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_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 */
621 /* cpu of this runqueue: */
625 unsigned long avg_load_per_task;
627 struct task_struct *migration_thread;
628 struct list_head migration_queue;
631 /* calc_load related fields */
632 unsigned long calc_load_update;
633 long calc_load_active;
635 #ifdef CONFIG_SCHED_HRTICK
637 int hrtick_csd_pending;
638 struct call_single_data hrtick_csd;
640 struct hrtimer hrtick_timer;
643 #ifdef CONFIG_SCHEDSTATS
645 struct sched_info rq_sched_info;
646 unsigned long long rq_cpu_time;
647 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
649 /* sys_sched_yield() stats */
650 unsigned int yld_count;
652 /* schedule() stats */
653 unsigned int sched_switch;
654 unsigned int sched_count;
655 unsigned int sched_goidle;
657 /* try_to_wake_up() stats */
658 unsigned int ttwu_count;
659 unsigned int ttwu_local;
662 unsigned int bkl_count;
666 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
668 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
670 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
673 static inline int cpu_of(struct rq *rq)
683 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
684 * See detach_destroy_domains: synchronize_sched for details.
686 * The domain tree of any CPU may only be accessed from within
687 * preempt-disabled sections.
689 #define for_each_domain(cpu, __sd) \
690 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
692 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
693 #define this_rq() (&__get_cpu_var(runqueues))
694 #define task_rq(p) cpu_rq(task_cpu(p))
695 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
696 #define raw_rq() (&__raw_get_cpu_var(runqueues))
698 inline void update_rq_clock(struct rq *rq)
700 rq->clock = sched_clock_cpu(cpu_of(rq));
704 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
706 #ifdef CONFIG_SCHED_DEBUG
707 # define const_debug __read_mostly
709 # define const_debug static const
715 * Returns true if the current cpu runqueue is locked.
716 * This interface allows printk to be called with the runqueue lock
717 * held and know whether or not it is OK to wake up the klogd.
719 int runqueue_is_locked(void)
722 struct rq *rq = cpu_rq(cpu);
725 ret = spin_is_locked(&rq->lock);
731 * Debugging: various feature bits
734 #define SCHED_FEAT(name, enabled) \
735 __SCHED_FEAT_##name ,
738 #include "sched_features.h"
743 #define SCHED_FEAT(name, enabled) \
744 (1UL << __SCHED_FEAT_##name) * enabled |
746 const_debug unsigned int sysctl_sched_features =
747 #include "sched_features.h"
752 #ifdef CONFIG_SCHED_DEBUG
753 #define SCHED_FEAT(name, enabled) \
756 static __read_mostly char *sched_feat_names[] = {
757 #include "sched_features.h"
763 static int sched_feat_show(struct seq_file *m, void *v)
767 for (i = 0; sched_feat_names[i]; i++) {
768 if (!(sysctl_sched_features & (1UL << i)))
770 seq_printf(m, "%s ", sched_feat_names[i]);
778 sched_feat_write(struct file *filp, const char __user *ubuf,
779 size_t cnt, loff_t *ppos)
789 if (copy_from_user(&buf, ubuf, cnt))
794 if (strncmp(buf, "NO_", 3) == 0) {
799 for (i = 0; sched_feat_names[i]; i++) {
800 int len = strlen(sched_feat_names[i]);
802 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
804 sysctl_sched_features &= ~(1UL << i);
806 sysctl_sched_features |= (1UL << i);
811 if (!sched_feat_names[i])
819 static int sched_feat_open(struct inode *inode, struct file *filp)
821 return single_open(filp, sched_feat_show, NULL);
824 static struct file_operations sched_feat_fops = {
825 .open = sched_feat_open,
826 .write = sched_feat_write,
829 .release = single_release,
832 static __init int sched_init_debug(void)
834 debugfs_create_file("sched_features", 0644, NULL, NULL,
839 late_initcall(sched_init_debug);
843 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
846 * Number of tasks to iterate in a single balance run.
847 * Limited because this is done with IRQs disabled.
849 const_debug unsigned int sysctl_sched_nr_migrate = 32;
852 * ratelimit for updating the group shares.
855 unsigned int sysctl_sched_shares_ratelimit = 250000;
858 * Inject some fuzzyness into changing the per-cpu group shares
859 * this avoids remote rq-locks at the expense of fairness.
862 unsigned int sysctl_sched_shares_thresh = 4;
865 * period over which we measure -rt task cpu usage in us.
868 unsigned int sysctl_sched_rt_period = 1000000;
870 static __read_mostly int scheduler_running;
873 * part of the period that we allow rt tasks to run in us.
876 int sysctl_sched_rt_runtime = 950000;
878 static inline u64 global_rt_period(void)
880 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
883 static inline u64 global_rt_runtime(void)
885 if (sysctl_sched_rt_runtime < 0)
888 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
891 #ifndef prepare_arch_switch
892 # define prepare_arch_switch(next) do { } while (0)
894 #ifndef finish_arch_switch
895 # define finish_arch_switch(prev) do { } while (0)
898 static inline int task_current(struct rq *rq, struct task_struct *p)
900 return rq->curr == p;
903 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
904 static inline int task_running(struct rq *rq, struct task_struct *p)
906 return task_current(rq, p);
909 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
913 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
915 #ifdef CONFIG_DEBUG_SPINLOCK
916 /* this is a valid case when another task releases the spinlock */
917 rq->lock.owner = current;
920 * If we are tracking spinlock dependencies then we have to
921 * fix up the runqueue lock - which gets 'carried over' from
924 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
926 spin_unlock_irq(&rq->lock);
929 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
930 static inline int task_running(struct rq *rq, struct task_struct *p)
935 return task_current(rq, p);
939 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
943 * We can optimise this out completely for !SMP, because the
944 * SMP rebalancing from interrupt is the only thing that cares
949 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
950 spin_unlock_irq(&rq->lock);
952 spin_unlock(&rq->lock);
956 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
960 * After ->oncpu is cleared, the task can be moved to a different CPU.
961 * We must ensure this doesn't happen until the switch is completely
967 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
971 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
974 * __task_rq_lock - lock the runqueue a given task resides on.
975 * Must be called interrupts disabled.
977 static inline struct rq *__task_rq_lock(struct task_struct *p)
981 struct rq *rq = task_rq(p);
982 spin_lock(&rq->lock);
983 if (likely(rq == task_rq(p)))
985 spin_unlock(&rq->lock);
990 * task_rq_lock - lock the runqueue a given task resides on and disable
991 * interrupts. Note the ordering: we can safely lookup the task_rq without
992 * explicitly disabling preemption.
994 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1000 local_irq_save(*flags);
1002 spin_lock(&rq->lock);
1003 if (likely(rq == task_rq(p)))
1005 spin_unlock_irqrestore(&rq->lock, *flags);
1009 void task_rq_unlock_wait(struct task_struct *p)
1011 struct rq *rq = task_rq(p);
1013 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1014 spin_unlock_wait(&rq->lock);
1017 static void __task_rq_unlock(struct rq *rq)
1018 __releases(rq->lock)
1020 spin_unlock(&rq->lock);
1023 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1024 __releases(rq->lock)
1026 spin_unlock_irqrestore(&rq->lock, *flags);
1030 * this_rq_lock - lock this runqueue and disable interrupts.
1032 static struct rq *this_rq_lock(void)
1033 __acquires(rq->lock)
1037 local_irq_disable();
1039 spin_lock(&rq->lock);
1044 #ifdef CONFIG_SCHED_HRTICK
1046 * Use HR-timers to deliver accurate preemption points.
1048 * Its all a bit involved since we cannot program an hrt while holding the
1049 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1052 * When we get rescheduled we reprogram the hrtick_timer outside of the
1058 * - enabled by features
1059 * - hrtimer is actually high res
1061 static inline int hrtick_enabled(struct rq *rq)
1063 if (!sched_feat(HRTICK))
1065 if (!cpu_active(cpu_of(rq)))
1067 return hrtimer_is_hres_active(&rq->hrtick_timer);
1070 static void hrtick_clear(struct rq *rq)
1072 if (hrtimer_active(&rq->hrtick_timer))
1073 hrtimer_cancel(&rq->hrtick_timer);
1077 * High-resolution timer tick.
1078 * Runs from hardirq context with interrupts disabled.
1080 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1082 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1084 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1086 spin_lock(&rq->lock);
1087 update_rq_clock(rq);
1088 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1089 spin_unlock(&rq->lock);
1091 return HRTIMER_NORESTART;
1096 * called from hardirq (IPI) context
1098 static void __hrtick_start(void *arg)
1100 struct rq *rq = arg;
1102 spin_lock(&rq->lock);
1103 hrtimer_restart(&rq->hrtick_timer);
1104 rq->hrtick_csd_pending = 0;
1105 spin_unlock(&rq->lock);
1109 * Called to set the hrtick timer state.
1111 * called with rq->lock held and irqs disabled
1113 static void hrtick_start(struct rq *rq, u64 delay)
1115 struct hrtimer *timer = &rq->hrtick_timer;
1116 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1118 hrtimer_set_expires(timer, time);
1120 if (rq == this_rq()) {
1121 hrtimer_restart(timer);
1122 } else if (!rq->hrtick_csd_pending) {
1123 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1124 rq->hrtick_csd_pending = 1;
1129 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1131 int cpu = (int)(long)hcpu;
1134 case CPU_UP_CANCELED:
1135 case CPU_UP_CANCELED_FROZEN:
1136 case CPU_DOWN_PREPARE:
1137 case CPU_DOWN_PREPARE_FROZEN:
1139 case CPU_DEAD_FROZEN:
1140 hrtick_clear(cpu_rq(cpu));
1147 static __init void init_hrtick(void)
1149 hotcpu_notifier(hotplug_hrtick, 0);
1153 * Called to set the hrtick timer state.
1155 * called with rq->lock held and irqs disabled
1157 static void hrtick_start(struct rq *rq, u64 delay)
1159 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1160 HRTIMER_MODE_REL_PINNED, 0);
1163 static inline void init_hrtick(void)
1166 #endif /* CONFIG_SMP */
1168 static void init_rq_hrtick(struct rq *rq)
1171 rq->hrtick_csd_pending = 0;
1173 rq->hrtick_csd.flags = 0;
1174 rq->hrtick_csd.func = __hrtick_start;
1175 rq->hrtick_csd.info = rq;
1178 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1179 rq->hrtick_timer.function = hrtick;
1181 #else /* CONFIG_SCHED_HRTICK */
1182 static inline void hrtick_clear(struct rq *rq)
1186 static inline void init_rq_hrtick(struct rq *rq)
1190 static inline void init_hrtick(void)
1193 #endif /* CONFIG_SCHED_HRTICK */
1196 * resched_task - mark a task 'to be rescheduled now'.
1198 * On UP this means the setting of the need_resched flag, on SMP it
1199 * might also involve a cross-CPU call to trigger the scheduler on
1204 #ifndef tsk_is_polling
1205 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1208 static void resched_task(struct task_struct *p)
1212 assert_spin_locked(&task_rq(p)->lock);
1214 if (test_tsk_need_resched(p))
1217 set_tsk_need_resched(p);
1220 if (cpu == smp_processor_id())
1223 /* NEED_RESCHED must be visible before we test polling */
1225 if (!tsk_is_polling(p))
1226 smp_send_reschedule(cpu);
1229 static void resched_cpu(int cpu)
1231 struct rq *rq = cpu_rq(cpu);
1232 unsigned long flags;
1234 if (!spin_trylock_irqsave(&rq->lock, flags))
1236 resched_task(cpu_curr(cpu));
1237 spin_unlock_irqrestore(&rq->lock, flags);
1242 * When add_timer_on() enqueues a timer into the timer wheel of an
1243 * idle CPU then this timer might expire before the next timer event
1244 * which is scheduled to wake up that CPU. In case of a completely
1245 * idle system the next event might even be infinite time into the
1246 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1247 * leaves the inner idle loop so the newly added timer is taken into
1248 * account when the CPU goes back to idle and evaluates the timer
1249 * wheel for the next timer event.
1251 void wake_up_idle_cpu(int cpu)
1253 struct rq *rq = cpu_rq(cpu);
1255 if (cpu == smp_processor_id())
1259 * This is safe, as this function is called with the timer
1260 * wheel base lock of (cpu) held. When the CPU is on the way
1261 * to idle and has not yet set rq->curr to idle then it will
1262 * be serialized on the timer wheel base lock and take the new
1263 * timer into account automatically.
1265 if (rq->curr != rq->idle)
1269 * We can set TIF_RESCHED on the idle task of the other CPU
1270 * lockless. The worst case is that the other CPU runs the
1271 * idle task through an additional NOOP schedule()
1273 set_tsk_need_resched(rq->idle);
1275 /* NEED_RESCHED must be visible before we test polling */
1277 if (!tsk_is_polling(rq->idle))
1278 smp_send_reschedule(cpu);
1280 #endif /* CONFIG_NO_HZ */
1282 #else /* !CONFIG_SMP */
1283 static void resched_task(struct task_struct *p)
1285 assert_spin_locked(&task_rq(p)->lock);
1286 set_tsk_need_resched(p);
1288 #endif /* CONFIG_SMP */
1290 #if BITS_PER_LONG == 32
1291 # define WMULT_CONST (~0UL)
1293 # define WMULT_CONST (1UL << 32)
1296 #define WMULT_SHIFT 32
1299 * Shift right and round:
1301 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1304 * delta *= weight / lw
1306 static unsigned long
1307 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1308 struct load_weight *lw)
1312 if (!lw->inv_weight) {
1313 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1316 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1320 tmp = (u64)delta_exec * weight;
1322 * Check whether we'd overflow the 64-bit multiplication:
1324 if (unlikely(tmp > WMULT_CONST))
1325 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1328 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1330 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1333 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1339 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1346 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1347 * of tasks with abnormal "nice" values across CPUs the contribution that
1348 * each task makes to its run queue's load is weighted according to its
1349 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1350 * scaled version of the new time slice allocation that they receive on time
1354 #define WEIGHT_IDLEPRIO 3
1355 #define WMULT_IDLEPRIO 1431655765
1358 * Nice levels are multiplicative, with a gentle 10% change for every
1359 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1360 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1361 * that remained on nice 0.
1363 * The "10% effect" is relative and cumulative: from _any_ nice level,
1364 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1365 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1366 * If a task goes up by ~10% and another task goes down by ~10% then
1367 * the relative distance between them is ~25%.)
1369 static const int prio_to_weight[40] = {
1370 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1371 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1372 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1373 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1374 /* 0 */ 1024, 820, 655, 526, 423,
1375 /* 5 */ 335, 272, 215, 172, 137,
1376 /* 10 */ 110, 87, 70, 56, 45,
1377 /* 15 */ 36, 29, 23, 18, 15,
1381 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1383 * In cases where the weight does not change often, we can use the
1384 * precalculated inverse to speed up arithmetics by turning divisions
1385 * into multiplications:
1387 static const u32 prio_to_wmult[40] = {
1388 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1389 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1390 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1391 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1392 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1393 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1394 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1395 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1398 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1401 * runqueue iterator, to support SMP load-balancing between different
1402 * scheduling classes, without having to expose their internal data
1403 * structures to the load-balancing proper:
1405 struct rq_iterator {
1407 struct task_struct *(*start)(void *);
1408 struct task_struct *(*next)(void *);
1412 static unsigned long
1413 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1414 unsigned long max_load_move, struct sched_domain *sd,
1415 enum cpu_idle_type idle, int *all_pinned,
1416 int *this_best_prio, struct rq_iterator *iterator);
1419 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1420 struct sched_domain *sd, enum cpu_idle_type idle,
1421 struct rq_iterator *iterator);
1424 /* Time spent by the tasks of the cpu accounting group executing in ... */
1425 enum cpuacct_stat_index {
1426 CPUACCT_STAT_USER, /* ... user mode */
1427 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1429 CPUACCT_STAT_NSTATS,
1432 #ifdef CONFIG_CGROUP_CPUACCT
1433 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1434 static void cpuacct_update_stats(struct task_struct *tsk,
1435 enum cpuacct_stat_index idx, cputime_t val);
1437 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1438 static inline void cpuacct_update_stats(struct task_struct *tsk,
1439 enum cpuacct_stat_index idx, cputime_t val) {}
1442 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1444 update_load_add(&rq->load, load);
1447 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1449 update_load_sub(&rq->load, load);
1452 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1453 typedef int (*tg_visitor)(struct task_group *, void *);
1456 * Iterate the full tree, calling @down when first entering a node and @up when
1457 * leaving it for the final time.
1459 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1461 struct task_group *parent, *child;
1465 parent = &root_task_group;
1467 ret = (*down)(parent, data);
1470 list_for_each_entry_rcu(child, &parent->children, siblings) {
1477 ret = (*up)(parent, data);
1482 parent = parent->parent;
1491 static int tg_nop(struct task_group *tg, void *data)
1498 static unsigned long source_load(int cpu, int type);
1499 static unsigned long target_load(int cpu, int type);
1500 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1502 static unsigned long cpu_avg_load_per_task(int cpu)
1504 struct rq *rq = cpu_rq(cpu);
1505 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1508 rq->avg_load_per_task = rq->load.weight / nr_running;
1510 rq->avg_load_per_task = 0;
1512 return rq->avg_load_per_task;
1515 #ifdef CONFIG_FAIR_GROUP_SCHED
1517 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1520 * Calculate and set the cpu's group shares.
1523 update_group_shares_cpu(struct task_group *tg, int cpu,
1524 unsigned long sd_shares, unsigned long sd_rq_weight)
1526 unsigned long rq_weight;
1527 unsigned long shares;
1533 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1536 rq_weight = NICE_0_LOAD;
1540 * \Sum shares * rq_weight
1541 * shares = -----------------------
1545 shares = (sd_shares * rq_weight) / sd_rq_weight;
1546 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1548 if (abs(shares - tg->se[cpu]->load.weight) >
1549 sysctl_sched_shares_thresh) {
1550 struct rq *rq = cpu_rq(cpu);
1551 unsigned long flags;
1553 spin_lock_irqsave(&rq->lock, flags);
1554 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1555 __set_se_shares(tg->se[cpu], shares);
1556 spin_unlock_irqrestore(&rq->lock, flags);
1561 * Re-compute the task group their per cpu shares over the given domain.
1562 * This needs to be done in a bottom-up fashion because the rq weight of a
1563 * parent group depends on the shares of its child groups.
1565 static int tg_shares_up(struct task_group *tg, void *data)
1567 unsigned long weight, rq_weight = 0, eff_weight = 0;
1568 unsigned long shares = 0;
1569 struct sched_domain *sd = data;
1572 for_each_cpu(i, sched_domain_span(sd)) {
1574 * If there are currently no tasks on the cpu pretend there
1575 * is one of average load so that when a new task gets to
1576 * run here it will not get delayed by group starvation.
1578 weight = tg->cfs_rq[i]->load.weight;
1579 tg->cfs_rq[i]->rq_weight = weight;
1580 rq_weight += weight;
1583 weight = NICE_0_LOAD;
1585 eff_weight += weight;
1586 shares += tg->cfs_rq[i]->shares;
1589 if ((!shares && rq_weight) || shares > tg->shares)
1590 shares = tg->shares;
1592 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1593 shares = tg->shares;
1595 for_each_cpu(i, sched_domain_span(sd)) {
1596 unsigned long sd_rq_weight = rq_weight;
1598 if (!tg->cfs_rq[i]->rq_weight)
1599 sd_rq_weight = eff_weight;
1601 update_group_shares_cpu(tg, i, shares, sd_rq_weight);
1608 * Compute the cpu's hierarchical load factor for each task group.
1609 * This needs to be done in a top-down fashion because the load of a child
1610 * group is a fraction of its parents load.
1612 static int tg_load_down(struct task_group *tg, void *data)
1615 long cpu = (long)data;
1618 load = cpu_rq(cpu)->load.weight;
1620 load = tg->parent->cfs_rq[cpu]->h_load;
1621 load *= tg->cfs_rq[cpu]->shares;
1622 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1625 tg->cfs_rq[cpu]->h_load = load;
1630 static void update_shares(struct sched_domain *sd)
1635 if (root_task_group_empty())
1638 now = cpu_clock(raw_smp_processor_id());
1639 elapsed = now - sd->last_update;
1641 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1642 sd->last_update = now;
1643 walk_tg_tree(tg_nop, tg_shares_up, sd);
1647 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1649 if (root_task_group_empty())
1652 spin_unlock(&rq->lock);
1654 spin_lock(&rq->lock);
1657 static void update_h_load(long cpu)
1659 if (root_task_group_empty())
1662 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1667 static inline void update_shares(struct sched_domain *sd)
1671 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1677 #ifdef CONFIG_PREEMPT
1680 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1681 * way at the expense of forcing extra atomic operations in all
1682 * invocations. This assures that the double_lock is acquired using the
1683 * same underlying policy as the spinlock_t on this architecture, which
1684 * reduces latency compared to the unfair variant below. However, it
1685 * also adds more overhead and therefore may reduce throughput.
1687 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1688 __releases(this_rq->lock)
1689 __acquires(busiest->lock)
1690 __acquires(this_rq->lock)
1692 spin_unlock(&this_rq->lock);
1693 double_rq_lock(this_rq, busiest);
1700 * Unfair double_lock_balance: Optimizes throughput at the expense of
1701 * latency by eliminating extra atomic operations when the locks are
1702 * already in proper order on entry. This favors lower cpu-ids and will
1703 * grant the double lock to lower cpus over higher ids under contention,
1704 * regardless of entry order into the function.
1706 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1707 __releases(this_rq->lock)
1708 __acquires(busiest->lock)
1709 __acquires(this_rq->lock)
1713 if (unlikely(!spin_trylock(&busiest->lock))) {
1714 if (busiest < this_rq) {
1715 spin_unlock(&this_rq->lock);
1716 spin_lock(&busiest->lock);
1717 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1720 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1725 #endif /* CONFIG_PREEMPT */
1728 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1730 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1732 if (unlikely(!irqs_disabled())) {
1733 /* printk() doesn't work good under rq->lock */
1734 spin_unlock(&this_rq->lock);
1738 return _double_lock_balance(this_rq, busiest);
1741 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1742 __releases(busiest->lock)
1744 spin_unlock(&busiest->lock);
1745 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1749 #ifdef CONFIG_FAIR_GROUP_SCHED
1750 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1753 cfs_rq->shares = shares;
1758 static void calc_load_account_active(struct rq *this_rq);
1760 #include "sched_stats.h"
1761 #include "sched_idletask.c"
1762 #include "sched_fair.c"
1763 #include "sched_rt.c"
1764 #ifdef CONFIG_SCHED_DEBUG
1765 # include "sched_debug.c"
1768 #define sched_class_highest (&rt_sched_class)
1769 #define for_each_class(class) \
1770 for (class = sched_class_highest; class; class = class->next)
1772 static void inc_nr_running(struct rq *rq)
1777 static void dec_nr_running(struct rq *rq)
1782 static void set_load_weight(struct task_struct *p)
1784 if (task_has_rt_policy(p)) {
1785 p->se.load.weight = prio_to_weight[0] * 2;
1786 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1791 * SCHED_IDLE tasks get minimal weight:
1793 if (p->policy == SCHED_IDLE) {
1794 p->se.load.weight = WEIGHT_IDLEPRIO;
1795 p->se.load.inv_weight = WMULT_IDLEPRIO;
1799 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1800 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1803 static void update_avg(u64 *avg, u64 sample)
1805 s64 diff = sample - *avg;
1809 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1812 p->se.start_runtime = p->se.sum_exec_runtime;
1814 sched_info_queued(p);
1815 p->sched_class->enqueue_task(rq, p, wakeup);
1819 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1822 if (p->se.last_wakeup) {
1823 update_avg(&p->se.avg_overlap,
1824 p->se.sum_exec_runtime - p->se.last_wakeup);
1825 p->se.last_wakeup = 0;
1827 update_avg(&p->se.avg_wakeup,
1828 sysctl_sched_wakeup_granularity);
1832 sched_info_dequeued(p);
1833 p->sched_class->dequeue_task(rq, p, sleep);
1838 * __normal_prio - return the priority that is based on the static prio
1840 static inline int __normal_prio(struct task_struct *p)
1842 return p->static_prio;
1846 * Calculate the expected normal priority: i.e. priority
1847 * without taking RT-inheritance into account. Might be
1848 * boosted by interactivity modifiers. Changes upon fork,
1849 * setprio syscalls, and whenever the interactivity
1850 * estimator recalculates.
1852 static inline int normal_prio(struct task_struct *p)
1856 if (task_has_rt_policy(p))
1857 prio = MAX_RT_PRIO-1 - p->rt_priority;
1859 prio = __normal_prio(p);
1864 * Calculate the current priority, i.e. the priority
1865 * taken into account by the scheduler. This value might
1866 * be boosted by RT tasks, or might be boosted by
1867 * interactivity modifiers. Will be RT if the task got
1868 * RT-boosted. If not then it returns p->normal_prio.
1870 static int effective_prio(struct task_struct *p)
1872 p->normal_prio = normal_prio(p);
1874 * If we are RT tasks or we were boosted to RT priority,
1875 * keep the priority unchanged. Otherwise, update priority
1876 * to the normal priority:
1878 if (!rt_prio(p->prio))
1879 return p->normal_prio;
1884 * activate_task - move a task to the runqueue.
1886 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1888 if (task_contributes_to_load(p))
1889 rq->nr_uninterruptible--;
1891 enqueue_task(rq, p, wakeup);
1896 * deactivate_task - remove a task from the runqueue.
1898 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1900 if (task_contributes_to_load(p))
1901 rq->nr_uninterruptible++;
1903 dequeue_task(rq, p, sleep);
1908 * task_curr - is this task currently executing on a CPU?
1909 * @p: the task in question.
1911 inline int task_curr(const struct task_struct *p)
1913 return cpu_curr(task_cpu(p)) == p;
1916 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1918 set_task_rq(p, cpu);
1921 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1922 * successfuly executed on another CPU. We must ensure that updates of
1923 * per-task data have been completed by this moment.
1926 task_thread_info(p)->cpu = cpu;
1930 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1931 const struct sched_class *prev_class,
1932 int oldprio, int running)
1934 if (prev_class != p->sched_class) {
1935 if (prev_class->switched_from)
1936 prev_class->switched_from(rq, p, running);
1937 p->sched_class->switched_to(rq, p, running);
1939 p->sched_class->prio_changed(rq, p, oldprio, running);
1944 /* Used instead of source_load when we know the type == 0 */
1945 static unsigned long weighted_cpuload(const int cpu)
1947 return cpu_rq(cpu)->load.weight;
1951 * Is this task likely cache-hot:
1954 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1959 * Buddy candidates are cache hot:
1961 if (sched_feat(CACHE_HOT_BUDDY) &&
1962 (&p->se == cfs_rq_of(&p->se)->next ||
1963 &p->se == cfs_rq_of(&p->se)->last))
1966 if (p->sched_class != &fair_sched_class)
1969 if (sysctl_sched_migration_cost == -1)
1971 if (sysctl_sched_migration_cost == 0)
1974 delta = now - p->se.exec_start;
1976 return delta < (s64)sysctl_sched_migration_cost;
1980 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1982 int old_cpu = task_cpu(p);
1983 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1984 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1985 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1988 clock_offset = old_rq->clock - new_rq->clock;
1990 trace_sched_migrate_task(p, new_cpu);
1992 #ifdef CONFIG_SCHEDSTATS
1993 if (p->se.wait_start)
1994 p->se.wait_start -= clock_offset;
1995 if (p->se.sleep_start)
1996 p->se.sleep_start -= clock_offset;
1997 if (p->se.block_start)
1998 p->se.block_start -= clock_offset;
2000 if (old_cpu != new_cpu) {
2001 p->se.nr_migrations++;
2002 new_rq->nr_migrations_in++;
2003 #ifdef CONFIG_SCHEDSTATS
2004 if (task_hot(p, old_rq->clock, NULL))
2005 schedstat_inc(p, se.nr_forced2_migrations);
2007 perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2010 p->se.vruntime -= old_cfsrq->min_vruntime -
2011 new_cfsrq->min_vruntime;
2013 __set_task_cpu(p, new_cpu);
2016 struct migration_req {
2017 struct list_head list;
2019 struct task_struct *task;
2022 struct completion done;
2026 * The task's runqueue lock must be held.
