_sibling_map, cpu_map,
8867 topology_thread_cpumask(cpu));
8868 if (cpu == cpumask_first(d->this_sibling_map))
8869 init_sched_build_groups(d->this_sibling_map, cpu_map,
8871 d->send_covered, d->tmpmask);
8874 #ifdef CONFIG_SCHED_MC
8875 case SD_LV_MC: /* set up multi-core groups */
8876 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8877 if (cpu == cpumask_first(d->this_core_map))
8878 init_sched_build_groups(d->this_core_map, cpu_map,
8880 d->send_covered, d->tmpmask);
8883 case SD_LV_CPU: /* set up physical groups */
8884 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8885 if (!cpumask_empty(d->nodemask))
8886 init_sched_build_groups(d->nodemask, cpu_map,
8888 d->send_covered, d->tmpmask);
8891 case SD_LV_ALLNODES:
8892 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8893 d->send_covered, d->tmpmask);
8902 * Build sched domains for a given set of cpus and attach the sched domains
8903 * to the individual cpus
8905 static int __build_sched_domains(const struct cpumask *cpu_map,
8906 struct sched_domain_attr *attr)
8908 enum s_alloc alloc_state = sa_none;
8910 struct sched_domain *sd;
8916 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8917 if (alloc_state != sa_rootdomain)
8919 alloc_state = sa_sched_groups;
8922 * Set up domains for cpus specified by the cpu_map.
8924 for_each_cpu(i, cpu_map) {
8925 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8928 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8929 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8930 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8931 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8934 for_each_cpu(i, cpu_map) {
8935 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8936 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8939 /* Set up physical groups */
8940 for (i = 0; i < nr_node_ids; i++)
8941 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8944 /* Set up node groups */
8946 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8948 for (i = 0; i < nr_node_ids; i++)
8949 if (build_numa_sched_groups(&d, cpu_map, i))
8953 /* Calculate CPU power for physical packages and nodes */
8954 #ifdef CONFIG_SCHED_SMT
8955 for_each_cpu(i, cpu_map) {
8956 sd = &per_cpu(cpu_domains, i).sd;
8957 init_sched_groups_power(i, sd);
8960 #ifdef CONFIG_SCHED_MC
8961 for_each_cpu(i, cpu_map) {
8962 sd = &per_cpu(core_domains, i).sd;
8963 init_sched_groups_power(i, sd);
8967 for_each_cpu(i, cpu_map) {
8968 sd = &per_cpu(phys_domains, i).sd;
8969 init_sched_groups_power(i, sd);
8973 for (i = 0; i < nr_node_ids; i++)
8974 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8976 if (d.sd_allnodes) {
8977 struct sched_group *sg;
8979 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8981 init_numa_sched_groups_power(sg);
8985 /* Attach the domains */
8986 for_each_cpu(i, cpu_map) {
8987 #ifdef CONFIG_SCHED_SMT
8988 sd = &per_cpu(cpu_domains, i).sd;
8989 #elif defined(CONFIG_SCHED_MC)
8990 sd = &per_cpu(core_domains, i).sd;
8992 sd = &per_cpu(phys_domains, i).sd;
8994 cpu_attach_domain(sd, d.rd, i);
8997 d.sched_group_nodes = NULL; /* don't free this we still need it */
8998 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
9002 __free_domain_allocs(&d, alloc_state, cpu_map);
9006 static int build_sched_domains(const struct cpumask *cpu_map)
9008 return __build_sched_domains(cpu_map, NULL);
9011 static struct cpumask *doms_cur; /* current sched domains */
9012 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
9013 static struct sched_domain_attr *dattr_cur;
9014 /* attribues of custom domains in 'doms_cur' */
9017 * Special case: If a kmalloc of a doms_cur partition (array of
9018 * cpumask) fails, then fallback to a single sched domain,
9019 * as determined by the single cpumask fallback_doms.
9021 static cpumask_var_t fallback_doms;
9024 * arch_update_cpu_topology lets virtualized architectures update the
9025 * cpu core maps. It is supposed to return 1 if the topology changed
9026 * or 0 if it stayed the same.
9028 int __attribute__((weak)) arch_update_cpu_topology(void)
9034 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9035 * For now this just excludes isolated cpus, but could be used to
9036 * exclude other special cases in the future.
9038 static int arch_init_sched_domains(const struct cpumask *cpu_map)
9042 arch_update_cpu_topology();
9044 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
9046 doms_cur = fallback_doms;
9047 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
9049 err = build_sched_domains(doms_cur);
9050 register_sched_domain_sysctl();
9055 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9056 struct cpumask *tmpmask)
9058 free_sched_groups(cpu_map, tmpmask);
9062 * Detach sched domains from a group of cpus specified in cpu_map
9063 * These cpus will now be attached to the NULL domain
9065 static void detach_destroy_domains(const struct cpumask *cpu_map)
9067 /* Save because hotplug lock held. */
9068 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9071 for_each_cpu(i, cpu_map)
9072 cpu_attach_domain(NULL, &def_root_domain, i);
9073 synchronize_sched();
9074 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9077 /* handle null as "default" */
9078 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9079 struct sched_domain_attr *new, int idx_new)
9081 struct sched_domain_attr tmp;
9088 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9089 new ? (new + idx_new) : &tmp,
9090 sizeof(struct sched_domain_attr));
9094 * Partition sched domains as specified by the 'ndoms_new'
9095 * cpumasks in the array doms_new[] of cpumasks. This compares
9096 * doms_new[] to the current sched domain partitioning, doms_cur[].
9097 * It destroys each deleted domain and builds each new domain.
