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 */
3556 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3558 return SCHED_LOAD_SCALE;
3561 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3563 return default_scale_freq_power(sd, cpu);
3566 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3568 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3569 unsigned long smt_gain = sd->smt_gain;
3576 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3578 return default_scale_smt_power(sd, cpu);
3581 unsigned long scale_rt_power(int cpu)
3583 struct rq *rq = cpu_rq(cpu);
3584 u64 total, available;
3586 sched_avg_update(rq);
3588 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3589 available = total - rq->rt_avg;
3591 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3592 total = SCHED_LOAD_SCALE;
3594 total >>= SCHED_LOAD_SHIFT;
3596 return div_u64(available, total);
3599 static void update_cpu_power(struct sched_domain *sd, int cpu)
3601 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3602 unsigned long power = SCHED_LOAD_SCALE;
3603 struct sched_group *sdg = sd->groups;
3605 power *= arch_scale_freq_power(sd, cpu);
3606 power >>= SCHED_LOAD_SHIFT;
3608 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3609 power *= arch_scale_smt_power(sd, cpu);
3610 power >>= SCHED_LOAD_SHIFT;
3613 power *= scale_rt_power(cpu);
3614 power >>= SCHED_LOAD_SHIFT;
3619 sdg->cpu_power = power;
3622 static void update_group_power(struct sched_domain *sd, int cpu)
3624 struct sched_domain *child = sd->child;
3625 struct sched_group *group, *sdg = sd->groups;
3626 unsigned long power;
3629 update_cpu_power(sd, cpu);
3635 group = child->groups;
3637 power += group->cpu_power;
3638 group = group->next;
3639 } while (group != child->groups);
3641 sdg->cpu_power = power;
3645 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3646 * @group: sched_group whose statistics are to be updated.
3647 * @this_cpu: Cpu for which load balance is currently performed.
3648 * @idle: Idle status of this_cpu
3649 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3650 * @sd_idle: Idle status of the sched_domain containing group.
3651 * @local_group: Does group contain this_cpu.
3652 * @cpus: Set of cpus considered for load balancing.
3653 * @balance: Should we balance.
3654 * @sgs: variable to hold the statistics for this group.
3656 static inline void update_sg_lb_stats(struct sched_domain *sd,
3657 struct sched_group *group, int this_cpu,
3658 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3659 int local_group, const struct cpumask *cpus,
3660 int *balance, struct sg_lb_stats *sgs)
3662 unsigned long load, max_cpu_load, min_cpu_load;
3664 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3665 unsigned long sum_avg_load_per_task;
3666 unsigned long avg_load_per_task;
3669 balance_cpu = group_first_cpu(group);
3670 if (balance_cpu == this_cpu)
3671 update_group_power(sd, this_cpu);
3674 /* Tally up the load of all CPUs in the group */
3675 sum_avg_load_per_task = avg_load_per_task = 0;
3677 min_cpu_load = ~0UL;
3679 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3680 struct rq *rq = cpu_rq(i);
3682 if (*sd_idle && rq->nr_running)
3685 /* Bias balancing toward cpus of our domain */
3687 if (idle_cpu(i) && !first_idle_cpu) {
3692 load = target_load(i, load_idx);
3694 load = source_load(i, load_idx);
3695 if (load > max_cpu_load)
3696 max_cpu_load = load;
3697 if (min_cpu_load > load)
3698 min_cpu_load = load;
3701 sgs->group_load += load;
3702 sgs->sum_nr_running += rq->nr_running;
3703 sgs->sum_weighted_load += weighted_cpuload(i);
3705 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3709 * First idle cpu or the first cpu(busiest) in this sched group
3710 * is eligible for doing load balancing at this and above
3711 * domains. In the newly idle case, we will allow all the cpu's
3712 * to do the newly idle load balance.
3714 if (idle != CPU_NEWLY_IDLE && local_group &&
3715 balance_cpu != this_cpu && balance) {
3720 /* Adjust by relative CPU power of the group */
3721 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3725 * Consider the group unbalanced when the imbalance is larger
3726 * than the average weight of two tasks.
3728 * APZ: with cgroup the avg task weight can vary wildly and
3729 * might not be a suitable number - should we keep a
3730 * normalized nr_running number somewhere that negates
3733 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3736 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3739 sgs->group_capacity =
3740 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3744 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3745 * @sd: sched_domain whose statistics are to be updated.
3746 * @this_cpu: Cpu for which load balance is currently performed.
3747 * @idle: Idle status of this_cpu
3748 * @sd_idle: Idle status of the sched_domain containing group.
3749 * @cpus: Set of cpus considered for load balancing.
3750 * @balance: Should we balance.
3751 * @sds: variable to hold the statistics for this sched_domain.
3753 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3754 enum cpu_idle_type idle, int *sd_idle,
3755 const struct cpumask *cpus, int *balance,
3756 struct sd_lb_stats *sds)
3758 struct sched_domain *child = sd->child;
3759 struct sched_group *group = sd->groups;
3760 struct sg_lb_stats sgs;
3761 int load_idx, prefer_sibling = 0;
3763 if (child && child->flags & SD_PREFER_SIBLING)
3766 init_sd_power_savings_stats(sd, sds, idle);
3767 load_idx = get_sd_load_idx(sd, idle);
3772 local_group = cpumask_test_cpu(this_cpu,
3773 sched_group_cpus(group));
3774 memset(&sgs, 0, sizeof(sgs));
3775 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3776 local_group, cpus, balance, &sgs);
3778 if (local_group && balance && !(*balance))
3781 sds->total_load += sgs.group_load;
3782 sds->total_pwr += group->cpu_power;
3785 * In case the child domain prefers tasks go to siblings
3786 * first, lower the group capacity to one so that we'll try
3787 * and move all the excess tasks away.
3790 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3793 sds->this_load = sgs.avg_load;
3795 sds->this_nr_running = sgs.sum_nr_running;
3796 sds->this_load_per_task = sgs.sum_weighted_load;
3797 } else if (sgs.avg_load > sds->max_load &&
3798 (sgs.sum_nr_running > sgs.group_capacity ||
3800 sds->max_load = sgs.avg_load;
3801 sds->busiest = group;
3802 sds->busiest_nr_running = sgs.sum_nr_running;
3803 sds->busiest_load_per_task = sgs.sum_weighted_load;
3804 sds->group_imb = sgs.group_imb;
3807 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3808 group = group->next;
3809 } while (group != sd->groups);
3813 * fix_small_imbalance - Calculate the minor imbalance that exists
3814 * amongst the groups of a sched_domain, during
3816 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3817 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3818 * @imbalance: Variable to store the imbalance.
3820 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3821 int this_cpu, unsigned long *imbalance)
3823 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3824 unsigned int imbn = 2;
3826 if (sds->this_nr_running) {
3827 sds->this_load_per_task /= sds->this_nr_running;
3828 if (sds->busiest_load_per_task >
3829 sds->this_load_per_task)
3832 sds->this_load_per_task =
3833 cpu_avg_load_per_task(this_cpu);
3835 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3836 sds->busiest_load_per_task * imbn) {
3837 *imbalance = sds->busiest_load_per_task;
3842 * OK, we don't have enough imbalance to justify moving tasks,
3843 * however we may be able to increase total CPU power used by
3847 pwr_now += sds->busiest->cpu_power *
3848 min(sds->busiest_load_per_task, sds->max_load);
3849 pwr_now += sds->this->cpu_power *
3850 min(sds->this_load_per_task, sds->this_load);
3851 pwr_now /= SCHED_LOAD_SCALE;
3853 /* Amount of load we'd subtract */
3854 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3855 sds->busiest->cpu_power;
3856 if (sds->max_load > tmp)
3857 pwr_move += sds->busiest->cpu_power *
3858 min(sds->busiest_load_per_task, sds->max_load - tmp);
3860 /* Amount of load we'd add */
3861 if (sds->max_load * sds->busiest->cpu_power <
3862 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3863 tmp = (sds->max_load * sds->busiest->cpu_power) /
3864 sds->this->cpu_power;
3866 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3867 sds->this->cpu_power;
3868 pwr_move += sds->this->cpu_power *
3869 min(sds->this_load_per_task, sds->this_load + tmp);
3870 pwr_move /= SCHED_LOAD_SCALE;
3872 /* Move if we gain throughput */
3873 if (pwr_move > pwr_now)
3874 *imbalance = sds->busiest_load_per_task;
3878 * calculate_imbalance - Calculate the amount of imbalance present within the
3879 * groups of a given sched_domain during load balance.
3880 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3881 * @this_cpu: Cpu for which currently load balance is being performed.
3882 * @imbalance: The variable to store the imbalance.
3884 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3885 unsigned long *imbalance)
3887 unsigned long max_pull;
3889 * In the presence of smp nice balancing, certain scenarios can have
3890 * max load less than avg load(as we skip the groups at or below
3891 * its cpu_power, while calculating max_load..)
3893 if (sds->max_load < sds->avg_load) {
3895 return fix_small_imbalance(sds, this_cpu, imbalance);
3898 /* Don't want to pull so many tasks that a group would go idle */
3899 max_pull = min(sds->max_load - sds->avg_load,
3900 sds->max_load - sds->busiest_load_per_task);
3902 /* How much load to actually move to equalise the imbalance */
3903 *imbalance = min(max_pull * sds->busiest->cpu_power,
3904 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3908 * if *imbalance is less than the average load per runnable task
3909 * there is no gaurantee that any tasks will be moved so we'll have
3910 * a think about bumping its value to force at least one task to be
3913 if (*imbalance < sds->busiest_load_per_task)
3914 return fix_small_imbalance(sds, this_cpu, imbalance);
3917 /******* find_busiest_group() helpers end here *********************/
3920 * find_busiest_group - Returns the busiest group within the sched_domain
3921 * if there is an imbalance. If there isn't an imbalance, and
3922 * the user has opted for power-savings, it returns a group whose
3923 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3924 * such a group exists.
3926 * Also calculates the amount of weighted load which should be moved
3927 * to restore balance.
3929 * @sd: The sched_domain whose busiest group is to be returned.
3930 * @this_cpu: The cpu for which load balancing is currently being performed.
3931 * @imbalance: Variable which stores amount of weighted load which should
3932 * be moved to restore balance/put a group to idle.
3933 * @idle: The idle status of this_cpu.
3934 * @sd_idle: The idleness of sd
3935 * @cpus: The set of CPUs under consideration for load-balancing.
3936 * @balance: Pointer to a variable indicating if this_cpu
3937 * is the appropriate cpu to perform load balancing at this_level.
3939 * Returns: - the busiest group if imbalance exists.
3940 * - If no imbalance and user has opted for power-savings balance,
3941 * return the least loaded group whose CPUs can be
3942 * put to idle by rebalancing its tasks onto our group.
3944 static struct sched_group *
3945 find_busiest_group(struct sched_domain *sd, int this_cpu,
3946 unsigned long *imbalance, enum cpu_idle_type idle,
3947 int *sd_idle, const struct cpumask *cpus, int *balance)
3949 struct sd_lb_stats sds;
3951 memset(&sds, 0, sizeof(sds));
3954 * Compute the various statistics relavent for load balancing at
3957 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3960 /* Cases where imbalance does not exist from POV of this_cpu */
3961 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3963 * 2) There is no busy sibling group to pull from.
3964 * 3) This group is the busiest group.
3965 * 4) This group is more busy than the avg busieness at this
3967 * 5) The imbalance is within the specified limit.
3968 * 6) Any rebalance would lead to ping-pong
3970 if (balance && !(*balance))
3973 if (!sds.busiest || sds.busiest_nr_running == 0)
3976 if (sds.this_load >= sds.max_load)
3979 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3981 if (sds.this_load >= sds.avg_load)
3984 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3987 sds.busiest_load_per_task /= sds.busiest_nr_running;
3989 sds.busiest_load_per_task =
3990 min(sds.busiest_load_per_task, sds.avg_load);
3993 * We're trying to get all the cpus to the average_load, so we don't
3994 * want to push ourselves above the average load, nor do we wish to
3995 * reduce the max loaded cpu below the average load, as either of these
3996 * actions would just result in more rebalancing later, and ping-pong
3997 * tasks around. Thus we look for the minimum possible imbalance.
3998 * Negative imbalances (*we* are more loaded than anyone else) will
3999 * be counted as no imbalance for these purposes -- we can't fix that
4000 * by pulling tasks to us. Be careful of negative numbers as they'll
4001 * appear as very large values with unsigned longs.
4003 if (sds.max_load <= sds.busiest_load_per_task)
4006 /* Looks like there is an imbalance. Compute it */
4007 calculate_imbalance(&sds, this_cpu, imbalance);
4012 * There is no obvious imbalance. But check if we can do some balancing
4015 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4023 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4026 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4027 unsigned long imbalance, const struct cpumask *cpus)
4029 struct rq *busiest = NULL, *rq;
4030 unsigned long max_load = 0;
4033 for_each_cpu(i, sched_group_cpus(group)) {
4034 unsigned long power = power_of(i);
4035 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4038 if (!cpumask_test_cpu(i, cpus))
4042 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4045 if (capacity && rq->nr_running == 1 && wl > imbalance)
4048 if (wl > max_load) {
4058 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4059 * so long as it is large enough.
4061 #define MAX_PINNED_INTERVAL 512
4063 /* Working cpumask for load_balance and load_balance_newidle. */
4064 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4067 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4068 * tasks if there is an imbalance.
4070 static int load_balance(int this_cpu, struct rq *this_rq,
4071 struct sched_domain *sd, enum cpu_idle_type idle,
4074 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4075 struct sched_group *group;
4076 unsigned long imbalance;
4078 unsigned long flags;
4079 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4081 cpumask_setall(cpus);
4084 * When power savings policy is enabled for the parent domain, idle
4085 * sibling can pick up load irrespective of busy siblings. In this case,
4086 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4087 * portraying it as CPU_NOT_IDLE.
4089 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4090 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4093 schedstat_inc(sd, lb_count[idle]);
4097 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4104 schedstat_inc(sd, lb_nobusyg[idle]);
4108 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4110 schedstat_inc(sd, lb_nobusyq[idle]);
4114 BUG_ON(busiest == this_rq);
4116 schedstat_add(sd, lb_imbalance[idle], imbalance);
4119 if (busiest->nr_running > 1) {
4121 * Attempt to move tasks. If find_busiest_group has found
4122 * an imbalance but busiest->nr_running <= 1, the group is
4123 * still unbalanced. ld_moved simply stays zero, so it is
4124 * correctly treated as an imbalance.
4126 local_irq_save(flags);
4127 double_rq_lock(this_rq, busiest);
4128 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4129 imbalance, sd, idle, &all_pinned);
4130 double_rq_unlock(this_rq, busiest);
4131 local_irq_restore(flags);
4134 * some other cpu did the load balance for us.
4136 if (ld_moved && this_cpu != smp_processor_id())
4137 resched_cpu(this_cpu);
4139 /* All tasks on this runqueue were pinned by CPU affinity */
4140 if (unlikely(all_pinned)) {
4141 cpumask_clear_cpu(cpu_of(busiest), cpus);
4142 if (!cpumask_empty(cpus))
4149 schedstat_inc(sd, lb_failed[idle]);
4150 sd->nr_balance_failed++;
4152 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4154 spin_lock_irqsave(&busiest->lock, flags);
4156 /* don't kick the migration_thread, if the curr
4157 * task on busiest cpu can't be moved to this_cpu
4159 if (!cpumask_test_cpu(this_cpu,
4160 &busiest->curr->cpus_allowed)) {
4161 spin_unlock_irqrestore(&busiest->lock, flags);
4163 goto out_one_pinned;
4166 if (!busiest->active_balance) {
4167 busiest->active_balance = 1;
4168 busiest->push_cpu = this_cpu;
4171 spin_unlock_irqrestore(&busiest->lock, flags);
4173 wake_up_process(busiest->migration_thread);
4176 * We've kicked active balancing, reset the failure
4179 sd->nr_balance_failed = sd->cache_nice_tries+1;
4182 sd->nr_balance_failed = 0;
4184 if (likely(!active_balance)) {
4185 /* We were unbalanced, so reset the balancing interval */
4186 sd->balance_interval = sd->min_interval;
4189 * If we've begun active balancing, start to back off. This
4190 * case may not be covered by the all_pinned logic if there
4191 * is only 1 task on the busy runqueue (because we don't call
4194 if (sd->balance_interval < sd->max_interval)
4195 sd->balance_interval *= 2;
4198 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4199 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4205 schedstat_inc(sd, lb_balanced[idle]);
4207 sd->nr_balance_failed = 0;
4210 /* tune up the balancing interval */
4211 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4212 (sd->balance_interval < sd->max_interval))
4213 sd->balance_interval *= 2;
4215 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4216 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4227 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4228 * tasks if there is an imbalance.
4230 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4231 * this_rq is locked.
4234 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4236 struct sched_group *group;
4237 struct rq *busiest = NULL;
4238 unsigned long imbalance;
4242 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4244 cpumask_setall(cpus);
4247 * When power savings policy is enabled for the parent domain, idle
4248 * sibling can pick up load irrespective of busy siblings. In this case,
4249 * let the state of idle sibling percolate up as IDLE, instead of
4250 * portraying it as CPU_NOT_IDLE.
4252 if (sd->flags & SD_SHARE_CPUPOWER &&
4253 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4256 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4258 update_shares_locked(this_rq, sd);
4259 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4260 &sd_idle, cpus, NULL);
4262 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4266 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4268 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4272 BUG_ON(busiest == this_rq);
4274 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4277 if (busiest->nr_running > 1) {
4278 /* Attempt to move tasks */
4279 double_lock_balance(this_rq, busiest);
4280 /* this_rq->clock is already updated */
4281 update_rq_clock(busiest);
4282 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4283 imbalance, sd, CPU_NEWLY_IDLE,
4285 double_unlock_balance(this_rq, busiest);
4287 if (unlikely(all_pinned)) {
4288 cpumask_clear_cpu(cpu_of(busiest), cpus);
4289 if (!cpumask_empty(cpus))
4295 int active_balance = 0;
4297 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4298 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4299 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4302 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4305 if (sd->nr_balance_failed++ < 2)
4309 * The only task running in a non-idle cpu can be moved to this
4310 * cpu in an attempt to completely freeup the other CPU
4311 * package. The same method used to move task in load_balance()
4312 * have been extended for load_balance_newidle() to speedup
4313 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4315 * The package power saving logic comes from
4316 * find_busiest_group(). If there are no imbalance, then
4317 * f_b_g() will return NULL. However when sched_mc={1,2} then
4318 * f_b_g() will select a group from which a running task may be
4319 * pulled to this cpu in order to make the other package idle.
4320 * If there is no opportunity to make a package idle and if
4321 * there are no imbalance, then f_b_g() will return NULL and no
4322 * action will be taken in load_balance_newidle().
4324 * Under normal task pull operation due to imbalance, there
4325 * will be more than one task in the source run queue and
4326 * move_tasks() will succeed. ld_moved will be true and this
4327 * active balance code will not be triggered.
4330 /* Lock busiest in correct order while this_rq is held */
4331 double_lock_balance(this_rq, busiest);
4334 * don't kick the migration_thread, if the curr
4335 * task on busiest cpu can't be moved to this_cpu
4337 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4338 double_unlock_balance(this_rq, busiest);
4343 if (!busiest->active_balance) {
4344 busiest->active_balance = 1;
4345 busiest->push_cpu = this_cpu;
4349 double_unlock_balance(this_rq, busiest);
4351 * Should not call ttwu while holding a rq->lock
4353 spin_unlock(&this_rq->lock);
4355 wake_up_process(busiest->migration_thread);
4356 spin_lock(&this_rq->lock);
4359 sd->nr_balance_failed = 0;
4361 update_shares_locked(this_rq, sd);
4365 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4366 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4367 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4369 sd->nr_balance_failed = 0;
4375 * idle_balance is called by schedule() if this_cpu is about to become
4376 * idle. Attempts to pull tasks from other CPUs.
