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_event.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/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy)
124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
129 static inline int task_has_rt_policy(struct task_struct *p)
131 return rt_policy(p->policy);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
142 struct rt_bandwidth {
143 /* nests inside the rq lock: */
144 spinlock_t rt_runtime_lock;
147 struct hrtimer rt_period_timer;
150 static struct rt_bandwidth def_rt_bandwidth;
152 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
154 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169 idle = do_sched_rt_period_timer(rt_b, overrun);
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
176 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
181 spin_lock_init(&rt_b->rt_runtime_lock);
183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime >= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
200 if (hrtimer_active(&rt_b->rt_period_timer))
203 spin_lock(&rt_b->rt_runtime_lock);
208 if (hrtimer_active(&rt_b->rt_period_timer))
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
218 HRTIMER_MODE_ABS_PINNED, 0);
220 spin_unlock(&rt_b->rt_runtime_lock);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
226 hrtimer_cancel(&rt_b->rt_period_timer);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex);
236 #ifdef CONFIG_GROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups);
244 /* task group related information */
246 #ifdef CONFIG_CGROUP_SCHED
247 struct cgroup_subsys_state css;
250 #ifdef CONFIG_USER_SCHED
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
266 struct rt_bandwidth rt_bandwidth;
270 struct list_head list;
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
277 #ifdef CONFIG_USER_SCHED
279 /* Helper function to pass uid information to create_sched_user() */
280 void set_tg_uid(struct user_struct *user)
282 user->tg->uid = user->uid;
287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
290 struct task_group root_task_group;
292 #ifdef CONFIG_FAIR_GROUP_SCHED
293 /* Default task group's sched entity on each cpu */
294 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
295 /* Default task group's cfs_rq on each cpu */
296 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq, init_tg_cfs_rq);
297 #endif /* CONFIG_FAIR_GROUP_SCHED */
299 #ifdef CONFIG_RT_GROUP_SCHED
300 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
301 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq, init_rt_rq);
302 #endif /* CONFIG_RT_GROUP_SCHED */
303 #else /* !CONFIG_USER_SCHED */
304 #define root_task_group init_task_group
305 #endif /* CONFIG_USER_SCHED */
307 /* task_group_lock serializes add/remove of task groups and also changes to
308 * a task group's cpu shares.
310 static DEFINE_SPINLOCK(task_group_lock);
312 #ifdef CONFIG_FAIR_GROUP_SCHED
315 static int root_task_group_empty(void)
317 return list_empty(&root_task_group.children);
321 #ifdef CONFIG_USER_SCHED
322 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
323 #else /* !CONFIG_USER_SCHED */
324 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
325 #endif /* CONFIG_USER_SCHED */
328 * A weight of 0 or 1 can cause arithmetics problems.
329 * A weight of a cfs_rq is the sum of weights of which entities
330 * are queued on this cfs_rq, so a weight of a entity should not be
331 * too large, so as the shares value of a task group.
332 * (The default weight is 1024 - so there's no practical
333 * limitation from this.)
336 #define MAX_SHARES (1UL << 18)
338 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
341 /* Default task group.
342 * Every task in system belong to this group at bootup.
344 struct task_group init_task_group;
346 /* return group to which a task belongs */
347 static inline struct task_group *task_group(struct task_struct *p)
349 struct task_group *tg;
351 #ifdef CONFIG_USER_SCHED
353 tg = __task_cred(p)->user->tg;
355 #elif defined(CONFIG_CGROUP_SCHED)
356 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
357 struct task_group, css);
359 tg = &init_task_group;
364 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
365 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
367 #ifdef CONFIG_FAIR_GROUP_SCHED
368 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
369 p->se.parent = task_group(p)->se[cpu];
372 #ifdef CONFIG_RT_GROUP_SCHED
373 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
374 p->rt.parent = task_group(p)->rt_se[cpu];
380 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
381 static inline struct task_group *task_group(struct task_struct *p)
386 #endif /* CONFIG_GROUP_SCHED */
388 /* CFS-related fields in a runqueue */
390 struct load_weight load;
391 unsigned long nr_running;
396 struct rb_root tasks_timeline;
397 struct rb_node *rb_leftmost;
399 struct list_head tasks;
400 struct list_head *balance_iterator;
403 * 'curr' points to currently running entity on this cfs_rq.
404 * It is set to NULL otherwise (i.e when none are currently running).
406 struct sched_entity *curr, *next, *last;
408 unsigned int nr_spread_over;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
421 struct list_head leaf_cfs_rq_list;
422 struct task_group *tg; /* group that "owns" this runqueue */
426 * the part of load.weight contributed by tasks
428 unsigned long task_weight;
431 * h_load = weight * f(tg)
433 * Where f(tg) is the recursive weight fraction assigned to
436 unsigned long h_load;
439 * this cpu's part of tg->shares
441 unsigned long shares;
444 * load.weight at the time we set shares
446 unsigned long rq_weight;
451 /* Real-Time classes' related field in a runqueue: */
453 struct rt_prio_array active;
454 unsigned long rt_nr_running;
455 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
457 int curr; /* highest queued rt task prio */
459 int next; /* next highest */
464 unsigned long rt_nr_migratory;
465 unsigned long rt_nr_total;
467 struct plist_head pushable_tasks;
472 /* Nests inside the rq lock: */
473 spinlock_t rt_runtime_lock;
475 #ifdef CONFIG_RT_GROUP_SCHED
476 unsigned long rt_nr_boosted;
479 struct list_head leaf_rt_rq_list;
480 struct task_group *tg;
481 struct sched_rt_entity *rt_se;
488 * We add the notion of a root-domain which will be used to define per-domain
489 * variables. Each exclusive cpuset essentially defines an island domain by
490 * fully partitioning the member cpus from any other cpuset. Whenever a new
491 * exclusive cpuset is created, we also create and attach a new root-domain
498 cpumask_var_t online;
501 * The "RT overload" flag: it gets set if a CPU has more than
502 * one runnable RT task.
504 cpumask_var_t rto_mask;
507 struct cpupri cpupri;
512 * By default the system creates a single root-domain with all cpus as
513 * members (mimicking the global state we have today).
515 static struct root_domain def_root_domain;
520 * This is the main, per-CPU runqueue data structure.
522 * Locking rule: those places that want to lock multiple runqueues
523 * (such as the load balancing or the thread migration code), lock
524 * acquire operations must be ordered by ascending &runqueue.
531 * nr_running and cpu_load should be in the same cacheline because
532 * remote CPUs use both these fields when doing load calculation.
534 unsigned long nr_running;
535 #define CPU_LOAD_IDX_MAX 5
536 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
538 unsigned long last_tick_seen;
539 unsigned char in_nohz_recently;
541 /* capture load from *all* tasks on this cpu: */
542 struct load_weight load;
543 unsigned long nr_load_updates;
549 #ifdef CONFIG_FAIR_GROUP_SCHED
550 /* list of leaf cfs_rq on this cpu: */
551 struct list_head leaf_cfs_rq_list;
553 #ifdef CONFIG_RT_GROUP_SCHED
554 struct list_head leaf_rt_rq_list;
558 * This is part of a global counter where only the total sum
559 * over all CPUs matters. A task can increase this counter on
560 * one CPU and if it got migrated afterwards it may decrease
561 * it on another CPU. Always updated under the runqueue lock:
563 unsigned long nr_uninterruptible;
565 struct task_struct *curr, *idle;
566 unsigned long next_balance;
567 struct mm_struct *prev_mm;
574 struct root_domain *rd;
575 struct sched_domain *sd;
577 unsigned char idle_at_tick;
578 /* For active balancing */
582 /* cpu of this runqueue: */
586 unsigned long avg_load_per_task;
588 struct task_struct *migration_thread;
589 struct list_head migration_queue;
597 /* calc_load related fields */
598 unsigned long calc_load_update;
599 long calc_load_active;
601 #ifdef CONFIG_SCHED_HRTICK
603 int hrtick_csd_pending;
604 struct call_single_data hrtick_csd;
606 struct hrtimer hrtick_timer;
609 #ifdef CONFIG_SCHEDSTATS
611 struct sched_info rq_sched_info;
612 unsigned long long rq_cpu_time;
613 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
615 /* sys_sched_yield() stats */
616 unsigned int yld_count;
618 /* schedule() stats */
619 unsigned int sched_switch;
620 unsigned int sched_count;
621 unsigned int sched_goidle;
623 /* try_to_wake_up() stats */
624 unsigned int ttwu_count;
625 unsigned int ttwu_local;
628 unsigned int bkl_count;
632 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
635 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
637 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
640 static inline int cpu_of(struct rq *rq)
650 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
651 * See detach_destroy_domains: synchronize_sched for details.
653 * The domain tree of any CPU may only be accessed from within
654 * preempt-disabled sections.
656 #define for_each_domain(cpu, __sd) \
657 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
659 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
660 #define this_rq() (&__get_cpu_var(runqueues))
661 #define task_rq(p) cpu_rq(task_cpu(p))
662 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
663 #define raw_rq() (&__raw_get_cpu_var(runqueues))
665 inline void update_rq_clock(struct rq *rq)
667 rq->clock = sched_clock_cpu(cpu_of(rq));
671 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
673 #ifdef CONFIG_SCHED_DEBUG
674 # define const_debug __read_mostly
676 # define const_debug static const
681 * @cpu: the processor in question.
683 * Returns true if the current cpu runqueue is locked.
684 * This interface allows printk to be called with the runqueue lock
685 * held and know whether or not it is OK to wake up the klogd.
687 int runqueue_is_locked(int cpu)
689 return spin_is_locked(&cpu_rq(cpu)->lock);
693 * Debugging: various feature bits
696 #define SCHED_FEAT(name, enabled) \
697 __SCHED_FEAT_##name ,
700 #include "sched_features.h"
705 #define SCHED_FEAT(name, enabled) \
706 (1UL << __SCHED_FEAT_##name) * enabled |
708 const_debug unsigned int sysctl_sched_features =
709 #include "sched_features.h"
714 #ifdef CONFIG_SCHED_DEBUG
715 #define SCHED_FEAT(name, enabled) \
718 static __read_mostly char *sched_feat_names[] = {
719 #include "sched_features.h"
725 static int sched_feat_show(struct seq_file *m, void *v)
729 for (i = 0; sched_feat_names[i]; i++) {
730 if (!(sysctl_sched_features & (1UL << i)))
732 seq_printf(m, "%s ", sched_feat_names[i]);
740 sched_feat_write(struct file *filp, const char __user *ubuf,
741 size_t cnt, loff_t *ppos)
751 if (copy_from_user(&buf, ubuf, cnt))
756 if (strncmp(buf, "NO_", 3) == 0) {
761 for (i = 0; sched_feat_names[i]; i++) {
762 int len = strlen(sched_feat_names[i]);
764 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
766 sysctl_sched_features &= ~(1UL << i);
768 sysctl_sched_features |= (1UL << i);
773 if (!sched_feat_names[i])
781 static int sched_feat_open(struct inode *inode, struct file *filp)
783 return single_open(filp, sched_feat_show, NULL);
786 static const struct file_operations sched_feat_fops = {
787 .open = sched_feat_open,
788 .write = sched_feat_write,
791 .release = single_release,
794 static __init int sched_init_debug(void)
796 debugfs_create_file("sched_features", 0644, NULL, NULL,
801 late_initcall(sched_init_debug);
805 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
808 * Number of tasks to iterate in a single balance run.
809 * Limited because this is done with IRQs disabled.
811 const_debug unsigned int sysctl_sched_nr_migrate = 32;
814 * ratelimit for updating the group shares.
817 unsigned int sysctl_sched_shares_ratelimit = 250000;
818 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
821 * Inject some fuzzyness into changing the per-cpu group shares
822 * this avoids remote rq-locks at the expense of fairness.
825 unsigned int sysctl_sched_shares_thresh = 4;
828 * period over which we average the RT time consumption, measured
833 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
836 * period over which we measure -rt task cpu usage in us.
839 unsigned int sysctl_sched_rt_period = 1000000;
841 static __read_mostly int scheduler_running;
844 * part of the period that we allow rt tasks to run in us.
847 int sysctl_sched_rt_runtime = 950000;
849 static inline u64 global_rt_period(void)
851 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
854 static inline u64 global_rt_runtime(void)
856 if (sysctl_sched_rt_runtime < 0)
859 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
862 #ifndef prepare_arch_switch
863 # define prepare_arch_switch(next) do { } while (0)
865 #ifndef finish_arch_switch
866 # define finish_arch_switch(prev) do { } while (0)
869 static inline int task_current(struct rq *rq, struct task_struct *p)
871 return rq->curr == p;
874 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
875 static inline int task_running(struct rq *rq, struct task_struct *p)
877 return task_current(rq, p);
880 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
884 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
886 #ifdef CONFIG_DEBUG_SPINLOCK
887 /* this is a valid case when another task releases the spinlock */
888 rq->lock.owner = current;
891 * If we are tracking spinlock dependencies then we have to
892 * fix up the runqueue lock - which gets 'carried over' from
895 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
897 spin_unlock_irq(&rq->lock);
900 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
901 static inline int task_running(struct rq *rq, struct task_struct *p)
906 return task_current(rq, p);
910 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
914 * We can optimise this out completely for !SMP, because the
915 * SMP rebalancing from interrupt is the only thing that cares
920 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
921 spin_unlock_irq(&rq->lock);
923 spin_unlock(&rq->lock);
927 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
931 * After ->oncpu is cleared, the task can be moved to a different CPU.
932 * We must ensure this doesn't happen until the switch is completely
938 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
942 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
945 * Check whether the task is waking, we use this to synchronize against
946 * ttwu() so that task_cpu() reports a stable number.
948 * We need to make an exception for PF_STARTING tasks because the fork
949 * path might require task_rq_lock() to work, eg. it can call
950 * set_cpus_allowed_ptr() from the cpuset clone_ns code.
952 static inline int task_is_waking(struct task_struct *p)
954 return unlikely((p->state == TASK_WAKING) && !(p->flags & PF_STARTING));
958 * __task_rq_lock - lock the runqueue a given task resides on.
959 * Must be called interrupts disabled.
961 static inline struct rq *__task_rq_lock(struct task_struct *p)
967 while (task_is_waking(p))
970 spin_lock(&rq->lock);
971 if (likely(rq == task_rq(p) && !task_is_waking(p)))
973 spin_unlock(&rq->lock);
978 * task_rq_lock - lock the runqueue a given task resides on and disable
979 * interrupts. Note the ordering: we can safely lookup the task_rq without
980 * explicitly disabling preemption.
982 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
988 while (task_is_waking(p))
990 local_irq_save(*flags);
992 spin_lock(&rq->lock);
993 if (likely(rq == task_rq(p) && !task_is_waking(p)))
995 spin_unlock_irqrestore(&rq->lock, *flags);
999 void task_rq_unlock_wait(struct task_struct *p)
1001 struct rq *rq = task_rq(p);
1003 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1004 spin_unlock_wait(&rq->lock);
1007 static void __task_rq_unlock(struct rq *rq)
1008 __releases(rq->lock)
1010 spin_unlock(&rq->lock);
1013 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1014 __releases(rq->lock)
1016 spin_unlock_irqrestore(&rq->lock, *flags);
1020 * this_rq_lock - lock this runqueue and disable interrupts.
1022 static struct rq *this_rq_lock(void)
1023 __acquires(rq->lock)
1027 local_irq_disable();
1029 spin_lock(&rq->lock);
1034 #ifdef CONFIG_SCHED_HRTICK
1036 * Use HR-timers to deliver accurate preemption points.
1038 * Its all a bit involved since we cannot program an hrt while holding the
1039 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1042 * When we get rescheduled we reprogram the hrtick_timer outside of the
1048 * - enabled by features
1049 * - hrtimer is actually high res
1051 static inline int hrtick_enabled(struct rq *rq)
1053 if (!sched_feat(HRTICK))
1055 if (!cpu_active(cpu_of(rq)))
1057 return hrtimer_is_hres_active(&rq->hrtick_timer);
1060 static void hrtick_clear(struct rq *rq)
1062 if (hrtimer_active(&rq->hrtick_timer))
1063 hrtimer_cancel(&rq->hrtick_timer);
1067 * High-resolution timer tick.
1068 * Runs from hardirq context with interrupts disabled.
1070 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1072 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1074 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1076 spin_lock(&rq->lock);
1077 update_rq_clock(rq);
1078 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1079 spin_unlock(&rq->lock);
1081 return HRTIMER_NORESTART;
1086 * called from hardirq (IPI) context
1088 static void __hrtick_start(void *arg)
1090 struct rq *rq = arg;
1092 spin_lock(&rq->lock);
1093 hrtimer_restart(&rq->hrtick_timer);
1094 rq->hrtick_csd_pending = 0;
1095 spin_unlock(&rq->lock);
1099 * Called to set the hrtick timer state.
1101 * called with rq->lock held and irqs disabled
1103 static void hrtick_start(struct rq *rq, u64 delay)
1105 struct hrtimer *timer = &rq->hrtick_timer;
1106 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1108 hrtimer_set_expires(timer, time);
1110 if (rq == this_rq()) {
1111 hrtimer_restart(timer);
1112 } else if (!rq->hrtick_csd_pending) {
1113 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1114 rq->hrtick_csd_pending = 1;
1119 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1121 int cpu = (int)(long)hcpu;
1124 case CPU_UP_CANCELED:
1125 case CPU_UP_CANCELED_FROZEN:
1126 case CPU_DOWN_PREPARE:
1127 case CPU_DOWN_PREPARE_FROZEN:
1129 case CPU_DEAD_FROZEN:
1130 hrtick_clear(cpu_rq(cpu));
1137 static __init void init_hrtick(void)
1139 hotcpu_notifier(hotplug_hrtick, 0);
1143 * Called to set the hrtick timer state.
1145 * called with rq->lock held and irqs disabled
1147 static void hrtick_start(struct rq *rq, u64 delay)
1149 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1150 HRTIMER_MODE_REL_PINNED, 0);
1153 static inline void init_hrtick(void)
1156 #endif /* CONFIG_SMP */
1158 static void init_rq_hrtick(struct rq *rq)
1161 rq->hrtick_csd_pending = 0;
1163 rq->hrtick_csd.flags = 0;
1164 rq->hrtick_csd.func = __hrtick_start;
1165 rq->hrtick_csd.info = rq;
1168 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1169 rq->hrtick_timer.function = hrtick;
1171 #else /* CONFIG_SCHED_HRTICK */
1172 static inline void hrtick_clear(struct rq *rq)
1176 static inline void init_rq_hrtick(struct rq *rq)
1180 static inline void init_hrtick(void)
1183 #endif /* CONFIG_SCHED_HRTICK */
1186 * resched_task - mark a task 'to be rescheduled now'.
1188 * On UP this means the setting of the need_resched flag, on SMP it
1189 * might also involve a cross-CPU call to trigger the scheduler on
1194 #ifndef tsk_is_polling
1195 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1198 static void resched_task(struct task_struct *p)
1202 assert_spin_locked(&task_rq(p)->lock);
1204 if (test_tsk_need_resched(p))
1207 set_tsk_need_resched(p);
1210 if (cpu == smp_processor_id())
1213 /* NEED_RESCHED must be visible before we test polling */
1215 if (!tsk_is_polling(p))
1216 smp_send_reschedule(cpu);
1219 static void resched_cpu(int cpu)
1221 struct rq *rq = cpu_rq(cpu);
1222 unsigned long flags;
1224 if (!spin_trylock_irqsave(&rq->lock, flags))
1226 resched_task(cpu_curr(cpu));
1227 spin_unlock_irqrestore(&rq->lock, flags);
1232 * When add_timer_on() enqueues a timer into the timer wheel of an
1233 * idle CPU then this timer might expire before the next timer event
1234 * which is scheduled to wake up that CPU. In case of a completely
1235 * idle system the next event might even be infinite time into the
1236 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1237 * leaves the inner idle loop so the newly added timer is taken into
1238 * account when the CPU goes back to idle and evaluates the timer
1239 * wheel for the next timer event.
1241 void wake_up_idle_cpu(int cpu)
1243 struct rq *rq = cpu_rq(cpu);
1245 if (cpu == smp_processor_id())
1249 * This is safe, as this function is called with the timer
1250 * wheel base lock of (cpu) held. When the CPU is on the way
1251 * to idle and has not yet set rq->curr to idle then it will
1252 * be serialized on the timer wheel base lock and take the new
1253 * timer into account automatically.
1255 if (rq->curr != rq->idle)
1259 * We can set TIF_RESCHED on the idle task of the other CPU
1260 * lockless. The worst case is that the other CPU runs the
1261 * idle task through an additional NOOP schedule()
1263 set_tsk_need_resched(rq->idle);
1265 /* NEED_RESCHED must be visible before we test polling */
1267 if (!tsk_is_polling(rq->idle))
1268 smp_send_reschedule(cpu);
1270 #endif /* CONFIG_NO_HZ */
1272 static u64 sched_avg_period(void)
1274 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1277 static void sched_avg_update(struct rq *rq)
1279 s64 period = sched_avg_period();
1281 while ((s64)(rq->clock - rq->age_stamp) > period) {
1283 * Inline assembly required to prevent the compiler
1284 * optimising this loop into a divmod call.
1285 * See __iter_div_u64_rem() for another example of this.
1287 asm("" : "+rm" (rq->age_stamp));
1288 rq->age_stamp += period;
1293 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1295 rq->rt_avg += rt_delta;
1296 sched_avg_update(rq);
1299 #else /* !CONFIG_SMP */
1300 static void resched_task(struct task_struct *p)
1302 assert_spin_locked(&task_rq(p)->lock);
1303 set_tsk_need_resched(p);
1306 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1309 #endif /* CONFIG_SMP */
1311 #if BITS_PER_LONG == 32
1312 # define WMULT_CONST (~0UL)
1314 # define WMULT_CONST (1UL << 32)
1317 #define WMULT_SHIFT 32
1320 * Shift right and round:
1322 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1325 * delta *= weight / lw
1327 static unsigned long
1328 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1329 struct load_weight *lw)
1333 if (!lw->inv_weight) {
1334 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1337 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1341 tmp = (u64)delta_exec * weight;
1343 * Check whether we'd overflow the 64-bit multiplication:
1345 if (unlikely(tmp > WMULT_CONST))
1346 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1349 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1351 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1354 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1360 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1367 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1368 * of tasks with abnormal "nice" values across CPUs the contribution that
1369 * each task makes to its run queue's load is weighted according to its
1370 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1371 * scaled version of the new time slice allocation that they receive on time
1375 #define WEIGHT_IDLEPRIO 3
1376 #define WMULT_IDLEPRIO 1431655765
1379 * Nice levels are multiplicative, with a gentle 10% change for every
1380 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1381 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1382 * that remained on nice 0.
1384 * The "10% effect" is relative and cumulative: from _any_ nice level,
1385 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1386 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1387 * If a task goes up by ~10% and another task goes down by ~10% then
1388 * the relative distance between them is ~25%.)
1390 static const int prio_to_weight[40] = {
1391 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1392 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1393 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1394 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1395 /* 0 */ 1024, 820, 655, 526, 423,
1396 /* 5 */ 335, 272, 215, 172, 137,
1397 /* 10 */ 110, 87, 70, 56, 45,
1398 /* 15 */ 36, 29, 23, 18, 15,
1402 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1404 * In cases where the weight does not change often, we can use the
1405 * precalculated inverse to speed up arithmetics by turning divisions
1406 * into multiplications:
1408 static const u32 prio_to_wmult[40] = {
1409 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1410 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1411 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1412 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1413 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1414 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1415 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1416 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1419 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1422 * runqueue iterator, to support SMP load-balancing between different
1423 * scheduling classes, without having to expose their internal data
1424 * structures to the load-balancing proper:
1426 struct rq_iterator {
1428 struct task_struct *(*start)(void *);
1429 struct task_struct *(*next)(void *);
1433 static unsigned long
1434 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1435 unsigned long max_load_move, struct sched_domain *sd,
1436 enum cpu_idle_type idle, int *all_pinned,
1437 int *this_best_prio, struct rq_iterator *iterator);
1440 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1441 struct sched_domain *sd, enum cpu_idle_type idle,
1442 struct rq_iterator *iterator);
1445 /* Time spent by the tasks of the cpu accounting group executing in ... */
1446 enum cpuacct_stat_index {
1447 CPUACCT_STAT_USER, /* ... user mode */
1448 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1450 CPUACCT_STAT_NSTATS,
1453 #ifdef CONFIG_CGROUP_CPUACCT
1454 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1455 static void cpuacct_update_stats(struct task_struct *tsk,
1456 enum cpuacct_stat_index idx, cputime_t val);
1458 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1459 static inline void cpuacct_update_stats(struct task_struct *tsk,
1460 enum cpuacct_stat_index idx, cputime_t val) {}
1463 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1465 update_load_add(&rq->load, load);
1468 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1470 update_load_sub(&rq->load, load);
1473 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1474 typedef int (*tg_visitor)(struct task_group *, void *);
1477 * Iterate the full tree, calling @down when first entering a node and @up when
1478 * leaving it for the final time.
1480 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1482 struct task_group *parent, *child;
1486 parent = &root_task_group;
1488 ret = (*down)(parent, data);
1491 list_for_each_entry_rcu(child, &parent->children, siblings) {
1498 ret = (*up)(parent, data);
1503 parent = parent->parent;
1512 static int tg_nop(struct task_group *tg, void *data)
1519 /* Used instead of source_load when we know the type == 0 */
1520 static unsigned long weighted_cpuload(const int cpu)
1522 return cpu_rq(cpu)->load.weight;
1526 * Return a low guess at the load of a migration-source cpu weighted
1527 * according to the scheduling class and "nice" value.
1529 * We want to under-estimate the load of migration sources, to
1530 * balance conservatively.
1532 static unsigned long source_load(int cpu, int type)
1534 struct rq *rq = cpu_rq(cpu);
1535 unsigned long total = weighted_cpuload(cpu);
1537 if (type == 0 || !sched_feat(LB_BIAS))
1540 return min(rq->cpu_load[type-1], total);
1544 * Return a high guess at the load of a migration-target cpu weighted
1545 * according to the scheduling class and "nice" value.
1547 static unsigned long target_load(int cpu, int type)
1549 struct rq *rq = cpu_rq(cpu);
1550 unsigned long total = weighted_cpuload(cpu);
1552 if (type == 0 || !sched_feat(LB_BIAS))
1555 return max(rq->cpu_load[type-1], total);
1558 static struct sched_group *group_of(int cpu)
1560 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1568 static unsigned long power_of(int cpu)
1570 struct sched_group *group = group_of(cpu);
1573 return SCHED_LOAD_SCALE;
1575 return group->cpu_power;
1578 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1580 static unsigned long cpu_avg_load_per_task(int cpu)
1582 struct rq *rq = cpu_rq(cpu);
1583 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1586 rq->avg_load_per_task = rq->load.weight / nr_running;
1588 rq->avg_load_per_task = 0;
1590 return rq->avg_load_per_task;
1593 #ifdef CONFIG_FAIR_GROUP_SCHED
1595 static __read_mostly unsigned long *update_shares_data;
1597 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1600 * Calculate and set the cpu's group shares.
1602 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1603 unsigned long sd_shares,
1604 unsigned long sd_rq_weight,
1605 unsigned long *usd_rq_weight)
1607 unsigned long shares, rq_weight;
1610 rq_weight = usd_rq_weight[cpu];
1613 rq_weight = NICE_0_LOAD;
1617 * \Sum_j shares_j * rq_weight_i
1618 * shares_i = -----------------------------
1619 * \Sum_j rq_weight_j
1621 shares = (sd_shares * rq_weight) / sd_rq_weight;
1622 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1624 if (abs(shares - tg->se[cpu]->load.weight) >
1625 sysctl_sched_shares_thresh) {
1626 struct rq *rq = cpu_rq(cpu);
1627 unsigned long flags;
1629 spin_lock_irqsave(&rq->lock, flags);
1630 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1631 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1632 __set_se_shares(tg->se[cpu], shares);
1633 spin_unlock_irqrestore(&rq->lock, flags);
1638 * Re-compute the task group their per cpu shares over the given domain.
1639 * This needs to be done in a bottom-up fashion because the rq weight of a
1640 * parent group depends on the shares of its child groups.
1642 static int tg_shares_up(struct task_group *tg, void *data)
1644 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1645 unsigned long *usd_rq_weight;
1646 struct sched_domain *sd = data;
1647 unsigned long flags;
1653 local_irq_save(flags);
1654 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1656 for_each_cpu(i, sched_domain_span(sd)) {
1657 weight = tg->cfs_rq[i]->load.weight;
1658 usd_rq_weight[i] = weight;
1660 rq_weight += weight;
1662 * If there are currently no tasks on the cpu pretend there
1663 * is one of average load so that when a new task gets to
1664 * run here it will not get delayed by group starvation.
1667 weight = NICE_0_LOAD;
1669 sum_weight += weight;
1670 shares += tg->cfs_rq[i]->shares;
1674 rq_weight = sum_weight;
1676 if ((!shares && rq_weight) || shares > tg->shares)
1677 shares = tg->shares;
1679 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1680 shares = tg->shares;
1682 for_each_cpu(i, sched_domain_span(sd))
1683 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1685 local_irq_restore(flags);
1691 * Compute the cpu's hierarchical load factor for each task group.
1692 * This needs to be done in a top-down fashion because the load of a child
1693 * group is a fraction of its parents load.
