Merge branch 'merge' of git://git.kernel.org/pub/scm/linux/kernel/git/paulus/powerpc
[linux-2.6/lfs.git] / kernel / sched.c
blobcfa222a91539e193a54d20d422fa92b43a19334b
1 /*
2 * kernel/sched.c
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
11 * by Andrea Arcangeli
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
22 * by Peter Williams
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
29 #include <linux/mm.h>
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>
74 #include <asm/tlb.h>
75 #include <asm/irq_regs.h>
78 * Convert user-nice values [ -20 ... 0 ... 19 ]
79 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
80 * and back.
82 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
83 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
84 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
87 * 'User priority' is the nice value converted to something we
88 * can work with better when scaling various scheduler parameters,
89 * it's a [ 0 ... 39 ] range.
91 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
92 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
93 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
96 * Helpers for converting nanosecond timing to jiffy resolution
98 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
100 #define NICE_0_LOAD SCHED_LOAD_SCALE
101 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
104 * These are the 'tuning knobs' of the scheduler:
106 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
107 * Timeslices get refilled after they expire.
109 #define DEF_TIMESLICE (100 * HZ / 1000)
112 * single value that denotes runtime == period, ie unlimited time.
114 #define RUNTIME_INF ((u64)~0ULL)
116 #ifdef CONFIG_SMP
118 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
119 * Since cpu_power is a 'constant', we can use a reciprocal divide.
121 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
123 return reciprocal_divide(load, sg->reciprocal_cpu_power);
127 * Each time a sched group cpu_power is changed,
128 * we must compute its reciprocal value
130 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
132 sg->__cpu_power += val;
133 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
135 #endif
137 static inline int rt_policy(int policy)
139 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
140 return 1;
141 return 0;
144 static inline int task_has_rt_policy(struct task_struct *p)
146 return rt_policy(p->policy);
150 * This is the priority-queue data structure of the RT scheduling class:
152 struct rt_prio_array {
153 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
154 struct list_head queue[MAX_RT_PRIO];
157 struct rt_bandwidth {
158 /* nests inside the rq lock: */
159 spinlock_t rt_runtime_lock;
160 ktime_t rt_period;
161 u64 rt_runtime;
162 struct hrtimer rt_period_timer;
165 static struct rt_bandwidth def_rt_bandwidth;
167 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
169 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
171 struct rt_bandwidth *rt_b =
172 container_of(timer, struct rt_bandwidth, rt_period_timer);
173 ktime_t now;
174 int overrun;
175 int idle = 0;
177 for (;;) {
178 now = hrtimer_cb_get_time(timer);
179 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
181 if (!overrun)
182 break;
184 idle = do_sched_rt_period_timer(rt_b, overrun);
187 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
190 static
191 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
193 rt_b->rt_period = ns_to_ktime(period);
194 rt_b->rt_runtime = runtime;
196 spin_lock_init(&rt_b->rt_runtime_lock);
198 hrtimer_init(&rt_b->rt_period_timer,
199 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
200 rt_b->rt_period_timer.function = sched_rt_period_timer;
201 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
204 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
206 ktime_t now;
208 if (rt_b->rt_runtime == RUNTIME_INF)
209 return;
211 if (hrtimer_active(&rt_b->rt_period_timer))
212 return;
214 spin_lock(&rt_b->rt_runtime_lock);
215 for (;;) {
216 if (hrtimer_active(&rt_b->rt_period_timer))
217 break;
219 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
220 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
221 hrtimer_start(&rt_b->rt_period_timer,
222 rt_b->rt_period_timer.expires,
223 HRTIMER_MODE_ABS);
225 spin_unlock(&rt_b->rt_runtime_lock);
228 #ifdef CONFIG_RT_GROUP_SCHED
229 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
231 hrtimer_cancel(&rt_b->rt_period_timer);
233 #endif
236 * sched_domains_mutex serializes calls to arch_init_sched_domains,
237 * detach_destroy_domains and partition_sched_domains.
239 static DEFINE_MUTEX(sched_domains_mutex);
241 #ifdef CONFIG_GROUP_SCHED
243 #include <linux/cgroup.h>
245 struct cfs_rq;
247 static LIST_HEAD(task_groups);
249 /* task group related information */
250 struct task_group {
251 #ifdef CONFIG_CGROUP_SCHED
252 struct cgroup_subsys_state css;
253 #endif
255 #ifdef CONFIG_FAIR_GROUP_SCHED
256 /* schedulable entities of this group on each cpu */
257 struct sched_entity **se;
258 /* runqueue "owned" by this group on each cpu */
259 struct cfs_rq **cfs_rq;
260 unsigned long shares;
261 #endif
263 #ifdef CONFIG_RT_GROUP_SCHED
264 struct sched_rt_entity **rt_se;
265 struct rt_rq **rt_rq;
267 struct rt_bandwidth rt_bandwidth;
268 #endif
270 struct rcu_head rcu;
271 struct list_head list;
273 struct task_group *parent;
274 struct list_head siblings;
275 struct list_head children;
278 #ifdef CONFIG_USER_SCHED
281 * Root task group.
282 * Every UID task group (including init_task_group aka UID-0) will
283 * be a child to this group.
285 struct task_group root_task_group;
287 #ifdef CONFIG_FAIR_GROUP_SCHED
288 /* Default task group's sched entity on each cpu */
289 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
290 /* Default task group's cfs_rq on each cpu */
291 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
292 #endif
294 #ifdef CONFIG_RT_GROUP_SCHED
295 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
296 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
297 #endif
298 #else
299 #define root_task_group init_task_group
300 #endif
302 /* task_group_lock serializes add/remove of task groups and also changes to
303 * a task group's cpu shares.
305 static DEFINE_SPINLOCK(task_group_lock);
307 #ifdef CONFIG_FAIR_GROUP_SCHED
308 #ifdef CONFIG_USER_SCHED
309 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
310 #else
311 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
312 #endif
315 * A weight of 0, 1 or ULONG_MAX can cause arithmetics problems.
316 * (The default weight is 1024 - so there's no practical
317 * limitation from this.)
319 #define MIN_SHARES 2
320 #define MAX_SHARES (ULONG_MAX - 1)
322 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
323 #endif
325 /* Default task group.
326 * Every task in system belong to this group at bootup.
328 struct task_group init_task_group;
330 /* return group to which a task belongs */
331 static inline struct task_group *task_group(struct task_struct *p)
333 struct task_group *tg;
335 #ifdef CONFIG_USER_SCHED
336 tg = p->user->tg;
337 #elif defined(CONFIG_CGROUP_SCHED)
338 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
339 struct task_group, css);
340 #else
341 tg = &init_task_group;
342 #endif
343 return tg;
346 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
347 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
349 #ifdef CONFIG_FAIR_GROUP_SCHED
350 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
351 p->se.parent = task_group(p)->se[cpu];
352 #endif
354 #ifdef CONFIG_RT_GROUP_SCHED
355 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
356 p->rt.parent = task_group(p)->rt_se[cpu];
357 #endif
360 #else
362 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
364 #endif /* CONFIG_GROUP_SCHED */
366 /* CFS-related fields in a runqueue */
367 struct cfs_rq {
368 struct load_weight load;
369 unsigned long nr_running;
371 u64 exec_clock;
372 u64 min_vruntime;
374 struct rb_root tasks_timeline;
375 struct rb_node *rb_leftmost;
377 struct list_head tasks;
378 struct list_head *balance_iterator;
381 * 'curr' points to currently running entity on this cfs_rq.
382 * It is set to NULL otherwise (i.e when none are currently running).
384 struct sched_entity *curr, *next;
386 unsigned long nr_spread_over;
388 #ifdef CONFIG_FAIR_GROUP_SCHED
389 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
392 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
393 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
394 * (like users, containers etc.)
396 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
397 * list is used during load balance.
399 struct list_head leaf_cfs_rq_list;
400 struct task_group *tg; /* group that "owns" this runqueue */
402 #ifdef CONFIG_SMP
403 unsigned long task_weight;
404 unsigned long shares;
406 * We need space to build a sched_domain wide view of the full task
407 * group tree, in order to avoid depending on dynamic memory allocation
408 * during the load balancing we place this in the per cpu task group
409 * hierarchy. This limits the load balancing to one instance per cpu,
410 * but more should not be needed anyway.
412 struct aggregate_struct {
414 * load = weight(cpus) * f(tg)
416 * Where f(tg) is the recursive weight fraction assigned to
417 * this group.
419 unsigned long load;
422 * part of the group weight distributed to this span.
424 unsigned long shares;
427 * The sum of all runqueue weights within this span.
429 unsigned long rq_weight;
432 * Weight contributed by tasks; this is the part we can
433 * influence by moving tasks around.
435 unsigned long task_weight;
436 } aggregate;
437 #endif
438 #endif
441 /* Real-Time classes' related field in a runqueue: */
442 struct rt_rq {
443 struct rt_prio_array active;
444 unsigned long rt_nr_running;
445 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
446 int highest_prio; /* highest queued rt task prio */
447 #endif
448 #ifdef CONFIG_SMP
449 unsigned long rt_nr_migratory;
450 int overloaded;
451 #endif
452 int rt_throttled;
453 u64 rt_time;
454 u64 rt_runtime;
455 /* Nests inside the rq lock: */
456 spinlock_t rt_runtime_lock;
458 #ifdef CONFIG_RT_GROUP_SCHED
459 unsigned long rt_nr_boosted;
461 struct rq *rq;
462 struct list_head leaf_rt_rq_list;
463 struct task_group *tg;
464 struct sched_rt_entity *rt_se;
465 #endif
468 #ifdef CONFIG_SMP
471 * We add the notion of a root-domain which will be used to define per-domain
472 * variables. Each exclusive cpuset essentially defines an island domain by
473 * fully partitioning the member cpus from any other cpuset. Whenever a new
474 * exclusive cpuset is created, we also create and attach a new root-domain
475 * object.
478 struct root_domain {
479 atomic_t refcount;
480 cpumask_t span;
481 cpumask_t online;
484 * The "RT overload" flag: it gets set if a CPU has more than
485 * one runnable RT task.
487 cpumask_t rto_mask;
488 atomic_t rto_count;
492 * By default the system creates a single root-domain with all cpus as
493 * members (mimicking the global state we have today).
495 static struct root_domain def_root_domain;
497 #endif
500 * This is the main, per-CPU runqueue data structure.
502 * Locking rule: those places that want to lock multiple runqueues
503 * (such as the load balancing or the thread migration code), lock
504 * acquire operations must be ordered by ascending &runqueue.
506 struct rq {
507 /* runqueue lock: */
508 spinlock_t lock;
511 * nr_running and cpu_load should be in the same cacheline because
512 * remote CPUs use both these fields when doing load calculation.
514 unsigned long nr_running;
515 #define CPU_LOAD_IDX_MAX 5
516 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
517 unsigned char idle_at_tick;
518 #ifdef CONFIG_NO_HZ
519 unsigned long last_tick_seen;
520 unsigned char in_nohz_recently;
521 #endif
522 /* capture load from *all* tasks on this cpu: */
523 struct load_weight load;
524 unsigned long nr_load_updates;
525 u64 nr_switches;
527 struct cfs_rq cfs;
528 struct rt_rq rt;
530 #ifdef CONFIG_FAIR_GROUP_SCHED
531 /* list of leaf cfs_rq on this cpu: */
532 struct list_head leaf_cfs_rq_list;
533 #endif
534 #ifdef CONFIG_RT_GROUP_SCHED
535 struct list_head leaf_rt_rq_list;
536 #endif
539 * This is part of a global counter where only the total sum
540 * over all CPUs matters. A task can increase this counter on
541 * one CPU and if it got migrated afterwards it may decrease
542 * it on another CPU. Always updated under the runqueue lock:
544 unsigned long nr_uninterruptible;
546 struct task_struct *curr, *idle;
547 unsigned long next_balance;
548 struct mm_struct *prev_mm;
550 u64 clock;
552 atomic_t nr_iowait;
554 #ifdef CONFIG_SMP
555 struct root_domain *rd;
556 struct sched_domain *sd;
558 /* For active balancing */
559 int active_balance;
560 int push_cpu;
561 /* cpu of this runqueue: */
562 int cpu;
564 struct task_struct *migration_thread;
565 struct list_head migration_queue;
566 #endif
568 #ifdef CONFIG_SCHED_HRTICK
569 unsigned long hrtick_flags;
570 ktime_t hrtick_expire;
571 struct hrtimer hrtick_timer;
572 #endif
574 #ifdef CONFIG_SCHEDSTATS
575 /* latency stats */
576 struct sched_info rq_sched_info;
578 /* sys_sched_yield() stats */
579 unsigned int yld_exp_empty;
580 unsigned int yld_act_empty;
581 unsigned int yld_both_empty;
582 unsigned int yld_count;
584 /* schedule() stats */
585 unsigned int sched_switch;
586 unsigned int sched_count;
587 unsigned int sched_goidle;
589 /* try_to_wake_up() stats */
590 unsigned int ttwu_count;
591 unsigned int ttwu_local;
593 /* BKL stats */
594 unsigned int bkl_count;
595 #endif
596 struct lock_class_key rq_lock_key;
599 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
601 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
603 rq->curr->sched_class->check_preempt_curr(rq, p);
606 static inline int cpu_of(struct rq *rq)
608 #ifdef CONFIG_SMP
609 return rq->cpu;
610 #else
611 return 0;
612 #endif
616 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
617 * See detach_destroy_domains: synchronize_sched for details.
619 * The domain tree of any CPU may only be accessed from within
620 * preempt-disabled sections.
622 #define for_each_domain(cpu, __sd) \
623 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
625 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
626 #define this_rq() (&__get_cpu_var(runqueues))
627 #define task_rq(p) cpu_rq(task_cpu(p))
628 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
630 static inline void update_rq_clock(struct rq *rq)
632 rq->clock = sched_clock_cpu(cpu_of(rq));
636 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
638 #ifdef CONFIG_SCHED_DEBUG
639 # define const_debug __read_mostly
640 #else
641 # define const_debug static const
642 #endif
645 * Debugging: various feature bits
648 #define SCHED_FEAT(name, enabled) \
649 __SCHED_FEAT_##name ,
651 enum {
652 #include "sched_features.h"
655 #undef SCHED_FEAT
657 #define SCHED_FEAT(name, enabled) \
658 (1UL << __SCHED_FEAT_##name) * enabled |
660 const_debug unsigned int sysctl_sched_features =
661 #include "sched_features.h"
664 #undef SCHED_FEAT
666 #ifdef CONFIG_SCHED_DEBUG
667 #define SCHED_FEAT(name, enabled) \
668 #name ,
670 static __read_mostly char *sched_feat_names[] = {
671 #include "sched_features.h"
672 NULL
675 #undef SCHED_FEAT
677 static int sched_feat_open(struct inode *inode, struct file *filp)
679 filp->private_data = inode->i_private;
680 return 0;
683 static ssize_t
684 sched_feat_read(struct file *filp, char __user *ubuf,
685 size_t cnt, loff_t *ppos)
687 char *buf;
688 int r = 0;
689 int len = 0;
690 int i;
692 for (i = 0; sched_feat_names[i]; i++) {
693 len += strlen(sched_feat_names[i]);
694 len += 4;
697 buf = kmalloc(len + 2, GFP_KERNEL);
698 if (!buf)
699 return -ENOMEM;
701 for (i = 0; sched_feat_names[i]; i++) {
702 if (sysctl_sched_features & (1UL << i))
703 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
704 else
705 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
708 r += sprintf(buf + r, "\n");
709 WARN_ON(r >= len + 2);
711 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
713 kfree(buf);
715 return r;
718 static ssize_t
719 sched_feat_write(struct file *filp, const char __user *ubuf,
720 size_t cnt, loff_t *ppos)
722 char buf[64];
723 char *cmp = buf;
724 int neg = 0;
725 int i;
727 if (cnt > 63)
728 cnt = 63;
730 if (copy_from_user(&buf, ubuf, cnt))
731 return -EFAULT;
733 buf[cnt] = 0;
735 if (strncmp(buf, "NO_", 3) == 0) {
736 neg = 1;
737 cmp += 3;
740 for (i = 0; sched_feat_names[i]; i++) {
741 int len = strlen(sched_feat_names[i]);
743 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
744 if (neg)
745 sysctl_sched_features &= ~(1UL << i);
746 else
747 sysctl_sched_features |= (1UL << i);
748 break;
752 if (!sched_feat_names[i])
753 return -EINVAL;
755 filp->f_pos += cnt;
757 return cnt;
760 static struct file_operations sched_feat_fops = {
761 .open = sched_feat_open,
762 .read = sched_feat_read,
763 .write = sched_feat_write,
766 static __init int sched_init_debug(void)
768 debugfs_create_file("sched_features", 0644, NULL, NULL,
769 &sched_feat_fops);
771 return 0;
773 late_initcall(sched_init_debug);
775 #endif
777 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
780 * Number of tasks to iterate in a single balance run.
781 * Limited because this is done with IRQs disabled.
783 const_debug unsigned int sysctl_sched_nr_migrate = 32;
786 * period over which we measure -rt task cpu usage in us.
787 * default: 1s
789 unsigned int sysctl_sched_rt_period = 1000000;
791 static __read_mostly int scheduler_running;
794 * part of the period that we allow rt tasks to run in us.
795 * default: 0.95s
797 int sysctl_sched_rt_runtime = 950000;
799 static inline u64 global_rt_period(void)
801 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
804 static inline u64 global_rt_runtime(void)
806 if (sysctl_sched_rt_period < 0)
807 return RUNTIME_INF;
809 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
812 unsigned long long time_sync_thresh = 100000;
814 static DEFINE_PER_CPU(unsigned long long, time_offset);
815 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
818 * Global lock which we take every now and then to synchronize
819 * the CPUs time. This method is not warp-safe, but it's good
820 * enough to synchronize slowly diverging time sources and thus
821 * it's good enough for tracing:
823 static DEFINE_SPINLOCK(time_sync_lock);
824 static unsigned long long prev_global_time;
826 static unsigned long long __sync_cpu_clock(unsigned long long time, int cpu)
829 * We want this inlined, to not get tracer function calls
830 * in this critical section:
832 spin_acquire(&time_sync_lock.dep_map, 0, 0, _THIS_IP_);
833 __raw_spin_lock(&time_sync_lock.raw_lock);
835 if (time < prev_global_time) {
836 per_cpu(time_offset, cpu) += prev_global_time - time;
837 time = prev_global_time;
838 } else {
839 prev_global_time = time;
842 __raw_spin_unlock(&time_sync_lock.raw_lock);
843 spin_release(&time_sync_lock.dep_map, 1, _THIS_IP_);
845 return time;
848 static unsigned long long __cpu_clock(int cpu)
850 unsigned long long now;
853 * Only call sched_clock() if the scheduler has already been
854 * initialized (some code might call cpu_clock() very early):
856 if (unlikely(!scheduler_running))
857 return 0;
859 now = sched_clock_cpu(cpu);
861 return now;
865 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
866 * clock constructed from sched_clock():
868 unsigned long long cpu_clock(int cpu)
870 unsigned long long prev_cpu_time, time, delta_time;
871 unsigned long flags;
873 local_irq_save(flags);
874 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
875 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
876 delta_time = time-prev_cpu_time;
878 if (unlikely(delta_time > time_sync_thresh)) {
879 time = __sync_cpu_clock(time, cpu);
880 per_cpu(prev_cpu_time, cpu) = time;
882 local_irq_restore(flags);
884 return time;
886 EXPORT_SYMBOL_GPL(cpu_clock);
888 #ifndef prepare_arch_switch
889 # define prepare_arch_switch(next) do { } while (0)
890 #endif
891 #ifndef finish_arch_switch
892 # define finish_arch_switch(prev) do { } while (0)
893 #endif
895 static inline int task_current(struct rq *rq, struct task_struct *p)
897 return rq->curr == p;
900 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
901 static inline int task_running(struct rq *rq, struct task_struct *p)
903 return task_current(rq, p);
906 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
910 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
912 #ifdef CONFIG_DEBUG_SPINLOCK
913 /* this is a valid case when another task releases the spinlock */
914 rq->lock.owner = current;
915 #endif
917 * If we are tracking spinlock dependencies then we have to
918 * fix up the runqueue lock - which gets 'carried over' from
919 * prev into current:
921 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
923 spin_unlock_irq(&rq->lock);
926 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
927 static inline int task_running(struct rq *rq, struct task_struct *p)
929 #ifdef CONFIG_SMP
930 return p->oncpu;
931 #else
932 return task_current(rq, p);
933 #endif
936 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
938 #ifdef CONFIG_SMP
940 * We can optimise this out completely for !SMP, because the
941 * SMP rebalancing from interrupt is the only thing that cares
942 * here.
944 next->oncpu = 1;
945 #endif
946 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
947 spin_unlock_irq(&rq->lock);
948 #else
949 spin_unlock(&rq->lock);
950 #endif
953 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
955 #ifdef CONFIG_SMP
957 * After ->oncpu is cleared, the task can be moved to a different CPU.
958 * We must ensure this doesn't happen until the switch is completely
959 * finished.
961 smp_wmb();
962 prev->oncpu = 0;
963 #endif
964 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
965 local_irq_enable();
966 #endif
968 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
971 * __task_rq_lock - lock the runqueue a given task resides on.
972 * Must be called interrupts disabled.
974 static inline struct rq *__task_rq_lock(struct task_struct *p)
975 __acquires(rq->lock)
977 for (;;) {
978 struct rq *rq = task_rq(p);
979 spin_lock(&rq->lock);
980 if (likely(rq == task_rq(p)))
981 return rq;
982 spin_unlock(&rq->lock);
987 * task_rq_lock - lock the runqueue a given task resides on and disable
988 * interrupts. Note the ordering: we can safely lookup the task_rq without
989 * explicitly disabling preemption.
991 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
992 __acquires(rq->lock)
994 struct rq *rq;
996 for (;;) {
997 local_irq_save(*flags);
998 rq = task_rq(p);
999 spin_lock(&rq->lock);
1000 if (likely(rq == task_rq(p)))
1001 return rq;
1002 spin_unlock_irqrestore(&rq->lock, *flags);
1006 static void __task_rq_unlock(struct rq *rq)
1007 __releases(rq->lock)
1009 spin_unlock(&rq->lock);
1012 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1013 __releases(rq->lock)
1015 spin_unlock_irqrestore(&rq->lock, *flags);
1019 * this_rq_lock - lock this runqueue and disable interrupts.
1021 static struct rq *this_rq_lock(void)
1022 __acquires(rq->lock)
1024 struct rq *rq;
1026 local_irq_disable();
1027 rq = this_rq();
1028 spin_lock(&rq->lock);
1030 return rq;
1033 static void __resched_task(struct task_struct *p, int tif_bit);
1035 static inline void resched_task(struct task_struct *p)
1037 __resched_task(p, TIF_NEED_RESCHED);
1040 #ifdef CONFIG_SCHED_HRTICK
1042 * Use HR-timers to deliver accurate preemption points.
1044 * Its all a bit involved since we cannot program an hrt while holding the
1045 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1046 * reschedule event.
1048 * When we get rescheduled we reprogram the hrtick_timer outside of the
1049 * rq->lock.
1051 static inline void resched_hrt(struct task_struct *p)
1053 __resched_task(p, TIF_HRTICK_RESCHED);
1056 static inline void resched_rq(struct rq *rq)
1058 unsigned long flags;
1060 spin_lock_irqsave(&rq->lock, flags);
1061 resched_task(rq->curr);
1062 spin_unlock_irqrestore(&rq->lock, flags);
1065 enum {
1066 HRTICK_SET, /* re-programm hrtick_timer */
1067 HRTICK_RESET, /* not a new slice */
1068 HRTICK_BLOCK, /* stop hrtick operations */
1072 * Use hrtick when:
1073 * - enabled by features
1074 * - hrtimer is actually high res
1076 static inline int hrtick_enabled(struct rq *rq)
1078 if (!sched_feat(HRTICK))
1079 return 0;
1080 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
1081 return 0;
1082 return hrtimer_is_hres_active(&rq->hrtick_timer);
1086 * Called to set the hrtick timer state.
1088 * called with rq->lock held and irqs disabled
1090 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1092 assert_spin_locked(&rq->lock);
1095 * preempt at: now + delay
1097 rq->hrtick_expire =
1098 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1100 * indicate we need to program the timer
1102 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1103 if (reset)
1104 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1107 * New slices are called from the schedule path and don't need a
1108 * forced reschedule.
1110 if (reset)
1111 resched_hrt(rq->curr);
1114 static void hrtick_clear(struct rq *rq)
1116 if (hrtimer_active(&rq->hrtick_timer))
1117 hrtimer_cancel(&rq->hrtick_timer);
1121 * Update the timer from the possible pending state.
1123 static void hrtick_set(struct rq *rq)
1125 ktime_t time;
1126 int set, reset;
1127 unsigned long flags;
1129 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1131 spin_lock_irqsave(&rq->lock, flags);
1132 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1133 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1134 time = rq->hrtick_expire;
1135 clear_thread_flag(TIF_HRTICK_RESCHED);
1136 spin_unlock_irqrestore(&rq->lock, flags);
1138 if (set) {
1139 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1140 if (reset && !hrtimer_active(&rq->hrtick_timer))
1141 resched_rq(rq);
1142 } else
1143 hrtick_clear(rq);
1147 * High-resolution timer tick.
1148 * Runs from hardirq context with interrupts disabled.
1150 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1152 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1154 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1156 spin_lock(&rq->lock);
1157 update_rq_clock(rq);
1158 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1159 spin_unlock(&rq->lock);
1161 return HRTIMER_NORESTART;
1164 static void hotplug_hrtick_disable(int cpu)
1166 struct rq *rq = cpu_rq(cpu);
1167 unsigned long flags;
1169 spin_lock_irqsave(&rq->lock, flags);
1170 rq->hrtick_flags = 0;
1171 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1172 spin_unlock_irqrestore(&rq->lock, flags);
1174 hrtick_clear(rq);
1177 static void hotplug_hrtick_enable(int cpu)
1179 struct rq *rq = cpu_rq(cpu);
1180 unsigned long flags;
1182 spin_lock_irqsave(&rq->lock, flags);
1183 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1184 spin_unlock_irqrestore(&rq->lock, flags);
1187 static int
1188 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1190 int cpu = (int)(long)hcpu;
1192 switch (action) {
1193 case CPU_UP_CANCELED:
1194 case CPU_UP_CANCELED_FROZEN:
1195 case CPU_DOWN_PREPARE:
1196 case CPU_DOWN_PREPARE_FROZEN:
1197 case CPU_DEAD:
1198 case CPU_DEAD_FROZEN:
1199 hotplug_hrtick_disable(cpu);
1200 return NOTIFY_OK;
1202 case CPU_UP_PREPARE:
1203 case CPU_UP_PREPARE_FROZEN:
1204 case CPU_DOWN_FAILED:
1205 case CPU_DOWN_FAILED_FROZEN:
1206 case CPU_ONLINE:
1207 case CPU_ONLINE_FROZEN:
1208 hotplug_hrtick_enable(cpu);
1209 return NOTIFY_OK;
1212 return NOTIFY_DONE;
1215 static void init_hrtick(void)
1217 hotcpu_notifier(hotplug_hrtick, 0);
1220 static void init_rq_hrtick(struct rq *rq)
1222 rq->hrtick_flags = 0;
1223 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1224 rq->hrtick_timer.function = hrtick;
1225 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1228 void hrtick_resched(void)
1230 struct rq *rq;
1231 unsigned long flags;
1233 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1234 return;
1236 local_irq_save(flags);
1237 rq = cpu_rq(smp_processor_id());
1238 hrtick_set(rq);
1239 local_irq_restore(flags);
1241 #else
1242 static inline void hrtick_clear(struct rq *rq)
1246 static inline void hrtick_set(struct rq *rq)
1250 static inline void init_rq_hrtick(struct rq *rq)
1254 void hrtick_resched(void)
1258 static inline void init_hrtick(void)
1261 #endif
1264 * resched_task - mark a task 'to be rescheduled now'.
1266 * On UP this means the setting of the need_resched flag, on SMP it
1267 * might also involve a cross-CPU call to trigger the scheduler on
1268 * the target CPU.
1270 #ifdef CONFIG_SMP
1272 #ifndef tsk_is_polling
1273 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1274 #endif
1276 static void __resched_task(struct task_struct *p, int tif_bit)
1278 int cpu;
1280 assert_spin_locked(&task_rq(p)->lock);
1282 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1283 return;
1285 set_tsk_thread_flag(p, tif_bit);
1287 cpu = task_cpu(p);
1288 if (cpu == smp_processor_id())
1289 return;
1291 /* NEED_RESCHED must be visible before we test polling */
1292 smp_mb();
1293 if (!tsk_is_polling(p))
1294 smp_send_reschedule(cpu);
1297 static void resched_cpu(int cpu)
1299 struct rq *rq = cpu_rq(cpu);
1300 unsigned long flags;
1302 if (!spin_trylock_irqsave(&rq->lock, flags))
1303 return;
1304 resched_task(cpu_curr(cpu));
1305 spin_unlock_irqrestore(&rq->lock, flags);
1308 #ifdef CONFIG_NO_HZ
1310 * When add_timer_on() enqueues a timer into the timer wheel of an
1311 * idle CPU then this timer might expire before the next timer event
1312 * which is scheduled to wake up that CPU. In case of a completely
1313 * idle system the next event might even be infinite time into the
1314 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1315 * leaves the inner idle loop so the newly added timer is taken into
1316 * account when the CPU goes back to idle and evaluates the timer
1317 * wheel for the next timer event.
1319 void wake_up_idle_cpu(int cpu)
1321 struct rq *rq = cpu_rq(cpu);
1323 if (cpu == smp_processor_id())
1324 return;
1327 * This is safe, as this function is called with the timer
1328 * wheel base lock of (cpu) held. When the CPU is on the way
1329 * to idle and has not yet set rq->curr to idle then it will
1330 * be serialized on the timer wheel base lock and take the new
1331 * timer into account automatically.