4378 static void idle_balance(int this_cpu, struct rq *this_rq)
4380 struct sched_domain *sd;
4381 int pulled_task = 0;
4382 unsigned long next_balance = jiffies + HZ;
4384 for_each_domain(this_cpu, sd) {
4385 unsigned long interval;
4387 if (!(sd->flags & SD_LOAD_BALANCE))
4390 if (sd->flags & SD_BALANCE_NEWIDLE)
4391 /* If we've pulled tasks over stop searching: */
4392 pulled_task = load_balance_newidle(this_cpu, this_rq,
4395 interval = msecs_to_jiffies(sd->balance_interval);
4396 if (time_after(next_balance, sd->last_balance + interval))
4397 next_balance = sd->last_balance + interval;
4401 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4403 * We are going idle. next_balance may be set based on
4404 * a busy processor. So reset next_balance.
4406 this_rq->next_balance = next_balance;
4411 * active_load_balance is run by migration threads. It pushes running tasks
4412 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4413 * running on each physical CPU where possible, and avoids physical /
4414 * logical imbalances.
4416 * Called with busiest_rq locked.
4418 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4420 int target_cpu = busiest_rq->push_cpu;
4421 struct sched_domain *sd;
4422 struct rq *target_rq;
4424 /* Is there any task to move? */
4425 if (busiest_rq->nr_running <= 1)
4428 target_rq = cpu_rq(target_cpu);
4431 * This condition is "impossible", if it occurs
4432 * we need to fix it. Originally reported by
4433 * Bjorn Helgaas on a 128-cpu setup.
4435 BUG_ON(busiest_rq == target_rq);
4437 /* move a task from busiest_rq to target_rq */
4438 double_lock_balance(busiest_rq, target_rq);
4439 update_rq_clock(busiest_rq);
4440 update_rq_clock(target_rq);
4442 /* Search for an sd spanning us and the target CPU. */
4443 for_each_domain(target_cpu, sd) {
4444 if ((sd->flags & SD_LOAD_BALANCE) &&
4445 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4450 schedstat_inc(sd, alb_count);
4452 if (move_one_task(target_rq, target_cpu, busiest_rq,
4454 schedstat_inc(sd, alb_pushed);
4456 schedstat_inc(sd, alb_failed);
4458 double_unlock_balance(busiest_rq, target_rq);
4463 atomic_t load_balancer;
4464 cpumask_var_t cpu_mask;
4465 cpumask_var_t ilb_grp_nohz_mask;
4466 } nohz ____cacheline_aligned = {
4467 .load_balancer = ATOMIC_INIT(-1),
4470 int get_nohz_load_balancer(void)
4472 return atomic_read(&nohz.load_balancer);
4475 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4477 * lowest_flag_domain - Return lowest sched_domain containing flag.
4478 * @cpu: The cpu whose lowest level of sched domain is to
4480 * @flag: The flag to check for the lowest sched_domain
4481 * for the given cpu.
4483 * Returns the lowest sched_domain of a cpu which contains the given flag.
4485 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4487 struct sched_domain *sd;
4489 for_each_domain(cpu, sd)
4490 if (sd && (sd->flags & flag))
4497 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4498 * @cpu: The cpu whose domains we're iterating over.
4499 * @sd: variable holding the value of the power_savings_sd
4501 * @flag: The flag to filter the sched_domains to be iterated.
4503 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4504 * set, starting from the lowest sched_domain to the highest.
4506 #define for_each_flag_domain(cpu, sd, flag) \
4507 for (sd = lowest_flag_domain(cpu, flag); \
4508 (sd && (sd->flags & flag)); sd = sd->parent)
4511 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4512 * @ilb_group: group to be checked for semi-idleness
4514 * Returns: 1 if the group is semi-idle. 0 otherwise.
4516 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4517 * and atleast one non-idle CPU. This helper function checks if the given
4518 * sched_group is semi-idle or not.
4520 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4522 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4523 sched_group_cpus(ilb_group));
4526 * A sched_group is semi-idle when it has atleast one busy cpu
4527 * and atleast one idle cpu.
4529 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4532 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4538 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4539 * @cpu: The cpu which is nominating a new idle_load_balancer.
4541 * Returns: Returns the id of the idle load balancer if it exists,
4542 * Else, returns >= nr_cpu_ids.
4544 * This algorithm picks the idle load balancer such that it belongs to a
4545 * semi-idle powersavings sched_domain. The idea is to try and avoid
4546 * completely idle packages/cores just for the purpose of idle load balancing
4547 * when there are other idle cpu's which are better suited for that job.
4549 static int find_new_ilb(int cpu)
4551 struct sched_domain *sd;
4552 struct sched_group *ilb_group;
4555 * Have idle load balancer selection from semi-idle packages only
4556 * when power-aware load balancing is enabled
4558 if (!(sched_smt_power_savings || sched_mc_power_savings))
4562 * Optimize for the case when we have no idle CPUs or only one
4563 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4565 if (cpumask_weight(nohz.cpu_mask) < 2)
4568 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4569 ilb_group = sd->groups;
4572 if (is_semi_idle_group(ilb_group))
4573 return cpumask_first(nohz.ilb_grp_nohz_mask);
4575 ilb_group = ilb_group->next;
4577 } while (ilb_group != sd->groups);
4581 return cpumask_first(nohz.cpu_mask);
4583 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4584 static inline int find_new_ilb(int call_cpu)
4586 return cpumask_first(nohz.cpu_mask);
4591 * This routine will try to nominate the ilb (idle load balancing)
4592 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4593 * load balancing on behalf of all those cpus. If all the cpus in the system
4594 * go into this tickless mode, then there will be no ilb owner (as there is
4595 * no need for one) and all the cpus will sleep till the next wakeup event
4598 * For the ilb owner, tick is not stopped. And this tick will be used
4599 * for idle load balancing. ilb owner will still be part of
4602 * While stopping the tick, this cpu will become the ilb owner if there
4603 * is no other owner. And will be the owner till that cpu becomes busy
4604 * or if all cpus in the system stop their ticks at which point
4605 * there is no need for ilb owner.
4607 * When the ilb owner becomes busy, it nominates another owner, during the
4608 * next busy scheduler_tick()
4610 int select_nohz_load_balancer(int stop_tick)
4612 int cpu = smp_processor_id();
4615 cpu_rq(cpu)->in_nohz_recently = 1;
4617 if (!cpu_active(cpu)) {
4618 if (atomic_read(&nohz.load_balancer) != cpu)
4622 * If we are going offline and still the leader,
4625 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4631 cpumask_set_cpu(cpu, nohz.cpu_mask);
4633 /* time for ilb owner also to sleep */
4634 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4635 if (atomic_read(&nohz.load_balancer) == cpu)
4636 atomic_set(&nohz.load_balancer, -1);
4640 if (atomic_read(&nohz.load_balancer) == -1) {
4641 /* make me the ilb owner */
4642 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4644 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4647 if (!(sched_smt_power_savings ||
4648 sched_mc_power_savings))
4651 * Check to see if there is a more power-efficient
4654 new_ilb = find_new_ilb(cpu);
4655 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4656 atomic_set(&nohz.load_balancer, -1);
4657 resched_cpu(new_ilb);
4663 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4666 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4668 if (atomic_read(&nohz.load_balancer) == cpu)
4669 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4676 static DEFINE_SPINLOCK(balancing);
4679 * It checks each scheduling domain to see if it is due to be balanced,
4680 * and initiates a balancing operation if so.
4682 * Balancing parameters are set up in arch_init_sched_domains.
4684 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4687 struct rq *rq = cpu_rq(cpu);
4688 unsigned long interval;
4689 struct sched_domain *sd;
4690 /* Earliest time when we have to do rebalance again */
4691 unsigned long next_balance = jiffies + 60*HZ;
4692 int update_next_balance = 0;
4695 for_each_domain(cpu, sd) {
4696 if (!(sd->flags & SD_LOAD_BALANCE))
4699 interval = sd->balance_interval;
4700 if (idle != CPU_IDLE)
4701 interval *= sd->busy_factor;
4703 /* scale ms to jiffies */
4704 interval = msecs_to_jiffies(interval);
4705 if (unlikely(!interval))
4707 if (interval > HZ*NR_CPUS/10)
4708 interval = HZ*NR_CPUS/10;
4710 need_serialize = sd->flags & SD_SERIALIZE;
4712 if (need_serialize) {
4713 if (!spin_trylock(&balancing))
4717 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4718 if (load_balance(cpu, rq, sd, idle, &balance)) {
4720 * We've pulled tasks over so either we're no
4721 * longer idle, or one of our SMT siblings is
4724 idle = CPU_NOT_IDLE;
4726 sd->last_balance = jiffies;
4729 spin_unlock(&balancing);
4731 if (time_after(next_balance, sd->last_balance + interval)) {
4732 next_balance = sd->last_balance + interval;
4733 update_next_balance = 1;
4737 * Stop the load balance at this level. There is another
4738 * CPU in our sched group which is doing load balancing more
4746 * next_balance will be updated only when there is a need.
4747 * When the cpu is attached to null domain for ex, it will not be
4750 if (likely(update_next_balance))
4751 rq->next_balance = next_balance;
4755 * run_rebalance_domains is triggered when needed from the scheduler tick.
4756 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4757 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4759 static void run_rebalance_domains(struct softirq_action *h)
4761 int this_cpu = smp_processor_id();
4762 struct rq *this_rq = cpu_rq(this_cpu);
4763 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4764 CPU_IDLE : CPU_NOT_IDLE;
4766 rebalance_domains(this_cpu, idle);
4770 * If this cpu is the owner for idle load balancing, then do the
4771 * balancing on behalf of the other idle cpus whose ticks are
4774 if (this_rq->idle_at_tick &&
4775 atomic_read(&nohz.load_balancer) == this_cpu) {
4779 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4780 if (balance_cpu == this_cpu)
4784 * If this cpu gets work to do, stop the load balancing
4785 * work being done for other cpus. Next load
4786 * balancing owner will pick it up.
4791 rebalance_domains(balance_cpu, CPU_IDLE);
4793 rq = cpu_rq(balance_cpu);
4794 if (time_after(this_rq->next_balance, rq->next_balance))
4795 this_rq->next_balance = rq->next_balance;
4801 static inline int on_null_domain(int cpu)
4803 return !rcu_dereference(cpu_rq(cpu)->sd);
4807 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4809 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4810 * idle load balancing owner or decide to stop the periodic load balancing,
4811 * if the whole system is idle.
4813 static inline void trigger_load_balance(struct rq *rq, int cpu)
4817 * If we were in the nohz mode recently and busy at the current
4818 * scheduler tick, then check if we need to nominate new idle
4821 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4822 rq->in_nohz_recently = 0;
4824 if (atomic_read(&nohz.load_balancer) == cpu) {
4825 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4826 atomic_set(&nohz.load_balancer, -1);
4829 if (atomic_read(&nohz.load_balancer) == -1) {
4830 int ilb = find_new_ilb(cpu);
4832 if (ilb < nr_cpu_ids)
4838 * If this cpu is idle and doing idle load balancing for all the
4839 * cpus with ticks stopped, is it time for that to stop?
4841 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4842 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4848 * If this cpu is idle and the idle load balancing is done by
4849 * someone else, then no need raise the SCHED_SOFTIRQ
4851 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4852 cpumask_test_cpu(cpu, nohz.cpu_mask))
4855 /* Don't need to rebalance while attached to NULL domain */
4856 if (time_after_eq(jiffies, rq->next_balance) &&
4857 likely(!on_null_domain(cpu)))
4858 raise_softirq(SCHED_SOFTIRQ);
4861 #else /* CONFIG_SMP */
4864 * on UP we do not need to balance between CPUs:
4866 static inline void idle_balance(int cpu, struct rq *rq)
4872 DEFINE_PER_CPU(struct kernel_stat, kstat);
4874 EXPORT_PER_CPU_SYMBOL(kstat);
4877 * Return any ns on the sched_clock that have not yet been accounted in
4878 * @p in case that task is currently running.
4880 * Called with task_rq_lock() held on @rq.
4882 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4886 if (task_current(rq, p)) {
4887 update_rq_clock(rq);
4888 ns = rq->clock - p->se.exec_start;
4896 unsigned long long task_delta_exec(struct task_struct *p)
4898 unsigned long flags;
4902 rq = task_rq_lock(p, &flags);
4903 ns = do_task_delta_exec(p, rq);
4904 task_rq_unlock(rq, &flags);
4910 * Return accounted runtime for the task.
4911 * In case the task is currently running, return the runtime plus current's
4912 * pending runtime that have not been accounted yet.
4914 unsigned long long task_sched_runtime(struct task_struct *p)
4916 unsigned long flags;
4920 rq = task_rq_lock(p, &flags);
4921 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4922 task_rq_unlock(rq, &flags);
4928 * Return sum_exec_runtime for the thread group.
4929 * In case the task is currently running, return the sum plus current's
4930 * pending runtime that have not been accounted yet.
4932 * Note that the thread group might have other running tasks as well,
4933 * so the return value not includes other pending runtime that other
4934 * running tasks might have.
4936 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4938 struct task_cputime totals;
4939 unsigned long flags;
4943 rq = task_rq_lock(p, &flags);
4944 thread_group_cputime(p, &totals);
4945 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4946 task_rq_unlock(rq, &flags);
4952 * Account user cpu time to a process.
4953 * @p: the process that the cpu time gets accounted to
4954 * @cputime: the cpu time spent in user space since the last update
4955 * @cputime_scaled: cputime scaled by cpu frequency
4957 void account_user_time(struct task_struct *p, cputime_t cputime,
4958 cputime_t cputime_scaled)
4960 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4963 /* Add user time to process. */
4964 p->utime = cputime_add(p->utime, cputime);
4965 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4966 account_group_user_time(p, cputime);
4968 /* Add user time to cpustat. */
4969 tmp = cputime_to_cputime64(cputime);
4970 if (TASK_NICE(p) > 0)
4971 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4973 cpustat->user = cputime64_add(cpustat->user, tmp);
4975 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4976 /* Account for user time used */
4977 acct_update_integrals(p);
4981 * Account guest cpu time to a process.
4982 * @p: the process that the cpu time gets accounted to
4983 * @cputime: the cpu time spent in virtual machine since the last update
4984 * @cputime_scaled: cputime scaled by cpu frequency
4986 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4987 cputime_t cputime_scaled)
4990 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4992 tmp = cputime_to_cputime64(cputime);
4994 /* Add guest time to process. */
4995 p->utime = cputime_add(p->utime, cputime);
4996 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4997 account_group_user_time(p, cputime);
4998 p->gtime = cputime_add(p->gtime, cputime);
5000 /* Add guest time to cpustat. */
5001 cpustat->user = cputime64_add(cpustat->user, tmp);
5002 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5006 * Account system cpu time to a process.
5007 * @p: the process that the cpu time gets accounted to
5008 * @hardirq_offset: the offset to subtract from hardirq_count()
5009 * @cputime: the cpu time spent in kernel space since the last update
5010 * @cputime_scaled: cputime scaled by cpu frequency
5012 void account_system_time(struct task_struct *p, int hardirq_offset,
5013 cputime_t cputime, cputime_t cputime_scaled)
5015 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5018 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5019 account_guest_time(p, cputime, cputime_scaled);
5023 /* Add system time to process. */
5024 p->stime = cputime_add(p->stime, cputime);
5025 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5026 account_group_system_time(p, cputime);
5028 /* Add system time to cpustat. */
5029 tmp = cputime_to_cputime64(cputime);
5030 if (hardirq_count() - hardirq_offset)
5031 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5032 else if (softirq_count())
5033 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5035 cpustat->system = cputime64_add(cpustat->system, tmp);
5037 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5039 /* Account for system time used */
5040 acct_update_integrals(p);
5044 * Account for involuntary wait time.
5045 * @steal: the cpu time spent in involuntary wait
5047 void account_steal_time(cputime_t cputime)
5049 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5050 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5052 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5056 * Account for idle time.
5057 * @cputime: the cpu time spent in idle wait
5059 void account_idle_time(cputime_t cputime)
5061 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5062 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5063 struct rq *rq = this_rq();
5065 if (atomic_read(&rq->nr_iowait) > 0)
5066 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5068 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5071 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5074 * Account a single tick of cpu time.
5075 * @p: the process that the cpu time gets accounted to
5076 * @user_tick: indicates if the tick is a user or a system tick
5078 void account_process_tick(struct task_struct *p, int user_tick)
5080 cputime_t one_jiffy = jiffies_to_cputime(1);
5081 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
5082 struct rq *rq = this_rq();
5085 account_user_time(p, one_jiffy, one_jiffy_scaled);
5086 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5087 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
5090 account_idle_time(one_jiffy);
5094 * Account multiple ticks of steal time.
5095 * @p: the process from which the cpu time has been stolen
5096 * @ticks: number of stolen ticks
5098 void account_steal_ticks(unsigned long ticks)
5100 account_steal_time(jiffies_to_cputime(ticks));
5104 * Account multiple ticks of idle time.
5105 * @ticks: number of stolen ticks
5107 void account_idle_ticks(unsigned long ticks)
5109 account_idle_time(jiffies_to_cputime(ticks));
5115 * Use precise platform statistics if available:
5117 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5118 cputime_t task_utime(struct task_struct *p)
5123 cputime_t task_stime(struct task_struct *p)
5128 cputime_t task_utime(struct task_struct *p)
5130 clock_t utime = cputime_to_clock_t(p->utime),
5131 total = utime + cputime_to_clock_t(p->stime);
5135 * Use CFS's precise accounting:
5137 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5141 do_div(temp, total);
5143 utime = (clock_t)temp;
5145 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5146 return p->prev_utime;
5149 cputime_t task_stime(struct task_struct *p)
5154 * Use CFS's precise accounting. (we subtract utime from
5155 * the total, to make sure the total observed by userspace
5156 * grows monotonically - apps rely on that):
5158 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5159 cputime_to_clock_t(task_utime(p));
5162 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5164 return p->prev_stime;
5168 inline cputime_t task_gtime(struct task_struct *p)
5174 * This function gets called by the timer code, with HZ frequency.
5175 * We call it with interrupts disabled.
5177 * It also gets called by the fork code, when changing the parent's
5180 void scheduler_tick(void)
5182 int cpu = smp_processor_id();
5183 struct rq *rq = cpu_rq(cpu);
5184 struct task_struct *curr = rq->curr;
5188 spin_lock(&rq->lock);
5189 update_rq_clock(rq);
5190 update_cpu_load(rq);
5191 curr->sched_class->task_tick(rq, curr, 0);
5192 spin_unlock(&rq->lock);
5194 perf_counter_task_tick(curr, cpu);
5197 rq->idle_at_tick = idle_cpu(cpu);
5198 trigger_load_balance(rq, cpu);
5202 notrace unsigned long get_parent_ip(unsigned long addr)
5204 if (in_lock_functions(addr)) {
5205 addr = CALLER_ADDR2;
5206 if (in_lock_functions(addr))
5207 addr = CALLER_ADDR3;
5212 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5213 defined(CONFIG_PREEMPT_TRACER))
5215 void __kprobes add_preempt_count(int val)
5217 #ifdef CONFIG_DEBUG_PREEMPT
5221 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5224 preempt_count() += val;
5225 #ifdef CONFIG_DEBUG_PREEMPT
5227 * Spinlock count overflowing soon?
5229 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5232 if (preempt_count() == val)
5233 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5235 EXPORT_SYMBOL(add_preempt_count);
5237 void __kprobes sub_preempt_count(int val)
5239 #ifdef CONFIG_DEBUG_PREEMPT
5243 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5246 * Is the spinlock portion underflowing?
