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);
313 static int root_task_group_empty(void)
315 return list_empty(&root_task_group.children);
319 #ifdef CONFIG_FAIR_GROUP_SCHED
320 #ifdef CONFIG_USER_SCHED
321 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
322 #else /* !CONFIG_USER_SCHED */
323 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
324 #endif /* CONFIG_USER_SCHED */
327 * A weight of 0 or 1 can cause arithmetics problems.
328 * A weight of a cfs_rq is the sum of weights of which entities
329 * are queued on this cfs_rq, so a weight of a entity should not be
330 * too large, so as the shares value of a task group.
331 * (The default weight is 1024 - so there's no practical
332 * limitation from this.)
335 #define MAX_SHARES (1UL << 18)
337 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
340 /* Default task group.
341 * Every task in system belong to this group at bootup.
343 struct task_group init_task_group;
345 /* return group to which a task belongs */
346 static inline struct task_group *task_group(struct task_struct *p)
348 struct task_group *tg;
350 #ifdef CONFIG_USER_SCHED
352 tg = __task_cred(p)->user->tg;
354 #elif defined(CONFIG_CGROUP_SCHED)
355 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
356 struct task_group, css);
358 tg = &init_task_group;
363 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
364 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
366 #ifdef CONFIG_FAIR_GROUP_SCHED
367 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
368 p->se.parent = task_group(p)->se[cpu];
371 #ifdef CONFIG_RT_GROUP_SCHED
372 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
373 p->rt.parent = task_group(p)->rt_se[cpu];
379 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
380 static inline struct task_group *task_group(struct task_struct *p)
385 #endif /* CONFIG_GROUP_SCHED */
387 /* CFS-related fields in a runqueue */
389 struct load_weight load;
390 unsigned long nr_running;
395 struct rb_root tasks_timeline;
396 struct rb_node *rb_leftmost;
398 struct list_head tasks;
399 struct list_head *balance_iterator;
402 * 'curr' points to currently running entity on this cfs_rq.
403 * It is set to NULL otherwise (i.e when none are currently running).
405 struct sched_entity *curr, *next, *last;
407 unsigned int nr_spread_over;
409 #ifdef CONFIG_FAIR_GROUP_SCHED
410 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
413 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
414 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
415 * (like users, containers etc.)
417 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
418 * list is used during load balance.
420 struct list_head leaf_cfs_rq_list;
421 struct task_group *tg; /* group that "owns" this runqueue */
425 * the part of load.weight contributed by tasks
427 unsigned long task_weight;
430 * h_load = weight * f(tg)
432 * Where f(tg) is the recursive weight fraction assigned to
435 unsigned long h_load;
438 * this cpu's part of tg->shares
440 unsigned long shares;
443 * load.weight at the time we set shares
445 unsigned long rq_weight;
450 /* Real-Time classes' related field in a runqueue: */
452 struct rt_prio_array active;
453 unsigned long rt_nr_running;
454 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
456 int curr; /* highest queued rt task prio */
458 int next; /* next highest */
463 unsigned long rt_nr_migratory;
464 unsigned long rt_nr_total;
466 struct plist_head pushable_tasks;
471 /* Nests inside the rq lock: */
472 spinlock_t rt_runtime_lock;
474 #ifdef CONFIG_RT_GROUP_SCHED
475 unsigned long rt_nr_boosted;
478 struct list_head leaf_rt_rq_list;
479 struct task_group *tg;
480 struct sched_rt_entity *rt_se;
487 * We add the notion of a root-domain which will be used to define per-domain
488 * variables. Each exclusive cpuset essentially defines an island domain by
489 * fully partitioning the member cpus from any other cpuset. Whenever a new
490 * exclusive cpuset is created, we also create and attach a new root-domain
497 cpumask_var_t online;
500 * The "RT overload" flag: it gets set if a CPU has more than
501 * one runnable RT task.
503 cpumask_var_t rto_mask;
506 struct cpupri cpupri;
511 * By default the system creates a single root-domain with all cpus as
512 * members (mimicking the global state we have today).
514 static struct root_domain def_root_domain;
519 * This is the main, per-CPU runqueue data structure.
521 * Locking rule: those places that want to lock multiple runqueues
522 * (such as the load balancing or the thread migration code), lock
523 * acquire operations must be ordered by ascending &runqueue.
530 * nr_running and cpu_load should be in the same cacheline because
531 * remote CPUs use both these fields when doing load calculation.
533 unsigned long nr_running;
534 #define CPU_LOAD_IDX_MAX 5
535 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
537 unsigned char in_nohz_recently;
539 /* capture load from *all* tasks on this cpu: */
540 struct load_weight load;
541 unsigned long nr_load_updates;
547 #ifdef CONFIG_FAIR_GROUP_SCHED
548 /* list of leaf cfs_rq on this cpu: */
549 struct list_head leaf_cfs_rq_list;
551 #ifdef CONFIG_RT_GROUP_SCHED
552 struct list_head leaf_rt_rq_list;
556 * This is part of a global counter where only the total sum
557 * over all CPUs matters. A task can increase this counter on
558 * one CPU and if it got migrated afterwards it may decrease
559 * it on another CPU. Always updated under the runqueue lock:
561 unsigned long nr_uninterruptible;
563 struct task_struct *curr, *idle;
564 unsigned long next_balance;
565 struct mm_struct *prev_mm;
572 struct root_domain *rd;
573 struct sched_domain *sd;
575 unsigned char idle_at_tick;
576 /* For active balancing */
580 /* cpu of this runqueue: */
584 unsigned long avg_load_per_task;
586 struct task_struct *migration_thread;
587 struct list_head migration_queue;
595 /* calc_load related fields */
596 unsigned long calc_load_update;
597 long calc_load_active;
599 #ifdef CONFIG_SCHED_HRTICK
601 int hrtick_csd_pending;
602 struct call_single_data hrtick_csd;
604 struct hrtimer hrtick_timer;
607 #ifdef CONFIG_SCHEDSTATS
609 struct sched_info rq_sched_info;
610 unsigned long long rq_cpu_time;
611 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
613 /* sys_sched_yield() stats */
614 unsigned int yld_count;
616 /* schedule() stats */
617 unsigned int sched_switch;
618 unsigned int sched_count;
619 unsigned int sched_goidle;
621 /* try_to_wake_up() stats */
622 unsigned int ttwu_count;
623 unsigned int ttwu_local;
626 unsigned int bkl_count;
630 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
633 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
635 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
638 static inline int cpu_of(struct rq *rq)
648 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
649 * See detach_destroy_domains: synchronize_sched for details.
651 * The domain tree of any CPU may only be accessed from within
652 * preempt-disabled sections.
654 #define for_each_domain(cpu, __sd) \
655 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
657 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
658 #define this_rq() (&__get_cpu_var(runqueues))
659 #define task_rq(p) cpu_rq(task_cpu(p))
660 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
661 #define raw_rq() (&__raw_get_cpu_var(runqueues))
663 inline void update_rq_clock(struct rq *rq)
665 rq->clock = sched_clock_cpu(cpu_of(rq));
669 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
671 #ifdef CONFIG_SCHED_DEBUG
672 # define const_debug __read_mostly
674 # define const_debug static const
679 * @cpu: the processor in question.
681 * Returns true if the current cpu runqueue is locked.
682 * This interface allows printk to be called with the runqueue lock
683 * held and know whether or not it is OK to wake up the klogd.
685 int runqueue_is_locked(int cpu)
687 return spin_is_locked(&cpu_rq(cpu)->lock);
691 * Debugging: various feature bits
694 #define SCHED_FEAT(name, enabled) \
695 __SCHED_FEAT_##name ,
698 #include "sched_features.h"
703 #define SCHED_FEAT(name, enabled) \
704 (1UL << __SCHED_FEAT_##name) * enabled |
706 const_debug unsigned int sysctl_sched_features =
707 #include "sched_features.h"
712 #ifdef CONFIG_SCHED_DEBUG
713 #define SCHED_FEAT(name, enabled) \
716 static __read_mostly char *sched_feat_names[] = {
717 #include "sched_features.h"
723 static int sched_feat_show(struct seq_file *m, void *v)
727 for (i = 0; sched_feat_names[i]; i++) {
728 if (!(sysctl_sched_features & (1UL << i)))
730 seq_printf(m, "%s ", sched_feat_names[i]);
738 sched_feat_write(struct file *filp, const char __user *ubuf,
739 size_t cnt, loff_t *ppos)
749 if (copy_from_user(&buf, ubuf, cnt))
754 if (strncmp(buf, "NO_", 3) == 0) {
759 for (i = 0; sched_feat_names[i]; i++) {
760 int len = strlen(sched_feat_names[i]);
762 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
764 sysctl_sched_features &= ~(1UL << i);
766 sysctl_sched_features |= (1UL << i);
771 if (!sched_feat_names[i])
779 static int sched_feat_open(struct inode *inode, struct file *filp)
781 return single_open(filp, sched_feat_show, NULL);
784 static const struct file_operations sched_feat_fops = {
785 .open = sched_feat_open,
786 .write = sched_feat_write,
789 .release = single_release,
792 static __init int sched_init_debug(void)
794 debugfs_create_file("sched_features", 0644, NULL, NULL,
799 late_initcall(sched_init_debug);
803 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
806 * Number of tasks to iterate in a single balance run.
807 * Limited because this is done with IRQs disabled.
809 const_debug unsigned int sysctl_sched_nr_migrate = 32;
812 * ratelimit for updating the group shares.
815 unsigned int sysctl_sched_shares_ratelimit = 250000;
818 * Inject some fuzzyness into changing the per-cpu group shares
819 * this avoids remote rq-locks at the expense of fairness.
822 unsigned int sysctl_sched_shares_thresh = 4;
825 * period over which we average the RT time consumption, measured
830 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
833 * period over which we measure -rt task cpu usage in us.
836 unsigned int sysctl_sched_rt_period = 1000000;
838 static __read_mostly int scheduler_running;
841 * part of the period that we allow rt tasks to run in us.
844 int sysctl_sched_rt_runtime = 950000;
846 static inline u64 global_rt_period(void)
848 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
851 static inline u64 global_rt_runtime(void)
853 if (sysctl_sched_rt_runtime < 0)
856 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
859 #ifndef prepare_arch_switch
860 # define prepare_arch_switch(next) do { } while (0)
862 #ifndef finish_arch_switch
863 # define finish_arch_switch(prev) do { } while (0)
866 static inline int task_current(struct rq *rq, struct task_struct *p)
868 return rq->curr == p;
871 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
872 static inline int task_running(struct rq *rq, struct task_struct *p)
874 return task_current(rq, p);
877 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
881 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
883 #ifdef CONFIG_DEBUG_SPINLOCK
884 /* this is a valid case when another task releases the spinlock */
885 rq->lock.owner = current;
888 * If we are tracking spinlock dependencies then we have to
889 * fix up the runqueue lock - which gets 'carried over' from
892 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
894 spin_unlock_irq(&rq->lock);
897 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
898 static inline int task_running(struct rq *rq, struct task_struct *p)
903 return task_current(rq, p);
907 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
911 * We can optimise this out completely for !SMP, because the
912 * SMP rebalancing from interrupt is the only thing that cares
917 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
918 spin_unlock_irq(&rq->lock);
920 spin_unlock(&rq->lock);
924 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
928 * After ->oncpu is cleared, the task can be moved to a different CPU.
929 * We must ensure this doesn't happen until the switch is completely
935 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
939 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
942 * __task_rq_lock - lock the runqueue a given task resides on.
943 * Must be called interrupts disabled.
945 static inline struct rq *__task_rq_lock(struct task_struct *p)
949 struct rq *rq = task_rq(p);
950 spin_lock(&rq->lock);
951 if (likely(rq == task_rq(p)))
953 spin_unlock(&rq->lock);
958 * task_rq_lock - lock the runqueue a given task resides on and disable
959 * interrupts. Note the ordering: we can safely lookup the task_rq without
960 * explicitly disabling preemption.
962 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
968 local_irq_save(*flags);
970 spin_lock(&rq->lock);
971 if (likely(rq == task_rq(p)))
973 spin_unlock_irqrestore(&rq->lock, *flags);
977 void task_rq_unlock_wait(struct task_struct *p)
979 struct rq *rq = task_rq(p);
981 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
982 spin_unlock_wait(&rq->lock);
985 static void __task_rq_unlock(struct rq *rq)
988 spin_unlock(&rq->lock);
991 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
994 spin_unlock_irqrestore(&rq->lock, *flags);
998 * this_rq_lock - lock this runqueue and disable interrupts.
1000 static struct rq *this_rq_lock(void)
1001 __acquires(rq->lock)
1005 local_irq_disable();
1007 spin_lock(&rq->lock);
1012 #ifdef CONFIG_SCHED_HRTICK
1014 * Use HR-timers to deliver accurate preemption points.
1016 * Its all a bit involved since we cannot program an hrt while holding the
1017 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1020 * When we get rescheduled we reprogram the hrtick_timer outside of the
1026 * - enabled by features
1027 * - hrtimer is actually high res
1029 static inline int hrtick_enabled(struct rq *rq)
1031 if (!sched_feat(HRTICK))
1033 if (!cpu_active(cpu_of(rq)))
1035 return hrtimer_is_hres_active(&rq->hrtick_timer);
1038 static void hrtick_clear(struct rq *rq)
1040 if (hrtimer_active(&rq->hrtick_timer))
1041 hrtimer_cancel(&rq->hrtick_timer);
1045 * High-resolution timer tick.
1046 * Runs from hardirq context with interrupts disabled.
1048 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1050 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1052 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1054 spin_lock(&rq->lock);
1055 update_rq_clock(rq);
1056 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1057 spin_unlock(&rq->lock);
1059 return HRTIMER_NORESTART;
1064 * called from hardirq (IPI) context
1066 static void __hrtick_start(void *arg)
1068 struct rq *rq = arg;
1070 spin_lock(&rq->lock);
1071 hrtimer_restart(&rq->hrtick_timer);
1072 rq->hrtick_csd_pending = 0;
1073 spin_unlock(&rq->lock);
1077 * Called to set the hrtick timer state.
1079 * called with rq->lock held and irqs disabled
1081 static void hrtick_start(struct rq *rq, u64 delay)
1083 struct hrtimer *timer = &rq->hrtick_timer;
1084 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1086 hrtimer_set_expires(timer, time);
1088 if (rq == this_rq()) {
1089 hrtimer_restart(timer);
1090 } else if (!rq->hrtick_csd_pending) {
1091 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1092 rq->hrtick_csd_pending = 1;
1097 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1099 int cpu = (int)(long)hcpu;
1102 case CPU_UP_CANCELED:
1103 case CPU_UP_CANCELED_FROZEN:
1104 case CPU_DOWN_PREPARE:
1105 case CPU_DOWN_PREPARE_FROZEN:
1107 case CPU_DEAD_FROZEN:
1108 hrtick_clear(cpu_rq(cpu));
1115 static __init void init_hrtick(void)
1117 hotcpu_notifier(hotplug_hrtick, 0);
1121 * Called to set the hrtick timer state.
1123 * called with rq->lock held and irqs disabled
1125 static void hrtick_start(struct rq *rq, u64 delay)
1127 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1128 HRTIMER_MODE_REL_PINNED, 0);
1131 static inline void init_hrtick(void)
1134 #endif /* CONFIG_SMP */
1136 static void init_rq_hrtick(struct rq *rq)
1139 rq->hrtick_csd_pending = 0;
1141 rq->hrtick_csd.flags = 0;
1142 rq->hrtick_csd.func = __hrtick_start;
1143 rq->hrtick_csd.info = rq;
1146 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1147 rq->hrtick_timer.function = hrtick;
1149 #else /* CONFIG_SCHED_HRTICK */
1150 static inline void hrtick_clear(struct rq *rq)
1154 static inline void init_rq_hrtick(struct rq *rq)
1158 static inline void init_hrtick(void)
1161 #endif /* CONFIG_SCHED_HRTICK */
1164 * resched_task - mark a task 'to be rescheduled now'.
1166 * On UP this means the setting of the need_resched flag, on SMP it
1167 * might also involve a cross-CPU call to trigger the scheduler on
1172 #ifndef tsk_is_polling
1173 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1176 static void resched_task(struct task_struct *p)
1180 assert_spin_locked(&task_rq(p)->lock);
1182 if (test_tsk_need_resched(p))
1185 set_tsk_need_resched(p);
1188 if (cpu == smp_processor_id())
1191 /* NEED_RESCHED must be visible before we test polling */
1193 if (!tsk_is_polling(p))
1194 smp_send_reschedule(cpu);
1197 static void resched_cpu(int cpu)
1199 struct rq *rq = cpu_rq(cpu);
1200 unsigned long flags;
1202 if (!spin_trylock_irqsave(&rq->lock, flags))
1204 resched_task(cpu_curr(cpu));
1205 spin_unlock_irqrestore(&rq->lock, flags);
1210 * When add_timer_on() enqueues a timer into the timer wheel of an
1211 * idle CPU then this timer might expire before the next timer event
1212 * which is scheduled to wake up that CPU. In case of a completely
1213 * idle system the next event might even be infinite time into the
1214 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1215 * leaves the inner idle loop so the newly added timer is taken into
1216 * account when the CPU goes back to idle and evaluates the timer
1217 * wheel for the next timer event.
1219 void wake_up_idle_cpu(int cpu)
1221 struct rq *rq = cpu_rq(cpu);
1223 if (cpu == smp_processor_id())
1227 * This is safe, as this function is called with the timer
1228 * wheel base lock of (cpu) held. When the CPU is on the way
1229 * to idle and has not yet set rq->curr to idle then it will
1230 * be serialized on the timer wheel base lock and take the new
1231 * timer into account automatically.
1233 if (rq->curr != rq->idle)
1237 * We can set TIF_RESCHED on the idle task of the other CPU
1238 * lockless. The worst case is that the other CPU runs the
1239 * idle task through an additional NOOP schedule()
1241 set_tsk_need_resched(rq->idle);
1243 /* NEED_RESCHED must be visible before we test polling */
1245 if (!tsk_is_polling(rq->idle))
1246 smp_send_reschedule(cpu);
1248 #endif /* CONFIG_NO_HZ */
1250 static u64 sched_avg_period(void)
1252 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1255 static void sched_avg_update(struct rq *rq)
1257 s64 period = sched_avg_period();
1259 while ((s64)(rq->clock - rq->age_stamp) > period) {
1260 rq->age_stamp += period;
1265 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1267 rq->rt_avg += rt_delta;
1268 sched_avg_update(rq);
1271 #else /* !CONFIG_SMP */
1272 static void resched_task(struct task_struct *p)
1274 assert_spin_locked(&task_rq(p)->lock);
1275 set_tsk_need_resched(p);
1278 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1281 #endif /* CONFIG_SMP */
1283 #if BITS_PER_LONG == 32
1284 # define WMULT_CONST (~0UL)
1286 # define WMULT_CONST (1UL << 32)
1289 #define WMULT_SHIFT 32
1292 * Shift right and round:
1294 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1297 * delta *= weight / lw
1299 static unsigned long
1300 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1301 struct load_weight *lw)
1305 if (!lw->inv_weight) {
1306 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1309 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1313 tmp = (u64)delta_exec * weight;
1315 * Check whether we'd overflow the 64-bit multiplication:
1317 if (unlikely(tmp > WMULT_CONST))
1318 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1321 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1323 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1326 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1332 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1339 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1340 * of tasks with abnormal "nice" values across CPUs the contribution that
1341 * each task makes to its run queue's load is weighted according to its
1342 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1343 * scaled version of the new time slice allocation that they receive on time
1347 #define WEIGHT_IDLEPRIO 3
1348 #define WMULT_IDLEPRIO 1431655765
1351 * Nice levels are multiplicative, with a gentle 10% change for every
1352 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1353 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1354 * that remained on nice 0.
1356 * The "10% effect" is relative and cumulative: from _any_ nice level,
1357 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1358 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1359 * If a task goes up by ~10% and another task goes down by ~10% then
1360 * the relative distance between them is ~25%.)
1362 static const int prio_to_weight[40] = {
1363 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1364 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1365 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1366 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1367 /* 0 */ 1024, 820, 655, 526, 423,
1368 /* 5 */ 335, 272, 215, 172, 137,
1369 /* 10 */ 110, 87, 70, 56, 45,
1370 /* 15 */ 36, 29, 23, 18, 15,
1374 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1376 * In cases where the weight does not change often, we can use the
1377 * precalculated inverse to speed up arithmetics by turning divisions
1378 * into multiplications:
1380 static const u32 prio_to_wmult[40] = {
1381 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1382 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1383 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1384 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1385 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1386 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1387 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1388 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1391 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1394 * runqueue iterator, to support SMP load-balancing between different
1395 * scheduling classes, without having to expose their internal data
1396 * structures to the load-balancing proper:
1398 struct rq_iterator {
1400 struct task_struct *(*start)(void *);
1401 struct task_struct *(*next)(void *);
1405 static unsigned long
1406 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1407 unsigned long max_load_move, struct sched_domain *sd,
1408 enum cpu_idle_type idle, int *all_pinned,
1409 int *this_best_prio, struct rq_iterator *iterator);
1412 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1413 struct sched_domain *sd, enum cpu_idle_type idle,
1414 struct rq_iterator *iterator);
1417 /* Time spent by the tasks of the cpu accounting group executing in ... */
1418 enum cpuacct_stat_index {
1419 CPUACCT_STAT_USER, /* ... user mode */
1420 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1422 CPUACCT_STAT_NSTATS,
1425 #ifdef CONFIG_CGROUP_CPUACCT
1426 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1427 static void cpuacct_update_stats(struct task_struct *tsk,
1428 enum cpuacct_stat_index idx, cputime_t val);
1430 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1431 static inline void cpuacct_update_stats(struct task_struct *tsk,
1432 enum cpuacct_stat_index idx, cputime_t val) {}
1435 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1437 update_load_add(&rq->load, load);
1440 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1442 update_load_sub(&rq->load, load);
1445 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1446 typedef int (*tg_visitor)(struct task_group *, void *);
1449 * Iterate the full tree, calling @down when first entering a node and @up when
1450 * leaving it for the final time.
1452 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1454 struct task_group *parent, *child;
1458 parent = &root_task_group;
1460 ret = (*down)(parent, data);
1463 list_for_each_entry_rcu(child, &parent->children, siblings) {
1470 ret = (*up)(parent, data);
1475 parent = parent->parent;
1484 static int tg_nop(struct task_group *tg, void *data)
1491 /* Used instead of source_load when we know the type == 0 */
1492 static unsigned long weighted_cpuload(const int cpu)
1494 return cpu_rq(cpu)->load.weight;
1498 * Return a low guess at the load of a migration-source cpu weighted
1499 * according to the scheduling class and "nice" value.
1501 * We want to under-estimate the load of migration sources, to
1502 * balance conservatively.
1504 static unsigned long source_load(int cpu, int type)
1506 struct rq *rq = cpu_rq(cpu);
1507 unsigned long total = weighted_cpuload(cpu);
1509 if (type == 0 || !sched_feat(LB_BIAS))
1512 return min(rq->cpu_load[type-1], total);
1516 * Return a high guess at the load of a migration-target cpu weighted
1517 * according to the scheduling class and "nice" value.
1519 static unsigned long target_load(int cpu, int type)
1521 struct rq *rq = cpu_rq(cpu);
1522 unsigned long total = weighted_cpuload(cpu);
1524 if (type == 0 || !sched_feat(LB_BIAS))
1527 return max(rq->cpu_load[type-1], total);
1530 static struct sched_group *group_of(int cpu)
1532 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1540 static unsigned long power_of(int cpu)
1542 struct sched_group *group = group_of(cpu);
1545 return SCHED_LOAD_SCALE;
1547 return group->cpu_power;
1550 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1552 static unsigned long cpu_avg_load_per_task(int cpu)
1554 struct rq *rq = cpu_rq(cpu);
1555 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1558 rq->avg_load_per_task = rq->load.weight / nr_running;
1560 rq->avg_load_per_task = 0;
1562 return rq->avg_load_per_task;
1565 #ifdef CONFIG_FAIR_GROUP_SCHED
1567 struct update_shares_data {
1568 unsigned long rq_weight[NR_CPUS];
1571 static DEFINE_PER_CPU(struct update_shares_data, update_shares_data);
1573 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1576 * Calculate and set the cpu's group shares.
1578 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1579 unsigned long sd_shares,
1580 unsigned long sd_rq_weight,
1581 struct update_shares_data *usd)
1583 unsigned long shares, rq_weight;
1586 rq_weight = usd->rq_weight[cpu];
1589 rq_weight = NICE_0_LOAD;
1593 * \Sum_j shares_j * rq_weight_i
1594 * shares_i = -----------------------------
1595 * \Sum_j rq_weight_j
1597 shares = (sd_shares * rq_weight) / sd_rq_weight;
1598 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1600 if (abs(shares - tg->se[cpu]->load.weight) >
1601 sysctl_sched_shares_thresh) {
1602 struct rq *rq = cpu_rq(cpu);
1603 unsigned long flags;
1605 spin_lock_irqsave(&rq->lock, flags);
1606 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1607 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1608 __set_se_shares(tg->se[cpu], shares);
1609 spin_unlock_irqrestore(&rq->lock, flags);
1614 * Re-compute the task group their per cpu shares over the given domain.
1615 * This needs to be done in a bottom-up fashion because the rq weight of a
1616 * parent group depends on the shares of its child groups.
1618 static int tg_shares_up(struct task_group *tg, void *data)
1620 unsigned long weight, rq_weight = 0, shares = 0;
1621 struct update_shares_data *usd;
1622 struct sched_domain *sd = data;
1623 unsigned long flags;
1629 local_irq_save(flags);
1630 usd = &__get_cpu_var(update_shares_data);
1632 for_each_cpu(i, sched_domain_span(sd)) {
1633 weight = tg->cfs_rq[i]->load.weight;
1634 usd->rq_weight[i] = weight;
1637 * If there are currently no tasks on the cpu pretend there
1638 * is one of average load so that when a new task gets to
1639 * run here it will not get delayed by group starvation.
1642 weight = NICE_0_LOAD;
1644 rq_weight += weight;
1645 shares += tg->cfs_rq[i]->shares;
1648 if ((!shares && rq_weight) || shares > tg->shares)
1649 shares = tg->shares;
1651 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1652 shares = tg->shares;
1654 for_each_cpu(i, sched_domain_span(sd))
1655 update_group_shares_cpu(tg, i, shares, rq_weight, usd);
1657 local_irq_restore(flags);
1663 * Compute the cpu's hierarchical load factor for each task group.
1664 * This needs to be done in a top-down fashion because the load of a child
1665 * group is a fraction of its parents load.
1667 static int tg_load_down(struct task_group *tg, void *data)
1670 long cpu = (long)data;
1673 load = cpu_rq(cpu)->load.weight;
1675 load = tg->parent->cfs_rq[cpu]->h_load;
1676 load *= tg->cfs_rq[cpu]->shares;
1677 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1680 tg->cfs_rq[cpu]->h_load = load;
1685 static void update_shares(struct sched_domain *sd)
1690 if (root_task_group_empty())
1693 now = cpu_clock(raw_smp_processor_id());
1694 elapsed = now - sd->last_update;
1696 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1697 sd->last_update = now;
1698 walk_tg_tree(tg_nop, tg_shares_up, sd);
1702 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1704 if (root_task_group_empty())
1707 spin_unlock(&rq->lock);
1709 spin_lock(&rq->lock);
1712 static void update_h_load(long cpu)
1714 if (root_task_group_empty())
1717 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1722 static inline void update_shares(struct sched_domain *sd)
1726 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1732 #ifdef CONFIG_PREEMPT
1734 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1737 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1738 * way at the expense of forcing extra atomic operations in all
1739 * invocations. This assures that the double_lock is acquired using the
1740 * same underlying policy as the spinlock_t on this architecture, which
1741 * reduces latency compared to the unfair variant below. However, it
1742 * also adds more overhead and therefore may reduce throughput.
1744 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1745 __releases(this_rq->lock)
1746 __acquires(busiest->lock)
1747 __acquires(this_rq->lock)
1749 spin_unlock(&this_rq->lock);
1750 double_rq_lock(this_rq, busiest);
1757 * Unfair double_lock_balance: Optimizes throughput at the expense of
1758 * latency by eliminating extra atomic operations when the locks are
1759 * already in proper order on entry. This favors lower cpu-ids and will
1760 * grant the double lock to lower cpus over higher ids under contention,
1761 * regardless of entry order into the function.