1333 if (rq->curr != rq->idle)
1334 return;
1337 * We can set TIF_RESCHED on the idle task of the other CPU
1338 * lockless. The worst case is that the other CPU runs the
1339 * idle task through an additional NOOP schedule()
1341 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1343 /* NEED_RESCHED must be visible before we test polling */
1344 smp_mb();
1345 if (!tsk_is_polling(rq->idle))
1346 smp_send_reschedule(cpu);
1348 #endif
1350 #else
1351 static void __resched_task(struct task_struct *p, int tif_bit)
1353 assert_spin_locked(&task_rq(p)->lock);
1354 set_tsk_thread_flag(p, tif_bit);
1356 #endif
1358 #if BITS_PER_LONG == 32
1359 # define WMULT_CONST (~0UL)
1360 #else
1361 # define WMULT_CONST (1UL << 32)
1362 #endif
1364 #define WMULT_SHIFT 32
1367 * Shift right and round:
1369 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1372 * delta *= weight / lw
1374 static unsigned long
1375 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1376 struct load_weight *lw)
1378 u64 tmp;
1380 if (!lw->inv_weight)
1381 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)/(lw->weight+1);
1383 tmp = (u64)delta_exec * weight;
1385 * Check whether we'd overflow the 64-bit multiplication:
1387 if (unlikely(tmp > WMULT_CONST))
1388 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1389 WMULT_SHIFT/2);
1390 else
1391 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1393 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1396 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1398 lw->weight += inc;
1399 lw->inv_weight = 0;
1402 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1404 lw->weight -= dec;
1405 lw->inv_weight = 0;
1409 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1410 * of tasks with abnormal "nice" values across CPUs the contribution that
1411 * each task makes to its run queue's load is weighted according to its
1412 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1413 * scaled version of the new time slice allocation that they receive on time
1414 * slice expiry etc.
1417 #define WEIGHT_IDLEPRIO 2
1418 #define WMULT_IDLEPRIO (1 << 31)
1421 * Nice levels are multiplicative, with a gentle 10% change for every
1422 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1423 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1424 * that remained on nice 0.
1426 * The "10% effect" is relative and cumulative: from _any_ nice level,
1427 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1428 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1429 * If a task goes up by ~10% and another task goes down by ~10% then
1430 * the relative distance between them is ~25%.)
1432 static const int prio_to_weight[40] = {
1433 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1434 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1435 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1436 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1437 /* 0 */ 1024, 820, 655, 526, 423,
1438 /* 5 */ 335, 272, 215, 172, 137,
1439 /* 10 */ 110, 87, 70, 56, 45,
1440 /* 15 */ 36, 29, 23, 18, 15,
1444 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1446 * In cases where the weight does not change often, we can use the
1447 * precalculated inverse to speed up arithmetics by turning divisions
1448 * into multiplications:
1450 static const u32 prio_to_wmult[40] = {
1451 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1452 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1453 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1454 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1455 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1456 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1457 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1458 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1461 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1464 * runqueue iterator, to support SMP load-balancing between different
1465 * scheduling classes, without having to expose their internal data
1466 * structures to the load-balancing proper:
1468 struct rq_iterator {
1469 void *arg;
1470 struct task_struct *(*start)(void *);
1471 struct task_struct *(*next)(void *);
1474 #ifdef CONFIG_SMP
1475 static unsigned long
1476 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1477 unsigned long max_load_move, struct sched_domain *sd,
1478 enum cpu_idle_type idle, int *all_pinned,
1479 int *this_best_prio, struct rq_iterator *iterator);
1481 static int
1482 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1483 struct sched_domain *sd, enum cpu_idle_type idle,
1484 struct rq_iterator *iterator);
1485 #endif
1487 #ifdef CONFIG_CGROUP_CPUACCT
1488 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1489 #else
1490 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1491 #endif
1493 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1495 update_load_add(&rq->load, load);
1498 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1500 update_load_sub(&rq->load, load);
1503 #ifdef CONFIG_SMP
1504 static unsigned long source_load(int cpu, int type);
1505 static unsigned long target_load(int cpu, int type);
1506 static unsigned long cpu_avg_load_per_task(int cpu);
1507 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1509 #ifdef CONFIG_FAIR_GROUP_SCHED
1512 * Group load balancing.
1514 * We calculate a few balance domain wide aggregate numbers; load and weight.
1515 * Given the pictures below, and assuming each item has equal weight:
1517 * root 1 - thread
1518 * / | \ A - group
1519 * A 1 B
1520 * /|\ / \
1521 * C 2 D 3 4
1522 * | |
1523 * 5 6
1525 * load:
1526 * A and B get 1/3-rd of the total load. C and D get 1/3-rd of A's 1/3-rd,
1527 * which equals 1/9-th of the total load.
1529 * shares:
1530 * The weight of this group on the selected cpus.
1532 * rq_weight:
1533 * Direct sum of all the cpu's their rq weight, e.g. A would get 3 while
1534 * B would get 2.
1536 * task_weight:
1537 * Part of the rq_weight contributed by tasks; all groups except B would
1538 * get 1, B gets 2.
1541 static inline struct aggregate_struct *
1542 aggregate(struct task_group *tg, struct sched_domain *sd)
1544 return &tg->cfs_rq[sd->first_cpu]->aggregate;
1547 typedef void (*aggregate_func)(struct task_group *, struct sched_domain *);
1550 * Iterate the full tree, calling @down when first entering a node and @up when
1551 * leaving it for the final time.
1553 static
1554 void aggregate_walk_tree(aggregate_func down, aggregate_func up,
1555 struct sched_domain *sd)
1557 struct task_group *parent, *child;
1559 rcu_read_lock();
1560 parent = &root_task_group;
1561 down:
1562 (*down)(parent, sd);
1563 list_for_each_entry_rcu(child, &parent->children, siblings) {
1564 parent = child;
1565 goto down;
1568 continue;
1570 (*up)(parent, sd);
1572 child = parent;
1573 parent = parent->parent;
1574 if (parent)
1575 goto up;
1576 rcu_read_unlock();
1580 * Calculate the aggregate runqueue weight.
1582 static
1583 void aggregate_group_weight(struct task_group *tg, struct sched_domain *sd)
1585 unsigned long rq_weight = 0;
1586 unsigned long task_weight = 0;
1587 int i;
1589 for_each_cpu_mask(i, sd->span) {
1590 rq_weight += tg->cfs_rq[i]->load.weight;
1591 task_weight += tg->cfs_rq[i]->task_weight;
1594 aggregate(tg, sd)->rq_weight = rq_weight;
1595 aggregate(tg, sd)->task_weight = task_weight;
1599 * Compute the weight of this group on the given cpus.
1601 static
1602 void aggregate_group_shares(struct task_group *tg, struct sched_domain *sd)
1604 unsigned long shares = 0;
1605 int i;
1607 for_each_cpu_mask(i, sd->span)
1608 shares += tg->cfs_rq[i]->shares;
1610 if ((!shares && aggregate(tg, sd)->rq_weight) || shares > tg->shares)
1611 shares = tg->shares;
1613 aggregate(tg, sd)->shares = shares;
1617 * Compute the load fraction assigned to this group, relies on the aggregate
1618 * weight and this group's parent's load, i.e. top-down.
1620 static
1621 void aggregate_group_load(struct task_group *tg, struct sched_domain *sd)
1623 unsigned long load;
1625 if (!tg->parent) {
1626 int i;
1628 load = 0;
1629 for_each_cpu_mask(i, sd->span)
1630 load += cpu_rq(i)->load.weight;
1632 } else {
1633 load = aggregate(tg->parent, sd)->load;
1636 * shares is our weight in the parent's rq so
1637 * shares/parent->rq_weight gives our fraction of the load
1639 load *= aggregate(tg, sd)->shares;
1640 load /= aggregate(tg->parent, sd)->rq_weight + 1;
1643 aggregate(tg, sd)->load = load;
1646 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1649 * Calculate and set the cpu's group shares.
1651 static void
1652 __update_group_shares_cpu(struct task_group *tg, struct sched_domain *sd,
1653 int tcpu)
1655 int boost = 0;
1656 unsigned long shares;
1657 unsigned long rq_weight;
1659 if (!tg->se[tcpu])
1660 return;
1662 rq_weight = tg->cfs_rq[tcpu]->load.weight;
1665 * If there are currently no tasks on the cpu pretend there is one of
1666 * average load so that when a new task gets to run here it will not
1667 * get delayed by group starvation.
1669 if (!rq_weight) {
1670 boost = 1;
1671 rq_weight = NICE_0_LOAD;
1675 * \Sum shares * rq_weight
1676 * shares = -----------------------
1677 * \Sum rq_weight
1680 shares = aggregate(tg, sd)->shares * rq_weight;
1681 shares /= aggregate(tg, sd)->rq_weight + 1;
1684 * record the actual number of shares, not the boosted amount.
1686 tg->cfs_rq[tcpu]->shares = boost ? 0 : shares;
1688 if (shares < MIN_SHARES)
1689 shares = MIN_SHARES;
1690 else if (shares > MAX_SHARES)
1691 shares = MAX_SHARES;
1693 __set_se_shares(tg->se[tcpu], shares);
1697 * Re-adjust the weights on the cpu the task came from and on the cpu the
1698 * task went to.
1700 static void
1701 __move_group_shares(struct task_group *tg, struct sched_domain *sd,
1702 int scpu, int dcpu)
1704 unsigned long shares;
1706 shares = tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1708 __update_group_shares_cpu(tg, sd, scpu);
1709 __update_group_shares_cpu(tg, sd, dcpu);
1712 * ensure we never loose shares due to rounding errors in the
1713 * above redistribution.
1715 shares -= tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1716 if (shares)
1717 tg->cfs_rq[dcpu]->shares += shares;
1721 * Because changing a group's shares changes the weight of the super-group
1722 * we need to walk up the tree and change all shares until we hit the root.
1724 static void
1725 move_group_shares(struct task_group *tg, struct sched_domain *sd,
1726 int scpu, int dcpu)
1728 while (tg) {
1729 __move_group_shares(tg, sd, scpu, dcpu);
1730 tg = tg->parent;
1734 static
1735 void aggregate_group_set_shares(struct task_group *tg, struct sched_domain *sd)
1737 unsigned long shares = aggregate(tg, sd)->shares;
1738 int i;
1740 for_each_cpu_mask(i, sd->span) {
1741 struct rq *rq = cpu_rq(i);
1742 unsigned long flags;
1744 spin_lock_irqsave(&rq->lock, flags);
1745 __update_group_shares_cpu(tg, sd, i);
1746 spin_unlock_irqrestore(&rq->lock, flags);
1749 aggregate_group_shares(tg, sd);
1752 * ensure we never loose shares due to rounding errors in the
1753 * above redistribution.
1755 shares -= aggregate(tg, sd)->shares;
1756 if (shares) {
1757 tg->cfs_rq[sd->first_cpu]->shares += shares;
1758 aggregate(tg, sd)->shares += shares;
1763 * Calculate the accumulative weight and recursive load of each task group
1764 * while walking down the tree.
1766 static
1767 void aggregate_get_down(struct task_group *tg, struct sched_domain *sd)
1769 aggregate_group_weight(tg, sd);
1770 aggregate_group_shares(tg, sd);
1771 aggregate_group_load(tg, sd);
1775 * Rebalance the cpu shares while walking back up the tree.
1777 static
1778 void aggregate_get_up(struct task_group *tg, struct sched_domain *sd)
1780 aggregate_group_set_shares(tg, sd);
1783 static DEFINE_PER_CPU(spinlock_t, aggregate_lock);
1785 static void __init init_aggregate(void)
1787 int i;
1789 for_each_possible_cpu(i)
1790 spin_lock_init(&per_cpu(aggregate_lock, i));
1793 static int get_aggregate(struct sched_domain *sd)
1795 if (!spin_trylock(&per_cpu(aggregate_lock, sd->first_cpu)))
1796 return 0;
1798 aggregate_walk_tree(aggregate_get_down, aggregate_get_up, sd);
1799 return 1;
1802 static void put_aggregate(struct sched_domain *sd)
1804 spin_unlock(&per_cpu(aggregate_lock, sd->first_cpu));
1807 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1809 cfs_rq->shares = shares;
1812 #else
1814 static inline void init_aggregate(void)
1818 static inline int get_aggregate(struct sched_domain *sd)
1820 return 0;
1823 static inline void put_aggregate(struct sched_domain *sd)
1826 #endif
1828 #else /* CONFIG_SMP */
1830 #ifdef CONFIG_FAIR_GROUP_SCHED
1831 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1834 #endif
1836 #endif /* CONFIG_SMP */
1838 #include "sched_stats.h"
1839 #include "sched_idletask.c"
1840 #include "sched_fair.c"
1841 #include "sched_rt.c"
1842 #ifdef CONFIG_SCHED_DEBUG
1843 # include "sched_debug.c"
1844 #endif
1846 #define sched_class_highest (&rt_sched_class)
1848 static void inc_nr_running(struct rq *rq)
1850 rq->nr_running++;
1853 static void dec_nr_running(struct rq *rq)
1855 rq->nr_running--;
1858 static void set_load_weight(struct task_struct *p)
1860 if (task_has_rt_policy(p)) {
1861 p->se.load.weight = prio_to_weight[0] * 2;
1862 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1863 return;
1867 * SCHED_IDLE tasks get minimal weight:
1869 if (p->policy == SCHED_IDLE) {
1870 p->se.load.weight = WEIGHT_IDLEPRIO;
1871 p->se.load.inv_weight = WMULT_IDLEPRIO;
1872 return;
1875 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1876 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1879 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1881 sched_info_queued(p);
1882 p->sched_class->enqueue_task(rq, p, wakeup);
1883 p->se.on_rq = 1;
1886 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1888 p->sched_class->dequeue_task(rq, p, sleep);
1889 p->se.on_rq = 0;
1893 * __normal_prio - return the priority that is based on the static prio
1895 static inline int __normal_prio(struct task_struct *p)
1897 return p->static_prio;
1901 * Calculate the expected normal priority: i.e. priority
1902 * without taking RT-inheritance into account. Might be
1903 * boosted by interactivity modifiers. Changes upon fork,
1904 * setprio syscalls, and whenever the interactivity
1905 * estimator recalculates.
1907 static inline int normal_prio(struct task_struct *p)
1909 int prio;
1911 if (task_has_rt_policy(p))
1912 prio = MAX_RT_PRIO-1 - p->rt_priority;
1913 else
1914 prio = __normal_prio(p);
1915 return prio;
1919 * Calculate the current priority, i.e. the priority
1920 * taken into account by the scheduler. This value might
1921 * be boosted by RT tasks, or might be boosted by
1922 * interactivity modifiers. Will be RT if the task got
1923 * RT-boosted. If not then it returns p->normal_prio.
1925 static int effective_prio(struct task_struct *p)
1927 p->normal_prio = normal_prio(p);
1929 * If we are RT tasks or we were boosted to RT priority,
1930 * keep the priority unchanged. Otherwise, update priority
1931 * to the normal priority:
1933 if (!rt_prio(p->prio))
1934 return p->normal_prio;
1935 return p->prio;
1939 * activate_task - move a task to the runqueue.
1941 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1943 if (task_contributes_to_load(p))
1944 rq->nr_uninterruptible--;
1946 enqueue_task(rq, p, wakeup);
1947 inc_nr_running(rq);
1951 * deactivate_task - remove a task from the runqueue.
1953 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1955 if (task_contributes_to_load(p))
1956 rq->nr_uninterruptible++;
1958 dequeue_task(rq, p, sleep);
1959 dec_nr_running(rq);
1963 * task_curr - is this task currently executing on a CPU?
1964 * @p: the task in question.
1966 inline int task_curr(const struct task_struct *p)
1968 return cpu_curr(task_cpu(p)) == p;
1971 /* Used instead of source_load when we know the type == 0 */
1972 unsigned long weighted_cpuload(const int cpu)
1974 return cpu_rq(cpu)->load.weight;
1977 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1979 set_task_rq(p, cpu);
1980 #ifdef CONFIG_SMP
1982 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1983 * successfuly executed on another CPU. We must ensure that updates of
1984 * per-task data have been completed by this moment.
1986 smp_wmb();
1987 task_thread_info(p)->cpu = cpu;
1988 #endif
1991 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1992 const struct sched_class *prev_class,
1993 int oldprio, int running)
1995 if (prev_class != p->sched_class) {
1996 if (prev_class->switched_from)
1997 prev_class->switched_from(rq, p, running);
1998 p->sched_class->switched_to(rq, p, running);
1999 } else
2000 p->sched_class->prio_changed(rq, p, oldprio, running);
2003 #ifdef CONFIG_SMP
2006 * Is this task likely cache-hot:
2008 static int
2009 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2011 s64 delta;
2014 * Buddy candidates are cache hot:
2016 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
2017 return 1;
2019 if (p->sched_class != &fair_sched_class)
2020 return 0;
2022 if (sysctl_sched_migration_cost == -1)
2023 return 1;
2024 if (sysctl_sched_migration_cost == 0)
2025 return 0;
2027 delta = now - p->se.exec_start;
2029 return delta < (s64)sysctl_sched_migration_cost;
2033 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2035 int old_cpu = task_cpu(p);
2036 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2037 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2038 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2039 u64 clock_offset;
2041 clock_offset = old_rq->clock - new_rq->clock;
2043 #ifdef CONFIG_SCHEDSTATS
2044 if (p->se.wait_start)
2045 p->se.wait_start -= clock_offset;
2046 if (p->se.sleep_start)
2047 p->se.sleep_start -= clock_offset;
2048 if (p->se.block_start)
2049 p->se.block_start -= clock_offset;
2050 if (old_cpu != new_cpu) {
2051 schedstat_inc(p, se.nr_migrations);
2052 if (task_hot(p, old_rq->clock, NULL))
2053 schedstat_inc(p, se.nr_forced2_migrations);
2055 #endif
2056 p->se.vruntime -= old_cfsrq->min_vruntime -
2057 new_cfsrq->min_vruntime;
2059 __set_task_cpu(p, new_cpu);
2062 struct migration_req {
2063 struct list_head list;
2065 struct task_struct *task;
2066 int dest_cpu;
2068 struct completion done;
2072 * The task's runqueue lock must be held.
2073 * Returns true if you have to wait for migration thread.
2075 static int
2076 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2078 struct rq *rq = task_rq(p);
2081 * If the task is not on a runqueue (and not running), then
2082 * it is sufficient to simply update the task's cpu field.
2084 if (!p->se.on_rq && !task_running(rq, p)) {
2085 set_task_cpu(p, dest_cpu);
2086 return 0;
2089 init_completion(&req->done);
2090 req->task = p;
2091 req->dest_cpu = dest_cpu;
2092 list_add(&req->list, &rq->migration_queue);
2094 return 1;
2098 * wait_task_inactive - wait for a thread to unschedule.
2100 * The caller must ensure that the task *will* unschedule sometime soon,
2101 * else this function might spin for a *long* time. This function can't
2102 * be called with interrupts off, or it may introduce deadlock with
2103 * smp_call_function() if an IPI is sent by the same process we are
2104 * waiting to become inactive.
2106 void wait_task_inactive(struct task_struct *p)
2108 unsigned long flags;
2109 int running, on_rq;
2110 struct rq *rq;
2112 for (;;) {
2114 * We do the initial early heuristics without holding
2115 * any task-queue locks at all. We'll only try to get
2116 * the runqueue lock when things look like they will
2117 * work out!
2119 rq = task_rq(p);
2122 * If the task is actively running on another CPU
2123 * still, just relax and busy-wait without holding
2124 * any locks.
2126 * NOTE! Since we don't hold any locks, it's not
2127 * even sure that "rq" stays as the right runqueue!
2128 * But we don't care, since "task_running()" will
2129 * return false if the runqueue has changed and p
2130 * is actually now running somewhere else!
2132 while (task_running(rq, p))
2133 cpu_relax();
2136 * Ok, time to look more closely! We need the rq
2137 * lock now, to be *sure*. If we're wrong, we'll
2138 * just go back and repeat.
2140 rq = task_rq_lock(p, &flags);
2141 running = task_running(rq, p);
2142 on_rq = p->se.on_rq;
2143 task_rq_unlock(rq, &flags);
2146 * Was it really running after all now that we
2147 * checked with the proper locks actually held?
2149 * Oops. Go back and try again..
2151 if (unlikely(running)) {
2152 cpu_relax();
2153 continue;
2157 * It's not enough that it's not actively running,
2158 * it must be off the runqueue _entirely_, and not
2159 * preempted!
2161 * So if it wa still runnable (but just not actively
2162 * running right now), it's preempted, and we should
2163 * yield - it could be a while.
2165 if (unlikely(on_rq)) {
2166 schedule_timeout_uninterruptible(1);
2167 continue;
2171 * Ahh, all good. It wasn't running, and it wasn't
2172 * runnable, which means that it will never become
2173 * running in the future either. We're all done!
2175 break;
2179 /***
2180 * kick_process - kick a running thread to enter/exit the kernel
2181 * @p: the to-be-kicked thread
2183 * Cause a process which is running on another CPU to enter
2184 * kernel-mode, without any delay. (to get signals handled.)
2186 * NOTE: this function doesnt have to take the runqueue lock,
2187 * because all it wants to ensure is that the remote task enters
2188 * the kernel. If the IPI races and the task has been migrated
2189 * to another CPU then no harm is done and the purpose has been
2190 * achieved as well.
2192 void kick_process(struct task_struct *p)
2194 int cpu;
2196 preempt_disable();
2197 cpu = task_cpu(p);
2198 if ((cpu != smp_processor_id()) && task_curr(p))
2199 smp_send_reschedule(cpu);
2200 preempt_enable();
2204 * Return a low guess at the load of a migration-source cpu weighted
2205 * according to the scheduling class and "nice" value.
2207 * We want to under-estimate the load of migration sources, to
2208 * balance conservatively.
2210 static unsigned long source_load(int cpu, int type)
2212 struct rq *rq = cpu_rq(cpu);
2213 unsigned long total = weighted_cpuload(cpu);
2215 if (type == 0)
2216 return total;
2218 return min(rq->cpu_load[type-1], total);
2222 * Return a high guess at the load of a migration-target cpu weighted
2223 * according to the scheduling class and "nice" value.
2225 static unsigned long target_load(int cpu, int type)
2227 struct rq *rq = cpu_rq(cpu);
2228 unsigned long total = weighted_cpuload(cpu);
2230 if (type == 0)
2231 return total;
2233 return max(rq->cpu_load[type-1], total);
2237 * Return the average load per task on the cpu's run queue
2239 static unsigned long cpu_avg_load_per_task(int cpu)
2241 struct rq *rq = cpu_rq(cpu);
2242 unsigned long total = weighted_cpuload(cpu);
2243 unsigned long n = rq->nr_running;
2245 return n ? total / n : SCHED_LOAD_SCALE;
2249 * find_idlest_group finds and returns the least busy CPU group within the
2250 * domain.
2252 static struct sched_group *
2253 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2255 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2256 unsigned long min_load = ULONG_MAX, this_load = 0;
2257 int load_idx = sd->forkexec_idx;
2258 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2260 do {
2261 unsigned long load, avg_load;
2262 int local_group;
2263 int i;
2265 /* Skip over this group if it has no CPUs allowed */
2266 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2267 continue;
2269 local_group = cpu_isset(this_cpu, group->cpumask);
2271 /* Tally up the load of all CPUs in the group */
2272 avg_load = 0;
2274 for_each_cpu_mask(i, group->cpumask) {
2275 /* Bias balancing toward cpus of our domain */
2276 if (local_group)
2277 load = source_load(i, load_idx);
2278 else
2279 load = target_load(i, load_idx);
2281 avg_load += load;
2284 /* Adjust by relative CPU power of the group */
2285 avg_load = sg_div_cpu_power(group,
2286 avg_load * SCHED_LOAD_SCALE);
2288 if (local_group) {
2289 this_load = avg_load;
2290 this = group;
2291 } else if (avg_load < min_load) {
2292 min_load = avg_load;
2293 idlest = group;
2295 } while (group = group->next, group != sd->groups);
2297 if (!idlest || 100*this_load < imbalance*min_load)
2298 return NULL;
2299 return idlest;
2303 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2305 static int
2306 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2307 cpumask_t *tmp)
2309 unsigned long load, min_load = ULONG_MAX;
2310 int idlest = -1;
2311 int i;
2313 /* Traverse only the allowed CPUs */
2314 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2316 for_each_cpu_mask(i, *tmp) {
2317 load = weighted_cpuload(i);
2319 if (load < min_load || (load == min_load && i == this_cpu)) {
2320 min_load = load;
2321 idlest = i;
2325 return idlest;
2329 * sched_balance_self: balance the current task (running on cpu) in domains
2330 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2331 * SD_BALANCE_EXEC.
2333 * Balance, ie. select the least loaded group.
2335 * Returns the target CPU number, or the same CPU if no balancing is needed.
2337 * preempt must be disabled.
2339 static int sched_balance_self(int cpu, int flag)
2341 struct task_struct *t = current;
2342 struct sched_domain *tmp, *sd = NULL;
2344 for_each_domain(cpu, tmp) {
2346 * If power savings logic is enabled for a domain, stop there.
2348 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2349 break;
2350 if (tmp->flags & flag)
2351 sd = tmp;
2354 while (sd) {
2355 cpumask_t span, tmpmask;
2356 struct sched_group *group;
2357 int new_cpu, weight;
2359 if (!(sd->flags & flag)) {
2360 sd = sd->child;
2361 continue;
2364 span = sd->span;
2365 group = find_idlest_group(sd, t, cpu);
2366 if (!group) {
2367 sd = sd->child;
2368 continue;
2371 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2372 if (new_cpu == -1 || new_cpu == cpu) {
2373 /* Now try balancing at a lower domain level of cpu */
2374 sd = sd->child;
2375 continue;
2378 /* Now try balancing at a lower domain level of new_cpu */
2379 cpu = new_cpu;
2380 sd = NULL;
2381 weight = cpus_weight(span);
2382 for_each_domain(cpu, tmp) {
2383 if (weight <= cpus_weight(tmp->span))
2384 break;
2385 if (tmp->flags & flag)
2386 sd = tmp;
2388 /* while loop will break here if sd == NULL */
2391 return cpu;
2394 #endif /* CONFIG_SMP */
2396 /***
2397 * try_to_wake_up - wake up a thread
2398 * @p: the to-be-woken-up thread
2399 * @state: the mask of task states that can be woken
2400 * @sync: do a synchronous wakeup?
2402 * Put it on the run-queue if it's not already there. The "current"
2403 * thread is always on the run-queue (except when the actual
2404 * re-schedule is in progress), and as such you're allowed to do
2405 * the simpler "current->state = TASK_RUNNING" to mark yourself
2406 * runnable without the overhead of this.
2408 * returns failure only if the task is already active.
2410 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2412 int cpu, orig_cpu, this_cpu, success = 0;
2413 unsigned long flags;
2414 long old_state;
2415 struct rq *rq;
2417 if (!sched_feat(SYNC_WAKEUPS))
2418 sync = 0;
2420 smp_wmb();
2421 rq = task_rq_lock(p, &flags);
2422 old_state = p->state;
2423 if (!(old_state & state))
2424 goto out;
2426 if (p->se.on_rq)
2427 goto out_running;
2429 cpu = task_cpu(p);
2430 orig_cpu = cpu;
2431 this_cpu = smp_processor_id();
2433 #ifdef CONFIG_SMP
2434 if (unlikely(task_running(rq, p)))
2435 goto out_activate;
2437 cpu = p->sched_class->select_task_rq(p, sync);
2438 if (cpu != orig_cpu) {
2439 set_task_cpu(p, cpu);
2440 task_rq_unlock(rq, &flags);
2441 /* might preempt at this point */
2442 rq = task_rq_lock(p, &flags);
2443 old_state = p->state;
2444 if (!(old_state & state))
2445 goto out;
2446 if (p->se.on_rq)
2447 goto out_running;
2449 this_cpu = smp_processor_id();
2450 cpu = task_cpu(p);
2453 #ifdef CONFIG_SCHEDSTATS
2454 schedstat_inc(rq, ttwu_count);
2455 if (cpu == this_cpu)
2456 schedstat_inc(rq, ttwu_local);
2457 else {
2458 struct sched_domain *sd;
2459 for_each_domain(this_cpu, sd) {
2460 if (cpu_isset(cpu, sd->span)) {
2461 schedstat_inc(sd, ttwu_wake_remote);
2462 break;
2466 #endif
2468 out_activate:
2469 #endif /* CONFIG_SMP */
2470 schedstat_inc(p, se.nr_wakeups);
2471 if (sync)
2472 schedstat_inc(p, se.nr_wakeups_sync);
2473 if (orig_cpu != cpu)
2474 schedstat_inc(p, se.nr_wakeups_migrate);
2475 if (cpu == this_cpu)
2476 schedstat_inc(p, se.nr_wakeups_local);
2477 else
2478 schedstat_inc(p, se.nr_wakeups_remote);
2479 update_rq_clock(rq);
2480 activate_task(rq, p, 1);
2481 success = 1;
2483 out_running:
2484 check_preempt_curr(rq, p);
2486 p->state = TASK_RUNNING;
2487 #ifdef CONFIG_SMP
2488 if (p->sched_class->task_wake_up)
2489 p->sched_class->task_wake_up(rq, p);
2490 #endif
2491 out:
2492 task_rq_unlock(rq, &flags);
2494 return success;
2497 int wake_up_process(struct task_struct *p)
2499 return try_to_wake_up(p, TASK_ALL, 0);
2501 EXPORT_SYMBOL(wake_up_process);
2503 int wake_up_state(struct task_struct *p, unsigned int state)
2505 return try_to_wake_up(p, state, 0);
2509 * Perform scheduler related setup for a newly forked process p.
2510 * p is forked by current.
