sched: fair group: fix overflow(was: fix divide by zero)
[linux-2.6/cjktty.git] / kernel / sched.c
blob6c1ecbdc0db9941ffa7c5698761e71aa36381009
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 static void hotplug_hrtick_disable(int cpu)
1132 struct rq *rq = cpu_rq(cpu);
1133 unsigned long flags;
1135 spin_lock_irqsave(&rq->lock, flags);
1136 rq->hrtick_flags = 0;
1137 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1138 spin_unlock_irqrestore(&rq->lock, flags);
1140 hrtick_clear(rq);
1143 static void hotplug_hrtick_enable(int cpu)
1145 struct rq *rq = cpu_rq(cpu);
1146 unsigned long flags;
1148 spin_lock_irqsave(&rq->lock, flags);
1149 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1150 spin_unlock_irqrestore(&rq->lock, flags);
1153 static int
1154 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1156 int cpu = (int)(long)hcpu;
1158 switch (action) {
1159 case CPU_UP_CANCELED:
1160 case CPU_UP_CANCELED_FROZEN:
1161 case CPU_DOWN_PREPARE:
1162 case CPU_DOWN_PREPARE_FROZEN:
1163 case CPU_DEAD:
1164 case CPU_DEAD_FROZEN:
1165 hotplug_hrtick_disable(cpu);
1166 return NOTIFY_OK;
1168 case CPU_UP_PREPARE:
1169 case CPU_UP_PREPARE_FROZEN:
1170 case CPU_DOWN_FAILED:
1171 case CPU_DOWN_FAILED_FROZEN:
1172 case CPU_ONLINE:
1173 case CPU_ONLINE_FROZEN:
1174 hotplug_hrtick_enable(cpu);
1175 return NOTIFY_OK;
1178 return NOTIFY_DONE;
1181 static void init_hrtick(void)
1183 hotcpu_notifier(hotplug_hrtick, 0);
1186 static void init_rq_hrtick(struct rq *rq)
1188 rq->hrtick_flags = 0;
1189 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1190 rq->hrtick_timer.function = hrtick;
1191 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1194 void hrtick_resched(void)
1196 struct rq *rq;
1197 unsigned long flags;
1199 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1200 return;
1202 local_irq_save(flags);
1203 rq = cpu_rq(smp_processor_id());
1204 hrtick_set(rq);
1205 local_irq_restore(flags);
1207 #else
1208 static inline void hrtick_clear(struct rq *rq)
1212 static inline void hrtick_set(struct rq *rq)
1216 static inline void init_rq_hrtick(struct rq *rq)
1220 void hrtick_resched(void)
1224 static inline void init_hrtick(void)
1227 #endif
1230 * resched_task - mark a task 'to be rescheduled now'.
1232 * On UP this means the setting of the need_resched flag, on SMP it
1233 * might also involve a cross-CPU call to trigger the scheduler on
1234 * the target CPU.
1236 #ifdef CONFIG_SMP
1238 #ifndef tsk_is_polling
1239 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1240 #endif
1242 static void __resched_task(struct task_struct *p, int tif_bit)
1244 int cpu;
1246 assert_spin_locked(&task_rq(p)->lock);
1248 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1249 return;
1251 set_tsk_thread_flag(p, tif_bit);
1253 cpu = task_cpu(p);
1254 if (cpu == smp_processor_id())
1255 return;
1257 /* NEED_RESCHED must be visible before we test polling */
1258 smp_mb();
1259 if (!tsk_is_polling(p))
1260 smp_send_reschedule(cpu);
1263 static void resched_cpu(int cpu)
1265 struct rq *rq = cpu_rq(cpu);
1266 unsigned long flags;
1268 if (!spin_trylock_irqsave(&rq->lock, flags))
1269 return;
1270 resched_task(cpu_curr(cpu));
1271 spin_unlock_irqrestore(&rq->lock, flags);
1274 #ifdef CONFIG_NO_HZ
1276 * When add_timer_on() enqueues a timer into the timer wheel of an
1277 * idle CPU then this timer might expire before the next timer event
1278 * which is scheduled to wake up that CPU. In case of a completely
1279 * idle system the next event might even be infinite time into the
1280 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1281 * leaves the inner idle loop so the newly added timer is taken into
1282 * account when the CPU goes back to idle and evaluates the timer
1283 * wheel for the next timer event.
1285 void wake_up_idle_cpu(int cpu)
1287 struct rq *rq = cpu_rq(cpu);
1289 if (cpu == smp_processor_id())
1290 return;
1293 * This is safe, as this function is called with the timer
1294 * wheel base lock of (cpu) held. When the CPU is on the way
1295 * to idle and has not yet set rq->curr to idle then it will
1296 * be serialized on the timer wheel base lock and take the new
1297 * timer into account automatically.
1299 if (rq->curr != rq->idle)
1300 return;
1303 * We can set TIF_RESCHED on the idle task of the other CPU
1304 * lockless. The worst case is that the other CPU runs the
1305 * idle task through an additional NOOP schedule()
1307 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1309 /* NEED_RESCHED must be visible before we test polling */
1310 smp_mb();
1311 if (!tsk_is_polling(rq->idle))
1312 smp_send_reschedule(cpu);
1314 #endif
1316 #else
1317 static void __resched_task(struct task_struct *p, int tif_bit)
1319 assert_spin_locked(&task_rq(p)->lock);
1320 set_tsk_thread_flag(p, tif_bit);
1322 #endif
1324 #if BITS_PER_LONG == 32
1325 # define WMULT_CONST (~0UL)
1326 #else
1327 # define WMULT_CONST (1UL << 32)
1328 #endif
1330 #define WMULT_SHIFT 32
1333 * Shift right and round:
1335 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1337 static unsigned long
1338 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1339 struct load_weight *lw)
1341 u64 tmp;
1343 if (!lw->inv_weight)
1344 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)/(lw->weight+1);
1346 tmp = (u64)delta_exec * weight;
1348 * Check whether we'd overflow the 64-bit multiplication:
1350 if (unlikely(tmp > WMULT_CONST))
1351 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1352 WMULT_SHIFT/2);
1353 else
1354 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1356 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1359 static inline unsigned long
1360 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1362 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1365 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1367 lw->weight += inc;
1368 lw->inv_weight = 0;
1371 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1373 lw->weight -= dec;
1374 lw->inv_weight = 0;
1378 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1379 * of tasks with abnormal "nice" values across CPUs the contribution that
1380 * each task makes to its run queue's load is weighted according to its
1381 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1382 * scaled version of the new time slice allocation that they receive on time
1383 * slice expiry etc.
1386 #define WEIGHT_IDLEPRIO 2
1387 #define WMULT_IDLEPRIO (1 << 31)
1390 * Nice levels are multiplicative, with a gentle 10% change for every
1391 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1392 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1393 * that remained on nice 0.
1395 * The "10% effect" is relative and cumulative: from _any_ nice level,
1396 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1397 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1398 * If a task goes up by ~10% and another task goes down by ~10% then
1399 * the relative distance between them is ~25%.)
1401 static const int prio_to_weight[40] = {
1402 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1403 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1404 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1405 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1406 /* 0 */ 1024, 820, 655, 526, 423,
1407 /* 5 */ 335, 272, 215, 172, 137,
1408 /* 10 */ 110, 87, 70, 56, 45,
1409 /* 15 */ 36, 29, 23, 18, 15,
1413 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1415 * In cases where the weight does not change often, we can use the
1416 * precalculated inverse to speed up arithmetics by turning divisions
1417 * into multiplications:
1419 static const u32 prio_to_wmult[40] = {
1420 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1421 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1422 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1423 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1424 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1425 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1426 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1427 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1430 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1433 * runqueue iterator, to support SMP load-balancing between different
1434 * scheduling classes, without having to expose their internal data
1435 * structures to the load-balancing proper:
1437 struct rq_iterator {
1438 void *arg;
1439 struct task_struct *(*start)(void *);
1440 struct task_struct *(*next)(void *);
1443 #ifdef CONFIG_SMP
1444 static unsigned long
1445 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1446 unsigned long max_load_move, struct sched_domain *sd,
1447 enum cpu_idle_type idle, int *all_pinned,
1448 int *this_best_prio, struct rq_iterator *iterator);
1450 static int
1451 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1452 struct sched_domain *sd, enum cpu_idle_type idle,
1453 struct rq_iterator *iterator);
1454 #endif
1456 #ifdef CONFIG_CGROUP_CPUACCT
1457 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1458 #else
1459 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1460 #endif
1462 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1464 update_load_add(&rq->load, load);
1467 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1469 update_load_sub(&rq->load, load);
1472 #ifdef CONFIG_SMP
1473 static unsigned long source_load(int cpu, int type);
1474 static unsigned long target_load(int cpu, int type);
1475 static unsigned long cpu_avg_load_per_task(int cpu);
1476 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1477 #else /* CONFIG_SMP */
1479 #ifdef CONFIG_FAIR_GROUP_SCHED
1480 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1483 #endif
1485 #endif /* CONFIG_SMP */
1487 #include "sched_stats.h"
1488 #include "sched_idletask.c"
1489 #include "sched_fair.c"
1490 #include "sched_rt.c"
1491 #ifdef CONFIG_SCHED_DEBUG
1492 # include "sched_debug.c"
1493 #endif
1495 #define sched_class_highest (&rt_sched_class)
1497 static inline void inc_load(struct rq *rq, const struct task_struct *p)
1499 update_load_add(&rq->load, p->se.load.weight);
1502 static inline void dec_load(struct rq *rq, const struct task_struct *p)
1504 update_load_sub(&rq->load, p->se.load.weight);
1507 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1509 rq->nr_running++;
1510 inc_load(rq, p);
1513 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1515 rq->nr_running--;
1516 dec_load(rq, p);
1519 static void set_load_weight(struct task_struct *p)
1521 if (task_has_rt_policy(p)) {
1522 p->se.load.weight = prio_to_weight[0] * 2;
1523 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1524 return;
1528 * SCHED_IDLE tasks get minimal weight:
1530 if (p->policy == SCHED_IDLE) {
1531 p->se.load.weight = WEIGHT_IDLEPRIO;
1532 p->se.load.inv_weight = WMULT_IDLEPRIO;
1533 return;
1536 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1537 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1540 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1542 sched_info_queued(p);
1543 p->sched_class->enqueue_task(rq, p, wakeup);
1544 p->se.on_rq = 1;
1547 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1549 p->sched_class->dequeue_task(rq, p, sleep);
1550 p->se.on_rq = 0;
1554 * __normal_prio - return the priority that is based on the static prio
1556 static inline int __normal_prio(struct task_struct *p)
1558 return p->static_prio;
1562 * Calculate the expected normal priority: i.e. priority
1563 * without taking RT-inheritance into account. Might be
1564 * boosted by interactivity modifiers. Changes upon fork,
1565 * setprio syscalls, and whenever the interactivity
1566 * estimator recalculates.
