Merge branch 'master' into for-next
[linux-2.6/mini2440.git] / kernel / sched.c
blob4e2f6033565687f648cbd27e8976ea6c647217fa
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 || 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 or 1 can cause arithmetics problems.
316 * A weight of a cfs_rq is the sum of weights of which entities
317 * are queued on this cfs_rq, so a weight of a entity should not be
318 * too large, so as the shares value of a task group.
319 * (The default weight is 1024 - so there's no practical
320 * limitation from this.)
322 #define MIN_SHARES 2
323 #define MAX_SHARES (1UL << 18)
325 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
326 #endif
328 /* Default task group.
329 * Every task in system belong to this group at bootup.
331 struct task_group init_task_group;
333 /* return group to which a task belongs */
334 static inline struct task_group *task_group(struct task_struct *p)
336 struct task_group *tg;
338 #ifdef CONFIG_USER_SCHED
339 tg = p->user->tg;
340 #elif defined(CONFIG_CGROUP_SCHED)
341 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
342 struct task_group, css);
343 #else
344 tg = &init_task_group;
345 #endif
346 return tg;
349 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
350 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
352 #ifdef CONFIG_FAIR_GROUP_SCHED
353 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
354 p->se.parent = task_group(p)->se[cpu];
355 #endif
357 #ifdef CONFIG_RT_GROUP_SCHED
358 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
359 p->rt.parent = task_group(p)->rt_se[cpu];
360 #endif
363 #else
365 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
367 #endif /* CONFIG_GROUP_SCHED */
369 /* CFS-related fields in a runqueue */
370 struct cfs_rq {
371 struct load_weight load;
372 unsigned long nr_running;
374 u64 exec_clock;
375 u64 min_vruntime;
377 struct rb_root tasks_timeline;
378 struct rb_node *rb_leftmost;
380 struct list_head tasks;
381 struct list_head *balance_iterator;
384 * 'curr' points to currently running entity on this cfs_rq.
385 * It is set to NULL otherwise (i.e when none are currently running).
387 struct sched_entity *curr, *next;
389 unsigned long nr_spread_over;
391 #ifdef CONFIG_FAIR_GROUP_SCHED
392 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
395 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
396 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
397 * (like users, containers etc.)
399 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
400 * list is used during load balance.
402 struct list_head leaf_cfs_rq_list;
403 struct task_group *tg; /* group that "owns" this runqueue */
404 #endif
407 /* Real-Time classes' related field in a runqueue: */
408 struct rt_rq {
409 struct rt_prio_array active;
410 unsigned long rt_nr_running;
411 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
412 int highest_prio; /* highest queued rt task prio */
413 #endif
414 #ifdef CONFIG_SMP
415 unsigned long rt_nr_migratory;
416 int overloaded;
417 #endif
418 int rt_throttled;
419 u64 rt_time;
420 u64 rt_runtime;
421 /* Nests inside the rq lock: */
422 spinlock_t rt_runtime_lock;
424 #ifdef CONFIG_RT_GROUP_SCHED
425 unsigned long rt_nr_boosted;
427 struct rq *rq;
428 struct list_head leaf_rt_rq_list;
429 struct task_group *tg;
430 struct sched_rt_entity *rt_se;
431 #endif
434 #ifdef CONFIG_SMP
437 * We add the notion of a root-domain which will be used to define per-domain
438 * variables. Each exclusive cpuset essentially defines an island domain by
439 * fully partitioning the member cpus from any other cpuset. Whenever a new
440 * exclusive cpuset is created, we also create and attach a new root-domain
441 * object.
444 struct root_domain {
445 atomic_t refcount;
446 cpumask_t span;
447 cpumask_t online;
450 * The "RT overload" flag: it gets set if a CPU has more than
451 * one runnable RT task.
453 cpumask_t rto_mask;
454 atomic_t rto_count;
458 * By default the system creates a single root-domain with all cpus as
459 * members (mimicking the global state we have today).
461 static struct root_domain def_root_domain;
463 #endif
466 * This is the main, per-CPU runqueue data structure.
468 * Locking rule: those places that want to lock multiple runqueues
469 * (such as the load balancing or the thread migration code), lock
470 * acquire operations must be ordered by ascending &runqueue.
472 struct rq {
473 /* runqueue lock: */
474 spinlock_t lock;
477 * nr_running and cpu_load should be in the same cacheline because
478 * remote CPUs use both these fields when doing load calculation.
480 unsigned long nr_running;
481 #define CPU_LOAD_IDX_MAX 5
482 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
483 unsigned char idle_at_tick;
484 #ifdef CONFIG_NO_HZ
485 unsigned long last_tick_seen;
486 unsigned char in_nohz_recently;
487 #endif
488 /* capture load from *all* tasks on this cpu: */
489 struct load_weight load;
490 unsigned long nr_load_updates;
491 u64 nr_switches;
493 struct cfs_rq cfs;
494 struct rt_rq rt;
496 #ifdef CONFIG_FAIR_GROUP_SCHED
497 /* list of leaf cfs_rq on this cpu: */
498 struct list_head leaf_cfs_rq_list;
499 #endif
500 #ifdef CONFIG_RT_GROUP_SCHED
501 struct list_head leaf_rt_rq_list;
502 #endif
505 * This is part of a global counter where only the total sum
506 * over all CPUs matters. A task can increase this counter on
507 * one CPU and if it got migrated afterwards it may decrease
508 * it on another CPU. Always updated under the runqueue lock:
510 unsigned long nr_uninterruptible;
512 struct task_struct *curr, *idle;
513 unsigned long next_balance;
514 struct mm_struct *prev_mm;
516 u64 clock;
518 atomic_t nr_iowait;
520 #ifdef CONFIG_SMP
521 struct root_domain *rd;
522 struct sched_domain *sd;
524 /* For active balancing */
525 int active_balance;
526 int push_cpu;
527 /* cpu of this runqueue: */
528 int cpu;
530 struct task_struct *migration_thread;
531 struct list_head migration_queue;
532 #endif
534 #ifdef CONFIG_SCHED_HRTICK
535 unsigned long hrtick_flags;
536 ktime_t hrtick_expire;
537 struct hrtimer hrtick_timer;
538 #endif
540 #ifdef CONFIG_SCHEDSTATS
541 /* latency stats */
542 struct sched_info rq_sched_info;
544 /* sys_sched_yield() stats */
545 unsigned int yld_exp_empty;
546 unsigned int yld_act_empty;
547 unsigned int yld_both_empty;
548 unsigned int yld_count;
550 /* schedule() stats */
551 unsigned int sched_switch;
552 unsigned int sched_count;
553 unsigned int sched_goidle;
555 /* try_to_wake_up() stats */
556 unsigned int ttwu_count;
557 unsigned int ttwu_local;
559 /* BKL stats */
560 unsigned int bkl_count;
561 #endif
562 struct lock_class_key rq_lock_key;
565 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
567 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
569 rq->curr->sched_class->check_preempt_curr(rq, p);
572 static inline int cpu_of(struct rq *rq)
574 #ifdef CONFIG_SMP
575 return rq->cpu;
576 #else
577 return 0;
578 #endif
582 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
583 * See detach_destroy_domains: synchronize_sched for details.
585 * The domain tree of any CPU may only be accessed from within
586 * preempt-disabled sections.
588 #define for_each_domain(cpu, __sd) \
589 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
591 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
592 #define this_rq() (&__get_cpu_var(runqueues))
593 #define task_rq(p) cpu_rq(task_cpu(p))
594 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
596 static inline void update_rq_clock(struct rq *rq)
598 rq->clock = sched_clock_cpu(cpu_of(rq));
602 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
604 #ifdef CONFIG_SCHED_DEBUG
605 # define const_debug __read_mostly
606 #else
607 # define const_debug static const
608 #endif
611 * Debugging: various feature bits
614 #define SCHED_FEAT(name, enabled) \
615 __SCHED_FEAT_##name ,
617 enum {
618 #include "sched_features.h"
621 #undef SCHED_FEAT
623 #define SCHED_FEAT(name, enabled) \
624 (1UL << __SCHED_FEAT_##name) * enabled |
626 const_debug unsigned int sysctl_sched_features =
627 #include "sched_features.h"
630 #undef SCHED_FEAT
632 #ifdef CONFIG_SCHED_DEBUG
633 #define SCHED_FEAT(name, enabled) \
634 #name ,
636 static __read_mostly char *sched_feat_names[] = {
637 #include "sched_features.h"
638 NULL
641 #undef SCHED_FEAT
643 static int sched_feat_open(struct inode *inode, struct file *filp)
645 filp->private_data = inode->i_private;
646 return 0;
649 static ssize_t
650 sched_feat_read(struct file *filp, char __user *ubuf,
651 size_t cnt, loff_t *ppos)
653 char *buf;
654 int r = 0;
655 int len = 0;
656 int i;
658 for (i = 0; sched_feat_names[i]; i++) {
659 len += strlen(sched_feat_names[i]);
660 len += 4;
663 buf = kmalloc(len + 2, GFP_KERNEL);
664 if (!buf)
665 return -ENOMEM;
667 for (i = 0; sched_feat_names[i]; i++) {
668 if (sysctl_sched_features & (1UL << i))
669 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
670 else
671 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
674 r += sprintf(buf + r, "\n");
675 WARN_ON(r >= len + 2);
677 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
679 kfree(buf);
681 return r;
684 static ssize_t
685 sched_feat_write(struct file *filp, const char __user *ubuf,
686 size_t cnt, loff_t *ppos)
688 char buf[64];
689 char *cmp = buf;
690 int neg = 0;
691 int i;
693 if (cnt > 63)
694 cnt = 63;
696 if (copy_from_user(&buf, ubuf, cnt))
697 return -EFAULT;
699 buf[cnt] = 0;
701 if (strncmp(buf, "NO_", 3) == 0) {
702 neg = 1;
703 cmp += 3;
706 for (i = 0; sched_feat_names[i]; i++) {
707 int len = strlen(sched_feat_names[i]);
709 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
710 if (neg)
711 sysctl_sched_features &= ~(1UL << i);
712 else
713 sysctl_sched_features |= (1UL << i);
714 break;
718 if (!sched_feat_names[i])
719 return -EINVAL;
721 filp->f_pos += cnt;
723 return cnt;
726 static struct file_operations sched_feat_fops = {
727 .open = sched_feat_open,
728 .read = sched_feat_read,
729 .write = sched_feat_write,
732 static __init int sched_init_debug(void)
734 debugfs_create_file("sched_features", 0644, NULL, NULL,
735 &sched_feat_fops);
737 return 0;
739 late_initcall(sched_init_debug);
741 #endif
743 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
746 * Number of tasks to iterate in a single balance run.
747 * Limited because this is done with IRQs disabled.
749 const_debug unsigned int sysctl_sched_nr_migrate = 32;
752 * period over which we measure -rt task cpu usage in us.
753 * default: 1s
755 unsigned int sysctl_sched_rt_period = 1000000;
757 static __read_mostly int scheduler_running;
760 * part of the period that we allow rt tasks to run in us.
761 * default: 0.95s
763 int sysctl_sched_rt_runtime = 950000;
765 static inline u64 global_rt_period(void)
767 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
770 static inline u64 global_rt_runtime(void)
772 if (sysctl_sched_rt_period < 0)
773 return RUNTIME_INF;
775 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
778 unsigned long long time_sync_thresh = 100000;
780 static DEFINE_PER_CPU(unsigned long long, time_offset);
781 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
784 * Global lock which we take every now and then to synchronize
785 * the CPUs time. This method is not warp-safe, but it's good
786 * enough to synchronize slowly diverging time sources and thus
787 * it's good enough for tracing:
789 static DEFINE_SPINLOCK(time_sync_lock);
790 static unsigned long long prev_global_time;
792 static unsigned long long __sync_cpu_clock(unsigned long long time, int cpu)
795 * We want this inlined, to not get tracer function calls
796 * in this critical section:
798 spin_acquire(&time_sync_lock.dep_map, 0, 0, _THIS_IP_);
799 __raw_spin_lock(&time_sync_lock.raw_lock);
801 if (time < prev_global_time) {
802 per_cpu(time_offset, cpu) += prev_global_time - time;
803 time = prev_global_time;
804 } else {
805 prev_global_time = time;
808 __raw_spin_unlock(&time_sync_lock.raw_lock);
809 spin_release(&time_sync_lock.dep_map, 1, _THIS_IP_);
811 return time;
814 static unsigned long long __cpu_clock(int cpu)
816 unsigned long long now;
819 * Only call sched_clock() if the scheduler has already been
820 * initialized (some code might call cpu_clock() very early):
822 if (unlikely(!scheduler_running))
823 return 0;
825 now = sched_clock_cpu(cpu);
827 return now;
831 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
832 * clock constructed from sched_clock():
834 unsigned long long cpu_clock(int cpu)
836 unsigned long long prev_cpu_time, time, delta_time;
837 unsigned long flags;
839 local_irq_save(flags);
840 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
841 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
842 delta_time = time-prev_cpu_time;
844 if (unlikely(delta_time > time_sync_thresh)) {
845 time = __sync_cpu_clock(time, cpu);
846 per_cpu(prev_cpu_time, cpu) = time;
848 local_irq_restore(flags);
850 return time;
852 EXPORT_SYMBOL_GPL(cpu_clock);
854 #ifndef prepare_arch_switch
855 # define prepare_arch_switch(next) do { } while (0)
856 #endif
857 #ifndef finish_arch_switch
858 # define finish_arch_switch(prev) do { } while (0)
859 #endif
861 static inline int task_current(struct rq *rq, struct task_struct *p)
863 return rq->curr == p;
866 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
867 static inline int task_running(struct rq *rq, struct task_struct *p)
869 return task_current(rq, p);
872 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
876 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
878 #ifdef CONFIG_DEBUG_SPINLOCK
879 /* this is a valid case when another task releases the spinlock */
880 rq->lock.owner = current;
881 #endif
883 * If we are tracking spinlock dependencies then we have to
884 * fix up the runqueue lock - which gets 'carried over' from
885 * prev into current:
887 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
889 spin_unlock_irq(&rq->lock);
892 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
893 static inline int task_running(struct rq *rq, struct task_struct *p)
895 #ifdef CONFIG_SMP
896 return p->oncpu;
897 #else
898 return task_current(rq, p);
899 #endif
902 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
904 #ifdef CONFIG_SMP
906 * We can optimise this out completely for !SMP, because the
907 * SMP rebalancing from interrupt is the only thing that cares
908 * here.
910 next->oncpu = 1;
911 #endif
912 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
913 spin_unlock_irq(&rq->lock);
914 #else
915 spin_unlock(&rq->lock);
916 #endif
919 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
921 #ifdef CONFIG_SMP
923 * After ->oncpu is cleared, the task can be moved to a different CPU.
924 * We must ensure this doesn't happen until the switch is completely
925 * finished.
927 smp_wmb();
928 prev->oncpu = 0;
929 #endif
930 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
931 local_irq_enable();
932 #endif
934 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
937 * __task_rq_lock - lock the runqueue a given task resides on.
938 * Must be called interrupts disabled.
940 static inline struct rq *__task_rq_lock(struct task_struct *p)
941 __acquires(rq->lock)
943 for (;;) {
944 struct rq *rq = task_rq(p);
945 spin_lock(&rq->lock);
946 if (likely(rq == task_rq(p)))
947 return rq;
948 spin_unlock(&rq->lock);
953 * task_rq_lock - lock the runqueue a given task resides on and disable
954 * interrupts. Note the ordering: we can safely lookup the task_rq without
955 * explicitly disabling preemption.
957 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
958 __acquires(rq->lock)
960 struct rq *rq;
962 for (;;) {
963 local_irq_save(*flags);
964 rq = task_rq(p);
965 spin_lock(&rq->lock);
966 if (likely(rq == task_rq(p)))
967 return rq;
968 spin_unlock_irqrestore(&rq->lock, *flags);
972 static void __task_rq_unlock(struct rq *rq)
973 __releases(rq->lock)
975 spin_unlock(&rq->lock);
978 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
979 __releases(rq->lock)
981 spin_unlock_irqrestore(&rq->lock, *flags);
985 * this_rq_lock - lock this runqueue and disable interrupts.
987 static struct rq *this_rq_lock(void)
988 __acquires(rq->lock)
990 struct rq *rq;
992 local_irq_disable();
993 rq = this_rq();
994 spin_lock(&rq->lock);
996 return rq;
999 static void __resched_task(struct task_struct *p, int tif_bit);
1001 static inline void resched_task(struct task_struct *p)
1003 __resched_task(p, TIF_NEED_RESCHED);
1006 #ifdef CONFIG_SCHED_HRTICK
1008 * Use HR-timers to deliver accurate preemption points.
1010 * Its all a bit involved since we cannot program an hrt while holding the
1011 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1012 * reschedule event.
1014 * When we get rescheduled we reprogram the hrtick_timer outside of the
1015 * rq->lock.
1017 static inline void resched_hrt(struct task_struct *p)
1019 __resched_task(p, TIF_HRTICK_RESCHED);
1022 static inline void resched_rq(struct rq *rq)
1024 unsigned long flags;
1026 spin_lock_irqsave(&rq->lock, flags);
1027 resched_task(rq->curr);
1028 spin_unlock_irqrestore(&rq->lock, flags);
1031 enum {
1032 HRTICK_SET, /* re-programm hrtick_timer */
1033 HRTICK_RESET, /* not a new slice */
1034 HRTICK_BLOCK, /* stop hrtick operations */
1038 * Use hrtick when:
1039 * - enabled by features
1040 * - hrtimer is actually high res
1042 static inline int hrtick_enabled(struct rq *rq)
1044 if (!sched_feat(HRTICK))
1045 return 0;
1046 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
1047 return 0;
1048 return hrtimer_is_hres_active(&rq->hrtick_timer);
1052 * Called to set the hrtick timer state.
1054 * called with rq->lock held and irqs disabled
1056 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1058 assert_spin_locked(&rq->lock);
1061 * preempt at: now + delay
1063 rq->hrtick_expire =
1064 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1066 * indicate we need to program the timer
1068 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1069 if (reset)
1070 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1073 * New slices are called from the schedule path and don't need a
1074 * forced reschedule.
1076 if (reset)
1077 resched_hrt(rq->curr);
1080 static void hrtick_clear(struct rq *rq)
1082 if (hrtimer_active(&rq->hrtick_timer))
1083 hrtimer_cancel(&rq->hrtick_timer);
1087 * Update the timer from the possible pending state.
1089 static void hrtick_set(struct rq *rq)
1091 ktime_t time;
1092 int set, reset;
1093 unsigned long flags;
1095 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1097 spin_lock_irqsave(&rq->lock, flags);
1098 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1099 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1100 time = rq->hrtick_expire;
1101 clear_thread_flag(TIF_HRTICK_RESCHED);
1102 spin_unlock_irqrestore(&rq->lock, flags);
1104 if (set) {
1105 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1106 if (reset && !hrtimer_active(&rq->hrtick_timer))
1107 resched_rq(rq);
1108 } else
1109 hrtick_clear(rq);
1113 * High-resolution timer tick.
1114 * Runs from hardirq context with interrupts disabled.
1116 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1118 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1120 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1122 spin_lock(&rq->lock);
1123 update_rq_clock(rq);
1124 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1125 spin_unlock(&rq->lock);
1127 return HRTIMER_NORESTART;
1130 #ifdef CONFIG_SMP
1131 static void hotplug_hrtick_disable(int cpu)
1133 struct rq *rq = cpu_rq(cpu);
1134 unsigned long flags;
1136 spin_lock_irqsave(&rq->lock, flags);
1137 rq->hrtick_flags = 0;
1138 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1139 spin_unlock_irqrestore(&rq->lock, flags);
1141 hrtick_clear(rq);
1144 static void hotplug_hrtick_enable(int cpu)
1146 struct rq *rq = cpu_rq(cpu);
1147 unsigned long flags;
1149 spin_lock_irqsave(&rq->lock, flags);
1150 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1151 spin_unlock_irqrestore(&rq->lock, flags);
1154 static int
1155 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1157 int cpu = (int)(long)hcpu;
1159 switch (action) {
1160 case CPU_UP_CANCELED:
1161 case CPU_UP_CANCELED_FROZEN:
1162 case CPU_DOWN_PREPARE:
1163 case CPU_DOWN_PREPARE_FROZEN:
1164 case CPU_DEAD:
1165 case CPU_DEAD_FROZEN:
1166 hotplug_hrtick_disable(cpu);
1167 return NOTIFY_OK;
1169 case CPU_UP_PREPARE:
1170 case CPU_UP_PREPARE_FROZEN:
1171 case CPU_DOWN_FAILED:
1172 case CPU_DOWN_FAILED_FROZEN:
1173 case CPU_ONLINE:
1174 case CPU_ONLINE_FROZEN:
1175 hotplug_hrtick_enable(cpu);
1176 return NOTIFY_OK;
1179 return NOTIFY_DONE;
1182 static void init_hrtick(void)
1184 hotcpu_notifier(hotplug_hrtick, 0);
1186 #endif /* CONFIG_SMP */
1188 static void init_rq_hrtick(struct rq *rq)
1190 rq->hrtick_flags = 0;
1191 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1192 rq->hrtick_timer.function = hrtick;
1193 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1196 void hrtick_resched(void)
1198 struct rq *rq;
1199 unsigned long flags;
1201 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1202 return;
1204 local_irq_save(flags);
1205 rq = cpu_rq(smp_processor_id());
1206 hrtick_set(rq);
1207 local_irq_restore(flags);
1209 #else
1210 static inline void hrtick_clear(struct rq *rq)
1214 static inline void hrtick_set(struct rq *rq)
1218 static inline void init_rq_hrtick(struct rq *rq)
1222 void hrtick_resched(void)
1226 static inline void init_hrtick(void)
1229 #endif
1232 * resched_task - mark a task 'to be rescheduled now'.
1234 * On UP this means the setting of the need_resched flag, on SMP it
1235 * might also involve a cross-CPU call to trigger the scheduler on
1236 * the target CPU.
1238 #ifdef CONFIG_SMP
1240 #ifndef tsk_is_polling
1241 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1242 #endif
1244 static void __resched_task(struct task_struct *p, int tif_bit)
1246 int cpu;
1248 assert_spin_locked(&task_rq(p)->lock);
1250 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1251 return;
1253 set_tsk_thread_flag(p, tif_bit);
1255 cpu = task_cpu(p);
1256 if (cpu == smp_processor_id())
1257 return;
1259 /* NEED_RESCHED must be visible before we test polling */
1260 smp_mb();
1261 if (!tsk_is_polling(p))
1262 smp_send_reschedule(cpu);
1265 static void resched_cpu(int cpu)
1267 struct rq *rq = cpu_rq(cpu);
1268 unsigned long flags;
1270 if (!spin_trylock_irqsave(&rq->lock, flags))
1271 return;
1272 resched_task(cpu_curr(cpu));
1273 spin_unlock_irqrestore(&rq->lock, flags);
1276 #ifdef CONFIG_NO_HZ
1278 * When add_timer_on() enqueues a timer into the timer wheel of an
1279 * idle CPU then this timer might expire before the next timer event
1280 * which is scheduled to wake up that CPU. In case of a completely
1281 * idle system the next event might even be infinite time into the
1282 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1283 * leaves the inner idle loop so the newly added timer is taken into
1284 * account when the CPU goes back to idle and evaluates the timer
1285 * wheel for the next timer event.
1287 void wake_up_idle_cpu(int cpu)
1289 struct rq *rq = cpu_rq(cpu);
1291 if (cpu == smp_processor_id())
1292 return;
1295 * This is safe, as this function is called with the timer
1296 * wheel base lock of (cpu) held. When the CPU is on the way
1297 * to idle and has not yet set rq->curr to idle then it will
1298 * be serialized on the timer wheel base lock and take the new
1299 * timer into account automatically.
1301 if (rq->curr != rq->idle)
1302 return;
1305 * We can set TIF_RESCHED on the idle task of the other CPU
1306 * lockless. The worst case is that the other CPU runs the
1307 * idle task through an additional NOOP schedule()
1309 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1311 /* NEED_RESCHED must be visible before we test polling */
1312 smp_mb();
1313 if (!tsk_is_polling(rq->idle))
1314 smp_send_reschedule(cpu);
1316 #endif
1318 #else
1319 static void __resched_task(struct task_struct *p, int tif_bit)
1321 assert_spin_locked(&task_rq(p)->lock);
1322 set_tsk_thread_flag(p, tif_bit);
1324 #endif
1326 #if BITS_PER_LONG == 32
1327 # define WMULT_CONST (~0UL)
1328 #else
1329 # define WMULT_CONST (1UL << 32)
1330 #endif
1332 #define WMULT_SHIFT 32
1335 * Shift right and round:
1337 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1339 static unsigned long
1340 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1341 struct load_weight *lw)
1343 u64 tmp;
1345 if (!lw->inv_weight) {
1346 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1347 lw->inv_weight = 1;
1348 else
1349 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1350 / (lw->weight+1);
1353 tmp = (u64)delta_exec * weight;
1355 * Check whether we'd overflow the 64-bit multiplication:
1357 if (unlikely(tmp > WMULT_CONST))
1358 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1359 WMULT_SHIFT/2);
1360 else
1361 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1363 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1366 static inline unsigned long
1367 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1369 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1372 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1374 lw->weight += inc;
1375 lw->inv_weight = 0;
1378 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1380 lw->weight -= dec;
1381 lw->inv_weight = 0;
1385 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1386 * of tasks with abnormal "nice" values across CPUs the contribution that
1387 * each task makes to its run queue's load is weighted according to its
1388 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1389 * scaled version of the new time slice allocation that they receive on time
1390 * slice expiry etc.
1393 #define WEIGHT_IDLEPRIO 2
1394 #define WMULT_IDLEPRIO (1 << 31)
1397 * Nice levels are multiplicative, with a gentle 10% change for every
1398 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1399 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1400 * that remained on nice 0.
1402 * The "10% effect" is relative and cumulative: from _any_ nice level,
1403 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1404 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1405 * If a task goes up by ~10% and another task goes down by ~10% then
1406 * the relative distance between them is ~25%.)
1408 static const int prio_to_weight[40] = {
1409 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1410 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1411 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1412 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1413 /* 0 */ 1024, 820, 655, 526, 423,
1414 /* 5 */ 335, 272, 215, 172, 137,
1415 /* 10 */ 110, 87, 70, 56, 45,
1416 /* 15 */ 36, 29, 23, 18, 15,
1420 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1422 * In cases where the weight does not change often, we can use the
1423 * precalculated inverse to speed up arithmetics by turning divisions
1424 * into multiplications:
1426 static const u32 prio_to_wmult[40] = {
1427 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1428 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1429 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1430 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1431 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1432 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1433 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1434 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1437 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1440 * runqueue iterator, to support SMP load-balancing between different
1441 * scheduling classes, without having to expose their internal data
1442 * structures to the load-balancing proper:
1444 struct rq_iterator {
1445 void *arg;
1446 struct task_struct *(*start)(void *);
1447 struct task_struct *(*next)(void *);
1450 #ifdef CONFIG_SMP
1451 static unsigned long
1452 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1453 unsigned long max_load_move, struct sched_domain *sd,
1454 enum cpu_idle_type idle, int *all_pinned,
1455 int *this_best_prio, struct rq_iterator *iterator);
1457 static int
1458 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1459 struct sched_domain *sd, enum cpu_idle_type idle,
1460 struct rq_iterator *iterator);
1461 #endif
1463 #ifdef CONFIG_CGROUP_CPUACCT
1464 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1465 #else
1466 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1467 #endif
1469 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1471 update_load_add(&rq->load, load);
1474 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1476 update_load_sub(&rq->load, load);
1479 #ifdef CONFIG_SMP
1480 static unsigned long source_load(int cpu, int type);
1481 static unsigned long target_load(int cpu, int type);
1482 static unsigned long cpu_avg_load_per_task(int cpu);
1483 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1484 #else /* CONFIG_SMP */
1486 #ifdef CONFIG_FAIR_GROUP_SCHED
1487 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1490 #endif
1492 #endif /* CONFIG_SMP */
1494 #include "sched_stats.h"
1495 #include "sched_idletask.c"
1496 #include "sched_fair.c"
1497 #include "sched_rt.c"
1498 #ifdef CONFIG_SCHED_DEBUG
1499 # include "sched_debug.c"
1500 #endif
1502 #define sched_class_highest (&rt_sched_class)
1504 static inline void inc_load(struct rq *rq, const struct task_struct *p)
1506 update_load_add(&rq->load, p->se.load.weight);
1509 static inline void dec_load(struct rq *rq, const struct task_struct *p)
1511 update_load_sub(&rq->load, p->se.load.weight);
1514 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1516 rq->nr_running++;
1517 inc_load(rq, p);
1520 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1522 rq->nr_running--;
1523 dec_load(rq, p);
1526 static void set_load_weight(struct task_struct *p)
1528 if (task_has_rt_policy(p)) {
1529 p->se.load.weight = prio_to_weight[0] * 2;
1530 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1531 return;
1535 * SCHED_IDLE tasks get minimal weight:
1537 if (p->policy == SCHED_IDLE) {
1538 p->se.load.weight = WEIGHT_IDLEPRIO;
1539 p->se.load.inv_weight = WMULT_IDLEPRIO;
1540 return;
1543 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1544 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1547 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1549 sched_info_queued(p);
1550 p->sched_class->enqueue_task(rq, p, wakeup);
1551 p->se.on_rq = 1;
1554 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1556 p->sched_class->dequeue_task(rq, p, sleep);
1557 p->se.on_rq = 0;
1561 * __normal_prio - return the priority that is based on the static prio
1563 static inline int __normal_prio(struct task_struct *p)
1565 return p->static_prio;
1569 * Calculate the expected normal priority: i.e. priority
1570 * without taking RT-inheritance into account. Might be
1571 * boosted by interactivity modifiers. Changes upon fork,
1572 * setprio syscalls, and whenever the interactivity
1573 * estimator recalculates.