2512 * __sched_fork() is basic setup used by init_idle() too:
2514 static void __sched_fork(struct task_struct *p)
2516 p->se.exec_start = 0;
2517 p->se.sum_exec_runtime = 0;
2518 p->se.prev_sum_exec_runtime = 0;
2519 p->se.last_wakeup = 0;
2520 p->se.avg_overlap = 0;
2522 #ifdef CONFIG_SCHEDSTATS
2523 p->se.wait_start = 0;
2524 p->se.sum_sleep_runtime = 0;
2525 p->se.sleep_start = 0;
2526 p->se.block_start = 0;
2527 p->se.sleep_max = 0;
2528 p->se.block_max = 0;
2529 p->se.exec_max = 0;
2530 p->se.slice_max = 0;
2531 p->se.wait_max = 0;
2532 #endif
2534 INIT_LIST_HEAD(&p->rt.run_list);
2535 p->se.on_rq = 0;
2536 INIT_LIST_HEAD(&p->se.group_node);
2538 #ifdef CONFIG_PREEMPT_NOTIFIERS
2539 INIT_HLIST_HEAD(&p->preempt_notifiers);
2540 #endif
2543 * We mark the process as running here, but have not actually
2544 * inserted it onto the runqueue yet. This guarantees that
2545 * nobody will actually run it, and a signal or other external
2546 * event cannot wake it up and insert it on the runqueue either.
2548 p->state = TASK_RUNNING;
2552 * fork()/clone()-time setup:
2554 void sched_fork(struct task_struct *p, int clone_flags)
2556 int cpu = get_cpu();
2558 __sched_fork(p);
2560 #ifdef CONFIG_SMP
2561 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2562 #endif
2563 set_task_cpu(p, cpu);
2566 * Make sure we do not leak PI boosting priority to the child:
2568 p->prio = current->normal_prio;
2569 if (!rt_prio(p->prio))
2570 p->sched_class = &fair_sched_class;
2572 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2573 if (likely(sched_info_on()))
2574 memset(&p->sched_info, 0, sizeof(p->sched_info));
2575 #endif
2576 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2577 p->oncpu = 0;
2578 #endif
2579 #ifdef CONFIG_PREEMPT
2580 /* Want to start with kernel preemption disabled. */
2581 task_thread_info(p)->preempt_count = 1;
2582 #endif
2583 put_cpu();
2587 * wake_up_new_task - wake up a newly created task for the first time.
2589 * This function will do some initial scheduler statistics housekeeping
2590 * that must be done for every newly created context, then puts the task
2591 * on the runqueue and wakes it.
2593 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2595 unsigned long flags;
2596 struct rq *rq;
2598 rq = task_rq_lock(p, &flags);
2599 BUG_ON(p->state != TASK_RUNNING);
2600 update_rq_clock(rq);
2602 p->prio = effective_prio(p);
2604 if (!p->sched_class->task_new || !current->se.on_rq) {
2605 activate_task(rq, p, 0);
2606 } else {
2608 * Let the scheduling class do new task startup
2609 * management (if any):
2611 p->sched_class->task_new(rq, p);
2612 inc_nr_running(rq);
2614 check_preempt_curr(rq, p);
2615 #ifdef CONFIG_SMP
2616 if (p->sched_class->task_wake_up)
2617 p->sched_class->task_wake_up(rq, p);
2618 #endif
2619 task_rq_unlock(rq, &flags);
2622 #ifdef CONFIG_PREEMPT_NOTIFIERS
2625 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2626 * @notifier: notifier struct to register
2628 void preempt_notifier_register(struct preempt_notifier *notifier)
2630 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2632 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2635 * preempt_notifier_unregister - no longer interested in preemption notifications
2636 * @notifier: notifier struct to unregister
2638 * This is safe to call from within a preemption notifier.
2640 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2642 hlist_del(&notifier->link);
2644 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2646 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2648 struct preempt_notifier *notifier;
2649 struct hlist_node *node;
2651 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2652 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2655 static void
2656 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2657 struct task_struct *next)
2659 struct preempt_notifier *notifier;
2660 struct hlist_node *node;
2662 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2663 notifier->ops->sched_out(notifier, next);
2666 #else
2668 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2672 static void
2673 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2674 struct task_struct *next)
2678 #endif
2681 * prepare_task_switch - prepare to switch tasks
2682 * @rq: the runqueue preparing to switch
2683 * @prev: the current task that is being switched out
2684 * @next: the task we are going to switch to.
2686 * This is called with the rq lock held and interrupts off. It must
2687 * be paired with a subsequent finish_task_switch after the context
2688 * switch.
2690 * prepare_task_switch sets up locking and calls architecture specific
2691 * hooks.
2693 static inline void
2694 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2695 struct task_struct *next)
2697 fire_sched_out_preempt_notifiers(prev, next);
2698 prepare_lock_switch(rq, next);
2699 prepare_arch_switch(next);
2703 * finish_task_switch - clean up after a task-switch
2704 * @rq: runqueue associated with task-switch
2705 * @prev: the thread we just switched away from.
2707 * finish_task_switch must be called after the context switch, paired
2708 * with a prepare_task_switch call before the context switch.
2709 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2710 * and do any other architecture-specific cleanup actions.
2712 * Note that we may have delayed dropping an mm in context_switch(). If
2713 * so, we finish that here outside of the runqueue lock. (Doing it
2714 * with the lock held can cause deadlocks; see schedule() for
2715 * details.)
2717 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2718 __releases(rq->lock)
2720 struct mm_struct *mm = rq->prev_mm;
2721 long prev_state;
2723 rq->prev_mm = NULL;
2726 * A task struct has one reference for the use as "current".
2727 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2728 * schedule one last time. The schedule call will never return, and
2729 * the scheduled task must drop that reference.
2730 * The test for TASK_DEAD must occur while the runqueue locks are
2731 * still held, otherwise prev could be scheduled on another cpu, die
2732 * there before we look at prev->state, and then the reference would
2733 * be dropped twice.
2734 * Manfred Spraul <manfred@colorfullife.com>
2736 prev_state = prev->state;
2737 finish_arch_switch(prev);
2738 finish_lock_switch(rq, prev);
2739 #ifdef CONFIG_SMP
2740 if (current->sched_class->post_schedule)
2741 current->sched_class->post_schedule(rq);
2742 #endif
2744 fire_sched_in_preempt_notifiers(current);
2745 if (mm)
2746 mmdrop(mm);
2747 if (unlikely(prev_state == TASK_DEAD)) {
2749 * Remove function-return probe instances associated with this
2750 * task and put them back on the free list.
2752 kprobe_flush_task(prev);
2753 put_task_struct(prev);
2758 * schedule_tail - first thing a freshly forked thread must call.
2759 * @prev: the thread we just switched away from.
2761 asmlinkage void schedule_tail(struct task_struct *prev)
2762 __releases(rq->lock)
2764 struct rq *rq = this_rq();
2766 finish_task_switch(rq, prev);
2767 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2768 /* In this case, finish_task_switch does not reenable preemption */
2769 preempt_enable();
2770 #endif
2771 if (current->set_child_tid)
2772 put_user(task_pid_vnr(current), current->set_child_tid);
2776 * context_switch - switch to the new MM and the new
2777 * thread's register state.
2779 static inline void
2780 context_switch(struct rq *rq, struct task_struct *prev,
2781 struct task_struct *next)
2783 struct mm_struct *mm, *oldmm;
2785 prepare_task_switch(rq, prev, next);
2786 mm = next->mm;
2787 oldmm = prev->active_mm;
2789 * For paravirt, this is coupled with an exit in switch_to to
2790 * combine the page table reload and the switch backend into
2791 * one hypercall.
2793 arch_enter_lazy_cpu_mode();
2795 if (unlikely(!mm)) {
2796 next->active_mm = oldmm;
2797 atomic_inc(&oldmm->mm_count);
2798 enter_lazy_tlb(oldmm, next);
2799 } else
2800 switch_mm(oldmm, mm, next);
2802 if (unlikely(!prev->mm)) {
2803 prev->active_mm = NULL;
2804 rq->prev_mm = oldmm;
2807 * Since the runqueue lock will be released by the next
2808 * task (which is an invalid locking op but in the case
2809 * of the scheduler it's an obvious special-case), so we
2810 * do an early lockdep release here:
2812 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2813 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2814 #endif
2816 /* Here we just switch the register state and the stack. */
2817 switch_to(prev, next, prev);
2819 barrier();
2821 * this_rq must be evaluated again because prev may have moved
2822 * CPUs since it called schedule(), thus the 'rq' on its stack
2823 * frame will be invalid.
2825 finish_task_switch(this_rq(), prev);
2829 * nr_running, nr_uninterruptible and nr_context_switches:
2831 * externally visible scheduler statistics: current number of runnable
2832 * threads, current number of uninterruptible-sleeping threads, total
2833 * number of context switches performed since bootup.
2835 unsigned long nr_running(void)
2837 unsigned long i, sum = 0;
2839 for_each_online_cpu(i)
2840 sum += cpu_rq(i)->nr_running;
2842 return sum;
2845 unsigned long nr_uninterruptible(void)
2847 unsigned long i, sum = 0;
2849 for_each_possible_cpu(i)
2850 sum += cpu_rq(i)->nr_uninterruptible;
2853 * Since we read the counters lockless, it might be slightly
2854 * inaccurate. Do not allow it to go below zero though:
2856 if (unlikely((long)sum < 0))
2857 sum = 0;
2859 return sum;
2862 unsigned long long nr_context_switches(void)
2864 int i;
2865 unsigned long long sum = 0;
2867 for_each_possible_cpu(i)
2868 sum += cpu_rq(i)->nr_switches;
2870 return sum;
2873 unsigned long nr_iowait(void)
2875 unsigned long i, sum = 0;
2877 for_each_possible_cpu(i)
2878 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2880 return sum;
2883 unsigned long nr_active(void)
2885 unsigned long i, running = 0, uninterruptible = 0;
2887 for_each_online_cpu(i) {
2888 running += cpu_rq(i)->nr_running;
2889 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2892 if (unlikely((long)uninterruptible < 0))
2893 uninterruptible = 0;
2895 return running + uninterruptible;
2899 * Update rq->cpu_load[] statistics. This function is usually called every
2900 * scheduler tick (TICK_NSEC).
2902 static void update_cpu_load(struct rq *this_rq)
2904 unsigned long this_load = this_rq->load.weight;
2905 int i, scale;
2907 this_rq->nr_load_updates++;
2909 /* Update our load: */
2910 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2911 unsigned long old_load, new_load;
2913 /* scale is effectively 1 << i now, and >> i divides by scale */
2915 old_load = this_rq->cpu_load[i];
2916 new_load = this_load;
2918 * Round up the averaging division if load is increasing. This
2919 * prevents us from getting stuck on 9 if the load is 10, for
2920 * example.
2922 if (new_load > old_load)
2923 new_load += scale-1;
2924 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2928 #ifdef CONFIG_SMP
2931 * double_rq_lock - safely lock two runqueues
2933 * Note this does not disable interrupts like task_rq_lock,
2934 * you need to do so manually before calling.
2936 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2937 __acquires(rq1->lock)
2938 __acquires(rq2->lock)
2940 BUG_ON(!irqs_disabled());
2941 if (rq1 == rq2) {
2942 spin_lock(&rq1->lock);
2943 __acquire(rq2->lock); /* Fake it out ;) */
2944 } else {
2945 if (rq1 < rq2) {
2946 spin_lock(&rq1->lock);
2947 spin_lock(&rq2->lock);
2948 } else {
2949 spin_lock(&rq2->lock);
2950 spin_lock(&rq1->lock);
2953 update_rq_clock(rq1);
2954 update_rq_clock(rq2);
2958 * double_rq_unlock - safely unlock two runqueues
2960 * Note this does not restore interrupts like task_rq_unlock,
2961 * you need to do so manually after calling.
2963 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2964 __releases(rq1->lock)
2965 __releases(rq2->lock)
2967 spin_unlock(&rq1->lock);
2968 if (rq1 != rq2)
2969 spin_unlock(&rq2->lock);
2970 else
2971 __release(rq2->lock);
2975 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2977 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2978 __releases(this_rq->lock)
2979 __acquires(busiest->lock)
2980 __acquires(this_rq->lock)
2982 int ret = 0;
2984 if (unlikely(!irqs_disabled())) {
2985 /* printk() doesn't work good under rq->lock */
2986 spin_unlock(&this_rq->lock);
2987 BUG_ON(1);
2989 if (unlikely(!spin_trylock(&busiest->lock))) {
2990 if (busiest < this_rq) {
2991 spin_unlock(&this_rq->lock);
2992 spin_lock(&busiest->lock);
2993 spin_lock(&this_rq->lock);
2994 ret = 1;
2995 } else
2996 spin_lock(&busiest->lock);
2998 return ret;
3002 * If dest_cpu is allowed for this process, migrate the task to it.
3003 * This is accomplished by forcing the cpu_allowed mask to only
3004 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3005 * the cpu_allowed mask is restored.
3007 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3009 struct migration_req req;
3010 unsigned long flags;
3011 struct rq *rq;
3013 rq = task_rq_lock(p, &flags);
3014 if (!cpu_isset(dest_cpu, p->cpus_allowed)
3015 || unlikely(cpu_is_offline(dest_cpu)))
3016 goto out;
3018 /* force the process onto the specified CPU */
3019 if (migrate_task(p, dest_cpu, &req)) {
3020 /* Need to wait for migration thread (might exit: take ref). */
3021 struct task_struct *mt = rq->migration_thread;
3023 get_task_struct(mt);
3024 task_rq_unlock(rq, &flags);
3025 wake_up_process(mt);
3026 put_task_struct(mt);
3027 wait_for_completion(&req.done);
3029 return;
3031 out:
3032 task_rq_unlock(rq, &flags);
3036 * sched_exec - execve() is a valuable balancing opportunity, because at
3037 * this point the task has the smallest effective memory and cache footprint.
3039 void sched_exec(void)
3041 int new_cpu, this_cpu = get_cpu();
3042 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3043 put_cpu();
3044 if (new_cpu != this_cpu)
3045 sched_migrate_task(current, new_cpu);
3049 * pull_task - move a task from a remote runqueue to the local runqueue.
3050 * Both runqueues must be locked.
3052 static void pull_task(struct rq *src_rq, struct task_struct *p,
3053 struct rq *this_rq, int this_cpu)
3055 deactivate_task(src_rq, p, 0);
3056 set_task_cpu(p, this_cpu);
3057 activate_task(this_rq, p, 0);
3059 * Note that idle threads have a prio of MAX_PRIO, for this test
3060 * to be always true for them.
3062 check_preempt_curr(this_rq, p);
3066 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3068 static
3069 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3070 struct sched_domain *sd, enum cpu_idle_type idle,
3071 int *all_pinned)
3074 * We do not migrate tasks that are:
3075 * 1) running (obviously), or
3076 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3077 * 3) are cache-hot on their current CPU.
3079 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
3080 schedstat_inc(p, se.nr_failed_migrations_affine);
3081 return 0;
3083 *all_pinned = 0;
3085 if (task_running(rq, p)) {
3086 schedstat_inc(p, se.nr_failed_migrations_running);
3087 return 0;
3091 * Aggressive migration if:
3092 * 1) task is cache cold, or
3093 * 2) too many balance attempts have failed.
3096 if (!task_hot(p, rq->clock, sd) ||
3097 sd->nr_balance_failed > sd->cache_nice_tries) {
3098 #ifdef CONFIG_SCHEDSTATS
3099 if (task_hot(p, rq->clock, sd)) {
3100 schedstat_inc(sd, lb_hot_gained[idle]);
3101 schedstat_inc(p, se.nr_forced_migrations);
3103 #endif
3104 return 1;
3107 if (task_hot(p, rq->clock, sd)) {
3108 schedstat_inc(p, se.nr_failed_migrations_hot);
3109 return 0;
3111 return 1;
3114 static unsigned long
3115 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3116 unsigned long max_load_move, struct sched_domain *sd,
3117 enum cpu_idle_type idle, int *all_pinned,
3118 int *this_best_prio, struct rq_iterator *iterator)
3120 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
3121 struct task_struct *p;
3122 long rem_load_move = max_load_move;
3124 if (max_load_move == 0)
3125 goto out;
3127 pinned = 1;
3130 * Start the load-balancing iterator:
3132 p = iterator->start(iterator->arg);
3133 next:
3134 if (!p || loops++ > sysctl_sched_nr_migrate)
3135 goto out;
3137 * To help distribute high priority tasks across CPUs we don't
3138 * skip a task if it will be the highest priority task (i.e. smallest
3139 * prio value) on its new queue regardless of its load weight
3141 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
3142 SCHED_LOAD_SCALE_FUZZ;
3143 if ((skip_for_load && p->prio >= *this_best_prio) ||
3144 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3145 p = iterator->next(iterator->arg);
3146 goto next;
3149 pull_task(busiest, p, this_rq, this_cpu);
3150 pulled++;
3151 rem_load_move -= p->se.load.weight;
3154 * We only want to steal up to the prescribed amount of weighted load.
3156 if (rem_load_move > 0) {
3157 if (p->prio < *this_best_prio)
3158 *this_best_prio = p->prio;
3159 p = iterator->next(iterator->arg);
3160 goto next;
3162 out:
3164 * Right now, this is one of only two places pull_task() is called,
3165 * so we can safely collect pull_task() stats here rather than
3166 * inside pull_task().
3168 schedstat_add(sd, lb_gained[idle], pulled);
3170 if (all_pinned)
3171 *all_pinned = pinned;
3173 return max_load_move - rem_load_move;
3177 * move_tasks tries to move up to max_load_move weighted load from busiest to
3178 * this_rq, as part of a balancing operation within domain "sd".
3179 * Returns 1 if successful and 0 otherwise.
3181 * Called with both runqueues locked.
3183 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3184 unsigned long max_load_move,
3185 struct sched_domain *sd, enum cpu_idle_type idle,
3186 int *all_pinned)
3188 const struct sched_class *class = sched_class_highest;
3189 unsigned long total_load_moved = 0;
3190 int this_best_prio = this_rq->curr->prio;
3192 do {
3193 total_load_moved +=
3194 class->load_balance(this_rq, this_cpu, busiest,
3195 max_load_move - total_load_moved,
3196 sd, idle, all_pinned, &this_best_prio);
3197 class = class->next;
3198 } while (class && max_load_move > total_load_moved);
3200 return total_load_moved > 0;
3203 static int
3204 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3205 struct sched_domain *sd, enum cpu_idle_type idle,
3206 struct rq_iterator *iterator)
3208 struct task_struct *p = iterator->start(iterator->arg);
3209 int pinned = 0;
3211 while (p) {
3212 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3213 pull_task(busiest, p, this_rq, this_cpu);
3215 * Right now, this is only the second place pull_task()
3216 * is called, so we can safely collect pull_task()
3217 * stats here rather than inside pull_task().
3219 schedstat_inc(sd, lb_gained[idle]);
3221 return 1;
3223 p = iterator->next(iterator->arg);
3226 return 0;
3230 * move_one_task tries to move exactly one task from busiest to this_rq, as
3231 * part of active balancing operations within "domain".
3232 * Returns 1 if successful and 0 otherwise.
3234 * Called with both runqueues locked.
3236 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3237 struct sched_domain *sd, enum cpu_idle_type idle)
3239 const struct sched_class *class;
3241 for (class = sched_class_highest; class; class = class->next)
3242 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3243 return 1;
3245 return 0;
3249 * find_busiest_group finds and returns the busiest CPU group within the
3250 * domain. It calculates and returns the amount of weighted load which
3251 * should be moved to restore balance via the imbalance parameter.
3253 static struct sched_group *
3254 find_busiest_group(struct sched_domain *sd, int this_cpu,
3255 unsigned long *imbalance, enum cpu_idle_type idle,
3256 int *sd_idle, const cpumask_t *cpus, int *balance)
3258 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3259 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3260 unsigned long max_pull;
3261 unsigned long busiest_load_per_task, busiest_nr_running;
3262 unsigned long this_load_per_task, this_nr_running;
3263 int load_idx, group_imb = 0;
3264 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3265 int power_savings_balance = 1;
3266 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3267 unsigned long min_nr_running = ULONG_MAX;
3268 struct sched_group *group_min = NULL, *group_leader = NULL;
3269 #endif
3271 max_load = this_load = total_load = total_pwr = 0;
3272 busiest_load_per_task = busiest_nr_running = 0;
3273 this_load_per_task = this_nr_running = 0;
3274 if (idle == CPU_NOT_IDLE)
3275 load_idx = sd->busy_idx;
3276 else if (idle == CPU_NEWLY_IDLE)
3277 load_idx = sd->newidle_idx;
3278 else
3279 load_idx = sd->idle_idx;
3281 do {
3282 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3283 int local_group;
3284 int i;
3285 int __group_imb = 0;
3286 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3287 unsigned long sum_nr_running, sum_weighted_load;
3289 local_group = cpu_isset(this_cpu, group->cpumask);
3291 if (local_group)
3292 balance_cpu = first_cpu(group->cpumask);
3294 /* Tally up the load of all CPUs in the group */
3295 sum_weighted_load = sum_nr_running = avg_load = 0;
3296 max_cpu_load = 0;
3297 min_cpu_load = ~0UL;
3299 for_each_cpu_mask(i, group->cpumask) {
3300 struct rq *rq;
3302 if (!cpu_isset(i, *cpus))
3303 continue;
3305 rq = cpu_rq(i);
3307 if (*sd_idle && rq->nr_running)
3308 *sd_idle = 0;
3310 /* Bias balancing toward cpus of our domain */
3311 if (local_group) {
3312 if (idle_cpu(i) && !first_idle_cpu) {
3313 first_idle_cpu = 1;
3314 balance_cpu = i;
3317 load = target_load(i, load_idx);
3318 } else {
3319 load = source_load(i, load_idx);
3320 if (load > max_cpu_load)
3321 max_cpu_load = load;
3322 if (min_cpu_load > load)
3323 min_cpu_load = load;
3326 avg_load += load;
3327 sum_nr_running += rq->nr_running;
3328 sum_weighted_load += weighted_cpuload(i);
3332 * First idle cpu or the first cpu(busiest) in this sched group
3333 * is eligible for doing load balancing at this and above
3334 * domains. In the newly idle case, we will allow all the cpu's
3335 * to do the newly idle load balance.
3337 if (idle != CPU_NEWLY_IDLE && local_group &&
3338 balance_cpu != this_cpu && balance) {
3339 *balance = 0;
3340 goto ret;
3343 total_load += avg_load;
3344 total_pwr += group->__cpu_power;
3346 /* Adjust by relative CPU power of the group */
3347 avg_load = sg_div_cpu_power(group,
3348 avg_load * SCHED_LOAD_SCALE);
3350 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3351 __group_imb = 1;
3353 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3355 if (local_group) {
3356 this_load = avg_load;
3357 this = group;
3358 this_nr_running = sum_nr_running;
3359 this_load_per_task = sum_weighted_load;
3360 } else if (avg_load > max_load &&
3361 (sum_nr_running > group_capacity || __group_imb)) {
3362 max_load = avg_load;
3363 busiest = group;
3364 busiest_nr_running = sum_nr_running;
3365 busiest_load_per_task = sum_weighted_load;
3366 group_imb = __group_imb;
3369 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3371 * Busy processors will not participate in power savings
3372 * balance.
3374 if (idle == CPU_NOT_IDLE ||
3375 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3376 goto group_next;
3379 * If the local group is idle or completely loaded
3380 * no need to do power savings balance at this domain
3382 if (local_group && (this_nr_running >= group_capacity ||
3383 !this_nr_running))
3384 power_savings_balance = 0;
3387 * If a group is already running at full capacity or idle,
3388 * don't include that group in power savings calculations
3390 if (!power_savings_balance || sum_nr_running >= group_capacity
3391 || !sum_nr_running)
3392 goto group_next;
3395 * Calculate the group which has the least non-idle load.
3396 * This is the group from where we need to pick up the load
3397 * for saving power
3399 if ((sum_nr_running < min_nr_running) ||
3400 (sum_nr_running == min_nr_running &&
3401 first_cpu(group->cpumask) <
3402 first_cpu(group_min->cpumask))) {
3403 group_min = group;
3404 min_nr_running = sum_nr_running;
3405 min_load_per_task = sum_weighted_load /
3406 sum_nr_running;
3410 * Calculate the group which is almost near its
3411 * capacity but still has some space to pick up some load
3412 * from other group and save more power
3414 if (sum_nr_running <= group_capacity - 1) {
3415 if (sum_nr_running > leader_nr_running ||
3416 (sum_nr_running == leader_nr_running &&
3417 first_cpu(group->cpumask) >
3418 first_cpu(group_leader->cpumask))) {
3419 group_leader = group;
3420 leader_nr_running = sum_nr_running;
3423 group_next:
3424 #endif
3425 group = group->next;
3426 } while (group != sd->groups);
3428 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3429 goto out_balanced;
3431 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3433 if (this_load >= avg_load ||
3434 100*max_load <= sd->imbalance_pct*this_load)
3435 goto out_balanced;
3437 busiest_load_per_task /= busiest_nr_running;
3438 if (group_imb)
3439 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3442 * We're trying to get all the cpus to the average_load, so we don't
3443 * want to push ourselves above the average load, nor do we wish to
3444 * reduce the max loaded cpu below the average load, as either of these
3445 * actions would just result in more rebalancing later, and ping-pong
3446 * tasks around. Thus we look for the minimum possible imbalance.
3447 * Negative imbalances (*we* are more loaded than anyone else) will
3448 * be counted as no imbalance for these purposes -- we can't fix that
3449 * by pulling tasks to us. Be careful of negative numbers as they'll
3450 * appear as very large values with unsigned longs.
3452 if (max_load <= busiest_load_per_task)
3453 goto out_balanced;
3456 * In the presence of smp nice balancing, certain scenarios can have
3457 * max load less than avg load(as we skip the groups at or below
3458 * its cpu_power, while calculating max_load..)
3460 if (max_load < avg_load) {
3461 *imbalance = 0;
3462 goto small_imbalance;
3465 /* Don't want to pull so many tasks that a group would go idle */
3466 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3468 /* How much load to actually move to equalise the imbalance */
3469 *imbalance = min(max_pull * busiest->__cpu_power,
3470 (avg_load - this_load) * this->__cpu_power)
3471 / SCHED_LOAD_SCALE;
3474 * if *imbalance is less than the average load per runnable task
3475 * there is no gaurantee that any tasks will be moved so we'll have
3476 * a think about bumping its value to force at least one task to be
3477 * moved
3479 if (*imbalance < busiest_load_per_task) {
3480 unsigned long tmp, pwr_now, pwr_move;
3481 unsigned int imbn;
3483 small_imbalance:
3484 pwr_move = pwr_now = 0;
3485 imbn = 2;
3486 if (this_nr_running) {
3487 this_load_per_task /= this_nr_running;
3488 if (busiest_load_per_task > this_load_per_task)
3489 imbn = 1;
3490 } else
3491 this_load_per_task = SCHED_LOAD_SCALE;
3493 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3494 busiest_load_per_task * imbn) {
3495 *imbalance = busiest_load_per_task;
3496 return busiest;
3500 * OK, we don't have enough imbalance to justify moving tasks,
3501 * however we may be able to increase total CPU power used by
3502 * moving them.
3505 pwr_now += busiest->__cpu_power *
3506 min(busiest_load_per_task, max_load);
3507 pwr_now += this->__cpu_power *
3508 min(this_load_per_task, this_load);
3509 pwr_now /= SCHED_LOAD_SCALE;
3511 /* Amount of load we'd subtract */
3512 tmp = sg_div_cpu_power(busiest,
3513 busiest_load_per_task * SCHED_LOAD_SCALE);
3514 if (max_load > tmp)
3515 pwr_move += busiest->__cpu_power *
3516 min(busiest_load_per_task, max_load - tmp);
3518 /* Amount of load we'd add */
3519 if (max_load * busiest->__cpu_power <
3520 busiest_load_per_task * SCHED_LOAD_SCALE)
3521 tmp = sg_div_cpu_power(this,
3522 max_load * busiest->__cpu_power);
3523 else
3524 tmp = sg_div_cpu_power(this,
3525 busiest_load_per_task * SCHED_LOAD_SCALE);
3526 pwr_move += this->__cpu_power *
3527 min(this_load_per_task, this_load + tmp);
3528 pwr_move /= SCHED_LOAD_SCALE;
3530 /* Move if we gain throughput */
3531 if (pwr_move > pwr_now)
3532 *imbalance = busiest_load_per_task;
3535 return busiest;
3537 out_balanced:
3538 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3539 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3540 goto ret;
3542 if (this == group_leader && group_leader != group_min) {
3543 *imbalance = min_load_per_task;
3544 return group_min;
3546 #endif
3547 ret:
3548 *imbalance = 0;
3549 return NULL;
3553 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3555 static struct rq *
3556 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3557 unsigned long imbalance, const cpumask_t *cpus)
3559 struct rq *busiest = NULL, *rq;
3560 unsigned long max_load = 0;
3561 int i;
3563 for_each_cpu_mask(i, group->cpumask) {
3564 unsigned long wl;
3566 if (!cpu_isset(i, *cpus))
3567 continue;
3569 rq = cpu_rq(i);
3570 wl = weighted_cpuload(i);
3572 if (rq->nr_running == 1 && wl > imbalance)
3573 continue;
3575 if (wl > max_load) {
3576 max_load = wl;
3577 busiest = rq;
3581 return busiest;
3585 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3586 * so long as it is large enough.
3588 #define MAX_PINNED_INTERVAL 512
3591 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3592 * tasks if there is an imbalance.
