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 raw_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 raw_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 raw_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 raw_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_var);
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 raw_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 char in_nohz_recently;
540 /* capture load from *all* tasks on this cpu: */
541 struct load_weight load;
542 unsigned long nr_load_updates;
548 #ifdef CONFIG_FAIR_GROUP_SCHED
549 /* list of leaf cfs_rq on this cpu: */
550 struct list_head leaf_cfs_rq_list;
552 #ifdef CONFIG_RT_GROUP_SCHED
553 struct list_head leaf_rt_rq_list;
557 * This is part of a global counter where only the total sum
558 * over all CPUs matters. A task can increase this counter on
559 * one CPU and if it got migrated afterwards it may decrease
560 * it on another CPU. Always updated under the runqueue lock:
562 unsigned long nr_uninterruptible;
564 struct task_struct *curr, *idle;
565 unsigned long next_balance;
566 struct mm_struct *prev_mm;
573 struct root_domain *rd;
574 struct sched_domain *sd;
576 unsigned char idle_at_tick;
577 /* For active balancing */
581 /* cpu of this runqueue: */
585 unsigned long avg_load_per_task;
587 struct task_struct *migration_thread;
588 struct list_head migration_queue;
596 /* calc_load related fields */
597 unsigned long calc_load_update;
598 long calc_load_active;
600 #ifdef CONFIG_SCHED_HRTICK
602 int hrtick_csd_pending;
603 struct call_single_data hrtick_csd;
605 struct hrtimer hrtick_timer;
608 #ifdef CONFIG_SCHEDSTATS
610 struct sched_info rq_sched_info;
611 unsigned long long rq_cpu_time;
612 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
614 /* sys_sched_yield() stats */
615 unsigned int yld_count;
617 /* schedule() stats */
618 unsigned int sched_switch;
619 unsigned int sched_count;
620 unsigned int sched_goidle;
622 /* try_to_wake_up() stats */
623 unsigned int ttwu_count;
624 unsigned int ttwu_local;
627 unsigned int bkl_count;
631 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
634 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
636 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
639 static inline int cpu_of(struct rq *rq)
649 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
650 * See detach_destroy_domains: synchronize_sched for details.
652 * The domain tree of any CPU may only be accessed from within
653 * preempt-disabled sections.
655 #define for_each_domain(cpu, __sd) \
656 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
658 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
659 #define this_rq() (&__get_cpu_var(runqueues))
660 #define task_rq(p) cpu_rq(task_cpu(p))
661 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
662 #define raw_rq() (&__raw_get_cpu_var(runqueues))
664 inline void update_rq_clock(struct rq *rq)
666 rq->clock = sched_clock_cpu(cpu_of(rq));
670 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
672 #ifdef CONFIG_SCHED_DEBUG
673 # define const_debug __read_mostly
675 # define const_debug static const
680 * @cpu: the processor in question.
682 * Returns true if the current cpu runqueue is locked.
683 * This interface allows printk to be called with the runqueue lock
684 * held and know whether or not it is OK to wake up the klogd.
686 int runqueue_is_locked(int cpu)
688 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
692 * Debugging: various feature bits
695 #define SCHED_FEAT(name, enabled) \
696 __SCHED_FEAT_##name ,
699 #include "sched_features.h"
704 #define SCHED_FEAT(name, enabled) \
705 (1UL << __SCHED_FEAT_##name) * enabled |
707 const_debug unsigned int sysctl_sched_features =
708 #include "sched_features.h"
713 #ifdef CONFIG_SCHED_DEBUG
714 #define SCHED_FEAT(name, enabled) \
717 static __read_mostly char *sched_feat_names[] = {
718 #include "sched_features.h"
724 static int sched_feat_show(struct seq_file *m, void *v)
728 for (i = 0; sched_feat_names[i]; i++) {
729 if (!(sysctl_sched_features & (1UL << i)))
731 seq_printf(m, "%s ", sched_feat_names[i]);
739 sched_feat_write(struct file *filp, const char __user *ubuf,
740 size_t cnt, loff_t *ppos)
750 if (copy_from_user(&buf, ubuf, cnt))
755 if (strncmp(buf, "NO_", 3) == 0) {
760 for (i = 0; sched_feat_names[i]; i++) {
761 int len = strlen(sched_feat_names[i]);
763 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
765 sysctl_sched_features &= ~(1UL << i);
767 sysctl_sched_features |= (1UL << i);
772 if (!sched_feat_names[i])
780 static int sched_feat_open(struct inode *inode, struct file *filp)
782 return single_open(filp, sched_feat_show, NULL);
785 static const struct file_operations sched_feat_fops = {
786 .open = sched_feat_open,
787 .write = sched_feat_write,
790 .release = single_release,
793 static __init int sched_init_debug(void)
795 debugfs_create_file("sched_features", 0644, NULL, NULL,
800 late_initcall(sched_init_debug);
804 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
807 * Number of tasks to iterate in a single balance run.
808 * Limited because this is done with IRQs disabled.
810 const_debug unsigned int sysctl_sched_nr_migrate = 32;
813 * ratelimit for updating the group shares.
816 unsigned int sysctl_sched_shares_ratelimit = 250000;
817 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
820 * Inject some fuzzyness into changing the per-cpu group shares
821 * this avoids remote rq-locks at the expense of fairness.
824 unsigned int sysctl_sched_shares_thresh = 4;
827 * period over which we average the RT time consumption, measured
832 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
835 * period over which we measure -rt task cpu usage in us.
838 unsigned int sysctl_sched_rt_period = 1000000;
840 static __read_mostly int scheduler_running;
843 * part of the period that we allow rt tasks to run in us.
846 int sysctl_sched_rt_runtime = 950000;
848 static inline u64 global_rt_period(void)
850 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
853 static inline u64 global_rt_runtime(void)
855 if (sysctl_sched_rt_runtime < 0)
858 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
861 #ifndef prepare_arch_switch
862 # define prepare_arch_switch(next) do { } while (0)
864 #ifndef finish_arch_switch
865 # define finish_arch_switch(prev) do { } while (0)
868 static inline int task_current(struct rq *rq, struct task_struct *p)
870 return rq->curr == p;
873 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
874 static inline int task_running(struct rq *rq, struct task_struct *p)
876 return task_current(rq, p);
879 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
883 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
885 #ifdef CONFIG_DEBUG_SPINLOCK
886 /* this is a valid case when another task releases the spinlock */
887 rq->lock.owner = current;
890 * If we are tracking spinlock dependencies then we have to
891 * fix up the runqueue lock - which gets 'carried over' from
894 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
896 raw_spin_unlock_irq(&rq->lock);
899 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
900 static inline int task_running(struct rq *rq, struct task_struct *p)
905 return task_current(rq, p);
909 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
913 * We can optimise this out completely for !SMP, because the
914 * SMP rebalancing from interrupt is the only thing that cares
919 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
920 raw_spin_unlock_irq(&rq->lock);
922 raw_spin_unlock(&rq->lock);
926 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
930 * After ->oncpu is cleared, the task can be moved to a different CPU.
931 * We must ensure this doesn't happen until the switch is completely
937 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
941 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
944 * Check whether the task is waking, we use this to synchronize against
945 * ttwu() so that task_cpu() reports a stable number.
947 * We need to make an exception for PF_STARTING tasks because the fork
948 * path might require task_rq_lock() to work, eg. it can call
949 * set_cpus_allowed_ptr() from the cpuset clone_ns code.
951 static inline int task_is_waking(struct task_struct *p)
953 return unlikely((p->state == TASK_WAKING) && !(p->flags & PF_STARTING));
957 * __task_rq_lock - lock the runqueue a given task resides on.
958 * Must be called interrupts disabled.
960 static inline struct rq *__task_rq_lock(struct task_struct *p)
966 while (task_is_waking(p))
969 raw_spin_lock(&rq->lock);
970 if (likely(rq == task_rq(p) && !task_is_waking(p)))
972 raw_spin_unlock(&rq->lock);
977 * task_rq_lock - lock the runqueue a given task resides on and disable
978 * interrupts. Note the ordering: we can safely lookup the task_rq without
979 * explicitly disabling preemption.
981 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
987 while (task_is_waking(p))
989 local_irq_save(*flags);
991 raw_spin_lock(&rq->lock);
992 if (likely(rq == task_rq(p) && !task_is_waking(p)))
994 raw_spin_unlock_irqrestore(&rq->lock, *flags);
998 void task_rq_unlock_wait(struct task_struct *p)
1000 struct rq *rq = task_rq(p);
1002 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1003 raw_spin_unlock_wait(&rq->lock);
1006 static void __task_rq_unlock(struct rq *rq)
1007 __releases(rq->lock)
1009 raw_spin_unlock(&rq->lock);
1012 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1013 __releases(rq->lock)
1015 raw_spin_unlock_irqrestore(&rq->lock, *flags);
1019 * this_rq_lock - lock this runqueue and disable interrupts.
1021 static struct rq *this_rq_lock(void)
1022 __acquires(rq->lock)
1026 local_irq_disable();
1028 raw_spin_lock(&rq->lock);
1033 #ifdef CONFIG_SCHED_HRTICK
1035 * Use HR-timers to deliver accurate preemption points.
1037 * Its all a bit involved since we cannot program an hrt while holding the
1038 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1041 * When we get rescheduled we reprogram the hrtick_timer outside of the
1047 * - enabled by features
1048 * - hrtimer is actually high res
1050 static inline int hrtick_enabled(struct rq *rq)
1052 if (!sched_feat(HRTICK))
1054 if (!cpu_active(cpu_of(rq)))
1056 return hrtimer_is_hres_active(&rq->hrtick_timer);
1059 static void hrtick_clear(struct rq *rq)
1061 if (hrtimer_active(&rq->hrtick_timer))
1062 hrtimer_cancel(&rq->hrtick_timer);
1066 * High-resolution timer tick.
1067 * Runs from hardirq context with interrupts disabled.
1069 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1071 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1073 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1075 raw_spin_lock(&rq->lock);
1076 update_rq_clock(rq);
1077 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1078 raw_spin_unlock(&rq->lock);
1080 return HRTIMER_NORESTART;
1085 * called from hardirq (IPI) context
1087 static void __hrtick_start(void *arg)
1089 struct rq *rq = arg;
1091 raw_spin_lock(&rq->lock);
1092 hrtimer_restart(&rq->hrtick_timer);
1093 rq->hrtick_csd_pending = 0;
1094 raw_spin_unlock(&rq->lock);
1098 * Called to set the hrtick timer state.
1100 * called with rq->lock held and irqs disabled
1102 static void hrtick_start(struct rq *rq, u64 delay)
1104 struct hrtimer *timer = &rq->hrtick_timer;
1105 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1107 hrtimer_set_expires(timer, time);
1109 if (rq == this_rq()) {
1110 hrtimer_restart(timer);
1111 } else if (!rq->hrtick_csd_pending) {
1112 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1113 rq->hrtick_csd_pending = 1;
1118 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1120 int cpu = (int)(long)hcpu;
1123 case CPU_UP_CANCELED:
1124 case CPU_UP_CANCELED_FROZEN:
1125 case CPU_DOWN_PREPARE:
1126 case CPU_DOWN_PREPARE_FROZEN:
1128 case CPU_DEAD_FROZEN:
1129 hrtick_clear(cpu_rq(cpu));
1136 static __init void init_hrtick(void)
1138 hotcpu_notifier(hotplug_hrtick, 0);
1142 * Called to set the hrtick timer state.
1144 * called with rq->lock held and irqs disabled
1146 static void hrtick_start(struct rq *rq, u64 delay)
1148 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1149 HRTIMER_MODE_REL_PINNED, 0);
1152 static inline void init_hrtick(void)
1155 #endif /* CONFIG_SMP */
1157 static void init_rq_hrtick(struct rq *rq)
1160 rq->hrtick_csd_pending = 0;
1162 rq->hrtick_csd.flags = 0;
1163 rq->hrtick_csd.func = __hrtick_start;
1164 rq->hrtick_csd.info = rq;
1167 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1168 rq->hrtick_timer.function = hrtick;
1170 #else /* CONFIG_SCHED_HRTICK */
1171 static inline void hrtick_clear(struct rq *rq)
1175 static inline void init_rq_hrtick(struct rq *rq)
1179 static inline void init_hrtick(void)
1182 #endif /* CONFIG_SCHED_HRTICK */
1185 * resched_task - mark a task 'to be rescheduled now'.
1187 * On UP this means the setting of the need_resched flag, on SMP it
1188 * might also involve a cross-CPU call to trigger the scheduler on
1193 #ifndef tsk_is_polling
1194 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1197 static void resched_task(struct task_struct *p)
1201 assert_raw_spin_locked(&task_rq(p)->lock);
1203 if (test_tsk_need_resched(p))
1206 set_tsk_need_resched(p);
1209 if (cpu == smp_processor_id())
1212 /* NEED_RESCHED must be visible before we test polling */
1214 if (!tsk_is_polling(p))
1215 smp_send_reschedule(cpu);
1218 static void resched_cpu(int cpu)
1220 struct rq *rq = cpu_rq(cpu);
1221 unsigned long flags;
1223 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1225 resched_task(cpu_curr(cpu));
1226 raw_spin_unlock_irqrestore(&rq->lock, flags);
1231 * When add_timer_on() enqueues a timer into the timer wheel of an
1232 * idle CPU then this timer might expire before the next timer event
1233 * which is scheduled to wake up that CPU. In case of a completely
1234 * idle system the next event might even be infinite time into the
1235 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1236 * leaves the inner idle loop so the newly added timer is taken into
1237 * account when the CPU goes back to idle and evaluates the timer
1238 * wheel for the next timer event.
1240 void wake_up_idle_cpu(int cpu)
1242 struct rq *rq = cpu_rq(cpu);
1244 if (cpu == smp_processor_id())
1248 * This is safe, as this function is called with the timer
1249 * wheel base lock of (cpu) held. When the CPU is on the way
1250 * to idle and has not yet set rq->curr to idle then it will
1251 * be serialized on the timer wheel base lock and take the new
1252 * timer into account automatically.
1254 if (rq->curr != rq->idle)
1258 * We can set TIF_RESCHED on the idle task of the other CPU
1259 * lockless. The worst case is that the other CPU runs the
1260 * idle task through an additional NOOP schedule()
1262 set_tsk_need_resched(rq->idle);
1264 /* NEED_RESCHED must be visible before we test polling */
1266 if (!tsk_is_polling(rq->idle))
1267 smp_send_reschedule(cpu);
1269 #endif /* CONFIG_NO_HZ */
1271 static u64 sched_avg_period(void)
1273 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1276 static void sched_avg_update(struct rq *rq)
1278 s64 period = sched_avg_period();
1280 while ((s64)(rq->clock - rq->age_stamp) > period) {
1281 rq->age_stamp += period;
1286 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1288 rq->rt_avg += rt_delta;
1289 sched_avg_update(rq);
1292 #else /* !CONFIG_SMP */
1293 static void resched_task(struct task_struct *p)
1295 assert_raw_spin_locked(&task_rq(p)->lock);
1296 set_tsk_need_resched(p);
1299 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1302 #endif /* CONFIG_SMP */
1304 #if BITS_PER_LONG == 32
1305 # define WMULT_CONST (~0UL)
1307 # define WMULT_CONST (1UL << 32)
1310 #define WMULT_SHIFT 32
1313 * Shift right and round:
1315 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1318 * delta *= weight / lw
1320 static unsigned long
1321 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1322 struct load_weight *lw)
1326 if (!lw->inv_weight) {
1327 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1330 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1334 tmp = (u64)delta_exec * weight;
1336 * Check whether we'd overflow the 64-bit multiplication:
1338 if (unlikely(tmp > WMULT_CONST))
1339 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1342 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1344 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1347 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1353 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1360 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1361 * of tasks with abnormal "nice" values across CPUs the contribution that
1362 * each task makes to its run queue's load is weighted according to its
1363 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1364 * scaled version of the new time slice allocation that they receive on time
1368 #define WEIGHT_IDLEPRIO 3
1369 #define WMULT_IDLEPRIO 1431655765
1372 * Nice levels are multiplicative, with a gentle 10% change for every
1373 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1374 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1375 * that remained on nice 0.
1377 * The "10% effect" is relative and cumulative: from _any_ nice level,
1378 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1379 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1380 * If a task goes up by ~10% and another task goes down by ~10% then
1381 * the relative distance between them is ~25%.)
1383 static const int prio_to_weight[40] = {
1384 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1385 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1386 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1387 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1388 /* 0 */ 1024, 820, 655, 526, 423,
1389 /* 5 */ 335, 272, 215, 172, 137,
1390 /* 10 */ 110, 87, 70, 56, 45,
1391 /* 15 */ 36, 29, 23, 18, 15,
1395 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1397 * In cases where the weight does not change often, we can use the
1398 * precalculated inverse to speed up arithmetics by turning divisions
1399 * into multiplications:
1401 static const u32 prio_to_wmult[40] = {
1402 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1403 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1404 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1405 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1406 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1407 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1408 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1409 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1412 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1415 * runqueue iterator, to support SMP load-balancing between different
1416 * scheduling classes, without having to expose their internal data
1417 * structures to the load-balancing proper:
1419 struct rq_iterator {
1421 struct task_struct *(*start)(void *);
1422 struct task_struct *(*next)(void *);
1426 static unsigned long
1427 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1428 unsigned long max_load_move, struct sched_domain *sd,
1429 enum cpu_idle_type idle, int *all_pinned,
1430 int *this_best_prio, struct rq_iterator *iterator);
1433 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1434 struct sched_domain *sd, enum cpu_idle_type idle,
1435 struct rq_iterator *iterator);
1438 /* Time spent by the tasks of the cpu accounting group executing in ... */
1439 enum cpuacct_stat_index {
1440 CPUACCT_STAT_USER, /* ... user mode */
1441 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1443 CPUACCT_STAT_NSTATS,
1446 #ifdef CONFIG_CGROUP_CPUACCT
1447 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1448 static void cpuacct_update_stats(struct task_struct *tsk,
1449 enum cpuacct_stat_index idx, cputime_t val);
1451 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1452 static inline void cpuacct_update_stats(struct task_struct *tsk,
1453 enum cpuacct_stat_index idx, cputime_t val) {}
1456 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1458 update_load_add(&rq->load, load);
1461 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1463 update_load_sub(&rq->load, load);
1466 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1467 typedef int (*tg_visitor)(struct task_group *, void *);
1470 * Iterate the full tree, calling @down when first entering a node and @up when
1471 * leaving it for the final time.
1473 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1475 struct task_group *parent, *child;
1479 parent = &root_task_group;
1481 ret = (*down)(parent, data);
1484 list_for_each_entry_rcu(child, &parent->children, siblings) {
1491 ret = (*up)(parent, data);
1496 parent = parent->parent;
1505 static int tg_nop(struct task_group *tg, void *data)
1512 /* Used instead of source_load when we know the type == 0 */
1513 static unsigned long weighted_cpuload(const int cpu)
1515 return cpu_rq(cpu)->load.weight;
1519 * Return a low guess at the load of a migration-source cpu weighted
1520 * according to the scheduling class and "nice" value.
1522 * We want to under-estimate the load of migration sources, to
1523 * balance conservatively.
1525 static unsigned long source_load(int cpu, int type)
1527 struct rq *rq = cpu_rq(cpu);
1528 unsigned long total = weighted_cpuload(cpu);
1530 if (type == 0 || !sched_feat(LB_BIAS))
1533 return min(rq->cpu_load[type-1], total);
1537 * Return a high guess at the load of a migration-target cpu weighted
1538 * according to the scheduling class and "nice" value.
1540 static unsigned long target_load(int cpu, int type)
1542 struct rq *rq = cpu_rq(cpu);
1543 unsigned long total = weighted_cpuload(cpu);
1545 if (type == 0 || !sched_feat(LB_BIAS))
1548 return max(rq->cpu_load[type-1], total);
1551 static struct sched_group *group_of(int cpu)
1553 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1561 static unsigned long power_of(int cpu)
1563 struct sched_group *group = group_of(cpu);
1566 return SCHED_LOAD_SCALE;
1568 return group->cpu_power;
1571 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1573 static unsigned long cpu_avg_load_per_task(int cpu)
1575 struct rq *rq = cpu_rq(cpu);
1576 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1579 rq->avg_load_per_task = rq->load.weight / nr_running;
1581 rq->avg_load_per_task = 0;
1583 return rq->avg_load_per_task;
1586 #ifdef CONFIG_FAIR_GROUP_SCHED
1588 static __read_mostly unsigned long *update_shares_data;
1590 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1593 * Calculate and set the cpu's group shares.
1595 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1596 unsigned long sd_shares,
1597 unsigned long sd_rq_weight,
1598 unsigned long *usd_rq_weight)
1600 unsigned long shares, rq_weight;
1603 rq_weight = usd_rq_weight[cpu];
1606 rq_weight = NICE_0_LOAD;
1610 * \Sum_j shares_j * rq_weight_i
1611 * shares_i = -----------------------------
1612 * \Sum_j rq_weight_j
1614 shares = (sd_shares * rq_weight) / sd_rq_weight;
1615 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1617 if (abs(shares - tg->se[cpu]->load.weight) >
1618 sysctl_sched_shares_thresh) {
1619 struct rq *rq = cpu_rq(cpu);
1620 unsigned long flags;
1622 raw_spin_lock_irqsave(&rq->lock, flags);
1623 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1624 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1625 __set_se_shares(tg->se[cpu], shares);
1626 raw_spin_unlock_irqrestore(&rq->lock, flags);
1631 * Re-compute the task group their per cpu shares over the given domain.
1632 * This needs to be done in a bottom-up fashion because the rq weight of a
1633 * parent group depends on the shares of its child groups.
1635 static int tg_shares_up(struct task_group *tg, void *data)
1637 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1638 unsigned long *usd_rq_weight;
1639 struct sched_domain *sd = data;
1640 unsigned long flags;
1646 local_irq_save(flags);
1647 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1649 for_each_cpu(i, sched_domain_span(sd)) {
1650 weight = tg->cfs_rq[i]->load.weight;
1651 usd_rq_weight[i] = weight;
1653 rq_weight += weight;
1655 * If there are currently no tasks on the cpu pretend there
1656 * is one of average load so that when a new task gets to
1657 * run here it will not get delayed by group starvation.
1660 weight = NICE_0_LOAD;
1662 sum_weight += weight;
1663 shares += tg->cfs_rq[i]->shares;
1667 rq_weight = sum_weight;
1669 if ((!shares && rq_weight) || shares > tg->shares)
1670 shares = tg->shares;
1672 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1673 shares = tg->shares;
1675 for_each_cpu(i, sched_domain_span(sd))
1676 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1678 local_irq_restore(flags);
1684 * Compute the cpu's hierarchical load factor for each task group.
1685 * This needs to be done in a top-down fashion because the load of a child
1686 * group is a fraction of its parents load.
1688 static int tg_load_down(struct task_group *tg, void *data)
1691 long cpu = (long)data;
1694 load = cpu_rq(cpu)->load.weight;
1696 load = tg->parent->cfs_rq[cpu]->h_load;
1697 load *= tg->cfs_rq[cpu]->shares;
1698 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1701 tg->cfs_rq[cpu]->h_load = load;
1706 static void update_shares(struct sched_domain *sd)
1711 if (root_task_group_empty())
1714 now = cpu_clock(raw_smp_processor_id());
1715 elapsed = now - sd->last_update;
1717 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1718 sd->last_update = now;
1719 walk_tg_tree(tg_nop, tg_shares_up, sd);
1723 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1725 if (root_task_group_empty())
1728 raw_spin_unlock(&rq->lock);
1730 raw_spin_lock(&rq->lock);
1733 static void update_h_load(long cpu)
1735 if (root_task_group_empty())
1738 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1743 static inline void update_shares(struct sched_domain *sd)
1747 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1753 #ifdef CONFIG_PREEMPT
1755 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1758 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1759 * way at the expense of forcing extra atomic operations in all
1760 * invocations. This assures that the double_lock is acquired using the
1761 * same underlying policy as the spinlock_t on this architecture, which
1762 * reduces latency compared to the unfair variant below. However, it
1763 * also adds more overhead and therefore may reduce throughput.
1765 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1766 __releases(this_rq->lock)
1767 __acquires(busiest->lock)
1768 __acquires(this_rq->lock)
1770 raw_spin_unlock(&this_rq->lock);
1771 double_rq_lock(this_rq, busiest);
1778 * Unfair double_lock_balance: Optimizes throughput at the expense of
1779 * latency by eliminating extra atomic operations when the locks are
1780 * already in proper order on entry. This favors lower cpu-ids and will
1781 * grant the double lock to lower cpus over higher ids under contention,
1782 * regardless of entry order into the function.
1784 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1785 __releases(this_rq->lock)
1786 __acquires(busiest->lock)
1787 __acquires(this_rq->lock)
1791 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1792 if (busiest < this_rq) {
1793 raw_spin_unlock(&this_rq->lock);
1794 raw_spin_lock(&busiest->lock);
1795 raw_spin_lock_nested(&this_rq->lock,
1796 SINGLE_DEPTH_NESTING);
1799 raw_spin_lock_nested(&busiest->lock,
1800 SINGLE_DEPTH_NESTING);
1805 #endif /* CONFIG_PREEMPT */
1808 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1810 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1812 if (unlikely(!irqs_disabled())) {
1813 /* printk() doesn't work good under rq->lock */
1814 raw_spin_unlock(&this_rq->lock);
1818 return _double_lock_balance(this_rq, busiest);
1821 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1822 __releases(busiest->lock)
1824 raw_spin_unlock(&busiest->lock);
1825 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1829 #ifdef CONFIG_FAIR_GROUP_SCHED
1830 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1833 cfs_rq->shares = shares;
1838 static void calc_load_account_active(struct rq *this_rq);
1839 static void update_sysctl(void);
1840 static int get_update_sysctl_factor(void);
1842 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1844 set_task_rq(p, cpu);
1847 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1848 * successfuly executed on another CPU. We must ensure that updates of
1849 * per-task data have been completed by this moment.
1852 task_thread_info(p)->cpu = cpu;
1856 #include "sched_stats.h"
1857 #include "sched_idletask.c"
1858 #include "sched_fair.c"
1859 #include "sched_rt.c"
1860 #ifdef CONFIG_SCHED_DEBUG
1861 # include "sched_debug.c"
1864 #define sched_class_highest (&rt_sched_class)
1865 #define for_each_class(class) \
1866 for (class = sched_class_highest; class; class = class->next)
1868 static void inc_nr_running(struct rq *rq)
1873 static void dec_nr_running(struct rq *rq)
1878 static void set_load_weight(struct task_struct *p)
1880 if (task_has_rt_policy(p)) {
1881 p->se.load.weight = prio_to_weight[0] * 2;
1882 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1887 * SCHED_IDLE tasks get minimal weight:
1889 if (p->policy == SCHED_IDLE) {
1890 p->se.load.weight = WEIGHT_IDLEPRIO;
1891 p->se.load.inv_weight = WMULT_IDLEPRIO;
1895 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1896 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1899 static void update_avg(u64 *avg, u64 sample)
1901 s64 diff = sample - *avg;
1905 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1908 p->se.start_runtime = p->se.sum_exec_runtime;
1910 sched_info_queued(p);
1911 p->sched_class->enqueue_task(rq, p, wakeup);
1915 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1918 if (p->se.last_wakeup) {
1919 update_avg(&p->se.avg_overlap,
1920 p->se.sum_exec_runtime - p->se.last_wakeup);
1921 p->se.last_wakeup = 0;
1923 update_avg(&p->se.avg_wakeup,
1924 sysctl_sched_wakeup_granularity);
1928 sched_info_dequeued(p);
1929 p->sched_class->dequeue_task(rq, p, sleep);
1934 * __normal_prio - return the priority that is based on the static prio
1936 static inline int __normal_prio(struct task_struct *p)
1938 return p->static_prio;
1942 * Calculate the expected normal priority: i.e. priority
1943 * without taking RT-inheritance into account. Might be
1944 * boosted by interactivity modifiers. Changes upon fork,
1945 * setprio syscalls, and whenever the interactivity
1946 * estimator recalculates.
1948 static inline int normal_prio(struct task_struct *p)
1952 if (task_has_rt_policy(p))
1953 prio = MAX_RT_PRIO-1 - p->rt_priority;
1955 prio = __normal_prio(p);
1960 * Calculate the current priority, i.e. the priority
1961 * taken into account by the scheduler. This value might
1962 * be boosted by RT tasks, or might be boosted by
1963 * interactivity modifiers. Will be RT if the task got
1964 * RT-boosted. If not then it returns p->normal_prio.