1575 static inline int normal_prio(struct task_struct *p)
1577 int prio;
1579 if (task_has_rt_policy(p))
1580 prio = MAX_RT_PRIO-1 - p->rt_priority;
1581 else
1582 prio = __normal_prio(p);
1583 return prio;
1587 * Calculate the current priority, i.e. the priority
1588 * taken into account by the scheduler. This value might
1589 * be boosted by RT tasks, or might be boosted by
1590 * interactivity modifiers. Will be RT if the task got
1591 * RT-boosted. If not then it returns p->normal_prio.
1593 static int effective_prio(struct task_struct *p)
1595 p->normal_prio = normal_prio(p);
1597 * If we are RT tasks or we were boosted to RT priority,
1598 * keep the priority unchanged. Otherwise, update priority
1599 * to the normal priority:
1601 if (!rt_prio(p->prio))
1602 return p->normal_prio;
1603 return p->prio;
1607 * activate_task - move a task to the runqueue.
1609 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1611 if (task_contributes_to_load(p))
1612 rq->nr_uninterruptible--;
1614 enqueue_task(rq, p, wakeup);
1615 inc_nr_running(p, rq);
1619 * deactivate_task - remove a task from the runqueue.
1621 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1623 if (task_contributes_to_load(p))
1624 rq->nr_uninterruptible++;
1626 dequeue_task(rq, p, sleep);
1627 dec_nr_running(p, rq);
1631 * task_curr - is this task currently executing on a CPU?
1632 * @p: the task in question.
1634 inline int task_curr(const struct task_struct *p)
1636 return cpu_curr(task_cpu(p)) == p;
1639 /* Used instead of source_load when we know the type == 0 */
1640 unsigned long weighted_cpuload(const int cpu)
1642 return cpu_rq(cpu)->load.weight;
1645 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1647 set_task_rq(p, cpu);
1648 #ifdef CONFIG_SMP
1650 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1651 * successfuly executed on another CPU. We must ensure that updates of
1652 * per-task data have been completed by this moment.
1654 smp_wmb();
1655 task_thread_info(p)->cpu = cpu;
1656 #endif
1659 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1660 const struct sched_class *prev_class,
1661 int oldprio, int running)
1663 if (prev_class != p->sched_class) {
1664 if (prev_class->switched_from)
1665 prev_class->switched_from(rq, p, running);
1666 p->sched_class->switched_to(rq, p, running);
1667 } else
1668 p->sched_class->prio_changed(rq, p, oldprio, running);
1671 #ifdef CONFIG_SMP
1674 * Is this task likely cache-hot:
1676 static int
1677 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1679 s64 delta;
1682 * Buddy candidates are cache hot:
1684 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1685 return 1;
1687 if (p->sched_class != &fair_sched_class)
1688 return 0;
1690 if (sysctl_sched_migration_cost == -1)
1691 return 1;
1692 if (sysctl_sched_migration_cost == 0)
1693 return 0;
1695 delta = now - p->se.exec_start;
1697 return delta < (s64)sysctl_sched_migration_cost;
1701 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1703 int old_cpu = task_cpu(p);
1704 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1705 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1706 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1707 u64 clock_offset;
1709 clock_offset = old_rq->clock - new_rq->clock;
1711 #ifdef CONFIG_SCHEDSTATS
1712 if (p->se.wait_start)
1713 p->se.wait_start -= clock_offset;
1714 if (p->se.sleep_start)
1715 p->se.sleep_start -= clock_offset;
1716 if (p->se.block_start)
1717 p->se.block_start -= clock_offset;
1718 if (old_cpu != new_cpu) {
1719 schedstat_inc(p, se.nr_migrations);
1720 if (task_hot(p, old_rq->clock, NULL))
1721 schedstat_inc(p, se.nr_forced2_migrations);
1723 #endif
1724 p->se.vruntime -= old_cfsrq->min_vruntime -
1725 new_cfsrq->min_vruntime;
1727 __set_task_cpu(p, new_cpu);
1730 struct migration_req {
1731 struct list_head list;
1733 struct task_struct *task;
1734 int dest_cpu;
1736 struct completion done;
1740 * The task's runqueue lock must be held.
1741 * Returns true if you have to wait for migration thread.
1743 static int
1744 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1746 struct rq *rq = task_rq(p);
1749 * If the task is not on a runqueue (and not running), then
1750 * it is sufficient to simply update the task's cpu field.
1752 if (!p->se.on_rq && !task_running(rq, p)) {
1753 set_task_cpu(p, dest_cpu);
1754 return 0;
1757 init_completion(&req->done);
1758 req->task = p;
1759 req->dest_cpu = dest_cpu;
1760 list_add(&req->list, &rq->migration_queue);
1762 return 1;
1766 * wait_task_inactive - wait for a thread to unschedule.
1768 * The caller must ensure that the task *will* unschedule sometime soon,
1769 * else this function might spin for a *long* time. This function can't
1770 * be called with interrupts off, or it may introduce deadlock with
1771 * smp_call_function() if an IPI is sent by the same process we are
1772 * waiting to become inactive.
1774 void wait_task_inactive(struct task_struct *p)
1776 unsigned long flags;
1777 int running, on_rq;
1778 struct rq *rq;
1780 for (;;) {
1782 * We do the initial early heuristics without holding
1783 * any task-queue locks at all. We'll only try to get
1784 * the runqueue lock when things look like they will
1785 * work out!
1787 rq = task_rq(p);
1790 * If the task is actively running on another CPU
1791 * still, just relax and busy-wait without holding
1792 * any locks.
1794 * NOTE! Since we don't hold any locks, it's not
1795 * even sure that "rq" stays as the right runqueue!
1796 * But we don't care, since "task_running()" will
1797 * return false if the runqueue has changed and p
1798 * is actually now running somewhere else!
1800 while (task_running(rq, p))
1801 cpu_relax();
1804 * Ok, time to look more closely! We need the rq
1805 * lock now, to be *sure*. If we're wrong, we'll
1806 * just go back and repeat.
1808 rq = task_rq_lock(p, &flags);
1809 running = task_running(rq, p);
1810 on_rq = p->se.on_rq;
1811 task_rq_unlock(rq, &flags);
1814 * Was it really running after all now that we
1815 * checked with the proper locks actually held?
1817 * Oops. Go back and try again..
1819 if (unlikely(running)) {
1820 cpu_relax();
1821 continue;
1825 * It's not enough that it's not actively running,
1826 * it must be off the runqueue _entirely_, and not
1827 * preempted!
1829 * So if it wa still runnable (but just not actively
1830 * running right now), it's preempted, and we should
1831 * yield - it could be a while.
1833 if (unlikely(on_rq)) {
1834 schedule_timeout_uninterruptible(1);
1835 continue;
1839 * Ahh, all good. It wasn't running, and it wasn't
1840 * runnable, which means that it will never become
1841 * running in the future either. We're all done!
1843 break;
1847 /***
1848 * kick_process - kick a running thread to enter/exit the kernel
1849 * @p: the to-be-kicked thread
1851 * Cause a process which is running on another CPU to enter
1852 * kernel-mode, without any delay. (to get signals handled.)
1854 * NOTE: this function doesnt have to take the runqueue lock,
1855 * because all it wants to ensure is that the remote task enters
1856 * the kernel. If the IPI races and the task has been migrated
1857 * to another CPU then no harm is done and the purpose has been
1858 * achieved as well.
1860 void kick_process(struct task_struct *p)
1862 int cpu;
1864 preempt_disable();
1865 cpu = task_cpu(p);
1866 if ((cpu != smp_processor_id()) && task_curr(p))
1867 smp_send_reschedule(cpu);
1868 preempt_enable();
1872 * Return a low guess at the load of a migration-source cpu weighted
1873 * according to the scheduling class and "nice" value.
1875 * We want to under-estimate the load of migration sources, to
1876 * balance conservatively.
1878 static unsigned long source_load(int cpu, int type)
1880 struct rq *rq = cpu_rq(cpu);
1881 unsigned long total = weighted_cpuload(cpu);
1883 if (type == 0)
1884 return total;
1886 return min(rq->cpu_load[type-1], total);
1890 * Return a high guess at the load of a migration-target cpu weighted
1891 * according to the scheduling class and "nice" value.
1893 static unsigned long target_load(int cpu, int type)
1895 struct rq *rq = cpu_rq(cpu);
1896 unsigned long total = weighted_cpuload(cpu);
1898 if (type == 0)
1899 return total;
1901 return max(rq->cpu_load[type-1], total);
1905 * Return the average load per task on the cpu's run queue
1907 static unsigned long cpu_avg_load_per_task(int cpu)
1909 struct rq *rq = cpu_rq(cpu);
1910 unsigned long total = weighted_cpuload(cpu);
1911 unsigned long n = rq->nr_running;
1913 return n ? total / n : SCHED_LOAD_SCALE;
1917 * find_idlest_group finds and returns the least busy CPU group within the
1918 * domain.
1920 static struct sched_group *
1921 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1923 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1924 unsigned long min_load = ULONG_MAX, this_load = 0;
1925 int load_idx = sd->forkexec_idx;
1926 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1928 do {
1929 unsigned long load, avg_load;
1930 int local_group;
1931 int i;
1933 /* Skip over this group if it has no CPUs allowed */
1934 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1935 continue;
1937 local_group = cpu_isset(this_cpu, group->cpumask);
1939 /* Tally up the load of all CPUs in the group */
1940 avg_load = 0;
1942 for_each_cpu_mask(i, group->cpumask) {
1943 /* Bias balancing toward cpus of our domain */
1944 if (local_group)
1945 load = source_load(i, load_idx);
1946 else
1947 load = target_load(i, load_idx);
1949 avg_load += load;
1952 /* Adjust by relative CPU power of the group */
1953 avg_load = sg_div_cpu_power(group,
1954 avg_load * SCHED_LOAD_SCALE);
1956 if (local_group) {
1957 this_load = avg_load;
1958 this = group;
1959 } else if (avg_load < min_load) {
1960 min_load = avg_load;
1961 idlest = group;
1963 } while (group = group->next, group != sd->groups);
1965 if (!idlest || 100*this_load < imbalance*min_load)
1966 return NULL;
1967 return idlest;
1971 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1973 static int
1974 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
1975 cpumask_t *tmp)
1977 unsigned long load, min_load = ULONG_MAX;
1978 int idlest = -1;
1979 int i;
1981 /* Traverse only the allowed CPUs */
1982 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
1984 for_each_cpu_mask(i, *tmp) {
1985 load = weighted_cpuload(i);
1987 if (load < min_load || (load == min_load && i == this_cpu)) {
1988 min_load = load;
1989 idlest = i;
1993 return idlest;
1997 * sched_balance_self: balance the current task (running on cpu) in domains
1998 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1999 * SD_BALANCE_EXEC.
2001 * Balance, ie. select the least loaded group.
2003 * Returns the target CPU number, or the same CPU if no balancing is needed.
2005 * preempt must be disabled.
2007 static int sched_balance_self(int cpu, int flag)
2009 struct task_struct *t = current;
2010 struct sched_domain *tmp, *sd = NULL;
2012 for_each_domain(cpu, tmp) {
2014 * If power savings logic is enabled for a domain, stop there.
2016 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2017 break;
2018 if (tmp->flags & flag)
2019 sd = tmp;
2022 while (sd) {
2023 cpumask_t span, tmpmask;
2024 struct sched_group *group;
2025 int new_cpu, weight;
2027 if (!(sd->flags & flag)) {
2028 sd = sd->child;
2029 continue;
2032 span = sd->span;
2033 group = find_idlest_group(sd, t, cpu);
2034 if (!group) {
2035 sd = sd->child;
2036 continue;
2039 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2040 if (new_cpu == -1 || new_cpu == cpu) {
2041 /* Now try balancing at a lower domain level of cpu */
2042 sd = sd->child;
2043 continue;
2046 /* Now try balancing at a lower domain level of new_cpu */
2047 cpu = new_cpu;
2048 sd = NULL;
2049 weight = cpus_weight(span);
2050 for_each_domain(cpu, tmp) {
2051 if (weight <= cpus_weight(tmp->span))
2052 break;
2053 if (tmp->flags & flag)
2054 sd = tmp;
2056 /* while loop will break here if sd == NULL */
2059 return cpu;
2062 #endif /* CONFIG_SMP */
2064 /***
2065 * try_to_wake_up - wake up a thread
2066 * @p: the to-be-woken-up thread
2067 * @state: the mask of task states that can be woken
2068 * @sync: do a synchronous wakeup?
2070 * Put it on the run-queue if it's not already there. The "current"
2071 * thread is always on the run-queue (except when the actual
2072 * re-schedule is in progress), and as such you're allowed to do
2073 * the simpler "current->state = TASK_RUNNING" to mark yourself
2074 * runnable without the overhead of this.
2076 * returns failure only if the task is already active.
2078 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2080 int cpu, orig_cpu, this_cpu, success = 0;
2081 unsigned long flags;
2082 long old_state;
2083 struct rq *rq;
2085 if (!sched_feat(SYNC_WAKEUPS))
2086 sync = 0;
2088 smp_wmb();
2089 rq = task_rq_lock(p, &flags);
2090 old_state = p->state;
2091 if (!(old_state & state))
2092 goto out;
2094 if (p->se.on_rq)
2095 goto out_running;
2097 cpu = task_cpu(p);
2098 orig_cpu = cpu;
2099 this_cpu = smp_processor_id();
2101 #ifdef CONFIG_SMP
2102 if (unlikely(task_running(rq, p)))
2103 goto out_activate;
2105 cpu = p->sched_class->select_task_rq(p, sync);
2106 if (cpu != orig_cpu) {
2107 set_task_cpu(p, cpu);
2108 task_rq_unlock(rq, &flags);
2109 /* might preempt at this point */
2110 rq = task_rq_lock(p, &flags);
2111 old_state = p->state;
2112 if (!(old_state & state))
2113 goto out;
2114 if (p->se.on_rq)
2115 goto out_running;
2117 this_cpu = smp_processor_id();
2118 cpu = task_cpu(p);
2121 #ifdef CONFIG_SCHEDSTATS
2122 schedstat_inc(rq, ttwu_count);
2123 if (cpu == this_cpu)
2124 schedstat_inc(rq, ttwu_local);
2125 else {
2126 struct sched_domain *sd;
2127 for_each_domain(this_cpu, sd) {
2128 if (cpu_isset(cpu, sd->span)) {
2129 schedstat_inc(sd, ttwu_wake_remote);
2130 break;
2134 #endif
2136 out_activate:
2137 #endif /* CONFIG_SMP */
2138 schedstat_inc(p, se.nr_wakeups);
2139 if (sync)
2140 schedstat_inc(p, se.nr_wakeups_sync);
2141 if (orig_cpu != cpu)
2142 schedstat_inc(p, se.nr_wakeups_migrate);
2143 if (cpu == this_cpu)
2144 schedstat_inc(p, se.nr_wakeups_local);
2145 else
2146 schedstat_inc(p, se.nr_wakeups_remote);
2147 update_rq_clock(rq);
2148 activate_task(rq, p, 1);
2149 success = 1;
2151 out_running:
2152 check_preempt_curr(rq, p);
2154 p->state = TASK_RUNNING;
2155 #ifdef CONFIG_SMP
2156 if (p->sched_class->task_wake_up)
2157 p->sched_class->task_wake_up(rq, p);
2158 #endif
2159 out:
2160 task_rq_unlock(rq, &flags);
2162 return success;
2165 int wake_up_process(struct task_struct *p)
2167 return try_to_wake_up(p, TASK_ALL, 0);
2169 EXPORT_SYMBOL(wake_up_process);
2171 int wake_up_state(struct task_struct *p, unsigned int state)
2173 return try_to_wake_up(p, state, 0);
2177 * Perform scheduler related setup for a newly forked process p.
2178 * p is forked by current.
2180 * __sched_fork() is basic setup used by init_idle() too:
2182 static void __sched_fork(struct task_struct *p)
2184 p->se.exec_start = 0;
2185 p->se.sum_exec_runtime = 0;
2186 p->se.prev_sum_exec_runtime = 0;
2187 p->se.last_wakeup = 0;
2188 p->se.avg_overlap = 0;
2190 #ifdef CONFIG_SCHEDSTATS
2191 p->se.wait_start = 0;
2192 p->se.sum_sleep_runtime = 0;
2193 p->se.sleep_start = 0;
2194 p->se.block_start = 0;
2195 p->se.sleep_max = 0;
2196 p->se.block_max = 0;
2197 p->se.exec_max = 0;
2198 p->se.slice_max = 0;
2199 p->se.wait_max = 0;
2200 #endif
2202 INIT_LIST_HEAD(&p->rt.run_list);
2203 p->se.on_rq = 0;
2204 INIT_LIST_HEAD(&p->se.group_node);
2206 #ifdef CONFIG_PREEMPT_NOTIFIERS
2207 INIT_HLIST_HEAD(&p->preempt_notifiers);
2208 #endif
2211 * We mark the process as running here, but have not actually
2212 * inserted it onto the runqueue yet. This guarantees that
2213 * nobody will actually run it, and a signal or other external
2214 * event cannot wake it up and insert it on the runqueue either.
2216 p->state = TASK_RUNNING;
2220 * fork()/clone()-time setup:
2222 void sched_fork(struct task_struct *p, int clone_flags)
2224 int cpu = get_cpu();
2226 __sched_fork(p);
2228 #ifdef CONFIG_SMP
2229 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2230 #endif
2231 set_task_cpu(p, cpu);
2234 * Make sure we do not leak PI boosting priority to the child:
2236 p->prio = current->normal_prio;
2237 if (!rt_prio(p->prio))
2238 p->sched_class = &fair_sched_class;
2240 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2241 if (likely(sched_info_on()))
2242 memset(&p->sched_info, 0, sizeof(p->sched_info));
2243 #endif
2244 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2245 p->oncpu = 0;
2246 #endif
2247 #ifdef CONFIG_PREEMPT
2248 /* Want to start with kernel preemption disabled. */
2249 task_thread_info(p)->preempt_count = 1;
2250 #endif
2251 put_cpu();
2255 * wake_up_new_task - wake up a newly created task for the first time.
2257 * This function will do some initial scheduler statistics housekeeping
2258 * that must be done for every newly created context, then puts the task
2259 * on the runqueue and wakes it.
2261 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2263 unsigned long flags;
2264 struct rq *rq;
2266 rq = task_rq_lock(p, &flags);
2267 BUG_ON(p->state != TASK_RUNNING);
2268 update_rq_clock(rq);
2270 p->prio = effective_prio(p);
2272 if (!p->sched_class->task_new || !current->se.on_rq) {
2273 activate_task(rq, p, 0);
2274 } else {
2276 * Let the scheduling class do new task startup
2277 * management (if any):
2279 p->sched_class->task_new(rq, p);
2280 inc_nr_running(p, rq);
2282 check_preempt_curr(rq, p);
2283 #ifdef CONFIG_SMP
2284 if (p->sched_class->task_wake_up)
2285 p->sched_class->task_wake_up(rq, p);
2286 #endif
2287 task_rq_unlock(rq, &flags);
2290 #ifdef CONFIG_PREEMPT_NOTIFIERS
2293 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2294 * @notifier: notifier struct to register
2296 void preempt_notifier_register(struct preempt_notifier *notifier)
2298 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2300 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2303 * preempt_notifier_unregister - no longer interested in preemption notifications
2304 * @notifier: notifier struct to unregister
2306 * This is safe to call from within a preemption notifier.
2308 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2310 hlist_del(&notifier->link);
2312 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2314 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2316 struct preempt_notifier *notifier;
2317 struct hlist_node *node;
2319 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2320 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2323 static void
2324 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2325 struct task_struct *next)
2327 struct preempt_notifier *notifier;
2328 struct hlist_node *node;
2330 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2331 notifier->ops->sched_out(notifier, next);
2334 #else
2336 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2340 static void
2341 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2342 struct task_struct *next)
2346 #endif
2349 * prepare_task_switch - prepare to switch tasks
2350 * @rq: the runqueue preparing to switch
2351 * @prev: the current task that is being switched out
2352 * @next: the task we are going to switch to.
2354 * This is called with the rq lock held and interrupts off. It must
2355 * be paired with a subsequent finish_task_switch after the context
2356 * switch.
2358 * prepare_task_switch sets up locking and calls architecture specific
2359 * hooks.
2361 static inline void
2362 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2363 struct task_struct *next)
2365 fire_sched_out_preempt_notifiers(prev, next);
2366 prepare_lock_switch(rq, next);
2367 prepare_arch_switch(next);
2371 * finish_task_switch - clean up after a task-switch
2372 * @rq: runqueue associated with task-switch
2373 * @prev: the thread we just switched away from.
2375 * finish_task_switch must be called after the context switch, paired
2376 * with a prepare_task_switch call before the context switch.
2377 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2378 * and do any other architecture-specific cleanup actions.
2380 * Note that we may have delayed dropping an mm in context_switch(). If
2381 * so, we finish that here outside of the runqueue lock. (Doing it
2382 * with the lock held can cause deadlocks; see schedule() for
2383 * details.)
2385 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2386 __releases(rq->lock)
2388 struct mm_struct *mm = rq->prev_mm;
2389 long prev_state;
2391 rq->prev_mm = NULL;
2394 * A task struct has one reference for the use as "current".
2395 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2396 * schedule one last time. The schedule call will never return, and
2397 * the scheduled task must drop that reference.
2398 * The test for TASK_DEAD must occur while the runqueue locks are
2399 * still held, otherwise prev could be scheduled on another cpu, die
2400 * there before we look at prev->state, and then the reference would
2401 * be dropped twice.
2402 * Manfred Spraul <manfred@colorfullife.com>
2404 prev_state = prev->state;
2405 finish_arch_switch(prev);
2406 finish_lock_switch(rq, prev);
2407 #ifdef CONFIG_SMP
2408 if (current->sched_class->post_schedule)
2409 current->sched_class->post_schedule(rq);
2410 #endif
2412 fire_sched_in_preempt_notifiers(current);
2413 if (mm)
2414 mmdrop(mm);
2415 if (unlikely(prev_state == TASK_DEAD)) {
2417 * Remove function-return probe instances associated with this
2418 * task and put them back on the free list.
2420 kprobe_flush_task(prev);
2421 put_task_struct(prev);
2426 * schedule_tail - first thing a freshly forked thread must call.
2427 * @prev: the thread we just switched away from.
2429 asmlinkage void schedule_tail(struct task_struct *prev)
2430 __releases(rq->lock)
2432 struct rq *rq = this_rq();
2434 finish_task_switch(rq, prev);
2435 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2436 /* In this case, finish_task_switch does not reenable preemption */
2437 preempt_enable();
2438 #endif
2439 if (current->set_child_tid)
2440 put_user(task_pid_vnr(current), current->set_child_tid);
2444 * context_switch - switch to the new MM and the new
2445 * thread's register state.
2447 static inline void
2448 context_switch(struct rq *rq, struct task_struct *prev,
2449 struct task_struct *next)
2451 struct mm_struct *mm, *oldmm;
2453 prepare_task_switch(rq, prev, next);
2454 mm = next->mm;
2455 oldmm = prev->active_mm;
2457 * For paravirt, this is coupled with an exit in switch_to to
2458 * combine the page table reload and the switch backend into
2459 * one hypercall.
2461 arch_enter_lazy_cpu_mode();
2463 if (unlikely(!mm)) {
2464 next->active_mm = oldmm;
2465 atomic_inc(&oldmm->mm_count);
2466 enter_lazy_tlb(oldmm, next);
2467 } else
2468 switch_mm(oldmm, mm, next);
2470 if (unlikely(!prev->mm)) {
2471 prev->active_mm = NULL;
2472 rq->prev_mm = oldmm;
2475 * Since the runqueue lock will be released by the next
2476 * task (which is an invalid locking op but in the case
2477 * of the scheduler it's an obvious special-case), so we
2478 * do an early lockdep release here:
2480 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2481 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2482 #endif
2484 /* Here we just switch the register state and the stack. */
2485 switch_to(prev, next, prev);
2487 barrier();
2489 * this_rq must be evaluated again because prev may have moved
2490 * CPUs since it called schedule(), thus the 'rq' on its stack
2491 * frame will be invalid.
2493 finish_task_switch(this_rq(), prev);
2497 * nr_running, nr_uninterruptible and nr_context_switches:
2499 * externally visible scheduler statistics: current number of runnable
2500 * threads, current number of uninterruptible-sleeping threads, total
2501 * number of context switches performed since bootup.
2503 unsigned long nr_running(void)
2505 unsigned long i, sum = 0;
2507 for_each_online_cpu(i)
2508 sum += cpu_rq(i)->nr_running;
2510 return sum;
2513 unsigned long nr_uninterruptible(void)
2515 unsigned long i, sum = 0;
2517 for_each_possible_cpu(i)
2518 sum += cpu_rq(i)->nr_uninterruptible;
2521 * Since we read the counters lockless, it might be slightly
2522 * inaccurate. Do not allow it to go below zero though:
2524 if (unlikely((long)sum < 0))
2525 sum = 0;
2527 return sum;
2530 unsigned long long nr_context_switches(void)
2532 int i;
2533 unsigned long long sum = 0;
2535 for_each_possible_cpu(i)
2536 sum += cpu_rq(i)->nr_switches;
2538 return sum;
2541 unsigned long nr_iowait(void)
2543 unsigned long i, sum = 0;
2545 for_each_possible_cpu(i)
2546 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2548 return sum;
2551 unsigned long nr_active(void)
2553 unsigned long i, running = 0, uninterruptible = 0;
2555 for_each_online_cpu(i) {
2556 running += cpu_rq(i)->nr_running;
2557 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2560 if (unlikely((long)uninterruptible < 0))
2561 uninterruptible = 0;
2563 return running + uninterruptible;
2567 * Update rq->cpu_load[] statistics. This function is usually called every
2568 * scheduler tick (TICK_NSEC).
2570 static void update_cpu_load(struct rq *this_rq)
2572 unsigned long this_load = this_rq->load.weight;
2573 int i, scale;
2575 this_rq->nr_load_updates++;
2577 /* Update our load: */
2578 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2579 unsigned long old_load, new_load;
2581 /* scale is effectively 1 << i now, and >> i divides by scale */
2583 old_load = this_rq->cpu_load[i];
2584 new_load = this_load;
2586 * Round up the averaging division if load is increasing. This
2587 * prevents us from getting stuck on 9 if the load is 10, for
2588 * example.