2027 * Returns true if you have to wait for migration thread.
2030 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2032 struct rq *rq = task_rq(p);
2035 * If the task is not on a runqueue (and not running), then
2036 * it is sufficient to simply update the task's cpu field.
2038 if (!p->se.on_rq && !task_running(rq, p)) {
2039 set_task_cpu(p, dest_cpu);
2043 init_completion(&req->done);
2045 req->dest_cpu = dest_cpu;
2046 list_add(&req->list, &rq->migration_queue);
2052 * wait_task_context_switch - wait for a thread to complete at least one
2055 * @p must not be current.
2057 void wait_task_context_switch(struct task_struct *p)
2059 unsigned long nvcsw, nivcsw, flags;
2067 * The runqueue is assigned before the actual context
2068 * switch. We need to take the runqueue lock.
2070 * We could check initially without the lock but it is
2071 * very likely that we need to take the lock in every
2074 rq = task_rq_lock(p, &flags);
2075 running = task_running(rq, p);
2076 task_rq_unlock(rq, &flags);
2078 if (likely(!running))
2081 * The switch count is incremented before the actual
2082 * context switch. We thus wait for two switches to be
2083 * sure at least one completed.
2085 if ((p->nvcsw - nvcsw) > 1)
2087 if ((p->nivcsw - nivcsw) > 1)
2095 * wait_task_inactive - wait for a thread to unschedule.
2097 * If @match_state is nonzero, it's the @p->state value just checked and
2098 * not expected to change. If it changes, i.e. @p might have woken up,
2099 * then return zero. When we succeed in waiting for @p to be off its CPU,
2100 * we return a positive number (its total switch count). If a second call
2101 * a short while later returns the same number, the caller can be sure that
2102 * @p has remained unscheduled the whole time.
2104 * The caller must ensure that the task *will* unschedule sometime soon,
2105 * else this function might spin for a *long* time. This function can't
2106 * be called with interrupts off, or it may introduce deadlock with
2107 * smp_call_function() if an IPI is sent by the same process we are
2108 * waiting to become inactive.
2110 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2112 unsigned long flags;
2119 * We do the initial early heuristics without holding
2120 * any task-queue locks at all. We'll only try to get
2121 * the runqueue lock when things look like they will
2127 * If the task is actively running on another CPU
2128 * still, just relax and busy-wait without holding
2131 * NOTE! Since we don't hold any locks, it's not
2132 * even sure that "rq" stays as the right runqueue!
2133 * But we don't care, since "task_running()" will
2134 * return false if the runqueue has changed and p
2135 * is actually now running somewhere else!
2137 while (task_running(rq, p)) {
2138 if (match_state && unlikely(p->state != match_state))
2144 * Ok, time to look more closely! We need the rq
2145 * lock now, to be *sure*. If we're wrong, we'll
2146 * just go back and repeat.
2148 rq = task_rq_lock(p, &flags);
2149 trace_sched_wait_task(rq, p);
2150 running = task_running(rq, p);
2151 on_rq = p->se.on_rq;
2153 if (!match_state || p->state == match_state)
2154 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2155 task_rq_unlock(rq, &flags);
2158 * If it changed from the expected state, bail out now.
2160 if (unlikely(!ncsw))
2164 * Was it really running after all now that we
2165 * checked with the proper locks actually held?
2167 * Oops. Go back and try again..
2169 if (unlikely(running)) {
2175 * It's not enough that it's not actively running,
2176 * it must be off the runqueue _entirely_, and not
2179 * So if it was still runnable (but just not actively
2180 * running right now), it's preempted, and we should
2181 * yield - it could be a while.
2183 if (unlikely(on_rq)) {
2184 schedule_timeout_uninterruptible(1);
2189 * Ahh, all good. It wasn't running, and it wasn't
2190 * runnable, which means that it will never become
2191 * running in the future either. We're all done!
2200 * kick_process - kick a running thread to enter/exit the kernel
2201 * @p: the to-be-kicked thread
2203 * Cause a process which is running on another CPU to enter
2204 * kernel-mode, without any delay. (to get signals handled.)
2206 * NOTE: this function doesnt have to take the runqueue lock,
2207 * because all it wants to ensure is that the remote task enters
2208 * the kernel. If the IPI races and the task has been migrated
2209 * to another CPU then no harm is done and the purpose has been
2212 void kick_process(struct task_struct *p)
2218 if ((cpu != smp_processor_id()) && task_curr(p))
2219 smp_send_reschedule(cpu);
2222 EXPORT_SYMBOL_GPL(kick_process);
2225 * Return a low guess at the load of a migration-source cpu weighted
2226 * according to the scheduling class and "nice" value.
2228 * We want to under-estimate the load of migration sources, to
2229 * balance conservatively.
2231 static unsigned long source_load(int cpu, int type)
2233 struct rq *rq = cpu_rq(cpu);
2234 unsigned long total = weighted_cpuload(cpu);
2236 if (type == 0 || !sched_feat(LB_BIAS))
2239 return min(rq->cpu_load[type-1], total);
2243 * Return a high guess at the load of a migration-target cpu weighted
2244 * according to the scheduling class and "nice" value.
2246 static unsigned long target_load(int cpu, int type)
2248 struct rq *rq = cpu_rq(cpu);
2249 unsigned long total = weighted_cpuload(cpu);
2251 if (type == 0 || !sched_feat(LB_BIAS))
2254 return max(rq->cpu_load[type-1], total);
2258 * find_idlest_group finds and returns the least busy CPU group within the
2261 static struct sched_group *
2262 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2264 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2265 unsigned long min_load = ULONG_MAX, this_load = 0;
2266 int load_idx = sd->forkexec_idx;
2267 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2270 unsigned long load, avg_load;
2274 /* Skip over this group if it has no CPUs allowed */
2275 if (!cpumask_intersects(sched_group_cpus(group),
2279 local_group = cpumask_test_cpu(this_cpu,
2280 sched_group_cpus(group));
2282 /* Tally up the load of all CPUs in the group */
2285 for_each_cpu(i, sched_group_cpus(group)) {
2286 /* Bias balancing toward cpus of our domain */
2288 load = source_load(i, load_idx);
2290 load = target_load(i, load_idx);
2295 /* Adjust by relative CPU power of the group */
2296 avg_load = sg_div_cpu_power(group,
2297 avg_load * SCHED_LOAD_SCALE);
2300 this_load = avg_load;
2302 } else if (avg_load < min_load) {
2303 min_load = avg_load;
2306 } while (group = group->next, group != sd->groups);
2308 if (!idlest || 100*this_load < imbalance*min_load)
2314 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2317 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2319 unsigned long load, min_load = ULONG_MAX;
2323 /* Traverse only the allowed CPUs */
2324 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2325 load = weighted_cpuload(i);
2327 if (load < min_load || (load == min_load && i == this_cpu)) {
2337 * sched_balance_self: balance the current task (running on cpu) in domains
2338 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2341 * Balance, ie. select the least loaded group.
2343 * Returns the target CPU number, or the same CPU if no balancing is needed.
2345 * preempt must be disabled.
2347 static int sched_balance_self(int cpu, int flag)
2349 struct task_struct *t = current;
2350 struct sched_domain *tmp, *sd = NULL;
2352 for_each_domain(cpu, tmp) {
2354 * If power savings logic is enabled for a domain, stop there.
2356 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2358 if (tmp->flags & flag)
2366 struct sched_group *group;
2367 int new_cpu, weight;
2369 if (!(sd->flags & flag)) {
2374 group = find_idlest_group(sd, t, cpu);
2380 new_cpu = find_idlest_cpu(group, t, cpu);
2381 if (new_cpu == -1 || new_cpu == cpu) {
2382 /* Now try balancing at a lower domain level of cpu */
2387 /* Now try balancing at a lower domain level of new_cpu */
2389 weight = cpumask_weight(sched_domain_span(sd));
2391 for_each_domain(cpu, tmp) {
2392 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2394 if (tmp->flags & flag)
2397 /* while loop will break here if sd == NULL */
2403 #endif /* CONFIG_SMP */
2406 * task_oncpu_function_call - call a function on the cpu on which a task runs
2407 * @p: the task to evaluate
2408 * @func: the function to be called
2409 * @info: the function call argument
2411 * Calls the function @func when the task is currently running. This might
2412 * be on the current CPU, which just calls the function directly
2414 void task_oncpu_function_call(struct task_struct *p,
2415 void (*func) (void *info), void *info)
2422 smp_call_function_single(cpu, func, info, 1);
2427 * try_to_wake_up - wake up a thread
2428 * @p: the to-be-woken-up thread
2429 * @state: the mask of task states that can be woken
2430 * @sync: do a synchronous wakeup?
2432 * Put it on the run-queue if it's not already there. The "current"
2433 * thread is always on the run-queue (except when the actual
2434 * re-schedule is in progress), and as such you're allowed to do
2435 * the simpler "current->state = TASK_RUNNING" to mark yourself
2436 * runnable without the overhead of this.
2438 * returns failure only if the task is already active.
2440 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2442 int cpu, orig_cpu, this_cpu, success = 0;
2443 unsigned long flags;
2447 if (!sched_feat(SYNC_WAKEUPS))
2451 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2452 struct sched_domain *sd;
2454 this_cpu = raw_smp_processor_id();
2457 for_each_domain(this_cpu, sd) {
2458 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2467 rq = task_rq_lock(p, &flags);
2468 update_rq_clock(rq);
2469 old_state = p->state;
2470 if (!(old_state & state))
2478 this_cpu = smp_processor_id();
2481 if (unlikely(task_running(rq, p)))
2484 cpu = p->sched_class->select_task_rq(p, sync);
2485 if (cpu != orig_cpu) {
2486 set_task_cpu(p, cpu);
2487 task_rq_unlock(rq, &flags);
2488 /* might preempt at this point */
2489 rq = task_rq_lock(p, &flags);
2490 old_state = p->state;
2491 if (!(old_state & state))
2496 this_cpu = smp_processor_id();
2500 #ifdef CONFIG_SCHEDSTATS
2501 schedstat_inc(rq, ttwu_count);
2502 if (cpu == this_cpu)
2503 schedstat_inc(rq, ttwu_local);
2505 struct sched_domain *sd;
2506 for_each_domain(this_cpu, sd) {
2507 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2508 schedstat_inc(sd, ttwu_wake_remote);
2513 #endif /* CONFIG_SCHEDSTATS */
2516 #endif /* CONFIG_SMP */
2517 schedstat_inc(p, se.nr_wakeups);
2519 schedstat_inc(p, se.nr_wakeups_sync);
2520 if (orig_cpu != cpu)
2521 schedstat_inc(p, se.nr_wakeups_migrate);
2522 if (cpu == this_cpu)
2523 schedstat_inc(p, se.nr_wakeups_local);
2525 schedstat_inc(p, se.nr_wakeups_remote);
2526 activate_task(rq, p, 1);
2530 * Only attribute actual wakeups done by this task.
2532 if (!in_interrupt()) {
2533 struct sched_entity *se = ¤t->se;
2534 u64 sample = se->sum_exec_runtime;
2536 if (se->last_wakeup)
2537 sample -= se->last_wakeup;
2539 sample -= se->start_runtime;
2540 update_avg(&se->avg_wakeup, sample);
2542 se->last_wakeup = se->sum_exec_runtime;
2546 trace_sched_wakeup(rq, p, success);
2547 check_preempt_curr(rq, p, sync);
2549 p->state = TASK_RUNNING;
2551 if (p->sched_class->task_wake_up)
2552 p->sched_class->task_wake_up(rq, p);
2555 task_rq_unlock(rq, &flags);
2561 * wake_up_process - Wake up a specific process
2562 * @p: The process to be woken up.
2564 * Attempt to wake up the nominated process and move it to the set of runnable
2565 * processes. Returns 1 if the process was woken up, 0 if it was already
2568 * It may be assumed that this function implies a write memory barrier before
2569 * changing the task state if and only if any tasks are woken up.
2571 int wake_up_process(struct task_struct *p)
2573 return try_to_wake_up(p, TASK_ALL, 0);
2575 EXPORT_SYMBOL(wake_up_process);
2577 int wake_up_state(struct task_struct *p, unsigned int state)
2579 return try_to_wake_up(p, state, 0);
2583 * Perform scheduler related setup for a newly forked process p.
2584 * p is forked by current.
2586 * __sched_fork() is basic setup used by init_idle() too:
2588 static void __sched_fork(struct task_struct *p)
2590 p->se.exec_start = 0;
2591 p->se.sum_exec_runtime = 0;
2592 p->se.prev_sum_exec_runtime = 0;
2593 p->se.nr_migrations = 0;
2594 p->se.last_wakeup = 0;
2595 p->se.avg_overlap = 0;
2596 p->se.start_runtime = 0;
2597 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2599 #ifdef CONFIG_SCHEDSTATS
2600 p->se.wait_start = 0;
2602 p->se.wait_count = 0;
2605 p->se.sleep_start = 0;
2606 p->se.sleep_max = 0;
2607 p->se.sum_sleep_runtime = 0;
2609 p->se.block_start = 0;
2610 p->se.block_max = 0;
2612 p->se.slice_max = 0;
2614 p->se.nr_migrations_cold = 0;
2615 p->se.nr_failed_migrations_affine = 0;
2616 p->se.nr_failed_migrations_running = 0;
2617 p->se.nr_failed_migrations_hot = 0;
2618 p->se.nr_forced_migrations = 0;
2619 p->se.nr_forced2_migrations = 0;
2621 p->se.nr_wakeups = 0;
2622 p->se.nr_wakeups_sync = 0;
2623 p->se.nr_wakeups_migrate = 0;
2624 p->se.nr_wakeups_local = 0;
2625 p->se.nr_wakeups_remote = 0;
2626 p->se.nr_wakeups_affine = 0;
2627 p->se.nr_wakeups_affine_attempts = 0;
2628 p->se.nr_wakeups_passive = 0;
2629 p->se.nr_wakeups_idle = 0;
2633 INIT_LIST_HEAD(&p->rt.run_list);
2635 INIT_LIST_HEAD(&p->se.group_node);
2637 #ifdef CONFIG_PREEMPT_NOTIFIERS
2638 INIT_HLIST_HEAD(&p->preempt_notifiers);
2642 * We mark the process as running here, but have not actually
2643 * inserted it onto the runqueue yet. This guarantees that
2644 * nobody will actually run it, and a signal or other external
2645 * event cannot wake it up and insert it on the runqueue either.
2647 p->state = TASK_RUNNING;
2651 * fork()/clone()-time setup:
2653 void sched_fork(struct task_struct *p, int clone_flags)
2655 int cpu = get_cpu();
2660 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2662 set_task_cpu(p, cpu);
2665 * Make sure we do not leak PI boosting priority to the child.
2667 p->prio = current->normal_prio;
2670 * Revert to default priority/policy on fork if requested.
2672 if (unlikely(p->sched_reset_on_fork)) {
2673 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR)
2674 p->policy = SCHED_NORMAL;
2676 if (p->normal_prio < DEFAULT_PRIO)
2677 p->prio = DEFAULT_PRIO;
2679 if (PRIO_TO_NICE(p->static_prio) < 0) {
2680 p->static_prio = NICE_TO_PRIO(0);
2685 * We don't need the reset flag anymore after the fork. It has
2686 * fulfilled its duty:
2688 p->sched_reset_on_fork = 0;
2691 if (!rt_prio(p->prio))
2692 p->sched_class = &fair_sched_class;
2694 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2695 if (likely(sched_info_on()))
2696 memset(&p->sched_info, 0, sizeof(p->sched_info));
2698 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2701 #ifdef CONFIG_PREEMPT
2702 /* Want to start with kernel preemption disabled. */
2703 task_thread_info(p)->preempt_count = 1;
2705 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2711 * wake_up_new_task - wake up a newly created task for the first time.
2713 * This function will do some initial scheduler statistics housekeeping
2714 * that must be done for every newly created context, then puts the task
2715 * on the runqueue and wakes it.
2717 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2719 unsigned long flags;
2722 rq = task_rq_lock(p, &flags);
2723 BUG_ON(p->state != TASK_RUNNING);
2724 update_rq_clock(rq);
2726 p->prio = effective_prio(p);
2728 if (!p->sched_class->task_new || !current->se.on_rq) {
2729 activate_task(rq, p, 0);
2732 * Let the scheduling class do new task startup
2733 * management (if any):
2735 p->sched_class->task_new(rq, p);
2738 trace_sched_wakeup_new(rq, p, 1);
2739 check_preempt_curr(rq, p, 0);
2741 if (p->sched_class->task_wake_up)
2742 p->sched_class->task_wake_up(rq, p);
2744 task_rq_unlock(rq, &flags);
2747 #ifdef CONFIG_PREEMPT_NOTIFIERS
2750 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2751 * @notifier: notifier struct to register
2753 void preempt_notifier_register(struct preempt_notifier *notifier)
2755 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2757 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2760 * preempt_notifier_unregister - no longer interested in preemption notifications
2761 * @notifier: notifier struct to unregister
2763 * This is safe to call from within a preemption notifier.
2765 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2767 hlist_del(¬ifier->link);
2769 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2771 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2773 struct preempt_notifier *notifier;
2774 struct hlist_node *node;
2776 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2777 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2781 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2782 struct task_struct *next)
2784 struct preempt_notifier *notifier;
2785 struct hlist_node *node;
2787 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2788 notifier->ops->sched_out(notifier, next);
2791 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2793 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2798 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2799 struct task_struct *next)
2803 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2806 * prepare_task_switch - prepare to switch tasks
2807 * @rq: the runqueue preparing to switch
2808 * @prev: the current task that is being switched out
2809 * @next: the task we are going to switch to.
2811 * This is called with the rq lock held and interrupts off. It must
2812 * be paired with a subsequent finish_task_switch after the context
2815 * prepare_task_switch sets up locking and calls architecture specific
2819 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2820 struct task_struct *next)
2822 fire_sched_out_preempt_notifiers(prev, next);
2823 prepare_lock_switch(rq, next);
2824 prepare_arch_switch(next);
2828 * finish_task_switch - clean up after a task-switch
2829 * @rq: runqueue associated with task-switch
2830 * @prev: the thread we just switched away from.
2832 * finish_task_switch must be called after the context switch, paired
2833 * with a prepare_task_switch call before the context switch.
2834 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2835 * and do any other architecture-specific cleanup actions.
2837 * Note that we may have delayed dropping an mm in context_switch(). If
2838 * so, we finish that here outside of the runqueue lock. (Doing it
2839 * with the lock held can cause deadlocks; see schedule() for
2842 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2843 __releases(rq->lock)
2845 struct mm_struct *mm = rq->prev_mm;
2848 int post_schedule = 0;
2850 if (current->sched_class->needs_post_schedule)
2851 post_schedule = current->sched_class->needs_post_schedule(rq);
2857 * A task struct has one reference for the use as "current".
2858 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2859 * schedule one last time. The schedule call will never return, and
2860 * the scheduled task must drop that reference.
2861 * The test for TASK_DEAD must occur while the runqueue locks are
2862 * still held, otherwise prev could be scheduled on another cpu, die
2863 * there before we look at prev->state, and then the reference would
2865 * Manfred Spraul <manfred@colorfullife.com>
2867 prev_state = prev->state;
2868 finish_arch_switch(prev);
2869 perf_counter_task_sched_in(current, cpu_of(rq));
2870 finish_lock_switch(rq, prev);
2873 current->sched_class->post_schedule(rq);
2876 fire_sched_in_preempt_notifiers(current);
2879 if (unlikely(prev_state == TASK_DEAD)) {
2881 * Remove function-return probe instances associated with this
2882 * task and put them back on the free list.
2884 kprobe_flush_task(prev);
2885 put_task_struct(prev);
2890 * schedule_tail - first thing a freshly forked thread must call.
2891 * @prev: the thread we just switched away from.
2893 asmlinkage void schedule_tail(struct task_struct *prev)
2894 __releases(rq->lock)
2896 struct rq *rq = this_rq();
2898 finish_task_switch(rq, prev);
2899 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2900 /* In this case, finish_task_switch does not reenable preemption */
2903 if (current->set_child_tid)
2904 put_user(task_pid_vnr(current), current->set_child_tid);
2908 * context_switch - switch to the new MM and the new
2909 * thread's register state.
2912 context_switch(struct rq *rq, struct task_struct *prev,
2913 struct task_struct *next)
2915 struct mm_struct *mm, *oldmm;
2917 prepare_task_switch(rq, prev, next);
2918 trace_sched_switch(rq, prev, next);
2920 oldmm = prev->active_mm;
2922 * For paravirt, this is coupled with an exit in switch_to to
2923 * combine the page table reload and the switch backend into
2926 arch_start_context_switch(prev);
2928 if (unlikely(!mm)) {
2929 next->active_mm = oldmm;
2930 atomic_inc(&oldmm->mm_count);
2931 enter_lazy_tlb(oldmm, next);
2933 switch_mm(oldmm, mm, next);
2935 if (unlikely(!prev->mm)) {
2936 prev->active_mm = NULL;
2937 rq->prev_mm = oldmm;
2940 * Since the runqueue lock will be released by the next
2941 * task (which is an invalid locking op but in the case
2942 * of the scheduler it's an obvious special-case), so we
2943 * do an early lockdep release here:
2945 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2946 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2949 /* Here we just switch the register state and the stack. */
2950 switch_to(prev, next, prev);
2954 * this_rq must be evaluated again because prev may have moved
2955 * CPUs since it called schedule(), thus the 'rq' on its stack
2956 * frame will be invalid.
2958 finish_task_switch(this_rq(), prev);
2962 * nr_running, nr_uninterruptible and nr_context_switches:
2964 * externally visible scheduler statistics: current number of runnable
2965 * threads, current number of uninterruptible-sleeping threads, total
2966 * number of context switches performed since bootup.
2968 unsigned long nr_running(void)
2970 unsigned long i, sum = 0;
2972 for_each_online_cpu(i)
2973 sum += cpu_rq(i)->nr_running;
2978 unsigned long nr_uninterruptible(void)
2980 unsigned long i, sum = 0;
2982 for_each_possible_cpu(i)
2983 sum += cpu_rq(i)->nr_uninterruptible;
2986 * Since we read the counters lockless, it might be slightly
2987 * inaccurate. Do not allow it to go below zero though:
2989 if (unlikely((long)sum < 0))
2995 unsigned long long nr_context_switches(void)
2998 unsigned long long sum = 0;
3000 for_each_possible_cpu(i)
3001 sum += cpu_rq(i)->nr_switches;
3006 unsigned long nr_iowait(void)
3008 unsigned long i, sum = 0;
3010 for_each_possible_cpu(i)
3011 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3016 /* Variables and functions for calc_load */
3017 static atomic_long_t calc_load_tasks;
3018 static unsigned long calc_load_update;
3019 unsigned long avenrun[3];
3020 EXPORT_SYMBOL(avenrun);
3023 * get_avenrun - get the load average array
3024 * @loads: pointer to dest load array
3025 * @offset: offset to add
3026 * @shift: shift count to shift the result left
3028 * These values are estimates at best, so no need for locking.
3030 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3032 loads[0] = (avenrun[0] + offset) << shift;
3033 loads[1] = (avenrun[1] + offset) << shift;
3034 loads[2] = (avenrun[2] + offset) << shift;
3037 static unsigned long
3038 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3041 load += active * (FIXED_1 - exp);
3042 return load >> FSHIFT;
3046 * calc_load - update the avenrun load estimates 10 ticks after the
3047 * CPUs have updated calc_load_tasks.
3049 void calc_global_load(void)
3051 unsigned long upd = calc_load_update + 10;
3054 if (time_before(jiffies, upd))
3057 active = atomic_long_read(&calc_load_tasks);
3058 active = active > 0 ? active * FIXED_1 : 0;
3060 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3061 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3062 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3064 calc_load_update += LOAD_FREQ;
3068 * Either called from update_cpu_load() or from a cpu going idle
3070 static void calc_load_account_active(struct rq *this_rq)
3072 long nr_active, delta;
3074 nr_active = this_rq->nr_running;
3075 nr_active += (long) this_rq->nr_uninterruptible;
3077 if (nr_active != this_rq->calc_load_active) {
3078 delta = nr_active - this_rq->calc_load_active;
3079 this_rq->calc_load_active = nr_active;
3080 atomic_long_add(delta, &calc_load_tasks);
3085 * Externally visible per-cpu scheduler statistics:
3086 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3088 u64 cpu_nr_migrations(int cpu)
3090 return cpu_rq(cpu)->nr_migrations_in;
3094 * Update rq->cpu_load[] statistics. This function is usually called every
3095 * scheduler tick (TICK_NSEC).
3097 static void update_cpu_load(struct rq *this_rq)
3099 unsigned long this_load = this_rq->load.weight;
3102 this_rq->nr_load_updates++;
3104 /* Update our load: */
3105 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3106 unsigned long old_load, new_load;
3108 /* scale is effectively 1 << i now, and >> i divides by scale */
3110 old_load = this_rq->cpu_load[i];
3111 new_load = this_load;
3113 * Round up the averaging division if load is increasing. This
3114 * prevents us from getting stuck on 9 if the load is 10, for
3117 if (new_load > old_load)
3118 new_load += scale-1;
3119 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3122 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3123 this_rq->calc_load_update += LOAD_FREQ;
3124 calc_load_account_active(this_rq);
3131 * double_rq_lock - safely lock two runqueues
3133 * Note this does not disable interrupts like task_rq_lock,
3134 * you need to do so manually before calling.
3136 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3137 __acquires(rq1->lock)
3138 __acquires(rq2->lock)
3140 BUG_ON(!irqs_disabled());
3142 spin_lock(&rq1->lock);
3143 __acquire(rq2->lock); /* Fake it out ;) */
3146 spin_lock(&rq1->lock);
3147 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3149 spin_lock(&rq2->lock);
3150 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3153 update_rq_clock(rq1);
3154 update_rq_clock(rq2);
3158 * double_rq_unlock - safely unlock two runqueues
3160 * Note this does not restore interrupts like task_rq_unlock,
3161 * you need to do so manually after calling.
3163 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3164 __releases(rq1->lock)
3165 __releases(rq2->lock)
3167 spin_unlock(&rq1->lock);
3169 spin_unlock(&rq2->lock);
3171 __release(rq2->lock);
3175 * If dest_cpu is allowed for this process, migrate the task to it.
3176 * This is accomplished by forcing the cpu_allowed mask to only
3177 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3178 * the cpu_allowed mask is restored.