5248 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5249 !(preempt_count() & PREEMPT_MASK)))
5253 if (preempt_count() == val)
5254 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5255 preempt_count() -= val;
5257 EXPORT_SYMBOL(sub_preempt_count);
5262 * Print scheduling while atomic bug:
5264 static noinline void __schedule_bug(struct task_struct *prev)
5266 struct pt_regs *regs = get_irq_regs();
5268 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5269 prev->comm, prev->pid, preempt_count());
5271 debug_show_held_locks(prev);
5273 if (irqs_disabled())
5274 print_irqtrace_events(prev);
5283 * Various schedule()-time debugging checks and statistics:
5285 static inline void schedule_debug(struct task_struct *prev)
5288 * Test if we are atomic. Since do_exit() needs to call into
5289 * schedule() atomically, we ignore that path for now.
5290 * Otherwise, whine if we are scheduling when we should not be.
5292 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5293 __schedule_bug(prev);
5295 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5297 schedstat_inc(this_rq(), sched_count);
5298 #ifdef CONFIG_SCHEDSTATS
5299 if (unlikely(prev->lock_depth >= 0)) {
5300 schedstat_inc(this_rq(), bkl_count);
5301 schedstat_inc(prev, sched_info.bkl_count);
5306 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5308 if (prev->state == TASK_RUNNING) {
5309 u64 runtime = prev->se.sum_exec_runtime;
5311 runtime -= prev->se.prev_sum_exec_runtime;
5312 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5315 * In order to avoid avg_overlap growing stale when we are
5316 * indeed overlapping and hence not getting put to sleep, grow
5317 * the avg_overlap on preemption.
5319 * We use the average preemption runtime because that
5320 * correlates to the amount of cache footprint a task can
5323 update_avg(&prev->se.avg_overlap, runtime);
5325 prev->sched_class->put_prev_task(rq, prev);
5329 * Pick up the highest-prio task:
5331 static inline struct task_struct *
5332 pick_next_task(struct rq *rq)
5334 const struct sched_class *class;
5335 struct task_struct *p;
5338 * Optimization: we know that if all tasks are in
5339 * the fair class we can call that function directly:
5341 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5342 p = fair_sched_class.pick_next_task(rq);
5347 class = sched_class_highest;
5349 p = class->pick_next_task(rq);
5353 * Will never be NULL as the idle class always
5354 * returns a non-NULL p:
5356 class = class->next;
5361 * schedule() is the main scheduler function.
5363 asmlinkage void __sched schedule(void)
5365 struct task_struct *prev, *next;
5366 unsigned long *switch_count;
5372 cpu = smp_processor_id();
5376 switch_count = &prev->nivcsw;
5378 release_kernel_lock(prev);
5379 need_resched_nonpreemptible:
5381 schedule_debug(prev);
5383 if (sched_feat(HRTICK))
5386 spin_lock_irq(&rq->lock);
5387 update_rq_clock(rq);
5388 clear_tsk_need_resched(prev);
5390 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5391 if (unlikely(signal_pending_state(prev->state, prev)))
5392 prev->state = TASK_RUNNING;
5394 deactivate_task(rq, prev, 1);
5395 switch_count = &prev->nvcsw;
5398 pre_schedule(rq, prev);
5400 if (unlikely(!rq->nr_running))
5401 idle_balance(cpu, rq);
5403 put_prev_task(rq, prev);
5404 next = pick_next_task(rq);
5406 if (likely(prev != next)) {
5407 sched_info_switch(prev, next);
5408 perf_counter_task_sched_out(prev, next, cpu);
5414 context_switch(rq, prev, next); /* unlocks the rq */
5416 * the context switch might have flipped the stack from under
5417 * us, hence refresh the local variables.
5419 cpu = smp_processor_id();
5422 spin_unlock_irq(&rq->lock);
5426 if (unlikely(reacquire_kernel_lock(current) < 0))
5427 goto need_resched_nonpreemptible;
5429 preempt_enable_no_resched();
5433 EXPORT_SYMBOL(schedule);
5437 * Look out! "owner" is an entirely speculative pointer
5438 * access and not reliable.
5440 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5445 if (!sched_feat(OWNER_SPIN))
5448 #ifdef CONFIG_DEBUG_PAGEALLOC
5450 * Need to access the cpu field knowing that
5451 * DEBUG_PAGEALLOC could have unmapped it if
5452 * the mutex owner just released it and exited.
5454 if (probe_kernel_address(&owner->cpu, cpu))
5461 * Even if the access succeeded (likely case),
5462 * the cpu field may no longer be valid.
5464 if (cpu >= nr_cpumask_bits)
5468 * We need to validate that we can do a
5469 * get_cpu() and that we have the percpu area.
5471 if (!cpu_online(cpu))
5478 * Owner changed, break to re-assess state.
5480 if (lock->owner != owner)
5484 * Is that owner really running on that cpu?
5486 if (task_thread_info(rq->curr) != owner || need_resched())
5496 #ifdef CONFIG_PREEMPT
5498 * this is the entry point to schedule() from in-kernel preemption
5499 * off of preempt_enable. Kernel preemptions off return from interrupt
5500 * occur there and call schedule directly.
5502 asmlinkage void __sched preempt_schedule(void)
5504 struct thread_info *ti = current_thread_info();
5507 * If there is a non-zero preempt_count or interrupts are disabled,
5508 * we do not want to preempt the current task. Just return..
5510 if (likely(ti->preempt_count || irqs_disabled()))
5514 add_preempt_count(PREEMPT_ACTIVE);
5516 sub_preempt_count(PREEMPT_ACTIVE);
5519 * Check again in case we missed a preemption opportunity
5520 * between schedule and now.
5523 } while (need_resched());
5525 EXPORT_SYMBOL(preempt_schedule);
5528 * this is the entry point to schedule() from kernel preemption
5529 * off of irq context.
5530 * Note, that this is called and return with irqs disabled. This will
5531 * protect us against recursive calling from irq.
5533 asmlinkage void __sched preempt_schedule_irq(void)
5535 struct thread_info *ti = current_thread_info();
5537 /* Catch callers which need to be fixed */
5538 BUG_ON(ti->preempt_count || !irqs_disabled());
5541 add_preempt_count(PREEMPT_ACTIVE);
5544 local_irq_disable();
5545 sub_preempt_count(PREEMPT_ACTIVE);
5548 * Check again in case we missed a preemption opportunity
5549 * between schedule and now.
5552 } while (need_resched());
5555 #endif /* CONFIG_PREEMPT */
5557 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5560 return try_to_wake_up(curr->private, mode, sync);
5562 EXPORT_SYMBOL(default_wake_function);
5565 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5566 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5567 * number) then we wake all the non-exclusive tasks and one exclusive task.
5569 * There are circumstances in which we can try to wake a task which has already
5570 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5571 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5573 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5574 int nr_exclusive, int sync, void *key)
5576 wait_queue_t *curr, *next;
5578 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5579 unsigned flags = curr->flags;
5581 if (curr->func(curr, mode, sync, key) &&
5582 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5588 * __wake_up - wake up threads blocked on a waitqueue.
5590 * @mode: which threads
5591 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5592 * @key: is directly passed to the wakeup function
5594 * It may be assumed that this function implies a write memory barrier before
5595 * changing the task state if and only if any tasks are woken up.
5597 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5598 int nr_exclusive, void *key)
5600 unsigned long flags;
5602 spin_lock_irqsave(&q->lock, flags);
5603 __wake_up_common(q, mode, nr_exclusive, 0, key);
5604 spin_unlock_irqrestore(&q->lock, flags);
5606 EXPORT_SYMBOL(__wake_up);
5609 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5611 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5613 __wake_up_common(q, mode, 1, 0, NULL);
5616 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5618 __wake_up_common(q, mode, 1, 0, key);
5622 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5624 * @mode: which threads
5625 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5626 * @key: opaque value to be passed to wakeup targets
5628 * The sync wakeup differs that the waker knows that it will schedule
5629 * away soon, so while the target thread will be woken up, it will not
5630 * be migrated to another CPU - ie. the two threads are 'synchronized'
5631 * with each other. This can prevent needless bouncing between CPUs.
5633 * On UP it can prevent extra preemption.
5635 * It may be assumed that this function implies a write memory barrier before
5636 * changing the task state if and only if any tasks are woken up.
5638 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5639 int nr_exclusive, void *key)
5641 unsigned long flags;
5647 if (unlikely(!nr_exclusive))
5650 spin_lock_irqsave(&q->lock, flags);
5651 __wake_up_common(q, mode, nr_exclusive, sync, key);
5652 spin_unlock_irqrestore(&q->lock, flags);
5654 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5657 * __wake_up_sync - see __wake_up_sync_key()
5659 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5661 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5663 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5666 * complete: - signals a single thread waiting on this completion
5667 * @x: holds the state of this particular completion
5669 * This will wake up a single thread waiting on this completion. Threads will be
5670 * awakened in the same order in which they were queued.
5672 * See also complete_all(), wait_for_completion() and related routines.
5674 * It may be assumed that this function implies a write memory barrier before
5675 * changing the task state if and only if any tasks are woken up.
5677 void complete(struct completion *x)
5679 unsigned long flags;
5681 spin_lock_irqsave(&x->wait.lock, flags);
5683 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5684 spin_unlock_irqrestore(&x->wait.lock, flags);
5686 EXPORT_SYMBOL(complete);
5689 * complete_all: - signals all threads waiting on this completion
5690 * @x: holds the state of this particular completion
5692 * This will wake up all threads waiting on this particular completion event.
5694 * It may be assumed that this function implies a write memory barrier before
5695 * changing the task state if and only if any tasks are woken up.
5697 void complete_all(struct completion *x)
5699 unsigned long flags;
5701 spin_lock_irqsave(&x->wait.lock, flags);
5702 x->done += UINT_MAX/2;
5703 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5704 spin_unlock_irqrestore(&x->wait.lock, flags);
5706 EXPORT_SYMBOL(complete_all);
5708 static inline long __sched
5709 do_wait_for_common(struct completion *x, long timeout, int state)
5712 DECLARE_WAITQUEUE(wait, current);
5714 wait.flags |= WQ_FLAG_EXCLUSIVE;
5715 __add_wait_queue_tail(&x->wait, &wait);
5717 if (signal_pending_state(state, current)) {
5718 timeout = -ERESTARTSYS;
5721 __set_current_state(state);
5722 spin_unlock_irq(&x->wait.lock);
5723 timeout = schedule_timeout(timeout);
5724 spin_lock_irq(&x->wait.lock);
5725 } while (!x->done && timeout);
5726 __remove_wait_queue(&x->wait, &wait);
5731 return timeout ?: 1;
5735 wait_for_common(struct completion *x, long timeout, int state)
5739 spin_lock_irq(&x->wait.lock);
5740 timeout = do_wait_for_common(x, timeout, state);
5741 spin_unlock_irq(&x->wait.lock);
5746 * wait_for_completion: - waits for completion of a task
5747 * @x: holds the state of this particular completion
5749 * This waits to be signaled for completion of a specific task. It is NOT
5750 * interruptible and there is no timeout.
5752 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5753 * and interrupt capability. Also see complete().
5755 void __sched wait_for_completion(struct completion *x)
5757 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5759 EXPORT_SYMBOL(wait_for_completion);
5762 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5763 * @x: holds the state of this particular completion
5764 * @timeout: timeout value in jiffies
5766 * This waits for either a completion of a specific task to be signaled or for a
5767 * specified timeout to expire. The timeout is in jiffies. It is not
5770 unsigned long __sched
5771 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5773 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5775 EXPORT_SYMBOL(wait_for_completion_timeout);
5778 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5779 * @x: holds the state of this particular completion
5781 * This waits for completion of a specific task to be signaled. It is
5784 int __sched wait_for_completion_interruptible(struct completion *x)
5786 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5787 if (t == -ERESTARTSYS)
5791 EXPORT_SYMBOL(wait_for_completion_interruptible);
5794 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5795 * @x: holds the state of this particular completion
5796 * @timeout: timeout value in jiffies
5798 * This waits for either a completion of a specific task to be signaled or for a
5799 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5801 unsigned long __sched
5802 wait_for_completion_interruptible_timeout(struct completion *x,
5803 unsigned long timeout)
5805 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5807 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5810 * wait_for_completion_killable: - waits for completion of a task (killable)
5811 * @x: holds the state of this particular completion
5813 * This waits to be signaled for completion of a specific task. It can be
5814 * interrupted by a kill signal.
5816 int __sched wait_for_completion_killable(struct completion *x)
5818 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5819 if (t == -ERESTARTSYS)
5823 EXPORT_SYMBOL(wait_for_completion_killable);
5826 * try_wait_for_completion - try to decrement a completion without blocking
5827 * @x: completion structure
5829 * Returns: 0 if a decrement cannot be done without blocking
5830 * 1 if a decrement succeeded.
5832 * If a completion is being used as a counting completion,
5833 * attempt to decrement the counter without blocking. This
5834 * enables us to avoid waiting if the resource the completion
5835 * is protecting is not available.
5837 bool try_wait_for_completion(struct completion *x)
5841 spin_lock_irq(&x->wait.lock);
5846 spin_unlock_irq(&x->wait.lock);
5849 EXPORT_SYMBOL(try_wait_for_completion);
5852 * completion_done - Test to see if a completion has any waiters
5853 * @x: completion structure
5855 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5856 * 1 if there are no waiters.
5859 bool completion_done(struct completion *x)
5863 spin_lock_irq(&x->wait.lock);
5866 spin_unlock_irq(&x->wait.lock);
5869 EXPORT_SYMBOL(completion_done);
5872 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5874 unsigned long flags;
5877 init_waitqueue_entry(&wait, current);
5879 __set_current_state(state);
5881 spin_lock_irqsave(&q->lock, flags);
5882 __add_wait_queue(q, &wait);
5883 spin_unlock(&q->lock);
5884 timeout = schedule_timeout(timeout);
5885 spin_lock_irq(&q->lock);
5886 __remove_wait_queue(q, &wait);
5887 spin_unlock_irqrestore(&q->lock, flags);
5892 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5894 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5896 EXPORT_SYMBOL(interruptible_sleep_on);
5899 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5901 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5903 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5905 void __sched sleep_on(wait_queue_head_t *q)
5907 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5909 EXPORT_SYMBOL(sleep_on);
5911 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5913 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5915 EXPORT_SYMBOL(sleep_on_timeout);
5917 #ifdef CONFIG_RT_MUTEXES
5920 * rt_mutex_setprio - set the current priority of a task
5922 * @prio: prio value (kernel-internal form)
5924 * This function changes the 'effective' priority of a task. It does
5925 * not touch ->normal_prio like __setscheduler().
5927 * Used by the rt_mutex code to implement priority inheritance logic.
5929 void rt_mutex_setprio(struct task_struct *p, int prio)
5931 unsigned long flags;
5932 int oldprio, on_rq, running;
5934 const struct sched_class *prev_class = p->sched_class;
5936 BUG_ON(prio < 0 || prio > MAX_PRIO);
5938 rq = task_rq_lock(p, &flags);
5939 update_rq_clock(rq);
5942 on_rq = p->se.on_rq;
5943 running = task_current(rq, p);
5945 dequeue_task(rq, p, 0);
5947 p->sched_class->put_prev_task(rq, p);
5950 p->sched_class = &rt_sched_class;
5952 p->sched_class = &fair_sched_class;
5957 p->sched_class->set_curr_task(rq);
5959 enqueue_task(rq, p, 0);
5961 check_class_changed(rq, p, prev_class, oldprio, running);
5963 task_rq_unlock(rq, &flags);
5968 void set_user_nice(struct task_struct *p, long nice)
5970 int old_prio, delta, on_rq;
5971 unsigned long flags;
5974 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5977 * We have to be careful, if called from sys_setpriority(),
5978 * the task might be in the middle of scheduling on another CPU.
5980 rq = task_rq_lock(p, &flags);
5981 update_rq_clock(rq);
5983 * The RT priorities are set via sched_setscheduler(), but we still
5984 * allow the 'normal' nice value to be set - but as expected
5985 * it wont have any effect on scheduling until the task is
5986 * SCHED_FIFO/SCHED_RR:
5988 if (task_has_rt_policy(p)) {
5989 p->static_prio = NICE_TO_PRIO(nice);
5992 on_rq = p->se.on_rq;
5994 dequeue_task(rq, p, 0);
5996 p->static_prio = NICE_TO_PRIO(nice);
5999 p->prio = effective_prio(p);
6000 delta = p->prio - old_prio;
6003 enqueue_task(rq, p, 0);
6005 * If the task increased its priority or is running and
6006 * lowered its priority, then reschedule its CPU:
6008 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6009 resched_task(rq->curr);
6012 task_rq_unlock(rq, &flags);
6014 EXPORT_SYMBOL(set_user_nice);
6017 * can_nice - check if a task can reduce its nice value
6021 int can_nice(const struct task_struct *p, const int nice)
6023 /* convert nice value [19,-20] to rlimit style value [1,40] */
6024 int nice_rlim = 20 - nice;
6026 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6027 capable(CAP_SYS_NICE));
6030 #ifdef __ARCH_WANT_SYS_NICE
6033 * sys_nice - change the priority of the current process.
6034 * @increment: priority increment
6036 * sys_setpriority is a more generic, but much slower function that
6037 * does similar things.
6039 SYSCALL_DEFINE1(nice, int, increment)
6044 * Setpriority might change our priority at the same moment.
6045 * We don't have to worry. Conceptually one call occurs first
6046 * and we have a single winner.
6048 if (increment < -40)
6053 nice = TASK_NICE(current) + increment;
6059 if (increment < 0 && !can_nice(current, nice))
6062 retval = security_task_setnice(current, nice);
6066 set_user_nice(current, nice);
6073 * task_prio - return the priority value of a given task.
6074 * @p: the task in question.
6076 * This is the priority value as seen by users in /proc.
6077 * RT tasks are offset by -200. Normal tasks are centered
6078 * around 0, value goes from -16 to +15.
6080 int task_prio(const struct task_struct *p)
6082 return p->prio - MAX_RT_PRIO;
6086 * task_nice - return the nice value of a given task.
6087 * @p: the task in question.
6089 int task_nice(const struct task_struct *p)
6091 return TASK_NICE(p);
6093 EXPORT_SYMBOL(task_nice);
6096 * idle_cpu - is a given cpu idle currently?
6097 * @cpu: the processor in question.
6099 int idle_cpu(int cpu)
6101 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6105 * idle_task - return the idle task for a given cpu.
6106 * @cpu: the processor in question.
6108 struct task_struct *idle_task(int cpu)
6110 return cpu_rq(cpu)->idle;
6114 * find_process_by_pid - find a process with a matching PID value.
6115 * @pid: the pid in question.
6117 static struct task_struct *find_process_by_pid(pid_t pid)
6119 return pid ? find_task_by_vpid(pid) : current;
6122 /* Actually do priority change: must hold rq lock. */
6124 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6126 BUG_ON(p->se.on_rq);
6129 switch (p->policy) {
6133 p->sched_class = &fair_sched_class;
6137 p->sched_class = &rt_sched_class;
6141 p->rt_priority = prio;
6142 p->normal_prio = normal_prio(p);
6143 /* we are holding p->pi_lock already */
6144 p->prio = rt_mutex_getprio(p);
6149 * check the target process has a UID that matches the current process's
6151 static bool check_same_owner(struct task_struct *p)
6153 const struct cred *cred = current_cred(), *pcred;
6157 pcred = __task_cred(p);
6158 match = (cred->euid == pcred->euid ||
6159 cred->euid == pcred->uid);
6164 static int __sched_setscheduler(struct task_struct *p, int policy,
6165 struct sched_param *param, bool user)
6167 int retval, oldprio, oldpolicy = -1, on_rq, running;
6168 unsigned long flags;
6169 const struct sched_class *prev_class = p->sched_class;
6173 /* may grab non-irq protected spin_locks */
6174 BUG_ON(in_interrupt());
6176 /* double check policy once rq lock held */
6178 reset_on_fork = p->sched_reset_on_fork;
6179 policy = oldpolicy = p->policy;
6181 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6182 policy &= ~SCHED_RESET_ON_FORK;
6184 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6185 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6186 policy != SCHED_IDLE)
6191 * Valid priorities for SCHED_FIFO and SCHED_RR are
6192 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6193 * SCHED_BATCH and SCHED_IDLE is 0.