1763 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1764 __releases(this_rq->lock)
1765 __acquires(busiest->lock)
1766 __acquires(this_rq->lock)
1770 if (unlikely(!spin_trylock(&busiest->lock))) {
1771 if (busiest < this_rq) {
1772 spin_unlock(&this_rq->lock);
1773 spin_lock(&busiest->lock);
1774 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1777 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1782 #endif /* CONFIG_PREEMPT */
1785 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1787 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1789 if (unlikely(!irqs_disabled())) {
1790 /* printk() doesn't work good under rq->lock */
1791 spin_unlock(&this_rq->lock);
1795 return _double_lock_balance(this_rq, busiest);
1798 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1799 __releases(busiest->lock)
1801 spin_unlock(&busiest->lock);
1802 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1806 #ifdef CONFIG_FAIR_GROUP_SCHED
1807 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1810 cfs_rq->shares = shares;
1815 static void calc_load_account_active(struct rq *this_rq);
1817 #include "sched_stats.h"
1818 #include "sched_idletask.c"
1819 #include "sched_fair.c"
1820 #include "sched_rt.c"
1821 #ifdef CONFIG_SCHED_DEBUG
1822 # include "sched_debug.c"
1825 #define sched_class_highest (&rt_sched_class)
1826 #define for_each_class(class) \
1827 for (class = sched_class_highest; class; class = class->next)
1829 static void inc_nr_running(struct rq *rq)
1834 static void dec_nr_running(struct rq *rq)
1839 static void set_load_weight(struct task_struct *p)
1841 if (task_has_rt_policy(p)) {
1842 p->se.load.weight = prio_to_weight[0] * 2;
1843 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1848 * SCHED_IDLE tasks get minimal weight:
1850 if (p->policy == SCHED_IDLE) {
1851 p->se.load.weight = WEIGHT_IDLEPRIO;
1852 p->se.load.inv_weight = WMULT_IDLEPRIO;
1856 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1857 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1860 static void update_avg(u64 *avg, u64 sample)
1862 s64 diff = sample - *avg;
1866 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1869 p->se.start_runtime = p->se.sum_exec_runtime;
1871 sched_info_queued(p);
1872 p->sched_class->enqueue_task(rq, p, wakeup);
1876 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1879 if (p->se.last_wakeup) {
1880 update_avg(&p->se.avg_overlap,
1881 p->se.sum_exec_runtime - p->se.last_wakeup);
1882 p->se.last_wakeup = 0;
1884 update_avg(&p->se.avg_wakeup,
1885 sysctl_sched_wakeup_granularity);
1889 sched_info_dequeued(p);
1890 p->sched_class->dequeue_task(rq, p, sleep);
1895 * __normal_prio - return the priority that is based on the static prio
1897 static inline int __normal_prio(struct task_struct *p)
1899 return p->static_prio;
1903 * Calculate the expected normal priority: i.e. priority
1904 * without taking RT-inheritance into account. Might be
1905 * boosted by interactivity modifiers. Changes upon fork,
1906 * setprio syscalls, and whenever the interactivity
1907 * estimator recalculates.
1909 static inline int normal_prio(struct task_struct *p)
1913 if (task_has_rt_policy(p))
1914 prio = MAX_RT_PRIO-1 - p->rt_priority;
1916 prio = __normal_prio(p);
1921 * Calculate the current priority, i.e. the priority
1922 * taken into account by the scheduler. This value might
1923 * be boosted by RT tasks, or might be boosted by
1924 * interactivity modifiers. Will be RT if the task got
1925 * RT-boosted. If not then it returns p->normal_prio.
1927 static int effective_prio(struct task_struct *p)
1929 p->normal_prio = normal_prio(p);
1931 * If we are RT tasks or we were boosted to RT priority,
1932 * keep the priority unchanged. Otherwise, update priority
1933 * to the normal priority:
1935 if (!rt_prio(p->prio))
1936 return p->normal_prio;
1941 * activate_task - move a task to the runqueue.
1943 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1945 if (task_contributes_to_load(p))
1946 rq->nr_uninterruptible--;
1948 enqueue_task(rq, p, wakeup);
1953 * deactivate_task - remove a task from the runqueue.
1955 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1957 if (task_contributes_to_load(p))
1958 rq->nr_uninterruptible++;
1960 dequeue_task(rq, p, sleep);
1965 * task_curr - is this task currently executing on a CPU?
1966 * @p: the task in question.
1968 inline int task_curr(const struct task_struct *p)
1970 return cpu_curr(task_cpu(p)) == p;
1973 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1975 set_task_rq(p, cpu);
1978 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1979 * successfuly executed on another CPU. We must ensure that updates of
1980 * per-task data have been completed by this moment.
1983 task_thread_info(p)->cpu = cpu;
1987 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1988 const struct sched_class *prev_class,
1989 int oldprio, int running)
1991 if (prev_class != p->sched_class) {
1992 if (prev_class->switched_from)
1993 prev_class->switched_from(rq, p, running);
1994 p->sched_class->switched_to(rq, p, running);
1996 p->sched_class->prio_changed(rq, p, oldprio, running);
2001 * Is this task likely cache-hot:
2004 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2009 * Buddy candidates are cache hot:
2011 if (sched_feat(CACHE_HOT_BUDDY) &&
2012 (&p->se == cfs_rq_of(&p->se)->next ||
2013 &p->se == cfs_rq_of(&p->se)->last))
2016 if (p->sched_class != &fair_sched_class)
2019 if (sysctl_sched_migration_cost == -1)
2021 if (sysctl_sched_migration_cost == 0)
2024 delta = now - p->se.exec_start;
2026 return delta < (s64)sysctl_sched_migration_cost;
2030 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2032 int old_cpu = task_cpu(p);
2033 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2034 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2035 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2038 clock_offset = old_rq->clock - new_rq->clock;
2040 trace_sched_migrate_task(p, new_cpu);
2042 #ifdef CONFIG_SCHEDSTATS
2043 if (p->se.wait_start)
2044 p->se.wait_start -= clock_offset;
2045 if (p->se.sleep_start)
2046 p->se.sleep_start -= clock_offset;
2047 if (p->se.block_start)
2048 p->se.block_start -= clock_offset;
2050 if (old_cpu != new_cpu) {
2051 p->se.nr_migrations++;
2052 #ifdef CONFIG_SCHEDSTATS
2053 if (task_hot(p, old_rq->clock, NULL))
2054 schedstat_inc(p, se.nr_forced2_migrations);
2056 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2059 p->se.vruntime -= old_cfsrq->min_vruntime -
2060 new_cfsrq->min_vruntime;
2062 __set_task_cpu(p, new_cpu);
2065 struct migration_req {
2066 struct list_head list;
2068 struct task_struct *task;
2071 struct completion done;
2075 * The task's runqueue lock must be held.
2076 * Returns true if you have to wait for migration thread.
2079 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2081 struct rq *rq = task_rq(p);
2084 * If the task is not on a runqueue (and not running), then
2085 * it is sufficient to simply update the task's cpu field.
2087 if (!p->se.on_rq && !task_running(rq, p)) {
2088 set_task_cpu(p, dest_cpu);
2092 init_completion(&req->done);
2094 req->dest_cpu = dest_cpu;
2095 list_add(&req->list, &rq->migration_queue);
2101 * wait_task_context_switch - wait for a thread to complete at least one
2104 * @p must not be current.
2106 void wait_task_context_switch(struct task_struct *p)
2108 unsigned long nvcsw, nivcsw, flags;
2116 * The runqueue is assigned before the actual context
2117 * switch. We need to take the runqueue lock.
2119 * We could check initially without the lock but it is
2120 * very likely that we need to take the lock in every
2123 rq = task_rq_lock(p, &flags);
2124 running = task_running(rq, p);
2125 task_rq_unlock(rq, &flags);
2127 if (likely(!running))
2130 * The switch count is incremented before the actual
2131 * context switch. We thus wait for two switches to be
2132 * sure at least one completed.
2134 if ((p->nvcsw - nvcsw) > 1)
2136 if ((p->nivcsw - nivcsw) > 1)
2144 * wait_task_inactive - wait for a thread to unschedule.
2146 * If @match_state is nonzero, it's the @p->state value just checked and
2147 * not expected to change. If it changes, i.e. @p might have woken up,
2148 * then return zero. When we succeed in waiting for @p to be off its CPU,
2149 * we return a positive number (its total switch count). If a second call
2150 * a short while later returns the same number, the caller can be sure that
2151 * @p has remained unscheduled the whole time.
2153 * The caller must ensure that the task *will* unschedule sometime soon,
2154 * else this function might spin for a *long* time. This function can't
2155 * be called with interrupts off, or it may introduce deadlock with
2156 * smp_call_function() if an IPI is sent by the same process we are
2157 * waiting to become inactive.
2159 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2161 unsigned long flags;
2168 * We do the initial early heuristics without holding
2169 * any task-queue locks at all. We'll only try to get
2170 * the runqueue lock when things look like they will
2176 * If the task is actively running on another CPU
2177 * still, just relax and busy-wait without holding
2180 * NOTE! Since we don't hold any locks, it's not
2181 * even sure that "rq" stays as the right runqueue!
2182 * But we don't care, since "task_running()" will
2183 * return false if the runqueue has changed and p
2184 * is actually now running somewhere else!
2186 while (task_running(rq, p)) {
2187 if (match_state && unlikely(p->state != match_state))
2193 * Ok, time to look more closely! We need the rq
2194 * lock now, to be *sure*. If we're wrong, we'll
2195 * just go back and repeat.
2197 rq = task_rq_lock(p, &flags);
2198 trace_sched_wait_task(rq, p);
2199 running = task_running(rq, p);
2200 on_rq = p->se.on_rq;
2202 if (!match_state || p->state == match_state)
2203 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2204 task_rq_unlock(rq, &flags);
2207 * If it changed from the expected state, bail out now.
2209 if (unlikely(!ncsw))
2213 * Was it really running after all now that we
2214 * checked with the proper locks actually held?
2216 * Oops. Go back and try again..
2218 if (unlikely(running)) {
2224 * It's not enough that it's not actively running,
2225 * it must be off the runqueue _entirely_, and not
2228 * So if it was still runnable (but just not actively
2229 * running right now), it's preempted, and we should
2230 * yield - it could be a while.
2232 if (unlikely(on_rq)) {
2233 schedule_timeout_uninterruptible(1);
2238 * Ahh, all good. It wasn't running, and it wasn't
2239 * runnable, which means that it will never become
2240 * running in the future either. We're all done!
2249 * kick_process - kick a running thread to enter/exit the kernel
2250 * @p: the to-be-kicked thread
2252 * Cause a process which is running on another CPU to enter
2253 * kernel-mode, without any delay. (to get signals handled.)
2255 * NOTE: this function doesnt have to take the runqueue lock,
2256 * because all it wants to ensure is that the remote task enters
2257 * the kernel. If the IPI races and the task has been migrated
2258 * to another CPU then no harm is done and the purpose has been
2261 void kick_process(struct task_struct *p)
2267 if ((cpu != smp_processor_id()) && task_curr(p))
2268 smp_send_reschedule(cpu);
2271 EXPORT_SYMBOL_GPL(kick_process);
2272 #endif /* CONFIG_SMP */
2275 * task_oncpu_function_call - call a function on the cpu on which a task runs
2276 * @p: the task to evaluate
2277 * @func: the function to be called
2278 * @info: the function call argument
2280 * Calls the function @func when the task is currently running. This might
2281 * be on the current CPU, which just calls the function directly
2283 void task_oncpu_function_call(struct task_struct *p,
2284 void (*func) (void *info), void *info)
2291 smp_call_function_single(cpu, func, info, 1);
2296 * try_to_wake_up - wake up a thread
2297 * @p: the to-be-woken-up thread
2298 * @state: the mask of task states that can be woken
2299 * @sync: do a synchronous wakeup?
2301 * Put it on the run-queue if it's not already there. The "current"
2302 * thread is always on the run-queue (except when the actual
2303 * re-schedule is in progress), and as such you're allowed to do
2304 * the simpler "current->state = TASK_RUNNING" to mark yourself
2305 * runnable without the overhead of this.
2307 * returns failure only if the task is already active.
2309 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2312 int cpu, orig_cpu, this_cpu, success = 0;
2313 unsigned long flags;
2314 struct rq *rq, *orig_rq;
2316 if (!sched_feat(SYNC_WAKEUPS))
2317 wake_flags &= ~WF_SYNC;
2319 this_cpu = get_cpu();
2322 rq = orig_rq = task_rq_lock(p, &flags);
2323 update_rq_clock(rq);
2324 if (!(p->state & state))
2334 if (unlikely(task_running(rq, p)))
2338 * In order to handle concurrent wakeups and release the rq->lock
2339 * we put the task in TASK_WAKING state.
2341 * First fix up the nr_uninterruptible count:
2343 if (task_contributes_to_load(p))
2344 rq->nr_uninterruptible--;
2345 p->state = TASK_WAKING;
2346 task_rq_unlock(rq, &flags);
2348 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2349 if (cpu != orig_cpu)
2350 set_task_cpu(p, cpu);
2352 rq = task_rq_lock(p, &flags);
2355 update_rq_clock(rq);
2357 if (rq->idle_stamp) {
2358 u64 delta = rq->clock - rq->idle_stamp;
2359 u64 max = 2*sysctl_sched_migration_cost;
2364 update_avg(&rq->avg_idle, delta);
2368 WARN_ON(p->state != TASK_WAKING);
2371 #ifdef CONFIG_SCHEDSTATS
2372 schedstat_inc(rq, ttwu_count);
2373 if (cpu == this_cpu)
2374 schedstat_inc(rq, ttwu_local);
2376 struct sched_domain *sd;
2377 for_each_domain(this_cpu, sd) {
2378 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2379 schedstat_inc(sd, ttwu_wake_remote);
2384 #endif /* CONFIG_SCHEDSTATS */
2387 #endif /* CONFIG_SMP */
2388 schedstat_inc(p, se.nr_wakeups);
2389 if (wake_flags & WF_SYNC)
2390 schedstat_inc(p, se.nr_wakeups_sync);
2391 if (orig_cpu != cpu)
2392 schedstat_inc(p, se.nr_wakeups_migrate);
2393 if (cpu == this_cpu)
2394 schedstat_inc(p, se.nr_wakeups_local);
2396 schedstat_inc(p, se.nr_wakeups_remote);
2397 activate_task(rq, p, 1);
2401 * Only attribute actual wakeups done by this task.
2403 if (!in_interrupt()) {
2404 struct sched_entity *se = ¤t->se;
2405 u64 sample = se->sum_exec_runtime;
2407 if (se->last_wakeup)
2408 sample -= se->last_wakeup;
2410 sample -= se->start_runtime;
2411 update_avg(&se->avg_wakeup, sample);
2413 se->last_wakeup = se->sum_exec_runtime;
2417 trace_sched_wakeup(rq, p, success);
2418 check_preempt_curr(rq, p, wake_flags);
2420 p->state = TASK_RUNNING;
2422 if (p->sched_class->task_wake_up)
2423 p->sched_class->task_wake_up(rq, p);
2426 task_rq_unlock(rq, &flags);
2433 * wake_up_process - Wake up a specific process
2434 * @p: The process to be woken up.
2436 * Attempt to wake up the nominated process and move it to the set of runnable
2437 * processes. Returns 1 if the process was woken up, 0 if it was already
2440 * It may be assumed that this function implies a write memory barrier before
2441 * changing the task state if and only if any tasks are woken up.
2443 int wake_up_process(struct task_struct *p)
2445 return try_to_wake_up(p, TASK_ALL, 0);
2447 EXPORT_SYMBOL(wake_up_process);
2449 int wake_up_state(struct task_struct *p, unsigned int state)
2451 return try_to_wake_up(p, state, 0);
2455 * Perform scheduler related setup for a newly forked process p.
2456 * p is forked by current.
2458 * __sched_fork() is basic setup used by init_idle() too:
2460 static void __sched_fork(struct task_struct *p)
2462 p->se.exec_start = 0;
2463 p->se.sum_exec_runtime = 0;
2464 p->se.prev_sum_exec_runtime = 0;
2465 p->se.nr_migrations = 0;
2466 p->se.last_wakeup = 0;
2467 p->se.avg_overlap = 0;
2468 p->se.start_runtime = 0;
2469 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2470 p->se.avg_running = 0;
2472 #ifdef CONFIG_SCHEDSTATS
2473 p->se.wait_start = 0;
2475 p->se.wait_count = 0;
2478 p->se.sleep_start = 0;
2479 p->se.sleep_max = 0;
2480 p->se.sum_sleep_runtime = 0;
2482 p->se.block_start = 0;
2483 p->se.block_max = 0;
2485 p->se.slice_max = 0;
2487 p->se.nr_migrations_cold = 0;
2488 p->se.nr_failed_migrations_affine = 0;
2489 p->se.nr_failed_migrations_running = 0;
2490 p->se.nr_failed_migrations_hot = 0;
2491 p->se.nr_forced_migrations = 0;
2492 p->se.nr_forced2_migrations = 0;
2494 p->se.nr_wakeups = 0;
2495 p->se.nr_wakeups_sync = 0;
2496 p->se.nr_wakeups_migrate = 0;
2497 p->se.nr_wakeups_local = 0;
2498 p->se.nr_wakeups_remote = 0;
2499 p->se.nr_wakeups_affine = 0;
2500 p->se.nr_wakeups_affine_attempts = 0;
2501 p->se.nr_wakeups_passive = 0;
2502 p->se.nr_wakeups_idle = 0;
2506 INIT_LIST_HEAD(&p->rt.run_list);
2508 INIT_LIST_HEAD(&p->se.group_node);
2510 #ifdef CONFIG_PREEMPT_NOTIFIERS
2511 INIT_HLIST_HEAD(&p->preempt_notifiers);
2515 * We mark the process as running here, but have not actually
2516 * inserted it onto the runqueue yet. This guarantees that
2517 * nobody will actually run it, and a signal or other external
2518 * event cannot wake it up and insert it on the runqueue either.
2520 p->state = TASK_RUNNING;
2524 * fork()/clone()-time setup:
2526 void sched_fork(struct task_struct *p, int clone_flags)
2528 int cpu = get_cpu();
2533 * Revert to default priority/policy on fork if requested.
2535 if (unlikely(p->sched_reset_on_fork)) {
2536 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2537 p->policy = SCHED_NORMAL;
2538 p->normal_prio = p->static_prio;
2541 if (PRIO_TO_NICE(p->static_prio) < 0) {
2542 p->static_prio = NICE_TO_PRIO(0);
2543 p->normal_prio = p->static_prio;
2548 * We don't need the reset flag anymore after the fork. It has
2549 * fulfilled its duty:
2551 p->sched_reset_on_fork = 0;
2555 * Make sure we do not leak PI boosting priority to the child.
2557 p->prio = current->normal_prio;
2559 if (!rt_prio(p->prio))
2560 p->sched_class = &fair_sched_class;
2563 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_FORK, 0);
2565 set_task_cpu(p, cpu);
2567 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2568 if (likely(sched_info_on()))
2569 memset(&p->sched_info, 0, sizeof(p->sched_info));
2571 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2574 #ifdef CONFIG_PREEMPT
2575 /* Want to start with kernel preemption disabled. */
2576 task_thread_info(p)->preempt_count = 1;
2578 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2584 * wake_up_new_task - wake up a newly created task for the first time.
2586 * This function will do some initial scheduler statistics housekeeping
2587 * that must be done for every newly created context, then puts the task
2588 * on the runqueue and wakes it.
2590 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2592 unsigned long flags;
2595 rq = task_rq_lock(p, &flags);
2596 BUG_ON(p->state != TASK_RUNNING);
2597 update_rq_clock(rq);
2599 if (!p->sched_class->task_new || !current->se.on_rq) {
2600 activate_task(rq, p, 0);
2603 * Let the scheduling class do new task startup
2604 * management (if any):
2606 p->sched_class->task_new(rq, p);
2609 trace_sched_wakeup_new(rq, p, 1);
2610 check_preempt_curr(rq, p, WF_FORK);
2612 if (p->sched_class->task_wake_up)
2613 p->sched_class->task_wake_up(rq, p);
2615 task_rq_unlock(rq, &flags);
2618 #ifdef CONFIG_PREEMPT_NOTIFIERS
2621 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2622 * @notifier: notifier struct to register
2624 void preempt_notifier_register(struct preempt_notifier *notifier)
2626 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2628 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2631 * preempt_notifier_unregister - no longer interested in preemption notifications
2632 * @notifier: notifier struct to unregister
2634 * This is safe to call from within a preemption notifier.
2636 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2638 hlist_del(¬ifier->link);
2640 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2642 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2644 struct preempt_notifier *notifier;
2645 struct hlist_node *node;
2647 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2648 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2652 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2653 struct task_struct *next)
2655 struct preempt_notifier *notifier;
2656 struct hlist_node *node;
2658 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2659 notifier->ops->sched_out(notifier, next);
2662 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2664 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2669 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2670 struct task_struct *next)
2674 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2677 * prepare_task_switch - prepare to switch tasks
2678 * @rq: the runqueue preparing to switch
2679 * @prev: the current task that is being switched out
2680 * @next: the task we are going to switch to.
2682 * This is called with the rq lock held and interrupts off. It must
2683 * be paired with a subsequent finish_task_switch after the context
2686 * prepare_task_switch sets up locking and calls architecture specific
2690 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2691 struct task_struct *next)
2693 fire_sched_out_preempt_notifiers(prev, next);
2694 prepare_lock_switch(rq, next);
2695 prepare_arch_switch(next);
2699 * finish_task_switch - clean up after a task-switch
2700 * @rq: runqueue associated with task-switch
2701 * @prev: the thread we just switched away from.
2703 * finish_task_switch must be called after the context switch, paired
2704 * with a prepare_task_switch call before the context switch.
2705 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2706 * and do any other architecture-specific cleanup actions.
2708 * Note that we may have delayed dropping an mm in context_switch(). If
2709 * so, we finish that here outside of the runqueue lock. (Doing it
2710 * with the lock held can cause deadlocks; see schedule() for
2713 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2714 __releases(rq->lock)
2716 struct mm_struct *mm = rq->prev_mm;
2722 * A task struct has one reference for the use as "current".
2723 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2724 * schedule one last time. The schedule call will never return, and
2725 * the scheduled task must drop that reference.
2726 * The test for TASK_DEAD must occur while the runqueue locks are
2727 * still held, otherwise prev could be scheduled on another cpu, die
2728 * there before we look at prev->state, and then the reference would
2730 * Manfred Spraul <manfred@colorfullife.com>
2732 prev_state = prev->state;
2733 finish_arch_switch(prev);
2734 perf_event_task_sched_in(current, cpu_of(rq));
2735 finish_lock_switch(rq, prev);
2737 fire_sched_in_preempt_notifiers(current);
2740 if (unlikely(prev_state == TASK_DEAD)) {
2742 * Remove function-return probe instances associated with this
2743 * task and put them back on the free list.
2745 kprobe_flush_task(prev);
2746 put_task_struct(prev);
2752 /* assumes rq->lock is held */
2753 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2755 if (prev->sched_class->pre_schedule)
2756 prev->sched_class->pre_schedule(rq, prev);
2759 /* rq->lock is NOT held, but preemption is disabled */
2760 static inline void post_schedule(struct rq *rq)
2762 if (rq->post_schedule) {
2763 unsigned long flags;
2765 spin_lock_irqsave(&rq->lock, flags);
2766 if (rq->curr->sched_class->post_schedule)
2767 rq->curr->sched_class->post_schedule(rq);
2768 spin_unlock_irqrestore(&rq->lock, flags);
2770 rq->post_schedule = 0;
2776 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2780 static inline void post_schedule(struct rq *rq)
2787 * schedule_tail - first thing a freshly forked thread must call.
2788 * @prev: the thread we just switched away from.
2790 asmlinkage void schedule_tail(struct task_struct *prev)
2791 __releases(rq->lock)
2793 struct rq *rq = this_rq();
2795 finish_task_switch(rq, prev);
2798 * FIXME: do we need to worry about rq being invalidated by the
2803 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2804 /* In this case, finish_task_switch does not reenable preemption */
2807 if (current->set_child_tid)
2808 put_user(task_pid_vnr(current), current->set_child_tid);
2812 * context_switch - switch to the new MM and the new
2813 * thread's register state.
2816 context_switch(struct rq *rq, struct task_struct *prev,
2817 struct task_struct *next)
2819 struct mm_struct *mm, *oldmm;
2821 prepare_task_switch(rq, prev, next);
2822 trace_sched_switch(rq, prev, next);
2824 oldmm = prev->active_mm;
2826 * For paravirt, this is coupled with an exit in switch_to to
2827 * combine the page table reload and the switch backend into
2830 arch_start_context_switch(prev);
2832 if (unlikely(!mm)) {
2833 next->active_mm = oldmm;
2834 atomic_inc(&oldmm->mm_count);
2835 enter_lazy_tlb(oldmm, next);
2837 switch_mm(oldmm, mm, next);
2839 if (unlikely(!prev->mm)) {
2840 prev->active_mm = NULL;
2841 rq->prev_mm = oldmm;
2844 * Since the runqueue lock will be released by the next
2845 * task (which is an invalid locking op but in the case
2846 * of the scheduler it's an obvious special-case), so we
2847 * do an early lockdep release here:
2849 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2850 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2853 /* Here we just switch the register state and the stack. */
2854 switch_to(prev, next, prev);
2858 * this_rq must be evaluated again because prev may have moved
2859 * CPUs since it called schedule(), thus the 'rq' on its stack
2860 * frame will be invalid.
2862 finish_task_switch(this_rq(), prev);
2866 * nr_running, nr_uninterruptible and nr_context_switches:
2868 * externally visible scheduler statistics: current number of runnable
2869 * threads, current number of uninterruptible-sleeping threads, total
2870 * number of context switches performed since bootup.
2872 unsigned long nr_running(void)
2874 unsigned long i, sum = 0;
2876 for_each_online_cpu(i)
2877 sum += cpu_rq(i)->nr_running;
2882 unsigned long nr_uninterruptible(void)
2884 unsigned long i, sum = 0;
2886 for_each_possible_cpu(i)
2887 sum += cpu_rq(i)->nr_uninterruptible;
2890 * Since we read the counters lockless, it might be slightly
2891 * inaccurate. Do not allow it to go below zero though:
2893 if (unlikely((long)sum < 0))
2899 unsigned long long nr_context_switches(void)
2902 unsigned long long sum = 0;
2904 for_each_possible_cpu(i)
2905 sum += cpu_rq(i)->nr_switches;
2910 unsigned long nr_iowait(void)
2912 unsigned long i, sum = 0;
2914 for_each_possible_cpu(i)
2915 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2920 unsigned long nr_iowait_cpu(void)
2922 struct rq *this = this_rq();
2923 return atomic_read(&this->nr_iowait);
2926 unsigned long this_cpu_load(void)
2928 struct rq *this = this_rq();
2929 return this->cpu_load[0];
2933 /* Variables and functions for calc_load */
2934 static atomic_long_t calc_load_tasks;
2935 static unsigned long calc_load_update;
2936 unsigned long avenrun[3];
2937 EXPORT_SYMBOL(avenrun);
2940 * get_avenrun - get the load average array
2941 * @loads: pointer to dest load array
2942 * @offset: offset to add
2943 * @shift: shift count to shift the result left
2945 * These values are estimates at best, so no need for locking.
2947 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2949 loads[0] = (avenrun[0] + offset) << shift;
2950 loads[1] = (avenrun[1] + offset) << shift;
2951 loads[2] = (avenrun[2] + offset) << shift;
2954 static unsigned long
2955 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2958 load += active * (FIXED_1 - exp);
2959 return load >> FSHIFT;
2963 * calc_load - update the avenrun load estimates 10 ticks after the
2964 * CPUs have updated calc_load_tasks.
2966 void calc_global_load(void)
2968 unsigned long upd = calc_load_update + 10;
2971 if (time_before(jiffies, upd))
2974 active = atomic_long_read(&calc_load_tasks);
2975 active = active > 0 ? active * FIXED_1 : 0;
2977 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2978 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2979 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2981 calc_load_update += LOAD_FREQ;
2985 * Either called from update_cpu_load() or from a cpu going idle
2987 static void calc_load_account_active(struct rq *this_rq)
2989 long nr_active, delta;
2991 nr_active = this_rq->nr_running;
2992 nr_active += (long) this_rq->nr_uninterruptible;
2994 if (nr_active != this_rq->calc_load_active) {
2995 delta = nr_active - this_rq->calc_load_active;
2996 this_rq->calc_load_active = nr_active;
2997 atomic_long_add(delta, &calc_load_tasks);
3002 * Update rq->cpu_load[] statistics. This function is usually called every
3003 * scheduler tick (TICK_NSEC).
