4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_counter.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/reciprocal_div.h>
68 #include <linux/unistd.h>
69 #include <linux/pagemap.h>
70 #include <linux/hrtimer.h>
71 #include <linux/tick.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/sched.h>
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * single value that denotes runtime == period, ie unlimited time.
121 #define RUNTIME_INF ((u64)~0ULL)
125 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
128 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
129 * Since cpu_power is a 'constant', we can use a reciprocal divide.
131 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
133 return reciprocal_divide(load, sg->reciprocal_cpu_power);
137 * Each time a sched group cpu_power is changed,
138 * we must compute its reciprocal value
140 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
142 sg->__cpu_power += val;
143 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
147 static inline int rt_policy(int policy)
149 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
154 static inline int task_has_rt_policy(struct task_struct *p)
156 return rt_policy(p->policy);
160 * This is the priority-queue data structure of the RT scheduling class:
162 struct rt_prio_array {
163 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
164 struct list_head queue[MAX_RT_PRIO];
167 struct rt_bandwidth {
168 /* nests inside the rq lock: */
169 spinlock_t rt_runtime_lock;
172 struct hrtimer rt_period_timer;
175 static struct rt_bandwidth def_rt_bandwidth;
177 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
179 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
181 struct rt_bandwidth *rt_b =
182 container_of(timer, struct rt_bandwidth, rt_period_timer);
188 now = hrtimer_cb_get_time(timer);
189 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
194 idle = do_sched_rt_period_timer(rt_b, overrun);
197 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
201 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
203 rt_b->rt_period = ns_to_ktime(period);
204 rt_b->rt_runtime = runtime;
206 spin_lock_init(&rt_b->rt_runtime_lock);
208 hrtimer_init(&rt_b->rt_period_timer,
209 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
210 rt_b->rt_period_timer.function = sched_rt_period_timer;
213 static inline int rt_bandwidth_enabled(void)
215 return sysctl_sched_rt_runtime >= 0;
218 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
222 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
225 if (hrtimer_active(&rt_b->rt_period_timer))
228 spin_lock(&rt_b->rt_runtime_lock);
233 if (hrtimer_active(&rt_b->rt_period_timer))
236 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
237 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
239 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
240 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
241 delta = ktime_to_ns(ktime_sub(hard, soft));
242 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
243 HRTIMER_MODE_ABS_PINNED, 0);
245 spin_unlock(&rt_b->rt_runtime_lock);
248 #ifdef CONFIG_RT_GROUP_SCHED
249 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
251 hrtimer_cancel(&rt_b->rt_period_timer);
256 * sched_domains_mutex serializes calls to arch_init_sched_domains,
257 * detach_destroy_domains and partition_sched_domains.
259 static DEFINE_MUTEX(sched_domains_mutex);
261 #ifdef CONFIG_GROUP_SCHED
263 #include <linux/cgroup.h>
267 static LIST_HEAD(task_groups);
269 /* task group related information */
271 #ifdef CONFIG_CGROUP_SCHED
272 struct cgroup_subsys_state css;
275 #ifdef CONFIG_USER_SCHED
279 #ifdef CONFIG_FAIR_GROUP_SCHED
280 /* schedulable entities of this group on each cpu */
281 struct sched_entity **se;
282 /* runqueue "owned" by this group on each cpu */
283 struct cfs_rq **cfs_rq;
284 unsigned long shares;
287 #ifdef CONFIG_RT_GROUP_SCHED
288 struct sched_rt_entity **rt_se;
289 struct rt_rq **rt_rq;
291 struct rt_bandwidth rt_bandwidth;
295 struct list_head list;
297 struct task_group *parent;
298 struct list_head siblings;
299 struct list_head children;
302 #ifdef CONFIG_USER_SCHED
304 /* Helper function to pass uid information to create_sched_user() */
305 void set_tg_uid(struct user_struct *user)
307 user->tg->uid = user->uid;
312 * Every UID task group (including init_task_group aka UID-0) will
313 * be a child to this group.
315 struct task_group root_task_group;
317 #ifdef CONFIG_FAIR_GROUP_SCHED
318 /* Default task group's sched entity on each cpu */
319 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
320 /* Default task group's cfs_rq on each cpu */
321 static DEFINE_PER_CPU(struct cfs_rq, init_tg_cfs_rq) ____cacheline_aligned_in_smp;
322 #endif /* CONFIG_FAIR_GROUP_SCHED */
324 #ifdef CONFIG_RT_GROUP_SCHED
325 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
326 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
327 #endif /* CONFIG_RT_GROUP_SCHED */
328 #else /* !CONFIG_USER_SCHED */
329 #define root_task_group init_task_group
330 #endif /* CONFIG_USER_SCHED */
332 /* task_group_lock serializes add/remove of task groups and also changes to
333 * a task group's cpu shares.
335 static DEFINE_SPINLOCK(task_group_lock);
338 static int root_task_group_empty(void)
340 return list_empty(&root_task_group.children);
344 #ifdef CONFIG_FAIR_GROUP_SCHED
345 #ifdef CONFIG_USER_SCHED
346 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
347 #else /* !CONFIG_USER_SCHED */
348 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
349 #endif /* CONFIG_USER_SCHED */
352 * A weight of 0 or 1 can cause arithmetics problems.
353 * A weight of a cfs_rq is the sum of weights of which entities
354 * are queued on this cfs_rq, so a weight of a entity should not be
355 * too large, so as the shares value of a task group.
356 * (The default weight is 1024 - so there's no practical
357 * limitation from this.)
360 #define MAX_SHARES (1UL << 18)
362 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
365 /* Default task group.
366 * Every task in system belong to this group at bootup.
368 struct task_group init_task_group;
370 /* return group to which a task belongs */
371 static inline struct task_group *task_group(struct task_struct *p)
373 struct task_group *tg;
375 #ifdef CONFIG_USER_SCHED
377 tg = __task_cred(p)->user->tg;
379 #elif defined(CONFIG_CGROUP_SCHED)
380 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
381 struct task_group, css);
383 tg = &init_task_group;
388 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
389 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
391 #ifdef CONFIG_FAIR_GROUP_SCHED
392 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
393 p->se.parent = task_group(p)->se[cpu];
396 #ifdef CONFIG_RT_GROUP_SCHED
397 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
398 p->rt.parent = task_group(p)->rt_se[cpu];
405 static int root_task_group_empty(void)
411 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
412 static inline struct task_group *task_group(struct task_struct *p)
417 #endif /* CONFIG_GROUP_SCHED */
419 /* CFS-related fields in a runqueue */
421 struct load_weight load;
422 unsigned long nr_running;
427 struct rb_root tasks_timeline;
428 struct rb_node *rb_leftmost;
430 struct list_head tasks;
431 struct list_head *balance_iterator;
434 * 'curr' points to currently running entity on this cfs_rq.
435 * It is set to NULL otherwise (i.e when none are currently running).
437 struct sched_entity *curr, *next, *last;
439 unsigned int nr_spread_over;
441 #ifdef CONFIG_FAIR_GROUP_SCHED
442 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
445 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
446 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
447 * (like users, containers etc.)
449 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
450 * list is used during load balance.
452 struct list_head leaf_cfs_rq_list;
453 struct task_group *tg; /* group that "owns" this runqueue */
457 * the part of load.weight contributed by tasks
459 unsigned long task_weight;
462 * h_load = weight * f(tg)
464 * Where f(tg) is the recursive weight fraction assigned to
467 unsigned long h_load;
470 * this cpu's part of tg->shares
472 unsigned long shares;
475 * load.weight at the time we set shares
477 unsigned long rq_weight;
482 /* Real-Time classes' related field in a runqueue: */
484 struct rt_prio_array active;
485 unsigned long rt_nr_running;
486 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
488 int curr; /* highest queued rt task prio */
490 int next; /* next highest */
495 unsigned long rt_nr_migratory;
496 unsigned long rt_nr_total;
498 struct plist_head pushable_tasks;
503 /* Nests inside the rq lock: */
504 spinlock_t rt_runtime_lock;
506 #ifdef CONFIG_RT_GROUP_SCHED
507 unsigned long rt_nr_boosted;
510 struct list_head leaf_rt_rq_list;
511 struct task_group *tg;
512 struct sched_rt_entity *rt_se;
519 * We add the notion of a root-domain which will be used to define per-domain
520 * variables. Each exclusive cpuset essentially defines an island domain by
521 * fully partitioning the member cpus from any other cpuset. Whenever a new
522 * exclusive cpuset is created, we also create and attach a new root-domain
529 cpumask_var_t online;
532 * The "RT overload" flag: it gets set if a CPU has more than
533 * one runnable RT task.
535 cpumask_var_t rto_mask;
538 struct cpupri cpupri;
540 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
542 * Preferred wake up cpu nominated by sched_mc balance that will be
543 * used when most cpus are idle in the system indicating overall very
544 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
546 unsigned int sched_mc_preferred_wakeup_cpu;
551 * By default the system creates a single root-domain with all cpus as
552 * members (mimicking the global state we have today).
554 static struct root_domain def_root_domain;
559 * This is the main, per-CPU runqueue data structure.
561 * Locking rule: those places that want to lock multiple runqueues
562 * (such as the load balancing or the thread migration code), lock
563 * acquire operations must be ordered by ascending &runqueue.
570 * nr_running and cpu_load should be in the same cacheline because
571 * remote CPUs use both these fields when doing load calculation.
573 unsigned long nr_running;
574 #define CPU_LOAD_IDX_MAX 5
575 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
577 unsigned long last_tick_seen;
578 unsigned char in_nohz_recently;
580 /* capture load from *all* tasks on this cpu: */
581 struct load_weight load;
582 unsigned long nr_load_updates;
584 u64 nr_migrations_in;
589 #ifdef CONFIG_FAIR_GROUP_SCHED
590 /* list of leaf cfs_rq on this cpu: */
591 struct list_head leaf_cfs_rq_list;
593 #ifdef CONFIG_RT_GROUP_SCHED
594 struct list_head leaf_rt_rq_list;
598 * This is part of a global counter where only the total sum
599 * over all CPUs matters. A task can increase this counter on
600 * one CPU and if it got migrated afterwards it may decrease
601 * it on another CPU. Always updated under the runqueue lock:
603 unsigned long nr_uninterruptible;
605 struct task_struct *curr, *idle;
606 unsigned long next_balance;
607 struct mm_struct *prev_mm;
614 struct root_domain *rd;
615 struct sched_domain *sd;
617 unsigned char idle_at_tick;
618 /* For active balancing */
622 /* cpu of this runqueue: */
626 unsigned long avg_load_per_task;
628 struct task_struct *migration_thread;
629 struct list_head migration_queue;
632 /* calc_load related fields */
633 unsigned long calc_load_update;
634 long calc_load_active;
636 #ifdef CONFIG_SCHED_HRTICK
638 int hrtick_csd_pending;
639 struct call_single_data hrtick_csd;
641 struct hrtimer hrtick_timer;
644 #ifdef CONFIG_SCHEDSTATS
646 struct sched_info rq_sched_info;
647 unsigned long long rq_cpu_time;
648 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
650 /* sys_sched_yield() stats */
651 unsigned int yld_count;
653 /* schedule() stats */
654 unsigned int sched_switch;
655 unsigned int sched_count;
656 unsigned int sched_goidle;
658 /* try_to_wake_up() stats */
659 unsigned int ttwu_count;
660 unsigned int ttwu_local;
663 unsigned int bkl_count;
667 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
669 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
671 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
674 static inline int cpu_of(struct rq *rq)
684 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
685 * See detach_destroy_domains: synchronize_sched for details.
687 * The domain tree of any CPU may only be accessed from within
688 * preempt-disabled sections.
690 #define for_each_domain(cpu, __sd) \
691 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
693 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
694 #define this_rq() (&__get_cpu_var(runqueues))
695 #define task_rq(p) cpu_rq(task_cpu(p))
696 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
697 #define raw_rq() (&__raw_get_cpu_var(runqueues))
699 inline void update_rq_clock(struct rq *rq)
701 rq->clock = sched_clock_cpu(cpu_of(rq));
705 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
707 #ifdef CONFIG_SCHED_DEBUG
708 # define const_debug __read_mostly
710 # define const_debug static const
716 * Returns true if the current cpu runqueue is locked.
717 * This interface allows printk to be called with the runqueue lock
718 * held and know whether or not it is OK to wake up the klogd.
720 int runqueue_is_locked(void)
723 struct rq *rq = cpu_rq(cpu);
726 ret = spin_is_locked(&rq->lock);
732 * Debugging: various feature bits
735 #define SCHED_FEAT(name, enabled) \
736 __SCHED_FEAT_##name ,
739 #include "sched_features.h"
744 #define SCHED_FEAT(name, enabled) \
745 (1UL << __SCHED_FEAT_##name) * enabled |
747 const_debug unsigned int sysctl_sched_features =
748 #include "sched_features.h"
753 #ifdef CONFIG_SCHED_DEBUG
754 #define SCHED_FEAT(name, enabled) \
757 static __read_mostly char *sched_feat_names[] = {
758 #include "sched_features.h"
764 static int sched_feat_show(struct seq_file *m, void *v)
768 for (i = 0; sched_feat_names[i]; i++) {
769 if (!(sysctl_sched_features & (1UL << i)))
771 seq_printf(m, "%s ", sched_feat_names[i]);
779 sched_feat_write(struct file *filp, const char __user *ubuf,
780 size_t cnt, loff_t *ppos)
790 if (copy_from_user(&buf, ubuf, cnt))
795 if (strncmp(buf, "NO_", 3) == 0) {
800 for (i = 0; sched_feat_names[i]; i++) {
801 int len = strlen(sched_feat_names[i]);
803 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
805 sysctl_sched_features &= ~(1UL << i);
807 sysctl_sched_features |= (1UL << i);
812 if (!sched_feat_names[i])
820 static int sched_feat_open(struct inode *inode, struct file *filp)
822 return single_open(filp, sched_feat_show, NULL);
825 static struct file_operations sched_feat_fops = {
826 .open = sched_feat_open,
827 .write = sched_feat_write,
830 .release = single_release,
833 static __init int sched_init_debug(void)
835 debugfs_create_file("sched_features", 0644, NULL, NULL,
840 late_initcall(sched_init_debug);
844 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
847 * Number of tasks to iterate in a single balance run.
848 * Limited because this is done with IRQs disabled.
850 const_debug unsigned int sysctl_sched_nr_migrate = 32;
853 * ratelimit for updating the group shares.
856 unsigned int sysctl_sched_shares_ratelimit = 250000;
859 * Inject some fuzzyness into changing the per-cpu group shares
860 * this avoids remote rq-locks at the expense of fairness.
863 unsigned int sysctl_sched_shares_thresh = 4;
866 * period over which we measure -rt task cpu usage in us.
869 unsigned int sysctl_sched_rt_period = 1000000;
871 static __read_mostly int scheduler_running;
874 * part of the period that we allow rt tasks to run in us.
877 int sysctl_sched_rt_runtime = 950000;
879 static inline u64 global_rt_period(void)
881 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
884 static inline u64 global_rt_runtime(void)
886 if (sysctl_sched_rt_runtime < 0)
889 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
892 #ifndef prepare_arch_switch
893 # define prepare_arch_switch(next) do { } while (0)
895 #ifndef finish_arch_switch
896 # define finish_arch_switch(prev) do { } while (0)
899 static inline int task_current(struct rq *rq, struct task_struct *p)
901 return rq->curr == p;
904 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
905 static inline int task_running(struct rq *rq, struct task_struct *p)
907 return task_current(rq, p);
910 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
914 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
916 #ifdef CONFIG_DEBUG_SPINLOCK
917 /* this is a valid case when another task releases the spinlock */
918 rq->lock.owner = current;
921 * If we are tracking spinlock dependencies then we have to
922 * fix up the runqueue lock - which gets 'carried over' from
925 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
927 spin_unlock_irq(&rq->lock);
930 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
931 static inline int task_running(struct rq *rq, struct task_struct *p)
936 return task_current(rq, p);
940 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
944 * We can optimise this out completely for !SMP, because the
945 * SMP rebalancing from interrupt is the only thing that cares
950 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
951 spin_unlock_irq(&rq->lock);
953 spin_unlock(&rq->lock);
957 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
961 * After ->oncpu is cleared, the task can be moved to a different CPU.
962 * We must ensure this doesn't happen until the switch is completely
968 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
972 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
975 * __task_rq_lock - lock the runqueue a given task resides on.
976 * Must be called interrupts disabled.
978 static inline struct rq *__task_rq_lock(struct task_struct *p)
982 struct rq *rq = task_rq(p);
983 spin_lock(&rq->lock);
984 if (likely(rq == task_rq(p)))
986 spin_unlock(&rq->lock);
991 * task_rq_lock - lock the runqueue a given task resides on and disable
992 * interrupts. Note the ordering: we can safely lookup the task_rq without
993 * explicitly disabling preemption.
995 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1001 local_irq_save(*flags);
1003 spin_lock(&rq->lock);
1004 if (likely(rq == task_rq(p)))
1006 spin_unlock_irqrestore(&rq->lock, *flags);
1010 void task_rq_unlock_wait(struct task_struct *p)
1012 struct rq *rq = task_rq(p);
1014 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1015 spin_unlock_wait(&rq->lock);
1018 static void __task_rq_unlock(struct rq *rq)
1019 __releases(rq->lock)
1021 spin_unlock(&rq->lock);
1024 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1025 __releases(rq->lock)
1027 spin_unlock_irqrestore(&rq->lock, *flags);
1031 * this_rq_lock - lock this runqueue and disable interrupts.
1033 static struct rq *this_rq_lock(void)
1034 __acquires(rq->lock)
1038 local_irq_disable();
1040 spin_lock(&rq->lock);
1045 #ifdef CONFIG_SCHED_HRTICK
1047 * Use HR-timers to deliver accurate preemption points.
1049 * Its all a bit involved since we cannot program an hrt while holding the
1050 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1053 * When we get rescheduled we reprogram the hrtick_timer outside of the
1059 * - enabled by features
1060 * - hrtimer is actually high res
1062 static inline int hrtick_enabled(struct rq *rq)
1064 if (!sched_feat(HRTICK))
1066 if (!cpu_active(cpu_of(rq)))
1068 return hrtimer_is_hres_active(&rq->hrtick_timer);
1071 static void hrtick_clear(struct rq *rq)
1073 if (hrtimer_active(&rq->hrtick_timer))
1074 hrtimer_cancel(&rq->hrtick_timer);
1078 * High-resolution timer tick.
1079 * Runs from hardirq context with interrupts disabled.
1081 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1083 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1085 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1087 spin_lock(&rq->lock);
1088 update_rq_clock(rq);
1089 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1090 spin_unlock(&rq->lock);
1092 return HRTIMER_NORESTART;
1097 * called from hardirq (IPI) context
1099 static void __hrtick_start(void *arg)
1101 struct rq *rq = arg;
1103 spin_lock(&rq->lock);
1104 hrtimer_restart(&rq->hrtick_timer);
1105 rq->hrtick_csd_pending = 0;
1106 spin_unlock(&rq->lock);
1110 * Called to set the hrtick timer state.
1112 * called with rq->lock held and irqs disabled
1114 static void hrtick_start(struct rq *rq, u64 delay)
1116 struct hrtimer *timer = &rq->hrtick_timer;
1117 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1119 hrtimer_set_expires(timer, time);
1121 if (rq == this_rq()) {
1122 hrtimer_restart(timer);
1123 } else if (!rq->hrtick_csd_pending) {
1124 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1125 rq->hrtick_csd_pending = 1;
1130 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1132 int cpu = (int)(long)hcpu;
1135 case CPU_UP_CANCELED:
1136 case CPU_UP_CANCELED_FROZEN:
1137 case CPU_DOWN_PREPARE:
1138 case CPU_DOWN_PREPARE_FROZEN:
1140 case CPU_DEAD_FROZEN:
1141 hrtick_clear(cpu_rq(cpu));
1148 static __init void init_hrtick(void)
1150 hotcpu_notifier(hotplug_hrtick, 0);
1154 * Called to set the hrtick timer state.
1156 * called with rq->lock held and irqs disabled
1158 static void hrtick_start(struct rq *rq, u64 delay)
1160 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1161 HRTIMER_MODE_REL_PINNED, 0);
1164 static inline void init_hrtick(void)
1167 #endif /* CONFIG_SMP */
1169 static void init_rq_hrtick(struct rq *rq)
1172 rq->hrtick_csd_pending = 0;
1174 rq->hrtick_csd.flags = 0;
1175 rq->hrtick_csd.func = __hrtick_start;
1176 rq->hrtick_csd.info = rq;
1179 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1180 rq->hrtick_timer.function = hrtick;
1182 #else /* CONFIG_SCHED_HRTICK */
1183 static inline void hrtick_clear(struct rq *rq)
1187 static inline void init_rq_hrtick(struct rq *rq)
1191 static inline void init_hrtick(void)
1194 #endif /* CONFIG_SCHED_HRTICK */
1197 * resched_task - mark a task 'to be rescheduled now'.
1199 * On UP this means the setting of the need_resched flag, on SMP it
1200 * might also involve a cross-CPU call to trigger the scheduler on
1205 #ifndef tsk_is_polling
1206 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1209 static void resched_task(struct task_struct *p)
1213 assert_spin_locked(&task_rq(p)->lock);
1215 if (test_tsk_need_resched(p))
1218 set_tsk_need_resched(p);
1221 if (cpu == smp_processor_id())
1224 /* NEED_RESCHED must be visible before we test polling */
1226 if (!tsk_is_polling(p))
1227 smp_send_reschedule(cpu);
1230 static void resched_cpu(int cpu)
1232 struct rq *rq = cpu_rq(cpu);
1233 unsigned long flags;
1235 if (!spin_trylock_irqsave(&rq->lock, flags))
1237 resched_task(cpu_curr(cpu));
1238 spin_unlock_irqrestore(&rq->lock, flags);
1243 * When add_timer_on() enqueues a timer into the timer wheel of an
1244 * idle CPU then this timer might expire before the next timer event
1245 * which is scheduled to wake up that CPU. In case of a completely
1246 * idle system the next event might even be infinite time into the
1247 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1248 * leaves the inner idle loop so the newly added timer is taken into
1249 * account when the CPU goes back to idle and evaluates the timer
1250 * wheel for the next timer event.
1252 void wake_up_idle_cpu(int cpu)
1254 struct rq *rq = cpu_rq(cpu);
1256 if (cpu == smp_processor_id())
1260 * This is safe, as this function is called with the timer
1261 * wheel base lock of (cpu) held. When the CPU is on the way
1262 * to idle and has not yet set rq->curr to idle then it will
1263 * be serialized on the timer wheel base lock and take the new
1264 * timer into account automatically.
1266 if (rq->curr != rq->idle)
1270 * We can set TIF_RESCHED on the idle task of the other CPU
1271 * lockless. The worst case is that the other CPU runs the
1272 * idle task through an additional NOOP schedule()
1274 set_tsk_need_resched(rq->idle);
1276 /* NEED_RESCHED must be visible before we test polling */
1278 if (!tsk_is_polling(rq->idle))
1279 smp_send_reschedule(cpu);
1281 #endif /* CONFIG_NO_HZ */
1283 #else /* !CONFIG_SMP */
1284 static void resched_task(struct task_struct *p)
1286 assert_spin_locked(&task_rq(p)->lock);
1287 set_tsk_need_resched(p);
1289 #endif /* CONFIG_SMP */
1291 #if BITS_PER_LONG == 32
1292 # define WMULT_CONST (~0UL)
1294 # define WMULT_CONST (1UL << 32)
1297 #define WMULT_SHIFT 32
1300 * Shift right and round:
1302 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1305 * delta *= weight / lw
1307 static unsigned long
1308 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1309 struct load_weight *lw)
1313 if (!lw->inv_weight) {
1314 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1317 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1321 tmp = (u64)delta_exec * weight;
1323 * Check whether we'd overflow the 64-bit multiplication:
1325 if (unlikely(tmp > WMULT_CONST))
1326 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1329 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1331 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1334 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1340 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1347 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1348 * of tasks with abnormal "nice" values across CPUs the contribution that
1349 * each task makes to its run queue's load is weighted according to its
1350 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1351 * scaled version of the new time slice allocation that they receive on time
1355 #define WEIGHT_IDLEPRIO 3
1356 #define WMULT_IDLEPRIO 1431655765
1359 * Nice levels are multiplicative, with a gentle 10% change for every
1360 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1361 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1362 * that remained on nice 0.
1364 * The "10% effect" is relative and cumulative: from _any_ nice level,
1365 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1366 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1367 * If a task goes up by ~10% and another task goes down by ~10% then
1368 * the relative distance between them is ~25%.)
1370 static const int prio_to_weight[40] = {
1371 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1372 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1373 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1374 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1375 /* 0 */ 1024, 820, 655, 526, 423,
1376 /* 5 */ 335, 272, 215, 172, 137,
1377 /* 10 */ 110, 87, 70, 56, 45,
1378 /* 15 */ 36, 29, 23, 18, 15,
1382 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1384 * In cases where the weight does not change often, we can use the
1385 * precalculated inverse to speed up arithmetics by turning divisions
1386 * into multiplications:
1388 static const u32 prio_to_wmult[40] = {
1389 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1390 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1391 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1392 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1393 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1394 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1395 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1396 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1399 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1402 * runqueue iterator, to support SMP load-balancing between different
1403 * scheduling classes, without having to expose their internal data
1404 * structures to the load-balancing proper:
1406 struct rq_iterator {
1408 struct task_struct *(*start)(void *);
1409 struct task_struct *(*next)(void *);
1413 static unsigned long
1414 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1415 unsigned long max_load_move, struct sched_domain *sd,
1416 enum cpu_idle_type idle, int *all_pinned,
1417 int *this_best_prio, struct rq_iterator *iterator);
1420 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1421 struct sched_domain *sd, enum cpu_idle_type idle,
1422 struct rq_iterator *iterator);
1425 /* Time spent by the tasks of the cpu accounting group executing in ... */
1426 enum cpuacct_stat_index {
1427 CPUACCT_STAT_USER, /* ... user mode */
1428 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1430 CPUACCT_STAT_NSTATS,
1433 #ifdef CONFIG_CGROUP_CPUACCT
1434 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1435 static void cpuacct_update_stats(struct task_struct *tsk,
1436 enum cpuacct_stat_index idx, cputime_t val);
1438 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1439 static inline void cpuacct_update_stats(struct task_struct *tsk,
1440 enum cpuacct_stat_index idx, cputime_t val) {}
1443 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1445 update_load_add(&rq->load, load);
1448 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1450 update_load_sub(&rq->load, load);
1453 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1454 typedef int (*tg_visitor)(struct task_group *, void *);
1457 * Iterate the full tree, calling @down when first entering a node and @up when
1458 * leaving it for the final time.
1460 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1462 struct task_group *parent, *child;
1466 parent = &root_task_group;
1468 ret = (*down)(parent, data);
1471 list_for_each_entry_rcu(child, &parent->children, siblings) {
1478 ret = (*up)(parent, data);
1483 parent = parent->parent;
1492 static int tg_nop(struct task_group *tg, void *data)
1499 static unsigned long source_load(int cpu, int type);
1500 static unsigned long target_load(int cpu, int type);
1501 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1503 static unsigned long cpu_avg_load_per_task(int cpu)
1505 struct rq *rq = cpu_rq(cpu);
1506 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1509 rq->avg_load_per_task = rq->load.weight / nr_running;
1511 rq->avg_load_per_task = 0;
1513 return rq->avg_load_per_task;
1516 #ifdef CONFIG_FAIR_GROUP_SCHED
1518 struct update_shares_data {
1519 unsigned long rq_weight[NR_CPUS];
1522 static DEFINE_PER_CPU(struct update_shares_data, update_shares_data);
1524 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1527 * Calculate and set the cpu's group shares.
1529 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1530 unsigned long sd_shares,
1531 unsigned long sd_rq_weight,
1532 struct update_shares_data *usd)
1534 unsigned long shares, rq_weight;
1537 rq_weight = usd->rq_weight[cpu];
1540 rq_weight = NICE_0_LOAD;
1544 * \Sum_j shares_j * rq_weight_i
1545 * shares_i = -----------------------------
1546 * \Sum_j rq_weight_j
1548 shares = (sd_shares * rq_weight) / sd_rq_weight;
1549 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1551 if (abs(shares - tg->se[cpu]->load.weight) >
1552 sysctl_sched_shares_thresh) {
1553 struct rq *rq = cpu_rq(cpu);
1554 unsigned long flags;
1556 spin_lock_irqsave(&rq->lock, flags);
1557 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1558 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1559 __set_se_shares(tg->se[cpu], shares);
1560 spin_unlock_irqrestore(&rq->lock, flags);
1565 * Re-compute the task group their per cpu shares over the given domain.
1566 * This needs to be done in a bottom-up fashion because the rq weight of a
1567 * parent group depends on the shares of its child groups.
