4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy)
124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
129 static inline int task_has_rt_policy(struct task_struct *p)
131 return rt_policy(p->policy);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
142 struct rt_bandwidth {
143 /* nests inside the rq lock: */
144 spinlock_t rt_runtime_lock;
147 struct hrtimer rt_period_timer;
150 static struct rt_bandwidth def_rt_bandwidth;
152 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
154 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169 idle = do_sched_rt_period_timer(rt_b, overrun);
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
176 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
181 spin_lock_init(&rt_b->rt_runtime_lock);
183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime >= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
200 if (hrtimer_active(&rt_b->rt_period_timer))
203 spin_lock(&rt_b->rt_runtime_lock);
208 if (hrtimer_active(&rt_b->rt_period_timer))
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
218 HRTIMER_MODE_ABS_PINNED, 0);
220 spin_unlock(&rt_b->rt_runtime_lock);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
226 hrtimer_cancel(&rt_b->rt_period_timer);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex);
236 #ifdef CONFIG_GROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups);
244 /* task group related information */
246 #ifdef CONFIG_CGROUP_SCHED
247 struct cgroup_subsys_state css;
250 #ifdef CONFIG_USER_SCHED
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
266 struct rt_bandwidth rt_bandwidth;
270 struct list_head list;
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
277 #ifdef CONFIG_USER_SCHED
279 /* Helper function to pass uid information to create_sched_user() */
280 void set_tg_uid(struct user_struct *user)
282 user->tg->uid = user->uid;
287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
290 struct task_group root_task_group;
292 #ifdef CONFIG_FAIR_GROUP_SCHED
293 /* Default task group's sched entity on each cpu */
294 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
295 /* Default task group's cfs_rq on each cpu */
296 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq, init_tg_cfs_rq);
297 #endif /* CONFIG_FAIR_GROUP_SCHED */
299 #ifdef CONFIG_RT_GROUP_SCHED
300 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
301 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq, init_rt_rq);
302 #endif /* CONFIG_RT_GROUP_SCHED */
303 #else /* !CONFIG_USER_SCHED */
304 #define root_task_group init_task_group
305 #endif /* CONFIG_USER_SCHED */
307 /* task_group_lock serializes add/remove of task groups and also changes to
308 * a task group's cpu shares.
310 static DEFINE_SPINLOCK(task_group_lock);
312 #ifdef CONFIG_FAIR_GROUP_SCHED
315 static int root_task_group_empty(void)
317 return list_empty(&root_task_group.children);
321 #ifdef CONFIG_USER_SCHED
322 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
323 #else /* !CONFIG_USER_SCHED */
324 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
325 #endif /* CONFIG_USER_SCHED */
328 * A weight of 0 or 1 can cause arithmetics problems.
329 * A weight of a cfs_rq is the sum of weights of which entities
330 * are queued on this cfs_rq, so a weight of a entity should not be
331 * too large, so as the shares value of a task group.
332 * (The default weight is 1024 - so there's no practical
333 * limitation from this.)
336 #define MAX_SHARES (1UL << 18)
338 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
341 /* Default task group.
342 * Every task in system belong to this group at bootup.
344 struct task_group init_task_group;
346 /* return group to which a task belongs */
347 static inline struct task_group *task_group(struct task_struct *p)
349 struct task_group *tg;
351 #ifdef CONFIG_USER_SCHED
353 tg = __task_cred(p)->user->tg;
355 #elif defined(CONFIG_CGROUP_SCHED)
356 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
357 struct task_group, css);
359 tg = &init_task_group;
364 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
365 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
367 #ifdef CONFIG_FAIR_GROUP_SCHED
368 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
369 p->se.parent = task_group(p)->se[cpu];
372 #ifdef CONFIG_RT_GROUP_SCHED
373 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
374 p->rt.parent = task_group(p)->rt_se[cpu];
380 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
381 static inline struct task_group *task_group(struct task_struct *p)
386 #endif /* CONFIG_GROUP_SCHED */
388 /* CFS-related fields in a runqueue */
390 struct load_weight load;
391 unsigned long nr_running;
396 struct rb_root tasks_timeline;
397 struct rb_node *rb_leftmost;
399 struct list_head tasks;
400 struct list_head *balance_iterator;
403 * 'curr' points to currently running entity on this cfs_rq.
404 * It is set to NULL otherwise (i.e when none are currently running).
406 struct sched_entity *curr, *next, *last;
408 unsigned int nr_spread_over;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
421 struct list_head leaf_cfs_rq_list;
422 struct task_group *tg; /* group that "owns" this runqueue */
426 * the part of load.weight contributed by tasks
428 unsigned long task_weight;
431 * h_load = weight * f(tg)
433 * Where f(tg) is the recursive weight fraction assigned to
436 unsigned long h_load;
439 * this cpu's part of tg->shares
441 unsigned long shares;
444 * load.weight at the time we set shares
446 unsigned long rq_weight;
451 /* Real-Time classes' related field in a runqueue: */
453 struct rt_prio_array active;
454 unsigned long rt_nr_running;
455 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
457 int curr; /* highest queued rt task prio */
459 int next; /* next highest */
464 unsigned long rt_nr_migratory;
465 unsigned long rt_nr_total;
467 struct plist_head pushable_tasks;
472 /* Nests inside the rq lock: */
473 spinlock_t rt_runtime_lock;
475 #ifdef CONFIG_RT_GROUP_SCHED
476 unsigned long rt_nr_boosted;
479 struct list_head leaf_rt_rq_list;
480 struct task_group *tg;
481 struct sched_rt_entity *rt_se;
488 * We add the notion of a root-domain which will be used to define per-domain
489 * variables. Each exclusive cpuset essentially defines an island domain by
490 * fully partitioning the member cpus from any other cpuset. Whenever a new
491 * exclusive cpuset is created, we also create and attach a new root-domain
498 cpumask_var_t online;
501 * The "RT overload" flag: it gets set if a CPU has more than
502 * one runnable RT task.
504 cpumask_var_t rto_mask;
507 struct cpupri cpupri;
512 * By default the system creates a single root-domain with all cpus as
513 * members (mimicking the global state we have today).
515 static struct root_domain def_root_domain;
520 * This is the main, per-CPU runqueue data structure.
522 * Locking rule: those places that want to lock multiple runqueues
523 * (such as the load balancing or the thread migration code), lock
524 * acquire operations must be ordered by ascending &runqueue.
531 * nr_running and cpu_load should be in the same cacheline because
532 * remote CPUs use both these fields when doing load calculation.
534 unsigned long nr_running;
535 #define CPU_LOAD_IDX_MAX 5
536 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
538 unsigned long last_tick_seen;
539 unsigned char in_nohz_recently;
541 /* capture load from *all* tasks on this cpu: */
542 struct load_weight load;
543 unsigned long nr_load_updates;
549 #ifdef CONFIG_FAIR_GROUP_SCHED
550 /* list of leaf cfs_rq on this cpu: */
551 struct list_head leaf_cfs_rq_list;
553 #ifdef CONFIG_RT_GROUP_SCHED
554 struct list_head leaf_rt_rq_list;
558 * This is part of a global counter where only the total sum
559 * over all CPUs matters. A task can increase this counter on
560 * one CPU and if it got migrated afterwards it may decrease
561 * it on another CPU. Always updated under the runqueue lock:
563 unsigned long nr_uninterruptible;
565 struct task_struct *curr, *idle;
566 unsigned long next_balance;
567 struct mm_struct *prev_mm;
574 struct root_domain *rd;
575 struct sched_domain *sd;
577 unsigned char idle_at_tick;
578 /* For active balancing */
582 /* cpu of this runqueue: */
586 unsigned long avg_load_per_task;
588 struct task_struct *migration_thread;
589 struct list_head migration_queue;
597 /* calc_load related fields */
598 unsigned long calc_load_update;
599 long calc_load_active;
601 #ifdef CONFIG_SCHED_HRTICK
603 int hrtick_csd_pending;
604 struct call_single_data hrtick_csd;
606 struct hrtimer hrtick_timer;
609 #ifdef CONFIG_SCHEDSTATS
611 struct sched_info rq_sched_info;
612 unsigned long long rq_cpu_time;
613 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
615 /* sys_sched_yield() stats */
616 unsigned int yld_count;
618 /* schedule() stats */
619 unsigned int sched_switch;
620 unsigned int sched_count;
621 unsigned int sched_goidle;
623 /* try_to_wake_up() stats */
624 unsigned int ttwu_count;
625 unsigned int ttwu_local;
628 unsigned int bkl_count;
632 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
635 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
637 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
640 static inline int cpu_of(struct rq *rq)
650 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
651 * See detach_destroy_domains: synchronize_sched for details.
653 * The domain tree of any CPU may only be accessed from within
654 * preempt-disabled sections.
656 #define for_each_domain(cpu, __sd) \
657 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
659 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
660 #define this_rq() (&__get_cpu_var(runqueues))
661 #define task_rq(p) cpu_rq(task_cpu(p))
662 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
663 #define raw_rq() (&__raw_get_cpu_var(runqueues))
665 inline void update_rq_clock(struct rq *rq)
667 rq->clock = sched_clock_cpu(cpu_of(rq));
671 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
673 #ifdef CONFIG_SCHED_DEBUG
674 # define const_debug __read_mostly
676 # define const_debug static const
681 * @cpu: the processor in question.
683 * Returns true if the current cpu runqueue is locked.
684 * This interface allows printk to be called with the runqueue lock
685 * held and know whether or not it is OK to wake up the klogd.
687 int runqueue_is_locked(int cpu)
689 return spin_is_locked(&cpu_rq(cpu)->lock);
693 * Debugging: various feature bits
696 #define SCHED_FEAT(name, enabled) \
697 __SCHED_FEAT_##name ,
700 #include "sched_features.h"
705 #define SCHED_FEAT(name, enabled) \
706 (1UL << __SCHED_FEAT_##name) * enabled |
708 const_debug unsigned int sysctl_sched_features =
709 #include "sched_features.h"
714 #ifdef CONFIG_SCHED_DEBUG
715 #define SCHED_FEAT(name, enabled) \
718 static __read_mostly char *sched_feat_names[] = {
719 #include "sched_features.h"
725 static int sched_feat_show(struct seq_file *m, void *v)
729 for (i = 0; sched_feat_names[i]; i++) {
730 if (!(sysctl_sched_features & (1UL << i)))
732 seq_printf(m, "%s ", sched_feat_names[i]);
740 sched_feat_write(struct file *filp, const char __user *ubuf,
741 size_t cnt, loff_t *ppos)
751 if (copy_from_user(&buf, ubuf, cnt))
757 if (strncmp(buf, "NO_", 3) == 0) {
762 for (i = 0; sched_feat_names[i]; i++) {
763 if (strcmp(cmp, sched_feat_names[i]) == 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 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 spin_unlock_irq(&rq->lock);
922 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 ->cpus_allowed
947 static inline int task_is_waking(struct task_struct *p)
949 return unlikely(p->state == TASK_WAKING);
953 * __task_rq_lock - lock the runqueue a given task resides on.
954 * Must be called interrupts disabled.
956 static inline struct rq *__task_rq_lock(struct task_struct *p)
963 spin_lock(&rq->lock);
964 if (likely(rq == task_rq(p)))
966 spin_unlock(&rq->lock);
971 * task_rq_lock - lock the runqueue a given task resides on and disable
972 * interrupts. Note the ordering: we can safely lookup the task_rq without
973 * explicitly disabling preemption.
975 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
981 local_irq_save(*flags);
983 spin_lock(&rq->lock);
984 if (likely(rq == task_rq(p)))
986 spin_unlock_irqrestore(&rq->lock, *flags);
990 void task_rq_unlock_wait(struct task_struct *p)
992 struct rq *rq = task_rq(p);
994 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
995 spin_unlock_wait(&rq->lock);
998 static void __task_rq_unlock(struct rq *rq)
1001 spin_unlock(&rq->lock);
1004 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1005 __releases(rq->lock)
1007 spin_unlock_irqrestore(&rq->lock, *flags);
1011 * this_rq_lock - lock this runqueue and disable interrupts.
1013 static struct rq *this_rq_lock(void)
1014 __acquires(rq->lock)
1018 local_irq_disable();
1020 spin_lock(&rq->lock);
1025 #ifdef CONFIG_SCHED_HRTICK
1027 * Use HR-timers to deliver accurate preemption points.
1029 * Its all a bit involved since we cannot program an hrt while holding the
1030 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1033 * When we get rescheduled we reprogram the hrtick_timer outside of the
1039 * - enabled by features
1040 * - hrtimer is actually high res
1042 static inline int hrtick_enabled(struct rq *rq)
1044 if (!sched_feat(HRTICK))
1046 if (!cpu_active(cpu_of(rq)))
1048 return hrtimer_is_hres_active(&rq->hrtick_timer);
1051 static void hrtick_clear(struct rq *rq)
1053 if (hrtimer_active(&rq->hrtick_timer))
1054 hrtimer_cancel(&rq->hrtick_timer);
1058 * High-resolution timer tick.
1059 * Runs from hardirq context with interrupts disabled.
1061 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1063 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1065 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1067 spin_lock(&rq->lock);
1068 update_rq_clock(rq);
1069 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1070 spin_unlock(&rq->lock);
1072 return HRTIMER_NORESTART;
1077 * called from hardirq (IPI) context
1079 static void __hrtick_start(void *arg)
1081 struct rq *rq = arg;
1083 spin_lock(&rq->lock);
1084 hrtimer_restart(&rq->hrtick_timer);
1085 rq->hrtick_csd_pending = 0;
1086 spin_unlock(&rq->lock);
1090 * Called to set the hrtick timer state.
1092 * called with rq->lock held and irqs disabled
1094 static void hrtick_start(struct rq *rq, u64 delay)
1096 struct hrtimer *timer = &rq->hrtick_timer;
1097 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1099 hrtimer_set_expires(timer, time);
1101 if (rq == this_rq()) {
1102 hrtimer_restart(timer);
1103 } else if (!rq->hrtick_csd_pending) {
1104 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1105 rq->hrtick_csd_pending = 1;
1110 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1112 int cpu = (int)(long)hcpu;
1115 case CPU_UP_CANCELED:
1116 case CPU_UP_CANCELED_FROZEN:
1117 case CPU_DOWN_PREPARE:
1118 case CPU_DOWN_PREPARE_FROZEN:
1120 case CPU_DEAD_FROZEN:
1121 hrtick_clear(cpu_rq(cpu));
1128 static __init void init_hrtick(void)
1130 hotcpu_notifier(hotplug_hrtick, 0);
1134 * Called to set the hrtick timer state.
1136 * called with rq->lock held and irqs disabled
1138 static void hrtick_start(struct rq *rq, u64 delay)
1140 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1141 HRTIMER_MODE_REL_PINNED, 0);
1144 static inline void init_hrtick(void)
1147 #endif /* CONFIG_SMP */
1149 static void init_rq_hrtick(struct rq *rq)
1152 rq->hrtick_csd_pending = 0;
1154 rq->hrtick_csd.flags = 0;
1155 rq->hrtick_csd.func = __hrtick_start;
1156 rq->hrtick_csd.info = rq;
1159 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1160 rq->hrtick_timer.function = hrtick;
1162 #else /* CONFIG_SCHED_HRTICK */
1163 static inline void hrtick_clear(struct rq *rq)
1167 static inline void init_rq_hrtick(struct rq *rq)
1171 static inline void init_hrtick(void)
1174 #endif /* CONFIG_SCHED_HRTICK */
1177 * resched_task - mark a task 'to be rescheduled now'.
1179 * On UP this means the setting of the need_resched flag, on SMP it
1180 * might also involve a cross-CPU call to trigger the scheduler on
1185 #ifndef tsk_is_polling
1186 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1189 static void resched_task(struct task_struct *p)
1193 assert_spin_locked(&task_rq(p)->lock);
1195 if (test_tsk_need_resched(p))
1198 set_tsk_need_resched(p);
1201 if (cpu == smp_processor_id())
1204 /* NEED_RESCHED must be visible before we test polling */
1206 if (!tsk_is_polling(p))
1207 smp_send_reschedule(cpu);
1210 static void resched_cpu(int cpu)
1212 struct rq *rq = cpu_rq(cpu);
1213 unsigned long flags;
1215 if (!spin_trylock_irqsave(&rq->lock, flags))
1217 resched_task(cpu_curr(cpu));
1218 spin_unlock_irqrestore(&rq->lock, flags);
1223 * When add_timer_on() enqueues a timer into the timer wheel of an
1224 * idle CPU then this timer might expire before the next timer event
1225 * which is scheduled to wake up that CPU. In case of a completely
1226 * idle system the next event might even be infinite time into the
1227 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1228 * leaves the inner idle loop so the newly added timer is taken into
1229 * account when the CPU goes back to idle and evaluates the timer
1230 * wheel for the next timer event.
1232 void wake_up_idle_cpu(int cpu)
1234 struct rq *rq = cpu_rq(cpu);
1236 if (cpu == smp_processor_id())
1240 * This is safe, as this function is called with the timer
1241 * wheel base lock of (cpu) held. When the CPU is on the way
1242 * to idle and has not yet set rq->curr to idle then it will
1243 * be serialized on the timer wheel base lock and take the new
1244 * timer into account automatically.
1246 if (rq->curr != rq->idle)
1250 * We can set TIF_RESCHED on the idle task of the other CPU
1251 * lockless. The worst case is that the other CPU runs the
1252 * idle task through an additional NOOP schedule()
1254 set_tsk_need_resched(rq->idle);
1256 /* NEED_RESCHED must be visible before we test polling */
1258 if (!tsk_is_polling(rq->idle))
1259 smp_send_reschedule(cpu);
1261 #endif /* CONFIG_NO_HZ */
1263 static u64 sched_avg_period(void)
1265 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1268 static void sched_avg_update(struct rq *rq)
1270 s64 period = sched_avg_period();
1272 while ((s64)(rq->clock - rq->age_stamp) > period) {
1274 * Inline assembly required to prevent the compiler
1275 * optimising this loop into a divmod call.
1276 * See __iter_div_u64_rem() for another example of this.
1278 asm("" : "+rm" (rq->age_stamp));
1279 rq->age_stamp += period;
1284 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1286 rq->rt_avg += rt_delta;
1287 sched_avg_update(rq);
1290 #else /* !CONFIG_SMP */
1291 static void resched_task(struct task_struct *p)
1293 assert_spin_locked(&task_rq(p)->lock);
1294 set_tsk_need_resched(p);
1297 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1300 #endif /* CONFIG_SMP */
1302 #if BITS_PER_LONG == 32
1303 # define WMULT_CONST (~0UL)
1305 # define WMULT_CONST (1UL << 32)
1308 #define WMULT_SHIFT 32
1311 * Shift right and round:
1313 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1316 * delta *= weight / lw
1318 static unsigned long
1319 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1320 struct load_weight *lw)
1324 if (!lw->inv_weight) {
1325 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1328 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1332 tmp = (u64)delta_exec * weight;
1334 * Check whether we'd overflow the 64-bit multiplication:
1336 if (unlikely(tmp > WMULT_CONST))
1337 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1340 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1342 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1345 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1351 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1358 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1359 * of tasks with abnormal "nice" values across CPUs the contribution that
1360 * each task makes to its run queue's load is weighted according to its
1361 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1362 * scaled version of the new time slice allocation that they receive on time
1366 #define WEIGHT_IDLEPRIO 3
1367 #define WMULT_IDLEPRIO 1431655765
1370 * Nice levels are multiplicative, with a gentle 10% change for every
1371 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1372 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1373 * that remained on nice 0.
1375 * The "10% effect" is relative and cumulative: from _any_ nice level,
1376 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1377 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1378 * If a task goes up by ~10% and another task goes down by ~10% then
1379 * the relative distance between them is ~25%.)
1381 static const int prio_to_weight[40] = {
1382 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1383 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1384 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1385 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1386 /* 0 */ 1024, 820, 655, 526, 423,
1387 /* 5 */ 335, 272, 215, 172, 137,
1388 /* 10 */ 110, 87, 70, 56, 45,
1389 /* 15 */ 36, 29, 23, 18, 15,
1393 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1395 * In cases where the weight does not change often, we can use the
1396 * precalculated inverse to speed up arithmetics by turning divisions
1397 * into multiplications:
1399 static const u32 prio_to_wmult[40] = {
1400 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1401 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1402 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1403 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1404 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1405 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1406 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1407 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1410 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1413 * runqueue iterator, to support SMP load-balancing between different
1414 * scheduling classes, without having to expose their internal data
1415 * structures to the load-balancing proper:
1417 struct rq_iterator {
1419 struct task_struct *(*start)(void *);
1420 struct task_struct *(*next)(void *);
1424 static unsigned long
1425 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1426 unsigned long max_load_move, struct sched_domain *sd,
1427 enum cpu_idle_type idle, int *all_pinned,
1428 int *this_best_prio, struct rq_iterator *iterator);
1431 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1432 struct sched_domain *sd, enum cpu_idle_type idle,
1433 struct rq_iterator *iterator);
1436 /* Time spent by the tasks of the cpu accounting group executing in ... */
1437 enum cpuacct_stat_index {
1438 CPUACCT_STAT_USER, /* ... user mode */
1439 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1441 CPUACCT_STAT_NSTATS,
1444 #ifdef CONFIG_CGROUP_CPUACCT
1445 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1446 static void cpuacct_update_stats(struct task_struct *tsk,
1447 enum cpuacct_stat_index idx, cputime_t val);
1449 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1450 static inline void cpuacct_update_stats(struct task_struct *tsk,
1451 enum cpuacct_stat_index idx, cputime_t val) {}
1454 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1456 update_load_add(&rq->load, load);
1459 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1461 update_load_sub(&rq->load, load);
1464 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1465 typedef int (*tg_visitor)(struct task_group *, void *);
1468 * Iterate the full tree, calling @down when first entering a node and @up when
1469 * leaving it for the final time.
1471 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1473 struct task_group *parent, *child;
1477 parent = &root_task_group;
1479 ret = (*down)(parent, data);
1482 list_for_each_entry_rcu(child, &parent->children, siblings) {
1489 ret = (*up)(parent, data);
1494 parent = parent->parent;
1503 static int tg_nop(struct task_group *tg, void *data)
1510 /* Used instead of source_load when we know the type == 0 */
1511 static unsigned long weighted_cpuload(const int cpu)
1513 return cpu_rq(cpu)->load.weight;
1517 * Return a low guess at the load of a migration-source cpu weighted
1518 * according to the scheduling class and "nice" value.
1520 * We want to under-estimate the load of migration sources, to
1521 * balance conservatively.
1523 static unsigned long source_load(int cpu, int type)
1525 struct rq *rq = cpu_rq(cpu);
1526 unsigned long total = weighted_cpuload(cpu);
1528 if (type == 0 || !sched_feat(LB_BIAS))
1531 return min(rq->cpu_load[type-1], total);
1535 * Return a high guess at the load of a migration-target cpu weighted
1536 * according to the scheduling class and "nice" value.
1538 static unsigned long target_load(int cpu, int type)
1540 struct rq *rq = cpu_rq(cpu);
1541 unsigned long total = weighted_cpuload(cpu);
1543 if (type == 0 || !sched_feat(LB_BIAS))
1546 return max(rq->cpu_load[type-1], total);
1549 static struct sched_group *group_of(int cpu)
1551 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1559 static unsigned long power_of(int cpu)
1561 struct sched_group *group = group_of(cpu);
1564 return SCHED_LOAD_SCALE;
1566 return group->cpu_power;
1569 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1571 static unsigned long cpu_avg_load_per_task(int cpu)
1573 struct rq *rq = cpu_rq(cpu);
1574 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1577 rq->avg_load_per_task = rq->load.weight / nr_running;
1579 rq->avg_load_per_task = 0;
1581 return rq->avg_load_per_task;
1584 #ifdef CONFIG_FAIR_GROUP_SCHED
1586 static __read_mostly unsigned long *update_shares_data;
1588 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1591 * Calculate and set the cpu's group shares.
1593 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1594 unsigned long sd_shares,
1595 unsigned long sd_rq_weight,
1596 unsigned long *usd_rq_weight)
1598 unsigned long shares, rq_weight;
1601 rq_weight = usd_rq_weight[cpu];
1604 rq_weight = NICE_0_LOAD;
1608 * \Sum_j shares_j * rq_weight_i
1609 * shares_i = -----------------------------
1610 * \Sum_j rq_weight_j
1612 shares = (sd_shares * rq_weight) / sd_rq_weight;
1613 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1615 if (abs(shares - tg->se[cpu]->load.weight) >
1616 sysctl_sched_shares_thresh) {
1617 struct rq *rq = cpu_rq(cpu);
1618 unsigned long flags;
1620 spin_lock_irqsave(&rq->lock, flags);
1621 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1622 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1623 __set_se_shares(tg->se[cpu], shares);
1624 spin_unlock_irqrestore(&rq->lock, flags);
1629 * Re-compute the task group their per cpu shares over the given domain.
1630 * This needs to be done in a bottom-up fashion because the rq weight of a
1631 * parent group depends on the shares of its child groups.
1633 static int tg_shares_up(struct task_group *tg, void *data)
1635 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1636 unsigned long *usd_rq_weight;
1637 struct sched_domain *sd = data;
1638 unsigned long flags;
1644 local_irq_save(flags);
1645 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1647 for_each_cpu(i, sched_domain_span(sd)) {
1648 weight = tg->cfs_rq[i]->load.weight;
1649 usd_rq_weight[i] = weight;
1651 rq_weight += weight;
1653 * If there are currently no tasks on the cpu pretend there
1654 * is one of average load so that when a new task gets to
1655 * run here it will not get delayed by group starvation.
1658 weight = NICE_0_LOAD;
1660 sum_weight += weight;
1661 shares += tg->cfs_rq[i]->shares;
1665 rq_weight = sum_weight;
1667 if ((!shares && rq_weight) || shares > tg->shares)
1668 shares = tg->shares;
1670 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1671 shares = tg->shares;
1673 for_each_cpu(i, sched_domain_span(sd))
1674 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1676 local_irq_restore(flags);
1682 * Compute the cpu's hierarchical load factor for each task group.
1683 * This needs to be done in a top-down fashion because the load of a child
1684 * group is a fraction of its parents load.
1686 static int tg_load_down(struct task_group *tg, void *data)
1689 long cpu = (long)data;
1692 load = cpu_rq(cpu)->load.weight;
1694 load = tg->parent->cfs_rq[cpu]->h_load;
1695 load *= tg->cfs_rq[cpu]->shares;
1696 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1699 tg->cfs_rq[cpu]->h_load = load;
1704 static void update_shares(struct sched_domain *sd)
1709 if (root_task_group_empty())
1712 now = cpu_clock(raw_smp_processor_id());
1713 elapsed = now - sd->last_update;
1715 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1716 sd->last_update = now;
1717 walk_tg_tree(tg_nop, tg_shares_up, sd);
1721 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1723 if (root_task_group_empty())
1726 spin_unlock(&rq->lock);
1728 spin_lock(&rq->lock);
1731 static void update_h_load(long cpu)
1733 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1738 static inline void update_shares(struct sched_domain *sd)
1742 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1748 #ifdef CONFIG_PREEMPT
1750 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1753 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1754 * way at the expense of forcing extra atomic operations in all
1755 * invocations. This assures that the double_lock is acquired using the
1756 * same underlying policy as the spinlock_t on this architecture, which
1757 * reduces latency compared to the unfair variant below. However, it
1758 * also adds more overhead and therefore may reduce throughput.
1760 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1761 __releases(this_rq->lock)
1762 __acquires(busiest->lock)
1763 __acquires(this_rq->lock)
1765 spin_unlock(&this_rq->lock);
1766 double_rq_lock(this_rq, busiest);
1773 * Unfair double_lock_balance: Optimizes throughput at the expense of
1774 * latency by eliminating extra atomic operations when the locks are
1775 * already in proper order on entry. This favors lower cpu-ids and will
1776 * grant the double lock to lower cpus over higher ids under contention,
1777 * regardless of entry order into the function.
1779 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1780 __releases(this_rq->lock)
1781 __acquires(busiest->lock)
1782 __acquires(this_rq->lock)
1786 if (unlikely(!spin_trylock(&busiest->lock))) {
1787 if (busiest < this_rq) {
1788 spin_unlock(&this_rq->lock);
1789 spin_lock(&busiest->lock);
1790 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1793 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1798 #endif /* CONFIG_PREEMPT */
1801 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1803 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1805 if (unlikely(!irqs_disabled())) {
1806 /* printk() doesn't work good under rq->lock */
1807 spin_unlock(&this_rq->lock);
1811 return _double_lock_balance(this_rq, busiest);
1814 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1815 __releases(busiest->lock)
1817 spin_unlock(&busiest->lock);
1818 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1822 #ifdef CONFIG_FAIR_GROUP_SCHED
1823 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1826 cfs_rq->shares = shares;
1831 static void calc_load_account_active(struct rq *this_rq);
1832 static void update_sysctl(void);
1834 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1836 set_task_rq(p, cpu);
1839 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1840 * successfuly executed on another CPU. We must ensure that updates of
1841 * per-task data have been completed by this moment.