3005 static void update_cpu_load(struct rq *this_rq)
3007 unsigned long this_load = this_rq->load.weight;
3010 this_rq->nr_load_updates++;
3012 /* Update our load: */
3013 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3014 unsigned long old_load, new_load;
3016 /* scale is effectively 1 << i now, and >> i divides by scale */
3018 old_load = this_rq->cpu_load[i];
3019 new_load = this_load;
3021 * Round up the averaging division if load is increasing. This
3022 * prevents us from getting stuck on 9 if the load is 10, for
3025 if (new_load > old_load)
3026 new_load += scale-1;
3027 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3030 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3031 this_rq->calc_load_update += LOAD_FREQ;
3032 calc_load_account_active(this_rq);
3039 * double_rq_lock - safely lock two runqueues
3041 * Note this does not disable interrupts like task_rq_lock,
3042 * you need to do so manually before calling.
3044 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3045 __acquires(rq1->lock)
3046 __acquires(rq2->lock)
3048 BUG_ON(!irqs_disabled());
3050 spin_lock(&rq1->lock);
3051 __acquire(rq2->lock); /* Fake it out ;) */
3054 spin_lock(&rq1->lock);
3055 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3057 spin_lock(&rq2->lock);
3058 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3061 update_rq_clock(rq1);
3062 update_rq_clock(rq2);
3066 * double_rq_unlock - safely unlock two runqueues
3068 * Note this does not restore interrupts like task_rq_unlock,
3069 * you need to do so manually after calling.
3071 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3072 __releases(rq1->lock)
3073 __releases(rq2->lock)
3075 spin_unlock(&rq1->lock);
3077 spin_unlock(&rq2->lock);
3079 __release(rq2->lock);
3083 * If dest_cpu is allowed for this process, migrate the task to it.
3084 * This is accomplished by forcing the cpu_allowed mask to only
3085 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3086 * the cpu_allowed mask is restored.
3088 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3090 struct migration_req req;
3091 unsigned long flags;
3094 rq = task_rq_lock(p, &flags);
3095 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3096 || unlikely(!cpu_active(dest_cpu)))
3099 /* force the process onto the specified CPU */
3100 if (migrate_task(p, dest_cpu, &req)) {
3101 /* Need to wait for migration thread (might exit: take ref). */
3102 struct task_struct *mt = rq->migration_thread;
3104 get_task_struct(mt);
3105 task_rq_unlock(rq, &flags);
3106 wake_up_process(mt);
3107 put_task_struct(mt);
3108 wait_for_completion(&req.done);
3113 task_rq_unlock(rq, &flags);
3117 * sched_exec - execve() is a valuable balancing opportunity, because at
3118 * this point the task has the smallest effective memory and cache footprint.
3120 void sched_exec(void)
3122 int new_cpu, this_cpu = get_cpu();
3123 new_cpu = current->sched_class->select_task_rq(current, SD_BALANCE_EXEC, 0);
3125 if (new_cpu != this_cpu)
3126 sched_migrate_task(current, new_cpu);
3130 * pull_task - move a task from a remote runqueue to the local runqueue.
3131 * Both runqueues must be locked.
3133 static void pull_task(struct rq *src_rq, struct task_struct *p,
3134 struct rq *this_rq, int this_cpu)
3136 deactivate_task(src_rq, p, 0);
3137 set_task_cpu(p, this_cpu);
3138 activate_task(this_rq, p, 0);
3140 * Note that idle threads have a prio of MAX_PRIO, for this test
3141 * to be always true for them.
3143 check_preempt_curr(this_rq, p, 0);
3147 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3150 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3151 struct sched_domain *sd, enum cpu_idle_type idle,
3154 int tsk_cache_hot = 0;
3156 * We do not migrate tasks that are:
3157 * 1) running (obviously), or
3158 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3159 * 3) are cache-hot on their current CPU.
3161 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3162 schedstat_inc(p, se.nr_failed_migrations_affine);
3167 if (task_running(rq, p)) {
3168 schedstat_inc(p, se.nr_failed_migrations_running);
3173 * Aggressive migration if:
3174 * 1) task is cache cold, or
3175 * 2) too many balance attempts have failed.
3178 tsk_cache_hot = task_hot(p, rq->clock, sd);
3179 if (!tsk_cache_hot ||
3180 sd->nr_balance_failed > sd->cache_nice_tries) {
3181 #ifdef CONFIG_SCHEDSTATS
3182 if (tsk_cache_hot) {
3183 schedstat_inc(sd, lb_hot_gained[idle]);
3184 schedstat_inc(p, se.nr_forced_migrations);
3190 if (tsk_cache_hot) {
3191 schedstat_inc(p, se.nr_failed_migrations_hot);
3197 static unsigned long
3198 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3199 unsigned long max_load_move, struct sched_domain *sd,
3200 enum cpu_idle_type idle, int *all_pinned,
3201 int *this_best_prio, struct rq_iterator *iterator)
3203 int loops = 0, pulled = 0, pinned = 0;
3204 struct task_struct *p;
3205 long rem_load_move = max_load_move;
3207 if (max_load_move == 0)
3213 * Start the load-balancing iterator:
3215 p = iterator->start(iterator->arg);
3217 if (!p || loops++ > sysctl_sched_nr_migrate)
3220 if ((p->se.load.weight >> 1) > rem_load_move ||
3221 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3222 p = iterator->next(iterator->arg);
3226 pull_task(busiest, p, this_rq, this_cpu);
3228 rem_load_move -= p->se.load.weight;
3230 #ifdef CONFIG_PREEMPT
3232 * NEWIDLE balancing is a source of latency, so preemptible kernels
3233 * will stop after the first task is pulled to minimize the critical
3236 if (idle == CPU_NEWLY_IDLE)
3241 * We only want to steal up to the prescribed amount of weighted load.
3243 if (rem_load_move > 0) {
3244 if (p->prio < *this_best_prio)
3245 *this_best_prio = p->prio;
3246 p = iterator->next(iterator->arg);
3251 * Right now, this is one of only two places pull_task() is called,
3252 * so we can safely collect pull_task() stats here rather than
3253 * inside pull_task().
3255 schedstat_add(sd, lb_gained[idle], pulled);
3258 *all_pinned = pinned;
3260 return max_load_move - rem_load_move;
3264 * move_tasks tries to move up to max_load_move weighted load from busiest to
3265 * this_rq, as part of a balancing operation within domain "sd".
3266 * Returns 1 if successful and 0 otherwise.
3268 * Called with both runqueues locked.
3270 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3271 unsigned long max_load_move,
3272 struct sched_domain *sd, enum cpu_idle_type idle,
3275 const struct sched_class *class = sched_class_highest;
3276 unsigned long total_load_moved = 0;
3277 int this_best_prio = this_rq->curr->prio;
3281 class->load_balance(this_rq, this_cpu, busiest,
3282 max_load_move - total_load_moved,
3283 sd, idle, all_pinned, &this_best_prio);
3284 class = class->next;
3286 #ifdef CONFIG_PREEMPT
3288 * NEWIDLE balancing is a source of latency, so preemptible
3289 * kernels will stop after the first task is pulled to minimize
3290 * the critical section.
3292 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3295 } while (class && max_load_move > total_load_moved);
3297 return total_load_moved > 0;
3301 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3302 struct sched_domain *sd, enum cpu_idle_type idle,
3303 struct rq_iterator *iterator)
3305 struct task_struct *p = iterator->start(iterator->arg);
3309 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3310 pull_task(busiest, p, this_rq, this_cpu);
3312 * Right now, this is only the second place pull_task()
3313 * is called, so we can safely collect pull_task()
3314 * stats here rather than inside pull_task().
3316 schedstat_inc(sd, lb_gained[idle]);
3320 p = iterator->next(iterator->arg);
3327 * move_one_task tries to move exactly one task from busiest to this_rq, as
3328 * part of active balancing operations within "domain".
3329 * Returns 1 if successful and 0 otherwise.
3331 * Called with both runqueues locked.
3333 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3334 struct sched_domain *sd, enum cpu_idle_type idle)
3336 const struct sched_class *class;
3338 for_each_class(class) {
3339 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3345 /********** Helpers for find_busiest_group ************************/
3347 * sd_lb_stats - Structure to store the statistics of a sched_domain
3348 * during load balancing.
3350 struct sd_lb_stats {
3351 struct sched_group *busiest; /* Busiest group in this sd */
3352 struct sched_group *this; /* Local group in this sd */
3353 unsigned long total_load; /* Total load of all groups in sd */
3354 unsigned long total_pwr; /* Total power of all groups in sd */
3355 unsigned long avg_load; /* Average load across all groups in sd */
3357 /** Statistics of this group */
3358 unsigned long this_load;
3359 unsigned long this_load_per_task;
3360 unsigned long this_nr_running;
3362 /* Statistics of the busiest group */
3363 unsigned long max_load;
3364 unsigned long busiest_load_per_task;
3365 unsigned long busiest_nr_running;
3367 int group_imb; /* Is there imbalance in this sd */
3368 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3369 int power_savings_balance; /* Is powersave balance needed for this sd */
3370 struct sched_group *group_min; /* Least loaded group in sd */
3371 struct sched_group *group_leader; /* Group which relieves group_min */
3372 unsigned long min_load_per_task; /* load_per_task in group_min */
3373 unsigned long leader_nr_running; /* Nr running of group_leader */
3374 unsigned long min_nr_running; /* Nr running of group_min */
3379 * sg_lb_stats - stats of a sched_group required for load_balancing
3381 struct sg_lb_stats {
3382 unsigned long avg_load; /*Avg load across the CPUs of the group */
3383 unsigned long group_load; /* Total load over the CPUs of the group */
3384 unsigned long sum_nr_running; /* Nr tasks running in the group */
3385 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3386 unsigned long group_capacity;
3387 int group_imb; /* Is there an imbalance in the group ? */
3391 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3392 * @group: The group whose first cpu is to be returned.
3394 static inline unsigned int group_first_cpu(struct sched_group *group)
3396 return cpumask_first(sched_group_cpus(group));
3400 * get_sd_load_idx - Obtain the load index for a given sched domain.
3401 * @sd: The sched_domain whose load_idx is to be obtained.
3402 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3404 static inline int get_sd_load_idx(struct sched_domain *sd,
3405 enum cpu_idle_type idle)
3411 load_idx = sd->busy_idx;
3414 case CPU_NEWLY_IDLE:
3415 load_idx = sd->newidle_idx;
3418 load_idx = sd->idle_idx;
3426 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3428 * init_sd_power_savings_stats - Initialize power savings statistics for
3429 * the given sched_domain, during load balancing.
3431 * @sd: Sched domain whose power-savings statistics are to be initialized.
3432 * @sds: Variable containing the statistics for sd.
3433 * @idle: Idle status of the CPU at which we're performing load-balancing.
3435 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3436 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3439 * Busy processors will not participate in power savings
3442 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3443 sds->power_savings_balance = 0;
3445 sds->power_savings_balance = 1;
3446 sds->min_nr_running = ULONG_MAX;
3447 sds->leader_nr_running = 0;
3452 * update_sd_power_savings_stats - Update the power saving stats for a
3453 * sched_domain while performing load balancing.
3455 * @group: sched_group belonging to the sched_domain under consideration.
3456 * @sds: Variable containing the statistics of the sched_domain
3457 * @local_group: Does group contain the CPU for which we're performing
3459 * @sgs: Variable containing the statistics of the group.
3461 static inline void update_sd_power_savings_stats(struct sched_group *group,
3462 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3465 if (!sds->power_savings_balance)
3469 * If the local group is idle or completely loaded
3470 * no need to do power savings balance at this domain
3472 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3473 !sds->this_nr_running))
3474 sds->power_savings_balance = 0;
3477 * If a group is already running at full capacity or idle,
3478 * don't include that group in power savings calculations
3480 if (!sds->power_savings_balance ||
3481 sgs->sum_nr_running >= sgs->group_capacity ||
3482 !sgs->sum_nr_running)
3486 * Calculate the group which has the least non-idle load.
3487 * This is the group from where we need to pick up the load
3490 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3491 (sgs->sum_nr_running == sds->min_nr_running &&
3492 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3493 sds->group_min = group;
3494 sds->min_nr_running = sgs->sum_nr_running;
3495 sds->min_load_per_task = sgs->sum_weighted_load /
3496 sgs->sum_nr_running;
3500 * Calculate the group which is almost near its
3501 * capacity but still has some space to pick up some load
3502 * from other group and save more power
3504 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3507 if (sgs->sum_nr_running > sds->leader_nr_running ||
3508 (sgs->sum_nr_running == sds->leader_nr_running &&
3509 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3510 sds->group_leader = group;
3511 sds->leader_nr_running = sgs->sum_nr_running;
3516 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3517 * @sds: Variable containing the statistics of the sched_domain
3518 * under consideration.
3519 * @this_cpu: Cpu at which we're currently performing load-balancing.
3520 * @imbalance: Variable to store the imbalance.
3523 * Check if we have potential to perform some power-savings balance.
3524 * If yes, set the busiest group to be the least loaded group in the
3525 * sched_domain, so that it's CPUs can be put to idle.
3527 * Returns 1 if there is potential to perform power-savings balance.
3530 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3531 int this_cpu, unsigned long *imbalance)
3533 if (!sds->power_savings_balance)
3536 if (sds->this != sds->group_leader ||
3537 sds->group_leader == sds->group_min)
3540 *imbalance = sds->min_load_per_task;
3541 sds->busiest = sds->group_min;
3546 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3547 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3548 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3553 static inline void update_sd_power_savings_stats(struct sched_group *group,
3554 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3559 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3560 int this_cpu, unsigned long *imbalance)
3564 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3567 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3569 return SCHED_LOAD_SCALE;
3572 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3574 return default_scale_freq_power(sd, cpu);
3577 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3579 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3580 unsigned long smt_gain = sd->smt_gain;
3587 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3589 return default_scale_smt_power(sd, cpu);
3592 unsigned long scale_rt_power(int cpu)
3594 struct rq *rq = cpu_rq(cpu);
3595 u64 total, available;
3597 sched_avg_update(rq);
3599 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3600 available = total - rq->rt_avg;
3602 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3603 total = SCHED_LOAD_SCALE;
3605 total >>= SCHED_LOAD_SHIFT;
3607 return div_u64(available, total);
3610 static void update_cpu_power(struct sched_domain *sd, int cpu)
3612 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3613 unsigned long power = SCHED_LOAD_SCALE;
3614 struct sched_group *sdg = sd->groups;
3616 if (sched_feat(ARCH_POWER))
3617 power *= arch_scale_freq_power(sd, cpu);
3619 power *= default_scale_freq_power(sd, cpu);
3621 power >>= SCHED_LOAD_SHIFT;
3623 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3624 if (sched_feat(ARCH_POWER))
3625 power *= arch_scale_smt_power(sd, cpu);
3627 power *= default_scale_smt_power(sd, cpu);
3629 power >>= SCHED_LOAD_SHIFT;
3632 power *= scale_rt_power(cpu);
3633 power >>= SCHED_LOAD_SHIFT;
3638 sdg->cpu_power = power;
3641 static void update_group_power(struct sched_domain *sd, int cpu)
3643 struct sched_domain *child = sd->child;
3644 struct sched_group *group, *sdg = sd->groups;
3645 unsigned long power;
3648 update_cpu_power(sd, cpu);
3654 group = child->groups;
3656 power += group->cpu_power;
3657 group = group->next;
3658 } while (group != child->groups);
3660 sdg->cpu_power = power;
3664 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3665 * @sd: The sched_domain whose statistics are to be updated.
3666 * @group: sched_group whose statistics are to be updated.
3667 * @this_cpu: Cpu for which load balance is currently performed.
3668 * @idle: Idle status of this_cpu
3669 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3670 * @sd_idle: Idle status of the sched_domain containing group.
3671 * @local_group: Does group contain this_cpu.
3672 * @cpus: Set of cpus considered for load balancing.
3673 * @balance: Should we balance.
3674 * @sgs: variable to hold the statistics for this group.
3676 static inline void update_sg_lb_stats(struct sched_domain *sd,
3677 struct sched_group *group, int this_cpu,
3678 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3679 int local_group, const struct cpumask *cpus,
3680 int *balance, struct sg_lb_stats *sgs)
3682 unsigned long load, max_cpu_load, min_cpu_load;
3684 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3685 unsigned long sum_avg_load_per_task;
3686 unsigned long avg_load_per_task;
3689 balance_cpu = group_first_cpu(group);
3690 if (balance_cpu == this_cpu)
3691 update_group_power(sd, this_cpu);
3694 /* Tally up the load of all CPUs in the group */
3695 sum_avg_load_per_task = avg_load_per_task = 0;
3697 min_cpu_load = ~0UL;
3699 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3700 struct rq *rq = cpu_rq(i);
3702 if (*sd_idle && rq->nr_running)
3705 /* Bias balancing toward cpus of our domain */
3707 if (idle_cpu(i) && !first_idle_cpu) {
3712 load = target_load(i, load_idx);
3714 load = source_load(i, load_idx);
3715 if (load > max_cpu_load)
3716 max_cpu_load = load;
3717 if (min_cpu_load > load)
3718 min_cpu_load = load;
3721 sgs->group_load += load;
3722 sgs->sum_nr_running += rq->nr_running;
3723 sgs->sum_weighted_load += weighted_cpuload(i);
3725 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3729 * First idle cpu or the first cpu(busiest) in this sched group
3730 * is eligible for doing load balancing at this and above
3731 * domains. In the newly idle case, we will allow all the cpu's
3732 * to do the newly idle load balance.
3734 if (idle != CPU_NEWLY_IDLE && local_group &&
3735 balance_cpu != this_cpu && balance) {
3740 /* Adjust by relative CPU power of the group */
3741 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3745 * Consider the group unbalanced when the imbalance is larger
3746 * than the average weight of two tasks.
3748 * APZ: with cgroup the avg task weight can vary wildly and
3749 * might not be a suitable number - should we keep a
3750 * normalized nr_running number somewhere that negates
3753 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3756 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3759 sgs->group_capacity =
3760 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3764 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3765 * @sd: sched_domain whose statistics are to be updated.
3766 * @this_cpu: Cpu for which load balance is currently performed.
3767 * @idle: Idle status of this_cpu
3768 * @sd_idle: Idle status of the sched_domain containing group.
3769 * @cpus: Set of cpus considered for load balancing.
3770 * @balance: Should we balance.
3771 * @sds: variable to hold the statistics for this sched_domain.
3773 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3774 enum cpu_idle_type idle, int *sd_idle,
3775 const struct cpumask *cpus, int *balance,
3776 struct sd_lb_stats *sds)
3778 struct sched_domain *child = sd->child;
3779 struct sched_group *group = sd->groups;
3780 struct sg_lb_stats sgs;
3781 int load_idx, prefer_sibling = 0;
3783 if (child && child->flags & SD_PREFER_SIBLING)
3786 init_sd_power_savings_stats(sd, sds, idle);
3787 load_idx = get_sd_load_idx(sd, idle);
3792 local_group = cpumask_test_cpu(this_cpu,
3793 sched_group_cpus(group));
3794 memset(&sgs, 0, sizeof(sgs));
3795 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3796 local_group, cpus, balance, &sgs);
3798 if (local_group && balance && !(*balance))
3801 sds->total_load += sgs.group_load;
3802 sds->total_pwr += group->cpu_power;
3805 * In case the child domain prefers tasks go to siblings
3806 * first, lower the group capacity to one so that we'll try
3807 * and move all the excess tasks away.
3810 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3813 sds->this_load = sgs.avg_load;
3815 sds->this_nr_running = sgs.sum_nr_running;
3816 sds->this_load_per_task = sgs.sum_weighted_load;
3817 } else if (sgs.avg_load > sds->max_load &&
3818 (sgs.sum_nr_running > sgs.group_capacity ||
3820 sds->max_load = sgs.avg_load;
3821 sds->busiest = group;
3822 sds->busiest_nr_running = sgs.sum_nr_running;
3823 sds->busiest_load_per_task = sgs.sum_weighted_load;
3824 sds->group_imb = sgs.group_imb;
3827 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3828 group = group->next;
3829 } while (group != sd->groups);
3833 * fix_small_imbalance - Calculate the minor imbalance that exists
3834 * amongst the groups of a sched_domain, during
3836 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3837 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3838 * @imbalance: Variable to store the imbalance.
3840 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3841 int this_cpu, unsigned long *imbalance)
3843 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3844 unsigned int imbn = 2;
3846 if (sds->this_nr_running) {
3847 sds->this_load_per_task /= sds->this_nr_running;
3848 if (sds->busiest_load_per_task >
3849 sds->this_load_per_task)
3852 sds->this_load_per_task =
3853 cpu_avg_load_per_task(this_cpu);
3855 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3856 sds->busiest_load_per_task * imbn) {
3857 *imbalance = sds->busiest_load_per_task;
3862 * OK, we don't have enough imbalance to justify moving tasks,
3863 * however we may be able to increase total CPU power used by
3867 pwr_now += sds->busiest->cpu_power *
3868 min(sds->busiest_load_per_task, sds->max_load);
3869 pwr_now += sds->this->cpu_power *
3870 min(sds->this_load_per_task, sds->this_load);
3871 pwr_now /= SCHED_LOAD_SCALE;
3873 /* Amount of load we'd subtract */
3874 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3875 sds->busiest->cpu_power;
3876 if (sds->max_load > tmp)
3877 pwr_move += sds->busiest->cpu_power *
3878 min(sds->busiest_load_per_task, sds->max_load - tmp);
3880 /* Amount of load we'd add */
3881 if (sds->max_load * sds->busiest->cpu_power <
3882 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3883 tmp = (sds->max_load * sds->busiest->cpu_power) /
3884 sds->this->cpu_power;
3886 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3887 sds->this->cpu_power;
3888 pwr_move += sds->this->cpu_power *
3889 min(sds->this_load_per_task, sds->this_load + tmp);
3890 pwr_move /= SCHED_LOAD_SCALE;
3892 /* Move if we gain throughput */
3893 if (pwr_move > pwr_now)
3894 *imbalance = sds->busiest_load_per_task;
3898 * calculate_imbalance - Calculate the amount of imbalance present within the
3899 * groups of a given sched_domain during load balance.
3900 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3901 * @this_cpu: Cpu for which currently load balance is being performed.
3902 * @imbalance: The variable to store the imbalance.
3904 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3905 unsigned long *imbalance)
3907 unsigned long max_pull;
3909 * In the presence of smp nice balancing, certain scenarios can have
3910 * max load less than avg load(as we skip the groups at or below
3911 * its cpu_power, while calculating max_load..)
3913 if (sds->max_load < sds->avg_load) {
3915 return fix_small_imbalance(sds, this_cpu, imbalance);
3918 /* Don't want to pull so many tasks that a group would go idle */
3919 max_pull = min(sds->max_load - sds->avg_load,
3920 sds->max_load - sds->busiest_load_per_task);
3922 /* How much load to actually move to equalise the imbalance */
3923 *imbalance = min(max_pull * sds->busiest->cpu_power,
3924 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3928 * if *imbalance is less than the average load per runnable task
3929 * there is no gaurantee that any tasks will be moved so we'll have
3930 * a think about bumping its value to force at least one task to be
3933 if (*imbalance < sds->busiest_load_per_task)
3934 return fix_small_imbalance(sds, this_cpu, imbalance);
3937 /******* find_busiest_group() helpers end here *********************/
3940 * find_busiest_group - Returns the busiest group within the sched_domain
3941 * if there is an imbalance. If there isn't an imbalance, and
3942 * the user has opted for power-savings, it returns a group whose
3943 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3944 * such a group exists.
3946 * Also calculates the amount of weighted load which should be moved
3947 * to restore balance.
3949 * @sd: The sched_domain whose busiest group is to be returned.
3950 * @this_cpu: The cpu for which load balancing is currently being performed.
3951 * @imbalance: Variable which stores amount of weighted load which should
3952 * be moved to restore balance/put a group to idle.
3953 * @idle: The idle status of this_cpu.
3954 * @sd_idle: The idleness of sd
3955 * @cpus: The set of CPUs under consideration for load-balancing.
3956 * @balance: Pointer to a variable indicating if this_cpu
3957 * is the appropriate cpu to perform load balancing at this_level.
3959 * Returns: - the busiest group if imbalance exists.
3960 * - If no imbalance and user has opted for power-savings balance,
3961 * return the least loaded group whose CPUs can be
3962 * put to idle by rebalancing its tasks onto our group.
3964 static struct sched_group *
3965 find_busiest_group(struct sched_domain *sd, int this_cpu,
3966 unsigned long *imbalance, enum cpu_idle_type idle,
3967 int *sd_idle, const struct cpumask *cpus, int *balance)
3969 struct sd_lb_stats sds;
3971 memset(&sds, 0, sizeof(sds));
3974 * Compute the various statistics relavent for load balancing at
3977 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3980 /* Cases where imbalance does not exist from POV of this_cpu */
3981 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3983 * 2) There is no busy sibling group to pull from.
3984 * 3) This group is the busiest group.
3985 * 4) This group is more busy than the avg busieness at this
3987 * 5) The imbalance is within the specified limit.
3988 * 6) Any rebalance would lead to ping-pong
3990 if (balance && !(*balance))
3993 if (!sds.busiest || sds.busiest_nr_running == 0)
3996 if (sds.this_load >= sds.max_load)
3999 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4001 if (sds.this_load >= sds.avg_load)
4004 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4007 sds.busiest_load_per_task /= sds.busiest_nr_running;
4009 sds.busiest_load_per_task =
4010 min(sds.busiest_load_per_task, sds.avg_load);
4013 * We're trying to get all the cpus to the average_load, so we don't
4014 * want to push ourselves above the average load, nor do we wish to
4015 * reduce the max loaded cpu below the average load, as either of these
4016 * actions would just result in more rebalancing later, and ping-pong
4017 * tasks around. Thus we look for the minimum possible imbalance.
4018 * Negative imbalances (*we* are more loaded than anyone else) will
4019 * be counted as no imbalance for these purposes -- we can't fix that
4020 * by pulling tasks to us. Be careful of negative numbers as they'll
4021 * appear as very large values with unsigned longs.
4023 if (sds.max_load <= sds.busiest_load_per_task)
4026 /* Looks like there is an imbalance. Compute it */
4027 calculate_imbalance(&sds, this_cpu, imbalance);
4032 * There is no obvious imbalance. But check if we can do some balancing
4035 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4043 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4046 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4047 unsigned long imbalance, const struct cpumask *cpus)
4049 struct rq *busiest = NULL, *rq;
4050 unsigned long max_load = 0;
4053 for_each_cpu(i, sched_group_cpus(group)) {
4054 unsigned long power = power_of(i);
4055 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4058 if (!cpumask_test_cpu(i, cpus))
4062 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4065 if (capacity && rq->nr_running == 1 && wl > imbalance)
4068 if (wl > max_load) {
4078 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4079 * so long as it is large enough.
4081 #define MAX_PINNED_INTERVAL 512
4083 /* Working cpumask for load_balance and load_balance_newidle. */
4084 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4087 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4088 * tasks if there is an imbalance.
4090 static int load_balance(int this_cpu, struct rq *this_rq,
4091 struct sched_domain *sd, enum cpu_idle_type idle,
4094 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4095 struct sched_group *group;
4096 unsigned long imbalance;
4098 unsigned long flags;
4099 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4101 cpumask_setall(cpus);
4104 * When power savings policy is enabled for the parent domain, idle
4105 * sibling can pick up load irrespective of busy siblings. In this case,
4106 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4107 * portraying it as CPU_NOT_IDLE.
4109 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4110 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4113 schedstat_inc(sd, lb_count[idle]);
4117 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4124 schedstat_inc(sd, lb_nobusyg[idle]);
4128 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4130 schedstat_inc(sd, lb_nobusyq[idle]);
4134 BUG_ON(busiest == this_rq);
4136 schedstat_add(sd, lb_imbalance[idle], imbalance);
4139 if (busiest->nr_running > 1) {
4141 * Attempt to move tasks. If find_busiest_group has found
4142 * an imbalance but busiest->nr_running <= 1, the group is
4143 * still unbalanced. ld_moved simply stays zero, so it is
4144 * correctly treated as an imbalance.
4146 local_irq_save(flags);
4147 double_rq_lock(this_rq, busiest);
4148 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4149 imbalance, sd, idle, &all_pinned);
4150 double_rq_unlock(this_rq, busiest);
4151 local_irq_restore(flags);
4154 * some other cpu did the load balance for us.
