4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy)
124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
129 static inline int task_has_rt_policy(struct task_struct *p)
131 return rt_policy(p->policy);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
142 struct rt_bandwidth {
143 /* nests inside the rq lock: */
144 raw_spinlock_t rt_runtime_lock;
147 struct hrtimer rt_period_timer;
150 static struct rt_bandwidth def_rt_bandwidth;
152 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
154 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169 idle = do_sched_rt_period_timer(rt_b, overrun);
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
176 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
181 raw_spin_lock_init(&rt_b->rt_runtime_lock);
183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime >= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
200 if (hrtimer_active(&rt_b->rt_period_timer))
203 raw_spin_lock(&rt_b->rt_runtime_lock);
208 if (hrtimer_active(&rt_b->rt_period_timer))
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
218 HRTIMER_MODE_ABS_PINNED, 0);
220 raw_spin_unlock(&rt_b->rt_runtime_lock);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
226 hrtimer_cancel(&rt_b->rt_period_timer);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex);
236 #ifdef CONFIG_GROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups);
244 /* task group related information */
246 #ifdef CONFIG_CGROUP_SCHED
247 struct cgroup_subsys_state css;
250 #ifdef CONFIG_USER_SCHED
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
266 struct rt_bandwidth rt_bandwidth;
270 struct list_head list;
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
277 #ifdef CONFIG_USER_SCHED
279 /* Helper function to pass uid information to create_sched_user() */
280 void set_tg_uid(struct user_struct *user)
282 user->tg->uid = user->uid;
287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
290 struct task_group root_task_group;
292 #ifdef CONFIG_FAIR_GROUP_SCHED
293 /* Default task group's sched entity on each cpu */
294 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
295 /* Default task group's cfs_rq on each cpu */
296 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq, init_tg_cfs_rq);
297 #endif /* CONFIG_FAIR_GROUP_SCHED */
299 #ifdef CONFIG_RT_GROUP_SCHED
300 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
301 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq, init_rt_rq_var);
302 #endif /* CONFIG_RT_GROUP_SCHED */
303 #else /* !CONFIG_USER_SCHED */
304 #define root_task_group init_task_group
305 #endif /* CONFIG_USER_SCHED */
307 /* task_group_lock serializes add/remove of task groups and also changes to
308 * a task group's cpu shares.
310 static DEFINE_SPINLOCK(task_group_lock);
312 #ifdef CONFIG_FAIR_GROUP_SCHED
315 static int root_task_group_empty(void)
317 return list_empty(&root_task_group.children);
321 #ifdef CONFIG_USER_SCHED
322 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
323 #else /* !CONFIG_USER_SCHED */
324 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
325 #endif /* CONFIG_USER_SCHED */
328 * A weight of 0 or 1 can cause arithmetics problems.
329 * A weight of a cfs_rq is the sum of weights of which entities
330 * are queued on this cfs_rq, so a weight of a entity should not be
331 * too large, so as the shares value of a task group.
332 * (The default weight is 1024 - so there's no practical
333 * limitation from this.)
336 #define MAX_SHARES (1UL << 18)
338 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
341 /* Default task group.
342 * Every task in system belong to this group at bootup.
344 struct task_group init_task_group;
346 /* return group to which a task belongs */
347 static inline struct task_group *task_group(struct task_struct *p)
349 struct task_group *tg;
351 #ifdef CONFIG_USER_SCHED
353 tg = __task_cred(p)->user->tg;
355 #elif defined(CONFIG_CGROUP_SCHED)
356 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
357 struct task_group, css);
359 tg = &init_task_group;
364 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
365 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
367 #ifdef CONFIG_FAIR_GROUP_SCHED
368 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
369 p->se.parent = task_group(p)->se[cpu];
372 #ifdef CONFIG_RT_GROUP_SCHED
373 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
374 p->rt.parent = task_group(p)->rt_se[cpu];
380 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
381 static inline struct task_group *task_group(struct task_struct *p)
386 #endif /* CONFIG_GROUP_SCHED */
388 /* CFS-related fields in a runqueue */
390 struct load_weight load;
391 unsigned long nr_running;
396 struct rb_root tasks_timeline;
397 struct rb_node *rb_leftmost;
399 struct list_head tasks;
400 struct list_head *balance_iterator;
403 * 'curr' points to currently running entity on this cfs_rq.
404 * It is set to NULL otherwise (i.e when none are currently running).
406 struct sched_entity *curr, *next, *last;
408 unsigned int nr_spread_over;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
421 struct list_head leaf_cfs_rq_list;
422 struct task_group *tg; /* group that "owns" this runqueue */
426 * the part of load.weight contributed by tasks
428 unsigned long task_weight;
431 * h_load = weight * f(tg)
433 * Where f(tg) is the recursive weight fraction assigned to
436 unsigned long h_load;
439 * this cpu's part of tg->shares
441 unsigned long shares;
444 * load.weight at the time we set shares
446 unsigned long rq_weight;
451 /* Real-Time classes' related field in a runqueue: */
453 struct rt_prio_array active;
454 unsigned long rt_nr_running;
455 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
457 int curr; /* highest queued rt task prio */
459 int next; /* next highest */
464 unsigned long rt_nr_migratory;
465 unsigned long rt_nr_total;
467 struct plist_head pushable_tasks;
472 /* Nests inside the rq lock: */
473 raw_spinlock_t rt_runtime_lock;
475 #ifdef CONFIG_RT_GROUP_SCHED
476 unsigned long rt_nr_boosted;
479 struct list_head leaf_rt_rq_list;
480 struct task_group *tg;
481 struct sched_rt_entity *rt_se;
488 * We add the notion of a root-domain which will be used to define per-domain
489 * variables. Each exclusive cpuset essentially defines an island domain by
490 * fully partitioning the member cpus from any other cpuset. Whenever a new
491 * exclusive cpuset is created, we also create and attach a new root-domain
498 cpumask_var_t online;
501 * The "RT overload" flag: it gets set if a CPU has more than
502 * one runnable RT task.
504 cpumask_var_t rto_mask;
507 struct cpupri cpupri;
512 * By default the system creates a single root-domain with all cpus as
513 * members (mimicking the global state we have today).
515 static struct root_domain def_root_domain;
520 * This is the main, per-CPU runqueue data structure.
522 * Locking rule: those places that want to lock multiple runqueues
523 * (such as the load balancing or the thread migration code), lock
524 * acquire operations must be ordered by ascending &runqueue.
531 * nr_running and cpu_load should be in the same cacheline because
532 * remote CPUs use both these fields when doing load calculation.
534 unsigned long nr_running;
535 #define CPU_LOAD_IDX_MAX 5
536 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
538 unsigned char in_nohz_recently;
540 /* capture load from *all* tasks on this cpu: */
541 struct load_weight load;
542 unsigned long nr_load_updates;
548 #ifdef CONFIG_FAIR_GROUP_SCHED
549 /* list of leaf cfs_rq on this cpu: */
550 struct list_head leaf_cfs_rq_list;
552 #ifdef CONFIG_RT_GROUP_SCHED
553 struct list_head leaf_rt_rq_list;
557 * This is part of a global counter where only the total sum
558 * over all CPUs matters. A task can increase this counter on
559 * one CPU and if it got migrated afterwards it may decrease
560 * it on another CPU. Always updated under the runqueue lock:
562 unsigned long nr_uninterruptible;
564 struct task_struct *curr, *idle;
565 unsigned long next_balance;
566 struct mm_struct *prev_mm;
573 struct root_domain *rd;
574 struct sched_domain *sd;
576 unsigned char idle_at_tick;
577 /* For active balancing */
581 /* cpu of this runqueue: */
585 unsigned long avg_load_per_task;
587 struct task_struct *migration_thread;
588 struct list_head migration_queue;
596 /* calc_load related fields */
597 unsigned long calc_load_update;
598 long calc_load_active;
600 #ifdef CONFIG_SCHED_HRTICK
602 int hrtick_csd_pending;
603 struct call_single_data hrtick_csd;
605 struct hrtimer hrtick_timer;
608 #ifdef CONFIG_SCHEDSTATS
610 struct sched_info rq_sched_info;
611 unsigned long long rq_cpu_time;
612 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
614 /* sys_sched_yield() stats */
615 unsigned int yld_count;
617 /* schedule() stats */
618 unsigned int sched_switch;
619 unsigned int sched_count;
620 unsigned int sched_goidle;
622 /* try_to_wake_up() stats */
623 unsigned int ttwu_count;
624 unsigned int ttwu_local;
627 unsigned int bkl_count;
631 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
634 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
636 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
639 static inline int cpu_of(struct rq *rq)
649 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
650 * See detach_destroy_domains: synchronize_sched for details.
652 * The domain tree of any CPU may only be accessed from within
653 * preempt-disabled sections.
655 #define for_each_domain(cpu, __sd) \
656 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
658 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
659 #define this_rq() (&__get_cpu_var(runqueues))
660 #define task_rq(p) cpu_rq(task_cpu(p))
661 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
662 #define raw_rq() (&__raw_get_cpu_var(runqueues))
664 inline void update_rq_clock(struct rq *rq)
666 rq->clock = sched_clock_cpu(cpu_of(rq));
670 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
672 #ifdef CONFIG_SCHED_DEBUG
673 # define const_debug __read_mostly
675 # define const_debug static const
680 * @cpu: the processor in question.
682 * Returns true if the current cpu runqueue is locked.
683 * This interface allows printk to be called with the runqueue lock
684 * held and know whether or not it is OK to wake up the klogd.
686 int runqueue_is_locked(int cpu)
688 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
692 * Debugging: various feature bits
695 #define SCHED_FEAT(name, enabled) \
696 __SCHED_FEAT_##name ,
699 #include "sched_features.h"
704 #define SCHED_FEAT(name, enabled) \
705 (1UL << __SCHED_FEAT_##name) * enabled |
707 const_debug unsigned int sysctl_sched_features =
708 #include "sched_features.h"
713 #ifdef CONFIG_SCHED_DEBUG
714 #define SCHED_FEAT(name, enabled) \
717 static __read_mostly char *sched_feat_names[] = {
718 #include "sched_features.h"
724 static int sched_feat_show(struct seq_file *m, void *v)
728 for (i = 0; sched_feat_names[i]; i++) {
729 if (!(sysctl_sched_features & (1UL << i)))
731 seq_printf(m, "%s ", sched_feat_names[i]);
739 sched_feat_write(struct file *filp, const char __user *ubuf,
740 size_t cnt, loff_t *ppos)
750 if (copy_from_user(&buf, ubuf, cnt))
755 if (strncmp(buf, "NO_", 3) == 0) {
760 for (i = 0; sched_feat_names[i]; i++) {
761 int len = strlen(sched_feat_names[i]);
763 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
765 sysctl_sched_features &= ~(1UL << i);
767 sysctl_sched_features |= (1UL << i);
772 if (!sched_feat_names[i])
780 static int sched_feat_open(struct inode *inode, struct file *filp)
782 return single_open(filp, sched_feat_show, NULL);
785 static const struct file_operations sched_feat_fops = {
786 .open = sched_feat_open,
787 .write = sched_feat_write,
790 .release = single_release,
793 static __init int sched_init_debug(void)
795 debugfs_create_file("sched_features", 0644, NULL, NULL,
800 late_initcall(sched_init_debug);
804 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
807 * Number of tasks to iterate in a single balance run.
808 * Limited because this is done with IRQs disabled.
810 const_debug unsigned int sysctl_sched_nr_migrate = 32;
813 * ratelimit for updating the group shares.
816 unsigned int sysctl_sched_shares_ratelimit = 250000;
817 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
820 * Inject some fuzzyness into changing the per-cpu group shares
821 * this avoids remote rq-locks at the expense of fairness.
824 unsigned int sysctl_sched_shares_thresh = 4;
827 * period over which we average the RT time consumption, measured
832 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
835 * period over which we measure -rt task cpu usage in us.
838 unsigned int sysctl_sched_rt_period = 1000000;
840 static __read_mostly int scheduler_running;
843 * part of the period that we allow rt tasks to run in us.
846 int sysctl_sched_rt_runtime = 950000;
848 static inline u64 global_rt_period(void)
850 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
853 static inline u64 global_rt_runtime(void)
855 if (sysctl_sched_rt_runtime < 0)
858 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
861 #ifndef prepare_arch_switch
862 # define prepare_arch_switch(next) do { } while (0)
864 #ifndef finish_arch_switch
865 # define finish_arch_switch(prev) do { } while (0)
868 static inline int task_current(struct rq *rq, struct task_struct *p)
870 return rq->curr == p;
873 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
874 static inline int task_running(struct rq *rq, struct task_struct *p)
876 return task_current(rq, p);
879 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
883 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
885 #ifdef CONFIG_DEBUG_SPINLOCK
886 /* this is a valid case when another task releases the spinlock */
887 rq->lock.owner = current;
890 * If we are tracking spinlock dependencies then we have to
891 * fix up the runqueue lock - which gets 'carried over' from
894 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
896 raw_spin_unlock_irq(&rq->lock);
899 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
900 static inline int task_running(struct rq *rq, struct task_struct *p)
905 return task_current(rq, p);
909 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
913 * We can optimise this out completely for !SMP, because the
914 * SMP rebalancing from interrupt is the only thing that cares
919 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
920 raw_spin_unlock_irq(&rq->lock);
922 raw_spin_unlock(&rq->lock);
926 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
930 * After ->oncpu is cleared, the task can be moved to a different CPU.
931 * We must ensure this doesn't happen until the switch is completely
937 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
941 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
944 * __task_rq_lock - lock the runqueue a given task resides on.
945 * Must be called interrupts disabled.
947 static inline struct rq *__task_rq_lock(struct task_struct *p)
951 struct rq *rq = task_rq(p);
952 raw_spin_lock(&rq->lock);
953 if (likely(rq == task_rq(p)))
955 raw_spin_unlock(&rq->lock);
960 * task_rq_lock - lock the runqueue a given task resides on and disable
961 * interrupts. Note the ordering: we can safely lookup the task_rq without
962 * explicitly disabling preemption.
964 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
970 local_irq_save(*flags);
972 raw_spin_lock(&rq->lock);
973 if (likely(rq == task_rq(p)))
975 raw_spin_unlock_irqrestore(&rq->lock, *flags);
979 void task_rq_unlock_wait(struct task_struct *p)
981 struct rq *rq = task_rq(p);
983 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
984 raw_spin_unlock_wait(&rq->lock);
987 static void __task_rq_unlock(struct rq *rq)
990 raw_spin_unlock(&rq->lock);
993 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
996 raw_spin_unlock_irqrestore(&rq->lock, *flags);
1000 * this_rq_lock - lock this runqueue and disable interrupts.
1002 static struct rq *this_rq_lock(void)
1003 __acquires(rq->lock)
1007 local_irq_disable();
1009 raw_spin_lock(&rq->lock);
1014 #ifdef CONFIG_SCHED_HRTICK
1016 * Use HR-timers to deliver accurate preemption points.
1018 * Its all a bit involved since we cannot program an hrt while holding the
1019 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1022 * When we get rescheduled we reprogram the hrtick_timer outside of the
1028 * - enabled by features
1029 * - hrtimer is actually high res
1031 static inline int hrtick_enabled(struct rq *rq)
1033 if (!sched_feat(HRTICK))
1035 if (!cpu_active(cpu_of(rq)))
1037 return hrtimer_is_hres_active(&rq->hrtick_timer);
1040 static void hrtick_clear(struct rq *rq)
1042 if (hrtimer_active(&rq->hrtick_timer))
1043 hrtimer_cancel(&rq->hrtick_timer);
1047 * High-resolution timer tick.
1048 * Runs from hardirq context with interrupts disabled.
1050 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1052 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1054 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1056 raw_spin_lock(&rq->lock);
1057 update_rq_clock(rq);
1058 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1059 raw_spin_unlock(&rq->lock);
1061 return HRTIMER_NORESTART;
1066 * called from hardirq (IPI) context
1068 static void __hrtick_start(void *arg)
1070 struct rq *rq = arg;
1072 raw_spin_lock(&rq->lock);
1073 hrtimer_restart(&rq->hrtick_timer);
1074 rq->hrtick_csd_pending = 0;
1075 raw_spin_unlock(&rq->lock);
1079 * Called to set the hrtick timer state.
1081 * called with rq->lock held and irqs disabled
1083 static void hrtick_start(struct rq *rq, u64 delay)
1085 struct hrtimer *timer = &rq->hrtick_timer;
1086 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1088 hrtimer_set_expires(timer, time);
1090 if (rq == this_rq()) {
1091 hrtimer_restart(timer);
1092 } else if (!rq->hrtick_csd_pending) {
1093 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1094 rq->hrtick_csd_pending = 1;
1099 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1101 int cpu = (int)(long)hcpu;
1104 case CPU_UP_CANCELED:
1105 case CPU_UP_CANCELED_FROZEN:
1106 case CPU_DOWN_PREPARE:
1107 case CPU_DOWN_PREPARE_FROZEN:
1109 case CPU_DEAD_FROZEN:
1110 hrtick_clear(cpu_rq(cpu));
1117 static __init void init_hrtick(void)
1119 hotcpu_notifier(hotplug_hrtick, 0);
1123 * Called to set the hrtick timer state.
1125 * called with rq->lock held and irqs disabled
1127 static void hrtick_start(struct rq *rq, u64 delay)
1129 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1130 HRTIMER_MODE_REL_PINNED, 0);
1133 static inline void init_hrtick(void)
1136 #endif /* CONFIG_SMP */
1138 static void init_rq_hrtick(struct rq *rq)
1141 rq->hrtick_csd_pending = 0;
1143 rq->hrtick_csd.flags = 0;
1144 rq->hrtick_csd.func = __hrtick_start;
1145 rq->hrtick_csd.info = rq;
1148 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1149 rq->hrtick_timer.function = hrtick;
1151 #else /* CONFIG_SCHED_HRTICK */
1152 static inline void hrtick_clear(struct rq *rq)
1156 static inline void init_rq_hrtick(struct rq *rq)
1160 static inline void init_hrtick(void)
1163 #endif /* CONFIG_SCHED_HRTICK */
1166 * resched_task - mark a task 'to be rescheduled now'.
1168 * On UP this means the setting of the need_resched flag, on SMP it
1169 * might also involve a cross-CPU call to trigger the scheduler on
1174 #ifndef tsk_is_polling
1175 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1178 static void resched_task(struct task_struct *p)
1182 assert_raw_spin_locked(&task_rq(p)->lock);
1184 if (test_tsk_need_resched(p))
1187 set_tsk_need_resched(p);
1190 if (cpu == smp_processor_id())
1193 /* NEED_RESCHED must be visible before we test polling */
1195 if (!tsk_is_polling(p))
1196 smp_send_reschedule(cpu);
1199 static void resched_cpu(int cpu)
1201 struct rq *rq = cpu_rq(cpu);
1202 unsigned long flags;
1204 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1206 resched_task(cpu_curr(cpu));
1207 raw_spin_unlock_irqrestore(&rq->lock, flags);
1212 * When add_timer_on() enqueues a timer into the timer wheel of an
1213 * idle CPU then this timer might expire before the next timer event
1214 * which is scheduled to wake up that CPU. In case of a completely
1215 * idle system the next event might even be infinite time into the
1216 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1217 * leaves the inner idle loop so the newly added timer is taken into
1218 * account when the CPU goes back to idle and evaluates the timer
1219 * wheel for the next timer event.
1221 void wake_up_idle_cpu(int cpu)
1223 struct rq *rq = cpu_rq(cpu);
1225 if (cpu == smp_processor_id())
1229 * This is safe, as this function is called with the timer
1230 * wheel base lock of (cpu) held. When the CPU is on the way
1231 * to idle and has not yet set rq->curr to idle then it will
1232 * be serialized on the timer wheel base lock and take the new
1233 * timer into account automatically.
1235 if (rq->curr != rq->idle)
1239 * We can set TIF_RESCHED on the idle task of the other CPU
1240 * lockless. The worst case is that the other CPU runs the
1241 * idle task through an additional NOOP schedule()
1243 set_tsk_need_resched(rq->idle);
1245 /* NEED_RESCHED must be visible before we test polling */
1247 if (!tsk_is_polling(rq->idle))
1248 smp_send_reschedule(cpu);
1250 #endif /* CONFIG_NO_HZ */
1252 static u64 sched_avg_period(void)
1254 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1257 static void sched_avg_update(struct rq *rq)
1259 s64 period = sched_avg_period();
1261 while ((s64)(rq->clock - rq->age_stamp) > period) {
1262 rq->age_stamp += period;
1267 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1269 rq->rt_avg += rt_delta;
1270 sched_avg_update(rq);
1273 #else /* !CONFIG_SMP */
1274 static void resched_task(struct task_struct *p)
1276 assert_raw_spin_locked(&task_rq(p)->lock);
1277 set_tsk_need_resched(p);
1280 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1283 #endif /* CONFIG_SMP */
1285 #if BITS_PER_LONG == 32
1286 # define WMULT_CONST (~0UL)
1288 # define WMULT_CONST (1UL << 32)
1291 #define WMULT_SHIFT 32
1294 * Shift right and round:
1296 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1299 * delta *= weight / lw
1301 static unsigned long
1302 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1303 struct load_weight *lw)
1307 if (!lw->inv_weight) {
1308 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1311 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1315 tmp = (u64)delta_exec * weight;
1317 * Check whether we'd overflow the 64-bit multiplication:
1319 if (unlikely(tmp > WMULT_CONST))
1320 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1323 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1325 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1328 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1334 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1341 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1342 * of tasks with abnormal "nice" values across CPUs the contribution that
1343 * each task makes to its run queue's load is weighted according to its
1344 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1345 * scaled version of the new time slice allocation that they receive on time
1349 #define WEIGHT_IDLEPRIO 3
1350 #define WMULT_IDLEPRIO 1431655765
1353 * Nice levels are multiplicative, with a gentle 10% change for every
1354 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1355 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1356 * that remained on nice 0.
1358 * The "10% effect" is relative and cumulative: from _any_ nice level,
1359 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1360 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1361 * If a task goes up by ~10% and another task goes down by ~10% then
1362 * the relative distance between them is ~25%.)
1364 static const int prio_to_weight[40] = {
1365 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1366 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1367 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1368 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1369 /* 0 */ 1024, 820, 655, 526, 423,
1370 /* 5 */ 335, 272, 215, 172, 137,
1371 /* 10 */ 110, 87, 70, 56, 45,
1372 /* 15 */ 36, 29, 23, 18, 15,
1376 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1378 * In cases where the weight does not change often, we can use the
1379 * precalculated inverse to speed up arithmetics by turning divisions
1380 * into multiplications:
1382 static const u32 prio_to_wmult[40] = {
1383 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1384 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1385 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1386 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1387 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1388 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1389 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1390 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1393 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1396 * runqueue iterator, to support SMP load-balancing between different
1397 * scheduling classes, without having to expose their internal data
1398 * structures to the load-balancing proper:
1400 struct rq_iterator {
1402 struct task_struct *(*start)(void *);
1403 struct task_struct *(*next)(void *);
1407 static unsigned long
1408 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1409 unsigned long max_load_move, struct sched_domain *sd,
1410 enum cpu_idle_type idle, int *all_pinned,
1411 int *this_best_prio, struct rq_iterator *iterator);
1414 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1415 struct sched_domain *sd, enum cpu_idle_type idle,
1416 struct rq_iterator *iterator);
1419 /* Time spent by the tasks of the cpu accounting group executing in ... */
1420 enum cpuacct_stat_index {
1421 CPUACCT_STAT_USER, /* ... user mode */
1422 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1424 CPUACCT_STAT_NSTATS,
1427 #ifdef CONFIG_CGROUP_CPUACCT
1428 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1429 static void cpuacct_update_stats(struct task_struct *tsk,
1430 enum cpuacct_stat_index idx, cputime_t val);
1432 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1433 static inline void cpuacct_update_stats(struct task_struct *tsk,
1434 enum cpuacct_stat_index idx, cputime_t val) {}
1437 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1439 update_load_add(&rq->load, load);
1442 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1444 update_load_sub(&rq->load, load);
1447 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1448 typedef int (*tg_visitor)(struct task_group *, void *);
1451 * Iterate the full tree, calling @down when first entering a node and @up when
1452 * leaving it for the final time.