3594 static int load_balance(int this_cpu, struct rq *this_rq,
3595 struct sched_domain *sd, enum cpu_idle_type idle,
3596 int *balance, cpumask_t *cpus)
3598 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3599 struct sched_group *group;
3600 unsigned long imbalance;
3601 struct rq *busiest;
3602 unsigned long flags;
3603 int unlock_aggregate;
3605 cpus_setall(*cpus);
3607 unlock_aggregate = get_aggregate(sd);
3610 * When power savings policy is enabled for the parent domain, idle
3611 * sibling can pick up load irrespective of busy siblings. In this case,
3612 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3613 * portraying it as CPU_NOT_IDLE.
3615 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3616 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3617 sd_idle = 1;
3619 schedstat_inc(sd, lb_count[idle]);
3621 redo:
3622 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3623 cpus, balance);
3625 if (*balance == 0)
3626 goto out_balanced;
3628 if (!group) {
3629 schedstat_inc(sd, lb_nobusyg[idle]);
3630 goto out_balanced;
3633 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3634 if (!busiest) {
3635 schedstat_inc(sd, lb_nobusyq[idle]);
3636 goto out_balanced;
3639 BUG_ON(busiest == this_rq);
3641 schedstat_add(sd, lb_imbalance[idle], imbalance);
3643 ld_moved = 0;
3644 if (busiest->nr_running > 1) {
3646 * Attempt to move tasks. If find_busiest_group has found
3647 * an imbalance but busiest->nr_running <= 1, the group is
3648 * still unbalanced. ld_moved simply stays zero, so it is
3649 * correctly treated as an imbalance.
3651 local_irq_save(flags);
3652 double_rq_lock(this_rq, busiest);
3653 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3654 imbalance, sd, idle, &all_pinned);
3655 double_rq_unlock(this_rq, busiest);
3656 local_irq_restore(flags);
3659 * some other cpu did the load balance for us.
3661 if (ld_moved && this_cpu != smp_processor_id())
3662 resched_cpu(this_cpu);
3664 /* All tasks on this runqueue were pinned by CPU affinity */
3665 if (unlikely(all_pinned)) {
3666 cpu_clear(cpu_of(busiest), *cpus);
3667 if (!cpus_empty(*cpus))
3668 goto redo;
3669 goto out_balanced;
3673 if (!ld_moved) {
3674 schedstat_inc(sd, lb_failed[idle]);
3675 sd->nr_balance_failed++;
3677 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3679 spin_lock_irqsave(&busiest->lock, flags);
3681 /* don't kick the migration_thread, if the curr
3682 * task on busiest cpu can't be moved to this_cpu
3684 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3685 spin_unlock_irqrestore(&busiest->lock, flags);
3686 all_pinned = 1;
3687 goto out_one_pinned;
3690 if (!busiest->active_balance) {
3691 busiest->active_balance = 1;
3692 busiest->push_cpu = this_cpu;
3693 active_balance = 1;
3695 spin_unlock_irqrestore(&busiest->lock, flags);
3696 if (active_balance)
3697 wake_up_process(busiest->migration_thread);
3700 * We've kicked active balancing, reset the failure
3701 * counter.
3703 sd->nr_balance_failed = sd->cache_nice_tries+1;
3705 } else
3706 sd->nr_balance_failed = 0;
3708 if (likely(!active_balance)) {
3709 /* We were unbalanced, so reset the balancing interval */
3710 sd->balance_interval = sd->min_interval;
3711 } else {
3713 * If we've begun active balancing, start to back off. This
3714 * case may not be covered by the all_pinned logic if there
3715 * is only 1 task on the busy runqueue (because we don't call
3716 * move_tasks).
3718 if (sd->balance_interval < sd->max_interval)
3719 sd->balance_interval *= 2;
3722 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3723 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3724 ld_moved = -1;
3726 goto out;
3728 out_balanced:
3729 schedstat_inc(sd, lb_balanced[idle]);
3731 sd->nr_balance_failed = 0;
3733 out_one_pinned:
3734 /* tune up the balancing interval */
3735 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3736 (sd->balance_interval < sd->max_interval))
3737 sd->balance_interval *= 2;
3739 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3740 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3741 ld_moved = -1;
3742 else
3743 ld_moved = 0;
3744 out:
3745 if (unlock_aggregate)
3746 put_aggregate(sd);
3747 return ld_moved;
3751 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3752 * tasks if there is an imbalance.
3754 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3755 * this_rq is locked.
3757 static int
3758 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3759 cpumask_t *cpus)
3761 struct sched_group *group;
3762 struct rq *busiest = NULL;
3763 unsigned long imbalance;
3764 int ld_moved = 0;
3765 int sd_idle = 0;
3766 int all_pinned = 0;
3768 cpus_setall(*cpus);
3771 * When power savings policy is enabled for the parent domain, idle
3772 * sibling can pick up load irrespective of busy siblings. In this case,
3773 * let the state of idle sibling percolate up as IDLE, instead of
3774 * portraying it as CPU_NOT_IDLE.
3776 if (sd->flags & SD_SHARE_CPUPOWER &&
3777 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3778 sd_idle = 1;
3780 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3781 redo:
3782 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3783 &sd_idle, cpus, NULL);
3784 if (!group) {
3785 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3786 goto out_balanced;
3789 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3790 if (!busiest) {
3791 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3792 goto out_balanced;
3795 BUG_ON(busiest == this_rq);
3797 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3799 ld_moved = 0;
3800 if (busiest->nr_running > 1) {
3801 /* Attempt to move tasks */
3802 double_lock_balance(this_rq, busiest);
3803 /* this_rq->clock is already updated */
3804 update_rq_clock(busiest);
3805 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3806 imbalance, sd, CPU_NEWLY_IDLE,
3807 &all_pinned);
3808 spin_unlock(&busiest->lock);
3810 if (unlikely(all_pinned)) {
3811 cpu_clear(cpu_of(busiest), *cpus);
3812 if (!cpus_empty(*cpus))
3813 goto redo;
3817 if (!ld_moved) {
3818 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3819 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3820 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3821 return -1;
3822 } else
3823 sd->nr_balance_failed = 0;
3825 return ld_moved;
3827 out_balanced:
3828 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3829 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3830 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3831 return -1;
3832 sd->nr_balance_failed = 0;
3834 return 0;
3838 * idle_balance is called by schedule() if this_cpu is about to become
3839 * idle. Attempts to pull tasks from other CPUs.
3841 static void idle_balance(int this_cpu, struct rq *this_rq)
3843 struct sched_domain *sd;
3844 int pulled_task = -1;
3845 unsigned long next_balance = jiffies + HZ;
3846 cpumask_t tmpmask;
3848 for_each_domain(this_cpu, sd) {
3849 unsigned long interval;
3851 if (!(sd->flags & SD_LOAD_BALANCE))
3852 continue;
3854 if (sd->flags & SD_BALANCE_NEWIDLE)
3855 /* If we've pulled tasks over stop searching: */
3856 pulled_task = load_balance_newidle(this_cpu, this_rq,
3857 sd, &tmpmask);
3859 interval = msecs_to_jiffies(sd->balance_interval);
3860 if (time_after(next_balance, sd->last_balance + interval))
3861 next_balance = sd->last_balance + interval;
3862 if (pulled_task)
3863 break;
3865 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3867 * We are going idle. next_balance may be set based on
3868 * a busy processor. So reset next_balance.
3870 this_rq->next_balance = next_balance;
3875 * active_load_balance is run by migration threads. It pushes running tasks
3876 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3877 * running on each physical CPU where possible, and avoids physical /
3878 * logical imbalances.
3880 * Called with busiest_rq locked.
3882 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3884 int target_cpu = busiest_rq->push_cpu;
3885 struct sched_domain *sd;
3886 struct rq *target_rq;
3888 /* Is there any task to move? */
3889 if (busiest_rq->nr_running <= 1)
3890 return;
3892 target_rq = cpu_rq(target_cpu);
3895 * This condition is "impossible", if it occurs
3896 * we need to fix it. Originally reported by
3897 * Bjorn Helgaas on a 128-cpu setup.
3899 BUG_ON(busiest_rq == target_rq);
3901 /* move a task from busiest_rq to target_rq */
3902 double_lock_balance(busiest_rq, target_rq);
3903 update_rq_clock(busiest_rq);
3904 update_rq_clock(target_rq);
3906 /* Search for an sd spanning us and the target CPU. */
3907 for_each_domain(target_cpu, sd) {
3908 if ((sd->flags & SD_LOAD_BALANCE) &&
3909 cpu_isset(busiest_cpu, sd->span))
3910 break;
3913 if (likely(sd)) {
3914 schedstat_inc(sd, alb_count);
3916 if (move_one_task(target_rq, target_cpu, busiest_rq,
3917 sd, CPU_IDLE))
3918 schedstat_inc(sd, alb_pushed);
3919 else
3920 schedstat_inc(sd, alb_failed);
3922 spin_unlock(&target_rq->lock);
3925 #ifdef CONFIG_NO_HZ
3926 static struct {
3927 atomic_t load_balancer;
3928 cpumask_t cpu_mask;
3929 } nohz ____cacheline_aligned = {
3930 .load_balancer = ATOMIC_INIT(-1),
3931 .cpu_mask = CPU_MASK_NONE,
3935 * This routine will try to nominate the ilb (idle load balancing)
3936 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3937 * load balancing on behalf of all those cpus. If all the cpus in the system
3938 * go into this tickless mode, then there will be no ilb owner (as there is
3939 * no need for one) and all the cpus will sleep till the next wakeup event
3940 * arrives...
3942 * For the ilb owner, tick is not stopped. And this tick will be used
3943 * for idle load balancing. ilb owner will still be part of
3944 * nohz.cpu_mask..
3946 * While stopping the tick, this cpu will become the ilb owner if there
3947 * is no other owner. And will be the owner till that cpu becomes busy
3948 * or if all cpus in the system stop their ticks at which point
3949 * there is no need for ilb owner.
3951 * When the ilb owner becomes busy, it nominates another owner, during the
3952 * next busy scheduler_tick()
3954 int select_nohz_load_balancer(int stop_tick)
3956 int cpu = smp_processor_id();
3958 if (stop_tick) {
3959 cpu_set(cpu, nohz.cpu_mask);
3960 cpu_rq(cpu)->in_nohz_recently = 1;
3963 * If we are going offline and still the leader, give up!
3965 if (cpu_is_offline(cpu) &&
3966 atomic_read(&nohz.load_balancer) == cpu) {
3967 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3968 BUG();
3969 return 0;
3972 /* time for ilb owner also to sleep */
3973 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3974 if (atomic_read(&nohz.load_balancer) == cpu)
3975 atomic_set(&nohz.load_balancer, -1);
3976 return 0;
3979 if (atomic_read(&nohz.load_balancer) == -1) {
3980 /* make me the ilb owner */
3981 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3982 return 1;
3983 } else if (atomic_read(&nohz.load_balancer) == cpu)
3984 return 1;
3985 } else {
3986 if (!cpu_isset(cpu, nohz.cpu_mask))
3987 return 0;
3989 cpu_clear(cpu, nohz.cpu_mask);
3991 if (atomic_read(&nohz.load_balancer) == cpu)
3992 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3993 BUG();
3995 return 0;
3997 #endif
3999 static DEFINE_SPINLOCK(balancing);
4002 * It checks each scheduling domain to see if it is due to be balanced,
4003 * and initiates a balancing operation if so.
4005 * Balancing parameters are set up in arch_init_sched_domains.
4007 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4009 int balance = 1;
4010 struct rq *rq = cpu_rq(cpu);
4011 unsigned long interval;
4012 struct sched_domain *sd;
4013 /* Earliest time when we have to do rebalance again */
4014 unsigned long next_balance = jiffies + 60*HZ;
4015 int update_next_balance = 0;
4016 cpumask_t tmp;
4018 for_each_domain(cpu, sd) {
4019 if (!(sd->flags & SD_LOAD_BALANCE))
4020 continue;
4022 interval = sd->balance_interval;
4023 if (idle != CPU_IDLE)
4024 interval *= sd->busy_factor;
4026 /* scale ms to jiffies */
4027 interval = msecs_to_jiffies(interval);
4028 if (unlikely(!interval))
4029 interval = 1;
4030 if (interval > HZ*NR_CPUS/10)
4031 interval = HZ*NR_CPUS/10;
4034 if (sd->flags & SD_SERIALIZE) {
4035 if (!spin_trylock(&balancing))
4036 goto out;
4039 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4040 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
4042 * We've pulled tasks over so either we're no
4043 * longer idle, or one of our SMT siblings is
4044 * not idle.
4046 idle = CPU_NOT_IDLE;
4048 sd->last_balance = jiffies;
4050 if (sd->flags & SD_SERIALIZE)
4051 spin_unlock(&balancing);
4052 out:
4053 if (time_after(next_balance, sd->last_balance + interval)) {
4054 next_balance = sd->last_balance + interval;
4055 update_next_balance = 1;
4059 * Stop the load balance at this level. There is another
4060 * CPU in our sched group which is doing load balancing more
4061 * actively.
4063 if (!balance)
4064 break;
4068 * next_balance will be updated only when there is a need.
4069 * When the cpu is attached to null domain for ex, it will not be
4070 * updated.
4072 if (likely(update_next_balance))
4073 rq->next_balance = next_balance;
4077 * run_rebalance_domains is triggered when needed from the scheduler tick.
4078 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4079 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4081 static void run_rebalance_domains(struct softirq_action *h)
4083 int this_cpu = smp_processor_id();
4084 struct rq *this_rq = cpu_rq(this_cpu);
4085 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4086 CPU_IDLE : CPU_NOT_IDLE;
4088 rebalance_domains(this_cpu, idle);
4090 #ifdef CONFIG_NO_HZ
4092 * If this cpu is the owner for idle load balancing, then do the
4093 * balancing on behalf of the other idle cpus whose ticks are
4094 * stopped.
4096 if (this_rq->idle_at_tick &&
4097 atomic_read(&nohz.load_balancer) == this_cpu) {
4098 cpumask_t cpus = nohz.cpu_mask;
4099 struct rq *rq;
4100 int balance_cpu;
4102 cpu_clear(this_cpu, cpus);
4103 for_each_cpu_mask(balance_cpu, cpus) {
4105 * If this cpu gets work to do, stop the load balancing
4106 * work being done for other cpus. Next load
4107 * balancing owner will pick it up.
4109 if (need_resched())
4110 break;
4112 rebalance_domains(balance_cpu, CPU_IDLE);
4114 rq = cpu_rq(balance_cpu);
4115 if (time_after(this_rq->next_balance, rq->next_balance))
4116 this_rq->next_balance = rq->next_balance;
4119 #endif
4123 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4125 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4126 * idle load balancing owner or decide to stop the periodic load balancing,
4127 * if the whole system is idle.
4129 static inline void trigger_load_balance(struct rq *rq, int cpu)
4131 #ifdef CONFIG_NO_HZ
4133 * If we were in the nohz mode recently and busy at the current
4134 * scheduler tick, then check if we need to nominate new idle
4135 * load balancer.
4137 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4138 rq->in_nohz_recently = 0;
4140 if (atomic_read(&nohz.load_balancer) == cpu) {
4141 cpu_clear(cpu, nohz.cpu_mask);
4142 atomic_set(&nohz.load_balancer, -1);
4145 if (atomic_read(&nohz.load_balancer) == -1) {
4147 * simple selection for now: Nominate the
4148 * first cpu in the nohz list to be the next
4149 * ilb owner.
4151 * TBD: Traverse the sched domains and nominate
4152 * the nearest cpu in the nohz.cpu_mask.
4154 int ilb = first_cpu(nohz.cpu_mask);
4156 if (ilb < nr_cpu_ids)
4157 resched_cpu(ilb);
4162 * If this cpu is idle and doing idle load balancing for all the
4163 * cpus with ticks stopped, is it time for that to stop?
4165 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4166 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4167 resched_cpu(cpu);
4168 return;
4172 * If this cpu is idle and the idle load balancing is done by
4173 * someone else, then no need raise the SCHED_SOFTIRQ
4175 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4176 cpu_isset(cpu, nohz.cpu_mask))
4177 return;
4178 #endif
4179 if (time_after_eq(jiffies, rq->next_balance))
4180 raise_softirq(SCHED_SOFTIRQ);
4183 #else /* CONFIG_SMP */
4186 * on UP we do not need to balance between CPUs:
4188 static inline void idle_balance(int cpu, struct rq *rq)
4192 #endif
4194 DEFINE_PER_CPU(struct kernel_stat, kstat);
4196 EXPORT_PER_CPU_SYMBOL(kstat);
4199 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4200 * that have not yet been banked in case the task is currently running.
4202 unsigned long long task_sched_runtime(struct task_struct *p)
4204 unsigned long flags;
4205 u64 ns, delta_exec;
4206 struct rq *rq;
4208 rq = task_rq_lock(p, &flags);
4209 ns = p->se.sum_exec_runtime;
4210 if (task_current(rq, p)) {
4211 update_rq_clock(rq);
4212 delta_exec = rq->clock - p->se.exec_start;
4213 if ((s64)delta_exec > 0)
4214 ns += delta_exec;
4216 task_rq_unlock(rq, &flags);
4218 return ns;
4222 * Account user cpu time to a process.
4223 * @p: the process that the cpu time gets accounted to
4224 * @cputime: the cpu time spent in user space since the last update
4226 void account_user_time(struct task_struct *p, cputime_t cputime)
4228 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4229 cputime64_t tmp;
4231 p->utime = cputime_add(p->utime, cputime);
4233 /* Add user time to cpustat. */
4234 tmp = cputime_to_cputime64(cputime);
4235 if (TASK_NICE(p) > 0)
4236 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4237 else
4238 cpustat->user = cputime64_add(cpustat->user, tmp);
4242 * Account guest cpu time to a process.
4243 * @p: the process that the cpu time gets accounted to
4244 * @cputime: the cpu time spent in virtual machine since the last update
4246 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4248 cputime64_t tmp;
4249 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4251 tmp = cputime_to_cputime64(cputime);
4253 p->utime = cputime_add(p->utime, cputime);
4254 p->gtime = cputime_add(p->gtime, cputime);
4256 cpustat->user = cputime64_add(cpustat->user, tmp);
4257 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4261 * Account scaled user cpu time to a process.
4262 * @p: the process that the cpu time gets accounted to
4263 * @cputime: the cpu time spent in user space since the last update
4265 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4267 p->utimescaled = cputime_add(p->utimescaled, cputime);
4271 * Account system cpu time to a process.
4272 * @p: the process that the cpu time gets accounted to
4273 * @hardirq_offset: the offset to subtract from hardirq_count()
4274 * @cputime: the cpu time spent in kernel space since the last update
4276 void account_system_time(struct task_struct *p, int hardirq_offset,
4277 cputime_t cputime)
4279 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4280 struct rq *rq = this_rq();
4281 cputime64_t tmp;
4283 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4284 account_guest_time(p, cputime);
4285 return;
4288 p->stime = cputime_add(p->stime, cputime);
4290 /* Add system time to cpustat. */
4291 tmp = cputime_to_cputime64(cputime);
4292 if (hardirq_count() - hardirq_offset)
4293 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4294 else if (softirq_count())
4295 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4296 else if (p != rq->idle)
4297 cpustat->system = cputime64_add(cpustat->system, tmp);
4298 else if (atomic_read(&rq->nr_iowait) > 0)
4299 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4300 else
4301 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4302 /* Account for system time used */
4303 acct_update_integrals(p);
4307 * Account scaled system cpu time to a process.
4308 * @p: the process that the cpu time gets accounted to
4309 * @hardirq_offset: the offset to subtract from hardirq_count()
4310 * @cputime: the cpu time spent in kernel space since the last update
4312 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4314 p->stimescaled = cputime_add(p->stimescaled, cputime);
4318 * Account for involuntary wait time.
4319 * @p: the process from which the cpu time has been stolen
4320 * @steal: the cpu time spent in involuntary wait
4322 void account_steal_time(struct task_struct *p, cputime_t steal)
4324 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4325 cputime64_t tmp = cputime_to_cputime64(steal);
4326 struct rq *rq = this_rq();
4328 if (p == rq->idle) {
4329 p->stime = cputime_add(p->stime, steal);
4330 if (atomic_read(&rq->nr_iowait) > 0)
4331 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4332 else
4333 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4334 } else
4335 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4339 * This function gets called by the timer code, with HZ frequency.
4340 * We call it with interrupts disabled.
4342 * It also gets called by the fork code, when changing the parent's
4343 * timeslices.
4345 void scheduler_tick(void)
4347 int cpu = smp_processor_id();
4348 struct rq *rq = cpu_rq(cpu);
4349 struct task_struct *curr = rq->curr;
4351 sched_clock_tick();
4353 spin_lock(&rq->lock);
4354 update_rq_clock(rq);
4355 update_cpu_load(rq);
4356 curr->sched_class->task_tick(rq, curr, 0);
4357 spin_unlock(&rq->lock);
4359 #ifdef CONFIG_SMP
4360 rq->idle_at_tick = idle_cpu(cpu);
4361 trigger_load_balance(rq, cpu);
4362 #endif
4365 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4367 void __kprobes add_preempt_count(int val)
4370 * Underflow?
4372 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4373 return;
4374 preempt_count() += val;
4376 * Spinlock count overflowing soon?
4378 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4379 PREEMPT_MASK - 10);
4381 EXPORT_SYMBOL(add_preempt_count);
4383 void __kprobes sub_preempt_count(int val)
4386 * Underflow?
4388 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4389 return;
4391 * Is the spinlock portion underflowing?
4393 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4394 !(preempt_count() & PREEMPT_MASK)))
4395 return;
4397 preempt_count() -= val;
4399 EXPORT_SYMBOL(sub_preempt_count);
4401 #endif
4404 * Print scheduling while atomic bug:
4406 static noinline void __schedule_bug(struct task_struct *prev)
4408 struct pt_regs *regs = get_irq_regs();
4410 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4411 prev->comm, prev->pid, preempt_count());
4413 debug_show_held_locks(prev);
4414 if (irqs_disabled())
4415 print_irqtrace_events(prev);
4417 if (regs)
4418 show_regs(regs);
4419 else
4420 dump_stack();
4424 * Various schedule()-time debugging checks and statistics:
4426 static inline void schedule_debug(struct task_struct *prev)
4429 * Test if we are atomic. Since do_exit() needs to call into
4430 * schedule() atomically, we ignore that path for now.
4431 * Otherwise, whine if we are scheduling when we should not be.
4433 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
4434 __schedule_bug(prev);
4436 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4438 schedstat_inc(this_rq(), sched_count);
4439 #ifdef CONFIG_SCHEDSTATS
4440 if (unlikely(prev->lock_depth >= 0)) {
4441 schedstat_inc(this_rq(), bkl_count);
4442 schedstat_inc(prev, sched_info.bkl_count);
4444 #endif
4448 * Pick up the highest-prio task:
4450 static inline struct task_struct *
4451 pick_next_task(struct rq *rq, struct task_struct *prev)
4453 const struct sched_class *class;
4454 struct task_struct *p;
4457 * Optimization: we know that if all tasks are in
4458 * the fair class we can call that function directly:
4460 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4461 p = fair_sched_class.pick_next_task(rq);
4462 if (likely(p))
4463 return p;
4466 class = sched_class_highest;
4467 for ( ; ; ) {
4468 p = class->pick_next_task(rq);
4469 if (p)
4470 return p;
4472 * Will never be NULL as the idle class always
4473 * returns a non-NULL p:
4475 class = class->next;
4480 * schedule() is the main scheduler function.
4482 asmlinkage void __sched schedule(void)
4484 struct task_struct *prev, *next;
4485 unsigned long *switch_count;
4486 struct rq *rq;
4487 int cpu;
4489 need_resched:
4490 preempt_disable();
4491 cpu = smp_processor_id();
4492 rq = cpu_rq(cpu);
4493 rcu_qsctr_inc(cpu);
4494 prev = rq->curr;
4495 switch_count = &prev->nivcsw;
4497 release_kernel_lock(prev);
4498 need_resched_nonpreemptible:
4500 schedule_debug(prev);
4502 hrtick_clear(rq);
4505 * Do the rq-clock update outside the rq lock:
4507 local_irq_disable();
4508 update_rq_clock(rq);
4509 spin_lock(&rq->lock);
4510 clear_tsk_need_resched(prev);
4512 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4513 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
4514 signal_pending(prev))) {
4515 prev->state = TASK_RUNNING;
4516 } else {
4517 deactivate_task(rq, prev, 1);
4519 switch_count = &prev->nvcsw;
4522 #ifdef CONFIG_SMP
4523 if (prev->sched_class->pre_schedule)
4524 prev->sched_class->pre_schedule(rq, prev);
4525 #endif
4527 if (unlikely(!rq->nr_running))
4528 idle_balance(cpu, rq);
4530 prev->sched_class->put_prev_task(rq, prev);
4531 next = pick_next_task(rq, prev);
4533 if (likely(prev != next)) {
4534 sched_info_switch(prev, next);
4536 rq->nr_switches++;
4537 rq->curr = next;
4538 ++*switch_count;
4540 context_switch(rq, prev, next); /* unlocks the rq */
4542 * the context switch might have flipped the stack from under
4543 * us, hence refresh the local variables.
4545 cpu = smp_processor_id();
4546 rq = cpu_rq(cpu);
4547 } else
4548 spin_unlock_irq(&rq->lock);
4550 hrtick_set(rq);
4552 if (unlikely(reacquire_kernel_lock(current) < 0))
4553 goto need_resched_nonpreemptible;
4555 preempt_enable_no_resched();
4556 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4557 goto need_resched;
4559 EXPORT_SYMBOL(schedule);
4561 #ifdef CONFIG_PREEMPT
4563 * this is the entry point to schedule() from in-kernel preemption
4564 * off of preempt_enable. Kernel preemptions off return from interrupt
4565 * occur there and call schedule directly.
4567 asmlinkage void __sched preempt_schedule(void)
4569 struct thread_info *ti = current_thread_info();
4572 * If there is a non-zero preempt_count or interrupts are disabled,
4573 * we do not want to preempt the current task. Just return..
4575 if (likely(ti->preempt_count || irqs_disabled()))
4576 return;
4578 do {
4579 add_preempt_count(PREEMPT_ACTIVE);
4580 schedule();
4581 sub_preempt_count(PREEMPT_ACTIVE);
4584 * Check again in case we missed a preemption opportunity
4585 * between schedule and now.
4587 barrier();
4588 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4590 EXPORT_SYMBOL(preempt_schedule);
4593 * this is the entry point to schedule() from kernel preemption
4594 * off of irq context.
4595 * Note, that this is called and return with irqs disabled. This will
4596 * protect us against recursive calling from irq.
4598 asmlinkage void __sched preempt_schedule_irq(void)
4600 struct thread_info *ti = current_thread_info();
4602 /* Catch callers which need to be fixed */
4603 BUG_ON(ti->preempt_count || !irqs_disabled());
4605 do {
4606 add_preempt_count(PREEMPT_ACTIVE);
4607 local_irq_enable();
4608 schedule();
4609 local_irq_disable();
4610 sub_preempt_count(PREEMPT_ACTIVE);
4613 * Check again in case we missed a preemption opportunity
4614 * between schedule and now.
4616 barrier();
4617 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4620 #endif /* CONFIG_PREEMPT */
4622 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4623 void *key)
4625 return try_to_wake_up(curr->private, mode, sync);
4627 EXPORT_SYMBOL(default_wake_function);
4630 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4631 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4632 * number) then we wake all the non-exclusive tasks and one exclusive task.
4634 * There are circumstances in which we can try to wake a task which has already
4635 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4636 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4638 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4639 int nr_exclusive, int sync, void *key)
4641 wait_queue_t *curr, *next;
4643 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4644 unsigned flags = curr->flags;
4646 if (curr->func(curr, mode, sync, key) &&
4647 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4648 break;
4653 * __wake_up - wake up threads blocked on a waitqueue.
4654 * @q: the waitqueue
4655 * @mode: which threads
4656 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4657 * @key: is directly passed to the wakeup function
4659 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4660 int nr_exclusive, void *key)
4662 unsigned long flags;
4664 spin_lock_irqsave(&q->lock, flags);
4665 __wake_up_common(q, mode, nr_exclusive, 0, key);
4666 spin_unlock_irqrestore(&q->lock, flags);
4668 EXPORT_SYMBOL(__wake_up);
4671 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4673 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4675 __wake_up_common(q, mode, 1, 0, NULL);
4679 * __wake_up_sync - wake up threads blocked on a waitqueue.
4680 * @q: the waitqueue
4681 * @mode: which threads
4682 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4684 * The sync wakeup differs that the waker knows that it will schedule
4685 * away soon, so while the target thread will be woken up, it will not
4686 * be migrated to another CPU - ie. the two threads are 'synchronized'
4687 * with each other. This can prevent needless bouncing between CPUs.
4689 * On UP it can prevent extra preemption.
