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_counter.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy)
124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
129 static inline int task_has_rt_policy(struct task_struct *p)
131 return rt_policy(p->policy);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
142 struct rt_bandwidth {
143 /* nests inside the rq lock: */
144 spinlock_t rt_runtime_lock;
147 struct hrtimer rt_period_timer;
150 static struct rt_bandwidth def_rt_bandwidth;
152 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
154 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169 idle = do_sched_rt_period_timer(rt_b, overrun);
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
176 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
181 spin_lock_init(&rt_b->rt_runtime_lock);
183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime >= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
200 if (hrtimer_active(&rt_b->rt_period_timer))
203 spin_lock(&rt_b->rt_runtime_lock);
208 if (hrtimer_active(&rt_b->rt_period_timer))
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
218 HRTIMER_MODE_ABS_PINNED, 0);
220 spin_unlock(&rt_b->rt_runtime_lock);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
226 hrtimer_cancel(&rt_b->rt_period_timer);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex);
236 #ifdef CONFIG_GROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups);
244 /* task group related information */
246 #ifdef CONFIG_CGROUP_SCHED
247 struct cgroup_subsys_state css;
250 #ifdef CONFIG_USER_SCHED
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
266 struct rt_bandwidth rt_bandwidth;
270 struct list_head list;
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
277 #ifdef CONFIG_USER_SCHED
279 /* Helper function to pass uid information to create_sched_user() */
280 void set_tg_uid(struct user_struct *user)
282 user->tg->uid = user->uid;
287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
290 struct task_group root_task_group;
292 #ifdef CONFIG_FAIR_GROUP_SCHED
293 /* Default task group's sched entity on each cpu */
294 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
295 /* Default task group's cfs_rq on each cpu */
296 static DEFINE_PER_CPU(struct cfs_rq, init_tg_cfs_rq) ____cacheline_aligned_in_smp;
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(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
302 #endif /* CONFIG_RT_GROUP_SCHED */
303 #else /* !CONFIG_USER_SCHED */
304 #define root_task_group init_task_group
305 #endif /* CONFIG_USER_SCHED */
307 /* task_group_lock serializes add/remove of task groups and also changes to
308 * a task group's cpu shares.
310 static DEFINE_SPINLOCK(task_group_lock);
313 static int root_task_group_empty(void)
315 return list_empty(&root_task_group.children);
319 #ifdef CONFIG_FAIR_GROUP_SCHED
320 #ifdef CONFIG_USER_SCHED
321 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
322 #else /* !CONFIG_USER_SCHED */
323 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
324 #endif /* CONFIG_USER_SCHED */
327 * A weight of 0 or 1 can cause arithmetics problems.
328 * A weight of a cfs_rq is the sum of weights of which entities
329 * are queued on this cfs_rq, so a weight of a entity should not be
330 * too large, so as the shares value of a task group.
331 * (The default weight is 1024 - so there's no practical
332 * limitation from this.)
335 #define MAX_SHARES (1UL << 18)
337 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
340 /* Default task group.
341 * Every task in system belong to this group at bootup.
343 struct task_group init_task_group;
345 /* return group to which a task belongs */
346 static inline struct task_group *task_group(struct task_struct *p)
348 struct task_group *tg;
350 #ifdef CONFIG_USER_SCHED
352 tg = __task_cred(p)->user->tg;
354 #elif defined(CONFIG_CGROUP_SCHED)
355 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
356 struct task_group, css);
358 tg = &init_task_group;
363 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
364 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
366 #ifdef CONFIG_FAIR_GROUP_SCHED
367 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
368 p->se.parent = task_group(p)->se[cpu];
371 #ifdef CONFIG_RT_GROUP_SCHED
372 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
373 p->rt.parent = task_group(p)->rt_se[cpu];
380 static int root_task_group_empty(void)
386 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
387 static inline struct task_group *task_group(struct task_struct *p)
392 #endif /* CONFIG_GROUP_SCHED */
394 /* CFS-related fields in a runqueue */
396 struct load_weight load;
397 unsigned long nr_running;
402 struct rb_root tasks_timeline;
403 struct rb_node *rb_leftmost;
405 struct list_head tasks;
406 struct list_head *balance_iterator;
409 * 'curr' points to currently running entity on this cfs_rq.
410 * It is set to NULL otherwise (i.e when none are currently running).
412 struct sched_entity *curr, *next, *last;
414 unsigned int nr_spread_over;
416 #ifdef CONFIG_FAIR_GROUP_SCHED
417 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
420 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
421 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
422 * (like users, containers etc.)
424 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
425 * list is used during load balance.
427 struct list_head leaf_cfs_rq_list;
428 struct task_group *tg; /* group that "owns" this runqueue */
432 * the part of load.weight contributed by tasks
434 unsigned long task_weight;
437 * h_load = weight * f(tg)
439 * Where f(tg) is the recursive weight fraction assigned to
442 unsigned long h_load;
445 * this cpu's part of tg->shares
447 unsigned long shares;
450 * load.weight at the time we set shares
452 unsigned long rq_weight;
457 /* Real-Time classes' related field in a runqueue: */
459 struct rt_prio_array active;
460 unsigned long rt_nr_running;
461 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
463 int curr; /* highest queued rt task prio */
465 int next; /* next highest */
470 unsigned long rt_nr_migratory;
471 unsigned long rt_nr_total;
473 struct plist_head pushable_tasks;
478 /* Nests inside the rq lock: */
479 spinlock_t rt_runtime_lock;
481 #ifdef CONFIG_RT_GROUP_SCHED
482 unsigned long rt_nr_boosted;
485 struct list_head leaf_rt_rq_list;
486 struct task_group *tg;
487 struct sched_rt_entity *rt_se;
494 * We add the notion of a root-domain which will be used to define per-domain
495 * variables. Each exclusive cpuset essentially defines an island domain by
496 * fully partitioning the member cpus from any other cpuset. Whenever a new
497 * exclusive cpuset is created, we also create and attach a new root-domain
504 cpumask_var_t online;
507 * The "RT overload" flag: it gets set if a CPU has more than
508 * one runnable RT task.
510 cpumask_var_t rto_mask;
513 struct cpupri cpupri;
518 * By default the system creates a single root-domain with all cpus as
519 * members (mimicking the global state we have today).
521 static struct root_domain def_root_domain;
526 * This is the main, per-CPU runqueue data structure.
528 * Locking rule: those places that want to lock multiple runqueues
529 * (such as the load balancing or the thread migration code), lock
530 * acquire operations must be ordered by ascending &runqueue.
537 * nr_running and cpu_load should be in the same cacheline because
538 * remote CPUs use both these fields when doing load calculation.
540 unsigned long nr_running;
541 #define CPU_LOAD_IDX_MAX 5
542 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
544 unsigned long last_tick_seen;
545 unsigned char in_nohz_recently;
547 /* capture load from *all* tasks on this cpu: */
548 struct load_weight load;
549 unsigned long nr_load_updates;
551 u64 nr_migrations_in;
556 #ifdef CONFIG_FAIR_GROUP_SCHED
557 /* list of leaf cfs_rq on this cpu: */
558 struct list_head leaf_cfs_rq_list;
560 #ifdef CONFIG_RT_GROUP_SCHED
561 struct list_head leaf_rt_rq_list;
565 * This is part of a global counter where only the total sum
566 * over all CPUs matters. A task can increase this counter on
567 * one CPU and if it got migrated afterwards it may decrease
568 * it on another CPU. Always updated under the runqueue lock:
570 unsigned long nr_uninterruptible;
572 struct task_struct *curr, *idle;
573 unsigned long next_balance;
574 struct mm_struct *prev_mm;
581 struct root_domain *rd;
582 struct sched_domain *sd;
584 unsigned char idle_at_tick;
585 /* For active balancing */
589 /* cpu of this runqueue: */
593 unsigned long avg_load_per_task;
595 struct task_struct *migration_thread;
596 struct list_head migration_queue;
602 /* calc_load related fields */
603 unsigned long calc_load_update;
604 long calc_load_active;
606 #ifdef CONFIG_SCHED_HRTICK
608 int hrtick_csd_pending;
609 struct call_single_data hrtick_csd;
611 struct hrtimer hrtick_timer;
614 #ifdef CONFIG_SCHEDSTATS
616 struct sched_info rq_sched_info;
617 unsigned long long rq_cpu_time;
618 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
620 /* sys_sched_yield() stats */
621 unsigned int yld_count;
623 /* schedule() stats */
624 unsigned int sched_switch;
625 unsigned int sched_count;
626 unsigned int sched_goidle;
628 /* try_to_wake_up() stats */
629 unsigned int ttwu_count;
630 unsigned int ttwu_local;
633 unsigned int bkl_count;
637 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
639 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
641 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
644 static inline int cpu_of(struct rq *rq)
654 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
655 * See detach_destroy_domains: synchronize_sched for details.
657 * The domain tree of any CPU may only be accessed from within
658 * preempt-disabled sections.
660 #define for_each_domain(cpu, __sd) \
661 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
663 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
664 #define this_rq() (&__get_cpu_var(runqueues))
665 #define task_rq(p) cpu_rq(task_cpu(p))
666 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
667 #define raw_rq() (&__raw_get_cpu_var(runqueues))
669 inline void update_rq_clock(struct rq *rq)
671 rq->clock = sched_clock_cpu(cpu_of(rq));
675 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
677 #ifdef CONFIG_SCHED_DEBUG
678 # define const_debug __read_mostly
680 # define const_debug static const
686 * Returns true if the current cpu runqueue is locked.
687 * This interface allows printk to be called with the runqueue lock
688 * held and know whether or not it is OK to wake up the klogd.
690 int runqueue_is_locked(void)
693 struct rq *rq = cpu_rq(cpu);
696 ret = spin_is_locked(&rq->lock);
702 * Debugging: various feature bits
705 #define SCHED_FEAT(name, enabled) \
706 __SCHED_FEAT_##name ,
709 #include "sched_features.h"
714 #define SCHED_FEAT(name, enabled) \
715 (1UL << __SCHED_FEAT_##name) * enabled |
717 const_debug unsigned int sysctl_sched_features =
718 #include "sched_features.h"
723 #ifdef CONFIG_SCHED_DEBUG
724 #define SCHED_FEAT(name, enabled) \
727 static __read_mostly char *sched_feat_names[] = {
728 #include "sched_features.h"
734 static int sched_feat_show(struct seq_file *m, void *v)
738 for (i = 0; sched_feat_names[i]; i++) {
739 if (!(sysctl_sched_features & (1UL << i)))
741 seq_printf(m, "%s ", sched_feat_names[i]);
749 sched_feat_write(struct file *filp, const char __user *ubuf,
750 size_t cnt, loff_t *ppos)
760 if (copy_from_user(&buf, ubuf, cnt))
765 if (strncmp(buf, "NO_", 3) == 0) {
770 for (i = 0; sched_feat_names[i]; i++) {
771 int len = strlen(sched_feat_names[i]);
773 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
775 sysctl_sched_features &= ~(1UL << i);
777 sysctl_sched_features |= (1UL << i);
782 if (!sched_feat_names[i])
790 static int sched_feat_open(struct inode *inode, struct file *filp)
792 return single_open(filp, sched_feat_show, NULL);
795 static struct file_operations sched_feat_fops = {
796 .open = sched_feat_open,
797 .write = sched_feat_write,
800 .release = single_release,
803 static __init int sched_init_debug(void)
805 debugfs_create_file("sched_features", 0644, NULL, NULL,
810 late_initcall(sched_init_debug);
814 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
817 * Number of tasks to iterate in a single balance run.
818 * Limited because this is done with IRQs disabled.
820 const_debug unsigned int sysctl_sched_nr_migrate = 32;
823 * ratelimit for updating the group shares.
826 unsigned int sysctl_sched_shares_ratelimit = 250000;
829 * Inject some fuzzyness into changing the per-cpu group shares
830 * this avoids remote rq-locks at the expense of fairness.
833 unsigned int sysctl_sched_shares_thresh = 4;
836 * period over which we average the RT time consumption, measured
841 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
844 * period over which we measure -rt task cpu usage in us.
847 unsigned int sysctl_sched_rt_period = 1000000;
849 static __read_mostly int scheduler_running;
852 * part of the period that we allow rt tasks to run in us.
855 int sysctl_sched_rt_runtime = 950000;
857 static inline u64 global_rt_period(void)
859 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
862 static inline u64 global_rt_runtime(void)
864 if (sysctl_sched_rt_runtime < 0)
867 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
870 #ifndef prepare_arch_switch
871 # define prepare_arch_switch(next) do { } while (0)
873 #ifndef finish_arch_switch
874 # define finish_arch_switch(prev) do { } while (0)
877 static inline int task_current(struct rq *rq, struct task_struct *p)
879 return rq->curr == p;
882 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
883 static inline int task_running(struct rq *rq, struct task_struct *p)
885 return task_current(rq, p);
888 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
892 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
894 #ifdef CONFIG_DEBUG_SPINLOCK
895 /* this is a valid case when another task releases the spinlock */
896 rq->lock.owner = current;
899 * If we are tracking spinlock dependencies then we have to
900 * fix up the runqueue lock - which gets 'carried over' from
903 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
905 spin_unlock_irq(&rq->lock);
908 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
909 static inline int task_running(struct rq *rq, struct task_struct *p)
914 return task_current(rq, p);
918 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
922 * We can optimise this out completely for !SMP, because the
923 * SMP rebalancing from interrupt is the only thing that cares
928 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
929 spin_unlock_irq(&rq->lock);
931 spin_unlock(&rq->lock);
935 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
939 * After ->oncpu is cleared, the task can be moved to a different CPU.
940 * We must ensure this doesn't happen until the switch is completely
946 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
950 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
953 * __task_rq_lock - lock the runqueue a given task resides on.
954 * Must be called interrupts disabled.
956 static inline struct rq *__task_rq_lock(struct task_struct *p)
960 struct rq *rq = task_rq(p);
961 spin_lock(&rq->lock);
962 if (likely(rq == task_rq(p)))
964 spin_unlock(&rq->lock);
969 * task_rq_lock - lock the runqueue a given task resides on and disable
970 * interrupts. Note the ordering: we can safely lookup the task_rq without
971 * explicitly disabling preemption.
973 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
979 local_irq_save(*flags);
981 spin_lock(&rq->lock);
982 if (likely(rq == task_rq(p)))
984 spin_unlock_irqrestore(&rq->lock, *flags);
988 void task_rq_unlock_wait(struct task_struct *p)
990 struct rq *rq = task_rq(p);
992 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
993 spin_unlock_wait(&rq->lock);
996 static void __task_rq_unlock(struct rq *rq)
999 spin_unlock(&rq->lock);
1002 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1003 __releases(rq->lock)
1005 spin_unlock_irqrestore(&rq->lock, *flags);
1009 * this_rq_lock - lock this runqueue and disable interrupts.
1011 static struct rq *this_rq_lock(void)
1012 __acquires(rq->lock)
1016 local_irq_disable();
1018 spin_lock(&rq->lock);
1023 #ifdef CONFIG_SCHED_HRTICK
1025 * Use HR-timers to deliver accurate preemption points.
1027 * Its all a bit involved since we cannot program an hrt while holding the
1028 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1031 * When we get rescheduled we reprogram the hrtick_timer outside of the
1037 * - enabled by features
1038 * - hrtimer is actually high res
1040 static inline int hrtick_enabled(struct rq *rq)
1042 if (!sched_feat(HRTICK))
1044 if (!cpu_active(cpu_of(rq)))
1046 return hrtimer_is_hres_active(&rq->hrtick_timer);
1049 static void hrtick_clear(struct rq *rq)
1051 if (hrtimer_active(&rq->hrtick_timer))
1052 hrtimer_cancel(&rq->hrtick_timer);
1056 * High-resolution timer tick.
1057 * Runs from hardirq context with interrupts disabled.
1059 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1061 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1063 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1065 spin_lock(&rq->lock);
1066 update_rq_clock(rq);
1067 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1068 spin_unlock(&rq->lock);
1070 return HRTIMER_NORESTART;
1075 * called from hardirq (IPI) context
1077 static void __hrtick_start(void *arg)
1079 struct rq *rq = arg;
1081 spin_lock(&rq->lock);
1082 hrtimer_restart(&rq->hrtick_timer);
1083 rq->hrtick_csd_pending = 0;
1084 spin_unlock(&rq->lock);
1088 * Called to set the hrtick timer state.
1090 * called with rq->lock held and irqs disabled
1092 static void hrtick_start(struct rq *rq, u64 delay)
1094 struct hrtimer *timer = &rq->hrtick_timer;
1095 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1097 hrtimer_set_expires(timer, time);
1099 if (rq == this_rq()) {
1100 hrtimer_restart(timer);
1101 } else if (!rq->hrtick_csd_pending) {
1102 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1103 rq->hrtick_csd_pending = 1;
1108 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1110 int cpu = (int)(long)hcpu;
1113 case CPU_UP_CANCELED:
1114 case CPU_UP_CANCELED_FROZEN:
1115 case CPU_DOWN_PREPARE:
1116 case CPU_DOWN_PREPARE_FROZEN:
1118 case CPU_DEAD_FROZEN:
1119 hrtick_clear(cpu_rq(cpu));
1126 static __init void init_hrtick(void)
1128 hotcpu_notifier(hotplug_hrtick, 0);
1132 * Called to set the hrtick timer state.
1134 * called with rq->lock held and irqs disabled
1136 static void hrtick_start(struct rq *rq, u64 delay)
1138 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1139 HRTIMER_MODE_REL_PINNED, 0);
1142 static inline void init_hrtick(void)
1145 #endif /* CONFIG_SMP */
1147 static void init_rq_hrtick(struct rq *rq)
1150 rq->hrtick_csd_pending = 0;
1152 rq->hrtick_csd.flags = 0;
1153 rq->hrtick_csd.func = __hrtick_start;
1154 rq->hrtick_csd.info = rq;
1157 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1158 rq->hrtick_timer.function = hrtick;
1160 #else /* CONFIG_SCHED_HRTICK */
1161 static inline void hrtick_clear(struct rq *rq)
1165 static inline void init_rq_hrtick(struct rq *rq)
1169 static inline void init_hrtick(void)
1172 #endif /* CONFIG_SCHED_HRTICK */
1175 * resched_task - mark a task 'to be rescheduled now'.
1177 * On UP this means the setting of the need_resched flag, on SMP it
1178 * might also involve a cross-CPU call to trigger the scheduler on
1183 #ifndef tsk_is_polling
1184 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1187 static void resched_task(struct task_struct *p)
1191 assert_spin_locked(&task_rq(p)->lock);
1193 if (test_tsk_need_resched(p))
1196 set_tsk_need_resched(p);
1199 if (cpu == smp_processor_id())
1202 /* NEED_RESCHED must be visible before we test polling */
1204 if (!tsk_is_polling(p))
1205 smp_send_reschedule(cpu);
1208 static void resched_cpu(int cpu)
1210 struct rq *rq = cpu_rq(cpu);
1211 unsigned long flags;
1213 if (!spin_trylock_irqsave(&rq->lock, flags))
1215 resched_task(cpu_curr(cpu));
1216 spin_unlock_irqrestore(&rq->lock, flags);
1221 * When add_timer_on() enqueues a timer into the timer wheel of an
1222 * idle CPU then this timer might expire before the next timer event
1223 * which is scheduled to wake up that CPU. In case of a completely
1224 * idle system the next event might even be infinite time into the
1225 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1226 * leaves the inner idle loop so the newly added timer is taken into
1227 * account when the CPU goes back to idle and evaluates the timer
1228 * wheel for the next timer event.
1230 void wake_up_idle_cpu(int cpu)
1232 struct rq *rq = cpu_rq(cpu);
1234 if (cpu == smp_processor_id())
1238 * This is safe, as this function is called with the timer
1239 * wheel base lock of (cpu) held. When the CPU is on the way
1240 * to idle and has not yet set rq->curr to idle then it will
1241 * be serialized on the timer wheel base lock and take the new
1242 * timer into account automatically.
1244 if (rq->curr != rq->idle)
1248 * We can set TIF_RESCHED on the idle task of the other CPU
1249 * lockless. The worst case is that the other CPU runs the
1250 * idle task through an additional NOOP schedule()
1252 set_tsk_need_resched(rq->idle);
1254 /* NEED_RESCHED must be visible before we test polling */
1256 if (!tsk_is_polling(rq->idle))
1257 smp_send_reschedule(cpu);
1259 #endif /* CONFIG_NO_HZ */
1261 static u64 sched_avg_period(void)
1263 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1266 static void sched_avg_update(struct rq *rq)
1268 s64 period = sched_avg_period();
1270 while ((s64)(rq->clock - rq->age_stamp) > period) {
1271 rq->age_stamp += period;
1276 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1278 rq->rt_avg += rt_delta;
1279 sched_avg_update(rq);
1282 #else /* !CONFIG_SMP */
1283 static void resched_task(struct task_struct *p)
1285 assert_spin_locked(&task_rq(p)->lock);
1286 set_tsk_need_resched(p);
1289 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1292 #endif /* CONFIG_SMP */
1294 #if BITS_PER_LONG == 32
1295 # define WMULT_CONST (~0UL)
1297 # define WMULT_CONST (1UL << 32)
1300 #define WMULT_SHIFT 32
1303 * Shift right and round:
1305 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1308 * delta *= weight / lw
1310 static unsigned long
1311 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1312 struct load_weight *lw)
1316 if (!lw->inv_weight) {
1317 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1320 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1324 tmp = (u64)delta_exec * weight;
1326 * Check whether we'd overflow the 64-bit multiplication:
1328 if (unlikely(tmp > WMULT_CONST))
1329 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1332 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1334 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1337 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1343 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1350 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1351 * of tasks with abnormal "nice" values across CPUs the contribution that
1352 * each task makes to its run queue's load is weighted according to its
1353 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1354 * scaled version of the new time slice allocation that they receive on time
1358 #define WEIGHT_IDLEPRIO 3
1359 #define WMULT_IDLEPRIO 1431655765
1362 * Nice levels are multiplicative, with a gentle 10% change for every
1363 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1364 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1365 * that remained on nice 0.
1367 * The "10% effect" is relative and cumulative: from _any_ nice level,
1368 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1369 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1370 * If a task goes up by ~10% and another task goes down by ~10% then
1371 * the relative distance between them is ~25%.)
1373 static const int prio_to_weight[40] = {
1374 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1375 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1376 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1377 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1378 /* 0 */ 1024, 820, 655, 526, 423,
1379 /* 5 */ 335, 272, 215, 172, 137,
1380 /* 10 */ 110, 87, 70, 56, 45,
1381 /* 15 */ 36, 29, 23, 18, 15,
1385 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1387 * In cases where the weight does not change often, we can use the
1388 * precalculated inverse to speed up arithmetics by turning divisions
1389 * into multiplications:
1391 static const u32 prio_to_wmult[40] = {
1392 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1393 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1394 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1395 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1396 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1397 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1398 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1399 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1402 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1405 * runqueue iterator, to support SMP load-balancing between different
1406 * scheduling classes, without having to expose their internal data
1407 * structures to the load-balancing proper:
1409 struct rq_iterator {
1411 struct task_struct *(*start)(void *);
1412 struct task_struct *(*next)(void *);
1416 static unsigned long
1417 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1418 unsigned long max_load_move, struct sched_domain *sd,
1419 enum cpu_idle_type idle, int *all_pinned,
1420 int *this_best_prio, struct rq_iterator *iterator);
1423 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1424 struct sched_domain *sd, enum cpu_idle_type idle,
1425 struct rq_iterator *iterator);
1428 /* Time spent by the tasks of the cpu accounting group executing in ... */
1429 enum cpuacct_stat_index {
1430 CPUACCT_STAT_USER, /* ... user mode */
1431 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1433 CPUACCT_STAT_NSTATS,
1436 #ifdef CONFIG_CGROUP_CPUACCT
1437 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1438 static void cpuacct_update_stats(struct task_struct *tsk,
1439 enum cpuacct_stat_index idx, cputime_t val);
1441 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1442 static inline void cpuacct_update_stats(struct task_struct *tsk,
1443 enum cpuacct_stat_index idx, cputime_t val) {}
1446 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1448 update_load_add(&rq->load, load);
1451 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1453 update_load_sub(&rq->load, load);
1456 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1457 typedef int (*tg_visitor)(struct task_group *, void *);
1460 * Iterate the full tree, calling @down when first entering a node and @up when
1461 * leaving it for the final time.
1463 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1465 struct task_group *parent, *child;
1469 parent = &root_task_group;
1471 ret = (*down)(parent, data);
1474 list_for_each_entry_rcu(child, &parent->children, siblings) {
1481 ret = (*up)(parent, data);
1486 parent = parent->parent;
1495 static int tg_nop(struct task_group *tg, void *data)
1502 /* Used instead of source_load when we know the type == 0 */
1503 static unsigned long weighted_cpuload(const int cpu)
1505 return cpu_rq(cpu)->load.weight;
1509 * Return a low guess at the load of a migration-source cpu weighted
1510 * according to the scheduling class and "nice" value.
1512 * We want to under-estimate the load of migration sources, to
1513 * balance conservatively.
1515 static unsigned long source_load(int cpu, int type)
1517 struct rq *rq = cpu_rq(cpu);
1518 unsigned long total = weighted_cpuload(cpu);
1520 if (type == 0 || !sched_feat(LB_BIAS))
1523 return min(rq->cpu_load[type-1], total);
1527 * Return a high guess at the load of a migration-target cpu weighted
1528 * according to the scheduling class and "nice" value.
1530 static unsigned long target_load(int cpu, int type)
1532 struct rq *rq = cpu_rq(cpu);
1533 unsigned long total = weighted_cpuload(cpu);
1535 if (type == 0 || !sched_feat(LB_BIAS))
1538 return max(rq->cpu_load[type-1], total);
1541 static struct sched_group *group_of(int cpu)
1543 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1551 static unsigned long power_of(int cpu)
1553 struct sched_group *group = group_of(cpu);
1556 return SCHED_LOAD_SCALE;
1558 return group->cpu_power;
1561 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1563 static unsigned long cpu_avg_load_per_task(int cpu)
1565 struct rq *rq = cpu_rq(cpu);
1566 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1569 rq->avg_load_per_task = rq->load.weight / nr_running;
1571 rq->avg_load_per_task = 0;
1573 return rq->avg_load_per_task;
1576 #ifdef CONFIG_FAIR_GROUP_SCHED
1578 struct update_shares_data {
1579 unsigned long rq_weight[NR_CPUS];
1582 static DEFINE_PER_CPU(struct update_shares_data, update_shares_data);
1584 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1587 * Calculate and set the cpu's group shares.
1589 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1590 unsigned long sd_shares,
1591 unsigned long sd_rq_weight,
1592 struct update_shares_data *usd)
1594 unsigned long shares, rq_weight;
1597 rq_weight = usd->rq_weight[cpu];
1600 rq_weight = NICE_0_LOAD;
1604 * \Sum_j shares_j * rq_weight_i
1605 * shares_i = -----------------------------
1606 * \Sum_j rq_weight_j
1608 shares = (sd_shares * rq_weight) / sd_rq_weight;
1609 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1611 if (abs(shares - tg->se[cpu]->load.weight) >
1612 sysctl_sched_shares_thresh) {
1613 struct rq *rq = cpu_rq(cpu);
1614 unsigned long flags;
1616 spin_lock_irqsave(&rq->lock, flags);
1617 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1618 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1619 __set_se_shares(tg->se[cpu], shares);
1620 spin_unlock_irqrestore(&rq->lock, flags);
1625 * Re-compute the task group their per cpu shares over the given domain.
1626 * This needs to be done in a bottom-up fashion because the rq weight of a
1627 * parent group depends on the shares of its child groups.
1629 static int tg_shares_up(struct task_group *tg, void *data)
1631 unsigned long weight, rq_weight = 0, shares = 0;
1632 struct update_shares_data *usd;
1633 struct sched_domain *sd = data;
1634 unsigned long flags;
1640 local_irq_save(flags);
1641 usd = &__get_cpu_var(update_shares_data);
1643 for_each_cpu(i, sched_domain_span(sd)) {
1644 weight = tg->cfs_rq[i]->load.weight;
1645 usd->rq_weight[i] = weight;
1648 * If there are currently no tasks on the cpu pretend there
1649 * is one of average load so that when a new task gets to
1650 * run here it will not get delayed by group starvation.
1653 weight = NICE_0_LOAD;
1655 rq_weight += weight;
1656 shares += tg->cfs_rq[i]->shares;
1659 if ((!shares && rq_weight) || shares > tg->shares)
1660 shares = tg->shares;
1662 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1663 shares = tg->shares;
1665 for_each_cpu(i, sched_domain_span(sd))
1666 update_group_shares_cpu(tg, i, shares, rq_weight, usd);
1668 local_irq_restore(flags);
1674 * Compute the cpu's hierarchical load factor for each task group.