4156 if (ld_moved && this_cpu != smp_processor_id())
4157 resched_cpu(this_cpu);
4159 /* All tasks on this runqueue were pinned by CPU affinity */
4160 if (unlikely(all_pinned)) {
4161 cpumask_clear_cpu(cpu_of(busiest), cpus);
4162 if (!cpumask_empty(cpus))
4169 schedstat_inc(sd, lb_failed[idle]);
4170 sd->nr_balance_failed++;
4172 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4174 spin_lock_irqsave(&busiest->lock, flags);
4176 /* don't kick the migration_thread, if the curr
4177 * task on busiest cpu can't be moved to this_cpu
4179 if (!cpumask_test_cpu(this_cpu,
4180 &busiest->curr->cpus_allowed)) {
4181 spin_unlock_irqrestore(&busiest->lock, flags);
4183 goto out_one_pinned;
4186 if (!busiest->active_balance) {
4187 busiest->active_balance = 1;
4188 busiest->push_cpu = this_cpu;
4191 spin_unlock_irqrestore(&busiest->lock, flags);
4193 wake_up_process(busiest->migration_thread);
4196 * We've kicked active balancing, reset the failure
4199 sd->nr_balance_failed = sd->cache_nice_tries+1;
4202 sd->nr_balance_failed = 0;
4204 if (likely(!active_balance)) {
4205 /* We were unbalanced, so reset the balancing interval */
4206 sd->balance_interval = sd->min_interval;
4209 * If we've begun active balancing, start to back off. This
4210 * case may not be covered by the all_pinned logic if there
4211 * is only 1 task on the busy runqueue (because we don't call
4214 if (sd->balance_interval < sd->max_interval)
4215 sd->balance_interval *= 2;
4218 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4219 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4225 schedstat_inc(sd, lb_balanced[idle]);
4227 sd->nr_balance_failed = 0;
4230 /* tune up the balancing interval */
4231 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4232 (sd->balance_interval < sd->max_interval))
4233 sd->balance_interval *= 2;
4235 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4236 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4247 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4248 * tasks if there is an imbalance.
4250 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4251 * this_rq is locked.
4254 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4256 struct sched_group *group;
4257 struct rq *busiest = NULL;
4258 unsigned long imbalance;
4262 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4264 cpumask_setall(cpus);
4267 * When power savings policy is enabled for the parent domain, idle
4268 * sibling can pick up load irrespective of busy siblings. In this case,
4269 * let the state of idle sibling percolate up as IDLE, instead of
4270 * portraying it as CPU_NOT_IDLE.
4272 if (sd->flags & SD_SHARE_CPUPOWER &&
4273 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4276 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4278 update_shares_locked(this_rq, sd);
4279 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4280 &sd_idle, cpus, NULL);
4282 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4286 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4288 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4292 BUG_ON(busiest == this_rq);
4294 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4297 if (busiest->nr_running > 1) {
4298 /* Attempt to move tasks */
4299 double_lock_balance(this_rq, busiest);
4300 /* this_rq->clock is already updated */
4301 update_rq_clock(busiest);
4302 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4303 imbalance, sd, CPU_NEWLY_IDLE,
4305 double_unlock_balance(this_rq, busiest);
4307 if (unlikely(all_pinned)) {
4308 cpumask_clear_cpu(cpu_of(busiest), cpus);
4309 if (!cpumask_empty(cpus))
4315 int active_balance = 0;
4317 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4318 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4319 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4322 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4325 if (sd->nr_balance_failed++ < 2)
4329 * The only task running in a non-idle cpu can be moved to this
4330 * cpu in an attempt to completely freeup the other CPU
4331 * package. The same method used to move task in load_balance()
4332 * have been extended for load_balance_newidle() to speedup
4333 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4335 * The package power saving logic comes from
4336 * find_busiest_group(). If there are no imbalance, then
4337 * f_b_g() will return NULL. However when sched_mc={1,2} then
4338 * f_b_g() will select a group from which a running task may be
4339 * pulled to this cpu in order to make the other package idle.
4340 * If there is no opportunity to make a package idle and if
4341 * there are no imbalance, then f_b_g() will return NULL and no
4342 * action will be taken in load_balance_newidle().
4344 * Under normal task pull operation due to imbalance, there
4345 * will be more than one task in the source run queue and
4346 * move_tasks() will succeed. ld_moved will be true and this
4347 * active balance code will not be triggered.
4350 /* Lock busiest in correct order while this_rq is held */
4351 double_lock_balance(this_rq, busiest);
4354 * don't kick the migration_thread, if the curr
4355 * task on busiest cpu can't be moved to this_cpu
4357 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4358 double_unlock_balance(this_rq, busiest);
4363 if (!busiest->active_balance) {
4364 busiest->active_balance = 1;
4365 busiest->push_cpu = this_cpu;
4369 double_unlock_balance(this_rq, busiest);
4371 * Should not call ttwu while holding a rq->lock
4373 spin_unlock(&this_rq->lock);
4375 wake_up_process(busiest->migration_thread);
4376 spin_lock(&this_rq->lock);
4379 sd->nr_balance_failed = 0;
4381 update_shares_locked(this_rq, sd);
4385 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4386 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4387 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4389 sd->nr_balance_failed = 0;
4395 * idle_balance is called by schedule() if this_cpu is about to become
4396 * idle. Attempts to pull tasks from other CPUs.
4398 static void idle_balance(int this_cpu, struct rq *this_rq)
4400 struct sched_domain *sd;
4401 int pulled_task = 0;
4402 unsigned long next_balance = jiffies + HZ;
4404 this_rq->idle_stamp = this_rq->clock;
4406 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4409 for_each_domain(this_cpu, sd) {
4410 unsigned long interval;
4412 if (!(sd->flags & SD_LOAD_BALANCE))
4415 if (sd->flags & SD_BALANCE_NEWIDLE)
4416 /* If we've pulled tasks over stop searching: */
4417 pulled_task = load_balance_newidle(this_cpu, this_rq,
4420 interval = msecs_to_jiffies(sd->balance_interval);
4421 if (time_after(next_balance, sd->last_balance + interval))
4422 next_balance = sd->last_balance + interval;
4424 this_rq->idle_stamp = 0;
4428 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4430 * We are going idle. next_balance may be set based on
4431 * a busy processor. So reset next_balance.
4433 this_rq->next_balance = next_balance;
4438 * active_load_balance is run by migration threads. It pushes running tasks
4439 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4440 * running on each physical CPU where possible, and avoids physical /
4441 * logical imbalances.
4443 * Called with busiest_rq locked.
4445 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4447 int target_cpu = busiest_rq->push_cpu;
4448 struct sched_domain *sd;
4449 struct rq *target_rq;
4451 /* Is there any task to move? */
4452 if (busiest_rq->nr_running <= 1)
4455 target_rq = cpu_rq(target_cpu);
4458 * This condition is "impossible", if it occurs
4459 * we need to fix it. Originally reported by
4460 * Bjorn Helgaas on a 128-cpu setup.
4462 BUG_ON(busiest_rq == target_rq);
4464 /* move a task from busiest_rq to target_rq */
4465 double_lock_balance(busiest_rq, target_rq);
4466 update_rq_clock(busiest_rq);
4467 update_rq_clock(target_rq);
4469 /* Search for an sd spanning us and the target CPU. */
4470 for_each_domain(target_cpu, sd) {
4471 if ((sd->flags & SD_LOAD_BALANCE) &&
4472 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4477 schedstat_inc(sd, alb_count);
4479 if (move_one_task(target_rq, target_cpu, busiest_rq,
4481 schedstat_inc(sd, alb_pushed);
4483 schedstat_inc(sd, alb_failed);
4485 double_unlock_balance(busiest_rq, target_rq);
4490 atomic_t load_balancer;
4491 cpumask_var_t cpu_mask;
4492 cpumask_var_t ilb_grp_nohz_mask;
4493 } nohz ____cacheline_aligned = {
4494 .load_balancer = ATOMIC_INIT(-1),
4497 int get_nohz_load_balancer(void)
4499 return atomic_read(&nohz.load_balancer);
4502 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4504 * lowest_flag_domain - Return lowest sched_domain containing flag.
4505 * @cpu: The cpu whose lowest level of sched domain is to
4507 * @flag: The flag to check for the lowest sched_domain
4508 * for the given cpu.
4510 * Returns the lowest sched_domain of a cpu which contains the given flag.
4512 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4514 struct sched_domain *sd;
4516 for_each_domain(cpu, sd)
4517 if (sd && (sd->flags & flag))
4524 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4525 * @cpu: The cpu whose domains we're iterating over.
4526 * @sd: variable holding the value of the power_savings_sd
4528 * @flag: The flag to filter the sched_domains to be iterated.
4530 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4531 * set, starting from the lowest sched_domain to the highest.
4533 #define for_each_flag_domain(cpu, sd, flag) \
4534 for (sd = lowest_flag_domain(cpu, flag); \
4535 (sd && (sd->flags & flag)); sd = sd->parent)
4538 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4539 * @ilb_group: group to be checked for semi-idleness
4541 * Returns: 1 if the group is semi-idle. 0 otherwise.
4543 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4544 * and atleast one non-idle CPU. This helper function checks if the given
4545 * sched_group is semi-idle or not.
4547 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4549 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4550 sched_group_cpus(ilb_group));
4553 * A sched_group is semi-idle when it has atleast one busy cpu
4554 * and atleast one idle cpu.
4556 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4559 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4565 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4566 * @cpu: The cpu which is nominating a new idle_load_balancer.
4568 * Returns: Returns the id of the idle load balancer if it exists,
4569 * Else, returns >= nr_cpu_ids.
4571 * This algorithm picks the idle load balancer such that it belongs to a
4572 * semi-idle powersavings sched_domain. The idea is to try and avoid
4573 * completely idle packages/cores just for the purpose of idle load balancing
4574 * when there are other idle cpu's which are better suited for that job.
4576 static int find_new_ilb(int cpu)
4578 struct sched_domain *sd;
4579 struct sched_group *ilb_group;
4582 * Have idle load balancer selection from semi-idle packages only
4583 * when power-aware load balancing is enabled
4585 if (!(sched_smt_power_savings || sched_mc_power_savings))
4589 * Optimize for the case when we have no idle CPUs or only one
4590 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4592 if (cpumask_weight(nohz.cpu_mask) < 2)
4595 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4596 ilb_group = sd->groups;
4599 if (is_semi_idle_group(ilb_group))
4600 return cpumask_first(nohz.ilb_grp_nohz_mask);
4602 ilb_group = ilb_group->next;
4604 } while (ilb_group != sd->groups);
4608 return cpumask_first(nohz.cpu_mask);
4610 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4611 static inline int find_new_ilb(int call_cpu)
4613 return cpumask_first(nohz.cpu_mask);
4618 * This routine will try to nominate the ilb (idle load balancing)
4619 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4620 * load balancing on behalf of all those cpus. If all the cpus in the system
4621 * go into this tickless mode, then there will be no ilb owner (as there is
4622 * no need for one) and all the cpus will sleep till the next wakeup event
4625 * For the ilb owner, tick is not stopped. And this tick will be used
4626 * for idle load balancing. ilb owner will still be part of
4629 * While stopping the tick, this cpu will become the ilb owner if there
4630 * is no other owner. And will be the owner till that cpu becomes busy
4631 * or if all cpus in the system stop their ticks at which point
4632 * there is no need for ilb owner.
4634 * When the ilb owner becomes busy, it nominates another owner, during the
4635 * next busy scheduler_tick()
4637 int select_nohz_load_balancer(int stop_tick)
4639 int cpu = smp_processor_id();
4642 cpu_rq(cpu)->in_nohz_recently = 1;
4644 if (!cpu_active(cpu)) {
4645 if (atomic_read(&nohz.load_balancer) != cpu)
4649 * If we are going offline and still the leader,
4652 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4658 cpumask_set_cpu(cpu, nohz.cpu_mask);
4660 /* time for ilb owner also to sleep */
4661 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4662 if (atomic_read(&nohz.load_balancer) == cpu)
4663 atomic_set(&nohz.load_balancer, -1);
4667 if (atomic_read(&nohz.load_balancer) == -1) {
4668 /* make me the ilb owner */
4669 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4671 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4674 if (!(sched_smt_power_savings ||
4675 sched_mc_power_savings))
4678 * Check to see if there is a more power-efficient
4681 new_ilb = find_new_ilb(cpu);
4682 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4683 atomic_set(&nohz.load_balancer, -1);
4684 resched_cpu(new_ilb);
4690 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4693 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4695 if (atomic_read(&nohz.load_balancer) == cpu)
4696 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4703 static DEFINE_SPINLOCK(balancing);
4706 * It checks each scheduling domain to see if it is due to be balanced,
4707 * and initiates a balancing operation if so.
4709 * Balancing parameters are set up in arch_init_sched_domains.
4711 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4714 struct rq *rq = cpu_rq(cpu);
4715 unsigned long interval;
4716 struct sched_domain *sd;
4717 /* Earliest time when we have to do rebalance again */
4718 unsigned long next_balance = jiffies + 60*HZ;
4719 int update_next_balance = 0;
4722 for_each_domain(cpu, sd) {
4723 if (!(sd->flags & SD_LOAD_BALANCE))
4726 interval = sd->balance_interval;
4727 if (idle != CPU_IDLE)
4728 interval *= sd->busy_factor;
4730 /* scale ms to jiffies */
4731 interval = msecs_to_jiffies(interval);
4732 if (unlikely(!interval))
4734 if (interval > HZ*NR_CPUS/10)
4735 interval = HZ*NR_CPUS/10;
4737 need_serialize = sd->flags & SD_SERIALIZE;
4739 if (need_serialize) {
4740 if (!spin_trylock(&balancing))
4744 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4745 if (load_balance(cpu, rq, sd, idle, &balance)) {
4747 * We've pulled tasks over so either we're no
4748 * longer idle, or one of our SMT siblings is
4751 idle = CPU_NOT_IDLE;
4753 sd->last_balance = jiffies;
4756 spin_unlock(&balancing);
4758 if (time_after(next_balance, sd->last_balance + interval)) {
4759 next_balance = sd->last_balance + interval;
4760 update_next_balance = 1;
4764 * Stop the load balance at this level. There is another
4765 * CPU in our sched group which is doing load balancing more
4773 * next_balance will be updated only when there is a need.
4774 * When the cpu is attached to null domain for ex, it will not be
4777 if (likely(update_next_balance))
4778 rq->next_balance = next_balance;
4782 * run_rebalance_domains is triggered when needed from the scheduler tick.
4783 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4784 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4786 static void run_rebalance_domains(struct softirq_action *h)
4788 int this_cpu = smp_processor_id();
4789 struct rq *this_rq = cpu_rq(this_cpu);
4790 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4791 CPU_IDLE : CPU_NOT_IDLE;
4793 rebalance_domains(this_cpu, idle);
4797 * If this cpu is the owner for idle load balancing, then do the
4798 * balancing on behalf of the other idle cpus whose ticks are
4801 if (this_rq->idle_at_tick &&
4802 atomic_read(&nohz.load_balancer) == this_cpu) {
4806 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4807 if (balance_cpu == this_cpu)
4811 * If this cpu gets work to do, stop the load balancing
4812 * work being done for other cpus. Next load
4813 * balancing owner will pick it up.
4818 rebalance_domains(balance_cpu, CPU_IDLE);
4820 rq = cpu_rq(balance_cpu);
4821 if (time_after(this_rq->next_balance, rq->next_balance))
4822 this_rq->next_balance = rq->next_balance;
4828 static inline int on_null_domain(int cpu)
4830 return !rcu_dereference(cpu_rq(cpu)->sd);
4834 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4836 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4837 * idle load balancing owner or decide to stop the periodic load balancing,
4838 * if the whole system is idle.
4840 static inline void trigger_load_balance(struct rq *rq, int cpu)
4844 * If we were in the nohz mode recently and busy at the current
4845 * scheduler tick, then check if we need to nominate new idle
4848 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4849 rq->in_nohz_recently = 0;
4851 if (atomic_read(&nohz.load_balancer) == cpu) {
4852 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4853 atomic_set(&nohz.load_balancer, -1);
4856 if (atomic_read(&nohz.load_balancer) == -1) {
4857 int ilb = find_new_ilb(cpu);
4859 if (ilb < nr_cpu_ids)
4865 * If this cpu is idle and doing idle load balancing for all the
4866 * cpus with ticks stopped, is it time for that to stop?
4868 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4869 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4875 * If this cpu is idle and the idle load balancing is done by
4876 * someone else, then no need raise the SCHED_SOFTIRQ
4878 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4879 cpumask_test_cpu(cpu, nohz.cpu_mask))
4882 /* Don't need to rebalance while attached to NULL domain */
4883 if (time_after_eq(jiffies, rq->next_balance) &&
4884 likely(!on_null_domain(cpu)))
4885 raise_softirq(SCHED_SOFTIRQ);
4888 #else /* CONFIG_SMP */
4891 * on UP we do not need to balance between CPUs:
4893 static inline void idle_balance(int cpu, struct rq *rq)
4899 DEFINE_PER_CPU(struct kernel_stat, kstat);
4901 EXPORT_PER_CPU_SYMBOL(kstat);
4904 * Return any ns on the sched_clock that have not yet been accounted in
4905 * @p in case that task is currently running.
4907 * Called with task_rq_lock() held on @rq.
4909 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4913 if (task_current(rq, p)) {
4914 update_rq_clock(rq);
4915 ns = rq->clock - p->se.exec_start;
4923 unsigned long long task_delta_exec(struct task_struct *p)
4925 unsigned long flags;
4929 rq = task_rq_lock(p, &flags);
4930 ns = do_task_delta_exec(p, rq);
4931 task_rq_unlock(rq, &flags);
4937 * Return accounted runtime for the task.
4938 * In case the task is currently running, return the runtime plus current's
4939 * pending runtime that have not been accounted yet.
4941 unsigned long long task_sched_runtime(struct task_struct *p)
4943 unsigned long flags;
4947 rq = task_rq_lock(p, &flags);
4948 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4949 task_rq_unlock(rq, &flags);
4955 * Return sum_exec_runtime for the thread group.
4956 * In case the task is currently running, return the sum plus current's
4957 * pending runtime that have not been accounted yet.
4959 * Note that the thread group might have other running tasks as well,
4960 * so the return value not includes other pending runtime that other
4961 * running tasks might have.
4963 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4965 struct task_cputime totals;
4966 unsigned long flags;
4970 rq = task_rq_lock(p, &flags);
4971 thread_group_cputime(p, &totals);
4972 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4973 task_rq_unlock(rq, &flags);
4979 * Account user cpu time to a process.
4980 * @p: the process that the cpu time gets accounted to
4981 * @cputime: the cpu time spent in user space since the last update
4982 * @cputime_scaled: cputime scaled by cpu frequency
4984 void account_user_time(struct task_struct *p, cputime_t cputime,
4985 cputime_t cputime_scaled)
4987 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4990 /* Add user time to process. */
4991 p->utime = cputime_add(p->utime, cputime);
4992 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4993 account_group_user_time(p, cputime);
4995 /* Add user time to cpustat. */
4996 tmp = cputime_to_cputime64(cputime);
4997 if (TASK_NICE(p) > 0)
4998 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5000 cpustat->user = cputime64_add(cpustat->user, tmp);
5002 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5003 /* Account for user time used */
5004 acct_update_integrals(p);
5008 * Account guest cpu time to a process.
5009 * @p: the process that the cpu time gets accounted to
5010 * @cputime: the cpu time spent in virtual machine since the last update
5011 * @cputime_scaled: cputime scaled by cpu frequency
5013 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5014 cputime_t cputime_scaled)
5017 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5019 tmp = cputime_to_cputime64(cputime);
5021 /* Add guest time to process. */
5022 p->utime = cputime_add(p->utime, cputime);
5023 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5024 account_group_user_time(p, cputime);
5025 p->gtime = cputime_add(p->gtime, cputime);
5027 /* Add guest time to cpustat. */
5028 if (TASK_NICE(p) > 0) {
5029 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5030 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
5032 cpustat->user = cputime64_add(cpustat->user, tmp);
5033 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5038 * Account system cpu time to a process.
5039 * @p: the process that the cpu time gets accounted to
5040 * @hardirq_offset: the offset to subtract from hardirq_count()
5041 * @cputime: the cpu time spent in kernel space since the last update
5042 * @cputime_scaled: cputime scaled by cpu frequency
5044 void account_system_time(struct task_struct *p, int hardirq_offset,
5045 cputime_t cputime, cputime_t cputime_scaled)
5047 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5050 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5051 account_guest_time(p, cputime, cputime_scaled);
5055 /* Add system time to process. */
5056 p->stime = cputime_add(p->stime, cputime);
5057 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5058 account_group_system_time(p, cputime);
5060 /* Add system time to cpustat. */
5061 tmp = cputime_to_cputime64(cputime);
5062 if (hardirq_count() - hardirq_offset)
5063 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5064 else if (softirq_count())
5065 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5067 cpustat->system = cputime64_add(cpustat->system, tmp);
5069 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5071 /* Account for system time used */
5072 acct_update_integrals(p);
5076 * Account for involuntary wait time.
5077 * @steal: the cpu time spent in involuntary wait
5079 void account_steal_time(cputime_t cputime)
5081 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5082 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5084 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5088 * Account for idle time.
5089 * @cputime: the cpu time spent in idle wait
5091 void account_idle_time(cputime_t cputime)
5093 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5094 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5095 struct rq *rq = this_rq();
5097 if (atomic_read(&rq->nr_iowait) > 0)
5098 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5100 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5103 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5106 * Account a single tick of cpu time.
5107 * @p: the process that the cpu time gets accounted to
5108 * @user_tick: indicates if the tick is a user or a system tick
5110 void account_process_tick(struct task_struct *p, int user_tick)
5112 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5113 struct rq *rq = this_rq();
5116 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5117 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5118 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5121 account_idle_time(cputime_one_jiffy);
5125 * Account multiple ticks of steal time.
5126 * @p: the process from which the cpu time has been stolen
5127 * @ticks: number of stolen ticks
5129 void account_steal_ticks(unsigned long ticks)
5131 account_steal_time(jiffies_to_cputime(ticks));
5135 * Account multiple ticks of idle time.
5136 * @ticks: number of stolen ticks
5138 void account_idle_ticks(unsigned long ticks)
5140 account_idle_time(jiffies_to_cputime(ticks));
5146 * Use precise platform statistics if available:
5148 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5149 cputime_t task_utime(struct task_struct *p)
5154 cputime_t task_stime(struct task_struct *p)
5159 cputime_t task_utime(struct task_struct *p)
5161 clock_t utime = cputime_to_clock_t(p->utime),
5162 total = utime + cputime_to_clock_t(p->stime);
5166 * Use CFS's precise accounting:
5168 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5172 do_div(temp, total);
5174 utime = (clock_t)temp;
5176 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5177 return p->prev_utime;
5180 cputime_t task_stime(struct task_struct *p)
5185 * Use CFS's precise accounting. (we subtract utime from
5186 * the total, to make sure the total observed by userspace
5187 * grows monotonically - apps rely on that):
5189 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5190 cputime_to_clock_t(task_utime(p));
5193 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5195 return p->prev_stime;
5199 inline cputime_t task_gtime(struct task_struct *p)
5205 * This function gets called by the timer code, with HZ frequency.
5206 * We call it with interrupts disabled.
5208 * It also gets called by the fork code, when changing the parent's
5211 void scheduler_tick(void)
5213 int cpu = smp_processor_id();
5214 struct rq *rq = cpu_rq(cpu);
5215 struct task_struct *curr = rq->curr;
5219 spin_lock(&rq->lock);
5220 update_rq_clock(rq);
5221 update_cpu_load(rq);
5222 curr->sched_class->task_tick(rq, curr, 0);
5223 spin_unlock(&rq->lock);
5225 perf_event_task_tick(curr, cpu);
5228 rq->idle_at_tick = idle_cpu(cpu);
5229 trigger_load_balance(rq, cpu);
5233 notrace unsigned long get_parent_ip(unsigned long addr)
5235 if (in_lock_functions(addr)) {
5236 addr = CALLER_ADDR2;
5237 if (in_lock_functions(addr))
5238 addr = CALLER_ADDR3;
5243 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5244 defined(CONFIG_PREEMPT_TRACER))
5246 void __kprobes add_preempt_count(int val)
5248 #ifdef CONFIG_DEBUG_PREEMPT
5252 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5255 preempt_count() += val;
5256 #ifdef CONFIG_DEBUG_PREEMPT
5258 * Spinlock count overflowing soon?
5260 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5263 if (preempt_count() == val)
5264 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5266 EXPORT_SYMBOL(add_preempt_count);
5268 void __kprobes sub_preempt_count(int val)
5270 #ifdef CONFIG_DEBUG_PREEMPT
5274 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5277 * Is the spinlock portion underflowing?
5279 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5280 !(preempt_count() & PREEMPT_MASK)))
5284 if (preempt_count() == val)
5285 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5286 preempt_count() -= val;
5288 EXPORT_SYMBOL(sub_preempt_count);
5293 * Print scheduling while atomic bug:
5295 static noinline void __schedule_bug(struct task_struct *prev)
5297 struct pt_regs *regs = get_irq_regs();
5299 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5300 prev->comm, prev->pid, preempt_count());
5302 debug_show_held_locks(prev);
5304 if (irqs_disabled())
5305 print_irqtrace_events(prev);
5314 * Various schedule()-time debugging checks and statistics:
5316 static inline void schedule_debug(struct task_struct *prev)
5319 * Test if we are atomic. Since do_exit() needs to call into
5320 * schedule() atomically, we ignore that path for now.
5321 * Otherwise, whine if we are scheduling when we should not be.
5323 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5324 __schedule_bug(prev);
5326 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5328 schedstat_inc(this_rq(), sched_count);
5329 #ifdef CONFIG_SCHEDSTATS
5330 if (unlikely(prev->lock_depth >= 0)) {
5331 schedstat_inc(this_rq(), bkl_count);
5332 schedstat_inc(prev, sched_info.bkl_count);
5337 static void put_prev_task(struct rq *rq, struct task_struct *p)
5339 u64 runtime = p->se.sum_exec_runtime - p->se.prev_sum_exec_runtime;
5341 update_avg(&p->se.avg_running, runtime);
5343 if (p->state == TASK_RUNNING) {
5345 * In order to avoid avg_overlap growing stale when we are
5346 * indeed overlapping and hence not getting put to sleep, grow
5347 * the avg_overlap on preemption.
5349 * We use the average preemption runtime because that
5350 * correlates to the amount of cache footprint a task can
5353 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5354 update_avg(&p->se.avg_overlap, runtime);
5356 update_avg(&p->se.avg_running, 0);
5358 p->sched_class->put_prev_task(rq, p);
5362 * Pick up the highest-prio task:
5364 static inline struct task_struct *
5365 pick_next_task(struct rq *rq)
5367 const struct sched_class *class;
5368 struct task_struct *p;
5371 * Optimization: we know that if all tasks are in
5372 * the fair class we can call that function directly:
5374 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5375 p = fair_sched_class.pick_next_task(rq);
5380 class = sched_class_highest;
5382 p = class->pick_next_task(rq);
5386 * Will never be NULL as the idle class always
5387 * returns a non-NULL p:
5389 class = class->next;
5394 * schedule() is the main scheduler function.
5396 asmlinkage void __sched schedule(void)
5398 struct task_struct *prev, *next;
5399 unsigned long *switch_count;
5405 cpu = smp_processor_id();
5409 switch_count = &prev->nivcsw;
5411 release_kernel_lock(prev);
5412 need_resched_nonpreemptible:
5414 schedule_debug(prev);
5416 if (sched_feat(HRTICK))
5419 spin_lock_irq(&rq->lock);
5420 update_rq_clock(rq);
5421 clear_tsk_need_resched(prev);
5423 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5424 if (unlikely(signal_pending_state(prev->state, prev)))
5425 prev->state = TASK_RUNNING;
5427 deactivate_task(rq, prev, 1);
5428 switch_count = &prev->nvcsw;
5431 pre_schedule(rq, prev);
5433 if (unlikely(!rq->nr_running))
5434 idle_balance(cpu, rq);
5436 put_prev_task(rq, prev);
5437 next = pick_next_task(rq);
5439 if (likely(prev != next)) {
5440 sched_info_switch(prev, next);
5441 perf_event_task_sched_out(prev, next, cpu);
5447 context_switch(rq, prev, next); /* unlocks the rq */
5449 * the context switch might have flipped the stack from under
5450 * us, hence refresh the local variables.
5452 cpu = smp_processor_id();
5455 spin_unlock_irq(&rq->lock);
5459 if (unlikely(reacquire_kernel_lock(current) < 0))
5460 goto need_resched_nonpreemptible;
5462 preempt_enable_no_resched();
5466 EXPORT_SYMBOL(schedule);
5470 * Look out! "owner" is an entirely speculative pointer
5471 * access and not reliable.
5473 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5478 if (!sched_feat(OWNER_SPIN))
5481 #ifdef CONFIG_DEBUG_PAGEALLOC
5483 * Need to access the cpu field knowing that
5484 * DEBUG_PAGEALLOC could have unmapped it if
5485 * the mutex owner just released it and exited.
5487 if (probe_kernel_address(&owner->cpu, cpu))
5494 * Even if the access succeeded (likely case),
5495 * the cpu field may no longer be valid.
5497 if (cpu >= nr_cpumask_bits)
5501 * We need to validate that we can do a
5502 * get_cpu() and that we have the percpu area.
