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/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/bootmem.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
79 * Convert user-nice values [ -20 ... 0 ... 19 ]
80 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
83 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
84 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
85 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
88 * 'User priority' is the nice value converted to something we
89 * can work with better when scaling various scheduler parameters,
90 * it's a [ 0 ... 39 ] range.
92 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
93 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
94 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
97 * Helpers for converting nanosecond timing to jiffy resolution
99 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
101 #define NICE_0_LOAD SCHED_LOAD_SCALE
102 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
105 * These are the 'tuning knobs' of the scheduler:
107 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
108 * Timeslices get refilled after they expire.
110 #define DEF_TIMESLICE (100 * HZ / 1000)
113 * single value that denotes runtime == period, ie unlimited time.
115 #define RUNTIME_INF ((u64)~0ULL)
119 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
120 * Since cpu_power is a 'constant', we can use a reciprocal divide.
122 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
124 return reciprocal_divide(load, sg->reciprocal_cpu_power);
128 * Each time a sched group cpu_power is changed,
129 * we must compute its reciprocal value
131 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
133 sg->__cpu_power += val;
134 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
138 static inline int rt_policy(int policy)
140 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
145 static inline int task_has_rt_policy(struct task_struct *p)
147 return rt_policy(p->policy);
151 * This is the priority-queue data structure of the RT scheduling class:
153 struct rt_prio_array {
154 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
155 struct list_head queue[MAX_RT_PRIO];
158 struct rt_bandwidth {
159 /* nests inside the rq lock: */
160 spinlock_t rt_runtime_lock;
163 struct hrtimer rt_period_timer;
166 static struct rt_bandwidth def_rt_bandwidth;
168 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
170 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
172 struct rt_bandwidth *rt_b =
173 container_of(timer, struct rt_bandwidth, rt_period_timer);
179 now = hrtimer_cb_get_time(timer);
180 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
185 idle = do_sched_rt_period_timer(rt_b, overrun);
188 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
192 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
194 rt_b->rt_period = ns_to_ktime(period);
195 rt_b->rt_runtime = runtime;
197 spin_lock_init(&rt_b->rt_runtime_lock);
199 hrtimer_init(&rt_b->rt_period_timer,
200 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
201 rt_b->rt_period_timer.function = sched_rt_period_timer;
202 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
205 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
209 if (rt_b->rt_runtime == RUNTIME_INF)
212 if (hrtimer_active(&rt_b->rt_period_timer))
215 spin_lock(&rt_b->rt_runtime_lock);
217 if (hrtimer_active(&rt_b->rt_period_timer))
220 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
221 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
222 hrtimer_start(&rt_b->rt_period_timer,
223 rt_b->rt_period_timer.expires,
226 spin_unlock(&rt_b->rt_runtime_lock);
229 #ifdef CONFIG_RT_GROUP_SCHED
230 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
232 hrtimer_cancel(&rt_b->rt_period_timer);
237 * sched_domains_mutex serializes calls to arch_init_sched_domains,
238 * detach_destroy_domains and partition_sched_domains.
240 static DEFINE_MUTEX(sched_domains_mutex);
242 #ifdef CONFIG_GROUP_SCHED
244 #include <linux/cgroup.h>
248 static LIST_HEAD(task_groups);
250 /* task group related information */
252 #ifdef CONFIG_CGROUP_SCHED
253 struct cgroup_subsys_state css;
256 #ifdef CONFIG_FAIR_GROUP_SCHED
257 /* schedulable entities of this group on each cpu */
258 struct sched_entity **se;
259 /* runqueue "owned" by this group on each cpu */
260 struct cfs_rq **cfs_rq;
261 unsigned long shares;
264 #ifdef CONFIG_RT_GROUP_SCHED
265 struct sched_rt_entity **rt_se;
266 struct rt_rq **rt_rq;
268 struct rt_bandwidth rt_bandwidth;
272 struct list_head list;
274 struct task_group *parent;
275 struct list_head siblings;
276 struct list_head children;
279 #ifdef CONFIG_USER_SCHED
283 * Every UID task group (including init_task_group aka UID-0) will
284 * be a child to this group.
286 struct task_group root_task_group;
288 #ifdef CONFIG_FAIR_GROUP_SCHED
289 /* Default task group's sched entity on each cpu */
290 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
291 /* Default task group's cfs_rq on each cpu */
292 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
295 #ifdef CONFIG_RT_GROUP_SCHED
296 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
297 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
300 #define root_task_group init_task_group
303 /* task_group_lock serializes add/remove of task groups and also changes to
304 * a task group's cpu shares.
306 static DEFINE_SPINLOCK(task_group_lock);
308 #ifdef CONFIG_FAIR_GROUP_SCHED
309 #ifdef CONFIG_USER_SCHED
310 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
312 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
316 * A weight of 0 or 1 can cause arithmetics problems.
317 * A weight of a cfs_rq is the sum of weights of which entities
318 * are queued on this cfs_rq, so a weight of a entity should not be
319 * too large, so as the shares value of a task group.
320 * (The default weight is 1024 - so there's no practical
321 * limitation from this.)
324 #define MAX_SHARES (1UL << 18)
326 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
329 /* Default task group.
330 * Every task in system belong to this group at bootup.
332 struct task_group init_task_group;
334 /* return group to which a task belongs */
335 static inline struct task_group *task_group(struct task_struct *p)
337 struct task_group *tg;
339 #ifdef CONFIG_USER_SCHED
341 #elif defined(CONFIG_CGROUP_SCHED)
342 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
343 struct task_group, css);
345 tg = &init_task_group;
350 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
351 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
353 #ifdef CONFIG_FAIR_GROUP_SCHED
354 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
355 p->se.parent = task_group(p)->se[cpu];
358 #ifdef CONFIG_RT_GROUP_SCHED
359 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
360 p->rt.parent = task_group(p)->rt_se[cpu];
366 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
368 #endif /* CONFIG_GROUP_SCHED */
370 /* CFS-related fields in a runqueue */
372 struct load_weight load;
373 unsigned long nr_running;
378 struct rb_root tasks_timeline;
379 struct rb_node *rb_leftmost;
381 struct list_head tasks;
382 struct list_head *balance_iterator;
385 * 'curr' points to currently running entity on this cfs_rq.
386 * It is set to NULL otherwise (i.e when none are currently running).
388 struct sched_entity *curr, *next;
390 unsigned long nr_spread_over;
392 #ifdef CONFIG_FAIR_GROUP_SCHED
393 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
396 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
397 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
398 * (like users, containers etc.)
400 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
401 * list is used during load balance.
403 struct list_head leaf_cfs_rq_list;
404 struct task_group *tg; /* group that "owns" this runqueue */
408 /* Real-Time classes' related field in a runqueue: */
410 struct rt_prio_array active;
411 unsigned long rt_nr_running;
412 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
413 int highest_prio; /* highest queued rt task prio */
416 unsigned long rt_nr_migratory;
422 /* Nests inside the rq lock: */
423 spinlock_t rt_runtime_lock;
425 #ifdef CONFIG_RT_GROUP_SCHED
426 unsigned long rt_nr_boosted;
429 struct list_head leaf_rt_rq_list;
430 struct task_group *tg;
431 struct sched_rt_entity *rt_se;
438 * We add the notion of a root-domain which will be used to define per-domain
439 * variables. Each exclusive cpuset essentially defines an island domain by
440 * fully partitioning the member cpus from any other cpuset. Whenever a new
441 * exclusive cpuset is created, we also create and attach a new root-domain
451 * The "RT overload" flag: it gets set if a CPU has more than
452 * one runnable RT task.
459 * By default the system creates a single root-domain with all cpus as
460 * members (mimicking the global state we have today).
462 static struct root_domain def_root_domain;
467 * This is the main, per-CPU runqueue data structure.
469 * Locking rule: those places that want to lock multiple runqueues
470 * (such as the load balancing or the thread migration code), lock
471 * acquire operations must be ordered by ascending &runqueue.
478 * nr_running and cpu_load should be in the same cacheline because
479 * remote CPUs use both these fields when doing load calculation.
481 unsigned long nr_running;
482 #define CPU_LOAD_IDX_MAX 5
483 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
484 unsigned char idle_at_tick;
486 unsigned long last_tick_seen;
487 unsigned char in_nohz_recently;
489 /* capture load from *all* tasks on this cpu: */
490 struct load_weight load;
491 unsigned long nr_load_updates;
497 #ifdef CONFIG_FAIR_GROUP_SCHED
498 /* list of leaf cfs_rq on this cpu: */
499 struct list_head leaf_cfs_rq_list;
501 #ifdef CONFIG_RT_GROUP_SCHED
502 struct list_head leaf_rt_rq_list;
506 * This is part of a global counter where only the total sum
507 * over all CPUs matters. A task can increase this counter on
508 * one CPU and if it got migrated afterwards it may decrease
509 * it on another CPU. Always updated under the runqueue lock:
511 unsigned long nr_uninterruptible;
513 struct task_struct *curr, *idle;
514 unsigned long next_balance;
515 struct mm_struct *prev_mm;
522 struct root_domain *rd;
523 struct sched_domain *sd;
525 /* For active balancing */
528 /* cpu of this runqueue: */
531 struct task_struct *migration_thread;
532 struct list_head migration_queue;
535 #ifdef CONFIG_SCHED_HRTICK
536 unsigned long hrtick_flags;
537 ktime_t hrtick_expire;
538 struct hrtimer hrtick_timer;
541 #ifdef CONFIG_SCHEDSTATS
543 struct sched_info rq_sched_info;
545 /* sys_sched_yield() stats */
546 unsigned int yld_exp_empty;
547 unsigned int yld_act_empty;
548 unsigned int yld_both_empty;
549 unsigned int yld_count;
551 /* schedule() stats */
552 unsigned int sched_switch;
553 unsigned int sched_count;
554 unsigned int sched_goidle;
556 /* try_to_wake_up() stats */
557 unsigned int ttwu_count;
558 unsigned int ttwu_local;
561 unsigned int bkl_count;
563 struct lock_class_key rq_lock_key;
566 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
568 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
570 rq->curr->sched_class->check_preempt_curr(rq, p);
573 static inline int cpu_of(struct rq *rq)
583 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
584 * See detach_destroy_domains: synchronize_sched for details.
586 * The domain tree of any CPU may only be accessed from within
587 * preempt-disabled sections.
589 #define for_each_domain(cpu, __sd) \
590 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
592 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
593 #define this_rq() (&__get_cpu_var(runqueues))
594 #define task_rq(p) cpu_rq(task_cpu(p))
595 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
597 static inline void update_rq_clock(struct rq *rq)
599 rq->clock = sched_clock_cpu(cpu_of(rq));
603 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
605 #ifdef CONFIG_SCHED_DEBUG
606 # define const_debug __read_mostly
608 # define const_debug static const
614 * Returns true if the current cpu runqueue is locked.
615 * This interface allows printk to be called with the runqueue lock
616 * held and know whether or not it is OK to wake up the klogd.
618 int runqueue_is_locked(void)
621 struct rq *rq = cpu_rq(cpu);
624 ret = spin_is_locked(&rq->lock);
630 * Debugging: various feature bits
633 #define SCHED_FEAT(name, enabled) \
634 __SCHED_FEAT_##name ,
637 #include "sched_features.h"
642 #define SCHED_FEAT(name, enabled) \
643 (1UL << __SCHED_FEAT_##name) * enabled |
645 const_debug unsigned int sysctl_sched_features =
646 #include "sched_features.h"
651 #ifdef CONFIG_SCHED_DEBUG
652 #define SCHED_FEAT(name, enabled) \
655 static __read_mostly char *sched_feat_names[] = {
656 #include "sched_features.h"
662 static int sched_feat_open(struct inode *inode, struct file *filp)
664 filp->private_data = inode->i_private;
669 sched_feat_read(struct file *filp, char __user *ubuf,
670 size_t cnt, loff_t *ppos)
677 for (i = 0; sched_feat_names[i]; i++) {
678 len += strlen(sched_feat_names[i]);
682 buf = kmalloc(len + 2, GFP_KERNEL);
686 for (i = 0; sched_feat_names[i]; i++) {
687 if (sysctl_sched_features & (1UL << i))
688 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
690 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
693 r += sprintf(buf + r, "\n");
694 WARN_ON(r >= len + 2);
696 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
704 sched_feat_write(struct file *filp, const char __user *ubuf,
705 size_t cnt, loff_t *ppos)
715 if (copy_from_user(&buf, ubuf, cnt))
720 if (strncmp(buf, "NO_", 3) == 0) {
725 for (i = 0; sched_feat_names[i]; i++) {
726 int len = strlen(sched_feat_names[i]);
728 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
730 sysctl_sched_features &= ~(1UL << i);
732 sysctl_sched_features |= (1UL << i);
737 if (!sched_feat_names[i])
745 static struct file_operations sched_feat_fops = {
746 .open = sched_feat_open,
747 .read = sched_feat_read,
748 .write = sched_feat_write,
751 static __init int sched_init_debug(void)
753 debugfs_create_file("sched_features", 0644, NULL, NULL,
758 late_initcall(sched_init_debug);
762 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
765 * Number of tasks to iterate in a single balance run.
766 * Limited because this is done with IRQs disabled.
768 const_debug unsigned int sysctl_sched_nr_migrate = 32;
771 * period over which we measure -rt task cpu usage in us.
774 unsigned int sysctl_sched_rt_period = 1000000;
776 static __read_mostly int scheduler_running;
779 * part of the period that we allow rt tasks to run in us.
782 int sysctl_sched_rt_runtime = 950000;
784 static inline u64 global_rt_period(void)
786 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
789 static inline u64 global_rt_runtime(void)
791 if (sysctl_sched_rt_period < 0)
794 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
797 unsigned long long time_sync_thresh = 100000;
799 static DEFINE_PER_CPU(unsigned long long, time_offset);
800 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
803 * Global lock which we take every now and then to synchronize
804 * the CPUs time. This method is not warp-safe, but it's good
805 * enough to synchronize slowly diverging time sources and thus
806 * it's good enough for tracing:
808 static DEFINE_SPINLOCK(time_sync_lock);
809 static unsigned long long prev_global_time;
811 static unsigned long long __sync_cpu_clock(unsigned long long time, int cpu)
814 * We want this inlined, to not get tracer function calls
815 * in this critical section:
817 spin_acquire(&time_sync_lock.dep_map, 0, 0, _THIS_IP_);
818 __raw_spin_lock(&time_sync_lock.raw_lock);
820 if (time < prev_global_time) {
821 per_cpu(time_offset, cpu) += prev_global_time - time;
822 time = prev_global_time;
824 prev_global_time = time;
827 __raw_spin_unlock(&time_sync_lock.raw_lock);
828 spin_release(&time_sync_lock.dep_map, 1, _THIS_IP_);
833 static unsigned long long __cpu_clock(int cpu)
835 unsigned long long now;
838 * Only call sched_clock() if the scheduler has already been
839 * initialized (some code might call cpu_clock() very early):
841 if (unlikely(!scheduler_running))
844 now = sched_clock_cpu(cpu);
850 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
851 * clock constructed from sched_clock():
853 unsigned long long notrace cpu_clock(int cpu)
855 unsigned long long prev_cpu_time, time, delta_time;
858 local_irq_save(flags);
859 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
860 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
861 delta_time = time-prev_cpu_time;
863 if (unlikely(delta_time > time_sync_thresh)) {
864 time = __sync_cpu_clock(time, cpu);
865 per_cpu(prev_cpu_time, cpu) = time;
867 local_irq_restore(flags);
871 EXPORT_SYMBOL_GPL(cpu_clock);
873 #ifndef prepare_arch_switch
874 # define prepare_arch_switch(next) do { } while (0)
876 #ifndef finish_arch_switch
877 # define finish_arch_switch(prev) do { } while (0)
880 static inline int task_current(struct rq *rq, struct task_struct *p)
882 return rq->curr == p;
885 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
886 static inline int task_running(struct rq *rq, struct task_struct *p)
888 return task_current(rq, p);
891 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
895 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
897 #ifdef CONFIG_DEBUG_SPINLOCK
898 /* this is a valid case when another task releases the spinlock */
899 rq->lock.owner = current;
902 * If we are tracking spinlock dependencies then we have to
903 * fix up the runqueue lock - which gets 'carried over' from
906 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
908 spin_unlock_irq(&rq->lock);
911 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
912 static inline int task_running(struct rq *rq, struct task_struct *p)
917 return task_current(rq, p);
921 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
925 * We can optimise this out completely for !SMP, because the
926 * SMP rebalancing from interrupt is the only thing that cares
931 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
932 spin_unlock_irq(&rq->lock);
934 spin_unlock(&rq->lock);
938 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
942 * After ->oncpu is cleared, the task can be moved to a different CPU.
943 * We must ensure this doesn't happen until the switch is completely
949 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
953 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
956 * __task_rq_lock - lock the runqueue a given task resides on.
957 * Must be called interrupts disabled.
959 static inline struct rq *__task_rq_lock(struct task_struct *p)
963 struct rq *rq = task_rq(p);
964 spin_lock(&rq->lock);
965 if (likely(rq == task_rq(p)))
967 spin_unlock(&rq->lock);
972 * task_rq_lock - lock the runqueue a given task resides on and disable
973 * interrupts. Note the ordering: we can safely lookup the task_rq without
974 * explicitly disabling preemption.
976 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
982 local_irq_save(*flags);
984 spin_lock(&rq->lock);
985 if (likely(rq == task_rq(p)))
987 spin_unlock_irqrestore(&rq->lock, *flags);
991 static void __task_rq_unlock(struct rq *rq)
994 spin_unlock(&rq->lock);
997 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1000 spin_unlock_irqrestore(&rq->lock, *flags);
1004 * this_rq_lock - lock this runqueue and disable interrupts.
1006 static struct rq *this_rq_lock(void)
1007 __acquires(rq->lock)
1011 local_irq_disable();
1013 spin_lock(&rq->lock);
1018 static void __resched_task(struct task_struct *p, int tif_bit);
1020 static inline void resched_task(struct task_struct *p)
1022 __resched_task(p, TIF_NEED_RESCHED);
1025 #ifdef CONFIG_SCHED_HRTICK
1027 * Use HR-timers to deliver accurate preemption points.
1029 * Its all a bit involved since we cannot program an hrt while holding the
1030 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1033 * When we get rescheduled we reprogram the hrtick_timer outside of the
1036 static inline void resched_hrt(struct task_struct *p)
1038 __resched_task(p, TIF_HRTICK_RESCHED);
1041 static inline void resched_rq(struct rq *rq)
1043 unsigned long flags;
1045 spin_lock_irqsave(&rq->lock, flags);
1046 resched_task(rq->curr);
1047 spin_unlock_irqrestore(&rq->lock, flags);
1051 HRTICK_SET, /* re-programm hrtick_timer */
1052 HRTICK_RESET, /* not a new slice */
1053 HRTICK_BLOCK, /* stop hrtick operations */
1058 * - enabled by features
1059 * - hrtimer is actually high res
1061 static inline int hrtick_enabled(struct rq *rq)
1063 if (!sched_feat(HRTICK))
1065 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
1067 return hrtimer_is_hres_active(&rq->hrtick_timer);
1071 * Called to set the hrtick timer state.
1073 * called with rq->lock held and irqs disabled
1075 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1077 assert_spin_locked(&rq->lock);
1080 * preempt at: now + delay
1083 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1085 * indicate we need to program the timer
1087 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1089 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1092 * New slices are called from the schedule path and don't need a
1093 * forced reschedule.
1096 resched_hrt(rq->curr);
1099 static void hrtick_clear(struct rq *rq)
1101 if (hrtimer_active(&rq->hrtick_timer))
1102 hrtimer_cancel(&rq->hrtick_timer);
1106 * Update the timer from the possible pending state.
1108 static void hrtick_set(struct rq *rq)
1112 unsigned long flags;
1114 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1116 spin_lock_irqsave(&rq->lock, flags);
1117 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1118 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1119 time = rq->hrtick_expire;
1120 clear_thread_flag(TIF_HRTICK_RESCHED);
1121 spin_unlock_irqrestore(&rq->lock, flags);
1124 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1125 if (reset && !hrtimer_active(&rq->hrtick_timer))
1132 * High-resolution timer tick.
1133 * Runs from hardirq context with interrupts disabled.
1135 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1137 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1139 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1141 spin_lock(&rq->lock);
1142 update_rq_clock(rq);
1143 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1144 spin_unlock(&rq->lock);
1146 return HRTIMER_NORESTART;
1149 static void hotplug_hrtick_disable(int cpu)
1151 struct rq *rq = cpu_rq(cpu);
1152 unsigned long flags;
1154 spin_lock_irqsave(&rq->lock, flags);
1155 rq->hrtick_flags = 0;
1156 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1157 spin_unlock_irqrestore(&rq->lock, flags);
1162 static void hotplug_hrtick_enable(int cpu)
1164 struct rq *rq = cpu_rq(cpu);
1165 unsigned long flags;
1167 spin_lock_irqsave(&rq->lock, flags);
1168 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1169 spin_unlock_irqrestore(&rq->lock, flags);
1173 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1175 int cpu = (int)(long)hcpu;
1178 case CPU_UP_CANCELED:
1179 case CPU_UP_CANCELED_FROZEN:
1180 case CPU_DOWN_PREPARE:
1181 case CPU_DOWN_PREPARE_FROZEN:
1183 case CPU_DEAD_FROZEN:
1184 hotplug_hrtick_disable(cpu);
1187 case CPU_UP_PREPARE:
1188 case CPU_UP_PREPARE_FROZEN:
1189 case CPU_DOWN_FAILED:
1190 case CPU_DOWN_FAILED_FROZEN:
1192 case CPU_ONLINE_FROZEN:
1193 hotplug_hrtick_enable(cpu);
1200 static void init_hrtick(void)
1202 hotcpu_notifier(hotplug_hrtick, 0);
1205 static void init_rq_hrtick(struct rq *rq)
1207 rq->hrtick_flags = 0;
1208 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1209 rq->hrtick_timer.function = hrtick;
1210 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1213 void hrtick_resched(void)
1216 unsigned long flags;
1218 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1221 local_irq_save(flags);
1222 rq = cpu_rq(smp_processor_id());
1224 local_irq_restore(flags);
1227 static inline void hrtick_clear(struct rq *rq)
1231 static inline void hrtick_set(struct rq *rq)
1235 static inline void init_rq_hrtick(struct rq *rq)
1239 void hrtick_resched(void)
1243 static inline void init_hrtick(void)
1249 * resched_task - mark a task 'to be rescheduled now'.
1251 * On UP this means the setting of the need_resched flag, on SMP it
1252 * might also involve a cross-CPU call to trigger the scheduler on
1257 #ifndef tsk_is_polling
1258 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1261 static void __resched_task(struct task_struct *p, int tif_bit)
1265 assert_spin_locked(&task_rq(p)->lock);
1267 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1270 set_tsk_thread_flag(p, tif_bit);
1273 if (cpu == smp_processor_id())
1276 /* NEED_RESCHED must be visible before we test polling */
1278 if (!tsk_is_polling(p))
1279 smp_send_reschedule(cpu);
1282 static void resched_cpu(int cpu)
1284 struct rq *rq = cpu_rq(cpu);
1285 unsigned long flags;
1287 if (!spin_trylock_irqsave(&rq->lock, flags))
1289 resched_task(cpu_curr(cpu));
1290 spin_unlock_irqrestore(&rq->lock, flags);
1295 * When add_timer_on() enqueues a timer into the timer wheel of an
1296 * idle CPU then this timer might expire before the next timer event
1297 * which is scheduled to wake up that CPU. In case of a completely
1298 * idle system the next event might even be infinite time into the
1299 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1300 * leaves the inner idle loop so the newly added timer is taken into
1301 * account when the CPU goes back to idle and evaluates the timer
1302 * wheel for the next timer event.
1304 void wake_up_idle_cpu(int cpu)
1306 struct rq *rq = cpu_rq(cpu);
1308 if (cpu == smp_processor_id())
1312 * This is safe, as this function is called with the timer
1313 * wheel base lock of (cpu) held. When the CPU is on the way
1314 * to idle and has not yet set rq->curr to idle then it will
1315 * be serialized on the timer wheel base lock and take the new
1316 * timer into account automatically.