9099 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
9100 * The masks don't intersect (don't overlap.) We should setup one
9101 * sched domain for each mask. CPUs not in any of the cpumasks will
9102 * not be load balanced. If the same cpumask appears both in the
9103 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9106 * The passed in 'doms_new' should be kmalloc'd. This routine takes
9107 * ownership of it and will kfree it when done with it. If the caller
9108 * failed the kmalloc call, then it can pass in doms_new == NULL &&
9109 * ndoms_new == 1, and partition_sched_domains() will fallback to
9110 * the single partition 'fallback_doms', it also forces the domains
9113 * If doms_new == NULL it will be replaced with cpu_online_mask.
9114 * ndoms_new == 0 is a special case for destroying existing domains,
9115 * and it will not create the default domain.
9117 * Call with hotplug lock held
9119 /* FIXME: Change to struct cpumask *doms_new[] */
9120 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
9121 struct sched_domain_attr *dattr_new)
9126 mutex_lock(&sched_domains_mutex);
9128 /* always unregister in case we don't destroy any domains */
9129 unregister_sched_domain_sysctl();
9131 /* Let architecture update cpu core mappings. */
9132 new_topology = arch_update_cpu_topology();
9134 n = doms_new ? ndoms_new : 0;
9136 /* Destroy deleted domains */
9137 for (i = 0; i < ndoms_cur; i++) {
9138 for (j = 0; j < n && !new_topology; j++) {
9139 if (cpumask_equal(&doms_cur[i], &doms_new[j])
9140 && dattrs_equal(dattr_cur, i, dattr_new, j))
9143 /* no match - a current sched domain not in new doms_new[] */
9144 detach_destroy_domains(doms_cur + i);
9149 if (doms_new == NULL) {
9151 doms_new = fallback_doms;
9152 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
9153 WARN_ON_ONCE(dattr_new);
9156 /* Build new domains */
9157 for (i = 0; i < ndoms_new; i++) {
9158 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9159 if (cpumask_equal(&doms_new[i], &doms_cur[j])
9160 && dattrs_equal(dattr_new, i, dattr_cur, j))
9163 /* no match - add a new doms_new */
9164 __build_sched_domains(doms_new + i,
9165 dattr_new ? dattr_new + i : NULL);
9170 /* Remember the new sched domains */
9171 if (doms_cur != fallback_doms)
9173 kfree(dattr_cur); /* kfree(NULL) is safe */
9174 doms_cur = doms_new;
9175 dattr_cur = dattr_new;
9176 ndoms_cur = ndoms_new;
9178 register_sched_domain_sysctl();
9180 mutex_unlock(&sched_domains_mutex);
9183 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9184 static void arch_reinit_sched_domains(void)
9188 /* Destroy domains first to force the rebuild */
9189 partition_sched_domains(0, NULL, NULL);
9191 rebuild_sched_domains();
9195 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9197 unsigned int level = 0;
9199 if (sscanf(buf, "%u", &level) != 1)
9203 * level is always be positive so don't check for
9204 * level < POWERSAVINGS_BALANCE_NONE which is 0
9205 * What happens on 0 or 1 byte write,
9206 * need to check for count as well?
9209 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9213 sched_smt_power_savings = level;
9215 sched_mc_power_savings = level;
9217 arch_reinit_sched_domains();
9222 #ifdef CONFIG_SCHED_MC
9223 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9226 return sprintf(page, "%u\n", sched_mc_power_savings);
9228 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9229 const char *buf, size_t count)
9231 return sched_power_savings_store(buf, count, 0);
9233 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9234 sched_mc_power_savings_show,
9235 sched_mc_power_savings_store);
9238 #ifdef CONFIG_SCHED_SMT
9239 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9242 return sprintf(page, "%u\n", sched_smt_power_savings);
9244 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9245 const char *buf, size_t count)
9247 return sched_power_savings_store(buf, count, 1);
9249 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9250 sched_smt_power_savings_show,
9251 sched_smt_power_savings_store);
9254 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9258 #ifdef CONFIG_SCHED_SMT
9260 err = sysfs_create_file(&cls->kset.kobj,
9261 &attr_sched_smt_power_savings.attr);
9263 #ifdef CONFIG_SCHED_MC
9264 if (!err && mc_capable())
9265 err = sysfs_create_file(&cls->kset.kobj,
9266 &attr_sched_mc_power_savings.attr);
9270 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9272 #ifndef CONFIG_CPUSETS
9274 * Add online and remove offline CPUs from the scheduler domains.
9275 * When cpusets are enabled they take over this function.
