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 int finish_task_switch(struct rq *rq, struct task_struct *prev)
2843 __releases(rq->lock)
2845 struct mm_struct *mm = rq->prev_mm;
2847 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);
2872 fire_sched_in_preempt_notifiers(current);
2875 if (unlikely(prev_state == TASK_DEAD)) {
2877 * Remove function-return probe instances associated with this
2878 * task and put them back on the free list.
2880 kprobe_flush_task(prev);
2881 put_task_struct(prev);
2884 return post_schedule;
2888 * schedule_tail - first thing a freshly forked thread must call.
2889 * @prev: the thread we just switched away from.
2891 asmlinkage void schedule_tail(struct task_struct *prev)
2892 __releases(rq->lock)
2894 struct rq *rq = this_rq();
2897 post_schedule = finish_task_switch(rq, prev);
2901 current->sched_class->post_schedule(rq);
2904 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2905 /* In this case, finish_task_switch does not reenable preemption */
2908 if (current->set_child_tid)
2909 put_user(task_pid_vnr(current), current->set_child_tid);
2913 * context_switch - switch to the new MM and the new
2914 * thread's register state.
2917 context_switch(struct rq *rq, struct task_struct *prev,
2918 struct task_struct *next)
2920 struct mm_struct *mm, *oldmm;
2922 prepare_task_switch(rq, prev, next);
2923 trace_sched_switch(rq, prev, next);
2925 oldmm = prev->active_mm;
2927 * For paravirt, this is coupled with an exit in switch_to to
2928 * combine the page table reload and the switch backend into
2931 arch_start_context_switch(prev);
2933 if (unlikely(!mm)) {
2934 next->active_mm = oldmm;
2935 atomic_inc(&oldmm->mm_count);
2936 enter_lazy_tlb(oldmm, next);
2938 switch_mm(oldmm, mm, next);
2940 if (unlikely(!prev->mm)) {
2941 prev->active_mm = NULL;
2942 rq->prev_mm = oldmm;
2945 * Since the runqueue lock will be released by the next
2946 * task (which is an invalid locking op but in the case
2947 * of the scheduler it's an obvious special-case), so we
2948 * do an early lockdep release here:
2950 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2951 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2954 /* Here we just switch the register state and the stack. */
2955 switch_to(prev, next, prev);
2959 * this_rq must be evaluated again because prev may have moved
2960 * CPUs since it called schedule(), thus the 'rq' on its stack
2961 * frame will be invalid.
2963 return finish_task_switch(this_rq(), prev);
2967 * nr_running, nr_uninterruptible and nr_context_switches:
2969 * externally visible scheduler statistics: current number of runnable
2970 * threads, current number of uninterruptible-sleeping threads, total
2971 * number of context switches performed since bootup.
2973 unsigned long nr_running(void)
2975 unsigned long i, sum = 0;
2977 for_each_online_cpu(i)
2978 sum += cpu_rq(i)->nr_running;
2983 unsigned long nr_uninterruptible(void)
2985 unsigned long i, sum = 0;
2987 for_each_possible_cpu(i)
2988 sum += cpu_rq(i)->nr_uninterruptible;
2991 * Since we read the counters lockless, it might be slightly
2992 * inaccurate. Do not allow it to go below zero though:
2994 if (unlikely((long)sum < 0))
3000 unsigned long long nr_context_switches(void)
3003 unsigned long long sum = 0;
3005 for_each_possible_cpu(i)
3006 sum += cpu_rq(i)->nr_switches;
3011 unsigned long nr_iowait(void)
3013 unsigned long i, sum = 0;
3015 for_each_possible_cpu(i)
3016 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3021 /* Variables and functions for calc_load */
3022 static atomic_long_t calc_load_tasks;
3023 static unsigned long calc_load_update;
3024 unsigned long avenrun[3];
3025 EXPORT_SYMBOL(avenrun);
3028 * get_avenrun - get the load average array
3029 * @loads: pointer to dest load array
3030 * @offset: offset to add
3031 * @shift: shift count to shift the result left
3033 * These values are estimates at best, so no need for locking.
3035 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3037 loads[0] = (avenrun[0] + offset) << shift;
3038 loads[1] = (avenrun[1] + offset) << shift;
3039 loads[2] = (avenrun[2] + offset) << shift;
3042 static unsigned long
3043 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3046 load += active * (FIXED_1 - exp);
3047 return load >> FSHIFT;
3051 * calc_load - update the avenrun load estimates 10 ticks after the
3052 * CPUs have updated calc_load_tasks.
3054 void calc_global_load(void)
3056 unsigned long upd = calc_load_update + 10;
3059 if (time_before(jiffies, upd))
3062 active = atomic_long_read(&calc_load_tasks);
3063 active = active > 0 ? active * FIXED_1 : 0;
3065 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3066 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3067 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3069 calc_load_update += LOAD_FREQ;
3073 * Either called from update_cpu_load() or from a cpu going idle
3075 static void calc_load_account_active(struct rq *this_rq)
3077 long nr_active, delta;
3079 nr_active = this_rq->nr_running;
3080 nr_active += (long) this_rq->nr_uninterruptible;
3082 if (nr_active != this_rq->calc_load_active) {
3083 delta = nr_active - this_rq->calc_load_active;
3084 this_rq->calc_load_active = nr_active;
3085 atomic_long_add(delta, &calc_load_tasks);
3090 * Externally visible per-cpu scheduler statistics:
3091 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3093 u64 cpu_nr_migrations(int cpu)
3095 return cpu_rq(cpu)->nr_migrations_in;
3099 * Update rq->cpu_load[] statistics. This function is usually called every
3100 * scheduler tick (TICK_NSEC).
3102 static void update_cpu_load(struct rq *this_rq)
3104 unsigned long this_load = this_rq->load.weight;
3107 this_rq->nr_load_updates++;
3109 /* Update our load: */
3110 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3111 unsigned long old_load, new_load;
3113 /* scale is effectively 1 << i now, and >> i divides by scale */
3115 old_load = this_rq->cpu_load[i];
3116 new_load = this_load;
3118 * Round up the averaging division if load is increasing. This
3119 * prevents us from getting stuck on 9 if the load is 10, for
3122 if (new_load > old_load)
3123 new_load += scale-1;
3124 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3127 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3128 this_rq->calc_load_update += LOAD_FREQ;
3129 calc_load_account_active(this_rq);
3136 * double_rq_lock - safely lock two runqueues
3138 * Note this does not disable interrupts like task_rq_lock,
3139 * you need to do so manually before calling.
3141 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3142 __acquires(rq1->lock)
3143 __acquires(rq2->lock)
3145 BUG_ON(!irqs_disabled());
3147 spin_lock(&rq1->lock);
3148 __acquire(rq2->lock); /* Fake it out ;) */
3151 spin_lock(&rq1->lock);
3152 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3154 spin_lock(&rq2->lock);
3155 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3158 update_rq_clock(rq1);
3159 update_rq_clock(rq2);
3163 * double_rq_unlock - safely unlock two runqueues
3165 * Note this does not restore interrupts like task_rq_unlock,
3166 * you need to do so manually after calling.
3168 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3169 __releases(rq1->lock)
3170 __releases(rq2->lock)
3172 spin_unlock(&rq1->lock);
3174 spin_unlock(&rq2->lock);
3176 __release(rq2->lock);
3180 * If dest_cpu is allowed for this process, migrate the task to it.
3181 * This is accomplished by forcing the cpu_allowed mask to only
3182 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3183 * the cpu_allowed mask is restored.
3185 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3187 struct migration_req req;
3188 unsigned long flags;
3191 rq = task_rq_lock(p, &flags);
3192 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3193 || unlikely(!cpu_active(dest_cpu)))
3196 /* force the process onto the specified CPU */
3197 if (migrate_task(p, dest_cpu, &req)) {
3198 /* Need to wait for migration thread (might exit: take ref). */
3199 struct task_struct *mt = rq->migration_thread;
3201 get_task_struct(mt);
3202 task_rq_unlock(rq, &flags);
3203 wake_up_process(mt);
3204 put_task_struct(mt);
3205 wait_for_completion(&req.done);
3210 task_rq_unlock(rq, &flags);
3214 * sched_exec - execve() is a valuable balancing opportunity, because at
3215 * this point the task has the smallest effective memory and cache footprint.
3217 void sched_exec(void)
3219 int new_cpu, this_cpu = get_cpu();
3220 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3222 if (new_cpu != this_cpu)
3223 sched_migrate_task(current, new_cpu);
3227 * pull_task - move a task from a remote runqueue to the local runqueue.
3228 * Both runqueues must be locked.
3230 static void pull_task(struct rq *src_rq, struct task_struct *p,
3231 struct rq *this_rq, int this_cpu)
3233 deactivate_task(src_rq, p, 0);
3234 set_task_cpu(p, this_cpu);
3235 activate_task(this_rq, p, 0);
3237 * Note that idle threads have a prio of MAX_PRIO, for this test
3238 * to be always true for them.
3240 check_preempt_curr(this_rq, p, 0);
3244 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3247 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3248 struct sched_domain *sd, enum cpu_idle_type idle,
3251 int tsk_cache_hot = 0;
3253 * We do not migrate tasks that are:
3254 * 1) running (obviously), or
3255 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3256 * 3) are cache-hot on their current CPU.
3258 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3259 schedstat_inc(p, se.nr_failed_migrations_affine);
3264 if (task_running(rq, p)) {
3265 schedstat_inc(p, se.nr_failed_migrations_running);
3270 * Aggressive migration if:
3271 * 1) task is cache cold, or
3272 * 2) too many balance attempts have failed.
3275 tsk_cache_hot = task_hot(p, rq->clock, sd);
3276 if (!tsk_cache_hot ||
3277 sd->nr_balance_failed > sd->cache_nice_tries) {
3278 #ifdef CONFIG_SCHEDSTATS
3279 if (tsk_cache_hot) {
3280 schedstat_inc(sd, lb_hot_gained[idle]);
3281 schedstat_inc(p, se.nr_forced_migrations);
3287 if (tsk_cache_hot) {
3288 schedstat_inc(p, se.nr_failed_migrations_hot);
3294 static unsigned long
3295 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3296 unsigned long max_load_move, struct sched_domain *sd,
3297 enum cpu_idle_type idle, int *all_pinned,
3298 int *this_best_prio, struct rq_iterator *iterator)
3300 int loops = 0, pulled = 0, pinned = 0;
3301 struct task_struct *p;
3302 long rem_load_move = max_load_move;
3304 if (max_load_move == 0)
3310 * Start the load-balancing iterator:
3312 p = iterator->start(iterator->arg);
3314 if (!p || loops++ > sysctl_sched_nr_migrate)
3317 if ((p->se.load.weight >> 1) > rem_load_move ||
3318 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3319 p = iterator->next(iterator->arg);
3323 pull_task(busiest, p, this_rq, this_cpu);
3325 rem_load_move -= p->se.load.weight;
3327 #ifdef CONFIG_PREEMPT
3329 * NEWIDLE balancing is a source of latency, so preemptible kernels
3330 * will stop after the first task is pulled to minimize the critical
3333 if (idle == CPU_NEWLY_IDLE)
3338 * We only want to steal up to the prescribed amount of weighted load.
3340 if (rem_load_move > 0) {
3341 if (p->prio < *this_best_prio)
3342 *this_best_prio = p->prio;
3343 p = iterator->next(iterator->arg);
3348 * Right now, this is one of only two places pull_task() is called,
3349 * so we can safely collect pull_task() stats here rather than
3350 * inside pull_task().
3352 schedstat_add(sd, lb_gained[idle], pulled);
3355 *all_pinned = pinned;
3357 return max_load_move - rem_load_move;
3361 * move_tasks tries to move up to max_load_move weighted load from busiest to
3362 * this_rq, as part of a balancing operation within domain "sd".
3363 * Returns 1 if successful and 0 otherwise.
3365 * Called with both runqueues locked.
3367 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3368 unsigned long max_load_move,
3369 struct sched_domain *sd, enum cpu_idle_type idle,
3372 const struct sched_class *class = sched_class_highest;
3373 unsigned long total_load_moved = 0;
3374 int this_best_prio = this_rq->curr->prio;
3378 class->load_balance(this_rq, this_cpu, busiest,
3379 max_load_move - total_load_moved,
3380 sd, idle, all_pinned, &this_best_prio);
3381 class = class->next;
3383 #ifdef CONFIG_PREEMPT
3385 * NEWIDLE balancing is a source of latency, so preemptible
3386 * kernels will stop after the first task is pulled to minimize
3387 * the critical section.
3389 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3392 } while (class && max_load_move > total_load_moved);
3394 return total_load_moved > 0;
3398 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3399 struct sched_domain *sd, enum cpu_idle_type idle,
3400 struct rq_iterator *iterator)
3402 struct task_struct *p = iterator->start(iterator->arg);
3406 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3407 pull_task(busiest, p, this_rq, this_cpu);
3409 * Right now, this is only the second place pull_task()
3410 * is called, so we can safely collect pull_task()
3411 * stats here rather than inside pull_task().
3413 schedstat_inc(sd, lb_gained[idle]);
3417 p = iterator->next(iterator->arg);
3424 * move_one_task tries to move exactly one task from busiest to this_rq, as
3425 * part of active balancing operations within "domain".
3426 * Returns 1 if successful and 0 otherwise.
3428 * Called with both runqueues locked.
3430 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3431 struct sched_domain *sd, enum cpu_idle_type idle)
3433 const struct sched_class *class;
3435 for (class = sched_class_highest; class; class = class->next)
3436 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3441 /********** Helpers for find_busiest_group ************************/
3443 * sd_lb_stats - Structure to store the statistics of a sched_domain
3444 * during load balancing.
3446 struct sd_lb_stats {
3447 struct sched_group *busiest; /* Busiest group in this sd */
3448 struct sched_group *this; /* Local group in this sd */
3449 unsigned long total_load; /* Total load of all groups in sd */
3450 unsigned long total_pwr; /* Total power of all groups in sd */
3451 unsigned long avg_load; /* Average load across all groups in sd */
3453 /** Statistics of this group */
3454 unsigned long this_load;
3455 unsigned long this_load_per_task;
3456 unsigned long this_nr_running;
3458 /* Statistics of the busiest group */
3459 unsigned long max_load;
3460 unsigned long busiest_load_per_task;
3461 unsigned long busiest_nr_running;
3463 int group_imb; /* Is there imbalance in this sd */
3464 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3465 int power_savings_balance; /* Is powersave balance needed for this sd */
3466 struct sched_group *group_min; /* Least loaded group in sd */
3467 struct sched_group *group_leader; /* Group which relieves group_min */
3468 unsigned long min_load_per_task; /* load_per_task in group_min */
3469 unsigned long leader_nr_running; /* Nr running of group_leader */
3470 unsigned long min_nr_running; /* Nr running of group_min */
3475 * sg_lb_stats - stats of a sched_group required for load_balancing
3477 struct sg_lb_stats {
3478 unsigned long avg_load; /*Avg load across the CPUs of the group */
3479 unsigned long group_load; /* Total load over the CPUs of the group */
3480 unsigned long sum_nr_running; /* Nr tasks running in the group */
3481 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3482 unsigned long group_capacity;
3483 int group_imb; /* Is there an imbalance in the group ? */
3487 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3488 * @group: The group whose first cpu is to be returned.
3490 static inline unsigned int group_first_cpu(struct sched_group *group)
3492 return cpumask_first(sched_group_cpus(group));
3496 * get_sd_load_idx - Obtain the load index for a given sched domain.
3497 * @sd: The sched_domain whose load_idx is to be obtained.
3498 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3500 static inline int get_sd_load_idx(struct sched_domain *sd,
3501 enum cpu_idle_type idle)
3507 load_idx = sd->busy_idx;
3510 case CPU_NEWLY_IDLE:
3511 load_idx = sd->newidle_idx;
3514 load_idx = sd->idle_idx;
3522 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3524 * init_sd_power_savings_stats - Initialize power savings statistics for
3525 * the given sched_domain, during load balancing.
3527 * @sd: Sched domain whose power-savings statistics are to be initialized.
3528 * @sds: Variable containing the statistics for sd.
3529 * @idle: Idle status of the CPU at which we're performing load-balancing.
3531 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3532 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3535 * Busy processors will not participate in power savings
3538 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3539 sds->power_savings_balance = 0;
3541 sds->power_savings_balance = 1;
3542 sds->min_nr_running = ULONG_MAX;
3543 sds->leader_nr_running = 0;
3548 * update_sd_power_savings_stats - Update the power saving stats for a
3549 * sched_domain while performing load balancing.
3551 * @group: sched_group belonging to the sched_domain under consideration.
3552 * @sds: Variable containing the statistics of the sched_domain
3553 * @local_group: Does group contain the CPU for which we're performing
3555 * @sgs: Variable containing the statistics of the group.
3557 static inline void update_sd_power_savings_stats(struct sched_group *group,
3558 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3561 if (!sds->power_savings_balance)
3565 * If the local group is idle or completely loaded
3566 * no need to do power savings balance at this domain
3568 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3569 !sds->this_nr_running))
3570 sds->power_savings_balance = 0;
3573 * If a group is already running at full capacity or idle,
3574 * don't include that group in power savings calculations
3576 if (!sds->power_savings_balance ||
3577 sgs->sum_nr_running >= sgs->group_capacity ||
3578 !sgs->sum_nr_running)
3582 * Calculate the group which has the least non-idle load.
3583 * This is the group from where we need to pick up the load
3586 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3587 (sgs->sum_nr_running == sds->min_nr_running &&
3588 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3589 sds->group_min = group;
3590 sds->min_nr_running = sgs->sum_nr_running;
3591 sds->min_load_per_task = sgs->sum_weighted_load /
3592 sgs->sum_nr_running;
3596 * Calculate the group which is almost near its
3597 * capacity but still has some space to pick up some load
3598 * from other group and save more power
3600 if (sgs->sum_nr_running > sgs->group_capacity - 1)
3603 if (sgs->sum_nr_running > sds->leader_nr_running ||
3604 (sgs->sum_nr_running == sds->leader_nr_running &&
3605 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3606 sds->group_leader = group;
3607 sds->leader_nr_running = sgs->sum_nr_running;
3612 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3613 * @sds: Variable containing the statistics of the sched_domain
3614 * under consideration.
3615 * @this_cpu: Cpu at which we're currently performing load-balancing.
3616 * @imbalance: Variable to store the imbalance.
3619 * Check if we have potential to perform some power-savings balance.
3620 * If yes, set the busiest group to be the least loaded group in the
3621 * sched_domain, so that it's CPUs can be put to idle.
3623 * Returns 1 if there is potential to perform power-savings balance.
3626 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3627 int this_cpu, unsigned long *imbalance)
3629 if (!sds->power_savings_balance)
3632 if (sds->this != sds->group_leader ||
3633 sds->group_leader == sds->group_min)
3636 *imbalance = sds->min_load_per_task;
3637 sds->busiest = sds->group_min;
3639 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3640 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3641 group_first_cpu(sds->group_leader);
3647 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3648 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3649 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3654 static inline void update_sd_power_savings_stats(struct sched_group *group,
3655 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3660 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3661 int this_cpu, unsigned long *imbalance)
3665 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3669 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3670 * @group: sched_group whose statistics are to be updated.
3671 * @this_cpu: Cpu for which load balance is currently performed.
3672 * @idle: Idle status of this_cpu
3673 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3674 * @sd_idle: Idle status of the sched_domain containing group.
3675 * @local_group: Does group contain this_cpu.
3676 * @cpus: Set of cpus considered for load balancing.
3677 * @balance: Should we balance.
3678 * @sgs: variable to hold the statistics for this group.
3680 static inline void update_sg_lb_stats(struct sched_group *group, int this_cpu,
3681 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3682 int local_group, const struct cpumask *cpus,
3683 int *balance, struct sg_lb_stats *sgs)
3685 unsigned long load, max_cpu_load, min_cpu_load;
3687 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3688 unsigned long sum_avg_load_per_task;
3689 unsigned long avg_load_per_task;
3692 balance_cpu = group_first_cpu(group);
3694 /* Tally up the load of all CPUs in the group */
3695 sum_avg_load_per_task = avg_load_per_task = 0;
3697 min_cpu_load = ~0UL;
3699 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3700 struct rq *rq = cpu_rq(i);
3702 if (*sd_idle && rq->nr_running)
3705 /* Bias balancing toward cpus of our domain */
3707 if (idle_cpu(i) && !first_idle_cpu) {
3712 load = target_load(i, load_idx);
3714 load = source_load(i, load_idx);
3715 if (load > max_cpu_load)
3716 max_cpu_load = load;
3717 if (min_cpu_load > load)
3718 min_cpu_load = load;
3721 sgs->group_load += load;
3722 sgs->sum_nr_running += rq->nr_running;
3723 sgs->sum_weighted_load += weighted_cpuload(i);
3725 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3729 * First idle cpu or the first cpu(busiest) in this sched group
3730 * is eligible for doing load balancing at this and above
3731 * domains. In the newly idle case, we will allow all the cpu's
3732 * to do the newly idle load balance.
3734 if (idle != CPU_NEWLY_IDLE && local_group &&
3735 balance_cpu != this_cpu && balance) {
3740 /* Adjust by relative CPU power of the group */
3741 sgs->avg_load = sg_div_cpu_power(group,
3742 sgs->group_load * SCHED_LOAD_SCALE);
3746 * Consider the group unbalanced when the imbalance is larger
3747 * than the average weight of two tasks.