2590 if (new_load > old_load)
2591 new_load += scale-1;
2592 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2596 #ifdef CONFIG_SMP
2599 * double_rq_lock - safely lock two runqueues
2601 * Note this does not disable interrupts like task_rq_lock,
2602 * you need to do so manually before calling.
2604 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2605 __acquires(rq1->lock)
2606 __acquires(rq2->lock)
2608 BUG_ON(!irqs_disabled());
2609 if (rq1 == rq2) {
2610 spin_lock(&rq1->lock);
2611 __acquire(rq2->lock); /* Fake it out ;) */
2612 } else {
2613 if (rq1 < rq2) {
2614 spin_lock(&rq1->lock);
2615 spin_lock(&rq2->lock);
2616 } else {
2617 spin_lock(&rq2->lock);
2618 spin_lock(&rq1->lock);
2621 update_rq_clock(rq1);
2622 update_rq_clock(rq2);
2626 * double_rq_unlock - safely unlock two runqueues
2628 * Note this does not restore interrupts like task_rq_unlock,
2629 * you need to do so manually after calling.
2631 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2632 __releases(rq1->lock)
2633 __releases(rq2->lock)
2635 spin_unlock(&rq1->lock);
2636 if (rq1 != rq2)
2637 spin_unlock(&rq2->lock);
2638 else
2639 __release(rq2->lock);
2643 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2645 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2646 __releases(this_rq->lock)
2647 __acquires(busiest->lock)
2648 __acquires(this_rq->lock)
2650 int ret = 0;
2652 if (unlikely(!irqs_disabled())) {
2653 /* printk() doesn't work good under rq->lock */
2654 spin_unlock(&this_rq->lock);
2655 BUG_ON(1);
2657 if (unlikely(!spin_trylock(&busiest->lock))) {
2658 if (busiest < this_rq) {
2659 spin_unlock(&this_rq->lock);
2660 spin_lock(&busiest->lock);
2661 spin_lock(&this_rq->lock);
2662 ret = 1;
2663 } else
2664 spin_lock(&busiest->lock);
2666 return ret;
2670 * If dest_cpu is allowed for this process, migrate the task to it.
2671 * This is accomplished by forcing the cpu_allowed mask to only
2672 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2673 * the cpu_allowed mask is restored.
2675 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2677 struct migration_req req;
2678 unsigned long flags;
2679 struct rq *rq;
2681 rq = task_rq_lock(p, &flags);
2682 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2683 || unlikely(cpu_is_offline(dest_cpu)))
2684 goto out;
2686 /* force the process onto the specified CPU */
2687 if (migrate_task(p, dest_cpu, &req)) {
2688 /* Need to wait for migration thread (might exit: take ref). */
2689 struct task_struct *mt = rq->migration_thread;
2691 get_task_struct(mt);
2692 task_rq_unlock(rq, &flags);
2693 wake_up_process(mt);
2694 put_task_struct(mt);
2695 wait_for_completion(&req.done);
2697 return;
2699 out:
2700 task_rq_unlock(rq, &flags);
2704 * sched_exec - execve() is a valuable balancing opportunity, because at
2705 * this point the task has the smallest effective memory and cache footprint.
2707 void sched_exec(void)
2709 int new_cpu, this_cpu = get_cpu();
2710 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2711 put_cpu();
2712 if (new_cpu != this_cpu)
2713 sched_migrate_task(current, new_cpu);
2717 * pull_task - move a task from a remote runqueue to the local runqueue.
2718 * Both runqueues must be locked.
2720 static void pull_task(struct rq *src_rq, struct task_struct *p,
2721 struct rq *this_rq, int this_cpu)
2723 deactivate_task(src_rq, p, 0);
2724 set_task_cpu(p, this_cpu);
2725 activate_task(this_rq, p, 0);
2727 * Note that idle threads have a prio of MAX_PRIO, for this test
2728 * to be always true for them.
2730 check_preempt_curr(this_rq, p);
2734 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2736 static
2737 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2738 struct sched_domain *sd, enum cpu_idle_type idle,
2739 int *all_pinned)
2742 * We do not migrate tasks that are:
2743 * 1) running (obviously), or
2744 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2745 * 3) are cache-hot on their current CPU.
2747 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2748 schedstat_inc(p, se.nr_failed_migrations_affine);
2749 return 0;
2751 *all_pinned = 0;
2753 if (task_running(rq, p)) {
2754 schedstat_inc(p, se.nr_failed_migrations_running);
2755 return 0;
2759 * Aggressive migration if:
2760 * 1) task is cache cold, or
2761 * 2) too many balance attempts have failed.
2764 if (!task_hot(p, rq->clock, sd) ||
2765 sd->nr_balance_failed > sd->cache_nice_tries) {
2766 #ifdef CONFIG_SCHEDSTATS
2767 if (task_hot(p, rq->clock, sd)) {
2768 schedstat_inc(sd, lb_hot_gained[idle]);
2769 schedstat_inc(p, se.nr_forced_migrations);
2771 #endif
2772 return 1;
2775 if (task_hot(p, rq->clock, sd)) {
2776 schedstat_inc(p, se.nr_failed_migrations_hot);
2777 return 0;
2779 return 1;
2782 static unsigned long
2783 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2784 unsigned long max_load_move, struct sched_domain *sd,
2785 enum cpu_idle_type idle, int *all_pinned,
2786 int *this_best_prio, struct rq_iterator *iterator)
2788 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2789 struct task_struct *p;
2790 long rem_load_move = max_load_move;
2792 if (max_load_move == 0)
2793 goto out;
2795 pinned = 1;
2798 * Start the load-balancing iterator:
2800 p = iterator->start(iterator->arg);
2801 next:
2802 if (!p || loops++ > sysctl_sched_nr_migrate)
2803 goto out;
2805 * To help distribute high priority tasks across CPUs we don't
2806 * skip a task if it will be the highest priority task (i.e. smallest
2807 * prio value) on its new queue regardless of its load weight
2809 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2810 SCHED_LOAD_SCALE_FUZZ;
2811 if ((skip_for_load && p->prio >= *this_best_prio) ||
2812 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2813 p = iterator->next(iterator->arg);
2814 goto next;
2817 pull_task(busiest, p, this_rq, this_cpu);
2818 pulled++;
2819 rem_load_move -= p->se.load.weight;
2822 * We only want to steal up to the prescribed amount of weighted load.
2824 if (rem_load_move > 0) {
2825 if (p->prio < *this_best_prio)
2826 *this_best_prio = p->prio;
2827 p = iterator->next(iterator->arg);
2828 goto next;
2830 out:
2832 * Right now, this is one of only two places pull_task() is called,
2833 * so we can safely collect pull_task() stats here rather than
2834 * inside pull_task().
2836 schedstat_add(sd, lb_gained[idle], pulled);
2838 if (all_pinned)
2839 *all_pinned = pinned;
2841 return max_load_move - rem_load_move;
2845 * move_tasks tries to move up to max_load_move weighted load from busiest to
2846 * this_rq, as part of a balancing operation within domain "sd".
2847 * Returns 1 if successful and 0 otherwise.
2849 * Called with both runqueues locked.
2851 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2852 unsigned long max_load_move,
2853 struct sched_domain *sd, enum cpu_idle_type idle,
2854 int *all_pinned)
2856 const struct sched_class *class = sched_class_highest;
2857 unsigned long total_load_moved = 0;
2858 int this_best_prio = this_rq->curr->prio;
2860 do {
2861 total_load_moved +=
2862 class->load_balance(this_rq, this_cpu, busiest,
2863 max_load_move - total_load_moved,
2864 sd, idle, all_pinned, &this_best_prio);
2865 class = class->next;
2866 } while (class && max_load_move > total_load_moved);
2868 return total_load_moved > 0;
2871 static int
2872 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2873 struct sched_domain *sd, enum cpu_idle_type idle,
2874 struct rq_iterator *iterator)
2876 struct task_struct *p = iterator->start(iterator->arg);
2877 int pinned = 0;
2879 while (p) {
2880 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2881 pull_task(busiest, p, this_rq, this_cpu);
2883 * Right now, this is only the second place pull_task()
2884 * is called, so we can safely collect pull_task()
2885 * stats here rather than inside pull_task().
2887 schedstat_inc(sd, lb_gained[idle]);
2889 return 1;
2891 p = iterator->next(iterator->arg);
2894 return 0;
2898 * move_one_task tries to move exactly one task from busiest to this_rq, as
2899 * part of active balancing operations within "domain".
2900 * Returns 1 if successful and 0 otherwise.
2902 * Called with both runqueues locked.
2904 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2905 struct sched_domain *sd, enum cpu_idle_type idle)
2907 const struct sched_class *class;
2909 for (class = sched_class_highest; class; class = class->next)
2910 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2911 return 1;
2913 return 0;
2917 * find_busiest_group finds and returns the busiest CPU group within the
2918 * domain. It calculates and returns the amount of weighted load which
2919 * should be moved to restore balance via the imbalance parameter.
2921 static struct sched_group *
2922 find_busiest_group(struct sched_domain *sd, int this_cpu,
2923 unsigned long *imbalance, enum cpu_idle_type idle,
2924 int *sd_idle, const cpumask_t *cpus, int *balance)
2926 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2927 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2928 unsigned long max_pull;
2929 unsigned long busiest_load_per_task, busiest_nr_running;
2930 unsigned long this_load_per_task, this_nr_running;
2931 int load_idx, group_imb = 0;
2932 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2933 int power_savings_balance = 1;
2934 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2935 unsigned long min_nr_running = ULONG_MAX;
2936 struct sched_group *group_min = NULL, *group_leader = NULL;
2937 #endif
2939 max_load = this_load = total_load = total_pwr = 0;
2940 busiest_load_per_task = busiest_nr_running = 0;
2941 this_load_per_task = this_nr_running = 0;
2942 if (idle == CPU_NOT_IDLE)
2943 load_idx = sd->busy_idx;
2944 else if (idle == CPU_NEWLY_IDLE)
2945 load_idx = sd->newidle_idx;
2946 else
2947 load_idx = sd->idle_idx;
2949 do {
2950 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2951 int local_group;
2952 int i;
2953 int __group_imb = 0;
2954 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2955 unsigned long sum_nr_running, sum_weighted_load;
2957 local_group = cpu_isset(this_cpu, group->cpumask);
2959 if (local_group)
2960 balance_cpu = first_cpu(group->cpumask);
2962 /* Tally up the load of all CPUs in the group */
2963 sum_weighted_load = sum_nr_running = avg_load = 0;
2964 max_cpu_load = 0;
2965 min_cpu_load = ~0UL;
2967 for_each_cpu_mask(i, group->cpumask) {
2968 struct rq *rq;
2970 if (!cpu_isset(i, *cpus))
2971 continue;
2973 rq = cpu_rq(i);
2975 if (*sd_idle && rq->nr_running)
2976 *sd_idle = 0;
2978 /* Bias balancing toward cpus of our domain */
2979 if (local_group) {
2980 if (idle_cpu(i) && !first_idle_cpu) {
2981 first_idle_cpu = 1;
2982 balance_cpu = i;
2985 load = target_load(i, load_idx);
2986 } else {
2987 load = source_load(i, load_idx);
2988 if (load > max_cpu_load)
2989 max_cpu_load = load;
2990 if (min_cpu_load > load)
2991 min_cpu_load = load;
2994 avg_load += load;
2995 sum_nr_running += rq->nr_running;
2996 sum_weighted_load += weighted_cpuload(i);
3000 * First idle cpu or the first cpu(busiest) in this sched group
3001 * is eligible for doing load balancing at this and above
3002 * domains. In the newly idle case, we will allow all the cpu's
3003 * to do the newly idle load balance.
3005 if (idle != CPU_NEWLY_IDLE && local_group &&
3006 balance_cpu != this_cpu && balance) {
3007 *balance = 0;
3008 goto ret;
3011 total_load += avg_load;
3012 total_pwr += group->__cpu_power;
3014 /* Adjust by relative CPU power of the group */
3015 avg_load = sg_div_cpu_power(group,
3016 avg_load * SCHED_LOAD_SCALE);
3018 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3019 __group_imb = 1;
3021 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3023 if (local_group) {
3024 this_load = avg_load;
3025 this = group;
3026 this_nr_running = sum_nr_running;
3027 this_load_per_task = sum_weighted_load;
3028 } else if (avg_load > max_load &&
3029 (sum_nr_running > group_capacity || __group_imb)) {
3030 max_load = avg_load;
3031 busiest = group;
3032 busiest_nr_running = sum_nr_running;
3033 busiest_load_per_task = sum_weighted_load;
3034 group_imb = __group_imb;
3037 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3039 * Busy processors will not participate in power savings
3040 * balance.
3042 if (idle == CPU_NOT_IDLE ||
3043 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3044 goto group_next;
3047 * If the local group is idle or completely loaded
3048 * no need to do power savings balance at this domain
3050 if (local_group && (this_nr_running >= group_capacity ||
3051 !this_nr_running))
3052 power_savings_balance = 0;
3055 * If a group is already running at full capacity or idle,
3056 * don't include that group in power savings calculations
3058 if (!power_savings_balance || sum_nr_running >= group_capacity
3059 || !sum_nr_running)
3060 goto group_next;
3063 * Calculate the group which has the least non-idle load.
3064 * This is the group from where we need to pick up the load
3065 * for saving power
3067 if ((sum_nr_running < min_nr_running) ||
3068 (sum_nr_running == min_nr_running &&
3069 first_cpu(group->cpumask) <
3070 first_cpu(group_min->cpumask))) {
3071 group_min = group;
3072 min_nr_running = sum_nr_running;
3073 min_load_per_task = sum_weighted_load /
3074 sum_nr_running;
3078 * Calculate the group which is almost near its
3079 * capacity but still has some space to pick up some load
3080 * from other group and save more power
3082 if (sum_nr_running <= group_capacity - 1) {
3083 if (sum_nr_running > leader_nr_running ||
3084 (sum_nr_running == leader_nr_running &&
3085 first_cpu(group->cpumask) >
3086 first_cpu(group_leader->cpumask))) {
3087 group_leader = group;
3088 leader_nr_running = sum_nr_running;
3091 group_next:
3092 #endif
3093 group = group->next;
3094 } while (group != sd->groups);
3096 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3097 goto out_balanced;
3099 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3101 if (this_load >= avg_load ||
3102 100*max_load <= sd->imbalance_pct*this_load)
3103 goto out_balanced;
3105 busiest_load_per_task /= busiest_nr_running;
3106 if (group_imb)
3107 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3110 * We're trying to get all the cpus to the average_load, so we don't
3111 * want to push ourselves above the average load, nor do we wish to
3112 * reduce the max loaded cpu below the average load, as either of these
3113 * actions would just result in more rebalancing later, and ping-pong
3114 * tasks around. Thus we look for the minimum possible imbalance.
3115 * Negative imbalances (*we* are more loaded than anyone else) will
3116 * be counted as no imbalance for these purposes -- we can't fix that
3117 * by pulling tasks to us. Be careful of negative numbers as they'll
3118 * appear as very large values with unsigned longs.
3120 if (max_load <= busiest_load_per_task)
3121 goto out_balanced;
3124 * In the presence of smp nice balancing, certain scenarios can have
3125 * max load less than avg load(as we skip the groups at or below
3126 * its cpu_power, while calculating max_load..)
3128 if (max_load < avg_load) {
3129 *imbalance = 0;
3130 goto small_imbalance;
3133 /* Don't want to pull so many tasks that a group would go idle */
3134 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3136 /* How much load to actually move to equalise the imbalance */
3137 *imbalance = min(max_pull * busiest->__cpu_power,
3138 (avg_load - this_load) * this->__cpu_power)
3139 / SCHED_LOAD_SCALE;
3142 * if *imbalance is less than the average load per runnable task
3143 * there is no gaurantee that any tasks will be moved so we'll have
3144 * a think about bumping its value to force at least one task to be
3145 * moved
3147 if (*imbalance < busiest_load_per_task) {
3148 unsigned long tmp, pwr_now, pwr_move;
3149 unsigned int imbn;
3151 small_imbalance:
3152 pwr_move = pwr_now = 0;
3153 imbn = 2;
3154 if (this_nr_running) {
3155 this_load_per_task /= this_nr_running;
3156 if (busiest_load_per_task > this_load_per_task)
3157 imbn = 1;
3158 } else
3159 this_load_per_task = SCHED_LOAD_SCALE;
3161 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3162 busiest_load_per_task * imbn) {
3163 *imbalance = busiest_load_per_task;
3164 return busiest;
3168 * OK, we don't have enough imbalance to justify moving tasks,
3169 * however we may be able to increase total CPU power used by
3170 * moving them.
3173 pwr_now += busiest->__cpu_power *
3174 min(busiest_load_per_task, max_load);
3175 pwr_now += this->__cpu_power *
3176 min(this_load_per_task, this_load);
3177 pwr_now /= SCHED_LOAD_SCALE;
3179 /* Amount of load we'd subtract */
3180 tmp = sg_div_cpu_power(busiest,
3181 busiest_load_per_task * SCHED_LOAD_SCALE);
3182 if (max_load > tmp)
3183 pwr_move += busiest->__cpu_power *
3184 min(busiest_load_per_task, max_load - tmp);
3186 /* Amount of load we'd add */
3187 if (max_load * busiest->__cpu_power <
3188 busiest_load_per_task * SCHED_LOAD_SCALE)
3189 tmp = sg_div_cpu_power(this,
3190 max_load * busiest->__cpu_power);
3191 else
3192 tmp = sg_div_cpu_power(this,
3193 busiest_load_per_task * SCHED_LOAD_SCALE);
3194 pwr_move += this->__cpu_power *
3195 min(this_load_per_task, this_load + tmp);
3196 pwr_move /= SCHED_LOAD_SCALE;
3198 /* Move if we gain throughput */
3199 if (pwr_move > pwr_now)
3200 *imbalance = busiest_load_per_task;
3203 return busiest;
3205 out_balanced:
3206 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3207 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3208 goto ret;
3210 if (this == group_leader && group_leader != group_min) {
3211 *imbalance = min_load_per_task;
3212 return group_min;
3214 #endif
3215 ret:
3216 *imbalance = 0;
3217 return NULL;
3221 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3223 static struct rq *
3224 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3225 unsigned long imbalance, const cpumask_t *cpus)
3227 struct rq *busiest = NULL, *rq;
3228 unsigned long max_load = 0;
3229 int i;
3231 for_each_cpu_mask(i, group->cpumask) {
3232 unsigned long wl;
3234 if (!cpu_isset(i, *cpus))
3235 continue;
3237 rq = cpu_rq(i);
3238 wl = weighted_cpuload(i);
3240 if (rq->nr_running == 1 && wl > imbalance)
3241 continue;
3243 if (wl > max_load) {
3244 max_load = wl;
3245 busiest = rq;
3249 return busiest;
3253 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3254 * so long as it is large enough.
3256 #define MAX_PINNED_INTERVAL 512
3259 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3260 * tasks if there is an imbalance.
3262 static int load_balance(int this_cpu, struct rq *this_rq,
3263 struct sched_domain *sd, enum cpu_idle_type idle,
3264 int *balance, cpumask_t *cpus)
3266 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3267 struct sched_group *group;
3268 unsigned long imbalance;
3269 struct rq *busiest;
3270 unsigned long flags;
3272 cpus_setall(*cpus);
3275 * When power savings policy is enabled for the parent domain, idle
3276 * sibling can pick up load irrespective of busy siblings. In this case,
3277 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3278 * portraying it as CPU_NOT_IDLE.
3280 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3281 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3282 sd_idle = 1;
3284 schedstat_inc(sd, lb_count[idle]);
3286 redo:
3287 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3288 cpus, balance);
3290 if (*balance == 0)
3291 goto out_balanced;
3293 if (!group) {
3294 schedstat_inc(sd, lb_nobusyg[idle]);
3295 goto out_balanced;
3298 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3299 if (!busiest) {
3300 schedstat_inc(sd, lb_nobusyq[idle]);
3301 goto out_balanced;
3304 BUG_ON(busiest == this_rq);
3306 schedstat_add(sd, lb_imbalance[idle], imbalance);
3308 ld_moved = 0;
3309 if (busiest->nr_running > 1) {
3311 * Attempt to move tasks. If find_busiest_group has found
3312 * an imbalance but busiest->nr_running <= 1, the group is
3313 * still unbalanced. ld_moved simply stays zero, so it is
3314 * correctly treated as an imbalance.
3316 local_irq_save(flags);
3317 double_rq_lock(this_rq, busiest);
3318 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3319 imbalance, sd, idle, &all_pinned);
3320 double_rq_unlock(this_rq, busiest);
3321 local_irq_restore(flags);
3324 * some other cpu did the load balance for us.
3326 if (ld_moved && this_cpu != smp_processor_id())
3327 resched_cpu(this_cpu);
3329 /* All tasks on this runqueue were pinned by CPU affinity */
3330 if (unlikely(all_pinned)) {
3331 cpu_clear(cpu_of(busiest), *cpus);
3332 if (!cpus_empty(*cpus))
3333 goto redo;
3334 goto out_balanced;
3338 if (!ld_moved) {
3339 schedstat_inc(sd, lb_failed[idle]);
3340 sd->nr_balance_failed++;
3342 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3344 spin_lock_irqsave(&busiest->lock, flags);
3346 /* don't kick the migration_thread, if the curr
3347 * task on busiest cpu can't be moved to this_cpu
3349 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3350 spin_unlock_irqrestore(&busiest->lock, flags);
3351 all_pinned = 1;
3352 goto out_one_pinned;
3355 if (!busiest->active_balance) {
3356 busiest->active_balance = 1;
3357 busiest->push_cpu = this_cpu;
3358 active_balance = 1;
3360 spin_unlock_irqrestore(&busiest->lock, flags);
3361 if (active_balance)
3362 wake_up_process(busiest->migration_thread);
3365 * We've kicked active balancing, reset the failure
3366 * counter.
3368 sd->nr_balance_failed = sd->cache_nice_tries+1;
3370 } else
3371 sd->nr_balance_failed = 0;
3373 if (likely(!active_balance)) {
3374 /* We were unbalanced, so reset the balancing interval */
3375 sd->balance_interval = sd->min_interval;
3376 } else {
3378 * If we've begun active balancing, start to back off. This
3379 * case may not be covered by the all_pinned logic if there
3380 * is only 1 task on the busy runqueue (because we don't call
3381 * move_tasks).
3383 if (sd->balance_interval < sd->max_interval)
3384 sd->balance_interval *= 2;
3387 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3388 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3389 return -1;
3390 return ld_moved;
3392 out_balanced:
3393 schedstat_inc(sd, lb_balanced[idle]);
3395 sd->nr_balance_failed = 0;
3397 out_one_pinned:
3398 /* tune up the balancing interval */
3399 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3400 (sd->balance_interval < sd->max_interval))
3401 sd->balance_interval *= 2;
3403 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3404 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3405 return -1;
3406 return 0;
3410 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3411 * tasks if there is an imbalance.
3413 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3414 * this_rq is locked.
3416 static int
3417 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3418 cpumask_t *cpus)
3420 struct sched_group *group;
3421 struct rq *busiest = NULL;
3422 unsigned long imbalance;
3423 int ld_moved = 0;
3424 int sd_idle = 0;
3425 int all_pinned = 0;
3427 cpus_setall(*cpus);
3430 * When power savings policy is enabled for the parent domain, idle
3431 * sibling can pick up load irrespective of busy siblings. In this case,
3432 * let the state of idle sibling percolate up as IDLE, instead of
3433 * portraying it as CPU_NOT_IDLE.
3435 if (sd->flags & SD_SHARE_CPUPOWER &&
3436 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3437 sd_idle = 1;
3439 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3440 redo:
3441 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3442 &sd_idle, cpus, NULL);
3443 if (!group) {
3444 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3445 goto out_balanced;
3448 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3449 if (!busiest) {
3450 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3451 goto out_balanced;
3454 BUG_ON(busiest == this_rq);
3456 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3458 ld_moved = 0;
3459 if (busiest->nr_running > 1) {
3460 /* Attempt to move tasks */
3461 double_lock_balance(this_rq, busiest);
3462 /* this_rq->clock is already updated */
3463 update_rq_clock(busiest);
3464 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3465 imbalance, sd, CPU_NEWLY_IDLE,
3466 &all_pinned);
3467 spin_unlock(&busiest->lock);
3469 if (unlikely(all_pinned)) {
3470 cpu_clear(cpu_of(busiest), *cpus);
3471 if (!cpus_empty(*cpus))
3472 goto redo;
3476 if (!ld_moved) {
3477 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3478 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3479 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3480 return -1;
3481 } else
3482 sd->nr_balance_failed = 0;
3484 return ld_moved;
3486 out_balanced:
3487 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3488 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3489 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3490 return -1;
3491 sd->nr_balance_failed = 0;
3493 return 0;
3497 * idle_balance is called by schedule() if this_cpu is about to become
3498 * idle. Attempts to pull tasks from other CPUs.
3500 static void idle_balance(int this_cpu, struct rq *this_rq)
3502 struct sched_domain *sd;
3503 int pulled_task = -1;
3504 unsigned long next_balance = jiffies + HZ;
3505 cpumask_t tmpmask;
3507 for_each_domain(this_cpu, sd) {
3508 unsigned long interval;
3510 if (!(sd->flags & SD_LOAD_BALANCE))
3511 continue;
3513 if (sd->flags & SD_BALANCE_NEWIDLE)
3514 /* If we've pulled tasks over stop searching: */
3515 pulled_task = load_balance_newidle(this_cpu, this_rq,
3516 sd, &tmpmask);
3518 interval = msecs_to_jiffies(sd->balance_interval);
3519 if (time_after(next_balance, sd->last_balance + interval))
3520 next_balance = sd->last_balance + interval;
3521 if (pulled_task)
3522 break;
3524 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3526 * We are going idle. next_balance may be set based on
3527 * a busy processor. So reset next_balance.
3529 this_rq->next_balance = next_balance;
3534 * active_load_balance is run by migration threads. It pushes running tasks
3535 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3536 * running on each physical CPU where possible, and avoids physical /
3537 * logical imbalances.
3539 * Called with busiest_rq locked.
3541 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3543 int target_cpu = busiest_rq->push_cpu;
3544 struct sched_domain *sd;
3545 struct rq *target_rq;
3547 /* Is there any task to move? */
3548 if (busiest_rq->nr_running <= 1)
3549 return;
3551 target_rq = cpu_rq(target_cpu);
3554 * This condition is "impossible", if it occurs
3555 * we need to fix it. Originally reported by
3556 * Bjorn Helgaas on a 128-cpu setup.
3558 BUG_ON(busiest_rq == target_rq);
3560 /* move a task from busiest_rq to target_rq */
3561 double_lock_balance(busiest_rq, target_rq);
3562 update_rq_clock(busiest_rq);
3563 update_rq_clock(target_rq);
3565 /* Search for an sd spanning us and the target CPU. */
3566 for_each_domain(target_cpu, sd) {
3567 if ((sd->flags & SD_LOAD_BALANCE) &&
3568 cpu_isset(busiest_cpu, sd->span))
3569 break;
3572 if (likely(sd)) {
3573 schedstat_inc(sd, alb_count);
3575 if (move_one_task(target_rq, target_cpu, busiest_rq,
3576 sd, CPU_IDLE))
3577 schedstat_inc(sd, alb_pushed);
3578 else
3579 schedstat_inc(sd, alb_failed);
3581 spin_unlock(&target_rq->lock);
3584 #ifdef CONFIG_NO_HZ
3585 static struct {
3586 atomic_t load_balancer;
3587 cpumask_t cpu_mask;
3588 } nohz ____cacheline_aligned = {
3589 .load_balancer = ATOMIC_INIT(-1),
3590 .cpu_mask = CPU_MASK_NONE,
3594 * This routine will try to nominate the ilb (idle load balancing)
3595 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3596 * load balancing on behalf of all those cpus. If all the cpus in the system
3597 * go into this tickless mode, then there will be no ilb owner (as there is
3598 * no need for one) and all the cpus will sleep till the next wakeup event
3599 * arrives...