9277 static int update_sched_domains(struct notifier_block *nfb,
9278 unsigned long action, void *hcpu)
9282 case CPU_ONLINE_FROZEN:
9284 case CPU_DEAD_FROZEN:
9285 partition_sched_domains(1, NULL, NULL);
9294 static int update_runtime(struct notifier_block *nfb,
9295 unsigned long action, void *hcpu)
9297 int cpu = (int)(long)hcpu;
9300 case CPU_DOWN_PREPARE:
9301 case CPU_DOWN_PREPARE_FROZEN:
9302 disable_runtime(cpu_rq(cpu));
9305 case CPU_DOWN_FAILED:
9306 case CPU_DOWN_FAILED_FROZEN:
9308 case CPU_ONLINE_FROZEN:
9309 enable_runtime(cpu_rq(cpu));
9317 void __init sched_init_smp(void)
9319 cpumask_var_t non_isolated_cpus;
9321 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9323 #if defined(CONFIG_NUMA)
9324 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9326 BUG_ON(sched_group_nodes_bycpu == NULL);
9329 mutex_lock(&sched_domains_mutex);
9330 arch_init_sched_domains(cpu_online_mask);
9331 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9332 if (cpumask_empty(non_isolated_cpus))
9333 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9334 mutex_unlock(&sched_domains_mutex);
9337 #ifndef CONFIG_CPUSETS
9338 /* XXX: Theoretical race here - CPU may be hotplugged now */
9339 hotcpu_notifier(update_sched_domains, 0);
9342 /* RT runtime code needs to handle some hotplug events */
9343 hotcpu_notifier(update_runtime, 0);
9347 /* Move init over to a non-isolated CPU */
9348 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9350 sched_init_granularity();
9351 free_cpumask_var(non_isolated_cpus);
9353 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9354 init_sched_rt_class();
9357 void __init sched_init_smp(void)
9359 sched_init_granularity();
9361 #endif /* CONFIG_SMP */
9363 const_debug unsigned int sysctl_timer_migration = 1;
9365 int in_sched_functions(unsigned long addr)
9367 return in_lock_functions(addr) ||
9368 (addr >= (unsigned long)__sched_text_start
9369 && addr < (unsigned long)__sched_text_end);
9372 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9374 cfs_rq->tasks_timeline = RB_ROOT;
9375 INIT_LIST_HEAD(&cfs_rq->tasks);
9376 #ifdef CONFIG_FAIR_GROUP_SCHED
9379 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9382 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9384 struct rt_prio_array *array;
9387 array = &rt_rq->active;
9388 for (i = 0; i < MAX_RT_PRIO; i++) {
9389 INIT_LIST_HEAD(array->queue + i);
9390 __clear_bit(i, array->bitmap);
9392 /* delimiter for bitsearch: */
9393 __set_bit(MAX_RT_PRIO, array->bitmap);
9395 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9396 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9398 rt_rq->highest_prio.next = MAX_RT_PRIO;
9402 rt_rq->rt_nr_migratory = 0;
9403 rt_rq->overloaded = 0;
9404 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9408 rt_rq->rt_throttled = 0;
9409 rt_rq->rt_runtime = 0;
9410 spin_lock_init(&rt_rq->rt_runtime_lock);
9412 #ifdef CONFIG_RT_GROUP_SCHED
9413 rt_rq->rt_nr_boosted = 0;
9418 #ifdef CONFIG_FAIR_GROUP_SCHED
9419 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9420 struct sched_entity *se, int cpu, int add,
9421 struct sched_entity *parent)
9423 struct rq *rq = cpu_rq(cpu);
9424 tg->cfs_rq[cpu] = cfs_rq;
9425 init_cfs_rq(cfs_rq, rq);
9428 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9431 /* se could be NULL for init_task_group */
9436 se->cfs_rq = &rq->cfs;
9438 se->cfs_rq = parent->my_q;
9441 se->load.weight = tg->shares;
9442 se->load.inv_weight = 0;
9443 se->parent = parent;
9447 #ifdef CONFIG_RT_GROUP_SCHED
9448 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9449 struct sched_rt_entity *rt_se, int cpu, int add,
9450 struct sched_rt_entity *parent)
9452 struct rq *rq = cpu_rq(cpu);
9454 tg->rt_rq[cpu] = rt_rq;
9455 init_rt_rq(rt_rq, rq);
9457 rt_rq->rt_se = rt_se;
9458 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9460 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9462 tg->rt_se[cpu] = rt_se;
9467 rt_se->rt_rq = &rq->rt;
9469 rt_se->rt_rq = parent->my_q;
9471 rt_se->my_q = rt_rq;
9472 rt_se->parent = parent;
9473 INIT_LIST_HEAD(&rt_se->run_list);
9477 void __init sched_init(void)
9480 unsigned long alloc_size = 0, ptr;
9482 #ifdef CONFIG_FAIR_GROUP_SCHED
9483 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9485 #ifdef CONFIG_RT_GROUP_SCHED
9486 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9488 #ifdef CONFIG_USER_SCHED
9491 #ifdef CONFIG_CPUMASK_OFFSTACK
9492 alloc_size += num_possible_cpus() * cpumask_size();
9495 * As sched_init() is called before page_alloc is setup,
9496 * we use alloc_bootmem().
9499 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9501 #ifdef CONFIG_FAIR_GROUP_SCHED
9502 init_task_group.se = (struct sched_entity **)ptr;
9503 ptr += nr_cpu_ids * sizeof(void **);
9505 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9506 ptr += nr_cpu_ids * sizeof(void **);
9508 #ifdef CONFIG_USER_SCHED
9509 root_task_group.se = (struct sched_entity **)ptr;
9510 ptr += nr_cpu_ids * sizeof(void **);
9512 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9513 ptr += nr_cpu_ids * sizeof(void **);
9514 #endif /* CONFIG_USER_SCHED */
9515 #endif /* CONFIG_FAIR_GROUP_SCHED */
9516 #ifdef CONFIG_RT_GROUP_SCHED
9517 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9518 ptr += nr_cpu_ids * sizeof(void **);
9520 init_task_group.rt_rq = (struct rt_rq **)ptr;
9521 ptr += nr_cpu_ids * sizeof(void **);
9523 #ifdef CONFIG_USER_SCHED
9524 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9525 ptr += nr_cpu_ids * sizeof(void **);
9527 root_task_group.rt_rq = (struct rt_rq **)ptr;
9528 ptr += nr_cpu_ids * sizeof(void **);
9529 #endif /* CONFIG_USER_SCHED */
9530 #endif /* CONFIG_RT_GROUP_SCHED */
9531 #ifdef CONFIG_CPUMASK_OFFSTACK
9532 for_each_possible_cpu(i) {
9533 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9534 ptr += cpumask_size();
9536 #endif /* CONFIG_CPUMASK_OFFSTACK */
9540 init_defrootdomain();
9543 init_rt_bandwidth(&def_rt_bandwidth,
9544 global_rt_period(), global_rt_runtime());
9546 #ifdef CONFIG_RT_GROUP_SCHED
9547 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9548 global_rt_period(), global_rt_runtime());
9549 #ifdef CONFIG_USER_SCHED
9550 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9551 global_rt_period(), RUNTIME_INF);
9552 #endif /* CONFIG_USER_SCHED */
9553 #endif /* CONFIG_RT_GROUP_SCHED */
9555 #ifdef CONFIG_GROUP_SCHED
9556 list_add(&init_task_group.list, &task_groups);
9557 INIT_LIST_HEAD(&init_task_group.children);
9559 #ifdef CONFIG_USER_SCHED
9560 INIT_LIST_HEAD(&root_task_group.children);
9561 init_task_group.parent = &root_task_group;
9562 list_add(&init_task_group.siblings, &root_task_group.children);
9563 #endif /* CONFIG_USER_SCHED */
9564 #endif /* CONFIG_GROUP_SCHED */
9566 for_each_possible_cpu(i) {
9570 spin_lock_init(&rq->lock);
9572 rq->calc_load_active = 0;
9573 rq->calc_load_update = jiffies + LOAD_FREQ;
9574 init_cfs_rq(&rq->cfs, rq);
9575 init_rt_rq(&rq->rt, rq);
9576 #ifdef CONFIG_FAIR_GROUP_SCHED
9577 init_task_group.shares = init_task_group_load;
9578 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9579 #ifdef CONFIG_CGROUP_SCHED
9581 * How much cpu bandwidth does init_task_group get?