1675 * This needs to be done in a top-down fashion because the load of a child
1676 * group is a fraction of its parents load.
1678 static int tg_load_down(struct task_group *tg, void *data)
1681 long cpu = (long)data;
1684 load = cpu_rq(cpu)->load.weight;
1686 load = tg->parent->cfs_rq[cpu]->h_load;
1687 load *= tg->cfs_rq[cpu]->shares;
1688 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1691 tg->cfs_rq[cpu]->h_load = load;
1696 static void update_shares(struct sched_domain *sd)
1701 if (root_task_group_empty())
1704 now = cpu_clock(raw_smp_processor_id());
1705 elapsed = now - sd->last_update;
1707 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1708 sd->last_update = now;
1709 walk_tg_tree(tg_nop, tg_shares_up, sd);
1713 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1715 if (root_task_group_empty())
1718 spin_unlock(&rq->lock);
1720 spin_lock(&rq->lock);
1723 static void update_h_load(long cpu)
1725 if (root_task_group_empty())
1728 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1733 static inline void update_shares(struct sched_domain *sd)
1737 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1743 #ifdef CONFIG_PREEMPT
1745 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1748 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1749 * way at the expense of forcing extra atomic operations in all
1750 * invocations. This assures that the double_lock is acquired using the
1751 * same underlying policy as the spinlock_t on this architecture, which
1752 * reduces latency compared to the unfair variant below. However, it
1753 * also adds more overhead and therefore may reduce throughput.
1755 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1756 __releases(this_rq->lock)
1757 __acquires(busiest->lock)
1758 __acquires(this_rq->lock)
1760 spin_unlock(&this_rq->lock);
1761 double_rq_lock(this_rq, busiest);
1768 * Unfair double_lock_balance: Optimizes throughput at the expense of
1769 * latency by eliminating extra atomic operations when the locks are
1770 * already in proper order on entry. This favors lower cpu-ids and will
1771 * grant the double lock to lower cpus over higher ids under contention,
1772 * regardless of entry order into the function.
1774 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1775 __releases(this_rq->lock)
1776 __acquires(busiest->lock)
1777 __acquires(this_rq->lock)
1781 if (unlikely(!spin_trylock(&busiest->lock))) {
1782 if (busiest < this_rq) {
1783 spin_unlock(&this_rq->lock);
1784 spin_lock(&busiest->lock);
1785 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1788 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1793 #endif /* CONFIG_PREEMPT */
1796 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1798 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1800 if (unlikely(!irqs_disabled())) {
1801 /* printk() doesn't work good under rq->lock */
1802 spin_unlock(&this_rq->lock);
1806 return _double_lock_balance(this_rq, busiest);
1809 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1810 __releases(busiest->lock)
1812 spin_unlock(&busiest->lock);
1813 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1817 #ifdef CONFIG_FAIR_GROUP_SCHED
1818 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1821 cfs_rq->shares = shares;
1826 static void calc_load_account_active(struct rq *this_rq);
1828 #include "sched_stats.h"
1829 #include "sched_idletask.c"
1830 #include "sched_fair.c"
1831 #include "sched_rt.c"
1832 #ifdef CONFIG_SCHED_DEBUG
1833 # include "sched_debug.c"
1836 #define sched_class_highest (&rt_sched_class)
1837 #define for_each_class(class) \
1838 for (class = sched_class_highest; class; class = class->next)
1840 static void inc_nr_running(struct rq *rq)
1845 static void dec_nr_running(struct rq *rq)
1850 static void set_load_weight(struct task_struct *p)
1852 if (task_has_rt_policy(p)) {
1853 p->se.load.weight = prio_to_weight[0] * 2;
1854 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1859 * SCHED_IDLE tasks get minimal weight:
1861 if (p->policy == SCHED_IDLE) {
1862 p->se.load.weight = WEIGHT_IDLEPRIO;
1863 p->se.load.inv_weight = WMULT_IDLEPRIO;
1867 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1868 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1871 static void update_avg(u64 *avg, u64 sample)
1873 s64 diff = sample - *avg;
1877 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1880 p->se.start_runtime = p->se.sum_exec_runtime;
1882 sched_info_queued(p);
1883 p->sched_class->enqueue_task(rq, p, wakeup);
1887 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1890 if (p->se.last_wakeup) {
1891 update_avg(&p->se.avg_overlap,
1892 p->se.sum_exec_runtime - p->se.last_wakeup);
1893 p->se.last_wakeup = 0;
1895 update_avg(&p->se.avg_wakeup,
1896 sysctl_sched_wakeup_granularity);
1900 sched_info_dequeued(p);
1901 p->sched_class->dequeue_task(rq, p, sleep);
1906 * __normal_prio - return the priority that is based on the static prio
1908 static inline int __normal_prio(struct task_struct *p)
1910 return p->static_prio;
1914 * Calculate the expected normal priority: i.e. priority
1915 * without taking RT-inheritance into account. Might be
1916 * boosted by interactivity modifiers. Changes upon fork,
1917 * setprio syscalls, and whenever the interactivity
1918 * estimator recalculates.
1920 static inline int normal_prio(struct task_struct *p)
1924 if (task_has_rt_policy(p))
1925 prio = MAX_RT_PRIO-1 - p->rt_priority;
1927 prio = __normal_prio(p);
1932 * Calculate the current priority, i.e. the priority
1933 * taken into account by the scheduler. This value might
1934 * be boosted by RT tasks, or might be boosted by
1935 * interactivity modifiers. Will be RT if the task got
1936 * RT-boosted. If not then it returns p->normal_prio.
1938 static int effective_prio(struct task_struct *p)
1940 p->normal_prio = normal_prio(p);
1942 * If we are RT tasks or we were boosted to RT priority,
1943 * keep the priority unchanged. Otherwise, update priority
1944 * to the normal priority:
1946 if (!rt_prio(p->prio))
1947 return p->normal_prio;
1952 * activate_task - move a task to the runqueue.
1954 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1956 if (task_contributes_to_load(p))
1957 rq->nr_uninterruptible--;
1959 enqueue_task(rq, p, wakeup);
1964 * deactivate_task - remove a task from the runqueue.
1966 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1968 if (task_contributes_to_load(p))
1969 rq->nr_uninterruptible++;
1971 dequeue_task(rq, p, sleep);
1976 * task_curr - is this task currently executing on a CPU?
1977 * @p: the task in question.
1979 inline int task_curr(const struct task_struct *p)
1981 return cpu_curr(task_cpu(p)) == p;
1984 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1986 set_task_rq(p, cpu);
1989 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1990 * successfuly executed on another CPU. We must ensure that updates of
1991 * per-task data have been completed by this moment.
1994 task_thread_info(p)->cpu = cpu;
1998 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1999 const struct sched_class *prev_class,
2000 int oldprio, int running)
2002 if (prev_class != p->sched_class) {
2003 if (prev_class->switched_from)
2004 prev_class->switched_from(rq, p, running);
2005 p->sched_class->switched_to(rq, p, running);
2007 p->sched_class->prio_changed(rq, p, oldprio, running);
2012 * Is this task likely cache-hot:
2015 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2020 * Buddy candidates are cache hot:
2022 if (sched_feat(CACHE_HOT_BUDDY) &&
2023 (&p->se == cfs_rq_of(&p->se)->next ||
2024 &p->se == cfs_rq_of(&p->se)->last))
2027 if (p->sched_class != &fair_sched_class)
2030 if (sysctl_sched_migration_cost == -1)
2032 if (sysctl_sched_migration_cost == 0)
2035 delta = now - p->se.exec_start;
2037 return delta < (s64)sysctl_sched_migration_cost;
2041 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2043 int old_cpu = task_cpu(p);
2044 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2045 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2046 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2049 clock_offset = old_rq->clock - new_rq->clock;
2051 trace_sched_migrate_task(p, new_cpu);
2053 #ifdef CONFIG_SCHEDSTATS
2054 if (p->se.wait_start)
2055 p->se.wait_start -= clock_offset;
2056 if (p->se.sleep_start)
2057 p->se.sleep_start -= clock_offset;
2058 if (p->se.block_start)
2059 p->se.block_start -= clock_offset;
2061 if (old_cpu != new_cpu) {
2062 p->se.nr_migrations++;
2063 new_rq->nr_migrations_in++;
2064 #ifdef CONFIG_SCHEDSTATS
2065 if (task_hot(p, old_rq->clock, NULL))
2066 schedstat_inc(p, se.nr_forced2_migrations);
2068 perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2071 p->se.vruntime -= old_cfsrq->min_vruntime -
2072 new_cfsrq->min_vruntime;
2074 __set_task_cpu(p, new_cpu);
2077 struct migration_req {
2078 struct list_head list;
2080 struct task_struct *task;
2083 struct completion done;
2087 * The task's runqueue lock must be held.
2088 * Returns true if you have to wait for migration thread.
2091 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2093 struct rq *rq = task_rq(p);
2096 * If the task is not on a runqueue (and not running), then
2097 * it is sufficient to simply update the task's cpu field.
2099 if (!p->se.on_rq && !task_running(rq, p)) {
2100 set_task_cpu(p, dest_cpu);
2104 init_completion(&req->done);
2106 req->dest_cpu = dest_cpu;
2107 list_add(&req->list, &rq->migration_queue);
2113 * wait_task_context_switch - wait for a thread to complete at least one
2116 * @p must not be current.
2118 void wait_task_context_switch(struct task_struct *p)
2120 unsigned long nvcsw, nivcsw, flags;
2128 * The runqueue is assigned before the actual context
2129 * switch. We need to take the runqueue lock.
2131 * We could check initially without the lock but it is
2132 * very likely that we need to take the lock in every
2135 rq = task_rq_lock(p, &flags);
2136 running = task_running(rq, p);
2137 task_rq_unlock(rq, &flags);
2139 if (likely(!running))
2142 * The switch count is incremented before the actual
2143 * context switch. We thus wait for two switches to be
2144 * sure at least one completed.
2146 if ((p->nvcsw - nvcsw) > 1)
2148 if ((p->nivcsw - nivcsw) > 1)
2156 * wait_task_inactive - wait for a thread to unschedule.
2158 * If @match_state is nonzero, it's the @p->state value just checked and
2159 * not expected to change. If it changes, i.e. @p might have woken up,
2160 * then return zero. When we succeed in waiting for @p to be off its CPU,
2161 * we return a positive number (its total switch count). If a second call
2162 * a short while later returns the same number, the caller can be sure that
2163 * @p has remained unscheduled the whole time.
2165 * The caller must ensure that the task *will* unschedule sometime soon,
2166 * else this function might spin for a *long* time. This function can't
2167 * be called with interrupts off, or it may introduce deadlock with
2168 * smp_call_function() if an IPI is sent by the same process we are
2169 * waiting to become inactive.
2171 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2173 unsigned long flags;
2180 * We do the initial early heuristics without holding
2181 * any task-queue locks at all. We'll only try to get
2182 * the runqueue lock when things look like they will
2188 * If the task is actively running on another CPU
2189 * still, just relax and busy-wait without holding
2192 * NOTE! Since we don't hold any locks, it's not
2193 * even sure that "rq" stays as the right runqueue!
2194 * But we don't care, since "task_running()" will
2195 * return false if the runqueue has changed and p
2196 * is actually now running somewhere else!
2198 while (task_running(rq, p)) {
2199 if (match_state && unlikely(p->state != match_state))
2205 * Ok, time to look more closely! We need the rq
2206 * lock now, to be *sure*. If we're wrong, we'll
2207 * just go back and repeat.
2209 rq = task_rq_lock(p, &flags);
2210 trace_sched_wait_task(rq, p);
2211 running = task_running(rq, p);
2212 on_rq = p->se.on_rq;
2214 if (!match_state || p->state == match_state)
2215 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2216 task_rq_unlock(rq, &flags);
2219 * If it changed from the expected state, bail out now.
2221 if (unlikely(!ncsw))
2225 * Was it really running after all now that we
2226 * checked with the proper locks actually held?
2228 * Oops. Go back and try again..
2230 if (unlikely(running)) {
2236 * It's not enough that it's not actively running,
2237 * it must be off the runqueue _entirely_, and not
2240 * So if it was still runnable (but just not actively
2241 * running right now), it's preempted, and we should
2242 * yield - it could be a while.
2244 if (unlikely(on_rq)) {
2245 schedule_timeout_uninterruptible(1);
2250 * Ahh, all good. It wasn't running, and it wasn't
2251 * runnable, which means that it will never become
2252 * running in the future either. We're all done!
2261 * kick_process - kick a running thread to enter/exit the kernel
2262 * @p: the to-be-kicked thread
2264 * Cause a process which is running on another CPU to enter
2265 * kernel-mode, without any delay. (to get signals handled.)
2267 * NOTE: this function doesnt have to take the runqueue lock,
2268 * because all it wants to ensure is that the remote task enters
2269 * the kernel. If the IPI races and the task has been migrated
2270 * to another CPU then no harm is done and the purpose has been
2273 void kick_process(struct task_struct *p)
2279 if ((cpu != smp_processor_id()) && task_curr(p))
2280 smp_send_reschedule(cpu);
2283 EXPORT_SYMBOL_GPL(kick_process);
2284 #endif /* CONFIG_SMP */
2287 * task_oncpu_function_call - call a function on the cpu on which a task runs
2288 * @p: the task to evaluate
2289 * @func: the function to be called
2290 * @info: the function call argument
2292 * Calls the function @func when the task is currently running. This might
2293 * be on the current CPU, which just calls the function directly
2295 void task_oncpu_function_call(struct task_struct *p,
2296 void (*func) (void *info), void *info)
2303 smp_call_function_single(cpu, func, info, 1);
2308 * try_to_wake_up - wake up a thread
2309 * @p: the to-be-woken-up thread
2310 * @state: the mask of task states that can be woken
2311 * @sync: do a synchronous wakeup?
2313 * Put it on the run-queue if it's not already there. The "current"
2314 * thread is always on the run-queue (except when the actual
2315 * re-schedule is in progress), and as such you're allowed to do
2316 * the simpler "current->state = TASK_RUNNING" to mark yourself
2317 * runnable without the overhead of this.
2319 * returns failure only if the task is already active.
2321 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2323 int cpu, orig_cpu, this_cpu, success = 0;
2324 unsigned long flags;
2327 if (!sched_feat(SYNC_WAKEUPS))
2330 this_cpu = get_cpu();
2333 rq = task_rq_lock(p, &flags);
2334 update_rq_clock(rq);
2335 if (!(p->state & state))
2345 if (unlikely(task_running(rq, p)))
2349 * In order to handle concurrent wakeups and release the rq->lock
2350 * we put the task in TASK_WAKING state.
2352 p->state = TASK_WAKING;
2353 task_rq_unlock(rq, &flags);
2355 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_WAKE, sync);
2356 if (cpu != orig_cpu)
2357 set_task_cpu(p, cpu);
2359 rq = task_rq_lock(p, &flags);
2360 WARN_ON(p->state != TASK_WAKING);
2363 #ifdef CONFIG_SCHEDSTATS
2364 schedstat_inc(rq, ttwu_count);
2365 if (cpu == this_cpu)
2366 schedstat_inc(rq, ttwu_local);
2368 struct sched_domain *sd;
2369 for_each_domain(this_cpu, sd) {
2370 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2371 schedstat_inc(sd, ttwu_wake_remote);
2376 #endif /* CONFIG_SCHEDSTATS */
2379 #endif /* CONFIG_SMP */
2380 schedstat_inc(p, se.nr_wakeups);
2382 schedstat_inc(p, se.nr_wakeups_sync);
2383 if (orig_cpu != cpu)
2384 schedstat_inc(p, se.nr_wakeups_migrate);
2385 if (cpu == this_cpu)
2386 schedstat_inc(p, se.nr_wakeups_local);
2388 schedstat_inc(p, se.nr_wakeups_remote);
2389 activate_task(rq, p, 1);
2393 * Only attribute actual wakeups done by this task.
2395 if (!in_interrupt()) {
2396 struct sched_entity *se = ¤t->se;
2397 u64 sample = se->sum_exec_runtime;
2399 if (se->last_wakeup)
2400 sample -= se->last_wakeup;
2402 sample -= se->start_runtime;
2403 update_avg(&se->avg_wakeup, sample);
2405 se->last_wakeup = se->sum_exec_runtime;
2409 trace_sched_wakeup(rq, p, success);
2410 check_preempt_curr(rq, p, sync);
2412 p->state = TASK_RUNNING;
2414 if (p->sched_class->task_wake_up)
2415 p->sched_class->task_wake_up(rq, p);
2418 task_rq_unlock(rq, &flags);
2425 * wake_up_process - Wake up a specific process
2426 * @p: The process to be woken up.
2428 * Attempt to wake up the nominated process and move it to the set of runnable
2429 * processes. Returns 1 if the process was woken up, 0 if it was already
2432 * It may be assumed that this function implies a write memory barrier before
2433 * changing the task state if and only if any tasks are woken up.
2435 int wake_up_process(struct task_struct *p)
2437 return try_to_wake_up(p, TASK_ALL, 0);
2439 EXPORT_SYMBOL(wake_up_process);
2441 int wake_up_state(struct task_struct *p, unsigned int state)
2443 return try_to_wake_up(p, state, 0);
2447 * Perform scheduler related setup for a newly forked process p.
2448 * p is forked by current.
2450 * __sched_fork() is basic setup used by init_idle() too:
2452 static void __sched_fork(struct task_struct *p)
2454 p->se.exec_start = 0;
2455 p->se.sum_exec_runtime = 0;
2456 p->se.prev_sum_exec_runtime = 0;
2457 p->se.nr_migrations = 0;
2458 p->se.last_wakeup = 0;
2459 p->se.avg_overlap = 0;
2460 p->se.start_runtime = 0;
2461 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2463 #ifdef CONFIG_SCHEDSTATS
2464 p->se.wait_start = 0;
2466 p->se.wait_count = 0;
2469 p->se.sleep_start = 0;
2470 p->se.sleep_max = 0;
2471 p->se.sum_sleep_runtime = 0;
2473 p->se.block_start = 0;
2474 p->se.block_max = 0;
2476 p->se.slice_max = 0;
2478 p->se.nr_migrations_cold = 0;
2479 p->se.nr_failed_migrations_affine = 0;
2480 p->se.nr_failed_migrations_running = 0;
2481 p->se.nr_failed_migrations_hot = 0;
2482 p->se.nr_forced_migrations = 0;
2483 p->se.nr_forced2_migrations = 0;
2485 p->se.nr_wakeups = 0;
2486 p->se.nr_wakeups_sync = 0;
2487 p->se.nr_wakeups_migrate = 0;
2488 p->se.nr_wakeups_local = 0;
2489 p->se.nr_wakeups_remote = 0;
2490 p->se.nr_wakeups_affine = 0;
2491 p->se.nr_wakeups_affine_attempts = 0;
2492 p->se.nr_wakeups_passive = 0;
2493 p->se.nr_wakeups_idle = 0;
2497 INIT_LIST_HEAD(&p->rt.run_list);
2499 INIT_LIST_HEAD(&p->se.group_node);
2501 #ifdef CONFIG_PREEMPT_NOTIFIERS
2502 INIT_HLIST_HEAD(&p->preempt_notifiers);
2506 * We mark the process as running here, but have not actually
2507 * inserted it onto the runqueue yet. This guarantees that
2508 * nobody will actually run it, and a signal or other external
2509 * event cannot wake it up and insert it on the runqueue either.
2511 p->state = TASK_RUNNING;
2515 * fork()/clone()-time setup:
2517 void sched_fork(struct task_struct *p, int clone_flags)
2519 int cpu = get_cpu();
2524 * Make sure we do not leak PI boosting priority to the child.
2526 p->prio = current->normal_prio;
2529 * Revert to default priority/policy on fork if requested.
2531 if (unlikely(p->sched_reset_on_fork)) {
2532 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR)
2533 p->policy = SCHED_NORMAL;
2535 if (p->normal_prio < DEFAULT_PRIO)
2536 p->prio = DEFAULT_PRIO;
2538 if (PRIO_TO_NICE(p->static_prio) < 0) {
2539 p->static_prio = NICE_TO_PRIO(0);
2544 * We don't need the reset flag anymore after the fork. It has
2545 * fulfilled its duty:
2547 p->sched_reset_on_fork = 0;
2550 if (!rt_prio(p->prio))
2551 p->sched_class = &fair_sched_class;
2554 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_FORK, 0);
2556 set_task_cpu(p, cpu);
2558 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2559 if (likely(sched_info_on()))
2560 memset(&p->sched_info, 0, sizeof(p->sched_info));
2562 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2565 #ifdef CONFIG_PREEMPT
2566 /* Want to start with kernel preemption disabled. */
2567 task_thread_info(p)->preempt_count = 1;
2569 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2575 * wake_up_new_task - wake up a newly created task for the first time.
2577 * This function will do some initial scheduler statistics housekeeping
2578 * that must be done for every newly created context, then puts the task
2579 * on the runqueue and wakes it.
2581 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2583 unsigned long flags;
2586 rq = task_rq_lock(p, &flags);
2587 BUG_ON(p->state != TASK_RUNNING);
2588 update_rq_clock(rq);
2590 p->prio = effective_prio(p);
2592 if (!p->sched_class->task_new || !current->se.on_rq) {
2593 activate_task(rq, p, 0);
2596 * Let the scheduling class do new task startup
2597 * management (if any):
2599 p->sched_class->task_new(rq, p);
2602 trace_sched_wakeup_new(rq, p, 1);
2603 check_preempt_curr(rq, p, 0);
2605 if (p->sched_class->task_wake_up)
2606 p->sched_class->task_wake_up(rq, p);
2608 task_rq_unlock(rq, &flags);
2611 #ifdef CONFIG_PREEMPT_NOTIFIERS
2614 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2615 * @notifier: notifier struct to register
2617 void preempt_notifier_register(struct preempt_notifier *notifier)
2619 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2621 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2624 * preempt_notifier_unregister - no longer interested in preemption notifications
2625 * @notifier: notifier struct to unregister
2627 * This is safe to call from within a preemption notifier.
2629 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2631 hlist_del(¬ifier->link);
2633 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2635 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2637 struct preempt_notifier *notifier;
2638 struct hlist_node *node;
2640 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2641 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2645 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2646 struct task_struct *next)
2648 struct preempt_notifier *notifier;
2649 struct hlist_node *node;
2651 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2652 notifier->ops->sched_out(notifier, next);
2655 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2657 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2662 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2663 struct task_struct *next)
2667 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2670 * prepare_task_switch - prepare to switch tasks
2671 * @rq: the runqueue preparing to switch
2672 * @prev: the current task that is being switched out
2673 * @next: the task we are going to switch to.
2675 * This is called with the rq lock held and interrupts off. It must
2676 * be paired with a subsequent finish_task_switch after the context
2679 * prepare_task_switch sets up locking and calls architecture specific
2683 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2684 struct task_struct *next)
2686 fire_sched_out_preempt_notifiers(prev, next);
2687 prepare_lock_switch(rq, next);
2688 prepare_arch_switch(next);
2692 * finish_task_switch - clean up after a task-switch
2693 * @rq: runqueue associated with task-switch
2694 * @prev: the thread we just switched away from.
2696 * finish_task_switch must be called after the context switch, paired
2697 * with a prepare_task_switch call before the context switch.
2698 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2699 * and do any other architecture-specific cleanup actions.
2701 * Note that we may have delayed dropping an mm in context_switch(). If
2702 * so, we finish that here outside of the runqueue lock. (Doing it
2703 * with the lock held can cause deadlocks; see schedule() for
2706 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2707 __releases(rq->lock)
2709 struct mm_struct *mm = rq->prev_mm;
2715 * A task struct has one reference for the use as "current".
2716 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2717 * schedule one last time. The schedule call will never return, and
2718 * the scheduled task must drop that reference.
2719 * The test for TASK_DEAD must occur while the runqueue locks are
2720 * still held, otherwise prev could be scheduled on another cpu, die
2721 * there before we look at prev->state, and then the reference would
2723 * Manfred Spraul <manfred@colorfullife.com>
2725 prev_state = prev->state;
2726 finish_arch_switch(prev);
2727 perf_counter_task_sched_in(current, cpu_of(rq));
2728 finish_lock_switch(rq, prev);
2730 fire_sched_in_preempt_notifiers(current);
2733 if (unlikely(prev_state == TASK_DEAD)) {
2735 * Remove function-return probe instances associated with this
2736 * task and put them back on the free list.
2738 kprobe_flush_task(prev);
2739 put_task_struct(prev);
2745 /* assumes rq->lock is held */
2746 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2748 if (prev->sched_class->pre_schedule)
2749 prev->sched_class->pre_schedule(rq, prev);
2752 /* rq->lock is NOT held, but preemption is disabled */
2753 static inline void post_schedule(struct rq *rq)
2755 if (rq->post_schedule) {
2756 unsigned long flags;
2758 spin_lock_irqsave(&rq->lock, flags);
2759 if (rq->curr->sched_class->post_schedule)
2760 rq->curr->sched_class->post_schedule(rq);
2761 spin_unlock_irqrestore(&rq->lock, flags);
2763 rq->post_schedule = 0;
2769 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2773 static inline void post_schedule(struct rq *rq)
2780 * schedule_tail - first thing a freshly forked thread must call.
2781 * @prev: the thread we just switched away from.
2783 asmlinkage void schedule_tail(struct task_struct *prev)
2784 __releases(rq->lock)
2786 struct rq *rq = this_rq();
2788 finish_task_switch(rq, prev);
2791 * FIXME: do we need to worry about rq being invalidated by the
2796 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2797 /* In this case, finish_task_switch does not reenable preemption */
2800 if (current->set_child_tid)
2801 put_user(task_pid_vnr(current), current->set_child_tid);
2805 * context_switch - switch to the new MM and the new
2806 * thread's register state.
2809 context_switch(struct rq *rq, struct task_struct *prev,
2810 struct task_struct *next)
2812 struct mm_struct *mm, *oldmm;
2814 prepare_task_switch(rq, prev, next);
2815 trace_sched_switch(rq, prev, next);
2817 oldmm = prev->active_mm;
2819 * For paravirt, this is coupled with an exit in switch_to to
2820 * combine the page table reload and the switch backend into
2823 arch_start_context_switch(prev);
2825 if (unlikely(!mm)) {
2826 next->active_mm = oldmm;
2827 atomic_inc(&oldmm->mm_count);
2828 enter_lazy_tlb(oldmm, next);
2830 switch_mm(oldmm, mm, next);
2832 if (unlikely(!prev->mm)) {
2833 prev->active_mm = NULL;
2834 rq->prev_mm = oldmm;
2837 * Since the runqueue lock will be released by the next
2838 * task (which is an invalid locking op but in the case
2839 * of the scheduler it's an obvious special-case), so we
2840 * do an early lockdep release here:
2842 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2843 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2846 /* Here we just switch the register state and the stack. */
2847 switch_to(prev, next, prev);
2851 * this_rq must be evaluated again because prev may have moved
2852 * CPUs since it called schedule(), thus the 'rq' on its stack
2853 * frame will be invalid.
2855 finish_task_switch(this_rq(), prev);
2859 * nr_running, nr_uninterruptible and nr_context_switches:
2861 * externally visible scheduler statistics: current number of runnable
2862 * threads, current number of uninterruptible-sleeping threads, total
2863 * number of context switches performed since bootup.
2865 unsigned long nr_running(void)
2867 unsigned long i, sum = 0;
2869 for_each_online_cpu(i)
2870 sum += cpu_rq(i)->nr_running;
2875 unsigned long nr_uninterruptible(void)
2877 unsigned long i, sum = 0;
2879 for_each_possible_cpu(i)
2880 sum += cpu_rq(i)->nr_uninterruptible;
2883 * Since we read the counters lockless, it might be slightly
2884 * inaccurate. Do not allow it to go below zero though:
2886 if (unlikely((long)sum < 0))
2892 unsigned long long nr_context_switches(void)
2895 unsigned long long sum = 0;
2897 for_each_possible_cpu(i)
2898 sum += cpu_rq(i)->nr_switches;
2903 unsigned long nr_iowait(void)
2905 unsigned long i, sum = 0;
2907 for_each_possible_cpu(i)
2908 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2913 /* Variables and functions for calc_load */
2914 static atomic_long_t calc_load_tasks;
2915 static unsigned long calc_load_update;
2916 unsigned long avenrun[3];
2917 EXPORT_SYMBOL(avenrun);
2920 * get_avenrun - get the load average array
2921 * @loads: pointer to dest load array
2922 * @offset: offset to add
2923 * @shift: shift count to shift the result left
2925 * These values are estimates at best, so no need for locking.
2927 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2929 loads[0] = (avenrun[0] + offset) << shift;
2930 loads[1] = (avenrun[1] + offset) << shift;
2931 loads[2] = (avenrun[2] + offset) << shift;
2934 static unsigned long
2935 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2938 load += active * (FIXED_1 - exp);
2939 return load >> FSHIFT;
2943 * calc_load - update the avenrun load estimates 10 ticks after the
2944 * CPUs have updated calc_load_tasks.