6195 if (param->sched_priority < 0 ||
6196 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6197 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6199 if (rt_policy(policy) != (param->sched_priority != 0))
6203 * Allow unprivileged RT tasks to decrease priority:
6205 if (user && !capable(CAP_SYS_NICE)) {
6206 if (rt_policy(policy)) {
6207 unsigned long rlim_rtprio;
6209 if (!lock_task_sighand(p, &flags))
6211 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6212 unlock_task_sighand(p, &flags);
6214 /* can't set/change the rt policy */
6215 if (policy != p->policy && !rlim_rtprio)
6218 /* can't increase priority */
6219 if (param->sched_priority > p->rt_priority &&
6220 param->sched_priority > rlim_rtprio)
6224 * Like positive nice levels, dont allow tasks to
6225 * move out of SCHED_IDLE either:
6227 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6230 /* can't change other user's priorities */
6231 if (!check_same_owner(p))
6234 /* Normal users shall not reset the sched_reset_on_fork flag */
6235 if (p->sched_reset_on_fork && !reset_on_fork)
6240 #ifdef CONFIG_RT_GROUP_SCHED
6242 * Do not allow realtime tasks into groups that have no runtime
6245 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6246 task_group(p)->rt_bandwidth.rt_runtime == 0)
6250 retval = security_task_setscheduler(p, policy, param);
6256 * make sure no PI-waiters arrive (or leave) while we are
6257 * changing the priority of the task:
6259 spin_lock_irqsave(&p->pi_lock, flags);
6261 * To be able to change p->policy safely, the apropriate
6262 * runqueue lock must be held.
6264 rq = __task_rq_lock(p);
6265 /* recheck policy now with rq lock held */
6266 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6267 policy = oldpolicy = -1;
6268 __task_rq_unlock(rq);
6269 spin_unlock_irqrestore(&p->pi_lock, flags);
6272 update_rq_clock(rq);
6273 on_rq = p->se.on_rq;
6274 running = task_current(rq, p);
6276 deactivate_task(rq, p, 0);
6278 p->sched_class->put_prev_task(rq, p);
6280 p->sched_reset_on_fork = reset_on_fork;
6283 __setscheduler(rq, p, policy, param->sched_priority);
6286 p->sched_class->set_curr_task(rq);
6288 activate_task(rq, p, 0);
6290 check_class_changed(rq, p, prev_class, oldprio, running);
6292 __task_rq_unlock(rq);
6293 spin_unlock_irqrestore(&p->pi_lock, flags);
6295 rt_mutex_adjust_pi(p);
6301 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6302 * @p: the task in question.
6303 * @policy: new policy.
6304 * @param: structure containing the new RT priority.
6306 * NOTE that the task may be already dead.
6308 int sched_setscheduler(struct task_struct *p, int policy,
6309 struct sched_param *param)
6311 return __sched_setscheduler(p, policy, param, true);
6313 EXPORT_SYMBOL_GPL(sched_setscheduler);
6316 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6317 * @p: the task in question.
6318 * @policy: new policy.
6319 * @param: structure containing the new RT priority.
6321 * Just like sched_setscheduler, only don't bother checking if the
6322 * current context has permission. For example, this is needed in
6323 * stop_machine(): we create temporary high priority worker threads,
6324 * but our caller might not have that capability.
6326 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6327 struct sched_param *param)
6329 return __sched_setscheduler(p, policy, param, false);
6333 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6335 struct sched_param lparam;
6336 struct task_struct *p;
6339 if (!param || pid < 0)
6341 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6346 p = find_process_by_pid(pid);
6348 retval = sched_setscheduler(p, policy, &lparam);
6355 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6356 * @pid: the pid in question.
6357 * @policy: new policy.
6358 * @param: structure containing the new RT priority.
6360 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6361 struct sched_param __user *, param)
6363 /* negative values for policy are not valid */
6367 return do_sched_setscheduler(pid, policy, param);
6371 * sys_sched_setparam - set/change the RT priority of a thread
6372 * @pid: the pid in question.
6373 * @param: structure containing the new RT priority.
6375 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6377 return do_sched_setscheduler(pid, -1, param);
6381 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6382 * @pid: the pid in question.
6384 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6386 struct task_struct *p;
6393 read_lock(&tasklist_lock);
6394 p = find_process_by_pid(pid);
6396 retval = security_task_getscheduler(p);
6399 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6401 read_unlock(&tasklist_lock);
6406 * sys_sched_getparam - get the RT priority of a thread
6407 * @pid: the pid in question.
6408 * @param: structure containing the RT priority.
6410 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6412 struct sched_param lp;
6413 struct task_struct *p;
6416 if (!param || pid < 0)
6419 read_lock(&tasklist_lock);
6420 p = find_process_by_pid(pid);
6425 retval = security_task_getscheduler(p);
6429 lp.sched_priority = p->rt_priority;
6430 read_unlock(&tasklist_lock);
6433 * This one might sleep, we cannot do it with a spinlock held ...
6435 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6440 read_unlock(&tasklist_lock);
6444 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6446 cpumask_var_t cpus_allowed, new_mask;
6447 struct task_struct *p;
6451 read_lock(&tasklist_lock);
6453 p = find_process_by_pid(pid);
6455 read_unlock(&tasklist_lock);
6461 * It is not safe to call set_cpus_allowed with the
6462 * tasklist_lock held. We will bump the task_struct's
6463 * usage count and then drop tasklist_lock.
6466 read_unlock(&tasklist_lock);
6468 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6472 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6474 goto out_free_cpus_allowed;
6477 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6480 retval = security_task_setscheduler(p, 0, NULL);
6484 cpuset_cpus_allowed(p, cpus_allowed);
6485 cpumask_and(new_mask, in_mask, cpus_allowed);
6487 retval = set_cpus_allowed_ptr(p, new_mask);
6490 cpuset_cpus_allowed(p, cpus_allowed);
6491 if (!cpumask_subset(new_mask, cpus_allowed)) {
6493 * We must have raced with a concurrent cpuset
6494 * update. Just reset the cpus_allowed to the
6495 * cpuset's cpus_allowed
6497 cpumask_copy(new_mask, cpus_allowed);
6502 free_cpumask_var(new_mask);
6503 out_free_cpus_allowed:
6504 free_cpumask_var(cpus_allowed);
6511 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6512 struct cpumask *new_mask)
6514 if (len < cpumask_size())
6515 cpumask_clear(new_mask);
6516 else if (len > cpumask_size())
6517 len = cpumask_size();
6519 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6523 * sys_sched_setaffinity - set the cpu affinity of a process
6524 * @pid: pid of the process
6525 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6526 * @user_mask_ptr: user-space pointer to the new cpu mask
6528 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6529 unsigned long __user *, user_mask_ptr)
6531 cpumask_var_t new_mask;
6534 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6537 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6539 retval = sched_setaffinity(pid, new_mask);
6540 free_cpumask_var(new_mask);
6544 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6546 struct task_struct *p;
6550 read_lock(&tasklist_lock);
6553 p = find_process_by_pid(pid);
6557 retval = security_task_getscheduler(p);
6561 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6564 read_unlock(&tasklist_lock);
6571 * sys_sched_getaffinity - get the cpu affinity of a process
6572 * @pid: pid of the process
6573 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6574 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6576 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6577 unsigned long __user *, user_mask_ptr)
6582 if (len < cpumask_size())
6585 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6588 ret = sched_getaffinity(pid, mask);
6590 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6593 ret = cpumask_size();
6595 free_cpumask_var(mask);
6601 * sys_sched_yield - yield the current processor to other threads.
6603 * This function yields the current CPU to other tasks. If there are no
6604 * other threads running on this CPU then this function will return.
6606 SYSCALL_DEFINE0(sched_yield)
6608 struct rq *rq = this_rq_lock();
6610 schedstat_inc(rq, yld_count);
6611 current->sched_class->yield_task(rq);
6614 * Since we are going to call schedule() anyway, there's
6615 * no need to preempt or enable interrupts:
6617 __release(rq->lock);
6618 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6619 _raw_spin_unlock(&rq->lock);
6620 preempt_enable_no_resched();
6627 static inline int should_resched(void)
6629 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6632 static void __cond_resched(void)
6634 add_preempt_count(PREEMPT_ACTIVE);
6636 sub_preempt_count(PREEMPT_ACTIVE);
6639 int __sched _cond_resched(void)
6641 if (should_resched()) {
6647 EXPORT_SYMBOL(_cond_resched);
6650 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6651 * call schedule, and on return reacquire the lock.
6653 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6654 * operations here to prevent schedule() from being called twice (once via
6655 * spin_unlock(), once by hand).
6657 int __cond_resched_lock(spinlock_t *lock)
6659 int resched = should_resched();
6662 lockdep_assert_held(lock);
6664 if (spin_needbreak(lock) || resched) {
6675 EXPORT_SYMBOL(__cond_resched_lock);
6677 int __sched __cond_resched_softirq(void)
6679 BUG_ON(!in_softirq());
6681 if (should_resched()) {
6689 EXPORT_SYMBOL(__cond_resched_softirq);
6692 * yield - yield the current processor to other threads.
6694 * This is a shortcut for kernel-space yielding - it marks the
6695 * thread runnable and calls sys_sched_yield().
6697 void __sched yield(void)
6699 set_current_state(TASK_RUNNING);
6702 EXPORT_SYMBOL(yield);
6705 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6706 * that process accounting knows that this is a task in IO wait state.
6708 * But don't do that if it is a deliberate, throttling IO wait (this task
6709 * has set its backing_dev_info: the queue against which it should throttle)
6711 void __sched io_schedule(void)
6713 struct rq *rq = raw_rq();
6715 delayacct_blkio_start();
6716 atomic_inc(&rq->nr_iowait);
6717 current->in_iowait = 1;
6719 current->in_iowait = 0;
6720 atomic_dec(&rq->nr_iowait);
6721 delayacct_blkio_end();
6723 EXPORT_SYMBOL(io_schedule);
6725 long __sched io_schedule_timeout(long timeout)
6727 struct rq *rq = raw_rq();
6730 delayacct_blkio_start();
6731 atomic_inc(&rq->nr_iowait);
6732 current->in_iowait = 1;
6733 ret = schedule_timeout(timeout);
6734 current->in_iowait = 0;
6735 atomic_dec(&rq->nr_iowait);
6736 delayacct_blkio_end();
6741 * sys_sched_get_priority_max - return maximum RT priority.
6742 * @policy: scheduling class.
6744 * this syscall returns the maximum rt_priority that can be used
6745 * by a given scheduling class.
6747 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6754 ret = MAX_USER_RT_PRIO-1;
6766 * sys_sched_get_priority_min - return minimum RT priority.
6767 * @policy: scheduling class.
6769 * this syscall returns the minimum rt_priority that can be used
6770 * by a given scheduling class.
6772 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6790 * sys_sched_rr_get_interval - return the default timeslice of a process.
6791 * @pid: pid of the process.
6792 * @interval: userspace pointer to the timeslice value.
6794 * this syscall writes the default timeslice value of a given process
6795 * into the user-space timespec buffer. A value of '0' means infinity.
6797 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6798 struct timespec __user *, interval)
6800 struct task_struct *p;
6801 unsigned int time_slice;
6809 read_lock(&tasklist_lock);
6810 p = find_process_by_pid(pid);
6814 retval = security_task_getscheduler(p);
6819 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6820 * tasks that are on an otherwise idle runqueue:
6823 if (p->policy == SCHED_RR) {
6824 time_slice = DEF_TIMESLICE;
6825 } else if (p->policy != SCHED_FIFO) {
6826 struct sched_entity *se = &p->se;
6827 unsigned long flags;
6830 rq = task_rq_lock(p, &flags);
6831 if (rq->cfs.load.weight)
6832 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6833 task_rq_unlock(rq, &flags);
6835 read_unlock(&tasklist_lock);
6836 jiffies_to_timespec(time_slice, &t);
6837 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6841 read_unlock(&tasklist_lock);
6845 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6847 void sched_show_task(struct task_struct *p)
6849 unsigned long free = 0;
6852 state = p->state ? __ffs(p->state) + 1 : 0;
6853 printk(KERN_INFO "%-13.13s %c", p->comm,
6854 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6855 #if BITS_PER_LONG == 32
6856 if (state == TASK_RUNNING)
6857 printk(KERN_CONT " running ");
6859 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6861 if (state == TASK_RUNNING)
6862 printk(KERN_CONT " running task ");
6864 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6866 #ifdef CONFIG_DEBUG_STACK_USAGE
6867 free = stack_not_used(p);
6869 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6870 task_pid_nr(p), task_pid_nr(p->real_parent),
6871 (unsigned long)task_thread_info(p)->flags);
6873 show_stack(p, NULL);
6876 void show_state_filter(unsigned long state_filter)
6878 struct task_struct *g, *p;
6880 #if BITS_PER_LONG == 32
6882 " task PC stack pid father\n");
6885 " task PC stack pid father\n");
6887 read_lock(&tasklist_lock);
6888 do_each_thread(g, p) {
6890 * reset the NMI-timeout, listing all files on a slow
6891 * console might take alot of time:
6893 touch_nmi_watchdog();
6894 if (!state_filter || (p->state & state_filter))
6896 } while_each_thread(g, p);
6898 touch_all_softlockup_watchdogs();
6900 #ifdef CONFIG_SCHED_DEBUG
6901 sysrq_sched_debug_show();
6903 read_unlock(&tasklist_lock);
6905 * Only show locks if all tasks are dumped:
6907 if (state_filter == -1)
6908 debug_show_all_locks();
6911 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6913 idle->sched_class = &idle_sched_class;
6917 * init_idle - set up an idle thread for a given CPU
6918 * @idle: task in question
6919 * @cpu: cpu the idle task belongs to
6921 * NOTE: this function does not set the idle thread's NEED_RESCHED
6922 * flag, to make booting more robust.
6924 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6926 struct rq *rq = cpu_rq(cpu);
6927 unsigned long flags;
6929 spin_lock_irqsave(&rq->lock, flags);
6932 idle->se.exec_start = sched_clock();
6934 idle->prio = idle->normal_prio = MAX_PRIO;
6935 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6936 __set_task_cpu(idle, cpu);
6938 rq->curr = rq->idle = idle;
6939 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6942 spin_unlock_irqrestore(&rq->lock, flags);
6944 /* Set the preempt count _outside_ the spinlocks! */
6945 #if defined(CONFIG_PREEMPT)
6946 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6948 task_thread_info(idle)->preempt_count = 0;
6951 * The idle tasks have their own, simple scheduling class:
6953 idle->sched_class = &idle_sched_class;
6954 ftrace_graph_init_task(idle);
6958 * In a system that switches off the HZ timer nohz_cpu_mask
6959 * indicates which cpus entered this state. This is used
6960 * in the rcu update to wait only for active cpus. For system
6961 * which do not switch off the HZ timer nohz_cpu_mask should
6962 * always be CPU_BITS_NONE.
6964 cpumask_var_t nohz_cpu_mask;
6967 * Increase the granularity value when there are more CPUs,
6968 * because with more CPUs the 'effective latency' as visible
6969 * to users decreases. But the relationship is not linear,
6970 * so pick a second-best guess by going with the log2 of the
6973 * This idea comes from the SD scheduler of Con Kolivas:
6975 static inline void sched_init_granularity(void)
6977 unsigned int factor = 1 + ilog2(num_online_cpus());
6978 const unsigned long limit = 200000000;
6980 sysctl_sched_min_granularity *= factor;
6981 if (sysctl_sched_min_granularity > limit)
6982 sysctl_sched_min_granularity = limit;
6984 sysctl_sched_latency *= factor;
6985 if (sysctl_sched_latency > limit)
6986 sysctl_sched_latency = limit;
6988 sysctl_sched_wakeup_granularity *= factor;
6990 sysctl_sched_shares_ratelimit *= factor;
6995 * This is how migration works:
6997 * 1) we queue a struct migration_req structure in the source CPU's
6998 * runqueue and wake up that CPU's migration thread.
6999 * 2) we down() the locked semaphore => thread blocks.
7000 * 3) migration thread wakes up (implicitly it forces the migrated
7001 * thread off the CPU)
7002 * 4) it gets the migration request and checks whether the migrated
7003 * task is still in the wrong runqueue.
7004 * 5) if it's in the wrong runqueue then the migration thread removes
7005 * it and puts it into the right queue.
7006 * 6) migration thread up()s the semaphore.
7007 * 7) we wake up and the migration is done.
7011 * Change a given task's CPU affinity. Migrate the thread to a
7012 * proper CPU and schedule it away if the CPU it's executing on
7013 * is removed from the allowed bitmask.
7015 * NOTE: the caller must have a valid reference to the task, the
7016 * task must not exit() & deallocate itself prematurely. The
7017 * call is not atomic; no spinlocks may be held.
7019 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7021 struct migration_req req;
7022 unsigned long flags;
7026 rq = task_rq_lock(p, &flags);
7027 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
7032 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7033 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7038 if (p->sched_class->set_cpus_allowed)
7039 p->sched_class->set_cpus_allowed(p, new_mask);
7041 cpumask_copy(&p->cpus_allowed, new_mask);
7042 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7045 /* Can the task run on the task's current CPU? If so, we're done */
7046 if (cpumask_test_cpu(task_cpu(p), new_mask))
7049 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
7050 /* Need help from migration thread: drop lock and wait. */
7051 struct task_struct *mt = rq->migration_thread;
7053 get_task_struct(mt);
7054 task_rq_unlock(rq, &flags);
7055 wake_up_process(rq->migration_thread);
7056 put_task_struct(mt);
7057 wait_for_completion(&req.done);
7058 tlb_migrate_finish(p->mm);
7062 task_rq_unlock(rq, &flags);
7066 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7069 * Move (not current) task off this cpu, onto dest cpu. We're doing
7070 * this because either it can't run here any more (set_cpus_allowed()
7071 * away from this CPU, or CPU going down), or because we're
7072 * attempting to rebalance this task on exec (sched_exec).
7074 * So we race with normal scheduler movements, but that's OK, as long
7075 * as the task is no longer on this CPU.
7077 * Returns non-zero if task was successfully migrated.
7079 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7081 struct rq *rq_dest, *rq_src;
7084 if (unlikely(!cpu_active(dest_cpu)))
7087 rq_src = cpu_rq(src_cpu);
7088 rq_dest = cpu_rq(dest_cpu);
7090 double_rq_lock(rq_src, rq_dest);
7091 /* Already moved. */
7092 if (task_cpu(p) != src_cpu)
7094 /* Affinity changed (again). */
7095 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7098 on_rq = p->se.on_rq;
7100 deactivate_task(rq_src, p, 0);
7102 set_task_cpu(p, dest_cpu);
7104 activate_task(rq_dest, p, 0);
7105 check_preempt_curr(rq_dest, p, 0);
7110 double_rq_unlock(rq_src, rq_dest);
7114 #define RCU_MIGRATION_IDLE 0
7115 #define RCU_MIGRATION_NEED_QS 1
7116 #define RCU_MIGRATION_GOT_QS 2
7117 #define RCU_MIGRATION_MUST_SYNC 3
7120 * migration_thread - this is a highprio system thread that performs
7121 * thread migration by bumping thread off CPU then 'pushing' onto
7124 static int migration_thread(void *data)
7127 int cpu = (long)data;
7131 BUG_ON(rq->migration_thread != current);
7133 set_current_state(TASK_INTERRUPTIBLE);
7134 while (!kthread_should_stop()) {
7135 struct migration_req *req;
7136 struct list_head *head;
7138 spin_lock_irq(&rq->lock);
7140 if (cpu_is_offline(cpu)) {
7141 spin_unlock_irq(&rq->lock);
7145 if (rq->active_balance) {
7146 active_load_balance(rq, cpu);
7147 rq->active_balance = 0;
7150 head = &rq->migration_queue;
7152 if (list_empty(head)) {
7153 spin_unlock_irq(&rq->lock);
7155 set_current_state(TASK_INTERRUPTIBLE);
7158 req = list_entry(head->next, struct migration_req, list);
7159 list_del_init(head->next);
7161 if (req->task != NULL) {
7162 spin_unlock(&rq->lock);
7163 __migrate_task(req->task, cpu, req->dest_cpu);
7164 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7165 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7166 spin_unlock(&rq->lock);
7168 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7169 spin_unlock(&rq->lock);
7170 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7174 complete(&req->done);
7176 __set_current_state(TASK_RUNNING);
7181 #ifdef CONFIG_HOTPLUG_CPU
7183 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7187 local_irq_disable();
7188 ret = __migrate_task(p, src_cpu, dest_cpu);
7194 * Figure out where task on dead CPU should go, use force if necessary.