1966 static int effective_prio(struct task_struct *p)
1968 p->normal_prio = normal_prio(p);
1970 * If we are RT tasks or we were boosted to RT priority,
1971 * keep the priority unchanged. Otherwise, update priority
1972 * to the normal priority:
1974 if (!rt_prio(p->prio))
1975 return p->normal_prio;
1980 * activate_task - move a task to the runqueue.
1982 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1984 if (task_contributes_to_load(p))
1985 rq->nr_uninterruptible--;
1987 enqueue_task(rq, p, wakeup);
1992 * deactivate_task - remove a task from the runqueue.
1994 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1996 if (task_contributes_to_load(p))
1997 rq->nr_uninterruptible++;
1999 dequeue_task(rq, p, sleep);
2004 * task_curr - is this task currently executing on a CPU?
2005 * @p: the task in question.
2007 inline int task_curr(const struct task_struct *p)
2009 return cpu_curr(task_cpu(p)) == p;
2012 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2013 const struct sched_class *prev_class,
2014 int oldprio, int running)
2016 if (prev_class != p->sched_class) {
2017 if (prev_class->switched_from)
2018 prev_class->switched_from(rq, p, running);
2019 p->sched_class->switched_to(rq, p, running);
2021 p->sched_class->prio_changed(rq, p, oldprio, running);
2026 * Is this task likely cache-hot:
2029 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2033 if (p->sched_class != &fair_sched_class)
2037 * Buddy candidates are cache hot:
2039 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2040 (&p->se == cfs_rq_of(&p->se)->next ||
2041 &p->se == cfs_rq_of(&p->se)->last))
2044 if (sysctl_sched_migration_cost == -1)
2046 if (sysctl_sched_migration_cost == 0)
2049 delta = now - p->se.exec_start;
2051 return delta < (s64)sysctl_sched_migration_cost;
2054 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2056 #ifdef CONFIG_SCHED_DEBUG
2058 * We should never call set_task_cpu() on a blocked task,
2059 * ttwu() will sort out the placement.
2061 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2062 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2065 trace_sched_migrate_task(p, new_cpu);
2067 if (task_cpu(p) != new_cpu) {
2068 p->se.nr_migrations++;
2069 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2072 __set_task_cpu(p, new_cpu);
2075 struct migration_req {
2076 struct list_head list;
2078 struct task_struct *task;
2081 struct completion done;
2085 * The task's runqueue lock must be held.
2086 * Returns true if you have to wait for migration thread.
2089 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2091 struct rq *rq = task_rq(p);
2094 * If the task is not on a runqueue (and not running), then
2095 * the next wake-up will properly place the task.
2097 if (!p->se.on_rq && !task_running(rq, p))
2100 init_completion(&req->done);
2102 req->dest_cpu = dest_cpu;
2103 list_add(&req->list, &rq->migration_queue);
2109 * wait_task_context_switch - wait for a thread to complete at least one
2112 * @p must not be current.
2114 void wait_task_context_switch(struct task_struct *p)
2116 unsigned long nvcsw, nivcsw, flags;
2124 * The runqueue is assigned before the actual context
2125 * switch. We need to take the runqueue lock.
2127 * We could check initially without the lock but it is
2128 * very likely that we need to take the lock in every
2131 rq = task_rq_lock(p, &flags);
2132 running = task_running(rq, p);
2133 task_rq_unlock(rq, &flags);
2135 if (likely(!running))
2138 * The switch count is incremented before the actual
2139 * context switch. We thus wait for two switches to be
2140 * sure at least one completed.
2142 if ((p->nvcsw - nvcsw) > 1)
2144 if ((p->nivcsw - nivcsw) > 1)
2152 * wait_task_inactive - wait for a thread to unschedule.
2154 * If @match_state is nonzero, it's the @p->state value just checked and
2155 * not expected to change. If it changes, i.e. @p might have woken up,
2156 * then return zero. When we succeed in waiting for @p to be off its CPU,
2157 * we return a positive number (its total switch count). If a second call
2158 * a short while later returns the same number, the caller can be sure that
2159 * @p has remained unscheduled the whole time.
2161 * The caller must ensure that the task *will* unschedule sometime soon,
2162 * else this function might spin for a *long* time. This function can't
2163 * be called with interrupts off, or it may introduce deadlock with
2164 * smp_call_function() if an IPI is sent by the same process we are
2165 * waiting to become inactive.
2167 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2169 unsigned long flags;
2176 * We do the initial early heuristics without holding
2177 * any task-queue locks at all. We'll only try to get
2178 * the runqueue lock when things look like they will
2184 * If the task is actively running on another CPU
2185 * still, just relax and busy-wait without holding
2188 * NOTE! Since we don't hold any locks, it's not
2189 * even sure that "rq" stays as the right runqueue!
2190 * But we don't care, since "task_running()" will
2191 * return false if the runqueue has changed and p
2192 * is actually now running somewhere else!
2194 while (task_running(rq, p)) {
2195 if (match_state && unlikely(p->state != match_state))
2201 * Ok, time to look more closely! We need the rq
2202 * lock now, to be *sure*. If we're wrong, we'll
2203 * just go back and repeat.
2205 rq = task_rq_lock(p, &flags);
2206 trace_sched_wait_task(rq, p);
2207 running = task_running(rq, p);
2208 on_rq = p->se.on_rq;
2210 if (!match_state || p->state == match_state)
2211 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2212 task_rq_unlock(rq, &flags);
2215 * If it changed from the expected state, bail out now.
2217 if (unlikely(!ncsw))
2221 * Was it really running after all now that we
2222 * checked with the proper locks actually held?
2224 * Oops. Go back and try again..
2226 if (unlikely(running)) {
2232 * It's not enough that it's not actively running,
2233 * it must be off the runqueue _entirely_, and not
2236 * So if it was still runnable (but just not actively
2237 * running right now), it's preempted, and we should
2238 * yield - it could be a while.
2240 if (unlikely(on_rq)) {
2241 schedule_timeout_uninterruptible(1);
2246 * Ahh, all good. It wasn't running, and it wasn't
2247 * runnable, which means that it will never become
2248 * running in the future either. We're all done!
2257 * kick_process - kick a running thread to enter/exit the kernel
2258 * @p: the to-be-kicked thread
2260 * Cause a process which is running on another CPU to enter
2261 * kernel-mode, without any delay. (to get signals handled.)
2263 * NOTE: this function doesnt have to take the runqueue lock,
2264 * because all it wants to ensure is that the remote task enters
2265 * the kernel. If the IPI races and the task has been migrated
2266 * to another CPU then no harm is done and the purpose has been
2269 void kick_process(struct task_struct *p)
2275 if ((cpu != smp_processor_id()) && task_curr(p))
2276 smp_send_reschedule(cpu);
2279 EXPORT_SYMBOL_GPL(kick_process);
2280 #endif /* CONFIG_SMP */
2283 * task_oncpu_function_call - call a function on the cpu on which a task runs
2284 * @p: the task to evaluate
2285 * @func: the function to be called
2286 * @info: the function call argument
2288 * Calls the function @func when the task is currently running. This might
2289 * be on the current CPU, which just calls the function directly
2291 void task_oncpu_function_call(struct task_struct *p,
2292 void (*func) (void *info), void *info)
2299 smp_call_function_single(cpu, func, info, 1);
2304 static int select_fallback_rq(int cpu, struct task_struct *p)
2307 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2309 /* Look for allowed, online CPU in same node. */
2310 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2311 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2314 /* Any allowed, online CPU? */
2315 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2316 if (dest_cpu < nr_cpu_ids)
2319 /* No more Mr. Nice Guy. */
2320 if (dest_cpu >= nr_cpu_ids) {
2322 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2324 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2327 * Don't tell them about moving exiting tasks or
2328 * kernel threads (both mm NULL), since they never
2331 if (p->mm && printk_ratelimit()) {
2332 printk(KERN_INFO "process %d (%s) no "
2333 "longer affine to cpu%d\n",
2334 task_pid_nr(p), p->comm, cpu);
2342 * Gets called from 3 sites (exec, fork, wakeup), since it is called without
2343 * holding rq->lock we need to ensure ->cpus_allowed is stable, this is done
2346 * exec: is unstable, retry loop
2347 * fork & wake-up: serialize ->cpus_allowed against TASK_WAKING
2350 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2352 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2355 * In order not to call set_task_cpu() on a blocking task we need
2356 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2359 * Since this is common to all placement strategies, this lives here.
2361 * [ this allows ->select_task() to simply return task_cpu(p) and
2362 * not worry about this generic constraint ]
2364 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2366 cpu = select_fallback_rq(task_cpu(p), p);
2373 * try_to_wake_up - wake up a thread
2374 * @p: the to-be-woken-up thread
2375 * @state: the mask of task states that can be woken
2376 * @sync: do a synchronous wakeup?
2378 * Put it on the run-queue if it's not already there. The "current"
2379 * thread is always on the run-queue (except when the actual
2380 * re-schedule is in progress), and as such you're allowed to do
2381 * the simpler "current->state = TASK_RUNNING" to mark yourself
2382 * runnable without the overhead of this.
2384 * returns failure only if the task is already active.
2386 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2389 int cpu, orig_cpu, this_cpu, success = 0;
2390 unsigned long flags;
2391 struct rq *rq, *orig_rq;
2393 if (!sched_feat(SYNC_WAKEUPS))
2394 wake_flags &= ~WF_SYNC;
2396 this_cpu = get_cpu();
2399 rq = orig_rq = task_rq_lock(p, &flags);
2400 update_rq_clock(rq);
2401 if (!(p->state & state))
2411 if (unlikely(task_running(rq, p)))
2415 * In order to handle concurrent wakeups and release the rq->lock
2416 * we put the task in TASK_WAKING state.
2418 * First fix up the nr_uninterruptible count:
2420 if (task_contributes_to_load(p))
2421 rq->nr_uninterruptible--;
2422 p->state = TASK_WAKING;
2424 if (p->sched_class->task_waking)
2425 p->sched_class->task_waking(rq, p);
2427 __task_rq_unlock(rq);
2429 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2430 if (cpu != orig_cpu) {
2432 * Since we migrate the task without holding any rq->lock,
2433 * we need to be careful with task_rq_lock(), since that
2434 * might end up locking an invalid rq.
2436 set_task_cpu(p, cpu);
2440 raw_spin_lock(&rq->lock);
2441 update_rq_clock(rq);
2444 * We migrated the task without holding either rq->lock, however
2445 * since the task is not on the task list itself, nobody else
2446 * will try and migrate the task, hence the rq should match the
2447 * cpu we just moved it to.
2449 WARN_ON(task_cpu(p) != cpu);
2450 WARN_ON(p->state != TASK_WAKING);
2452 #ifdef CONFIG_SCHEDSTATS
2453 schedstat_inc(rq, ttwu_count);
2454 if (cpu == this_cpu)
2455 schedstat_inc(rq, ttwu_local);
2457 struct sched_domain *sd;
2458 for_each_domain(this_cpu, sd) {
2459 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2460 schedstat_inc(sd, ttwu_wake_remote);
2465 #endif /* CONFIG_SCHEDSTATS */
2468 #endif /* CONFIG_SMP */
2469 schedstat_inc(p, se.nr_wakeups);
2470 if (wake_flags & WF_SYNC)
2471 schedstat_inc(p, se.nr_wakeups_sync);
2472 if (orig_cpu != cpu)
2473 schedstat_inc(p, se.nr_wakeups_migrate);
2474 if (cpu == this_cpu)
2475 schedstat_inc(p, se.nr_wakeups_local);
2477 schedstat_inc(p, se.nr_wakeups_remote);
2478 activate_task(rq, p, 1);
2482 * Only attribute actual wakeups done by this task.
2484 if (!in_interrupt()) {
2485 struct sched_entity *se = ¤t->se;
2486 u64 sample = se->sum_exec_runtime;
2488 if (se->last_wakeup)
2489 sample -= se->last_wakeup;
2491 sample -= se->start_runtime;
2492 update_avg(&se->avg_wakeup, sample);
2494 se->last_wakeup = se->sum_exec_runtime;
2498 trace_sched_wakeup(rq, p, success);
2499 check_preempt_curr(rq, p, wake_flags);
2501 p->state = TASK_RUNNING;
2503 if (p->sched_class->task_woken)
2504 p->sched_class->task_woken(rq, p);
2506 if (unlikely(rq->idle_stamp)) {
2507 u64 delta = rq->clock - rq->idle_stamp;
2508 u64 max = 2*sysctl_sched_migration_cost;
2513 update_avg(&rq->avg_idle, delta);
2518 task_rq_unlock(rq, &flags);
2525 * wake_up_process - Wake up a specific process
2526 * @p: The process to be woken up.
2528 * Attempt to wake up the nominated process and move it to the set of runnable
2529 * processes. Returns 1 if the process was woken up, 0 if it was already
2532 * It may be assumed that this function implies a write memory barrier before
2533 * changing the task state if and only if any tasks are woken up.
2535 int wake_up_process(struct task_struct *p)
2537 return try_to_wake_up(p, TASK_ALL, 0);
2539 EXPORT_SYMBOL(wake_up_process);
2541 int wake_up_state(struct task_struct *p, unsigned int state)
2543 return try_to_wake_up(p, state, 0);
2547 * Perform scheduler related setup for a newly forked process p.
2548 * p is forked by current.
2550 * __sched_fork() is basic setup used by init_idle() too:
2552 static void __sched_fork(struct task_struct *p)
2554 p->se.exec_start = 0;
2555 p->se.sum_exec_runtime = 0;
2556 p->se.prev_sum_exec_runtime = 0;
2557 p->se.nr_migrations = 0;
2558 p->se.last_wakeup = 0;
2559 p->se.avg_overlap = 0;
2560 p->se.start_runtime = 0;
2561 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2563 #ifdef CONFIG_SCHEDSTATS
2564 p->se.wait_start = 0;
2566 p->se.wait_count = 0;
2569 p->se.sleep_start = 0;
2570 p->se.sleep_max = 0;
2571 p->se.sum_sleep_runtime = 0;
2573 p->se.block_start = 0;
2574 p->se.block_max = 0;
2576 p->se.slice_max = 0;
2578 p->se.nr_migrations_cold = 0;
2579 p->se.nr_failed_migrations_affine = 0;
2580 p->se.nr_failed_migrations_running = 0;
2581 p->se.nr_failed_migrations_hot = 0;
2582 p->se.nr_forced_migrations = 0;
2584 p->se.nr_wakeups = 0;
2585 p->se.nr_wakeups_sync = 0;
2586 p->se.nr_wakeups_migrate = 0;
2587 p->se.nr_wakeups_local = 0;
2588 p->se.nr_wakeups_remote = 0;
2589 p->se.nr_wakeups_affine = 0;
2590 p->se.nr_wakeups_affine_attempts = 0;
2591 p->se.nr_wakeups_passive = 0;
2592 p->se.nr_wakeups_idle = 0;
2596 INIT_LIST_HEAD(&p->rt.run_list);
2598 INIT_LIST_HEAD(&p->se.group_node);
2600 #ifdef CONFIG_PREEMPT_NOTIFIERS
2601 INIT_HLIST_HEAD(&p->preempt_notifiers);
2606 * fork()/clone()-time setup:
2608 void sched_fork(struct task_struct *p, int clone_flags)
2610 int cpu = get_cpu();
2614 * We mark the process as waking here. This guarantees that
2615 * nobody will actually run it, and a signal or other external
2616 * event cannot wake it up and insert it on the runqueue either.
2618 p->state = TASK_WAKING;
2621 * Revert to default priority/policy on fork if requested.
2623 if (unlikely(p->sched_reset_on_fork)) {
2624 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2625 p->policy = SCHED_NORMAL;
2626 p->normal_prio = p->static_prio;
2629 if (PRIO_TO_NICE(p->static_prio) < 0) {
2630 p->static_prio = NICE_TO_PRIO(0);
2631 p->normal_prio = p->static_prio;
2636 * We don't need the reset flag anymore after the fork. It has
2637 * fulfilled its duty:
2639 p->sched_reset_on_fork = 0;
2643 * Make sure we do not leak PI boosting priority to the child.
2645 p->prio = current->normal_prio;
2647 if (!rt_prio(p->prio))
2648 p->sched_class = &fair_sched_class;
2650 if (p->sched_class->task_fork)
2651 p->sched_class->task_fork(p);
2653 set_task_cpu(p, cpu);
2655 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2656 if (likely(sched_info_on()))
2657 memset(&p->sched_info, 0, sizeof(p->sched_info));
2659 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2662 #ifdef CONFIG_PREEMPT
2663 /* Want to start with kernel preemption disabled. */
2664 task_thread_info(p)->preempt_count = 1;
2666 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2672 * wake_up_new_task - wake up a newly created task for the first time.
2674 * This function will do some initial scheduler statistics housekeeping
2675 * that must be done for every newly created context, then puts the task
2676 * on the runqueue and wakes it.
2678 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2680 unsigned long flags;
2682 int cpu = get_cpu();
2686 * Fork balancing, do it here and not earlier because:
2687 * - cpus_allowed can change in the fork path
2688 * - any previously selected cpu might disappear through hotplug
2690 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2691 * ->cpus_allowed is stable, we have preemption disabled, meaning
2692 * cpu_online_mask is stable.
2694 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2695 set_task_cpu(p, cpu);
2699 * Since the task is not on the rq and we still have TASK_WAKING set
2700 * nobody else will migrate this task.
2703 raw_spin_lock_irqsave(&rq->lock, flags);
2705 BUG_ON(p->state != TASK_WAKING);
2706 p->state = TASK_RUNNING;
2707 update_rq_clock(rq);
2708 activate_task(rq, p, 0);
2709 trace_sched_wakeup_new(rq, p, 1);
2710 check_preempt_curr(rq, p, WF_FORK);
2712 if (p->sched_class->task_woken)
2713 p->sched_class->task_woken(rq, p);
2715 task_rq_unlock(rq, &flags);
2719 #ifdef CONFIG_PREEMPT_NOTIFIERS
2722 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2723 * @notifier: notifier struct to register
2725 void preempt_notifier_register(struct preempt_notifier *notifier)
2727 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2729 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2732 * preempt_notifier_unregister - no longer interested in preemption notifications
2733 * @notifier: notifier struct to unregister
2735 * This is safe to call from within a preemption notifier.
2737 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2739 hlist_del(¬ifier->link);
2741 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2743 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2745 struct preempt_notifier *notifier;
2746 struct hlist_node *node;
2748 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2749 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2753 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2754 struct task_struct *next)
2756 struct preempt_notifier *notifier;
2757 struct hlist_node *node;
2759 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2760 notifier->ops->sched_out(notifier, next);
2763 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2765 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2770 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2771 struct task_struct *next)
2775 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2778 * prepare_task_switch - prepare to switch tasks
2779 * @rq: the runqueue preparing to switch
2780 * @prev: the current task that is being switched out
2781 * @next: the task we are going to switch to.
2783 * This is called with the rq lock held and interrupts off. It must
2784 * be paired with a subsequent finish_task_switch after the context
2787 * prepare_task_switch sets up locking and calls architecture specific
2791 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2792 struct task_struct *next)
2794 fire_sched_out_preempt_notifiers(prev, next);
2795 prepare_lock_switch(rq, next);
2796 prepare_arch_switch(next);
2800 * finish_task_switch - clean up after a task-switch
2801 * @rq: runqueue associated with task-switch
2802 * @prev: the thread we just switched away from.
2804 * finish_task_switch must be called after the context switch, paired
2805 * with a prepare_task_switch call before the context switch.
2806 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2807 * and do any other architecture-specific cleanup actions.
2809 * Note that we may have delayed dropping an mm in context_switch(). If
2810 * so, we finish that here outside of the runqueue lock. (Doing it
2811 * with the lock held can cause deadlocks; see schedule() for
2814 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2815 __releases(rq->lock)
2817 struct mm_struct *mm = rq->prev_mm;
2823 * A task struct has one reference for the use as "current".
2824 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2825 * schedule one last time. The schedule call will never return, and
2826 * the scheduled task must drop that reference.
2827 * The test for TASK_DEAD must occur while the runqueue locks are
2828 * still held, otherwise prev could be scheduled on another cpu, die
2829 * there before we look at prev->state, and then the reference would
2831 * Manfred Spraul <manfred@colorfullife.com>
2833 prev_state = prev->state;
2834 finish_arch_switch(prev);
2835 perf_event_task_sched_in(current, cpu_of(rq));
2836 finish_lock_switch(rq, prev);
2838 fire_sched_in_preempt_notifiers(current);
2841 if (unlikely(prev_state == TASK_DEAD)) {
2843 * Remove function-return probe instances associated with this
2844 * task and put them back on the free list.
2846 kprobe_flush_task(prev);
2847 put_task_struct(prev);
2853 /* assumes rq->lock is held */
2854 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2856 if (prev->sched_class->pre_schedule)
2857 prev->sched_class->pre_schedule(rq, prev);
2860 /* rq->lock is NOT held, but preemption is disabled */
2861 static inline void post_schedule(struct rq *rq)
2863 if (rq->post_schedule) {
2864 unsigned long flags;
2866 raw_spin_lock_irqsave(&rq->lock, flags);
2867 if (rq->curr->sched_class->post_schedule)
2868 rq->curr->sched_class->post_schedule(rq);
2869 raw_spin_unlock_irqrestore(&rq->lock, flags);
2871 rq->post_schedule = 0;
2877 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2881 static inline void post_schedule(struct rq *rq)
2888 * schedule_tail - first thing a freshly forked thread must call.
2889 * @prev: the thread we just switched away from.
2891 asmlinkage void schedule_tail(struct task_struct *prev)
2892 __releases(rq->lock)
2894 struct rq *rq = this_rq();
2896 finish_task_switch(rq, prev);
2899 * FIXME: do we need to worry about rq being invalidated by the
2904 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2905 /* In this case, finish_task_switch does not reenable preemption */
2908 if (current->set_child_tid)
2909 put_user(task_pid_vnr(current), current->set_child_tid);
2913 * context_switch - switch to the new MM and the new
2914 * thread's register state.
2917 context_switch(struct rq *rq, struct task_struct *prev,
2918 struct task_struct *next)
2920 struct mm_struct *mm, *oldmm;
2922 prepare_task_switch(rq, prev, next);
2923 trace_sched_switch(rq, prev, next);
2925 oldmm = prev->active_mm;
2927 * For paravirt, this is coupled with an exit in switch_to to
2928 * combine the page table reload and the switch backend into
2931 arch_start_context_switch(prev);
2934 next->active_mm = oldmm;
2935 atomic_inc(&oldmm->mm_count);
2936 enter_lazy_tlb(oldmm, next);
2938 switch_mm(oldmm, mm, next);
2940 if (likely(!prev->mm)) {
2941 prev->active_mm = NULL;
2942 rq->prev_mm = oldmm;
2945 * Since the runqueue lock will be released by the next
2946 * task (which is an invalid locking op but in the case
2947 * of the scheduler it's an obvious special-case), so we
2948 * do an early lockdep release here:
2950 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2951 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2954 /* Here we just switch the register state and the stack. */
2955 switch_to(prev, next, prev);
2959 * this_rq must be evaluated again because prev may have moved
2960 * CPUs since it called schedule(), thus the 'rq' on its stack
2961 * frame will be invalid.
2963 finish_task_switch(this_rq(), prev);
2967 * nr_running, nr_uninterruptible and nr_context_switches:
2969 * externally visible scheduler statistics: current number of runnable
2970 * threads, current number of uninterruptible-sleeping threads, total
2971 * number of context switches performed since bootup.
2973 unsigned long nr_running(void)
2975 unsigned long i, sum = 0;
2977 for_each_online_cpu(i)
2978 sum += cpu_rq(i)->nr_running;
2983 unsigned long nr_uninterruptible(void)
2985 unsigned long i, sum = 0;
2987 for_each_possible_cpu(i)
2988 sum += cpu_rq(i)->nr_uninterruptible;
2991 * Since we read the counters lockless, it might be slightly
2992 * inaccurate. Do not allow it to go below zero though:
2994 if (unlikely((long)sum < 0))
3000 unsigned long long nr_context_switches(void)
3003 unsigned long long sum = 0;
3005 for_each_possible_cpu(i)
3006 sum += cpu_rq(i)->nr_switches;
3011 unsigned long nr_iowait(void)
3013 unsigned long i, sum = 0;
3015 for_each_possible_cpu(i)
3016 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3021 unsigned long nr_iowait_cpu(void)
3023 struct rq *this = this_rq();
3024 return atomic_read(&this->nr_iowait);
3027 unsigned long this_cpu_load(void)
3029 struct rq *this = this_rq();
3030 return this->cpu_load[0];
3034 /* Variables and functions for calc_load */
3035 static atomic_long_t calc_load_tasks;
3036 static unsigned long calc_load_update;
3037 unsigned long avenrun[3];
3038 EXPORT_SYMBOL(avenrun);
3041 * get_avenrun - get the load average array
3042 * @loads: pointer to dest load array
3043 * @offset: offset to add
3044 * @shift: shift count to shift the result left
3046 * These values are estimates at best, so no need for locking.
3048 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3050 loads[0] = (avenrun[0] + offset) << shift;
3051 loads[1] = (avenrun[1] + offset) << shift;
3052 loads[2] = (avenrun[2] + offset) << shift;
3055 static unsigned long
3056 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3059 load += active * (FIXED_1 - exp);
3060 return load >> FSHIFT;
3064 * calc_load - update the avenrun load estimates 10 ticks after the
3065 * CPUs have updated calc_load_tasks.
3067 void calc_global_load(void)
3069 unsigned long upd = calc_load_update + 10;
3072 if (time_before(jiffies, upd))
3075 active = atomic_long_read(&calc_load_tasks);
3076 active = active > 0 ? active * FIXED_1 : 0;
3078 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3079 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3080 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3082 calc_load_update += LOAD_FREQ;
3086 * Either called from update_cpu_load() or from a cpu going idle
3088 static void calc_load_account_active(struct rq *this_rq)
3090 long nr_active, delta;
3092 nr_active = this_rq->nr_running;
3093 nr_active += (long) this_rq->nr_uninterruptible;
3095 if (nr_active != this_rq->calc_load_active) {
3096 delta = nr_active - this_rq->calc_load_active;
3097 this_rq->calc_load_active = nr_active;
3098 atomic_long_add(delta, &calc_load_tasks);
3103 * Update rq->cpu_load[] statistics. This function is usually called every
3104 * scheduler tick (TICK_NSEC).
3106 static void update_cpu_load(struct rq *this_rq)
3108 unsigned long this_load = this_rq->load.weight;
3111 this_rq->nr_load_updates++;
3113 /* Update our load: */
3114 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3115 unsigned long old_load, new_load;
3117 /* scale is effectively 1 << i now, and >> i divides by scale */
3119 old_load = this_rq->cpu_load[i];
3120 new_load = this_load;
3122 * Round up the averaging division if load is increasing. This
3123 * prevents us from getting stuck on 9 if the load is 10, for
3126 if (new_load > old_load)
3127 new_load += scale-1;
3128 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3131 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3132 this_rq->calc_load_update += LOAD_FREQ;
3133 calc_load_account_active(this_rq);
3140 * double_rq_lock - safely lock two runqueues
3142 * Note this does not disable interrupts like task_rq_lock,
3143 * you need to do so manually before calling.
3145 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3146 __acquires(rq1->lock)
3147 __acquires(rq2->lock)
3149 BUG_ON(!irqs_disabled());
3151 raw_spin_lock(&rq1->lock);
3152 __acquire(rq2->lock); /* Fake it out ;) */
3155 raw_spin_lock(&rq1->lock);
3156 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3158 raw_spin_lock(&rq2->lock);
3159 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3162 update_rq_clock(rq1);
3163 update_rq_clock(rq2);
3167 * double_rq_unlock - safely unlock two runqueues
3169 * Note this does not restore interrupts like task_rq_unlock,
3170 * you need to do so manually after calling.
3172 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3173 __releases(rq1->lock)
3174 __releases(rq2->lock)
3176 raw_spin_unlock(&rq1->lock);
3178 raw_spin_unlock(&rq2->lock);
3180 __release(rq2->lock);
3184 * sched_exec - execve() is a valuable balancing opportunity, because at
3185 * this point the task has the smallest effective memory and cache footprint.