3749 * APZ: with cgroup the avg task weight can vary wildly and
3750 * might not be a suitable number - should we keep a
3751 * normalized nr_running number somewhere that negates
3754 avg_load_per_task = sg_div_cpu_power(group,
3755 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3757 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3760 sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3765 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3766 * @sd: sched_domain whose statistics are to be updated.
3767 * @this_cpu: Cpu for which load balance is currently performed.
3768 * @idle: Idle status of this_cpu
3769 * @sd_idle: Idle status of the sched_domain containing group.
3770 * @cpus: Set of cpus considered for load balancing.
3771 * @balance: Should we balance.
3772 * @sds: variable to hold the statistics for this sched_domain.
3774 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3775 enum cpu_idle_type idle, int *sd_idle,
3776 const struct cpumask *cpus, int *balance,
3777 struct sd_lb_stats *sds)
3779 struct sched_group *group = sd->groups;
3780 struct sg_lb_stats sgs;
3783 init_sd_power_savings_stats(sd, sds, idle);
3784 load_idx = get_sd_load_idx(sd, idle);
3789 local_group = cpumask_test_cpu(this_cpu,
3790 sched_group_cpus(group));
3791 memset(&sgs, 0, sizeof(sgs));
3792 update_sg_lb_stats(group, this_cpu, idle, load_idx, sd_idle,
3793 local_group, cpus, balance, &sgs);
3795 if (local_group && balance && !(*balance))
3798 sds->total_load += sgs.group_load;
3799 sds->total_pwr += group->__cpu_power;
3802 sds->this_load = sgs.avg_load;
3804 sds->this_nr_running = sgs.sum_nr_running;
3805 sds->this_load_per_task = sgs.sum_weighted_load;
3806 } else if (sgs.avg_load > sds->max_load &&
3807 (sgs.sum_nr_running > sgs.group_capacity ||
3809 sds->max_load = sgs.avg_load;
3810 sds->busiest = group;
3811 sds->busiest_nr_running = sgs.sum_nr_running;
3812 sds->busiest_load_per_task = sgs.sum_weighted_load;
3813 sds->group_imb = sgs.group_imb;
3816 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3817 group = group->next;
3818 } while (group != sd->groups);
3823 * fix_small_imbalance - Calculate the minor imbalance that exists
3824 * amongst the groups of a sched_domain, during
3826 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3827 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3828 * @imbalance: Variable to store the imbalance.
3830 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3831 int this_cpu, unsigned long *imbalance)
3833 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3834 unsigned int imbn = 2;
3836 if (sds->this_nr_running) {
3837 sds->this_load_per_task /= sds->this_nr_running;
3838 if (sds->busiest_load_per_task >
3839 sds->this_load_per_task)
3842 sds->this_load_per_task =
3843 cpu_avg_load_per_task(this_cpu);
3845 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3846 sds->busiest_load_per_task * imbn) {
3847 *imbalance = sds->busiest_load_per_task;
3852 * OK, we don't have enough imbalance to justify moving tasks,
3853 * however we may be able to increase total CPU power used by
3857 pwr_now += sds->busiest->__cpu_power *
3858 min(sds->busiest_load_per_task, sds->max_load);
3859 pwr_now += sds->this->__cpu_power *
3860 min(sds->this_load_per_task, sds->this_load);
3861 pwr_now /= SCHED_LOAD_SCALE;
3863 /* Amount of load we'd subtract */
3864 tmp = sg_div_cpu_power(sds->busiest,
3865 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3866 if (sds->max_load > tmp)
3867 pwr_move += sds->busiest->__cpu_power *
3868 min(sds->busiest_load_per_task, sds->max_load - tmp);
3870 /* Amount of load we'd add */
3871 if (sds->max_load * sds->busiest->__cpu_power <
3872 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3873 tmp = sg_div_cpu_power(sds->this,
3874 sds->max_load * sds->busiest->__cpu_power);
3876 tmp = sg_div_cpu_power(sds->this,
3877 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3878 pwr_move += sds->this->__cpu_power *
3879 min(sds->this_load_per_task, sds->this_load + tmp);
3880 pwr_move /= SCHED_LOAD_SCALE;
3882 /* Move if we gain throughput */
3883 if (pwr_move > pwr_now)
3884 *imbalance = sds->busiest_load_per_task;
3888 * calculate_imbalance - Calculate the amount of imbalance present within the
3889 * groups of a given sched_domain during load balance.
3890 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3891 * @this_cpu: Cpu for which currently load balance is being performed.
3892 * @imbalance: The variable to store the imbalance.
3894 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3895 unsigned long *imbalance)
3897 unsigned long max_pull;
3899 * In the presence of smp nice balancing, certain scenarios can have
3900 * max load less than avg load(as we skip the groups at or below
3901 * its cpu_power, while calculating max_load..)
3903 if (sds->max_load < sds->avg_load) {
3905 return fix_small_imbalance(sds, this_cpu, imbalance);
3908 /* Don't want to pull so many tasks that a group would go idle */
3909 max_pull = min(sds->max_load - sds->avg_load,
3910 sds->max_load - sds->busiest_load_per_task);
3912 /* How much load to actually move to equalise the imbalance */
3913 *imbalance = min(max_pull * sds->busiest->__cpu_power,
3914 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
3918 * if *imbalance is less than the average load per runnable task
3919 * there is no gaurantee that any tasks will be moved so we'll have
3920 * a think about bumping its value to force at least one task to be
3923 if (*imbalance < sds->busiest_load_per_task)
3924 return fix_small_imbalance(sds, this_cpu, imbalance);
3927 /******* find_busiest_group() helpers end here *********************/
3930 * find_busiest_group - Returns the busiest group within the sched_domain
3931 * if there is an imbalance. If there isn't an imbalance, and
3932 * the user has opted for power-savings, it returns a group whose
3933 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3934 * such a group exists.
3936 * Also calculates the amount of weighted load which should be moved
3937 * to restore balance.
3939 * @sd: The sched_domain whose busiest group is to be returned.
3940 * @this_cpu: The cpu for which load balancing is currently being performed.
3941 * @imbalance: Variable which stores amount of weighted load which should
3942 * be moved to restore balance/put a group to idle.
3943 * @idle: The idle status of this_cpu.
3944 * @sd_idle: The idleness of sd
3945 * @cpus: The set of CPUs under consideration for load-balancing.
3946 * @balance: Pointer to a variable indicating if this_cpu
3947 * is the appropriate cpu to perform load balancing at this_level.
3949 * Returns: - the busiest group if imbalance exists.
3950 * - If no imbalance and user has opted for power-savings balance,
3951 * return the least loaded group whose CPUs can be
3952 * put to idle by rebalancing its tasks onto our group.
3954 static struct sched_group *
3955 find_busiest_group(struct sched_domain *sd, int this_cpu,
3956 unsigned long *imbalance, enum cpu_idle_type idle,
3957 int *sd_idle, const struct cpumask *cpus, int *balance)
3959 struct sd_lb_stats sds;
3961 memset(&sds, 0, sizeof(sds));
3964 * Compute the various statistics relavent for load balancing at
3967 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3970 /* Cases where imbalance does not exist from POV of this_cpu */
3971 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3973 * 2) There is no busy sibling group to pull from.
3974 * 3) This group is the busiest group.
3975 * 4) This group is more busy than the avg busieness at this
3977 * 5) The imbalance is within the specified limit.
3978 * 6) Any rebalance would lead to ping-pong
3980 if (balance && !(*balance))
3983 if (!sds.busiest || sds.busiest_nr_running == 0)
3986 if (sds.this_load >= sds.max_load)
3989 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3991 if (sds.this_load >= sds.avg_load)
3994 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3997 sds.busiest_load_per_task /= sds.busiest_nr_running;
3999 sds.busiest_load_per_task =
4000 min(sds.busiest_load_per_task, sds.avg_load);
4003 * We're trying to get all the cpus to the average_load, so we don't
4004 * want to push ourselves above the average load, nor do we wish to
4005 * reduce the max loaded cpu below the average load, as either of these
4006 * actions would just result in more rebalancing later, and ping-pong
4007 * tasks around. Thus we look for the minimum possible imbalance.
4008 * Negative imbalances (*we* are more loaded than anyone else) will
4009 * be counted as no imbalance for these purposes -- we can't fix that
4010 * by pulling tasks to us. Be careful of negative numbers as they'll
4011 * appear as very large values with unsigned longs.
4013 if (sds.max_load <= sds.busiest_load_per_task)
4016 /* Looks like there is an imbalance. Compute it */
4017 calculate_imbalance(&sds, this_cpu, imbalance);
4022 * There is no obvious imbalance. But check if we can do some balancing
4025 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4033 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4036 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4037 unsigned long imbalance, const struct cpumask *cpus)
4039 struct rq *busiest = NULL, *rq;
4040 unsigned long max_load = 0;
4043 for_each_cpu(i, sched_group_cpus(group)) {
4046 if (!cpumask_test_cpu(i, cpus))
4050 wl = weighted_cpuload(i);
4052 if (rq->nr_running == 1 && wl > imbalance)
4055 if (wl > max_load) {
4065 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4066 * so long as it is large enough.
4068 #define MAX_PINNED_INTERVAL 512
4070 /* Working cpumask for load_balance and load_balance_newidle. */
4071 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4074 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4075 * tasks if there is an imbalance.
4077 static int load_balance(int this_cpu, struct rq *this_rq,
4078 struct sched_domain *sd, enum cpu_idle_type idle,
4081 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4082 struct sched_group *group;
4083 unsigned long imbalance;
4085 unsigned long flags;
4086 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4088 cpumask_setall(cpus);
4091 * When power savings policy is enabled for the parent domain, idle
4092 * sibling can pick up load irrespective of busy siblings. In this case,
4093 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4094 * portraying it as CPU_NOT_IDLE.
4096 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4097 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4100 schedstat_inc(sd, lb_count[idle]);
4104 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4111 schedstat_inc(sd, lb_nobusyg[idle]);
4115 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4117 schedstat_inc(sd, lb_nobusyq[idle]);
4121 BUG_ON(busiest == this_rq);
4123 schedstat_add(sd, lb_imbalance[idle], imbalance);
4126 if (busiest->nr_running > 1) {
4128 * Attempt to move tasks. If find_busiest_group has found
4129 * an imbalance but busiest->nr_running <= 1, the group is
4130 * still unbalanced. ld_moved simply stays zero, so it is
4131 * correctly treated as an imbalance.
4133 local_irq_save(flags);
4134 double_rq_lock(this_rq, busiest);
4135 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4136 imbalance, sd, idle, &all_pinned);
4137 double_rq_unlock(this_rq, busiest);
4138 local_irq_restore(flags);
4141 * some other cpu did the load balance for us.
4143 if (ld_moved && this_cpu != smp_processor_id())
4144 resched_cpu(this_cpu);
4146 /* All tasks on this runqueue were pinned by CPU affinity */
4147 if (unlikely(all_pinned)) {
4148 cpumask_clear_cpu(cpu_of(busiest), cpus);
4149 if (!cpumask_empty(cpus))
4156 schedstat_inc(sd, lb_failed[idle]);
4157 sd->nr_balance_failed++;
4159 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4161 spin_lock_irqsave(&busiest->lock, flags);
4163 /* don't kick the migration_thread, if the curr
4164 * task on busiest cpu can't be moved to this_cpu
4166 if (!cpumask_test_cpu(this_cpu,
4167 &busiest->curr->cpus_allowed)) {
4168 spin_unlock_irqrestore(&busiest->lock, flags);
4170 goto out_one_pinned;
4173 if (!busiest->active_balance) {
4174 busiest->active_balance = 1;
4175 busiest->push_cpu = this_cpu;
4178 spin_unlock_irqrestore(&busiest->lock, flags);
4180 wake_up_process(busiest->migration_thread);
4183 * We've kicked active balancing, reset the failure
4186 sd->nr_balance_failed = sd->cache_nice_tries+1;
4189 sd->nr_balance_failed = 0;
4191 if (likely(!active_balance)) {
4192 /* We were unbalanced, so reset the balancing interval */
4193 sd->balance_interval = sd->min_interval;
4196 * If we've begun active balancing, start to back off. This
4197 * case may not be covered by the all_pinned logic if there
4198 * is only 1 task on the busy runqueue (because we don't call
4201 if (sd->balance_interval < sd->max_interval)
4202 sd->balance_interval *= 2;
4205 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4206 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4212 schedstat_inc(sd, lb_balanced[idle]);
4214 sd->nr_balance_failed = 0;
4217 /* tune up the balancing interval */
4218 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4219 (sd->balance_interval < sd->max_interval))
4220 sd->balance_interval *= 2;
4222 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4223 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4234 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4235 * tasks if there is an imbalance.
4237 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4238 * this_rq is locked.
4241 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4243 struct sched_group *group;
4244 struct rq *busiest = NULL;
4245 unsigned long imbalance;
4249 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4251 cpumask_setall(cpus);
4254 * When power savings policy is enabled for the parent domain, idle
4255 * sibling can pick up load irrespective of busy siblings. In this case,
4256 * let the state of idle sibling percolate up as IDLE, instead of
4257 * portraying it as CPU_NOT_IDLE.
4259 if (sd->flags & SD_SHARE_CPUPOWER &&
4260 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4263 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4265 update_shares_locked(this_rq, sd);
4266 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4267 &sd_idle, cpus, NULL);
4269 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4273 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4275 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4279 BUG_ON(busiest == this_rq);
4281 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4284 if (busiest->nr_running > 1) {
4285 /* Attempt to move tasks */
4286 double_lock_balance(this_rq, busiest);
4287 /* this_rq->clock is already updated */
4288 update_rq_clock(busiest);
4289 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4290 imbalance, sd, CPU_NEWLY_IDLE,
4292 double_unlock_balance(this_rq, busiest);
4294 if (unlikely(all_pinned)) {
4295 cpumask_clear_cpu(cpu_of(busiest), cpus);
4296 if (!cpumask_empty(cpus))
4302 int active_balance = 0;
4304 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4305 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4306 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4309 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4312 if (sd->nr_balance_failed++ < 2)
4316 * The only task running in a non-idle cpu can be moved to this
4317 * cpu in an attempt to completely freeup the other CPU
4318 * package. The same method used to move task in load_balance()
4319 * have been extended for load_balance_newidle() to speedup
4320 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4322 * The package power saving logic comes from
4323 * find_busiest_group(). If there are no imbalance, then
4324 * f_b_g() will return NULL. However when sched_mc={1,2} then
4325 * f_b_g() will select a group from which a running task may be
4326 * pulled to this cpu in order to make the other package idle.
4327 * If there is no opportunity to make a package idle and if
4328 * there are no imbalance, then f_b_g() will return NULL and no
4329 * action will be taken in load_balance_newidle().
4331 * Under normal task pull operation due to imbalance, there
4332 * will be more than one task in the source run queue and
4333 * move_tasks() will succeed. ld_moved will be true and this
4334 * active balance code will not be triggered.
4337 /* Lock busiest in correct order while this_rq is held */
4338 double_lock_balance(this_rq, busiest);
4341 * don't kick the migration_thread, if the curr
4342 * task on busiest cpu can't be moved to this_cpu
4344 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4345 double_unlock_balance(this_rq, busiest);
4350 if (!busiest->active_balance) {
4351 busiest->active_balance = 1;
4352 busiest->push_cpu = this_cpu;
4356 double_unlock_balance(this_rq, busiest);
4358 * Should not call ttwu while holding a rq->lock
4360 spin_unlock(&this_rq->lock);
4362 wake_up_process(busiest->migration_thread);
4363 spin_lock(&this_rq->lock);
4366 sd->nr_balance_failed = 0;
4368 update_shares_locked(this_rq, sd);
4372 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4373 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4374 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4376 sd->nr_balance_failed = 0;
4382 * idle_balance is called by schedule() if this_cpu is about to become
4383 * idle. Attempts to pull tasks from other CPUs.
4385 static void idle_balance(int this_cpu, struct rq *this_rq)
4387 struct sched_domain *sd;
4388 int pulled_task = 0;
4389 unsigned long next_balance = jiffies + HZ;
4391 for_each_domain(this_cpu, sd) {
4392 unsigned long interval;
4394 if (!(sd->flags & SD_LOAD_BALANCE))
4397 if (sd->flags & SD_BALANCE_NEWIDLE)
4398 /* If we've pulled tasks over stop searching: */
4399 pulled_task = load_balance_newidle(this_cpu, this_rq,
4402 interval = msecs_to_jiffies(sd->balance_interval);
4403 if (time_after(next_balance, sd->last_balance + interval))
4404 next_balance = sd->last_balance + interval;
4408 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4410 * We are going idle. next_balance may be set based on
4411 * a busy processor. So reset next_balance.
4413 this_rq->next_balance = next_balance;
4418 * active_load_balance is run by migration threads. It pushes running tasks
4419 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4420 * running on each physical CPU where possible, and avoids physical /
4421 * logical imbalances.
4423 * Called with busiest_rq locked.
4425 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4427 int target_cpu = busiest_rq->push_cpu;
4428 struct sched_domain *sd;
4429 struct rq *target_rq;
4431 /* Is there any task to move? */
4432 if (busiest_rq->nr_running <= 1)
4435 target_rq = cpu_rq(target_cpu);
4438 * This condition is "impossible", if it occurs
4439 * we need to fix it. Originally reported by
4440 * Bjorn Helgaas on a 128-cpu setup.
4442 BUG_ON(busiest_rq == target_rq);
4444 /* move a task from busiest_rq to target_rq */
4445 double_lock_balance(busiest_rq, target_rq);
4446 update_rq_clock(busiest_rq);
4447 update_rq_clock(target_rq);
4449 /* Search for an sd spanning us and the target CPU. */
4450 for_each_domain(target_cpu, sd) {
4451 if ((sd->flags & SD_LOAD_BALANCE) &&
4452 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4457 schedstat_inc(sd, alb_count);
4459 if (move_one_task(target_rq, target_cpu, busiest_rq,
4461 schedstat_inc(sd, alb_pushed);
4463 schedstat_inc(sd, alb_failed);
4465 double_unlock_balance(busiest_rq, target_rq);
4470 atomic_t load_balancer;
4471 cpumask_var_t cpu_mask;
4472 cpumask_var_t ilb_grp_nohz_mask;
4473 } nohz ____cacheline_aligned = {
4474 .load_balancer = ATOMIC_INIT(-1),
4477 int get_nohz_load_balancer(void)
4479 return atomic_read(&nohz.load_balancer);
4482 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4484 * lowest_flag_domain - Return lowest sched_domain containing flag.
4485 * @cpu: The cpu whose lowest level of sched domain is to
4487 * @flag: The flag to check for the lowest sched_domain
4488 * for the given cpu.
4490 * Returns the lowest sched_domain of a cpu which contains the given flag.
4492 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4494 struct sched_domain *sd;
4496 for_each_domain(cpu, sd)
4497 if (sd && (sd->flags & flag))
4504 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4505 * @cpu: The cpu whose domains we're iterating over.
4506 * @sd: variable holding the value of the power_savings_sd
4508 * @flag: The flag to filter the sched_domains to be iterated.
4510 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4511 * set, starting from the lowest sched_domain to the highest.
4513 #define for_each_flag_domain(cpu, sd, flag) \
4514 for (sd = lowest_flag_domain(cpu, flag); \
4515 (sd && (sd->flags & flag)); sd = sd->parent)
4518 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4519 * @ilb_group: group to be checked for semi-idleness
4521 * Returns: 1 if the group is semi-idle. 0 otherwise.
4523 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4524 * and atleast one non-idle CPU. This helper function checks if the given
4525 * sched_group is semi-idle or not.
4527 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4529 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4530 sched_group_cpus(ilb_group));
4533 * A sched_group is semi-idle when it has atleast one busy cpu
4534 * and atleast one idle cpu.
4536 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4539 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4545 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4546 * @cpu: The cpu which is nominating a new idle_load_balancer.
4548 * Returns: Returns the id of the idle load balancer if it exists,
4549 * Else, returns >= nr_cpu_ids.
4551 * This algorithm picks the idle load balancer such that it belongs to a
4552 * semi-idle powersavings sched_domain. The idea is to try and avoid
4553 * completely idle packages/cores just for the purpose of idle load balancing
4554 * when there are other idle cpu's which are better suited for that job.
4556 static int find_new_ilb(int cpu)
4558 struct sched_domain *sd;
4559 struct sched_group *ilb_group;
4562 * Have idle load balancer selection from semi-idle packages only
4563 * when power-aware load balancing is enabled
4565 if (!(sched_smt_power_savings || sched_mc_power_savings))
4569 * Optimize for the case when we have no idle CPUs or only one
4570 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4572 if (cpumask_weight(nohz.cpu_mask) < 2)
4575 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4576 ilb_group = sd->groups;
4579 if (is_semi_idle_group(ilb_group))
4580 return cpumask_first(nohz.ilb_grp_nohz_mask);
4582 ilb_group = ilb_group->next;
4584 } while (ilb_group != sd->groups);
4588 return cpumask_first(nohz.cpu_mask);
4590 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4591 static inline int find_new_ilb(int call_cpu)
4593 return cpumask_first(nohz.cpu_mask);
4598 * This routine will try to nominate the ilb (idle load balancing)
4599 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4600 * load balancing on behalf of all those cpus. If all the cpus in the system
4601 * go into this tickless mode, then there will be no ilb owner (as there is
4602 * no need for one) and all the cpus will sleep till the next wakeup event
4605 * For the ilb owner, tick is not stopped. And this tick will be used
4606 * for idle load balancing. ilb owner will still be part of
4609 * While stopping the tick, this cpu will become the ilb owner if there
4610 * is no other owner. And will be the owner till that cpu becomes busy
4611 * or if all cpus in the system stop their ticks at which point
4612 * there is no need for ilb owner.