1695 static int tg_load_down(struct task_group *tg, void *data)
1698 long cpu = (long)data;
1701 load = cpu_rq(cpu)->load.weight;
1703 load = tg->parent->cfs_rq[cpu]->h_load;
1704 load *= tg->cfs_rq[cpu]->shares;
1705 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1708 tg->cfs_rq[cpu]->h_load = load;
1713 static void update_shares(struct sched_domain *sd)
1718 if (root_task_group_empty())
1721 now = cpu_clock(raw_smp_processor_id());
1722 elapsed = now - sd->last_update;
1724 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1725 sd->last_update = now;
1726 walk_tg_tree(tg_nop, tg_shares_up, sd);
1730 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1732 if (root_task_group_empty())
1735 spin_unlock(&rq->lock);
1737 spin_lock(&rq->lock);
1740 static void update_h_load(long cpu)
1742 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1747 static inline void update_shares(struct sched_domain *sd)
1751 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1757 #ifdef CONFIG_PREEMPT
1759 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1762 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1763 * way at the expense of forcing extra atomic operations in all
1764 * invocations. This assures that the double_lock is acquired using the
1765 * same underlying policy as the spinlock_t on this architecture, which
1766 * reduces latency compared to the unfair variant below. However, it
1767 * also adds more overhead and therefore may reduce throughput.
1769 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1770 __releases(this_rq->lock)
1771 __acquires(busiest->lock)
1772 __acquires(this_rq->lock)
1774 spin_unlock(&this_rq->lock);
1775 double_rq_lock(this_rq, busiest);
1782 * Unfair double_lock_balance: Optimizes throughput at the expense of
1783 * latency by eliminating extra atomic operations when the locks are
1784 * already in proper order on entry. This favors lower cpu-ids and will
1785 * grant the double lock to lower cpus over higher ids under contention,
1786 * regardless of entry order into the function.
1788 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1789 __releases(this_rq->lock)
1790 __acquires(busiest->lock)
1791 __acquires(this_rq->lock)
1795 if (unlikely(!spin_trylock(&busiest->lock))) {
1796 if (busiest < this_rq) {
1797 spin_unlock(&this_rq->lock);
1798 spin_lock(&busiest->lock);
1799 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1802 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1807 #endif /* CONFIG_PREEMPT */
1810 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1812 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1814 if (unlikely(!irqs_disabled())) {
1815 /* printk() doesn't work good under rq->lock */
1816 spin_unlock(&this_rq->lock);
1820 return _double_lock_balance(this_rq, busiest);
1823 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1824 __releases(busiest->lock)
1826 spin_unlock(&busiest->lock);
1827 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1831 #ifdef CONFIG_FAIR_GROUP_SCHED
1832 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1835 cfs_rq->shares = shares;
1840 static void calc_load_account_active(struct rq *this_rq);
1841 static void update_sysctl(void);
1843 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1845 set_task_rq(p, cpu);
1848 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1849 * successfuly executed on another CPU. We must ensure that updates of
1850 * per-task data have been completed by this moment.
1853 task_thread_info(p)->cpu = cpu;
1857 #include "sched_stats.h"
1858 #include "sched_idletask.c"
1859 #include "sched_fair.c"
1860 #include "sched_rt.c"
1861 #ifdef CONFIG_SCHED_DEBUG
1862 # include "sched_debug.c"
1865 #define sched_class_highest (&rt_sched_class)
1866 #define for_each_class(class) \
1867 for (class = sched_class_highest; class; class = class->next)
1869 static void inc_nr_running(struct rq *rq)
1874 static void dec_nr_running(struct rq *rq)
1879 static void set_load_weight(struct task_struct *p)
1881 if (task_has_rt_policy(p)) {
1882 p->se.load.weight = prio_to_weight[0] * 2;
1883 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1888 * SCHED_IDLE tasks get minimal weight:
1890 if (p->policy == SCHED_IDLE) {
1891 p->se.load.weight = WEIGHT_IDLEPRIO;
1892 p->se.load.inv_weight = WMULT_IDLEPRIO;
1896 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1897 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1900 static void update_avg(u64 *avg, u64 sample)
1902 s64 diff = sample - *avg;
1907 enqueue_task(struct rq *rq, struct task_struct *p, int wakeup, bool head)
1910 p->se.start_runtime = p->se.sum_exec_runtime;
1912 sched_info_queued(p);
1913 p->sched_class->enqueue_task(rq, p, wakeup, head);
1917 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1920 if (p->se.last_wakeup) {
1921 update_avg(&p->se.avg_overlap,
1922 p->se.sum_exec_runtime - p->se.last_wakeup);
1923 p->se.last_wakeup = 0;
1925 update_avg(&p->se.avg_wakeup,
1926 sysctl_sched_wakeup_granularity);
1930 sched_info_dequeued(p);
1931 p->sched_class->dequeue_task(rq, p, sleep);
1936 * __normal_prio - return the priority that is based on the static prio
1938 static inline int __normal_prio(struct task_struct *p)
1940 return p->static_prio;
1944 * Calculate the expected normal priority: i.e. priority
1945 * without taking RT-inheritance into account. Might be
1946 * boosted by interactivity modifiers. Changes upon fork,
1947 * setprio syscalls, and whenever the interactivity
1948 * estimator recalculates.
1950 static inline int normal_prio(struct task_struct *p)
1954 if (task_has_rt_policy(p))
1955 prio = MAX_RT_PRIO-1 - p->rt_priority;
1957 prio = __normal_prio(p);
1962 * Calculate the current priority, i.e. the priority
1963 * taken into account by the scheduler. This value might
1964 * be boosted by RT tasks, or might be boosted by
1965 * interactivity modifiers. Will be RT if the task got
1966 * RT-boosted. If not then it returns p->normal_prio.
1968 static int effective_prio(struct task_struct *p)
1970 p->normal_prio = normal_prio(p);
1972 * If we are RT tasks or we were boosted to RT priority,
1973 * keep the priority unchanged. Otherwise, update priority
1974 * to the normal priority:
1976 if (!rt_prio(p->prio))
1977 return p->normal_prio;
1982 * activate_task - move a task to the runqueue.
1984 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1986 if (task_contributes_to_load(p))
1987 rq->nr_uninterruptible--;
1989 enqueue_task(rq, p, wakeup, false);
1994 * deactivate_task - remove a task from the runqueue.
1996 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1998 if (task_contributes_to_load(p))
1999 rq->nr_uninterruptible++;
2001 dequeue_task(rq, p, sleep);
2006 * task_curr - is this task currently executing on a CPU?
2007 * @p: the task in question.
2009 inline int task_curr(const struct task_struct *p)
2011 return cpu_curr(task_cpu(p)) == p;
2014 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2015 const struct sched_class *prev_class,
2016 int oldprio, int running)
2018 if (prev_class != p->sched_class) {
2019 if (prev_class->switched_from)
2020 prev_class->switched_from(rq, p, running);
2021 p->sched_class->switched_to(rq, p, running);
2023 p->sched_class->prio_changed(rq, p, oldprio, running);
2027 * kthread_bind - bind a just-created kthread to a cpu.
2028 * @p: thread created by kthread_create().
2029 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2031 * Description: This function is equivalent to set_cpus_allowed(),
2032 * except that @cpu doesn't need to be online, and the thread must be
2033 * stopped (i.e., just returned from kthread_create()).
2035 * Function lives here instead of kthread.c because it messes with
2036 * scheduler internals which require locking.
2038 void kthread_bind(struct task_struct *p, unsigned int cpu)
2040 /* Must have done schedule() in kthread() before we set_task_cpu */
2041 if (!wait_task_inactive(p, TASK_UNINTERRUPTIBLE)) {
2046 p->cpus_allowed = cpumask_of_cpu(cpu);
2047 p->rt.nr_cpus_allowed = 1;
2048 p->flags |= PF_THREAD_BOUND;
2050 EXPORT_SYMBOL(kthread_bind);
2054 * Is this task likely cache-hot:
2057 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2061 if (p->sched_class != &fair_sched_class)
2065 * Buddy candidates are cache hot:
2067 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2068 (&p->se == cfs_rq_of(&p->se)->next ||
2069 &p->se == cfs_rq_of(&p->se)->last))
2072 if (sysctl_sched_migration_cost == -1)
2074 if (sysctl_sched_migration_cost == 0)
2077 delta = now - p->se.exec_start;
2079 return delta < (s64)sysctl_sched_migration_cost;
2083 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2085 int old_cpu = task_cpu(p);
2087 #ifdef CONFIG_SCHED_DEBUG
2089 * We should never call set_task_cpu() on a blocked task,
2090 * ttwu() will sort out the placement.
2092 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2093 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2096 trace_sched_migrate_task(p, new_cpu);
2098 if (old_cpu != new_cpu) {
2099 p->se.nr_migrations++;
2100 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2104 __set_task_cpu(p, new_cpu);
2107 struct migration_req {
2108 struct list_head list;
2110 struct task_struct *task;
2113 struct completion done;
2117 * The task's runqueue lock must be held.
2118 * Returns true if you have to wait for migration thread.
2121 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2123 struct rq *rq = task_rq(p);
2126 * If the task is not on a runqueue (and not running), then
2127 * the next wake-up will properly place the task.
2129 if (!p->se.on_rq && !task_running(rq, p))
2132 init_completion(&req->done);
2134 req->dest_cpu = dest_cpu;
2135 list_add(&req->list, &rq->migration_queue);
2141 * wait_task_context_switch - wait for a thread to complete at least one
2144 * @p must not be current.
2146 void wait_task_context_switch(struct task_struct *p)
2148 unsigned long nvcsw, nivcsw, flags;
2156 * The runqueue is assigned before the actual context
2157 * switch. We need to take the runqueue lock.
2159 * We could check initially without the lock but it is
2160 * very likely that we need to take the lock in every
2163 rq = task_rq_lock(p, &flags);
2164 running = task_running(rq, p);
2165 task_rq_unlock(rq, &flags);
2167 if (likely(!running))
2170 * The switch count is incremented before the actual
2171 * context switch. We thus wait for two switches to be
2172 * sure at least one completed.
2174 if ((p->nvcsw - nvcsw) > 1)
2176 if ((p->nivcsw - nivcsw) > 1)
2184 * wait_task_inactive - wait for a thread to unschedule.
2186 * If @match_state is nonzero, it's the @p->state value just checked and
2187 * not expected to change. If it changes, i.e. @p might have woken up,
2188 * then return zero. When we succeed in waiting for @p to be off its CPU,
2189 * we return a positive number (its total switch count). If a second call
2190 * a short while later returns the same number, the caller can be sure that
2191 * @p has remained unscheduled the whole time.
2193 * The caller must ensure that the task *will* unschedule sometime soon,
2194 * else this function might spin for a *long* time. This function can't
2195 * be called with interrupts off, or it may introduce deadlock with
2196 * smp_call_function() if an IPI is sent by the same process we are
2197 * waiting to become inactive.
2199 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2201 unsigned long flags;
2208 * We do the initial early heuristics without holding
2209 * any task-queue locks at all. We'll only try to get
2210 * the runqueue lock when things look like they will
2216 * If the task is actively running on another CPU
2217 * still, just relax and busy-wait without holding
2220 * NOTE! Since we don't hold any locks, it's not
2221 * even sure that "rq" stays as the right runqueue!
2222 * But we don't care, since "task_running()" will
2223 * return false if the runqueue has changed and p
2224 * is actually now running somewhere else!
2226 while (task_running(rq, p)) {
2227 if (match_state && unlikely(p->state != match_state))
2233 * Ok, time to look more closely! We need the rq
2234 * lock now, to be *sure*. If we're wrong, we'll
2235 * just go back and repeat.
2237 rq = task_rq_lock(p, &flags);
2238 trace_sched_wait_task(rq, p);
2239 running = task_running(rq, p);
2240 on_rq = p->se.on_rq;
2242 if (!match_state || p->state == match_state)
2243 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2244 task_rq_unlock(rq, &flags);
2247 * If it changed from the expected state, bail out now.
2249 if (unlikely(!ncsw))
2253 * Was it really running after all now that we
2254 * checked with the proper locks actually held?
2256 * Oops. Go back and try again..
2258 if (unlikely(running)) {
2264 * It's not enough that it's not actively running,
2265 * it must be off the runqueue _entirely_, and not
2268 * So if it was still runnable (but just not actively
2269 * running right now), it's preempted, and we should
2270 * yield - it could be a while.
2272 if (unlikely(on_rq)) {
2273 schedule_timeout_uninterruptible(1);
2278 * Ahh, all good. It wasn't running, and it wasn't
2279 * runnable, which means that it will never become
2280 * running in the future either. We're all done!
2289 * kick_process - kick a running thread to enter/exit the kernel
2290 * @p: the to-be-kicked thread
2292 * Cause a process which is running on another CPU to enter
2293 * kernel-mode, without any delay. (to get signals handled.)
2295 * NOTE: this function doesnt have to take the runqueue lock,
2296 * because all it wants to ensure is that the remote task enters
2297 * the kernel. If the IPI races and the task has been migrated
2298 * to another CPU then no harm is done and the purpose has been
2301 void kick_process(struct task_struct *p)
2307 if ((cpu != smp_processor_id()) && task_curr(p))
2308 smp_send_reschedule(cpu);
2311 EXPORT_SYMBOL_GPL(kick_process);
2312 #endif /* CONFIG_SMP */
2315 * task_oncpu_function_call - call a function on the cpu on which a task runs
2316 * @p: the task to evaluate
2317 * @func: the function to be called
2318 * @info: the function call argument
2320 * Calls the function @func when the task is currently running. This might
2321 * be on the current CPU, which just calls the function directly
2323 void task_oncpu_function_call(struct task_struct *p,
2324 void (*func) (void *info), void *info)
2331 smp_call_function_single(cpu, func, info, 1);
2337 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2339 static int select_fallback_rq(int cpu, struct task_struct *p)
2342 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2344 /* Look for allowed, online CPU in same node. */
2345 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2346 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2349 /* Any allowed, online CPU? */
2350 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2351 if (dest_cpu < nr_cpu_ids)
2354 /* No more Mr. Nice Guy. */
2355 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2356 dest_cpu = cpuset_cpus_allowed_fallback(p);
2358 * Don't tell them about moving exiting tasks or
2359 * kernel threads (both mm NULL), since they never
2362 if (p->mm && printk_ratelimit()) {
2363 printk(KERN_INFO "process %d (%s) no "
2364 "longer affine to cpu%d\n",
2365 task_pid_nr(p), p->comm, cpu);
2373 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2376 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2378 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2381 * In order not to call set_task_cpu() on a blocking task we need
2382 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2385 * Since this is common to all placement strategies, this lives here.
2387 * [ this allows ->select_task() to simply return task_cpu(p) and
2388 * not worry about this generic constraint ]
2390 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2392 cpu = select_fallback_rq(task_cpu(p), p);
2399 * try_to_wake_up - wake up a thread
2400 * @p: the to-be-woken-up thread
2401 * @state: the mask of task states that can be woken
2402 * @sync: do a synchronous wakeup?
2404 * Put it on the run-queue if it's not already there. The "current"
2405 * thread is always on the run-queue (except when the actual
2406 * re-schedule is in progress), and as such you're allowed to do
2407 * the simpler "current->state = TASK_RUNNING" to mark yourself
2408 * runnable without the overhead of this.
2410 * returns failure only if the task is already active.
2412 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2415 int cpu, orig_cpu, this_cpu, success = 0;
2416 unsigned long flags;
2417 struct rq *rq, *orig_rq;
2419 if (!sched_feat(SYNC_WAKEUPS))
2420 wake_flags &= ~WF_SYNC;
2422 this_cpu = get_cpu();
2425 rq = orig_rq = task_rq_lock(p, &flags);
2426 update_rq_clock(rq);
2427 if (!(p->state & state))
2437 if (unlikely(task_running(rq, p)))
2441 * In order to handle concurrent wakeups and release the rq->lock
2442 * we put the task in TASK_WAKING state.
2444 * First fix up the nr_uninterruptible count:
2446 if (task_contributes_to_load(p))
2447 rq->nr_uninterruptible--;
2448 p->state = TASK_WAKING;
2450 if (p->sched_class->task_waking)
2451 p->sched_class->task_waking(rq, p);
2453 __task_rq_unlock(rq);
2455 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2456 if (cpu != orig_cpu) {
2458 * Since we migrate the task without holding any rq->lock,
2459 * we need to be careful with task_rq_lock(), since that
2460 * might end up locking an invalid rq.
2462 set_task_cpu(p, cpu);
2466 spin_lock(&rq->lock);
2467 update_rq_clock(rq);
2470 * We migrated the task without holding either rq->lock, however
2471 * since the task is not on the task list itself, nobody else
2472 * will try and migrate the task, hence the rq should match the
2473 * cpu we just moved it to.
2475 WARN_ON(task_cpu(p) != cpu);
2476 WARN_ON(p->state != TASK_WAKING);
2478 #ifdef CONFIG_SCHEDSTATS
2479 schedstat_inc(rq, ttwu_count);
2480 if (cpu == this_cpu)
2481 schedstat_inc(rq, ttwu_local);
2483 struct sched_domain *sd;
2484 for_each_domain(this_cpu, sd) {
2485 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2486 schedstat_inc(sd, ttwu_wake_remote);
2491 #endif /* CONFIG_SCHEDSTATS */
2494 #endif /* CONFIG_SMP */
2495 schedstat_inc(p, se.nr_wakeups);
2496 if (wake_flags & WF_SYNC)
2497 schedstat_inc(p, se.nr_wakeups_sync);
2498 if (orig_cpu != cpu)
2499 schedstat_inc(p, se.nr_wakeups_migrate);
2500 if (cpu == this_cpu)
2501 schedstat_inc(p, se.nr_wakeups_local);
2503 schedstat_inc(p, se.nr_wakeups_remote);
2504 activate_task(rq, p, 1);
2508 * Only attribute actual wakeups done by this task.
2510 if (!in_interrupt()) {
2511 struct sched_entity *se = ¤t->se;
2512 u64 sample = se->sum_exec_runtime;
2514 if (se->last_wakeup)
2515 sample -= se->last_wakeup;
2517 sample -= se->start_runtime;
2518 update_avg(&se->avg_wakeup, sample);
2520 se->last_wakeup = se->sum_exec_runtime;
2524 trace_sched_wakeup(rq, p, success);
2525 check_preempt_curr(rq, p, wake_flags);
2527 p->state = TASK_RUNNING;
2529 if (p->sched_class->task_woken)
2530 p->sched_class->task_woken(rq, p);
2532 if (unlikely(rq->idle_stamp)) {
2533 u64 delta = rq->clock - rq->idle_stamp;
2534 u64 max = 2*sysctl_sched_migration_cost;
2539 update_avg(&rq->avg_idle, delta);
2544 task_rq_unlock(rq, &flags);
2551 * wake_up_process - Wake up a specific process
2552 * @p: The process to be woken up.
2554 * Attempt to wake up the nominated process and move it to the set of runnable
2555 * processes. Returns 1 if the process was woken up, 0 if it was already
2558 * It may be assumed that this function implies a write memory barrier before
2559 * changing the task state if and only if any tasks are woken up.
2561 int wake_up_process(struct task_struct *p)
2563 return try_to_wake_up(p, TASK_ALL, 0);
2565 EXPORT_SYMBOL(wake_up_process);
2567 int wake_up_state(struct task_struct *p, unsigned int state)
2569 return try_to_wake_up(p, state, 0);
2573 * Perform scheduler related setup for a newly forked process p.
2574 * p is forked by current.
2576 * __sched_fork() is basic setup used by init_idle() too:
2578 static void __sched_fork(struct task_struct *p)
2580 p->se.exec_start = 0;
2581 p->se.sum_exec_runtime = 0;
2582 p->se.prev_sum_exec_runtime = 0;
2583 p->se.nr_migrations = 0;
2584 p->se.last_wakeup = 0;
2585 p->se.avg_overlap = 0;
2586 p->se.start_runtime = 0;
2587 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2588 p->se.avg_running = 0;
2590 #ifdef CONFIG_SCHEDSTATS
2591 p->se.wait_start = 0;
2593 p->se.wait_count = 0;
2596 p->se.sleep_start = 0;
2597 p->se.sleep_max = 0;
2598 p->se.sum_sleep_runtime = 0;
2600 p->se.block_start = 0;
2601 p->se.block_max = 0;
2603 p->se.slice_max = 0;
2605 p->se.nr_migrations_cold = 0;
2606 p->se.nr_failed_migrations_affine = 0;
2607 p->se.nr_failed_migrations_running = 0;
2608 p->se.nr_failed_migrations_hot = 0;
2609 p->se.nr_forced_migrations = 0;
2611 p->se.nr_wakeups = 0;
2612 p->se.nr_wakeups_sync = 0;
2613 p->se.nr_wakeups_migrate = 0;
2614 p->se.nr_wakeups_local = 0;
2615 p->se.nr_wakeups_remote = 0;
2616 p->se.nr_wakeups_affine = 0;
2617 p->se.nr_wakeups_affine_attempts = 0;
2618 p->se.nr_wakeups_passive = 0;
2619 p->se.nr_wakeups_idle = 0;
2623 INIT_LIST_HEAD(&p->rt.run_list);
2625 INIT_LIST_HEAD(&p->se.group_node);
2627 #ifdef CONFIG_PREEMPT_NOTIFIERS
2628 INIT_HLIST_HEAD(&p->preempt_notifiers);
2633 * fork()/clone()-time setup:
2635 void sched_fork(struct task_struct *p, int clone_flags)
2637 int cpu = get_cpu();
2641 * We mark the process as waking here. This guarantees that
2642 * nobody will actually run it, and a signal or other external
2643 * event cannot wake it up and insert it on the runqueue either.
2645 p->state = TASK_WAKING;
2648 * Revert to default priority/policy on fork if requested.
2650 if (unlikely(p->sched_reset_on_fork)) {
2651 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2652 p->policy = SCHED_NORMAL;
2653 p->normal_prio = p->static_prio;
2656 if (PRIO_TO_NICE(p->static_prio) < 0) {
2657 p->static_prio = NICE_TO_PRIO(0);
2658 p->normal_prio = p->static_prio;
2663 * We don't need the reset flag anymore after the fork. It has
2664 * fulfilled its duty:
2666 p->sched_reset_on_fork = 0;
2670 * Make sure we do not leak PI boosting priority to the child.
2672 p->prio = current->normal_prio;
2674 if (!rt_prio(p->prio))
2675 p->sched_class = &fair_sched_class;
2677 if (p->sched_class->task_fork)
2678 p->sched_class->task_fork(p);
2680 set_task_cpu(p, cpu);
2682 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2683 if (likely(sched_info_on()))
2684 memset(&p->sched_info, 0, sizeof(p->sched_info));
2686 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2689 #ifdef CONFIG_PREEMPT
2690 /* Want to start with kernel preemption disabled. */
2691 task_thread_info(p)->preempt_count = 1;
2693 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2699 * wake_up_new_task - wake up a newly created task for the first time.
2701 * This function will do some initial scheduler statistics housekeeping
2702 * that must be done for every newly created context, then puts the task
2703 * on the runqueue and wakes it.
2705 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2707 unsigned long flags;
2709 int cpu = get_cpu();
2713 * Fork balancing, do it here and not earlier because:
2714 * - cpus_allowed can change in the fork path
2715 * - any previously selected cpu might disappear through hotplug
2717 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2718 * ->cpus_allowed is stable, we have preemption disabled, meaning
2719 * cpu_online_mask is stable.
2721 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2722 set_task_cpu(p, cpu);
2726 * Since the task is not on the rq and we still have TASK_WAKING set
2727 * nobody else will migrate this task.
2730 spin_lock_irqsave(&rq->lock, flags);
2732 BUG_ON(p->state != TASK_WAKING);
2733 p->state = TASK_RUNNING;
2734 update_rq_clock(rq);
2735 activate_task(rq, p, 0);
2736 trace_sched_wakeup_new(rq, p, 1);
2737 check_preempt_curr(rq, p, WF_FORK);
2739 if (p->sched_class->task_woken)
2740 p->sched_class->task_woken(rq, p);
2742 task_rq_unlock(rq, &flags);
2746 #ifdef CONFIG_PREEMPT_NOTIFIERS
2749 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2750 * @notifier: notifier struct to register
2752 void preempt_notifier_register(struct preempt_notifier *notifier)
2754 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2756 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2759 * preempt_notifier_unregister - no longer interested in preemption notifications
2760 * @notifier: notifier struct to unregister
2762 * This is safe to call from within a preemption notifier.
2764 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2766 hlist_del(¬ifier->link);
2768 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2770 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2772 struct preempt_notifier *notifier;
2773 struct hlist_node *node;
2775 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2776 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2780 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2781 struct task_struct *next)
2783 struct preempt_notifier *notifier;
2784 struct hlist_node *node;
2786 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2787 notifier->ops->sched_out(notifier, next);
2790 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2792 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2797 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2798 struct task_struct *next)
2802 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2805 * prepare_task_switch - prepare to switch tasks
2806 * @rq: the runqueue preparing to switch
2807 * @prev: the current task that is being switched out
2808 * @next: the task we are going to switch to.
2810 * This is called with the rq lock held and interrupts off. It must
2811 * be paired with a subsequent finish_task_switch after the context
2814 * prepare_task_switch sets up locking and calls architecture specific
2818 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2819 struct task_struct *next)
2821 fire_sched_out_preempt_notifiers(prev, next);
2822 prepare_lock_switch(rq, next);
2823 prepare_arch_switch(next);
2827 * finish_task_switch - clean up after a task-switch
2828 * @rq: runqueue associated with task-switch
2829 * @prev: the thread we just switched away from.
2831 * finish_task_switch must be called after the context switch, paired
2832 * with a prepare_task_switch call before the context switch.
2833 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2834 * and do any other architecture-specific cleanup actions.
2836 * Note that we may have delayed dropping an mm in context_switch(). If
2837 * so, we finish that here outside of the runqueue lock. (Doing it
2838 * with the lock held can cause deadlocks; see schedule() for
2841 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2842 __releases(rq->lock)
2844 struct mm_struct *mm = rq->prev_mm;
2850 * A task struct has one reference for the use as "current".
2851 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2852 * schedule one last time. The schedule call will never return, and
2853 * the scheduled task must drop that reference.
2854 * The test for TASK_DEAD must occur while the runqueue locks are
2855 * still held, otherwise prev could be scheduled on another cpu, die
2856 * there before we look at prev->state, and then the reference would
2858 * Manfred Spraul <manfred@colorfullife.com>
2860 prev_state = prev->state;
2861 finish_arch_switch(prev);
2862 perf_event_task_sched_in(current, cpu_of(rq));
2863 finish_lock_switch(rq, prev);
2865 fire_sched_in_preempt_notifiers(current);
2868 if (unlikely(prev_state == TASK_DEAD)) {
2870 * Remove function-return probe instances associated with this
2871 * task and put them back on the free list.
2873 kprobe_flush_task(prev);
2874 put_task_struct(prev);
2880 /* assumes rq->lock is held */
2881 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2883 if (prev->sched_class->pre_schedule)
2884 prev->sched_class->pre_schedule(rq, prev);
2887 /* rq->lock is NOT held, but preemption is disabled */
2888 static inline void post_schedule(struct rq *rq)
2890 if (rq->post_schedule) {
2891 unsigned long flags;
2893 spin_lock_irqsave(&rq->lock, flags);
2894 if (rq->curr->sched_class->post_schedule)
2895 rq->curr->sched_class->post_schedule(rq);
2896 spin_unlock_irqrestore(&rq->lock, flags);
2898 rq->post_schedule = 0;
2904 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2908 static inline void post_schedule(struct rq *rq)
2915 * schedule_tail - first thing a freshly forked thread must call.
2916 * @prev: the thread we just switched away from.
2918 asmlinkage void schedule_tail(struct task_struct *prev)
2919 __releases(rq->lock)
2921 struct rq *rq = this_rq();
2923 finish_task_switch(rq, prev);
2926 * FIXME: do we need to worry about rq being invalidated by the
2931 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2932 /* In this case, finish_task_switch does not reenable preemption */
2935 if (current->set_child_tid)
2936 put_user(task_pid_vnr(current), current->set_child_tid);
2940 * context_switch - switch to the new MM and the new
2941 * thread's register state.
2944 context_switch(struct rq *rq, struct task_struct *prev,
2945 struct task_struct *next)
2947 struct mm_struct *mm, *oldmm;
2949 prepare_task_switch(rq, prev, next);
2950 trace_sched_switch(rq, prev, next);
2952 oldmm = prev->active_mm;
2954 * For paravirt, this is coupled with an exit in switch_to to
2955 * combine the page table reload and the switch backend into
2958 arch_start_context_switch(prev);
2960 if (unlikely(!mm)) {
2961 next->active_mm = oldmm;
2962 atomic_inc(&oldmm->mm_count);
2963 enter_lazy_tlb(oldmm, next);
2965 switch_mm(oldmm, mm, next);
2967 if (unlikely(!prev->mm)) {
2968 prev->active_mm = NULL;
2969 rq->prev_mm = oldmm;
2972 * Since the runqueue lock will be released by the next
2973 * task (which is an invalid locking op but in the case
2974 * of the scheduler it's an obvious special-case), so we
2975 * do an early lockdep release here:
2977 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2978 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2981 /* Here we just switch the register state and the stack. */
2982 switch_to(prev, next, prev);
2986 * this_rq must be evaluated again because prev may have moved
2987 * CPUs since it called schedule(), thus the 'rq' on its stack
2988 * frame will be invalid.
2990 finish_task_switch(this_rq(), prev);
2994 * nr_running, nr_uninterruptible and nr_context_switches:
2996 * externally visible scheduler statistics: current number of runnable
2997 * threads, current number of uninterruptible-sleeping threads, total
2998 * number of context switches performed since bootup.