3187 void sched_exec(void)
3189 struct task_struct *p = current;
3190 struct migration_req req;
3191 int dest_cpu, this_cpu;
3192 unsigned long flags;
3196 this_cpu = get_cpu();
3197 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3198 if (dest_cpu == this_cpu) {
3203 rq = task_rq_lock(p, &flags);
3207 * select_task_rq() can race against ->cpus_allowed
3209 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3210 || unlikely(!cpu_active(dest_cpu))) {
3211 task_rq_unlock(rq, &flags);
3215 /* force the process onto the specified CPU */
3216 if (migrate_task(p, dest_cpu, &req)) {
3217 /* Need to wait for migration thread (might exit: take ref). */
3218 struct task_struct *mt = rq->migration_thread;
3220 get_task_struct(mt);
3221 task_rq_unlock(rq, &flags);
3222 wake_up_process(mt);
3223 put_task_struct(mt);
3224 wait_for_completion(&req.done);
3228 task_rq_unlock(rq, &flags);
3232 * pull_task - move a task from a remote runqueue to the local runqueue.
3233 * Both runqueues must be locked.
3235 static void pull_task(struct rq *src_rq, struct task_struct *p,
3236 struct rq *this_rq, int this_cpu)
3238 deactivate_task(src_rq, p, 0);
3239 set_task_cpu(p, this_cpu);
3240 activate_task(this_rq, p, 0);
3241 check_preempt_curr(this_rq, p, 0);
3245 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3248 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3249 struct sched_domain *sd, enum cpu_idle_type idle,
3252 int tsk_cache_hot = 0;
3254 * We do not migrate tasks that are:
3255 * 1) running (obviously), or
3256 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3257 * 3) are cache-hot on their current CPU.
3259 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3260 schedstat_inc(p, se.nr_failed_migrations_affine);
3265 if (task_running(rq, p)) {
3266 schedstat_inc(p, se.nr_failed_migrations_running);
3271 * Aggressive migration if:
3272 * 1) task is cache cold, or
3273 * 2) too many balance attempts have failed.
3276 tsk_cache_hot = task_hot(p, rq->clock, sd);
3277 if (!tsk_cache_hot ||
3278 sd->nr_balance_failed > sd->cache_nice_tries) {
3279 #ifdef CONFIG_SCHEDSTATS
3280 if (tsk_cache_hot) {
3281 schedstat_inc(sd, lb_hot_gained[idle]);
3282 schedstat_inc(p, se.nr_forced_migrations);
3288 if (tsk_cache_hot) {
3289 schedstat_inc(p, se.nr_failed_migrations_hot);
3295 static unsigned long
3296 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3297 unsigned long max_load_move, struct sched_domain *sd,
3298 enum cpu_idle_type idle, int *all_pinned,
3299 int *this_best_prio, struct rq_iterator *iterator)
3301 int loops = 0, pulled = 0, pinned = 0;
3302 struct task_struct *p;
3303 long rem_load_move = max_load_move;
3305 if (max_load_move == 0)
3311 * Start the load-balancing iterator:
3313 p = iterator->start(iterator->arg);
3315 if (!p || loops++ > sysctl_sched_nr_migrate)
3318 if ((p->se.load.weight >> 1) > rem_load_move ||
3319 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3320 p = iterator->next(iterator->arg);
3324 pull_task(busiest, p, this_rq, this_cpu);
3326 rem_load_move -= p->se.load.weight;
3328 #ifdef CONFIG_PREEMPT
3330 * NEWIDLE balancing is a source of latency, so preemptible kernels
3331 * will stop after the first task is pulled to minimize the critical
3334 if (idle == CPU_NEWLY_IDLE)
3339 * We only want to steal up to the prescribed amount of weighted load.
3341 if (rem_load_move > 0) {
3342 if (p->prio < *this_best_prio)
3343 *this_best_prio = p->prio;
3344 p = iterator->next(iterator->arg);
3349 * Right now, this is one of only two places pull_task() is called,
3350 * so we can safely collect pull_task() stats here rather than
3351 * inside pull_task().
3353 schedstat_add(sd, lb_gained[idle], pulled);
3356 *all_pinned = pinned;
3358 return max_load_move - rem_load_move;
3362 * move_tasks tries to move up to max_load_move weighted load from busiest to
3363 * this_rq, as part of a balancing operation within domain "sd".
3364 * Returns 1 if successful and 0 otherwise.
3366 * Called with both runqueues locked.
3368 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3369 unsigned long max_load_move,
3370 struct sched_domain *sd, enum cpu_idle_type idle,
3373 const struct sched_class *class = sched_class_highest;
3374 unsigned long total_load_moved = 0;
3375 int this_best_prio = this_rq->curr->prio;
3379 class->load_balance(this_rq, this_cpu, busiest,
3380 max_load_move - total_load_moved,
3381 sd, idle, all_pinned, &this_best_prio);
3382 class = class->next;
3384 #ifdef CONFIG_PREEMPT
3386 * NEWIDLE balancing is a source of latency, so preemptible
3387 * kernels will stop after the first task is pulled to minimize
3388 * the critical section.
3390 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3393 } while (class && max_load_move > total_load_moved);
3395 return total_load_moved > 0;
3399 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3400 struct sched_domain *sd, enum cpu_idle_type idle,
3401 struct rq_iterator *iterator)
3403 struct task_struct *p = iterator->start(iterator->arg);
3407 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3408 pull_task(busiest, p, this_rq, this_cpu);
3410 * Right now, this is only the second place pull_task()
3411 * is called, so we can safely collect pull_task()
3412 * stats here rather than inside pull_task().
3414 schedstat_inc(sd, lb_gained[idle]);
3418 p = iterator->next(iterator->arg);
3425 * move_one_task tries to move exactly one task from busiest to this_rq, as
3426 * part of active balancing operations within "domain".
3427 * Returns 1 if successful and 0 otherwise.
3429 * Called with both runqueues locked.
3431 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3432 struct sched_domain *sd, enum cpu_idle_type idle)
3434 const struct sched_class *class;
3436 for_each_class(class) {
3437 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3443 /********** Helpers for find_busiest_group ************************/
3445 * sd_lb_stats - Structure to store the statistics of a sched_domain
3446 * during load balancing.
3448 struct sd_lb_stats {
3449 struct sched_group *busiest; /* Busiest group in this sd */
3450 struct sched_group *this; /* Local group in this sd */
3451 unsigned long total_load; /* Total load of all groups in sd */
3452 unsigned long total_pwr; /* Total power of all groups in sd */
3453 unsigned long avg_load; /* Average load across all groups in sd */
3455 /** Statistics of this group */
3456 unsigned long this_load;
3457 unsigned long this_load_per_task;
3458 unsigned long this_nr_running;
3460 /* Statistics of the busiest group */
3461 unsigned long max_load;
3462 unsigned long busiest_load_per_task;
3463 unsigned long busiest_nr_running;
3465 int group_imb; /* Is there imbalance in this sd */
3466 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3467 int power_savings_balance; /* Is powersave balance needed for this sd */
3468 struct sched_group *group_min; /* Least loaded group in sd */
3469 struct sched_group *group_leader; /* Group which relieves group_min */
3470 unsigned long min_load_per_task; /* load_per_task in group_min */
3471 unsigned long leader_nr_running; /* Nr running of group_leader */
3472 unsigned long min_nr_running; /* Nr running of group_min */
3477 * sg_lb_stats - stats of a sched_group required for load_balancing
3479 struct sg_lb_stats {
3480 unsigned long avg_load; /*Avg load across the CPUs of the group */
3481 unsigned long group_load; /* Total load over the CPUs of the group */
3482 unsigned long sum_nr_running; /* Nr tasks running in the group */
3483 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3484 unsigned long group_capacity;
3485 int group_imb; /* Is there an imbalance in the group ? */
3489 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3490 * @group: The group whose first cpu is to be returned.
3492 static inline unsigned int group_first_cpu(struct sched_group *group)
3494 return cpumask_first(sched_group_cpus(group));
3498 * get_sd_load_idx - Obtain the load index for a given sched domain.
3499 * @sd: The sched_domain whose load_idx is to be obtained.
3500 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3502 static inline int get_sd_load_idx(struct sched_domain *sd,
3503 enum cpu_idle_type idle)
3509 load_idx = sd->busy_idx;
3512 case CPU_NEWLY_IDLE:
3513 load_idx = sd->newidle_idx;
3516 load_idx = sd->idle_idx;
3524 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3526 * init_sd_power_savings_stats - Initialize power savings statistics for
3527 * the given sched_domain, during load balancing.
3529 * @sd: Sched domain whose power-savings statistics are to be initialized.
3530 * @sds: Variable containing the statistics for sd.
3531 * @idle: Idle status of the CPU at which we're performing load-balancing.
3533 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3534 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3537 * Busy processors will not participate in power savings
3540 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3541 sds->power_savings_balance = 0;
3543 sds->power_savings_balance = 1;
3544 sds->min_nr_running = ULONG_MAX;
3545 sds->leader_nr_running = 0;
3550 * update_sd_power_savings_stats - Update the power saving stats for a
3551 * sched_domain while performing load balancing.
3553 * @group: sched_group belonging to the sched_domain under consideration.
3554 * @sds: Variable containing the statistics of the sched_domain
3555 * @local_group: Does group contain the CPU for which we're performing
3557 * @sgs: Variable containing the statistics of the group.
3559 static inline void update_sd_power_savings_stats(struct sched_group *group,
3560 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3563 if (!sds->power_savings_balance)
3567 * If the local group is idle or completely loaded
3568 * no need to do power savings balance at this domain
3570 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3571 !sds->this_nr_running))
3572 sds->power_savings_balance = 0;
3575 * If a group is already running at full capacity or idle,
3576 * don't include that group in power savings calculations
3578 if (!sds->power_savings_balance ||
3579 sgs->sum_nr_running >= sgs->group_capacity ||
3580 !sgs->sum_nr_running)
3584 * Calculate the group which has the least non-idle load.
3585 * This is the group from where we need to pick up the load
3588 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3589 (sgs->sum_nr_running == sds->min_nr_running &&
3590 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3591 sds->group_min = group;
3592 sds->min_nr_running = sgs->sum_nr_running;
3593 sds->min_load_per_task = sgs->sum_weighted_load /
3594 sgs->sum_nr_running;
3598 * Calculate the group which is almost near its
3599 * capacity but still has some space to pick up some load
3600 * from other group and save more power
3602 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3605 if (sgs->sum_nr_running > sds->leader_nr_running ||
3606 (sgs->sum_nr_running == sds->leader_nr_running &&
3607 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3608 sds->group_leader = group;
3609 sds->leader_nr_running = sgs->sum_nr_running;
3614 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3615 * @sds: Variable containing the statistics of the sched_domain
3616 * under consideration.
3617 * @this_cpu: Cpu at which we're currently performing load-balancing.
3618 * @imbalance: Variable to store the imbalance.
3621 * Check if we have potential to perform some power-savings balance.
3622 * If yes, set the busiest group to be the least loaded group in the
3623 * sched_domain, so that it's CPUs can be put to idle.
3625 * Returns 1 if there is potential to perform power-savings balance.
3628 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3629 int this_cpu, unsigned long *imbalance)
3631 if (!sds->power_savings_balance)
3634 if (sds->this != sds->group_leader ||
3635 sds->group_leader == sds->group_min)
3638 *imbalance = sds->min_load_per_task;
3639 sds->busiest = sds->group_min;
3644 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3645 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3646 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3651 static inline void update_sd_power_savings_stats(struct sched_group *group,
3652 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3657 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3658 int this_cpu, unsigned long *imbalance)
3662 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3665 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3667 return SCHED_LOAD_SCALE;
3670 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3672 return default_scale_freq_power(sd, cpu);
3675 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3677 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3678 unsigned long smt_gain = sd->smt_gain;
3685 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3687 return default_scale_smt_power(sd, cpu);
3690 unsigned long scale_rt_power(int cpu)
3692 struct rq *rq = cpu_rq(cpu);
3693 u64 total, available;
3695 sched_avg_update(rq);
3697 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3698 available = total - rq->rt_avg;
3700 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3701 total = SCHED_LOAD_SCALE;
3703 total >>= SCHED_LOAD_SHIFT;
3705 return div_u64(available, total);
3708 static void update_cpu_power(struct sched_domain *sd, int cpu)
3710 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3711 unsigned long power = SCHED_LOAD_SCALE;
3712 struct sched_group *sdg = sd->groups;
3714 if (sched_feat(ARCH_POWER))
3715 power *= arch_scale_freq_power(sd, cpu);
3717 power *= default_scale_freq_power(sd, cpu);
3719 power >>= SCHED_LOAD_SHIFT;
3721 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3722 if (sched_feat(ARCH_POWER))
3723 power *= arch_scale_smt_power(sd, cpu);
3725 power *= default_scale_smt_power(sd, cpu);
3727 power >>= SCHED_LOAD_SHIFT;
3730 power *= scale_rt_power(cpu);
3731 power >>= SCHED_LOAD_SHIFT;
3736 sdg->cpu_power = power;
3739 static void update_group_power(struct sched_domain *sd, int cpu)
3741 struct sched_domain *child = sd->child;
3742 struct sched_group *group, *sdg = sd->groups;
3743 unsigned long power;
3746 update_cpu_power(sd, cpu);
3752 group = child->groups;
3754 power += group->cpu_power;
3755 group = group->next;
3756 } while (group != child->groups);
3758 sdg->cpu_power = power;
3762 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3763 * @sd: The sched_domain whose statistics are to be updated.
3764 * @group: sched_group whose statistics are to be updated.
3765 * @this_cpu: Cpu for which load balance is currently performed.
3766 * @idle: Idle status of this_cpu
3767 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3768 * @sd_idle: Idle status of the sched_domain containing group.
3769 * @local_group: Does group contain this_cpu.
3770 * @cpus: Set of cpus considered for load balancing.
3771 * @balance: Should we balance.
3772 * @sgs: variable to hold the statistics for this group.
3774 static inline void update_sg_lb_stats(struct sched_domain *sd,
3775 struct sched_group *group, int this_cpu,
3776 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3777 int local_group, const struct cpumask *cpus,
3778 int *balance, struct sg_lb_stats *sgs)
3780 unsigned long load, max_cpu_load, min_cpu_load;
3782 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3783 unsigned long sum_avg_load_per_task;
3784 unsigned long avg_load_per_task;
3787 balance_cpu = group_first_cpu(group);
3788 if (balance_cpu == this_cpu)
3789 update_group_power(sd, this_cpu);
3792 /* Tally up the load of all CPUs in the group */
3793 sum_avg_load_per_task = avg_load_per_task = 0;
3795 min_cpu_load = ~0UL;
3797 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3798 struct rq *rq = cpu_rq(i);
3800 if (*sd_idle && rq->nr_running)
3803 /* Bias balancing toward cpus of our domain */
3805 if (idle_cpu(i) && !first_idle_cpu) {
3810 load = target_load(i, load_idx);
3812 load = source_load(i, load_idx);
3813 if (load > max_cpu_load)
3814 max_cpu_load = load;
3815 if (min_cpu_load > load)
3816 min_cpu_load = load;
3819 sgs->group_load += load;
3820 sgs->sum_nr_running += rq->nr_running;
3821 sgs->sum_weighted_load += weighted_cpuload(i);
3823 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3827 * First idle cpu or the first cpu(busiest) in this sched group
3828 * is eligible for doing load balancing at this and above
3829 * domains. In the newly idle case, we will allow all the cpu's
3830 * to do the newly idle load balance.
3832 if (idle != CPU_NEWLY_IDLE && local_group &&
3833 balance_cpu != this_cpu && balance) {
3838 /* Adjust by relative CPU power of the group */
3839 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3843 * Consider the group unbalanced when the imbalance is larger
3844 * than the average weight of two tasks.
3846 * APZ: with cgroup the avg task weight can vary wildly and
3847 * might not be a suitable number - should we keep a
3848 * normalized nr_running number somewhere that negates
3851 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3854 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3857 sgs->group_capacity =
3858 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3862 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3863 * @sd: sched_domain whose statistics are to be updated.
3864 * @this_cpu: Cpu for which load balance is currently performed.
3865 * @idle: Idle status of this_cpu
3866 * @sd_idle: Idle status of the sched_domain containing group.
3867 * @cpus: Set of cpus considered for load balancing.
3868 * @balance: Should we balance.
3869 * @sds: variable to hold the statistics for this sched_domain.
3871 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3872 enum cpu_idle_type idle, int *sd_idle,
3873 const struct cpumask *cpus, int *balance,
3874 struct sd_lb_stats *sds)
3876 struct sched_domain *child = sd->child;
3877 struct sched_group *group = sd->groups;
3878 struct sg_lb_stats sgs;
3879 int load_idx, prefer_sibling = 0;
3881 if (child && child->flags & SD_PREFER_SIBLING)
3884 init_sd_power_savings_stats(sd, sds, idle);
3885 load_idx = get_sd_load_idx(sd, idle);
3890 local_group = cpumask_test_cpu(this_cpu,
3891 sched_group_cpus(group));
3892 memset(&sgs, 0, sizeof(sgs));
3893 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3894 local_group, cpus, balance, &sgs);
3896 if (local_group && balance && !(*balance))
3899 sds->total_load += sgs.group_load;
3900 sds->total_pwr += group->cpu_power;
3903 * In case the child domain prefers tasks go to siblings
3904 * first, lower the group capacity to one so that we'll try
3905 * and move all the excess tasks away.
3908 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3911 sds->this_load = sgs.avg_load;
3913 sds->this_nr_running = sgs.sum_nr_running;
3914 sds->this_load_per_task = sgs.sum_weighted_load;
3915 } else if (sgs.avg_load > sds->max_load &&
3916 (sgs.sum_nr_running > sgs.group_capacity ||
3918 sds->max_load = sgs.avg_load;
3919 sds->busiest = group;
3920 sds->busiest_nr_running = sgs.sum_nr_running;
3921 sds->busiest_load_per_task = sgs.sum_weighted_load;
3922 sds->group_imb = sgs.group_imb;
3925 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3926 group = group->next;
3927 } while (group != sd->groups);
3931 * fix_small_imbalance - Calculate the minor imbalance that exists
3932 * amongst the groups of a sched_domain, during
3934 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3935 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3936 * @imbalance: Variable to store the imbalance.
3938 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3939 int this_cpu, unsigned long *imbalance)
3941 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3942 unsigned int imbn = 2;
3944 if (sds->this_nr_running) {
3945 sds->this_load_per_task /= sds->this_nr_running;
3946 if (sds->busiest_load_per_task >
3947 sds->this_load_per_task)
3950 sds->this_load_per_task =
3951 cpu_avg_load_per_task(this_cpu);
3953 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3954 sds->busiest_load_per_task * imbn) {
3955 *imbalance = sds->busiest_load_per_task;
3960 * OK, we don't have enough imbalance to justify moving tasks,
3961 * however we may be able to increase total CPU power used by
3965 pwr_now += sds->busiest->cpu_power *
3966 min(sds->busiest_load_per_task, sds->max_load);
3967 pwr_now += sds->this->cpu_power *
3968 min(sds->this_load_per_task, sds->this_load);
3969 pwr_now /= SCHED_LOAD_SCALE;
3971 /* Amount of load we'd subtract */
3972 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3973 sds->busiest->cpu_power;
3974 if (sds->max_load > tmp)
3975 pwr_move += sds->busiest->cpu_power *
3976 min(sds->busiest_load_per_task, sds->max_load - tmp);
3978 /* Amount of load we'd add */
3979 if (sds->max_load * sds->busiest->cpu_power <
3980 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3981 tmp = (sds->max_load * sds->busiest->cpu_power) /
3982 sds->this->cpu_power;
3984 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3985 sds->this->cpu_power;
3986 pwr_move += sds->this->cpu_power *
3987 min(sds->this_load_per_task, sds->this_load + tmp);
3988 pwr_move /= SCHED_LOAD_SCALE;
3990 /* Move if we gain throughput */
3991 if (pwr_move > pwr_now)
3992 *imbalance = sds->busiest_load_per_task;
3996 * calculate_imbalance - Calculate the amount of imbalance present within the
3997 * groups of a given sched_domain during load balance.
3998 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3999 * @this_cpu: Cpu for which currently load balance is being performed.
4000 * @imbalance: The variable to store the imbalance.
4002 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
4003 unsigned long *imbalance)
4005 unsigned long max_pull;
4007 * In the presence of smp nice balancing, certain scenarios can have
4008 * max load less than avg load(as we skip the groups at or below
4009 * its cpu_power, while calculating max_load..)
4011 if (sds->max_load < sds->avg_load) {
4013 return fix_small_imbalance(sds, this_cpu, imbalance);
4016 /* Don't want to pull so many tasks that a group would go idle */
4017 max_pull = min(sds->max_load - sds->avg_load,
4018 sds->max_load - sds->busiest_load_per_task);
4020 /* How much load to actually move to equalise the imbalance */
4021 *imbalance = min(max_pull * sds->busiest->cpu_power,
4022 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
4026 * if *imbalance is less than the average load per runnable task
4027 * there is no gaurantee that any tasks will be moved so we'll have
4028 * a think about bumping its value to force at least one task to be
4031 if (*imbalance < sds->busiest_load_per_task)
4032 return fix_small_imbalance(sds, this_cpu, imbalance);
4035 /******* find_busiest_group() helpers end here *********************/
4038 * find_busiest_group - Returns the busiest group within the sched_domain
4039 * if there is an imbalance. If there isn't an imbalance, and
4040 * the user has opted for power-savings, it returns a group whose
4041 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4042 * such a group exists.
4044 * Also calculates the amount of weighted load which should be moved
4045 * to restore balance.
4047 * @sd: The sched_domain whose busiest group is to be returned.
4048 * @this_cpu: The cpu for which load balancing is currently being performed.
4049 * @imbalance: Variable which stores amount of weighted load which should
4050 * be moved to restore balance/put a group to idle.
4051 * @idle: The idle status of this_cpu.
4052 * @sd_idle: The idleness of sd
4053 * @cpus: The set of CPUs under consideration for load-balancing.
4054 * @balance: Pointer to a variable indicating if this_cpu
4055 * is the appropriate cpu to perform load balancing at this_level.
4057 * Returns: - the busiest group if imbalance exists.
4058 * - If no imbalance and user has opted for power-savings balance,
4059 * return the least loaded group whose CPUs can be
4060 * put to idle by rebalancing its tasks onto our group.
4062 static struct sched_group *
4063 find_busiest_group(struct sched_domain *sd, int this_cpu,
4064 unsigned long *imbalance, enum cpu_idle_type idle,
4065 int *sd_idle, const struct cpumask *cpus, int *balance)
4067 struct sd_lb_stats sds;
4069 memset(&sds, 0, sizeof(sds));
4072 * Compute the various statistics relavent for load balancing at
4075 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4078 /* Cases where imbalance does not exist from POV of this_cpu */
4079 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4081 * 2) There is no busy sibling group to pull from.
4082 * 3) This group is the busiest group.
4083 * 4) This group is more busy than the avg busieness at this
4085 * 5) The imbalance is within the specified limit.
4086 * 6) Any rebalance would lead to ping-pong
4088 if (balance && !(*balance))
4091 if (!sds.busiest || sds.busiest_nr_running == 0)
4094 if (sds.this_load >= sds.max_load)
4097 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4099 if (sds.this_load >= sds.avg_load)
4102 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4105 sds.busiest_load_per_task /= sds.busiest_nr_running;
4107 sds.busiest_load_per_task =
4108 min(sds.busiest_load_per_task, sds.avg_load);
4111 * We're trying to get all the cpus to the average_load, so we don't
4112 * want to push ourselves above the average load, nor do we wish to
4113 * reduce the max loaded cpu below the average load, as either of these
4114 * actions would just result in more rebalancing later, and ping-pong
4115 * tasks around. Thus we look for the minimum possible imbalance.
4116 * Negative imbalances (*we* are more loaded than anyone else) will
4117 * be counted as no imbalance for these purposes -- we can't fix that
4118 * by pulling tasks to us. Be careful of negative numbers as they'll
4119 * appear as very large values with unsigned longs.
4121 if (sds.max_load <= sds.busiest_load_per_task)
4124 /* Looks like there is an imbalance. Compute it */
4125 calculate_imbalance(&sds, this_cpu, imbalance);
4130 * There is no obvious imbalance. But check if we can do some balancing
4133 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4141 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4144 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4145 unsigned long imbalance, const struct cpumask *cpus)
4147 struct rq *busiest = NULL, *rq;
4148 unsigned long max_load = 0;
4151 for_each_cpu(i, sched_group_cpus(group)) {
4152 unsigned long power = power_of(i);
4153 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4156 if (!cpumask_test_cpu(i, cpus))
4160 wl = weighted_cpuload(i);
4163 * When comparing with imbalance, use weighted_cpuload()
4164 * which is not scaled with the cpu power.
4166 if (capacity && rq->nr_running == 1 && wl > imbalance)
4170 * For the load comparisons with the other cpu's, consider
4171 * the weighted_cpuload() scaled with the cpu power, so that
4172 * the load can be moved away from the cpu that is potentially
4173 * running at a lower capacity.
4175 wl = (wl * SCHED_LOAD_SCALE) / power;
4177 if (wl > max_load) {
4187 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4188 * so long as it is large enough.
4190 #define MAX_PINNED_INTERVAL 512
4192 /* Working cpumask for load_balance and load_balance_newidle. */
4193 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4196 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4197 * tasks if there is an imbalance.
4199 static int load_balance(int this_cpu, struct rq *this_rq,
4200 struct sched_domain *sd, enum cpu_idle_type idle,
4203 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4204 struct sched_group *group;
4205 unsigned long imbalance;
4207 unsigned long flags;
4208 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4210 cpumask_copy(cpus, cpu_active_mask);
4213 * When power savings policy is enabled for the parent domain, idle
4214 * sibling can pick up load irrespective of busy siblings. In this case,
4215 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4216 * portraying it as CPU_NOT_IDLE.
4218 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4219 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4222 schedstat_inc(sd, lb_count[idle]);
4226 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4233 schedstat_inc(sd, lb_nobusyg[idle]);
4237 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4239 schedstat_inc(sd, lb_nobusyq[idle]);
4243 BUG_ON(busiest == this_rq);
4245 schedstat_add(sd, lb_imbalance[idle], imbalance);
4248 if (busiest->nr_running > 1) {
4250 * Attempt to move tasks. If find_busiest_group has found
4251 * an imbalance but busiest->nr_running <= 1, the group is
4252 * still unbalanced. ld_moved simply stays zero, so it is
4253 * correctly treated as an imbalance.
4255 local_irq_save(flags);
4256 double_rq_lock(this_rq, busiest);
4257 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4258 imbalance, sd, idle, &all_pinned);
4259 double_rq_unlock(this_rq, busiest);
4260 local_irq_restore(flags);
4263 * some other cpu did the load balance for us.
4265 if (ld_moved && this_cpu != smp_processor_id())
4266 resched_cpu(this_cpu);
4268 /* All tasks on this runqueue were pinned by CPU affinity */
4269 if (unlikely(all_pinned)) {
4270 cpumask_clear_cpu(cpu_of(busiest), cpus);
4271 if (!cpumask_empty(cpus))
4278 schedstat_inc(sd, lb_failed[idle]);
4279 sd->nr_balance_failed++;
4281 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4283 raw_spin_lock_irqsave(&busiest->lock, flags);
4285 /* don't kick the migration_thread, if the curr
4286 * task on busiest cpu can't be moved to this_cpu
4288 if (!cpumask_test_cpu(this_cpu,
4289 &busiest->curr->cpus_allowed)) {
4290 raw_spin_unlock_irqrestore(&busiest->lock,
4293 goto out_one_pinned;
4296 if (!busiest->active_balance) {
4297 busiest->active_balance = 1;
4298 busiest->push_cpu = this_cpu;
4301 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4303 wake_up_process(busiest->migration_thread);
4306 * We've kicked active balancing, reset the failure
4309 sd->nr_balance_failed = sd->cache_nice_tries+1;
4312 sd->nr_balance_failed = 0;
4314 if (likely(!active_balance)) {
4315 /* We were unbalanced, so reset the balancing interval */
4316 sd->balance_interval = sd->min_interval;
4319 * If we've begun active balancing, start to back off. This
4320 * case may not be covered by the all_pinned logic if there
4321 * is only 1 task on the busy runqueue (because we don't call
4324 if (sd->balance_interval < sd->max_interval)
4325 sd->balance_interval *= 2;
4328 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4329 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4335 schedstat_inc(sd, lb_balanced[idle]);
4337 sd->nr_balance_failed = 0;
4340 /* tune up the balancing interval */
4341 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4342 (sd->balance_interval < sd->max_interval))
4343 sd->balance_interval *= 2;
4345 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4346 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4357 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4358 * tasks if there is an imbalance.
4360 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4361 * this_rq is locked.