2946 void calc_global_load(void)
2948 unsigned long upd = calc_load_update + 10;
2951 if (time_before(jiffies, upd))
2954 active = atomic_long_read(&calc_load_tasks);
2955 active = active > 0 ? active * FIXED_1 : 0;
2957 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2958 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2959 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2961 calc_load_update += LOAD_FREQ;
2965 * Either called from update_cpu_load() or from a cpu going idle
2967 static void calc_load_account_active(struct rq *this_rq)
2969 long nr_active, delta;
2971 nr_active = this_rq->nr_running;
2972 nr_active += (long) this_rq->nr_uninterruptible;
2974 if (nr_active != this_rq->calc_load_active) {
2975 delta = nr_active - this_rq->calc_load_active;
2976 this_rq->calc_load_active = nr_active;
2977 atomic_long_add(delta, &calc_load_tasks);
2982 * Externally visible per-cpu scheduler statistics:
2983 * cpu_nr_migrations(cpu) - number of migrations into that cpu
2985 u64 cpu_nr_migrations(int cpu)
2987 return cpu_rq(cpu)->nr_migrations_in;
2991 * Update rq->cpu_load[] statistics. This function is usually called every
2992 * scheduler tick (TICK_NSEC).
2994 static void update_cpu_load(struct rq *this_rq)
2996 unsigned long this_load = this_rq->load.weight;
2999 this_rq->nr_load_updates++;
3001 /* Update our load: */
3002 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3003 unsigned long old_load, new_load;
3005 /* scale is effectively 1 << i now, and >> i divides by scale */
3007 old_load = this_rq->cpu_load[i];
3008 new_load = this_load;
3010 * Round up the averaging division if load is increasing. This
3011 * prevents us from getting stuck on 9 if the load is 10, for
3014 if (new_load > old_load)
3015 new_load += scale-1;
3016 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3019 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3020 this_rq->calc_load_update += LOAD_FREQ;
3021 calc_load_account_active(this_rq);
3028 * double_rq_lock - safely lock two runqueues
3030 * Note this does not disable interrupts like task_rq_lock,
3031 * you need to do so manually before calling.
3033 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3034 __acquires(rq1->lock)
3035 __acquires(rq2->lock)
3037 BUG_ON(!irqs_disabled());
3039 spin_lock(&rq1->lock);
3040 __acquire(rq2->lock); /* Fake it out ;) */
3043 spin_lock(&rq1->lock);
3044 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3046 spin_lock(&rq2->lock);
3047 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3050 update_rq_clock(rq1);
3051 update_rq_clock(rq2);
3055 * double_rq_unlock - safely unlock two runqueues
3057 * Note this does not restore interrupts like task_rq_unlock,
3058 * you need to do so manually after calling.
3060 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3061 __releases(rq1->lock)
3062 __releases(rq2->lock)
3064 spin_unlock(&rq1->lock);
3066 spin_unlock(&rq2->lock);
3068 __release(rq2->lock);
3072 * If dest_cpu is allowed for this process, migrate the task to it.
3073 * This is accomplished by forcing the cpu_allowed mask to only
3074 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3075 * the cpu_allowed mask is restored.
3077 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3079 struct migration_req req;
3080 unsigned long flags;
3083 rq = task_rq_lock(p, &flags);
3084 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3085 || unlikely(!cpu_active(dest_cpu)))
3088 /* force the process onto the specified CPU */
3089 if (migrate_task(p, dest_cpu, &req)) {
3090 /* Need to wait for migration thread (might exit: take ref). */
3091 struct task_struct *mt = rq->migration_thread;
3093 get_task_struct(mt);
3094 task_rq_unlock(rq, &flags);
3095 wake_up_process(mt);
3096 put_task_struct(mt);
3097 wait_for_completion(&req.done);
3102 task_rq_unlock(rq, &flags);
3106 * sched_exec - execve() is a valuable balancing opportunity, because at
3107 * this point the task has the smallest effective memory and cache footprint.
3109 void sched_exec(void)
3111 int new_cpu, this_cpu = get_cpu();
3112 new_cpu = current->sched_class->select_task_rq(current, SD_BALANCE_EXEC, 0);
3114 if (new_cpu != this_cpu)
3115 sched_migrate_task(current, new_cpu);
3119 * pull_task - move a task from a remote runqueue to the local runqueue.
3120 * Both runqueues must be locked.
3122 static void pull_task(struct rq *src_rq, struct task_struct *p,
3123 struct rq *this_rq, int this_cpu)
3125 deactivate_task(src_rq, p, 0);
3126 set_task_cpu(p, this_cpu);
3127 activate_task(this_rq, p, 0);
3129 * Note that idle threads have a prio of MAX_PRIO, for this test
3130 * to be always true for them.
3132 check_preempt_curr(this_rq, p, 0);
3136 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3139 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3140 struct sched_domain *sd, enum cpu_idle_type idle,
3143 int tsk_cache_hot = 0;
3145 * We do not migrate tasks that are:
3146 * 1) running (obviously), or
3147 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3148 * 3) are cache-hot on their current CPU.
3150 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3151 schedstat_inc(p, se.nr_failed_migrations_affine);
3156 if (task_running(rq, p)) {
3157 schedstat_inc(p, se.nr_failed_migrations_running);
3162 * Aggressive migration if:
3163 * 1) task is cache cold, or
3164 * 2) too many balance attempts have failed.
3167 tsk_cache_hot = task_hot(p, rq->clock, sd);
3168 if (!tsk_cache_hot ||
3169 sd->nr_balance_failed > sd->cache_nice_tries) {
3170 #ifdef CONFIG_SCHEDSTATS
3171 if (tsk_cache_hot) {
3172 schedstat_inc(sd, lb_hot_gained[idle]);
3173 schedstat_inc(p, se.nr_forced_migrations);
3179 if (tsk_cache_hot) {
3180 schedstat_inc(p, se.nr_failed_migrations_hot);
3186 static unsigned long
3187 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3188 unsigned long max_load_move, struct sched_domain *sd,
3189 enum cpu_idle_type idle, int *all_pinned,
3190 int *this_best_prio, struct rq_iterator *iterator)
3192 int loops = 0, pulled = 0, pinned = 0;
3193 struct task_struct *p;
3194 long rem_load_move = max_load_move;
3196 if (max_load_move == 0)
3202 * Start the load-balancing iterator:
3204 p = iterator->start(iterator->arg);
3206 if (!p || loops++ > sysctl_sched_nr_migrate)
3209 if ((p->se.load.weight >> 1) > rem_load_move ||
3210 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3211 p = iterator->next(iterator->arg);
3215 pull_task(busiest, p, this_rq, this_cpu);
3217 rem_load_move -= p->se.load.weight;
3219 #ifdef CONFIG_PREEMPT
3221 * NEWIDLE balancing is a source of latency, so preemptible kernels
3222 * will stop after the first task is pulled to minimize the critical
3225 if (idle == CPU_NEWLY_IDLE)
3230 * We only want to steal up to the prescribed amount of weighted load.
3232 if (rem_load_move > 0) {
3233 if (p->prio < *this_best_prio)
3234 *this_best_prio = p->prio;
3235 p = iterator->next(iterator->arg);
3240 * Right now, this is one of only two places pull_task() is called,
3241 * so we can safely collect pull_task() stats here rather than
3242 * inside pull_task().
3244 schedstat_add(sd, lb_gained[idle], pulled);
3247 *all_pinned = pinned;
3249 return max_load_move - rem_load_move;
3253 * move_tasks tries to move up to max_load_move weighted load from busiest to
3254 * this_rq, as part of a balancing operation within domain "sd".
3255 * Returns 1 if successful and 0 otherwise.
3257 * Called with both runqueues locked.
3259 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3260 unsigned long max_load_move,
3261 struct sched_domain *sd, enum cpu_idle_type idle,
3264 const struct sched_class *class = sched_class_highest;
3265 unsigned long total_load_moved = 0;
3266 int this_best_prio = this_rq->curr->prio;
3270 class->load_balance(this_rq, this_cpu, busiest,
3271 max_load_move - total_load_moved,
3272 sd, idle, all_pinned, &this_best_prio);
3273 class = class->next;
3275 #ifdef CONFIG_PREEMPT
3277 * NEWIDLE balancing is a source of latency, so preemptible
3278 * kernels will stop after the first task is pulled to minimize
3279 * the critical section.
3281 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3284 } while (class && max_load_move > total_load_moved);
3286 return total_load_moved > 0;
3290 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3291 struct sched_domain *sd, enum cpu_idle_type idle,
3292 struct rq_iterator *iterator)
3294 struct task_struct *p = iterator->start(iterator->arg);
3298 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3299 pull_task(busiest, p, this_rq, this_cpu);
3301 * Right now, this is only the second place pull_task()
3302 * is called, so we can safely collect pull_task()
3303 * stats here rather than inside pull_task().
3305 schedstat_inc(sd, lb_gained[idle]);
3309 p = iterator->next(iterator->arg);
3316 * move_one_task tries to move exactly one task from busiest to this_rq, as
3317 * part of active balancing operations within "domain".
3318 * Returns 1 if successful and 0 otherwise.
3320 * Called with both runqueues locked.
3322 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3323 struct sched_domain *sd, enum cpu_idle_type idle)
3325 const struct sched_class *class;
3327 for_each_class(class) {
3328 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3334 /********** Helpers for find_busiest_group ************************/
3336 * sd_lb_stats - Structure to store the statistics of a sched_domain
3337 * during load balancing.
3339 struct sd_lb_stats {
3340 struct sched_group *busiest; /* Busiest group in this sd */
3341 struct sched_group *this; /* Local group in this sd */
3342 unsigned long total_load; /* Total load of all groups in sd */
3343 unsigned long total_pwr; /* Total power of all groups in sd */
3344 unsigned long avg_load; /* Average load across all groups in sd */
3346 /** Statistics of this group */
3347 unsigned long this_load;
3348 unsigned long this_load_per_task;
3349 unsigned long this_nr_running;
3351 /* Statistics of the busiest group */
3352 unsigned long max_load;
3353 unsigned long busiest_load_per_task;
3354 unsigned long busiest_nr_running;
3356 int group_imb; /* Is there imbalance in this sd */
3357 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3358 int power_savings_balance; /* Is powersave balance needed for this sd */
3359 struct sched_group *group_min; /* Least loaded group in sd */
3360 struct sched_group *group_leader; /* Group which relieves group_min */
3361 unsigned long min_load_per_task; /* load_per_task in group_min */
3362 unsigned long leader_nr_running; /* Nr running of group_leader */
3363 unsigned long min_nr_running; /* Nr running of group_min */
3368 * sg_lb_stats - stats of a sched_group required for load_balancing
3370 struct sg_lb_stats {
3371 unsigned long avg_load; /*Avg load across the CPUs of the group */
3372 unsigned long group_load; /* Total load over the CPUs of the group */
3373 unsigned long sum_nr_running; /* Nr tasks running in the group */
3374 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3375 unsigned long group_capacity;
3376 int group_imb; /* Is there an imbalance in the group ? */
3380 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3381 * @group: The group whose first cpu is to be returned.
3383 static inline unsigned int group_first_cpu(struct sched_group *group)
3385 return cpumask_first(sched_group_cpus(group));
3389 * get_sd_load_idx - Obtain the load index for a given sched domain.
3390 * @sd: The sched_domain whose load_idx is to be obtained.
3391 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3393 static inline int get_sd_load_idx(struct sched_domain *sd,
3394 enum cpu_idle_type idle)
3400 load_idx = sd->busy_idx;
3403 case CPU_NEWLY_IDLE:
3404 load_idx = sd->newidle_idx;
3407 load_idx = sd->idle_idx;
3415 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3417 * init_sd_power_savings_stats - Initialize power savings statistics for
3418 * the given sched_domain, during load balancing.
3420 * @sd: Sched domain whose power-savings statistics are to be initialized.
3421 * @sds: Variable containing the statistics for sd.
3422 * @idle: Idle status of the CPU at which we're performing load-balancing.
3424 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3425 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3428 * Busy processors will not participate in power savings
3431 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3432 sds->power_savings_balance = 0;
3434 sds->power_savings_balance = 1;
3435 sds->min_nr_running = ULONG_MAX;
3436 sds->leader_nr_running = 0;
3441 * update_sd_power_savings_stats - Update the power saving stats for a
3442 * sched_domain while performing load balancing.
3444 * @group: sched_group belonging to the sched_domain under consideration.
3445 * @sds: Variable containing the statistics of the sched_domain
3446 * @local_group: Does group contain the CPU for which we're performing
3448 * @sgs: Variable containing the statistics of the group.
3450 static inline void update_sd_power_savings_stats(struct sched_group *group,
3451 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3454 if (!sds->power_savings_balance)
3458 * If the local group is idle or completely loaded
3459 * no need to do power savings balance at this domain
3461 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3462 !sds->this_nr_running))
3463 sds->power_savings_balance = 0;
3466 * If a group is already running at full capacity or idle,
3467 * don't include that group in power savings calculations
3469 if (!sds->power_savings_balance ||
3470 sgs->sum_nr_running >= sgs->group_capacity ||
3471 !sgs->sum_nr_running)
3475 * Calculate the group which has the least non-idle load.
3476 * This is the group from where we need to pick up the load
3479 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3480 (sgs->sum_nr_running == sds->min_nr_running &&
3481 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3482 sds->group_min = group;
3483 sds->min_nr_running = sgs->sum_nr_running;
3484 sds->min_load_per_task = sgs->sum_weighted_load /
3485 sgs->sum_nr_running;
3489 * Calculate the group which is almost near its
3490 * capacity but still has some space to pick up some load
3491 * from other group and save more power
3493 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3496 if (sgs->sum_nr_running > sds->leader_nr_running ||
3497 (sgs->sum_nr_running == sds->leader_nr_running &&
3498 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3499 sds->group_leader = group;
3500 sds->leader_nr_running = sgs->sum_nr_running;
3505 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3506 * @sds: Variable containing the statistics of the sched_domain
3507 * under consideration.
3508 * @this_cpu: Cpu at which we're currently performing load-balancing.
3509 * @imbalance: Variable to store the imbalance.
3512 * Check if we have potential to perform some power-savings balance.
3513 * If yes, set the busiest group to be the least loaded group in the
3514 * sched_domain, so that it's CPUs can be put to idle.
3516 * Returns 1 if there is potential to perform power-savings balance.
3519 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3520 int this_cpu, unsigned long *imbalance)
3522 if (!sds->power_savings_balance)
3525 if (sds->this != sds->group_leader ||
3526 sds->group_leader == sds->group_min)
3529 *imbalance = sds->min_load_per_task;
3530 sds->busiest = sds->group_min;
3535 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3536 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3537 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3542 static inline void update_sd_power_savings_stats(struct sched_group *group,
3543 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3548 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3549 int this_cpu, unsigned long *imbalance)
3553 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3555 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3557 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3558 unsigned long smt_gain = sd->smt_gain;
3565 unsigned long scale_rt_power(int cpu)
3567 struct rq *rq = cpu_rq(cpu);
3568 u64 total, available;
3570 sched_avg_update(rq);
3572 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3573 available = total - rq->rt_avg;
3575 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3576 total = SCHED_LOAD_SCALE;
3578 total >>= SCHED_LOAD_SHIFT;
3580 return div_u64(available, total);
3583 static void update_cpu_power(struct sched_domain *sd, int cpu)
3585 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3586 unsigned long power = SCHED_LOAD_SCALE;
3587 struct sched_group *sdg = sd->groups;
3589 /* here we could scale based on cpufreq */
3591 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3592 power *= arch_scale_smt_power(sd, cpu);
3593 power >>= SCHED_LOAD_SHIFT;
3596 power *= scale_rt_power(cpu);
3597 power >>= SCHED_LOAD_SHIFT;
3602 sdg->cpu_power = power;
3605 static void update_group_power(struct sched_domain *sd, int cpu)
3607 struct sched_domain *child = sd->child;
3608 struct sched_group *group, *sdg = sd->groups;
3609 unsigned long power;
3612 update_cpu_power(sd, cpu);
3618 group = child->groups;
3620 power += group->cpu_power;
3621 group = group->next;
3622 } while (group != child->groups);
3624 sdg->cpu_power = power;
3628 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3629 * @group: sched_group whose statistics are to be updated.
3630 * @this_cpu: Cpu for which load balance is currently performed.
3631 * @idle: Idle status of this_cpu
3632 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3633 * @sd_idle: Idle status of the sched_domain containing group.
3634 * @local_group: Does group contain this_cpu.
3635 * @cpus: Set of cpus considered for load balancing.
3636 * @balance: Should we balance.
3637 * @sgs: variable to hold the statistics for this group.
3639 static inline void update_sg_lb_stats(struct sched_domain *sd,
3640 struct sched_group *group, int this_cpu,
3641 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3642 int local_group, const struct cpumask *cpus,
3643 int *balance, struct sg_lb_stats *sgs)
3645 unsigned long load, max_cpu_load, min_cpu_load;
3647 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3648 unsigned long sum_avg_load_per_task;
3649 unsigned long avg_load_per_task;
3652 balance_cpu = group_first_cpu(group);
3653 if (balance_cpu == this_cpu)
3654 update_group_power(sd, this_cpu);
3657 /* Tally up the load of all CPUs in the group */
3658 sum_avg_load_per_task = avg_load_per_task = 0;
3660 min_cpu_load = ~0UL;
3662 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3663 struct rq *rq = cpu_rq(i);
3665 if (*sd_idle && rq->nr_running)
3668 /* Bias balancing toward cpus of our domain */
3670 if (idle_cpu(i) && !first_idle_cpu) {
3675 load = target_load(i, load_idx);
3677 load = source_load(i, load_idx);
3678 if (load > max_cpu_load)
3679 max_cpu_load = load;
3680 if (min_cpu_load > load)
3681 min_cpu_load = load;
3684 sgs->group_load += load;
3685 sgs->sum_nr_running += rq->nr_running;
3686 sgs->sum_weighted_load += weighted_cpuload(i);
3688 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3692 * First idle cpu or the first cpu(busiest) in this sched group
3693 * is eligible for doing load balancing at this and above
3694 * domains. In the newly idle case, we will allow all the cpu's
3695 * to do the newly idle load balance.
3697 if (idle != CPU_NEWLY_IDLE && local_group &&
3698 balance_cpu != this_cpu && balance) {
3703 /* Adjust by relative CPU power of the group */
3704 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3708 * Consider the group unbalanced when the imbalance is larger
3709 * than the average weight of two tasks.
3711 * APZ: with cgroup the avg task weight can vary wildly and
3712 * might not be a suitable number - should we keep a
3713 * normalized nr_running number somewhere that negates
3716 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3719 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3722 sgs->group_capacity =
3723 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3727 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3728 * @sd: sched_domain whose statistics are to be updated.
3729 * @this_cpu: Cpu for which load balance is currently performed.
3730 * @idle: Idle status of this_cpu
3731 * @sd_idle: Idle status of the sched_domain containing group.
3732 * @cpus: Set of cpus considered for load balancing.
3733 * @balance: Should we balance.
3734 * @sds: variable to hold the statistics for this sched_domain.
3736 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3737 enum cpu_idle_type idle, int *sd_idle,
3738 const struct cpumask *cpus, int *balance,
3739 struct sd_lb_stats *sds)
3741 struct sched_domain *child = sd->child;
3742 struct sched_group *group = sd->groups;
3743 struct sg_lb_stats sgs;
3744 int load_idx, prefer_sibling = 0;
3746 if (child && child->flags & SD_PREFER_SIBLING)
3749 init_sd_power_savings_stats(sd, sds, idle);
3750 load_idx = get_sd_load_idx(sd, idle);
3755 local_group = cpumask_test_cpu(this_cpu,
3756 sched_group_cpus(group));
3757 memset(&sgs, 0, sizeof(sgs));
3758 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3759 local_group, cpus, balance, &sgs);
3761 if (local_group && balance && !(*balance))
3764 sds->total_load += sgs.group_load;
3765 sds->total_pwr += group->cpu_power;
3768 * In case the child domain prefers tasks go to siblings
3769 * first, lower the group capacity to one so that we'll try
3770 * and move all the excess tasks away.
3773 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3776 sds->this_load = sgs.avg_load;
3778 sds->this_nr_running = sgs.sum_nr_running;
3779 sds->this_load_per_task = sgs.sum_weighted_load;
3780 } else if (sgs.avg_load > sds->max_load &&
3781 (sgs.sum_nr_running > sgs.group_capacity ||
3783 sds->max_load = sgs.avg_load;
3784 sds->busiest = group;
3785 sds->busiest_nr_running = sgs.sum_nr_running;
3786 sds->busiest_load_per_task = sgs.sum_weighted_load;
3787 sds->group_imb = sgs.group_imb;
3790 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3791 group = group->next;
3792 } while (group != sd->groups);
3796 * fix_small_imbalance - Calculate the minor imbalance that exists
3797 * amongst the groups of a sched_domain, during
3799 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3800 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3801 * @imbalance: Variable to store the imbalance.
3803 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3804 int this_cpu, unsigned long *imbalance)
3806 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3807 unsigned int imbn = 2;
3809 if (sds->this_nr_running) {
3810 sds->this_load_per_task /= sds->this_nr_running;
3811 if (sds->busiest_load_per_task >
3812 sds->this_load_per_task)
3815 sds->this_load_per_task =
3816 cpu_avg_load_per_task(this_cpu);
3818 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3819 sds->busiest_load_per_task * imbn) {
3820 *imbalance = sds->busiest_load_per_task;
3825 * OK, we don't have enough imbalance to justify moving tasks,
3826 * however we may be able to increase total CPU power used by
3830 pwr_now += sds->busiest->cpu_power *
3831 min(sds->busiest_load_per_task, sds->max_load);
3832 pwr_now += sds->this->cpu_power *
3833 min(sds->this_load_per_task, sds->this_load);
3834 pwr_now /= SCHED_LOAD_SCALE;
3836 /* Amount of load we'd subtract */
3837 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3838 sds->busiest->cpu_power;
3839 if (sds->max_load > tmp)
3840 pwr_move += sds->busiest->cpu_power *
3841 min(sds->busiest_load_per_task, sds->max_load - tmp);
3843 /* Amount of load we'd add */
3844 if (sds->max_load * sds->busiest->cpu_power <
3845 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3846 tmp = (sds->max_load * sds->busiest->cpu_power) /
3847 sds->this->cpu_power;
3849 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3850 sds->this->cpu_power;
3851 pwr_move += sds->this->cpu_power *
3852 min(sds->this_load_per_task, sds->this_load + tmp);
3853 pwr_move /= SCHED_LOAD_SCALE;
3855 /* Move if we gain throughput */
3856 if (pwr_move > pwr_now)
3857 *imbalance = sds->busiest_load_per_task;
3861 * calculate_imbalance - Calculate the amount of imbalance present within the
3862 * groups of a given sched_domain during load balance.
3863 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3864 * @this_cpu: Cpu for which currently load balance is being performed.
3865 * @imbalance: The variable to store the imbalance.
3867 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3868 unsigned long *imbalance)
3870 unsigned long max_pull;
3872 * In the presence of smp nice balancing, certain scenarios can have
3873 * max load less than avg load(as we skip the groups at or below
3874 * its cpu_power, while calculating max_load..)
3876 if (sds->max_load < sds->avg_load) {
3878 return fix_small_imbalance(sds, this_cpu, imbalance);
3881 /* Don't want to pull so many tasks that a group would go idle */
3882 max_pull = min(sds->max_load - sds->avg_load,
3883 sds->max_load - sds->busiest_load_per_task);
3885 /* How much load to actually move to equalise the imbalance */
3886 *imbalance = min(max_pull * sds->busiest->cpu_power,
3887 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3891 * if *imbalance is less than the average load per runnable task
3892 * there is no gaurantee that any tasks will be moved so we'll have
3893 * a think about bumping its value to force at least one task to be
3896 if (*imbalance < sds->busiest_load_per_task)
3897 return fix_small_imbalance(sds, this_cpu, imbalance);
3900 /******* find_busiest_group() helpers end here *********************/
3903 * find_busiest_group - Returns the busiest group within the sched_domain
3904 * if there is an imbalance. If there isn't an imbalance, and
3905 * the user has opted for power-savings, it returns a group whose
3906 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3907 * such a group exists.
3909 * Also calculates the amount of weighted load which should be moved
3910 * to restore balance.
3912 * @sd: The sched_domain whose busiest group is to be returned.
3913 * @this_cpu: The cpu for which load balancing is currently being performed.
3914 * @imbalance: Variable which stores amount of weighted load which should
3915 * be moved to restore balance/put a group to idle.
3916 * @idle: The idle status of this_cpu.
3917 * @sd_idle: The idleness of sd
3918 * @cpus: The set of CPUs under consideration for load-balancing.
3919 * @balance: Pointer to a variable indicating if this_cpu
3920 * is the appropriate cpu to perform load balancing at this_level.
3922 * Returns: - the busiest group if imbalance exists.
3923 * - If no imbalance and user has opted for power-savings balance,
3924 * return the least loaded group whose CPUs can be
3925 * put to idle by rebalancing its tasks onto our group.
3927 static struct sched_group *
3928 find_busiest_group(struct sched_domain *sd, int this_cpu,
3929 unsigned long *imbalance, enum cpu_idle_type idle,
3930 int *sd_idle, const struct cpumask *cpus, int *balance)
3932 struct sd_lb_stats sds;
3934 memset(&sds, 0, sizeof(sds));
3937 * Compute the various statistics relavent for load balancing at
3940 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3943 /* Cases where imbalance does not exist from POV of this_cpu */
3944 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3946 * 2) There is no busy sibling group to pull from.
3947 * 3) This group is the busiest group.
3948 * 4) This group is more busy than the avg busieness at this
3950 * 5) The imbalance is within the specified limit.
3951 * 6) Any rebalance would lead to ping-pong
3953 if (balance && !(*balance))
3956 if (!sds.busiest || sds.busiest_nr_running == 0)
3959 if (sds.this_load >= sds.max_load)
3962 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3964 if (sds.this_load >= sds.avg_load)
3967 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3970 sds.busiest_load_per_task /= sds.busiest_nr_running;
3972 sds.busiest_load_per_task =
3973 min(sds.busiest_load_per_task, sds.avg_load);
3976 * We're trying to get all the cpus to the average_load, so we don't
3977 * want to push ourselves above the average load, nor do we wish to
3978 * reduce the max loaded cpu below the average load, as either of these
3979 * actions would just result in more rebalancing later, and ping-pong
3980 * tasks around. Thus we look for the minimum possible imbalance.
3981 * Negative imbalances (*we* are more loaded than anyone else) will
3982 * be counted as no imbalance for these purposes -- we can't fix that
3983 * by pulling tasks to us. Be careful of negative numbers as they'll
3984 * appear as very large values with unsigned longs.
3986 if (sds.max_load <= sds.busiest_load_per_task)
3989 /* Looks like there is an imbalance. Compute it */
3990 calculate_imbalance(&sds, this_cpu, imbalance);
3995 * There is no obvious imbalance. But check if we can do some balancing
3998 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4006 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4009 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4010 unsigned long imbalance, const struct cpumask *cpus)
4012 struct rq *busiest = NULL, *rq;
4013 unsigned long max_load = 0;
4016 for_each_cpu(i, sched_group_cpus(group)) {
4017 unsigned long power = power_of(i);
4018 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4021 if (!cpumask_test_cpu(i, cpus))
4025 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4028 if (capacity && rq->nr_running == 1 && wl > imbalance)
4031 if (wl > max_load) {
4041 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4042 * so long as it is large enough.
4044 #define MAX_PINNED_INTERVAL 512
4046 /* Working cpumask for load_balance and load_balance_newidle. */
4047 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4050 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4051 * tasks if there is an imbalance.
4053 static int load_balance(int this_cpu, struct rq *this_rq,
4054 struct sched_domain *sd, enum cpu_idle_type idle,
4057 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4058 struct sched_group *group;
4059 unsigned long imbalance;
4061 unsigned long flags;
4062 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4064 cpumask_setall(cpus);
4067 * When power savings policy is enabled for the parent domain, idle
4068 * sibling can pick up load irrespective of busy siblings. In this case,
4069 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4070 * portraying it as CPU_NOT_IDLE.
4072 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4073 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4076 schedstat_inc(sd, lb_count[idle]);
4080 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4087 schedstat_inc(sd, lb_nobusyg[idle]);
4091 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4093 schedstat_inc(sd, lb_nobusyq[idle]);
4097 BUG_ON(busiest == this_rq);
4099 schedstat_add(sd, lb_imbalance[idle], imbalance);
4102 if (busiest->nr_running > 1) {
4104 * Attempt to move tasks. If find_busiest_group has found
4105 * an imbalance but busiest->nr_running <= 1, the group is
4106 * still unbalanced. ld_moved simply stays zero, so it is
4107 * correctly treated as an imbalance.