1318 if (rq->curr != rq->idle)
1322 * We can set TIF_RESCHED on the idle task of the other CPU
1323 * lockless. The worst case is that the other CPU runs the
1324 * idle task through an additional NOOP schedule()
1326 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1328 /* NEED_RESCHED must be visible before we test polling */
1330 if (!tsk_is_polling(rq->idle))
1331 smp_send_reschedule(cpu);
1336 static void __resched_task(struct task_struct *p, int tif_bit)
1338 assert_spin_locked(&task_rq(p)->lock);
1339 set_tsk_thread_flag(p, tif_bit);
1343 #if BITS_PER_LONG == 32
1344 # define WMULT_CONST (~0UL)
1346 # define WMULT_CONST (1UL << 32)
1349 #define WMULT_SHIFT 32
1352 * Shift right and round:
1354 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1356 static unsigned long
1357 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1358 struct load_weight *lw)
1362 if (!lw->inv_weight) {
1363 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1366 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1370 tmp = (u64)delta_exec * weight;
1372 * Check whether we'd overflow the 64-bit multiplication:
1374 if (unlikely(tmp > WMULT_CONST))
1375 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1378 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1380 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1383 static inline unsigned long
1384 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1386 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1389 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1395 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1402 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1403 * of tasks with abnormal "nice" values across CPUs the contribution that
1404 * each task makes to its run queue's load is weighted according to its
1405 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1406 * scaled version of the new time slice allocation that they receive on time
1410 #define WEIGHT_IDLEPRIO 2
1411 #define WMULT_IDLEPRIO (1 << 31)
1414 * Nice levels are multiplicative, with a gentle 10% change for every
1415 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1416 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1417 * that remained on nice 0.
1419 * The "10% effect" is relative and cumulative: from _any_ nice level,
1420 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1421 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1422 * If a task goes up by ~10% and another task goes down by ~10% then
1423 * the relative distance between them is ~25%.)
1425 static const int prio_to_weight[40] = {
1426 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1427 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1428 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1429 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1430 /* 0 */ 1024, 820, 655, 526, 423,
1431 /* 5 */ 335, 272, 215, 172, 137,
1432 /* 10 */ 110, 87, 70, 56, 45,
1433 /* 15 */ 36, 29, 23, 18, 15,
1437 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1439 * In cases where the weight does not change often, we can use the
1440 * precalculated inverse to speed up arithmetics by turning divisions
1441 * into multiplications:
1443 static const u32 prio_to_wmult[40] = {
1444 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1445 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1446 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1447 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1448 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1449 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1450 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1451 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1454 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1457 * runqueue iterator, to support SMP load-balancing between different
1458 * scheduling classes, without having to expose their internal data
1459 * structures to the load-balancing proper:
1461 struct rq_iterator {
1463 struct task_struct *(*start)(void *);
1464 struct task_struct *(*next)(void *);
1468 static unsigned long
1469 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1470 unsigned long max_load_move, struct sched_domain *sd,
1471 enum cpu_idle_type idle, int *all_pinned,
1472 int *this_best_prio, struct rq_iterator *iterator);
1475 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1476 struct sched_domain *sd, enum cpu_idle_type idle,
1477 struct rq_iterator *iterator);
1480 #ifdef CONFIG_CGROUP_CPUACCT
1481 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1483 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1486 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1488 update_load_add(&rq->load, load);
1491 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1493 update_load_sub(&rq->load, load);
1497 static unsigned long source_load(int cpu, int type);
1498 static unsigned long target_load(int cpu, int type);
1499 static unsigned long cpu_avg_load_per_task(int cpu);
1500 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1501 #else /* CONFIG_SMP */
1503 #ifdef CONFIG_FAIR_GROUP_SCHED
1504 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1509 #endif /* CONFIG_SMP */
1511 #include "sched_stats.h"
1512 #include "sched_idletask.c"
1513 #include "sched_fair.c"
1514 #include "sched_rt.c"
1515 #ifdef CONFIG_SCHED_DEBUG
1516 # include "sched_debug.c"
1519 #define sched_class_highest (&rt_sched_class)
1521 static inline void inc_load(struct rq *rq, const struct task_struct *p)
1523 update_load_add(&rq->load, p->se.load.weight);
1526 static inline void dec_load(struct rq *rq, const struct task_struct *p)
1528 update_load_sub(&rq->load, p->se.load.weight);
1531 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1537 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1543 static void set_load_weight(struct task_struct *p)
1545 if (task_has_rt_policy(p)) {
1546 p->se.load.weight = prio_to_weight[0] * 2;
1547 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1552 * SCHED_IDLE tasks get minimal weight:
1554 if (p->policy == SCHED_IDLE) {
1555 p->se.load.weight = WEIGHT_IDLEPRIO;
1556 p->se.load.inv_weight = WMULT_IDLEPRIO;
1560 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1561 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1564 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1566 sched_info_queued(p);
1567 p->sched_class->enqueue_task(rq, p, wakeup);
1571 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1573 p->sched_class->dequeue_task(rq, p, sleep);
1578 * __normal_prio - return the priority that is based on the static prio
1580 static inline int __normal_prio(struct task_struct *p)
1582 return p->static_prio;
1586 * Calculate the expected normal priority: i.e. priority
1587 * without taking RT-inheritance into account. Might be
1588 * boosted by interactivity modifiers. Changes upon fork,
1589 * setprio syscalls, and whenever the interactivity
1590 * estimator recalculates.
1592 static inline int normal_prio(struct task_struct *p)
1596 if (task_has_rt_policy(p))
1597 prio = MAX_RT_PRIO-1 - p->rt_priority;
1599 prio = __normal_prio(p);
1604 * Calculate the current priority, i.e. the priority
1605 * taken into account by the scheduler. This value might
1606 * be boosted by RT tasks, or might be boosted by
1607 * interactivity modifiers. Will be RT if the task got
1608 * RT-boosted. If not then it returns p->normal_prio.
1610 static int effective_prio(struct task_struct *p)
1612 p->normal_prio = normal_prio(p);
1614 * If we are RT tasks or we were boosted to RT priority,
1615 * keep the priority unchanged. Otherwise, update priority
1616 * to the normal priority:
1618 if (!rt_prio(p->prio))
1619 return p->normal_prio;
1624 * activate_task - move a task to the runqueue.
1626 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1628 if (task_contributes_to_load(p))
1629 rq->nr_uninterruptible--;
1631 enqueue_task(rq, p, wakeup);
1632 inc_nr_running(p, rq);
1636 * deactivate_task - remove a task from the runqueue.
1638 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1640 if (task_contributes_to_load(p))
1641 rq->nr_uninterruptible++;
1643 dequeue_task(rq, p, sleep);
1644 dec_nr_running(p, rq);
1648 * task_curr - is this task currently executing on a CPU?
1649 * @p: the task in question.
1651 inline int task_curr(const struct task_struct *p)
1653 return cpu_curr(task_cpu(p)) == p;
1656 /* Used instead of source_load when we know the type == 0 */
1657 unsigned long weighted_cpuload(const int cpu)
1659 return cpu_rq(cpu)->load.weight;
1662 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1664 set_task_rq(p, cpu);
1667 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1668 * successfuly executed on another CPU. We must ensure that updates of
1669 * per-task data have been completed by this moment.
1672 task_thread_info(p)->cpu = cpu;
1676 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1677 const struct sched_class *prev_class,
1678 int oldprio, int running)
1680 if (prev_class != p->sched_class) {
1681 if (prev_class->switched_from)
1682 prev_class->switched_from(rq, p, running);
1683 p->sched_class->switched_to(rq, p, running);
1685 p->sched_class->prio_changed(rq, p, oldprio, running);
1691 * Is this task likely cache-hot:
1694 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1699 * Buddy candidates are cache hot:
1701 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1704 if (p->sched_class != &fair_sched_class)
1707 if (sysctl_sched_migration_cost == -1)
1709 if (sysctl_sched_migration_cost == 0)
1712 delta = now - p->se.exec_start;
1714 return delta < (s64)sysctl_sched_migration_cost;
1718 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1720 int old_cpu = task_cpu(p);
1721 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1722 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1723 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1726 clock_offset = old_rq->clock - new_rq->clock;
1728 #ifdef CONFIG_SCHEDSTATS
1729 if (p->se.wait_start)
1730 p->se.wait_start -= clock_offset;
1731 if (p->se.sleep_start)
1732 p->se.sleep_start -= clock_offset;
1733 if (p->se.block_start)
1734 p->se.block_start -= clock_offset;
1735 if (old_cpu != new_cpu) {
1736 schedstat_inc(p, se.nr_migrations);
1737 if (task_hot(p, old_rq->clock, NULL))
1738 schedstat_inc(p, se.nr_forced2_migrations);
1741 p->se.vruntime -= old_cfsrq->min_vruntime -
1742 new_cfsrq->min_vruntime;
1744 __set_task_cpu(p, new_cpu);
1747 struct migration_req {
1748 struct list_head list;
1750 struct task_struct *task;
1753 struct completion done;
1757 * The task's runqueue lock must be held.
1758 * Returns true if you have to wait for migration thread.
1761 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1763 struct rq *rq = task_rq(p);
1766 * If the task is not on a runqueue (and not running), then
1767 * it is sufficient to simply update the task's cpu field.
1769 if (!p->se.on_rq && !task_running(rq, p)) {
1770 set_task_cpu(p, dest_cpu);
1774 init_completion(&req->done);
1776 req->dest_cpu = dest_cpu;
1777 list_add(&req->list, &rq->migration_queue);
1783 * wait_task_inactive - wait for a thread to unschedule.
1785 * The caller must ensure that the task *will* unschedule sometime soon,
1786 * else this function might spin for a *long* time. This function can't
1787 * be called with interrupts off, or it may introduce deadlock with
1788 * smp_call_function() if an IPI is sent by the same process we are
1789 * waiting to become inactive.
1791 void wait_task_inactive(struct task_struct *p)
1793 unsigned long flags;
1799 * We do the initial early heuristics without holding
1800 * any task-queue locks at all. We'll only try to get
1801 * the runqueue lock when things look like they will
1807 * If the task is actively running on another CPU
1808 * still, just relax and busy-wait without holding
1811 * NOTE! Since we don't hold any locks, it's not
1812 * even sure that "rq" stays as the right runqueue!
1813 * But we don't care, since "task_running()" will
1814 * return false if the runqueue has changed and p
1815 * is actually now running somewhere else!
1817 while (task_running(rq, p))
1821 * Ok, time to look more closely! We need the rq
1822 * lock now, to be *sure*. If we're wrong, we'll
1823 * just go back and repeat.
1825 rq = task_rq_lock(p, &flags);
1826 running = task_running(rq, p);
1827 on_rq = p->se.on_rq;
1828 task_rq_unlock(rq, &flags);
1831 * Was it really running after all now that we
1832 * checked with the proper locks actually held?
1834 * Oops. Go back and try again..
1836 if (unlikely(running)) {
1842 * It's not enough that it's not actively running,
1843 * it must be off the runqueue _entirely_, and not
1846 * So if it wa still runnable (but just not actively
1847 * running right now), it's preempted, and we should
1848 * yield - it could be a while.
1850 if (unlikely(on_rq)) {
1851 schedule_timeout_uninterruptible(1);
1856 * Ahh, all good. It wasn't running, and it wasn't
1857 * runnable, which means that it will never become
1858 * running in the future either. We're all done!
1865 * kick_process - kick a running thread to enter/exit the kernel
1866 * @p: the to-be-kicked thread
1868 * Cause a process which is running on another CPU to enter
1869 * kernel-mode, without any delay. (to get signals handled.)
1871 * NOTE: this function doesnt have to take the runqueue lock,
1872 * because all it wants to ensure is that the remote task enters
1873 * the kernel. If the IPI races and the task has been migrated
1874 * to another CPU then no harm is done and the purpose has been
1877 void kick_process(struct task_struct *p)
1883 if ((cpu != smp_processor_id()) && task_curr(p))
1884 smp_send_reschedule(cpu);
1889 * Return a low guess at the load of a migration-source cpu weighted
1890 * according to the scheduling class and "nice" value.
1892 * We want to under-estimate the load of migration sources, to
1893 * balance conservatively.
1895 static unsigned long source_load(int cpu, int type)
1897 struct rq *rq = cpu_rq(cpu);
1898 unsigned long total = weighted_cpuload(cpu);
1903 return min(rq->cpu_load[type-1], total);
1907 * Return a high guess at the load of a migration-target cpu weighted
1908 * according to the scheduling class and "nice" value.
1910 static unsigned long target_load(int cpu, int type)
1912 struct rq *rq = cpu_rq(cpu);
1913 unsigned long total = weighted_cpuload(cpu);
1918 return max(rq->cpu_load[type-1], total);
1922 * Return the average load per task on the cpu's run queue
1924 static unsigned long cpu_avg_load_per_task(int cpu)
1926 struct rq *rq = cpu_rq(cpu);
1927 unsigned long total = weighted_cpuload(cpu);
1928 unsigned long n = rq->nr_running;
1930 return n ? total / n : SCHED_LOAD_SCALE;
1934 * find_idlest_group finds and returns the least busy CPU group within the
1937 static struct sched_group *
1938 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1940 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1941 unsigned long min_load = ULONG_MAX, this_load = 0;
1942 int load_idx = sd->forkexec_idx;
1943 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1946 unsigned long load, avg_load;
1950 /* Skip over this group if it has no CPUs allowed */
1951 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1954 local_group = cpu_isset(this_cpu, group->cpumask);
1956 /* Tally up the load of all CPUs in the group */
1959 for_each_cpu_mask(i, group->cpumask) {
1960 /* Bias balancing toward cpus of our domain */
1962 load = source_load(i, load_idx);
1964 load = target_load(i, load_idx);
1969 /* Adjust by relative CPU power of the group */
1970 avg_load = sg_div_cpu_power(group,
1971 avg_load * SCHED_LOAD_SCALE);
1974 this_load = avg_load;
1976 } else if (avg_load < min_load) {
1977 min_load = avg_load;
1980 } while (group = group->next, group != sd->groups);
1982 if (!idlest || 100*this_load < imbalance*min_load)
1988 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1991 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
1994 unsigned long load, min_load = ULONG_MAX;
1998 /* Traverse only the allowed CPUs */
1999 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2001 for_each_cpu_mask(i, *tmp) {
2002 load = weighted_cpuload(i);
2004 if (load < min_load || (load == min_load && i == this_cpu)) {
2014 * sched_balance_self: balance the current task (running on cpu) in domains
2015 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2018 * Balance, ie. select the least loaded group.
2020 * Returns the target CPU number, or the same CPU if no balancing is needed.
2022 * preempt must be disabled.
2024 static int sched_balance_self(int cpu, int flag)
2026 struct task_struct *t = current;
2027 struct sched_domain *tmp, *sd = NULL;
2029 for_each_domain(cpu, tmp) {
2031 * If power savings logic is enabled for a domain, stop there.
2033 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2035 if (tmp->flags & flag)
2040 cpumask_t span, tmpmask;
2041 struct sched_group *group;
2042 int new_cpu, weight;
2044 if (!(sd->flags & flag)) {
2050 group = find_idlest_group(sd, t, cpu);
2056 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2057 if (new_cpu == -1 || new_cpu == cpu) {
2058 /* Now try balancing at a lower domain level of cpu */
2063 /* Now try balancing at a lower domain level of new_cpu */
2066 weight = cpus_weight(span);
2067 for_each_domain(cpu, tmp) {
2068 if (weight <= cpus_weight(tmp->span))
2070 if (tmp->flags & flag)
2073 /* while loop will break here if sd == NULL */
2079 #endif /* CONFIG_SMP */
2082 * try_to_wake_up - wake up a thread
2083 * @p: the to-be-woken-up thread
2084 * @state: the mask of task states that can be woken
2085 * @sync: do a synchronous wakeup?
2087 * Put it on the run-queue if it's not already there. The "current"
2088 * thread is always on the run-queue (except when the actual
2089 * re-schedule is in progress), and as such you're allowed to do
2090 * the simpler "current->state = TASK_RUNNING" to mark yourself
2091 * runnable without the overhead of this.
2093 * returns failure only if the task is already active.
2095 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2097 int cpu, orig_cpu, this_cpu, success = 0;
2098 unsigned long flags;
2102 if (!sched_feat(SYNC_WAKEUPS))
2106 rq = task_rq_lock(p, &flags);
2107 old_state = p->state;
2108 if (!(old_state & state))
2116 this_cpu = smp_processor_id();
2119 if (unlikely(task_running(rq, p)))
2122 cpu = p->sched_class->select_task_rq(p, sync);
2123 if (cpu != orig_cpu) {
2124 set_task_cpu(p, cpu);
2125 task_rq_unlock(rq, &flags);
2126 /* might preempt at this point */
2127 rq = task_rq_lock(p, &flags);
2128 old_state = p->state;
2129 if (!(old_state & state))
2134 this_cpu = smp_processor_id();
2138 #ifdef CONFIG_SCHEDSTATS
2139 schedstat_inc(rq, ttwu_count);
2140 if (cpu == this_cpu)
2141 schedstat_inc(rq, ttwu_local);
2143 struct sched_domain *sd;
2144 for_each_domain(this_cpu, sd) {
2145 if (cpu_isset(cpu, sd->span)) {
2146 schedstat_inc(sd, ttwu_wake_remote);
2154 #endif /* CONFIG_SMP */
2155 schedstat_inc(p, se.nr_wakeups);
2157 schedstat_inc(p, se.nr_wakeups_sync);
2158 if (orig_cpu != cpu)
2159 schedstat_inc(p, se.nr_wakeups_migrate);
2160 if (cpu == this_cpu)
2161 schedstat_inc(p, se.nr_wakeups_local);
2163 schedstat_inc(p, se.nr_wakeups_remote);
2164 update_rq_clock(rq);
2165 activate_task(rq, p, 1);
2169 trace_mark(kernel_sched_wakeup,
2170 "pid %d state %ld ## rq %p task %p rq->curr %p",
2171 p->pid, p->state, rq, p, rq->curr);
2172 check_preempt_curr(rq, p);
2174 p->state = TASK_RUNNING;
2176 if (p->sched_class->task_wake_up)
2177 p->sched_class->task_wake_up(rq, p);
2180 task_rq_unlock(rq, &flags);
2185 int wake_up_process(struct task_struct *p)
2187 return try_to_wake_up(p, TASK_ALL, 0);
2189 EXPORT_SYMBOL(wake_up_process);
2191 int wake_up_state(struct task_struct *p, unsigned int state)
2193 return try_to_wake_up(p, state, 0);
2197 * Perform scheduler related setup for a newly forked process p.
2198 * p is forked by current.
2200 * __sched_fork() is basic setup used by init_idle() too:
2202 static void __sched_fork(struct task_struct *p)
2204 p->se.exec_start = 0;
2205 p->se.sum_exec_runtime = 0;
2206 p->se.prev_sum_exec_runtime = 0;
2207 p->se.last_wakeup = 0;
2208 p->se.avg_overlap = 0;
2210 #ifdef CONFIG_SCHEDSTATS
2211 p->se.wait_start = 0;
2212 p->se.sum_sleep_runtime = 0;
2213 p->se.sleep_start = 0;
2214 p->se.block_start = 0;
2215 p->se.sleep_max = 0;
2216 p->se.block_max = 0;
2218 p->se.slice_max = 0;
2222 INIT_LIST_HEAD(&p->rt.run_list);
2224 INIT_LIST_HEAD(&p->se.group_node);
2226 #ifdef CONFIG_PREEMPT_NOTIFIERS
2227 INIT_HLIST_HEAD(&p->preempt_notifiers);
2231 * We mark the process as running here, but have not actually
2232 * inserted it onto the runqueue yet. This guarantees that
2233 * nobody will actually run it, and a signal or other external
2234 * event cannot wake it up and insert it on the runqueue either.
2236 p->state = TASK_RUNNING;
2240 * fork()/clone()-time setup:
2242 void sched_fork(struct task_struct *p, int clone_flags)
2244 int cpu = get_cpu();
2249 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2251 set_task_cpu(p, cpu);
2254 * Make sure we do not leak PI boosting priority to the child:
2256 p->prio = current->normal_prio;
2257 if (!rt_prio(p->prio))
2258 p->sched_class = &fair_sched_class;
2260 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2261 if (likely(sched_info_on()))
2262 memset(&p->sched_info, 0, sizeof(p->sched_info));
2264 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2267 #ifdef CONFIG_PREEMPT
2268 /* Want to start with kernel preemption disabled. */
2269 task_thread_info(p)->preempt_count = 1;
2275 * wake_up_new_task - wake up a newly created task for the first time.
2277 * This function will do some initial scheduler statistics housekeeping
2278 * that must be done for every newly created context, then puts the task
2279 * on the runqueue and wakes it.
2281 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2283 unsigned long flags;
2286 rq = task_rq_lock(p, &flags);
2287 BUG_ON(p->state != TASK_RUNNING);
2288 update_rq_clock(rq);
2290 p->prio = effective_prio(p);
2292 if (!p->sched_class->task_new || !current->se.on_rq) {
2293 activate_task(rq, p, 0);
2296 * Let the scheduling class do new task startup
2297 * management (if any):
2299 p->sched_class->task_new(rq, p);
2300 inc_nr_running(p, rq);
2302 trace_mark(kernel_sched_wakeup_new,
2303 "pid %d state %ld ## rq %p task %p rq->curr %p",
2304 p->pid, p->state, rq, p, rq->curr);
2305 check_preempt_curr(rq, p);
2307 if (p->sched_class->task_wake_up)
2308 p->sched_class->task_wake_up(rq, p);
2310 task_rq_unlock(rq, &flags);
2313 #ifdef CONFIG_PREEMPT_NOTIFIERS
2316 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2317 * @notifier: notifier struct to register
2319 void preempt_notifier_register(struct preempt_notifier *notifier)
2321 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2323 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2326 * preempt_notifier_unregister - no longer interested in preemption notifications
2327 * @notifier: notifier struct to unregister
2329 * This is safe to call from within a preemption notifier.
2331 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2333 hlist_del(¬ifier->link);
2335 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2337 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2339 struct preempt_notifier *notifier;
2340 struct hlist_node *node;
2342 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2343 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2347 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2348 struct task_struct *next)
2350 struct preempt_notifier *notifier;
2351 struct hlist_node *node;
2353 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2354 notifier->ops->sched_out(notifier, next);
2359 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2364 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2365 struct task_struct *next)
2372 * prepare_task_switch - prepare to switch tasks
2373 * @rq: the runqueue preparing to switch
2374 * @prev: the current task that is being switched out
2375 * @next: the task we are going to switch to.
2377 * This is called with the rq lock held and interrupts off. It must
2378 * be paired with a subsequent finish_task_switch after the context
2381 * prepare_task_switch sets up locking and calls architecture specific
2385 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2386 struct task_struct *next)
2388 fire_sched_out_preempt_notifiers(prev, next);
2389 prepare_lock_switch(rq, next);
2390 prepare_arch_switch(next);
2394 * finish_task_switch - clean up after a task-switch
2395 * @rq: runqueue associated with task-switch
2396 * @prev: the thread we just switched away from.
2398 * finish_task_switch must be called after the context switch, paired
2399 * with a prepare_task_switch call before the context switch.
2400 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2401 * and do any other architecture-specific cleanup actions.
2403 * Note that we may have delayed dropping an mm in context_switch(). If
2404 * so, we finish that here outside of the runqueue lock. (Doing it
2405 * with the lock held can cause deadlocks; see schedule() for
2408 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2409 __releases(rq->lock)
2411 struct mm_struct *mm = rq->prev_mm;
2417 * A task struct has one reference for the use as "current".
2418 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2419 * schedule one last time. The schedule call will never return, and
2420 * the scheduled task must drop that reference.