1568 static inline int normal_prio(struct task_struct *p)
1570 int prio;
1572 if (task_has_rt_policy(p))
1573 prio = MAX_RT_PRIO-1 - p->rt_priority;
1574 else
1575 prio = __normal_prio(p);
1576 return prio;
1580 * Calculate the current priority, i.e. the priority
1581 * taken into account by the scheduler. This value might
1582 * be boosted by RT tasks, or might be boosted by
1583 * interactivity modifiers. Will be RT if the task got
1584 * RT-boosted. If not then it returns p->normal_prio.
1586 static int effective_prio(struct task_struct *p)
1588 p->normal_prio = normal_prio(p);
1590 * If we are RT tasks or we were boosted to RT priority,
1591 * keep the priority unchanged. Otherwise, update priority
1592 * to the normal priority:
1594 if (!rt_prio(p->prio))
1595 return p->normal_prio;
1596 return p->prio;
1600 * activate_task - move a task to the runqueue.
1602 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1604 if (task_contributes_to_load(p))
1605 rq->nr_uninterruptible--;
1607 enqueue_task(rq, p, wakeup);
1608 inc_nr_running(p, rq);
1612 * deactivate_task - remove a task from the runqueue.
1614 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1616 if (task_contributes_to_load(p))
1617 rq->nr_uninterruptible++;
1619 dequeue_task(rq, p, sleep);
1620 dec_nr_running(p, rq);
1624 * task_curr - is this task currently executing on a CPU?
1625 * @p: the task in question.
1627 inline int task_curr(const struct task_struct *p)
1629 return cpu_curr(task_cpu(p)) == p;
1632 /* Used instead of source_load when we know the type == 0 */
1633 unsigned long weighted_cpuload(const int cpu)
1635 return cpu_rq(cpu)->load.weight;
1638 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1640 set_task_rq(p, cpu);
1641 #ifdef CONFIG_SMP
1643 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1644 * successfuly executed on another CPU. We must ensure that updates of
1645 * per-task data have been completed by this moment.
1647 smp_wmb();
1648 task_thread_info(p)->cpu = cpu;
1649 #endif
1652 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1653 const struct sched_class *prev_class,
1654 int oldprio, int running)
1656 if (prev_class != p->sched_class) {
1657 if (prev_class->switched_from)
1658 prev_class->switched_from(rq, p, running);
1659 p->sched_class->switched_to(rq, p, running);
1660 } else
1661 p->sched_class->prio_changed(rq, p, oldprio, running);
1664 #ifdef CONFIG_SMP
1667 * Is this task likely cache-hot:
1669 static int
1670 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1672 s64 delta;
1675 * Buddy candidates are cache hot:
1677 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1678 return 1;
1680 if (p->sched_class != &fair_sched_class)
1681 return 0;
1683 if (sysctl_sched_migration_cost == -1)
1684 return 1;
1685 if (sysctl_sched_migration_cost == 0)
1686 return 0;
1688 delta = now - p->se.exec_start;
1690 return delta < (s64)sysctl_sched_migration_cost;
1694 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1696 int old_cpu = task_cpu(p);
1697 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1698 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1699 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1700 u64 clock_offset;
1702 clock_offset = old_rq->clock - new_rq->clock;
1704 #ifdef CONFIG_SCHEDSTATS
1705 if (p->se.wait_start)
1706 p->se.wait_start -= clock_offset;
1707 if (p->se.sleep_start)
1708 p->se.sleep_start -= clock_offset;
1709 if (p->se.block_start)
1710 p->se.block_start -= clock_offset;
1711 if (old_cpu != new_cpu) {
1712 schedstat_inc(p, se.nr_migrations);
1713 if (task_hot(p, old_rq->clock, NULL))
1714 schedstat_inc(p, se.nr_forced2_migrations);
1716 #endif
1717 p->se.vruntime -= old_cfsrq->min_vruntime -
1718 new_cfsrq->min_vruntime;
1720 __set_task_cpu(p, new_cpu);
1723 struct migration_req {
1724 struct list_head list;
1726 struct task_struct *task;
1727 int dest_cpu;
1729 struct completion done;
1733 * The task's runqueue lock must be held.
1734 * Returns true if you have to wait for migration thread.
1736 static int
1737 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1739 struct rq *rq = task_rq(p);
1742 * If the task is not on a runqueue (and not running), then
1743 * it is sufficient to simply update the task's cpu field.
1745 if (!p->se.on_rq && !task_running(rq, p)) {
1746 set_task_cpu(p, dest_cpu);
1747 return 0;
1750 init_completion(&req->done);
1751 req->task = p;
1752 req->dest_cpu = dest_cpu;
1753 list_add(&req->list, &rq->migration_queue);
1755 return 1;
1759 * wait_task_inactive - wait for a thread to unschedule.
1761 * The caller must ensure that the task *will* unschedule sometime soon,
1762 * else this function might spin for a *long* time. This function can't
1763 * be called with interrupts off, or it may introduce deadlock with
1764 * smp_call_function() if an IPI is sent by the same process we are
1765 * waiting to become inactive.
1767 void wait_task_inactive(struct task_struct *p)
1769 unsigned long flags;
1770 int running, on_rq;
1771 struct rq *rq;
1773 for (;;) {
1775 * We do the initial early heuristics without holding
1776 * any task-queue locks at all. We'll only try to get
1777 * the runqueue lock when things look like they will
1778 * work out!
1780 rq = task_rq(p);
1783 * If the task is actively running on another CPU
1784 * still, just relax and busy-wait without holding
1785 * any locks.
1787 * NOTE! Since we don't hold any locks, it's not
1788 * even sure that "rq" stays as the right runqueue!
1789 * But we don't care, since "task_running()" will
1790 * return false if the runqueue has changed and p
1791 * is actually now running somewhere else!
1793 while (task_running(rq, p))
1794 cpu_relax();
1797 * Ok, time to look more closely! We need the rq
1798 * lock now, to be *sure*. If we're wrong, we'll
1799 * just go back and repeat.
1801 rq = task_rq_lock(p, &flags);
1802 running = task_running(rq, p);
1803 on_rq = p->se.on_rq;
1804 task_rq_unlock(rq, &flags);
1807 * Was it really running after all now that we
1808 * checked with the proper locks actually held?
1810 * Oops. Go back and try again..
1812 if (unlikely(running)) {
1813 cpu_relax();
1814 continue;
1818 * It's not enough that it's not actively running,
1819 * it must be off the runqueue _entirely_, and not
1820 * preempted!
1822 * So if it wa still runnable (but just not actively
1823 * running right now), it's preempted, and we should
1824 * yield - it could be a while.
1826 if (unlikely(on_rq)) {
1827 schedule_timeout_uninterruptible(1);
1828 continue;
1832 * Ahh, all good. It wasn't running, and it wasn't
1833 * runnable, which means that it will never become
1834 * running in the future either. We're all done!
1836 break;
1840 /***
1841 * kick_process - kick a running thread to enter/exit the kernel
1842 * @p: the to-be-kicked thread
1844 * Cause a process which is running on another CPU to enter
1845 * kernel-mode, without any delay. (to get signals handled.)
1847 * NOTE: this function doesnt have to take the runqueue lock,
1848 * because all it wants to ensure is that the remote task enters
1849 * the kernel. If the IPI races and the task has been migrated
1850 * to another CPU then no harm is done and the purpose has been
1851 * achieved as well.
1853 void kick_process(struct task_struct *p)
1855 int cpu;
1857 preempt_disable();
1858 cpu = task_cpu(p);
1859 if ((cpu != smp_processor_id()) && task_curr(p))
1860 smp_send_reschedule(cpu);
1861 preempt_enable();
1865 * Return a low guess at the load of a migration-source cpu weighted
1866 * according to the scheduling class and "nice" value.
1868 * We want to under-estimate the load of migration sources, to
1869 * balance conservatively.
1871 static unsigned long source_load(int cpu, int type)
1873 struct rq *rq = cpu_rq(cpu);
1874 unsigned long total = weighted_cpuload(cpu);
1876 if (type == 0)
1877 return total;
1879 return min(rq->cpu_load[type-1], total);
1883 * Return a high guess at the load of a migration-target cpu weighted
1884 * according to the scheduling class and "nice" value.
1886 static unsigned long target_load(int cpu, int type)
1888 struct rq *rq = cpu_rq(cpu);
1889 unsigned long total = weighted_cpuload(cpu);
1891 if (type == 0)
1892 return total;
1894 return max(rq->cpu_load[type-1], total);
1898 * Return the average load per task on the cpu's run queue
1900 static unsigned long cpu_avg_load_per_task(int cpu)
1902 struct rq *rq = cpu_rq(cpu);
1903 unsigned long total = weighted_cpuload(cpu);
1904 unsigned long n = rq->nr_running;
1906 return n ? total / n : SCHED_LOAD_SCALE;
1910 * find_idlest_group finds and returns the least busy CPU group within the
1911 * domain.