4364 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4366 struct sched_group *group;
4367 struct rq *busiest = NULL;
4368 unsigned long imbalance;
4372 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4374 cpumask_copy(cpus, cpu_active_mask);
4377 * When power savings policy is enabled for the parent domain, idle
4378 * sibling can pick up load irrespective of busy siblings. In this case,
4379 * let the state of idle sibling percolate up as IDLE, instead of
4380 * portraying it as CPU_NOT_IDLE.
4382 if (sd->flags & SD_SHARE_CPUPOWER &&
4383 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4386 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4388 update_shares_locked(this_rq, sd);
4389 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4390 &sd_idle, cpus, NULL);
4392 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4396 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4398 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4402 BUG_ON(busiest == this_rq);
4404 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4407 if (busiest->nr_running > 1) {
4408 /* Attempt to move tasks */
4409 double_lock_balance(this_rq, busiest);
4410 /* this_rq->clock is already updated */
4411 update_rq_clock(busiest);
4412 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4413 imbalance, sd, CPU_NEWLY_IDLE,
4415 double_unlock_balance(this_rq, busiest);
4417 if (unlikely(all_pinned)) {
4418 cpumask_clear_cpu(cpu_of(busiest), cpus);
4419 if (!cpumask_empty(cpus))
4425 int active_balance = 0;
4427 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4428 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4429 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4432 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4435 if (sd->nr_balance_failed++ < 2)
4439 * The only task running in a non-idle cpu can be moved to this
4440 * cpu in an attempt to completely freeup the other CPU
4441 * package. The same method used to move task in load_balance()
4442 * have been extended for load_balance_newidle() to speedup
4443 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4445 * The package power saving logic comes from
4446 * find_busiest_group(). If there are no imbalance, then
4447 * f_b_g() will return NULL. However when sched_mc={1,2} then
4448 * f_b_g() will select a group from which a running task may be
4449 * pulled to this cpu in order to make the other package idle.
4450 * If there is no opportunity to make a package idle and if
4451 * there are no imbalance, then f_b_g() will return NULL and no
4452 * action will be taken in load_balance_newidle().
4454 * Under normal task pull operation due to imbalance, there
4455 * will be more than one task in the source run queue and
4456 * move_tasks() will succeed. ld_moved will be true and this
4457 * active balance code will not be triggered.
4460 /* Lock busiest in correct order while this_rq is held */
4461 double_lock_balance(this_rq, busiest);
4464 * don't kick the migration_thread, if the curr
4465 * task on busiest cpu can't be moved to this_cpu
4467 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4468 double_unlock_balance(this_rq, busiest);
4473 if (!busiest->active_balance) {
4474 busiest->active_balance = 1;
4475 busiest->push_cpu = this_cpu;
4479 double_unlock_balance(this_rq, busiest);
4481 * Should not call ttwu while holding a rq->lock
4483 raw_spin_unlock(&this_rq->lock);
4485 wake_up_process(busiest->migration_thread);
4486 raw_spin_lock(&this_rq->lock);
4489 sd->nr_balance_failed = 0;
4491 update_shares_locked(this_rq, sd);
4495 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4496 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4497 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4499 sd->nr_balance_failed = 0;
4505 * idle_balance is called by schedule() if this_cpu is about to become
4506 * idle. Attempts to pull tasks from other CPUs.
4508 static void idle_balance(int this_cpu, struct rq *this_rq)
4510 struct sched_domain *sd;
4511 int pulled_task = 0;
4512 unsigned long next_balance = jiffies + HZ;
4514 this_rq->idle_stamp = this_rq->clock;
4516 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4519 for_each_domain(this_cpu, sd) {
4520 unsigned long interval;
4522 if (!(sd->flags & SD_LOAD_BALANCE))
4525 if (sd->flags & SD_BALANCE_NEWIDLE)
4526 /* If we've pulled tasks over stop searching: */
4527 pulled_task = load_balance_newidle(this_cpu, this_rq,
4530 interval = msecs_to_jiffies(sd->balance_interval);
4531 if (time_after(next_balance, sd->last_balance + interval))
4532 next_balance = sd->last_balance + interval;
4534 this_rq->idle_stamp = 0;
4538 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4540 * We are going idle. next_balance may be set based on
4541 * a busy processor. So reset next_balance.
4543 this_rq->next_balance = next_balance;
4548 * active_load_balance is run by migration threads. It pushes running tasks
4549 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4550 * running on each physical CPU where possible, and avoids physical /
4551 * logical imbalances.
4553 * Called with busiest_rq locked.
4555 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4557 int target_cpu = busiest_rq->push_cpu;
4558 struct sched_domain *sd;
4559 struct rq *target_rq;
4561 /* Is there any task to move? */
4562 if (busiest_rq->nr_running <= 1)
4565 target_rq = cpu_rq(target_cpu);
4568 * This condition is "impossible", if it occurs
4569 * we need to fix it. Originally reported by
4570 * Bjorn Helgaas on a 128-cpu setup.
4572 BUG_ON(busiest_rq == target_rq);
4574 /* move a task from busiest_rq to target_rq */
4575 double_lock_balance(busiest_rq, target_rq);
4576 update_rq_clock(busiest_rq);
4577 update_rq_clock(target_rq);
4579 /* Search for an sd spanning us and the target CPU. */
4580 for_each_domain(target_cpu, sd) {
4581 if ((sd->flags & SD_LOAD_BALANCE) &&
4582 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4587 schedstat_inc(sd, alb_count);
4589 if (move_one_task(target_rq, target_cpu, busiest_rq,
4591 schedstat_inc(sd, alb_pushed);
4593 schedstat_inc(sd, alb_failed);
4595 double_unlock_balance(busiest_rq, target_rq);
4600 atomic_t load_balancer;
4601 cpumask_var_t cpu_mask;
4602 cpumask_var_t ilb_grp_nohz_mask;
4603 } nohz ____cacheline_aligned = {
4604 .load_balancer = ATOMIC_INIT(-1),
4607 int get_nohz_load_balancer(void)
4609 return atomic_read(&nohz.load_balancer);
4612 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4614 * lowest_flag_domain - Return lowest sched_domain containing flag.
4615 * @cpu: The cpu whose lowest level of sched domain is to
4617 * @flag: The flag to check for the lowest sched_domain
4618 * for the given cpu.
4620 * Returns the lowest sched_domain of a cpu which contains the given flag.
4622 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4624 struct sched_domain *sd;
4626 for_each_domain(cpu, sd)
4627 if (sd && (sd->flags & flag))
4634 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4635 * @cpu: The cpu whose domains we're iterating over.
4636 * @sd: variable holding the value of the power_savings_sd
4638 * @flag: The flag to filter the sched_domains to be iterated.
4640 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4641 * set, starting from the lowest sched_domain to the highest.
4643 #define for_each_flag_domain(cpu, sd, flag) \
4644 for (sd = lowest_flag_domain(cpu, flag); \
4645 (sd && (sd->flags & flag)); sd = sd->parent)
4648 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4649 * @ilb_group: group to be checked for semi-idleness
4651 * Returns: 1 if the group is semi-idle. 0 otherwise.
4653 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4654 * and atleast one non-idle CPU. This helper function checks if the given
4655 * sched_group is semi-idle or not.
4657 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4659 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4660 sched_group_cpus(ilb_group));
4663 * A sched_group is semi-idle when it has atleast one busy cpu
4664 * and atleast one idle cpu.
4666 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4669 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4675 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4676 * @cpu: The cpu which is nominating a new idle_load_balancer.
4678 * Returns: Returns the id of the idle load balancer if it exists,
4679 * Else, returns >= nr_cpu_ids.
4681 * This algorithm picks the idle load balancer such that it belongs to a
4682 * semi-idle powersavings sched_domain. The idea is to try and avoid
4683 * completely idle packages/cores just for the purpose of idle load balancing
4684 * when there are other idle cpu's which are better suited for that job.
4686 static int find_new_ilb(int cpu)
4688 struct sched_domain *sd;
4689 struct sched_group *ilb_group;
4692 * Have idle load balancer selection from semi-idle packages only
4693 * when power-aware load balancing is enabled
4695 if (!(sched_smt_power_savings || sched_mc_power_savings))
4699 * Optimize for the case when we have no idle CPUs or only one
4700 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4702 if (cpumask_weight(nohz.cpu_mask) < 2)
4705 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4706 ilb_group = sd->groups;
4709 if (is_semi_idle_group(ilb_group))
4710 return cpumask_first(nohz.ilb_grp_nohz_mask);
4712 ilb_group = ilb_group->next;
4714 } while (ilb_group != sd->groups);
4718 return cpumask_first(nohz.cpu_mask);
4720 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4721 static inline int find_new_ilb(int call_cpu)
4723 return cpumask_first(nohz.cpu_mask);
4728 * This routine will try to nominate the ilb (idle load balancing)
4729 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4730 * load balancing on behalf of all those cpus. If all the cpus in the system
4731 * go into this tickless mode, then there will be no ilb owner (as there is
4732 * no need for one) and all the cpus will sleep till the next wakeup event
4735 * For the ilb owner, tick is not stopped. And this tick will be used
4736 * for idle load balancing. ilb owner will still be part of
4739 * While stopping the tick, this cpu will become the ilb owner if there
4740 * is no other owner. And will be the owner till that cpu becomes busy
4741 * or if all cpus in the system stop their ticks at which point
4742 * there is no need for ilb owner.
4744 * When the ilb owner becomes busy, it nominates another owner, during the
4745 * next busy scheduler_tick()
4747 int select_nohz_load_balancer(int stop_tick)
4749 int cpu = smp_processor_id();
4752 cpu_rq(cpu)->in_nohz_recently = 1;
4754 if (!cpu_active(cpu)) {
4755 if (atomic_read(&nohz.load_balancer) != cpu)
4759 * If we are going offline and still the leader,
4762 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4768 cpumask_set_cpu(cpu, nohz.cpu_mask);
4770 /* time for ilb owner also to sleep */
4771 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4772 if (atomic_read(&nohz.load_balancer) == cpu)
4773 atomic_set(&nohz.load_balancer, -1);
4777 if (atomic_read(&nohz.load_balancer) == -1) {
4778 /* make me the ilb owner */
4779 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4781 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4784 if (!(sched_smt_power_savings ||
4785 sched_mc_power_savings))
4788 * Check to see if there is a more power-efficient
4791 new_ilb = find_new_ilb(cpu);
4792 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4793 atomic_set(&nohz.load_balancer, -1);
4794 resched_cpu(new_ilb);
4800 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4803 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4805 if (atomic_read(&nohz.load_balancer) == cpu)
4806 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4813 static DEFINE_SPINLOCK(balancing);
4816 * It checks each scheduling domain to see if it is due to be balanced,
4817 * and initiates a balancing operation if so.
4819 * Balancing parameters are set up in arch_init_sched_domains.
4821 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4824 struct rq *rq = cpu_rq(cpu);
4825 unsigned long interval;
4826 struct sched_domain *sd;
4827 /* Earliest time when we have to do rebalance again */
4828 unsigned long next_balance = jiffies + 60*HZ;
4829 int update_next_balance = 0;
4832 for_each_domain(cpu, sd) {
4833 if (!(sd->flags & SD_LOAD_BALANCE))
4836 interval = sd->balance_interval;
4837 if (idle != CPU_IDLE)
4838 interval *= sd->busy_factor;
4840 /* scale ms to jiffies */
4841 interval = msecs_to_jiffies(interval);
4842 if (unlikely(!interval))
4844 if (interval > HZ*NR_CPUS/10)
4845 interval = HZ*NR_CPUS/10;
4847 need_serialize = sd->flags & SD_SERIALIZE;
4849 if (need_serialize) {
4850 if (!spin_trylock(&balancing))
4854 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4855 if (load_balance(cpu, rq, sd, idle, &balance)) {
4857 * We've pulled tasks over so either we're no
4858 * longer idle, or one of our SMT siblings is
4861 idle = CPU_NOT_IDLE;
4863 sd->last_balance = jiffies;
4866 spin_unlock(&balancing);
4868 if (time_after(next_balance, sd->last_balance + interval)) {
4869 next_balance = sd->last_balance + interval;
4870 update_next_balance = 1;
4874 * Stop the load balance at this level. There is another
4875 * CPU in our sched group which is doing load balancing more
4883 * next_balance will be updated only when there is a need.
4884 * When the cpu is attached to null domain for ex, it will not be
4887 if (likely(update_next_balance))
4888 rq->next_balance = next_balance;
4892 * run_rebalance_domains is triggered when needed from the scheduler tick.
4893 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4894 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4896 static void run_rebalance_domains(struct softirq_action *h)
4898 int this_cpu = smp_processor_id();
4899 struct rq *this_rq = cpu_rq(this_cpu);
4900 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4901 CPU_IDLE : CPU_NOT_IDLE;
4903 rebalance_domains(this_cpu, idle);
4907 * If this cpu is the owner for idle load balancing, then do the
4908 * balancing on behalf of the other idle cpus whose ticks are
4911 if (this_rq->idle_at_tick &&
4912 atomic_read(&nohz.load_balancer) == this_cpu) {
4916 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4917 if (balance_cpu == this_cpu)
4921 * If this cpu gets work to do, stop the load balancing
4922 * work being done for other cpus. Next load
4923 * balancing owner will pick it up.
4928 rebalance_domains(balance_cpu, CPU_IDLE);
4930 rq = cpu_rq(balance_cpu);
4931 if (time_after(this_rq->next_balance, rq->next_balance))
4932 this_rq->next_balance = rq->next_balance;
4938 static inline int on_null_domain(int cpu)
4940 return !rcu_dereference(cpu_rq(cpu)->sd);
4944 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4946 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4947 * idle load balancing owner or decide to stop the periodic load balancing,
4948 * if the whole system is idle.
4950 static inline void trigger_load_balance(struct rq *rq, int cpu)
4954 * If we were in the nohz mode recently and busy at the current
4955 * scheduler tick, then check if we need to nominate new idle
4958 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4959 rq->in_nohz_recently = 0;
4961 if (atomic_read(&nohz.load_balancer) == cpu) {
4962 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4963 atomic_set(&nohz.load_balancer, -1);
4966 if (atomic_read(&nohz.load_balancer) == -1) {
4967 int ilb = find_new_ilb(cpu);
4969 if (ilb < nr_cpu_ids)
4975 * If this cpu is idle and doing idle load balancing for all the
4976 * cpus with ticks stopped, is it time for that to stop?
4978 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4979 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4985 * If this cpu is idle and the idle load balancing is done by
4986 * someone else, then no need raise the SCHED_SOFTIRQ
4988 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4989 cpumask_test_cpu(cpu, nohz.cpu_mask))
4992 /* Don't need to rebalance while attached to NULL domain */
4993 if (time_after_eq(jiffies, rq->next_balance) &&
4994 likely(!on_null_domain(cpu)))
4995 raise_softirq(SCHED_SOFTIRQ);
4998 #else /* CONFIG_SMP */
5001 * on UP we do not need to balance between CPUs:
5003 static inline void idle_balance(int cpu, struct rq *rq)
5009 DEFINE_PER_CPU(struct kernel_stat, kstat);
5011 EXPORT_PER_CPU_SYMBOL(kstat);
5014 * Return any ns on the sched_clock that have not yet been accounted in
5015 * @p in case that task is currently running.
5017 * Called with task_rq_lock() held on @rq.
5019 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
5023 if (task_current(rq, p)) {
5024 update_rq_clock(rq);
5025 ns = rq->clock - p->se.exec_start;
5033 unsigned long long task_delta_exec(struct task_struct *p)
5035 unsigned long flags;
5039 rq = task_rq_lock(p, &flags);
5040 ns = do_task_delta_exec(p, rq);
5041 task_rq_unlock(rq, &flags);
5047 * Return accounted runtime for the task.
5048 * In case the task is currently running, return the runtime plus current's
5049 * pending runtime that have not been accounted yet.
5051 unsigned long long task_sched_runtime(struct task_struct *p)
5053 unsigned long flags;
5057 rq = task_rq_lock(p, &flags);
5058 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
5059 task_rq_unlock(rq, &flags);
5065 * Return sum_exec_runtime for the thread group.
5066 * In case the task is currently running, return the sum plus current's
5067 * pending runtime that have not been accounted yet.
5069 * Note that the thread group might have other running tasks as well,
5070 * so the return value not includes other pending runtime that other
5071 * running tasks might have.
5073 unsigned long long thread_group_sched_runtime(struct task_struct *p)
5075 struct task_cputime totals;
5076 unsigned long flags;
5080 rq = task_rq_lock(p, &flags);
5081 thread_group_cputime(p, &totals);
5082 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
5083 task_rq_unlock(rq, &flags);
5089 * Account user cpu time to a process.
5090 * @p: the process that the cpu time gets accounted to
5091 * @cputime: the cpu time spent in user space since the last update
5092 * @cputime_scaled: cputime scaled by cpu frequency
5094 void account_user_time(struct task_struct *p, cputime_t cputime,
5095 cputime_t cputime_scaled)
5097 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5100 /* Add user time to process. */
5101 p->utime = cputime_add(p->utime, cputime);
5102 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5103 account_group_user_time(p, cputime);
5105 /* Add user time to cpustat. */
5106 tmp = cputime_to_cputime64(cputime);
5107 if (TASK_NICE(p) > 0)
5108 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5110 cpustat->user = cputime64_add(cpustat->user, tmp);
5112 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5113 /* Account for user time used */
5114 acct_update_integrals(p);
5118 * Account guest cpu time to a process.
5119 * @p: the process that the cpu time gets accounted to
5120 * @cputime: the cpu time spent in virtual machine since the last update
5121 * @cputime_scaled: cputime scaled by cpu frequency
5123 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5124 cputime_t cputime_scaled)
5127 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5129 tmp = cputime_to_cputime64(cputime);
5131 /* Add guest time to process. */
5132 p->utime = cputime_add(p->utime, cputime);
5133 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5134 account_group_user_time(p, cputime);
5135 p->gtime = cputime_add(p->gtime, cputime);
5137 /* Add guest time to cpustat. */
5138 if (TASK_NICE(p) > 0) {
5139 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5140 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
5142 cpustat->user = cputime64_add(cpustat->user, tmp);
5143 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5148 * Account system cpu time to a process.
5149 * @p: the process that the cpu time gets accounted to
5150 * @hardirq_offset: the offset to subtract from hardirq_count()
5151 * @cputime: the cpu time spent in kernel space since the last update
5152 * @cputime_scaled: cputime scaled by cpu frequency
5154 void account_system_time(struct task_struct *p, int hardirq_offset,
5155 cputime_t cputime, cputime_t cputime_scaled)
5157 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5160 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5161 account_guest_time(p, cputime, cputime_scaled);
5165 /* Add system time to process. */
5166 p->stime = cputime_add(p->stime, cputime);
5167 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5168 account_group_system_time(p, cputime);
5170 /* Add system time to cpustat. */
5171 tmp = cputime_to_cputime64(cputime);
5172 if (hardirq_count() - hardirq_offset)
5173 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5174 else if (softirq_count())
5175 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5177 cpustat->system = cputime64_add(cpustat->system, tmp);
5179 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5181 /* Account for system time used */
5182 acct_update_integrals(p);
5186 * Account for involuntary wait time.
5187 * @steal: the cpu time spent in involuntary wait
5189 void account_steal_time(cputime_t cputime)
5191 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5192 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5194 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5198 * Account for idle time.
5199 * @cputime: the cpu time spent in idle wait
5201 void account_idle_time(cputime_t cputime)
5203 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5204 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5205 struct rq *rq = this_rq();
5207 if (atomic_read(&rq->nr_iowait) > 0)
5208 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5210 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5213 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5216 * Account a single tick of cpu time.
5217 * @p: the process that the cpu time gets accounted to
5218 * @user_tick: indicates if the tick is a user or a system tick
5220 void account_process_tick(struct task_struct *p, int user_tick)
5222 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5223 struct rq *rq = this_rq();
5226 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5227 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5228 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5231 account_idle_time(cputime_one_jiffy);
5235 * Account multiple ticks of steal time.
5236 * @p: the process from which the cpu time has been stolen
5237 * @ticks: number of stolen ticks
5239 void account_steal_ticks(unsigned long ticks)
5241 account_steal_time(jiffies_to_cputime(ticks));
5245 * Account multiple ticks of idle time.
5246 * @ticks: number of stolen ticks
5248 void account_idle_ticks(unsigned long ticks)
5250 account_idle_time(jiffies_to_cputime(ticks));
5256 * Use precise platform statistics if available:
5258 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5259 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5265 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5267 struct task_cputime cputime;
5269 thread_group_cputime(p, &cputime);
5271 *ut = cputime.utime;
5272 *st = cputime.stime;
5276 #ifndef nsecs_to_cputime
5277 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
5280 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5282 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
5285 * Use CFS's precise accounting:
5287 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
5292 temp = (u64)(rtime * utime);
5293 do_div(temp, total);
5294 utime = (cputime_t)temp;
5299 * Compare with previous values, to keep monotonicity:
5301 p->prev_utime = max(p->prev_utime, utime);
5302 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
5304 *ut = p->prev_utime;
5305 *st = p->prev_stime;
5309 * Must be called with siglock held.
5311 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5313 struct signal_struct *sig = p->signal;
5314 struct task_cputime cputime;
5315 cputime_t rtime, utime, total;
5317 thread_group_cputime(p, &cputime);
5319 total = cputime_add(cputime.utime, cputime.stime);
5320 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5325 temp = (u64)(rtime * cputime.utime);
5326 do_div(temp, total);
5327 utime = (cputime_t)temp;
5331 sig->prev_utime = max(sig->prev_utime, utime);
5332 sig->prev_stime = max(sig->prev_stime,
5333 cputime_sub(rtime, sig->prev_utime));
5335 *ut = sig->prev_utime;
5336 *st = sig->prev_stime;
5341 * This function gets called by the timer code, with HZ frequency.
5342 * We call it with interrupts disabled.
5344 * It also gets called by the fork code, when changing the parent's
5347 void scheduler_tick(void)
5349 int cpu = smp_processor_id();
5350 struct rq *rq = cpu_rq(cpu);
5351 struct task_struct *curr = rq->curr;
5355 raw_spin_lock(&rq->lock);
5356 update_rq_clock(rq);
5357 update_cpu_load(rq);
5358 curr->sched_class->task_tick(rq, curr, 0);
5359 raw_spin_unlock(&rq->lock);
5361 perf_event_task_tick(curr, cpu);
5364 rq->idle_at_tick = idle_cpu(cpu);
5365 trigger_load_balance(rq, cpu);
5369 notrace unsigned long get_parent_ip(unsigned long addr)
5371 if (in_lock_functions(addr)) {
5372 addr = CALLER_ADDR2;
5373 if (in_lock_functions(addr))
5374 addr = CALLER_ADDR3;
5379 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5380 defined(CONFIG_PREEMPT_TRACER))
5382 void __kprobes add_preempt_count(int val)
5384 #ifdef CONFIG_DEBUG_PREEMPT
5388 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5391 preempt_count() += val;
5392 #ifdef CONFIG_DEBUG_PREEMPT
5394 * Spinlock count overflowing soon?
5396 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5399 if (preempt_count() == val)
5400 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5402 EXPORT_SYMBOL(add_preempt_count);
5404 void __kprobes sub_preempt_count(int val)
5406 #ifdef CONFIG_DEBUG_PREEMPT
5410 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5413 * Is the spinlock portion underflowing?
5415 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5416 !(preempt_count() & PREEMPT_MASK)))
5420 if (preempt_count() == val)
5421 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5422 preempt_count() -= val;
5424 EXPORT_SYMBOL(sub_preempt_count);
5429 * Print scheduling while atomic bug:
5431 static noinline void __schedule_bug(struct task_struct *prev)
5433 struct pt_regs *regs = get_irq_regs();
5435 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5436 prev->comm, prev->pid, preempt_count());
5438 debug_show_held_locks(prev);
5440 if (irqs_disabled())
5441 print_irqtrace_events(prev);
5450 * Various schedule()-time debugging checks and statistics:
5452 static inline void schedule_debug(struct task_struct *prev)
5455 * Test if we are atomic. Since do_exit() needs to call into
5456 * schedule() atomically, we ignore that path for now.
5457 * Otherwise, whine if we are scheduling when we should not be.
5459 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5460 __schedule_bug(prev);
5462 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5464 schedstat_inc(this_rq(), sched_count);
5465 #ifdef CONFIG_SCHEDSTATS
5466 if (unlikely(prev->lock_depth >= 0)) {
5467 schedstat_inc(this_rq(), bkl_count);
5468 schedstat_inc(prev, sched_info.bkl_count);
5473 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5475 if (prev->state == TASK_RUNNING) {
5476 u64 runtime = prev->se.sum_exec_runtime;
5478 runtime -= prev->se.prev_sum_exec_runtime;
5479 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5482 * In order to avoid avg_overlap growing stale when we are
5483 * indeed overlapping and hence not getting put to sleep, grow
5484 * the avg_overlap on preemption.
5486 * We use the average preemption runtime because that
5487 * correlates to the amount of cache footprint a task can
5490 update_avg(&prev->se.avg_overlap, runtime);
5492 prev->sched_class->put_prev_task(rq, prev);
5496 * Pick up the highest-prio task:
5498 static inline struct task_struct *
5499 pick_next_task(struct rq *rq)
5501 const struct sched_class *class;
5502 struct task_struct *p;
5505 * Optimization: we know that if all tasks are in
5506 * the fair class we can call that function directly:
5508 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5509 p = fair_sched_class.pick_next_task(rq);
5514 class = sched_class_highest;
5516 p = class->pick_next_task(rq);
5520 * Will never be NULL as the idle class always
5521 * returns a non-NULL p:
5523 class = class->next;
5528 * schedule() is the main scheduler function.
5530 asmlinkage void __sched schedule(void)
5532 struct task_struct *prev, *next;
5533 unsigned long *switch_count;
5539 cpu = smp_processor_id();
5543 switch_count = &prev->nivcsw;
5545 release_kernel_lock(prev);
5546 need_resched_nonpreemptible:
5548 schedule_debug(prev);
5550 if (sched_feat(HRTICK))
5553 raw_spin_lock_irq(&rq->lock);
5554 update_rq_clock(rq);
5555 clear_tsk_need_resched(prev);
5557 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5558 if (unlikely(signal_pending_state(prev->state, prev)))
5559 prev->state = TASK_RUNNING;
5561 deactivate_task(rq, prev, 1);
5562 switch_count = &prev->nvcsw;
5565 pre_schedule(rq, prev);
5567 if (unlikely(!rq->nr_running))
5568 idle_balance(cpu, rq);
5570 put_prev_task(rq, prev);
5571 next = pick_next_task(rq);
5573 if (likely(prev != next)) {
5574 sched_info_switch(prev, next);
5575 perf_event_task_sched_out(prev, next, cpu);
5581 context_switch(rq, prev, next); /* unlocks the rq */
5583 * the context switch might have flipped the stack from under
5584 * us, hence refresh the local variables.
5586 cpu = smp_processor_id();
5589 raw_spin_unlock_irq(&rq->lock);
5593 if (unlikely(reacquire_kernel_lock(current) < 0)) {
5595 switch_count = &prev->nivcsw;
5596 goto need_resched_nonpreemptible;
5599 preempt_enable_no_resched();
5603 EXPORT_SYMBOL(schedule);
5605 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5607 * Look out! "owner" is an entirely speculative pointer
5608 * access and not reliable.
5610 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5615 if (!sched_feat(OWNER_SPIN))
5618 #ifdef CONFIG_DEBUG_PAGEALLOC
5620 * Need to access the cpu field knowing that
5621 * DEBUG_PAGEALLOC could have unmapped it if
5622 * the mutex owner just released it and exited.
5624 if (probe_kernel_address(&owner->cpu, cpu))
5631 * Even if the access succeeded (likely case),
5632 * the cpu field may no longer be valid.
5634 if (cpu >= nr_cpumask_bits)
5638 * We need to validate that we can do a
5639 * get_cpu() and that we have the percpu area.
5641 if (!cpu_online(cpu))
5648 * Owner changed, break to re-assess state.
5650 if (lock->owner != owner)
5654 * Is that owner really running on that cpu?
5656 if (task_thread_info(rq->curr) != owner || need_resched())
5666 #ifdef CONFIG_PREEMPT
5668 * this is the entry point to schedule() from in-kernel preemption
5669 * off of preempt_enable. Kernel preemptions off return from interrupt
5670 * occur there and call schedule directly.
5672 asmlinkage void __sched preempt_schedule(void)
5674 struct thread_info *ti = current_thread_info();
5677 * If there is a non-zero preempt_count or interrupts are disabled,
5678 * we do not want to preempt the current task. Just return..
5680 if (likely(ti->preempt_count || irqs_disabled()))
5684 add_preempt_count(PREEMPT_ACTIVE);
5686 sub_preempt_count(PREEMPT_ACTIVE);
5689 * Check again in case we missed a preemption opportunity
5690 * between schedule and now.