1454 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1456 struct task_group *parent, *child;
1460 parent = &root_task_group;
1462 ret = (*down)(parent, data);
1465 list_for_each_entry_rcu(child, &parent->children, siblings) {
1472 ret = (*up)(parent, data);
1477 parent = parent->parent;
1486 static int tg_nop(struct task_group *tg, void *data)
1493 /* Used instead of source_load when we know the type == 0 */
1494 static unsigned long weighted_cpuload(const int cpu)
1496 return cpu_rq(cpu)->load.weight;
1500 * Return a low guess at the load of a migration-source cpu weighted
1501 * according to the scheduling class and "nice" value.
1503 * We want to under-estimate the load of migration sources, to
1504 * balance conservatively.
1506 static unsigned long source_load(int cpu, int type)
1508 struct rq *rq = cpu_rq(cpu);
1509 unsigned long total = weighted_cpuload(cpu);
1511 if (type == 0 || !sched_feat(LB_BIAS))
1514 return min(rq->cpu_load[type-1], total);
1518 * Return a high guess at the load of a migration-target cpu weighted
1519 * according to the scheduling class and "nice" value.
1521 static unsigned long target_load(int cpu, int type)
1523 struct rq *rq = cpu_rq(cpu);
1524 unsigned long total = weighted_cpuload(cpu);
1526 if (type == 0 || !sched_feat(LB_BIAS))
1529 return max(rq->cpu_load[type-1], total);
1532 static struct sched_group *group_of(int cpu)
1534 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1542 static unsigned long power_of(int cpu)
1544 struct sched_group *group = group_of(cpu);
1547 return SCHED_LOAD_SCALE;
1549 return group->cpu_power;
1552 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1554 static unsigned long cpu_avg_load_per_task(int cpu)
1556 struct rq *rq = cpu_rq(cpu);
1557 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1560 rq->avg_load_per_task = rq->load.weight / nr_running;
1562 rq->avg_load_per_task = 0;
1564 return rq->avg_load_per_task;
1567 #ifdef CONFIG_FAIR_GROUP_SCHED
1569 static __read_mostly unsigned long *update_shares_data;
1571 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1574 * Calculate and set the cpu's group shares.
1576 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1577 unsigned long sd_shares,
1578 unsigned long sd_rq_weight,
1579 unsigned long *usd_rq_weight)
1581 unsigned long shares, rq_weight;
1584 rq_weight = usd_rq_weight[cpu];
1587 rq_weight = NICE_0_LOAD;
1591 * \Sum_j shares_j * rq_weight_i
1592 * shares_i = -----------------------------
1593 * \Sum_j rq_weight_j
1595 shares = (sd_shares * rq_weight) / sd_rq_weight;
1596 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1598 if (abs(shares - tg->se[cpu]->load.weight) >
1599 sysctl_sched_shares_thresh) {
1600 struct rq *rq = cpu_rq(cpu);
1601 unsigned long flags;
1603 raw_spin_lock_irqsave(&rq->lock, flags);
1604 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1605 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1606 __set_se_shares(tg->se[cpu], shares);
1607 raw_spin_unlock_irqrestore(&rq->lock, flags);
1612 * Re-compute the task group their per cpu shares over the given domain.
1613 * This needs to be done in a bottom-up fashion because the rq weight of a
1614 * parent group depends on the shares of its child groups.
1616 static int tg_shares_up(struct task_group *tg, void *data)
1618 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1619 unsigned long *usd_rq_weight;
1620 struct sched_domain *sd = data;
1621 unsigned long flags;
1627 local_irq_save(flags);
1628 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1630 for_each_cpu(i, sched_domain_span(sd)) {
1631 weight = tg->cfs_rq[i]->load.weight;
1632 usd_rq_weight[i] = weight;
1634 rq_weight += weight;
1636 * If there are currently no tasks on the cpu pretend there
1637 * is one of average load so that when a new task gets to
1638 * run here it will not get delayed by group starvation.
1641 weight = NICE_0_LOAD;
1643 sum_weight += weight;
1644 shares += tg->cfs_rq[i]->shares;
1648 rq_weight = sum_weight;
1650 if ((!shares && rq_weight) || shares > tg->shares)
1651 shares = tg->shares;
1653 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1654 shares = tg->shares;
1656 for_each_cpu(i, sched_domain_span(sd))
1657 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1659 local_irq_restore(flags);
1665 * Compute the cpu's hierarchical load factor for each task group.
1666 * This needs to be done in a top-down fashion because the load of a child
1667 * group is a fraction of its parents load.
1669 static int tg_load_down(struct task_group *tg, void *data)
1672 long cpu = (long)data;
1675 load = cpu_rq(cpu)->load.weight;
1677 load = tg->parent->cfs_rq[cpu]->h_load;
1678 load *= tg->cfs_rq[cpu]->shares;
1679 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1682 tg->cfs_rq[cpu]->h_load = load;
1687 static void update_shares(struct sched_domain *sd)
1692 if (root_task_group_empty())
1695 now = cpu_clock(raw_smp_processor_id());
1696 elapsed = now - sd->last_update;
1698 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1699 sd->last_update = now;
1700 walk_tg_tree(tg_nop, tg_shares_up, sd);
1704 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1706 if (root_task_group_empty())
1709 raw_spin_unlock(&rq->lock);
1711 raw_spin_lock(&rq->lock);
1714 static void update_h_load(long cpu)
1716 if (root_task_group_empty())
1719 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1724 static inline void update_shares(struct sched_domain *sd)
1728 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1734 #ifdef CONFIG_PREEMPT
1736 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1739 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1740 * way at the expense of forcing extra atomic operations in all
1741 * invocations. This assures that the double_lock is acquired using the
1742 * same underlying policy as the spinlock_t on this architecture, which
1743 * reduces latency compared to the unfair variant below. However, it
1744 * also adds more overhead and therefore may reduce throughput.
1746 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1747 __releases(this_rq->lock)
1748 __acquires(busiest->lock)
1749 __acquires(this_rq->lock)
1751 raw_spin_unlock(&this_rq->lock);
1752 double_rq_lock(this_rq, busiest);
1759 * Unfair double_lock_balance: Optimizes throughput at the expense of
1760 * latency by eliminating extra atomic operations when the locks are
1761 * already in proper order on entry. This favors lower cpu-ids and will
1762 * grant the double lock to lower cpus over higher ids under contention,
1763 * regardless of entry order into the function.
1765 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1766 __releases(this_rq->lock)
1767 __acquires(busiest->lock)
1768 __acquires(this_rq->lock)
1772 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1773 if (busiest < this_rq) {
1774 raw_spin_unlock(&this_rq->lock);
1775 raw_spin_lock(&busiest->lock);
1776 raw_spin_lock_nested(&this_rq->lock,
1777 SINGLE_DEPTH_NESTING);
1780 raw_spin_lock_nested(&busiest->lock,
1781 SINGLE_DEPTH_NESTING);
1786 #endif /* CONFIG_PREEMPT */
1789 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1791 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1793 if (unlikely(!irqs_disabled())) {
1794 /* printk() doesn't work good under rq->lock */
1795 raw_spin_unlock(&this_rq->lock);
1799 return _double_lock_balance(this_rq, busiest);
1802 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1803 __releases(busiest->lock)
1805 raw_spin_unlock(&busiest->lock);
1806 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1810 * double_rq_lock - safely lock two runqueues
1812 * Note this does not disable interrupts like task_rq_lock,
1813 * you need to do so manually before calling.
1815 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1816 __acquires(rq1->lock)
1817 __acquires(rq2->lock)
1819 BUG_ON(!irqs_disabled());
1821 raw_spin_lock(&rq1->lock);
1822 __acquire(rq2->lock); /* Fake it out ;) */
1825 raw_spin_lock(&rq1->lock);
1826 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1828 raw_spin_lock(&rq2->lock);
1829 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1832 update_rq_clock(rq1);
1833 update_rq_clock(rq2);
1837 * double_rq_unlock - safely unlock two runqueues
1839 * Note this does not restore interrupts like task_rq_unlock,
1840 * you need to do so manually after calling.
1842 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1843 __releases(rq1->lock)
1844 __releases(rq2->lock)
1846 raw_spin_unlock(&rq1->lock);
1848 raw_spin_unlock(&rq2->lock);
1850 __release(rq2->lock);
1855 #ifdef CONFIG_FAIR_GROUP_SCHED
1856 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1859 cfs_rq->shares = shares;
1864 static void calc_load_account_active(struct rq *this_rq);
1865 static void update_sysctl(void);
1866 static int get_update_sysctl_factor(void);
1868 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1870 set_task_rq(p, cpu);
1873 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1874 * successfuly executed on another CPU. We must ensure that updates of
1875 * per-task data have been completed by this moment.
1878 task_thread_info(p)->cpu = cpu;
1882 static const struct sched_class rt_sched_class;
1884 #define sched_class_highest (&rt_sched_class)
1885 #define for_each_class(class) \
1886 for (class = sched_class_highest; class; class = class->next)
1888 #include "sched_stats.h"
1890 static void inc_nr_running(struct rq *rq)
1895 static void dec_nr_running(struct rq *rq)
1900 static void set_load_weight(struct task_struct *p)
1902 if (task_has_rt_policy(p)) {
1903 p->se.load.weight = prio_to_weight[0] * 2;
1904 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1909 * SCHED_IDLE tasks get minimal weight:
1911 if (p->policy == SCHED_IDLE) {
1912 p->se.load.weight = WEIGHT_IDLEPRIO;
1913 p->se.load.inv_weight = WMULT_IDLEPRIO;
1917 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1918 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1921 static void update_avg(u64 *avg, u64 sample)
1923 s64 diff = sample - *avg;
1927 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1930 p->se.start_runtime = p->se.sum_exec_runtime;
1932 sched_info_queued(p);
1933 p->sched_class->enqueue_task(rq, p, wakeup);
1937 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1940 if (p->se.last_wakeup) {
1941 update_avg(&p->se.avg_overlap,
1942 p->se.sum_exec_runtime - p->se.last_wakeup);
1943 p->se.last_wakeup = 0;
1945 update_avg(&p->se.avg_wakeup,
1946 sysctl_sched_wakeup_granularity);
1950 sched_info_dequeued(p);
1951 p->sched_class->dequeue_task(rq, p, sleep);
1956 * activate_task - move a task to the runqueue.
1958 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1960 if (task_contributes_to_load(p))
1961 rq->nr_uninterruptible--;
1963 enqueue_task(rq, p, wakeup);
1968 * deactivate_task - remove a task from the runqueue.
1970 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1972 if (task_contributes_to_load(p))
1973 rq->nr_uninterruptible++;
1975 dequeue_task(rq, p, sleep);
1979 #include "sched_idletask.c"
1980 #include "sched_fair.c"
1981 #include "sched_rt.c"
1982 #ifdef CONFIG_SCHED_DEBUG
1983 # include "sched_debug.c"
1987 * __normal_prio - return the priority that is based on the static prio
1989 static inline int __normal_prio(struct task_struct *p)
1991 return p->static_prio;
1995 * Calculate the expected normal priority: i.e. priority
1996 * without taking RT-inheritance into account. Might be
1997 * boosted by interactivity modifiers. Changes upon fork,
1998 * setprio syscalls, and whenever the interactivity
1999 * estimator recalculates.
2001 static inline int normal_prio(struct task_struct *p)
2005 if (task_has_rt_policy(p))
2006 prio = MAX_RT_PRIO-1 - p->rt_priority;
2008 prio = __normal_prio(p);
2013 * Calculate the current priority, i.e. the priority
2014 * taken into account by the scheduler. This value might
2015 * be boosted by RT tasks, or might be boosted by
2016 * interactivity modifiers. Will be RT if the task got
2017 * RT-boosted. If not then it returns p->normal_prio.
2019 static int effective_prio(struct task_struct *p)
2021 p->normal_prio = normal_prio(p);
2023 * If we are RT tasks or we were boosted to RT priority,
2024 * keep the priority unchanged. Otherwise, update priority
2025 * to the normal priority:
2027 if (!rt_prio(p->prio))
2028 return p->normal_prio;
2033 * task_curr - is this task currently executing on a CPU?
2034 * @p: the task in question.
2036 inline int task_curr(const struct task_struct *p)
2038 return cpu_curr(task_cpu(p)) == p;
2041 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2042 const struct sched_class *prev_class,
2043 int oldprio, int running)
2045 if (prev_class != p->sched_class) {
2046 if (prev_class->switched_from)
2047 prev_class->switched_from(rq, p, running);
2048 p->sched_class->switched_to(rq, p, running);
2050 p->sched_class->prio_changed(rq, p, oldprio, running);
2055 * Is this task likely cache-hot:
2058 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2062 if (p->sched_class != &fair_sched_class)
2066 * Buddy candidates are cache hot:
2068 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2069 (&p->se == cfs_rq_of(&p->se)->next ||
2070 &p->se == cfs_rq_of(&p->se)->last))
2073 if (sysctl_sched_migration_cost == -1)
2075 if (sysctl_sched_migration_cost == 0)
2078 delta = now - p->se.exec_start;
2080 return delta < (s64)sysctl_sched_migration_cost;
2083 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2085 #ifdef CONFIG_SCHED_DEBUG
2087 * We should never call set_task_cpu() on a blocked task,
2088 * ttwu() will sort out the placement.
2090 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2091 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2094 trace_sched_migrate_task(p, new_cpu);
2096 if (task_cpu(p) != new_cpu) {
2097 p->se.nr_migrations++;
2098 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2101 __set_task_cpu(p, new_cpu);
2104 struct migration_req {
2105 struct list_head list;
2107 struct task_struct *task;
2110 struct completion done;
2114 * The task's runqueue lock must be held.
2115 * Returns true if you have to wait for migration thread.
2118 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2120 struct rq *rq = task_rq(p);
2123 * If the task is not on a runqueue (and not running), then
2124 * the next wake-up will properly place the task.
2126 if (!p->se.on_rq && !task_running(rq, p))
2129 init_completion(&req->done);
2131 req->dest_cpu = dest_cpu;
2132 list_add(&req->list, &rq->migration_queue);
2138 * wait_task_context_switch - wait for a thread to complete at least one
2141 * @p must not be current.
2143 void wait_task_context_switch(struct task_struct *p)
2145 unsigned long nvcsw, nivcsw, flags;
2153 * The runqueue is assigned before the actual context
2154 * switch. We need to take the runqueue lock.
2156 * We could check initially without the lock but it is
2157 * very likely that we need to take the lock in every
2160 rq = task_rq_lock(p, &flags);
2161 running = task_running(rq, p);
2162 task_rq_unlock(rq, &flags);
2164 if (likely(!running))
2167 * The switch count is incremented before the actual
2168 * context switch. We thus wait for two switches to be
2169 * sure at least one completed.
2171 if ((p->nvcsw - nvcsw) > 1)
2173 if ((p->nivcsw - nivcsw) > 1)
2181 * wait_task_inactive - wait for a thread to unschedule.
2183 * If @match_state is nonzero, it's the @p->state value just checked and
2184 * not expected to change. If it changes, i.e. @p might have woken up,
2185 * then return zero. When we succeed in waiting for @p to be off its CPU,
2186 * we return a positive number (its total switch count). If a second call
2187 * a short while later returns the same number, the caller can be sure that
2188 * @p has remained unscheduled the whole time.
2190 * The caller must ensure that the task *will* unschedule sometime soon,
2191 * else this function might spin for a *long* time. This function can't
2192 * be called with interrupts off, or it may introduce deadlock with
2193 * smp_call_function() if an IPI is sent by the same process we are
2194 * waiting to become inactive.
2196 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2198 unsigned long flags;
2205 * We do the initial early heuristics without holding
2206 * any task-queue locks at all. We'll only try to get
2207 * the runqueue lock when things look like they will
2213 * If the task is actively running on another CPU
2214 * still, just relax and busy-wait without holding
2217 * NOTE! Since we don't hold any locks, it's not
2218 * even sure that "rq" stays as the right runqueue!
2219 * But we don't care, since "task_running()" will
2220 * return false if the runqueue has changed and p
2221 * is actually now running somewhere else!
2223 while (task_running(rq, p)) {
2224 if (match_state && unlikely(p->state != match_state))
2230 * Ok, time to look more closely! We need the rq
2231 * lock now, to be *sure*. If we're wrong, we'll
2232 * just go back and repeat.
2234 rq = task_rq_lock(p, &flags);
2235 trace_sched_wait_task(rq, p);
2236 running = task_running(rq, p);
2237 on_rq = p->se.on_rq;
2239 if (!match_state || p->state == match_state)
2240 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2241 task_rq_unlock(rq, &flags);
2244 * If it changed from the expected state, bail out now.
2246 if (unlikely(!ncsw))
2250 * Was it really running after all now that we
2251 * checked with the proper locks actually held?
2253 * Oops. Go back and try again..
2255 if (unlikely(running)) {
2261 * It's not enough that it's not actively running,
2262 * it must be off the runqueue _entirely_, and not
2265 * So if it was still runnable (but just not actively
2266 * running right now), it's preempted, and we should
2267 * yield - it could be a while.
2269 if (unlikely(on_rq)) {
2270 schedule_timeout_uninterruptible(1);
2275 * Ahh, all good. It wasn't running, and it wasn't
2276 * runnable, which means that it will never become
2277 * running in the future either. We're all done!
2286 * kick_process - kick a running thread to enter/exit the kernel
2287 * @p: the to-be-kicked thread
2289 * Cause a process which is running on another CPU to enter
2290 * kernel-mode, without any delay. (to get signals handled.)
2292 * NOTE: this function doesnt have to take the runqueue lock,
2293 * because all it wants to ensure is that the remote task enters
2294 * the kernel. If the IPI races and the task has been migrated
2295 * to another CPU then no harm is done and the purpose has been
2298 void kick_process(struct task_struct *p)
2304 if ((cpu != smp_processor_id()) && task_curr(p))
2305 smp_send_reschedule(cpu);
2308 EXPORT_SYMBOL_GPL(kick_process);
2309 #endif /* CONFIG_SMP */
2312 * task_oncpu_function_call - call a function on the cpu on which a task runs
2313 * @p: the task to evaluate
2314 * @func: the function to be called
2315 * @info: the function call argument
2317 * Calls the function @func when the task is currently running. This might
2318 * be on the current CPU, which just calls the function directly
2320 void task_oncpu_function_call(struct task_struct *p,
2321 void (*func) (void *info), void *info)
2328 smp_call_function_single(cpu, func, info, 1);
2333 static int select_fallback_rq(int cpu, struct task_struct *p)
2336 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2338 /* Look for allowed, online CPU in same node. */
2339 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2340 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2343 /* Any allowed, online CPU? */
2344 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2345 if (dest_cpu < nr_cpu_ids)
2348 /* No more Mr. Nice Guy. */
2349 if (dest_cpu >= nr_cpu_ids) {
2351 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2353 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2356 * Don't tell them about moving exiting tasks or
2357 * kernel threads (both mm NULL), since they never
2360 if (p->mm && printk_ratelimit()) {
2361 printk(KERN_INFO "process %d (%s) no "
2362 "longer affine to cpu%d\n",
2363 task_pid_nr(p), p->comm, cpu);
2373 * - fork, @p is stable because it isn't on the tasklist yet
2375 * - exec, @p is unstable, retry loop
2377 * - wake-up, we serialize ->cpus_allowed against TASK_WAKING so
2378 * we should be good.
2381 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2383 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2386 * In order not to call set_task_cpu() on a blocking task we need
2387 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2390 * Since this is common to all placement strategies, this lives here.
2392 * [ this allows ->select_task() to simply return task_cpu(p) and
2393 * not worry about this generic constraint ]
2395 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2397 cpu = select_fallback_rq(task_cpu(p), p);
2404 * try_to_wake_up - wake up a thread
2405 * @p: the to-be-woken-up thread
2406 * @state: the mask of task states that can be woken
2407 * @sync: do a synchronous wakeup?
2409 * Put it on the run-queue if it's not already there. The "current"
2410 * thread is always on the run-queue (except when the actual
2411 * re-schedule is in progress), and as such you're allowed to do
2412 * the simpler "current->state = TASK_RUNNING" to mark yourself
2413 * runnable without the overhead of this.
2415 * returns failure only if the task is already active.
2417 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2420 int cpu, orig_cpu, this_cpu, success = 0;
2421 unsigned long flags;
2422 struct rq *rq, *orig_rq;
2424 if (!sched_feat(SYNC_WAKEUPS))
2425 wake_flags &= ~WF_SYNC;
2427 this_cpu = get_cpu();
2430 rq = orig_rq = task_rq_lock(p, &flags);
2431 update_rq_clock(rq);
2432 if (!(p->state & state))
2442 if (unlikely(task_running(rq, p)))
2446 * In order to handle concurrent wakeups and release the rq->lock
2447 * we put the task in TASK_WAKING state.
2449 * First fix up the nr_uninterruptible count:
2451 if (task_contributes_to_load(p))
2452 rq->nr_uninterruptible--;
2453 p->state = TASK_WAKING;
2455 if (p->sched_class->task_waking)
2456 p->sched_class->task_waking(rq, p);
2458 __task_rq_unlock(rq);
2460 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2461 if (cpu != orig_cpu)
2462 set_task_cpu(p, cpu);
2464 rq = __task_rq_lock(p);
2465 update_rq_clock(rq);
2467 WARN_ON(p->state != TASK_WAKING);
2470 #ifdef CONFIG_SCHEDSTATS
2471 schedstat_inc(rq, ttwu_count);
2472 if (cpu == this_cpu)
2473 schedstat_inc(rq, ttwu_local);
2475 struct sched_domain *sd;
2476 for_each_domain(this_cpu, sd) {
2477 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2478 schedstat_inc(sd, ttwu_wake_remote);
2483 #endif /* CONFIG_SCHEDSTATS */
2486 #endif /* CONFIG_SMP */
2487 schedstat_inc(p, se.nr_wakeups);
2488 if (wake_flags & WF_SYNC)
2489 schedstat_inc(p, se.nr_wakeups_sync);
2490 if (orig_cpu != cpu)
2491 schedstat_inc(p, se.nr_wakeups_migrate);
2492 if (cpu == this_cpu)
2493 schedstat_inc(p, se.nr_wakeups_local);
2495 schedstat_inc(p, se.nr_wakeups_remote);
2496 activate_task(rq, p, 1);
2500 * Only attribute actual wakeups done by this task.
2502 if (!in_interrupt()) {
2503 struct sched_entity *se = ¤t->se;
2504 u64 sample = se->sum_exec_runtime;
2506 if (se->last_wakeup)
2507 sample -= se->last_wakeup;
2509 sample -= se->start_runtime;
2510 update_avg(&se->avg_wakeup, sample);
2512 se->last_wakeup = se->sum_exec_runtime;
2516 trace_sched_wakeup(rq, p, success);
2517 check_preempt_curr(rq, p, wake_flags);
2519 p->state = TASK_RUNNING;
2521 if (p->sched_class->task_woken)
2522 p->sched_class->task_woken(rq, p);
2524 if (unlikely(rq->idle_stamp)) {
2525 u64 delta = rq->clock - rq->idle_stamp;
2526 u64 max = 2*sysctl_sched_migration_cost;
2531 update_avg(&rq->avg_idle, delta);
2536 task_rq_unlock(rq, &flags);
2543 * wake_up_process - Wake up a specific process
2544 * @p: The process to be woken up.
2546 * Attempt to wake up the nominated process and move it to the set of runnable
2547 * processes. Returns 1 if the process was woken up, 0 if it was already
2550 * It may be assumed that this function implies a write memory barrier before
2551 * changing the task state if and only if any tasks are woken up.