1569 static int tg_shares_up(struct task_group *tg, void *data)
1571 unsigned long weight, rq_weight = 0, shares = 0;
1572 struct update_shares_data *usd;
1573 struct sched_domain *sd = data;
1574 unsigned long flags;
1580 local_irq_save(flags);
1581 usd = &__get_cpu_var(update_shares_data);
1583 for_each_cpu(i, sched_domain_span(sd)) {
1584 weight = tg->cfs_rq[i]->load.weight;
1585 usd->rq_weight[i] = weight;
1588 * If there are currently no tasks on the cpu pretend there
1589 * is one of average load so that when a new task gets to
1590 * run here it will not get delayed by group starvation.
1593 weight = NICE_0_LOAD;
1595 rq_weight += weight;
1596 shares += tg->cfs_rq[i]->shares;
1599 if ((!shares && rq_weight) || shares > tg->shares)
1600 shares = tg->shares;
1602 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1603 shares = tg->shares;
1605 for_each_cpu(i, sched_domain_span(sd))
1606 update_group_shares_cpu(tg, i, shares, rq_weight, usd);
1608 local_irq_restore(flags);
1614 * Compute the cpu's hierarchical load factor for each task group.
1615 * This needs to be done in a top-down fashion because the load of a child
1616 * group is a fraction of its parents load.
1618 static int tg_load_down(struct task_group *tg, void *data)
1621 long cpu = (long)data;
1624 load = cpu_rq(cpu)->load.weight;
1626 load = tg->parent->cfs_rq[cpu]->h_load;
1627 load *= tg->cfs_rq[cpu]->shares;
1628 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1631 tg->cfs_rq[cpu]->h_load = load;
1636 static void update_shares(struct sched_domain *sd)
1641 if (root_task_group_empty())
1644 now = cpu_clock(raw_smp_processor_id());
1645 elapsed = now - sd->last_update;
1647 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1648 sd->last_update = now;
1649 walk_tg_tree(tg_nop, tg_shares_up, sd);
1653 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1655 if (root_task_group_empty())
1658 spin_unlock(&rq->lock);
1660 spin_lock(&rq->lock);
1663 static void update_h_load(long cpu)
1665 if (root_task_group_empty())
1668 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1673 static inline void update_shares(struct sched_domain *sd)
1677 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1683 #ifdef CONFIG_PREEMPT
1686 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1687 * way at the expense of forcing extra atomic operations in all
1688 * invocations. This assures that the double_lock is acquired using the
1689 * same underlying policy as the spinlock_t on this architecture, which
1690 * reduces latency compared to the unfair variant below. However, it
1691 * also adds more overhead and therefore may reduce throughput.
1693 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1694 __releases(this_rq->lock)
1695 __acquires(busiest->lock)
1696 __acquires(this_rq->lock)
1698 spin_unlock(&this_rq->lock);
1699 double_rq_lock(this_rq, busiest);
1706 * Unfair double_lock_balance: Optimizes throughput at the expense of
1707 * latency by eliminating extra atomic operations when the locks are
1708 * already in proper order on entry. This favors lower cpu-ids and will
1709 * grant the double lock to lower cpus over higher ids under contention,
1710 * regardless of entry order into the function.
1712 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1713 __releases(this_rq->lock)
1714 __acquires(busiest->lock)
1715 __acquires(this_rq->lock)
1719 if (unlikely(!spin_trylock(&busiest->lock))) {
1720 if (busiest < this_rq) {
1721 spin_unlock(&this_rq->lock);
1722 spin_lock(&busiest->lock);
1723 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1726 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1731 #endif /* CONFIG_PREEMPT */
1734 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1736 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1738 if (unlikely(!irqs_disabled())) {
1739 /* printk() doesn't work good under rq->lock */
1740 spin_unlock(&this_rq->lock);
1744 return _double_lock_balance(this_rq, busiest);
1747 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1748 __releases(busiest->lock)
1750 spin_unlock(&busiest->lock);
1751 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1755 #ifdef CONFIG_FAIR_GROUP_SCHED
1756 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1759 cfs_rq->shares = shares;
1764 static void calc_load_account_active(struct rq *this_rq);
1766 #include "sched_stats.h"
1767 #include "sched_idletask.c"
1768 #include "sched_fair.c"
1769 #include "sched_rt.c"
1770 #ifdef CONFIG_SCHED_DEBUG
1771 # include "sched_debug.c"
1774 #define sched_class_highest (&rt_sched_class)
1775 #define for_each_class(class) \
1776 for (class = sched_class_highest; class; class = class->next)
1778 static void inc_nr_running(struct rq *rq)
1783 static void dec_nr_running(struct rq *rq)
1788 static void set_load_weight(struct task_struct *p)
1790 if (task_has_rt_policy(p)) {
1791 p->se.load.weight = prio_to_weight[0] * 2;
1792 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1797 * SCHED_IDLE tasks get minimal weight:
1799 if (p->policy == SCHED_IDLE) {
1800 p->se.load.weight = WEIGHT_IDLEPRIO;
1801 p->se.load.inv_weight = WMULT_IDLEPRIO;
1805 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1806 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1809 static void update_avg(u64 *avg, u64 sample)
1811 s64 diff = sample - *avg;
1815 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1818 p->se.start_runtime = p->se.sum_exec_runtime;
1820 sched_info_queued(p);
1821 p->sched_class->enqueue_task(rq, p, wakeup);
1825 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1828 if (p->se.last_wakeup) {
1829 update_avg(&p->se.avg_overlap,
1830 p->se.sum_exec_runtime - p->se.last_wakeup);
1831 p->se.last_wakeup = 0;
1833 update_avg(&p->se.avg_wakeup,
1834 sysctl_sched_wakeup_granularity);
1838 sched_info_dequeued(p);
1839 p->sched_class->dequeue_task(rq, p, sleep);
1844 * __normal_prio - return the priority that is based on the static prio
1846 static inline int __normal_prio(struct task_struct *p)
1848 return p->static_prio;
1852 * Calculate the expected normal priority: i.e. priority
1853 * without taking RT-inheritance into account. Might be
1854 * boosted by interactivity modifiers. Changes upon fork,
1855 * setprio syscalls, and whenever the interactivity
1856 * estimator recalculates.
1858 static inline int normal_prio(struct task_struct *p)
1862 if (task_has_rt_policy(p))
1863 prio = MAX_RT_PRIO-1 - p->rt_priority;
1865 prio = __normal_prio(p);
1870 * Calculate the current priority, i.e. the priority
1871 * taken into account by the scheduler. This value might
1872 * be boosted by RT tasks, or might be boosted by
1873 * interactivity modifiers. Will be RT if the task got
1874 * RT-boosted. If not then it returns p->normal_prio.
1876 static int effective_prio(struct task_struct *p)
1878 p->normal_prio = normal_prio(p);
1880 * If we are RT tasks or we were boosted to RT priority,
1881 * keep the priority unchanged. Otherwise, update priority
1882 * to the normal priority:
1884 if (!rt_prio(p->prio))
1885 return p->normal_prio;
1890 * activate_task - move a task to the runqueue.
1892 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1894 if (task_contributes_to_load(p))
1895 rq->nr_uninterruptible--;
1897 enqueue_task(rq, p, wakeup);
1902 * deactivate_task - remove a task from the runqueue.
1904 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1906 if (task_contributes_to_load(p))
1907 rq->nr_uninterruptible++;
1909 dequeue_task(rq, p, sleep);
1914 * task_curr - is this task currently executing on a CPU?
1915 * @p: the task in question.
1917 inline int task_curr(const struct task_struct *p)
1919 return cpu_curr(task_cpu(p)) == p;
1922 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1924 set_task_rq(p, cpu);
1927 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1928 * successfuly executed on another CPU. We must ensure that updates of
1929 * per-task data have been completed by this moment.
1932 task_thread_info(p)->cpu = cpu;
1936 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1937 const struct sched_class *prev_class,
1938 int oldprio, int running)
1940 if (prev_class != p->sched_class) {
1941 if (prev_class->switched_from)
1942 prev_class->switched_from(rq, p, running);
1943 p->sched_class->switched_to(rq, p, running);
1945 p->sched_class->prio_changed(rq, p, oldprio, running);
1950 /* Used instead of source_load when we know the type == 0 */
1951 static unsigned long weighted_cpuload(const int cpu)
1953 return cpu_rq(cpu)->load.weight;
1957 * Is this task likely cache-hot:
1960 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1965 * Buddy candidates are cache hot:
1967 if (sched_feat(CACHE_HOT_BUDDY) &&
1968 (&p->se == cfs_rq_of(&p->se)->next ||
1969 &p->se == cfs_rq_of(&p->se)->last))
1972 if (p->sched_class != &fair_sched_class)
1975 if (sysctl_sched_migration_cost == -1)
1977 if (sysctl_sched_migration_cost == 0)
1980 delta = now - p->se.exec_start;
1982 return delta < (s64)sysctl_sched_migration_cost;
1986 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1988 int old_cpu = task_cpu(p);
1989 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1990 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1991 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1994 clock_offset = old_rq->clock - new_rq->clock;
1996 trace_sched_migrate_task(p, new_cpu);
1998 #ifdef CONFIG_SCHEDSTATS
1999 if (p->se.wait_start)
2000 p->se.wait_start -= clock_offset;
2001 if (p->se.sleep_start)
2002 p->se.sleep_start -= clock_offset;
2003 if (p->se.block_start)
2004 p->se.block_start -= clock_offset;
2006 if (old_cpu != new_cpu) {
2007 p->se.nr_migrations++;
2008 new_rq->nr_migrations_in++;
2009 #ifdef CONFIG_SCHEDSTATS
2010 if (task_hot(p, old_rq->clock, NULL))
2011 schedstat_inc(p, se.nr_forced2_migrations);
2013 perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2016 p->se.vruntime -= old_cfsrq->min_vruntime -
2017 new_cfsrq->min_vruntime;
2019 __set_task_cpu(p, new_cpu);
2022 struct migration_req {
2023 struct list_head list;
2025 struct task_struct *task;
2028 struct completion done;
2032 * The task's runqueue lock must be held.
2033 * Returns true if you have to wait for migration thread.
2036 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2038 struct rq *rq = task_rq(p);
2041 * If the task is not on a runqueue (and not running), then
2042 * it is sufficient to simply update the task's cpu field.
2044 if (!p->se.on_rq && !task_running(rq, p)) {
2045 set_task_cpu(p, dest_cpu);
2049 init_completion(&req->done);
2051 req->dest_cpu = dest_cpu;
2052 list_add(&req->list, &rq->migration_queue);
2058 * wait_task_context_switch - wait for a thread to complete at least one
2061 * @p must not be current.
2063 void wait_task_context_switch(struct task_struct *p)
2065 unsigned long nvcsw, nivcsw, flags;
2073 * The runqueue is assigned before the actual context
2074 * switch. We need to take the runqueue lock.
2076 * We could check initially without the lock but it is
2077 * very likely that we need to take the lock in every
2080 rq = task_rq_lock(p, &flags);
2081 running = task_running(rq, p);
2082 task_rq_unlock(rq, &flags);
2084 if (likely(!running))
2087 * The switch count is incremented before the actual
2088 * context switch. We thus wait for two switches to be
2089 * sure at least one completed.
2091 if ((p->nvcsw - nvcsw) > 1)
2093 if ((p->nivcsw - nivcsw) > 1)
2101 * wait_task_inactive - wait for a thread to unschedule.
2103 * If @match_state is nonzero, it's the @p->state value just checked and
2104 * not expected to change. If it changes, i.e. @p might have woken up,
2105 * then return zero. When we succeed in waiting for @p to be off its CPU,
2106 * we return a positive number (its total switch count). If a second call
2107 * a short while later returns the same number, the caller can be sure that
2108 * @p has remained unscheduled the whole time.
2110 * The caller must ensure that the task *will* unschedule sometime soon,
2111 * else this function might spin for a *long* time. This function can't
2112 * be called with interrupts off, or it may introduce deadlock with
2113 * smp_call_function() if an IPI is sent by the same process we are
2114 * waiting to become inactive.
2116 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2118 unsigned long flags;
2125 * We do the initial early heuristics without holding
2126 * any task-queue locks at all. We'll only try to get
2127 * the runqueue lock when things look like they will
2133 * If the task is actively running on another CPU
2134 * still, just relax and busy-wait without holding
2137 * NOTE! Since we don't hold any locks, it's not
2138 * even sure that "rq" stays as the right runqueue!
2139 * But we don't care, since "task_running()" will
2140 * return false if the runqueue has changed and p
2141 * is actually now running somewhere else!
2143 while (task_running(rq, p)) {
2144 if (match_state && unlikely(p->state != match_state))
2150 * Ok, time to look more closely! We need the rq
2151 * lock now, to be *sure*. If we're wrong, we'll
2152 * just go back and repeat.
2154 rq = task_rq_lock(p, &flags);
2155 trace_sched_wait_task(rq, p);
2156 running = task_running(rq, p);
2157 on_rq = p->se.on_rq;
2159 if (!match_state || p->state == match_state)
2160 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2161 task_rq_unlock(rq, &flags);
2164 * If it changed from the expected state, bail out now.
2166 if (unlikely(!ncsw))
2170 * Was it really running after all now that we
2171 * checked with the proper locks actually held?
2173 * Oops. Go back and try again..
2175 if (unlikely(running)) {
2181 * It's not enough that it's not actively running,
2182 * it must be off the runqueue _entirely_, and not
2185 * So if it was still runnable (but just not actively
2186 * running right now), it's preempted, and we should
2187 * yield - it could be a while.
2189 if (unlikely(on_rq)) {
2190 schedule_timeout_uninterruptible(1);
2195 * Ahh, all good. It wasn't running, and it wasn't
2196 * runnable, which means that it will never become
2197 * running in the future either. We're all done!
2206 * kick_process - kick a running thread to enter/exit the kernel
2207 * @p: the to-be-kicked thread
2209 * Cause a process which is running on another CPU to enter
2210 * kernel-mode, without any delay. (to get signals handled.)
2212 * NOTE: this function doesnt have to take the runqueue lock,
2213 * because all it wants to ensure is that the remote task enters
2214 * the kernel. If the IPI races and the task has been migrated
2215 * to another CPU then no harm is done and the purpose has been
2218 void kick_process(struct task_struct *p)
2224 if ((cpu != smp_processor_id()) && task_curr(p))
2225 smp_send_reschedule(cpu);
2228 EXPORT_SYMBOL_GPL(kick_process);
2231 * Return a low guess at the load of a migration-source cpu weighted
2232 * according to the scheduling class and "nice" value.
2234 * We want to under-estimate the load of migration sources, to
2235 * balance conservatively.
2237 static unsigned long source_load(int cpu, int type)
2239 struct rq *rq = cpu_rq(cpu);
2240 unsigned long total = weighted_cpuload(cpu);
2242 if (type == 0 || !sched_feat(LB_BIAS))
2245 return min(rq->cpu_load[type-1], total);
2249 * Return a high guess at the load of a migration-target cpu weighted
2250 * according to the scheduling class and "nice" value.
2252 static unsigned long target_load(int cpu, int type)
2254 struct rq *rq = cpu_rq(cpu);
2255 unsigned long total = weighted_cpuload(cpu);
2257 if (type == 0 || !sched_feat(LB_BIAS))
2260 return max(rq->cpu_load[type-1], total);
2264 * find_idlest_group finds and returns the least busy CPU group within the
2267 static struct sched_group *
2268 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2270 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2271 unsigned long min_load = ULONG_MAX, this_load = 0;
2272 int load_idx = sd->forkexec_idx;
2273 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2276 unsigned long load, avg_load;
2280 /* Skip over this group if it has no CPUs allowed */
2281 if (!cpumask_intersects(sched_group_cpus(group),
2285 local_group = cpumask_test_cpu(this_cpu,
2286 sched_group_cpus(group));
2288 /* Tally up the load of all CPUs in the group */
2291 for_each_cpu(i, sched_group_cpus(group)) {
2292 /* Bias balancing toward cpus of our domain */
2294 load = source_load(i, load_idx);
2296 load = target_load(i, load_idx);
2301 /* Adjust by relative CPU power of the group */
2302 avg_load = sg_div_cpu_power(group,
2303 avg_load * SCHED_LOAD_SCALE);
2306 this_load = avg_load;
2308 } else if (avg_load < min_load) {
2309 min_load = avg_load;
2312 } while (group = group->next, group != sd->groups);
2314 if (!idlest || 100*this_load < imbalance*min_load)
2320 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2323 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2325 unsigned long load, min_load = ULONG_MAX;
2329 /* Traverse only the allowed CPUs */
2330 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2331 load = weighted_cpuload(i);
2333 if (load < min_load || (load == min_load && i == this_cpu)) {
2343 * sched_balance_self: balance the current task (running on cpu) in domains
2344 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2347 * Balance, ie. select the least loaded group.
2349 * Returns the target CPU number, or the same CPU if no balancing is needed.
2351 * preempt must be disabled.
2353 static int sched_balance_self(int cpu, int flag)
2355 struct task_struct *t = current;
2356 struct sched_domain *tmp, *sd = NULL;
2358 for_each_domain(cpu, tmp) {
2360 * If power savings logic is enabled for a domain, stop there.
2362 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2364 if (tmp->flags & flag)
2372 struct sched_group *group;
2373 int new_cpu, weight;
2375 if (!(sd->flags & flag)) {
2380 group = find_idlest_group(sd, t, cpu);
2386 new_cpu = find_idlest_cpu(group, t, cpu);
2387 if (new_cpu == -1 || new_cpu == cpu) {
2388 /* Now try balancing at a lower domain level of cpu */
2393 /* Now try balancing at a lower domain level of new_cpu */
2395 weight = cpumask_weight(sched_domain_span(sd));
2397 for_each_domain(cpu, tmp) {
2398 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2400 if (tmp->flags & flag)
2403 /* while loop will break here if sd == NULL */
2409 #endif /* CONFIG_SMP */
2412 * task_oncpu_function_call - call a function on the cpu on which a task runs
2413 * @p: the task to evaluate
2414 * @func: the function to be called
2415 * @info: the function call argument
2417 * Calls the function @func when the task is currently running. This might
2418 * be on the current CPU, which just calls the function directly
2420 void task_oncpu_function_call(struct task_struct *p,
2421 void (*func) (void *info), void *info)
2428 smp_call_function_single(cpu, func, info, 1);
2433 * try_to_wake_up - wake up a thread
2434 * @p: the to-be-woken-up thread
2435 * @state: the mask of task states that can be woken
2436 * @sync: do a synchronous wakeup?
2438 * Put it on the run-queue if it's not already there. The "current"
2439 * thread is always on the run-queue (except when the actual
2440 * re-schedule is in progress), and as such you're allowed to do
2441 * the simpler "current->state = TASK_RUNNING" to mark yourself
2442 * runnable without the overhead of this.
2444 * returns failure only if the task is already active.
2446 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2448 int cpu, orig_cpu, this_cpu, success = 0;
2449 unsigned long flags;
2453 if (!sched_feat(SYNC_WAKEUPS))
2457 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2458 struct sched_domain *sd;
2460 this_cpu = raw_smp_processor_id();
2463 for_each_domain(this_cpu, sd) {
2464 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2473 rq = task_rq_lock(p, &flags);
2474 update_rq_clock(rq);
2475 old_state = p->state;
2476 if (!(old_state & state))
2484 this_cpu = smp_processor_id();
2487 if (unlikely(task_running(rq, p)))
2490 cpu = p->sched_class->select_task_rq(p, sync);
2491 if (cpu != orig_cpu) {
2492 set_task_cpu(p, cpu);
2493 task_rq_unlock(rq, &flags);
2494 /* might preempt at this point */
2495 rq = task_rq_lock(p, &flags);
2496 old_state = p->state;
2497 if (!(old_state & state))
2502 this_cpu = smp_processor_id();
2506 #ifdef CONFIG_SCHEDSTATS
2507 schedstat_inc(rq, ttwu_count);
2508 if (cpu == this_cpu)
2509 schedstat_inc(rq, ttwu_local);
2511 struct sched_domain *sd;
2512 for_each_domain(this_cpu, sd) {
2513 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2514 schedstat_inc(sd, ttwu_wake_remote);
2519 #endif /* CONFIG_SCHEDSTATS */
2522 #endif /* CONFIG_SMP */
2523 schedstat_inc(p, se.nr_wakeups);
2525 schedstat_inc(p, se.nr_wakeups_sync);
2526 if (orig_cpu != cpu)
2527 schedstat_inc(p, se.nr_wakeups_migrate);
2528 if (cpu == this_cpu)
2529 schedstat_inc(p, se.nr_wakeups_local);
2531 schedstat_inc(p, se.nr_wakeups_remote);
2532 activate_task(rq, p, 1);
2536 * Only attribute actual wakeups done by this task.
2538 if (!in_interrupt()) {
2539 struct sched_entity *se = ¤t->se;
2540 u64 sample = se->sum_exec_runtime;
2542 if (se->last_wakeup)
2543 sample -= se->last_wakeup;
2545 sample -= se->start_runtime;
2546 update_avg(&se->avg_wakeup, sample);
2548 se->last_wakeup = se->sum_exec_runtime;
2552 trace_sched_wakeup(rq, p, success);
2553 check_preempt_curr(rq, p, sync);
2555 p->state = TASK_RUNNING;
2557 if (p->sched_class->task_wake_up)
2558 p->sched_class->task_wake_up(rq, p);
2561 task_rq_unlock(rq, &flags);
2567 * wake_up_process - Wake up a specific process
2568 * @p: The process to be woken up.
2570 * Attempt to wake up the nominated process and move it to the set of runnable
2571 * processes. Returns 1 if the process was woken up, 0 if it was already
2574 * It may be assumed that this function implies a write memory barrier before
2575 * changing the task state if and only if any tasks are woken up.
2577 int wake_up_process(struct task_struct *p)
2579 return try_to_wake_up(p, TASK_ALL, 0);
2581 EXPORT_SYMBOL(wake_up_process);
2583 int wake_up_state(struct task_struct *p, unsigned int state)
2585 return try_to_wake_up(p, state, 0);
2589 * Perform scheduler related setup for a newly forked process p.
2590 * p is forked by current.
2592 * __sched_fork() is basic setup used by init_idle() too:
2594 static void __sched_fork(struct task_struct *p)
2596 p->se.exec_start = 0;
2597 p->se.sum_exec_runtime = 0;
2598 p->se.prev_sum_exec_runtime = 0;
2599 p->se.nr_migrations = 0;
2600 p->se.last_wakeup = 0;
2601 p->se.avg_overlap = 0;
2602 p->se.start_runtime = 0;
2603 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2605 #ifdef CONFIG_SCHEDSTATS
2606 p->se.wait_start = 0;
2608 p->se.wait_count = 0;
2611 p->se.sleep_start = 0;
2612 p->se.sleep_max = 0;
2613 p->se.sum_sleep_runtime = 0;
2615 p->se.block_start = 0;
2616 p->se.block_max = 0;
2618 p->se.slice_max = 0;
2620 p->se.nr_migrations_cold = 0;
2621 p->se.nr_failed_migrations_affine = 0;
2622 p->se.nr_failed_migrations_running = 0;
2623 p->se.nr_failed_migrations_hot = 0;
2624 p->se.nr_forced_migrations = 0;
2625 p->se.nr_forced2_migrations = 0;
2627 p->se.nr_wakeups = 0;
2628 p->se.nr_wakeups_sync = 0;
2629 p->se.nr_wakeups_migrate = 0;
2630 p->se.nr_wakeups_local = 0;
2631 p->se.nr_wakeups_remote = 0;
2632 p->se.nr_wakeups_affine = 0;
2633 p->se.nr_wakeups_affine_attempts = 0;
2634 p->se.nr_wakeups_passive = 0;
2635 p->se.nr_wakeups_idle = 0;
2639 INIT_LIST_HEAD(&p->rt.run_list);
2641 INIT_LIST_HEAD(&p->se.group_node);
2643 #ifdef CONFIG_PREEMPT_NOTIFIERS
2644 INIT_HLIST_HEAD(&p->preempt_notifiers);
2648 * We mark the process as running here, but have not actually
2649 * inserted it onto the runqueue yet. This guarantees that
2650 * nobody will actually run it, and a signal or other external
2651 * event cannot wake it up and insert it on the runqueue either.
2653 p->state = TASK_RUNNING;
2657 * fork()/clone()-time setup:
2659 void sched_fork(struct task_struct *p, int clone_flags)
2661 int cpu = get_cpu();
2666 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2668 set_task_cpu(p, cpu);
2671 * Make sure we do not leak PI boosting priority to the child.
2673 p->prio = current->normal_prio;
2676 * Revert to default priority/policy on fork if requested.
2678 if (unlikely(p->sched_reset_on_fork)) {
2679 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR)
2680 p->policy = SCHED_NORMAL;
2682 if (p->normal_prio < DEFAULT_PRIO)
2683 p->prio = DEFAULT_PRIO;
2685 if (PRIO_TO_NICE(p->static_prio) < 0) {
2686 p->static_prio = NICE_TO_PRIO(0);
2691 * We don't need the reset flag anymore after the fork. It has
2692 * fulfilled its duty:
2694 p->sched_reset_on_fork = 0;
2697 if (!rt_prio(p->prio))
2698 p->sched_class = &fair_sched_class;
2700 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2701 if (likely(sched_info_on()))
2702 memset(&p->sched_info, 0, sizeof(p->sched_info));
2704 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2707 #ifdef CONFIG_PREEMPT
2708 /* Want to start with kernel preemption disabled. */
2709 task_thread_info(p)->preempt_count = 1;
2711 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2717 * wake_up_new_task - wake up a newly created task for the first time.
2719 * This function will do some initial scheduler statistics housekeeping
2720 * that must be done for every newly created context, then puts the task
2721 * on the runqueue and wakes it.
2723 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2725 unsigned long flags;
2728 rq = task_rq_lock(p, &flags);
2729 BUG_ON(p->state != TASK_RUNNING);
2730 update_rq_clock(rq);
2732 p->prio = effective_prio(p);
2734 if (!p->sched_class->task_new || !current->se.on_rq) {
2735 activate_task(rq, p, 0);
2738 * Let the scheduling class do new task startup
2739 * management (if any):
2741 p->sched_class->task_new(rq, p);
2744 trace_sched_wakeup_new(rq, p, 1);
2745 check_preempt_curr(rq, p, 0);
2747 if (p->sched_class->task_wake_up)
2748 p->sched_class->task_wake_up(rq, p);
2750 task_rq_unlock(rq, &flags);
2753 #ifdef CONFIG_PREEMPT_NOTIFIERS
2756 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2757 * @notifier: notifier struct to register
2759 void preempt_notifier_register(struct preempt_notifier *notifier)
2761 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2763 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2766 * preempt_notifier_unregister - no longer interested in preemption notifications
2767 * @notifier: notifier struct to unregister
2769 * This is safe to call from within a preemption notifier.
2771 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2773 hlist_del(¬ifier->link);
2775 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2777 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2779 struct preempt_notifier *notifier;
2780 struct hlist_node *node;
2782 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2783 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2787 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2788 struct task_struct *next)
2790 struct preempt_notifier *notifier;
2791 struct hlist_node *node;
2793 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2794 notifier->ops->sched_out(notifier, next);
2797 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2799 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2804 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2805 struct task_struct *next)
2809 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2812 * prepare_task_switch - prepare to switch tasks
2813 * @rq: the runqueue preparing to switch
2814 * @prev: the current task that is being switched out
2815 * @next: the task we are going to switch to.
2817 * This is called with the rq lock held and interrupts off. It must
2818 * be paired with a subsequent finish_task_switch after the context
2821 * prepare_task_switch sets up locking and calls architecture specific
2825 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2826 struct task_struct *next)
2828 fire_sched_out_preempt_notifiers(prev, next);
2829 prepare_lock_switch(rq, next);
2830 prepare_arch_switch(next);
2834 * finish_task_switch - clean up after a task-switch
2835 * @rq: runqueue associated with task-switch
2836 * @prev: the thread we just switched away from.
2838 * finish_task_switch must be called after the context switch, paired
2839 * with a prepare_task_switch call before the context switch.
2840 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2841 * and do any other architecture-specific cleanup actions.
2843 * Note that we may have delayed dropping an mm in context_switch(). If
2844 * so, we finish that here outside of the runqueue lock. (Doing it
2845 * with the lock held can cause deadlocks; see schedule() for
2848 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2849 __releases(rq->lock)
2851 struct mm_struct *mm = rq->prev_mm;
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);
2887 /* assumes rq->lock is held */
2888 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2890 if (prev->sched_class->pre_schedule)
2891 prev->sched_class->pre_schedule(rq, prev);
2894 /* rq->lock is NOT held, but preemption is disabled */
2895 static inline void post_schedule(struct rq *rq)
2897 if (rq->post_schedule) {
2898 unsigned long flags;
2900 spin_lock_irqsave(&rq->lock, flags);
2901 if (rq->curr->sched_class->post_schedule)
2902 rq->curr->sched_class->post_schedule(rq);
2903 spin_unlock_irqrestore(&rq->lock, flags);
2905 rq->post_schedule = 0;
2911 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2915 static inline void post_schedule(struct rq *rq)
2922 * schedule_tail - first thing a freshly forked thread must call.
2923 * @prev: the thread we just switched away from.