4691 void
4692 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4694 unsigned long flags;
4695 int sync = 1;
4697 if (unlikely(!q))
4698 return;
4700 if (unlikely(!nr_exclusive))
4701 sync = 0;
4703 spin_lock_irqsave(&q->lock, flags);
4704 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4705 spin_unlock_irqrestore(&q->lock, flags);
4707 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4709 void complete(struct completion *x)
4711 unsigned long flags;
4713 spin_lock_irqsave(&x->wait.lock, flags);
4714 x->done++;
4715 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4716 spin_unlock_irqrestore(&x->wait.lock, flags);
4718 EXPORT_SYMBOL(complete);
4720 void complete_all(struct completion *x)
4722 unsigned long flags;
4724 spin_lock_irqsave(&x->wait.lock, flags);
4725 x->done += UINT_MAX/2;
4726 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4727 spin_unlock_irqrestore(&x->wait.lock, flags);
4729 EXPORT_SYMBOL(complete_all);
4731 static inline long __sched
4732 do_wait_for_common(struct completion *x, long timeout, int state)
4734 if (!x->done) {
4735 DECLARE_WAITQUEUE(wait, current);
4737 wait.flags |= WQ_FLAG_EXCLUSIVE;
4738 __add_wait_queue_tail(&x->wait, &wait);
4739 do {
4740 if ((state == TASK_INTERRUPTIBLE &&
4741 signal_pending(current)) ||
4742 (state == TASK_KILLABLE &&
4743 fatal_signal_pending(current))) {
4744 __remove_wait_queue(&x->wait, &wait);
4745 return -ERESTARTSYS;
4747 __set_current_state(state);
4748 spin_unlock_irq(&x->wait.lock);
4749 timeout = schedule_timeout(timeout);
4750 spin_lock_irq(&x->wait.lock);
4751 if (!timeout) {
4752 __remove_wait_queue(&x->wait, &wait);
4753 return timeout;
4755 } while (!x->done);
4756 __remove_wait_queue(&x->wait, &wait);
4758 x->done--;
4759 return timeout;
4762 static long __sched
4763 wait_for_common(struct completion *x, long timeout, int state)
4765 might_sleep();
4767 spin_lock_irq(&x->wait.lock);
4768 timeout = do_wait_for_common(x, timeout, state);
4769 spin_unlock_irq(&x->wait.lock);
4770 return timeout;
4773 void __sched wait_for_completion(struct completion *x)
4775 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4777 EXPORT_SYMBOL(wait_for_completion);
4779 unsigned long __sched
4780 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4782 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4784 EXPORT_SYMBOL(wait_for_completion_timeout);
4786 int __sched wait_for_completion_interruptible(struct completion *x)
4788 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4789 if (t == -ERESTARTSYS)
4790 return t;
4791 return 0;
4793 EXPORT_SYMBOL(wait_for_completion_interruptible);
4795 unsigned long __sched
4796 wait_for_completion_interruptible_timeout(struct completion *x,
4797 unsigned long timeout)
4799 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4801 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4803 int __sched wait_for_completion_killable(struct completion *x)
4805 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4806 if (t == -ERESTARTSYS)
4807 return t;
4808 return 0;
4810 EXPORT_SYMBOL(wait_for_completion_killable);
4812 static long __sched
4813 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4815 unsigned long flags;
4816 wait_queue_t wait;
4818 init_waitqueue_entry(&wait, current);
4820 __set_current_state(state);
4822 spin_lock_irqsave(&q->lock, flags);
4823 __add_wait_queue(q, &wait);
4824 spin_unlock(&q->lock);
4825 timeout = schedule_timeout(timeout);
4826 spin_lock_irq(&q->lock);
4827 __remove_wait_queue(q, &wait);
4828 spin_unlock_irqrestore(&q->lock, flags);
4830 return timeout;
4833 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4835 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4837 EXPORT_SYMBOL(interruptible_sleep_on);
4839 long __sched
4840 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4842 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4844 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4846 void __sched sleep_on(wait_queue_head_t *q)
4848 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4850 EXPORT_SYMBOL(sleep_on);
4852 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4854 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4856 EXPORT_SYMBOL(sleep_on_timeout);
4858 #ifdef CONFIG_RT_MUTEXES
4861 * rt_mutex_setprio - set the current priority of a task
4862 * @p: task
4863 * @prio: prio value (kernel-internal form)
4865 * This function changes the 'effective' priority of a task. It does
4866 * not touch ->normal_prio like __setscheduler().
4868 * Used by the rt_mutex code to implement priority inheritance logic.
4870 void rt_mutex_setprio(struct task_struct *p, int prio)
4872 unsigned long flags;
4873 int oldprio, on_rq, running;
4874 struct rq *rq;
4875 const struct sched_class *prev_class = p->sched_class;
4877 BUG_ON(prio < 0 || prio > MAX_PRIO);
4879 rq = task_rq_lock(p, &flags);
4880 update_rq_clock(rq);
4882 oldprio = p->prio;
4883 on_rq = p->se.on_rq;
4884 running = task_current(rq, p);
4885 if (on_rq)
4886 dequeue_task(rq, p, 0);
4887 if (running)
4888 p->sched_class->put_prev_task(rq, p);
4890 if (rt_prio(prio))
4891 p->sched_class = &rt_sched_class;
4892 else
4893 p->sched_class = &fair_sched_class;
4895 p->prio = prio;
4897 if (running)
4898 p->sched_class->set_curr_task(rq);
4899 if (on_rq) {
4900 enqueue_task(rq, p, 0);
4902 check_class_changed(rq, p, prev_class, oldprio, running);
4904 task_rq_unlock(rq, &flags);
4907 #endif
4909 void set_user_nice(struct task_struct *p, long nice)
4911 int old_prio, delta, on_rq;
4912 unsigned long flags;
4913 struct rq *rq;
4915 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4916 return;
4918 * We have to be careful, if called from sys_setpriority(),
4919 * the task might be in the middle of scheduling on another CPU.
4921 rq = task_rq_lock(p, &flags);
4922 update_rq_clock(rq);
4924 * The RT priorities are set via sched_setscheduler(), but we still
4925 * allow the 'normal' nice value to be set - but as expected
4926 * it wont have any effect on scheduling until the task is
4927 * SCHED_FIFO/SCHED_RR:
4929 if (task_has_rt_policy(p)) {
4930 p->static_prio = NICE_TO_PRIO(nice);
4931 goto out_unlock;
4933 on_rq = p->se.on_rq;
4934 if (on_rq)
4935 dequeue_task(rq, p, 0);
4937 p->static_prio = NICE_TO_PRIO(nice);
4938 set_load_weight(p);
4939 old_prio = p->prio;
4940 p->prio = effective_prio(p);
4941 delta = p->prio - old_prio;
4943 if (on_rq) {
4944 enqueue_task(rq, p, 0);
4946 * If the task increased its priority or is running and
4947 * lowered its priority, then reschedule its CPU:
4949 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4950 resched_task(rq->curr);
4952 out_unlock:
4953 task_rq_unlock(rq, &flags);
4955 EXPORT_SYMBOL(set_user_nice);
4958 * can_nice - check if a task can reduce its nice value
4959 * @p: task
4960 * @nice: nice value
4962 int can_nice(const struct task_struct *p, const int nice)
4964 /* convert nice value [19,-20] to rlimit style value [1,40] */
4965 int nice_rlim = 20 - nice;
4967 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4968 capable(CAP_SYS_NICE));
4971 #ifdef __ARCH_WANT_SYS_NICE
4974 * sys_nice - change the priority of the current process.
4975 * @increment: priority increment
4977 * sys_setpriority is a more generic, but much slower function that
4978 * does similar things.
4980 asmlinkage long sys_nice(int increment)
4982 long nice, retval;
4985 * Setpriority might change our priority at the same moment.
4986 * We don't have to worry. Conceptually one call occurs first
4987 * and we have a single winner.
4989 if (increment < -40)
4990 increment = -40;
4991 if (increment > 40)
4992 increment = 40;
4994 nice = PRIO_TO_NICE(current->static_prio) + increment;
4995 if (nice < -20)
4996 nice = -20;
4997 if (nice > 19)
4998 nice = 19;
5000 if (increment < 0 && !can_nice(current, nice))
5001 return -EPERM;
5003 retval = security_task_setnice(current, nice);
5004 if (retval)
5005 return retval;
5007 set_user_nice(current, nice);
5008 return 0;
5011 #endif
5014 * task_prio - return the priority value of a given task.
5015 * @p: the task in question.
5017 * This is the priority value as seen by users in /proc.
5018 * RT tasks are offset by -200. Normal tasks are centered
5019 * around 0, value goes from -16 to +15.
5021 int task_prio(const struct task_struct *p)
5023 return p->prio - MAX_RT_PRIO;
5027 * task_nice - return the nice value of a given task.
5028 * @p: the task in question.
5030 int task_nice(const struct task_struct *p)
5032 return TASK_NICE(p);
5034 EXPORT_SYMBOL(task_nice);
5037 * idle_cpu - is a given cpu idle currently?
5038 * @cpu: the processor in question.
5040 int idle_cpu(int cpu)
5042 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5046 * idle_task - return the idle task for a given cpu.
5047 * @cpu: the processor in question.
5049 struct task_struct *idle_task(int cpu)
5051 return cpu_rq(cpu)->idle;
5055 * find_process_by_pid - find a process with a matching PID value.
5056 * @pid: the pid in question.
5058 static struct task_struct *find_process_by_pid(pid_t pid)
5060 return pid ? find_task_by_vpid(pid) : current;
5063 /* Actually do priority change: must hold rq lock. */
5064 static void
5065 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5067 BUG_ON(p->se.on_rq);
5069 p->policy = policy;
5070 switch (p->policy) {
5071 case SCHED_NORMAL:
5072 case SCHED_BATCH:
5073 case SCHED_IDLE:
5074 p->sched_class = &fair_sched_class;
5075 break;
5076 case SCHED_FIFO:
5077 case SCHED_RR:
5078 p->sched_class = &rt_sched_class;
5079 break;
5082 p->rt_priority = prio;
5083 p->normal_prio = normal_prio(p);
5084 /* we are holding p->pi_lock already */
5085 p->prio = rt_mutex_getprio(p);
5086 set_load_weight(p);
5090 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5091 * @p: the task in question.
5092 * @policy: new policy.
5093 * @param: structure containing the new RT priority.
5095 * NOTE that the task may be already dead.
5097 int sched_setscheduler(struct task_struct *p, int policy,
5098 struct sched_param *param)
5100 int retval, oldprio, oldpolicy = -1, on_rq, running;
5101 unsigned long flags;
5102 const struct sched_class *prev_class = p->sched_class;
5103 struct rq *rq;
5105 /* may grab non-irq protected spin_locks */
5106 BUG_ON(in_interrupt());
5107 recheck:
5108 /* double check policy once rq lock held */
5109 if (policy < 0)
5110 policy = oldpolicy = p->policy;
5111 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5112 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5113 policy != SCHED_IDLE)
5114 return -EINVAL;
5116 * Valid priorities for SCHED_FIFO and SCHED_RR are
5117 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5118 * SCHED_BATCH and SCHED_IDLE is 0.
5120 if (param->sched_priority < 0 ||
5121 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5122 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5123 return -EINVAL;
5124 if (rt_policy(policy) != (param->sched_priority != 0))
5125 return -EINVAL;
5128 * Allow unprivileged RT tasks to decrease priority:
5130 if (!capable(CAP_SYS_NICE)) {
5131 if (rt_policy(policy)) {
5132 unsigned long rlim_rtprio;
5134 if (!lock_task_sighand(p, &flags))
5135 return -ESRCH;
5136 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5137 unlock_task_sighand(p, &flags);
5139 /* can't set/change the rt policy */
5140 if (policy != p->policy && !rlim_rtprio)
5141 return -EPERM;
5143 /* can't increase priority */
5144 if (param->sched_priority > p->rt_priority &&
5145 param->sched_priority > rlim_rtprio)
5146 return -EPERM;
5149 * Like positive nice levels, dont allow tasks to
5150 * move out of SCHED_IDLE either:
5152 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5153 return -EPERM;
5155 /* can't change other user's priorities */
5156 if ((current->euid != p->euid) &&
5157 (current->euid != p->uid))
5158 return -EPERM;
5161 #ifdef CONFIG_RT_GROUP_SCHED
5163 * Do not allow realtime tasks into groups that have no runtime
5164 * assigned.
5166 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
5167 return -EPERM;
5168 #endif
5170 retval = security_task_setscheduler(p, policy, param);
5171 if (retval)
5172 return retval;
5174 * make sure no PI-waiters arrive (or leave) while we are
5175 * changing the priority of the task:
5177 spin_lock_irqsave(&p->pi_lock, flags);
5179 * To be able to change p->policy safely, the apropriate
5180 * runqueue lock must be held.
5182 rq = __task_rq_lock(p);
5183 /* recheck policy now with rq lock held */
5184 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5185 policy = oldpolicy = -1;
5186 __task_rq_unlock(rq);
5187 spin_unlock_irqrestore(&p->pi_lock, flags);
5188 goto recheck;
5190 update_rq_clock(rq);
5191 on_rq = p->se.on_rq;
5192 running = task_current(rq, p);
5193 if (on_rq)
5194 deactivate_task(rq, p, 0);
5195 if (running)
5196 p->sched_class->put_prev_task(rq, p);
5198 oldprio = p->prio;
5199 __setscheduler(rq, p, policy, param->sched_priority);
5201 if (running)
5202 p->sched_class->set_curr_task(rq);
5203 if (on_rq) {
5204 activate_task(rq, p, 0);
5206 check_class_changed(rq, p, prev_class, oldprio, running);
5208 __task_rq_unlock(rq);
5209 spin_unlock_irqrestore(&p->pi_lock, flags);
5211 rt_mutex_adjust_pi(p);
5213 return 0;
5215 EXPORT_SYMBOL_GPL(sched_setscheduler);
5217 static int
5218 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5220 struct sched_param lparam;
5221 struct task_struct *p;
5222 int retval;
5224 if (!param || pid < 0)
5225 return -EINVAL;
5226 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5227 return -EFAULT;
5229 rcu_read_lock();
5230 retval = -ESRCH;
5231 p = find_process_by_pid(pid);
5232 if (p != NULL)
5233 retval = sched_setscheduler(p, policy, &lparam);
5234 rcu_read_unlock();
5236 return retval;
5240 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5241 * @pid: the pid in question.
5242 * @policy: new policy.
5243 * @param: structure containing the new RT priority.
5245 asmlinkage long
5246 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5248 /* negative values for policy are not valid */
5249 if (policy < 0)
5250 return -EINVAL;
5252 return do_sched_setscheduler(pid, policy, param);
5256 * sys_sched_setparam - set/change the RT priority of a thread
5257 * @pid: the pid in question.
5258 * @param: structure containing the new RT priority.
5260 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5262 return do_sched_setscheduler(pid, -1, param);
5266 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5267 * @pid: the pid in question.
5269 asmlinkage long sys_sched_getscheduler(pid_t pid)
5271 struct task_struct *p;
5272 int retval;
5274 if (pid < 0)
5275 return -EINVAL;
5277 retval = -ESRCH;
5278 read_lock(&tasklist_lock);
5279 p = find_process_by_pid(pid);
5280 if (p) {
5281 retval = security_task_getscheduler(p);
5282 if (!retval)
5283 retval = p->policy;
5285 read_unlock(&tasklist_lock);
5286 return retval;
5290 * sys_sched_getscheduler - get the RT priority of a thread
5291 * @pid: the pid in question.
5292 * @param: structure containing the RT priority.
5294 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5296 struct sched_param lp;
5297 struct task_struct *p;
5298 int retval;
5300 if (!param || pid < 0)
5301 return -EINVAL;
5303 read_lock(&tasklist_lock);
5304 p = find_process_by_pid(pid);
5305 retval = -ESRCH;
5306 if (!p)
5307 goto out_unlock;
5309 retval = security_task_getscheduler(p);
5310 if (retval)
5311 goto out_unlock;
5313 lp.sched_priority = p->rt_priority;
5314 read_unlock(&tasklist_lock);
5317 * This one might sleep, we cannot do it with a spinlock held ...
5319 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5321 return retval;
5323 out_unlock:
5324 read_unlock(&tasklist_lock);
5325 return retval;
5328 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5330 cpumask_t cpus_allowed;
5331 cpumask_t new_mask = *in_mask;
5332 struct task_struct *p;
5333 int retval;
5335 get_online_cpus();
5336 read_lock(&tasklist_lock);
5338 p = find_process_by_pid(pid);
5339 if (!p) {
5340 read_unlock(&tasklist_lock);
5341 put_online_cpus();
5342 return -ESRCH;
5346 * It is not safe to call set_cpus_allowed with the
5347 * tasklist_lock held. We will bump the task_struct's
5348 * usage count and then drop tasklist_lock.
5350 get_task_struct(p);
5351 read_unlock(&tasklist_lock);
5353 retval = -EPERM;
5354 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5355 !capable(CAP_SYS_NICE))
5356 goto out_unlock;
5358 retval = security_task_setscheduler(p, 0, NULL);
5359 if (retval)
5360 goto out_unlock;
5362 cpuset_cpus_allowed(p, &cpus_allowed);
5363 cpus_and(new_mask, new_mask, cpus_allowed);
5364 again:
5365 retval = set_cpus_allowed_ptr(p, &new_mask);
5367 if (!retval) {
5368 cpuset_cpus_allowed(p, &cpus_allowed);
5369 if (!cpus_subset(new_mask, cpus_allowed)) {
5371 * We must have raced with a concurrent cpuset
5372 * update. Just reset the cpus_allowed to the
5373 * cpuset's cpus_allowed
5375 new_mask = cpus_allowed;
5376 goto again;
5379 out_unlock:
5380 put_task_struct(p);
5381 put_online_cpus();
5382 return retval;
5385 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5386 cpumask_t *new_mask)
5388 if (len < sizeof(cpumask_t)) {
5389 memset(new_mask, 0, sizeof(cpumask_t));
5390 } else if (len > sizeof(cpumask_t)) {
5391 len = sizeof(cpumask_t);
5393 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5397 * sys_sched_setaffinity - set the cpu affinity of a process
5398 * @pid: pid of the process
5399 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5400 * @user_mask_ptr: user-space pointer to the new cpu mask
5402 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5403 unsigned long __user *user_mask_ptr)
5405 cpumask_t new_mask;
5406 int retval;
5408 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5409 if (retval)
5410 return retval;
5412 return sched_setaffinity(pid, &new_mask);
5416 * Represents all cpu's present in the system
5417 * In systems capable of hotplug, this map could dynamically grow
5418 * as new cpu's are detected in the system via any platform specific
5419 * method, such as ACPI for e.g.
5422 cpumask_t cpu_present_map __read_mostly;
5423 EXPORT_SYMBOL(cpu_present_map);
5425 #ifndef CONFIG_SMP
5426 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
5427 EXPORT_SYMBOL(cpu_online_map);
5429 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
5430 EXPORT_SYMBOL(cpu_possible_map);
5431 #endif
5433 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5435 struct task_struct *p;
5436 int retval;
5438 get_online_cpus();
5439 read_lock(&tasklist_lock);
5441 retval = -ESRCH;
5442 p = find_process_by_pid(pid);
5443 if (!p)
5444 goto out_unlock;
5446 retval = security_task_getscheduler(p);
5447 if (retval)
5448 goto out_unlock;
5450 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5452 out_unlock:
5453 read_unlock(&tasklist_lock);
5454 put_online_cpus();
5456 return retval;
5460 * sys_sched_getaffinity - get the cpu affinity of a process
5461 * @pid: pid of the process
5462 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5463 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5465 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5466 unsigned long __user *user_mask_ptr)
5468 int ret;
5469 cpumask_t mask;
5471 if (len < sizeof(cpumask_t))
5472 return -EINVAL;
5474 ret = sched_getaffinity(pid, &mask);
5475 if (ret < 0)
5476 return ret;
5478 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5479 return -EFAULT;
5481 return sizeof(cpumask_t);
5485 * sys_sched_yield - yield the current processor to other threads.
5487 * This function yields the current CPU to other tasks. If there are no
5488 * other threads running on this CPU then this function will return.
5490 asmlinkage long sys_sched_yield(void)
5492 struct rq *rq = this_rq_lock();
5494 schedstat_inc(rq, yld_count);
5495 current->sched_class->yield_task(rq);
5498 * Since we are going to call schedule() anyway, there's
5499 * no need to preempt or enable interrupts:
5501 __release(rq->lock);
5502 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5503 _raw_spin_unlock(&rq->lock);
5504 preempt_enable_no_resched();
5506 schedule();
5508 return 0;
5511 static void __cond_resched(void)
5513 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5514 __might_sleep(__FILE__, __LINE__);
5515 #endif
5517 * The BKS might be reacquired before we have dropped
5518 * PREEMPT_ACTIVE, which could trigger a second
5519 * cond_resched() call.
5521 do {
5522 add_preempt_count(PREEMPT_ACTIVE);
5523 schedule();
5524 sub_preempt_count(PREEMPT_ACTIVE);
5525 } while (need_resched());
5528 int __sched _cond_resched(void)
5530 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5531 system_state == SYSTEM_RUNNING) {
5532 __cond_resched();
5533 return 1;
5535 return 0;
5537 EXPORT_SYMBOL(_cond_resched);
5540 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5541 * call schedule, and on return reacquire the lock.
5543 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5544 * operations here to prevent schedule() from being called twice (once via
5545 * spin_unlock(), once by hand).
5547 int cond_resched_lock(spinlock_t *lock)
5549 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5550 int ret = 0;
5552 if (spin_needbreak(lock) || resched) {
5553 spin_unlock(lock);
5554 if (resched && need_resched())
5555 __cond_resched();
5556 else
5557 cpu_relax();
5558 ret = 1;
5559 spin_lock(lock);
5561 return ret;
5563 EXPORT_SYMBOL(cond_resched_lock);
5565 int __sched cond_resched_softirq(void)
5567 BUG_ON(!in_softirq());
5569 if (need_resched() && system_state == SYSTEM_RUNNING) {
5570 local_bh_enable();
5571 __cond_resched();
5572 local_bh_disable();
5573 return 1;
5575 return 0;
5577 EXPORT_SYMBOL(cond_resched_softirq);
5580 * yield - yield the current processor to other threads.
5582 * This is a shortcut for kernel-space yielding - it marks the
5583 * thread runnable and calls sys_sched_yield().
5585 void __sched yield(void)
5587 set_current_state(TASK_RUNNING);
5588 sys_sched_yield();
5590 EXPORT_SYMBOL(yield);
5593 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5594 * that process accounting knows that this is a task in IO wait state.
5596 * But don't do that if it is a deliberate, throttling IO wait (this task
5597 * has set its backing_dev_info: the queue against which it should throttle)
5599 void __sched io_schedule(void)
5601 struct rq *rq = &__raw_get_cpu_var(runqueues);
5603 delayacct_blkio_start();
5604 atomic_inc(&rq->nr_iowait);
5605 schedule();
5606 atomic_dec(&rq->nr_iowait);
5607 delayacct_blkio_end();
5609 EXPORT_SYMBOL(io_schedule);
5611 long __sched io_schedule_timeout(long timeout)
5613 struct rq *rq = &__raw_get_cpu_var(runqueues);
5614 long ret;
5616 delayacct_blkio_start();
5617 atomic_inc(&rq->nr_iowait);
5618 ret = schedule_timeout(timeout);
5619 atomic_dec(&rq->nr_iowait);
5620 delayacct_blkio_end();
5621 return ret;
5625 * sys_sched_get_priority_max - return maximum RT priority.
5626 * @policy: scheduling class.
5628 * this syscall returns the maximum rt_priority that can be used
5629 * by a given scheduling class.
5631 asmlinkage long sys_sched_get_priority_max(int policy)
5633 int ret = -EINVAL;
5635 switch (policy) {
5636 case SCHED_FIFO:
5637 case SCHED_RR:
5638 ret = MAX_USER_RT_PRIO-1;
5639 break;
5640 case SCHED_NORMAL:
5641 case SCHED_BATCH:
5642 case SCHED_IDLE:
5643 ret = 0;
5644 break;
5646 return ret;
5650 * sys_sched_get_priority_min - return minimum RT priority.
5651 * @policy: scheduling class.
5653 * this syscall returns the minimum rt_priority that can be used
5654 * by a given scheduling class.
5656 asmlinkage long sys_sched_get_priority_min(int policy)
5658 int ret = -EINVAL;
5660 switch (policy) {
5661 case SCHED_FIFO:
5662 case SCHED_RR:
5663 ret = 1;
5664 break;
5665 case SCHED_NORMAL:
5666 case SCHED_BATCH:
5667 case SCHED_IDLE:
5668 ret = 0;
5670 return ret;
5674 * sys_sched_rr_get_interval - return the default timeslice of a process.
5675 * @pid: pid of the process.
5676 * @interval: userspace pointer to the timeslice value.
5678 * this syscall writes the default timeslice value of a given process
5679 * into the user-space timespec buffer. A value of '0' means infinity.
5681 asmlinkage
5682 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5684 struct task_struct *p;
5685 unsigned int time_slice;
5686 int retval;
5687 struct timespec t;
5689 if (pid < 0)
5690 return -EINVAL;
5692 retval = -ESRCH;
5693 read_lock(&tasklist_lock);
5694 p = find_process_by_pid(pid);
5695 if (!p)
5696 goto out_unlock;
5698 retval = security_task_getscheduler(p);
5699 if (retval)
5700 goto out_unlock;
5703 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5704 * tasks that are on an otherwise idle runqueue:
5706 time_slice = 0;
5707 if (p->policy == SCHED_RR) {
5708 time_slice = DEF_TIMESLICE;
5709 } else if (p->policy != SCHED_FIFO) {
5710 struct sched_entity *se = &p->se;
5711 unsigned long flags;
5712 struct rq *rq;
5714 rq = task_rq_lock(p, &flags);
5715 if (rq->cfs.load.weight)
5716 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5717 task_rq_unlock(rq, &flags);
5719 read_unlock(&tasklist_lock);
5720 jiffies_to_timespec(time_slice, &t);
5721 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5722 return retval;
5724 out_unlock:
5725 read_unlock(&tasklist_lock);
5726 return retval;
5729 static const char stat_nam[] = "RSDTtZX";
5731 void sched_show_task(struct task_struct *p)
5733 unsigned long free = 0;
5734 unsigned state;
5736 state = p->state ? __ffs(p->state) + 1 : 0;
5737 printk(KERN_INFO "%-13.13s %c", p->comm,
5738 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5739 #if BITS_PER_LONG == 32
5740 if (state == TASK_RUNNING)
5741 printk(KERN_CONT " running ");
5742 else
5743 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5744 #else
5745 if (state == TASK_RUNNING)
5746 printk(KERN_CONT " running task ");
5747 else
5748 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5749 #endif
5750 #ifdef CONFIG_DEBUG_STACK_USAGE
5752 unsigned long *n = end_of_stack(p);
5753 while (!*n)
5754 n++;
5755 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5757 #endif
5758 printk(KERN_CONT "%5lu %5d %6d\n", free,
5759 task_pid_nr(p), task_pid_nr(p->real_parent));
5761 show_stack(p, NULL);
5764 void show_state_filter(unsigned long state_filter)
5766 struct task_struct *g, *p;
5768 #if BITS_PER_LONG == 32
5769 printk(KERN_INFO
5770 " task PC stack pid father\n");
5771 #else
5772 printk(KERN_INFO
5773 " task PC stack pid father\n");
5774 #endif
5775 read_lock(&tasklist_lock);
5776 do_each_thread(g, p) {
5778 * reset the NMI-timeout, listing all files on a slow
5779 * console might take alot of time:
5781 touch_nmi_watchdog();
5782 if (!state_filter || (p->state & state_filter))
5783 sched_show_task(p);
5784 } while_each_thread(g, p);
5786 touch_all_softlockup_watchdogs();
5788 #ifdef CONFIG_SCHED_DEBUG
5789 sysrq_sched_debug_show();
5790 #endif
5791 read_unlock(&tasklist_lock);
5793 * Only show locks if all tasks are dumped:
5795 if (state_filter == -1)
5796 debug_show_all_locks();
5799 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5801 idle->sched_class = &idle_sched_class;
5805 * init_idle - set up an idle thread for a given CPU
5806 * @idle: task in question
5807 * @cpu: cpu the idle task belongs to
5809 * NOTE: this function does not set the idle thread's NEED_RESCHED
5810 * flag, to make booting more robust.
5812 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5814 struct rq *rq = cpu_rq(cpu);
5815 unsigned long flags;
5817 __sched_fork(idle);
5818 idle->se.exec_start = sched_clock();
5820 idle->prio = idle->normal_prio = MAX_PRIO;
5821 idle->cpus_allowed = cpumask_of_cpu(cpu);
5822 __set_task_cpu(idle, cpu);
5824 spin_lock_irqsave(&rq->lock, flags);
5825 rq->curr = rq->idle = idle;
5826 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5827 idle->oncpu = 1;
5828 #endif
5829 spin_unlock_irqrestore(&rq->lock, flags);
5831 /* Set the preempt count _outside_ the spinlocks! */
5832 #if defined(CONFIG_PREEMPT)
5833 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5834 #else
5835 task_thread_info(idle)->preempt_count = 0;
5836 #endif
5838 * The idle tasks have their own, simple scheduling class:
5840 idle->sched_class = &idle_sched_class;
5844 * In a system that switches off the HZ timer nohz_cpu_mask
5845 * indicates which cpus entered this state. This is used
5846 * in the rcu update to wait only for active cpus. For system
5847 * which do not switch off the HZ timer nohz_cpu_mask should
5848 * always be CPU_MASK_NONE.
5850 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5853 * Increase the granularity value when there are more CPUs,
5854 * because with more CPUs the 'effective latency' as visible
5855 * to users decreases. But the relationship is not linear,
5856 * so pick a second-best guess by going with the log2 of the
5857 * number of CPUs.
5859 * This idea comes from the SD scheduler of Con Kolivas:
5861 static inline void sched_init_granularity(void)
5863 unsigned int factor = 1 + ilog2(num_online_cpus());
5864 const unsigned long limit = 200000000;
5866 sysctl_sched_min_granularity *= factor;
5867 if (sysctl_sched_min_granularity > limit)
5868 sysctl_sched_min_granularity = limit;
5870 sysctl_sched_latency *= factor;
5871 if (sysctl_sched_latency > limit)
5872 sysctl_sched_latency = limit;
5874 sysctl_sched_wakeup_granularity *= factor;
5877 #ifdef CONFIG_SMP
5879 * This is how migration works:
5881 * 1) we queue a struct migration_req structure in the source CPU's
5882 * runqueue and wake up that CPU's migration thread.
5883 * 2) we down() the locked semaphore => thread blocks.
5884 * 3) migration thread wakes up (implicitly it forces the migrated
5885 * thread off the CPU)
5886 * 4) it gets the migration request and checks whether the migrated
5887 * task is still in the wrong runqueue.
5888 * 5) if it's in the wrong runqueue then the migration thread removes
5889 * it and puts it into the right queue.
5890 * 6) migration thread up()s the semaphore.
5891 * 7) we wake up and the migration is done.