5504 if (!cpu_online(cpu))
5511 * Owner changed, break to re-assess state.
5513 if (lock->owner != owner)
5517 * Is that owner really running on that cpu?
5519 if (task_thread_info(rq->curr) != owner || need_resched())
5529 #ifdef CONFIG_PREEMPT
5531 * this is the entry point to schedule() from in-kernel preemption
5532 * off of preempt_enable. Kernel preemptions off return from interrupt
5533 * occur there and call schedule directly.
5535 asmlinkage void __sched preempt_schedule(void)
5537 struct thread_info *ti = current_thread_info();
5540 * If there is a non-zero preempt_count or interrupts are disabled,
5541 * we do not want to preempt the current task. Just return..
5543 if (likely(ti->preempt_count || irqs_disabled()))
5547 add_preempt_count(PREEMPT_ACTIVE);
5549 sub_preempt_count(PREEMPT_ACTIVE);
5552 * Check again in case we missed a preemption opportunity
5553 * between schedule and now.
5556 } while (need_resched());
5558 EXPORT_SYMBOL(preempt_schedule);
5561 * this is the entry point to schedule() from kernel preemption
5562 * off of irq context.
5563 * Note, that this is called and return with irqs disabled. This will
5564 * protect us against recursive calling from irq.
5566 asmlinkage void __sched preempt_schedule_irq(void)
5568 struct thread_info *ti = current_thread_info();
5570 /* Catch callers which need to be fixed */
5571 BUG_ON(ti->preempt_count || !irqs_disabled());
5574 add_preempt_count(PREEMPT_ACTIVE);
5577 local_irq_disable();
5578 sub_preempt_count(PREEMPT_ACTIVE);
5581 * Check again in case we missed a preemption opportunity
5582 * between schedule and now.
5585 } while (need_resched());
5588 #endif /* CONFIG_PREEMPT */
5590 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5593 return try_to_wake_up(curr->private, mode, wake_flags);
5595 EXPORT_SYMBOL(default_wake_function);
5598 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5599 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5600 * number) then we wake all the non-exclusive tasks and one exclusive task.
5602 * There are circumstances in which we can try to wake a task which has already
5603 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5604 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5606 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5607 int nr_exclusive, int wake_flags, void *key)
5609 wait_queue_t *curr, *next;
5611 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5612 unsigned flags = curr->flags;
5614 if (curr->func(curr, mode, wake_flags, key) &&
5615 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5621 * __wake_up - wake up threads blocked on a waitqueue.
5623 * @mode: which threads
5624 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5625 * @key: is directly passed to the wakeup function
5627 * It may be assumed that this function implies a write memory barrier before
5628 * changing the task state if and only if any tasks are woken up.
5630 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5631 int nr_exclusive, void *key)
5633 unsigned long flags;
5635 spin_lock_irqsave(&q->lock, flags);
5636 __wake_up_common(q, mode, nr_exclusive, 0, key);
5637 spin_unlock_irqrestore(&q->lock, flags);
5639 EXPORT_SYMBOL(__wake_up);
5642 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5644 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5646 __wake_up_common(q, mode, 1, 0, NULL);
5649 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5651 __wake_up_common(q, mode, 1, 0, key);
5655 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5657 * @mode: which threads
5658 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5659 * @key: opaque value to be passed to wakeup targets
5661 * The sync wakeup differs that the waker knows that it will schedule
5662 * away soon, so while the target thread will be woken up, it will not
5663 * be migrated to another CPU - ie. the two threads are 'synchronized'
5664 * with each other. This can prevent needless bouncing between CPUs.
5666 * On UP it can prevent extra preemption.
5668 * It may be assumed that this function implies a write memory barrier before
5669 * changing the task state if and only if any tasks are woken up.
5671 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5672 int nr_exclusive, void *key)
5674 unsigned long flags;
5675 int wake_flags = WF_SYNC;
5680 if (unlikely(!nr_exclusive))
5683 spin_lock_irqsave(&q->lock, flags);
5684 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5685 spin_unlock_irqrestore(&q->lock, flags);
5687 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5690 * __wake_up_sync - see __wake_up_sync_key()
5692 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5694 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5696 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5699 * complete: - signals a single thread waiting on this completion
5700 * @x: holds the state of this particular completion
5702 * This will wake up a single thread waiting on this completion. Threads will be
5703 * awakened in the same order in which they were queued.
5705 * See also complete_all(), wait_for_completion() and related routines.
5707 * It may be assumed that this function implies a write memory barrier before
5708 * changing the task state if and only if any tasks are woken up.
5710 void complete(struct completion *x)
5712 unsigned long flags;
5714 spin_lock_irqsave(&x->wait.lock, flags);
5716 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5717 spin_unlock_irqrestore(&x->wait.lock, flags);
5719 EXPORT_SYMBOL(complete);
5722 * complete_all: - signals all threads waiting on this completion
5723 * @x: holds the state of this particular completion
5725 * This will wake up all threads waiting on this particular completion event.
5727 * It may be assumed that this function implies a write memory barrier before
5728 * changing the task state if and only if any tasks are woken up.
5730 void complete_all(struct completion *x)
5732 unsigned long flags;
5734 spin_lock_irqsave(&x->wait.lock, flags);
5735 x->done += UINT_MAX/2;
5736 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5737 spin_unlock_irqrestore(&x->wait.lock, flags);
5739 EXPORT_SYMBOL(complete_all);
5741 static inline long __sched
5742 do_wait_for_common(struct completion *x, long timeout, int state)
5745 DECLARE_WAITQUEUE(wait, current);
5747 wait.flags |= WQ_FLAG_EXCLUSIVE;
5748 __add_wait_queue_tail(&x->wait, &wait);
5750 if (signal_pending_state(state, current)) {
5751 timeout = -ERESTARTSYS;
5754 __set_current_state(state);
5755 spin_unlock_irq(&x->wait.lock);
5756 timeout = schedule_timeout(timeout);
5757 spin_lock_irq(&x->wait.lock);
5758 } while (!x->done && timeout);
5759 __remove_wait_queue(&x->wait, &wait);
5764 return timeout ?: 1;
5768 wait_for_common(struct completion *x, long timeout, int state)
5772 spin_lock_irq(&x->wait.lock);
5773 timeout = do_wait_for_common(x, timeout, state);
5774 spin_unlock_irq(&x->wait.lock);
5779 * wait_for_completion: - waits for completion of a task
5780 * @x: holds the state of this particular completion
5782 * This waits to be signaled for completion of a specific task. It is NOT
5783 * interruptible and there is no timeout.
5785 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5786 * and interrupt capability. Also see complete().
5788 void __sched wait_for_completion(struct completion *x)
5790 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5792 EXPORT_SYMBOL(wait_for_completion);
5795 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5796 * @x: holds the state of this particular completion
5797 * @timeout: timeout value in jiffies
5799 * This waits for either a completion of a specific task to be signaled or for a
5800 * specified timeout to expire. The timeout is in jiffies. It is not
5803 unsigned long __sched
5804 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5806 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5808 EXPORT_SYMBOL(wait_for_completion_timeout);
5811 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5812 * @x: holds the state of this particular completion
5814 * This waits for completion of a specific task to be signaled. It is
5817 int __sched wait_for_completion_interruptible(struct completion *x)
5819 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5820 if (t == -ERESTARTSYS)
5824 EXPORT_SYMBOL(wait_for_completion_interruptible);
5827 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5828 * @x: holds the state of this particular completion
5829 * @timeout: timeout value in jiffies
5831 * This waits for either a completion of a specific task to be signaled or for a
5832 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5834 unsigned long __sched
5835 wait_for_completion_interruptible_timeout(struct completion *x,
5836 unsigned long timeout)
5838 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5840 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5843 * wait_for_completion_killable: - waits for completion of a task (killable)
5844 * @x: holds the state of this particular completion
5846 * This waits to be signaled for completion of a specific task. It can be
5847 * interrupted by a kill signal.
5849 int __sched wait_for_completion_killable(struct completion *x)
5851 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5852 if (t == -ERESTARTSYS)
5856 EXPORT_SYMBOL(wait_for_completion_killable);
5859 * try_wait_for_completion - try to decrement a completion without blocking
5860 * @x: completion structure
5862 * Returns: 0 if a decrement cannot be done without blocking
5863 * 1 if a decrement succeeded.
5865 * If a completion is being used as a counting completion,
5866 * attempt to decrement the counter without blocking. This
5867 * enables us to avoid waiting if the resource the completion
5868 * is protecting is not available.
5870 bool try_wait_for_completion(struct completion *x)
5874 spin_lock_irq(&x->wait.lock);
5879 spin_unlock_irq(&x->wait.lock);
5882 EXPORT_SYMBOL(try_wait_for_completion);
5885 * completion_done - Test to see if a completion has any waiters
5886 * @x: completion structure
5888 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5889 * 1 if there are no waiters.
5892 bool completion_done(struct completion *x)
5896 spin_lock_irq(&x->wait.lock);
5899 spin_unlock_irq(&x->wait.lock);
5902 EXPORT_SYMBOL(completion_done);
5905 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5907 unsigned long flags;
5910 init_waitqueue_entry(&wait, current);
5912 __set_current_state(state);
5914 spin_lock_irqsave(&q->lock, flags);
5915 __add_wait_queue(q, &wait);
5916 spin_unlock(&q->lock);
5917 timeout = schedule_timeout(timeout);
5918 spin_lock_irq(&q->lock);
5919 __remove_wait_queue(q, &wait);
5920 spin_unlock_irqrestore(&q->lock, flags);
5925 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5927 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5929 EXPORT_SYMBOL(interruptible_sleep_on);
5932 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5934 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5936 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5938 void __sched sleep_on(wait_queue_head_t *q)
5940 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5942 EXPORT_SYMBOL(sleep_on);
5944 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5946 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5948 EXPORT_SYMBOL(sleep_on_timeout);
5950 #ifdef CONFIG_RT_MUTEXES
5953 * rt_mutex_setprio - set the current priority of a task
5955 * @prio: prio value (kernel-internal form)
5957 * This function changes the 'effective' priority of a task. It does
5958 * not touch ->normal_prio like __setscheduler().
5960 * Used by the rt_mutex code to implement priority inheritance logic.
5962 void rt_mutex_setprio(struct task_struct *p, int prio)
5964 unsigned long flags;
5965 int oldprio, on_rq, running;
5967 const struct sched_class *prev_class = p->sched_class;
5969 BUG_ON(prio < 0 || prio > MAX_PRIO);
5971 rq = task_rq_lock(p, &flags);
5972 update_rq_clock(rq);
5975 on_rq = p->se.on_rq;
5976 running = task_current(rq, p);
5978 dequeue_task(rq, p, 0);
5980 p->sched_class->put_prev_task(rq, p);
5983 p->sched_class = &rt_sched_class;
5985 p->sched_class = &fair_sched_class;
5990 p->sched_class->set_curr_task(rq);
5992 enqueue_task(rq, p, 0);
5994 check_class_changed(rq, p, prev_class, oldprio, running);
5996 task_rq_unlock(rq, &flags);
6001 void set_user_nice(struct task_struct *p, long nice)
6003 int old_prio, delta, on_rq;
6004 unsigned long flags;
6007 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6010 * We have to be careful, if called from sys_setpriority(),
6011 * the task might be in the middle of scheduling on another CPU.
6013 rq = task_rq_lock(p, &flags);
6014 update_rq_clock(rq);
6016 * The RT priorities are set via sched_setscheduler(), but we still
6017 * allow the 'normal' nice value to be set - but as expected
6018 * it wont have any effect on scheduling until the task is
6019 * SCHED_FIFO/SCHED_RR:
6021 if (task_has_rt_policy(p)) {
6022 p->static_prio = NICE_TO_PRIO(nice);
6025 on_rq = p->se.on_rq;
6027 dequeue_task(rq, p, 0);
6029 p->static_prio = NICE_TO_PRIO(nice);
6032 p->prio = effective_prio(p);
6033 delta = p->prio - old_prio;
6036 enqueue_task(rq, p, 0);
6038 * If the task increased its priority or is running and
6039 * lowered its priority, then reschedule its CPU:
6041 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6042 resched_task(rq->curr);
6045 task_rq_unlock(rq, &flags);
6047 EXPORT_SYMBOL(set_user_nice);
6050 * can_nice - check if a task can reduce its nice value
6054 int can_nice(const struct task_struct *p, const int nice)
6056 /* convert nice value [19,-20] to rlimit style value [1,40] */
6057 int nice_rlim = 20 - nice;
6059 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6060 capable(CAP_SYS_NICE));
6063 #ifdef __ARCH_WANT_SYS_NICE
6066 * sys_nice - change the priority of the current process.
6067 * @increment: priority increment
6069 * sys_setpriority is a more generic, but much slower function that
6070 * does similar things.
6072 SYSCALL_DEFINE1(nice, int, increment)
6077 * Setpriority might change our priority at the same moment.
6078 * We don't have to worry. Conceptually one call occurs first
6079 * and we have a single winner.
6081 if (increment < -40)
6086 nice = TASK_NICE(current) + increment;
6092 if (increment < 0 && !can_nice(current, nice))
6095 retval = security_task_setnice(current, nice);
6099 set_user_nice(current, nice);
6106 * task_prio - return the priority value of a given task.
6107 * @p: the task in question.
6109 * This is the priority value as seen by users in /proc.
6110 * RT tasks are offset by -200. Normal tasks are centered
6111 * around 0, value goes from -16 to +15.
6113 int task_prio(const struct task_struct *p)
6115 return p->prio - MAX_RT_PRIO;
6119 * task_nice - return the nice value of a given task.
6120 * @p: the task in question.
6122 int task_nice(const struct task_struct *p)
6124 return TASK_NICE(p);
6126 EXPORT_SYMBOL(task_nice);
6129 * idle_cpu - is a given cpu idle currently?
6130 * @cpu: the processor in question.
6132 int idle_cpu(int cpu)
6134 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6138 * idle_task - return the idle task for a given cpu.
6139 * @cpu: the processor in question.
6141 struct task_struct *idle_task(int cpu)
6143 return cpu_rq(cpu)->idle;
6147 * find_process_by_pid - find a process with a matching PID value.
6148 * @pid: the pid in question.
6150 static struct task_struct *find_process_by_pid(pid_t pid)
6152 return pid ? find_task_by_vpid(pid) : current;
6155 /* Actually do priority change: must hold rq lock. */
6157 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6159 BUG_ON(p->se.on_rq);
6162 switch (p->policy) {
6166 p->sched_class = &fair_sched_class;
6170 p->sched_class = &rt_sched_class;
6174 p->rt_priority = prio;
6175 p->normal_prio = normal_prio(p);
6176 /* we are holding p->pi_lock already */
6177 p->prio = rt_mutex_getprio(p);
6182 * check the target process has a UID that matches the current process's
6184 static bool check_same_owner(struct task_struct *p)
6186 const struct cred *cred = current_cred(), *pcred;
6190 pcred = __task_cred(p);
6191 match = (cred->euid == pcred->euid ||
6192 cred->euid == pcred->uid);
6197 static int __sched_setscheduler(struct task_struct *p, int policy,
6198 struct sched_param *param, bool user)
6200 int retval, oldprio, oldpolicy = -1, on_rq, running;
6201 unsigned long flags;
6202 const struct sched_class *prev_class = p->sched_class;
6206 /* may grab non-irq protected spin_locks */
6207 BUG_ON(in_interrupt());
6209 /* double check policy once rq lock held */
6211 reset_on_fork = p->sched_reset_on_fork;
6212 policy = oldpolicy = p->policy;
6214 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6215 policy &= ~SCHED_RESET_ON_FORK;
6217 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6218 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6219 policy != SCHED_IDLE)
6224 * Valid priorities for SCHED_FIFO and SCHED_RR are
6225 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6226 * SCHED_BATCH and SCHED_IDLE is 0.
6228 if (param->sched_priority < 0 ||
6229 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6230 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6232 if (rt_policy(policy) != (param->sched_priority != 0))
6236 * Allow unprivileged RT tasks to decrease priority:
6238 if (user && !capable(CAP_SYS_NICE)) {
6239 if (rt_policy(policy)) {
6240 unsigned long rlim_rtprio;
6242 if (!lock_task_sighand(p, &flags))
6244 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6245 unlock_task_sighand(p, &flags);
6247 /* can't set/change the rt policy */
6248 if (policy != p->policy && !rlim_rtprio)
6251 /* can't increase priority */
6252 if (param->sched_priority > p->rt_priority &&
6253 param->sched_priority > rlim_rtprio)
6257 * Like positive nice levels, dont allow tasks to
6258 * move out of SCHED_IDLE either:
6260 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6263 /* can't change other user's priorities */
6264 if (!check_same_owner(p))
6267 /* Normal users shall not reset the sched_reset_on_fork flag */
6268 if (p->sched_reset_on_fork && !reset_on_fork)
6273 #ifdef CONFIG_RT_GROUP_SCHED
6275 * Do not allow realtime tasks into groups that have no runtime
6278 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6279 task_group(p)->rt_bandwidth.rt_runtime == 0)
6283 retval = security_task_setscheduler(p, policy, param);
6289 * make sure no PI-waiters arrive (or leave) while we are
6290 * changing the priority of the task:
6292 spin_lock_irqsave(&p->pi_lock, flags);
6294 * To be able to change p->policy safely, the apropriate
6295 * runqueue lock must be held.
6297 rq = __task_rq_lock(p);
6298 /* recheck policy now with rq lock held */
6299 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6300 policy = oldpolicy = -1;
6301 __task_rq_unlock(rq);
6302 spin_unlock_irqrestore(&p->pi_lock, flags);
6305 update_rq_clock(rq);
6306 on_rq = p->se.on_rq;
6307 running = task_current(rq, p);
6309 deactivate_task(rq, p, 0);
6311 p->sched_class->put_prev_task(rq, p);
6313 p->sched_reset_on_fork = reset_on_fork;
6316 __setscheduler(rq, p, policy, param->sched_priority);
6319 p->sched_class->set_curr_task(rq);
6321 activate_task(rq, p, 0);
6323 check_class_changed(rq, p, prev_class, oldprio, running);
6325 __task_rq_unlock(rq);
6326 spin_unlock_irqrestore(&p->pi_lock, flags);
6328 rt_mutex_adjust_pi(p);
6334 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6335 * @p: the task in question.
6336 * @policy: new policy.
6337 * @param: structure containing the new RT priority.
6339 * NOTE that the task may be already dead.
6341 int sched_setscheduler(struct task_struct *p, int policy,
6342 struct sched_param *param)
6344 return __sched_setscheduler(p, policy, param, true);
6346 EXPORT_SYMBOL_GPL(sched_setscheduler);
6349 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6350 * @p: the task in question.
6351 * @policy: new policy.
6352 * @param: structure containing the new RT priority.
6354 * Just like sched_setscheduler, only don't bother checking if the
6355 * current context has permission. For example, this is needed in
6356 * stop_machine(): we create temporary high priority worker threads,
6357 * but our caller might not have that capability.
6359 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6360 struct sched_param *param)
6362 return __sched_setscheduler(p, policy, param, false);
6366 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6368 struct sched_param lparam;
6369 struct task_struct *p;
6372 if (!param || pid < 0)
6374 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6379 p = find_process_by_pid(pid);
6381 retval = sched_setscheduler(p, policy, &lparam);
6388 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6389 * @pid: the pid in question.
6390 * @policy: new policy.
6391 * @param: structure containing the new RT priority.
6393 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6394 struct sched_param __user *, param)
6396 /* negative values for policy are not valid */
6400 return do_sched_setscheduler(pid, policy, param);
6404 * sys_sched_setparam - set/change the RT priority of a thread
6405 * @pid: the pid in question.
6406 * @param: structure containing the new RT priority.
6408 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6410 return do_sched_setscheduler(pid, -1, param);
6414 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6415 * @pid: the pid in question.
6417 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6419 struct task_struct *p;
6426 read_lock(&tasklist_lock);
6427 p = find_process_by_pid(pid);
6429 retval = security_task_getscheduler(p);
6432 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6434 read_unlock(&tasklist_lock);
6439 * sys_sched_getparam - get the RT priority of a thread
6440 * @pid: the pid in question.
6441 * @param: structure containing the RT priority.
6443 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6445 struct sched_param lp;
6446 struct task_struct *p;
6449 if (!param || pid < 0)
6452 read_lock(&tasklist_lock);
6453 p = find_process_by_pid(pid);
6458 retval = security_task_getscheduler(p);
6462 lp.sched_priority = p->rt_priority;
6463 read_unlock(&tasklist_lock);
6466 * This one might sleep, we cannot do it with a spinlock held ...
6468 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6473 read_unlock(&tasklist_lock);
6477 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6479 cpumask_var_t cpus_allowed, new_mask;
6480 struct task_struct *p;
6484 read_lock(&tasklist_lock);
6486 p = find_process_by_pid(pid);
6488 read_unlock(&tasklist_lock);
6494 * It is not safe to call set_cpus_allowed with the
6495 * tasklist_lock held. We will bump the task_struct's
6496 * usage count and then drop tasklist_lock.
6499 read_unlock(&tasklist_lock);
6501 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6505 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6507 goto out_free_cpus_allowed;
6510 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6513 retval = security_task_setscheduler(p, 0, NULL);
6517 cpuset_cpus_allowed(p, cpus_allowed);
6518 cpumask_and(new_mask, in_mask, cpus_allowed);
6520 retval = set_cpus_allowed_ptr(p, new_mask);
6523 cpuset_cpus_allowed(p, cpus_allowed);
6524 if (!cpumask_subset(new_mask, cpus_allowed)) {
6526 * We must have raced with a concurrent cpuset
6527 * update. Just reset the cpus_allowed to the
6528 * cpuset's cpus_allowed
6530 cpumask_copy(new_mask, cpus_allowed);
6535 free_cpumask_var(new_mask);
6536 out_free_cpus_allowed:
6537 free_cpumask_var(cpus_allowed);
6544 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6545 struct cpumask *new_mask)
6547 if (len < cpumask_size())
6548 cpumask_clear(new_mask);
6549 else if (len > cpumask_size())
6550 len = cpumask_size();
6552 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6556 * sys_sched_setaffinity - set the cpu affinity of a process
6557 * @pid: pid of the process
6558 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6559 * @user_mask_ptr: user-space pointer to the new cpu mask
6561 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6562 unsigned long __user *, user_mask_ptr)
6564 cpumask_var_t new_mask;
6567 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6570 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6572 retval = sched_setaffinity(pid, new_mask);
6573 free_cpumask_var(new_mask);
6577 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6579 struct task_struct *p;
6583 read_lock(&tasklist_lock);
6586 p = find_process_by_pid(pid);
6590 retval = security_task_getscheduler(p);
6594 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6597 read_unlock(&tasklist_lock);
6604 * sys_sched_getaffinity - get the cpu affinity of a process
6605 * @pid: pid of the process
6606 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6607 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6609 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6610 unsigned long __user *, user_mask_ptr)
6615 if (len < cpumask_size())
6618 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6621 ret = sched_getaffinity(pid, mask);
6623 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6626 ret = cpumask_size();
6628 free_cpumask_var(mask);
6634 * sys_sched_yield - yield the current processor to other threads.
6636 * This function yields the current CPU to other tasks. If there are no
6637 * other threads running on this CPU then this function will return.
6639 SYSCALL_DEFINE0(sched_yield)
6641 struct rq *rq = this_rq_lock();
6643 schedstat_inc(rq, yld_count);
6644 current->sched_class->yield_task(rq);
6647 * Since we are going to call schedule() anyway, there's
6648 * no need to preempt or enable interrupts:
6650 __release(rq->lock);
6651 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6652 _raw_spin_unlock(&rq->lock);
6653 preempt_enable_no_resched();
6660 static inline int should_resched(void)
6662 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6665 static void __cond_resched(void)
6667 add_preempt_count(PREEMPT_ACTIVE);
6669 sub_preempt_count(PREEMPT_ACTIVE);
6672 int __sched _cond_resched(void)
6674 if (should_resched()) {
6680 EXPORT_SYMBOL(_cond_resched);
6683 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6684 * call schedule, and on return reacquire the lock.
6686 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6687 * operations here to prevent schedule() from being called twice (once via
6688 * spin_unlock(), once by hand).
6690 int __cond_resched_lock(spinlock_t *lock)
6692 int resched = should_resched();
6695 lockdep_assert_held(lock);
6697 if (spin_needbreak(lock) || resched) {
6708 EXPORT_SYMBOL(__cond_resched_lock);
6710 int __sched __cond_resched_softirq(void)
6712 BUG_ON(!in_softirq());
6714 if (should_resched()) {
6722 EXPORT_SYMBOL(__cond_resched_softirq);
6725 * yield - yield the current processor to other threads.
6727 * This is a shortcut for kernel-space yielding - it marks the
6728 * thread runnable and calls sys_sched_yield().
6730 void __sched yield(void)
6732 set_current_state(TASK_RUNNING);
6735 EXPORT_SYMBOL(yield);
6738 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6739 * that process accounting knows that this is a task in IO wait state.
6741 void __sched io_schedule(void)
6743 struct rq *rq = raw_rq();
6745 delayacct_blkio_start();
6746 atomic_inc(&rq->nr_iowait);
6747 current->in_iowait = 1;
6749 current->in_iowait = 0;
6750 atomic_dec(&rq->nr_iowait);
6751 delayacct_blkio_end();
6753 EXPORT_SYMBOL(io_schedule);
6755 long __sched io_schedule_timeout(long timeout)
6757 struct rq *rq = raw_rq();
6760 delayacct_blkio_start();
6761 atomic_inc(&rq->nr_iowait);
6762 current->in_iowait = 1;
6763 ret = schedule_timeout(timeout);
6764 current->in_iowait = 0;
6765 atomic_dec(&rq->nr_iowait);
6766 delayacct_blkio_end();
6771 * sys_sched_get_priority_max - return maximum RT priority.
6772 * @policy: scheduling class.
6774 * this syscall returns the maximum rt_priority that can be used
6775 * by a given scheduling class.
6777 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6784 ret = MAX_USER_RT_PRIO-1;
6796 * sys_sched_get_priority_min - return minimum RT priority.
6797 * @policy: scheduling class.
6799 * this syscall returns the minimum rt_priority that can be used
6800 * by a given scheduling class.
6802 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6820 * sys_sched_rr_get_interval - return the default timeslice of a process.
6821 * @pid: pid of the process.
6822 * @interval: userspace pointer to the timeslice value.
6824 * this syscall writes the default timeslice value of a given process
6825 * into the user-space timespec buffer. A value of '0' means infinity.
6827 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6828 struct timespec __user *, interval)
6830 struct task_struct *p;
6831 unsigned int time_slice;
6839 read_lock(&tasklist_lock);
6840 p = find_process_by_pid(pid);
6844 retval = security_task_getscheduler(p);
6848 time_slice = p->sched_class->get_rr_interval(p);
6850 read_unlock(&tasklist_lock);
6851 jiffies_to_timespec(time_slice, &t);
6852 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6856 read_unlock(&tasklist_lock);
6860 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6862 void sched_show_task(struct task_struct *p)
6864 unsigned long free = 0;
6867 state = p->state ? __ffs(p->state) + 1 : 0;
6868 printk(KERN_INFO "%-13.13s %c", p->comm,
6869 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6870 #if BITS_PER_LONG == 32
6871 if (state == TASK_RUNNING)
6872 printk(KERN_CONT " running ");
6874 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6876 if (state == TASK_RUNNING)
6877 printk(KERN_CONT " running task ");
6879 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6881 #ifdef CONFIG_DEBUG_STACK_USAGE
6882 free = stack_not_used(p);
6884 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6885 task_pid_nr(p), task_pid_nr(p->real_parent),
6886 (unsigned long)task_thread_info(p)->flags);
6888 show_stack(p, NULL);
6891 void show_state_filter(unsigned long state_filter)
6893 struct task_struct *g, *p;
6895 #if BITS_PER_LONG == 32
6897 " task PC stack pid father\n");
6900 " task PC stack pid father\n");
6902 read_lock(&tasklist_lock);
6903 do_each_thread(g, p) {
6905 * reset the NMI-timeout, listing all files on a slow
6906 * console might take alot of time:
6908 touch_nmi_watchdog();
6909 if (!state_filter || (p->state & state_filter))
6911 } while_each_thread(g, p);
6913 touch_all_softlockup_watchdogs();
6915 #ifdef CONFIG_SCHED_DEBUG
6916 sysrq_sched_debug_show();
6918 read_unlock(&tasklist_lock);
6920 * Only show locks if all tasks are dumped:
6922 if (state_filter == -1)
6923 debug_show_all_locks();
6926 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6928 idle->sched_class = &idle_sched_class;
6932 * init_idle - set up an idle thread for a given CPU
6933 * @idle: task in question
6934 * @cpu: cpu the idle task belongs to
6936 * NOTE: this function does not set the idle thread's NEED_RESCHED
6937 * flag, to make booting more robust.