7196 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7199 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7202 /* Look for allowed, online CPU in same node. */
7203 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7204 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7207 /* Any allowed, online CPU? */
7208 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7209 if (dest_cpu < nr_cpu_ids)
7212 /* No more Mr. Nice Guy. */
7213 if (dest_cpu >= nr_cpu_ids) {
7214 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7215 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7218 * Don't tell them about moving exiting tasks or
7219 * kernel threads (both mm NULL), since they never
7222 if (p->mm && printk_ratelimit()) {
7223 printk(KERN_INFO "process %d (%s) no "
7224 "longer affine to cpu%d\n",
7225 task_pid_nr(p), p->comm, dead_cpu);
7230 /* It can have affinity changed while we were choosing. */
7231 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7236 * While a dead CPU has no uninterruptible tasks queued at this point,
7237 * it might still have a nonzero ->nr_uninterruptible counter, because
7238 * for performance reasons the counter is not stricly tracking tasks to
7239 * their home CPUs. So we just add the counter to another CPU's counter,
7240 * to keep the global sum constant after CPU-down:
7242 static void migrate_nr_uninterruptible(struct rq *rq_src)
7244 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7245 unsigned long flags;
7247 local_irq_save(flags);
7248 double_rq_lock(rq_src, rq_dest);
7249 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7250 rq_src->nr_uninterruptible = 0;
7251 double_rq_unlock(rq_src, rq_dest);
7252 local_irq_restore(flags);
7255 /* Run through task list and migrate tasks from the dead cpu. */
7256 static void migrate_live_tasks(int src_cpu)
7258 struct task_struct *p, *t;
7260 read_lock(&tasklist_lock);
7262 do_each_thread(t, p) {
7266 if (task_cpu(p) == src_cpu)
7267 move_task_off_dead_cpu(src_cpu, p);
7268 } while_each_thread(t, p);
7270 read_unlock(&tasklist_lock);
7274 * Schedules idle task to be the next runnable task on current CPU.
7275 * It does so by boosting its priority to highest possible.
7276 * Used by CPU offline code.
7278 void sched_idle_next(void)
7280 int this_cpu = smp_processor_id();
7281 struct rq *rq = cpu_rq(this_cpu);
7282 struct task_struct *p = rq->idle;
7283 unsigned long flags;
7285 /* cpu has to be offline */
7286 BUG_ON(cpu_online(this_cpu));
7289 * Strictly not necessary since rest of the CPUs are stopped by now
7290 * and interrupts disabled on the current cpu.
7292 spin_lock_irqsave(&rq->lock, flags);
7294 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7296 update_rq_clock(rq);
7297 activate_task(rq, p, 0);
7299 spin_unlock_irqrestore(&rq->lock, flags);
7303 * Ensures that the idle task is using init_mm right before its cpu goes
7306 void idle_task_exit(void)
7308 struct mm_struct *mm = current->active_mm;
7310 BUG_ON(cpu_online(smp_processor_id()));
7313 switch_mm(mm, &init_mm, current);
7317 /* called under rq->lock with disabled interrupts */
7318 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7320 struct rq *rq = cpu_rq(dead_cpu);
7322 /* Must be exiting, otherwise would be on tasklist. */
7323 BUG_ON(!p->exit_state);
7325 /* Cannot have done final schedule yet: would have vanished. */
7326 BUG_ON(p->state == TASK_DEAD);
7331 * Drop lock around migration; if someone else moves it,
7332 * that's OK. No task can be added to this CPU, so iteration is
7335 spin_unlock_irq(&rq->lock);
7336 move_task_off_dead_cpu(dead_cpu, p);
7337 spin_lock_irq(&rq->lock);
7342 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7343 static void migrate_dead_tasks(unsigned int dead_cpu)
7345 struct rq *rq = cpu_rq(dead_cpu);
7346 struct task_struct *next;
7349 if (!rq->nr_running)
7351 update_rq_clock(rq);
7352 next = pick_next_task(rq);
7355 next->sched_class->put_prev_task(rq, next);
7356 migrate_dead(dead_cpu, next);
7362 * remove the tasks which were accounted by rq from calc_load_tasks.
7364 static void calc_global_load_remove(struct rq *rq)
7366 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7367 rq->calc_load_active = 0;
7369 #endif /* CONFIG_HOTPLUG_CPU */
7371 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7373 static struct ctl_table sd_ctl_dir[] = {
7375 .procname = "sched_domain",
7381 static struct ctl_table sd_ctl_root[] = {
7383 .ctl_name = CTL_KERN,
7384 .procname = "kernel",
7386 .child = sd_ctl_dir,
7391 static struct ctl_table *sd_alloc_ctl_entry(int n)
7393 struct ctl_table *entry =
7394 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7399 static void sd_free_ctl_entry(struct ctl_table **tablep)
7401 struct ctl_table *entry;
7404 * In the intermediate directories, both the child directory and
7405 * procname are dynamically allocated and could fail but the mode
7406 * will always be set. In the lowest directory the names are
7407 * static strings and all have proc handlers.
7409 for (entry = *tablep; entry->mode; entry++) {
7411 sd_free_ctl_entry(&entry->child);
7412 if (entry->proc_handler == NULL)
7413 kfree(entry->procname);
7421 set_table_entry(struct ctl_table *entry,
7422 const char *procname, void *data, int maxlen,
7423 mode_t mode, proc_handler *proc_handler)
7425 entry->procname = procname;
7427 entry->maxlen = maxlen;
7429 entry->proc_handler = proc_handler;
7432 static struct ctl_table *
7433 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7435 struct ctl_table *table = sd_alloc_ctl_entry(13);
7440 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7441 sizeof(long), 0644, proc_doulongvec_minmax);
7442 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7443 sizeof(long), 0644, proc_doulongvec_minmax);
7444 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7445 sizeof(int), 0644, proc_dointvec_minmax);
7446 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7447 sizeof(int), 0644, proc_dointvec_minmax);
7448 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7449 sizeof(int), 0644, proc_dointvec_minmax);
7450 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7451 sizeof(int), 0644, proc_dointvec_minmax);
7452 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7453 sizeof(int), 0644, proc_dointvec_minmax);
7454 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7455 sizeof(int), 0644, proc_dointvec_minmax);
7456 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7457 sizeof(int), 0644, proc_dointvec_minmax);
7458 set_table_entry(&table[9], "cache_nice_tries",
7459 &sd->cache_nice_tries,
7460 sizeof(int), 0644, proc_dointvec_minmax);
7461 set_table_entry(&table[10], "flags", &sd->flags,
7462 sizeof(int), 0644, proc_dointvec_minmax);
7463 set_table_entry(&table[11], "name", sd->name,
7464 CORENAME_MAX_SIZE, 0444, proc_dostring);
7465 /* &table[12] is terminator */
7470 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7472 struct ctl_table *entry, *table;
7473 struct sched_domain *sd;
7474 int domain_num = 0, i;
7477 for_each_domain(cpu, sd)
7479 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7484 for_each_domain(cpu, sd) {
7485 snprintf(buf, 32, "domain%d", i);
7486 entry->procname = kstrdup(buf, GFP_KERNEL);
7488 entry->child = sd_alloc_ctl_domain_table(sd);
7495 static struct ctl_table_header *sd_sysctl_header;
7496 static void register_sched_domain_sysctl(void)
7498 int i, cpu_num = num_online_cpus();
7499 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7502 WARN_ON(sd_ctl_dir[0].child);
7503 sd_ctl_dir[0].child = entry;
7508 for_each_online_cpu(i) {
7509 snprintf(buf, 32, "cpu%d", i);
7510 entry->procname = kstrdup(buf, GFP_KERNEL);
7512 entry->child = sd_alloc_ctl_cpu_table(i);
7516 WARN_ON(sd_sysctl_header);
7517 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7520 /* may be called multiple times per register */
7521 static void unregister_sched_domain_sysctl(void)
7523 if (sd_sysctl_header)
7524 unregister_sysctl_table(sd_sysctl_header);
7525 sd_sysctl_header = NULL;
7526 if (sd_ctl_dir[0].child)
7527 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7530 static void register_sched_domain_sysctl(void)
7533 static void unregister_sched_domain_sysctl(void)
7538 static void set_rq_online(struct rq *rq)
7541 const struct sched_class *class;
7543 cpumask_set_cpu(rq->cpu, rq->rd->online);
7546 for_each_class(class) {
7547 if (class->rq_online)
7548 class->rq_online(rq);
7553 static void set_rq_offline(struct rq *rq)
7556 const struct sched_class *class;
7558 for_each_class(class) {
7559 if (class->rq_offline)
7560 class->rq_offline(rq);
7563 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7569 * migration_call - callback that gets triggered when a CPU is added.
7570 * Here we can start up the necessary migration thread for the new CPU.
7572 static int __cpuinit
7573 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7575 struct task_struct *p;
7576 int cpu = (long)hcpu;
7577 unsigned long flags;
7582 case CPU_UP_PREPARE:
7583 case CPU_UP_PREPARE_FROZEN:
7584 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7587 kthread_bind(p, cpu);
7588 /* Must be high prio: stop_machine expects to yield to it. */
7589 rq = task_rq_lock(p, &flags);
7590 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7591 task_rq_unlock(rq, &flags);
7593 cpu_rq(cpu)->migration_thread = p;
7594 rq->calc_load_update = calc_load_update;
7598 case CPU_ONLINE_FROZEN:
7599 /* Strictly unnecessary, as first user will wake it. */
7600 wake_up_process(cpu_rq(cpu)->migration_thread);
7602 /* Update our root-domain */
7604 spin_lock_irqsave(&rq->lock, flags);
7606 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7610 spin_unlock_irqrestore(&rq->lock, flags);
7613 #ifdef CONFIG_HOTPLUG_CPU
7614 case CPU_UP_CANCELED:
7615 case CPU_UP_CANCELED_FROZEN:
7616 if (!cpu_rq(cpu)->migration_thread)
7618 /* Unbind it from offline cpu so it can run. Fall thru. */
7619 kthread_bind(cpu_rq(cpu)->migration_thread,
7620 cpumask_any(cpu_online_mask));
7621 kthread_stop(cpu_rq(cpu)->migration_thread);
7622 put_task_struct(cpu_rq(cpu)->migration_thread);
7623 cpu_rq(cpu)->migration_thread = NULL;
7627 case CPU_DEAD_FROZEN:
7628 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7629 migrate_live_tasks(cpu);
7631 kthread_stop(rq->migration_thread);
7632 put_task_struct(rq->migration_thread);
7633 rq->migration_thread = NULL;
7634 /* Idle task back to normal (off runqueue, low prio) */
7635 spin_lock_irq(&rq->lock);
7636 update_rq_clock(rq);
7637 deactivate_task(rq, rq->idle, 0);
7638 rq->idle->static_prio = MAX_PRIO;
7639 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7640 rq->idle->sched_class = &idle_sched_class;
7641 migrate_dead_tasks(cpu);
7642 spin_unlock_irq(&rq->lock);
7644 migrate_nr_uninterruptible(rq);
7645 BUG_ON(rq->nr_running != 0);
7646 calc_global_load_remove(rq);
7648 * No need to migrate the tasks: it was best-effort if
7649 * they didn't take sched_hotcpu_mutex. Just wake up
7652 spin_lock_irq(&rq->lock);
7653 while (!list_empty(&rq->migration_queue)) {
7654 struct migration_req *req;
7656 req = list_entry(rq->migration_queue.next,
7657 struct migration_req, list);
7658 list_del_init(&req->list);
7659 spin_unlock_irq(&rq->lock);
7660 complete(&req->done);
7661 spin_lock_irq(&rq->lock);
7663 spin_unlock_irq(&rq->lock);
7667 case CPU_DYING_FROZEN:
7668 /* Update our root-domain */
7670 spin_lock_irqsave(&rq->lock, flags);
7672 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7675 spin_unlock_irqrestore(&rq->lock, flags);
7683 * Register at high priority so that task migration (migrate_all_tasks)
7684 * happens before everything else. This has to be lower priority than
7685 * the notifier in the perf_counter subsystem, though.
7687 static struct notifier_block __cpuinitdata migration_notifier = {
7688 .notifier_call = migration_call,
7692 static int __init migration_init(void)
7694 void *cpu = (void *)(long)smp_processor_id();
7697 /* Start one for the boot CPU: */
7698 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7699 BUG_ON(err == NOTIFY_BAD);
7700 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7701 register_cpu_notifier(&migration_notifier);
7705 early_initcall(migration_init);
7710 #ifdef CONFIG_SCHED_DEBUG
7712 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7713 struct cpumask *groupmask)
7715 struct sched_group *group = sd->groups;
7718 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7719 cpumask_clear(groupmask);
7721 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7723 if (!(sd->flags & SD_LOAD_BALANCE)) {
7724 printk("does not load-balance\n");
7726 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7731 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7733 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7734 printk(KERN_ERR "ERROR: domain->span does not contain "
7737 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7738 printk(KERN_ERR "ERROR: domain->groups does not contain"
7742 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7746 printk(KERN_ERR "ERROR: group is NULL\n");
7750 if (!group->cpu_power) {
7751 printk(KERN_CONT "\n");
7752 printk(KERN_ERR "ERROR: domain->cpu_power not "
7757 if (!cpumask_weight(sched_group_cpus(group))) {
7758 printk(KERN_CONT "\n");
7759 printk(KERN_ERR "ERROR: empty group\n");
7763 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7764 printk(KERN_CONT "\n");
7765 printk(KERN_ERR "ERROR: repeated CPUs\n");
7769 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7771 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7773 printk(KERN_CONT " %s", str);
7774 if (group->cpu_power != SCHED_LOAD_SCALE) {
7775 printk(KERN_CONT " (cpu_power = %d)",
7779 group = group->next;
7780 } while (group != sd->groups);
7781 printk(KERN_CONT "\n");
7783 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7784 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7787 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7788 printk(KERN_ERR "ERROR: parent span is not a superset "
7789 "of domain->span\n");
7793 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7795 cpumask_var_t groupmask;
7799 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7803 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7805 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7806 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7811 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7818 free_cpumask_var(groupmask);
7820 #else /* !CONFIG_SCHED_DEBUG */
7821 # define sched_domain_debug(sd, cpu) do { } while (0)
7822 #endif /* CONFIG_SCHED_DEBUG */
7824 static int sd_degenerate(struct sched_domain *sd)
7826 if (cpumask_weight(sched_domain_span(sd)) == 1)
7829 /* Following flags need at least 2 groups */
7830 if (sd->flags & (SD_LOAD_BALANCE |
7831 SD_BALANCE_NEWIDLE |
7835 SD_SHARE_PKG_RESOURCES)) {
7836 if (sd->groups != sd->groups->next)
7840 /* Following flags don't use groups */
7841 if (sd->flags & (SD_WAKE_AFFINE))
7848 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7850 unsigned long cflags = sd->flags, pflags = parent->flags;
7852 if (sd_degenerate(parent))
7855 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7858 /* Flags needing groups don't count if only 1 group in parent */
7859 if (parent->groups == parent->groups->next) {
7860 pflags &= ~(SD_LOAD_BALANCE |
7861 SD_BALANCE_NEWIDLE |
7865 SD_SHARE_PKG_RESOURCES);
7866 if (nr_node_ids == 1)
7867 pflags &= ~SD_SERIALIZE;
7869 if (~cflags & pflags)
7875 static void free_rootdomain(struct root_domain *rd)
7877 cpupri_cleanup(&rd->cpupri);
7879 free_cpumask_var(rd->rto_mask);
7880 free_cpumask_var(rd->online);
7881 free_cpumask_var(rd->span);
7885 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7887 struct root_domain *old_rd = NULL;
7888 unsigned long flags;
7890 spin_lock_irqsave(&rq->lock, flags);
7895 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7898 cpumask_clear_cpu(rq->cpu, old_rd->span);
7901 * If we dont want to free the old_rt yet then
7902 * set old_rd to NULL to skip the freeing later
7905 if (!atomic_dec_and_test(&old_rd->refcount))
7909 atomic_inc(&rd->refcount);
7912 cpumask_set_cpu(rq->cpu, rd->span);
7913 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7916 spin_unlock_irqrestore(&rq->lock, flags);
7919 free_rootdomain(old_rd);
7922 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7924 gfp_t gfp = GFP_KERNEL;
7926 memset(rd, 0, sizeof(*rd));
7931 if (!alloc_cpumask_var(&rd->span, gfp))
7933 if (!alloc_cpumask_var(&rd->online, gfp))
7935 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
7938 if (cpupri_init(&rd->cpupri, bootmem) != 0)
7943 free_cpumask_var(rd->rto_mask);
7945 free_cpumask_var(rd->online);
7947 free_cpumask_var(rd->span);
7952 static void init_defrootdomain(void)
7954 init_rootdomain(&def_root_domain, true);
7956 atomic_set(&def_root_domain.refcount, 1);
7959 static struct root_domain *alloc_rootdomain(void)
7961 struct root_domain *rd;
7963 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7967 if (init_rootdomain(rd, false) != 0) {
7976 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7977 * hold the hotplug lock.
7980 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7982 struct rq *rq = cpu_rq(cpu);
7983 struct sched_domain *tmp;
7985 /* Remove the sched domains which do not contribute to scheduling. */
7986 for (tmp = sd; tmp; ) {
7987 struct sched_domain *parent = tmp->parent;
7991 if (sd_parent_degenerate(tmp, parent)) {
7992 tmp->parent = parent->parent;
7994 parent->parent->child = tmp;
7999 if (sd && sd_degenerate(sd)) {
8005 sched_domain_debug(sd, cpu);
8007 rq_attach_root(rq, rd);
8008 rcu_assign_pointer(rq->sd, sd);
8011 /* cpus with isolated domains */
8012 static cpumask_var_t cpu_isolated_map;
8014 /* Setup the mask of cpus configured for isolated domains */
8015 static int __init isolated_cpu_setup(char *str)
8017 cpulist_parse(str, cpu_isolated_map);
8021 __setup("isolcpus=", isolated_cpu_setup);
8024 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8025 * to a function which identifies what group(along with sched group) a CPU
8026 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8027 * (due to the fact that we keep track of groups covered with a struct cpumask).
8029 * init_sched_build_groups will build a circular linked list of the groups
8030 * covered by the given span, and will set each group's ->cpumask correctly,
8031 * and ->cpu_power to 0.
8034 init_sched_build_groups(const struct cpumask *span,
8035 const struct cpumask *cpu_map,
8036 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8037 struct sched_group **sg,
8038 struct cpumask *tmpmask),
8039 struct cpumask *covered, struct cpumask *tmpmask)
8041 struct sched_group *first = NULL, *last = NULL;
8044 cpumask_clear(covered);
8046 for_each_cpu(i, span) {
8047 struct sched_group *sg;
8048 int group = group_fn(i, cpu_map, &sg, tmpmask);
8051 if (cpumask_test_cpu(i, covered))
8054 cpumask_clear(sched_group_cpus(sg));
8057 for_each_cpu(j, span) {
8058 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8061 cpumask_set_cpu(j, covered);
8062 cpumask_set_cpu(j, sched_group_cpus(sg));
8073 #define SD_NODES_PER_DOMAIN 16
8078 * find_next_best_node - find the next node to include in a sched_domain
8079 * @node: node whose sched_domain we're building
8080 * @used_nodes: nodes already in the sched_domain
8082 * Find the next node to include in a given scheduling domain. Simply
8083 * finds the closest node not already in the @used_nodes map.