5693 } while (need_resched());
5695 EXPORT_SYMBOL(preempt_schedule);
5698 * this is the entry point to schedule() from kernel preemption
5699 * off of irq context.
5700 * Note, that this is called and return with irqs disabled. This will
5701 * protect us against recursive calling from irq.
5703 asmlinkage void __sched preempt_schedule_irq(void)
5705 struct thread_info *ti = current_thread_info();
5707 /* Catch callers which need to be fixed */
5708 BUG_ON(ti->preempt_count || !irqs_disabled());
5711 add_preempt_count(PREEMPT_ACTIVE);
5714 local_irq_disable();
5715 sub_preempt_count(PREEMPT_ACTIVE);
5718 * Check again in case we missed a preemption opportunity
5719 * between schedule and now.
5722 } while (need_resched());
5725 #endif /* CONFIG_PREEMPT */
5727 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5730 return try_to_wake_up(curr->private, mode, wake_flags);
5732 EXPORT_SYMBOL(default_wake_function);
5735 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5736 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5737 * number) then we wake all the non-exclusive tasks and one exclusive task.
5739 * There are circumstances in which we can try to wake a task which has already
5740 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5741 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5743 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5744 int nr_exclusive, int wake_flags, void *key)
5746 wait_queue_t *curr, *next;
5748 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5749 unsigned flags = curr->flags;
5751 if (curr->func(curr, mode, wake_flags, key) &&
5752 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5758 * __wake_up - wake up threads blocked on a waitqueue.
5760 * @mode: which threads
5761 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5762 * @key: is directly passed to the wakeup function
5764 * It may be assumed that this function implies a write memory barrier before
5765 * changing the task state if and only if any tasks are woken up.
5767 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5768 int nr_exclusive, void *key)
5770 unsigned long flags;
5772 spin_lock_irqsave(&q->lock, flags);
5773 __wake_up_common(q, mode, nr_exclusive, 0, key);
5774 spin_unlock_irqrestore(&q->lock, flags);
5776 EXPORT_SYMBOL(__wake_up);
5779 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5781 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5783 __wake_up_common(q, mode, 1, 0, NULL);
5786 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5788 __wake_up_common(q, mode, 1, 0, key);
5792 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5794 * @mode: which threads
5795 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5796 * @key: opaque value to be passed to wakeup targets
5798 * The sync wakeup differs that the waker knows that it will schedule
5799 * away soon, so while the target thread will be woken up, it will not
5800 * be migrated to another CPU - ie. the two threads are 'synchronized'
5801 * with each other. This can prevent needless bouncing between CPUs.
5803 * On UP it can prevent extra preemption.
5805 * It may be assumed that this function implies a write memory barrier before
5806 * changing the task state if and only if any tasks are woken up.
5808 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5809 int nr_exclusive, void *key)
5811 unsigned long flags;
5812 int wake_flags = WF_SYNC;
5817 if (unlikely(!nr_exclusive))
5820 spin_lock_irqsave(&q->lock, flags);
5821 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5822 spin_unlock_irqrestore(&q->lock, flags);
5824 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5827 * __wake_up_sync - see __wake_up_sync_key()
5829 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5831 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5833 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5836 * complete: - signals a single thread waiting on this completion
5837 * @x: holds the state of this particular completion
5839 * This will wake up a single thread waiting on this completion. Threads will be
5840 * awakened in the same order in which they were queued.
5842 * See also complete_all(), wait_for_completion() and related routines.
5844 * It may be assumed that this function implies a write memory barrier before
5845 * changing the task state if and only if any tasks are woken up.
5847 void complete(struct completion *x)
5849 unsigned long flags;
5851 spin_lock_irqsave(&x->wait.lock, flags);
5853 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5854 spin_unlock_irqrestore(&x->wait.lock, flags);
5856 EXPORT_SYMBOL(complete);
5859 * complete_all: - signals all threads waiting on this completion
5860 * @x: holds the state of this particular completion
5862 * This will wake up all threads waiting on this particular completion event.
5864 * It may be assumed that this function implies a write memory barrier before
5865 * changing the task state if and only if any tasks are woken up.
5867 void complete_all(struct completion *x)
5869 unsigned long flags;
5871 spin_lock_irqsave(&x->wait.lock, flags);
5872 x->done += UINT_MAX/2;
5873 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5874 spin_unlock_irqrestore(&x->wait.lock, flags);
5876 EXPORT_SYMBOL(complete_all);
5878 static inline long __sched
5879 do_wait_for_common(struct completion *x, long timeout, int state)
5882 DECLARE_WAITQUEUE(wait, current);
5884 wait.flags |= WQ_FLAG_EXCLUSIVE;
5885 __add_wait_queue_tail(&x->wait, &wait);
5887 if (signal_pending_state(state, current)) {
5888 timeout = -ERESTARTSYS;
5891 __set_current_state(state);
5892 spin_unlock_irq(&x->wait.lock);
5893 timeout = schedule_timeout(timeout);
5894 spin_lock_irq(&x->wait.lock);
5895 } while (!x->done && timeout);
5896 __remove_wait_queue(&x->wait, &wait);
5901 return timeout ?: 1;
5905 wait_for_common(struct completion *x, long timeout, int state)
5909 spin_lock_irq(&x->wait.lock);
5910 timeout = do_wait_for_common(x, timeout, state);
5911 spin_unlock_irq(&x->wait.lock);
5916 * wait_for_completion: - waits for completion of a task
5917 * @x: holds the state of this particular completion
5919 * This waits to be signaled for completion of a specific task. It is NOT
5920 * interruptible and there is no timeout.
5922 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5923 * and interrupt capability. Also see complete().
5925 void __sched wait_for_completion(struct completion *x)
5927 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5929 EXPORT_SYMBOL(wait_for_completion);
5932 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5933 * @x: holds the state of this particular completion
5934 * @timeout: timeout value in jiffies
5936 * This waits for either a completion of a specific task to be signaled or for a
5937 * specified timeout to expire. The timeout is in jiffies. It is not
5940 unsigned long __sched
5941 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5943 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5945 EXPORT_SYMBOL(wait_for_completion_timeout);
5948 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5949 * @x: holds the state of this particular completion
5951 * This waits for completion of a specific task to be signaled. It is
5954 int __sched wait_for_completion_interruptible(struct completion *x)
5956 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5957 if (t == -ERESTARTSYS)
5961 EXPORT_SYMBOL(wait_for_completion_interruptible);
5964 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5965 * @x: holds the state of this particular completion
5966 * @timeout: timeout value in jiffies
5968 * This waits for either a completion of a specific task to be signaled or for a
5969 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5971 unsigned long __sched
5972 wait_for_completion_interruptible_timeout(struct completion *x,
5973 unsigned long timeout)
5975 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5977 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5980 * wait_for_completion_killable: - waits for completion of a task (killable)
5981 * @x: holds the state of this particular completion
5983 * This waits to be signaled for completion of a specific task. It can be
5984 * interrupted by a kill signal.
5986 int __sched wait_for_completion_killable(struct completion *x)
5988 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5989 if (t == -ERESTARTSYS)
5993 EXPORT_SYMBOL(wait_for_completion_killable);
5996 * try_wait_for_completion - try to decrement a completion without blocking
5997 * @x: completion structure
5999 * Returns: 0 if a decrement cannot be done without blocking
6000 * 1 if a decrement succeeded.
6002 * If a completion is being used as a counting completion,
6003 * attempt to decrement the counter without blocking. This
6004 * enables us to avoid waiting if the resource the completion
6005 * is protecting is not available.
6007 bool try_wait_for_completion(struct completion *x)
6009 unsigned long flags;
6012 spin_lock_irqsave(&x->wait.lock, flags);
6017 spin_unlock_irqrestore(&x->wait.lock, flags);
6020 EXPORT_SYMBOL(try_wait_for_completion);
6023 * completion_done - Test to see if a completion has any waiters
6024 * @x: completion structure
6026 * Returns: 0 if there are waiters (wait_for_completion() in progress)
6027 * 1 if there are no waiters.
6030 bool completion_done(struct completion *x)
6032 unsigned long flags;
6035 spin_lock_irqsave(&x->wait.lock, flags);
6038 spin_unlock_irqrestore(&x->wait.lock, flags);
6041 EXPORT_SYMBOL(completion_done);
6044 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
6046 unsigned long flags;
6049 init_waitqueue_entry(&wait, current);
6051 __set_current_state(state);
6053 spin_lock_irqsave(&q->lock, flags);
6054 __add_wait_queue(q, &wait);
6055 spin_unlock(&q->lock);
6056 timeout = schedule_timeout(timeout);
6057 spin_lock_irq(&q->lock);
6058 __remove_wait_queue(q, &wait);
6059 spin_unlock_irqrestore(&q->lock, flags);
6064 void __sched interruptible_sleep_on(wait_queue_head_t *q)
6066 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6068 EXPORT_SYMBOL(interruptible_sleep_on);
6071 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
6073 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
6075 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
6077 void __sched sleep_on(wait_queue_head_t *q)
6079 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6081 EXPORT_SYMBOL(sleep_on);
6083 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
6085 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
6087 EXPORT_SYMBOL(sleep_on_timeout);
6089 #ifdef CONFIG_RT_MUTEXES
6092 * rt_mutex_setprio - set the current priority of a task
6094 * @prio: prio value (kernel-internal form)
6096 * This function changes the 'effective' priority of a task. It does
6097 * not touch ->normal_prio like __setscheduler().
6099 * Used by the rt_mutex code to implement priority inheritance logic.
6101 void rt_mutex_setprio(struct task_struct *p, int prio)
6103 unsigned long flags;
6104 int oldprio, on_rq, running;
6106 const struct sched_class *prev_class = p->sched_class;
6108 BUG_ON(prio < 0 || prio > MAX_PRIO);
6110 rq = task_rq_lock(p, &flags);
6111 update_rq_clock(rq);
6114 on_rq = p->se.on_rq;
6115 running = task_current(rq, p);
6117 dequeue_task(rq, p, 0);
6119 p->sched_class->put_prev_task(rq, p);
6122 p->sched_class = &rt_sched_class;
6124 p->sched_class = &fair_sched_class;
6129 p->sched_class->set_curr_task(rq);
6131 enqueue_task(rq, p, 0);
6133 check_class_changed(rq, p, prev_class, oldprio, running);
6135 task_rq_unlock(rq, &flags);
6140 void set_user_nice(struct task_struct *p, long nice)
6142 int old_prio, delta, on_rq;
6143 unsigned long flags;
6146 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6149 * We have to be careful, if called from sys_setpriority(),
6150 * the task might be in the middle of scheduling on another CPU.
6152 rq = task_rq_lock(p, &flags);
6153 update_rq_clock(rq);
6155 * The RT priorities are set via sched_setscheduler(), but we still
6156 * allow the 'normal' nice value to be set - but as expected
6157 * it wont have any effect on scheduling until the task is
6158 * SCHED_FIFO/SCHED_RR:
6160 if (task_has_rt_policy(p)) {
6161 p->static_prio = NICE_TO_PRIO(nice);
6164 on_rq = p->se.on_rq;
6166 dequeue_task(rq, p, 0);
6168 p->static_prio = NICE_TO_PRIO(nice);
6171 p->prio = effective_prio(p);
6172 delta = p->prio - old_prio;
6175 enqueue_task(rq, p, 0);
6177 * If the task increased its priority or is running and
6178 * lowered its priority, then reschedule its CPU:
6180 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6181 resched_task(rq->curr);
6184 task_rq_unlock(rq, &flags);
6186 EXPORT_SYMBOL(set_user_nice);
6189 * can_nice - check if a task can reduce its nice value
6193 int can_nice(const struct task_struct *p, const int nice)
6195 /* convert nice value [19,-20] to rlimit style value [1,40] */
6196 int nice_rlim = 20 - nice;
6198 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6199 capable(CAP_SYS_NICE));
6202 #ifdef __ARCH_WANT_SYS_NICE
6205 * sys_nice - change the priority of the current process.
6206 * @increment: priority increment
6208 * sys_setpriority is a more generic, but much slower function that
6209 * does similar things.
6211 SYSCALL_DEFINE1(nice, int, increment)
6216 * Setpriority might change our priority at the same moment.
6217 * We don't have to worry. Conceptually one call occurs first
6218 * and we have a single winner.
6220 if (increment < -40)
6225 nice = TASK_NICE(current) + increment;
6231 if (increment < 0 && !can_nice(current, nice))
6234 retval = security_task_setnice(current, nice);
6238 set_user_nice(current, nice);
6245 * task_prio - return the priority value of a given task.
6246 * @p: the task in question.
6248 * This is the priority value as seen by users in /proc.
6249 * RT tasks are offset by -200. Normal tasks are centered
6250 * around 0, value goes from -16 to +15.
6252 int task_prio(const struct task_struct *p)
6254 return p->prio - MAX_RT_PRIO;
6258 * task_nice - return the nice value of a given task.
6259 * @p: the task in question.
6261 int task_nice(const struct task_struct *p)
6263 return TASK_NICE(p);
6265 EXPORT_SYMBOL(task_nice);
6268 * idle_cpu - is a given cpu idle currently?
6269 * @cpu: the processor in question.
6271 int idle_cpu(int cpu)
6273 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6277 * idle_task - return the idle task for a given cpu.
6278 * @cpu: the processor in question.
6280 struct task_struct *idle_task(int cpu)
6282 return cpu_rq(cpu)->idle;
6286 * find_process_by_pid - find a process with a matching PID value.
6287 * @pid: the pid in question.
6289 static struct task_struct *find_process_by_pid(pid_t pid)
6291 return pid ? find_task_by_vpid(pid) : current;
6294 /* Actually do priority change: must hold rq lock. */
6296 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6298 BUG_ON(p->se.on_rq);
6301 p->rt_priority = prio;
6302 p->normal_prio = normal_prio(p);
6303 /* we are holding p->pi_lock already */
6304 p->prio = rt_mutex_getprio(p);
6305 if (rt_prio(p->prio))
6306 p->sched_class = &rt_sched_class;
6308 p->sched_class = &fair_sched_class;
6313 * check the target process has a UID that matches the current process's
6315 static bool check_same_owner(struct task_struct *p)
6317 const struct cred *cred = current_cred(), *pcred;
6321 pcred = __task_cred(p);
6322 match = (cred->euid == pcred->euid ||
6323 cred->euid == pcred->uid);
6328 static int __sched_setscheduler(struct task_struct *p, int policy,
6329 struct sched_param *param, bool user)
6331 int retval, oldprio, oldpolicy = -1, on_rq, running;
6332 unsigned long flags;
6333 const struct sched_class *prev_class = p->sched_class;
6337 /* may grab non-irq protected spin_locks */
6338 BUG_ON(in_interrupt());
6340 /* double check policy once rq lock held */
6342 reset_on_fork = p->sched_reset_on_fork;
6343 policy = oldpolicy = p->policy;
6345 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6346 policy &= ~SCHED_RESET_ON_FORK;
6348 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6349 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6350 policy != SCHED_IDLE)
6355 * Valid priorities for SCHED_FIFO and SCHED_RR are
6356 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6357 * SCHED_BATCH and SCHED_IDLE is 0.
6359 if (param->sched_priority < 0 ||
6360 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6361 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6363 if (rt_policy(policy) != (param->sched_priority != 0))
6367 * Allow unprivileged RT tasks to decrease priority:
6369 if (user && !capable(CAP_SYS_NICE)) {
6370 if (rt_policy(policy)) {
6371 unsigned long rlim_rtprio;
6373 if (!lock_task_sighand(p, &flags))
6375 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6376 unlock_task_sighand(p, &flags);
6378 /* can't set/change the rt policy */
6379 if (policy != p->policy && !rlim_rtprio)
6382 /* can't increase priority */
6383 if (param->sched_priority > p->rt_priority &&
6384 param->sched_priority > rlim_rtprio)
6388 * Like positive nice levels, dont allow tasks to
6389 * move out of SCHED_IDLE either:
6391 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6394 /* can't change other user's priorities */
6395 if (!check_same_owner(p))
6398 /* Normal users shall not reset the sched_reset_on_fork flag */
6399 if (p->sched_reset_on_fork && !reset_on_fork)
6404 #ifdef CONFIG_RT_GROUP_SCHED
6406 * Do not allow realtime tasks into groups that have no runtime
6409 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6410 task_group(p)->rt_bandwidth.rt_runtime == 0)
6414 retval = security_task_setscheduler(p, policy, param);
6420 * make sure no PI-waiters arrive (or leave) while we are
6421 * changing the priority of the task:
6423 raw_spin_lock_irqsave(&p->pi_lock, flags);
6425 * To be able to change p->policy safely, the apropriate
6426 * runqueue lock must be held.
6428 rq = __task_rq_lock(p);
6429 /* recheck policy now with rq lock held */
6430 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6431 policy = oldpolicy = -1;
6432 __task_rq_unlock(rq);
6433 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6436 update_rq_clock(rq);
6437 on_rq = p->se.on_rq;
6438 running = task_current(rq, p);
6440 deactivate_task(rq, p, 0);
6442 p->sched_class->put_prev_task(rq, p);
6444 p->sched_reset_on_fork = reset_on_fork;
6447 __setscheduler(rq, p, policy, param->sched_priority);
6450 p->sched_class->set_curr_task(rq);
6452 activate_task(rq, p, 0);
6454 check_class_changed(rq, p, prev_class, oldprio, running);
6456 __task_rq_unlock(rq);
6457 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6459 rt_mutex_adjust_pi(p);
6465 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6466 * @p: the task in question.
6467 * @policy: new policy.
6468 * @param: structure containing the new RT priority.
6470 * NOTE that the task may be already dead.
6472 int sched_setscheduler(struct task_struct *p, int policy,
6473 struct sched_param *param)
6475 return __sched_setscheduler(p, policy, param, true);
6477 EXPORT_SYMBOL_GPL(sched_setscheduler);
6480 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6481 * @p: the task in question.
6482 * @policy: new policy.
6483 * @param: structure containing the new RT priority.
6485 * Just like sched_setscheduler, only don't bother checking if the
6486 * current context has permission. For example, this is needed in
6487 * stop_machine(): we create temporary high priority worker threads,
6488 * but our caller might not have that capability.
6490 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6491 struct sched_param *param)
6493 return __sched_setscheduler(p, policy, param, false);
6497 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6499 struct sched_param lparam;
6500 struct task_struct *p;
6503 if (!param || pid < 0)
6505 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6510 p = find_process_by_pid(pid);
6512 retval = sched_setscheduler(p, policy, &lparam);
6519 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6520 * @pid: the pid in question.
6521 * @policy: new policy.
6522 * @param: structure containing the new RT priority.
6524 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6525 struct sched_param __user *, param)
6527 /* negative values for policy are not valid */
6531 return do_sched_setscheduler(pid, policy, param);
6535 * sys_sched_setparam - set/change the RT priority of a thread
6536 * @pid: the pid in question.
6537 * @param: structure containing the new RT priority.
6539 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6541 return do_sched_setscheduler(pid, -1, param);
6545 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6546 * @pid: the pid in question.
6548 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6550 struct task_struct *p;
6558 p = find_process_by_pid(pid);
6560 retval = security_task_getscheduler(p);
6563 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6570 * sys_sched_getparam - get the RT priority of a thread
6571 * @pid: the pid in question.
6572 * @param: structure containing the RT priority.
6574 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6576 struct sched_param lp;
6577 struct task_struct *p;
6580 if (!param || pid < 0)
6584 p = find_process_by_pid(pid);
6589 retval = security_task_getscheduler(p);
6593 lp.sched_priority = p->rt_priority;
6597 * This one might sleep, we cannot do it with a spinlock held ...
6599 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6608 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6610 cpumask_var_t cpus_allowed, new_mask;
6611 struct task_struct *p;
6617 p = find_process_by_pid(pid);
6624 /* Prevent p going away */
6628 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6632 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6634 goto out_free_cpus_allowed;
6637 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6640 retval = security_task_setscheduler(p, 0, NULL);
6644 cpuset_cpus_allowed(p, cpus_allowed);
6645 cpumask_and(new_mask, in_mask, cpus_allowed);
6647 retval = set_cpus_allowed_ptr(p, new_mask);
6650 cpuset_cpus_allowed(p, cpus_allowed);
6651 if (!cpumask_subset(new_mask, cpus_allowed)) {
6653 * We must have raced with a concurrent cpuset
6654 * update. Just reset the cpus_allowed to the
6655 * cpuset's cpus_allowed
6657 cpumask_copy(new_mask, cpus_allowed);
6662 free_cpumask_var(new_mask);
6663 out_free_cpus_allowed:
6664 free_cpumask_var(cpus_allowed);
6671 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6672 struct cpumask *new_mask)
6674 if (len < cpumask_size())
6675 cpumask_clear(new_mask);
6676 else if (len > cpumask_size())
6677 len = cpumask_size();
6679 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6683 * sys_sched_setaffinity - set the cpu affinity of a process
6684 * @pid: pid of the process
6685 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6686 * @user_mask_ptr: user-space pointer to the new cpu mask
6688 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6689 unsigned long __user *, user_mask_ptr)
6691 cpumask_var_t new_mask;
6694 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6697 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6699 retval = sched_setaffinity(pid, new_mask);
6700 free_cpumask_var(new_mask);
6704 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6706 struct task_struct *p;
6707 unsigned long flags;
6715 p = find_process_by_pid(pid);
6719 retval = security_task_getscheduler(p);
6723 rq = task_rq_lock(p, &flags);
6724 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6725 task_rq_unlock(rq, &flags);
6735 * sys_sched_getaffinity - get 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 hold the current cpu mask
6740 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6741 unsigned long __user *, user_mask_ptr)
6746 if (len < cpumask_size())
6749 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6752 ret = sched_getaffinity(pid, mask);
6754 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6757 ret = cpumask_size();
6759 free_cpumask_var(mask);
6765 * sys_sched_yield - yield the current processor to other threads.
6767 * This function yields the current CPU to other tasks. If there are no
6768 * other threads running on this CPU then this function will return.
6770 SYSCALL_DEFINE0(sched_yield)
6772 struct rq *rq = this_rq_lock();
6774 schedstat_inc(rq, yld_count);
6775 current->sched_class->yield_task(rq);
6778 * Since we are going to call schedule() anyway, there's
6779 * no need to preempt or enable interrupts:
6781 __release(rq->lock);
6782 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6783 do_raw_spin_unlock(&rq->lock);
6784 preempt_enable_no_resched();
6791 static inline int should_resched(void)
6793 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6796 static void __cond_resched(void)
6798 add_preempt_count(PREEMPT_ACTIVE);
6800 sub_preempt_count(PREEMPT_ACTIVE);
6803 int __sched _cond_resched(void)
6805 if (should_resched()) {
6811 EXPORT_SYMBOL(_cond_resched);
6814 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6815 * call schedule, and on return reacquire the lock.
6817 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6818 * operations here to prevent schedule() from being called twice (once via
6819 * spin_unlock(), once by hand).
6821 int __cond_resched_lock(spinlock_t *lock)
6823 int resched = should_resched();
6826 lockdep_assert_held(lock);
6828 if (spin_needbreak(lock) || resched) {
6839 EXPORT_SYMBOL(__cond_resched_lock);
6841 int __sched __cond_resched_softirq(void)
6843 BUG_ON(!in_softirq());
6845 if (should_resched()) {
6853 EXPORT_SYMBOL(__cond_resched_softirq);
6856 * yield - yield the current processor to other threads.
6858 * This is a shortcut for kernel-space yielding - it marks the
6859 * thread runnable and calls sys_sched_yield().
6861 void __sched yield(void)
6863 set_current_state(TASK_RUNNING);
6866 EXPORT_SYMBOL(yield);
6869 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6870 * that process accounting knows that this is a task in IO wait state.
6872 void __sched io_schedule(void)
6874 struct rq *rq = raw_rq();
6876 delayacct_blkio_start();
6877 atomic_inc(&rq->nr_iowait);
6878 current->in_iowait = 1;
6880 current->in_iowait = 0;
6881 atomic_dec(&rq->nr_iowait);
6882 delayacct_blkio_end();
6884 EXPORT_SYMBOL(io_schedule);
6886 long __sched io_schedule_timeout(long timeout)
6888 struct rq *rq = raw_rq();
6891 delayacct_blkio_start();
6892 atomic_inc(&rq->nr_iowait);
6893 current->in_iowait = 1;
6894 ret = schedule_timeout(timeout);
6895 current->in_iowait = 0;
6896 atomic_dec(&rq->nr_iowait);
6897 delayacct_blkio_end();
6902 * sys_sched_get_priority_max - return maximum RT priority.
6903 * @policy: scheduling class.
6905 * this syscall returns the maximum rt_priority that can be used
6906 * by a given scheduling class.
6908 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6915 ret = MAX_USER_RT_PRIO-1;
6927 * sys_sched_get_priority_min - return minimum RT priority.
6928 * @policy: scheduling class.
6930 * this syscall returns the minimum rt_priority that can be used
6931 * by a given scheduling class.
6933 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6951 * sys_sched_rr_get_interval - return the default timeslice of a process.
6952 * @pid: pid of the process.
6953 * @interval: userspace pointer to the timeslice value.
6955 * this syscall writes the default timeslice value of a given process
6956 * into the user-space timespec buffer. A value of '0' means infinity.
6958 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6959 struct timespec __user *, interval)
6961 struct task_struct *p;
6962 unsigned int time_slice;
6963 unsigned long flags;
6973 p = find_process_by_pid(pid);
6977 retval = security_task_getscheduler(p);
6981 rq = task_rq_lock(p, &flags);
6982 time_slice = p->sched_class->get_rr_interval(rq, p);
6983 task_rq_unlock(rq, &flags);
6986 jiffies_to_timespec(time_slice, &t);
6987 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6995 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6997 void sched_show_task(struct task_struct *p)
6999 unsigned long free = 0;
7002 state = p->state ? __ffs(p->state) + 1 : 0;
7003 printk(KERN_INFO "%-13.13s %c", p->comm,
7004 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
7005 #if BITS_PER_LONG == 32
7006 if (state == TASK_RUNNING)
7007 printk(KERN_CONT " running ");
7009 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
7011 if (state == TASK_RUNNING)
7012 printk(KERN_CONT " running task ");
7014 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
7016 #ifdef CONFIG_DEBUG_STACK_USAGE
7017 free = stack_not_used(p);
7019 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
7020 task_pid_nr(p), task_pid_nr(p->real_parent),
7021 (unsigned long)task_thread_info(p)->flags);
7023 show_stack(p, NULL);
7026 void show_state_filter(unsigned long state_filter)
7028 struct task_struct *g, *p;
7030 #if BITS_PER_LONG == 32
7032 " task PC stack pid father\n");
7035 " task PC stack pid father\n");
7037 read_lock(&tasklist_lock);
7038 do_each_thread(g, p) {
7040 * reset the NMI-timeout, listing all files on a slow
7041 * console might take alot of time:
7043 touch_nmi_watchdog();
7044 if (!state_filter || (p->state & state_filter))
7046 } while_each_thread(g, p);
7048 touch_all_softlockup_watchdogs();
7050 #ifdef CONFIG_SCHED_DEBUG
7051 sysrq_sched_debug_show();
7053 read_unlock(&tasklist_lock);
7055 * Only show locks if all tasks are dumped:
7058 debug_show_all_locks();
7061 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
7063 idle->sched_class = &idle_sched_class;
7067 * init_idle - set up an idle thread for a given CPU
7068 * @idle: task in question
7069 * @cpu: cpu the idle task belongs to
7071 * NOTE: this function does not set the idle thread's NEED_RESCHED
7072 * flag, to make booting more robust.
7074 void __cpuinit init_idle(struct task_struct *idle, int cpu)
7076 struct rq *rq = cpu_rq(cpu);
7077 unsigned long flags;
7079 raw_spin_lock_irqsave(&rq->lock, flags);
7082 idle->state = TASK_RUNNING;
7083 idle->se.exec_start = sched_clock();
7085 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
7086 __set_task_cpu(idle, cpu);
7088 rq->curr = rq->idle = idle;
7089 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7092 raw_spin_unlock_irqrestore(&rq->lock, flags);
7094 /* Set the preempt count _outside_ the spinlocks! */
7095 #if defined(CONFIG_PREEMPT)
7096 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7098 task_thread_info(idle)->preempt_count = 0;
7101 * The idle tasks have their own, simple scheduling class:
7103 idle->sched_class = &idle_sched_class;
7104 ftrace_graph_init_task(idle);
7108 * In a system that switches off the HZ timer nohz_cpu_mask
7109 * indicates which cpus entered this state. This is used
7110 * in the rcu update to wait only for active cpus. For system
7111 * which do not switch off the HZ timer nohz_cpu_mask should
7112 * always be CPU_BITS_NONE.