4614 * When the ilb owner becomes busy, it nominates another owner, during the
4615 * next busy scheduler_tick()
4617 int select_nohz_load_balancer(int stop_tick)
4619 int cpu = smp_processor_id();
4622 cpu_rq(cpu)->in_nohz_recently = 1;
4624 if (!cpu_active(cpu)) {
4625 if (atomic_read(&nohz.load_balancer) != cpu)
4629 * If we are going offline and still the leader,
4632 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4638 cpumask_set_cpu(cpu, nohz.cpu_mask);
4640 /* time for ilb owner also to sleep */
4641 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4642 if (atomic_read(&nohz.load_balancer) == cpu)
4643 atomic_set(&nohz.load_balancer, -1);
4647 if (atomic_read(&nohz.load_balancer) == -1) {
4648 /* make me the ilb owner */
4649 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4651 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4654 if (!(sched_smt_power_savings ||
4655 sched_mc_power_savings))
4658 * Check to see if there is a more power-efficient
4661 new_ilb = find_new_ilb(cpu);
4662 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4663 atomic_set(&nohz.load_balancer, -1);
4664 resched_cpu(new_ilb);
4670 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4673 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4675 if (atomic_read(&nohz.load_balancer) == cpu)
4676 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4683 static DEFINE_SPINLOCK(balancing);
4686 * It checks each scheduling domain to see if it is due to be balanced,
4687 * and initiates a balancing operation if so.
4689 * Balancing parameters are set up in arch_init_sched_domains.
4691 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4694 struct rq *rq = cpu_rq(cpu);
4695 unsigned long interval;
4696 struct sched_domain *sd;
4697 /* Earliest time when we have to do rebalance again */
4698 unsigned long next_balance = jiffies + 60*HZ;
4699 int update_next_balance = 0;
4702 for_each_domain(cpu, sd) {
4703 if (!(sd->flags & SD_LOAD_BALANCE))
4706 interval = sd->balance_interval;
4707 if (idle != CPU_IDLE)
4708 interval *= sd->busy_factor;
4710 /* scale ms to jiffies */
4711 interval = msecs_to_jiffies(interval);
4712 if (unlikely(!interval))
4714 if (interval > HZ*NR_CPUS/10)
4715 interval = HZ*NR_CPUS/10;
4717 need_serialize = sd->flags & SD_SERIALIZE;
4719 if (need_serialize) {
4720 if (!spin_trylock(&balancing))
4724 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4725 if (load_balance(cpu, rq, sd, idle, &balance)) {
4727 * We've pulled tasks over so either we're no
4728 * longer idle, or one of our SMT siblings is
4731 idle = CPU_NOT_IDLE;
4733 sd->last_balance = jiffies;
4736 spin_unlock(&balancing);
4738 if (time_after(next_balance, sd->last_balance + interval)) {
4739 next_balance = sd->last_balance + interval;
4740 update_next_balance = 1;
4744 * Stop the load balance at this level. There is another
4745 * CPU in our sched group which is doing load balancing more
4753 * next_balance will be updated only when there is a need.
4754 * When the cpu is attached to null domain for ex, it will not be
4757 if (likely(update_next_balance))
4758 rq->next_balance = next_balance;
4762 * run_rebalance_domains is triggered when needed from the scheduler tick.
4763 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4764 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4766 static void run_rebalance_domains(struct softirq_action *h)
4768 int this_cpu = smp_processor_id();
4769 struct rq *this_rq = cpu_rq(this_cpu);
4770 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4771 CPU_IDLE : CPU_NOT_IDLE;
4773 rebalance_domains(this_cpu, idle);
4777 * If this cpu is the owner for idle load balancing, then do the
4778 * balancing on behalf of the other idle cpus whose ticks are
4781 if (this_rq->idle_at_tick &&
4782 atomic_read(&nohz.load_balancer) == this_cpu) {
4786 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4787 if (balance_cpu == this_cpu)
4791 * If this cpu gets work to do, stop the load balancing
4792 * work being done for other cpus. Next load
4793 * balancing owner will pick it up.
4798 rebalance_domains(balance_cpu, CPU_IDLE);
4800 rq = cpu_rq(balance_cpu);
4801 if (time_after(this_rq->next_balance, rq->next_balance))
4802 this_rq->next_balance = rq->next_balance;
4808 static inline int on_null_domain(int cpu)
4810 return !rcu_dereference(cpu_rq(cpu)->sd);
4814 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4816 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4817 * idle load balancing owner or decide to stop the periodic load balancing,
4818 * if the whole system is idle.
4820 static inline void trigger_load_balance(struct rq *rq, int cpu)
4824 * If we were in the nohz mode recently and busy at the current
4825 * scheduler tick, then check if we need to nominate new idle
4828 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4829 rq->in_nohz_recently = 0;
4831 if (atomic_read(&nohz.load_balancer) == cpu) {
4832 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4833 atomic_set(&nohz.load_balancer, -1);
4836 if (atomic_read(&nohz.load_balancer) == -1) {
4837 int ilb = find_new_ilb(cpu);
4839 if (ilb < nr_cpu_ids)
4845 * If this cpu is idle and doing idle load balancing for all the
4846 * cpus with ticks stopped, is it time for that to stop?
4848 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4849 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4855 * If this cpu is idle and the idle load balancing is done by
4856 * someone else, then no need raise the SCHED_SOFTIRQ
4858 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4859 cpumask_test_cpu(cpu, nohz.cpu_mask))
4862 /* Don't need to rebalance while attached to NULL domain */
4863 if (time_after_eq(jiffies, rq->next_balance) &&
4864 likely(!on_null_domain(cpu)))
4865 raise_softirq(SCHED_SOFTIRQ);
4868 #else /* CONFIG_SMP */
4871 * on UP we do not need to balance between CPUs:
4873 static inline void idle_balance(int cpu, struct rq *rq)
4879 DEFINE_PER_CPU(struct kernel_stat, kstat);
4881 EXPORT_PER_CPU_SYMBOL(kstat);
4884 * Return any ns on the sched_clock that have not yet been accounted in
4885 * @p in case that task is currently running.
4887 * Called with task_rq_lock() held on @rq.
4889 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4893 if (task_current(rq, p)) {
4894 update_rq_clock(rq);
4895 ns = rq->clock - p->se.exec_start;
4903 unsigned long long task_delta_exec(struct task_struct *p)
4905 unsigned long flags;
4909 rq = task_rq_lock(p, &flags);
4910 ns = do_task_delta_exec(p, rq);
4911 task_rq_unlock(rq, &flags);
4917 * Return accounted runtime for the task.
4918 * In case the task is currently running, return the runtime plus current's
4919 * pending runtime that have not been accounted yet.
4921 unsigned long long task_sched_runtime(struct task_struct *p)
4923 unsigned long flags;
4927 rq = task_rq_lock(p, &flags);
4928 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4929 task_rq_unlock(rq, &flags);
4935 * Return sum_exec_runtime for the thread group.
4936 * In case the task is currently running, return the sum plus current's
4937 * pending runtime that have not been accounted yet.
4939 * Note that the thread group might have other running tasks as well,
4940 * so the return value not includes other pending runtime that other
4941 * running tasks might have.
4943 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4945 struct task_cputime totals;
4946 unsigned long flags;
4950 rq = task_rq_lock(p, &flags);
4951 thread_group_cputime(p, &totals);
4952 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4953 task_rq_unlock(rq, &flags);
4959 * Account user cpu time to a process.
4960 * @p: the process that the cpu time gets accounted to
4961 * @cputime: the cpu time spent in user space since the last update
4962 * @cputime_scaled: cputime scaled by cpu frequency
4964 void account_user_time(struct task_struct *p, cputime_t cputime,
4965 cputime_t cputime_scaled)
4967 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4970 /* Add user time to process. */
4971 p->utime = cputime_add(p->utime, cputime);
4972 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4973 account_group_user_time(p, cputime);
4975 /* Add user time to cpustat. */
4976 tmp = cputime_to_cputime64(cputime);
4977 if (TASK_NICE(p) > 0)
4978 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4980 cpustat->user = cputime64_add(cpustat->user, tmp);
4982 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4983 /* Account for user time used */
4984 acct_update_integrals(p);
4988 * Account guest cpu time to a process.
4989 * @p: the process that the cpu time gets accounted to
4990 * @cputime: the cpu time spent in virtual machine since the last update
4991 * @cputime_scaled: cputime scaled by cpu frequency
4993 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4994 cputime_t cputime_scaled)
4997 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4999 tmp = cputime_to_cputime64(cputime);
5001 /* Add guest time to process. */
5002 p->utime = cputime_add(p->utime, cputime);
5003 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5004 account_group_user_time(p, cputime);
5005 p->gtime = cputime_add(p->gtime, cputime);
5007 /* Add guest time to cpustat. */
5008 cpustat->user = cputime64_add(cpustat->user, tmp);
5009 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5013 * Account system cpu time to a process.
5014 * @p: the process that the cpu time gets accounted to
5015 * @hardirq_offset: the offset to subtract from hardirq_count()
5016 * @cputime: the cpu time spent in kernel space since the last update
5017 * @cputime_scaled: cputime scaled by cpu frequency
5019 void account_system_time(struct task_struct *p, int hardirq_offset,
5020 cputime_t cputime, cputime_t cputime_scaled)
5022 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5025 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5026 account_guest_time(p, cputime, cputime_scaled);
5030 /* Add system time to process. */
5031 p->stime = cputime_add(p->stime, cputime);
5032 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5033 account_group_system_time(p, cputime);
5035 /* Add system time to cpustat. */
5036 tmp = cputime_to_cputime64(cputime);
5037 if (hardirq_count() - hardirq_offset)
5038 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5039 else if (softirq_count())
5040 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5042 cpustat->system = cputime64_add(cpustat->system, tmp);
5044 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5046 /* Account for system time used */
5047 acct_update_integrals(p);
5051 * Account for involuntary wait time.
5052 * @steal: the cpu time spent in involuntary wait
5054 void account_steal_time(cputime_t cputime)
5056 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5057 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5059 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5063 * Account for idle time.
5064 * @cputime: the cpu time spent in idle wait
5066 void account_idle_time(cputime_t cputime)
5068 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5069 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5070 struct rq *rq = this_rq();
5072 if (atomic_read(&rq->nr_iowait) > 0)
5073 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5075 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5078 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5081 * Account a single tick of cpu time.
5082 * @p: the process that the cpu time gets accounted to
5083 * @user_tick: indicates if the tick is a user or a system tick
5085 void account_process_tick(struct task_struct *p, int user_tick)
5087 cputime_t one_jiffy = jiffies_to_cputime(1);
5088 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
5089 struct rq *rq = this_rq();
5092 account_user_time(p, one_jiffy, one_jiffy_scaled);
5093 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5094 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
5097 account_idle_time(one_jiffy);
5101 * Account multiple ticks of steal time.
5102 * @p: the process from which the cpu time has been stolen
5103 * @ticks: number of stolen ticks
5105 void account_steal_ticks(unsigned long ticks)
5107 account_steal_time(jiffies_to_cputime(ticks));
5111 * Account multiple ticks of idle time.
5112 * @ticks: number of stolen ticks
5114 void account_idle_ticks(unsigned long ticks)
5116 account_idle_time(jiffies_to_cputime(ticks));
5122 * Use precise platform statistics if available:
5124 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5125 cputime_t task_utime(struct task_struct *p)
5130 cputime_t task_stime(struct task_struct *p)
5135 cputime_t task_utime(struct task_struct *p)
5137 clock_t utime = cputime_to_clock_t(p->utime),
5138 total = utime + cputime_to_clock_t(p->stime);
5142 * Use CFS's precise accounting:
5144 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5148 do_div(temp, total);
5150 utime = (clock_t)temp;
5152 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5153 return p->prev_utime;
5156 cputime_t task_stime(struct task_struct *p)
5161 * Use CFS's precise accounting. (we subtract utime from
5162 * the total, to make sure the total observed by userspace
5163 * grows monotonically - apps rely on that):
5165 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5166 cputime_to_clock_t(task_utime(p));
5169 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5171 return p->prev_stime;
5175 inline cputime_t task_gtime(struct task_struct *p)
5181 * This function gets called by the timer code, with HZ frequency.
5182 * We call it with interrupts disabled.
5184 * It also gets called by the fork code, when changing the parent's
5187 void scheduler_tick(void)
5189 int cpu = smp_processor_id();
5190 struct rq *rq = cpu_rq(cpu);
5191 struct task_struct *curr = rq->curr;
5195 spin_lock(&rq->lock);
5196 update_rq_clock(rq);
5197 update_cpu_load(rq);
5198 curr->sched_class->task_tick(rq, curr, 0);
5199 spin_unlock(&rq->lock);
5201 perf_counter_task_tick(curr, cpu);
5204 rq->idle_at_tick = idle_cpu(cpu);
5205 trigger_load_balance(rq, cpu);
5209 notrace unsigned long get_parent_ip(unsigned long addr)
5211 if (in_lock_functions(addr)) {
5212 addr = CALLER_ADDR2;
5213 if (in_lock_functions(addr))
5214 addr = CALLER_ADDR3;
5219 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5220 defined(CONFIG_PREEMPT_TRACER))
5222 void __kprobes add_preempt_count(int val)
5224 #ifdef CONFIG_DEBUG_PREEMPT
5228 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5231 preempt_count() += val;
5232 #ifdef CONFIG_DEBUG_PREEMPT
5234 * Spinlock count overflowing soon?
5236 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5239 if (preempt_count() == val)
5240 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5242 EXPORT_SYMBOL(add_preempt_count);
5244 void __kprobes sub_preempt_count(int val)
5246 #ifdef CONFIG_DEBUG_PREEMPT
5250 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5253 * Is the spinlock portion underflowing?
5255 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5256 !(preempt_count() & PREEMPT_MASK)))
5260 if (preempt_count() == val)
5261 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5262 preempt_count() -= val;
5264 EXPORT_SYMBOL(sub_preempt_count);
5269 * Print scheduling while atomic bug:
5271 static noinline void __schedule_bug(struct task_struct *prev)
5273 struct pt_regs *regs = get_irq_regs();
5275 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5276 prev->comm, prev->pid, preempt_count());
5278 debug_show_held_locks(prev);
5280 if (irqs_disabled())
5281 print_irqtrace_events(prev);
5290 * Various schedule()-time debugging checks and statistics:
5292 static inline void schedule_debug(struct task_struct *prev)
5295 * Test if we are atomic. Since do_exit() needs to call into
5296 * schedule() atomically, we ignore that path for now.
5297 * Otherwise, whine if we are scheduling when we should not be.
5299 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5300 __schedule_bug(prev);
5302 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5304 schedstat_inc(this_rq(), sched_count);
5305 #ifdef CONFIG_SCHEDSTATS
5306 if (unlikely(prev->lock_depth >= 0)) {
5307 schedstat_inc(this_rq(), bkl_count);
5308 schedstat_inc(prev, sched_info.bkl_count);
5313 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5315 if (prev->state == TASK_RUNNING) {
5316 u64 runtime = prev->se.sum_exec_runtime;
5318 runtime -= prev->se.prev_sum_exec_runtime;
5319 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5322 * In order to avoid avg_overlap growing stale when we are
5323 * indeed overlapping and hence not getting put to sleep, grow
5324 * the avg_overlap on preemption.
5326 * We use the average preemption runtime because that
5327 * correlates to the amount of cache footprint a task can
5330 update_avg(&prev->se.avg_overlap, runtime);
5332 prev->sched_class->put_prev_task(rq, prev);
5336 * Pick up the highest-prio task:
5338 static inline struct task_struct *
5339 pick_next_task(struct rq *rq)
5341 const struct sched_class *class;
5342 struct task_struct *p;
5345 * Optimization: we know that if all tasks are in
5346 * the fair class we can call that function directly:
5348 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5349 p = fair_sched_class.pick_next_task(rq);
5354 class = sched_class_highest;
5356 p = class->pick_next_task(rq);
5360 * Will never be NULL as the idle class always
5361 * returns a non-NULL p:
5363 class = class->next;
5368 * schedule() is the main scheduler function.
5370 asmlinkage void __sched schedule(void)
5372 struct task_struct *prev, *next;
5373 unsigned long *switch_count;
5374 int post_schedule = 0;
5380 cpu = smp_processor_id();
5384 switch_count = &prev->nivcsw;
5386 release_kernel_lock(prev);
5387 need_resched_nonpreemptible:
5389 schedule_debug(prev);
5391 if (sched_feat(HRTICK))
5394 spin_lock_irq(&rq->lock);
5395 update_rq_clock(rq);
5396 clear_tsk_need_resched(prev);
5398 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5399 if (unlikely(signal_pending_state(prev->state, prev)))
5400 prev->state = TASK_RUNNING;
5402 deactivate_task(rq, prev, 1);
5403 switch_count = &prev->nvcsw;
5407 if (prev->sched_class->pre_schedule)
5408 prev->sched_class->pre_schedule(rq, prev);
5411 if (unlikely(!rq->nr_running))
5412 idle_balance(cpu, rq);
5414 put_prev_task(rq, prev);
5415 next = pick_next_task(rq);
5417 if (likely(prev != next)) {
5418 sched_info_switch(prev, next);
5419 perf_counter_task_sched_out(prev, next, cpu);
5425 post_schedule = context_switch(rq, prev, next); /* unlocks the rq */
5427 * the context switch might have flipped the stack from under
5428 * us, hence refresh the local variables.
5430 cpu = smp_processor_id();
5434 if (current->sched_class->needs_post_schedule)
5435 post_schedule = current->sched_class->needs_post_schedule(rq);
5437 spin_unlock_irq(&rq->lock);
5442 current->sched_class->post_schedule(rq);
5445 if (unlikely(reacquire_kernel_lock(current) < 0))
5446 goto need_resched_nonpreemptible;
5448 preempt_enable_no_resched();
5452 EXPORT_SYMBOL(schedule);
5456 * Look out! "owner" is an entirely speculative pointer
5457 * access and not reliable.
5459 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5464 if (!sched_feat(OWNER_SPIN))
5467 #ifdef CONFIG_DEBUG_PAGEALLOC
5469 * Need to access the cpu field knowing that
5470 * DEBUG_PAGEALLOC could have unmapped it if
5471 * the mutex owner just released it and exited.
5473 if (probe_kernel_address(&owner->cpu, cpu))
5480 * Even if the access succeeded (likely case),
5481 * the cpu field may no longer be valid.
5483 if (cpu >= nr_cpumask_bits)
5487 * We need to validate that we can do a
5488 * get_cpu() and that we have the percpu area.
5490 if (!cpu_online(cpu))
5497 * Owner changed, break to re-assess state.
5499 if (lock->owner != owner)
5503 * Is that owner really running on that cpu?
5505 if (task_thread_info(rq->curr) != owner || need_resched())
5515 #ifdef CONFIG_PREEMPT
5517 * this is the entry point to schedule() from in-kernel preemption
5518 * off of preempt_enable. Kernel preemptions off return from interrupt
5519 * occur there and call schedule directly.
5521 asmlinkage void __sched preempt_schedule(void)
5523 struct thread_info *ti = current_thread_info();
5526 * If there is a non-zero preempt_count or interrupts are disabled,
5527 * we do not want to preempt the current task. Just return..
5529 if (likely(ti->preempt_count || irqs_disabled()))
5533 add_preempt_count(PREEMPT_ACTIVE);
5535 sub_preempt_count(PREEMPT_ACTIVE);
5538 * Check again in case we missed a preemption opportunity
5539 * between schedule and now.
5542 } while (need_resched());
5544 EXPORT_SYMBOL(preempt_schedule);
5547 * this is the entry point to schedule() from kernel preemption
5548 * off of irq context.
5549 * Note, that this is called and return with irqs disabled. This will
5550 * protect us against recursive calling from irq.
5552 asmlinkage void __sched preempt_schedule_irq(void)
5554 struct thread_info *ti = current_thread_info();
5556 /* Catch callers which need to be fixed */
5557 BUG_ON(ti->preempt_count || !irqs_disabled());
5560 add_preempt_count(PREEMPT_ACTIVE);
5563 local_irq_disable();
5564 sub_preempt_count(PREEMPT_ACTIVE);
5567 * Check again in case we missed a preemption opportunity
5568 * between schedule and now.
5571 } while (need_resched());
5574 #endif /* CONFIG_PREEMPT */
5576 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5579 return try_to_wake_up(curr->private, mode, sync);
5581 EXPORT_SYMBOL(default_wake_function);
5584 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5585 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5586 * number) then we wake all the non-exclusive tasks and one exclusive task.
5588 * There are circumstances in which we can try to wake a task which has already
5589 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5590 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5592 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5593 int nr_exclusive, int sync, void *key)
5595 wait_queue_t *curr, *next;
5597 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5598 unsigned flags = curr->flags;
5600 if (curr->func(curr, mode, sync, key) &&
5601 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5607 * __wake_up - wake up threads blocked on a waitqueue.
5609 * @mode: which threads
5610 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5611 * @key: is directly passed to the wakeup function
5613 * It may be assumed that this function implies a write memory barrier before
5614 * changing the task state if and only if any tasks are woken up.