6939 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6941 struct rq *rq = cpu_rq(cpu);
6942 unsigned long flags;
6944 spin_lock_irqsave(&rq->lock, flags);
6947 idle->se.exec_start = sched_clock();
6949 idle->prio = idle->normal_prio = MAX_PRIO;
6950 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6951 __set_task_cpu(idle, cpu);
6953 rq->curr = rq->idle = idle;
6954 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6957 spin_unlock_irqrestore(&rq->lock, flags);
6959 /* Set the preempt count _outside_ the spinlocks! */
6960 #if defined(CONFIG_PREEMPT)
6961 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6963 task_thread_info(idle)->preempt_count = 0;
6966 * The idle tasks have their own, simple scheduling class:
6968 idle->sched_class = &idle_sched_class;
6969 ftrace_graph_init_task(idle);
6973 * In a system that switches off the HZ timer nohz_cpu_mask
6974 * indicates which cpus entered this state. This is used
6975 * in the rcu update to wait only for active cpus. For system
6976 * which do not switch off the HZ timer nohz_cpu_mask should
6977 * always be CPU_BITS_NONE.
6979 cpumask_var_t nohz_cpu_mask;
6982 * Increase the granularity value when there are more CPUs,
6983 * because with more CPUs the 'effective latency' as visible
6984 * to users decreases. But the relationship is not linear,
6985 * so pick a second-best guess by going with the log2 of the
6988 * This idea comes from the SD scheduler of Con Kolivas:
6990 static inline void sched_init_granularity(void)
6992 unsigned int factor = 1 + ilog2(num_online_cpus());
6993 const unsigned long limit = 200000000;
6995 sysctl_sched_min_granularity *= factor;
6996 if (sysctl_sched_min_granularity > limit)
6997 sysctl_sched_min_granularity = limit;
6999 sysctl_sched_latency *= factor;
7000 if (sysctl_sched_latency > limit)
7001 sysctl_sched_latency = limit;
7003 sysctl_sched_wakeup_granularity *= factor;
7005 sysctl_sched_shares_ratelimit *= factor;
7010 * This is how migration works:
7012 * 1) we queue a struct migration_req structure in the source CPU's
7013 * runqueue and wake up that CPU's migration thread.
7014 * 2) we down() the locked semaphore => thread blocks.
7015 * 3) migration thread wakes up (implicitly it forces the migrated
7016 * thread off the CPU)
7017 * 4) it gets the migration request and checks whether the migrated
7018 * task is still in the wrong runqueue.
7019 * 5) if it's in the wrong runqueue then the migration thread removes
7020 * it and puts it into the right queue.
7021 * 6) migration thread up()s the semaphore.
7022 * 7) we wake up and the migration is done.
7026 * Change a given task's CPU affinity. Migrate the thread to a
7027 * proper CPU and schedule it away if the CPU it's executing on
7028 * is removed from the allowed bitmask.
7030 * NOTE: the caller must have a valid reference to the task, the
7031 * task must not exit() & deallocate itself prematurely. The
7032 * call is not atomic; no spinlocks may be held.
7034 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7036 struct migration_req req;
7037 unsigned long flags;
7041 rq = task_rq_lock(p, &flags);
7042 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
7047 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7048 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7053 if (p->sched_class->set_cpus_allowed)
7054 p->sched_class->set_cpus_allowed(p, new_mask);
7056 cpumask_copy(&p->cpus_allowed, new_mask);
7057 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7060 /* Can the task run on the task's current CPU? If so, we're done */
7061 if (cpumask_test_cpu(task_cpu(p), new_mask))
7064 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
7065 /* Need help from migration thread: drop lock and wait. */
7066 struct task_struct *mt = rq->migration_thread;
7068 get_task_struct(mt);
7069 task_rq_unlock(rq, &flags);
7070 wake_up_process(rq->migration_thread);
7071 put_task_struct(mt);
7072 wait_for_completion(&req.done);
7073 tlb_migrate_finish(p->mm);
7077 task_rq_unlock(rq, &flags);
7081 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7084 * Move (not current) task off this cpu, onto dest cpu. We're doing
7085 * this because either it can't run here any more (set_cpus_allowed()
7086 * away from this CPU, or CPU going down), or because we're
7087 * attempting to rebalance this task on exec (sched_exec).
7089 * So we race with normal scheduler movements, but that's OK, as long
7090 * as the task is no longer on this CPU.
7092 * Returns non-zero if task was successfully migrated.
7094 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7096 struct rq *rq_dest, *rq_src;
7099 if (unlikely(!cpu_active(dest_cpu)))
7102 rq_src = cpu_rq(src_cpu);
7103 rq_dest = cpu_rq(dest_cpu);
7105 double_rq_lock(rq_src, rq_dest);
7106 /* Already moved. */
7107 if (task_cpu(p) != src_cpu)
7109 /* Affinity changed (again). */
7110 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7113 on_rq = p->se.on_rq;
7115 deactivate_task(rq_src, p, 0);
7117 set_task_cpu(p, dest_cpu);
7119 activate_task(rq_dest, p, 0);
7120 check_preempt_curr(rq_dest, p, 0);
7125 double_rq_unlock(rq_src, rq_dest);
7129 #define RCU_MIGRATION_IDLE 0
7130 #define RCU_MIGRATION_NEED_QS 1
7131 #define RCU_MIGRATION_GOT_QS 2
7132 #define RCU_MIGRATION_MUST_SYNC 3
7135 * migration_thread - this is a highprio system thread that performs
7136 * thread migration by bumping thread off CPU then 'pushing' onto
7139 static int migration_thread(void *data)
7142 int cpu = (long)data;
7146 BUG_ON(rq->migration_thread != current);
7148 set_current_state(TASK_INTERRUPTIBLE);
7149 while (!kthread_should_stop()) {
7150 struct migration_req *req;
7151 struct list_head *head;
7153 spin_lock_irq(&rq->lock);
7155 if (cpu_is_offline(cpu)) {
7156 spin_unlock_irq(&rq->lock);
7160 if (rq->active_balance) {
7161 active_load_balance(rq, cpu);
7162 rq->active_balance = 0;
7165 head = &rq->migration_queue;
7167 if (list_empty(head)) {
7168 spin_unlock_irq(&rq->lock);
7170 set_current_state(TASK_INTERRUPTIBLE);
7173 req = list_entry(head->next, struct migration_req, list);
7174 list_del_init(head->next);
7176 if (req->task != NULL) {
7177 spin_unlock(&rq->lock);
7178 __migrate_task(req->task, cpu, req->dest_cpu);
7179 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7180 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7181 spin_unlock(&rq->lock);
7183 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7184 spin_unlock(&rq->lock);
7185 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7189 complete(&req->done);
7191 __set_current_state(TASK_RUNNING);
7196 #ifdef CONFIG_HOTPLUG_CPU
7198 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7202 local_irq_disable();
7203 ret = __migrate_task(p, src_cpu, dest_cpu);
7209 * Figure out where task on dead CPU should go, use force if necessary.
7211 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7214 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7217 /* Look for allowed, online CPU in same node. */
7218 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7219 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7222 /* Any allowed, online CPU? */
7223 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7224 if (dest_cpu < nr_cpu_ids)
7227 /* No more Mr. Nice Guy. */
7228 if (dest_cpu >= nr_cpu_ids) {
7229 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7230 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7233 * Don't tell them about moving exiting tasks or
7234 * kernel threads (both mm NULL), since they never
7237 if (p->mm && printk_ratelimit()) {
7238 printk(KERN_INFO "process %d (%s) no "
7239 "longer affine to cpu%d\n",
7240 task_pid_nr(p), p->comm, dead_cpu);
7245 /* It can have affinity changed while we were choosing. */
7246 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7251 * While a dead CPU has no uninterruptible tasks queued at this point,
7252 * it might still have a nonzero ->nr_uninterruptible counter, because
7253 * for performance reasons the counter is not stricly tracking tasks to
7254 * their home CPUs. So we just add the counter to another CPU's counter,
7255 * to keep the global sum constant after CPU-down:
7257 static void migrate_nr_uninterruptible(struct rq *rq_src)
7259 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7260 unsigned long flags;
7262 local_irq_save(flags);
7263 double_rq_lock(rq_src, rq_dest);
7264 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7265 rq_src->nr_uninterruptible = 0;
7266 double_rq_unlock(rq_src, rq_dest);
7267 local_irq_restore(flags);
7270 /* Run through task list and migrate tasks from the dead cpu. */
7271 static void migrate_live_tasks(int src_cpu)
7273 struct task_struct *p, *t;
7275 read_lock(&tasklist_lock);
7277 do_each_thread(t, p) {
7281 if (task_cpu(p) == src_cpu)
7282 move_task_off_dead_cpu(src_cpu, p);
7283 } while_each_thread(t, p);
7285 read_unlock(&tasklist_lock);
7289 * Schedules idle task to be the next runnable task on current CPU.
7290 * It does so by boosting its priority to highest possible.
7291 * Used by CPU offline code.
7293 void sched_idle_next(void)
7295 int this_cpu = smp_processor_id();
7296 struct rq *rq = cpu_rq(this_cpu);
7297 struct task_struct *p = rq->idle;
7298 unsigned long flags;
7300 /* cpu has to be offline */
7301 BUG_ON(cpu_online(this_cpu));
7304 * Strictly not necessary since rest of the CPUs are stopped by now
7305 * and interrupts disabled on the current cpu.
7307 spin_lock_irqsave(&rq->lock, flags);
7309 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7311 update_rq_clock(rq);
7312 activate_task(rq, p, 0);
7314 spin_unlock_irqrestore(&rq->lock, flags);
7318 * Ensures that the idle task is using init_mm right before its cpu goes
7321 void idle_task_exit(void)
7323 struct mm_struct *mm = current->active_mm;
7325 BUG_ON(cpu_online(smp_processor_id()));
7328 switch_mm(mm, &init_mm, current);
7332 /* called under rq->lock with disabled interrupts */
7333 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7335 struct rq *rq = cpu_rq(dead_cpu);
7337 /* Must be exiting, otherwise would be on tasklist. */
7338 BUG_ON(!p->exit_state);
7340 /* Cannot have done final schedule yet: would have vanished. */
7341 BUG_ON(p->state == TASK_DEAD);
7346 * Drop lock around migration; if someone else moves it,
7347 * that's OK. No task can be added to this CPU, so iteration is
7350 spin_unlock_irq(&rq->lock);
7351 move_task_off_dead_cpu(dead_cpu, p);
7352 spin_lock_irq(&rq->lock);
7357 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7358 static void migrate_dead_tasks(unsigned int dead_cpu)
7360 struct rq *rq = cpu_rq(dead_cpu);
7361 struct task_struct *next;
7364 if (!rq->nr_running)
7366 update_rq_clock(rq);
7367 next = pick_next_task(rq);
7370 next->sched_class->put_prev_task(rq, next);
7371 migrate_dead(dead_cpu, next);
7377 * remove the tasks which were accounted by rq from calc_load_tasks.
7379 static void calc_global_load_remove(struct rq *rq)
7381 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7382 rq->calc_load_active = 0;
7384 #endif /* CONFIG_HOTPLUG_CPU */
7386 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7388 static struct ctl_table sd_ctl_dir[] = {
7390 .procname = "sched_domain",
7396 static struct ctl_table sd_ctl_root[] = {
7398 .ctl_name = CTL_KERN,
7399 .procname = "kernel",
7401 .child = sd_ctl_dir,
7406 static struct ctl_table *sd_alloc_ctl_entry(int n)
7408 struct ctl_table *entry =
7409 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7414 static void sd_free_ctl_entry(struct ctl_table **tablep)
7416 struct ctl_table *entry;
7419 * In the intermediate directories, both the child directory and
7420 * procname are dynamically allocated and could fail but the mode
7421 * will always be set. In the lowest directory the names are
7422 * static strings and all have proc handlers.
7424 for (entry = *tablep; entry->mode; entry++) {
7426 sd_free_ctl_entry(&entry->child);
7427 if (entry->proc_handler == NULL)
7428 kfree(entry->procname);
7436 set_table_entry(struct ctl_table *entry,
7437 const char *procname, void *data, int maxlen,
7438 mode_t mode, proc_handler *proc_handler)
7440 entry->procname = procname;
7442 entry->maxlen = maxlen;
7444 entry->proc_handler = proc_handler;
7447 static struct ctl_table *
7448 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7450 struct ctl_table *table = sd_alloc_ctl_entry(13);
7455 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7456 sizeof(long), 0644, proc_doulongvec_minmax);
7457 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7458 sizeof(long), 0644, proc_doulongvec_minmax);
7459 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7460 sizeof(int), 0644, proc_dointvec_minmax);
7461 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7462 sizeof(int), 0644, proc_dointvec_minmax);
7463 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7464 sizeof(int), 0644, proc_dointvec_minmax);
7465 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7466 sizeof(int), 0644, proc_dointvec_minmax);
7467 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7468 sizeof(int), 0644, proc_dointvec_minmax);
7469 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7470 sizeof(int), 0644, proc_dointvec_minmax);
7471 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7472 sizeof(int), 0644, proc_dointvec_minmax);
7473 set_table_entry(&table[9], "cache_nice_tries",
7474 &sd->cache_nice_tries,
7475 sizeof(int), 0644, proc_dointvec_minmax);
7476 set_table_entry(&table[10], "flags", &sd->flags,
7477 sizeof(int), 0644, proc_dointvec_minmax);
7478 set_table_entry(&table[11], "name", sd->name,
7479 CORENAME_MAX_SIZE, 0444, proc_dostring);
7480 /* &table[12] is terminator */
7485 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7487 struct ctl_table *entry, *table;
7488 struct sched_domain *sd;
7489 int domain_num = 0, i;
7492 for_each_domain(cpu, sd)
7494 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7499 for_each_domain(cpu, sd) {
7500 snprintf(buf, 32, "domain%d", i);
7501 entry->procname = kstrdup(buf, GFP_KERNEL);
7503 entry->child = sd_alloc_ctl_domain_table(sd);
7510 static struct ctl_table_header *sd_sysctl_header;
7511 static void register_sched_domain_sysctl(void)
7513 int i, cpu_num = num_online_cpus();
7514 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7517 WARN_ON(sd_ctl_dir[0].child);
7518 sd_ctl_dir[0].child = entry;
7523 for_each_online_cpu(i) {
7524 snprintf(buf, 32, "cpu%d", i);
7525 entry->procname = kstrdup(buf, GFP_KERNEL);
7527 entry->child = sd_alloc_ctl_cpu_table(i);
7531 WARN_ON(sd_sysctl_header);
7532 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7535 /* may be called multiple times per register */
7536 static void unregister_sched_domain_sysctl(void)
7538 if (sd_sysctl_header)
7539 unregister_sysctl_table(sd_sysctl_header);
7540 sd_sysctl_header = NULL;
7541 if (sd_ctl_dir[0].child)
7542 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7545 static void register_sched_domain_sysctl(void)
7548 static void unregister_sched_domain_sysctl(void)
7553 static void set_rq_online(struct rq *rq)
7556 const struct sched_class *class;
7558 cpumask_set_cpu(rq->cpu, rq->rd->online);
7561 for_each_class(class) {
7562 if (class->rq_online)
7563 class->rq_online(rq);
7568 static void set_rq_offline(struct rq *rq)
7571 const struct sched_class *class;
7573 for_each_class(class) {
7574 if (class->rq_offline)
7575 class->rq_offline(rq);
7578 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7584 * migration_call - callback that gets triggered when a CPU is added.
7585 * Here we can start up the necessary migration thread for the new CPU.
7587 static int __cpuinit
7588 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7590 struct task_struct *p;
7591 int cpu = (long)hcpu;
7592 unsigned long flags;
7597 case CPU_UP_PREPARE:
7598 case CPU_UP_PREPARE_FROZEN:
7599 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7602 kthread_bind(p, cpu);
7603 /* Must be high prio: stop_machine expects to yield to it. */
7604 rq = task_rq_lock(p, &flags);
7605 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7606 task_rq_unlock(rq, &flags);
7608 cpu_rq(cpu)->migration_thread = p;
7609 rq->calc_load_update = calc_load_update;
7613 case CPU_ONLINE_FROZEN:
7614 /* Strictly unnecessary, as first user will wake it. */
7615 wake_up_process(cpu_rq(cpu)->migration_thread);
7617 /* Update our root-domain */
7619 spin_lock_irqsave(&rq->lock, flags);
7621 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7625 spin_unlock_irqrestore(&rq->lock, flags);
7628 #ifdef CONFIG_HOTPLUG_CPU
7629 case CPU_UP_CANCELED:
7630 case CPU_UP_CANCELED_FROZEN:
7631 if (!cpu_rq(cpu)->migration_thread)
7633 /* Unbind it from offline cpu so it can run. Fall thru. */
7634 kthread_bind(cpu_rq(cpu)->migration_thread,
7635 cpumask_any(cpu_online_mask));
7636 kthread_stop(cpu_rq(cpu)->migration_thread);
7637 put_task_struct(cpu_rq(cpu)->migration_thread);
7638 cpu_rq(cpu)->migration_thread = NULL;
7642 case CPU_DEAD_FROZEN:
7643 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7644 migrate_live_tasks(cpu);
7646 kthread_stop(rq->migration_thread);
7647 put_task_struct(rq->migration_thread);
7648 rq->migration_thread = NULL;
7649 /* Idle task back to normal (off runqueue, low prio) */
7650 spin_lock_irq(&rq->lock);
7651 update_rq_clock(rq);
7652 deactivate_task(rq, rq->idle, 0);
7653 rq->idle->static_prio = MAX_PRIO;
7654 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7655 rq->idle->sched_class = &idle_sched_class;
7656 migrate_dead_tasks(cpu);
7657 spin_unlock_irq(&rq->lock);
7659 migrate_nr_uninterruptible(rq);
7660 BUG_ON(rq->nr_running != 0);
7661 calc_global_load_remove(rq);
7663 * No need to migrate the tasks: it was best-effort if
7664 * they didn't take sched_hotcpu_mutex. Just wake up
7667 spin_lock_irq(&rq->lock);
7668 while (!list_empty(&rq->migration_queue)) {
7669 struct migration_req *req;
7671 req = list_entry(rq->migration_queue.next,
7672 struct migration_req, list);
7673 list_del_init(&req->list);
7674 spin_unlock_irq(&rq->lock);
7675 complete(&req->done);
7676 spin_lock_irq(&rq->lock);
7678 spin_unlock_irq(&rq->lock);
7682 case CPU_DYING_FROZEN:
7683 /* Update our root-domain */
7685 spin_lock_irqsave(&rq->lock, flags);
7687 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7690 spin_unlock_irqrestore(&rq->lock, flags);
7698 * Register at high priority so that task migration (migrate_all_tasks)
7699 * happens before everything else. This has to be lower priority than
7700 * the notifier in the perf_event subsystem, though.
7702 static struct notifier_block __cpuinitdata migration_notifier = {
7703 .notifier_call = migration_call,
7707 static int __init migration_init(void)
7709 void *cpu = (void *)(long)smp_processor_id();
7712 /* Start one for the boot CPU: */
7713 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7714 BUG_ON(err == NOTIFY_BAD);
7715 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7716 register_cpu_notifier(&migration_notifier);
7720 early_initcall(migration_init);
7725 #ifdef CONFIG_SCHED_DEBUG
7727 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7728 struct cpumask *groupmask)
7730 struct sched_group *group = sd->groups;
7733 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7734 cpumask_clear(groupmask);
7736 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7738 if (!(sd->flags & SD_LOAD_BALANCE)) {
7739 printk("does not load-balance\n");
7741 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7746 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7748 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7749 printk(KERN_ERR "ERROR: domain->span does not contain "
7752 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7753 printk(KERN_ERR "ERROR: domain->groups does not contain"
7757 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7761 printk(KERN_ERR "ERROR: group is NULL\n");
7765 if (!group->cpu_power) {
7766 printk(KERN_CONT "\n");
7767 printk(KERN_ERR "ERROR: domain->cpu_power not "
7772 if (!cpumask_weight(sched_group_cpus(group))) {
7773 printk(KERN_CONT "\n");
7774 printk(KERN_ERR "ERROR: empty group\n");
7778 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7779 printk(KERN_CONT "\n");
7780 printk(KERN_ERR "ERROR: repeated CPUs\n");
7784 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7786 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7788 printk(KERN_CONT " %s", str);
7789 if (group->cpu_power != SCHED_LOAD_SCALE) {
7790 printk(KERN_CONT " (cpu_power = %d)",
7794 group = group->next;
7795 } while (group != sd->groups);
7796 printk(KERN_CONT "\n");
7798 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7799 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7802 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7803 printk(KERN_ERR "ERROR: parent span is not a superset "
7804 "of domain->span\n");
7808 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7810 cpumask_var_t groupmask;
7814 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7818 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7820 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7821 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7826 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7833 free_cpumask_var(groupmask);
7835 #else /* !CONFIG_SCHED_DEBUG */
7836 # define sched_domain_debug(sd, cpu) do { } while (0)
7837 #endif /* CONFIG_SCHED_DEBUG */
7839 static int sd_degenerate(struct sched_domain *sd)
7841 if (cpumask_weight(sched_domain_span(sd)) == 1)
7844 /* Following flags need at least 2 groups */
7845 if (sd->flags & (SD_LOAD_BALANCE |
7846 SD_BALANCE_NEWIDLE |
7850 SD_SHARE_PKG_RESOURCES)) {
7851 if (sd->groups != sd->groups->next)
7855 /* Following flags don't use groups */
7856 if (sd->flags & (SD_WAKE_AFFINE))
7863 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7865 unsigned long cflags = sd->flags, pflags = parent->flags;
7867 if (sd_degenerate(parent))
7870 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7873 /* Flags needing groups don't count if only 1 group in parent */
7874 if (parent->groups == parent->groups->next) {
7875 pflags &= ~(SD_LOAD_BALANCE |
7876 SD_BALANCE_NEWIDLE |
7880 SD_SHARE_PKG_RESOURCES);
7881 if (nr_node_ids == 1)
7882 pflags &= ~SD_SERIALIZE;
7884 if (~cflags & pflags)
7890 static void free_rootdomain(struct root_domain *rd)
7892 cpupri_cleanup(&rd->cpupri);
7894 free_cpumask_var(rd->rto_mask);
7895 free_cpumask_var(rd->online);
7896 free_cpumask_var(rd->span);
7900 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7902 struct root_domain *old_rd = NULL;
7903 unsigned long flags;
7905 spin_lock_irqsave(&rq->lock, flags);
7910 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7913 cpumask_clear_cpu(rq->cpu, old_rd->span);
7916 * If we dont want to free the old_rt yet then
7917 * set old_rd to NULL to skip the freeing later
7920 if (!atomic_dec_and_test(&old_rd->refcount))
7924 atomic_inc(&rd->refcount);
7927 cpumask_set_cpu(rq->cpu, rd->span);
7928 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7931 spin_unlock_irqrestore(&rq->lock, flags);
7934 free_rootdomain(old_rd);
7937 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7939 gfp_t gfp = GFP_KERNEL;
7941 memset(rd, 0, sizeof(*rd));
7946 if (!alloc_cpumask_var(&rd->span, gfp))
7948 if (!alloc_cpumask_var(&rd->online, gfp))
7950 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
7953 if (cpupri_init(&rd->cpupri, bootmem) != 0)
7958 free_cpumask_var(rd->rto_mask);
7960 free_cpumask_var(rd->online);
7962 free_cpumask_var(rd->span);
7967 static void init_defrootdomain(void)
7969 init_rootdomain(&def_root_domain, true);
7971 atomic_set(&def_root_domain.refcount, 1);
7974 static struct root_domain *alloc_rootdomain(void)
7976 struct root_domain *rd;
7978 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7982 if (init_rootdomain(rd, false) != 0) {
7991 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7992 * hold the hotplug lock.
7995 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7997 struct rq *rq = cpu_rq(cpu);
7998 struct sched_domain *tmp;
8000 /* Remove the sched domains which do not contribute to scheduling. */
8001 for (tmp = sd; tmp; ) {
8002 struct sched_domain *parent = tmp->parent;
8006 if (sd_parent_degenerate(tmp, parent)) {
8007 tmp->parent = parent->parent;
8009 parent->parent->child = tmp;
8014 if (sd && sd_degenerate(sd)) {
8020 sched_domain_debug(sd, cpu);
8022 rq_attach_root(rq, rd);
8023 rcu_assign_pointer(rq->sd, sd);
8026 /* cpus with isolated domains */
8027 static cpumask_var_t cpu_isolated_map;
8029 /* Setup the mask of cpus configured for isolated domains */
8030 static int __init isolated_cpu_setup(char *str)
8032 cpulist_parse(str, cpu_isolated_map);
8036 __setup("isolcpus=", isolated_cpu_setup);
8039 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8040 * to a function which identifies what group(along with sched group) a CPU
8041 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8042 * (due to the fact that we keep track of groups covered with a struct cpumask).
8044 * init_sched_build_groups will build a circular linked list of the groups
8045 * covered by the given span, and will set each group's ->cpumask correctly,
8046 * and ->cpu_power to 0.
8049 init_sched_build_groups(const struct cpumask *span,
8050 const struct cpumask *cpu_map,
8051 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8052 struct sched_group **sg,
8053 struct cpumask *tmpmask),
8054 struct cpumask *covered, struct cpumask *tmpmask)
8056 struct sched_group *first = NULL, *last = NULL;
8059 cpumask_clear(covered);
8061 for_each_cpu(i, span) {
8062 struct sched_group *sg;
8063 int group = group_fn(i, cpu_map, &sg, tmpmask);
8066 if (cpumask_test_cpu(i, covered))
8069 cpumask_clear(sched_group_cpus(sg));
8072 for_each_cpu(j, span) {
8073 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8076 cpumask_set_cpu(j, covered);
8077 cpumask_set_cpu(j, sched_group_cpus(sg));
8088 #define SD_NODES_PER_DOMAIN 16
8093 * find_next_best_node - find the next node to include in a sched_domain
8094 * @node: node whose sched_domain we're building
8095 * @used_nodes: nodes already in the sched_domain
8097 * Find the next node to include in a given scheduling domain. Simply
8098 * finds the closest node not already in the @used_nodes map.
8100 * Should use nodemask_t.
8102 static int find_next_best_node(int node, nodemask_t *used_nodes)
8104 int i, n, val, min_val, best_node = 0;
8108 for (i = 0; i < nr_node_ids; i++) {
8109 /* Start at @node */
8110 n = (node + i) % nr_node_ids;
8112 if (!nr_cpus_node(n))
8115 /* Skip already used nodes */
8116 if (node_isset(n, *used_nodes))
8119 /* Simple min distance search */
8120 val = node_distance(node, n);
8122 if (val < min_val) {
8128 node_set(best_node, *used_nodes);
8133 * sched_domain_node_span - get a cpumask for a node's sched_domain
8134 * @node: node whose cpumask we're constructing
8135 * @span: resulting cpumask
8137 * Given a node, construct a good cpumask for its sched_domain to span. It
8138 * should be one that prevents unnecessary balancing, but also spreads tasks
8141 static void sched_domain_node_span(int node, struct cpumask *span)
8143 nodemask_t used_nodes;
8146 cpumask_clear(span);
8147 nodes_clear(used_nodes);
8149 cpumask_or(span, span, cpumask_of_node(node));
8150 node_set(node, used_nodes);
8152 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8153 int next_node = find_next_best_node(node, &used_nodes);
8155 cpumask_or(span, span, cpumask_of_node(next_node));
8158 #endif /* CONFIG_NUMA */
8160 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8163 * The cpus mask in sched_group and sched_domain hangs off the end.