1913 static struct sched_group *
1914 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1916 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1917 unsigned long min_load = ULONG_MAX, this_load = 0;
1918 int load_idx = sd->forkexec_idx;
1919 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1921 do {
1922 unsigned long load, avg_load;
1923 int local_group;
1924 int i;
1926 /* Skip over this group if it has no CPUs allowed */
1927 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1928 continue;
1930 local_group = cpu_isset(this_cpu, group->cpumask);
1932 /* Tally up the load of all CPUs in the group */
1933 avg_load = 0;
1935 for_each_cpu_mask(i, group->cpumask) {
1936 /* Bias balancing toward cpus of our domain */
1937 if (local_group)
1938 load = source_load(i, load_idx);
1939 else
1940 load = target_load(i, load_idx);
1942 avg_load += load;
1945 /* Adjust by relative CPU power of the group */
1946 avg_load = sg_div_cpu_power(group,
1947 avg_load * SCHED_LOAD_SCALE);
1949 if (local_group) {
1950 this_load = avg_load;
1951 this = group;
1952 } else if (avg_load < min_load) {
1953 min_load = avg_load;
1954 idlest = group;
1956 } while (group = group->next, group != sd->groups);
1958 if (!idlest || 100*this_load < imbalance*min_load)
1959 return NULL;
1960 return idlest;
1964 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1966 static int
1967 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
1968 cpumask_t *tmp)
1970 unsigned long load, min_load = ULONG_MAX;
1971 int idlest = -1;
1972 int i;
1974 /* Traverse only the allowed CPUs */
1975 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
1977 for_each_cpu_mask(i, *tmp) {
1978 load = weighted_cpuload(i);
1980 if (load < min_load || (load == min_load && i == this_cpu)) {
1981 min_load = load;
1982 idlest = i;
1986 return idlest;
1990 * sched_balance_self: balance the current task (running on cpu) in domains
1991 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1992 * SD_BALANCE_EXEC.
1994 * Balance, ie. select the least loaded group.
1996 * Returns the target CPU number, or the same CPU if no balancing is needed.
1998 * preempt must be disabled.
2000 static int sched_balance_self(int cpu, int flag)
2002 struct task_struct *t = current;
2003 struct sched_domain *tmp, *sd = NULL;
2005 for_each_domain(cpu, tmp) {
2007 * If power savings logic is enabled for a domain, stop there.
2009 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2010 break;
2011 if (tmp->flags & flag)
2012 sd = tmp;
2015 while (sd) {
2016 cpumask_t span, tmpmask;
2017 struct sched_group *group;
2018 int new_cpu, weight;
2020 if (!(sd->flags & flag)) {
2021 sd = sd->child;
2022 continue;
2025 span = sd->span;
2026 group = find_idlest_group(sd, t, cpu);
2027 if (!group) {
2028 sd = sd->child;
2029 continue;
2032 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2033 if (new_cpu == -1 || new_cpu == cpu) {
2034 /* Now try balancing at a lower domain level of cpu */
2035 sd = sd->child;
2036 continue;
2039 /* Now try balancing at a lower domain level of new_cpu */
2040 cpu = new_cpu;
2041 sd = NULL;
2042 weight = cpus_weight(span);
2043 for_each_domain(cpu, tmp) {
2044 if (weight <= cpus_weight(tmp->span))
2045 break;
2046 if (tmp->flags & flag)
2047 sd = tmp;
2049 /* while loop will break here if sd == NULL */
2052 return cpu;
2055 #endif /* CONFIG_SMP */
2057 /***
2058 * try_to_wake_up - wake up a thread
2059 * @p: the to-be-woken-up thread
2060 * @state: the mask of task states that can be woken
2061 * @sync: do a synchronous wakeup?
2063 * Put it on the run-queue if it's not already there. The "current"
2064 * thread is always on the run-queue (except when the actual
2065 * re-schedule is in progress), and as such you're allowed to do
2066 * the simpler "current->state = TASK_RUNNING" to mark yourself
2067 * runnable without the overhead of this.
2069 * returns failure only if the task is already active.
2071 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2073 int cpu, orig_cpu, this_cpu, success = 0;
2074 unsigned long flags;
2075 long old_state;
2076 struct rq *rq;
2078 if (!sched_feat(SYNC_WAKEUPS))
2079 sync = 0;
2081 smp_wmb();
2082 rq = task_rq_lock(p, &flags);
2083 old_state = p->state;
2084 if (!(old_state & state))
2085 goto out;
2087 if (p->se.on_rq)
2088 goto out_running;
2090 cpu = task_cpu(p);
2091 orig_cpu = cpu;
2092 this_cpu = smp_processor_id();
2094 #ifdef CONFIG_SMP
2095 if (unlikely(task_running(rq, p)))
2096 goto out_activate;
2098 cpu = p->sched_class->select_task_rq(p, sync);
2099 if (cpu != orig_cpu) {
2100 set_task_cpu(p, cpu);
2101 task_rq_unlock(rq, &flags);
2102 /* might preempt at this point */
2103 rq = task_rq_lock(p, &flags);
2104 old_state = p->state;
2105 if (!(old_state & state))
2106 goto out;
2107 if (p->se.on_rq)
2108 goto out_running;
2110 this_cpu = smp_processor_id();
2111 cpu = task_cpu(p);
2114 #ifdef CONFIG_SCHEDSTATS
2115 schedstat_inc(rq, ttwu_count);
2116 if (cpu == this_cpu)
2117 schedstat_inc(rq, ttwu_local);
2118 else {
2119 struct sched_domain *sd;
2120 for_each_domain(this_cpu, sd) {
2121 if (cpu_isset(cpu, sd->span)) {
2122 schedstat_inc(sd, ttwu_wake_remote);
2123 break;
2127 #endif
2129 out_activate:
2130 #endif /* CONFIG_SMP */
2131 schedstat_inc(p, se.nr_wakeups);
2132 if (sync)
2133 schedstat_inc(p, se.nr_wakeups_sync);
2134 if (orig_cpu != cpu)
2135 schedstat_inc(p, se.nr_wakeups_migrate);
2136 if (cpu == this_cpu)
2137 schedstat_inc(p, se.nr_wakeups_local);
2138 else
2139 schedstat_inc(p, se.nr_wakeups_remote);
2140 update_rq_clock(rq);
2141 activate_task(rq, p, 1);
2142 success = 1;
2144 out_running:
2145 check_preempt_curr(rq, p);
2147 p->state = TASK_RUNNING;
2148 #ifdef CONFIG_SMP
2149 if (p->sched_class->task_wake_up)
2150 p->sched_class->task_wake_up(rq, p);
2151 #endif
2152 out:
2153 task_rq_unlock(rq, &flags);
2155 return success;
2158 int wake_up_process(struct task_struct *p)
2160 return try_to_wake_up(p, TASK_ALL, 0);
2162 EXPORT_SYMBOL(wake_up_process);
2164 int wake_up_state(struct task_struct *p, unsigned int state)
2166 return try_to_wake_up(p, state, 0);
2170 * Perform scheduler related setup for a newly forked process p.
2171 * p is forked by current.
2173 * __sched_fork() is basic setup used by init_idle() too:
2175 static void __sched_fork(struct task_struct *p)
2177 p->se.exec_start = 0;
2178 p->se.sum_exec_runtime = 0;
2179 p->se.prev_sum_exec_runtime = 0;
2180 p->se.last_wakeup = 0;
2181 p->se.avg_overlap = 0;
2183 #ifdef CONFIG_SCHEDSTATS
2184 p->se.wait_start = 0;
2185 p->se.sum_sleep_runtime = 0;
2186 p->se.sleep_start = 0;
2187 p->se.block_start = 0;
2188 p->se.sleep_max = 0;
2189 p->se.block_max = 0;
2190 p->se.exec_max = 0;
2191 p->se.slice_max = 0;
2192 p->se.wait_max = 0;
2193 #endif
2195 INIT_LIST_HEAD(&p->rt.run_list);
2196 p->se.on_rq = 0;
2197 INIT_LIST_HEAD(&p->se.group_node);
2199 #ifdef CONFIG_PREEMPT_NOTIFIERS
2200 INIT_HLIST_HEAD(&p->preempt_notifiers);
2201 #endif
2204 * We mark the process as running here, but have not actually
2205 * inserted it onto the runqueue yet. This guarantees that
2206 * nobody will actually run it, and a signal or other external
2207 * event cannot wake it up and insert it on the runqueue either.
2209 p->state = TASK_RUNNING;
2213 * fork()/clone()-time setup:
2215 void sched_fork(struct task_struct *p, int clone_flags)
2217 int cpu = get_cpu();
2219 __sched_fork(p);
2221 #ifdef CONFIG_SMP
2222 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2223 #endif
2224 set_task_cpu(p, cpu);
2227 * Make sure we do not leak PI boosting priority to the child:
2229 p->prio = current->normal_prio;
2230 if (!rt_prio(p->prio))
2231 p->sched_class = &fair_sched_class;
2233 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2234 if (likely(sched_info_on()))
2235 memset(&p->sched_info, 0, sizeof(p->sched_info));
2236 #endif
2237 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2238 p->oncpu = 0;
2239 #endif
2240 #ifdef CONFIG_PREEMPT
2241 /* Want to start with kernel preemption disabled. */
2242 task_thread_info(p)->preempt_count = 1;
2243 #endif
2244 put_cpu();
2248 * wake_up_new_task - wake up a newly created task for the first time.
2250 * This function will do some initial scheduler statistics housekeeping
2251 * that must be done for every newly created context, then puts the task
2252 * on the runqueue and wakes it.