9583 * In case of task-groups formed thr' the cgroup filesystem, it
9584 * gets 100% of the cpu resources in the system. This overall
9585 * system cpu resource is divided among the tasks of
9586 * init_task_group and its child task-groups in a fair manner,
9587 * based on each entity's (task or task-group's) weight
9588 * (se->load.weight).
9590 * In other words, if init_task_group has 10 tasks of weight
9591 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9592 * then A0's share of the cpu resource is:
9594 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9596 * We achieve this by letting init_task_group's tasks sit
9597 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9599 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9600 #elif defined CONFIG_USER_SCHED
9601 root_task_group.shares = NICE_0_LOAD;
9602 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9604 * In case of task-groups formed thr' the user id of tasks,
9605 * init_task_group represents tasks belonging to root user.
9606 * Hence it forms a sibling of all subsequent groups formed.
9607 * In this case, init_task_group gets only a fraction of overall
9608 * system cpu resource, based on the weight assigned to root
9609 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9610 * by letting tasks of init_task_group sit in a separate cfs_rq
9611 * (init_tg_cfs_rq) and having one entity represent this group of
9612 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9614 init_tg_cfs_entry(&init_task_group,
9615 &per_cpu(init_tg_cfs_rq, i),
9616 &per_cpu(init_sched_entity, i), i, 1,
9617 root_task_group.se[i]);
9620 #endif /* CONFIG_FAIR_GROUP_SCHED */
9622 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9623 #ifdef CONFIG_RT_GROUP_SCHED
9624 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9625 #ifdef CONFIG_CGROUP_SCHED
9626 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9627 #elif defined CONFIG_USER_SCHED
9628 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9629 init_tg_rt_entry(&init_task_group,
9630 &per_cpu(init_rt_rq, i),
9631 &per_cpu(init_sched_rt_entity, i), i, 1,
9632 root_task_group.rt_se[i]);
9636 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9637 rq->cpu_load[j] = 0;
9641 rq->post_schedule = 0;
9642 rq->active_balance = 0;
9643 rq->next_balance = jiffies;
9647 rq->migration_thread = NULL;
9648 INIT_LIST_HEAD(&rq->migration_queue);
9649 rq_attach_root(rq, &def_root_domain);
9652 atomic_set(&rq->nr_iowait, 0);
9655 set_load_weight(&init_task);
9657 #ifdef CONFIG_PREEMPT_NOTIFIERS
9658 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9662 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9665 #ifdef CONFIG_RT_MUTEXES
9666 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9670 * The boot idle thread does lazy MMU switching as well:
9672 atomic_inc(&init_mm.mm_count);
9673 enter_lazy_tlb(&init_mm, current);
9676 * Make us the idle thread. Technically, schedule() should not be
9677 * called from this thread, however somewhere below it might be,
9678 * but because we are the idle thread, we just pick up running again
9679 * when this runqueue becomes "idle".
9681 init_idle(current, smp_processor_id());
9683 calc_load_update = jiffies + LOAD_FREQ;
9686 * During early bootup we pretend to be a normal task:
9688 current->sched_class = &fair_sched_class;
9690 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9691 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9694 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9695 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9697 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9700 perf_counter_init();
9702 scheduler_running = 1;
9705 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9706 static inline int preempt_count_equals(int preempt_offset)
9708 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9710 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9713 void __might_sleep(char *file, int line, int preempt_offset)
9716 static unsigned long prev_jiffy; /* ratelimiting */
9718 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9719 system_state != SYSTEM_RUNNING || oops_in_progress)
9721 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9723 prev_jiffy = jiffies;
9726 "BUG: sleeping function called from invalid context at %s:%d\n",
9729 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9730 in_atomic(), irqs_disabled(),
9731 current->pid, current->comm);
9733 debug_show_held_locks(current);
9734 if (irqs_disabled())
9735 print_irqtrace_events(current);
9739 EXPORT_SYMBOL(__might_sleep);
9742 #ifdef CONFIG_MAGIC_SYSRQ
9743 static void normalize_task(struct rq *rq, struct task_struct *p)
9747 update_rq_clock(rq);
9748 on_rq = p->se.on_rq;
9750 deactivate_task(rq, p, 0);
9751 __setscheduler(rq, p, SCHED_NORMAL, 0);
9753 activate_task(rq, p, 0);
9754 resched_task(rq->curr);
9758 void normalize_rt_tasks(void)
9760 struct task_struct *g, *p;
9761 unsigned long flags;
9764 read_lock_irqsave(&tasklist_lock, flags);
9765 do_each_thread(g, p) {
9767 * Only normalize user tasks:
9772 p->se.exec_start = 0;
9773 #ifdef CONFIG_SCHEDSTATS
9774 p->se.wait_start = 0;
9775 p->se.sleep_start = 0;
9776 p->se.block_start = 0;
9781 * Renice negative nice level userspace
9784 if (TASK_NICE(p) < 0 && p->mm)
9785 set_user_nice(p, 0);
9789 spin_lock(&p->pi_lock);
9790 rq = __task_rq_lock(p);
9792 normalize_task(rq, p);
9794 __task_rq_unlock(rq);
9795 spin_unlock(&p->pi_lock);
9796 } while_each_thread(g, p);
9798 read_unlock_irqrestore(&tasklist_lock, flags);
9801 #endif /* CONFIG_MAGIC_SYSRQ */
9805 * These functions are only useful for the IA64 MCA handling.