4109 local_irq_save(flags);
4110 double_rq_lock(this_rq, busiest);
4111 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4112 imbalance, sd, idle, &all_pinned);
4113 double_rq_unlock(this_rq, busiest);
4114 local_irq_restore(flags);
4117 * some other cpu did the load balance for us.
4119 if (ld_moved && this_cpu != smp_processor_id())
4120 resched_cpu(this_cpu);
4122 /* All tasks on this runqueue were pinned by CPU affinity */
4123 if (unlikely(all_pinned)) {
4124 cpumask_clear_cpu(cpu_of(busiest), cpus);
4125 if (!cpumask_empty(cpus))
4132 schedstat_inc(sd, lb_failed[idle]);
4133 sd->nr_balance_failed++;
4135 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4137 spin_lock_irqsave(&busiest->lock, flags);
4139 /* don't kick the migration_thread, if the curr
4140 * task on busiest cpu can't be moved to this_cpu
4142 if (!cpumask_test_cpu(this_cpu,
4143 &busiest->curr->cpus_allowed)) {
4144 spin_unlock_irqrestore(&busiest->lock, flags);
4146 goto out_one_pinned;
4149 if (!busiest->active_balance) {
4150 busiest->active_balance = 1;
4151 busiest->push_cpu = this_cpu;
4154 spin_unlock_irqrestore(&busiest->lock, flags);
4156 wake_up_process(busiest->migration_thread);
4159 * We've kicked active balancing, reset the failure
4162 sd->nr_balance_failed = sd->cache_nice_tries+1;
4165 sd->nr_balance_failed = 0;
4167 if (likely(!active_balance)) {
4168 /* We were unbalanced, so reset the balancing interval */
4169 sd->balance_interval = sd->min_interval;
4172 * If we've begun active balancing, start to back off. This
4173 * case may not be covered by the all_pinned logic if there
4174 * is only 1 task on the busy runqueue (because we don't call
4177 if (sd->balance_interval < sd->max_interval)
4178 sd->balance_interval *= 2;
4181 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4182 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4188 schedstat_inc(sd, lb_balanced[idle]);
4190 sd->nr_balance_failed = 0;
4193 /* tune up the balancing interval */
4194 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4195 (sd->balance_interval < sd->max_interval))
4196 sd->balance_interval *= 2;
4198 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4199 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4210 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4211 * tasks if there is an imbalance.
4213 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4214 * this_rq is locked.
4217 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4219 struct sched_group *group;
4220 struct rq *busiest = NULL;
4221 unsigned long imbalance;
4225 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4227 cpumask_setall(cpus);
4230 * When power savings policy is enabled for the parent domain, idle
4231 * sibling can pick up load irrespective of busy siblings. In this case,
4232 * let the state of idle sibling percolate up as IDLE, instead of
4233 * portraying it as CPU_NOT_IDLE.
4235 if (sd->flags & SD_SHARE_CPUPOWER &&
4236 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4239 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4241 update_shares_locked(this_rq, sd);
4242 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4243 &sd_idle, cpus, NULL);
4245 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4249 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4251 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4255 BUG_ON(busiest == this_rq);
4257 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4260 if (busiest->nr_running > 1) {
4261 /* Attempt to move tasks */
4262 double_lock_balance(this_rq, busiest);
4263 /* this_rq->clock is already updated */
4264 update_rq_clock(busiest);
4265 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4266 imbalance, sd, CPU_NEWLY_IDLE,
4268 double_unlock_balance(this_rq, busiest);
4270 if (unlikely(all_pinned)) {
4271 cpumask_clear_cpu(cpu_of(busiest), cpus);
4272 if (!cpumask_empty(cpus))
4278 int active_balance = 0;
4280 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4281 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4282 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4285 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4288 if (sd->nr_balance_failed++ < 2)
4292 * The only task running in a non-idle cpu can be moved to this
4293 * cpu in an attempt to completely freeup the other CPU
4294 * package. The same method used to move task in load_balance()
4295 * have been extended for load_balance_newidle() to speedup
4296 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4298 * The package power saving logic comes from
4299 * find_busiest_group(). If there are no imbalance, then
4300 * f_b_g() will return NULL. However when sched_mc={1,2} then
4301 * f_b_g() will select a group from which a running task may be
4302 * pulled to this cpu in order to make the other package idle.
4303 * If there is no opportunity to make a package idle and if
4304 * there are no imbalance, then f_b_g() will return NULL and no
4305 * action will be taken in load_balance_newidle().
4307 * Under normal task pull operation due to imbalance, there
4308 * will be more than one task in the source run queue and
4309 * move_tasks() will succeed. ld_moved will be true and this
4310 * active balance code will not be triggered.
4313 /* Lock busiest in correct order while this_rq is held */
4314 double_lock_balance(this_rq, busiest);
4317 * don't kick the migration_thread, if the curr
4318 * task on busiest cpu can't be moved to this_cpu
4320 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4321 double_unlock_balance(this_rq, busiest);
4326 if (!busiest->active_balance) {
4327 busiest->active_balance = 1;
4328 busiest->push_cpu = this_cpu;
4332 double_unlock_balance(this_rq, busiest);
4334 * Should not call ttwu while holding a rq->lock
4336 spin_unlock(&this_rq->lock);
4338 wake_up_process(busiest->migration_thread);
4339 spin_lock(&this_rq->lock);
4342 sd->nr_balance_failed = 0;
4344 update_shares_locked(this_rq, sd);
4348 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4349 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4350 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4352 sd->nr_balance_failed = 0;
4358 * idle_balance is called by schedule() if this_cpu is about to become
4359 * idle. Attempts to pull tasks from other CPUs.
4361 static void idle_balance(int this_cpu, struct rq *this_rq)
4363 struct sched_domain *sd;
4364 int pulled_task = 0;
4365 unsigned long next_balance = jiffies + HZ;
4367 for_each_domain(this_cpu, sd) {
4368 unsigned long interval;
4370 if (!(sd->flags & SD_LOAD_BALANCE))
4373 if (sd->flags & SD_BALANCE_NEWIDLE)
4374 /* If we've pulled tasks over stop searching: */
4375 pulled_task = load_balance_newidle(this_cpu, this_rq,
4378 interval = msecs_to_jiffies(sd->balance_interval);
4379 if (time_after(next_balance, sd->last_balance + interval))
4380 next_balance = sd->last_balance + interval;
4384 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4386 * We are going idle. next_balance may be set based on
4387 * a busy processor. So reset next_balance.
4389 this_rq->next_balance = next_balance;
4394 * active_load_balance is run by migration threads. It pushes running tasks
4395 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4396 * running on each physical CPU where possible, and avoids physical /
4397 * logical imbalances.
4399 * Called with busiest_rq locked.
4401 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4403 int target_cpu = busiest_rq->push_cpu;
4404 struct sched_domain *sd;
4405 struct rq *target_rq;
4407 /* Is there any task to move? */
4408 if (busiest_rq->nr_running <= 1)
4411 target_rq = cpu_rq(target_cpu);
4414 * This condition is "impossible", if it occurs
4415 * we need to fix it. Originally reported by
4416 * Bjorn Helgaas on a 128-cpu setup.
4418 BUG_ON(busiest_rq == target_rq);
4420 /* move a task from busiest_rq to target_rq */
4421 double_lock_balance(busiest_rq, target_rq);
4422 update_rq_clock(busiest_rq);
4423 update_rq_clock(target_rq);
4425 /* Search for an sd spanning us and the target CPU. */
4426 for_each_domain(target_cpu, sd) {
4427 if ((sd->flags & SD_LOAD_BALANCE) &&
4428 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4433 schedstat_inc(sd, alb_count);
4435 if (move_one_task(target_rq, target_cpu, busiest_rq,
4437 schedstat_inc(sd, alb_pushed);
4439 schedstat_inc(sd, alb_failed);
4441 double_unlock_balance(busiest_rq, target_rq);
4446 atomic_t load_balancer;
4447 cpumask_var_t cpu_mask;
4448 cpumask_var_t ilb_grp_nohz_mask;
4449 } nohz ____cacheline_aligned = {
4450 .load_balancer = ATOMIC_INIT(-1),
4453 int get_nohz_load_balancer(void)
4455 return atomic_read(&nohz.load_balancer);
4458 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4460 * lowest_flag_domain - Return lowest sched_domain containing flag.
4461 * @cpu: The cpu whose lowest level of sched domain is to
4463 * @flag: The flag to check for the lowest sched_domain
4464 * for the given cpu.
4466 * Returns the lowest sched_domain of a cpu which contains the given flag.
4468 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4470 struct sched_domain *sd;
4472 for_each_domain(cpu, sd)
4473 if (sd && (sd->flags & flag))
4480 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4481 * @cpu: The cpu whose domains we're iterating over.
4482 * @sd: variable holding the value of the power_savings_sd
4484 * @flag: The flag to filter the sched_domains to be iterated.
4486 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4487 * set, starting from the lowest sched_domain to the highest.
4489 #define for_each_flag_domain(cpu, sd, flag) \
4490 for (sd = lowest_flag_domain(cpu, flag); \
4491 (sd && (sd->flags & flag)); sd = sd->parent)
4494 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4495 * @ilb_group: group to be checked for semi-idleness
4497 * Returns: 1 if the group is semi-idle. 0 otherwise.
4499 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4500 * and atleast one non-idle CPU. This helper function checks if the given
4501 * sched_group is semi-idle or not.
4503 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4505 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4506 sched_group_cpus(ilb_group));
4509 * A sched_group is semi-idle when it has atleast one busy cpu
4510 * and atleast one idle cpu.
4512 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4515 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4521 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4522 * @cpu: The cpu which is nominating a new idle_load_balancer.
4524 * Returns: Returns the id of the idle load balancer if it exists,
4525 * Else, returns >= nr_cpu_ids.
4527 * This algorithm picks the idle load balancer such that it belongs to a
4528 * semi-idle powersavings sched_domain. The idea is to try and avoid
4529 * completely idle packages/cores just for the purpose of idle load balancing
4530 * when there are other idle cpu's which are better suited for that job.
4532 static int find_new_ilb(int cpu)
4534 struct sched_domain *sd;
4535 struct sched_group *ilb_group;
4538 * Have idle load balancer selection from semi-idle packages only
4539 * when power-aware load balancing is enabled
4541 if (!(sched_smt_power_savings || sched_mc_power_savings))
4545 * Optimize for the case when we have no idle CPUs or only one
4546 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4548 if (cpumask_weight(nohz.cpu_mask) < 2)
4551 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4552 ilb_group = sd->groups;
4555 if (is_semi_idle_group(ilb_group))
4556 return cpumask_first(nohz.ilb_grp_nohz_mask);
4558 ilb_group = ilb_group->next;
4560 } while (ilb_group != sd->groups);
4564 return cpumask_first(nohz.cpu_mask);
4566 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4567 static inline int find_new_ilb(int call_cpu)
4569 return cpumask_first(nohz.cpu_mask);
4574 * This routine will try to nominate the ilb (idle load balancing)
4575 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4576 * load balancing on behalf of all those cpus. If all the cpus in the system
4577 * go into this tickless mode, then there will be no ilb owner (as there is
4578 * no need for one) and all the cpus will sleep till the next wakeup event
4581 * For the ilb owner, tick is not stopped. And this tick will be used
4582 * for idle load balancing. ilb owner will still be part of
4585 * While stopping the tick, this cpu will become the ilb owner if there
4586 * is no other owner. And will be the owner till that cpu becomes busy
4587 * or if all cpus in the system stop their ticks at which point
4588 * there is no need for ilb owner.
4590 * When the ilb owner becomes busy, it nominates another owner, during the
4591 * next busy scheduler_tick()
4593 int select_nohz_load_balancer(int stop_tick)
4595 int cpu = smp_processor_id();
4598 cpu_rq(cpu)->in_nohz_recently = 1;
4600 if (!cpu_active(cpu)) {
4601 if (atomic_read(&nohz.load_balancer) != cpu)
4605 * If we are going offline and still the leader,
4608 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4614 cpumask_set_cpu(cpu, nohz.cpu_mask);
4616 /* time for ilb owner also to sleep */
4617 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4618 if (atomic_read(&nohz.load_balancer) == cpu)
4619 atomic_set(&nohz.load_balancer, -1);
4623 if (atomic_read(&nohz.load_balancer) == -1) {
4624 /* make me the ilb owner */
4625 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4627 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4630 if (!(sched_smt_power_savings ||
4631 sched_mc_power_savings))
4634 * Check to see if there is a more power-efficient
4637 new_ilb = find_new_ilb(cpu);
4638 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4639 atomic_set(&nohz.load_balancer, -1);
4640 resched_cpu(new_ilb);
4646 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4649 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4651 if (atomic_read(&nohz.load_balancer) == cpu)
4652 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4659 static DEFINE_SPINLOCK(balancing);
4662 * It checks each scheduling domain to see if it is due to be balanced,
4663 * and initiates a balancing operation if so.
4665 * Balancing parameters are set up in arch_init_sched_domains.
4667 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4670 struct rq *rq = cpu_rq(cpu);
4671 unsigned long interval;
4672 struct sched_domain *sd;
4673 /* Earliest time when we have to do rebalance again */
4674 unsigned long next_balance = jiffies + 60*HZ;
4675 int update_next_balance = 0;
4678 for_each_domain(cpu, sd) {
4679 if (!(sd->flags & SD_LOAD_BALANCE))
4682 interval = sd->balance_interval;
4683 if (idle != CPU_IDLE)
4684 interval *= sd->busy_factor;
4686 /* scale ms to jiffies */
4687 interval = msecs_to_jiffies(interval);
4688 if (unlikely(!interval))
4690 if (interval > HZ*NR_CPUS/10)
4691 interval = HZ*NR_CPUS/10;
4693 need_serialize = sd->flags & SD_SERIALIZE;
4695 if (need_serialize) {
4696 if (!spin_trylock(&balancing))
4700 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4701 if (load_balance(cpu, rq, sd, idle, &balance)) {
4703 * We've pulled tasks over so either we're no
4704 * longer idle, or one of our SMT siblings is
4707 idle = CPU_NOT_IDLE;
4709 sd->last_balance = jiffies;
4712 spin_unlock(&balancing);
4714 if (time_after(next_balance, sd->last_balance + interval)) {
4715 next_balance = sd->last_balance + interval;
4716 update_next_balance = 1;
4720 * Stop the load balance at this level. There is another
4721 * CPU in our sched group which is doing load balancing more
4729 * next_balance will be updated only when there is a need.
4730 * When the cpu is attached to null domain for ex, it will not be
4733 if (likely(update_next_balance))
4734 rq->next_balance = next_balance;
4738 * run_rebalance_domains is triggered when needed from the scheduler tick.
4739 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4740 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4742 static void run_rebalance_domains(struct softirq_action *h)
4744 int this_cpu = smp_processor_id();
4745 struct rq *this_rq = cpu_rq(this_cpu);
4746 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4747 CPU_IDLE : CPU_NOT_IDLE;
4749 rebalance_domains(this_cpu, idle);
4753 * If this cpu is the owner for idle load balancing, then do the
4754 * balancing on behalf of the other idle cpus whose ticks are
4757 if (this_rq->idle_at_tick &&
4758 atomic_read(&nohz.load_balancer) == this_cpu) {
4762 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4763 if (balance_cpu == this_cpu)
4767 * If this cpu gets work to do, stop the load balancing
4768 * work being done for other cpus. Next load
4769 * balancing owner will pick it up.
4774 rebalance_domains(balance_cpu, CPU_IDLE);
4776 rq = cpu_rq(balance_cpu);
4777 if (time_after(this_rq->next_balance, rq->next_balance))
4778 this_rq->next_balance = rq->next_balance;
4784 static inline int on_null_domain(int cpu)
4786 return !rcu_dereference(cpu_rq(cpu)->sd);
4790 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4792 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4793 * idle load balancing owner or decide to stop the periodic load balancing,
4794 * if the whole system is idle.
4796 static inline void trigger_load_balance(struct rq *rq, int cpu)
4800 * If we were in the nohz mode recently and busy at the current
4801 * scheduler tick, then check if we need to nominate new idle
4804 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4805 rq->in_nohz_recently = 0;
4807 if (atomic_read(&nohz.load_balancer) == cpu) {
4808 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4809 atomic_set(&nohz.load_balancer, -1);
4812 if (atomic_read(&nohz.load_balancer) == -1) {
4813 int ilb = find_new_ilb(cpu);
4815 if (ilb < nr_cpu_ids)
4821 * If this cpu is idle and doing idle load balancing for all the
4822 * cpus with ticks stopped, is it time for that to stop?
4824 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4825 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4831 * If this cpu is idle and the idle load balancing is done by
4832 * someone else, then no need raise the SCHED_SOFTIRQ
4834 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4835 cpumask_test_cpu(cpu, nohz.cpu_mask))
4838 /* Don't need to rebalance while attached to NULL domain */
4839 if (time_after_eq(jiffies, rq->next_balance) &&
4840 likely(!on_null_domain(cpu)))
4841 raise_softirq(SCHED_SOFTIRQ);
4844 #else /* CONFIG_SMP */
4847 * on UP we do not need to balance between CPUs:
4849 static inline void idle_balance(int cpu, struct rq *rq)
4855 DEFINE_PER_CPU(struct kernel_stat, kstat);
4857 EXPORT_PER_CPU_SYMBOL(kstat);
4860 * Return any ns on the sched_clock that have not yet been accounted in
4861 * @p in case that task is currently running.
4863 * Called with task_rq_lock() held on @rq.
4865 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4869 if (task_current(rq, p)) {
4870 update_rq_clock(rq);
4871 ns = rq->clock - p->se.exec_start;
4879 unsigned long long task_delta_exec(struct task_struct *p)
4881 unsigned long flags;
4885 rq = task_rq_lock(p, &flags);
4886 ns = do_task_delta_exec(p, rq);
4887 task_rq_unlock(rq, &flags);
4893 * Return accounted runtime for the task.
4894 * In case the task is currently running, return the runtime plus current's
4895 * pending runtime that have not been accounted yet.
4897 unsigned long long task_sched_runtime(struct task_struct *p)
4899 unsigned long flags;
4903 rq = task_rq_lock(p, &flags);
4904 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4905 task_rq_unlock(rq, &flags);
4911 * Return sum_exec_runtime for the thread group.
4912 * In case the task is currently running, return the sum plus current's
4913 * pending runtime that have not been accounted yet.
4915 * Note that the thread group might have other running tasks as well,
4916 * so the return value not includes other pending runtime that other
4917 * running tasks might have.
4919 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4921 struct task_cputime totals;
4922 unsigned long flags;
4926 rq = task_rq_lock(p, &flags);
4927 thread_group_cputime(p, &totals);
4928 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4929 task_rq_unlock(rq, &flags);
4935 * Account user cpu time to a process.
4936 * @p: the process that the cpu time gets accounted to
4937 * @cputime: the cpu time spent in user space since the last update
4938 * @cputime_scaled: cputime scaled by cpu frequency
4940 void account_user_time(struct task_struct *p, cputime_t cputime,
4941 cputime_t cputime_scaled)
4943 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4946 /* Add user time to process. */
4947 p->utime = cputime_add(p->utime, cputime);
4948 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4949 account_group_user_time(p, cputime);
4951 /* Add user time to cpustat. */
4952 tmp = cputime_to_cputime64(cputime);
4953 if (TASK_NICE(p) > 0)
4954 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4956 cpustat->user = cputime64_add(cpustat->user, tmp);
4958 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4959 /* Account for user time used */
4960 acct_update_integrals(p);
4964 * Account guest cpu time to a process.
4965 * @p: the process that the cpu time gets accounted to
4966 * @cputime: the cpu time spent in virtual machine since the last update
4967 * @cputime_scaled: cputime scaled by cpu frequency
4969 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4970 cputime_t cputime_scaled)
4973 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4975 tmp = cputime_to_cputime64(cputime);
4977 /* Add guest time to process. */
4978 p->utime = cputime_add(p->utime, cputime);
4979 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4980 account_group_user_time(p, cputime);
4981 p->gtime = cputime_add(p->gtime, cputime);
4983 /* Add guest time to cpustat. */
4984 cpustat->user = cputime64_add(cpustat->user, tmp);
4985 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4989 * Account system cpu time to a process.
4990 * @p: the process that the cpu time gets accounted to
4991 * @hardirq_offset: the offset to subtract from hardirq_count()
4992 * @cputime: the cpu time spent in kernel space since the last update
4993 * @cputime_scaled: cputime scaled by cpu frequency
4995 void account_system_time(struct task_struct *p, int hardirq_offset,
4996 cputime_t cputime, cputime_t cputime_scaled)
4998 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5001 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5002 account_guest_time(p, cputime, cputime_scaled);
5006 /* Add system time to process. */
5007 p->stime = cputime_add(p->stime, cputime);
5008 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5009 account_group_system_time(p, cputime);
5011 /* Add system time to cpustat. */
5012 tmp = cputime_to_cputime64(cputime);
5013 if (hardirq_count() - hardirq_offset)
5014 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5015 else if (softirq_count())
5016 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5018 cpustat->system = cputime64_add(cpustat->system, tmp);
5020 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5022 /* Account for system time used */
5023 acct_update_integrals(p);
5027 * Account for involuntary wait time.
5028 * @steal: the cpu time spent in involuntary wait
5030 void account_steal_time(cputime_t cputime)
5032 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5033 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5035 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5039 * Account for idle time.
5040 * @cputime: the cpu time spent in idle wait
5042 void account_idle_time(cputime_t cputime)
5044 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5045 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5046 struct rq *rq = this_rq();
5048 if (atomic_read(&rq->nr_iowait) > 0)
5049 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5051 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5054 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5057 * Account a single tick of cpu time.
5058 * @p: the process that the cpu time gets accounted to
5059 * @user_tick: indicates if the tick is a user or a system tick
5061 void account_process_tick(struct task_struct *p, int user_tick)
5063 cputime_t one_jiffy = jiffies_to_cputime(1);
5064 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
5065 struct rq *rq = this_rq();
5068 account_user_time(p, one_jiffy, one_jiffy_scaled);
5069 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5070 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
5073 account_idle_time(one_jiffy);
5077 * Account multiple ticks of steal time.
5078 * @p: the process from which the cpu time has been stolen
5079 * @ticks: number of stolen ticks
5081 void account_steal_ticks(unsigned long ticks)
5083 account_steal_time(jiffies_to_cputime(ticks));
5087 * Account multiple ticks of idle time.
5088 * @ticks: number of stolen ticks
5090 void account_idle_ticks(unsigned long ticks)
5092 account_idle_time(jiffies_to_cputime(ticks));
5098 * Use precise platform statistics if available:
5100 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5101 cputime_t task_utime(struct task_struct *p)
5106 cputime_t task_stime(struct task_struct *p)
5111 cputime_t task_utime(struct task_struct *p)
5113 clock_t utime = cputime_to_clock_t(p->utime),
5114 total = utime + cputime_to_clock_t(p->stime);
5118 * Use CFS's precise accounting:
5120 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5124 do_div(temp, total);
5126 utime = (clock_t)temp;
5128 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5129 return p->prev_utime;
5132 cputime_t task_stime(struct task_struct *p)
5137 * Use CFS's precise accounting. (we subtract utime from
5138 * the total, to make sure the total observed by userspace
5139 * grows monotonically - apps rely on that):
5141 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5142 cputime_to_clock_t(task_utime(p));
5145 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5147 return p->prev_stime;
5151 inline cputime_t task_gtime(struct task_struct *p)
5157 * This function gets called by the timer code, with HZ frequency.
5158 * We call it with interrupts disabled.
5160 * It also gets called by the fork code, when changing the parent's
5163 void scheduler_tick(void)
5165 int cpu = smp_processor_id();
5166 struct rq *rq = cpu_rq(cpu);
5167 struct task_struct *curr = rq->curr;
5171 spin_lock(&rq->lock);
5172 update_rq_clock(rq);
5173 update_cpu_load(rq);
5174 curr->sched_class->task_tick(rq, curr, 0);
5175 spin_unlock(&rq->lock);
5177 perf_counter_task_tick(curr, cpu);
5180 rq->idle_at_tick = idle_cpu(cpu);
5181 trigger_load_balance(rq, cpu);
5185 notrace unsigned long get_parent_ip(unsigned long addr)
5187 if (in_lock_functions(addr)) {
5188 addr = CALLER_ADDR2;
5189 if (in_lock_functions(addr))
5190 addr = CALLER_ADDR3;
5195 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5196 defined(CONFIG_PREEMPT_TRACER))
5198 void __kprobes add_preempt_count(int val)
5200 #ifdef CONFIG_DEBUG_PREEMPT
5204 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5207 preempt_count() += val;
5208 #ifdef CONFIG_DEBUG_PREEMPT
5210 * Spinlock count overflowing soon?
5212 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5215 if (preempt_count() == val)
5216 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5218 EXPORT_SYMBOL(add_preempt_count);
5220 void __kprobes sub_preempt_count(int val)
5222 #ifdef CONFIG_DEBUG_PREEMPT
5226 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5229 * Is the spinlock portion underflowing?
5231 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5232 !(preempt_count() & PREEMPT_MASK)))
5236 if (preempt_count() == val)
5237 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5238 preempt_count() -= val;
5240 EXPORT_SYMBOL(sub_preempt_count);
5245 * Print scheduling while atomic bug:
5247 static noinline void __schedule_bug(struct task_struct *prev)
5249 struct pt_regs *regs = get_irq_regs();
5251 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5252 prev->comm, prev->pid, preempt_count());
5254 debug_show_held_locks(prev);
5256 if (irqs_disabled())
5257 print_irqtrace_events(prev);
5266 * Various schedule()-time debugging checks and statistics:
5268 static inline void schedule_debug(struct task_struct *prev)
5271 * Test if we are atomic. Since do_exit() needs to call into
5272 * schedule() atomically, we ignore that path for now.
5273 * Otherwise, whine if we are scheduling when we should not be.
5275 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5276 __schedule_bug(prev);
5278 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5280 schedstat_inc(this_rq(), sched_count);
5281 #ifdef CONFIG_SCHEDSTATS
5282 if (unlikely(prev->lock_depth >= 0)) {
5283 schedstat_inc(this_rq(), bkl_count);
5284 schedstat_inc(prev, sched_info.bkl_count);
5289 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5291 if (prev->state == TASK_RUNNING) {
5292 u64 runtime = prev->se.sum_exec_runtime;
5294 runtime -= prev->se.prev_sum_exec_runtime;
5295 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5298 * In order to avoid avg_overlap growing stale when we are
5299 * indeed overlapping and hence not getting put to sleep, grow
5300 * the avg_overlap on preemption.
5302 * We use the average preemption runtime because that
5303 * correlates to the amount of cache footprint a task can
5306 update_avg(&prev->se.avg_overlap, runtime);
5308 prev->sched_class->put_prev_task(rq, prev);
5312 * Pick up the highest-prio task:
5314 static inline struct task_struct *
5315 pick_next_task(struct rq *rq)
5317 const struct sched_class *class;
5318 struct task_struct *p;
5321 * Optimization: we know that if all tasks are in
5322 * the fair class we can call that function directly:
5324 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5325 p = fair_sched_class.pick_next_task(rq);
5330 class = sched_class_highest;
5332 p = class->pick_next_task(rq);
5336 * Will never be NULL as the idle class always
5337 * returns a non-NULL p:
5339 class = class->next;
5344 * schedule() is the main scheduler function.
5346 asmlinkage void __sched schedule(void)
5348 struct task_struct *prev, *next;
5349 unsigned long *switch_count;
5355 cpu = smp_processor_id();
5359 switch_count = &prev->nivcsw;
5361 release_kernel_lock(prev);
5362 need_resched_nonpreemptible:
5364 schedule_debug(prev);
5366 if (sched_feat(HRTICK))
5369 spin_lock_irq(&rq->lock);
5370 update_rq_clock(rq);
5371 clear_tsk_need_resched(prev);
5373 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5374 if (unlikely(signal_pending_state(prev->state, prev)))
5375 prev->state = TASK_RUNNING;
5377 deactivate_task(rq, prev, 1);
5378 switch_count = &prev->nvcsw;
5381 pre_schedule(rq, prev);
5383 if (unlikely(!rq->nr_running))
5384 idle_balance(cpu, rq);
5386 put_prev_task(rq, prev);
5387 next = pick_next_task(rq);
5389 if (likely(prev != next)) {
5390 sched_info_switch(prev, next);
5391 perf_counter_task_sched_out(prev, next, cpu);
5397 context_switch(rq, prev, next); /* unlocks the rq */
5399 * the context switch might have flipped the stack from under
5400 * us, hence refresh the local variables.
5402 cpu = smp_processor_id();
5405 spin_unlock_irq(&rq->lock);
5409 if (unlikely(reacquire_kernel_lock(current) < 0))
5410 goto need_resched_nonpreemptible;
5412 preempt_enable_no_resched();
5416 EXPORT_SYMBOL(schedule);
5420 * Look out! "owner" is an entirely speculative pointer
5421 * access and not reliable.