3180 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3182 struct migration_req req;
3183 unsigned long flags;
3186 rq = task_rq_lock(p, &flags);
3187 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3188 || unlikely(!cpu_active(dest_cpu)))
3191 /* force the process onto the specified CPU */
3192 if (migrate_task(p, dest_cpu, &req)) {
3193 /* Need to wait for migration thread (might exit: take ref). */
3194 struct task_struct *mt = rq->migration_thread;
3196 get_task_struct(mt);
3197 task_rq_unlock(rq, &flags);
3198 wake_up_process(mt);
3199 put_task_struct(mt);
3200 wait_for_completion(&req.done);
3205 task_rq_unlock(rq, &flags);
3209 * sched_exec - execve() is a valuable balancing opportunity, because at
3210 * this point the task has the smallest effective memory and cache footprint.
3212 void sched_exec(void)
3214 int new_cpu, this_cpu = get_cpu();
3215 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3217 if (new_cpu != this_cpu)
3218 sched_migrate_task(current, new_cpu);
3222 * pull_task - move a task from a remote runqueue to the local runqueue.
3223 * Both runqueues must be locked.
3225 static void pull_task(struct rq *src_rq, struct task_struct *p,
3226 struct rq *this_rq, int this_cpu)
3228 deactivate_task(src_rq, p, 0);
3229 set_task_cpu(p, this_cpu);
3230 activate_task(this_rq, p, 0);
3232 * Note that idle threads have a prio of MAX_PRIO, for this test
3233 * to be always true for them.
3235 check_preempt_curr(this_rq, p, 0);
3239 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3242 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3243 struct sched_domain *sd, enum cpu_idle_type idle,
3246 int tsk_cache_hot = 0;
3248 * We do not migrate tasks that are:
3249 * 1) running (obviously), or
3250 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3251 * 3) are cache-hot on their current CPU.
3253 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3254 schedstat_inc(p, se.nr_failed_migrations_affine);
3259 if (task_running(rq, p)) {
3260 schedstat_inc(p, se.nr_failed_migrations_running);
3265 * Aggressive migration if:
3266 * 1) task is cache cold, or
3267 * 2) too many balance attempts have failed.
3270 tsk_cache_hot = task_hot(p, rq->clock, sd);
3271 if (!tsk_cache_hot ||
3272 sd->nr_balance_failed > sd->cache_nice_tries) {
3273 #ifdef CONFIG_SCHEDSTATS
3274 if (tsk_cache_hot) {
3275 schedstat_inc(sd, lb_hot_gained[idle]);
3276 schedstat_inc(p, se.nr_forced_migrations);
3282 if (tsk_cache_hot) {
3283 schedstat_inc(p, se.nr_failed_migrations_hot);
3289 static unsigned long
3290 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3291 unsigned long max_load_move, struct sched_domain *sd,
3292 enum cpu_idle_type idle, int *all_pinned,
3293 int *this_best_prio, struct rq_iterator *iterator)
3295 int loops = 0, pulled = 0, pinned = 0;
3296 struct task_struct *p;
3297 long rem_load_move = max_load_move;
3299 if (max_load_move == 0)
3305 * Start the load-balancing iterator:
3307 p = iterator->start(iterator->arg);
3309 if (!p || loops++ > sysctl_sched_nr_migrate)
3312 if ((p->se.load.weight >> 1) > rem_load_move ||
3313 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3314 p = iterator->next(iterator->arg);
3318 pull_task(busiest, p, this_rq, this_cpu);
3320 rem_load_move -= p->se.load.weight;
3322 #ifdef CONFIG_PREEMPT
3324 * NEWIDLE balancing is a source of latency, so preemptible kernels
3325 * will stop after the first task is pulled to minimize the critical
3328 if (idle == CPU_NEWLY_IDLE)
3333 * We only want to steal up to the prescribed amount of weighted load.
3335 if (rem_load_move > 0) {
3336 if (p->prio < *this_best_prio)
3337 *this_best_prio = p->prio;
3338 p = iterator->next(iterator->arg);
3343 * Right now, this is one of only two places pull_task() is called,
3344 * so we can safely collect pull_task() stats here rather than
3345 * inside pull_task().
3347 schedstat_add(sd, lb_gained[idle], pulled);
3350 *all_pinned = pinned;
3352 return max_load_move - rem_load_move;
3356 * move_tasks tries to move up to max_load_move weighted load from busiest to
3357 * this_rq, as part of a balancing operation within domain "sd".
3358 * Returns 1 if successful and 0 otherwise.
3360 * Called with both runqueues locked.
3362 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3363 unsigned long max_load_move,
3364 struct sched_domain *sd, enum cpu_idle_type idle,
3367 const struct sched_class *class = sched_class_highest;
3368 unsigned long total_load_moved = 0;
3369 int this_best_prio = this_rq->curr->prio;
3373 class->load_balance(this_rq, this_cpu, busiest,
3374 max_load_move - total_load_moved,
3375 sd, idle, all_pinned, &this_best_prio);
3376 class = class->next;
3378 #ifdef CONFIG_PREEMPT
3380 * NEWIDLE balancing is a source of latency, so preemptible
3381 * kernels will stop after the first task is pulled to minimize
3382 * the critical section.
3384 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3387 } while (class && max_load_move > total_load_moved);
3389 return total_load_moved > 0;
3393 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3394 struct sched_domain *sd, enum cpu_idle_type idle,
3395 struct rq_iterator *iterator)
3397 struct task_struct *p = iterator->start(iterator->arg);
3401 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3402 pull_task(busiest, p, this_rq, this_cpu);
3404 * Right now, this is only the second place pull_task()
3405 * is called, so we can safely collect pull_task()
3406 * stats here rather than inside pull_task().
3408 schedstat_inc(sd, lb_gained[idle]);
3412 p = iterator->next(iterator->arg);
3419 * move_one_task tries to move exactly one task from busiest to this_rq, as
3420 * part of active balancing operations within "domain".
3421 * Returns 1 if successful and 0 otherwise.
3423 * Called with both runqueues locked.
3425 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3426 struct sched_domain *sd, enum cpu_idle_type idle)
3428 const struct sched_class *class;
3430 for (class = sched_class_highest; class; class = class->next)
3431 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3436 /********** Helpers for find_busiest_group ************************/
3438 * sd_lb_stats - Structure to store the statistics of a sched_domain
3439 * during load balancing.
3441 struct sd_lb_stats {
3442 struct sched_group *busiest; /* Busiest group in this sd */
3443 struct sched_group *this; /* Local group in this sd */
3444 unsigned long total_load; /* Total load of all groups in sd */
3445 unsigned long total_pwr; /* Total power of all groups in sd */
3446 unsigned long avg_load; /* Average load across all groups in sd */
3448 /** Statistics of this group */
3449 unsigned long this_load;
3450 unsigned long this_load_per_task;
3451 unsigned long this_nr_running;
3453 /* Statistics of the busiest group */
3454 unsigned long max_load;
3455 unsigned long busiest_load_per_task;
3456 unsigned long busiest_nr_running;
3458 int group_imb; /* Is there imbalance in this sd */
3459 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3460 int power_savings_balance; /* Is powersave balance needed for this sd */
3461 struct sched_group *group_min; /* Least loaded group in sd */
3462 struct sched_group *group_leader; /* Group which relieves group_min */
3463 unsigned long min_load_per_task; /* load_per_task in group_min */
3464 unsigned long leader_nr_running; /* Nr running of group_leader */
3465 unsigned long min_nr_running; /* Nr running of group_min */
3470 * sg_lb_stats - stats of a sched_group required for load_balancing
3472 struct sg_lb_stats {
3473 unsigned long avg_load; /*Avg load across the CPUs of the group */
3474 unsigned long group_load; /* Total load over the CPUs of the group */
3475 unsigned long sum_nr_running; /* Nr tasks running in the group */
3476 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3477 unsigned long group_capacity;
3478 int group_imb; /* Is there an imbalance in the group ? */
3482 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3483 * @group: The group whose first cpu is to be returned.
3485 static inline unsigned int group_first_cpu(struct sched_group *group)
3487 return cpumask_first(sched_group_cpus(group));
3491 * get_sd_load_idx - Obtain the load index for a given sched domain.
3492 * @sd: The sched_domain whose load_idx is to be obtained.
3493 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3495 static inline int get_sd_load_idx(struct sched_domain *sd,
3496 enum cpu_idle_type idle)
3502 load_idx = sd->busy_idx;
3505 case CPU_NEWLY_IDLE:
3506 load_idx = sd->newidle_idx;
3509 load_idx = sd->idle_idx;
3517 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3519 * init_sd_power_savings_stats - Initialize power savings statistics for
3520 * the given sched_domain, during load balancing.
3522 * @sd: Sched domain whose power-savings statistics are to be initialized.
3523 * @sds: Variable containing the statistics for sd.
3524 * @idle: Idle status of the CPU at which we're performing load-balancing.
3526 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3527 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3530 * Busy processors will not participate in power savings
3533 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3534 sds->power_savings_balance = 0;
3536 sds->power_savings_balance = 1;
3537 sds->min_nr_running = ULONG_MAX;
3538 sds->leader_nr_running = 0;
3543 * update_sd_power_savings_stats - Update the power saving stats for a
3544 * sched_domain while performing load balancing.
3546 * @group: sched_group belonging to the sched_domain under consideration.
3547 * @sds: Variable containing the statistics of the sched_domain
3548 * @local_group: Does group contain the CPU for which we're performing
3550 * @sgs: Variable containing the statistics of the group.
3552 static inline void update_sd_power_savings_stats(struct sched_group *group,
3553 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3556 if (!sds->power_savings_balance)
3560 * If the local group is idle or completely loaded
3561 * no need to do power savings balance at this domain
3563 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3564 !sds->this_nr_running))
3565 sds->power_savings_balance = 0;
3568 * If a group is already running at full capacity or idle,
3569 * don't include that group in power savings calculations
3571 if (!sds->power_savings_balance ||
3572 sgs->sum_nr_running >= sgs->group_capacity ||
3573 !sgs->sum_nr_running)
3577 * Calculate the group which has the least non-idle load.
3578 * This is the group from where we need to pick up the load
3581 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3582 (sgs->sum_nr_running == sds->min_nr_running &&
3583 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3584 sds->group_min = group;
3585 sds->min_nr_running = sgs->sum_nr_running;
3586 sds->min_load_per_task = sgs->sum_weighted_load /
3587 sgs->sum_nr_running;
3591 * Calculate the group which is almost near its
3592 * capacity but still has some space to pick up some load
3593 * from other group and save more power
3595 if (sgs->sum_nr_running > sgs->group_capacity - 1)
3598 if (sgs->sum_nr_running > sds->leader_nr_running ||
3599 (sgs->sum_nr_running == sds->leader_nr_running &&
3600 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3601 sds->group_leader = group;
3602 sds->leader_nr_running = sgs->sum_nr_running;
3607 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3608 * @sds: Variable containing the statistics of the sched_domain
3609 * under consideration.
3610 * @this_cpu: Cpu at which we're currently performing load-balancing.
3611 * @imbalance: Variable to store the imbalance.
3614 * Check if we have potential to perform some power-savings balance.
3615 * If yes, set the busiest group to be the least loaded group in the
3616 * sched_domain, so that it's CPUs can be put to idle.
3618 * Returns 1 if there is potential to perform power-savings balance.
3621 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3622 int this_cpu, unsigned long *imbalance)
3624 if (!sds->power_savings_balance)
3627 if (sds->this != sds->group_leader ||
3628 sds->group_leader == sds->group_min)
3631 *imbalance = sds->min_load_per_task;
3632 sds->busiest = sds->group_min;
3634 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3635 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3636 group_first_cpu(sds->group_leader);
3642 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3643 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3644 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3649 static inline void update_sd_power_savings_stats(struct sched_group *group,
3650 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3655 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3656 int this_cpu, unsigned long *imbalance)
3660 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3664 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3665 * @group: sched_group whose statistics are to be updated.
3666 * @this_cpu: Cpu for which load balance is currently performed.
3667 * @idle: Idle status of this_cpu
3668 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3669 * @sd_idle: Idle status of the sched_domain containing group.
3670 * @local_group: Does group contain this_cpu.
3671 * @cpus: Set of cpus considered for load balancing.
3672 * @balance: Should we balance.
3673 * @sgs: variable to hold the statistics for this group.
3675 static inline void update_sg_lb_stats(struct sched_group *group, int this_cpu,
3676 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3677 int local_group, const struct cpumask *cpus,
3678 int *balance, struct sg_lb_stats *sgs)
3680 unsigned long load, max_cpu_load, min_cpu_load;
3682 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3683 unsigned long sum_avg_load_per_task;
3684 unsigned long avg_load_per_task;
3687 balance_cpu = group_first_cpu(group);
3689 /* Tally up the load of all CPUs in the group */
3690 sum_avg_load_per_task = avg_load_per_task = 0;
3692 min_cpu_load = ~0UL;
3694 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3695 struct rq *rq = cpu_rq(i);
3697 if (*sd_idle && rq->nr_running)
3700 /* Bias balancing toward cpus of our domain */
3702 if (idle_cpu(i) && !first_idle_cpu) {
3707 load = target_load(i, load_idx);
3709 load = source_load(i, load_idx);
3710 if (load > max_cpu_load)
3711 max_cpu_load = load;
3712 if (min_cpu_load > load)
3713 min_cpu_load = load;
3716 sgs->group_load += load;
3717 sgs->sum_nr_running += rq->nr_running;
3718 sgs->sum_weighted_load += weighted_cpuload(i);
3720 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3724 * First idle cpu or the first cpu(busiest) in this sched group
3725 * is eligible for doing load balancing at this and above
3726 * domains. In the newly idle case, we will allow all the cpu's
3727 * to do the newly idle load balance.
3729 if (idle != CPU_NEWLY_IDLE && local_group &&
3730 balance_cpu != this_cpu && balance) {
3735 /* Adjust by relative CPU power of the group */
3736 sgs->avg_load = sg_div_cpu_power(group,
3737 sgs->group_load * SCHED_LOAD_SCALE);
3741 * Consider the group unbalanced when the imbalance is larger
3742 * than the average weight of two tasks.
3744 * APZ: with cgroup the avg task weight can vary wildly and
3745 * might not be a suitable number - should we keep a
3746 * normalized nr_running number somewhere that negates
3749 avg_load_per_task = sg_div_cpu_power(group,
3750 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3752 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3755 sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3760 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3761 * @sd: sched_domain whose statistics are to be updated.
3762 * @this_cpu: Cpu for which load balance is currently performed.
3763 * @idle: Idle status of this_cpu
3764 * @sd_idle: Idle status of the sched_domain containing group.
3765 * @cpus: Set of cpus considered for load balancing.
3766 * @balance: Should we balance.
3767 * @sds: variable to hold the statistics for this sched_domain.
3769 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3770 enum cpu_idle_type idle, int *sd_idle,
3771 const struct cpumask *cpus, int *balance,
3772 struct sd_lb_stats *sds)
3774 struct sched_group *group = sd->groups;
3775 struct sg_lb_stats sgs;
3778 init_sd_power_savings_stats(sd, sds, idle);
3779 load_idx = get_sd_load_idx(sd, idle);
3784 local_group = cpumask_test_cpu(this_cpu,
3785 sched_group_cpus(group));
3786 memset(&sgs, 0, sizeof(sgs));
3787 update_sg_lb_stats(group, this_cpu, idle, load_idx, sd_idle,
3788 local_group, cpus, balance, &sgs);
3790 if (local_group && balance && !(*balance))
3793 sds->total_load += sgs.group_load;
3794 sds->total_pwr += group->__cpu_power;
3797 sds->this_load = sgs.avg_load;
3799 sds->this_nr_running = sgs.sum_nr_running;
3800 sds->this_load_per_task = sgs.sum_weighted_load;
3801 } else if (sgs.avg_load > sds->max_load &&
3802 (sgs.sum_nr_running > sgs.group_capacity ||
3804 sds->max_load = sgs.avg_load;
3805 sds->busiest = group;
3806 sds->busiest_nr_running = sgs.sum_nr_running;
3807 sds->busiest_load_per_task = sgs.sum_weighted_load;
3808 sds->group_imb = sgs.group_imb;
3811 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3812 group = group->next;
3813 } while (group != sd->groups);
3818 * fix_small_imbalance - Calculate the minor imbalance that exists
3819 * amongst the groups of a sched_domain, during
3821 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3822 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3823 * @imbalance: Variable to store the imbalance.
3825 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3826 int this_cpu, unsigned long *imbalance)
3828 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3829 unsigned int imbn = 2;
3831 if (sds->this_nr_running) {
3832 sds->this_load_per_task /= sds->this_nr_running;
3833 if (sds->busiest_load_per_task >
3834 sds->this_load_per_task)
3837 sds->this_load_per_task =
3838 cpu_avg_load_per_task(this_cpu);
3840 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3841 sds->busiest_load_per_task * imbn) {
3842 *imbalance = sds->busiest_load_per_task;
3847 * OK, we don't have enough imbalance to justify moving tasks,
3848 * however we may be able to increase total CPU power used by
3852 pwr_now += sds->busiest->__cpu_power *
3853 min(sds->busiest_load_per_task, sds->max_load);
3854 pwr_now += sds->this->__cpu_power *
3855 min(sds->this_load_per_task, sds->this_load);
3856 pwr_now /= SCHED_LOAD_SCALE;
3858 /* Amount of load we'd subtract */
3859 tmp = sg_div_cpu_power(sds->busiest,
3860 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3861 if (sds->max_load > tmp)
3862 pwr_move += sds->busiest->__cpu_power *
3863 min(sds->busiest_load_per_task, sds->max_load - tmp);
3865 /* Amount of load we'd add */
3866 if (sds->max_load * sds->busiest->__cpu_power <
3867 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3868 tmp = sg_div_cpu_power(sds->this,
3869 sds->max_load * sds->busiest->__cpu_power);
3871 tmp = sg_div_cpu_power(sds->this,
3872 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3873 pwr_move += sds->this->__cpu_power *
3874 min(sds->this_load_per_task, sds->this_load + tmp);
3875 pwr_move /= SCHED_LOAD_SCALE;
3877 /* Move if we gain throughput */
3878 if (pwr_move > pwr_now)
3879 *imbalance = sds->busiest_load_per_task;
3883 * calculate_imbalance - Calculate the amount of imbalance present within the
3884 * groups of a given sched_domain during load balance.
3885 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3886 * @this_cpu: Cpu for which currently load balance is being performed.
3887 * @imbalance: The variable to store the imbalance.
3889 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3890 unsigned long *imbalance)
3892 unsigned long max_pull;
3894 * In the presence of smp nice balancing, certain scenarios can have
3895 * max load less than avg load(as we skip the groups at or below
3896 * its cpu_power, while calculating max_load..)
3898 if (sds->max_load < sds->avg_load) {
3900 return fix_small_imbalance(sds, this_cpu, imbalance);
3903 /* Don't want to pull so many tasks that a group would go idle */
3904 max_pull = min(sds->max_load - sds->avg_load,
3905 sds->max_load - sds->busiest_load_per_task);
3907 /* How much load to actually move to equalise the imbalance */
3908 *imbalance = min(max_pull * sds->busiest->__cpu_power,
3909 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
3913 * if *imbalance is less than the average load per runnable task
3914 * there is no gaurantee that any tasks will be moved so we'll have
3915 * a think about bumping its value to force at least one task to be
3918 if (*imbalance < sds->busiest_load_per_task)
3919 return fix_small_imbalance(sds, this_cpu, imbalance);
3922 /******* find_busiest_group() helpers end here *********************/
3925 * find_busiest_group - Returns the busiest group within the sched_domain
3926 * if there is an imbalance. If there isn't an imbalance, and
3927 * the user has opted for power-savings, it returns a group whose
3928 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3929 * such a group exists.
3931 * Also calculates the amount of weighted load which should be moved
3932 * to restore balance.
3934 * @sd: The sched_domain whose busiest group is to be returned.
3935 * @this_cpu: The cpu for which load balancing is currently being performed.
3936 * @imbalance: Variable which stores amount of weighted load which should
3937 * be moved to restore balance/put a group to idle.
3938 * @idle: The idle status of this_cpu.
3939 * @sd_idle: The idleness of sd
3940 * @cpus: The set of CPUs under consideration for load-balancing.
3941 * @balance: Pointer to a variable indicating if this_cpu
3942 * is the appropriate cpu to perform load balancing at this_level.
3944 * Returns: - the busiest group if imbalance exists.
3945 * - If no imbalance and user has opted for power-savings balance,
3946 * return the least loaded group whose CPUs can be
3947 * put to idle by rebalancing its tasks onto our group.
3949 static struct sched_group *
3950 find_busiest_group(struct sched_domain *sd, int this_cpu,
3951 unsigned long *imbalance, enum cpu_idle_type idle,
3952 int *sd_idle, const struct cpumask *cpus, int *balance)
3954 struct sd_lb_stats sds;
3956 memset(&sds, 0, sizeof(sds));
3959 * Compute the various statistics relavent for load balancing at
3962 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3965 /* Cases where imbalance does not exist from POV of this_cpu */
3966 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3968 * 2) There is no busy sibling group to pull from.
3969 * 3) This group is the busiest group.
3970 * 4) This group is more busy than the avg busieness at this
3972 * 5) The imbalance is within the specified limit.
3973 * 6) Any rebalance would lead to ping-pong
3975 if (balance && !(*balance))
3978 if (!sds.busiest || sds.busiest_nr_running == 0)
3981 if (sds.this_load >= sds.max_load)
3984 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3986 if (sds.this_load >= sds.avg_load)
3989 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3992 sds.busiest_load_per_task /= sds.busiest_nr_running;
3994 sds.busiest_load_per_task =
3995 min(sds.busiest_load_per_task, sds.avg_load);
3998 * We're trying to get all the cpus to the average_load, so we don't
3999 * want to push ourselves above the average load, nor do we wish to
4000 * reduce the max loaded cpu below the average load, as either of these
4001 * actions would just result in more rebalancing later, and ping-pong
4002 * tasks around. Thus we look for the minimum possible imbalance.
4003 * Negative imbalances (*we* are more loaded than anyone else) will
4004 * be counted as no imbalance for these purposes -- we can't fix that
4005 * by pulling tasks to us. Be careful of negative numbers as they'll
4006 * appear as very large values with unsigned longs.
4008 if (sds.max_load <= sds.busiest_load_per_task)
4011 /* Looks like there is an imbalance. Compute it */
4012 calculate_imbalance(&sds, this_cpu, imbalance);
4017 * There is no obvious imbalance. But check if we can do some balancing
4020 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4028 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4031 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4032 unsigned long imbalance, const struct cpumask *cpus)
4034 struct rq *busiest = NULL, *rq;
4035 unsigned long max_load = 0;
4038 for_each_cpu(i, sched_group_cpus(group)) {
4041 if (!cpumask_test_cpu(i, cpus))
4045 wl = weighted_cpuload(i);
4047 if (rq->nr_running == 1 && wl > imbalance)
4050 if (wl > max_load) {
4060 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4061 * so long as it is large enough.
4063 #define MAX_PINNED_INTERVAL 512
4065 /* Working cpumask for load_balance and load_balance_newidle. */
4066 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4069 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4070 * tasks if there is an imbalance.
4072 static int load_balance(int this_cpu, struct rq *this_rq,
4073 struct sched_domain *sd, enum cpu_idle_type idle,
4076 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4077 struct sched_group *group;
4078 unsigned long imbalance;
4080 unsigned long flags;
4081 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4083 cpumask_setall(cpus);
4086 * When power savings policy is enabled for the parent domain, idle
4087 * sibling can pick up load irrespective of busy siblings. In this case,
4088 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4089 * portraying it as CPU_NOT_IDLE.
4091 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4092 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4095 schedstat_inc(sd, lb_count[idle]);
4099 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4106 schedstat_inc(sd, lb_nobusyg[idle]);
4110 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4112 schedstat_inc(sd, lb_nobusyq[idle]);
4116 BUG_ON(busiest == this_rq);
4118 schedstat_add(sd, lb_imbalance[idle], imbalance);
4121 if (busiest->nr_running > 1) {
4123 * Attempt to move tasks. If find_busiest_group has found
4124 * an imbalance but busiest->nr_running <= 1, the group is
4125 * still unbalanced. ld_moved simply stays zero, so it is
4126 * correctly treated as an imbalance.
4128 local_irq_save(flags);
4129 double_rq_lock(this_rq, busiest);
4130 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4131 imbalance, sd, idle, &all_pinned);
4132 double_rq_unlock(this_rq, busiest);
4133 local_irq_restore(flags);
4136 * some other cpu did the load balance for us.
4138 if (ld_moved && this_cpu != smp_processor_id())
4139 resched_cpu(this_cpu);
4141 /* All tasks on this runqueue were pinned by CPU affinity */
4142 if (unlikely(all_pinned)) {
4143 cpumask_clear_cpu(cpu_of(busiest), cpus);
4144 if (!cpumask_empty(cpus))
4151 schedstat_inc(sd, lb_failed[idle]);
4152 sd->nr_balance_failed++;
4154 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4156 spin_lock_irqsave(&busiest->lock, flags);
4158 /* don't kick the migration_thread, if the curr
4159 * task on busiest cpu can't be moved to this_cpu
4161 if (!cpumask_test_cpu(this_cpu,
4162 &busiest->curr->cpus_allowed)) {
4163 spin_unlock_irqrestore(&busiest->lock, flags);
4165 goto out_one_pinned;
4168 if (!busiest->active_balance) {
4169 busiest->active_balance = 1;
4170 busiest->push_cpu = this_cpu;
4173 spin_unlock_irqrestore(&busiest->lock, flags);
4175 wake_up_process(busiest->migration_thread);
4178 * We've kicked active balancing, reset the failure
4181 sd->nr_balance_failed = sd->cache_nice_tries+1;
4184 sd->nr_balance_failed = 0;
4186 if (likely(!active_balance)) {
4187 /* We were unbalanced, so reset the balancing interval */
4188 sd->balance_interval = sd->min_interval;
4191 * If we've begun active balancing, start to back off. This
4192 * case may not be covered by the all_pinned logic if there
4193 * is only 1 task on the busy runqueue (because we don't call
4196 if (sd->balance_interval < sd->max_interval)
4197 sd->balance_interval *= 2;
4200 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4201 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4207 schedstat_inc(sd, lb_balanced[idle]);
4209 sd->nr_balance_failed = 0;
4212 /* tune up the balancing interval */
4213 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4214 (sd->balance_interval < sd->max_interval))
4215 sd->balance_interval *= 2;
4217 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4218 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4229 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4230 * tasks if there is an imbalance.
4232 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4233 * this_rq is locked.
4236 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4238 struct sched_group *group;
4239 struct rq *busiest = NULL;
4240 unsigned long imbalance;
4244 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4246 cpumask_setall(cpus);
4249 * When power savings policy is enabled for the parent domain, idle
4250 * sibling can pick up load irrespective of busy siblings. In this case,
4251 * let the state of idle sibling percolate up as IDLE, instead of
4252 * portraying it as CPU_NOT_IDLE.