2925 asmlinkage void schedule_tail(struct task_struct *prev)
2926 __releases(rq->lock)
2928 struct rq *rq = this_rq();
2930 finish_task_switch(rq, prev);
2933 * FIXME: do we need to worry about rq being invalidated by the
2938 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2939 /* In this case, finish_task_switch does not reenable preemption */
2942 if (current->set_child_tid)
2943 put_user(task_pid_vnr(current), current->set_child_tid);
2947 * context_switch - switch to the new MM and the new
2948 * thread's register state.
2951 context_switch(struct rq *rq, struct task_struct *prev,
2952 struct task_struct *next)
2954 struct mm_struct *mm, *oldmm;
2956 prepare_task_switch(rq, prev, next);
2957 trace_sched_switch(rq, prev, next);
2959 oldmm = prev->active_mm;
2961 * For paravirt, this is coupled with an exit in switch_to to
2962 * combine the page table reload and the switch backend into
2965 arch_start_context_switch(prev);
2967 if (unlikely(!mm)) {
2968 next->active_mm = oldmm;
2969 atomic_inc(&oldmm->mm_count);
2970 enter_lazy_tlb(oldmm, next);
2972 switch_mm(oldmm, mm, next);
2974 if (unlikely(!prev->mm)) {
2975 prev->active_mm = NULL;
2976 rq->prev_mm = oldmm;
2979 * Since the runqueue lock will be released by the next
2980 * task (which is an invalid locking op but in the case
2981 * of the scheduler it's an obvious special-case), so we
2982 * do an early lockdep release here:
2984 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2985 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2988 /* Here we just switch the register state and the stack. */
2989 switch_to(prev, next, prev);
2993 * this_rq must be evaluated again because prev may have moved
2994 * CPUs since it called schedule(), thus the 'rq' on its stack
2995 * frame will be invalid.
2997 finish_task_switch(this_rq(), prev);
3001 * nr_running, nr_uninterruptible and nr_context_switches:
3003 * externally visible scheduler statistics: current number of runnable
3004 * threads, current number of uninterruptible-sleeping threads, total
3005 * number of context switches performed since bootup.
3007 unsigned long nr_running(void)
3009 unsigned long i, sum = 0;
3011 for_each_online_cpu(i)
3012 sum += cpu_rq(i)->nr_running;
3017 unsigned long nr_uninterruptible(void)
3019 unsigned long i, sum = 0;
3021 for_each_possible_cpu(i)
3022 sum += cpu_rq(i)->nr_uninterruptible;
3025 * Since we read the counters lockless, it might be slightly
3026 * inaccurate. Do not allow it to go below zero though:
3028 if (unlikely((long)sum < 0))
3034 unsigned long long nr_context_switches(void)
3037 unsigned long long sum = 0;
3039 for_each_possible_cpu(i)
3040 sum += cpu_rq(i)->nr_switches;
3045 unsigned long nr_iowait(void)
3047 unsigned long i, sum = 0;
3049 for_each_possible_cpu(i)
3050 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3055 /* Variables and functions for calc_load */
3056 static atomic_long_t calc_load_tasks;
3057 static unsigned long calc_load_update;
3058 unsigned long avenrun[3];
3059 EXPORT_SYMBOL(avenrun);
3062 * get_avenrun - get the load average array
3063 * @loads: pointer to dest load array
3064 * @offset: offset to add
3065 * @shift: shift count to shift the result left
3067 * These values are estimates at best, so no need for locking.
3069 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3071 loads[0] = (avenrun[0] + offset) << shift;
3072 loads[1] = (avenrun[1] + offset) << shift;
3073 loads[2] = (avenrun[2] + offset) << shift;
3076 static unsigned long
3077 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3080 load += active * (FIXED_1 - exp);
3081 return load >> FSHIFT;
3085 * calc_load - update the avenrun load estimates 10 ticks after the
3086 * CPUs have updated calc_load_tasks.
3088 void calc_global_load(void)
3090 unsigned long upd = calc_load_update + 10;
3093 if (time_before(jiffies, upd))
3096 active = atomic_long_read(&calc_load_tasks);
3097 active = active > 0 ? active * FIXED_1 : 0;
3099 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3100 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3101 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3103 calc_load_update += LOAD_FREQ;
3107 * Either called from update_cpu_load() or from a cpu going idle
3109 static void calc_load_account_active(struct rq *this_rq)
3111 long nr_active, delta;
3113 nr_active = this_rq->nr_running;
3114 nr_active += (long) this_rq->nr_uninterruptible;
3116 if (nr_active != this_rq->calc_load_active) {
3117 delta = nr_active - this_rq->calc_load_active;
3118 this_rq->calc_load_active = nr_active;
3119 atomic_long_add(delta, &calc_load_tasks);
3124 * Externally visible per-cpu scheduler statistics:
3125 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3127 u64 cpu_nr_migrations(int cpu)
3129 return cpu_rq(cpu)->nr_migrations_in;
3133 * Update rq->cpu_load[] statistics. This function is usually called every
3134 * scheduler tick (TICK_NSEC).
3136 static void update_cpu_load(struct rq *this_rq)
3138 unsigned long this_load = this_rq->load.weight;
3141 this_rq->nr_load_updates++;
3143 /* Update our load: */
3144 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3145 unsigned long old_load, new_load;
3147 /* scale is effectively 1 << i now, and >> i divides by scale */
3149 old_load = this_rq->cpu_load[i];
3150 new_load = this_load;
3152 * Round up the averaging division if load is increasing. This
3153 * prevents us from getting stuck on 9 if the load is 10, for
3156 if (new_load > old_load)
3157 new_load += scale-1;
3158 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3161 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3162 this_rq->calc_load_update += LOAD_FREQ;
3163 calc_load_account_active(this_rq);
3170 * double_rq_lock - safely lock two runqueues
3172 * Note this does not disable interrupts like task_rq_lock,
3173 * you need to do so manually before calling.
3175 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3176 __acquires(rq1->lock)
3177 __acquires(rq2->lock)
3179 BUG_ON(!irqs_disabled());
3181 spin_lock(&rq1->lock);
3182 __acquire(rq2->lock); /* Fake it out ;) */
3185 spin_lock(&rq1->lock);
3186 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3188 spin_lock(&rq2->lock);
3189 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3192 update_rq_clock(rq1);
3193 update_rq_clock(rq2);
3197 * double_rq_unlock - safely unlock two runqueues
3199 * Note this does not restore interrupts like task_rq_unlock,
3200 * you need to do so manually after calling.
3202 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3203 __releases(rq1->lock)
3204 __releases(rq2->lock)
3206 spin_unlock(&rq1->lock);
3208 spin_unlock(&rq2->lock);
3210 __release(rq2->lock);
3214 * If dest_cpu is allowed for this process, migrate the task to it.
3215 * This is accomplished by forcing the cpu_allowed mask to only
3216 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3217 * the cpu_allowed mask is restored.
3219 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3221 struct migration_req req;
3222 unsigned long flags;
3225 rq = task_rq_lock(p, &flags);
3226 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3227 || unlikely(!cpu_active(dest_cpu)))
3230 /* force the process onto the specified CPU */
3231 if (migrate_task(p, dest_cpu, &req)) {
3232 /* Need to wait for migration thread (might exit: take ref). */
3233 struct task_struct *mt = rq->migration_thread;
3235 get_task_struct(mt);
3236 task_rq_unlock(rq, &flags);
3237 wake_up_process(mt);
3238 put_task_struct(mt);
3239 wait_for_completion(&req.done);
3244 task_rq_unlock(rq, &flags);
3248 * sched_exec - execve() is a valuable balancing opportunity, because at
3249 * this point the task has the smallest effective memory and cache footprint.
3251 void sched_exec(void)
3253 int new_cpu, this_cpu = get_cpu();
3254 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3256 if (new_cpu != this_cpu)
3257 sched_migrate_task(current, new_cpu);
3261 * pull_task - move a task from a remote runqueue to the local runqueue.
3262 * Both runqueues must be locked.
3264 static void pull_task(struct rq *src_rq, struct task_struct *p,
3265 struct rq *this_rq, int this_cpu)
3267 deactivate_task(src_rq, p, 0);
3268 set_task_cpu(p, this_cpu);
3269 activate_task(this_rq, p, 0);
3271 * Note that idle threads have a prio of MAX_PRIO, for this test
3272 * to be always true for them.
3274 check_preempt_curr(this_rq, p, 0);
3278 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3281 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3282 struct sched_domain *sd, enum cpu_idle_type idle,
3285 int tsk_cache_hot = 0;
3287 * We do not migrate tasks that are:
3288 * 1) running (obviously), or
3289 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3290 * 3) are cache-hot on their current CPU.
3292 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3293 schedstat_inc(p, se.nr_failed_migrations_affine);
3298 if (task_running(rq, p)) {
3299 schedstat_inc(p, se.nr_failed_migrations_running);
3304 * Aggressive migration if:
3305 * 1) task is cache cold, or
3306 * 2) too many balance attempts have failed.
3309 tsk_cache_hot = task_hot(p, rq->clock, sd);
3310 if (!tsk_cache_hot ||
3311 sd->nr_balance_failed > sd->cache_nice_tries) {
3312 #ifdef CONFIG_SCHEDSTATS
3313 if (tsk_cache_hot) {
3314 schedstat_inc(sd, lb_hot_gained[idle]);
3315 schedstat_inc(p, se.nr_forced_migrations);
3321 if (tsk_cache_hot) {
3322 schedstat_inc(p, se.nr_failed_migrations_hot);
3328 static unsigned long
3329 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3330 unsigned long max_load_move, struct sched_domain *sd,
3331 enum cpu_idle_type idle, int *all_pinned,
3332 int *this_best_prio, struct rq_iterator *iterator)
3334 int loops = 0, pulled = 0, pinned = 0;
3335 struct task_struct *p;
3336 long rem_load_move = max_load_move;
3338 if (max_load_move == 0)
3344 * Start the load-balancing iterator:
3346 p = iterator->start(iterator->arg);
3348 if (!p || loops++ > sysctl_sched_nr_migrate)
3351 if ((p->se.load.weight >> 1) > rem_load_move ||
3352 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3353 p = iterator->next(iterator->arg);
3357 pull_task(busiest, p, this_rq, this_cpu);
3359 rem_load_move -= p->se.load.weight;
3361 #ifdef CONFIG_PREEMPT
3363 * NEWIDLE balancing is a source of latency, so preemptible kernels
3364 * will stop after the first task is pulled to minimize the critical
3367 if (idle == CPU_NEWLY_IDLE)
3372 * We only want to steal up to the prescribed amount of weighted load.
3374 if (rem_load_move > 0) {
3375 if (p->prio < *this_best_prio)
3376 *this_best_prio = p->prio;
3377 p = iterator->next(iterator->arg);
3382 * Right now, this is one of only two places pull_task() is called,
3383 * so we can safely collect pull_task() stats here rather than
3384 * inside pull_task().
3386 schedstat_add(sd, lb_gained[idle], pulled);
3389 *all_pinned = pinned;
3391 return max_load_move - rem_load_move;
3395 * move_tasks tries to move up to max_load_move weighted load from busiest to
3396 * this_rq, as part of a balancing operation within domain "sd".
3397 * Returns 1 if successful and 0 otherwise.
3399 * Called with both runqueues locked.
3401 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3402 unsigned long max_load_move,
3403 struct sched_domain *sd, enum cpu_idle_type idle,
3406 const struct sched_class *class = sched_class_highest;
3407 unsigned long total_load_moved = 0;
3408 int this_best_prio = this_rq->curr->prio;
3412 class->load_balance(this_rq, this_cpu, busiest,
3413 max_load_move - total_load_moved,
3414 sd, idle, all_pinned, &this_best_prio);
3415 class = class->next;
3417 #ifdef CONFIG_PREEMPT
3419 * NEWIDLE balancing is a source of latency, so preemptible
3420 * kernels will stop after the first task is pulled to minimize
3421 * the critical section.
3423 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3426 } while (class && max_load_move > total_load_moved);
3428 return total_load_moved > 0;
3432 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3433 struct sched_domain *sd, enum cpu_idle_type idle,
3434 struct rq_iterator *iterator)
3436 struct task_struct *p = iterator->start(iterator->arg);
3440 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3441 pull_task(busiest, p, this_rq, this_cpu);
3443 * Right now, this is only the second place pull_task()
3444 * is called, so we can safely collect pull_task()
3445 * stats here rather than inside pull_task().
3447 schedstat_inc(sd, lb_gained[idle]);
3451 p = iterator->next(iterator->arg);
3458 * move_one_task tries to move exactly one task from busiest to this_rq, as
3459 * part of active balancing operations within "domain".
3460 * Returns 1 if successful and 0 otherwise.
3462 * Called with both runqueues locked.
3464 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3465 struct sched_domain *sd, enum cpu_idle_type idle)
3467 const struct sched_class *class;
3469 for_each_class(class) {
3470 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3476 /********** Helpers for find_busiest_group ************************/
3478 * sd_lb_stats - Structure to store the statistics of a sched_domain
3479 * during load balancing.
3481 struct sd_lb_stats {
3482 struct sched_group *busiest; /* Busiest group in this sd */
3483 struct sched_group *this; /* Local group in this sd */
3484 unsigned long total_load; /* Total load of all groups in sd */
3485 unsigned long total_pwr; /* Total power of all groups in sd */
3486 unsigned long avg_load; /* Average load across all groups in sd */
3488 /** Statistics of this group */
3489 unsigned long this_load;
3490 unsigned long this_load_per_task;
3491 unsigned long this_nr_running;
3493 /* Statistics of the busiest group */
3494 unsigned long max_load;
3495 unsigned long busiest_load_per_task;
3496 unsigned long busiest_nr_running;
3498 int group_imb; /* Is there imbalance in this sd */
3499 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3500 int power_savings_balance; /* Is powersave balance needed for this sd */
3501 struct sched_group *group_min; /* Least loaded group in sd */
3502 struct sched_group *group_leader; /* Group which relieves group_min */
3503 unsigned long min_load_per_task; /* load_per_task in group_min */
3504 unsigned long leader_nr_running; /* Nr running of group_leader */
3505 unsigned long min_nr_running; /* Nr running of group_min */
3510 * sg_lb_stats - stats of a sched_group required for load_balancing
3512 struct sg_lb_stats {
3513 unsigned long avg_load; /*Avg load across the CPUs of the group */
3514 unsigned long group_load; /* Total load over the CPUs of the group */
3515 unsigned long sum_nr_running; /* Nr tasks running in the group */
3516 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3517 unsigned long group_capacity;
3518 int group_imb; /* Is there an imbalance in the group ? */
3522 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3523 * @group: The group whose first cpu is to be returned.
3525 static inline unsigned int group_first_cpu(struct sched_group *group)
3527 return cpumask_first(sched_group_cpus(group));
3531 * get_sd_load_idx - Obtain the load index for a given sched domain.
3532 * @sd: The sched_domain whose load_idx is to be obtained.
3533 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3535 static inline int get_sd_load_idx(struct sched_domain *sd,
3536 enum cpu_idle_type idle)
3542 load_idx = sd->busy_idx;
3545 case CPU_NEWLY_IDLE:
3546 load_idx = sd->newidle_idx;
3549 load_idx = sd->idle_idx;
3557 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3559 * init_sd_power_savings_stats - Initialize power savings statistics for
3560 * the given sched_domain, during load balancing.
3562 * @sd: Sched domain whose power-savings statistics are to be initialized.
3563 * @sds: Variable containing the statistics for sd.
3564 * @idle: Idle status of the CPU at which we're performing load-balancing.
3566 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3567 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3570 * Busy processors will not participate in power savings
3573 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3574 sds->power_savings_balance = 0;
3576 sds->power_savings_balance = 1;
3577 sds->min_nr_running = ULONG_MAX;
3578 sds->leader_nr_running = 0;
3583 * update_sd_power_savings_stats - Update the power saving stats for a
3584 * sched_domain while performing load balancing.
3586 * @group: sched_group belonging to the sched_domain under consideration.
3587 * @sds: Variable containing the statistics of the sched_domain
3588 * @local_group: Does group contain the CPU for which we're performing
3590 * @sgs: Variable containing the statistics of the group.
3592 static inline void update_sd_power_savings_stats(struct sched_group *group,
3593 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3596 if (!sds->power_savings_balance)
3600 * If the local group is idle or completely loaded
3601 * no need to do power savings balance at this domain
3603 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3604 !sds->this_nr_running))
3605 sds->power_savings_balance = 0;
3608 * If a group is already running at full capacity or idle,
3609 * don't include that group in power savings calculations
3611 if (!sds->power_savings_balance ||
3612 sgs->sum_nr_running >= sgs->group_capacity ||
3613 !sgs->sum_nr_running)
3617 * Calculate the group which has the least non-idle load.
3618 * This is the group from where we need to pick up the load
3621 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3622 (sgs->sum_nr_running == sds->min_nr_running &&
3623 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3624 sds->group_min = group;
3625 sds->min_nr_running = sgs->sum_nr_running;
3626 sds->min_load_per_task = sgs->sum_weighted_load /
3627 sgs->sum_nr_running;
3631 * Calculate the group which is almost near its
3632 * capacity but still has some space to pick up some load
3633 * from other group and save more power
3635 if (sgs->sum_nr_running > sgs->group_capacity - 1)
3638 if (sgs->sum_nr_running > sds->leader_nr_running ||
3639 (sgs->sum_nr_running == sds->leader_nr_running &&
3640 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3641 sds->group_leader = group;
3642 sds->leader_nr_running = sgs->sum_nr_running;
3647 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3648 * @sds: Variable containing the statistics of the sched_domain
3649 * under consideration.
3650 * @this_cpu: Cpu at which we're currently performing load-balancing.
3651 * @imbalance: Variable to store the imbalance.
3654 * Check if we have potential to perform some power-savings balance.
3655 * If yes, set the busiest group to be the least loaded group in the
3656 * sched_domain, so that it's CPUs can be put to idle.
3658 * Returns 1 if there is potential to perform power-savings balance.
3661 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3662 int this_cpu, unsigned long *imbalance)
3664 if (!sds->power_savings_balance)
3667 if (sds->this != sds->group_leader ||
3668 sds->group_leader == sds->group_min)
3671 *imbalance = sds->min_load_per_task;
3672 sds->busiest = sds->group_min;
3674 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3675 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3676 group_first_cpu(sds->group_leader);
3682 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3683 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3684 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3689 static inline void update_sd_power_savings_stats(struct sched_group *group,
3690 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3695 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3696 int this_cpu, unsigned long *imbalance)
3700 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3702 unsigned long __weak arch_smt_gain(struct sched_domain *sd, int cpu)
3704 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3705 unsigned long smt_gain = sd->smt_gain;
3712 static void update_cpu_power(struct sched_domain *sd, int cpu)
3714 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3715 unsigned long power = SCHED_LOAD_SCALE;
3716 struct sched_group *sdg = sd->groups;
3717 unsigned long old = sdg->__cpu_power;
3719 /* here we could scale based on cpufreq */
3721 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3722 power *= arch_smt_gain(sd, cpu);
3723 power >>= SCHED_LOAD_SHIFT;
3726 /* here we could scale based on RT time */
3729 sdg->__cpu_power = power;
3730 sdg->reciprocal_cpu_power = reciprocal_value(power);
3734 static void update_group_power(struct sched_domain *sd, int cpu)
3736 struct sched_domain *child = sd->child;
3737 struct sched_group *group, *sdg = sd->groups;
3738 unsigned long power = sdg->__cpu_power;
3741 update_cpu_power(sd, cpu);
3745 sdg->__cpu_power = 0;
3747 group = child->groups;
3749 sdg->__cpu_power += group->__cpu_power;
3750 group = group->next;
3751 } while (group != child->groups);
3753 if (power != sdg->__cpu_power)
3754 sdg->reciprocal_cpu_power = reciprocal_value(sdg->__cpu_power);
3758 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3759 * @group: sched_group whose statistics are to be updated.
3760 * @this_cpu: Cpu for which load balance is currently performed.
3761 * @idle: Idle status of this_cpu
3762 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3763 * @sd_idle: Idle status of the sched_domain containing group.
3764 * @local_group: Does group contain this_cpu.
3765 * @cpus: Set of cpus considered for load balancing.
3766 * @balance: Should we balance.
3767 * @sgs: variable to hold the statistics for this group.
3769 static inline void update_sg_lb_stats(struct sched_domain *sd,
3770 struct sched_group *group, int this_cpu,
3771 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3772 int local_group, const struct cpumask *cpus,
3773 int *balance, struct sg_lb_stats *sgs)
3775 unsigned long load, max_cpu_load, min_cpu_load;
3777 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3778 unsigned long sum_avg_load_per_task;
3779 unsigned long avg_load_per_task;
3782 balance_cpu = group_first_cpu(group);
3783 if (balance_cpu == this_cpu)
3784 update_group_power(sd, this_cpu);
3787 /* Tally up the load of all CPUs in the group */
3788 sum_avg_load_per_task = avg_load_per_task = 0;
3790 min_cpu_load = ~0UL;
3792 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3793 struct rq *rq = cpu_rq(i);
3795 if (*sd_idle && rq->nr_running)
3798 /* Bias balancing toward cpus of our domain */
3800 if (idle_cpu(i) && !first_idle_cpu) {
3805 load = target_load(i, load_idx);
3807 load = source_load(i, load_idx);
3808 if (load > max_cpu_load)
3809 max_cpu_load = load;
3810 if (min_cpu_load > load)
3811 min_cpu_load = load;
3814 sgs->group_load += load;
3815 sgs->sum_nr_running += rq->nr_running;
3816 sgs->sum_weighted_load += weighted_cpuload(i);
3818 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3822 * First idle cpu or the first cpu(busiest) in this sched group
3823 * is eligible for doing load balancing at this and above
3824 * domains. In the newly idle case, we will allow all the cpu's
3825 * to do the newly idle load balance.
3827 if (idle != CPU_NEWLY_IDLE && local_group &&
3828 balance_cpu != this_cpu && balance) {
3833 /* Adjust by relative CPU power of the group */
3834 sgs->avg_load = sg_div_cpu_power(group,
3835 sgs->group_load * SCHED_LOAD_SCALE);
3839 * Consider the group unbalanced when the imbalance is larger
3840 * than the average weight of two tasks.
3842 * APZ: with cgroup the avg task weight can vary wildly and
3843 * might not be a suitable number - should we keep a
3844 * normalized nr_running number somewhere that negates
3847 avg_load_per_task = sg_div_cpu_power(group,
3848 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3850 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3853 sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3858 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3859 * @sd: sched_domain whose statistics are to be updated.
3860 * @this_cpu: Cpu for which load balance is currently performed.
3861 * @idle: Idle status of this_cpu
3862 * @sd_idle: Idle status of the sched_domain containing group.
3863 * @cpus: Set of cpus considered for load balancing.
3864 * @balance: Should we balance.
3865 * @sds: variable to hold the statistics for this sched_domain.
3867 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3868 enum cpu_idle_type idle, int *sd_idle,
3869 const struct cpumask *cpus, int *balance,
3870 struct sd_lb_stats *sds)
3872 struct sched_domain *child = sd->child;
3873 struct sched_group *group = sd->groups;
3874 struct sg_lb_stats sgs;
3875 int load_idx, prefer_sibling = 0;
3877 if (child && child->flags & SD_PREFER_SIBLING)
3880 init_sd_power_savings_stats(sd, sds, idle);
3881 load_idx = get_sd_load_idx(sd, idle);
3886 local_group = cpumask_test_cpu(this_cpu,
3887 sched_group_cpus(group));
3888 memset(&sgs, 0, sizeof(sgs));
3889 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3890 local_group, cpus, balance, &sgs);
3892 if (local_group && balance && !(*balance))
3895 sds->total_load += sgs.group_load;
3896 sds->total_pwr += group->__cpu_power;
3899 * In case the child domain prefers tasks go to siblings
3900 * first, lower the group capacity to one so that we'll try
3901 * and move all the excess tasks away.
3904 sgs.group_capacity = 1;
3907 sds->this_load = sgs.avg_load;
3909 sds->this_nr_running = sgs.sum_nr_running;
3910 sds->this_load_per_task = sgs.sum_weighted_load;
3911 } else if (sgs.avg_load > sds->max_load &&
3912 (sgs.sum_nr_running > sgs.group_capacity ||
3914 sds->max_load = sgs.avg_load;
3915 sds->busiest = group;
3916 sds->busiest_nr_running = sgs.sum_nr_running;
3917 sds->busiest_load_per_task = sgs.sum_weighted_load;
3918 sds->group_imb = sgs.group_imb;
3921 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3922 group = group->next;
3923 } while (group != sd->groups);
3927 * fix_small_imbalance - Calculate the minor imbalance that exists
3928 * amongst the groups of a sched_domain, during
3930 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3931 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3932 * @imbalance: Variable to store the imbalance.
3934 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3935 int this_cpu, unsigned long *imbalance)
3937 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3938 unsigned int imbn = 2;
3940 if (sds->this_nr_running) {
3941 sds->this_load_per_task /= sds->this_nr_running;
3942 if (sds->busiest_load_per_task >
3943 sds->this_load_per_task)
3946 sds->this_load_per_task =
3947 cpu_avg_load_per_task(this_cpu);
3949 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3950 sds->busiest_load_per_task * imbn) {
3951 *imbalance = sds->busiest_load_per_task;
3956 * OK, we don't have enough imbalance to justify moving tasks,
3957 * however we may be able to increase total CPU power used by
3961 pwr_now += sds->busiest->__cpu_power *
3962 min(sds->busiest_load_per_task, sds->max_load);
3963 pwr_now += sds->this->__cpu_power *
3964 min(sds->this_load_per_task, sds->this_load);
3965 pwr_now /= SCHED_LOAD_SCALE;
3967 /* Amount of load we'd subtract */
3968 tmp = sg_div_cpu_power(sds->busiest,
3969 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3970 if (sds->max_load > tmp)
3971 pwr_move += sds->busiest->__cpu_power *
3972 min(sds->busiest_load_per_task, sds->max_load - tmp);
3974 /* Amount of load we'd add */
3975 if (sds->max_load * sds->busiest->__cpu_power <
3976 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3977 tmp = sg_div_cpu_power(sds->this,
3978 sds->max_load * sds->busiest->__cpu_power);
3980 tmp = sg_div_cpu_power(sds->this,
3981 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3982 pwr_move += sds->this->__cpu_power *
3983 min(sds->this_load_per_task, sds->this_load + tmp);
3984 pwr_move /= SCHED_LOAD_SCALE;
3986 /* Move if we gain throughput */
3987 if (pwr_move > pwr_now)
3988 *imbalance = sds->busiest_load_per_task;
3992 * calculate_imbalance - Calculate the amount of imbalance present within the
3993 * groups of a given sched_domain during load balance.
3994 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3995 * @this_cpu: Cpu for which currently load balance is being performed.
3996 * @imbalance: The variable to store the imbalance.
3998 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3999 unsigned long *imbalance)
4001 unsigned long max_pull;
4003 * In the presence of smp nice balancing, certain scenarios can have
4004 * max load less than avg load(as we skip the groups at or below
4005 * its cpu_power, while calculating max_load..)
4007 if (sds->max_load < sds->avg_load) {
4009 return fix_small_imbalance(sds, this_cpu, imbalance);
4012 /* Don't want to pull so many tasks that a group would go idle */
4013 max_pull = min(sds->max_load - sds->avg_load,
4014 sds->max_load - sds->busiest_load_per_task);
4016 /* How much load to actually move to equalise the imbalance */
4017 *imbalance = min(max_pull * sds->busiest->__cpu_power,
4018 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
4022 * if *imbalance is less than the average load per runnable task
4023 * there is no gaurantee that any tasks will be moved so we'll have
4024 * a think about bumping its value to force at least one task to be
4027 if (*imbalance < sds->busiest_load_per_task)
4028 return fix_small_imbalance(sds, this_cpu, imbalance);
4031 /******* find_busiest_group() helpers end here *********************/
4034 * find_busiest_group - Returns the busiest group within the sched_domain
4035 * if there is an imbalance. If there isn't an imbalance, and
4036 * the user has opted for power-savings, it returns a group whose
4037 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4038 * such a group exists.
4040 * Also calculates the amount of weighted load which should be moved
4041 * to restore balance.
4043 * @sd: The sched_domain whose busiest group is to be returned.
4044 * @this_cpu: The cpu for which load balancing is currently being performed.
4045 * @imbalance: Variable which stores amount of weighted load which should
4046 * be moved to restore balance/put a group to idle.
4047 * @idle: The idle status of this_cpu.
4048 * @sd_idle: The idleness of sd
4049 * @cpus: The set of CPUs under consideration for load-balancing.
4050 * @balance: Pointer to a variable indicating if this_cpu
4051 * is the appropriate cpu to perform load balancing at this_level.
4053 * Returns: - the busiest group if imbalance exists.
4054 * - If no imbalance and user has opted for power-savings balance,
4055 * return the least loaded group whose CPUs can be
4056 * put to idle by rebalancing its tasks onto our group.
4058 static struct sched_group *
4059 find_busiest_group(struct sched_domain *sd, int this_cpu,
4060 unsigned long *imbalance, enum cpu_idle_type idle,
4061 int *sd_idle, const struct cpumask *cpus, int *balance)
4063 struct sd_lb_stats sds;
4065 memset(&sds, 0, sizeof(sds));
4068 * Compute the various statistics relavent for load balancing at
4071 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4074 /* Cases where imbalance does not exist from POV of this_cpu */
4075 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4077 * 2) There is no busy sibling group to pull from.