5895 * Change a given task's CPU affinity. Migrate the thread to a
5896 * proper CPU and schedule it away if the CPU it's executing on
5897 * is removed from the allowed bitmask.
5899 * NOTE: the caller must have a valid reference to the task, the
5900 * task must not exit() & deallocate itself prematurely. The
5901 * call is not atomic; no spinlocks may be held.
5903 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5905 struct migration_req req;
5906 unsigned long flags;
5907 struct rq *rq;
5908 int ret = 0;
5910 rq = task_rq_lock(p, &flags);
5911 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5912 ret = -EINVAL;
5913 goto out;
5916 if (p->sched_class->set_cpus_allowed)
5917 p->sched_class->set_cpus_allowed(p, new_mask);
5918 else {
5919 p->cpus_allowed = *new_mask;
5920 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5923 /* Can the task run on the task's current CPU? If so, we're done */
5924 if (cpu_isset(task_cpu(p), *new_mask))
5925 goto out;
5927 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5928 /* Need help from migration thread: drop lock and wait. */
5929 task_rq_unlock(rq, &flags);
5930 wake_up_process(rq->migration_thread);
5931 wait_for_completion(&req.done);
5932 tlb_migrate_finish(p->mm);
5933 return 0;
5935 out:
5936 task_rq_unlock(rq, &flags);
5938 return ret;
5940 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5943 * Move (not current) task off this cpu, onto dest cpu. We're doing
5944 * this because either it can't run here any more (set_cpus_allowed()
5945 * away from this CPU, or CPU going down), or because we're
5946 * attempting to rebalance this task on exec (sched_exec).
5948 * So we race with normal scheduler movements, but that's OK, as long
5949 * as the task is no longer on this CPU.
5951 * Returns non-zero if task was successfully migrated.
5953 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5955 struct rq *rq_dest, *rq_src;
5956 int ret = 0, on_rq;
5958 if (unlikely(cpu_is_offline(dest_cpu)))
5959 return ret;
5961 rq_src = cpu_rq(src_cpu);
5962 rq_dest = cpu_rq(dest_cpu);
5964 double_rq_lock(rq_src, rq_dest);
5965 /* Already moved. */
5966 if (task_cpu(p) != src_cpu)
5967 goto out;
5968 /* Affinity changed (again). */
5969 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5970 goto out;
5972 on_rq = p->se.on_rq;
5973 if (on_rq)
5974 deactivate_task(rq_src, p, 0);
5976 set_task_cpu(p, dest_cpu);
5977 if (on_rq) {
5978 activate_task(rq_dest, p, 0);
5979 check_preempt_curr(rq_dest, p);
5981 ret = 1;
5982 out:
5983 double_rq_unlock(rq_src, rq_dest);
5984 return ret;
5988 * migration_thread - this is a highprio system thread that performs
5989 * thread migration by bumping thread off CPU then 'pushing' onto
5990 * another runqueue.
5992 static int migration_thread(void *data)
5994 int cpu = (long)data;
5995 struct rq *rq;
5997 rq = cpu_rq(cpu);
5998 BUG_ON(rq->migration_thread != current);
6000 set_current_state(TASK_INTERRUPTIBLE);
6001 while (!kthread_should_stop()) {
6002 struct migration_req *req;
6003 struct list_head *head;
6005 spin_lock_irq(&rq->lock);
6007 if (cpu_is_offline(cpu)) {
6008 spin_unlock_irq(&rq->lock);
6009 goto wait_to_die;
6012 if (rq->active_balance) {
6013 active_load_balance(rq, cpu);
6014 rq->active_balance = 0;
6017 head = &rq->migration_queue;
6019 if (list_empty(head)) {
6020 spin_unlock_irq(&rq->lock);
6021 schedule();
6022 set_current_state(TASK_INTERRUPTIBLE);
6023 continue;
6025 req = list_entry(head->next, struct migration_req, list);
6026 list_del_init(head->next);
6028 spin_unlock(&rq->lock);
6029 __migrate_task(req->task, cpu, req->dest_cpu);
6030 local_irq_enable();
6032 complete(&req->done);
6034 __set_current_state(TASK_RUNNING);
6035 return 0;
6037 wait_to_die:
6038 /* Wait for kthread_stop */
6039 set_current_state(TASK_INTERRUPTIBLE);
6040 while (!kthread_should_stop()) {
6041 schedule();
6042 set_current_state(TASK_INTERRUPTIBLE);
6044 __set_current_state(TASK_RUNNING);
6045 return 0;
6048 #ifdef CONFIG_HOTPLUG_CPU
6050 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6052 int ret;
6054 local_irq_disable();
6055 ret = __migrate_task(p, src_cpu, dest_cpu);
6056 local_irq_enable();
6057 return ret;
6061 * Figure out where task on dead CPU should go, use force if necessary.
6062 * NOTE: interrupts should be disabled by the caller
6064 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6066 unsigned long flags;
6067 cpumask_t mask;
6068 struct rq *rq;
6069 int dest_cpu;
6071 do {
6072 /* On same node? */
6073 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6074 cpus_and(mask, mask, p->cpus_allowed);
6075 dest_cpu = any_online_cpu(mask);
6077 /* On any allowed CPU? */
6078 if (dest_cpu >= nr_cpu_ids)
6079 dest_cpu = any_online_cpu(p->cpus_allowed);
6081 /* No more Mr. Nice Guy. */
6082 if (dest_cpu >= nr_cpu_ids) {
6083 cpumask_t cpus_allowed;
6085 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6087 * Try to stay on the same cpuset, where the
6088 * current cpuset may be a subset of all cpus.
6089 * The cpuset_cpus_allowed_locked() variant of
6090 * cpuset_cpus_allowed() will not block. It must be
6091 * called within calls to cpuset_lock/cpuset_unlock.
6093 rq = task_rq_lock(p, &flags);
6094 p->cpus_allowed = cpus_allowed;
6095 dest_cpu = any_online_cpu(p->cpus_allowed);
6096 task_rq_unlock(rq, &flags);
6099 * Don't tell them about moving exiting tasks or
6100 * kernel threads (both mm NULL), since they never
6101 * leave kernel.
6103 if (p->mm && printk_ratelimit()) {
6104 printk(KERN_INFO "process %d (%s) no "
6105 "longer affine to cpu%d\n",
6106 task_pid_nr(p), p->comm, dead_cpu);
6109 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6113 * While a dead CPU has no uninterruptible tasks queued at this point,
6114 * it might still have a nonzero ->nr_uninterruptible counter, because
6115 * for performance reasons the counter is not stricly tracking tasks to
6116 * their home CPUs. So we just add the counter to another CPU's counter,
6117 * to keep the global sum constant after CPU-down:
6119 static void migrate_nr_uninterruptible(struct rq *rq_src)
6121 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6122 unsigned long flags;
6124 local_irq_save(flags);
6125 double_rq_lock(rq_src, rq_dest);
6126 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6127 rq_src->nr_uninterruptible = 0;
6128 double_rq_unlock(rq_src, rq_dest);
6129 local_irq_restore(flags);
6132 /* Run through task list and migrate tasks from the dead cpu. */
6133 static void migrate_live_tasks(int src_cpu)
6135 struct task_struct *p, *t;
6137 read_lock(&tasklist_lock);
6139 do_each_thread(t, p) {
6140 if (p == current)
6141 continue;
6143 if (task_cpu(p) == src_cpu)
6144 move_task_off_dead_cpu(src_cpu, p);
6145 } while_each_thread(t, p);
6147 read_unlock(&tasklist_lock);
6151 * Schedules idle task to be the next runnable task on current CPU.
6152 * It does so by boosting its priority to highest possible.
6153 * Used by CPU offline code.
6155 void sched_idle_next(void)
6157 int this_cpu = smp_processor_id();
6158 struct rq *rq = cpu_rq(this_cpu);
6159 struct task_struct *p = rq->idle;
6160 unsigned long flags;
6162 /* cpu has to be offline */
6163 BUG_ON(cpu_online(this_cpu));
6166 * Strictly not necessary since rest of the CPUs are stopped by now
6167 * and interrupts disabled on the current cpu.
6169 spin_lock_irqsave(&rq->lock, flags);
6171 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6173 update_rq_clock(rq);
6174 activate_task(rq, p, 0);
6176 spin_unlock_irqrestore(&rq->lock, flags);
6180 * Ensures that the idle task is using init_mm right before its cpu goes
6181 * offline.
6183 void idle_task_exit(void)
6185 struct mm_struct *mm = current->active_mm;
6187 BUG_ON(cpu_online(smp_processor_id()));
6189 if (mm != &init_mm)
6190 switch_mm(mm, &init_mm, current);
6191 mmdrop(mm);
6194 /* called under rq->lock with disabled interrupts */
6195 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6197 struct rq *rq = cpu_rq(dead_cpu);
6199 /* Must be exiting, otherwise would be on tasklist. */
6200 BUG_ON(!p->exit_state);
6202 /* Cannot have done final schedule yet: would have vanished. */
6203 BUG_ON(p->state == TASK_DEAD);
6205 get_task_struct(p);
6208 * Drop lock around migration; if someone else moves it,
6209 * that's OK. No task can be added to this CPU, so iteration is
6210 * fine.
6212 spin_unlock_irq(&rq->lock);
6213 move_task_off_dead_cpu(dead_cpu, p);
6214 spin_lock_irq(&rq->lock);
6216 put_task_struct(p);
6219 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6220 static void migrate_dead_tasks(unsigned int dead_cpu)
6222 struct rq *rq = cpu_rq(dead_cpu);
6223 struct task_struct *next;
6225 for ( ; ; ) {
6226 if (!rq->nr_running)
6227 break;
6228 update_rq_clock(rq);
6229 next = pick_next_task(rq, rq->curr);
6230 if (!next)
6231 break;
6232 migrate_dead(dead_cpu, next);
6236 #endif /* CONFIG_HOTPLUG_CPU */
6238 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6240 static struct ctl_table sd_ctl_dir[] = {
6242 .procname = "sched_domain",
6243 .mode = 0555,
6245 {0, },
6248 static struct ctl_table sd_ctl_root[] = {
6250 .ctl_name = CTL_KERN,
6251 .procname = "kernel",
6252 .mode = 0555,
6253 .child = sd_ctl_dir,
6255 {0, },
6258 static struct ctl_table *sd_alloc_ctl_entry(int n)
6260 struct ctl_table *entry =
6261 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6263 return entry;
6266 static void sd_free_ctl_entry(struct ctl_table **tablep)
6268 struct ctl_table *entry;
6271 * In the intermediate directories, both the child directory and
6272 * procname are dynamically allocated and could fail but the mode
6273 * will always be set. In the lowest directory the names are
6274 * static strings and all have proc handlers.
6276 for (entry = *tablep; entry->mode; entry++) {
6277 if (entry->child)
6278 sd_free_ctl_entry(&entry->child);
6279 if (entry->proc_handler == NULL)
6280 kfree(entry->procname);
6283 kfree(*tablep);
6284 *tablep = NULL;
6287 static void
6288 set_table_entry(struct ctl_table *entry,
6289 const char *procname, void *data, int maxlen,
6290 mode_t mode, proc_handler *proc_handler)
6292 entry->procname = procname;
6293 entry->data = data;
6294 entry->maxlen = maxlen;
6295 entry->mode = mode;
6296 entry->proc_handler = proc_handler;
6299 static struct ctl_table *
6300 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6302 struct ctl_table *table = sd_alloc_ctl_entry(12);
6304 if (table == NULL)
6305 return NULL;
6307 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6308 sizeof(long), 0644, proc_doulongvec_minmax);
6309 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6310 sizeof(long), 0644, proc_doulongvec_minmax);
6311 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6312 sizeof(int), 0644, proc_dointvec_minmax);
6313 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6314 sizeof(int), 0644, proc_dointvec_minmax);
6315 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6316 sizeof(int), 0644, proc_dointvec_minmax);
6317 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6318 sizeof(int), 0644, proc_dointvec_minmax);
6319 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6320 sizeof(int), 0644, proc_dointvec_minmax);
6321 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6322 sizeof(int), 0644, proc_dointvec_minmax);
6323 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6324 sizeof(int), 0644, proc_dointvec_minmax);
6325 set_table_entry(&table[9], "cache_nice_tries",
6326 &sd->cache_nice_tries,
6327 sizeof(int), 0644, proc_dointvec_minmax);
6328 set_table_entry(&table[10], "flags", &sd->flags,
6329 sizeof(int), 0644, proc_dointvec_minmax);
6330 /* &table[11] is terminator */
6332 return table;
6335 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6337 struct ctl_table *entry, *table;
6338 struct sched_domain *sd;
6339 int domain_num = 0, i;
6340 char buf[32];
6342 for_each_domain(cpu, sd)
6343 domain_num++;
6344 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6345 if (table == NULL)
6346 return NULL;
6348 i = 0;
6349 for_each_domain(cpu, sd) {
6350 snprintf(buf, 32, "domain%d", i);
6351 entry->procname = kstrdup(buf, GFP_KERNEL);
6352 entry->mode = 0555;
6353 entry->child = sd_alloc_ctl_domain_table(sd);
6354 entry++;
6355 i++;
6357 return table;
6360 static struct ctl_table_header *sd_sysctl_header;
6361 static void register_sched_domain_sysctl(void)
6363 int i, cpu_num = num_online_cpus();
6364 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6365 char buf[32];
6367 WARN_ON(sd_ctl_dir[0].child);
6368 sd_ctl_dir[0].child = entry;
6370 if (entry == NULL)
6371 return;
6373 for_each_online_cpu(i) {
6374 snprintf(buf, 32, "cpu%d", i);
6375 entry->procname = kstrdup(buf, GFP_KERNEL);
6376 entry->mode = 0555;
6377 entry->child = sd_alloc_ctl_cpu_table(i);
6378 entry++;
6381 WARN_ON(sd_sysctl_header);
6382 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6385 /* may be called multiple times per register */
6386 static void unregister_sched_domain_sysctl(void)
6388 if (sd_sysctl_header)
6389 unregister_sysctl_table(sd_sysctl_header);
6390 sd_sysctl_header = NULL;
6391 if (sd_ctl_dir[0].child)
6392 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6394 #else
6395 static void register_sched_domain_sysctl(void)
6398 static void unregister_sched_domain_sysctl(void)
6401 #endif
6404 * migration_call - callback that gets triggered when a CPU is added.
6405 * Here we can start up the necessary migration thread for the new CPU.
6407 static int __cpuinit
6408 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6410 struct task_struct *p;
6411 int cpu = (long)hcpu;
6412 unsigned long flags;
6413 struct rq *rq;
6415 switch (action) {
6417 case CPU_UP_PREPARE:
6418 case CPU_UP_PREPARE_FROZEN:
6419 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6420 if (IS_ERR(p))
6421 return NOTIFY_BAD;
6422 kthread_bind(p, cpu);
6423 /* Must be high prio: stop_machine expects to yield to it. */
6424 rq = task_rq_lock(p, &flags);
6425 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6426 task_rq_unlock(rq, &flags);
6427 cpu_rq(cpu)->migration_thread = p;
6428 break;
6430 case CPU_ONLINE:
6431 case CPU_ONLINE_FROZEN:
6432 /* Strictly unnecessary, as first user will wake it. */
6433 wake_up_process(cpu_rq(cpu)->migration_thread);
6435 /* Update our root-domain */
6436 rq = cpu_rq(cpu);
6437 spin_lock_irqsave(&rq->lock, flags);
6438 if (rq->rd) {
6439 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6440 cpu_set(cpu, rq->rd->online);
6442 spin_unlock_irqrestore(&rq->lock, flags);
6443 break;
6445 #ifdef CONFIG_HOTPLUG_CPU
6446 case CPU_UP_CANCELED:
6447 case CPU_UP_CANCELED_FROZEN:
6448 if (!cpu_rq(cpu)->migration_thread)
6449 break;
6450 /* Unbind it from offline cpu so it can run. Fall thru. */
6451 kthread_bind(cpu_rq(cpu)->migration_thread,
6452 any_online_cpu(cpu_online_map));
6453 kthread_stop(cpu_rq(cpu)->migration_thread);
6454 cpu_rq(cpu)->migration_thread = NULL;
6455 break;
6457 case CPU_DEAD:
6458 case CPU_DEAD_FROZEN:
6459 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6460 migrate_live_tasks(cpu);
6461 rq = cpu_rq(cpu);
6462 kthread_stop(rq->migration_thread);
6463 rq->migration_thread = NULL;
6464 /* Idle task back to normal (off runqueue, low prio) */
6465 spin_lock_irq(&rq->lock);
6466 update_rq_clock(rq);
6467 deactivate_task(rq, rq->idle, 0);
6468 rq->idle->static_prio = MAX_PRIO;
6469 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6470 rq->idle->sched_class = &idle_sched_class;
6471 migrate_dead_tasks(cpu);
6472 spin_unlock_irq(&rq->lock);
6473 cpuset_unlock();
6474 migrate_nr_uninterruptible(rq);
6475 BUG_ON(rq->nr_running != 0);
6478 * No need to migrate the tasks: it was best-effort if
6479 * they didn't take sched_hotcpu_mutex. Just wake up
6480 * the requestors.
6482 spin_lock_irq(&rq->lock);
6483 while (!list_empty(&rq->migration_queue)) {
6484 struct migration_req *req;
6486 req = list_entry(rq->migration_queue.next,
6487 struct migration_req, list);
6488 list_del_init(&req->list);
6489 complete(&req->done);
6491 spin_unlock_irq(&rq->lock);
6492 break;
6494 case CPU_DYING:
6495 case CPU_DYING_FROZEN:
6496 /* Update our root-domain */
6497 rq = cpu_rq(cpu);
6498 spin_lock_irqsave(&rq->lock, flags);
6499 if (rq->rd) {
6500 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6501 cpu_clear(cpu, rq->rd->online);
6503 spin_unlock_irqrestore(&rq->lock, flags);
6504 break;
6505 #endif
6507 return NOTIFY_OK;
6510 /* Register at highest priority so that task migration (migrate_all_tasks)
6511 * happens before everything else.
6513 static struct notifier_block __cpuinitdata migration_notifier = {
6514 .notifier_call = migration_call,
6515 .priority = 10
6518 void __init migration_init(void)
6520 void *cpu = (void *)(long)smp_processor_id();
6521 int err;
6523 /* Start one for the boot CPU: */
6524 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6525 BUG_ON(err == NOTIFY_BAD);
6526 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6527 register_cpu_notifier(&migration_notifier);
6529 #endif
6531 #ifdef CONFIG_SMP
6533 #ifdef CONFIG_SCHED_DEBUG
6535 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6536 cpumask_t *groupmask)
6538 struct sched_group *group = sd->groups;
6539 char str[256];
6541 cpulist_scnprintf(str, sizeof(str), sd->span);
6542 cpus_clear(*groupmask);
6544 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6546 if (!(sd->flags & SD_LOAD_BALANCE)) {
6547 printk("does not load-balance\n");
6548 if (sd->parent)
6549 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6550 " has parent");
6551 return -1;
6554 printk(KERN_CONT "span %s\n", str);
6556 if (!cpu_isset(cpu, sd->span)) {
6557 printk(KERN_ERR "ERROR: domain->span does not contain "
6558 "CPU%d\n", cpu);
6560 if (!cpu_isset(cpu, group->cpumask)) {
6561 printk(KERN_ERR "ERROR: domain->groups does not contain"
6562 " CPU%d\n", cpu);
6565 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6566 do {
6567 if (!group) {
6568 printk("\n");
6569 printk(KERN_ERR "ERROR: group is NULL\n");
6570 break;
6573 if (!group->__cpu_power) {
6574 printk(KERN_CONT "\n");
6575 printk(KERN_ERR "ERROR: domain->cpu_power not "
6576 "set\n");
6577 break;
6580 if (!cpus_weight(group->cpumask)) {
6581 printk(KERN_CONT "\n");
6582 printk(KERN_ERR "ERROR: empty group\n");
6583 break;
6586 if (cpus_intersects(*groupmask, group->cpumask)) {
6587 printk(KERN_CONT "\n");
6588 printk(KERN_ERR "ERROR: repeated CPUs\n");
6589 break;
6592 cpus_or(*groupmask, *groupmask, group->cpumask);
6594 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6595 printk(KERN_CONT " %s", str);
6597 group = group->next;
6598 } while (group != sd->groups);
6599 printk(KERN_CONT "\n");
6601 if (!cpus_equal(sd->span, *groupmask))
6602 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6604 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6605 printk(KERN_ERR "ERROR: parent span is not a superset "
6606 "of domain->span\n");
6607 return 0;
6610 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6612 cpumask_t *groupmask;
6613 int level = 0;
6615 if (!sd) {
6616 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6617 return;
6620 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6622 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6623 if (!groupmask) {
6624 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6625 return;
6628 for (;;) {
6629 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6630 break;
6631 level++;
6632 sd = sd->parent;
6633 if (!sd)
6634 break;
6636 kfree(groupmask);
6638 #else
6639 # define sched_domain_debug(sd, cpu) do { } while (0)
6640 #endif
6642 static int sd_degenerate(struct sched_domain *sd)
6644 if (cpus_weight(sd->span) == 1)
6645 return 1;
6647 /* Following flags need at least 2 groups */
6648 if (sd->flags & (SD_LOAD_BALANCE |
6649 SD_BALANCE_NEWIDLE |
6650 SD_BALANCE_FORK |
6651 SD_BALANCE_EXEC |
6652 SD_SHARE_CPUPOWER |
6653 SD_SHARE_PKG_RESOURCES)) {
6654 if (sd->groups != sd->groups->next)
6655 return 0;
6658 /* Following flags don't use groups */
6659 if (sd->flags & (SD_WAKE_IDLE |
6660 SD_WAKE_AFFINE |
6661 SD_WAKE_BALANCE))
6662 return 0;
6664 return 1;
6667 static int
6668 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6670 unsigned long cflags = sd->flags, pflags = parent->flags;
6672 if (sd_degenerate(parent))
6673 return 1;
6675 if (!cpus_equal(sd->span, parent->span))
6676 return 0;
6678 /* Does parent contain flags not in child? */
6679 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6680 if (cflags & SD_WAKE_AFFINE)
6681 pflags &= ~SD_WAKE_BALANCE;
6682 /* Flags needing groups don't count if only 1 group in parent */
6683 if (parent->groups == parent->groups->next) {
6684 pflags &= ~(SD_LOAD_BALANCE |
6685 SD_BALANCE_NEWIDLE |
6686 SD_BALANCE_FORK |
6687 SD_BALANCE_EXEC |
6688 SD_SHARE_CPUPOWER |
6689 SD_SHARE_PKG_RESOURCES);
6691 if (~cflags & pflags)
6692 return 0;
6694 return 1;
6697 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6699 unsigned long flags;
6700 const struct sched_class *class;
6702 spin_lock_irqsave(&rq->lock, flags);
6704 if (rq->rd) {
6705 struct root_domain *old_rd = rq->rd;
6707 for (class = sched_class_highest; class; class = class->next) {
6708 if (class->leave_domain)
6709 class->leave_domain(rq);
6712 cpu_clear(rq->cpu, old_rd->span);
6713 cpu_clear(rq->cpu, old_rd->online);
6715 if (atomic_dec_and_test(&old_rd->refcount))
6716 kfree(old_rd);
6719 atomic_inc(&rd->refcount);
6720 rq->rd = rd;
6722 cpu_set(rq->cpu, rd->span);
6723 if (cpu_isset(rq->cpu, cpu_online_map))
6724 cpu_set(rq->cpu, rd->online);
6726 for (class = sched_class_highest; class; class = class->next) {
6727 if (class->join_domain)
6728 class->join_domain(rq);
6731 spin_unlock_irqrestore(&rq->lock, flags);
6734 static void init_rootdomain(struct root_domain *rd)
6736 memset(rd, 0, sizeof(*rd));
6738 cpus_clear(rd->span);
6739 cpus_clear(rd->online);
6742 static void init_defrootdomain(void)
6744 init_rootdomain(&def_root_domain);
6745 atomic_set(&def_root_domain.refcount, 1);
6748 static struct root_domain *alloc_rootdomain(void)
6750 struct root_domain *rd;
6752 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6753 if (!rd)
6754 return NULL;
6756 init_rootdomain(rd);
6758 return rd;
6762 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6763 * hold the hotplug lock.
6765 static void
6766 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6768 struct rq *rq = cpu_rq(cpu);
6769 struct sched_domain *tmp;
6771 /* Remove the sched domains which do not contribute to scheduling. */
6772 for (tmp = sd; tmp; tmp = tmp->parent) {
6773 struct sched_domain *parent = tmp->parent;
6774 if (!parent)
6775 break;
6776 if (sd_parent_degenerate(tmp, parent)) {
6777 tmp->parent = parent->parent;
6778 if (parent->parent)
6779 parent->parent->child = tmp;
6783 if (sd && sd_degenerate(sd)) {
6784 sd = sd->parent;
6785 if (sd)
6786 sd->child = NULL;
6789 sched_domain_debug(sd, cpu);
6791 rq_attach_root(rq, rd);
6792 rcu_assign_pointer(rq->sd, sd);
6795 /* cpus with isolated domains */
6796 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6798 /* Setup the mask of cpus configured for isolated domains */
6799 static int __init isolated_cpu_setup(char *str)
6801 int ints[NR_CPUS], i;
6803 str = get_options(str, ARRAY_SIZE(ints), ints);
6804 cpus_clear(cpu_isolated_map);
6805 for (i = 1; i <= ints[0]; i++)
6806 if (ints[i] < NR_CPUS)
6807 cpu_set(ints[i], cpu_isolated_map);
6808 return 1;
6811 __setup("isolcpus=", isolated_cpu_setup);
6814 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6815 * to a function which identifies what group(along with sched group) a CPU
6816 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6817 * (due to the fact that we keep track of groups covered with a cpumask_t).
6819 * init_sched_build_groups will build a circular linked list of the groups
6820 * covered by the given span, and will set each group's ->cpumask correctly,
6821 * and ->cpu_power to 0.
6823 static void
6824 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6825 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6826 struct sched_group **sg,
6827 cpumask_t *tmpmask),
6828 cpumask_t *covered, cpumask_t *tmpmask)
6830 struct sched_group *first = NULL, *last = NULL;
6831 int i;
6833 cpus_clear(*covered);
6835 for_each_cpu_mask(i, *span) {
6836 struct sched_group *sg;
6837 int group = group_fn(i, cpu_map, &sg, tmpmask);
6838 int j;
6840 if (cpu_isset(i, *covered))
6841 continue;
6843 cpus_clear(sg->cpumask);
6844 sg->__cpu_power = 0;
6846 for_each_cpu_mask(j, *span) {
6847 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6848 continue;
6850 cpu_set(j, *covered);
6851 cpu_set(j, sg->cpumask);
6853 if (!first)
6854 first = sg;
6855 if (last)
6856 last->next = sg;
6857 last = sg;
6859 last->next = first;
6862 #define SD_NODES_PER_DOMAIN 16
6864 #ifdef CONFIG_NUMA
6867 * find_next_best_node - find the next node to include in a sched_domain
6868 * @node: node whose sched_domain we're building
6869 * @used_nodes: nodes already in the sched_domain
6871 * Find the next node to include in a given scheduling domain. Simply
6872 * finds the closest node not already in the @used_nodes map.
6874 * Should use nodemask_t.
6876 static int find_next_best_node(int node, nodemask_t *used_nodes)
6878 int i, n, val, min_val, best_node = 0;
6880 min_val = INT_MAX;
6882 for (i = 0; i < MAX_NUMNODES; i++) {
6883 /* Start at @node */
6884 n = (node + i) % MAX_NUMNODES;
6886 if (!nr_cpus_node(n))
6887 continue;
6889 /* Skip already used nodes */
6890 if (node_isset(n, *used_nodes))
6891 continue;
6893 /* Simple min distance search */
6894 val = node_distance(node, n);
6896 if (val < min_val) {
6897 min_val = val;
6898 best_node = n;
6902 node_set(best_node, *used_nodes);
6903 return best_node;
6907 * sched_domain_node_span - get a cpumask for a node's sched_domain
6908 * @node: node whose cpumask we're constructing
6909 * @span: resulting cpumask
6911 * Given a node, construct a good cpumask for its sched_domain to span. It
6912 * should be one that prevents unnecessary balancing, but also spreads tasks
6913 * out optimally.
6915 static void sched_domain_node_span(int node, cpumask_t *span)
6917 nodemask_t used_nodes;
6918 node_to_cpumask_ptr(nodemask, node);
6919 int i;
6921 cpus_clear(*span);
6922 nodes_clear(used_nodes);
6924 cpus_or(*span, *span, *nodemask);
6925 node_set(node, used_nodes);
6927 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6928 int next_node = find_next_best_node(node, &used_nodes);
6930 node_to_cpumask_ptr_next(nodemask, next_node);
6931 cpus_or(*span, *span, *nodemask);
6934 #endif
6936 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6939 * SMT sched-domains:
6941 #ifdef CONFIG_SCHED_SMT
6942 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6943 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6945 static int
6946 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6947 cpumask_t *unused)
6949 if (sg)
6950 *sg = &per_cpu(sched_group_cpus, cpu);
6951 return cpu;
6953 #endif
6956 * multi-core sched-domains:
6958 #ifdef CONFIG_SCHED_MC
6959 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6960 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6961 #endif
6963 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6964 static int
6965 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6966 cpumask_t *mask)
6968 int group;
6970 *mask = per_cpu(cpu_sibling_map, cpu);
6971 cpus_and(*mask, *mask, *cpu_map);
6972 group = first_cpu(*mask);
6973 if (sg)
6974 *sg = &per_cpu(sched_group_core, group);
6975 return group;
6977 #elif defined(CONFIG_SCHED_MC)
6978 static int
6979 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6980 cpumask_t *unused)
6982 if (sg)
6983 *sg = &per_cpu(sched_group_core, cpu);
6984 return cpu;
6986 #endif
6988 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6989 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6991 static int
6992 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6993 cpumask_t *mask)
6995 int group;
6996 #ifdef CONFIG_SCHED_MC
6997 *mask = cpu_coregroup_map(cpu);
6998 cpus_and(*mask, *mask, *cpu_map);
6999 group = first_cpu(*mask);
7000 #elif defined(CONFIG_SCHED_SMT)
7001 *mask = per_cpu(cpu_sibling_map, cpu);
7002 cpus_and(*mask, *mask, *cpu_map);
7003 group = first_cpu(*mask);
7004 #else
7005 group = cpu;
7006 #endif
7007 if (sg)
7008 *sg = &per_cpu(sched_group_phys, group);
7009 return group;
7012 #ifdef CONFIG_NUMA
7014 * The init_sched_build_groups can't handle what we want to do with node
7015 * groups, so roll our own. Now each node has its own list of groups which
7016 * gets dynamically allocated.