2421 * The test for TASK_DEAD must occur while the runqueue locks are
2422 * still held, otherwise prev could be scheduled on another cpu, die
2423 * there before we look at prev->state, and then the reference would
2425 * Manfred Spraul <manfred@colorfullife.com>
2427 prev_state = prev->state;
2428 finish_arch_switch(prev);
2429 finish_lock_switch(rq, prev);
2431 if (current->sched_class->post_schedule)
2432 current->sched_class->post_schedule(rq);
2435 fire_sched_in_preempt_notifiers(current);
2438 if (unlikely(prev_state == TASK_DEAD)) {
2440 * Remove function-return probe instances associated with this
2441 * task and put them back on the free list.
2443 kprobe_flush_task(prev);
2444 put_task_struct(prev);
2449 * schedule_tail - first thing a freshly forked thread must call.
2450 * @prev: the thread we just switched away from.
2452 asmlinkage void schedule_tail(struct task_struct *prev)
2453 __releases(rq->lock)
2455 struct rq *rq = this_rq();
2457 finish_task_switch(rq, prev);
2458 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2459 /* In this case, finish_task_switch does not reenable preemption */
2462 if (current->set_child_tid)
2463 put_user(task_pid_vnr(current), current->set_child_tid);
2467 * context_switch - switch to the new MM and the new
2468 * thread's register state.
2471 context_switch(struct rq *rq, struct task_struct *prev,
2472 struct task_struct *next)
2474 struct mm_struct *mm, *oldmm;
2476 prepare_task_switch(rq, prev, next);
2477 trace_mark(kernel_sched_schedule,
2478 "prev_pid %d next_pid %d prev_state %ld "
2479 "## rq %p prev %p next %p",
2480 prev->pid, next->pid, prev->state,
2483 oldmm = prev->active_mm;
2485 * For paravirt, this is coupled with an exit in switch_to to
2486 * combine the page table reload and the switch backend into
2489 arch_enter_lazy_cpu_mode();
2491 if (unlikely(!mm)) {
2492 next->active_mm = oldmm;
2493 atomic_inc(&oldmm->mm_count);
2494 enter_lazy_tlb(oldmm, next);
2496 switch_mm(oldmm, mm, next);
2498 if (unlikely(!prev->mm)) {
2499 prev->active_mm = NULL;
2500 rq->prev_mm = oldmm;
2503 * Since the runqueue lock will be released by the next
2504 * task (which is an invalid locking op but in the case
2505 * of the scheduler it's an obvious special-case), so we
2506 * do an early lockdep release here:
2508 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2509 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2512 /* Here we just switch the register state and the stack. */
2513 switch_to(prev, next, prev);
2517 * this_rq must be evaluated again because prev may have moved
2518 * CPUs since it called schedule(), thus the 'rq' on its stack
2519 * frame will be invalid.
2521 finish_task_switch(this_rq(), prev);
2525 * nr_running, nr_uninterruptible and nr_context_switches:
2527 * externally visible scheduler statistics: current number of runnable
2528 * threads, current number of uninterruptible-sleeping threads, total
2529 * number of context switches performed since bootup.
2531 unsigned long nr_running(void)
2533 unsigned long i, sum = 0;
2535 for_each_online_cpu(i)
2536 sum += cpu_rq(i)->nr_running;
2541 unsigned long nr_uninterruptible(void)
2543 unsigned long i, sum = 0;
2545 for_each_possible_cpu(i)
2546 sum += cpu_rq(i)->nr_uninterruptible;
2549 * Since we read the counters lockless, it might be slightly
2550 * inaccurate. Do not allow it to go below zero though:
2552 if (unlikely((long)sum < 0))
2558 unsigned long long nr_context_switches(void)
2561 unsigned long long sum = 0;
2563 for_each_possible_cpu(i)
2564 sum += cpu_rq(i)->nr_switches;
2569 unsigned long nr_iowait(void)
2571 unsigned long i, sum = 0;
2573 for_each_possible_cpu(i)
2574 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2579 unsigned long nr_active(void)
2581 unsigned long i, running = 0, uninterruptible = 0;
2583 for_each_online_cpu(i) {
2584 running += cpu_rq(i)->nr_running;
2585 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2588 if (unlikely((long)uninterruptible < 0))
2589 uninterruptible = 0;
2591 return running + uninterruptible;
2595 * Update rq->cpu_load[] statistics. This function is usually called every
2596 * scheduler tick (TICK_NSEC).
2598 static void update_cpu_load(struct rq *this_rq)
2600 unsigned long this_load = this_rq->load.weight;
2603 this_rq->nr_load_updates++;
2605 /* Update our load: */
2606 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2607 unsigned long old_load, new_load;
2609 /* scale is effectively 1 << i now, and >> i divides by scale */
2611 old_load = this_rq->cpu_load[i];
2612 new_load = this_load;
2614 * Round up the averaging division if load is increasing. This
2615 * prevents us from getting stuck on 9 if the load is 10, for
2618 if (new_load > old_load)
2619 new_load += scale-1;
2620 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2627 * double_rq_lock - safely lock two runqueues
2629 * Note this does not disable interrupts like task_rq_lock,
2630 * you need to do so manually before calling.
2632 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2633 __acquires(rq1->lock)
2634 __acquires(rq2->lock)
2636 BUG_ON(!irqs_disabled());
2638 spin_lock(&rq1->lock);
2639 __acquire(rq2->lock); /* Fake it out ;) */
2642 spin_lock(&rq1->lock);
2643 spin_lock(&rq2->lock);
2645 spin_lock(&rq2->lock);
2646 spin_lock(&rq1->lock);
2649 update_rq_clock(rq1);
2650 update_rq_clock(rq2);
2654 * double_rq_unlock - safely unlock two runqueues
2656 * Note this does not restore interrupts like task_rq_unlock,
2657 * you need to do so manually after calling.
2659 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2660 __releases(rq1->lock)
2661 __releases(rq2->lock)
2663 spin_unlock(&rq1->lock);
2665 spin_unlock(&rq2->lock);
2667 __release(rq2->lock);
2671 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2673 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2674 __releases(this_rq->lock)
2675 __acquires(busiest->lock)
2676 __acquires(this_rq->lock)
2680 if (unlikely(!irqs_disabled())) {
2681 /* printk() doesn't work good under rq->lock */
2682 spin_unlock(&this_rq->lock);
2685 if (unlikely(!spin_trylock(&busiest->lock))) {
2686 if (busiest < this_rq) {
2687 spin_unlock(&this_rq->lock);
2688 spin_lock(&busiest->lock);
2689 spin_lock(&this_rq->lock);
2692 spin_lock(&busiest->lock);
2698 * If dest_cpu is allowed for this process, migrate the task to it.
2699 * This is accomplished by forcing the cpu_allowed mask to only
2700 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2701 * the cpu_allowed mask is restored.
2703 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2705 struct migration_req req;
2706 unsigned long flags;
2709 rq = task_rq_lock(p, &flags);
2710 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2711 || unlikely(cpu_is_offline(dest_cpu)))
2714 /* force the process onto the specified CPU */
2715 if (migrate_task(p, dest_cpu, &req)) {
2716 /* Need to wait for migration thread (might exit: take ref). */
2717 struct task_struct *mt = rq->migration_thread;
2719 get_task_struct(mt);
2720 task_rq_unlock(rq, &flags);
2721 wake_up_process(mt);
2722 put_task_struct(mt);
2723 wait_for_completion(&req.done);
2728 task_rq_unlock(rq, &flags);
2732 * sched_exec - execve() is a valuable balancing opportunity, because at
2733 * this point the task has the smallest effective memory and cache footprint.
2735 void sched_exec(void)
2737 int new_cpu, this_cpu = get_cpu();
2738 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2740 if (new_cpu != this_cpu)
2741 sched_migrate_task(current, new_cpu);
2745 * pull_task - move a task from a remote runqueue to the local runqueue.
2746 * Both runqueues must be locked.
2748 static void pull_task(struct rq *src_rq, struct task_struct *p,
2749 struct rq *this_rq, int this_cpu)
2751 deactivate_task(src_rq, p, 0);
2752 set_task_cpu(p, this_cpu);
2753 activate_task(this_rq, p, 0);
2755 * Note that idle threads have a prio of MAX_PRIO, for this test
2756 * to be always true for them.
2758 check_preempt_curr(this_rq, p);
2762 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2765 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2766 struct sched_domain *sd, enum cpu_idle_type idle,
2770 * We do not migrate tasks that are:
2771 * 1) running (obviously), or
2772 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2773 * 3) are cache-hot on their current CPU.
2775 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2776 schedstat_inc(p, se.nr_failed_migrations_affine);
2781 if (task_running(rq, p)) {
2782 schedstat_inc(p, se.nr_failed_migrations_running);
2787 * Aggressive migration if:
2788 * 1) task is cache cold, or
2789 * 2) too many balance attempts have failed.
2792 if (!task_hot(p, rq->clock, sd) ||
2793 sd->nr_balance_failed > sd->cache_nice_tries) {
2794 #ifdef CONFIG_SCHEDSTATS
2795 if (task_hot(p, rq->clock, sd)) {
2796 schedstat_inc(sd, lb_hot_gained[idle]);
2797 schedstat_inc(p, se.nr_forced_migrations);
2803 if (task_hot(p, rq->clock, sd)) {
2804 schedstat_inc(p, se.nr_failed_migrations_hot);
2810 static unsigned long
2811 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2812 unsigned long max_load_move, struct sched_domain *sd,
2813 enum cpu_idle_type idle, int *all_pinned,
2814 int *this_best_prio, struct rq_iterator *iterator)
2816 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2817 struct task_struct *p;
2818 long rem_load_move = max_load_move;
2820 if (max_load_move == 0)
2826 * Start the load-balancing iterator:
2828 p = iterator->start(iterator->arg);
2830 if (!p || loops++ > sysctl_sched_nr_migrate)
2833 * To help distribute high priority tasks across CPUs we don't
2834 * skip a task if it will be the highest priority task (i.e. smallest
2835 * prio value) on its new queue regardless of its load weight
2837 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2838 SCHED_LOAD_SCALE_FUZZ;
2839 if ((skip_for_load && p->prio >= *this_best_prio) ||
2840 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2841 p = iterator->next(iterator->arg);
2845 pull_task(busiest, p, this_rq, this_cpu);
2847 rem_load_move -= p->se.load.weight;
2850 * We only want to steal up to the prescribed amount of weighted load.
2852 if (rem_load_move > 0) {
2853 if (p->prio < *this_best_prio)
2854 *this_best_prio = p->prio;
2855 p = iterator->next(iterator->arg);
2860 * Right now, this is one of only two places pull_task() is called,
2861 * so we can safely collect pull_task() stats here rather than
2862 * inside pull_task().
2864 schedstat_add(sd, lb_gained[idle], pulled);
2867 *all_pinned = pinned;
2869 return max_load_move - rem_load_move;
2873 * move_tasks tries to move up to max_load_move weighted load from busiest to
2874 * this_rq, as part of a balancing operation within domain "sd".
2875 * Returns 1 if successful and 0 otherwise.
2877 * Called with both runqueues locked.
2879 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2880 unsigned long max_load_move,
2881 struct sched_domain *sd, enum cpu_idle_type idle,
2884 const struct sched_class *class = sched_class_highest;
2885 unsigned long total_load_moved = 0;
2886 int this_best_prio = this_rq->curr->prio;
2890 class->load_balance(this_rq, this_cpu, busiest,
2891 max_load_move - total_load_moved,
2892 sd, idle, all_pinned, &this_best_prio);
2893 class = class->next;
2894 } while (class && max_load_move > total_load_moved);
2896 return total_load_moved > 0;
2900 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2901 struct sched_domain *sd, enum cpu_idle_type idle,
2902 struct rq_iterator *iterator)
2904 struct task_struct *p = iterator->start(iterator->arg);
2908 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2909 pull_task(busiest, p, this_rq, this_cpu);
2911 * Right now, this is only the second place pull_task()
2912 * is called, so we can safely collect pull_task()
2913 * stats here rather than inside pull_task().
2915 schedstat_inc(sd, lb_gained[idle]);
2919 p = iterator->next(iterator->arg);
2926 * move_one_task tries to move exactly one task from busiest to this_rq, as
2927 * part of active balancing operations within "domain".
2928 * Returns 1 if successful and 0 otherwise.
2930 * Called with both runqueues locked.
2932 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2933 struct sched_domain *sd, enum cpu_idle_type idle)
2935 const struct sched_class *class;
2937 for (class = sched_class_highest; class; class = class->next)
2938 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2945 * find_busiest_group finds and returns the busiest CPU group within the
2946 * domain. It calculates and returns the amount of weighted load which
2947 * should be moved to restore balance via the imbalance parameter.
2949 static struct sched_group *
2950 find_busiest_group(struct sched_domain *sd, int this_cpu,
2951 unsigned long *imbalance, enum cpu_idle_type idle,
2952 int *sd_idle, const cpumask_t *cpus, int *balance)
2954 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2955 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2956 unsigned long max_pull;
2957 unsigned long busiest_load_per_task, busiest_nr_running;
2958 unsigned long this_load_per_task, this_nr_running;
2959 int load_idx, group_imb = 0;
2960 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2961 int power_savings_balance = 1;
2962 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2963 unsigned long min_nr_running = ULONG_MAX;
2964 struct sched_group *group_min = NULL, *group_leader = NULL;
2967 max_load = this_load = total_load = total_pwr = 0;
2968 busiest_load_per_task = busiest_nr_running = 0;
2969 this_load_per_task = this_nr_running = 0;
2970 if (idle == CPU_NOT_IDLE)
2971 load_idx = sd->busy_idx;
2972 else if (idle == CPU_NEWLY_IDLE)
2973 load_idx = sd->newidle_idx;
2975 load_idx = sd->idle_idx;
2978 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2981 int __group_imb = 0;
2982 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2983 unsigned long sum_nr_running, sum_weighted_load;
2985 local_group = cpu_isset(this_cpu, group->cpumask);
2988 balance_cpu = first_cpu(group->cpumask);
2990 /* Tally up the load of all CPUs in the group */
2991 sum_weighted_load = sum_nr_running = avg_load = 0;
2993 min_cpu_load = ~0UL;
2995 for_each_cpu_mask(i, group->cpumask) {
2998 if (!cpu_isset(i, *cpus))
3003 if (*sd_idle && rq->nr_running)
3006 /* Bias balancing toward cpus of our domain */
3008 if (idle_cpu(i) && !first_idle_cpu) {
3013 load = target_load(i, load_idx);
3015 load = source_load(i, load_idx);
3016 if (load > max_cpu_load)
3017 max_cpu_load = load;
3018 if (min_cpu_load > load)
3019 min_cpu_load = load;
3023 sum_nr_running += rq->nr_running;
3024 sum_weighted_load += weighted_cpuload(i);
3028 * First idle cpu or the first cpu(busiest) in this sched group
3029 * is eligible for doing load balancing at this and above
3030 * domains. In the newly idle case, we will allow all the cpu's
3031 * to do the newly idle load balance.
3033 if (idle != CPU_NEWLY_IDLE && local_group &&
3034 balance_cpu != this_cpu && balance) {
3039 total_load += avg_load;
3040 total_pwr += group->__cpu_power;
3042 /* Adjust by relative CPU power of the group */
3043 avg_load = sg_div_cpu_power(group,
3044 avg_load * SCHED_LOAD_SCALE);
3046 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3049 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3052 this_load = avg_load;
3054 this_nr_running = sum_nr_running;
3055 this_load_per_task = sum_weighted_load;
3056 } else if (avg_load > max_load &&
3057 (sum_nr_running > group_capacity || __group_imb)) {
3058 max_load = avg_load;
3060 busiest_nr_running = sum_nr_running;
3061 busiest_load_per_task = sum_weighted_load;
3062 group_imb = __group_imb;
3065 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3067 * Busy processors will not participate in power savings
3070 if (idle == CPU_NOT_IDLE ||
3071 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3075 * If the local group is idle or completely loaded
3076 * no need to do power savings balance at this domain
3078 if (local_group && (this_nr_running >= group_capacity ||
3080 power_savings_balance = 0;
3083 * If a group is already running at full capacity or idle,
3084 * don't include that group in power savings calculations
3086 if (!power_savings_balance || sum_nr_running >= group_capacity
3091 * Calculate the group which has the least non-idle load.
3092 * This is the group from where we need to pick up the load
3095 if ((sum_nr_running < min_nr_running) ||
3096 (sum_nr_running == min_nr_running &&
3097 first_cpu(group->cpumask) <
3098 first_cpu(group_min->cpumask))) {
3100 min_nr_running = sum_nr_running;
3101 min_load_per_task = sum_weighted_load /
3106 * Calculate the group which is almost near its
3107 * capacity but still has some space to pick up some load
3108 * from other group and save more power
3110 if (sum_nr_running <= group_capacity - 1) {
3111 if (sum_nr_running > leader_nr_running ||
3112 (sum_nr_running == leader_nr_running &&
3113 first_cpu(group->cpumask) >
3114 first_cpu(group_leader->cpumask))) {
3115 group_leader = group;
3116 leader_nr_running = sum_nr_running;
3121 group = group->next;
3122 } while (group != sd->groups);
3124 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3127 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3129 if (this_load >= avg_load ||
3130 100*max_load <= sd->imbalance_pct*this_load)
3133 busiest_load_per_task /= busiest_nr_running;
3135 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3138 * We're trying to get all the cpus to the average_load, so we don't
3139 * want to push ourselves above the average load, nor do we wish to
3140 * reduce the max loaded cpu below the average load, as either of these
3141 * actions would just result in more rebalancing later, and ping-pong
3142 * tasks around. Thus we look for the minimum possible imbalance.
3143 * Negative imbalances (*we* are more loaded than anyone else) will
3144 * be counted as no imbalance for these purposes -- we can't fix that
3145 * by pulling tasks to us. Be careful of negative numbers as they'll
3146 * appear as very large values with unsigned longs.
3148 if (max_load <= busiest_load_per_task)
3152 * In the presence of smp nice balancing, certain scenarios can have
3153 * max load less than avg load(as we skip the groups at or below
3154 * its cpu_power, while calculating max_load..)
3156 if (max_load < avg_load) {
3158 goto small_imbalance;
3161 /* Don't want to pull so many tasks that a group would go idle */
3162 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3164 /* How much load to actually move to equalise the imbalance */
3165 *imbalance = min(max_pull * busiest->__cpu_power,
3166 (avg_load - this_load) * this->__cpu_power)
3170 * if *imbalance is less than the average load per runnable task
3171 * there is no gaurantee that any tasks will be moved so we'll have
3172 * a think about bumping its value to force at least one task to be
3175 if (*imbalance < busiest_load_per_task) {
3176 unsigned long tmp, pwr_now, pwr_move;
3180 pwr_move = pwr_now = 0;
3182 if (this_nr_running) {
3183 this_load_per_task /= this_nr_running;
3184 if (busiest_load_per_task > this_load_per_task)
3187 this_load_per_task = SCHED_LOAD_SCALE;
3189 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3190 busiest_load_per_task * imbn) {
3191 *imbalance = busiest_load_per_task;
3196 * OK, we don't have enough imbalance to justify moving tasks,
3197 * however we may be able to increase total CPU power used by
3201 pwr_now += busiest->__cpu_power *
3202 min(busiest_load_per_task, max_load);
3203 pwr_now += this->__cpu_power *
3204 min(this_load_per_task, this_load);
3205 pwr_now /= SCHED_LOAD_SCALE;
3207 /* Amount of load we'd subtract */
3208 tmp = sg_div_cpu_power(busiest,
3209 busiest_load_per_task * SCHED_LOAD_SCALE);
3211 pwr_move += busiest->__cpu_power *
3212 min(busiest_load_per_task, max_load - tmp);
3214 /* Amount of load we'd add */
3215 if (max_load * busiest->__cpu_power <
3216 busiest_load_per_task * SCHED_LOAD_SCALE)
3217 tmp = sg_div_cpu_power(this,
3218 max_load * busiest->__cpu_power);
3220 tmp = sg_div_cpu_power(this,
3221 busiest_load_per_task * SCHED_LOAD_SCALE);
3222 pwr_move += this->__cpu_power *
3223 min(this_load_per_task, this_load + tmp);
3224 pwr_move /= SCHED_LOAD_SCALE;
3226 /* Move if we gain throughput */
3227 if (pwr_move > pwr_now)
3228 *imbalance = busiest_load_per_task;
3234 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3235 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3238 if (this == group_leader && group_leader != group_min) {
3239 *imbalance = min_load_per_task;
3249 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3252 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3253 unsigned long imbalance, const cpumask_t *cpus)
3255 struct rq *busiest = NULL, *rq;
3256 unsigned long max_load = 0;
3259 for_each_cpu_mask(i, group->cpumask) {
3262 if (!cpu_isset(i, *cpus))
3266 wl = weighted_cpuload(i);
3268 if (rq->nr_running == 1 && wl > imbalance)
3271 if (wl > max_load) {
3281 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3282 * so long as it is large enough.
3284 #define MAX_PINNED_INTERVAL 512
3287 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3288 * tasks if there is an imbalance.
3290 static int load_balance(int this_cpu, struct rq *this_rq,
3291 struct sched_domain *sd, enum cpu_idle_type idle,
3292 int *balance, cpumask_t *cpus)
3294 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3295 struct sched_group *group;
3296 unsigned long imbalance;
3298 unsigned long flags;
3303 * When power savings policy is enabled for the parent domain, idle
3304 * sibling can pick up load irrespective of busy siblings. In this case,
3305 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3306 * portraying it as CPU_NOT_IDLE.
3308 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3309 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3312 schedstat_inc(sd, lb_count[idle]);
3315 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3322 schedstat_inc(sd, lb_nobusyg[idle]);
3326 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3328 schedstat_inc(sd, lb_nobusyq[idle]);
3332 BUG_ON(busiest == this_rq);
3334 schedstat_add(sd, lb_imbalance[idle], imbalance);
3337 if (busiest->nr_running > 1) {
3339 * Attempt to move tasks. If find_busiest_group has found
3340 * an imbalance but busiest->nr_running <= 1, the group is
3341 * still unbalanced. ld_moved simply stays zero, so it is
3342 * correctly treated as an imbalance.
3344 local_irq_save(flags);
3345 double_rq_lock(this_rq, busiest);
3346 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3347 imbalance, sd, idle, &all_pinned);
3348 double_rq_unlock(this_rq, busiest);
3349 local_irq_restore(flags);
3352 * some other cpu did the load balance for us.
3354 if (ld_moved && this_cpu != smp_processor_id())
3355 resched_cpu(this_cpu);
3357 /* All tasks on this runqueue were pinned by CPU affinity */
3358 if (unlikely(all_pinned)) {
3359 cpu_clear(cpu_of(busiest), *cpus);
3360 if (!cpus_empty(*cpus))
3367 schedstat_inc(sd, lb_failed[idle]);
3368 sd->nr_balance_failed++;
3370 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3372 spin_lock_irqsave(&busiest->lock, flags);
3374 /* don't kick the migration_thread, if the curr
3375 * task on busiest cpu can't be moved to this_cpu
3377 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3378 spin_unlock_irqrestore(&busiest->lock, flags);
3380 goto out_one_pinned;
3383 if (!busiest->active_balance) {
3384 busiest->active_balance = 1;
3385 busiest->push_cpu = this_cpu;
3388 spin_unlock_irqrestore(&busiest->lock, flags);
3390 wake_up_process(busiest->migration_thread);
3393 * We've kicked active balancing, reset the failure
3396 sd->nr_balance_failed = sd->cache_nice_tries+1;
3399 sd->nr_balance_failed = 0;
3401 if (likely(!active_balance)) {
3402 /* We were unbalanced, so reset the balancing interval */
3403 sd->balance_interval = sd->min_interval;
3406 * If we've begun active balancing, start to back off. This
3407 * case may not be covered by the all_pinned logic if there
3408 * is only 1 task on the busy runqueue (because we don't call
3411 if (sd->balance_interval < sd->max_interval)
3412 sd->balance_interval *= 2;
3415 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3416 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3421 schedstat_inc(sd, lb_balanced[idle]);
3423 sd->nr_balance_failed = 0;
3426 /* tune up the balancing interval */
3427 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3428 (sd->balance_interval < sd->max_interval))
3429 sd->balance_interval *= 2;
3431 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3432 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3438 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3439 * tasks if there is an imbalance.