4078 * 3) This group is the busiest group.
4079 * 4) This group is more busy than the avg busieness at this
4081 * 5) The imbalance is within the specified limit.
4082 * 6) Any rebalance would lead to ping-pong
4084 if (balance && !(*balance))
4087 if (!sds.busiest || sds.busiest_nr_running == 0)
4090 if (sds.this_load >= sds.max_load)
4093 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4095 if (sds.this_load >= sds.avg_load)
4098 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4101 sds.busiest_load_per_task /= sds.busiest_nr_running;
4103 sds.busiest_load_per_task =
4104 min(sds.busiest_load_per_task, sds.avg_load);
4107 * We're trying to get all the cpus to the average_load, so we don't
4108 * want to push ourselves above the average load, nor do we wish to
4109 * reduce the max loaded cpu below the average load, as either of these
4110 * actions would just result in more rebalancing later, and ping-pong
4111 * tasks around. Thus we look for the minimum possible imbalance.
4112 * Negative imbalances (*we* are more loaded than anyone else) will
4113 * be counted as no imbalance for these purposes -- we can't fix that
4114 * by pulling tasks to us. Be careful of negative numbers as they'll
4115 * appear as very large values with unsigned longs.
4117 if (sds.max_load <= sds.busiest_load_per_task)
4120 /* Looks like there is an imbalance. Compute it */
4121 calculate_imbalance(&sds, this_cpu, imbalance);
4126 * There is no obvious imbalance. But check if we can do some balancing
4129 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4137 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4140 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4141 unsigned long imbalance, const struct cpumask *cpus)
4143 struct rq *busiest = NULL, *rq;
4144 unsigned long max_load = 0;
4147 for_each_cpu(i, sched_group_cpus(group)) {
4150 if (!cpumask_test_cpu(i, cpus))
4154 wl = weighted_cpuload(i);
4156 if (rq->nr_running == 1 && wl > imbalance)
4159 if (wl > max_load) {
4169 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4170 * so long as it is large enough.
4172 #define MAX_PINNED_INTERVAL 512
4174 /* Working cpumask for load_balance and load_balance_newidle. */
4175 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4178 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4179 * tasks if there is an imbalance.
4181 static int load_balance(int this_cpu, struct rq *this_rq,
4182 struct sched_domain *sd, enum cpu_idle_type idle,
4185 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4186 struct sched_group *group;
4187 unsigned long imbalance;
4189 unsigned long flags;
4190 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4192 cpumask_setall(cpus);
4195 * When power savings policy is enabled for the parent domain, idle
4196 * sibling can pick up load irrespective of busy siblings. In this case,
4197 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4198 * portraying it as CPU_NOT_IDLE.
4200 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4201 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4204 schedstat_inc(sd, lb_count[idle]);
4208 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4215 schedstat_inc(sd, lb_nobusyg[idle]);
4219 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4221 schedstat_inc(sd, lb_nobusyq[idle]);
4225 BUG_ON(busiest == this_rq);
4227 schedstat_add(sd, lb_imbalance[idle], imbalance);
4230 if (busiest->nr_running > 1) {
4232 * Attempt to move tasks. If find_busiest_group has found
4233 * an imbalance but busiest->nr_running <= 1, the group is
4234 * still unbalanced. ld_moved simply stays zero, so it is
4235 * correctly treated as an imbalance.
4237 local_irq_save(flags);
4238 double_rq_lock(this_rq, busiest);
4239 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4240 imbalance, sd, idle, &all_pinned);
4241 double_rq_unlock(this_rq, busiest);
4242 local_irq_restore(flags);
4245 * some other cpu did the load balance for us.
4247 if (ld_moved && this_cpu != smp_processor_id())
4248 resched_cpu(this_cpu);
4250 /* All tasks on this runqueue were pinned by CPU affinity */
4251 if (unlikely(all_pinned)) {
4252 cpumask_clear_cpu(cpu_of(busiest), cpus);
4253 if (!cpumask_empty(cpus))
4260 schedstat_inc(sd, lb_failed[idle]);
4261 sd->nr_balance_failed++;
4263 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4265 spin_lock_irqsave(&busiest->lock, flags);
4267 /* don't kick the migration_thread, if the curr
4268 * task on busiest cpu can't be moved to this_cpu
4270 if (!cpumask_test_cpu(this_cpu,
4271 &busiest->curr->cpus_allowed)) {
4272 spin_unlock_irqrestore(&busiest->lock, flags);
4274 goto out_one_pinned;
4277 if (!busiest->active_balance) {
4278 busiest->active_balance = 1;
4279 busiest->push_cpu = this_cpu;
4282 spin_unlock_irqrestore(&busiest->lock, flags);
4284 wake_up_process(busiest->migration_thread);
4287 * We've kicked active balancing, reset the failure
4290 sd->nr_balance_failed = sd->cache_nice_tries+1;
4293 sd->nr_balance_failed = 0;
4295 if (likely(!active_balance)) {
4296 /* We were unbalanced, so reset the balancing interval */
4297 sd->balance_interval = sd->min_interval;
4300 * If we've begun active balancing, start to back off. This
4301 * case may not be covered by the all_pinned logic if there
4302 * is only 1 task on the busy runqueue (because we don't call
4305 if (sd->balance_interval < sd->max_interval)
4306 sd->balance_interval *= 2;
4309 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4310 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4316 schedstat_inc(sd, lb_balanced[idle]);
4318 sd->nr_balance_failed = 0;
4321 /* tune up the balancing interval */
4322 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4323 (sd->balance_interval < sd->max_interval))
4324 sd->balance_interval *= 2;
4326 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4327 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4338 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4339 * tasks if there is an imbalance.
4341 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4342 * this_rq is locked.
4345 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4347 struct sched_group *group;
4348 struct rq *busiest = NULL;
4349 unsigned long imbalance;
4353 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4355 cpumask_setall(cpus);
4358 * When power savings policy is enabled for the parent domain, idle
4359 * sibling can pick up load irrespective of busy siblings. In this case,
4360 * let the state of idle sibling percolate up as IDLE, instead of
4361 * portraying it as CPU_NOT_IDLE.
4363 if (sd->flags & SD_SHARE_CPUPOWER &&
4364 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4367 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4369 update_shares_locked(this_rq, sd);
4370 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4371 &sd_idle, cpus, NULL);
4373 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4377 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4379 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4383 BUG_ON(busiest == this_rq);
4385 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4388 if (busiest->nr_running > 1) {
4389 /* Attempt to move tasks */
4390 double_lock_balance(this_rq, busiest);
4391 /* this_rq->clock is already updated */
4392 update_rq_clock(busiest);
4393 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4394 imbalance, sd, CPU_NEWLY_IDLE,
4396 double_unlock_balance(this_rq, busiest);
4398 if (unlikely(all_pinned)) {
4399 cpumask_clear_cpu(cpu_of(busiest), cpus);
4400 if (!cpumask_empty(cpus))
4406 int active_balance = 0;
4408 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4409 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4410 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4413 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4416 if (sd->nr_balance_failed++ < 2)
4420 * The only task running in a non-idle cpu can be moved to this
4421 * cpu in an attempt to completely freeup the other CPU
4422 * package. The same method used to move task in load_balance()
4423 * have been extended for load_balance_newidle() to speedup
4424 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4426 * The package power saving logic comes from
4427 * find_busiest_group(). If there are no imbalance, then
4428 * f_b_g() will return NULL. However when sched_mc={1,2} then
4429 * f_b_g() will select a group from which a running task may be
4430 * pulled to this cpu in order to make the other package idle.
4431 * If there is no opportunity to make a package idle and if
4432 * there are no imbalance, then f_b_g() will return NULL and no
4433 * action will be taken in load_balance_newidle().
4435 * Under normal task pull operation due to imbalance, there
4436 * will be more than one task in the source run queue and
4437 * move_tasks() will succeed. ld_moved will be true and this
4438 * active balance code will not be triggered.
4441 /* Lock busiest in correct order while this_rq is held */
4442 double_lock_balance(this_rq, busiest);
4445 * don't kick the migration_thread, if the curr
4446 * task on busiest cpu can't be moved to this_cpu
4448 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4449 double_unlock_balance(this_rq, busiest);
4454 if (!busiest->active_balance) {
4455 busiest->active_balance = 1;
4456 busiest->push_cpu = this_cpu;
4460 double_unlock_balance(this_rq, busiest);
4462 * Should not call ttwu while holding a rq->lock
4464 spin_unlock(&this_rq->lock);
4466 wake_up_process(busiest->migration_thread);
4467 spin_lock(&this_rq->lock);
4470 sd->nr_balance_failed = 0;
4472 update_shares_locked(this_rq, sd);
4476 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4477 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4478 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4480 sd->nr_balance_failed = 0;
4486 * idle_balance is called by schedule() if this_cpu is about to become
4487 * idle. Attempts to pull tasks from other CPUs.
4489 static void idle_balance(int this_cpu, struct rq *this_rq)
4491 struct sched_domain *sd;
4492 int pulled_task = 0;
4493 unsigned long next_balance = jiffies + HZ;
4495 for_each_domain(this_cpu, sd) {
4496 unsigned long interval;
4498 if (!(sd->flags & SD_LOAD_BALANCE))
4501 if (sd->flags & SD_BALANCE_NEWIDLE)
4502 /* If we've pulled tasks over stop searching: */
4503 pulled_task = load_balance_newidle(this_cpu, this_rq,
4506 interval = msecs_to_jiffies(sd->balance_interval);
4507 if (time_after(next_balance, sd->last_balance + interval))
4508 next_balance = sd->last_balance + interval;
4512 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4514 * We are going idle. next_balance may be set based on
4515 * a busy processor. So reset next_balance.
4517 this_rq->next_balance = next_balance;
4522 * active_load_balance is run by migration threads. It pushes running tasks
4523 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4524 * running on each physical CPU where possible, and avoids physical /
4525 * logical imbalances.
4527 * Called with busiest_rq locked.
4529 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4531 int target_cpu = busiest_rq->push_cpu;
4532 struct sched_domain *sd;
4533 struct rq *target_rq;
4535 /* Is there any task to move? */
4536 if (busiest_rq->nr_running <= 1)
4539 target_rq = cpu_rq(target_cpu);
4542 * This condition is "impossible", if it occurs
4543 * we need to fix it. Originally reported by
4544 * Bjorn Helgaas on a 128-cpu setup.
4546 BUG_ON(busiest_rq == target_rq);
4548 /* move a task from busiest_rq to target_rq */
4549 double_lock_balance(busiest_rq, target_rq);
4550 update_rq_clock(busiest_rq);
4551 update_rq_clock(target_rq);
4553 /* Search for an sd spanning us and the target CPU. */
4554 for_each_domain(target_cpu, sd) {
4555 if ((sd->flags & SD_LOAD_BALANCE) &&
4556 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4561 schedstat_inc(sd, alb_count);
4563 if (move_one_task(target_rq, target_cpu, busiest_rq,
4565 schedstat_inc(sd, alb_pushed);
4567 schedstat_inc(sd, alb_failed);
4569 double_unlock_balance(busiest_rq, target_rq);
4574 atomic_t load_balancer;
4575 cpumask_var_t cpu_mask;
4576 cpumask_var_t ilb_grp_nohz_mask;
4577 } nohz ____cacheline_aligned = {
4578 .load_balancer = ATOMIC_INIT(-1),
4581 int get_nohz_load_balancer(void)
4583 return atomic_read(&nohz.load_balancer);
4586 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4588 * lowest_flag_domain - Return lowest sched_domain containing flag.
4589 * @cpu: The cpu whose lowest level of sched domain is to
4591 * @flag: The flag to check for the lowest sched_domain
4592 * for the given cpu.
4594 * Returns the lowest sched_domain of a cpu which contains the given flag.
4596 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4598 struct sched_domain *sd;
4600 for_each_domain(cpu, sd)
4601 if (sd && (sd->flags & flag))
4608 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4609 * @cpu: The cpu whose domains we're iterating over.
4610 * @sd: variable holding the value of the power_savings_sd
4612 * @flag: The flag to filter the sched_domains to be iterated.
4614 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4615 * set, starting from the lowest sched_domain to the highest.
4617 #define for_each_flag_domain(cpu, sd, flag) \
4618 for (sd = lowest_flag_domain(cpu, flag); \
4619 (sd && (sd->flags & flag)); sd = sd->parent)
4622 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4623 * @ilb_group: group to be checked for semi-idleness
4625 * Returns: 1 if the group is semi-idle. 0 otherwise.
4627 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4628 * and atleast one non-idle CPU. This helper function checks if the given
4629 * sched_group is semi-idle or not.
4631 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4633 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4634 sched_group_cpus(ilb_group));
4637 * A sched_group is semi-idle when it has atleast one busy cpu
4638 * and atleast one idle cpu.
4640 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4643 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4649 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4650 * @cpu: The cpu which is nominating a new idle_load_balancer.
4652 * Returns: Returns the id of the idle load balancer if it exists,
4653 * Else, returns >= nr_cpu_ids.
4655 * This algorithm picks the idle load balancer such that it belongs to a
4656 * semi-idle powersavings sched_domain. The idea is to try and avoid
4657 * completely idle packages/cores just for the purpose of idle load balancing
4658 * when there are other idle cpu's which are better suited for that job.
4660 static int find_new_ilb(int cpu)
4662 struct sched_domain *sd;
4663 struct sched_group *ilb_group;
4666 * Have idle load balancer selection from semi-idle packages only
4667 * when power-aware load balancing is enabled
4669 if (!(sched_smt_power_savings || sched_mc_power_savings))
4673 * Optimize for the case when we have no idle CPUs or only one
4674 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4676 if (cpumask_weight(nohz.cpu_mask) < 2)
4679 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4680 ilb_group = sd->groups;
4683 if (is_semi_idle_group(ilb_group))
4684 return cpumask_first(nohz.ilb_grp_nohz_mask);
4686 ilb_group = ilb_group->next;
4688 } while (ilb_group != sd->groups);
4692 return cpumask_first(nohz.cpu_mask);
4694 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4695 static inline int find_new_ilb(int call_cpu)
4697 return cpumask_first(nohz.cpu_mask);
4702 * This routine will try to nominate the ilb (idle load balancing)
4703 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4704 * load balancing on behalf of all those cpus. If all the cpus in the system
4705 * go into this tickless mode, then there will be no ilb owner (as there is
4706 * no need for one) and all the cpus will sleep till the next wakeup event
4709 * For the ilb owner, tick is not stopped. And this tick will be used
4710 * for idle load balancing. ilb owner will still be part of
4713 * While stopping the tick, this cpu will become the ilb owner if there
4714 * is no other owner. And will be the owner till that cpu becomes busy
4715 * or if all cpus in the system stop their ticks at which point
4716 * there is no need for ilb owner.
4718 * When the ilb owner becomes busy, it nominates another owner, during the
4719 * next busy scheduler_tick()
4721 int select_nohz_load_balancer(int stop_tick)
4723 int cpu = smp_processor_id();
4726 cpu_rq(cpu)->in_nohz_recently = 1;
4728 if (!cpu_active(cpu)) {
4729 if (atomic_read(&nohz.load_balancer) != cpu)
4733 * If we are going offline and still the leader,
4736 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4742 cpumask_set_cpu(cpu, nohz.cpu_mask);
4744 /* time for ilb owner also to sleep */
4745 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4746 if (atomic_read(&nohz.load_balancer) == cpu)
4747 atomic_set(&nohz.load_balancer, -1);
4751 if (atomic_read(&nohz.load_balancer) == -1) {
4752 /* make me the ilb owner */
4753 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4755 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4758 if (!(sched_smt_power_savings ||
4759 sched_mc_power_savings))
4762 * Check to see if there is a more power-efficient
4765 new_ilb = find_new_ilb(cpu);
4766 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4767 atomic_set(&nohz.load_balancer, -1);
4768 resched_cpu(new_ilb);
4774 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4777 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4779 if (atomic_read(&nohz.load_balancer) == cpu)
4780 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4787 static DEFINE_SPINLOCK(balancing);
4790 * It checks each scheduling domain to see if it is due to be balanced,
4791 * and initiates a balancing operation if so.
4793 * Balancing parameters are set up in arch_init_sched_domains.
4795 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4798 struct rq *rq = cpu_rq(cpu);
4799 unsigned long interval;
4800 struct sched_domain *sd;
4801 /* Earliest time when we have to do rebalance again */
4802 unsigned long next_balance = jiffies + 60*HZ;
4803 int update_next_balance = 0;
4806 for_each_domain(cpu, sd) {
4807 if (!(sd->flags & SD_LOAD_BALANCE))
4810 interval = sd->balance_interval;
4811 if (idle != CPU_IDLE)
4812 interval *= sd->busy_factor;
4814 /* scale ms to jiffies */
4815 interval = msecs_to_jiffies(interval);
4816 if (unlikely(!interval))
4818 if (interval > HZ*NR_CPUS/10)
4819 interval = HZ*NR_CPUS/10;
4821 need_serialize = sd->flags & SD_SERIALIZE;
4823 if (need_serialize) {
4824 if (!spin_trylock(&balancing))
4828 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4829 if (load_balance(cpu, rq, sd, idle, &balance)) {
4831 * We've pulled tasks over so either we're no
4832 * longer idle, or one of our SMT siblings is
4835 idle = CPU_NOT_IDLE;
4837 sd->last_balance = jiffies;
4840 spin_unlock(&balancing);
4842 if (time_after(next_balance, sd->last_balance + interval)) {
4843 next_balance = sd->last_balance + interval;
4844 update_next_balance = 1;
4848 * Stop the load balance at this level. There is another
4849 * CPU in our sched group which is doing load balancing more
4857 * next_balance will be updated only when there is a need.
4858 * When the cpu is attached to null domain for ex, it will not be
4861 if (likely(update_next_balance))
4862 rq->next_balance = next_balance;
4866 * run_rebalance_domains is triggered when needed from the scheduler tick.
4867 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4868 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4870 static void run_rebalance_domains(struct softirq_action *h)
4872 int this_cpu = smp_processor_id();
4873 struct rq *this_rq = cpu_rq(this_cpu);
4874 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4875 CPU_IDLE : CPU_NOT_IDLE;
4877 rebalance_domains(this_cpu, idle);
4881 * If this cpu is the owner for idle load balancing, then do the
4882 * balancing on behalf of the other idle cpus whose ticks are
4885 if (this_rq->idle_at_tick &&
4886 atomic_read(&nohz.load_balancer) == this_cpu) {
4890 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4891 if (balance_cpu == this_cpu)
4895 * If this cpu gets work to do, stop the load balancing
4896 * work being done for other cpus. Next load
4897 * balancing owner will pick it up.
4902 rebalance_domains(balance_cpu, CPU_IDLE);
4904 rq = cpu_rq(balance_cpu);
4905 if (time_after(this_rq->next_balance, rq->next_balance))
4906 this_rq->next_balance = rq->next_balance;
4912 static inline int on_null_domain(int cpu)
4914 return !rcu_dereference(cpu_rq(cpu)->sd);
4918 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4920 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4921 * idle load balancing owner or decide to stop the periodic load balancing,
4922 * if the whole system is idle.
4924 static inline void trigger_load_balance(struct rq *rq, int cpu)
4928 * If we were in the nohz mode recently and busy at the current
4929 * scheduler tick, then check if we need to nominate new idle
4932 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4933 rq->in_nohz_recently = 0;
4935 if (atomic_read(&nohz.load_balancer) == cpu) {
4936 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4937 atomic_set(&nohz.load_balancer, -1);
4940 if (atomic_read(&nohz.load_balancer) == -1) {
4941 int ilb = find_new_ilb(cpu);
4943 if (ilb < nr_cpu_ids)
4949 * If this cpu is idle and doing idle load balancing for all the
4950 * cpus with ticks stopped, is it time for that to stop?
4952 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4953 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4959 * If this cpu is idle and the idle load balancing is done by
4960 * someone else, then no need raise the SCHED_SOFTIRQ
4962 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4963 cpumask_test_cpu(cpu, nohz.cpu_mask))
4966 /* Don't need to rebalance while attached to NULL domain */
4967 if (time_after_eq(jiffies, rq->next_balance) &&
4968 likely(!on_null_domain(cpu)))
4969 raise_softirq(SCHED_SOFTIRQ);
4972 #else /* CONFIG_SMP */
4975 * on UP we do not need to balance between CPUs:
4977 static inline void idle_balance(int cpu, struct rq *rq)
4983 DEFINE_PER_CPU(struct kernel_stat, kstat);
4985 EXPORT_PER_CPU_SYMBOL(kstat);
4988 * Return any ns on the sched_clock that have not yet been accounted in
4989 * @p in case that task is currently running.
4991 * Called with task_rq_lock() held on @rq.
4993 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4997 if (task_current(rq, p)) {
4998 update_rq_clock(rq);
4999 ns = rq->clock - p->se.exec_start;
5007 unsigned long long task_delta_exec(struct task_struct *p)
5009 unsigned long flags;
5013 rq = task_rq_lock(p, &flags);
5014 ns = do_task_delta_exec(p, rq);
5015 task_rq_unlock(rq, &flags);
5021 * Return accounted runtime for the task.
5022 * In case the task is currently running, return the runtime plus current's
5023 * pending runtime that have not been accounted yet.
5025 unsigned long long task_sched_runtime(struct task_struct *p)
5027 unsigned long flags;
5031 rq = task_rq_lock(p, &flags);
5032 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
5033 task_rq_unlock(rq, &flags);
5039 * Return sum_exec_runtime for the thread group.
5040 * In case the task is currently running, return the sum plus current's
5041 * pending runtime that have not been accounted yet.
5043 * Note that the thread group might have other running tasks as well,
5044 * so the return value not includes other pending runtime that other
5045 * running tasks might have.
5047 unsigned long long thread_group_sched_runtime(struct task_struct *p)
5049 struct task_cputime totals;
5050 unsigned long flags;
5054 rq = task_rq_lock(p, &flags);
5055 thread_group_cputime(p, &totals);
5056 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
5057 task_rq_unlock(rq, &flags);
5063 * Account user cpu time to a process.
5064 * @p: the process that the cpu time gets accounted to
5065 * @cputime: the cpu time spent in user space since the last update
5066 * @cputime_scaled: cputime scaled by cpu frequency
5068 void account_user_time(struct task_struct *p, cputime_t cputime,
5069 cputime_t cputime_scaled)
5071 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5074 /* Add user time to process. */
5075 p->utime = cputime_add(p->utime, cputime);
5076 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5077 account_group_user_time(p, cputime);
5079 /* Add user time to cpustat. */
5080 tmp = cputime_to_cputime64(cputime);
5081 if (TASK_NICE(p) > 0)
5082 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5084 cpustat->user = cputime64_add(cpustat->user, tmp);
5086 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5087 /* Account for user time used */
5088 acct_update_integrals(p);
5092 * Account guest cpu time to a process.
5093 * @p: the process that the cpu time gets accounted to
5094 * @cputime: the cpu time spent in virtual machine since the last update
5095 * @cputime_scaled: cputime scaled by cpu frequency
5097 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5098 cputime_t cputime_scaled)
5101 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5103 tmp = cputime_to_cputime64(cputime);
5105 /* Add guest time to process. */
5106 p->utime = cputime_add(p->utime, cputime);
5107 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5108 account_group_user_time(p, cputime);
5109 p->gtime = cputime_add(p->gtime, cputime);
5111 /* Add guest time to cpustat. */
5112 cpustat->user = cputime64_add(cpustat->user, tmp);
5113 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5117 * Account system cpu time to a process.
5118 * @p: the process that the cpu time gets accounted to
5119 * @hardirq_offset: the offset to subtract from hardirq_count()
5120 * @cputime: the cpu time spent in kernel space since the last update
5121 * @cputime_scaled: cputime scaled by cpu frequency
5123 void account_system_time(struct task_struct *p, int hardirq_offset,
5124 cputime_t cputime, cputime_t cputime_scaled)
5126 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5129 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5130 account_guest_time(p, cputime, cputime_scaled);
5134 /* Add system time to process. */
5135 p->stime = cputime_add(p->stime, cputime);
5136 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5137 account_group_system_time(p, cputime);
5139 /* Add system time to cpustat. */
5140 tmp = cputime_to_cputime64(cputime);
5141 if (hardirq_count() - hardirq_offset)
5142 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5143 else if (softirq_count())
5144 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5146 cpustat->system = cputime64_add(cpustat->system, tmp);
5148 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5150 /* Account for system time used */
5151 acct_update_integrals(p);
5155 * Account for involuntary wait time.
5156 * @steal: the cpu time spent in involuntary wait
5158 void account_steal_time(cputime_t cputime)
5160 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5161 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5163 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5167 * Account for idle time.
5168 * @cputime: the cpu time spent in idle wait
5170 void account_idle_time(cputime_t cputime)
5172 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5173 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5174 struct rq *rq = this_rq();
5176 if (atomic_read(&rq->nr_iowait) > 0)
5177 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5179 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5182 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5185 * Account a single tick of cpu time.
5186 * @p: the process that the cpu time gets accounted to
5187 * @user_tick: indicates if the tick is a user or a system tick
5189 void account_process_tick(struct task_struct *p, int user_tick)
5191 cputime_t one_jiffy = jiffies_to_cputime(1);
5192 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
5193 struct rq *rq = this_rq();
5196 account_user_time(p, one_jiffy, one_jiffy_scaled);
5197 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5198 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
5201 account_idle_time(one_jiffy);
5205 * Account multiple ticks of steal time.
5206 * @p: the process from which the cpu time has been stolen
5207 * @ticks: number of stolen ticks
5209 void account_steal_ticks(unsigned long ticks)
5211 account_steal_time(jiffies_to_cputime(ticks));
5215 * Account multiple ticks of idle time.
5216 * @ticks: number of stolen ticks
5218 void account_idle_ticks(unsigned long ticks)
5220 account_idle_time(jiffies_to_cputime(ticks));
5226 * Use precise platform statistics if available:
5228 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5229 cputime_t task_utime(struct task_struct *p)
5234 cputime_t task_stime(struct task_struct *p)
5239 cputime_t task_utime(struct task_struct *p)
5241 clock_t utime = cputime_to_clock_t(p->utime),
5242 total = utime + cputime_to_clock_t(p->stime);
5246 * Use CFS's precise accounting:
5248 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5252 do_div(temp, total);
5254 utime = (clock_t)temp;
5256 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5257 return p->prev_utime;
5260 cputime_t task_stime(struct task_struct *p)
5265 * Use CFS's precise accounting. (we subtract utime from
5266 * the total, to make sure the total observed by userspace
5267 * grows monotonically - apps rely on that):
5269 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5270 cputime_to_clock_t(task_utime(p));
5273 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5275 return p->prev_stime;
5279 inline cputime_t task_gtime(struct task_struct *p)
5285 * This function gets called by the timer code, with HZ frequency.
5286 * We call it with interrupts disabled.
5288 * It also gets called by the fork code, when changing the parent's
5291 void scheduler_tick(void)
5293 int cpu = smp_processor_id();
5294 struct rq *rq = cpu_rq(cpu);
5295 struct task_struct *curr = rq->curr;
5299 spin_lock(&rq->lock);
5300 update_rq_clock(rq);
5301 update_cpu_load(rq);
5302 curr->sched_class->task_tick(rq, curr, 0);
5303 spin_unlock(&rq->lock);
5305 perf_counter_task_tick(curr, cpu);
5308 rq->idle_at_tick = idle_cpu(cpu);
5309 trigger_load_balance(rq, cpu);
5313 notrace unsigned long get_parent_ip(unsigned long addr)
5315 if (in_lock_functions(addr)) {
5316 addr = CALLER_ADDR2;
5317 if (in_lock_functions(addr))
5318 addr = CALLER_ADDR3;
5323 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5324 defined(CONFIG_PREEMPT_TRACER))
5326 void __kprobes add_preempt_count(int val)
5328 #ifdef CONFIG_DEBUG_PREEMPT
5332 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5335 preempt_count() += val;
5336 #ifdef CONFIG_DEBUG_PREEMPT
5338 * Spinlock count overflowing soon?
5340 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5343 if (preempt_count() == val)
5344 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5346 EXPORT_SYMBOL(add_preempt_count);
5348 void __kprobes sub_preempt_count(int val)
5350 #ifdef CONFIG_DEBUG_PREEMPT
5354 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5357 * Is the spinlock portion underflowing?
5359 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5360 !(preempt_count() & PREEMPT_MASK)))
5364 if (preempt_count() == val)
5365 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5366 preempt_count() -= val;
5368 EXPORT_SYMBOL(sub_preempt_count);
5373 * Print scheduling while atomic bug:
5375 static noinline void __schedule_bug(struct task_struct *prev)
5377 struct pt_regs *regs = get_irq_regs();
5379 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5380 prev->comm, prev->pid, preempt_count());
5382 debug_show_held_locks(prev);
5384 if (irqs_disabled())
5385 print_irqtrace_events(prev);
5394 * Various schedule()-time debugging checks and statistics:
5396 static inline void schedule_debug(struct task_struct *prev)
5399 * Test if we are atomic. Since do_exit() needs to call into
5400 * schedule() atomically, we ignore that path for now.