3000 unsigned long nr_running(void)
3002 unsigned long i, sum = 0;
3004 for_each_online_cpu(i)
3005 sum += cpu_rq(i)->nr_running;
3010 unsigned long nr_uninterruptible(void)
3012 unsigned long i, sum = 0;
3014 for_each_possible_cpu(i)
3015 sum += cpu_rq(i)->nr_uninterruptible;
3018 * Since we read the counters lockless, it might be slightly
3019 * inaccurate. Do not allow it to go below zero though:
3021 if (unlikely((long)sum < 0))
3027 unsigned long long nr_context_switches(void)
3030 unsigned long long sum = 0;
3032 for_each_possible_cpu(i)
3033 sum += cpu_rq(i)->nr_switches;
3038 unsigned long nr_iowait(void)
3040 unsigned long i, sum = 0;
3042 for_each_possible_cpu(i)
3043 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3048 unsigned long nr_iowait_cpu(void)
3050 struct rq *this = this_rq();
3051 return atomic_read(&this->nr_iowait);
3054 unsigned long this_cpu_load(void)
3056 struct rq *this = this_rq();
3057 return this->cpu_load[0];
3061 /* Variables and functions for calc_load */
3062 static atomic_long_t calc_load_tasks;
3063 static unsigned long calc_load_update;
3064 unsigned long avenrun[3];
3065 EXPORT_SYMBOL(avenrun);
3068 * get_avenrun - get the load average array
3069 * @loads: pointer to dest load array
3070 * @offset: offset to add
3071 * @shift: shift count to shift the result left
3073 * These values are estimates at best, so no need for locking.
3075 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3077 loads[0] = (avenrun[0] + offset) << shift;
3078 loads[1] = (avenrun[1] + offset) << shift;
3079 loads[2] = (avenrun[2] + offset) << shift;
3082 static unsigned long
3083 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3086 load += active * (FIXED_1 - exp);
3087 return load >> FSHIFT;
3091 * calc_load - update the avenrun load estimates 10 ticks after the
3092 * CPUs have updated calc_load_tasks.
3094 void calc_global_load(void)
3096 unsigned long upd = calc_load_update + 10;
3099 if (time_before(jiffies, upd))
3102 active = atomic_long_read(&calc_load_tasks);
3103 active = active > 0 ? active * FIXED_1 : 0;
3105 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3106 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3107 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3109 calc_load_update += LOAD_FREQ;
3113 * Either called from update_cpu_load() or from a cpu going idle
3115 static void calc_load_account_active(struct rq *this_rq)
3117 long nr_active, delta;
3119 nr_active = this_rq->nr_running;
3120 nr_active += (long) this_rq->nr_uninterruptible;
3122 if (nr_active != this_rq->calc_load_active) {
3123 delta = nr_active - this_rq->calc_load_active;
3124 this_rq->calc_load_active = nr_active;
3125 atomic_long_add(delta, &calc_load_tasks);
3130 * Update rq->cpu_load[] statistics. This function is usually called every
3131 * scheduler tick (TICK_NSEC).
3133 static void update_cpu_load(struct rq *this_rq)
3135 unsigned long this_load = this_rq->load.weight;
3138 this_rq->nr_load_updates++;
3140 /* Update our load: */
3141 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3142 unsigned long old_load, new_load;
3144 /* scale is effectively 1 << i now, and >> i divides by scale */
3146 old_load = this_rq->cpu_load[i];
3147 new_load = this_load;
3149 * Round up the averaging division if load is increasing. This
3150 * prevents us from getting stuck on 9 if the load is 10, for
3153 if (new_load > old_load)
3154 new_load += scale-1;
3155 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3158 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3159 this_rq->calc_load_update += LOAD_FREQ;
3160 calc_load_account_active(this_rq);
3167 * double_rq_lock - safely lock two runqueues
3169 * Note this does not disable interrupts like task_rq_lock,
3170 * you need to do so manually before calling.
3172 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3173 __acquires(rq1->lock)
3174 __acquires(rq2->lock)
3176 BUG_ON(!irqs_disabled());
3178 spin_lock(&rq1->lock);
3179 __acquire(rq2->lock); /* Fake it out ;) */
3182 spin_lock(&rq1->lock);
3183 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3185 spin_lock(&rq2->lock);
3186 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3189 update_rq_clock(rq1);
3190 update_rq_clock(rq2);
3194 * double_rq_unlock - safely unlock two runqueues
3196 * Note this does not restore interrupts like task_rq_unlock,
3197 * you need to do so manually after calling.
3199 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3200 __releases(rq1->lock)
3201 __releases(rq2->lock)
3203 spin_unlock(&rq1->lock);
3205 spin_unlock(&rq2->lock);
3207 __release(rq2->lock);
3211 * sched_exec - execve() is a valuable balancing opportunity, because at
3212 * this point the task has the smallest effective memory and cache footprint.
3214 void sched_exec(void)
3216 struct task_struct *p = current;
3217 struct migration_req req;
3218 int dest_cpu, this_cpu;
3219 unsigned long flags;
3222 this_cpu = get_cpu();
3223 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3224 if (dest_cpu == this_cpu) {
3229 rq = task_rq_lock(p, &flags);
3232 * select_task_rq() can race against ->cpus_allowed
3234 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3235 likely(cpu_active(dest_cpu)) &&
3236 migrate_task(p, dest_cpu, &req)) {
3237 /* Need to wait for migration thread (might exit: take ref). */
3238 struct task_struct *mt = rq->migration_thread;
3240 get_task_struct(mt);
3241 task_rq_unlock(rq, &flags);
3242 wake_up_process(mt);
3243 put_task_struct(mt);
3244 wait_for_completion(&req.done);
3248 task_rq_unlock(rq, &flags);
3252 * pull_task - move a task from a remote runqueue to the local runqueue.
3253 * Both runqueues must be locked.
3255 static void pull_task(struct rq *src_rq, struct task_struct *p,
3256 struct rq *this_rq, int this_cpu)
3258 deactivate_task(src_rq, p, 0);
3259 set_task_cpu(p, this_cpu);
3260 activate_task(this_rq, p, 0);
3261 check_preempt_curr(this_rq, p, 0);
3265 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3268 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3269 struct sched_domain *sd, enum cpu_idle_type idle,
3272 int tsk_cache_hot = 0;
3274 * We do not migrate tasks that are:
3275 * 1) running (obviously), or
3276 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3277 * 3) are cache-hot on their current CPU.
3279 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3280 schedstat_inc(p, se.nr_failed_migrations_affine);
3285 if (task_running(rq, p)) {
3286 schedstat_inc(p, se.nr_failed_migrations_running);
3291 * Aggressive migration if:
3292 * 1) task is cache cold, or
3293 * 2) too many balance attempts have failed.
3296 tsk_cache_hot = task_hot(p, rq->clock, sd);
3297 if (!tsk_cache_hot ||
3298 sd->nr_balance_failed > sd->cache_nice_tries) {
3299 #ifdef CONFIG_SCHEDSTATS
3300 if (tsk_cache_hot) {
3301 schedstat_inc(sd, lb_hot_gained[idle]);
3302 schedstat_inc(p, se.nr_forced_migrations);
3308 if (tsk_cache_hot) {
3309 schedstat_inc(p, se.nr_failed_migrations_hot);
3315 static unsigned long
3316 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3317 unsigned long max_load_move, struct sched_domain *sd,
3318 enum cpu_idle_type idle, int *all_pinned,
3319 int *this_best_prio, struct rq_iterator *iterator)
3321 int loops = 0, pulled = 0, pinned = 0;
3322 struct task_struct *p;
3323 long rem_load_move = max_load_move;
3325 if (max_load_move == 0)
3331 * Start the load-balancing iterator:
3333 p = iterator->start(iterator->arg);
3335 if (!p || loops++ > sysctl_sched_nr_migrate)
3338 if ((p->se.load.weight >> 1) > rem_load_move ||
3339 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3340 p = iterator->next(iterator->arg);
3344 pull_task(busiest, p, this_rq, this_cpu);
3346 rem_load_move -= p->se.load.weight;
3348 #ifdef CONFIG_PREEMPT
3350 * NEWIDLE balancing is a source of latency, so preemptible kernels
3351 * will stop after the first task is pulled to minimize the critical
3354 if (idle == CPU_NEWLY_IDLE)
3359 * We only want to steal up to the prescribed amount of weighted load.
3361 if (rem_load_move > 0) {
3362 if (p->prio < *this_best_prio)
3363 *this_best_prio = p->prio;
3364 p = iterator->next(iterator->arg);
3369 * Right now, this is one of only two places pull_task() is called,
3370 * so we can safely collect pull_task() stats here rather than
3371 * inside pull_task().
3373 schedstat_add(sd, lb_gained[idle], pulled);
3376 *all_pinned = pinned;
3378 return max_load_move - rem_load_move;
3382 * move_tasks tries to move up to max_load_move weighted load from busiest to
3383 * this_rq, as part of a balancing operation within domain "sd".
3384 * Returns 1 if successful and 0 otherwise.
3386 * Called with both runqueues locked.
3388 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3389 unsigned long max_load_move,
3390 struct sched_domain *sd, enum cpu_idle_type idle,
3393 const struct sched_class *class = sched_class_highest;
3394 unsigned long total_load_moved = 0;
3395 int this_best_prio = this_rq->curr->prio;
3399 class->load_balance(this_rq, this_cpu, busiest,
3400 max_load_move - total_load_moved,
3401 sd, idle, all_pinned, &this_best_prio);
3402 class = class->next;
3404 #ifdef CONFIG_PREEMPT
3406 * NEWIDLE balancing is a source of latency, so preemptible
3407 * kernels will stop after the first task is pulled to minimize
3408 * the critical section.
3410 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3413 } while (class && max_load_move > total_load_moved);
3415 return total_load_moved > 0;
3419 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3420 struct sched_domain *sd, enum cpu_idle_type idle,
3421 struct rq_iterator *iterator)
3423 struct task_struct *p = iterator->start(iterator->arg);
3427 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3428 pull_task(busiest, p, this_rq, this_cpu);
3430 * Right now, this is only the second place pull_task()
3431 * is called, so we can safely collect pull_task()
3432 * stats here rather than inside pull_task().
3434 schedstat_inc(sd, lb_gained[idle]);
3438 p = iterator->next(iterator->arg);
3445 * move_one_task tries to move exactly one task from busiest to this_rq, as
3446 * part of active balancing operations within "domain".
3447 * Returns 1 if successful and 0 otherwise.
3449 * Called with both runqueues locked.
3451 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3452 struct sched_domain *sd, enum cpu_idle_type idle)
3454 const struct sched_class *class;
3456 for_each_class(class) {
3457 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3463 /********** Helpers for find_busiest_group ************************/
3465 * sd_lb_stats - Structure to store the statistics of a sched_domain
3466 * during load balancing.
3468 struct sd_lb_stats {
3469 struct sched_group *busiest; /* Busiest group in this sd */
3470 struct sched_group *this; /* Local group in this sd */
3471 unsigned long total_load; /* Total load of all groups in sd */
3472 unsigned long total_pwr; /* Total power of all groups in sd */
3473 unsigned long avg_load; /* Average load across all groups in sd */
3475 /** Statistics of this group */
3476 unsigned long this_load;
3477 unsigned long this_load_per_task;
3478 unsigned long this_nr_running;
3480 /* Statistics of the busiest group */
3481 unsigned long max_load;
3482 unsigned long busiest_load_per_task;
3483 unsigned long busiest_nr_running;
3484 unsigned long busiest_group_capacity;
3486 int group_imb; /* Is there imbalance in this sd */
3487 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3488 int power_savings_balance; /* Is powersave balance needed for this sd */
3489 struct sched_group *group_min; /* Least loaded group in sd */
3490 struct sched_group *group_leader; /* Group which relieves group_min */
3491 unsigned long min_load_per_task; /* load_per_task in group_min */
3492 unsigned long leader_nr_running; /* Nr running of group_leader */
3493 unsigned long min_nr_running; /* Nr running of group_min */
3498 * sg_lb_stats - stats of a sched_group required for load_balancing
3500 struct sg_lb_stats {
3501 unsigned long avg_load; /*Avg load across the CPUs of the group */
3502 unsigned long group_load; /* Total load over the CPUs of the group */
3503 unsigned long sum_nr_running; /* Nr tasks running in the group */
3504 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3505 unsigned long group_capacity;
3506 int group_imb; /* Is there an imbalance in the group ? */
3510 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3511 * @group: The group whose first cpu is to be returned.
3513 static inline unsigned int group_first_cpu(struct sched_group *group)
3515 return cpumask_first(sched_group_cpus(group));
3519 * get_sd_load_idx - Obtain the load index for a given sched domain.
3520 * @sd: The sched_domain whose load_idx is to be obtained.
3521 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3523 static inline int get_sd_load_idx(struct sched_domain *sd,
3524 enum cpu_idle_type idle)
3530 load_idx = sd->busy_idx;
3533 case CPU_NEWLY_IDLE:
3534 load_idx = sd->newidle_idx;
3537 load_idx = sd->idle_idx;
3545 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3547 * init_sd_power_savings_stats - Initialize power savings statistics for
3548 * the given sched_domain, during load balancing.
3550 * @sd: Sched domain whose power-savings statistics are to be initialized.
3551 * @sds: Variable containing the statistics for sd.
3552 * @idle: Idle status of the CPU at which we're performing load-balancing.
3554 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3555 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3558 * Busy processors will not participate in power savings
3561 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3562 sds->power_savings_balance = 0;
3564 sds->power_savings_balance = 1;
3565 sds->min_nr_running = ULONG_MAX;
3566 sds->leader_nr_running = 0;
3571 * update_sd_power_savings_stats - Update the power saving stats for a
3572 * sched_domain while performing load balancing.
3574 * @group: sched_group belonging to the sched_domain under consideration.
3575 * @sds: Variable containing the statistics of the sched_domain
3576 * @local_group: Does group contain the CPU for which we're performing
3578 * @sgs: Variable containing the statistics of the group.
3580 static inline void update_sd_power_savings_stats(struct sched_group *group,
3581 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3584 if (!sds->power_savings_balance)
3588 * If the local group is idle or completely loaded
3589 * no need to do power savings balance at this domain
3591 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3592 !sds->this_nr_running))
3593 sds->power_savings_balance = 0;
3596 * If a group is already running at full capacity or idle,
3597 * don't include that group in power savings calculations
3599 if (!sds->power_savings_balance ||
3600 sgs->sum_nr_running >= sgs->group_capacity ||
3601 !sgs->sum_nr_running)
3605 * Calculate the group which has the least non-idle load.
3606 * This is the group from where we need to pick up the load
3609 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3610 (sgs->sum_nr_running == sds->min_nr_running &&
3611 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3612 sds->group_min = group;
3613 sds->min_nr_running = sgs->sum_nr_running;
3614 sds->min_load_per_task = sgs->sum_weighted_load /
3615 sgs->sum_nr_running;
3619 * Calculate the group which is almost near its
3620 * capacity but still has some space to pick up some load
3621 * from other group and save more power
3623 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3626 if (sgs->sum_nr_running > sds->leader_nr_running ||
3627 (sgs->sum_nr_running == sds->leader_nr_running &&
3628 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3629 sds->group_leader = group;
3630 sds->leader_nr_running = sgs->sum_nr_running;
3635 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3636 * @sds: Variable containing the statistics of the sched_domain
3637 * under consideration.
3638 * @this_cpu: Cpu at which we're currently performing load-balancing.
3639 * @imbalance: Variable to store the imbalance.
3642 * Check if we have potential to perform some power-savings balance.
3643 * If yes, set the busiest group to be the least loaded group in the
3644 * sched_domain, so that it's CPUs can be put to idle.
3646 * Returns 1 if there is potential to perform power-savings balance.
3649 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3650 int this_cpu, unsigned long *imbalance)
3652 if (!sds->power_savings_balance)
3655 if (sds->this != sds->group_leader ||
3656 sds->group_leader == sds->group_min)
3659 *imbalance = sds->min_load_per_task;
3660 sds->busiest = sds->group_min;
3665 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3666 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3667 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3672 static inline void update_sd_power_savings_stats(struct sched_group *group,
3673 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3678 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3679 int this_cpu, unsigned long *imbalance)
3683 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3686 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3688 return SCHED_LOAD_SCALE;
3691 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3693 return default_scale_freq_power(sd, cpu);
3696 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3698 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3699 unsigned long smt_gain = sd->smt_gain;
3706 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3708 return default_scale_smt_power(sd, cpu);
3711 unsigned long scale_rt_power(int cpu)
3713 struct rq *rq = cpu_rq(cpu);
3714 u64 total, available;
3716 sched_avg_update(rq);
3718 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3719 available = total - rq->rt_avg;
3721 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3722 total = SCHED_LOAD_SCALE;
3724 total >>= SCHED_LOAD_SHIFT;
3726 return div_u64(available, total);
3729 static void update_cpu_power(struct sched_domain *sd, int cpu)
3731 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3732 unsigned long power = SCHED_LOAD_SCALE;
3733 struct sched_group *sdg = sd->groups;
3735 if (sched_feat(ARCH_POWER))
3736 power *= arch_scale_freq_power(sd, cpu);
3738 power *= default_scale_freq_power(sd, cpu);
3740 power >>= SCHED_LOAD_SHIFT;
3742 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3743 if (sched_feat(ARCH_POWER))
3744 power *= arch_scale_smt_power(sd, cpu);
3746 power *= default_scale_smt_power(sd, cpu);
3748 power >>= SCHED_LOAD_SHIFT;
3751 power *= scale_rt_power(cpu);
3752 power >>= SCHED_LOAD_SHIFT;
3757 sdg->cpu_power = power;
3760 static void update_group_power(struct sched_domain *sd, int cpu)
3762 struct sched_domain *child = sd->child;
3763 struct sched_group *group, *sdg = sd->groups;
3764 unsigned long power;
3767 update_cpu_power(sd, cpu);
3773 group = child->groups;
3775 power += group->cpu_power;
3776 group = group->next;
3777 } while (group != child->groups);
3779 sdg->cpu_power = power;
3783 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3784 * @sd: The sched_domain whose statistics are to be updated.
3785 * @group: sched_group whose statistics are to be updated.
3786 * @this_cpu: Cpu for which load balance is currently performed.
3787 * @idle: Idle status of this_cpu
3788 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3789 * @sd_idle: Idle status of the sched_domain containing group.
3790 * @local_group: Does group contain this_cpu.
3791 * @cpus: Set of cpus considered for load balancing.
3792 * @balance: Should we balance.
3793 * @sgs: variable to hold the statistics for this group.
3795 static inline void update_sg_lb_stats(struct sched_domain *sd,
3796 struct sched_group *group, int this_cpu,
3797 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3798 int local_group, const struct cpumask *cpus,
3799 int *balance, struct sg_lb_stats *sgs)
3801 unsigned long load, max_cpu_load, min_cpu_load;
3803 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3804 unsigned long avg_load_per_task = 0;
3807 balance_cpu = group_first_cpu(group);
3808 if (balance_cpu == this_cpu)
3809 update_group_power(sd, this_cpu);
3812 /* Tally up the load of all CPUs in the group */
3814 min_cpu_load = ~0UL;
3816 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3817 struct rq *rq = cpu_rq(i);
3819 if (*sd_idle && rq->nr_running)
3822 /* Bias balancing toward cpus of our domain */
3824 if (idle_cpu(i) && !first_idle_cpu) {
3829 load = target_load(i, load_idx);
3831 load = source_load(i, load_idx);
3832 if (load > max_cpu_load)
3833 max_cpu_load = load;
3834 if (min_cpu_load > load)
3835 min_cpu_load = load;
3838 sgs->group_load += load;
3839 sgs->sum_nr_running += rq->nr_running;
3840 sgs->sum_weighted_load += weighted_cpuload(i);
3845 * First idle cpu or the first cpu(busiest) in this sched group
3846 * is eligible for doing load balancing at this and above
3847 * domains. In the newly idle case, we will allow all the cpu's
3848 * to do the newly idle load balance.
3850 if (idle != CPU_NEWLY_IDLE && local_group &&
3851 balance_cpu != this_cpu && balance) {
3856 /* Adjust by relative CPU power of the group */
3857 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3860 * Consider the group unbalanced when the imbalance is larger
3861 * than the average weight of two tasks.
3863 * APZ: with cgroup the avg task weight can vary wildly and
3864 * might not be a suitable number - should we keep a
3865 * normalized nr_running number somewhere that negates
3868 if (sgs->sum_nr_running)
3869 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3871 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3874 sgs->group_capacity =
3875 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3879 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3880 * @sd: sched_domain whose statistics are to be updated.
3881 * @this_cpu: Cpu for which load balance is currently performed.
3882 * @idle: Idle status of this_cpu
3883 * @sd_idle: Idle status of the sched_domain containing group.
3884 * @cpus: Set of cpus considered for load balancing.
3885 * @balance: Should we balance.
3886 * @sds: variable to hold the statistics for this sched_domain.
3888 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3889 enum cpu_idle_type idle, int *sd_idle,
3890 const struct cpumask *cpus, int *balance,
3891 struct sd_lb_stats *sds)
3893 struct sched_domain *child = sd->child;
3894 struct sched_group *group = sd->groups;
3895 struct sg_lb_stats sgs;
3896 int load_idx, prefer_sibling = 0;
3898 if (child && child->flags & SD_PREFER_SIBLING)
3901 init_sd_power_savings_stats(sd, sds, idle);
3902 load_idx = get_sd_load_idx(sd, idle);
3907 local_group = cpumask_test_cpu(this_cpu,
3908 sched_group_cpus(group));
3909 memset(&sgs, 0, sizeof(sgs));
3910 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3911 local_group, cpus, balance, &sgs);
3913 if (local_group && balance && !(*balance))
3916 sds->total_load += sgs.group_load;
3917 sds->total_pwr += group->cpu_power;
3920 * In case the child domain prefers tasks go to siblings
3921 * first, lower the group capacity to one so that we'll try
3922 * and move all the excess tasks away.
3925 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3928 sds->this_load = sgs.avg_load;
3930 sds->this_nr_running = sgs.sum_nr_running;
3931 sds->this_load_per_task = sgs.sum_weighted_load;
3932 } else if (sgs.avg_load > sds->max_load &&
3933 (sgs.sum_nr_running > sgs.group_capacity ||
3935 sds->max_load = sgs.avg_load;
3936 sds->busiest = group;
3937 sds->busiest_nr_running = sgs.sum_nr_running;
3938 sds->busiest_group_capacity = sgs.group_capacity;
3939 sds->busiest_load_per_task = sgs.sum_weighted_load;
3940 sds->group_imb = sgs.group_imb;
3943 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3944 group = group->next;
3945 } while (group != sd->groups);
3949 * fix_small_imbalance - Calculate the minor imbalance that exists
3950 * amongst the groups of a sched_domain, during
3952 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3953 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3954 * @imbalance: Variable to store the imbalance.
3956 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3957 int this_cpu, unsigned long *imbalance)
3959 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3960 unsigned int imbn = 2;
3961 unsigned long scaled_busy_load_per_task;
3963 if (sds->this_nr_running) {
3964 sds->this_load_per_task /= sds->this_nr_running;
3965 if (sds->busiest_load_per_task >
3966 sds->this_load_per_task)
3969 sds->this_load_per_task =
3970 cpu_avg_load_per_task(this_cpu);
3972 scaled_busy_load_per_task = sds->busiest_load_per_task
3974 scaled_busy_load_per_task /= sds->busiest->cpu_power;
3976 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
3977 (scaled_busy_load_per_task * imbn)) {
3978 *imbalance = sds->busiest_load_per_task;
3983 * OK, we don't have enough imbalance to justify moving tasks,
3984 * however we may be able to increase total CPU power used by
3988 pwr_now += sds->busiest->cpu_power *
3989 min(sds->busiest_load_per_task, sds->max_load);
3990 pwr_now += sds->this->cpu_power *
3991 min(sds->this_load_per_task, sds->this_load);
3992 pwr_now /= SCHED_LOAD_SCALE;
3994 /* Amount of load we'd subtract */
3995 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3996 sds->busiest->cpu_power;
3997 if (sds->max_load > tmp)
3998 pwr_move += sds->busiest->cpu_power *
3999 min(sds->busiest_load_per_task, sds->max_load - tmp);
4001 /* Amount of load we'd add */
4002 if (sds->max_load * sds->busiest->cpu_power <
4003 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
4004 tmp = (sds->max_load * sds->busiest->cpu_power) /
4005 sds->this->cpu_power;
4007 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
4008 sds->this->cpu_power;
4009 pwr_move += sds->this->cpu_power *
4010 min(sds->this_load_per_task, sds->this_load + tmp);
4011 pwr_move /= SCHED_LOAD_SCALE;
4013 /* Move if we gain throughput */
4014 if (pwr_move > pwr_now)
4015 *imbalance = sds->busiest_load_per_task;
4019 * calculate_imbalance - Calculate the amount of imbalance present within the
4020 * groups of a given sched_domain during load balance.
4021 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4022 * @this_cpu: Cpu for which currently load balance is being performed.
4023 * @imbalance: The variable to store the imbalance.
4025 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
4026 unsigned long *imbalance)
4028 unsigned long max_pull, load_above_capacity = ~0UL;
4030 sds->busiest_load_per_task /= sds->busiest_nr_running;
4031 if (sds->group_imb) {
4032 sds->busiest_load_per_task =
4033 min(sds->busiest_load_per_task, sds->avg_load);
4037 * In the presence of smp nice balancing, certain scenarios can have
4038 * max load less than avg load(as we skip the groups at or below
4039 * its cpu_power, while calculating max_load..)
4041 if (sds->max_load < sds->avg_load) {
4043 return fix_small_imbalance(sds, this_cpu, imbalance);
4046 if (!sds->group_imb) {
4048 * Don't want to pull so many tasks that a group would go idle.
4050 load_above_capacity = (sds->busiest_nr_running -
4051 sds->busiest_group_capacity);
4053 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE);
4055 load_above_capacity /= sds->busiest->cpu_power;
4059 * We're trying to get all the cpus to the average_load, so we don't
4060 * want to push ourselves above the average load, nor do we wish to
4061 * reduce the max loaded cpu below the average load. At the same time,
4062 * we also don't want to reduce the group load below the group capacity
4063 * (so that we can implement power-savings policies etc). Thus we look
4064 * for the minimum possible imbalance.
4065 * Be careful of negative numbers as they'll appear as very large values
4066 * with unsigned longs.
4068 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4070 /* How much load to actually move to equalise the imbalance */
4071 *imbalance = min(max_pull * sds->busiest->cpu_power,
4072 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
4076 * if *imbalance is less than the average load per runnable task
4077 * there is no gaurantee that any tasks will be moved so we'll have
4078 * a think about bumping its value to force at least one task to be
4081 if (*imbalance < sds->busiest_load_per_task)
4082 return fix_small_imbalance(sds, this_cpu, imbalance);
4085 /******* find_busiest_group() helpers end here *********************/
4088 * find_busiest_group - Returns the busiest group within the sched_domain
4089 * if there is an imbalance. If there isn't an imbalance, and
4090 * the user has opted for power-savings, it returns a group whose
4091 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4092 * such a group exists.
4094 * Also calculates the amount of weighted load which should be moved
4095 * to restore balance.
4097 * @sd: The sched_domain whose busiest group is to be returned.
4098 * @this_cpu: The cpu for which load balancing is currently being performed.
4099 * @imbalance: Variable which stores amount of weighted load which should
4100 * be moved to restore balance/put a group to idle.
4101 * @idle: The idle status of this_cpu.
4102 * @sd_idle: The idleness of sd
4103 * @cpus: The set of CPUs under consideration for load-balancing.
4104 * @balance: Pointer to a variable indicating if this_cpu
4105 * is the appropriate cpu to perform load balancing at this_level.
4107 * Returns: - the busiest group if imbalance exists.
4108 * - If no imbalance and user has opted for power-savings balance,
4109 * return the least loaded group whose CPUs can be
4110 * put to idle by rebalancing its tasks onto our group.
4112 static struct sched_group *
4113 find_busiest_group(struct sched_domain *sd, int this_cpu,
4114 unsigned long *imbalance, enum cpu_idle_type idle,
4115 int *sd_idle, const struct cpumask *cpus, int *balance)
4117 struct sd_lb_stats sds;
4119 memset(&sds, 0, sizeof(sds));
4122 * Compute the various statistics relavent for load balancing at
4125 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4128 /* Cases where imbalance does not exist from POV of this_cpu */
4129 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4131 * 2) There is no busy sibling group to pull from.
4132 * 3) This group is the busiest group.
4133 * 4) This group is more busy than the avg busieness at this
4135 * 5) The imbalance is within the specified limit.
4137 if (balance && !(*balance))
4140 if (!sds.busiest || sds.busiest_nr_running == 0)
4143 if (sds.this_load >= sds.max_load)
4146 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4148 if (sds.this_load >= sds.avg_load)
4151 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4154 /* Looks like there is an imbalance. Compute it */
4155 calculate_imbalance(&sds, this_cpu, imbalance);
4160 * There is no obvious imbalance. But check if we can do some balancing
4163 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4171 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4174 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4175 unsigned long imbalance, const struct cpumask *cpus)
4177 struct rq *busiest = NULL, *rq;
4178 unsigned long max_load = 0;
4181 for_each_cpu(i, sched_group_cpus(group)) {
4182 unsigned long power = power_of(i);
4183 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4186 if (!cpumask_test_cpu(i, cpus))
4190 wl = weighted_cpuload(i);
4193 * When comparing with imbalance, use weighted_cpuload()
4194 * which is not scaled with the cpu power.
4196 if (capacity && rq->nr_running == 1 && wl > imbalance)
4200 * For the load comparisons with the other cpu's, consider
4201 * the weighted_cpuload() scaled with the cpu power, so that
4202 * the load can be moved away from the cpu that is potentially
4203 * running at a lower capacity.