5423 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5428 if (!sched_feat(OWNER_SPIN))
5431 #ifdef CONFIG_DEBUG_PAGEALLOC
5433 * Need to access the cpu field knowing that
5434 * DEBUG_PAGEALLOC could have unmapped it if
5435 * the mutex owner just released it and exited.
5437 if (probe_kernel_address(&owner->cpu, cpu))
5444 * Even if the access succeeded (likely case),
5445 * the cpu field may no longer be valid.
5447 if (cpu >= nr_cpumask_bits)
5451 * We need to validate that we can do a
5452 * get_cpu() and that we have the percpu area.
5454 if (!cpu_online(cpu))
5461 * Owner changed, break to re-assess state.
5463 if (lock->owner != owner)
5467 * Is that owner really running on that cpu?
5469 if (task_thread_info(rq->curr) != owner || need_resched())
5479 #ifdef CONFIG_PREEMPT
5481 * this is the entry point to schedule() from in-kernel preemption
5482 * off of preempt_enable. Kernel preemptions off return from interrupt
5483 * occur there and call schedule directly.
5485 asmlinkage void __sched preempt_schedule(void)
5487 struct thread_info *ti = current_thread_info();
5490 * If there is a non-zero preempt_count or interrupts are disabled,
5491 * we do not want to preempt the current task. Just return..
5493 if (likely(ti->preempt_count || irqs_disabled()))
5497 add_preempt_count(PREEMPT_ACTIVE);
5499 sub_preempt_count(PREEMPT_ACTIVE);
5502 * Check again in case we missed a preemption opportunity
5503 * between schedule and now.
5506 } while (need_resched());
5508 EXPORT_SYMBOL(preempt_schedule);
5511 * this is the entry point to schedule() from kernel preemption
5512 * off of irq context.
5513 * Note, that this is called and return with irqs disabled. This will
5514 * protect us against recursive calling from irq.
5516 asmlinkage void __sched preempt_schedule_irq(void)
5518 struct thread_info *ti = current_thread_info();
5520 /* Catch callers which need to be fixed */
5521 BUG_ON(ti->preempt_count || !irqs_disabled());
5524 add_preempt_count(PREEMPT_ACTIVE);
5527 local_irq_disable();
5528 sub_preempt_count(PREEMPT_ACTIVE);
5531 * Check again in case we missed a preemption opportunity
5532 * between schedule and now.
5535 } while (need_resched());
5538 #endif /* CONFIG_PREEMPT */
5540 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5543 return try_to_wake_up(curr->private, mode, sync);
5545 EXPORT_SYMBOL(default_wake_function);
5548 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5549 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5550 * number) then we wake all the non-exclusive tasks and one exclusive task.
5552 * There are circumstances in which we can try to wake a task which has already
5553 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5554 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5556 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5557 int nr_exclusive, int sync, void *key)
5559 wait_queue_t *curr, *next;
5561 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5562 unsigned flags = curr->flags;
5564 if (curr->func(curr, mode, sync, key) &&
5565 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5571 * __wake_up - wake up threads blocked on a waitqueue.
5573 * @mode: which threads
5574 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5575 * @key: is directly passed to the wakeup function
5577 * It may be assumed that this function implies a write memory barrier before
5578 * changing the task state if and only if any tasks are woken up.
5580 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5581 int nr_exclusive, void *key)
5583 unsigned long flags;
5585 spin_lock_irqsave(&q->lock, flags);
5586 __wake_up_common(q, mode, nr_exclusive, 0, key);
5587 spin_unlock_irqrestore(&q->lock, flags);
5589 EXPORT_SYMBOL(__wake_up);
5592 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5594 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5596 __wake_up_common(q, mode, 1, 0, NULL);
5599 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5601 __wake_up_common(q, mode, 1, 0, key);
5605 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5607 * @mode: which threads
5608 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5609 * @key: opaque value to be passed to wakeup targets
5611 * The sync wakeup differs that the waker knows that it will schedule
5612 * away soon, so while the target thread will be woken up, it will not
5613 * be migrated to another CPU - ie. the two threads are 'synchronized'
5614 * with each other. This can prevent needless bouncing between CPUs.
5616 * On UP it can prevent extra preemption.
5618 * It may be assumed that this function implies a write memory barrier before
5619 * changing the task state if and only if any tasks are woken up.
5621 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5622 int nr_exclusive, void *key)
5624 unsigned long flags;
5630 if (unlikely(!nr_exclusive))
5633 spin_lock_irqsave(&q->lock, flags);
5634 __wake_up_common(q, mode, nr_exclusive, sync, key);
5635 spin_unlock_irqrestore(&q->lock, flags);
5637 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5640 * __wake_up_sync - see __wake_up_sync_key()
5642 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5644 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5646 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5649 * complete: - signals a single thread waiting on this completion
5650 * @x: holds the state of this particular completion
5652 * This will wake up a single thread waiting on this completion. Threads will be
5653 * awakened in the same order in which they were queued.
5655 * See also complete_all(), wait_for_completion() and related routines.
5657 * It may be assumed that this function implies a write memory barrier before
5658 * changing the task state if and only if any tasks are woken up.
5660 void complete(struct completion *x)
5662 unsigned long flags;
5664 spin_lock_irqsave(&x->wait.lock, flags);
5666 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5667 spin_unlock_irqrestore(&x->wait.lock, flags);
5669 EXPORT_SYMBOL(complete);
5672 * complete_all: - signals all threads waiting on this completion
5673 * @x: holds the state of this particular completion
5675 * This will wake up all threads waiting on this particular completion event.
5677 * It may be assumed that this function implies a write memory barrier before
5678 * changing the task state if and only if any tasks are woken up.
5680 void complete_all(struct completion *x)
5682 unsigned long flags;
5684 spin_lock_irqsave(&x->wait.lock, flags);
5685 x->done += UINT_MAX/2;
5686 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5687 spin_unlock_irqrestore(&x->wait.lock, flags);
5689 EXPORT_SYMBOL(complete_all);
5691 static inline long __sched
5692 do_wait_for_common(struct completion *x, long timeout, int state)
5695 DECLARE_WAITQUEUE(wait, current);
5697 wait.flags |= WQ_FLAG_EXCLUSIVE;
5698 __add_wait_queue_tail(&x->wait, &wait);
5700 if (signal_pending_state(state, current)) {
5701 timeout = -ERESTARTSYS;
5704 __set_current_state(state);
5705 spin_unlock_irq(&x->wait.lock);
5706 timeout = schedule_timeout(timeout);
5707 spin_lock_irq(&x->wait.lock);
5708 } while (!x->done && timeout);
5709 __remove_wait_queue(&x->wait, &wait);
5714 return timeout ?: 1;
5718 wait_for_common(struct completion *x, long timeout, int state)
5722 spin_lock_irq(&x->wait.lock);
5723 timeout = do_wait_for_common(x, timeout, state);
5724 spin_unlock_irq(&x->wait.lock);
5729 * wait_for_completion: - waits for completion of a task
5730 * @x: holds the state of this particular completion
5732 * This waits to be signaled for completion of a specific task. It is NOT
5733 * interruptible and there is no timeout.
5735 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5736 * and interrupt capability. Also see complete().
5738 void __sched wait_for_completion(struct completion *x)
5740 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5742 EXPORT_SYMBOL(wait_for_completion);
5745 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5746 * @x: holds the state of this particular completion
5747 * @timeout: timeout value in jiffies
5749 * This waits for either a completion of a specific task to be signaled or for a
5750 * specified timeout to expire. The timeout is in jiffies. It is not
5753 unsigned long __sched
5754 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5756 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5758 EXPORT_SYMBOL(wait_for_completion_timeout);
5761 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5762 * @x: holds the state of this particular completion
5764 * This waits for completion of a specific task to be signaled. It is
5767 int __sched wait_for_completion_interruptible(struct completion *x)
5769 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5770 if (t == -ERESTARTSYS)
5774 EXPORT_SYMBOL(wait_for_completion_interruptible);
5777 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5778 * @x: holds the state of this particular completion
5779 * @timeout: timeout value in jiffies
5781 * This waits for either a completion of a specific task to be signaled or for a
5782 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5784 unsigned long __sched
5785 wait_for_completion_interruptible_timeout(struct completion *x,
5786 unsigned long timeout)
5788 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5790 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5793 * wait_for_completion_killable: - waits for completion of a task (killable)
5794 * @x: holds the state of this particular completion
5796 * This waits to be signaled for completion of a specific task. It can be
5797 * interrupted by a kill signal.
5799 int __sched wait_for_completion_killable(struct completion *x)
5801 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5802 if (t == -ERESTARTSYS)
5806 EXPORT_SYMBOL(wait_for_completion_killable);
5809 * try_wait_for_completion - try to decrement a completion without blocking
5810 * @x: completion structure
5812 * Returns: 0 if a decrement cannot be done without blocking
5813 * 1 if a decrement succeeded.
5815 * If a completion is being used as a counting completion,
5816 * attempt to decrement the counter without blocking. This
5817 * enables us to avoid waiting if the resource the completion
5818 * is protecting is not available.
5820 bool try_wait_for_completion(struct completion *x)
5824 spin_lock_irq(&x->wait.lock);
5829 spin_unlock_irq(&x->wait.lock);
5832 EXPORT_SYMBOL(try_wait_for_completion);
5835 * completion_done - Test to see if a completion has any waiters
5836 * @x: completion structure
5838 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5839 * 1 if there are no waiters.
5842 bool completion_done(struct completion *x)
5846 spin_lock_irq(&x->wait.lock);
5849 spin_unlock_irq(&x->wait.lock);
5852 EXPORT_SYMBOL(completion_done);
5855 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5857 unsigned long flags;
5860 init_waitqueue_entry(&wait, current);
5862 __set_current_state(state);
5864 spin_lock_irqsave(&q->lock, flags);
5865 __add_wait_queue(q, &wait);
5866 spin_unlock(&q->lock);
5867 timeout = schedule_timeout(timeout);
5868 spin_lock_irq(&q->lock);
5869 __remove_wait_queue(q, &wait);
5870 spin_unlock_irqrestore(&q->lock, flags);
5875 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5877 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5879 EXPORT_SYMBOL(interruptible_sleep_on);
5882 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5884 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5886 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5888 void __sched sleep_on(wait_queue_head_t *q)
5890 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5892 EXPORT_SYMBOL(sleep_on);
5894 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5896 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5898 EXPORT_SYMBOL(sleep_on_timeout);
5900 #ifdef CONFIG_RT_MUTEXES
5903 * rt_mutex_setprio - set the current priority of a task
5905 * @prio: prio value (kernel-internal form)
5907 * This function changes the 'effective' priority of a task. It does
5908 * not touch ->normal_prio like __setscheduler().
5910 * Used by the rt_mutex code to implement priority inheritance logic.
5912 void rt_mutex_setprio(struct task_struct *p, int prio)
5914 unsigned long flags;
5915 int oldprio, on_rq, running;
5917 const struct sched_class *prev_class = p->sched_class;
5919 BUG_ON(prio < 0 || prio > MAX_PRIO);
5921 rq = task_rq_lock(p, &flags);
5922 update_rq_clock(rq);
5925 on_rq = p->se.on_rq;
5926 running = task_current(rq, p);
5928 dequeue_task(rq, p, 0);
5930 p->sched_class->put_prev_task(rq, p);
5933 p->sched_class = &rt_sched_class;
5935 p->sched_class = &fair_sched_class;
5940 p->sched_class->set_curr_task(rq);
5942 enqueue_task(rq, p, 0);
5944 check_class_changed(rq, p, prev_class, oldprio, running);
5946 task_rq_unlock(rq, &flags);
5951 void set_user_nice(struct task_struct *p, long nice)
5953 int old_prio, delta, on_rq;
5954 unsigned long flags;
5957 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5960 * We have to be careful, if called from sys_setpriority(),
5961 * the task might be in the middle of scheduling on another CPU.
5963 rq = task_rq_lock(p, &flags);
5964 update_rq_clock(rq);
5966 * The RT priorities are set via sched_setscheduler(), but we still
5967 * allow the 'normal' nice value to be set - but as expected
5968 * it wont have any effect on scheduling until the task is
5969 * SCHED_FIFO/SCHED_RR:
5971 if (task_has_rt_policy(p)) {
5972 p->static_prio = NICE_TO_PRIO(nice);
5975 on_rq = p->se.on_rq;
5977 dequeue_task(rq, p, 0);
5979 p->static_prio = NICE_TO_PRIO(nice);
5982 p->prio = effective_prio(p);
5983 delta = p->prio - old_prio;
5986 enqueue_task(rq, p, 0);
5988 * If the task increased its priority or is running and
5989 * lowered its priority, then reschedule its CPU:
5991 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5992 resched_task(rq->curr);
5995 task_rq_unlock(rq, &flags);
5997 EXPORT_SYMBOL(set_user_nice);
6000 * can_nice - check if a task can reduce its nice value
6004 int can_nice(const struct task_struct *p, const int nice)
6006 /* convert nice value [19,-20] to rlimit style value [1,40] */
6007 int nice_rlim = 20 - nice;
6009 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6010 capable(CAP_SYS_NICE));
6013 #ifdef __ARCH_WANT_SYS_NICE
6016 * sys_nice - change the priority of the current process.
6017 * @increment: priority increment
6019 * sys_setpriority is a more generic, but much slower function that
6020 * does similar things.
6022 SYSCALL_DEFINE1(nice, int, increment)
6027 * Setpriority might change our priority at the same moment.
6028 * We don't have to worry. Conceptually one call occurs first
6029 * and we have a single winner.
6031 if (increment < -40)
6036 nice = TASK_NICE(current) + increment;
6042 if (increment < 0 && !can_nice(current, nice))
6045 retval = security_task_setnice(current, nice);
6049 set_user_nice(current, nice);
6056 * task_prio - return the priority value of a given task.
6057 * @p: the task in question.
6059 * This is the priority value as seen by users in /proc.
6060 * RT tasks are offset by -200. Normal tasks are centered
6061 * around 0, value goes from -16 to +15.
6063 int task_prio(const struct task_struct *p)
6065 return p->prio - MAX_RT_PRIO;
6069 * task_nice - return the nice value of a given task.
6070 * @p: the task in question.
6072 int task_nice(const struct task_struct *p)
6074 return TASK_NICE(p);
6076 EXPORT_SYMBOL(task_nice);
6079 * idle_cpu - is a given cpu idle currently?
6080 * @cpu: the processor in question.
6082 int idle_cpu(int cpu)
6084 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6088 * idle_task - return the idle task for a given cpu.
6089 * @cpu: the processor in question.
6091 struct task_struct *idle_task(int cpu)
6093 return cpu_rq(cpu)->idle;
6097 * find_process_by_pid - find a process with a matching PID value.
6098 * @pid: the pid in question.
6100 static struct task_struct *find_process_by_pid(pid_t pid)
6102 return pid ? find_task_by_vpid(pid) : current;
6105 /* Actually do priority change: must hold rq lock. */
6107 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6109 BUG_ON(p->se.on_rq);
6112 switch (p->policy) {
6116 p->sched_class = &fair_sched_class;
6120 p->sched_class = &rt_sched_class;
6124 p->rt_priority = prio;
6125 p->normal_prio = normal_prio(p);
6126 /* we are holding p->pi_lock already */
6127 p->prio = rt_mutex_getprio(p);
6132 * check the target process has a UID that matches the current process's
6134 static bool check_same_owner(struct task_struct *p)
6136 const struct cred *cred = current_cred(), *pcred;
6140 pcred = __task_cred(p);
6141 match = (cred->euid == pcred->euid ||
6142 cred->euid == pcred->uid);
6147 static int __sched_setscheduler(struct task_struct *p, int policy,
6148 struct sched_param *param, bool user)
6150 int retval, oldprio, oldpolicy = -1, on_rq, running;
6151 unsigned long flags;
6152 const struct sched_class *prev_class = p->sched_class;
6156 /* may grab non-irq protected spin_locks */
6157 BUG_ON(in_interrupt());
6159 /* double check policy once rq lock held */
6161 reset_on_fork = p->sched_reset_on_fork;
6162 policy = oldpolicy = p->policy;
6164 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6165 policy &= ~SCHED_RESET_ON_FORK;
6167 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6168 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6169 policy != SCHED_IDLE)
6174 * Valid priorities for SCHED_FIFO and SCHED_RR are
6175 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6176 * SCHED_BATCH and SCHED_IDLE is 0.
6178 if (param->sched_priority < 0 ||
6179 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6180 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6182 if (rt_policy(policy) != (param->sched_priority != 0))
6186 * Allow unprivileged RT tasks to decrease priority:
6188 if (user && !capable(CAP_SYS_NICE)) {
6189 if (rt_policy(policy)) {
6190 unsigned long rlim_rtprio;
6192 if (!lock_task_sighand(p, &flags))
6194 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6195 unlock_task_sighand(p, &flags);
6197 /* can't set/change the rt policy */
6198 if (policy != p->policy && !rlim_rtprio)
6201 /* can't increase priority */
6202 if (param->sched_priority > p->rt_priority &&
6203 param->sched_priority > rlim_rtprio)
6207 * Like positive nice levels, dont allow tasks to
6208 * move out of SCHED_IDLE either:
6210 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6213 /* can't change other user's priorities */
6214 if (!check_same_owner(p))
6217 /* Normal users shall not reset the sched_reset_on_fork flag */
6218 if (p->sched_reset_on_fork && !reset_on_fork)
6223 #ifdef CONFIG_RT_GROUP_SCHED
6225 * Do not allow realtime tasks into groups that have no runtime
6228 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6229 task_group(p)->rt_bandwidth.rt_runtime == 0)
6233 retval = security_task_setscheduler(p, policy, param);
6239 * make sure no PI-waiters arrive (or leave) while we are
6240 * changing the priority of the task:
6242 spin_lock_irqsave(&p->pi_lock, flags);
6244 * To be able to change p->policy safely, the apropriate
6245 * runqueue lock must be held.
6247 rq = __task_rq_lock(p);
6248 /* recheck policy now with rq lock held */
6249 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6250 policy = oldpolicy = -1;
6251 __task_rq_unlock(rq);
6252 spin_unlock_irqrestore(&p->pi_lock, flags);
6255 update_rq_clock(rq);
6256 on_rq = p->se.on_rq;
6257 running = task_current(rq, p);
6259 deactivate_task(rq, p, 0);
6261 p->sched_class->put_prev_task(rq, p);
6263 p->sched_reset_on_fork = reset_on_fork;
6266 __setscheduler(rq, p, policy, param->sched_priority);
6269 p->sched_class->set_curr_task(rq);
6271 activate_task(rq, p, 0);
6273 check_class_changed(rq, p, prev_class, oldprio, running);
6275 __task_rq_unlock(rq);
6276 spin_unlock_irqrestore(&p->pi_lock, flags);
6278 rt_mutex_adjust_pi(p);
6284 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6285 * @p: the task in question.
6286 * @policy: new policy.
6287 * @param: structure containing the new RT priority.
6289 * NOTE that the task may be already dead.
6291 int sched_setscheduler(struct task_struct *p, int policy,
6292 struct sched_param *param)
6294 return __sched_setscheduler(p, policy, param, true);
6296 EXPORT_SYMBOL_GPL(sched_setscheduler);
6299 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6300 * @p: the task in question.
6301 * @policy: new policy.
6302 * @param: structure containing the new RT priority.
6304 * Just like sched_setscheduler, only don't bother checking if the
6305 * current context has permission. For example, this is needed in
6306 * stop_machine(): we create temporary high priority worker threads,
6307 * but our caller might not have that capability.
6309 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6310 struct sched_param *param)
6312 return __sched_setscheduler(p, policy, param, false);
6316 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6318 struct sched_param lparam;
6319 struct task_struct *p;
6322 if (!param || pid < 0)
6324 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6329 p = find_process_by_pid(pid);
6331 retval = sched_setscheduler(p, policy, &lparam);
6338 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6339 * @pid: the pid in question.
6340 * @policy: new policy.
6341 * @param: structure containing the new RT priority.
6343 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6344 struct sched_param __user *, param)
6346 /* negative values for policy are not valid */
6350 return do_sched_setscheduler(pid, policy, param);
6354 * sys_sched_setparam - set/change the RT priority of a thread
6355 * @pid: the pid in question.
6356 * @param: structure containing the new RT priority.
6358 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6360 return do_sched_setscheduler(pid, -1, param);
6364 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6365 * @pid: the pid in question.
6367 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6369 struct task_struct *p;
6376 read_lock(&tasklist_lock);
6377 p = find_process_by_pid(pid);
6379 retval = security_task_getscheduler(p);
6382 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6384 read_unlock(&tasklist_lock);
6389 * sys_sched_getparam - get the RT priority of a thread
6390 * @pid: the pid in question.
6391 * @param: structure containing the RT priority.
6393 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6395 struct sched_param lp;
6396 struct task_struct *p;
6399 if (!param || pid < 0)
6402 read_lock(&tasklist_lock);
6403 p = find_process_by_pid(pid);
6408 retval = security_task_getscheduler(p);
6412 lp.sched_priority = p->rt_priority;
6413 read_unlock(&tasklist_lock);
6416 * This one might sleep, we cannot do it with a spinlock held ...
6418 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6423 read_unlock(&tasklist_lock);
6427 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6429 cpumask_var_t cpus_allowed, new_mask;
6430 struct task_struct *p;
6434 read_lock(&tasklist_lock);
6436 p = find_process_by_pid(pid);
6438 read_unlock(&tasklist_lock);
6444 * It is not safe to call set_cpus_allowed with the
6445 * tasklist_lock held. We will bump the task_struct's
6446 * usage count and then drop tasklist_lock.
6449 read_unlock(&tasklist_lock);
6451 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6455 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6457 goto out_free_cpus_allowed;
6460 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6463 retval = security_task_setscheduler(p, 0, NULL);
6467 cpuset_cpus_allowed(p, cpus_allowed);
6468 cpumask_and(new_mask, in_mask, cpus_allowed);
6470 retval = set_cpus_allowed_ptr(p, new_mask);
6473 cpuset_cpus_allowed(p, cpus_allowed);
6474 if (!cpumask_subset(new_mask, cpus_allowed)) {
6476 * We must have raced with a concurrent cpuset
6477 * update. Just reset the cpus_allowed to the
6478 * cpuset's cpus_allowed
6480 cpumask_copy(new_mask, cpus_allowed);
6485 free_cpumask_var(new_mask);
6486 out_free_cpus_allowed:
6487 free_cpumask_var(cpus_allowed);
6494 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6495 struct cpumask *new_mask)
6497 if (len < cpumask_size())
6498 cpumask_clear(new_mask);
6499 else if (len > cpumask_size())
6500 len = cpumask_size();
6502 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6506 * sys_sched_setaffinity - set the cpu affinity of a process
6507 * @pid: pid of the process
6508 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6509 * @user_mask_ptr: user-space pointer to the new cpu mask
6511 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6512 unsigned long __user *, user_mask_ptr)
6514 cpumask_var_t new_mask;
6517 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6520 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6522 retval = sched_setaffinity(pid, new_mask);
6523 free_cpumask_var(new_mask);
6527 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6529 struct task_struct *p;
6533 read_lock(&tasklist_lock);
6536 p = find_process_by_pid(pid);
6540 retval = security_task_getscheduler(p);
6544 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6547 read_unlock(&tasklist_lock);
6554 * sys_sched_getaffinity - get the cpu affinity of a process
6555 * @pid: pid of the process
6556 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6557 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6559 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6560 unsigned long __user *, user_mask_ptr)
6565 if (len < cpumask_size())
6568 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6571 ret = sched_getaffinity(pid, mask);
6573 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6576 ret = cpumask_size();
6578 free_cpumask_var(mask);
6584 * sys_sched_yield - yield the current processor to other threads.
6586 * This function yields the current CPU to other tasks. If there are no
6587 * other threads running on this CPU then this function will return.
6589 SYSCALL_DEFINE0(sched_yield)
6591 struct rq *rq = this_rq_lock();
6593 schedstat_inc(rq, yld_count);
6594 current->sched_class->yield_task(rq);
6597 * Since we are going to call schedule() anyway, there's
6598 * no need to preempt or enable interrupts:
6600 __release(rq->lock);
6601 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6602 _raw_spin_unlock(&rq->lock);
6603 preempt_enable_no_resched();
6610 static inline int should_resched(void)
6612 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6615 static void __cond_resched(void)
6617 add_preempt_count(PREEMPT_ACTIVE);
6619 sub_preempt_count(PREEMPT_ACTIVE);
6622 int __sched _cond_resched(void)
6624 if (should_resched()) {
6630 EXPORT_SYMBOL(_cond_resched);
6633 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6634 * call schedule, and on return reacquire the lock.
6636 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6637 * operations here to prevent schedule() from being called twice (once via
6638 * spin_unlock(), once by hand).
6640 int __cond_resched_lock(spinlock_t *lock)
6642 int resched = should_resched();
6645 lockdep_assert_held(lock);
6647 if (spin_needbreak(lock) || resched) {
6658 EXPORT_SYMBOL(__cond_resched_lock);
6660 int __sched __cond_resched_softirq(void)
6662 BUG_ON(!in_softirq());
6664 if (should_resched()) {
6672 EXPORT_SYMBOL(__cond_resched_softirq);
6675 * yield - yield the current processor to other threads.
6677 * This is a shortcut for kernel-space yielding - it marks the
6678 * thread runnable and calls sys_sched_yield().
6680 void __sched yield(void)
6682 set_current_state(TASK_RUNNING);
6685 EXPORT_SYMBOL(yield);
6688 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6689 * that process accounting knows that this is a task in IO wait state.
6691 * But don't do that if it is a deliberate, throttling IO wait (this task
6692 * has set its backing_dev_info: the queue against which it should throttle)
6694 void __sched io_schedule(void)
6696 struct rq *rq = raw_rq();
6698 delayacct_blkio_start();
6699 atomic_inc(&rq->nr_iowait);
6700 current->in_iowait = 1;
6702 current->in_iowait = 0;
6703 atomic_dec(&rq->nr_iowait);
6704 delayacct_blkio_end();
6706 EXPORT_SYMBOL(io_schedule);
6708 long __sched io_schedule_timeout(long timeout)
6710 struct rq *rq = raw_rq();
6713 delayacct_blkio_start();
6714 atomic_inc(&rq->nr_iowait);
6715 current->in_iowait = 1;
6716 ret = schedule_timeout(timeout);
6717 current->in_iowait = 0;
6718 atomic_dec(&rq->nr_iowait);
6719 delayacct_blkio_end();
6724 * sys_sched_get_priority_max - return maximum RT priority.
6725 * @policy: scheduling class.
6727 * this syscall returns the maximum rt_priority that can be used
6728 * by a given scheduling class.
6730 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6737 ret = MAX_USER_RT_PRIO-1;
6749 * sys_sched_get_priority_min - return minimum RT priority.
6750 * @policy: scheduling class.
6752 * this syscall returns the minimum rt_priority that can be used
6753 * by a given scheduling class.
6755 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6773 * sys_sched_rr_get_interval - return the default timeslice of a process.
6774 * @pid: pid of the process.
6775 * @interval: userspace pointer to the timeslice value.
6777 * this syscall writes the default timeslice value of a given process
6778 * into the user-space timespec buffer. A value of '0' means infinity.
6780 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6781 struct timespec __user *, interval)
6783 struct task_struct *p;
6784 unsigned int time_slice;
6792 read_lock(&tasklist_lock);
6793 p = find_process_by_pid(pid);
6797 retval = security_task_getscheduler(p);
6802 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6803 * tasks that are on an otherwise idle runqueue:
6806 if (p->policy == SCHED_RR) {
6807 time_slice = DEF_TIMESLICE;
6808 } else if (p->policy != SCHED_FIFO) {
6809 struct sched_entity *se = &p->se;
6810 unsigned long flags;
6813 rq = task_rq_lock(p, &flags);
6814 if (rq->cfs.load.weight)
6815 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6816 task_rq_unlock(rq, &flags);
6818 read_unlock(&tasklist_lock);
6819 jiffies_to_timespec(time_slice, &t);
6820 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6824 read_unlock(&tasklist_lock);
6828 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6830 void sched_show_task(struct task_struct *p)
6832 unsigned long free = 0;
6835 state = p->state ? __ffs(p->state) + 1 : 0;
6836 printk(KERN_INFO "%-13.13s %c", p->comm,
6837 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6838 #if BITS_PER_LONG == 32
6839 if (state == TASK_RUNNING)
6840 printk(KERN_CONT " running ");
6842 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6844 if (state == TASK_RUNNING)
6845 printk(KERN_CONT " running task ");
6847 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6849 #ifdef CONFIG_DEBUG_STACK_USAGE
6850 free = stack_not_used(p);
6852 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6853 task_pid_nr(p), task_pid_nr(p->real_parent),
6854 (unsigned long)task_thread_info(p)->flags);
6856 show_stack(p, NULL);
6859 void show_state_filter(unsigned long state_filter)
6861 struct task_struct *g, *p;
6863 #if BITS_PER_LONG == 32
6865 " task PC stack pid father\n");
6868 " task PC stack pid father\n");
6870 read_lock(&tasklist_lock);
6871 do_each_thread(g, p) {
6873 * reset the NMI-timeout, listing all files on a slow
6874 * console might take alot of time:
6876 touch_nmi_watchdog();
6877 if (!state_filter || (p->state & state_filter))
6879 } while_each_thread(g, p);
6881 touch_all_softlockup_watchdogs();
6883 #ifdef CONFIG_SCHED_DEBUG
6884 sysrq_sched_debug_show();
6886 read_unlock(&tasklist_lock);
6888 * Only show locks if all tasks are dumped:
6890 if (state_filter == -1)
6891 debug_show_all_locks();
6894 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6896 idle->sched_class = &idle_sched_class;
6900 * init_idle - set up an idle thread for a given CPU
6901 * @idle: task in question
6902 * @cpu: cpu the idle task belongs to
6904 * NOTE: this function does not set the idle thread's NEED_RESCHED
6905 * flag, to make booting more robust.