9807 * They can only be called when the whole system has been
9808 * stopped - every CPU needs to be quiescent, and no scheduling
9809 * activity can take place. Using them for anything else would
9810 * be a serious bug, and as a result, they aren't even visible
9811 * under any other configuration.
9815 * curr_task - return the current task for a given cpu.
9816 * @cpu: the processor in question.
9818 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9820 struct task_struct *curr_task(int cpu)
9822 return cpu_curr(cpu);
9826 * set_curr_task - set the current task for a given cpu.
9827 * @cpu: the processor in question.
9828 * @p: the task pointer to set.
9830 * Description: This function must only be used when non-maskable interrupts
9831 * are serviced on a separate stack. It allows the architecture to switch the
9832 * notion of the current task on a cpu in a non-blocking manner. This function
9833 * must be called with all CPU's synchronized, and interrupts disabled, the
9834 * and caller must save the original value of the current task (see
9835 * curr_task() above) and restore that value before reenabling interrupts and
9836 * re-starting the system.
9838 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9840 void set_curr_task(int cpu, struct task_struct *p)
9847 #ifdef CONFIG_FAIR_GROUP_SCHED
9848 static void free_fair_sched_group(struct task_group *tg)
9852 for_each_possible_cpu(i) {
9854 kfree(tg->cfs_rq[i]);
9864 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9866 struct cfs_rq *cfs_rq;
9867 struct sched_entity *se;
9871 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9874 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9878 tg->shares = NICE_0_LOAD;
9880 for_each_possible_cpu(i) {
9883 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9884 GFP_KERNEL, cpu_to_node(i));
9888 se = kzalloc_node(sizeof(struct sched_entity),
9889 GFP_KERNEL, cpu_to_node(i));
9893 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9902 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9904 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9905 &cpu_rq(cpu)->leaf_cfs_rq_list);
9908 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9910 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9912 #else /* !CONFG_FAIR_GROUP_SCHED */
9913 static inline void free_fair_sched_group(struct task_group *tg)
9918 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9923 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9927 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9930 #endif /* CONFIG_FAIR_GROUP_SCHED */
9932 #ifdef CONFIG_RT_GROUP_SCHED
9933 static void free_rt_sched_group(struct task_group *tg)
9937 destroy_rt_bandwidth(&tg->rt_bandwidth);
9939 for_each_possible_cpu(i) {
9941 kfree(tg->rt_rq[i]);
9943 kfree(tg->rt_se[i]);
9951 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9953 struct rt_rq *rt_rq;
9954 struct sched_rt_entity *rt_se;
9958 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9961 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9965 init_rt_bandwidth(&tg->rt_bandwidth,
9966 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9968 for_each_possible_cpu(i) {
9971 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9972 GFP_KERNEL, cpu_to_node(i));
9976 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9977 GFP_KERNEL, cpu_to_node(i));
9981 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9990 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9992 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9993 &cpu_rq(cpu)->leaf_rt_rq_list);
9996 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9998 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
10000 #else /* !CONFIG_RT_GROUP_SCHED */
10001 static inline void free_rt_sched_group(struct task_group *tg)
10006 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
10011 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10015 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10018 #endif /* CONFIG_RT_GROUP_SCHED */
10020 #ifdef CONFIG_GROUP_SCHED
10021 static void free_sched_group(struct task_group *tg)
10023 free_fair_sched_group(tg);
10024 free_rt_sched_group(tg);
10028 /* allocate runqueue etc for a new task group */
10029 struct task_group *sched_create_group(struct task_group *parent)
10031 struct task_group *tg;
10032 unsigned long flags;
10035 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
10037 return ERR_PTR(-ENOMEM);
10039 if (!alloc_fair_sched_group(tg, parent))
10042 if (!alloc_rt_sched_group(tg, parent))
10045 spin_lock_irqsave(&task_group_lock, flags);
10046 for_each_possible_cpu(i) {
10047 register_fair_sched_group(tg, i);
10048 register_rt_sched_group(tg, i);
10050 list_add_rcu(&tg->list, &task_groups);
10052 WARN_ON(!parent); /* root should already exist */
10054 tg->parent = parent;
10055 INIT_LIST_HEAD(&tg->children);
10056 list_add_rcu(&tg->siblings, &parent->children);
10057 spin_unlock_irqrestore(&task_group_lock, flags);
10062 free_sched_group(tg);
10063 return ERR_PTR(-ENOMEM);
10066 /* rcu callback to free various structures associated with a task group */
10067 static void free_sched_group_rcu(struct rcu_head *rhp)
10069 /* now it should be safe to free those cfs_rqs */
10070 free_sched_group(container_of(rhp, struct task_group, rcu));
10073 /* Destroy runqueue etc associated with a task group */
10074 void sched_destroy_group(struct task_group *tg)
10076 unsigned long flags;
10079 spin_lock_irqsave(&task_group_lock, flags);
10080 for_each_possible_cpu(i) {
10081 unregister_fair_sched_group(tg, i);
10082 unregister_rt_sched_group(tg, i);
10084 list_del_rcu(&tg->list);
10085 list_del_rcu(&tg->siblings);
10086 spin_unlock_irqrestore(&task_group_lock, flags);
10088 /* wait for possible concurrent references to cfs_rqs complete */
10089 call_rcu(&tg->rcu, free_sched_group_rcu);
10092 /* change task's runqueue when it moves between groups.