7018 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7019 static struct sched_group ***sched_group_nodes_bycpu;
7021 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7022 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7024 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7025 struct sched_group **sg, cpumask_t *nodemask)
7027 int group;
7029 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7030 cpus_and(*nodemask, *nodemask, *cpu_map);
7031 group = first_cpu(*nodemask);
7033 if (sg)
7034 *sg = &per_cpu(sched_group_allnodes, group);
7035 return group;
7038 static void init_numa_sched_groups_power(struct sched_group *group_head)
7040 struct sched_group *sg = group_head;
7041 int j;
7043 if (!sg)
7044 return;
7045 do {
7046 for_each_cpu_mask(j, sg->cpumask) {
7047 struct sched_domain *sd;
7049 sd = &per_cpu(phys_domains, j);
7050 if (j != first_cpu(sd->groups->cpumask)) {
7052 * Only add "power" once for each
7053 * physical package.
7055 continue;
7058 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7060 sg = sg->next;
7061 } while (sg != group_head);
7063 #endif
7065 #ifdef CONFIG_NUMA
7066 /* Free memory allocated for various sched_group structures */
7067 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7069 int cpu, i;
7071 for_each_cpu_mask(cpu, *cpu_map) {
7072 struct sched_group **sched_group_nodes
7073 = sched_group_nodes_bycpu[cpu];
7075 if (!sched_group_nodes)
7076 continue;
7078 for (i = 0; i < MAX_NUMNODES; i++) {
7079 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7081 *nodemask = node_to_cpumask(i);
7082 cpus_and(*nodemask, *nodemask, *cpu_map);
7083 if (cpus_empty(*nodemask))
7084 continue;
7086 if (sg == NULL)
7087 continue;
7088 sg = sg->next;
7089 next_sg:
7090 oldsg = sg;
7091 sg = sg->next;
7092 kfree(oldsg);
7093 if (oldsg != sched_group_nodes[i])
7094 goto next_sg;
7096 kfree(sched_group_nodes);
7097 sched_group_nodes_bycpu[cpu] = NULL;
7100 #else
7101 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7104 #endif
7107 * Initialize sched groups cpu_power.
7109 * cpu_power indicates the capacity of sched group, which is used while
7110 * distributing the load between different sched groups in a sched domain.
7111 * Typically cpu_power for all the groups in a sched domain will be same unless
7112 * there are asymmetries in the topology. If there are asymmetries, group
7113 * having more cpu_power will pickup more load compared to the group having
7114 * less cpu_power.
7116 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7117 * the maximum number of tasks a group can handle in the presence of other idle
7118 * or lightly loaded groups in the same sched domain.
7120 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7122 struct sched_domain *child;
7123 struct sched_group *group;
7125 WARN_ON(!sd || !sd->groups);
7127 if (cpu != first_cpu(sd->groups->cpumask))
7128 return;
7130 child = sd->child;
7132 sd->groups->__cpu_power = 0;
7135 * For perf policy, if the groups in child domain share resources
7136 * (for example cores sharing some portions of the cache hierarchy
7137 * or SMT), then set this domain groups cpu_power such that each group
7138 * can handle only one task, when there are other idle groups in the
7139 * same sched domain.
7141 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7142 (child->flags &
7143 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7144 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7145 return;
7149 * add cpu_power of each child group to this groups cpu_power
7151 group = child->groups;
7152 do {
7153 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7154 group = group->next;
7155 } while (group != child->groups);
7159 * Initializers for schedule domains
7160 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7163 #define SD_INIT(sd, type) sd_init_##type(sd)
7164 #define SD_INIT_FUNC(type) \
7165 static noinline void sd_init_##type(struct sched_domain *sd) \
7167 memset(sd, 0, sizeof(*sd)); \
7168 *sd = SD_##type##_INIT; \
7169 sd->level = SD_LV_##type; \
7172 SD_INIT_FUNC(CPU)
7173 #ifdef CONFIG_NUMA
7174 SD_INIT_FUNC(ALLNODES)
7175 SD_INIT_FUNC(NODE)
7176 #endif
7177 #ifdef CONFIG_SCHED_SMT
7178 SD_INIT_FUNC(SIBLING)
7179 #endif
7180 #ifdef CONFIG_SCHED_MC
7181 SD_INIT_FUNC(MC)
7182 #endif
7185 * To minimize stack usage kmalloc room for cpumasks and share the
7186 * space as the usage in build_sched_domains() dictates. Used only
7187 * if the amount of space is significant.
7189 struct allmasks {
7190 cpumask_t tmpmask; /* make this one first */
7191 union {
7192 cpumask_t nodemask;
7193 cpumask_t this_sibling_map;
7194 cpumask_t this_core_map;
7196 cpumask_t send_covered;
7198 #ifdef CONFIG_NUMA
7199 cpumask_t domainspan;
7200 cpumask_t covered;
7201 cpumask_t notcovered;
7202 #endif
7205 #if NR_CPUS > 128
7206 #define SCHED_CPUMASK_ALLOC 1
7207 #define SCHED_CPUMASK_FREE(v) kfree(v)
7208 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7209 #else
7210 #define SCHED_CPUMASK_ALLOC 0
7211 #define SCHED_CPUMASK_FREE(v)
7212 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7213 #endif
7215 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7216 ((unsigned long)(a) + offsetof(struct allmasks, v))
7218 static int default_relax_domain_level = -1;
7220 static int __init setup_relax_domain_level(char *str)
7222 default_relax_domain_level = simple_strtoul(str, NULL, 0);
7223 return 1;
7225 __setup("relax_domain_level=", setup_relax_domain_level);
7227 static void set_domain_attribute(struct sched_domain *sd,
7228 struct sched_domain_attr *attr)
7230 int request;
7232 if (!attr || attr->relax_domain_level < 0) {
7233 if (default_relax_domain_level < 0)
7234 return;
7235 else
7236 request = default_relax_domain_level;
7237 } else
7238 request = attr->relax_domain_level;
7239 if (request < sd->level) {
7240 /* turn off idle balance on this domain */
7241 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7242 } else {
7243 /* turn on idle balance on this domain */
7244 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7249 * Build sched domains for a given set of cpus and attach the sched domains
7250 * to the individual cpus
7252 static int __build_sched_domains(const cpumask_t *cpu_map,
7253 struct sched_domain_attr *attr)
7255 int i;
7256 struct root_domain *rd;
7257 SCHED_CPUMASK_DECLARE(allmasks);
7258 cpumask_t *tmpmask;
7259 #ifdef CONFIG_NUMA
7260 struct sched_group **sched_group_nodes = NULL;
7261 int sd_allnodes = 0;
7264 * Allocate the per-node list of sched groups
7266 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
7267 GFP_KERNEL);
7268 if (!sched_group_nodes) {
7269 printk(KERN_WARNING "Can not alloc sched group node list\n");
7270 return -ENOMEM;
7272 #endif
7274 rd = alloc_rootdomain();
7275 if (!rd) {
7276 printk(KERN_WARNING "Cannot alloc root domain\n");
7277 #ifdef CONFIG_NUMA
7278 kfree(sched_group_nodes);
7279 #endif
7280 return -ENOMEM;
7283 #if SCHED_CPUMASK_ALLOC
7284 /* get space for all scratch cpumask variables */
7285 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7286 if (!allmasks) {
7287 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7288 kfree(rd);
7289 #ifdef CONFIG_NUMA
7290 kfree(sched_group_nodes);
7291 #endif
7292 return -ENOMEM;
7294 #endif
7295 tmpmask = (cpumask_t *)allmasks;
7298 #ifdef CONFIG_NUMA
7299 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7300 #endif
7303 * Set up domains for cpus specified by the cpu_map.
7305 for_each_cpu_mask(i, *cpu_map) {
7306 struct sched_domain *sd = NULL, *p;
7307 SCHED_CPUMASK_VAR(nodemask, allmasks);
7309 *nodemask = node_to_cpumask(cpu_to_node(i));
7310 cpus_and(*nodemask, *nodemask, *cpu_map);
7312 #ifdef CONFIG_NUMA
7313 if (cpus_weight(*cpu_map) >
7314 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7315 sd = &per_cpu(allnodes_domains, i);
7316 SD_INIT(sd, ALLNODES);
7317 set_domain_attribute(sd, attr);
7318 sd->span = *cpu_map;
7319 sd->first_cpu = first_cpu(sd->span);
7320 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7321 p = sd;
7322 sd_allnodes = 1;
7323 } else
7324 p = NULL;
7326 sd = &per_cpu(node_domains, i);
7327 SD_INIT(sd, NODE);
7328 set_domain_attribute(sd, attr);
7329 sched_domain_node_span(cpu_to_node(i), &sd->span);
7330 sd->first_cpu = first_cpu(sd->span);
7331 sd->parent = p;
7332 if (p)
7333 p->child = sd;
7334 cpus_and(sd->span, sd->span, *cpu_map);
7335 #endif
7337 p = sd;
7338 sd = &per_cpu(phys_domains, i);
7339 SD_INIT(sd, CPU);
7340 set_domain_attribute(sd, attr);
7341 sd->span = *nodemask;
7342 sd->first_cpu = first_cpu(sd->span);
7343 sd->parent = p;
7344 if (p)
7345 p->child = sd;
7346 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7348 #ifdef CONFIG_SCHED_MC
7349 p = sd;
7350 sd = &per_cpu(core_domains, i);
7351 SD_INIT(sd, MC);
7352 set_domain_attribute(sd, attr);
7353 sd->span = cpu_coregroup_map(i);
7354 sd->first_cpu = first_cpu(sd->span);
7355 cpus_and(sd->span, sd->span, *cpu_map);
7356 sd->parent = p;
7357 p->child = sd;
7358 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7359 #endif
7361 #ifdef CONFIG_SCHED_SMT
7362 p = sd;
7363 sd = &per_cpu(cpu_domains, i);
7364 SD_INIT(sd, SIBLING);
7365 set_domain_attribute(sd, attr);
7366 sd->span = per_cpu(cpu_sibling_map, i);
7367 sd->first_cpu = first_cpu(sd->span);
7368 cpus_and(sd->span, sd->span, *cpu_map);
7369 sd->parent = p;
7370 p->child = sd;
7371 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7372 #endif
7375 #ifdef CONFIG_SCHED_SMT
7376 /* Set up CPU (sibling) groups */
7377 for_each_cpu_mask(i, *cpu_map) {
7378 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7379 SCHED_CPUMASK_VAR(send_covered, allmasks);
7381 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7382 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7383 if (i != first_cpu(*this_sibling_map))
7384 continue;
7386 init_sched_build_groups(this_sibling_map, cpu_map,
7387 &cpu_to_cpu_group,
7388 send_covered, tmpmask);
7390 #endif
7392 #ifdef CONFIG_SCHED_MC
7393 /* Set up multi-core groups */
7394 for_each_cpu_mask(i, *cpu_map) {
7395 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7396 SCHED_CPUMASK_VAR(send_covered, allmasks);
7398 *this_core_map = cpu_coregroup_map(i);
7399 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7400 if (i != first_cpu(*this_core_map))
7401 continue;
7403 init_sched_build_groups(this_core_map, cpu_map,
7404 &cpu_to_core_group,
7405 send_covered, tmpmask);
7407 #endif
7409 /* Set up physical groups */
7410 for (i = 0; i < MAX_NUMNODES; i++) {
7411 SCHED_CPUMASK_VAR(nodemask, allmasks);
7412 SCHED_CPUMASK_VAR(send_covered, allmasks);
7414 *nodemask = node_to_cpumask(i);
7415 cpus_and(*nodemask, *nodemask, *cpu_map);
7416 if (cpus_empty(*nodemask))
7417 continue;
7419 init_sched_build_groups(nodemask, cpu_map,
7420 &cpu_to_phys_group,
7421 send_covered, tmpmask);
7424 #ifdef CONFIG_NUMA
7425 /* Set up node groups */
7426 if (sd_allnodes) {
7427 SCHED_CPUMASK_VAR(send_covered, allmasks);
7429 init_sched_build_groups(cpu_map, cpu_map,
7430 &cpu_to_allnodes_group,
7431 send_covered, tmpmask);
7434 for (i = 0; i < MAX_NUMNODES; i++) {
7435 /* Set up node groups */
7436 struct sched_group *sg, *prev;
7437 SCHED_CPUMASK_VAR(nodemask, allmasks);
7438 SCHED_CPUMASK_VAR(domainspan, allmasks);
7439 SCHED_CPUMASK_VAR(covered, allmasks);
7440 int j;
7442 *nodemask = node_to_cpumask(i);
7443 cpus_clear(*covered);
7445 cpus_and(*nodemask, *nodemask, *cpu_map);
7446 if (cpus_empty(*nodemask)) {
7447 sched_group_nodes[i] = NULL;
7448 continue;
7451 sched_domain_node_span(i, domainspan);
7452 cpus_and(*domainspan, *domainspan, *cpu_map);
7454 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7455 if (!sg) {
7456 printk(KERN_WARNING "Can not alloc domain group for "
7457 "node %d\n", i);
7458 goto error;
7460 sched_group_nodes[i] = sg;
7461 for_each_cpu_mask(j, *nodemask) {
7462 struct sched_domain *sd;
7464 sd = &per_cpu(node_domains, j);
7465 sd->groups = sg;
7467 sg->__cpu_power = 0;
7468 sg->cpumask = *nodemask;
7469 sg->next = sg;
7470 cpus_or(*covered, *covered, *nodemask);
7471 prev = sg;
7473 for (j = 0; j < MAX_NUMNODES; j++) {
7474 SCHED_CPUMASK_VAR(notcovered, allmasks);
7475 int n = (i + j) % MAX_NUMNODES;
7476 node_to_cpumask_ptr(pnodemask, n);
7478 cpus_complement(*notcovered, *covered);
7479 cpus_and(*tmpmask, *notcovered, *cpu_map);
7480 cpus_and(*tmpmask, *tmpmask, *domainspan);
7481 if (cpus_empty(*tmpmask))
7482 break;
7484 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7485 if (cpus_empty(*tmpmask))
7486 continue;
7488 sg = kmalloc_node(sizeof(struct sched_group),
7489 GFP_KERNEL, i);
7490 if (!sg) {
7491 printk(KERN_WARNING
7492 "Can not alloc domain group for node %d\n", j);
7493 goto error;
7495 sg->__cpu_power = 0;
7496 sg->cpumask = *tmpmask;
7497 sg->next = prev->next;
7498 cpus_or(*covered, *covered, *tmpmask);
7499 prev->next = sg;
7500 prev = sg;
7503 #endif
7505 /* Calculate CPU power for physical packages and nodes */
7506 #ifdef CONFIG_SCHED_SMT
7507 for_each_cpu_mask(i, *cpu_map) {
7508 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7510 init_sched_groups_power(i, sd);
7512 #endif
7513 #ifdef CONFIG_SCHED_MC
7514 for_each_cpu_mask(i, *cpu_map) {
7515 struct sched_domain *sd = &per_cpu(core_domains, i);
7517 init_sched_groups_power(i, sd);
7519 #endif
7521 for_each_cpu_mask(i, *cpu_map) {
7522 struct sched_domain *sd = &per_cpu(phys_domains, i);
7524 init_sched_groups_power(i, sd);
7527 #ifdef CONFIG_NUMA
7528 for (i = 0; i < MAX_NUMNODES; i++)
7529 init_numa_sched_groups_power(sched_group_nodes[i]);
7531 if (sd_allnodes) {
7532 struct sched_group *sg;
7534 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7535 tmpmask);
7536 init_numa_sched_groups_power(sg);
7538 #endif
7540 /* Attach the domains */
7541 for_each_cpu_mask(i, *cpu_map) {
7542 struct sched_domain *sd;
7543 #ifdef CONFIG_SCHED_SMT
7544 sd = &per_cpu(cpu_domains, i);
7545 #elif defined(CONFIG_SCHED_MC)
7546 sd = &per_cpu(core_domains, i);
7547 #else
7548 sd = &per_cpu(phys_domains, i);
7549 #endif
7550 cpu_attach_domain(sd, rd, i);
7553 SCHED_CPUMASK_FREE((void *)allmasks);
7554 return 0;
7556 #ifdef CONFIG_NUMA
7557 error:
7558 free_sched_groups(cpu_map, tmpmask);
7559 SCHED_CPUMASK_FREE((void *)allmasks);
7560 return -ENOMEM;
7561 #endif
7564 static int build_sched_domains(const cpumask_t *cpu_map)
7566 return __build_sched_domains(cpu_map, NULL);
7569 static cpumask_t *doms_cur; /* current sched domains */
7570 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7571 static struct sched_domain_attr *dattr_cur; /* attribues of custom domains
7572 in 'doms_cur' */
7575 * Special case: If a kmalloc of a doms_cur partition (array of
7576 * cpumask_t) fails, then fallback to a single sched domain,
7577 * as determined by the single cpumask_t fallback_doms.
7579 static cpumask_t fallback_doms;
7581 void __attribute__((weak)) arch_update_cpu_topology(void)
7586 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7587 * For now this just excludes isolated cpus, but could be used to
7588 * exclude other special cases in the future.
7590 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7592 int err;
7594 arch_update_cpu_topology();
7595 ndoms_cur = 1;
7596 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7597 if (!doms_cur)
7598 doms_cur = &fallback_doms;
7599 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7600 dattr_cur = NULL;
7601 err = build_sched_domains(doms_cur);
7602 register_sched_domain_sysctl();
7604 return err;
7607 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7608 cpumask_t *tmpmask)
7610 free_sched_groups(cpu_map, tmpmask);
7614 * Detach sched domains from a group of cpus specified in cpu_map
7615 * These cpus will now be attached to the NULL domain
7617 static void detach_destroy_domains(const cpumask_t *cpu_map)
7619 cpumask_t tmpmask;
7620 int i;
7622 unregister_sched_domain_sysctl();
7624 for_each_cpu_mask(i, *cpu_map)
7625 cpu_attach_domain(NULL, &def_root_domain, i);
7626 synchronize_sched();
7627 arch_destroy_sched_domains(cpu_map, &tmpmask);
7630 /* handle null as "default" */
7631 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7632 struct sched_domain_attr *new, int idx_new)
7634 struct sched_domain_attr tmp;
7636 /* fast path */
7637 if (!new && !cur)
7638 return 1;
7640 tmp = SD_ATTR_INIT;
7641 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7642 new ? (new + idx_new) : &tmp,
7643 sizeof(struct sched_domain_attr));
7647 * Partition sched domains as specified by the 'ndoms_new'
7648 * cpumasks in the array doms_new[] of cpumasks. This compares
7649 * doms_new[] to the current sched domain partitioning, doms_cur[].
7650 * It destroys each deleted domain and builds each new domain.
7652 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7653 * The masks don't intersect (don't overlap.) We should setup one
7654 * sched domain for each mask. CPUs not in any of the cpumasks will
7655 * not be load balanced. If the same cpumask appears both in the
7656 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7657 * it as it is.
7659 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7660 * ownership of it and will kfree it when done with it. If the caller
7661 * failed the kmalloc call, then it can pass in doms_new == NULL,
7662 * and partition_sched_domains() will fallback to the single partition
7663 * 'fallback_doms'.
7665 * Call with hotplug lock held
7667 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7668 struct sched_domain_attr *dattr_new)
7670 int i, j;
7672 mutex_lock(&sched_domains_mutex);
7674 /* always unregister in case we don't destroy any domains */
7675 unregister_sched_domain_sysctl();
7677 if (doms_new == NULL) {
7678 ndoms_new = 1;
7679 doms_new = &fallback_doms;
7680 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7681 dattr_new = NULL;
7684 /* Destroy deleted domains */
7685 for (i = 0; i < ndoms_cur; i++) {
7686 for (j = 0; j < ndoms_new; j++) {
7687 if (cpus_equal(doms_cur[i], doms_new[j])
7688 && dattrs_equal(dattr_cur, i, dattr_new, j))
7689 goto match1;
7691 /* no match - a current sched domain not in new doms_new[] */
7692 detach_destroy_domains(doms_cur + i);
7693 match1:
7697 /* Build new domains */
7698 for (i = 0; i < ndoms_new; i++) {
7699 for (j = 0; j < ndoms_cur; j++) {
7700 if (cpus_equal(doms_new[i], doms_cur[j])
7701 && dattrs_equal(dattr_new, i, dattr_cur, j))
7702 goto match2;
7704 /* no match - add a new doms_new */
7705 __build_sched_domains(doms_new + i,
7706 dattr_new ? dattr_new + i : NULL);
7707 match2:
7711 /* Remember the new sched domains */
7712 if (doms_cur != &fallback_doms)
7713 kfree(doms_cur);
7714 kfree(dattr_cur); /* kfree(NULL) is safe */
7715 doms_cur = doms_new;
7716 dattr_cur = dattr_new;
7717 ndoms_cur = ndoms_new;
7719 register_sched_domain_sysctl();
7721 mutex_unlock(&sched_domains_mutex);
7724 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7725 int arch_reinit_sched_domains(void)
7727 int err;
7729 get_online_cpus();
7730 mutex_lock(&sched_domains_mutex);
7731 detach_destroy_domains(&cpu_online_map);
7732 err = arch_init_sched_domains(&cpu_online_map);
7733 mutex_unlock(&sched_domains_mutex);
7734 put_online_cpus();
7736 return err;
7739 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7741 int ret;
7743 if (buf[0] != '0' && buf[0] != '1')
7744 return -EINVAL;
7746 if (smt)
7747 sched_smt_power_savings = (buf[0] == '1');
7748 else
7749 sched_mc_power_savings = (buf[0] == '1');
7751 ret = arch_reinit_sched_domains();
7753 return ret ? ret : count;
7756 #ifdef CONFIG_SCHED_MC
7757 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7759 return sprintf(page, "%u\n", sched_mc_power_savings);
7761 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7762 const char *buf, size_t count)
7764 return sched_power_savings_store(buf, count, 0);
7766 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7767 sched_mc_power_savings_store);
7768 #endif
7770 #ifdef CONFIG_SCHED_SMT
7771 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7773 return sprintf(page, "%u\n", sched_smt_power_savings);
7775 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7776 const char *buf, size_t count)
7778 return sched_power_savings_store(buf, count, 1);
7780 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7781 sched_smt_power_savings_store);
7782 #endif
7784 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7786 int err = 0;
7788 #ifdef CONFIG_SCHED_SMT
7789 if (smt_capable())
7790 err = sysfs_create_file(&cls->kset.kobj,
7791 &attr_sched_smt_power_savings.attr);
7792 #endif
7793 #ifdef CONFIG_SCHED_MC
7794 if (!err && mc_capable())
7795 err = sysfs_create_file(&cls->kset.kobj,
7796 &attr_sched_mc_power_savings.attr);
7797 #endif
7798 return err;
7800 #endif
7803 * Force a reinitialization of the sched domains hierarchy. The domains
7804 * and groups cannot be updated in place without racing with the balancing
7805 * code, so we temporarily attach all running cpus to the NULL domain
7806 * which will prevent rebalancing while the sched domains are recalculated.
7808 static int update_sched_domains(struct notifier_block *nfb,
7809 unsigned long action, void *hcpu)
7811 switch (action) {
7812 case CPU_UP_PREPARE:
7813 case CPU_UP_PREPARE_FROZEN:
7814 case CPU_DOWN_PREPARE:
7815 case CPU_DOWN_PREPARE_FROZEN:
7816 detach_destroy_domains(&cpu_online_map);
7817 return NOTIFY_OK;
7819 case CPU_UP_CANCELED:
7820 case CPU_UP_CANCELED_FROZEN:
7821 case CPU_DOWN_FAILED:
7822 case CPU_DOWN_FAILED_FROZEN:
7823 case CPU_ONLINE:
7824 case CPU_ONLINE_FROZEN:
7825 case CPU_DEAD:
7826 case CPU_DEAD_FROZEN:
7828 * Fall through and re-initialise the domains.
7830 break;
7831 default:
7832 return NOTIFY_DONE;
7835 /* The hotplug lock is already held by cpu_up/cpu_down */
7836 arch_init_sched_domains(&cpu_online_map);
7838 return NOTIFY_OK;
7841 void __init sched_init_smp(void)
7843 cpumask_t non_isolated_cpus;
7845 #if defined(CONFIG_NUMA)
7846 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7847 GFP_KERNEL);
7848 BUG_ON(sched_group_nodes_bycpu == NULL);
7849 #endif
7850 get_online_cpus();
7851 mutex_lock(&sched_domains_mutex);
7852 arch_init_sched_domains(&cpu_online_map);
7853 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7854 if (cpus_empty(non_isolated_cpus))
7855 cpu_set(smp_processor_id(), non_isolated_cpus);
7856 mutex_unlock(&sched_domains_mutex);
7857 put_online_cpus();
7858 /* XXX: Theoretical race here - CPU may be hotplugged now */
7859 hotcpu_notifier(update_sched_domains, 0);
7860 init_hrtick();
7862 /* Move init over to a non-isolated CPU */
7863 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7864 BUG();
7865 sched_init_granularity();
7867 #else
7868 void __init sched_init_smp(void)
7870 sched_init_granularity();
7872 #endif /* CONFIG_SMP */
7874 int in_sched_functions(unsigned long addr)
7876 return in_lock_functions(addr) ||
7877 (addr >= (unsigned long)__sched_text_start
7878 && addr < (unsigned long)__sched_text_end);
7881 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7883 cfs_rq->tasks_timeline = RB_ROOT;
7884 INIT_LIST_HEAD(&cfs_rq->tasks);
7885 #ifdef CONFIG_FAIR_GROUP_SCHED
7886 cfs_rq->rq = rq;
7887 #endif
7888 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7891 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7893 struct rt_prio_array *array;
7894 int i;
7896 array = &rt_rq->active;
7897 for (i = 0; i < MAX_RT_PRIO; i++) {
7898 INIT_LIST_HEAD(array->queue + i);
7899 __clear_bit(i, array->bitmap);
7901 /* delimiter for bitsearch: */
7902 __set_bit(MAX_RT_PRIO, array->bitmap);
7904 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7905 rt_rq->highest_prio = MAX_RT_PRIO;
7906 #endif
7907 #ifdef CONFIG_SMP
7908 rt_rq->rt_nr_migratory = 0;
7909 rt_rq->overloaded = 0;
7910 #endif
7912 rt_rq->rt_time = 0;
7913 rt_rq->rt_throttled = 0;
7914 rt_rq->rt_runtime = 0;
7915 spin_lock_init(&rt_rq->rt_runtime_lock);
7917 #ifdef CONFIG_RT_GROUP_SCHED
7918 rt_rq->rt_nr_boosted = 0;
7919 rt_rq->rq = rq;
7920 #endif
7923 #ifdef CONFIG_FAIR_GROUP_SCHED
7924 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7925 struct sched_entity *se, int cpu, int add,
7926 struct sched_entity *parent)
7928 struct rq *rq = cpu_rq(cpu);
7929 tg->cfs_rq[cpu] = cfs_rq;
7930 init_cfs_rq(cfs_rq, rq);
7931 cfs_rq->tg = tg;
7932 if (add)
7933 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7935 tg->se[cpu] = se;
7936 /* se could be NULL for init_task_group */
7937 if (!se)
7938 return;
7940 if (!parent)
7941 se->cfs_rq = &rq->cfs;
7942 else
7943 se->cfs_rq = parent->my_q;
7945 se->my_q = cfs_rq;
7946 se->load.weight = tg->shares;
7947 se->load.inv_weight = 0;
7948 se->parent = parent;
7950 #endif
7952 #ifdef CONFIG_RT_GROUP_SCHED
7953 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7954 struct sched_rt_entity *rt_se, int cpu, int add,
7955 struct sched_rt_entity *parent)
7957 struct rq *rq = cpu_rq(cpu);
7959 tg->rt_rq[cpu] = rt_rq;
7960 init_rt_rq(rt_rq, rq);
7961 rt_rq->tg = tg;
7962 rt_rq->rt_se = rt_se;
7963 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7964 if (add)
7965 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7967 tg->rt_se[cpu] = rt_se;
7968 if (!rt_se)
7969 return;
7971 if (!parent)
7972 rt_se->rt_rq = &rq->rt;
7973 else
7974 rt_se->rt_rq = parent->my_q;
7976 rt_se->rt_rq = &rq->rt;
7977 rt_se->my_q = rt_rq;
7978 rt_se->parent = parent;
7979 INIT_LIST_HEAD(&rt_se->run_list);
7981 #endif
7983 void __init sched_init(void)
7985 int i, j;
7986 unsigned long alloc_size = 0, ptr;
7988 #ifdef CONFIG_FAIR_GROUP_SCHED
7989 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7990 #endif
7991 #ifdef CONFIG_RT_GROUP_SCHED
7992 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7993 #endif
7994 #ifdef CONFIG_USER_SCHED
7995 alloc_size *= 2;
7996 #endif
7998 * As sched_init() is called before page_alloc is setup,
7999 * we use alloc_bootmem().