1844 task_thread_info(p)->cpu = cpu;
1848 #include "sched_stats.h"
1849 #include "sched_idletask.c"
1850 #include "sched_fair.c"
1851 #include "sched_rt.c"
1852 #ifdef CONFIG_SCHED_DEBUG
1853 # include "sched_debug.c"
1856 #define sched_class_highest (&rt_sched_class)
1857 #define for_each_class(class) \
1858 for (class = sched_class_highest; class; class = class->next)
1860 static void inc_nr_running(struct rq *rq)
1865 static void dec_nr_running(struct rq *rq)
1870 static void set_load_weight(struct task_struct *p)
1872 if (task_has_rt_policy(p)) {
1873 p->se.load.weight = prio_to_weight[0] * 2;
1874 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1879 * SCHED_IDLE tasks get minimal weight:
1881 if (p->policy == SCHED_IDLE) {
1882 p->se.load.weight = WEIGHT_IDLEPRIO;
1883 p->se.load.inv_weight = WMULT_IDLEPRIO;
1887 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1888 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1891 static void update_avg(u64 *avg, u64 sample)
1893 s64 diff = sample - *avg;
1898 enqueue_task(struct rq *rq, struct task_struct *p, int wakeup, bool head)
1901 p->se.start_runtime = p->se.sum_exec_runtime;
1903 sched_info_queued(p);
1904 p->sched_class->enqueue_task(rq, p, wakeup, head);
1908 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1911 if (p->se.last_wakeup) {
1912 update_avg(&p->se.avg_overlap,
1913 p->se.sum_exec_runtime - p->se.last_wakeup);
1914 p->se.last_wakeup = 0;
1916 update_avg(&p->se.avg_wakeup,
1917 sysctl_sched_wakeup_granularity);
1921 sched_info_dequeued(p);
1922 p->sched_class->dequeue_task(rq, p, sleep);
1927 * __normal_prio - return the priority that is based on the static prio
1929 static inline int __normal_prio(struct task_struct *p)
1931 return p->static_prio;
1935 * Calculate the expected normal priority: i.e. priority
1936 * without taking RT-inheritance into account. Might be
1937 * boosted by interactivity modifiers. Changes upon fork,
1938 * setprio syscalls, and whenever the interactivity
1939 * estimator recalculates.
1941 static inline int normal_prio(struct task_struct *p)
1945 if (task_has_rt_policy(p))
1946 prio = MAX_RT_PRIO-1 - p->rt_priority;
1948 prio = __normal_prio(p);
1953 * Calculate the current priority, i.e. the priority
1954 * taken into account by the scheduler. This value might
1955 * be boosted by RT tasks, or might be boosted by
1956 * interactivity modifiers. Will be RT if the task got
1957 * RT-boosted. If not then it returns p->normal_prio.
1959 static int effective_prio(struct task_struct *p)
1961 p->normal_prio = normal_prio(p);
1963 * If we are RT tasks or we were boosted to RT priority,
1964 * keep the priority unchanged. Otherwise, update priority
1965 * to the normal priority:
1967 if (!rt_prio(p->prio))
1968 return p->normal_prio;
1973 * activate_task - move a task to the runqueue.
1975 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1977 if (task_contributes_to_load(p))
1978 rq->nr_uninterruptible--;
1980 enqueue_task(rq, p, wakeup, false);
1985 * deactivate_task - remove a task from the runqueue.
1987 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1989 if (task_contributes_to_load(p))
1990 rq->nr_uninterruptible++;
1992 dequeue_task(rq, p, sleep);
1997 * task_curr - is this task currently executing on a CPU?
1998 * @p: the task in question.
2000 inline int task_curr(const struct task_struct *p)
2002 return cpu_curr(task_cpu(p)) == p;
2005 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2006 const struct sched_class *prev_class,
2007 int oldprio, int running)
2009 if (prev_class != p->sched_class) {
2010 if (prev_class->switched_from)
2011 prev_class->switched_from(rq, p, running);
2012 p->sched_class->switched_to(rq, p, running);
2014 p->sched_class->prio_changed(rq, p, oldprio, running);
2018 * kthread_bind - bind a just-created kthread to a cpu.
2019 * @p: thread created by kthread_create().
2020 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2022 * Description: This function is equivalent to set_cpus_allowed(),
2023 * except that @cpu doesn't need to be online, and the thread must be
2024 * stopped (i.e., just returned from kthread_create()).
2026 * Function lives here instead of kthread.c because it messes with
2027 * scheduler internals which require locking.
2029 void kthread_bind(struct task_struct *p, unsigned int cpu)
2031 /* Must have done schedule() in kthread() before we set_task_cpu */
2032 if (!wait_task_inactive(p, TASK_UNINTERRUPTIBLE)) {
2037 p->cpus_allowed = cpumask_of_cpu(cpu);
2038 p->rt.nr_cpus_allowed = 1;
2039 p->flags |= PF_THREAD_BOUND;
2041 EXPORT_SYMBOL(kthread_bind);
2045 * Is this task likely cache-hot:
2048 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2052 if (p->sched_class != &fair_sched_class)
2056 * Buddy candidates are cache hot:
2058 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2059 (&p->se == cfs_rq_of(&p->se)->next ||
2060 &p->se == cfs_rq_of(&p->se)->last))
2063 if (sysctl_sched_migration_cost == -1)
2065 if (sysctl_sched_migration_cost == 0)
2068 delta = now - p->se.exec_start;
2070 return delta < (s64)sysctl_sched_migration_cost;
2074 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2076 int old_cpu = task_cpu(p);
2078 #ifdef CONFIG_SCHED_DEBUG
2080 * We should never call set_task_cpu() on a blocked task,
2081 * ttwu() will sort out the placement.
2083 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2084 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2087 trace_sched_migrate_task(p, new_cpu);
2089 if (old_cpu != new_cpu) {
2090 p->se.nr_migrations++;
2091 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2095 __set_task_cpu(p, new_cpu);
2098 struct migration_req {
2099 struct list_head list;
2101 struct task_struct *task;
2104 struct completion done;
2108 * The task's runqueue lock must be held.
2109 * Returns true if you have to wait for migration thread.
2112 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2114 struct rq *rq = task_rq(p);
2117 * If the task is not on a runqueue (and not running), then
2118 * the next wake-up will properly place the task.
2120 if (!p->se.on_rq && !task_running(rq, p))
2123 init_completion(&req->done);
2125 req->dest_cpu = dest_cpu;
2126 list_add(&req->list, &rq->migration_queue);
2132 * wait_task_context_switch - wait for a thread to complete at least one
2135 * @p must not be current.
2137 void wait_task_context_switch(struct task_struct *p)
2139 unsigned long nvcsw, nivcsw, flags;
2147 * The runqueue is assigned before the actual context
2148 * switch. We need to take the runqueue lock.
2150 * We could check initially without the lock but it is
2151 * very likely that we need to take the lock in every
2154 rq = task_rq_lock(p, &flags);
2155 running = task_running(rq, p);
2156 task_rq_unlock(rq, &flags);
2158 if (likely(!running))
2161 * The switch count is incremented before the actual
2162 * context switch. We thus wait for two switches to be
2163 * sure at least one completed.
2165 if ((p->nvcsw - nvcsw) > 1)
2167 if ((p->nivcsw - nivcsw) > 1)
2175 * wait_task_inactive - wait for a thread to unschedule.
2177 * If @match_state is nonzero, it's the @p->state value just checked and
2178 * not expected to change. If it changes, i.e. @p might have woken up,
2179 * then return zero. When we succeed in waiting for @p to be off its CPU,
2180 * we return a positive number (its total switch count). If a second call
2181 * a short while later returns the same number, the caller can be sure that
2182 * @p has remained unscheduled the whole time.
2184 * The caller must ensure that the task *will* unschedule sometime soon,
2185 * else this function might spin for a *long* time. This function can't
2186 * be called with interrupts off, or it may introduce deadlock with
2187 * smp_call_function() if an IPI is sent by the same process we are
2188 * waiting to become inactive.
2190 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2192 unsigned long flags;
2199 * We do the initial early heuristics without holding
2200 * any task-queue locks at all. We'll only try to get
2201 * the runqueue lock when things look like they will
2207 * If the task is actively running on another CPU
2208 * still, just relax and busy-wait without holding
2211 * NOTE! Since we don't hold any locks, it's not
2212 * even sure that "rq" stays as the right runqueue!
2213 * But we don't care, since "task_running()" will
2214 * return false if the runqueue has changed and p
2215 * is actually now running somewhere else!
2217 while (task_running(rq, p)) {
2218 if (match_state && unlikely(p->state != match_state))
2224 * Ok, time to look more closely! We need the rq
2225 * lock now, to be *sure*. If we're wrong, we'll
2226 * just go back and repeat.
2228 rq = task_rq_lock(p, &flags);
2229 trace_sched_wait_task(rq, p);
2230 running = task_running(rq, p);
2231 on_rq = p->se.on_rq;
2233 if (!match_state || p->state == match_state)
2234 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2235 task_rq_unlock(rq, &flags);
2238 * If it changed from the expected state, bail out now.
2240 if (unlikely(!ncsw))
2244 * Was it really running after all now that we
2245 * checked with the proper locks actually held?
2247 * Oops. Go back and try again..
2249 if (unlikely(running)) {
2255 * It's not enough that it's not actively running,
2256 * it must be off the runqueue _entirely_, and not
2259 * So if it was still runnable (but just not actively
2260 * running right now), it's preempted, and we should
2261 * yield - it could be a while.
2263 if (unlikely(on_rq)) {
2264 schedule_timeout_uninterruptible(1);
2269 * Ahh, all good. It wasn't running, and it wasn't
2270 * runnable, which means that it will never become
2271 * running in the future either. We're all done!
2280 * kick_process - kick a running thread to enter/exit the kernel
2281 * @p: the to-be-kicked thread
2283 * Cause a process which is running on another CPU to enter
2284 * kernel-mode, without any delay. (to get signals handled.)
2286 * NOTE: this function doesnt have to take the runqueue lock,
2287 * because all it wants to ensure is that the remote task enters
2288 * the kernel. If the IPI races and the task has been migrated
2289 * to another CPU then no harm is done and the purpose has been
2292 void kick_process(struct task_struct *p)
2298 if ((cpu != smp_processor_id()) && task_curr(p))
2299 smp_send_reschedule(cpu);
2302 EXPORT_SYMBOL_GPL(kick_process);
2303 #endif /* CONFIG_SMP */
2306 * task_oncpu_function_call - call a function on the cpu on which a task runs
2307 * @p: the task to evaluate
2308 * @func: the function to be called
2309 * @info: the function call argument
2311 * Calls the function @func when the task is currently running. This might
2312 * be on the current CPU, which just calls the function directly
2314 void task_oncpu_function_call(struct task_struct *p,
2315 void (*func) (void *info), void *info)
2322 smp_call_function_single(cpu, func, info, 1);
2328 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2330 static int select_fallback_rq(int cpu, struct task_struct *p)
2333 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2335 /* Look for allowed, online CPU in same node. */
2336 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2337 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2340 /* Any allowed, online CPU? */
2341 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2342 if (dest_cpu < nr_cpu_ids)
2345 /* No more Mr. Nice Guy. */
2346 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2347 dest_cpu = cpuset_cpus_allowed_fallback(p);
2349 * Don't tell them about moving exiting tasks or
2350 * kernel threads (both mm NULL), since they never
2353 if (p->mm && printk_ratelimit()) {
2354 printk(KERN_INFO "process %d (%s) no "
2355 "longer affine to cpu%d\n",
2356 task_pid_nr(p), p->comm, cpu);
2364 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2367 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2369 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2372 * In order not to call set_task_cpu() on a blocking task we need
2373 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2376 * Since this is common to all placement strategies, this lives here.
2378 * [ this allows ->select_task() to simply return task_cpu(p) and
2379 * not worry about this generic constraint ]
2381 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2383 cpu = select_fallback_rq(task_cpu(p), p);
2390 * try_to_wake_up - wake up a thread
2391 * @p: the to-be-woken-up thread
2392 * @state: the mask of task states that can be woken
2393 * @sync: do a synchronous wakeup?
2395 * Put it on the run-queue if it's not already there. The "current"
2396 * thread is always on the run-queue (except when the actual
2397 * re-schedule is in progress), and as such you're allowed to do
2398 * the simpler "current->state = TASK_RUNNING" to mark yourself
2399 * runnable without the overhead of this.
2401 * returns failure only if the task is already active.
2403 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2406 int cpu, orig_cpu, this_cpu, success = 0;
2407 unsigned long flags;
2408 struct rq *rq, *orig_rq;
2410 if (!sched_feat(SYNC_WAKEUPS))
2411 wake_flags &= ~WF_SYNC;
2413 this_cpu = get_cpu();
2416 rq = orig_rq = task_rq_lock(p, &flags);
2417 update_rq_clock(rq);
2418 if (!(p->state & state))
2428 if (unlikely(task_running(rq, p)))
2432 * In order to handle concurrent wakeups and release the rq->lock
2433 * we put the task in TASK_WAKING state.
2435 * First fix up the nr_uninterruptible count:
2437 if (task_contributes_to_load(p)) {
2438 if (likely(cpu_online(orig_cpu)))
2439 rq->nr_uninterruptible--;
2441 this_rq()->nr_uninterruptible--;
2443 p->state = TASK_WAKING;
2445 if (p->sched_class->task_waking)
2446 p->sched_class->task_waking(rq, p);
2448 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2449 if (cpu != orig_cpu)
2450 set_task_cpu(p, cpu);
2451 __task_rq_unlock(rq);
2454 spin_lock(&rq->lock);
2455 update_rq_clock(rq);
2458 * We migrated the task without holding either rq->lock, however
2459 * since the task is not on the task list itself, nobody else
2460 * will try and migrate the task, hence the rq should match the
2461 * cpu we just moved it to.
2463 WARN_ON(task_cpu(p) != cpu);
2464 WARN_ON(p->state != TASK_WAKING);
2466 #ifdef CONFIG_SCHEDSTATS
2467 schedstat_inc(rq, ttwu_count);
2468 if (cpu == this_cpu)
2469 schedstat_inc(rq, ttwu_local);
2471 struct sched_domain *sd;
2472 for_each_domain(this_cpu, sd) {
2473 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2474 schedstat_inc(sd, ttwu_wake_remote);
2479 #endif /* CONFIG_SCHEDSTATS */
2482 #endif /* CONFIG_SMP */
2483 schedstat_inc(p, se.nr_wakeups);
2484 if (wake_flags & WF_SYNC)
2485 schedstat_inc(p, se.nr_wakeups_sync);
2486 if (orig_cpu != cpu)
2487 schedstat_inc(p, se.nr_wakeups_migrate);
2488 if (cpu == this_cpu)
2489 schedstat_inc(p, se.nr_wakeups_local);
2491 schedstat_inc(p, se.nr_wakeups_remote);
2492 activate_task(rq, p, 1);
2496 * Only attribute actual wakeups done by this task.
2498 if (!in_interrupt()) {
2499 struct sched_entity *se = ¤t->se;
2500 u64 sample = se->sum_exec_runtime;
2502 if (se->last_wakeup)
2503 sample -= se->last_wakeup;
2505 sample -= se->start_runtime;
2506 update_avg(&se->avg_wakeup, sample);
2508 se->last_wakeup = se->sum_exec_runtime;
2512 trace_sched_wakeup(rq, p, success);
2513 check_preempt_curr(rq, p, wake_flags);
2515 p->state = TASK_RUNNING;
2517 if (p->sched_class->task_woken)
2518 p->sched_class->task_woken(rq, p);
2520 if (unlikely(rq->idle_stamp)) {
2521 u64 delta = rq->clock - rq->idle_stamp;
2522 u64 max = 2*sysctl_sched_migration_cost;
2527 update_avg(&rq->avg_idle, delta);
2532 task_rq_unlock(rq, &flags);
2539 * wake_up_process - Wake up a specific process
2540 * @p: The process to be woken up.
2542 * Attempt to wake up the nominated process and move it to the set of runnable
2543 * processes. Returns 1 if the process was woken up, 0 if it was already
2546 * It may be assumed that this function implies a write memory barrier before
2547 * changing the task state if and only if any tasks are woken up.
2549 int wake_up_process(struct task_struct *p)
2551 return try_to_wake_up(p, TASK_ALL, 0);
2553 EXPORT_SYMBOL(wake_up_process);
2555 int wake_up_state(struct task_struct *p, unsigned int state)
2557 return try_to_wake_up(p, state, 0);
2561 * Perform scheduler related setup for a newly forked process p.
2562 * p is forked by current.
2564 * __sched_fork() is basic setup used by init_idle() too:
2566 static void __sched_fork(struct task_struct *p)
2568 p->se.exec_start = 0;
2569 p->se.sum_exec_runtime = 0;
2570 p->se.prev_sum_exec_runtime = 0;
2571 p->se.nr_migrations = 0;
2572 p->se.last_wakeup = 0;
2573 p->se.avg_overlap = 0;
2574 p->se.start_runtime = 0;
2575 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2576 p->se.avg_running = 0;
2578 #ifdef CONFIG_SCHEDSTATS
2579 p->se.wait_start = 0;
2581 p->se.wait_count = 0;
2584 p->se.sleep_start = 0;
2585 p->se.sleep_max = 0;
2586 p->se.sum_sleep_runtime = 0;
2588 p->se.block_start = 0;
2589 p->se.block_max = 0;
2591 p->se.slice_max = 0;
2593 p->se.nr_migrations_cold = 0;
2594 p->se.nr_failed_migrations_affine = 0;
2595 p->se.nr_failed_migrations_running = 0;
2596 p->se.nr_failed_migrations_hot = 0;
2597 p->se.nr_forced_migrations = 0;
2599 p->se.nr_wakeups = 0;
2600 p->se.nr_wakeups_sync = 0;
2601 p->se.nr_wakeups_migrate = 0;
2602 p->se.nr_wakeups_local = 0;
2603 p->se.nr_wakeups_remote = 0;
2604 p->se.nr_wakeups_affine = 0;
2605 p->se.nr_wakeups_affine_attempts = 0;
2606 p->se.nr_wakeups_passive = 0;
2607 p->se.nr_wakeups_idle = 0;
2611 INIT_LIST_HEAD(&p->rt.run_list);
2613 INIT_LIST_HEAD(&p->se.group_node);
2615 #ifdef CONFIG_PREEMPT_NOTIFIERS
2616 INIT_HLIST_HEAD(&p->preempt_notifiers);
2621 * fork()/clone()-time setup:
2623 void sched_fork(struct task_struct *p, int clone_flags)
2625 int cpu = get_cpu();
2629 * We mark the process as running here. This guarantees that
2630 * nobody will actually run it, and a signal or other external
2631 * event cannot wake it up and insert it on the runqueue either.
2633 p->state = TASK_RUNNING;
2636 * Revert to default priority/policy on fork if requested.
2638 if (unlikely(p->sched_reset_on_fork)) {
2639 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2640 p->policy = SCHED_NORMAL;
2641 p->normal_prio = p->static_prio;
2644 if (PRIO_TO_NICE(p->static_prio) < 0) {
2645 p->static_prio = NICE_TO_PRIO(0);
2646 p->normal_prio = p->static_prio;
2651 * We don't need the reset flag anymore after the fork. It has
2652 * fulfilled its duty:
2654 p->sched_reset_on_fork = 0;
2658 * Make sure we do not leak PI boosting priority to the child.
2660 p->prio = current->normal_prio;
2662 if (!rt_prio(p->prio))
2663 p->sched_class = &fair_sched_class;
2665 if (p->sched_class->task_fork)
2666 p->sched_class->task_fork(p);
2668 set_task_cpu(p, cpu);
2670 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2671 if (likely(sched_info_on()))
2672 memset(&p->sched_info, 0, sizeof(p->sched_info));
2674 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2677 #ifdef CONFIG_PREEMPT
2678 /* Want to start with kernel preemption disabled. */
2679 task_thread_info(p)->preempt_count = 1;
2681 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2687 * wake_up_new_task - wake up a newly created task for the first time.
2689 * This function will do some initial scheduler statistics housekeeping
2690 * that must be done for every newly created context, then puts the task
2691 * on the runqueue and wakes it.
2693 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2695 unsigned long flags;
2697 int cpu = get_cpu();
2700 rq = task_rq_lock(p, &flags);
2701 p->state = TASK_WAKING;
2704 * Fork balancing, do it here and not earlier because:
2705 * - cpus_allowed can change in the fork path
2706 * - any previously selected cpu might disappear through hotplug
2708 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2709 * without people poking at ->cpus_allowed.
2711 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2712 set_task_cpu(p, cpu);
2714 p->state = TASK_RUNNING;
2715 task_rq_unlock(rq, &flags);
2718 rq = task_rq_lock(p, &flags);
2719 update_rq_clock(rq);
2720 activate_task(rq, p, 0);
2721 trace_sched_wakeup_new(rq, p, 1);
2722 check_preempt_curr(rq, p, WF_FORK);
2724 if (p->sched_class->task_woken)
2725 p->sched_class->task_woken(rq, p);
2727 task_rq_unlock(rq, &flags);
2731 #ifdef CONFIG_PREEMPT_NOTIFIERS
2734 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2735 * @notifier: notifier struct to register
2737 void preempt_notifier_register(struct preempt_notifier *notifier)
2739 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2741 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2744 * preempt_notifier_unregister - no longer interested in preemption notifications
2745 * @notifier: notifier struct to unregister
2747 * This is safe to call from within a preemption notifier.
2749 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2751 hlist_del(¬ifier->link);
2753 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2755 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2757 struct preempt_notifier *notifier;
2758 struct hlist_node *node;
2760 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2761 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2765 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2766 struct task_struct *next)
2768 struct preempt_notifier *notifier;
2769 struct hlist_node *node;
2771 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2772 notifier->ops->sched_out(notifier, next);
2775 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2777 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2782 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2783 struct task_struct *next)
2787 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2790 * prepare_task_switch - prepare to switch tasks
2791 * @rq: the runqueue preparing to switch
2792 * @prev: the current task that is being switched out
2793 * @next: the task we are going to switch to.
2795 * This is called with the rq lock held and interrupts off. It must
2796 * be paired with a subsequent finish_task_switch after the context
2799 * prepare_task_switch sets up locking and calls architecture specific
2803 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2804 struct task_struct *next)
2806 fire_sched_out_preempt_notifiers(prev, next);
2807 prepare_lock_switch(rq, next);
2808 prepare_arch_switch(next);
2812 * finish_task_switch - clean up after a task-switch
2813 * @rq: runqueue associated with task-switch
2814 * @prev: the thread we just switched away from.
2816 * finish_task_switch must be called after the context switch, paired
2817 * with a prepare_task_switch call before the context switch.
2818 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2819 * and do any other architecture-specific cleanup actions.
2821 * Note that we may have delayed dropping an mm in context_switch(). If
2822 * so, we finish that here outside of the runqueue lock. (Doing it
2823 * with the lock held can cause deadlocks; see schedule() for
2826 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2827 __releases(rq->lock)
2829 struct mm_struct *mm = rq->prev_mm;
2835 * A task struct has one reference for the use as "current".
2836 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2837 * schedule one last time. The schedule call will never return, and
2838 * the scheduled task must drop that reference.
2839 * The test for TASK_DEAD must occur while the runqueue locks are
2840 * still held, otherwise prev could be scheduled on another cpu, die
2841 * there before we look at prev->state, and then the reference would
2843 * Manfred Spraul <manfred@colorfullife.com>
2845 prev_state = prev->state;
2846 finish_arch_switch(prev);
2847 perf_event_task_sched_in(current, cpu_of(rq));
2848 finish_lock_switch(rq, prev);
2850 fire_sched_in_preempt_notifiers(current);
2853 if (unlikely(prev_state == TASK_DEAD)) {
2855 * Remove function-return probe instances associated with this
2856 * task and put them back on the free list.
2858 kprobe_flush_task(prev);
2859 put_task_struct(prev);
2865 /* assumes rq->lock is held */
2866 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2868 if (prev->sched_class->pre_schedule)
2869 prev->sched_class->pre_schedule(rq, prev);
2872 /* rq->lock is NOT held, but preemption is disabled */
2873 static inline void post_schedule(struct rq *rq)
2875 if (rq->post_schedule) {
2876 unsigned long flags;
2878 spin_lock_irqsave(&rq->lock, flags);
2879 if (rq->curr->sched_class->post_schedule)
2880 rq->curr->sched_class->post_schedule(rq);
2881 spin_unlock_irqrestore(&rq->lock, flags);
2883 rq->post_schedule = 0;
2889 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2893 static inline void post_schedule(struct rq *rq)
2900 * schedule_tail - first thing a freshly forked thread must call.
2901 * @prev: the thread we just switched away from.
2903 asmlinkage void schedule_tail(struct task_struct *prev)
2904 __releases(rq->lock)
2906 struct rq *rq = this_rq();
2908 finish_task_switch(rq, prev);
2911 * FIXME: do we need to worry about rq being invalidated by the
2916 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2917 /* In this case, finish_task_switch does not reenable preemption */
2920 if (current->set_child_tid)
2921 put_user(task_pid_vnr(current), current->set_child_tid);
2925 * context_switch - switch to the new MM and the new
2926 * thread's register state.
2929 context_switch(struct rq *rq, struct task_struct *prev,
2930 struct task_struct *next)
2932 struct mm_struct *mm, *oldmm;
2934 prepare_task_switch(rq, prev, next);
2935 trace_sched_switch(rq, prev, next);
2937 oldmm = prev->active_mm;
2939 * For paravirt, this is coupled with an exit in switch_to to
2940 * combine the page table reload and the switch backend into
2943 arch_start_context_switch(prev);
2945 if (unlikely(!mm)) {
2946 next->active_mm = oldmm;
2947 atomic_inc(&oldmm->mm_count);
2948 enter_lazy_tlb(oldmm, next);
2950 switch_mm(oldmm, mm, next);
2952 if (unlikely(!prev->mm)) {
2953 prev->active_mm = NULL;
2954 rq->prev_mm = oldmm;
2957 * Since the runqueue lock will be released by the next
2958 * task (which is an invalid locking op but in the case
2959 * of the scheduler it's an obvious special-case), so we
2960 * do an early lockdep release here:
2962 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2963 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2966 /* Here we just switch the register state and the stack. */
2967 switch_to(prev, next, prev);
2971 * this_rq must be evaluated again because prev may have moved
2972 * CPUs since it called schedule(), thus the 'rq' on its stack
2973 * frame will be invalid.
2975 finish_task_switch(this_rq(), prev);
2979 * nr_running, nr_uninterruptible and nr_context_switches:
2981 * externally visible scheduler statistics: current number of runnable
2982 * threads, current number of uninterruptible-sleeping threads, total
2983 * number of context switches performed since bootup.
2985 unsigned long nr_running(void)
2987 unsigned long i, sum = 0;
2989 for_each_online_cpu(i)
2990 sum += cpu_rq(i)->nr_running;
2995 unsigned long nr_uninterruptible(void)
2997 unsigned long i, sum = 0;
2999 for_each_possible_cpu(i)
3000 sum += cpu_rq(i)->nr_uninterruptible;
3003 * Since we read the counters lockless, it might be slightly
3004 * inaccurate. Do not allow it to go below zero though:
3006 if (unlikely((long)sum < 0))
3012 unsigned long long nr_context_switches(void)
3015 unsigned long long sum = 0;
3017 for_each_possible_cpu(i)
3018 sum += cpu_rq(i)->nr_switches;
3023 unsigned long nr_iowait(void)
3025 unsigned long i, sum = 0;
3027 for_each_possible_cpu(i)
3028 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3033 unsigned long nr_iowait_cpu(void)
3035 struct rq *this = this_rq();
3036 return atomic_read(&this->nr_iowait);
3039 unsigned long this_cpu_load(void)
3041 struct rq *this = this_rq();
3042 return this->cpu_load[0];
3046 /* Variables and functions for calc_load */
3047 static atomic_long_t calc_load_tasks;
3048 static unsigned long calc_load_update;
3049 unsigned long avenrun[3];
3050 EXPORT_SYMBOL(avenrun);
3053 * get_avenrun - get the load average array
3054 * @loads: pointer to dest load array
3055 * @offset: offset to add
3056 * @shift: shift count to shift the result left
3058 * These values are estimates at best, so no need for locking.
3060 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3062 loads[0] = (avenrun[0] + offset) << shift;
3063 loads[1] = (avenrun[1] + offset) << shift;
3064 loads[2] = (avenrun[2] + offset) << shift;
3067 static unsigned long
3068 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3071 load += active * (FIXED_1 - exp);
3072 return load >> FSHIFT;
3076 * calc_load - update the avenrun load estimates 10 ticks after the
3077 * CPUs have updated calc_load_tasks.
3079 void calc_global_load(void)
3081 unsigned long upd = calc_load_update + 10;
3084 if (time_before(jiffies, upd))
3087 active = atomic_long_read(&calc_load_tasks);
3088 active = active > 0 ? active * FIXED_1 : 0;
3090 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3091 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3092 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3094 calc_load_update += LOAD_FREQ;
3098 * Either called from update_cpu_load() or from a cpu going idle
3100 static void calc_load_account_active(struct rq *this_rq)
3102 long nr_active, delta;
3104 nr_active = this_rq->nr_running;
3105 nr_active += (long) this_rq->nr_uninterruptible;
3107 if (nr_active != this_rq->calc_load_active) {
3108 delta = nr_active - this_rq->calc_load_active;
3109 this_rq->calc_load_active = nr_active;
3110 atomic_long_add(delta, &calc_load_tasks);
3115 * Update rq->cpu_load[] statistics. This function is usually called every
3116 * scheduler tick (TICK_NSEC).