2553 int wake_up_process(struct task_struct *p)
2555 return try_to_wake_up(p, TASK_ALL, 0);
2557 EXPORT_SYMBOL(wake_up_process);
2559 int wake_up_state(struct task_struct *p, unsigned int state)
2561 return try_to_wake_up(p, state, 0);
2565 * Perform scheduler related setup for a newly forked process p.
2566 * p is forked by current.
2568 * __sched_fork() is basic setup used by init_idle() too:
2570 static void __sched_fork(struct task_struct *p)
2572 p->se.exec_start = 0;
2573 p->se.sum_exec_runtime = 0;
2574 p->se.prev_sum_exec_runtime = 0;
2575 p->se.nr_migrations = 0;
2576 p->se.last_wakeup = 0;
2577 p->se.avg_overlap = 0;
2578 p->se.start_runtime = 0;
2579 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2581 #ifdef CONFIG_SCHEDSTATS
2582 p->se.wait_start = 0;
2584 p->se.wait_count = 0;
2587 p->se.sleep_start = 0;
2588 p->se.sleep_max = 0;
2589 p->se.sum_sleep_runtime = 0;
2591 p->se.block_start = 0;
2592 p->se.block_max = 0;
2594 p->se.slice_max = 0;
2596 p->se.nr_migrations_cold = 0;
2597 p->se.nr_failed_migrations_affine = 0;
2598 p->se.nr_failed_migrations_running = 0;
2599 p->se.nr_failed_migrations_hot = 0;
2600 p->se.nr_forced_migrations = 0;
2602 p->se.nr_wakeups = 0;
2603 p->se.nr_wakeups_sync = 0;
2604 p->se.nr_wakeups_migrate = 0;
2605 p->se.nr_wakeups_local = 0;
2606 p->se.nr_wakeups_remote = 0;
2607 p->se.nr_wakeups_affine = 0;
2608 p->se.nr_wakeups_affine_attempts = 0;
2609 p->se.nr_wakeups_passive = 0;
2610 p->se.nr_wakeups_idle = 0;
2614 INIT_LIST_HEAD(&p->rt.run_list);
2616 INIT_LIST_HEAD(&p->se.group_node);
2618 #ifdef CONFIG_PREEMPT_NOTIFIERS
2619 INIT_HLIST_HEAD(&p->preempt_notifiers);
2624 * fork()/clone()-time setup:
2626 void sched_fork(struct task_struct *p, int clone_flags)
2628 int cpu = get_cpu();
2632 * We mark the process as waking here. This guarantees that
2633 * nobody will actually run it, and a signal or other external
2634 * event cannot wake it up and insert it on the runqueue either.
2636 p->state = TASK_WAKING;
2639 * Revert to default priority/policy on fork if requested.
2641 if (unlikely(p->sched_reset_on_fork)) {
2642 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2643 p->policy = SCHED_NORMAL;
2644 p->normal_prio = p->static_prio;
2647 if (PRIO_TO_NICE(p->static_prio) < 0) {
2648 p->static_prio = NICE_TO_PRIO(0);
2649 p->normal_prio = p->static_prio;
2654 * We don't need the reset flag anymore after the fork. It has
2655 * fulfilled its duty:
2657 p->sched_reset_on_fork = 0;
2661 * Make sure we do not leak PI boosting priority to the child.
2663 p->prio = current->normal_prio;
2665 if (!rt_prio(p->prio))
2666 p->sched_class = &fair_sched_class;
2668 if (p->sched_class->task_fork)
2669 p->sched_class->task_fork(p);
2672 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2674 set_task_cpu(p, cpu);
2676 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2677 if (likely(sched_info_on()))
2678 memset(&p->sched_info, 0, sizeof(p->sched_info));
2680 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2683 #ifdef CONFIG_PREEMPT
2684 /* Want to start with kernel preemption disabled. */
2685 task_thread_info(p)->preempt_count = 1;
2687 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2693 * wake_up_new_task - wake up a newly created task for the first time.
2695 * This function will do some initial scheduler statistics housekeeping
2696 * that must be done for every newly created context, then puts the task
2697 * on the runqueue and wakes it.
2699 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2701 unsigned long flags;
2704 rq = task_rq_lock(p, &flags);
2705 BUG_ON(p->state != TASK_WAKING);
2706 p->state = TASK_RUNNING;
2707 update_rq_clock(rq);
2708 activate_task(rq, p, 0);
2709 trace_sched_wakeup_new(rq, p, 1);
2710 check_preempt_curr(rq, p, WF_FORK);
2712 if (p->sched_class->task_woken)
2713 p->sched_class->task_woken(rq, p);
2715 task_rq_unlock(rq, &flags);
2718 #ifdef CONFIG_PREEMPT_NOTIFIERS
2721 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2722 * @notifier: notifier struct to register
2724 void preempt_notifier_register(struct preempt_notifier *notifier)
2726 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2728 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2731 * preempt_notifier_unregister - no longer interested in preemption notifications
2732 * @notifier: notifier struct to unregister
2734 * This is safe to call from within a preemption notifier.
2736 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2738 hlist_del(¬ifier->link);
2740 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2742 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2744 struct preempt_notifier *notifier;
2745 struct hlist_node *node;
2747 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2748 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2752 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2753 struct task_struct *next)
2755 struct preempt_notifier *notifier;
2756 struct hlist_node *node;
2758 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2759 notifier->ops->sched_out(notifier, next);
2762 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2764 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2769 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2770 struct task_struct *next)
2774 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2777 * prepare_task_switch - prepare to switch tasks
2778 * @rq: the runqueue preparing to switch
2779 * @prev: the current task that is being switched out
2780 * @next: the task we are going to switch to.
2782 * This is called with the rq lock held and interrupts off. It must
2783 * be paired with a subsequent finish_task_switch after the context
2786 * prepare_task_switch sets up locking and calls architecture specific
2790 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2791 struct task_struct *next)
2793 fire_sched_out_preempt_notifiers(prev, next);
2794 prepare_lock_switch(rq, next);
2795 prepare_arch_switch(next);
2799 * finish_task_switch - clean up after a task-switch
2800 * @rq: runqueue associated with task-switch
2801 * @prev: the thread we just switched away from.
2803 * finish_task_switch must be called after the context switch, paired
2804 * with a prepare_task_switch call before the context switch.
2805 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2806 * and do any other architecture-specific cleanup actions.
2808 * Note that we may have delayed dropping an mm in context_switch(). If
2809 * so, we finish that here outside of the runqueue lock. (Doing it
2810 * with the lock held can cause deadlocks; see schedule() for
2813 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2814 __releases(rq->lock)
2816 struct mm_struct *mm = rq->prev_mm;
2822 * A task struct has one reference for the use as "current".
2823 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2824 * schedule one last time. The schedule call will never return, and
2825 * the scheduled task must drop that reference.
2826 * The test for TASK_DEAD must occur while the runqueue locks are
2827 * still held, otherwise prev could be scheduled on another cpu, die
2828 * there before we look at prev->state, and then the reference would
2830 * Manfred Spraul <manfred@colorfullife.com>
2832 prev_state = prev->state;
2833 finish_arch_switch(prev);
2834 perf_event_task_sched_in(current, cpu_of(rq));
2835 finish_lock_switch(rq, prev);
2837 fire_sched_in_preempt_notifiers(current);
2840 if (unlikely(prev_state == TASK_DEAD)) {
2842 * Remove function-return probe instances associated with this
2843 * task and put them back on the free list.
2845 kprobe_flush_task(prev);
2846 put_task_struct(prev);
2852 /* assumes rq->lock is held */
2853 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2855 if (prev->sched_class->pre_schedule)
2856 prev->sched_class->pre_schedule(rq, prev);
2859 /* rq->lock is NOT held, but preemption is disabled */
2860 static inline void post_schedule(struct rq *rq)
2862 if (rq->post_schedule) {
2863 unsigned long flags;
2865 raw_spin_lock_irqsave(&rq->lock, flags);
2866 if (rq->curr->sched_class->post_schedule)
2867 rq->curr->sched_class->post_schedule(rq);
2868 raw_spin_unlock_irqrestore(&rq->lock, flags);
2870 rq->post_schedule = 0;
2876 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2880 static inline void post_schedule(struct rq *rq)
2887 * schedule_tail - first thing a freshly forked thread must call.
2888 * @prev: the thread we just switched away from.
2890 asmlinkage void schedule_tail(struct task_struct *prev)
2891 __releases(rq->lock)
2893 struct rq *rq = this_rq();
2895 finish_task_switch(rq, prev);
2898 * FIXME: do we need to worry about rq being invalidated by the
2903 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2904 /* In this case, finish_task_switch does not reenable preemption */
2907 if (current->set_child_tid)
2908 put_user(task_pid_vnr(current), current->set_child_tid);
2912 * context_switch - switch to the new MM and the new
2913 * thread's register state.
2916 context_switch(struct rq *rq, struct task_struct *prev,
2917 struct task_struct *next)
2919 struct mm_struct *mm, *oldmm;
2921 prepare_task_switch(rq, prev, next);
2922 trace_sched_switch(rq, prev, next);
2924 oldmm = prev->active_mm;
2926 * For paravirt, this is coupled with an exit in switch_to to
2927 * combine the page table reload and the switch backend into
2930 arch_start_context_switch(prev);
2933 next->active_mm = oldmm;
2934 atomic_inc(&oldmm->mm_count);
2935 enter_lazy_tlb(oldmm, next);
2937 switch_mm(oldmm, mm, next);
2939 if (likely(!prev->mm)) {
2940 prev->active_mm = NULL;
2941 rq->prev_mm = oldmm;
2944 * Since the runqueue lock will be released by the next
2945 * task (which is an invalid locking op but in the case
2946 * of the scheduler it's an obvious special-case), so we
2947 * do an early lockdep release here:
2949 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2950 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2953 /* Here we just switch the register state and the stack. */
2954 switch_to(prev, next, prev);
2958 * this_rq must be evaluated again because prev may have moved
2959 * CPUs since it called schedule(), thus the 'rq' on its stack
2960 * frame will be invalid.
2962 finish_task_switch(this_rq(), prev);
2966 * nr_running, nr_uninterruptible and nr_context_switches:
2968 * externally visible scheduler statistics: current number of runnable
2969 * threads, current number of uninterruptible-sleeping threads, total
2970 * number of context switches performed since bootup.
2972 unsigned long nr_running(void)
2974 unsigned long i, sum = 0;
2976 for_each_online_cpu(i)
2977 sum += cpu_rq(i)->nr_running;
2982 unsigned long nr_uninterruptible(void)
2984 unsigned long i, sum = 0;
2986 for_each_possible_cpu(i)
2987 sum += cpu_rq(i)->nr_uninterruptible;
2990 * Since we read the counters lockless, it might be slightly
2991 * inaccurate. Do not allow it to go below zero though:
2993 if (unlikely((long)sum < 0))
2999 unsigned long long nr_context_switches(void)
3002 unsigned long long sum = 0;
3004 for_each_possible_cpu(i)
3005 sum += cpu_rq(i)->nr_switches;
3010 unsigned long nr_iowait(void)
3012 unsigned long i, sum = 0;
3014 for_each_possible_cpu(i)
3015 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3020 unsigned long nr_iowait_cpu(void)
3022 struct rq *this = this_rq();
3023 return atomic_read(&this->nr_iowait);
3026 unsigned long this_cpu_load(void)
3028 struct rq *this = this_rq();
3029 return this->cpu_load[0];
3033 /* Variables and functions for calc_load */
3034 static atomic_long_t calc_load_tasks;
3035 static unsigned long calc_load_update;
3036 unsigned long avenrun[3];
3037 EXPORT_SYMBOL(avenrun);
3040 * get_avenrun - get the load average array
3041 * @loads: pointer to dest load array
3042 * @offset: offset to add
3043 * @shift: shift count to shift the result left
3045 * These values are estimates at best, so no need for locking.
3047 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3049 loads[0] = (avenrun[0] + offset) << shift;
3050 loads[1] = (avenrun[1] + offset) << shift;
3051 loads[2] = (avenrun[2] + offset) << shift;
3054 static unsigned long
3055 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3058 load += active * (FIXED_1 - exp);
3059 return load >> FSHIFT;
3063 * calc_load - update the avenrun load estimates 10 ticks after the
3064 * CPUs have updated calc_load_tasks.
3066 void calc_global_load(void)
3068 unsigned long upd = calc_load_update + 10;
3071 if (time_before(jiffies, upd))
3074 active = atomic_long_read(&calc_load_tasks);
3075 active = active > 0 ? active * FIXED_1 : 0;
3077 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3078 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3079 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3081 calc_load_update += LOAD_FREQ;
3085 * Either called from update_cpu_load() or from a cpu going idle
3087 static void calc_load_account_active(struct rq *this_rq)
3089 long nr_active, delta;
3091 nr_active = this_rq->nr_running;
3092 nr_active += (long) this_rq->nr_uninterruptible;
3094 if (nr_active != this_rq->calc_load_active) {
3095 delta = nr_active - this_rq->calc_load_active;
3096 this_rq->calc_load_active = nr_active;
3097 atomic_long_add(delta, &calc_load_tasks);
3102 * Update rq->cpu_load[] statistics. This function is usually called every
3103 * scheduler tick (TICK_NSEC).
3105 static void update_cpu_load(struct rq *this_rq)
3107 unsigned long this_load = this_rq->load.weight;
3110 this_rq->nr_load_updates++;
3112 /* Update our load: */
3113 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3114 unsigned long old_load, new_load;
3116 /* scale is effectively 1 << i now, and >> i divides by scale */
3118 old_load = this_rq->cpu_load[i];
3119 new_load = this_load;
3121 * Round up the averaging division if load is increasing. This
3122 * prevents us from getting stuck on 9 if the load is 10, for
3125 if (new_load > old_load)
3126 new_load += scale-1;
3127 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3130 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3131 this_rq->calc_load_update += LOAD_FREQ;
3132 calc_load_account_active(this_rq);
3139 * sched_exec - execve() is a valuable balancing opportunity, because at
3140 * this point the task has the smallest effective memory and cache footprint.
3142 void sched_exec(void)
3144 struct task_struct *p = current;
3145 struct migration_req req;
3146 int dest_cpu, this_cpu;
3147 unsigned long flags;
3151 this_cpu = get_cpu();
3152 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3153 if (dest_cpu == this_cpu) {
3158 rq = task_rq_lock(p, &flags);
3162 * select_task_rq() can race against ->cpus_allowed
3164 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3165 || unlikely(!cpu_active(dest_cpu))) {
3166 task_rq_unlock(rq, &flags);
3170 /* force the process onto the specified CPU */
3171 if (migrate_task(p, dest_cpu, &req)) {
3172 /* Need to wait for migration thread (might exit: take ref). */
3173 struct task_struct *mt = rq->migration_thread;
3175 get_task_struct(mt);
3176 task_rq_unlock(rq, &flags);
3177 wake_up_process(mt);
3178 put_task_struct(mt);
3179 wait_for_completion(&req.done);
3183 task_rq_unlock(rq, &flags);
3188 DEFINE_PER_CPU(struct kernel_stat, kstat);
3190 EXPORT_PER_CPU_SYMBOL(kstat);
3193 * Return any ns on the sched_clock that have not yet been accounted in
3194 * @p in case that task is currently running.
3196 * Called with task_rq_lock() held on @rq.
3198 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3202 if (task_current(rq, p)) {
3203 update_rq_clock(rq);
3204 ns = rq->clock - p->se.exec_start;
3212 unsigned long long task_delta_exec(struct task_struct *p)
3214 unsigned long flags;
3218 rq = task_rq_lock(p, &flags);
3219 ns = do_task_delta_exec(p, rq);
3220 task_rq_unlock(rq, &flags);
3226 * Return accounted runtime for the task.
3227 * In case the task is currently running, return the runtime plus current's
3228 * pending runtime that have not been accounted yet.
3230 unsigned long long task_sched_runtime(struct task_struct *p)
3232 unsigned long flags;
3236 rq = task_rq_lock(p, &flags);
3237 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3238 task_rq_unlock(rq, &flags);
3244 * Return sum_exec_runtime for the thread group.
3245 * In case the task is currently running, return the sum plus current's
3246 * pending runtime that have not been accounted yet.
3248 * Note that the thread group might have other running tasks as well,
3249 * so the return value not includes other pending runtime that other
3250 * running tasks might have.
3252 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3254 struct task_cputime totals;
3255 unsigned long flags;
3259 rq = task_rq_lock(p, &flags);
3260 thread_group_cputime(p, &totals);
3261 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3262 task_rq_unlock(rq, &flags);
3268 * Account user cpu time to a process.
3269 * @p: the process that the cpu time gets accounted to
3270 * @cputime: the cpu time spent in user space since the last update
3271 * @cputime_scaled: cputime scaled by cpu frequency
3273 void account_user_time(struct task_struct *p, cputime_t cputime,
3274 cputime_t cputime_scaled)
3276 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3279 /* Add user time to process. */
3280 p->utime = cputime_add(p->utime, cputime);
3281 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3282 account_group_user_time(p, cputime);
3284 /* Add user time to cpustat. */
3285 tmp = cputime_to_cputime64(cputime);
3286 if (TASK_NICE(p) > 0)
3287 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3289 cpustat->user = cputime64_add(cpustat->user, tmp);
3291 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3292 /* Account for user time used */
3293 acct_update_integrals(p);
3297 * Account guest cpu time to a process.
3298 * @p: the process that the cpu time gets accounted to
3299 * @cputime: the cpu time spent in virtual machine since the last update
3300 * @cputime_scaled: cputime scaled by cpu frequency
3302 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3303 cputime_t cputime_scaled)
3306 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3308 tmp = cputime_to_cputime64(cputime);
3310 /* Add guest time to process. */
3311 p->utime = cputime_add(p->utime, cputime);
3312 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3313 account_group_user_time(p, cputime);
3314 p->gtime = cputime_add(p->gtime, cputime);
3316 /* Add guest time to cpustat. */
3317 if (TASK_NICE(p) > 0) {
3318 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3319 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3321 cpustat->user = cputime64_add(cpustat->user, tmp);
3322 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3327 * Account system cpu time to a process.
3328 * @p: the process that the cpu time gets accounted to
3329 * @hardirq_offset: the offset to subtract from hardirq_count()
3330 * @cputime: the cpu time spent in kernel space since the last update
3331 * @cputime_scaled: cputime scaled by cpu frequency
3333 void account_system_time(struct task_struct *p, int hardirq_offset,
3334 cputime_t cputime, cputime_t cputime_scaled)
3336 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3339 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3340 account_guest_time(p, cputime, cputime_scaled);
3344 /* Add system time to process. */
3345 p->stime = cputime_add(p->stime, cputime);
3346 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3347 account_group_system_time(p, cputime);
3349 /* Add system time to cpustat. */
3350 tmp = cputime_to_cputime64(cputime);
3351 if (hardirq_count() - hardirq_offset)
3352 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3353 else if (softirq_count())
3354 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3356 cpustat->system = cputime64_add(cpustat->system, tmp);
3358 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3360 /* Account for system time used */
3361 acct_update_integrals(p);
3365 * Account for involuntary wait time.
3366 * @steal: the cpu time spent in involuntary wait
3368 void account_steal_time(cputime_t cputime)
3370 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3371 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3373 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3377 * Account for idle time.
3378 * @cputime: the cpu time spent in idle wait
3380 void account_idle_time(cputime_t cputime)
3382 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3383 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3384 struct rq *rq = this_rq();
3386 if (atomic_read(&rq->nr_iowait) > 0)
3387 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3389 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3392 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3395 * Account a single tick of cpu time.
3396 * @p: the process that the cpu time gets accounted to
3397 * @user_tick: indicates if the tick is a user or a system tick
3399 void account_process_tick(struct task_struct *p, int user_tick)
3401 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3402 struct rq *rq = this_rq();
3405 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3406 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3407 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3410 account_idle_time(cputime_one_jiffy);
3414 * Account multiple ticks of steal time.
3415 * @p: the process from which the cpu time has been stolen
3416 * @ticks: number of stolen ticks
3418 void account_steal_ticks(unsigned long ticks)
3420 account_steal_time(jiffies_to_cputime(ticks));
3424 * Account multiple ticks of idle time.
3425 * @ticks: number of stolen ticks
3427 void account_idle_ticks(unsigned long ticks)
3429 account_idle_time(jiffies_to_cputime(ticks));
3435 * Use precise platform statistics if available:
3437 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3438 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3444 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3446 struct task_cputime cputime;
3448 thread_group_cputime(p, &cputime);
3450 *ut = cputime.utime;
3451 *st = cputime.stime;
3455 #ifndef nsecs_to_cputime
3456 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3459 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3461 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3464 * Use CFS's precise accounting:
3466 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3471 temp = (u64)(rtime * utime);
3472 do_div(temp, total);
3473 utime = (cputime_t)temp;
3478 * Compare with previous values, to keep monotonicity:
3480 p->prev_utime = max(p->prev_utime, utime);
3481 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3483 *ut = p->prev_utime;
3484 *st = p->prev_stime;
3488 * Must be called with siglock held.
3490 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3492 struct signal_struct *sig = p->signal;
3493 struct task_cputime cputime;
3494 cputime_t rtime, utime, total;
3496 thread_group_cputime(p, &cputime);
3498 total = cputime_add(cputime.utime, cputime.stime);
3499 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3504 temp = (u64)(rtime * cputime.utime);
3505 do_div(temp, total);
3506 utime = (cputime_t)temp;
3510 sig->prev_utime = max(sig->prev_utime, utime);
3511 sig->prev_stime = max(sig->prev_stime,
3512 cputime_sub(rtime, sig->prev_utime));
3514 *ut = sig->prev_utime;
3515 *st = sig->prev_stime;
3520 * This function gets called by the timer code, with HZ frequency.
3521 * We call it with interrupts disabled.
3523 * It also gets called by the fork code, when changing the parent's
3526 void scheduler_tick(void)
3528 int cpu = smp_processor_id();
3529 struct rq *rq = cpu_rq(cpu);
3530 struct task_struct *curr = rq->curr;
3534 raw_spin_lock(&rq->lock);
3535 update_rq_clock(rq);
3536 update_cpu_load(rq);
3537 curr->sched_class->task_tick(rq, curr, 0);
3538 raw_spin_unlock(&rq->lock);
3540 perf_event_task_tick(curr, cpu);
3543 rq->idle_at_tick = idle_cpu(cpu);
3544 trigger_load_balance(rq, cpu);
3548 notrace unsigned long get_parent_ip(unsigned long addr)
3550 if (in_lock_functions(addr)) {
3551 addr = CALLER_ADDR2;
3552 if (in_lock_functions(addr))
3553 addr = CALLER_ADDR3;
3558 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3559 defined(CONFIG_PREEMPT_TRACER))
3561 void __kprobes add_preempt_count(int val)
3563 #ifdef CONFIG_DEBUG_PREEMPT
3567 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3570 preempt_count() += val;
3571 #ifdef CONFIG_DEBUG_PREEMPT
3573 * Spinlock count overflowing soon?
3575 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3578 if (preempt_count() == val)
3579 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3581 EXPORT_SYMBOL(add_preempt_count);
3583 void __kprobes sub_preempt_count(int val)
3585 #ifdef CONFIG_DEBUG_PREEMPT
3589 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3592 * Is the spinlock portion underflowing?
3594 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3595 !(preempt_count() & PREEMPT_MASK)))
3599 if (preempt_count() == val)
3600 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3601 preempt_count() -= val;
3603 EXPORT_SYMBOL(sub_preempt_count);
3608 * Print scheduling while atomic bug:
3610 static noinline void __schedule_bug(struct task_struct *prev)
3612 struct pt_regs *regs = get_irq_regs();
3614 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3615 prev->comm, prev->pid, preempt_count());
3617 debug_show_held_locks(prev);
3619 if (irqs_disabled())
3620 print_irqtrace_events(prev);
3629 * Various schedule()-time debugging checks and statistics:
3631 static inline void schedule_debug(struct task_struct *prev)
3634 * Test if we are atomic. Since do_exit() needs to call into
3635 * schedule() atomically, we ignore that path for now.