3441 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3442 * this_rq is locked.
3445 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3448 struct sched_group *group;
3449 struct rq *busiest = NULL;
3450 unsigned long imbalance;
3458 * When power savings policy is enabled for the parent domain, idle
3459 * sibling can pick up load irrespective of busy siblings. In this case,
3460 * let the state of idle sibling percolate up as IDLE, instead of
3461 * portraying it as CPU_NOT_IDLE.
3463 if (sd->flags & SD_SHARE_CPUPOWER &&
3464 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3467 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3469 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3470 &sd_idle, cpus, NULL);
3472 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3476 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3478 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3482 BUG_ON(busiest == this_rq);
3484 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3487 if (busiest->nr_running > 1) {
3488 /* Attempt to move tasks */
3489 double_lock_balance(this_rq, busiest);
3490 /* this_rq->clock is already updated */
3491 update_rq_clock(busiest);
3492 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3493 imbalance, sd, CPU_NEWLY_IDLE,
3495 spin_unlock(&busiest->lock);
3497 if (unlikely(all_pinned)) {
3498 cpu_clear(cpu_of(busiest), *cpus);
3499 if (!cpus_empty(*cpus))
3505 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3506 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3507 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3510 sd->nr_balance_failed = 0;
3515 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3516 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3517 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3519 sd->nr_balance_failed = 0;
3525 * idle_balance is called by schedule() if this_cpu is about to become
3526 * idle. Attempts to pull tasks from other CPUs.
3528 static void idle_balance(int this_cpu, struct rq *this_rq)
3530 struct sched_domain *sd;
3531 int pulled_task = -1;
3532 unsigned long next_balance = jiffies + HZ;
3535 for_each_domain(this_cpu, sd) {
3536 unsigned long interval;
3538 if (!(sd->flags & SD_LOAD_BALANCE))
3541 if (sd->flags & SD_BALANCE_NEWIDLE)
3542 /* If we've pulled tasks over stop searching: */
3543 pulled_task = load_balance_newidle(this_cpu, this_rq,
3546 interval = msecs_to_jiffies(sd->balance_interval);
3547 if (time_after(next_balance, sd->last_balance + interval))
3548 next_balance = sd->last_balance + interval;
3552 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3554 * We are going idle. next_balance may be set based on
3555 * a busy processor. So reset next_balance.
3557 this_rq->next_balance = next_balance;
3562 * active_load_balance is run by migration threads. It pushes running tasks
3563 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3564 * running on each physical CPU where possible, and avoids physical /
3565 * logical imbalances.
3567 * Called with busiest_rq locked.
3569 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3571 int target_cpu = busiest_rq->push_cpu;
3572 struct sched_domain *sd;
3573 struct rq *target_rq;
3575 /* Is there any task to move? */
3576 if (busiest_rq->nr_running <= 1)
3579 target_rq = cpu_rq(target_cpu);
3582 * This condition is "impossible", if it occurs
3583 * we need to fix it. Originally reported by
3584 * Bjorn Helgaas on a 128-cpu setup.
3586 BUG_ON(busiest_rq == target_rq);
3588 /* move a task from busiest_rq to target_rq */
3589 double_lock_balance(busiest_rq, target_rq);
3590 update_rq_clock(busiest_rq);
3591 update_rq_clock(target_rq);
3593 /* Search for an sd spanning us and the target CPU. */
3594 for_each_domain(target_cpu, sd) {
3595 if ((sd->flags & SD_LOAD_BALANCE) &&
3596 cpu_isset(busiest_cpu, sd->span))
3601 schedstat_inc(sd, alb_count);
3603 if (move_one_task(target_rq, target_cpu, busiest_rq,
3605 schedstat_inc(sd, alb_pushed);
3607 schedstat_inc(sd, alb_failed);
3609 spin_unlock(&target_rq->lock);
3614 atomic_t load_balancer;
3616 } nohz ____cacheline_aligned = {
3617 .load_balancer = ATOMIC_INIT(-1),
3618 .cpu_mask = CPU_MASK_NONE,
3622 * This routine will try to nominate the ilb (idle load balancing)
3623 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3624 * load balancing on behalf of all those cpus. If all the cpus in the system
3625 * go into this tickless mode, then there will be no ilb owner (as there is
3626 * no need for one) and all the cpus will sleep till the next wakeup event
3629 * For the ilb owner, tick is not stopped. And this tick will be used
3630 * for idle load balancing. ilb owner will still be part of
3633 * While stopping the tick, this cpu will become the ilb owner if there
3634 * is no other owner. And will be the owner till that cpu becomes busy
3635 * or if all cpus in the system stop their ticks at which point
3636 * there is no need for ilb owner.
3638 * When the ilb owner becomes busy, it nominates another owner, during the
3639 * next busy scheduler_tick()
3641 int select_nohz_load_balancer(int stop_tick)
3643 int cpu = smp_processor_id();
3646 cpu_set(cpu, nohz.cpu_mask);
3647 cpu_rq(cpu)->in_nohz_recently = 1;
3650 * If we are going offline and still the leader, give up!
3652 if (cpu_is_offline(cpu) &&
3653 atomic_read(&nohz.load_balancer) == cpu) {
3654 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3659 /* time for ilb owner also to sleep */
3660 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3661 if (atomic_read(&nohz.load_balancer) == cpu)
3662 atomic_set(&nohz.load_balancer, -1);
3666 if (atomic_read(&nohz.load_balancer) == -1) {
3667 /* make me the ilb owner */
3668 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3670 } else if (atomic_read(&nohz.load_balancer) == cpu)
3673 if (!cpu_isset(cpu, nohz.cpu_mask))
3676 cpu_clear(cpu, nohz.cpu_mask);
3678 if (atomic_read(&nohz.load_balancer) == cpu)
3679 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3686 static DEFINE_SPINLOCK(balancing);
3689 * It checks each scheduling domain to see if it is due to be balanced,
3690 * and initiates a balancing operation if so.
3692 * Balancing parameters are set up in arch_init_sched_domains.
3694 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3697 struct rq *rq = cpu_rq(cpu);
3698 unsigned long interval;
3699 struct sched_domain *sd;
3700 /* Earliest time when we have to do rebalance again */
3701 unsigned long next_balance = jiffies + 60*HZ;
3702 int update_next_balance = 0;
3705 for_each_domain(cpu, sd) {
3706 if (!(sd->flags & SD_LOAD_BALANCE))
3709 interval = sd->balance_interval;
3710 if (idle != CPU_IDLE)
3711 interval *= sd->busy_factor;
3713 /* scale ms to jiffies */
3714 interval = msecs_to_jiffies(interval);
3715 if (unlikely(!interval))
3717 if (interval > HZ*NR_CPUS/10)
3718 interval = HZ*NR_CPUS/10;
3721 if (sd->flags & SD_SERIALIZE) {
3722 if (!spin_trylock(&balancing))
3726 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3727 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3729 * We've pulled tasks over so either we're no
3730 * longer idle, or one of our SMT siblings is
3733 idle = CPU_NOT_IDLE;
3735 sd->last_balance = jiffies;
3737 if (sd->flags & SD_SERIALIZE)
3738 spin_unlock(&balancing);
3740 if (time_after(next_balance, sd->last_balance + interval)) {
3741 next_balance = sd->last_balance + interval;
3742 update_next_balance = 1;
3746 * Stop the load balance at this level. There is another
3747 * CPU in our sched group which is doing load balancing more
3755 * next_balance will be updated only when there is a need.
3756 * When the cpu is attached to null domain for ex, it will not be
3759 if (likely(update_next_balance))
3760 rq->next_balance = next_balance;
3764 * run_rebalance_domains is triggered when needed from the scheduler tick.
3765 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3766 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3768 static void run_rebalance_domains(struct softirq_action *h)
3770 int this_cpu = smp_processor_id();
3771 struct rq *this_rq = cpu_rq(this_cpu);
3772 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3773 CPU_IDLE : CPU_NOT_IDLE;
3775 rebalance_domains(this_cpu, idle);
3779 * If this cpu is the owner for idle load balancing, then do the
3780 * balancing on behalf of the other idle cpus whose ticks are
3783 if (this_rq->idle_at_tick &&
3784 atomic_read(&nohz.load_balancer) == this_cpu) {
3785 cpumask_t cpus = nohz.cpu_mask;
3789 cpu_clear(this_cpu, cpus);
3790 for_each_cpu_mask(balance_cpu, cpus) {
3792 * If this cpu gets work to do, stop the load balancing
3793 * work being done for other cpus. Next load
3794 * balancing owner will pick it up.
3799 rebalance_domains(balance_cpu, CPU_IDLE);
3801 rq = cpu_rq(balance_cpu);
3802 if (time_after(this_rq->next_balance, rq->next_balance))
3803 this_rq->next_balance = rq->next_balance;
3810 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3812 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3813 * idle load balancing owner or decide to stop the periodic load balancing,
3814 * if the whole system is idle.
3816 static inline void trigger_load_balance(struct rq *rq, int cpu)
3820 * If we were in the nohz mode recently and busy at the current
3821 * scheduler tick, then check if we need to nominate new idle
3824 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3825 rq->in_nohz_recently = 0;
3827 if (atomic_read(&nohz.load_balancer) == cpu) {
3828 cpu_clear(cpu, nohz.cpu_mask);
3829 atomic_set(&nohz.load_balancer, -1);
3832 if (atomic_read(&nohz.load_balancer) == -1) {
3834 * simple selection for now: Nominate the
3835 * first cpu in the nohz list to be the next
3838 * TBD: Traverse the sched domains and nominate
3839 * the nearest cpu in the nohz.cpu_mask.
3841 int ilb = first_cpu(nohz.cpu_mask);
3843 if (ilb < nr_cpu_ids)
3849 * If this cpu is idle and doing idle load balancing for all the
3850 * cpus with ticks stopped, is it time for that to stop?
3852 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3853 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3859 * If this cpu is idle and the idle load balancing is done by
3860 * someone else, then no need raise the SCHED_SOFTIRQ
3862 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3863 cpu_isset(cpu, nohz.cpu_mask))
3866 if (time_after_eq(jiffies, rq->next_balance))
3867 raise_softirq(SCHED_SOFTIRQ);
3870 #else /* CONFIG_SMP */
3873 * on UP we do not need to balance between CPUs:
3875 static inline void idle_balance(int cpu, struct rq *rq)
3881 DEFINE_PER_CPU(struct kernel_stat, kstat);
3883 EXPORT_PER_CPU_SYMBOL(kstat);
3886 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3887 * that have not yet been banked in case the task is currently running.
3889 unsigned long long task_sched_runtime(struct task_struct *p)
3891 unsigned long flags;
3895 rq = task_rq_lock(p, &flags);
3896 ns = p->se.sum_exec_runtime;
3897 if (task_current(rq, p)) {
3898 update_rq_clock(rq);
3899 delta_exec = rq->clock - p->se.exec_start;
3900 if ((s64)delta_exec > 0)
3903 task_rq_unlock(rq, &flags);
3909 * Account user cpu time to a process.
3910 * @p: the process that the cpu time gets accounted to
3911 * @cputime: the cpu time spent in user space since the last update
3913 void account_user_time(struct task_struct *p, cputime_t cputime)
3915 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3918 p->utime = cputime_add(p->utime, cputime);
3920 /* Add user time to cpustat. */
3921 tmp = cputime_to_cputime64(cputime);
3922 if (TASK_NICE(p) > 0)
3923 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3925 cpustat->user = cputime64_add(cpustat->user, tmp);
3929 * Account guest cpu time to a process.
3930 * @p: the process that the cpu time gets accounted to
3931 * @cputime: the cpu time spent in virtual machine since the last update
3933 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3936 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3938 tmp = cputime_to_cputime64(cputime);
3940 p->utime = cputime_add(p->utime, cputime);
3941 p->gtime = cputime_add(p->gtime, cputime);
3943 cpustat->user = cputime64_add(cpustat->user, tmp);
3944 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3948 * Account scaled user cpu time to a process.
3949 * @p: the process that the cpu time gets accounted to
3950 * @cputime: the cpu time spent in user space since the last update
3952 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3954 p->utimescaled = cputime_add(p->utimescaled, cputime);
3958 * Account system cpu time to a process.
3959 * @p: the process that the cpu time gets accounted to
3960 * @hardirq_offset: the offset to subtract from hardirq_count()
3961 * @cputime: the cpu time spent in kernel space since the last update
3963 void account_system_time(struct task_struct *p, int hardirq_offset,
3966 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3967 struct rq *rq = this_rq();
3970 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3971 account_guest_time(p, cputime);
3975 p->stime = cputime_add(p->stime, cputime);
3977 /* Add system time to cpustat. */
3978 tmp = cputime_to_cputime64(cputime);
3979 if (hardirq_count() - hardirq_offset)
3980 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3981 else if (softirq_count())
3982 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3983 else if (p != rq->idle)
3984 cpustat->system = cputime64_add(cpustat->system, tmp);
3985 else if (atomic_read(&rq->nr_iowait) > 0)
3986 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3988 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3989 /* Account for system time used */
3990 acct_update_integrals(p);
3994 * Account scaled system cpu time to a process.
3995 * @p: the process that the cpu time gets accounted to
3996 * @hardirq_offset: the offset to subtract from hardirq_count()
3997 * @cputime: the cpu time spent in kernel space since the last update
3999 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4001 p->stimescaled = cputime_add(p->stimescaled, cputime);
4005 * Account for involuntary wait time.
4006 * @p: the process from which the cpu time has been stolen
4007 * @steal: the cpu time spent in involuntary wait
4009 void account_steal_time(struct task_struct *p, cputime_t steal)
4011 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4012 cputime64_t tmp = cputime_to_cputime64(steal);
4013 struct rq *rq = this_rq();
4015 if (p == rq->idle) {
4016 p->stime = cputime_add(p->stime, steal);
4017 if (atomic_read(&rq->nr_iowait) > 0)
4018 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4020 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4022 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4026 * This function gets called by the timer code, with HZ frequency.
4027 * We call it with interrupts disabled.
4029 * It also gets called by the fork code, when changing the parent's
4032 void scheduler_tick(void)
4034 int cpu = smp_processor_id();
4035 struct rq *rq = cpu_rq(cpu);
4036 struct task_struct *curr = rq->curr;
4040 spin_lock(&rq->lock);
4041 update_rq_clock(rq);
4042 update_cpu_load(rq);
4043 curr->sched_class->task_tick(rq, curr, 0);
4044 spin_unlock(&rq->lock);
4047 rq->idle_at_tick = idle_cpu(cpu);
4048 trigger_load_balance(rq, cpu);
4052 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4053 defined(CONFIG_PREEMPT_TRACER))
4055 static inline unsigned long get_parent_ip(unsigned long addr)
4057 if (in_lock_functions(addr)) {
4058 addr = CALLER_ADDR2;
4059 if (in_lock_functions(addr))
4060 addr = CALLER_ADDR3;
4065 void __kprobes add_preempt_count(int val)
4067 #ifdef CONFIG_DEBUG_PREEMPT
4071 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4074 preempt_count() += val;
4075 #ifdef CONFIG_DEBUG_PREEMPT
4077 * Spinlock count overflowing soon?
4079 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4082 if (preempt_count() == val)
4083 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4085 EXPORT_SYMBOL(add_preempt_count);
4087 void __kprobes sub_preempt_count(int val)
4089 #ifdef CONFIG_DEBUG_PREEMPT
4093 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4096 * Is the spinlock portion underflowing?
4098 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4099 !(preempt_count() & PREEMPT_MASK)))
4103 if (preempt_count() == val)
4104 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4105 preempt_count() -= val;
4107 EXPORT_SYMBOL(sub_preempt_count);
4112 * Print scheduling while atomic bug:
4114 static noinline void __schedule_bug(struct task_struct *prev)
4116 struct pt_regs *regs = get_irq_regs();
4118 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4119 prev->comm, prev->pid, preempt_count());
4121 debug_show_held_locks(prev);
4122 if (irqs_disabled())
4123 print_irqtrace_events(prev);
4132 * Various schedule()-time debugging checks and statistics:
4134 static inline void schedule_debug(struct task_struct *prev)
4137 * Test if we are atomic. Since do_exit() needs to call into
4138 * schedule() atomically, we ignore that path for now.
4139 * Otherwise, whine if we are scheduling when we should not be.
4141 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4142 __schedule_bug(prev);
4144 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4146 schedstat_inc(this_rq(), sched_count);
4147 #ifdef CONFIG_SCHEDSTATS
4148 if (unlikely(prev->lock_depth >= 0)) {
4149 schedstat_inc(this_rq(), bkl_count);
4150 schedstat_inc(prev, sched_info.bkl_count);
4156 * Pick up the highest-prio task:
4158 static inline struct task_struct *
4159 pick_next_task(struct rq *rq, struct task_struct *prev)
4161 const struct sched_class *class;
4162 struct task_struct *p;
4165 * Optimization: we know that if all tasks are in
4166 * the fair class we can call that function directly:
4168 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4169 p = fair_sched_class.pick_next_task(rq);
4174 class = sched_class_highest;
4176 p = class->pick_next_task(rq);
4180 * Will never be NULL as the idle class always
4181 * returns a non-NULL p:
4183 class = class->next;
4188 * schedule() is the main scheduler function.
4190 asmlinkage void __sched schedule(void)
4192 struct task_struct *prev, *next;
4193 unsigned long *switch_count;
4199 cpu = smp_processor_id();
4203 switch_count = &prev->nivcsw;
4205 release_kernel_lock(prev);
4206 need_resched_nonpreemptible:
4208 schedule_debug(prev);
4213 * Do the rq-clock update outside the rq lock:
4215 local_irq_disable();
4216 update_rq_clock(rq);
4217 spin_lock(&rq->lock);
4218 clear_tsk_need_resched(prev);
4220 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4221 if (unlikely(signal_pending_state(prev->state, prev)))
4222 prev->state = TASK_RUNNING;
4224 deactivate_task(rq, prev, 1);
4225 switch_count = &prev->nvcsw;
4229 if (prev->sched_class->pre_schedule)
4230 prev->sched_class->pre_schedule(rq, prev);
4233 if (unlikely(!rq->nr_running))
4234 idle_balance(cpu, rq);
4236 prev->sched_class->put_prev_task(rq, prev);
4237 next = pick_next_task(rq, prev);
4239 if (likely(prev != next)) {
4240 sched_info_switch(prev, next);
4246 context_switch(rq, prev, next); /* unlocks the rq */
4248 * the context switch might have flipped the stack from under
4249 * us, hence refresh the local variables.
4251 cpu = smp_processor_id();
4254 spin_unlock_irq(&rq->lock);
4258 if (unlikely(reacquire_kernel_lock(current) < 0))
4259 goto need_resched_nonpreemptible;
4261 preempt_enable_no_resched();
4262 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4265 EXPORT_SYMBOL(schedule);
4267 #ifdef CONFIG_PREEMPT
4269 * this is the entry point to schedule() from in-kernel preemption
4270 * off of preempt_enable. Kernel preemptions off return from interrupt
4271 * occur there and call schedule directly.
4273 asmlinkage void __sched preempt_schedule(void)
4275 struct thread_info *ti = current_thread_info();
4278 * If there is a non-zero preempt_count or interrupts are disabled,
4279 * we do not want to preempt the current task. Just return..
4281 if (likely(ti->preempt_count || irqs_disabled()))
4285 add_preempt_count(PREEMPT_ACTIVE);
4287 sub_preempt_count(PREEMPT_ACTIVE);
4290 * Check again in case we missed a preemption opportunity
4291 * between schedule and now.
4294 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4296 EXPORT_SYMBOL(preempt_schedule);
4299 * this is the entry point to schedule() from kernel preemption
4300 * off of irq context.
4301 * Note, that this is called and return with irqs disabled. This will
4302 * protect us against recursive calling from irq.
4304 asmlinkage void __sched preempt_schedule_irq(void)
4306 struct thread_info *ti = current_thread_info();
4308 /* Catch callers which need to be fixed */
4309 BUG_ON(ti->preempt_count || !irqs_disabled());
4312 add_preempt_count(PREEMPT_ACTIVE);
4315 local_irq_disable();
4316 sub_preempt_count(PREEMPT_ACTIVE);
4319 * Check again in case we missed a preemption opportunity
4320 * between schedule and now.
4323 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4326 #endif /* CONFIG_PREEMPT */
4328 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4331 return try_to_wake_up(curr->private, mode, sync);
4333 EXPORT_SYMBOL(default_wake_function);
4336 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4337 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4338 * number) then we wake all the non-exclusive tasks and one exclusive task.
4340 * There are circumstances in which we can try to wake a task which has already
4341 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4342 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4344 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4345 int nr_exclusive, int sync, void *key)
4347 wait_queue_t *curr, *next;
4349 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4350 unsigned flags = curr->flags;
4352 if (curr->func(curr, mode, sync, key) &&
4353 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4359 * __wake_up - wake up threads blocked on a waitqueue.
4361 * @mode: which threads
4362 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4363 * @key: is directly passed to the wakeup function
4365 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4366 int nr_exclusive, void *key)
4368 unsigned long flags;
4370 spin_lock_irqsave(&q->lock, flags);
4371 __wake_up_common(q, mode, nr_exclusive, 0, key);
4372 spin_unlock_irqrestore(&q->lock, flags);
4374 EXPORT_SYMBOL(__wake_up);
4377 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4379 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4381 __wake_up_common(q, mode, 1, 0, NULL);
4385 * __wake_up_sync - wake up threads blocked on a waitqueue.
4387 * @mode: which threads
4388 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4390 * The sync wakeup differs that the waker knows that it will schedule
4391 * away soon, so while the target thread will be woken up, it will not
4392 * be migrated to another CPU - ie. the two threads are 'synchronized'
4393 * with each other. This can prevent needless bouncing between CPUs.
4395 * On UP it can prevent extra preemption.
4398 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4400 unsigned long flags;
4406 if (unlikely(!nr_exclusive))
4409 spin_lock_irqsave(&q->lock, flags);
4410 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4411 spin_unlock_irqrestore(&q->lock, flags);
4413 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4415 void complete(struct completion *x)
4417 unsigned long flags;
4419 spin_lock_irqsave(&x->wait.lock, flags);
4421 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4422 spin_unlock_irqrestore(&x->wait.lock, flags);
4424 EXPORT_SYMBOL(complete);
4426 void complete_all(struct completion *x)
4428 unsigned long flags;
4430 spin_lock_irqsave(&x->wait.lock, flags);
4431 x->done += UINT_MAX/2;
4432 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4433 spin_unlock_irqrestore(&x->wait.lock, flags);
4435 EXPORT_SYMBOL(complete_all);
4437 static inline long __sched
4438 do_wait_for_common(struct completion *x, long timeout, int state)
4441 DECLARE_WAITQUEUE(wait, current);
4443 wait.flags |= WQ_FLAG_EXCLUSIVE;
4444 __add_wait_queue_tail(&x->wait, &wait);
4446 if ((state == TASK_INTERRUPTIBLE &&
4447 signal_pending(current)) ||
4448 (state == TASK_KILLABLE &&
4449 fatal_signal_pending(current))) {
4450 __remove_wait_queue(&x->wait, &wait);
4451 return -ERESTARTSYS;
4453 __set_current_state(state);
4454 spin_unlock_irq(&x->wait.lock);
4455 timeout = schedule_timeout(timeout);
4456 spin_lock_irq(&x->wait.lock);
4458 __remove_wait_queue(&x->wait, &wait);
4462 __remove_wait_queue(&x->wait, &wait);
4469 wait_for_common(struct completion *x, long timeout, int state)
4473 spin_lock_irq(&x->wait.lock);
4474 timeout = do_wait_for_common(x, timeout, state);
4475 spin_unlock_irq(&x->wait.lock);
4479 void __sched wait_for_completion(struct completion *x)
4481 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4483 EXPORT_SYMBOL(wait_for_completion);
4485 unsigned long __sched
4486 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4488 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4490 EXPORT_SYMBOL(wait_for_completion_timeout);
4492 int __sched wait_for_completion_interruptible(struct completion *x)
4494 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4495 if (t == -ERESTARTSYS)
4499 EXPORT_SYMBOL(wait_for_completion_interruptible);
4501 unsigned long __sched
4502 wait_for_completion_interruptible_timeout(struct completion *x,
4503 unsigned long timeout)
4505 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4507 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4509 int __sched wait_for_completion_killable(struct completion *x)
4511 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4512 if (t == -ERESTARTSYS)
4516 EXPORT_SYMBOL(wait_for_completion_killable);
4519 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4521 unsigned long flags;
4524 init_waitqueue_entry(&wait, current);
4526 __set_current_state(state);
4528 spin_lock_irqsave(&q->lock, flags);
4529 __add_wait_queue(q, &wait);
4530 spin_unlock(&q->lock);
4531 timeout = schedule_timeout(timeout);
4532 spin_lock_irq(&q->lock);
4533 __remove_wait_queue(q, &wait);
4534 spin_unlock_irqrestore(&q->lock, flags);
4539 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4541 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4543 EXPORT_SYMBOL(interruptible_sleep_on);
4546 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4548 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4550 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4552 void __sched sleep_on(wait_queue_head_t *q)
4554 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4556 EXPORT_SYMBOL(sleep_on);
4558 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4560 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4562 EXPORT_SYMBOL(sleep_on_timeout);
4564 #ifdef CONFIG_RT_MUTEXES
4567 * rt_mutex_setprio - set the current priority of a task
4569 * @prio: prio value (kernel-internal form)
4571 * This function changes the 'effective' priority of a task. It does
4572 * not touch ->normal_prio like __setscheduler().