3601 * For the ilb owner, tick is not stopped. And this tick will be used
3602 * for idle load balancing. ilb owner will still be part of
3603 * nohz.cpu_mask..
3605 * While stopping the tick, this cpu will become the ilb owner if there
3606 * is no other owner. And will be the owner till that cpu becomes busy
3607 * or if all cpus in the system stop their ticks at which point
3608 * there is no need for ilb owner.
3610 * When the ilb owner becomes busy, it nominates another owner, during the
3611 * next busy scheduler_tick()
3613 int select_nohz_load_balancer(int stop_tick)
3615 int cpu = smp_processor_id();
3617 if (stop_tick) {
3618 cpu_set(cpu, nohz.cpu_mask);
3619 cpu_rq(cpu)->in_nohz_recently = 1;
3622 * If we are going offline and still the leader, give up!
3624 if (cpu_is_offline(cpu) &&
3625 atomic_read(&nohz.load_balancer) == cpu) {
3626 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3627 BUG();
3628 return 0;
3631 /* time for ilb owner also to sleep */
3632 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3633 if (atomic_read(&nohz.load_balancer) == cpu)
3634 atomic_set(&nohz.load_balancer, -1);
3635 return 0;
3638 if (atomic_read(&nohz.load_balancer) == -1) {
3639 /* make me the ilb owner */
3640 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3641 return 1;
3642 } else if (atomic_read(&nohz.load_balancer) == cpu)
3643 return 1;
3644 } else {
3645 if (!cpu_isset(cpu, nohz.cpu_mask))
3646 return 0;
3648 cpu_clear(cpu, nohz.cpu_mask);
3650 if (atomic_read(&nohz.load_balancer) == cpu)
3651 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3652 BUG();
3654 return 0;
3656 #endif
3658 static DEFINE_SPINLOCK(balancing);
3661 * It checks each scheduling domain to see if it is due to be balanced,
3662 * and initiates a balancing operation if so.
3664 * Balancing parameters are set up in arch_init_sched_domains.
3666 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3668 int balance = 1;
3669 struct rq *rq = cpu_rq(cpu);
3670 unsigned long interval;
3671 struct sched_domain *sd;
3672 /* Earliest time when we have to do rebalance again */
3673 unsigned long next_balance = jiffies + 60*HZ;
3674 int update_next_balance = 0;
3675 cpumask_t tmp;
3677 for_each_domain(cpu, sd) {
3678 if (!(sd->flags & SD_LOAD_BALANCE))
3679 continue;
3681 interval = sd->balance_interval;
3682 if (idle != CPU_IDLE)
3683 interval *= sd->busy_factor;
3685 /* scale ms to jiffies */
3686 interval = msecs_to_jiffies(interval);
3687 if (unlikely(!interval))
3688 interval = 1;
3689 if (interval > HZ*NR_CPUS/10)
3690 interval = HZ*NR_CPUS/10;
3693 if (sd->flags & SD_SERIALIZE) {
3694 if (!spin_trylock(&balancing))
3695 goto out;
3698 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3699 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3701 * We've pulled tasks over so either we're no
3702 * longer idle, or one of our SMT siblings is
3703 * not idle.
3705 idle = CPU_NOT_IDLE;
3707 sd->last_balance = jiffies;
3709 if (sd->flags & SD_SERIALIZE)
3710 spin_unlock(&balancing);
3711 out:
3712 if (time_after(next_balance, sd->last_balance + interval)) {
3713 next_balance = sd->last_balance + interval;
3714 update_next_balance = 1;
3718 * Stop the load balance at this level. There is another
3719 * CPU in our sched group which is doing load balancing more
3720 * actively.
3722 if (!balance)
3723 break;
3727 * next_balance will be updated only when there is a need.
3728 * When the cpu is attached to null domain for ex, it will not be
3729 * updated.
3731 if (likely(update_next_balance))
3732 rq->next_balance = next_balance;
3736 * run_rebalance_domains is triggered when needed from the scheduler tick.
3737 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3738 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3740 static void run_rebalance_domains(struct softirq_action *h)
3742 int this_cpu = smp_processor_id();
3743 struct rq *this_rq = cpu_rq(this_cpu);
3744 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3745 CPU_IDLE : CPU_NOT_IDLE;
3747 rebalance_domains(this_cpu, idle);
3749 #ifdef CONFIG_NO_HZ
3751 * If this cpu is the owner for idle load balancing, then do the
3752 * balancing on behalf of the other idle cpus whose ticks are
3753 * stopped.
3755 if (this_rq->idle_at_tick &&
3756 atomic_read(&nohz.load_balancer) == this_cpu) {
3757 cpumask_t cpus = nohz.cpu_mask;
3758 struct rq *rq;
3759 int balance_cpu;
3761 cpu_clear(this_cpu, cpus);
3762 for_each_cpu_mask(balance_cpu, cpus) {
3764 * If this cpu gets work to do, stop the load balancing
3765 * work being done for other cpus. Next load
3766 * balancing owner will pick it up.
3768 if (need_resched())
3769 break;
3771 rebalance_domains(balance_cpu, CPU_IDLE);
3773 rq = cpu_rq(balance_cpu);
3774 if (time_after(this_rq->next_balance, rq->next_balance))
3775 this_rq->next_balance = rq->next_balance;
3778 #endif
3782 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3784 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3785 * idle load balancing owner or decide to stop the periodic load balancing,
3786 * if the whole system is idle.
3788 static inline void trigger_load_balance(struct rq *rq, int cpu)
3790 #ifdef CONFIG_NO_HZ
3792 * If we were in the nohz mode recently and busy at the current
3793 * scheduler tick, then check if we need to nominate new idle
3794 * load balancer.
3796 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3797 rq->in_nohz_recently = 0;
3799 if (atomic_read(&nohz.load_balancer) == cpu) {
3800 cpu_clear(cpu, nohz.cpu_mask);
3801 atomic_set(&nohz.load_balancer, -1);
3804 if (atomic_read(&nohz.load_balancer) == -1) {
3806 * simple selection for now: Nominate the
3807 * first cpu in the nohz list to be the next
3808 * ilb owner.
3810 * TBD: Traverse the sched domains and nominate
3811 * the nearest cpu in the nohz.cpu_mask.
3813 int ilb = first_cpu(nohz.cpu_mask);
3815 if (ilb < nr_cpu_ids)
3816 resched_cpu(ilb);
3821 * If this cpu is idle and doing idle load balancing for all the
3822 * cpus with ticks stopped, is it time for that to stop?
3824 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3825 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3826 resched_cpu(cpu);
3827 return;
3831 * If this cpu is idle and the idle load balancing is done by
3832 * someone else, then no need raise the SCHED_SOFTIRQ
3834 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3835 cpu_isset(cpu, nohz.cpu_mask))
3836 return;
3837 #endif
3838 if (time_after_eq(jiffies, rq->next_balance))
3839 raise_softirq(SCHED_SOFTIRQ);
3842 #else /* CONFIG_SMP */
3845 * on UP we do not need to balance between CPUs:
3847 static inline void idle_balance(int cpu, struct rq *rq)
3851 #endif
3853 DEFINE_PER_CPU(struct kernel_stat, kstat);
3855 EXPORT_PER_CPU_SYMBOL(kstat);
3858 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3859 * that have not yet been banked in case the task is currently running.
3861 unsigned long long task_sched_runtime(struct task_struct *p)
3863 unsigned long flags;
3864 u64 ns, delta_exec;
3865 struct rq *rq;
3867 rq = task_rq_lock(p, &flags);
3868 ns = p->se.sum_exec_runtime;
3869 if (task_current(rq, p)) {
3870 update_rq_clock(rq);
3871 delta_exec = rq->clock - p->se.exec_start;
3872 if ((s64)delta_exec > 0)
3873 ns += delta_exec;
3875 task_rq_unlock(rq, &flags);
3877 return ns;
3881 * Account user cpu time to a process.
3882 * @p: the process that the cpu time gets accounted to
3883 * @cputime: the cpu time spent in user space since the last update
3885 void account_user_time(struct task_struct *p, cputime_t cputime)
3887 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3888 cputime64_t tmp;
3890 p->utime = cputime_add(p->utime, cputime);
3892 /* Add user time to cpustat. */
3893 tmp = cputime_to_cputime64(cputime);
3894 if (TASK_NICE(p) > 0)
3895 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3896 else
3897 cpustat->user = cputime64_add(cpustat->user, tmp);
3901 * Account guest cpu time to a process.
3902 * @p: the process that the cpu time gets accounted to
3903 * @cputime: the cpu time spent in virtual machine since the last update
3905 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3907 cputime64_t tmp;
3908 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3910 tmp = cputime_to_cputime64(cputime);
3912 p->utime = cputime_add(p->utime, cputime);
3913 p->gtime = cputime_add(p->gtime, cputime);
3915 cpustat->user = cputime64_add(cpustat->user, tmp);
3916 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3920 * Account scaled user cpu time to a process.
3921 * @p: the process that the cpu time gets accounted to
3922 * @cputime: the cpu time spent in user space since the last update
3924 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3926 p->utimescaled = cputime_add(p->utimescaled, cputime);
3930 * Account system cpu time to a process.
3931 * @p: the process that the cpu time gets accounted to
3932 * @hardirq_offset: the offset to subtract from hardirq_count()
3933 * @cputime: the cpu time spent in kernel space since the last update
3935 void account_system_time(struct task_struct *p, int hardirq_offset,
3936 cputime_t cputime)
3938 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3939 struct rq *rq = this_rq();
3940 cputime64_t tmp;
3942 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3943 account_guest_time(p, cputime);
3944 return;
3947 p->stime = cputime_add(p->stime, cputime);
3949 /* Add system time to cpustat. */
3950 tmp = cputime_to_cputime64(cputime);
3951 if (hardirq_count() - hardirq_offset)
3952 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3953 else if (softirq_count())
3954 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3955 else if (p != rq->idle)
3956 cpustat->system = cputime64_add(cpustat->system, tmp);
3957 else if (atomic_read(&rq->nr_iowait) > 0)
3958 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3959 else
3960 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3961 /* Account for system time used */
3962 acct_update_integrals(p);
3966 * Account scaled system cpu time to a process.
3967 * @p: the process that the cpu time gets accounted to
3968 * @hardirq_offset: the offset to subtract from hardirq_count()
3969 * @cputime: the cpu time spent in kernel space since the last update
3971 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3973 p->stimescaled = cputime_add(p->stimescaled, cputime);
3977 * Account for involuntary wait time.
3978 * @p: the process from which the cpu time has been stolen
3979 * @steal: the cpu time spent in involuntary wait
3981 void account_steal_time(struct task_struct *p, cputime_t steal)
3983 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3984 cputime64_t tmp = cputime_to_cputime64(steal);
3985 struct rq *rq = this_rq();
3987 if (p == rq->idle) {
3988 p->stime = cputime_add(p->stime, steal);
3989 if (atomic_read(&rq->nr_iowait) > 0)
3990 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3991 else
3992 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3993 } else
3994 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3998 * This function gets called by the timer code, with HZ frequency.
3999 * We call it with interrupts disabled.
4001 * It also gets called by the fork code, when changing the parent's
4002 * timeslices.
4004 void scheduler_tick(void)
4006 int cpu = smp_processor_id();
4007 struct rq *rq = cpu_rq(cpu);
4008 struct task_struct *curr = rq->curr;
4010 sched_clock_tick();
4012 spin_lock(&rq->lock);
4013 update_rq_clock(rq);
4014 update_cpu_load(rq);
4015 curr->sched_class->task_tick(rq, curr, 0);
4016 spin_unlock(&rq->lock);
4018 #ifdef CONFIG_SMP
4019 rq->idle_at_tick = idle_cpu(cpu);
4020 trigger_load_balance(rq, cpu);
4021 #endif
4024 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4026 void __kprobes add_preempt_count(int val)
4029 * Underflow?
4031 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4032 return;
4033 preempt_count() += val;
4035 * Spinlock count overflowing soon?
4037 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4038 PREEMPT_MASK - 10);
4040 EXPORT_SYMBOL(add_preempt_count);
4042 void __kprobes sub_preempt_count(int val)
4045 * Underflow?
4047 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4048 return;
4050 * Is the spinlock portion underflowing?
4052 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4053 !(preempt_count() & PREEMPT_MASK)))
4054 return;
4056 preempt_count() -= val;
4058 EXPORT_SYMBOL(sub_preempt_count);
4060 #endif
4063 * Print scheduling while atomic bug:
4065 static noinline void __schedule_bug(struct task_struct *prev)
4067 struct pt_regs *regs = get_irq_regs();
4069 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4070 prev->comm, prev->pid, preempt_count());
4072 debug_show_held_locks(prev);
4073 if (irqs_disabled())
4074 print_irqtrace_events(prev);
4076 if (regs)
4077 show_regs(regs);
4078 else
4079 dump_stack();
4083 * Various schedule()-time debugging checks and statistics:
4085 static inline void schedule_debug(struct task_struct *prev)
4088 * Test if we are atomic. Since do_exit() needs to call into
4089 * schedule() atomically, we ignore that path for now.
4090 * Otherwise, whine if we are scheduling when we should not be.
4092 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4093 __schedule_bug(prev);
4095 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4097 schedstat_inc(this_rq(), sched_count);
4098 #ifdef CONFIG_SCHEDSTATS
4099 if (unlikely(prev->lock_depth >= 0)) {
4100 schedstat_inc(this_rq(), bkl_count);
4101 schedstat_inc(prev, sched_info.bkl_count);
4103 #endif
4107 * Pick up the highest-prio task:
4109 static inline struct task_struct *
4110 pick_next_task(struct rq *rq, struct task_struct *prev)
4112 const struct sched_class *class;
4113 struct task_struct *p;
4116 * Optimization: we know that if all tasks are in
4117 * the fair class we can call that function directly:
4119 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4120 p = fair_sched_class.pick_next_task(rq);
4121 if (likely(p))
4122 return p;
4125 class = sched_class_highest;
4126 for ( ; ; ) {
4127 p = class->pick_next_task(rq);
4128 if (p)
4129 return p;
4131 * Will never be NULL as the idle class always
4132 * returns a non-NULL p:
4134 class = class->next;
4139 * schedule() is the main scheduler function.
4141 asmlinkage void __sched schedule(void)
4143 struct task_struct *prev, *next;
4144 unsigned long *switch_count;
4145 struct rq *rq;
4146 int cpu;
4148 need_resched:
4149 preempt_disable();
4150 cpu = smp_processor_id();
4151 rq = cpu_rq(cpu);
4152 rcu_qsctr_inc(cpu);
4153 prev = rq->curr;
4154 switch_count = &prev->nivcsw;
4156 release_kernel_lock(prev);
4157 need_resched_nonpreemptible:
4159 schedule_debug(prev);
4161 hrtick_clear(rq);
4164 * Do the rq-clock update outside the rq lock:
4166 local_irq_disable();
4167 update_rq_clock(rq);
4168 spin_lock(&rq->lock);
4169 clear_tsk_need_resched(prev);
4171 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4172 if (unlikely(signal_pending_state(prev->state, prev)))
4173 prev->state = TASK_RUNNING;
4174 else
4175 deactivate_task(rq, prev, 1);
4176 switch_count = &prev->nvcsw;
4179 #ifdef CONFIG_SMP
4180 if (prev->sched_class->pre_schedule)
4181 prev->sched_class->pre_schedule(rq, prev);
4182 #endif
4184 if (unlikely(!rq->nr_running))
4185 idle_balance(cpu, rq);
4187 prev->sched_class->put_prev_task(rq, prev);
4188 next = pick_next_task(rq, prev);
4190 if (likely(prev != next)) {
4191 sched_info_switch(prev, next);
4193 rq->nr_switches++;
4194 rq->curr = next;
4195 ++*switch_count;
4197 context_switch(rq, prev, next); /* unlocks the rq */
4199 * the context switch might have flipped the stack from under
4200 * us, hence refresh the local variables.
4202 cpu = smp_processor_id();
4203 rq = cpu_rq(cpu);
4204 } else
4205 spin_unlock_irq(&rq->lock);
4207 hrtick_set(rq);
4209 if (unlikely(reacquire_kernel_lock(current) < 0))
4210 goto need_resched_nonpreemptible;
4212 preempt_enable_no_resched();
4213 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4214 goto need_resched;
4216 EXPORT_SYMBOL(schedule);
4218 #ifdef CONFIG_PREEMPT
4220 * this is the entry point to schedule() from in-kernel preemption
4221 * off of preempt_enable. Kernel preemptions off return from interrupt
4222 * occur there and call schedule directly.
4224 asmlinkage void __sched preempt_schedule(void)
4226 struct thread_info *ti = current_thread_info();
4229 * If there is a non-zero preempt_count or interrupts are disabled,
4230 * we do not want to preempt the current task. Just return..
4232 if (likely(ti->preempt_count || irqs_disabled()))
4233 return;
4235 do {
4236 add_preempt_count(PREEMPT_ACTIVE);
4237 schedule();
4238 sub_preempt_count(PREEMPT_ACTIVE);
4241 * Check again in case we missed a preemption opportunity
4242 * between schedule and now.
4244 barrier();
4245 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4247 EXPORT_SYMBOL(preempt_schedule);
4250 * this is the entry point to schedule() from kernel preemption
4251 * off of irq context.
4252 * Note, that this is called and return with irqs disabled. This will
4253 * protect us against recursive calling from irq.
4255 asmlinkage void __sched preempt_schedule_irq(void)
4257 struct thread_info *ti = current_thread_info();
4259 /* Catch callers which need to be fixed */
4260 BUG_ON(ti->preempt_count || !irqs_disabled());
4262 do {
4263 add_preempt_count(PREEMPT_ACTIVE);
4264 local_irq_enable();
4265 schedule();
4266 local_irq_disable();
4267 sub_preempt_count(PREEMPT_ACTIVE);
4270 * Check again in case we missed a preemption opportunity
4271 * between schedule and now.
4273 barrier();
4274 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4277 #endif /* CONFIG_PREEMPT */
4279 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4280 void *key)
4282 return try_to_wake_up(curr->private, mode, sync);
4284 EXPORT_SYMBOL(default_wake_function);
4287 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4288 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4289 * number) then we wake all the non-exclusive tasks and one exclusive task.
4291 * There are circumstances in which we can try to wake a task which has already
4292 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4293 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4295 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4296 int nr_exclusive, int sync, void *key)
4298 wait_queue_t *curr, *next;
4300 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4301 unsigned flags = curr->flags;
4303 if (curr->func(curr, mode, sync, key) &&
4304 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4305 break;
4310 * __wake_up - wake up threads blocked on a waitqueue.
4311 * @q: the waitqueue
4312 * @mode: which threads
4313 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4314 * @key: is directly passed to the wakeup function
4316 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4317 int nr_exclusive, void *key)
4319 unsigned long flags;
4321 spin_lock_irqsave(&q->lock, flags);
4322 __wake_up_common(q, mode, nr_exclusive, 0, key);
4323 spin_unlock_irqrestore(&q->lock, flags);
4325 EXPORT_SYMBOL(__wake_up);
4328 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4330 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4332 __wake_up_common(q, mode, 1, 0, NULL);
4336 * __wake_up_sync - wake up threads blocked on a waitqueue.
4337 * @q: the waitqueue
4338 * @mode: which threads
4339 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4341 * The sync wakeup differs that the waker knows that it will schedule
4342 * away soon, so while the target thread will be woken up, it will not
4343 * be migrated to another CPU - ie. the two threads are 'synchronized'
4344 * with each other. This can prevent needless bouncing between CPUs.
4346 * On UP it can prevent extra preemption.
4348 void
4349 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4351 unsigned long flags;
4352 int sync = 1;
4354 if (unlikely(!q))
4355 return;
4357 if (unlikely(!nr_exclusive))
4358 sync = 0;
4360 spin_lock_irqsave(&q->lock, flags);
4361 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4362 spin_unlock_irqrestore(&q->lock, flags);
4364 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4366 void complete(struct completion *x)
4368 unsigned long flags;
4370 spin_lock_irqsave(&x->wait.lock, flags);
4371 x->done++;
4372 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4373 spin_unlock_irqrestore(&x->wait.lock, flags);
4375 EXPORT_SYMBOL(complete);
4377 void complete_all(struct completion *x)
4379 unsigned long flags;
4381 spin_lock_irqsave(&x->wait.lock, flags);
4382 x->done += UINT_MAX/2;
4383 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4384 spin_unlock_irqrestore(&x->wait.lock, flags);
4386 EXPORT_SYMBOL(complete_all);
4388 static inline long __sched
4389 do_wait_for_common(struct completion *x, long timeout, int state)
4391 if (!x->done) {
4392 DECLARE_WAITQUEUE(wait, current);
4394 wait.flags |= WQ_FLAG_EXCLUSIVE;
4395 __add_wait_queue_tail(&x->wait, &wait);
4396 do {
4397 if ((state == TASK_INTERRUPTIBLE &&
4398 signal_pending(current)) ||
4399 (state == TASK_KILLABLE &&
4400 fatal_signal_pending(current))) {
4401 timeout = -ERESTARTSYS;
4402 break;
4404 __set_current_state(state);
4405 spin_unlock_irq(&x->wait.lock);
4406 timeout = schedule_timeout(timeout);
4407 spin_lock_irq(&x->wait.lock);
4408 } while (!x->done && timeout);
4409 __remove_wait_queue(&x->wait, &wait);
4410 if (!x->done)
4411 return timeout;
4413 x->done--;
4414 return timeout ?: 1;
4417 static long __sched
4418 wait_for_common(struct completion *x, long timeout, int state)
4420 might_sleep();
4422 spin_lock_irq(&x->wait.lock);
4423 timeout = do_wait_for_common(x, timeout, state);
4424 spin_unlock_irq(&x->wait.lock);
4425 return timeout;
4428 void __sched wait_for_completion(struct completion *x)
4430 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4432 EXPORT_SYMBOL(wait_for_completion);
4434 unsigned long __sched
4435 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4437 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4439 EXPORT_SYMBOL(wait_for_completion_timeout);
4441 int __sched wait_for_completion_interruptible(struct completion *x)
4443 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4444 if (t == -ERESTARTSYS)
4445 return t;
4446 return 0;
4448 EXPORT_SYMBOL(wait_for_completion_interruptible);
4450 unsigned long __sched
4451 wait_for_completion_interruptible_timeout(struct completion *x,
4452 unsigned long timeout)
4454 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4456 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4458 int __sched wait_for_completion_killable(struct completion *x)
4460 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4461 if (t == -ERESTARTSYS)
4462 return t;
4463 return 0;
4465 EXPORT_SYMBOL(wait_for_completion_killable);
4467 static long __sched
4468 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4470 unsigned long flags;
4471 wait_queue_t wait;
4473 init_waitqueue_entry(&wait, current);
4475 __set_current_state(state);
4477 spin_lock_irqsave(&q->lock, flags);
4478 __add_wait_queue(q, &wait);
4479 spin_unlock(&q->lock);
4480 timeout = schedule_timeout(timeout);
4481 spin_lock_irq(&q->lock);
4482 __remove_wait_queue(q, &wait);
4483 spin_unlock_irqrestore(&q->lock, flags);
4485 return timeout;
4488 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4490 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4492 EXPORT_SYMBOL(interruptible_sleep_on);
4494 long __sched
4495 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4497 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4499 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4501 void __sched sleep_on(wait_queue_head_t *q)
4503 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4505 EXPORT_SYMBOL(sleep_on);
4507 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4509 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4511 EXPORT_SYMBOL(sleep_on_timeout);
4513 #ifdef CONFIG_RT_MUTEXES
4516 * rt_mutex_setprio - set the current priority of a task
4517 * @p: task
4518 * @prio: prio value (kernel-internal form)
4520 * This function changes the 'effective' priority of a task. It does
4521 * not touch ->normal_prio like __setscheduler().
4523 * Used by the rt_mutex code to implement priority inheritance logic.
4525 void rt_mutex_setprio(struct task_struct *p, int prio)
4527 unsigned long flags;
4528 int oldprio, on_rq, running;
4529 struct rq *rq;
4530 const struct sched_class *prev_class = p->sched_class;
4532 BUG_ON(prio < 0 || prio > MAX_PRIO);
4534 rq = task_rq_lock(p, &flags);
4535 update_rq_clock(rq);
4537 oldprio = p->prio;
4538 on_rq = p->se.on_rq;
4539 running = task_current(rq, p);
4540 if (on_rq)
4541 dequeue_task(rq, p, 0);
4542 if (running)
4543 p->sched_class->put_prev_task(rq, p);
4545 if (rt_prio(prio))
4546 p->sched_class = &rt_sched_class;
4547 else
4548 p->sched_class = &fair_sched_class;
4550 p->prio = prio;
4552 if (running)
4553 p->sched_class->set_curr_task(rq);
4554 if (on_rq) {
4555 enqueue_task(rq, p, 0);
4557 check_class_changed(rq, p, prev_class, oldprio, running);
4559 task_rq_unlock(rq, &flags);
4562 #endif
4564 void set_user_nice(struct task_struct *p, long nice)
4566 int old_prio, delta, on_rq;
4567 unsigned long flags;
4568 struct rq *rq;
4570 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4571 return;
4573 * We have to be careful, if called from sys_setpriority(),
4574 * the task might be in the middle of scheduling on another CPU.
4576 rq = task_rq_lock(p, &flags);
4577 update_rq_clock(rq);
4579 * The RT priorities are set via sched_setscheduler(), but we still
4580 * allow the 'normal' nice value to be set - but as expected
4581 * it wont have any effect on scheduling until the task is
4582 * SCHED_FIFO/SCHED_RR:
4584 if (task_has_rt_policy(p)) {
4585 p->static_prio = NICE_TO_PRIO(nice);
4586 goto out_unlock;
4588 on_rq = p->se.on_rq;
4589 if (on_rq) {
4590 dequeue_task(rq, p, 0);
4591 dec_load(rq, p);
4594 p->static_prio = NICE_TO_PRIO(nice);
4595 set_load_weight(p);
4596 old_prio = p->prio;
4597 p->prio = effective_prio(p);
4598 delta = p->prio - old_prio;
4600 if (on_rq) {
4601 enqueue_task(rq, p, 0);
4602 inc_load(rq, p);
4604 * If the task increased its priority or is running and
4605 * lowered its priority, then reschedule its CPU:
4607 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4608 resched_task(rq->curr);
4610 out_unlock:
4611 task_rq_unlock(rq, &flags);
4613 EXPORT_SYMBOL(set_user_nice);
4616 * can_nice - check if a task can reduce its nice value
4617 * @p: task
4618 * @nice: nice value
4620 int can_nice(const struct task_struct *p, const int nice)
4622 /* convert nice value [19,-20] to rlimit style value [1,40] */
4623 int nice_rlim = 20 - nice;
4625 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4626 capable(CAP_SYS_NICE));
4629 #ifdef __ARCH_WANT_SYS_NICE
4632 * sys_nice - change the priority of the current process.
4633 * @increment: priority increment
4635 * sys_setpriority is a more generic, but much slower function that
4636 * does similar things.
4638 asmlinkage long sys_nice(int increment)
4640 long nice, retval;
4643 * Setpriority might change our priority at the same moment.
4644 * We don't have to worry. Conceptually one call occurs first
4645 * and we have a single winner.
4647 if (increment < -40)
4648 increment = -40;
4649 if (increment > 40)
4650 increment = 40;
4652 nice = PRIO_TO_NICE(current->static_prio) + increment;
4653 if (nice < -20)
4654 nice = -20;
4655 if (nice > 19)
4656 nice = 19;
4658 if (increment < 0 && !can_nice(current, nice))
4659 return -EPERM;
4661 retval = security_task_setnice(current, nice);
4662 if (retval)
4663 return retval;
4665 set_user_nice(current, nice);
4666 return 0;
4669 #endif
4672 * task_prio - return the priority value of a given task.
4673 * @p: the task in question.
4675 * This is the priority value as seen by users in /proc.
4676 * RT tasks are offset by -200. Normal tasks are centered
4677 * around 0, value goes from -16 to +15.
4679 int task_prio(const struct task_struct *p)
4681 return p->prio - MAX_RT_PRIO;
4685 * task_nice - return the nice value of a given task.
4686 * @p: the task in question.