5616 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5617 int nr_exclusive, void *key)
5619 unsigned long flags;
5621 spin_lock_irqsave(&q->lock, flags);
5622 __wake_up_common(q, mode, nr_exclusive, 0, key);
5623 spin_unlock_irqrestore(&q->lock, flags);
5625 EXPORT_SYMBOL(__wake_up);
5628 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5630 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5632 __wake_up_common(q, mode, 1, 0, NULL);
5635 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5637 __wake_up_common(q, mode, 1, 0, key);
5641 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5643 * @mode: which threads
5644 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5645 * @key: opaque value to be passed to wakeup targets
5647 * The sync wakeup differs that the waker knows that it will schedule
5648 * away soon, so while the target thread will be woken up, it will not
5649 * be migrated to another CPU - ie. the two threads are 'synchronized'
5650 * with each other. This can prevent needless bouncing between CPUs.
5652 * On UP it can prevent extra preemption.
5654 * It may be assumed that this function implies a write memory barrier before
5655 * changing the task state if and only if any tasks are woken up.
5657 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5658 int nr_exclusive, void *key)
5660 unsigned long flags;
5666 if (unlikely(!nr_exclusive))
5669 spin_lock_irqsave(&q->lock, flags);
5670 __wake_up_common(q, mode, nr_exclusive, sync, key);
5671 spin_unlock_irqrestore(&q->lock, flags);
5673 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5676 * __wake_up_sync - see __wake_up_sync_key()
5678 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5680 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5682 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5685 * complete: - signals a single thread waiting on this completion
5686 * @x: holds the state of this particular completion
5688 * This will wake up a single thread waiting on this completion. Threads will be
5689 * awakened in the same order in which they were queued.
5691 * See also complete_all(), wait_for_completion() and related routines.
5693 * It may be assumed that this function implies a write memory barrier before
5694 * changing the task state if and only if any tasks are woken up.
5696 void complete(struct completion *x)
5698 unsigned long flags;
5700 spin_lock_irqsave(&x->wait.lock, flags);
5702 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5703 spin_unlock_irqrestore(&x->wait.lock, flags);
5705 EXPORT_SYMBOL(complete);
5708 * complete_all: - signals all threads waiting on this completion
5709 * @x: holds the state of this particular completion
5711 * This will wake up all threads waiting on this particular completion event.
5713 * It may be assumed that this function implies a write memory barrier before
5714 * changing the task state if and only if any tasks are woken up.
5716 void complete_all(struct completion *x)
5718 unsigned long flags;
5720 spin_lock_irqsave(&x->wait.lock, flags);
5721 x->done += UINT_MAX/2;
5722 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5723 spin_unlock_irqrestore(&x->wait.lock, flags);
5725 EXPORT_SYMBOL(complete_all);
5727 static inline long __sched
5728 do_wait_for_common(struct completion *x, long timeout, int state)
5731 DECLARE_WAITQUEUE(wait, current);
5733 wait.flags |= WQ_FLAG_EXCLUSIVE;
5734 __add_wait_queue_tail(&x->wait, &wait);
5736 if (signal_pending_state(state, current)) {
5737 timeout = -ERESTARTSYS;
5740 __set_current_state(state);
5741 spin_unlock_irq(&x->wait.lock);
5742 timeout = schedule_timeout(timeout);
5743 spin_lock_irq(&x->wait.lock);
5744 } while (!x->done && timeout);
5745 __remove_wait_queue(&x->wait, &wait);
5750 return timeout ?: 1;
5754 wait_for_common(struct completion *x, long timeout, int state)
5758 spin_lock_irq(&x->wait.lock);
5759 timeout = do_wait_for_common(x, timeout, state);
5760 spin_unlock_irq(&x->wait.lock);
5765 * wait_for_completion: - waits for completion of a task
5766 * @x: holds the state of this particular completion
5768 * This waits to be signaled for completion of a specific task. It is NOT
5769 * interruptible and there is no timeout.
5771 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5772 * and interrupt capability. Also see complete().
5774 void __sched wait_for_completion(struct completion *x)
5776 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5778 EXPORT_SYMBOL(wait_for_completion);
5781 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5782 * @x: holds the state of this particular completion
5783 * @timeout: timeout value in jiffies
5785 * This waits for either a completion of a specific task to be signaled or for a
5786 * specified timeout to expire. The timeout is in jiffies. It is not
5789 unsigned long __sched
5790 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5792 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5794 EXPORT_SYMBOL(wait_for_completion_timeout);
5797 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5798 * @x: holds the state of this particular completion
5800 * This waits for completion of a specific task to be signaled. It is
5803 int __sched wait_for_completion_interruptible(struct completion *x)
5805 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5806 if (t == -ERESTARTSYS)
5810 EXPORT_SYMBOL(wait_for_completion_interruptible);
5813 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5814 * @x: holds the state of this particular completion
5815 * @timeout: timeout value in jiffies
5817 * This waits for either a completion of a specific task to be signaled or for a
5818 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5820 unsigned long __sched
5821 wait_for_completion_interruptible_timeout(struct completion *x,
5822 unsigned long timeout)
5824 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5826 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5829 * wait_for_completion_killable: - waits for completion of a task (killable)
5830 * @x: holds the state of this particular completion
5832 * This waits to be signaled for completion of a specific task. It can be
5833 * interrupted by a kill signal.
5835 int __sched wait_for_completion_killable(struct completion *x)
5837 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5838 if (t == -ERESTARTSYS)
5842 EXPORT_SYMBOL(wait_for_completion_killable);
5845 * try_wait_for_completion - try to decrement a completion without blocking
5846 * @x: completion structure
5848 * Returns: 0 if a decrement cannot be done without blocking
5849 * 1 if a decrement succeeded.
5851 * If a completion is being used as a counting completion,
5852 * attempt to decrement the counter without blocking. This
5853 * enables us to avoid waiting if the resource the completion
5854 * is protecting is not available.
5856 bool try_wait_for_completion(struct completion *x)
5860 spin_lock_irq(&x->wait.lock);
5865 spin_unlock_irq(&x->wait.lock);
5868 EXPORT_SYMBOL(try_wait_for_completion);
5871 * completion_done - Test to see if a completion has any waiters
5872 * @x: completion structure
5874 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5875 * 1 if there are no waiters.
5878 bool completion_done(struct completion *x)
5882 spin_lock_irq(&x->wait.lock);
5885 spin_unlock_irq(&x->wait.lock);
5888 EXPORT_SYMBOL(completion_done);
5891 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5893 unsigned long flags;
5896 init_waitqueue_entry(&wait, current);
5898 __set_current_state(state);
5900 spin_lock_irqsave(&q->lock, flags);
5901 __add_wait_queue(q, &wait);
5902 spin_unlock(&q->lock);
5903 timeout = schedule_timeout(timeout);
5904 spin_lock_irq(&q->lock);
5905 __remove_wait_queue(q, &wait);
5906 spin_unlock_irqrestore(&q->lock, flags);
5911 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5913 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5915 EXPORT_SYMBOL(interruptible_sleep_on);
5918 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5920 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5922 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5924 void __sched sleep_on(wait_queue_head_t *q)
5926 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5928 EXPORT_SYMBOL(sleep_on);
5930 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5932 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5934 EXPORT_SYMBOL(sleep_on_timeout);
5936 #ifdef CONFIG_RT_MUTEXES
5939 * rt_mutex_setprio - set the current priority of a task
5941 * @prio: prio value (kernel-internal form)
5943 * This function changes the 'effective' priority of a task. It does
5944 * not touch ->normal_prio like __setscheduler().
5946 * Used by the rt_mutex code to implement priority inheritance logic.
5948 void rt_mutex_setprio(struct task_struct *p, int prio)
5950 unsigned long flags;
5951 int oldprio, on_rq, running;
5953 const struct sched_class *prev_class = p->sched_class;
5955 BUG_ON(prio < 0 || prio > MAX_PRIO);
5957 rq = task_rq_lock(p, &flags);
5958 update_rq_clock(rq);
5961 on_rq = p->se.on_rq;
5962 running = task_current(rq, p);
5964 dequeue_task(rq, p, 0);
5966 p->sched_class->put_prev_task(rq, p);
5969 p->sched_class = &rt_sched_class;
5971 p->sched_class = &fair_sched_class;
5976 p->sched_class->set_curr_task(rq);
5978 enqueue_task(rq, p, 0);
5980 check_class_changed(rq, p, prev_class, oldprio, running);
5982 task_rq_unlock(rq, &flags);
5987 void set_user_nice(struct task_struct *p, long nice)
5989 int old_prio, delta, on_rq;
5990 unsigned long flags;
5993 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5996 * We have to be careful, if called from sys_setpriority(),
5997 * the task might be in the middle of scheduling on another CPU.
5999 rq = task_rq_lock(p, &flags);
6000 update_rq_clock(rq);
6002 * The RT priorities are set via sched_setscheduler(), but we still
6003 * allow the 'normal' nice value to be set - but as expected
6004 * it wont have any effect on scheduling until the task is
6005 * SCHED_FIFO/SCHED_RR:
6007 if (task_has_rt_policy(p)) {
6008 p->static_prio = NICE_TO_PRIO(nice);
6011 on_rq = p->se.on_rq;
6013 dequeue_task(rq, p, 0);
6015 p->static_prio = NICE_TO_PRIO(nice);
6018 p->prio = effective_prio(p);
6019 delta = p->prio - old_prio;
6022 enqueue_task(rq, p, 0);
6024 * If the task increased its priority or is running and
6025 * lowered its priority, then reschedule its CPU:
6027 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6028 resched_task(rq->curr);
6031 task_rq_unlock(rq, &flags);
6033 EXPORT_SYMBOL(set_user_nice);
6036 * can_nice - check if a task can reduce its nice value
6040 int can_nice(const struct task_struct *p, const int nice)
6042 /* convert nice value [19,-20] to rlimit style value [1,40] */
6043 int nice_rlim = 20 - nice;
6045 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6046 capable(CAP_SYS_NICE));
6049 #ifdef __ARCH_WANT_SYS_NICE
6052 * sys_nice - change the priority of the current process.
6053 * @increment: priority increment
6055 * sys_setpriority is a more generic, but much slower function that
6056 * does similar things.
6058 SYSCALL_DEFINE1(nice, int, increment)
6063 * Setpriority might change our priority at the same moment.
6064 * We don't have to worry. Conceptually one call occurs first
6065 * and we have a single winner.
6067 if (increment < -40)
6072 nice = TASK_NICE(current) + increment;
6078 if (increment < 0 && !can_nice(current, nice))
6081 retval = security_task_setnice(current, nice);
6085 set_user_nice(current, nice);
6092 * task_prio - return the priority value of a given task.
6093 * @p: the task in question.
6095 * This is the priority value as seen by users in /proc.
6096 * RT tasks are offset by -200. Normal tasks are centered
6097 * around 0, value goes from -16 to +15.
6099 int task_prio(const struct task_struct *p)
6101 return p->prio - MAX_RT_PRIO;
6105 * task_nice - return the nice value of a given task.
6106 * @p: the task in question.
6108 int task_nice(const struct task_struct *p)
6110 return TASK_NICE(p);
6112 EXPORT_SYMBOL(task_nice);
6115 * idle_cpu - is a given cpu idle currently?
6116 * @cpu: the processor in question.
6118 int idle_cpu(int cpu)
6120 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6124 * idle_task - return the idle task for a given cpu.
6125 * @cpu: the processor in question.
6127 struct task_struct *idle_task(int cpu)
6129 return cpu_rq(cpu)->idle;
6133 * find_process_by_pid - find a process with a matching PID value.
6134 * @pid: the pid in question.
6136 static struct task_struct *find_process_by_pid(pid_t pid)
6138 return pid ? find_task_by_vpid(pid) : current;
6141 /* Actually do priority change: must hold rq lock. */
6143 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6145 BUG_ON(p->se.on_rq);
6148 switch (p->policy) {
6152 p->sched_class = &fair_sched_class;
6156 p->sched_class = &rt_sched_class;
6160 p->rt_priority = prio;
6161 p->normal_prio = normal_prio(p);
6162 /* we are holding p->pi_lock already */
6163 p->prio = rt_mutex_getprio(p);
6168 * check the target process has a UID that matches the current process's
6170 static bool check_same_owner(struct task_struct *p)
6172 const struct cred *cred = current_cred(), *pcred;
6176 pcred = __task_cred(p);
6177 match = (cred->euid == pcred->euid ||
6178 cred->euid == pcred->uid);
6183 static int __sched_setscheduler(struct task_struct *p, int policy,
6184 struct sched_param *param, bool user)
6186 int retval, oldprio, oldpolicy = -1, on_rq, running;
6187 unsigned long flags;
6188 const struct sched_class *prev_class = p->sched_class;
6192 /* may grab non-irq protected spin_locks */
6193 BUG_ON(in_interrupt());
6195 /* double check policy once rq lock held */
6197 reset_on_fork = p->sched_reset_on_fork;
6198 policy = oldpolicy = p->policy;
6200 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6201 policy &= ~SCHED_RESET_ON_FORK;
6203 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6204 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6205 policy != SCHED_IDLE)
6210 * Valid priorities for SCHED_FIFO and SCHED_RR are
6211 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6212 * SCHED_BATCH and SCHED_IDLE is 0.
6214 if (param->sched_priority < 0 ||
6215 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6216 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6218 if (rt_policy(policy) != (param->sched_priority != 0))
6222 * Allow unprivileged RT tasks to decrease priority:
6224 if (user && !capable(CAP_SYS_NICE)) {
6225 if (rt_policy(policy)) {
6226 unsigned long rlim_rtprio;
6228 if (!lock_task_sighand(p, &flags))
6230 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6231 unlock_task_sighand(p, &flags);
6233 /* can't set/change the rt policy */
6234 if (policy != p->policy && !rlim_rtprio)
6237 /* can't increase priority */
6238 if (param->sched_priority > p->rt_priority &&
6239 param->sched_priority > rlim_rtprio)
6243 * Like positive nice levels, dont allow tasks to
6244 * move out of SCHED_IDLE either:
6246 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6249 /* can't change other user's priorities */
6250 if (!check_same_owner(p))
6253 /* Normal users shall not reset the sched_reset_on_fork flag */
6254 if (p->sched_reset_on_fork && !reset_on_fork)
6259 #ifdef CONFIG_RT_GROUP_SCHED
6261 * Do not allow realtime tasks into groups that have no runtime
6264 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6265 task_group(p)->rt_bandwidth.rt_runtime == 0)
6269 retval = security_task_setscheduler(p, policy, param);
6275 * make sure no PI-waiters arrive (or leave) while we are
6276 * changing the priority of the task:
6278 spin_lock_irqsave(&p->pi_lock, flags);
6280 * To be able to change p->policy safely, the apropriate
6281 * runqueue lock must be held.
6283 rq = __task_rq_lock(p);
6284 /* recheck policy now with rq lock held */
6285 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6286 policy = oldpolicy = -1;
6287 __task_rq_unlock(rq);
6288 spin_unlock_irqrestore(&p->pi_lock, flags);
6291 update_rq_clock(rq);
6292 on_rq = p->se.on_rq;
6293 running = task_current(rq, p);
6295 deactivate_task(rq, p, 0);
6297 p->sched_class->put_prev_task(rq, p);
6299 p->sched_reset_on_fork = reset_on_fork;
6302 __setscheduler(rq, p, policy, param->sched_priority);
6305 p->sched_class->set_curr_task(rq);
6307 activate_task(rq, p, 0);
6309 check_class_changed(rq, p, prev_class, oldprio, running);
6311 __task_rq_unlock(rq);
6312 spin_unlock_irqrestore(&p->pi_lock, flags);
6314 rt_mutex_adjust_pi(p);
6320 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6321 * @p: the task in question.
6322 * @policy: new policy.
6323 * @param: structure containing the new RT priority.
6325 * NOTE that the task may be already dead.
6327 int sched_setscheduler(struct task_struct *p, int policy,
6328 struct sched_param *param)
6330 return __sched_setscheduler(p, policy, param, true);
6332 EXPORT_SYMBOL_GPL(sched_setscheduler);
6335 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6336 * @p: the task in question.
6337 * @policy: new policy.
6338 * @param: structure containing the new RT priority.
6340 * Just like sched_setscheduler, only don't bother checking if the
6341 * current context has permission. For example, this is needed in
6342 * stop_machine(): we create temporary high priority worker threads,
6343 * but our caller might not have that capability.
6345 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6346 struct sched_param *param)
6348 return __sched_setscheduler(p, policy, param, false);
6352 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6354 struct sched_param lparam;
6355 struct task_struct *p;
6358 if (!param || pid < 0)
6360 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6365 p = find_process_by_pid(pid);
6367 retval = sched_setscheduler(p, policy, &lparam);
6374 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6375 * @pid: the pid in question.
6376 * @policy: new policy.
6377 * @param: structure containing the new RT priority.
6379 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6380 struct sched_param __user *, param)
6382 /* negative values for policy are not valid */
6386 return do_sched_setscheduler(pid, policy, param);
6390 * sys_sched_setparam - set/change the RT priority of a thread
6391 * @pid: the pid in question.
6392 * @param: structure containing the new RT priority.
6394 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6396 return do_sched_setscheduler(pid, -1, param);
6400 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6401 * @pid: the pid in question.
6403 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6405 struct task_struct *p;
6412 read_lock(&tasklist_lock);
6413 p = find_process_by_pid(pid);
6415 retval = security_task_getscheduler(p);
6418 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6420 read_unlock(&tasklist_lock);
6425 * sys_sched_getparam - get the RT priority of a thread
6426 * @pid: the pid in question.
6427 * @param: structure containing the RT priority.
6429 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6431 struct sched_param lp;
6432 struct task_struct *p;
6435 if (!param || pid < 0)
6438 read_lock(&tasklist_lock);
6439 p = find_process_by_pid(pid);
6444 retval = security_task_getscheduler(p);
6448 lp.sched_priority = p->rt_priority;
6449 read_unlock(&tasklist_lock);
6452 * This one might sleep, we cannot do it with a spinlock held ...
6454 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6459 read_unlock(&tasklist_lock);
6463 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6465 cpumask_var_t cpus_allowed, new_mask;
6466 struct task_struct *p;
6470 read_lock(&tasklist_lock);
6472 p = find_process_by_pid(pid);
6474 read_unlock(&tasklist_lock);
6480 * It is not safe to call set_cpus_allowed with the
6481 * tasklist_lock held. We will bump the task_struct's
6482 * usage count and then drop tasklist_lock.
6485 read_unlock(&tasklist_lock);
6487 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6491 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6493 goto out_free_cpus_allowed;
6496 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6499 retval = security_task_setscheduler(p, 0, NULL);
6503 cpuset_cpus_allowed(p, cpus_allowed);
6504 cpumask_and(new_mask, in_mask, cpus_allowed);
6506 retval = set_cpus_allowed_ptr(p, new_mask);
6509 cpuset_cpus_allowed(p, cpus_allowed);
6510 if (!cpumask_subset(new_mask, cpus_allowed)) {
6512 * We must have raced with a concurrent cpuset
6513 * update. Just reset the cpus_allowed to the
6514 * cpuset's cpus_allowed
6516 cpumask_copy(new_mask, cpus_allowed);
6521 free_cpumask_var(new_mask);
6522 out_free_cpus_allowed:
6523 free_cpumask_var(cpus_allowed);
6530 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6531 struct cpumask *new_mask)
6533 if (len < cpumask_size())
6534 cpumask_clear(new_mask);
6535 else if (len > cpumask_size())
6536 len = cpumask_size();
6538 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6542 * sys_sched_setaffinity - set the cpu affinity of a process
6543 * @pid: pid of the process
6544 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6545 * @user_mask_ptr: user-space pointer to the new cpu mask
6547 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6548 unsigned long __user *, user_mask_ptr)
6550 cpumask_var_t new_mask;
6553 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6556 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6558 retval = sched_setaffinity(pid, new_mask);
6559 free_cpumask_var(new_mask);
6563 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6565 struct task_struct *p;
6569 read_lock(&tasklist_lock);
6572 p = find_process_by_pid(pid);
6576 retval = security_task_getscheduler(p);
6580 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6583 read_unlock(&tasklist_lock);
6590 * sys_sched_getaffinity - get the cpu affinity of a process
6591 * @pid: pid of the process
6592 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6593 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6595 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6596 unsigned long __user *, user_mask_ptr)
6601 if (len < cpumask_size())
6604 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6607 ret = sched_getaffinity(pid, mask);
6609 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6612 ret = cpumask_size();
6614 free_cpumask_var(mask);
6620 * sys_sched_yield - yield the current processor to other threads.
6622 * This function yields the current CPU to other tasks. If there are no
6623 * other threads running on this CPU then this function will return.
6625 SYSCALL_DEFINE0(sched_yield)
6627 struct rq *rq = this_rq_lock();
6629 schedstat_inc(rq, yld_count);
6630 current->sched_class->yield_task(rq);
6633 * Since we are going to call schedule() anyway, there's
6634 * no need to preempt or enable interrupts:
6636 __release(rq->lock);
6637 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6638 _raw_spin_unlock(&rq->lock);
6639 preempt_enable_no_resched();
6646 static inline int should_resched(void)
6648 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6651 static void __cond_resched(void)
6653 add_preempt_count(PREEMPT_ACTIVE);
6655 sub_preempt_count(PREEMPT_ACTIVE);
6658 int __sched _cond_resched(void)
6660 if (should_resched()) {
6666 EXPORT_SYMBOL(_cond_resched);
6669 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6670 * call schedule, and on return reacquire the lock.
6672 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6673 * operations here to prevent schedule() from being called twice (once via
6674 * spin_unlock(), once by hand).
6676 int __cond_resched_lock(spinlock_t *lock)
6678 int resched = should_resched();
6681 if (spin_needbreak(lock) || resched) {
6692 EXPORT_SYMBOL(__cond_resched_lock);
6694 int __sched __cond_resched_softirq(void)
6696 BUG_ON(!in_softirq());
6698 if (should_resched()) {
6706 EXPORT_SYMBOL(__cond_resched_softirq);
6709 * yield - yield the current processor to other threads.