4254 if (sd->flags & SD_SHARE_CPUPOWER &&
4255 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4258 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4260 update_shares_locked(this_rq, sd);
4261 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4262 &sd_idle, cpus, NULL);
4264 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4268 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4270 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4274 BUG_ON(busiest == this_rq);
4276 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4279 if (busiest->nr_running > 1) {
4280 /* Attempt to move tasks */
4281 double_lock_balance(this_rq, busiest);
4282 /* this_rq->clock is already updated */
4283 update_rq_clock(busiest);
4284 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4285 imbalance, sd, CPU_NEWLY_IDLE,
4287 double_unlock_balance(this_rq, busiest);
4289 if (unlikely(all_pinned)) {
4290 cpumask_clear_cpu(cpu_of(busiest), cpus);
4291 if (!cpumask_empty(cpus))
4297 int active_balance = 0;
4299 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4300 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4301 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4304 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4307 if (sd->nr_balance_failed++ < 2)
4311 * The only task running in a non-idle cpu can be moved to this
4312 * cpu in an attempt to completely freeup the other CPU
4313 * package. The same method used to move task in load_balance()
4314 * have been extended for load_balance_newidle() to speedup
4315 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4317 * The package power saving logic comes from
4318 * find_busiest_group(). If there are no imbalance, then
4319 * f_b_g() will return NULL. However when sched_mc={1,2} then
4320 * f_b_g() will select a group from which a running task may be
4321 * pulled to this cpu in order to make the other package idle.
4322 * If there is no opportunity to make a package idle and if
4323 * there are no imbalance, then f_b_g() will return NULL and no
4324 * action will be taken in load_balance_newidle().
4326 * Under normal task pull operation due to imbalance, there
4327 * will be more than one task in the source run queue and
4328 * move_tasks() will succeed. ld_moved will be true and this
4329 * active balance code will not be triggered.
4332 /* Lock busiest in correct order while this_rq is held */
4333 double_lock_balance(this_rq, busiest);
4336 * don't kick the migration_thread, if the curr
4337 * task on busiest cpu can't be moved to this_cpu
4339 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4340 double_unlock_balance(this_rq, busiest);
4345 if (!busiest->active_balance) {
4346 busiest->active_balance = 1;
4347 busiest->push_cpu = this_cpu;
4351 double_unlock_balance(this_rq, busiest);
4353 * Should not call ttwu while holding a rq->lock
4355 spin_unlock(&this_rq->lock);
4357 wake_up_process(busiest->migration_thread);
4358 spin_lock(&this_rq->lock);
4361 sd->nr_balance_failed = 0;
4363 update_shares_locked(this_rq, sd);
4367 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4368 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4369 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4371 sd->nr_balance_failed = 0;
4377 * idle_balance is called by schedule() if this_cpu is about to become
4378 * idle. Attempts to pull tasks from other CPUs.
4380 static void idle_balance(int this_cpu, struct rq *this_rq)
4382 struct sched_domain *sd;
4383 int pulled_task = 0;
4384 unsigned long next_balance = jiffies + HZ;
4386 for_each_domain(this_cpu, sd) {
4387 unsigned long interval;
4389 if (!(sd->flags & SD_LOAD_BALANCE))
4392 if (sd->flags & SD_BALANCE_NEWIDLE)
4393 /* If we've pulled tasks over stop searching: */
4394 pulled_task = load_balance_newidle(this_cpu, this_rq,
4397 interval = msecs_to_jiffies(sd->balance_interval);
4398 if (time_after(next_balance, sd->last_balance + interval))
4399 next_balance = sd->last_balance + interval;
4403 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4405 * We are going idle. next_balance may be set based on
4406 * a busy processor. So reset next_balance.
4408 this_rq->next_balance = next_balance;
4413 * active_load_balance is run by migration threads. It pushes running tasks
4414 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4415 * running on each physical CPU where possible, and avoids physical /
4416 * logical imbalances.
4418 * Called with busiest_rq locked.
4420 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4422 int target_cpu = busiest_rq->push_cpu;
4423 struct sched_domain *sd;
4424 struct rq *target_rq;
4426 /* Is there any task to move? */
4427 if (busiest_rq->nr_running <= 1)
4430 target_rq = cpu_rq(target_cpu);
4433 * This condition is "impossible", if it occurs
4434 * we need to fix it. Originally reported by
4435 * Bjorn Helgaas on a 128-cpu setup.
4437 BUG_ON(busiest_rq == target_rq);
4439 /* move a task from busiest_rq to target_rq */
4440 double_lock_balance(busiest_rq, target_rq);
4441 update_rq_clock(busiest_rq);
4442 update_rq_clock(target_rq);
4444 /* Search for an sd spanning us and the target CPU. */
4445 for_each_domain(target_cpu, sd) {
4446 if ((sd->flags & SD_LOAD_BALANCE) &&
4447 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4452 schedstat_inc(sd, alb_count);
4454 if (move_one_task(target_rq, target_cpu, busiest_rq,
4456 schedstat_inc(sd, alb_pushed);
4458 schedstat_inc(sd, alb_failed);
4460 double_unlock_balance(busiest_rq, target_rq);
4465 atomic_t load_balancer;
4466 cpumask_var_t cpu_mask;
4467 cpumask_var_t ilb_grp_nohz_mask;
4468 } nohz ____cacheline_aligned = {
4469 .load_balancer = ATOMIC_INIT(-1),
4472 int get_nohz_load_balancer(void)
4474 return atomic_read(&nohz.load_balancer);
4477 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4479 * lowest_flag_domain - Return lowest sched_domain containing flag.
4480 * @cpu: The cpu whose lowest level of sched domain is to
4482 * @flag: The flag to check for the lowest sched_domain
4483 * for the given cpu.
4485 * Returns the lowest sched_domain of a cpu which contains the given flag.
4487 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4489 struct sched_domain *sd;
4491 for_each_domain(cpu, sd)
4492 if (sd && (sd->flags & flag))
4499 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4500 * @cpu: The cpu whose domains we're iterating over.
4501 * @sd: variable holding the value of the power_savings_sd
4503 * @flag: The flag to filter the sched_domains to be iterated.
4505 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4506 * set, starting from the lowest sched_domain to the highest.
4508 #define for_each_flag_domain(cpu, sd, flag) \
4509 for (sd = lowest_flag_domain(cpu, flag); \
4510 (sd && (sd->flags & flag)); sd = sd->parent)
4513 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4514 * @ilb_group: group to be checked for semi-idleness
4516 * Returns: 1 if the group is semi-idle. 0 otherwise.
4518 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4519 * and atleast one non-idle CPU. This helper function checks if the given
4520 * sched_group is semi-idle or not.
4522 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4524 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4525 sched_group_cpus(ilb_group));
4528 * A sched_group is semi-idle when it has atleast one busy cpu
4529 * and atleast one idle cpu.
4531 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4534 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4540 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4541 * @cpu: The cpu which is nominating a new idle_load_balancer.
4543 * Returns: Returns the id of the idle load balancer if it exists,
4544 * Else, returns >= nr_cpu_ids.
4546 * This algorithm picks the idle load balancer such that it belongs to a
4547 * semi-idle powersavings sched_domain. The idea is to try and avoid
4548 * completely idle packages/cores just for the purpose of idle load balancing
4549 * when there are other idle cpu's which are better suited for that job.
4551 static int find_new_ilb(int cpu)
4553 struct sched_domain *sd;
4554 struct sched_group *ilb_group;
4557 * Have idle load balancer selection from semi-idle packages only
4558 * when power-aware load balancing is enabled
4560 if (!(sched_smt_power_savings || sched_mc_power_savings))
4564 * Optimize for the case when we have no idle CPUs or only one
4565 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4567 if (cpumask_weight(nohz.cpu_mask) < 2)
4570 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4571 ilb_group = sd->groups;
4574 if (is_semi_idle_group(ilb_group))
4575 return cpumask_first(nohz.ilb_grp_nohz_mask);
4577 ilb_group = ilb_group->next;
4579 } while (ilb_group != sd->groups);
4583 return cpumask_first(nohz.cpu_mask);
4585 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4586 static inline int find_new_ilb(int call_cpu)
4588 return cpumask_first(nohz.cpu_mask);
4593 * This routine will try to nominate the ilb (idle load balancing)
4594 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4595 * load balancing on behalf of all those cpus. If all the cpus in the system
4596 * go into this tickless mode, then there will be no ilb owner (as there is
4597 * no need for one) and all the cpus will sleep till the next wakeup event
4600 * For the ilb owner, tick is not stopped. And this tick will be used
4601 * for idle load balancing. ilb owner will still be part of
4604 * While stopping the tick, this cpu will become the ilb owner if there
4605 * is no other owner. And will be the owner till that cpu becomes busy
4606 * or if all cpus in the system stop their ticks at which point
4607 * there is no need for ilb owner.
4609 * When the ilb owner becomes busy, it nominates another owner, during the
4610 * next busy scheduler_tick()
4612 int select_nohz_load_balancer(int stop_tick)
4614 int cpu = smp_processor_id();
4617 cpu_rq(cpu)->in_nohz_recently = 1;
4619 if (!cpu_active(cpu)) {
4620 if (atomic_read(&nohz.load_balancer) != cpu)
4624 * If we are going offline and still the leader,
4627 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4633 cpumask_set_cpu(cpu, nohz.cpu_mask);
4635 /* time for ilb owner also to sleep */
4636 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4637 if (atomic_read(&nohz.load_balancer) == cpu)
4638 atomic_set(&nohz.load_balancer, -1);
4642 if (atomic_read(&nohz.load_balancer) == -1) {
4643 /* make me the ilb owner */
4644 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4646 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4649 if (!(sched_smt_power_savings ||
4650 sched_mc_power_savings))
4653 * Check to see if there is a more power-efficient
4656 new_ilb = find_new_ilb(cpu);
4657 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4658 atomic_set(&nohz.load_balancer, -1);
4659 resched_cpu(new_ilb);
4665 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4668 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4670 if (atomic_read(&nohz.load_balancer) == cpu)
4671 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4678 static DEFINE_SPINLOCK(balancing);
4681 * It checks each scheduling domain to see if it is due to be balanced,
4682 * and initiates a balancing operation if so.
4684 * Balancing parameters are set up in arch_init_sched_domains.
4686 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4689 struct rq *rq = cpu_rq(cpu);
4690 unsigned long interval;
4691 struct sched_domain *sd;
4692 /* Earliest time when we have to do rebalance again */
4693 unsigned long next_balance = jiffies + 60*HZ;
4694 int update_next_balance = 0;
4697 for_each_domain(cpu, sd) {
4698 if (!(sd->flags & SD_LOAD_BALANCE))
4701 interval = sd->balance_interval;
4702 if (idle != CPU_IDLE)
4703 interval *= sd->busy_factor;
4705 /* scale ms to jiffies */
4706 interval = msecs_to_jiffies(interval);
4707 if (unlikely(!interval))
4709 if (interval > HZ*NR_CPUS/10)
4710 interval = HZ*NR_CPUS/10;
4712 need_serialize = sd->flags & SD_SERIALIZE;
4714 if (need_serialize) {
4715 if (!spin_trylock(&balancing))
4719 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4720 if (load_balance(cpu, rq, sd, idle, &balance)) {
4722 * We've pulled tasks over so either we're no
4723 * longer idle, or one of our SMT siblings is
4726 idle = CPU_NOT_IDLE;
4728 sd->last_balance = jiffies;
4731 spin_unlock(&balancing);
4733 if (time_after(next_balance, sd->last_balance + interval)) {
4734 next_balance = sd->last_balance + interval;
4735 update_next_balance = 1;
4739 * Stop the load balance at this level. There is another
4740 * CPU in our sched group which is doing load balancing more
4748 * next_balance will be updated only when there is a need.
4749 * When the cpu is attached to null domain for ex, it will not be
4752 if (likely(update_next_balance))
4753 rq->next_balance = next_balance;
4757 * run_rebalance_domains is triggered when needed from the scheduler tick.
4758 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4759 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4761 static void run_rebalance_domains(struct softirq_action *h)
4763 int this_cpu = smp_processor_id();
4764 struct rq *this_rq = cpu_rq(this_cpu);
4765 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4766 CPU_IDLE : CPU_NOT_IDLE;
4768 rebalance_domains(this_cpu, idle);
4772 * If this cpu is the owner for idle load balancing, then do the
4773 * balancing on behalf of the other idle cpus whose ticks are
4776 if (this_rq->idle_at_tick &&
4777 atomic_read(&nohz.load_balancer) == this_cpu) {
4781 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4782 if (balance_cpu == this_cpu)
4786 * If this cpu gets work to do, stop the load balancing
4787 * work being done for other cpus. Next load
4788 * balancing owner will pick it up.
4793 rebalance_domains(balance_cpu, CPU_IDLE);
4795 rq = cpu_rq(balance_cpu);
4796 if (time_after(this_rq->next_balance, rq->next_balance))
4797 this_rq->next_balance = rq->next_balance;
4803 static inline int on_null_domain(int cpu)
4805 return !rcu_dereference(cpu_rq(cpu)->sd);
4809 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4811 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4812 * idle load balancing owner or decide to stop the periodic load balancing,
4813 * if the whole system is idle.
4815 static inline void trigger_load_balance(struct rq *rq, int cpu)
4819 * If we were in the nohz mode recently and busy at the current
4820 * scheduler tick, then check if we need to nominate new idle
4823 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4824 rq->in_nohz_recently = 0;
4826 if (atomic_read(&nohz.load_balancer) == cpu) {
4827 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4828 atomic_set(&nohz.load_balancer, -1);
4831 if (atomic_read(&nohz.load_balancer) == -1) {
4832 int ilb = find_new_ilb(cpu);
4834 if (ilb < nr_cpu_ids)
4840 * If this cpu is idle and doing idle load balancing for all the
4841 * cpus with ticks stopped, is it time for that to stop?
4843 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4844 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4850 * If this cpu is idle and the idle load balancing is done by
4851 * someone else, then no need raise the SCHED_SOFTIRQ
4853 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4854 cpumask_test_cpu(cpu, nohz.cpu_mask))
4857 /* Don't need to rebalance while attached to NULL domain */
4858 if (time_after_eq(jiffies, rq->next_balance) &&
4859 likely(!on_null_domain(cpu)))
4860 raise_softirq(SCHED_SOFTIRQ);
4863 #else /* CONFIG_SMP */
4866 * on UP we do not need to balance between CPUs:
4868 static inline void idle_balance(int cpu, struct rq *rq)
4874 DEFINE_PER_CPU(struct kernel_stat, kstat);
4876 EXPORT_PER_CPU_SYMBOL(kstat);
4879 * Return any ns on the sched_clock that have not yet been accounted in
4880 * @p in case that task is currently running.
4882 * Called with task_rq_lock() held on @rq.
4884 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4888 if (task_current(rq, p)) {
4889 update_rq_clock(rq);
4890 ns = rq->clock - p->se.exec_start;
4898 unsigned long long task_delta_exec(struct task_struct *p)
4900 unsigned long flags;
4904 rq = task_rq_lock(p, &flags);
4905 ns = do_task_delta_exec(p, rq);
4906 task_rq_unlock(rq, &flags);
4912 * Return accounted runtime for the task.
4913 * In case the task is currently running, return the runtime plus current's
4914 * pending runtime that have not been accounted yet.
4916 unsigned long long task_sched_runtime(struct task_struct *p)
4918 unsigned long flags;
4922 rq = task_rq_lock(p, &flags);
4923 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4924 task_rq_unlock(rq, &flags);
4930 * Return sum_exec_runtime for the thread group.
4931 * In case the task is currently running, return the sum plus current's
4932 * pending runtime that have not been accounted yet.
4934 * Note that the thread group might have other running tasks as well,
4935 * so the return value not includes other pending runtime that other
4936 * running tasks might have.
4938 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4940 struct task_cputime totals;
4941 unsigned long flags;
4945 rq = task_rq_lock(p, &flags);
4946 thread_group_cputime(p, &totals);
4947 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4948 task_rq_unlock(rq, &flags);
4954 * Account user cpu time to a process.
4955 * @p: the process that the cpu time gets accounted to
4956 * @cputime: the cpu time spent in user space since the last update
4957 * @cputime_scaled: cputime scaled by cpu frequency
4959 void account_user_time(struct task_struct *p, cputime_t cputime,
4960 cputime_t cputime_scaled)
4962 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4965 /* Add user time to process. */
4966 p->utime = cputime_add(p->utime, cputime);
4967 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4968 account_group_user_time(p, cputime);
4970 /* Add user time to cpustat. */
4971 tmp = cputime_to_cputime64(cputime);
4972 if (TASK_NICE(p) > 0)
4973 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4975 cpustat->user = cputime64_add(cpustat->user, tmp);
4977 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4978 /* Account for user time used */
4979 acct_update_integrals(p);
4983 * Account guest cpu time to a process.
4984 * @p: the process that the cpu time gets accounted to
4985 * @cputime: the cpu time spent in virtual machine since the last update
4986 * @cputime_scaled: cputime scaled by cpu frequency
4988 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4989 cputime_t cputime_scaled)
4992 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4994 tmp = cputime_to_cputime64(cputime);
4996 /* Add guest time to process. */
4997 p->utime = cputime_add(p->utime, cputime);
4998 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4999 account_group_user_time(p, cputime);
5000 p->gtime = cputime_add(p->gtime, cputime);
5002 /* Add guest time to cpustat. */
5003 cpustat->user = cputime64_add(cpustat->user, tmp);
5004 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5008 * Account system cpu time to a process.
5009 * @p: the process that the cpu time gets accounted to
5010 * @hardirq_offset: the offset to subtract from hardirq_count()
5011 * @cputime: the cpu time spent in kernel space since the last update
5012 * @cputime_scaled: cputime scaled by cpu frequency
5014 void account_system_time(struct task_struct *p, int hardirq_offset,
5015 cputime_t cputime, cputime_t cputime_scaled)
5017 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5020 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5021 account_guest_time(p, cputime, cputime_scaled);
5025 /* Add system time to process. */
5026 p->stime = cputime_add(p->stime, cputime);
5027 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5028 account_group_system_time(p, cputime);
5030 /* Add system time to cpustat. */
5031 tmp = cputime_to_cputime64(cputime);
5032 if (hardirq_count() - hardirq_offset)
5033 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5034 else if (softirq_count())
5035 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5037 cpustat->system = cputime64_add(cpustat->system, tmp);
5039 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5041 /* Account for system time used */
5042 acct_update_integrals(p);
5046 * Account for involuntary wait time.
5047 * @steal: the cpu time spent in involuntary wait
5049 void account_steal_time(cputime_t cputime)
5051 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5052 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5054 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5058 * Account for idle time.
5059 * @cputime: the cpu time spent in idle wait
5061 void account_idle_time(cputime_t cputime)
5063 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5064 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5065 struct rq *rq = this_rq();
5067 if (atomic_read(&rq->nr_iowait) > 0)
5068 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5070 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5073 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5076 * Account a single tick of cpu time.
5077 * @p: the process that the cpu time gets accounted to
5078 * @user_tick: indicates if the tick is a user or a system tick
5080 void account_process_tick(struct task_struct *p, int user_tick)
5082 cputime_t one_jiffy = jiffies_to_cputime(1);
5083 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
5084 struct rq *rq = this_rq();
5087 account_user_time(p, one_jiffy, one_jiffy_scaled);
5088 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5089 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
5092 account_idle_time(one_jiffy);
5096 * Account multiple ticks of steal time.
5097 * @p: the process from which the cpu time has been stolen
5098 * @ticks: number of stolen ticks
5100 void account_steal_ticks(unsigned long ticks)
5102 account_steal_time(jiffies_to_cputime(ticks));
5106 * Account multiple ticks of idle time.
5107 * @ticks: number of stolen ticks
5109 void account_idle_ticks(unsigned long ticks)
5111 account_idle_time(jiffies_to_cputime(ticks));
5117 * Use precise platform statistics if available:
5119 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5120 cputime_t task_utime(struct task_struct *p)
5125 cputime_t task_stime(struct task_struct *p)
5130 cputime_t task_utime(struct task_struct *p)
5132 clock_t utime = cputime_to_clock_t(p->utime),
5133 total = utime + cputime_to_clock_t(p->stime);
5137 * Use CFS's precise accounting:
5139 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5143 do_div(temp, total);
5145 utime = (clock_t)temp;
5147 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5148 return p->prev_utime;
5151 cputime_t task_stime(struct task_struct *p)
5156 * Use CFS's precise accounting. (we subtract utime from
5157 * the total, to make sure the total observed by userspace
5158 * grows monotonically - apps rely on that):
5160 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5161 cputime_to_clock_t(task_utime(p));
5164 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5166 return p->prev_stime;
5170 inline cputime_t task_gtime(struct task_struct *p)
5176 * This function gets called by the timer code, with HZ frequency.
5177 * We call it with interrupts disabled.
5179 * It also gets called by the fork code, when changing the parent's
5182 void scheduler_tick(void)
5184 int cpu = smp_processor_id();
5185 struct rq *rq = cpu_rq(cpu);
5186 struct task_struct *curr = rq->curr;
5190 spin_lock(&rq->lock);
5191 update_rq_clock(rq);
5192 update_cpu_load(rq);
5193 curr->sched_class->task_tick(rq, curr, 0);
5194 spin_unlock(&rq->lock);
5196 perf_counter_task_tick(curr, cpu);
5199 rq->idle_at_tick = idle_cpu(cpu);
5200 trigger_load_balance(rq, cpu);
5204 notrace unsigned long get_parent_ip(unsigned long addr)
5206 if (in_lock_functions(addr)) {
5207 addr = CALLER_ADDR2;
5208 if (in_lock_functions(addr))
5209 addr = CALLER_ADDR3;
5214 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5215 defined(CONFIG_PREEMPT_TRACER))
5217 void __kprobes add_preempt_count(int val)
5219 #ifdef CONFIG_DEBUG_PREEMPT
5223 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5226 preempt_count() += val;
5227 #ifdef CONFIG_DEBUG_PREEMPT
5229 * Spinlock count overflowing soon?
5231 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5234 if (preempt_count() == val)
5235 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5237 EXPORT_SYMBOL(add_preempt_count);
5239 void __kprobes sub_preempt_count(int val)
5241 #ifdef CONFIG_DEBUG_PREEMPT
5245 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5248 * Is the spinlock portion underflowing?
5250 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5251 !(preempt_count() & PREEMPT_MASK)))
5255 if (preempt_count() == val)
5256 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5257 preempt_count() -= val;
5259 EXPORT_SYMBOL(sub_preempt_count);
5264 * Print scheduling while atomic bug:
5266 static noinline void __schedule_bug(struct task_struct *prev)
5268 struct pt_regs *regs = get_irq_regs();
5270 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5271 prev->comm, prev->pid, preempt_count());
5273 debug_show_held_locks(prev);
5275 if (irqs_disabled())
5276 print_irqtrace_events(prev);
5285 * Various schedule()-time debugging checks and statistics:
5287 static inline void schedule_debug(struct task_struct *prev)
5290 * Test if we are atomic. Since do_exit() needs to call into
5291 * schedule() atomically, we ignore that path for now.
5292 * Otherwise, whine if we are scheduling when we should not be.
5294 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5295 __schedule_bug(prev);
5297 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5299 schedstat_inc(this_rq(), sched_count);
5300 #ifdef CONFIG_SCHEDSTATS
5301 if (unlikely(prev->lock_depth >= 0)) {
5302 schedstat_inc(this_rq(), bkl_count);
5303 schedstat_inc(prev, sched_info.bkl_count);
5308 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5310 if (prev->state == TASK_RUNNING) {
5311 u64 runtime = prev->se.sum_exec_runtime;
5313 runtime -= prev->se.prev_sum_exec_runtime;
5314 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5317 * In order to avoid avg_overlap growing stale when we are
5318 * indeed overlapping and hence not getting put to sleep, grow
5319 * the avg_overlap on preemption.
5321 * We use the average preemption runtime because that
5322 * correlates to the amount of cache footprint a task can
5325 update_avg(&prev->se.avg_overlap, runtime);
5327 prev->sched_class->put_prev_task(rq, prev);
5331 * Pick up the highest-prio task:
5333 static inline struct task_struct *
5334 pick_next_task(struct rq *rq)
5336 const struct sched_class *class;
5337 struct task_struct *p;
5340 * Optimization: we know that if all tasks are in
5341 * the fair class we can call that function directly:
5343 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5344 p = fair_sched_class.pick_next_task(rq);
5349 class = sched_class_highest;
5351 p = class->pick_next_task(rq);
5355 * Will never be NULL as the idle class always
5356 * returns a non-NULL p:
5358 class = class->next;
5363 * schedule() is the main scheduler function.
5365 asmlinkage void __sched schedule(void)
5367 struct task_struct *prev, *next;
5368 unsigned long *switch_count;
5374 cpu = smp_processor_id();
5378 switch_count = &prev->nivcsw;
5380 release_kernel_lock(prev);
5381 need_resched_nonpreemptible:
5383 schedule_debug(prev);
5385 if (sched_feat(HRTICK))
5388 spin_lock_irq(&rq->lock);
5389 update_rq_clock(rq);
5390 clear_tsk_need_resched(prev);
5392 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5393 if (unlikely(signal_pending_state(prev->state, prev)))
5394 prev->state = TASK_RUNNING;
5396 deactivate_task(rq, prev, 1);
5397 switch_count = &prev->nvcsw;
5401 if (prev->sched_class->pre_schedule)
5402 prev->sched_class->pre_schedule(rq, prev);
5405 if (unlikely(!rq->nr_running))
5406 idle_balance(cpu, rq);
5408 put_prev_task(rq, prev);
5409 next = pick_next_task(rq);
5411 if (likely(prev != next)) {
5412 sched_info_switch(prev, next);
5413 perf_counter_task_sched_out(prev, next, cpu);
5419 context_switch(rq, prev, next); /* unlocks the rq */
5421 * the context switch might have flipped the stack from under
5422 * us, hence refresh the local variables.
5424 cpu = smp_processor_id();
5427 spin_unlock_irq(&rq->lock);
5429 if (unlikely(reacquire_kernel_lock(current) < 0))
5430 goto need_resched_nonpreemptible;
5432 preempt_enable_no_resched();
5436 EXPORT_SYMBOL(schedule);
5440 * Look out! "owner" is an entirely speculative pointer
5441 * access and not reliable.
5443 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5448 if (!sched_feat(OWNER_SPIN))
5451 #ifdef CONFIG_DEBUG_PAGEALLOC
5453 * Need to access the cpu field knowing that
5454 * DEBUG_PAGEALLOC could have unmapped it if
5455 * the mutex owner just released it and exited.
5457 if (probe_kernel_address(&owner->cpu, cpu))
5464 * Even if the access succeeded (likely case),
5465 * the cpu field may no longer be valid.
5467 if (cpu >= nr_cpumask_bits)
5471 * We need to validate that we can do a
5472 * get_cpu() and that we have the percpu area.
5474 if (!cpu_online(cpu))
5481 * Owner changed, break to re-assess state.
5483 if (lock->owner != owner)
5487 * Is that owner really running on that cpu?
5489 if (task_thread_info(rq->curr) != owner || need_resched())
5499 #ifdef CONFIG_PREEMPT
5501 * this is the entry point to schedule() from in-kernel preemption
5502 * off of preempt_enable. Kernel preemptions off return from interrupt
5503 * occur there and call schedule directly.
5505 asmlinkage void __sched preempt_schedule(void)
5507 struct thread_info *ti = current_thread_info();
5510 * If there is a non-zero preempt_count or interrupts are disabled,
5511 * we do not want to preempt the current task. Just return..
5513 if (likely(ti->preempt_count || irqs_disabled()))
5517 add_preempt_count(PREEMPT_ACTIVE);
5519 sub_preempt_count(PREEMPT_ACTIVE);
5522 * Check again in case we missed a preemption opportunity
5523 * between schedule and now.