4205 wl = (wl * SCHED_LOAD_SCALE) / power;
4207 if (wl > max_load) {
4217 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4218 * so long as it is large enough.
4220 #define MAX_PINNED_INTERVAL 512
4222 /* Working cpumask for load_balance and load_balance_newidle. */
4223 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4226 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4227 * tasks if there is an imbalance.
4229 static int load_balance(int this_cpu, struct rq *this_rq,
4230 struct sched_domain *sd, enum cpu_idle_type idle,
4233 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4234 struct sched_group *group;
4235 unsigned long imbalance;
4237 unsigned long flags;
4238 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4240 cpumask_copy(cpus, cpu_active_mask);
4243 * When power savings policy is enabled for the parent domain, idle
4244 * sibling can pick up load irrespective of busy siblings. In this case,
4245 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4246 * portraying it as CPU_NOT_IDLE.
4248 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4249 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4252 schedstat_inc(sd, lb_count[idle]);
4256 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4263 schedstat_inc(sd, lb_nobusyg[idle]);
4267 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4269 schedstat_inc(sd, lb_nobusyq[idle]);
4273 BUG_ON(busiest == this_rq);
4275 schedstat_add(sd, lb_imbalance[idle], imbalance);
4278 if (busiest->nr_running > 1) {
4280 * Attempt to move tasks. If find_busiest_group has found
4281 * an imbalance but busiest->nr_running <= 1, the group is
4282 * still unbalanced. ld_moved simply stays zero, so it is
4283 * correctly treated as an imbalance.
4285 local_irq_save(flags);
4286 double_rq_lock(this_rq, busiest);
4287 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4288 imbalance, sd, idle, &all_pinned);
4289 double_rq_unlock(this_rq, busiest);
4290 local_irq_restore(flags);
4293 * some other cpu did the load balance for us.
4295 if (ld_moved && this_cpu != smp_processor_id())
4296 resched_cpu(this_cpu);
4298 /* All tasks on this runqueue were pinned by CPU affinity */
4299 if (unlikely(all_pinned)) {
4300 cpumask_clear_cpu(cpu_of(busiest), cpus);
4301 if (!cpumask_empty(cpus))
4308 schedstat_inc(sd, lb_failed[idle]);
4309 sd->nr_balance_failed++;
4311 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4313 spin_lock_irqsave(&busiest->lock, flags);
4315 /* don't kick the migration_thread, if the curr
4316 * task on busiest cpu can't be moved to this_cpu
4318 if (!cpumask_test_cpu(this_cpu,
4319 &busiest->curr->cpus_allowed)) {
4320 spin_unlock_irqrestore(&busiest->lock, flags);
4322 goto out_one_pinned;
4325 if (!busiest->active_balance) {
4326 busiest->active_balance = 1;
4327 busiest->push_cpu = this_cpu;
4330 spin_unlock_irqrestore(&busiest->lock, flags);
4332 wake_up_process(busiest->migration_thread);
4335 * We've kicked active balancing, reset the failure
4338 sd->nr_balance_failed = sd->cache_nice_tries+1;
4341 sd->nr_balance_failed = 0;
4343 if (likely(!active_balance)) {
4344 /* We were unbalanced, so reset the balancing interval */
4345 sd->balance_interval = sd->min_interval;
4348 * If we've begun active balancing, start to back off. This
4349 * case may not be covered by the all_pinned logic if there
4350 * is only 1 task on the busy runqueue (because we don't call
4353 if (sd->balance_interval < sd->max_interval)
4354 sd->balance_interval *= 2;
4357 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4358 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4364 schedstat_inc(sd, lb_balanced[idle]);
4366 sd->nr_balance_failed = 0;
4369 /* tune up the balancing interval */
4370 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4371 (sd->balance_interval < sd->max_interval))
4372 sd->balance_interval *= 2;
4374 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4375 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4386 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4387 * tasks if there is an imbalance.
4389 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4390 * this_rq is locked.
4393 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4395 struct sched_group *group;
4396 struct rq *busiest = NULL;
4397 unsigned long imbalance;
4401 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4403 cpumask_copy(cpus, cpu_active_mask);
4406 * When power savings policy is enabled for the parent domain, idle
4407 * sibling can pick up load irrespective of busy siblings. In this case,
4408 * let the state of idle sibling percolate up as IDLE, instead of
4409 * portraying it as CPU_NOT_IDLE.
4411 if (sd->flags & SD_SHARE_CPUPOWER &&
4412 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4415 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4417 update_shares_locked(this_rq, sd);
4418 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4419 &sd_idle, cpus, NULL);
4421 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4425 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4427 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4431 BUG_ON(busiest == this_rq);
4433 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4436 if (busiest->nr_running > 1) {
4437 /* Attempt to move tasks */
4438 double_lock_balance(this_rq, busiest);
4439 /* this_rq->clock is already updated */
4440 update_rq_clock(busiest);
4441 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4442 imbalance, sd, CPU_NEWLY_IDLE,
4444 double_unlock_balance(this_rq, busiest);
4446 if (unlikely(all_pinned)) {
4447 cpumask_clear_cpu(cpu_of(busiest), cpus);
4448 if (!cpumask_empty(cpus))
4454 int active_balance = 0;
4456 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4457 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4458 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4461 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4464 if (sd->nr_balance_failed++ < 2)
4468 * The only task running in a non-idle cpu can be moved to this
4469 * cpu in an attempt to completely freeup the other CPU
4470 * package. The same method used to move task in load_balance()
4471 * have been extended for load_balance_newidle() to speedup
4472 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4474 * The package power saving logic comes from
4475 * find_busiest_group(). If there are no imbalance, then
4476 * f_b_g() will return NULL. However when sched_mc={1,2} then
4477 * f_b_g() will select a group from which a running task may be
4478 * pulled to this cpu in order to make the other package idle.
4479 * If there is no opportunity to make a package idle and if
4480 * there are no imbalance, then f_b_g() will return NULL and no
4481 * action will be taken in load_balance_newidle().
4483 * Under normal task pull operation due to imbalance, there
4484 * will be more than one task in the source run queue and
4485 * move_tasks() will succeed. ld_moved will be true and this
4486 * active balance code will not be triggered.
4489 /* Lock busiest in correct order while this_rq is held */
4490 double_lock_balance(this_rq, busiest);
4493 * don't kick the migration_thread, if the curr
4494 * task on busiest cpu can't be moved to this_cpu
4496 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4497 double_unlock_balance(this_rq, busiest);
4502 if (!busiest->active_balance) {
4503 busiest->active_balance = 1;
4504 busiest->push_cpu = this_cpu;
4508 double_unlock_balance(this_rq, busiest);
4510 * Should not call ttwu while holding a rq->lock
4512 spin_unlock(&this_rq->lock);
4514 wake_up_process(busiest->migration_thread);
4515 spin_lock(&this_rq->lock);
4518 sd->nr_balance_failed = 0;
4520 update_shares_locked(this_rq, sd);
4524 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4525 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4526 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4528 sd->nr_balance_failed = 0;
4534 * idle_balance is called by schedule() if this_cpu is about to become
4535 * idle. Attempts to pull tasks from other CPUs.
4537 static void idle_balance(int this_cpu, struct rq *this_rq)
4539 struct sched_domain *sd;
4540 int pulled_task = 0;
4541 unsigned long next_balance = jiffies + HZ;
4543 this_rq->idle_stamp = this_rq->clock;
4545 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4548 for_each_domain(this_cpu, sd) {
4549 unsigned long interval;
4551 if (!(sd->flags & SD_LOAD_BALANCE))
4554 if (sd->flags & SD_BALANCE_NEWIDLE)
4555 /* If we've pulled tasks over stop searching: */
4556 pulled_task = load_balance_newidle(this_cpu, this_rq,
4559 interval = msecs_to_jiffies(sd->balance_interval);
4560 if (time_after(next_balance, sd->last_balance + interval))
4561 next_balance = sd->last_balance + interval;
4563 this_rq->idle_stamp = 0;
4567 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4569 * We are going idle. next_balance may be set based on
4570 * a busy processor. So reset next_balance.
4572 this_rq->next_balance = next_balance;
4577 * active_load_balance is run by migration threads. It pushes running tasks
4578 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4579 * running on each physical CPU where possible, and avoids physical /
4580 * logical imbalances.
4582 * Called with busiest_rq locked.
4584 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4586 int target_cpu = busiest_rq->push_cpu;
4587 struct sched_domain *sd;
4588 struct rq *target_rq;
4590 /* Is there any task to move? */
4591 if (busiest_rq->nr_running <= 1)
4594 target_rq = cpu_rq(target_cpu);
4597 * This condition is "impossible", if it occurs
4598 * we need to fix it. Originally reported by
4599 * Bjorn Helgaas on a 128-cpu setup.
4601 BUG_ON(busiest_rq == target_rq);
4603 /* move a task from busiest_rq to target_rq */
4604 double_lock_balance(busiest_rq, target_rq);
4605 update_rq_clock(busiest_rq);
4606 update_rq_clock(target_rq);
4608 /* Search for an sd spanning us and the target CPU. */
4609 for_each_domain(target_cpu, sd) {
4610 if ((sd->flags & SD_LOAD_BALANCE) &&
4611 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4616 schedstat_inc(sd, alb_count);
4618 if (move_one_task(target_rq, target_cpu, busiest_rq,
4620 schedstat_inc(sd, alb_pushed);
4622 schedstat_inc(sd, alb_failed);
4624 double_unlock_balance(busiest_rq, target_rq);
4629 atomic_t load_balancer;
4630 cpumask_var_t cpu_mask;
4631 cpumask_var_t ilb_grp_nohz_mask;
4632 } nohz ____cacheline_aligned = {
4633 .load_balancer = ATOMIC_INIT(-1),
4636 int get_nohz_load_balancer(void)
4638 return atomic_read(&nohz.load_balancer);
4641 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4643 * lowest_flag_domain - Return lowest sched_domain containing flag.
4644 * @cpu: The cpu whose lowest level of sched domain is to
4646 * @flag: The flag to check for the lowest sched_domain
4647 * for the given cpu.
4649 * Returns the lowest sched_domain of a cpu which contains the given flag.
4651 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4653 struct sched_domain *sd;
4655 for_each_domain(cpu, sd)
4656 if (sd && (sd->flags & flag))
4663 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4664 * @cpu: The cpu whose domains we're iterating over.
4665 * @sd: variable holding the value of the power_savings_sd
4667 * @flag: The flag to filter the sched_domains to be iterated.
4669 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4670 * set, starting from the lowest sched_domain to the highest.
4672 #define for_each_flag_domain(cpu, sd, flag) \
4673 for (sd = lowest_flag_domain(cpu, flag); \
4674 (sd && (sd->flags & flag)); sd = sd->parent)
4677 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4678 * @ilb_group: group to be checked for semi-idleness
4680 * Returns: 1 if the group is semi-idle. 0 otherwise.
4682 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4683 * and atleast one non-idle CPU. This helper function checks if the given
4684 * sched_group is semi-idle or not.
4686 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4688 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4689 sched_group_cpus(ilb_group));
4692 * A sched_group is semi-idle when it has atleast one busy cpu
4693 * and atleast one idle cpu.
4695 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4698 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4704 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4705 * @cpu: The cpu which is nominating a new idle_load_balancer.
4707 * Returns: Returns the id of the idle load balancer if it exists,
4708 * Else, returns >= nr_cpu_ids.
4710 * This algorithm picks the idle load balancer such that it belongs to a
4711 * semi-idle powersavings sched_domain. The idea is to try and avoid
4712 * completely idle packages/cores just for the purpose of idle load balancing
4713 * when there are other idle cpu's which are better suited for that job.
4715 static int find_new_ilb(int cpu)
4717 struct sched_domain *sd;
4718 struct sched_group *ilb_group;
4721 * Have idle load balancer selection from semi-idle packages only
4722 * when power-aware load balancing is enabled
4724 if (!(sched_smt_power_savings || sched_mc_power_savings))
4728 * Optimize for the case when we have no idle CPUs or only one
4729 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4731 if (cpumask_weight(nohz.cpu_mask) < 2)
4734 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4735 ilb_group = sd->groups;
4738 if (is_semi_idle_group(ilb_group))
4739 return cpumask_first(nohz.ilb_grp_nohz_mask);
4741 ilb_group = ilb_group->next;
4743 } while (ilb_group != sd->groups);
4747 return cpumask_first(nohz.cpu_mask);
4749 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4750 static inline int find_new_ilb(int call_cpu)
4752 return cpumask_first(nohz.cpu_mask);
4757 * This routine will try to nominate the ilb (idle load balancing)
4758 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4759 * load balancing on behalf of all those cpus. If all the cpus in the system
4760 * go into this tickless mode, then there will be no ilb owner (as there is
4761 * no need for one) and all the cpus will sleep till the next wakeup event
4764 * For the ilb owner, tick is not stopped. And this tick will be used
4765 * for idle load balancing. ilb owner will still be part of
4768 * While stopping the tick, this cpu will become the ilb owner if there
4769 * is no other owner. And will be the owner till that cpu becomes busy
4770 * or if all cpus in the system stop their ticks at which point
4771 * there is no need for ilb owner.
4773 * When the ilb owner becomes busy, it nominates another owner, during the
4774 * next busy scheduler_tick()
4776 int select_nohz_load_balancer(int stop_tick)
4778 int cpu = smp_processor_id();
4781 cpu_rq(cpu)->in_nohz_recently = 1;
4783 if (!cpu_active(cpu)) {
4784 if (atomic_read(&nohz.load_balancer) != cpu)
4788 * If we are going offline and still the leader,
4791 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4797 cpumask_set_cpu(cpu, nohz.cpu_mask);
4799 /* time for ilb owner also to sleep */
4800 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4801 if (atomic_read(&nohz.load_balancer) == cpu)
4802 atomic_set(&nohz.load_balancer, -1);
4806 if (atomic_read(&nohz.load_balancer) == -1) {
4807 /* make me the ilb owner */
4808 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4810 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4813 if (!(sched_smt_power_savings ||
4814 sched_mc_power_savings))
4817 * Check to see if there is a more power-efficient
4820 new_ilb = find_new_ilb(cpu);
4821 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4822 atomic_set(&nohz.load_balancer, -1);
4823 resched_cpu(new_ilb);
4829 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4832 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4834 if (atomic_read(&nohz.load_balancer) == cpu)
4835 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4842 static DEFINE_SPINLOCK(balancing);
4845 * It checks each scheduling domain to see if it is due to be balanced,
4846 * and initiates a balancing operation if so.
4848 * Balancing parameters are set up in arch_init_sched_domains.
4850 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4853 struct rq *rq = cpu_rq(cpu);
4854 unsigned long interval;
4855 struct sched_domain *sd;
4856 /* Earliest time when we have to do rebalance again */
4857 unsigned long next_balance = jiffies + 60*HZ;
4858 int update_next_balance = 0;
4861 for_each_domain(cpu, sd) {
4862 if (!(sd->flags & SD_LOAD_BALANCE))
4865 interval = sd->balance_interval;
4866 if (idle != CPU_IDLE)
4867 interval *= sd->busy_factor;
4869 /* scale ms to jiffies */
4870 interval = msecs_to_jiffies(interval);
4871 if (unlikely(!interval))
4873 if (interval > HZ*NR_CPUS/10)
4874 interval = HZ*NR_CPUS/10;
4876 need_serialize = sd->flags & SD_SERIALIZE;
4878 if (need_serialize) {
4879 if (!spin_trylock(&balancing))
4883 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4884 if (load_balance(cpu, rq, sd, idle, &balance)) {
4886 * We've pulled tasks over so either we're no
4887 * longer idle, or one of our SMT siblings is
4890 idle = CPU_NOT_IDLE;
4892 sd->last_balance = jiffies;
4895 spin_unlock(&balancing);
4897 if (time_after(next_balance, sd->last_balance + interval)) {
4898 next_balance = sd->last_balance + interval;
4899 update_next_balance = 1;
4903 * Stop the load balance at this level. There is another
4904 * CPU in our sched group which is doing load balancing more
4912 * next_balance will be updated only when there is a need.
4913 * When the cpu is attached to null domain for ex, it will not be
4916 if (likely(update_next_balance))
4917 rq->next_balance = next_balance;
4921 * run_rebalance_domains is triggered when needed from the scheduler tick.
4922 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4923 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4925 static void run_rebalance_domains(struct softirq_action *h)
4927 int this_cpu = smp_processor_id();
4928 struct rq *this_rq = cpu_rq(this_cpu);
4929 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4930 CPU_IDLE : CPU_NOT_IDLE;
4932 rebalance_domains(this_cpu, idle);
4936 * If this cpu is the owner for idle load balancing, then do the
4937 * balancing on behalf of the other idle cpus whose ticks are
4940 if (this_rq->idle_at_tick &&
4941 atomic_read(&nohz.load_balancer) == this_cpu) {
4945 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4946 if (balance_cpu == this_cpu)
4950 * If this cpu gets work to do, stop the load balancing
4951 * work being done for other cpus. Next load
4952 * balancing owner will pick it up.
4957 rebalance_domains(balance_cpu, CPU_IDLE);
4959 rq = cpu_rq(balance_cpu);
4960 if (time_after(this_rq->next_balance, rq->next_balance))
4961 this_rq->next_balance = rq->next_balance;
4967 static inline int on_null_domain(int cpu)
4969 return !rcu_dereference(cpu_rq(cpu)->sd);
4973 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4975 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4976 * idle load balancing owner or decide to stop the periodic load balancing,
4977 * if the whole system is idle.
4979 static inline void trigger_load_balance(struct rq *rq, int cpu)
4983 * If we were in the nohz mode recently and busy at the current
4984 * scheduler tick, then check if we need to nominate new idle
4987 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4988 rq->in_nohz_recently = 0;
4990 if (atomic_read(&nohz.load_balancer) == cpu) {
4991 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4992 atomic_set(&nohz.load_balancer, -1);
4995 if (atomic_read(&nohz.load_balancer) == -1) {
4996 int ilb = find_new_ilb(cpu);
4998 if (ilb < nr_cpu_ids)
5004 * If this cpu is idle and doing idle load balancing for all the
5005 * cpus with ticks stopped, is it time for that to stop?
5007 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
5008 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
5014 * If this cpu is idle and the idle load balancing is done by
5015 * someone else, then no need raise the SCHED_SOFTIRQ
5017 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
5018 cpumask_test_cpu(cpu, nohz.cpu_mask))
5021 /* Don't need to rebalance while attached to NULL domain */
5022 if (time_after_eq(jiffies, rq->next_balance) &&
5023 likely(!on_null_domain(cpu)))
5024 raise_softirq(SCHED_SOFTIRQ);
5027 #else /* CONFIG_SMP */
5030 * on UP we do not need to balance between CPUs:
5032 static inline void idle_balance(int cpu, struct rq *rq)
5038 DEFINE_PER_CPU(struct kernel_stat, kstat);
5040 EXPORT_PER_CPU_SYMBOL(kstat);
5043 * Return any ns on the sched_clock that have not yet been accounted in
5044 * @p in case that task is currently running.
5046 * Called with task_rq_lock() held on @rq.
5048 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
5052 if (task_current(rq, p)) {
5053 update_rq_clock(rq);
5054 ns = rq->clock - p->se.exec_start;
5062 unsigned long long task_delta_exec(struct task_struct *p)
5064 unsigned long flags;
5068 rq = task_rq_lock(p, &flags);
5069 ns = do_task_delta_exec(p, rq);
5070 task_rq_unlock(rq, &flags);
5076 * Return accounted runtime for the task.
5077 * In case the task is currently running, return the runtime plus current's
5078 * pending runtime that have not been accounted yet.
5080 unsigned long long task_sched_runtime(struct task_struct *p)
5082 unsigned long flags;
5086 rq = task_rq_lock(p, &flags);
5087 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
5088 task_rq_unlock(rq, &flags);
5094 * Return sum_exec_runtime for the thread group.
5095 * In case the task is currently running, return the sum plus current's
5096 * pending runtime that have not been accounted yet.
5098 * Note that the thread group might have other running tasks as well,
5099 * so the return value not includes other pending runtime that other
5100 * running tasks might have.
5102 unsigned long long thread_group_sched_runtime(struct task_struct *p)
5104 struct task_cputime totals;
5105 unsigned long flags;
5109 rq = task_rq_lock(p, &flags);
5110 thread_group_cputime(p, &totals);
5111 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
5112 task_rq_unlock(rq, &flags);
5118 * Account user cpu time to a process.
5119 * @p: the process that the cpu time gets accounted to
5120 * @cputime: the cpu time spent in user space since the last update
5121 * @cputime_scaled: cputime scaled by cpu frequency
5123 void account_user_time(struct task_struct *p, cputime_t cputime,
5124 cputime_t cputime_scaled)
5126 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5129 /* Add user time to process. */
5130 p->utime = cputime_add(p->utime, cputime);
5131 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5132 account_group_user_time(p, cputime);
5134 /* Add user time to cpustat. */
5135 tmp = cputime_to_cputime64(cputime);
5136 if (TASK_NICE(p) > 0)
5137 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5139 cpustat->user = cputime64_add(cpustat->user, tmp);
5141 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5142 /* Account for user time used */
5143 acct_update_integrals(p);
5147 * Account guest cpu time to a process.
5148 * @p: the process that the cpu time gets accounted to
5149 * @cputime: the cpu time spent in virtual machine since the last update
5150 * @cputime_scaled: cputime scaled by cpu frequency
5152 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5153 cputime_t cputime_scaled)
5156 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5158 tmp = cputime_to_cputime64(cputime);
5160 /* Add guest time to process. */
5161 p->utime = cputime_add(p->utime, cputime);
5162 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5163 account_group_user_time(p, cputime);
5164 p->gtime = cputime_add(p->gtime, cputime);
5166 /* Add guest time to cpustat. */
5167 cpustat->user = cputime64_add(cpustat->user, tmp);
5168 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5172 * Account system cpu time to a process.
5173 * @p: the process that the cpu time gets accounted to
5174 * @hardirq_offset: the offset to subtract from hardirq_count()
5175 * @cputime: the cpu time spent in kernel space since the last update
5176 * @cputime_scaled: cputime scaled by cpu frequency
5178 void account_system_time(struct task_struct *p, int hardirq_offset,
5179 cputime_t cputime, cputime_t cputime_scaled)
5181 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5184 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5185 account_guest_time(p, cputime, cputime_scaled);
5189 /* Add system time to process. */
5190 p->stime = cputime_add(p->stime, cputime);
5191 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5192 account_group_system_time(p, cputime);
5194 /* Add system time to cpustat. */
5195 tmp = cputime_to_cputime64(cputime);
5196 if (hardirq_count() - hardirq_offset)
5197 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5198 else if (softirq_count())
5199 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5201 cpustat->system = cputime64_add(cpustat->system, tmp);
5203 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5205 /* Account for system time used */
5206 acct_update_integrals(p);
5210 * Account for involuntary wait time.
5211 * @steal: the cpu time spent in involuntary wait
5213 void account_steal_time(cputime_t cputime)
5215 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5216 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5218 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5222 * Account for idle time.
5223 * @cputime: the cpu time spent in idle wait
5225 void account_idle_time(cputime_t cputime)
5227 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5228 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5229 struct rq *rq = this_rq();
5231 if (atomic_read(&rq->nr_iowait) > 0)
5232 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5234 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5237 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5240 * Account a single tick of cpu time.
5241 * @p: the process that the cpu time gets accounted to
5242 * @user_tick: indicates if the tick is a user or a system tick
5244 void account_process_tick(struct task_struct *p, int user_tick)
5246 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5247 struct rq *rq = this_rq();
5250 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5251 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5252 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5255 account_idle_time(cputime_one_jiffy);
5259 * Account multiple ticks of steal time.
5260 * @p: the process from which the cpu time has been stolen
5261 * @ticks: number of stolen ticks
5263 void account_steal_ticks(unsigned long ticks)
5265 account_steal_time(jiffies_to_cputime(ticks));
5269 * Account multiple ticks of idle time.
5270 * @ticks: number of stolen ticks
5272 void account_idle_ticks(unsigned long ticks)
5274 account_idle_time(jiffies_to_cputime(ticks));
5280 * Use precise platform statistics if available:
5282 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5283 cputime_t task_utime(struct task_struct *p)
5288 cputime_t task_stime(struct task_struct *p)
5293 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5295 struct task_cputime cputime;
5297 thread_group_cputime(p, &cputime);
5299 *ut = cputime.utime;
5300 *st = cputime.stime;
5304 #ifndef nsecs_to_cputime
5305 # define nsecs_to_cputime(__nsecs) \
5306 msecs_to_cputime(div_u64((__nsecs), NSEC_PER_MSEC))
5309 cputime_t task_utime(struct task_struct *p)
5311 cputime_t utime = p->utime, total = utime + p->stime;
5315 * Use CFS's precise accounting:
5317 temp = (u64)nsecs_to_cputime(p->se.sum_exec_runtime);
5321 do_div(temp, total);
5323 utime = (cputime_t)temp;
5325 p->prev_utime = max(p->prev_utime, utime);
5326 return p->prev_utime;
5329 cputime_t task_stime(struct task_struct *p)
5334 * Use CFS's precise accounting. (we subtract utime from
5335 * the total, to make sure the total observed by userspace
5336 * grows monotonically - apps rely on that):
5338 stime = nsecs_to_cputime(p->se.sum_exec_runtime) - task_utime(p);
5341 p->prev_stime = max(p->prev_stime, stime);
5343 return p->prev_stime;
5347 * Must be called with siglock held.
5349 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5351 struct signal_struct *sig = p->signal;
5352 struct task_cputime cputime;
5353 cputime_t rtime, utime, total;
5355 thread_group_cputime(p, &cputime);
5357 total = cputime_add(cputime.utime, cputime.stime);
5358 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5363 temp = (u64)(rtime * cputime.utime);
5364 do_div(temp, total);
5365 utime = (cputime_t)temp;
5369 sig->prev_utime = max(sig->prev_utime, utime);
5370 sig->prev_stime = max(sig->prev_stime,
5371 cputime_sub(rtime, sig->prev_utime));
5373 *ut = sig->prev_utime;
5374 *st = sig->prev_stime;
5378 inline cputime_t task_gtime(struct task_struct *p)
5384 * This function gets called by the timer code, with HZ frequency.
5385 * We call it with interrupts disabled.
5387 * It also gets called by the fork code, when changing the parent's
5390 void scheduler_tick(void)
5392 int cpu = smp_processor_id();
5393 struct rq *rq = cpu_rq(cpu);
5394 struct task_struct *curr = rq->curr;
5398 spin_lock(&rq->lock);
5399 update_rq_clock(rq);
5400 update_cpu_load(rq);
5401 curr->sched_class->task_tick(rq, curr, 0);
5402 spin_unlock(&rq->lock);
5404 perf_event_task_tick(curr, cpu);
5407 rq->idle_at_tick = idle_cpu(cpu);
5408 trigger_load_balance(rq, cpu);
5412 notrace unsigned long get_parent_ip(unsigned long addr)
5414 if (in_lock_functions(addr)) {
5415 addr = CALLER_ADDR2;
5416 if (in_lock_functions(addr))
5417 addr = CALLER_ADDR3;
5422 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5423 defined(CONFIG_PREEMPT_TRACER))
5425 void __kprobes add_preempt_count(int val)
5427 #ifdef CONFIG_DEBUG_PREEMPT
5431 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5434 preempt_count() += val;
5435 #ifdef CONFIG_DEBUG_PREEMPT
5437 * Spinlock count overflowing soon?
5439 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5442 if (preempt_count() == val)
5443 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5445 EXPORT_SYMBOL(add_preempt_count);
5447 void __kprobes sub_preempt_count(int val)
5449 #ifdef CONFIG_DEBUG_PREEMPT
5453 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5456 * Is the spinlock portion underflowing?
5458 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5459 !(preempt_count() & PREEMPT_MASK)))
5463 if (preempt_count() == val)
5464 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5465 preempt_count() -= val;
5467 EXPORT_SYMBOL(sub_preempt_count);
5472 * Print scheduling while atomic bug:
5474 static noinline void __schedule_bug(struct task_struct *prev)
5476 struct pt_regs *regs = get_irq_regs();
5478 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5479 prev->comm, prev->pid, preempt_count());
5481 debug_show_held_locks(prev);
5483 if (irqs_disabled())
5484 print_irqtrace_events(prev);
5493 * Various schedule()-time debugging checks and statistics:
5495 static inline void schedule_debug(struct task_struct *prev)
5498 * Test if we are atomic. Since do_exit() needs to call into
5499 * schedule() atomically, we ignore that path for now.
5500 * Otherwise, whine if we are scheduling when we should not be.
5502 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5503 __schedule_bug(prev);
5505 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5507 schedstat_inc(this_rq(), sched_count);
5508 #ifdef CONFIG_SCHEDSTATS
5509 if (unlikely(prev->lock_depth >= 0)) {
5510 schedstat_inc(this_rq(), bkl_count);
5511 schedstat_inc(prev, sched_info.bkl_count);
5516 static void put_prev_task(struct rq *rq, struct task_struct *p)
5518 u64 runtime = p->se.sum_exec_runtime - p->se.prev_sum_exec_runtime;
5520 update_avg(&p->se.avg_running, runtime);
5522 if (p->state == TASK_RUNNING) {
5524 * In order to avoid avg_overlap growing stale when we are
5525 * indeed overlapping and hence not getting put to sleep, grow
5526 * the avg_overlap on preemption.