4574 * Used by the rt_mutex code to implement priority inheritance logic.
4576 void rt_mutex_setprio(struct task_struct *p, int prio)
4578 unsigned long flags;
4579 int oldprio, on_rq, running;
4581 const struct sched_class *prev_class = p->sched_class;
4583 BUG_ON(prio < 0 || prio > MAX_PRIO);
4585 rq = task_rq_lock(p, &flags);
4586 update_rq_clock(rq);
4589 on_rq = p->se.on_rq;
4590 running = task_current(rq, p);
4592 dequeue_task(rq, p, 0);
4594 p->sched_class->put_prev_task(rq, p);
4597 p->sched_class = &rt_sched_class;
4599 p->sched_class = &fair_sched_class;
4604 p->sched_class->set_curr_task(rq);
4606 enqueue_task(rq, p, 0);
4608 check_class_changed(rq, p, prev_class, oldprio, running);
4610 task_rq_unlock(rq, &flags);
4615 void set_user_nice(struct task_struct *p, long nice)
4617 int old_prio, delta, on_rq;
4618 unsigned long flags;
4621 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4624 * We have to be careful, if called from sys_setpriority(),
4625 * the task might be in the middle of scheduling on another CPU.
4627 rq = task_rq_lock(p, &flags);
4628 update_rq_clock(rq);
4630 * The RT priorities are set via sched_setscheduler(), but we still
4631 * allow the 'normal' nice value to be set - but as expected
4632 * it wont have any effect on scheduling until the task is
4633 * SCHED_FIFO/SCHED_RR:
4635 if (task_has_rt_policy(p)) {
4636 p->static_prio = NICE_TO_PRIO(nice);
4639 on_rq = p->se.on_rq;
4641 dequeue_task(rq, p, 0);
4645 p->static_prio = NICE_TO_PRIO(nice);
4648 p->prio = effective_prio(p);
4649 delta = p->prio - old_prio;
4652 enqueue_task(rq, p, 0);
4655 * If the task increased its priority or is running and
4656 * lowered its priority, then reschedule its CPU:
4658 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4659 resched_task(rq->curr);
4662 task_rq_unlock(rq, &flags);
4664 EXPORT_SYMBOL(set_user_nice);
4667 * can_nice - check if a task can reduce its nice value
4671 int can_nice(const struct task_struct *p, const int nice)
4673 /* convert nice value [19,-20] to rlimit style value [1,40] */
4674 int nice_rlim = 20 - nice;
4676 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4677 capable(CAP_SYS_NICE));
4680 #ifdef __ARCH_WANT_SYS_NICE
4683 * sys_nice - change the priority of the current process.
4684 * @increment: priority increment
4686 * sys_setpriority is a more generic, but much slower function that
4687 * does similar things.
4689 asmlinkage long sys_nice(int increment)
4694 * Setpriority might change our priority at the same moment.
4695 * We don't have to worry. Conceptually one call occurs first
4696 * and we have a single winner.
4698 if (increment < -40)
4703 nice = PRIO_TO_NICE(current->static_prio) + increment;
4709 if (increment < 0 && !can_nice(current, nice))
4712 retval = security_task_setnice(current, nice);
4716 set_user_nice(current, nice);
4723 * task_prio - return the priority value of a given task.
4724 * @p: the task in question.
4726 * This is the priority value as seen by users in /proc.
4727 * RT tasks are offset by -200. Normal tasks are centered
4728 * around 0, value goes from -16 to +15.
4730 int task_prio(const struct task_struct *p)
4732 return p->prio - MAX_RT_PRIO;
4736 * task_nice - return the nice value of a given task.
4737 * @p: the task in question.
4739 int task_nice(const struct task_struct *p)
4741 return TASK_NICE(p);
4743 EXPORT_SYMBOL(task_nice);
4746 * idle_cpu - is a given cpu idle currently?
4747 * @cpu: the processor in question.
4749 int idle_cpu(int cpu)
4751 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4755 * idle_task - return the idle task for a given cpu.
4756 * @cpu: the processor in question.
4758 struct task_struct *idle_task(int cpu)
4760 return cpu_rq(cpu)->idle;
4764 * find_process_by_pid - find a process with a matching PID value.
4765 * @pid: the pid in question.
4767 static struct task_struct *find_process_by_pid(pid_t pid)
4769 return pid ? find_task_by_vpid(pid) : current;
4772 /* Actually do priority change: must hold rq lock. */
4774 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4776 BUG_ON(p->se.on_rq);
4779 switch (p->policy) {
4783 p->sched_class = &fair_sched_class;
4787 p->sched_class = &rt_sched_class;
4791 p->rt_priority = prio;
4792 p->normal_prio = normal_prio(p);
4793 /* we are holding p->pi_lock already */
4794 p->prio = rt_mutex_getprio(p);
4799 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4800 * @p: the task in question.
4801 * @policy: new policy.
4802 * @param: structure containing the new RT priority.
4804 * NOTE that the task may be already dead.
4806 int sched_setscheduler(struct task_struct *p, int policy,
4807 struct sched_param *param)
4809 int retval, oldprio, oldpolicy = -1, on_rq, running;
4810 unsigned long flags;
4811 const struct sched_class *prev_class = p->sched_class;
4814 /* may grab non-irq protected spin_locks */
4815 BUG_ON(in_interrupt());
4817 /* double check policy once rq lock held */
4819 policy = oldpolicy = p->policy;
4820 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4821 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4822 policy != SCHED_IDLE)
4825 * Valid priorities for SCHED_FIFO and SCHED_RR are
4826 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4827 * SCHED_BATCH and SCHED_IDLE is 0.
4829 if (param->sched_priority < 0 ||
4830 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4831 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4833 if (rt_policy(policy) != (param->sched_priority != 0))
4837 * Allow unprivileged RT tasks to decrease priority:
4839 if (!capable(CAP_SYS_NICE)) {
4840 if (rt_policy(policy)) {
4841 unsigned long rlim_rtprio;
4843 if (!lock_task_sighand(p, &flags))
4845 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4846 unlock_task_sighand(p, &flags);
4848 /* can't set/change the rt policy */
4849 if (policy != p->policy && !rlim_rtprio)
4852 /* can't increase priority */
4853 if (param->sched_priority > p->rt_priority &&
4854 param->sched_priority > rlim_rtprio)
4858 * Like positive nice levels, dont allow tasks to
4859 * move out of SCHED_IDLE either:
4861 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4864 /* can't change other user's priorities */
4865 if ((current->euid != p->euid) &&
4866 (current->euid != p->uid))
4870 #ifdef CONFIG_RT_GROUP_SCHED
4872 * Do not allow realtime tasks into groups that have no runtime
4875 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
4879 retval = security_task_setscheduler(p, policy, param);
4883 * make sure no PI-waiters arrive (or leave) while we are
4884 * changing the priority of the task:
4886 spin_lock_irqsave(&p->pi_lock, flags);
4888 * To be able to change p->policy safely, the apropriate
4889 * runqueue lock must be held.
4891 rq = __task_rq_lock(p);
4892 /* recheck policy now with rq lock held */
4893 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4894 policy = oldpolicy = -1;
4895 __task_rq_unlock(rq);
4896 spin_unlock_irqrestore(&p->pi_lock, flags);
4899 update_rq_clock(rq);
4900 on_rq = p->se.on_rq;
4901 running = task_current(rq, p);
4903 deactivate_task(rq, p, 0);
4905 p->sched_class->put_prev_task(rq, p);
4908 __setscheduler(rq, p, policy, param->sched_priority);
4911 p->sched_class->set_curr_task(rq);
4913 activate_task(rq, p, 0);
4915 check_class_changed(rq, p, prev_class, oldprio, running);
4917 __task_rq_unlock(rq);
4918 spin_unlock_irqrestore(&p->pi_lock, flags);
4920 rt_mutex_adjust_pi(p);
4924 EXPORT_SYMBOL_GPL(sched_setscheduler);
4927 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4929 struct sched_param lparam;
4930 struct task_struct *p;
4933 if (!param || pid < 0)
4935 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4940 p = find_process_by_pid(pid);
4942 retval = sched_setscheduler(p, policy, &lparam);
4949 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4950 * @pid: the pid in question.
4951 * @policy: new policy.
4952 * @param: structure containing the new RT priority.
4955 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4957 /* negative values for policy are not valid */
4961 return do_sched_setscheduler(pid, policy, param);
4965 * sys_sched_setparam - set/change the RT priority of a thread
4966 * @pid: the pid in question.
4967 * @param: structure containing the new RT priority.
4969 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4971 return do_sched_setscheduler(pid, -1, param);
4975 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4976 * @pid: the pid in question.
4978 asmlinkage long sys_sched_getscheduler(pid_t pid)
4980 struct task_struct *p;
4987 read_lock(&tasklist_lock);
4988 p = find_process_by_pid(pid);
4990 retval = security_task_getscheduler(p);
4994 read_unlock(&tasklist_lock);
4999 * sys_sched_getscheduler - get the RT priority of a thread
5000 * @pid: the pid in question.
5001 * @param: structure containing the RT priority.
5003 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5005 struct sched_param lp;
5006 struct task_struct *p;
5009 if (!param || pid < 0)
5012 read_lock(&tasklist_lock);
5013 p = find_process_by_pid(pid);
5018 retval = security_task_getscheduler(p);
5022 lp.sched_priority = p->rt_priority;
5023 read_unlock(&tasklist_lock);
5026 * This one might sleep, we cannot do it with a spinlock held ...
5028 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5033 read_unlock(&tasklist_lock);
5037 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5039 cpumask_t cpus_allowed;
5040 cpumask_t new_mask = *in_mask;
5041 struct task_struct *p;
5045 read_lock(&tasklist_lock);
5047 p = find_process_by_pid(pid);
5049 read_unlock(&tasklist_lock);
5055 * It is not safe to call set_cpus_allowed with the
5056 * tasklist_lock held. We will bump the task_struct's
5057 * usage count and then drop tasklist_lock.
5060 read_unlock(&tasklist_lock);
5063 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5064 !capable(CAP_SYS_NICE))
5067 retval = security_task_setscheduler(p, 0, NULL);
5071 cpuset_cpus_allowed(p, &cpus_allowed);
5072 cpus_and(new_mask, new_mask, cpus_allowed);
5074 retval = set_cpus_allowed_ptr(p, &new_mask);
5077 cpuset_cpus_allowed(p, &cpus_allowed);
5078 if (!cpus_subset(new_mask, cpus_allowed)) {
5080 * We must have raced with a concurrent cpuset
5081 * update. Just reset the cpus_allowed to the
5082 * cpuset's cpus_allowed
5084 new_mask = cpus_allowed;
5094 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5095 cpumask_t *new_mask)
5097 if (len < sizeof(cpumask_t)) {
5098 memset(new_mask, 0, sizeof(cpumask_t));
5099 } else if (len > sizeof(cpumask_t)) {
5100 len = sizeof(cpumask_t);
5102 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5106 * sys_sched_setaffinity - set the cpu affinity of a process
5107 * @pid: pid of the process
5108 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5109 * @user_mask_ptr: user-space pointer to the new cpu mask
5111 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5112 unsigned long __user *user_mask_ptr)
5117 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5121 return sched_setaffinity(pid, &new_mask);
5125 * Represents all cpu's present in the system
5126 * In systems capable of hotplug, this map could dynamically grow
5127 * as new cpu's are detected in the system via any platform specific
5128 * method, such as ACPI for e.g.
5131 cpumask_t cpu_present_map __read_mostly;
5132 EXPORT_SYMBOL(cpu_present_map);
5135 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
5136 EXPORT_SYMBOL(cpu_online_map);
5138 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
5139 EXPORT_SYMBOL(cpu_possible_map);
5142 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5144 struct task_struct *p;
5148 read_lock(&tasklist_lock);
5151 p = find_process_by_pid(pid);
5155 retval = security_task_getscheduler(p);
5159 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5162 read_unlock(&tasklist_lock);
5169 * sys_sched_getaffinity - get the cpu affinity of a process
5170 * @pid: pid of the process
5171 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5172 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5174 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5175 unsigned long __user *user_mask_ptr)
5180 if (len < sizeof(cpumask_t))
5183 ret = sched_getaffinity(pid, &mask);
5187 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5190 return sizeof(cpumask_t);
5194 * sys_sched_yield - yield the current processor to other threads.
5196 * This function yields the current CPU to other tasks. If there are no
5197 * other threads running on this CPU then this function will return.
5199 asmlinkage long sys_sched_yield(void)
5201 struct rq *rq = this_rq_lock();
5203 schedstat_inc(rq, yld_count);
5204 current->sched_class->yield_task(rq);
5207 * Since we are going to call schedule() anyway, there's
5208 * no need to preempt or enable interrupts:
5210 __release(rq->lock);
5211 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5212 _raw_spin_unlock(&rq->lock);
5213 preempt_enable_no_resched();
5220 static void __cond_resched(void)
5222 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5223 __might_sleep(__FILE__, __LINE__);
5226 * The BKS might be reacquired before we have dropped
5227 * PREEMPT_ACTIVE, which could trigger a second
5228 * cond_resched() call.
5231 add_preempt_count(PREEMPT_ACTIVE);
5233 sub_preempt_count(PREEMPT_ACTIVE);
5234 } while (need_resched());
5237 int __sched _cond_resched(void)
5239 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5240 system_state == SYSTEM_RUNNING) {
5246 EXPORT_SYMBOL(_cond_resched);
5249 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5250 * call schedule, and on return reacquire the lock.
5252 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5253 * operations here to prevent schedule() from being called twice (once via
5254 * spin_unlock(), once by hand).
5256 int cond_resched_lock(spinlock_t *lock)
5258 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5261 if (spin_needbreak(lock) || resched) {
5263 if (resched && need_resched())
5272 EXPORT_SYMBOL(cond_resched_lock);
5274 int __sched cond_resched_softirq(void)
5276 BUG_ON(!in_softirq());
5278 if (need_resched() && system_state == SYSTEM_RUNNING) {
5286 EXPORT_SYMBOL(cond_resched_softirq);
5289 * yield - yield the current processor to other threads.
5291 * This is a shortcut for kernel-space yielding - it marks the
5292 * thread runnable and calls sys_sched_yield().
5294 void __sched yield(void)
5296 set_current_state(TASK_RUNNING);
5299 EXPORT_SYMBOL(yield);
5302 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5303 * that process accounting knows that this is a task in IO wait state.
5305 * But don't do that if it is a deliberate, throttling IO wait (this task
5306 * has set its backing_dev_info: the queue against which it should throttle)
5308 void __sched io_schedule(void)
5310 struct rq *rq = &__raw_get_cpu_var(runqueues);
5312 delayacct_blkio_start();
5313 atomic_inc(&rq->nr_iowait);
5315 atomic_dec(&rq->nr_iowait);
5316 delayacct_blkio_end();
5318 EXPORT_SYMBOL(io_schedule);
5320 long __sched io_schedule_timeout(long timeout)
5322 struct rq *rq = &__raw_get_cpu_var(runqueues);
5325 delayacct_blkio_start();
5326 atomic_inc(&rq->nr_iowait);
5327 ret = schedule_timeout(timeout);
5328 atomic_dec(&rq->nr_iowait);
5329 delayacct_blkio_end();
5334 * sys_sched_get_priority_max - return maximum RT priority.
5335 * @policy: scheduling class.
5337 * this syscall returns the maximum rt_priority that can be used
5338 * by a given scheduling class.
5340 asmlinkage long sys_sched_get_priority_max(int policy)
5347 ret = MAX_USER_RT_PRIO-1;
5359 * sys_sched_get_priority_min - return minimum RT priority.
5360 * @policy: scheduling class.
5362 * this syscall returns the minimum rt_priority that can be used
5363 * by a given scheduling class.
5365 asmlinkage long sys_sched_get_priority_min(int policy)
5383 * sys_sched_rr_get_interval - return the default timeslice of a process.
5384 * @pid: pid of the process.
5385 * @interval: userspace pointer to the timeslice value.
5387 * this syscall writes the default timeslice value of a given process
5388 * into the user-space timespec buffer. A value of '0' means infinity.
5391 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5393 struct task_struct *p;
5394 unsigned int time_slice;
5402 read_lock(&tasklist_lock);
5403 p = find_process_by_pid(pid);
5407 retval = security_task_getscheduler(p);
5412 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5413 * tasks that are on an otherwise idle runqueue:
5416 if (p->policy == SCHED_RR) {
5417 time_slice = DEF_TIMESLICE;
5418 } else if (p->policy != SCHED_FIFO) {
5419 struct sched_entity *se = &p->se;
5420 unsigned long flags;
5423 rq = task_rq_lock(p, &flags);
5424 if (rq->cfs.load.weight)
5425 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5426 task_rq_unlock(rq, &flags);
5428 read_unlock(&tasklist_lock);
5429 jiffies_to_timespec(time_slice, &t);
5430 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5434 read_unlock(&tasklist_lock);
5438 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5440 void sched_show_task(struct task_struct *p)
5442 unsigned long free = 0;
5445 state = p->state ? __ffs(p->state) + 1 : 0;
5446 printk(KERN_INFO "%-13.13s %c", p->comm,
5447 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5448 #if BITS_PER_LONG == 32
5449 if (state == TASK_RUNNING)
5450 printk(KERN_CONT " running ");
5452 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5454 if (state == TASK_RUNNING)
5455 printk(KERN_CONT " running task ");
5457 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5459 #ifdef CONFIG_DEBUG_STACK_USAGE
5461 unsigned long *n = end_of_stack(p);
5464 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5467 printk(KERN_CONT "%5lu %5d %6d\n", free,
5468 task_pid_nr(p), task_pid_nr(p->real_parent));
5470 show_stack(p, NULL);
5473 void show_state_filter(unsigned long state_filter)
5475 struct task_struct *g, *p;
5477 #if BITS_PER_LONG == 32
5479 " task PC stack pid father\n");
5482 " task PC stack pid father\n");
5484 read_lock(&tasklist_lock);
5485 do_each_thread(g, p) {
5487 * reset the NMI-timeout, listing all files on a slow
5488 * console might take alot of time:
5490 touch_nmi_watchdog();
5491 if (!state_filter || (p->state & state_filter))
5493 } while_each_thread(g, p);
5495 touch_all_softlockup_watchdogs();
5497 #ifdef CONFIG_SCHED_DEBUG
5498 sysrq_sched_debug_show();
5500 read_unlock(&tasklist_lock);
5502 * Only show locks if all tasks are dumped:
5504 if (state_filter == -1)
5505 debug_show_all_locks();
5508 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5510 idle->sched_class = &idle_sched_class;
5514 * init_idle - set up an idle thread for a given CPU
5515 * @idle: task in question
5516 * @cpu: cpu the idle task belongs to
5518 * NOTE: this function does not set the idle thread's NEED_RESCHED
5519 * flag, to make booting more robust.
5521 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5523 struct rq *rq = cpu_rq(cpu);
5524 unsigned long flags;
5527 idle->se.exec_start = sched_clock();
5529 idle->prio = idle->normal_prio = MAX_PRIO;
5530 idle->cpus_allowed = cpumask_of_cpu(cpu);
5531 __set_task_cpu(idle, cpu);
5533 spin_lock_irqsave(&rq->lock, flags);
5534 rq->curr = rq->idle = idle;
5535 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5538 spin_unlock_irqrestore(&rq->lock, flags);
5540 /* Set the preempt count _outside_ the spinlocks! */
5541 #if defined(CONFIG_PREEMPT)
5542 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5544 task_thread_info(idle)->preempt_count = 0;
5547 * The idle tasks have their own, simple scheduling class:
5549 idle->sched_class = &idle_sched_class;
5553 * In a system that switches off the HZ timer nohz_cpu_mask
5554 * indicates which cpus entered this state. This is used
5555 * in the rcu update to wait only for active cpus. For system
5556 * which do not switch off the HZ timer nohz_cpu_mask should
5557 * always be CPU_MASK_NONE.
5559 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5562 * Increase the granularity value when there are more CPUs,
5563 * because with more CPUs the 'effective latency' as visible
5564 * to users decreases. But the relationship is not linear,
5565 * so pick a second-best guess by going with the log2 of the
5568 * This idea comes from the SD scheduler of Con Kolivas:
5570 static inline void sched_init_granularity(void)
5572 unsigned int factor = 1 + ilog2(num_online_cpus());
5573 const unsigned long limit = 200000000;
5575 sysctl_sched_min_granularity *= factor;
5576 if (sysctl_sched_min_granularity > limit)
5577 sysctl_sched_min_granularity = limit;
5579 sysctl_sched_latency *= factor;
5580 if (sysctl_sched_latency > limit)
5581 sysctl_sched_latency = limit;
5583 sysctl_sched_wakeup_granularity *= factor;
5588 * This is how migration works:
5590 * 1) we queue a struct migration_req structure in the source CPU's
5591 * runqueue and wake up that CPU's migration thread.
5592 * 2) we down() the locked semaphore => thread blocks.
5593 * 3) migration thread wakes up (implicitly it forces the migrated
5594 * thread off the CPU)
5595 * 4) it gets the migration request and checks whether the migrated
5596 * task is still in the wrong runqueue.
5597 * 5) if it's in the wrong runqueue then the migration thread removes
5598 * it and puts it into the right queue.
5599 * 6) migration thread up()s the semaphore.
5600 * 7) we wake up and the migration is done.
5604 * Change a given task's CPU affinity. Migrate the thread to a
5605 * proper CPU and schedule it away if the CPU it's executing on
5606 * is removed from the allowed bitmask.
5608 * NOTE: the caller must have a valid reference to the task, the
5609 * task must not exit() & deallocate itself prematurely. The
5610 * call is not atomic; no spinlocks may be held.
5612 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5614 struct migration_req req;
5615 unsigned long flags;
5619 rq = task_rq_lock(p, &flags);
5620 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5625 if (p->sched_class->set_cpus_allowed)
5626 p->sched_class->set_cpus_allowed(p, new_mask);
5628 p->cpus_allowed = *new_mask;
5629 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5632 /* Can the task run on the task's current CPU? If so, we're done */
5633 if (cpu_isset(task_cpu(p), *new_mask))
5636 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5637 /* Need help from migration thread: drop lock and wait. */
5638 task_rq_unlock(rq, &flags);
5639 wake_up_process(rq->migration_thread);
5640 wait_for_completion(&req.done);
5641 tlb_migrate_finish(p->mm);
5645 task_rq_unlock(rq, &flags);
5649 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5652 * Move (not current) task off this cpu, onto dest cpu. We're doing
5653 * this because either it can't run here any more (set_cpus_allowed()
5654 * away from this CPU, or CPU going down), or because we're
5655 * attempting to rebalance this task on exec (sched_exec).