6711 * This is a shortcut for kernel-space yielding - it marks the
6712 * thread runnable and calls sys_sched_yield().
6714 void __sched yield(void)
6716 set_current_state(TASK_RUNNING);
6719 EXPORT_SYMBOL(yield);
6722 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6723 * that process accounting knows that this is a task in IO wait state.
6725 * But don't do that if it is a deliberate, throttling IO wait (this task
6726 * has set its backing_dev_info: the queue against which it should throttle)
6728 void __sched io_schedule(void)
6730 struct rq *rq = raw_rq();
6732 delayacct_blkio_start();
6733 atomic_inc(&rq->nr_iowait);
6735 atomic_dec(&rq->nr_iowait);
6736 delayacct_blkio_end();
6738 EXPORT_SYMBOL(io_schedule);
6740 long __sched io_schedule_timeout(long timeout)
6742 struct rq *rq = raw_rq();
6745 delayacct_blkio_start();
6746 atomic_inc(&rq->nr_iowait);
6747 ret = schedule_timeout(timeout);
6748 atomic_dec(&rq->nr_iowait);
6749 delayacct_blkio_end();
6754 * sys_sched_get_priority_max - return maximum RT priority.
6755 * @policy: scheduling class.
6757 * this syscall returns the maximum rt_priority that can be used
6758 * by a given scheduling class.
6760 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6767 ret = MAX_USER_RT_PRIO-1;
6779 * sys_sched_get_priority_min - return minimum RT priority.
6780 * @policy: scheduling class.
6782 * this syscall returns the minimum rt_priority that can be used
6783 * by a given scheduling class.
6785 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6803 * sys_sched_rr_get_interval - return the default timeslice of a process.
6804 * @pid: pid of the process.
6805 * @interval: userspace pointer to the timeslice value.
6807 * this syscall writes the default timeslice value of a given process
6808 * into the user-space timespec buffer. A value of '0' means infinity.
6810 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6811 struct timespec __user *, interval)
6813 struct task_struct *p;
6814 unsigned int time_slice;
6822 read_lock(&tasklist_lock);
6823 p = find_process_by_pid(pid);
6827 retval = security_task_getscheduler(p);
6832 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6833 * tasks that are on an otherwise idle runqueue:
6836 if (p->policy == SCHED_RR) {
6837 time_slice = DEF_TIMESLICE;
6838 } else if (p->policy != SCHED_FIFO) {
6839 struct sched_entity *se = &p->se;
6840 unsigned long flags;
6843 rq = task_rq_lock(p, &flags);
6844 if (rq->cfs.load.weight)
6845 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6846 task_rq_unlock(rq, &flags);
6848 read_unlock(&tasklist_lock);
6849 jiffies_to_timespec(time_slice, &t);
6850 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6854 read_unlock(&tasklist_lock);
6858 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6860 void sched_show_task(struct task_struct *p)
6862 unsigned long free = 0;
6865 state = p->state ? __ffs(p->state) + 1 : 0;
6866 printk(KERN_INFO "%-13.13s %c", p->comm,
6867 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6868 #if BITS_PER_LONG == 32
6869 if (state == TASK_RUNNING)
6870 printk(KERN_CONT " running ");
6872 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6874 if (state == TASK_RUNNING)
6875 printk(KERN_CONT " running task ");
6877 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6879 #ifdef CONFIG_DEBUG_STACK_USAGE
6880 free = stack_not_used(p);
6882 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6883 task_pid_nr(p), task_pid_nr(p->real_parent),
6884 (unsigned long)task_thread_info(p)->flags);
6886 show_stack(p, NULL);
6889 void show_state_filter(unsigned long state_filter)
6891 struct task_struct *g, *p;
6893 #if BITS_PER_LONG == 32
6895 " task PC stack pid father\n");
6898 " task PC stack pid father\n");
6900 read_lock(&tasklist_lock);
6901 do_each_thread(g, p) {
6903 * reset the NMI-timeout, listing all files on a slow
6904 * console might take alot of time:
6906 touch_nmi_watchdog();
6907 if (!state_filter || (p->state & state_filter))
6909 } while_each_thread(g, p);
6911 touch_all_softlockup_watchdogs();
6913 #ifdef CONFIG_SCHED_DEBUG
6914 sysrq_sched_debug_show();
6916 read_unlock(&tasklist_lock);
6918 * Only show locks if all tasks are dumped:
6920 if (state_filter == -1)
6921 debug_show_all_locks();
6924 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6926 idle->sched_class = &idle_sched_class;
6930 * init_idle - set up an idle thread for a given CPU
6931 * @idle: task in question
6932 * @cpu: cpu the idle task belongs to
6934 * NOTE: this function does not set the idle thread's NEED_RESCHED
6935 * flag, to make booting more robust.
6937 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6939 struct rq *rq = cpu_rq(cpu);
6940 unsigned long flags;
6942 spin_lock_irqsave(&rq->lock, flags);
6945 idle->se.exec_start = sched_clock();
6947 idle->prio = idle->normal_prio = MAX_PRIO;
6948 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6949 __set_task_cpu(idle, cpu);
6951 rq->curr = rq->idle = idle;
6952 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6955 spin_unlock_irqrestore(&rq->lock, flags);
6957 /* Set the preempt count _outside_ the spinlocks! */
6958 #if defined(CONFIG_PREEMPT)
6959 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6961 task_thread_info(idle)->preempt_count = 0;
6964 * The idle tasks have their own, simple scheduling class:
6966 idle->sched_class = &idle_sched_class;
6967 ftrace_graph_init_task(idle);
6971 * In a system that switches off the HZ timer nohz_cpu_mask
6972 * indicates which cpus entered this state. This is used
6973 * in the rcu update to wait only for active cpus. For system
6974 * which do not switch off the HZ timer nohz_cpu_mask should
6975 * always be CPU_BITS_NONE.
6977 cpumask_var_t nohz_cpu_mask;
6980 * Increase the granularity value when there are more CPUs,
6981 * because with more CPUs the 'effective latency' as visible
6982 * to users decreases. But the relationship is not linear,
6983 * so pick a second-best guess by going with the log2 of the
6986 * This idea comes from the SD scheduler of Con Kolivas:
6988 static inline void sched_init_granularity(void)
6990 unsigned int factor = 1 + ilog2(num_online_cpus());
6991 const unsigned long limit = 200000000;
6993 sysctl_sched_min_granularity *= factor;
6994 if (sysctl_sched_min_granularity > limit)
6995 sysctl_sched_min_granularity = limit;
6997 sysctl_sched_latency *= factor;
6998 if (sysctl_sched_latency > limit)
6999 sysctl_sched_latency = limit;
7001 sysctl_sched_wakeup_granularity *= factor;
7003 sysctl_sched_shares_ratelimit *= factor;
7008 * This is how migration works:
7010 * 1) we queue a struct migration_req structure in the source CPU's
7011 * runqueue and wake up that CPU's migration thread.
7012 * 2) we down() the locked semaphore => thread blocks.
7013 * 3) migration thread wakes up (implicitly it forces the migrated
7014 * thread off the CPU)
7015 * 4) it gets the migration request and checks whether the migrated
7016 * task is still in the wrong runqueue.
7017 * 5) if it's in the wrong runqueue then the migration thread removes
7018 * it and puts it into the right queue.
7019 * 6) migration thread up()s the semaphore.
7020 * 7) we wake up and the migration is done.
7024 * Change a given task's CPU affinity. Migrate the thread to a
7025 * proper CPU and schedule it away if the CPU it's executing on
7026 * is removed from the allowed bitmask.
7028 * NOTE: the caller must have a valid reference to the task, the
7029 * task must not exit() & deallocate itself prematurely. The
7030 * call is not atomic; no spinlocks may be held.
7032 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7034 struct migration_req req;
7035 unsigned long flags;
7039 rq = task_rq_lock(p, &flags);
7040 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
7045 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7046 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7051 if (p->sched_class->set_cpus_allowed)
7052 p->sched_class->set_cpus_allowed(p, new_mask);
7054 cpumask_copy(&p->cpus_allowed, new_mask);
7055 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7058 /* Can the task run on the task's current CPU? If so, we're done */
7059 if (cpumask_test_cpu(task_cpu(p), new_mask))
7062 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
7063 /* Need help from migration thread: drop lock and wait. */
7064 task_rq_unlock(rq, &flags);
7065 wake_up_process(rq->migration_thread);
7066 wait_for_completion(&req.done);
7067 tlb_migrate_finish(p->mm);
7071 task_rq_unlock(rq, &flags);
7075 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7078 * Move (not current) task off this cpu, onto dest cpu. We're doing
7079 * this because either it can't run here any more (set_cpus_allowed()
7080 * away from this CPU, or CPU going down), or because we're
7081 * attempting to rebalance this task on exec (sched_exec).
7083 * So we race with normal scheduler movements, but that's OK, as long
7084 * as the task is no longer on this CPU.
7086 * Returns non-zero if task was successfully migrated.
7088 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7090 struct rq *rq_dest, *rq_src;
7093 if (unlikely(!cpu_active(dest_cpu)))
7096 rq_src = cpu_rq(src_cpu);
7097 rq_dest = cpu_rq(dest_cpu);
7099 double_rq_lock(rq_src, rq_dest);
7100 /* Already moved. */
7101 if (task_cpu(p) != src_cpu)
7103 /* Affinity changed (again). */
7104 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7107 on_rq = p->se.on_rq;
7109 deactivate_task(rq_src, p, 0);
7111 set_task_cpu(p, dest_cpu);
7113 activate_task(rq_dest, p, 0);
7114 check_preempt_curr(rq_dest, p, 0);
7119 double_rq_unlock(rq_src, rq_dest);
7124 * migration_thread - this is a highprio system thread that performs
7125 * thread migration by bumping thread off CPU then 'pushing' onto
7128 static int migration_thread(void *data)
7130 int cpu = (long)data;
7134 BUG_ON(rq->migration_thread != current);
7136 set_current_state(TASK_INTERRUPTIBLE);
7137 while (!kthread_should_stop()) {
7138 struct migration_req *req;
7139 struct list_head *head;
7141 spin_lock_irq(&rq->lock);
7143 if (cpu_is_offline(cpu)) {
7144 spin_unlock_irq(&rq->lock);
7148 if (rq->active_balance) {
7149 active_load_balance(rq, cpu);
7150 rq->active_balance = 0;
7153 head = &rq->migration_queue;
7155 if (list_empty(head)) {
7156 spin_unlock_irq(&rq->lock);
7158 set_current_state(TASK_INTERRUPTIBLE);
7161 req = list_entry(head->next, struct migration_req, list);
7162 list_del_init(head->next);
7164 spin_unlock(&rq->lock);
7165 __migrate_task(req->task, cpu, req->dest_cpu);
7168 complete(&req->done);
7170 __set_current_state(TASK_RUNNING);
7175 #ifdef CONFIG_HOTPLUG_CPU
7177 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7181 local_irq_disable();
7182 ret = __migrate_task(p, src_cpu, dest_cpu);
7188 * Figure out where task on dead CPU should go, use force if necessary.
7190 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7193 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7196 /* Look for allowed, online CPU in same node. */
7197 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7198 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7201 /* Any allowed, online CPU? */
7202 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7203 if (dest_cpu < nr_cpu_ids)
7206 /* No more Mr. Nice Guy. */
7207 if (dest_cpu >= nr_cpu_ids) {
7208 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7209 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7212 * Don't tell them about moving exiting tasks or
7213 * kernel threads (both mm NULL), since they never
7216 if (p->mm && printk_ratelimit()) {
7217 printk(KERN_INFO "process %d (%s) no "
7218 "longer affine to cpu%d\n",
7219 task_pid_nr(p), p->comm, dead_cpu);
7224 /* It can have affinity changed while we were choosing. */
7225 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7230 * While a dead CPU has no uninterruptible tasks queued at this point,
7231 * it might still have a nonzero ->nr_uninterruptible counter, because
7232 * for performance reasons the counter is not stricly tracking tasks to
7233 * their home CPUs. So we just add the counter to another CPU's counter,
7234 * to keep the global sum constant after CPU-down:
7236 static void migrate_nr_uninterruptible(struct rq *rq_src)
7238 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7239 unsigned long flags;
7241 local_irq_save(flags);
7242 double_rq_lock(rq_src, rq_dest);
7243 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7244 rq_src->nr_uninterruptible = 0;
7245 double_rq_unlock(rq_src, rq_dest);
7246 local_irq_restore(flags);
7249 /* Run through task list and migrate tasks from the dead cpu. */
7250 static void migrate_live_tasks(int src_cpu)
7252 struct task_struct *p, *t;
7254 read_lock(&tasklist_lock);
7256 do_each_thread(t, p) {
7260 if (task_cpu(p) == src_cpu)
7261 move_task_off_dead_cpu(src_cpu, p);
7262 } while_each_thread(t, p);
7264 read_unlock(&tasklist_lock);
7268 * Schedules idle task to be the next runnable task on current CPU.
7269 * It does so by boosting its priority to highest possible.
7270 * Used by CPU offline code.
7272 void sched_idle_next(void)
7274 int this_cpu = smp_processor_id();
7275 struct rq *rq = cpu_rq(this_cpu);
7276 struct task_struct *p = rq->idle;
7277 unsigned long flags;
7279 /* cpu has to be offline */
7280 BUG_ON(cpu_online(this_cpu));
7283 * Strictly not necessary since rest of the CPUs are stopped by now
7284 * and interrupts disabled on the current cpu.
7286 spin_lock_irqsave(&rq->lock, flags);
7288 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7290 update_rq_clock(rq);
7291 activate_task(rq, p, 0);
7293 spin_unlock_irqrestore(&rq->lock, flags);
7297 * Ensures that the idle task is using init_mm right before its cpu goes
7300 void idle_task_exit(void)
7302 struct mm_struct *mm = current->active_mm;
7304 BUG_ON(cpu_online(smp_processor_id()));
7307 switch_mm(mm, &init_mm, current);
7311 /* called under rq->lock with disabled interrupts */
7312 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7314 struct rq *rq = cpu_rq(dead_cpu);
7316 /* Must be exiting, otherwise would be on tasklist. */
7317 BUG_ON(!p->exit_state);
7319 /* Cannot have done final schedule yet: would have vanished. */
7320 BUG_ON(p->state == TASK_DEAD);
7325 * Drop lock around migration; if someone else moves it,
7326 * that's OK. No task can be added to this CPU, so iteration is
7329 spin_unlock_irq(&rq->lock);
7330 move_task_off_dead_cpu(dead_cpu, p);
7331 spin_lock_irq(&rq->lock);
7336 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7337 static void migrate_dead_tasks(unsigned int dead_cpu)
7339 struct rq *rq = cpu_rq(dead_cpu);
7340 struct task_struct *next;
7343 if (!rq->nr_running)
7345 update_rq_clock(rq);
7346 next = pick_next_task(rq);
7349 next->sched_class->put_prev_task(rq, next);
7350 migrate_dead(dead_cpu, next);
7356 * remove the tasks which were accounted by rq from calc_load_tasks.
7358 static void calc_global_load_remove(struct rq *rq)
7360 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7361 rq->calc_load_active = 0;
7363 #endif /* CONFIG_HOTPLUG_CPU */
7365 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7367 static struct ctl_table sd_ctl_dir[] = {
7369 .procname = "sched_domain",
7375 static struct ctl_table sd_ctl_root[] = {
7377 .ctl_name = CTL_KERN,
7378 .procname = "kernel",
7380 .child = sd_ctl_dir,
7385 static struct ctl_table *sd_alloc_ctl_entry(int n)
7387 struct ctl_table *entry =
7388 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7393 static void sd_free_ctl_entry(struct ctl_table **tablep)
7395 struct ctl_table *entry;
7398 * In the intermediate directories, both the child directory and
7399 * procname are dynamically allocated and could fail but the mode
7400 * will always be set. In the lowest directory the names are
7401 * static strings and all have proc handlers.
7403 for (entry = *tablep; entry->mode; entry++) {
7405 sd_free_ctl_entry(&entry->child);
7406 if (entry->proc_handler == NULL)
7407 kfree(entry->procname);
7415 set_table_entry(struct ctl_table *entry,
7416 const char *procname, void *data, int maxlen,
7417 mode_t mode, proc_handler *proc_handler)
7419 entry->procname = procname;
7421 entry->maxlen = maxlen;
7423 entry->proc_handler = proc_handler;
7426 static struct ctl_table *
7427 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7429 struct ctl_table *table = sd_alloc_ctl_entry(13);
7434 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7435 sizeof(long), 0644, proc_doulongvec_minmax);
7436 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7437 sizeof(long), 0644, proc_doulongvec_minmax);
7438 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7439 sizeof(int), 0644, proc_dointvec_minmax);
7440 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7441 sizeof(int), 0644, proc_dointvec_minmax);
7442 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7443 sizeof(int), 0644, proc_dointvec_minmax);
7444 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7445 sizeof(int), 0644, proc_dointvec_minmax);
7446 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7447 sizeof(int), 0644, proc_dointvec_minmax);
7448 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7449 sizeof(int), 0644, proc_dointvec_minmax);
7450 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7451 sizeof(int), 0644, proc_dointvec_minmax);
7452 set_table_entry(&table[9], "cache_nice_tries",
7453 &sd->cache_nice_tries,
7454 sizeof(int), 0644, proc_dointvec_minmax);
7455 set_table_entry(&table[10], "flags", &sd->flags,
7456 sizeof(int), 0644, proc_dointvec_minmax);
7457 set_table_entry(&table[11], "name", sd->name,
7458 CORENAME_MAX_SIZE, 0444, proc_dostring);
7459 /* &table[12] is terminator */
7464 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7466 struct ctl_table *entry, *table;
7467 struct sched_domain *sd;
7468 int domain_num = 0, i;
7471 for_each_domain(cpu, sd)
7473 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7478 for_each_domain(cpu, sd) {
7479 snprintf(buf, 32, "domain%d", i);
7480 entry->procname = kstrdup(buf, GFP_KERNEL);
7482 entry->child = sd_alloc_ctl_domain_table(sd);
7489 static struct ctl_table_header *sd_sysctl_header;
7490 static void register_sched_domain_sysctl(void)
7492 int i, cpu_num = num_online_cpus();
7493 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7496 WARN_ON(sd_ctl_dir[0].child);
7497 sd_ctl_dir[0].child = entry;
7502 for_each_online_cpu(i) {
7503 snprintf(buf, 32, "cpu%d", i);
7504 entry->procname = kstrdup(buf, GFP_KERNEL);
7506 entry->child = sd_alloc_ctl_cpu_table(i);
7510 WARN_ON(sd_sysctl_header);
7511 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7514 /* may be called multiple times per register */
7515 static void unregister_sched_domain_sysctl(void)
7517 if (sd_sysctl_header)
7518 unregister_sysctl_table(sd_sysctl_header);
7519 sd_sysctl_header = NULL;
7520 if (sd_ctl_dir[0].child)
7521 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7524 static void register_sched_domain_sysctl(void)
7527 static void unregister_sched_domain_sysctl(void)
7532 static void set_rq_online(struct rq *rq)
7535 const struct sched_class *class;
7537 cpumask_set_cpu(rq->cpu, rq->rd->online);
7540 for_each_class(class) {
7541 if (class->rq_online)
7542 class->rq_online(rq);
7547 static void set_rq_offline(struct rq *rq)
7550 const struct sched_class *class;
7552 for_each_class(class) {
7553 if (class->rq_offline)
7554 class->rq_offline(rq);
7557 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7563 * migration_call - callback that gets triggered when a CPU is added.
7564 * Here we can start up the necessary migration thread for the new CPU.
7566 static int __cpuinit
7567 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7569 struct task_struct *p;
7570 int cpu = (long)hcpu;
7571 unsigned long flags;
7576 case CPU_UP_PREPARE:
7577 case CPU_UP_PREPARE_FROZEN:
7578 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7581 kthread_bind(p, cpu);
7582 /* Must be high prio: stop_machine expects to yield to it. */
7583 rq = task_rq_lock(p, &flags);
7584 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7585 task_rq_unlock(rq, &flags);
7587 cpu_rq(cpu)->migration_thread = p;
7588 rq->calc_load_update = calc_load_update;
7592 case CPU_ONLINE_FROZEN:
7593 /* Strictly unnecessary, as first user will wake it. */
7594 wake_up_process(cpu_rq(cpu)->migration_thread);
7596 /* Update our root-domain */
7598 spin_lock_irqsave(&rq->lock, flags);
7600 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7604 spin_unlock_irqrestore(&rq->lock, flags);
7607 #ifdef CONFIG_HOTPLUG_CPU
7608 case CPU_UP_CANCELED:
7609 case CPU_UP_CANCELED_FROZEN:
7610 if (!cpu_rq(cpu)->migration_thread)
7612 /* Unbind it from offline cpu so it can run. Fall thru. */
7613 kthread_bind(cpu_rq(cpu)->migration_thread,
7614 cpumask_any(cpu_online_mask));
7615 kthread_stop(cpu_rq(cpu)->migration_thread);
7616 put_task_struct(cpu_rq(cpu)->migration_thread);
7617 cpu_rq(cpu)->migration_thread = NULL;
7621 case CPU_DEAD_FROZEN:
7622 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7623 migrate_live_tasks(cpu);
7625 kthread_stop(rq->migration_thread);
7626 put_task_struct(rq->migration_thread);
7627 rq->migration_thread = NULL;
7628 /* Idle task back to normal (off runqueue, low prio) */
7629 spin_lock_irq(&rq->lock);
7630 update_rq_clock(rq);
7631 deactivate_task(rq, rq->idle, 0);
7632 rq->idle->static_prio = MAX_PRIO;
7633 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7634 rq->idle->sched_class = &idle_sched_class;
7635 migrate_dead_tasks(cpu);
7636 spin_unlock_irq(&rq->lock);
7638 migrate_nr_uninterruptible(rq);
7639 BUG_ON(rq->nr_running != 0);
7640 calc_global_load_remove(rq);
7642 * No need to migrate the tasks: it was best-effort if
7643 * they didn't take sched_hotcpu_mutex. Just wake up
7646 spin_lock_irq(&rq->lock);
7647 while (!list_empty(&rq->migration_queue)) {
7648 struct migration_req *req;
7650 req = list_entry(rq->migration_queue.next,
7651 struct migration_req, list);
7652 list_del_init(&req->list);
7653 spin_unlock_irq(&rq->lock);
7654 complete(&req->done);
7655 spin_lock_irq(&rq->lock);
7657 spin_unlock_irq(&rq->lock);
7661 case CPU_DYING_FROZEN:
7662 /* Update our root-domain */
7664 spin_lock_irqsave(&rq->lock, flags);
7666 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7669 spin_unlock_irqrestore(&rq->lock, flags);
7677 * Register at high priority so that task migration (migrate_all_tasks)
7678 * happens before everything else. This has to be lower priority than
7679 * the notifier in the perf_counter subsystem, though.