3636 * Otherwise, whine if we are scheduling when we should not be.
3638 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3639 __schedule_bug(prev);
3641 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3643 schedstat_inc(this_rq(), sched_count);
3644 #ifdef CONFIG_SCHEDSTATS
3645 if (unlikely(prev->lock_depth >= 0)) {
3646 schedstat_inc(this_rq(), bkl_count);
3647 schedstat_inc(prev, sched_info.bkl_count);
3652 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3654 if (prev->state == TASK_RUNNING) {
3655 u64 runtime = prev->se.sum_exec_runtime;
3657 runtime -= prev->se.prev_sum_exec_runtime;
3658 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
3661 * In order to avoid avg_overlap growing stale when we are
3662 * indeed overlapping and hence not getting put to sleep, grow
3663 * the avg_overlap on preemption.
3665 * We use the average preemption runtime because that
3666 * correlates to the amount of cache footprint a task can
3669 update_avg(&prev->se.avg_overlap, runtime);
3671 prev->sched_class->put_prev_task(rq, prev);
3675 * Pick up the highest-prio task:
3677 static inline struct task_struct *
3678 pick_next_task(struct rq *rq)
3680 const struct sched_class *class;
3681 struct task_struct *p;
3684 * Optimization: we know that if all tasks are in
3685 * the fair class we can call that function directly:
3687 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3688 p = fair_sched_class.pick_next_task(rq);
3693 class = sched_class_highest;
3695 p = class->pick_next_task(rq);
3699 * Will never be NULL as the idle class always
3700 * returns a non-NULL p:
3702 class = class->next;
3707 * schedule() is the main scheduler function.
3709 asmlinkage void __sched schedule(void)
3711 struct task_struct *prev, *next;
3712 unsigned long *switch_count;
3718 cpu = smp_processor_id();
3722 switch_count = &prev->nivcsw;
3724 release_kernel_lock(prev);
3725 need_resched_nonpreemptible:
3727 schedule_debug(prev);
3729 if (sched_feat(HRTICK))
3732 raw_spin_lock_irq(&rq->lock);
3733 update_rq_clock(rq);
3734 clear_tsk_need_resched(prev);
3736 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3737 if (unlikely(signal_pending_state(prev->state, prev)))
3738 prev->state = TASK_RUNNING;
3740 deactivate_task(rq, prev, 1);
3741 switch_count = &prev->nvcsw;
3744 pre_schedule(rq, prev);
3746 if (unlikely(!rq->nr_running))
3747 idle_balance(cpu, rq);
3749 put_prev_task(rq, prev);
3750 next = pick_next_task(rq);
3752 if (likely(prev != next)) {
3753 sched_info_switch(prev, next);
3754 perf_event_task_sched_out(prev, next, cpu);
3760 context_switch(rq, prev, next); /* unlocks the rq */
3762 * the context switch might have flipped the stack from under
3763 * us, hence refresh the local variables.
3765 cpu = smp_processor_id();
3768 raw_spin_unlock_irq(&rq->lock);
3772 if (unlikely(reacquire_kernel_lock(current) < 0))
3773 goto need_resched_nonpreemptible;
3775 preempt_enable_no_resched();
3779 EXPORT_SYMBOL(schedule);
3781 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3783 * Look out! "owner" is an entirely speculative pointer
3784 * access and not reliable.
3786 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3791 if (!sched_feat(OWNER_SPIN))
3794 #ifdef CONFIG_DEBUG_PAGEALLOC
3796 * Need to access the cpu field knowing that
3797 * DEBUG_PAGEALLOC could have unmapped it if
3798 * the mutex owner just released it and exited.
3800 if (probe_kernel_address(&owner->cpu, cpu))
3807 * Even if the access succeeded (likely case),
3808 * the cpu field may no longer be valid.
3810 if (cpu >= nr_cpumask_bits)
3814 * We need to validate that we can do a
3815 * get_cpu() and that we have the percpu area.
3817 if (!cpu_online(cpu))
3824 * Owner changed, break to re-assess state.
3826 if (lock->owner != owner)
3830 * Is that owner really running on that cpu?
3832 if (task_thread_info(rq->curr) != owner || need_resched())
3842 #ifdef CONFIG_PREEMPT
3844 * this is the entry point to schedule() from in-kernel preemption
3845 * off of preempt_enable. Kernel preemptions off return from interrupt
3846 * occur there and call schedule directly.
3848 asmlinkage void __sched preempt_schedule(void)
3850 struct thread_info *ti = current_thread_info();
3853 * If there is a non-zero preempt_count or interrupts are disabled,
3854 * we do not want to preempt the current task. Just return..
3856 if (likely(ti->preempt_count || irqs_disabled()))
3860 add_preempt_count(PREEMPT_ACTIVE);
3862 sub_preempt_count(PREEMPT_ACTIVE);
3865 * Check again in case we missed a preemption opportunity
3866 * between schedule and now.
3869 } while (need_resched());
3871 EXPORT_SYMBOL(preempt_schedule);
3874 * this is the entry point to schedule() from kernel preemption
3875 * off of irq context.
3876 * Note, that this is called and return with irqs disabled. This will
3877 * protect us against recursive calling from irq.
3879 asmlinkage void __sched preempt_schedule_irq(void)
3881 struct thread_info *ti = current_thread_info();
3883 /* Catch callers which need to be fixed */
3884 BUG_ON(ti->preempt_count || !irqs_disabled());
3887 add_preempt_count(PREEMPT_ACTIVE);
3890 local_irq_disable();
3891 sub_preempt_count(PREEMPT_ACTIVE);
3894 * Check again in case we missed a preemption opportunity
3895 * between schedule and now.
3898 } while (need_resched());
3901 #endif /* CONFIG_PREEMPT */
3903 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3906 return try_to_wake_up(curr->private, mode, wake_flags);
3908 EXPORT_SYMBOL(default_wake_function);
3911 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3912 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3913 * number) then we wake all the non-exclusive tasks and one exclusive task.
3915 * There are circumstances in which we can try to wake a task which has already
3916 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3917 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3919 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3920 int nr_exclusive, int wake_flags, void *key)
3922 wait_queue_t *curr, *next;
3924 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3925 unsigned flags = curr->flags;
3927 if (curr->func(curr, mode, wake_flags, key) &&
3928 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3934 * __wake_up - wake up threads blocked on a waitqueue.
3936 * @mode: which threads
3937 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3938 * @key: is directly passed to the wakeup function
3940 * It may be assumed that this function implies a write memory barrier before
3941 * changing the task state if and only if any tasks are woken up.
3943 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3944 int nr_exclusive, void *key)
3946 unsigned long flags;
3948 spin_lock_irqsave(&q->lock, flags);
3949 __wake_up_common(q, mode, nr_exclusive, 0, key);
3950 spin_unlock_irqrestore(&q->lock, flags);
3952 EXPORT_SYMBOL(__wake_up);
3955 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3957 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3959 __wake_up_common(q, mode, 1, 0, NULL);
3962 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3964 __wake_up_common(q, mode, 1, 0, key);
3968 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3970 * @mode: which threads
3971 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3972 * @key: opaque value to be passed to wakeup targets
3974 * The sync wakeup differs that the waker knows that it will schedule
3975 * away soon, so while the target thread will be woken up, it will not
3976 * be migrated to another CPU - ie. the two threads are 'synchronized'
3977 * with each other. This can prevent needless bouncing between CPUs.
3979 * On UP it can prevent extra preemption.
3981 * It may be assumed that this function implies a write memory barrier before
3982 * changing the task state if and only if any tasks are woken up.
3984 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3985 int nr_exclusive, void *key)
3987 unsigned long flags;
3988 int wake_flags = WF_SYNC;
3993 if (unlikely(!nr_exclusive))
3996 spin_lock_irqsave(&q->lock, flags);
3997 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3998 spin_unlock_irqrestore(&q->lock, flags);
4000 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4003 * __wake_up_sync - see __wake_up_sync_key()
4005 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4007 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4009 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4012 * complete: - signals a single thread waiting on this completion
4013 * @x: holds the state of this particular completion
4015 * This will wake up a single thread waiting on this completion. Threads will be
4016 * awakened in the same order in which they were queued.
4018 * See also complete_all(), wait_for_completion() and related routines.
4020 * It may be assumed that this function implies a write memory barrier before
4021 * changing the task state if and only if any tasks are woken up.
4023 void complete(struct completion *x)
4025 unsigned long flags;
4027 spin_lock_irqsave(&x->wait.lock, flags);
4029 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4030 spin_unlock_irqrestore(&x->wait.lock, flags);
4032 EXPORT_SYMBOL(complete);
4035 * complete_all: - signals all threads waiting on this completion
4036 * @x: holds the state of this particular completion
4038 * This will wake up all threads waiting on this particular completion event.
4040 * It may be assumed that this function implies a write memory barrier before
4041 * changing the task state if and only if any tasks are woken up.
4043 void complete_all(struct completion *x)
4045 unsigned long flags;
4047 spin_lock_irqsave(&x->wait.lock, flags);
4048 x->done += UINT_MAX/2;
4049 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4050 spin_unlock_irqrestore(&x->wait.lock, flags);
4052 EXPORT_SYMBOL(complete_all);
4054 static inline long __sched
4055 do_wait_for_common(struct completion *x, long timeout, int state)
4058 DECLARE_WAITQUEUE(wait, current);
4060 wait.flags |= WQ_FLAG_EXCLUSIVE;
4061 __add_wait_queue_tail(&x->wait, &wait);
4063 if (signal_pending_state(state, current)) {
4064 timeout = -ERESTARTSYS;
4067 __set_current_state(state);
4068 spin_unlock_irq(&x->wait.lock);
4069 timeout = schedule_timeout(timeout);
4070 spin_lock_irq(&x->wait.lock);
4071 } while (!x->done && timeout);
4072 __remove_wait_queue(&x->wait, &wait);
4077 return timeout ?: 1;
4081 wait_for_common(struct completion *x, long timeout, int state)
4085 spin_lock_irq(&x->wait.lock);
4086 timeout = do_wait_for_common(x, timeout, state);
4087 spin_unlock_irq(&x->wait.lock);
4092 * wait_for_completion: - waits for completion of a task
4093 * @x: holds the state of this particular completion
4095 * This waits to be signaled for completion of a specific task. It is NOT
4096 * interruptible and there is no timeout.
4098 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4099 * and interrupt capability. Also see complete().
4101 void __sched wait_for_completion(struct completion *x)
4103 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4105 EXPORT_SYMBOL(wait_for_completion);
4108 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4109 * @x: holds the state of this particular completion
4110 * @timeout: timeout value in jiffies
4112 * This waits for either a completion of a specific task to be signaled or for a
4113 * specified timeout to expire. The timeout is in jiffies. It is not
4116 unsigned long __sched
4117 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4119 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4121 EXPORT_SYMBOL(wait_for_completion_timeout);
4124 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4125 * @x: holds the state of this particular completion
4127 * This waits for completion of a specific task to be signaled. It is
4130 int __sched wait_for_completion_interruptible(struct completion *x)
4132 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4133 if (t == -ERESTARTSYS)
4137 EXPORT_SYMBOL(wait_for_completion_interruptible);
4140 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4141 * @x: holds the state of this particular completion
4142 * @timeout: timeout value in jiffies
4144 * This waits for either a completion of a specific task to be signaled or for a
4145 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4147 unsigned long __sched
4148 wait_for_completion_interruptible_timeout(struct completion *x,
4149 unsigned long timeout)
4151 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4153 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4156 * wait_for_completion_killable: - waits for completion of a task (killable)
4157 * @x: holds the state of this particular completion
4159 * This waits to be signaled for completion of a specific task. It can be
4160 * interrupted by a kill signal.
4162 int __sched wait_for_completion_killable(struct completion *x)
4164 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4165 if (t == -ERESTARTSYS)
4169 EXPORT_SYMBOL(wait_for_completion_killable);
4172 * try_wait_for_completion - try to decrement a completion without blocking
4173 * @x: completion structure
4175 * Returns: 0 if a decrement cannot be done without blocking
4176 * 1 if a decrement succeeded.
4178 * If a completion is being used as a counting completion,
4179 * attempt to decrement the counter without blocking. This
4180 * enables us to avoid waiting if the resource the completion
4181 * is protecting is not available.
4183 bool try_wait_for_completion(struct completion *x)
4185 unsigned long flags;
4188 spin_lock_irqsave(&x->wait.lock, flags);
4193 spin_unlock_irqrestore(&x->wait.lock, flags);
4196 EXPORT_SYMBOL(try_wait_for_completion);
4199 * completion_done - Test to see if a completion has any waiters
4200 * @x: completion structure
4202 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4203 * 1 if there are no waiters.
4206 bool completion_done(struct completion *x)
4208 unsigned long flags;
4211 spin_lock_irqsave(&x->wait.lock, flags);
4214 spin_unlock_irqrestore(&x->wait.lock, flags);
4217 EXPORT_SYMBOL(completion_done);
4220 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4222 unsigned long flags;
4225 init_waitqueue_entry(&wait, current);
4227 __set_current_state(state);
4229 spin_lock_irqsave(&q->lock, flags);
4230 __add_wait_queue(q, &wait);
4231 spin_unlock(&q->lock);
4232 timeout = schedule_timeout(timeout);
4233 spin_lock_irq(&q->lock);
4234 __remove_wait_queue(q, &wait);
4235 spin_unlock_irqrestore(&q->lock, flags);
4240 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4242 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4244 EXPORT_SYMBOL(interruptible_sleep_on);
4247 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4249 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4251 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4253 void __sched sleep_on(wait_queue_head_t *q)
4255 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4257 EXPORT_SYMBOL(sleep_on);
4259 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4261 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4263 EXPORT_SYMBOL(sleep_on_timeout);
4265 #ifdef CONFIG_RT_MUTEXES
4268 * rt_mutex_setprio - set the current priority of a task
4270 * @prio: prio value (kernel-internal form)
4272 * This function changes the 'effective' priority of a task. It does
4273 * not touch ->normal_prio like __setscheduler().
4275 * Used by the rt_mutex code to implement priority inheritance logic.
4277 void rt_mutex_setprio(struct task_struct *p, int prio)
4279 unsigned long flags;
4280 int oldprio, on_rq, running;
4282 const struct sched_class *prev_class = p->sched_class;
4284 BUG_ON(prio < 0 || prio > MAX_PRIO);
4286 rq = task_rq_lock(p, &flags);
4287 update_rq_clock(rq);
4290 on_rq = p->se.on_rq;
4291 running = task_current(rq, p);
4293 dequeue_task(rq, p, 0);
4295 p->sched_class->put_prev_task(rq, p);
4298 p->sched_class = &rt_sched_class;
4300 p->sched_class = &fair_sched_class;
4305 p->sched_class->set_curr_task(rq);
4307 enqueue_task(rq, p, 0);
4309 check_class_changed(rq, p, prev_class, oldprio, running);
4311 task_rq_unlock(rq, &flags);
4316 void set_user_nice(struct task_struct *p, long nice)
4318 int old_prio, delta, on_rq;
4319 unsigned long flags;
4322 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4325 * We have to be careful, if called from sys_setpriority(),
4326 * the task might be in the middle of scheduling on another CPU.
4328 rq = task_rq_lock(p, &flags);
4329 update_rq_clock(rq);
4331 * The RT priorities are set via sched_setscheduler(), but we still
4332 * allow the 'normal' nice value to be set - but as expected
4333 * it wont have any effect on scheduling until the task is
4334 * SCHED_FIFO/SCHED_RR:
4336 if (task_has_rt_policy(p)) {
4337 p->static_prio = NICE_TO_PRIO(nice);
4340 on_rq = p->se.on_rq;
4342 dequeue_task(rq, p, 0);
4344 p->static_prio = NICE_TO_PRIO(nice);
4347 p->prio = effective_prio(p);
4348 delta = p->prio - old_prio;
4351 enqueue_task(rq, p, 0);
4353 * If the task increased its priority or is running and
4354 * lowered its priority, then reschedule its CPU:
4356 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4357 resched_task(rq->curr);
4360 task_rq_unlock(rq, &flags);
4362 EXPORT_SYMBOL(set_user_nice);
4365 * can_nice - check if a task can reduce its nice value
4369 int can_nice(const struct task_struct *p, const int nice)
4371 /* convert nice value [19,-20] to rlimit style value [1,40] */
4372 int nice_rlim = 20 - nice;
4374 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4375 capable(CAP_SYS_NICE));
4378 #ifdef __ARCH_WANT_SYS_NICE
4381 * sys_nice - change the priority of the current process.
4382 * @increment: priority increment
4384 * sys_setpriority is a more generic, but much slower function that
4385 * does similar things.
4387 SYSCALL_DEFINE1(nice, int, increment)
4392 * Setpriority might change our priority at the same moment.
4393 * We don't have to worry. Conceptually one call occurs first
4394 * and we have a single winner.
4396 if (increment < -40)
4401 nice = TASK_NICE(current) + increment;
4407 if (increment < 0 && !can_nice(current, nice))
4410 retval = security_task_setnice(current, nice);
4414 set_user_nice(current, nice);
4421 * task_prio - return the priority value of a given task.
4422 * @p: the task in question.
4424 * This is the priority value as seen by users in /proc.
4425 * RT tasks are offset by -200. Normal tasks are centered
4426 * around 0, value goes from -16 to +15.
4428 int task_prio(const struct task_struct *p)
4430 return p->prio - MAX_RT_PRIO;
4434 * task_nice - return the nice value of a given task.
4435 * @p: the task in question.
4437 int task_nice(const struct task_struct *p)
4439 return TASK_NICE(p);
4441 EXPORT_SYMBOL(task_nice);
4444 * idle_cpu - is a given cpu idle currently?
4445 * @cpu: the processor in question.
4447 int idle_cpu(int cpu)
4449 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4453 * idle_task - return the idle task for a given cpu.
4454 * @cpu: the processor in question.
4456 struct task_struct *idle_task(int cpu)
4458 return cpu_rq(cpu)->idle;
4462 * find_process_by_pid - find a process with a matching PID value.
4463 * @pid: the pid in question.
4465 static struct task_struct *find_process_by_pid(pid_t pid)
4467 return pid ? find_task_by_vpid(pid) : current;
4470 /* Actually do priority change: must hold rq lock. */
4472 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4474 BUG_ON(p->se.on_rq);
4477 p->rt_priority = prio;
4478 p->normal_prio = normal_prio(p);
4479 /* we are holding p->pi_lock already */
4480 p->prio = rt_mutex_getprio(p);
4481 if (rt_prio(p->prio))
4482 p->sched_class = &rt_sched_class;
4484 p->sched_class = &fair_sched_class;
4489 * check the target process has a UID that matches the current process's
4491 static bool check_same_owner(struct task_struct *p)
4493 const struct cred *cred = current_cred(), *pcred;
4497 pcred = __task_cred(p);
4498 match = (cred->euid == pcred->euid ||
4499 cred->euid == pcred->uid);
4504 static int __sched_setscheduler(struct task_struct *p, int policy,
4505 struct sched_param *param, bool user)
4507 int retval, oldprio, oldpolicy = -1, on_rq, running;
4508 unsigned long flags;
4509 const struct sched_class *prev_class = p->sched_class;
4513 /* may grab non-irq protected spin_locks */
4514 BUG_ON(in_interrupt());
4516 /* double check policy once rq lock held */
4518 reset_on_fork = p->sched_reset_on_fork;
4519 policy = oldpolicy = p->policy;
4521 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4522 policy &= ~SCHED_RESET_ON_FORK;
4524 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4525 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4526 policy != SCHED_IDLE)
4531 * Valid priorities for SCHED_FIFO and SCHED_RR are
4532 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4533 * SCHED_BATCH and SCHED_IDLE is 0.
4535 if (param->sched_priority < 0 ||
4536 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4537 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4539 if (rt_policy(policy) != (param->sched_priority != 0))
4543 * Allow unprivileged RT tasks to decrease priority:
4545 if (user && !capable(CAP_SYS_NICE)) {
4546 if (rt_policy(policy)) {
4547 unsigned long rlim_rtprio;
4549 if (!lock_task_sighand(p, &flags))
4551 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4552 unlock_task_sighand(p, &flags);
4554 /* can't set/change the rt policy */
4555 if (policy != p->policy && !rlim_rtprio)
4558 /* can't increase priority */
4559 if (param->sched_priority > p->rt_priority &&
4560 param->sched_priority > rlim_rtprio)
4564 * Like positive nice levels, dont allow tasks to
4565 * move out of SCHED_IDLE either:
4567 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4570 /* can't change other user's priorities */
4571 if (!check_same_owner(p))
4574 /* Normal users shall not reset the sched_reset_on_fork flag */
4575 if (p->sched_reset_on_fork && !reset_on_fork)
4580 #ifdef CONFIG_RT_GROUP_SCHED
4582 * Do not allow realtime tasks into groups that have no runtime
4585 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4586 task_group(p)->rt_bandwidth.rt_runtime == 0)
4590 retval = security_task_setscheduler(p, policy, param);
4596 * make sure no PI-waiters arrive (or leave) while we are
4597 * changing the priority of the task:
4599 raw_spin_lock_irqsave(&p->pi_lock, flags);
4601 * To be able to change p->policy safely, the apropriate
4602 * runqueue lock must be held.
4604 rq = __task_rq_lock(p);
4605 /* recheck policy now with rq lock held */
4606 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4607 policy = oldpolicy = -1;
4608 __task_rq_unlock(rq);
4609 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4612 update_rq_clock(rq);
4613 on_rq = p->se.on_rq;
4614 running = task_current(rq, p);
4616 deactivate_task(rq, p, 0);
4618 p->sched_class->put_prev_task(rq, p);
4620 p->sched_reset_on_fork = reset_on_fork;
4623 __setscheduler(rq, p, policy, param->sched_priority);
4626 p->sched_class->set_curr_task(rq);
4628 activate_task(rq, p, 0);
4630 check_class_changed(rq, p, prev_class, oldprio, running);
4632 __task_rq_unlock(rq);
4633 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4635 rt_mutex_adjust_pi(p);
4641 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4642 * @p: the task in question.
4643 * @policy: new policy.
4644 * @param: structure containing the new RT priority.
4646 * NOTE that the task may be already dead.
4648 int sched_setscheduler(struct task_struct *p, int policy,
4649 struct sched_param *param)
4651 return __sched_setscheduler(p, policy, param, true);
4653 EXPORT_SYMBOL_GPL(sched_setscheduler);
4656 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4657 * @p: the task in question.
4658 * @policy: new policy.
4659 * @param: structure containing the new RT priority.
4661 * Just like sched_setscheduler, only don't bother checking if the
4662 * current context has permission. For example, this is needed in
4663 * stop_machine(): we create temporary high priority worker threads,
4664 * but our caller might not have that capability.
4666 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4667 struct sched_param *param)
4669 return __sched_setscheduler(p, policy, param, false);
4673 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4675 struct sched_param lparam;
4676 struct task_struct *p;
4679 if (!param || pid < 0)
4681 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4686 p = find_process_by_pid(pid);
4688 retval = sched_setscheduler(p, policy, &lparam);
4695 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4696 * @pid: the pid in question.
4697 * @policy: new policy.
4698 * @param: structure containing the new RT priority.
4700 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4701 struct sched_param __user *, param)
4703 /* negative values for policy are not valid */
4707 return do_sched_setscheduler(pid, policy, param);
4711 * sys_sched_setparam - set/change the RT priority of a thread
4712 * @pid: the pid in question.
4713 * @param: structure containing the new RT priority.