2254 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2256 unsigned long flags;
2257 struct rq *rq;
2259 rq = task_rq_lock(p, &flags);
2260 BUG_ON(p->state != TASK_RUNNING);
2261 update_rq_clock(rq);
2263 p->prio = effective_prio(p);
2265 if (!p->sched_class->task_new || !current->se.on_rq) {
2266 activate_task(rq, p, 0);
2267 } else {
2269 * Let the scheduling class do new task startup
2270 * management (if any):
2272 p->sched_class->task_new(rq, p);
2273 inc_nr_running(p, rq);
2275 check_preempt_curr(rq, p);
2276 #ifdef CONFIG_SMP
2277 if (p->sched_class->task_wake_up)
2278 p->sched_class->task_wake_up(rq, p);
2279 #endif
2280 task_rq_unlock(rq, &flags);
2283 #ifdef CONFIG_PREEMPT_NOTIFIERS
2286 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2287 * @notifier: notifier struct to register
2289 void preempt_notifier_register(struct preempt_notifier *notifier)
2291 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2293 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2296 * preempt_notifier_unregister - no longer interested in preemption notifications
2297 * @notifier: notifier struct to unregister
2299 * This is safe to call from within a preemption notifier.
2301 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2303 hlist_del(&notifier->link);
2305 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2307 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2309 struct preempt_notifier *notifier;
2310 struct hlist_node *node;
2312 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2313 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2316 static void
2317 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2318 struct task_struct *next)
2320 struct preempt_notifier *notifier;
2321 struct hlist_node *node;
2323 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2324 notifier->ops->sched_out(notifier, next);
2327 #else
2329 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2333 static void
2334 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2335 struct task_struct *next)
2339 #endif
2342 * prepare_task_switch - prepare to switch tasks
2343 * @rq: the runqueue preparing to switch
2344 * @prev: the current task that is being switched out
2345 * @next: the task we are going to switch to.
2347 * This is called with the rq lock held and interrupts off. It must
2348 * be paired with a subsequent finish_task_switch after the context
2349 * switch.
2351 * prepare_task_switch sets up locking and calls architecture specific
2352 * hooks.
2354 static inline void
2355 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2356 struct task_struct *next)
2358 fire_sched_out_preempt_notifiers(prev, next);
2359 prepare_lock_switch(rq, next);
2360 prepare_arch_switch(next);
2364 * finish_task_switch - clean up after a task-switch
2365 * @rq: runqueue associated with task-switch
2366 * @prev: the thread we just switched away from.
2368 * finish_task_switch must be called after the context switch, paired
2369 * with a prepare_task_switch call before the context switch.
2370 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2371 * and do any other architecture-specific cleanup actions.
2373 * Note that we may have delayed dropping an mm in context_switch(). If
2374 * so, we finish that here outside of the runqueue lock. (Doing it
2375 * with the lock held can cause deadlocks; see schedule() for
2376 * details.)
2378 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2379 __releases(rq->lock)
2381 struct mm_struct *mm = rq->prev_mm;
2382 long prev_state;
2384 rq->prev_mm = NULL;
2387 * A task struct has one reference for the use as "current".
2388 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2389 * schedule one last time. The schedule call will never return, and
2390 * the scheduled task must drop that reference.
2391 * The test for TASK_DEAD must occur while the runqueue locks are
2392 * still held, otherwise prev could be scheduled on another cpu, die
2393 * there before we look at prev->state, and then the reference would
2394 * be dropped twice.
2395 * Manfred Spraul <manfred@colorfullife.com>
2397 prev_state = prev->state;
2398 finish_arch_switch(prev);
2399 finish_lock_switch(rq, prev);
2400 #ifdef CONFIG_SMP
2401 if (current->sched_class->post_schedule)
2402 current->sched_class->post_schedule(rq);
2403 #endif
2405 fire_sched_in_preempt_notifiers(current);
2406 if (mm)
2407 mmdrop(mm);
2408 if (unlikely(prev_state == TASK_DEAD)) {
2410 * Remove function-return probe instances associated with this
2411 * task and put them back on the free list.
2413 kprobe_flush_task(prev);
2414 put_task_struct(prev);
2419 * schedule_tail - first thing a freshly forked thread must call.
2420 * @prev: the thread we just switched away from.
2422 asmlinkage void schedule_tail(struct task_struct *prev)
2423 __releases(rq->lock)
2425 struct rq *rq = this_rq();
2427 finish_task_switch(rq, prev);
2428 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2429 /* In this case, finish_task_switch does not reenable preemption */
2430 preempt_enable();
2431 #endif
2432 if (current->set_child_tid)
2433 put_user(task_pid_vnr(current), current->set_child_tid);
2437 * context_switch - switch to the new MM and the new
2438 * thread's register state.
2440 static inline void
2441 context_switch(struct rq *rq, struct task_struct *prev,
2442 struct task_struct *next)
2444 struct mm_struct *mm, *oldmm;
2446 prepare_task_switch(rq, prev, next);
2447 mm = next->mm;
2448 oldmm = prev->active_mm;
2450 * For paravirt, this is coupled with an exit in switch_to to
2451 * combine the page table reload and the switch backend into
2452 * one hypercall.
2454 arch_enter_lazy_cpu_mode();
2456 if (unlikely(!mm)) {
2457 next->active_mm = oldmm;
2458 atomic_inc(&oldmm->mm_count);
2459 enter_lazy_tlb(oldmm, next);
2460 } else
2461 switch_mm(oldmm, mm, next);
2463 if (unlikely(!prev->mm)) {
2464 prev->active_mm = NULL;
2465 rq->prev_mm = oldmm;
2468 * Since the runqueue lock will be released by the next
2469 * task (which is an invalid locking op but in the case
2470 * of the scheduler it's an obvious special-case), so we
2471 * do an early lockdep release here:
2473 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2474 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2475 #endif
2477 /* Here we just switch the register state and the stack. */
2478 switch_to(prev, next, prev);
2480 barrier();
2482 * this_rq must be evaluated again because prev may have moved
2483 * CPUs since it called schedule(), thus the 'rq' on its stack
2484 * frame will be invalid.
2486 finish_task_switch(this_rq(), prev);
2490 * nr_running, nr_uninterruptible and nr_context_switches:
2492 * externally visible scheduler statistics: current number of runnable
2493 * threads, current number of uninterruptible-sleeping threads, total
2494 * number of context switches performed since bootup.
2496 unsigned long nr_running(void)
2498 unsigned long i, sum = 0;
2500 for_each_online_cpu(i)
2501 sum += cpu_rq(i)->nr_running;
2503 return sum;
2506 unsigned long nr_uninterruptible(void)
2508 unsigned long i, sum = 0;
2510 for_each_possible_cpu(i)
2511 sum += cpu_rq(i)->nr_uninterruptible;
2514 * Since we read the counters lockless, it might be slightly
2515 * inaccurate. Do not allow it to go below zero though:
2517 if (unlikely((long)sum < 0))
2518 sum = 0;
2520 return sum;
2523 unsigned long long nr_context_switches(void)
2525 int i;
2526 unsigned long long sum = 0;
2528 for_each_possible_cpu(i)
2529 sum += cpu_rq(i)->nr_switches;
2531 return sum;
2534 unsigned long nr_iowait(void)
2536 unsigned long i, sum = 0;
2538 for_each_possible_cpu(i)
2539 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2541 return sum;
2544 unsigned long nr_active(void)
2546 unsigned long i, running = 0, uninterruptible = 0;
2548 for_each_online_cpu(i) {
2549 running += cpu_rq(i)->nr_running;
2550 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2553 if (unlikely((long)uninterruptible < 0))
2554 uninterruptible = 0;
2556 return running + uninterruptible;
2560 * Update rq->cpu_load[] statistics. This function is usually called every
2561 * scheduler tick (TICK_NSEC).
2563 static void update_cpu_load(struct rq *this_rq)
2565 unsigned long this_load = this_rq->load.weight;
2566 int i, scale;
2568 this_rq->nr_load_updates++;
2570 /* Update our load: */
2571 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2572 unsigned long old_load, new_load;
2574 /* scale is effectively 1 << i now, and >> i divides by scale */
2576 old_load = this_rq->cpu_load[i];
2577 new_load = this_load;
2579 * Round up the averaging division if load is increasing. This
2580 * prevents us from getting stuck on 9 if the load is 10, for
2581 * example.
2583 if (new_load > old_load)
2584 new_load += scale-1;
2585 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2589 #ifdef CONFIG_SMP
2592 * double_rq_lock - safely lock two runqueues
2594 * Note this does not disable interrupts like task_rq_lock,
2595 * you need to do so manually before calling.
2597 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2598 __acquires(rq1->lock)
2599 __acquires(rq2->lock)
2601 BUG_ON(!irqs_disabled());
2602 if (rq1 == rq2) {
2603 spin_lock(&rq1->lock);
2604 __acquire(rq2->lock); /* Fake it out ;) */
2605 } else {
2606 if (rq1 < rq2) {
2607 spin_lock(&rq1->lock);
2608 spin_lock(&rq2->lock);
2609 } else {
2610 spin_lock(&rq2->lock);
2611 spin_lock(&rq1->lock);
2614 update_rq_clock(rq1);
2615 update_rq_clock(rq2);
2619 * double_rq_unlock - safely unlock two runqueues
2621 * Note this does not restore interrupts like task_rq_unlock,
2622 * you need to do so manually after calling.
2624 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2625 __releases(rq1->lock)
2626 __releases(rq2->lock)
2628 spin_unlock(&rq1->lock);
2629 if (rq1 != rq2)
2630 spin_unlock(&rq2->lock);
2631 else
2632 __release(rq2->lock);
2636 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2638 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2639 __releases(this_rq->lock)
2640 __acquires(busiest->lock)
2641 __acquires(this_rq->lock)
2643 int ret = 0;
2645 if (unlikely(!irqs_disabled())) {
2646 /* printk() doesn't work good under rq->lock */
2647 spin_unlock(&this_rq->lock);
2648 BUG_ON(1);
2650 if (unlikely(!spin_trylock(&busiest->lock))) {
2651 if (busiest < this_rq) {
2652 spin_unlock(&this_rq->lock);
2653 spin_lock(&busiest->lock);
2654 spin_lock(&this_rq->lock);
2655 ret = 1;
2656 } else
2657 spin_lock(&busiest->lock);
2659 return ret;
2663 * If dest_cpu is allowed for this process, migrate the task to it.