7681 static struct notifier_block __cpuinitdata migration_notifier = {
7682 .notifier_call = migration_call,
7686 static int __init migration_init(void)
7688 void *cpu = (void *)(long)smp_processor_id();
7691 /* Start one for the boot CPU: */
7692 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7693 BUG_ON(err == NOTIFY_BAD);
7694 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7695 register_cpu_notifier(&migration_notifier);
7699 early_initcall(migration_init);
7704 #ifdef CONFIG_SCHED_DEBUG
7706 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7707 struct cpumask *groupmask)
7709 struct sched_group *group = sd->groups;
7712 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7713 cpumask_clear(groupmask);
7715 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7717 if (!(sd->flags & SD_LOAD_BALANCE)) {
7718 printk("does not load-balance\n");
7720 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7725 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7727 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7728 printk(KERN_ERR "ERROR: domain->span does not contain "
7731 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7732 printk(KERN_ERR "ERROR: domain->groups does not contain"
7736 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7740 printk(KERN_ERR "ERROR: group is NULL\n");
7744 if (!group->__cpu_power) {
7745 printk(KERN_CONT "\n");
7746 printk(KERN_ERR "ERROR: domain->cpu_power not "
7751 if (!cpumask_weight(sched_group_cpus(group))) {
7752 printk(KERN_CONT "\n");
7753 printk(KERN_ERR "ERROR: empty group\n");
7757 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7758 printk(KERN_CONT "\n");
7759 printk(KERN_ERR "ERROR: repeated CPUs\n");
7763 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7765 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7767 printk(KERN_CONT " %s", str);
7768 if (group->__cpu_power != SCHED_LOAD_SCALE) {
7769 printk(KERN_CONT " (__cpu_power = %d)",
7770 group->__cpu_power);
7773 group = group->next;
7774 } while (group != sd->groups);
7775 printk(KERN_CONT "\n");
7777 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7778 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7781 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7782 printk(KERN_ERR "ERROR: parent span is not a superset "
7783 "of domain->span\n");
7787 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7789 cpumask_var_t groupmask;
7793 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7797 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7799 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7800 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7805 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7812 free_cpumask_var(groupmask);
7814 #else /* !CONFIG_SCHED_DEBUG */
7815 # define sched_domain_debug(sd, cpu) do { } while (0)
7816 #endif /* CONFIG_SCHED_DEBUG */
7818 static int sd_degenerate(struct sched_domain *sd)
7820 if (cpumask_weight(sched_domain_span(sd)) == 1)
7823 /* Following flags need at least 2 groups */
7824 if (sd->flags & (SD_LOAD_BALANCE |
7825 SD_BALANCE_NEWIDLE |
7829 SD_SHARE_PKG_RESOURCES)) {
7830 if (sd->groups != sd->groups->next)
7834 /* Following flags don't use groups */
7835 if (sd->flags & (SD_WAKE_IDLE |
7844 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7846 unsigned long cflags = sd->flags, pflags = parent->flags;
7848 if (sd_degenerate(parent))
7851 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7854 /* Does parent contain flags not in child? */
7855 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7856 if (cflags & SD_WAKE_AFFINE)
7857 pflags &= ~SD_WAKE_BALANCE;
7858 /* Flags needing groups don't count if only 1 group in parent */
7859 if (parent->groups == parent->groups->next) {
7860 pflags &= ~(SD_LOAD_BALANCE |
7861 SD_BALANCE_NEWIDLE |
7865 SD_SHARE_PKG_RESOURCES);
7866 if (nr_node_ids == 1)
7867 pflags &= ~SD_SERIALIZE;
7869 if (~cflags & pflags)
7875 static void free_rootdomain(struct root_domain *rd)
7877 cpupri_cleanup(&rd->cpupri);
7879 free_cpumask_var(rd->rto_mask);
7880 free_cpumask_var(rd->online);
7881 free_cpumask_var(rd->span);
7885 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7887 struct root_domain *old_rd = NULL;
7888 unsigned long flags;
7890 spin_lock_irqsave(&rq->lock, flags);
7895 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7898 cpumask_clear_cpu(rq->cpu, old_rd->span);
7901 * If we dont want to free the old_rt yet then
7902 * set old_rd to NULL to skip the freeing later
7905 if (!atomic_dec_and_test(&old_rd->refcount))
7909 atomic_inc(&rd->refcount);
7912 cpumask_set_cpu(rq->cpu, rd->span);
7913 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7916 spin_unlock_irqrestore(&rq->lock, flags);
7919 free_rootdomain(old_rd);
7922 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7924 gfp_t gfp = GFP_KERNEL;
7926 memset(rd, 0, sizeof(*rd));
7931 if (!alloc_cpumask_var(&rd->span, gfp))
7933 if (!alloc_cpumask_var(&rd->online, gfp))
7935 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
7938 if (cpupri_init(&rd->cpupri, bootmem) != 0)
7943 free_cpumask_var(rd->rto_mask);
7945 free_cpumask_var(rd->online);
7947 free_cpumask_var(rd->span);
7952 static void init_defrootdomain(void)
7954 init_rootdomain(&def_root_domain, true);
7956 atomic_set(&def_root_domain.refcount, 1);
7959 static struct root_domain *alloc_rootdomain(void)
7961 struct root_domain *rd;
7963 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7967 if (init_rootdomain(rd, false) != 0) {
7976 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7977 * hold the hotplug lock.
7980 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7982 struct rq *rq = cpu_rq(cpu);
7983 struct sched_domain *tmp;
7985 /* Remove the sched domains which do not contribute to scheduling. */
7986 for (tmp = sd; tmp; ) {
7987 struct sched_domain *parent = tmp->parent;
7991 if (sd_parent_degenerate(tmp, parent)) {
7992 tmp->parent = parent->parent;
7994 parent->parent->child = tmp;
7999 if (sd && sd_degenerate(sd)) {
8005 sched_domain_debug(sd, cpu);
8007 rq_attach_root(rq, rd);
8008 rcu_assign_pointer(rq->sd, sd);
8011 /* cpus with isolated domains */
8012 static cpumask_var_t cpu_isolated_map;
8014 /* Setup the mask of cpus configured for isolated domains */
8015 static int __init isolated_cpu_setup(char *str)
8017 cpulist_parse(str, cpu_isolated_map);
8021 __setup("isolcpus=", isolated_cpu_setup);
8024 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8025 * to a function which identifies what group(along with sched group) a CPU
8026 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8027 * (due to the fact that we keep track of groups covered with a struct cpumask).
8029 * init_sched_build_groups will build a circular linked list of the groups
8030 * covered by the given span, and will set each group's ->cpumask correctly,
8031 * and ->cpu_power to 0.
8034 init_sched_build_groups(const struct cpumask *span,
8035 const struct cpumask *cpu_map,
8036 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8037 struct sched_group **sg,
8038 struct cpumask *tmpmask),
8039 struct cpumask *covered, struct cpumask *tmpmask)
8041 struct sched_group *first = NULL, *last = NULL;
8044 cpumask_clear(covered);
8046 for_each_cpu(i, span) {
8047 struct sched_group *sg;
8048 int group = group_fn(i, cpu_map, &sg, tmpmask);
8051 if (cpumask_test_cpu(i, covered))
8054 cpumask_clear(sched_group_cpus(sg));
8055 sg->__cpu_power = 0;
8057 for_each_cpu(j, span) {
8058 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8061 cpumask_set_cpu(j, covered);
8062 cpumask_set_cpu(j, sched_group_cpus(sg));
8073 #define SD_NODES_PER_DOMAIN 16
8078 * find_next_best_node - find the next node to include in a sched_domain
8079 * @node: node whose sched_domain we're building
8080 * @used_nodes: nodes already in the sched_domain
8082 * Find the next node to include in a given scheduling domain. Simply
8083 * finds the closest node not already in the @used_nodes map.
8085 * Should use nodemask_t.
8087 static int find_next_best_node(int node, nodemask_t *used_nodes)
8089 int i, n, val, min_val, best_node = 0;
8093 for (i = 0; i < nr_node_ids; i++) {
8094 /* Start at @node */
8095 n = (node + i) % nr_node_ids;
8097 if (!nr_cpus_node(n))
8100 /* Skip already used nodes */
8101 if (node_isset(n, *used_nodes))
8104 /* Simple min distance search */
8105 val = node_distance(node, n);
8107 if (val < min_val) {
8113 node_set(best_node, *used_nodes);
8118 * sched_domain_node_span - get a cpumask for a node's sched_domain
8119 * @node: node whose cpumask we're constructing
8120 * @span: resulting cpumask
8122 * Given a node, construct a good cpumask for its sched_domain to span. It
8123 * should be one that prevents unnecessary balancing, but also spreads tasks
8126 static void sched_domain_node_span(int node, struct cpumask *span)
8128 nodemask_t used_nodes;
8131 cpumask_clear(span);
8132 nodes_clear(used_nodes);
8134 cpumask_or(span, span, cpumask_of_node(node));
8135 node_set(node, used_nodes);
8137 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8138 int next_node = find_next_best_node(node, &used_nodes);
8140 cpumask_or(span, span, cpumask_of_node(next_node));
8143 #endif /* CONFIG_NUMA */
8145 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8148 * The cpus mask in sched_group and sched_domain hangs off the end.
8150 * ( See the the comments in include/linux/sched.h:struct sched_group
8151 * and struct sched_domain. )
8153 struct static_sched_group {
8154 struct sched_group sg;
8155 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8158 struct static_sched_domain {
8159 struct sched_domain sd;
8160 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8164 * SMT sched-domains:
8166 #ifdef CONFIG_SCHED_SMT
8167 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8168 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8171 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8172 struct sched_group **sg, struct cpumask *unused)
8175 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8178 #endif /* CONFIG_SCHED_SMT */
8181 * multi-core sched-domains:
8183 #ifdef CONFIG_SCHED_MC
8184 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8185 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8186 #endif /* CONFIG_SCHED_MC */
8188 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8190 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8191 struct sched_group **sg, struct cpumask *mask)
8195 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8196 group = cpumask_first(mask);
8198 *sg = &per_cpu(sched_group_core, group).sg;
8201 #elif defined(CONFIG_SCHED_MC)
8203 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8204 struct sched_group **sg, struct cpumask *unused)
8207 *sg = &per_cpu(sched_group_core, cpu).sg;
8212 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8213 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8216 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8217 struct sched_group **sg, struct cpumask *mask)
8220 #ifdef CONFIG_SCHED_MC
8221 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8222 group = cpumask_first(mask);
8223 #elif defined(CONFIG_SCHED_SMT)
8224 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8225 group = cpumask_first(mask);
8230 *sg = &per_cpu(sched_group_phys, group).sg;
8236 * The init_sched_build_groups can't handle what we want to do with node
8237 * groups, so roll our own. Now each node has its own list of groups which
8238 * gets dynamically allocated.
8240 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8241 static struct sched_group ***sched_group_nodes_bycpu;
8243 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8244 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8246 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8247 struct sched_group **sg,
8248 struct cpumask *nodemask)
8252 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8253 group = cpumask_first(nodemask);
8256 *sg = &per_cpu(sched_group_allnodes, group).sg;
8260 static void init_numa_sched_groups_power(struct sched_group *group_head)
8262 struct sched_group *sg = group_head;
8268 for_each_cpu(j, sched_group_cpus(sg)) {
8269 struct sched_domain *sd;
8271 sd = &per_cpu(phys_domains, j).sd;
8272 if (j != group_first_cpu(sd->groups)) {
8274 * Only add "power" once for each
8280 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
8283 } while (sg != group_head);
8285 #endif /* CONFIG_NUMA */
8288 /* Free memory allocated for various sched_group structures */
8289 static void free_sched_groups(const struct cpumask *cpu_map,
8290 struct cpumask *nodemask)
8294 for_each_cpu(cpu, cpu_map) {
8295 struct sched_group **sched_group_nodes
8296 = sched_group_nodes_bycpu[cpu];
8298 if (!sched_group_nodes)
8301 for (i = 0; i < nr_node_ids; i++) {
8302 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8304 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8305 if (cpumask_empty(nodemask))
8315 if (oldsg != sched_group_nodes[i])
8318 kfree(sched_group_nodes);
8319 sched_group_nodes_bycpu[cpu] = NULL;
8322 #else /* !CONFIG_NUMA */
8323 static void free_sched_groups(const struct cpumask *cpu_map,
8324 struct cpumask *nodemask)
8327 #endif /* CONFIG_NUMA */
8330 * Initialize sched groups cpu_power.
8332 * cpu_power indicates the capacity of sched group, which is used while
8333 * distributing the load between different sched groups in a sched domain.
8334 * Typically cpu_power for all the groups in a sched domain will be same unless
8335 * there are asymmetries in the topology. If there are asymmetries, group
8336 * having more cpu_power will pickup more load compared to the group having
8339 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
8340 * the maximum number of tasks a group can handle in the presence of other idle
8341 * or lightly loaded groups in the same sched domain.
8343 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8345 struct sched_domain *child;
8346 struct sched_group *group;
8348 WARN_ON(!sd || !sd->groups);
8350 if (cpu != group_first_cpu(sd->groups))
8355 sd->groups->__cpu_power = 0;
8358 * For perf policy, if the groups in child domain share resources
8359 * (for example cores sharing some portions of the cache hierarchy
8360 * or SMT), then set this domain groups cpu_power such that each group
8361 * can handle only one task, when there are other idle groups in the
8362 * same sched domain.
8364 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
8366 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
8367 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
8372 * add cpu_power of each child group to this groups cpu_power
8374 group = child->groups;
8376 sg_inc_cpu_power(sd->groups, group->__cpu_power);
8377 group = group->next;
8378 } while (group != child->groups);
8382 * Initializers for schedule domains
8383 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8386 #ifdef CONFIG_SCHED_DEBUG
8387 # define SD_INIT_NAME(sd, type) sd->name = #type
8389 # define SD_INIT_NAME(sd, type) do { } while (0)
8392 #define SD_INIT(sd, type) sd_init_##type(sd)
8394 #define SD_INIT_FUNC(type) \
8395 static noinline void sd_init_##type(struct sched_domain *sd) \
8397 memset(sd, 0, sizeof(*sd)); \
8398 *sd = SD_##type##_INIT; \
8399 sd->level = SD_LV_##type; \
8400 SD_INIT_NAME(sd, type); \
8405 SD_INIT_FUNC(ALLNODES)
8408 #ifdef CONFIG_SCHED_SMT
8409 SD_INIT_FUNC(SIBLING)
8411 #ifdef CONFIG_SCHED_MC
8415 static int default_relax_domain_level = -1;
8417 static int __init setup_relax_domain_level(char *str)
8421 val = simple_strtoul(str, NULL, 0);
8422 if (val < SD_LV_MAX)
8423 default_relax_domain_level = val;
8427 __setup("relax_domain_level=", setup_relax_domain_level);
8429 static void set_domain_attribute(struct sched_domain *sd,
8430 struct sched_domain_attr *attr)
8434 if (!attr || attr->relax_domain_level < 0) {
8435 if (default_relax_domain_level < 0)
8438 request = default_relax_domain_level;
8440 request = attr->relax_domain_level;
8441 if (request < sd->level) {
8442 /* turn off idle balance on this domain */
8443 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
8445 /* turn on idle balance on this domain */
8446 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
8451 * Build sched domains for a given set of cpus and attach the sched domains
8452 * to the individual cpus
8454 static int __build_sched_domains(const struct cpumask *cpu_map,
8455 struct sched_domain_attr *attr)
8457 int i, err = -ENOMEM;
8458 struct root_domain *rd;
8459 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
8462 cpumask_var_t domainspan, covered, notcovered;
8463 struct sched_group **sched_group_nodes = NULL;
8464 int sd_allnodes = 0;
8466 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
8468 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
8469 goto free_domainspan;
8470 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
8474 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
8475 goto free_notcovered;
8476 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
8478 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
8479 goto free_this_sibling_map;
8480 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
8481 goto free_this_core_map;
8482 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
8483 goto free_send_covered;
8487 * Allocate the per-node list of sched groups
8489 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
8491 if (!sched_group_nodes) {
8492 printk(KERN_WARNING "Can not alloc sched group node list\n");
8497 rd = alloc_rootdomain();
8499 printk(KERN_WARNING "Cannot alloc root domain\n");
8500 goto free_sched_groups;
8504 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
8508 * Set up domains for cpus specified by the cpu_map.
8510 for_each_cpu(i, cpu_map) {
8511 struct sched_domain *sd = NULL, *p;
8513 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
8516 if (cpumask_weight(cpu_map) >
8517 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
8518 sd = &per_cpu(allnodes_domains, i).sd;
8519 SD_INIT(sd, ALLNODES);
8520 set_domain_attribute(sd, attr);
8521 cpumask_copy(sched_domain_span(sd), cpu_map);
8522 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
8528 sd = &per_cpu(node_domains, i).sd;
8530 set_domain_attribute(sd, attr);
8531 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8535 cpumask_and(sched_domain_span(sd),
8536 sched_domain_span(sd), cpu_map);
8540 sd = &per_cpu(phys_domains, i).sd;
8542 set_domain_attribute(sd, attr);
8543 cpumask_copy(sched_domain_span(sd), nodemask);
8547 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
8549 #ifdef CONFIG_SCHED_MC
8551 sd = &per_cpu(core_domains, i).sd;
8553 set_domain_attribute(sd, attr);
8554 cpumask_and(sched_domain_span(sd), cpu_map,
8555 cpu_coregroup_mask(i));
8558 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
8561 #ifdef CONFIG_SCHED_SMT
8563 sd = &per_cpu(cpu_domains, i).sd;
8564 SD_INIT(sd, SIBLING);
8565 set_domain_attribute(sd, attr);
8566 cpumask_and(sched_domain_span(sd),
8567 topology_thread_cpumask(i), cpu_map);
8570 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
8574 #ifdef CONFIG_SCHED_SMT
8575 /* Set up CPU (sibling) groups */
8576 for_each_cpu(i, cpu_map) {
8577 cpumask_and(this_sibling_map,
8578 topology_thread_cpumask(i), cpu_map);
8579 if (i != cpumask_first(this_sibling_map))
8582 init_sched_build_groups(this_sibling_map, cpu_map,
8584 send_covered, tmpmask);
8588 #ifdef CONFIG_SCHED_MC
8589 /* Set up multi-core groups */
8590 for_each_cpu(i, cpu_map) {
8591 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
8592 if (i != cpumask_first(this_core_map))
8595 init_sched_build_groups(this_core_map, cpu_map,
8597 send_covered, tmpmask);
8601 /* Set up physical groups */
8602 for (i = 0; i < nr_node_ids; i++) {
8603 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8604 if (cpumask_empty(nodemask))
8607 init_sched_build_groups(nodemask, cpu_map,
8609 send_covered, tmpmask);
8613 /* Set up node groups */
8615 init_sched_build_groups(cpu_map, cpu_map,
8616 &cpu_to_allnodes_group,
8617 send_covered, tmpmask);
8620 for (i = 0; i < nr_node_ids; i++) {
8621 /* Set up node groups */
8622 struct sched_group *sg, *prev;
8625 cpumask_clear(covered);
8626 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8627 if (cpumask_empty(nodemask)) {
8628 sched_group_nodes[i] = NULL;
8632 sched_domain_node_span(i, domainspan);
8633 cpumask_and(domainspan, domainspan, cpu_map);
8635 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8638 printk(KERN_WARNING "Can not alloc domain group for "
8642 sched_group_nodes[i] = sg;
8643 for_each_cpu(j, nodemask) {
8644 struct sched_domain *sd;
8646 sd = &per_cpu(node_domains, j).sd;
8649 sg->__cpu_power = 0;
8650 cpumask_copy(sched_group_cpus(sg), nodemask);
8652 cpumask_or(covered, covered, nodemask);
8655 for (j = 0; j < nr_node_ids; j++) {
8656 int n = (i + j) % nr_node_ids;
8658 cpumask_complement(notcovered, covered);
8659 cpumask_and(tmpmask, notcovered, cpu_map);
8660 cpumask_and(tmpmask, tmpmask, domainspan);
8661 if (cpumask_empty(tmpmask))
8664 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
8665 if (cpumask_empty(tmpmask))
8668 sg = kmalloc_node(sizeof(struct sched_group) +
8673 "Can not alloc domain group for node %d\n", j);
8676 sg->__cpu_power = 0;
8677 cpumask_copy(sched_group_cpus(sg), tmpmask);
8678 sg->next = prev->next;
8679 cpumask_or(covered, covered, tmpmask);
8686 /* Calculate CPU power for physical packages and nodes */
8687 #ifdef CONFIG_SCHED_SMT
8688 for_each_cpu(i, cpu_map) {
8689 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
8691 init_sched_groups_power(i, sd);
8694 #ifdef CONFIG_SCHED_MC
8695 for_each_cpu(i, cpu_map) {
8696 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
8698 init_sched_groups_power(i, sd);
8702 for_each_cpu(i, cpu_map) {
8703 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
8705 init_sched_groups_power(i, sd);
8709 for (i = 0; i < nr_node_ids; i++)
8710 init_numa_sched_groups_power(sched_group_nodes[i]);
8713 struct sched_group *sg;
8715 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8717 init_numa_sched_groups_power(sg);
8721 /* Attach the domains */
8722 for_each_cpu(i, cpu_map) {
8723 struct sched_domain *sd;
8724 #ifdef CONFIG_SCHED_SMT
8725 sd = &per_cpu(cpu_domains, i).sd;
8726 #elif defined(CONFIG_SCHED_MC)
8727 sd = &per_cpu(core_domains, i).sd;
8729 sd = &per_cpu(phys_domains, i).sd;
8731 cpu_attach_domain(sd, rd, i);
8737 free_cpumask_var(tmpmask);
8739 free_cpumask_var(send_covered);
8741 free_cpumask_var(this_core_map);
8742 free_this_sibling_map:
8743 free_cpumask_var(this_sibling_map);
8745 free_cpumask_var(nodemask);
8748 free_cpumask_var(notcovered);
8750 free_cpumask_var(covered);
8752 free_cpumask_var(domainspan);
8759 kfree(sched_group_nodes);
8765 free_sched_groups(cpu_map, tmpmask);
8766 free_rootdomain(rd);
8771 static int build_sched_domains(const struct cpumask *cpu_map)
8773 return __build_sched_domains(cpu_map, NULL);
8776 static struct cpumask *doms_cur; /* current sched domains */
8777 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8778 static struct sched_domain_attr *dattr_cur;
8779 /* attribues of custom domains in 'doms_cur' */
8782 * Special case: If a kmalloc of a doms_cur partition (array of
8783 * cpumask) fails, then fallback to a single sched domain,
8784 * as determined by the single cpumask fallback_doms.