4715 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4717 return do_sched_setscheduler(pid, -1, param);
4721 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4722 * @pid: the pid in question.
4724 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4726 struct task_struct *p;
4734 p = find_process_by_pid(pid);
4736 retval = security_task_getscheduler(p);
4739 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4746 * sys_sched_getparam - get the RT priority of a thread
4747 * @pid: the pid in question.
4748 * @param: structure containing the RT priority.
4750 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4752 struct sched_param lp;
4753 struct task_struct *p;
4756 if (!param || pid < 0)
4760 p = find_process_by_pid(pid);
4765 retval = security_task_getscheduler(p);
4769 lp.sched_priority = p->rt_priority;
4773 * This one might sleep, we cannot do it with a spinlock held ...
4775 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4784 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4786 cpumask_var_t cpus_allowed, new_mask;
4787 struct task_struct *p;
4793 p = find_process_by_pid(pid);
4800 /* Prevent p going away */
4804 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4808 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4810 goto out_free_cpus_allowed;
4813 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4816 retval = security_task_setscheduler(p, 0, NULL);
4820 cpuset_cpus_allowed(p, cpus_allowed);
4821 cpumask_and(new_mask, in_mask, cpus_allowed);
4823 retval = set_cpus_allowed_ptr(p, new_mask);
4826 cpuset_cpus_allowed(p, cpus_allowed);
4827 if (!cpumask_subset(new_mask, cpus_allowed)) {
4829 * We must have raced with a concurrent cpuset
4830 * update. Just reset the cpus_allowed to the
4831 * cpuset's cpus_allowed
4833 cpumask_copy(new_mask, cpus_allowed);
4838 free_cpumask_var(new_mask);
4839 out_free_cpus_allowed:
4840 free_cpumask_var(cpus_allowed);
4847 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4848 struct cpumask *new_mask)
4850 if (len < cpumask_size())
4851 cpumask_clear(new_mask);
4852 else if (len > cpumask_size())
4853 len = cpumask_size();
4855 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4859 * sys_sched_setaffinity - set the cpu affinity of a process
4860 * @pid: pid of the process
4861 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4862 * @user_mask_ptr: user-space pointer to the new cpu mask
4864 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4865 unsigned long __user *, user_mask_ptr)
4867 cpumask_var_t new_mask;
4870 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4873 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4875 retval = sched_setaffinity(pid, new_mask);
4876 free_cpumask_var(new_mask);
4880 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4882 struct task_struct *p;
4883 unsigned long flags;
4891 p = find_process_by_pid(pid);
4895 retval = security_task_getscheduler(p);
4899 rq = task_rq_lock(p, &flags);
4900 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4901 task_rq_unlock(rq, &flags);
4911 * sys_sched_getaffinity - get the cpu affinity of a process
4912 * @pid: pid of the process
4913 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4914 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4916 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4917 unsigned long __user *, user_mask_ptr)
4922 if (len < cpumask_size())
4925 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4928 ret = sched_getaffinity(pid, mask);
4930 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
4933 ret = cpumask_size();
4935 free_cpumask_var(mask);
4941 * sys_sched_yield - yield the current processor to other threads.
4943 * This function yields the current CPU to other tasks. If there are no
4944 * other threads running on this CPU then this function will return.
4946 SYSCALL_DEFINE0(sched_yield)
4948 struct rq *rq = this_rq_lock();
4950 schedstat_inc(rq, yld_count);
4951 current->sched_class->yield_task(rq);
4954 * Since we are going to call schedule() anyway, there's
4955 * no need to preempt or enable interrupts:
4957 __release(rq->lock);
4958 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4959 do_raw_spin_unlock(&rq->lock);
4960 preempt_enable_no_resched();
4967 static inline int should_resched(void)
4969 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4972 static void __cond_resched(void)
4974 add_preempt_count(PREEMPT_ACTIVE);
4976 sub_preempt_count(PREEMPT_ACTIVE);
4979 int __sched _cond_resched(void)
4981 if (should_resched()) {
4987 EXPORT_SYMBOL(_cond_resched);
4990 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4991 * call schedule, and on return reacquire the lock.
4993 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4994 * operations here to prevent schedule() from being called twice (once via
4995 * spin_unlock(), once by hand).
4997 int __cond_resched_lock(spinlock_t *lock)
4999 int resched = should_resched();
5002 lockdep_assert_held(lock);
5004 if (spin_needbreak(lock) || resched) {
5015 EXPORT_SYMBOL(__cond_resched_lock);
5017 int __sched __cond_resched_softirq(void)
5019 BUG_ON(!in_softirq());
5021 if (should_resched()) {
5029 EXPORT_SYMBOL(__cond_resched_softirq);
5032 * yield - yield the current processor to other threads.
5034 * This is a shortcut for kernel-space yielding - it marks the
5035 * thread runnable and calls sys_sched_yield().
5037 void __sched yield(void)
5039 set_current_state(TASK_RUNNING);
5042 EXPORT_SYMBOL(yield);
5045 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5046 * that process accounting knows that this is a task in IO wait state.
5048 void __sched io_schedule(void)
5050 struct rq *rq = raw_rq();
5052 delayacct_blkio_start();
5053 atomic_inc(&rq->nr_iowait);
5054 current->in_iowait = 1;
5056 current->in_iowait = 0;
5057 atomic_dec(&rq->nr_iowait);
5058 delayacct_blkio_end();
5060 EXPORT_SYMBOL(io_schedule);
5062 long __sched io_schedule_timeout(long timeout)
5064 struct rq *rq = raw_rq();
5067 delayacct_blkio_start();
5068 atomic_inc(&rq->nr_iowait);
5069 current->in_iowait = 1;
5070 ret = schedule_timeout(timeout);
5071 current->in_iowait = 0;
5072 atomic_dec(&rq->nr_iowait);
5073 delayacct_blkio_end();
5078 * sys_sched_get_priority_max - return maximum RT priority.
5079 * @policy: scheduling class.
5081 * this syscall returns the maximum rt_priority that can be used
5082 * by a given scheduling class.
5084 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5091 ret = MAX_USER_RT_PRIO-1;
5103 * sys_sched_get_priority_min - return minimum RT priority.
5104 * @policy: scheduling class.
5106 * this syscall returns the minimum rt_priority that can be used
5107 * by a given scheduling class.
5109 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5127 * sys_sched_rr_get_interval - return the default timeslice of a process.
5128 * @pid: pid of the process.
5129 * @interval: userspace pointer to the timeslice value.
5131 * this syscall writes the default timeslice value of a given process
5132 * into the user-space timespec buffer. A value of '0' means infinity.
5134 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5135 struct timespec __user *, interval)
5137 struct task_struct *p;
5138 unsigned int time_slice;
5139 unsigned long flags;
5149 p = find_process_by_pid(pid);
5153 retval = security_task_getscheduler(p);
5157 rq = task_rq_lock(p, &flags);
5158 time_slice = p->sched_class->get_rr_interval(rq, p);
5159 task_rq_unlock(rq, &flags);
5162 jiffies_to_timespec(time_slice, &t);
5163 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5171 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5173 void sched_show_task(struct task_struct *p)
5175 unsigned long free = 0;
5178 state = p->state ? __ffs(p->state) + 1 : 0;
5179 printk(KERN_INFO "%-13.13s %c", p->comm,
5180 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5181 #if BITS_PER_LONG == 32
5182 if (state == TASK_RUNNING)
5183 printk(KERN_CONT " running ");
5185 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5187 if (state == TASK_RUNNING)
5188 printk(KERN_CONT " running task ");
5190 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5192 #ifdef CONFIG_DEBUG_STACK_USAGE
5193 free = stack_not_used(p);
5195 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5196 task_pid_nr(p), task_pid_nr(p->real_parent),
5197 (unsigned long)task_thread_info(p)->flags);
5199 show_stack(p, NULL);
5202 void show_state_filter(unsigned long state_filter)
5204 struct task_struct *g, *p;
5206 #if BITS_PER_LONG == 32
5208 " task PC stack pid father\n");
5211 " task PC stack pid father\n");
5213 read_lock(&tasklist_lock);
5214 do_each_thread(g, p) {
5216 * reset the NMI-timeout, listing all files on a slow
5217 * console might take alot of time:
5219 touch_nmi_watchdog();
5220 if (!state_filter || (p->state & state_filter))
5222 } while_each_thread(g, p);
5224 touch_all_softlockup_watchdogs();
5226 #ifdef CONFIG_SCHED_DEBUG
5227 sysrq_sched_debug_show();
5229 read_unlock(&tasklist_lock);
5231 * Only show locks if all tasks are dumped:
5234 debug_show_all_locks();
5237 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5239 idle->sched_class = &idle_sched_class;
5243 * init_idle - set up an idle thread for a given CPU
5244 * @idle: task in question
5245 * @cpu: cpu the idle task belongs to
5247 * NOTE: this function does not set the idle thread's NEED_RESCHED
5248 * flag, to make booting more robust.
5250 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5252 struct rq *rq = cpu_rq(cpu);
5253 unsigned long flags;
5255 raw_spin_lock_irqsave(&rq->lock, flags);
5258 idle->state = TASK_RUNNING;
5259 idle->se.exec_start = sched_clock();
5261 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5262 __set_task_cpu(idle, cpu);
5264 rq->curr = rq->idle = idle;
5265 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5268 raw_spin_unlock_irqrestore(&rq->lock, flags);
5270 /* Set the preempt count _outside_ the spinlocks! */
5271 #if defined(CONFIG_PREEMPT)
5272 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5274 task_thread_info(idle)->preempt_count = 0;
5277 * The idle tasks have their own, simple scheduling class:
5279 idle->sched_class = &idle_sched_class;
5280 ftrace_graph_init_task(idle);
5284 * In a system that switches off the HZ timer nohz_cpu_mask
5285 * indicates which cpus entered this state. This is used
5286 * in the rcu update to wait only for active cpus. For system
5287 * which do not switch off the HZ timer nohz_cpu_mask should
5288 * always be CPU_BITS_NONE.
5290 cpumask_var_t nohz_cpu_mask;
5293 * Increase the granularity value when there are more CPUs,
5294 * because with more CPUs the 'effective latency' as visible
5295 * to users decreases. But the relationship is not linear,
5296 * so pick a second-best guess by going with the log2 of the
5299 * This idea comes from the SD scheduler of Con Kolivas:
5301 static int get_update_sysctl_factor(void)
5303 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5304 unsigned int factor;
5306 switch (sysctl_sched_tunable_scaling) {
5307 case SCHED_TUNABLESCALING_NONE:
5310 case SCHED_TUNABLESCALING_LINEAR:
5313 case SCHED_TUNABLESCALING_LOG:
5315 factor = 1 + ilog2(cpus);
5322 static void update_sysctl(void)
5324 unsigned int factor = get_update_sysctl_factor();
5326 #define SET_SYSCTL(name) \
5327 (sysctl_##name = (factor) * normalized_sysctl_##name)
5328 SET_SYSCTL(sched_min_granularity);
5329 SET_SYSCTL(sched_latency);
5330 SET_SYSCTL(sched_wakeup_granularity);
5331 SET_SYSCTL(sched_shares_ratelimit);
5335 static inline void sched_init_granularity(void)
5342 * This is how migration works:
5344 * 1) we queue a struct migration_req structure in the source CPU's
5345 * runqueue and wake up that CPU's migration thread.
5346 * 2) we down() the locked semaphore => thread blocks.
5347 * 3) migration thread wakes up (implicitly it forces the migrated
5348 * thread off the CPU)
5349 * 4) it gets the migration request and checks whether the migrated
5350 * task is still in the wrong runqueue.
5351 * 5) if it's in the wrong runqueue then the migration thread removes
5352 * it and puts it into the right queue.
5353 * 6) migration thread up()s the semaphore.
5354 * 7) we wake up and the migration is done.
5358 * Change a given task's CPU affinity. Migrate the thread to a
5359 * proper CPU and schedule it away if the CPU it's executing on
5360 * is removed from the allowed bitmask.
5362 * NOTE: the caller must have a valid reference to the task, the
5363 * task must not exit() & deallocate itself prematurely. The
5364 * call is not atomic; no spinlocks may be held.
5366 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5368 struct migration_req req;
5369 unsigned long flags;
5374 * Since we rely on wake-ups to migrate sleeping tasks, don't change
5375 * the ->cpus_allowed mask from under waking tasks, which would be
5376 * possible when we change rq->lock in ttwu(), so synchronize against
5377 * TASK_WAKING to avoid that.
5380 while (p->state == TASK_WAKING)
5383 rq = task_rq_lock(p, &flags);
5385 if (p->state == TASK_WAKING) {
5386 task_rq_unlock(rq, &flags);
5390 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5395 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5396 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5401 if (p->sched_class->set_cpus_allowed)
5402 p->sched_class->set_cpus_allowed(p, new_mask);
5404 cpumask_copy(&p->cpus_allowed, new_mask);
5405 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5408 /* Can the task run on the task's current CPU? If so, we're done */
5409 if (cpumask_test_cpu(task_cpu(p), new_mask))
5412 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
5413 /* Need help from migration thread: drop lock and wait. */
5414 struct task_struct *mt = rq->migration_thread;
5416 get_task_struct(mt);
5417 task_rq_unlock(rq, &flags);
5418 wake_up_process(rq->migration_thread);
5419 put_task_struct(mt);
5420 wait_for_completion(&req.done);
5421 tlb_migrate_finish(p->mm);
5425 task_rq_unlock(rq, &flags);
5429 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5432 * Move (not current) task off this cpu, onto dest cpu. We're doing
5433 * this because either it can't run here any more (set_cpus_allowed()
5434 * away from this CPU, or CPU going down), or because we're
5435 * attempting to rebalance this task on exec (sched_exec).
5437 * So we race with normal scheduler movements, but that's OK, as long
5438 * as the task is no longer on this CPU.
5440 * Returns non-zero if task was successfully migrated.
5442 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5444 struct rq *rq_dest, *rq_src;
5447 if (unlikely(!cpu_active(dest_cpu)))
5450 rq_src = cpu_rq(src_cpu);
5451 rq_dest = cpu_rq(dest_cpu);
5453 double_rq_lock(rq_src, rq_dest);
5454 /* Already moved. */
5455 if (task_cpu(p) != src_cpu)
5457 /* Affinity changed (again). */
5458 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5462 * If we're not on a rq, the next wake-up will ensure we're
5466 deactivate_task(rq_src, p, 0);
5467 set_task_cpu(p, dest_cpu);
5468 activate_task(rq_dest, p, 0);
5469 check_preempt_curr(rq_dest, p, 0);
5474 double_rq_unlock(rq_src, rq_dest);
5478 #define RCU_MIGRATION_IDLE 0
5479 #define RCU_MIGRATION_NEED_QS 1
5480 #define RCU_MIGRATION_GOT_QS 2
5481 #define RCU_MIGRATION_MUST_SYNC 3
5484 * migration_thread - this is a highprio system thread that performs
5485 * thread migration by bumping thread off CPU then 'pushing' onto
5488 static int migration_thread(void *data)
5491 int cpu = (long)data;
5495 BUG_ON(rq->migration_thread != current);
5497 set_current_state(TASK_INTERRUPTIBLE);
5498 while (!kthread_should_stop()) {
5499 struct migration_req *req;
5500 struct list_head *head;
5502 raw_spin_lock_irq(&rq->lock);
5504 if (cpu_is_offline(cpu)) {
5505 raw_spin_unlock_irq(&rq->lock);
5509 if (rq->active_balance) {
5510 active_load_balance(rq, cpu);
5511 rq->active_balance = 0;
5514 head = &rq->migration_queue;
5516 if (list_empty(head)) {
5517 raw_spin_unlock_irq(&rq->lock);
5519 set_current_state(TASK_INTERRUPTIBLE);
5522 req = list_entry(head->next, struct migration_req, list);
5523 list_del_init(head->next);
5525 if (req->task != NULL) {
5526 raw_spin_unlock(&rq->lock);
5527 __migrate_task(req->task, cpu, req->dest_cpu);
5528 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
5529 req->dest_cpu = RCU_MIGRATION_GOT_QS;
5530 raw_spin_unlock(&rq->lock);
5532 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
5533 raw_spin_unlock(&rq->lock);
5534 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
5538 complete(&req->done);
5540 __set_current_state(TASK_RUNNING);
5545 #ifdef CONFIG_HOTPLUG_CPU
5547 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5551 local_irq_disable();
5552 ret = __migrate_task(p, src_cpu, dest_cpu);
5558 * Figure out where task on dead CPU should go, use force if necessary.
5560 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5565 dest_cpu = select_fallback_rq(dead_cpu, p);
5567 /* It can have affinity changed while we were choosing. */
5568 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
5573 * While a dead CPU has no uninterruptible tasks queued at this point,
5574 * it might still have a nonzero ->nr_uninterruptible counter, because
5575 * for performance reasons the counter is not stricly tracking tasks to
5576 * their home CPUs. So we just add the counter to another CPU's counter,
5577 * to keep the global sum constant after CPU-down:
5579 static void migrate_nr_uninterruptible(struct rq *rq_src)
5581 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5582 unsigned long flags;
5584 local_irq_save(flags);
5585 double_rq_lock(rq_src, rq_dest);
5586 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5587 rq_src->nr_uninterruptible = 0;
5588 double_rq_unlock(rq_src, rq_dest);
5589 local_irq_restore(flags);
5592 /* Run through task list and migrate tasks from the dead cpu. */
5593 static void migrate_live_tasks(int src_cpu)
5595 struct task_struct *p, *t;
5597 read_lock(&tasklist_lock);
5599 do_each_thread(t, p) {
5603 if (task_cpu(p) == src_cpu)
5604 move_task_off_dead_cpu(src_cpu, p);
5605 } while_each_thread(t, p);
5607 read_unlock(&tasklist_lock);
5611 * Schedules idle task to be the next runnable task on current CPU.
5612 * It does so by boosting its priority to highest possible.
5613 * Used by CPU offline code.
5615 void sched_idle_next(void)
5617 int this_cpu = smp_processor_id();
5618 struct rq *rq = cpu_rq(this_cpu);
5619 struct task_struct *p = rq->idle;
5620 unsigned long flags;
5622 /* cpu has to be offline */
5623 BUG_ON(cpu_online(this_cpu));
5626 * Strictly not necessary since rest of the CPUs are stopped by now
5627 * and interrupts disabled on the current cpu.
5629 raw_spin_lock_irqsave(&rq->lock, flags);
5631 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5633 update_rq_clock(rq);
5634 activate_task(rq, p, 0);
5636 raw_spin_unlock_irqrestore(&rq->lock, flags);
5640 * Ensures that the idle task is using init_mm right before its cpu goes
5643 void idle_task_exit(void)
5645 struct mm_struct *mm = current->active_mm;
5647 BUG_ON(cpu_online(smp_processor_id()));
5650 switch_mm(mm, &init_mm, current);
5654 /* called under rq->lock with disabled interrupts */
5655 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5657 struct rq *rq = cpu_rq(dead_cpu);
5659 /* Must be exiting, otherwise would be on tasklist. */
5660 BUG_ON(!p->exit_state);
5662 /* Cannot have done final schedule yet: would have vanished. */
5663 BUG_ON(p->state == TASK_DEAD);
5668 * Drop lock around migration; if someone else moves it,
5669 * that's OK. No task can be added to this CPU, so iteration is
5672 raw_spin_unlock_irq(&rq->lock);
5673 move_task_off_dead_cpu(dead_cpu, p);
5674 raw_spin_lock_irq(&rq->lock);
5679 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5680 static void migrate_dead_tasks(unsigned int dead_cpu)
5682 struct rq *rq = cpu_rq(dead_cpu);
5683 struct task_struct *next;
5686 if (!rq->nr_running)
5688 update_rq_clock(rq);
5689 next = pick_next_task(rq);
5692 next->sched_class->put_prev_task(rq, next);
5693 migrate_dead(dead_cpu, next);
5699 * remove the tasks which were accounted by rq from calc_load_tasks.
5701 static void calc_global_load_remove(struct rq *rq)
5703 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5704 rq->calc_load_active = 0;
5706 #endif /* CONFIG_HOTPLUG_CPU */
5708 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5710 static struct ctl_table sd_ctl_dir[] = {
5712 .procname = "sched_domain",
5718 static struct ctl_table sd_ctl_root[] = {
5720 .procname = "kernel",
5722 .child = sd_ctl_dir,
5727 static struct ctl_table *sd_alloc_ctl_entry(int n)
5729 struct ctl_table *entry =
5730 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5735 static void sd_free_ctl_entry(struct ctl_table **tablep)
5737 struct ctl_table *entry;
5740 * In the intermediate directories, both the child directory and
5741 * procname are dynamically allocated and could fail but the mode
5742 * will always be set. In the lowest directory the names are
5743 * static strings and all have proc handlers.
5745 for (entry = *tablep; entry->mode; entry++) {
5747 sd_free_ctl_entry(&entry->child);
5748 if (entry->proc_handler == NULL)
5749 kfree(entry->procname);
5757 set_table_entry(struct ctl_table *entry,
5758 const char *procname, void *data, int maxlen,
5759 mode_t mode, proc_handler *proc_handler)
5761 entry->procname = procname;
5763 entry->maxlen = maxlen;
5765 entry->proc_handler = proc_handler;
5768 static struct ctl_table *
5769 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5771 struct ctl_table *table = sd_alloc_ctl_entry(13);
5776 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5777 sizeof(long), 0644, proc_doulongvec_minmax);
5778 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5779 sizeof(long), 0644, proc_doulongvec_minmax);
5780 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5781 sizeof(int), 0644, proc_dointvec_minmax);
5782 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5783 sizeof(int), 0644, proc_dointvec_minmax);
5784 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5785 sizeof(int), 0644, proc_dointvec_minmax);
5786 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5787 sizeof(int), 0644, proc_dointvec_minmax);
5788 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5789 sizeof(int), 0644, proc_dointvec_minmax);
5790 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5791 sizeof(int), 0644, proc_dointvec_minmax);
5792 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5793 sizeof(int), 0644, proc_dointvec_minmax);
5794 set_table_entry(&table[9], "cache_nice_tries",
5795 &sd->cache_nice_tries,
5796 sizeof(int), 0644, proc_dointvec_minmax);
5797 set_table_entry(&table[10], "flags", &sd->flags,
5798 sizeof(int), 0644, proc_dointvec_minmax);
5799 set_table_entry(&table[11], "name", sd->name,
5800 CORENAME_MAX_SIZE, 0444, proc_dostring);
5801 /* &table[12] is terminator */
5806 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5808 struct ctl_table *entry, *table;
5809 struct sched_domain *sd;
5810 int domain_num = 0, i;
5813 for_each_domain(cpu, sd)
5815 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5820 for_each_domain(cpu, sd) {
5821 snprintf(buf, 32, "domain%d", i);
5822 entry->procname = kstrdup(buf, GFP_KERNEL);
5824 entry->child = sd_alloc_ctl_domain_table(sd);
5831 static struct ctl_table_header *sd_sysctl_header;
5832 static void register_sched_domain_sysctl(void)
5834 int i, cpu_num = num_possible_cpus();
5835 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5838 WARN_ON(sd_ctl_dir[0].child);
5839 sd_ctl_dir[0].child = entry;
5844 for_each_possible_cpu(i) {
5845 snprintf(buf, 32, "cpu%d", i);
5846 entry->procname = kstrdup(buf, GFP_KERNEL);
5848 entry->child = sd_alloc_ctl_cpu_table(i);
5852 WARN_ON(sd_sysctl_header);
5853 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5856 /* may be called multiple times per register */
5857 static void unregister_sched_domain_sysctl(void)
5859 if (sd_sysctl_header)
5860 unregister_sysctl_table(sd_sysctl_header);
5861 sd_sysctl_header = NULL;
5862 if (sd_ctl_dir[0].child)
5863 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5866 static void register_sched_domain_sysctl(void)
5869 static void unregister_sched_domain_sysctl(void)
5874 static void set_rq_online(struct rq *rq)
5877 const struct sched_class *class;
5879 cpumask_set_cpu(rq->cpu, rq->rd->online);
5882 for_each_class(class) {
5883 if (class->rq_online)
5884 class->rq_online(rq);
5889 static void set_rq_offline(struct rq *rq)
5892 const struct sched_class *class;
5894 for_each_class(class) {
5895 if (class->rq_offline)
5896 class->rq_offline(rq);
5899 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5905 * migration_call - callback that gets triggered when a CPU is added.