5657 * So we race with normal scheduler movements, but that's OK, as long
5658 * as the task is no longer on this CPU.
5660 * Returns non-zero if task was successfully migrated.
5662 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5664 struct rq *rq_dest, *rq_src;
5667 if (unlikely(cpu_is_offline(dest_cpu)))
5670 rq_src = cpu_rq(src_cpu);
5671 rq_dest = cpu_rq(dest_cpu);
5673 double_rq_lock(rq_src, rq_dest);
5674 /* Already moved. */
5675 if (task_cpu(p) != src_cpu)
5677 /* Affinity changed (again). */
5678 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5681 on_rq = p->se.on_rq;
5683 deactivate_task(rq_src, p, 0);
5685 set_task_cpu(p, dest_cpu);
5687 activate_task(rq_dest, p, 0);
5688 check_preempt_curr(rq_dest, p);
5692 double_rq_unlock(rq_src, rq_dest);
5697 * migration_thread - this is a highprio system thread that performs
5698 * thread migration by bumping thread off CPU then 'pushing' onto
5701 static int migration_thread(void *data)
5703 int cpu = (long)data;
5707 BUG_ON(rq->migration_thread != current);
5709 set_current_state(TASK_INTERRUPTIBLE);
5710 while (!kthread_should_stop()) {
5711 struct migration_req *req;
5712 struct list_head *head;
5714 spin_lock_irq(&rq->lock);
5716 if (cpu_is_offline(cpu)) {
5717 spin_unlock_irq(&rq->lock);
5721 if (rq->active_balance) {
5722 active_load_balance(rq, cpu);
5723 rq->active_balance = 0;
5726 head = &rq->migration_queue;
5728 if (list_empty(head)) {
5729 spin_unlock_irq(&rq->lock);
5731 set_current_state(TASK_INTERRUPTIBLE);
5734 req = list_entry(head->next, struct migration_req, list);
5735 list_del_init(head->next);
5737 spin_unlock(&rq->lock);
5738 __migrate_task(req->task, cpu, req->dest_cpu);
5741 complete(&req->done);
5743 __set_current_state(TASK_RUNNING);
5747 /* Wait for kthread_stop */
5748 set_current_state(TASK_INTERRUPTIBLE);
5749 while (!kthread_should_stop()) {
5751 set_current_state(TASK_INTERRUPTIBLE);
5753 __set_current_state(TASK_RUNNING);
5757 #ifdef CONFIG_HOTPLUG_CPU
5759 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5763 local_irq_disable();
5764 ret = __migrate_task(p, src_cpu, dest_cpu);
5770 * Figure out where task on dead CPU should go, use force if necessary.
5771 * NOTE: interrupts should be disabled by the caller
5773 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5775 unsigned long flags;
5782 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5783 cpus_and(mask, mask, p->cpus_allowed);
5784 dest_cpu = any_online_cpu(mask);
5786 /* On any allowed CPU? */
5787 if (dest_cpu >= nr_cpu_ids)
5788 dest_cpu = any_online_cpu(p->cpus_allowed);
5790 /* No more Mr. Nice Guy. */
5791 if (dest_cpu >= nr_cpu_ids) {
5792 cpumask_t cpus_allowed;
5794 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5796 * Try to stay on the same cpuset, where the
5797 * current cpuset may be a subset of all cpus.
5798 * The cpuset_cpus_allowed_locked() variant of
5799 * cpuset_cpus_allowed() will not block. It must be
5800 * called within calls to cpuset_lock/cpuset_unlock.
5802 rq = task_rq_lock(p, &flags);
5803 p->cpus_allowed = cpus_allowed;
5804 dest_cpu = any_online_cpu(p->cpus_allowed);
5805 task_rq_unlock(rq, &flags);
5808 * Don't tell them about moving exiting tasks or
5809 * kernel threads (both mm NULL), since they never
5812 if (p->mm && printk_ratelimit()) {
5813 printk(KERN_INFO "process %d (%s) no "
5814 "longer affine to cpu%d\n",
5815 task_pid_nr(p), p->comm, dead_cpu);
5818 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5822 * While a dead CPU has no uninterruptible tasks queued at this point,
5823 * it might still have a nonzero ->nr_uninterruptible counter, because
5824 * for performance reasons the counter is not stricly tracking tasks to
5825 * their home CPUs. So we just add the counter to another CPU's counter,
5826 * to keep the global sum constant after CPU-down:
5828 static void migrate_nr_uninterruptible(struct rq *rq_src)
5830 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5831 unsigned long flags;
5833 local_irq_save(flags);
5834 double_rq_lock(rq_src, rq_dest);
5835 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5836 rq_src->nr_uninterruptible = 0;
5837 double_rq_unlock(rq_src, rq_dest);
5838 local_irq_restore(flags);
5841 /* Run through task list and migrate tasks from the dead cpu. */
5842 static void migrate_live_tasks(int src_cpu)
5844 struct task_struct *p, *t;
5846 read_lock(&tasklist_lock);
5848 do_each_thread(t, p) {
5852 if (task_cpu(p) == src_cpu)
5853 move_task_off_dead_cpu(src_cpu, p);
5854 } while_each_thread(t, p);
5856 read_unlock(&tasklist_lock);
5860 * Schedules idle task to be the next runnable task on current CPU.
5861 * It does so by boosting its priority to highest possible.
5862 * Used by CPU offline code.
5864 void sched_idle_next(void)
5866 int this_cpu = smp_processor_id();
5867 struct rq *rq = cpu_rq(this_cpu);
5868 struct task_struct *p = rq->idle;
5869 unsigned long flags;
5871 /* cpu has to be offline */
5872 BUG_ON(cpu_online(this_cpu));
5875 * Strictly not necessary since rest of the CPUs are stopped by now
5876 * and interrupts disabled on the current cpu.
5878 spin_lock_irqsave(&rq->lock, flags);
5880 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5882 update_rq_clock(rq);
5883 activate_task(rq, p, 0);
5885 spin_unlock_irqrestore(&rq->lock, flags);
5889 * Ensures that the idle task is using init_mm right before its cpu goes
5892 void idle_task_exit(void)
5894 struct mm_struct *mm = current->active_mm;
5896 BUG_ON(cpu_online(smp_processor_id()));
5899 switch_mm(mm, &init_mm, current);
5903 /* called under rq->lock with disabled interrupts */
5904 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5906 struct rq *rq = cpu_rq(dead_cpu);
5908 /* Must be exiting, otherwise would be on tasklist. */
5909 BUG_ON(!p->exit_state);
5911 /* Cannot have done final schedule yet: would have vanished. */
5912 BUG_ON(p->state == TASK_DEAD);
5917 * Drop lock around migration; if someone else moves it,
5918 * that's OK. No task can be added to this CPU, so iteration is
5921 spin_unlock_irq(&rq->lock);
5922 move_task_off_dead_cpu(dead_cpu, p);
5923 spin_lock_irq(&rq->lock);
5928 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5929 static void migrate_dead_tasks(unsigned int dead_cpu)
5931 struct rq *rq = cpu_rq(dead_cpu);
5932 struct task_struct *next;
5935 if (!rq->nr_running)
5937 update_rq_clock(rq);
5938 next = pick_next_task(rq, rq->curr);
5941 migrate_dead(dead_cpu, next);
5945 #endif /* CONFIG_HOTPLUG_CPU */
5947 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5949 static struct ctl_table sd_ctl_dir[] = {
5951 .procname = "sched_domain",
5957 static struct ctl_table sd_ctl_root[] = {
5959 .ctl_name = CTL_KERN,
5960 .procname = "kernel",
5962 .child = sd_ctl_dir,
5967 static struct ctl_table *sd_alloc_ctl_entry(int n)
5969 struct ctl_table *entry =
5970 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5975 static void sd_free_ctl_entry(struct ctl_table **tablep)
5977 struct ctl_table *entry;
5980 * In the intermediate directories, both the child directory and
5981 * procname are dynamically allocated and could fail but the mode
5982 * will always be set. In the lowest directory the names are
5983 * static strings and all have proc handlers.
5985 for (entry = *tablep; entry->mode; entry++) {
5987 sd_free_ctl_entry(&entry->child);
5988 if (entry->proc_handler == NULL)
5989 kfree(entry->procname);
5997 set_table_entry(struct ctl_table *entry,
5998 const char *procname, void *data, int maxlen,
5999 mode_t mode, proc_handler *proc_handler)
6001 entry->procname = procname;
6003 entry->maxlen = maxlen;
6005 entry->proc_handler = proc_handler;
6008 static struct ctl_table *
6009 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6011 struct ctl_table *table = sd_alloc_ctl_entry(12);
6016 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6017 sizeof(long), 0644, proc_doulongvec_minmax);
6018 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6019 sizeof(long), 0644, proc_doulongvec_minmax);
6020 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6021 sizeof(int), 0644, proc_dointvec_minmax);
6022 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6023 sizeof(int), 0644, proc_dointvec_minmax);
6024 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6025 sizeof(int), 0644, proc_dointvec_minmax);
6026 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6027 sizeof(int), 0644, proc_dointvec_minmax);
6028 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6029 sizeof(int), 0644, proc_dointvec_minmax);
6030 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6031 sizeof(int), 0644, proc_dointvec_minmax);
6032 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6033 sizeof(int), 0644, proc_dointvec_minmax);
6034 set_table_entry(&table[9], "cache_nice_tries",
6035 &sd->cache_nice_tries,
6036 sizeof(int), 0644, proc_dointvec_minmax);
6037 set_table_entry(&table[10], "flags", &sd->flags,
6038 sizeof(int), 0644, proc_dointvec_minmax);
6039 /* &table[11] is terminator */
6044 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6046 struct ctl_table *entry, *table;
6047 struct sched_domain *sd;
6048 int domain_num = 0, i;
6051 for_each_domain(cpu, sd)
6053 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6058 for_each_domain(cpu, sd) {
6059 snprintf(buf, 32, "domain%d", i);
6060 entry->procname = kstrdup(buf, GFP_KERNEL);
6062 entry->child = sd_alloc_ctl_domain_table(sd);
6069 static struct ctl_table_header *sd_sysctl_header;
6070 static void register_sched_domain_sysctl(void)
6072 int i, cpu_num = num_online_cpus();
6073 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6076 WARN_ON(sd_ctl_dir[0].child);
6077 sd_ctl_dir[0].child = entry;
6082 for_each_online_cpu(i) {
6083 snprintf(buf, 32, "cpu%d", i);
6084 entry->procname = kstrdup(buf, GFP_KERNEL);
6086 entry->child = sd_alloc_ctl_cpu_table(i);
6090 WARN_ON(sd_sysctl_header);
6091 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6094 /* may be called multiple times per register */
6095 static void unregister_sched_domain_sysctl(void)
6097 if (sd_sysctl_header)
6098 unregister_sysctl_table(sd_sysctl_header);
6099 sd_sysctl_header = NULL;
6100 if (sd_ctl_dir[0].child)
6101 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6104 static void register_sched_domain_sysctl(void)
6107 static void unregister_sched_domain_sysctl(void)
6113 * migration_call - callback that gets triggered when a CPU is added.
6114 * Here we can start up the necessary migration thread for the new CPU.
6116 static int __cpuinit
6117 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6119 struct task_struct *p;
6120 int cpu = (long)hcpu;
6121 unsigned long flags;
6126 case CPU_UP_PREPARE:
6127 case CPU_UP_PREPARE_FROZEN:
6128 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6131 kthread_bind(p, cpu);
6132 /* Must be high prio: stop_machine expects to yield to it. */
6133 rq = task_rq_lock(p, &flags);
6134 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6135 task_rq_unlock(rq, &flags);
6136 cpu_rq(cpu)->migration_thread = p;
6140 case CPU_ONLINE_FROZEN:
6141 /* Strictly unnecessary, as first user will wake it. */
6142 wake_up_process(cpu_rq(cpu)->migration_thread);
6144 /* Update our root-domain */
6146 spin_lock_irqsave(&rq->lock, flags);
6148 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6149 cpu_set(cpu, rq->rd->online);
6151 spin_unlock_irqrestore(&rq->lock, flags);
6154 #ifdef CONFIG_HOTPLUG_CPU
6155 case CPU_UP_CANCELED:
6156 case CPU_UP_CANCELED_FROZEN:
6157 if (!cpu_rq(cpu)->migration_thread)
6159 /* Unbind it from offline cpu so it can run. Fall thru. */
6160 kthread_bind(cpu_rq(cpu)->migration_thread,
6161 any_online_cpu(cpu_online_map));
6162 kthread_stop(cpu_rq(cpu)->migration_thread);
6163 cpu_rq(cpu)->migration_thread = NULL;
6167 case CPU_DEAD_FROZEN:
6168 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6169 migrate_live_tasks(cpu);
6171 kthread_stop(rq->migration_thread);
6172 rq->migration_thread = NULL;
6173 /* Idle task back to normal (off runqueue, low prio) */
6174 spin_lock_irq(&rq->lock);
6175 update_rq_clock(rq);
6176 deactivate_task(rq, rq->idle, 0);
6177 rq->idle->static_prio = MAX_PRIO;
6178 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6179 rq->idle->sched_class = &idle_sched_class;
6180 migrate_dead_tasks(cpu);
6181 spin_unlock_irq(&rq->lock);
6183 migrate_nr_uninterruptible(rq);
6184 BUG_ON(rq->nr_running != 0);
6187 * No need to migrate the tasks: it was best-effort if
6188 * they didn't take sched_hotcpu_mutex. Just wake up
6191 spin_lock_irq(&rq->lock);
6192 while (!list_empty(&rq->migration_queue)) {
6193 struct migration_req *req;
6195 req = list_entry(rq->migration_queue.next,
6196 struct migration_req, list);
6197 list_del_init(&req->list);
6198 complete(&req->done);
6200 spin_unlock_irq(&rq->lock);
6204 case CPU_DYING_FROZEN:
6205 /* Update our root-domain */
6207 spin_lock_irqsave(&rq->lock, flags);
6209 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6210 cpu_clear(cpu, rq->rd->online);
6212 spin_unlock_irqrestore(&rq->lock, flags);
6219 /* Register at highest priority so that task migration (migrate_all_tasks)
6220 * happens before everything else.
6222 static struct notifier_block __cpuinitdata migration_notifier = {
6223 .notifier_call = migration_call,
6227 void __init migration_init(void)
6229 void *cpu = (void *)(long)smp_processor_id();
6232 /* Start one for the boot CPU: */
6233 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6234 BUG_ON(err == NOTIFY_BAD);
6235 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6236 register_cpu_notifier(&migration_notifier);
6242 #ifdef CONFIG_SCHED_DEBUG
6244 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6245 cpumask_t *groupmask)
6247 struct sched_group *group = sd->groups;
6250 cpulist_scnprintf(str, sizeof(str), sd->span);
6251 cpus_clear(*groupmask);
6253 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6255 if (!(sd->flags & SD_LOAD_BALANCE)) {
6256 printk("does not load-balance\n");
6258 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6263 printk(KERN_CONT "span %s\n", str);
6265 if (!cpu_isset(cpu, sd->span)) {
6266 printk(KERN_ERR "ERROR: domain->span does not contain "
6269 if (!cpu_isset(cpu, group->cpumask)) {
6270 printk(KERN_ERR "ERROR: domain->groups does not contain"
6274 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6278 printk(KERN_ERR "ERROR: group is NULL\n");
6282 if (!group->__cpu_power) {
6283 printk(KERN_CONT "\n");
6284 printk(KERN_ERR "ERROR: domain->cpu_power not "
6289 if (!cpus_weight(group->cpumask)) {
6290 printk(KERN_CONT "\n");
6291 printk(KERN_ERR "ERROR: empty group\n");
6295 if (cpus_intersects(*groupmask, group->cpumask)) {
6296 printk(KERN_CONT "\n");
6297 printk(KERN_ERR "ERROR: repeated CPUs\n");
6301 cpus_or(*groupmask, *groupmask, group->cpumask);
6303 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6304 printk(KERN_CONT " %s", str);
6306 group = group->next;
6307 } while (group != sd->groups);
6308 printk(KERN_CONT "\n");
6310 if (!cpus_equal(sd->span, *groupmask))
6311 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6313 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6314 printk(KERN_ERR "ERROR: parent span is not a superset "
6315 "of domain->span\n");
6319 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6321 cpumask_t *groupmask;
6325 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6329 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6331 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6333 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6338 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6348 # define sched_domain_debug(sd, cpu) do { } while (0)
6351 static int sd_degenerate(struct sched_domain *sd)
6353 if (cpus_weight(sd->span) == 1)
6356 /* Following flags need at least 2 groups */
6357 if (sd->flags & (SD_LOAD_BALANCE |
6358 SD_BALANCE_NEWIDLE |
6362 SD_SHARE_PKG_RESOURCES)) {
6363 if (sd->groups != sd->groups->next)
6367 /* Following flags don't use groups */
6368 if (sd->flags & (SD_WAKE_IDLE |
6377 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6379 unsigned long cflags = sd->flags, pflags = parent->flags;
6381 if (sd_degenerate(parent))
6384 if (!cpus_equal(sd->span, parent->span))
6387 /* Does parent contain flags not in child? */
6388 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6389 if (cflags & SD_WAKE_AFFINE)
6390 pflags &= ~SD_WAKE_BALANCE;
6391 /* Flags needing groups don't count if only 1 group in parent */
6392 if (parent->groups == parent->groups->next) {
6393 pflags &= ~(SD_LOAD_BALANCE |
6394 SD_BALANCE_NEWIDLE |
6398 SD_SHARE_PKG_RESOURCES);
6400 if (~cflags & pflags)
6406 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6408 unsigned long flags;
6409 const struct sched_class *class;
6411 spin_lock_irqsave(&rq->lock, flags);
6414 struct root_domain *old_rd = rq->rd;
6416 for (class = sched_class_highest; class; class = class->next) {
6417 if (class->leave_domain)
6418 class->leave_domain(rq);
6421 cpu_clear(rq->cpu, old_rd->span);
6422 cpu_clear(rq->cpu, old_rd->online);
6424 if (atomic_dec_and_test(&old_rd->refcount))
6428 atomic_inc(&rd->refcount);
6431 cpu_set(rq->cpu, rd->span);
6432 if (cpu_isset(rq->cpu, cpu_online_map))
6433 cpu_set(rq->cpu, rd->online);
6435 for (class = sched_class_highest; class; class = class->next) {
6436 if (class->join_domain)
6437 class->join_domain(rq);
6440 spin_unlock_irqrestore(&rq->lock, flags);
6443 static void init_rootdomain(struct root_domain *rd)
6445 memset(rd, 0, sizeof(*rd));
6447 cpus_clear(rd->span);
6448 cpus_clear(rd->online);
6451 static void init_defrootdomain(void)
6453 init_rootdomain(&def_root_domain);
6454 atomic_set(&def_root_domain.refcount, 1);
6457 static struct root_domain *alloc_rootdomain(void)
6459 struct root_domain *rd;
6461 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6465 init_rootdomain(rd);
6471 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6472 * hold the hotplug lock.
6475 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6477 struct rq *rq = cpu_rq(cpu);
6478 struct sched_domain *tmp;
6480 /* Remove the sched domains which do not contribute to scheduling. */
6481 for (tmp = sd; tmp; tmp = tmp->parent) {
6482 struct sched_domain *parent = tmp->parent;
6485 if (sd_parent_degenerate(tmp, parent)) {
6486 tmp->parent = parent->parent;
6488 parent->parent->child = tmp;
6492 if (sd && sd_degenerate(sd)) {
6498 sched_domain_debug(sd, cpu);
6500 rq_attach_root(rq, rd);
6501 rcu_assign_pointer(rq->sd, sd);
6504 /* cpus with isolated domains */
6505 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6507 /* Setup the mask of cpus configured for isolated domains */
6508 static int __init isolated_cpu_setup(char *str)
6510 int ints[NR_CPUS], i;
6512 str = get_options(str, ARRAY_SIZE(ints), ints);
6513 cpus_clear(cpu_isolated_map);
6514 for (i = 1; i <= ints[0]; i++)
6515 if (ints[i] < NR_CPUS)
6516 cpu_set(ints[i], cpu_isolated_map);
6520 __setup("isolcpus=", isolated_cpu_setup);
6523 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6524 * to a function which identifies what group(along with sched group) a CPU
6525 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6526 * (due to the fact that we keep track of groups covered with a cpumask_t).
6528 * init_sched_build_groups will build a circular linked list of the groups
6529 * covered by the given span, and will set each group's ->cpumask correctly,
6530 * and ->cpu_power to 0.
6533 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6534 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6535 struct sched_group **sg,
6536 cpumask_t *tmpmask),
6537 cpumask_t *covered, cpumask_t *tmpmask)
6539 struct sched_group *first = NULL, *last = NULL;
6542 cpus_clear(*covered);
6544 for_each_cpu_mask(i, *span) {
6545 struct sched_group *sg;
6546 int group = group_fn(i, cpu_map, &sg, tmpmask);
6549 if (cpu_isset(i, *covered))
6552 cpus_clear(sg->cpumask);
6553 sg->__cpu_power = 0;
6555 for_each_cpu_mask(j, *span) {
6556 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6559 cpu_set(j, *covered);
6560 cpu_set(j, sg->cpumask);
6571 #define SD_NODES_PER_DOMAIN 16
6576 * find_next_best_node - find the next node to include in a sched_domain
6577 * @node: node whose sched_domain we're building
6578 * @used_nodes: nodes already in the sched_domain
6580 * Find the next node to include in a given scheduling domain. Simply
6581 * finds the closest node not already in the @used_nodes map.
6583 * Should use nodemask_t.
6585 static int find_next_best_node(int node, nodemask_t *used_nodes)
6587 int i, n, val, min_val, best_node = 0;
6591 for (i = 0; i < MAX_NUMNODES; i++) {
6592 /* Start at @node */
6593 n = (node + i) % MAX_NUMNODES;
6595 if (!nr_cpus_node(n))
6598 /* Skip already used nodes */
6599 if (node_isset(n, *used_nodes))
6602 /* Simple min distance search */
6603 val = node_distance(node, n);
6605 if (val < min_val) {
6611 node_set(best_node, *used_nodes);
6616 * sched_domain_node_span - get a cpumask for a node's sched_domain
6617 * @node: node whose cpumask we're constructing
6618 * @span: resulting cpumask
6620 * Given a node, construct a good cpumask for its sched_domain to span. It
6621 * should be one that prevents unnecessary balancing, but also spreads tasks
6624 static void sched_domain_node_span(int node, cpumask_t *span)
6626 nodemask_t used_nodes;
6627 node_to_cpumask_ptr(nodemask, node);
6631 nodes_clear(used_nodes);
6633 cpus_or(*span, *span, *nodemask);
6634 node_set(node, used_nodes);
6636 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6637 int next_node = find_next_best_node(node, &used_nodes);
6639 node_to_cpumask_ptr_next(nodemask, next_node);
6640 cpus_or(*span, *span, *nodemask);
6645 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6648 * SMT sched-domains:
6650 #ifdef CONFIG_SCHED_SMT
6651 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6652 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6655 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6659 *sg = &per_cpu(sched_group_cpus, cpu);
6665 * multi-core sched-domains:
6667 #ifdef CONFIG_SCHED_MC
6668 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6669 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6672 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6674 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6679 *mask = per_cpu(cpu_sibling_map, cpu);
6680 cpus_and(*mask, *mask, *cpu_map);
6681 group = first_cpu(*mask);
6683 *sg = &per_cpu(sched_group_core, group);
6686 #elif defined(CONFIG_SCHED_MC)
6688 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6692 *sg = &per_cpu(sched_group_core, cpu);
6697 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6698 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6701 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6705 #ifdef CONFIG_SCHED_MC
6706 *mask = cpu_coregroup_map(cpu);
6707 cpus_and(*mask, *mask, *cpu_map);
6708 group = first_cpu(*mask);
6709 #elif defined(CONFIG_SCHED_SMT)
6710 *mask = per_cpu(cpu_sibling_map, cpu);
6711 cpus_and(*mask, *mask, *cpu_map);
6712 group = first_cpu(*mask);
6717 *sg = &per_cpu(sched_group_phys, group);
6723 * The init_sched_build_groups can't handle what we want to do with node
6724 * groups, so roll our own. Now each node has its own list of groups which
6725 * gets dynamically allocated.