5528 * We use the average preemption runtime because that
5529 * correlates to the amount of cache footprint a task can
5532 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5533 update_avg(&p->se.avg_overlap, runtime);
5535 update_avg(&p->se.avg_running, 0);
5537 p->sched_class->put_prev_task(rq, p);
5541 * Pick up the highest-prio task:
5543 static inline struct task_struct *
5544 pick_next_task(struct rq *rq)
5546 const struct sched_class *class;
5547 struct task_struct *p;
5550 * Optimization: we know that if all tasks are in
5551 * the fair class we can call that function directly:
5553 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5554 p = fair_sched_class.pick_next_task(rq);
5559 class = sched_class_highest;
5561 p = class->pick_next_task(rq);
5565 * Will never be NULL as the idle class always
5566 * returns a non-NULL p:
5568 class = class->next;
5573 * schedule() is the main scheduler function.
5575 asmlinkage void __sched schedule(void)
5577 struct task_struct *prev, *next;
5578 unsigned long *switch_count;
5584 cpu = smp_processor_id();
5588 switch_count = &prev->nivcsw;
5590 release_kernel_lock(prev);
5591 need_resched_nonpreemptible:
5593 schedule_debug(prev);
5595 if (sched_feat(HRTICK))
5598 spin_lock_irq(&rq->lock);
5599 update_rq_clock(rq);
5600 clear_tsk_need_resched(prev);
5602 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5603 if (unlikely(signal_pending_state(prev->state, prev)))
5604 prev->state = TASK_RUNNING;
5606 deactivate_task(rq, prev, 1);
5607 switch_count = &prev->nvcsw;
5610 pre_schedule(rq, prev);
5612 if (unlikely(!rq->nr_running))
5613 idle_balance(cpu, rq);
5615 put_prev_task(rq, prev);
5616 next = pick_next_task(rq);
5618 if (likely(prev != next)) {
5619 sched_info_switch(prev, next);
5620 perf_event_task_sched_out(prev, next, cpu);
5626 context_switch(rq, prev, next); /* unlocks the rq */
5628 * the context switch might have flipped the stack from under
5629 * us, hence refresh the local variables.
5631 cpu = smp_processor_id();
5634 spin_unlock_irq(&rq->lock);
5638 if (unlikely(reacquire_kernel_lock(current) < 0))
5639 goto need_resched_nonpreemptible;
5641 preempt_enable_no_resched();
5645 EXPORT_SYMBOL(schedule);
5649 * Look out! "owner" is an entirely speculative pointer
5650 * access and not reliable.
5652 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5657 if (!sched_feat(OWNER_SPIN))
5660 #ifdef CONFIG_DEBUG_PAGEALLOC
5662 * Need to access the cpu field knowing that
5663 * DEBUG_PAGEALLOC could have unmapped it if
5664 * the mutex owner just released it and exited.
5666 if (probe_kernel_address(&owner->cpu, cpu))
5673 * Even if the access succeeded (likely case),
5674 * the cpu field may no longer be valid.
5676 if (cpu >= nr_cpumask_bits)
5680 * We need to validate that we can do a
5681 * get_cpu() and that we have the percpu area.
5683 if (!cpu_online(cpu))
5690 * Owner changed, break to re-assess state.
5692 if (lock->owner != owner)
5696 * Is that owner really running on that cpu?
5698 if (task_thread_info(rq->curr) != owner || need_resched())
5708 #ifdef CONFIG_PREEMPT
5710 * this is the entry point to schedule() from in-kernel preemption
5711 * off of preempt_enable. Kernel preemptions off return from interrupt
5712 * occur there and call schedule directly.
5714 asmlinkage void __sched preempt_schedule(void)
5716 struct thread_info *ti = current_thread_info();
5719 * If there is a non-zero preempt_count or interrupts are disabled,
5720 * we do not want to preempt the current task. Just return..
5722 if (likely(ti->preempt_count || irqs_disabled()))
5726 add_preempt_count(PREEMPT_ACTIVE);
5728 sub_preempt_count(PREEMPT_ACTIVE);
5731 * Check again in case we missed a preemption opportunity
5732 * between schedule and now.
5735 } while (need_resched());
5737 EXPORT_SYMBOL(preempt_schedule);
5740 * this is the entry point to schedule() from kernel preemption
5741 * off of irq context.
5742 * Note, that this is called and return with irqs disabled. This will
5743 * protect us against recursive calling from irq.
5745 asmlinkage void __sched preempt_schedule_irq(void)
5747 struct thread_info *ti = current_thread_info();
5749 /* Catch callers which need to be fixed */
5750 BUG_ON(ti->preempt_count || !irqs_disabled());
5753 add_preempt_count(PREEMPT_ACTIVE);
5756 local_irq_disable();
5757 sub_preempt_count(PREEMPT_ACTIVE);
5760 * Check again in case we missed a preemption opportunity
5761 * between schedule and now.
5764 } while (need_resched());
5767 #endif /* CONFIG_PREEMPT */
5769 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5772 return try_to_wake_up(curr->private, mode, wake_flags);
5774 EXPORT_SYMBOL(default_wake_function);
5777 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5778 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5779 * number) then we wake all the non-exclusive tasks and one exclusive task.
5781 * There are circumstances in which we can try to wake a task which has already
5782 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5783 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5785 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5786 int nr_exclusive, int wake_flags, void *key)
5788 wait_queue_t *curr, *next;
5790 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5791 unsigned flags = curr->flags;
5793 if (curr->func(curr, mode, wake_flags, key) &&
5794 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5800 * __wake_up - wake up threads blocked on a waitqueue.
5802 * @mode: which threads
5803 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5804 * @key: is directly passed to the wakeup function
5806 * It may be assumed that this function implies a write memory barrier before
5807 * changing the task state if and only if any tasks are woken up.
5809 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5810 int nr_exclusive, void *key)
5812 unsigned long flags;
5814 spin_lock_irqsave(&q->lock, flags);
5815 __wake_up_common(q, mode, nr_exclusive, 0, key);
5816 spin_unlock_irqrestore(&q->lock, flags);
5818 EXPORT_SYMBOL(__wake_up);
5821 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5823 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5825 __wake_up_common(q, mode, 1, 0, NULL);
5828 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5830 __wake_up_common(q, mode, 1, 0, key);
5834 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5836 * @mode: which threads
5837 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5838 * @key: opaque value to be passed to wakeup targets
5840 * The sync wakeup differs that the waker knows that it will schedule
5841 * away soon, so while the target thread will be woken up, it will not
5842 * be migrated to another CPU - ie. the two threads are 'synchronized'
5843 * with each other. This can prevent needless bouncing between CPUs.
5845 * On UP it can prevent extra preemption.
5847 * It may be assumed that this function implies a write memory barrier before
5848 * changing the task state if and only if any tasks are woken up.
5850 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5851 int nr_exclusive, void *key)
5853 unsigned long flags;
5854 int wake_flags = WF_SYNC;
5859 if (unlikely(!nr_exclusive))
5862 spin_lock_irqsave(&q->lock, flags);
5863 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5864 spin_unlock_irqrestore(&q->lock, flags);
5866 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5869 * __wake_up_sync - see __wake_up_sync_key()
5871 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5873 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5875 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5878 * complete: - signals a single thread waiting on this completion
5879 * @x: holds the state of this particular completion
5881 * This will wake up a single thread waiting on this completion. Threads will be
5882 * awakened in the same order in which they were queued.
5884 * See also complete_all(), wait_for_completion() and related routines.
5886 * It may be assumed that this function implies a write memory barrier before
5887 * changing the task state if and only if any tasks are woken up.
5889 void complete(struct completion *x)
5891 unsigned long flags;
5893 spin_lock_irqsave(&x->wait.lock, flags);
5895 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5896 spin_unlock_irqrestore(&x->wait.lock, flags);
5898 EXPORT_SYMBOL(complete);
5901 * complete_all: - signals all threads waiting on this completion
5902 * @x: holds the state of this particular completion
5904 * This will wake up all threads waiting on this particular completion event.
5906 * It may be assumed that this function implies a write memory barrier before
5907 * changing the task state if and only if any tasks are woken up.
5909 void complete_all(struct completion *x)
5911 unsigned long flags;
5913 spin_lock_irqsave(&x->wait.lock, flags);
5914 x->done += UINT_MAX/2;
5915 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5916 spin_unlock_irqrestore(&x->wait.lock, flags);
5918 EXPORT_SYMBOL(complete_all);
5920 static inline long __sched
5921 do_wait_for_common(struct completion *x, long timeout, int state)
5924 DECLARE_WAITQUEUE(wait, current);
5926 wait.flags |= WQ_FLAG_EXCLUSIVE;
5927 __add_wait_queue_tail(&x->wait, &wait);
5929 if (signal_pending_state(state, current)) {
5930 timeout = -ERESTARTSYS;
5933 __set_current_state(state);
5934 spin_unlock_irq(&x->wait.lock);
5935 timeout = schedule_timeout(timeout);
5936 spin_lock_irq(&x->wait.lock);
5937 } while (!x->done && timeout);
5938 __remove_wait_queue(&x->wait, &wait);
5943 return timeout ?: 1;
5947 wait_for_common(struct completion *x, long timeout, int state)
5951 spin_lock_irq(&x->wait.lock);
5952 timeout = do_wait_for_common(x, timeout, state);
5953 spin_unlock_irq(&x->wait.lock);
5958 * wait_for_completion: - waits for completion of a task
5959 * @x: holds the state of this particular completion
5961 * This waits to be signaled for completion of a specific task. It is NOT
5962 * interruptible and there is no timeout.
5964 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5965 * and interrupt capability. Also see complete().
5967 void __sched wait_for_completion(struct completion *x)
5969 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5971 EXPORT_SYMBOL(wait_for_completion);
5974 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5975 * @x: holds the state of this particular completion
5976 * @timeout: timeout value in jiffies
5978 * This waits for either a completion of a specific task to be signaled or for a
5979 * specified timeout to expire. The timeout is in jiffies. It is not
5982 unsigned long __sched
5983 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5985 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5987 EXPORT_SYMBOL(wait_for_completion_timeout);
5990 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5991 * @x: holds the state of this particular completion
5993 * This waits for completion of a specific task to be signaled. It is
5996 int __sched wait_for_completion_interruptible(struct completion *x)
5998 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5999 if (t == -ERESTARTSYS)
6003 EXPORT_SYMBOL(wait_for_completion_interruptible);
6006 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
6007 * @x: holds the state of this particular completion
6008 * @timeout: timeout value in jiffies
6010 * This waits for either a completion of a specific task to be signaled or for a
6011 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
6013 unsigned long __sched
6014 wait_for_completion_interruptible_timeout(struct completion *x,
6015 unsigned long timeout)
6017 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
6019 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
6022 * wait_for_completion_killable: - waits for completion of a task (killable)
6023 * @x: holds the state of this particular completion
6025 * This waits to be signaled for completion of a specific task. It can be
6026 * interrupted by a kill signal.
6028 int __sched wait_for_completion_killable(struct completion *x)
6030 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
6031 if (t == -ERESTARTSYS)
6035 EXPORT_SYMBOL(wait_for_completion_killable);
6038 * try_wait_for_completion - try to decrement a completion without blocking
6039 * @x: completion structure
6041 * Returns: 0 if a decrement cannot be done without blocking
6042 * 1 if a decrement succeeded.
6044 * If a completion is being used as a counting completion,
6045 * attempt to decrement the counter without blocking. This
6046 * enables us to avoid waiting if the resource the completion
6047 * is protecting is not available.
6049 bool try_wait_for_completion(struct completion *x)
6051 unsigned long flags;
6054 spin_lock_irqsave(&x->wait.lock, flags);
6059 spin_unlock_irqrestore(&x->wait.lock, flags);
6062 EXPORT_SYMBOL(try_wait_for_completion);
6065 * completion_done - Test to see if a completion has any waiters
6066 * @x: completion structure
6068 * Returns: 0 if there are waiters (wait_for_completion() in progress)
6069 * 1 if there are no waiters.
6072 bool completion_done(struct completion *x)
6074 unsigned long flags;
6077 spin_lock_irqsave(&x->wait.lock, flags);
6080 spin_unlock_irqrestore(&x->wait.lock, flags);
6083 EXPORT_SYMBOL(completion_done);
6086 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
6088 unsigned long flags;
6091 init_waitqueue_entry(&wait, current);
6093 __set_current_state(state);
6095 spin_lock_irqsave(&q->lock, flags);
6096 __add_wait_queue(q, &wait);
6097 spin_unlock(&q->lock);
6098 timeout = schedule_timeout(timeout);
6099 spin_lock_irq(&q->lock);
6100 __remove_wait_queue(q, &wait);
6101 spin_unlock_irqrestore(&q->lock, flags);
6106 void __sched interruptible_sleep_on(wait_queue_head_t *q)
6108 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6110 EXPORT_SYMBOL(interruptible_sleep_on);
6113 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
6115 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
6117 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
6119 void __sched sleep_on(wait_queue_head_t *q)
6121 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6123 EXPORT_SYMBOL(sleep_on);
6125 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
6127 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
6129 EXPORT_SYMBOL(sleep_on_timeout);
6131 #ifdef CONFIG_RT_MUTEXES
6134 * rt_mutex_setprio - set the current priority of a task
6136 * @prio: prio value (kernel-internal form)
6138 * This function changes the 'effective' priority of a task. It does
6139 * not touch ->normal_prio like __setscheduler().
6141 * Used by the rt_mutex code to implement priority inheritance logic.
6143 void rt_mutex_setprio(struct task_struct *p, int prio)
6145 unsigned long flags;
6146 int oldprio, on_rq, running;
6148 const struct sched_class *prev_class;
6150 BUG_ON(prio < 0 || prio > MAX_PRIO);
6152 rq = task_rq_lock(p, &flags);
6153 update_rq_clock(rq);
6156 prev_class = p->sched_class;
6157 on_rq = p->se.on_rq;
6158 running = task_current(rq, p);
6160 dequeue_task(rq, p, 0);
6162 p->sched_class->put_prev_task(rq, p);
6165 p->sched_class = &rt_sched_class;
6167 p->sched_class = &fair_sched_class;
6172 p->sched_class->set_curr_task(rq);
6174 enqueue_task(rq, p, 0, oldprio < prio);
6176 check_class_changed(rq, p, prev_class, oldprio, running);
6178 task_rq_unlock(rq, &flags);
6183 void set_user_nice(struct task_struct *p, long nice)
6185 int old_prio, delta, on_rq;
6186 unsigned long flags;
6189 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6192 * We have to be careful, if called from sys_setpriority(),
6193 * the task might be in the middle of scheduling on another CPU.
6195 rq = task_rq_lock(p, &flags);
6196 update_rq_clock(rq);
6198 * The RT priorities are set via sched_setscheduler(), but we still
6199 * allow the 'normal' nice value to be set - but as expected
6200 * it wont have any effect on scheduling until the task is
6201 * SCHED_FIFO/SCHED_RR:
6203 if (task_has_rt_policy(p)) {
6204 p->static_prio = NICE_TO_PRIO(nice);
6207 on_rq = p->se.on_rq;
6209 dequeue_task(rq, p, 0);
6211 p->static_prio = NICE_TO_PRIO(nice);
6214 p->prio = effective_prio(p);
6215 delta = p->prio - old_prio;
6218 enqueue_task(rq, p, 0, false);
6220 * If the task increased its priority or is running and
6221 * lowered its priority, then reschedule its CPU:
6223 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6224 resched_task(rq->curr);
6227 task_rq_unlock(rq, &flags);
6229 EXPORT_SYMBOL(set_user_nice);
6232 * can_nice - check if a task can reduce its nice value
6236 int can_nice(const struct task_struct *p, const int nice)
6238 /* convert nice value [19,-20] to rlimit style value [1,40] */
6239 int nice_rlim = 20 - nice;
6241 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6242 capable(CAP_SYS_NICE));
6245 #ifdef __ARCH_WANT_SYS_NICE
6248 * sys_nice - change the priority of the current process.
6249 * @increment: priority increment
6251 * sys_setpriority is a more generic, but much slower function that
6252 * does similar things.
6254 SYSCALL_DEFINE1(nice, int, increment)
6259 * Setpriority might change our priority at the same moment.
6260 * We don't have to worry. Conceptually one call occurs first
6261 * and we have a single winner.
6263 if (increment < -40)
6268 nice = TASK_NICE(current) + increment;
6274 if (increment < 0 && !can_nice(current, nice))
6277 retval = security_task_setnice(current, nice);
6281 set_user_nice(current, nice);
6288 * task_prio - return the priority value of a given task.
6289 * @p: the task in question.
6291 * This is the priority value as seen by users in /proc.
6292 * RT tasks are offset by -200. Normal tasks are centered
6293 * around 0, value goes from -16 to +15.
6295 int task_prio(const struct task_struct *p)
6297 return p->prio - MAX_RT_PRIO;
6301 * task_nice - return the nice value of a given task.
6302 * @p: the task in question.
6304 int task_nice(const struct task_struct *p)
6306 return TASK_NICE(p);
6308 EXPORT_SYMBOL(task_nice);
6311 * idle_cpu - is a given cpu idle currently?
6312 * @cpu: the processor in question.
6314 int idle_cpu(int cpu)
6316 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6320 * idle_task - return the idle task for a given cpu.
6321 * @cpu: the processor in question.
6323 struct task_struct *idle_task(int cpu)
6325 return cpu_rq(cpu)->idle;
6329 * find_process_by_pid - find a process with a matching PID value.
6330 * @pid: the pid in question.
6332 static struct task_struct *find_process_by_pid(pid_t pid)
6334 return pid ? find_task_by_vpid(pid) : current;
6337 /* Actually do priority change: must hold rq lock. */
6339 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6341 BUG_ON(p->se.on_rq);
6344 switch (p->policy) {
6348 p->sched_class = &fair_sched_class;
6352 p->sched_class = &rt_sched_class;
6356 p->rt_priority = prio;
6357 p->normal_prio = normal_prio(p);
6358 /* we are holding p->pi_lock already */
6359 p->prio = rt_mutex_getprio(p);
6364 * check the target process has a UID that matches the current process's
6366 static bool check_same_owner(struct task_struct *p)
6368 const struct cred *cred = current_cred(), *pcred;
6372 pcred = __task_cred(p);
6373 match = (cred->euid == pcred->euid ||
6374 cred->euid == pcred->uid);
6379 static int __sched_setscheduler(struct task_struct *p, int policy,
6380 struct sched_param *param, bool user)
6382 int retval, oldprio, oldpolicy = -1, on_rq, running;
6383 unsigned long flags;
6384 const struct sched_class *prev_class;
6388 /* may grab non-irq protected spin_locks */
6389 BUG_ON(in_interrupt());
6391 /* double check policy once rq lock held */
6393 reset_on_fork = p->sched_reset_on_fork;
6394 policy = oldpolicy = p->policy;
6396 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6397 policy &= ~SCHED_RESET_ON_FORK;
6399 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6400 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6401 policy != SCHED_IDLE)
6406 * Valid priorities for SCHED_FIFO and SCHED_RR are
6407 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6408 * SCHED_BATCH and SCHED_IDLE is 0.
6410 if (param->sched_priority < 0 ||
6411 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6412 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6414 if (rt_policy(policy) != (param->sched_priority != 0))
6418 * Allow unprivileged RT tasks to decrease priority:
6420 if (user && !capable(CAP_SYS_NICE)) {
6421 if (rt_policy(policy)) {
6422 unsigned long rlim_rtprio;
6424 if (!lock_task_sighand(p, &flags))
6426 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6427 unlock_task_sighand(p, &flags);
6429 /* can't set/change the rt policy */
6430 if (policy != p->policy && !rlim_rtprio)
6433 /* can't increase priority */
6434 if (param->sched_priority > p->rt_priority &&
6435 param->sched_priority > rlim_rtprio)
6439 * Like positive nice levels, dont allow tasks to
6440 * move out of SCHED_IDLE either:
6442 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6445 /* can't change other user's priorities */
6446 if (!check_same_owner(p))
6449 /* Normal users shall not reset the sched_reset_on_fork flag */
6450 if (p->sched_reset_on_fork && !reset_on_fork)
6455 #ifdef CONFIG_RT_GROUP_SCHED
6457 * Do not allow realtime tasks into groups that have no runtime
6460 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6461 task_group(p)->rt_bandwidth.rt_runtime == 0)
6465 retval = security_task_setscheduler(p, policy, param);
6471 * make sure no PI-waiters arrive (or leave) while we are
6472 * changing the priority of the task:
6474 spin_lock_irqsave(&p->pi_lock, flags);
6476 * To be able to change p->policy safely, the apropriate
6477 * runqueue lock must be held.
6479 rq = __task_rq_lock(p);
6480 /* recheck policy now with rq lock held */
6481 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6482 policy = oldpolicy = -1;
6483 __task_rq_unlock(rq);
6484 spin_unlock_irqrestore(&p->pi_lock, flags);
6487 update_rq_clock(rq);
6488 on_rq = p->se.on_rq;
6489 running = task_current(rq, p);
6491 deactivate_task(rq, p, 0);
6493 p->sched_class->put_prev_task(rq, p);
6495 p->sched_reset_on_fork = reset_on_fork;
6498 prev_class = p->sched_class;
6499 __setscheduler(rq, p, policy, param->sched_priority);
6502 p->sched_class->set_curr_task(rq);
6504 activate_task(rq, p, 0);
6506 check_class_changed(rq, p, prev_class, oldprio, running);
6508 __task_rq_unlock(rq);
6509 spin_unlock_irqrestore(&p->pi_lock, flags);
6511 rt_mutex_adjust_pi(p);
6517 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6518 * @p: the task in question.
6519 * @policy: new policy.
6520 * @param: structure containing the new RT priority.
6522 * NOTE that the task may be already dead.
6524 int sched_setscheduler(struct task_struct *p, int policy,
6525 struct sched_param *param)
6527 return __sched_setscheduler(p, policy, param, true);
6529 EXPORT_SYMBOL_GPL(sched_setscheduler);
6532 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6533 * @p: the task in question.
6534 * @policy: new policy.
6535 * @param: structure containing the new RT priority.
6537 * Just like sched_setscheduler, only don't bother checking if the
6538 * current context has permission. For example, this is needed in
6539 * stop_machine(): we create temporary high priority worker threads,
6540 * but our caller might not have that capability.
6542 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6543 struct sched_param *param)
6545 return __sched_setscheduler(p, policy, param, false);
6549 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6551 struct sched_param lparam;
6552 struct task_struct *p;
6555 if (!param || pid < 0)
6557 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6562 p = find_process_by_pid(pid);
6564 retval = sched_setscheduler(p, policy, &lparam);
6571 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6572 * @pid: the pid in question.
6573 * @policy: new policy.
6574 * @param: structure containing the new RT priority.
6576 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6577 struct sched_param __user *, param)
6579 /* negative values for policy are not valid */
6583 return do_sched_setscheduler(pid, policy, param);
6587 * sys_sched_setparam - set/change the RT priority of a thread
6588 * @pid: the pid in question.
6589 * @param: structure containing the new RT priority.
6591 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6593 return do_sched_setscheduler(pid, -1, param);
6597 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6598 * @pid: the pid in question.
6600 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6602 struct task_struct *p;
6610 p = find_process_by_pid(pid);
6612 retval = security_task_getscheduler(p);
6615 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6622 * sys_sched_getparam - get the RT priority of a thread
6623 * @pid: the pid in question.
6624 * @param: structure containing the RT priority.
6626 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6628 struct sched_param lp;
6629 struct task_struct *p;
6632 if (!param || pid < 0)
6636 p = find_process_by_pid(pid);
6641 retval = security_task_getscheduler(p);
6645 lp.sched_priority = p->rt_priority;
6649 * This one might sleep, we cannot do it with a spinlock held ...
6651 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6660 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6662 cpumask_var_t cpus_allowed, new_mask;
6663 struct task_struct *p;
6669 p = find_process_by_pid(pid);
6676 /* Prevent p going away */
6680 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6684 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6686 goto out_free_cpus_allowed;
6689 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6692 retval = security_task_setscheduler(p, 0, NULL);
6696 cpuset_cpus_allowed(p, cpus_allowed);
6697 cpumask_and(new_mask, in_mask, cpus_allowed);
6699 retval = set_cpus_allowed_ptr(p, new_mask);
6702 cpuset_cpus_allowed(p, cpus_allowed);
6703 if (!cpumask_subset(new_mask, cpus_allowed)) {
6705 * We must have raced with a concurrent cpuset
6706 * update. Just reset the cpus_allowed to the
6707 * cpuset's cpus_allowed
6709 cpumask_copy(new_mask, cpus_allowed);
6714 free_cpumask_var(new_mask);
6715 out_free_cpus_allowed:
6716 free_cpumask_var(cpus_allowed);
6723 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6724 struct cpumask *new_mask)
6726 if (len < cpumask_size())
6727 cpumask_clear(new_mask);
6728 else if (len > cpumask_size())
6729 len = cpumask_size();
6731 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6735 * sys_sched_setaffinity - set the cpu affinity of a process
6736 * @pid: pid of the process
6737 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6738 * @user_mask_ptr: user-space pointer to the new cpu mask
6740 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6741 unsigned long __user *, user_mask_ptr)
6743 cpumask_var_t new_mask;
6746 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6749 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6751 retval = sched_setaffinity(pid, new_mask);
6752 free_cpumask_var(new_mask);
6756 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6758 struct task_struct *p;
6759 unsigned long flags;
6767 p = find_process_by_pid(pid);
6771 retval = security_task_getscheduler(p);
6775 rq = task_rq_lock(p, &flags);
6776 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6777 task_rq_unlock(rq, &flags);
6787 * sys_sched_getaffinity - get the cpu affinity of a process
6788 * @pid: pid of the process
6789 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6790 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6792 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6793 unsigned long __user *, user_mask_ptr)
6798 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
6800 if (len & (sizeof(unsigned long)-1))
6803 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6806 ret = sched_getaffinity(pid, mask);
6808 size_t retlen = min_t(size_t, len, cpumask_size());
6810 if (copy_to_user(user_mask_ptr, mask, retlen))
6815 free_cpumask_var(mask);
6821 * sys_sched_yield - yield the current processor to other threads.
6823 * This function yields the current CPU to other tasks. If there are no
6824 * other threads running on this CPU then this function will return.
6826 SYSCALL_DEFINE0(sched_yield)
6828 struct rq *rq = this_rq_lock();
6830 schedstat_inc(rq, yld_count);
6831 current->sched_class->yield_task(rq);
6834 * Since we are going to call schedule() anyway, there's
6835 * no need to preempt or enable interrupts:
6837 __release(rq->lock);
6838 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6839 _raw_spin_unlock(&rq->lock);
6840 preempt_enable_no_resched();
6847 static inline int should_resched(void)
6849 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6852 static void __cond_resched(void)
6854 add_preempt_count(PREEMPT_ACTIVE);
6856 sub_preempt_count(PREEMPT_ACTIVE);
6859 int __sched _cond_resched(void)
6861 if (should_resched()) {
6867 EXPORT_SYMBOL(_cond_resched);
6870 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6871 * call schedule, and on return reacquire the lock.
6873 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6874 * operations here to prevent schedule() from being called twice (once via
6875 * spin_unlock(), once by hand).
6877 int __cond_resched_lock(spinlock_t *lock)
6879 int resched = should_resched();
6882 lockdep_assert_held(lock);
6884 if (spin_needbreak(lock) || resched) {
6895 EXPORT_SYMBOL(__cond_resched_lock);
6897 int __sched __cond_resched_softirq(void)
6899 BUG_ON(!in_softirq());
6901 if (should_resched()) {
6909 EXPORT_SYMBOL(__cond_resched_softirq);
6912 * yield - yield the current processor to other threads.
6914 * This is a shortcut for kernel-space yielding - it marks the
6915 * thread runnable and calls sys_sched_yield().
6917 void __sched yield(void)
6919 set_current_state(TASK_RUNNING);
6922 EXPORT_SYMBOL(yield);
6925 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6926 * that process accounting knows that this is a task in IO wait state.
6928 void __sched io_schedule(void)
6930 struct rq *rq = raw_rq();
6932 delayacct_blkio_start();
6933 atomic_inc(&rq->nr_iowait);
6934 current->in_iowait = 1;
6936 current->in_iowait = 0;
6937 atomic_dec(&rq->nr_iowait);
6938 delayacct_blkio_end();
6940 EXPORT_SYMBOL(io_schedule);
6942 long __sched io_schedule_timeout(long timeout)
6944 struct rq *rq = raw_rq();
6947 delayacct_blkio_start();
6948 atomic_inc(&rq->nr_iowait);
6949 current->in_iowait = 1;
6950 ret = schedule_timeout(timeout);
6951 current->in_iowait = 0;
6952 atomic_dec(&rq->nr_iowait);
6953 delayacct_blkio_end();
6958 * sys_sched_get_priority_max - return maximum RT priority.
6959 * @policy: scheduling class.
6961 * this syscall returns the maximum rt_priority that can be used
6962 * by a given scheduling class.
6964 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6971 ret = MAX_USER_RT_PRIO-1;
6983 * sys_sched_get_priority_min - return minimum RT priority.
6984 * @policy: scheduling class.
6986 * this syscall returns the minimum rt_priority that can be used
6987 * by a given scheduling class.
6989 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
7007 * sys_sched_rr_get_interval - return the default timeslice of a process.
7008 * @pid: pid of the process.
7009 * @interval: userspace pointer to the timeslice value.
7011 * this syscall writes the default timeslice value of a given process
7012 * into the user-space timespec buffer. A value of '0' means infinity.