5906 * Here we can start up the necessary migration thread for the new CPU.
5908 static int __cpuinit
5909 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5911 struct task_struct *p;
5912 int cpu = (long)hcpu;
5913 unsigned long flags;
5918 case CPU_UP_PREPARE:
5919 case CPU_UP_PREPARE_FROZEN:
5920 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5923 kthread_bind(p, cpu);
5924 /* Must be high prio: stop_machine expects to yield to it. */
5925 rq = task_rq_lock(p, &flags);
5926 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5927 task_rq_unlock(rq, &flags);
5929 cpu_rq(cpu)->migration_thread = p;
5930 rq->calc_load_update = calc_load_update;
5934 case CPU_ONLINE_FROZEN:
5935 /* Strictly unnecessary, as first user will wake it. */
5936 wake_up_process(cpu_rq(cpu)->migration_thread);
5938 /* Update our root-domain */
5940 raw_spin_lock_irqsave(&rq->lock, flags);
5942 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5946 raw_spin_unlock_irqrestore(&rq->lock, flags);
5949 #ifdef CONFIG_HOTPLUG_CPU
5950 case CPU_UP_CANCELED:
5951 case CPU_UP_CANCELED_FROZEN:
5952 if (!cpu_rq(cpu)->migration_thread)
5954 /* Unbind it from offline cpu so it can run. Fall thru. */
5955 kthread_bind(cpu_rq(cpu)->migration_thread,
5956 cpumask_any(cpu_online_mask));
5957 kthread_stop(cpu_rq(cpu)->migration_thread);
5958 put_task_struct(cpu_rq(cpu)->migration_thread);
5959 cpu_rq(cpu)->migration_thread = NULL;
5963 case CPU_DEAD_FROZEN:
5964 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5965 migrate_live_tasks(cpu);
5967 kthread_stop(rq->migration_thread);
5968 put_task_struct(rq->migration_thread);
5969 rq->migration_thread = NULL;
5970 /* Idle task back to normal (off runqueue, low prio) */
5971 raw_spin_lock_irq(&rq->lock);
5972 update_rq_clock(rq);
5973 deactivate_task(rq, rq->idle, 0);
5974 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5975 rq->idle->sched_class = &idle_sched_class;
5976 migrate_dead_tasks(cpu);
5977 raw_spin_unlock_irq(&rq->lock);
5979 migrate_nr_uninterruptible(rq);
5980 BUG_ON(rq->nr_running != 0);
5981 calc_global_load_remove(rq);
5983 * No need to migrate the tasks: it was best-effort if
5984 * they didn't take sched_hotcpu_mutex. Just wake up
5987 raw_spin_lock_irq(&rq->lock);
5988 while (!list_empty(&rq->migration_queue)) {
5989 struct migration_req *req;
5991 req = list_entry(rq->migration_queue.next,
5992 struct migration_req, list);
5993 list_del_init(&req->list);
5994 raw_spin_unlock_irq(&rq->lock);
5995 complete(&req->done);
5996 raw_spin_lock_irq(&rq->lock);
5998 raw_spin_unlock_irq(&rq->lock);
6002 case CPU_DYING_FROZEN:
6003 /* Update our root-domain */
6005 raw_spin_lock_irqsave(&rq->lock, flags);
6007 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6010 raw_spin_unlock_irqrestore(&rq->lock, flags);
6018 * Register at high priority so that task migration (migrate_all_tasks)
6019 * happens before everything else. This has to be lower priority than
6020 * the notifier in the perf_event subsystem, though.
6022 static struct notifier_block __cpuinitdata migration_notifier = {
6023 .notifier_call = migration_call,
6027 static int __init migration_init(void)
6029 void *cpu = (void *)(long)smp_processor_id();
6032 /* Start one for the boot CPU: */
6033 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6034 BUG_ON(err == NOTIFY_BAD);
6035 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6036 register_cpu_notifier(&migration_notifier);
6040 early_initcall(migration_init);
6045 #ifdef CONFIG_SCHED_DEBUG
6047 static __read_mostly int sched_domain_debug_enabled;
6049 static int __init sched_domain_debug_setup(char *str)
6051 sched_domain_debug_enabled = 1;
6055 early_param("sched_debug", sched_domain_debug_setup);
6057 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6058 struct cpumask *groupmask)
6060 struct sched_group *group = sd->groups;
6063 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6064 cpumask_clear(groupmask);
6066 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6068 if (!(sd->flags & SD_LOAD_BALANCE)) {
6069 printk("does not load-balance\n");
6071 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6076 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6078 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6079 printk(KERN_ERR "ERROR: domain->span does not contain "
6082 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6083 printk(KERN_ERR "ERROR: domain->groups does not contain"
6087 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6091 printk(KERN_ERR "ERROR: group is NULL\n");
6095 if (!group->cpu_power) {
6096 printk(KERN_CONT "\n");
6097 printk(KERN_ERR "ERROR: domain->cpu_power not "
6102 if (!cpumask_weight(sched_group_cpus(group))) {
6103 printk(KERN_CONT "\n");
6104 printk(KERN_ERR "ERROR: empty group\n");
6108 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6109 printk(KERN_CONT "\n");
6110 printk(KERN_ERR "ERROR: repeated CPUs\n");
6114 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6116 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6118 printk(KERN_CONT " %s", str);
6119 if (group->cpu_power != SCHED_LOAD_SCALE) {
6120 printk(KERN_CONT " (cpu_power = %d)",
6124 group = group->next;
6125 } while (group != sd->groups);
6126 printk(KERN_CONT "\n");
6128 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6129 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6132 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6133 printk(KERN_ERR "ERROR: parent span is not a superset "
6134 "of domain->span\n");
6138 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6140 cpumask_var_t groupmask;
6143 if (!sched_domain_debug_enabled)
6147 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6151 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6153 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6154 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6159 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6166 free_cpumask_var(groupmask);
6168 #else /* !CONFIG_SCHED_DEBUG */
6169 # define sched_domain_debug(sd, cpu) do { } while (0)
6170 #endif /* CONFIG_SCHED_DEBUG */
6172 static int sd_degenerate(struct sched_domain *sd)
6174 if (cpumask_weight(sched_domain_span(sd)) == 1)
6177 /* Following flags need at least 2 groups */
6178 if (sd->flags & (SD_LOAD_BALANCE |
6179 SD_BALANCE_NEWIDLE |
6183 SD_SHARE_PKG_RESOURCES)) {
6184 if (sd->groups != sd->groups->next)
6188 /* Following flags don't use groups */
6189 if (sd->flags & (SD_WAKE_AFFINE))
6196 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6198 unsigned long cflags = sd->flags, pflags = parent->flags;
6200 if (sd_degenerate(parent))
6203 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6206 /* Flags needing groups don't count if only 1 group in parent */
6207 if (parent->groups == parent->groups->next) {
6208 pflags &= ~(SD_LOAD_BALANCE |
6209 SD_BALANCE_NEWIDLE |
6213 SD_SHARE_PKG_RESOURCES);
6214 if (nr_node_ids == 1)
6215 pflags &= ~SD_SERIALIZE;
6217 if (~cflags & pflags)
6223 static void free_rootdomain(struct root_domain *rd)
6225 synchronize_sched();
6227 cpupri_cleanup(&rd->cpupri);
6229 free_cpumask_var(rd->rto_mask);
6230 free_cpumask_var(rd->online);
6231 free_cpumask_var(rd->span);
6235 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6237 struct root_domain *old_rd = NULL;
6238 unsigned long flags;
6240 raw_spin_lock_irqsave(&rq->lock, flags);
6245 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6248 cpumask_clear_cpu(rq->cpu, old_rd->span);
6251 * If we dont want to free the old_rt yet then
6252 * set old_rd to NULL to skip the freeing later
6255 if (!atomic_dec_and_test(&old_rd->refcount))
6259 atomic_inc(&rd->refcount);
6262 cpumask_set_cpu(rq->cpu, rd->span);
6263 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6266 raw_spin_unlock_irqrestore(&rq->lock, flags);
6269 free_rootdomain(old_rd);
6272 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6274 gfp_t gfp = GFP_KERNEL;
6276 memset(rd, 0, sizeof(*rd));
6281 if (!alloc_cpumask_var(&rd->span, gfp))
6283 if (!alloc_cpumask_var(&rd->online, gfp))
6285 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6288 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6293 free_cpumask_var(rd->rto_mask);
6295 free_cpumask_var(rd->online);
6297 free_cpumask_var(rd->span);
6302 static void init_defrootdomain(void)
6304 init_rootdomain(&def_root_domain, true);
6306 atomic_set(&def_root_domain.refcount, 1);
6309 static struct root_domain *alloc_rootdomain(void)
6311 struct root_domain *rd;
6313 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6317 if (init_rootdomain(rd, false) != 0) {
6326 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6327 * hold the hotplug lock.
6330 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6332 struct rq *rq = cpu_rq(cpu);
6333 struct sched_domain *tmp;
6335 /* Remove the sched domains which do not contribute to scheduling. */
6336 for (tmp = sd; tmp; ) {
6337 struct sched_domain *parent = tmp->parent;
6341 if (sd_parent_degenerate(tmp, parent)) {
6342 tmp->parent = parent->parent;
6344 parent->parent->child = tmp;
6349 if (sd && sd_degenerate(sd)) {
6355 sched_domain_debug(sd, cpu);
6357 rq_attach_root(rq, rd);
6358 rcu_assign_pointer(rq->sd, sd);
6361 /* cpus with isolated domains */
6362 static cpumask_var_t cpu_isolated_map;
6364 /* Setup the mask of cpus configured for isolated domains */
6365 static int __init isolated_cpu_setup(char *str)
6367 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6368 cpulist_parse(str, cpu_isolated_map);
6372 __setup("isolcpus=", isolated_cpu_setup);
6375 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6376 * to a function which identifies what group(along with sched group) a CPU
6377 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6378 * (due to the fact that we keep track of groups covered with a struct cpumask).
6380 * init_sched_build_groups will build a circular linked list of the groups
6381 * covered by the given span, and will set each group's ->cpumask correctly,
6382 * and ->cpu_power to 0.
6385 init_sched_build_groups(const struct cpumask *span,
6386 const struct cpumask *cpu_map,
6387 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6388 struct sched_group **sg,
6389 struct cpumask *tmpmask),
6390 struct cpumask *covered, struct cpumask *tmpmask)
6392 struct sched_group *first = NULL, *last = NULL;
6395 cpumask_clear(covered);
6397 for_each_cpu(i, span) {
6398 struct sched_group *sg;
6399 int group = group_fn(i, cpu_map, &sg, tmpmask);
6402 if (cpumask_test_cpu(i, covered))
6405 cpumask_clear(sched_group_cpus(sg));
6408 for_each_cpu(j, span) {
6409 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6412 cpumask_set_cpu(j, covered);
6413 cpumask_set_cpu(j, sched_group_cpus(sg));
6424 #define SD_NODES_PER_DOMAIN 16
6429 * find_next_best_node - find the next node to include in a sched_domain
6430 * @node: node whose sched_domain we're building
6431 * @used_nodes: nodes already in the sched_domain
6433 * Find the next node to include in a given scheduling domain. Simply
6434 * finds the closest node not already in the @used_nodes map.
6436 * Should use nodemask_t.
6438 static int find_next_best_node(int node, nodemask_t *used_nodes)
6440 int i, n, val, min_val, best_node = 0;
6444 for (i = 0; i < nr_node_ids; i++) {
6445 /* Start at @node */
6446 n = (node + i) % nr_node_ids;
6448 if (!nr_cpus_node(n))
6451 /* Skip already used nodes */
6452 if (node_isset(n, *used_nodes))
6455 /* Simple min distance search */
6456 val = node_distance(node, n);
6458 if (val < min_val) {
6464 node_set(best_node, *used_nodes);
6469 * sched_domain_node_span - get a cpumask for a node's sched_domain
6470 * @node: node whose cpumask we're constructing
6471 * @span: resulting cpumask
6473 * Given a node, construct a good cpumask for its sched_domain to span. It
6474 * should be one that prevents unnecessary balancing, but also spreads tasks
6477 static void sched_domain_node_span(int node, struct cpumask *span)
6479 nodemask_t used_nodes;
6482 cpumask_clear(span);
6483 nodes_clear(used_nodes);
6485 cpumask_or(span, span, cpumask_of_node(node));
6486 node_set(node, used_nodes);
6488 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6489 int next_node = find_next_best_node(node, &used_nodes);
6491 cpumask_or(span, span, cpumask_of_node(next_node));
6494 #endif /* CONFIG_NUMA */
6496 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6499 * The cpus mask in sched_group and sched_domain hangs off the end.
6501 * ( See the the comments in include/linux/sched.h:struct sched_group
6502 * and struct sched_domain. )
6504 struct static_sched_group {
6505 struct sched_group sg;
6506 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6509 struct static_sched_domain {
6510 struct sched_domain sd;
6511 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6517 cpumask_var_t domainspan;
6518 cpumask_var_t covered;
6519 cpumask_var_t notcovered;
6521 cpumask_var_t nodemask;
6522 cpumask_var_t this_sibling_map;
6523 cpumask_var_t this_core_map;
6524 cpumask_var_t send_covered;
6525 cpumask_var_t tmpmask;
6526 struct sched_group **sched_group_nodes;
6527 struct root_domain *rd;
6531 sa_sched_groups = 0,
6536 sa_this_sibling_map,
6538 sa_sched_group_nodes,
6548 * SMT sched-domains:
6550 #ifdef CONFIG_SCHED_SMT
6551 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6552 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6555 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6556 struct sched_group **sg, struct cpumask *unused)
6559 *sg = &per_cpu(sched_groups, cpu).sg;
6562 #endif /* CONFIG_SCHED_SMT */
6565 * multi-core sched-domains:
6567 #ifdef CONFIG_SCHED_MC
6568 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6569 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6570 #endif /* CONFIG_SCHED_MC */
6572 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6574 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6575 struct sched_group **sg, struct cpumask *mask)
6579 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6580 group = cpumask_first(mask);
6582 *sg = &per_cpu(sched_group_core, group).sg;
6585 #elif defined(CONFIG_SCHED_MC)
6587 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6588 struct sched_group **sg, struct cpumask *unused)
6591 *sg = &per_cpu(sched_group_core, cpu).sg;
6596 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6597 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6600 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6601 struct sched_group **sg, struct cpumask *mask)
6604 #ifdef CONFIG_SCHED_MC
6605 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6606 group = cpumask_first(mask);
6607 #elif defined(CONFIG_SCHED_SMT)
6608 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6609 group = cpumask_first(mask);
6614 *sg = &per_cpu(sched_group_phys, group).sg;
6620 * The init_sched_build_groups can't handle what we want to do with node
6621 * groups, so roll our own. Now each node has its own list of groups which
6622 * gets dynamically allocated.
6624 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6625 static struct sched_group ***sched_group_nodes_bycpu;
6627 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6628 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6630 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6631 struct sched_group **sg,
6632 struct cpumask *nodemask)
6636 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6637 group = cpumask_first(nodemask);
6640 *sg = &per_cpu(sched_group_allnodes, group).sg;
6644 static void init_numa_sched_groups_power(struct sched_group *group_head)
6646 struct sched_group *sg = group_head;
6652 for_each_cpu(j, sched_group_cpus(sg)) {
6653 struct sched_domain *sd;
6655 sd = &per_cpu(phys_domains, j).sd;
6656 if (j != group_first_cpu(sd->groups)) {
6658 * Only add "power" once for each
6664 sg->cpu_power += sd->groups->cpu_power;
6667 } while (sg != group_head);
6670 static int build_numa_sched_groups(struct s_data *d,
6671 const struct cpumask *cpu_map, int num)
6673 struct sched_domain *sd;
6674 struct sched_group *sg, *prev;
6677 cpumask_clear(d->covered);
6678 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6679 if (cpumask_empty(d->nodemask)) {
6680 d->sched_group_nodes[num] = NULL;
6684 sched_domain_node_span(num, d->domainspan);
6685 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6687 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6690 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6694 d->sched_group_nodes[num] = sg;
6696 for_each_cpu(j, d->nodemask) {
6697 sd = &per_cpu(node_domains, j).sd;
6702 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6704 cpumask_or(d->covered, d->covered, d->nodemask);
6707 for (j = 0; j < nr_node_ids; j++) {
6708 n = (num + j) % nr_node_ids;
6709 cpumask_complement(d->notcovered, d->covered);
6710 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6711 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6712 if (cpumask_empty(d->tmpmask))
6714 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6715 if (cpumask_empty(d->tmpmask))
6717 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6721 "Can not alloc domain group for node %d\n", j);
6725 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6726 sg->next = prev->next;
6727 cpumask_or(d->covered, d->covered, d->tmpmask);
6734 #endif /* CONFIG_NUMA */
6737 /* Free memory allocated for various sched_group structures */
6738 static void free_sched_groups(const struct cpumask *cpu_map,
6739 struct cpumask *nodemask)
6743 for_each_cpu(cpu, cpu_map) {
6744 struct sched_group **sched_group_nodes
6745 = sched_group_nodes_bycpu[cpu];
6747 if (!sched_group_nodes)
6750 for (i = 0; i < nr_node_ids; i++) {
6751 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6753 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6754 if (cpumask_empty(nodemask))
6764 if (oldsg != sched_group_nodes[i])
6767 kfree(sched_group_nodes);
6768 sched_group_nodes_bycpu[cpu] = NULL;
6771 #else /* !CONFIG_NUMA */
6772 static void free_sched_groups(const struct cpumask *cpu_map,
6773 struct cpumask *nodemask)
6776 #endif /* CONFIG_NUMA */
6779 * Initialize sched groups cpu_power.
6781 * cpu_power indicates the capacity of sched group, which is used while
6782 * distributing the load between different sched groups in a sched domain.
6783 * Typically cpu_power for all the groups in a sched domain will be same unless
6784 * there are asymmetries in the topology. If there are asymmetries, group
6785 * having more cpu_power will pickup more load compared to the group having
6788 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6790 struct sched_domain *child;
6791 struct sched_group *group;
6795 WARN_ON(!sd || !sd->groups);
6797 if (cpu != group_first_cpu(sd->groups))
6802 sd->groups->cpu_power = 0;
6805 power = SCHED_LOAD_SCALE;
6806 weight = cpumask_weight(sched_domain_span(sd));
6808 * SMT siblings share the power of a single core.
6809 * Usually multiple threads get a better yield out of
6810 * that one core than a single thread would have,
6811 * reflect that in sd->smt_gain.
6813 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6814 power *= sd->smt_gain;
6816 power >>= SCHED_LOAD_SHIFT;
6818 sd->groups->cpu_power += power;
6823 * Add cpu_power of each child group to this groups cpu_power.
6825 group = child->groups;
6827 sd->groups->cpu_power += group->cpu_power;
6828 group = group->next;
6829 } while (group != child->groups);
6833 * Initializers for schedule domains
6834 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6837 #ifdef CONFIG_SCHED_DEBUG
6838 # define SD_INIT_NAME(sd, type) sd->name = #type
6840 # define SD_INIT_NAME(sd, type) do { } while (0)
6843 #define SD_INIT(sd, type) sd_init_##type(sd)
6845 #define SD_INIT_FUNC(type) \
6846 static noinline void sd_init_##type(struct sched_domain *sd) \
6848 memset(sd, 0, sizeof(*sd)); \
6849 *sd = SD_##type##_INIT; \
6850 sd->level = SD_LV_##type; \
6851 SD_INIT_NAME(sd, type); \
6856 SD_INIT_FUNC(ALLNODES)
6859 #ifdef CONFIG_SCHED_SMT
6860 SD_INIT_FUNC(SIBLING)
6862 #ifdef CONFIG_SCHED_MC
6866 static int default_relax_domain_level = -1;
6868 static int __init setup_relax_domain_level(char *str)
6872 val = simple_strtoul(str, NULL, 0);
6873 if (val < SD_LV_MAX)
6874 default_relax_domain_level = val;
6878 __setup("relax_domain_level=", setup_relax_domain_level);
6880 static void set_domain_attribute(struct sched_domain *sd,
6881 struct sched_domain_attr *attr)
6885 if (!attr || attr->relax_domain_level < 0) {
6886 if (default_relax_domain_level < 0)
6889 request = default_relax_domain_level;
6891 request = attr->relax_domain_level;
6892 if (request < sd->level) {
6893 /* turn off idle balance on this domain */
6894 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6896 /* turn on idle balance on this domain */
6897 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6901 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6902 const struct cpumask *cpu_map)
6905 case sa_sched_groups:
6906 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6907 d->sched_group_nodes = NULL;
6909 free_rootdomain(d->rd); /* fall through */
6911 free_cpumask_var(d->tmpmask); /* fall through */
6912 case sa_send_covered:
6913 free_cpumask_var(d->send_covered); /* fall through */
6914 case sa_this_core_map:
6915 free_cpumask_var(d->this_core_map); /* fall through */
6916 case sa_this_sibling_map:
6917 free_cpumask_var(d->this_sibling_map); /* fall through */
6919 free_cpumask_var(d->nodemask); /* fall through */
6920 case sa_sched_group_nodes:
6922 kfree(d->sched_group_nodes); /* fall through */
6924 free_cpumask_var(d->notcovered); /* fall through */
6926 free_cpumask_var(d->covered); /* fall through */
6928 free_cpumask_var(d->domainspan); /* fall through */
6935 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6936 const struct cpumask *cpu_map)
6939 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6941 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6942 return sa_domainspan;
6943 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6945 /* Allocate the per-node list of sched groups */
6946 d->sched_group_nodes = kcalloc(nr_node_ids,
6947 sizeof(struct sched_group *), GFP_KERNEL);
6948 if (!d->sched_group_nodes) {
6949 printk(KERN_WARNING "Can not alloc sched group node list\n");
6950 return sa_notcovered;
6952 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6954 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6955 return sa_sched_group_nodes;
6956 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6958 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6959 return sa_this_sibling_map;
6960 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6961 return sa_this_core_map;
6962 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6963 return sa_send_covered;
6964 d->rd = alloc_rootdomain();
6966 printk(KERN_WARNING "Cannot alloc root domain\n");
6969 return sa_rootdomain;
6972 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6973 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6975 struct sched_domain *sd = NULL;
6977 struct sched_domain *parent;
6980 if (cpumask_weight(cpu_map) >
6981 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6982 sd = &per_cpu(allnodes_domains, i).sd;
6983 SD_INIT(sd, ALLNODES);
6984 set_domain_attribute(sd, attr);
6985 cpumask_copy(sched_domain_span(sd), cpu_map);
6986 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6991 sd = &per_cpu(node_domains, i).sd;
6993 set_domain_attribute(sd, attr);
6994 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6995 sd->parent = parent;
6998 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7003 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7004 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7005 struct sched_domain *parent, int i)
7007 struct sched_domain *sd;
7008 sd = &per_cpu(phys_domains, i).sd;
7010 set_domain_attribute(sd, attr);
7011 cpumask_copy(sched_domain_span(sd), d->nodemask);
7012 sd->parent = parent;
7015 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7019 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7020 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7021 struct sched_domain *parent, int i)
7023 struct sched_domain *sd = parent;
7024 #ifdef CONFIG_SCHED_MC
7025 sd = &per_cpu(core_domains, i).sd;
7027 set_domain_attribute(sd, attr);
7028 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7029 sd->parent = parent;
7031 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7036 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7037 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7038 struct sched_domain *parent, int i)
7040 struct sched_domain *sd = parent;
7041 #ifdef CONFIG_SCHED_SMT
7042 sd = &per_cpu(cpu_domains, i).sd;
7043 SD_INIT(sd, SIBLING);
7044 set_domain_attribute(sd, attr);
7045 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7046 sd->parent = parent;
7048 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7053 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7054 const struct cpumask *cpu_map, int cpu)
7057 #ifdef CONFIG_SCHED_SMT
7058 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7059 cpumask_and(d->this_sibling_map, cpu_map,
7060 topology_thread_cpumask(cpu));
7061 if (cpu == cpumask_first(d->this_sibling_map))
7062 init_sched_build_groups(d->this_sibling_map, cpu_map,
7064 d->send_covered, d->tmpmask);
7067 #ifdef CONFIG_SCHED_MC
7068 case SD_LV_MC: /* set up multi-core groups */
7069 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7070 if (cpu == cpumask_first(d->this_core_map))
7071 init_sched_build_groups(d->this_core_map, cpu_map,
7073 d->send_covered, d->tmpmask);
7076 case SD_LV_CPU: /* set up physical groups */
7077 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7078 if (!cpumask_empty(d->nodemask))
7079 init_sched_build_groups(d->nodemask, cpu_map,
7081 d->send_covered, d->tmpmask);
7084 case SD_LV_ALLNODES:
7085 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7086 d->send_covered, d->tmpmask);
7095 * Build sched domains for a given set of cpus and attach the sched domains
7096 * to the individual cpus
7098 static int __build_sched_domains(const struct cpumask *cpu_map,
7099 struct sched_domain_attr *attr)
7101 enum s_alloc alloc_state = sa_none;
7103 struct sched_domain *sd;
7109 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7110 if (alloc_state != sa_rootdomain)
7112 alloc_state = sa_sched_groups;
7115 * Set up domains for cpus specified by the cpu_map.