8786 static cpumask_var_t fallback_doms;
8789 * arch_update_cpu_topology lets virtualized architectures update the
8790 * cpu core maps. It is supposed to return 1 if the topology changed
8791 * or 0 if it stayed the same.
8793 int __attribute__((weak)) arch_update_cpu_topology(void)
8799 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8800 * For now this just excludes isolated cpus, but could be used to
8801 * exclude other special cases in the future.
8803 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8807 arch_update_cpu_topology();
8809 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8811 doms_cur = fallback_doms;
8812 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8814 err = build_sched_domains(doms_cur);
8815 register_sched_domain_sysctl();
8820 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8821 struct cpumask *tmpmask)
8823 free_sched_groups(cpu_map, tmpmask);
8827 * Detach sched domains from a group of cpus specified in cpu_map
8828 * These cpus will now be attached to the NULL domain
8830 static void detach_destroy_domains(const struct cpumask *cpu_map)
8832 /* Save because hotplug lock held. */
8833 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8836 for_each_cpu(i, cpu_map)
8837 cpu_attach_domain(NULL, &def_root_domain, i);
8838 synchronize_sched();
8839 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8842 /* handle null as "default" */
8843 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8844 struct sched_domain_attr *new, int idx_new)
8846 struct sched_domain_attr tmp;
8853 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8854 new ? (new + idx_new) : &tmp,
8855 sizeof(struct sched_domain_attr));
8859 * Partition sched domains as specified by the 'ndoms_new'
8860 * cpumasks in the array doms_new[] of cpumasks. This compares
8861 * doms_new[] to the current sched domain partitioning, doms_cur[].
8862 * It destroys each deleted domain and builds each new domain.
8864 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8865 * The masks don't intersect (don't overlap.) We should setup one
8866 * sched domain for each mask. CPUs not in any of the cpumasks will
8867 * not be load balanced. If the same cpumask appears both in the
8868 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8871 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8872 * ownership of it and will kfree it when done with it. If the caller
8873 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8874 * ndoms_new == 1, and partition_sched_domains() will fallback to
8875 * the single partition 'fallback_doms', it also forces the domains
8878 * If doms_new == NULL it will be replaced with cpu_online_mask.
8879 * ndoms_new == 0 is a special case for destroying existing domains,
8880 * and it will not create the default domain.
8882 * Call with hotplug lock held
8884 /* FIXME: Change to struct cpumask *doms_new[] */
8885 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8886 struct sched_domain_attr *dattr_new)
8891 mutex_lock(&sched_domains_mutex);
8893 /* always unregister in case we don't destroy any domains */
8894 unregister_sched_domain_sysctl();
8896 /* Let architecture update cpu core mappings. */
8897 new_topology = arch_update_cpu_topology();
8899 n = doms_new ? ndoms_new : 0;
8901 /* Destroy deleted domains */
8902 for (i = 0; i < ndoms_cur; i++) {
8903 for (j = 0; j < n && !new_topology; j++) {
8904 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8905 && dattrs_equal(dattr_cur, i, dattr_new, j))
8908 /* no match - a current sched domain not in new doms_new[] */
8909 detach_destroy_domains(doms_cur + i);
8914 if (doms_new == NULL) {
8916 doms_new = fallback_doms;
8917 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8918 WARN_ON_ONCE(dattr_new);
8921 /* Build new domains */
8922 for (i = 0; i < ndoms_new; i++) {
8923 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8924 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8925 && dattrs_equal(dattr_new, i, dattr_cur, j))
8928 /* no match - add a new doms_new */
8929 __build_sched_domains(doms_new + i,
8930 dattr_new ? dattr_new + i : NULL);
8935 /* Remember the new sched domains */
8936 if (doms_cur != fallback_doms)
8938 kfree(dattr_cur); /* kfree(NULL) is safe */
8939 doms_cur = doms_new;
8940 dattr_cur = dattr_new;
8941 ndoms_cur = ndoms_new;
8943 register_sched_domain_sysctl();
8945 mutex_unlock(&sched_domains_mutex);
8948 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8949 static void arch_reinit_sched_domains(void)
8953 /* Destroy domains first to force the rebuild */
8954 partition_sched_domains(0, NULL, NULL);
8956 rebuild_sched_domains();
8960 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8962 unsigned int level = 0;
8964 if (sscanf(buf, "%u", &level) != 1)
8968 * level is always be positive so don't check for
8969 * level < POWERSAVINGS_BALANCE_NONE which is 0
8970 * What happens on 0 or 1 byte write,
8971 * need to check for count as well?
8974 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8978 sched_smt_power_savings = level;
8980 sched_mc_power_savings = level;
8982 arch_reinit_sched_domains();
8987 #ifdef CONFIG_SCHED_MC
8988 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8991 return sprintf(page, "%u\n", sched_mc_power_savings);
8993 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8994 const char *buf, size_t count)
8996 return sched_power_savings_store(buf, count, 0);
8998 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8999 sched_mc_power_savings_show,
9000 sched_mc_power_savings_store);
9003 #ifdef CONFIG_SCHED_SMT
9004 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9007 return sprintf(page, "%u\n", sched_smt_power_savings);
9009 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9010 const char *buf, size_t count)
9012 return sched_power_savings_store(buf, count, 1);
9014 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9015 sched_smt_power_savings_show,
9016 sched_smt_power_savings_store);
9019 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9023 #ifdef CONFIG_SCHED_SMT
9025 err = sysfs_create_file(&cls->kset.kobj,
9026 &attr_sched_smt_power_savings.attr);
9028 #ifdef CONFIG_SCHED_MC
9029 if (!err && mc_capable())
9030 err = sysfs_create_file(&cls->kset.kobj,
9031 &attr_sched_mc_power_savings.attr);
9035 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9037 #ifndef CONFIG_CPUSETS
9039 * Add online and remove offline CPUs from the scheduler domains.
9040 * When cpusets are enabled they take over this function.
9042 static int update_sched_domains(struct notifier_block *nfb,
9043 unsigned long action, void *hcpu)
9047 case CPU_ONLINE_FROZEN:
9049 case CPU_DEAD_FROZEN:
9050 partition_sched_domains(1, NULL, NULL);
9059 static int update_runtime(struct notifier_block *nfb,
9060 unsigned long action, void *hcpu)
9062 int cpu = (int)(long)hcpu;
9065 case CPU_DOWN_PREPARE:
9066 case CPU_DOWN_PREPARE_FROZEN:
9067 disable_runtime(cpu_rq(cpu));
9070 case CPU_DOWN_FAILED:
9071 case CPU_DOWN_FAILED_FROZEN:
9073 case CPU_ONLINE_FROZEN:
9074 enable_runtime(cpu_rq(cpu));
9082 void __init sched_init_smp(void)
9084 cpumask_var_t non_isolated_cpus;
9086 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9088 #if defined(CONFIG_NUMA)
9089 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9091 BUG_ON(sched_group_nodes_bycpu == NULL);
9094 mutex_lock(&sched_domains_mutex);
9095 arch_init_sched_domains(cpu_online_mask);
9096 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9097 if (cpumask_empty(non_isolated_cpus))
9098 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9099 mutex_unlock(&sched_domains_mutex);
9102 #ifndef CONFIG_CPUSETS
9103 /* XXX: Theoretical race here - CPU may be hotplugged now */
9104 hotcpu_notifier(update_sched_domains, 0);
9107 /* RT runtime code needs to handle some hotplug events */
9108 hotcpu_notifier(update_runtime, 0);
9112 /* Move init over to a non-isolated CPU */
9113 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9115 sched_init_granularity();
9116 free_cpumask_var(non_isolated_cpus);
9118 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9119 init_sched_rt_class();
9122 void __init sched_init_smp(void)
9124 sched_init_granularity();
9126 #endif /* CONFIG_SMP */
9128 const_debug unsigned int sysctl_timer_migration = 1;
9130 int in_sched_functions(unsigned long addr)
9132 return in_lock_functions(addr) ||
9133 (addr >= (unsigned long)__sched_text_start
9134 && addr < (unsigned long)__sched_text_end);
9137 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9139 cfs_rq->tasks_timeline = RB_ROOT;
9140 INIT_LIST_HEAD(&cfs_rq->tasks);
9141 #ifdef CONFIG_FAIR_GROUP_SCHED
9144 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9147 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9149 struct rt_prio_array *array;
9152 array = &rt_rq->active;
9153 for (i = 0; i < MAX_RT_PRIO; i++) {
9154 INIT_LIST_HEAD(array->queue + i);
9155 __clear_bit(i, array->bitmap);
9157 /* delimiter for bitsearch: */
9158 __set_bit(MAX_RT_PRIO, array->bitmap);
9160 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9161 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9163 rt_rq->highest_prio.next = MAX_RT_PRIO;
9167 rt_rq->rt_nr_migratory = 0;
9168 rt_rq->overloaded = 0;
9169 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9173 rt_rq->rt_throttled = 0;
9174 rt_rq->rt_runtime = 0;
9175 spin_lock_init(&rt_rq->rt_runtime_lock);
9177 #ifdef CONFIG_RT_GROUP_SCHED
9178 rt_rq->rt_nr_boosted = 0;
9183 #ifdef CONFIG_FAIR_GROUP_SCHED
9184 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9185 struct sched_entity *se, int cpu, int add,
9186 struct sched_entity *parent)
9188 struct rq *rq = cpu_rq(cpu);
9189 tg->cfs_rq[cpu] = cfs_rq;
9190 init_cfs_rq(cfs_rq, rq);
9193 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9196 /* se could be NULL for init_task_group */
9201 se->cfs_rq = &rq->cfs;
9203 se->cfs_rq = parent->my_q;
9206 se->load.weight = tg->shares;
9207 se->load.inv_weight = 0;
9208 se->parent = parent;
9212 #ifdef CONFIG_RT_GROUP_SCHED
9213 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9214 struct sched_rt_entity *rt_se, int cpu, int add,
9215 struct sched_rt_entity *parent)
9217 struct rq *rq = cpu_rq(cpu);
9219 tg->rt_rq[cpu] = rt_rq;
9220 init_rt_rq(rt_rq, rq);
9222 rt_rq->rt_se = rt_se;
9223 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9225 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9227 tg->rt_se[cpu] = rt_se;
9232 rt_se->rt_rq = &rq->rt;
9234 rt_se->rt_rq = parent->my_q;
9236 rt_se->my_q = rt_rq;
9237 rt_se->parent = parent;
9238 INIT_LIST_HEAD(&rt_se->run_list);
9242 void __init sched_init(void)
9245 unsigned long alloc_size = 0, ptr;
9247 #ifdef CONFIG_FAIR_GROUP_SCHED
9248 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9250 #ifdef CONFIG_RT_GROUP_SCHED
9251 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9253 #ifdef CONFIG_USER_SCHED
9256 #ifdef CONFIG_CPUMASK_OFFSTACK
9257 alloc_size += num_possible_cpus() * cpumask_size();
9260 * As sched_init() is called before page_alloc is setup,
9261 * we use alloc_bootmem().
9264 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9266 #ifdef CONFIG_FAIR_GROUP_SCHED
9267 init_task_group.se = (struct sched_entity **)ptr;
9268 ptr += nr_cpu_ids * sizeof(void **);
9270 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9271 ptr += nr_cpu_ids * sizeof(void **);
9273 #ifdef CONFIG_USER_SCHED
9274 root_task_group.se = (struct sched_entity **)ptr;
9275 ptr += nr_cpu_ids * sizeof(void **);
9277 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9278 ptr += nr_cpu_ids * sizeof(void **);
9279 #endif /* CONFIG_USER_SCHED */
9280 #endif /* CONFIG_FAIR_GROUP_SCHED */
9281 #ifdef CONFIG_RT_GROUP_SCHED
9282 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9283 ptr += nr_cpu_ids * sizeof(void **);
9285 init_task_group.rt_rq = (struct rt_rq **)ptr;
9286 ptr += nr_cpu_ids * sizeof(void **);
9288 #ifdef CONFIG_USER_SCHED
9289 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9290 ptr += nr_cpu_ids * sizeof(void **);
9292 root_task_group.rt_rq = (struct rt_rq **)ptr;
9293 ptr += nr_cpu_ids * sizeof(void **);
9294 #endif /* CONFIG_USER_SCHED */
9295 #endif /* CONFIG_RT_GROUP_SCHED */
9296 #ifdef CONFIG_CPUMASK_OFFSTACK
9297 for_each_possible_cpu(i) {
9298 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9299 ptr += cpumask_size();
9301 #endif /* CONFIG_CPUMASK_OFFSTACK */
9305 init_defrootdomain();
9308 init_rt_bandwidth(&def_rt_bandwidth,
9309 global_rt_period(), global_rt_runtime());
9311 #ifdef CONFIG_RT_GROUP_SCHED
9312 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9313 global_rt_period(), global_rt_runtime());
9314 #ifdef CONFIG_USER_SCHED
9315 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9316 global_rt_period(), RUNTIME_INF);
9317 #endif /* CONFIG_USER_SCHED */
9318 #endif /* CONFIG_RT_GROUP_SCHED */
9320 #ifdef CONFIG_GROUP_SCHED
9321 list_add(&init_task_group.list, &task_groups);
9322 INIT_LIST_HEAD(&init_task_group.children);
9324 #ifdef CONFIG_USER_SCHED
9325 INIT_LIST_HEAD(&root_task_group.children);
9326 init_task_group.parent = &root_task_group;
9327 list_add(&init_task_group.siblings, &root_task_group.children);
9328 #endif /* CONFIG_USER_SCHED */
9329 #endif /* CONFIG_GROUP_SCHED */
9331 for_each_possible_cpu(i) {
9335 spin_lock_init(&rq->lock);
9337 rq->calc_load_active = 0;
9338 rq->calc_load_update = jiffies + LOAD_FREQ;
9339 init_cfs_rq(&rq->cfs, rq);
9340 init_rt_rq(&rq->rt, rq);
9341 #ifdef CONFIG_FAIR_GROUP_SCHED
9342 init_task_group.shares = init_task_group_load;
9343 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9344 #ifdef CONFIG_CGROUP_SCHED
9346 * How much cpu bandwidth does init_task_group get?
9348 * In case of task-groups formed thr' the cgroup filesystem, it
9349 * gets 100% of the cpu resources in the system. This overall
9350 * system cpu resource is divided among the tasks of
9351 * init_task_group and its child task-groups in a fair manner,
9352 * based on each entity's (task or task-group's) weight
9353 * (se->load.weight).
9355 * In other words, if init_task_group has 10 tasks of weight
9356 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9357 * then A0's share of the cpu resource is:
9359 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9361 * We achieve this by letting init_task_group's tasks sit
9362 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9364 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9365 #elif defined CONFIG_USER_SCHED
9366 root_task_group.shares = NICE_0_LOAD;
9367 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9369 * In case of task-groups formed thr' the user id of tasks,
9370 * init_task_group represents tasks belonging to root user.
9371 * Hence it forms a sibling of all subsequent groups formed.
9372 * In this case, init_task_group gets only a fraction of overall
9373 * system cpu resource, based on the weight assigned to root
9374 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9375 * by letting tasks of init_task_group sit in a separate cfs_rq
9376 * (init_cfs_rq) and having one entity represent this group of
9377 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9379 init_tg_cfs_entry(&init_task_group,
9380 &per_cpu(init_cfs_rq, i),
9381 &per_cpu(init_sched_entity, i), i, 1,
9382 root_task_group.se[i]);
9385 #endif /* CONFIG_FAIR_GROUP_SCHED */
9387 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9388 #ifdef CONFIG_RT_GROUP_SCHED
9389 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9390 #ifdef CONFIG_CGROUP_SCHED
9391 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9392 #elif defined CONFIG_USER_SCHED
9393 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9394 init_tg_rt_entry(&init_task_group,
9395 &per_cpu(init_rt_rq, i),
9396 &per_cpu(init_sched_rt_entity, i), i, 1,
9397 root_task_group.rt_se[i]);
9401 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9402 rq->cpu_load[j] = 0;
9406 rq->active_balance = 0;
9407 rq->next_balance = jiffies;
9411 rq->migration_thread = NULL;
9412 INIT_LIST_HEAD(&rq->migration_queue);
9413 rq_attach_root(rq, &def_root_domain);
9416 atomic_set(&rq->nr_iowait, 0);
9419 set_load_weight(&init_task);
9421 #ifdef CONFIG_PREEMPT_NOTIFIERS
9422 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9426 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9429 #ifdef CONFIG_RT_MUTEXES
9430 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9434 * The boot idle thread does lazy MMU switching as well:
9436 atomic_inc(&init_mm.mm_count);
9437 enter_lazy_tlb(&init_mm, current);
9440 * Make us the idle thread. Technically, schedule() should not be
9441 * called from this thread, however somewhere below it might be,
9442 * but because we are the idle thread, we just pick up running again
9443 * when this runqueue becomes "idle".
9445 init_idle(current, smp_processor_id());
9447 calc_load_update = jiffies + LOAD_FREQ;
9450 * During early bootup we pretend to be a normal task:
9452 current->sched_class = &fair_sched_class;
9454 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9455 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9458 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9459 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9461 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9464 perf_counter_init();
9466 scheduler_running = 1;
9469 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9470 static inline int preempt_count_equals(int preempt_offset)
9472 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9474 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9477 void __might_sleep(char *file, int line, int preempt_offset)
9480 static unsigned long prev_jiffy; /* ratelimiting */
9482 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9483 system_state != SYSTEM_RUNNING || oops_in_progress)
9485 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9487 prev_jiffy = jiffies;
9490 "BUG: sleeping function called from invalid context at %s:%d\n",
9493 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9494 in_atomic(), irqs_disabled(),
9495 current->pid, current->comm);
9497 debug_show_held_locks(current);
9498 if (irqs_disabled())
9499 print_irqtrace_events(current);
9503 EXPORT_SYMBOL(__might_sleep);
9506 #ifdef CONFIG_MAGIC_SYSRQ
9507 static void normalize_task(struct rq *rq, struct task_struct *p)
9511 update_rq_clock(rq);
9512 on_rq = p->se.on_rq;
9514 deactivate_task(rq, p, 0);
9515 __setscheduler(rq, p, SCHED_NORMAL, 0);
9517 activate_task(rq, p, 0);
9518 resched_task(rq->curr);
9522 void normalize_rt_tasks(void)
9524 struct task_struct *g, *p;
9525 unsigned long flags;
9528 read_lock_irqsave(&tasklist_lock, flags);
9529 do_each_thread(g, p) {
9531 * Only normalize user tasks:
9536 p->se.exec_start = 0;
9537 #ifdef CONFIG_SCHEDSTATS
9538 p->se.wait_start = 0;
9539 p->se.sleep_start = 0;
9540 p->se.block_start = 0;
9545 * Renice negative nice level userspace
9548 if (TASK_NICE(p) < 0 && p->mm)
9549 set_user_nice(p, 0);
9553 spin_lock(&p->pi_lock);
9554 rq = __task_rq_lock(p);
9556 normalize_task(rq, p);
9558 __task_rq_unlock(rq);
9559 spin_unlock(&p->pi_lock);
9560 } while_each_thread(g, p);
9562 read_unlock_irqrestore(&tasklist_lock, flags);
9565 #endif /* CONFIG_MAGIC_SYSRQ */
9569 * These functions are only useful for the IA64 MCA handling.