7014 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
7015 struct timespec __user *, interval)
7017 struct task_struct *p;
7018 unsigned int time_slice;
7019 unsigned long flags;
7029 p = find_process_by_pid(pid);
7033 retval = security_task_getscheduler(p);
7037 rq = task_rq_lock(p, &flags);
7038 time_slice = p->sched_class->get_rr_interval(rq, p);
7039 task_rq_unlock(rq, &flags);
7042 jiffies_to_timespec(time_slice, &t);
7043 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
7051 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
7053 void sched_show_task(struct task_struct *p)
7055 unsigned long free = 0;
7058 state = p->state ? __ffs(p->state) + 1 : 0;
7059 printk(KERN_INFO "%-13.13s %c", p->comm,
7060 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
7061 #if BITS_PER_LONG == 32
7062 if (state == TASK_RUNNING)
7063 printk(KERN_CONT " running ");
7065 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
7067 if (state == TASK_RUNNING)
7068 printk(KERN_CONT " running task ");
7070 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
7072 #ifdef CONFIG_DEBUG_STACK_USAGE
7073 free = stack_not_used(p);
7075 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
7076 task_pid_nr(p), task_pid_nr(p->real_parent),
7077 (unsigned long)task_thread_info(p)->flags);
7079 show_stack(p, NULL);
7082 void show_state_filter(unsigned long state_filter)
7084 struct task_struct *g, *p;
7086 #if BITS_PER_LONG == 32
7088 " task PC stack pid father\n");
7091 " task PC stack pid father\n");
7093 read_lock(&tasklist_lock);
7094 do_each_thread(g, p) {
7096 * reset the NMI-timeout, listing all files on a slow
7097 * console might take alot of time:
7099 touch_nmi_watchdog();
7100 if (!state_filter || (p->state & state_filter))
7102 } while_each_thread(g, p);
7104 touch_all_softlockup_watchdogs();
7106 #ifdef CONFIG_SCHED_DEBUG
7107 sysrq_sched_debug_show();
7109 read_unlock(&tasklist_lock);
7111 * Only show locks if all tasks are dumped:
7113 if (state_filter == -1)
7114 debug_show_all_locks();
7117 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
7119 idle->sched_class = &idle_sched_class;
7123 * init_idle - set up an idle thread for a given CPU
7124 * @idle: task in question
7125 * @cpu: cpu the idle task belongs to
7127 * NOTE: this function does not set the idle thread's NEED_RESCHED
7128 * flag, to make booting more robust.
7130 void __cpuinit init_idle(struct task_struct *idle, int cpu)
7132 struct rq *rq = cpu_rq(cpu);
7133 unsigned long flags;
7135 spin_lock_irqsave(&rq->lock, flags);
7138 idle->state = TASK_RUNNING;
7139 idle->se.exec_start = sched_clock();
7141 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
7142 __set_task_cpu(idle, cpu);
7144 rq->curr = rq->idle = idle;
7145 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7148 spin_unlock_irqrestore(&rq->lock, flags);
7150 /* Set the preempt count _outside_ the spinlocks! */
7151 #if defined(CONFIG_PREEMPT)
7152 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7154 task_thread_info(idle)->preempt_count = 0;
7157 * The idle tasks have their own, simple scheduling class:
7159 idle->sched_class = &idle_sched_class;
7160 ftrace_graph_init_task(idle);
7164 * In a system that switches off the HZ timer nohz_cpu_mask
7165 * indicates which cpus entered this state. This is used
7166 * in the rcu update to wait only for active cpus. For system
7167 * which do not switch off the HZ timer nohz_cpu_mask should
7168 * always be CPU_BITS_NONE.
7170 cpumask_var_t nohz_cpu_mask;
7173 * Increase the granularity value when there are more CPUs,
7174 * because with more CPUs the 'effective latency' as visible
7175 * to users decreases. But the relationship is not linear,
7176 * so pick a second-best guess by going with the log2 of the
7179 * This idea comes from the SD scheduler of Con Kolivas:
7181 static void update_sysctl(void)
7183 unsigned int cpus = min(num_online_cpus(), 8U);
7184 unsigned int factor = 1 + ilog2(cpus);
7186 #define SET_SYSCTL(name) \
7187 (sysctl_##name = (factor) * normalized_sysctl_##name)
7188 SET_SYSCTL(sched_min_granularity);
7189 SET_SYSCTL(sched_latency);
7190 SET_SYSCTL(sched_wakeup_granularity);
7191 SET_SYSCTL(sched_shares_ratelimit);
7195 static inline void sched_init_granularity(void)
7202 * This is how migration works:
7204 * 1) we queue a struct migration_req structure in the source CPU's
7205 * runqueue and wake up that CPU's migration thread.
7206 * 2) we down() the locked semaphore => thread blocks.
7207 * 3) migration thread wakes up (implicitly it forces the migrated
7208 * thread off the CPU)
7209 * 4) it gets the migration request and checks whether the migrated
7210 * task is still in the wrong runqueue.
7211 * 5) if it's in the wrong runqueue then the migration thread removes
7212 * it and puts it into the right queue.
7213 * 6) migration thread up()s the semaphore.
7214 * 7) we wake up and the migration is done.
7218 * Change a given task's CPU affinity. Migrate the thread to a
7219 * proper CPU and schedule it away if the CPU it's executing on
7220 * is removed from the allowed bitmask.
7222 * NOTE: the caller must have a valid reference to the task, the
7223 * task must not exit() & deallocate itself prematurely. The
7224 * call is not atomic; no spinlocks may be held.
7226 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7228 struct migration_req req;
7229 unsigned long flags;
7233 rq = task_rq_lock(p, &flags);
7235 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7240 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7241 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7246 if (p->sched_class->set_cpus_allowed)
7247 p->sched_class->set_cpus_allowed(p, new_mask);
7249 cpumask_copy(&p->cpus_allowed, new_mask);
7250 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7253 /* Can the task run on the task's current CPU? If so, we're done */
7254 if (cpumask_test_cpu(task_cpu(p), new_mask))
7257 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7258 /* Need help from migration thread: drop lock and wait. */
7259 struct task_struct *mt = rq->migration_thread;
7261 get_task_struct(mt);
7262 task_rq_unlock(rq, &flags);
7263 wake_up_process(mt);
7264 put_task_struct(mt);
7265 wait_for_completion(&req.done);
7266 tlb_migrate_finish(p->mm);
7270 task_rq_unlock(rq, &flags);
7274 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7277 * Move (not current) task off this cpu, onto dest cpu. We're doing
7278 * this because either it can't run here any more (set_cpus_allowed()
7279 * away from this CPU, or CPU going down), or because we're
7280 * attempting to rebalance this task on exec (sched_exec).
7282 * So we race with normal scheduler movements, but that's OK, as long
7283 * as the task is no longer on this CPU.
7285 * Returns non-zero if task was successfully migrated.
7287 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7289 struct rq *rq_dest, *rq_src;
7292 if (unlikely(!cpu_active(dest_cpu)))
7295 rq_src = cpu_rq(src_cpu);
7296 rq_dest = cpu_rq(dest_cpu);
7298 double_rq_lock(rq_src, rq_dest);
7299 /* Already moved. */
7300 if (task_cpu(p) != src_cpu)
7302 /* Affinity changed (again). */
7303 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7307 * If we're not on a rq, the next wake-up will ensure we're
7311 deactivate_task(rq_src, p, 0);
7312 set_task_cpu(p, dest_cpu);
7313 activate_task(rq_dest, p, 0);
7314 check_preempt_curr(rq_dest, p, 0);
7319 double_rq_unlock(rq_src, rq_dest);
7323 #define RCU_MIGRATION_IDLE 0
7324 #define RCU_MIGRATION_NEED_QS 1
7325 #define RCU_MIGRATION_GOT_QS 2
7326 #define RCU_MIGRATION_MUST_SYNC 3
7329 * migration_thread - this is a highprio system thread that performs
7330 * thread migration by bumping thread off CPU then 'pushing' onto
7333 static int migration_thread(void *data)
7336 int cpu = (long)data;
7340 BUG_ON(rq->migration_thread != current);
7342 set_current_state(TASK_INTERRUPTIBLE);
7343 while (!kthread_should_stop()) {
7344 struct migration_req *req;
7345 struct list_head *head;
7347 spin_lock_irq(&rq->lock);
7349 if (cpu_is_offline(cpu)) {
7350 spin_unlock_irq(&rq->lock);
7354 if (rq->active_balance) {
7355 active_load_balance(rq, cpu);
7356 rq->active_balance = 0;
7359 head = &rq->migration_queue;
7361 if (list_empty(head)) {
7362 spin_unlock_irq(&rq->lock);
7364 set_current_state(TASK_INTERRUPTIBLE);
7367 req = list_entry(head->next, struct migration_req, list);
7368 list_del_init(head->next);
7370 if (req->task != NULL) {
7371 spin_unlock(&rq->lock);
7372 __migrate_task(req->task, cpu, req->dest_cpu);
7373 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7374 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7375 spin_unlock(&rq->lock);
7377 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7378 spin_unlock(&rq->lock);
7379 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7383 complete(&req->done);
7385 __set_current_state(TASK_RUNNING);
7390 #ifdef CONFIG_HOTPLUG_CPU
7392 * Figure out where task on dead CPU should go, use force if necessary.
7394 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7396 struct rq *rq = cpu_rq(dead_cpu);
7397 int needs_cpu, uninitialized_var(dest_cpu);
7398 unsigned long flags;
7400 local_irq_save(flags);
7402 spin_lock(&rq->lock);
7403 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
7405 dest_cpu = select_fallback_rq(dead_cpu, p);
7406 spin_unlock(&rq->lock);
7408 * It can only fail if we race with set_cpus_allowed(),
7409 * in the racer should migrate the task anyway.
7412 __migrate_task(p, dead_cpu, dest_cpu);
7413 local_irq_restore(flags);
7417 * While a dead CPU has no uninterruptible tasks queued at this point,
7418 * it might still have a nonzero ->nr_uninterruptible counter, because
7419 * for performance reasons the counter is not stricly tracking tasks to
7420 * their home CPUs. So we just add the counter to another CPU's counter,
7421 * to keep the global sum constant after CPU-down:
7423 static void migrate_nr_uninterruptible(struct rq *rq_src)
7425 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7426 unsigned long flags;
7428 local_irq_save(flags);
7429 double_rq_lock(rq_src, rq_dest);
7430 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7431 rq_src->nr_uninterruptible = 0;
7432 double_rq_unlock(rq_src, rq_dest);
7433 local_irq_restore(flags);
7436 /* Run through task list and migrate tasks from the dead cpu. */
7437 static void migrate_live_tasks(int src_cpu)
7439 struct task_struct *p, *t;
7441 read_lock(&tasklist_lock);
7443 do_each_thread(t, p) {
7447 if (task_cpu(p) == src_cpu)
7448 move_task_off_dead_cpu(src_cpu, p);
7449 } while_each_thread(t, p);
7451 read_unlock(&tasklist_lock);
7455 * Schedules idle task to be the next runnable task on current CPU.
7456 * It does so by boosting its priority to highest possible.
7457 * Used by CPU offline code.
7459 void sched_idle_next(void)
7461 int this_cpu = smp_processor_id();
7462 struct rq *rq = cpu_rq(this_cpu);
7463 struct task_struct *p = rq->idle;
7464 unsigned long flags;
7466 /* cpu has to be offline */
7467 BUG_ON(cpu_online(this_cpu));
7470 * Strictly not necessary since rest of the CPUs are stopped by now
7471 * and interrupts disabled on the current cpu.
7473 spin_lock_irqsave(&rq->lock, flags);
7475 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7477 update_rq_clock(rq);
7478 activate_task(rq, p, 0);
7480 spin_unlock_irqrestore(&rq->lock, flags);
7484 * Ensures that the idle task is using init_mm right before its cpu goes
7487 void idle_task_exit(void)
7489 struct mm_struct *mm = current->active_mm;
7491 BUG_ON(cpu_online(smp_processor_id()));
7494 switch_mm(mm, &init_mm, current);
7498 /* called under rq->lock with disabled interrupts */
7499 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7501 struct rq *rq = cpu_rq(dead_cpu);
7503 /* Must be exiting, otherwise would be on tasklist. */
7504 BUG_ON(!p->exit_state);
7506 /* Cannot have done final schedule yet: would have vanished. */
7507 BUG_ON(p->state == TASK_DEAD);
7512 * Drop lock around migration; if someone else moves it,
7513 * that's OK. No task can be added to this CPU, so iteration is
7516 spin_unlock_irq(&rq->lock);
7517 move_task_off_dead_cpu(dead_cpu, p);
7518 spin_lock_irq(&rq->lock);
7523 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7524 static void migrate_dead_tasks(unsigned int dead_cpu)
7526 struct rq *rq = cpu_rq(dead_cpu);
7527 struct task_struct *next;
7530 if (!rq->nr_running)
7532 update_rq_clock(rq);
7533 next = pick_next_task(rq);
7536 next->sched_class->put_prev_task(rq, next);
7537 migrate_dead(dead_cpu, next);
7543 * remove the tasks which were accounted by rq from calc_load_tasks.
7545 static void calc_global_load_remove(struct rq *rq)
7547 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7548 rq->calc_load_active = 0;
7550 #endif /* CONFIG_HOTPLUG_CPU */
7552 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7554 static struct ctl_table sd_ctl_dir[] = {
7556 .procname = "sched_domain",
7562 static struct ctl_table sd_ctl_root[] = {
7564 .ctl_name = CTL_KERN,
7565 .procname = "kernel",
7567 .child = sd_ctl_dir,
7572 static struct ctl_table *sd_alloc_ctl_entry(int n)
7574 struct ctl_table *entry =
7575 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7580 static void sd_free_ctl_entry(struct ctl_table **tablep)
7582 struct ctl_table *entry;
7585 * In the intermediate directories, both the child directory and
7586 * procname are dynamically allocated and could fail but the mode
7587 * will always be set. In the lowest directory the names are
7588 * static strings and all have proc handlers.
7590 for (entry = *tablep; entry->mode; entry++) {
7592 sd_free_ctl_entry(&entry->child);
7593 if (entry->proc_handler == NULL)
7594 kfree(entry->procname);
7602 set_table_entry(struct ctl_table *entry,
7603 const char *procname, void *data, int maxlen,
7604 mode_t mode, proc_handler *proc_handler)
7606 entry->procname = procname;
7608 entry->maxlen = maxlen;
7610 entry->proc_handler = proc_handler;
7613 static struct ctl_table *
7614 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7616 struct ctl_table *table = sd_alloc_ctl_entry(13);
7621 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7622 sizeof(long), 0644, proc_doulongvec_minmax);
7623 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7624 sizeof(long), 0644, proc_doulongvec_minmax);
7625 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7626 sizeof(int), 0644, proc_dointvec_minmax);
7627 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7628 sizeof(int), 0644, proc_dointvec_minmax);
7629 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7630 sizeof(int), 0644, proc_dointvec_minmax);
7631 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7632 sizeof(int), 0644, proc_dointvec_minmax);
7633 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7634 sizeof(int), 0644, proc_dointvec_minmax);
7635 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7636 sizeof(int), 0644, proc_dointvec_minmax);
7637 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7638 sizeof(int), 0644, proc_dointvec_minmax);
7639 set_table_entry(&table[9], "cache_nice_tries",
7640 &sd->cache_nice_tries,
7641 sizeof(int), 0644, proc_dointvec_minmax);
7642 set_table_entry(&table[10], "flags", &sd->flags,
7643 sizeof(int), 0644, proc_dointvec_minmax);
7644 set_table_entry(&table[11], "name", sd->name,
7645 CORENAME_MAX_SIZE, 0444, proc_dostring);
7646 /* &table[12] is terminator */
7651 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7653 struct ctl_table *entry, *table;
7654 struct sched_domain *sd;
7655 int domain_num = 0, i;
7658 for_each_domain(cpu, sd)
7660 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7665 for_each_domain(cpu, sd) {
7666 snprintf(buf, 32, "domain%d", i);
7667 entry->procname = kstrdup(buf, GFP_KERNEL);
7669 entry->child = sd_alloc_ctl_domain_table(sd);
7676 static struct ctl_table_header *sd_sysctl_header;
7677 static void register_sched_domain_sysctl(void)
7679 int i, cpu_num = num_possible_cpus();
7680 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7683 WARN_ON(sd_ctl_dir[0].child);
7684 sd_ctl_dir[0].child = entry;
7689 for_each_possible_cpu(i) {
7690 snprintf(buf, 32, "cpu%d", i);
7691 entry->procname = kstrdup(buf, GFP_KERNEL);
7693 entry->child = sd_alloc_ctl_cpu_table(i);
7697 WARN_ON(sd_sysctl_header);
7698 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7701 /* may be called multiple times per register */
7702 static void unregister_sched_domain_sysctl(void)
7704 if (sd_sysctl_header)
7705 unregister_sysctl_table(sd_sysctl_header);
7706 sd_sysctl_header = NULL;
7707 if (sd_ctl_dir[0].child)
7708 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7711 static void register_sched_domain_sysctl(void)
7714 static void unregister_sched_domain_sysctl(void)
7719 static void set_rq_online(struct rq *rq)
7722 const struct sched_class *class;
7724 cpumask_set_cpu(rq->cpu, rq->rd->online);
7727 for_each_class(class) {
7728 if (class->rq_online)
7729 class->rq_online(rq);
7734 static void set_rq_offline(struct rq *rq)
7737 const struct sched_class *class;
7739 for_each_class(class) {
7740 if (class->rq_offline)
7741 class->rq_offline(rq);
7744 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7750 * migration_call - callback that gets triggered when a CPU is added.
7751 * Here we can start up the necessary migration thread for the new CPU.
7753 static int __cpuinit
7754 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7756 struct task_struct *p;
7757 int cpu = (long)hcpu;
7758 unsigned long flags;
7763 case CPU_UP_PREPARE:
7764 case CPU_UP_PREPARE_FROZEN:
7765 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7768 kthread_bind(p, cpu);
7769 /* Must be high prio: stop_machine expects to yield to it. */
7770 rq = task_rq_lock(p, &flags);
7771 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7772 task_rq_unlock(rq, &flags);
7774 cpu_rq(cpu)->migration_thread = p;
7775 rq->calc_load_update = calc_load_update;
7779 case CPU_ONLINE_FROZEN:
7780 /* Strictly unnecessary, as first user will wake it. */
7781 wake_up_process(cpu_rq(cpu)->migration_thread);
7783 /* Update our root-domain */
7785 spin_lock_irqsave(&rq->lock, flags);
7787 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7791 spin_unlock_irqrestore(&rq->lock, flags);
7794 #ifdef CONFIG_HOTPLUG_CPU
7795 case CPU_UP_CANCELED:
7796 case CPU_UP_CANCELED_FROZEN:
7797 if (!cpu_rq(cpu)->migration_thread)
7799 /* Unbind it from offline cpu so it can run. Fall thru. */
7800 kthread_bind(cpu_rq(cpu)->migration_thread,
7801 cpumask_any(cpu_online_mask));
7802 kthread_stop(cpu_rq(cpu)->migration_thread);
7803 put_task_struct(cpu_rq(cpu)->migration_thread);
7804 cpu_rq(cpu)->migration_thread = NULL;
7809 * Bring the migration thread down in CPU_POST_DEAD event,
7810 * since the timers should have got migrated by now and thus
7811 * we should not see a deadlock between trying to kill the
7812 * migration thread and the sched_rt_period_timer.
7815 kthread_stop(rq->migration_thread);
7816 put_task_struct(rq->migration_thread);
7817 rq->migration_thread = NULL;
7821 case CPU_DEAD_FROZEN:
7822 migrate_live_tasks(cpu);
7824 /* Idle task back to normal (off runqueue, low prio) */
7825 spin_lock_irq(&rq->lock);
7826 update_rq_clock(rq);
7827 deactivate_task(rq, rq->idle, 0);
7828 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7829 rq->idle->sched_class = &idle_sched_class;
7830 migrate_dead_tasks(cpu);
7831 spin_unlock_irq(&rq->lock);
7832 migrate_nr_uninterruptible(rq);
7833 BUG_ON(rq->nr_running != 0);
7834 calc_global_load_remove(rq);
7836 * No need to migrate the tasks: it was best-effort if
7837 * they didn't take sched_hotcpu_mutex. Just wake up
7840 spin_lock_irq(&rq->lock);
7841 while (!list_empty(&rq->migration_queue)) {
7842 struct migration_req *req;
7844 req = list_entry(rq->migration_queue.next,
7845 struct migration_req, list);
7846 list_del_init(&req->list);
7847 spin_unlock_irq(&rq->lock);
7848 complete(&req->done);
7849 spin_lock_irq(&rq->lock);
7851 spin_unlock_irq(&rq->lock);
7855 case CPU_DYING_FROZEN:
7856 /* Update our root-domain */
7858 spin_lock_irqsave(&rq->lock, flags);
7860 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7863 spin_unlock_irqrestore(&rq->lock, flags);
7871 * Register at high priority so that task migration (migrate_all_tasks)
7872 * happens before everything else. This has to be lower priority than
7873 * the notifier in the perf_event subsystem, though.
7875 static struct notifier_block __cpuinitdata migration_notifier = {
7876 .notifier_call = migration_call,
7880 static int __init migration_init(void)
7882 void *cpu = (void *)(long)smp_processor_id();
7885 /* Start one for the boot CPU: */
7886 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7887 BUG_ON(err == NOTIFY_BAD);
7888 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7889 register_cpu_notifier(&migration_notifier);
7893 early_initcall(migration_init);
7898 #ifdef CONFIG_SCHED_DEBUG
7900 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7901 struct cpumask *groupmask)
7903 struct sched_group *group = sd->groups;
7906 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7907 cpumask_clear(groupmask);
7909 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7911 if (!(sd->flags & SD_LOAD_BALANCE)) {
7912 printk("does not load-balance\n");
7914 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7919 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7921 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7922 printk(KERN_ERR "ERROR: domain->span does not contain "
7925 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7926 printk(KERN_ERR "ERROR: domain->groups does not contain"
7930 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7934 printk(KERN_ERR "ERROR: group is NULL\n");
7938 if (!group->cpu_power) {
7939 printk(KERN_CONT "\n");
7940 printk(KERN_ERR "ERROR: domain->cpu_power not "
7945 if (!cpumask_weight(sched_group_cpus(group))) {
7946 printk(KERN_CONT "\n");
7947 printk(KERN_ERR "ERROR: empty group\n");
7951 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7952 printk(KERN_CONT "\n");
7953 printk(KERN_ERR "ERROR: repeated CPUs\n");
7957 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7959 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7961 printk(KERN_CONT " %s", str);
7962 if (group->cpu_power != SCHED_LOAD_SCALE) {
7963 printk(KERN_CONT " (cpu_power = %d)",
7967 group = group->next;
7968 } while (group != sd->groups);
7969 printk(KERN_CONT "\n");
7971 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7972 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7975 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7976 printk(KERN_ERR "ERROR: parent span is not a superset "
7977 "of domain->span\n");
7981 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7983 cpumask_var_t groupmask;
7987 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7991 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7993 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7994 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7999 if (sched_domain_debug_one(sd, cpu, level, groupmask))
8006 free_cpumask_var(groupmask);
8008 #else /* !CONFIG_SCHED_DEBUG */
8009 # define sched_domain_debug(sd, cpu) do { } while (0)
8010 #endif /* CONFIG_SCHED_DEBUG */
8012 static int sd_degenerate(struct sched_domain *sd)
8014 if (cpumask_weight(sched_domain_span(sd)) == 1)
8017 /* Following flags need at least 2 groups */
8018 if (sd->flags & (SD_LOAD_BALANCE |
8019 SD_BALANCE_NEWIDLE |
8023 SD_SHARE_PKG_RESOURCES)) {
8024 if (sd->groups != sd->groups->next)
8028 /* Following flags don't use groups */
8029 if (sd->flags & (SD_WAKE_AFFINE))
8036 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
8038 unsigned long cflags = sd->flags, pflags = parent->flags;
8040 if (sd_degenerate(parent))
8043 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
8046 /* Flags needing groups don't count if only 1 group in parent */
8047 if (parent->groups == parent->groups->next) {
8048 pflags &= ~(SD_LOAD_BALANCE |
8049 SD_BALANCE_NEWIDLE |
8053 SD_SHARE_PKG_RESOURCES);
8054 if (nr_node_ids == 1)
8055 pflags &= ~SD_SERIALIZE;
8057 if (~cflags & pflags)
8063 static void free_rootdomain(struct root_domain *rd)
8065 synchronize_sched();
8067 cpupri_cleanup(&rd->cpupri);
8069 free_cpumask_var(rd->rto_mask);
8070 free_cpumask_var(rd->online);
8071 free_cpumask_var(rd->span);
8075 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
8077 struct root_domain *old_rd = NULL;
8078 unsigned long flags;
8080 spin_lock_irqsave(&rq->lock, flags);
8085 if (cpumask_test_cpu(rq->cpu, old_rd->online))
8088 cpumask_clear_cpu(rq->cpu, old_rd->span);
8091 * If we dont want to free the old_rt yet then
8092 * set old_rd to NULL to skip the freeing later
8095 if (!atomic_dec_and_test(&old_rd->refcount))
8099 atomic_inc(&rd->refcount);
8102 cpumask_set_cpu(rq->cpu, rd->span);
8103 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
8106 spin_unlock_irqrestore(&rq->lock, flags);
8109 free_rootdomain(old_rd);
8112 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8114 gfp_t gfp = GFP_KERNEL;
8116 memset(rd, 0, sizeof(*rd));
8121 if (!alloc_cpumask_var(&rd->span, gfp))
8123 if (!alloc_cpumask_var(&rd->online, gfp))
8125 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8128 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8133 free_cpumask_var(rd->rto_mask);
8135 free_cpumask_var(rd->online);
8137 free_cpumask_var(rd->span);
8142 static void init_defrootdomain(void)
8144 init_rootdomain(&def_root_domain, true);
8146 atomic_set(&def_root_domain.refcount, 1);
8149 static struct root_domain *alloc_rootdomain(void)
8151 struct root_domain *rd;
8153 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8157 if (init_rootdomain(rd, false) != 0) {
8166 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8167 * hold the hotplug lock.
8170 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8172 struct rq *rq = cpu_rq(cpu);
8173 struct sched_domain *tmp;
8175 /* Remove the sched domains which do not contribute to scheduling. */
8176 for (tmp = sd; tmp; ) {
8177 struct sched_domain *parent = tmp->parent;
8181 if (sd_parent_degenerate(tmp, parent)) {
8182 tmp->parent = parent->parent;
8184 parent->parent->child = tmp;
8189 if (sd && sd_degenerate(sd)) {
8195 sched_domain_debug(sd, cpu);
8197 rq_attach_root(rq, rd);
8198 rcu_assign_pointer(rq->sd, sd);
8201 /* cpus with isolated domains */
8202 static cpumask_var_t cpu_isolated_map;
8204 /* Setup the mask of cpus configured for isolated domains */
8205 static int __init isolated_cpu_setup(char *str)
8207 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8208 cpulist_parse(str, cpu_isolated_map);
8212 __setup("isolcpus=", isolated_cpu_setup);
8215 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8216 * to a function which identifies what group(along with sched group) a CPU
8217 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8218 * (due to the fact that we keep track of groups covered with a struct cpumask).
8220 * init_sched_build_groups will build a circular linked list of the groups
8221 * covered by the given span, and will set each group's ->cpumask correctly,
8222 * and ->cpu_power to 0.
8225 init_sched_build_groups(const struct cpumask *span,
8226 const struct cpumask *cpu_map,
8227 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8228 struct sched_group **sg,
8229 struct cpumask *tmpmask),
8230 struct cpumask *covered, struct cpumask *tmpmask)
8232 struct sched_group *first = NULL, *last = NULL;
8235 cpumask_clear(covered);
8237 for_each_cpu(i, span) {
8238 struct sched_group *sg;
8239 int group = group_fn(i, cpu_map, &sg, tmpmask);
8242 if (cpumask_test_cpu(i, covered))
8245 cpumask_clear(sched_group_cpus(sg));
8248 for_each_cpu(j, span) {
8249 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8252 cpumask_set_cpu(j, covered);
8253 cpumask_set_cpu(j, sched_group_cpus(sg));
8264 #define SD_NODES_PER_DOMAIN 16
8269 * find_next_best_node - find the next node to include in a sched_domain
8270 * @node: node whose sched_domain we're building
8271 * @used_nodes: nodes already in the sched_domain
8273 * Find the next node to include in a given scheduling domain. Simply
8274 * finds the closest node not already in the @used_nodes map.
8276 * Should use nodemask_t.
8278 static int find_next_best_node(int node, nodemask_t *used_nodes)
8280 int i, n, val, min_val, best_node = 0;
8284 for (i = 0; i < nr_node_ids; i++) {
8285 /* Start at @node */
8286 n = (node + i) % nr_node_ids;
8288 if (!nr_cpus_node(n))
8291 /* Skip already used nodes */
8292 if (node_isset(n, *used_nodes))
8295 /* Simple min distance search */
8296 val = node_distance(node, n);
8298 if (val < min_val) {
8304 node_set(best_node, *used_nodes);
8309 * sched_domain_node_span - get a cpumask for a node's sched_domain
8310 * @node: node whose cpumask we're constructing
8311 * @span: resulting cpumask
8313 * Given a node, construct a good cpumask for its sched_domain to span. It
8314 * should be one that prevents unnecessary balancing, but also spreads tasks
8317 static void sched_domain_node_span(int node, struct cpumask *span)
8319 nodemask_t used_nodes;
8322 cpumask_clear(span);
8323 nodes_clear(used_nodes);
8325 cpumask_or(span, span, cpumask_of_node(node));
8326 node_set(node, used_nodes);
8328 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8329 int next_node = find_next_best_node(node, &used_nodes);
8331 cpumask_or(span, span, cpumask_of_node(next_node));
8334 #endif /* CONFIG_NUMA */
8336 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8339 * The cpus mask in sched_group and sched_domain hangs off the end.