5401 * Otherwise, whine if we are scheduling when we should not be.
5403 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5404 __schedule_bug(prev);
5406 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5408 schedstat_inc(this_rq(), sched_count);
5409 #ifdef CONFIG_SCHEDSTATS
5410 if (unlikely(prev->lock_depth >= 0)) {
5411 schedstat_inc(this_rq(), bkl_count);
5412 schedstat_inc(prev, sched_info.bkl_count);
5417 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5419 if (prev->state == TASK_RUNNING) {
5420 u64 runtime = prev->se.sum_exec_runtime;
5422 runtime -= prev->se.prev_sum_exec_runtime;
5423 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5426 * In order to avoid avg_overlap growing stale when we are
5427 * indeed overlapping and hence not getting put to sleep, grow
5428 * the avg_overlap on preemption.
5430 * We use the average preemption runtime because that
5431 * correlates to the amount of cache footprint a task can
5434 update_avg(&prev->se.avg_overlap, runtime);
5436 prev->sched_class->put_prev_task(rq, prev);
5440 * Pick up the highest-prio task:
5442 static inline struct task_struct *
5443 pick_next_task(struct rq *rq)
5445 const struct sched_class *class;
5446 struct task_struct *p;
5449 * Optimization: we know that if all tasks are in
5450 * the fair class we can call that function directly:
5452 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5453 p = fair_sched_class.pick_next_task(rq);
5458 class = sched_class_highest;
5460 p = class->pick_next_task(rq);
5464 * Will never be NULL as the idle class always
5465 * returns a non-NULL p:
5467 class = class->next;
5472 * schedule() is the main scheduler function.
5474 asmlinkage void __sched schedule(void)
5476 struct task_struct *prev, *next;
5477 unsigned long *switch_count;
5483 cpu = smp_processor_id();
5487 switch_count = &prev->nivcsw;
5489 release_kernel_lock(prev);
5490 need_resched_nonpreemptible:
5492 schedule_debug(prev);
5494 if (sched_feat(HRTICK))
5497 spin_lock_irq(&rq->lock);
5498 update_rq_clock(rq);
5499 clear_tsk_need_resched(prev);
5501 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5502 if (unlikely(signal_pending_state(prev->state, prev)))
5503 prev->state = TASK_RUNNING;
5505 deactivate_task(rq, prev, 1);
5506 switch_count = &prev->nvcsw;
5509 pre_schedule(rq, prev);
5511 if (unlikely(!rq->nr_running))
5512 idle_balance(cpu, rq);
5514 put_prev_task(rq, prev);
5515 next = pick_next_task(rq);
5517 if (likely(prev != next)) {
5518 sched_info_switch(prev, next);
5519 perf_counter_task_sched_out(prev, next, cpu);
5525 context_switch(rq, prev, next); /* unlocks the rq */
5527 * the context switch might have flipped the stack from under
5528 * us, hence refresh the local variables.
5530 cpu = smp_processor_id();
5533 spin_unlock_irq(&rq->lock);
5537 if (unlikely(reacquire_kernel_lock(current) < 0))
5538 goto need_resched_nonpreemptible;
5540 preempt_enable_no_resched();
5544 EXPORT_SYMBOL(schedule);
5548 * Look out! "owner" is an entirely speculative pointer
5549 * access and not reliable.
5551 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5556 if (!sched_feat(OWNER_SPIN))
5559 #ifdef CONFIG_DEBUG_PAGEALLOC
5561 * Need to access the cpu field knowing that
5562 * DEBUG_PAGEALLOC could have unmapped it if
5563 * the mutex owner just released it and exited.
5565 if (probe_kernel_address(&owner->cpu, cpu))
5572 * Even if the access succeeded (likely case),
5573 * the cpu field may no longer be valid.
5575 if (cpu >= nr_cpumask_bits)
5579 * We need to validate that we can do a
5580 * get_cpu() and that we have the percpu area.
5582 if (!cpu_online(cpu))
5589 * Owner changed, break to re-assess state.
5591 if (lock->owner != owner)
5595 * Is that owner really running on that cpu?
5597 if (task_thread_info(rq->curr) != owner || need_resched())
5607 #ifdef CONFIG_PREEMPT
5609 * this is the entry point to schedule() from in-kernel preemption
5610 * off of preempt_enable. Kernel preemptions off return from interrupt
5611 * occur there and call schedule directly.
5613 asmlinkage void __sched preempt_schedule(void)
5615 struct thread_info *ti = current_thread_info();
5618 * If there is a non-zero preempt_count or interrupts are disabled,
5619 * we do not want to preempt the current task. Just return..
5621 if (likely(ti->preempt_count || irqs_disabled()))
5625 add_preempt_count(PREEMPT_ACTIVE);
5627 sub_preempt_count(PREEMPT_ACTIVE);
5630 * Check again in case we missed a preemption opportunity
5631 * between schedule and now.
5634 } while (need_resched());
5636 EXPORT_SYMBOL(preempt_schedule);
5639 * this is the entry point to schedule() from kernel preemption
5640 * off of irq context.
5641 * Note, that this is called and return with irqs disabled. This will
5642 * protect us against recursive calling from irq.
5644 asmlinkage void __sched preempt_schedule_irq(void)
5646 struct thread_info *ti = current_thread_info();
5648 /* Catch callers which need to be fixed */
5649 BUG_ON(ti->preempt_count || !irqs_disabled());
5652 add_preempt_count(PREEMPT_ACTIVE);
5655 local_irq_disable();
5656 sub_preempt_count(PREEMPT_ACTIVE);
5659 * Check again in case we missed a preemption opportunity
5660 * between schedule and now.
5663 } while (need_resched());
5666 #endif /* CONFIG_PREEMPT */
5668 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5671 return try_to_wake_up(curr->private, mode, sync);
5673 EXPORT_SYMBOL(default_wake_function);
5676 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5677 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5678 * number) then we wake all the non-exclusive tasks and one exclusive task.
5680 * There are circumstances in which we can try to wake a task which has already
5681 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5682 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5684 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5685 int nr_exclusive, int sync, void *key)
5687 wait_queue_t *curr, *next;
5689 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5690 unsigned flags = curr->flags;
5692 if (curr->func(curr, mode, sync, key) &&
5693 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5699 * __wake_up - wake up threads blocked on a waitqueue.
5701 * @mode: which threads
5702 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5703 * @key: is directly passed to the wakeup function
5705 * It may be assumed that this function implies a write memory barrier before
5706 * changing the task state if and only if any tasks are woken up.
5708 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5709 int nr_exclusive, void *key)
5711 unsigned long flags;
5713 spin_lock_irqsave(&q->lock, flags);
5714 __wake_up_common(q, mode, nr_exclusive, 0, key);
5715 spin_unlock_irqrestore(&q->lock, flags);
5717 EXPORT_SYMBOL(__wake_up);
5720 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5722 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5724 __wake_up_common(q, mode, 1, 0, NULL);
5727 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5729 __wake_up_common(q, mode, 1, 0, key);
5733 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5735 * @mode: which threads
5736 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5737 * @key: opaque value to be passed to wakeup targets
5739 * The sync wakeup differs that the waker knows that it will schedule
5740 * away soon, so while the target thread will be woken up, it will not
5741 * be migrated to another CPU - ie. the two threads are 'synchronized'
5742 * with each other. This can prevent needless bouncing between CPUs.
5744 * On UP it can prevent extra preemption.
5746 * It may be assumed that this function implies a write memory barrier before
5747 * changing the task state if and only if any tasks are woken up.
5749 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5750 int nr_exclusive, void *key)
5752 unsigned long flags;
5758 if (unlikely(!nr_exclusive))
5761 spin_lock_irqsave(&q->lock, flags);
5762 __wake_up_common(q, mode, nr_exclusive, sync, key);
5763 spin_unlock_irqrestore(&q->lock, flags);
5765 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5768 * __wake_up_sync - see __wake_up_sync_key()
5770 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5772 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5774 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5777 * complete: - signals a single thread waiting on this completion
5778 * @x: holds the state of this particular completion
5780 * This will wake up a single thread waiting on this completion. Threads will be
5781 * awakened in the same order in which they were queued.
5783 * See also complete_all(), wait_for_completion() and related routines.
5785 * It may be assumed that this function implies a write memory barrier before
5786 * changing the task state if and only if any tasks are woken up.
5788 void complete(struct completion *x)
5790 unsigned long flags;
5792 spin_lock_irqsave(&x->wait.lock, flags);
5794 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5795 spin_unlock_irqrestore(&x->wait.lock, flags);
5797 EXPORT_SYMBOL(complete);
5800 * complete_all: - signals all threads waiting on this completion
5801 * @x: holds the state of this particular completion
5803 * This will wake up all threads waiting on this particular completion event.
5805 * It may be assumed that this function implies a write memory barrier before
5806 * changing the task state if and only if any tasks are woken up.
5808 void complete_all(struct completion *x)
5810 unsigned long flags;
5812 spin_lock_irqsave(&x->wait.lock, flags);
5813 x->done += UINT_MAX/2;
5814 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5815 spin_unlock_irqrestore(&x->wait.lock, flags);
5817 EXPORT_SYMBOL(complete_all);
5819 static inline long __sched
5820 do_wait_for_common(struct completion *x, long timeout, int state)
5823 DECLARE_WAITQUEUE(wait, current);
5825 wait.flags |= WQ_FLAG_EXCLUSIVE;
5826 __add_wait_queue_tail(&x->wait, &wait);
5828 if (signal_pending_state(state, current)) {
5829 timeout = -ERESTARTSYS;
5832 __set_current_state(state);
5833 spin_unlock_irq(&x->wait.lock);
5834 timeout = schedule_timeout(timeout);
5835 spin_lock_irq(&x->wait.lock);
5836 } while (!x->done && timeout);
5837 __remove_wait_queue(&x->wait, &wait);
5842 return timeout ?: 1;
5846 wait_for_common(struct completion *x, long timeout, int state)
5850 spin_lock_irq(&x->wait.lock);
5851 timeout = do_wait_for_common(x, timeout, state);
5852 spin_unlock_irq(&x->wait.lock);
5857 * wait_for_completion: - waits for completion of a task
5858 * @x: holds the state of this particular completion
5860 * This waits to be signaled for completion of a specific task. It is NOT
5861 * interruptible and there is no timeout.
5863 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5864 * and interrupt capability. Also see complete().
5866 void __sched wait_for_completion(struct completion *x)
5868 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5870 EXPORT_SYMBOL(wait_for_completion);
5873 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5874 * @x: holds the state of this particular completion
5875 * @timeout: timeout value in jiffies
5877 * This waits for either a completion of a specific task to be signaled or for a
5878 * specified timeout to expire. The timeout is in jiffies. It is not
5881 unsigned long __sched
5882 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5884 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5886 EXPORT_SYMBOL(wait_for_completion_timeout);
5889 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5890 * @x: holds the state of this particular completion
5892 * This waits for completion of a specific task to be signaled. It is
5895 int __sched wait_for_completion_interruptible(struct completion *x)
5897 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5898 if (t == -ERESTARTSYS)
5902 EXPORT_SYMBOL(wait_for_completion_interruptible);
5905 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5906 * @x: holds the state of this particular completion
5907 * @timeout: timeout value in jiffies
5909 * This waits for either a completion of a specific task to be signaled or for a
5910 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5912 unsigned long __sched
5913 wait_for_completion_interruptible_timeout(struct completion *x,
5914 unsigned long timeout)
5916 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5918 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5921 * wait_for_completion_killable: - waits for completion of a task (killable)
5922 * @x: holds the state of this particular completion
5924 * This waits to be signaled for completion of a specific task. It can be
5925 * interrupted by a kill signal.
5927 int __sched wait_for_completion_killable(struct completion *x)
5929 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5930 if (t == -ERESTARTSYS)
5934 EXPORT_SYMBOL(wait_for_completion_killable);
5937 * try_wait_for_completion - try to decrement a completion without blocking
5938 * @x: completion structure
5940 * Returns: 0 if a decrement cannot be done without blocking
5941 * 1 if a decrement succeeded.
5943 * If a completion is being used as a counting completion,
5944 * attempt to decrement the counter without blocking. This
5945 * enables us to avoid waiting if the resource the completion
5946 * is protecting is not available.
5948 bool try_wait_for_completion(struct completion *x)
5952 spin_lock_irq(&x->wait.lock);
5957 spin_unlock_irq(&x->wait.lock);
5960 EXPORT_SYMBOL(try_wait_for_completion);
5963 * completion_done - Test to see if a completion has any waiters
5964 * @x: completion structure
5966 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5967 * 1 if there are no waiters.
5970 bool completion_done(struct completion *x)
5974 spin_lock_irq(&x->wait.lock);
5977 spin_unlock_irq(&x->wait.lock);
5980 EXPORT_SYMBOL(completion_done);
5983 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5985 unsigned long flags;
5988 init_waitqueue_entry(&wait, current);
5990 __set_current_state(state);
5992 spin_lock_irqsave(&q->lock, flags);
5993 __add_wait_queue(q, &wait);
5994 spin_unlock(&q->lock);
5995 timeout = schedule_timeout(timeout);
5996 spin_lock_irq(&q->lock);
5997 __remove_wait_queue(q, &wait);
5998 spin_unlock_irqrestore(&q->lock, flags);
6003 void __sched interruptible_sleep_on(wait_queue_head_t *q)
6005 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6007 EXPORT_SYMBOL(interruptible_sleep_on);
6010 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
6012 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
6014 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
6016 void __sched sleep_on(wait_queue_head_t *q)
6018 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6020 EXPORT_SYMBOL(sleep_on);
6022 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
6024 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
6026 EXPORT_SYMBOL(sleep_on_timeout);
6028 #ifdef CONFIG_RT_MUTEXES
6031 * rt_mutex_setprio - set the current priority of a task
6033 * @prio: prio value (kernel-internal form)
6035 * This function changes the 'effective' priority of a task. It does
6036 * not touch ->normal_prio like __setscheduler().
6038 * Used by the rt_mutex code to implement priority inheritance logic.
6040 void rt_mutex_setprio(struct task_struct *p, int prio)
6042 unsigned long flags;
6043 int oldprio, on_rq, running;
6045 const struct sched_class *prev_class = p->sched_class;
6047 BUG_ON(prio < 0 || prio > MAX_PRIO);
6049 rq = task_rq_lock(p, &flags);
6050 update_rq_clock(rq);
6053 on_rq = p->se.on_rq;
6054 running = task_current(rq, p);
6056 dequeue_task(rq, p, 0);
6058 p->sched_class->put_prev_task(rq, p);
6061 p->sched_class = &rt_sched_class;
6063 p->sched_class = &fair_sched_class;
6068 p->sched_class->set_curr_task(rq);
6070 enqueue_task(rq, p, 0);
6072 check_class_changed(rq, p, prev_class, oldprio, running);
6074 task_rq_unlock(rq, &flags);
6079 void set_user_nice(struct task_struct *p, long nice)
6081 int old_prio, delta, on_rq;
6082 unsigned long flags;
6085 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6088 * We have to be careful, if called from sys_setpriority(),
6089 * the task might be in the middle of scheduling on another CPU.
6091 rq = task_rq_lock(p, &flags);
6092 update_rq_clock(rq);
6094 * The RT priorities are set via sched_setscheduler(), but we still
6095 * allow the 'normal' nice value to be set - but as expected
6096 * it wont have any effect on scheduling until the task is
6097 * SCHED_FIFO/SCHED_RR:
6099 if (task_has_rt_policy(p)) {
6100 p->static_prio = NICE_TO_PRIO(nice);
6103 on_rq = p->se.on_rq;
6105 dequeue_task(rq, p, 0);
6107 p->static_prio = NICE_TO_PRIO(nice);
6110 p->prio = effective_prio(p);
6111 delta = p->prio - old_prio;
6114 enqueue_task(rq, p, 0);
6116 * If the task increased its priority or is running and
6117 * lowered its priority, then reschedule its CPU:
6119 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6120 resched_task(rq->curr);
6123 task_rq_unlock(rq, &flags);
6125 EXPORT_SYMBOL(set_user_nice);
6128 * can_nice - check if a task can reduce its nice value
6132 int can_nice(const struct task_struct *p, const int nice)
6134 /* convert nice value [19,-20] to rlimit style value [1,40] */
6135 int nice_rlim = 20 - nice;
6137 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6138 capable(CAP_SYS_NICE));
6141 #ifdef __ARCH_WANT_SYS_NICE
6144 * sys_nice - change the priority of the current process.
6145 * @increment: priority increment
6147 * sys_setpriority is a more generic, but much slower function that
6148 * does similar things.
6150 SYSCALL_DEFINE1(nice, int, increment)
6155 * Setpriority might change our priority at the same moment.
6156 * We don't have to worry. Conceptually one call occurs first
6157 * and we have a single winner.
6159 if (increment < -40)
6164 nice = TASK_NICE(current) + increment;
6170 if (increment < 0 && !can_nice(current, nice))
6173 retval = security_task_setnice(current, nice);
6177 set_user_nice(current, nice);
6184 * task_prio - return the priority value of a given task.
6185 * @p: the task in question.
6187 * This is the priority value as seen by users in /proc.
6188 * RT tasks are offset by -200. Normal tasks are centered
6189 * around 0, value goes from -16 to +15.
6191 int task_prio(const struct task_struct *p)
6193 return p->prio - MAX_RT_PRIO;
6197 * task_nice - return the nice value of a given task.
6198 * @p: the task in question.
6200 int task_nice(const struct task_struct *p)
6202 return TASK_NICE(p);
6204 EXPORT_SYMBOL(task_nice);
6207 * idle_cpu - is a given cpu idle currently?
6208 * @cpu: the processor in question.
6210 int idle_cpu(int cpu)
6212 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6216 * idle_task - return the idle task for a given cpu.
6217 * @cpu: the processor in question.
6219 struct task_struct *idle_task(int cpu)
6221 return cpu_rq(cpu)->idle;
6225 * find_process_by_pid - find a process with a matching PID value.
6226 * @pid: the pid in question.
6228 static struct task_struct *find_process_by_pid(pid_t pid)
6230 return pid ? find_task_by_vpid(pid) : current;
6233 /* Actually do priority change: must hold rq lock. */
6235 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6237 BUG_ON(p->se.on_rq);
6240 switch (p->policy) {
6244 p->sched_class = &fair_sched_class;
6248 p->sched_class = &rt_sched_class;
6252 p->rt_priority = prio;
6253 p->normal_prio = normal_prio(p);
6254 /* we are holding p->pi_lock already */
6255 p->prio = rt_mutex_getprio(p);
6260 * check the target process has a UID that matches the current process's
6262 static bool check_same_owner(struct task_struct *p)
6264 const struct cred *cred = current_cred(), *pcred;
6268 pcred = __task_cred(p);
6269 match = (cred->euid == pcred->euid ||
6270 cred->euid == pcred->uid);
6275 static int __sched_setscheduler(struct task_struct *p, int policy,
6276 struct sched_param *param, bool user)
6278 int retval, oldprio, oldpolicy = -1, on_rq, running;
6279 unsigned long flags;
6280 const struct sched_class *prev_class = p->sched_class;
6284 /* may grab non-irq protected spin_locks */
6285 BUG_ON(in_interrupt());
6287 /* double check policy once rq lock held */
6289 reset_on_fork = p->sched_reset_on_fork;
6290 policy = oldpolicy = p->policy;
6292 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6293 policy &= ~SCHED_RESET_ON_FORK;
6295 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6296 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6297 policy != SCHED_IDLE)
6302 * Valid priorities for SCHED_FIFO and SCHED_RR are
6303 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6304 * SCHED_BATCH and SCHED_IDLE is 0.
6306 if (param->sched_priority < 0 ||
6307 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6308 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6310 if (rt_policy(policy) != (param->sched_priority != 0))
6314 * Allow unprivileged RT tasks to decrease priority:
6316 if (user && !capable(CAP_SYS_NICE)) {
6317 if (rt_policy(policy)) {
6318 unsigned long rlim_rtprio;
6320 if (!lock_task_sighand(p, &flags))
6322 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6323 unlock_task_sighand(p, &flags);
6325 /* can't set/change the rt policy */
6326 if (policy != p->policy && !rlim_rtprio)
6329 /* can't increase priority */
6330 if (param->sched_priority > p->rt_priority &&
6331 param->sched_priority > rlim_rtprio)
6335 * Like positive nice levels, dont allow tasks to
6336 * move out of SCHED_IDLE either:
6338 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6341 /* can't change other user's priorities */
6342 if (!check_same_owner(p))
6345 /* Normal users shall not reset the sched_reset_on_fork flag */
6346 if (p->sched_reset_on_fork && !reset_on_fork)
6351 #ifdef CONFIG_RT_GROUP_SCHED
6353 * Do not allow realtime tasks into groups that have no runtime
6356 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6357 task_group(p)->rt_bandwidth.rt_runtime == 0)
6361 retval = security_task_setscheduler(p, policy, param);
6367 * make sure no PI-waiters arrive (or leave) while we are
6368 * changing the priority of the task:
6370 spin_lock_irqsave(&p->pi_lock, flags);
6372 * To be able to change p->policy safely, the apropriate
6373 * runqueue lock must be held.
6375 rq = __task_rq_lock(p);
6376 /* recheck policy now with rq lock held */
6377 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6378 policy = oldpolicy = -1;
6379 __task_rq_unlock(rq);
6380 spin_unlock_irqrestore(&p->pi_lock, flags);
6383 update_rq_clock(rq);
6384 on_rq = p->se.on_rq;
6385 running = task_current(rq, p);
6387 deactivate_task(rq, p, 0);
6389 p->sched_class->put_prev_task(rq, p);
6391 p->sched_reset_on_fork = reset_on_fork;
6394 __setscheduler(rq, p, policy, param->sched_priority);
6397 p->sched_class->set_curr_task(rq);
6399 activate_task(rq, p, 0);
6401 check_class_changed(rq, p, prev_class, oldprio, running);
6403 __task_rq_unlock(rq);
6404 spin_unlock_irqrestore(&p->pi_lock, flags);
6406 rt_mutex_adjust_pi(p);
6412 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6413 * @p: the task in question.
6414 * @policy: new policy.
6415 * @param: structure containing the new RT priority.
6417 * NOTE that the task may be already dead.
6419 int sched_setscheduler(struct task_struct *p, int policy,
6420 struct sched_param *param)
6422 return __sched_setscheduler(p, policy, param, true);
6424 EXPORT_SYMBOL_GPL(sched_setscheduler);
6427 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6428 * @p: the task in question.
6429 * @policy: new policy.
6430 * @param: structure containing the new RT priority.
6432 * Just like sched_setscheduler, only don't bother checking if the
6433 * current context has permission. For example, this is needed in
6434 * stop_machine(): we create temporary high priority worker threads,
6435 * but our caller might not have that capability.
6437 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6438 struct sched_param *param)
6440 return __sched_setscheduler(p, policy, param, false);
6444 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6446 struct sched_param lparam;
6447 struct task_struct *p;
6450 if (!param || pid < 0)
6452 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6457 p = find_process_by_pid(pid);
6459 retval = sched_setscheduler(p, policy, &lparam);
6466 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6467 * @pid: the pid in question.
6468 * @policy: new policy.
6469 * @param: structure containing the new RT priority.
6471 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6472 struct sched_param __user *, param)
6474 /* negative values for policy are not valid */
6478 return do_sched_setscheduler(pid, policy, param);
6482 * sys_sched_setparam - set/change the RT priority of a thread
6483 * @pid: the pid in question.
6484 * @param: structure containing the new RT priority.
6486 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6488 return do_sched_setscheduler(pid, -1, param);
6492 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6493 * @pid: the pid in question.
6495 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6497 struct task_struct *p;
6504 read_lock(&tasklist_lock);
6505 p = find_process_by_pid(pid);
6507 retval = security_task_getscheduler(p);
6510 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6512 read_unlock(&tasklist_lock);
6517 * sys_sched_getparam - get the RT priority of a thread
6518 * @pid: the pid in question.
6519 * @param: structure containing the RT priority.
6521 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6523 struct sched_param lp;
6524 struct task_struct *p;
6527 if (!param || pid < 0)
6530 read_lock(&tasklist_lock);
6531 p = find_process_by_pid(pid);
6536 retval = security_task_getscheduler(p);
6540 lp.sched_priority = p->rt_priority;
6541 read_unlock(&tasklist_lock);
6544 * This one might sleep, we cannot do it with a spinlock held ...
6546 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6551 read_unlock(&tasklist_lock);
6555 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6557 cpumask_var_t cpus_allowed, new_mask;
6558 struct task_struct *p;
6562 read_lock(&tasklist_lock);
6564 p = find_process_by_pid(pid);
6566 read_unlock(&tasklist_lock);
6572 * It is not safe to call set_cpus_allowed with the
6573 * tasklist_lock held. We will bump the task_struct's
6574 * usage count and then drop tasklist_lock.
6577 read_unlock(&tasklist_lock);
6579 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6583 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6585 goto out_free_cpus_allowed;
6588 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6591 retval = security_task_setscheduler(p, 0, NULL);
6595 cpuset_cpus_allowed(p, cpus_allowed);
6596 cpumask_and(new_mask, in_mask, cpus_allowed);
6598 retval = set_cpus_allowed_ptr(p, new_mask);
6601 cpuset_cpus_allowed(p, cpus_allowed);
6602 if (!cpumask_subset(new_mask, cpus_allowed)) {
6604 * We must have raced with a concurrent cpuset
6605 * update. Just reset the cpus_allowed to the
6606 * cpuset's cpus_allowed
6608 cpumask_copy(new_mask, cpus_allowed);
6613 free_cpumask_var(new_mask);
6614 out_free_cpus_allowed:
6615 free_cpumask_var(cpus_allowed);
6622 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6623 struct cpumask *new_mask)
6625 if (len < cpumask_size())
6626 cpumask_clear(new_mask);
6627 else if (len > cpumask_size())
6628 len = cpumask_size();
6630 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6634 * sys_sched_setaffinity - set the cpu affinity of a process
6635 * @pid: pid of the process
6636 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6637 * @user_mask_ptr: user-space pointer to the new cpu mask
6639 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6640 unsigned long __user *, user_mask_ptr)
6642 cpumask_var_t new_mask;
6645 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6648 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6650 retval = sched_setaffinity(pid, new_mask);
6651 free_cpumask_var(new_mask);
6655 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6657 struct task_struct *p;
6661 read_lock(&tasklist_lock);
6664 p = find_process_by_pid(pid);
6668 retval = security_task_getscheduler(p);
6672 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6675 read_unlock(&tasklist_lock);
6682 * sys_sched_getaffinity - get the cpu affinity of a process
6683 * @pid: pid of the process
6684 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6685 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6687 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6688 unsigned long __user *, user_mask_ptr)
6693 if (len < cpumask_size())
6696 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6699 ret = sched_getaffinity(pid, mask);
6701 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6704 ret = cpumask_size();
6706 free_cpumask_var(mask);
6712 * sys_sched_yield - yield the current processor to other threads.
6714 * This function yields the current CPU to other tasks. If there are no
6715 * other threads running on this CPU then this function will return.
6717 SYSCALL_DEFINE0(sched_yield)
6719 struct rq *rq = this_rq_lock();
6721 schedstat_inc(rq, yld_count);
6722 current->sched_class->yield_task(rq);
6725 * Since we are going to call schedule() anyway, there's
6726 * no need to preempt or enable interrupts:
6728 __release(rq->lock);
6729 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6730 _raw_spin_unlock(&rq->lock);
6731 preempt_enable_no_resched();
6738 static inline int should_resched(void)
6740 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6743 static void __cond_resched(void)
6745 add_preempt_count(PREEMPT_ACTIVE);
6747 sub_preempt_count(PREEMPT_ACTIVE);
6750 int __sched _cond_resched(void)
6752 if (should_resched()) {
6758 EXPORT_SYMBOL(_cond_resched);
6761 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6762 * call schedule, and on return reacquire the lock.
6764 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6765 * operations here to prevent schedule() from being called twice (once via
6766 * spin_unlock(), once by hand).
6768 int __cond_resched_lock(spinlock_t *lock)
6770 int resched = should_resched();
6773 if (spin_needbreak(lock) || resched) {
6784 EXPORT_SYMBOL(__cond_resched_lock);
6786 int __sched __cond_resched_softirq(void)
6788 BUG_ON(!in_softirq());
6790 if (should_resched()) {
6798 EXPORT_SYMBOL(__cond_resched_softirq);
6801 * yield - yield the current processor to other threads.
6803 * This is a shortcut for kernel-space yielding - it marks the
6804 * thread runnable and calls sys_sched_yield().
6806 void __sched yield(void)
6808 set_current_state(TASK_RUNNING);
6811 EXPORT_SYMBOL(yield);
6814 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6815 * that process accounting knows that this is a task in IO wait state.