2664 * This is accomplished by forcing the cpu_allowed mask to only
2665 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2666 * the cpu_allowed mask is restored.
2668 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2670 struct migration_req req;
2671 unsigned long flags;
2672 struct rq *rq;
2674 rq = task_rq_lock(p, &flags);
2675 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2676 || unlikely(cpu_is_offline(dest_cpu)))
2677 goto out;
2679 /* force the process onto the specified CPU */
2680 if (migrate_task(p, dest_cpu, &req)) {
2681 /* Need to wait for migration thread (might exit: take ref). */
2682 struct task_struct *mt = rq->migration_thread;
2684 get_task_struct(mt);
2685 task_rq_unlock(rq, &flags);
2686 wake_up_process(mt);
2687 put_task_struct(mt);
2688 wait_for_completion(&req.done);
2690 return;
2692 out:
2693 task_rq_unlock(rq, &flags);
2697 * sched_exec - execve() is a valuable balancing opportunity, because at
2698 * this point the task has the smallest effective memory and cache footprint.
2700 void sched_exec(void)
2702 int new_cpu, this_cpu = get_cpu();
2703 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2704 put_cpu();
2705 if (new_cpu != this_cpu)
2706 sched_migrate_task(current, new_cpu);
2710 * pull_task - move a task from a remote runqueue to the local runqueue.
2711 * Both runqueues must be locked.
2713 static void pull_task(struct rq *src_rq, struct task_struct *p,
2714 struct rq *this_rq, int this_cpu)
2716 deactivate_task(src_rq, p, 0);
2717 set_task_cpu(p, this_cpu);
2718 activate_task(this_rq, p, 0);
2720 * Note that idle threads have a prio of MAX_PRIO, for this test
2721 * to be always true for them.
2723 check_preempt_curr(this_rq, p);
2727 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2729 static
2730 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2731 struct sched_domain *sd, enum cpu_idle_type idle,
2732 int *all_pinned)
2735 * We do not migrate tasks that are:
2736 * 1) running (obviously), or
2737 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2738 * 3) are cache-hot on their current CPU.
2740 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2741 schedstat_inc(p, se.nr_failed_migrations_affine);
2742 return 0;
2744 *all_pinned = 0;
2746 if (task_running(rq, p)) {
2747 schedstat_inc(p, se.nr_failed_migrations_running);
2748 return 0;
2752 * Aggressive migration if:
2753 * 1) task is cache cold, or
2754 * 2) too many balance attempts have failed.
2757 if (!task_hot(p, rq->clock, sd) ||
2758 sd->nr_balance_failed > sd->cache_nice_tries) {
2759 #ifdef CONFIG_SCHEDSTATS
2760 if (task_hot(p, rq->clock, sd)) {
2761 schedstat_inc(sd, lb_hot_gained[idle]);
2762 schedstat_inc(p, se.nr_forced_migrations);
2764 #endif
2765 return 1;
2768 if (task_hot(p, rq->clock, sd)) {
2769 schedstat_inc(p, se.nr_failed_migrations_hot);
2770 return 0;
2772 return 1;
2775 static unsigned long
2776 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2777 unsigned long max_load_move, struct sched_domain *sd,
2778 enum cpu_idle_type idle, int *all_pinned,
2779 int *this_best_prio, struct rq_iterator *iterator)
2781 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2782 struct task_struct *p;
2783 long rem_load_move = max_load_move;
2785 if (max_load_move == 0)
2786 goto out;
2788 pinned = 1;
2791 * Start the load-balancing iterator:
2793 p = iterator->start(iterator->arg);
2794 next:
2795 if (!p || loops++ > sysctl_sched_nr_migrate)
2796 goto out;
2798 * To help distribute high priority tasks across CPUs we don't
2799 * skip a task if it will be the highest priority task (i.e. smallest
2800 * prio value) on its new queue regardless of its load weight
2802 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2803 SCHED_LOAD_SCALE_FUZZ;
2804 if ((skip_for_load && p->prio >= *this_best_prio) ||
2805 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2806 p = iterator->next(iterator->arg);
2807 goto next;
2810 pull_task(busiest, p, this_rq, this_cpu);
2811 pulled++;
2812 rem_load_move -= p->se.load.weight;
2815 * We only want to steal up to the prescribed amount of weighted load.
2817 if (rem_load_move > 0) {
2818 if (p->prio < *this_best_prio)
2819 *this_best_prio = p->prio;
2820 p = iterator->next(iterator->arg);
2821 goto next;
2823 out:
2825 * Right now, this is one of only two places pull_task() is called,
2826 * so we can safely collect pull_task() stats here rather than
2827 * inside pull_task().
2829 schedstat_add(sd, lb_gained[idle], pulled);
2831 if (all_pinned)
2832 *all_pinned = pinned;
2834 return max_load_move - rem_load_move;
2838 * move_tasks tries to move up to max_load_move weighted load from busiest to
2839 * this_rq, as part of a balancing operation within domain "sd".
2840 * Returns 1 if successful and 0 otherwise.
2842 * Called with both runqueues locked.
2844 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2845 unsigned long max_load_move,
2846 struct sched_domain *sd, enum cpu_idle_type idle,
2847 int *all_pinned)
2849 const struct sched_class *class = sched_class_highest;
2850 unsigned long total_load_moved = 0;
2851 int this_best_prio = this_rq->curr->prio;
2853 do {
2854 total_load_moved +=
2855 class->load_balance(this_rq, this_cpu, busiest,
2856 max_load_move - total_load_moved,
2857 sd, idle, all_pinned, &this_best_prio);
2858 class = class->next;
2859 } while (class && max_load_move > total_load_moved);
2861 return total_load_moved > 0;
2864 static int
2865 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2866 struct sched_domain *sd, enum cpu_idle_type idle,
2867 struct rq_iterator *iterator)
2869 struct task_struct *p = iterator->start(iterator->arg);
2870 int pinned = 0;
2872 while (p) {
2873 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2874 pull_task(busiest, p, this_rq, this_cpu);
2876 * Right now, this is only the second place pull_task()
2877 * is called, so we can safely collect pull_task()
2878 * stats here rather than inside pull_task().
2880 schedstat_inc(sd, lb_gained[idle]);
2882 return 1;
2884 p = iterator->next(iterator->arg);
2887 return 0;
2891 * move_one_task tries to move exactly one task from busiest to this_rq, as
2892 * part of active balancing operations within "domain".
2893 * Returns 1 if successful and 0 otherwise.
2895 * Called with both runqueues locked.
2897 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2898 struct sched_domain *sd, enum cpu_idle_type idle)
2900 const struct sched_class *class;
2902 for (class = sched_class_highest; class; class = class->next)
2903 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2904 return 1;
2906 return 0;
2910 * find_busiest_group finds and returns the busiest CPU group within the
2911 * domain. It calculates and returns the amount of weighted load which
2912 * should be moved to restore balance via the imbalance parameter.
2914 static struct sched_group *
2915 find_busiest_group(struct sched_domain *sd, int this_cpu,
2916 unsigned long *imbalance, enum cpu_idle_type idle,
2917 int *sd_idle, const cpumask_t *cpus, int *balance)
2919 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2920 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2921 unsigned long max_pull;
2922 unsigned long busiest_load_per_task, busiest_nr_running;
2923 unsigned long this_load_per_task, this_nr_running;
2924 int load_idx, group_imb = 0;
2925 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2926 int power_savings_balance = 1;
2927 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2928 unsigned long min_nr_running = ULONG_MAX;
2929 struct sched_group *group_min = NULL, *group_leader = NULL;
2930 #endif
2932 max_load = this_load = total_load = total_pwr = 0;
2933 busiest_load_per_task = busiest_nr_running = 0;
2934 this_load_per_task = this_nr_running = 0;
2935 if (idle == CPU_NOT_IDLE)
2936 load_idx = sd->busy_idx;
2937 else if (idle == CPU_NEWLY_IDLE)
2938 load_idx = sd->newidle_idx;
2939 else
2940 load_idx = sd->idle_idx;
2942 do {
2943 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2944 int local_group;
2945 int i;
2946 int __group_imb = 0;
2947 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2948 unsigned long sum_nr_running, sum_weighted_load;
2950 local_group = cpu_isset(this_cpu, group->cpumask);
2952 if (local_group)
2953 balance_cpu = first_cpu(group->cpumask);
2955 /* Tally up the load of all CPUs in the group */
2956 sum_weighted_load = sum_nr_running = avg_load = 0;
2957 max_cpu_load = 0;
2958 min_cpu_load = ~0UL;
2960 for_each_cpu_mask(i, group->cpumask) {
2961 struct rq *rq;
2963 if (!cpu_isset(i, *cpus))
2964 continue;
2966 rq = cpu_rq(i);
2968 if (*sd_idle && rq->nr_running)
2969 *sd_idle = 0;
2971 /* Bias balancing toward cpus of our domain */
2972 if (local_group) {
2973 if (idle_cpu(i) && !first_idle_cpu) {
2974 first_idle_cpu = 1;
2975 balance_cpu = i;
2978 load = target_load(i, load_idx);
2979 } else {
2980 load = source_load(i, load_idx);
2981 if (load > max_cpu_load)
2982 max_cpu_load = load;
2983 if (min_cpu_load > load)
2984 min_cpu_load = load;
2987 avg_load += load;
2988 sum_nr_running += rq->nr_running;
2989 sum_weighted_load += weighted_cpuload(i);
2993 * First idle cpu or the first cpu(busiest) in this sched group
2994 * is eligible for doing load balancing at this and above
2995 * domains. In the newly idle case, we will allow all the cpu's
2996 * to do the newly idle load balance.