6907 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6909 struct rq *rq = cpu_rq(cpu);
6910 unsigned long flags;
6912 spin_lock_irqsave(&rq->lock, flags);
6915 idle->se.exec_start = sched_clock();
6917 idle->prio = idle->normal_prio = MAX_PRIO;
6918 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6919 __set_task_cpu(idle, cpu);
6921 rq->curr = rq->idle = idle;
6922 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6925 spin_unlock_irqrestore(&rq->lock, flags);
6927 /* Set the preempt count _outside_ the spinlocks! */
6928 #if defined(CONFIG_PREEMPT)
6929 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6931 task_thread_info(idle)->preempt_count = 0;
6934 * The idle tasks have their own, simple scheduling class:
6936 idle->sched_class = &idle_sched_class;
6937 ftrace_graph_init_task(idle);
6941 * In a system that switches off the HZ timer nohz_cpu_mask
6942 * indicates which cpus entered this state. This is used
6943 * in the rcu update to wait only for active cpus. For system
6944 * which do not switch off the HZ timer nohz_cpu_mask should
6945 * always be CPU_BITS_NONE.
6947 cpumask_var_t nohz_cpu_mask;
6950 * Increase the granularity value when there are more CPUs,
6951 * because with more CPUs the 'effective latency' as visible
6952 * to users decreases. But the relationship is not linear,
6953 * so pick a second-best guess by going with the log2 of the
6956 * This idea comes from the SD scheduler of Con Kolivas:
6958 static inline void sched_init_granularity(void)
6960 unsigned int factor = 1 + ilog2(num_online_cpus());
6961 const unsigned long limit = 200000000;
6963 sysctl_sched_min_granularity *= factor;
6964 if (sysctl_sched_min_granularity > limit)
6965 sysctl_sched_min_granularity = limit;
6967 sysctl_sched_latency *= factor;
6968 if (sysctl_sched_latency > limit)
6969 sysctl_sched_latency = limit;
6971 sysctl_sched_wakeup_granularity *= factor;
6973 sysctl_sched_shares_ratelimit *= factor;
6978 * This is how migration works:
6980 * 1) we queue a struct migration_req structure in the source CPU's
6981 * runqueue and wake up that CPU's migration thread.
6982 * 2) we down() the locked semaphore => thread blocks.
6983 * 3) migration thread wakes up (implicitly it forces the migrated
6984 * thread off the CPU)
6985 * 4) it gets the migration request and checks whether the migrated
6986 * task is still in the wrong runqueue.
6987 * 5) if it's in the wrong runqueue then the migration thread removes
6988 * it and puts it into the right queue.
6989 * 6) migration thread up()s the semaphore.
6990 * 7) we wake up and the migration is done.
6994 * Change a given task's CPU affinity. Migrate the thread to a
6995 * proper CPU and schedule it away if the CPU it's executing on
6996 * is removed from the allowed bitmask.
6998 * NOTE: the caller must have a valid reference to the task, the
6999 * task must not exit() & deallocate itself prematurely. The
7000 * call is not atomic; no spinlocks may be held.
7002 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7004 struct migration_req req;
7005 unsigned long flags;
7009 rq = task_rq_lock(p, &flags);
7010 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
7015 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7016 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7021 if (p->sched_class->set_cpus_allowed)
7022 p->sched_class->set_cpus_allowed(p, new_mask);
7024 cpumask_copy(&p->cpus_allowed, new_mask);
7025 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7028 /* Can the task run on the task's current CPU? If so, we're done */
7029 if (cpumask_test_cpu(task_cpu(p), new_mask))
7032 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
7033 /* Need help from migration thread: drop lock and wait. */
7034 struct task_struct *mt = rq->migration_thread;
7036 get_task_struct(mt);
7037 task_rq_unlock(rq, &flags);
7038 wake_up_process(rq->migration_thread);
7039 put_task_struct(mt);
7040 wait_for_completion(&req.done);
7041 tlb_migrate_finish(p->mm);
7045 task_rq_unlock(rq, &flags);
7049 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7052 * Move (not current) task off this cpu, onto dest cpu. We're doing
7053 * this because either it can't run here any more (set_cpus_allowed()
7054 * away from this CPU, or CPU going down), or because we're
7055 * attempting to rebalance this task on exec (sched_exec).
7057 * So we race with normal scheduler movements, but that's OK, as long
7058 * as the task is no longer on this CPU.
7060 * Returns non-zero if task was successfully migrated.
7062 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7064 struct rq *rq_dest, *rq_src;
7067 if (unlikely(!cpu_active(dest_cpu)))
7070 rq_src = cpu_rq(src_cpu);
7071 rq_dest = cpu_rq(dest_cpu);
7073 double_rq_lock(rq_src, rq_dest);
7074 /* Already moved. */
7075 if (task_cpu(p) != src_cpu)
7077 /* Affinity changed (again). */
7078 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7081 on_rq = p->se.on_rq;
7083 deactivate_task(rq_src, p, 0);
7085 set_task_cpu(p, dest_cpu);
7087 activate_task(rq_dest, p, 0);
7088 check_preempt_curr(rq_dest, p, 0);
7093 double_rq_unlock(rq_src, rq_dest);
7097 #define RCU_MIGRATION_IDLE 0
7098 #define RCU_MIGRATION_NEED_QS 1
7099 #define RCU_MIGRATION_GOT_QS 2
7100 #define RCU_MIGRATION_MUST_SYNC 3
7103 * migration_thread - this is a highprio system thread that performs
7104 * thread migration by bumping thread off CPU then 'pushing' onto
7107 static int migration_thread(void *data)
7110 int cpu = (long)data;
7114 BUG_ON(rq->migration_thread != current);
7116 set_current_state(TASK_INTERRUPTIBLE);
7117 while (!kthread_should_stop()) {
7118 struct migration_req *req;
7119 struct list_head *head;
7121 spin_lock_irq(&rq->lock);
7123 if (cpu_is_offline(cpu)) {
7124 spin_unlock_irq(&rq->lock);
7128 if (rq->active_balance) {
7129 active_load_balance(rq, cpu);
7130 rq->active_balance = 0;
7133 head = &rq->migration_queue;
7135 if (list_empty(head)) {
7136 spin_unlock_irq(&rq->lock);
7138 set_current_state(TASK_INTERRUPTIBLE);
7141 req = list_entry(head->next, struct migration_req, list);
7142 list_del_init(head->next);
7144 if (req->task != NULL) {
7145 spin_unlock(&rq->lock);
7146 __migrate_task(req->task, cpu, req->dest_cpu);
7147 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7148 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7149 spin_unlock(&rq->lock);
7151 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7152 spin_unlock(&rq->lock);
7153 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7157 complete(&req->done);
7159 __set_current_state(TASK_RUNNING);
7164 #ifdef CONFIG_HOTPLUG_CPU
7166 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7170 local_irq_disable();
7171 ret = __migrate_task(p, src_cpu, dest_cpu);
7177 * Figure out where task on dead CPU should go, use force if necessary.
7179 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7182 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7185 /* Look for allowed, online CPU in same node. */
7186 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7187 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7190 /* Any allowed, online CPU? */
7191 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7192 if (dest_cpu < nr_cpu_ids)
7195 /* No more Mr. Nice Guy. */
7196 if (dest_cpu >= nr_cpu_ids) {
7197 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7198 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7201 * Don't tell them about moving exiting tasks or
7202 * kernel threads (both mm NULL), since they never
7205 if (p->mm && printk_ratelimit()) {
7206 printk(KERN_INFO "process %d (%s) no "
7207 "longer affine to cpu%d\n",
7208 task_pid_nr(p), p->comm, dead_cpu);
7213 /* It can have affinity changed while we were choosing. */
7214 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7219 * While a dead CPU has no uninterruptible tasks queued at this point,
7220 * it might still have a nonzero ->nr_uninterruptible counter, because
7221 * for performance reasons the counter is not stricly tracking tasks to
7222 * their home CPUs. So we just add the counter to another CPU's counter,
7223 * to keep the global sum constant after CPU-down:
7225 static void migrate_nr_uninterruptible(struct rq *rq_src)
7227 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7228 unsigned long flags;
7230 local_irq_save(flags);
7231 double_rq_lock(rq_src, rq_dest);
7232 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7233 rq_src->nr_uninterruptible = 0;
7234 double_rq_unlock(rq_src, rq_dest);
7235 local_irq_restore(flags);
7238 /* Run through task list and migrate tasks from the dead cpu. */
7239 static void migrate_live_tasks(int src_cpu)
7241 struct task_struct *p, *t;
7243 read_lock(&tasklist_lock);
7245 do_each_thread(t, p) {
7249 if (task_cpu(p) == src_cpu)
7250 move_task_off_dead_cpu(src_cpu, p);
7251 } while_each_thread(t, p);
7253 read_unlock(&tasklist_lock);
7257 * Schedules idle task to be the next runnable task on current CPU.
7258 * It does so by boosting its priority to highest possible.
7259 * Used by CPU offline code.
7261 void sched_idle_next(void)
7263 int this_cpu = smp_processor_id();
7264 struct rq *rq = cpu_rq(this_cpu);
7265 struct task_struct *p = rq->idle;
7266 unsigned long flags;
7268 /* cpu has to be offline */
7269 BUG_ON(cpu_online(this_cpu));
7272 * Strictly not necessary since rest of the CPUs are stopped by now
7273 * and interrupts disabled on the current cpu.
7275 spin_lock_irqsave(&rq->lock, flags);
7277 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7279 update_rq_clock(rq);
7280 activate_task(rq, p, 0);
7282 spin_unlock_irqrestore(&rq->lock, flags);
7286 * Ensures that the idle task is using init_mm right before its cpu goes
7289 void idle_task_exit(void)
7291 struct mm_struct *mm = current->active_mm;
7293 BUG_ON(cpu_online(smp_processor_id()));
7296 switch_mm(mm, &init_mm, current);
7300 /* called under rq->lock with disabled interrupts */
7301 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7303 struct rq *rq = cpu_rq(dead_cpu);
7305 /* Must be exiting, otherwise would be on tasklist. */
7306 BUG_ON(!p->exit_state);
7308 /* Cannot have done final schedule yet: would have vanished. */
7309 BUG_ON(p->state == TASK_DEAD);
7314 * Drop lock around migration; if someone else moves it,
7315 * that's OK. No task can be added to this CPU, so iteration is
7318 spin_unlock_irq(&rq->lock);
7319 move_task_off_dead_cpu(dead_cpu, p);
7320 spin_lock_irq(&rq->lock);
7325 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7326 static void migrate_dead_tasks(unsigned int dead_cpu)
7328 struct rq *rq = cpu_rq(dead_cpu);
7329 struct task_struct *next;
7332 if (!rq->nr_running)
7334 update_rq_clock(rq);
7335 next = pick_next_task(rq);
7338 next->sched_class->put_prev_task(rq, next);
7339 migrate_dead(dead_cpu, next);
7345 * remove the tasks which were accounted by rq from calc_load_tasks.
7347 static void calc_global_load_remove(struct rq *rq)
7349 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7350 rq->calc_load_active = 0;
7352 #endif /* CONFIG_HOTPLUG_CPU */
7354 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7356 static struct ctl_table sd_ctl_dir[] = {
7358 .procname = "sched_domain",
7364 static struct ctl_table sd_ctl_root[] = {
7366 .ctl_name = CTL_KERN,
7367 .procname = "kernel",
7369 .child = sd_ctl_dir,
7374 static struct ctl_table *sd_alloc_ctl_entry(int n)
7376 struct ctl_table *entry =
7377 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7382 static void sd_free_ctl_entry(struct ctl_table **tablep)
7384 struct ctl_table *entry;
7387 * In the intermediate directories, both the child directory and
7388 * procname are dynamically allocated and could fail but the mode
7389 * will always be set. In the lowest directory the names are
7390 * static strings and all have proc handlers.
7392 for (entry = *tablep; entry->mode; entry++) {
7394 sd_free_ctl_entry(&entry->child);
7395 if (entry->proc_handler == NULL)
7396 kfree(entry->procname);
7404 set_table_entry(struct ctl_table *entry,
7405 const char *procname, void *data, int maxlen,
7406 mode_t mode, proc_handler *proc_handler)
7408 entry->procname = procname;
7410 entry->maxlen = maxlen;
7412 entry->proc_handler = proc_handler;
7415 static struct ctl_table *
7416 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7418 struct ctl_table *table = sd_alloc_ctl_entry(13);
7423 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7424 sizeof(long), 0644, proc_doulongvec_minmax);
7425 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7426 sizeof(long), 0644, proc_doulongvec_minmax);
7427 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7428 sizeof(int), 0644, proc_dointvec_minmax);
7429 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7430 sizeof(int), 0644, proc_dointvec_minmax);
7431 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7432 sizeof(int), 0644, proc_dointvec_minmax);
7433 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7434 sizeof(int), 0644, proc_dointvec_minmax);
7435 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7436 sizeof(int), 0644, proc_dointvec_minmax);
7437 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7438 sizeof(int), 0644, proc_dointvec_minmax);
7439 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7440 sizeof(int), 0644, proc_dointvec_minmax);
7441 set_table_entry(&table[9], "cache_nice_tries",
7442 &sd->cache_nice_tries,
7443 sizeof(int), 0644, proc_dointvec_minmax);
7444 set_table_entry(&table[10], "flags", &sd->flags,
7445 sizeof(int), 0644, proc_dointvec_minmax);
7446 set_table_entry(&table[11], "name", sd->name,
7447 CORENAME_MAX_SIZE, 0444, proc_dostring);
7448 /* &table[12] is terminator */
7453 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7455 struct ctl_table *entry, *table;
7456 struct sched_domain *sd;
7457 int domain_num = 0, i;
7460 for_each_domain(cpu, sd)
7462 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7467 for_each_domain(cpu, sd) {
7468 snprintf(buf, 32, "domain%d", i);
7469 entry->procname = kstrdup(buf, GFP_KERNEL);
7471 entry->child = sd_alloc_ctl_domain_table(sd);
7478 static struct ctl_table_header *sd_sysctl_header;
7479 static void register_sched_domain_sysctl(void)
7481 int i, cpu_num = num_online_cpus();
7482 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7485 WARN_ON(sd_ctl_dir[0].child);
7486 sd_ctl_dir[0].child = entry;
7491 for_each_online_cpu(i) {
7492 snprintf(buf, 32, "cpu%d", i);
7493 entry->procname = kstrdup(buf, GFP_KERNEL);
7495 entry->child = sd_alloc_ctl_cpu_table(i);
7499 WARN_ON(sd_sysctl_header);
7500 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7503 /* may be called multiple times per register */
7504 static void unregister_sched_domain_sysctl(void)
7506 if (sd_sysctl_header)
7507 unregister_sysctl_table(sd_sysctl_header);
7508 sd_sysctl_header = NULL;
7509 if (sd_ctl_dir[0].child)
7510 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7513 static void register_sched_domain_sysctl(void)
7516 static void unregister_sched_domain_sysctl(void)
7521 static void set_rq_online(struct rq *rq)
7524 const struct sched_class *class;
7526 cpumask_set_cpu(rq->cpu, rq->rd->online);
7529 for_each_class(class) {
7530 if (class->rq_online)
7531 class->rq_online(rq);
7536 static void set_rq_offline(struct rq *rq)
7539 const struct sched_class *class;
7541 for_each_class(class) {
7542 if (class->rq_offline)
7543 class->rq_offline(rq);
7546 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7552 * migration_call - callback that gets triggered when a CPU is added.
7553 * Here we can start up the necessary migration thread for the new CPU.
7555 static int __cpuinit
7556 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7558 struct task_struct *p;
7559 int cpu = (long)hcpu;
7560 unsigned long flags;
7565 case CPU_UP_PREPARE:
7566 case CPU_UP_PREPARE_FROZEN:
7567 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7570 kthread_bind(p, cpu);
7571 /* Must be high prio: stop_machine expects to yield to it. */
7572 rq = task_rq_lock(p, &flags);
7573 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7574 task_rq_unlock(rq, &flags);
7576 cpu_rq(cpu)->migration_thread = p;
7577 rq->calc_load_update = calc_load_update;
7581 case CPU_ONLINE_FROZEN:
7582 /* Strictly unnecessary, as first user will wake it. */
7583 wake_up_process(cpu_rq(cpu)->migration_thread);
7585 /* Update our root-domain */
7587 spin_lock_irqsave(&rq->lock, flags);
7589 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7593 spin_unlock_irqrestore(&rq->lock, flags);
7596 #ifdef CONFIG_HOTPLUG_CPU
7597 case CPU_UP_CANCELED:
7598 case CPU_UP_CANCELED_FROZEN:
7599 if (!cpu_rq(cpu)->migration_thread)
7601 /* Unbind it from offline cpu so it can run. Fall thru. */
7602 kthread_bind(cpu_rq(cpu)->migration_thread,
7603 cpumask_any(cpu_online_mask));
7604 kthread_stop(cpu_rq(cpu)->migration_thread);
7605 put_task_struct(cpu_rq(cpu)->migration_thread);
7606 cpu_rq(cpu)->migration_thread = NULL;
7610 case CPU_DEAD_FROZEN:
7611 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7612 migrate_live_tasks(cpu);
7614 kthread_stop(rq->migration_thread);
7615 put_task_struct(rq->migration_thread);
7616 rq->migration_thread = NULL;
7617 /* Idle task back to normal (off runqueue, low prio) */
7618 spin_lock_irq(&rq->lock);
7619 update_rq_clock(rq);
7620 deactivate_task(rq, rq->idle, 0);
7621 rq->idle->static_prio = MAX_PRIO;
7622 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7623 rq->idle->sched_class = &idle_sched_class;
7624 migrate_dead_tasks(cpu);
7625 spin_unlock_irq(&rq->lock);
7627 migrate_nr_uninterruptible(rq);
7628 BUG_ON(rq->nr_running != 0);
7629 calc_global_load_remove(rq);
7631 * No need to migrate the tasks: it was best-effort if
7632 * they didn't take sched_hotcpu_mutex. Just wake up
7635 spin_lock_irq(&rq->lock);
7636 while (!list_empty(&rq->migration_queue)) {
7637 struct migration_req *req;
7639 req = list_entry(rq->migration_queue.next,
7640 struct migration_req, list);
7641 list_del_init(&req->list);
7642 spin_unlock_irq(&rq->lock);
7643 complete(&req->done);
7644 spin_lock_irq(&rq->lock);
7646 spin_unlock_irq(&rq->lock);
7650 case CPU_DYING_FROZEN:
7651 /* Update our root-domain */
7653 spin_lock_irqsave(&rq->lock, flags);
7655 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7658 spin_unlock_irqrestore(&rq->lock, flags);
7666 * Register at high priority so that task migration (migrate_all_tasks)
7667 * happens before everything else. This has to be lower priority than
7668 * the notifier in the perf_counter subsystem, though.
7670 static struct notifier_block __cpuinitdata migration_notifier = {
7671 .notifier_call = migration_call,
7675 static int __init migration_init(void)
7677 void *cpu = (void *)(long)smp_processor_id();
7680 /* Start one for the boot CPU: */
7681 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7682 BUG_ON(err == NOTIFY_BAD);
7683 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7684 register_cpu_notifier(&migration_notifier);
7688 early_initcall(migration_init);
7693 #ifdef CONFIG_SCHED_DEBUG
7695 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7696 struct cpumask *groupmask)
7698 struct sched_group *group = sd->groups;
7701 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7702 cpumask_clear(groupmask);
7704 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7706 if (!(sd->flags & SD_LOAD_BALANCE)) {
7707 printk("does not load-balance\n");
7709 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7714 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7716 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7717 printk(KERN_ERR "ERROR: domain->span does not contain "
7720 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7721 printk(KERN_ERR "ERROR: domain->groups does not contain"
7725 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7729 printk(KERN_ERR "ERROR: group is NULL\n");
7733 if (!group->cpu_power) {
7734 printk(KERN_CONT "\n");
7735 printk(KERN_ERR "ERROR: domain->cpu_power not "
7740 if (!cpumask_weight(sched_group_cpus(group))) {
7741 printk(KERN_CONT "\n");
7742 printk(KERN_ERR "ERROR: empty group\n");
7746 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7747 printk(KERN_CONT "\n");
7748 printk(KERN_ERR "ERROR: repeated CPUs\n");
7752 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7754 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7756 printk(KERN_CONT " %s", str);
7757 if (group->cpu_power != SCHED_LOAD_SCALE) {
7758 printk(KERN_CONT " (cpu_power = %d)",
7762 group = group->next;
7763 } while (group != sd->groups);
7764 printk(KERN_CONT "\n");
7766 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7767 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7770 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7771 printk(KERN_ERR "ERROR: parent span is not a superset "
7772 "of domain->span\n");
7776 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7778 cpumask_var_t groupmask;
7782 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7786 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7788 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7789 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7794 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7801 free_cpumask_var(groupmask);
7803 #else /* !CONFIG_SCHED_DEBUG */
7804 # define sched_domain_debug(sd, cpu) do { } while (0)
7805 #endif /* CONFIG_SCHED_DEBUG */
7807 static int sd_degenerate(struct sched_domain *sd)
7809 if (cpumask_weight(sched_domain_span(sd)) == 1)
7812 /* Following flags need at least 2 groups */
7813 if (sd->flags & (SD_LOAD_BALANCE |
7814 SD_BALANCE_NEWIDLE |
7818 SD_SHARE_PKG_RESOURCES)) {
7819 if (sd->groups != sd->groups->next)
7823 /* Following flags don't use groups */
7824 if (sd->flags & (SD_WAKE_AFFINE))
7831 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7833 unsigned long cflags = sd->flags, pflags = parent->flags;
7835 if (sd_degenerate(parent))
7838 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7841 /* Flags needing groups don't count if only 1 group in parent */
7842 if (parent->groups == parent->groups->next) {
7843 pflags &= ~(SD_LOAD_BALANCE |
7844 SD_BALANCE_NEWIDLE |
7848 SD_SHARE_PKG_RESOURCES);
7849 if (nr_node_ids == 1)
7850 pflags &= ~SD_SERIALIZE;
7852 if (~cflags & pflags)
7858 static void free_rootdomain(struct root_domain *rd)
7860 cpupri_cleanup(&rd->cpupri);
7862 free_cpumask_var(rd->rto_mask);
7863 free_cpumask_var(rd->online);
7864 free_cpumask_var(rd->span);
7868 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7870 struct root_domain *old_rd = NULL;
7871 unsigned long flags;
7873 spin_lock_irqsave(&rq->lock, flags);
7878 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7881 cpumask_clear_cpu(rq->cpu, old_rd->span);
7884 * If we dont want to free the old_rt yet then
7885 * set old_rd to NULL to skip the freeing later
7888 if (!atomic_dec_and_test(&old_rd->refcount))
7892 atomic_inc(&rd->refcount);
7895 cpumask_set_cpu(rq->cpu, rd->span);
7896 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7899 spin_unlock_irqrestore(&rq->lock, flags);
7902 free_rootdomain(old_rd);
7905 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7907 gfp_t gfp = GFP_KERNEL;
7909 memset(rd, 0, sizeof(*rd));
7914 if (!alloc_cpumask_var(&rd->span, gfp))
7916 if (!alloc_cpumask_var(&rd->online, gfp))
7918 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
7921 if (cpupri_init(&rd->cpupri, bootmem) != 0)
7926 free_cpumask_var(rd->rto_mask);
7928 free_cpumask_var(rd->online);
7930 free_cpumask_var(rd->span);
7935 static void init_defrootdomain(void)
7937 init_rootdomain(&def_root_domain, true);
7939 atomic_set(&def_root_domain.refcount, 1);
7942 static struct root_domain *alloc_rootdomain(void)
7944 struct root_domain *rd;
7946 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7950 if (init_rootdomain(rd, false) != 0) {
7959 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7960 * hold the hotplug lock.
7963 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7965 struct rq *rq = cpu_rq(cpu);
7966 struct sched_domain *tmp;
7968 /* Remove the sched domains which do not contribute to scheduling. */
7969 for (tmp = sd; tmp; ) {
7970 struct sched_domain *parent = tmp->parent;
7974 if (sd_parent_degenerate(tmp, parent)) {
7975 tmp->parent = parent->parent;
7977 parent->parent->child = tmp;
7982 if (sd && sd_degenerate(sd)) {
7988 sched_domain_debug(sd, cpu);
7990 rq_attach_root(rq, rd);
7991 rcu_assign_pointer(rq->sd, sd);
7994 /* cpus with isolated domains */
7995 static cpumask_var_t cpu_isolated_map;
7997 /* Setup the mask of cpus configured for isolated domains */
7998 static int __init isolated_cpu_setup(char *str)
8000 cpulist_parse(str, cpu_isolated_map);
8004 __setup("isolcpus=", isolated_cpu_setup);
8007 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8008 * to a function which identifies what group(along with sched group) a CPU
8009 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8010 * (due to the fact that we keep track of groups covered with a struct cpumask).
8012 * init_sched_build_groups will build a circular linked list of the groups
8013 * covered by the given span, and will set each group's ->cpumask correctly,
8014 * and ->cpu_power to 0.
8017 init_sched_build_groups(const struct cpumask *span,
8018 const struct cpumask *cpu_map,
8019 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8020 struct sched_group **sg,
8021 struct cpumask *tmpmask),
8022 struct cpumask *covered, struct cpumask *tmpmask)
8024 struct sched_group *first = NULL, *last = NULL;
8027 cpumask_clear(covered);
8029 for_each_cpu(i, span) {
8030 struct sched_group *sg;
8031 int group = group_fn(i, cpu_map, &sg, tmpmask);
8034 if (cpumask_test_cpu(i, covered))
8037 cpumask_clear(sched_group_cpus(sg));
8040 for_each_cpu(j, span) {
8041 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8044 cpumask_set_cpu(j, covered);
8045 cpumask_set_cpu(j, sched_group_cpus(sg));
8056 #define SD_NODES_PER_DOMAIN 16
8061 * find_next_best_node - find the next node to include in a sched_domain
8062 * @node: node whose sched_domain we're building
8063 * @used_nodes: nodes already in the sched_domain
8065 * Find the next node to include in a given scheduling domain. Simply
8066 * finds the closest node not already in the @used_nodes map.
8068 * Should use nodemask_t.
8070 static int find_next_best_node(int node, nodemask_t *used_nodes)
8072 int i, n, val, min_val, best_node = 0;
8076 for (i = 0; i < nr_node_ids; i++) {
8077 /* Start at @node */
8078 n = (node + i) % nr_node_ids;
8080 if (!nr_cpus_node(n))
8083 /* Skip already used nodes */
8084 if (node_isset(n, *used_nodes))
8087 /* Simple min distance search */
8088 val = node_distance(node, n);
8090 if (val < min_val) {
8096 node_set(best_node, *used_nodes);
8101 * sched_domain_node_span - get a cpumask for a node's sched_domain
8102 * @node: node whose cpumask we're constructing
8103 * @span: resulting cpumask
8105 * Given a node, construct a good cpumask for its sched_domain to span. It
8106 * should be one that prevents unnecessary balancing, but also spreads tasks
8109 static void sched_domain_node_span(int node, struct cpumask *span)
8111 nodemask_t used_nodes;
8114 cpumask_clear(span);
8115 nodes_clear(used_nodes);
8117 cpumask_or(span, span, cpumask_of_node(node));
8118 node_set(node, used_nodes);
8120 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8121 int next_node = find_next_best_node(node, &used_nodes);
8123 cpumask_or(span, span, cpumask_of_node(next_node));
8126 #endif /* CONFIG_NUMA */
8128 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8131 * The cpus mask in sched_group and sched_domain hangs off the end.