5526 } while (need_resched());
5528 EXPORT_SYMBOL(preempt_schedule);
5531 * this is the entry point to schedule() from kernel preemption
5532 * off of irq context.
5533 * Note, that this is called and return with irqs disabled. This will
5534 * protect us against recursive calling from irq.
5536 asmlinkage void __sched preempt_schedule_irq(void)
5538 struct thread_info *ti = current_thread_info();
5540 /* Catch callers which need to be fixed */
5541 BUG_ON(ti->preempt_count || !irqs_disabled());
5544 add_preempt_count(PREEMPT_ACTIVE);
5547 local_irq_disable();
5548 sub_preempt_count(PREEMPT_ACTIVE);
5551 * Check again in case we missed a preemption opportunity
5552 * between schedule and now.
5555 } while (need_resched());
5558 #endif /* CONFIG_PREEMPT */
5560 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5563 return try_to_wake_up(curr->private, mode, sync);
5565 EXPORT_SYMBOL(default_wake_function);
5568 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5569 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5570 * number) then we wake all the non-exclusive tasks and one exclusive task.
5572 * There are circumstances in which we can try to wake a task which has already
5573 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5574 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5576 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5577 int nr_exclusive, int sync, void *key)
5579 wait_queue_t *curr, *next;
5581 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5582 unsigned flags = curr->flags;
5584 if (curr->func(curr, mode, sync, key) &&
5585 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5591 * __wake_up - wake up threads blocked on a waitqueue.
5593 * @mode: which threads
5594 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5595 * @key: is directly passed to the wakeup function
5597 * It may be assumed that this function implies a write memory barrier before
5598 * changing the task state if and only if any tasks are woken up.
5600 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5601 int nr_exclusive, void *key)
5603 unsigned long flags;
5605 spin_lock_irqsave(&q->lock, flags);
5606 __wake_up_common(q, mode, nr_exclusive, 0, key);
5607 spin_unlock_irqrestore(&q->lock, flags);
5609 EXPORT_SYMBOL(__wake_up);
5612 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5614 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5616 __wake_up_common(q, mode, 1, 0, NULL);
5619 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5621 __wake_up_common(q, mode, 1, 0, key);
5625 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5627 * @mode: which threads
5628 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5629 * @key: opaque value to be passed to wakeup targets
5631 * The sync wakeup differs that the waker knows that it will schedule
5632 * away soon, so while the target thread will be woken up, it will not
5633 * be migrated to another CPU - ie. the two threads are 'synchronized'
5634 * with each other. This can prevent needless bouncing between CPUs.
5636 * On UP it can prevent extra preemption.
5638 * It may be assumed that this function implies a write memory barrier before
5639 * changing the task state if and only if any tasks are woken up.
5641 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5642 int nr_exclusive, void *key)
5644 unsigned long flags;
5650 if (unlikely(!nr_exclusive))
5653 spin_lock_irqsave(&q->lock, flags);
5654 __wake_up_common(q, mode, nr_exclusive, sync, key);
5655 spin_unlock_irqrestore(&q->lock, flags);
5657 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5660 * __wake_up_sync - see __wake_up_sync_key()
5662 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5664 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5666 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5669 * complete: - signals a single thread waiting on this completion
5670 * @x: holds the state of this particular completion
5672 * This will wake up a single thread waiting on this completion. Threads will be
5673 * awakened in the same order in which they were queued.
5675 * See also complete_all(), wait_for_completion() and related routines.
5677 * It may be assumed that this function implies a write memory barrier before
5678 * changing the task state if and only if any tasks are woken up.
5680 void complete(struct completion *x)
5682 unsigned long flags;
5684 spin_lock_irqsave(&x->wait.lock, flags);
5686 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5687 spin_unlock_irqrestore(&x->wait.lock, flags);
5689 EXPORT_SYMBOL(complete);
5692 * complete_all: - signals all threads waiting on this completion
5693 * @x: holds the state of this particular completion
5695 * This will wake up all threads waiting on this particular completion event.
5697 * It may be assumed that this function implies a write memory barrier before
5698 * changing the task state if and only if any tasks are woken up.
5700 void complete_all(struct completion *x)
5702 unsigned long flags;
5704 spin_lock_irqsave(&x->wait.lock, flags);
5705 x->done += UINT_MAX/2;
5706 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5707 spin_unlock_irqrestore(&x->wait.lock, flags);
5709 EXPORT_SYMBOL(complete_all);
5711 static inline long __sched
5712 do_wait_for_common(struct completion *x, long timeout, int state)
5715 DECLARE_WAITQUEUE(wait, current);
5717 wait.flags |= WQ_FLAG_EXCLUSIVE;
5718 __add_wait_queue_tail(&x->wait, &wait);
5720 if (signal_pending_state(state, current)) {
5721 timeout = -ERESTARTSYS;
5724 __set_current_state(state);
5725 spin_unlock_irq(&x->wait.lock);
5726 timeout = schedule_timeout(timeout);
5727 spin_lock_irq(&x->wait.lock);
5728 } while (!x->done && timeout);
5729 __remove_wait_queue(&x->wait, &wait);
5734 return timeout ?: 1;
5738 wait_for_common(struct completion *x, long timeout, int state)
5742 spin_lock_irq(&x->wait.lock);
5743 timeout = do_wait_for_common(x, timeout, state);
5744 spin_unlock_irq(&x->wait.lock);
5749 * wait_for_completion: - waits for completion of a task
5750 * @x: holds the state of this particular completion
5752 * This waits to be signaled for completion of a specific task. It is NOT
5753 * interruptible and there is no timeout.
5755 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5756 * and interrupt capability. Also see complete().
5758 void __sched wait_for_completion(struct completion *x)
5760 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5762 EXPORT_SYMBOL(wait_for_completion);
5765 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5766 * @x: holds the state of this particular completion
5767 * @timeout: timeout value in jiffies
5769 * This waits for either a completion of a specific task to be signaled or for a
5770 * specified timeout to expire. The timeout is in jiffies. It is not
5773 unsigned long __sched
5774 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5776 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5778 EXPORT_SYMBOL(wait_for_completion_timeout);
5781 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5782 * @x: holds the state of this particular completion
5784 * This waits for completion of a specific task to be signaled. It is
5787 int __sched wait_for_completion_interruptible(struct completion *x)
5789 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5790 if (t == -ERESTARTSYS)
5794 EXPORT_SYMBOL(wait_for_completion_interruptible);
5797 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5798 * @x: holds the state of this particular completion
5799 * @timeout: timeout value in jiffies
5801 * This waits for either a completion of a specific task to be signaled or for a
5802 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5804 unsigned long __sched
5805 wait_for_completion_interruptible_timeout(struct completion *x,
5806 unsigned long timeout)
5808 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5810 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5813 * wait_for_completion_killable: - waits for completion of a task (killable)
5814 * @x: holds the state of this particular completion
5816 * This waits to be signaled for completion of a specific task. It can be
5817 * interrupted by a kill signal.
5819 int __sched wait_for_completion_killable(struct completion *x)
5821 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5822 if (t == -ERESTARTSYS)
5826 EXPORT_SYMBOL(wait_for_completion_killable);
5829 * try_wait_for_completion - try to decrement a completion without blocking
5830 * @x: completion structure
5832 * Returns: 0 if a decrement cannot be done without blocking
5833 * 1 if a decrement succeeded.
5835 * If a completion is being used as a counting completion,
5836 * attempt to decrement the counter without blocking. This
5837 * enables us to avoid waiting if the resource the completion
5838 * is protecting is not available.
5840 bool try_wait_for_completion(struct completion *x)
5844 spin_lock_irq(&x->wait.lock);
5849 spin_unlock_irq(&x->wait.lock);
5852 EXPORT_SYMBOL(try_wait_for_completion);
5855 * completion_done - Test to see if a completion has any waiters
5856 * @x: completion structure
5858 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5859 * 1 if there are no waiters.
5862 bool completion_done(struct completion *x)
5866 spin_lock_irq(&x->wait.lock);
5869 spin_unlock_irq(&x->wait.lock);
5872 EXPORT_SYMBOL(completion_done);
5875 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5877 unsigned long flags;
5880 init_waitqueue_entry(&wait, current);
5882 __set_current_state(state);
5884 spin_lock_irqsave(&q->lock, flags);
5885 __add_wait_queue(q, &wait);
5886 spin_unlock(&q->lock);
5887 timeout = schedule_timeout(timeout);
5888 spin_lock_irq(&q->lock);
5889 __remove_wait_queue(q, &wait);
5890 spin_unlock_irqrestore(&q->lock, flags);
5895 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5897 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5899 EXPORT_SYMBOL(interruptible_sleep_on);
5902 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5904 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5906 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5908 void __sched sleep_on(wait_queue_head_t *q)
5910 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5912 EXPORT_SYMBOL(sleep_on);
5914 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5916 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5918 EXPORT_SYMBOL(sleep_on_timeout);
5920 #ifdef CONFIG_RT_MUTEXES
5923 * rt_mutex_setprio - set the current priority of a task
5925 * @prio: prio value (kernel-internal form)
5927 * This function changes the 'effective' priority of a task. It does
5928 * not touch ->normal_prio like __setscheduler().
5930 * Used by the rt_mutex code to implement priority inheritance logic.
5932 void rt_mutex_setprio(struct task_struct *p, int prio)
5934 unsigned long flags;
5935 int oldprio, on_rq, running;
5937 const struct sched_class *prev_class = p->sched_class;
5939 BUG_ON(prio < 0 || prio > MAX_PRIO);
5941 rq = task_rq_lock(p, &flags);
5942 update_rq_clock(rq);
5945 on_rq = p->se.on_rq;
5946 running = task_current(rq, p);
5948 dequeue_task(rq, p, 0);
5950 p->sched_class->put_prev_task(rq, p);
5953 p->sched_class = &rt_sched_class;
5955 p->sched_class = &fair_sched_class;
5960 p->sched_class->set_curr_task(rq);
5962 enqueue_task(rq, p, 0);
5964 check_class_changed(rq, p, prev_class, oldprio, running);
5966 task_rq_unlock(rq, &flags);
5971 void set_user_nice(struct task_struct *p, long nice)
5973 int old_prio, delta, on_rq;
5974 unsigned long flags;
5977 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5980 * We have to be careful, if called from sys_setpriority(),
5981 * the task might be in the middle of scheduling on another CPU.
5983 rq = task_rq_lock(p, &flags);
5984 update_rq_clock(rq);
5986 * The RT priorities are set via sched_setscheduler(), but we still
5987 * allow the 'normal' nice value to be set - but as expected
5988 * it wont have any effect on scheduling until the task is
5989 * SCHED_FIFO/SCHED_RR:
5991 if (task_has_rt_policy(p)) {
5992 p->static_prio = NICE_TO_PRIO(nice);
5995 on_rq = p->se.on_rq;
5997 dequeue_task(rq, p, 0);
5999 p->static_prio = NICE_TO_PRIO(nice);
6002 p->prio = effective_prio(p);
6003 delta = p->prio - old_prio;
6006 enqueue_task(rq, p, 0);
6008 * If the task increased its priority or is running and
6009 * lowered its priority, then reschedule its CPU:
6011 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6012 resched_task(rq->curr);
6015 task_rq_unlock(rq, &flags);
6017 EXPORT_SYMBOL(set_user_nice);
6020 * can_nice - check if a task can reduce its nice value
6024 int can_nice(const struct task_struct *p, const int nice)
6026 /* convert nice value [19,-20] to rlimit style value [1,40] */
6027 int nice_rlim = 20 - nice;
6029 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6030 capable(CAP_SYS_NICE));
6033 #ifdef __ARCH_WANT_SYS_NICE
6036 * sys_nice - change the priority of the current process.
6037 * @increment: priority increment
6039 * sys_setpriority is a more generic, but much slower function that
6040 * does similar things.
6042 SYSCALL_DEFINE1(nice, int, increment)
6047 * Setpriority might change our priority at the same moment.
6048 * We don't have to worry. Conceptually one call occurs first
6049 * and we have a single winner.
6051 if (increment < -40)
6056 nice = TASK_NICE(current) + increment;
6062 if (increment < 0 && !can_nice(current, nice))
6065 retval = security_task_setnice(current, nice);
6069 set_user_nice(current, nice);
6076 * task_prio - return the priority value of a given task.
6077 * @p: the task in question.
6079 * This is the priority value as seen by users in /proc.
6080 * RT tasks are offset by -200. Normal tasks are centered
6081 * around 0, value goes from -16 to +15.
6083 int task_prio(const struct task_struct *p)
6085 return p->prio - MAX_RT_PRIO;
6089 * task_nice - return the nice value of a given task.
6090 * @p: the task in question.
6092 int task_nice(const struct task_struct *p)
6094 return TASK_NICE(p);
6096 EXPORT_SYMBOL(task_nice);
6099 * idle_cpu - is a given cpu idle currently?
6100 * @cpu: the processor in question.
6102 int idle_cpu(int cpu)
6104 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6108 * idle_task - return the idle task for a given cpu.
6109 * @cpu: the processor in question.
6111 struct task_struct *idle_task(int cpu)
6113 return cpu_rq(cpu)->idle;
6117 * find_process_by_pid - find a process with a matching PID value.
6118 * @pid: the pid in question.
6120 static struct task_struct *find_process_by_pid(pid_t pid)
6122 return pid ? find_task_by_vpid(pid) : current;
6125 /* Actually do priority change: must hold rq lock. */
6127 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6129 BUG_ON(p->se.on_rq);
6132 switch (p->policy) {
6136 p->sched_class = &fair_sched_class;
6140 p->sched_class = &rt_sched_class;
6144 p->rt_priority = prio;
6145 p->normal_prio = normal_prio(p);
6146 /* we are holding p->pi_lock already */
6147 p->prio = rt_mutex_getprio(p);
6152 * check the target process has a UID that matches the current process's
6154 static bool check_same_owner(struct task_struct *p)
6156 const struct cred *cred = current_cred(), *pcred;
6160 pcred = __task_cred(p);
6161 match = (cred->euid == pcred->euid ||
6162 cred->euid == pcred->uid);
6167 static int __sched_setscheduler(struct task_struct *p, int policy,
6168 struct sched_param *param, bool user)
6170 int retval, oldprio, oldpolicy = -1, on_rq, running;
6171 unsigned long flags;
6172 const struct sched_class *prev_class = p->sched_class;
6176 /* may grab non-irq protected spin_locks */
6177 BUG_ON(in_interrupt());
6179 /* double check policy once rq lock held */
6181 reset_on_fork = p->sched_reset_on_fork;
6182 policy = oldpolicy = p->policy;
6184 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6185 policy &= ~SCHED_RESET_ON_FORK;
6187 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6188 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6189 policy != SCHED_IDLE)
6194 * Valid priorities for SCHED_FIFO and SCHED_RR are
6195 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6196 * SCHED_BATCH and SCHED_IDLE is 0.
6198 if (param->sched_priority < 0 ||
6199 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6200 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6202 if (rt_policy(policy) != (param->sched_priority != 0))
6206 * Allow unprivileged RT tasks to decrease priority:
6208 if (user && !capable(CAP_SYS_NICE)) {
6209 if (rt_policy(policy)) {
6210 unsigned long rlim_rtprio;
6212 if (!lock_task_sighand(p, &flags))
6214 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6215 unlock_task_sighand(p, &flags);
6217 /* can't set/change the rt policy */
6218 if (policy != p->policy && !rlim_rtprio)
6221 /* can't increase priority */
6222 if (param->sched_priority > p->rt_priority &&
6223 param->sched_priority > rlim_rtprio)
6227 * Like positive nice levels, dont allow tasks to
6228 * move out of SCHED_IDLE either:
6230 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6233 /* can't change other user's priorities */
6234 if (!check_same_owner(p))
6237 /* Normal users shall not reset the sched_reset_on_fork flag */
6238 if (p->sched_reset_on_fork && !reset_on_fork)
6243 #ifdef CONFIG_RT_GROUP_SCHED
6245 * Do not allow realtime tasks into groups that have no runtime
6248 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6249 task_group(p)->rt_bandwidth.rt_runtime == 0)
6253 retval = security_task_setscheduler(p, policy, param);
6259 * make sure no PI-waiters arrive (or leave) while we are
6260 * changing the priority of the task:
6262 spin_lock_irqsave(&p->pi_lock, flags);
6264 * To be able to change p->policy safely, the apropriate
6265 * runqueue lock must be held.
6267 rq = __task_rq_lock(p);
6268 /* recheck policy now with rq lock held */
6269 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6270 policy = oldpolicy = -1;
6271 __task_rq_unlock(rq);
6272 spin_unlock_irqrestore(&p->pi_lock, flags);
6275 update_rq_clock(rq);
6276 on_rq = p->se.on_rq;
6277 running = task_current(rq, p);
6279 deactivate_task(rq, p, 0);
6281 p->sched_class->put_prev_task(rq, p);
6283 p->sched_reset_on_fork = reset_on_fork;
6286 __setscheduler(rq, p, policy, param->sched_priority);
6289 p->sched_class->set_curr_task(rq);
6291 activate_task(rq, p, 0);
6293 check_class_changed(rq, p, prev_class, oldprio, running);
6295 __task_rq_unlock(rq);
6296 spin_unlock_irqrestore(&p->pi_lock, flags);
6298 rt_mutex_adjust_pi(p);
6304 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6305 * @p: the task in question.
6306 * @policy: new policy.
6307 * @param: structure containing the new RT priority.
6309 * NOTE that the task may be already dead.
6311 int sched_setscheduler(struct task_struct *p, int policy,
6312 struct sched_param *param)
6314 return __sched_setscheduler(p, policy, param, true);
6316 EXPORT_SYMBOL_GPL(sched_setscheduler);
6319 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6320 * @p: the task in question.
6321 * @policy: new policy.
6322 * @param: structure containing the new RT priority.
6324 * Just like sched_setscheduler, only don't bother checking if the
6325 * current context has permission. For example, this is needed in
6326 * stop_machine(): we create temporary high priority worker threads,
6327 * but our caller might not have that capability.
6329 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6330 struct sched_param *param)
6332 return __sched_setscheduler(p, policy, param, false);
6336 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6338 struct sched_param lparam;
6339 struct task_struct *p;
6342 if (!param || pid < 0)
6344 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6349 p = find_process_by_pid(pid);
6351 retval = sched_setscheduler(p, policy, &lparam);
6358 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6359 * @pid: the pid in question.
6360 * @policy: new policy.
6361 * @param: structure containing the new RT priority.
6363 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6364 struct sched_param __user *, param)
6366 /* negative values for policy are not valid */
6370 return do_sched_setscheduler(pid, policy, param);
6374 * sys_sched_setparam - set/change the RT priority of a thread
6375 * @pid: the pid in question.
6376 * @param: structure containing the new RT priority.
6378 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6380 return do_sched_setscheduler(pid, -1, param);
6384 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6385 * @pid: the pid in question.
6387 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6389 struct task_struct *p;
6396 read_lock(&tasklist_lock);
6397 p = find_process_by_pid(pid);
6399 retval = security_task_getscheduler(p);
6402 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6404 read_unlock(&tasklist_lock);
6409 * sys_sched_getparam - get the RT priority of a thread
6410 * @pid: the pid in question.
6411 * @param: structure containing the RT priority.
6413 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6415 struct sched_param lp;
6416 struct task_struct *p;
6419 if (!param || pid < 0)
6422 read_lock(&tasklist_lock);
6423 p = find_process_by_pid(pid);
6428 retval = security_task_getscheduler(p);
6432 lp.sched_priority = p->rt_priority;
6433 read_unlock(&tasklist_lock);
6436 * This one might sleep, we cannot do it with a spinlock held ...
6438 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6443 read_unlock(&tasklist_lock);
6447 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6449 cpumask_var_t cpus_allowed, new_mask;
6450 struct task_struct *p;
6454 read_lock(&tasklist_lock);
6456 p = find_process_by_pid(pid);
6458 read_unlock(&tasklist_lock);
6464 * It is not safe to call set_cpus_allowed with the
6465 * tasklist_lock held. We will bump the task_struct's
6466 * usage count and then drop tasklist_lock.
6469 read_unlock(&tasklist_lock);
6471 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6475 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6477 goto out_free_cpus_allowed;
6480 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6483 retval = security_task_setscheduler(p, 0, NULL);
6487 cpuset_cpus_allowed(p, cpus_allowed);
6488 cpumask_and(new_mask, in_mask, cpus_allowed);
6490 retval = set_cpus_allowed_ptr(p, new_mask);
6493 cpuset_cpus_allowed(p, cpus_allowed);
6494 if (!cpumask_subset(new_mask, cpus_allowed)) {
6496 * We must have raced with a concurrent cpuset
6497 * update. Just reset the cpus_allowed to the
6498 * cpuset's cpus_allowed
6500 cpumask_copy(new_mask, cpus_allowed);
6505 free_cpumask_var(new_mask);
6506 out_free_cpus_allowed:
6507 free_cpumask_var(cpus_allowed);
6514 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6515 struct cpumask *new_mask)
6517 if (len < cpumask_size())
6518 cpumask_clear(new_mask);
6519 else if (len > cpumask_size())
6520 len = cpumask_size();
6522 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6526 * sys_sched_setaffinity - set the cpu affinity of a process
6527 * @pid: pid of the process
6528 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6529 * @user_mask_ptr: user-space pointer to the new cpu mask
6531 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6532 unsigned long __user *, user_mask_ptr)
6534 cpumask_var_t new_mask;
6537 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6540 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6542 retval = sched_setaffinity(pid, new_mask);
6543 free_cpumask_var(new_mask);
6547 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6549 struct task_struct *p;
6553 read_lock(&tasklist_lock);
6556 p = find_process_by_pid(pid);
6560 retval = security_task_getscheduler(p);
6564 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6567 read_unlock(&tasklist_lock);
6574 * sys_sched_getaffinity - get the cpu affinity of a process
6575 * @pid: pid of the process
6576 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6577 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6579 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6580 unsigned long __user *, user_mask_ptr)
6585 if (len < cpumask_size())
6588 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6591 ret = sched_getaffinity(pid, mask);
6593 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6596 ret = cpumask_size();
6598 free_cpumask_var(mask);
6604 * sys_sched_yield - yield the current processor to other threads.
6606 * This function yields the current CPU to other tasks. If there are no
6607 * other threads running on this CPU then this function will return.
6609 SYSCALL_DEFINE0(sched_yield)
6611 struct rq *rq = this_rq_lock();
6613 schedstat_inc(rq, yld_count);
6614 current->sched_class->yield_task(rq);
6617 * Since we are going to call schedule() anyway, there's
6618 * no need to preempt or enable interrupts:
6620 __release(rq->lock);
6621 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6622 _raw_spin_unlock(&rq->lock);
6623 preempt_enable_no_resched();
6630 static inline int should_resched(void)
6632 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6635 static void __cond_resched(void)
6637 add_preempt_count(PREEMPT_ACTIVE);
6639 sub_preempt_count(PREEMPT_ACTIVE);
6642 int __sched _cond_resched(void)
6644 if (should_resched()) {
6650 EXPORT_SYMBOL(_cond_resched);
6653 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6654 * call schedule, and on return reacquire the lock.
6656 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6657 * operations here to prevent schedule() from being called twice (once via
6658 * spin_unlock(), once by hand).
6660 int __cond_resched_lock(spinlock_t *lock)
6662 int resched = should_resched();
6665 if (spin_needbreak(lock) || resched) {
6676 EXPORT_SYMBOL(__cond_resched_lock);
6678 int __sched __cond_resched_softirq(void)
6680 BUG_ON(!in_softirq());
6682 if (should_resched()) {
6690 EXPORT_SYMBOL(__cond_resched_softirq);
6693 * yield - yield the current processor to other threads.
6695 * This is a shortcut for kernel-space yielding - it marks the
6696 * thread runnable and calls sys_sched_yield().
6698 void __sched yield(void)
6700 set_current_state(TASK_RUNNING);
6703 EXPORT_SYMBOL(yield);
6706 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6707 * that process accounting knows that this is a task in IO wait state.
6709 * But don't do that if it is a deliberate, throttling IO wait (this task
6710 * has set its backing_dev_info: the queue against which it should throttle)
6712 void __sched io_schedule(void)
6714 struct rq *rq = raw_rq();
6716 delayacct_blkio_start();
6717 atomic_inc(&rq->nr_iowait);
6719 atomic_dec(&rq->nr_iowait);
6720 delayacct_blkio_end();
6722 EXPORT_SYMBOL(io_schedule);
6724 long __sched io_schedule_timeout(long timeout)
6726 struct rq *rq = raw_rq();
6729 delayacct_blkio_start();
6730 atomic_inc(&rq->nr_iowait);
6731 ret = schedule_timeout(timeout);
6732 atomic_dec(&rq->nr_iowait);
6733 delayacct_blkio_end();
6738 * sys_sched_get_priority_max - return maximum RT priority.
6739 * @policy: scheduling class.
6741 * this syscall returns the maximum rt_priority that can be used
6742 * by a given scheduling class.
6744 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6751 ret = MAX_USER_RT_PRIO-1;
6763 * sys_sched_get_priority_min - return minimum RT priority.
6764 * @policy: scheduling class.
6766 * this syscall returns the minimum rt_priority that can be used
6767 * by a given scheduling class.
6769 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6787 * sys_sched_rr_get_interval - return the default timeslice of a process.
6788 * @pid: pid of the process.
6789 * @interval: userspace pointer to the timeslice value.
6791 * this syscall writes the default timeslice value of a given process
6792 * into the user-space timespec buffer. A value of '0' means infinity.
6794 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6795 struct timespec __user *, interval)
6797 struct task_struct *p;
6798 unsigned int time_slice;
6806 read_lock(&tasklist_lock);
6807 p = find_process_by_pid(pid);
6811 retval = security_task_getscheduler(p);
6816 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6817 * tasks that are on an otherwise idle runqueue:
6820 if (p->policy == SCHED_RR) {
6821 time_slice = DEF_TIMESLICE;
6822 } else if (p->policy != SCHED_FIFO) {
6823 struct sched_entity *se = &p->se;
6824 unsigned long flags;
6827 rq = task_rq_lock(p, &flags);
6828 if (rq->cfs.load.weight)
6829 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6830 task_rq_unlock(rq, &flags);
6832 read_unlock(&tasklist_lock);
6833 jiffies_to_timespec(time_slice, &t);
6834 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6838 read_unlock(&tasklist_lock);
6842 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6844 void sched_show_task(struct task_struct *p)
6846 unsigned long free = 0;
6849 state = p->state ? __ffs(p->state) + 1 : 0;
6850 printk(KERN_INFO "%-13.13s %c", p->comm,
6851 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6852 #if BITS_PER_LONG == 32
6853 if (state == TASK_RUNNING)
6854 printk(KERN_CONT " running ");
6856 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6858 if (state == TASK_RUNNING)
6859 printk(KERN_CONT " running task ");
6861 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6863 #ifdef CONFIG_DEBUG_STACK_USAGE
6864 free = stack_not_used(p);
6866 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6867 task_pid_nr(p), task_pid_nr(p->real_parent),
6868 (unsigned long)task_thread_info(p)->flags);
6870 show_stack(p, NULL);
6873 void show_state_filter(unsigned long state_filter)
6875 struct task_struct *g, *p;
6877 #if BITS_PER_LONG == 32
6879 " task PC stack pid father\n");
6882 " task PC stack pid father\n");
6884 read_lock(&tasklist_lock);
6885 do_each_thread(g, p) {
6887 * reset the NMI-timeout, listing all files on a slow
6888 * console might take alot of time:
6890 touch_nmi_watchdog();
6891 if (!state_filter || (p->state & state_filter))
6893 } while_each_thread(g, p);
6895 touch_all_softlockup_watchdogs();
6897 #ifdef CONFIG_SCHED_DEBUG
6898 sysrq_sched_debug_show();
6900 read_unlock(&tasklist_lock);
6902 * Only show locks if all tasks are dumped:
6904 if (state_filter == -1)
6905 debug_show_all_locks();
6908 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6910 idle->sched_class = &idle_sched_class;
6914 * init_idle - set up an idle thread for a given CPU
6915 * @idle: task in question
6916 * @cpu: cpu the idle task belongs to
6918 * NOTE: this function does not set the idle thread's NEED_RESCHED
6919 * flag, to make booting more robust.