6817 * But don't do that if it is a deliberate, throttling IO wait (this task
6818 * has set its backing_dev_info: the queue against which it should throttle)
6820 void __sched io_schedule(void)
6822 struct rq *rq = raw_rq();
6824 delayacct_blkio_start();
6825 atomic_inc(&rq->nr_iowait);
6826 current->in_iowait = 1;
6828 current->in_iowait = 0;
6829 atomic_dec(&rq->nr_iowait);
6830 delayacct_blkio_end();
6832 EXPORT_SYMBOL(io_schedule);
6834 long __sched io_schedule_timeout(long timeout)
6836 struct rq *rq = raw_rq();
6839 delayacct_blkio_start();
6840 atomic_inc(&rq->nr_iowait);
6841 current->in_iowait = 1;
6842 ret = schedule_timeout(timeout);
6843 current->in_iowait = 0;
6844 atomic_dec(&rq->nr_iowait);
6845 delayacct_blkio_end();
6850 * sys_sched_get_priority_max - return maximum RT priority.
6851 * @policy: scheduling class.
6853 * this syscall returns the maximum rt_priority that can be used
6854 * by a given scheduling class.
6856 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6863 ret = MAX_USER_RT_PRIO-1;
6875 * sys_sched_get_priority_min - return minimum RT priority.
6876 * @policy: scheduling class.
6878 * this syscall returns the minimum rt_priority that can be used
6879 * by a given scheduling class.
6881 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6899 * sys_sched_rr_get_interval - return the default timeslice of a process.
6900 * @pid: pid of the process.
6901 * @interval: userspace pointer to the timeslice value.
6903 * this syscall writes the default timeslice value of a given process
6904 * into the user-space timespec buffer. A value of '0' means infinity.
6906 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6907 struct timespec __user *, interval)
6909 struct task_struct *p;
6910 unsigned int time_slice;
6918 read_lock(&tasklist_lock);
6919 p = find_process_by_pid(pid);
6923 retval = security_task_getscheduler(p);
6928 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6929 * tasks that are on an otherwise idle runqueue:
6932 if (p->policy == SCHED_RR) {
6933 time_slice = DEF_TIMESLICE;
6934 } else if (p->policy != SCHED_FIFO) {
6935 struct sched_entity *se = &p->se;
6936 unsigned long flags;
6939 rq = task_rq_lock(p, &flags);
6940 if (rq->cfs.load.weight)
6941 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6942 task_rq_unlock(rq, &flags);
6944 read_unlock(&tasklist_lock);
6945 jiffies_to_timespec(time_slice, &t);
6946 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6950 read_unlock(&tasklist_lock);
6954 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6956 void sched_show_task(struct task_struct *p)
6958 unsigned long free = 0;
6961 state = p->state ? __ffs(p->state) + 1 : 0;
6962 printk(KERN_INFO "%-13.13s %c", p->comm,
6963 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6964 #if BITS_PER_LONG == 32
6965 if (state == TASK_RUNNING)
6966 printk(KERN_CONT " running ");
6968 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6970 if (state == TASK_RUNNING)
6971 printk(KERN_CONT " running task ");
6973 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6975 #ifdef CONFIG_DEBUG_STACK_USAGE
6976 free = stack_not_used(p);
6978 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6979 task_pid_nr(p), task_pid_nr(p->real_parent),
6980 (unsigned long)task_thread_info(p)->flags);
6982 show_stack(p, NULL);
6985 void show_state_filter(unsigned long state_filter)
6987 struct task_struct *g, *p;
6989 #if BITS_PER_LONG == 32
6991 " task PC stack pid father\n");
6994 " task PC stack pid father\n");
6996 read_lock(&tasklist_lock);
6997 do_each_thread(g, p) {
6999 * reset the NMI-timeout, listing all files on a slow
7000 * console might take alot of time:
7002 touch_nmi_watchdog();
7003 if (!state_filter || (p->state & state_filter))
7005 } while_each_thread(g, p);
7007 touch_all_softlockup_watchdogs();
7009 #ifdef CONFIG_SCHED_DEBUG
7010 sysrq_sched_debug_show();
7012 read_unlock(&tasklist_lock);
7014 * Only show locks if all tasks are dumped:
7016 if (state_filter == -1)
7017 debug_show_all_locks();
7020 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
7022 idle->sched_class = &idle_sched_class;
7026 * init_idle - set up an idle thread for a given CPU
7027 * @idle: task in question
7028 * @cpu: cpu the idle task belongs to
7030 * NOTE: this function does not set the idle thread's NEED_RESCHED
7031 * flag, to make booting more robust.
7033 void __cpuinit init_idle(struct task_struct *idle, int cpu)
7035 struct rq *rq = cpu_rq(cpu);
7036 unsigned long flags;
7038 spin_lock_irqsave(&rq->lock, flags);
7041 idle->se.exec_start = sched_clock();
7043 idle->prio = idle->normal_prio = MAX_PRIO;
7044 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
7045 __set_task_cpu(idle, cpu);
7047 rq->curr = rq->idle = idle;
7048 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7051 spin_unlock_irqrestore(&rq->lock, flags);
7053 /* Set the preempt count _outside_ the spinlocks! */
7054 #if defined(CONFIG_PREEMPT)
7055 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7057 task_thread_info(idle)->preempt_count = 0;
7060 * The idle tasks have their own, simple scheduling class:
7062 idle->sched_class = &idle_sched_class;
7063 ftrace_graph_init_task(idle);
7067 * In a system that switches off the HZ timer nohz_cpu_mask
7068 * indicates which cpus entered this state. This is used
7069 * in the rcu update to wait only for active cpus. For system
7070 * which do not switch off the HZ timer nohz_cpu_mask should
7071 * always be CPU_BITS_NONE.
7073 cpumask_var_t nohz_cpu_mask;
7076 * Increase the granularity value when there are more CPUs,
7077 * because with more CPUs the 'effective latency' as visible
7078 * to users decreases. But the relationship is not linear,
7079 * so pick a second-best guess by going with the log2 of the
7082 * This idea comes from the SD scheduler of Con Kolivas:
7084 static inline void sched_init_granularity(void)
7086 unsigned int factor = 1 + ilog2(num_online_cpus());
7087 const unsigned long limit = 200000000;
7089 sysctl_sched_min_granularity *= factor;
7090 if (sysctl_sched_min_granularity > limit)
7091 sysctl_sched_min_granularity = limit;
7093 sysctl_sched_latency *= factor;
7094 if (sysctl_sched_latency > limit)
7095 sysctl_sched_latency = limit;
7097 sysctl_sched_wakeup_granularity *= factor;
7099 sysctl_sched_shares_ratelimit *= factor;
7104 * This is how migration works:
7106 * 1) we queue a struct migration_req structure in the source CPU's
7107 * runqueue and wake up that CPU's migration thread.
7108 * 2) we down() the locked semaphore => thread blocks.
7109 * 3) migration thread wakes up (implicitly it forces the migrated
7110 * thread off the CPU)
7111 * 4) it gets the migration request and checks whether the migrated
7112 * task is still in the wrong runqueue.
7113 * 5) if it's in the wrong runqueue then the migration thread removes
7114 * it and puts it into the right queue.
7115 * 6) migration thread up()s the semaphore.
7116 * 7) we wake up and the migration is done.
7120 * Change a given task's CPU affinity. Migrate the thread to a
7121 * proper CPU and schedule it away if the CPU it's executing on
7122 * is removed from the allowed bitmask.
7124 * NOTE: the caller must have a valid reference to the task, the
7125 * task must not exit() & deallocate itself prematurely. The
7126 * call is not atomic; no spinlocks may be held.
7128 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7130 struct migration_req req;
7131 unsigned long flags;
7135 rq = task_rq_lock(p, &flags);
7136 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
7141 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7142 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7147 if (p->sched_class->set_cpus_allowed)
7148 p->sched_class->set_cpus_allowed(p, new_mask);
7150 cpumask_copy(&p->cpus_allowed, new_mask);
7151 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7154 /* Can the task run on the task's current CPU? If so, we're done */
7155 if (cpumask_test_cpu(task_cpu(p), new_mask))
7158 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
7159 /* Need help from migration thread: drop lock and wait. */
7160 struct task_struct *mt = rq->migration_thread;
7162 get_task_struct(mt);
7163 task_rq_unlock(rq, &flags);
7164 wake_up_process(rq->migration_thread);
7165 put_task_struct(mt);
7166 wait_for_completion(&req.done);
7167 tlb_migrate_finish(p->mm);
7171 task_rq_unlock(rq, &flags);
7175 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7178 * Move (not current) task off this cpu, onto dest cpu. We're doing
7179 * this because either it can't run here any more (set_cpus_allowed()
7180 * away from this CPU, or CPU going down), or because we're
7181 * attempting to rebalance this task on exec (sched_exec).
7183 * So we race with normal scheduler movements, but that's OK, as long
7184 * as the task is no longer on this CPU.
7186 * Returns non-zero if task was successfully migrated.
7188 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7190 struct rq *rq_dest, *rq_src;
7193 if (unlikely(!cpu_active(dest_cpu)))
7196 rq_src = cpu_rq(src_cpu);
7197 rq_dest = cpu_rq(dest_cpu);
7199 double_rq_lock(rq_src, rq_dest);
7200 /* Already moved. */
7201 if (task_cpu(p) != src_cpu)
7203 /* Affinity changed (again). */
7204 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7207 on_rq = p->se.on_rq;
7209 deactivate_task(rq_src, p, 0);
7211 set_task_cpu(p, dest_cpu);
7213 activate_task(rq_dest, p, 0);
7214 check_preempt_curr(rq_dest, p, 0);
7219 double_rq_unlock(rq_src, rq_dest);
7224 * migration_thread - this is a highprio system thread that performs
7225 * thread migration by bumping thread off CPU then 'pushing' onto
7228 static int migration_thread(void *data)
7230 int cpu = (long)data;
7234 BUG_ON(rq->migration_thread != current);
7236 set_current_state(TASK_INTERRUPTIBLE);
7237 while (!kthread_should_stop()) {
7238 struct migration_req *req;
7239 struct list_head *head;
7241 spin_lock_irq(&rq->lock);
7243 if (cpu_is_offline(cpu)) {
7244 spin_unlock_irq(&rq->lock);
7248 if (rq->active_balance) {
7249 active_load_balance(rq, cpu);
7250 rq->active_balance = 0;
7253 head = &rq->migration_queue;
7255 if (list_empty(head)) {
7256 spin_unlock_irq(&rq->lock);
7258 set_current_state(TASK_INTERRUPTIBLE);
7261 req = list_entry(head->next, struct migration_req, list);
7262 list_del_init(head->next);
7264 spin_unlock(&rq->lock);
7265 __migrate_task(req->task, cpu, req->dest_cpu);
7268 complete(&req->done);
7270 __set_current_state(TASK_RUNNING);
7275 #ifdef CONFIG_HOTPLUG_CPU
7277 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7281 local_irq_disable();
7282 ret = __migrate_task(p, src_cpu, dest_cpu);
7288 * Figure out where task on dead CPU should go, use force if necessary.
7290 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7293 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7296 /* Look for allowed, online CPU in same node. */
7297 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7298 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7301 /* Any allowed, online CPU? */
7302 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7303 if (dest_cpu < nr_cpu_ids)
7306 /* No more Mr. Nice Guy. */
7307 if (dest_cpu >= nr_cpu_ids) {
7308 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7309 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7312 * Don't tell them about moving exiting tasks or
7313 * kernel threads (both mm NULL), since they never
7316 if (p->mm && printk_ratelimit()) {
7317 printk(KERN_INFO "process %d (%s) no "
7318 "longer affine to cpu%d\n",
7319 task_pid_nr(p), p->comm, dead_cpu);
7324 /* It can have affinity changed while we were choosing. */
7325 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7330 * While a dead CPU has no uninterruptible tasks queued at this point,
7331 * it might still have a nonzero ->nr_uninterruptible counter, because
7332 * for performance reasons the counter is not stricly tracking tasks to
7333 * their home CPUs. So we just add the counter to another CPU's counter,
7334 * to keep the global sum constant after CPU-down:
7336 static void migrate_nr_uninterruptible(struct rq *rq_src)
7338 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7339 unsigned long flags;
7341 local_irq_save(flags);
7342 double_rq_lock(rq_src, rq_dest);
7343 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7344 rq_src->nr_uninterruptible = 0;
7345 double_rq_unlock(rq_src, rq_dest);
7346 local_irq_restore(flags);
7349 /* Run through task list and migrate tasks from the dead cpu. */
7350 static void migrate_live_tasks(int src_cpu)
7352 struct task_struct *p, *t;
7354 read_lock(&tasklist_lock);
7356 do_each_thread(t, p) {
7360 if (task_cpu(p) == src_cpu)
7361 move_task_off_dead_cpu(src_cpu, p);
7362 } while_each_thread(t, p);
7364 read_unlock(&tasklist_lock);
7368 * Schedules idle task to be the next runnable task on current CPU.
7369 * It does so by boosting its priority to highest possible.
7370 * Used by CPU offline code.
7372 void sched_idle_next(void)
7374 int this_cpu = smp_processor_id();
7375 struct rq *rq = cpu_rq(this_cpu);
7376 struct task_struct *p = rq->idle;
7377 unsigned long flags;
7379 /* cpu has to be offline */
7380 BUG_ON(cpu_online(this_cpu));
7383 * Strictly not necessary since rest of the CPUs are stopped by now
7384 * and interrupts disabled on the current cpu.
7386 spin_lock_irqsave(&rq->lock, flags);
7388 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7390 update_rq_clock(rq);
7391 activate_task(rq, p, 0);
7393 spin_unlock_irqrestore(&rq->lock, flags);
7397 * Ensures that the idle task is using init_mm right before its cpu goes
7400 void idle_task_exit(void)
7402 struct mm_struct *mm = current->active_mm;
7404 BUG_ON(cpu_online(smp_processor_id()));
7407 switch_mm(mm, &init_mm, current);
7411 /* called under rq->lock with disabled interrupts */
7412 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7414 struct rq *rq = cpu_rq(dead_cpu);
7416 /* Must be exiting, otherwise would be on tasklist. */
7417 BUG_ON(!p->exit_state);
7419 /* Cannot have done final schedule yet: would have vanished. */
7420 BUG_ON(p->state == TASK_DEAD);
7425 * Drop lock around migration; if someone else moves it,
7426 * that's OK. No task can be added to this CPU, so iteration is
7429 spin_unlock_irq(&rq->lock);
7430 move_task_off_dead_cpu(dead_cpu, p);
7431 spin_lock_irq(&rq->lock);
7436 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7437 static void migrate_dead_tasks(unsigned int dead_cpu)
7439 struct rq *rq = cpu_rq(dead_cpu);
7440 struct task_struct *next;
7443 if (!rq->nr_running)
7445 update_rq_clock(rq);
7446 next = pick_next_task(rq);
7449 next->sched_class->put_prev_task(rq, next);
7450 migrate_dead(dead_cpu, next);
7456 * remove the tasks which were accounted by rq from calc_load_tasks.
7458 static void calc_global_load_remove(struct rq *rq)
7460 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7461 rq->calc_load_active = 0;
7463 #endif /* CONFIG_HOTPLUG_CPU */
7465 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7467 static struct ctl_table sd_ctl_dir[] = {
7469 .procname = "sched_domain",
7475 static struct ctl_table sd_ctl_root[] = {
7477 .ctl_name = CTL_KERN,
7478 .procname = "kernel",
7480 .child = sd_ctl_dir,
7485 static struct ctl_table *sd_alloc_ctl_entry(int n)
7487 struct ctl_table *entry =
7488 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7493 static void sd_free_ctl_entry(struct ctl_table **tablep)
7495 struct ctl_table *entry;
7498 * In the intermediate directories, both the child directory and
7499 * procname are dynamically allocated and could fail but the mode
7500 * will always be set. In the lowest directory the names are
7501 * static strings and all have proc handlers.
7503 for (entry = *tablep; entry->mode; entry++) {
7505 sd_free_ctl_entry(&entry->child);
7506 if (entry->proc_handler == NULL)
7507 kfree(entry->procname);
7515 set_table_entry(struct ctl_table *entry,
7516 const char *procname, void *data, int maxlen,
7517 mode_t mode, proc_handler *proc_handler)
7519 entry->procname = procname;
7521 entry->maxlen = maxlen;
7523 entry->proc_handler = proc_handler;
7526 static struct ctl_table *
7527 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7529 struct ctl_table *table = sd_alloc_ctl_entry(13);
7534 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7535 sizeof(long), 0644, proc_doulongvec_minmax);
7536 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7537 sizeof(long), 0644, proc_doulongvec_minmax);
7538 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7539 sizeof(int), 0644, proc_dointvec_minmax);
7540 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7541 sizeof(int), 0644, proc_dointvec_minmax);
7542 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7543 sizeof(int), 0644, proc_dointvec_minmax);
7544 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7545 sizeof(int), 0644, proc_dointvec_minmax);
7546 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7547 sizeof(int), 0644, proc_dointvec_minmax);
7548 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7549 sizeof(int), 0644, proc_dointvec_minmax);
7550 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7551 sizeof(int), 0644, proc_dointvec_minmax);
7552 set_table_entry(&table[9], "cache_nice_tries",
7553 &sd->cache_nice_tries,
7554 sizeof(int), 0644, proc_dointvec_minmax);
7555 set_table_entry(&table[10], "flags", &sd->flags,
7556 sizeof(int), 0644, proc_dointvec_minmax);
7557 set_table_entry(&table[11], "name", sd->name,
7558 CORENAME_MAX_SIZE, 0444, proc_dostring);
7559 /* &table[12] is terminator */
7564 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7566 struct ctl_table *entry, *table;
7567 struct sched_domain *sd;
7568 int domain_num = 0, i;
7571 for_each_domain(cpu, sd)
7573 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7578 for_each_domain(cpu, sd) {
7579 snprintf(buf, 32, "domain%d", i);
7580 entry->procname = kstrdup(buf, GFP_KERNEL);
7582 entry->child = sd_alloc_ctl_domain_table(sd);
7589 static struct ctl_table_header *sd_sysctl_header;
7590 static void register_sched_domain_sysctl(void)
7592 int i, cpu_num = num_online_cpus();
7593 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7596 WARN_ON(sd_ctl_dir[0].child);
7597 sd_ctl_dir[0].child = entry;
7602 for_each_online_cpu(i) {
7603 snprintf(buf, 32, "cpu%d", i);
7604 entry->procname = kstrdup(buf, GFP_KERNEL);
7606 entry->child = sd_alloc_ctl_cpu_table(i);
7610 WARN_ON(sd_sysctl_header);
7611 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7614 /* may be called multiple times per register */
7615 static void unregister_sched_domain_sysctl(void)
7617 if (sd_sysctl_header)
7618 unregister_sysctl_table(sd_sysctl_header);
7619 sd_sysctl_header = NULL;
7620 if (sd_ctl_dir[0].child)
7621 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7624 static void register_sched_domain_sysctl(void)
7627 static void unregister_sched_domain_sysctl(void)
7632 static void set_rq_online(struct rq *rq)
7635 const struct sched_class *class;
7637 cpumask_set_cpu(rq->cpu, rq->rd->online);
7640 for_each_class(class) {
7641 if (class->rq_online)
7642 class->rq_online(rq);
7647 static void set_rq_offline(struct rq *rq)
7650 const struct sched_class *class;
7652 for_each_class(class) {
7653 if (class->rq_offline)
7654 class->rq_offline(rq);
7657 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7663 * migration_call - callback that gets triggered when a CPU is added.
7664 * Here we can start up the necessary migration thread for the new CPU.
7666 static int __cpuinit
7667 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7669 struct task_struct *p;
7670 int cpu = (long)hcpu;
7671 unsigned long flags;
7676 case CPU_UP_PREPARE:
7677 case CPU_UP_PREPARE_FROZEN:
7678 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7681 kthread_bind(p, cpu);
7682 /* Must be high prio: stop_machine expects to yield to it. */
7683 rq = task_rq_lock(p, &flags);
7684 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7685 task_rq_unlock(rq, &flags);
7687 cpu_rq(cpu)->migration_thread = p;
7688 rq->calc_load_update = calc_load_update;
7692 case CPU_ONLINE_FROZEN:
7693 /* Strictly unnecessary, as first user will wake it. */
7694 wake_up_process(cpu_rq(cpu)->migration_thread);
7696 /* Update our root-domain */
7698 spin_lock_irqsave(&rq->lock, flags);
7700 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7704 spin_unlock_irqrestore(&rq->lock, flags);
7707 #ifdef CONFIG_HOTPLUG_CPU
7708 case CPU_UP_CANCELED:
7709 case CPU_UP_CANCELED_FROZEN:
7710 if (!cpu_rq(cpu)->migration_thread)
7712 /* Unbind it from offline cpu so it can run. Fall thru. */
7713 kthread_bind(cpu_rq(cpu)->migration_thread,
7714 cpumask_any(cpu_online_mask));
7715 kthread_stop(cpu_rq(cpu)->migration_thread);
7716 put_task_struct(cpu_rq(cpu)->migration_thread);
7717 cpu_rq(cpu)->migration_thread = NULL;
7721 case CPU_DEAD_FROZEN:
7722 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7723 migrate_live_tasks(cpu);
7725 kthread_stop(rq->migration_thread);
7726 put_task_struct(rq->migration_thread);
7727 rq->migration_thread = NULL;
7728 /* Idle task back to normal (off runqueue, low prio) */
7729 spin_lock_irq(&rq->lock);
7730 update_rq_clock(rq);
7731 deactivate_task(rq, rq->idle, 0);
7732 rq->idle->static_prio = MAX_PRIO;
7733 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7734 rq->idle->sched_class = &idle_sched_class;
7735 migrate_dead_tasks(cpu);
7736 spin_unlock_irq(&rq->lock);
7738 migrate_nr_uninterruptible(rq);
7739 BUG_ON(rq->nr_running != 0);
7740 calc_global_load_remove(rq);
7742 * No need to migrate the tasks: it was best-effort if
7743 * they didn't take sched_hotcpu_mutex. Just wake up
7746 spin_lock_irq(&rq->lock);
7747 while (!list_empty(&rq->migration_queue)) {
7748 struct migration_req *req;
7750 req = list_entry(rq->migration_queue.next,
7751 struct migration_req, list);
7752 list_del_init(&req->list);
7753 spin_unlock_irq(&rq->lock);
7754 complete(&req->done);
7755 spin_lock_irq(&rq->lock);
7757 spin_unlock_irq(&rq->lock);
7761 case CPU_DYING_FROZEN:
7762 /* Update our root-domain */
7764 spin_lock_irqsave(&rq->lock, flags);
7766 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7769 spin_unlock_irqrestore(&rq->lock, flags);
7777 * Register at high priority so that task migration (migrate_all_tasks)
7778 * happens before everything else. This has to be lower priority than
7779 * the notifier in the perf_counter subsystem, though.
7781 static struct notifier_block __cpuinitdata migration_notifier = {
7782 .notifier_call = migration_call,
7786 static int __init migration_init(void)
7788 void *cpu = (void *)(long)smp_processor_id();
7791 /* Start one for the boot CPU: */
7792 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7793 BUG_ON(err == NOTIFY_BAD);
7794 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7795 register_cpu_notifier(&migration_notifier);
7799 early_initcall(migration_init);
7804 #ifdef CONFIG_SCHED_DEBUG
7806 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7807 struct cpumask *groupmask)
7809 struct sched_group *group = sd->groups;
7812 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7813 cpumask_clear(groupmask);
7815 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7817 if (!(sd->flags & SD_LOAD_BALANCE)) {
7818 printk("does not load-balance\n");
7820 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7825 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7827 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7828 printk(KERN_ERR "ERROR: domain->span does not contain "
7831 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7832 printk(KERN_ERR "ERROR: domain->groups does not contain"
7836 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7840 printk(KERN_ERR "ERROR: group is NULL\n");
7844 if (!group->__cpu_power) {
7845 printk(KERN_CONT "\n");
7846 printk(KERN_ERR "ERROR: domain->cpu_power not "
7851 if (!cpumask_weight(sched_group_cpus(group))) {
7852 printk(KERN_CONT "\n");
7853 printk(KERN_ERR "ERROR: empty group\n");
7857 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7858 printk(KERN_CONT "\n");
7859 printk(KERN_ERR "ERROR: repeated CPUs\n");
7863 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7865 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7867 printk(KERN_CONT " %s", str);
7868 if (group->__cpu_power != SCHED_LOAD_SCALE) {
7869 printk(KERN_CONT " (__cpu_power = %d)",
7870 group->__cpu_power);
7873 group = group->next;
7874 } while (group != sd->groups);
7875 printk(KERN_CONT "\n");
7877 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7878 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7881 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7882 printk(KERN_ERR "ERROR: parent span is not a superset "
7883 "of domain->span\n");
7887 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7889 cpumask_var_t groupmask;
7893 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7897 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7899 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7900 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7905 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7912 free_cpumask_var(groupmask);
7914 #else /* !CONFIG_SCHED_DEBUG */
7915 # define sched_domain_debug(sd, cpu) do { } while (0)
7916 #endif /* CONFIG_SCHED_DEBUG */
7918 static int sd_degenerate(struct sched_domain *sd)
7920 if (cpumask_weight(sched_domain_span(sd)) == 1)
7923 /* Following flags need at least 2 groups */
7924 if (sd->flags & (SD_LOAD_BALANCE |
7925 SD_BALANCE_NEWIDLE |
7929 SD_SHARE_PKG_RESOURCES)) {
7930 if (sd->groups != sd->groups->next)
7934 /* Following flags don't use groups */
7935 if (sd->flags & (SD_WAKE_IDLE |
7944 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7946 unsigned long cflags = sd->flags, pflags = parent->flags;
7948 if (sd_degenerate(parent))
7951 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7954 /* Does parent contain flags not in child? */
7955 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7956 if (cflags & SD_WAKE_AFFINE)
7957 pflags &= ~SD_WAKE_BALANCE;
7958 /* Flags needing groups don't count if only 1 group in parent */
7959 if (parent->groups == parent->groups->next) {
7960 pflags &= ~(SD_LOAD_BALANCE |
7961 SD_BALANCE_NEWIDLE |
7965 SD_SHARE_PKG_RESOURCES);
7966 if (nr_node_ids == 1)
7967 pflags &= ~SD_SERIALIZE;
7969 if (~cflags & pflags)
7975 static void free_rootdomain(struct root_domain *rd)
7977 cpupri_cleanup(&rd->cpupri);
7979 free_cpumask_var(rd->rto_mask);
7980 free_cpumask_var(rd->online);
7981 free_cpumask_var(rd->span);
7985 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7987 struct root_domain *old_rd = NULL;
7988 unsigned long flags;
7990 spin_lock_irqsave(&rq->lock, flags);
7995 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7998 cpumask_clear_cpu(rq->cpu, old_rd->span);
8001 * If we dont want to free the old_rt yet then
8002 * set old_rd to NULL to skip the freeing later
8005 if (!atomic_dec_and_test(&old_rd->refcount))
8009 atomic_inc(&rd->refcount);
8012 cpumask_set_cpu(rq->cpu, rd->span);
8013 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
8016 spin_unlock_irqrestore(&rq->lock, flags);
8019 free_rootdomain(old_rd);
8022 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8024 gfp_t gfp = GFP_KERNEL;
8026 memset(rd, 0, sizeof(*rd));
8031 if (!alloc_cpumask_var(&rd->span, gfp))
8033 if (!alloc_cpumask_var(&rd->online, gfp))
8035 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8038 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8043 free_cpumask_var(rd->rto_mask);
8045 free_cpumask_var(rd->online);
8047 free_cpumask_var(rd->span);
8052 static void init_defrootdomain(void)
8054 init_rootdomain(&def_root_domain, true);
8056 atomic_set(&def_root_domain.refcount, 1);
8059 static struct root_domain *alloc_rootdomain(void)
8061 struct root_domain *rd;
8063 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8067 if (init_rootdomain(rd, false) != 0) {
8076 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8077 * hold the hotplug lock.
8080 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8082 struct rq *rq = cpu_rq(cpu);
8083 struct sched_domain *tmp;
8085 /* Remove the sched domains which do not contribute to scheduling. */
8086 for (tmp = sd; tmp; ) {
8087 struct sched_domain *parent = tmp->parent;
8091 if (sd_parent_degenerate(tmp, parent)) {
8092 tmp->parent = parent->parent;
8094 parent->parent->child = tmp;
8099 if (sd && sd_degenerate(sd)) {
8105 sched_domain_debug(sd, cpu);
8107 rq_attach_root(rq, rd);
8108 rcu_assign_pointer(rq->sd, sd);
8111 /* cpus with isolated domains */
8112 static cpumask_var_t cpu_isolated_map;
8114 /* Setup the mask of cpus configured for isolated domains */
8115 static int __init isolated_cpu_setup(char *str)
8117 cpulist_parse(str, cpu_isolated_map);
8121 __setup("isolcpus=", isolated_cpu_setup);
8124 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8125 * to a function which identifies what group(along with sched group) a CPU
8126 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8127 * (due to the fact that we keep track of groups covered with a struct cpumask).
8129 * init_sched_build_groups will build a circular linked list of the groups
8130 * covered by the given span, and will set each group's ->cpumask correctly,
8131 * and ->cpu_power to 0.