7114 cpumask_var_t nohz_cpu_mask;
7117 * Increase the granularity value when there are more CPUs,
7118 * because with more CPUs the 'effective latency' as visible
7119 * to users decreases. But the relationship is not linear,
7120 * so pick a second-best guess by going with the log2 of the
7123 * This idea comes from the SD scheduler of Con Kolivas:
7125 static int get_update_sysctl_factor(void)
7127 unsigned int cpus = min_t(int, num_online_cpus(), 8);
7128 unsigned int factor;
7130 switch (sysctl_sched_tunable_scaling) {
7131 case SCHED_TUNABLESCALING_NONE:
7134 case SCHED_TUNABLESCALING_LINEAR:
7137 case SCHED_TUNABLESCALING_LOG:
7139 factor = 1 + ilog2(cpus);
7146 static void update_sysctl(void)
7148 unsigned int factor = get_update_sysctl_factor();
7150 #define SET_SYSCTL(name) \
7151 (sysctl_##name = (factor) * normalized_sysctl_##name)
7152 SET_SYSCTL(sched_min_granularity);
7153 SET_SYSCTL(sched_latency);
7154 SET_SYSCTL(sched_wakeup_granularity);
7155 SET_SYSCTL(sched_shares_ratelimit);
7159 static inline void sched_init_granularity(void)
7166 * This is how migration works:
7168 * 1) we queue a struct migration_req structure in the source CPU's
7169 * runqueue and wake up that CPU's migration thread.
7170 * 2) we down() the locked semaphore => thread blocks.
7171 * 3) migration thread wakes up (implicitly it forces the migrated
7172 * thread off the CPU)
7173 * 4) it gets the migration request and checks whether the migrated
7174 * task is still in the wrong runqueue.
7175 * 5) if it's in the wrong runqueue then the migration thread removes
7176 * it and puts it into the right queue.
7177 * 6) migration thread up()s the semaphore.
7178 * 7) we wake up and the migration is done.
7182 * Change a given task's CPU affinity. Migrate the thread to a
7183 * proper CPU and schedule it away if the CPU it's executing on
7184 * is removed from the allowed bitmask.
7186 * NOTE: the caller must have a valid reference to the task, the
7187 * task must not exit() & deallocate itself prematurely. The
7188 * call is not atomic; no spinlocks may be held.
7190 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7192 struct migration_req req;
7193 unsigned long flags;
7197 rq = task_rq_lock(p, &flags);
7199 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7204 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7205 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7210 if (p->sched_class->set_cpus_allowed)
7211 p->sched_class->set_cpus_allowed(p, new_mask);
7213 cpumask_copy(&p->cpus_allowed, new_mask);
7214 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7217 /* Can the task run on the task's current CPU? If so, we're done */
7218 if (cpumask_test_cpu(task_cpu(p), new_mask))
7221 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7222 /* Need help from migration thread: drop lock and wait. */
7223 struct task_struct *mt = rq->migration_thread;
7225 get_task_struct(mt);
7226 task_rq_unlock(rq, &flags);
7227 wake_up_process(rq->migration_thread);
7228 put_task_struct(mt);
7229 wait_for_completion(&req.done);
7230 tlb_migrate_finish(p->mm);
7234 task_rq_unlock(rq, &flags);
7238 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7241 * Move (not current) task off this cpu, onto dest cpu. We're doing
7242 * this because either it can't run here any more (set_cpus_allowed()
7243 * away from this CPU, or CPU going down), or because we're
7244 * attempting to rebalance this task on exec (sched_exec).
7246 * So we race with normal scheduler movements, but that's OK, as long
7247 * as the task is no longer on this CPU.
7249 * Returns non-zero if task was successfully migrated.
7251 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7253 struct rq *rq_dest, *rq_src;
7256 if (unlikely(!cpu_active(dest_cpu)))
7259 rq_src = cpu_rq(src_cpu);
7260 rq_dest = cpu_rq(dest_cpu);
7262 double_rq_lock(rq_src, rq_dest);
7263 /* Already moved. */
7264 if (task_cpu(p) != src_cpu)
7266 /* Affinity changed (again). */
7267 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7271 * If we're not on a rq, the next wake-up will ensure we're
7275 deactivate_task(rq_src, p, 0);
7276 set_task_cpu(p, dest_cpu);
7277 activate_task(rq_dest, p, 0);
7278 check_preempt_curr(rq_dest, p, 0);
7283 double_rq_unlock(rq_src, rq_dest);
7287 #define RCU_MIGRATION_IDLE 0
7288 #define RCU_MIGRATION_NEED_QS 1
7289 #define RCU_MIGRATION_GOT_QS 2
7290 #define RCU_MIGRATION_MUST_SYNC 3
7293 * migration_thread - this is a highprio system thread that performs
7294 * thread migration by bumping thread off CPU then 'pushing' onto
7297 static int migration_thread(void *data)
7300 int cpu = (long)data;
7304 BUG_ON(rq->migration_thread != current);
7306 set_current_state(TASK_INTERRUPTIBLE);
7307 while (!kthread_should_stop()) {
7308 struct migration_req *req;
7309 struct list_head *head;
7311 raw_spin_lock_irq(&rq->lock);
7313 if (cpu_is_offline(cpu)) {
7314 raw_spin_unlock_irq(&rq->lock);
7318 if (rq->active_balance) {
7319 active_load_balance(rq, cpu);
7320 rq->active_balance = 0;
7323 head = &rq->migration_queue;
7325 if (list_empty(head)) {
7326 raw_spin_unlock_irq(&rq->lock);
7328 set_current_state(TASK_INTERRUPTIBLE);
7331 req = list_entry(head->next, struct migration_req, list);
7332 list_del_init(head->next);
7334 if (req->task != NULL) {
7335 raw_spin_unlock(&rq->lock);
7336 __migrate_task(req->task, cpu, req->dest_cpu);
7337 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7338 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7339 raw_spin_unlock(&rq->lock);
7341 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7342 raw_spin_unlock(&rq->lock);
7343 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7347 complete(&req->done);
7349 __set_current_state(TASK_RUNNING);
7354 #ifdef CONFIG_HOTPLUG_CPU
7356 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7360 local_irq_disable();
7361 ret = __migrate_task(p, src_cpu, dest_cpu);
7367 * Figure out where task on dead CPU should go, use force if necessary.
7369 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7374 dest_cpu = select_fallback_rq(dead_cpu, p);
7376 /* It can have affinity changed while we were choosing. */
7377 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7382 * While a dead CPU has no uninterruptible tasks queued at this point,
7383 * it might still have a nonzero ->nr_uninterruptible counter, because
7384 * for performance reasons the counter is not stricly tracking tasks to
7385 * their home CPUs. So we just add the counter to another CPU's counter,
7386 * to keep the global sum constant after CPU-down:
7388 static void migrate_nr_uninterruptible(struct rq *rq_src)
7390 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7391 unsigned long flags;
7393 local_irq_save(flags);
7394 double_rq_lock(rq_src, rq_dest);
7395 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7396 rq_src->nr_uninterruptible = 0;
7397 double_rq_unlock(rq_src, rq_dest);
7398 local_irq_restore(flags);
7401 /* Run through task list and migrate tasks from the dead cpu. */
7402 static void migrate_live_tasks(int src_cpu)
7404 struct task_struct *p, *t;
7406 read_lock(&tasklist_lock);
7408 do_each_thread(t, p) {
7412 if (task_cpu(p) == src_cpu)
7413 move_task_off_dead_cpu(src_cpu, p);
7414 } while_each_thread(t, p);
7416 read_unlock(&tasklist_lock);
7420 * Schedules idle task to be the next runnable task on current CPU.
7421 * It does so by boosting its priority to highest possible.
7422 * Used by CPU offline code.
7424 void sched_idle_next(void)
7426 int this_cpu = smp_processor_id();
7427 struct rq *rq = cpu_rq(this_cpu);
7428 struct task_struct *p = rq->idle;
7429 unsigned long flags;
7431 /* cpu has to be offline */
7432 BUG_ON(cpu_online(this_cpu));
7435 * Strictly not necessary since rest of the CPUs are stopped by now
7436 * and interrupts disabled on the current cpu.
7438 raw_spin_lock_irqsave(&rq->lock, flags);
7440 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7442 update_rq_clock(rq);
7443 activate_task(rq, p, 0);
7445 raw_spin_unlock_irqrestore(&rq->lock, flags);
7449 * Ensures that the idle task is using init_mm right before its cpu goes
7452 void idle_task_exit(void)
7454 struct mm_struct *mm = current->active_mm;
7456 BUG_ON(cpu_online(smp_processor_id()));
7459 switch_mm(mm, &init_mm, current);
7463 /* called under rq->lock with disabled interrupts */
7464 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7466 struct rq *rq = cpu_rq(dead_cpu);
7468 /* Must be exiting, otherwise would be on tasklist. */
7469 BUG_ON(!p->exit_state);
7471 /* Cannot have done final schedule yet: would have vanished. */
7472 BUG_ON(p->state == TASK_DEAD);
7477 * Drop lock around migration; if someone else moves it,
7478 * that's OK. No task can be added to this CPU, so iteration is
7481 raw_spin_unlock_irq(&rq->lock);
7482 move_task_off_dead_cpu(dead_cpu, p);
7483 raw_spin_lock_irq(&rq->lock);
7488 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7489 static void migrate_dead_tasks(unsigned int dead_cpu)
7491 struct rq *rq = cpu_rq(dead_cpu);
7492 struct task_struct *next;
7495 if (!rq->nr_running)
7497 update_rq_clock(rq);
7498 next = pick_next_task(rq);
7501 next->sched_class->put_prev_task(rq, next);
7502 migrate_dead(dead_cpu, next);
7508 * remove the tasks which were accounted by rq from calc_load_tasks.
7510 static void calc_global_load_remove(struct rq *rq)
7512 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7513 rq->calc_load_active = 0;
7515 #endif /* CONFIG_HOTPLUG_CPU */
7517 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7519 static struct ctl_table sd_ctl_dir[] = {
7521 .procname = "sched_domain",
7527 static struct ctl_table sd_ctl_root[] = {
7529 .procname = "kernel",
7531 .child = sd_ctl_dir,
7536 static struct ctl_table *sd_alloc_ctl_entry(int n)
7538 struct ctl_table *entry =
7539 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7544 static void sd_free_ctl_entry(struct ctl_table **tablep)
7546 struct ctl_table *entry;
7549 * In the intermediate directories, both the child directory and
7550 * procname are dynamically allocated and could fail but the mode
7551 * will always be set. In the lowest directory the names are
7552 * static strings and all have proc handlers.
7554 for (entry = *tablep; entry->mode; entry++) {
7556 sd_free_ctl_entry(&entry->child);
7557 if (entry->proc_handler == NULL)
7558 kfree(entry->procname);
7566 set_table_entry(struct ctl_table *entry,
7567 const char *procname, void *data, int maxlen,
7568 mode_t mode, proc_handler *proc_handler)
7570 entry->procname = procname;
7572 entry->maxlen = maxlen;
7574 entry->proc_handler = proc_handler;
7577 static struct ctl_table *
7578 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7580 struct ctl_table *table = sd_alloc_ctl_entry(13);
7585 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7586 sizeof(long), 0644, proc_doulongvec_minmax);
7587 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7588 sizeof(long), 0644, proc_doulongvec_minmax);
7589 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7590 sizeof(int), 0644, proc_dointvec_minmax);
7591 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7592 sizeof(int), 0644, proc_dointvec_minmax);
7593 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7594 sizeof(int), 0644, proc_dointvec_minmax);
7595 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7596 sizeof(int), 0644, proc_dointvec_minmax);
7597 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7598 sizeof(int), 0644, proc_dointvec_minmax);
7599 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7600 sizeof(int), 0644, proc_dointvec_minmax);
7601 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7602 sizeof(int), 0644, proc_dointvec_minmax);
7603 set_table_entry(&table[9], "cache_nice_tries",
7604 &sd->cache_nice_tries,
7605 sizeof(int), 0644, proc_dointvec_minmax);
7606 set_table_entry(&table[10], "flags", &sd->flags,
7607 sizeof(int), 0644, proc_dointvec_minmax);
7608 set_table_entry(&table[11], "name", sd->name,
7609 CORENAME_MAX_SIZE, 0444, proc_dostring);
7610 /* &table[12] is terminator */
7615 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7617 struct ctl_table *entry, *table;
7618 struct sched_domain *sd;
7619 int domain_num = 0, i;
7622 for_each_domain(cpu, sd)
7624 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7629 for_each_domain(cpu, sd) {
7630 snprintf(buf, 32, "domain%d", i);
7631 entry->procname = kstrdup(buf, GFP_KERNEL);
7633 entry->child = sd_alloc_ctl_domain_table(sd);
7640 static struct ctl_table_header *sd_sysctl_header;
7641 static void register_sched_domain_sysctl(void)
7643 int i, cpu_num = num_possible_cpus();
7644 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7647 WARN_ON(sd_ctl_dir[0].child);
7648 sd_ctl_dir[0].child = entry;
7653 for_each_possible_cpu(i) {
7654 snprintf(buf, 32, "cpu%d", i);
7655 entry->procname = kstrdup(buf, GFP_KERNEL);
7657 entry->child = sd_alloc_ctl_cpu_table(i);
7661 WARN_ON(sd_sysctl_header);
7662 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7665 /* may be called multiple times per register */
7666 static void unregister_sched_domain_sysctl(void)
7668 if (sd_sysctl_header)
7669 unregister_sysctl_table(sd_sysctl_header);
7670 sd_sysctl_header = NULL;
7671 if (sd_ctl_dir[0].child)
7672 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7675 static void register_sched_domain_sysctl(void)
7678 static void unregister_sched_domain_sysctl(void)
7683 static void set_rq_online(struct rq *rq)
7686 const struct sched_class *class;
7688 cpumask_set_cpu(rq->cpu, rq->rd->online);
7691 for_each_class(class) {
7692 if (class->rq_online)
7693 class->rq_online(rq);
7698 static void set_rq_offline(struct rq *rq)
7701 const struct sched_class *class;
7703 for_each_class(class) {
7704 if (class->rq_offline)
7705 class->rq_offline(rq);
7708 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7714 * migration_call - callback that gets triggered when a CPU is added.
7715 * Here we can start up the necessary migration thread for the new CPU.
7717 static int __cpuinit
7718 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7720 struct task_struct *p;
7721 int cpu = (long)hcpu;
7722 unsigned long flags;
7727 case CPU_UP_PREPARE:
7728 case CPU_UP_PREPARE_FROZEN:
7729 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7732 kthread_bind(p, cpu);
7733 /* Must be high prio: stop_machine expects to yield to it. */
7734 rq = task_rq_lock(p, &flags);
7735 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7736 task_rq_unlock(rq, &flags);
7738 cpu_rq(cpu)->migration_thread = p;
7739 rq->calc_load_update = calc_load_update;
7743 case CPU_ONLINE_FROZEN:
7744 /* Strictly unnecessary, as first user will wake it. */
7745 wake_up_process(cpu_rq(cpu)->migration_thread);
7747 /* Update our root-domain */
7749 raw_spin_lock_irqsave(&rq->lock, flags);
7751 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7755 raw_spin_unlock_irqrestore(&rq->lock, flags);
7758 #ifdef CONFIG_HOTPLUG_CPU
7759 case CPU_UP_CANCELED:
7760 case CPU_UP_CANCELED_FROZEN:
7761 if (!cpu_rq(cpu)->migration_thread)
7763 /* Unbind it from offline cpu so it can run. Fall thru. */
7764 kthread_bind(cpu_rq(cpu)->migration_thread,
7765 cpumask_any(cpu_online_mask));
7766 kthread_stop(cpu_rq(cpu)->migration_thread);
7767 put_task_struct(cpu_rq(cpu)->migration_thread);
7768 cpu_rq(cpu)->migration_thread = NULL;
7772 case CPU_DEAD_FROZEN:
7773 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7774 migrate_live_tasks(cpu);
7776 kthread_stop(rq->migration_thread);
7777 put_task_struct(rq->migration_thread);
7778 rq->migration_thread = NULL;
7779 /* Idle task back to normal (off runqueue, low prio) */
7780 raw_spin_lock_irq(&rq->lock);
7781 update_rq_clock(rq);
7782 deactivate_task(rq, rq->idle, 0);
7783 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7784 rq->idle->sched_class = &idle_sched_class;
7785 migrate_dead_tasks(cpu);
7786 raw_spin_unlock_irq(&rq->lock);
7788 migrate_nr_uninterruptible(rq);
7789 BUG_ON(rq->nr_running != 0);
7790 calc_global_load_remove(rq);
7792 * No need to migrate the tasks: it was best-effort if
7793 * they didn't take sched_hotcpu_mutex. Just wake up
7796 raw_spin_lock_irq(&rq->lock);
7797 while (!list_empty(&rq->migration_queue)) {
7798 struct migration_req *req;
7800 req = list_entry(rq->migration_queue.next,
7801 struct migration_req, list);
7802 list_del_init(&req->list);
7803 raw_spin_unlock_irq(&rq->lock);
7804 complete(&req->done);
7805 raw_spin_lock_irq(&rq->lock);
7807 raw_spin_unlock_irq(&rq->lock);
7811 case CPU_DYING_FROZEN:
7812 /* Update our root-domain */
7814 raw_spin_lock_irqsave(&rq->lock, flags);
7816 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7819 raw_spin_unlock_irqrestore(&rq->lock, flags);
7827 * Register at high priority so that task migration (migrate_all_tasks)
7828 * happens before everything else. This has to be lower priority than
7829 * the notifier in the perf_event subsystem, though.
7831 static struct notifier_block __cpuinitdata migration_notifier = {
7832 .notifier_call = migration_call,
7836 static int __init migration_init(void)
7838 void *cpu = (void *)(long)smp_processor_id();
7841 /* Start one for the boot CPU: */
7842 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7843 BUG_ON(err == NOTIFY_BAD);
7844 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7845 register_cpu_notifier(&migration_notifier);
7849 early_initcall(migration_init);
7854 #ifdef CONFIG_SCHED_DEBUG
7856 static __read_mostly int sched_domain_debug_enabled;
7858 static int __init sched_domain_debug_setup(char *str)
7860 sched_domain_debug_enabled = 1;
7864 early_param("sched_debug", sched_domain_debug_setup);
7866 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7867 struct cpumask *groupmask)
7869 struct sched_group *group = sd->groups;
7872 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7873 cpumask_clear(groupmask);
7875 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7877 if (!(sd->flags & SD_LOAD_BALANCE)) {
7878 printk("does not load-balance\n");
7880 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7885 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7887 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7888 printk(KERN_ERR "ERROR: domain->span does not contain "
7891 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7892 printk(KERN_ERR "ERROR: domain->groups does not contain"
7896 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7900 printk(KERN_ERR "ERROR: group is NULL\n");
7904 if (!group->cpu_power) {
7905 printk(KERN_CONT "\n");
7906 printk(KERN_ERR "ERROR: domain->cpu_power not "
7911 if (!cpumask_weight(sched_group_cpus(group))) {
7912 printk(KERN_CONT "\n");
7913 printk(KERN_ERR "ERROR: empty group\n");
7917 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7918 printk(KERN_CONT "\n");
7919 printk(KERN_ERR "ERROR: repeated CPUs\n");
7923 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7925 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7927 printk(KERN_CONT " %s", str);
7928 if (group->cpu_power != SCHED_LOAD_SCALE) {
7929 printk(KERN_CONT " (cpu_power = %d)",
7933 group = group->next;
7934 } while (group != sd->groups);
7935 printk(KERN_CONT "\n");
7937 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7938 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7941 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7942 printk(KERN_ERR "ERROR: parent span is not a superset "
7943 "of domain->span\n");
7947 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7949 cpumask_var_t groupmask;
7952 if (!sched_domain_debug_enabled)
7956 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7960 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7962 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7963 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7968 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7975 free_cpumask_var(groupmask);
7977 #else /* !CONFIG_SCHED_DEBUG */
7978 # define sched_domain_debug(sd, cpu) do { } while (0)
7979 #endif /* CONFIG_SCHED_DEBUG */
7981 static int sd_degenerate(struct sched_domain *sd)
7983 if (cpumask_weight(sched_domain_span(sd)) == 1)
7986 /* Following flags need at least 2 groups */
7987 if (sd->flags & (SD_LOAD_BALANCE |
7988 SD_BALANCE_NEWIDLE |
7992 SD_SHARE_PKG_RESOURCES)) {
7993 if (sd->groups != sd->groups->next)
7997 /* Following flags don't use groups */
7998 if (sd->flags & (SD_WAKE_AFFINE))
8005 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
8007 unsigned long cflags = sd->flags, pflags = parent->flags;
8009 if (sd_degenerate(parent))
8012 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
8015 /* Flags needing groups don't count if only 1 group in parent */
8016 if (parent->groups == parent->groups->next) {
8017 pflags &= ~(SD_LOAD_BALANCE |
8018 SD_BALANCE_NEWIDLE |
8022 SD_SHARE_PKG_RESOURCES);
8023 if (nr_node_ids == 1)
8024 pflags &= ~SD_SERIALIZE;
8026 if (~cflags & pflags)
8032 static void free_rootdomain(struct root_domain *rd)
8034 synchronize_sched();
8036 cpupri_cleanup(&rd->cpupri);
8038 free_cpumask_var(rd->rto_mask);
8039 free_cpumask_var(rd->online);
8040 free_cpumask_var(rd->span);
8044 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
8046 struct root_domain *old_rd = NULL;
8047 unsigned long flags;
8049 raw_spin_lock_irqsave(&rq->lock, flags);
8054 if (cpumask_test_cpu(rq->cpu, old_rd->online))
8057 cpumask_clear_cpu(rq->cpu, old_rd->span);
8060 * If we dont want to free the old_rt yet then
8061 * set old_rd to NULL to skip the freeing later
8064 if (!atomic_dec_and_test(&old_rd->refcount))
8068 atomic_inc(&rd->refcount);
8071 cpumask_set_cpu(rq->cpu, rd->span);
8072 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
8075 raw_spin_unlock_irqrestore(&rq->lock, flags);
8078 free_rootdomain(old_rd);
8081 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8083 gfp_t gfp = GFP_KERNEL;
8085 memset(rd, 0, sizeof(*rd));
8090 if (!alloc_cpumask_var(&rd->span, gfp))
8092 if (!alloc_cpumask_var(&rd->online, gfp))
8094 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8097 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8102 free_cpumask_var(rd->rto_mask);
8104 free_cpumask_var(rd->online);
8106 free_cpumask_var(rd->span);
8111 static void init_defrootdomain(void)
8113 init_rootdomain(&def_root_domain, true);
8115 atomic_set(&def_root_domain.refcount, 1);
8118 static struct root_domain *alloc_rootdomain(void)
8120 struct root_domain *rd;
8122 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8126 if (init_rootdomain(rd, false) != 0) {
8135 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8136 * hold the hotplug lock.
8139 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8141 struct rq *rq = cpu_rq(cpu);
8142 struct sched_domain *tmp;
8144 /* Remove the sched domains which do not contribute to scheduling. */
8145 for (tmp = sd; tmp; ) {
8146 struct sched_domain *parent = tmp->parent;
8150 if (sd_parent_degenerate(tmp, parent)) {
8151 tmp->parent = parent->parent;
8153 parent->parent->child = tmp;
8158 if (sd && sd_degenerate(sd)) {
8164 sched_domain_debug(sd, cpu);
8166 rq_attach_root(rq, rd);
8167 rcu_assign_pointer(rq->sd, sd);
8170 /* cpus with isolated domains */
8171 static cpumask_var_t cpu_isolated_map;
8173 /* Setup the mask of cpus configured for isolated domains */
8174 static int __init isolated_cpu_setup(char *str)
8176 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8177 cpulist_parse(str, cpu_isolated_map);
8181 __setup("isolcpus=", isolated_cpu_setup);
8184 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8185 * to a function which identifies what group(along with sched group) a CPU
8186 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8187 * (due to the fact that we keep track of groups covered with a struct cpumask).
8189 * init_sched_build_groups will build a circular linked list of the groups
8190 * covered by the given span, and will set each group's ->cpumask correctly,
8191 * and ->cpu_power to 0.
8194 init_sched_build_groups(const struct cpumask *span,
8195 const struct cpumask *cpu_map,
8196 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8197 struct sched_group **sg,
8198 struct cpumask *tmpmask),
8199 struct cpumask *covered, struct cpumask *tmpmask)
8201 struct sched_group *first = NULL, *last = NULL;
8204 cpumask_clear(covered);
8206 for_each_cpu(i, span) {
8207 struct sched_group *sg;
8208 int group = group_fn(i, cpu_map, &sg, tmpmask);
8211 if (cpumask_test_cpu(i, covered))
8214 cpumask_clear(sched_group_cpus(sg));
8217 for_each_cpu(j, span) {
8218 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8221 cpumask_set_cpu(j, covered);
8222 cpumask_set_cpu(j, sched_group_cpus(sg));
8233 #define SD_NODES_PER_DOMAIN 16
8238 * find_next_best_node - find the next node to include in a sched_domain
8239 * @node: node whose sched_domain we're building
8240 * @used_nodes: nodes already in the sched_domain
8242 * Find the next node to include in a given scheduling domain. Simply
8243 * finds the closest node not already in the @used_nodes map.
8245 * Should use nodemask_t.
8247 static int find_next_best_node(int node, nodemask_t *used_nodes)
8249 int i, n, val, min_val, best_node = 0;
8253 for (i = 0; i < nr_node_ids; i++) {
8254 /* Start at @node */
8255 n = (node + i) % nr_node_ids;
8257 if (!nr_cpus_node(n))
8260 /* Skip already used nodes */
8261 if (node_isset(n, *used_nodes))
8264 /* Simple min distance search */
8265 val = node_distance(node, n);
8267 if (val < min_val) {
8273 node_set(best_node, *used_nodes);
8278 * sched_domain_node_span - get a cpumask for a node's sched_domain
8279 * @node: node whose cpumask we're constructing
8280 * @span: resulting cpumask
8282 * Given a node, construct a good cpumask for its sched_domain to span. It
8283 * should be one that prevents unnecessary balancing, but also spreads tasks
8286 static void sched_domain_node_span(int node, struct cpumask *span)
8288 nodemask_t used_nodes;
8291 cpumask_clear(span);
8292 nodes_clear(used_nodes);
8294 cpumask_or(span, span, cpumask_of_node(node));
8295 node_set(node, used_nodes);
8297 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8298 int next_node = find_next_best_node(node, &used_nodes);
8300 cpumask_or(span, span, cpumask_of_node(next_node));
8303 #endif /* CONFIG_NUMA */
8305 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8308 * The cpus mask in sched_group and sched_domain hangs off the end.
8310 * ( See the the comments in include/linux/sched.h:struct sched_group
8311 * and struct sched_domain. )
8313 struct static_sched_group {
8314 struct sched_group sg;
8315 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8318 struct static_sched_domain {
8319 struct sched_domain sd;
8320 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8326 cpumask_var_t domainspan;
8327 cpumask_var_t covered;
8328 cpumask_var_t notcovered;
8330 cpumask_var_t nodemask;
8331 cpumask_var_t this_sibling_map;
8332 cpumask_var_t this_core_map;
8333 cpumask_var_t send_covered;
8334 cpumask_var_t tmpmask;
8335 struct sched_group **sched_group_nodes;
8336 struct root_domain *rd;
8340 sa_sched_groups = 0,
8345 sa_this_sibling_map,
8347 sa_sched_group_nodes,
8357 * SMT sched-domains:
8359 #ifdef CONFIG_SCHED_SMT
8360 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8361 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
8364 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8365 struct sched_group **sg, struct cpumask *unused)
8368 *sg = &per_cpu(sched_groups, cpu).sg;
8371 #endif /* CONFIG_SCHED_SMT */
8374 * multi-core sched-domains:
8376 #ifdef CONFIG_SCHED_MC
8377 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8378 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8379 #endif /* CONFIG_SCHED_MC */
8381 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8383 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8384 struct sched_group **sg, struct cpumask *mask)
8388 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8389 group = cpumask_first(mask);
8391 *sg = &per_cpu(sched_group_core, group).sg;
8394 #elif defined(CONFIG_SCHED_MC)
8396 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8397 struct sched_group **sg, struct cpumask *unused)
8400 *sg = &per_cpu(sched_group_core, cpu).sg;
8405 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8406 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8409 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8410 struct sched_group **sg, struct cpumask *mask)
8413 #ifdef CONFIG_SCHED_MC
8414 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8415 group = cpumask_first(mask);
8416 #elif defined(CONFIG_SCHED_SMT)
8417 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8418 group = cpumask_first(mask);
8423 *sg = &per_cpu(sched_group_phys, group).sg;
8429 * The init_sched_build_groups can't handle what we want to do with node
8430 * groups, so roll our own. Now each node has its own list of groups which
8431 * gets dynamically allocated.