4688 int task_nice(const struct task_struct *p)
4690 return TASK_NICE(p);
4692 EXPORT_SYMBOL(task_nice);
4695 * idle_cpu - is a given cpu idle currently?
4696 * @cpu: the processor in question.
4698 int idle_cpu(int cpu)
4700 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4704 * idle_task - return the idle task for a given cpu.
4705 * @cpu: the processor in question.
4707 struct task_struct *idle_task(int cpu)
4709 return cpu_rq(cpu)->idle;
4713 * find_process_by_pid - find a process with a matching PID value.
4714 * @pid: the pid in question.
4716 static struct task_struct *find_process_by_pid(pid_t pid)
4718 return pid ? find_task_by_vpid(pid) : current;
4721 /* Actually do priority change: must hold rq lock. */
4722 static void
4723 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4725 BUG_ON(p->se.on_rq);
4727 p->policy = policy;
4728 switch (p->policy) {
4729 case SCHED_NORMAL:
4730 case SCHED_BATCH:
4731 case SCHED_IDLE:
4732 p->sched_class = &fair_sched_class;
4733 break;
4734 case SCHED_FIFO:
4735 case SCHED_RR:
4736 p->sched_class = &rt_sched_class;
4737 break;
4740 p->rt_priority = prio;
4741 p->normal_prio = normal_prio(p);
4742 /* we are holding p->pi_lock already */
4743 p->prio = rt_mutex_getprio(p);
4744 set_load_weight(p);
4748 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4749 * @p: the task in question.
4750 * @policy: new policy.
4751 * @param: structure containing the new RT priority.
4753 * NOTE that the task may be already dead.
4755 int sched_setscheduler(struct task_struct *p, int policy,
4756 struct sched_param *param)
4758 int retval, oldprio, oldpolicy = -1, on_rq, running;
4759 unsigned long flags;
4760 const struct sched_class *prev_class = p->sched_class;
4761 struct rq *rq;
4763 /* may grab non-irq protected spin_locks */
4764 BUG_ON(in_interrupt());
4765 recheck:
4766 /* double check policy once rq lock held */
4767 if (policy < 0)
4768 policy = oldpolicy = p->policy;
4769 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4770 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4771 policy != SCHED_IDLE)
4772 return -EINVAL;
4774 * Valid priorities for SCHED_FIFO and SCHED_RR are
4775 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4776 * SCHED_BATCH and SCHED_IDLE is 0.
4778 if (param->sched_priority < 0 ||
4779 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4780 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4781 return -EINVAL;
4782 if (rt_policy(policy) != (param->sched_priority != 0))
4783 return -EINVAL;
4786 * Allow unprivileged RT tasks to decrease priority:
4788 if (!capable(CAP_SYS_NICE)) {
4789 if (rt_policy(policy)) {
4790 unsigned long rlim_rtprio;
4792 if (!lock_task_sighand(p, &flags))
4793 return -ESRCH;
4794 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4795 unlock_task_sighand(p, &flags);
4797 /* can't set/change the rt policy */
4798 if (policy != p->policy && !rlim_rtprio)
4799 return -EPERM;
4801 /* can't increase priority */
4802 if (param->sched_priority > p->rt_priority &&
4803 param->sched_priority > rlim_rtprio)
4804 return -EPERM;
4807 * Like positive nice levels, dont allow tasks to
4808 * move out of SCHED_IDLE either:
4810 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4811 return -EPERM;
4813 /* can't change other user's priorities */
4814 if ((current->euid != p->euid) &&
4815 (current->euid != p->uid))
4816 return -EPERM;
4819 #ifdef CONFIG_RT_GROUP_SCHED
4821 * Do not allow realtime tasks into groups that have no runtime
4822 * assigned.
4824 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
4825 return -EPERM;
4826 #endif
4828 retval = security_task_setscheduler(p, policy, param);
4829 if (retval)
4830 return retval;
4832 * make sure no PI-waiters arrive (or leave) while we are
4833 * changing the priority of the task:
4835 spin_lock_irqsave(&p->pi_lock, flags);
4837 * To be able to change p->policy safely, the apropriate
4838 * runqueue lock must be held.
4840 rq = __task_rq_lock(p);
4841 /* recheck policy now with rq lock held */
4842 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4843 policy = oldpolicy = -1;
4844 __task_rq_unlock(rq);
4845 spin_unlock_irqrestore(&p->pi_lock, flags);
4846 goto recheck;
4848 update_rq_clock(rq);
4849 on_rq = p->se.on_rq;
4850 running = task_current(rq, p);
4851 if (on_rq)
4852 deactivate_task(rq, p, 0);
4853 if (running)
4854 p->sched_class->put_prev_task(rq, p);
4856 oldprio = p->prio;
4857 __setscheduler(rq, p, policy, param->sched_priority);
4859 if (running)
4860 p->sched_class->set_curr_task(rq);
4861 if (on_rq) {
4862 activate_task(rq, p, 0);
4864 check_class_changed(rq, p, prev_class, oldprio, running);
4866 __task_rq_unlock(rq);
4867 spin_unlock_irqrestore(&p->pi_lock, flags);
4869 rt_mutex_adjust_pi(p);
4871 return 0;
4873 EXPORT_SYMBOL_GPL(sched_setscheduler);
4875 static int
4876 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4878 struct sched_param lparam;
4879 struct task_struct *p;
4880 int retval;
4882 if (!param || pid < 0)
4883 return -EINVAL;
4884 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4885 return -EFAULT;
4887 rcu_read_lock();
4888 retval = -ESRCH;
4889 p = find_process_by_pid(pid);
4890 if (p != NULL)
4891 retval = sched_setscheduler(p, policy, &lparam);
4892 rcu_read_unlock();
4894 return retval;
4898 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4899 * @pid: the pid in question.
4900 * @policy: new policy.
4901 * @param: structure containing the new RT priority.
4903 asmlinkage long
4904 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4906 /* negative values for policy are not valid */
4907 if (policy < 0)
4908 return -EINVAL;
4910 return do_sched_setscheduler(pid, policy, param);
4914 * sys_sched_setparam - set/change the RT priority of a thread
4915 * @pid: the pid in question.
4916 * @param: structure containing the new RT priority.
4918 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4920 return do_sched_setscheduler(pid, -1, param);
4924 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4925 * @pid: the pid in question.
4927 asmlinkage long sys_sched_getscheduler(pid_t pid)
4929 struct task_struct *p;
4930 int retval;
4932 if (pid < 0)
4933 return -EINVAL;
4935 retval = -ESRCH;
4936 read_lock(&tasklist_lock);
4937 p = find_process_by_pid(pid);
4938 if (p) {
4939 retval = security_task_getscheduler(p);
4940 if (!retval)
4941 retval = p->policy;
4943 read_unlock(&tasklist_lock);
4944 return retval;
4948 * sys_sched_getscheduler - get the RT priority of a thread
4949 * @pid: the pid in question.
4950 * @param: structure containing the RT priority.
4952 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4954 struct sched_param lp;
4955 struct task_struct *p;
4956 int retval;
4958 if (!param || pid < 0)
4959 return -EINVAL;
4961 read_lock(&tasklist_lock);
4962 p = find_process_by_pid(pid);
4963 retval = -ESRCH;
4964 if (!p)
4965 goto out_unlock;
4967 retval = security_task_getscheduler(p);
4968 if (retval)
4969 goto out_unlock;
4971 lp.sched_priority = p->rt_priority;
4972 read_unlock(&tasklist_lock);
4975 * This one might sleep, we cannot do it with a spinlock held ...
4977 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4979 return retval;
4981 out_unlock:
4982 read_unlock(&tasklist_lock);
4983 return retval;
4986 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
4988 cpumask_t cpus_allowed;
4989 cpumask_t new_mask = *in_mask;
4990 struct task_struct *p;
4991 int retval;
4993 get_online_cpus();
4994 read_lock(&tasklist_lock);
4996 p = find_process_by_pid(pid);
4997 if (!p) {
4998 read_unlock(&tasklist_lock);
4999 put_online_cpus();
5000 return -ESRCH;
5004 * It is not safe to call set_cpus_allowed with the
5005 * tasklist_lock held. We will bump the task_struct's
5006 * usage count and then drop tasklist_lock.
5008 get_task_struct(p);
5009 read_unlock(&tasklist_lock);
5011 retval = -EPERM;
5012 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5013 !capable(CAP_SYS_NICE))
5014 goto out_unlock;
5016 retval = security_task_setscheduler(p, 0, NULL);
5017 if (retval)
5018 goto out_unlock;
5020 cpuset_cpus_allowed(p, &cpus_allowed);
5021 cpus_and(new_mask, new_mask, cpus_allowed);
5022 again:
5023 retval = set_cpus_allowed_ptr(p, &new_mask);
5025 if (!retval) {
5026 cpuset_cpus_allowed(p, &cpus_allowed);
5027 if (!cpus_subset(new_mask, cpus_allowed)) {
5029 * We must have raced with a concurrent cpuset
5030 * update. Just reset the cpus_allowed to the
5031 * cpuset's cpus_allowed
5033 new_mask = cpus_allowed;
5034 goto again;
5037 out_unlock:
5038 put_task_struct(p);
5039 put_online_cpus();
5040 return retval;
5043 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5044 cpumask_t *new_mask)
5046 if (len < sizeof(cpumask_t)) {
5047 memset(new_mask, 0, sizeof(cpumask_t));
5048 } else if (len > sizeof(cpumask_t)) {
5049 len = sizeof(cpumask_t);
5051 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5055 * sys_sched_setaffinity - set the cpu affinity of a process
5056 * @pid: pid of the process
5057 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5058 * @user_mask_ptr: user-space pointer to the new cpu mask
5060 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5061 unsigned long __user *user_mask_ptr)
5063 cpumask_t new_mask;
5064 int retval;
5066 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5067 if (retval)
5068 return retval;
5070 return sched_setaffinity(pid, &new_mask);
5074 * Represents all cpu's present in the system
5075 * In systems capable of hotplug, this map could dynamically grow
5076 * as new cpu's are detected in the system via any platform specific
5077 * method, such as ACPI for e.g.
5080 cpumask_t cpu_present_map __read_mostly;
5081 EXPORT_SYMBOL(cpu_present_map);
5083 #ifndef CONFIG_SMP
5084 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
5085 EXPORT_SYMBOL(cpu_online_map);
5087 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
5088 EXPORT_SYMBOL(cpu_possible_map);
5089 #endif
5091 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5093 struct task_struct *p;
5094 int retval;
5096 get_online_cpus();
5097 read_lock(&tasklist_lock);
5099 retval = -ESRCH;
5100 p = find_process_by_pid(pid);
5101 if (!p)
5102 goto out_unlock;
5104 retval = security_task_getscheduler(p);
5105 if (retval)
5106 goto out_unlock;
5108 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5110 out_unlock:
5111 read_unlock(&tasklist_lock);
5112 put_online_cpus();
5114 return retval;
5118 * sys_sched_getaffinity - get the cpu affinity of a process
5119 * @pid: pid of the process
5120 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5121 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5123 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5124 unsigned long __user *user_mask_ptr)
5126 int ret;
5127 cpumask_t mask;
5129 if (len < sizeof(cpumask_t))
5130 return -EINVAL;
5132 ret = sched_getaffinity(pid, &mask);
5133 if (ret < 0)
5134 return ret;
5136 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5137 return -EFAULT;
5139 return sizeof(cpumask_t);
5143 * sys_sched_yield - yield the current processor to other threads.
5145 * This function yields the current CPU to other tasks. If there are no
5146 * other threads running on this CPU then this function will return.
5148 asmlinkage long sys_sched_yield(void)
5150 struct rq *rq = this_rq_lock();
5152 schedstat_inc(rq, yld_count);
5153 current->sched_class->yield_task(rq);
5156 * Since we are going to call schedule() anyway, there's
5157 * no need to preempt or enable interrupts:
5159 __release(rq->lock);
5160 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5161 _raw_spin_unlock(&rq->lock);
5162 preempt_enable_no_resched();
5164 schedule();
5166 return 0;
5169 static void __cond_resched(void)
5171 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5172 __might_sleep(__FILE__, __LINE__);
5173 #endif
5175 * The BKS might be reacquired before we have dropped
5176 * PREEMPT_ACTIVE, which could trigger a second
5177 * cond_resched() call.
5179 do {
5180 add_preempt_count(PREEMPT_ACTIVE);
5181 schedule();
5182 sub_preempt_count(PREEMPT_ACTIVE);
5183 } while (need_resched());
5186 int __sched _cond_resched(void)
5188 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5189 system_state == SYSTEM_RUNNING) {
5190 __cond_resched();
5191 return 1;
5193 return 0;
5195 EXPORT_SYMBOL(_cond_resched);
5198 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5199 * call schedule, and on return reacquire the lock.
5201 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5202 * operations here to prevent schedule() from being called twice (once via
5203 * spin_unlock(), once by hand).
5205 int cond_resched_lock(spinlock_t *lock)
5207 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5208 int ret = 0;
5210 if (spin_needbreak(lock) || resched) {
5211 spin_unlock(lock);
5212 if (resched && need_resched())
5213 __cond_resched();
5214 else
5215 cpu_relax();
5216 ret = 1;
5217 spin_lock(lock);
5219 return ret;
5221 EXPORT_SYMBOL(cond_resched_lock);
5223 int __sched cond_resched_softirq(void)
5225 BUG_ON(!in_softirq());
5227 if (need_resched() && system_state == SYSTEM_RUNNING) {
5228 local_bh_enable();
5229 __cond_resched();
5230 local_bh_disable();
5231 return 1;
5233 return 0;
5235 EXPORT_SYMBOL(cond_resched_softirq);
5238 * yield - yield the current processor to other threads.
5240 * This is a shortcut for kernel-space yielding - it marks the
5241 * thread runnable and calls sys_sched_yield().
5243 void __sched yield(void)
5245 set_current_state(TASK_RUNNING);
5246 sys_sched_yield();
5248 EXPORT_SYMBOL(yield);
5251 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5252 * that process accounting knows that this is a task in IO wait state.
5254 * But don't do that if it is a deliberate, throttling IO wait (this task
5255 * has set its backing_dev_info: the queue against which it should throttle)
5257 void __sched io_schedule(void)
5259 struct rq *rq = &__raw_get_cpu_var(runqueues);
5261 delayacct_blkio_start();
5262 atomic_inc(&rq->nr_iowait);
5263 schedule();
5264 atomic_dec(&rq->nr_iowait);
5265 delayacct_blkio_end();
5267 EXPORT_SYMBOL(io_schedule);
5269 long __sched io_schedule_timeout(long timeout)
5271 struct rq *rq = &__raw_get_cpu_var(runqueues);
5272 long ret;
5274 delayacct_blkio_start();
5275 atomic_inc(&rq->nr_iowait);
5276 ret = schedule_timeout(timeout);
5277 atomic_dec(&rq->nr_iowait);
5278 delayacct_blkio_end();
5279 return ret;
5283 * sys_sched_get_priority_max - return maximum RT priority.
5284 * @policy: scheduling class.
5286 * this syscall returns the maximum rt_priority that can be used
5287 * by a given scheduling class.
5289 asmlinkage long sys_sched_get_priority_max(int policy)
5291 int ret = -EINVAL;
5293 switch (policy) {
5294 case SCHED_FIFO:
5295 case SCHED_RR:
5296 ret = MAX_USER_RT_PRIO-1;
5297 break;
5298 case SCHED_NORMAL:
5299 case SCHED_BATCH:
5300 case SCHED_IDLE:
5301 ret = 0;
5302 break;
5304 return ret;
5308 * sys_sched_get_priority_min - return minimum RT priority.
5309 * @policy: scheduling class.
5311 * this syscall returns the minimum rt_priority that can be used
5312 * by a given scheduling class.
5314 asmlinkage long sys_sched_get_priority_min(int policy)
5316 int ret = -EINVAL;
5318 switch (policy) {
5319 case SCHED_FIFO:
5320 case SCHED_RR:
5321 ret = 1;
5322 break;
5323 case SCHED_NORMAL:
5324 case SCHED_BATCH:
5325 case SCHED_IDLE:
5326 ret = 0;
5328 return ret;
5332 * sys_sched_rr_get_interval - return the default timeslice of a process.
5333 * @pid: pid of the process.
5334 * @interval: userspace pointer to the timeslice value.
5336 * this syscall writes the default timeslice value of a given process
5337 * into the user-space timespec buffer. A value of '0' means infinity.
5339 asmlinkage
5340 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5342 struct task_struct *p;
5343 unsigned int time_slice;
5344 int retval;
5345 struct timespec t;
5347 if (pid < 0)
5348 return -EINVAL;
5350 retval = -ESRCH;
5351 read_lock(&tasklist_lock);
5352 p = find_process_by_pid(pid);
5353 if (!p)
5354 goto out_unlock;
5356 retval = security_task_getscheduler(p);
5357 if (retval)
5358 goto out_unlock;
5361 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5362 * tasks that are on an otherwise idle runqueue:
5364 time_slice = 0;
5365 if (p->policy == SCHED_RR) {
5366 time_slice = DEF_TIMESLICE;
5367 } else if (p->policy != SCHED_FIFO) {
5368 struct sched_entity *se = &p->se;
5369 unsigned long flags;
5370 struct rq *rq;
5372 rq = task_rq_lock(p, &flags);
5373 if (rq->cfs.load.weight)
5374 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5375 task_rq_unlock(rq, &flags);
5377 read_unlock(&tasklist_lock);
5378 jiffies_to_timespec(time_slice, &t);
5379 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5380 return retval;
5382 out_unlock:
5383 read_unlock(&tasklist_lock);
5384 return retval;
5387 static const char stat_nam[] = "RSDTtZX";
5389 void sched_show_task(struct task_struct *p)
5391 unsigned long free = 0;
5392 unsigned state;
5394 state = p->state ? __ffs(p->state) + 1 : 0;
5395 printk(KERN_INFO "%-13.13s %c", p->comm,
5396 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5397 #if BITS_PER_LONG == 32
5398 if (state == TASK_RUNNING)
5399 printk(KERN_CONT " running ");
5400 else
5401 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5402 #else
5403 if (state == TASK_RUNNING)
5404 printk(KERN_CONT " running task ");
5405 else
5406 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5407 #endif
5408 #ifdef CONFIG_DEBUG_STACK_USAGE
5410 unsigned long *n = end_of_stack(p);
5411 while (!*n)
5412 n++;
5413 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5415 #endif
5416 printk(KERN_CONT "%5lu %5d %6d\n", free,
5417 task_pid_nr(p), task_pid_nr(p->real_parent));
5419 show_stack(p, NULL);
5422 void show_state_filter(unsigned long state_filter)
5424 struct task_struct *g, *p;
5426 #if BITS_PER_LONG == 32
5427 printk(KERN_INFO
5428 " task PC stack pid father\n");
5429 #else
5430 printk(KERN_INFO
5431 " task PC stack pid father\n");
5432 #endif
5433 read_lock(&tasklist_lock);
5434 do_each_thread(g, p) {
5436 * reset the NMI-timeout, listing all files on a slow
5437 * console might take alot of time:
5439 touch_nmi_watchdog();
5440 if (!state_filter || (p->state & state_filter))
5441 sched_show_task(p);
5442 } while_each_thread(g, p);
5444 touch_all_softlockup_watchdogs();
5446 #ifdef CONFIG_SCHED_DEBUG
5447 sysrq_sched_debug_show();
5448 #endif
5449 read_unlock(&tasklist_lock);
5451 * Only show locks if all tasks are dumped:
5453 if (state_filter == -1)
5454 debug_show_all_locks();
5457 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5459 idle->sched_class = &idle_sched_class;
5463 * init_idle - set up an idle thread for a given CPU
5464 * @idle: task in question
5465 * @cpu: cpu the idle task belongs to
5467 * NOTE: this function does not set the idle thread's NEED_RESCHED
5468 * flag, to make booting more robust.
5470 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5472 struct rq *rq = cpu_rq(cpu);
5473 unsigned long flags;
5475 __sched_fork(idle);
5476 idle->se.exec_start = sched_clock();
5478 idle->prio = idle->normal_prio = MAX_PRIO;
5479 idle->cpus_allowed = cpumask_of_cpu(cpu);
5480 __set_task_cpu(idle, cpu);
5482 spin_lock_irqsave(&rq->lock, flags);
5483 rq->curr = rq->idle = idle;
5484 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5485 idle->oncpu = 1;
5486 #endif
5487 spin_unlock_irqrestore(&rq->lock, flags);
5489 /* Set the preempt count _outside_ the spinlocks! */
5490 #if defined(CONFIG_PREEMPT)
5491 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5492 #else
5493 task_thread_info(idle)->preempt_count = 0;
5494 #endif
5496 * The idle tasks have their own, simple scheduling class:
5498 idle->sched_class = &idle_sched_class;
5502 * In a system that switches off the HZ timer nohz_cpu_mask
5503 * indicates which cpus entered this state. This is used
5504 * in the rcu update to wait only for active cpus. For system
5505 * which do not switch off the HZ timer nohz_cpu_mask should
5506 * always be CPU_MASK_NONE.
5508 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5511 * Increase the granularity value when there are more CPUs,
5512 * because with more CPUs the 'effective latency' as visible
5513 * to users decreases. But the relationship is not linear,
5514 * so pick a second-best guess by going with the log2 of the
5515 * number of CPUs.
5517 * This idea comes from the SD scheduler of Con Kolivas:
5519 static inline void sched_init_granularity(void)
5521 unsigned int factor = 1 + ilog2(num_online_cpus());
5522 const unsigned long limit = 200000000;
5524 sysctl_sched_min_granularity *= factor;
5525 if (sysctl_sched_min_granularity > limit)
5526 sysctl_sched_min_granularity = limit;
5528 sysctl_sched_latency *= factor;
5529 if (sysctl_sched_latency > limit)
5530 sysctl_sched_latency = limit;
5532 sysctl_sched_wakeup_granularity *= factor;
5535 #ifdef CONFIG_SMP
5537 * This is how migration works:
5539 * 1) we queue a struct migration_req structure in the source CPU's
5540 * runqueue and wake up that CPU's migration thread.
5541 * 2) we down() the locked semaphore => thread blocks.
5542 * 3) migration thread wakes up (implicitly it forces the migrated
5543 * thread off the CPU)
5544 * 4) it gets the migration request and checks whether the migrated
5545 * task is still in the wrong runqueue.
5546 * 5) if it's in the wrong runqueue then the migration thread removes
5547 * it and puts it into the right queue.
5548 * 6) migration thread up()s the semaphore.
5549 * 7) we wake up and the migration is done.
5553 * Change a given task's CPU affinity. Migrate the thread to a
5554 * proper CPU and schedule it away if the CPU it's executing on
5555 * is removed from the allowed bitmask.
5557 * NOTE: the caller must have a valid reference to the task, the
5558 * task must not exit() & deallocate itself prematurely. The
5559 * call is not atomic; no spinlocks may be held.
5561 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5563 struct migration_req req;
5564 unsigned long flags;
5565 struct rq *rq;
5566 int ret = 0;
5568 rq = task_rq_lock(p, &flags);
5569 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5570 ret = -EINVAL;
5571 goto out;
5574 if (p->sched_class->set_cpus_allowed)
5575 p->sched_class->set_cpus_allowed(p, new_mask);
5576 else {
5577 p->cpus_allowed = *new_mask;
5578 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5581 /* Can the task run on the task's current CPU? If so, we're done */
5582 if (cpu_isset(task_cpu(p), *new_mask))
5583 goto out;
5585 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5586 /* Need help from migration thread: drop lock and wait. */
5587 task_rq_unlock(rq, &flags);
5588 wake_up_process(rq->migration_thread);
5589 wait_for_completion(&req.done);
5590 tlb_migrate_finish(p->mm);
5591 return 0;
5593 out:
5594 task_rq_unlock(rq, &flags);
5596 return ret;
5598 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5601 * Move (not current) task off this cpu, onto dest cpu. We're doing
5602 * this because either it can't run here any more (set_cpus_allowed()
5603 * away from this CPU, or CPU going down), or because we're
5604 * attempting to rebalance this task on exec (sched_exec).
5606 * So we race with normal scheduler movements, but that's OK, as long
5607 * as the task is no longer on this CPU.
5609 * Returns non-zero if task was successfully migrated.
5611 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5613 struct rq *rq_dest, *rq_src;
5614 int ret = 0, on_rq;
5616 if (unlikely(cpu_is_offline(dest_cpu)))
5617 return ret;
5619 rq_src = cpu_rq(src_cpu);
5620 rq_dest = cpu_rq(dest_cpu);
5622 double_rq_lock(rq_src, rq_dest);
5623 /* Already moved. */
5624 if (task_cpu(p) != src_cpu)
5625 goto done;
5626 /* Affinity changed (again). */
5627 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5628 goto fail;
5630 on_rq = p->se.on_rq;
5631 if (on_rq)
5632 deactivate_task(rq_src, p, 0);
5634 set_task_cpu(p, dest_cpu);
5635 if (on_rq) {
5636 activate_task(rq_dest, p, 0);
5637 check_preempt_curr(rq_dest, p);
5639 done:
5640 ret = 1;
5641 fail:
5642 double_rq_unlock(rq_src, rq_dest);
5643 return ret;
5647 * migration_thread - this is a highprio system thread that performs
5648 * thread migration by bumping thread off CPU then 'pushing' onto
5649 * another runqueue.
5651 static int migration_thread(void *data)
5653 int cpu = (long)data;
5654 struct rq *rq;
5656 rq = cpu_rq(cpu);
5657 BUG_ON(rq->migration_thread != current);
5659 set_current_state(TASK_INTERRUPTIBLE);
5660 while (!kthread_should_stop()) {
5661 struct migration_req *req;
5662 struct list_head *head;
5664 spin_lock_irq(&rq->lock);
5666 if (cpu_is_offline(cpu)) {
5667 spin_unlock_irq(&rq->lock);
5668 goto wait_to_die;
5671 if (rq->active_balance) {
5672 active_load_balance(rq, cpu);
5673 rq->active_balance = 0;
5676 head = &rq->migration_queue;
5678 if (list_empty(head)) {
5679 spin_unlock_irq(&rq->lock);
5680 schedule();
5681 set_current_state(TASK_INTERRUPTIBLE);
5682 continue;
5684 req = list_entry(head->next, struct migration_req, list);
5685 list_del_init(head->next);
5687 spin_unlock(&rq->lock);
5688 __migrate_task(req->task, cpu, req->dest_cpu);
5689 local_irq_enable();
5691 complete(&req->done);
5693 __set_current_state(TASK_RUNNING);
5694 return 0;
5696 wait_to_die:
5697 /* Wait for kthread_stop */
5698 set_current_state(TASK_INTERRUPTIBLE);
5699 while (!kthread_should_stop()) {
5700 schedule();
5701 set_current_state(TASK_INTERRUPTIBLE);
5703 __set_current_state(TASK_RUNNING);
5704 return 0;
5707 #ifdef CONFIG_HOTPLUG_CPU
5709 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5711 int ret;
5713 local_irq_disable();
5714 ret = __migrate_task(p, src_cpu, dest_cpu);
5715 local_irq_enable();
5716 return ret;
5720 * Figure out where task on dead CPU should go, use force if necessary.
5721 * NOTE: interrupts should be disabled by the caller
5723 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5725 unsigned long flags;
5726 cpumask_t mask;
5727 struct rq *rq;
5728 int dest_cpu;
5730 do {
5731 /* On same node? */
5732 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5733 cpus_and(mask, mask, p->cpus_allowed);
5734 dest_cpu = any_online_cpu(mask);
5736 /* On any allowed CPU? */
5737 if (dest_cpu >= nr_cpu_ids)
5738 dest_cpu = any_online_cpu(p->cpus_allowed);
5740 /* No more Mr. Nice Guy. */
5741 if (dest_cpu >= nr_cpu_ids) {
5742 cpumask_t cpus_allowed;
5744 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5746 * Try to stay on the same cpuset, where the
5747 * current cpuset may be a subset of all cpus.
5748 * The cpuset_cpus_allowed_locked() variant of
5749 * cpuset_cpus_allowed() will not block. It must be
5750 * called within calls to cpuset_lock/cpuset_unlock.