8001 if (alloc_size) {
8002 ptr = (unsigned long)alloc_bootmem(alloc_size);
8004 #ifdef CONFIG_FAIR_GROUP_SCHED
8005 init_task_group.se = (struct sched_entity **)ptr;
8006 ptr += nr_cpu_ids * sizeof(void **);
8008 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8009 ptr += nr_cpu_ids * sizeof(void **);
8011 #ifdef CONFIG_USER_SCHED
8012 root_task_group.se = (struct sched_entity **)ptr;
8013 ptr += nr_cpu_ids * sizeof(void **);
8015 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8016 ptr += nr_cpu_ids * sizeof(void **);
8017 #endif
8018 #endif
8019 #ifdef CONFIG_RT_GROUP_SCHED
8020 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8021 ptr += nr_cpu_ids * sizeof(void **);
8023 init_task_group.rt_rq = (struct rt_rq **)ptr;
8024 ptr += nr_cpu_ids * sizeof(void **);
8026 #ifdef CONFIG_USER_SCHED
8027 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8028 ptr += nr_cpu_ids * sizeof(void **);
8030 root_task_group.rt_rq = (struct rt_rq **)ptr;
8031 ptr += nr_cpu_ids * sizeof(void **);
8032 #endif
8033 #endif
8036 #ifdef CONFIG_SMP
8037 init_aggregate();
8038 init_defrootdomain();
8039 #endif
8041 init_rt_bandwidth(&def_rt_bandwidth,
8042 global_rt_period(), global_rt_runtime());
8044 #ifdef CONFIG_RT_GROUP_SCHED
8045 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8046 global_rt_period(), global_rt_runtime());
8047 #ifdef CONFIG_USER_SCHED
8048 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8049 global_rt_period(), RUNTIME_INF);
8050 #endif
8051 #endif
8053 #ifdef CONFIG_GROUP_SCHED
8054 list_add(&init_task_group.list, &task_groups);
8055 INIT_LIST_HEAD(&init_task_group.children);
8057 #ifdef CONFIG_USER_SCHED
8058 INIT_LIST_HEAD(&root_task_group.children);
8059 init_task_group.parent = &root_task_group;
8060 list_add(&init_task_group.siblings, &root_task_group.children);
8061 #endif
8062 #endif
8064 for_each_possible_cpu(i) {
8065 struct rq *rq;
8067 rq = cpu_rq(i);
8068 spin_lock_init(&rq->lock);
8069 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
8070 rq->nr_running = 0;
8071 init_cfs_rq(&rq->cfs, rq);
8072 init_rt_rq(&rq->rt, rq);
8073 #ifdef CONFIG_FAIR_GROUP_SCHED
8074 init_task_group.shares = init_task_group_load;
8075 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8076 #ifdef CONFIG_CGROUP_SCHED
8078 * How much cpu bandwidth does init_task_group get?
8080 * In case of task-groups formed thr' the cgroup filesystem, it
8081 * gets 100% of the cpu resources in the system. This overall
8082 * system cpu resource is divided among the tasks of
8083 * init_task_group and its child task-groups in a fair manner,
8084 * based on each entity's (task or task-group's) weight
8085 * (se->load.weight).
8087 * In other words, if init_task_group has 10 tasks of weight
8088 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8089 * then A0's share of the cpu resource is:
8091 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8093 * We achieve this by letting init_task_group's tasks sit
8094 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8096 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8097 #elif defined CONFIG_USER_SCHED
8098 root_task_group.shares = NICE_0_LOAD;
8099 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8101 * In case of task-groups formed thr' the user id of tasks,
8102 * init_task_group represents tasks belonging to root user.
8103 * Hence it forms a sibling of all subsequent groups formed.
8104 * In this case, init_task_group gets only a fraction of overall
8105 * system cpu resource, based on the weight assigned to root
8106 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8107 * by letting tasks of init_task_group sit in a separate cfs_rq
8108 * (init_cfs_rq) and having one entity represent this group of
8109 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8111 init_tg_cfs_entry(&init_task_group,
8112 &per_cpu(init_cfs_rq, i),
8113 &per_cpu(init_sched_entity, i), i, 1,
8114 root_task_group.se[i]);
8116 #endif
8117 #endif /* CONFIG_FAIR_GROUP_SCHED */
8119 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8120 #ifdef CONFIG_RT_GROUP_SCHED
8121 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8122 #ifdef CONFIG_CGROUP_SCHED
8123 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8124 #elif defined CONFIG_USER_SCHED
8125 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8126 init_tg_rt_entry(&init_task_group,
8127 &per_cpu(init_rt_rq, i),
8128 &per_cpu(init_sched_rt_entity, i), i, 1,
8129 root_task_group.rt_se[i]);
8130 #endif
8131 #endif
8133 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8134 rq->cpu_load[j] = 0;
8135 #ifdef CONFIG_SMP
8136 rq->sd = NULL;
8137 rq->rd = NULL;
8138 rq->active_balance = 0;
8139 rq->next_balance = jiffies;
8140 rq->push_cpu = 0;
8141 rq->cpu = i;
8142 rq->migration_thread = NULL;
8143 INIT_LIST_HEAD(&rq->migration_queue);
8144 rq_attach_root(rq, &def_root_domain);
8145 #endif
8146 init_rq_hrtick(rq);
8147 atomic_set(&rq->nr_iowait, 0);
8150 set_load_weight(&init_task);
8152 #ifdef CONFIG_PREEMPT_NOTIFIERS
8153 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8154 #endif
8156 #ifdef CONFIG_SMP
8157 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
8158 #endif
8160 #ifdef CONFIG_RT_MUTEXES
8161 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8162 #endif
8165 * The boot idle thread does lazy MMU switching as well:
8167 atomic_inc(&init_mm.mm_count);
8168 enter_lazy_tlb(&init_mm, current);
8171 * Make us the idle thread. Technically, schedule() should not be
8172 * called from this thread, however somewhere below it might be,
8173 * but because we are the idle thread, we just pick up running again
8174 * when this runqueue becomes "idle".
8176 init_idle(current, smp_processor_id());
8178 * During early bootup we pretend to be a normal task:
8180 current->sched_class = &fair_sched_class;
8182 scheduler_running = 1;
8185 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8186 void __might_sleep(char *file, int line)
8188 #ifdef in_atomic
8189 static unsigned long prev_jiffy; /* ratelimiting */
8191 if ((in_atomic() || irqs_disabled()) &&
8192 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8193 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8194 return;
8195 prev_jiffy = jiffies;
8196 printk(KERN_ERR "BUG: sleeping function called from invalid"
8197 " context at %s:%d\n", file, line);
8198 printk("in_atomic():%d, irqs_disabled():%d\n",
8199 in_atomic(), irqs_disabled());
8200 debug_show_held_locks(current);
8201 if (irqs_disabled())
8202 print_irqtrace_events(current);
8203 dump_stack();
8205 #endif
8207 EXPORT_SYMBOL(__might_sleep);
8208 #endif
8210 #ifdef CONFIG_MAGIC_SYSRQ
8211 static void normalize_task(struct rq *rq, struct task_struct *p)
8213 int on_rq;
8215 update_rq_clock(rq);
8216 on_rq = p->se.on_rq;
8217 if (on_rq)
8218 deactivate_task(rq, p, 0);
8219 __setscheduler(rq, p, SCHED_NORMAL, 0);
8220 if (on_rq) {
8221 activate_task(rq, p, 0);
8222 resched_task(rq->curr);
8226 void normalize_rt_tasks(void)
8228 struct task_struct *g, *p;
8229 unsigned long flags;
8230 struct rq *rq;
8232 read_lock_irqsave(&tasklist_lock, flags);
8233 do_each_thread(g, p) {
8235 * Only normalize user tasks:
8237 if (!p->mm)
8238 continue;
8240 p->se.exec_start = 0;
8241 #ifdef CONFIG_SCHEDSTATS
8242 p->se.wait_start = 0;
8243 p->se.sleep_start = 0;
8244 p->se.block_start = 0;
8245 #endif
8247 if (!rt_task(p)) {
8249 * Renice negative nice level userspace
8250 * tasks back to 0:
8252 if (TASK_NICE(p) < 0 && p->mm)
8253 set_user_nice(p, 0);
8254 continue;
8257 spin_lock(&p->pi_lock);
8258 rq = __task_rq_lock(p);
8260 normalize_task(rq, p);
8262 __task_rq_unlock(rq);
8263 spin_unlock(&p->pi_lock);
8264 } while_each_thread(g, p);
8266 read_unlock_irqrestore(&tasklist_lock, flags);
8269 #endif /* CONFIG_MAGIC_SYSRQ */
8271 #ifdef CONFIG_IA64
8273 * These functions are only useful for the IA64 MCA handling.
8275 * They can only be called when the whole system has been
8276 * stopped - every CPU needs to be quiescent, and no scheduling
8277 * activity can take place. Using them for anything else would
8278 * be a serious bug, and as a result, they aren't even visible
8279 * under any other configuration.
8283 * curr_task - return the current task for a given cpu.
8284 * @cpu: the processor in question.
8286 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8288 struct task_struct *curr_task(int cpu)
8290 return cpu_curr(cpu);
8294 * set_curr_task - set the current task for a given cpu.
8295 * @cpu: the processor in question.
8296 * @p: the task pointer to set.
8298 * Description: This function must only be used when non-maskable interrupts
8299 * are serviced on a separate stack. It allows the architecture to switch the
8300 * notion of the current task on a cpu in a non-blocking manner. This function
8301 * must be called with all CPU's synchronized, and interrupts disabled, the
8302 * and caller must save the original value of the current task (see
8303 * curr_task() above) and restore that value before reenabling interrupts and
8304 * re-starting the system.
8306 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8308 void set_curr_task(int cpu, struct task_struct *p)
8310 cpu_curr(cpu) = p;
8313 #endif
8315 #ifdef CONFIG_FAIR_GROUP_SCHED
8316 static void free_fair_sched_group(struct task_group *tg)
8318 int i;
8320 for_each_possible_cpu(i) {
8321 if (tg->cfs_rq)
8322 kfree(tg->cfs_rq[i]);
8323 if (tg->se)
8324 kfree(tg->se[i]);
8327 kfree(tg->cfs_rq);
8328 kfree(tg->se);
8331 static
8332 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8334 struct cfs_rq *cfs_rq;
8335 struct sched_entity *se, *parent_se;
8336 struct rq *rq;
8337 int i;
8339 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8340 if (!tg->cfs_rq)
8341 goto err;
8342 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8343 if (!tg->se)
8344 goto err;
8346 tg->shares = NICE_0_LOAD;
8348 for_each_possible_cpu(i) {
8349 rq = cpu_rq(i);
8351 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8352 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8353 if (!cfs_rq)
8354 goto err;
8356 se = kmalloc_node(sizeof(struct sched_entity),
8357 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8358 if (!se)
8359 goto err;
8361 parent_se = parent ? parent->se[i] : NULL;
8362 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8365 return 1;
8367 err:
8368 return 0;
8371 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8373 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8374 &cpu_rq(cpu)->leaf_cfs_rq_list);
8377 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8379 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8381 #else
8382 static inline void free_fair_sched_group(struct task_group *tg)
8386 static inline
8387 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8389 return 1;
8392 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8396 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8399 #endif
8401 #ifdef CONFIG_RT_GROUP_SCHED
8402 static void free_rt_sched_group(struct task_group *tg)
8404 int i;
8406 destroy_rt_bandwidth(&tg->rt_bandwidth);
8408 for_each_possible_cpu(i) {
8409 if (tg->rt_rq)
8410 kfree(tg->rt_rq[i]);
8411 if (tg->rt_se)
8412 kfree(tg->rt_se[i]);
8415 kfree(tg->rt_rq);
8416 kfree(tg->rt_se);
8419 static
8420 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8422 struct rt_rq *rt_rq;
8423 struct sched_rt_entity *rt_se, *parent_se;
8424 struct rq *rq;
8425 int i;
8427 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8428 if (!tg->rt_rq)
8429 goto err;
8430 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8431 if (!tg->rt_se)
8432 goto err;
8434 init_rt_bandwidth(&tg->rt_bandwidth,
8435 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8437 for_each_possible_cpu(i) {
8438 rq = cpu_rq(i);
8440 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8441 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8442 if (!rt_rq)
8443 goto err;
8445 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8446 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8447 if (!rt_se)
8448 goto err;
8450 parent_se = parent ? parent->rt_se[i] : NULL;
8451 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8454 return 1;
8456 err:
8457 return 0;
8460 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8462 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8463 &cpu_rq(cpu)->leaf_rt_rq_list);
8466 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8468 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8470 #else
8471 static inline void free_rt_sched_group(struct task_group *tg)
8475 static inline
8476 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8478 return 1;
8481 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8485 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8488 #endif
8490 #ifdef CONFIG_GROUP_SCHED
8491 static void free_sched_group(struct task_group *tg)
8493 free_fair_sched_group(tg);
8494 free_rt_sched_group(tg);
8495 kfree(tg);
8498 /* allocate runqueue etc for a new task group */
8499 struct task_group *sched_create_group(struct task_group *parent)
8501 struct task_group *tg;
8502 unsigned long flags;
8503 int i;
8505 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8506 if (!tg)
8507 return ERR_PTR(-ENOMEM);
8509 if (!alloc_fair_sched_group(tg, parent))
8510 goto err;
8512 if (!alloc_rt_sched_group(tg, parent))
8513 goto err;
8515 spin_lock_irqsave(&task_group_lock, flags);
8516 for_each_possible_cpu(i) {
8517 register_fair_sched_group(tg, i);
8518 register_rt_sched_group(tg, i);
8520 list_add_rcu(&tg->list, &task_groups);
8522 WARN_ON(!parent); /* root should already exist */
8524 tg->parent = parent;
8525 list_add_rcu(&tg->siblings, &parent->children);
8526 INIT_LIST_HEAD(&tg->children);
8527 spin_unlock_irqrestore(&task_group_lock, flags);
8529 return tg;
8531 err:
8532 free_sched_group(tg);
8533 return ERR_PTR(-ENOMEM);
8536 /* rcu callback to free various structures associated with a task group */
8537 static void free_sched_group_rcu(struct rcu_head *rhp)
8539 /* now it should be safe to free those cfs_rqs */
8540 free_sched_group(container_of(rhp, struct task_group, rcu));
8543 /* Destroy runqueue etc associated with a task group */
8544 void sched_destroy_group(struct task_group *tg)
8546 unsigned long flags;
8547 int i;
8549 spin_lock_irqsave(&task_group_lock, flags);
8550 for_each_possible_cpu(i) {
8551 unregister_fair_sched_group(tg, i);
8552 unregister_rt_sched_group(tg, i);
8554 list_del_rcu(&tg->list);
8555 list_del_rcu(&tg->siblings);
8556 spin_unlock_irqrestore(&task_group_lock, flags);
8558 /* wait for possible concurrent references to cfs_rqs complete */
8559 call_rcu(&tg->rcu, free_sched_group_rcu);
8562 /* change task's runqueue when it moves between groups.
8563 * The caller of this function should have put the task in its new group
8564 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8565 * reflect its new group.
8567 void sched_move_task(struct task_struct *tsk)
8569 int on_rq, running;
8570 unsigned long flags;
8571 struct rq *rq;
8573 rq = task_rq_lock(tsk, &flags);
8575 update_rq_clock(rq);
8577 running = task_current(rq, tsk);
8578 on_rq = tsk->se.on_rq;
8580 if (on_rq)
8581 dequeue_task(rq, tsk, 0);
8582 if (unlikely(running))
8583 tsk->sched_class->put_prev_task(rq, tsk);
8585 set_task_rq(tsk, task_cpu(tsk));
8587 #ifdef CONFIG_FAIR_GROUP_SCHED
8588 if (tsk->sched_class->moved_group)
8589 tsk->sched_class->moved_group(tsk);
8590 #endif
8592 if (unlikely(running))
8593 tsk->sched_class->set_curr_task(rq);
8594 if (on_rq)
8595 enqueue_task(rq, tsk, 0);
8597 task_rq_unlock(rq, &flags);
8599 #endif
8601 #ifdef CONFIG_FAIR_GROUP_SCHED
8602 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8604 struct cfs_rq *cfs_rq = se->cfs_rq;
8605 int on_rq;
8607 on_rq = se->on_rq;
8608 if (on_rq)
8609 dequeue_entity(cfs_rq, se, 0);
8611 se->load.weight = shares;
8612 se->load.inv_weight = 0;
8614 if (on_rq)
8615 enqueue_entity(cfs_rq, se, 0);
8618 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8620 struct cfs_rq *cfs_rq = se->cfs_rq;
8621 struct rq *rq = cfs_rq->rq;
8622 unsigned long flags;
8624 spin_lock_irqsave(&rq->lock, flags);
8625 __set_se_shares(se, shares);
8626 spin_unlock_irqrestore(&rq->lock, flags);
8629 static DEFINE_MUTEX(shares_mutex);
8631 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8633 int i;
8634 unsigned long flags;
8637 * We can't change the weight of the root cgroup.
8639 if (!tg->se[0])
8640 return -EINVAL;
8642 if (shares < MIN_SHARES)
8643 shares = MIN_SHARES;
8644 else if (shares > MAX_SHARES)
8645 shares = MAX_SHARES;
8647 mutex_lock(&shares_mutex);
8648 if (tg->shares == shares)
8649 goto done;
8651 spin_lock_irqsave(&task_group_lock, flags);
8652 for_each_possible_cpu(i)
8653 unregister_fair_sched_group(tg, i);
8654 list_del_rcu(&tg->siblings);
8655 spin_unlock_irqrestore(&task_group_lock, flags);
8657 /* wait for any ongoing reference to this group to finish */
8658 synchronize_sched();
8661 * Now we are free to modify the group's share on each cpu
8662 * w/o tripping rebalance_share or load_balance_fair.
8664 tg->shares = shares;
8665 for_each_possible_cpu(i) {
8667 * force a rebalance
8669 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8670 set_se_shares(tg->se[i], shares);
8674 * Enable load balance activity on this group, by inserting it back on
8675 * each cpu's rq->leaf_cfs_rq_list.
8677 spin_lock_irqsave(&task_group_lock, flags);
8678 for_each_possible_cpu(i)
8679 register_fair_sched_group(tg, i);
8680 list_add_rcu(&tg->siblings, &tg->parent->children);
8681 spin_unlock_irqrestore(&task_group_lock, flags);
8682 done:
8683 mutex_unlock(&shares_mutex);
8684 return 0;
8687 unsigned long sched_group_shares(struct task_group *tg)
8689 return tg->shares;
8691 #endif
8693 #ifdef CONFIG_RT_GROUP_SCHED
8695 * Ensure that the real time constraints are schedulable.
8697 static DEFINE_MUTEX(rt_constraints_mutex);
8699 static unsigned long to_ratio(u64 period, u64 runtime)
8701 if (runtime == RUNTIME_INF)
8702 return 1ULL << 16;
8704 return div64_u64(runtime << 16, period);
8707 #ifdef CONFIG_CGROUP_SCHED
8708 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8710 struct task_group *tgi, *parent = tg->parent;
8711 unsigned long total = 0;
8713 if (!parent) {
8714 if (global_rt_period() < period)
8715 return 0;
8717 return to_ratio(period, runtime) <
8718 to_ratio(global_rt_period(), global_rt_runtime());
8721 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8722 return 0;
8724 rcu_read_lock();
8725 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8726 if (tgi == tg)
8727 continue;
8729 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8730 tgi->rt_bandwidth.rt_runtime);
8732 rcu_read_unlock();
8734 return total + to_ratio(period, runtime) <
8735 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8736 parent->rt_bandwidth.rt_runtime);
8738 #elif defined CONFIG_USER_SCHED
8739 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8741 struct task_group *tgi;
8742 unsigned long total = 0;
8743 unsigned long global_ratio =
8744 to_ratio(global_rt_period(), global_rt_runtime());
8746 rcu_read_lock();
8747 list_for_each_entry_rcu(tgi, &task_groups, list) {
8748 if (tgi == tg)
8749 continue;
8751 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8752 tgi->rt_bandwidth.rt_runtime);
8754 rcu_read_unlock();
8756 return total + to_ratio(period, runtime) < global_ratio;
8758 #endif
8760 /* Must be called with tasklist_lock held */
8761 static inline int tg_has_rt_tasks(struct task_group *tg)
8763 struct task_struct *g, *p;
8764 do_each_thread(g, p) {
8765 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8766 return 1;
8767 } while_each_thread(g, p);
8768 return 0;
8771 static int tg_set_bandwidth(struct task_group *tg,
8772 u64 rt_period, u64 rt_runtime)
8774 int i, err = 0;
8776 mutex_lock(&rt_constraints_mutex);
8777 read_lock(&tasklist_lock);
8778 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8779 err = -EBUSY;
8780 goto unlock;
8782 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8783 err = -EINVAL;
8784 goto unlock;
8787 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8788 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8789 tg->rt_bandwidth.rt_runtime = rt_runtime;
8791 for_each_possible_cpu(i) {
8792 struct rt_rq *rt_rq = tg->rt_rq[i];
8794 spin_lock(&rt_rq->rt_runtime_lock);
8795 rt_rq->rt_runtime = rt_runtime;
8796 spin_unlock(&rt_rq->rt_runtime_lock);
8798 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8799 unlock:
8800 read_unlock(&tasklist_lock);
8801 mutex_unlock(&rt_constraints_mutex);
8803 return err;
8806 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8808 u64 rt_runtime, rt_period;
8810 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8811 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8812 if (rt_runtime_us < 0)
8813 rt_runtime = RUNTIME_INF;
8815 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8818 long sched_group_rt_runtime(struct task_group *tg)
8820 u64 rt_runtime_us;
8822 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8823 return -1;
8825 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8826 do_div(rt_runtime_us, NSEC_PER_USEC);
8827 return rt_runtime_us;
8830 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8832 u64 rt_runtime, rt_period;
8834 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8835 rt_runtime = tg->rt_bandwidth.rt_runtime;
8837 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8840 long sched_group_rt_period(struct task_group *tg)
8842 u64 rt_period_us;
8844 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8845 do_div(rt_period_us, NSEC_PER_USEC);
8846 return rt_period_us;
8849 static int sched_rt_global_constraints(void)
8851 int ret = 0;
8853 mutex_lock(&rt_constraints_mutex);
8854 if (!__rt_schedulable(NULL, 1, 0))
8855 ret = -EINVAL;
8856 mutex_unlock(&rt_constraints_mutex);
8858 return ret;
8860 #else
8861 static int sched_rt_global_constraints(void)
8863 unsigned long flags;
8864 int i;
8866 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8867 for_each_possible_cpu(i) {
8868 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8870 spin_lock(&rt_rq->rt_runtime_lock);
8871 rt_rq->rt_runtime = global_rt_runtime();
8872 spin_unlock(&rt_rq->rt_runtime_lock);
8874 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8876 return 0;
8878 #endif
8880 int sched_rt_handler(struct ctl_table *table, int write,
8881 struct file *filp, void __user *buffer, size_t *lenp,
8882 loff_t *ppos)
8884 int ret;
8885 int old_period, old_runtime;
8886 static DEFINE_MUTEX(mutex);
8888 mutex_lock(&mutex);
8889 old_period = sysctl_sched_rt_period;
8890 old_runtime = sysctl_sched_rt_runtime;
8892 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8894 if (!ret && write) {
8895 ret = sched_rt_global_constraints();
8896 if (ret) {
8897 sysctl_sched_rt_period = old_period;
8898 sysctl_sched_rt_runtime = old_runtime;
8899 } else {
8900 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8901 def_rt_bandwidth.rt_period =
8902 ns_to_ktime(global_rt_period());
8905 mutex_unlock(&mutex);
8907 return ret;
8910 #ifdef CONFIG_CGROUP_SCHED
8912 /* return corresponding task_group object of a cgroup */
8913 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8915 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8916 struct task_group, css);
8919 static struct cgroup_subsys_state *
8920 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8922 struct task_group *tg, *parent;
8924 if (!cgrp->parent) {
8925 /* This is early initialization for the top cgroup */
8926 init_task_group.css.cgroup = cgrp;
8927 return &init_task_group.css;
8930 parent = cgroup_tg(cgrp->parent);
8931 tg = sched_create_group(parent);
8932 if (IS_ERR(tg))
8933 return ERR_PTR(-ENOMEM);
8935 /* Bind the cgroup to task_group object we just created */
8936 tg->css.cgroup = cgrp;
8938 return &tg->css;
8941 static void
8942 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8944 struct task_group *tg = cgroup_tg(cgrp);
8946 sched_destroy_group(tg);
8949 static int
8950 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8951 struct task_struct *tsk)
8953 #ifdef CONFIG_RT_GROUP_SCHED
8954 /* Don't accept realtime tasks when there is no way for them to run */
8955 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8956 return -EINVAL;
8957 #else
8958 /* We don't support RT-tasks being in separate groups */
8959 if (tsk->sched_class != &fair_sched_class)
8960 return -EINVAL;
8961 #endif
8963 return 0;
8966 static void
8967 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8968 struct cgroup *old_cont, struct task_struct *tsk)
8970 sched_move_task(tsk);
8973 #ifdef CONFIG_FAIR_GROUP_SCHED
8974 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8975 u64 shareval)
8977 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8980 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8982 struct task_group *tg = cgroup_tg(cgrp);
8984 return (u64) tg->shares;
8986 #endif
8988 #ifdef CONFIG_RT_GROUP_SCHED
8989 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8990 s64 val)
8992 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8995 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8997 return sched_group_rt_runtime(cgroup_tg(cgrp));
9000 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9001 u64 rt_period_us)
9003 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9006 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9008 return sched_group_rt_period(cgroup_tg(cgrp));
9010 #endif
9012 static struct cftype cpu_files[] = {
9013 #ifdef CONFIG_FAIR_GROUP_SCHED
9015 .name = "shares",
9016 .read_u64 = cpu_shares_read_u64,
9017 .write_u64 = cpu_shares_write_u64,
9019 #endif
9020 #ifdef CONFIG_RT_GROUP_SCHED
9022 .name = "rt_runtime_us",
9023 .read_s64 = cpu_rt_runtime_read,
9024 .write_s64 = cpu_rt_runtime_write,
9027 .name = "rt_period_us",
9028 .read_u64 = cpu_rt_period_read_uint,
9029 .write_u64 = cpu_rt_period_write_uint,
9031 #endif
9034 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9036 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9039 struct cgroup_subsys cpu_cgroup_subsys = {
9040 .name = "cpu",
9041 .create = cpu_cgroup_create,
9042 .destroy = cpu_cgroup_destroy,
9043 .can_attach = cpu_cgroup_can_attach,
9044 .attach = cpu_cgroup_attach,
9045 .populate = cpu_cgroup_populate,
9046 .subsys_id = cpu_cgroup_subsys_id,
9047 .early_init = 1,
9050 #endif /* CONFIG_CGROUP_SCHED */
9052 #ifdef CONFIG_CGROUP_CPUACCT
9055 * CPU accounting code for task groups.
9057 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9058 * (balbir@in.ibm.com).
9061 /* track cpu usage of a group of tasks */
9062 struct cpuacct {
9063 struct cgroup_subsys_state css;
9064 /* cpuusage holds pointer to a u64-type object on every cpu */
9065 u64 *cpuusage;
9068 struct cgroup_subsys cpuacct_subsys;
9070 /* return cpu accounting group corresponding to this container */
9071 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9073 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9074 struct cpuacct, css);
9077 /* return cpu accounting group to which this task belongs */
9078 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9080 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9081 struct cpuacct, css);
9084 /* create a new cpu accounting group */
9085 static struct cgroup_subsys_state *cpuacct_create(
9086 struct cgroup_subsys *ss, struct cgroup *cgrp)
9088 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9090 if (!ca)
9091 return ERR_PTR(-ENOMEM);
9093 ca->cpuusage = alloc_percpu(u64);
9094 if (!ca->cpuusage) {
9095 kfree(ca);
9096 return ERR_PTR(-ENOMEM);
9099 return &ca->css;
9102 /* destroy an existing cpu accounting group */
9103 static void
9104 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9106 struct cpuacct *ca = cgroup_ca(cgrp);
9108 free_percpu(ca->cpuusage);
9109 kfree(ca);
9112 /* return total cpu usage (in nanoseconds) of a group */
9113 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9115 struct cpuacct *ca = cgroup_ca(cgrp);
9116 u64 totalcpuusage = 0;
9117 int i;
9119 for_each_possible_cpu(i) {
9120 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9123 * Take rq->lock to make 64-bit addition safe on 32-bit
9124 * platforms.
9126 spin_lock_irq(&cpu_rq(i)->lock);
9127 totalcpuusage += *cpuusage;
9128 spin_unlock_irq(&cpu_rq(i)->lock);
9131 return totalcpuusage;
9134 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9135 u64 reset)
9137 struct cpuacct *ca = cgroup_ca(cgrp);
9138 int err = 0;
9139 int i;
9141 if (reset) {
9142 err = -EINVAL;
9143 goto out;
9146 for_each_possible_cpu(i) {
9147 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9149 spin_lock_irq(&cpu_rq(i)->lock);
9150 *cpuusage = 0;
9151 spin_unlock_irq(&cpu_rq(i)->lock);
9153 out:
9154 return err;
9157 static struct cftype files[] = {
9159 .name = "usage",
9160 .read_u64 = cpuusage_read,
9161 .write_u64 = cpuusage_write,
9165 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9167 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9171 * charge this task's execution time to its accounting group.
9173 * called with rq->lock held.
9175 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9177 struct cpuacct *ca;
9179 if (!cpuacct_subsys.active)
9180 return;
9182 ca = task_ca(tsk);
9183 if (ca) {
9184 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9186 *cpuusage += cputime;
9190 struct cgroup_subsys cpuacct_subsys = {
9191 .name = "cpuacct",
9192 .create = cpuacct_create,
9193 .destroy = cpuacct_destroy,
9194 .populate = cpuacct_populate,
9195 .subsys_id = cpuacct_subsys_id,
9197 #endif /* CONFIG_CGROUP_CPUACCT */