3118 static void update_cpu_load(struct rq *this_rq)
3120 unsigned long this_load = this_rq->load.weight;
3123 this_rq->nr_load_updates++;
3125 /* Update our load: */
3126 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3127 unsigned long old_load, new_load;
3129 /* scale is effectively 1 << i now, and >> i divides by scale */
3131 old_load = this_rq->cpu_load[i];
3132 new_load = this_load;
3134 * Round up the averaging division if load is increasing. This
3135 * prevents us from getting stuck on 9 if the load is 10, for
3138 if (new_load > old_load)
3139 new_load += scale-1;
3140 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3143 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3144 this_rq->calc_load_update += LOAD_FREQ;
3145 calc_load_account_active(this_rq);
3152 * double_rq_lock - safely lock two runqueues
3154 * Note this does not disable interrupts like task_rq_lock,
3155 * you need to do so manually before calling.
3157 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3158 __acquires(rq1->lock)
3159 __acquires(rq2->lock)
3161 BUG_ON(!irqs_disabled());
3163 spin_lock(&rq1->lock);
3164 __acquire(rq2->lock); /* Fake it out ;) */
3167 spin_lock(&rq1->lock);
3168 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3170 spin_lock(&rq2->lock);
3171 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3174 update_rq_clock(rq1);
3175 update_rq_clock(rq2);
3179 * double_rq_unlock - safely unlock two runqueues
3181 * Note this does not restore interrupts like task_rq_unlock,
3182 * you need to do so manually after calling.
3184 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3185 __releases(rq1->lock)
3186 __releases(rq2->lock)
3188 spin_unlock(&rq1->lock);
3190 spin_unlock(&rq2->lock);
3192 __release(rq2->lock);
3196 * sched_exec - execve() is a valuable balancing opportunity, because at
3197 * this point the task has the smallest effective memory and cache footprint.
3199 void sched_exec(void)
3201 struct task_struct *p = current;
3202 struct migration_req req;
3203 unsigned long flags;
3207 rq = task_rq_lock(p, &flags);
3208 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3209 if (dest_cpu == smp_processor_id())
3213 * select_task_rq() can race against ->cpus_allowed
3215 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3216 likely(cpu_active(dest_cpu)) &&
3217 migrate_task(p, dest_cpu, &req)) {
3218 /* Need to wait for migration thread (might exit: take ref). */
3219 struct task_struct *mt = rq->migration_thread;
3221 get_task_struct(mt);
3222 task_rq_unlock(rq, &flags);
3223 wake_up_process(mt);
3224 put_task_struct(mt);
3225 wait_for_completion(&req.done);
3230 task_rq_unlock(rq, &flags);
3234 * pull_task - move a task from a remote runqueue to the local runqueue.
3235 * Both runqueues must be locked.
3237 static void pull_task(struct rq *src_rq, struct task_struct *p,
3238 struct rq *this_rq, int this_cpu)
3240 deactivate_task(src_rq, p, 0);
3241 set_task_cpu(p, this_cpu);
3242 activate_task(this_rq, p, 0);
3243 check_preempt_curr(this_rq, p, 0);
3247 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3250 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3251 struct sched_domain *sd, enum cpu_idle_type idle,
3254 int tsk_cache_hot = 0;
3256 * We do not migrate tasks that are:
3257 * 1) running (obviously), or
3258 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3259 * 3) are cache-hot on their current CPU.
3261 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3262 schedstat_inc(p, se.nr_failed_migrations_affine);
3267 if (task_running(rq, p)) {
3268 schedstat_inc(p, se.nr_failed_migrations_running);
3273 * Aggressive migration if:
3274 * 1) task is cache cold, or
3275 * 2) too many balance attempts have failed.
3278 tsk_cache_hot = task_hot(p, rq->clock, sd);
3279 if (!tsk_cache_hot ||
3280 sd->nr_balance_failed > sd->cache_nice_tries) {
3281 #ifdef CONFIG_SCHEDSTATS
3282 if (tsk_cache_hot) {
3283 schedstat_inc(sd, lb_hot_gained[idle]);
3284 schedstat_inc(p, se.nr_forced_migrations);
3290 if (tsk_cache_hot) {
3291 schedstat_inc(p, se.nr_failed_migrations_hot);
3297 static unsigned long
3298 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3299 unsigned long max_load_move, struct sched_domain *sd,
3300 enum cpu_idle_type idle, int *all_pinned,
3301 int *this_best_prio, struct rq_iterator *iterator)
3303 int loops = 0, pulled = 0, pinned = 0;
3304 struct task_struct *p;
3305 long rem_load_move = max_load_move;
3307 if (max_load_move == 0)
3313 * Start the load-balancing iterator:
3315 p = iterator->start(iterator->arg);
3317 if (!p || loops++ > sysctl_sched_nr_migrate)
3320 if ((p->se.load.weight >> 1) > rem_load_move ||
3321 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3322 p = iterator->next(iterator->arg);
3326 pull_task(busiest, p, this_rq, this_cpu);
3328 rem_load_move -= p->se.load.weight;
3330 #ifdef CONFIG_PREEMPT
3332 * NEWIDLE balancing is a source of latency, so preemptible kernels
3333 * will stop after the first task is pulled to minimize the critical
3336 if (idle == CPU_NEWLY_IDLE)
3341 * We only want to steal up to the prescribed amount of weighted load.
3343 if (rem_load_move > 0) {
3344 if (p->prio < *this_best_prio)
3345 *this_best_prio = p->prio;
3346 p = iterator->next(iterator->arg);
3351 * Right now, this is one of only two places pull_task() is called,
3352 * so we can safely collect pull_task() stats here rather than
3353 * inside pull_task().
3355 schedstat_add(sd, lb_gained[idle], pulled);
3358 *all_pinned = pinned;
3360 return max_load_move - rem_load_move;
3364 * move_tasks tries to move up to max_load_move weighted load from busiest to
3365 * this_rq, as part of a balancing operation within domain "sd".
3366 * Returns 1 if successful and 0 otherwise.
3368 * Called with both runqueues locked.
3370 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3371 unsigned long max_load_move,
3372 struct sched_domain *sd, enum cpu_idle_type idle,
3375 const struct sched_class *class = sched_class_highest;
3376 unsigned long total_load_moved = 0;
3377 int this_best_prio = this_rq->curr->prio;
3381 class->load_balance(this_rq, this_cpu, busiest,
3382 max_load_move - total_load_moved,
3383 sd, idle, all_pinned, &this_best_prio);
3384 class = class->next;
3386 #ifdef CONFIG_PREEMPT
3388 * NEWIDLE balancing is a source of latency, so preemptible
3389 * kernels will stop after the first task is pulled to minimize
3390 * the critical section.
3392 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3395 } while (class && max_load_move > total_load_moved);
3397 return total_load_moved > 0;
3401 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3402 struct sched_domain *sd, enum cpu_idle_type idle,
3403 struct rq_iterator *iterator)
3405 struct task_struct *p = iterator->start(iterator->arg);
3409 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3410 pull_task(busiest, p, this_rq, this_cpu);
3412 * Right now, this is only the second place pull_task()
3413 * is called, so we can safely collect pull_task()
3414 * stats here rather than inside pull_task().
3416 schedstat_inc(sd, lb_gained[idle]);
3420 p = iterator->next(iterator->arg);
3427 * move_one_task tries to move exactly one task from busiest to this_rq, as
3428 * part of active balancing operations within "domain".
3429 * Returns 1 if successful and 0 otherwise.
3431 * Called with both runqueues locked.
3433 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3434 struct sched_domain *sd, enum cpu_idle_type idle)
3436 const struct sched_class *class;
3438 for_each_class(class) {
3439 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3445 /********** Helpers for find_busiest_group ************************/
3447 * sd_lb_stats - Structure to store the statistics of a sched_domain
3448 * during load balancing.
3450 struct sd_lb_stats {
3451 struct sched_group *busiest; /* Busiest group in this sd */
3452 struct sched_group *this; /* Local group in this sd */
3453 unsigned long total_load; /* Total load of all groups in sd */
3454 unsigned long total_pwr; /* Total power of all groups in sd */
3455 unsigned long avg_load; /* Average load across all groups in sd */
3457 /** Statistics of this group */
3458 unsigned long this_load;
3459 unsigned long this_load_per_task;
3460 unsigned long this_nr_running;
3462 /* Statistics of the busiest group */
3463 unsigned long max_load;
3464 unsigned long busiest_load_per_task;
3465 unsigned long busiest_nr_running;
3466 unsigned long busiest_group_capacity;
3468 int group_imb; /* Is there imbalance in this sd */
3469 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3470 int power_savings_balance; /* Is powersave balance needed for this sd */
3471 struct sched_group *group_min; /* Least loaded group in sd */
3472 struct sched_group *group_leader; /* Group which relieves group_min */
3473 unsigned long min_load_per_task; /* load_per_task in group_min */
3474 unsigned long leader_nr_running; /* Nr running of group_leader */
3475 unsigned long min_nr_running; /* Nr running of group_min */
3480 * sg_lb_stats - stats of a sched_group required for load_balancing
3482 struct sg_lb_stats {
3483 unsigned long avg_load; /*Avg load across the CPUs of the group */
3484 unsigned long group_load; /* Total load over the CPUs of the group */
3485 unsigned long sum_nr_running; /* Nr tasks running in the group */
3486 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3487 unsigned long group_capacity;
3488 int group_imb; /* Is there an imbalance in the group ? */
3492 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3493 * @group: The group whose first cpu is to be returned.
3495 static inline unsigned int group_first_cpu(struct sched_group *group)
3497 return cpumask_first(sched_group_cpus(group));
3501 * get_sd_load_idx - Obtain the load index for a given sched domain.
3502 * @sd: The sched_domain whose load_idx is to be obtained.
3503 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3505 static inline int get_sd_load_idx(struct sched_domain *sd,
3506 enum cpu_idle_type idle)
3512 load_idx = sd->busy_idx;
3515 case CPU_NEWLY_IDLE:
3516 load_idx = sd->newidle_idx;
3519 load_idx = sd->idle_idx;
3527 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3529 * init_sd_power_savings_stats - Initialize power savings statistics for
3530 * the given sched_domain, during load balancing.
3532 * @sd: Sched domain whose power-savings statistics are to be initialized.
3533 * @sds: Variable containing the statistics for sd.
3534 * @idle: Idle status of the CPU at which we're performing load-balancing.
3536 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3537 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3540 * Busy processors will not participate in power savings
3543 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3544 sds->power_savings_balance = 0;
3546 sds->power_savings_balance = 1;
3547 sds->min_nr_running = ULONG_MAX;
3548 sds->leader_nr_running = 0;
3553 * update_sd_power_savings_stats - Update the power saving stats for a
3554 * sched_domain while performing load balancing.
3556 * @group: sched_group belonging to the sched_domain under consideration.
3557 * @sds: Variable containing the statistics of the sched_domain
3558 * @local_group: Does group contain the CPU for which we're performing
3560 * @sgs: Variable containing the statistics of the group.
3562 static inline void update_sd_power_savings_stats(struct sched_group *group,
3563 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3566 if (!sds->power_savings_balance)
3570 * If the local group is idle or completely loaded
3571 * no need to do power savings balance at this domain
3573 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3574 !sds->this_nr_running))
3575 sds->power_savings_balance = 0;
3578 * If a group is already running at full capacity or idle,
3579 * don't include that group in power savings calculations
3581 if (!sds->power_savings_balance ||
3582 sgs->sum_nr_running >= sgs->group_capacity ||
3583 !sgs->sum_nr_running)
3587 * Calculate the group which has the least non-idle load.
3588 * This is the group from where we need to pick up the load
3591 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3592 (sgs->sum_nr_running == sds->min_nr_running &&
3593 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3594 sds->group_min = group;
3595 sds->min_nr_running = sgs->sum_nr_running;
3596 sds->min_load_per_task = sgs->sum_weighted_load /
3597 sgs->sum_nr_running;
3601 * Calculate the group which is almost near its
3602 * capacity but still has some space to pick up some load
3603 * from other group and save more power
3605 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3608 if (sgs->sum_nr_running > sds->leader_nr_running ||
3609 (sgs->sum_nr_running == sds->leader_nr_running &&
3610 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3611 sds->group_leader = group;
3612 sds->leader_nr_running = sgs->sum_nr_running;
3617 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3618 * @sds: Variable containing the statistics of the sched_domain
3619 * under consideration.
3620 * @this_cpu: Cpu at which we're currently performing load-balancing.
3621 * @imbalance: Variable to store the imbalance.
3624 * Check if we have potential to perform some power-savings balance.
3625 * If yes, set the busiest group to be the least loaded group in the
3626 * sched_domain, so that it's CPUs can be put to idle.
3628 * Returns 1 if there is potential to perform power-savings balance.
3631 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3632 int this_cpu, unsigned long *imbalance)
3634 if (!sds->power_savings_balance)
3637 if (sds->this != sds->group_leader ||
3638 sds->group_leader == sds->group_min)
3641 *imbalance = sds->min_load_per_task;
3642 sds->busiest = sds->group_min;
3647 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3648 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3649 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3654 static inline void update_sd_power_savings_stats(struct sched_group *group,
3655 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3660 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3661 int this_cpu, unsigned long *imbalance)
3665 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3668 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3670 return SCHED_LOAD_SCALE;
3673 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3675 return default_scale_freq_power(sd, cpu);
3678 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3680 unsigned long weight = sd->span_weight;
3681 unsigned long smt_gain = sd->smt_gain;
3688 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3690 return default_scale_smt_power(sd, cpu);
3693 unsigned long scale_rt_power(int cpu)
3695 struct rq *rq = cpu_rq(cpu);
3696 u64 total, available;
3698 sched_avg_update(rq);
3700 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3701 available = total - rq->rt_avg;
3703 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3704 total = SCHED_LOAD_SCALE;
3706 total >>= SCHED_LOAD_SHIFT;
3708 return div_u64(available, total);
3711 static void update_cpu_power(struct sched_domain *sd, int cpu)
3713 unsigned long weight = sd->span_weight;
3714 unsigned long power = SCHED_LOAD_SCALE;
3715 struct sched_group *sdg = sd->groups;
3717 if (sched_feat(ARCH_POWER))
3718 power *= arch_scale_freq_power(sd, cpu);
3720 power *= default_scale_freq_power(sd, cpu);
3722 power >>= SCHED_LOAD_SHIFT;
3724 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3725 if (sched_feat(ARCH_POWER))
3726 power *= arch_scale_smt_power(sd, cpu);
3728 power *= default_scale_smt_power(sd, cpu);
3730 power >>= SCHED_LOAD_SHIFT;
3733 power *= scale_rt_power(cpu);
3734 power >>= SCHED_LOAD_SHIFT;
3739 sdg->cpu_power = power;
3742 static void update_group_power(struct sched_domain *sd, int cpu)
3744 struct sched_domain *child = sd->child;
3745 struct sched_group *group, *sdg = sd->groups;
3746 unsigned long power;
3749 update_cpu_power(sd, cpu);
3755 group = child->groups;
3757 power += group->cpu_power;
3758 group = group->next;
3759 } while (group != child->groups);
3761 sdg->cpu_power = power;
3765 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3766 * @sd: The sched_domain whose statistics are to be updated.
3767 * @group: sched_group whose statistics are to be updated.
3768 * @this_cpu: Cpu for which load balance is currently performed.
3769 * @idle: Idle status of this_cpu
3770 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3771 * @sd_idle: Idle status of the sched_domain containing group.
3772 * @local_group: Does group contain this_cpu.
3773 * @cpus: Set of cpus considered for load balancing.
3774 * @balance: Should we balance.
3775 * @sgs: variable to hold the statistics for this group.
3777 static inline void update_sg_lb_stats(struct sched_domain *sd,
3778 struct sched_group *group, int this_cpu,
3779 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3780 int local_group, const struct cpumask *cpus,
3781 int *balance, struct sg_lb_stats *sgs)
3783 unsigned long load, max_cpu_load, min_cpu_load;
3785 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3786 unsigned long avg_load_per_task = 0;
3789 balance_cpu = group_first_cpu(group);
3790 if (balance_cpu == this_cpu)
3791 update_group_power(sd, this_cpu);
3794 /* Tally up the load of all CPUs in the group */
3796 min_cpu_load = ~0UL;
3798 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3799 struct rq *rq = cpu_rq(i);
3801 if (*sd_idle && rq->nr_running)
3804 /* Bias balancing toward cpus of our domain */
3806 if (idle_cpu(i) && !first_idle_cpu) {
3811 load = target_load(i, load_idx);
3813 load = source_load(i, load_idx);
3814 if (load > max_cpu_load)
3815 max_cpu_load = load;
3816 if (min_cpu_load > load)
3817 min_cpu_load = load;
3820 sgs->group_load += load;
3821 sgs->sum_nr_running += rq->nr_running;
3822 sgs->sum_weighted_load += weighted_cpuload(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;
3842 * Consider the group unbalanced when the imbalance is larger
3843 * than the average weight of two tasks.
3845 * APZ: with cgroup the avg task weight can vary wildly and
3846 * might not be a suitable number - should we keep a
3847 * normalized nr_running number somewhere that negates
3850 if (sgs->sum_nr_running)
3851 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3853 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3856 sgs->group_capacity =
3857 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3861 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3862 * @sd: sched_domain whose statistics are to be updated.
3863 * @this_cpu: Cpu for which load balance is currently performed.
3864 * @idle: Idle status of this_cpu
3865 * @sd_idle: Idle status of the sched_domain containing group.
3866 * @cpus: Set of cpus considered for load balancing.
3867 * @balance: Should we balance.
3868 * @sds: variable to hold the statistics for this sched_domain.
3870 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3871 enum cpu_idle_type idle, int *sd_idle,
3872 const struct cpumask *cpus, int *balance,
3873 struct sd_lb_stats *sds)
3875 struct sched_domain *child = sd->child;
3876 struct sched_group *group = sd->groups;
3877 struct sg_lb_stats sgs;
3878 int load_idx, prefer_sibling = 0;
3880 if (child && child->flags & SD_PREFER_SIBLING)
3883 init_sd_power_savings_stats(sd, sds, idle);
3884 load_idx = get_sd_load_idx(sd, idle);
3889 local_group = cpumask_test_cpu(this_cpu,
3890 sched_group_cpus(group));
3891 memset(&sgs, 0, sizeof(sgs));
3892 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3893 local_group, cpus, balance, &sgs);
3895 if (local_group && balance && !(*balance))
3898 sds->total_load += sgs.group_load;
3899 sds->total_pwr += group->cpu_power;
3902 * In case the child domain prefers tasks go to siblings
3903 * first, lower the group capacity to one so that we'll try
3904 * and move all the excess tasks away.
3907 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3910 sds->this_load = sgs.avg_load;
3912 sds->this_nr_running = sgs.sum_nr_running;
3913 sds->this_load_per_task = sgs.sum_weighted_load;
3914 } else if (sgs.avg_load > sds->max_load &&
3915 (sgs.sum_nr_running > sgs.group_capacity ||
3917 sds->max_load = sgs.avg_load;
3918 sds->busiest = group;
3919 sds->busiest_nr_running = sgs.sum_nr_running;
3920 sds->busiest_group_capacity = sgs.group_capacity;
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;
3943 unsigned long scaled_busy_load_per_task;
3945 if (sds->this_nr_running) {
3946 sds->this_load_per_task /= sds->this_nr_running;
3947 if (sds->busiest_load_per_task >
3948 sds->this_load_per_task)
3951 sds->this_load_per_task =
3952 cpu_avg_load_per_task(this_cpu);
3954 scaled_busy_load_per_task = sds->busiest_load_per_task
3956 scaled_busy_load_per_task /= sds->busiest->cpu_power;
3958 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
3959 (scaled_busy_load_per_task * imbn)) {
3960 *imbalance = sds->busiest_load_per_task;
3965 * OK, we don't have enough imbalance to justify moving tasks,
3966 * however we may be able to increase total CPU power used by
3970 pwr_now += sds->busiest->cpu_power *
3971 min(sds->busiest_load_per_task, sds->max_load);
3972 pwr_now += sds->this->cpu_power *
3973 min(sds->this_load_per_task, sds->this_load);
3974 pwr_now /= SCHED_LOAD_SCALE;
3976 /* Amount of load we'd subtract */
3977 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3978 sds->busiest->cpu_power;
3979 if (sds->max_load > tmp)
3980 pwr_move += sds->busiest->cpu_power *
3981 min(sds->busiest_load_per_task, sds->max_load - tmp);
3983 /* Amount of load we'd add */
3984 if (sds->max_load * sds->busiest->cpu_power <
3985 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3986 tmp = (sds->max_load * sds->busiest->cpu_power) /
3987 sds->this->cpu_power;
3989 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3990 sds->this->cpu_power;
3991 pwr_move += sds->this->cpu_power *
3992 min(sds->this_load_per_task, sds->this_load + tmp);
3993 pwr_move /= SCHED_LOAD_SCALE;
3995 /* Move if we gain throughput */
3996 if (pwr_move > pwr_now)
3997 *imbalance = sds->busiest_load_per_task;
4001 * calculate_imbalance - Calculate the amount of imbalance present within the
4002 * groups of a given sched_domain during load balance.
4003 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4004 * @this_cpu: Cpu for which currently load balance is being performed.
4005 * @imbalance: The variable to store the imbalance.
4007 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
4008 unsigned long *imbalance)
4010 unsigned long max_pull, load_above_capacity = ~0UL;
4012 sds->busiest_load_per_task /= sds->busiest_nr_running;
4013 if (sds->group_imb) {
4014 sds->busiest_load_per_task =
4015 min(sds->busiest_load_per_task, sds->avg_load);
4019 * In the presence of smp nice balancing, certain scenarios can have
4020 * max load less than avg load(as we skip the groups at or below
4021 * its cpu_power, while calculating max_load..)
4023 if (sds->max_load < sds->avg_load) {
4025 return fix_small_imbalance(sds, this_cpu, imbalance);
4028 if (!sds->group_imb) {
4030 * Don't want to pull so many tasks that a group would go idle.
4032 load_above_capacity = (sds->busiest_nr_running -
4033 sds->busiest_group_capacity);
4035 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE);
4037 load_above_capacity /= sds->busiest->cpu_power;
4041 * We're trying to get all the cpus to the average_load, so we don't
4042 * want to push ourselves above the average load, nor do we wish to
4043 * reduce the max loaded cpu below the average load. At the same time,
4044 * we also don't want to reduce the group load below the group capacity
4045 * (so that we can implement power-savings policies etc). Thus we look
4046 * for the minimum possible imbalance.
4047 * Be careful of negative numbers as they'll appear as very large values
4048 * with unsigned longs.
4050 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4052 /* How much load to actually move to equalise the imbalance */
4053 *imbalance = min(max_pull * sds->busiest->cpu_power,
4054 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
4058 * if *imbalance is less than the average load per runnable task
4059 * there is no gaurantee that any tasks will be moved so we'll have
4060 * a think about bumping its value to force at least one task to be
4063 if (*imbalance < sds->busiest_load_per_task)
4064 return fix_small_imbalance(sds, this_cpu, imbalance);
4067 /******* find_busiest_group() helpers end here *********************/
4070 * find_busiest_group - Returns the busiest group within the sched_domain
4071 * if there is an imbalance. If there isn't an imbalance, and
4072 * the user has opted for power-savings, it returns a group whose
4073 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4074 * such a group exists.
4076 * Also calculates the amount of weighted load which should be moved
4077 * to restore balance.
4079 * @sd: The sched_domain whose busiest group is to be returned.
4080 * @this_cpu: The cpu for which load balancing is currently being performed.
4081 * @imbalance: Variable which stores amount of weighted load which should
4082 * be moved to restore balance/put a group to idle.
4083 * @idle: The idle status of this_cpu.
4084 * @sd_idle: The idleness of sd
4085 * @cpus: The set of CPUs under consideration for load-balancing.
4086 * @balance: Pointer to a variable indicating if this_cpu
4087 * is the appropriate cpu to perform load balancing at this_level.
4089 * Returns: - the busiest group if imbalance exists.
4090 * - If no imbalance and user has opted for power-savings balance,
4091 * return the least loaded group whose CPUs can be
4092 * put to idle by rebalancing its tasks onto our group.
4094 static struct sched_group *
4095 find_busiest_group(struct sched_domain *sd, int this_cpu,
4096 unsigned long *imbalance, enum cpu_idle_type idle,
4097 int *sd_idle, const struct cpumask *cpus, int *balance)
4099 struct sd_lb_stats sds;
4101 memset(&sds, 0, sizeof(sds));
4104 * Compute the various statistics relavent for load balancing at
4107 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4110 /* Cases where imbalance does not exist from POV of this_cpu */
4111 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4113 * 2) There is no busy sibling group to pull from.
4114 * 3) This group is the busiest group.
4115 * 4) This group is more busy than the avg busieness at this
4117 * 5) The imbalance is within the specified limit.
4119 if (balance && !(*balance))
4122 if (!sds.busiest || sds.busiest_nr_running == 0)
4125 if (sds.this_load >= sds.max_load)
4128 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4130 if (sds.this_load >= sds.avg_load)
4133 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4136 /* Looks like there is an imbalance. Compute it */
4137 calculate_imbalance(&sds, this_cpu, imbalance);
4142 * There is no obvious imbalance. But check if we can do some balancing
4145 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4153 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4156 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4157 unsigned long imbalance, const struct cpumask *cpus)
4159 struct rq *busiest = NULL, *rq;
4160 unsigned long max_load = 0;
4163 for_each_cpu(i, sched_group_cpus(group)) {
4164 unsigned long power = power_of(i);
4165 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4168 if (!cpumask_test_cpu(i, cpus))
4172 wl = weighted_cpuload(i);
4175 * When comparing with imbalance, use weighted_cpuload()
4176 * which is not scaled with the cpu power.
4178 if (capacity && rq->nr_running == 1 && wl > imbalance)
4182 * For the load comparisons with the other cpu's, consider
4183 * the weighted_cpuload() scaled with the cpu power, so that
4184 * the load can be moved away from the cpu that is potentially
4185 * running at a lower capacity.
4187 wl = (wl * SCHED_LOAD_SCALE) / power;
4189 if (wl > max_load) {
4199 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4200 * so long as it is large enough.
4202 #define MAX_PINNED_INTERVAL 512
4204 /* Working cpumask for load_balance and load_balance_newidle. */
4205 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4208 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4209 * tasks if there is an imbalance.
4211 static int load_balance(int this_cpu, struct rq *this_rq,
4212 struct sched_domain *sd, enum cpu_idle_type idle,
4215 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4216 struct sched_group *group;
4217 unsigned long imbalance;
4219 unsigned long flags;
4220 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4222 cpumask_copy(cpus, cpu_active_mask);
4225 * When power savings policy is enabled for the parent domain, idle
4226 * sibling can pick up load irrespective of busy siblings. In this case,
4227 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4228 * portraying it as CPU_NOT_IDLE.
4230 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4231 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4234 schedstat_inc(sd, lb_count[idle]);
4238 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4245 schedstat_inc(sd, lb_nobusyg[idle]);
4249 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4251 schedstat_inc(sd, lb_nobusyq[idle]);
4255 BUG_ON(busiest == this_rq);
4257 schedstat_add(sd, lb_imbalance[idle], imbalance);
4260 if (busiest->nr_running > 1) {
4262 * Attempt to move tasks. If find_busiest_group has found
4263 * an imbalance but busiest->nr_running <= 1, the group is
4264 * still unbalanced. ld_moved simply stays zero, so it is
4265 * correctly treated as an imbalance.
4267 local_irq_save(flags);
4268 double_rq_lock(this_rq, busiest);
4269 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4270 imbalance, sd, idle, &all_pinned);
4271 double_rq_unlock(this_rq, busiest);
4272 local_irq_restore(flags);
4275 * some other cpu did the load balance for us.
4277 if (ld_moved && this_cpu != smp_processor_id())
4278 resched_cpu(this_cpu);
4280 /* All tasks on this runqueue were pinned by CPU affinity */
4281 if (unlikely(all_pinned)) {
4282 cpumask_clear_cpu(cpu_of(busiest), cpus);
4283 if (!cpumask_empty(cpus))
4290 schedstat_inc(sd, lb_failed[idle]);
4291 sd->nr_balance_failed++;
4293 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4295 spin_lock_irqsave(&busiest->lock, flags);
4297 /* don't kick the migration_thread, if the curr
4298 * task on busiest cpu can't be moved to this_cpu
4300 if (!cpumask_test_cpu(this_cpu,
4301 &busiest->curr->cpus_allowed)) {
4302 spin_unlock_irqrestore(&busiest->lock, flags);
4304 goto out_one_pinned;
4307 if (!busiest->active_balance) {
4308 busiest->active_balance = 1;
4309 busiest->push_cpu = this_cpu;
4312 spin_unlock_irqrestore(&busiest->lock, flags);
4314 wake_up_process(busiest->migration_thread);
4317 * We've kicked active balancing, reset the failure
4320 sd->nr_balance_failed = sd->cache_nice_tries+1;
4323 sd->nr_balance_failed = 0;
4325 if (likely(!active_balance)) {
4326 /* We were unbalanced, so reset the balancing interval */
4327 sd->balance_interval = sd->min_interval;
4330 * If we've begun active balancing, start to back off. This
4331 * case may not be covered by the all_pinned logic if there
4332 * is only 1 task on the busy runqueue (because we don't call
4335 if (sd->balance_interval < sd->max_interval)
4336 sd->balance_interval *= 2;
4339 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4340 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4346 schedstat_inc(sd, lb_balanced[idle]);
4348 sd->nr_balance_failed = 0;
4351 /* tune up the balancing interval */
4352 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4353 (sd->balance_interval < sd->max_interval))
4354 sd->balance_interval *= 2;
4356 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4357 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4368 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4369 * tasks if there is an imbalance.