10093 * The caller of this function should have put the task in its new group
10094 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10095 * reflect its new group.
10097 void sched_move_task(struct task_struct *tsk)
10099 int on_rq, running;
10100 unsigned long flags;
10103 rq = task_rq_lock(tsk, &flags);
10105 update_rq_clock(rq);
10107 running = task_current(rq, tsk);
10108 on_rq = tsk->se.on_rq;
10111 dequeue_task(rq, tsk, 0);
10112 if (unlikely(running))
10113 tsk->sched_class->put_prev_task(rq, tsk);
10115 set_task_rq(tsk, task_cpu(tsk));
10117 #ifdef CONFIG_FAIR_GROUP_SCHED
10118 if (tsk->sched_class->moved_group)
10119 tsk->sched_class->moved_group(tsk);
10122 if (unlikely(running))
10123 tsk->sched_class->set_curr_task(rq);
10125 enqueue_task(rq, tsk, 0);
10127 task_rq_unlock(rq, &flags);
10129 #endif /* CONFIG_GROUP_SCHED */
10131 #ifdef CONFIG_FAIR_GROUP_SCHED
10132 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10134 struct cfs_rq *cfs_rq = se->cfs_rq;
10139 dequeue_entity(cfs_rq, se, 0);
10141 se->load.weight = shares;
10142 se->load.inv_weight = 0;
10145 enqueue_entity(cfs_rq, se, 0);
10148 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10150 struct cfs_rq *cfs_rq = se->cfs_rq;
10151 struct rq *rq = cfs_rq->rq;
10152 unsigned long flags;
10154 spin_lock_irqsave(&rq->lock, flags);
10155 __set_se_shares(se, shares);
10156 spin_unlock_irqrestore(&rq->lock, flags);
10159 static DEFINE_MUTEX(shares_mutex);
10161 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10164 unsigned long flags;
10167 * We can't change the weight of the root cgroup.
10172 if (shares < MIN_SHARES)
10173 shares = MIN_SHARES;
10174 else if (shares > MAX_SHARES)
10175 shares = MAX_SHARES;
10177 mutex_lock(&shares_mutex);
10178 if (tg->shares == shares)
10181 spin_lock_irqsave(&task_group_lock, flags);
10182 for_each_possible_cpu(i)
10183 unregister_fair_sched_group(tg, i);
10184 list_del_rcu(&tg->siblings);
10185 spin_unlock_irqrestore(&task_group_lock, flags);
10187 /* wait for any ongoing reference to this group to finish */
10188 synchronize_sched();
10191 * Now we are free to modify the group's share on each cpu
10192 * w/o tripping rebalance_share or load_balance_fair.
10194 tg->shares = shares;
10195 for_each_possible_cpu(i) {
10197 * force a rebalance
10199 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10200 set_se_shares(tg->se[i], shares);
10204 * Enable load balance activity on this group, by inserting it back on
10205 * each cpu's rq->leaf_cfs_rq_list.
10207 spin_lock_irqsave(&task_group_lock, flags);
10208 for_each_possible_cpu(i)
10209 register_fair_sched_group(tg, i);
10210 list_add_rcu(&tg->siblings, &tg->parent->children);
10211 spin_unlock_irqrestore(&task_group_lock, flags);
10213 mutex_unlock(&shares_mutex);
10217 unsigned long sched_group_shares(struct task_group *tg)
10223 #ifdef CONFIG_RT_GROUP_SCHED
10225 * Ensure that the real time constraints are schedulable.
10227 static DEFINE_MUTEX(rt_constraints_mutex);
10229 static unsigned long to_ratio(u64 period, u64 runtime)
10231 if (runtime == RUNTIME_INF)
10234 return div64_u64(runtime << 20, period);
10237 /* Must be called with tasklist_lock held */
10238 static inline int tg_has_rt_tasks(struct task_group *tg)
10240 struct task_struct *g, *p;
10242 do_each_thread(g, p) {
10243 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10245 } while_each_thread(g, p);
10250 struct rt_schedulable_data {
10251 struct task_group *tg;
10256 static int tg_schedulable(struct task_group *tg, void *data)
10258 struct rt_schedulable_data *d = data;
10259 struct task_group *child;
10260 unsigned long total, sum = 0;
10261 u64 period, runtime;
10263 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10264 runtime = tg->rt_bandwidth.rt_runtime;
10267 period = d->rt_period;
10268 runtime = d->rt_runtime;
10271 #ifdef CONFIG_USER_SCHED
10272 if (tg == &root_task_group) {
10273 period = global_rt_period();
10274 runtime = global_rt_runtime();
10279 * Cannot have more runtime than the period.
10281 if (runtime > period && runtime != RUNTIME_INF)
10285 * Ensure we don't starve existing RT tasks.
10287 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10290 total = to_ratio(period, runtime);
10293 * Nobody can have more than the global setting allows.
10295 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10299 * The sum of our children's runtime should not exceed our own.