8133 * ( See the the comments in include/linux/sched.h:struct sched_group
8134 * and struct sched_domain. )
8136 struct static_sched_group {
8137 struct sched_group sg;
8138 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8141 struct static_sched_domain {
8142 struct sched_domain sd;
8143 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8149 cpumask_var_t domainspan;
8150 cpumask_var_t covered;
8151 cpumask_var_t notcovered;
8153 cpumask_var_t nodemask;
8154 cpumask_var_t this_sibling_map;
8155 cpumask_var_t this_core_map;
8156 cpumask_var_t send_covered;
8157 cpumask_var_t tmpmask;
8158 struct sched_group **sched_group_nodes;
8159 struct root_domain *rd;
8163 sa_sched_groups = 0,
8168 sa_this_sibling_map,
8170 sa_sched_group_nodes,
8180 * SMT sched-domains:
8182 #ifdef CONFIG_SCHED_SMT
8183 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8184 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8187 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8188 struct sched_group **sg, struct cpumask *unused)
8191 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8194 #endif /* CONFIG_SCHED_SMT */
8197 * multi-core sched-domains:
8199 #ifdef CONFIG_SCHED_MC
8200 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8201 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8202 #endif /* CONFIG_SCHED_MC */
8204 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8206 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8207 struct sched_group **sg, struct cpumask *mask)
8211 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8212 group = cpumask_first(mask);
8214 *sg = &per_cpu(sched_group_core, group).sg;
8217 #elif defined(CONFIG_SCHED_MC)
8219 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8220 struct sched_group **sg, struct cpumask *unused)
8223 *sg = &per_cpu(sched_group_core, cpu).sg;
8228 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8229 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8232 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8233 struct sched_group **sg, struct cpumask *mask)
8236 #ifdef CONFIG_SCHED_MC
8237 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8238 group = cpumask_first(mask);
8239 #elif defined(CONFIG_SCHED_SMT)
8240 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8241 group = cpumask_first(mask);
8246 *sg = &per_cpu(sched_group_phys, group).sg;
8252 * The init_sched_build_groups can't handle what we want to do with node
8253 * groups, so roll our own. Now each node has its own list of groups which
8254 * gets dynamically allocated.
8256 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8257 static struct sched_group ***sched_group_nodes_bycpu;
8259 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8260 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8262 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8263 struct sched_group **sg,
8264 struct cpumask *nodemask)
8268 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8269 group = cpumask_first(nodemask);
8272 *sg = &per_cpu(sched_group_allnodes, group).sg;
8276 static void init_numa_sched_groups_power(struct sched_group *group_head)
8278 struct sched_group *sg = group_head;
8284 for_each_cpu(j, sched_group_cpus(sg)) {
8285 struct sched_domain *sd;
8287 sd = &per_cpu(phys_domains, j).sd;
8288 if (j != group_first_cpu(sd->groups)) {
8290 * Only add "power" once for each
8296 sg->cpu_power += sd->groups->cpu_power;
8299 } while (sg != group_head);
8302 static int build_numa_sched_groups(struct s_data *d,
8303 const struct cpumask *cpu_map, int num)
8305 struct sched_domain *sd;
8306 struct sched_group *sg, *prev;
8309 cpumask_clear(d->covered);
8310 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8311 if (cpumask_empty(d->nodemask)) {
8312 d->sched_group_nodes[num] = NULL;
8316 sched_domain_node_span(num, d->domainspan);
8317 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8319 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8322 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8326 d->sched_group_nodes[num] = sg;
8328 for_each_cpu(j, d->nodemask) {
8329 sd = &per_cpu(node_domains, j).sd;
8334 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8336 cpumask_or(d->covered, d->covered, d->nodemask);
8339 for (j = 0; j < nr_node_ids; j++) {
8340 n = (num + j) % nr_node_ids;
8341 cpumask_complement(d->notcovered, d->covered);
8342 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8343 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8344 if (cpumask_empty(d->tmpmask))
8346 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8347 if (cpumask_empty(d->tmpmask))
8349 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8353 "Can not alloc domain group for node %d\n", j);
8357 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8358 sg->next = prev->next;
8359 cpumask_or(d->covered, d->covered, d->tmpmask);
8366 #endif /* CONFIG_NUMA */
8369 /* Free memory allocated for various sched_group structures */
8370 static void free_sched_groups(const struct cpumask *cpu_map,
8371 struct cpumask *nodemask)
8375 for_each_cpu(cpu, cpu_map) {
8376 struct sched_group **sched_group_nodes
8377 = sched_group_nodes_bycpu[cpu];
8379 if (!sched_group_nodes)
8382 for (i = 0; i < nr_node_ids; i++) {
8383 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8385 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8386 if (cpumask_empty(nodemask))
8396 if (oldsg != sched_group_nodes[i])
8399 kfree(sched_group_nodes);
8400 sched_group_nodes_bycpu[cpu] = NULL;
8403 #else /* !CONFIG_NUMA */
8404 static void free_sched_groups(const struct cpumask *cpu_map,
8405 struct cpumask *nodemask)
8408 #endif /* CONFIG_NUMA */
8411 * Initialize sched groups cpu_power.
8413 * cpu_power indicates the capacity of sched group, which is used while
8414 * distributing the load between different sched groups in a sched domain.
8415 * Typically cpu_power for all the groups in a sched domain will be same unless
8416 * there are asymmetries in the topology. If there are asymmetries, group
8417 * having more cpu_power will pickup more load compared to the group having
8420 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8422 struct sched_domain *child;
8423 struct sched_group *group;
8427 WARN_ON(!sd || !sd->groups);
8429 if (cpu != group_first_cpu(sd->groups))
8434 sd->groups->cpu_power = 0;
8437 power = SCHED_LOAD_SCALE;
8438 weight = cpumask_weight(sched_domain_span(sd));
8440 * SMT siblings share the power of a single core.
8441 * Usually multiple threads get a better yield out of
8442 * that one core than a single thread would have,
8443 * reflect that in sd->smt_gain.
8445 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8446 power *= sd->smt_gain;
8448 power >>= SCHED_LOAD_SHIFT;
8450 sd->groups->cpu_power += power;
8455 * Add cpu_power of each child group to this groups cpu_power.
8457 group = child->groups;
8459 sd->groups->cpu_power += group->cpu_power;
8460 group = group->next;
8461 } while (group != child->groups);
8465 * Initializers for schedule domains
8466 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8469 #ifdef CONFIG_SCHED_DEBUG
8470 # define SD_INIT_NAME(sd, type) sd->name = #type
8472 # define SD_INIT_NAME(sd, type) do { } while (0)
8475 #define SD_INIT(sd, type) sd_init_##type(sd)
8477 #define SD_INIT_FUNC(type) \
8478 static noinline void sd_init_##type(struct sched_domain *sd) \
8480 memset(sd, 0, sizeof(*sd)); \
8481 *sd = SD_##type##_INIT; \
8482 sd->level = SD_LV_##type; \
8483 SD_INIT_NAME(sd, type); \
8488 SD_INIT_FUNC(ALLNODES)
8491 #ifdef CONFIG_SCHED_SMT
8492 SD_INIT_FUNC(SIBLING)
8494 #ifdef CONFIG_SCHED_MC
8498 static int default_relax_domain_level = -1;
8500 static int __init setup_relax_domain_level(char *str)
8504 val = simple_strtoul(str, NULL, 0);
8505 if (val < SD_LV_MAX)
8506 default_relax_domain_level = val;
8510 __setup("relax_domain_level=", setup_relax_domain_level);
8512 static void set_domain_attribute(struct sched_domain *sd,
8513 struct sched_domain_attr *attr)
8517 if (!attr || attr->relax_domain_level < 0) {
8518 if (default_relax_domain_level < 0)
8521 request = default_relax_domain_level;
8523 request = attr->relax_domain_level;
8524 if (request < sd->level) {
8525 /* turn off idle balance on this domain */
8526 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8528 /* turn on idle balance on this domain */
8529 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8533 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8534 const struct cpumask *cpu_map)
8537 case sa_sched_groups:
8538 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8539 d->sched_group_nodes = NULL;
8541 free_rootdomain(d->rd); /* fall through */
8543 free_cpumask_var(d->tmpmask); /* fall through */
8544 case sa_send_covered:
8545 free_cpumask_var(d->send_covered); /* fall through */
8546 case sa_this_core_map:
8547 free_cpumask_var(d->this_core_map); /* fall through */
8548 case sa_this_sibling_map:
8549 free_cpumask_var(d->this_sibling_map); /* fall through */
8551 free_cpumask_var(d->nodemask); /* fall through */
8552 case sa_sched_group_nodes:
8554 kfree(d->sched_group_nodes); /* fall through */
8556 free_cpumask_var(d->notcovered); /* fall through */
8558 free_cpumask_var(d->covered); /* fall through */
8560 free_cpumask_var(d->domainspan); /* fall through */
8567 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8568 const struct cpumask *cpu_map)
8571 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8573 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8574 return sa_domainspan;
8575 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8577 /* Allocate the per-node list of sched groups */
8578 d->sched_group_nodes = kcalloc(nr_node_ids,
8579 sizeof(struct sched_group *), GFP_KERNEL);
8580 if (!d->sched_group_nodes) {
8581 printk(KERN_WARNING "Can not alloc sched group node list\n");
8582 return sa_notcovered;
8584 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8586 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8587 return sa_sched_group_nodes;
8588 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8590 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8591 return sa_this_sibling_map;
8592 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8593 return sa_this_core_map;
8594 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8595 return sa_send_covered;
8596 d->rd = alloc_rootdomain();
8598 printk(KERN_WARNING "Cannot alloc root domain\n");
8601 return sa_rootdomain;
8604 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8605 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8607 struct sched_domain *sd = NULL;
8609 struct sched_domain *parent;
8612 if (cpumask_weight(cpu_map) >
8613 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8614 sd = &per_cpu(allnodes_domains, i).sd;
8615 SD_INIT(sd, ALLNODES);
8616 set_domain_attribute(sd, attr);
8617 cpumask_copy(sched_domain_span(sd), cpu_map);
8618 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8623 sd = &per_cpu(node_domains, i).sd;
8625 set_domain_attribute(sd, attr);
8626 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8627 sd->parent = parent;
8630 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8635 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8636 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8637 struct sched_domain *parent, int i)
8639 struct sched_domain *sd;
8640 sd = &per_cpu(phys_domains, i).sd;
8642 set_domain_attribute(sd, attr);
8643 cpumask_copy(sched_domain_span(sd), d->nodemask);
8644 sd->parent = parent;
8647 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8651 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8652 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8653 struct sched_domain *parent, int i)
8655 struct sched_domain *sd = parent;
8656 #ifdef CONFIG_SCHED_MC
8657 sd = &per_cpu(core_domains, i).sd;
8659 set_domain_attribute(sd, attr);
8660 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8661 sd->parent = parent;
8663 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8668 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8669 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8670 struct sched_domain *parent, int i)
8672 struct sched_domain *sd = parent;
8673 #ifdef CONFIG_SCHED_SMT
8674 sd = &per_cpu(cpu_domains, i).sd;
8675 SD_INIT(sd, SIBLING);
8676 set_domain_attribute(sd, attr);
8677 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8678 sd->parent = parent;
8680 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8685 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8686 const struct cpumask *cpu_map, int cpu)
8689 #ifdef CONFIG_SCHED_SMT
8690 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8691 cpumask_and(d->this_sibling_map, cpu_map,
8692 topology_thread_cpumask(cpu));
8693 if (cpu == cpumask_first(d->this_sibling_map))
8694 init_sched_build_groups(d->this_sibling_map, cpu_map,
8696 d->send_covered, d->tmpmask);
8699 #ifdef CONFIG_SCHED_MC
8700 case SD_LV_MC: /* set up multi-core groups */
8701 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8702 if (cpu == cpumask_first(d->this_core_map))
8703 init_sched_build_groups(d->this_core_map, cpu_map,
8705 d->send_covered, d->tmpmask);
8708 case SD_LV_CPU: /* set up physical groups */
8709 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8710 if (!cpumask_empty(d->nodemask))
8711 init_sched_build_groups(d->nodemask, cpu_map,
8713 d->send_covered, d->tmpmask);
8716 case SD_LV_ALLNODES:
8717 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8718 d->send_covered, d->tmpmask);
8727 * Build sched domains for a given set of cpus and attach the sched domains
8728 * to the individual cpus
8730 static int __build_sched_domains(const struct cpumask *cpu_map,
8731 struct sched_domain_attr *attr)
8733 enum s_alloc alloc_state = sa_none;
8735 struct sched_domain *sd;
8741 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8742 if (alloc_state != sa_rootdomain)
8744 alloc_state = sa_sched_groups;
8747 * Set up domains for cpus specified by the cpu_map.
8749 for_each_cpu(i, cpu_map) {
8750 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8753 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8754 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8755 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8756 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8759 for_each_cpu(i, cpu_map) {
8760 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8761 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8764 /* Set up physical groups */
8765 for (i = 0; i < nr_node_ids; i++)
8766 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8769 /* Set up node groups */
8771 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8773 for (i = 0; i < nr_node_ids; i++)
8774 if (build_numa_sched_groups(&d, cpu_map, i))
8778 /* Calculate CPU power for physical packages and nodes */
8779 #ifdef CONFIG_SCHED_SMT
8780 for_each_cpu(i, cpu_map) {
8781 sd = &per_cpu(cpu_domains, i).sd;
8782 init_sched_groups_power(i, sd);
8785 #ifdef CONFIG_SCHED_MC
8786 for_each_cpu(i, cpu_map) {
8787 sd = &per_cpu(core_domains, i).sd;
8788 init_sched_groups_power(i, sd);
8792 for_each_cpu(i, cpu_map) {
8793 sd = &per_cpu(phys_domains, i).sd;
8794 init_sched_groups_power(i, sd);
8798 for (i = 0; i < nr_node_ids; i++)
8799 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8801 if (d.sd_allnodes) {
8802 struct sched_group *sg;
8804 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8806 init_numa_sched_groups_power(sg);
8810 /* Attach the domains */
8811 for_each_cpu(i, cpu_map) {
8812 #ifdef CONFIG_SCHED_SMT
8813 sd = &per_cpu(cpu_domains, i).sd;
8814 #elif defined(CONFIG_SCHED_MC)
8815 sd = &per_cpu(core_domains, i).sd;
8817 sd = &per_cpu(phys_domains, i).sd;
8819 cpu_attach_domain(sd, d.rd, i);
8822 d.sched_group_nodes = NULL; /* don't free this we still need it */
8823 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8827 __free_domain_allocs(&d, alloc_state, cpu_map);
8831 static int build_sched_domains(const struct cpumask *cpu_map)
8833 return __build_sched_domains(cpu_map, NULL);
8836 static struct cpumask *doms_cur; /* current sched domains */
8837 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8838 static struct sched_domain_attr *dattr_cur;
8839 /* attribues of custom domains in 'doms_cur' */
8842 * Special case: If a kmalloc of a doms_cur partition (array of
8843 * cpumask) fails, then fallback to a single sched domain,
8844 * as determined by the single cpumask fallback_doms.
8846 static cpumask_var_t fallback_doms;
8849 * arch_update_cpu_topology lets virtualized architectures update the
8850 * cpu core maps. It is supposed to return 1 if the topology changed
8851 * or 0 if it stayed the same.
8853 int __attribute__((weak)) arch_update_cpu_topology(void)
8859 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8860 * For now this just excludes isolated cpus, but could be used to
8861 * exclude other special cases in the future.
8863 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8867 arch_update_cpu_topology();
8869 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8871 doms_cur = fallback_doms;
8872 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8874 err = build_sched_domains(doms_cur);
8875 register_sched_domain_sysctl();
8880 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8881 struct cpumask *tmpmask)
8883 free_sched_groups(cpu_map, tmpmask);
8887 * Detach sched domains from a group of cpus specified in cpu_map
8888 * These cpus will now be attached to the NULL domain
8890 static void detach_destroy_domains(const struct cpumask *cpu_map)
8892 /* Save because hotplug lock held. */
8893 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8896 for_each_cpu(i, cpu_map)
8897 cpu_attach_domain(NULL, &def_root_domain, i);
8898 synchronize_sched();
8899 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8902 /* handle null as "default" */
8903 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8904 struct sched_domain_attr *new, int idx_new)
8906 struct sched_domain_attr tmp;
8913 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8914 new ? (new + idx_new) : &tmp,
8915 sizeof(struct sched_domain_attr));
8919 * Partition sched domains as specified by the 'ndoms_new'
8920 * cpumasks in the array doms_new[] of cpumasks. This compares
8921 * doms_new[] to the current sched domain partitioning, doms_cur[].
8922 * It destroys each deleted domain and builds each new domain.
8924 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8925 * The masks don't intersect (don't overlap.) We should setup one
8926 * sched domain for each mask. CPUs not in any of the cpumasks will
8927 * not be load balanced. If the same cpumask appears both in the
8928 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8931 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8932 * ownership of it and will kfree it when done with it. If the caller
8933 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8934 * ndoms_new == 1, and partition_sched_domains() will fallback to
8935 * the single partition 'fallback_doms', it also forces the domains
8938 * If doms_new == NULL it will be replaced with cpu_online_mask.
8939 * ndoms_new == 0 is a special case for destroying existing domains,
8940 * and it will not create the default domain.
8942 * Call with hotplug lock held
8944 /* FIXME: Change to struct cpumask *doms_new[] */
8945 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8946 struct sched_domain_attr *dattr_new)
8951 mutex_lock(&sched_domains_mutex);
8953 /* always unregister in case we don't destroy any domains */
8954 unregister_sched_domain_sysctl();
8956 /* Let architecture update cpu core mappings. */
8957 new_topology = arch_update_cpu_topology();
8959 n = doms_new ? ndoms_new : 0;
8961 /* Destroy deleted domains */
8962 for (i = 0; i < ndoms_cur; i++) {
8963 for (j = 0; j < n && !new_topology; j++) {
8964 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8965 && dattrs_equal(dattr_cur, i, dattr_new, j))
8968 /* no match - a current sched domain not in new doms_new[] */
8969 detach_destroy_domains(doms_cur + i);
8974 if (doms_new == NULL) {
8976 doms_new = fallback_doms;
8977 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8978 WARN_ON_ONCE(dattr_new);
8981 /* Build new domains */
8982 for (i = 0; i < ndoms_new; i++) {
8983 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8984 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8985 && dattrs_equal(dattr_new, i, dattr_cur, j))
8988 /* no match - add a new doms_new */
8989 __build_sched_domains(doms_new + i,
8990 dattr_new ? dattr_new + i : NULL);
8995 /* Remember the new sched domains */
8996 if (doms_cur != fallback_doms)
8998 kfree(dattr_cur); /* kfree(NULL) is safe */
8999 doms_cur = doms_new;
9000 dattr_cur = dattr_new;
9001 ndoms_cur = ndoms_new;
9003 register_sched_domain_sysctl();
9005 mutex_unlock(&sched_domains_mutex);
9008 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9009 static void arch_reinit_sched_domains(void)
9013 /* Destroy domains first to force the rebuild */
9014 partition_sched_domains(0, NULL, NULL);
9016 rebuild_sched_domains();
9020 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9022 unsigned int level = 0;
9024 if (sscanf(buf, "%u", &level) != 1)
9028 * level is always be positive so don't check for
9029 * level < POWERSAVINGS_BALANCE_NONE which is 0
9030 * What happens on 0 or 1 byte write,
9031 * need to check for count as well?
9034 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9038 sched_smt_power_savings = level;
9040 sched_mc_power_savings = level;
9042 arch_reinit_sched_domains();
9047 #ifdef CONFIG_SCHED_MC
9048 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9051 return sprintf(page, "%u\n", sched_mc_power_savings);
9053 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9054 const char *buf, size_t count)
9056 return sched_power_savings_store(buf, count, 0);
9058 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9059 sched_mc_power_savings_show,
9060 sched_mc_power_savings_store);
9063 #ifdef CONFIG_SCHED_SMT
9064 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9067 return sprintf(page, "%u\n", sched_smt_power_savings);
9069 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9070 const char *buf, size_t count)
9072 return sched_power_savings_store(buf, count, 1);
9074 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9075 sched_smt_power_savings_show,
9076 sched_smt_power_savings_store);
9079 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9083 #ifdef CONFIG_SCHED_SMT
9085 err = sysfs_create_file(&cls->kset.kobj,
9086 &attr_sched_smt_power_savings.attr);
9088 #ifdef CONFIG_SCHED_MC
9089 if (!err && mc_capable())
9090 err = sysfs_create_file(&cls->kset.kobj,
9091 &attr_sched_mc_power_savings.attr);
9095 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9097 #ifndef CONFIG_CPUSETS
9099 * Add online and remove offline CPUs from the scheduler domains.
9100 * When cpusets are enabled they take over this function.
9102 static int update_sched_domains(struct notifier_block *nfb,
9103 unsigned long action, void *hcpu)
9107 case CPU_ONLINE_FROZEN:
9109 case CPU_DEAD_FROZEN:
9110 partition_sched_domains(1, NULL, NULL);
9119 static int update_runtime(struct notifier_block *nfb,
9120 unsigned long action, void *hcpu)
9122 int cpu = (int)(long)hcpu;
9125 case CPU_DOWN_PREPARE:
9126 case CPU_DOWN_PREPARE_FROZEN:
9127 disable_runtime(cpu_rq(cpu));
9130 case CPU_DOWN_FAILED:
9131 case CPU_DOWN_FAILED_FROZEN:
9133 case CPU_ONLINE_FROZEN:
9134 enable_runtime(cpu_rq(cpu));
9142 void __init sched_init_smp(void)
9144 cpumask_var_t non_isolated_cpus;
9146 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9148 #if defined(CONFIG_NUMA)
9149 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9151 BUG_ON(sched_group_nodes_bycpu == NULL);
9154 mutex_lock(&sched_domains_mutex);
9155 arch_init_sched_domains(cpu_online_mask);
9156 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9157 if (cpumask_empty(non_isolated_cpus))
9158 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9159 mutex_unlock(&sched_domains_mutex);
9162 #ifndef CONFIG_CPUSETS
9163 /* XXX: Theoretical race here - CPU may be hotplugged now */
9164 hotcpu_notifier(update_sched_domains, 0);
9167 /* RT runtime code needs to handle some hotplug events */
9168 hotcpu_notifier(update_runtime, 0);
9172 /* Move init over to a non-isolated CPU */
9173 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9175 sched_init_granularity();
9176 free_cpumask_var(non_isolated_cpus);
9178 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9179 init_sched_rt_class();
9182 void __init sched_init_smp(void)
9184 sched_init_granularity();
9186 #endif /* CONFIG_SMP */
9188 const_debug unsigned int sysctl_timer_migration = 1;
9190 int in_sched_functions(unsigned long addr)
9192 return in_lock_functions(addr) ||
9193 (addr >= (unsigned long)__sched_text_start
9194 && addr < (unsigned long)__sched_text_end);
9197 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9199 cfs_rq->tasks_timeline = RB_ROOT;
9200 INIT_LIST_HEAD(&cfs_rq->tasks);
9201 #ifdef CONFIG_FAIR_GROUP_SCHED
9204 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9207 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9209 struct rt_prio_array *array;
9212 array = &rt_rq->active;
9213 for (i = 0; i < MAX_RT_PRIO; i++) {
9214 INIT_LIST_HEAD(array->queue + i);
9215 __clear_bit(i, array->bitmap);
9217 /* delimiter for bitsearch: */
9218 __set_bit(MAX_RT_PRIO, array->bitmap);
9220 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9221 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9223 rt_rq->highest_prio.next = MAX_RT_PRIO;
9227 rt_rq->rt_nr_migratory = 0;
9228 rt_rq->overloaded = 0;
9229 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9233 rt_rq->rt_throttled = 0;
9234 rt_rq->rt_runtime = 0;
9235 spin_lock_init(&rt_rq->rt_runtime_lock);
9237 #ifdef CONFIG_RT_GROUP_SCHED
9238 rt_rq->rt_nr_boosted = 0;
9243 #ifdef CONFIG_FAIR_GROUP_SCHED
9244 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9245 struct sched_entity *se, int cpu, int add,
9246 struct sched_entity *parent)
9248 struct rq *rq = cpu_rq(cpu);
9249 tg->cfs_rq[cpu] = cfs_rq;
9250 init_cfs_rq(cfs_rq, rq);
9253 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9256 /* se could be NULL for init_task_group */
9261 se->cfs_rq = &rq->cfs;
9263 se->cfs_rq = parent->my_q;
9266 se->load.weight = tg->shares;
9267 se->load.inv_weight = 0;
9268 se->parent = parent;
9272 #ifdef CONFIG_RT_GROUP_SCHED
9273 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9274 struct sched_rt_entity *rt_se, int cpu, int add,
9275 struct sched_rt_entity *parent)
9277 struct rq *rq = cpu_rq(cpu);
9279 tg->rt_rq[cpu] = rt_rq;
9280 init_rt_rq(rt_rq, rq);
9282 rt_rq->rt_se = rt_se;
9283 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9285 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9287 tg->rt_se[cpu] = rt_se;
9292 rt_se->rt_rq = &rq->rt;
9294 rt_se->rt_rq = parent->my_q;
9296 rt_se->my_q = rt_rq;
9297 rt_se->parent = parent;
9298 INIT_LIST_HEAD(&rt_se->run_list);
9302 void __init sched_init(void)
9305 unsigned long alloc_size = 0, ptr;
9307 #ifdef CONFIG_FAIR_GROUP_SCHED
9308 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9310 #ifdef CONFIG_RT_GROUP_SCHED
9311 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9313 #ifdef CONFIG_USER_SCHED
9316 #ifdef CONFIG_CPUMASK_OFFSTACK
9317 alloc_size += num_possible_cpus() * cpumask_size();
9320 * As sched_init() is called before page_alloc is setup,
9321 * we use alloc_bootmem().
9324 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9326 #ifdef CONFIG_FAIR_GROUP_SCHED
9327 init_task_group.se = (struct sched_entity **)ptr;
9328 ptr += nr_cpu_ids * sizeof(void **);
9330 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9331 ptr += nr_cpu_ids * sizeof(void **);
9333 #ifdef CONFIG_USER_SCHED
9334 root_task_group.se = (struct sched_entity **)ptr;
9335 ptr += nr_cpu_ids * sizeof(void **);
9337 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9338 ptr += nr_cpu_ids * sizeof(void **);
9339 #endif /* CONFIG_USER_SCHED */
9340 #endif /* CONFIG_FAIR_GROUP_SCHED */
9341 #ifdef CONFIG_RT_GROUP_SCHED
9342 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9343 ptr += nr_cpu_ids * sizeof(void **);
9345 init_task_group.rt_rq = (struct rt_rq **)ptr;
9346 ptr += nr_cpu_ids * sizeof(void **);
9348 #ifdef CONFIG_USER_SCHED
9349 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9350 ptr += nr_cpu_ids * sizeof(void **);
9352 root_task_group.rt_rq = (struct rt_rq **)ptr;
9353 ptr += nr_cpu_ids * sizeof(void **);
9354 #endif /* CONFIG_USER_SCHED */
9355 #endif /* CONFIG_RT_GROUP_SCHED */
9356 #ifdef CONFIG_CPUMASK_OFFSTACK
9357 for_each_possible_cpu(i) {
9358 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9359 ptr += cpumask_size();
9361 #endif /* CONFIG_CPUMASK_OFFSTACK */
9365 init_defrootdomain();
9368 init_rt_bandwidth(&def_rt_bandwidth,
9369 global_rt_period(), global_rt_runtime());
9371 #ifdef CONFIG_RT_GROUP_SCHED
9372 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9373 global_rt_period(), global_rt_runtime());
9374 #ifdef CONFIG_USER_SCHED
9375 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9376 global_rt_period(), RUNTIME_INF);
9377 #endif /* CONFIG_USER_SCHED */
9378 #endif /* CONFIG_RT_GROUP_SCHED */
9380 #ifdef CONFIG_GROUP_SCHED
9381 list_add(&init_task_group.list, &task_groups);
9382 INIT_LIST_HEAD(&init_task_group.children);
9384 #ifdef CONFIG_USER_SCHED
9385 INIT_LIST_HEAD(&root_task_group.children);
9386 init_task_group.parent = &root_task_group;
9387 list_add(&init_task_group.siblings, &root_task_group.children);
9388 #endif /* CONFIG_USER_SCHED */
9389 #endif /* CONFIG_GROUP_SCHED */
9391 for_each_possible_cpu(i) {
9395 spin_lock_init(&rq->lock);
9397 rq->calc_load_active = 0;
9398 rq->calc_load_update = jiffies + LOAD_FREQ;
9399 init_cfs_rq(&rq->cfs, rq);
9400 init_rt_rq(&rq->rt, rq);
9401 #ifdef CONFIG_FAIR_GROUP_SCHED
9402 init_task_group.shares = init_task_group_load;
9403 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9404 #ifdef CONFIG_CGROUP_SCHED
9406 * How much cpu bandwidth does init_task_group get?