8341 * ( See the the comments in include/linux/sched.h:struct sched_group
8342 * and struct sched_domain. )
8344 struct static_sched_group {
8345 struct sched_group sg;
8346 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8349 struct static_sched_domain {
8350 struct sched_domain sd;
8351 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8357 cpumask_var_t domainspan;
8358 cpumask_var_t covered;
8359 cpumask_var_t notcovered;
8361 cpumask_var_t nodemask;
8362 cpumask_var_t this_sibling_map;
8363 cpumask_var_t this_core_map;
8364 cpumask_var_t send_covered;
8365 cpumask_var_t tmpmask;
8366 struct sched_group **sched_group_nodes;
8367 struct root_domain *rd;
8371 sa_sched_groups = 0,
8376 sa_this_sibling_map,
8378 sa_sched_group_nodes,
8388 * SMT sched-domains:
8390 #ifdef CONFIG_SCHED_SMT
8391 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8392 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8395 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8396 struct sched_group **sg, struct cpumask *unused)
8399 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8402 #endif /* CONFIG_SCHED_SMT */
8405 * multi-core sched-domains:
8407 #ifdef CONFIG_SCHED_MC
8408 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8409 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8410 #endif /* CONFIG_SCHED_MC */
8412 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8414 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8415 struct sched_group **sg, struct cpumask *mask)
8419 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8420 group = cpumask_first(mask);
8422 *sg = &per_cpu(sched_group_core, group).sg;
8425 #elif defined(CONFIG_SCHED_MC)
8427 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8428 struct sched_group **sg, struct cpumask *unused)
8431 *sg = &per_cpu(sched_group_core, cpu).sg;
8436 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8437 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8440 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8441 struct sched_group **sg, struct cpumask *mask)
8444 #ifdef CONFIG_SCHED_MC
8445 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8446 group = cpumask_first(mask);
8447 #elif defined(CONFIG_SCHED_SMT)
8448 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8449 group = cpumask_first(mask);
8454 *sg = &per_cpu(sched_group_phys, group).sg;
8460 * The init_sched_build_groups can't handle what we want to do with node
8461 * groups, so roll our own. Now each node has its own list of groups which
8462 * gets dynamically allocated.
8464 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8465 static struct sched_group ***sched_group_nodes_bycpu;
8467 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8468 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8470 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8471 struct sched_group **sg,
8472 struct cpumask *nodemask)
8476 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8477 group = cpumask_first(nodemask);
8480 *sg = &per_cpu(sched_group_allnodes, group).sg;
8484 static void init_numa_sched_groups_power(struct sched_group *group_head)
8486 struct sched_group *sg = group_head;
8492 for_each_cpu(j, sched_group_cpus(sg)) {
8493 struct sched_domain *sd;
8495 sd = &per_cpu(phys_domains, j).sd;
8496 if (j != group_first_cpu(sd->groups)) {
8498 * Only add "power" once for each
8504 sg->cpu_power += sd->groups->cpu_power;
8507 } while (sg != group_head);
8510 static int build_numa_sched_groups(struct s_data *d,
8511 const struct cpumask *cpu_map, int num)
8513 struct sched_domain *sd;
8514 struct sched_group *sg, *prev;
8517 cpumask_clear(d->covered);
8518 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8519 if (cpumask_empty(d->nodemask)) {
8520 d->sched_group_nodes[num] = NULL;
8524 sched_domain_node_span(num, d->domainspan);
8525 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8527 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8530 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8534 d->sched_group_nodes[num] = sg;
8536 for_each_cpu(j, d->nodemask) {
8537 sd = &per_cpu(node_domains, j).sd;
8542 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8544 cpumask_or(d->covered, d->covered, d->nodemask);
8547 for (j = 0; j < nr_node_ids; j++) {
8548 n = (num + j) % nr_node_ids;
8549 cpumask_complement(d->notcovered, d->covered);
8550 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8551 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8552 if (cpumask_empty(d->tmpmask))
8554 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8555 if (cpumask_empty(d->tmpmask))
8557 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8561 "Can not alloc domain group for node %d\n", j);
8565 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8566 sg->next = prev->next;
8567 cpumask_or(d->covered, d->covered, d->tmpmask);
8574 #endif /* CONFIG_NUMA */
8577 /* Free memory allocated for various sched_group structures */
8578 static void free_sched_groups(const struct cpumask *cpu_map,
8579 struct cpumask *nodemask)
8583 for_each_cpu(cpu, cpu_map) {
8584 struct sched_group **sched_group_nodes
8585 = sched_group_nodes_bycpu[cpu];
8587 if (!sched_group_nodes)
8590 for (i = 0; i < nr_node_ids; i++) {
8591 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8593 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8594 if (cpumask_empty(nodemask))
8604 if (oldsg != sched_group_nodes[i])
8607 kfree(sched_group_nodes);
8608 sched_group_nodes_bycpu[cpu] = NULL;
8611 #else /* !CONFIG_NUMA */
8612 static void free_sched_groups(const struct cpumask *cpu_map,
8613 struct cpumask *nodemask)
8616 #endif /* CONFIG_NUMA */
8619 * Initialize sched groups cpu_power.
8621 * cpu_power indicates the capacity of sched group, which is used while
8622 * distributing the load between different sched groups in a sched domain.
8623 * Typically cpu_power for all the groups in a sched domain will be same unless
8624 * there are asymmetries in the topology. If there are asymmetries, group
8625 * having more cpu_power will pickup more load compared to the group having
8628 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8630 struct sched_domain *child;
8631 struct sched_group *group;
8635 WARN_ON(!sd || !sd->groups);
8637 if (cpu != group_first_cpu(sd->groups))
8642 sd->groups->cpu_power = 0;
8645 power = SCHED_LOAD_SCALE;
8646 weight = cpumask_weight(sched_domain_span(sd));
8648 * SMT siblings share the power of a single core.
8649 * Usually multiple threads get a better yield out of
8650 * that one core than a single thread would have,
8651 * reflect that in sd->smt_gain.
8653 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8654 power *= sd->smt_gain;
8656 power >>= SCHED_LOAD_SHIFT;
8658 sd->groups->cpu_power += power;
8663 * Add cpu_power of each child group to this groups cpu_power.
8665 group = child->groups;
8667 sd->groups->cpu_power += group->cpu_power;
8668 group = group->next;
8669 } while (group != child->groups);
8673 * Initializers for schedule domains
8674 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8677 #ifdef CONFIG_SCHED_DEBUG
8678 # define SD_INIT_NAME(sd, type) sd->name = #type
8680 # define SD_INIT_NAME(sd, type) do { } while (0)
8683 #define SD_INIT(sd, type) sd_init_##type(sd)
8685 #define SD_INIT_FUNC(type) \
8686 static noinline void sd_init_##type(struct sched_domain *sd) \
8688 memset(sd, 0, sizeof(*sd)); \
8689 *sd = SD_##type##_INIT; \
8690 sd->level = SD_LV_##type; \
8691 SD_INIT_NAME(sd, type); \
8696 SD_INIT_FUNC(ALLNODES)
8699 #ifdef CONFIG_SCHED_SMT
8700 SD_INIT_FUNC(SIBLING)
8702 #ifdef CONFIG_SCHED_MC
8706 static int default_relax_domain_level = -1;
8708 static int __init setup_relax_domain_level(char *str)
8712 val = simple_strtoul(str, NULL, 0);
8713 if (val < SD_LV_MAX)
8714 default_relax_domain_level = val;
8718 __setup("relax_domain_level=", setup_relax_domain_level);
8720 static void set_domain_attribute(struct sched_domain *sd,
8721 struct sched_domain_attr *attr)
8725 if (!attr || attr->relax_domain_level < 0) {
8726 if (default_relax_domain_level < 0)
8729 request = default_relax_domain_level;
8731 request = attr->relax_domain_level;
8732 if (request < sd->level) {
8733 /* turn off idle balance on this domain */
8734 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8736 /* turn on idle balance on this domain */
8737 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8741 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8742 const struct cpumask *cpu_map)
8745 case sa_sched_groups:
8746 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8747 d->sched_group_nodes = NULL;
8749 free_rootdomain(d->rd); /* fall through */
8751 free_cpumask_var(d->tmpmask); /* fall through */
8752 case sa_send_covered:
8753 free_cpumask_var(d->send_covered); /* fall through */
8754 case sa_this_core_map:
8755 free_cpumask_var(d->this_core_map); /* fall through */
8756 case sa_this_sibling_map:
8757 free_cpumask_var(d->this_sibling_map); /* fall through */
8759 free_cpumask_var(d->nodemask); /* fall through */
8760 case sa_sched_group_nodes:
8762 kfree(d->sched_group_nodes); /* fall through */
8764 free_cpumask_var(d->notcovered); /* fall through */
8766 free_cpumask_var(d->covered); /* fall through */
8768 free_cpumask_var(d->domainspan); /* fall through */
8775 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8776 const struct cpumask *cpu_map)
8779 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8781 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8782 return sa_domainspan;
8783 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8785 /* Allocate the per-node list of sched groups */
8786 d->sched_group_nodes = kcalloc(nr_node_ids,
8787 sizeof(struct sched_group *), GFP_KERNEL);
8788 if (!d->sched_group_nodes) {
8789 printk(KERN_WARNING "Can not alloc sched group node list\n");
8790 return sa_notcovered;
8792 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8794 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8795 return sa_sched_group_nodes;
8796 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8798 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8799 return sa_this_sibling_map;
8800 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8801 return sa_this_core_map;
8802 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8803 return sa_send_covered;
8804 d->rd = alloc_rootdomain();
8806 printk(KERN_WARNING "Cannot alloc root domain\n");
8809 return sa_rootdomain;
8812 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8813 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8815 struct sched_domain *sd = NULL;
8817 struct sched_domain *parent;
8820 if (cpumask_weight(cpu_map) >
8821 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8822 sd = &per_cpu(allnodes_domains, i).sd;
8823 SD_INIT(sd, ALLNODES);
8824 set_domain_attribute(sd, attr);
8825 cpumask_copy(sched_domain_span(sd), cpu_map);
8826 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8831 sd = &per_cpu(node_domains, i).sd;
8833 set_domain_attribute(sd, attr);
8834 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8835 sd->parent = parent;
8838 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8843 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8844 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8845 struct sched_domain *parent, int i)
8847 struct sched_domain *sd;
8848 sd = &per_cpu(phys_domains, i).sd;
8850 set_domain_attribute(sd, attr);
8851 cpumask_copy(sched_domain_span(sd), d->nodemask);
8852 sd->parent = parent;
8855 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8859 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8860 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8861 struct sched_domain *parent, int i)
8863 struct sched_domain *sd = parent;
8864 #ifdef CONFIG_SCHED_MC
8865 sd = &per_cpu(core_domains, i).sd;
8867 set_domain_attribute(sd, attr);
8868 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8869 sd->parent = parent;
8871 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8876 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8877 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8878 struct sched_domain *parent, int i)
8880 struct sched_domain *sd = parent;
8881 #ifdef CONFIG_SCHED_SMT
8882 sd = &per_cpu(cpu_domains, i).sd;
8883 SD_INIT(sd, SIBLING);
8884 set_domain_attribute(sd, attr);
8885 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8886 sd->parent = parent;
8888 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8893 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8894 const struct cpumask *cpu_map, int cpu)
8897 #ifdef CONFIG_SCHED_SMT
8898 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8899 cpumask_and(d->this_sibling_map, cpu_map,
8900 topology_thread_cpumask(cpu));
8901 if (cpu == cpumask_first(d->this_sibling_map))
8902 init_sched_build_groups(d->this_sibling_map, cpu_map,
8904 d->send_covered, d->tmpmask);
8907 #ifdef CONFIG_SCHED_MC
8908 case SD_LV_MC: /* set up multi-core groups */
8909 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8910 if (cpu == cpumask_first(d->this_core_map))
8911 init_sched_build_groups(d->this_core_map, cpu_map,
8913 d->send_covered, d->tmpmask);
8916 case SD_LV_CPU: /* set up physical groups */
8917 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8918 if (!cpumask_empty(d->nodemask))
8919 init_sched_build_groups(d->nodemask, cpu_map,
8921 d->send_covered, d->tmpmask);
8924 case SD_LV_ALLNODES:
8925 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8926 d->send_covered, d->tmpmask);
8935 * Build sched domains for a given set of cpus and attach the sched domains
8936 * to the individual cpus
8938 static int __build_sched_domains(const struct cpumask *cpu_map,
8939 struct sched_domain_attr *attr)
8941 enum s_alloc alloc_state = sa_none;
8943 struct sched_domain *sd;
8949 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8950 if (alloc_state != sa_rootdomain)
8952 alloc_state = sa_sched_groups;
8955 * Set up domains for cpus specified by the cpu_map.
8957 for_each_cpu(i, cpu_map) {
8958 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8961 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8962 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8963 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8964 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8967 for_each_cpu(i, cpu_map) {
8968 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8969 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8972 /* Set up physical groups */
8973 for (i = 0; i < nr_node_ids; i++)
8974 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8977 /* Set up node groups */
8979 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8981 for (i = 0; i < nr_node_ids; i++)
8982 if (build_numa_sched_groups(&d, cpu_map, i))
8986 /* Calculate CPU power for physical packages and nodes */
8987 #ifdef CONFIG_SCHED_SMT
8988 for_each_cpu(i, cpu_map) {
8989 sd = &per_cpu(cpu_domains, i).sd;
8990 init_sched_groups_power(i, sd);
8993 #ifdef CONFIG_SCHED_MC
8994 for_each_cpu(i, cpu_map) {
8995 sd = &per_cpu(core_domains, i).sd;
8996 init_sched_groups_power(i, sd);
9000 for_each_cpu(i, cpu_map) {
9001 sd = &per_cpu(phys_domains, i).sd;
9002 init_sched_groups_power(i, sd);
9006 for (i = 0; i < nr_node_ids; i++)
9007 init_numa_sched_groups_power(d.sched_group_nodes[i]);
9009 if (d.sd_allnodes) {
9010 struct sched_group *sg;
9012 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
9014 init_numa_sched_groups_power(sg);
9018 /* Attach the domains */
9019 for_each_cpu(i, cpu_map) {
9020 #ifdef CONFIG_SCHED_SMT
9021 sd = &per_cpu(cpu_domains, i).sd;
9022 #elif defined(CONFIG_SCHED_MC)
9023 sd = &per_cpu(core_domains, i).sd;
9025 sd = &per_cpu(phys_domains, i).sd;
9027 cpu_attach_domain(sd, d.rd, i);
9030 d.sched_group_nodes = NULL; /* don't free this we still need it */
9031 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
9035 __free_domain_allocs(&d, alloc_state, cpu_map);
9039 static int build_sched_domains(const struct cpumask *cpu_map)
9041 return __build_sched_domains(cpu_map, NULL);
9044 static struct cpumask *doms_cur; /* current sched domains */
9045 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
9046 static struct sched_domain_attr *dattr_cur;
9047 /* attribues of custom domains in 'doms_cur' */
9050 * Special case: If a kmalloc of a doms_cur partition (array of
9051 * cpumask) fails, then fallback to a single sched domain,
9052 * as determined by the single cpumask fallback_doms.
9054 static cpumask_var_t fallback_doms;
9057 * arch_update_cpu_topology lets virtualized architectures update the
9058 * cpu core maps. It is supposed to return 1 if the topology changed
9059 * or 0 if it stayed the same.
9061 int __attribute__((weak)) arch_update_cpu_topology(void)
9067 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9068 * For now this just excludes isolated cpus, but could be used to
9069 * exclude other special cases in the future.
9071 static int arch_init_sched_domains(const struct cpumask *cpu_map)
9075 arch_update_cpu_topology();
9077 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
9079 doms_cur = fallback_doms;
9080 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
9082 err = build_sched_domains(doms_cur);
9083 register_sched_domain_sysctl();
9088 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9089 struct cpumask *tmpmask)
9091 free_sched_groups(cpu_map, tmpmask);
9095 * Detach sched domains from a group of cpus specified in cpu_map
9096 * These cpus will now be attached to the NULL domain
9098 static void detach_destroy_domains(const struct cpumask *cpu_map)
9100 /* Save because hotplug lock held. */
9101 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9104 for_each_cpu(i, cpu_map)
9105 cpu_attach_domain(NULL, &def_root_domain, i);
9106 synchronize_sched();
9107 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9110 /* handle null as "default" */
9111 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9112 struct sched_domain_attr *new, int idx_new)
9114 struct sched_domain_attr tmp;
9121 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9122 new ? (new + idx_new) : &tmp,
9123 sizeof(struct sched_domain_attr));
9127 * Partition sched domains as specified by the 'ndoms_new'
9128 * cpumasks in the array doms_new[] of cpumasks. This compares
9129 * doms_new[] to the current sched domain partitioning, doms_cur[].
9130 * It destroys each deleted domain and builds each new domain.
9132 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
9133 * The masks don't intersect (don't overlap.) We should setup one
9134 * sched domain for each mask. CPUs not in any of the cpumasks will
9135 * not be load balanced. If the same cpumask appears both in the
9136 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9139 * The passed in 'doms_new' should be kmalloc'd. This routine takes
9140 * ownership of it and will kfree it when done with it. If the caller
9141 * failed the kmalloc call, then it can pass in doms_new == NULL &&
9142 * ndoms_new == 1, and partition_sched_domains() will fallback to
9143 * the single partition 'fallback_doms', it also forces the domains
9146 * If doms_new == NULL it will be replaced with cpu_online_mask.
9147 * ndoms_new == 0 is a special case for destroying existing domains,
9148 * and it will not create the default domain.
9150 * Call with hotplug lock held
9152 /* FIXME: Change to struct cpumask *doms_new[] */
9153 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
9154 struct sched_domain_attr *dattr_new)
9159 mutex_lock(&sched_domains_mutex);
9161 /* always unregister in case we don't destroy any domains */
9162 unregister_sched_domain_sysctl();
9164 /* Let architecture update cpu core mappings. */
9165 new_topology = arch_update_cpu_topology();
9167 n = doms_new ? ndoms_new : 0;
9169 /* Destroy deleted domains */
9170 for (i = 0; i < ndoms_cur; i++) {
9171 for (j = 0; j < n && !new_topology; j++) {
9172 if (cpumask_equal(&doms_cur[i], &doms_new[j])
9173 && dattrs_equal(dattr_cur, i, dattr_new, j))
9176 /* no match - a current sched domain not in new doms_new[] */
9177 detach_destroy_domains(doms_cur + i);
9182 if (doms_new == NULL) {
9184 doms_new = fallback_doms;
9185 cpumask_andnot(&doms_new[0], cpu_active_mask, cpu_isolated_map);
9186 WARN_ON_ONCE(dattr_new);
9189 /* Build new domains */
9190 for (i = 0; i < ndoms_new; i++) {
9191 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9192 if (cpumask_equal(&doms_new[i], &doms_cur[j])
9193 && dattrs_equal(dattr_new, i, dattr_cur, j))
9196 /* no match - add a new doms_new */
9197 __build_sched_domains(doms_new + i,
9198 dattr_new ? dattr_new + i : NULL);
9203 /* Remember the new sched domains */
9204 if (doms_cur != fallback_doms)
9206 kfree(dattr_cur); /* kfree(NULL) is safe */
9207 doms_cur = doms_new;
9208 dattr_cur = dattr_new;
9209 ndoms_cur = ndoms_new;
9211 register_sched_domain_sysctl();
9213 mutex_unlock(&sched_domains_mutex);
9216 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9217 static void arch_reinit_sched_domains(void)
9221 /* Destroy domains first to force the rebuild */
9222 partition_sched_domains(0, NULL, NULL);
9224 rebuild_sched_domains();
9228 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9230 unsigned int level = 0;
9232 if (sscanf(buf, "%u", &level) != 1)
9236 * level is always be positive so don't check for
9237 * level < POWERSAVINGS_BALANCE_NONE which is 0
9238 * What happens on 0 or 1 byte write,
9239 * need to check for count as well?
9242 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9246 sched_smt_power_savings = level;
9248 sched_mc_power_savings = level;
9250 arch_reinit_sched_domains();
9255 #ifdef CONFIG_SCHED_MC
9256 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9259 return sprintf(page, "%u\n", sched_mc_power_savings);
9261 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9262 const char *buf, size_t count)
9264 return sched_power_savings_store(buf, count, 0);
9266 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9267 sched_mc_power_savings_show,
9268 sched_mc_power_savings_store);
9271 #ifdef CONFIG_SCHED_SMT
9272 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9275 return sprintf(page, "%u\n", sched_smt_power_savings);
9277 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9278 const char *buf, size_t count)
9280 return sched_power_savings_store(buf, count, 1);
9282 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9283 sched_smt_power_savings_show,
9284 sched_smt_power_savings_store);
9287 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9291 #ifdef CONFIG_SCHED_SMT
9293 err = sysfs_create_file(&cls->kset.kobj,
9294 &attr_sched_smt_power_savings.attr);
9296 #ifdef CONFIG_SCHED_MC
9297 if (!err && mc_capable())
9298 err = sysfs_create_file(&cls->kset.kobj,
9299 &attr_sched_mc_power_savings.attr);
9303 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9305 #ifndef CONFIG_CPUSETS
9307 * Add online and remove offline CPUs from the scheduler domains.
9308 * When cpusets are enabled they take over this function.
9310 static int update_sched_domains(struct notifier_block *nfb,
9311 unsigned long action, void *hcpu)
9315 case CPU_ONLINE_FROZEN:
9316 case CPU_DOWN_PREPARE:
9317 case CPU_DOWN_PREPARE_FROZEN:
9318 case CPU_DOWN_FAILED:
9319 case CPU_DOWN_FAILED_FROZEN:
9320 partition_sched_domains(1, NULL, NULL);
9329 static int update_runtime(struct notifier_block *nfb,
9330 unsigned long action, void *hcpu)
9332 int cpu = (int)(long)hcpu;
9335 case CPU_DOWN_PREPARE:
9336 case CPU_DOWN_PREPARE_FROZEN:
9337 disable_runtime(cpu_rq(cpu));
9340 case CPU_DOWN_FAILED:
9341 case CPU_DOWN_FAILED_FROZEN:
9343 case CPU_ONLINE_FROZEN:
9344 enable_runtime(cpu_rq(cpu));
9352 void __init sched_init_smp(void)
9354 cpumask_var_t non_isolated_cpus;
9356 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9357 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9359 #if defined(CONFIG_NUMA)
9360 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9362 BUG_ON(sched_group_nodes_bycpu == NULL);
9365 mutex_lock(&sched_domains_mutex);
9366 arch_init_sched_domains(cpu_active_mask);
9367 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9368 if (cpumask_empty(non_isolated_cpus))
9369 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9370 mutex_unlock(&sched_domains_mutex);
9373 #ifndef CONFIG_CPUSETS
9374 /* XXX: Theoretical race here - CPU may be hotplugged now */
9375 hotcpu_notifier(update_sched_domains, 0);
9378 /* RT runtime code needs to handle some hotplug events */
9379 hotcpu_notifier(update_runtime, 0);
9383 /* Move init over to a non-isolated CPU */
9384 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9386 sched_init_granularity();
9387 free_cpumask_var(non_isolated_cpus);
9389 init_sched_rt_class();
9392 void __init sched_init_smp(void)
9394 sched_init_granularity();
9396 #endif /* CONFIG_SMP */
9398 const_debug unsigned int sysctl_timer_migration = 1;
9400 int in_sched_functions(unsigned long addr)
9402 return in_lock_functions(addr) ||
9403 (addr >= (unsigned long)__sched_text_start
9404 && addr < (unsigned long)__sched_text_end);
9407 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9409 cfs_rq->tasks_timeline = RB_ROOT;
9410 INIT_LIST_HEAD(&cfs_rq->tasks);
9411 #ifdef CONFIG_FAIR_GROUP_SCHED
9414 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9417 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9419 struct rt_prio_array *array;
9422 array = &rt_rq->active;
9423 for (i = 0; i < MAX_RT_PRIO; i++) {
9424 INIT_LIST_HEAD(array->queue + i);
9425 __clear_bit(i, array->bitmap);
9427 /* delimiter for bitsearch: */
9428 __set_bit(MAX_RT_PRIO, array->bitmap);
9430 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9431 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9433 rt_rq->highest_prio.next = MAX_RT_PRIO;
9437 rt_rq->rt_nr_migratory = 0;
9438 rt_rq->overloaded = 0;
9439 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9443 rt_rq->rt_throttled = 0;
9444 rt_rq->rt_runtime = 0;
9445 spin_lock_init(&rt_rq->rt_runtime_lock);
9447 #ifdef CONFIG_RT_GROUP_SCHED
9448 rt_rq->rt_nr_boosted = 0;
9453 #ifdef CONFIG_FAIR_GROUP_SCHED
9454 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9455 struct sched_entity *se, int cpu, int add,
9456 struct sched_entity *parent)
9458 struct rq *rq = cpu_rq(cpu);
9459 tg->cfs_rq[cpu] = cfs_rq;
9460 init_cfs_rq(cfs_rq, rq);
9463 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9466 /* se could be NULL for init_task_group */
9471 se->cfs_rq = &rq->cfs;
9473 se->cfs_rq = parent->my_q;
9476 se->load.weight = tg->shares;
9477 se->load.inv_weight = 0;
9478 se->parent = parent;
9482 #ifdef CONFIG_RT_GROUP_SCHED
9483 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9484 struct sched_rt_entity *rt_se, int cpu, int add,
9485 struct sched_rt_entity *parent)
9487 struct rq *rq = cpu_rq(cpu);
9489 tg->rt_rq[cpu] = rt_rq;
9490 init_rt_rq(rt_rq, rq);
9492 rt_rq->rt_se = rt_se;
9493 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9495 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9497 tg->rt_se[cpu] = rt_se;
9502 rt_se->rt_rq = &rq->rt;
9504 rt_se->rt_rq = parent->my_q;
9506 rt_se->my_q = rt_rq;
9507 rt_se->parent = parent;
9508 INIT_LIST_HEAD(&rt_se->run_list);
9512 void __init sched_init(void)
9515 unsigned long alloc_size = 0, ptr;
9517 #ifdef CONFIG_FAIR_GROUP_SCHED
9518 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9520 #ifdef CONFIG_RT_GROUP_SCHED
9521 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9523 #ifdef CONFIG_USER_SCHED
9526 #ifdef CONFIG_CPUMASK_OFFSTACK
9527 alloc_size += num_possible_cpus() * cpumask_size();
9530 * As sched_init() is called before page_alloc is setup,
9531 * we use alloc_bootmem().
9534 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9536 #ifdef CONFIG_FAIR_GROUP_SCHED
9537 init_task_group.se = (struct sched_entity **)ptr;
9538 ptr += nr_cpu_ids * sizeof(void **);
9540 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9541 ptr += nr_cpu_ids * sizeof(void **);
9543 #ifdef CONFIG_USER_SCHED
9544 root_task_group.se = (struct sched_entity **)ptr;
9545 ptr += nr_cpu_ids * sizeof(void **);
9547 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9548 ptr += nr_cpu_ids * sizeof(void **);
9549 #endif /* CONFIG_USER_SCHED */
9550 #endif /* CONFIG_FAIR_GROUP_SCHED */
9551 #ifdef CONFIG_RT_GROUP_SCHED
9552 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9553 ptr += nr_cpu_ids * sizeof(void **);
9555 init_task_group.rt_rq = (struct rt_rq **)ptr;
9556 ptr += nr_cpu_ids * sizeof(void **);
9558 #ifdef CONFIG_USER_SCHED
9559 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9560 ptr += nr_cpu_ids * sizeof(void **);
9562 root_task_group.rt_rq = (struct rt_rq **)ptr;
9563 ptr += nr_cpu_ids * sizeof(void **);
9564 #endif /* CONFIG_USER_SCHED */
9565 #endif /* CONFIG_RT_GROUP_SCHED */
9566 #ifdef CONFIG_CPUMASK_OFFSTACK
9567 for_each_possible_cpu(i) {
9568 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9569 ptr += cpumask_size();
9571 #endif /* CONFIG_CPUMASK_OFFSTACK */
9575 init_defrootdomain();
9578 init_rt_bandwidth(&def_rt_bandwidth,
9579 global_rt_period(), global_rt_runtime());
9581 #ifdef CONFIG_RT_GROUP_SCHED
9582 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9583 global_rt_period(), global_rt_runtime());
9584 #ifdef CONFIG_USER_SCHED
9585 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9586 global_rt_period(), RUNTIME_INF);
9587 #endif /* CONFIG_USER_SCHED */
9588 #endif /* CONFIG_RT_GROUP_SCHED */
9590 #ifdef CONFIG_GROUP_SCHED
9591 list_add(&init_task_group.list, &task_groups);
9592 INIT_LIST_HEAD(&init_task_group.children);
9594 #ifdef CONFIG_USER_SCHED
9595 INIT_LIST_HEAD(&root_task_group.children);
9596 init_task_group.parent = &root_task_group;
9597 list_add(&init_task_group.siblings, &root_task_group.children);
9598 #endif /* CONFIG_USER_SCHED */
9599 #endif /* CONFIG_GROUP_SCHED */
9601 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9602 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9603 __alignof__(unsigned long));
9605 for_each_possible_cpu(i) {
9609 spin_lock_init(&rq->lock);
9611 rq->calc_load_active = 0;
9612 rq->calc_load_update = jiffies + LOAD_FREQ;
9613 init_cfs_rq(&rq->cfs, rq);
9614 init_rt_rq(&rq->rt, rq);
9615 #ifdef CONFIG_FAIR_GROUP_SCHED
9616 init_task_group.shares = init_task_group_load;
9617 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9618 #ifdef CONFIG_CGROUP_SCHED
9620 * How much cpu bandwidth does init_task_group get?