4371 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4372 * this_rq is locked.
4375 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4377 struct sched_group *group;
4378 struct rq *busiest = NULL;
4379 unsigned long imbalance;
4383 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4385 cpumask_copy(cpus, cpu_active_mask);
4388 * When power savings policy is enabled for the parent domain, idle
4389 * sibling can pick up load irrespective of busy siblings. In this case,
4390 * let the state of idle sibling percolate up as IDLE, instead of
4391 * portraying it as CPU_NOT_IDLE.
4393 if (sd->flags & SD_SHARE_CPUPOWER &&
4394 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4397 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4399 update_shares_locked(this_rq, sd);
4400 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4401 &sd_idle, cpus, NULL);
4403 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4407 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4409 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4413 BUG_ON(busiest == this_rq);
4415 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4418 if (busiest->nr_running > 1) {
4419 /* Attempt to move tasks */
4420 double_lock_balance(this_rq, busiest);
4421 /* this_rq->clock is already updated */
4422 update_rq_clock(busiest);
4423 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4424 imbalance, sd, CPU_NEWLY_IDLE,
4426 double_unlock_balance(this_rq, busiest);
4428 if (unlikely(all_pinned)) {
4429 cpumask_clear_cpu(cpu_of(busiest), cpus);
4430 if (!cpumask_empty(cpus))
4436 int active_balance = 0;
4438 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4439 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4440 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4443 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4446 if (sd->nr_balance_failed++ < 2)
4450 * The only task running in a non-idle cpu can be moved to this
4451 * cpu in an attempt to completely freeup the other CPU
4452 * package. The same method used to move task in load_balance()
4453 * have been extended for load_balance_newidle() to speedup
4454 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4456 * The package power saving logic comes from
4457 * find_busiest_group(). If there are no imbalance, then
4458 * f_b_g() will return NULL. However when sched_mc={1,2} then
4459 * f_b_g() will select a group from which a running task may be
4460 * pulled to this cpu in order to make the other package idle.
4461 * If there is no opportunity to make a package idle and if
4462 * there are no imbalance, then f_b_g() will return NULL and no
4463 * action will be taken in load_balance_newidle().
4465 * Under normal task pull operation due to imbalance, there
4466 * will be more than one task in the source run queue and
4467 * move_tasks() will succeed. ld_moved will be true and this
4468 * active balance code will not be triggered.
4471 /* Lock busiest in correct order while this_rq is held */
4472 double_lock_balance(this_rq, busiest);
4475 * don't kick the migration_thread, if the curr
4476 * task on busiest cpu can't be moved to this_cpu
4478 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4479 double_unlock_balance(this_rq, busiest);
4484 if (!busiest->active_balance) {
4485 busiest->active_balance = 1;
4486 busiest->push_cpu = this_cpu;
4490 double_unlock_balance(this_rq, busiest);
4492 * Should not call ttwu while holding a rq->lock
4494 spin_unlock(&this_rq->lock);
4496 wake_up_process(busiest->migration_thread);
4497 spin_lock(&this_rq->lock);
4500 sd->nr_balance_failed = 0;
4502 update_shares_locked(this_rq, sd);
4506 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4507 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4508 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4510 sd->nr_balance_failed = 0;
4516 * idle_balance is called by schedule() if this_cpu is about to become
4517 * idle. Attempts to pull tasks from other CPUs.
4519 static void idle_balance(int this_cpu, struct rq *this_rq)
4521 struct sched_domain *sd;
4522 int pulled_task = 0;
4523 unsigned long next_balance = jiffies + HZ;
4525 this_rq->idle_stamp = this_rq->clock;
4527 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4530 for_each_domain(this_cpu, sd) {
4531 unsigned long interval;
4533 if (!(sd->flags & SD_LOAD_BALANCE))
4536 if (sd->flags & SD_BALANCE_NEWIDLE)
4537 /* If we've pulled tasks over stop searching: */
4538 pulled_task = load_balance_newidle(this_cpu, this_rq,
4541 interval = msecs_to_jiffies(sd->balance_interval);
4542 if (time_after(next_balance, sd->last_balance + interval))
4543 next_balance = sd->last_balance + interval;
4545 this_rq->idle_stamp = 0;
4549 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4551 * We are going idle. next_balance may be set based on
4552 * a busy processor. So reset next_balance.
4554 this_rq->next_balance = next_balance;
4559 * active_load_balance is run by migration threads. It pushes running tasks
4560 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4561 * running on each physical CPU where possible, and avoids physical /
4562 * logical imbalances.
4564 * Called with busiest_rq locked.
4566 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4568 int target_cpu = busiest_rq->push_cpu;
4569 struct sched_domain *sd;
4570 struct rq *target_rq;
4572 /* Is there any task to move? */
4573 if (busiest_rq->nr_running <= 1)
4576 target_rq = cpu_rq(target_cpu);
4579 * This condition is "impossible", if it occurs
4580 * we need to fix it. Originally reported by
4581 * Bjorn Helgaas on a 128-cpu setup.
4583 BUG_ON(busiest_rq == target_rq);
4585 /* move a task from busiest_rq to target_rq */
4586 double_lock_balance(busiest_rq, target_rq);
4587 update_rq_clock(busiest_rq);
4588 update_rq_clock(target_rq);
4590 /* Search for an sd spanning us and the target CPU. */
4591 for_each_domain(target_cpu, sd) {
4592 if ((sd->flags & SD_LOAD_BALANCE) &&
4593 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4598 schedstat_inc(sd, alb_count);
4600 if (move_one_task(target_rq, target_cpu, busiest_rq,
4602 schedstat_inc(sd, alb_pushed);
4604 schedstat_inc(sd, alb_failed);
4606 double_unlock_balance(busiest_rq, target_rq);
4611 atomic_t load_balancer;
4612 cpumask_var_t cpu_mask;
4613 cpumask_var_t ilb_grp_nohz_mask;
4614 } nohz ____cacheline_aligned = {
4615 .load_balancer = ATOMIC_INIT(-1),
4618 int get_nohz_load_balancer(void)
4620 return atomic_read(&nohz.load_balancer);
4623 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4625 * lowest_flag_domain - Return lowest sched_domain containing flag.
4626 * @cpu: The cpu whose lowest level of sched domain is to
4628 * @flag: The flag to check for the lowest sched_domain
4629 * for the given cpu.
4631 * Returns the lowest sched_domain of a cpu which contains the given flag.
4633 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4635 struct sched_domain *sd;
4637 for_each_domain(cpu, sd)
4638 if (sd && (sd->flags & flag))
4645 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4646 * @cpu: The cpu whose domains we're iterating over.
4647 * @sd: variable holding the value of the power_savings_sd
4649 * @flag: The flag to filter the sched_domains to be iterated.
4651 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4652 * set, starting from the lowest sched_domain to the highest.
4654 #define for_each_flag_domain(cpu, sd, flag) \
4655 for (sd = lowest_flag_domain(cpu, flag); \
4656 (sd && (sd->flags & flag)); sd = sd->parent)
4659 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4660 * @ilb_group: group to be checked for semi-idleness
4662 * Returns: 1 if the group is semi-idle. 0 otherwise.
4664 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4665 * and atleast one non-idle CPU. This helper function checks if the given
4666 * sched_group is semi-idle or not.
4668 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4670 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4671 sched_group_cpus(ilb_group));
4674 * A sched_group is semi-idle when it has atleast one busy cpu
4675 * and atleast one idle cpu.
4677 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4680 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4686 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4687 * @cpu: The cpu which is nominating a new idle_load_balancer.
4689 * Returns: Returns the id of the idle load balancer if it exists,
4690 * Else, returns >= nr_cpu_ids.
4692 * This algorithm picks the idle load balancer such that it belongs to a
4693 * semi-idle powersavings sched_domain. The idea is to try and avoid
4694 * completely idle packages/cores just for the purpose of idle load balancing
4695 * when there are other idle cpu's which are better suited for that job.
4697 static int find_new_ilb(int cpu)
4699 struct sched_domain *sd;
4700 struct sched_group *ilb_group;
4703 * Have idle load balancer selection from semi-idle packages only
4704 * when power-aware load balancing is enabled
4706 if (!(sched_smt_power_savings || sched_mc_power_savings))
4710 * Optimize for the case when we have no idle CPUs or only one
4711 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4713 if (cpumask_weight(nohz.cpu_mask) < 2)
4716 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4717 ilb_group = sd->groups;
4720 if (is_semi_idle_group(ilb_group))
4721 return cpumask_first(nohz.ilb_grp_nohz_mask);
4723 ilb_group = ilb_group->next;
4725 } while (ilb_group != sd->groups);
4729 return cpumask_first(nohz.cpu_mask);
4731 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4732 static inline int find_new_ilb(int call_cpu)
4734 return cpumask_first(nohz.cpu_mask);
4739 * This routine will try to nominate the ilb (idle load balancing)
4740 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4741 * load balancing on behalf of all those cpus. If all the cpus in the system
4742 * go into this tickless mode, then there will be no ilb owner (as there is
4743 * no need for one) and all the cpus will sleep till the next wakeup event
4746 * For the ilb owner, tick is not stopped. And this tick will be used
4747 * for idle load balancing. ilb owner will still be part of
4750 * While stopping the tick, this cpu will become the ilb owner if there
4751 * is no other owner. And will be the owner till that cpu becomes busy
4752 * or if all cpus in the system stop their ticks at which point
4753 * there is no need for ilb owner.
4755 * When the ilb owner becomes busy, it nominates another owner, during the
4756 * next busy scheduler_tick()
4758 int select_nohz_load_balancer(int stop_tick)
4760 int cpu = smp_processor_id();
4763 cpu_rq(cpu)->in_nohz_recently = 1;
4765 if (!cpu_active(cpu)) {
4766 if (atomic_read(&nohz.load_balancer) != cpu)
4770 * If we are going offline and still the leader,
4773 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4779 cpumask_set_cpu(cpu, nohz.cpu_mask);
4781 /* time for ilb owner also to sleep */
4782 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4783 if (atomic_read(&nohz.load_balancer) == cpu)
4784 atomic_set(&nohz.load_balancer, -1);
4788 if (atomic_read(&nohz.load_balancer) == -1) {
4789 /* make me the ilb owner */
4790 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4792 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4795 if (!(sched_smt_power_savings ||
4796 sched_mc_power_savings))
4799 * Check to see if there is a more power-efficient
4802 new_ilb = find_new_ilb(cpu);
4803 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4804 atomic_set(&nohz.load_balancer, -1);
4805 resched_cpu(new_ilb);
4811 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4814 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4816 if (atomic_read(&nohz.load_balancer) == cpu)
4817 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4824 static DEFINE_SPINLOCK(balancing);
4827 * It checks each scheduling domain to see if it is due to be balanced,
4828 * and initiates a balancing operation if so.
4830 * Balancing parameters are set up in arch_init_sched_domains.
4832 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4835 struct rq *rq = cpu_rq(cpu);
4836 unsigned long interval;
4837 struct sched_domain *sd;
4838 /* Earliest time when we have to do rebalance again */
4839 unsigned long next_balance = jiffies + 60*HZ;
4840 int update_next_balance = 0;
4843 for_each_domain(cpu, sd) {
4844 if (!(sd->flags & SD_LOAD_BALANCE))
4847 interval = sd->balance_interval;
4848 if (idle != CPU_IDLE)
4849 interval *= sd->busy_factor;
4851 /* scale ms to jiffies */
4852 interval = msecs_to_jiffies(interval);
4853 if (unlikely(!interval))
4855 if (interval > HZ*NR_CPUS/10)
4856 interval = HZ*NR_CPUS/10;
4858 need_serialize = sd->flags & SD_SERIALIZE;
4860 if (need_serialize) {
4861 if (!spin_trylock(&balancing))
4865 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4866 if (load_balance(cpu, rq, sd, idle, &balance)) {
4868 * We've pulled tasks over so either we're no
4869 * longer idle, or one of our SMT siblings is
4872 idle = CPU_NOT_IDLE;
4874 sd->last_balance = jiffies;
4877 spin_unlock(&balancing);
4879 if (time_after(next_balance, sd->last_balance + interval)) {
4880 next_balance = sd->last_balance + interval;
4881 update_next_balance = 1;
4885 * Stop the load balance at this level. There is another
4886 * CPU in our sched group which is doing load balancing more
4894 * next_balance will be updated only when there is a need.
4895 * When the cpu is attached to null domain for ex, it will not be
4898 if (likely(update_next_balance))
4899 rq->next_balance = next_balance;
4903 * run_rebalance_domains is triggered when needed from the scheduler tick.
4904 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4905 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4907 static void run_rebalance_domains(struct softirq_action *h)
4909 int this_cpu = smp_processor_id();
4910 struct rq *this_rq = cpu_rq(this_cpu);
4911 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4912 CPU_IDLE : CPU_NOT_IDLE;
4914 rebalance_domains(this_cpu, idle);
4918 * If this cpu is the owner for idle load balancing, then do the
4919 * balancing on behalf of the other idle cpus whose ticks are
4922 if (this_rq->idle_at_tick &&
4923 atomic_read(&nohz.load_balancer) == this_cpu) {
4927 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4928 if (balance_cpu == this_cpu)
4932 * If this cpu gets work to do, stop the load balancing
4933 * work being done for other cpus. Next load
4934 * balancing owner will pick it up.
4939 rebalance_domains(balance_cpu, CPU_IDLE);
4941 rq = cpu_rq(balance_cpu);
4942 if (time_after(this_rq->next_balance, rq->next_balance))
4943 this_rq->next_balance = rq->next_balance;
4949 static inline int on_null_domain(int cpu)
4951 return !rcu_dereference(cpu_rq(cpu)->sd);
4955 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4957 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4958 * idle load balancing owner or decide to stop the periodic load balancing,
4959 * if the whole system is idle.
4961 static inline void trigger_load_balance(struct rq *rq, int cpu)
4965 * If we were in the nohz mode recently and busy at the current
4966 * scheduler tick, then check if we need to nominate new idle
4969 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4970 rq->in_nohz_recently = 0;
4972 if (atomic_read(&nohz.load_balancer) == cpu) {
4973 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4974 atomic_set(&nohz.load_balancer, -1);
4977 if (atomic_read(&nohz.load_balancer) == -1) {
4978 int ilb = find_new_ilb(cpu);
4980 if (ilb < nr_cpu_ids)
4986 * If this cpu is idle and doing idle load balancing for all the
4987 * cpus with ticks stopped, is it time for that to stop?
4989 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4990 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4996 * If this cpu is idle and the idle load balancing is done by
4997 * someone else, then no need raise the SCHED_SOFTIRQ
4999 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
5000 cpumask_test_cpu(cpu, nohz.cpu_mask))
5003 /* Don't need to rebalance while attached to NULL domain */
5004 if (time_after_eq(jiffies, rq->next_balance) &&
5005 likely(!on_null_domain(cpu)))
5006 raise_softirq(SCHED_SOFTIRQ);
5009 #else /* CONFIG_SMP */
5012 * on UP we do not need to balance between CPUs:
5014 static inline void idle_balance(int cpu, struct rq *rq)
5020 DEFINE_PER_CPU(struct kernel_stat, kstat);
5022 EXPORT_PER_CPU_SYMBOL(kstat);
5025 * Return any ns on the sched_clock that have not yet been accounted in
5026 * @p in case that task is currently running.
5028 * Called with task_rq_lock() held on @rq.
5030 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
5034 if (task_current(rq, p)) {
5035 update_rq_clock(rq);
5036 ns = rq->clock - p->se.exec_start;
5044 unsigned long long task_delta_exec(struct task_struct *p)
5046 unsigned long flags;
5050 rq = task_rq_lock(p, &flags);
5051 ns = do_task_delta_exec(p, rq);
5052 task_rq_unlock(rq, &flags);
5058 * Return accounted runtime for the task.
5059 * In case the task is currently running, return the runtime plus current's
5060 * pending runtime that have not been accounted yet.
5062 unsigned long long task_sched_runtime(struct task_struct *p)
5064 unsigned long flags;
5068 rq = task_rq_lock(p, &flags);
5069 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
5070 task_rq_unlock(rq, &flags);
5076 * Return sum_exec_runtime for the thread group.
5077 * In case the task is currently running, return the sum plus current's
5078 * pending runtime that have not been accounted yet.
5080 * Note that the thread group might have other running tasks as well,
5081 * so the return value not includes other pending runtime that other
5082 * running tasks might have.
5084 unsigned long long thread_group_sched_runtime(struct task_struct *p)
5086 struct task_cputime totals;
5087 unsigned long flags;
5091 rq = task_rq_lock(p, &flags);
5092 thread_group_cputime(p, &totals);
5093 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
5094 task_rq_unlock(rq, &flags);
5100 * Account user cpu time to a process.
5101 * @p: the process that the cpu time gets accounted to
5102 * @cputime: the cpu time spent in user space since the last update
5103 * @cputime_scaled: cputime scaled by cpu frequency
5105 void account_user_time(struct task_struct *p, cputime_t cputime,
5106 cputime_t cputime_scaled)
5108 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5111 /* Add user time to process. */
5112 p->utime = cputime_add(p->utime, cputime);
5113 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5114 account_group_user_time(p, cputime);
5116 /* Add user time to cpustat. */
5117 tmp = cputime_to_cputime64(cputime);
5118 if (TASK_NICE(p) > 0)
5119 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5121 cpustat->user = cputime64_add(cpustat->user, tmp);
5123 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5124 /* Account for user time used */
5125 acct_update_integrals(p);
5129 * Account guest cpu time to a process.
5130 * @p: the process that the cpu time gets accounted to
5131 * @cputime: the cpu time spent in virtual machine since the last update
5132 * @cputime_scaled: cputime scaled by cpu frequency
5134 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5135 cputime_t cputime_scaled)
5138 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5140 tmp = cputime_to_cputime64(cputime);
5142 /* Add guest time to process. */
5143 p->utime = cputime_add(p->utime, cputime);
5144 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5145 account_group_user_time(p, cputime);
5146 p->gtime = cputime_add(p->gtime, cputime);
5148 /* Add guest time to cpustat. */
5149 cpustat->user = cputime64_add(cpustat->user, tmp);
5150 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5154 * Account system cpu time to a process.
5155 * @p: the process that the cpu time gets accounted to
5156 * @hardirq_offset: the offset to subtract from hardirq_count()
5157 * @cputime: the cpu time spent in kernel space since the last update
5158 * @cputime_scaled: cputime scaled by cpu frequency
5160 void account_system_time(struct task_struct *p, int hardirq_offset,
5161 cputime_t cputime, cputime_t cputime_scaled)
5163 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5166 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5167 account_guest_time(p, cputime, cputime_scaled);
5171 /* Add system time to process. */
5172 p->stime = cputime_add(p->stime, cputime);
5173 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5174 account_group_system_time(p, cputime);
5176 /* Add system time to cpustat. */
5177 tmp = cputime_to_cputime64(cputime);
5178 if (hardirq_count() - hardirq_offset)
5179 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5180 else if (softirq_count())
5181 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5183 cpustat->system = cputime64_add(cpustat->system, tmp);
5185 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5187 /* Account for system time used */
5188 acct_update_integrals(p);
5192 * Account for involuntary wait time.
5193 * @steal: the cpu time spent in involuntary wait
5195 void account_steal_time(cputime_t cputime)
5197 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5198 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5200 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5204 * Account for idle time.
5205 * @cputime: the cpu time spent in idle wait
5207 void account_idle_time(cputime_t cputime)
5209 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5210 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5211 struct rq *rq = this_rq();
5213 if (atomic_read(&rq->nr_iowait) > 0)
5214 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5216 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5219 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5222 * Account a single tick of cpu time.
5223 * @p: the process that the cpu time gets accounted to
5224 * @user_tick: indicates if the tick is a user or a system tick
5226 void account_process_tick(struct task_struct *p, int user_tick)
5228 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5229 struct rq *rq = this_rq();
5232 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5233 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5234 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5237 account_idle_time(cputime_one_jiffy);
5241 * Account multiple ticks of steal time.
5242 * @p: the process from which the cpu time has been stolen
5243 * @ticks: number of stolen ticks
5245 void account_steal_ticks(unsigned long ticks)
5247 account_steal_time(jiffies_to_cputime(ticks));
5251 * Account multiple ticks of idle time.
5252 * @ticks: number of stolen ticks
5254 void account_idle_ticks(unsigned long ticks)
5256 account_idle_time(jiffies_to_cputime(ticks));
5262 * Use precise platform statistics if available:
5264 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5265 cputime_t task_utime(struct task_struct *p)
5270 cputime_t task_stime(struct task_struct *p)
5275 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5277 struct task_cputime cputime;
5279 thread_group_cputime(p, &cputime);
5281 *ut = cputime.utime;
5282 *st = cputime.stime;
5286 #ifndef nsecs_to_cputime
5287 # define nsecs_to_cputime(__nsecs) \
5288 msecs_to_cputime(div_u64((__nsecs), NSEC_PER_MSEC))
5291 cputime_t task_utime(struct task_struct *p)
5293 cputime_t utime = p->utime, total = utime + p->stime;
5297 * Use CFS's precise accounting:
5299 temp = (u64)nsecs_to_cputime(p->se.sum_exec_runtime);
5303 do_div(temp, total);
5305 utime = (cputime_t)temp;
5307 p->prev_utime = max(p->prev_utime, utime);
5308 return p->prev_utime;
5311 cputime_t task_stime(struct task_struct *p)
5316 * Use CFS's precise accounting. (we subtract utime from
5317 * the total, to make sure the total observed by userspace
5318 * grows monotonically - apps rely on that):
5320 stime = nsecs_to_cputime(p->se.sum_exec_runtime) - task_utime(p);
5323 p->prev_stime = max(p->prev_stime, stime);
5325 return p->prev_stime;
5329 * Must be called with siglock held.
5331 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5333 struct signal_struct *sig = p->signal;
5334 struct task_cputime cputime;
5335 cputime_t rtime, utime, total;
5337 thread_group_cputime(p, &cputime);
5339 total = cputime_add(cputime.utime, cputime.stime);
5340 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5345 temp *= cputime.utime;
5346 do_div(temp, total);
5347 utime = (cputime_t)temp;
5351 sig->prev_utime = max(sig->prev_utime, utime);
5352 sig->prev_stime = max(sig->prev_stime,
5353 cputime_sub(rtime, sig->prev_utime));
5355 *ut = sig->prev_utime;
5356 *st = sig->prev_stime;
5360 inline cputime_t task_gtime(struct task_struct *p)
5366 * This function gets called by the timer code, with HZ frequency.
5367 * We call it with interrupts disabled.
5369 * It also gets called by the fork code, when changing the parent's
5372 void scheduler_tick(void)
5374 int cpu = smp_processor_id();
5375 struct rq *rq = cpu_rq(cpu);
5376 struct task_struct *curr = rq->curr;
5380 spin_lock(&rq->lock);
5381 update_rq_clock(rq);
5382 update_cpu_load(rq);
5383 curr->sched_class->task_tick(rq, curr, 0);
5384 spin_unlock(&rq->lock);
5386 perf_event_task_tick(curr, cpu);
5389 rq->idle_at_tick = idle_cpu(cpu);
5390 trigger_load_balance(rq, cpu);
5394 notrace unsigned long get_parent_ip(unsigned long addr)
5396 if (in_lock_functions(addr)) {
5397 addr = CALLER_ADDR2;
5398 if (in_lock_functions(addr))
5399 addr = CALLER_ADDR3;
5404 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5405 defined(CONFIG_PREEMPT_TRACER))
5407 void __kprobes add_preempt_count(int val)
5409 #ifdef CONFIG_DEBUG_PREEMPT
5413 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5416 preempt_count() += val;
5417 #ifdef CONFIG_DEBUG_PREEMPT
5419 * Spinlock count overflowing soon?
5421 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5424 if (preempt_count() == val)
5425 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5427 EXPORT_SYMBOL(add_preempt_count);
5429 void __kprobes sub_preempt_count(int val)
5431 #ifdef CONFIG_DEBUG_PREEMPT
5435 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5438 * Is the spinlock portion underflowing?
5440 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5441 !(preempt_count() & PREEMPT_MASK)))
5445 if (preempt_count() == val)
5446 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5447 preempt_count() -= val;
5449 EXPORT_SYMBOL(sub_preempt_count);
5454 * Print scheduling while atomic bug:
5456 static noinline void __schedule_bug(struct task_struct *prev)
5458 struct pt_regs *regs = get_irq_regs();
5460 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5461 prev->comm, prev->pid, preempt_count());
5463 debug_show_held_locks(prev);
5465 if (irqs_disabled())
5466 print_irqtrace_events(prev);
5475 * Various schedule()-time debugging checks and statistics:
5477 static inline void schedule_debug(struct task_struct *prev)
5480 * Test if we are atomic. Since do_exit() needs to call into
5481 * schedule() atomically, we ignore that path for now.
5482 * Otherwise, whine if we are scheduling when we should not be.
5484 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5485 __schedule_bug(prev);
5487 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5489 schedstat_inc(this_rq(), sched_count);
5490 #ifdef CONFIG_SCHEDSTATS
5491 if (unlikely(prev->lock_depth >= 0)) {
5492 schedstat_inc(this_rq(), bkl_count);
5493 schedstat_inc(prev, sched_info.bkl_count);
5498 static void put_prev_task(struct rq *rq, struct task_struct *p)
5500 u64 runtime = p->se.sum_exec_runtime - p->se.prev_sum_exec_runtime;
5502 update_avg(&p->se.avg_running, runtime);
5504 if (p->state == TASK_RUNNING) {
5506 * In order to avoid avg_overlap growing stale when we are
5507 * indeed overlapping and hence not getting put to sleep, grow
5508 * the avg_overlap on preemption.
5510 * We use the average preemption runtime because that
5511 * correlates to the amount of cache footprint a task can
5514 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5515 update_avg(&p->se.avg_overlap, runtime);
5517 update_avg(&p->se.avg_running, 0);
5519 p->sched_class->put_prev_task(rq, p);
5523 * Pick up the highest-prio task:
5525 static inline struct task_struct *
5526 pick_next_task(struct rq *rq)
5528 const struct sched_class *class;
5529 struct task_struct *p;
5532 * Optimization: we know that if all tasks are in
5533 * the fair class we can call that function directly:
5535 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5536 p = fair_sched_class.pick_next_task(rq);
5541 class = sched_class_highest;
5543 p = class->pick_next_task(rq);
5547 * Will never be NULL as the idle class always
5548 * returns a non-NULL p:
5550 class = class->next;
5555 * schedule() is the main scheduler function.
5557 asmlinkage void __sched schedule(void)
5559 struct task_struct *prev, *next;
5560 unsigned long *switch_count;
5566 cpu = smp_processor_id();
5570 switch_count = &prev->nivcsw;
5572 release_kernel_lock(prev);
5573 need_resched_nonpreemptible:
5575 schedule_debug(prev);
5577 if (sched_feat(HRTICK))
5580 spin_lock_irq(&rq->lock);
5581 update_rq_clock(rq);
5582 clear_tsk_need_resched(prev);
5584 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5585 if (unlikely(signal_pending_state(prev->state, prev)))
5586 prev->state = TASK_RUNNING;
5588 deactivate_task(rq, prev, 1);
5589 switch_count = &prev->nvcsw;
5592 pre_schedule(rq, prev);
5594 if (unlikely(!rq->nr_running))
5595 idle_balance(cpu, rq);
5597 put_prev_task(rq, prev);
5598 next = pick_next_task(rq);
5600 if (likely(prev != next)) {
5601 sched_info_switch(prev, next);
5602 perf_event_task_sched_out(prev, next, cpu);
5608 context_switch(rq, prev, next); /* unlocks the rq */
5610 * the context switch might have flipped the stack from under
5611 * us, hence refresh the local variables.
5613 cpu = smp_processor_id();
5616 spin_unlock_irq(&rq->lock);
5620 if (unlikely(reacquire_kernel_lock(current) < 0))
5621 goto need_resched_nonpreemptible;
5623 preempt_enable_no_resched();
5627 EXPORT_SYMBOL(schedule);
5631 * Look out! "owner" is an entirely speculative pointer
5632 * access and not reliable.
5634 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5639 if (!sched_feat(OWNER_SPIN))
5642 #ifdef CONFIG_DEBUG_PAGEALLOC
5644 * Need to access the cpu field knowing that
5645 * DEBUG_PAGEALLOC could have unmapped it if
5646 * the mutex owner just released it and exited.