6921 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6923 struct rq *rq = cpu_rq(cpu);
6924 unsigned long flags;
6926 spin_lock_irqsave(&rq->lock, flags);
6929 idle->se.exec_start = sched_clock();
6931 idle->prio = idle->normal_prio = MAX_PRIO;
6932 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6933 __set_task_cpu(idle, cpu);
6935 rq->curr = rq->idle = idle;
6936 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6939 spin_unlock_irqrestore(&rq->lock, flags);
6941 /* Set the preempt count _outside_ the spinlocks! */
6942 #if defined(CONFIG_PREEMPT)
6943 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6945 task_thread_info(idle)->preempt_count = 0;
6948 * The idle tasks have their own, simple scheduling class:
6950 idle->sched_class = &idle_sched_class;
6951 ftrace_graph_init_task(idle);
6955 * In a system that switches off the HZ timer nohz_cpu_mask
6956 * indicates which cpus entered this state. This is used
6957 * in the rcu update to wait only for active cpus. For system
6958 * which do not switch off the HZ timer nohz_cpu_mask should
6959 * always be CPU_BITS_NONE.
6961 cpumask_var_t nohz_cpu_mask;
6964 * Increase the granularity value when there are more CPUs,
6965 * because with more CPUs the 'effective latency' as visible
6966 * to users decreases. But the relationship is not linear,
6967 * so pick a second-best guess by going with the log2 of the
6970 * This idea comes from the SD scheduler of Con Kolivas:
6972 static inline void sched_init_granularity(void)
6974 unsigned int factor = 1 + ilog2(num_online_cpus());
6975 const unsigned long limit = 200000000;
6977 sysctl_sched_min_granularity *= factor;
6978 if (sysctl_sched_min_granularity > limit)
6979 sysctl_sched_min_granularity = limit;
6981 sysctl_sched_latency *= factor;
6982 if (sysctl_sched_latency > limit)
6983 sysctl_sched_latency = limit;
6985 sysctl_sched_wakeup_granularity *= factor;
6987 sysctl_sched_shares_ratelimit *= factor;
6992 * This is how migration works:
6994 * 1) we queue a struct migration_req structure in the source CPU's
6995 * runqueue and wake up that CPU's migration thread.
6996 * 2) we down() the locked semaphore => thread blocks.
6997 * 3) migration thread wakes up (implicitly it forces the migrated
6998 * thread off the CPU)
6999 * 4) it gets the migration request and checks whether the migrated
7000 * task is still in the wrong runqueue.
7001 * 5) if it's in the wrong runqueue then the migration thread removes
7002 * it and puts it into the right queue.
7003 * 6) migration thread up()s the semaphore.
7004 * 7) we wake up and the migration is done.
7008 * Change a given task's CPU affinity. Migrate the thread to a
7009 * proper CPU and schedule it away if the CPU it's executing on
7010 * is removed from the allowed bitmask.
7012 * NOTE: the caller must have a valid reference to the task, the
7013 * task must not exit() & deallocate itself prematurely. The
7014 * call is not atomic; no spinlocks may be held.
7016 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7018 struct migration_req req;
7019 unsigned long flags;
7023 rq = task_rq_lock(p, &flags);
7024 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
7029 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7030 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7035 if (p->sched_class->set_cpus_allowed)
7036 p->sched_class->set_cpus_allowed(p, new_mask);
7038 cpumask_copy(&p->cpus_allowed, new_mask);
7039 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7042 /* Can the task run on the task's current CPU? If so, we're done */
7043 if (cpumask_test_cpu(task_cpu(p), new_mask))
7046 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
7047 /* Need help from migration thread: drop lock and wait. */
7048 task_rq_unlock(rq, &flags);
7049 wake_up_process(rq->migration_thread);
7050 wait_for_completion(&req.done);
7051 tlb_migrate_finish(p->mm);
7055 task_rq_unlock(rq, &flags);
7059 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7062 * Move (not current) task off this cpu, onto dest cpu. We're doing
7063 * this because either it can't run here any more (set_cpus_allowed()
7064 * away from this CPU, or CPU going down), or because we're
7065 * attempting to rebalance this task on exec (sched_exec).
7067 * So we race with normal scheduler movements, but that's OK, as long
7068 * as the task is no longer on this CPU.
7070 * Returns non-zero if task was successfully migrated.
7072 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7074 struct rq *rq_dest, *rq_src;
7077 if (unlikely(!cpu_active(dest_cpu)))
7080 rq_src = cpu_rq(src_cpu);
7081 rq_dest = cpu_rq(dest_cpu);
7083 double_rq_lock(rq_src, rq_dest);
7084 /* Already moved. */
7085 if (task_cpu(p) != src_cpu)
7087 /* Affinity changed (again). */
7088 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7091 on_rq = p->se.on_rq;
7093 deactivate_task(rq_src, p, 0);
7095 set_task_cpu(p, dest_cpu);
7097 activate_task(rq_dest, p, 0);
7098 check_preempt_curr(rq_dest, p, 0);
7103 double_rq_unlock(rq_src, rq_dest);
7108 * migration_thread - this is a highprio system thread that performs
7109 * thread migration by bumping thread off CPU then 'pushing' onto
7112 static int migration_thread(void *data)
7114 int cpu = (long)data;
7118 BUG_ON(rq->migration_thread != current);
7120 set_current_state(TASK_INTERRUPTIBLE);
7121 while (!kthread_should_stop()) {
7122 struct migration_req *req;
7123 struct list_head *head;
7125 spin_lock_irq(&rq->lock);
7127 if (cpu_is_offline(cpu)) {
7128 spin_unlock_irq(&rq->lock);
7132 if (rq->active_balance) {
7133 active_load_balance(rq, cpu);
7134 rq->active_balance = 0;
7137 head = &rq->migration_queue;
7139 if (list_empty(head)) {
7140 spin_unlock_irq(&rq->lock);
7142 set_current_state(TASK_INTERRUPTIBLE);
7145 req = list_entry(head->next, struct migration_req, list);
7146 list_del_init(head->next);
7148 spin_unlock(&rq->lock);
7149 __migrate_task(req->task, cpu, req->dest_cpu);
7152 complete(&req->done);
7154 __set_current_state(TASK_RUNNING);
7159 #ifdef CONFIG_HOTPLUG_CPU
7161 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7165 local_irq_disable();
7166 ret = __migrate_task(p, src_cpu, dest_cpu);
7172 * Figure out where task on dead CPU should go, use force if necessary.
7174 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7177 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7180 /* Look for allowed, online CPU in same node. */
7181 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7182 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7185 /* Any allowed, online CPU? */
7186 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7187 if (dest_cpu < nr_cpu_ids)
7190 /* No more Mr. Nice Guy. */
7191 if (dest_cpu >= nr_cpu_ids) {
7192 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7193 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7196 * Don't tell them about moving exiting tasks or
7197 * kernel threads (both mm NULL), since they never
7200 if (p->mm && printk_ratelimit()) {
7201 printk(KERN_INFO "process %d (%s) no "
7202 "longer affine to cpu%d\n",
7203 task_pid_nr(p), p->comm, dead_cpu);
7208 /* It can have affinity changed while we were choosing. */
7209 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7214 * While a dead CPU has no uninterruptible tasks queued at this point,
7215 * it might still have a nonzero ->nr_uninterruptible counter, because
7216 * for performance reasons the counter is not stricly tracking tasks to
7217 * their home CPUs. So we just add the counter to another CPU's counter,
7218 * to keep the global sum constant after CPU-down:
7220 static void migrate_nr_uninterruptible(struct rq *rq_src)
7222 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7223 unsigned long flags;
7225 local_irq_save(flags);
7226 double_rq_lock(rq_src, rq_dest);
7227 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7228 rq_src->nr_uninterruptible = 0;
7229 double_rq_unlock(rq_src, rq_dest);
7230 local_irq_restore(flags);
7233 /* Run through task list and migrate tasks from the dead cpu. */
7234 static void migrate_live_tasks(int src_cpu)
7236 struct task_struct *p, *t;
7238 read_lock(&tasklist_lock);
7240 do_each_thread(t, p) {
7244 if (task_cpu(p) == src_cpu)
7245 move_task_off_dead_cpu(src_cpu, p);
7246 } while_each_thread(t, p);
7248 read_unlock(&tasklist_lock);
7252 * Schedules idle task to be the next runnable task on current CPU.
7253 * It does so by boosting its priority to highest possible.
7254 * Used by CPU offline code.
7256 void sched_idle_next(void)
7258 int this_cpu = smp_processor_id();
7259 struct rq *rq = cpu_rq(this_cpu);
7260 struct task_struct *p = rq->idle;
7261 unsigned long flags;
7263 /* cpu has to be offline */
7264 BUG_ON(cpu_online(this_cpu));
7267 * Strictly not necessary since rest of the CPUs are stopped by now
7268 * and interrupts disabled on the current cpu.
7270 spin_lock_irqsave(&rq->lock, flags);
7272 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7274 update_rq_clock(rq);
7275 activate_task(rq, p, 0);
7277 spin_unlock_irqrestore(&rq->lock, flags);
7281 * Ensures that the idle task is using init_mm right before its cpu goes
7284 void idle_task_exit(void)
7286 struct mm_struct *mm = current->active_mm;
7288 BUG_ON(cpu_online(smp_processor_id()));
7291 switch_mm(mm, &init_mm, current);
7295 /* called under rq->lock with disabled interrupts */
7296 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7298 struct rq *rq = cpu_rq(dead_cpu);
7300 /* Must be exiting, otherwise would be on tasklist. */
7301 BUG_ON(!p->exit_state);
7303 /* Cannot have done final schedule yet: would have vanished. */
7304 BUG_ON(p->state == TASK_DEAD);
7309 * Drop lock around migration; if someone else moves it,
7310 * that's OK. No task can be added to this CPU, so iteration is
7313 spin_unlock_irq(&rq->lock);
7314 move_task_off_dead_cpu(dead_cpu, p);
7315 spin_lock_irq(&rq->lock);
7320 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7321 static void migrate_dead_tasks(unsigned int dead_cpu)
7323 struct rq *rq = cpu_rq(dead_cpu);
7324 struct task_struct *next;
7327 if (!rq->nr_running)
7329 update_rq_clock(rq);
7330 next = pick_next_task(rq);
7333 next->sched_class->put_prev_task(rq, next);
7334 migrate_dead(dead_cpu, next);
7340 * remove the tasks which were accounted by rq from calc_load_tasks.
7342 static void calc_global_load_remove(struct rq *rq)
7344 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7345 rq->calc_load_active = 0;
7347 #endif /* CONFIG_HOTPLUG_CPU */
7349 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7351 static struct ctl_table sd_ctl_dir[] = {
7353 .procname = "sched_domain",
7359 static struct ctl_table sd_ctl_root[] = {
7361 .ctl_name = CTL_KERN,
7362 .procname = "kernel",
7364 .child = sd_ctl_dir,
7369 static struct ctl_table *sd_alloc_ctl_entry(int n)
7371 struct ctl_table *entry =
7372 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7377 static void sd_free_ctl_entry(struct ctl_table **tablep)
7379 struct ctl_table *entry;
7382 * In the intermediate directories, both the child directory and
7383 * procname are dynamically allocated and could fail but the mode
7384 * will always be set. In the lowest directory the names are
7385 * static strings and all have proc handlers.
7387 for (entry = *tablep; entry->mode; entry++) {
7389 sd_free_ctl_entry(&entry->child);
7390 if (entry->proc_handler == NULL)
7391 kfree(entry->procname);
7399 set_table_entry(struct ctl_table *entry,
7400 const char *procname, void *data, int maxlen,
7401 mode_t mode, proc_handler *proc_handler)
7403 entry->procname = procname;
7405 entry->maxlen = maxlen;
7407 entry->proc_handler = proc_handler;
7410 static struct ctl_table *
7411 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7413 struct ctl_table *table = sd_alloc_ctl_entry(13);
7418 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7419 sizeof(long), 0644, proc_doulongvec_minmax);
7420 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7421 sizeof(long), 0644, proc_doulongvec_minmax);
7422 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7423 sizeof(int), 0644, proc_dointvec_minmax);
7424 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7425 sizeof(int), 0644, proc_dointvec_minmax);
7426 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7427 sizeof(int), 0644, proc_dointvec_minmax);
7428 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7429 sizeof(int), 0644, proc_dointvec_minmax);
7430 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7431 sizeof(int), 0644, proc_dointvec_minmax);
7432 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7433 sizeof(int), 0644, proc_dointvec_minmax);
7434 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7435 sizeof(int), 0644, proc_dointvec_minmax);
7436 set_table_entry(&table[9], "cache_nice_tries",
7437 &sd->cache_nice_tries,
7438 sizeof(int), 0644, proc_dointvec_minmax);
7439 set_table_entry(&table[10], "flags", &sd->flags,
7440 sizeof(int), 0644, proc_dointvec_minmax);
7441 set_table_entry(&table[11], "name", sd->name,
7442 CORENAME_MAX_SIZE, 0444, proc_dostring);
7443 /* &table[12] is terminator */
7448 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7450 struct ctl_table *entry, *table;
7451 struct sched_domain *sd;
7452 int domain_num = 0, i;
7455 for_each_domain(cpu, sd)
7457 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7462 for_each_domain(cpu, sd) {
7463 snprintf(buf, 32, "domain%d", i);
7464 entry->procname = kstrdup(buf, GFP_KERNEL);
7466 entry->child = sd_alloc_ctl_domain_table(sd);
7473 static struct ctl_table_header *sd_sysctl_header;
7474 static void register_sched_domain_sysctl(void)
7476 int i, cpu_num = num_online_cpus();
7477 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7480 WARN_ON(sd_ctl_dir[0].child);
7481 sd_ctl_dir[0].child = entry;
7486 for_each_online_cpu(i) {
7487 snprintf(buf, 32, "cpu%d", i);
7488 entry->procname = kstrdup(buf, GFP_KERNEL);
7490 entry->child = sd_alloc_ctl_cpu_table(i);
7494 WARN_ON(sd_sysctl_header);
7495 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7498 /* may be called multiple times per register */
7499 static void unregister_sched_domain_sysctl(void)
7501 if (sd_sysctl_header)
7502 unregister_sysctl_table(sd_sysctl_header);
7503 sd_sysctl_header = NULL;
7504 if (sd_ctl_dir[0].child)
7505 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7508 static void register_sched_domain_sysctl(void)
7511 static void unregister_sched_domain_sysctl(void)
7516 static void set_rq_online(struct rq *rq)
7519 const struct sched_class *class;
7521 cpumask_set_cpu(rq->cpu, rq->rd->online);
7524 for_each_class(class) {
7525 if (class->rq_online)
7526 class->rq_online(rq);
7531 static void set_rq_offline(struct rq *rq)
7534 const struct sched_class *class;
7536 for_each_class(class) {
7537 if (class->rq_offline)
7538 class->rq_offline(rq);
7541 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7547 * migration_call - callback that gets triggered when a CPU is added.
7548 * Here we can start up the necessary migration thread for the new CPU.
7550 static int __cpuinit
7551 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7553 struct task_struct *p;
7554 int cpu = (long)hcpu;
7555 unsigned long flags;
7560 case CPU_UP_PREPARE:
7561 case CPU_UP_PREPARE_FROZEN:
7562 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7565 kthread_bind(p, cpu);
7566 /* Must be high prio: stop_machine expects to yield to it. */
7567 rq = task_rq_lock(p, &flags);
7568 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7569 task_rq_unlock(rq, &flags);
7571 cpu_rq(cpu)->migration_thread = p;
7572 rq->calc_load_update = calc_load_update;
7576 case CPU_ONLINE_FROZEN:
7577 /* Strictly unnecessary, as first user will wake it. */
7578 wake_up_process(cpu_rq(cpu)->migration_thread);
7580 /* Update our root-domain */
7582 spin_lock_irqsave(&rq->lock, flags);
7584 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7588 spin_unlock_irqrestore(&rq->lock, flags);
7591 #ifdef CONFIG_HOTPLUG_CPU
7592 case CPU_UP_CANCELED:
7593 case CPU_UP_CANCELED_FROZEN:
7594 if (!cpu_rq(cpu)->migration_thread)
7596 /* Unbind it from offline cpu so it can run. Fall thru. */
7597 kthread_bind(cpu_rq(cpu)->migration_thread,
7598 cpumask_any(cpu_online_mask));
7599 kthread_stop(cpu_rq(cpu)->migration_thread);
7600 put_task_struct(cpu_rq(cpu)->migration_thread);
7601 cpu_rq(cpu)->migration_thread = NULL;
7605 case CPU_DEAD_FROZEN:
7606 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7607 migrate_live_tasks(cpu);
7609 kthread_stop(rq->migration_thread);
7610 put_task_struct(rq->migration_thread);
7611 rq->migration_thread = NULL;
7612 /* Idle task back to normal (off runqueue, low prio) */
7613 spin_lock_irq(&rq->lock);
7614 update_rq_clock(rq);
7615 deactivate_task(rq, rq->idle, 0);
7616 rq->idle->static_prio = MAX_PRIO;
7617 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7618 rq->idle->sched_class = &idle_sched_class;
7619 migrate_dead_tasks(cpu);
7620 spin_unlock_irq(&rq->lock);
7622 migrate_nr_uninterruptible(rq);
7623 BUG_ON(rq->nr_running != 0);
7624 calc_global_load_remove(rq);
7626 * No need to migrate the tasks: it was best-effort if
7627 * they didn't take sched_hotcpu_mutex. Just wake up
7630 spin_lock_irq(&rq->lock);
7631 while (!list_empty(&rq->migration_queue)) {
7632 struct migration_req *req;
7634 req = list_entry(rq->migration_queue.next,
7635 struct migration_req, list);
7636 list_del_init(&req->list);
7637 spin_unlock_irq(&rq->lock);
7638 complete(&req->done);
7639 spin_lock_irq(&rq->lock);
7641 spin_unlock_irq(&rq->lock);
7645 case CPU_DYING_FROZEN:
7646 /* Update our root-domain */
7648 spin_lock_irqsave(&rq->lock, flags);
7650 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7653 spin_unlock_irqrestore(&rq->lock, flags);
7661 * Register at high priority so that task migration (migrate_all_tasks)
7662 * happens before everything else. This has to be lower priority than
7663 * the notifier in the perf_counter subsystem, though.
7665 static struct notifier_block __cpuinitdata migration_notifier = {
7666 .notifier_call = migration_call,
7670 static int __init migration_init(void)
7672 void *cpu = (void *)(long)smp_processor_id();
7675 /* Start one for the boot CPU: */
7676 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7677 BUG_ON(err == NOTIFY_BAD);
7678 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7679 register_cpu_notifier(&migration_notifier);
7683 early_initcall(migration_init);
7688 #ifdef CONFIG_SCHED_DEBUG
7690 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7691 struct cpumask *groupmask)
7693 struct sched_group *group = sd->groups;
7696 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7697 cpumask_clear(groupmask);
7699 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7701 if (!(sd->flags & SD_LOAD_BALANCE)) {
7702 printk("does not load-balance\n");
7704 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7709 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7711 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7712 printk(KERN_ERR "ERROR: domain->span does not contain "
7715 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7716 printk(KERN_ERR "ERROR: domain->groups does not contain"
7720 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7724 printk(KERN_ERR "ERROR: group is NULL\n");
7728 if (!group->__cpu_power) {
7729 printk(KERN_CONT "\n");
7730 printk(KERN_ERR "ERROR: domain->cpu_power not "
7735 if (!cpumask_weight(sched_group_cpus(group))) {
7736 printk(KERN_CONT "\n");
7737 printk(KERN_ERR "ERROR: empty group\n");
7741 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7742 printk(KERN_CONT "\n");
7743 printk(KERN_ERR "ERROR: repeated CPUs\n");
7747 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7749 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7751 printk(KERN_CONT " %s", str);
7752 if (group->__cpu_power != SCHED_LOAD_SCALE) {
7753 printk(KERN_CONT " (__cpu_power = %d)",
7754 group->__cpu_power);
7757 group = group->next;
7758 } while (group != sd->groups);
7759 printk(KERN_CONT "\n");
7761 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7762 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7765 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7766 printk(KERN_ERR "ERROR: parent span is not a superset "
7767 "of domain->span\n");
7771 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7773 cpumask_var_t groupmask;
7777 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7781 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7783 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7784 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7789 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7796 free_cpumask_var(groupmask);
7798 #else /* !CONFIG_SCHED_DEBUG */
7799 # define sched_domain_debug(sd, cpu) do { } while (0)
7800 #endif /* CONFIG_SCHED_DEBUG */
7802 static int sd_degenerate(struct sched_domain *sd)
7804 if (cpumask_weight(sched_domain_span(sd)) == 1)
7807 /* Following flags need at least 2 groups */
7808 if (sd->flags & (SD_LOAD_BALANCE |
7809 SD_BALANCE_NEWIDLE |
7813 SD_SHARE_PKG_RESOURCES)) {
7814 if (sd->groups != sd->groups->next)
7818 /* Following flags don't use groups */
7819 if (sd->flags & (SD_WAKE_IDLE |
7828 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7830 unsigned long cflags = sd->flags, pflags = parent->flags;
7832 if (sd_degenerate(parent))
7835 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7838 /* Does parent contain flags not in child? */
7839 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7840 if (cflags & SD_WAKE_AFFINE)
7841 pflags &= ~SD_WAKE_BALANCE;
7842 /* Flags needing groups don't count if only 1 group in parent */
7843 if (parent->groups == parent->groups->next) {
7844 pflags &= ~(SD_LOAD_BALANCE |
7845 SD_BALANCE_NEWIDLE |
7849 SD_SHARE_PKG_RESOURCES);
7850 if (nr_node_ids == 1)
7851 pflags &= ~SD_SERIALIZE;
7853 if (~cflags & pflags)
7859 static void free_rootdomain(struct root_domain *rd)
7861 cpupri_cleanup(&rd->cpupri);
7863 free_cpumask_var(rd->rto_mask);
7864 free_cpumask_var(rd->online);
7865 free_cpumask_var(rd->span);
7869 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7871 struct root_domain *old_rd = NULL;
7872 unsigned long flags;
7874 spin_lock_irqsave(&rq->lock, flags);
7879 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7882 cpumask_clear_cpu(rq->cpu, old_rd->span);
7885 * If we dont want to free the old_rt yet then
7886 * set old_rd to NULL to skip the freeing later
7889 if (!atomic_dec_and_test(&old_rd->refcount))
7893 atomic_inc(&rd->refcount);
7896 cpumask_set_cpu(rq->cpu, rd->span);
7897 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7900 spin_unlock_irqrestore(&rq->lock, flags);
7903 free_rootdomain(old_rd);
7906 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7908 gfp_t gfp = GFP_KERNEL;
7910 memset(rd, 0, sizeof(*rd));
7915 if (!alloc_cpumask_var(&rd->span, gfp))
7917 if (!alloc_cpumask_var(&rd->online, gfp))
7919 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
7922 if (cpupri_init(&rd->cpupri, bootmem) != 0)
7927 free_cpumask_var(rd->rto_mask);
7929 free_cpumask_var(rd->online);
7931 free_cpumask_var(rd->span);
7936 static void init_defrootdomain(void)
7938 init_rootdomain(&def_root_domain, true);
7940 atomic_set(&def_root_domain.refcount, 1);
7943 static struct root_domain *alloc_rootdomain(void)
7945 struct root_domain *rd;
7947 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7951 if (init_rootdomain(rd, false) != 0) {
7960 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7961 * hold the hotplug lock.
7964 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7966 struct rq *rq = cpu_rq(cpu);
7967 struct sched_domain *tmp;
7969 /* Remove the sched domains which do not contribute to scheduling. */
7970 for (tmp = sd; tmp; ) {
7971 struct sched_domain *parent = tmp->parent;
7975 if (sd_parent_degenerate(tmp, parent)) {
7976 tmp->parent = parent->parent;
7978 parent->parent->child = tmp;
7983 if (sd && sd_degenerate(sd)) {
7989 sched_domain_debug(sd, cpu);
7991 rq_attach_root(rq, rd);
7992 rcu_assign_pointer(rq->sd, sd);
7995 /* cpus with isolated domains */
7996 static cpumask_var_t cpu_isolated_map;
7998 /* Setup the mask of cpus configured for isolated domains */
7999 static int __init isolated_cpu_setup(char *str)
8001 cpulist_parse(str, cpu_isolated_map);
8005 __setup("isolcpus=", isolated_cpu_setup);
8008 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8009 * to a function which identifies what group(along with sched group) a CPU
8010 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8011 * (due to the fact that we keep track of groups covered with a struct cpumask).
8013 * init_sched_build_groups will build a circular linked list of the groups
8014 * covered by the given span, and will set each group's ->cpumask correctly,
8015 * and ->cpu_power to 0.
8018 init_sched_build_groups(const struct cpumask *span,
8019 const struct cpumask *cpu_map,
8020 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8021 struct sched_group **sg,
8022 struct cpumask *tmpmask),
8023 struct cpumask *covered, struct cpumask *tmpmask)
8025 struct sched_group *first = NULL, *last = NULL;
8028 cpumask_clear(covered);
8030 for_each_cpu(i, span) {
8031 struct sched_group *sg;
8032 int group = group_fn(i, cpu_map, &sg, tmpmask);
8035 if (cpumask_test_cpu(i, covered))
8038 cpumask_clear(sched_group_cpus(sg));
8039 sg->__cpu_power = 0;
8041 for_each_cpu(j, span) {
8042 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8045 cpumask_set_cpu(j, covered);
8046 cpumask_set_cpu(j, sched_group_cpus(sg));
8057 #define SD_NODES_PER_DOMAIN 16
8062 * find_next_best_node - find the next node to include in a sched_domain
8063 * @node: node whose sched_domain we're building
8064 * @used_nodes: nodes already in the sched_domain
8066 * Find the next node to include in a given scheduling domain. Simply
8067 * finds the closest node not already in the @used_nodes map.
8069 * Should use nodemask_t.