8433 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8434 static struct sched_group ***sched_group_nodes_bycpu;
8436 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8437 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8439 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8440 struct sched_group **sg,
8441 struct cpumask *nodemask)
8445 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8446 group = cpumask_first(nodemask);
8449 *sg = &per_cpu(sched_group_allnodes, group).sg;
8453 static void init_numa_sched_groups_power(struct sched_group *group_head)
8455 struct sched_group *sg = group_head;
8461 for_each_cpu(j, sched_group_cpus(sg)) {
8462 struct sched_domain *sd;
8464 sd = &per_cpu(phys_domains, j).sd;
8465 if (j != group_first_cpu(sd->groups)) {
8467 * Only add "power" once for each
8473 sg->cpu_power += sd->groups->cpu_power;
8476 } while (sg != group_head);
8479 static int build_numa_sched_groups(struct s_data *d,
8480 const struct cpumask *cpu_map, int num)
8482 struct sched_domain *sd;
8483 struct sched_group *sg, *prev;
8486 cpumask_clear(d->covered);
8487 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8488 if (cpumask_empty(d->nodemask)) {
8489 d->sched_group_nodes[num] = NULL;
8493 sched_domain_node_span(num, d->domainspan);
8494 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8496 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8499 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8503 d->sched_group_nodes[num] = sg;
8505 for_each_cpu(j, d->nodemask) {
8506 sd = &per_cpu(node_domains, j).sd;
8511 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8513 cpumask_or(d->covered, d->covered, d->nodemask);
8516 for (j = 0; j < nr_node_ids; j++) {
8517 n = (num + j) % nr_node_ids;
8518 cpumask_complement(d->notcovered, d->covered);
8519 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8520 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8521 if (cpumask_empty(d->tmpmask))
8523 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8524 if (cpumask_empty(d->tmpmask))
8526 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8530 "Can not alloc domain group for node %d\n", j);
8534 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8535 sg->next = prev->next;
8536 cpumask_or(d->covered, d->covered, d->tmpmask);
8543 #endif /* CONFIG_NUMA */
8546 /* Free memory allocated for various sched_group structures */
8547 static void free_sched_groups(const struct cpumask *cpu_map,
8548 struct cpumask *nodemask)
8552 for_each_cpu(cpu, cpu_map) {
8553 struct sched_group **sched_group_nodes
8554 = sched_group_nodes_bycpu[cpu];
8556 if (!sched_group_nodes)
8559 for (i = 0; i < nr_node_ids; i++) {
8560 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8562 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8563 if (cpumask_empty(nodemask))
8573 if (oldsg != sched_group_nodes[i])
8576 kfree(sched_group_nodes);
8577 sched_group_nodes_bycpu[cpu] = NULL;
8580 #else /* !CONFIG_NUMA */
8581 static void free_sched_groups(const struct cpumask *cpu_map,
8582 struct cpumask *nodemask)
8585 #endif /* CONFIG_NUMA */
8588 * Initialize sched groups cpu_power.
8590 * cpu_power indicates the capacity of sched group, which is used while
8591 * distributing the load between different sched groups in a sched domain.
8592 * Typically cpu_power for all the groups in a sched domain will be same unless
8593 * there are asymmetries in the topology. If there are asymmetries, group
8594 * having more cpu_power will pickup more load compared to the group having
8597 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8599 struct sched_domain *child;
8600 struct sched_group *group;
8604 WARN_ON(!sd || !sd->groups);
8606 if (cpu != group_first_cpu(sd->groups))
8611 sd->groups->cpu_power = 0;
8614 power = SCHED_LOAD_SCALE;
8615 weight = cpumask_weight(sched_domain_span(sd));
8617 * SMT siblings share the power of a single core.
8618 * Usually multiple threads get a better yield out of
8619 * that one core than a single thread would have,
8620 * reflect that in sd->smt_gain.
8622 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8623 power *= sd->smt_gain;
8625 power >>= SCHED_LOAD_SHIFT;
8627 sd->groups->cpu_power += power;
8632 * Add cpu_power of each child group to this groups cpu_power.
8634 group = child->groups;
8636 sd->groups->cpu_power += group->cpu_power;
8637 group = group->next;
8638 } while (group != child->groups);
8642 * Initializers for schedule domains
8643 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8646 #ifdef CONFIG_SCHED_DEBUG
8647 # define SD_INIT_NAME(sd, type) sd->name = #type
8649 # define SD_INIT_NAME(sd, type) do { } while (0)
8652 #define SD_INIT(sd, type) sd_init_##type(sd)
8654 #define SD_INIT_FUNC(type) \
8655 static noinline void sd_init_##type(struct sched_domain *sd) \
8657 memset(sd, 0, sizeof(*sd)); \
8658 *sd = SD_##type##_INIT; \
8659 sd->level = SD_LV_##type; \
8660 SD_INIT_NAME(sd, type); \
8665 SD_INIT_FUNC(ALLNODES)
8668 #ifdef CONFIG_SCHED_SMT
8669 SD_INIT_FUNC(SIBLING)
8671 #ifdef CONFIG_SCHED_MC
8675 static int default_relax_domain_level = -1;
8677 static int __init setup_relax_domain_level(char *str)
8681 val = simple_strtoul(str, NULL, 0);
8682 if (val < SD_LV_MAX)
8683 default_relax_domain_level = val;
8687 __setup("relax_domain_level=", setup_relax_domain_level);
8689 static void set_domain_attribute(struct sched_domain *sd,
8690 struct sched_domain_attr *attr)
8694 if (!attr || attr->relax_domain_level < 0) {
8695 if (default_relax_domain_level < 0)
8698 request = default_relax_domain_level;
8700 request = attr->relax_domain_level;
8701 if (request < sd->level) {
8702 /* turn off idle balance on this domain */
8703 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8705 /* turn on idle balance on this domain */
8706 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8710 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8711 const struct cpumask *cpu_map)
8714 case sa_sched_groups:
8715 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8716 d->sched_group_nodes = NULL;
8718 free_rootdomain(d->rd); /* fall through */
8720 free_cpumask_var(d->tmpmask); /* fall through */
8721 case sa_send_covered:
8722 free_cpumask_var(d->send_covered); /* fall through */
8723 case sa_this_core_map:
8724 free_cpumask_var(d->this_core_map); /* fall through */
8725 case sa_this_sibling_map:
8726 free_cpumask_var(d->this_sibling_map); /* fall through */
8728 free_cpumask_var(d->nodemask); /* fall through */
8729 case sa_sched_group_nodes:
8731 kfree(d->sched_group_nodes); /* fall through */
8733 free_cpumask_var(d->notcovered); /* fall through */
8735 free_cpumask_var(d->covered); /* fall through */
8737 free_cpumask_var(d->domainspan); /* fall through */
8744 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8745 const struct cpumask *cpu_map)
8748 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8750 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8751 return sa_domainspan;
8752 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8754 /* Allocate the per-node list of sched groups */
8755 d->sched_group_nodes = kcalloc(nr_node_ids,
8756 sizeof(struct sched_group *), GFP_KERNEL);
8757 if (!d->sched_group_nodes) {
8758 printk(KERN_WARNING "Can not alloc sched group node list\n");
8759 return sa_notcovered;
8761 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8763 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8764 return sa_sched_group_nodes;
8765 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8767 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8768 return sa_this_sibling_map;
8769 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8770 return sa_this_core_map;
8771 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8772 return sa_send_covered;
8773 d->rd = alloc_rootdomain();
8775 printk(KERN_WARNING "Cannot alloc root domain\n");
8778 return sa_rootdomain;
8781 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8782 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8784 struct sched_domain *sd = NULL;
8786 struct sched_domain *parent;
8789 if (cpumask_weight(cpu_map) >
8790 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8791 sd = &per_cpu(allnodes_domains, i).sd;
8792 SD_INIT(sd, ALLNODES);
8793 set_domain_attribute(sd, attr);
8794 cpumask_copy(sched_domain_span(sd), cpu_map);
8795 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8800 sd = &per_cpu(node_domains, i).sd;
8802 set_domain_attribute(sd, attr);
8803 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8804 sd->parent = parent;
8807 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8812 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8813 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8814 struct sched_domain *parent, int i)
8816 struct sched_domain *sd;
8817 sd = &per_cpu(phys_domains, i).sd;
8819 set_domain_attribute(sd, attr);
8820 cpumask_copy(sched_domain_span(sd), d->nodemask);
8821 sd->parent = parent;
8824 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8828 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8829 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8830 struct sched_domain *parent, int i)
8832 struct sched_domain *sd = parent;
8833 #ifdef CONFIG_SCHED_MC
8834 sd = &per_cpu(core_domains, i).sd;
8836 set_domain_attribute(sd, attr);
8837 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8838 sd->parent = parent;
8840 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8845 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8846 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8847 struct sched_domain *parent, int i)
8849 struct sched_domain *sd = parent;
8850 #ifdef CONFIG_SCHED_SMT
8851 sd = &per_cpu(cpu_domains, i).sd;
8852 SD_INIT(sd, SIBLING);
8853 set_domain_attribute(sd, attr);
8854 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8855 sd->parent = parent;
8857 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8862 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8863 const struct cpumask *cpu_map, int cpu)
8866 #ifdef CONFIG_SCHED_SMT
8867 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8868 cpumask_and(d->this_sibling_map, cpu_map,
8869 topology_thread_cpumask(cpu));
8870 if (cpu == cpumask_first(d->this_sibling_map))
8871 init_sched_build_groups(d->this_sibling_map, cpu_map,
8873 d->send_covered, d->tmpmask);
8876 #ifdef CONFIG_SCHED_MC
8877 case SD_LV_MC: /* set up multi-core groups */
8878 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8879 if (cpu == cpumask_first(d->this_core_map))
8880 init_sched_build_groups(d->this_core_map, cpu_map,
8882 d->send_covered, d->tmpmask);
8885 case SD_LV_CPU: /* set up physical groups */
8886 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8887 if (!cpumask_empty(d->nodemask))
8888 init_sched_build_groups(d->nodemask, cpu_map,
8890 d->send_covered, d->tmpmask);
8893 case SD_LV_ALLNODES:
8894 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8895 d->send_covered, d->tmpmask);
8904 * Build sched domains for a given set of cpus and attach the sched domains
8905 * to the individual cpus
8907 static int __build_sched_domains(const struct cpumask *cpu_map,
8908 struct sched_domain_attr *attr)
8910 enum s_alloc alloc_state = sa_none;
8912 struct sched_domain *sd;
8918 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8919 if (alloc_state != sa_rootdomain)
8921 alloc_state = sa_sched_groups;
8924 * Set up domains for cpus specified by the cpu_map.
8926 for_each_cpu(i, cpu_map) {
8927 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8930 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8931 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8932 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8933 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8936 for_each_cpu(i, cpu_map) {
8937 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8938 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8941 /* Set up physical groups */
8942 for (i = 0; i < nr_node_ids; i++)
8943 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8946 /* Set up node groups */
8948 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8950 for (i = 0; i < nr_node_ids; i++)
8951 if (build_numa_sched_groups(&d, cpu_map, i))
8955 /* Calculate CPU power for physical packages and nodes */
8956 #ifdef CONFIG_SCHED_SMT
8957 for_each_cpu(i, cpu_map) {
8958 sd = &per_cpu(cpu_domains, i).sd;
8959 init_sched_groups_power(i, sd);
8962 #ifdef CONFIG_SCHED_MC
8963 for_each_cpu(i, cpu_map) {
8964 sd = &per_cpu(core_domains, i).sd;
8965 init_sched_groups_power(i, sd);
8969 for_each_cpu(i, cpu_map) {
8970 sd = &per_cpu(phys_domains, i).sd;
8971 init_sched_groups_power(i, sd);
8975 for (i = 0; i < nr_node_ids; i++)
8976 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8978 if (d.sd_allnodes) {
8979 struct sched_group *sg;
8981 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8983 init_numa_sched_groups_power(sg);
8987 /* Attach the domains */
8988 for_each_cpu(i, cpu_map) {
8989 #ifdef CONFIG_SCHED_SMT
8990 sd = &per_cpu(cpu_domains, i).sd;
8991 #elif defined(CONFIG_SCHED_MC)
8992 sd = &per_cpu(core_domains, i).sd;
8994 sd = &per_cpu(phys_domains, i).sd;
8996 cpu_attach_domain(sd, d.rd, i);
8999 d.sched_group_nodes = NULL; /* don't free this we still need it */
9000 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
9004 __free_domain_allocs(&d, alloc_state, cpu_map);
9008 static int build_sched_domains(const struct cpumask *cpu_map)
9010 return __build_sched_domains(cpu_map, NULL);
9013 static cpumask_var_t *doms_cur; /* current sched domains */
9014 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
9015 static struct sched_domain_attr *dattr_cur;
9016 /* attribues of custom domains in 'doms_cur' */
9019 * Special case: If a kmalloc of a doms_cur partition (array of
9020 * cpumask) fails, then fallback to a single sched domain,
9021 * as determined by the single cpumask fallback_doms.
9023 static cpumask_var_t fallback_doms;
9026 * arch_update_cpu_topology lets virtualized architectures update the
9027 * cpu core maps. It is supposed to return 1 if the topology changed
9028 * or 0 if it stayed the same.
9030 int __attribute__((weak)) arch_update_cpu_topology(void)
9035 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
9038 cpumask_var_t *doms;
9040 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
9043 for (i = 0; i < ndoms; i++) {
9044 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
9045 free_sched_domains(doms, i);
9052 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
9055 for (i = 0; i < ndoms; i++)
9056 free_cpumask_var(doms[i]);
9061 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9062 * For now this just excludes isolated cpus, but could be used to
9063 * exclude other special cases in the future.
9065 static int arch_init_sched_domains(const struct cpumask *cpu_map)
9069 arch_update_cpu_topology();
9071 doms_cur = alloc_sched_domains(ndoms_cur);
9073 doms_cur = &fallback_doms;
9074 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
9076 err = build_sched_domains(doms_cur[0]);
9077 register_sched_domain_sysctl();
9082 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9083 struct cpumask *tmpmask)
9085 free_sched_groups(cpu_map, tmpmask);
9089 * Detach sched domains from a group of cpus specified in cpu_map
9090 * These cpus will now be attached to the NULL domain
9092 static void detach_destroy_domains(const struct cpumask *cpu_map)
9094 /* Save because hotplug lock held. */
9095 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9098 for_each_cpu(i, cpu_map)
9099 cpu_attach_domain(NULL, &def_root_domain, i);
9100 synchronize_sched();
9101 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9104 /* handle null as "default" */
9105 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9106 struct sched_domain_attr *new, int idx_new)
9108 struct sched_domain_attr tmp;
9115 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9116 new ? (new + idx_new) : &tmp,
9117 sizeof(struct sched_domain_attr));
9121 * Partition sched domains as specified by the 'ndoms_new'
9122 * cpumasks in the array doms_new[] of cpumasks. This compares
9123 * doms_new[] to the current sched domain partitioning, doms_cur[].
9124 * It destroys each deleted domain and builds each new domain.
9126 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
9127 * The masks don't intersect (don't overlap.) We should setup one
9128 * sched domain for each mask. CPUs not in any of the cpumasks will
9129 * not be load balanced. If the same cpumask appears both in the
9130 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9133 * The passed in 'doms_new' should be allocated using
9134 * alloc_sched_domains. This routine takes ownership of it and will
9135 * free_sched_domains it when done with it. If the caller failed the
9136 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9137 * and partition_sched_domains() will fallback to the single partition
9138 * 'fallback_doms', it also forces the domains to be rebuilt.
9140 * If doms_new == NULL it will be replaced with cpu_online_mask.
9141 * ndoms_new == 0 is a special case for destroying existing domains,
9142 * and it will not create the default domain.
9144 * Call with hotplug lock held
9146 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
9147 struct sched_domain_attr *dattr_new)
9152 mutex_lock(&sched_domains_mutex);
9154 /* always unregister in case we don't destroy any domains */
9155 unregister_sched_domain_sysctl();
9157 /* Let architecture update cpu core mappings. */
9158 new_topology = arch_update_cpu_topology();
9160 n = doms_new ? ndoms_new : 0;
9162 /* Destroy deleted domains */
9163 for (i = 0; i < ndoms_cur; i++) {
9164 for (j = 0; j < n && !new_topology; j++) {
9165 if (cpumask_equal(doms_cur[i], doms_new[j])
9166 && dattrs_equal(dattr_cur, i, dattr_new, j))
9169 /* no match - a current sched domain not in new doms_new[] */
9170 detach_destroy_domains(doms_cur[i]);
9175 if (doms_new == NULL) {
9177 doms_new = &fallback_doms;
9178 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
9179 WARN_ON_ONCE(dattr_new);
9182 /* Build new domains */
9183 for (i = 0; i < ndoms_new; i++) {
9184 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9185 if (cpumask_equal(doms_new[i], doms_cur[j])
9186 && dattrs_equal(dattr_new, i, dattr_cur, j))
9189 /* no match - add a new doms_new */
9190 __build_sched_domains(doms_new[i],
9191 dattr_new ? dattr_new + i : NULL);
9196 /* Remember the new sched domains */
9197 if (doms_cur != &fallback_doms)
9198 free_sched_domains(doms_cur, ndoms_cur);
9199 kfree(dattr_cur); /* kfree(NULL) is safe */
9200 doms_cur = doms_new;
9201 dattr_cur = dattr_new;
9202 ndoms_cur = ndoms_new;
9204 register_sched_domain_sysctl();
9206 mutex_unlock(&sched_domains_mutex);
9209 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9210 static void arch_reinit_sched_domains(void)
9214 /* Destroy domains first to force the rebuild */
9215 partition_sched_domains(0, NULL, NULL);
9217 rebuild_sched_domains();
9221 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9223 unsigned int level = 0;
9225 if (sscanf(buf, "%u", &level) != 1)
9229 * level is always be positive so don't check for
9230 * level < POWERSAVINGS_BALANCE_NONE which is 0
9231 * What happens on 0 or 1 byte write,
9232 * need to check for count as well?
9235 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9239 sched_smt_power_savings = level;
9241 sched_mc_power_savings = level;
9243 arch_reinit_sched_domains();
9248 #ifdef CONFIG_SCHED_MC
9249 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9252 return sprintf(page, "%u\n", sched_mc_power_savings);
9254 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9255 const char *buf, size_t count)
9257 return sched_power_savings_store(buf, count, 0);
9259 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9260 sched_mc_power_savings_show,
9261 sched_mc_power_savings_store);
9264 #ifdef CONFIG_SCHED_SMT
9265 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9268 return sprintf(page, "%u\n", sched_smt_power_savings);
9270 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9271 const char *buf, size_t count)
9273 return sched_power_savings_store(buf, count, 1);
9275 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9276 sched_smt_power_savings_show,
9277 sched_smt_power_savings_store);
9280 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9284 #ifdef CONFIG_SCHED_SMT
9286 err = sysfs_create_file(&cls->kset.kobj,
9287 &attr_sched_smt_power_savings.attr);
9289 #ifdef CONFIG_SCHED_MC
9290 if (!err && mc_capable())
9291 err = sysfs_create_file(&cls->kset.kobj,
9292 &attr_sched_mc_power_savings.attr);
9296 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9298 #ifndef CONFIG_CPUSETS
9300 * Add online and remove offline CPUs from the scheduler domains.
9301 * When cpusets are enabled they take over this function.
9303 static int update_sched_domains(struct notifier_block *nfb,
9304 unsigned long action, void *hcpu)
9308 case CPU_ONLINE_FROZEN:
9309 case CPU_DOWN_PREPARE:
9310 case CPU_DOWN_PREPARE_FROZEN:
9311 case CPU_DOWN_FAILED:
9312 case CPU_DOWN_FAILED_FROZEN:
9313 partition_sched_domains(1, NULL, NULL);
9322 static int update_runtime(struct notifier_block *nfb,
9323 unsigned long action, void *hcpu)
9325 int cpu = (int)(long)hcpu;
9328 case CPU_DOWN_PREPARE:
9329 case CPU_DOWN_PREPARE_FROZEN:
9330 disable_runtime(cpu_rq(cpu));
9333 case CPU_DOWN_FAILED:
9334 case CPU_DOWN_FAILED_FROZEN:
9336 case CPU_ONLINE_FROZEN:
9337 enable_runtime(cpu_rq(cpu));
9345 void __init sched_init_smp(void)
9347 cpumask_var_t non_isolated_cpus;
9349 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9350 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9352 #if defined(CONFIG_NUMA)
9353 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9355 BUG_ON(sched_group_nodes_bycpu == NULL);
9358 mutex_lock(&sched_domains_mutex);
9359 arch_init_sched_domains(cpu_active_mask);
9360 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9361 if (cpumask_empty(non_isolated_cpus))
9362 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9363 mutex_unlock(&sched_domains_mutex);
9366 #ifndef CONFIG_CPUSETS
9367 /* XXX: Theoretical race here - CPU may be hotplugged now */
9368 hotcpu_notifier(update_sched_domains, 0);
9371 /* RT runtime code needs to handle some hotplug events */
9372 hotcpu_notifier(update_runtime, 0);
9376 /* Move init over to a non-isolated CPU */
9377 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9379 sched_init_granularity();
9380 free_cpumask_var(non_isolated_cpus);
9382 init_sched_rt_class();
9385 void __init sched_init_smp(void)
9387 sched_init_granularity();
9389 #endif /* CONFIG_SMP */
9391 const_debug unsigned int sysctl_timer_migration = 1;
9393 int in_sched_functions(unsigned long addr)
9395 return in_lock_functions(addr) ||
9396 (addr >= (unsigned long)__sched_text_start
9397 && addr < (unsigned long)__sched_text_end);
9400 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9402 cfs_rq->tasks_timeline = RB_ROOT;
9403 INIT_LIST_HEAD(&cfs_rq->tasks);
9404 #ifdef CONFIG_FAIR_GROUP_SCHED
9407 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9410 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9412 struct rt_prio_array *array;
9415 array = &rt_rq->active;
9416 for (i = 0; i < MAX_RT_PRIO; i++) {
9417 INIT_LIST_HEAD(array->queue + i);
9418 __clear_bit(i, array->bitmap);
9420 /* delimiter for bitsearch: */
9421 __set_bit(MAX_RT_PRIO, array->bitmap);
9423 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9424 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9426 rt_rq->highest_prio.next = MAX_RT_PRIO;
9430 rt_rq->rt_nr_migratory = 0;
9431 rt_rq->overloaded = 0;
9432 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
9436 rt_rq->rt_throttled = 0;
9437 rt_rq->rt_runtime = 0;
9438 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
9440 #ifdef CONFIG_RT_GROUP_SCHED
9441 rt_rq->rt_nr_boosted = 0;
9446 #ifdef CONFIG_FAIR_GROUP_SCHED
9447 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9448 struct sched_entity *se, int cpu, int add,
9449 struct sched_entity *parent)
9451 struct rq *rq = cpu_rq(cpu);
9452 tg->cfs_rq[cpu] = cfs_rq;
9453 init_cfs_rq(cfs_rq, rq);
9456 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9459 /* se could be NULL for init_task_group */
9464 se->cfs_rq = &rq->cfs;
9466 se->cfs_rq = parent->my_q;
9469 se->load.weight = tg->shares;
9470 se->load.inv_weight = 0;
9471 se->parent = parent;
9475 #ifdef CONFIG_RT_GROUP_SCHED
9476 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9477 struct sched_rt_entity *rt_se, int cpu, int add,
9478 struct sched_rt_entity *parent)
9480 struct rq *rq = cpu_rq(cpu);
9482 tg->rt_rq[cpu] = rt_rq;
9483 init_rt_rq(rt_rq, rq);
9485 rt_rq->rt_se = rt_se;
9486 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9488 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9490 tg->rt_se[cpu] = rt_se;
9495 rt_se->rt_rq = &rq->rt;
9497 rt_se->rt_rq = parent->my_q;
9499 rt_se->my_q = rt_rq;
9500 rt_se->parent = parent;
9501 INIT_LIST_HEAD(&rt_se->run_list);
9505 void __init sched_init(void)
9508 unsigned long alloc_size = 0, ptr;
9510 #ifdef CONFIG_FAIR_GROUP_SCHED
9511 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9513 #ifdef CONFIG_RT_GROUP_SCHED
9514 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9516 #ifdef CONFIG_USER_SCHED
9519 #ifdef CONFIG_CPUMASK_OFFSTACK
9520 alloc_size += num_possible_cpus() * cpumask_size();
9523 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9525 #ifdef CONFIG_FAIR_GROUP_SCHED
9526 init_task_group.se = (struct sched_entity **)ptr;
9527 ptr += nr_cpu_ids * sizeof(void **);
9529 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9530 ptr += nr_cpu_ids * sizeof(void **);
9532 #ifdef CONFIG_USER_SCHED
9533 root_task_group.se = (struct sched_entity **)ptr;
9534 ptr += nr_cpu_ids * sizeof(void **);
9536 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9537 ptr += nr_cpu_ids * sizeof(void **);
9538 #endif /* CONFIG_USER_SCHED */
9539 #endif /* CONFIG_FAIR_GROUP_SCHED */
9540 #ifdef CONFIG_RT_GROUP_SCHED
9541 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9542 ptr += nr_cpu_ids * sizeof(void **);
9544 init_task_group.rt_rq = (struct rt_rq **)ptr;
9545 ptr += nr_cpu_ids * sizeof(void **);
9547 #ifdef CONFIG_USER_SCHED
9548 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9549 ptr += nr_cpu_ids * sizeof(void **);
9551 root_task_group.rt_rq = (struct rt_rq **)ptr;
9552 ptr += nr_cpu_ids * sizeof(void **);
9553 #endif /* CONFIG_USER_SCHED */
9554 #endif /* CONFIG_RT_GROUP_SCHED */
9555 #ifdef CONFIG_CPUMASK_OFFSTACK
9556 for_each_possible_cpu(i) {
9557 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9558 ptr += cpumask_size();
9560 #endif /* CONFIG_CPUMASK_OFFSTACK */
9564 init_defrootdomain();
9567 init_rt_bandwidth(&def_rt_bandwidth,
9568 global_rt_period(), global_rt_runtime());
9570 #ifdef CONFIG_RT_GROUP_SCHED
9571 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9572 global_rt_period(), global_rt_runtime());
9573 #ifdef CONFIG_USER_SCHED
9574 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9575 global_rt_period(), RUNTIME_INF);
9576 #endif /* CONFIG_USER_SCHED */
9577 #endif /* CONFIG_RT_GROUP_SCHED */
9579 #ifdef CONFIG_GROUP_SCHED
9580 list_add(&init_task_group.list, &task_groups);
9581 INIT_LIST_HEAD(&init_task_group.children);
9583 #ifdef CONFIG_USER_SCHED
9584 INIT_LIST_HEAD(&root_task_group.children);
9585 init_task_group.parent = &root_task_group;
9586 list_add(&init_task_group.siblings, &root_task_group.children);
9587 #endif /* CONFIG_USER_SCHED */
9588 #endif /* CONFIG_GROUP_SCHED */
9590 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9591 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9592 __alignof__(unsigned long));
9594 for_each_possible_cpu(i) {
9598 raw_spin_lock_init(&rq->lock);
9600 rq->calc_load_active = 0;
9601 rq->calc_load_update = jiffies + LOAD_FREQ;
9602 init_cfs_rq(&rq->cfs, rq);
9603 init_rt_rq(&rq->rt, rq);
9604 #ifdef CONFIG_FAIR_GROUP_SCHED
9605 init_task_group.shares = init_task_group_load;
9606 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9607 #ifdef CONFIG_CGROUP_SCHED
9609 * How much cpu bandwidth does init_task_group get?
9611 * In case of task-groups formed thr' the cgroup filesystem, it
9612 * gets 100% of the cpu resources in the system. This overall
9613 * system cpu resource is divided among the tasks of
9614 * init_task_group and its child task-groups in a fair manner,
9615 * based on each entity's (task or task-group's) weight
9616 * (se->load.weight).