7117 for_each_cpu(i, cpu_map) {
7118 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7121 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7122 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7123 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7124 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7127 for_each_cpu(i, cpu_map) {
7128 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7129 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7132 /* Set up physical groups */
7133 for (i = 0; i < nr_node_ids; i++)
7134 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7137 /* Set up node groups */
7139 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7141 for (i = 0; i < nr_node_ids; i++)
7142 if (build_numa_sched_groups(&d, cpu_map, i))
7146 /* Calculate CPU power for physical packages and nodes */
7147 #ifdef CONFIG_SCHED_SMT
7148 for_each_cpu(i, cpu_map) {
7149 sd = &per_cpu(cpu_domains, i).sd;
7150 init_sched_groups_power(i, sd);
7153 #ifdef CONFIG_SCHED_MC
7154 for_each_cpu(i, cpu_map) {
7155 sd = &per_cpu(core_domains, i).sd;
7156 init_sched_groups_power(i, sd);
7160 for_each_cpu(i, cpu_map) {
7161 sd = &per_cpu(phys_domains, i).sd;
7162 init_sched_groups_power(i, sd);
7166 for (i = 0; i < nr_node_ids; i++)
7167 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7169 if (d.sd_allnodes) {
7170 struct sched_group *sg;
7172 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7174 init_numa_sched_groups_power(sg);
7178 /* Attach the domains */
7179 for_each_cpu(i, cpu_map) {
7180 #ifdef CONFIG_SCHED_SMT
7181 sd = &per_cpu(cpu_domains, i).sd;
7182 #elif defined(CONFIG_SCHED_MC)
7183 sd = &per_cpu(core_domains, i).sd;
7185 sd = &per_cpu(phys_domains, i).sd;
7187 cpu_attach_domain(sd, d.rd, i);
7190 d.sched_group_nodes = NULL; /* don't free this we still need it */
7191 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7195 __free_domain_allocs(&d, alloc_state, cpu_map);
7199 static int build_sched_domains(const struct cpumask *cpu_map)
7201 return __build_sched_domains(cpu_map, NULL);
7204 static cpumask_var_t *doms_cur; /* current sched domains */
7205 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7206 static struct sched_domain_attr *dattr_cur;
7207 /* attribues of custom domains in 'doms_cur' */
7210 * Special case: If a kmalloc of a doms_cur partition (array of
7211 * cpumask) fails, then fallback to a single sched domain,
7212 * as determined by the single cpumask fallback_doms.
7214 static cpumask_var_t fallback_doms;
7217 * arch_update_cpu_topology lets virtualized architectures update the
7218 * cpu core maps. It is supposed to return 1 if the topology changed
7219 * or 0 if it stayed the same.
7221 int __attribute__((weak)) arch_update_cpu_topology(void)
7226 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7229 cpumask_var_t *doms;
7231 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7234 for (i = 0; i < ndoms; i++) {
7235 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7236 free_sched_domains(doms, i);
7243 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7246 for (i = 0; i < ndoms; i++)
7247 free_cpumask_var(doms[i]);
7252 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7253 * For now this just excludes isolated cpus, but could be used to
7254 * exclude other special cases in the future.
7256 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7260 arch_update_cpu_topology();
7262 doms_cur = alloc_sched_domains(ndoms_cur);
7264 doms_cur = &fallback_doms;
7265 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7267 err = build_sched_domains(doms_cur[0]);
7268 register_sched_domain_sysctl();
7273 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7274 struct cpumask *tmpmask)
7276 free_sched_groups(cpu_map, tmpmask);
7280 * Detach sched domains from a group of cpus specified in cpu_map
7281 * These cpus will now be attached to the NULL domain
7283 static void detach_destroy_domains(const struct cpumask *cpu_map)
7285 /* Save because hotplug lock held. */
7286 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7289 for_each_cpu(i, cpu_map)
7290 cpu_attach_domain(NULL, &def_root_domain, i);
7291 synchronize_sched();
7292 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7295 /* handle null as "default" */
7296 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7297 struct sched_domain_attr *new, int idx_new)
7299 struct sched_domain_attr tmp;
7306 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7307 new ? (new + idx_new) : &tmp,
7308 sizeof(struct sched_domain_attr));
7312 * Partition sched domains as specified by the 'ndoms_new'
7313 * cpumasks in the array doms_new[] of cpumasks. This compares
7314 * doms_new[] to the current sched domain partitioning, doms_cur[].
7315 * It destroys each deleted domain and builds each new domain.
7317 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7318 * The masks don't intersect (don't overlap.) We should setup one
7319 * sched domain for each mask. CPUs not in any of the cpumasks will
7320 * not be load balanced. If the same cpumask appears both in the
7321 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7324 * The passed in 'doms_new' should be allocated using
7325 * alloc_sched_domains. This routine takes ownership of it and will
7326 * free_sched_domains it when done with it. If the caller failed the
7327 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7328 * and partition_sched_domains() will fallback to the single partition
7329 * 'fallback_doms', it also forces the domains to be rebuilt.
7331 * If doms_new == NULL it will be replaced with cpu_online_mask.
7332 * ndoms_new == 0 is a special case for destroying existing domains,
7333 * and it will not create the default domain.
7335 * Call with hotplug lock held
7337 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7338 struct sched_domain_attr *dattr_new)
7343 mutex_lock(&sched_domains_mutex);
7345 /* always unregister in case we don't destroy any domains */
7346 unregister_sched_domain_sysctl();
7348 /* Let architecture update cpu core mappings. */
7349 new_topology = arch_update_cpu_topology();
7351 n = doms_new ? ndoms_new : 0;
7353 /* Destroy deleted domains */
7354 for (i = 0; i < ndoms_cur; i++) {
7355 for (j = 0; j < n && !new_topology; j++) {
7356 if (cpumask_equal(doms_cur[i], doms_new[j])
7357 && dattrs_equal(dattr_cur, i, dattr_new, j))
7360 /* no match - a current sched domain not in new doms_new[] */
7361 detach_destroy_domains(doms_cur[i]);
7366 if (doms_new == NULL) {
7368 doms_new = &fallback_doms;
7369 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7370 WARN_ON_ONCE(dattr_new);
7373 /* Build new domains */
7374 for (i = 0; i < ndoms_new; i++) {
7375 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7376 if (cpumask_equal(doms_new[i], doms_cur[j])
7377 && dattrs_equal(dattr_new, i, dattr_cur, j))
7380 /* no match - add a new doms_new */
7381 __build_sched_domains(doms_new[i],
7382 dattr_new ? dattr_new + i : NULL);
7387 /* Remember the new sched domains */
7388 if (doms_cur != &fallback_doms)
7389 free_sched_domains(doms_cur, ndoms_cur);
7390 kfree(dattr_cur); /* kfree(NULL) is safe */
7391 doms_cur = doms_new;
7392 dattr_cur = dattr_new;
7393 ndoms_cur = ndoms_new;
7395 register_sched_domain_sysctl();
7397 mutex_unlock(&sched_domains_mutex);
7400 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7401 static void arch_reinit_sched_domains(void)
7405 /* Destroy domains first to force the rebuild */
7406 partition_sched_domains(0, NULL, NULL);
7408 rebuild_sched_domains();
7412 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7414 unsigned int level = 0;
7416 if (sscanf(buf, "%u", &level) != 1)
7420 * level is always be positive so don't check for
7421 * level < POWERSAVINGS_BALANCE_NONE which is 0
7422 * What happens on 0 or 1 byte write,
7423 * need to check for count as well?
7426 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7430 sched_smt_power_savings = level;
7432 sched_mc_power_savings = level;
7434 arch_reinit_sched_domains();
7439 #ifdef CONFIG_SCHED_MC
7440 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7443 return sprintf(page, "%u\n", sched_mc_power_savings);
7445 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7446 const char *buf, size_t count)
7448 return sched_power_savings_store(buf, count, 0);
7450 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7451 sched_mc_power_savings_show,
7452 sched_mc_power_savings_store);
7455 #ifdef CONFIG_SCHED_SMT
7456 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7459 return sprintf(page, "%u\n", sched_smt_power_savings);
7461 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7462 const char *buf, size_t count)
7464 return sched_power_savings_store(buf, count, 1);
7466 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7467 sched_smt_power_savings_show,
7468 sched_smt_power_savings_store);
7471 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7475 #ifdef CONFIG_SCHED_SMT
7477 err = sysfs_create_file(&cls->kset.kobj,
7478 &attr_sched_smt_power_savings.attr);
7480 #ifdef CONFIG_SCHED_MC
7481 if (!err && mc_capable())
7482 err = sysfs_create_file(&cls->kset.kobj,
7483 &attr_sched_mc_power_savings.attr);
7487 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7489 #ifndef CONFIG_CPUSETS
7491 * Add online and remove offline CPUs from the scheduler domains.
7492 * When cpusets are enabled they take over this function.
7494 static int update_sched_domains(struct notifier_block *nfb,
7495 unsigned long action, void *hcpu)
7499 case CPU_ONLINE_FROZEN:
7500 case CPU_DOWN_PREPARE:
7501 case CPU_DOWN_PREPARE_FROZEN:
7502 case CPU_DOWN_FAILED:
7503 case CPU_DOWN_FAILED_FROZEN:
7504 partition_sched_domains(1, NULL, NULL);
7513 static int update_runtime(struct notifier_block *nfb,
7514 unsigned long action, void *hcpu)
7516 int cpu = (int)(long)hcpu;
7519 case CPU_DOWN_PREPARE:
7520 case CPU_DOWN_PREPARE_FROZEN:
7521 disable_runtime(cpu_rq(cpu));
7524 case CPU_DOWN_FAILED:
7525 case CPU_DOWN_FAILED_FROZEN:
7527 case CPU_ONLINE_FROZEN:
7528 enable_runtime(cpu_rq(cpu));
7536 void __init sched_init_smp(void)
7538 cpumask_var_t non_isolated_cpus;
7540 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7541 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7543 #if defined(CONFIG_NUMA)
7544 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7546 BUG_ON(sched_group_nodes_bycpu == NULL);
7549 mutex_lock(&sched_domains_mutex);
7550 arch_init_sched_domains(cpu_active_mask);
7551 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7552 if (cpumask_empty(non_isolated_cpus))
7553 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7554 mutex_unlock(&sched_domains_mutex);
7557 #ifndef CONFIG_CPUSETS
7558 /* XXX: Theoretical race here - CPU may be hotplugged now */
7559 hotcpu_notifier(update_sched_domains, 0);
7562 /* RT runtime code needs to handle some hotplug events */
7563 hotcpu_notifier(update_runtime, 0);
7567 /* Move init over to a non-isolated CPU */
7568 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7570 sched_init_granularity();
7571 free_cpumask_var(non_isolated_cpus);
7573 init_sched_rt_class();
7576 void __init sched_init_smp(void)
7578 sched_init_granularity();
7580 #endif /* CONFIG_SMP */
7582 const_debug unsigned int sysctl_timer_migration = 1;
7584 int in_sched_functions(unsigned long addr)
7586 return in_lock_functions(addr) ||
7587 (addr >= (unsigned long)__sched_text_start
7588 && addr < (unsigned long)__sched_text_end);
7591 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7593 cfs_rq->tasks_timeline = RB_ROOT;
7594 INIT_LIST_HEAD(&cfs_rq->tasks);
7595 #ifdef CONFIG_FAIR_GROUP_SCHED
7598 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7601 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7603 struct rt_prio_array *array;
7606 array = &rt_rq->active;
7607 for (i = 0; i < MAX_RT_PRIO; i++) {
7608 INIT_LIST_HEAD(array->queue + i);
7609 __clear_bit(i, array->bitmap);
7611 /* delimiter for bitsearch: */
7612 __set_bit(MAX_RT_PRIO, array->bitmap);
7614 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7615 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7617 rt_rq->highest_prio.next = MAX_RT_PRIO;
7621 rt_rq->rt_nr_migratory = 0;
7622 rt_rq->overloaded = 0;
7623 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7627 rt_rq->rt_throttled = 0;
7628 rt_rq->rt_runtime = 0;
7629 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7631 #ifdef CONFIG_RT_GROUP_SCHED
7632 rt_rq->rt_nr_boosted = 0;
7637 #ifdef CONFIG_FAIR_GROUP_SCHED
7638 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7639 struct sched_entity *se, int cpu, int add,
7640 struct sched_entity *parent)
7642 struct rq *rq = cpu_rq(cpu);
7643 tg->cfs_rq[cpu] = cfs_rq;
7644 init_cfs_rq(cfs_rq, rq);
7647 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7650 /* se could be NULL for init_task_group */
7655 se->cfs_rq = &rq->cfs;
7657 se->cfs_rq = parent->my_q;
7660 se->load.weight = tg->shares;
7661 se->load.inv_weight = 0;
7662 se->parent = parent;
7666 #ifdef CONFIG_RT_GROUP_SCHED
7667 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7668 struct sched_rt_entity *rt_se, int cpu, int add,
7669 struct sched_rt_entity *parent)
7671 struct rq *rq = cpu_rq(cpu);
7673 tg->rt_rq[cpu] = rt_rq;
7674 init_rt_rq(rt_rq, rq);
7676 rt_rq->rt_se = rt_se;
7677 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7679 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7681 tg->rt_se[cpu] = rt_se;
7686 rt_se->rt_rq = &rq->rt;
7688 rt_se->rt_rq = parent->my_q;
7690 rt_se->my_q = rt_rq;
7691 rt_se->parent = parent;
7692 INIT_LIST_HEAD(&rt_se->run_list);
7696 void __init sched_init(void)
7699 unsigned long alloc_size = 0, ptr;
7701 #ifdef CONFIG_FAIR_GROUP_SCHED
7702 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7704 #ifdef CONFIG_RT_GROUP_SCHED
7705 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7707 #ifdef CONFIG_USER_SCHED
7710 #ifdef CONFIG_CPUMASK_OFFSTACK
7711 alloc_size += num_possible_cpus() * cpumask_size();
7714 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7716 #ifdef CONFIG_FAIR_GROUP_SCHED
7717 init_task_group.se = (struct sched_entity **)ptr;
7718 ptr += nr_cpu_ids * sizeof(void **);
7720 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7721 ptr += nr_cpu_ids * sizeof(void **);
7723 #ifdef CONFIG_USER_SCHED
7724 root_task_group.se = (struct sched_entity **)ptr;
7725 ptr += nr_cpu_ids * sizeof(void **);
7727 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7728 ptr += nr_cpu_ids * sizeof(void **);
7729 #endif /* CONFIG_USER_SCHED */
7730 #endif /* CONFIG_FAIR_GROUP_SCHED */
7731 #ifdef CONFIG_RT_GROUP_SCHED
7732 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7733 ptr += nr_cpu_ids * sizeof(void **);
7735 init_task_group.rt_rq = (struct rt_rq **)ptr;
7736 ptr += nr_cpu_ids * sizeof(void **);
7738 #ifdef CONFIG_USER_SCHED
7739 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7740 ptr += nr_cpu_ids * sizeof(void **);
7742 root_task_group.rt_rq = (struct rt_rq **)ptr;
7743 ptr += nr_cpu_ids * sizeof(void **);
7744 #endif /* CONFIG_USER_SCHED */
7745 #endif /* CONFIG_RT_GROUP_SCHED */
7746 #ifdef CONFIG_CPUMASK_OFFSTACK
7747 for_each_possible_cpu(i) {
7748 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7749 ptr += cpumask_size();
7751 #endif /* CONFIG_CPUMASK_OFFSTACK */
7755 init_defrootdomain();
7758 init_rt_bandwidth(&def_rt_bandwidth,
7759 global_rt_period(), global_rt_runtime());
7761 #ifdef CONFIG_RT_GROUP_SCHED
7762 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7763 global_rt_period(), global_rt_runtime());
7764 #ifdef CONFIG_USER_SCHED
7765 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7766 global_rt_period(), RUNTIME_INF);
7767 #endif /* CONFIG_USER_SCHED */
7768 #endif /* CONFIG_RT_GROUP_SCHED */
7770 #ifdef CONFIG_GROUP_SCHED
7771 list_add(&init_task_group.list, &task_groups);
7772 INIT_LIST_HEAD(&init_task_group.children);
7774 #ifdef CONFIG_USER_SCHED
7775 INIT_LIST_HEAD(&root_task_group.children);
7776 init_task_group.parent = &root_task_group;
7777 list_add(&init_task_group.siblings, &root_task_group.children);
7778 #endif /* CONFIG_USER_SCHED */
7779 #endif /* CONFIG_GROUP_SCHED */
7781 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7782 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7783 __alignof__(unsigned long));
7785 for_each_possible_cpu(i) {
7789 raw_spin_lock_init(&rq->lock);
7791 rq->calc_load_active = 0;
7792 rq->calc_load_update = jiffies + LOAD_FREQ;
7793 init_cfs_rq(&rq->cfs, rq);
7794 init_rt_rq(&rq->rt, rq);
7795 #ifdef CONFIG_FAIR_GROUP_SCHED
7796 init_task_group.shares = init_task_group_load;
7797 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7798 #ifdef CONFIG_CGROUP_SCHED
7800 * How much cpu bandwidth does init_task_group get?
7802 * In case of task-groups formed thr' the cgroup filesystem, it
7803 * gets 100% of the cpu resources in the system. This overall
7804 * system cpu resource is divided among the tasks of
7805 * init_task_group and its child task-groups in a fair manner,
7806 * based on each entity's (task or task-group's) weight
7807 * (se->load.weight).
7809 * In other words, if init_task_group has 10 tasks of weight
7810 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7811 * then A0's share of the cpu resource is:
7813 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7815 * We achieve this by letting init_task_group's tasks sit
7816 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7818 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7819 #elif defined CONFIG_USER_SCHED
7820 root_task_group.shares = NICE_0_LOAD;
7821 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
7823 * In case of task-groups formed thr' the user id of tasks,
7824 * init_task_group represents tasks belonging to root user.
7825 * Hence it forms a sibling of all subsequent groups formed.
7826 * In this case, init_task_group gets only a fraction of overall
7827 * system cpu resource, based on the weight assigned to root
7828 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7829 * by letting tasks of init_task_group sit in a separate cfs_rq
7830 * (init_tg_cfs_rq) and having one entity represent this group of
7831 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7833 init_tg_cfs_entry(&init_task_group,
7834 &per_cpu(init_tg_cfs_rq, i),
7835 &per_cpu(init_sched_entity, i), i, 1,
7836 root_task_group.se[i]);
7839 #endif /* CONFIG_FAIR_GROUP_SCHED */
7841 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7842 #ifdef CONFIG_RT_GROUP_SCHED
7843 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7844 #ifdef CONFIG_CGROUP_SCHED
7845 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7846 #elif defined CONFIG_USER_SCHED
7847 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
7848 init_tg_rt_entry(&init_task_group,
7849 &per_cpu(init_rt_rq_var, i),
7850 &per_cpu(init_sched_rt_entity, i), i, 1,
7851 root_task_group.rt_se[i]);
7855 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7856 rq->cpu_load[j] = 0;
7860 rq->post_schedule = 0;
7861 rq->active_balance = 0;
7862 rq->next_balance = jiffies;
7866 rq->migration_thread = NULL;
7868 rq->avg_idle = 2*sysctl_sched_migration_cost;
7869 INIT_LIST_HEAD(&rq->migration_queue);
7870 rq_attach_root(rq, &def_root_domain);
7873 atomic_set(&rq->nr_iowait, 0);
7876 set_load_weight(&init_task);
7878 #ifdef CONFIG_PREEMPT_NOTIFIERS
7879 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7883 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7886 #ifdef CONFIG_RT_MUTEXES
7887 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7891 * The boot idle thread does lazy MMU switching as well:
7893 atomic_inc(&init_mm.mm_count);
7894 enter_lazy_tlb(&init_mm, current);
7897 * Make us the idle thread. Technically, schedule() should not be
7898 * called from this thread, however somewhere below it might be,
7899 * but because we are the idle thread, we just pick up running again
7900 * when this runqueue becomes "idle".
7902 init_idle(current, smp_processor_id());
7904 calc_load_update = jiffies + LOAD_FREQ;
7907 * During early bootup we pretend to be a normal task:
7909 current->sched_class = &fair_sched_class;
7911 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7912 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7915 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7916 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7918 /* May be allocated at isolcpus cmdline parse time */
7919 if (cpu_isolated_map == NULL)
7920 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7925 scheduler_running = 1;
7928 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7929 static inline int preempt_count_equals(int preempt_offset)
7931 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7933 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7936 void __might_sleep(const char *file, int line, int preempt_offset)
7939 static unsigned long prev_jiffy; /* ratelimiting */
7941 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7942 system_state != SYSTEM_RUNNING || oops_in_progress)
7944 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7946 prev_jiffy = jiffies;
7949 "BUG: sleeping function called from invalid context at %s:%d\n",
7952 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7953 in_atomic(), irqs_disabled(),
7954 current->pid, current->comm);
7956 debug_show_held_locks(current);
7957 if (irqs_disabled())
7958 print_irqtrace_events(current);
7962 EXPORT_SYMBOL(__might_sleep);
7965 #ifdef CONFIG_MAGIC_SYSRQ
7966 static void normalize_task(struct rq *rq, struct task_struct *p)
7970 update_rq_clock(rq);
7971 on_rq = p->se.on_rq;
7973 deactivate_task(rq, p, 0);
7974 __setscheduler(rq, p, SCHED_NORMAL, 0);
7976 activate_task(rq, p, 0);
7977 resched_task(rq->curr);
7981 void normalize_rt_tasks(void)
7983 struct task_struct *g, *p;
7984 unsigned long flags;
7987 read_lock_irqsave(&tasklist_lock, flags);
7988 do_each_thread(g, p) {
7990 * Only normalize user tasks:
7995 p->se.exec_start = 0;
7996 #ifdef CONFIG_SCHEDSTATS
7997 p->se.wait_start = 0;
7998 p->se.sleep_start = 0;
7999 p->se.block_start = 0;
8004 * Renice negative nice level userspace
8007 if (TASK_NICE(p) < 0 && p->mm)
8008 set_user_nice(p, 0);
8012 raw_spin_lock(&p->pi_lock);
8013 rq = __task_rq_lock(p);
8015 normalize_task(rq, p);
8017 __task_rq_unlock(rq);
8018 raw_spin_unlock(&p->pi_lock);
8019 } while_each_thread(g, p);
8021 read_unlock_irqrestore(&tasklist_lock, flags);
8024 #endif /* CONFIG_MAGIC_SYSRQ */
8028 * These functions are only useful for the IA64 MCA handling.