8071 static int find_next_best_node(int node, nodemask_t *used_nodes)
8073 int i, n, val, min_val, best_node = 0;
8077 for (i = 0; i < nr_node_ids; i++) {
8078 /* Start at @node */
8079 n = (node + i) % nr_node_ids;
8081 if (!nr_cpus_node(n))
8084 /* Skip already used nodes */
8085 if (node_isset(n, *used_nodes))
8088 /* Simple min distance search */
8089 val = node_distance(node, n);
8091 if (val < min_val) {
8097 node_set(best_node, *used_nodes);
8102 * sched_domain_node_span - get a cpumask for a node's sched_domain
8103 * @node: node whose cpumask we're constructing
8104 * @span: resulting cpumask
8106 * Given a node, construct a good cpumask for its sched_domain to span. It
8107 * should be one that prevents unnecessary balancing, but also spreads tasks
8110 static void sched_domain_node_span(int node, struct cpumask *span)
8112 nodemask_t used_nodes;
8115 cpumask_clear(span);
8116 nodes_clear(used_nodes);
8118 cpumask_or(span, span, cpumask_of_node(node));
8119 node_set(node, used_nodes);
8121 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8122 int next_node = find_next_best_node(node, &used_nodes);
8124 cpumask_or(span, span, cpumask_of_node(next_node));
8127 #endif /* CONFIG_NUMA */
8129 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8132 * The cpus mask in sched_group and sched_domain hangs off the end.
8134 * ( See the the comments in include/linux/sched.h:struct sched_group
8135 * and struct sched_domain. )
8137 struct static_sched_group {
8138 struct sched_group sg;
8139 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8142 struct static_sched_domain {
8143 struct sched_domain sd;
8144 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8148 * SMT sched-domains:
8150 #ifdef CONFIG_SCHED_SMT
8151 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8152 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8155 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8156 struct sched_group **sg, struct cpumask *unused)
8159 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8162 #endif /* CONFIG_SCHED_SMT */
8165 * multi-core sched-domains:
8167 #ifdef CONFIG_SCHED_MC
8168 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8169 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8170 #endif /* CONFIG_SCHED_MC */
8172 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8174 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8175 struct sched_group **sg, struct cpumask *mask)
8179 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8180 group = cpumask_first(mask);
8182 *sg = &per_cpu(sched_group_core, group).sg;
8185 #elif defined(CONFIG_SCHED_MC)
8187 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8188 struct sched_group **sg, struct cpumask *unused)
8191 *sg = &per_cpu(sched_group_core, cpu).sg;
8196 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8197 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8200 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8201 struct sched_group **sg, struct cpumask *mask)
8204 #ifdef CONFIG_SCHED_MC
8205 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8206 group = cpumask_first(mask);
8207 #elif defined(CONFIG_SCHED_SMT)
8208 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8209 group = cpumask_first(mask);
8214 *sg = &per_cpu(sched_group_phys, group).sg;
8220 * The init_sched_build_groups can't handle what we want to do with node
8221 * groups, so roll our own. Now each node has its own list of groups which
8222 * gets dynamically allocated.
8224 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8225 static struct sched_group ***sched_group_nodes_bycpu;
8227 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8228 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8230 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8231 struct sched_group **sg,
8232 struct cpumask *nodemask)
8236 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8237 group = cpumask_first(nodemask);
8240 *sg = &per_cpu(sched_group_allnodes, group).sg;
8244 static void init_numa_sched_groups_power(struct sched_group *group_head)
8246 struct sched_group *sg = group_head;
8252 for_each_cpu(j, sched_group_cpus(sg)) {
8253 struct sched_domain *sd;
8255 sd = &per_cpu(phys_domains, j).sd;
8256 if (j != group_first_cpu(sd->groups)) {
8258 * Only add "power" once for each
8264 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
8267 } while (sg != group_head);
8269 #endif /* CONFIG_NUMA */
8272 /* Free memory allocated for various sched_group structures */
8273 static void free_sched_groups(const struct cpumask *cpu_map,
8274 struct cpumask *nodemask)
8278 for_each_cpu(cpu, cpu_map) {
8279 struct sched_group **sched_group_nodes
8280 = sched_group_nodes_bycpu[cpu];
8282 if (!sched_group_nodes)
8285 for (i = 0; i < nr_node_ids; i++) {
8286 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8288 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8289 if (cpumask_empty(nodemask))
8299 if (oldsg != sched_group_nodes[i])
8302 kfree(sched_group_nodes);
8303 sched_group_nodes_bycpu[cpu] = NULL;
8306 #else /* !CONFIG_NUMA */
8307 static void free_sched_groups(const struct cpumask *cpu_map,
8308 struct cpumask *nodemask)
8311 #endif /* CONFIG_NUMA */
8314 * Initialize sched groups cpu_power.
8316 * cpu_power indicates the capacity of sched group, which is used while
8317 * distributing the load between different sched groups in a sched domain.
8318 * Typically cpu_power for all the groups in a sched domain will be same unless
8319 * there are asymmetries in the topology. If there are asymmetries, group
8320 * having more cpu_power will pickup more load compared to the group having
8323 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
8324 * the maximum number of tasks a group can handle in the presence of other idle
8325 * or lightly loaded groups in the same sched domain.
8327 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8329 struct sched_domain *child;
8330 struct sched_group *group;
8332 WARN_ON(!sd || !sd->groups);
8334 if (cpu != group_first_cpu(sd->groups))
8339 sd->groups->__cpu_power = 0;
8342 * For perf policy, if the groups in child domain share resources
8343 * (for example cores sharing some portions of the cache hierarchy
8344 * or SMT), then set this domain groups cpu_power such that each group
8345 * can handle only one task, when there are other idle groups in the
8346 * same sched domain.
8348 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
8350 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
8351 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
8356 * add cpu_power of each child group to this groups cpu_power
8358 group = child->groups;
8360 sg_inc_cpu_power(sd->groups, group->__cpu_power);
8361 group = group->next;
8362 } while (group != child->groups);
8366 * Initializers for schedule domains
8367 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8370 #ifdef CONFIG_SCHED_DEBUG
8371 # define SD_INIT_NAME(sd, type) sd->name = #type
8373 # define SD_INIT_NAME(sd, type) do { } while (0)
8376 #define SD_INIT(sd, type) sd_init_##type(sd)
8378 #define SD_INIT_FUNC(type) \
8379 static noinline void sd_init_##type(struct sched_domain *sd) \
8381 memset(sd, 0, sizeof(*sd)); \
8382 *sd = SD_##type##_INIT; \
8383 sd->level = SD_LV_##type; \
8384 SD_INIT_NAME(sd, type); \
8389 SD_INIT_FUNC(ALLNODES)
8392 #ifdef CONFIG_SCHED_SMT
8393 SD_INIT_FUNC(SIBLING)
8395 #ifdef CONFIG_SCHED_MC
8399 static int default_relax_domain_level = -1;
8401 static int __init setup_relax_domain_level(char *str)
8405 val = simple_strtoul(str, NULL, 0);
8406 if (val < SD_LV_MAX)
8407 default_relax_domain_level = val;
8411 __setup("relax_domain_level=", setup_relax_domain_level);
8413 static void set_domain_attribute(struct sched_domain *sd,
8414 struct sched_domain_attr *attr)
8418 if (!attr || attr->relax_domain_level < 0) {
8419 if (default_relax_domain_level < 0)
8422 request = default_relax_domain_level;
8424 request = attr->relax_domain_level;
8425 if (request < sd->level) {
8426 /* turn off idle balance on this domain */
8427 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
8429 /* turn on idle balance on this domain */
8430 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
8435 * Build sched domains for a given set of cpus and attach the sched domains
8436 * to the individual cpus
8438 static int __build_sched_domains(const struct cpumask *cpu_map,
8439 struct sched_domain_attr *attr)
8441 int i, err = -ENOMEM;
8442 struct root_domain *rd;
8443 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
8446 cpumask_var_t domainspan, covered, notcovered;
8447 struct sched_group **sched_group_nodes = NULL;
8448 int sd_allnodes = 0;
8450 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
8452 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
8453 goto free_domainspan;
8454 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
8458 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
8459 goto free_notcovered;
8460 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
8462 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
8463 goto free_this_sibling_map;
8464 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
8465 goto free_this_core_map;
8466 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
8467 goto free_send_covered;
8471 * Allocate the per-node list of sched groups
8473 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
8475 if (!sched_group_nodes) {
8476 printk(KERN_WARNING "Can not alloc sched group node list\n");
8481 rd = alloc_rootdomain();
8483 printk(KERN_WARNING "Cannot alloc root domain\n");
8484 goto free_sched_groups;
8488 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
8492 * Set up domains for cpus specified by the cpu_map.
8494 for_each_cpu(i, cpu_map) {
8495 struct sched_domain *sd = NULL, *p;
8497 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
8500 if (cpumask_weight(cpu_map) >
8501 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
8502 sd = &per_cpu(allnodes_domains, i).sd;
8503 SD_INIT(sd, ALLNODES);
8504 set_domain_attribute(sd, attr);
8505 cpumask_copy(sched_domain_span(sd), cpu_map);
8506 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
8512 sd = &per_cpu(node_domains, i).sd;
8514 set_domain_attribute(sd, attr);
8515 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8519 cpumask_and(sched_domain_span(sd),
8520 sched_domain_span(sd), cpu_map);
8524 sd = &per_cpu(phys_domains, i).sd;
8526 set_domain_attribute(sd, attr);
8527 cpumask_copy(sched_domain_span(sd), nodemask);
8531 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
8533 #ifdef CONFIG_SCHED_MC
8535 sd = &per_cpu(core_domains, i).sd;
8537 set_domain_attribute(sd, attr);
8538 cpumask_and(sched_domain_span(sd), cpu_map,
8539 cpu_coregroup_mask(i));
8542 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
8545 #ifdef CONFIG_SCHED_SMT
8547 sd = &per_cpu(cpu_domains, i).sd;
8548 SD_INIT(sd, SIBLING);
8549 set_domain_attribute(sd, attr);
8550 cpumask_and(sched_domain_span(sd),
8551 topology_thread_cpumask(i), cpu_map);
8554 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
8558 #ifdef CONFIG_SCHED_SMT
8559 /* Set up CPU (sibling) groups */
8560 for_each_cpu(i, cpu_map) {
8561 cpumask_and(this_sibling_map,
8562 topology_thread_cpumask(i), cpu_map);
8563 if (i != cpumask_first(this_sibling_map))
8566 init_sched_build_groups(this_sibling_map, cpu_map,
8568 send_covered, tmpmask);
8572 #ifdef CONFIG_SCHED_MC
8573 /* Set up multi-core groups */
8574 for_each_cpu(i, cpu_map) {
8575 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
8576 if (i != cpumask_first(this_core_map))
8579 init_sched_build_groups(this_core_map, cpu_map,
8581 send_covered, tmpmask);
8585 /* Set up physical groups */
8586 for (i = 0; i < nr_node_ids; i++) {
8587 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8588 if (cpumask_empty(nodemask))
8591 init_sched_build_groups(nodemask, cpu_map,
8593 send_covered, tmpmask);
8597 /* Set up node groups */
8599 init_sched_build_groups(cpu_map, cpu_map,
8600 &cpu_to_allnodes_group,
8601 send_covered, tmpmask);
8604 for (i = 0; i < nr_node_ids; i++) {
8605 /* Set up node groups */
8606 struct sched_group *sg, *prev;
8609 cpumask_clear(covered);
8610 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8611 if (cpumask_empty(nodemask)) {
8612 sched_group_nodes[i] = NULL;
8616 sched_domain_node_span(i, domainspan);
8617 cpumask_and(domainspan, domainspan, cpu_map);
8619 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8622 printk(KERN_WARNING "Can not alloc domain group for "
8626 sched_group_nodes[i] = sg;
8627 for_each_cpu(j, nodemask) {
8628 struct sched_domain *sd;
8630 sd = &per_cpu(node_domains, j).sd;
8633 sg->__cpu_power = 0;
8634 cpumask_copy(sched_group_cpus(sg), nodemask);
8636 cpumask_or(covered, covered, nodemask);
8639 for (j = 0; j < nr_node_ids; j++) {
8640 int n = (i + j) % nr_node_ids;
8642 cpumask_complement(notcovered, covered);
8643 cpumask_and(tmpmask, notcovered, cpu_map);
8644 cpumask_and(tmpmask, tmpmask, domainspan);
8645 if (cpumask_empty(tmpmask))
8648 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
8649 if (cpumask_empty(tmpmask))
8652 sg = kmalloc_node(sizeof(struct sched_group) +
8657 "Can not alloc domain group for node %d\n", j);
8660 sg->__cpu_power = 0;
8661 cpumask_copy(sched_group_cpus(sg), tmpmask);
8662 sg->next = prev->next;
8663 cpumask_or(covered, covered, tmpmask);
8670 /* Calculate CPU power for physical packages and nodes */
8671 #ifdef CONFIG_SCHED_SMT
8672 for_each_cpu(i, cpu_map) {
8673 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
8675 init_sched_groups_power(i, sd);
8678 #ifdef CONFIG_SCHED_MC
8679 for_each_cpu(i, cpu_map) {
8680 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
8682 init_sched_groups_power(i, sd);
8686 for_each_cpu(i, cpu_map) {
8687 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
8689 init_sched_groups_power(i, sd);
8693 for (i = 0; i < nr_node_ids; i++)
8694 init_numa_sched_groups_power(sched_group_nodes[i]);
8697 struct sched_group *sg;
8699 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8701 init_numa_sched_groups_power(sg);
8705 /* Attach the domains */
8706 for_each_cpu(i, cpu_map) {
8707 struct sched_domain *sd;
8708 #ifdef CONFIG_SCHED_SMT
8709 sd = &per_cpu(cpu_domains, i).sd;
8710 #elif defined(CONFIG_SCHED_MC)
8711 sd = &per_cpu(core_domains, i).sd;
8713 sd = &per_cpu(phys_domains, i).sd;
8715 cpu_attach_domain(sd, rd, i);
8721 free_cpumask_var(tmpmask);
8723 free_cpumask_var(send_covered);
8725 free_cpumask_var(this_core_map);
8726 free_this_sibling_map:
8727 free_cpumask_var(this_sibling_map);
8729 free_cpumask_var(nodemask);
8732 free_cpumask_var(notcovered);
8734 free_cpumask_var(covered);
8736 free_cpumask_var(domainspan);
8743 kfree(sched_group_nodes);
8749 free_sched_groups(cpu_map, tmpmask);
8750 free_rootdomain(rd);
8755 static int build_sched_domains(const struct cpumask *cpu_map)
8757 return __build_sched_domains(cpu_map, NULL);
8760 static struct cpumask *doms_cur; /* current sched domains */
8761 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8762 static struct sched_domain_attr *dattr_cur;
8763 /* attribues of custom domains in 'doms_cur' */
8766 * Special case: If a kmalloc of a doms_cur partition (array of
8767 * cpumask) fails, then fallback to a single sched domain,
8768 * as determined by the single cpumask fallback_doms.
8770 static cpumask_var_t fallback_doms;
8773 * arch_update_cpu_topology lets virtualized architectures update the
8774 * cpu core maps. It is supposed to return 1 if the topology changed
8775 * or 0 if it stayed the same.
8777 int __attribute__((weak)) arch_update_cpu_topology(void)
8783 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8784 * For now this just excludes isolated cpus, but could be used to
8785 * exclude other special cases in the future.
8787 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8791 arch_update_cpu_topology();
8793 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8795 doms_cur = fallback_doms;
8796 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8798 err = build_sched_domains(doms_cur);
8799 register_sched_domain_sysctl();
8804 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8805 struct cpumask *tmpmask)
8807 free_sched_groups(cpu_map, tmpmask);
8811 * Detach sched domains from a group of cpus specified in cpu_map
8812 * These cpus will now be attached to the NULL domain
8814 static void detach_destroy_domains(const struct cpumask *cpu_map)
8816 /* Save because hotplug lock held. */
8817 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8820 for_each_cpu(i, cpu_map)
8821 cpu_attach_domain(NULL, &def_root_domain, i);
8822 synchronize_sched();
8823 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8826 /* handle null as "default" */
8827 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8828 struct sched_domain_attr *new, int idx_new)
8830 struct sched_domain_attr tmp;
8837 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8838 new ? (new + idx_new) : &tmp,
8839 sizeof(struct sched_domain_attr));
8843 * Partition sched domains as specified by the 'ndoms_new'
8844 * cpumasks in the array doms_new[] of cpumasks. This compares
8845 * doms_new[] to the current sched domain partitioning, doms_cur[].
8846 * It destroys each deleted domain and builds each new domain.
8848 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8849 * The masks don't intersect (don't overlap.) We should setup one
8850 * sched domain for each mask. CPUs not in any of the cpumasks will
8851 * not be load balanced. If the same cpumask appears both in the
8852 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8855 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8856 * ownership of it and will kfree it when done with it. If the caller
8857 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8858 * ndoms_new == 1, and partition_sched_domains() will fallback to
8859 * the single partition 'fallback_doms', it also forces the domains
8862 * If doms_new == NULL it will be replaced with cpu_online_mask.
8863 * ndoms_new == 0 is a special case for destroying existing domains,
8864 * and it will not create the default domain.
8866 * Call with hotplug lock held
8868 /* FIXME: Change to struct cpumask *doms_new[] */
8869 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8870 struct sched_domain_attr *dattr_new)
8875 mutex_lock(&sched_domains_mutex);
8877 /* always unregister in case we don't destroy any domains */
8878 unregister_sched_domain_sysctl();
8880 /* Let architecture update cpu core mappings. */
8881 new_topology = arch_update_cpu_topology();
8883 n = doms_new ? ndoms_new : 0;
8885 /* Destroy deleted domains */
8886 for (i = 0; i < ndoms_cur; i++) {
8887 for (j = 0; j < n && !new_topology; j++) {
8888 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8889 && dattrs_equal(dattr_cur, i, dattr_new, j))
8892 /* no match - a current sched domain not in new doms_new[] */
8893 detach_destroy_domains(doms_cur + i);
8898 if (doms_new == NULL) {
8900 doms_new = fallback_doms;
8901 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8902 WARN_ON_ONCE(dattr_new);
8905 /* Build new domains */
8906 for (i = 0; i < ndoms_new; i++) {
8907 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8908 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8909 && dattrs_equal(dattr_new, i, dattr_cur, j))
8912 /* no match - add a new doms_new */
8913 __build_sched_domains(doms_new + i,
8914 dattr_new ? dattr_new + i : NULL);
8919 /* Remember the new sched domains */
8920 if (doms_cur != fallback_doms)
8922 kfree(dattr_cur); /* kfree(NULL) is safe */
8923 doms_cur = doms_new;
8924 dattr_cur = dattr_new;
8925 ndoms_cur = ndoms_new;
8927 register_sched_domain_sysctl();
8929 mutex_unlock(&sched_domains_mutex);
8932 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8933 static void arch_reinit_sched_domains(void)
8937 /* Destroy domains first to force the rebuild */
8938 partition_sched_domains(0, NULL, NULL);
8940 rebuild_sched_domains();
8944 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8946 unsigned int level = 0;
8948 if (sscanf(buf, "%u", &level) != 1)
8952 * level is always be positive so don't check for
8953 * level < POWERSAVINGS_BALANCE_NONE which is 0
8954 * What happens on 0 or 1 byte write,
8955 * need to check for count as well?
8958 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8962 sched_smt_power_savings = level;
8964 sched_mc_power_savings = level;
8966 arch_reinit_sched_domains();
8971 #ifdef CONFIG_SCHED_MC
8972 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8975 return sprintf(page, "%u\n", sched_mc_power_savings);
8977 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8978 const char *buf, size_t count)
8980 return sched_power_savings_store(buf, count, 0);
8982 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8983 sched_mc_power_savings_show,
8984 sched_mc_power_savings_store);
8987 #ifdef CONFIG_SCHED_SMT
8988 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8991 return sprintf(page, "%u\n", sched_smt_power_savings);
8993 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8994 const char *buf, size_t count)
8996 return sched_power_savings_store(buf, count, 1);
8998 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8999 sched_smt_power_savings_show,
9000 sched_smt_power_savings_store);
9003 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9007 #ifdef CONFIG_SCHED_SMT
9009 err = sysfs_create_file(&cls->kset.kobj,
9010 &attr_sched_smt_power_savings.attr);
9012 #ifdef CONFIG_SCHED_MC
9013 if (!err && mc_capable())
9014 err = sysfs_create_file(&cls->kset.kobj,
9015 &attr_sched_mc_power_savings.attr);
9019 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9021 #ifndef CONFIG_CPUSETS
9023 * Add online and remove offline CPUs from the scheduler domains.
9024 * When cpusets are enabled they take over this function.
9026 static int update_sched_domains(struct notifier_block *nfb,
9027 unsigned long action, void *hcpu)
9031 case CPU_ONLINE_FROZEN:
9033 case CPU_DEAD_FROZEN:
9034 partition_sched_domains(1, NULL, NULL);
9043 static int update_runtime(struct notifier_block *nfb,
9044 unsigned long action, void *hcpu)
9046 int cpu = (int)(long)hcpu;
9049 case CPU_DOWN_PREPARE:
9050 case CPU_DOWN_PREPARE_FROZEN:
9051 disable_runtime(cpu_rq(cpu));
9054 case CPU_DOWN_FAILED:
9055 case CPU_DOWN_FAILED_FROZEN:
9057 case CPU_ONLINE_FROZEN:
9058 enable_runtime(cpu_rq(cpu));
9066 void __init sched_init_smp(void)
9068 cpumask_var_t non_isolated_cpus;
9070 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9072 #if defined(CONFIG_NUMA)
9073 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9075 BUG_ON(sched_group_nodes_bycpu == NULL);
9078 mutex_lock(&sched_domains_mutex);
9079 arch_init_sched_domains(cpu_online_mask);
9080 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9081 if (cpumask_empty(non_isolated_cpus))
9082 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9083 mutex_unlock(&sched_domains_mutex);
9086 #ifndef CONFIG_CPUSETS
9087 /* XXX: Theoretical race here - CPU may be hotplugged now */
9088 hotcpu_notifier(update_sched_domains, 0);
9091 /* RT runtime code needs to handle some hotplug events */
9092 hotcpu_notifier(update_runtime, 0);
9096 /* Move init over to a non-isolated CPU */
9097 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9099 sched_init_granularity();
9100 free_cpumask_var(non_isolated_cpus);
9102 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9103 init_sched_rt_class();
9106 void __init sched_init_smp(void)
9108 sched_init_granularity();
9110 #endif /* CONFIG_SMP */
9112 const_debug unsigned int sysctl_timer_migration = 1;
9114 int in_sched_functions(unsigned long addr)
9116 return in_lock_functions(addr) ||
9117 (addr >= (unsigned long)__sched_text_start
9118 && addr < (unsigned long)__sched_text_end);
9121 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9123 cfs_rq->tasks_timeline = RB_ROOT;
9124 INIT_LIST_HEAD(&cfs_rq->tasks);
9125 #ifdef CONFIG_FAIR_GROUP_SCHED
9128 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9131 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9133 struct rt_prio_array *array;
9136 array = &rt_rq->active;
9137 for (i = 0; i < MAX_RT_PRIO; i++) {
9138 INIT_LIST_HEAD(array->queue + i);
9139 __clear_bit(i, array->bitmap);
9141 /* delimiter for bitsearch: */
9142 __set_bit(MAX_RT_PRIO, array->bitmap);
9144 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9145 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9147 rt_rq->highest_prio.next = MAX_RT_PRIO;
9151 rt_rq->rt_nr_migratory = 0;
9152 rt_rq->overloaded = 0;
9153 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9157 rt_rq->rt_throttled = 0;
9158 rt_rq->rt_runtime = 0;
9159 spin_lock_init(&rt_rq->rt_runtime_lock);
9161 #ifdef CONFIG_RT_GROUP_SCHED
9162 rt_rq->rt_nr_boosted = 0;
9167 #ifdef CONFIG_FAIR_GROUP_SCHED
9168 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9169 struct sched_entity *se, int cpu, int add,
9170 struct sched_entity *parent)
9172 struct rq *rq = cpu_rq(cpu);
9173 tg->cfs_rq[cpu] = cfs_rq;
9174 init_cfs_rq(cfs_rq, rq);
9177 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9180 /* se could be NULL for init_task_group */
9185 se->cfs_rq = &rq->cfs;
9187 se->cfs_rq = parent->my_q;
9190 se->load.weight = tg->shares;
9191 se->load.inv_weight = 0;
9192 se->parent = parent;
9196 #ifdef CONFIG_RT_GROUP_SCHED
9197 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9198 struct sched_rt_entity *rt_se, int cpu, int add,
9199 struct sched_rt_entity *parent)
9201 struct rq *rq = cpu_rq(cpu);
9203 tg->rt_rq[cpu] = rt_rq;
9204 init_rt_rq(rt_rq, rq);
9206 rt_rq->rt_se = rt_se;
9207 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9209 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9211 tg->rt_se[cpu] = rt_se;
9216 rt_se->rt_rq = &rq->rt;
9218 rt_se->rt_rq = parent->my_q;
9220 rt_se->my_q = rt_rq;
9221 rt_se->parent = parent;
9222 INIT_LIST_HEAD(&rt_se->run_list);
9226 void __init sched_init(void)
9229 unsigned long alloc_size = 0, ptr;
9231 #ifdef CONFIG_FAIR_GROUP_SCHED
9232 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9234 #ifdef CONFIG_RT_GROUP_SCHED
9235 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9237 #ifdef CONFIG_USER_SCHED
9240 #ifdef CONFIG_CPUMASK_OFFSTACK
9241 alloc_size += num_possible_cpus() * cpumask_size();
9244 * As sched_init() is called before page_alloc is setup,
9245 * we use alloc_bootmem().
9248 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9250 #ifdef CONFIG_FAIR_GROUP_SCHED
9251 init_task_group.se = (struct sched_entity **)ptr;
9252 ptr += nr_cpu_ids * sizeof(void **);
9254 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9255 ptr += nr_cpu_ids * sizeof(void **);
9257 #ifdef CONFIG_USER_SCHED
9258 root_task_group.se = (struct sched_entity **)ptr;
9259 ptr += nr_cpu_ids * sizeof(void **);
9261 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9262 ptr += nr_cpu_ids * sizeof(void **);
9263 #endif /* CONFIG_USER_SCHED */
9264 #endif /* CONFIG_FAIR_GROUP_SCHED */
9265 #ifdef CONFIG_RT_GROUP_SCHED
9266 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9267 ptr += nr_cpu_ids * sizeof(void **);
9269 init_task_group.rt_rq = (struct rt_rq **)ptr;
9270 ptr += nr_cpu_ids * sizeof(void **);
9272 #ifdef CONFIG_USER_SCHED
9273 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9274 ptr += nr_cpu_ids * sizeof(void **);
9276 root_task_group.rt_rq = (struct rt_rq **)ptr;
9277 ptr += nr_cpu_ids * sizeof(void **);
9278 #endif /* CONFIG_USER_SCHED */
9279 #endif /* CONFIG_RT_GROUP_SCHED */
9280 #ifdef CONFIG_CPUMASK_OFFSTACK
9281 for_each_possible_cpu(i) {
9282 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9283 ptr += cpumask_size();
9285 #endif /* CONFIG_CPUMASK_OFFSTACK */
9289 init_defrootdomain();
9292 init_rt_bandwidth(&def_rt_bandwidth,
9293 global_rt_period(), global_rt_runtime());
9295 #ifdef CONFIG_RT_GROUP_SCHED
9296 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9297 global_rt_period(), global_rt_runtime());
9298 #ifdef CONFIG_USER_SCHED
9299 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9300 global_rt_period(), RUNTIME_INF);
9301 #endif /* CONFIG_USER_SCHED */
9302 #endif /* CONFIG_RT_GROUP_SCHED */
9304 #ifdef CONFIG_GROUP_SCHED
9305 list_add(&init_task_group.list, &task_groups);
9306 INIT_LIST_HEAD(&init_task_group.children);
9308 #ifdef CONFIG_USER_SCHED
9309 INIT_LIST_HEAD(&root_task_group.children);
9310 init_task_group.parent = &root_task_group;
9311 list_add(&init_task_group.siblings, &root_task_group.children);
9312 #endif /* CONFIG_USER_SCHED */
9313 #endif /* CONFIG_GROUP_SCHED */
9315 for_each_possible_cpu(i) {
9319 spin_lock_init(&rq->lock);
9321 rq->calc_load_active = 0;
9322 rq->calc_load_update = jiffies + LOAD_FREQ;
9323 init_cfs_rq(&rq->cfs, rq);
9324 init_rt_rq(&rq->rt, rq);
9325 #ifdef CONFIG_FAIR_GROUP_SCHED
9326 init_task_group.shares = init_task_group_load;
9327 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9328 #ifdef CONFIG_CGROUP_SCHED
9330 * How much cpu bandwidth does init_task_group get?