8165 * ( See the the comments in include/linux/sched.h:struct sched_group
8166 * and struct sched_domain. )
8168 struct static_sched_group {
8169 struct sched_group sg;
8170 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8173 struct static_sched_domain {
8174 struct sched_domain sd;
8175 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8181 cpumask_var_t domainspan;
8182 cpumask_var_t covered;
8183 cpumask_var_t notcovered;
8185 cpumask_var_t nodemask;
8186 cpumask_var_t this_sibling_map;
8187 cpumask_var_t this_core_map;
8188 cpumask_var_t send_covered;
8189 cpumask_var_t tmpmask;
8190 struct sched_group **sched_group_nodes;
8191 struct root_domain *rd;
8195 sa_sched_groups = 0,
8200 sa_this_sibling_map,
8202 sa_sched_group_nodes,
8212 * SMT sched-domains:
8214 #ifdef CONFIG_SCHED_SMT
8215 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8216 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8219 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8220 struct sched_group **sg, struct cpumask *unused)
8223 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8226 #endif /* CONFIG_SCHED_SMT */
8229 * multi-core sched-domains:
8231 #ifdef CONFIG_SCHED_MC
8232 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8233 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8234 #endif /* CONFIG_SCHED_MC */
8236 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8238 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8239 struct sched_group **sg, struct cpumask *mask)
8243 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8244 group = cpumask_first(mask);
8246 *sg = &per_cpu(sched_group_core, group).sg;
8249 #elif defined(CONFIG_SCHED_MC)
8251 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8252 struct sched_group **sg, struct cpumask *unused)
8255 *sg = &per_cpu(sched_group_core, cpu).sg;
8260 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8261 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8264 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8265 struct sched_group **sg, struct cpumask *mask)
8268 #ifdef CONFIG_SCHED_MC
8269 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8270 group = cpumask_first(mask);
8271 #elif defined(CONFIG_SCHED_SMT)
8272 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8273 group = cpumask_first(mask);
8278 *sg = &per_cpu(sched_group_phys, group).sg;
8284 * The init_sched_build_groups can't handle what we want to do with node
8285 * groups, so roll our own. Now each node has its own list of groups which
8286 * gets dynamically allocated.
8288 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8289 static struct sched_group ***sched_group_nodes_bycpu;
8291 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8292 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8294 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8295 struct sched_group **sg,
8296 struct cpumask *nodemask)
8300 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8301 group = cpumask_first(nodemask);
8304 *sg = &per_cpu(sched_group_allnodes, group).sg;
8308 static void init_numa_sched_groups_power(struct sched_group *group_head)
8310 struct sched_group *sg = group_head;
8316 for_each_cpu(j, sched_group_cpus(sg)) {
8317 struct sched_domain *sd;
8319 sd = &per_cpu(phys_domains, j).sd;
8320 if (j != group_first_cpu(sd->groups)) {
8322 * Only add "power" once for each
8328 sg->cpu_power += sd->groups->cpu_power;
8331 } while (sg != group_head);
8334 static int build_numa_sched_groups(struct s_data *d,
8335 const struct cpumask *cpu_map, int num)
8337 struct sched_domain *sd;
8338 struct sched_group *sg, *prev;
8341 cpumask_clear(d->covered);
8342 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8343 if (cpumask_empty(d->nodemask)) {
8344 d->sched_group_nodes[num] = NULL;
8348 sched_domain_node_span(num, d->domainspan);
8349 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8351 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8354 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8358 d->sched_group_nodes[num] = sg;
8360 for_each_cpu(j, d->nodemask) {
8361 sd = &per_cpu(node_domains, j).sd;
8366 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8368 cpumask_or(d->covered, d->covered, d->nodemask);
8371 for (j = 0; j < nr_node_ids; j++) {
8372 n = (num + j) % nr_node_ids;
8373 cpumask_complement(d->notcovered, d->covered);
8374 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8375 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8376 if (cpumask_empty(d->tmpmask))
8378 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8379 if (cpumask_empty(d->tmpmask))
8381 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8385 "Can not alloc domain group for node %d\n", j);
8389 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8390 sg->next = prev->next;
8391 cpumask_or(d->covered, d->covered, d->tmpmask);
8398 #endif /* CONFIG_NUMA */
8401 /* Free memory allocated for various sched_group structures */
8402 static void free_sched_groups(const struct cpumask *cpu_map,
8403 struct cpumask *nodemask)
8407 for_each_cpu(cpu, cpu_map) {
8408 struct sched_group **sched_group_nodes
8409 = sched_group_nodes_bycpu[cpu];
8411 if (!sched_group_nodes)
8414 for (i = 0; i < nr_node_ids; i++) {
8415 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8417 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8418 if (cpumask_empty(nodemask))
8428 if (oldsg != sched_group_nodes[i])
8431 kfree(sched_group_nodes);
8432 sched_group_nodes_bycpu[cpu] = NULL;
8435 #else /* !CONFIG_NUMA */
8436 static void free_sched_groups(const struct cpumask *cpu_map,
8437 struct cpumask *nodemask)
8440 #endif /* CONFIG_NUMA */
8443 * Initialize sched groups cpu_power.
8445 * cpu_power indicates the capacity of sched group, which is used while
8446 * distributing the load between different sched groups in a sched domain.
8447 * Typically cpu_power for all the groups in a sched domain will be same unless
8448 * there are asymmetries in the topology. If there are asymmetries, group
8449 * having more cpu_power will pickup more load compared to the group having
8452 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8454 struct sched_domain *child;
8455 struct sched_group *group;
8459 WARN_ON(!sd || !sd->groups);
8461 if (cpu != group_first_cpu(sd->groups))
8466 sd->groups->cpu_power = 0;
8469 power = SCHED_LOAD_SCALE;
8470 weight = cpumask_weight(sched_domain_span(sd));
8472 * SMT siblings share the power of a single core.
8473 * Usually multiple threads get a better yield out of
8474 * that one core than a single thread would have,
8475 * reflect that in sd->smt_gain.
8477 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8478 power *= sd->smt_gain;
8480 power >>= SCHED_LOAD_SHIFT;
8482 sd->groups->cpu_power += power;
8487 * Add cpu_power of each child group to this groups cpu_power.
8489 group = child->groups;
8491 sd->groups->cpu_power += group->cpu_power;
8492 group = group->next;
8493 } while (group != child->groups);
8497 * Initializers for schedule domains
8498 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8501 #ifdef CONFIG_SCHED_DEBUG
8502 # define SD_INIT_NAME(sd, type) sd->name = #type
8504 # define SD_INIT_NAME(sd, type) do { } while (0)
8507 #define SD_INIT(sd, type) sd_init_##type(sd)
8509 #define SD_INIT_FUNC(type) \
8510 static noinline void sd_init_##type(struct sched_domain *sd) \
8512 memset(sd, 0, sizeof(*sd)); \
8513 *sd = SD_##type##_INIT; \
8514 sd->level = SD_LV_##type; \
8515 SD_INIT_NAME(sd, type); \
8520 SD_INIT_FUNC(ALLNODES)
8523 #ifdef CONFIG_SCHED_SMT
8524 SD_INIT_FUNC(SIBLING)
8526 #ifdef CONFIG_SCHED_MC
8530 static int default_relax_domain_level = -1;
8532 static int __init setup_relax_domain_level(char *str)
8536 val = simple_strtoul(str, NULL, 0);
8537 if (val < SD_LV_MAX)
8538 default_relax_domain_level = val;
8542 __setup("relax_domain_level=", setup_relax_domain_level);
8544 static void set_domain_attribute(struct sched_domain *sd,
8545 struct sched_domain_attr *attr)
8549 if (!attr || attr->relax_domain_level < 0) {
8550 if (default_relax_domain_level < 0)
8553 request = default_relax_domain_level;
8555 request = attr->relax_domain_level;
8556 if (request < sd->level) {
8557 /* turn off idle balance on this domain */
8558 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8560 /* turn on idle balance on this domain */
8561 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8565 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8566 const struct cpumask *cpu_map)
8569 case sa_sched_groups:
8570 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8571 d->sched_group_nodes = NULL;
8573 free_rootdomain(d->rd); /* fall through */
8575 free_cpumask_var(d->tmpmask); /* fall through */
8576 case sa_send_covered:
8577 free_cpumask_var(d->send_covered); /* fall through */
8578 case sa_this_core_map:
8579 free_cpumask_var(d->this_core_map); /* fall through */
8580 case sa_this_sibling_map:
8581 free_cpumask_var(d->this_sibling_map); /* fall through */
8583 free_cpumask_var(d->nodemask); /* fall through */
8584 case sa_sched_group_nodes:
8586 kfree(d->sched_group_nodes); /* fall through */
8588 free_cpumask_var(d->notcovered); /* fall through */
8590 free_cpumask_var(d->covered); /* fall through */
8592 free_cpumask_var(d->domainspan); /* fall through */
8599 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8600 const struct cpumask *cpu_map)
8603 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8605 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8606 return sa_domainspan;
8607 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8609 /* Allocate the per-node list of sched groups */
8610 d->sched_group_nodes = kcalloc(nr_node_ids,
8611 sizeof(struct sched_group *), GFP_KERNEL);
8612 if (!d->sched_group_nodes) {
8613 printk(KERN_WARNING "Can not alloc sched group node list\n");
8614 return sa_notcovered;
8616 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8618 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8619 return sa_sched_group_nodes;
8620 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8622 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8623 return sa_this_sibling_map;
8624 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8625 return sa_this_core_map;
8626 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8627 return sa_send_covered;
8628 d->rd = alloc_rootdomain();
8630 printk(KERN_WARNING "Cannot alloc root domain\n");
8633 return sa_rootdomain;
8636 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8637 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8639 struct sched_domain *sd = NULL;
8641 struct sched_domain *parent;
8644 if (cpumask_weight(cpu_map) >
8645 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8646 sd = &per_cpu(allnodes_domains, i).sd;
8647 SD_INIT(sd, ALLNODES);
8648 set_domain_attribute(sd, attr);
8649 cpumask_copy(sched_domain_span(sd), cpu_map);
8650 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8655 sd = &per_cpu(node_domains, i).sd;
8657 set_domain_attribute(sd, attr);
8658 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8659 sd->parent = parent;
8662 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8667 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8668 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8669 struct sched_domain *parent, int i)
8671 struct sched_domain *sd;
8672 sd = &per_cpu(phys_domains, i).sd;
8674 set_domain_attribute(sd, attr);
8675 cpumask_copy(sched_domain_span(sd), d->nodemask);
8676 sd->parent = parent;
8679 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8683 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8684 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8685 struct sched_domain *parent, int i)
8687 struct sched_domain *sd = parent;
8688 #ifdef CONFIG_SCHED_MC
8689 sd = &per_cpu(core_domains, i).sd;
8691 set_domain_attribute(sd, attr);
8692 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8693 sd->parent = parent;
8695 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8700 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8701 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8702 struct sched_domain *parent, int i)
8704 struct sched_domain *sd = parent;
8705 #ifdef CONFIG_SCHED_SMT
8706 sd = &per_cpu(cpu_domains, i).sd;
8707 SD_INIT(sd, SIBLING);
8708 set_domain_attribute(sd, attr);
8709 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8710 sd->parent = parent;
8712 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8717 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8718 const struct cpumask *cpu_map, int cpu)
8721 #ifdef CONFIG_SCHED_SMT
8722 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8723 cpumask_and(d->this_sibling_map, cpu_map,
8724 topology_thread_cpumask(cpu));
8725 if (cpu == cpumask_first(d->this_sibling_map))
8726 init_sched_build_groups(d->this_sibling_map, cpu_map,
8728 d->send_covered, d->tmpmask);
8731 #ifdef CONFIG_SCHED_MC
8732 case SD_LV_MC: /* set up multi-core groups */
8733 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8734 if (cpu == cpumask_first(d->this_core_map))
8735 init_sched_build_groups(d->this_core_map, cpu_map,
8737 d->send_covered, d->tmpmask);
8740 case SD_LV_CPU: /* set up physical groups */
8741 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8742 if (!cpumask_empty(d->nodemask))
8743 init_sched_build_groups(d->nodemask, cpu_map,
8745 d->send_covered, d->tmpmask);
8748 case SD_LV_ALLNODES:
8749 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8750 d->send_covered, d->tmpmask);
8759 * Build sched domains for a given set of cpus and attach the sched domains
8760 * to the individual cpus
8762 static int __build_sched_domains(const struct cpumask *cpu_map,
8763 struct sched_domain_attr *attr)
8765 enum s_alloc alloc_state = sa_none;
8767 struct sched_domain *sd;
8773 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8774 if (alloc_state != sa_rootdomain)
8776 alloc_state = sa_sched_groups;
8779 * Set up domains for cpus specified by the cpu_map.
8781 for_each_cpu(i, cpu_map) {
8782 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8785 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8786 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8787 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8788 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8791 for_each_cpu(i, cpu_map) {
8792 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8793 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8796 /* Set up physical groups */
8797 for (i = 0; i < nr_node_ids; i++)
8798 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8801 /* Set up node groups */
8803 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8805 for (i = 0; i < nr_node_ids; i++)
8806 if (build_numa_sched_groups(&d, cpu_map, i))
8810 /* Calculate CPU power for physical packages and nodes */
8811 #ifdef CONFIG_SCHED_SMT
8812 for_each_cpu(i, cpu_map) {
8813 sd = &per_cpu(cpu_domains, i).sd;
8814 init_sched_groups_power(i, sd);
8817 #ifdef CONFIG_SCHED_MC
8818 for_each_cpu(i, cpu_map) {
8819 sd = &per_cpu(core_domains, i).sd;
8820 init_sched_groups_power(i, sd);
8824 for_each_cpu(i, cpu_map) {
8825 sd = &per_cpu(phys_domains, i).sd;
8826 init_sched_groups_power(i, sd);
8830 for (i = 0; i < nr_node_ids; i++)
8831 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8833 if (d.sd_allnodes) {
8834 struct sched_group *sg;
8836 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8838 init_numa_sched_groups_power(sg);
8842 /* Attach the domains */
8843 for_each_cpu(i, cpu_map) {
8844 #ifdef CONFIG_SCHED_SMT
8845 sd = &per_cpu(cpu_domains, i).sd;
8846 #elif defined(CONFIG_SCHED_MC)
8847 sd = &per_cpu(core_domains, i).sd;
8849 sd = &per_cpu(phys_domains, i).sd;
8851 cpu_attach_domain(sd, d.rd, i);
8854 d.sched_group_nodes = NULL; /* don't free this we still need it */
8855 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8859 __free_domain_allocs(&d, alloc_state, cpu_map);
8863 static int build_sched_domains(const struct cpumask *cpu_map)
8865 return __build_sched_domains(cpu_map, NULL);
8868 static cpumask_var_t *doms_cur; /* current sched domains */
8869 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8870 static struct sched_domain_attr *dattr_cur;
8871 /* attribues of custom domains in 'doms_cur' */
8874 * Special case: If a kmalloc of a doms_cur partition (array of
8875 * cpumask) fails, then fallback to a single sched domain,
8876 * as determined by the single cpumask fallback_doms.
8878 static cpumask_var_t fallback_doms;
8881 * arch_update_cpu_topology lets virtualized architectures update the
8882 * cpu core maps. It is supposed to return 1 if the topology changed
8883 * or 0 if it stayed the same.
8885 int __attribute__((weak)) arch_update_cpu_topology(void)
8890 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
8893 cpumask_var_t *doms;
8895 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
8898 for (i = 0; i < ndoms; i++) {
8899 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
8900 free_sched_domains(doms, i);
8907 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
8910 for (i = 0; i < ndoms; i++)
8911 free_cpumask_var(doms[i]);
8916 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8917 * For now this just excludes isolated cpus, but could be used to
8918 * exclude other special cases in the future.
8920 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8924 arch_update_cpu_topology();
8926 doms_cur = alloc_sched_domains(ndoms_cur);
8928 doms_cur = &fallback_doms;
8929 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
8931 err = build_sched_domains(doms_cur[0]);
8932 register_sched_domain_sysctl();
8937 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8938 struct cpumask *tmpmask)
8940 free_sched_groups(cpu_map, tmpmask);
8944 * Detach sched domains from a group of cpus specified in cpu_map
8945 * These cpus will now be attached to the NULL domain
8947 static void detach_destroy_domains(const struct cpumask *cpu_map)
8949 /* Save because hotplug lock held. */
8950 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8953 for_each_cpu(i, cpu_map)
8954 cpu_attach_domain(NULL, &def_root_domain, i);
8955 synchronize_sched();
8956 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8959 /* handle null as "default" */
8960 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8961 struct sched_domain_attr *new, int idx_new)
8963 struct sched_domain_attr tmp;
8970 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8971 new ? (new + idx_new) : &tmp,
8972 sizeof(struct sched_domain_attr));
8976 * Partition sched domains as specified by the 'ndoms_new'
8977 * cpumasks in the array doms_new[] of cpumasks. This compares
8978 * doms_new[] to the current sched domain partitioning, doms_cur[].
8979 * It destroys each deleted domain and builds each new domain.
8981 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
8982 * The masks don't intersect (don't overlap.) We should setup one
8983 * sched domain for each mask. CPUs not in any of the cpumasks will
8984 * not be load balanced. If the same cpumask appears both in the
8985 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8988 * The passed in 'doms_new' should be allocated using
8989 * alloc_sched_domains. This routine takes ownership of it and will
8990 * free_sched_domains it when done with it. If the caller failed the
8991 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
8992 * and partition_sched_domains() will fallback to the single partition
8993 * 'fallback_doms', it also forces the domains to be rebuilt.
8995 * If doms_new == NULL it will be replaced with cpu_online_mask.
8996 * ndoms_new == 0 is a special case for destroying existing domains,
8997 * and it will not create the default domain.
8999 * Call with hotplug lock held
9001 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
9002 struct sched_domain_attr *dattr_new)
9007 mutex_lock(&sched_domains_mutex);
9009 /* always unregister in case we don't destroy any domains */
9010 unregister_sched_domain_sysctl();
9012 /* Let architecture update cpu core mappings. */
9013 new_topology = arch_update_cpu_topology();
9015 n = doms_new ? ndoms_new : 0;
9017 /* Destroy deleted domains */
9018 for (i = 0; i < ndoms_cur; i++) {
9019 for (j = 0; j < n && !new_topology; j++) {
9020 if (cpumask_equal(doms_cur[i], doms_new[j])
9021 && dattrs_equal(dattr_cur, i, dattr_new, j))
9024 /* no match - a current sched domain not in new doms_new[] */
9025 detach_destroy_domains(doms_cur[i]);
9030 if (doms_new == NULL) {
9032 doms_new = &fallback_doms;
9033 cpumask_andnot(doms_new[0], cpu_online_mask, cpu_isolated_map);
9034 WARN_ON_ONCE(dattr_new);
9037 /* Build new domains */
9038 for (i = 0; i < ndoms_new; i++) {
9039 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9040 if (cpumask_equal(doms_new[i], doms_cur[j])
9041 && dattrs_equal(dattr_new, i, dattr_cur, j))
9044 /* no match - add a new doms_new */
9045 __build_sched_domains(doms_new[i],
9046 dattr_new ? dattr_new + i : NULL);
9051 /* Remember the new sched domains */
9052 if (doms_cur != &fallback_doms)
9053 free_sched_domains(doms_cur, ndoms_cur);
9054 kfree(dattr_cur); /* kfree(NULL) is safe */
9055 doms_cur = doms_new;
9056 dattr_cur = dattr_new;
9057 ndoms_cur = ndoms_new;
9059 register_sched_domain_sysctl();
9061 mutex_unlock(&sched_domains_mutex);
9064 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9065 static void arch_reinit_sched_domains(void)
9069 /* Destroy domains first to force the rebuild */
9070 partition_sched_domains(0, NULL, NULL);
9072 rebuild_sched_domains();
9076 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9078 unsigned int level = 0;
9080 if (sscanf(buf, "%u", &level) != 1)
9084 * level is always be positive so don't check for
9085 * level < POWERSAVINGS_BALANCE_NONE which is 0
9086 * What happens on 0 or 1 byte write,
9087 * need to check for count as well?
9090 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9094 sched_smt_power_savings = level;
9096 sched_mc_power_savings = level;
9098 arch_reinit_sched_domains();
9103 #ifdef CONFIG_SCHED_MC
9104 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9107 return sprintf(page, "%u\n", sched_mc_power_savings);
9109 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9110 const char *buf, size_t count)
9112 return sched_power_savings_store(buf, count, 0);
9114 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9115 sched_mc_power_savings_show,
9116 sched_mc_power_savings_store);
9119 #ifdef CONFIG_SCHED_SMT
9120 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9123 return sprintf(page, "%u\n", sched_smt_power_savings);
9125 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9126 const char *buf, size_t count)
9128 return sched_power_savings_store(buf, count, 1);
9130 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9131 sched_smt_power_savings_show,
9132 sched_smt_power_savings_store);
9135 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9139 #ifdef CONFIG_SCHED_SMT
9141 err = sysfs_create_file(&cls->kset.kobj,
9142 &attr_sched_smt_power_savings.attr);
9144 #ifdef CONFIG_SCHED_MC
9145 if (!err && mc_capable())
9146 err = sysfs_create_file(&cls->kset.kobj,
9147 &attr_sched_mc_power_savings.attr);
9151 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9153 #ifndef CONFIG_CPUSETS
9155 * Add online and remove offline CPUs from the scheduler domains.
9156 * When cpusets are enabled they take over this function.
9158 static int update_sched_domains(struct notifier_block *nfb,
9159 unsigned long action, void *hcpu)
9163 case CPU_ONLINE_FROZEN:
9165 case CPU_DEAD_FROZEN:
9166 partition_sched_domains(1, NULL, NULL);
9175 static int update_runtime(struct notifier_block *nfb,
9176 unsigned long action, void *hcpu)
9178 int cpu = (int)(long)hcpu;
9181 case CPU_DOWN_PREPARE:
9182 case CPU_DOWN_PREPARE_FROZEN:
9183 disable_runtime(cpu_rq(cpu));
9186 case CPU_DOWN_FAILED:
9187 case CPU_DOWN_FAILED_FROZEN:
9189 case CPU_ONLINE_FROZEN:
9190 enable_runtime(cpu_rq(cpu));
9198 void __init sched_init_smp(void)
9200 cpumask_var_t non_isolated_cpus;
9202 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9203 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9205 #if defined(CONFIG_NUMA)
9206 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9208 BUG_ON(sched_group_nodes_bycpu == NULL);
9211 mutex_lock(&sched_domains_mutex);
9212 arch_init_sched_domains(cpu_online_mask);
9213 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9214 if (cpumask_empty(non_isolated_cpus))
9215 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9216 mutex_unlock(&sched_domains_mutex);
9219 #ifndef CONFIG_CPUSETS
9220 /* XXX: Theoretical race here - CPU may be hotplugged now */
9221 hotcpu_notifier(update_sched_domains, 0);
9224 /* RT runtime code needs to handle some hotplug events */
9225 hotcpu_notifier(update_runtime, 0);
9229 /* Move init over to a non-isolated CPU */
9230 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9232 sched_init_granularity();
9233 free_cpumask_var(non_isolated_cpus);
9235 init_sched_rt_class();
9238 void __init sched_init_smp(void)
9240 sched_init_granularity();
9242 #endif /* CONFIG_SMP */
9244 const_debug unsigned int sysctl_timer_migration = 1;
9246 int in_sched_functions(unsigned long addr)
9248 return in_lock_functions(addr) ||
9249 (addr >= (unsigned long)__sched_text_start
9250 && addr < (unsigned long)__sched_text_end);
9253 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9255 cfs_rq->tasks_timeline = RB_ROOT;
9256 INIT_LIST_HEAD(&cfs_rq->tasks);
9257 #ifdef CONFIG_FAIR_GROUP_SCHED
9260 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9263 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9265 struct rt_prio_array *array;
9268 array = &rt_rq->active;
9269 for (i = 0; i < MAX_RT_PRIO; i++) {
9270 INIT_LIST_HEAD(array->queue + i);
9271 __clear_bit(i, array->bitmap);
9273 /* delimiter for bitsearch: */
9274 __set_bit(MAX_RT_PRIO, array->bitmap);
9276 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9277 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9279 rt_rq->highest_prio.next = MAX_RT_PRIO;
9283 rt_rq->rt_nr_migratory = 0;
9284 rt_rq->overloaded = 0;
9285 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9289 rt_rq->rt_throttled = 0;
9290 rt_rq->rt_runtime = 0;
9291 spin_lock_init(&rt_rq->rt_runtime_lock);
9293 #ifdef CONFIG_RT_GROUP_SCHED
9294 rt_rq->rt_nr_boosted = 0;
9299 #ifdef CONFIG_FAIR_GROUP_SCHED
9300 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9301 struct sched_entity *se, int cpu, int add,
9302 struct sched_entity *parent)
9304 struct rq *rq = cpu_rq(cpu);
9305 tg->cfs_rq[cpu] = cfs_rq;
9306 init_cfs_rq(cfs_rq, rq);
9309 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9312 /* se could be NULL for init_task_group */
9317 se->cfs_rq = &rq->cfs;
9319 se->cfs_rq = parent->my_q;
9322 se->load.weight = tg->shares;
9323 se->load.inv_weight = 0;
9324 se->parent = parent;
9328 #ifdef CONFIG_RT_GROUP_SCHED
9329 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9330 struct sched_rt_entity *rt_se, int cpu, int add,
9331 struct sched_rt_entity *parent)
9333 struct rq *rq = cpu_rq(cpu);
9335 tg->rt_rq[cpu] = rt_rq;
9336 init_rt_rq(rt_rq, rq);
9338 rt_rq->rt_se = rt_se;
9339 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9341 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9343 tg->rt_se[cpu] = rt_se;
9348 rt_se->rt_rq = &rq->rt;
9350 rt_se->rt_rq = parent->my_q;
9352 rt_se->my_q = rt_rq;
9353 rt_se->parent = parent;
9354 INIT_LIST_HEAD(&rt_se->run_list);
9358 void __init sched_init(void)
9361 unsigned long alloc_size = 0, ptr;
9363 #ifdef CONFIG_FAIR_GROUP_SCHED
9364 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9366 #ifdef CONFIG_RT_GROUP_SCHED
9367 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9369 #ifdef CONFIG_USER_SCHED
9372 #ifdef CONFIG_CPUMASK_OFFSTACK
9373 alloc_size += num_possible_cpus() * cpumask_size();
9376 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9378 #ifdef CONFIG_FAIR_GROUP_SCHED
9379 init_task_group.se = (struct sched_entity **)ptr;
9380 ptr += nr_cpu_ids * sizeof(void **);
9382 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9383 ptr += nr_cpu_ids * sizeof(void **);
9385 #ifdef CONFIG_USER_SCHED
9386 root_task_group.se = (struct sched_entity **)ptr;
9387 ptr += nr_cpu_ids * sizeof(void **);
9389 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9390 ptr += nr_cpu_ids * sizeof(void **);
9391 #endif /* CONFIG_USER_SCHED */
9392 #endif /* CONFIG_FAIR_GROUP_SCHED */
9393 #ifdef CONFIG_RT_GROUP_SCHED
9394 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9395 ptr += nr_cpu_ids * sizeof(void **);
9397 init_task_group.rt_rq = (struct rt_rq **)ptr;
9398 ptr += nr_cpu_ids * sizeof(void **);
9400 #ifdef CONFIG_USER_SCHED
9401 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9402 ptr += nr_cpu_ids * sizeof(void **);
9404 root_task_group.rt_rq = (struct rt_rq **)ptr;
9405 ptr += nr_cpu_ids * sizeof(void **);
9406 #endif /* CONFIG_USER_SCHED */
9407 #endif /* CONFIG_RT_GROUP_SCHED */
9408 #ifdef CONFIG_CPUMASK_OFFSTACK
9409 for_each_possible_cpu(i) {
9410 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9411 ptr += cpumask_size();
9413 #endif /* CONFIG_CPUMASK_OFFSTACK */
9417 init_defrootdomain();
9420 init_rt_bandwidth(&def_rt_bandwidth,
9421 global_rt_period(), global_rt_runtime());
9423 #ifdef CONFIG_RT_GROUP_SCHED
9424 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9425 global_rt_period(), global_rt_runtime());
9426 #ifdef CONFIG_USER_SCHED
9427 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9428 global_rt_period(), RUNTIME_INF);
9429 #endif /* CONFIG_USER_SCHED */
9430 #endif /* CONFIG_RT_GROUP_SCHED */
9432 #ifdef CONFIG_GROUP_SCHED
9433 list_add(&init_task_group.list, &task_groups);
9434 INIT_LIST_HEAD(&init_task_group.children);
9436 #ifdef CONFIG_USER_SCHED
9437 INIT_LIST_HEAD(&root_task_group.children);
9438 init_task_group.parent = &root_task_group;
9439 list_add(&init_task_group.siblings, &root_task_group.children);
9440 #endif /* CONFIG_USER_SCHED */
9441 #endif /* CONFIG_GROUP_SCHED */
9443 for_each_possible_cpu(i) {
9447 spin_lock_init(&rq->lock);
9449 rq->calc_load_active = 0;
9450 rq->calc_load_update = jiffies + LOAD_FREQ;
9451 init_cfs_rq(&rq->cfs, rq);
9452 init_rt_rq(&rq->rt, rq);
9453 #ifdef CONFIG_FAIR_GROUP_SCHED
9454 init_task_group.shares = init_task_group_load;
9455 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9456 #ifdef CONFIG_CGROUP_SCHED
9458 * How much cpu bandwidth does init_task_group get?
9460 * In case of task-groups formed thr' the cgroup filesystem, it
9461 * gets 100% of the cpu resources in the system. This overall
9462 * system cpu resource is divided among the tasks of
9463 * init_task_group and its child task-groups in a fair manner,
9464 * based on each entity's (task or task-group's) weight
9465 * (se->load.weight).