8030 * They can only be called when the whole system has been
8031 * stopped - every CPU needs to be quiescent, and no scheduling
8032 * activity can take place. Using them for anything else would
8033 * be a serious bug, and as a result, they aren't even visible
8034 * under any other configuration.
8038 * curr_task - return the current task for a given cpu.
8039 * @cpu: the processor in question.
8041 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8043 struct task_struct *curr_task(int cpu)
8045 return cpu_curr(cpu);
8049 * set_curr_task - set the current task for a given cpu.
8050 * @cpu: the processor in question.
8051 * @p: the task pointer to set.
8053 * Description: This function must only be used when non-maskable interrupts
8054 * are serviced on a separate stack. It allows the architecture to switch the
8055 * notion of the current task on a cpu in a non-blocking manner. This function
8056 * must be called with all CPU's synchronized, and interrupts disabled, the
8057 * and caller must save the original value of the current task (see
8058 * curr_task() above) and restore that value before reenabling interrupts and
8059 * re-starting the system.
8061 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8063 void set_curr_task(int cpu, struct task_struct *p)
8070 #ifdef CONFIG_FAIR_GROUP_SCHED
8071 static void free_fair_sched_group(struct task_group *tg)
8075 for_each_possible_cpu(i) {
8077 kfree(tg->cfs_rq[i]);
8087 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8089 struct cfs_rq *cfs_rq;
8090 struct sched_entity *se;
8094 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8097 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8101 tg->shares = NICE_0_LOAD;
8103 for_each_possible_cpu(i) {
8106 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8107 GFP_KERNEL, cpu_to_node(i));
8111 se = kzalloc_node(sizeof(struct sched_entity),
8112 GFP_KERNEL, cpu_to_node(i));
8116 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8127 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8129 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8130 &cpu_rq(cpu)->leaf_cfs_rq_list);
8133 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8135 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8137 #else /* !CONFG_FAIR_GROUP_SCHED */
8138 static inline void free_fair_sched_group(struct task_group *tg)
8143 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8148 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8152 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8155 #endif /* CONFIG_FAIR_GROUP_SCHED */
8157 #ifdef CONFIG_RT_GROUP_SCHED
8158 static void free_rt_sched_group(struct task_group *tg)
8162 destroy_rt_bandwidth(&tg->rt_bandwidth);
8164 for_each_possible_cpu(i) {
8166 kfree(tg->rt_rq[i]);
8168 kfree(tg->rt_se[i]);
8176 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8178 struct rt_rq *rt_rq;
8179 struct sched_rt_entity *rt_se;
8183 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8186 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8190 init_rt_bandwidth(&tg->rt_bandwidth,
8191 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8193 for_each_possible_cpu(i) {
8196 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8197 GFP_KERNEL, cpu_to_node(i));
8201 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8202 GFP_KERNEL, cpu_to_node(i));
8206 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8217 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8219 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8220 &cpu_rq(cpu)->leaf_rt_rq_list);
8223 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8225 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8227 #else /* !CONFIG_RT_GROUP_SCHED */
8228 static inline void free_rt_sched_group(struct task_group *tg)
8233 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8238 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8242 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8245 #endif /* CONFIG_RT_GROUP_SCHED */
8247 #ifdef CONFIG_GROUP_SCHED
8248 static void free_sched_group(struct task_group *tg)
8250 free_fair_sched_group(tg);
8251 free_rt_sched_group(tg);
8255 /* allocate runqueue etc for a new task group */
8256 struct task_group *sched_create_group(struct task_group *parent)
8258 struct task_group *tg;
8259 unsigned long flags;
8262 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8264 return ERR_PTR(-ENOMEM);
8266 if (!alloc_fair_sched_group(tg, parent))
8269 if (!alloc_rt_sched_group(tg, parent))
8272 spin_lock_irqsave(&task_group_lock, flags);
8273 for_each_possible_cpu(i) {
8274 register_fair_sched_group(tg, i);
8275 register_rt_sched_group(tg, i);
8277 list_add_rcu(&tg->list, &task_groups);
8279 WARN_ON(!parent); /* root should already exist */
8281 tg->parent = parent;
8282 INIT_LIST_HEAD(&tg->children);
8283 list_add_rcu(&tg->siblings, &parent->children);
8284 spin_unlock_irqrestore(&task_group_lock, flags);
8289 free_sched_group(tg);
8290 return ERR_PTR(-ENOMEM);
8293 /* rcu callback to free various structures associated with a task group */
8294 static void free_sched_group_rcu(struct rcu_head *rhp)
8296 /* now it should be safe to free those cfs_rqs */
8297 free_sched_group(container_of(rhp, struct task_group, rcu));
8300 /* Destroy runqueue etc associated with a task group */
8301 void sched_destroy_group(struct task_group *tg)
8303 unsigned long flags;
8306 spin_lock_irqsave(&task_group_lock, flags);
8307 for_each_possible_cpu(i) {
8308 unregister_fair_sched_group(tg, i);
8309 unregister_rt_sched_group(tg, i);
8311 list_del_rcu(&tg->list);
8312 list_del_rcu(&tg->siblings);
8313 spin_unlock_irqrestore(&task_group_lock, flags);
8315 /* wait for possible concurrent references to cfs_rqs complete */
8316 call_rcu(&tg->rcu, free_sched_group_rcu);
8319 /* change task's runqueue when it moves between groups.
8320 * The caller of this function should have put the task in its new group
8321 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8322 * reflect its new group.
8324 void sched_move_task(struct task_struct *tsk)
8327 unsigned long flags;
8330 rq = task_rq_lock(tsk, &flags);
8332 update_rq_clock(rq);
8334 running = task_current(rq, tsk);
8335 on_rq = tsk->se.on_rq;
8338 dequeue_task(rq, tsk, 0);
8339 if (unlikely(running))
8340 tsk->sched_class->put_prev_task(rq, tsk);
8342 set_task_rq(tsk, task_cpu(tsk));
8344 #ifdef CONFIG_FAIR_GROUP_SCHED
8345 if (tsk->sched_class->moved_group)
8346 tsk->sched_class->moved_group(tsk, on_rq);
8349 if (unlikely(running))
8350 tsk->sched_class->set_curr_task(rq);
8352 enqueue_task(rq, tsk, 0);
8354 task_rq_unlock(rq, &flags);
8356 #endif /* CONFIG_GROUP_SCHED */
8358 #ifdef CONFIG_FAIR_GROUP_SCHED
8359 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8361 struct cfs_rq *cfs_rq = se->cfs_rq;
8366 dequeue_entity(cfs_rq, se, 0);
8368 se->load.weight = shares;
8369 se->load.inv_weight = 0;
8372 enqueue_entity(cfs_rq, se, 0);
8375 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8377 struct cfs_rq *cfs_rq = se->cfs_rq;
8378 struct rq *rq = cfs_rq->rq;
8379 unsigned long flags;
8381 raw_spin_lock_irqsave(&rq->lock, flags);
8382 __set_se_shares(se, shares);
8383 raw_spin_unlock_irqrestore(&rq->lock, flags);
8386 static DEFINE_MUTEX(shares_mutex);
8388 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8391 unsigned long flags;
8394 * We can't change the weight of the root cgroup.
8399 if (shares < MIN_SHARES)
8400 shares = MIN_SHARES;
8401 else if (shares > MAX_SHARES)
8402 shares = MAX_SHARES;
8404 mutex_lock(&shares_mutex);
8405 if (tg->shares == shares)
8408 spin_lock_irqsave(&task_group_lock, flags);
8409 for_each_possible_cpu(i)
8410 unregister_fair_sched_group(tg, i);
8411 list_del_rcu(&tg->siblings);
8412 spin_unlock_irqrestore(&task_group_lock, flags);
8414 /* wait for any ongoing reference to this group to finish */
8415 synchronize_sched();
8418 * Now we are free to modify the group's share on each cpu
8419 * w/o tripping rebalance_share or load_balance_fair.
8421 tg->shares = shares;
8422 for_each_possible_cpu(i) {
8426 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8427 set_se_shares(tg->se[i], shares);
8431 * Enable load balance activity on this group, by inserting it back on
8432 * each cpu's rq->leaf_cfs_rq_list.
8434 spin_lock_irqsave(&task_group_lock, flags);
8435 for_each_possible_cpu(i)
8436 register_fair_sched_group(tg, i);
8437 list_add_rcu(&tg->siblings, &tg->parent->children);
8438 spin_unlock_irqrestore(&task_group_lock, flags);
8440 mutex_unlock(&shares_mutex);
8444 unsigned long sched_group_shares(struct task_group *tg)
8450 #ifdef CONFIG_RT_GROUP_SCHED
8452 * Ensure that the real time constraints are schedulable.
8454 static DEFINE_MUTEX(rt_constraints_mutex);
8456 static unsigned long to_ratio(u64 period, u64 runtime)
8458 if (runtime == RUNTIME_INF)
8461 return div64_u64(runtime << 20, period);
8464 /* Must be called with tasklist_lock held */
8465 static inline int tg_has_rt_tasks(struct task_group *tg)
8467 struct task_struct *g, *p;
8469 do_each_thread(g, p) {
8470 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8472 } while_each_thread(g, p);
8477 struct rt_schedulable_data {
8478 struct task_group *tg;
8483 static int tg_schedulable(struct task_group *tg, void *data)
8485 struct rt_schedulable_data *d = data;
8486 struct task_group *child;
8487 unsigned long total, sum = 0;
8488 u64 period, runtime;
8490 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8491 runtime = tg->rt_bandwidth.rt_runtime;
8494 period = d->rt_period;
8495 runtime = d->rt_runtime;
8498 #ifdef CONFIG_USER_SCHED
8499 if (tg == &root_task_group) {
8500 period = global_rt_period();
8501 runtime = global_rt_runtime();
8506 * Cannot have more runtime than the period.
8508 if (runtime > period && runtime != RUNTIME_INF)
8512 * Ensure we don't starve existing RT tasks.
8514 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8517 total = to_ratio(period, runtime);
8520 * Nobody can have more than the global setting allows.
8522 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8526 * The sum of our children's runtime should not exceed our own.
8528 list_for_each_entry_rcu(child, &tg->children, siblings) {
8529 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8530 runtime = child->rt_bandwidth.rt_runtime;
8532 if (child == d->tg) {
8533 period = d->rt_period;
8534 runtime = d->rt_runtime;
8537 sum += to_ratio(period, runtime);
8546 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8548 struct rt_schedulable_data data = {
8550 .rt_period = period,
8551 .rt_runtime = runtime,
8554 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8557 static int tg_set_bandwidth(struct task_group *tg,
8558 u64 rt_period, u64 rt_runtime)
8562 mutex_lock(&rt_constraints_mutex);
8563 read_lock(&tasklist_lock);
8564 err = __rt_schedulable(tg, rt_period, rt_runtime);
8568 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8569 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8570 tg->rt_bandwidth.rt_runtime = rt_runtime;
8572 for_each_possible_cpu(i) {
8573 struct rt_rq *rt_rq = tg->rt_rq[i];
8575 raw_spin_lock(&rt_rq->rt_runtime_lock);
8576 rt_rq->rt_runtime = rt_runtime;
8577 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8579 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8581 read_unlock(&tasklist_lock);
8582 mutex_unlock(&rt_constraints_mutex);
8587 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8589 u64 rt_runtime, rt_period;
8591 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8592 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8593 if (rt_runtime_us < 0)
8594 rt_runtime = RUNTIME_INF;
8596 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8599 long sched_group_rt_runtime(struct task_group *tg)
8603 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8606 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8607 do_div(rt_runtime_us, NSEC_PER_USEC);
8608 return rt_runtime_us;
8611 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8613 u64 rt_runtime, rt_period;
8615 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8616 rt_runtime = tg->rt_bandwidth.rt_runtime;
8621 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8624 long sched_group_rt_period(struct task_group *tg)
8628 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8629 do_div(rt_period_us, NSEC_PER_USEC);
8630 return rt_period_us;
8633 static int sched_rt_global_constraints(void)
8635 u64 runtime, period;
8638 if (sysctl_sched_rt_period <= 0)
8641 runtime = global_rt_runtime();
8642 period = global_rt_period();
8645 * Sanity check on the sysctl variables.
8647 if (runtime > period && runtime != RUNTIME_INF)
8650 mutex_lock(&rt_constraints_mutex);
8651 read_lock(&tasklist_lock);
8652 ret = __rt_schedulable(NULL, 0, 0);
8653 read_unlock(&tasklist_lock);
8654 mutex_unlock(&rt_constraints_mutex);
8659 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8661 /* Don't accept realtime tasks when there is no way for them to run */
8662 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8668 #else /* !CONFIG_RT_GROUP_SCHED */
8669 static int sched_rt_global_constraints(void)
8671 unsigned long flags;
8674 if (sysctl_sched_rt_period <= 0)
8678 * There's always some RT tasks in the root group
8679 * -- migration, kstopmachine etc..
8681 if (sysctl_sched_rt_runtime == 0)
8684 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8685 for_each_possible_cpu(i) {
8686 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8688 raw_spin_lock(&rt_rq->rt_runtime_lock);
8689 rt_rq->rt_runtime = global_rt_runtime();
8690 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8692 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8696 #endif /* CONFIG_RT_GROUP_SCHED */
8698 int sched_rt_handler(struct ctl_table *table, int write,
8699 void __user *buffer, size_t *lenp,
8703 int old_period, old_runtime;
8704 static DEFINE_MUTEX(mutex);
8707 old_period = sysctl_sched_rt_period;
8708 old_runtime = sysctl_sched_rt_runtime;
8710 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8712 if (!ret && write) {
8713 ret = sched_rt_global_constraints();
8715 sysctl_sched_rt_period = old_period;
8716 sysctl_sched_rt_runtime = old_runtime;
8718 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8719 def_rt_bandwidth.rt_period =
8720 ns_to_ktime(global_rt_period());
8723 mutex_unlock(&mutex);
8728 #ifdef CONFIG_CGROUP_SCHED
8730 /* return corresponding task_group object of a cgroup */
8731 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8733 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8734 struct task_group, css);
8737 static struct cgroup_subsys_state *
8738 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8740 struct task_group *tg, *parent;
8742 if (!cgrp->parent) {
8743 /* This is early initialization for the top cgroup */
8744 return &init_task_group.css;
8747 parent = cgroup_tg(cgrp->parent);
8748 tg = sched_create_group(parent);
8750 return ERR_PTR(-ENOMEM);
8756 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8758 struct task_group *tg = cgroup_tg(cgrp);
8760 sched_destroy_group(tg);
8764 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8766 #ifdef CONFIG_RT_GROUP_SCHED
8767 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8770 /* We don't support RT-tasks being in separate groups */
8771 if (tsk->sched_class != &fair_sched_class)
8778 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8779 struct task_struct *tsk, bool threadgroup)
8781 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8785 struct task_struct *c;
8787 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8788 retval = cpu_cgroup_can_attach_task(cgrp, c);
8800 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8801 struct cgroup *old_cont, struct task_struct *tsk,
8804 sched_move_task(tsk);
8806 struct task_struct *c;
8808 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8815 #ifdef CONFIG_FAIR_GROUP_SCHED
8816 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8819 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8822 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8824 struct task_group *tg = cgroup_tg(cgrp);
8826 return (u64) tg->shares;
8828 #endif /* CONFIG_FAIR_GROUP_SCHED */
8830 #ifdef CONFIG_RT_GROUP_SCHED
8831 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8834 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8837 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8839 return sched_group_rt_runtime(cgroup_tg(cgrp));
8842 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8845 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8848 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8850 return sched_group_rt_period(cgroup_tg(cgrp));
8852 #endif /* CONFIG_RT_GROUP_SCHED */
8854 static struct cftype cpu_files[] = {
8855 #ifdef CONFIG_FAIR_GROUP_SCHED
8858 .read_u64 = cpu_shares_read_u64,
8859 .write_u64 = cpu_shares_write_u64,
8862 #ifdef CONFIG_RT_GROUP_SCHED
8864 .name = "rt_runtime_us",
8865 .read_s64 = cpu_rt_runtime_read,
8866 .write_s64 = cpu_rt_runtime_write,
8869 .name = "rt_period_us",
8870 .read_u64 = cpu_rt_period_read_uint,
8871 .write_u64 = cpu_rt_period_write_uint,
8876 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8878 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8881 struct cgroup_subsys cpu_cgroup_subsys = {
8883 .create = cpu_cgroup_create,
8884 .destroy = cpu_cgroup_destroy,
8885 .can_attach = cpu_cgroup_can_attach,
8886 .attach = cpu_cgroup_attach,
8887 .populate = cpu_cgroup_populate,
8888 .subsys_id = cpu_cgroup_subsys_id,
8892 #endif /* CONFIG_CGROUP_SCHED */
8894 #ifdef CONFIG_CGROUP_CPUACCT
8897 * CPU accounting code for task groups.
8899 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8900 * (balbir@in.ibm.com).
8903 /* track cpu usage of a group of tasks and its child groups */
8905 struct cgroup_subsys_state css;
8906 /* cpuusage holds pointer to a u64-type object on every cpu */
8908 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8909 struct cpuacct *parent;
8912 struct cgroup_subsys cpuacct_subsys;
8914 /* return cpu accounting group corresponding to this container */
8915 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8917 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8918 struct cpuacct, css);
8921 /* return cpu accounting group to which this task belongs */
8922 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8924 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8925 struct cpuacct, css);
8928 /* create a new cpu accounting group */
8929 static struct cgroup_subsys_state *cpuacct_create(
8930 struct cgroup_subsys *ss, struct cgroup *cgrp)
8932 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8938 ca->cpuusage = alloc_percpu(u64);
8942 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8943 if (percpu_counter_init(&ca->cpustat[i], 0))
8944 goto out_free_counters;
8947 ca->parent = cgroup_ca(cgrp->parent);
8953 percpu_counter_destroy(&ca->cpustat[i]);
8954 free_percpu(ca->cpuusage);
8958 return ERR_PTR(-ENOMEM);
8961 /* destroy an existing cpu accounting group */
8963 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8965 struct cpuacct *ca = cgroup_ca(cgrp);
8968 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8969 percpu_counter_destroy(&ca->cpustat[i]);
8970 free_percpu(ca->cpuusage);
8974 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8976 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8979 #ifndef CONFIG_64BIT
8981 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8983 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8985 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8993 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8995 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8997 #ifndef CONFIG_64BIT
8999 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9001 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9003 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9009 /* return total cpu usage (in nanoseconds) of a group */
9010 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9012 struct cpuacct *ca = cgroup_ca(cgrp);
9013 u64 totalcpuusage = 0;
9016 for_each_present_cpu(i)
9017 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9019 return totalcpuusage;
9022 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9025 struct cpuacct *ca = cgroup_ca(cgrp);
9034 for_each_present_cpu(i)
9035 cpuacct_cpuusage_write(ca, i, 0);
9041 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9044 struct cpuacct *ca = cgroup_ca(cgroup);
9048 for_each_present_cpu(i) {
9049 percpu = cpuacct_cpuusage_read(ca, i);
9050 seq_printf(m, "%llu ", (unsigned long long) percpu);
9052 seq_printf(m, "\n");
9056 static const char *cpuacct_stat_desc[] = {
9057 [CPUACCT_STAT_USER] = "user",
9058 [CPUACCT_STAT_SYSTEM] = "system",
9061 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9062 struct cgroup_map_cb *cb)
9064 struct cpuacct *ca = cgroup_ca(cgrp);
9067 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9068 s64 val = percpu_counter_read(&ca->cpustat[i]);
9069 val = cputime64_to_clock_t(val);
9070 cb->fill(cb, cpuacct_stat_desc[i], val);
9075 static struct cftype files[] = {
9078 .read_u64 = cpuusage_read,
9079 .write_u64 = cpuusage_write,
9082 .name = "usage_percpu",
9083 .read_seq_string = cpuacct_percpu_seq_read,
9087 .read_map = cpuacct_stats_show,
9091 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9093 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9097 * charge this task's execution time to its accounting group.
9099 * called with rq->lock held.
9101 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9106 if (unlikely(!cpuacct_subsys.active))
9109 cpu = task_cpu(tsk);
9115 for (; ca; ca = ca->parent) {
9116 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9117 *cpuusage += cputime;
9124 * Charge the system/user time to the task's accounting group.
9126 static void cpuacct_update_stats(struct task_struct *tsk,
9127 enum cpuacct_stat_index idx, cputime_t val)
9131 if (unlikely(!cpuacct_subsys.active))
9138 percpu_counter_add(&ca->cpustat[idx], val);
9144 struct cgroup_subsys cpuacct_subsys = {
9146 .create = cpuacct_create,
9147 .destroy = cpuacct_destroy,
9148 .populate = cpuacct_populate,
9149 .subsys_id = cpuacct_subsys_id,
9151 #endif /* CONFIG_CGROUP_CPUACCT */
9155 int rcu_expedited_torture_stats(char *page)
9159 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9161 void synchronize_sched_expedited(void)
9164 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9166 #else /* #ifndef CONFIG_SMP */
9168 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
9169 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
9171 #define RCU_EXPEDITED_STATE_POST -2
9172 #define RCU_EXPEDITED_STATE_IDLE -1
9174 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9176 int rcu_expedited_torture_stats(char *page)
9181 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
9182 for_each_online_cpu(cpu) {
9183 cnt += sprintf(&page[cnt], " %d:%d",
9184 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
9186 cnt += sprintf(&page[cnt], "\n");
9189 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9191 static long synchronize_sched_expedited_count;
9194 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9195 * approach to force grace period to end quickly. This consumes
9196 * significant time on all CPUs, and is thus not recommended for
9197 * any sort of common-case code.
9199 * Note that it is illegal to call this function while holding any
9200 * lock that is acquired by a CPU-hotplug notifier. Failing to
9201 * observe this restriction will result in deadlock.
9203 void synchronize_sched_expedited(void)
9206 unsigned long flags;
9207 bool need_full_sync = 0;
9209 struct migration_req *req;
9213 smp_mb(); /* ensure prior mod happens before capturing snap. */
9214 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
9216 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
9218 if (trycount++ < 10)
9219 udelay(trycount * num_online_cpus());
9221 synchronize_sched();
9224 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
9225 smp_mb(); /* ensure test happens before caller kfree */
9230 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
9231 for_each_online_cpu(cpu) {
9233 req = &per_cpu(rcu_migration_req, cpu);
9234 init_completion(&req->done);
9236 req->dest_cpu = RCU_MIGRATION_NEED_QS;
9237 raw_spin_lock_irqsave(&rq->lock, flags);
9238 list_add(&req->list, &rq->migration_queue);
9239 raw_spin_unlock_irqrestore(&rq->lock, flags);
9240 wake_up_process(rq->migration_thread);
9242 for_each_online_cpu(cpu) {
9243 rcu_expedited_state = cpu;
9244 req = &per_cpu(rcu_migration_req, cpu);
9246 wait_for_completion(&req->done);
9247 raw_spin_lock_irqsave(&rq->lock, flags);
9248 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
9250 req->dest_cpu = RCU_MIGRATION_IDLE;
9251 raw_spin_unlock_irqrestore(&rq->lock, flags);
9253 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9254 synchronize_sched_expedited_count++;
9255 mutex_unlock(&rcu_sched_expedited_mutex);
9258 synchronize_sched();
9260 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9262 #endif /* #else #ifndef CONFIG_SMP */