8134 init_sched_build_groups(const struct cpumask *span,
8135 const struct cpumask *cpu_map,
8136 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8137 struct sched_group **sg,
8138 struct cpumask *tmpmask),
8139 struct cpumask *covered, struct cpumask *tmpmask)
8141 struct sched_group *first = NULL, *last = NULL;
8144 cpumask_clear(covered);
8146 for_each_cpu(i, span) {
8147 struct sched_group *sg;
8148 int group = group_fn(i, cpu_map, &sg, tmpmask);
8151 if (cpumask_test_cpu(i, covered))
8154 cpumask_clear(sched_group_cpus(sg));
8155 sg->__cpu_power = 0;
8157 for_each_cpu(j, span) {
8158 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8161 cpumask_set_cpu(j, covered);
8162 cpumask_set_cpu(j, sched_group_cpus(sg));
8173 #define SD_NODES_PER_DOMAIN 16
8178 * find_next_best_node - find the next node to include in a sched_domain
8179 * @node: node whose sched_domain we're building
8180 * @used_nodes: nodes already in the sched_domain
8182 * Find the next node to include in a given scheduling domain. Simply
8183 * finds the closest node not already in the @used_nodes map.
8185 * Should use nodemask_t.
8187 static int find_next_best_node(int node, nodemask_t *used_nodes)
8189 int i, n, val, min_val, best_node = 0;
8193 for (i = 0; i < nr_node_ids; i++) {
8194 /* Start at @node */
8195 n = (node + i) % nr_node_ids;
8197 if (!nr_cpus_node(n))
8200 /* Skip already used nodes */
8201 if (node_isset(n, *used_nodes))
8204 /* Simple min distance search */
8205 val = node_distance(node, n);
8207 if (val < min_val) {
8213 node_set(best_node, *used_nodes);
8218 * sched_domain_node_span - get a cpumask for a node's sched_domain
8219 * @node: node whose cpumask we're constructing
8220 * @span: resulting cpumask
8222 * Given a node, construct a good cpumask for its sched_domain to span. It
8223 * should be one that prevents unnecessary balancing, but also spreads tasks
8226 static void sched_domain_node_span(int node, struct cpumask *span)
8228 nodemask_t used_nodes;
8231 cpumask_clear(span);
8232 nodes_clear(used_nodes);
8234 cpumask_or(span, span, cpumask_of_node(node));
8235 node_set(node, used_nodes);
8237 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8238 int next_node = find_next_best_node(node, &used_nodes);
8240 cpumask_or(span, span, cpumask_of_node(next_node));
8243 #endif /* CONFIG_NUMA */
8245 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8248 * The cpus mask in sched_group and sched_domain hangs off the end.
8250 * ( See the the comments in include/linux/sched.h:struct sched_group
8251 * and struct sched_domain. )
8253 struct static_sched_group {
8254 struct sched_group sg;
8255 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8258 struct static_sched_domain {
8259 struct sched_domain sd;
8260 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8266 cpumask_var_t domainspan;
8267 cpumask_var_t covered;
8268 cpumask_var_t notcovered;
8270 cpumask_var_t nodemask;
8271 cpumask_var_t this_sibling_map;
8272 cpumask_var_t this_core_map;
8273 cpumask_var_t send_covered;
8274 cpumask_var_t tmpmask;
8275 struct sched_group **sched_group_nodes;
8276 struct root_domain *rd;
8280 sa_sched_groups = 0,
8285 sa_this_sibling_map,
8287 sa_sched_group_nodes,
8297 * SMT sched-domains:
8299 #ifdef CONFIG_SCHED_SMT
8300 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8301 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8304 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8305 struct sched_group **sg, struct cpumask *unused)
8308 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8311 #endif /* CONFIG_SCHED_SMT */
8314 * multi-core sched-domains:
8316 #ifdef CONFIG_SCHED_MC
8317 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8318 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8319 #endif /* CONFIG_SCHED_MC */
8321 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8323 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8324 struct sched_group **sg, struct cpumask *mask)
8328 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8329 group = cpumask_first(mask);
8331 *sg = &per_cpu(sched_group_core, group).sg;
8334 #elif defined(CONFIG_SCHED_MC)
8336 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8337 struct sched_group **sg, struct cpumask *unused)
8340 *sg = &per_cpu(sched_group_core, cpu).sg;
8345 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8346 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8349 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8350 struct sched_group **sg, struct cpumask *mask)
8353 #ifdef CONFIG_SCHED_MC
8354 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8355 group = cpumask_first(mask);
8356 #elif defined(CONFIG_SCHED_SMT)
8357 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8358 group = cpumask_first(mask);
8363 *sg = &per_cpu(sched_group_phys, group).sg;
8369 * The init_sched_build_groups can't handle what we want to do with node
8370 * groups, so roll our own. Now each node has its own list of groups which
8371 * gets dynamically allocated.
8373 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8374 static struct sched_group ***sched_group_nodes_bycpu;
8376 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8377 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8379 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8380 struct sched_group **sg,
8381 struct cpumask *nodemask)
8385 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8386 group = cpumask_first(nodemask);
8389 *sg = &per_cpu(sched_group_allnodes, group).sg;
8393 static void init_numa_sched_groups_power(struct sched_group *group_head)
8395 struct sched_group *sg = group_head;
8401 for_each_cpu(j, sched_group_cpus(sg)) {
8402 struct sched_domain *sd;
8404 sd = &per_cpu(phys_domains, j).sd;
8405 if (j != group_first_cpu(sd->groups)) {
8407 * Only add "power" once for each
8413 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
8416 } while (sg != group_head);
8419 static int build_numa_sched_groups(struct s_data *d,
8420 const struct cpumask *cpu_map, int num)
8422 struct sched_domain *sd;
8423 struct sched_group *sg, *prev;
8426 cpumask_clear(d->covered);
8427 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8428 if (cpumask_empty(d->nodemask)) {
8429 d->sched_group_nodes[num] = NULL;
8433 sched_domain_node_span(num, d->domainspan);
8434 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8436 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8439 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8443 d->sched_group_nodes[num] = sg;
8445 for_each_cpu(j, d->nodemask) {
8446 sd = &per_cpu(node_domains, j).sd;
8450 sg->__cpu_power = 0;
8451 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8453 cpumask_or(d->covered, d->covered, d->nodemask);
8456 for (j = 0; j < nr_node_ids; j++) {
8457 n = (num + j) % nr_node_ids;
8458 cpumask_complement(d->notcovered, d->covered);
8459 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8460 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8461 if (cpumask_empty(d->tmpmask))
8463 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8464 if (cpumask_empty(d->tmpmask))
8466 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8470 "Can not alloc domain group for node %d\n", j);
8473 sg->__cpu_power = 0;
8474 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8475 sg->next = prev->next;
8476 cpumask_or(d->covered, d->covered, d->tmpmask);
8483 #endif /* CONFIG_NUMA */
8486 /* Free memory allocated for various sched_group structures */
8487 static void free_sched_groups(const struct cpumask *cpu_map,
8488 struct cpumask *nodemask)
8492 for_each_cpu(cpu, cpu_map) {
8493 struct sched_group **sched_group_nodes
8494 = sched_group_nodes_bycpu[cpu];
8496 if (!sched_group_nodes)
8499 for (i = 0; i < nr_node_ids; i++) {
8500 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8502 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8503 if (cpumask_empty(nodemask))
8513 if (oldsg != sched_group_nodes[i])
8516 kfree(sched_group_nodes);
8517 sched_group_nodes_bycpu[cpu] = NULL;
8520 #else /* !CONFIG_NUMA */
8521 static void free_sched_groups(const struct cpumask *cpu_map,
8522 struct cpumask *nodemask)
8525 #endif /* CONFIG_NUMA */
8528 * Initialize sched groups cpu_power.
8530 * cpu_power indicates the capacity of sched group, which is used while
8531 * distributing the load between different sched groups in a sched domain.
8532 * Typically cpu_power for all the groups in a sched domain will be same unless
8533 * there are asymmetries in the topology. If there are asymmetries, group
8534 * having more cpu_power will pickup more load compared to the group having
8537 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8539 struct sched_domain *child;
8540 struct sched_group *group;
8544 WARN_ON(!sd || !sd->groups);
8546 if (cpu != group_first_cpu(sd->groups))
8551 sd->groups->__cpu_power = 0;
8554 power = SCHED_LOAD_SCALE;
8555 weight = cpumask_weight(sched_domain_span(sd));
8557 * SMT siblings share the power of a single core.
8558 * Usually multiple threads get a better yield out of
8559 * that one core than a single thread would have,
8560 * reflect that in sd->smt_gain.
8562 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8563 power *= sd->smt_gain;
8565 power >>= SCHED_LOAD_SHIFT;
8567 sg_inc_cpu_power(sd->groups, power);
8572 * Add cpu_power of each child group to this groups cpu_power.
8574 group = child->groups;
8576 sg_inc_cpu_power(sd->groups, group->__cpu_power);
8577 group = group->next;
8578 } while (group != child->groups);
8582 * Initializers for schedule domains
8583 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8586 #ifdef CONFIG_SCHED_DEBUG
8587 # define SD_INIT_NAME(sd, type) sd->name = #type
8589 # define SD_INIT_NAME(sd, type) do { } while (0)
8592 #define SD_INIT(sd, type) sd_init_##type(sd)
8594 #define SD_INIT_FUNC(type) \
8595 static noinline void sd_init_##type(struct sched_domain *sd) \
8597 memset(sd, 0, sizeof(*sd)); \
8598 *sd = SD_##type##_INIT; \
8599 sd->level = SD_LV_##type; \
8600 SD_INIT_NAME(sd, type); \
8605 SD_INIT_FUNC(ALLNODES)
8608 #ifdef CONFIG_SCHED_SMT
8609 SD_INIT_FUNC(SIBLING)
8611 #ifdef CONFIG_SCHED_MC
8615 static int default_relax_domain_level = -1;
8617 static int __init setup_relax_domain_level(char *str)
8621 val = simple_strtoul(str, NULL, 0);
8622 if (val < SD_LV_MAX)
8623 default_relax_domain_level = val;
8627 __setup("relax_domain_level=", setup_relax_domain_level);
8629 static void set_domain_attribute(struct sched_domain *sd,
8630 struct sched_domain_attr *attr)
8634 if (!attr || attr->relax_domain_level < 0) {
8635 if (default_relax_domain_level < 0)
8638 request = default_relax_domain_level;
8640 request = attr->relax_domain_level;
8641 if (request < sd->level) {
8642 /* turn off idle balance on this domain */
8643 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
8645 /* turn on idle balance on this domain */
8646 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
8650 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8651 const struct cpumask *cpu_map)
8654 case sa_sched_groups:
8655 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8656 d->sched_group_nodes = NULL;
8658 free_rootdomain(d->rd); /* fall through */
8660 free_cpumask_var(d->tmpmask); /* fall through */
8661 case sa_send_covered:
8662 free_cpumask_var(d->send_covered); /* fall through */
8663 case sa_this_core_map:
8664 free_cpumask_var(d->this_core_map); /* fall through */
8665 case sa_this_sibling_map:
8666 free_cpumask_var(d->this_sibling_map); /* fall through */
8668 free_cpumask_var(d->nodemask); /* fall through */
8669 case sa_sched_group_nodes:
8671 kfree(d->sched_group_nodes); /* fall through */
8673 free_cpumask_var(d->notcovered); /* fall through */
8675 free_cpumask_var(d->covered); /* fall through */
8677 free_cpumask_var(d->domainspan); /* fall through */
8684 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8685 const struct cpumask *cpu_map)
8688 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8690 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8691 return sa_domainspan;
8692 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8694 /* Allocate the per-node list of sched groups */
8695 d->sched_group_nodes = kcalloc(nr_node_ids,
8696 sizeof(struct sched_group *), GFP_KERNEL);
8697 if (!d->sched_group_nodes) {
8698 printk(KERN_WARNING "Can not alloc sched group node list\n");
8699 return sa_notcovered;
8701 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8703 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8704 return sa_sched_group_nodes;
8705 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8707 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8708 return sa_this_sibling_map;
8709 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8710 return sa_this_core_map;
8711 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8712 return sa_send_covered;
8713 d->rd = alloc_rootdomain();
8715 printk(KERN_WARNING "Cannot alloc root domain\n");
8718 return sa_rootdomain;
8721 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8722 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8724 struct sched_domain *sd = NULL;
8726 struct sched_domain *parent;
8729 if (cpumask_weight(cpu_map) >
8730 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8731 sd = &per_cpu(allnodes_domains, i).sd;
8732 SD_INIT(sd, ALLNODES);
8733 set_domain_attribute(sd, attr);
8734 cpumask_copy(sched_domain_span(sd), cpu_map);
8735 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8740 sd = &per_cpu(node_domains, i).sd;
8742 set_domain_attribute(sd, attr);
8743 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8744 sd->parent = parent;
8747 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8752 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8753 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8754 struct sched_domain *parent, int i)
8756 struct sched_domain *sd;
8757 sd = &per_cpu(phys_domains, i).sd;
8759 set_domain_attribute(sd, attr);
8760 cpumask_copy(sched_domain_span(sd), d->nodemask);
8761 sd->parent = parent;
8764 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8768 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8769 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8770 struct sched_domain *parent, int i)
8772 struct sched_domain *sd = parent;
8773 #ifdef CONFIG_SCHED_MC
8774 sd = &per_cpu(core_domains, i).sd;
8776 set_domain_attribute(sd, attr);
8777 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8778 sd->parent = parent;
8780 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8785 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8786 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8787 struct sched_domain *parent, int i)
8789 struct sched_domain *sd = parent;
8790 #ifdef CONFIG_SCHED_SMT
8791 sd = &per_cpu(cpu_domains, i).sd;
8792 SD_INIT(sd, SIBLING);
8793 set_domain_attribute(sd, attr);
8794 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8795 sd->parent = parent;
8797 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8802 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8803 const struct cpumask *cpu_map, int cpu)
8806 #ifdef CONFIG_SCHED_SMT
8807 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8808 cpumask_and(d->this_sibling_map, cpu_map,
8809 topology_thread_cpumask(cpu));
8810 if (cpu == cpumask_first(d->this_sibling_map))
8811 init_sched_build_groups(d->this_sibling_map, cpu_map,
8813 d->send_covered, d->tmpmask);
8816 #ifdef CONFIG_SCHED_MC
8817 case SD_LV_MC: /* set up multi-core groups */
8818 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8819 if (cpu == cpumask_first(d->this_core_map))
8820 init_sched_build_groups(d->this_core_map, cpu_map,
8822 d->send_covered, d->tmpmask);
8825 case SD_LV_CPU: /* set up physical groups */
8826 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8827 if (!cpumask_empty(d->nodemask))
8828 init_sched_build_groups(d->nodemask, cpu_map,
8830 d->send_covered, d->tmpmask);
8833 case SD_LV_ALLNODES:
8834 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8835 d->send_covered, d->tmpmask);
8844 * Build sched domains for a given set of cpus and attach the sched domains
8845 * to the individual cpus
8847 static int __build_sched_domains(const struct cpumask *cpu_map,
8848 struct sched_domain_attr *attr)
8850 enum s_alloc alloc_state = sa_none;
8852 struct sched_domain *sd;
8858 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8859 if (alloc_state != sa_rootdomain)
8861 alloc_state = sa_sched_groups;
8864 * Set up domains for cpus specified by the cpu_map.
8866 for_each_cpu(i, cpu_map) {
8867 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8870 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8871 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8872 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8873 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8876 for_each_cpu(i, cpu_map) {
8877 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8878 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8881 /* Set up physical groups */
8882 for (i = 0; i < nr_node_ids; i++)
8883 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8886 /* Set up node groups */
8888 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8890 for (i = 0; i < nr_node_ids; i++)
8891 if (build_numa_sched_groups(&d, cpu_map, i))
8895 /* Calculate CPU power for physical packages and nodes */
8896 #ifdef CONFIG_SCHED_SMT
8897 for_each_cpu(i, cpu_map) {
8898 sd = &per_cpu(cpu_domains, i).sd;
8899 init_sched_groups_power(i, sd);
8902 #ifdef CONFIG_SCHED_MC
8903 for_each_cpu(i, cpu_map) {
8904 sd = &per_cpu(core_domains, i).sd;
8905 init_sched_groups_power(i, sd);
8909 for_each_cpu(i, cpu_map) {
8910 sd = &per_cpu(phys_domains, i).sd;
8911 init_sched_groups_power(i, sd);
8915 for (i = 0; i < nr_node_ids; i++)
8916 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8918 if (d.sd_allnodes) {
8919 struct sched_group *sg;
8921 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8923 init_numa_sched_groups_power(sg);
8927 /* Attach the domains */
8928 for_each_cpu(i, cpu_map) {
8929 #ifdef CONFIG_SCHED_SMT
8930 sd = &per_cpu(cpu_domains, i).sd;
8931 #elif defined(CONFIG_SCHED_MC)
8932 sd = &per_cpu(core_domains, i).sd;
8934 sd = &per_cpu(phys_domains, i).sd;
8936 cpu_attach_domain(sd, d.rd, i);
8939 d.sched_group_nodes = NULL; /* don't free this we still need it */
8940 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8944 __free_domain_allocs(&d, alloc_state, cpu_map);
8948 static int build_sched_domains(const struct cpumask *cpu_map)
8950 return __build_sched_domains(cpu_map, NULL);
8953 static struct cpumask *doms_cur; /* current sched domains */
8954 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8955 static struct sched_domain_attr *dattr_cur;
8956 /* attribues of custom domains in 'doms_cur' */
8959 * Special case: If a kmalloc of a doms_cur partition (array of
8960 * cpumask) fails, then fallback to a single sched domain,
8961 * as determined by the single cpumask fallback_doms.
8963 static cpumask_var_t fallback_doms;
8966 * arch_update_cpu_topology lets virtualized architectures update the
8967 * cpu core maps. It is supposed to return 1 if the topology changed
8968 * or 0 if it stayed the same.
8970 int __attribute__((weak)) arch_update_cpu_topology(void)
8976 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8977 * For now this just excludes isolated cpus, but could be used to
8978 * exclude other special cases in the future.
8980 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8984 arch_update_cpu_topology();
8986 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8988 doms_cur = fallback_doms;
8989 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8991 err = build_sched_domains(doms_cur);
8992 register_sched_domain_sysctl();
8997 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8998 struct cpumask *tmpmask)
9000 free_sched_groups(cpu_map, tmpmask);
9004 * Detach sched domains from a group of cpus specified in cpu_map
9005 * These cpus will now be attached to the NULL domain
9007 static void detach_destroy_domains(const struct cpumask *cpu_map)
9009 /* Save because hotplug lock held. */
9010 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9013 for_each_cpu(i, cpu_map)
9014 cpu_attach_domain(NULL, &def_root_domain, i);
9015 synchronize_sched();
9016 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9019 /* handle null as "default" */
9020 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9021 struct sched_domain_attr *new, int idx_new)
9023 struct sched_domain_attr tmp;
9030 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9031 new ? (new + idx_new) : &tmp,
9032 sizeof(struct sched_domain_attr));
9036 * Partition sched domains as specified by the 'ndoms_new'
9037 * cpumasks in the array doms_new[] of cpumasks. This compares
9038 * doms_new[] to the current sched domain partitioning, doms_cur[].
9039 * It destroys each deleted domain and builds each new domain.
9041 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
9042 * The masks don't intersect (don't overlap.) We should setup one
9043 * sched domain for each mask. CPUs not in any of the cpumasks will
9044 * not be load balanced. If the same cpumask appears both in the
9045 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9048 * The passed in 'doms_new' should be kmalloc'd. This routine takes
9049 * ownership of it and will kfree it when done with it. If the caller
9050 * failed the kmalloc call, then it can pass in doms_new == NULL &&
9051 * ndoms_new == 1, and partition_sched_domains() will fallback to
9052 * the single partition 'fallback_doms', it also forces the domains
9055 * If doms_new == NULL it will be replaced with cpu_online_mask.
9056 * ndoms_new == 0 is a special case for destroying existing domains,
9057 * and it will not create the default domain.
9059 * Call with hotplug lock held
9061 /* FIXME: Change to struct cpumask *doms_new[] */
9062 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
9063 struct sched_domain_attr *dattr_new)
9068 mutex_lock(&sched_domains_mutex);
9070 /* always unregister in case we don't destroy any domains */
9071 unregister_sched_domain_sysctl();
9073 /* Let architecture update cpu core mappings. */
9074 new_topology = arch_update_cpu_topology();
9076 n = doms_new ? ndoms_new : 0;
9078 /* Destroy deleted domains */
9079 for (i = 0; i < ndoms_cur; i++) {
9080 for (j = 0; j < n && !new_topology; j++) {
9081 if (cpumask_equal(&doms_cur[i], &doms_new[j])
9082 && dattrs_equal(dattr_cur, i, dattr_new, j))
9085 /* no match - a current sched domain not in new doms_new[] */
9086 detach_destroy_domains(doms_cur + i);
9091 if (doms_new == NULL) {
9093 doms_new = fallback_doms;
9094 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
9095 WARN_ON_ONCE(dattr_new);
9098 /* Build new domains */
9099 for (i = 0; i < ndoms_new; i++) {
9100 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9101 if (cpumask_equal(&doms_new[i], &doms_cur[j])
9102 && dattrs_equal(dattr_new, i, dattr_cur, j))
9105 /* no match - add a new doms_new */
9106 __build_sched_domains(doms_new + i,
9107 dattr_new ? dattr_new + i : NULL);
9112 /* Remember the new sched domains */
9113 if (doms_cur != fallback_doms)
9115 kfree(dattr_cur); /* kfree(NULL) is safe */
9116 doms_cur = doms_new;
9117 dattr_cur = dattr_new;
9118 ndoms_cur = ndoms_new;
9120 register_sched_domain_sysctl();
9122 mutex_unlock(&sched_domains_mutex);
9125 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9126 static void arch_reinit_sched_domains(void)
9130 /* Destroy domains first to force the rebuild */
9131 partition_sched_domains(0, NULL, NULL);
9133 rebuild_sched_domains();
9137 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9139 unsigned int level = 0;
9141 if (sscanf(buf, "%u", &level) != 1)
9145 * level is always be positive so don't check for
9146 * level < POWERSAVINGS_BALANCE_NONE which is 0
9147 * What happens on 0 or 1 byte write,
9148 * need to check for count as well?
9151 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9155 sched_smt_power_savings = level;
9157 sched_mc_power_savings = level;
9159 arch_reinit_sched_domains();
9164 #ifdef CONFIG_SCHED_MC
9165 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9168 return sprintf(page, "%u\n", sched_mc_power_savings);
9170 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9171 const char *buf, size_t count)
9173 return sched_power_savings_store(buf, count, 0);
9175 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9176 sched_mc_power_savings_show,
9177 sched_mc_power_savings_store);
9180 #ifdef CONFIG_SCHED_SMT
9181 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9184 return sprintf(page, "%u\n", sched_smt_power_savings);
9186 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9187 const char *buf, size_t count)
9189 return sched_power_savings_store(buf, count, 1);
9191 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9192 sched_smt_power_savings_show,
9193 sched_smt_power_savings_store);
9196 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9200 #ifdef CONFIG_SCHED_SMT
9202 err = sysfs_create_file(&cls->kset.kobj,
9203 &attr_sched_smt_power_savings.attr);
9205 #ifdef CONFIG_SCHED_MC
9206 if (!err && mc_capable())
9207 err = sysfs_create_file(&cls->kset.kobj,
9208 &attr_sched_mc_power_savings.attr);
9212 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9214 #ifndef CONFIG_CPUSETS
9216 * Add online and remove offline CPUs from the scheduler domains.
9217 * When cpusets are enabled they take over this function.
9219 static int update_sched_domains(struct notifier_block *nfb,
9220 unsigned long action, void *hcpu)
9224 case CPU_ONLINE_FROZEN:
9226 case CPU_DEAD_FROZEN:
9227 partition_sched_domains(1, NULL, NULL);
9236 static int update_runtime(struct notifier_block *nfb,
9237 unsigned long action, void *hcpu)
9239 int cpu = (int)(long)hcpu;
9242 case CPU_DOWN_PREPARE:
9243 case CPU_DOWN_PREPARE_FROZEN:
9244 disable_runtime(cpu_rq(cpu));
9247 case CPU_DOWN_FAILED:
9248 case CPU_DOWN_FAILED_FROZEN:
9250 case CPU_ONLINE_FROZEN:
9251 enable_runtime(cpu_rq(cpu));
9259 void __init sched_init_smp(void)
9261 cpumask_var_t non_isolated_cpus;
9263 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9265 #if defined(CONFIG_NUMA)
9266 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9268 BUG_ON(sched_group_nodes_bycpu == NULL);
9271 mutex_lock(&sched_domains_mutex);
9272 arch_init_sched_domains(cpu_online_mask);
9273 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9274 if (cpumask_empty(non_isolated_cpus))
9275 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9276 mutex_unlock(&sched_domains_mutex);
9279 #ifndef CONFIG_CPUSETS
9280 /* XXX: Theoretical race here - CPU may be hotplugged now */
9281 hotcpu_notifier(update_sched_domains, 0);
9284 /* RT runtime code needs to handle some hotplug events */
9285 hotcpu_notifier(update_runtime, 0);
9289 /* Move init over to a non-isolated CPU */
9290 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9292 sched_init_granularity();
9293 free_cpumask_var(non_isolated_cpus);
9295 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9296 init_sched_rt_class();
9299 void __init sched_init_smp(void)
9301 sched_init_granularity();
9303 #endif /* CONFIG_SMP */
9305 const_debug unsigned int sysctl_timer_migration = 1;
9307 int in_sched_functions(unsigned long addr)
9309 return in_lock_functions(addr) ||
9310 (addr >= (unsigned long)__sched_text_start
9311 && addr < (unsigned long)__sched_text_end);
9314 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9316 cfs_rq->tasks_timeline = RB_ROOT;
9317 INIT_LIST_HEAD(&cfs_rq->tasks);
9318 #ifdef CONFIG_FAIR_GROUP_SCHED
9321 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9324 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9326 struct rt_prio_array *array;
9329 array = &rt_rq->active;
9330 for (i = 0; i < MAX_RT_PRIO; i++) {
9331 INIT_LIST_HEAD(array->queue + i);
9332 __clear_bit(i, array->bitmap);
9334 /* delimiter for bitsearch: */
9335 __set_bit(MAX_RT_PRIO, array->bitmap);
9337 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9338 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9340 rt_rq->highest_prio.next = MAX_RT_PRIO;
9344 rt_rq->rt_nr_migratory = 0;
9345 rt_rq->overloaded = 0;
9346 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9350 rt_rq->rt_throttled = 0;
9351 rt_rq->rt_runtime = 0;
9352 spin_lock_init(&rt_rq->rt_runtime_lock);
9354 #ifdef CONFIG_RT_GROUP_SCHED
9355 rt_rq->rt_nr_boosted = 0;
9360 #ifdef CONFIG_FAIR_GROUP_SCHED
9361 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9362 struct sched_entity *se, int cpu, int add,
9363 struct sched_entity *parent)
9365 struct rq *rq = cpu_rq(cpu);
9366 tg->cfs_rq[cpu] = cfs_rq;
9367 init_cfs_rq(cfs_rq, rq);
9370 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9373 /* se could be NULL for init_task_group */
9378 se->cfs_rq = &rq->cfs;
9380 se->cfs_rq = parent->my_q;
9383 se->load.weight = tg->shares;
9384 se->load.inv_weight = 0;
9385 se->parent = parent;
9389 #ifdef CONFIG_RT_GROUP_SCHED
9390 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9391 struct sched_rt_entity *rt_se, int cpu, int add,
9392 struct sched_rt_entity *parent)
9394 struct rq *rq = cpu_rq(cpu);
9396 tg->rt_rq[cpu] = rt_rq;
9397 init_rt_rq(rt_rq, rq);
9399 rt_rq->rt_se = rt_se;
9400 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9402 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9404 tg->rt_se[cpu] = rt_se;
9409 rt_se->rt_rq = &rq->rt;
9411 rt_se->rt_rq = parent->my_q;
9413 rt_se->my_q = rt_rq;
9414 rt_se->parent = parent;
9415 INIT_LIST_HEAD(&rt_se->run_list);
9419 void __init sched_init(void)
9422 unsigned long alloc_size = 0, ptr;
9424 #ifdef CONFIG_FAIR_GROUP_SCHED
9425 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9427 #ifdef CONFIG_RT_GROUP_SCHED
9428 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9430 #ifdef CONFIG_USER_SCHED
9433 #ifdef CONFIG_CPUMASK_OFFSTACK
9434 alloc_size += num_possible_cpus() * cpumask_size();
9437 * As sched_init() is called before page_alloc is setup,
9438 * we use alloc_bootmem().