8085 * Should use nodemask_t.
8087 static int find_next_best_node(int node, nodemask_t *used_nodes)
8089 int i, n, val, min_val, best_node = 0;
8093 for (i = 0; i < nr_node_ids; i++) {
8094 /* Start at @node */
8095 n = (node + i) % nr_node_ids;
8097 if (!nr_cpus_node(n))
8100 /* Skip already used nodes */
8101 if (node_isset(n, *used_nodes))
8104 /* Simple min distance search */
8105 val = node_distance(node, n);
8107 if (val < min_val) {
8113 node_set(best_node, *used_nodes);
8118 * sched_domain_node_span - get a cpumask for a node's sched_domain
8119 * @node: node whose cpumask we're constructing
8120 * @span: resulting cpumask
8122 * Given a node, construct a good cpumask for its sched_domain to span. It
8123 * should be one that prevents unnecessary balancing, but also spreads tasks
8126 static void sched_domain_node_span(int node, struct cpumask *span)
8128 nodemask_t used_nodes;
8131 cpumask_clear(span);
8132 nodes_clear(used_nodes);
8134 cpumask_or(span, span, cpumask_of_node(node));
8135 node_set(node, used_nodes);
8137 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8138 int next_node = find_next_best_node(node, &used_nodes);
8140 cpumask_or(span, span, cpumask_of_node(next_node));
8143 #endif /* CONFIG_NUMA */
8145 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8148 * The cpus mask in sched_group and sched_domain hangs off the end.
8150 * ( See the the comments in include/linux/sched.h:struct sched_group
8151 * and struct sched_domain. )
8153 struct static_sched_group {
8154 struct sched_group sg;
8155 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8158 struct static_sched_domain {
8159 struct sched_domain sd;
8160 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8166 cpumask_var_t domainspan;
8167 cpumask_var_t covered;
8168 cpumask_var_t notcovered;
8170 cpumask_var_t nodemask;
8171 cpumask_var_t this_sibling_map;
8172 cpumask_var_t this_core_map;
8173 cpumask_var_t send_covered;
8174 cpumask_var_t tmpmask;
8175 struct sched_group **sched_group_nodes;
8176 struct root_domain *rd;
8180 sa_sched_groups = 0,
8185 sa_this_sibling_map,
8187 sa_sched_group_nodes,
8197 * SMT sched-domains:
8199 #ifdef CONFIG_SCHED_SMT
8200 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8201 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8204 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8205 struct sched_group **sg, struct cpumask *unused)
8208 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8211 #endif /* CONFIG_SCHED_SMT */
8214 * multi-core sched-domains:
8216 #ifdef CONFIG_SCHED_MC
8217 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8218 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8219 #endif /* CONFIG_SCHED_MC */
8221 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8223 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8224 struct sched_group **sg, struct cpumask *mask)
8228 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8229 group = cpumask_first(mask);
8231 *sg = &per_cpu(sched_group_core, group).sg;
8234 #elif defined(CONFIG_SCHED_MC)
8236 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8237 struct sched_group **sg, struct cpumask *unused)
8240 *sg = &per_cpu(sched_group_core, cpu).sg;
8245 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8246 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8249 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8250 struct sched_group **sg, struct cpumask *mask)
8253 #ifdef CONFIG_SCHED_MC
8254 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8255 group = cpumask_first(mask);
8256 #elif defined(CONFIG_SCHED_SMT)
8257 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8258 group = cpumask_first(mask);
8263 *sg = &per_cpu(sched_group_phys, group).sg;
8269 * The init_sched_build_groups can't handle what we want to do with node
8270 * groups, so roll our own. Now each node has its own list of groups which
8271 * gets dynamically allocated.
8273 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8274 static struct sched_group ***sched_group_nodes_bycpu;
8276 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8277 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8279 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8280 struct sched_group **sg,
8281 struct cpumask *nodemask)
8285 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8286 group = cpumask_first(nodemask);
8289 *sg = &per_cpu(sched_group_allnodes, group).sg;
8293 static void init_numa_sched_groups_power(struct sched_group *group_head)
8295 struct sched_group *sg = group_head;
8301 for_each_cpu(j, sched_group_cpus(sg)) {
8302 struct sched_domain *sd;
8304 sd = &per_cpu(phys_domains, j).sd;
8305 if (j != group_first_cpu(sd->groups)) {
8307 * Only add "power" once for each
8313 sg->cpu_power += sd->groups->cpu_power;
8316 } while (sg != group_head);
8319 static int build_numa_sched_groups(struct s_data *d,
8320 const struct cpumask *cpu_map, int num)
8322 struct sched_domain *sd;
8323 struct sched_group *sg, *prev;
8326 cpumask_clear(d->covered);
8327 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8328 if (cpumask_empty(d->nodemask)) {
8329 d->sched_group_nodes[num] = NULL;
8333 sched_domain_node_span(num, d->domainspan);
8334 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8336 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8339 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8343 d->sched_group_nodes[num] = sg;
8345 for_each_cpu(j, d->nodemask) {
8346 sd = &per_cpu(node_domains, j).sd;
8351 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8353 cpumask_or(d->covered, d->covered, d->nodemask);
8356 for (j = 0; j < nr_node_ids; j++) {
8357 n = (num + j) % nr_node_ids;
8358 cpumask_complement(d->notcovered, d->covered);
8359 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8360 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8361 if (cpumask_empty(d->tmpmask))
8363 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8364 if (cpumask_empty(d->tmpmask))
8366 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8370 "Can not alloc domain group for node %d\n", j);
8374 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8375 sg->next = prev->next;
8376 cpumask_or(d->covered, d->covered, d->tmpmask);
8383 #endif /* CONFIG_NUMA */
8386 /* Free memory allocated for various sched_group structures */
8387 static void free_sched_groups(const struct cpumask *cpu_map,
8388 struct cpumask *nodemask)
8392 for_each_cpu(cpu, cpu_map) {
8393 struct sched_group **sched_group_nodes
8394 = sched_group_nodes_bycpu[cpu];
8396 if (!sched_group_nodes)
8399 for (i = 0; i < nr_node_ids; i++) {
8400 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8402 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8403 if (cpumask_empty(nodemask))
8413 if (oldsg != sched_group_nodes[i])
8416 kfree(sched_group_nodes);
8417 sched_group_nodes_bycpu[cpu] = NULL;
8420 #else /* !CONFIG_NUMA */
8421 static void free_sched_groups(const struct cpumask *cpu_map,
8422 struct cpumask *nodemask)
8425 #endif /* CONFIG_NUMA */
8428 * Initialize sched groups cpu_power.
8430 * cpu_power indicates the capacity of sched group, which is used while
8431 * distributing the load between different sched groups in a sched domain.
8432 * Typically cpu_power for all the groups in a sched domain will be same unless
8433 * there are asymmetries in the topology. If there are asymmetries, group
8434 * having more cpu_power will pickup more load compared to the group having
8437 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8439 struct sched_domain *child;
8440 struct sched_group *group;
8444 WARN_ON(!sd || !sd->groups);
8446 if (cpu != group_first_cpu(sd->groups))
8451 sd->groups->cpu_power = 0;
8454 power = SCHED_LOAD_SCALE;
8455 weight = cpumask_weight(sched_domain_span(sd));
8457 * SMT siblings share the power of a single core.
8458 * Usually multiple threads get a better yield out of
8459 * that one core than a single thread would have,
8460 * reflect that in sd->smt_gain.
8462 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8463 power *= sd->smt_gain;
8465 power >>= SCHED_LOAD_SHIFT;
8467 sd->groups->cpu_power += power;
8472 * Add cpu_power of each child group to this groups cpu_power.
8474 group = child->groups;
8476 sd->groups->cpu_power += group->cpu_power;
8477 group = group->next;
8478 } while (group != child->groups);
8482 * Initializers for schedule domains
8483 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8486 #ifdef CONFIG_SCHED_DEBUG
8487 # define SD_INIT_NAME(sd, type) sd->name = #type
8489 # define SD_INIT_NAME(sd, type) do { } while (0)
8492 #define SD_INIT(sd, type) sd_init_##type(sd)
8494 #define SD_INIT_FUNC(type) \
8495 static noinline void sd_init_##type(struct sched_domain *sd) \
8497 memset(sd, 0, sizeof(*sd)); \
8498 *sd = SD_##type##_INIT; \
8499 sd->level = SD_LV_##type; \
8500 SD_INIT_NAME(sd, type); \
8505 SD_INIT_FUNC(ALLNODES)
8508 #ifdef CONFIG_SCHED_SMT
8509 SD_INIT_FUNC(SIBLING)
8511 #ifdef CONFIG_SCHED_MC
8515 static int default_relax_domain_level = -1;
8517 static int __init setup_relax_domain_level(char *str)
8521 val = simple_strtoul(str, NULL, 0);
8522 if (val < SD_LV_MAX)
8523 default_relax_domain_level = val;
8527 __setup("relax_domain_level=", setup_relax_domain_level);
8529 static void set_domain_attribute(struct sched_domain *sd,
8530 struct sched_domain_attr *attr)
8534 if (!attr || attr->relax_domain_level < 0) {
8535 if (default_relax_domain_level < 0)
8538 request = default_relax_domain_level;
8540 request = attr->relax_domain_level;
8541 if (request < sd->level) {
8542 /* turn off idle balance on this domain */
8543 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8545 /* turn on idle balance on this domain */
8546 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8550 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8551 const struct cpumask *cpu_map)
8554 case sa_sched_groups:
8555 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8556 d->sched_group_nodes = NULL;
8558 free_rootdomain(d->rd); /* fall through */
8560 free_cpumask_var(d->tmpmask); /* fall through */
8561 case sa_send_covered:
8562 free_cpumask_var(d->send_covered); /* fall through */
8563 case sa_this_core_map:
8564 free_cpumask_var(d->this_core_map); /* fall through */
8565 case sa_this_sibling_map:
8566 free_cpumask_var(d->this_sibling_map); /* fall through */
8568 free_cpumask_var(d->nodemask); /* fall through */
8569 case sa_sched_group_nodes:
8571 kfree(d->sched_group_nodes); /* fall through */
8573 free_cpumask_var(d->notcovered); /* fall through */
8575 free_cpumask_var(d->covered); /* fall through */
8577 free_cpumask_var(d->domainspan); /* fall through */
8584 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8585 const struct cpumask *cpu_map)
8588 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8590 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8591 return sa_domainspan;
8592 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8594 /* Allocate the per-node list of sched groups */
8595 d->sched_group_nodes = kcalloc(nr_node_ids,
8596 sizeof(struct sched_group *), GFP_KERNEL);
8597 if (!d->sched_group_nodes) {
8598 printk(KERN_WARNING "Can not alloc sched group node list\n");
8599 return sa_notcovered;
8601 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8603 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8604 return sa_sched_group_nodes;
8605 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8607 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8608 return sa_this_sibling_map;
8609 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8610 return sa_this_core_map;
8611 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8612 return sa_send_covered;
8613 d->rd = alloc_rootdomain();
8615 printk(KERN_WARNING "Cannot alloc root domain\n");
8618 return sa_rootdomain;
8621 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8622 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8624 struct sched_domain *sd = NULL;
8626 struct sched_domain *parent;
8629 if (cpumask_weight(cpu_map) >
8630 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8631 sd = &per_cpu(allnodes_domains, i).sd;
8632 SD_INIT(sd, ALLNODES);
8633 set_domain_attribute(sd, attr);
8634 cpumask_copy(sched_domain_span(sd), cpu_map);
8635 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8640 sd = &per_cpu(node_domains, i).sd;
8642 set_domain_attribute(sd, attr);
8643 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8644 sd->parent = parent;
8647 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8652 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8653 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8654 struct sched_domain *parent, int i)
8656 struct sched_domain *sd;
8657 sd = &per_cpu(phys_domains, i).sd;
8659 set_domain_attribute(sd, attr);
8660 cpumask_copy(sched_domain_span(sd), d->nodemask);
8661 sd->parent = parent;
8664 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8668 static struct sched_domain *__build_mc_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_MC
8674 sd = &per_cpu(core_domains, i).sd;
8676 set_domain_attribute(sd, attr);
8677 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8678 sd->parent = parent;
8680 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8685 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8686 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8687 struct sched_domain *parent, int i)
8689 struct sched_domain *sd = parent;
8690 #ifdef CONFIG_SCHED_SMT
8691 sd = &per_cpu(cpu_domains, i).sd;
8692 SD_INIT(sd, SIBLING);
8693 set_domain_attribute(sd, attr);
8694 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8695 sd->parent = parent;
8697 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8702 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8703 const struct cpumask *cpu_map, int cpu)
8706 #ifdef CONFIG_SCHED_SMT
8707 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8708 cpumask_and(d->this_sibling_map, cpu_map,
8709 topology_thread_cpumask(cpu));
8710 if (cpu == cpumask_first(d->this_sibling_map))
8711 init_sched_build_groups(d->this_sibling_map, cpu_map,
8713 d->send_covered, d->tmpmask);
8716 #ifdef CONFIG_SCHED_MC
8717 case SD_LV_MC: /* set up multi-core groups */
8718 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8719 if (cpu == cpumask_first(d->this_core_map))
8720 init_sched_build_groups(d->this_core_map, cpu_map,
8722 d->send_covered, d->tmpmask);
8725 case SD_LV_CPU: /* set up physical groups */
8726 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8727 if (!cpumask_empty(d->nodemask))
8728 init_sched_build_groups(d->nodemask, cpu_map,
8730 d->send_covered, d->tmpmask);
8733 case SD_LV_ALLNODES:
8734 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8735 d->send_covered, d->tmpmask);
8744 * Build sched domains for a given set of cpus and attach the sched domains
8745 * to the individual cpus
8747 static int __build_sched_domains(const struct cpumask *cpu_map,
8748 struct sched_domain_attr *attr)
8750 enum s_alloc alloc_state = sa_none;
8752 struct sched_domain *sd;
8758 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8759 if (alloc_state != sa_rootdomain)
8761 alloc_state = sa_sched_groups;
8764 * Set up domains for cpus specified by the cpu_map.
8766 for_each_cpu(i, cpu_map) {
8767 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8770 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8771 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8772 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8773 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8776 for_each_cpu(i, cpu_map) {
8777 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8778 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8781 /* Set up physical groups */
8782 for (i = 0; i < nr_node_ids; i++)
8783 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8786 /* Set up node groups */
8788 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8790 for (i = 0; i < nr_node_ids; i++)
8791 if (build_numa_sched_groups(&d, cpu_map, i))
8795 /* Calculate CPU power for physical packages and nodes */
8796 #ifdef CONFIG_SCHED_SMT
8797 for_each_cpu(i, cpu_map) {
8798 sd = &per_cpu(cpu_domains, i).sd;
8799 init_sched_groups_power(i, sd);
8802 #ifdef CONFIG_SCHED_MC
8803 for_each_cpu(i, cpu_map) {
8804 sd = &per_cpu(core_domains, i).sd;
8805 init_sched_groups_power(i, sd);
8809 for_each_cpu(i, cpu_map) {
8810 sd = &per_cpu(phys_domains, i).sd;
8811 init_sched_groups_power(i, sd);
8815 for (i = 0; i < nr_node_ids; i++)
8816 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8818 if (d.sd_allnodes) {
8819 struct sched_group *sg;
8821 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8823 init_numa_sched_groups_power(sg);
8827 /* Attach the domains */
8828 for_each_cpu(i, cpu_map) {
8829 #ifdef CONFIG_SCHED_SMT
8830 sd = &per_cpu(cpu_domains, i).sd;
8831 #elif defined(CONFIG_SCHED_MC)
8832 sd = &per_cpu(core_domains, i).sd;
8834 sd = &per_cpu(phys_domains, i).sd;
8836 cpu_attach_domain(sd, d.rd, i);
8839 d.sched_group_nodes = NULL; /* don't free this we still need it */
8840 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8844 __free_domain_allocs(&d, alloc_state, cpu_map);
8848 static int build_sched_domains(const struct cpumask *cpu_map)
8850 return __build_sched_domains(cpu_map, NULL);
8853 static struct cpumask *doms_cur; /* current sched domains */
8854 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8855 static struct sched_domain_attr *dattr_cur;
8856 /* attribues of custom domains in 'doms_cur' */
8859 * Special case: If a kmalloc of a doms_cur partition (array of
8860 * cpumask) fails, then fallback to a single sched domain,
8861 * as determined by the single cpumask fallback_doms.
8863 static cpumask_var_t fallback_doms;
8866 * arch_update_cpu_topology lets virtualized architectures update the
8867 * cpu core maps. It is supposed to return 1 if the topology changed
8868 * or 0 if it stayed the same.
8870 int __attribute__((weak)) arch_update_cpu_topology(void)
8876 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8877 * For now this just excludes isolated cpus, but could be used to
8878 * exclude other special cases in the future.
8880 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8884 arch_update_cpu_topology();
8886 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8888 doms_cur = fallback_doms;
8889 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8891 err = build_sched_domains(doms_cur);
8892 register_sched_domain_sysctl();
8897 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8898 struct cpumask *tmpmask)
8900 free_sched_groups(cpu_map, tmpmask);
8904 * Detach sched domains from a group of cpus specified in cpu_map
8905 * These cpus will now be attached to the NULL domain
8907 static void detach_destroy_domains(const struct cpumask *cpu_map)
8909 /* Save because hotplug lock held. */
8910 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8913 for_each_cpu(i, cpu_map)
8914 cpu_attach_domain(NULL, &def_root_domain, i);
8915 synchronize_sched();
8916 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8919 /* handle null as "default" */
8920 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8921 struct sched_domain_attr *new, int idx_new)
8923 struct sched_domain_attr tmp;
8930 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8931 new ? (new + idx_new) : &tmp,
8932 sizeof(struct sched_domain_attr));
8936 * Partition sched domains as specified by the 'ndoms_new'
8937 * cpumasks in the array doms_new[] of cpumasks. This compares
8938 * doms_new[] to the current sched domain partitioning, doms_cur[].
8939 * It destroys each deleted domain and builds each new domain.
8941 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8942 * The masks don't intersect (don't overlap.) We should setup one
8943 * sched domain for each mask. CPUs not in any of the cpumasks will
8944 * not be load balanced. If the same cpumask appears both in the
8945 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8948 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8949 * ownership of it and will kfree it when done with it. If the caller
8950 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8951 * ndoms_new == 1, and partition_sched_domains() will fallback to
8952 * the single partition 'fallback_doms', it also forces the domains
8955 * If doms_new == NULL it will be replaced with cpu_online_mask.
8956 * ndoms_new == 0 is a special case for destroying existing domains,
8957 * and it will not create the default domain.