2998 if (idle != CPU_NEWLY_IDLE && local_group &&
2999 balance_cpu != this_cpu && balance) {
3000 *balance = 0;
3001 goto ret;
3004 total_load += avg_load;
3005 total_pwr += group->__cpu_power;
3007 /* Adjust by relative CPU power of the group */
3008 avg_load = sg_div_cpu_power(group,
3009 avg_load * SCHED_LOAD_SCALE);
3011 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3012 __group_imb = 1;
3014 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3016 if (local_group) {
3017 this_load = avg_load;
3018 this = group;
3019 this_nr_running = sum_nr_running;
3020 this_load_per_task = sum_weighted_load;
3021 } else if (avg_load > max_load &&
3022 (sum_nr_running > group_capacity || __group_imb)) {
3023 max_load = avg_load;
3024 busiest = group;
3025 busiest_nr_running = sum_nr_running;
3026 busiest_load_per_task = sum_weighted_load;
3027 group_imb = __group_imb;
3030 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3032 * Busy processors will not participate in power savings
3033 * balance.
3035 if (idle == CPU_NOT_IDLE ||
3036 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3037 goto group_next;
3040 * If the local group is idle or completely loaded
3041 * no need to do power savings balance at this domain
3043 if (local_group && (this_nr_running >= group_capacity ||
3044 !this_nr_running))
3045 power_savings_balance = 0;
3048 * If a group is already running at full capacity or idle,
3049 * don't include that group in power savings calculations
3051 if (!power_savings_balance || sum_nr_running >= group_capacity
3052 || !sum_nr_running)
3053 goto group_next;
3056 * Calculate the group which has the least non-idle load.
3057 * This is the group from where we need to pick up the load
3058 * for saving power
3060 if ((sum_nr_running < min_nr_running) ||
3061 (sum_nr_running == min_nr_running &&
3062 first_cpu(group->cpumask) <
3063 first_cpu(group_min->cpumask))) {
3064 group_min = group;
3065 min_nr_running = sum_nr_running;
3066 min_load_per_task = sum_weighted_load /
3067 sum_nr_running;
3071 * Calculate the group which is almost near its
3072 * capacity but still has some space to pick up some load
3073 * from other group and save more power
3075 if (sum_nr_running <= group_capacity - 1) {
3076 if (sum_nr_running > leader_nr_running ||
3077 (sum_nr_running == leader_nr_running &&
3078 first_cpu(group->cpumask) >
3079 first_cpu(group_leader->cpumask))) {
3080 group_leader = group;
3081 leader_nr_running = sum_nr_running;
3084 group_next:
3085 #endif
3086 group = group->next;
3087 } while (group != sd->groups);
3089 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3090 goto out_balanced;
3092 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3094 if (this_load >= avg_load ||
3095 100*max_load <= sd->imbalance_pct*this_load)
3096 goto out_balanced;
3098 busiest_load_per_task /= busiest_nr_running;
3099 if (group_imb)
3100 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3103 * We're trying to get all the cpus to the average_load, so we don't
3104 * want to push ourselves above the average load, nor do we wish to
3105 * reduce the max loaded cpu below the average load, as either of these
3106 * actions would just result in more rebalancing later, and ping-pong
3107 * tasks around. Thus we look for the minimum possible imbalance.
3108 * Negative imbalances (*we* are more loaded than anyone else) will
3109 * be counted as no imbalance for these purposes -- we can't fix that
3110 * by pulling tasks to us. Be careful of negative numbers as they'll
3111 * appear as very large values with unsigned longs.
3113 if (max_load <= busiest_load_per_task)
3114 goto out_balanced;
3117 * In the presence of smp nice balancing, certain scenarios can have
3118 * max load less than avg load(as we skip the groups at or below
3119 * its cpu_power, while calculating max_load..)
3121 if (max_load < avg_load) {
3122 *imbalance = 0;
3123 goto small_imbalance;
3126 /* Don't want to pull so many tasks that a group would go idle */
3127 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3129 /* How much load to actually move to equalise the imbalance */
3130 *imbalance = min(max_pull * busiest->__cpu_power,
3131 (avg_load - this_load) * this->__cpu_power)
3132 / SCHED_LOAD_SCALE;
3135 * if *imbalance is less than the average load per runnable task
3136 * there is no gaurantee that any tasks will be moved so we'll have
3137 * a think about bumping its value to force at least one task to be
3138 * moved
3140 if (*imbalance < busiest_load_per_task) {
3141 unsigned long tmp, pwr_now, pwr_move;
3142 unsigned int imbn;
3144 small_imbalance:
3145 pwr_move = pwr_now = 0;
3146 imbn = 2;
3147 if (this_nr_running) {
3148 this_load_per_task /= this_nr_running;
3149 if (busiest_load_per_task > this_load_per_task)
3150 imbn = 1;
3151 } else
3152 this_load_per_task = SCHED_LOAD_SCALE;
3154 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3155 busiest_load_per_task * imbn) {
3156 *imbalance = busiest_load_per_task;
3157 return busiest;
3161 * OK, we don't have enough imbalance to justify moving tasks,
3162 * however we may be able to increase total CPU power used by
3163 * moving them.
3166 pwr_now += busiest->__cpu_power *
3167 min(busiest_load_per_task, max_load);
3168 pwr_now += this->__cpu_power *
3169 min(this_load_per_task, this_load);
3170 pwr_now /= SCHED_LOAD_SCALE;
3172 /* Amount of load we'd subtract */
3173 tmp = sg_div_cpu_power(busiest,
3174 busiest_load_per_task * SCHED_LOAD_SCALE);
3175 if (max_load > tmp)
3176 pwr_move += busiest->__cpu_power *
3177 min(busiest_load_per_task, max_load - tmp);
3179 /* Amount of load we'd add */
3180 if (max_load * busiest->__cpu_power <
3181 busiest_load_per_task * SCHED_LOAD_SCALE)
3182 tmp = sg_div_cpu_power(this,
3183 max_load * busiest->__cpu_power);
3184 else
3185 tmp = sg_div_cpu_power(this,
3186 busiest_load_per_task * SCHED_LOAD_SCALE);
3187 pwr_move += this->__cpu_power *
3188 min(this_load_per_task, this_load + tmp);
3189 pwr_move /= SCHED_LOAD_SCALE;
3191 /* Move if we gain throughput */
3192 if (pwr_move > pwr_now)
3193 *imbalance = busiest_load_per_task;
3196 return busiest;
3198 out_balanced:
3199 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3200 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3201 goto ret;
3203 if (this == group_leader && group_leader != group_min) {
3204 *imbalance = min_load_per_task;
3205 return group_min;
3207 #endif
3208 ret:
3209 *imbalance = 0;
3210 return NULL;
3214 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3216 static struct rq *
3217 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3218 unsigned long imbalance, const cpumask_t *cpus)
3220 struct rq *busiest = NULL, *rq;
3221 unsigned long max_load = 0;
3222 int i;
3224 for_each_cpu_mask(i, group->cpumask) {
3225 unsigned long wl;
3227 if (!cpu_isset(i, *cpus))
3228 continue;
3230 rq = cpu_rq(i);
3231 wl = weighted_cpuload(i);
3233 if (rq->nr_running == 1 && wl > imbalance)
3234 continue;
3236 if (wl > max_load) {
3237 max_load = wl;
3238 busiest = rq;
3242 return busiest;
3246 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3247 * so long as it is large enough.
3249 #define MAX_PINNED_INTERVAL 512
3252 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3253 * tasks if there is an imbalance.
3255 static int load_balance(int this_cpu, struct rq *this_rq,
3256 struct sched_domain *sd, enum cpu_idle_type idle,
3257 int *balance, cpumask_t *cpus)
3259 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3260 struct sched_group *group;
3261 unsigned long imbalance;
3262 struct rq *busiest;
3263 unsigned long flags;
3265 cpus_setall(*cpus);
3268 * When power savings policy is enabled for the parent domain, idle
3269 * sibling can pick up load irrespective of busy siblings. In this case,
3270 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3271 * portraying it as CPU_NOT_IDLE.
3273 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3274 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3275 sd_idle = 1;
3277 schedstat_inc(sd, lb_count[idle]);
3279 redo:
3280 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3281 cpus, balance);
3283 if (*balance == 0)
3284 goto out_balanced;
3286 if (!group) {
3287 schedstat_inc(sd, lb_nobusyg[idle]);
3288 goto out_balanced;
3291 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3292 if (!busiest) {
3293 schedstat_inc(sd, lb_nobusyq[idle]);
3294 goto out_balanced;
3297 BUG_ON(busiest == this_rq);
3299 schedstat_add(sd, lb_imbalance[idle], imbalance);
3301 ld_moved = 0;
3302 if (busiest->nr_running > 1) {
3304 * Attempt to move tasks. If find_busiest_group has found
3305 * an imbalance but busiest->nr_running <= 1, the group is
3306 * still unbalanced. ld_moved simply stays zero, so it is
3307 * correctly treated as an imbalance.
3309 local_irq_save(flags);
3310 double_rq_lock(this_rq, busiest);
3311 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3312 imbalance, sd, idle, &all_pinned);
3313 double_rq_unlock(this_rq, busiest);
3314 local_irq_restore(flags);
3317 * some other cpu did the load balance for us.
3319 if (ld_moved && this_cpu != smp_processor_id())
3320 resched_cpu(this_cpu);
3322 /* All tasks on this runqueue were pinned by CPU affinity */
3323 if (unlikely(all_pinned)) {
3324 cpu_clear(cpu_of(busiest), *cpus);
3325 if (!cpus_empty(*cpus))
3326 goto redo;
3327 goto out_balanced;
3331 if (!ld_moved) {
3332 schedstat_inc(sd, lb_failed[idle]);
3333 sd->nr_balance_failed++;
3335 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3337 spin_lock_irqsave(&busiest->lock, flags);
3339 /* don't kick the migration_thread, if the curr
3340 * task on busiest cpu can't be moved to this_cpu
3342 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3343 spin_unlock_irqrestore(&busiest->lock, flags);
3344 all_pinned = 1;
3345 goto out_one_pinned;
3348 if (!busiest->active_balance) {
3349 busiest->active_balance = 1;
3350 busiest->push_cpu = this_cpu;
3351 active_balance = 1;
3353 spin_unlock_irqrestore(&busiest->lock, flags);
3354 if (active_balance)
3355 wake_up_process(busiest->migration_thread);
3358 * We've kicked active balancing, reset the failure
3359 * counter.