6727 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6728 static struct sched_group ***sched_group_nodes_bycpu;
6730 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6731 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6733 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6734 struct sched_group **sg, cpumask_t *nodemask)
6738 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6739 cpus_and(*nodemask, *nodemask, *cpu_map);
6740 group = first_cpu(*nodemask);
6743 *sg = &per_cpu(sched_group_allnodes, group);
6747 static void init_numa_sched_groups_power(struct sched_group *group_head)
6749 struct sched_group *sg = group_head;
6755 for_each_cpu_mask(j, sg->cpumask) {
6756 struct sched_domain *sd;
6758 sd = &per_cpu(phys_domains, j);
6759 if (j != first_cpu(sd->groups->cpumask)) {
6761 * Only add "power" once for each
6767 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6770 } while (sg != group_head);
6775 /* Free memory allocated for various sched_group structures */
6776 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6780 for_each_cpu_mask(cpu, *cpu_map) {
6781 struct sched_group **sched_group_nodes
6782 = sched_group_nodes_bycpu[cpu];
6784 if (!sched_group_nodes)
6787 for (i = 0; i < MAX_NUMNODES; i++) {
6788 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6790 *nodemask = node_to_cpumask(i);
6791 cpus_and(*nodemask, *nodemask, *cpu_map);
6792 if (cpus_empty(*nodemask))
6802 if (oldsg != sched_group_nodes[i])
6805 kfree(sched_group_nodes);
6806 sched_group_nodes_bycpu[cpu] = NULL;
6810 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6816 * Initialize sched groups cpu_power.
6818 * cpu_power indicates the capacity of sched group, which is used while
6819 * distributing the load between different sched groups in a sched domain.
6820 * Typically cpu_power for all the groups in a sched domain will be same unless
6821 * there are asymmetries in the topology. If there are asymmetries, group
6822 * having more cpu_power will pickup more load compared to the group having
6825 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6826 * the maximum number of tasks a group can handle in the presence of other idle
6827 * or lightly loaded groups in the same sched domain.
6829 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6831 struct sched_domain *child;
6832 struct sched_group *group;
6834 WARN_ON(!sd || !sd->groups);
6836 if (cpu != first_cpu(sd->groups->cpumask))
6841 sd->groups->__cpu_power = 0;
6844 * For perf policy, if the groups in child domain share resources
6845 * (for example cores sharing some portions of the cache hierarchy
6846 * or SMT), then set this domain groups cpu_power such that each group
6847 * can handle only one task, when there are other idle groups in the
6848 * same sched domain.
6850 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6852 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6853 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6858 * add cpu_power of each child group to this groups cpu_power
6860 group = child->groups;
6862 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6863 group = group->next;
6864 } while (group != child->groups);
6868 * Initializers for schedule domains
6869 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6872 #define SD_INIT(sd, type) sd_init_##type(sd)
6873 #define SD_INIT_FUNC(type) \
6874 static noinline void sd_init_##type(struct sched_domain *sd) \
6876 memset(sd, 0, sizeof(*sd)); \
6877 *sd = SD_##type##_INIT; \
6878 sd->level = SD_LV_##type; \
6883 SD_INIT_FUNC(ALLNODES)
6886 #ifdef CONFIG_SCHED_SMT
6887 SD_INIT_FUNC(SIBLING)
6889 #ifdef CONFIG_SCHED_MC
6894 * To minimize stack usage kmalloc room for cpumasks and share the
6895 * space as the usage in build_sched_domains() dictates. Used only
6896 * if the amount of space is significant.
6899 cpumask_t tmpmask; /* make this one first */
6902 cpumask_t this_sibling_map;
6903 cpumask_t this_core_map;
6905 cpumask_t send_covered;
6908 cpumask_t domainspan;
6910 cpumask_t notcovered;
6915 #define SCHED_CPUMASK_ALLOC 1
6916 #define SCHED_CPUMASK_FREE(v) kfree(v)
6917 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
6919 #define SCHED_CPUMASK_ALLOC 0
6920 #define SCHED_CPUMASK_FREE(v)
6921 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
6924 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
6925 ((unsigned long)(a) + offsetof(struct allmasks, v))
6927 static int default_relax_domain_level = -1;
6929 static int __init setup_relax_domain_level(char *str)
6931 default_relax_domain_level = simple_strtoul(str, NULL, 0);
6934 __setup("relax_domain_level=", setup_relax_domain_level);
6936 static void set_domain_attribute(struct sched_domain *sd,
6937 struct sched_domain_attr *attr)
6941 if (!attr || attr->relax_domain_level < 0) {
6942 if (default_relax_domain_level < 0)
6945 request = default_relax_domain_level;
6947 request = attr->relax_domain_level;
6948 if (request < sd->level) {
6949 /* turn off idle balance on this domain */
6950 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
6952 /* turn on idle balance on this domain */
6953 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
6958 * Build sched domains for a given set of cpus and attach the sched domains
6959 * to the individual cpus
6961 static int __build_sched_domains(const cpumask_t *cpu_map,
6962 struct sched_domain_attr *attr)
6965 struct root_domain *rd;
6966 SCHED_CPUMASK_DECLARE(allmasks);
6969 struct sched_group **sched_group_nodes = NULL;
6970 int sd_allnodes = 0;
6973 * Allocate the per-node list of sched groups
6975 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6977 if (!sched_group_nodes) {
6978 printk(KERN_WARNING "Can not alloc sched group node list\n");
6983 rd = alloc_rootdomain();
6985 printk(KERN_WARNING "Cannot alloc root domain\n");
6987 kfree(sched_group_nodes);
6992 #if SCHED_CPUMASK_ALLOC
6993 /* get space for all scratch cpumask variables */
6994 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
6996 printk(KERN_WARNING "Cannot alloc cpumask array\n");
6999 kfree(sched_group_nodes);
7004 tmpmask = (cpumask_t *)allmasks;
7008 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7012 * Set up domains for cpus specified by the cpu_map.
7014 for_each_cpu_mask(i, *cpu_map) {
7015 struct sched_domain *sd = NULL, *p;
7016 SCHED_CPUMASK_VAR(nodemask, allmasks);
7018 *nodemask = node_to_cpumask(cpu_to_node(i));
7019 cpus_and(*nodemask, *nodemask, *cpu_map);
7022 if (cpus_weight(*cpu_map) >
7023 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7024 sd = &per_cpu(allnodes_domains, i);
7025 SD_INIT(sd, ALLNODES);
7026 set_domain_attribute(sd, attr);
7027 sd->span = *cpu_map;
7028 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7034 sd = &per_cpu(node_domains, i);
7036 set_domain_attribute(sd, attr);
7037 sched_domain_node_span(cpu_to_node(i), &sd->span);
7041 cpus_and(sd->span, sd->span, *cpu_map);
7045 sd = &per_cpu(phys_domains, i);
7047 set_domain_attribute(sd, attr);
7048 sd->span = *nodemask;
7052 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7054 #ifdef CONFIG_SCHED_MC
7056 sd = &per_cpu(core_domains, i);
7058 set_domain_attribute(sd, attr);
7059 sd->span = cpu_coregroup_map(i);
7060 cpus_and(sd->span, sd->span, *cpu_map);
7063 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7066 #ifdef CONFIG_SCHED_SMT
7068 sd = &per_cpu(cpu_domains, i);
7069 SD_INIT(sd, SIBLING);
7070 set_domain_attribute(sd, attr);
7071 sd->span = per_cpu(cpu_sibling_map, i);
7072 cpus_and(sd->span, sd->span, *cpu_map);
7075 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7079 #ifdef CONFIG_SCHED_SMT
7080 /* Set up CPU (sibling) groups */
7081 for_each_cpu_mask(i, *cpu_map) {
7082 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7083 SCHED_CPUMASK_VAR(send_covered, allmasks);
7085 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7086 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7087 if (i != first_cpu(*this_sibling_map))
7090 init_sched_build_groups(this_sibling_map, cpu_map,
7092 send_covered, tmpmask);
7096 #ifdef CONFIG_SCHED_MC
7097 /* Set up multi-core groups */
7098 for_each_cpu_mask(i, *cpu_map) {
7099 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7100 SCHED_CPUMASK_VAR(send_covered, allmasks);
7102 *this_core_map = cpu_coregroup_map(i);
7103 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7104 if (i != first_cpu(*this_core_map))
7107 init_sched_build_groups(this_core_map, cpu_map,
7109 send_covered, tmpmask);
7113 /* Set up physical groups */
7114 for (i = 0; i < MAX_NUMNODES; i++) {
7115 SCHED_CPUMASK_VAR(nodemask, allmasks);
7116 SCHED_CPUMASK_VAR(send_covered, allmasks);
7118 *nodemask = node_to_cpumask(i);
7119 cpus_and(*nodemask, *nodemask, *cpu_map);
7120 if (cpus_empty(*nodemask))
7123 init_sched_build_groups(nodemask, cpu_map,
7125 send_covered, tmpmask);
7129 /* Set up node groups */
7131 SCHED_CPUMASK_VAR(send_covered, allmasks);
7133 init_sched_build_groups(cpu_map, cpu_map,
7134 &cpu_to_allnodes_group,
7135 send_covered, tmpmask);
7138 for (i = 0; i < MAX_NUMNODES; i++) {
7139 /* Set up node groups */
7140 struct sched_group *sg, *prev;
7141 SCHED_CPUMASK_VAR(nodemask, allmasks);
7142 SCHED_CPUMASK_VAR(domainspan, allmasks);
7143 SCHED_CPUMASK_VAR(covered, allmasks);
7146 *nodemask = node_to_cpumask(i);
7147 cpus_clear(*covered);
7149 cpus_and(*nodemask, *nodemask, *cpu_map);
7150 if (cpus_empty(*nodemask)) {
7151 sched_group_nodes[i] = NULL;
7155 sched_domain_node_span(i, domainspan);
7156 cpus_and(*domainspan, *domainspan, *cpu_map);
7158 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7160 printk(KERN_WARNING "Can not alloc domain group for "
7164 sched_group_nodes[i] = sg;
7165 for_each_cpu_mask(j, *nodemask) {
7166 struct sched_domain *sd;
7168 sd = &per_cpu(node_domains, j);
7171 sg->__cpu_power = 0;
7172 sg->cpumask = *nodemask;
7174 cpus_or(*covered, *covered, *nodemask);
7177 for (j = 0; j < MAX_NUMNODES; j++) {
7178 SCHED_CPUMASK_VAR(notcovered, allmasks);
7179 int n = (i + j) % MAX_NUMNODES;
7180 node_to_cpumask_ptr(pnodemask, n);
7182 cpus_complement(*notcovered, *covered);
7183 cpus_and(*tmpmask, *notcovered, *cpu_map);
7184 cpus_and(*tmpmask, *tmpmask, *domainspan);
7185 if (cpus_empty(*tmpmask))
7188 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7189 if (cpus_empty(*tmpmask))
7192 sg = kmalloc_node(sizeof(struct sched_group),
7196 "Can not alloc domain group for node %d\n", j);
7199 sg->__cpu_power = 0;
7200 sg->cpumask = *tmpmask;
7201 sg->next = prev->next;
7202 cpus_or(*covered, *covered, *tmpmask);
7209 /* Calculate CPU power for physical packages and nodes */
7210 #ifdef CONFIG_SCHED_SMT
7211 for_each_cpu_mask(i, *cpu_map) {
7212 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7214 init_sched_groups_power(i, sd);
7217 #ifdef CONFIG_SCHED_MC
7218 for_each_cpu_mask(i, *cpu_map) {
7219 struct sched_domain *sd = &per_cpu(core_domains, i);
7221 init_sched_groups_power(i, sd);
7225 for_each_cpu_mask(i, *cpu_map) {
7226 struct sched_domain *sd = &per_cpu(phys_domains, i);
7228 init_sched_groups_power(i, sd);
7232 for (i = 0; i < MAX_NUMNODES; i++)
7233 init_numa_sched_groups_power(sched_group_nodes[i]);
7236 struct sched_group *sg;
7238 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7240 init_numa_sched_groups_power(sg);
7244 /* Attach the domains */
7245 for_each_cpu_mask(i, *cpu_map) {
7246 struct sched_domain *sd;
7247 #ifdef CONFIG_SCHED_SMT
7248 sd = &per_cpu(cpu_domains, i);
7249 #elif defined(CONFIG_SCHED_MC)
7250 sd = &per_cpu(core_domains, i);
7252 sd = &per_cpu(phys_domains, i);
7254 cpu_attach_domain(sd, rd, i);
7257 SCHED_CPUMASK_FREE((void *)allmasks);
7262 free_sched_groups(cpu_map, tmpmask);
7263 SCHED_CPUMASK_FREE((void *)allmasks);
7268 static int build_sched_domains(const cpumask_t *cpu_map)
7270 return __build_sched_domains(cpu_map, NULL);
7273 static cpumask_t *doms_cur; /* current sched domains */
7274 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7275 static struct sched_domain_attr *dattr_cur;
7276 /* attribues of custom domains in 'doms_cur' */
7279 * Special case: If a kmalloc of a doms_cur partition (array of
7280 * cpumask_t) fails, then fallback to a single sched domain,
7281 * as determined by the single cpumask_t fallback_doms.
7283 static cpumask_t fallback_doms;
7285 void __attribute__((weak)) arch_update_cpu_topology(void)
7290 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7291 * For now this just excludes isolated cpus, but could be used to
7292 * exclude other special cases in the future.
7294 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7298 arch_update_cpu_topology();
7300 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7302 doms_cur = &fallback_doms;
7303 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7305 err = build_sched_domains(doms_cur);
7306 register_sched_domain_sysctl();
7311 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7314 free_sched_groups(cpu_map, tmpmask);
7318 * Detach sched domains from a group of cpus specified in cpu_map
7319 * These cpus will now be attached to the NULL domain
7321 static void detach_destroy_domains(const cpumask_t *cpu_map)
7326 unregister_sched_domain_sysctl();
7328 for_each_cpu_mask(i, *cpu_map)
7329 cpu_attach_domain(NULL, &def_root_domain, i);
7330 synchronize_sched();
7331 arch_destroy_sched_domains(cpu_map, &tmpmask);
7334 /* handle null as "default" */
7335 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7336 struct sched_domain_attr *new, int idx_new)
7338 struct sched_domain_attr tmp;
7345 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7346 new ? (new + idx_new) : &tmp,
7347 sizeof(struct sched_domain_attr));
7351 * Partition sched domains as specified by the 'ndoms_new'
7352 * cpumasks in the array doms_new[] of cpumasks. This compares
7353 * doms_new[] to the current sched domain partitioning, doms_cur[].
7354 * It destroys each deleted domain and builds each new domain.
7356 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7357 * The masks don't intersect (don't overlap.) We should setup one
7358 * sched domain for each mask. CPUs not in any of the cpumasks will
7359 * not be load balanced. If the same cpumask appears both in the
7360 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7363 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7364 * ownership of it and will kfree it when done with it. If the caller
7365 * failed the kmalloc call, then it can pass in doms_new == NULL,
7366 * and partition_sched_domains() will fallback to the single partition
7369 * Call with hotplug lock held
7371 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7372 struct sched_domain_attr *dattr_new)
7376 mutex_lock(&sched_domains_mutex);
7378 /* always unregister in case we don't destroy any domains */
7379 unregister_sched_domain_sysctl();
7381 if (doms_new == NULL) {
7383 doms_new = &fallback_doms;
7384 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7388 /* Destroy deleted domains */
7389 for (i = 0; i < ndoms_cur; i++) {
7390 for (j = 0; j < ndoms_new; j++) {
7391 if (cpus_equal(doms_cur[i], doms_new[j])
7392 && dattrs_equal(dattr_cur, i, dattr_new, j))
7395 /* no match - a current sched domain not in new doms_new[] */
7396 detach_destroy_domains(doms_cur + i);
7401 /* Build new domains */
7402 for (i = 0; i < ndoms_new; i++) {
7403 for (j = 0; j < ndoms_cur; j++) {
7404 if (cpus_equal(doms_new[i], doms_cur[j])
7405 && dattrs_equal(dattr_new, i, dattr_cur, j))
7408 /* no match - add a new doms_new */
7409 __build_sched_domains(doms_new + i,
7410 dattr_new ? dattr_new + i : NULL);
7415 /* Remember the new sched domains */
7416 if (doms_cur != &fallback_doms)
7418 kfree(dattr_cur); /* kfree(NULL) is safe */
7419 doms_cur = doms_new;
7420 dattr_cur = dattr_new;
7421 ndoms_cur = ndoms_new;
7423 register_sched_domain_sysctl();
7425 mutex_unlock(&sched_domains_mutex);
7428 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7429 int arch_reinit_sched_domains(void)
7434 mutex_lock(&sched_domains_mutex);
7435 detach_destroy_domains(&cpu_online_map);
7436 err = arch_init_sched_domains(&cpu_online_map);
7437 mutex_unlock(&sched_domains_mutex);
7443 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7447 if (buf[0] != '0' && buf[0] != '1')
7451 sched_smt_power_savings = (buf[0] == '1');
7453 sched_mc_power_savings = (buf[0] == '1');
7455 ret = arch_reinit_sched_domains();
7457 return ret ? ret : count;
7460 #ifdef CONFIG_SCHED_MC
7461 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7463 return sprintf(page, "%u\n", sched_mc_power_savings);
7465 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7466 const char *buf, size_t count)
7468 return sched_power_savings_store(buf, count, 0);
7470 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7471 sched_mc_power_savings_store);
7474 #ifdef CONFIG_SCHED_SMT
7475 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7477 return sprintf(page, "%u\n", sched_smt_power_savings);
7479 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7480 const char *buf, size_t count)
7482 return sched_power_savings_store(buf, count, 1);
7484 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7485 sched_smt_power_savings_store);
7488 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7492 #ifdef CONFIG_SCHED_SMT
7494 err = sysfs_create_file(&cls->kset.kobj,
7495 &attr_sched_smt_power_savings.attr);
7497 #ifdef CONFIG_SCHED_MC
7498 if (!err && mc_capable())
7499 err = sysfs_create_file(&cls->kset.kobj,
7500 &attr_sched_mc_power_savings.attr);
7507 * Force a reinitialization of the sched domains hierarchy. The domains
7508 * and groups cannot be updated in place without racing with the balancing
7509 * code, so we temporarily attach all running cpus to the NULL domain
7510 * which will prevent rebalancing while the sched domains are recalculated.
7512 static int update_sched_domains(struct notifier_block *nfb,
7513 unsigned long action, void *hcpu)
7516 case CPU_UP_PREPARE:
7517 case CPU_UP_PREPARE_FROZEN:
7518 case CPU_DOWN_PREPARE:
7519 case CPU_DOWN_PREPARE_FROZEN:
7520 detach_destroy_domains(&cpu_online_map);
7523 case CPU_UP_CANCELED:
7524 case CPU_UP_CANCELED_FROZEN:
7525 case CPU_DOWN_FAILED:
7526 case CPU_DOWN_FAILED_FROZEN:
7528 case CPU_ONLINE_FROZEN:
7530 case CPU_DEAD_FROZEN:
7532 * Fall through and re-initialise the domains.
7539 /* The hotplug lock is already held by cpu_up/cpu_down */
7540 arch_init_sched_domains(&cpu_online_map);
7545 void __init sched_init_smp(void)
7547 cpumask_t non_isolated_cpus;
7549 #if defined(CONFIG_NUMA)
7550 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7552 BUG_ON(sched_group_nodes_bycpu == NULL);
7555 mutex_lock(&sched_domains_mutex);
7556 arch_init_sched_domains(&cpu_online_map);
7557 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7558 if (cpus_empty(non_isolated_cpus))
7559 cpu_set(smp_processor_id(), non_isolated_cpus);
7560 mutex_unlock(&sched_domains_mutex);
7562 /* XXX: Theoretical race here - CPU may be hotplugged now */
7563 hotcpu_notifier(update_sched_domains, 0);
7566 /* Move init over to a non-isolated CPU */
7567 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7569 sched_init_granularity();
7572 void __init sched_init_smp(void)
7574 sched_init_granularity();
7576 #endif /* CONFIG_SMP */
7578 int in_sched_functions(unsigned long addr)
7580 return in_lock_functions(addr) ||
7581 (addr >= (unsigned long)__sched_text_start
7582 && addr < (unsigned long)__sched_text_end);
7585 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7587 cfs_rq->tasks_timeline = RB_ROOT;
7588 INIT_LIST_HEAD(&cfs_rq->tasks);
7589 #ifdef CONFIG_FAIR_GROUP_SCHED
7592 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7595 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7597 struct rt_prio_array *array;
7600 array = &rt_rq->active;
7601 for (i = 0; i < MAX_RT_PRIO; i++) {
7602 INIT_LIST_HEAD(array->queue + i);
7603 __clear_bit(i, array->bitmap);
7605 /* delimiter for bitsearch: */
7606 __set_bit(MAX_RT_PRIO, array->bitmap);
7608 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7609 rt_rq->highest_prio = MAX_RT_PRIO;
7612 rt_rq->rt_nr_migratory = 0;
7613 rt_rq->overloaded = 0;
7617 rt_rq->rt_throttled = 0;
7618 rt_rq->rt_runtime = 0;
7619 spin_lock_init(&rt_rq->rt_runtime_lock);
7621 #ifdef CONFIG_RT_GROUP_SCHED
7622 rt_rq->rt_nr_boosted = 0;
7627 #ifdef CONFIG_FAIR_GROUP_SCHED
7628 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7629 struct sched_entity *se, int cpu, int add,
7630 struct sched_entity *parent)
7632 struct rq *rq = cpu_rq(cpu);
7633 tg->cfs_rq[cpu] = cfs_rq;
7634 init_cfs_rq(cfs_rq, rq);
7637 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7640 /* se could be NULL for init_task_group */
7645 se->cfs_rq = &rq->cfs;
7647 se->cfs_rq = parent->my_q;
7650 se->load.weight = tg->shares;
7651 se->load.inv_weight = 0;
7652 se->parent = parent;
7656 #ifdef CONFIG_RT_GROUP_SCHED
7657 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7658 struct sched_rt_entity *rt_se, int cpu, int add,
7659 struct sched_rt_entity *parent)
7661 struct rq *rq = cpu_rq(cpu);
7663 tg->rt_rq[cpu] = rt_rq;
7664 init_rt_rq(rt_rq, rq);
7666 rt_rq->rt_se = rt_se;
7667 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7669 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7671 tg->rt_se[cpu] = rt_se;
7676 rt_se->rt_rq = &rq->rt;
7678 rt_se->rt_rq = parent->my_q;
7680 rt_se->rt_rq = &rq->rt;
7681 rt_se->my_q = rt_rq;
7682 rt_se->parent = parent;
7683 INIT_LIST_HEAD(&rt_se->run_list);
7687 void __init sched_init(void)
7690 unsigned long alloc_size = 0, ptr;
7692 #ifdef CONFIG_FAIR_GROUP_SCHED
7693 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7695 #ifdef CONFIG_RT_GROUP_SCHED
7696 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7698 #ifdef CONFIG_USER_SCHED
7702 * As sched_init() is called before page_alloc is setup,
7703 * we use alloc_bootmem().