8959 * Call with hotplug lock held
8961 /* FIXME: Change to struct cpumask *doms_new[] */
8962 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8963 struct sched_domain_attr *dattr_new)
8968 mutex_lock(&sched_domains_mutex);
8970 /* always unregister in case we don't destroy any domains */
8971 unregister_sched_domain_sysctl();
8973 /* Let architecture update cpu core mappings. */
8974 new_topology = arch_update_cpu_topology();
8976 n = doms_new ? ndoms_new : 0;
8978 /* Destroy deleted domains */
8979 for (i = 0; i < ndoms_cur; i++) {
8980 for (j = 0; j < n && !new_topology; j++) {
8981 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8982 && dattrs_equal(dattr_cur, i, dattr_new, j))
8985 /* no match - a current sched domain not in new doms_new[] */
8986 detach_destroy_domains(doms_cur + i);
8991 if (doms_new == NULL) {
8993 doms_new = fallback_doms;
8994 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8995 WARN_ON_ONCE(dattr_new);
8998 /* Build new domains */
8999 for (i = 0; i < ndoms_new; i++) {
9000 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9001 if (cpumask_equal(&doms_new[i], &doms_cur[j])
9002 && dattrs_equal(dattr_new, i, dattr_cur, j))
9005 /* no match - add a new doms_new */
9006 __build_sched_domains(doms_new + i,
9007 dattr_new ? dattr_new + i : NULL);
9012 /* Remember the new sched domains */
9013 if (doms_cur != fallback_doms)
9015 kfree(dattr_cur); /* kfree(NULL) is safe */
9016 doms_cur = doms_new;
9017 dattr_cur = dattr_new;
9018 ndoms_cur = ndoms_new;
9020 register_sched_domain_sysctl();
9022 mutex_unlock(&sched_domains_mutex);
9025 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9026 static void arch_reinit_sched_domains(void)
9030 /* Destroy domains first to force the rebuild */
9031 partition_sched_domains(0, NULL, NULL);
9033 rebuild_sched_domains();
9037 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9039 unsigned int level = 0;
9041 if (sscanf(buf, "%u", &level) != 1)
9045 * level is always be positive so don't check for
9046 * level < POWERSAVINGS_BALANCE_NONE which is 0
9047 * What happens on 0 or 1 byte write,
9048 * need to check for count as well?
9051 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9055 sched_smt_power_savings = level;
9057 sched_mc_power_savings = level;
9059 arch_reinit_sched_domains();
9064 #ifdef CONFIG_SCHED_MC
9065 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9068 return sprintf(page, "%u\n", sched_mc_power_savings);
9070 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9071 const char *buf, size_t count)
9073 return sched_power_savings_store(buf, count, 0);
9075 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9076 sched_mc_power_savings_show,
9077 sched_mc_power_savings_store);
9080 #ifdef CONFIG_SCHED_SMT
9081 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9084 return sprintf(page, "%u\n", sched_smt_power_savings);
9086 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9087 const char *buf, size_t count)
9089 return sched_power_savings_store(buf, count, 1);
9091 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9092 sched_smt_power_savings_show,
9093 sched_smt_power_savings_store);
9096 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9100 #ifdef CONFIG_SCHED_SMT
9102 err = sysfs_create_file(&cls->kset.kobj,
9103 &attr_sched_smt_power_savings.attr);
9105 #ifdef CONFIG_SCHED_MC
9106 if (!err && mc_capable())
9107 err = sysfs_create_file(&cls->kset.kobj,
9108 &attr_sched_mc_power_savings.attr);
9112 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9114 #ifndef CONFIG_CPUSETS
9116 * Add online and remove offline CPUs from the scheduler domains.
9117 * When cpusets are enabled they take over this function.
9119 static int update_sched_domains(struct notifier_block *nfb,
9120 unsigned long action, void *hcpu)
9124 case CPU_ONLINE_FROZEN:
9126 case CPU_DEAD_FROZEN:
9127 partition_sched_domains(1, NULL, NULL);
9136 static int update_runtime(struct notifier_block *nfb,
9137 unsigned long action, void *hcpu)
9139 int cpu = (int)(long)hcpu;
9142 case CPU_DOWN_PREPARE:
9143 case CPU_DOWN_PREPARE_FROZEN:
9144 disable_runtime(cpu_rq(cpu));
9147 case CPU_DOWN_FAILED:
9148 case CPU_DOWN_FAILED_FROZEN:
9150 case CPU_ONLINE_FROZEN:
9151 enable_runtime(cpu_rq(cpu));
9159 void __init sched_init_smp(void)
9161 cpumask_var_t non_isolated_cpus;
9163 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9165 #if defined(CONFIG_NUMA)
9166 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9168 BUG_ON(sched_group_nodes_bycpu == NULL);
9171 mutex_lock(&sched_domains_mutex);
9172 arch_init_sched_domains(cpu_online_mask);
9173 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9174 if (cpumask_empty(non_isolated_cpus))
9175 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9176 mutex_unlock(&sched_domains_mutex);
9179 #ifndef CONFIG_CPUSETS
9180 /* XXX: Theoretical race here - CPU may be hotplugged now */
9181 hotcpu_notifier(update_sched_domains, 0);
9184 /* RT runtime code needs to handle some hotplug events */
9185 hotcpu_notifier(update_runtime, 0);
9189 /* Move init over to a non-isolated CPU */
9190 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9192 sched_init_granularity();
9193 free_cpumask_var(non_isolated_cpus);
9195 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9196 init_sched_rt_class();
9199 void __init sched_init_smp(void)
9201 sched_init_granularity();
9203 #endif /* CONFIG_SMP */
9205 const_debug unsigned int sysctl_timer_migration = 1;
9207 int in_sched_functions(unsigned long addr)
9209 return in_lock_functions(addr) ||
9210 (addr >= (unsigned long)__sched_text_start
9211 && addr < (unsigned long)__sched_text_end);
9214 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9216 cfs_rq->tasks_timeline = RB_ROOT;
9217 INIT_LIST_HEAD(&cfs_rq->tasks);
9218 #ifdef CONFIG_FAIR_GROUP_SCHED
9221 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9224 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9226 struct rt_prio_array *array;
9229 array = &rt_rq->active;
9230 for (i = 0; i < MAX_RT_PRIO; i++) {
9231 INIT_LIST_HEAD(array->queue + i);
9232 __clear_bit(i, array->bitmap);
9234 /* delimiter for bitsearch: */
9235 __set_bit(MAX_RT_PRIO, array->bitmap);
9237 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9238 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9240 rt_rq->highest_prio.next = MAX_RT_PRIO;
9244 rt_rq->rt_nr_migratory = 0;
9245 rt_rq->overloaded = 0;
9246 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9250 rt_rq->rt_throttled = 0;
9251 rt_rq->rt_runtime = 0;
9252 spin_lock_init(&rt_rq->rt_runtime_lock);
9254 #ifdef CONFIG_RT_GROUP_SCHED
9255 rt_rq->rt_nr_boosted = 0;
9260 #ifdef CONFIG_FAIR_GROUP_SCHED
9261 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9262 struct sched_entity *se, int cpu, int add,
9263 struct sched_entity *parent)
9265 struct rq *rq = cpu_rq(cpu);
9266 tg->cfs_rq[cpu] = cfs_rq;
9267 init_cfs_rq(cfs_rq, rq);
9270 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9273 /* se could be NULL for init_task_group */
9278 se->cfs_rq = &rq->cfs;
9280 se->cfs_rq = parent->my_q;
9283 se->load.weight = tg->shares;
9284 se->load.inv_weight = 0;
9285 se->parent = parent;
9289 #ifdef CONFIG_RT_GROUP_SCHED
9290 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9291 struct sched_rt_entity *rt_se, int cpu, int add,
9292 struct sched_rt_entity *parent)
9294 struct rq *rq = cpu_rq(cpu);
9296 tg->rt_rq[cpu] = rt_rq;
9297 init_rt_rq(rt_rq, rq);
9299 rt_rq->rt_se = rt_se;
9300 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9302 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9304 tg->rt_se[cpu] = rt_se;
9309 rt_se->rt_rq = &rq->rt;
9311 rt_se->rt_rq = parent->my_q;
9313 rt_se->my_q = rt_rq;
9314 rt_se->parent = parent;
9315 INIT_LIST_HEAD(&rt_se->run_list);
9319 void __init sched_init(void)
9322 unsigned long alloc_size = 0, ptr;
9324 #ifdef CONFIG_FAIR_GROUP_SCHED
9325 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9327 #ifdef CONFIG_RT_GROUP_SCHED
9328 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9330 #ifdef CONFIG_USER_SCHED
9333 #ifdef CONFIG_CPUMASK_OFFSTACK
9334 alloc_size += num_possible_cpus() * cpumask_size();
9337 * As sched_init() is called before page_alloc is setup,
9338 * we use alloc_bootmem().
9341 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9343 #ifdef CONFIG_FAIR_GROUP_SCHED
9344 init_task_group.se = (struct sched_entity **)ptr;
9345 ptr += nr_cpu_ids * sizeof(void **);
9347 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9348 ptr += nr_cpu_ids * sizeof(void **);
9350 #ifdef CONFIG_USER_SCHED
9351 root_task_group.se = (struct sched_entity **)ptr;
9352 ptr += nr_cpu_ids * sizeof(void **);
9354 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9355 ptr += nr_cpu_ids * sizeof(void **);
9356 #endif /* CONFIG_USER_SCHED */
9357 #endif /* CONFIG_FAIR_GROUP_SCHED */
9358 #ifdef CONFIG_RT_GROUP_SCHED
9359 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9360 ptr += nr_cpu_ids * sizeof(void **);
9362 init_task_group.rt_rq = (struct rt_rq **)ptr;
9363 ptr += nr_cpu_ids * sizeof(void **);
9365 #ifdef CONFIG_USER_SCHED
9366 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9367 ptr += nr_cpu_ids * sizeof(void **);
9369 root_task_group.rt_rq = (struct rt_rq **)ptr;
9370 ptr += nr_cpu_ids * sizeof(void **);
9371 #endif /* CONFIG_USER_SCHED */
9372 #endif /* CONFIG_RT_GROUP_SCHED */
9373 #ifdef CONFIG_CPUMASK_OFFSTACK
9374 for_each_possible_cpu(i) {
9375 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9376 ptr += cpumask_size();
9378 #endif /* CONFIG_CPUMASK_OFFSTACK */
9382 init_defrootdomain();
9385 init_rt_bandwidth(&def_rt_bandwidth,
9386 global_rt_period(), global_rt_runtime());
9388 #ifdef CONFIG_RT_GROUP_SCHED
9389 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9390 global_rt_period(), global_rt_runtime());
9391 #ifdef CONFIG_USER_SCHED
9392 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9393 global_rt_period(), RUNTIME_INF);
9394 #endif /* CONFIG_USER_SCHED */
9395 #endif /* CONFIG_RT_GROUP_SCHED */
9397 #ifdef CONFIG_GROUP_SCHED
9398 list_add(&init_task_group.list, &task_groups);
9399 INIT_LIST_HEAD(&init_task_group.children);
9401 #ifdef CONFIG_USER_SCHED
9402 INIT_LIST_HEAD(&root_task_group.children);
9403 init_task_group.parent = &root_task_group;
9404 list_add(&init_task_group.siblings, &root_task_group.children);
9405 #endif /* CONFIG_USER_SCHED */
9406 #endif /* CONFIG_GROUP_SCHED */
9408 for_each_possible_cpu(i) {
9412 spin_lock_init(&rq->lock);
9414 rq->calc_load_active = 0;
9415 rq->calc_load_update = jiffies + LOAD_FREQ;
9416 init_cfs_rq(&rq->cfs, rq);
9417 init_rt_rq(&rq->rt, rq);
9418 #ifdef CONFIG_FAIR_GROUP_SCHED
9419 init_task_group.shares = init_task_group_load;
9420 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9421 #ifdef CONFIG_CGROUP_SCHED
9423 * How much cpu bandwidth does init_task_group get?
9425 * In case of task-groups formed thr' the cgroup filesystem, it
9426 * gets 100% of the cpu resources in the system. This overall
9427 * system cpu resource is divided among the tasks of
9428 * init_task_group and its child task-groups in a fair manner,
9429 * based on each entity's (task or task-group's) weight
9430 * (se->load.weight).
9432 * In other words, if init_task_group has 10 tasks of weight
9433 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9434 * then A0's share of the cpu resource is:
9436 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9438 * We achieve this by letting init_task_group's tasks sit
9439 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9441 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9442 #elif defined CONFIG_USER_SCHED
9443 root_task_group.shares = NICE_0_LOAD;
9444 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9446 * In case of task-groups formed thr' the user id of tasks,
9447 * init_task_group represents tasks belonging to root user.
9448 * Hence it forms a sibling of all subsequent groups formed.
9449 * In this case, init_task_group gets only a fraction of overall
9450 * system cpu resource, based on the weight assigned to root
9451 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9452 * by letting tasks of init_task_group sit in a separate cfs_rq
9453 * (init_tg_cfs_rq) and having one entity represent this group of
9454 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9456 init_tg_cfs_entry(&init_task_group,
9457 &per_cpu(init_tg_cfs_rq, i),
9458 &per_cpu(init_sched_entity, i), i, 1,
9459 root_task_group.se[i]);
9462 #endif /* CONFIG_FAIR_GROUP_SCHED */
9464 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9465 #ifdef CONFIG_RT_GROUP_SCHED
9466 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9467 #ifdef CONFIG_CGROUP_SCHED
9468 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9469 #elif defined CONFIG_USER_SCHED
9470 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9471 init_tg_rt_entry(&init_task_group,
9472 &per_cpu(init_rt_rq, i),
9473 &per_cpu(init_sched_rt_entity, i), i, 1,
9474 root_task_group.rt_se[i]);
9478 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9479 rq->cpu_load[j] = 0;
9483 rq->post_schedule = 0;
9484 rq->active_balance = 0;
9485 rq->next_balance = jiffies;
9489 rq->migration_thread = NULL;
9490 INIT_LIST_HEAD(&rq->migration_queue);
9491 rq_attach_root(rq, &def_root_domain);
9494 atomic_set(&rq->nr_iowait, 0);
9497 set_load_weight(&init_task);
9499 #ifdef CONFIG_PREEMPT_NOTIFIERS
9500 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9504 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9507 #ifdef CONFIG_RT_MUTEXES
9508 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9512 * The boot idle thread does lazy MMU switching as well:
9514 atomic_inc(&init_mm.mm_count);
9515 enter_lazy_tlb(&init_mm, current);
9518 * Make us the idle thread. Technically, schedule() should not be
9519 * called from this thread, however somewhere below it might be,
9520 * but because we are the idle thread, we just pick up running again
9521 * when this runqueue becomes "idle".
9523 init_idle(current, smp_processor_id());
9525 calc_load_update = jiffies + LOAD_FREQ;
9528 * During early bootup we pretend to be a normal task:
9530 current->sched_class = &fair_sched_class;
9532 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9533 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9536 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9537 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9539 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9542 perf_counter_init();
9544 scheduler_running = 1;
9547 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9548 static inline int preempt_count_equals(int preempt_offset)
9550 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9552 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9555 void __might_sleep(char *file, int line, int preempt_offset)
9558 static unsigned long prev_jiffy; /* ratelimiting */
9560 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9561 system_state != SYSTEM_RUNNING || oops_in_progress)
9563 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9565 prev_jiffy = jiffies;
9568 "BUG: sleeping function called from invalid context at %s:%d\n",
9571 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9572 in_atomic(), irqs_disabled(),
9573 current->pid, current->comm);
9575 debug_show_held_locks(current);
9576 if (irqs_disabled())
9577 print_irqtrace_events(current);
9581 EXPORT_SYMBOL(__might_sleep);
9584 #ifdef CONFIG_MAGIC_SYSRQ
9585 static void normalize_task(struct rq *rq, struct task_struct *p)
9589 update_rq_clock(rq);
9590 on_rq = p->se.on_rq;
9592 deactivate_task(rq, p, 0);
9593 __setscheduler(rq, p, SCHED_NORMAL, 0);
9595 activate_task(rq, p, 0);
9596 resched_task(rq->curr);
9600 void normalize_rt_tasks(void)
9602 struct task_struct *g, *p;
9603 unsigned long flags;
9606 read_lock_irqsave(&tasklist_lock, flags);
9607 do_each_thread(g, p) {
9609 * Only normalize user tasks:
9614 p->se.exec_start = 0;
9615 #ifdef CONFIG_SCHEDSTATS
9616 p->se.wait_start = 0;
9617 p->se.sleep_start = 0;
9618 p->se.block_start = 0;
9623 * Renice negative nice level userspace
9626 if (TASK_NICE(p) < 0 && p->mm)
9627 set_user_nice(p, 0);
9631 spin_lock(&p->pi_lock);
9632 rq = __task_rq_lock(p);
9634 normalize_task(rq, p);
9636 __task_rq_unlock(rq);
9637 spin_unlock(&p->pi_lock);
9638 } while_each_thread(g, p);
9640 read_unlock_irqrestore(&tasklist_lock, flags);
9643 #endif /* CONFIG_MAGIC_SYSRQ */
9647 * These functions are only useful for the IA64 MCA handling.
9649 * They can only be called when the whole system has been
9650 * stopped - every CPU needs to be quiescent, and no scheduling
9651 * activity can take place. Using them for anything else would
9652 * be a serious bug, and as a result, they aren't even visible
9653 * under any other configuration.
9657 * curr_task - return the current task for a given cpu.
9658 * @cpu: the processor in question.
9660 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9662 struct task_struct *curr_task(int cpu)
9664 return cpu_curr(cpu);
9668 * set_curr_task - set the current task for a given cpu.
9669 * @cpu: the processor in question.
9670 * @p: the task pointer to set.
9672 * Description: This function must only be used when non-maskable interrupts
9673 * are serviced on a separate stack. It allows the architecture to switch the
9674 * notion of the current task on a cpu in a non-blocking manner. This function
9675 * must be called with all CPU's synchronized, and interrupts disabled, the
9676 * and caller must save the original value of the current task (see
9677 * curr_task() above) and restore that value before reenabling interrupts and
9678 * re-starting the system.
9680 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9682 void set_curr_task(int cpu, struct task_struct *p)
9689 #ifdef CONFIG_FAIR_GROUP_SCHED
9690 static void free_fair_sched_group(struct task_group *tg)
9694 for_each_possible_cpu(i) {
9696 kfree(tg->cfs_rq[i]);
9706 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9708 struct cfs_rq *cfs_rq;
9709 struct sched_entity *se;
9713 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9716 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9720 tg->shares = NICE_0_LOAD;
9722 for_each_possible_cpu(i) {
9725 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9726 GFP_KERNEL, cpu_to_node(i));
9730 se = kzalloc_node(sizeof(struct sched_entity),
9731 GFP_KERNEL, cpu_to_node(i));
9735 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9744 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9746 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9747 &cpu_rq(cpu)->leaf_cfs_rq_list);
9750 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9752 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9754 #else /* !CONFG_FAIR_GROUP_SCHED */
9755 static inline void free_fair_sched_group(struct task_group *tg)
9760 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9765 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9769 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9772 #endif /* CONFIG_FAIR_GROUP_SCHED */
9774 #ifdef CONFIG_RT_GROUP_SCHED
9775 static void free_rt_sched_group(struct task_group *tg)
9779 destroy_rt_bandwidth(&tg->rt_bandwidth);
9781 for_each_possible_cpu(i) {
9783 kfree(tg->rt_rq[i]);
9785 kfree(tg->rt_se[i]);
9793 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9795 struct rt_rq *rt_rq;
9796 struct sched_rt_entity *rt_se;
9800 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9803 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9807 init_rt_bandwidth(&tg->rt_bandwidth,
9808 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9810 for_each_possible_cpu(i) {
9813 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9814 GFP_KERNEL, cpu_to_node(i));
9818 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9819 GFP_KERNEL, cpu_to_node(i));
9823 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9832 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9834 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9835 &cpu_rq(cpu)->leaf_rt_rq_list);
9838 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9840 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9842 #else /* !CONFIG_RT_GROUP_SCHED */
9843 static inline void free_rt_sched_group(struct task_group *tg)
9848 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9853 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9857 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9860 #endif /* CONFIG_RT_GROUP_SCHED */
9862 #ifdef CONFIG_GROUP_SCHED
9863 static void free_sched_group(struct task_group *tg)
9865 free_fair_sched_group(tg);
9866 free_rt_sched_group(tg);
9870 /* allocate runqueue etc for a new task group */
9871 struct task_group *sched_create_group(struct task_group *parent)
9873 struct task_group *tg;
9874 unsigned long flags;
9877 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9879 return ERR_PTR(-ENOMEM);
9881 if (!alloc_fair_sched_group(tg, parent))
9884 if (!alloc_rt_sched_group(tg, parent))
9887 spin_lock_irqsave(&task_group_lock, flags);
9888 for_each_possible_cpu(i) {
9889 register_fair_sched_group(tg, i);
9890 register_rt_sched_group(tg, i);
9892 list_add_rcu(&tg->list, &task_groups);
9894 WARN_ON(!parent); /* root should already exist */
9896 tg->parent = parent;
9897 INIT_LIST_HEAD(&tg->children);
9898 list_add_rcu(&tg->siblings, &parent->children);
9899 spin_unlock_irqrestore(&task_group_lock, flags);
9904 free_sched_group(tg);
9905 return ERR_PTR(-ENOMEM);
9908 /* rcu callback to free various structures associated with a task group */
9909 static void free_sched_group_rcu(struct rcu_head *rhp)
9911 /* now it should be safe to free those cfs_rqs */
9912 free_sched_group(container_of(rhp, struct task_group, rcu));
9915 /* Destroy runqueue etc associated with a task group */
9916 void sched_destroy_group(struct task_group *tg)
9918 unsigned long flags;
9921 spin_lock_irqsave(&task_group_lock, flags);
9922 for_each_possible_cpu(i) {
9923 unregister_fair_sched_group(tg, i);
9924 unregister_rt_sched_group(tg, i);
9926 list_del_rcu(&tg->list);
9927 list_del_rcu(&tg->siblings);
9928 spin_unlock_irqrestore(&task_group_lock, flags);
9930 /* wait for possible concurrent references to cfs_rqs complete */
9931 call_rcu(&tg->rcu, free_sched_group_rcu);
9934 /* change task's runqueue when it moves between groups.