5648 if (probe_kernel_address(&owner->cpu, cpu))
5655 * Even if the access succeeded (likely case),
5656 * the cpu field may no longer be valid.
5658 if (cpu >= nr_cpumask_bits)
5662 * We need to validate that we can do a
5663 * get_cpu() and that we have the percpu area.
5665 if (!cpu_online(cpu))
5672 * Owner changed, break to re-assess state.
5674 if (lock->owner != owner)
5678 * Is that owner really running on that cpu?
5680 if (task_thread_info(rq->curr) != owner || need_resched())
5690 #ifdef CONFIG_PREEMPT
5692 * this is the entry point to schedule() from in-kernel preemption
5693 * off of preempt_enable. Kernel preemptions off return from interrupt
5694 * occur there and call schedule directly.
5696 asmlinkage void __sched preempt_schedule(void)
5698 struct thread_info *ti = current_thread_info();
5701 * If there is a non-zero preempt_count or interrupts are disabled,
5702 * we do not want to preempt the current task. Just return..
5704 if (likely(ti->preempt_count || irqs_disabled()))
5708 add_preempt_count(PREEMPT_ACTIVE);
5710 sub_preempt_count(PREEMPT_ACTIVE);
5713 * Check again in case we missed a preemption opportunity
5714 * between schedule and now.
5717 } while (need_resched());
5719 EXPORT_SYMBOL(preempt_schedule);
5722 * this is the entry point to schedule() from kernel preemption
5723 * off of irq context.
5724 * Note, that this is called and return with irqs disabled. This will
5725 * protect us against recursive calling from irq.
5727 asmlinkage void __sched preempt_schedule_irq(void)
5729 struct thread_info *ti = current_thread_info();
5731 /* Catch callers which need to be fixed */
5732 BUG_ON(ti->preempt_count || !irqs_disabled());
5735 add_preempt_count(PREEMPT_ACTIVE);
5738 local_irq_disable();
5739 sub_preempt_count(PREEMPT_ACTIVE);
5742 * Check again in case we missed a preemption opportunity
5743 * between schedule and now.
5746 } while (need_resched());
5749 #endif /* CONFIG_PREEMPT */
5751 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5754 return try_to_wake_up(curr->private, mode, wake_flags);
5756 EXPORT_SYMBOL(default_wake_function);
5759 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5760 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5761 * number) then we wake all the non-exclusive tasks and one exclusive task.
5763 * There are circumstances in which we can try to wake a task which has already
5764 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5765 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5767 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5768 int nr_exclusive, int wake_flags, void *key)
5770 wait_queue_t *curr, *next;
5772 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5773 unsigned flags = curr->flags;
5775 if (curr->func(curr, mode, wake_flags, key) &&
5776 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5782 * __wake_up - wake up threads blocked on a waitqueue.
5784 * @mode: which threads
5785 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5786 * @key: is directly passed to the wakeup function
5788 * It may be assumed that this function implies a write memory barrier before
5789 * changing the task state if and only if any tasks are woken up.
5791 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5792 int nr_exclusive, void *key)
5794 unsigned long flags;
5796 spin_lock_irqsave(&q->lock, flags);
5797 __wake_up_common(q, mode, nr_exclusive, 0, key);
5798 spin_unlock_irqrestore(&q->lock, flags);
5800 EXPORT_SYMBOL(__wake_up);
5803 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5805 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5807 __wake_up_common(q, mode, 1, 0, NULL);
5810 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5812 __wake_up_common(q, mode, 1, 0, key);
5816 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5818 * @mode: which threads
5819 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5820 * @key: opaque value to be passed to wakeup targets
5822 * The sync wakeup differs that the waker knows that it will schedule
5823 * away soon, so while the target thread will be woken up, it will not
5824 * be migrated to another CPU - ie. the two threads are 'synchronized'
5825 * with each other. This can prevent needless bouncing between CPUs.
5827 * On UP it can prevent extra preemption.
5829 * It may be assumed that this function implies a write memory barrier before
5830 * changing the task state if and only if any tasks are woken up.
5832 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5833 int nr_exclusive, void *key)
5835 unsigned long flags;
5836 int wake_flags = WF_SYNC;
5841 if (unlikely(!nr_exclusive))
5844 spin_lock_irqsave(&q->lock, flags);
5845 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5846 spin_unlock_irqrestore(&q->lock, flags);
5848 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5851 * __wake_up_sync - see __wake_up_sync_key()
5853 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5855 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5857 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5860 * complete: - signals a single thread waiting on this completion
5861 * @x: holds the state of this particular completion
5863 * This will wake up a single thread waiting on this completion. Threads will be
5864 * awakened in the same order in which they were queued.
5866 * See also complete_all(), wait_for_completion() and related routines.
5868 * It may be assumed that this function implies a write memory barrier before
5869 * changing the task state if and only if any tasks are woken up.
5871 void complete(struct completion *x)
5873 unsigned long flags;
5875 spin_lock_irqsave(&x->wait.lock, flags);
5877 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5878 spin_unlock_irqrestore(&x->wait.lock, flags);
5880 EXPORT_SYMBOL(complete);
5883 * complete_all: - signals all threads waiting on this completion
5884 * @x: holds the state of this particular completion
5886 * This will wake up all threads waiting on this particular completion event.
5888 * It may be assumed that this function implies a write memory barrier before
5889 * changing the task state if and only if any tasks are woken up.
5891 void complete_all(struct completion *x)
5893 unsigned long flags;
5895 spin_lock_irqsave(&x->wait.lock, flags);
5896 x->done += UINT_MAX/2;
5897 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5898 spin_unlock_irqrestore(&x->wait.lock, flags);
5900 EXPORT_SYMBOL(complete_all);
5902 static inline long __sched
5903 do_wait_for_common(struct completion *x, long timeout, int state)
5906 DECLARE_WAITQUEUE(wait, current);
5908 wait.flags |= WQ_FLAG_EXCLUSIVE;
5909 __add_wait_queue_tail(&x->wait, &wait);
5911 if (signal_pending_state(state, current)) {
5912 timeout = -ERESTARTSYS;
5915 __set_current_state(state);
5916 spin_unlock_irq(&x->wait.lock);
5917 timeout = schedule_timeout(timeout);
5918 spin_lock_irq(&x->wait.lock);
5919 } while (!x->done && timeout);
5920 __remove_wait_queue(&x->wait, &wait);
5925 return timeout ?: 1;
5929 wait_for_common(struct completion *x, long timeout, int state)
5933 spin_lock_irq(&x->wait.lock);
5934 timeout = do_wait_for_common(x, timeout, state);
5935 spin_unlock_irq(&x->wait.lock);
5940 * wait_for_completion: - waits for completion of a task
5941 * @x: holds the state of this particular completion
5943 * This waits to be signaled for completion of a specific task. It is NOT
5944 * interruptible and there is no timeout.
5946 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5947 * and interrupt capability. Also see complete().
5949 void __sched wait_for_completion(struct completion *x)
5951 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5953 EXPORT_SYMBOL(wait_for_completion);
5956 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5957 * @x: holds the state of this particular completion
5958 * @timeout: timeout value in jiffies
5960 * This waits for either a completion of a specific task to be signaled or for a
5961 * specified timeout to expire. The timeout is in jiffies. It is not
5964 unsigned long __sched
5965 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5967 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5969 EXPORT_SYMBOL(wait_for_completion_timeout);
5972 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5973 * @x: holds the state of this particular completion
5975 * This waits for completion of a specific task to be signaled. It is
5978 int __sched wait_for_completion_interruptible(struct completion *x)
5980 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5981 if (t == -ERESTARTSYS)
5985 EXPORT_SYMBOL(wait_for_completion_interruptible);
5988 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5989 * @x: holds the state of this particular completion
5990 * @timeout: timeout value in jiffies
5992 * This waits for either a completion of a specific task to be signaled or for a
5993 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5995 unsigned long __sched
5996 wait_for_completion_interruptible_timeout(struct completion *x,
5997 unsigned long timeout)
5999 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
6001 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
6004 * wait_for_completion_killable: - waits for completion of a task (killable)
6005 * @x: holds the state of this particular completion
6007 * This waits to be signaled for completion of a specific task. It can be
6008 * interrupted by a kill signal.
6010 int __sched wait_for_completion_killable(struct completion *x)
6012 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
6013 if (t == -ERESTARTSYS)
6017 EXPORT_SYMBOL(wait_for_completion_killable);
6020 * try_wait_for_completion - try to decrement a completion without blocking
6021 * @x: completion structure
6023 * Returns: 0 if a decrement cannot be done without blocking
6024 * 1 if a decrement succeeded.
6026 * If a completion is being used as a counting completion,
6027 * attempt to decrement the counter without blocking. This
6028 * enables us to avoid waiting if the resource the completion
6029 * is protecting is not available.
6031 bool try_wait_for_completion(struct completion *x)
6033 unsigned long flags;
6036 spin_lock_irqsave(&x->wait.lock, flags);
6041 spin_unlock_irqrestore(&x->wait.lock, flags);
6044 EXPORT_SYMBOL(try_wait_for_completion);
6047 * completion_done - Test to see if a completion has any waiters
6048 * @x: completion structure
6050 * Returns: 0 if there are waiters (wait_for_completion() in progress)
6051 * 1 if there are no waiters.
6054 bool completion_done(struct completion *x)
6056 unsigned long flags;
6059 spin_lock_irqsave(&x->wait.lock, flags);
6062 spin_unlock_irqrestore(&x->wait.lock, flags);
6065 EXPORT_SYMBOL(completion_done);
6068 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
6070 unsigned long flags;
6073 init_waitqueue_entry(&wait, current);
6075 __set_current_state(state);
6077 spin_lock_irqsave(&q->lock, flags);
6078 __add_wait_queue(q, &wait);
6079 spin_unlock(&q->lock);
6080 timeout = schedule_timeout(timeout);
6081 spin_lock_irq(&q->lock);
6082 __remove_wait_queue(q, &wait);
6083 spin_unlock_irqrestore(&q->lock, flags);
6088 void __sched interruptible_sleep_on(wait_queue_head_t *q)
6090 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6092 EXPORT_SYMBOL(interruptible_sleep_on);
6095 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
6097 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
6099 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
6101 void __sched sleep_on(wait_queue_head_t *q)
6103 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6105 EXPORT_SYMBOL(sleep_on);
6107 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
6109 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
6111 EXPORT_SYMBOL(sleep_on_timeout);
6113 #ifdef CONFIG_RT_MUTEXES
6116 * rt_mutex_setprio - set the current priority of a task
6118 * @prio: prio value (kernel-internal form)
6120 * This function changes the 'effective' priority of a task. It does
6121 * not touch ->normal_prio like __setscheduler().
6123 * Used by the rt_mutex code to implement priority inheritance logic.
6125 void rt_mutex_setprio(struct task_struct *p, int prio)
6127 unsigned long flags;
6128 int oldprio, on_rq, running;
6130 const struct sched_class *prev_class;
6132 BUG_ON(prio < 0 || prio > MAX_PRIO);
6134 rq = task_rq_lock(p, &flags);
6135 update_rq_clock(rq);
6138 prev_class = p->sched_class;
6139 on_rq = p->se.on_rq;
6140 running = task_current(rq, p);
6142 dequeue_task(rq, p, 0);
6144 p->sched_class->put_prev_task(rq, p);
6147 p->sched_class = &rt_sched_class;
6149 p->sched_class = &fair_sched_class;
6154 p->sched_class->set_curr_task(rq);
6156 enqueue_task(rq, p, 0, oldprio < prio);
6158 check_class_changed(rq, p, prev_class, oldprio, running);
6160 task_rq_unlock(rq, &flags);
6165 void set_user_nice(struct task_struct *p, long nice)
6167 int old_prio, delta, on_rq;
6168 unsigned long flags;
6171 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6174 * We have to be careful, if called from sys_setpriority(),
6175 * the task might be in the middle of scheduling on another CPU.
6177 rq = task_rq_lock(p, &flags);
6178 update_rq_clock(rq);
6180 * The RT priorities are set via sched_setscheduler(), but we still
6181 * allow the 'normal' nice value to be set - but as expected
6182 * it wont have any effect on scheduling until the task is
6183 * SCHED_FIFO/SCHED_RR:
6185 if (task_has_rt_policy(p)) {
6186 p->static_prio = NICE_TO_PRIO(nice);
6189 on_rq = p->se.on_rq;
6191 dequeue_task(rq, p, 0);
6193 p->static_prio = NICE_TO_PRIO(nice);
6196 p->prio = effective_prio(p);
6197 delta = p->prio - old_prio;
6200 enqueue_task(rq, p, 0, false);
6202 * If the task increased its priority or is running and
6203 * lowered its priority, then reschedule its CPU:
6205 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6206 resched_task(rq->curr);
6209 task_rq_unlock(rq, &flags);
6211 EXPORT_SYMBOL(set_user_nice);
6214 * can_nice - check if a task can reduce its nice value
6218 int can_nice(const struct task_struct *p, const int nice)
6220 /* convert nice value [19,-20] to rlimit style value [1,40] */
6221 int nice_rlim = 20 - nice;
6223 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6224 capable(CAP_SYS_NICE));
6227 #ifdef __ARCH_WANT_SYS_NICE
6230 * sys_nice - change the priority of the current process.
6231 * @increment: priority increment
6233 * sys_setpriority is a more generic, but much slower function that
6234 * does similar things.
6236 SYSCALL_DEFINE1(nice, int, increment)
6241 * Setpriority might change our priority at the same moment.
6242 * We don't have to worry. Conceptually one call occurs first
6243 * and we have a single winner.
6245 if (increment < -40)
6250 nice = TASK_NICE(current) + increment;
6256 if (increment < 0 && !can_nice(current, nice))
6259 retval = security_task_setnice(current, nice);
6263 set_user_nice(current, nice);
6270 * task_prio - return the priority value of a given task.
6271 * @p: the task in question.
6273 * This is the priority value as seen by users in /proc.
6274 * RT tasks are offset by -200. Normal tasks are centered
6275 * around 0, value goes from -16 to +15.
6277 int task_prio(const struct task_struct *p)
6279 return p->prio - MAX_RT_PRIO;
6283 * task_nice - return the nice value of a given task.
6284 * @p: the task in question.
6286 int task_nice(const struct task_struct *p)
6288 return TASK_NICE(p);
6290 EXPORT_SYMBOL(task_nice);
6293 * idle_cpu - is a given cpu idle currently?
6294 * @cpu: the processor in question.
6296 int idle_cpu(int cpu)
6298 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6302 * idle_task - return the idle task for a given cpu.
6303 * @cpu: the processor in question.
6305 struct task_struct *idle_task(int cpu)
6307 return cpu_rq(cpu)->idle;
6311 * find_process_by_pid - find a process with a matching PID value.
6312 * @pid: the pid in question.
6314 static struct task_struct *find_process_by_pid(pid_t pid)
6316 return pid ? find_task_by_vpid(pid) : current;
6319 /* Actually do priority change: must hold rq lock. */
6321 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6323 BUG_ON(p->se.on_rq);
6326 switch (p->policy) {
6330 p->sched_class = &fair_sched_class;
6334 p->sched_class = &rt_sched_class;
6338 p->rt_priority = prio;
6339 p->normal_prio = normal_prio(p);
6340 /* we are holding p->pi_lock already */
6341 p->prio = rt_mutex_getprio(p);
6346 * check the target process has a UID that matches the current process's
6348 static bool check_same_owner(struct task_struct *p)
6350 const struct cred *cred = current_cred(), *pcred;
6354 pcred = __task_cred(p);
6355 match = (cred->euid == pcred->euid ||
6356 cred->euid == pcred->uid);
6361 static int __sched_setscheduler(struct task_struct *p, int policy,
6362 struct sched_param *param, bool user)
6364 int retval, oldprio, oldpolicy = -1, on_rq, running;
6365 unsigned long flags;
6366 const struct sched_class *prev_class;
6370 /* may grab non-irq protected spin_locks */
6371 BUG_ON(in_interrupt());
6373 /* double check policy once rq lock held */
6375 reset_on_fork = p->sched_reset_on_fork;
6376 policy = oldpolicy = p->policy;
6378 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6379 policy &= ~SCHED_RESET_ON_FORK;
6381 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6382 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6383 policy != SCHED_IDLE)
6388 * Valid priorities for SCHED_FIFO and SCHED_RR are
6389 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6390 * SCHED_BATCH and SCHED_IDLE is 0.
6392 if (param->sched_priority < 0 ||
6393 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6394 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6396 if (rt_policy(policy) != (param->sched_priority != 0))
6400 * Allow unprivileged RT tasks to decrease priority:
6402 if (user && !capable(CAP_SYS_NICE)) {
6403 if (rt_policy(policy)) {
6404 unsigned long rlim_rtprio;
6406 if (!lock_task_sighand(p, &flags))
6408 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6409 unlock_task_sighand(p, &flags);
6411 /* can't set/change the rt policy */
6412 if (policy != p->policy && !rlim_rtprio)
6415 /* can't increase priority */
6416 if (param->sched_priority > p->rt_priority &&
6417 param->sched_priority > rlim_rtprio)
6421 * Like positive nice levels, dont allow tasks to
6422 * move out of SCHED_IDLE either:
6424 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6427 /* can't change other user's priorities */
6428 if (!check_same_owner(p))
6431 /* Normal users shall not reset the sched_reset_on_fork flag */
6432 if (p->sched_reset_on_fork && !reset_on_fork)
6437 #ifdef CONFIG_RT_GROUP_SCHED
6439 * Do not allow realtime tasks into groups that have no runtime
6442 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6443 task_group(p)->rt_bandwidth.rt_runtime == 0)
6447 retval = security_task_setscheduler(p, policy, param);
6453 * make sure no PI-waiters arrive (or leave) while we are
6454 * changing the priority of the task:
6456 spin_lock_irqsave(&p->pi_lock, flags);
6458 * To be able to change p->policy safely, the apropriate
6459 * runqueue lock must be held.
6461 rq = __task_rq_lock(p);
6462 /* recheck policy now with rq lock held */
6463 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6464 policy = oldpolicy = -1;
6465 __task_rq_unlock(rq);
6466 spin_unlock_irqrestore(&p->pi_lock, flags);
6469 update_rq_clock(rq);
6470 on_rq = p->se.on_rq;
6471 running = task_current(rq, p);
6473 deactivate_task(rq, p, 0);
6475 p->sched_class->put_prev_task(rq, p);
6477 p->sched_reset_on_fork = reset_on_fork;
6480 prev_class = p->sched_class;
6481 __setscheduler(rq, p, policy, param->sched_priority);
6484 p->sched_class->set_curr_task(rq);
6486 activate_task(rq, p, 0);
6488 check_class_changed(rq, p, prev_class, oldprio, running);
6490 __task_rq_unlock(rq);
6491 spin_unlock_irqrestore(&p->pi_lock, flags);
6493 rt_mutex_adjust_pi(p);
6499 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6500 * @p: the task in question.
6501 * @policy: new policy.
6502 * @param: structure containing the new RT priority.
6504 * NOTE that the task may be already dead.
6506 int sched_setscheduler(struct task_struct *p, int policy,
6507 struct sched_param *param)
6509 return __sched_setscheduler(p, policy, param, true);
6511 EXPORT_SYMBOL_GPL(sched_setscheduler);
6514 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6515 * @p: the task in question.
6516 * @policy: new policy.
6517 * @param: structure containing the new RT priority.
6519 * Just like sched_setscheduler, only don't bother checking if the
6520 * current context has permission. For example, this is needed in
6521 * stop_machine(): we create temporary high priority worker threads,
6522 * but our caller might not have that capability.
6524 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6525 struct sched_param *param)
6527 return __sched_setscheduler(p, policy, param, false);
6531 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6533 struct sched_param lparam;
6534 struct task_struct *p;
6537 if (!param || pid < 0)
6539 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6544 p = find_process_by_pid(pid);
6546 retval = sched_setscheduler(p, policy, &lparam);
6553 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6554 * @pid: the pid in question.
6555 * @policy: new policy.
6556 * @param: structure containing the new RT priority.
6558 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6559 struct sched_param __user *, param)
6561 /* negative values for policy are not valid */
6565 return do_sched_setscheduler(pid, policy, param);
6569 * sys_sched_setparam - set/change the RT priority of a thread
6570 * @pid: the pid in question.
6571 * @param: structure containing the new RT priority.
6573 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6575 return do_sched_setscheduler(pid, -1, param);
6579 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6580 * @pid: the pid in question.
6582 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6584 struct task_struct *p;
6592 p = find_process_by_pid(pid);
6594 retval = security_task_getscheduler(p);
6597 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6604 * sys_sched_getparam - get the RT priority of a thread
6605 * @pid: the pid in question.
6606 * @param: structure containing the RT priority.
6608 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6610 struct sched_param lp;
6611 struct task_struct *p;
6614 if (!param || pid < 0)
6618 p = find_process_by_pid(pid);
6623 retval = security_task_getscheduler(p);
6627 lp.sched_priority = p->rt_priority;
6631 * This one might sleep, we cannot do it with a spinlock held ...
6633 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6642 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6644 cpumask_var_t cpus_allowed, new_mask;
6645 struct task_struct *p;
6651 p = find_process_by_pid(pid);
6658 /* Prevent p going away */
6662 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6666 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6668 goto out_free_cpus_allowed;
6671 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6674 retval = security_task_setscheduler(p, 0, NULL);
6678 cpuset_cpus_allowed(p, cpus_allowed);
6679 cpumask_and(new_mask, in_mask, cpus_allowed);
6681 retval = set_cpus_allowed_ptr(p, new_mask);
6684 cpuset_cpus_allowed(p, cpus_allowed);
6685 if (!cpumask_subset(new_mask, cpus_allowed)) {
6687 * We must have raced with a concurrent cpuset
6688 * update. Just reset the cpus_allowed to the
6689 * cpuset's cpus_allowed
6691 cpumask_copy(new_mask, cpus_allowed);
6696 free_cpumask_var(new_mask);
6697 out_free_cpus_allowed:
6698 free_cpumask_var(cpus_allowed);
6705 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6706 struct cpumask *new_mask)
6708 if (len < cpumask_size())
6709 cpumask_clear(new_mask);
6710 else if (len > cpumask_size())
6711 len = cpumask_size();
6713 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6717 * sys_sched_setaffinity - set the cpu affinity of a process
6718 * @pid: pid of the process
6719 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6720 * @user_mask_ptr: user-space pointer to the new cpu mask
6722 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6723 unsigned long __user *, user_mask_ptr)
6725 cpumask_var_t new_mask;
6728 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6731 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6733 retval = sched_setaffinity(pid, new_mask);
6734 free_cpumask_var(new_mask);
6738 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6740 struct task_struct *p;
6741 unsigned long flags;
6749 p = find_process_by_pid(pid);
6753 retval = security_task_getscheduler(p);
6757 rq = task_rq_lock(p, &flags);
6758 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6759 task_rq_unlock(rq, &flags);
6769 * sys_sched_getaffinity - get the cpu affinity of a process
6770 * @pid: pid of the process
6771 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6772 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6774 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6775 unsigned long __user *, user_mask_ptr)
6780 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
6782 if (len & (sizeof(unsigned long)-1))
6785 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6788 ret = sched_getaffinity(pid, mask);
6790 size_t retlen = min_t(size_t, len, cpumask_size());
6792 if (copy_to_user(user_mask_ptr, mask, retlen))
6797 free_cpumask_var(mask);
6803 * sys_sched_yield - yield the current processor to other threads.
6805 * This function yields the current CPU to other tasks. If there are no
6806 * other threads running on this CPU then this function will return.
6808 SYSCALL_DEFINE0(sched_yield)
6810 struct rq *rq = this_rq_lock();
6812 schedstat_inc(rq, yld_count);
6813 current->sched_class->yield_task(rq);
6816 * Since we are going to call schedule() anyway, there's
6817 * no need to preempt or enable interrupts:
6819 __release(rq->lock);
6820 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6821 _raw_spin_unlock(&rq->lock);
6822 preempt_enable_no_resched();
6829 static inline int should_resched(void)
6831 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6834 static void __cond_resched(void)
6836 add_preempt_count(PREEMPT_ACTIVE);
6838 sub_preempt_count(PREEMPT_ACTIVE);
6841 int __sched _cond_resched(void)
6843 if (should_resched()) {
6849 EXPORT_SYMBOL(_cond_resched);
6852 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6853 * call schedule, and on return reacquire the lock.
6855 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6856 * operations here to prevent schedule() from being called twice (once via
6857 * spin_unlock(), once by hand).
6859 int __cond_resched_lock(spinlock_t *lock)
6861 int resched = should_resched();
6864 lockdep_assert_held(lock);
6866 if (spin_needbreak(lock) || resched) {
6877 EXPORT_SYMBOL(__cond_resched_lock);
6879 int __sched __cond_resched_softirq(void)
6881 BUG_ON(!in_softirq());
6883 if (should_resched()) {
6891 EXPORT_SYMBOL(__cond_resched_softirq);
6894 * yield - yield the current processor to other threads.
6896 * This is a shortcut for kernel-space yielding - it marks the
6897 * thread runnable and calls sys_sched_yield().
6899 void __sched yield(void)
6901 set_current_state(TASK_RUNNING);
6904 EXPORT_SYMBOL(yield);
6907 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6908 * that process accounting knows that this is a task in IO wait state.
6910 void __sched io_schedule(void)
6912 struct rq *rq = raw_rq();
6914 delayacct_blkio_start();
6915 atomic_inc(&rq->nr_iowait);
6916 current->in_iowait = 1;
6918 current->in_iowait = 0;
6919 atomic_dec(&rq->nr_iowait);
6920 delayacct_blkio_end();
6922 EXPORT_SYMBOL(io_schedule);
6924 long __sched io_schedule_timeout(long timeout)
6926 struct rq *rq = raw_rq();
6929 delayacct_blkio_start();
6930 atomic_inc(&rq->nr_iowait);
6931 current->in_iowait = 1;
6932 ret = schedule_timeout(timeout);
6933 current->in_iowait = 0;
6934 atomic_dec(&rq->nr_iowait);
6935 delayacct_blkio_end();
6940 * sys_sched_get_priority_max - return maximum RT priority.
6941 * @policy: scheduling class.
6943 * this syscall returns the maximum rt_priority that can be used
6944 * by a given scheduling class.
6946 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6953 ret = MAX_USER_RT_PRIO-1;
6965 * sys_sched_get_priority_min - return minimum RT priority.
6966 * @policy: scheduling class.
6968 * this syscall returns the minimum rt_priority that can be used
6969 * by a given scheduling class.
6971 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6989 * sys_sched_rr_get_interval - return the default timeslice of a process.
6990 * @pid: pid of the process.
6991 * @interval: userspace pointer to the timeslice value.
6993 * this syscall writes the default timeslice value of a given process
6994 * into the user-space timespec buffer. A value of '0' means infinity.
6996 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6997 struct timespec __user *, interval)
6999 struct task_struct *p;
7000 unsigned int time_slice;
7001 unsigned long flags;
7011 p = find_process_by_pid(pid);
7015 retval = security_task_getscheduler(p);
7019 rq = task_rq_lock(p, &flags);
7020 time_slice = p->sched_class->get_rr_interval(rq, p);
7021 task_rq_unlock(rq, &flags);
7024 jiffies_to_timespec(time_slice, &t);
7025 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
7033 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
7035 void sched_show_task(struct task_struct *p)
7037 unsigned long free = 0;
7040 state = p->state ? __ffs(p->state) + 1 : 0;
7041 printk(KERN_INFO "%-15.15s %c", p->comm,
7042 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
7043 #if BITS_PER_LONG == 32
7044 if (state == TASK_RUNNING)
7045 printk(KERN_CONT " running ");
7047 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
7049 if (state == TASK_RUNNING)
7050 printk(KERN_CONT " running task ");
7052 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
7054 #ifdef CONFIG_DEBUG_STACK_USAGE
7055 free = stack_not_used(p);
7057 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
7058 task_pid_nr(p), task_pid_nr(p->real_parent),
7059 (unsigned long)task_thread_info(p)->flags);
7061 show_stack(p, NULL);
7064 void show_state_filter(unsigned long state_filter)
7066 struct task_struct *g, *p;
7068 #if BITS_PER_LONG == 32
7070 " task PC stack pid father\n");
7073 " task PC stack pid father\n");
7075 read_lock(&tasklist_lock);
7076 do_each_thread(g, p) {
7078 * reset the NMI-timeout, listing all files on a slow
7079 * console might take alot of time:
7081 touch_nmi_watchdog();
7082 if (!state_filter || (p->state & state_filter))
7084 } while_each_thread(g, p);
7086 touch_all_softlockup_watchdogs();
7088 #ifdef CONFIG_SCHED_DEBUG
7089 sysrq_sched_debug_show();
7091 read_unlock(&tasklist_lock);
7093 * Only show locks if all tasks are dumped:
7095 if (state_filter == -1)
7096 debug_show_all_locks();
7099 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
7101 idle->sched_class = &idle_sched_class;
7105 * init_idle - set up an idle thread for a given CPU
7106 * @idle: task in question
7107 * @cpu: cpu the idle task belongs to
7109 * NOTE: this function does not set the idle thread's NEED_RESCHED
7110 * flag, to make booting more robust.