9618 * In other words, if init_task_group has 10 tasks of weight
9619 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9620 * then A0's share of the cpu resource is:
9622 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9624 * We achieve this by letting init_task_group's tasks sit
9625 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9627 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9628 #elif defined CONFIG_USER_SCHED
9629 root_task_group.shares = NICE_0_LOAD;
9630 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9632 * In case of task-groups formed thr' the user id of tasks,
9633 * init_task_group represents tasks belonging to root user.
9634 * Hence it forms a sibling of all subsequent groups formed.
9635 * In this case, init_task_group gets only a fraction of overall
9636 * system cpu resource, based on the weight assigned to root
9637 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9638 * by letting tasks of init_task_group sit in a separate cfs_rq
9639 * (init_tg_cfs_rq) and having one entity represent this group of
9640 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9642 init_tg_cfs_entry(&init_task_group,
9643 &per_cpu(init_tg_cfs_rq, i),
9644 &per_cpu(init_sched_entity, i), i, 1,
9645 root_task_group.se[i]);
9648 #endif /* CONFIG_FAIR_GROUP_SCHED */
9650 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9651 #ifdef CONFIG_RT_GROUP_SCHED
9652 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9653 #ifdef CONFIG_CGROUP_SCHED
9654 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9655 #elif defined CONFIG_USER_SCHED
9656 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9657 init_tg_rt_entry(&init_task_group,
9658 &per_cpu(init_rt_rq_var, i),
9659 &per_cpu(init_sched_rt_entity, i), i, 1,
9660 root_task_group.rt_se[i]);
9664 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9665 rq->cpu_load[j] = 0;
9669 rq->post_schedule = 0;
9670 rq->active_balance = 0;
9671 rq->next_balance = jiffies;
9675 rq->migration_thread = NULL;
9677 rq->avg_idle = 2*sysctl_sched_migration_cost;
9678 INIT_LIST_HEAD(&rq->migration_queue);
9679 rq_attach_root(rq, &def_root_domain);
9682 atomic_set(&rq->nr_iowait, 0);
9685 set_load_weight(&init_task);
9687 #ifdef CONFIG_PREEMPT_NOTIFIERS
9688 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9692 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9695 #ifdef CONFIG_RT_MUTEXES
9696 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
9700 * The boot idle thread does lazy MMU switching as well:
9702 atomic_inc(&init_mm.mm_count);
9703 enter_lazy_tlb(&init_mm, current);
9706 * Make us the idle thread. Technically, schedule() should not be
9707 * called from this thread, however somewhere below it might be,
9708 * but because we are the idle thread, we just pick up running again
9709 * when this runqueue becomes "idle".
9711 init_idle(current, smp_processor_id());
9713 calc_load_update = jiffies + LOAD_FREQ;
9716 * During early bootup we pretend to be a normal task:
9718 current->sched_class = &fair_sched_class;
9720 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9721 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9724 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9725 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9727 /* May be allocated at isolcpus cmdline parse time */
9728 if (cpu_isolated_map == NULL)
9729 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9734 scheduler_running = 1;
9737 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9738 static inline int preempt_count_equals(int preempt_offset)
9740 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
9742 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9745 void __might_sleep(char *file, int line, int preempt_offset)
9748 static unsigned long prev_jiffy; /* ratelimiting */
9750 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9751 system_state != SYSTEM_RUNNING || oops_in_progress)
9753 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9755 prev_jiffy = jiffies;
9758 "BUG: sleeping function called from invalid context at %s:%d\n",
9761 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9762 in_atomic(), irqs_disabled(),
9763 current->pid, current->comm);
9765 debug_show_held_locks(current);
9766 if (irqs_disabled())
9767 print_irqtrace_events(current);
9771 EXPORT_SYMBOL(__might_sleep);
9774 #ifdef CONFIG_MAGIC_SYSRQ
9775 static void normalize_task(struct rq *rq, struct task_struct *p)
9779 update_rq_clock(rq);
9780 on_rq = p->se.on_rq;
9782 deactivate_task(rq, p, 0);
9783 __setscheduler(rq, p, SCHED_NORMAL, 0);
9785 activate_task(rq, p, 0);
9786 resched_task(rq->curr);
9790 void normalize_rt_tasks(void)
9792 struct task_struct *g, *p;
9793 unsigned long flags;
9796 read_lock_irqsave(&tasklist_lock, flags);
9797 do_each_thread(g, p) {
9799 * Only normalize user tasks:
9804 p->se.exec_start = 0;
9805 #ifdef CONFIG_SCHEDSTATS
9806 p->se.wait_start = 0;
9807 p->se.sleep_start = 0;
9808 p->se.block_start = 0;
9813 * Renice negative nice level userspace
9816 if (TASK_NICE(p) < 0 && p->mm)
9817 set_user_nice(p, 0);
9821 raw_spin_lock(&p->pi_lock);
9822 rq = __task_rq_lock(p);
9824 normalize_task(rq, p);
9826 __task_rq_unlock(rq);
9827 raw_spin_unlock(&p->pi_lock);
9828 } while_each_thread(g, p);
9830 read_unlock_irqrestore(&tasklist_lock, flags);
9833 #endif /* CONFIG_MAGIC_SYSRQ */
9837 * These functions are only useful for the IA64 MCA handling.
9839 * They can only be called when the whole system has been
9840 * stopped - every CPU needs to be quiescent, and no scheduling
9841 * activity can take place. Using them for anything else would
9842 * be a serious bug, and as a result, they aren't even visible
9843 * under any other configuration.
9847 * curr_task - return the current task for a given cpu.
9848 * @cpu: the processor in question.
9850 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9852 struct task_struct *curr_task(int cpu)
9854 return cpu_curr(cpu);
9858 * set_curr_task - set the current task for a given cpu.
9859 * @cpu: the processor in question.
9860 * @p: the task pointer to set.
9862 * Description: This function must only be used when non-maskable interrupts
9863 * are serviced on a separate stack. It allows the architecture to switch the
9864 * notion of the current task on a cpu in a non-blocking manner. This function
9865 * must be called with all CPU's synchronized, and interrupts disabled, the
9866 * and caller must save the original value of the current task (see
9867 * curr_task() above) and restore that value before reenabling interrupts and
9868 * re-starting the system.
9870 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9872 void set_curr_task(int cpu, struct task_struct *p)
9879 #ifdef CONFIG_FAIR_GROUP_SCHED
9880 static void free_fair_sched_group(struct task_group *tg)
9884 for_each_possible_cpu(i) {
9886 kfree(tg->cfs_rq[i]);
9896 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9898 struct cfs_rq *cfs_rq;
9899 struct sched_entity *se;
9903 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9906 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9910 tg->shares = NICE_0_LOAD;
9912 for_each_possible_cpu(i) {
9915 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9916 GFP_KERNEL, cpu_to_node(i));
9920 se = kzalloc_node(sizeof(struct sched_entity),
9921 GFP_KERNEL, cpu_to_node(i));
9925 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9936 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9938 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9939 &cpu_rq(cpu)->leaf_cfs_rq_list);
9942 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9944 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9946 #else /* !CONFG_FAIR_GROUP_SCHED */
9947 static inline void free_fair_sched_group(struct task_group *tg)
9952 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9957 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9961 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9964 #endif /* CONFIG_FAIR_GROUP_SCHED */
9966 #ifdef CONFIG_RT_GROUP_SCHED
9967 static void free_rt_sched_group(struct task_group *tg)
9971 destroy_rt_bandwidth(&tg->rt_bandwidth);
9973 for_each_possible_cpu(i) {
9975 kfree(tg->rt_rq[i]);
9977 kfree(tg->rt_se[i]);
9985 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9987 struct rt_rq *rt_rq;
9988 struct sched_rt_entity *rt_se;
9992 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9995 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9999 init_rt_bandwidth(&tg->rt_bandwidth,
10000 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
10002 for_each_possible_cpu(i) {
10005 rt_rq = kzalloc_node(sizeof(struct rt_rq),
10006 GFP_KERNEL, cpu_to_node(i));
10010 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
10011 GFP_KERNEL, cpu_to_node(i));
10015 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
10026 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10028 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
10029 &cpu_rq(cpu)->leaf_rt_rq_list);
10032 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10034 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
10036 #else /* !CONFIG_RT_GROUP_SCHED */
10037 static inline void free_rt_sched_group(struct task_group *tg)
10042 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
10047 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10051 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10054 #endif /* CONFIG_RT_GROUP_SCHED */
10056 #ifdef CONFIG_GROUP_SCHED
10057 static void free_sched_group(struct task_group *tg)
10059 free_fair_sched_group(tg);
10060 free_rt_sched_group(tg);
10064 /* allocate runqueue etc for a new task group */
10065 struct task_group *sched_create_group(struct task_group *parent)
10067 struct task_group *tg;
10068 unsigned long flags;
10071 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
10073 return ERR_PTR(-ENOMEM);
10075 if (!alloc_fair_sched_group(tg, parent))
10078 if (!alloc_rt_sched_group(tg, parent))
10081 spin_lock_irqsave(&task_group_lock, flags);
10082 for_each_possible_cpu(i) {
10083 register_fair_sched_group(tg, i);
10084 register_rt_sched_group(tg, i);
10086 list_add_rcu(&tg->list, &task_groups);
10088 WARN_ON(!parent); /* root should already exist */
10090 tg->parent = parent;
10091 INIT_LIST_HEAD(&tg->children);
10092 list_add_rcu(&tg->siblings, &parent->children);
10093 spin_unlock_irqrestore(&task_group_lock, flags);
10098 free_sched_group(tg);
10099 return ERR_PTR(-ENOMEM);
10102 /* rcu callback to free various structures associated with a task group */
10103 static void free_sched_group_rcu(struct rcu_head *rhp)
10105 /* now it should be safe to free those cfs_rqs */
10106 free_sched_group(container_of(rhp, struct task_group, rcu));
10109 /* Destroy runqueue etc associated with a task group */
10110 void sched_destroy_group(struct task_group *tg)
10112 unsigned long flags;
10115 spin_lock_irqsave(&task_group_lock, flags);
10116 for_each_possible_cpu(i) {
10117 unregister_fair_sched_group(tg, i);
10118 unregister_rt_sched_group(tg, i);
10120 list_del_rcu(&tg->list);
10121 list_del_rcu(&tg->siblings);
10122 spin_unlock_irqrestore(&task_group_lock, flags);
10124 /* wait for possible concurrent references to cfs_rqs complete */
10125 call_rcu(&tg->rcu, free_sched_group_rcu);
10128 /* change task's runqueue when it moves between groups.
10129 * The caller of this function should have put the task in its new group
10130 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10131 * reflect its new group.
10133 void sched_move_task(struct task_struct *tsk)
10135 int on_rq, running;
10136 unsigned long flags;
10139 rq = task_rq_lock(tsk, &flags);
10141 update_rq_clock(rq);
10143 running = task_current(rq, tsk);
10144 on_rq = tsk->se.on_rq;
10147 dequeue_task(rq, tsk, 0);
10148 if (unlikely(running))
10149 tsk->sched_class->put_prev_task(rq, tsk);
10151 set_task_rq(tsk, task_cpu(tsk));
10153 #ifdef CONFIG_FAIR_GROUP_SCHED
10154 if (tsk->sched_class->moved_group)
10155 tsk->sched_class->moved_group(tsk, on_rq);
10158 if (unlikely(running))
10159 tsk->sched_class->set_curr_task(rq);
10161 enqueue_task(rq, tsk, 0);
10163 task_rq_unlock(rq, &flags);
10165 #endif /* CONFIG_GROUP_SCHED */
10167 #ifdef CONFIG_FAIR_GROUP_SCHED
10168 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10170 struct cfs_rq *cfs_rq = se->cfs_rq;
10175 dequeue_entity(cfs_rq, se, 0);
10177 se->load.weight = shares;
10178 se->load.inv_weight = 0;
10181 enqueue_entity(cfs_rq, se, 0);
10184 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10186 struct cfs_rq *cfs_rq = se->cfs_rq;
10187 struct rq *rq = cfs_rq->rq;
10188 unsigned long flags;
10190 raw_spin_lock_irqsave(&rq->lock, flags);
10191 __set_se_shares(se, shares);
10192 raw_spin_unlock_irqrestore(&rq->lock, flags);
10195 static DEFINE_MUTEX(shares_mutex);
10197 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10200 unsigned long flags;
10203 * We can't change the weight of the root cgroup.
10208 if (shares < MIN_SHARES)
10209 shares = MIN_SHARES;
10210 else if (shares > MAX_SHARES)
10211 shares = MAX_SHARES;
10213 mutex_lock(&shares_mutex);
10214 if (tg->shares == shares)
10217 spin_lock_irqsave(&task_group_lock, flags);
10218 for_each_possible_cpu(i)
10219 unregister_fair_sched_group(tg, i);
10220 list_del_rcu(&tg->siblings);
10221 spin_unlock_irqrestore(&task_group_lock, flags);
10223 /* wait for any ongoing reference to this group to finish */
10224 synchronize_sched();
10227 * Now we are free to modify the group's share on each cpu
10228 * w/o tripping rebalance_share or load_balance_fair.
10230 tg->shares = shares;
10231 for_each_possible_cpu(i) {
10233 * force a rebalance
10235 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10236 set_se_shares(tg->se[i], shares);
10240 * Enable load balance activity on this group, by inserting it back on
10241 * each cpu's rq->leaf_cfs_rq_list.
10243 spin_lock_irqsave(&task_group_lock, flags);
10244 for_each_possible_cpu(i)
10245 register_fair_sched_group(tg, i);
10246 list_add_rcu(&tg->siblings, &tg->parent->children);
10247 spin_unlock_irqrestore(&task_group_lock, flags);
10249 mutex_unlock(&shares_mutex);
10253 unsigned long sched_group_shares(struct task_group *tg)
10259 #ifdef CONFIG_RT_GROUP_SCHED
10261 * Ensure that the real time constraints are schedulable.
10263 static DEFINE_MUTEX(rt_constraints_mutex);
10265 static unsigned long to_ratio(u64 period, u64 runtime)
10267 if (runtime == RUNTIME_INF)
10270 return div64_u64(runtime << 20, period);
10273 /* Must be called with tasklist_lock held */
10274 static inline int tg_has_rt_tasks(struct task_group *tg)
10276 struct task_struct *g, *p;
10278 do_each_thread(g, p) {
10279 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10281 } while_each_thread(g, p);
10286 struct rt_schedulable_data {
10287 struct task_group *tg;
10292 static int tg_schedulable(struct task_group *tg, void *data)
10294 struct rt_schedulable_data *d = data;
10295 struct task_group *child;
10296 unsigned long total, sum = 0;
10297 u64 period, runtime;
10299 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10300 runtime = tg->rt_bandwidth.rt_runtime;
10303 period = d->rt_period;
10304 runtime = d->rt_runtime;
10307 #ifdef CONFIG_USER_SCHED
10308 if (tg == &root_task_group) {
10309 period = global_rt_period();
10310 runtime = global_rt_runtime();
10315 * Cannot have more runtime than the period.
10317 if (runtime > period && runtime != RUNTIME_INF)
10321 * Ensure we don't starve existing RT tasks.
10323 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10326 total = to_ratio(period, runtime);
10329 * Nobody can have more than the global setting allows.
10331 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10335 * The sum of our children's runtime should not exceed our own.
10337 list_for_each_entry_rcu(child, &tg->children, siblings) {
10338 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10339 runtime = child->rt_bandwidth.rt_runtime;
10341 if (child == d->tg) {
10342 period = d->rt_period;
10343 runtime = d->rt_runtime;
10346 sum += to_ratio(period, runtime);
10355 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10357 struct rt_schedulable_data data = {
10359 .rt_period = period,
10360 .rt_runtime = runtime,
10363 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10366 static int tg_set_bandwidth(struct task_group *tg,
10367 u64 rt_period, u64 rt_runtime)
10371 mutex_lock(&rt_constraints_mutex);
10372 read_lock(&tasklist_lock);
10373 err = __rt_schedulable(tg, rt_period, rt_runtime);
10377 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10378 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10379 tg->rt_bandwidth.rt_runtime = rt_runtime;
10381 for_each_possible_cpu(i) {
10382 struct rt_rq *rt_rq = tg->rt_rq[i];
10384 raw_spin_lock(&rt_rq->rt_runtime_lock);
10385 rt_rq->rt_runtime = rt_runtime;
10386 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10388 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10390 read_unlock(&tasklist_lock);
10391 mutex_unlock(&rt_constraints_mutex);
10396 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10398 u64 rt_runtime, rt_period;
10400 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10401 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10402 if (rt_runtime_us < 0)
10403 rt_runtime = RUNTIME_INF;
10405 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10408 long sched_group_rt_runtime(struct task_group *tg)
10412 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10415 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10416 do_div(rt_runtime_us, NSEC_PER_USEC);
10417 return rt_runtime_us;
10420 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10422 u64 rt_runtime, rt_period;
10424 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10425 rt_runtime = tg->rt_bandwidth.rt_runtime;
10427 if (rt_period == 0)
10430 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10433 long sched_group_rt_period(struct task_group *tg)
10437 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10438 do_div(rt_period_us, NSEC_PER_USEC);
10439 return rt_period_us;
10442 static int sched_rt_global_constraints(void)
10444 u64 runtime, period;
10447 if (sysctl_sched_rt_period <= 0)
10450 runtime = global_rt_runtime();
10451 period = global_rt_period();
10454 * Sanity check on the sysctl variables.
10456 if (runtime > period && runtime != RUNTIME_INF)
10459 mutex_lock(&rt_constraints_mutex);
10460 read_lock(&tasklist_lock);
10461 ret = __rt_schedulable(NULL, 0, 0);
10462 read_unlock(&tasklist_lock);
10463 mutex_unlock(&rt_constraints_mutex);
10468 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10470 /* Don't accept realtime tasks when there is no way for them to run */
10471 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10477 #else /* !CONFIG_RT_GROUP_SCHED */
10478 static int sched_rt_global_constraints(void)
10480 unsigned long flags;
10483 if (sysctl_sched_rt_period <= 0)
10487 * There's always some RT tasks in the root group
10488 * -- migration, kstopmachine etc..
10490 if (sysctl_sched_rt_runtime == 0)
10493 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10494 for_each_possible_cpu(i) {
10495 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10497 raw_spin_lock(&rt_rq->rt_runtime_lock);
10498 rt_rq->rt_runtime = global_rt_runtime();
10499 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10501 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10505 #endif /* CONFIG_RT_GROUP_SCHED */
10507 int sched_rt_handler(struct ctl_table *table, int write,
10508 void __user *buffer, size_t *lenp,
10512 int old_period, old_runtime;
10513 static DEFINE_MUTEX(mutex);
10515 mutex_lock(&mutex);
10516 old_period = sysctl_sched_rt_period;
10517 old_runtime = sysctl_sched_rt_runtime;
10519 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10521 if (!ret && write) {
10522 ret = sched_rt_global_constraints();
10524 sysctl_sched_rt_period = old_period;
10525 sysctl_sched_rt_runtime = old_runtime;
10527 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10528 def_rt_bandwidth.rt_period =
10529 ns_to_ktime(global_rt_period());
10532 mutex_unlock(&mutex);
10537 #ifdef CONFIG_CGROUP_SCHED
10539 /* return corresponding task_group object of a cgroup */
10540 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10542 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10543 struct task_group, css);
10546 static struct cgroup_subsys_state *
10547 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10549 struct task_group *tg, *parent;
10551 if (!cgrp->parent) {
10552 /* This is early initialization for the top cgroup */
10553 return &init_task_group.css;
10556 parent = cgroup_tg(cgrp->parent);
10557 tg = sched_create_group(parent);
10559 return ERR_PTR(-ENOMEM);
10565 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10567 struct task_group *tg = cgroup_tg(cgrp);
10569 sched_destroy_group(tg);
10573 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10575 #ifdef CONFIG_RT_GROUP_SCHED
10576 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10579 /* We don't support RT-tasks being in separate groups */
10580 if (tsk->sched_class != &fair_sched_class)
10587 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10588 struct task_struct *tsk, bool threadgroup)
10590 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10594 struct task_struct *c;
10596 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10597 retval = cpu_cgroup_can_attach_task(cgrp, c);
10609 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10610 struct cgroup *old_cont, struct task_struct *tsk,
10613 sched_move_task(tsk);
10615 struct task_struct *c;
10617 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10618 sched_move_task(c);
10624 #ifdef CONFIG_FAIR_GROUP_SCHED
10625 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10628 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10631 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10633 struct task_group *tg = cgroup_tg(cgrp);
10635 return (u64) tg->shares;
10637 #endif /* CONFIG_FAIR_GROUP_SCHED */
10639 #ifdef CONFIG_RT_GROUP_SCHED
10640 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10643 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10646 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10648 return sched_group_rt_runtime(cgroup_tg(cgrp));
10651 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10654 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10657 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10659 return sched_group_rt_period(cgroup_tg(cgrp));
10661 #endif /* CONFIG_RT_GROUP_SCHED */
10663 static struct cftype cpu_files[] = {
10664 #ifdef CONFIG_FAIR_GROUP_SCHED
10667 .read_u64 = cpu_shares_read_u64,
10668 .write_u64 = cpu_shares_write_u64,
10671 #ifdef CONFIG_RT_GROUP_SCHED
10673 .name = "rt_runtime_us",
10674 .read_s64 = cpu_rt_runtime_read,
10675 .write_s64 = cpu_rt_runtime_write,
10678 .name = "rt_period_us",
10679 .read_u64 = cpu_rt_period_read_uint,
10680 .write_u64 = cpu_rt_period_write_uint,
10685 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10687 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10690 struct cgroup_subsys cpu_cgroup_subsys = {
10692 .create = cpu_cgroup_create,
10693 .destroy = cpu_cgroup_destroy,
10694 .can_attach = cpu_cgroup_can_attach,
10695 .attach = cpu_cgroup_attach,
10696 .populate = cpu_cgroup_populate,
10697 .subsys_id = cpu_cgroup_subsys_id,
10701 #endif /* CONFIG_CGROUP_SCHED */
10703 #ifdef CONFIG_CGROUP_CPUACCT
10706 * CPU accounting code for task groups.
10708 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10709 * (balbir@in.ibm.com).
10712 /* track cpu usage of a group of tasks and its child groups */
10714 struct cgroup_subsys_state css;
10715 /* cpuusage holds pointer to a u64-type object on every cpu */
10717 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10718 struct cpuacct *parent;
10721 struct cgroup_subsys cpuacct_subsys;
10723 /* return cpu accounting group corresponding to this container */
10724 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10726 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10727 struct cpuacct, css);
10730 /* return cpu accounting group to which this task belongs */
10731 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10733 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10734 struct cpuacct, css);
10737 /* create a new cpu accounting group */
10738 static struct cgroup_subsys_state *cpuacct_create(
10739 struct cgroup_subsys *ss, struct cgroup *cgrp)
10741 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10747 ca->cpuusage = alloc_percpu(u64);
10751 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10752 if (percpu_counter_init(&ca->cpustat[i], 0))
10753 goto out_free_counters;
10756 ca->parent = cgroup_ca(cgrp->parent);
10762 percpu_counter_destroy(&ca->cpustat[i]);
10763 free_percpu(ca->cpuusage);
10767 return ERR_PTR(-ENOMEM);
10770 /* destroy an existing cpu accounting group */
10772 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10774 struct cpuacct *ca = cgroup_ca(cgrp);
10777 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10778 percpu_counter_destroy(&ca->cpustat[i]);
10779 free_percpu(ca->cpuusage);
10783 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10785 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10788 #ifndef CONFIG_64BIT
10790 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10792 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10794 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10802 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10804 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10806 #ifndef CONFIG_64BIT
10808 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10810 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10812 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10818 /* return total cpu usage (in nanoseconds) of a group */
10819 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10821 struct cpuacct *ca = cgroup_ca(cgrp);
10822 u64 totalcpuusage = 0;
10825 for_each_present_cpu(i)
10826 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10828 return totalcpuusage;
10831 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10834 struct cpuacct *ca = cgroup_ca(cgrp);
10843 for_each_present_cpu(i)
10844 cpuacct_cpuusage_write(ca, i, 0);
10850 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10851 struct seq_file *m)
10853 struct cpuacct *ca = cgroup_ca(cgroup);
10857 for_each_present_cpu(i) {
10858 percpu = cpuacct_cpuusage_read(ca, i);
10859 seq_printf(m, "%llu ", (unsigned long long) percpu);
10861 seq_printf(m, "\n");
10865 static const char *cpuacct_stat_desc[] = {
10866 [CPUACCT_STAT_USER] = "user",
10867 [CPUACCT_STAT_SYSTEM] = "system",
10870 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10871 struct cgroup_map_cb *cb)
10873 struct cpuacct *ca = cgroup_ca(cgrp);
10876 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10877 s64 val = percpu_counter_read(&ca->cpustat[i]);
10878 val = cputime64_to_clock_t(val);
10879 cb->fill(cb, cpuacct_stat_desc[i], val);
10884 static struct cftype files[] = {
10887 .read_u64 = cpuusage_read,
10888 .write_u64 = cpuusage_write,
10891 .name = "usage_percpu",
10892 .read_seq_string = cpuacct_percpu_seq_read,
10896 .read_map = cpuacct_stats_show,
10900 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10902 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10906 * charge this task's execution time to its accounting group.
10908 * called with rq->lock held.
10910 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10912 struct cpuacct *ca;
10915 if (unlikely(!cpuacct_subsys.active))
10918 cpu = task_cpu(tsk);
10924 for (; ca; ca = ca->parent) {
10925 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10926 *cpuusage += cputime;
10933 * Charge the system/user time to the task's accounting group.
10935 static void cpuacct_update_stats(struct task_struct *tsk,
10936 enum cpuacct_stat_index idx, cputime_t val)
10938 struct cpuacct *ca;
10940 if (unlikely(!cpuacct_subsys.active))
10947 percpu_counter_add(&ca->cpustat[idx], val);
10953 struct cgroup_subsys cpuacct_subsys = {
10955 .create = cpuacct_create,
10956 .destroy = cpuacct_destroy,
10957 .populate = cpuacct_populate,
10958 .subsys_id = cpuacct_subsys_id,
10960 #endif /* CONFIG_CGROUP_CPUACCT */
10964 int rcu_expedited_torture_stats(char *page)
10968 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10970 void synchronize_sched_expedited(void)
10973 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10975 #else /* #ifndef CONFIG_SMP */
10977 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10978 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10980 #define RCU_EXPEDITED_STATE_POST -2
10981 #define RCU_EXPEDITED_STATE_IDLE -1
10983 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10985 int rcu_expedited_torture_stats(char *page)
10990 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10991 for_each_online_cpu(cpu) {
10992 cnt += sprintf(&page[cnt], " %d:%d",
10993 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10995 cnt += sprintf(&page[cnt], "\n");
10998 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
11000 static long synchronize_sched_expedited_count;
11003 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
11004 * approach to force grace period to end quickly. This consumes
11005 * significant time on all CPUs, and is thus not recommended for
11006 * any sort of common-case code.
11008 * Note that it is illegal to call this function while holding any
11009 * lock that is acquired by a CPU-hotplug notifier. Failing to
11010 * observe this restriction will result in deadlock.
11012 void synchronize_sched_expedited(void)
11015 unsigned long flags;
11016 bool need_full_sync = 0;
11018 struct migration_req *req;
11022 smp_mb(); /* ensure prior mod happens before capturing snap. */
11023 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
11025 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
11027 if (trycount++ < 10)
11028 udelay(trycount * num_online_cpus());
11030 synchronize_sched();
11033 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
11034 smp_mb(); /* ensure test happens before caller kfree */
11039 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
11040 for_each_online_cpu(cpu) {
11042 req = &per_cpu(rcu_migration_req, cpu);
11043 init_completion(&req->done);
11045 req->dest_cpu = RCU_MIGRATION_NEED_QS;
11046 raw_spin_lock_irqsave(&rq->lock, flags);
11047 list_add(&req->list, &rq->migration_queue);
11048 raw_spin_unlock_irqrestore(&rq->lock, flags);
11049 wake_up_process(rq->migration_thread);
11051 for_each_online_cpu(cpu) {
11052 rcu_expedited_state = cpu;
11053 req = &per_cpu(rcu_migration_req, cpu);
11055 wait_for_completion(&req->done);
11056 raw_spin_lock_irqsave(&rq->lock, flags);
11057 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
11058 need_full_sync = 1;
11059 req->dest_cpu = RCU_MIGRATION_IDLE;
11060 raw_spin_unlock_irqrestore(&rq->lock, flags);
11062 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
11063 synchronize_sched_expedited_count++;
11064 mutex_unlock(&rcu_sched_expedited_mutex);
11066 if (need_full_sync)
11067 synchronize_sched();
11069 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
11071 #endif /* #else #ifndef CONFIG_SMP */