9441 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9443 #ifdef CONFIG_FAIR_GROUP_SCHED
9444 init_task_group.se = (struct sched_entity **)ptr;
9445 ptr += nr_cpu_ids * sizeof(void **);
9447 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9448 ptr += nr_cpu_ids * sizeof(void **);
9450 #ifdef CONFIG_USER_SCHED
9451 root_task_group.se = (struct sched_entity **)ptr;
9452 ptr += nr_cpu_ids * sizeof(void **);
9454 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9455 ptr += nr_cpu_ids * sizeof(void **);
9456 #endif /* CONFIG_USER_SCHED */
9457 #endif /* CONFIG_FAIR_GROUP_SCHED */
9458 #ifdef CONFIG_RT_GROUP_SCHED
9459 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9460 ptr += nr_cpu_ids * sizeof(void **);
9462 init_task_group.rt_rq = (struct rt_rq **)ptr;
9463 ptr += nr_cpu_ids * sizeof(void **);
9465 #ifdef CONFIG_USER_SCHED
9466 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9467 ptr += nr_cpu_ids * sizeof(void **);
9469 root_task_group.rt_rq = (struct rt_rq **)ptr;
9470 ptr += nr_cpu_ids * sizeof(void **);
9471 #endif /* CONFIG_USER_SCHED */
9472 #endif /* CONFIG_RT_GROUP_SCHED */
9473 #ifdef CONFIG_CPUMASK_OFFSTACK
9474 for_each_possible_cpu(i) {
9475 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9476 ptr += cpumask_size();
9478 #endif /* CONFIG_CPUMASK_OFFSTACK */
9482 init_defrootdomain();
9485 init_rt_bandwidth(&def_rt_bandwidth,
9486 global_rt_period(), global_rt_runtime());
9488 #ifdef CONFIG_RT_GROUP_SCHED
9489 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9490 global_rt_period(), global_rt_runtime());
9491 #ifdef CONFIG_USER_SCHED
9492 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9493 global_rt_period(), RUNTIME_INF);
9494 #endif /* CONFIG_USER_SCHED */
9495 #endif /* CONFIG_RT_GROUP_SCHED */
9497 #ifdef CONFIG_GROUP_SCHED
9498 list_add(&init_task_group.list, &task_groups);
9499 INIT_LIST_HEAD(&init_task_group.children);
9501 #ifdef CONFIG_USER_SCHED
9502 INIT_LIST_HEAD(&root_task_group.children);
9503 init_task_group.parent = &root_task_group;
9504 list_add(&init_task_group.siblings, &root_task_group.children);
9505 #endif /* CONFIG_USER_SCHED */
9506 #endif /* CONFIG_GROUP_SCHED */
9508 for_each_possible_cpu(i) {
9512 spin_lock_init(&rq->lock);
9514 rq->calc_load_active = 0;
9515 rq->calc_load_update = jiffies + LOAD_FREQ;
9516 init_cfs_rq(&rq->cfs, rq);
9517 init_rt_rq(&rq->rt, rq);
9518 #ifdef CONFIG_FAIR_GROUP_SCHED
9519 init_task_group.shares = init_task_group_load;
9520 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9521 #ifdef CONFIG_CGROUP_SCHED
9523 * How much cpu bandwidth does init_task_group get?
9525 * In case of task-groups formed thr' the cgroup filesystem, it
9526 * gets 100% of the cpu resources in the system. This overall
9527 * system cpu resource is divided among the tasks of
9528 * init_task_group and its child task-groups in a fair manner,
9529 * based on each entity's (task or task-group's) weight
9530 * (se->load.weight).
9532 * In other words, if init_task_group has 10 tasks of weight
9533 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9534 * then A0's share of the cpu resource is:
9536 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9538 * We achieve this by letting init_task_group's tasks sit
9539 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9541 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9542 #elif defined CONFIG_USER_SCHED
9543 root_task_group.shares = NICE_0_LOAD;
9544 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9546 * In case of task-groups formed thr' the user id of tasks,
9547 * init_task_group represents tasks belonging to root user.
9548 * Hence it forms a sibling of all subsequent groups formed.
9549 * In this case, init_task_group gets only a fraction of overall
9550 * system cpu resource, based on the weight assigned to root
9551 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9552 * by letting tasks of init_task_group sit in a separate cfs_rq
9553 * (init_tg_cfs_rq) and having one entity represent this group of
9554 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9556 init_tg_cfs_entry(&init_task_group,
9557 &per_cpu(init_tg_cfs_rq, i),
9558 &per_cpu(init_sched_entity, i), i, 1,
9559 root_task_group.se[i]);
9562 #endif /* CONFIG_FAIR_GROUP_SCHED */
9564 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9565 #ifdef CONFIG_RT_GROUP_SCHED
9566 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9567 #ifdef CONFIG_CGROUP_SCHED
9568 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9569 #elif defined CONFIG_USER_SCHED
9570 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9571 init_tg_rt_entry(&init_task_group,
9572 &per_cpu(init_rt_rq, i),
9573 &per_cpu(init_sched_rt_entity, i), i, 1,
9574 root_task_group.rt_se[i]);
9578 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9579 rq->cpu_load[j] = 0;
9583 rq->post_schedule = 0;
9584 rq->active_balance = 0;
9585 rq->next_balance = jiffies;
9589 rq->migration_thread = NULL;
9590 INIT_LIST_HEAD(&rq->migration_queue);
9591 rq_attach_root(rq, &def_root_domain);
9594 atomic_set(&rq->nr_iowait, 0);
9597 set_load_weight(&init_task);
9599 #ifdef CONFIG_PREEMPT_NOTIFIERS
9600 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9604 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9607 #ifdef CONFIG_RT_MUTEXES
9608 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9612 * The boot idle thread does lazy MMU switching as well:
9614 atomic_inc(&init_mm.mm_count);
9615 enter_lazy_tlb(&init_mm, current);
9618 * Make us the idle thread. Technically, schedule() should not be
9619 * called from this thread, however somewhere below it might be,
9620 * but because we are the idle thread, we just pick up running again
9621 * when this runqueue becomes "idle".
9623 init_idle(current, smp_processor_id());
9625 calc_load_update = jiffies + LOAD_FREQ;
9628 * During early bootup we pretend to be a normal task:
9630 current->sched_class = &fair_sched_class;
9632 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9633 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9636 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9637 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9639 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9642 perf_counter_init();
9644 scheduler_running = 1;
9647 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9648 static inline int preempt_count_equals(int preempt_offset)
9650 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9652 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9655 void __might_sleep(char *file, int line, int preempt_offset)
9658 static unsigned long prev_jiffy; /* ratelimiting */
9660 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9661 system_state != SYSTEM_RUNNING || oops_in_progress)
9663 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9665 prev_jiffy = jiffies;
9668 "BUG: sleeping function called from invalid context at %s:%d\n",
9671 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9672 in_atomic(), irqs_disabled(),
9673 current->pid, current->comm);
9675 debug_show_held_locks(current);
9676 if (irqs_disabled())
9677 print_irqtrace_events(current);
9681 EXPORT_SYMBOL(__might_sleep);
9684 #ifdef CONFIG_MAGIC_SYSRQ
9685 static void normalize_task(struct rq *rq, struct task_struct *p)
9689 update_rq_clock(rq);
9690 on_rq = p->se.on_rq;
9692 deactivate_task(rq, p, 0);
9693 __setscheduler(rq, p, SCHED_NORMAL, 0);
9695 activate_task(rq, p, 0);
9696 resched_task(rq->curr);
9700 void normalize_rt_tasks(void)
9702 struct task_struct *g, *p;
9703 unsigned long flags;
9706 read_lock_irqsave(&tasklist_lock, flags);
9707 do_each_thread(g, p) {
9709 * Only normalize user tasks:
9714 p->se.exec_start = 0;
9715 #ifdef CONFIG_SCHEDSTATS
9716 p->se.wait_start = 0;
9717 p->se.sleep_start = 0;
9718 p->se.block_start = 0;
9723 * Renice negative nice level userspace
9726 if (TASK_NICE(p) < 0 && p->mm)
9727 set_user_nice(p, 0);
9731 spin_lock(&p->pi_lock);
9732 rq = __task_rq_lock(p);
9734 normalize_task(rq, p);
9736 __task_rq_unlock(rq);
9737 spin_unlock(&p->pi_lock);
9738 } while_each_thread(g, p);
9740 read_unlock_irqrestore(&tasklist_lock, flags);
9743 #endif /* CONFIG_MAGIC_SYSRQ */
9747 * These functions are only useful for the IA64 MCA handling.
9749 * They can only be called when the whole system has been
9750 * stopped - every CPU needs to be quiescent, and no scheduling
9751 * activity can take place. Using them for anything else would
9752 * be a serious bug, and as a result, they aren't even visible
9753 * under any other configuration.
9757 * curr_task - return the current task for a given cpu.
9758 * @cpu: the processor in question.
9760 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9762 struct task_struct *curr_task(int cpu)
9764 return cpu_curr(cpu);
9768 * set_curr_task - set the current task for a given cpu.
9769 * @cpu: the processor in question.
9770 * @p: the task pointer to set.
9772 * Description: This function must only be used when non-maskable interrupts
9773 * are serviced on a separate stack. It allows the architecture to switch the
9774 * notion of the current task on a cpu in a non-blocking manner. This function
9775 * must be called with all CPU's synchronized, and interrupts disabled, the
9776 * and caller must save the original value of the current task (see
9777 * curr_task() above) and restore that value before reenabling interrupts and
9778 * re-starting the system.
9780 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9782 void set_curr_task(int cpu, struct task_struct *p)
9789 #ifdef CONFIG_FAIR_GROUP_SCHED
9790 static void free_fair_sched_group(struct task_group *tg)
9794 for_each_possible_cpu(i) {
9796 kfree(tg->cfs_rq[i]);
9806 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9808 struct cfs_rq *cfs_rq;
9809 struct sched_entity *se;
9813 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9816 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9820 tg->shares = NICE_0_LOAD;
9822 for_each_possible_cpu(i) {
9825 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9826 GFP_KERNEL, cpu_to_node(i));
9830 se = kzalloc_node(sizeof(struct sched_entity),
9831 GFP_KERNEL, cpu_to_node(i));
9835 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9844 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9846 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9847 &cpu_rq(cpu)->leaf_cfs_rq_list);
9850 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9852 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9854 #else /* !CONFG_FAIR_GROUP_SCHED */
9855 static inline void free_fair_sched_group(struct task_group *tg)
9860 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9865 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9869 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9872 #endif /* CONFIG_FAIR_GROUP_SCHED */
9874 #ifdef CONFIG_RT_GROUP_SCHED
9875 static void free_rt_sched_group(struct task_group *tg)
9879 destroy_rt_bandwidth(&tg->rt_bandwidth);
9881 for_each_possible_cpu(i) {
9883 kfree(tg->rt_rq[i]);
9885 kfree(tg->rt_se[i]);
9893 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9895 struct rt_rq *rt_rq;
9896 struct sched_rt_entity *rt_se;
9900 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9903 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9907 init_rt_bandwidth(&tg->rt_bandwidth,
9908 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9910 for_each_possible_cpu(i) {
9913 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9914 GFP_KERNEL, cpu_to_node(i));
9918 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9919 GFP_KERNEL, cpu_to_node(i));
9923 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9932 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9934 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9935 &cpu_rq(cpu)->leaf_rt_rq_list);
9938 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9940 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9942 #else /* !CONFIG_RT_GROUP_SCHED */
9943 static inline void free_rt_sched_group(struct task_group *tg)
9948 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9953 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9957 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9960 #endif /* CONFIG_RT_GROUP_SCHED */
9962 #ifdef CONFIG_GROUP_SCHED
9963 static void free_sched_group(struct task_group *tg)
9965 free_fair_sched_group(tg);
9966 free_rt_sched_group(tg);
9970 /* allocate runqueue etc for a new task group */
9971 struct task_group *sched_create_group(struct task_group *parent)
9973 struct task_group *tg;
9974 unsigned long flags;
9977 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9979 return ERR_PTR(-ENOMEM);
9981 if (!alloc_fair_sched_group(tg, parent))
9984 if (!alloc_rt_sched_group(tg, parent))
9987 spin_lock_irqsave(&task_group_lock, flags);
9988 for_each_possible_cpu(i) {
9989 register_fair_sched_group(tg, i);
9990 register_rt_sched_group(tg, i);
9992 list_add_rcu(&tg->list, &task_groups);
9994 WARN_ON(!parent); /* root should already exist */
9996 tg->parent = parent;
9997 INIT_LIST_HEAD(&tg->children);
9998 list_add_rcu(&tg->siblings, &parent->children);
9999 spin_unlock_irqrestore(&task_group_lock, flags);
10004 free_sched_group(tg);
10005 return ERR_PTR(-ENOMEM);
10008 /* rcu callback to free various structures associated with a task group */
10009 static void free_sched_group_rcu(struct rcu_head *rhp)
10011 /* now it should be safe to free those cfs_rqs */
10012 free_sched_group(container_of(rhp, struct task_group, rcu));
10015 /* Destroy runqueue etc associated with a task group */
10016 void sched_destroy_group(struct task_group *tg)
10018 unsigned long flags;
10021 spin_lock_irqsave(&task_group_lock, flags);
10022 for_each_possible_cpu(i) {
10023 unregister_fair_sched_group(tg, i);
10024 unregister_rt_sched_group(tg, i);
10026 list_del_rcu(&tg->list);
10027 list_del_rcu(&tg->siblings);
10028 spin_unlock_irqrestore(&task_group_lock, flags);
10030 /* wait for possible concurrent references to cfs_rqs complete */
10031 call_rcu(&tg->rcu, free_sched_group_rcu);
10034 /* change task's runqueue when it moves between groups.
10035 * The caller of this function should have put the task in its new group
10036 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10037 * reflect its new group.
10039 void sched_move_task(struct task_struct *tsk)
10041 int on_rq, running;
10042 unsigned long flags;
10045 rq = task_rq_lock(tsk, &flags);
10047 update_rq_clock(rq);
10049 running = task_current(rq, tsk);
10050 on_rq = tsk->se.on_rq;
10053 dequeue_task(rq, tsk, 0);
10054 if (unlikely(running))
10055 tsk->sched_class->put_prev_task(rq, tsk);
10057 set_task_rq(tsk, task_cpu(tsk));
10059 #ifdef CONFIG_FAIR_GROUP_SCHED
10060 if (tsk->sched_class->moved_group)
10061 tsk->sched_class->moved_group(tsk);
10064 if (unlikely(running))
10065 tsk->sched_class->set_curr_task(rq);
10067 enqueue_task(rq, tsk, 0);
10069 task_rq_unlock(rq, &flags);
10071 #endif /* CONFIG_GROUP_SCHED */
10073 #ifdef CONFIG_FAIR_GROUP_SCHED
10074 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10076 struct cfs_rq *cfs_rq = se->cfs_rq;
10081 dequeue_entity(cfs_rq, se, 0);
10083 se->load.weight = shares;
10084 se->load.inv_weight = 0;
10087 enqueue_entity(cfs_rq, se, 0);
10090 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10092 struct cfs_rq *cfs_rq = se->cfs_rq;
10093 struct rq *rq = cfs_rq->rq;
10094 unsigned long flags;
10096 spin_lock_irqsave(&rq->lock, flags);
10097 __set_se_shares(se, shares);
10098 spin_unlock_irqrestore(&rq->lock, flags);
10101 static DEFINE_MUTEX(shares_mutex);
10103 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10106 unsigned long flags;
10109 * We can't change the weight of the root cgroup.
10114 if (shares < MIN_SHARES)
10115 shares = MIN_SHARES;
10116 else if (shares > MAX_SHARES)
10117 shares = MAX_SHARES;
10119 mutex_lock(&shares_mutex);
10120 if (tg->shares == shares)
10123 spin_lock_irqsave(&task_group_lock, flags);
10124 for_each_possible_cpu(i)
10125 unregister_fair_sched_group(tg, i);
10126 list_del_rcu(&tg->siblings);
10127 spin_unlock_irqrestore(&task_group_lock, flags);
10129 /* wait for any ongoing reference to this group to finish */
10130 synchronize_sched();
10133 * Now we are free to modify the group's share on each cpu
10134 * w/o tripping rebalance_share or load_balance_fair.
10136 tg->shares = shares;
10137 for_each_possible_cpu(i) {
10139 * force a rebalance
10141 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10142 set_se_shares(tg->se[i], shares);
10146 * Enable load balance activity on this group, by inserting it back on
10147 * each cpu's rq->leaf_cfs_rq_list.
10149 spin_lock_irqsave(&task_group_lock, flags);
10150 for_each_possible_cpu(i)
10151 register_fair_sched_group(tg, i);
10152 list_add_rcu(&tg->siblings, &tg->parent->children);
10153 spin_unlock_irqrestore(&task_group_lock, flags);
10155 mutex_unlock(&shares_mutex);
10159 unsigned long sched_group_shares(struct task_group *tg)
10165 #ifdef CONFIG_RT_GROUP_SCHED
10167 * Ensure that the real time constraints are schedulable.
10169 static DEFINE_MUTEX(rt_constraints_mutex);
10171 static unsigned long to_ratio(u64 period, u64 runtime)
10173 if (runtime == RUNTIME_INF)
10176 return div64_u64(runtime << 20, period);
10179 /* Must be called with tasklist_lock held */
10180 static inline int tg_has_rt_tasks(struct task_group *tg)
10182 struct task_struct *g, *p;
10184 do_each_thread(g, p) {
10185 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10187 } while_each_thread(g, p);
10192 struct rt_schedulable_data {
10193 struct task_group *tg;
10198 static int tg_schedulable(struct task_group *tg, void *data)
10200 struct rt_schedulable_data *d = data;
10201 struct task_group *child;
10202 unsigned long total, sum = 0;
10203 u64 period, runtime;
10205 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10206 runtime = tg->rt_bandwidth.rt_runtime;
10209 period = d->rt_period;
10210 runtime = d->rt_runtime;
10213 #ifdef CONFIG_USER_SCHED
10214 if (tg == &root_task_group) {
10215 period = global_rt_period();
10216 runtime = global_rt_runtime();
10221 * Cannot have more runtime than the period.
10223 if (runtime > period && runtime != RUNTIME_INF)
10227 * Ensure we don't starve existing RT tasks.
10229 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10232 total = to_ratio(period, runtime);
10235 * Nobody can have more than the global setting allows.
10237 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10241 * The sum of our children's runtime should not exceed our own.
10243 list_for_each_entry_rcu(child, &tg->children, siblings) {
10244 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10245 runtime = child->rt_bandwidth.rt_runtime;
10247 if (child == d->tg) {
10248 period = d->rt_period;
10249 runtime = d->rt_runtime;
10252 sum += to_ratio(period, runtime);
10261 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10263 struct rt_schedulable_data data = {
10265 .rt_period = period,
10266 .rt_runtime = runtime,
10269 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10272 static int tg_set_bandwidth(struct task_group *tg,
10273 u64 rt_period, u64 rt_runtime)
10277 mutex_lock(&rt_constraints_mutex);
10278 read_lock(&tasklist_lock);
10279 err = __rt_schedulable(tg, rt_period, rt_runtime);
10283 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10284 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10285 tg->rt_bandwidth.rt_runtime = rt_runtime;
10287 for_each_possible_cpu(i) {
10288 struct rt_rq *rt_rq = tg->rt_rq[i];
10290 spin_lock(&rt_rq->rt_runtime_lock);
10291 rt_rq->rt_runtime = rt_runtime;
10292 spin_unlock(&rt_rq->rt_runtime_lock);
10294 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10296 read_unlock(&tasklist_lock);
10297 mutex_unlock(&rt_constraints_mutex);
10302 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10304 u64 rt_runtime, rt_period;
10306 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10307 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10308 if (rt_runtime_us < 0)
10309 rt_runtime = RUNTIME_INF;
10311 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10314 long sched_group_rt_runtime(struct task_group *tg)
10318 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10321 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10322 do_div(rt_runtime_us, NSEC_PER_USEC);
10323 return rt_runtime_us;
10326 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10328 u64 rt_runtime, rt_period;
10330 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10331 rt_runtime = tg->rt_bandwidth.rt_runtime;
10333 if (rt_period == 0)
10336 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10339 long sched_group_rt_period(struct task_group *tg)
10343 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10344 do_div(rt_period_us, NSEC_PER_USEC);
10345 return rt_period_us;
10348 static int sched_rt_global_constraints(void)
10350 u64 runtime, period;
10353 if (sysctl_sched_rt_period <= 0)
10356 runtime = global_rt_runtime();
10357 period = global_rt_period();
10360 * Sanity check on the sysctl variables.
10362 if (runtime > period && runtime != RUNTIME_INF)
10365 mutex_lock(&rt_constraints_mutex);
10366 read_lock(&tasklist_lock);
10367 ret = __rt_schedulable(NULL, 0, 0);
10368 read_unlock(&tasklist_lock);
10369 mutex_unlock(&rt_constraints_mutex);
10374 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10376 /* Don't accept realtime tasks when there is no way for them to run */
10377 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10383 #else /* !CONFIG_RT_GROUP_SCHED */
10384 static int sched_rt_global_constraints(void)
10386 unsigned long flags;
10389 if (sysctl_sched_rt_period <= 0)
10393 * There's always some RT tasks in the root group
10394 * -- migration, kstopmachine etc..
10396 if (sysctl_sched_rt_runtime == 0)
10399 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10400 for_each_possible_cpu(i) {
10401 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10403 spin_lock(&rt_rq->rt_runtime_lock);
10404 rt_rq->rt_runtime = global_rt_runtime();
10405 spin_unlock(&rt_rq->rt_runtime_lock);
10407 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10411 #endif /* CONFIG_RT_GROUP_SCHED */
10413 int sched_rt_handler(struct ctl_table *table, int write,
10414 struct file *filp, void __user *buffer, size_t *lenp,
10418 int old_period, old_runtime;
10419 static DEFINE_MUTEX(mutex);
10421 mutex_lock(&mutex);
10422 old_period = sysctl_sched_rt_period;
10423 old_runtime = sysctl_sched_rt_runtime;
10425 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
10427 if (!ret && write) {
10428 ret = sched_rt_global_constraints();
10430 sysctl_sched_rt_period = old_period;
10431 sysctl_sched_rt_runtime = old_runtime;
10433 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10434 def_rt_bandwidth.rt_period =
10435 ns_to_ktime(global_rt_period());
10438 mutex_unlock(&mutex);
10443 #ifdef CONFIG_CGROUP_SCHED
10445 /* return corresponding task_group object of a cgroup */
10446 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10448 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10449 struct task_group, css);
10452 static struct cgroup_subsys_state *
10453 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10455 struct task_group *tg, *parent;
10457 if (!cgrp->parent) {
10458 /* This is early initialization for the top cgroup */
10459 return &init_task_group.css;
10462 parent = cgroup_tg(cgrp->parent);
10463 tg = sched_create_group(parent);
10465 return ERR_PTR(-ENOMEM);
10471 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10473 struct task_group *tg = cgroup_tg(cgrp);
10475 sched_destroy_group(tg);
10479 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10480 struct task_struct *tsk)
10482 #ifdef CONFIG_RT_GROUP_SCHED
10483 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10486 /* We don't support RT-tasks being in separate groups */
10487 if (tsk->sched_class != &fair_sched_class)
10495 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10496 struct cgroup *old_cont, struct task_struct *tsk)
10498 sched_move_task(tsk);
10501 #ifdef CONFIG_FAIR_GROUP_SCHED
10502 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10505 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10508 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10510 struct task_group *tg = cgroup_tg(cgrp);
10512 return (u64) tg->shares;
10514 #endif /* CONFIG_FAIR_GROUP_SCHED */
10516 #ifdef CONFIG_RT_GROUP_SCHED
10517 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10520 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10523 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10525 return sched_group_rt_runtime(cgroup_tg(cgrp));
10528 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10531 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10534 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10536 return sched_group_rt_period(cgroup_tg(cgrp));
10538 #endif /* CONFIG_RT_GROUP_SCHED */
10540 static struct cftype cpu_files[] = {
10541 #ifdef CONFIG_FAIR_GROUP_SCHED
10544 .read_u64 = cpu_shares_read_u64,
10545 .write_u64 = cpu_shares_write_u64,
10548 #ifdef CONFIG_RT_GROUP_SCHED
10550 .name = "rt_runtime_us",
10551 .read_s64 = cpu_rt_runtime_read,
10552 .write_s64 = cpu_rt_runtime_write,
10555 .name = "rt_period_us",
10556 .read_u64 = cpu_rt_period_read_uint,
10557 .write_u64 = cpu_rt_period_write_uint,
10562 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10564 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10567 struct cgroup_subsys cpu_cgroup_subsys = {
10569 .create = cpu_cgroup_create,
10570 .destroy = cpu_cgroup_destroy,
10571 .can_attach = cpu_cgroup_can_attach,
10572 .attach = cpu_cgroup_attach,
10573 .populate = cpu_cgroup_populate,
10574 .subsys_id = cpu_cgroup_subsys_id,
10578 #endif /* CONFIG_CGROUP_SCHED */
10580 #ifdef CONFIG_CGROUP_CPUACCT
10583 * CPU accounting code for task groups.
10585 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10586 * (balbir@in.ibm.com).
10589 /* track cpu usage of a group of tasks and its child groups */
10591 struct cgroup_subsys_state css;
10592 /* cpuusage holds pointer to a u64-type object on every cpu */
10594 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10595 struct cpuacct *parent;
10598 struct cgroup_subsys cpuacct_subsys;
10600 /* return cpu accounting group corresponding to this container */
10601 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10603 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10604 struct cpuacct, css);
10607 /* return cpu accounting group to which this task belongs */
10608 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10610 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10611 struct cpuacct, css);
10614 /* create a new cpu accounting group */
10615 static struct cgroup_subsys_state *cpuacct_create(
10616 struct cgroup_subsys *ss, struct cgroup *cgrp)
10618 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10624 ca->cpuusage = alloc_percpu(u64);
10628 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10629 if (percpu_counter_init(&ca->cpustat[i], 0))
10630 goto out_free_counters;
10633 ca->parent = cgroup_ca(cgrp->parent);
10639 percpu_counter_destroy(&ca->cpustat[i]);
10640 free_percpu(ca->cpuusage);
10644 return ERR_PTR(-ENOMEM);
10647 /* destroy an existing cpu accounting group */
10649 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10651 struct cpuacct *ca = cgroup_ca(cgrp);
10654 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10655 percpu_counter_destroy(&ca->cpustat[i]);
10656 free_percpu(ca->cpuusage);
10660 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10662 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10665 #ifndef CONFIG_64BIT
10667 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10669 spin_lock_irq(&cpu_rq(cpu)->lock);
10671 spin_unlock_irq(&cpu_rq(cpu)->lock);
10679 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10681 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10683 #ifndef CONFIG_64BIT
10685 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10687 spin_lock_irq(&cpu_rq(cpu)->lock);
10689 spin_unlock_irq(&cpu_rq(cpu)->lock);
10695 /* return total cpu usage (in nanoseconds) of a group */
10696 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10698 struct cpuacct *ca = cgroup_ca(cgrp);
10699 u64 totalcpuusage = 0;
10702 for_each_present_cpu(i)
10703 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10705 return totalcpuusage;
10708 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10711 struct cpuacct *ca = cgroup_ca(cgrp);
10720 for_each_present_cpu(i)
10721 cpuacct_cpuusage_write(ca, i, 0);
10727 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10728 struct seq_file *m)
10730 struct cpuacct *ca = cgroup_ca(cgroup);
10734 for_each_present_cpu(i) {
10735 percpu = cpuacct_cpuusage_read(ca, i);
10736 seq_printf(m, "%llu ", (unsigned long long) percpu);
10738 seq_printf(m, "\n");
10742 static const char *cpuacct_stat_desc[] = {
10743 [CPUACCT_STAT_USER] = "user",
10744 [CPUACCT_STAT_SYSTEM] = "system",
10747 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10748 struct cgroup_map_cb *cb)
10750 struct cpuacct *ca = cgroup_ca(cgrp);
10753 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10754 s64 val = percpu_counter_read(&ca->cpustat[i]);
10755 val = cputime64_to_clock_t(val);
10756 cb->fill(cb, cpuacct_stat_desc[i], val);
10761 static struct cftype files[] = {
10764 .read_u64 = cpuusage_read,
10765 .write_u64 = cpuusage_write,
10768 .name = "usage_percpu",
10769 .read_seq_string = cpuacct_percpu_seq_read,
10773 .read_map = cpuacct_stats_show,
10777 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10779 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10783 * charge this task's execution time to its accounting group.
10785 * called with rq->lock held.
10787 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10789 struct cpuacct *ca;
10792 if (unlikely(!cpuacct_subsys.active))
10795 cpu = task_cpu(tsk);
10801 for (; ca; ca = ca->parent) {
10802 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10803 *cpuusage += cputime;
10810 * Charge the system/user time to the task's accounting group.
10812 static void cpuacct_update_stats(struct task_struct *tsk,
10813 enum cpuacct_stat_index idx, cputime_t val)
10815 struct cpuacct *ca;
10817 if (unlikely(!cpuacct_subsys.active))
10824 percpu_counter_add(&ca->cpustat[idx], val);
10830 struct cgroup_subsys cpuacct_subsys = {
10832 .create = cpuacct_create,
10833 .destroy = cpuacct_destroy,
10834 .populate = cpuacct_populate,
10835 .subsys_id = cpuacct_subsys_id,
10837 #endif /* CONFIG_CGROUP_CPUACCT */