9408 * In case of task-groups formed thr' the cgroup filesystem, it
9409 * gets 100% of the cpu resources in the system. This overall
9410 * system cpu resource is divided among the tasks of
9411 * init_task_group and its child task-groups in a fair manner,
9412 * based on each entity's (task or task-group's) weight
9413 * (se->load.weight).
9415 * In other words, if init_task_group has 10 tasks of weight
9416 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9417 * then A0's share of the cpu resource is:
9419 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9421 * We achieve this by letting init_task_group's tasks sit
9422 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9424 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9425 #elif defined CONFIG_USER_SCHED
9426 root_task_group.shares = NICE_0_LOAD;
9427 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9429 * In case of task-groups formed thr' the user id of tasks,
9430 * init_task_group represents tasks belonging to root user.
9431 * Hence it forms a sibling of all subsequent groups formed.
9432 * In this case, init_task_group gets only a fraction of overall
9433 * system cpu resource, based on the weight assigned to root
9434 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9435 * by letting tasks of init_task_group sit in a separate cfs_rq
9436 * (init_tg_cfs_rq) and having one entity represent this group of
9437 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9439 init_tg_cfs_entry(&init_task_group,
9440 &per_cpu(init_tg_cfs_rq, i),
9441 &per_cpu(init_sched_entity, i), i, 1,
9442 root_task_group.se[i]);
9445 #endif /* CONFIG_FAIR_GROUP_SCHED */
9447 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9448 #ifdef CONFIG_RT_GROUP_SCHED
9449 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9450 #ifdef CONFIG_CGROUP_SCHED
9451 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9452 #elif defined CONFIG_USER_SCHED
9453 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9454 init_tg_rt_entry(&init_task_group,
9455 &per_cpu(init_rt_rq, i),
9456 &per_cpu(init_sched_rt_entity, i), i, 1,
9457 root_task_group.rt_se[i]);
9461 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9462 rq->cpu_load[j] = 0;
9466 rq->post_schedule = 0;
9467 rq->active_balance = 0;
9468 rq->next_balance = jiffies;
9472 rq->migration_thread = NULL;
9473 INIT_LIST_HEAD(&rq->migration_queue);
9474 rq_attach_root(rq, &def_root_domain);
9477 atomic_set(&rq->nr_iowait, 0);
9480 set_load_weight(&init_task);
9482 #ifdef CONFIG_PREEMPT_NOTIFIERS
9483 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9487 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9490 #ifdef CONFIG_RT_MUTEXES
9491 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9495 * The boot idle thread does lazy MMU switching as well:
9497 atomic_inc(&init_mm.mm_count);
9498 enter_lazy_tlb(&init_mm, current);
9501 * Make us the idle thread. Technically, schedule() should not be
9502 * called from this thread, however somewhere below it might be,
9503 * but because we are the idle thread, we just pick up running again
9504 * when this runqueue becomes "idle".
9506 init_idle(current, smp_processor_id());
9508 calc_load_update = jiffies + LOAD_FREQ;
9511 * During early bootup we pretend to be a normal task:
9513 current->sched_class = &fair_sched_class;
9515 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9516 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9519 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9520 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9522 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9525 perf_counter_init();
9527 scheduler_running = 1;
9530 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9531 static inline int preempt_count_equals(int preempt_offset)
9533 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9535 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9538 void __might_sleep(char *file, int line, int preempt_offset)
9541 static unsigned long prev_jiffy; /* ratelimiting */
9543 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9544 system_state != SYSTEM_RUNNING || oops_in_progress)
9546 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9548 prev_jiffy = jiffies;
9551 "BUG: sleeping function called from invalid context at %s:%d\n",
9554 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9555 in_atomic(), irqs_disabled(),
9556 current->pid, current->comm);
9558 debug_show_held_locks(current);
9559 if (irqs_disabled())
9560 print_irqtrace_events(current);
9564 EXPORT_SYMBOL(__might_sleep);
9567 #ifdef CONFIG_MAGIC_SYSRQ
9568 static void normalize_task(struct rq *rq, struct task_struct *p)
9572 update_rq_clock(rq);
9573 on_rq = p->se.on_rq;
9575 deactivate_task(rq, p, 0);
9576 __setscheduler(rq, p, SCHED_NORMAL, 0);
9578 activate_task(rq, p, 0);
9579 resched_task(rq->curr);
9583 void normalize_rt_tasks(void)
9585 struct task_struct *g, *p;
9586 unsigned long flags;
9589 read_lock_irqsave(&tasklist_lock, flags);
9590 do_each_thread(g, p) {
9592 * Only normalize user tasks:
9597 p->se.exec_start = 0;
9598 #ifdef CONFIG_SCHEDSTATS
9599 p->se.wait_start = 0;
9600 p->se.sleep_start = 0;
9601 p->se.block_start = 0;
9606 * Renice negative nice level userspace
9609 if (TASK_NICE(p) < 0 && p->mm)
9610 set_user_nice(p, 0);
9614 spin_lock(&p->pi_lock);
9615 rq = __task_rq_lock(p);
9617 normalize_task(rq, p);
9619 __task_rq_unlock(rq);
9620 spin_unlock(&p->pi_lock);
9621 } while_each_thread(g, p);
9623 read_unlock_irqrestore(&tasklist_lock, flags);
9626 #endif /* CONFIG_MAGIC_SYSRQ */
9630 * These functions are only useful for the IA64 MCA handling.
9632 * They can only be called when the whole system has been
9633 * stopped - every CPU needs to be quiescent, and no scheduling
9634 * activity can take place. Using them for anything else would
9635 * be a serious bug, and as a result, they aren't even visible
9636 * under any other configuration.
9640 * curr_task - return the current task for a given cpu.
9641 * @cpu: the processor in question.
9643 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9645 struct task_struct *curr_task(int cpu)
9647 return cpu_curr(cpu);
9651 * set_curr_task - set the current task for a given cpu.
9652 * @cpu: the processor in question.
9653 * @p: the task pointer to set.
9655 * Description: This function must only be used when non-maskable interrupts
9656 * are serviced on a separate stack. It allows the architecture to switch the
9657 * notion of the current task on a cpu in a non-blocking manner. This function
9658 * must be called with all CPU's synchronized, and interrupts disabled, the
9659 * and caller must save the original value of the current task (see
9660 * curr_task() above) and restore that value before reenabling interrupts and
9661 * re-starting the system.
9663 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9665 void set_curr_task(int cpu, struct task_struct *p)
9672 #ifdef CONFIG_FAIR_GROUP_SCHED
9673 static void free_fair_sched_group(struct task_group *tg)
9677 for_each_possible_cpu(i) {
9679 kfree(tg->cfs_rq[i]);
9689 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9691 struct cfs_rq *cfs_rq;
9692 struct sched_entity *se;
9696 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9699 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9703 tg->shares = NICE_0_LOAD;
9705 for_each_possible_cpu(i) {
9708 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9709 GFP_KERNEL, cpu_to_node(i));
9713 se = kzalloc_node(sizeof(struct sched_entity),
9714 GFP_KERNEL, cpu_to_node(i));
9718 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9727 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9729 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9730 &cpu_rq(cpu)->leaf_cfs_rq_list);
9733 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9735 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9737 #else /* !CONFG_FAIR_GROUP_SCHED */
9738 static inline void free_fair_sched_group(struct task_group *tg)
9743 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9748 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9752 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9755 #endif /* CONFIG_FAIR_GROUP_SCHED */
9757 #ifdef CONFIG_RT_GROUP_SCHED
9758 static void free_rt_sched_group(struct task_group *tg)
9762 destroy_rt_bandwidth(&tg->rt_bandwidth);
9764 for_each_possible_cpu(i) {
9766 kfree(tg->rt_rq[i]);
9768 kfree(tg->rt_se[i]);
9776 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9778 struct rt_rq *rt_rq;
9779 struct sched_rt_entity *rt_se;
9783 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9786 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9790 init_rt_bandwidth(&tg->rt_bandwidth,
9791 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9793 for_each_possible_cpu(i) {
9796 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9797 GFP_KERNEL, cpu_to_node(i));
9801 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9802 GFP_KERNEL, cpu_to_node(i));
9806 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9815 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9817 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9818 &cpu_rq(cpu)->leaf_rt_rq_list);
9821 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9823 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9825 #else /* !CONFIG_RT_GROUP_SCHED */
9826 static inline void free_rt_sched_group(struct task_group *tg)
9831 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9836 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9840 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9843 #endif /* CONFIG_RT_GROUP_SCHED */
9845 #ifdef CONFIG_GROUP_SCHED
9846 static void free_sched_group(struct task_group *tg)
9848 free_fair_sched_group(tg);
9849 free_rt_sched_group(tg);
9853 /* allocate runqueue etc for a new task group */
9854 struct task_group *sched_create_group(struct task_group *parent)
9856 struct task_group *tg;
9857 unsigned long flags;
9860 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9862 return ERR_PTR(-ENOMEM);
9864 if (!alloc_fair_sched_group(tg, parent))
9867 if (!alloc_rt_sched_group(tg, parent))
9870 spin_lock_irqsave(&task_group_lock, flags);
9871 for_each_possible_cpu(i) {
9872 register_fair_sched_group(tg, i);
9873 register_rt_sched_group(tg, i);
9875 list_add_rcu(&tg->list, &task_groups);
9877 WARN_ON(!parent); /* root should already exist */
9879 tg->parent = parent;
9880 INIT_LIST_HEAD(&tg->children);
9881 list_add_rcu(&tg->siblings, &parent->children);
9882 spin_unlock_irqrestore(&task_group_lock, flags);
9887 free_sched_group(tg);
9888 return ERR_PTR(-ENOMEM);
9891 /* rcu callback to free various structures associated with a task group */
9892 static void free_sched_group_rcu(struct rcu_head *rhp)
9894 /* now it should be safe to free those cfs_rqs */
9895 free_sched_group(container_of(rhp, struct task_group, rcu));
9898 /* Destroy runqueue etc associated with a task group */
9899 void sched_destroy_group(struct task_group *tg)
9901 unsigned long flags;
9904 spin_lock_irqsave(&task_group_lock, flags);
9905 for_each_possible_cpu(i) {
9906 unregister_fair_sched_group(tg, i);
9907 unregister_rt_sched_group(tg, i);
9909 list_del_rcu(&tg->list);
9910 list_del_rcu(&tg->siblings);
9911 spin_unlock_irqrestore(&task_group_lock, flags);
9913 /* wait for possible concurrent references to cfs_rqs complete */
9914 call_rcu(&tg->rcu, free_sched_group_rcu);
9917 /* change task's runqueue when it moves between groups.
9918 * The caller of this function should have put the task in its new group
9919 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9920 * reflect its new group.
9922 void sched_move_task(struct task_struct *tsk)
9925 unsigned long flags;
9928 rq = task_rq_lock(tsk, &flags);
9930 update_rq_clock(rq);
9932 running = task_current(rq, tsk);
9933 on_rq = tsk->se.on_rq;
9936 dequeue_task(rq, tsk, 0);
9937 if (unlikely(running))
9938 tsk->sched_class->put_prev_task(rq, tsk);
9940 set_task_rq(tsk, task_cpu(tsk));
9942 #ifdef CONFIG_FAIR_GROUP_SCHED
9943 if (tsk->sched_class->moved_group)
9944 tsk->sched_class->moved_group(tsk);
9947 if (unlikely(running))
9948 tsk->sched_class->set_curr_task(rq);
9950 enqueue_task(rq, tsk, 0);
9952 task_rq_unlock(rq, &flags);
9954 #endif /* CONFIG_GROUP_SCHED */
9956 #ifdef CONFIG_FAIR_GROUP_SCHED
9957 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9959 struct cfs_rq *cfs_rq = se->cfs_rq;
9964 dequeue_entity(cfs_rq, se, 0);
9966 se->load.weight = shares;
9967 se->load.inv_weight = 0;
9970 enqueue_entity(cfs_rq, se, 0);
9973 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9975 struct cfs_rq *cfs_rq = se->cfs_rq;
9976 struct rq *rq = cfs_rq->rq;
9977 unsigned long flags;
9979 spin_lock_irqsave(&rq->lock, flags);
9980 __set_se_shares(se, shares);
9981 spin_unlock_irqrestore(&rq->lock, flags);
9984 static DEFINE_MUTEX(shares_mutex);
9986 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9989 unsigned long flags;
9992 * We can't change the weight of the root cgroup.
9997 if (shares < MIN_SHARES)
9998 shares = MIN_SHARES;
9999 else if (shares > MAX_SHARES)
10000 shares = MAX_SHARES;
10002 mutex_lock(&shares_mutex);
10003 if (tg->shares == shares)
10006 spin_lock_irqsave(&task_group_lock, flags);
10007 for_each_possible_cpu(i)
10008 unregister_fair_sched_group(tg, i);
10009 list_del_rcu(&tg->siblings);
10010 spin_unlock_irqrestore(&task_group_lock, flags);
10012 /* wait for any ongoing reference to this group to finish */
10013 synchronize_sched();
10016 * Now we are free to modify the group's share on each cpu
10017 * w/o tripping rebalance_share or load_balance_fair.
10019 tg->shares = shares;
10020 for_each_possible_cpu(i) {
10022 * force a rebalance
10024 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10025 set_se_shares(tg->se[i], shares);
10029 * Enable load balance activity on this group, by inserting it back on
10030 * each cpu's rq->leaf_cfs_rq_list.
10032 spin_lock_irqsave(&task_group_lock, flags);
10033 for_each_possible_cpu(i)
10034 register_fair_sched_group(tg, i);
10035 list_add_rcu(&tg->siblings, &tg->parent->children);
10036 spin_unlock_irqrestore(&task_group_lock, flags);
10038 mutex_unlock(&shares_mutex);
10042 unsigned long sched_group_shares(struct task_group *tg)
10048 #ifdef CONFIG_RT_GROUP_SCHED
10050 * Ensure that the real time constraints are schedulable.
10052 static DEFINE_MUTEX(rt_constraints_mutex);
10054 static unsigned long to_ratio(u64 period, u64 runtime)
10056 if (runtime == RUNTIME_INF)
10059 return div64_u64(runtime << 20, period);
10062 /* Must be called with tasklist_lock held */
10063 static inline int tg_has_rt_tasks(struct task_group *tg)
10065 struct task_struct *g, *p;
10067 do_each_thread(g, p) {
10068 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10070 } while_each_thread(g, p);
10075 struct rt_schedulable_data {
10076 struct task_group *tg;
10081 static int tg_schedulable(struct task_group *tg, void *data)
10083 struct rt_schedulable_data *d = data;
10084 struct task_group *child;
10085 unsigned long total, sum = 0;
10086 u64 period, runtime;
10088 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10089 runtime = tg->rt_bandwidth.rt_runtime;
10092 period = d->rt_period;
10093 runtime = d->rt_runtime;
10096 #ifdef CONFIG_USER_SCHED
10097 if (tg == &root_task_group) {
10098 period = global_rt_period();
10099 runtime = global_rt_runtime();
10104 * Cannot have more runtime than the period.
10106 if (runtime > period && runtime != RUNTIME_INF)
10110 * Ensure we don't starve existing RT tasks.
10112 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10115 total = to_ratio(period, runtime);
10118 * Nobody can have more than the global setting allows.
10120 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10124 * The sum of our children's runtime should not exceed our own.
10126 list_for_each_entry_rcu(child, &tg->children, siblings) {
10127 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10128 runtime = child->rt_bandwidth.rt_runtime;
10130 if (child == d->tg) {
10131 period = d->rt_period;
10132 runtime = d->rt_runtime;
10135 sum += to_ratio(period, runtime);
10144 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10146 struct rt_schedulable_data data = {
10148 .rt_period = period,
10149 .rt_runtime = runtime,
10152 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10155 static int tg_set_bandwidth(struct task_group *tg,
10156 u64 rt_period, u64 rt_runtime)
10160 mutex_lock(&rt_constraints_mutex);
10161 read_lock(&tasklist_lock);
10162 err = __rt_schedulable(tg, rt_period, rt_runtime);
10166 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10167 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10168 tg->rt_bandwidth.rt_runtime = rt_runtime;
10170 for_each_possible_cpu(i) {
10171 struct rt_rq *rt_rq = tg->rt_rq[i];
10173 spin_lock(&rt_rq->rt_runtime_lock);
10174 rt_rq->rt_runtime = rt_runtime;
10175 spin_unlock(&rt_rq->rt_runtime_lock);
10177 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10179 read_unlock(&tasklist_lock);
10180 mutex_unlock(&rt_constraints_mutex);
10185 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10187 u64 rt_runtime, rt_period;
10189 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10190 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10191 if (rt_runtime_us < 0)
10192 rt_runtime = RUNTIME_INF;
10194 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10197 long sched_group_rt_runtime(struct task_group *tg)
10201 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10204 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10205 do_div(rt_runtime_us, NSEC_PER_USEC);
10206 return rt_runtime_us;
10209 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10211 u64 rt_runtime, rt_period;
10213 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10214 rt_runtime = tg->rt_bandwidth.rt_runtime;
10216 if (rt_period == 0)
10219 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10222 long sched_group_rt_period(struct task_group *tg)
10226 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10227 do_div(rt_period_us, NSEC_PER_USEC);
10228 return rt_period_us;
10231 static int sched_rt_global_constraints(void)
10233 u64 runtime, period;
10236 if (sysctl_sched_rt_period <= 0)
10239 runtime = global_rt_runtime();
10240 period = global_rt_period();
10243 * Sanity check on the sysctl variables.
10245 if (runtime > period && runtime != RUNTIME_INF)
10248 mutex_lock(&rt_constraints_mutex);
10249 read_lock(&tasklist_lock);
10250 ret = __rt_schedulable(NULL, 0, 0);
10251 read_unlock(&tasklist_lock);
10252 mutex_unlock(&rt_constraints_mutex);
10257 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10259 /* Don't accept realtime tasks when there is no way for them to run */
10260 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10266 #else /* !CONFIG_RT_GROUP_SCHED */
10267 static int sched_rt_global_constraints(void)
10269 unsigned long flags;
10272 if (sysctl_sched_rt_period <= 0)
10276 * There's always some RT tasks in the root group
10277 * -- migration, kstopmachine etc..
10279 if (sysctl_sched_rt_runtime == 0)
10282 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10283 for_each_possible_cpu(i) {
10284 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10286 spin_lock(&rt_rq->rt_runtime_lock);
10287 rt_rq->rt_runtime = global_rt_runtime();
10288 spin_unlock(&rt_rq->rt_runtime_lock);
10290 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10294 #endif /* CONFIG_RT_GROUP_SCHED */
10296 int sched_rt_handler(struct ctl_table *table, int write,
10297 struct file *filp, void __user *buffer, size_t *lenp,
10301 int old_period, old_runtime;
10302 static DEFINE_MUTEX(mutex);
10304 mutex_lock(&mutex);
10305 old_period = sysctl_sched_rt_period;
10306 old_runtime = sysctl_sched_rt_runtime;
10308 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
10310 if (!ret && write) {
10311 ret = sched_rt_global_constraints();
10313 sysctl_sched_rt_period = old_period;
10314 sysctl_sched_rt_runtime = old_runtime;
10316 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10317 def_rt_bandwidth.rt_period =
10318 ns_to_ktime(global_rt_period());
10321 mutex_unlock(&mutex);
10326 #ifdef CONFIG_CGROUP_SCHED
10328 /* return corresponding task_group object of a cgroup */
10329 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10331 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10332 struct task_group, css);
10335 static struct cgroup_subsys_state *
10336 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10338 struct task_group *tg, *parent;
10340 if (!cgrp->parent) {
10341 /* This is early initialization for the top cgroup */
10342 return &init_task_group.css;
10345 parent = cgroup_tg(cgrp->parent);
10346 tg = sched_create_group(parent);
10348 return ERR_PTR(-ENOMEM);
10354 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10356 struct task_group *tg = cgroup_tg(cgrp);
10358 sched_destroy_group(tg);
10362 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10363 struct task_struct *tsk)
10365 #ifdef CONFIG_RT_GROUP_SCHED
10366 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10369 /* We don't support RT-tasks being in separate groups */
10370 if (tsk->sched_class != &fair_sched_class)
10378 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10379 struct cgroup *old_cont, struct task_struct *tsk)
10381 sched_move_task(tsk);
10384 #ifdef CONFIG_FAIR_GROUP_SCHED
10385 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10388 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10391 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10393 struct task_group *tg = cgroup_tg(cgrp);
10395 return (u64) tg->shares;
10397 #endif /* CONFIG_FAIR_GROUP_SCHED */
10399 #ifdef CONFIG_RT_GROUP_SCHED
10400 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10403 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10406 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10408 return sched_group_rt_runtime(cgroup_tg(cgrp));
10411 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10414 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10417 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10419 return sched_group_rt_period(cgroup_tg(cgrp));
10421 #endif /* CONFIG_RT_GROUP_SCHED */
10423 static struct cftype cpu_files[] = {
10424 #ifdef CONFIG_FAIR_GROUP_SCHED
10427 .read_u64 = cpu_shares_read_u64,
10428 .write_u64 = cpu_shares_write_u64,
10431 #ifdef CONFIG_RT_GROUP_SCHED
10433 .name = "rt_runtime_us",
10434 .read_s64 = cpu_rt_runtime_read,
10435 .write_s64 = cpu_rt_runtime_write,
10438 .name = "rt_period_us",
10439 .read_u64 = cpu_rt_period_read_uint,
10440 .write_u64 = cpu_rt_period_write_uint,
10445 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10447 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10450 struct cgroup_subsys cpu_cgroup_subsys = {
10452 .create = cpu_cgroup_create,
10453 .destroy = cpu_cgroup_destroy,
10454 .can_attach = cpu_cgroup_can_attach,
10455 .attach = cpu_cgroup_attach,
10456 .populate = cpu_cgroup_populate,
10457 .subsys_id = cpu_cgroup_subsys_id,
10461 #endif /* CONFIG_CGROUP_SCHED */
10463 #ifdef CONFIG_CGROUP_CPUACCT
10466 * CPU accounting code for task groups.
10468 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10469 * (balbir@in.ibm.com).
10472 /* track cpu usage of a group of tasks and its child groups */
10474 struct cgroup_subsys_state css;
10475 /* cpuusage holds pointer to a u64-type object on every cpu */
10477 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10478 struct cpuacct *parent;
10481 struct cgroup_subsys cpuacct_subsys;
10483 /* return cpu accounting group corresponding to this container */
10484 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10486 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10487 struct cpuacct, css);
10490 /* return cpu accounting group to which this task belongs */
10491 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10493 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10494 struct cpuacct, css);
10497 /* create a new cpu accounting group */
10498 static struct cgroup_subsys_state *cpuacct_create(
10499 struct cgroup_subsys *ss, struct cgroup *cgrp)
10501 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10507 ca->cpuusage = alloc_percpu(u64);
10511 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10512 if (percpu_counter_init(&ca->cpustat[i], 0))
10513 goto out_free_counters;
10516 ca->parent = cgroup_ca(cgrp->parent);
10522 percpu_counter_destroy(&ca->cpustat[i]);
10523 free_percpu(ca->cpuusage);
10527 return ERR_PTR(-ENOMEM);
10530 /* destroy an existing cpu accounting group */
10532 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10534 struct cpuacct *ca = cgroup_ca(cgrp);
10537 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10538 percpu_counter_destroy(&ca->cpustat[i]);
10539 free_percpu(ca->cpuusage);
10543 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10545 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10548 #ifndef CONFIG_64BIT
10550 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10552 spin_lock_irq(&cpu_rq(cpu)->lock);
10554 spin_unlock_irq(&cpu_rq(cpu)->lock);
10562 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10564 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10566 #ifndef CONFIG_64BIT
10568 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10570 spin_lock_irq(&cpu_rq(cpu)->lock);
10572 spin_unlock_irq(&cpu_rq(cpu)->lock);
10578 /* return total cpu usage (in nanoseconds) of a group */
10579 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10581 struct cpuacct *ca = cgroup_ca(cgrp);
10582 u64 totalcpuusage = 0;
10585 for_each_present_cpu(i)
10586 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10588 return totalcpuusage;
10591 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10594 struct cpuacct *ca = cgroup_ca(cgrp);
10603 for_each_present_cpu(i)
10604 cpuacct_cpuusage_write(ca, i, 0);
10610 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10611 struct seq_file *m)
10613 struct cpuacct *ca = cgroup_ca(cgroup);
10617 for_each_present_cpu(i) {
10618 percpu = cpuacct_cpuusage_read(ca, i);
10619 seq_printf(m, "%llu ", (unsigned long long) percpu);
10621 seq_printf(m, "\n");
10625 static const char *cpuacct_stat_desc[] = {
10626 [CPUACCT_STAT_USER] = "user",
10627 [CPUACCT_STAT_SYSTEM] = "system",
10630 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10631 struct cgroup_map_cb *cb)
10633 struct cpuacct *ca = cgroup_ca(cgrp);
10636 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10637 s64 val = percpu_counter_read(&ca->cpustat[i]);
10638 val = cputime64_to_clock_t(val);
10639 cb->fill(cb, cpuacct_stat_desc[i], val);
10644 static struct cftype files[] = {
10647 .read_u64 = cpuusage_read,
10648 .write_u64 = cpuusage_write,
10651 .name = "usage_percpu",
10652 .read_seq_string = cpuacct_percpu_seq_read,
10656 .read_map = cpuacct_stats_show,
10660 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10662 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10666 * charge this task's execution time to its accounting group.
10668 * called with rq->lock held.
10670 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10672 struct cpuacct *ca;
10675 if (unlikely(!cpuacct_subsys.active))
10678 cpu = task_cpu(tsk);
10684 for (; ca; ca = ca->parent) {
10685 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10686 *cpuusage += cputime;
10693 * Charge the system/user time to the task's accounting group.
10695 static void cpuacct_update_stats(struct task_struct *tsk,
10696 enum cpuacct_stat_index idx, cputime_t val)
10698 struct cpuacct *ca;
10700 if (unlikely(!cpuacct_subsys.active))
10707 percpu_counter_add(&ca->cpustat[idx], val);
10713 struct cgroup_subsys cpuacct_subsys = {
10715 .create = cpuacct_create,
10716 .destroy = cpuacct_destroy,
10717 .populate = cpuacct_populate,
10718 .subsys_id = cpuacct_subsys_id,
10720 #endif /* CONFIG_CGROUP_CPUACCT */
10724 int rcu_expedited_torture_stats(char *page)
10728 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10730 void synchronize_sched_expedited(void)
10733 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10735 #else /* #ifndef CONFIG_SMP */
10737 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10738 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10740 #define RCU_EXPEDITED_STATE_POST -2
10741 #define RCU_EXPEDITED_STATE_IDLE -1
10743 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10745 int rcu_expedited_torture_stats(char *page)
10750 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10751 for_each_online_cpu(cpu) {
10752 cnt += sprintf(&page[cnt], " %d:%d",
10753 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10755 cnt += sprintf(&page[cnt], "\n");
10758 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10760 static long synchronize_sched_expedited_count;
10763 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10764 * approach to force grace period to end quickly. This consumes
10765 * significant time on all CPUs, and is thus not recommended for
10766 * any sort of common-case code.
10768 * Note that it is illegal to call this function while holding any
10769 * lock that is acquired by a CPU-hotplug notifier. Failing to
10770 * observe this restriction will result in deadlock.
10772 void synchronize_sched_expedited(void)
10775 unsigned long flags;
10776 bool need_full_sync = 0;
10778 struct migration_req *req;
10782 smp_mb(); /* ensure prior mod happens before capturing snap. */
10783 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10785 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10787 if (trycount++ < 10)
10788 udelay(trycount * num_online_cpus());
10790 synchronize_sched();
10793 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10794 smp_mb(); /* ensure test happens before caller kfree */
10799 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10800 for_each_online_cpu(cpu) {
10802 req = &per_cpu(rcu_migration_req, cpu);
10803 init_completion(&req->done);
10805 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10806 spin_lock_irqsave(&rq->lock, flags);
10807 list_add(&req->list, &rq->migration_queue);
10808 spin_unlock_irqrestore(&rq->lock, flags);
10809 wake_up_process(rq->migration_thread);
10811 for_each_online_cpu(cpu) {
10812 rcu_expedited_state = cpu;
10813 req = &per_cpu(rcu_migration_req, cpu);
10815 wait_for_completion(&req->done);
10816 spin_lock_irqsave(&rq->lock, flags);
10817 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10818 need_full_sync = 1;
10819 req->dest_cpu = RCU_MIGRATION_IDLE;
10820 spin_unlock_irqrestore(&rq->lock, flags);
10822 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10823 mutex_unlock(&rcu_sched_expedited_mutex);
10825 if (need_full_sync)
10826 synchronize_sched();
10828 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10830 #endif /* #else #ifndef CONFIG_SMP */