5752 rq = task_rq_lock(p, &flags);
5753 p->cpus_allowed = cpus_allowed;
5754 dest_cpu = any_online_cpu(p->cpus_allowed);
5755 task_rq_unlock(rq, &flags);
5758 * Don't tell them about moving exiting tasks or
5759 * kernel threads (both mm NULL), since they never
5760 * leave kernel.
5762 if (p->mm && printk_ratelimit()) {
5763 printk(KERN_INFO "process %d (%s) no "
5764 "longer affine to cpu%d\n",
5765 task_pid_nr(p), p->comm, dead_cpu);
5768 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5772 * While a dead CPU has no uninterruptible tasks queued at this point,
5773 * it might still have a nonzero ->nr_uninterruptible counter, because
5774 * for performance reasons the counter is not stricly tracking tasks to
5775 * their home CPUs. So we just add the counter to another CPU's counter,
5776 * to keep the global sum constant after CPU-down:
5778 static void migrate_nr_uninterruptible(struct rq *rq_src)
5780 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5781 unsigned long flags;
5783 local_irq_save(flags);
5784 double_rq_lock(rq_src, rq_dest);
5785 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5786 rq_src->nr_uninterruptible = 0;
5787 double_rq_unlock(rq_src, rq_dest);
5788 local_irq_restore(flags);
5791 /* Run through task list and migrate tasks from the dead cpu. */
5792 static void migrate_live_tasks(int src_cpu)
5794 struct task_struct *p, *t;
5796 read_lock(&tasklist_lock);
5798 do_each_thread(t, p) {
5799 if (p == current)
5800 continue;
5802 if (task_cpu(p) == src_cpu)
5803 move_task_off_dead_cpu(src_cpu, p);
5804 } while_each_thread(t, p);
5806 read_unlock(&tasklist_lock);
5810 * Schedules idle task to be the next runnable task on current CPU.
5811 * It does so by boosting its priority to highest possible.
5812 * Used by CPU offline code.
5814 void sched_idle_next(void)
5816 int this_cpu = smp_processor_id();
5817 struct rq *rq = cpu_rq(this_cpu);
5818 struct task_struct *p = rq->idle;
5819 unsigned long flags;
5821 /* cpu has to be offline */
5822 BUG_ON(cpu_online(this_cpu));
5825 * Strictly not necessary since rest of the CPUs are stopped by now
5826 * and interrupts disabled on the current cpu.
5828 spin_lock_irqsave(&rq->lock, flags);
5830 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5832 update_rq_clock(rq);
5833 activate_task(rq, p, 0);
5835 spin_unlock_irqrestore(&rq->lock, flags);
5839 * Ensures that the idle task is using init_mm right before its cpu goes
5840 * offline.
5842 void idle_task_exit(void)
5844 struct mm_struct *mm = current->active_mm;
5846 BUG_ON(cpu_online(smp_processor_id()));
5848 if (mm != &init_mm)
5849 switch_mm(mm, &init_mm, current);
5850 mmdrop(mm);
5853 /* called under rq->lock with disabled interrupts */
5854 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5856 struct rq *rq = cpu_rq(dead_cpu);
5858 /* Must be exiting, otherwise would be on tasklist. */
5859 BUG_ON(!p->exit_state);
5861 /* Cannot have done final schedule yet: would have vanished. */
5862 BUG_ON(p->state == TASK_DEAD);
5864 get_task_struct(p);
5867 * Drop lock around migration; if someone else moves it,
5868 * that's OK. No task can be added to this CPU, so iteration is
5869 * fine.
5871 spin_unlock_irq(&rq->lock);
5872 move_task_off_dead_cpu(dead_cpu, p);
5873 spin_lock_irq(&rq->lock);
5875 put_task_struct(p);
5878 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5879 static void migrate_dead_tasks(unsigned int dead_cpu)
5881 struct rq *rq = cpu_rq(dead_cpu);
5882 struct task_struct *next;
5884 for ( ; ; ) {
5885 if (!rq->nr_running)
5886 break;
5887 update_rq_clock(rq);
5888 next = pick_next_task(rq, rq->curr);
5889 if (!next)
5890 break;
5891 next->sched_class->put_prev_task(rq, next);
5892 migrate_dead(dead_cpu, next);
5896 #endif /* CONFIG_HOTPLUG_CPU */
5898 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5900 static struct ctl_table sd_ctl_dir[] = {
5902 .procname = "sched_domain",
5903 .mode = 0555,
5905 {0, },
5908 static struct ctl_table sd_ctl_root[] = {
5910 .ctl_name = CTL_KERN,
5911 .procname = "kernel",
5912 .mode = 0555,
5913 .child = sd_ctl_dir,
5915 {0, },
5918 static struct ctl_table *sd_alloc_ctl_entry(int n)
5920 struct ctl_table *entry =
5921 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5923 return entry;
5926 static void sd_free_ctl_entry(struct ctl_table **tablep)
5928 struct ctl_table *entry;
5931 * In the intermediate directories, both the child directory and
5932 * procname are dynamically allocated and could fail but the mode
5933 * will always be set. In the lowest directory the names are
5934 * static strings and all have proc handlers.
5936 for (entry = *tablep; entry->mode; entry++) {
5937 if (entry->child)
5938 sd_free_ctl_entry(&entry->child);
5939 if (entry->proc_handler == NULL)
5940 kfree(entry->procname);
5943 kfree(*tablep);
5944 *tablep = NULL;
5947 static void
5948 set_table_entry(struct ctl_table *entry,
5949 const char *procname, void *data, int maxlen,
5950 mode_t mode, proc_handler *proc_handler)
5952 entry->procname = procname;
5953 entry->data = data;
5954 entry->maxlen = maxlen;
5955 entry->mode = mode;
5956 entry->proc_handler = proc_handler;
5959 static struct ctl_table *
5960 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5962 struct ctl_table *table = sd_alloc_ctl_entry(12);
5964 if (table == NULL)
5965 return NULL;
5967 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5968 sizeof(long), 0644, proc_doulongvec_minmax);
5969 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5970 sizeof(long), 0644, proc_doulongvec_minmax);
5971 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5972 sizeof(int), 0644, proc_dointvec_minmax);
5973 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5974 sizeof(int), 0644, proc_dointvec_minmax);
5975 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5976 sizeof(int), 0644, proc_dointvec_minmax);
5977 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5978 sizeof(int), 0644, proc_dointvec_minmax);
5979 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5980 sizeof(int), 0644, proc_dointvec_minmax);
5981 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5982 sizeof(int), 0644, proc_dointvec_minmax);
5983 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5984 sizeof(int), 0644, proc_dointvec_minmax);
5985 set_table_entry(&table[9], "cache_nice_tries",
5986 &sd->cache_nice_tries,
5987 sizeof(int), 0644, proc_dointvec_minmax);
5988 set_table_entry(&table[10], "flags", &sd->flags,
5989 sizeof(int), 0644, proc_dointvec_minmax);
5990 /* &table[11] is terminator */
5992 return table;
5995 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5997 struct ctl_table *entry, *table;
5998 struct sched_domain *sd;
5999 int domain_num = 0, i;
6000 char buf[32];
6002 for_each_domain(cpu, sd)
6003 domain_num++;
6004 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6005 if (table == NULL)
6006 return NULL;
6008 i = 0;
6009 for_each_domain(cpu, sd) {
6010 snprintf(buf, 32, "domain%d", i);
6011 entry->procname = kstrdup(buf, GFP_KERNEL);
6012 entry->mode = 0555;
6013 entry->child = sd_alloc_ctl_domain_table(sd);
6014 entry++;
6015 i++;
6017 return table;
6020 static struct ctl_table_header *sd_sysctl_header;
6021 static void register_sched_domain_sysctl(void)
6023 int i, cpu_num = num_online_cpus();
6024 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6025 char buf[32];
6027 WARN_ON(sd_ctl_dir[0].child);
6028 sd_ctl_dir[0].child = entry;
6030 if (entry == NULL)
6031 return;
6033 for_each_online_cpu(i) {
6034 snprintf(buf, 32, "cpu%d", i);
6035 entry->procname = kstrdup(buf, GFP_KERNEL);
6036 entry->mode = 0555;
6037 entry->child = sd_alloc_ctl_cpu_table(i);
6038 entry++;
6041 WARN_ON(sd_sysctl_header);
6042 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6045 /* may be called multiple times per register */
6046 static void unregister_sched_domain_sysctl(void)
6048 if (sd_sysctl_header)
6049 unregister_sysctl_table(sd_sysctl_header);
6050 sd_sysctl_header = NULL;
6051 if (sd_ctl_dir[0].child)
6052 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6054 #else
6055 static void register_sched_domain_sysctl(void)
6058 static void unregister_sched_domain_sysctl(void)
6061 #endif
6064 * migration_call - callback that gets triggered when a CPU is added.
6065 * Here we can start up the necessary migration thread for the new CPU.
6067 static int __cpuinit
6068 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6070 struct task_struct *p;
6071 int cpu = (long)hcpu;
6072 unsigned long flags;
6073 struct rq *rq;
6075 switch (action) {
6077 case CPU_UP_PREPARE:
6078 case CPU_UP_PREPARE_FROZEN:
6079 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6080 if (IS_ERR(p))
6081 return NOTIFY_BAD;
6082 kthread_bind(p, cpu);
6083 /* Must be high prio: stop_machine expects to yield to it. */
6084 rq = task_rq_lock(p, &flags);
6085 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6086 task_rq_unlock(rq, &flags);
6087 cpu_rq(cpu)->migration_thread = p;
6088 break;
6090 case CPU_ONLINE:
6091 case CPU_ONLINE_FROZEN:
6092 /* Strictly unnecessary, as first user will wake it. */
6093 wake_up_process(cpu_rq(cpu)->migration_thread);
6095 /* Update our root-domain */
6096 rq = cpu_rq(cpu);
6097 spin_lock_irqsave(&rq->lock, flags);
6098 if (rq->rd) {
6099 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6100 cpu_set(cpu, rq->rd->online);
6102 spin_unlock_irqrestore(&rq->lock, flags);
6103 break;
6105 #ifdef CONFIG_HOTPLUG_CPU
6106 case CPU_UP_CANCELED:
6107 case CPU_UP_CANCELED_FROZEN:
6108 if (!cpu_rq(cpu)->migration_thread)
6109 break;
6110 /* Unbind it from offline cpu so it can run. Fall thru. */
6111 kthread_bind(cpu_rq(cpu)->migration_thread,
6112 any_online_cpu(cpu_online_map));
6113 kthread_stop(cpu_rq(cpu)->migration_thread);
6114 cpu_rq(cpu)->migration_thread = NULL;
6115 break;
6117 case CPU_DEAD:
6118 case CPU_DEAD_FROZEN:
6119 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6120 migrate_live_tasks(cpu);
6121 rq = cpu_rq(cpu);
6122 kthread_stop(rq->migration_thread);
6123 rq->migration_thread = NULL;
6124 /* Idle task back to normal (off runqueue, low prio) */
6125 spin_lock_irq(&rq->lock);
6126 update_rq_clock(rq);
6127 deactivate_task(rq, rq->idle, 0);
6128 rq->idle->static_prio = MAX_PRIO;
6129 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6130 rq->idle->sched_class = &idle_sched_class;
6131 migrate_dead_tasks(cpu);
6132 spin_unlock_irq(&rq->lock);
6133 cpuset_unlock();
6134 migrate_nr_uninterruptible(rq);
6135 BUG_ON(rq->nr_running != 0);
6138 * No need to migrate the tasks: it was best-effort if
6139 * they didn't take sched_hotcpu_mutex. Just wake up
6140 * the requestors.
6142 spin_lock_irq(&rq->lock);
6143 while (!list_empty(&rq->migration_queue)) {
6144 struct migration_req *req;
6146 req = list_entry(rq->migration_queue.next,
6147 struct migration_req, list);
6148 list_del_init(&req->list);
6149 complete(&req->done);
6151 spin_unlock_irq(&rq->lock);
6152 break;
6154 case CPU_DYING:
6155 case CPU_DYING_FROZEN:
6156 /* Update our root-domain */
6157 rq = cpu_rq(cpu);
6158 spin_lock_irqsave(&rq->lock, flags);
6159 if (rq->rd) {
6160 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6161 cpu_clear(cpu, rq->rd->online);
6163 spin_unlock_irqrestore(&rq->lock, flags);
6164 break;
6165 #endif
6167 return NOTIFY_OK;
6170 /* Register at highest priority so that task migration (migrate_all_tasks)
6171 * happens before everything else.
6173 static struct notifier_block __cpuinitdata migration_notifier = {
6174 .notifier_call = migration_call,
6175 .priority = 10
6178 void __init migration_init(void)
6180 void *cpu = (void *)(long)smp_processor_id();
6181 int err;
6183 /* Start one for the boot CPU: */
6184 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6185 BUG_ON(err == NOTIFY_BAD);
6186 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6187 register_cpu_notifier(&migration_notifier);
6189 #endif
6191 #ifdef CONFIG_SMP
6193 #ifdef CONFIG_SCHED_DEBUG
6195 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6196 cpumask_t *groupmask)
6198 struct sched_group *group = sd->groups;
6199 char str[256];
6201 cpulist_scnprintf(str, sizeof(str), sd->span);
6202 cpus_clear(*groupmask);
6204 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6206 if (!(sd->flags & SD_LOAD_BALANCE)) {
6207 printk("does not load-balance\n");
6208 if (sd->parent)
6209 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6210 " has parent");
6211 return -1;
6214 printk(KERN_CONT "span %s\n", str);
6216 if (!cpu_isset(cpu, sd->span)) {
6217 printk(KERN_ERR "ERROR: domain->span does not contain "
6218 "CPU%d\n", cpu);
6220 if (!cpu_isset(cpu, group->cpumask)) {
6221 printk(KERN_ERR "ERROR: domain->groups does not contain"
6222 " CPU%d\n", cpu);
6225 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6226 do {
6227 if (!group) {
6228 printk("\n");
6229 printk(KERN_ERR "ERROR: group is NULL\n");
6230 break;
6233 if (!group->__cpu_power) {
6234 printk(KERN_CONT "\n");
6235 printk(KERN_ERR "ERROR: domain->cpu_power not "
6236 "set\n");
6237 break;
6240 if (!cpus_weight(group->cpumask)) {
6241 printk(KERN_CONT "\n");
6242 printk(KERN_ERR "ERROR: empty group\n");
6243 break;
6246 if (cpus_intersects(*groupmask, group->cpumask)) {
6247 printk(KERN_CONT "\n");
6248 printk(KERN_ERR "ERROR: repeated CPUs\n");
6249 break;
6252 cpus_or(*groupmask, *groupmask, group->cpumask);
6254 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6255 printk(KERN_CONT " %s", str);
6257 group = group->next;
6258 } while (group != sd->groups);
6259 printk(KERN_CONT "\n");
6261 if (!cpus_equal(sd->span, *groupmask))
6262 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6264 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6265 printk(KERN_ERR "ERROR: parent span is not a superset "
6266 "of domain->span\n");
6267 return 0;
6270 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6272 cpumask_t *groupmask;
6273 int level = 0;
6275 if (!sd) {
6276 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6277 return;
6280 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6282 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6283 if (!groupmask) {
6284 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6285 return;
6288 for (;;) {
6289 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6290 break;
6291 level++;
6292 sd = sd->parent;
6293 if (!sd)
6294 break;
6296 kfree(groupmask);
6298 #else
6299 # define sched_domain_debug(sd, cpu) do { } while (0)
6300 #endif
6302 static int sd_degenerate(struct sched_domain *sd)
6304 if (cpus_weight(sd->span) == 1)
6305 return 1;
6307 /* Following flags need at least 2 groups */
6308 if (sd->flags & (SD_LOAD_BALANCE |
6309 SD_BALANCE_NEWIDLE |
6310 SD_BALANCE_FORK |
6311 SD_BALANCE_EXEC |
6312 SD_SHARE_CPUPOWER |
6313 SD_SHARE_PKG_RESOURCES)) {
6314 if (sd->groups != sd->groups->next)
6315 return 0;
6318 /* Following flags don't use groups */
6319 if (sd->flags & (SD_WAKE_IDLE |
6320 SD_WAKE_AFFINE |
6321 SD_WAKE_BALANCE))
6322 return 0;
6324 return 1;
6327 static int
6328 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6330 unsigned long cflags = sd->flags, pflags = parent->flags;
6332 if (sd_degenerate(parent))
6333 return 1;
6335 if (!cpus_equal(sd->span, parent->span))
6336 return 0;
6338 /* Does parent contain flags not in child? */
6339 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6340 if (cflags & SD_WAKE_AFFINE)
6341 pflags &= ~SD_WAKE_BALANCE;
6342 /* Flags needing groups don't count if only 1 group in parent */
6343 if (parent->groups == parent->groups->next) {
6344 pflags &= ~(SD_LOAD_BALANCE |
6345 SD_BALANCE_NEWIDLE |
6346 SD_BALANCE_FORK |
6347 SD_BALANCE_EXEC |
6348 SD_SHARE_CPUPOWER |
6349 SD_SHARE_PKG_RESOURCES);
6351 if (~cflags & pflags)
6352 return 0;
6354 return 1;
6357 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6359 unsigned long flags;
6360 const struct sched_class *class;
6362 spin_lock_irqsave(&rq->lock, flags);
6364 if (rq->rd) {
6365 struct root_domain *old_rd = rq->rd;
6367 for (class = sched_class_highest; class; class = class->next) {
6368 if (class->leave_domain)
6369 class->leave_domain(rq);
6372 cpu_clear(rq->cpu, old_rd->span);
6373 cpu_clear(rq->cpu, old_rd->online);
6375 if (atomic_dec_and_test(&old_rd->refcount))
6376 kfree(old_rd);
6379 atomic_inc(&rd->refcount);
6380 rq->rd = rd;
6382 cpu_set(rq->cpu, rd->span);
6383 if (cpu_isset(rq->cpu, cpu_online_map))
6384 cpu_set(rq->cpu, rd->online);
6386 for (class = sched_class_highest; class; class = class->next) {
6387 if (class->join_domain)
6388 class->join_domain(rq);
6391 spin_unlock_irqrestore(&rq->lock, flags);
6394 static void init_rootdomain(struct root_domain *rd)
6396 memset(rd, 0, sizeof(*rd));
6398 cpus_clear(rd->span);
6399 cpus_clear(rd->online);
6402 static void init_defrootdomain(void)
6404 init_rootdomain(&def_root_domain);
6405 atomic_set(&def_root_domain.refcount, 1);
6408 static struct root_domain *alloc_rootdomain(void)
6410 struct root_domain *rd;
6412 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6413 if (!rd)
6414 return NULL;
6416 init_rootdomain(rd);
6418 return rd;
6422 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6423 * hold the hotplug lock.
6425 static void
6426 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6428 struct rq *rq = cpu_rq(cpu);
6429 struct sched_domain *tmp;
6431 /* Remove the sched domains which do not contribute to scheduling. */
6432 for (tmp = sd; tmp; tmp = tmp->parent) {
6433 struct sched_domain *parent = tmp->parent;
6434 if (!parent)
6435 break;
6436 if (sd_parent_degenerate(tmp, parent)) {
6437 tmp->parent = parent->parent;
6438 if (parent->parent)
6439 parent->parent->child = tmp;
6443 if (sd && sd_degenerate(sd)) {
6444 sd = sd->parent;
6445 if (sd)
6446 sd->child = NULL;
6449 sched_domain_debug(sd, cpu);
6451 rq_attach_root(rq, rd);
6452 rcu_assign_pointer(rq->sd, sd);
6455 /* cpus with isolated domains */
6456 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6458 /* Setup the mask of cpus configured for isolated domains */
6459 static int __init isolated_cpu_setup(char *str)
6461 int ints[NR_CPUS], i;
6463 str = get_options(str, ARRAY_SIZE(ints), ints);
6464 cpus_clear(cpu_isolated_map);
6465 for (i = 1; i <= ints[0]; i++)
6466 if (ints[i] < NR_CPUS)
6467 cpu_set(ints[i], cpu_isolated_map);
6468 return 1;
6471 __setup("isolcpus=", isolated_cpu_setup);
6474 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6475 * to a function which identifies what group(along with sched group) a CPU
6476 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6477 * (due to the fact that we keep track of groups covered with a cpumask_t).
6479 * init_sched_build_groups will build a circular linked list of the groups
6480 * covered by the given span, and will set each group's ->cpumask correctly,
6481 * and ->cpu_power to 0.
6483 static void
6484 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6485 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6486 struct sched_group **sg,
6487 cpumask_t *tmpmask),
6488 cpumask_t *covered, cpumask_t *tmpmask)
6490 struct sched_group *first = NULL, *last = NULL;
6491 int i;
6493 cpus_clear(*covered);
6495 for_each_cpu_mask(i, *span) {
6496 struct sched_group *sg;
6497 int group = group_fn(i, cpu_map, &sg, tmpmask);
6498 int j;
6500 if (cpu_isset(i, *covered))
6501 continue;
6503 cpus_clear(sg->cpumask);
6504 sg->__cpu_power = 0;
6506 for_each_cpu_mask(j, *span) {
6507 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6508 continue;
6510 cpu_set(j, *covered);
6511 cpu_set(j, sg->cpumask);
6513 if (!first)
6514 first = sg;
6515 if (last)
6516 last->next = sg;
6517 last = sg;
6519 last->next = first;
6522 #define SD_NODES_PER_DOMAIN 16
6524 #ifdef CONFIG_NUMA
6527 * find_next_best_node - find the next node to include in a sched_domain
6528 * @node: node whose sched_domain we're building
6529 * @used_nodes: nodes already in the sched_domain
6531 * Find the next node to include in a given scheduling domain. Simply
6532 * finds the closest node not already in the @used_nodes map.
6534 * Should use nodemask_t.
6536 static int find_next_best_node(int node, nodemask_t *used_nodes)
6538 int i, n, val, min_val, best_node = 0;
6540 min_val = INT_MAX;
6542 for (i = 0; i < MAX_NUMNODES; i++) {
6543 /* Start at @node */
6544 n = (node + i) % MAX_NUMNODES;
6546 if (!nr_cpus_node(n))
6547 continue;
6549 /* Skip already used nodes */
6550 if (node_isset(n, *used_nodes))
6551 continue;
6553 /* Simple min distance search */
6554 val = node_distance(node, n);
6556 if (val < min_val) {
6557 min_val = val;
6558 best_node = n;
6562 node_set(best_node, *used_nodes);
6563 return best_node;
6567 * sched_domain_node_span - get a cpumask for a node's sched_domain
6568 * @node: node whose cpumask we're constructing
6569 * @span: resulting cpumask
6571 * Given a node, construct a good cpumask for its sched_domain to span. It
6572 * should be one that prevents unnecessary balancing, but also spreads tasks
6573 * out optimally.
6575 static void sched_domain_node_span(int node, cpumask_t *span)
6577 nodemask_t used_nodes;
6578 node_to_cpumask_ptr(nodemask, node);
6579 int i;
6581 cpus_clear(*span);
6582 nodes_clear(used_nodes);
6584 cpus_or(*span, *span, *nodemask);
6585 node_set(node, used_nodes);
6587 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6588 int next_node = find_next_best_node(node, &used_nodes);
6590 node_to_cpumask_ptr_next(nodemask, next_node);
6591 cpus_or(*span, *span, *nodemask);
6594 #endif
6596 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6599 * SMT sched-domains:
6601 #ifdef CONFIG_SCHED_SMT
6602 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6603 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6605 static int
6606 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6607 cpumask_t *unused)
6609 if (sg)
6610 *sg = &per_cpu(sched_group_cpus, cpu);
6611 return cpu;
6613 #endif
6616 * multi-core sched-domains:
6618 #ifdef CONFIG_SCHED_MC
6619 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6620 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6621 #endif
6623 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6624 static int
6625 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6626 cpumask_t *mask)
6628 int group;
6630 *mask = per_cpu(cpu_sibling_map, cpu);
6631 cpus_and(*mask, *mask, *cpu_map);
6632 group = first_cpu(*mask);
6633 if (sg)
6634 *sg = &per_cpu(sched_group_core, group);
6635 return group;
6637 #elif defined(CONFIG_SCHED_MC)
6638 static int
6639 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6640 cpumask_t *unused)
6642 if (sg)
6643 *sg = &per_cpu(sched_group_core, cpu);
6644 return cpu;
6646 #endif
6648 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6649 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6651 static int
6652 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6653 cpumask_t *mask)
6655 int group;
6656 #ifdef CONFIG_SCHED_MC
6657 *mask = cpu_coregroup_map(cpu);
6658 cpus_and(*mask, *mask, *cpu_map);
6659 group = first_cpu(*mask);
6660 #elif defined(CONFIG_SCHED_SMT)
6661 *mask = per_cpu(cpu_sibling_map, cpu);
6662 cpus_and(*mask, *mask, *cpu_map);
6663 group = first_cpu(*mask);
6664 #else
6665 group = cpu;
6666 #endif
6667 if (sg)
6668 *sg = &per_cpu(sched_group_phys, group);
6669 return group;
6672 #ifdef CONFIG_NUMA
6674 * The init_sched_build_groups can't handle what we want to do with node
6675 * groups, so roll our own. Now each node has its own list of groups which
6676 * gets dynamically allocated.
6678 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6679 static struct sched_group ***sched_group_nodes_bycpu;
6681 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6682 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6684 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6685 struct sched_group **sg, cpumask_t *nodemask)
6687 int group;
6689 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6690 cpus_and(*nodemask, *nodemask, *cpu_map);
6691 group = first_cpu(*nodemask);
6693 if (sg)
6694 *sg = &per_cpu(sched_group_allnodes, group);
6695 return group;
6698 static void init_numa_sched_groups_power(struct sched_group *group_head)
6700 struct sched_group *sg = group_head;
6701 int j;
6703 if (!sg)
6704 return;
6705 do {
6706 for_each_cpu_mask(j, sg->cpumask) {
6707 struct sched_domain *sd;
6709 sd = &per_cpu(phys_domains, j);
6710 if (j != first_cpu(sd->groups->cpumask)) {
6712 * Only add "power" once for each
6713 * physical package.
6715 continue;
6718 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6720 sg = sg->next;
6721 } while (sg != group_head);
6723 #endif
6725 #ifdef CONFIG_NUMA
6726 /* Free memory allocated for various sched_group structures */
6727 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6729 int cpu, i;
6731 for_each_cpu_mask(cpu, *cpu_map) {
6732 struct sched_group **sched_group_nodes
6733 = sched_group_nodes_bycpu[cpu];
6735 if (!sched_group_nodes)
6736 continue;
6738 for (i = 0; i < MAX_NUMNODES; i++) {
6739 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6741 *nodemask = node_to_cpumask(i);
6742 cpus_and(*nodemask, *nodemask, *cpu_map);
6743 if (cpus_empty(*nodemask))
6744 continue;
6746 if (sg == NULL)
6747 continue;
6748 sg = sg->next;
6749 next_sg:
6750 oldsg = sg;
6751 sg = sg->next;
6752 kfree(oldsg);
6753 if (oldsg != sched_group_nodes[i])
6754 goto next_sg;
6756 kfree(sched_group_nodes);
6757 sched_group_nodes_bycpu[cpu] = NULL;
6760 #else
6761 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6764 #endif
6767 * Initialize sched groups cpu_power.
6769 * cpu_power indicates the capacity of sched group, which is used while
6770 * distributing the load between different sched groups in a sched domain.
6771 * Typically cpu_power for all the groups in a sched domain will be same unless
6772 * there are asymmetries in the topology. If there are asymmetries, group
6773 * having more cpu_power will pickup more load compared to the group having
6774 * less cpu_power.
6776 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6777 * the maximum number of tasks a group can handle in the presence of other idle
6778 * or lightly loaded groups in the same sched domain.
6780 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6782 struct sched_domain *child;
6783 struct sched_group *group;
6785 WARN_ON(!sd || !sd->groups);
6787 if (cpu != first_cpu(sd->groups->cpumask))
6788 return;
6790 child = sd->child;
6792 sd->groups->__cpu_power = 0;
6795 * For perf policy, if the groups in child domain share resources
6796 * (for example cores sharing some portions of the cache hierarchy
6797 * or SMT), then set this domain groups cpu_power such that each group
6798 * can handle only one task, when there are other idle groups in the
6799 * same sched domain.