7112 void __cpuinit init_idle(struct task_struct *idle, int cpu)
7114 struct rq *rq = cpu_rq(cpu);
7115 unsigned long flags;
7117 spin_lock_irqsave(&rq->lock, flags);
7120 idle->state = TASK_RUNNING;
7121 idle->se.exec_start = sched_clock();
7123 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
7124 __set_task_cpu(idle, cpu);
7126 rq->curr = rq->idle = idle;
7127 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7130 spin_unlock_irqrestore(&rq->lock, flags);
7132 /* Set the preempt count _outside_ the spinlocks! */
7133 #if defined(CONFIG_PREEMPT)
7134 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7136 task_thread_info(idle)->preempt_count = 0;
7139 * The idle tasks have their own, simple scheduling class:
7141 idle->sched_class = &idle_sched_class;
7142 ftrace_graph_init_task(idle);
7146 * In a system that switches off the HZ timer nohz_cpu_mask
7147 * indicates which cpus entered this state. This is used
7148 * in the rcu update to wait only for active cpus. For system
7149 * which do not switch off the HZ timer nohz_cpu_mask should
7150 * always be CPU_BITS_NONE.
7152 cpumask_var_t nohz_cpu_mask;
7155 * Increase the granularity value when there are more CPUs,
7156 * because with more CPUs the 'effective latency' as visible
7157 * to users decreases. But the relationship is not linear,
7158 * so pick a second-best guess by going with the log2 of the
7161 * This idea comes from the SD scheduler of Con Kolivas:
7163 static void update_sysctl(void)
7165 unsigned int cpus = min(num_online_cpus(), 8U);
7166 unsigned int factor = 1 + ilog2(cpus);
7168 #define SET_SYSCTL(name) \
7169 (sysctl_##name = (factor) * normalized_sysctl_##name)
7170 SET_SYSCTL(sched_min_granularity);
7171 SET_SYSCTL(sched_latency);
7172 SET_SYSCTL(sched_wakeup_granularity);
7173 SET_SYSCTL(sched_shares_ratelimit);
7177 static inline void sched_init_granularity(void)
7184 * This is how migration works:
7186 * 1) we queue a struct migration_req structure in the source CPU's
7187 * runqueue and wake up that CPU's migration thread.
7188 * 2) we down() the locked semaphore => thread blocks.
7189 * 3) migration thread wakes up (implicitly it forces the migrated
7190 * thread off the CPU)
7191 * 4) it gets the migration request and checks whether the migrated
7192 * task is still in the wrong runqueue.
7193 * 5) if it's in the wrong runqueue then the migration thread removes
7194 * it and puts it into the right queue.
7195 * 6) migration thread up()s the semaphore.
7196 * 7) we wake up and the migration is done.
7200 * Change a given task's CPU affinity. Migrate the thread to a
7201 * proper CPU and schedule it away if the CPU it's executing on
7202 * is removed from the allowed bitmask.
7204 * NOTE: the caller must have a valid reference to the task, the
7205 * task must not exit() & deallocate itself prematurely. The
7206 * call is not atomic; no spinlocks may be held.
7208 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7210 struct migration_req req;
7211 unsigned long flags;
7216 * Serialize against TASK_WAKING so that ttwu() and wunt() can
7217 * drop the rq->lock and still rely on ->cpus_allowed.
7220 while (task_is_waking(p))
7222 rq = task_rq_lock(p, &flags);
7223 if (task_is_waking(p)) {
7224 task_rq_unlock(rq, &flags);
7228 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7233 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7234 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7239 if (p->sched_class->set_cpus_allowed)
7240 p->sched_class->set_cpus_allowed(p, new_mask);
7242 cpumask_copy(&p->cpus_allowed, new_mask);
7243 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7246 /* Can the task run on the task's current CPU? If so, we're done */
7247 if (cpumask_test_cpu(task_cpu(p), new_mask))
7250 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7251 /* Need help from migration thread: drop lock and wait. */
7252 struct task_struct *mt = rq->migration_thread;
7254 get_task_struct(mt);
7255 task_rq_unlock(rq, &flags);
7256 wake_up_process(mt);
7257 put_task_struct(mt);
7258 wait_for_completion(&req.done);
7259 tlb_migrate_finish(p->mm);
7263 task_rq_unlock(rq, &flags);
7267 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7270 * Move (not current) task off this cpu, onto dest cpu. We're doing
7271 * this because either it can't run here any more (set_cpus_allowed()
7272 * away from this CPU, or CPU going down), or because we're
7273 * attempting to rebalance this task on exec (sched_exec).
7275 * So we race with normal scheduler movements, but that's OK, as long
7276 * as the task is no longer on this CPU.
7278 * Returns non-zero if task was successfully migrated.
7280 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7282 struct rq *rq_dest, *rq_src;
7285 if (unlikely(!cpu_active(dest_cpu)))
7288 rq_src = cpu_rq(src_cpu);
7289 rq_dest = cpu_rq(dest_cpu);
7291 double_rq_lock(rq_src, rq_dest);
7292 /* Already moved. */
7293 if (task_cpu(p) != src_cpu)
7295 /* Affinity changed (again). */
7296 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7300 * If we're not on a rq, the next wake-up will ensure we're
7304 deactivate_task(rq_src, p, 0);
7305 set_task_cpu(p, dest_cpu);
7306 activate_task(rq_dest, p, 0);
7307 check_preempt_curr(rq_dest, p, 0);
7312 double_rq_unlock(rq_src, rq_dest);
7316 #define RCU_MIGRATION_IDLE 0
7317 #define RCU_MIGRATION_NEED_QS 1
7318 #define RCU_MIGRATION_GOT_QS 2
7319 #define RCU_MIGRATION_MUST_SYNC 3
7322 * migration_thread - this is a highprio system thread that performs
7323 * thread migration by bumping thread off CPU then 'pushing' onto
7326 static int migration_thread(void *data)
7329 int cpu = (long)data;
7333 BUG_ON(rq->migration_thread != current);
7335 set_current_state(TASK_INTERRUPTIBLE);
7336 while (!kthread_should_stop()) {
7337 struct migration_req *req;
7338 struct list_head *head;
7340 spin_lock_irq(&rq->lock);
7342 if (cpu_is_offline(cpu)) {
7343 spin_unlock_irq(&rq->lock);
7347 if (rq->active_balance) {
7348 active_load_balance(rq, cpu);
7349 rq->active_balance = 0;
7352 head = &rq->migration_queue;
7354 if (list_empty(head)) {
7355 spin_unlock_irq(&rq->lock);
7357 set_current_state(TASK_INTERRUPTIBLE);
7360 req = list_entry(head->next, struct migration_req, list);
7361 list_del_init(head->next);
7363 if (req->task != NULL) {
7364 spin_unlock(&rq->lock);
7365 __migrate_task(req->task, cpu, req->dest_cpu);
7366 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7367 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7368 spin_unlock(&rq->lock);
7370 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7371 spin_unlock(&rq->lock);
7372 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7376 complete(&req->done);
7378 __set_current_state(TASK_RUNNING);
7383 #ifdef CONFIG_HOTPLUG_CPU
7385 * Figure out where task on dead CPU should go, use force if necessary.
7387 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7389 struct rq *rq = cpu_rq(dead_cpu);
7390 int needs_cpu, uninitialized_var(dest_cpu);
7391 unsigned long flags;
7393 local_irq_save(flags);
7395 spin_lock(&rq->lock);
7396 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
7398 dest_cpu = select_fallback_rq(dead_cpu, p);
7399 spin_unlock(&rq->lock);
7401 * It can only fail if we race with set_cpus_allowed(),
7402 * in the racer should migrate the task anyway.
7405 __migrate_task(p, dead_cpu, dest_cpu);
7406 local_irq_restore(flags);
7410 * While a dead CPU has no uninterruptible tasks queued at this point,
7411 * it might still have a nonzero ->nr_uninterruptible counter, because
7412 * for performance reasons the counter is not stricly tracking tasks to
7413 * their home CPUs. So we just add the counter to another CPU's counter,
7414 * to keep the global sum constant after CPU-down:
7416 static void migrate_nr_uninterruptible(struct rq *rq_src)
7418 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7419 unsigned long flags;
7421 local_irq_save(flags);
7422 double_rq_lock(rq_src, rq_dest);
7423 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7424 rq_src->nr_uninterruptible = 0;
7425 double_rq_unlock(rq_src, rq_dest);
7426 local_irq_restore(flags);
7429 /* Run through task list and migrate tasks from the dead cpu. */
7430 static void migrate_live_tasks(int src_cpu)
7432 struct task_struct *p, *t;
7434 read_lock(&tasklist_lock);
7436 do_each_thread(t, p) {
7440 if (task_cpu(p) == src_cpu)
7441 move_task_off_dead_cpu(src_cpu, p);
7442 } while_each_thread(t, p);
7444 read_unlock(&tasklist_lock);
7448 * Schedules idle task to be the next runnable task on current CPU.
7449 * It does so by boosting its priority to highest possible.
7450 * Used by CPU offline code.
7452 void sched_idle_next(void)
7454 int this_cpu = smp_processor_id();
7455 struct rq *rq = cpu_rq(this_cpu);
7456 struct task_struct *p = rq->idle;
7457 unsigned long flags;
7459 /* cpu has to be offline */
7460 BUG_ON(cpu_online(this_cpu));
7463 * Strictly not necessary since rest of the CPUs are stopped by now
7464 * and interrupts disabled on the current cpu.
7466 spin_lock_irqsave(&rq->lock, flags);
7468 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7470 update_rq_clock(rq);
7471 activate_task(rq, p, 0);
7473 spin_unlock_irqrestore(&rq->lock, flags);
7477 * Ensures that the idle task is using init_mm right before its cpu goes
7480 void idle_task_exit(void)
7482 struct mm_struct *mm = current->active_mm;
7484 BUG_ON(cpu_online(smp_processor_id()));
7487 switch_mm(mm, &init_mm, current);
7491 /* called under rq->lock with disabled interrupts */
7492 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7494 struct rq *rq = cpu_rq(dead_cpu);
7496 /* Must be exiting, otherwise would be on tasklist. */
7497 BUG_ON(!p->exit_state);
7499 /* Cannot have done final schedule yet: would have vanished. */
7500 BUG_ON(p->state == TASK_DEAD);
7505 * Drop lock around migration; if someone else moves it,
7506 * that's OK. No task can be added to this CPU, so iteration is
7509 spin_unlock_irq(&rq->lock);
7510 move_task_off_dead_cpu(dead_cpu, p);
7511 spin_lock_irq(&rq->lock);
7516 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7517 static void migrate_dead_tasks(unsigned int dead_cpu)
7519 struct rq *rq = cpu_rq(dead_cpu);
7520 struct task_struct *next;
7523 if (!rq->nr_running)
7525 update_rq_clock(rq);
7526 next = pick_next_task(rq);
7529 next->sched_class->put_prev_task(rq, next);
7530 migrate_dead(dead_cpu, next);
7536 * remove the tasks which were accounted by rq from calc_load_tasks.
7538 static void calc_global_load_remove(struct rq *rq)
7540 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7541 rq->calc_load_active = 0;
7543 #endif /* CONFIG_HOTPLUG_CPU */
7545 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7547 static struct ctl_table sd_ctl_dir[] = {
7549 .procname = "sched_domain",
7555 static struct ctl_table sd_ctl_root[] = {
7557 .ctl_name = CTL_KERN,
7558 .procname = "kernel",
7560 .child = sd_ctl_dir,
7565 static struct ctl_table *sd_alloc_ctl_entry(int n)
7567 struct ctl_table *entry =
7568 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7573 static void sd_free_ctl_entry(struct ctl_table **tablep)
7575 struct ctl_table *entry;
7578 * In the intermediate directories, both the child directory and
7579 * procname are dynamically allocated and could fail but the mode
7580 * will always be set. In the lowest directory the names are
7581 * static strings and all have proc handlers.
7583 for (entry = *tablep; entry->mode; entry++) {
7585 sd_free_ctl_entry(&entry->child);
7586 if (entry->proc_handler == NULL)
7587 kfree(entry->procname);
7595 set_table_entry(struct ctl_table *entry,
7596 const char *procname, void *data, int maxlen,
7597 mode_t mode, proc_handler *proc_handler)
7599 entry->procname = procname;
7601 entry->maxlen = maxlen;
7603 entry->proc_handler = proc_handler;
7606 static struct ctl_table *
7607 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7609 struct ctl_table *table = sd_alloc_ctl_entry(13);
7614 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7615 sizeof(long), 0644, proc_doulongvec_minmax);
7616 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7617 sizeof(long), 0644, proc_doulongvec_minmax);
7618 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7619 sizeof(int), 0644, proc_dointvec_minmax);
7620 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7621 sizeof(int), 0644, proc_dointvec_minmax);
7622 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7623 sizeof(int), 0644, proc_dointvec_minmax);
7624 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7625 sizeof(int), 0644, proc_dointvec_minmax);
7626 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7627 sizeof(int), 0644, proc_dointvec_minmax);
7628 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7629 sizeof(int), 0644, proc_dointvec_minmax);
7630 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7631 sizeof(int), 0644, proc_dointvec_minmax);
7632 set_table_entry(&table[9], "cache_nice_tries",
7633 &sd->cache_nice_tries,
7634 sizeof(int), 0644, proc_dointvec_minmax);
7635 set_table_entry(&table[10], "flags", &sd->flags,
7636 sizeof(int), 0644, proc_dointvec_minmax);
7637 set_table_entry(&table[11], "name", sd->name,
7638 CORENAME_MAX_SIZE, 0444, proc_dostring);
7639 /* &table[12] is terminator */
7644 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7646 struct ctl_table *entry, *table;
7647 struct sched_domain *sd;
7648 int domain_num = 0, i;
7651 for_each_domain(cpu, sd)
7653 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7658 for_each_domain(cpu, sd) {
7659 snprintf(buf, 32, "domain%d", i);
7660 entry->procname = kstrdup(buf, GFP_KERNEL);
7662 entry->child = sd_alloc_ctl_domain_table(sd);
7669 static struct ctl_table_header *sd_sysctl_header;
7670 static void register_sched_domain_sysctl(void)
7672 int i, cpu_num = num_possible_cpus();
7673 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7676 WARN_ON(sd_ctl_dir[0].child);
7677 sd_ctl_dir[0].child = entry;
7682 for_each_possible_cpu(i) {
7683 snprintf(buf, 32, "cpu%d", i);
7684 entry->procname = kstrdup(buf, GFP_KERNEL);
7686 entry->child = sd_alloc_ctl_cpu_table(i);
7690 WARN_ON(sd_sysctl_header);
7691 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7694 /* may be called multiple times per register */
7695 static void unregister_sched_domain_sysctl(void)
7697 if (sd_sysctl_header)
7698 unregister_sysctl_table(sd_sysctl_header);
7699 sd_sysctl_header = NULL;
7700 if (sd_ctl_dir[0].child)
7701 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7704 static void register_sched_domain_sysctl(void)
7707 static void unregister_sched_domain_sysctl(void)
7712 static void set_rq_online(struct rq *rq)
7715 const struct sched_class *class;
7717 cpumask_set_cpu(rq->cpu, rq->rd->online);
7720 for_each_class(class) {
7721 if (class->rq_online)
7722 class->rq_online(rq);
7727 static void set_rq_offline(struct rq *rq)
7730 const struct sched_class *class;
7732 for_each_class(class) {
7733 if (class->rq_offline)
7734 class->rq_offline(rq);
7737 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7743 * migration_call - callback that gets triggered when a CPU is added.
7744 * Here we can start up the necessary migration thread for the new CPU.
7746 static int __cpuinit
7747 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7749 struct task_struct *p;
7750 int cpu = (long)hcpu;
7751 unsigned long flags;
7754 switch (action & ~CPU_TASKS_FROZEN) {
7756 case CPU_UP_PREPARE:
7757 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7760 kthread_bind(p, cpu);
7761 /* Must be high prio: stop_machine expects to yield to it. */
7762 rq = task_rq_lock(p, &flags);
7763 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7764 task_rq_unlock(rq, &flags);
7766 cpu_rq(cpu)->migration_thread = p;
7767 rq->calc_load_update = calc_load_update;
7771 /* Strictly unnecessary, as first user will wake it. */
7772 wake_up_process(cpu_rq(cpu)->migration_thread);
7774 /* Update our root-domain */
7776 spin_lock_irqsave(&rq->lock, flags);
7778 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7782 spin_unlock_irqrestore(&rq->lock, flags);
7785 #ifdef CONFIG_HOTPLUG_CPU
7786 case CPU_UP_CANCELED:
7787 if (!cpu_rq(cpu)->migration_thread)
7789 /* Unbind it from offline cpu so it can run. Fall thru. */
7790 kthread_bind(cpu_rq(cpu)->migration_thread,
7791 cpumask_any(cpu_online_mask));
7792 kthread_stop(cpu_rq(cpu)->migration_thread);
7793 put_task_struct(cpu_rq(cpu)->migration_thread);
7794 cpu_rq(cpu)->migration_thread = NULL;
7799 * Bring the migration thread down in CPU_POST_DEAD event,
7800 * since the timers should have got migrated by now and thus
7801 * we should not see a deadlock between trying to kill the
7802 * migration thread and the sched_rt_period_timer.
7805 kthread_stop(rq->migration_thread);
7806 put_task_struct(rq->migration_thread);
7807 rq->migration_thread = NULL;
7811 migrate_live_tasks(cpu);
7813 /* Idle task back to normal (off runqueue, low prio) */
7814 spin_lock_irq(&rq->lock);
7815 update_rq_clock(rq);
7816 deactivate_task(rq, rq->idle, 0);
7817 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7818 rq->idle->sched_class = &idle_sched_class;
7819 migrate_dead_tasks(cpu);
7820 spin_unlock_irq(&rq->lock);
7821 migrate_nr_uninterruptible(rq);
7822 BUG_ON(rq->nr_running != 0);
7823 calc_global_load_remove(rq);
7825 * No need to migrate the tasks: it was best-effort if
7826 * they didn't take sched_hotcpu_mutex. Just wake up
7829 spin_lock_irq(&rq->lock);
7830 while (!list_empty(&rq->migration_queue)) {
7831 struct migration_req *req;
7833 req = list_entry(rq->migration_queue.next,
7834 struct migration_req, list);
7835 list_del_init(&req->list);
7836 spin_unlock_irq(&rq->lock);
7837 complete(&req->done);
7838 spin_lock_irq(&rq->lock);
7840 spin_unlock_irq(&rq->lock);
7844 /* Update our root-domain */
7846 spin_lock_irqsave(&rq->lock, flags);
7848 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7851 spin_unlock_irqrestore(&rq->lock, flags);
7859 * Register at high priority so that task migration (migrate_all_tasks)
7860 * happens before everything else. This has to be lower priority than
7861 * the notifier in the perf_event subsystem, though.
7863 static struct notifier_block __cpuinitdata migration_notifier = {
7864 .notifier_call = migration_call,
7868 static int __init migration_init(void)
7870 void *cpu = (void *)(long)smp_processor_id();
7873 /* Start one for the boot CPU: */
7874 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7875 BUG_ON(err == NOTIFY_BAD);
7876 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7877 register_cpu_notifier(&migration_notifier);
7881 early_initcall(migration_init);
7886 #ifdef CONFIG_SCHED_DEBUG
7888 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7889 struct cpumask *groupmask)
7891 struct sched_group *group = sd->groups;
7894 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7895 cpumask_clear(groupmask);
7897 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7899 if (!(sd->flags & SD_LOAD_BALANCE)) {
7900 printk("does not load-balance\n");
7902 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7907 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7909 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7910 printk(KERN_ERR "ERROR: domain->span does not contain "
7913 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7914 printk(KERN_ERR "ERROR: domain->groups does not contain"
7918 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7922 printk(KERN_ERR "ERROR: group is NULL\n");
7926 if (!group->cpu_power) {
7927 printk(KERN_CONT "\n");
7928 printk(KERN_ERR "ERROR: domain->cpu_power not "
7933 if (!cpumask_weight(sched_group_cpus(group))) {
7934 printk(KERN_CONT "\n");
7935 printk(KERN_ERR "ERROR: empty group\n");
7939 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7940 printk(KERN_CONT "\n");
7941 printk(KERN_ERR "ERROR: repeated CPUs\n");
7945 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7947 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7949 printk(KERN_CONT " %s", str);
7950 if (group->cpu_power != SCHED_LOAD_SCALE) {
7951 printk(KERN_CONT " (cpu_power = %d)",
7955 group = group->next;
7956 } while (group != sd->groups);
7957 printk(KERN_CONT "\n");
7959 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7960 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7963 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7964 printk(KERN_ERR "ERROR: parent span is not a superset "
7965 "of domain->span\n");
7969 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7971 cpumask_var_t groupmask;
7975 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7979 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7981 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7982 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7987 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7994 free_cpumask_var(groupmask);
7996 #else /* !CONFIG_SCHED_DEBUG */
7997 # define sched_domain_debug(sd, cpu) do { } while (0)
7998 #endif /* CONFIG_SCHED_DEBUG */
8000 static int sd_degenerate(struct sched_domain *sd)
8002 if (cpumask_weight(sched_domain_span(sd)) == 1)
8005 /* Following flags need at least 2 groups */
8006 if (sd->flags & (SD_LOAD_BALANCE |
8007 SD_BALANCE_NEWIDLE |
8011 SD_SHARE_PKG_RESOURCES)) {
8012 if (sd->groups != sd->groups->next)
8016 /* Following flags don't use groups */
8017 if (sd->flags & (SD_WAKE_AFFINE))
8024 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
8026 unsigned long cflags = sd->flags, pflags = parent->flags;
8028 if (sd_degenerate(parent))
8031 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
8034 /* Flags needing groups don't count if only 1 group in parent */
8035 if (parent->groups == parent->groups->next) {
8036 pflags &= ~(SD_LOAD_BALANCE |
8037 SD_BALANCE_NEWIDLE |
8041 SD_SHARE_PKG_RESOURCES);
8042 if (nr_node_ids == 1)
8043 pflags &= ~SD_SERIALIZE;
8045 if (~cflags & pflags)
8051 static void free_rootdomain(struct root_domain *rd)
8053 synchronize_sched();
8055 cpupri_cleanup(&rd->cpupri);
8057 free_cpumask_var(rd->rto_mask);
8058 free_cpumask_var(rd->online);
8059 free_cpumask_var(rd->span);
8063 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
8065 struct root_domain *old_rd = NULL;
8066 unsigned long flags;
8068 spin_lock_irqsave(&rq->lock, flags);
8073 if (cpumask_test_cpu(rq->cpu, old_rd->online))
8076 cpumask_clear_cpu(rq->cpu, old_rd->span);
8079 * If we dont want to free the old_rt yet then
8080 * set old_rd to NULL to skip the freeing later
8083 if (!atomic_dec_and_test(&old_rd->refcount))
8087 atomic_inc(&rd->refcount);
8090 cpumask_set_cpu(rq->cpu, rd->span);
8091 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
8094 spin_unlock_irqrestore(&rq->lock, flags);
8097 free_rootdomain(old_rd);
8100 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8102 gfp_t gfp = GFP_KERNEL;
8104 memset(rd, 0, sizeof(*rd));
8109 if (!alloc_cpumask_var(&rd->span, gfp))
8111 if (!alloc_cpumask_var(&rd->online, gfp))
8113 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8116 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8121 free_cpumask_var(rd->rto_mask);
8123 free_cpumask_var(rd->online);
8125 free_cpumask_var(rd->span);
8130 static void init_defrootdomain(void)
8132 init_rootdomain(&def_root_domain, true);
8134 atomic_set(&def_root_domain.refcount, 1);
8137 static struct root_domain *alloc_rootdomain(void)
8139 struct root_domain *rd;
8141 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8145 if (init_rootdomain(rd, false) != 0) {
8154 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8155 * hold the hotplug lock.
8158 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8160 struct rq *rq = cpu_rq(cpu);
8161 struct sched_domain *tmp;
8163 for (tmp = sd; tmp; tmp = tmp->parent)
8164 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
8166 /* Remove the sched domains which do not contribute to scheduling. */
8167 for (tmp = sd; tmp; ) {
8168 struct sched_domain *parent = tmp->parent;
8172 if (sd_parent_degenerate(tmp, parent)) {
8173 tmp->parent = parent->parent;
8175 parent->parent->child = tmp;
8180 if (sd && sd_degenerate(sd)) {
8186 sched_domain_debug(sd, cpu);
8188 rq_attach_root(rq, rd);
8189 rcu_assign_pointer(rq->sd, sd);
8192 /* cpus with isolated domains */
8193 static cpumask_var_t cpu_isolated_map;
8195 /* Setup the mask of cpus configured for isolated domains */
8196 static int __init isolated_cpu_setup(char *str)
8198 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8199 cpulist_parse(str, cpu_isolated_map);
8203 __setup("isolcpus=", isolated_cpu_setup);
8206 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8207 * to a function which identifies what group(along with sched group) a CPU
8208 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8209 * (due to the fact that we keep track of groups covered with a struct cpumask).
8211 * init_sched_build_groups will build a circular linked list of the groups
8212 * covered by the given span, and will set each group's ->cpumask correctly,
8213 * and ->cpu_power to 0.
8216 init_sched_build_groups(const struct cpumask *span,
8217 const struct cpumask *cpu_map,
8218 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8219 struct sched_group **sg,
8220 struct cpumask *tmpmask),
8221 struct cpumask *covered, struct cpumask *tmpmask)
8223 struct sched_group *first = NULL, *last = NULL;
8226 cpumask_clear(covered);
8228 for_each_cpu(i, span) {
8229 struct sched_group *sg;
8230 int group = group_fn(i, cpu_map, &sg, tmpmask);
8233 if (cpumask_test_cpu(i, covered))
8236 cpumask_clear(sched_group_cpus(sg));
8239 for_each_cpu(j, span) {
8240 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8243 cpumask_set_cpu(j, covered);
8244 cpumask_set_cpu(j, sched_group_cpus(sg));
8255 #define SD_NODES_PER_DOMAIN 16
8260 * find_next_best_node - find the next node to include in a sched_domain
8261 * @node: node whose sched_domain we're building
8262 * @used_nodes: nodes already in the sched_domain
8264 * Find the next node to include in a given scheduling domain. Simply
8265 * finds the closest node not already in the @used_nodes map.
8267 * Should use nodemask_t.
8269 static int find_next_best_node(int node, nodemask_t *used_nodes)
8271 int i, n, val, min_val, best_node = 0;
8275 for (i = 0; i < nr_node_ids; i++) {
8276 /* Start at @node */
8277 n = (node + i) % nr_node_ids;
8279 if (!nr_cpus_node(n))
8282 /* Skip already used nodes */
8283 if (node_isset(n, *used_nodes))
8286 /* Simple min distance search */
8287 val = node_distance(node, n);
8289 if (val < min_val) {
8295 node_set(best_node, *used_nodes);
8300 * sched_domain_node_span - get a cpumask for a node's sched_domain
8301 * @node: node whose cpumask we're constructing
8302 * @span: resulting cpumask
8304 * Given a node, construct a good cpumask for its sched_domain to span. It
8305 * should be one that prevents unnecessary balancing, but also spreads tasks
8308 static void sched_domain_node_span(int node, struct cpumask *span)
8310 nodemask_t used_nodes;
8313 cpumask_clear(span);
8314 nodes_clear(used_nodes);
8316 cpumask_or(span, span, cpumask_of_node(node));
8317 node_set(node, used_nodes);
8319 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8320 int next_node = find_next_best_node(node, &used_nodes);
8322 cpumask_or(span, span, cpumask_of_node(next_node));
8325 #endif /* CONFIG_NUMA */
8327 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8330 * The cpus mask in sched_group and sched_domain hangs off the end.
8332 * ( See the the comments in include/linux/sched.h:struct sched_group
8333 * and struct sched_domain. )
8335 struct static_sched_group {
8336 struct sched_group sg;
8337 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8340 struct static_sched_domain {
8341 struct sched_domain sd;
8342 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8348 cpumask_var_t domainspan;
8349 cpumask_var_t covered;
8350 cpumask_var_t notcovered;
8352 cpumask_var_t nodemask;
8353 cpumask_var_t this_sibling_map;
8354 cpumask_var_t this_core_map;
8355 cpumask_var_t send_covered;
8356 cpumask_var_t tmpmask;
8357 struct sched_group **sched_group_nodes;
8358 struct root_domain *rd;
8362 sa_sched_groups = 0,
8367 sa_this_sibling_map,
8369 sa_sched_group_nodes,
8379 * SMT sched-domains:
8381 #ifdef CONFIG_SCHED_SMT
8382 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8383 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8386 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8387 struct sched_group **sg, struct cpumask *unused)
8390 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8393 #endif /* CONFIG_SCHED_SMT */
8396 * multi-core sched-domains:
8398 #ifdef CONFIG_SCHED_MC
8399 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8400 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8401 #endif /* CONFIG_SCHED_MC */
8403 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8405 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8406 struct sched_group **sg, struct cpumask *mask)
8410 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8411 group = cpumask_first(mask);
8413 *sg = &per_cpu(sched_group_core, group).sg;
8416 #elif defined(CONFIG_SCHED_MC)
8418 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8419 struct sched_group **sg, struct cpumask *unused)
8422 *sg = &per_cpu(sched_group_core, cpu).sg;
8427 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8428 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8431 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8432 struct sched_group **sg, struct cpumask *mask)
8435 #ifdef CONFIG_SCHED_MC
8436 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8437 group = cpumask_first(mask);
8438 #elif defined(CONFIG_SCHED_SMT)
8439 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8440 group = cpumask_first(mask);
8445 *sg = &per_cpu(sched_group_phys, group).sg;
8451 * The init_sched_build_groups can't handle what we want to do with node
8452 * groups, so roll our own. Now each node has its own list of groups which
8453 * gets dynamically allocated.