10301 list_for_each_entry_rcu(child, &tg->children, siblings) {
10302 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10303 runtime = child->rt_bandwidth.rt_runtime;
10305 if (child == d->tg) {
10306 period = d->rt_period;
10307 runtime = d->rt_runtime;
10310 sum += to_ratio(period, runtime);
10319 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10321 struct rt_schedulable_data data = {
10323 .rt_period = period,
10324 .rt_runtime = runtime,
10327 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10330 static int tg_set_bandwidth(struct task_group *tg,
10331 u64 rt_period, u64 rt_runtime)
10335 mutex_lock(&rt_constraints_mutex);
10336 read_lock(&tasklist_lock);
10337 err = __rt_schedulable(tg, rt_period, rt_runtime);
10341 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10342 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10343 tg->rt_bandwidth.rt_runtime = rt_runtime;
10345 for_each_possible_cpu(i) {
10346 struct rt_rq *rt_rq = tg->rt_rq[i];
10348 spin_lock(&rt_rq->rt_runtime_lock);
10349 rt_rq->rt_runtime = rt_runtime;
10350 spin_unlock(&rt_rq->rt_runtime_lock);
10352 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10354 read_unlock(&tasklist_lock);
10355 mutex_unlock(&rt_constraints_mutex);
10360 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10362 u64 rt_runtime, rt_period;
10364 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10365 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10366 if (rt_runtime_us < 0)
10367 rt_runtime = RUNTIME_INF;
10369 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10372 long sched_group_rt_runtime(struct task_group *tg)
10376 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10379 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10380 do_div(rt_runtime_us, NSEC_PER_USEC);
10381 return rt_runtime_us;
10384 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10386 u64 rt_runtime, rt_period;
10388 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10389 rt_runtime = tg->rt_bandwidth.rt_runtime;
10391 if (rt_period == 0)
10394 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10397 long sched_group_rt_period(struct task_group *tg)
10401 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10402 do_div(rt_period_us, NSEC_PER_USEC);
10403 return rt_period_us;
10406 static int sched_rt_global_constraints(void)
10408 u64 runtime, period;
10411 if (sysctl_sched_rt_period <= 0)
10414 runtime = global_rt_runtime();
10415 period = global_rt_period();
10418 * Sanity check on the sysctl variables.
10420 if (runtime > period && runtime != RUNTIME_INF)
10423 mutex_lock(&rt_constraints_mutex);
10424 read_lock(&tasklist_lock);
10425 ret = __rt_schedulable(NULL, 0, 0);
10426 read_unlock(&tasklist_lock);
10427 mutex_unlock(&rt_constraints_mutex);
10432 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10434 /* Don't accept realtime tasks when there is no way for them to run */
10435 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10441 #else /* !CONFIG_RT_GROUP_SCHED */
10442 static int sched_rt_global_constraints(void)
10444 unsigned long flags;
10447 if (sysctl_sched_rt_period <= 0)
10451 * There's always some RT tasks in the root group
10452 * -- migration, kstopmachine etc..
10454 if (sysctl_sched_rt_runtime == 0)
10457 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10458 for_each_possible_cpu(i) {
10459 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10461 spin_lock(&rt_rq->rt_runtime_lock);
10462 rt_rq->rt_runtime = global_rt_runtime();
10463 spin_unlock(&rt_rq->rt_runtime_lock);
10465 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10469 #endif /* CONFIG_RT_GROUP_SCHED */
10471 int sched_rt_handler(struct ctl_table *table, int write,
10472 struct file *filp, void __user *buffer, size_t *lenp,
10476 int old_period, old_runtime;
10477 static DEFINE_MUTEX(mutex);
10479 mutex_lock(&mutex);
10480 old_period = sysctl_sched_rt_period;
10481 old_runtime = sysctl_sched_rt_runtime;
10483 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
10485 if (!ret && write) {
10486 ret = sched_rt_global_constraints();
10488 sysctl_sched_rt_period = old_period;
10489 sysctl_sched_rt_runtime = old_runtime;
10491 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10492 def_rt_bandwidth.rt_period =
10493 ns_to_ktime(global_rt_period());
10496 mutex_unlock(&mutex);
10501 #ifdef CONFIG_CGROUP_SCHED
10503 /* return corresponding task_group object of a cgroup */
10504 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10506 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10507 struct task_group, css);
10510 static struct cgroup_subsys_state *
10511 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10513 struct task_group *tg, *parent;
10515 if (!cgrp->parent) {
10516 /* This is early initialization for the top cgroup */
10517 return &init_task_group.css;
10520 parent = cgroup_tg(cgrp->parent);
10521 tg = sched_create_group(parent);
10523 return ERR_PTR(-ENOMEM);
10529 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10531 struct task_group *tg = cgroup_tg(cgrp);
10533 sched_destroy_group(tg);
10537 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10538 struct task_struct *tsk)
10540 #ifdef CONFIG_RT_GROUP_SCHED
10541 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10544 /* We don't support RT-tasks being in separate groups */
10545 if (tsk->sched_class != &fair_sched_class)
10553 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10554 struct cgroup *old_cont, struct task_struct *tsk)
10556 sched_move_task(tsk);
10559 #ifdef CONFIG_FAIR_GROUP_SCHED
10560 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10563 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10566 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10568 struct task_group *tg = cgroup_tg(cgrp);
10570 return (u64) tg->shares;
10572 #endif /* CONFIG_FAIR_GROUP_SCHED */
10574 #ifdef CONFIG_RT_GROUP_SCHED
10575 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10578 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10581 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10583 return sched_group_rt_runtime(cgroup_tg(cgrp));
10586 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10589 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10592 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10594 return sched_group_rt_period(cgroup_tg(cgrp));
10596 #endif /* CONFIG_RT_GROUP_SCHED */
10598 static struct cftype cpu_files[] = {
10599 #ifdef CONFIG_FAIR_GROUP_SCHED
10602 .read_u64 = cpu_shares_read_u64,
10603 .write_u64 = cpu_shares_write_u64,
10606 #ifdef CONFIG_RT_GROUP_SCHED
10608 .name = "rt_runtime_us",
10609 .read_s64 = cpu_rt_runtime_read,
10610 .write_s64 = cpu_rt_runtime_write,
10613 .name = "rt_period_us",
10614 .read_u64 = cpu_rt_period_read_uint,
10615 .write_u64 = cpu_rt_period_write_uint,
10620 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10622 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10625 struct cgroup_subsys cpu_cgroup_subsys = {
10627 .create = cpu_cgroup_create,
10628 .destroy = cpu_cgroup_destroy,
10629 .can_attach = cpu_cgroup_can_attach,
10630 .attach = cpu_cgroup_attach,
10631 .populate = cpu_cgroup_populate,
10632 .subsys_id = cpu_cgroup_subsys_id,
10636 #endif /* CONFIG_CGROUP_SCHED */
10638 #ifdef CONFIG_CGROUP_CPUACCT
10641 * CPU accounting code for task groups.