6801 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6802 (child->flags &
6803 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6804 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6805 return;
6809 * add cpu_power of each child group to this groups cpu_power
6811 group = child->groups;
6812 do {
6813 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6814 group = group->next;
6815 } while (group != child->groups);
6819 * Initializers for schedule domains
6820 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6823 #define SD_INIT(sd, type) sd_init_##type(sd)
6824 #define SD_INIT_FUNC(type) \
6825 static noinline void sd_init_##type(struct sched_domain *sd) \
6827 memset(sd, 0, sizeof(*sd)); \
6828 *sd = SD_##type##_INIT; \
6829 sd->level = SD_LV_##type; \
6832 SD_INIT_FUNC(CPU)
6833 #ifdef CONFIG_NUMA
6834 SD_INIT_FUNC(ALLNODES)
6835 SD_INIT_FUNC(NODE)
6836 #endif
6837 #ifdef CONFIG_SCHED_SMT
6838 SD_INIT_FUNC(SIBLING)
6839 #endif
6840 #ifdef CONFIG_SCHED_MC
6841 SD_INIT_FUNC(MC)
6842 #endif
6845 * To minimize stack usage kmalloc room for cpumasks and share the
6846 * space as the usage in build_sched_domains() dictates. Used only
6847 * if the amount of space is significant.
6849 struct allmasks {
6850 cpumask_t tmpmask; /* make this one first */
6851 union {
6852 cpumask_t nodemask;
6853 cpumask_t this_sibling_map;
6854 cpumask_t this_core_map;
6856 cpumask_t send_covered;
6858 #ifdef CONFIG_NUMA
6859 cpumask_t domainspan;
6860 cpumask_t covered;
6861 cpumask_t notcovered;
6862 #endif
6865 #if NR_CPUS > 128
6866 #define SCHED_CPUMASK_ALLOC 1
6867 #define SCHED_CPUMASK_FREE(v) kfree(v)
6868 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
6869 #else
6870 #define SCHED_CPUMASK_ALLOC 0
6871 #define SCHED_CPUMASK_FREE(v)
6872 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
6873 #endif
6875 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
6876 ((unsigned long)(a) + offsetof(struct allmasks, v))
6878 static int default_relax_domain_level = -1;
6880 static int __init setup_relax_domain_level(char *str)
6882 unsigned long val;
6884 val = simple_strtoul(str, NULL, 0);
6885 if (val < SD_LV_MAX)
6886 default_relax_domain_level = val;
6888 return 1;
6890 __setup("relax_domain_level=", setup_relax_domain_level);
6892 static void set_domain_attribute(struct sched_domain *sd,
6893 struct sched_domain_attr *attr)
6895 int request;
6897 if (!attr || attr->relax_domain_level < 0) {
6898 if (default_relax_domain_level < 0)
6899 return;
6900 else
6901 request = default_relax_domain_level;
6902 } else
6903 request = attr->relax_domain_level;
6904 if (request < sd->level) {
6905 /* turn off idle balance on this domain */
6906 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
6907 } else {
6908 /* turn on idle balance on this domain */
6909 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
6914 * Build sched domains for a given set of cpus and attach the sched domains
6915 * to the individual cpus
6917 static int __build_sched_domains(const cpumask_t *cpu_map,
6918 struct sched_domain_attr *attr)
6920 int i;
6921 struct root_domain *rd;
6922 SCHED_CPUMASK_DECLARE(allmasks);
6923 cpumask_t *tmpmask;
6924 #ifdef CONFIG_NUMA
6925 struct sched_group **sched_group_nodes = NULL;
6926 int sd_allnodes = 0;
6929 * Allocate the per-node list of sched groups
6931 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6932 GFP_KERNEL);
6933 if (!sched_group_nodes) {
6934 printk(KERN_WARNING "Can not alloc sched group node list\n");
6935 return -ENOMEM;
6937 #endif
6939 rd = alloc_rootdomain();
6940 if (!rd) {
6941 printk(KERN_WARNING "Cannot alloc root domain\n");
6942 #ifdef CONFIG_NUMA
6943 kfree(sched_group_nodes);
6944 #endif
6945 return -ENOMEM;
6948 #if SCHED_CPUMASK_ALLOC
6949 /* get space for all scratch cpumask variables */
6950 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
6951 if (!allmasks) {
6952 printk(KERN_WARNING "Cannot alloc cpumask array\n");
6953 kfree(rd);
6954 #ifdef CONFIG_NUMA
6955 kfree(sched_group_nodes);
6956 #endif
6957 return -ENOMEM;
6959 #endif
6960 tmpmask = (cpumask_t *)allmasks;
6963 #ifdef CONFIG_NUMA
6964 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6965 #endif
6968 * Set up domains for cpus specified by the cpu_map.
6970 for_each_cpu_mask(i, *cpu_map) {
6971 struct sched_domain *sd = NULL, *p;
6972 SCHED_CPUMASK_VAR(nodemask, allmasks);
6974 *nodemask = node_to_cpumask(cpu_to_node(i));
6975 cpus_and(*nodemask, *nodemask, *cpu_map);
6977 #ifdef CONFIG_NUMA
6978 if (cpus_weight(*cpu_map) >
6979 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
6980 sd = &per_cpu(allnodes_domains, i);
6981 SD_INIT(sd, ALLNODES);
6982 set_domain_attribute(sd, attr);
6983 sd->span = *cpu_map;
6984 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
6985 p = sd;
6986 sd_allnodes = 1;
6987 } else
6988 p = NULL;
6990 sd = &per_cpu(node_domains, i);
6991 SD_INIT(sd, NODE);
6992 set_domain_attribute(sd, attr);
6993 sched_domain_node_span(cpu_to_node(i), &sd->span);
6994 sd->parent = p;
6995 if (p)
6996 p->child = sd;
6997 cpus_and(sd->span, sd->span, *cpu_map);
6998 #endif
7000 p = sd;
7001 sd = &per_cpu(phys_domains, i);
7002 SD_INIT(sd, CPU);
7003 set_domain_attribute(sd, attr);
7004 sd->span = *nodemask;
7005 sd->parent = p;
7006 if (p)
7007 p->child = sd;
7008 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7010 #ifdef CONFIG_SCHED_MC
7011 p = sd;
7012 sd = &per_cpu(core_domains, i);
7013 SD_INIT(sd, MC);
7014 set_domain_attribute(sd, attr);
7015 sd->span = cpu_coregroup_map(i);
7016 cpus_and(sd->span, sd->span, *cpu_map);
7017 sd->parent = p;
7018 p->child = sd;
7019 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7020 #endif
7022 #ifdef CONFIG_SCHED_SMT
7023 p = sd;
7024 sd = &per_cpu(cpu_domains, i);
7025 SD_INIT(sd, SIBLING);
7026 set_domain_attribute(sd, attr);
7027 sd->span = per_cpu(cpu_sibling_map, i);
7028 cpus_and(sd->span, sd->span, *cpu_map);
7029 sd->parent = p;
7030 p->child = sd;
7031 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7032 #endif
7035 #ifdef CONFIG_SCHED_SMT
7036 /* Set up CPU (sibling) groups */
7037 for_each_cpu_mask(i, *cpu_map) {
7038 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7039 SCHED_CPUMASK_VAR(send_covered, allmasks);
7041 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7042 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7043 if (i != first_cpu(*this_sibling_map))
7044 continue;
7046 init_sched_build_groups(this_sibling_map, cpu_map,
7047 &cpu_to_cpu_group,
7048 send_covered, tmpmask);
7050 #endif
7052 #ifdef CONFIG_SCHED_MC
7053 /* Set up multi-core groups */
7054 for_each_cpu_mask(i, *cpu_map) {
7055 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7056 SCHED_CPUMASK_VAR(send_covered, allmasks);
7058 *this_core_map = cpu_coregroup_map(i);
7059 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7060 if (i != first_cpu(*this_core_map))
7061 continue;
7063 init_sched_build_groups(this_core_map, cpu_map,
7064 &cpu_to_core_group,
7065 send_covered, tmpmask);
7067 #endif
7069 /* Set up physical groups */
7070 for (i = 0; i < MAX_NUMNODES; i++) {
7071 SCHED_CPUMASK_VAR(nodemask, allmasks);
7072 SCHED_CPUMASK_VAR(send_covered, allmasks);
7074 *nodemask = node_to_cpumask(i);
7075 cpus_and(*nodemask, *nodemask, *cpu_map);
7076 if (cpus_empty(*nodemask))
7077 continue;
7079 init_sched_build_groups(nodemask, cpu_map,
7080 &cpu_to_phys_group,
7081 send_covered, tmpmask);
7084 #ifdef CONFIG_NUMA
7085 /* Set up node groups */
7086 if (sd_allnodes) {
7087 SCHED_CPUMASK_VAR(send_covered, allmasks);
7089 init_sched_build_groups(cpu_map, cpu_map,
7090 &cpu_to_allnodes_group,
7091 send_covered, tmpmask);
7094 for (i = 0; i < MAX_NUMNODES; i++) {
7095 /* Set up node groups */
7096 struct sched_group *sg, *prev;
7097 SCHED_CPUMASK_VAR(nodemask, allmasks);
7098 SCHED_CPUMASK_VAR(domainspan, allmasks);
7099 SCHED_CPUMASK_VAR(covered, allmasks);
7100 int j;
7102 *nodemask = node_to_cpumask(i);
7103 cpus_clear(*covered);
7105 cpus_and(*nodemask, *nodemask, *cpu_map);
7106 if (cpus_empty(*nodemask)) {
7107 sched_group_nodes[i] = NULL;
7108 continue;
7111 sched_domain_node_span(i, domainspan);
7112 cpus_and(*domainspan, *domainspan, *cpu_map);
7114 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7115 if (!sg) {
7116 printk(KERN_WARNING "Can not alloc domain group for "
7117 "node %d\n", i);
7118 goto error;
7120 sched_group_nodes[i] = sg;
7121 for_each_cpu_mask(j, *nodemask) {
7122 struct sched_domain *sd;
7124 sd = &per_cpu(node_domains, j);
7125 sd->groups = sg;
7127 sg->__cpu_power = 0;
7128 sg->cpumask = *nodemask;
7129 sg->next = sg;
7130 cpus_or(*covered, *covered, *nodemask);
7131 prev = sg;
7133 for (j = 0; j < MAX_NUMNODES; j++) {
7134 SCHED_CPUMASK_VAR(notcovered, allmasks);
7135 int n = (i + j) % MAX_NUMNODES;
7136 node_to_cpumask_ptr(pnodemask, n);
7138 cpus_complement(*notcovered, *covered);
7139 cpus_and(*tmpmask, *notcovered, *cpu_map);
7140 cpus_and(*tmpmask, *tmpmask, *domainspan);
7141 if (cpus_empty(*tmpmask))
7142 break;
7144 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7145 if (cpus_empty(*tmpmask))
7146 continue;
7148 sg = kmalloc_node(sizeof(struct sched_group),
7149 GFP_KERNEL, i);
7150 if (!sg) {
7151 printk(KERN_WARNING
7152 "Can not alloc domain group for node %d\n", j);
7153 goto error;
7155 sg->__cpu_power = 0;
7156 sg->cpumask = *tmpmask;
7157 sg->next = prev->next;
7158 cpus_or(*covered, *covered, *tmpmask);
7159 prev->next = sg;
7160 prev = sg;
7163 #endif
7165 /* Calculate CPU power for physical packages and nodes */
7166 #ifdef CONFIG_SCHED_SMT
7167 for_each_cpu_mask(i, *cpu_map) {
7168 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7170 init_sched_groups_power(i, sd);
7172 #endif
7173 #ifdef CONFIG_SCHED_MC
7174 for_each_cpu_mask(i, *cpu_map) {
7175 struct sched_domain *sd = &per_cpu(core_domains, i);
7177 init_sched_groups_power(i, sd);
7179 #endif
7181 for_each_cpu_mask(i, *cpu_map) {
7182 struct sched_domain *sd = &per_cpu(phys_domains, i);
7184 init_sched_groups_power(i, sd);
7187 #ifdef CONFIG_NUMA
7188 for (i = 0; i < MAX_NUMNODES; i++)
7189 init_numa_sched_groups_power(sched_group_nodes[i]);
7191 if (sd_allnodes) {
7192 struct sched_group *sg;
7194 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7195 tmpmask);
7196 init_numa_sched_groups_power(sg);
7198 #endif
7200 /* Attach the domains */
7201 for_each_cpu_mask(i, *cpu_map) {
7202 struct sched_domain *sd;
7203 #ifdef CONFIG_SCHED_SMT
7204 sd = &per_cpu(cpu_domains, i);
7205 #elif defined(CONFIG_SCHED_MC)
7206 sd = &per_cpu(core_domains, i);
7207 #else
7208 sd = &per_cpu(phys_domains, i);
7209 #endif
7210 cpu_attach_domain(sd, rd, i);
7213 SCHED_CPUMASK_FREE((void *)allmasks);
7214 return 0;
7216 #ifdef CONFIG_NUMA
7217 error:
7218 free_sched_groups(cpu_map, tmpmask);
7219 SCHED_CPUMASK_FREE((void *)allmasks);
7220 return -ENOMEM;
7221 #endif
7224 static int build_sched_domains(const cpumask_t *cpu_map)
7226 return __build_sched_domains(cpu_map, NULL);
7229 static cpumask_t *doms_cur; /* current sched domains */
7230 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7231 static struct sched_domain_attr *dattr_cur;
7232 /* attribues of custom domains in 'doms_cur' */
7235 * Special case: If a kmalloc of a doms_cur partition (array of
7236 * cpumask_t) fails, then fallback to a single sched domain,
7237 * as determined by the single cpumask_t fallback_doms.
7239 static cpumask_t fallback_doms;
7241 void __attribute__((weak)) arch_update_cpu_topology(void)
7246 * Free current domain masks.
7247 * Called after all cpus are attached to NULL domain.
7249 static void free_sched_domains(void)
7251 ndoms_cur = 0;
7252 if (doms_cur != &fallback_doms)
7253 kfree(doms_cur);
7254 doms_cur = &fallback_doms;
7258 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7259 * For now this just excludes isolated cpus, but could be used to
7260 * exclude other special cases in the future.
7262 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7264 int err;
7266 arch_update_cpu_topology();
7267 ndoms_cur = 1;
7268 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7269 if (!doms_cur)
7270 doms_cur = &fallback_doms;
7271 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7272 dattr_cur = NULL;
7273 err = build_sched_domains(doms_cur);
7274 register_sched_domain_sysctl();
7276 return err;
7279 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7280 cpumask_t *tmpmask)
7282 free_sched_groups(cpu_map, tmpmask);
7286 * Detach sched domains from a group of cpus specified in cpu_map
7287 * These cpus will now be attached to the NULL domain
7289 static void detach_destroy_domains(const cpumask_t *cpu_map)
7291 cpumask_t tmpmask;
7292 int i;
7294 unregister_sched_domain_sysctl();
7296 for_each_cpu_mask(i, *cpu_map)
7297 cpu_attach_domain(NULL, &def_root_domain, i);
7298 synchronize_sched();
7299 arch_destroy_sched_domains(cpu_map, &tmpmask);
7302 /* handle null as "default" */
7303 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7304 struct sched_domain_attr *new, int idx_new)
7306 struct sched_domain_attr tmp;
7308 /* fast path */
7309 if (!new && !cur)
7310 return 1;
7312 tmp = SD_ATTR_INIT;
7313 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7314 new ? (new + idx_new) : &tmp,
7315 sizeof(struct sched_domain_attr));
7319 * Partition sched domains as specified by the 'ndoms_new'
7320 * cpumasks in the array doms_new[] of cpumasks. This compares
7321 * doms_new[] to the current sched domain partitioning, doms_cur[].
7322 * It destroys each deleted domain and builds each new domain.
7324 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7325 * The masks don't intersect (don't overlap.) We should setup one
7326 * sched domain for each mask. CPUs not in any of the cpumasks will
7327 * not be load balanced. If the same cpumask appears both in the
7328 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7329 * it as it is.
7331 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7332 * ownership of it and will kfree it when done with it. If the caller
7333 * failed the kmalloc call, then it can pass in doms_new == NULL,
7334 * and partition_sched_domains() will fallback to the single partition
7335 * 'fallback_doms'.
7337 * Call with hotplug lock held
7339 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7340 struct sched_domain_attr *dattr_new)
7342 int i, j;
7344 mutex_lock(&sched_domains_mutex);
7346 /* always unregister in case we don't destroy any domains */
7347 unregister_sched_domain_sysctl();
7349 if (doms_new == NULL) {
7350 ndoms_new = 1;
7351 doms_new = &fallback_doms;
7352 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7353 dattr_new = NULL;
7356 /* Destroy deleted domains */
7357 for (i = 0; i < ndoms_cur; i++) {
7358 for (j = 0; j < ndoms_new; j++) {
7359 if (cpus_equal(doms_cur[i], doms_new[j])
7360 && dattrs_equal(dattr_cur, i, dattr_new, j))
7361 goto match1;
7363 /* no match - a current sched domain not in new doms_new[] */
7364 detach_destroy_domains(doms_cur + i);
7365 match1:
7369 /* Build new domains */
7370 for (i = 0; i < ndoms_new; i++) {
7371 for (j = 0; j < ndoms_cur; j++) {
7372 if (cpus_equal(doms_new[i], doms_cur[j])
7373 && dattrs_equal(dattr_new, i, dattr_cur, j))
7374 goto match2;
7376 /* no match - add a new doms_new */
7377 __build_sched_domains(doms_new + i,
7378 dattr_new ? dattr_new + i : NULL);
7379 match2:
7383 /* Remember the new sched domains */
7384 if (doms_cur != &fallback_doms)
7385 kfree(doms_cur);
7386 kfree(dattr_cur); /* kfree(NULL) is safe */
7387 doms_cur = doms_new;
7388 dattr_cur = dattr_new;
7389 ndoms_cur = ndoms_new;
7391 register_sched_domain_sysctl();
7393 mutex_unlock(&sched_domains_mutex);
7396 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7397 int arch_reinit_sched_domains(void)
7399 int err;
7401 get_online_cpus();
7402 mutex_lock(&sched_domains_mutex);
7403 detach_destroy_domains(&cpu_online_map);
7404 free_sched_domains();
7405 err = arch_init_sched_domains(&cpu_online_map);
7406 mutex_unlock(&sched_domains_mutex);
7407 put_online_cpus();
7409 return err;
7412 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7414 int ret;
7416 if (buf[0] != '0' && buf[0] != '1')
7417 return -EINVAL;
7419 if (smt)
7420 sched_smt_power_savings = (buf[0] == '1');
7421 else
7422 sched_mc_power_savings = (buf[0] == '1');
7424 ret = arch_reinit_sched_domains();
7426 return ret ? ret : count;
7429 #ifdef CONFIG_SCHED_MC
7430 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7432 return sprintf(page, "%u\n", sched_mc_power_savings);
7434 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7435 const char *buf, size_t count)
7437 return sched_power_savings_store(buf, count, 0);
7439 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7440 sched_mc_power_savings_store);
7441 #endif
7443 #ifdef CONFIG_SCHED_SMT
7444 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7446 return sprintf(page, "%u\n", sched_smt_power_savings);
7448 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7449 const char *buf, size_t count)
7451 return sched_power_savings_store(buf, count, 1);
7453 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7454 sched_smt_power_savings_store);
7455 #endif
7457 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7459 int err = 0;
7461 #ifdef CONFIG_SCHED_SMT
7462 if (smt_capable())
7463 err = sysfs_create_file(&cls->kset.kobj,
7464 &attr_sched_smt_power_savings.attr);
7465 #endif
7466 #ifdef CONFIG_SCHED_MC
7467 if (!err && mc_capable())
7468 err = sysfs_create_file(&cls->kset.kobj,
7469 &attr_sched_mc_power_savings.attr);
7470 #endif
7471 return err;
7473 #endif
7476 * Force a reinitialization of the sched domains hierarchy. The domains
7477 * and groups cannot be updated in place without racing with the balancing
7478 * code, so we temporarily attach all running cpus to the NULL domain
7479 * which will prevent rebalancing while the sched domains are recalculated.
7481 static int update_sched_domains(struct notifier_block *nfb,
7482 unsigned long action, void *hcpu)
7484 switch (action) {
7485 case CPU_UP_PREPARE:
7486 case CPU_UP_PREPARE_FROZEN:
7487 case CPU_DOWN_PREPARE:
7488 case CPU_DOWN_PREPARE_FROZEN:
7489 detach_destroy_domains(&cpu_online_map);
7490 free_sched_domains();
7491 return NOTIFY_OK;
7493 case CPU_UP_CANCELED:
7494 case CPU_UP_CANCELED_FROZEN:
7495 case CPU_DOWN_FAILED:
7496 case CPU_DOWN_FAILED_FROZEN:
7497 case CPU_ONLINE:
7498 case CPU_ONLINE_FROZEN:
7499 case CPU_DEAD:
7500 case CPU_DEAD_FROZEN:
7502 * Fall through and re-initialise the domains.
7504 break;
7505 default:
7506 return NOTIFY_DONE;
7509 #ifndef CONFIG_CPUSETS
7511 * Create default domain partitioning if cpusets are disabled.
7512 * Otherwise we let cpusets rebuild the domains based on the
7513 * current setup.
7516 /* The hotplug lock is already held by cpu_up/cpu_down */
7517 arch_init_sched_domains(&cpu_online_map);
7518 #endif
7520 return NOTIFY_OK;
7523 void __init sched_init_smp(void)
7525 cpumask_t non_isolated_cpus;
7527 #if defined(CONFIG_NUMA)
7528 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7529 GFP_KERNEL);
7530 BUG_ON(sched_group_nodes_bycpu == NULL);
7531 #endif
7532 get_online_cpus();
7533 mutex_lock(&sched_domains_mutex);
7534 arch_init_sched_domains(&cpu_online_map);
7535 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7536 if (cpus_empty(non_isolated_cpus))
7537 cpu_set(smp_processor_id(), non_isolated_cpus);
7538 mutex_unlock(&sched_domains_mutex);
7539 put_online_cpus();
7540 /* XXX: Theoretical race here - CPU may be hotplugged now */
7541 hotcpu_notifier(update_sched_domains, 0);
7542 init_hrtick();
7544 /* Move init over to a non-isolated CPU */
7545 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7546 BUG();
7547 sched_init_granularity();
7549 #else
7550 void __init sched_init_smp(void)
7552 sched_init_granularity();
7554 #endif /* CONFIG_SMP */
7556 int in_sched_functions(unsigned long addr)
7558 return in_lock_functions(addr) ||
7559 (addr >= (unsigned long)__sched_text_start
7560 && addr < (unsigned long)__sched_text_end);
7563 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7565 cfs_rq->tasks_timeline = RB_ROOT;
7566 INIT_LIST_HEAD(&cfs_rq->tasks);
7567 #ifdef CONFIG_FAIR_GROUP_SCHED
7568 cfs_rq->rq = rq;
7569 #endif
7570 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7573 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7575 struct rt_prio_array *array;
7576 int i;
7578 array = &rt_rq->active;
7579 for (i = 0; i < MAX_RT_PRIO; i++) {
7580 INIT_LIST_HEAD(array->queue + i);
7581 __clear_bit(i, array->bitmap);
7583 /* delimiter for bitsearch: */
7584 __set_bit(MAX_RT_PRIO, array->bitmap);
7586 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7587 rt_rq->highest_prio = MAX_RT_PRIO;
7588 #endif
7589 #ifdef CONFIG_SMP
7590 rt_rq->rt_nr_migratory = 0;
7591 rt_rq->overloaded = 0;
7592 #endif
7594 rt_rq->rt_time = 0;
7595 rt_rq->rt_throttled = 0;
7596 rt_rq->rt_runtime = 0;
7597 spin_lock_init(&rt_rq->rt_runtime_lock);
7599 #ifdef CONFIG_RT_GROUP_SCHED
7600 rt_rq->rt_nr_boosted = 0;
7601 rt_rq->rq = rq;
7602 #endif
7605 #ifdef CONFIG_FAIR_GROUP_SCHED
7606 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7607 struct sched_entity *se, int cpu, int add,
7608 struct sched_entity *parent)
7610 struct rq *rq = cpu_rq(cpu);
7611 tg->cfs_rq[cpu] = cfs_rq;
7612 init_cfs_rq(cfs_rq, rq);
7613 cfs_rq->tg = tg;
7614 if (add)
7615 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7617 tg->se[cpu] = se;
7618 /* se could be NULL for init_task_group */
7619 if (!se)
7620 return;
7622 if (!parent)
7623 se->cfs_rq = &rq->cfs;
7624 else
7625 se->cfs_rq = parent->my_q;
7627 se->my_q = cfs_rq;
7628 se->load.weight = tg->shares;
7629 se->load.inv_weight = 0;
7630 se->parent = parent;
7632 #endif
7634 #ifdef CONFIG_RT_GROUP_SCHED
7635 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7636 struct sched_rt_entity *rt_se, int cpu, int add,
7637 struct sched_rt_entity *parent)
7639 struct rq *rq = cpu_rq(cpu);
7641 tg->rt_rq[cpu] = rt_rq;
7642 init_rt_rq(rt_rq, rq);
7643 rt_rq->tg = tg;
7644 rt_rq->rt_se = rt_se;
7645 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7646 if (add)
7647 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7649 tg->rt_se[cpu] = rt_se;
7650 if (!rt_se)
7651 return;
7653 if (!parent)
7654 rt_se->rt_rq = &rq->rt;
7655 else
7656 rt_se->rt_rq = parent->my_q;
7658 rt_se->my_q = rt_rq;
7659 rt_se->parent = parent;
7660 INIT_LIST_HEAD(&rt_se->run_list);
7662 #endif
7664 void __init sched_init(void)
7666 int i, j;
7667 unsigned long alloc_size = 0, ptr;
7669 #ifdef CONFIG_FAIR_GROUP_SCHED
7670 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7671 #endif
7672 #ifdef CONFIG_RT_GROUP_SCHED
7673 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7674 #endif
7675 #ifdef CONFIG_USER_SCHED
7676 alloc_size *= 2;
7677 #endif
7679 * As sched_init() is called before page_alloc is setup,
7680 * we use alloc_bootmem().
7682 if (alloc_size) {
7683 ptr = (unsigned long)alloc_bootmem(alloc_size);
7685 #ifdef CONFIG_FAIR_GROUP_SCHED
7686 init_task_group.se = (struct sched_entity **)ptr;
7687 ptr += nr_cpu_ids * sizeof(void **);
7689 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7690 ptr += nr_cpu_ids * sizeof(void **);
7692 #ifdef CONFIG_USER_SCHED
7693 root_task_group.se = (struct sched_entity **)ptr;
7694 ptr += nr_cpu_ids * sizeof(void **);
7696 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7697 ptr += nr_cpu_ids * sizeof(void **);
7698 #endif
7699 #endif
7700 #ifdef CONFIG_RT_GROUP_SCHED
7701 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7702 ptr += nr_cpu_ids * sizeof(void **);
7704 init_task_group.rt_rq = (struct rt_rq **)ptr;
7705 ptr += nr_cpu_ids * sizeof(void **);
7707 #ifdef CONFIG_USER_SCHED
7708 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7709 ptr += nr_cpu_ids * sizeof(void **);
7711 root_task_group.rt_rq = (struct rt_rq **)ptr;
7712 ptr += nr_cpu_ids * sizeof(void **);
7713 #endif
7714 #endif
7717 #ifdef CONFIG_SMP
7718 init_defrootdomain();
7719 #endif
7721 init_rt_bandwidth(&def_rt_bandwidth,
7722 global_rt_period(), global_rt_runtime());
7724 #ifdef CONFIG_RT_GROUP_SCHED
7725 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7726 global_rt_period(), global_rt_runtime());
7727 #ifdef CONFIG_USER_SCHED
7728 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7729 global_rt_period(), RUNTIME_INF);
7730 #endif
7731 #endif
7733 #ifdef CONFIG_GROUP_SCHED
7734 list_add(&init_task_group.list, &task_groups);
7735 INIT_LIST_HEAD(&init_task_group.children);
7737 #ifdef CONFIG_USER_SCHED
7738 INIT_LIST_HEAD(&root_task_group.children);
7739 init_task_group.parent = &root_task_group;
7740 list_add(&init_task_group.siblings, &root_task_group.children);
7741 #endif
7742 #endif
7744 for_each_possible_cpu(i) {
7745 struct rq *rq;
7747 rq = cpu_rq(i);
7748 spin_lock_init(&rq->lock);
7749 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7750 rq->nr_running = 0;
7751 init_cfs_rq(&rq->cfs, rq);
7752 init_rt_rq(&rq->rt, rq);
7753 #ifdef CONFIG_FAIR_GROUP_SCHED
7754 init_task_group.shares = init_task_group_load;
7755 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7756 #ifdef CONFIG_CGROUP_SCHED
7758 * How much cpu bandwidth does init_task_group get?