8455 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8456 static struct sched_group ***sched_group_nodes_bycpu;
8458 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8459 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8461 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8462 struct sched_group **sg,
8463 struct cpumask *nodemask)
8467 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8468 group = cpumask_first(nodemask);
8471 *sg = &per_cpu(sched_group_allnodes, group).sg;
8475 static void init_numa_sched_groups_power(struct sched_group *group_head)
8477 struct sched_group *sg = group_head;
8483 for_each_cpu(j, sched_group_cpus(sg)) {
8484 struct sched_domain *sd;
8486 sd = &per_cpu(phys_domains, j).sd;
8487 if (j != group_first_cpu(sd->groups)) {
8489 * Only add "power" once for each
8495 sg->cpu_power += sd->groups->cpu_power;
8498 } while (sg != group_head);
8501 static int build_numa_sched_groups(struct s_data *d,
8502 const struct cpumask *cpu_map, int num)
8504 struct sched_domain *sd;
8505 struct sched_group *sg, *prev;
8508 cpumask_clear(d->covered);
8509 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8510 if (cpumask_empty(d->nodemask)) {
8511 d->sched_group_nodes[num] = NULL;
8515 sched_domain_node_span(num, d->domainspan);
8516 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8518 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8521 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8525 d->sched_group_nodes[num] = sg;
8527 for_each_cpu(j, d->nodemask) {
8528 sd = &per_cpu(node_domains, j).sd;
8533 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8535 cpumask_or(d->covered, d->covered, d->nodemask);
8538 for (j = 0; j < nr_node_ids; j++) {
8539 n = (num + j) % nr_node_ids;
8540 cpumask_complement(d->notcovered, d->covered);
8541 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8542 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8543 if (cpumask_empty(d->tmpmask))
8545 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8546 if (cpumask_empty(d->tmpmask))
8548 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8552 "Can not alloc domain group for node %d\n", j);
8556 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8557 sg->next = prev->next;
8558 cpumask_or(d->covered, d->covered, d->tmpmask);
8565 #endif /* CONFIG_NUMA */
8568 /* Free memory allocated for various sched_group structures */
8569 static void free_sched_groups(const struct cpumask *cpu_map,
8570 struct cpumask *nodemask)
8574 for_each_cpu(cpu, cpu_map) {
8575 struct sched_group **sched_group_nodes
8576 = sched_group_nodes_bycpu[cpu];
8578 if (!sched_group_nodes)
8581 for (i = 0; i < nr_node_ids; i++) {
8582 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8584 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8585 if (cpumask_empty(nodemask))
8595 if (oldsg != sched_group_nodes[i])
8598 kfree(sched_group_nodes);
8599 sched_group_nodes_bycpu[cpu] = NULL;
8602 #else /* !CONFIG_NUMA */
8603 static void free_sched_groups(const struct cpumask *cpu_map,
8604 struct cpumask *nodemask)
8607 #endif /* CONFIG_NUMA */
8610 * Initialize sched groups cpu_power.
8612 * cpu_power indicates the capacity of sched group, which is used while
8613 * distributing the load between different sched groups in a sched domain.
8614 * Typically cpu_power for all the groups in a sched domain will be same unless
8615 * there are asymmetries in the topology. If there are asymmetries, group
8616 * having more cpu_power will pickup more load compared to the group having
8619 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8621 struct sched_domain *child;
8622 struct sched_group *group;
8626 WARN_ON(!sd || !sd->groups);
8628 if (cpu != group_first_cpu(sd->groups))
8633 sd->groups->cpu_power = 0;
8636 power = SCHED_LOAD_SCALE;
8637 weight = cpumask_weight(sched_domain_span(sd));
8639 * SMT siblings share the power of a single core.
8640 * Usually multiple threads get a better yield out of
8641 * that one core than a single thread would have,
8642 * reflect that in sd->smt_gain.
8644 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8645 power *= sd->smt_gain;
8647 power >>= SCHED_LOAD_SHIFT;
8649 sd->groups->cpu_power += power;
8654 * Add cpu_power of each child group to this groups cpu_power.
8656 group = child->groups;
8658 sd->groups->cpu_power += group->cpu_power;
8659 group = group->next;
8660 } while (group != child->groups);
8664 * Initializers for schedule domains
8665 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8668 #ifdef CONFIG_SCHED_DEBUG
8669 # define SD_INIT_NAME(sd, type) sd->name = #type
8671 # define SD_INIT_NAME(sd, type) do { } while (0)
8674 #define SD_INIT(sd, type) sd_init_##type(sd)
8676 #define SD_INIT_FUNC(type) \
8677 static noinline void sd_init_##type(struct sched_domain *sd) \
8679 memset(sd, 0, sizeof(*sd)); \
8680 *sd = SD_##type##_INIT; \
8681 sd->level = SD_LV_##type; \
8682 SD_INIT_NAME(sd, type); \
8687 SD_INIT_FUNC(ALLNODES)
8690 #ifdef CONFIG_SCHED_SMT
8691 SD_INIT_FUNC(SIBLING)
8693 #ifdef CONFIG_SCHED_MC
8697 static int default_relax_domain_level = -1;
8699 static int __init setup_relax_domain_level(char *str)
8703 val = simple_strtoul(str, NULL, 0);
8704 if (val < SD_LV_MAX)
8705 default_relax_domain_level = val;
8709 __setup("relax_domain_level=", setup_relax_domain_level);
8711 static void set_domain_attribute(struct sched_domain *sd,
8712 struct sched_domain_attr *attr)
8716 if (!attr || attr->relax_domain_level < 0) {
8717 if (default_relax_domain_level < 0)
8720 request = default_relax_domain_level;
8722 request = attr->relax_domain_level;
8723 if (request < sd->level) {
8724 /* turn off idle balance on this domain */
8725 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8727 /* turn on idle balance on this domain */
8728 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8732 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8733 const struct cpumask *cpu_map)
8736 case sa_sched_groups:
8737 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8738 d->sched_group_nodes = NULL;
8740 free_rootdomain(d->rd); /* fall through */
8742 free_cpumask_var(d->tmpmask); /* fall through */
8743 case sa_send_covered:
8744 free_cpumask_var(d->send_covered); /* fall through */
8745 case sa_this_core_map:
8746 free_cpumask_var(d->this_core_map); /* fall through */
8747 case sa_this_sibling_map:
8748 free_cpumask_var(d->this_sibling_map); /* fall through */
8750 free_cpumask_var(d->nodemask); /* fall through */
8751 case sa_sched_group_nodes:
8753 kfree(d->sched_group_nodes); /* fall through */
8755 free_cpumask_var(d->notcovered); /* fall through */
8757 free_cpumask_var(d->covered); /* fall through */
8759 free_cpumask_var(d->domainspan); /* fall through */
8766 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8767 const struct cpumask *cpu_map)
8770 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8772 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8773 return sa_domainspan;
8774 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8776 /* Allocate the per-node list of sched groups */
8777 d->sched_group_nodes = kcalloc(nr_node_ids,
8778 sizeof(struct sched_group *), GFP_KERNEL);
8779 if (!d->sched_group_nodes) {
8780 printk(KERN_WARNING "Can not alloc sched group node list\n");
8781 return sa_notcovered;
8783 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8785 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8786 return sa_sched_group_nodes;
8787 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8789 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8790 return sa_this_sibling_map;
8791 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8792 return sa_this_core_map;
8793 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8794 return sa_send_covered;
8795 d->rd = alloc_rootdomain();
8797 printk(KERN_WARNING "Cannot alloc root domain\n");
8800 return sa_rootdomain;
8803 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8804 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8806 struct sched_domain *sd = NULL;
8808 struct sched_domain *parent;
8811 if (cpumask_weight(cpu_map) >
8812 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8813 sd = &per_cpu(allnodes_domains, i).sd;
8814 SD_INIT(sd, ALLNODES);
8815 set_domain_attribute(sd, attr);
8816 cpumask_copy(sched_domain_span(sd), cpu_map);
8817 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8822 sd = &per_cpu(node_domains, i).sd;
8824 set_domain_attribute(sd, attr);
8825 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8826 sd->parent = parent;
8829 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8834 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8835 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8836 struct sched_domain *parent, int i)
8838 struct sched_domain *sd;
8839 sd = &per_cpu(phys_domains, i).sd;
8841 set_domain_attribute(sd, attr);
8842 cpumask_copy(sched_domain_span(sd), d->nodemask);
8843 sd->parent = parent;
8846 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8850 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8851 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8852 struct sched_domain *parent, int i)
8854 struct sched_domain *sd = parent;
8855 #ifdef CONFIG_SCHED_MC
8856 sd = &per_cpu(core_domains, i).sd;
8858 set_domain_attribute(sd, attr);
8859 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8860 sd->parent = parent;
8862 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8867 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8868 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8869 struct sched_domain *parent, int i)
8871 struct sched_domain *sd = parent;
8872 #ifdef CONFIG_SCHED_SMT
8873 sd = &per_cpu(cpu_domains, i).sd;
8874 SD_INIT(sd, SIBLING);
8875 set_domain_attribute(sd, attr);
8876 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8877 sd->parent = parent;
8879 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8884 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8885 const struct cpumask *cpu_map, int cpu)
8888 #ifdef CONFIG_SCHED_SMT
8889 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8890 cpumask_and(d->this_sibling_map, cpu_map,
8891 topology_thread_cpumask(cpu));
8892 if (cpu == cpumask_first(d->this_sibling_map))
8893 init_sched_build_groups(d->this_sibling_map, cpu_map,
8895 d->send_covered, d->tmpmask);
8898 #ifdef CONFIG_SCHED_MC
8899 case SD_LV_MC: /* set up multi-core groups */
8900 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8901 if (cpu == cpumask_first(d->this_core_map))
8902 init_sched_build_groups(d->this_core_map, cpu_map,
8904 d->send_covered, d->tmpmask);
8907 case SD_LV_CPU: /* set up physical groups */
8908 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8909 if (!cpumask_empty(d->nodemask))
8910 init_sched_build_groups(d->nodemask, cpu_map,
8912 d->send_covered, d->tmpmask);
8915 case SD_LV_ALLNODES:
8916 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8917 d->send_covered, d->tmpmask);
8926 * Build sched domains for a given set of cpus and attach the sched domains
8927 * to the individual cpus
8929 static int __build_sched_domains(const struct cpumask *cpu_map,
8930 struct sched_domain_attr *attr)
8932 enum s_alloc alloc_state = sa_none;
8934 struct sched_domain *sd;
8940 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8941 if (alloc_state != sa_rootdomain)
8943 alloc_state = sa_sched_groups;
8946 * Set up domains for cpus specified by the cpu_map.
8948 for_each_cpu(i, cpu_map) {
8949 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8952 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8953 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8954 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8955 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8958 for_each_cpu(i, cpu_map) {
8959 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8960 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8963 /* Set up physical groups */
8964 for (i = 0; i < nr_node_ids; i++)
8965 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8968 /* Set up node groups */
8970 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8972 for (i = 0; i < nr_node_ids; i++)
8973 if (build_numa_sched_groups(&d, cpu_map, i))
8977 /* Calculate CPU power for physical packages and nodes */
8978 #ifdef CONFIG_SCHED_SMT
8979 for_each_cpu(i, cpu_map) {
8980 sd = &per_cpu(cpu_domains, i).sd;
8981 init_sched_groups_power(i, sd);
8984 #ifdef CONFIG_SCHED_MC
8985 for_each_cpu(i, cpu_map) {
8986 sd = &per_cpu(core_domains, i).sd;
8987 init_sched_groups_power(i, sd);
8991 for_each_cpu(i, cpu_map) {
8992 sd = &per_cpu(phys_domains, i).sd;
8993 init_sched_groups_power(i, sd);
8997 for (i = 0; i < nr_node_ids; i++)
8998 init_numa_sched_groups_power(d.sched_group_nodes[i]);
9000 if (d.sd_allnodes) {
9001 struct sched_group *sg;
9003 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
9005 init_numa_sched_groups_power(sg);
9009 /* Attach the domains */
9010 for_each_cpu(i, cpu_map) {
9011 #ifdef CONFIG_SCHED_SMT
9012 sd = &per_cpu(cpu_domains, i).sd;
9013 #elif defined(CONFIG_SCHED_MC)
9014 sd = &per_cpu(core_domains, i).sd;
9016 sd = &per_cpu(phys_domains, i).sd;
9018 cpu_attach_domain(sd, d.rd, i);
9021 d.sched_group_nodes = NULL; /* don't free this we still need it */
9022 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
9026 __free_domain_allocs(&d, alloc_state, cpu_map);
9030 static int build_sched_domains(const struct cpumask *cpu_map)
9032 return __build_sched_domains(cpu_map, NULL);
9035 static struct cpumask *doms_cur; /* current sched domains */
9036 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
9037 static struct sched_domain_attr *dattr_cur;
9038 /* attribues of custom domains in 'doms_cur' */
9041 * Special case: If a kmalloc of a doms_cur partition (array of
9042 * cpumask) fails, then fallback to a single sched domain,
9043 * as determined by the single cpumask fallback_doms.
9045 static cpumask_var_t fallback_doms;
9048 * arch_update_cpu_topology lets virtualized architectures update the
9049 * cpu core maps. It is supposed to return 1 if the topology changed
9050 * or 0 if it stayed the same.
9052 int __attribute__((weak)) arch_update_cpu_topology(void)
9058 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9059 * For now this just excludes isolated cpus, but could be used to
9060 * exclude other special cases in the future.
9062 static int arch_init_sched_domains(const struct cpumask *cpu_map)
9066 arch_update_cpu_topology();
9068 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
9070 doms_cur = fallback_doms;
9071 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
9073 err = build_sched_domains(doms_cur);
9074 register_sched_domain_sysctl();
9079 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9080 struct cpumask *tmpmask)
9082 free_sched_groups(cpu_map, tmpmask);
9086 * Detach sched domains from a group of cpus specified in cpu_map
9087 * These cpus will now be attached to the NULL domain
9089 static void detach_destroy_domains(const struct cpumask *cpu_map)
9091 /* Save because hotplug lock held. */
9092 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9095 for_each_cpu(i, cpu_map)
9096 cpu_attach_domain(NULL, &def_root_domain, i);
9097 synchronize_sched();
9098 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9101 /* handle null as "default" */
9102 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9103 struct sched_domain_attr *new, int idx_new)
9105 struct sched_domain_attr tmp;
9112 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9113 new ? (new + idx_new) : &tmp,
9114 sizeof(struct sched_domain_attr));
9118 * Partition sched domains as specified by the 'ndoms_new'
9119 * cpumasks in the array doms_new[] of cpumasks. This compares
9120 * doms_new[] to the current sched domain partitioning, doms_cur[].
9121 * It destroys each deleted domain and builds each new domain.
9123 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
9124 * The masks don't intersect (don't overlap.) We should setup one
9125 * sched domain for each mask. CPUs not in any of the cpumasks will
9126 * not be load balanced. If the same cpumask appears both in the
9127 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9130 * The passed in 'doms_new' should be kmalloc'd. This routine takes
9131 * ownership of it and will kfree it when done with it. If the caller
9132 * failed the kmalloc call, then it can pass in doms_new == NULL &&
9133 * ndoms_new == 1, and partition_sched_domains() will fallback to
9134 * the single partition 'fallback_doms', it also forces the domains
9137 * If doms_new == NULL it will be replaced with cpu_online_mask.
9138 * ndoms_new == 0 is a special case for destroying existing domains,
9139 * and it will not create the default domain.
9141 * Call with hotplug lock held
9143 /* FIXME: Change to struct cpumask *doms_new[] */
9144 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
9145 struct sched_domain_attr *dattr_new)
9150 mutex_lock(&sched_domains_mutex);
9152 /* always unregister in case we don't destroy any domains */
9153 unregister_sched_domain_sysctl();
9155 /* Let architecture update cpu core mappings. */
9156 new_topology = arch_update_cpu_topology();
9158 n = doms_new ? ndoms_new : 0;
9160 /* Destroy deleted domains */
9161 for (i = 0; i < ndoms_cur; i++) {
9162 for (j = 0; j < n && !new_topology; j++) {
9163 if (cpumask_equal(&doms_cur[i], &doms_new[j])
9164 && dattrs_equal(dattr_cur, i, dattr_new, j))
9167 /* no match - a current sched domain not in new doms_new[] */
9168 detach_destroy_domains(doms_cur + i);
9173 if (doms_new == NULL) {
9175 doms_new = fallback_doms;
9176 cpumask_andnot(&doms_new[0], cpu_active_mask, cpu_isolated_map);
9177 WARN_ON_ONCE(dattr_new);
9180 /* Build new domains */
9181 for (i = 0; i < ndoms_new; i++) {
9182 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9183 if (cpumask_equal(&doms_new[i], &doms_cur[j])
9184 && dattrs_equal(dattr_new, i, dattr_cur, j))
9187 /* no match - add a new doms_new */
9188 __build_sched_domains(doms_new + i,
9189 dattr_new ? dattr_new + i : NULL);
9194 /* Remember the new sched domains */
9195 if (doms_cur != fallback_doms)
9197 kfree(dattr_cur); /* kfree(NULL) is safe */
9198 doms_cur = doms_new;
9199 dattr_cur = dattr_new;
9200 ndoms_cur = ndoms_new;
9202 register_sched_domain_sysctl();
9204 mutex_unlock(&sched_domains_mutex);
9207 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9208 static void arch_reinit_sched_domains(void)
9212 /* Destroy domains first to force the rebuild */
9213 partition_sched_domains(0, NULL, NULL);
9215 rebuild_sched_domains();
9219 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9221 unsigned int level = 0;
9223 if (sscanf(buf, "%u", &level) != 1)
9227 * level is always be positive so don't check for
9228 * level < POWERSAVINGS_BALANCE_NONE which is 0
9229 * What happens on 0 or 1 byte write,
9230 * need to check for count as well?
9233 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9237 sched_smt_power_savings = level;
9239 sched_mc_power_savings = level;
9241 arch_reinit_sched_domains();
9246 #ifdef CONFIG_SCHED_MC
9247 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9250 return sprintf(page, "%u\n", sched_mc_power_savings);
9252 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9253 const char *buf, size_t count)
9255 return sched_power_savings_store(buf, count, 0);
9257 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9258 sched_mc_power_savings_show,
9259 sched_mc_power_savings_store);
9262 #ifdef CONFIG_SCHED_SMT
9263 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9266 return sprintf(page, "%u\n", sched_smt_power_savings);
9268 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9269 const char *buf, size_t count)
9271 return sched_power_savings_store(buf, count, 1);
9273 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9274 sched_smt_power_savings_show,
9275 sched_smt_power_savings_store);
9278 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9282 #ifdef CONFIG_SCHED_SMT
9284 err = sysfs_create_file(&cls->kset.kobj,
9285 &attr_sched_smt_power_savings.attr);
9287 #ifdef CONFIG_SCHED_MC
9288 if (!err && mc_capable())
9289 err = sysfs_create_file(&cls->kset.kobj,
9290 &attr_sched_mc_power_savings.attr);
9294 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9296 #ifndef CONFIG_CPUSETS
9298 * Add online and remove offline CPUs from the scheduler domains.
9299 * When cpusets are enabled they take over this function.
9301 static int update_sched_domains(struct notifier_block *nfb,
9302 unsigned long action, void *hcpu)
9306 case CPU_ONLINE_FROZEN:
9307 case CPU_DOWN_PREPARE:
9308 case CPU_DOWN_PREPARE_FROZEN:
9309 case CPU_DOWN_FAILED:
9310 case CPU_DOWN_FAILED_FROZEN:
9311 partition_sched_domains(1, NULL, NULL);
9320 static int update_runtime(struct notifier_block *nfb,
9321 unsigned long action, void *hcpu)
9323 int cpu = (int)(long)hcpu;
9326 case CPU_DOWN_PREPARE:
9327 case CPU_DOWN_PREPARE_FROZEN:
9328 disable_runtime(cpu_rq(cpu));
9331 case CPU_DOWN_FAILED:
9332 case CPU_DOWN_FAILED_FROZEN:
9334 case CPU_ONLINE_FROZEN:
9335 enable_runtime(cpu_rq(cpu));
9343 void __init sched_init_smp(void)
9345 cpumask_var_t non_isolated_cpus;
9347 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9348 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9350 #if defined(CONFIG_NUMA)
9351 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9353 BUG_ON(sched_group_nodes_bycpu == NULL);
9356 mutex_lock(&sched_domains_mutex);
9357 arch_init_sched_domains(cpu_active_mask);
9358 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9359 if (cpumask_empty(non_isolated_cpus))
9360 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9361 mutex_unlock(&sched_domains_mutex);
9364 #ifndef CONFIG_CPUSETS
9365 /* XXX: Theoretical race here - CPU may be hotplugged now */
9366 hotcpu_notifier(update_sched_domains, 0);
9369 /* RT runtime code needs to handle some hotplug events */
9370 hotcpu_notifier(update_runtime, 0);
9374 /* Move init over to a non-isolated CPU */
9375 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9377 sched_init_granularity();
9378 free_cpumask_var(non_isolated_cpus);
9380 init_sched_rt_class();
9383 void __init sched_init_smp(void)
9385 sched_init_granularity();
9387 #endif /* CONFIG_SMP */
9389 const_debug unsigned int sysctl_timer_migration = 1;
9391 int in_sched_functions(unsigned long addr)
9393 return in_lock_functions(addr) ||
9394 (addr >= (unsigned long)__sched_text_start
9395 && addr < (unsigned long)__sched_text_end);
9398 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9400 cfs_rq->tasks_timeline = RB_ROOT;
9401 INIT_LIST_HEAD(&cfs_rq->tasks);
9402 #ifdef CONFIG_FAIR_GROUP_SCHED
9405 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9408 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9410 struct rt_prio_array *array;
9413 array = &rt_rq->active;
9414 for (i = 0; i < MAX_RT_PRIO; i++) {
9415 INIT_LIST_HEAD(array->queue + i);
9416 __clear_bit(i, array->bitmap);
9418 /* delimiter for bitsearch: */
9419 __set_bit(MAX_RT_PRIO, array->bitmap);
9421 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9422 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9424 rt_rq->highest_prio.next = MAX_RT_PRIO;
9428 rt_rq->rt_nr_migratory = 0;
9429 rt_rq->overloaded = 0;
9430 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9434 rt_rq->rt_throttled = 0;
9435 rt_rq->rt_runtime = 0;
9436 spin_lock_init(&rt_rq->rt_runtime_lock);
9438 #ifdef CONFIG_RT_GROUP_SCHED
9439 rt_rq->rt_nr_boosted = 0;
9444 #ifdef CONFIG_FAIR_GROUP_SCHED
9445 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9446 struct sched_entity *se, int cpu, int add,
9447 struct sched_entity *parent)
9449 struct rq *rq = cpu_rq(cpu);
9450 tg->cfs_rq[cpu] = cfs_rq;
9451 init_cfs_rq(cfs_rq, rq);
9454 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9457 /* se could be NULL for init_task_group */
9462 se->cfs_rq = &rq->cfs;
9464 se->cfs_rq = parent->my_q;
9467 se->load.weight = tg->shares;
9468 se->load.inv_weight = 0;
9469 se->parent = parent;
9473 #ifdef CONFIG_RT_GROUP_SCHED
9474 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9475 struct sched_rt_entity *rt_se, int cpu, int add,
9476 struct sched_rt_entity *parent)
9478 struct rq *rq = cpu_rq(cpu);
9480 tg->rt_rq[cpu] = rt_rq;
9481 init_rt_rq(rt_rq, rq);
9483 rt_rq->rt_se = rt_se;
9484 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9486 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9488 tg->rt_se[cpu] = rt_se;
9493 rt_se->rt_rq = &rq->rt;
9495 rt_se->rt_rq = parent->my_q;
9497 rt_se->my_q = rt_rq;
9498 rt_se->parent = parent;
9499 INIT_LIST_HEAD(&rt_se->run_list);
9503 void __init sched_init(void)
9506 unsigned long alloc_size = 0, ptr;
9508 #ifdef CONFIG_FAIR_GROUP_SCHED
9509 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9511 #ifdef CONFIG_RT_GROUP_SCHED
9512 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9514 #ifdef CONFIG_USER_SCHED
9517 #ifdef CONFIG_CPUMASK_OFFSTACK
9518 alloc_size += num_possible_cpus() * cpumask_size();
9521 * As sched_init() is called before page_alloc is setup,
9522 * we use alloc_bootmem().
9525 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9527 #ifdef CONFIG_FAIR_GROUP_SCHED
9528 init_task_group.se = (struct sched_entity **)ptr;
9529 ptr += nr_cpu_ids * sizeof(void **);
9531 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9532 ptr += nr_cpu_ids * sizeof(void **);
9534 #ifdef CONFIG_USER_SCHED
9535 root_task_group.se = (struct sched_entity **)ptr;
9536 ptr += nr_cpu_ids * sizeof(void **);
9538 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9539 ptr += nr_cpu_ids * sizeof(void **);
9540 #endif /* CONFIG_USER_SCHED */
9541 #endif /* CONFIG_FAIR_GROUP_SCHED */
9542 #ifdef CONFIG_RT_GROUP_SCHED
9543 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9544 ptr += nr_cpu_ids * sizeof(void **);
9546 init_task_group.rt_rq = (struct rt_rq **)ptr;
9547 ptr += nr_cpu_ids * sizeof(void **);
9549 #ifdef CONFIG_USER_SCHED
9550 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9551 ptr += nr_cpu_ids * sizeof(void **);
9553 root_task_group.rt_rq = (struct rt_rq **)ptr;
9554 ptr += nr_cpu_ids * sizeof(void **);
9555 #endif /* CONFIG_USER_SCHED */
9556 #endif /* CONFIG_RT_GROUP_SCHED */
9557 #ifdef CONFIG_CPUMASK_OFFSTACK
9558 for_each_possible_cpu(i) {
9559 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9560 ptr += cpumask_size();
9562 #endif /* CONFIG_CPUMASK_OFFSTACK */
9566 init_defrootdomain();
9569 init_rt_bandwidth(&def_rt_bandwidth,
9570 global_rt_period(), global_rt_runtime());
9572 #ifdef CONFIG_RT_GROUP_SCHED
9573 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9574 global_rt_period(), global_rt_runtime());
9575 #ifdef CONFIG_USER_SCHED
9576 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9577 global_rt_period(), RUNTIME_INF);
9578 #endif /* CONFIG_USER_SCHED */
9579 #endif /* CONFIG_RT_GROUP_SCHED */
9581 #ifdef CONFIG_GROUP_SCHED
9582 list_add(&init_task_group.list, &task_groups);
9583 INIT_LIST_HEAD(&init_task_group.children);
9585 #ifdef CONFIG_USER_SCHED
9586 INIT_LIST_HEAD(&root_task_group.children);
9587 init_task_group.parent = &root_task_group;
9588 list_add(&init_task_group.siblings, &root_task_group.children);
9589 #endif /* CONFIG_USER_SCHED */
9590 #endif /* CONFIG_GROUP_SCHED */
9592 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9593 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9594 __alignof__(unsigned long));
9596 for_each_possible_cpu(i) {
9600 spin_lock_init(&rq->lock);
9602 rq->calc_load_active = 0;
9603 rq->calc_load_update = jiffies + LOAD_FREQ;
9604 init_cfs_rq(&rq->cfs, rq);
9605 init_rt_rq(&rq->rt, rq);
9606 #ifdef CONFIG_FAIR_GROUP_SCHED
9607 init_task_group.shares = init_task_group_load;
9608 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9609 #ifdef CONFIG_CGROUP_SCHED
9611 * How much cpu bandwidth does init_task_group get?
9613 * In case of task-groups formed thr' the cgroup filesystem, it
9614 * gets 100% of the cpu resources in the system. This overall
9615 * system cpu resource is divided among the tasks of
9616 * init_task_group and its child task-groups in a fair manner,
9617 * based on each entity's (task or task-group's) weight
9618 * (se->load.weight).
9620 * In other words, if init_task_group has 10 tasks of weight
9621 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9622 * then A0's share of the cpu resource is:
9624 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9626 * We achieve this by letting init_task_group's tasks sit
9627 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9629 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9630 #elif defined CONFIG_USER_SCHED
9631 root_task_group.shares = NICE_0_LOAD;
9632 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9634 * In case of task-groups formed thr' the user id of tasks,
9635 * init_task_group represents tasks belonging to root user.
9636 * Hence it forms a sibling of all subsequent groups formed.