9935 * The caller of this function should have put the task in its new group
9936 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9937 * reflect its new group.
9939 void sched_move_task(struct task_struct *tsk)
9942 unsigned long flags;
9945 rq = task_rq_lock(tsk, &flags);
9947 update_rq_clock(rq);
9949 running = task_current(rq, tsk);
9950 on_rq = tsk->se.on_rq;
9953 dequeue_task(rq, tsk, 0);
9954 if (unlikely(running))
9955 tsk->sched_class->put_prev_task(rq, tsk);
9957 set_task_rq(tsk, task_cpu(tsk));
9959 #ifdef CONFIG_FAIR_GROUP_SCHED
9960 if (tsk->sched_class->moved_group)
9961 tsk->sched_class->moved_group(tsk);
9964 if (unlikely(running))
9965 tsk->sched_class->set_curr_task(rq);
9967 enqueue_task(rq, tsk, 0);
9969 task_rq_unlock(rq, &flags);
9971 #endif /* CONFIG_GROUP_SCHED */
9973 #ifdef CONFIG_FAIR_GROUP_SCHED
9974 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9976 struct cfs_rq *cfs_rq = se->cfs_rq;
9981 dequeue_entity(cfs_rq, se, 0);
9983 se->load.weight = shares;
9984 se->load.inv_weight = 0;
9987 enqueue_entity(cfs_rq, se, 0);
9990 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9992 struct cfs_rq *cfs_rq = se->cfs_rq;
9993 struct rq *rq = cfs_rq->rq;
9994 unsigned long flags;
9996 spin_lock_irqsave(&rq->lock, flags);
9997 __set_se_shares(se, shares);
9998 spin_unlock_irqrestore(&rq->lock, flags);
10001 static DEFINE_MUTEX(shares_mutex);
10003 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10006 unsigned long flags;
10009 * We can't change the weight of the root cgroup.
10014 if (shares < MIN_SHARES)
10015 shares = MIN_SHARES;
10016 else if (shares > MAX_SHARES)
10017 shares = MAX_SHARES;
10019 mutex_lock(&shares_mutex);
10020 if (tg->shares == shares)
10023 spin_lock_irqsave(&task_group_lock, flags);
10024 for_each_possible_cpu(i)
10025 unregister_fair_sched_group(tg, i);
10026 list_del_rcu(&tg->siblings);
10027 spin_unlock_irqrestore(&task_group_lock, flags);
10029 /* wait for any ongoing reference to this group to finish */
10030 synchronize_sched();
10033 * Now we are free to modify the group's share on each cpu
10034 * w/o tripping rebalance_share or load_balance_fair.
10036 tg->shares = shares;
10037 for_each_possible_cpu(i) {
10039 * force a rebalance
10041 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10042 set_se_shares(tg->se[i], shares);
10046 * Enable load balance activity on this group, by inserting it back on
10047 * each cpu's rq->leaf_cfs_rq_list.
10049 spin_lock_irqsave(&task_group_lock, flags);
10050 for_each_possible_cpu(i)
10051 register_fair_sched_group(tg, i);
10052 list_add_rcu(&tg->siblings, &tg->parent->children);
10053 spin_unlock_irqrestore(&task_group_lock, flags);
10055 mutex_unlock(&shares_mutex);
10059 unsigned long sched_group_shares(struct task_group *tg)
10065 #ifdef CONFIG_RT_GROUP_SCHED
10067 * Ensure that the real time constraints are schedulable.
10069 static DEFINE_MUTEX(rt_constraints_mutex);
10071 static unsigned long to_ratio(u64 period, u64 runtime)
10073 if (runtime == RUNTIME_INF)
10076 return div64_u64(runtime << 20, period);
10079 /* Must be called with tasklist_lock held */
10080 static inline int tg_has_rt_tasks(struct task_group *tg)
10082 struct task_struct *g, *p;
10084 do_each_thread(g, p) {
10085 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10087 } while_each_thread(g, p);
10092 struct rt_schedulable_data {
10093 struct task_group *tg;
10098 static int tg_schedulable(struct task_group *tg, void *data)
10100 struct rt_schedulable_data *d = data;
10101 struct task_group *child;
10102 unsigned long total, sum = 0;
10103 u64 period, runtime;
10105 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10106 runtime = tg->rt_bandwidth.rt_runtime;
10109 period = d->rt_period;
10110 runtime = d->rt_runtime;
10113 #ifdef CONFIG_USER_SCHED
10114 if (tg == &root_task_group) {
10115 period = global_rt_period();
10116 runtime = global_rt_runtime();
10121 * Cannot have more runtime than the period.
10123 if (runtime > period && runtime != RUNTIME_INF)
10127 * Ensure we don't starve existing RT tasks.
10129 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10132 total = to_ratio(period, runtime);
10135 * Nobody can have more than the global setting allows.
10137 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10141 * The sum of our children's runtime should not exceed our own.
10143 list_for_each_entry_rcu(child, &tg->children, siblings) {
10144 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10145 runtime = child->rt_bandwidth.rt_runtime;
10147 if (child == d->tg) {
10148 period = d->rt_period;
10149 runtime = d->rt_runtime;
10152 sum += to_ratio(period, runtime);
10161 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10163 struct rt_schedulable_data data = {
10165 .rt_period = period,
10166 .rt_runtime = runtime,
10169 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10172 static int tg_set_bandwidth(struct task_group *tg,
10173 u64 rt_period, u64 rt_runtime)
10177 mutex_lock(&rt_constraints_mutex);
10178 read_lock(&tasklist_lock);
10179 err = __rt_schedulable(tg, rt_period, rt_runtime);
10183 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10184 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10185 tg->rt_bandwidth.rt_runtime = rt_runtime;
10187 for_each_possible_cpu(i) {
10188 struct rt_rq *rt_rq = tg->rt_rq[i];
10190 spin_lock(&rt_rq->rt_runtime_lock);
10191 rt_rq->rt_runtime = rt_runtime;
10192 spin_unlock(&rt_rq->rt_runtime_lock);
10194 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10196 read_unlock(&tasklist_lock);
10197 mutex_unlock(&rt_constraints_mutex);
10202 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10204 u64 rt_runtime, rt_period;
10206 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10207 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10208 if (rt_runtime_us < 0)
10209 rt_runtime = RUNTIME_INF;
10211 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10214 long sched_group_rt_runtime(struct task_group *tg)
10218 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10221 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10222 do_div(rt_runtime_us, NSEC_PER_USEC);
10223 return rt_runtime_us;
10226 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10228 u64 rt_runtime, rt_period;
10230 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10231 rt_runtime = tg->rt_bandwidth.rt_runtime;
10233 if (rt_period == 0)
10236 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10239 long sched_group_rt_period(struct task_group *tg)
10243 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10244 do_div(rt_period_us, NSEC_PER_USEC);
10245 return rt_period_us;
10248 static int sched_rt_global_constraints(void)
10250 u64 runtime, period;
10253 if (sysctl_sched_rt_period <= 0)
10256 runtime = global_rt_runtime();
10257 period = global_rt_period();
10260 * Sanity check on the sysctl variables.
10262 if (runtime > period && runtime != RUNTIME_INF)
10265 mutex_lock(&rt_constraints_mutex);
10266 read_lock(&tasklist_lock);
10267 ret = __rt_schedulable(NULL, 0, 0);
10268 read_unlock(&tasklist_lock);
10269 mutex_unlock(&rt_constraints_mutex);
10274 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10276 /* Don't accept realtime tasks when there is no way for them to run */
10277 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10283 #else /* !CONFIG_RT_GROUP_SCHED */
10284 static int sched_rt_global_constraints(void)
10286 unsigned long flags;
10289 if (sysctl_sched_rt_period <= 0)
10293 * There's always some RT tasks in the root group
10294 * -- migration, kstopmachine etc..
10296 if (sysctl_sched_rt_runtime == 0)
10299 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10300 for_each_possible_cpu(i) {
10301 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10303 spin_lock(&rt_rq->rt_runtime_lock);
10304 rt_rq->rt_runtime = global_rt_runtime();
10305 spin_unlock(&rt_rq->rt_runtime_lock);
10307 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10311 #endif /* CONFIG_RT_GROUP_SCHED */
10313 int sched_rt_handler(struct ctl_table *table, int write,
10314 struct file *filp, void __user *buffer, size_t *lenp,
10318 int old_period, old_runtime;
10319 static DEFINE_MUTEX(mutex);
10321 mutex_lock(&mutex);
10322 old_period = sysctl_sched_rt_period;
10323 old_runtime = sysctl_sched_rt_runtime;
10325 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
10327 if (!ret && write) {
10328 ret = sched_rt_global_constraints();
10330 sysctl_sched_rt_period = old_period;
10331 sysctl_sched_rt_runtime = old_runtime;
10333 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10334 def_rt_bandwidth.rt_period =
10335 ns_to_ktime(global_rt_period());
10338 mutex_unlock(&mutex);
10343 #ifdef CONFIG_CGROUP_SCHED
10345 /* return corresponding task_group object of a cgroup */
10346 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10348 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10349 struct task_group, css);
10352 static struct cgroup_subsys_state *
10353 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10355 struct task_group *tg, *parent;
10357 if (!cgrp->parent) {
10358 /* This is early initialization for the top cgroup */
10359 return &init_task_group.css;
10362 parent = cgroup_tg(cgrp->parent);
10363 tg = sched_create_group(parent);
10365 return ERR_PTR(-ENOMEM);
10371 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10373 struct task_group *tg = cgroup_tg(cgrp);
10375 sched_destroy_group(tg);
10379 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10380 struct task_struct *tsk)
10382 #ifdef CONFIG_RT_GROUP_SCHED
10383 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10386 /* We don't support RT-tasks being in separate groups */
10387 if (tsk->sched_class != &fair_sched_class)
10395 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10396 struct cgroup *old_cont, struct task_struct *tsk)
10398 sched_move_task(tsk);
10401 #ifdef CONFIG_FAIR_GROUP_SCHED
10402 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10405 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10408 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10410 struct task_group *tg = cgroup_tg(cgrp);
10412 return (u64) tg->shares;
10414 #endif /* CONFIG_FAIR_GROUP_SCHED */
10416 #ifdef CONFIG_RT_GROUP_SCHED
10417 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10420 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10423 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10425 return sched_group_rt_runtime(cgroup_tg(cgrp));
10428 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10431 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10434 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10436 return sched_group_rt_period(cgroup_tg(cgrp));
10438 #endif /* CONFIG_RT_GROUP_SCHED */
10440 static struct cftype cpu_files[] = {
10441 #ifdef CONFIG_FAIR_GROUP_SCHED
10444 .read_u64 = cpu_shares_read_u64,
10445 .write_u64 = cpu_shares_write_u64,
10448 #ifdef CONFIG_RT_GROUP_SCHED
10450 .name = "rt_runtime_us",
10451 .read_s64 = cpu_rt_runtime_read,
10452 .write_s64 = cpu_rt_runtime_write,
10455 .name = "rt_period_us",
10456 .read_u64 = cpu_rt_period_read_uint,
10457 .write_u64 = cpu_rt_period_write_uint,
10462 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10464 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10467 struct cgroup_subsys cpu_cgroup_subsys = {
10469 .create = cpu_cgroup_create,
10470 .destroy = cpu_cgroup_destroy,
10471 .can_attach = cpu_cgroup_can_attach,
10472 .attach = cpu_cgroup_attach,
10473 .populate = cpu_cgroup_populate,
10474 .subsys_id = cpu_cgroup_subsys_id,
10478 #endif /* CONFIG_CGROUP_SCHED */
10480 #ifdef CONFIG_CGROUP_CPUACCT
10483 * CPU accounting code for task groups.
10485 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10486 * (balbir@in.ibm.com).
10489 /* track cpu usage of a group of tasks and its child groups */
10491 struct cgroup_subsys_state css;
10492 /* cpuusage holds pointer to a u64-type object on every cpu */
10494 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10495 struct cpuacct *parent;
10498 struct cgroup_subsys cpuacct_subsys;
10500 /* return cpu accounting group corresponding to this container */
10501 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10503 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10504 struct cpuacct, css);
10507 /* return cpu accounting group to which this task belongs */
10508 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10510 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10511 struct cpuacct, css);
10514 /* create a new cpu accounting group */
10515 static struct cgroup_subsys_state *cpuacct_create(
10516 struct cgroup_subsys *ss, struct cgroup *cgrp)
10518 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10524 ca->cpuusage = alloc_percpu(u64);
10528 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10529 if (percpu_counter_init(&ca->cpustat[i], 0))
10530 goto out_free_counters;
10533 ca->parent = cgroup_ca(cgrp->parent);
10539 percpu_counter_destroy(&ca->cpustat[i]);
10540 free_percpu(ca->cpuusage);
10544 return ERR_PTR(-ENOMEM);
10547 /* destroy an existing cpu accounting group */
10549 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10551 struct cpuacct *ca = cgroup_ca(cgrp);
10554 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10555 percpu_counter_destroy(&ca->cpustat[i]);
10556 free_percpu(ca->cpuusage);
10560 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10562 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10565 #ifndef CONFIG_64BIT
10567 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10569 spin_lock_irq(&cpu_rq(cpu)->lock);
10571 spin_unlock_irq(&cpu_rq(cpu)->lock);
10579 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10581 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10583 #ifndef CONFIG_64BIT
10585 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10587 spin_lock_irq(&cpu_rq(cpu)->lock);
10589 spin_unlock_irq(&cpu_rq(cpu)->lock);
10595 /* return total cpu usage (in nanoseconds) of a group */
10596 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10598 struct cpuacct *ca = cgroup_ca(cgrp);
10599 u64 totalcpuusage = 0;
10602 for_each_present_cpu(i)
10603 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10605 return totalcpuusage;
10608 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10611 struct cpuacct *ca = cgroup_ca(cgrp);
10620 for_each_present_cpu(i)
10621 cpuacct_cpuusage_write(ca, i, 0);
10627 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10628 struct seq_file *m)
10630 struct cpuacct *ca = cgroup_ca(cgroup);
10634 for_each_present_cpu(i) {
10635 percpu = cpuacct_cpuusage_read(ca, i);
10636 seq_printf(m, "%llu ", (unsigned long long) percpu);
10638 seq_printf(m, "\n");
10642 static const char *cpuacct_stat_desc[] = {
10643 [CPUACCT_STAT_USER] = "user",
10644 [CPUACCT_STAT_SYSTEM] = "system",
10647 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10648 struct cgroup_map_cb *cb)
10650 struct cpuacct *ca = cgroup_ca(cgrp);
10653 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10654 s64 val = percpu_counter_read(&ca->cpustat[i]);
10655 val = cputime64_to_clock_t(val);
10656 cb->fill(cb, cpuacct_stat_desc[i], val);
10661 static struct cftype files[] = {
10664 .read_u64 = cpuusage_read,
10665 .write_u64 = cpuusage_write,
10668 .name = "usage_percpu",
10669 .read_seq_string = cpuacct_percpu_seq_read,
10673 .read_map = cpuacct_stats_show,
10677 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10679 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10683 * charge this task's execution time to its accounting group.
10685 * called with rq->lock held.
10687 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10689 struct cpuacct *ca;
10692 if (unlikely(!cpuacct_subsys.active))
10695 cpu = task_cpu(tsk);
10701 for (; ca; ca = ca->parent) {
10702 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10703 *cpuusage += cputime;
10710 * Charge the system/user time to the task's accounting group.
10712 static void cpuacct_update_stats(struct task_struct *tsk,
10713 enum cpuacct_stat_index idx, cputime_t val)
10715 struct cpuacct *ca;
10717 if (unlikely(!cpuacct_subsys.active))
10724 percpu_counter_add(&ca->cpustat[idx], val);
10730 struct cgroup_subsys cpuacct_subsys = {
10732 .create = cpuacct_create,
10733 .destroy = cpuacct_destroy,
10734 .populate = cpuacct_populate,
10735 .subsys_id = cpuacct_subsys_id,
10737 #endif /* CONFIG_CGROUP_CPUACCT */
10741 int rcu_expedited_torture_stats(char *page)
10745 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10747 void synchronize_sched_expedited(void)
10750 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10752 #else /* #ifndef CONFIG_SMP */
10754 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10755 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10757 #define RCU_EXPEDITED_STATE_POST -2
10758 #define RCU_EXPEDITED_STATE_IDLE -1
10760 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10762 int rcu_expedited_torture_stats(char *page)
10767 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10768 for_each_online_cpu(cpu) {
10769 cnt += sprintf(&page[cnt], " %d:%d",
10770 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10772 cnt += sprintf(&page[cnt], "\n");
10775 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10777 static long synchronize_sched_expedited_count;
10780 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10781 * approach to force grace period to end quickly. This consumes
10782 * significant time on all CPUs, and is thus not recommended for
10783 * any sort of common-case code.
10785 * Note that it is illegal to call this function while holding any
10786 * lock that is acquired by a CPU-hotplug notifier. Failing to
10787 * observe this restriction will result in deadlock.
10789 void synchronize_sched_expedited(void)
10792 unsigned long flags;
10793 bool need_full_sync = 0;
10795 struct migration_req *req;
10799 smp_mb(); /* ensure prior mod happens before capturing snap. */
10800 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10802 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10804 if (trycount++ < 10)
10805 udelay(trycount * num_online_cpus());
10807 synchronize_sched();
10810 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10811 smp_mb(); /* ensure test happens before caller kfree */
10816 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10817 for_each_online_cpu(cpu) {
10819 req = &per_cpu(rcu_migration_req, cpu);
10820 init_completion(&req->done);
10822 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10823 spin_lock_irqsave(&rq->lock, flags);
10824 list_add(&req->list, &rq->migration_queue);
10825 spin_unlock_irqrestore(&rq->lock, flags);
10826 wake_up_process(rq->migration_thread);
10828 for_each_online_cpu(cpu) {
10829 rcu_expedited_state = cpu;
10830 req = &per_cpu(rcu_migration_req, cpu);
10832 wait_for_completion(&req->done);
10833 spin_lock_irqsave(&rq->lock, flags);
10834 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10835 need_full_sync = 1;
10836 req->dest_cpu = RCU_MIGRATION_IDLE;
10837 spin_unlock_irqrestore(&rq->lock, flags);
10839 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10840 mutex_unlock(&rcu_sched_expedited_mutex);
10842 if (need_full_sync)
10843 synchronize_sched();
10845 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10847 #endif /* #else #ifndef CONFIG_SMP */