3361 sd->nr_balance_failed = sd->cache_nice_tries+1;
3363 } else
3364 sd->nr_balance_failed = 0;
3366 if (likely(!active_balance)) {
3367 /* We were unbalanced, so reset the balancing interval */
3368 sd->balance_interval = sd->min_interval;
3369 } else {
3371 * If we've begun active balancing, start to back off. This
3372 * case may not be covered by the all_pinned logic if there
3373 * is only 1 task on the busy runqueue (because we don't call
3374 * move_tasks).
3376 if (sd->balance_interval < sd->max_interval)
3377 sd->balance_interval *= 2;
3380 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3381 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3382 return -1;
3383 return ld_moved;
3385 out_balanced:
3386 schedstat_inc(sd, lb_balanced[idle]);
3388 sd->nr_balance_failed = 0;
3390 out_one_pinned:
3391 /* tune up the balancing interval */
3392 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3393 (sd->balance_interval < sd->max_interval))
3394 sd->balance_interval *= 2;
3396 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3397 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3398 return -1;
3399 return 0;
3403 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3404 * tasks if there is an imbalance.
3406 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3407 * this_rq is locked.
3409 static int
3410 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3411 cpumask_t *cpus)
3413 struct sched_group *group;
3414 struct rq *busiest = NULL;
3415 unsigned long imbalance;
3416 int ld_moved = 0;
3417 int sd_idle = 0;
3418 int all_pinned = 0;
3420 cpus_setall(*cpus);
3423 * When power savings policy is enabled for the parent domain, idle
3424 * sibling can pick up load irrespective of busy siblings. In this case,
3425 * let the state of idle sibling percolate up as IDLE, instead of
3426 * portraying it as CPU_NOT_IDLE.
3428 if (sd->flags & SD_SHARE_CPUPOWER &&
3429 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3430 sd_idle = 1;
3432 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3433 redo:
3434 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3435 &sd_idle, cpus, NULL);
3436 if (!group) {
3437 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3438 goto out_balanced;
3441 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3442 if (!busiest) {
3443 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3444 goto out_balanced;
3447 BUG_ON(busiest == this_rq);
3449 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3451 ld_moved = 0;
3452 if (busiest->nr_running > 1) {
3453 /* Attempt to move tasks */
3454 double_lock_balance(this_rq, busiest);
3455 /* this_rq->clock is already updated */
3456 update_rq_clock(busiest);
3457 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3458 imbalance, sd, CPU_NEWLY_IDLE,
3459 &all_pinned);
3460 spin_unlock(&busiest->lock);
3462 if (unlikely(all_pinned)) {
3463 cpu_clear(cpu_of(busiest), *cpus);
3464 if (!cpus_empty(*cpus))
3465 goto redo;
3469 if (!ld_moved) {
3470 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3471 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3472 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3473 return -1;
3474 } else
3475 sd->nr_balance_failed = 0;
3477 return ld_moved;
3479 out_balanced:
3480 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3481 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3482 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3483 return -1;
3484 sd->nr_balance_failed = 0;
3486 return 0;
3490 * idle_balance is called by schedule() if this_cpu is about to become
3491 * idle. Attempts to pull tasks from other CPUs.
3493 static void idle_balance(int this_cpu, struct rq *this_rq)
3495 struct sched_domain *sd;
3496 int pulled_task = -1;
3497 unsigned long next_balance = jiffies + HZ;
3498 cpumask_t tmpmask;
3500 for_each_domain(this_cpu, sd) {
3501 unsigned long interval;
3503 if (!(sd->flags & SD_LOAD_BALANCE))
3504 continue;
3506 if (sd->flags & SD_BALANCE_NEWIDLE)
3507 /* If we've pulled tasks over stop searching: */
3508 pulled_task = load_balance_newidle(this_cpu, this_rq,
3509 sd, &tmpmask);
3511 interval = msecs_to_jiffies(sd->balance_interval);
3512 if (time_after(next_balance, sd->last_balance + interval))
3513 next_balance = sd->last_balance + interval;
3514 if (pulled_task)
3515 break;
3517 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3519 * We are going idle. next_balance may be set based on
3520 * a busy processor. So reset next_balance.
3522 this_rq->next_balance = next_balance;
3527 * active_load_balance is run by migration threads. It pushes running tasks
3528 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3529 * running on each physical CPU where possible, and avoids physical /
3530 * logical imbalances.
3532 * Called with busiest_rq locked.
3534 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3536 int target_cpu = busiest_rq->push_cpu;
3537 struct sched_domain *sd;
3538 struct rq *target_rq;
3540 /* Is there any task to move? */
3541 if (busiest_rq->nr_running <= 1)
3542 return;
3544 target_rq = cpu_rq(target_cpu);
3547 * This condition is "impossible", if it occurs
3548 * we need to fix it. Originally reported by
3549 * Bjorn Helgaas on a 128-cpu setup.
3551 BUG_ON(busiest_rq == target_rq);
3553 /* move a task from busiest_rq to target_rq */
3554 double_lock_balance(busiest_rq, target_rq);
3555 update_rq_clock(busiest_rq);
3556 update_rq_clock(target_rq);
3558 /* Search for an sd spanning us and the target CPU. */
3559 for_each_domain(target_cpu, sd) {
3560 if ((sd->flags & SD_LOAD_BALANCE) &&
3561 cpu_isset(busiest_cpu, sd->span))
3562 break;
3565 if (likely(sd)) {
3566 schedstat_inc(sd, alb_count);
3568 if (move_one_task(target_rq, target_cpu, busiest_rq,
3569 sd, CPU_IDLE))
3570 schedstat_inc(sd, alb_pushed);
3571 else
3572 schedstat_inc(sd, alb_failed);
3574 spin_unlock(&target_rq->lock);
3577 #ifdef CONFIG_NO_HZ
3578 static struct {
3579 atomic_t load_balancer;
3580 cpumask_t cpu_mask;
3581 } nohz ____cacheline_aligned = {
3582 .load_balancer = ATOMIC_INIT(-1),
3583 .cpu_mask = CPU_MASK_NONE,
3587 * This routine will try to nominate the ilb (idle load balancing)
3588 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3589 * load balancing on behalf of all those cpus. If all the cpus in the system
3590 * go into this tickless mode, then there will be no ilb owner (as there is
3591 * no need for one) and all the cpus will sleep till the next wakeup event
3592 * arrives...
3594 * For the ilb owner, tick is not stopped. And this tick will be used
3595 * for idle load balancing. ilb owner will still be part of
3596 * nohz.cpu_mask..
3598 * While stopping the tick, this cpu will become the ilb owner if there
3599 * is no other owner. And will be the owner till that cpu becomes busy
3600 * or if all cpus in the system stop their ticks at which point
3601 * there is no need for ilb owner.
3603 * When the ilb owner becomes busy, it nominates another owner, during the
3604 * next busy scheduler_tick()
3606 int select_nohz_load_balancer(int stop_tick)
3608 int cpu = smp_processor_id();
3610 if (stop_tick) {
3611 cpu_set(cpu, nohz.cpu_mask);
3612 cpu_rq(cpu)->in_nohz_recently = 1;
3615 * If we are going offline and still the leader, give up!
3617 if (cpu_is_offline(cpu) &&
3618 atomic_read(&nohz.load_balancer) == cpu) {
3619 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3620 BUG();
3621 return 0;
3624 /* time for ilb owner also to sleep */
3625 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3626 if (atomic_read(&nohz.load_balancer) == cpu)
3627 atomic_set(&nohz.load_balancer, -1);
3628 return 0;
3631 if (atomic_read(&nohz.load_balancer) == -1) {
3632 /* make me the ilb owner */
3633 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3634 return 1;
3635 } else if (atomic_read(&nohz.load_balancer) == cpu)
3636 return 1;
3637 } else {
3638 if (!cpu_isset(cpu, nohz.cpu_mask))
3639 return 0;
3641 cpu_clear(cpu, nohz.cpu_mask);
3643 if (atomic_read(&nohz.load_balancer) == cpu)
3644 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3645 BUG();
3647 return 0;
3649 #endif
3651 static DEFINE_SPINLOCK(balancing);
3654 * It checks each scheduling domain to see if it is due to be balanced,
3655 * and initiates a balancing operation if so.
3657 * Balancing parameters are set up in arch_init_sched_domains.
3659 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3661 int balance = 1;
3662 struct rq *rq = cpu_rq(cpu);
3663 unsigned long interval;
3664 struct sched_domain *sd;
3665 /* Earliest time when we have to do rebalance again */
3666 unsigned long next_balance = jiffies + 60*HZ;
3667 int update_next_balance = 0;
3668 cpumask_t tmp;
3670 for_each_domain(cpu, sd) {
3671 if (!(sd->flags & SD_LOAD_BALANCE))
3672 continue;
3674 interval = sd->balance_interval;
3675 if (idle != CPU_IDLE)
3676 interval *= sd->busy_factor;
3678 /* scale ms to jiffies */
3679 interval = msecs_to_jiffies(interval);
3680 if (unlikely(!interval))
3681 interval = 1;
3682 if (interval > HZ*NR_CPUS/10)
3683 interval = HZ*NR_CPUS/10;
3686 if (sd->flags & SD_SERIALIZE) {
3687 if (!spin_trylock(&balancing))
3688 goto out;
3691 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3692 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3694 * We've pulled tasks over so either we're no
3695 * longer idle, or one of our SMT siblings is
3696 * not idle.