9637 * In this case, init_task_group gets only a fraction of overall
9638 * system cpu resource, based on the weight assigned to root
9639 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9640 * by letting tasks of init_task_group sit in a separate cfs_rq
9641 * (init_tg_cfs_rq) and having one entity represent this group of
9642 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9644 init_tg_cfs_entry(&init_task_group,
9645 &per_cpu(init_tg_cfs_rq, i),
9646 &per_cpu(init_sched_entity, i), i, 1,
9647 root_task_group.se[i]);
9650 #endif /* CONFIG_FAIR_GROUP_SCHED */
9652 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9653 #ifdef CONFIG_RT_GROUP_SCHED
9654 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9655 #ifdef CONFIG_CGROUP_SCHED
9656 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9657 #elif defined CONFIG_USER_SCHED
9658 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9659 init_tg_rt_entry(&init_task_group,
9660 &per_cpu(init_rt_rq, i),
9661 &per_cpu(init_sched_rt_entity, i), i, 1,
9662 root_task_group.rt_se[i]);
9666 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9667 rq->cpu_load[j] = 0;
9671 rq->post_schedule = 0;
9672 rq->active_balance = 0;
9673 rq->next_balance = jiffies;
9677 rq->migration_thread = NULL;
9679 rq->avg_idle = 2*sysctl_sched_migration_cost;
9680 INIT_LIST_HEAD(&rq->migration_queue);
9681 rq_attach_root(rq, &def_root_domain);
9684 atomic_set(&rq->nr_iowait, 0);
9687 set_load_weight(&init_task);
9689 #ifdef CONFIG_PREEMPT_NOTIFIERS
9690 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9694 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9697 #ifdef CONFIG_RT_MUTEXES
9698 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9702 * The boot idle thread does lazy MMU switching as well:
9704 atomic_inc(&init_mm.mm_count);
9705 enter_lazy_tlb(&init_mm, current);
9708 * Make us the idle thread. Technically, schedule() should not be
9709 * called from this thread, however somewhere below it might be,
9710 * but because we are the idle thread, we just pick up running again
9711 * when this runqueue becomes "idle".
9713 init_idle(current, smp_processor_id());
9715 calc_load_update = jiffies + LOAD_FREQ;
9718 * During early bootup we pretend to be a normal task:
9720 current->sched_class = &fair_sched_class;
9722 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9723 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9726 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9727 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9729 /* May be allocated at isolcpus cmdline parse time */
9730 if (cpu_isolated_map == NULL)
9731 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9736 scheduler_running = 1;
9739 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9740 static inline int preempt_count_equals(int preempt_offset)
9742 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9744 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9747 static int __might_sleep_init_called;
9748 int __init __might_sleep_init(void)
9750 __might_sleep_init_called = 1;
9753 early_initcall(__might_sleep_init);
9755 void __might_sleep(char *file, int line, int preempt_offset)
9758 static unsigned long prev_jiffy; /* ratelimiting */
9760 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9763 if (system_state != SYSTEM_RUNNING &&
9764 (!__might_sleep_init_called || system_state != SYSTEM_BOOTING))
9766 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9768 prev_jiffy = jiffies;
9771 "BUG: sleeping function called from invalid context at %s:%d\n",
9774 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9775 in_atomic(), irqs_disabled(),
9776 current->pid, current->comm);
9778 debug_show_held_locks(current);
9779 if (irqs_disabled())
9780 print_irqtrace_events(current);
9784 EXPORT_SYMBOL(__might_sleep);
9787 #ifdef CONFIG_MAGIC_SYSRQ
9788 static void normalize_task(struct rq *rq, struct task_struct *p)
9792 update_rq_clock(rq);
9793 on_rq = p->se.on_rq;
9795 deactivate_task(rq, p, 0);
9796 __setscheduler(rq, p, SCHED_NORMAL, 0);
9798 activate_task(rq, p, 0);
9799 resched_task(rq->curr);
9803 void normalize_rt_tasks(void)
9805 struct task_struct *g, *p;
9806 unsigned long flags;
9809 read_lock_irqsave(&tasklist_lock, flags);
9810 do_each_thread(g, p) {
9812 * Only normalize user tasks:
9817 p->se.exec_start = 0;
9818 #ifdef CONFIG_SCHEDSTATS
9819 p->se.wait_start = 0;
9820 p->se.sleep_start = 0;
9821 p->se.block_start = 0;
9826 * Renice negative nice level userspace
9829 if (TASK_NICE(p) < 0 && p->mm)
9830 set_user_nice(p, 0);
9834 spin_lock(&p->pi_lock);
9835 rq = __task_rq_lock(p);
9837 normalize_task(rq, p);
9839 __task_rq_unlock(rq);
9840 spin_unlock(&p->pi_lock);
9841 } while_each_thread(g, p);
9843 read_unlock_irqrestore(&tasklist_lock, flags);
9846 #endif /* CONFIG_MAGIC_SYSRQ */
9850 * These functions are only useful for the IA64 MCA handling.
9852 * They can only be called when the whole system has been
9853 * stopped - every CPU needs to be quiescent, and no scheduling
9854 * activity can take place. Using them for anything else would
9855 * be a serious bug, and as a result, they aren't even visible
9856 * under any other configuration.
9860 * curr_task - return the current task for a given cpu.
9861 * @cpu: the processor in question.
9863 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9865 struct task_struct *curr_task(int cpu)
9867 return cpu_curr(cpu);
9871 * set_curr_task - set the current task for a given cpu.
9872 * @cpu: the processor in question.
9873 * @p: the task pointer to set.
9875 * Description: This function must only be used when non-maskable interrupts
9876 * are serviced on a separate stack. It allows the architecture to switch the
9877 * notion of the current task on a cpu in a non-blocking manner. This function
9878 * must be called with all CPU's synchronized, and interrupts disabled, the
9879 * and caller must save the original value of the current task (see
9880 * curr_task() above) and restore that value before reenabling interrupts and
9881 * re-starting the system.
9883 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9885 void set_curr_task(int cpu, struct task_struct *p)
9892 #ifdef CONFIG_FAIR_GROUP_SCHED
9893 static void free_fair_sched_group(struct task_group *tg)
9897 for_each_possible_cpu(i) {
9899 kfree(tg->cfs_rq[i]);
9909 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9911 struct cfs_rq *cfs_rq;
9912 struct sched_entity *se;
9916 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9919 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9923 tg->shares = NICE_0_LOAD;
9925 for_each_possible_cpu(i) {
9928 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9929 GFP_KERNEL, cpu_to_node(i));
9933 se = kzalloc_node(sizeof(struct sched_entity),
9934 GFP_KERNEL, cpu_to_node(i));
9938 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9947 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9949 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9950 &cpu_rq(cpu)->leaf_cfs_rq_list);
9953 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9955 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9957 #else /* !CONFG_FAIR_GROUP_SCHED */
9958 static inline void free_fair_sched_group(struct task_group *tg)
9963 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9968 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9972 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9975 #endif /* CONFIG_FAIR_GROUP_SCHED */
9977 #ifdef CONFIG_RT_GROUP_SCHED
9978 static void free_rt_sched_group(struct task_group *tg)
9982 destroy_rt_bandwidth(&tg->rt_bandwidth);
9984 for_each_possible_cpu(i) {
9986 kfree(tg->rt_rq[i]);
9988 kfree(tg->rt_se[i]);
9996 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9998 struct rt_rq *rt_rq;
9999 struct sched_rt_entity *rt_se;
10003 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
10006 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
10010 init_rt_bandwidth(&tg->rt_bandwidth,
10011 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
10013 for_each_possible_cpu(i) {
10016 rt_rq = kzalloc_node(sizeof(struct rt_rq),
10017 GFP_KERNEL, cpu_to_node(i));
10021 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
10022 GFP_KERNEL, cpu_to_node(i));
10026 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
10035 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10037 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
10038 &cpu_rq(cpu)->leaf_rt_rq_list);
10041 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10043 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
10045 #else /* !CONFIG_RT_GROUP_SCHED */
10046 static inline void free_rt_sched_group(struct task_group *tg)
10051 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
10056 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10060 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10063 #endif /* CONFIG_RT_GROUP_SCHED */
10065 #ifdef CONFIG_GROUP_SCHED
10066 static void free_sched_group(struct task_group *tg)
10068 free_fair_sched_group(tg);
10069 free_rt_sched_group(tg);
10073 /* allocate runqueue etc for a new task group */
10074 struct task_group *sched_create_group(struct task_group *parent)
10076 struct task_group *tg;
10077 unsigned long flags;
10080 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
10082 return ERR_PTR(-ENOMEM);
10084 if (!alloc_fair_sched_group(tg, parent))
10087 if (!alloc_rt_sched_group(tg, parent))
10090 spin_lock_irqsave(&task_group_lock, flags);
10091 for_each_possible_cpu(i) {
10092 register_fair_sched_group(tg, i);
10093 register_rt_sched_group(tg, i);
10095 list_add_rcu(&tg->list, &task_groups);
10097 WARN_ON(!parent); /* root should already exist */
10099 tg->parent = parent;
10100 INIT_LIST_HEAD(&tg->children);
10101 list_add_rcu(&tg->siblings, &parent->children);
10102 spin_unlock_irqrestore(&task_group_lock, flags);
10107 free_sched_group(tg);
10108 return ERR_PTR(-ENOMEM);
10111 /* rcu callback to free various structures associated with a task group */
10112 static void free_sched_group_rcu(struct rcu_head *rhp)
10114 /* now it should be safe to free those cfs_rqs */
10115 free_sched_group(container_of(rhp, struct task_group, rcu));
10118 /* Destroy runqueue etc associated with a task group */
10119 void sched_destroy_group(struct task_group *tg)
10121 unsigned long flags;
10124 spin_lock_irqsave(&task_group_lock, flags);
10125 for_each_possible_cpu(i) {
10126 unregister_fair_sched_group(tg, i);
10127 unregister_rt_sched_group(tg, i);
10129 list_del_rcu(&tg->list);
10130 list_del_rcu(&tg->siblings);
10131 spin_unlock_irqrestore(&task_group_lock, flags);
10133 /* wait for possible concurrent references to cfs_rqs complete */
10134 call_rcu(&tg->rcu, free_sched_group_rcu);
10137 /* change task's runqueue when it moves between groups.
10138 * The caller of this function should have put the task in its new group
10139 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10140 * reflect its new group.
10142 void sched_move_task(struct task_struct *tsk)
10144 int on_rq, running;
10145 unsigned long flags;
10148 rq = task_rq_lock(tsk, &flags);
10150 update_rq_clock(rq);
10152 running = task_current(rq, tsk);
10153 on_rq = tsk->se.on_rq;
10156 dequeue_task(rq, tsk, 0);
10157 if (unlikely(running))
10158 tsk->sched_class->put_prev_task(rq, tsk);
10160 set_task_rq(tsk, task_cpu(tsk));
10162 #ifdef CONFIG_FAIR_GROUP_SCHED
10163 if (tsk->sched_class->moved_group)
10164 tsk->sched_class->moved_group(tsk, on_rq);
10167 if (unlikely(running))
10168 tsk->sched_class->set_curr_task(rq);
10170 enqueue_task(rq, tsk, 0, false);
10172 task_rq_unlock(rq, &flags);
10174 #endif /* CONFIG_GROUP_SCHED */
10176 #ifdef CONFIG_FAIR_GROUP_SCHED
10177 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10179 struct cfs_rq *cfs_rq = se->cfs_rq;
10184 dequeue_entity(cfs_rq, se, 0);
10186 se->load.weight = shares;
10187 se->load.inv_weight = 0;
10190 enqueue_entity(cfs_rq, se, 0);
10193 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10195 struct cfs_rq *cfs_rq = se->cfs_rq;
10196 struct rq *rq = cfs_rq->rq;
10197 unsigned long flags;
10199 spin_lock_irqsave(&rq->lock, flags);
10200 __set_se_shares(se, shares);
10201 spin_unlock_irqrestore(&rq->lock, flags);
10204 static DEFINE_MUTEX(shares_mutex);
10206 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10209 unsigned long flags;
10212 * We can't change the weight of the root cgroup.
10217 if (shares < MIN_SHARES)
10218 shares = MIN_SHARES;
10219 else if (shares > MAX_SHARES)
10220 shares = MAX_SHARES;
10222 mutex_lock(&shares_mutex);
10223 if (tg->shares == shares)
10226 spin_lock_irqsave(&task_group_lock, flags);
10227 for_each_possible_cpu(i)
10228 unregister_fair_sched_group(tg, i);
10229 list_del_rcu(&tg->siblings);
10230 spin_unlock_irqrestore(&task_group_lock, flags);
10232 /* wait for any ongoing reference to this group to finish */
10233 synchronize_sched();
10236 * Now we are free to modify the group's share on each cpu
10237 * w/o tripping rebalance_share or load_balance_fair.
10239 tg->shares = shares;
10240 for_each_possible_cpu(i) {
10242 * force a rebalance
10244 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10245 set_se_shares(tg->se[i], shares);
10249 * Enable load balance activity on this group, by inserting it back on
10250 * each cpu's rq->leaf_cfs_rq_list.
10252 spin_lock_irqsave(&task_group_lock, flags);
10253 for_each_possible_cpu(i)
10254 register_fair_sched_group(tg, i);
10255 list_add_rcu(&tg->siblings, &tg->parent->children);
10256 spin_unlock_irqrestore(&task_group_lock, flags);
10258 mutex_unlock(&shares_mutex);
10262 unsigned long sched_group_shares(struct task_group *tg)
10268 #ifdef CONFIG_RT_GROUP_SCHED
10270 * Ensure that the real time constraints are schedulable.
10272 static DEFINE_MUTEX(rt_constraints_mutex);
10274 static unsigned long to_ratio(u64 period, u64 runtime)
10276 if (runtime == RUNTIME_INF)
10279 return div64_u64(runtime << 20, period);
10282 /* Must be called with tasklist_lock held */
10283 static inline int tg_has_rt_tasks(struct task_group *tg)
10285 struct task_struct *g, *p;
10287 do_each_thread(g, p) {
10288 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10290 } while_each_thread(g, p);
10295 struct rt_schedulable_data {
10296 struct task_group *tg;
10301 static int tg_schedulable(struct task_group *tg, void *data)
10303 struct rt_schedulable_data *d = data;
10304 struct task_group *child;
10305 unsigned long total, sum = 0;
10306 u64 period, runtime;
10308 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10309 runtime = tg->rt_bandwidth.rt_runtime;
10312 period = d->rt_period;
10313 runtime = d->rt_runtime;
10316 #ifdef CONFIG_USER_SCHED
10317 if (tg == &root_task_group) {
10318 period = global_rt_period();
10319 runtime = global_rt_runtime();
10324 * Cannot have more runtime than the period.
10326 if (runtime > period && runtime != RUNTIME_INF)
10330 * Ensure we don't starve existing RT tasks.
10332 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10335 total = to_ratio(period, runtime);
10338 * Nobody can have more than the global setting allows.
10340 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10344 * The sum of our children's runtime should not exceed our own.
10346 list_for_each_entry_rcu(child, &tg->children, siblings) {
10347 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10348 runtime = child->rt_bandwidth.rt_runtime;
10350 if (child == d->tg) {
10351 period = d->rt_period;
10352 runtime = d->rt_runtime;
10355 sum += to_ratio(period, runtime);
10364 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10366 struct rt_schedulable_data data = {
10368 .rt_period = period,
10369 .rt_runtime = runtime,
10372 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10375 static int tg_set_bandwidth(struct task_group *tg,
10376 u64 rt_period, u64 rt_runtime)
10380 mutex_lock(&rt_constraints_mutex);
10381 read_lock(&tasklist_lock);
10382 err = __rt_schedulable(tg, rt_period, rt_runtime);
10386 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10387 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10388 tg->rt_bandwidth.rt_runtime = rt_runtime;
10390 for_each_possible_cpu(i) {
10391 struct rt_rq *rt_rq = tg->rt_rq[i];
10393 spin_lock(&rt_rq->rt_runtime_lock);
10394 rt_rq->rt_runtime = rt_runtime;
10395 spin_unlock(&rt_rq->rt_runtime_lock);
10397 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10399 read_unlock(&tasklist_lock);
10400 mutex_unlock(&rt_constraints_mutex);
10405 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10407 u64 rt_runtime, rt_period;
10409 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10410 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10411 if (rt_runtime_us < 0)
10412 rt_runtime = RUNTIME_INF;
10414 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10417 long sched_group_rt_runtime(struct task_group *tg)
10421 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10424 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10425 do_div(rt_runtime_us, NSEC_PER_USEC);
10426 return rt_runtime_us;
10429 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10431 u64 rt_runtime, rt_period;
10433 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10434 rt_runtime = tg->rt_bandwidth.rt_runtime;
10436 if (rt_period == 0)
10439 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10442 long sched_group_rt_period(struct task_group *tg)
10446 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10447 do_div(rt_period_us, NSEC_PER_USEC);
10448 return rt_period_us;
10451 static int sched_rt_global_constraints(void)
10453 u64 runtime, period;
10456 if (sysctl_sched_rt_period <= 0)
10459 runtime = global_rt_runtime();
10460 period = global_rt_period();
10463 * Sanity check on the sysctl variables.
10465 if (runtime > period && runtime != RUNTIME_INF)
10468 mutex_lock(&rt_constraints_mutex);
10469 read_lock(&tasklist_lock);
10470 ret = __rt_schedulable(NULL, 0, 0);
10471 read_unlock(&tasklist_lock);
10472 mutex_unlock(&rt_constraints_mutex);
10477 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10479 /* Don't accept realtime tasks when there is no way for them to run */
10480 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10486 #else /* !CONFIG_RT_GROUP_SCHED */
10487 static int sched_rt_global_constraints(void)
10489 unsigned long flags;
10492 if (sysctl_sched_rt_period <= 0)
10496 * There's always some RT tasks in the root group
10497 * -- migration, kstopmachine etc..
10499 if (sysctl_sched_rt_runtime == 0)
10502 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10503 for_each_possible_cpu(i) {
10504 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10506 spin_lock(&rt_rq->rt_runtime_lock);
10507 rt_rq->rt_runtime = global_rt_runtime();
10508 spin_unlock(&rt_rq->rt_runtime_lock);
10510 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10514 #endif /* CONFIG_RT_GROUP_SCHED */
10516 int sched_rt_handler(struct ctl_table *table, int write,
10517 void __user *buffer, size_t *lenp,
10521 int old_period, old_runtime;
10522 static DEFINE_MUTEX(mutex);
10524 mutex_lock(&mutex);
10525 old_period = sysctl_sched_rt_period;
10526 old_runtime = sysctl_sched_rt_runtime;
10528 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10530 if (!ret && write) {
10531 ret = sched_rt_global_constraints();
10533 sysctl_sched_rt_period = old_period;
10534 sysctl_sched_rt_runtime = old_runtime;
10536 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10537 def_rt_bandwidth.rt_period =
10538 ns_to_ktime(global_rt_period());
10541 mutex_unlock(&mutex);
10546 #ifdef CONFIG_CGROUP_SCHED
10548 /* return corresponding task_group object of a cgroup */
10549 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10551 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10552 struct task_group, css);
10555 static struct cgroup_subsys_state *
10556 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10558 struct task_group *tg, *parent;
10560 if (!cgrp->parent) {
10561 /* This is early initialization for the top cgroup */
10562 return &init_task_group.css;
10565 parent = cgroup_tg(cgrp->parent);
10566 tg = sched_create_group(parent);
10568 return ERR_PTR(-ENOMEM);
10574 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10576 struct task_group *tg = cgroup_tg(cgrp);
10578 sched_destroy_group(tg);
10582 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10584 if ((current != tsk) && (!capable(CAP_SYS_NICE))) {
10585 const struct cred *cred = current_cred(), *tcred;
10587 tcred = __task_cred(tsk);
10589 if (cred->euid != tcred->uid && cred->euid != tcred->suid)
10593 #ifdef CONFIG_RT_GROUP_SCHED
10594 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10597 /* We don't support RT-tasks being in separate groups */
10598 if (tsk->sched_class != &fair_sched_class)
10605 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10606 struct task_struct *tsk, bool threadgroup)
10608 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10612 struct task_struct *c;
10614 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10615 retval = cpu_cgroup_can_attach_task(cgrp, c);
10627 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10628 struct cgroup *old_cont, struct task_struct *tsk,
10631 sched_move_task(tsk);
10633 struct task_struct *c;
10635 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10636 sched_move_task(c);
10642 #ifdef CONFIG_FAIR_GROUP_SCHED
10643 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10646 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10649 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10651 struct task_group *tg = cgroup_tg(cgrp);
10653 return (u64) tg->shares;
10655 #endif /* CONFIG_FAIR_GROUP_SCHED */
10657 #ifdef CONFIG_RT_GROUP_SCHED
10658 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10661 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10664 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10666 return sched_group_rt_runtime(cgroup_tg(cgrp));
10669 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10672 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10675 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10677 return sched_group_rt_period(cgroup_tg(cgrp));
10679 #endif /* CONFIG_RT_GROUP_SCHED */
10681 static struct cftype cpu_files[] = {
10682 #ifdef CONFIG_FAIR_GROUP_SCHED
10685 .read_u64 = cpu_shares_read_u64,
10686 .write_u64 = cpu_shares_write_u64,
10689 #ifdef CONFIG_RT_GROUP_SCHED
10691 .name = "rt_runtime_us",
10692 .read_s64 = cpu_rt_runtime_read,
10693 .write_s64 = cpu_rt_runtime_write,
10696 .name = "rt_period_us",
10697 .read_u64 = cpu_rt_period_read_uint,
10698 .write_u64 = cpu_rt_period_write_uint,
10703 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10705 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10708 struct cgroup_subsys cpu_cgroup_subsys = {
10710 .create = cpu_cgroup_create,
10711 .destroy = cpu_cgroup_destroy,
10712 .can_attach = cpu_cgroup_can_attach,
10713 .attach = cpu_cgroup_attach,
10714 .populate = cpu_cgroup_populate,
10715 .subsys_id = cpu_cgroup_subsys_id,
10719 #endif /* CONFIG_CGROUP_SCHED */
10721 #ifdef CONFIG_CGROUP_CPUACCT
10724 * CPU accounting code for task groups.
10726 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10727 * (balbir@in.ibm.com).
10730 /* track cpu usage of a group of tasks and its child groups */
10732 struct cgroup_subsys_state css;
10733 /* cpuusage holds pointer to a u64-type object on every cpu */
10735 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10736 struct cpuacct *parent;
10739 struct cgroup_subsys cpuacct_subsys;
10741 /* return cpu accounting group corresponding to this container */
10742 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10744 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10745 struct cpuacct, css);
10748 /* return cpu accounting group to which this task belongs */
10749 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10751 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10752 struct cpuacct, css);
10755 /* create a new cpu accounting group */
10756 static struct cgroup_subsys_state *cpuacct_create(
10757 struct cgroup_subsys *ss, struct cgroup *cgrp)
10759 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10765 ca->cpuusage = alloc_percpu(u64);
10769 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10770 if (percpu_counter_init(&ca->cpustat[i], 0))
10771 goto out_free_counters;
10774 ca->parent = cgroup_ca(cgrp->parent);
10780 percpu_counter_destroy(&ca->cpustat[i]);
10781 free_percpu(ca->cpuusage);
10785 return ERR_PTR(-ENOMEM);
10788 /* destroy an existing cpu accounting group */
10790 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10792 struct cpuacct *ca = cgroup_ca(cgrp);
10795 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10796 percpu_counter_destroy(&ca->cpustat[i]);
10797 free_percpu(ca->cpuusage);
10801 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10803 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10806 #ifndef CONFIG_64BIT
10808 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10810 spin_lock_irq(&cpu_rq(cpu)->lock);
10812 spin_unlock_irq(&cpu_rq(cpu)->lock);
10820 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10822 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10824 #ifndef CONFIG_64BIT
10826 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10828 spin_lock_irq(&cpu_rq(cpu)->lock);
10830 spin_unlock_irq(&cpu_rq(cpu)->lock);
10836 /* return total cpu usage (in nanoseconds) of a group */
10837 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10839 struct cpuacct *ca = cgroup_ca(cgrp);
10840 u64 totalcpuusage = 0;
10843 for_each_present_cpu(i)
10844 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10846 return totalcpuusage;
10849 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10852 struct cpuacct *ca = cgroup_ca(cgrp);
10861 for_each_present_cpu(i)
10862 cpuacct_cpuusage_write(ca, i, 0);
10868 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10869 struct seq_file *m)
10871 struct cpuacct *ca = cgroup_ca(cgroup);
10875 for_each_present_cpu(i) {
10876 percpu = cpuacct_cpuusage_read(ca, i);
10877 seq_printf(m, "%llu ", (unsigned long long) percpu);
10879 seq_printf(m, "\n");
10883 static const char *cpuacct_stat_desc[] = {
10884 [CPUACCT_STAT_USER] = "user",
10885 [CPUACCT_STAT_SYSTEM] = "system",
10888 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10889 struct cgroup_map_cb *cb)
10891 struct cpuacct *ca = cgroup_ca(cgrp);
10894 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10895 s64 val = percpu_counter_read(&ca->cpustat[i]);
10896 val = cputime64_to_clock_t(val);
10897 cb->fill(cb, cpuacct_stat_desc[i], val);
10902 static struct cftype files[] = {
10905 .read_u64 = cpuusage_read,
10906 .write_u64 = cpuusage_write,
10909 .name = "usage_percpu",
10910 .read_seq_string = cpuacct_percpu_seq_read,
10914 .read_map = cpuacct_stats_show,
10918 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10920 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10924 * charge this task's execution time to its accounting group.
10926 * called with rq->lock held.
10928 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10930 struct cpuacct *ca;
10933 if (unlikely(!cpuacct_subsys.active))
10936 cpu = task_cpu(tsk);
10942 for (; ca; ca = ca->parent) {
10943 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10944 *cpuusage += cputime;
10951 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
10952 * in cputime_t units. As a result, cpuacct_update_stats calls
10953 * percpu_counter_add with values large enough to always overflow the
10954 * per cpu batch limit causing bad SMP scalability.
10956 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
10957 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
10958 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
10961 #define CPUACCT_BATCH \
10962 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
10964 #define CPUACCT_BATCH 0
10968 * Charge the system/user time to the task's accounting group.
10970 static void cpuacct_update_stats(struct task_struct *tsk,
10971 enum cpuacct_stat_index idx, cputime_t val)
10973 struct cpuacct *ca;
10974 int batch = CPUACCT_BATCH;
10976 if (unlikely(!cpuacct_subsys.active))
10983 __percpu_counter_add(&ca->cpustat[idx], val, batch);
10989 struct cgroup_subsys cpuacct_subsys = {
10991 .create = cpuacct_create,
10992 .destroy = cpuacct_destroy,
10993 .populate = cpuacct_populate,
10994 .subsys_id = cpuacct_subsys_id,
10996 #endif /* CONFIG_CGROUP_CPUACCT */
11000 int rcu_expedited_torture_stats(char *page)
11004 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
11006 void synchronize_sched_expedited(void)
11009 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
11011 #else /* #ifndef CONFIG_SMP */
11013 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
11014 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
11016 #define RCU_EXPEDITED_STATE_POST -2
11017 #define RCU_EXPEDITED_STATE_IDLE -1
11019 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
11021 int rcu_expedited_torture_stats(char *page)
11026 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
11027 for_each_online_cpu(cpu) {
11028 cnt += sprintf(&page[cnt], " %d:%d",
11029 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
11031 cnt += sprintf(&page[cnt], "\n");
11034 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
11036 static long synchronize_sched_expedited_count;
11039 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
11040 * approach to force grace period to end quickly. This consumes
11041 * significant time on all CPUs, and is thus not recommended for
11042 * any sort of common-case code.
11044 * Note that it is illegal to call this function while holding any
11045 * lock that is acquired by a CPU-hotplug notifier. Failing to
11046 * observe this restriction will result in deadlock.
11048 void synchronize_sched_expedited(void)
11051 unsigned long flags;
11052 bool need_full_sync = 0;
11054 struct migration_req *req;
11058 smp_mb(); /* ensure prior mod happens before capturing snap. */
11059 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
11061 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
11063 if (trycount++ < 10)
11064 udelay(trycount * num_online_cpus());
11066 synchronize_sched();
11069 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
11070 smp_mb(); /* ensure test happens before caller kfree */
11075 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
11076 for_each_online_cpu(cpu) {
11078 req = &per_cpu(rcu_migration_req, cpu);
11079 init_completion(&req->done);
11081 req->dest_cpu = RCU_MIGRATION_NEED_QS;
11082 spin_lock_irqsave(&rq->lock, flags);
11083 list_add(&req->list, &rq->migration_queue);
11084 spin_unlock_irqrestore(&rq->lock, flags);
11085 wake_up_process(rq->migration_thread);
11087 for_each_online_cpu(cpu) {
11088 rcu_expedited_state = cpu;
11089 req = &per_cpu(rcu_migration_req, cpu);
11091 wait_for_completion(&req->done);
11092 spin_lock_irqsave(&rq->lock, flags);
11093 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
11094 need_full_sync = 1;
11095 req->dest_cpu = RCU_MIGRATION_IDLE;
11096 spin_unlock_irqrestore(&rq->lock, flags);
11098 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
11099 mutex_unlock(&rcu_sched_expedited_mutex);
11101 if (need_full_sync)
11102 synchronize_sched();
11104 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
11106 #endif /* #else #ifndef CONFIG_SMP */