9622 * In case of task-groups formed thr' the cgroup filesystem, it
9623 * gets 100% of the cpu resources in the system. This overall
9624 * system cpu resource is divided among the tasks of
9625 * init_task_group and its child task-groups in a fair manner,
9626 * based on each entity's (task or task-group's) weight
9627 * (se->load.weight).
9629 * In other words, if init_task_group has 10 tasks of weight
9630 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9631 * then A0's share of the cpu resource is:
9633 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9635 * We achieve this by letting init_task_group's tasks sit
9636 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9638 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9639 #elif defined CONFIG_USER_SCHED
9640 root_task_group.shares = NICE_0_LOAD;
9641 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9643 * In case of task-groups formed thr' the user id of tasks,
9644 * init_task_group represents tasks belonging to root user.
9645 * Hence it forms a sibling of all subsequent groups formed.
9646 * In this case, init_task_group gets only a fraction of overall
9647 * system cpu resource, based on the weight assigned to root
9648 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9649 * by letting tasks of init_task_group sit in a separate cfs_rq
9650 * (init_tg_cfs_rq) and having one entity represent this group of
9651 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9653 init_tg_cfs_entry(&init_task_group,
9654 &per_cpu(init_tg_cfs_rq, i),
9655 &per_cpu(init_sched_entity, i), i, 1,
9656 root_task_group.se[i]);
9659 #endif /* CONFIG_FAIR_GROUP_SCHED */
9661 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9662 #ifdef CONFIG_RT_GROUP_SCHED
9663 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9664 #ifdef CONFIG_CGROUP_SCHED
9665 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9666 #elif defined CONFIG_USER_SCHED
9667 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9668 init_tg_rt_entry(&init_task_group,
9669 &per_cpu(init_rt_rq, i),
9670 &per_cpu(init_sched_rt_entity, i), i, 1,
9671 root_task_group.rt_se[i]);
9675 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9676 rq->cpu_load[j] = 0;
9680 rq->post_schedule = 0;
9681 rq->active_balance = 0;
9682 rq->next_balance = jiffies;
9686 rq->migration_thread = NULL;
9688 rq->avg_idle = 2*sysctl_sched_migration_cost;
9689 INIT_LIST_HEAD(&rq->migration_queue);
9690 rq_attach_root(rq, &def_root_domain);
9693 atomic_set(&rq->nr_iowait, 0);
9696 set_load_weight(&init_task);
9698 #ifdef CONFIG_PREEMPT_NOTIFIERS
9699 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9703 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9706 #ifdef CONFIG_RT_MUTEXES
9707 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9711 * The boot idle thread does lazy MMU switching as well:
9713 atomic_inc(&init_mm.mm_count);
9714 enter_lazy_tlb(&init_mm, current);
9717 * Make us the idle thread. Technically, schedule() should not be
9718 * called from this thread, however somewhere below it might be,
9719 * but because we are the idle thread, we just pick up running again
9720 * when this runqueue becomes "idle".
9722 init_idle(current, smp_processor_id());
9724 calc_load_update = jiffies + LOAD_FREQ;
9727 * During early bootup we pretend to be a normal task:
9729 current->sched_class = &fair_sched_class;
9731 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9732 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9735 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9736 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9738 /* May be allocated at isolcpus cmdline parse time */
9739 if (cpu_isolated_map == NULL)
9740 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9745 scheduler_running = 1;
9748 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9749 static inline int preempt_count_equals(int preempt_offset)
9751 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9753 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9756 void __might_sleep(char *file, int line, int preempt_offset)
9759 static unsigned long prev_jiffy; /* ratelimiting */
9761 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9762 system_state != SYSTEM_RUNNING || oops_in_progress)
9764 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9766 prev_jiffy = jiffies;
9769 "BUG: sleeping function called from invalid context at %s:%d\n",
9772 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9773 in_atomic(), irqs_disabled(),
9774 current->pid, current->comm);
9776 debug_show_held_locks(current);
9777 if (irqs_disabled())
9778 print_irqtrace_events(current);
9782 EXPORT_SYMBOL(__might_sleep);
9785 #ifdef CONFIG_MAGIC_SYSRQ
9786 static void normalize_task(struct rq *rq, struct task_struct *p)
9790 update_rq_clock(rq);
9791 on_rq = p->se.on_rq;
9793 deactivate_task(rq, p, 0);
9794 __setscheduler(rq, p, SCHED_NORMAL, 0);
9796 activate_task(rq, p, 0);
9797 resched_task(rq->curr);
9801 void normalize_rt_tasks(void)
9803 struct task_struct *g, *p;
9804 unsigned long flags;
9807 read_lock_irqsave(&tasklist_lock, flags);
9808 do_each_thread(g, p) {
9810 * Only normalize user tasks:
9815 p->se.exec_start = 0;
9816 #ifdef CONFIG_SCHEDSTATS
9817 p->se.wait_start = 0;
9818 p->se.sleep_start = 0;
9819 p->se.block_start = 0;
9824 * Renice negative nice level userspace
9827 if (TASK_NICE(p) < 0 && p->mm)
9828 set_user_nice(p, 0);
9832 spin_lock(&p->pi_lock);
9833 rq = __task_rq_lock(p);
9835 normalize_task(rq, p);
9837 __task_rq_unlock(rq);
9838 spin_unlock(&p->pi_lock);
9839 } while_each_thread(g, p);
9841 read_unlock_irqrestore(&tasklist_lock, flags);
9844 #endif /* CONFIG_MAGIC_SYSRQ */
9848 * These functions are only useful for the IA64 MCA handling.
9850 * They can only be called when the whole system has been
9851 * stopped - every CPU needs to be quiescent, and no scheduling
9852 * activity can take place. Using them for anything else would
9853 * be a serious bug, and as a result, they aren't even visible
9854 * under any other configuration.
9858 * curr_task - return the current task for a given cpu.
9859 * @cpu: the processor in question.
9861 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9863 struct task_struct *curr_task(int cpu)
9865 return cpu_curr(cpu);
9869 * set_curr_task - set the current task for a given cpu.
9870 * @cpu: the processor in question.
9871 * @p: the task pointer to set.
9873 * Description: This function must only be used when non-maskable interrupts
9874 * are serviced on a separate stack. It allows the architecture to switch the
9875 * notion of the current task on a cpu in a non-blocking manner. This function
9876 * must be called with all CPU's synchronized, and interrupts disabled, the
9877 * and caller must save the original value of the current task (see
9878 * curr_task() above) and restore that value before reenabling interrupts and
9879 * re-starting the system.
9881 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9883 void set_curr_task(int cpu, struct task_struct *p)
9890 #ifdef CONFIG_FAIR_GROUP_SCHED
9891 static void free_fair_sched_group(struct task_group *tg)
9895 for_each_possible_cpu(i) {
9897 kfree(tg->cfs_rq[i]);
9907 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9909 struct cfs_rq *cfs_rq;
9910 struct sched_entity *se;
9914 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9917 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9921 tg->shares = NICE_0_LOAD;
9923 for_each_possible_cpu(i) {
9926 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9927 GFP_KERNEL, cpu_to_node(i));
9931 se = kzalloc_node(sizeof(struct sched_entity),
9932 GFP_KERNEL, cpu_to_node(i));
9936 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9945 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9947 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9948 &cpu_rq(cpu)->leaf_cfs_rq_list);
9951 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9953 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9955 #else /* !CONFG_FAIR_GROUP_SCHED */
9956 static inline void free_fair_sched_group(struct task_group *tg)
9961 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9966 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9970 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9973 #endif /* CONFIG_FAIR_GROUP_SCHED */
9975 #ifdef CONFIG_RT_GROUP_SCHED
9976 static void free_rt_sched_group(struct task_group *tg)
9980 destroy_rt_bandwidth(&tg->rt_bandwidth);
9982 for_each_possible_cpu(i) {
9984 kfree(tg->rt_rq[i]);
9986 kfree(tg->rt_se[i]);
9994 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9996 struct rt_rq *rt_rq;
9997 struct sched_rt_entity *rt_se;
10001 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
10004 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
10008 init_rt_bandwidth(&tg->rt_bandwidth,
10009 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
10011 for_each_possible_cpu(i) {
10014 rt_rq = kzalloc_node(sizeof(struct rt_rq),
10015 GFP_KERNEL, cpu_to_node(i));
10019 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
10020 GFP_KERNEL, cpu_to_node(i));
10024 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
10033 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10035 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
10036 &cpu_rq(cpu)->leaf_rt_rq_list);
10039 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10041 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
10043 #else /* !CONFIG_RT_GROUP_SCHED */
10044 static inline void free_rt_sched_group(struct task_group *tg)
10049 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
10054 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10058 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10061 #endif /* CONFIG_RT_GROUP_SCHED */
10063 #ifdef CONFIG_GROUP_SCHED
10064 static void free_sched_group(struct task_group *tg)
10066 free_fair_sched_group(tg);
10067 free_rt_sched_group(tg);
10071 /* allocate runqueue etc for a new task group */
10072 struct task_group *sched_create_group(struct task_group *parent)
10074 struct task_group *tg;
10075 unsigned long flags;
10078 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
10080 return ERR_PTR(-ENOMEM);
10082 if (!alloc_fair_sched_group(tg, parent))
10085 if (!alloc_rt_sched_group(tg, parent))
10088 spin_lock_irqsave(&task_group_lock, flags);
10089 for_each_possible_cpu(i) {
10090 register_fair_sched_group(tg, i);
10091 register_rt_sched_group(tg, i);
10093 list_add_rcu(&tg->list, &task_groups);
10095 WARN_ON(!parent); /* root should already exist */
10097 tg->parent = parent;
10098 INIT_LIST_HEAD(&tg->children);
10099 list_add_rcu(&tg->siblings, &parent->children);
10100 spin_unlock_irqrestore(&task_group_lock, flags);
10105 free_sched_group(tg);
10106 return ERR_PTR(-ENOMEM);
10109 /* rcu callback to free various structures associated with a task group */
10110 static void free_sched_group_rcu(struct rcu_head *rhp)
10112 /* now it should be safe to free those cfs_rqs */
10113 free_sched_group(container_of(rhp, struct task_group, rcu));
10116 /* Destroy runqueue etc associated with a task group */
10117 void sched_destroy_group(struct task_group *tg)
10119 unsigned long flags;
10122 spin_lock_irqsave(&task_group_lock, flags);
10123 for_each_possible_cpu(i) {
10124 unregister_fair_sched_group(tg, i);
10125 unregister_rt_sched_group(tg, i);
10127 list_del_rcu(&tg->list);
10128 list_del_rcu(&tg->siblings);
10129 spin_unlock_irqrestore(&task_group_lock, flags);
10131 /* wait for possible concurrent references to cfs_rqs complete */
10132 call_rcu(&tg->rcu, free_sched_group_rcu);
10135 /* change task's runqueue when it moves between groups.
10136 * The caller of this function should have put the task in its new group
10137 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10138 * reflect its new group.
10140 void sched_move_task(struct task_struct *tsk)
10142 int on_rq, running;
10143 unsigned long flags;
10146 rq = task_rq_lock(tsk, &flags);
10148 update_rq_clock(rq);
10150 running = task_current(rq, tsk);
10151 on_rq = tsk->se.on_rq;
10154 dequeue_task(rq, tsk, 0);
10155 if (unlikely(running))
10156 tsk->sched_class->put_prev_task(rq, tsk);
10158 set_task_rq(tsk, task_cpu(tsk));
10160 #ifdef CONFIG_FAIR_GROUP_SCHED
10161 if (tsk->sched_class->moved_group)
10162 tsk->sched_class->moved_group(tsk, on_rq);
10165 if (unlikely(running))
10166 tsk->sched_class->set_curr_task(rq);
10168 enqueue_task(rq, tsk, 0, false);
10170 task_rq_unlock(rq, &flags);
10172 #endif /* CONFIG_GROUP_SCHED */
10174 #ifdef CONFIG_FAIR_GROUP_SCHED
10175 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10177 struct cfs_rq *cfs_rq = se->cfs_rq;
10182 dequeue_entity(cfs_rq, se, 0);
10184 se->load.weight = shares;
10185 se->load.inv_weight = 0;
10188 enqueue_entity(cfs_rq, se, 0);
10191 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10193 struct cfs_rq *cfs_rq = se->cfs_rq;
10194 struct rq *rq = cfs_rq->rq;
10195 unsigned long flags;
10197 spin_lock_irqsave(&rq->lock, flags);
10198 __set_se_shares(se, shares);
10199 spin_unlock_irqrestore(&rq->lock, flags);
10202 static DEFINE_MUTEX(shares_mutex);
10204 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10207 unsigned long flags;
10210 * We can't change the weight of the root cgroup.
10215 if (shares < MIN_SHARES)
10216 shares = MIN_SHARES;
10217 else if (shares > MAX_SHARES)
10218 shares = MAX_SHARES;
10220 mutex_lock(&shares_mutex);
10221 if (tg->shares == shares)
10224 spin_lock_irqsave(&task_group_lock, flags);
10225 for_each_possible_cpu(i)
10226 unregister_fair_sched_group(tg, i);
10227 list_del_rcu(&tg->siblings);
10228 spin_unlock_irqrestore(&task_group_lock, flags);
10230 /* wait for any ongoing reference to this group to finish */
10231 synchronize_sched();
10234 * Now we are free to modify the group's share on each cpu
10235 * w/o tripping rebalance_share or load_balance_fair.
10237 tg->shares = shares;
10238 for_each_possible_cpu(i) {
10240 * force a rebalance
10242 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10243 set_se_shares(tg->se[i], shares);
10247 * Enable load balance activity on this group, by inserting it back on
10248 * each cpu's rq->leaf_cfs_rq_list.
10250 spin_lock_irqsave(&task_group_lock, flags);
10251 for_each_possible_cpu(i)
10252 register_fair_sched_group(tg, i);
10253 list_add_rcu(&tg->siblings, &tg->parent->children);
10254 spin_unlock_irqrestore(&task_group_lock, flags);
10256 mutex_unlock(&shares_mutex);
10260 unsigned long sched_group_shares(struct task_group *tg)
10266 #ifdef CONFIG_RT_GROUP_SCHED
10268 * Ensure that the real time constraints are schedulable.
10270 static DEFINE_MUTEX(rt_constraints_mutex);
10272 static unsigned long to_ratio(u64 period, u64 runtime)
10274 if (runtime == RUNTIME_INF)
10277 return div64_u64(runtime << 20, period);
10280 /* Must be called with tasklist_lock held */
10281 static inline int tg_has_rt_tasks(struct task_group *tg)
10283 struct task_struct *g, *p;
10285 do_each_thread(g, p) {
10286 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10288 } while_each_thread(g, p);
10293 struct rt_schedulable_data {
10294 struct task_group *tg;
10299 static int tg_schedulable(struct task_group *tg, void *data)
10301 struct rt_schedulable_data *d = data;
10302 struct task_group *child;
10303 unsigned long total, sum = 0;
10304 u64 period, runtime;
10306 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10307 runtime = tg->rt_bandwidth.rt_runtime;
10310 period = d->rt_period;
10311 runtime = d->rt_runtime;
10314 #ifdef CONFIG_USER_SCHED
10315 if (tg == &root_task_group) {
10316 period = global_rt_period();
10317 runtime = global_rt_runtime();
10322 * Cannot have more runtime than the period.
10324 if (runtime > period && runtime != RUNTIME_INF)
10328 * Ensure we don't starve existing RT tasks.
10330 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10333 total = to_ratio(period, runtime);
10336 * Nobody can have more than the global setting allows.
10338 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10342 * The sum of our children's runtime should not exceed our own.
10344 list_for_each_entry_rcu(child, &tg->children, siblings) {
10345 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10346 runtime = child->rt_bandwidth.rt_runtime;
10348 if (child == d->tg) {
10349 period = d->rt_period;
10350 runtime = d->rt_runtime;
10353 sum += to_ratio(period, runtime);
10362 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10364 struct rt_schedulable_data data = {
10366 .rt_period = period,
10367 .rt_runtime = runtime,
10370 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10373 static int tg_set_bandwidth(struct task_group *tg,
10374 u64 rt_period, u64 rt_runtime)
10378 mutex_lock(&rt_constraints_mutex);
10379 read_lock(&tasklist_lock);
10380 err = __rt_schedulable(tg, rt_period, rt_runtime);
10384 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10385 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10386 tg->rt_bandwidth.rt_runtime = rt_runtime;
10388 for_each_possible_cpu(i) {
10389 struct rt_rq *rt_rq = tg->rt_rq[i];
10391 spin_lock(&rt_rq->rt_runtime_lock);
10392 rt_rq->rt_runtime = rt_runtime;
10393 spin_unlock(&rt_rq->rt_runtime_lock);
10395 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10397 read_unlock(&tasklist_lock);
10398 mutex_unlock(&rt_constraints_mutex);
10403 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10405 u64 rt_runtime, rt_period;
10407 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10408 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10409 if (rt_runtime_us < 0)
10410 rt_runtime = RUNTIME_INF;
10412 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10415 long sched_group_rt_runtime(struct task_group *tg)
10419 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10422 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10423 do_div(rt_runtime_us, NSEC_PER_USEC);
10424 return rt_runtime_us;
10427 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10429 u64 rt_runtime, rt_period;
10431 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10432 rt_runtime = tg->rt_bandwidth.rt_runtime;
10434 if (rt_period == 0)
10437 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10440 long sched_group_rt_period(struct task_group *tg)
10444 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10445 do_div(rt_period_us, NSEC_PER_USEC);
10446 return rt_period_us;
10449 static int sched_rt_global_constraints(void)
10451 u64 runtime, period;
10454 if (sysctl_sched_rt_period <= 0)
10457 runtime = global_rt_runtime();
10458 period = global_rt_period();
10461 * Sanity check on the sysctl variables.
10463 if (runtime > period && runtime != RUNTIME_INF)
10466 mutex_lock(&rt_constraints_mutex);
10467 read_lock(&tasklist_lock);
10468 ret = __rt_schedulable(NULL, 0, 0);
10469 read_unlock(&tasklist_lock);
10470 mutex_unlock(&rt_constraints_mutex);
10475 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10477 /* Don't accept realtime tasks when there is no way for them to run */
10478 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10484 #else /* !CONFIG_RT_GROUP_SCHED */
10485 static int sched_rt_global_constraints(void)
10487 unsigned long flags;
10490 if (sysctl_sched_rt_period <= 0)
10494 * There's always some RT tasks in the root group
10495 * -- migration, kstopmachine etc..
10497 if (sysctl_sched_rt_runtime == 0)
10500 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10501 for_each_possible_cpu(i) {
10502 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10504 spin_lock(&rt_rq->rt_runtime_lock);
10505 rt_rq->rt_runtime = global_rt_runtime();
10506 spin_unlock(&rt_rq->rt_runtime_lock);
10508 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10512 #endif /* CONFIG_RT_GROUP_SCHED */
10514 int sched_rt_handler(struct ctl_table *table, int write,
10515 void __user *buffer, size_t *lenp,
10519 int old_period, old_runtime;
10520 static DEFINE_MUTEX(mutex);
10522 mutex_lock(&mutex);
10523 old_period = sysctl_sched_rt_period;
10524 old_runtime = sysctl_sched_rt_runtime;
10526 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10528 if (!ret && write) {
10529 ret = sched_rt_global_constraints();
10531 sysctl_sched_rt_period = old_period;
10532 sysctl_sched_rt_runtime = old_runtime;
10534 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10535 def_rt_bandwidth.rt_period =
10536 ns_to_ktime(global_rt_period());
10539 mutex_unlock(&mutex);
10544 #ifdef CONFIG_CGROUP_SCHED
10546 /* return corresponding task_group object of a cgroup */
10547 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10549 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10550 struct task_group, css);
10553 static struct cgroup_subsys_state *
10554 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10556 struct task_group *tg, *parent;
10558 if (!cgrp->parent) {
10559 /* This is early initialization for the top cgroup */
10560 return &init_task_group.css;
10563 parent = cgroup_tg(cgrp->parent);
10564 tg = sched_create_group(parent);
10566 return ERR_PTR(-ENOMEM);
10572 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10574 struct task_group *tg = cgroup_tg(cgrp);
10576 sched_destroy_group(tg);
10580 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10582 #ifdef CONFIG_RT_GROUP_SCHED
10583 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10586 /* We don't support RT-tasks being in separate groups */
10587 if (tsk->sched_class != &fair_sched_class)
10594 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10595 struct task_struct *tsk, bool threadgroup)
10597 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10601 struct task_struct *c;
10603 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10604 retval = cpu_cgroup_can_attach_task(cgrp, c);
10616 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10617 struct cgroup *old_cont, struct task_struct *tsk,
10620 sched_move_task(tsk);
10622 struct task_struct *c;
10624 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10625 sched_move_task(c);
10631 #ifdef CONFIG_FAIR_GROUP_SCHED
10632 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10635 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10638 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10640 struct task_group *tg = cgroup_tg(cgrp);
10642 return (u64) tg->shares;
10644 #endif /* CONFIG_FAIR_GROUP_SCHED */
10646 #ifdef CONFIG_RT_GROUP_SCHED
10647 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10650 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10653 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10655 return sched_group_rt_runtime(cgroup_tg(cgrp));
10658 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10661 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10664 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10666 return sched_group_rt_period(cgroup_tg(cgrp));
10668 #endif /* CONFIG_RT_GROUP_SCHED */
10670 static struct cftype cpu_files[] = {
10671 #ifdef CONFIG_FAIR_GROUP_SCHED
10674 .read_u64 = cpu_shares_read_u64,
10675 .write_u64 = cpu_shares_write_u64,
10678 #ifdef CONFIG_RT_GROUP_SCHED
10680 .name = "rt_runtime_us",
10681 .read_s64 = cpu_rt_runtime_read,
10682 .write_s64 = cpu_rt_runtime_write,
10685 .name = "rt_period_us",
10686 .read_u64 = cpu_rt_period_read_uint,
10687 .write_u64 = cpu_rt_period_write_uint,
10692 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10694 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10697 struct cgroup_subsys cpu_cgroup_subsys = {
10699 .create = cpu_cgroup_create,
10700 .destroy = cpu_cgroup_destroy,
10701 .can_attach = cpu_cgroup_can_attach,
10702 .attach = cpu_cgroup_attach,
10703 .populate = cpu_cgroup_populate,
10704 .subsys_id = cpu_cgroup_subsys_id,
10708 #endif /* CONFIG_CGROUP_SCHED */
10710 #ifdef CONFIG_CGROUP_CPUACCT
10713 * CPU accounting code for task groups.
10715 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10716 * (balbir@in.ibm.com).
10719 /* track cpu usage of a group of tasks and its child groups */
10721 struct cgroup_subsys_state css;
10722 /* cpuusage holds pointer to a u64-type object on every cpu */
10724 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10725 struct cpuacct *parent;
10728 struct cgroup_subsys cpuacct_subsys;
10730 /* return cpu accounting group corresponding to this container */
10731 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10733 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10734 struct cpuacct, css);
10737 /* return cpu accounting group to which this task belongs */
10738 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10740 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10741 struct cpuacct, css);
10744 /* create a new cpu accounting group */
10745 static struct cgroup_subsys_state *cpuacct_create(
10746 struct cgroup_subsys *ss, struct cgroup *cgrp)
10748 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10754 ca->cpuusage = alloc_percpu(u64);
10758 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10759 if (percpu_counter_init(&ca->cpustat[i], 0))
10760 goto out_free_counters;
10763 ca->parent = cgroup_ca(cgrp->parent);
10769 percpu_counter_destroy(&ca->cpustat[i]);
10770 free_percpu(ca->cpuusage);
10774 return ERR_PTR(-ENOMEM);
10777 /* destroy an existing cpu accounting group */
10779 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10781 struct cpuacct *ca = cgroup_ca(cgrp);
10784 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10785 percpu_counter_destroy(&ca->cpustat[i]);
10786 free_percpu(ca->cpuusage);
10790 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10792 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10795 #ifndef CONFIG_64BIT
10797 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10799 spin_lock_irq(&cpu_rq(cpu)->lock);
10801 spin_unlock_irq(&cpu_rq(cpu)->lock);
10809 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10811 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10813 #ifndef CONFIG_64BIT
10815 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10817 spin_lock_irq(&cpu_rq(cpu)->lock);
10819 spin_unlock_irq(&cpu_rq(cpu)->lock);
10825 /* return total cpu usage (in nanoseconds) of a group */
10826 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10828 struct cpuacct *ca = cgroup_ca(cgrp);
10829 u64 totalcpuusage = 0;
10832 for_each_present_cpu(i)
10833 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10835 return totalcpuusage;
10838 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10841 struct cpuacct *ca = cgroup_ca(cgrp);
10850 for_each_present_cpu(i)
10851 cpuacct_cpuusage_write(ca, i, 0);
10857 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10858 struct seq_file *m)
10860 struct cpuacct *ca = cgroup_ca(cgroup);
10864 for_each_present_cpu(i) {
10865 percpu = cpuacct_cpuusage_read(ca, i);
10866 seq_printf(m, "%llu ", (unsigned long long) percpu);
10868 seq_printf(m, "\n");
10872 static const char *cpuacct_stat_desc[] = {
10873 [CPUACCT_STAT_USER] = "user",
10874 [CPUACCT_STAT_SYSTEM] = "system",
10877 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10878 struct cgroup_map_cb *cb)
10880 struct cpuacct *ca = cgroup_ca(cgrp);
10883 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10884 s64 val = percpu_counter_read(&ca->cpustat[i]);
10885 val = cputime64_to_clock_t(val);
10886 cb->fill(cb, cpuacct_stat_desc[i], val);
10891 static struct cftype files[] = {
10894 .read_u64 = cpuusage_read,
10895 .write_u64 = cpuusage_write,
10898 .name = "usage_percpu",
10899 .read_seq_string = cpuacct_percpu_seq_read,
10903 .read_map = cpuacct_stats_show,
10907 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10909 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10913 * charge this task's execution time to its accounting group.
10915 * called with rq->lock held.
10917 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10919 struct cpuacct *ca;
10922 if (unlikely(!cpuacct_subsys.active))
10925 cpu = task_cpu(tsk);
10931 for (; ca; ca = ca->parent) {
10932 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10933 *cpuusage += cputime;
10940 * Charge the system/user time to the task's accounting group.
10942 static void cpuacct_update_stats(struct task_struct *tsk,
10943 enum cpuacct_stat_index idx, cputime_t val)
10945 struct cpuacct *ca;
10947 if (unlikely(!cpuacct_subsys.active))
10954 percpu_counter_add(&ca->cpustat[idx], val);
10960 struct cgroup_subsys cpuacct_subsys = {
10962 .create = cpuacct_create,
10963 .destroy = cpuacct_destroy,
10964 .populate = cpuacct_populate,
10965 .subsys_id = cpuacct_subsys_id,
10967 #endif /* CONFIG_CGROUP_CPUACCT */
10971 int rcu_expedited_torture_stats(char *page)
10975 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10977 void synchronize_sched_expedited(void)
10980 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10982 #else /* #ifndef CONFIG_SMP */
10984 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10985 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10987 #define RCU_EXPEDITED_STATE_POST -2
10988 #define RCU_EXPEDITED_STATE_IDLE -1
10990 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10992 int rcu_expedited_torture_stats(char *page)
10997 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10998 for_each_online_cpu(cpu) {
10999 cnt += sprintf(&page[cnt], " %d:%d",
11000 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
11002 cnt += sprintf(&page[cnt], "\n");
11005 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
11007 static long synchronize_sched_expedited_count;
11010 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
11011 * approach to force grace period to end quickly. This consumes
11012 * significant time on all CPUs, and is thus not recommended for
11013 * any sort of common-case code.
11015 * Note that it is illegal to call this function while holding any
11016 * lock that is acquired by a CPU-hotplug notifier. Failing to
11017 * observe this restriction will result in deadlock.
11019 void synchronize_sched_expedited(void)
11022 unsigned long flags;
11023 bool need_full_sync = 0;
11025 struct migration_req *req;
11029 smp_mb(); /* ensure prior mod happens before capturing snap. */
11030 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
11032 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
11034 if (trycount++ < 10)
11035 udelay(trycount * num_online_cpus());
11037 synchronize_sched();
11040 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
11041 smp_mb(); /* ensure test happens before caller kfree */
11046 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
11047 for_each_online_cpu(cpu) {
11049 req = &per_cpu(rcu_migration_req, cpu);
11050 init_completion(&req->done);
11052 req->dest_cpu = RCU_MIGRATION_NEED_QS;
11053 spin_lock_irqsave(&rq->lock, flags);
11054 list_add(&req->list, &rq->migration_queue);
11055 spin_unlock_irqrestore(&rq->lock, flags);
11056 wake_up_process(rq->migration_thread);
11058 for_each_online_cpu(cpu) {
11059 rcu_expedited_state = cpu;
11060 req = &per_cpu(rcu_migration_req, cpu);
11062 wait_for_completion(&req->done);
11063 spin_lock_irqsave(&rq->lock, flags);
11064 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
11065 need_full_sync = 1;
11066 req->dest_cpu = RCU_MIGRATION_IDLE;
11067 spin_unlock_irqrestore(&rq->lock, flags);
11069 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
11070 mutex_unlock(&rcu_sched_expedited_mutex);
11072 if (need_full_sync)
11073 synchronize_sched();
11075 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
11077 #endif /* #else #ifndef CONFIG_SMP */