10643 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10644 * (balbir@in.ibm.com).
10647 /* track cpu usage of a group of tasks and its child groups */
10649 struct cgroup_subsys_state css;
10650 /* cpuusage holds pointer to a u64-type object on every cpu */
10652 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10653 struct cpuacct *parent;
10656 struct cgroup_subsys cpuacct_subsys;
10658 /* return cpu accounting group corresponding to this container */
10659 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10661 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10662 struct cpuacct, css);
10665 /* return cpu accounting group to which this task belongs */
10666 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10668 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10669 struct cpuacct, css);
10672 /* create a new cpu accounting group */
10673 static struct cgroup_subsys_state *cpuacct_create(
10674 struct cgroup_subsys *ss, struct cgroup *cgrp)
10676 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10682 ca->cpuusage = alloc_percpu(u64);
10686 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10687 if (percpu_counter_init(&ca->cpustat[i], 0))
10688 goto out_free_counters;
10691 ca->parent = cgroup_ca(cgrp->parent);
10697 percpu_counter_destroy(&ca->cpustat[i]);
10698 free_percpu(ca->cpuusage);
10702 return ERR_PTR(-ENOMEM);
10705 /* destroy an existing cpu accounting group */
10707 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10709 struct cpuacct *ca = cgroup_ca(cgrp);
10712 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10713 percpu_counter_destroy(&ca->cpustat[i]);
10714 free_percpu(ca->cpuusage);
10718 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10720 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10723 #ifndef CONFIG_64BIT
10725 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10727 spin_lock_irq(&cpu_rq(cpu)->lock);
10729 spin_unlock_irq(&cpu_rq(cpu)->lock);
10737 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10739 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10741 #ifndef CONFIG_64BIT
10743 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10745 spin_lock_irq(&cpu_rq(cpu)->lock);
10747 spin_unlock_irq(&cpu_rq(cpu)->lock);
10753 /* return total cpu usage (in nanoseconds) of a group */
10754 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10756 struct cpuacct *ca = cgroup_ca(cgrp);
10757 u64 totalcpuusage = 0;
10760 for_each_present_cpu(i)
10761 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10763 return totalcpuusage;
10766 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10769 struct cpuacct *ca = cgroup_ca(cgrp);
10778 for_each_present_cpu(i)
10779 cpuacct_cpuusage_write(ca, i, 0);
10785 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10786 struct seq_file *m)
10788 struct cpuacct *ca = cgroup_ca(cgroup);
10792 for_each_present_cpu(i) {
10793 percpu = cpuacct_cpuusage_read(ca, i);
10794 seq_printf(m, "%llu ", (unsigned long long) percpu);
10796 seq_printf(m, "\n");
10800 static const char *cpuacct_stat_desc[] = {
10801 [CPUACCT_STAT_USER] = "user",
10802 [CPUACCT_STAT_SYSTEM] = "system",
10805 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10806 struct cgroup_map_cb *cb)
10808 struct cpuacct *ca = cgroup_ca(cgrp);
10811 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10812 s64 val = percpu_counter_read(&ca->cpustat[i]);
10813 val = cputime64_to_clock_t(val);
10814 cb->fill(cb, cpuacct_stat_desc[i], val);
10819 static struct cftype files[] = {
10822 .read_u64 = cpuusage_read,
10823 .write_u64 = cpuusage_write,
10826 .name = "usage_percpu",
10827 .read_seq_string = cpuacct_percpu_seq_read,
10831 .read_map = cpuacct_stats_show,
10835 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10837 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10841 * charge this task's execution time to its accounting group.
10843 * called with rq->lock held.
10845 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10847 struct cpuacct *ca;
10850 if (unlikely(!cpuacct_subsys.active))
10853 cpu = task_cpu(tsk);
10859 for (; ca; ca = ca->parent) {
10860 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10861 *cpuusage += cputime;
10868 * Charge the system/user time to the task's accounting group.
10870 static void cpuacct_update_stats(struct task_struct *tsk,
10871 enum cpuacct_stat_index idx, cputime_t val)
10873 struct cpuacct *ca;
10875 if (unlikely(!cpuacct_subsys.active))
10882 percpu_counter_add(&ca->cpustat[idx], val);
10888 struct cgroup_subsys cpuacct_subsys = {
10890 .create = cpuacct_create,
10891 .destroy = cpuacct_destroy,
10892 .populate = cpuacct_populate,
10893 .subsys_id = cpuacct_subsys_id,
10895 #endif /* CONFIG_CGROUP_CPUACCT */