9467 * In other words, if init_task_group has 10 tasks of weight
9468 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9469 * then A0's share of the cpu resource is:
9471 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9473 * We achieve this by letting init_task_group's tasks sit
9474 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9476 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9477 #elif defined CONFIG_USER_SCHED
9478 root_task_group.shares = NICE_0_LOAD;
9479 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9481 * In case of task-groups formed thr' the user id of tasks,
9482 * init_task_group represents tasks belonging to root user.
9483 * Hence it forms a sibling of all subsequent groups formed.
9484 * In this case, init_task_group gets only a fraction of overall
9485 * system cpu resource, based on the weight assigned to root
9486 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9487 * by letting tasks of init_task_group sit in a separate cfs_rq
9488 * (init_tg_cfs_rq) and having one entity represent this group of
9489 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9491 init_tg_cfs_entry(&init_task_group,
9492 &per_cpu(init_tg_cfs_rq, i),
9493 &per_cpu(init_sched_entity, i), i, 1,
9494 root_task_group.se[i]);
9497 #endif /* CONFIG_FAIR_GROUP_SCHED */
9499 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9500 #ifdef CONFIG_RT_GROUP_SCHED
9501 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9502 #ifdef CONFIG_CGROUP_SCHED
9503 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9504 #elif defined CONFIG_USER_SCHED
9505 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9506 init_tg_rt_entry(&init_task_group,
9507 &per_cpu(init_rt_rq, i),
9508 &per_cpu(init_sched_rt_entity, i), i, 1,
9509 root_task_group.rt_se[i]);
9513 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9514 rq->cpu_load[j] = 0;
9518 rq->post_schedule = 0;
9519 rq->active_balance = 0;
9520 rq->next_balance = jiffies;
9524 rq->migration_thread = NULL;
9525 INIT_LIST_HEAD(&rq->migration_queue);
9526 rq_attach_root(rq, &def_root_domain);
9529 atomic_set(&rq->nr_iowait, 0);
9532 set_load_weight(&init_task);
9534 #ifdef CONFIG_PREEMPT_NOTIFIERS
9535 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9539 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9542 #ifdef CONFIG_RT_MUTEXES
9543 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9547 * The boot idle thread does lazy MMU switching as well:
9549 atomic_inc(&init_mm.mm_count);
9550 enter_lazy_tlb(&init_mm, current);
9553 * Make us the idle thread. Technically, schedule() should not be
9554 * called from this thread, however somewhere below it might be,
9555 * but because we are the idle thread, we just pick up running again
9556 * when this runqueue becomes "idle".
9558 init_idle(current, smp_processor_id());
9560 calc_load_update = jiffies + LOAD_FREQ;
9563 * During early bootup we pretend to be a normal task:
9565 current->sched_class = &fair_sched_class;
9567 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9568 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9571 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9572 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9574 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9579 scheduler_running = 1;
9582 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9583 static inline int preempt_count_equals(int preempt_offset)
9585 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9587 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9590 void __might_sleep(char *file, int line, int preempt_offset)
9593 static unsigned long prev_jiffy; /* ratelimiting */
9595 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9596 system_state != SYSTEM_RUNNING || oops_in_progress)
9598 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9600 prev_jiffy = jiffies;
9603 "BUG: sleeping function called from invalid context at %s:%d\n",
9606 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9607 in_atomic(), irqs_disabled(),
9608 current->pid, current->comm);
9610 debug_show_held_locks(current);
9611 if (irqs_disabled())
9612 print_irqtrace_events(current);
9616 EXPORT_SYMBOL(__might_sleep);
9619 #ifdef CONFIG_MAGIC_SYSRQ
9620 static void normalize_task(struct rq *rq, struct task_struct *p)
9624 update_rq_clock(rq);
9625 on_rq = p->se.on_rq;
9627 deactivate_task(rq, p, 0);
9628 __setscheduler(rq, p, SCHED_NORMAL, 0);
9630 activate_task(rq, p, 0);
9631 resched_task(rq->curr);
9635 void normalize_rt_tasks(void)
9637 struct task_struct *g, *p;
9638 unsigned long flags;
9641 read_lock_irqsave(&tasklist_lock, flags);
9642 do_each_thread(g, p) {
9644 * Only normalize user tasks:
9649 p->se.exec_start = 0;
9650 #ifdef CONFIG_SCHEDSTATS
9651 p->se.wait_start = 0;
9652 p->se.sleep_start = 0;
9653 p->se.block_start = 0;
9658 * Renice negative nice level userspace
9661 if (TASK_NICE(p) < 0 && p->mm)
9662 set_user_nice(p, 0);
9666 spin_lock(&p->pi_lock);
9667 rq = __task_rq_lock(p);
9669 normalize_task(rq, p);
9671 __task_rq_unlock(rq);
9672 spin_unlock(&p->pi_lock);
9673 } while_each_thread(g, p);
9675 read_unlock_irqrestore(&tasklist_lock, flags);
9678 #endif /* CONFIG_MAGIC_SYSRQ */
9682 * These functions are only useful for the IA64 MCA handling.
9684 * They can only be called when the whole system has been
9685 * stopped - every CPU needs to be quiescent, and no scheduling
9686 * activity can take place. Using them for anything else would
9687 * be a serious bug, and as a result, they aren't even visible
9688 * under any other configuration.
9692 * curr_task - return the current task for a given cpu.
9693 * @cpu: the processor in question.
9695 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9697 struct task_struct *curr_task(int cpu)
9699 return cpu_curr(cpu);
9703 * set_curr_task - set the current task for a given cpu.
9704 * @cpu: the processor in question.
9705 * @p: the task pointer to set.
9707 * Description: This function must only be used when non-maskable interrupts
9708 * are serviced on a separate stack. It allows the architecture to switch the
9709 * notion of the current task on a cpu in a non-blocking manner. This function
9710 * must be called with all CPU's synchronized, and interrupts disabled, the
9711 * and caller must save the original value of the current task (see
9712 * curr_task() above) and restore that value before reenabling interrupts and
9713 * re-starting the system.
9715 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9717 void set_curr_task(int cpu, struct task_struct *p)
9724 #ifdef CONFIG_FAIR_GROUP_SCHED
9725 static void free_fair_sched_group(struct task_group *tg)
9729 for_each_possible_cpu(i) {
9731 kfree(tg->cfs_rq[i]);
9741 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9743 struct cfs_rq *cfs_rq;
9744 struct sched_entity *se;
9748 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9751 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9755 tg->shares = NICE_0_LOAD;
9757 for_each_possible_cpu(i) {
9760 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9761 GFP_KERNEL, cpu_to_node(i));
9765 se = kzalloc_node(sizeof(struct sched_entity),
9766 GFP_KERNEL, cpu_to_node(i));
9770 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9779 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9781 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9782 &cpu_rq(cpu)->leaf_cfs_rq_list);
9785 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9787 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9789 #else /* !CONFG_FAIR_GROUP_SCHED */
9790 static inline void free_fair_sched_group(struct task_group *tg)
9795 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9800 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9804 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9807 #endif /* CONFIG_FAIR_GROUP_SCHED */
9809 #ifdef CONFIG_RT_GROUP_SCHED
9810 static void free_rt_sched_group(struct task_group *tg)
9814 destroy_rt_bandwidth(&tg->rt_bandwidth);
9816 for_each_possible_cpu(i) {
9818 kfree(tg->rt_rq[i]);
9820 kfree(tg->rt_se[i]);
9828 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9830 struct rt_rq *rt_rq;
9831 struct sched_rt_entity *rt_se;
9835 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9838 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9842 init_rt_bandwidth(&tg->rt_bandwidth,
9843 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9845 for_each_possible_cpu(i) {
9848 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9849 GFP_KERNEL, cpu_to_node(i));
9853 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9854 GFP_KERNEL, cpu_to_node(i));
9858 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9867 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9869 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9870 &cpu_rq(cpu)->leaf_rt_rq_list);
9873 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9875 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9877 #else /* !CONFIG_RT_GROUP_SCHED */
9878 static inline void free_rt_sched_group(struct task_group *tg)
9883 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9888 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9892 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9895 #endif /* CONFIG_RT_GROUP_SCHED */
9897 #ifdef CONFIG_GROUP_SCHED
9898 static void free_sched_group(struct task_group *tg)
9900 free_fair_sched_group(tg);
9901 free_rt_sched_group(tg);
9905 /* allocate runqueue etc for a new task group */
9906 struct task_group *sched_create_group(struct task_group *parent)
9908 struct task_group *tg;
9909 unsigned long flags;
9912 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9914 return ERR_PTR(-ENOMEM);
9916 if (!alloc_fair_sched_group(tg, parent))
9919 if (!alloc_rt_sched_group(tg, parent))
9922 spin_lock_irqsave(&task_group_lock, flags);
9923 for_each_possible_cpu(i) {
9924 register_fair_sched_group(tg, i);
9925 register_rt_sched_group(tg, i);
9927 list_add_rcu(&tg->list, &task_groups);
9929 WARN_ON(!parent); /* root should already exist */
9931 tg->parent = parent;
9932 INIT_LIST_HEAD(&tg->children);
9933 list_add_rcu(&tg->siblings, &parent->children);
9934 spin_unlock_irqrestore(&task_group_lock, flags);
9939 free_sched_group(tg);
9940 return ERR_PTR(-ENOMEM);
9943 /* rcu callback to free various structures associated with a task group */
9944 static void free_sched_group_rcu(struct rcu_head *rhp)
9946 /* now it should be safe to free those cfs_rqs */
9947 free_sched_group(container_of(rhp, struct task_group, rcu));
9950 /* Destroy runqueue etc associated with a task group */
9951 void sched_destroy_group(struct task_group *tg)
9953 unsigned long flags;
9956 spin_lock_irqsave(&task_group_lock, flags);
9957 for_each_possible_cpu(i) {
9958 unregister_fair_sched_group(tg, i);
9959 unregister_rt_sched_group(tg, i);
9961 list_del_rcu(&tg->list);
9962 list_del_rcu(&tg->siblings);
9963 spin_unlock_irqrestore(&task_group_lock, flags);
9965 /* wait for possible concurrent references to cfs_rqs complete */
9966 call_rcu(&tg->rcu, free_sched_group_rcu);
9969 /* change task's runqueue when it moves between groups.
9970 * The caller of this function should have put the task in its new group
9971 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9972 * reflect its new group.
9974 void sched_move_task(struct task_struct *tsk)
9977 unsigned long flags;
9980 rq = task_rq_lock(tsk, &flags);
9982 update_rq_clock(rq);
9984 running = task_current(rq, tsk);
9985 on_rq = tsk->se.on_rq;
9988 dequeue_task(rq, tsk, 0);
9989 if (unlikely(running))
9990 tsk->sched_class->put_prev_task(rq, tsk);
9992 set_task_rq(tsk, task_cpu(tsk));
9994 #ifdef CONFIG_FAIR_GROUP_SCHED
9995 if (tsk->sched_class->moved_group)
9996 tsk->sched_class->moved_group(tsk);
9999 if (unlikely(running))
10000 tsk->sched_class->set_curr_task(rq);
10002 enqueue_task(rq, tsk, 0);
10004 task_rq_unlock(rq, &flags);
10006 #endif /* CONFIG_GROUP_SCHED */
10008 #ifdef CONFIG_FAIR_GROUP_SCHED
10009 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10011 struct cfs_rq *cfs_rq = se->cfs_rq;
10016 dequeue_entity(cfs_rq, se, 0);
10018 se->load.weight = shares;
10019 se->load.inv_weight = 0;
10022 enqueue_entity(cfs_rq, se, 0);
10025 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10027 struct cfs_rq *cfs_rq = se->cfs_rq;
10028 struct rq *rq = cfs_rq->rq;
10029 unsigned long flags;
10031 spin_lock_irqsave(&rq->lock, flags);
10032 __set_se_shares(se, shares);
10033 spin_unlock_irqrestore(&rq->lock, flags);
10036 static DEFINE_MUTEX(shares_mutex);
10038 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10041 unsigned long flags;
10044 * We can't change the weight of the root cgroup.
10049 if (shares < MIN_SHARES)
10050 shares = MIN_SHARES;
10051 else if (shares > MAX_SHARES)
10052 shares = MAX_SHARES;
10054 mutex_lock(&shares_mutex);
10055 if (tg->shares == shares)
10058 spin_lock_irqsave(&task_group_lock, flags);
10059 for_each_possible_cpu(i)
10060 unregister_fair_sched_group(tg, i);
10061 list_del_rcu(&tg->siblings);
10062 spin_unlock_irqrestore(&task_group_lock, flags);
10064 /* wait for any ongoing reference to this group to finish */
10065 synchronize_sched();
10068 * Now we are free to modify the group's share on each cpu
10069 * w/o tripping rebalance_share or load_balance_fair.
10071 tg->shares = shares;
10072 for_each_possible_cpu(i) {
10074 * force a rebalance
10076 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10077 set_se_shares(tg->se[i], shares);
10081 * Enable load balance activity on this group, by inserting it back on
10082 * each cpu's rq->leaf_cfs_rq_list.
10084 spin_lock_irqsave(&task_group_lock, flags);
10085 for_each_possible_cpu(i)
10086 register_fair_sched_group(tg, i);
10087 list_add_rcu(&tg->siblings, &tg->parent->children);
10088 spin_unlock_irqrestore(&task_group_lock, flags);
10090 mutex_unlock(&shares_mutex);
10094 unsigned long sched_group_shares(struct task_group *tg)
10100 #ifdef CONFIG_RT_GROUP_SCHED
10102 * Ensure that the real time constraints are schedulable.
10104 static DEFINE_MUTEX(rt_constraints_mutex);
10106 static unsigned long to_ratio(u64 period, u64 runtime)
10108 if (runtime == RUNTIME_INF)
10111 return div64_u64(runtime << 20, period);
10114 /* Must be called with tasklist_lock held */
10115 static inline int tg_has_rt_tasks(struct task_group *tg)
10117 struct task_struct *g, *p;
10119 do_each_thread(g, p) {
10120 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10122 } while_each_thread(g, p);
10127 struct rt_schedulable_data {
10128 struct task_group *tg;
10133 static int tg_schedulable(struct task_group *tg, void *data)
10135 struct rt_schedulable_data *d = data;
10136 struct task_group *child;
10137 unsigned long total, sum = 0;
10138 u64 period, runtime;
10140 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10141 runtime = tg->rt_bandwidth.rt_runtime;
10144 period = d->rt_period;
10145 runtime = d->rt_runtime;
10148 #ifdef CONFIG_USER_SCHED
10149 if (tg == &root_task_group) {
10150 period = global_rt_period();
10151 runtime = global_rt_runtime();
10156 * Cannot have more runtime than the period.
10158 if (runtime > period && runtime != RUNTIME_INF)
10162 * Ensure we don't starve existing RT tasks.
10164 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10167 total = to_ratio(period, runtime);
10170 * Nobody can have more than the global setting allows.
10172 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10176 * The sum of our children's runtime should not exceed our own.
10178 list_for_each_entry_rcu(child, &tg->children, siblings) {
10179 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10180 runtime = child->rt_bandwidth.rt_runtime;
10182 if (child == d->tg) {
10183 period = d->rt_period;
10184 runtime = d->rt_runtime;
10187 sum += to_ratio(period, runtime);
10196 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10198 struct rt_schedulable_data data = {
10200 .rt_period = period,
10201 .rt_runtime = runtime,
10204 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10207 static int tg_set_bandwidth(struct task_group *tg,
10208 u64 rt_period, u64 rt_runtime)
10212 mutex_lock(&rt_constraints_mutex);
10213 read_lock(&tasklist_lock);
10214 err = __rt_schedulable(tg, rt_period, rt_runtime);
10218 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10219 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10220 tg->rt_bandwidth.rt_runtime = rt_runtime;
10222 for_each_possible_cpu(i) {
10223 struct rt_rq *rt_rq = tg->rt_rq[i];
10225 spin_lock(&rt_rq->rt_runtime_lock);
10226 rt_rq->rt_runtime = rt_runtime;
10227 spin_unlock(&rt_rq->rt_runtime_lock);
10229 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10231 read_unlock(&tasklist_lock);
10232 mutex_unlock(&rt_constraints_mutex);
10237 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10239 u64 rt_runtime, rt_period;
10241 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10242 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10243 if (rt_runtime_us < 0)
10244 rt_runtime = RUNTIME_INF;
10246 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10249 long sched_group_rt_runtime(struct task_group *tg)
10253 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10256 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10257 do_div(rt_runtime_us, NSEC_PER_USEC);
10258 return rt_runtime_us;
10261 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10263 u64 rt_runtime, rt_period;
10265 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10266 rt_runtime = tg->rt_bandwidth.rt_runtime;
10268 if (rt_period == 0)
10271 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10274 long sched_group_rt_period(struct task_group *tg)
10278 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10279 do_div(rt_period_us, NSEC_PER_USEC);
10280 return rt_period_us;
10283 static int sched_rt_global_constraints(void)
10285 u64 runtime, period;
10288 if (sysctl_sched_rt_period <= 0)
10291 runtime = global_rt_runtime();
10292 period = global_rt_period();
10295 * Sanity check on the sysctl variables.
10297 if (runtime > period && runtime != RUNTIME_INF)
10300 mutex_lock(&rt_constraints_mutex);
10301 read_lock(&tasklist_lock);
10302 ret = __rt_schedulable(NULL, 0, 0);
10303 read_unlock(&tasklist_lock);
10304 mutex_unlock(&rt_constraints_mutex);
10309 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10311 /* Don't accept realtime tasks when there is no way for them to run */
10312 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10318 #else /* !CONFIG_RT_GROUP_SCHED */
10319 static int sched_rt_global_constraints(void)
10321 unsigned long flags;
10324 if (sysctl_sched_rt_period <= 0)
10328 * There's always some RT tasks in the root group
10329 * -- migration, kstopmachine etc..
10331 if (sysctl_sched_rt_runtime == 0)
10334 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10335 for_each_possible_cpu(i) {
10336 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10338 spin_lock(&rt_rq->rt_runtime_lock);
10339 rt_rq->rt_runtime = global_rt_runtime();
10340 spin_unlock(&rt_rq->rt_runtime_lock);
10342 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10346 #endif /* CONFIG_RT_GROUP_SCHED */
10348 int sched_rt_handler(struct ctl_table *table, int write,
10349 void __user *buffer, size_t *lenp,
10353 int old_period, old_runtime;
10354 static DEFINE_MUTEX(mutex);
10356 mutex_lock(&mutex);
10357 old_period = sysctl_sched_rt_period;
10358 old_runtime = sysctl_sched_rt_runtime;
10360 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10362 if (!ret && write) {
10363 ret = sched_rt_global_constraints();
10365 sysctl_sched_rt_period = old_period;
10366 sysctl_sched_rt_runtime = old_runtime;
10368 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10369 def_rt_bandwidth.rt_period =
10370 ns_to_ktime(global_rt_period());
10373 mutex_unlock(&mutex);
10378 #ifdef CONFIG_CGROUP_SCHED
10380 /* return corresponding task_group object of a cgroup */
10381 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10383 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10384 struct task_group, css);
10387 static struct cgroup_subsys_state *
10388 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10390 struct task_group *tg, *parent;
10392 if (!cgrp->parent) {
10393 /* This is early initialization for the top cgroup */
10394 return &init_task_group.css;
10397 parent = cgroup_tg(cgrp->parent);
10398 tg = sched_create_group(parent);
10400 return ERR_PTR(-ENOMEM);
10406 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10408 struct task_group *tg = cgroup_tg(cgrp);
10410 sched_destroy_group(tg);
10414 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10416 #ifdef CONFIG_RT_GROUP_SCHED
10417 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10420 /* We don't support RT-tasks being in separate groups */
10421 if (tsk->sched_class != &fair_sched_class)
10428 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10429 struct task_struct *tsk, bool threadgroup)
10431 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10435 struct task_struct *c;
10437 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10438 retval = cpu_cgroup_can_attach_task(cgrp, c);
10450 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10451 struct cgroup *old_cont, struct task_struct *tsk,
10454 sched_move_task(tsk);
10456 struct task_struct *c;
10458 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10459 sched_move_task(c);
10465 #ifdef CONFIG_FAIR_GROUP_SCHED
10466 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10469 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10472 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10474 struct task_group *tg = cgroup_tg(cgrp);
10476 return (u64) tg->shares;
10478 #endif /* CONFIG_FAIR_GROUP_SCHED */
10480 #ifdef CONFIG_RT_GROUP_SCHED
10481 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10484 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10487 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10489 return sched_group_rt_runtime(cgroup_tg(cgrp));
10492 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10495 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10498 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10500 return sched_group_rt_period(cgroup_tg(cgrp));
10502 #endif /* CONFIG_RT_GROUP_SCHED */
10504 static struct cftype cpu_files[] = {
10505 #ifdef CONFIG_FAIR_GROUP_SCHED
10508 .read_u64 = cpu_shares_read_u64,
10509 .write_u64 = cpu_shares_write_u64,
10512 #ifdef CONFIG_RT_GROUP_SCHED
10514 .name = "rt_runtime_us",
10515 .read_s64 = cpu_rt_runtime_read,
10516 .write_s64 = cpu_rt_runtime_write,
10519 .name = "rt_period_us",
10520 .read_u64 = cpu_rt_period_read_uint,
10521 .write_u64 = cpu_rt_period_write_uint,
10526 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10528 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10531 struct cgroup_subsys cpu_cgroup_subsys = {
10533 .create = cpu_cgroup_create,
10534 .destroy = cpu_cgroup_destroy,
10535 .can_attach = cpu_cgroup_can_attach,
10536 .attach = cpu_cgroup_attach,
10537 .populate = cpu_cgroup_populate,
10538 .subsys_id = cpu_cgroup_subsys_id,
10542 #endif /* CONFIG_CGROUP_SCHED */
10544 #ifdef CONFIG_CGROUP_CPUACCT
10547 * CPU accounting code for task groups.
10549 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10550 * (balbir@in.ibm.com).
10553 /* track cpu usage of a group of tasks and its child groups */
10555 struct cgroup_subsys_state css;
10556 /* cpuusage holds pointer to a u64-type object on every cpu */
10558 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10559 struct cpuacct *parent;
10562 struct cgroup_subsys cpuacct_subsys;
10564 /* return cpu accounting group corresponding to this container */
10565 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10567 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10568 struct cpuacct, css);
10571 /* return cpu accounting group to which this task belongs */
10572 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10574 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10575 struct cpuacct, css);
10578 /* create a new cpu accounting group */
10579 static struct cgroup_subsys_state *cpuacct_create(
10580 struct cgroup_subsys *ss, struct cgroup *cgrp)
10582 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10588 ca->cpuusage = alloc_percpu(u64);
10592 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10593 if (percpu_counter_init(&ca->cpustat[i], 0))
10594 goto out_free_counters;
10597 ca->parent = cgroup_ca(cgrp->parent);
10603 percpu_counter_destroy(&ca->cpustat[i]);
10604 free_percpu(ca->cpuusage);
10608 return ERR_PTR(-ENOMEM);
10611 /* destroy an existing cpu accounting group */
10613 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10615 struct cpuacct *ca = cgroup_ca(cgrp);
10618 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10619 percpu_counter_destroy(&ca->cpustat[i]);
10620 free_percpu(ca->cpuusage);
10624 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10626 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10629 #ifndef CONFIG_64BIT
10631 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10633 spin_lock_irq(&cpu_rq(cpu)->lock);
10635 spin_unlock_irq(&cpu_rq(cpu)->lock);
10643 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10645 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10647 #ifndef CONFIG_64BIT
10649 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10651 spin_lock_irq(&cpu_rq(cpu)->lock);
10653 spin_unlock_irq(&cpu_rq(cpu)->lock);
10659 /* return total cpu usage (in nanoseconds) of a group */
10660 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10662 struct cpuacct *ca = cgroup_ca(cgrp);
10663 u64 totalcpuusage = 0;
10666 for_each_present_cpu(i)
10667 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10669 return totalcpuusage;
10672 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10675 struct cpuacct *ca = cgroup_ca(cgrp);
10684 for_each_present_cpu(i)
10685 cpuacct_cpuusage_write(ca, i, 0);
10691 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10692 struct seq_file *m)
10694 struct cpuacct *ca = cgroup_ca(cgroup);
10698 for_each_present_cpu(i) {
10699 percpu = cpuacct_cpuusage_read(ca, i);
10700 seq_printf(m, "%llu ", (unsigned long long) percpu);
10702 seq_printf(m, "\n");
10706 static const char *cpuacct_stat_desc[] = {
10707 [CPUACCT_STAT_USER] = "user",
10708 [CPUACCT_STAT_SYSTEM] = "system",
10711 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10712 struct cgroup_map_cb *cb)
10714 struct cpuacct *ca = cgroup_ca(cgrp);
10717 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10718 s64 val = percpu_counter_read(&ca->cpustat[i]);
10719 val = cputime64_to_clock_t(val);
10720 cb->fill(cb, cpuacct_stat_desc[i], val);
10725 static struct cftype files[] = {
10728 .read_u64 = cpuusage_read,
10729 .write_u64 = cpuusage_write,
10732 .name = "usage_percpu",
10733 .read_seq_string = cpuacct_percpu_seq_read,
10737 .read_map = cpuacct_stats_show,
10741 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10743 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10747 * charge this task's execution time to its accounting group.
10749 * called with rq->lock held.
10751 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10753 struct cpuacct *ca;
10756 if (unlikely(!cpuacct_subsys.active))
10759 cpu = task_cpu(tsk);
10765 for (; ca; ca = ca->parent) {
10766 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10767 *cpuusage += cputime;
10774 * Charge the system/user time to the task's accounting group.
10776 static void cpuacct_update_stats(struct task_struct *tsk,
10777 enum cpuacct_stat_index idx, cputime_t val)
10779 struct cpuacct *ca;
10781 if (unlikely(!cpuacct_subsys.active))
10788 percpu_counter_add(&ca->cpustat[idx], val);
10794 struct cgroup_subsys cpuacct_subsys = {
10796 .create = cpuacct_create,
10797 .destroy = cpuacct_destroy,
10798 .populate = cpuacct_populate,
10799 .subsys_id = cpuacct_subsys_id,
10801 #endif /* CONFIG_CGROUP_CPUACCT */
10805 int rcu_expedited_torture_stats(char *page)
10809 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10811 void synchronize_sched_expedited(void)
10814 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10816 #else /* #ifndef CONFIG_SMP */
10818 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10819 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10821 #define RCU_EXPEDITED_STATE_POST -2
10822 #define RCU_EXPEDITED_STATE_IDLE -1
10824 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10826 int rcu_expedited_torture_stats(char *page)
10831 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10832 for_each_online_cpu(cpu) {
10833 cnt += sprintf(&page[cnt], " %d:%d",
10834 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10836 cnt += sprintf(&page[cnt], "\n");
10839 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10841 static long synchronize_sched_expedited_count;
10844 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10845 * approach to force grace period to end quickly. This consumes
10846 * significant time on all CPUs, and is thus not recommended for
10847 * any sort of common-case code.
10849 * Note that it is illegal to call this function while holding any
10850 * lock that is acquired by a CPU-hotplug notifier. Failing to
10851 * observe this restriction will result in deadlock.
10853 void synchronize_sched_expedited(void)
10856 unsigned long flags;
10857 bool need_full_sync = 0;
10859 struct migration_req *req;
10863 smp_mb(); /* ensure prior mod happens before capturing snap. */
10864 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10866 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10868 if (trycount++ < 10)
10869 udelay(trycount * num_online_cpus());
10871 synchronize_sched();
10874 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10875 smp_mb(); /* ensure test happens before caller kfree */
10880 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10881 for_each_online_cpu(cpu) {
10883 req = &per_cpu(rcu_migration_req, cpu);
10884 init_completion(&req->done);
10886 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10887 spin_lock_irqsave(&rq->lock, flags);
10888 list_add(&req->list, &rq->migration_queue);
10889 spin_unlock_irqrestore(&rq->lock, flags);
10890 wake_up_process(rq->migration_thread);
10892 for_each_online_cpu(cpu) {
10893 rcu_expedited_state = cpu;
10894 req = &per_cpu(rcu_migration_req, cpu);
10896 wait_for_completion(&req->done);
10897 spin_lock_irqsave(&rq->lock, flags);
10898 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10899 need_full_sync = 1;
10900 req->dest_cpu = RCU_MIGRATION_IDLE;
10901 spin_unlock_irqrestore(&rq->lock, flags);
10903 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10904 mutex_unlock(&rcu_sched_expedited_mutex);
10906 if (need_full_sync)
10907 synchronize_sched();
10909 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10911 #endif /* #else #ifndef CONFIG_SMP */