7706 ptr = (unsigned long)alloc_bootmem(alloc_size);
7708 #ifdef CONFIG_FAIR_GROUP_SCHED
7709 init_task_group.se = (struct sched_entity **)ptr;
7710 ptr += nr_cpu_ids * sizeof(void **);
7712 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7713 ptr += nr_cpu_ids * sizeof(void **);
7715 #ifdef CONFIG_USER_SCHED
7716 root_task_group.se = (struct sched_entity **)ptr;
7717 ptr += nr_cpu_ids * sizeof(void **);
7719 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7720 ptr += nr_cpu_ids * sizeof(void **);
7723 #ifdef CONFIG_RT_GROUP_SCHED
7724 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7725 ptr += nr_cpu_ids * sizeof(void **);
7727 init_task_group.rt_rq = (struct rt_rq **)ptr;
7728 ptr += nr_cpu_ids * sizeof(void **);
7730 #ifdef CONFIG_USER_SCHED
7731 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7732 ptr += nr_cpu_ids * sizeof(void **);
7734 root_task_group.rt_rq = (struct rt_rq **)ptr;
7735 ptr += nr_cpu_ids * sizeof(void **);
7741 init_defrootdomain();
7744 init_rt_bandwidth(&def_rt_bandwidth,
7745 global_rt_period(), global_rt_runtime());
7747 #ifdef CONFIG_RT_GROUP_SCHED
7748 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7749 global_rt_period(), global_rt_runtime());
7750 #ifdef CONFIG_USER_SCHED
7751 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7752 global_rt_period(), RUNTIME_INF);
7756 #ifdef CONFIG_GROUP_SCHED
7757 list_add(&init_task_group.list, &task_groups);
7758 INIT_LIST_HEAD(&init_task_group.children);
7760 #ifdef CONFIG_USER_SCHED
7761 INIT_LIST_HEAD(&root_task_group.children);
7762 init_task_group.parent = &root_task_group;
7763 list_add(&init_task_group.siblings, &root_task_group.children);
7767 for_each_possible_cpu(i) {
7771 spin_lock_init(&rq->lock);
7772 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7774 init_cfs_rq(&rq->cfs, rq);
7775 init_rt_rq(&rq->rt, rq);
7776 #ifdef CONFIG_FAIR_GROUP_SCHED
7777 init_task_group.shares = init_task_group_load;
7778 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7779 #ifdef CONFIG_CGROUP_SCHED
7781 * How much cpu bandwidth does init_task_group get?
7783 * In case of task-groups formed thr' the cgroup filesystem, it
7784 * gets 100% of the cpu resources in the system. This overall
7785 * system cpu resource is divided among the tasks of
7786 * init_task_group and its child task-groups in a fair manner,
7787 * based on each entity's (task or task-group's) weight
7788 * (se->load.weight).
7790 * In other words, if init_task_group has 10 tasks of weight
7791 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7792 * then A0's share of the cpu resource is:
7794 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7796 * We achieve this by letting init_task_group's tasks sit
7797 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7799 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7800 #elif defined CONFIG_USER_SCHED
7801 root_task_group.shares = NICE_0_LOAD;
7802 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
7804 * In case of task-groups formed thr' the user id of tasks,
7805 * init_task_group represents tasks belonging to root user.
7806 * Hence it forms a sibling of all subsequent groups formed.
7807 * In this case, init_task_group gets only a fraction of overall
7808 * system cpu resource, based on the weight assigned to root
7809 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7810 * by letting tasks of init_task_group sit in a separate cfs_rq
7811 * (init_cfs_rq) and having one entity represent this group of
7812 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7814 init_tg_cfs_entry(&init_task_group,
7815 &per_cpu(init_cfs_rq, i),
7816 &per_cpu(init_sched_entity, i), i, 1,
7817 root_task_group.se[i]);
7820 #endif /* CONFIG_FAIR_GROUP_SCHED */
7822 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7823 #ifdef CONFIG_RT_GROUP_SCHED
7824 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7825 #ifdef CONFIG_CGROUP_SCHED
7826 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7827 #elif defined CONFIG_USER_SCHED
7828 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
7829 init_tg_rt_entry(&init_task_group,
7830 &per_cpu(init_rt_rq, i),
7831 &per_cpu(init_sched_rt_entity, i), i, 1,
7832 root_task_group.rt_se[i]);
7836 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7837 rq->cpu_load[j] = 0;
7841 rq->active_balance = 0;
7842 rq->next_balance = jiffies;
7845 rq->migration_thread = NULL;
7846 INIT_LIST_HEAD(&rq->migration_queue);
7847 rq_attach_root(rq, &def_root_domain);
7850 atomic_set(&rq->nr_iowait, 0);
7853 set_load_weight(&init_task);
7855 #ifdef CONFIG_PREEMPT_NOTIFIERS
7856 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7860 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7863 #ifdef CONFIG_RT_MUTEXES
7864 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7868 * The boot idle thread does lazy MMU switching as well:
7870 atomic_inc(&init_mm.mm_count);
7871 enter_lazy_tlb(&init_mm, current);
7874 * Make us the idle thread. Technically, schedule() should not be
7875 * called from this thread, however somewhere below it might be,
7876 * but because we are the idle thread, we just pick up running again
7877 * when this runqueue becomes "idle".
7879 init_idle(current, smp_processor_id());
7881 * During early bootup we pretend to be a normal task:
7883 current->sched_class = &fair_sched_class;
7885 scheduler_running = 1;
7888 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7889 void __might_sleep(char *file, int line)
7892 static unsigned long prev_jiffy; /* ratelimiting */
7894 if ((in_atomic() || irqs_disabled()) &&
7895 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7896 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7898 prev_jiffy = jiffies;
7899 printk(KERN_ERR "BUG: sleeping function called from invalid"
7900 " context at %s:%d\n", file, line);
7901 printk("in_atomic():%d, irqs_disabled():%d\n",
7902 in_atomic(), irqs_disabled());
7903 debug_show_held_locks(current);
7904 if (irqs_disabled())
7905 print_irqtrace_events(current);
7910 EXPORT_SYMBOL(__might_sleep);
7913 #ifdef CONFIG_MAGIC_SYSRQ
7914 static void normalize_task(struct rq *rq, struct task_struct *p)
7918 update_rq_clock(rq);
7919 on_rq = p->se.on_rq;
7921 deactivate_task(rq, p, 0);
7922 __setscheduler(rq, p, SCHED_NORMAL, 0);
7924 activate_task(rq, p, 0);
7925 resched_task(rq->curr);
7929 void normalize_rt_tasks(void)
7931 struct task_struct *g, *p;
7932 unsigned long flags;
7935 read_lock_irqsave(&tasklist_lock, flags);
7936 do_each_thread(g, p) {
7938 * Only normalize user tasks:
7943 p->se.exec_start = 0;
7944 #ifdef CONFIG_SCHEDSTATS
7945 p->se.wait_start = 0;
7946 p->se.sleep_start = 0;
7947 p->se.block_start = 0;
7952 * Renice negative nice level userspace
7955 if (TASK_NICE(p) < 0 && p->mm)
7956 set_user_nice(p, 0);
7960 spin_lock(&p->pi_lock);
7961 rq = __task_rq_lock(p);
7963 normalize_task(rq, p);
7965 __task_rq_unlock(rq);
7966 spin_unlock(&p->pi_lock);
7967 } while_each_thread(g, p);
7969 read_unlock_irqrestore(&tasklist_lock, flags);
7972 #endif /* CONFIG_MAGIC_SYSRQ */
7976 * These functions are only useful for the IA64 MCA handling.
7978 * They can only be called when the whole system has been
7979 * stopped - every CPU needs to be quiescent, and no scheduling
7980 * activity can take place. Using them for anything else would
7981 * be a serious bug, and as a result, they aren't even visible
7982 * under any other configuration.
7986 * curr_task - return the current task for a given cpu.
7987 * @cpu: the processor in question.
7989 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7991 struct task_struct *curr_task(int cpu)
7993 return cpu_curr(cpu);
7997 * set_curr_task - set the current task for a given cpu.
7998 * @cpu: the processor in question.
7999 * @p: the task pointer to set.
8001 * Description: This function must only be used when non-maskable interrupts
8002 * are serviced on a separate stack. It allows the architecture to switch the
8003 * notion of the current task on a cpu in a non-blocking manner. This function
8004 * must be called with all CPU's synchronized, and interrupts disabled, the
8005 * and caller must save the original value of the current task (see
8006 * curr_task() above) and restore that value before reenabling interrupts and
8007 * re-starting the system.
8009 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8011 void set_curr_task(int cpu, struct task_struct *p)
8018 #ifdef CONFIG_FAIR_GROUP_SCHED
8019 static void free_fair_sched_group(struct task_group *tg)
8023 for_each_possible_cpu(i) {
8025 kfree(tg->cfs_rq[i]);
8035 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8037 struct cfs_rq *cfs_rq;
8038 struct sched_entity *se, *parent_se;
8042 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8045 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8049 tg->shares = NICE_0_LOAD;
8051 for_each_possible_cpu(i) {
8054 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8055 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8059 se = kmalloc_node(sizeof(struct sched_entity),
8060 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8064 parent_se = parent ? parent->se[i] : NULL;
8065 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8074 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8076 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8077 &cpu_rq(cpu)->leaf_cfs_rq_list);
8080 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8082 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8085 static inline void free_fair_sched_group(struct task_group *tg)
8090 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8095 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8099 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8104 #ifdef CONFIG_RT_GROUP_SCHED
8105 static void free_rt_sched_group(struct task_group *tg)
8109 destroy_rt_bandwidth(&tg->rt_bandwidth);
8111 for_each_possible_cpu(i) {
8113 kfree(tg->rt_rq[i]);
8115 kfree(tg->rt_se[i]);
8123 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8125 struct rt_rq *rt_rq;
8126 struct sched_rt_entity *rt_se, *parent_se;
8130 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8133 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8137 init_rt_bandwidth(&tg->rt_bandwidth,
8138 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8140 for_each_possible_cpu(i) {
8143 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8144 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8148 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8149 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8153 parent_se = parent ? parent->rt_se[i] : NULL;
8154 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8163 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8165 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8166 &cpu_rq(cpu)->leaf_rt_rq_list);
8169 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8171 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8174 static inline void free_rt_sched_group(struct task_group *tg)
8179 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8184 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8188 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8193 #ifdef CONFIG_GROUP_SCHED
8194 static void free_sched_group(struct task_group *tg)
8196 free_fair_sched_group(tg);
8197 free_rt_sched_group(tg);
8201 /* allocate runqueue etc for a new task group */
8202 struct task_group *sched_create_group(struct task_group *parent)
8204 struct task_group *tg;
8205 unsigned long flags;
8208 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8210 return ERR_PTR(-ENOMEM);
8212 if (!alloc_fair_sched_group(tg, parent))
8215 if (!alloc_rt_sched_group(tg, parent))
8218 spin_lock_irqsave(&task_group_lock, flags);
8219 for_each_possible_cpu(i) {
8220 register_fair_sched_group(tg, i);
8221 register_rt_sched_group(tg, i);
8223 list_add_rcu(&tg->list, &task_groups);
8225 WARN_ON(!parent); /* root should already exist */
8227 tg->parent = parent;
8228 list_add_rcu(&tg->siblings, &parent->children);
8229 INIT_LIST_HEAD(&tg->children);
8230 spin_unlock_irqrestore(&task_group_lock, flags);
8235 free_sched_group(tg);
8236 return ERR_PTR(-ENOMEM);
8239 /* rcu callback to free various structures associated with a task group */
8240 static void free_sched_group_rcu(struct rcu_head *rhp)
8242 /* now it should be safe to free those cfs_rqs */
8243 free_sched_group(container_of(rhp, struct task_group, rcu));
8246 /* Destroy runqueue etc associated with a task group */
8247 void sched_destroy_group(struct task_group *tg)
8249 unsigned long flags;
8252 spin_lock_irqsave(&task_group_lock, flags);
8253 for_each_possible_cpu(i) {
8254 unregister_fair_sched_group(tg, i);
8255 unregister_rt_sched_group(tg, i);
8257 list_del_rcu(&tg->list);
8258 list_del_rcu(&tg->siblings);
8259 spin_unlock_irqrestore(&task_group_lock, flags);
8261 /* wait for possible concurrent references to cfs_rqs complete */
8262 call_rcu(&tg->rcu, free_sched_group_rcu);
8265 /* change task's runqueue when it moves between groups.
8266 * The caller of this function should have put the task in its new group
8267 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8268 * reflect its new group.
8270 void sched_move_task(struct task_struct *tsk)
8273 unsigned long flags;
8276 rq = task_rq_lock(tsk, &flags);
8278 update_rq_clock(rq);
8280 running = task_current(rq, tsk);
8281 on_rq = tsk->se.on_rq;
8284 dequeue_task(rq, tsk, 0);
8285 if (unlikely(running))
8286 tsk->sched_class->put_prev_task(rq, tsk);
8288 set_task_rq(tsk, task_cpu(tsk));
8290 #ifdef CONFIG_FAIR_GROUP_SCHED
8291 if (tsk->sched_class->moved_group)
8292 tsk->sched_class->moved_group(tsk);
8295 if (unlikely(running))
8296 tsk->sched_class->set_curr_task(rq);
8298 enqueue_task(rq, tsk, 0);
8300 task_rq_unlock(rq, &flags);
8304 #ifdef CONFIG_FAIR_GROUP_SCHED
8305 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8307 struct cfs_rq *cfs_rq = se->cfs_rq;
8308 struct rq *rq = cfs_rq->rq;
8311 spin_lock_irq(&rq->lock);
8315 dequeue_entity(cfs_rq, se, 0);
8317 se->load.weight = shares;
8318 se->load.inv_weight = 0;
8321 enqueue_entity(cfs_rq, se, 0);
8323 spin_unlock_irq(&rq->lock);
8326 static DEFINE_MUTEX(shares_mutex);
8328 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8331 unsigned long flags;
8334 * We can't change the weight of the root cgroup.
8339 if (shares < MIN_SHARES)
8340 shares = MIN_SHARES;
8341 else if (shares > MAX_SHARES)
8342 shares = MAX_SHARES;
8344 mutex_lock(&shares_mutex);
8345 if (tg->shares == shares)
8348 spin_lock_irqsave(&task_group_lock, flags);
8349 for_each_possible_cpu(i)
8350 unregister_fair_sched_group(tg, i);
8351 list_del_rcu(&tg->siblings);
8352 spin_unlock_irqrestore(&task_group_lock, flags);
8354 /* wait for any ongoing reference to this group to finish */
8355 synchronize_sched();
8358 * Now we are free to modify the group's share on each cpu
8359 * w/o tripping rebalance_share or load_balance_fair.
8361 tg->shares = shares;
8362 for_each_possible_cpu(i)
8363 set_se_shares(tg->se[i], shares);
8366 * Enable load balance activity on this group, by inserting it back on
8367 * each cpu's rq->leaf_cfs_rq_list.
8369 spin_lock_irqsave(&task_group_lock, flags);
8370 for_each_possible_cpu(i)
8371 register_fair_sched_group(tg, i);
8372 list_add_rcu(&tg->siblings, &tg->parent->children);
8373 spin_unlock_irqrestore(&task_group_lock, flags);
8375 mutex_unlock(&shares_mutex);
8379 unsigned long sched_group_shares(struct task_group *tg)
8385 #ifdef CONFIG_RT_GROUP_SCHED
8387 * Ensure that the real time constraints are schedulable.
8389 static DEFINE_MUTEX(rt_constraints_mutex);
8391 static unsigned long to_ratio(u64 period, u64 runtime)
8393 if (runtime == RUNTIME_INF)
8396 return div64_u64(runtime << 16, period);
8399 #ifdef CONFIG_CGROUP_SCHED
8400 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8402 struct task_group *tgi, *parent = tg->parent;
8403 unsigned long total = 0;
8406 if (global_rt_period() < period)
8409 return to_ratio(period, runtime) <
8410 to_ratio(global_rt_period(), global_rt_runtime());
8413 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8417 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8421 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8422 tgi->rt_bandwidth.rt_runtime);
8426 return total + to_ratio(period, runtime) <
8427 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8428 parent->rt_bandwidth.rt_runtime);
8430 #elif defined CONFIG_USER_SCHED
8431 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8433 struct task_group *tgi;
8434 unsigned long total = 0;
8435 unsigned long global_ratio =
8436 to_ratio(global_rt_period(), global_rt_runtime());
8439 list_for_each_entry_rcu(tgi, &task_groups, list) {
8443 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8444 tgi->rt_bandwidth.rt_runtime);
8448 return total + to_ratio(period, runtime) < global_ratio;
8452 /* Must be called with tasklist_lock held */
8453 static inline int tg_has_rt_tasks(struct task_group *tg)
8455 struct task_struct *g, *p;
8456 do_each_thread(g, p) {
8457 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8459 } while_each_thread(g, p);
8463 static int tg_set_bandwidth(struct task_group *tg,
8464 u64 rt_period, u64 rt_runtime)
8468 mutex_lock(&rt_constraints_mutex);
8469 read_lock(&tasklist_lock);
8470 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8474 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8479 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8480 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8481 tg->rt_bandwidth.rt_runtime = rt_runtime;
8483 for_each_possible_cpu(i) {
8484 struct rt_rq *rt_rq = tg->rt_rq[i];
8486 spin_lock(&rt_rq->rt_runtime_lock);
8487 rt_rq->rt_runtime = rt_runtime;
8488 spin_unlock(&rt_rq->rt_runtime_lock);
8490 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8492 read_unlock(&tasklist_lock);
8493 mutex_unlock(&rt_constraints_mutex);
8498 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8500 u64 rt_runtime, rt_period;
8502 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8503 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8504 if (rt_runtime_us < 0)
8505 rt_runtime = RUNTIME_INF;
8507 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8510 long sched_group_rt_runtime(struct task_group *tg)
8514 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8517 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8518 do_div(rt_runtime_us, NSEC_PER_USEC);
8519 return rt_runtime_us;
8522 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8524 u64 rt_runtime, rt_period;
8526 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8527 rt_runtime = tg->rt_bandwidth.rt_runtime;
8529 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8532 long sched_group_rt_period(struct task_group *tg)
8536 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8537 do_div(rt_period_us, NSEC_PER_USEC);
8538 return rt_period_us;
8541 static int sched_rt_global_constraints(void)
8545 mutex_lock(&rt_constraints_mutex);
8546 if (!__rt_schedulable(NULL, 1, 0))
8548 mutex_unlock(&rt_constraints_mutex);
8553 static int sched_rt_global_constraints(void)
8555 unsigned long flags;
8558 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8559 for_each_possible_cpu(i) {
8560 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8562 spin_lock(&rt_rq->rt_runtime_lock);
8563 rt_rq->rt_runtime = global_rt_runtime();
8564 spin_unlock(&rt_rq->rt_runtime_lock);
8566 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8572 int sched_rt_handler(struct ctl_table *table, int write,
8573 struct file *filp, void __user *buffer, size_t *lenp,
8577 int old_period, old_runtime;
8578 static DEFINE_MUTEX(mutex);
8581 old_period = sysctl_sched_rt_period;
8582 old_runtime = sysctl_sched_rt_runtime;
8584 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8586 if (!ret && write) {
8587 ret = sched_rt_global_constraints();
8589 sysctl_sched_rt_period = old_period;
8590 sysctl_sched_rt_runtime = old_runtime;
8592 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8593 def_rt_bandwidth.rt_period =
8594 ns_to_ktime(global_rt_period());
8597 mutex_unlock(&mutex);
8602 #ifdef CONFIG_CGROUP_SCHED
8604 /* return corresponding task_group object of a cgroup */
8605 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8607 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8608 struct task_group, css);
8611 static struct cgroup_subsys_state *
8612 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8614 struct task_group *tg, *parent;
8616 if (!cgrp->parent) {
8617 /* This is early initialization for the top cgroup */
8618 init_task_group.css.cgroup = cgrp;
8619 return &init_task_group.css;
8622 parent = cgroup_tg(cgrp->parent);
8623 tg = sched_create_group(parent);
8625 return ERR_PTR(-ENOMEM);
8627 /* Bind the cgroup to task_group object we just created */
8628 tg->css.cgroup = cgrp;
8634 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8636 struct task_group *tg = cgroup_tg(cgrp);
8638 sched_destroy_group(tg);
8642 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8643 struct task_struct *tsk)
8645 #ifdef CONFIG_RT_GROUP_SCHED
8646 /* Don't accept realtime tasks when there is no way for them to run */
8647 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8650 /* We don't support RT-tasks being in separate groups */
8651 if (tsk->sched_class != &fair_sched_class)
8659 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8660 struct cgroup *old_cont, struct task_struct *tsk)
8662 sched_move_task(tsk);
8665 #ifdef CONFIG_FAIR_GROUP_SCHED
8666 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8669 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8672 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8674 struct task_group *tg = cgroup_tg(cgrp);
8676 return (u64) tg->shares;
8680 #ifdef CONFIG_RT_GROUP_SCHED
8681 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8684 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8687 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8689 return sched_group_rt_runtime(cgroup_tg(cgrp));
8692 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8695 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8698 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8700 return sched_group_rt_period(cgroup_tg(cgrp));
8704 static struct cftype cpu_files[] = {
8705 #ifdef CONFIG_FAIR_GROUP_SCHED
8708 .read_u64 = cpu_shares_read_u64,
8709 .write_u64 = cpu_shares_write_u64,
8712 #ifdef CONFIG_RT_GROUP_SCHED
8714 .name = "rt_runtime_us",
8715 .read_s64 = cpu_rt_runtime_read,
8716 .write_s64 = cpu_rt_runtime_write,
8719 .name = "rt_period_us",
8720 .read_u64 = cpu_rt_period_read_uint,
8721 .write_u64 = cpu_rt_period_write_uint,
8726 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8728 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8731 struct cgroup_subsys cpu_cgroup_subsys = {
8733 .create = cpu_cgroup_create,
8734 .destroy = cpu_cgroup_destroy,
8735 .can_attach = cpu_cgroup_can_attach,
8736 .attach = cpu_cgroup_attach,
8737 .populate = cpu_cgroup_populate,
8738 .subsys_id = cpu_cgroup_subsys_id,
8742 #endif /* CONFIG_CGROUP_SCHED */
8744 #ifdef CONFIG_CGROUP_CPUACCT
8747 * CPU accounting code for task groups.
8749 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8750 * (balbir@in.ibm.com).
8753 /* track cpu usage of a group of tasks */
8755 struct cgroup_subsys_state css;
8756 /* cpuusage holds pointer to a u64-type object on every cpu */
8760 struct cgroup_subsys cpuacct_subsys;
8762 /* return cpu accounting group corresponding to this container */
8763 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8765 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8766 struct cpuacct, css);
8769 /* return cpu accounting group to which this task belongs */
8770 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8772 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8773 struct cpuacct, css);
8776 /* create a new cpu accounting group */
8777 static struct cgroup_subsys_state *cpuacct_create(
8778 struct cgroup_subsys *ss, struct cgroup *cgrp)
8780 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8783 return ERR_PTR(-ENOMEM);
8785 ca->cpuusage = alloc_percpu(u64);
8786 if (!ca->cpuusage) {
8788 return ERR_PTR(-ENOMEM);
8794 /* destroy an existing cpu accounting group */
8796 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8798 struct cpuacct *ca = cgroup_ca(cgrp);
8800 free_percpu(ca->cpuusage);
8804 /* return total cpu usage (in nanoseconds) of a group */
8805 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8807 struct cpuacct *ca = cgroup_ca(cgrp);
8808 u64 totalcpuusage = 0;
8811 for_each_possible_cpu(i) {
8812 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8815 * Take rq->lock to make 64-bit addition safe on 32-bit
8818 spin_lock_irq(&cpu_rq(i)->lock);
8819 totalcpuusage += *cpuusage;
8820 spin_unlock_irq(&cpu_rq(i)->lock);
8823 return totalcpuusage;
8826 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8829 struct cpuacct *ca = cgroup_ca(cgrp);
8838 for_each_possible_cpu(i) {
8839 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8841 spin_lock_irq(&cpu_rq(i)->lock);
8843 spin_unlock_irq(&cpu_rq(i)->lock);
8849 static struct cftype files[] = {
8852 .read_u64 = cpuusage_read,
8853 .write_u64 = cpuusage_write,
8857 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8859 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8863 * charge this task's execution time to its accounting group.
8865 * called with rq->lock held.
8867 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8871 if (!cpuacct_subsys.active)
8876 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8878 *cpuusage += cputime;
8882 struct cgroup_subsys cpuacct_subsys = {
8884 .create = cpuacct_create,
8885 .destroy = cpuacct_destroy,
8886 .populate = cpuacct_populate,
8887 .subsys_id = cpuacct_subsys_id,
8889 #endif /* CONFIG_CGROUP_CPUACCT */