9571 * They can only be called when the whole system has been
9572 * stopped - every CPU needs to be quiescent, and no scheduling
9573 * activity can take place. Using them for anything else would
9574 * be a serious bug, and as a result, they aren't even visible
9575 * under any other configuration.
9579 * curr_task - return the current task for a given cpu.
9580 * @cpu: the processor in question.
9582 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9584 struct task_struct *curr_task(int cpu)
9586 return cpu_curr(cpu);
9590 * set_curr_task - set the current task for a given cpu.
9591 * @cpu: the processor in question.
9592 * @p: the task pointer to set.
9594 * Description: This function must only be used when non-maskable interrupts
9595 * are serviced on a separate stack. It allows the architecture to switch the
9596 * notion of the current task on a cpu in a non-blocking manner. This function
9597 * must be called with all CPU's synchronized, and interrupts disabled, the
9598 * and caller must save the original value of the current task (see
9599 * curr_task() above) and restore that value before reenabling interrupts and
9600 * re-starting the system.
9602 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9604 void set_curr_task(int cpu, struct task_struct *p)
9611 #ifdef CONFIG_FAIR_GROUP_SCHED
9612 static void free_fair_sched_group(struct task_group *tg)
9616 for_each_possible_cpu(i) {
9618 kfree(tg->cfs_rq[i]);
9628 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9630 struct cfs_rq *cfs_rq;
9631 struct sched_entity *se;
9635 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9638 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9642 tg->shares = NICE_0_LOAD;
9644 for_each_possible_cpu(i) {
9647 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9648 GFP_KERNEL, cpu_to_node(i));
9652 se = kzalloc_node(sizeof(struct sched_entity),
9653 GFP_KERNEL, cpu_to_node(i));
9657 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9666 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9668 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9669 &cpu_rq(cpu)->leaf_cfs_rq_list);
9672 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9674 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9676 #else /* !CONFG_FAIR_GROUP_SCHED */
9677 static inline void free_fair_sched_group(struct task_group *tg)
9682 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9687 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9691 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9694 #endif /* CONFIG_FAIR_GROUP_SCHED */
9696 #ifdef CONFIG_RT_GROUP_SCHED
9697 static void free_rt_sched_group(struct task_group *tg)
9701 destroy_rt_bandwidth(&tg->rt_bandwidth);
9703 for_each_possible_cpu(i) {
9705 kfree(tg->rt_rq[i]);
9707 kfree(tg->rt_se[i]);
9715 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9717 struct rt_rq *rt_rq;
9718 struct sched_rt_entity *rt_se;
9722 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9725 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9729 init_rt_bandwidth(&tg->rt_bandwidth,
9730 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9732 for_each_possible_cpu(i) {
9735 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9736 GFP_KERNEL, cpu_to_node(i));
9740 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9741 GFP_KERNEL, cpu_to_node(i));
9745 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9754 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9756 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9757 &cpu_rq(cpu)->leaf_rt_rq_list);
9760 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9762 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9764 #else /* !CONFIG_RT_GROUP_SCHED */
9765 static inline void free_rt_sched_group(struct task_group *tg)
9770 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9775 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9779 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9782 #endif /* CONFIG_RT_GROUP_SCHED */
9784 #ifdef CONFIG_GROUP_SCHED
9785 static void free_sched_group(struct task_group *tg)
9787 free_fair_sched_group(tg);
9788 free_rt_sched_group(tg);
9792 /* allocate runqueue etc for a new task group */
9793 struct task_group *sched_create_group(struct task_group *parent)
9795 struct task_group *tg;
9796 unsigned long flags;
9799 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9801 return ERR_PTR(-ENOMEM);
9803 if (!alloc_fair_sched_group(tg, parent))
9806 if (!alloc_rt_sched_group(tg, parent))
9809 spin_lock_irqsave(&task_group_lock, flags);
9810 for_each_possible_cpu(i) {
9811 register_fair_sched_group(tg, i);
9812 register_rt_sched_group(tg, i);
9814 list_add_rcu(&tg->list, &task_groups);
9816 WARN_ON(!parent); /* root should already exist */
9818 tg->parent = parent;
9819 INIT_LIST_HEAD(&tg->children);
9820 list_add_rcu(&tg->siblings, &parent->children);
9821 spin_unlock_irqrestore(&task_group_lock, flags);
9826 free_sched_group(tg);
9827 return ERR_PTR(-ENOMEM);
9830 /* rcu callback to free various structures associated with a task group */
9831 static void free_sched_group_rcu(struct rcu_head *rhp)
9833 /* now it should be safe to free those cfs_rqs */
9834 free_sched_group(container_of(rhp, struct task_group, rcu));
9837 /* Destroy runqueue etc associated with a task group */
9838 void sched_destroy_group(struct task_group *tg)
9840 unsigned long flags;
9843 spin_lock_irqsave(&task_group_lock, flags);
9844 for_each_possible_cpu(i) {
9845 unregister_fair_sched_group(tg, i);
9846 unregister_rt_sched_group(tg, i);
9848 list_del_rcu(&tg->list);
9849 list_del_rcu(&tg->siblings);
9850 spin_unlock_irqrestore(&task_group_lock, flags);
9852 /* wait for possible concurrent references to cfs_rqs complete */
9853 call_rcu(&tg->rcu, free_sched_group_rcu);
9856 /* change task's runqueue when it moves between groups.
9857 * The caller of this function should have put the task in its new group
9858 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9859 * reflect its new group.
9861 void sched_move_task(struct task_struct *tsk)
9864 unsigned long flags;
9867 rq = task_rq_lock(tsk, &flags);
9869 update_rq_clock(rq);
9871 running = task_current(rq, tsk);
9872 on_rq = tsk->se.on_rq;
9875 dequeue_task(rq, tsk, 0);
9876 if (unlikely(running))
9877 tsk->sched_class->put_prev_task(rq, tsk);
9879 set_task_rq(tsk, task_cpu(tsk));
9881 #ifdef CONFIG_FAIR_GROUP_SCHED
9882 if (tsk->sched_class->moved_group)
9883 tsk->sched_class->moved_group(tsk);
9886 if (unlikely(running))
9887 tsk->sched_class->set_curr_task(rq);
9889 enqueue_task(rq, tsk, 0);
9891 task_rq_unlock(rq, &flags);
9893 #endif /* CONFIG_GROUP_SCHED */
9895 #ifdef CONFIG_FAIR_GROUP_SCHED
9896 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9898 struct cfs_rq *cfs_rq = se->cfs_rq;
9903 dequeue_entity(cfs_rq, se, 0);
9905 se->load.weight = shares;
9906 se->load.inv_weight = 0;
9909 enqueue_entity(cfs_rq, se, 0);
9912 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9914 struct cfs_rq *cfs_rq = se->cfs_rq;
9915 struct rq *rq = cfs_rq->rq;
9916 unsigned long flags;
9918 spin_lock_irqsave(&rq->lock, flags);
9919 __set_se_shares(se, shares);
9920 spin_unlock_irqrestore(&rq->lock, flags);
9923 static DEFINE_MUTEX(shares_mutex);
9925 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9928 unsigned long flags;
9931 * We can't change the weight of the root cgroup.
9936 if (shares < MIN_SHARES)
9937 shares = MIN_SHARES;
9938 else if (shares > MAX_SHARES)
9939 shares = MAX_SHARES;
9941 mutex_lock(&shares_mutex);
9942 if (tg->shares == shares)
9945 spin_lock_irqsave(&task_group_lock, flags);
9946 for_each_possible_cpu(i)
9947 unregister_fair_sched_group(tg, i);
9948 list_del_rcu(&tg->siblings);
9949 spin_unlock_irqrestore(&task_group_lock, flags);
9951 /* wait for any ongoing reference to this group to finish */
9952 synchronize_sched();
9955 * Now we are free to modify the group's share on each cpu
9956 * w/o tripping rebalance_share or load_balance_fair.
9958 tg->shares = shares;
9959 for_each_possible_cpu(i) {
9963 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9964 set_se_shares(tg->se[i], shares);
9968 * Enable load balance activity on this group, by inserting it back on
9969 * each cpu's rq->leaf_cfs_rq_list.
9971 spin_lock_irqsave(&task_group_lock, flags);
9972 for_each_possible_cpu(i)
9973 register_fair_sched_group(tg, i);
9974 list_add_rcu(&tg->siblings, &tg->parent->children);
9975 spin_unlock_irqrestore(&task_group_lock, flags);
9977 mutex_unlock(&shares_mutex);
9981 unsigned long sched_group_shares(struct task_group *tg)
9987 #ifdef CONFIG_RT_GROUP_SCHED
9989 * Ensure that the real time constraints are schedulable.
9991 static DEFINE_MUTEX(rt_constraints_mutex);
9993 static unsigned long to_ratio(u64 period, u64 runtime)
9995 if (runtime == RUNTIME_INF)
9998 return div64_u64(runtime << 20, period);
10001 /* Must be called with tasklist_lock held */
10002 static inline int tg_has_rt_tasks(struct task_group *tg)
10004 struct task_struct *g, *p;
10006 do_each_thread(g, p) {
10007 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10009 } while_each_thread(g, p);
10014 struct rt_schedulable_data {
10015 struct task_group *tg;
10020 static int tg_schedulable(struct task_group *tg, void *data)
10022 struct rt_schedulable_data *d = data;
10023 struct task_group *child;
10024 unsigned long total, sum = 0;
10025 u64 period, runtime;
10027 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10028 runtime = tg->rt_bandwidth.rt_runtime;
10031 period = d->rt_period;
10032 runtime = d->rt_runtime;
10035 #ifdef CONFIG_USER_SCHED
10036 if (tg == &root_task_group) {
10037 period = global_rt_period();
10038 runtime = global_rt_runtime();
10043 * Cannot have more runtime than the period.
10045 if (runtime > period && runtime != RUNTIME_INF)
10049 * Ensure we don't starve existing RT tasks.
10051 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10054 total = to_ratio(period, runtime);
10057 * Nobody can have more than the global setting allows.
10059 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10063 * The sum of our children's runtime should not exceed our own.
10065 list_for_each_entry_rcu(child, &tg->children, siblings) {
10066 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10067 runtime = child->rt_bandwidth.rt_runtime;
10069 if (child == d->tg) {
10070 period = d->rt_period;
10071 runtime = d->rt_runtime;
10074 sum += to_ratio(period, runtime);
10083 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10085 struct rt_schedulable_data data = {
10087 .rt_period = period,
10088 .rt_runtime = runtime,
10091 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10094 static int tg_set_bandwidth(struct task_group *tg,
10095 u64 rt_period, u64 rt_runtime)
10099 mutex_lock(&rt_constraints_mutex);
10100 read_lock(&tasklist_lock);
10101 err = __rt_schedulable(tg, rt_period, rt_runtime);
10105 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10106 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10107 tg->rt_bandwidth.rt_runtime = rt_runtime;
10109 for_each_possible_cpu(i) {
10110 struct rt_rq *rt_rq = tg->rt_rq[i];
10112 spin_lock(&rt_rq->rt_runtime_lock);
10113 rt_rq->rt_runtime = rt_runtime;
10114 spin_unlock(&rt_rq->rt_runtime_lock);
10116 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10118 read_unlock(&tasklist_lock);
10119 mutex_unlock(&rt_constraints_mutex);
10124 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10126 u64 rt_runtime, rt_period;
10128 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10129 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10130 if (rt_runtime_us < 0)
10131 rt_runtime = RUNTIME_INF;
10133 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10136 long sched_group_rt_runtime(struct task_group *tg)
10140 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10143 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10144 do_div(rt_runtime_us, NSEC_PER_USEC);
10145 return rt_runtime_us;
10148 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10150 u64 rt_runtime, rt_period;
10152 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10153 rt_runtime = tg->rt_bandwidth.rt_runtime;
10155 if (rt_period == 0)
10158 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10161 long sched_group_rt_period(struct task_group *tg)
10165 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10166 do_div(rt_period_us, NSEC_PER_USEC);
10167 return rt_period_us;
10170 static int sched_rt_global_constraints(void)
10172 u64 runtime, period;
10175 if (sysctl_sched_rt_period <= 0)
10178 runtime = global_rt_runtime();
10179 period = global_rt_period();
10182 * Sanity check on the sysctl variables.
10184 if (runtime > period && runtime != RUNTIME_INF)
10187 mutex_lock(&rt_constraints_mutex);
10188 read_lock(&tasklist_lock);
10189 ret = __rt_schedulable(NULL, 0, 0);
10190 read_unlock(&tasklist_lock);
10191 mutex_unlock(&rt_constraints_mutex);
10196 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10198 /* Don't accept realtime tasks when there is no way for them to run */
10199 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10205 #else /* !CONFIG_RT_GROUP_SCHED */
10206 static int sched_rt_global_constraints(void)
10208 unsigned long flags;
10211 if (sysctl_sched_rt_period <= 0)
10215 * There's always some RT tasks in the root group
10216 * -- migration, kstopmachine etc..
10218 if (sysctl_sched_rt_runtime == 0)
10221 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10222 for_each_possible_cpu(i) {
10223 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10225 spin_lock(&rt_rq->rt_runtime_lock);
10226 rt_rq->rt_runtime = global_rt_runtime();
10227 spin_unlock(&rt_rq->rt_runtime_lock);
10229 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10233 #endif /* CONFIG_RT_GROUP_SCHED */
10235 int sched_rt_handler(struct ctl_table *table, int write,
10236 struct file *filp, void __user *buffer, size_t *lenp,
10240 int old_period, old_runtime;
10241 static DEFINE_MUTEX(mutex);
10243 mutex_lock(&mutex);
10244 old_period = sysctl_sched_rt_period;
10245 old_runtime = sysctl_sched_rt_runtime;
10247 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
10249 if (!ret && write) {
10250 ret = sched_rt_global_constraints();
10252 sysctl_sched_rt_period = old_period;
10253 sysctl_sched_rt_runtime = old_runtime;
10255 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10256 def_rt_bandwidth.rt_period =
10257 ns_to_ktime(global_rt_period());
10260 mutex_unlock(&mutex);
10265 #ifdef CONFIG_CGROUP_SCHED
10267 /* return corresponding task_group object of a cgroup */
10268 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10270 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10271 struct task_group, css);
10274 static struct cgroup_subsys_state *
10275 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10277 struct task_group *tg, *parent;
10279 if (!cgrp->parent) {
10280 /* This is early initialization for the top cgroup */
10281 return &init_task_group.css;
10284 parent = cgroup_tg(cgrp->parent);
10285 tg = sched_create_group(parent);
10287 return ERR_PTR(-ENOMEM);
10293 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10295 struct task_group *tg = cgroup_tg(cgrp);
10297 sched_destroy_group(tg);
10301 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10302 struct task_struct *tsk)
10304 #ifdef CONFIG_RT_GROUP_SCHED
10305 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10308 /* We don't support RT-tasks being in separate groups */
10309 if (tsk->sched_class != &fair_sched_class)
10317 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10318 struct cgroup *old_cont, struct task_struct *tsk)
10320 sched_move_task(tsk);
10323 #ifdef CONFIG_FAIR_GROUP_SCHED
10324 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10327 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10330 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10332 struct task_group *tg = cgroup_tg(cgrp);
10334 return (u64) tg->shares;
10336 #endif /* CONFIG_FAIR_GROUP_SCHED */
10338 #ifdef CONFIG_RT_GROUP_SCHED
10339 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10342 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10345 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10347 return sched_group_rt_runtime(cgroup_tg(cgrp));
10350 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10353 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10356 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10358 return sched_group_rt_period(cgroup_tg(cgrp));
10360 #endif /* CONFIG_RT_GROUP_SCHED */
10362 static struct cftype cpu_files[] = {
10363 #ifdef CONFIG_FAIR_GROUP_SCHED
10366 .read_u64 = cpu_shares_read_u64,
10367 .write_u64 = cpu_shares_write_u64,
10370 #ifdef CONFIG_RT_GROUP_SCHED
10372 .name = "rt_runtime_us",
10373 .read_s64 = cpu_rt_runtime_read,
10374 .write_s64 = cpu_rt_runtime_write,
10377 .name = "rt_period_us",
10378 .read_u64 = cpu_rt_period_read_uint,
10379 .write_u64 = cpu_rt_period_write_uint,
10384 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10386 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10389 struct cgroup_subsys cpu_cgroup_subsys = {
10391 .create = cpu_cgroup_create,
10392 .destroy = cpu_cgroup_destroy,
10393 .can_attach = cpu_cgroup_can_attach,
10394 .attach = cpu_cgroup_attach,
10395 .populate = cpu_cgroup_populate,
10396 .subsys_id = cpu_cgroup_subsys_id,
10400 #endif /* CONFIG_CGROUP_SCHED */
10402 #ifdef CONFIG_CGROUP_CPUACCT
10405 * CPU accounting code for task groups.
10407 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10408 * (balbir@in.ibm.com).
10411 /* track cpu usage of a group of tasks and its child groups */
10413 struct cgroup_subsys_state css;
10414 /* cpuusage holds pointer to a u64-type object on every cpu */
10416 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10417 struct cpuacct *parent;
10420 struct cgroup_subsys cpuacct_subsys;
10422 /* return cpu accounting group corresponding to this container */
10423 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10425 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10426 struct cpuacct, css);
10429 /* return cpu accounting group to which this task belongs */
10430 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10432 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10433 struct cpuacct, css);
10436 /* create a new cpu accounting group */
10437 static struct cgroup_subsys_state *cpuacct_create(
10438 struct cgroup_subsys *ss, struct cgroup *cgrp)
10440 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10446 ca->cpuusage = alloc_percpu(u64);
10450 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10451 if (percpu_counter_init(&ca->cpustat[i], 0))
10452 goto out_free_counters;
10455 ca->parent = cgroup_ca(cgrp->parent);
10461 percpu_counter_destroy(&ca->cpustat[i]);
10462 free_percpu(ca->cpuusage);
10466 return ERR_PTR(-ENOMEM);
10469 /* destroy an existing cpu accounting group */
10471 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10473 struct cpuacct *ca = cgroup_ca(cgrp);
10476 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10477 percpu_counter_destroy(&ca->cpustat[i]);
10478 free_percpu(ca->cpuusage);
10482 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10484 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10487 #ifndef CONFIG_64BIT
10489 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10491 spin_lock_irq(&cpu_rq(cpu)->lock);
10493 spin_unlock_irq(&cpu_rq(cpu)->lock);
10501 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10503 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10505 #ifndef CONFIG_64BIT
10507 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10509 spin_lock_irq(&cpu_rq(cpu)->lock);
10511 spin_unlock_irq(&cpu_rq(cpu)->lock);
10517 /* return total cpu usage (in nanoseconds) of a group */
10518 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10520 struct cpuacct *ca = cgroup_ca(cgrp);
10521 u64 totalcpuusage = 0;
10524 for_each_present_cpu(i)
10525 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10527 return totalcpuusage;
10530 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10533 struct cpuacct *ca = cgroup_ca(cgrp);
10542 for_each_present_cpu(i)
10543 cpuacct_cpuusage_write(ca, i, 0);
10549 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10550 struct seq_file *m)
10552 struct cpuacct *ca = cgroup_ca(cgroup);
10556 for_each_present_cpu(i) {
10557 percpu = cpuacct_cpuusage_read(ca, i);
10558 seq_printf(m, "%llu ", (unsigned long long) percpu);
10560 seq_printf(m, "\n");
10564 static const char *cpuacct_stat_desc[] = {
10565 [CPUACCT_STAT_USER] = "user",
10566 [CPUACCT_STAT_SYSTEM] = "system",
10569 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10570 struct cgroup_map_cb *cb)
10572 struct cpuacct *ca = cgroup_ca(cgrp);
10575 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10576 s64 val = percpu_counter_read(&ca->cpustat[i]);
10577 val = cputime64_to_clock_t(val);
10578 cb->fill(cb, cpuacct_stat_desc[i], val);
10583 static struct cftype files[] = {
10586 .read_u64 = cpuusage_read,
10587 .write_u64 = cpuusage_write,
10590 .name = "usage_percpu",
10591 .read_seq_string = cpuacct_percpu_seq_read,
10595 .read_map = cpuacct_stats_show,
10599 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10601 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10605 * charge this task's execution time to its accounting group.
10607 * called with rq->lock held.
10609 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10611 struct cpuacct *ca;
10614 if (unlikely(!cpuacct_subsys.active))
10617 cpu = task_cpu(tsk);
10623 for (; ca; ca = ca->parent) {
10624 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10625 *cpuusage += cputime;
10632 * Charge the system/user time to the task's accounting group.
10634 static void cpuacct_update_stats(struct task_struct *tsk,
10635 enum cpuacct_stat_index idx, cputime_t val)
10637 struct cpuacct *ca;
10639 if (unlikely(!cpuacct_subsys.active))
10646 percpu_counter_add(&ca->cpustat[idx], val);
10652 struct cgroup_subsys cpuacct_subsys = {
10654 .create = cpuacct_create,
10655 .destroy = cpuacct_destroy,
10656 .populate = cpuacct_populate,
10657 .subsys_id = cpuacct_subsys_id,
10659 #endif /* CONFIG_CGROUP_CPUACCT */