9332 * In case of task-groups formed thr' the cgroup filesystem, it
9333 * gets 100% of the cpu resources in the system. This overall
9334 * system cpu resource is divided among the tasks of
9335 * init_task_group and its child task-groups in a fair manner,
9336 * based on each entity's (task or task-group's) weight
9337 * (se->load.weight).
9339 * In other words, if init_task_group has 10 tasks of weight
9340 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9341 * then A0's share of the cpu resource is:
9343 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9345 * We achieve this by letting init_task_group's tasks sit
9346 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9348 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9349 #elif defined CONFIG_USER_SCHED
9350 root_task_group.shares = NICE_0_LOAD;
9351 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9353 * In case of task-groups formed thr' the user id of tasks,
9354 * init_task_group represents tasks belonging to root user.
9355 * Hence it forms a sibling of all subsequent groups formed.
9356 * In this case, init_task_group gets only a fraction of overall
9357 * system cpu resource, based on the weight assigned to root
9358 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9359 * by letting tasks of init_task_group sit in a separate cfs_rq
9360 * (init_cfs_rq) and having one entity represent this group of
9361 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9363 init_tg_cfs_entry(&init_task_group,
9364 &per_cpu(init_cfs_rq, i),
9365 &per_cpu(init_sched_entity, i), i, 1,
9366 root_task_group.se[i]);
9369 #endif /* CONFIG_FAIR_GROUP_SCHED */
9371 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9372 #ifdef CONFIG_RT_GROUP_SCHED
9373 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9374 #ifdef CONFIG_CGROUP_SCHED
9375 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9376 #elif defined CONFIG_USER_SCHED
9377 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9378 init_tg_rt_entry(&init_task_group,
9379 &per_cpu(init_rt_rq, i),
9380 &per_cpu(init_sched_rt_entity, i), i, 1,
9381 root_task_group.rt_se[i]);
9385 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9386 rq->cpu_load[j] = 0;
9390 rq->active_balance = 0;
9391 rq->next_balance = jiffies;
9395 rq->migration_thread = NULL;
9396 INIT_LIST_HEAD(&rq->migration_queue);
9397 rq_attach_root(rq, &def_root_domain);
9400 atomic_set(&rq->nr_iowait, 0);
9403 set_load_weight(&init_task);
9405 #ifdef CONFIG_PREEMPT_NOTIFIERS
9406 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9410 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9413 #ifdef CONFIG_RT_MUTEXES
9414 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9418 * The boot idle thread does lazy MMU switching as well:
9420 atomic_inc(&init_mm.mm_count);
9421 enter_lazy_tlb(&init_mm, current);
9424 * Make us the idle thread. Technically, schedule() should not be
9425 * called from this thread, however somewhere below it might be,
9426 * but because we are the idle thread, we just pick up running again
9427 * when this runqueue becomes "idle".
9429 init_idle(current, smp_processor_id());
9431 calc_load_update = jiffies + LOAD_FREQ;
9434 * During early bootup we pretend to be a normal task:
9436 current->sched_class = &fair_sched_class;
9438 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9439 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9442 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9443 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9445 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9448 perf_counter_init();
9450 scheduler_running = 1;
9453 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9454 static inline int preempt_count_equals(int preempt_offset)
9456 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9458 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9461 void __might_sleep(char *file, int line, int preempt_offset)
9464 static unsigned long prev_jiffy; /* ratelimiting */
9466 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9467 system_state != SYSTEM_RUNNING || oops_in_progress)
9469 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9471 prev_jiffy = jiffies;
9474 "BUG: sleeping function called from invalid context at %s:%d\n",
9477 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9478 in_atomic(), irqs_disabled(),
9479 current->pid, current->comm);
9481 debug_show_held_locks(current);
9482 if (irqs_disabled())
9483 print_irqtrace_events(current);
9487 EXPORT_SYMBOL(__might_sleep);
9490 #ifdef CONFIG_MAGIC_SYSRQ
9491 static void normalize_task(struct rq *rq, struct task_struct *p)
9495 update_rq_clock(rq);
9496 on_rq = p->se.on_rq;
9498 deactivate_task(rq, p, 0);
9499 __setscheduler(rq, p, SCHED_NORMAL, 0);
9501 activate_task(rq, p, 0);
9502 resched_task(rq->curr);
9506 void normalize_rt_tasks(void)
9508 struct task_struct *g, *p;
9509 unsigned long flags;
9512 read_lock_irqsave(&tasklist_lock, flags);
9513 do_each_thread(g, p) {
9515 * Only normalize user tasks:
9520 p->se.exec_start = 0;
9521 #ifdef CONFIG_SCHEDSTATS
9522 p->se.wait_start = 0;
9523 p->se.sleep_start = 0;
9524 p->se.block_start = 0;
9529 * Renice negative nice level userspace
9532 if (TASK_NICE(p) < 0 && p->mm)
9533 set_user_nice(p, 0);
9537 spin_lock(&p->pi_lock);
9538 rq = __task_rq_lock(p);
9540 normalize_task(rq, p);
9542 __task_rq_unlock(rq);
9543 spin_unlock(&p->pi_lock);
9544 } while_each_thread(g, p);
9546 read_unlock_irqrestore(&tasklist_lock, flags);
9549 #endif /* CONFIG_MAGIC_SYSRQ */
9553 * These functions are only useful for the IA64 MCA handling.
9555 * They can only be called when the whole system has been
9556 * stopped - every CPU needs to be quiescent, and no scheduling
9557 * activity can take place. Using them for anything else would
9558 * be a serious bug, and as a result, they aren't even visible
9559 * under any other configuration.
9563 * curr_task - return the current task for a given cpu.
9564 * @cpu: the processor in question.
9566 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9568 struct task_struct *curr_task(int cpu)
9570 return cpu_curr(cpu);
9574 * set_curr_task - set the current task for a given cpu.
9575 * @cpu: the processor in question.
9576 * @p: the task pointer to set.
9578 * Description: This function must only be used when non-maskable interrupts
9579 * are serviced on a separate stack. It allows the architecture to switch the
9580 * notion of the current task on a cpu in a non-blocking manner. This function
9581 * must be called with all CPU's synchronized, and interrupts disabled, the
9582 * and caller must save the original value of the current task (see
9583 * curr_task() above) and restore that value before reenabling interrupts and
9584 * re-starting the system.
9586 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9588 void set_curr_task(int cpu, struct task_struct *p)
9595 #ifdef CONFIG_FAIR_GROUP_SCHED
9596 static void free_fair_sched_group(struct task_group *tg)
9600 for_each_possible_cpu(i) {
9602 kfree(tg->cfs_rq[i]);
9612 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9614 struct cfs_rq *cfs_rq;
9615 struct sched_entity *se;
9619 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9622 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9626 tg->shares = NICE_0_LOAD;
9628 for_each_possible_cpu(i) {
9631 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9632 GFP_KERNEL, cpu_to_node(i));
9636 se = kzalloc_node(sizeof(struct sched_entity),
9637 GFP_KERNEL, cpu_to_node(i));
9641 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9650 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9652 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9653 &cpu_rq(cpu)->leaf_cfs_rq_list);
9656 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9658 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9660 #else /* !CONFG_FAIR_GROUP_SCHED */
9661 static inline void free_fair_sched_group(struct task_group *tg)
9666 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9671 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9675 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9678 #endif /* CONFIG_FAIR_GROUP_SCHED */
9680 #ifdef CONFIG_RT_GROUP_SCHED
9681 static void free_rt_sched_group(struct task_group *tg)
9685 destroy_rt_bandwidth(&tg->rt_bandwidth);
9687 for_each_possible_cpu(i) {
9689 kfree(tg->rt_rq[i]);
9691 kfree(tg->rt_se[i]);
9699 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9701 struct rt_rq *rt_rq;
9702 struct sched_rt_entity *rt_se;
9706 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9709 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9713 init_rt_bandwidth(&tg->rt_bandwidth,
9714 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9716 for_each_possible_cpu(i) {
9719 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9720 GFP_KERNEL, cpu_to_node(i));
9724 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9725 GFP_KERNEL, cpu_to_node(i));
9729 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9738 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9740 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9741 &cpu_rq(cpu)->leaf_rt_rq_list);
9744 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9746 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9748 #else /* !CONFIG_RT_GROUP_SCHED */
9749 static inline void free_rt_sched_group(struct task_group *tg)
9754 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9759 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9763 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9766 #endif /* CONFIG_RT_GROUP_SCHED */
9768 #ifdef CONFIG_GROUP_SCHED
9769 static void free_sched_group(struct task_group *tg)
9771 free_fair_sched_group(tg);
9772 free_rt_sched_group(tg);
9776 /* allocate runqueue etc for a new task group */
9777 struct task_group *sched_create_group(struct task_group *parent)
9779 struct task_group *tg;
9780 unsigned long flags;
9783 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9785 return ERR_PTR(-ENOMEM);
9787 if (!alloc_fair_sched_group(tg, parent))
9790 if (!alloc_rt_sched_group(tg, parent))
9793 spin_lock_irqsave(&task_group_lock, flags);
9794 for_each_possible_cpu(i) {
9795 register_fair_sched_group(tg, i);
9796 register_rt_sched_group(tg, i);
9798 list_add_rcu(&tg->list, &task_groups);
9800 WARN_ON(!parent); /* root should already exist */
9802 tg->parent = parent;
9803 INIT_LIST_HEAD(&tg->children);
9804 list_add_rcu(&tg->siblings, &parent->children);
9805 spin_unlock_irqrestore(&task_group_lock, flags);
9810 free_sched_group(tg);
9811 return ERR_PTR(-ENOMEM);
9814 /* rcu callback to free various structures associated with a task group */
9815 static void free_sched_group_rcu(struct rcu_head *rhp)
9817 /* now it should be safe to free those cfs_rqs */
9818 free_sched_group(container_of(rhp, struct task_group, rcu));
9821 /* Destroy runqueue etc associated with a task group */
9822 void sched_destroy_group(struct task_group *tg)
9824 unsigned long flags;
9827 spin_lock_irqsave(&task_group_lock, flags);
9828 for_each_possible_cpu(i) {
9829 unregister_fair_sched_group(tg, i);
9830 unregister_rt_sched_group(tg, i);
9832 list_del_rcu(&tg->list);
9833 list_del_rcu(&tg->siblings);
9834 spin_unlock_irqrestore(&task_group_lock, flags);
9836 /* wait for possible concurrent references to cfs_rqs complete */
9837 call_rcu(&tg->rcu, free_sched_group_rcu);
9840 /* change task's runqueue when it moves between groups.
9841 * The caller of this function should have put the task in its new group
9842 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9843 * reflect its new group.
9845 void sched_move_task(struct task_struct *tsk)
9848 unsigned long flags;
9851 rq = task_rq_lock(tsk, &flags);
9853 update_rq_clock(rq);
9855 running = task_current(rq, tsk);
9856 on_rq = tsk->se.on_rq;
9859 dequeue_task(rq, tsk, 0);
9860 if (unlikely(running))
9861 tsk->sched_class->put_prev_task(rq, tsk);
9863 set_task_rq(tsk, task_cpu(tsk));
9865 #ifdef CONFIG_FAIR_GROUP_SCHED
9866 if (tsk->sched_class->moved_group)
9867 tsk->sched_class->moved_group(tsk);
9870 if (unlikely(running))
9871 tsk->sched_class->set_curr_task(rq);
9873 enqueue_task(rq, tsk, 0);
9875 task_rq_unlock(rq, &flags);
9877 #endif /* CONFIG_GROUP_SCHED */
9879 #ifdef CONFIG_FAIR_GROUP_SCHED
9880 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9882 struct cfs_rq *cfs_rq = se->cfs_rq;
9887 dequeue_entity(cfs_rq, se, 0);
9889 se->load.weight = shares;
9890 se->load.inv_weight = 0;
9893 enqueue_entity(cfs_rq, se, 0);
9896 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9898 struct cfs_rq *cfs_rq = se->cfs_rq;
9899 struct rq *rq = cfs_rq->rq;
9900 unsigned long flags;
9902 spin_lock_irqsave(&rq->lock, flags);
9903 __set_se_shares(se, shares);
9904 spin_unlock_irqrestore(&rq->lock, flags);
9907 static DEFINE_MUTEX(shares_mutex);
9909 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9912 unsigned long flags;
9915 * We can't change the weight of the root cgroup.
9920 if (shares < MIN_SHARES)
9921 shares = MIN_SHARES;
9922 else if (shares > MAX_SHARES)
9923 shares = MAX_SHARES;
9925 mutex_lock(&shares_mutex);
9926 if (tg->shares == shares)
9929 spin_lock_irqsave(&task_group_lock, flags);
9930 for_each_possible_cpu(i)
9931 unregister_fair_sched_group(tg, i);
9932 list_del_rcu(&tg->siblings);
9933 spin_unlock_irqrestore(&task_group_lock, flags);
9935 /* wait for any ongoing reference to this group to finish */
9936 synchronize_sched();
9939 * Now we are free to modify the group's share on each cpu
9940 * w/o tripping rebalance_share or load_balance_fair.
9942 tg->shares = shares;
9943 for_each_possible_cpu(i) {
9947 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9948 set_se_shares(tg->se[i], shares);
9952 * Enable load balance activity on this group, by inserting it back on
9953 * each cpu's rq->leaf_cfs_rq_list.
9955 spin_lock_irqsave(&task_group_lock, flags);
9956 for_each_possible_cpu(i)
9957 register_fair_sched_group(tg, i);
9958 list_add_rcu(&tg->siblings, &tg->parent->children);
9959 spin_unlock_irqrestore(&task_group_lock, flags);
9961 mutex_unlock(&shares_mutex);
9965 unsigned long sched_group_shares(struct task_group *tg)
9971 #ifdef CONFIG_RT_GROUP_SCHED
9973 * Ensure that the real time constraints are schedulable.
9975 static DEFINE_MUTEX(rt_constraints_mutex);
9977 static unsigned long to_ratio(u64 period, u64 runtime)
9979 if (runtime == RUNTIME_INF)
9982 return div64_u64(runtime << 20, period);
9985 /* Must be called with tasklist_lock held */
9986 static inline int tg_has_rt_tasks(struct task_group *tg)
9988 struct task_struct *g, *p;
9990 do_each_thread(g, p) {
9991 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9993 } while_each_thread(g, p);
9998 struct rt_schedulable_data {
9999 struct task_group *tg;
10004 static int tg_schedulable(struct task_group *tg, void *data)
10006 struct rt_schedulable_data *d = data;
10007 struct task_group *child;
10008 unsigned long total, sum = 0;
10009 u64 period, runtime;
10011 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10012 runtime = tg->rt_bandwidth.rt_runtime;
10015 period = d->rt_period;
10016 runtime = d->rt_runtime;
10019 #ifdef CONFIG_USER_SCHED
10020 if (tg == &root_task_group) {
10021 period = global_rt_period();
10022 runtime = global_rt_runtime();
10027 * Cannot have more runtime than the period.
10029 if (runtime > period && runtime != RUNTIME_INF)
10033 * Ensure we don't starve existing RT tasks.
10035 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10038 total = to_ratio(period, runtime);
10041 * Nobody can have more than the global setting allows.
10043 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10047 * The sum of our children's runtime should not exceed our own.
10049 list_for_each_entry_rcu(child, &tg->children, siblings) {
10050 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10051 runtime = child->rt_bandwidth.rt_runtime;
10053 if (child == d->tg) {
10054 period = d->rt_period;
10055 runtime = d->rt_runtime;
10058 sum += to_ratio(period, runtime);
10067 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10069 struct rt_schedulable_data data = {
10071 .rt_period = period,
10072 .rt_runtime = runtime,
10075 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10078 static int tg_set_bandwidth(struct task_group *tg,
10079 u64 rt_period, u64 rt_runtime)
10083 mutex_lock(&rt_constraints_mutex);
10084 read_lock(&tasklist_lock);
10085 err = __rt_schedulable(tg, rt_period, rt_runtime);
10089 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10090 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10091 tg->rt_bandwidth.rt_runtime = rt_runtime;
10093 for_each_possible_cpu(i) {
10094 struct rt_rq *rt_rq = tg->rt_rq[i];
10096 spin_lock(&rt_rq->rt_runtime_lock);
10097 rt_rq->rt_runtime = rt_runtime;
10098 spin_unlock(&rt_rq->rt_runtime_lock);
10100 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10102 read_unlock(&tasklist_lock);
10103 mutex_unlock(&rt_constraints_mutex);
10108 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10110 u64 rt_runtime, rt_period;
10112 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10113 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10114 if (rt_runtime_us < 0)
10115 rt_runtime = RUNTIME_INF;
10117 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10120 long sched_group_rt_runtime(struct task_group *tg)
10124 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10127 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10128 do_div(rt_runtime_us, NSEC_PER_USEC);
10129 return rt_runtime_us;
10132 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10134 u64 rt_runtime, rt_period;
10136 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10137 rt_runtime = tg->rt_bandwidth.rt_runtime;
10139 if (rt_period == 0)
10142 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10145 long sched_group_rt_period(struct task_group *tg)
10149 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10150 do_div(rt_period_us, NSEC_PER_USEC);
10151 return rt_period_us;
10154 static int sched_rt_global_constraints(void)
10156 u64 runtime, period;
10159 if (sysctl_sched_rt_period <= 0)
10162 runtime = global_rt_runtime();
10163 period = global_rt_period();
10166 * Sanity check on the sysctl variables.
10168 if (runtime > period && runtime != RUNTIME_INF)
10171 mutex_lock(&rt_constraints_mutex);
10172 read_lock(&tasklist_lock);
10173 ret = __rt_schedulable(NULL, 0, 0);
10174 read_unlock(&tasklist_lock);
10175 mutex_unlock(&rt_constraints_mutex);
10180 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10182 /* Don't accept realtime tasks when there is no way for them to run */
10183 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10189 #else /* !CONFIG_RT_GROUP_SCHED */
10190 static int sched_rt_global_constraints(void)
10192 unsigned long flags;
10195 if (sysctl_sched_rt_period <= 0)
10199 * There's always some RT tasks in the root group
10200 * -- migration, kstopmachine etc..
10202 if (sysctl_sched_rt_runtime == 0)
10205 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10206 for_each_possible_cpu(i) {
10207 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10209 spin_lock(&rt_rq->rt_runtime_lock);
10210 rt_rq->rt_runtime = global_rt_runtime();
10211 spin_unlock(&rt_rq->rt_runtime_lock);
10213 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10217 #endif /* CONFIG_RT_GROUP_SCHED */
10219 int sched_rt_handler(struct ctl_table *table, int write,
10220 struct file *filp, void __user *buffer, size_t *lenp,
10224 int old_period, old_runtime;
10225 static DEFINE_MUTEX(mutex);
10227 mutex_lock(&mutex);
10228 old_period = sysctl_sched_rt_period;
10229 old_runtime = sysctl_sched_rt_runtime;
10231 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
10233 if (!ret && write) {
10234 ret = sched_rt_global_constraints();
10236 sysctl_sched_rt_period = old_period;
10237 sysctl_sched_rt_runtime = old_runtime;
10239 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10240 def_rt_bandwidth.rt_period =
10241 ns_to_ktime(global_rt_period());
10244 mutex_unlock(&mutex);
10249 #ifdef CONFIG_CGROUP_SCHED
10251 /* return corresponding task_group object of a cgroup */
10252 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10254 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10255 struct task_group, css);
10258 static struct cgroup_subsys_state *
10259 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10261 struct task_group *tg, *parent;
10263 if (!cgrp->parent) {
10264 /* This is early initialization for the top cgroup */
10265 return &init_task_group.css;
10268 parent = cgroup_tg(cgrp->parent);
10269 tg = sched_create_group(parent);
10271 return ERR_PTR(-ENOMEM);
10277 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10279 struct task_group *tg = cgroup_tg(cgrp);
10281 sched_destroy_group(tg);
10285 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10286 struct task_struct *tsk)
10288 #ifdef CONFIG_RT_GROUP_SCHED
10289 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10292 /* We don't support RT-tasks being in separate groups */
10293 if (tsk->sched_class != &fair_sched_class)
10301 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10302 struct cgroup *old_cont, struct task_struct *tsk)
10304 sched_move_task(tsk);
10307 #ifdef CONFIG_FAIR_GROUP_SCHED
10308 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10311 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10314 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10316 struct task_group *tg = cgroup_tg(cgrp);
10318 return (u64) tg->shares;
10320 #endif /* CONFIG_FAIR_GROUP_SCHED */
10322 #ifdef CONFIG_RT_GROUP_SCHED
10323 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10326 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10329 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10331 return sched_group_rt_runtime(cgroup_tg(cgrp));
10334 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10337 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10340 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10342 return sched_group_rt_period(cgroup_tg(cgrp));
10344 #endif /* CONFIG_RT_GROUP_SCHED */
10346 static struct cftype cpu_files[] = {
10347 #ifdef CONFIG_FAIR_GROUP_SCHED
10350 .read_u64 = cpu_shares_read_u64,
10351 .write_u64 = cpu_shares_write_u64,
10354 #ifdef CONFIG_RT_GROUP_SCHED
10356 .name = "rt_runtime_us",
10357 .read_s64 = cpu_rt_runtime_read,
10358 .write_s64 = cpu_rt_runtime_write,
10361 .name = "rt_period_us",
10362 .read_u64 = cpu_rt_period_read_uint,
10363 .write_u64 = cpu_rt_period_write_uint,
10368 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10370 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10373 struct cgroup_subsys cpu_cgroup_subsys = {
10375 .create = cpu_cgroup_create,
10376 .destroy = cpu_cgroup_destroy,
10377 .can_attach = cpu_cgroup_can_attach,
10378 .attach = cpu_cgroup_attach,
10379 .populate = cpu_cgroup_populate,
10380 .subsys_id = cpu_cgroup_subsys_id,
10384 #endif /* CONFIG_CGROUP_SCHED */
10386 #ifdef CONFIG_CGROUP_CPUACCT
10389 * CPU accounting code for task groups.
10391 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10392 * (balbir@in.ibm.com).
10395 /* track cpu usage of a group of tasks and its child groups */
10397 struct cgroup_subsys_state css;
10398 /* cpuusage holds pointer to a u64-type object on every cpu */
10400 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10401 struct cpuacct *parent;
10404 struct cgroup_subsys cpuacct_subsys;
10406 /* return cpu accounting group corresponding to this container */
10407 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10409 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10410 struct cpuacct, css);
10413 /* return cpu accounting group to which this task belongs */
10414 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10416 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10417 struct cpuacct, css);
10420 /* create a new cpu accounting group */
10421 static struct cgroup_subsys_state *cpuacct_create(
10422 struct cgroup_subsys *ss, struct cgroup *cgrp)
10424 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10430 ca->cpuusage = alloc_percpu(u64);
10434 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10435 if (percpu_counter_init(&ca->cpustat[i], 0))
10436 goto out_free_counters;
10439 ca->parent = cgroup_ca(cgrp->parent);
10445 percpu_counter_destroy(&ca->cpustat[i]);
10446 free_percpu(ca->cpuusage);
10450 return ERR_PTR(-ENOMEM);
10453 /* destroy an existing cpu accounting group */
10455 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10457 struct cpuacct *ca = cgroup_ca(cgrp);
10460 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10461 percpu_counter_destroy(&ca->cpustat[i]);
10462 free_percpu(ca->cpuusage);
10466 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10468 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10471 #ifndef CONFIG_64BIT
10473 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10475 spin_lock_irq(&cpu_rq(cpu)->lock);
10477 spin_unlock_irq(&cpu_rq(cpu)->lock);
10485 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10487 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10489 #ifndef CONFIG_64BIT
10491 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10493 spin_lock_irq(&cpu_rq(cpu)->lock);
10495 spin_unlock_irq(&cpu_rq(cpu)->lock);
10501 /* return total cpu usage (in nanoseconds) of a group */
10502 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10504 struct cpuacct *ca = cgroup_ca(cgrp);
10505 u64 totalcpuusage = 0;
10508 for_each_present_cpu(i)
10509 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10511 return totalcpuusage;
10514 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10517 struct cpuacct *ca = cgroup_ca(cgrp);
10526 for_each_present_cpu(i)
10527 cpuacct_cpuusage_write(ca, i, 0);
10533 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10534 struct seq_file *m)
10536 struct cpuacct *ca = cgroup_ca(cgroup);
10540 for_each_present_cpu(i) {
10541 percpu = cpuacct_cpuusage_read(ca, i);
10542 seq_printf(m, "%llu ", (unsigned long long) percpu);
10544 seq_printf(m, "\n");
10548 static const char *cpuacct_stat_desc[] = {
10549 [CPUACCT_STAT_USER] = "user",
10550 [CPUACCT_STAT_SYSTEM] = "system",
10553 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10554 struct cgroup_map_cb *cb)
10556 struct cpuacct *ca = cgroup_ca(cgrp);
10559 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10560 s64 val = percpu_counter_read(&ca->cpustat[i]);
10561 val = cputime64_to_clock_t(val);
10562 cb->fill(cb, cpuacct_stat_desc[i], val);
10567 static struct cftype files[] = {
10570 .read_u64 = cpuusage_read,
10571 .write_u64 = cpuusage_write,
10574 .name = "usage_percpu",
10575 .read_seq_string = cpuacct_percpu_seq_read,
10579 .read_map = cpuacct_stats_show,
10583 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10585 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10589 * charge this task's execution time to its accounting group.
10591 * called with rq->lock held.
10593 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10595 struct cpuacct *ca;
10598 if (unlikely(!cpuacct_subsys.active))
10601 cpu = task_cpu(tsk);
10607 for (; ca; ca = ca->parent) {
10608 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10609 *cpuusage += cputime;
10616 * Charge the system/user time to the task's accounting group.
10618 static void cpuacct_update_stats(struct task_struct *tsk,
10619 enum cpuacct_stat_index idx, cputime_t val)
10621 struct cpuacct *ca;
10623 if (unlikely(!cpuacct_subsys.active))
10630 percpu_counter_add(&ca->cpustat[idx], val);
10636 struct cgroup_subsys cpuacct_subsys = {
10638 .create = cpuacct_create,
10639 .destroy = cpuacct_destroy,
10640 .populate = cpuacct_populate,
10641 .subsys_id = cpuacct_subsys_id,
10643 #endif /* CONFIG_CGROUP_CPUACCT */