7760 * In case of task-groups formed thr' the cgroup filesystem, it
7761 * gets 100% of the cpu resources in the system. This overall
7762 * system cpu resource is divided among the tasks of
7763 * init_task_group and its child task-groups in a fair manner,
7764 * based on each entity's (task or task-group's) weight
7765 * (se->load.weight).
7767 * In other words, if init_task_group has 10 tasks of weight
7768 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7769 * then A0's share of the cpu resource is:
7771 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7773 * We achieve this by letting init_task_group's tasks sit
7774 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7776 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7777 #elif defined CONFIG_USER_SCHED
7778 root_task_group.shares = NICE_0_LOAD;
7779 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
7781 * In case of task-groups formed thr' the user id of tasks,
7782 * init_task_group represents tasks belonging to root user.
7783 * Hence it forms a sibling of all subsequent groups formed.
7784 * In this case, init_task_group gets only a fraction of overall
7785 * system cpu resource, based on the weight assigned to root
7786 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7787 * by letting tasks of init_task_group sit in a separate cfs_rq
7788 * (init_cfs_rq) and having one entity represent this group of
7789 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7791 init_tg_cfs_entry(&init_task_group,
7792 &per_cpu(init_cfs_rq, i),
7793 &per_cpu(init_sched_entity, i), i, 1,
7794 root_task_group.se[i]);
7796 #endif
7797 #endif /* CONFIG_FAIR_GROUP_SCHED */
7799 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7800 #ifdef CONFIG_RT_GROUP_SCHED
7801 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7802 #ifdef CONFIG_CGROUP_SCHED
7803 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7804 #elif defined CONFIG_USER_SCHED
7805 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
7806 init_tg_rt_entry(&init_task_group,
7807 &per_cpu(init_rt_rq, i),
7808 &per_cpu(init_sched_rt_entity, i), i, 1,
7809 root_task_group.rt_se[i]);
7810 #endif
7811 #endif
7813 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7814 rq->cpu_load[j] = 0;
7815 #ifdef CONFIG_SMP
7816 rq->sd = NULL;
7817 rq->rd = NULL;
7818 rq->active_balance = 0;
7819 rq->next_balance = jiffies;
7820 rq->push_cpu = 0;
7821 rq->cpu = i;
7822 rq->migration_thread = NULL;
7823 INIT_LIST_HEAD(&rq->migration_queue);
7824 rq_attach_root(rq, &def_root_domain);
7825 #endif
7826 init_rq_hrtick(rq);
7827 atomic_set(&rq->nr_iowait, 0);
7830 set_load_weight(&init_task);
7832 #ifdef CONFIG_PREEMPT_NOTIFIERS
7833 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7834 #endif
7836 #ifdef CONFIG_SMP
7837 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7838 #endif
7840 #ifdef CONFIG_RT_MUTEXES
7841 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7842 #endif
7845 * The boot idle thread does lazy MMU switching as well:
7847 atomic_inc(&init_mm.mm_count);
7848 enter_lazy_tlb(&init_mm, current);
7851 * Make us the idle thread. Technically, schedule() should not be
7852 * called from this thread, however somewhere below it might be,
7853 * but because we are the idle thread, we just pick up running again
7854 * when this runqueue becomes "idle".
7856 init_idle(current, smp_processor_id());
7858 * During early bootup we pretend to be a normal task:
7860 current->sched_class = &fair_sched_class;
7862 scheduler_running = 1;
7865 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7866 void __might_sleep(char *file, int line)
7868 #ifdef in_atomic
7869 static unsigned long prev_jiffy; /* ratelimiting */
7871 if ((in_atomic() || irqs_disabled()) &&
7872 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7873 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7874 return;
7875 prev_jiffy = jiffies;
7876 printk(KERN_ERR "BUG: sleeping function called from invalid"
7877 " context at %s:%d\n", file, line);
7878 printk("in_atomic():%d, irqs_disabled():%d\n",
7879 in_atomic(), irqs_disabled());
7880 debug_show_held_locks(current);
7881 if (irqs_disabled())
7882 print_irqtrace_events(current);
7883 dump_stack();
7885 #endif
7887 EXPORT_SYMBOL(__might_sleep);
7888 #endif
7890 #ifdef CONFIG_MAGIC_SYSRQ
7891 static void normalize_task(struct rq *rq, struct task_struct *p)
7893 int on_rq;
7895 update_rq_clock(rq);
7896 on_rq = p->se.on_rq;
7897 if (on_rq)
7898 deactivate_task(rq, p, 0);
7899 __setscheduler(rq, p, SCHED_NORMAL, 0);
7900 if (on_rq) {
7901 activate_task(rq, p, 0);
7902 resched_task(rq->curr);
7906 void normalize_rt_tasks(void)
7908 struct task_struct *g, *p;
7909 unsigned long flags;
7910 struct rq *rq;
7912 read_lock_irqsave(&tasklist_lock, flags);
7913 do_each_thread(g, p) {
7915 * Only normalize user tasks:
7917 if (!p->mm)
7918 continue;
7920 p->se.exec_start = 0;
7921 #ifdef CONFIG_SCHEDSTATS
7922 p->se.wait_start = 0;
7923 p->se.sleep_start = 0;
7924 p->se.block_start = 0;
7925 #endif
7927 if (!rt_task(p)) {
7929 * Renice negative nice level userspace
7930 * tasks back to 0:
7932 if (TASK_NICE(p) < 0 && p->mm)
7933 set_user_nice(p, 0);
7934 continue;
7937 spin_lock(&p->pi_lock);
7938 rq = __task_rq_lock(p);
7940 normalize_task(rq, p);
7942 __task_rq_unlock(rq);
7943 spin_unlock(&p->pi_lock);
7944 } while_each_thread(g, p);
7946 read_unlock_irqrestore(&tasklist_lock, flags);
7949 #endif /* CONFIG_MAGIC_SYSRQ */
7951 #ifdef CONFIG_IA64
7953 * These functions are only useful for the IA64 MCA handling.
7955 * They can only be called when the whole system has been
7956 * stopped - every CPU needs to be quiescent, and no scheduling
7957 * activity can take place. Using them for anything else would
7958 * be a serious bug, and as a result, they aren't even visible
7959 * under any other configuration.
7963 * curr_task - return the current task for a given cpu.
7964 * @cpu: the processor in question.
7966 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7968 struct task_struct *curr_task(int cpu)
7970 return cpu_curr(cpu);
7974 * set_curr_task - set the current task for a given cpu.
7975 * @cpu: the processor in question.
7976 * @p: the task pointer to set.
7978 * Description: This function must only be used when non-maskable interrupts
7979 * are serviced on a separate stack. It allows the architecture to switch the
7980 * notion of the current task on a cpu in a non-blocking manner. This function
7981 * must be called with all CPU's synchronized, and interrupts disabled, the
7982 * and caller must save the original value of the current task (see
7983 * curr_task() above) and restore that value before reenabling interrupts and
7984 * re-starting the system.
7986 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7988 void set_curr_task(int cpu, struct task_struct *p)
7990 cpu_curr(cpu) = p;
7993 #endif
7995 #ifdef CONFIG_FAIR_GROUP_SCHED
7996 static void free_fair_sched_group(struct task_group *tg)
7998 int i;
8000 for_each_possible_cpu(i) {
8001 if (tg->cfs_rq)
8002 kfree(tg->cfs_rq[i]);
8003 if (tg->se)
8004 kfree(tg->se[i]);
8007 kfree(tg->cfs_rq);
8008 kfree(tg->se);
8011 static
8012 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8014 struct cfs_rq *cfs_rq;
8015 struct sched_entity *se, *parent_se;
8016 struct rq *rq;
8017 int i;
8019 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8020 if (!tg->cfs_rq)
8021 goto err;
8022 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8023 if (!tg->se)
8024 goto err;
8026 tg->shares = NICE_0_LOAD;
8028 for_each_possible_cpu(i) {
8029 rq = cpu_rq(i);
8031 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8032 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8033 if (!cfs_rq)
8034 goto err;
8036 se = kmalloc_node(sizeof(struct sched_entity),
8037 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8038 if (!se)
8039 goto err;
8041 parent_se = parent ? parent->se[i] : NULL;
8042 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8045 return 1;
8047 err:
8048 return 0;
8051 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8053 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8054 &cpu_rq(cpu)->leaf_cfs_rq_list);
8057 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8059 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8061 #else
8062 static inline void free_fair_sched_group(struct task_group *tg)
8066 static inline
8067 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8069 return 1;
8072 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8076 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8079 #endif
8081 #ifdef CONFIG_RT_GROUP_SCHED
8082 static void free_rt_sched_group(struct task_group *tg)
8084 int i;
8086 destroy_rt_bandwidth(&tg->rt_bandwidth);
8088 for_each_possible_cpu(i) {
8089 if (tg->rt_rq)
8090 kfree(tg->rt_rq[i]);
8091 if (tg->rt_se)
8092 kfree(tg->rt_se[i]);
8095 kfree(tg->rt_rq);
8096 kfree(tg->rt_se);
8099 static
8100 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8102 struct rt_rq *rt_rq;
8103 struct sched_rt_entity *rt_se, *parent_se;
8104 struct rq *rq;
8105 int i;
8107 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8108 if (!tg->rt_rq)
8109 goto err;
8110 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8111 if (!tg->rt_se)
8112 goto err;
8114 init_rt_bandwidth(&tg->rt_bandwidth,
8115 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8117 for_each_possible_cpu(i) {
8118 rq = cpu_rq(i);
8120 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8121 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8122 if (!rt_rq)
8123 goto err;
8125 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8126 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8127 if (!rt_se)
8128 goto err;
8130 parent_se = parent ? parent->rt_se[i] : NULL;
8131 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8134 return 1;
8136 err:
8137 return 0;
8140 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8142 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8143 &cpu_rq(cpu)->leaf_rt_rq_list);
8146 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8148 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8150 #else
8151 static inline void free_rt_sched_group(struct task_group *tg)
8155 static inline
8156 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8158 return 1;
8161 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8165 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8168 #endif
8170 #ifdef CONFIG_GROUP_SCHED
8171 static void free_sched_group(struct task_group *tg)
8173 free_fair_sched_group(tg);
8174 free_rt_sched_group(tg);
8175 kfree(tg);
8178 /* allocate runqueue etc for a new task group */
8179 struct task_group *sched_create_group(struct task_group *parent)
8181 struct task_group *tg;
8182 unsigned long flags;
8183 int i;
8185 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8186 if (!tg)
8187 return ERR_PTR(-ENOMEM);
8189 if (!alloc_fair_sched_group(tg, parent))
8190 goto err;
8192 if (!alloc_rt_sched_group(tg, parent))
8193 goto err;
8195 spin_lock_irqsave(&task_group_lock, flags);
8196 for_each_possible_cpu(i) {
8197 register_fair_sched_group(tg, i);
8198 register_rt_sched_group(tg, i);
8200 list_add_rcu(&tg->list, &task_groups);
8202 WARN_ON(!parent); /* root should already exist */
8204 tg->parent = parent;
8205 list_add_rcu(&tg->siblings, &parent->children);
8206 INIT_LIST_HEAD(&tg->children);
8207 spin_unlock_irqrestore(&task_group_lock, flags);
8209 return tg;
8211 err:
8212 free_sched_group(tg);
8213 return ERR_PTR(-ENOMEM);
8216 /* rcu callback to free various structures associated with a task group */
8217 static void free_sched_group_rcu(struct rcu_head *rhp)
8219 /* now it should be safe to free those cfs_rqs */
8220 free_sched_group(container_of(rhp, struct task_group, rcu));
8223 /* Destroy runqueue etc associated with a task group */
8224 void sched_destroy_group(struct task_group *tg)
8226 unsigned long flags;
8227 int i;
8229 spin_lock_irqsave(&task_group_lock, flags);
8230 for_each_possible_cpu(i) {
8231 unregister_fair_sched_group(tg, i);
8232 unregister_rt_sched_group(tg, i);
8234 list_del_rcu(&tg->list);
8235 list_del_rcu(&tg->siblings);
8236 spin_unlock_irqrestore(&task_group_lock, flags);
8238 /* wait for possible concurrent references to cfs_rqs complete */
8239 call_rcu(&tg->rcu, free_sched_group_rcu);
8242 /* change task's runqueue when it moves between groups.
8243 * The caller of this function should have put the task in its new group
8244 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8245 * reflect its new group.
8247 void sched_move_task(struct task_struct *tsk)
8249 int on_rq, running;
8250 unsigned long flags;
8251 struct rq *rq;
8253 rq = task_rq_lock(tsk, &flags);
8255 update_rq_clock(rq);
8257 running = task_current(rq, tsk);
8258 on_rq = tsk->se.on_rq;
8260 if (on_rq)
8261 dequeue_task(rq, tsk, 0);
8262 if (unlikely(running))
8263 tsk->sched_class->put_prev_task(rq, tsk);
8265 set_task_rq(tsk, task_cpu(tsk));
8267 #ifdef CONFIG_FAIR_GROUP_SCHED
8268 if (tsk->sched_class->moved_group)
8269 tsk->sched_class->moved_group(tsk);
8270 #endif
8272 if (unlikely(running))
8273 tsk->sched_class->set_curr_task(rq);
8274 if (on_rq)
8275 enqueue_task(rq, tsk, 0);
8277 task_rq_unlock(rq, &flags);
8279 #endif
8281 #ifdef CONFIG_FAIR_GROUP_SCHED
8282 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8284 struct cfs_rq *cfs_rq = se->cfs_rq;
8285 struct rq *rq = cfs_rq->rq;
8286 int on_rq;
8288 spin_lock_irq(&rq->lock);
8290 on_rq = se->on_rq;
8291 if (on_rq)
8292 dequeue_entity(cfs_rq, se, 0);
8294 se->load.weight = shares;
8295 se->load.inv_weight = 0;
8297 if (on_rq)
8298 enqueue_entity(cfs_rq, se, 0);
8300 spin_unlock_irq(&rq->lock);
8303 static DEFINE_MUTEX(shares_mutex);
8305 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8307 int i;
8308 unsigned long flags;
8311 * We can't change the weight of the root cgroup.
8313 if (!tg->se[0])
8314 return -EINVAL;
8316 if (shares < MIN_SHARES)
8317 shares = MIN_SHARES;
8318 else if (shares > MAX_SHARES)
8319 shares = MAX_SHARES;
8321 mutex_lock(&shares_mutex);
8322 if (tg->shares == shares)
8323 goto done;
8325 spin_lock_irqsave(&task_group_lock, flags);
8326 for_each_possible_cpu(i)
8327 unregister_fair_sched_group(tg, i);
8328 list_del_rcu(&tg->siblings);
8329 spin_unlock_irqrestore(&task_group_lock, flags);
8331 /* wait for any ongoing reference to this group to finish */
8332 synchronize_sched();
8335 * Now we are free to modify the group's share on each cpu
8336 * w/o tripping rebalance_share or load_balance_fair.
8338 tg->shares = shares;
8339 for_each_possible_cpu(i)
8340 set_se_shares(tg->se[i], shares);
8343 * Enable load balance activity on this group, by inserting it back on
8344 * each cpu's rq->leaf_cfs_rq_list.
8346 spin_lock_irqsave(&task_group_lock, flags);
8347 for_each_possible_cpu(i)
8348 register_fair_sched_group(tg, i);
8349 list_add_rcu(&tg->siblings, &tg->parent->children);
8350 spin_unlock_irqrestore(&task_group_lock, flags);
8351 done:
8352 mutex_unlock(&shares_mutex);
8353 return 0;
8356 unsigned long sched_group_shares(struct task_group *tg)
8358 return tg->shares;
8360 #endif
8362 #ifdef CONFIG_RT_GROUP_SCHED
8364 * Ensure that the real time constraints are schedulable.
8366 static DEFINE_MUTEX(rt_constraints_mutex);
8368 static unsigned long to_ratio(u64 period, u64 runtime)
8370 if (runtime == RUNTIME_INF)
8371 return 1ULL << 16;
8373 return div64_u64(runtime << 16, period);
8376 #ifdef CONFIG_CGROUP_SCHED
8377 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8379 struct task_group *tgi, *parent = tg ? tg->parent : NULL;
8380 unsigned long total = 0;
8382 if (!parent) {
8383 if (global_rt_period() < period)
8384 return 0;
8386 return to_ratio(period, runtime) <
8387 to_ratio(global_rt_period(), global_rt_runtime());
8390 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8391 return 0;
8393 rcu_read_lock();
8394 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8395 if (tgi == tg)
8396 continue;
8398 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8399 tgi->rt_bandwidth.rt_runtime);
8401 rcu_read_unlock();
8403 return total + to_ratio(period, runtime) <
8404 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8405 parent->rt_bandwidth.rt_runtime);
8407 #elif defined CONFIG_USER_SCHED
8408 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8410 struct task_group *tgi;
8411 unsigned long total = 0;
8412 unsigned long global_ratio =
8413 to_ratio(global_rt_period(), global_rt_runtime());
8415 rcu_read_lock();
8416 list_for_each_entry_rcu(tgi, &task_groups, list) {
8417 if (tgi == tg)
8418 continue;
8420 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8421 tgi->rt_bandwidth.rt_runtime);
8423 rcu_read_unlock();
8425 return total + to_ratio(period, runtime) < global_ratio;
8427 #endif
8429 /* Must be called with tasklist_lock held */
8430 static inline int tg_has_rt_tasks(struct task_group *tg)
8432 struct task_struct *g, *p;
8433 do_each_thread(g, p) {
8434 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8435 return 1;
8436 } while_each_thread(g, p);
8437 return 0;
8440 static int tg_set_bandwidth(struct task_group *tg,
8441 u64 rt_period, u64 rt_runtime)
8443 int i, err = 0;
8445 mutex_lock(&rt_constraints_mutex);
8446 read_lock(&tasklist_lock);
8447 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8448 err = -EBUSY;
8449 goto unlock;
8451 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8452 err = -EINVAL;
8453 goto unlock;
8456 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8457 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8458 tg->rt_bandwidth.rt_runtime = rt_runtime;
8460 for_each_possible_cpu(i) {
8461 struct rt_rq *rt_rq = tg->rt_rq[i];
8463 spin_lock(&rt_rq->rt_runtime_lock);
8464 rt_rq->rt_runtime = rt_runtime;
8465 spin_unlock(&rt_rq->rt_runtime_lock);
8467 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8468 unlock:
8469 read_unlock(&tasklist_lock);
8470 mutex_unlock(&rt_constraints_mutex);
8472 return err;
8475 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8477 u64 rt_runtime, rt_period;
8479 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8480 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8481 if (rt_runtime_us < 0)
8482 rt_runtime = RUNTIME_INF;
8484 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8487 long sched_group_rt_runtime(struct task_group *tg)
8489 u64 rt_runtime_us;
8491 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8492 return -1;
8494 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8495 do_div(rt_runtime_us, NSEC_PER_USEC);
8496 return rt_runtime_us;
8499 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8501 u64 rt_runtime, rt_period;
8503 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8504 rt_runtime = tg->rt_bandwidth.rt_runtime;
8506 if (rt_period == 0)
8507 return -EINVAL;
8509 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8512 long sched_group_rt_period(struct task_group *tg)
8514 u64 rt_period_us;
8516 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8517 do_div(rt_period_us, NSEC_PER_USEC);
8518 return rt_period_us;
8521 static int sched_rt_global_constraints(void)
8523 int ret = 0;
8525 mutex_lock(&rt_constraints_mutex);
8526 if (!__rt_schedulable(NULL, 1, 0))
8527 ret = -EINVAL;
8528 mutex_unlock(&rt_constraints_mutex);
8530 return ret;
8532 #else
8533 static int sched_rt_global_constraints(void)
8535 unsigned long flags;
8536 int i;
8538 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8539 for_each_possible_cpu(i) {
8540 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8542 spin_lock(&rt_rq->rt_runtime_lock);
8543 rt_rq->rt_runtime = global_rt_runtime();
8544 spin_unlock(&rt_rq->rt_runtime_lock);
8546 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8548 return 0;
8550 #endif
8552 int sched_rt_handler(struct ctl_table *table, int write,
8553 struct file *filp, void __user *buffer, size_t *lenp,
8554 loff_t *ppos)
8556 int ret;
8557 int old_period, old_runtime;
8558 static DEFINE_MUTEX(mutex);
8560 mutex_lock(&mutex);
8561 old_period = sysctl_sched_rt_period;
8562 old_runtime = sysctl_sched_rt_runtime;
8564 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8566 if (!ret && write) {
8567 ret = sched_rt_global_constraints();
8568 if (ret) {
8569 sysctl_sched_rt_period = old_period;
8570 sysctl_sched_rt_runtime = old_runtime;
8571 } else {
8572 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8573 def_rt_bandwidth.rt_period =
8574 ns_to_ktime(global_rt_period());
8577 mutex_unlock(&mutex);
8579 return ret;
8582 #ifdef CONFIG_CGROUP_SCHED
8584 /* return corresponding task_group object of a cgroup */
8585 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8587 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8588 struct task_group, css);
8591 static struct cgroup_subsys_state *
8592 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8594 struct task_group *tg, *parent;
8596 if (!cgrp->parent) {
8597 /* This is early initialization for the top cgroup */
8598 init_task_group.css.cgroup = cgrp;
8599 return &init_task_group.css;
8602 parent = cgroup_tg(cgrp->parent);
8603 tg = sched_create_group(parent);
8604 if (IS_ERR(tg))
8605 return ERR_PTR(-ENOMEM);
8607 /* Bind the cgroup to task_group object we just created */
8608 tg->css.cgroup = cgrp;
8610 return &tg->css;
8613 static void
8614 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8616 struct task_group *tg = cgroup_tg(cgrp);
8618 sched_destroy_group(tg);
8621 static int
8622 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8623 struct task_struct *tsk)
8625 #ifdef CONFIG_RT_GROUP_SCHED
8626 /* Don't accept realtime tasks when there is no way for them to run */
8627 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8628 return -EINVAL;
8629 #else
8630 /* We don't support RT-tasks being in separate groups */
8631 if (tsk->sched_class != &fair_sched_class)
8632 return -EINVAL;
8633 #endif
8635 return 0;
8638 static void
8639 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8640 struct cgroup *old_cont, struct task_struct *tsk)
8642 sched_move_task(tsk);
8645 #ifdef CONFIG_FAIR_GROUP_SCHED
8646 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8647 u64 shareval)
8649 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8652 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8654 struct task_group *tg = cgroup_tg(cgrp);
8656 return (u64) tg->shares;
8658 #endif
8660 #ifdef CONFIG_RT_GROUP_SCHED
8661 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8662 s64 val)
8664 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8667 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8669 return sched_group_rt_runtime(cgroup_tg(cgrp));
8672 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8673 u64 rt_period_us)
8675 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8678 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8680 return sched_group_rt_period(cgroup_tg(cgrp));
8682 #endif
8684 static struct cftype cpu_files[] = {
8685 #ifdef CONFIG_FAIR_GROUP_SCHED
8687 .name = "shares",
8688 .read_u64 = cpu_shares_read_u64,
8689 .write_u64 = cpu_shares_write_u64,
8691 #endif
8692 #ifdef CONFIG_RT_GROUP_SCHED
8694 .name = "rt_runtime_us",
8695 .read_s64 = cpu_rt_runtime_read,
8696 .write_s64 = cpu_rt_runtime_write,
8699 .name = "rt_period_us",
8700 .read_u64 = cpu_rt_period_read_uint,
8701 .write_u64 = cpu_rt_period_write_uint,
8703 #endif
8706 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8708 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8711 struct cgroup_subsys cpu_cgroup_subsys = {
8712 .name = "cpu",
8713 .create = cpu_cgroup_create,
8714 .destroy = cpu_cgroup_destroy,
8715 .can_attach = cpu_cgroup_can_attach,
8716 .attach = cpu_cgroup_attach,
8717 .populate = cpu_cgroup_populate,
8718 .subsys_id = cpu_cgroup_subsys_id,
8719 .early_init = 1,
8722 #endif /* CONFIG_CGROUP_SCHED */
8724 #ifdef CONFIG_CGROUP_CPUACCT
8727 * CPU accounting code for task groups.
8729 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8730 * (balbir@in.ibm.com).
8733 /* track cpu usage of a group of tasks */
8734 struct cpuacct {
8735 struct cgroup_subsys_state css;
8736 /* cpuusage holds pointer to a u64-type object on every cpu */
8737 u64 *cpuusage;
8740 struct cgroup_subsys cpuacct_subsys;
8742 /* return cpu accounting group corresponding to this container */
8743 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8745 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8746 struct cpuacct, css);
8749 /* return cpu accounting group to which this task belongs */
8750 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8752 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8753 struct cpuacct, css);
8756 /* create a new cpu accounting group */
8757 static struct cgroup_subsys_state *cpuacct_create(
8758 struct cgroup_subsys *ss, struct cgroup *cgrp)
8760 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8762 if (!ca)
8763 return ERR_PTR(-ENOMEM);
8765 ca->cpuusage = alloc_percpu(u64);
8766 if (!ca->cpuusage) {
8767 kfree(ca);
8768 return ERR_PTR(-ENOMEM);
8771 return &ca->css;
8774 /* destroy an existing cpu accounting group */
8775 static void
8776 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8778 struct cpuacct *ca = cgroup_ca(cgrp);
8780 free_percpu(ca->cpuusage);
8781 kfree(ca);
8784 /* return total cpu usage (in nanoseconds) of a group */
8785 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8787 struct cpuacct *ca = cgroup_ca(cgrp);
8788 u64 totalcpuusage = 0;
8789 int i;
8791 for_each_possible_cpu(i) {
8792 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8795 * Take rq->lock to make 64-bit addition safe on 32-bit
8796 * platforms.
8798 spin_lock_irq(&cpu_rq(i)->lock);
8799 totalcpuusage += *cpuusage;
8800 spin_unlock_irq(&cpu_rq(i)->lock);
8803 return totalcpuusage;
8806 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8807 u64 reset)
8809 struct cpuacct *ca = cgroup_ca(cgrp);
8810 int err = 0;
8811 int i;
8813 if (reset) {
8814 err = -EINVAL;
8815 goto out;
8818 for_each_possible_cpu(i) {
8819 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8821 spin_lock_irq(&cpu_rq(i)->lock);
8822 *cpuusage = 0;
8823 spin_unlock_irq(&cpu_rq(i)->lock);
8825 out:
8826 return err;
8829 static struct cftype files[] = {
8831 .name = "usage",
8832 .read_u64 = cpuusage_read,
8833 .write_u64 = cpuusage_write,
8837 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8839 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8843 * charge this task's execution time to its accounting group.
8845 * called with rq->lock held.
8847 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8849 struct cpuacct *ca;
8851 if (!cpuacct_subsys.active)
8852 return;
8854 ca = task_ca(tsk);
8855 if (ca) {
8856 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8858 *cpuusage += cputime;
8862 struct cgroup_subsys cpuacct_subsys = {
8863 .name = "cpuacct",
8864 .create = cpuacct_create,
8865 .destroy = cpuacct_destroy,
8866 .populate = cpuacct_populate,
8867 .subsys_id = cpuacct_subsys_id,
8869 #endif /* CONFIG_CGROUP_CPUACCT */