3698 idle = CPU_NOT_IDLE;
3700 sd->last_balance = jiffies;
3702 if (sd->flags & SD_SERIALIZE)
3703 spin_unlock(&balancing);
3704 out:
3705 if (time_after(next_balance, sd->last_balance + interval)) {
3706 next_balance = sd->last_balance + interval;
3707 update_next_balance = 1;
3711 * Stop the load balance at this level. There is another
3712 * CPU in our sched group which is doing load balancing more
3713 * actively.
3715 if (!balance)
3716 break;
3720 * next_balance will be updated only when there is a need.
3721 * When the cpu is attached to null domain for ex, it will not be
3722 * updated.
3724 if (likely(update_next_balance))
3725 rq->next_balance = next_balance;
3729 * run_rebalance_domains is triggered when needed from the scheduler tick.
3730 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3731 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3733 static void run_rebalance_domains(struct softirq_action *h)
3735 int this_cpu = smp_processor_id();
3736 struct rq *this_rq = cpu_rq(this_cpu);
3737 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3738 CPU_IDLE : CPU_NOT_IDLE;
3740 rebalance_domains(this_cpu, idle);
3742 #ifdef CONFIG_NO_HZ
3744 * If this cpu is the owner for idle load balancing, then do the
3745 * balancing on behalf of the other idle cpus whose ticks are
3746 * stopped.
3748 if (this_rq->idle_at_tick &&
3749 atomic_read(&nohz.load_balancer) == this_cpu) {
3750 cpumask_t cpus = nohz.cpu_mask;
3751 struct rq *rq;
3752 int balance_cpu;
3754 cpu_clear(this_cpu, cpus);
3755 for_each_cpu_mask(balance_cpu, cpus) {
3757 * If this cpu gets work to do, stop the load balancing
3758 * work being done for other cpus. Next load
3759 * balancing owner will pick it up.
3761 if (need_resched())
3762 break;
3764 rebalance_domains(balance_cpu, CPU_IDLE);
3766 rq = cpu_rq(balance_cpu);
3767 if (time_after(this_rq->next_balance, rq->next_balance))
3768 this_rq->next_balance = rq->next_balance;
3771 #endif
3775 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3777 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3778 * idle load balancing owner or decide to stop the periodic load balancing,
3779 * if the whole system is idle.
3781 static inline void trigger_load_balance(struct rq *rq, int cpu)
3783 #ifdef CONFIG_NO_HZ
3785 * If we were in the nohz mode recently and busy at the current
3786 * scheduler tick, then check if we need to nominate new idle
3787 * load balancer.
3789 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3790 rq->in_nohz_recently = 0;
3792 if (atomic_read(&nohz.load_balancer) == cpu) {
3793 cpu_clear(cpu, nohz.cpu_mask);
3794 atomic_set(&nohz.load_balancer, -1);
3797 if (atomic_read(&nohz.load_balancer) == -1) {
3799 * simple selection for now: Nominate the
3800 * first cpu in the nohz list to be the next
3801 * ilb owner.
3803 * TBD: Traverse the sched domains and nominate
3804 * the nearest cpu in the nohz.cpu_mask.
3806 int ilb = first_cpu(nohz.cpu_mask);
3808 if (ilb < nr_cpu_ids)
3809 resched_cpu(ilb);
3814 * If this cpu is idle and doing idle load balancing for all the
3815 * cpus with ticks stopped, is it time for that to stop?
3817 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3818 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3819 resched_cpu(cpu);
3820 return;
3824 * If this cpu is idle and the idle load balancing is done by
3825 * someone else, then no need raise the SCHED_SOFTIRQ
3827 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3828 cpu_isset(cpu, nohz.cpu_mask))
3829 return;
3830 #endif
3831 if (time_after_eq(jiffies, rq->next_balance))
3832 raise_softirq(SCHED_SOFTIRQ);
3835 #else /* CONFIG_SMP */
3838 * on UP we do not need to balance between CPUs:
3840 static inline void idle_balance(int cpu, struct rq *rq)
3844 #endif
3846 DEFINE_PER_CPU(struct kernel_stat, kstat);
3848 EXPORT_PER_CPU_SYMBOL(kstat);
3851 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3852 * that have not yet been banked in case the task is currently running.
3854 unsigned long long task_sched_runtime(struct task_struct *p)
3856 unsigned long flags;
3857 u64 ns, delta_exec;
3858 struct rq *rq;
3860 rq = task_rq_lock(p, &flags);
3861 ns = p->se.sum_exec_runtime;
3862 if (task_current(rq, p)) {
3863 update_rq_clock(rq);
3864 delta_exec = rq->clock - p->se.exec_start;
3865 if ((s64)delta_exec > 0)
3866 ns += delta_exec;
3868 task_rq_unlock(rq, &flags);
3870 return ns;
3874 * Account user cpu time to a process.
3875 * @p: the process that the cpu time gets accounted to
3876 * @cputime: the cpu time spent in user space since the last update
3878 void account_user_time(struct task_struct *p, cputime_t cputime)
3880 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3881 cputime64_t tmp;
3883 p->utime = cputime_add(p->utime, cputime);
3885 /* Add user time to cpustat. */
3886 tmp = cputime_to_cputime64(cputime);
3887 if (TASK_NICE(p) > 0)
3888 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3889 else
3890 cpustat->user = cputime64_add(cpustat->user, tmp);
3894 * Account guest cpu time to a process.
3895 * @p: the process that the cpu time gets accounted to
3896 * @cputime: the cpu time spent in virtual machine since the last update
3898 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3900 cputime64_t tmp;
3901 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3903 tmp = cputime_to_cputime64(cputime);
3905 p->utime = cputime_add(p->utime, cputime);
3906 p->gtime = cputime_add(p->gtime, cputime);
3908 cpustat->user = cputime64_add(cpustat->user, tmp);
3909 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3913 * Account scaled user cpu time to a process.
3914 * @p: the process that the cpu time gets accounted to
3915 * @cputime: the cpu time spent in user space since the last update
3917 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3919 p->utimescaled = cputime_add(p->utimescaled, cputime);
3923 * Account system cpu time to a process.
3924 * @p: the process that the cpu time gets accounted to
3925 * @hardirq_offset: the offset to subtract from hardirq_count()
3926 * @cputime: the cpu time spent in kernel space since the last update
3928 void account_system_time(struct task_struct *p, int hardirq_offset,
3929 cputime_t cputime)
3931 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3932 struct rq *rq = this_rq();
3933 cputime64_t tmp;
3935 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3936 account_guest_time(p, cputime);
3937 return;
3940 p->stime = cputime_add(p->stime, cputime);
3942 /* Add system time to cpustat. */
3943 tmp = cputime_to_cputime64(cputime);
3944 if (hardirq_count() - hardirq_offset)
3945 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3946 else if (softirq_count())
3947 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3948 else if (p != rq->idle)
3949 cpustat->system = cputime64_add(cpustat->system, tmp);
3950 else if (atomic_read(&rq->nr_iowait) > 0)
3951 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3952 else
3953 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3954 /* Account for system time used */
3955 acct_update_integrals(p);
3959 * Account scaled system cpu time to a process.
3960 * @p: the process that the cpu time gets accounted to
3961 * @hardirq_offset: the offset to subtract from hardirq_count()
3962 * @cputime: the cpu time spent in kernel space since the last update
3964 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3966 p->stimescaled = cputime_add(p->stimescaled, cputime);
3970 * Account for involuntary wait time.
3971 * @p: the process from which the cpu time has been stolen
3972 * @steal: the cpu time spent in involuntary wait
3974 void account_steal_time(struct task_struct *p, cputime_t steal)
3976 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3977 cputime64_t tmp = cputime_to_cputime64(steal);
3978 struct rq *rq = this_rq();
3980 if (p == rq->idle) {
3981 p->stime = cputime_add(p->stime, steal);
3982 if (atomic_read(&rq->nr_iowait) > 0)
3983 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3984 else
3985 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3986 } else
3987 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3991 * This function gets called by the timer code, with HZ frequency.
3992 * We call it with interrupts disabled.
3994 * It also gets called by the fork code, when changing the parent's
3995 * timeslices.
3997 void scheduler_tick(void)
3999 int cpu = smp_processor_id();
4000 struct rq *rq = cpu_rq(cpu);
4001 struct task_struct *curr = rq->curr;
4003 sched_clock_tick();
4005 spin_lock(&rq->lock);
4006 update_rq_clock(rq);
4007 update_cpu_load(rq);
4008 curr->sched_class->task_tick(rq, curr, 0);
4009 spin_unlock(&rq->lock);
4011 #ifdef CONFIG_SMP
4012 rq->idle_at_tick = idle_cpu(cpu);
4013 trigger_load_balance(rq, cpu);
4014 #endif
4017 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4019 void __kprobes add_preempt_count(int val)
4022 * Underflow?
4024 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4025 return;
4026 preempt_count() += val;
4028 * Spinlock count overflowing soon?
4030 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4031 PREEMPT_MASK - 10);
4033 EXPORT_SYMBOL(add_preempt_count);
4035 void __kprobes sub_preempt_count(int val)
4038 * Underflow?
4040 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4041 return;
4043 * Is the spinlock portion underflowing?
4045 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4046 !(preempt_count() & PREEMPT_MASK)))
4047 return;
4049 preempt_count() -= val;
4051 EXPORT_SYMBOL(sub_preempt_count);
4053 #endif
4056 * Print scheduling while atomic bug:
4058 static noinline void __schedule_bug(struct task_struct *prev)
4060 struct pt_regs *regs = get_irq_regs();
4062 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4063 prev->comm, prev->pid, preempt_count());
4065 debug_show_held_locks(prev);
4066 if (irqs_disabled())
4067 print_irqtrace_events(prev);
4069 if (regs)
4070 show_regs(regs);
4071 else
4072 dump_stack();
4076 * Various schedule()-time debugging checks and statistics:
4078 static inline void schedule_debug(struct task_struct *prev)
4081 * Test if we are atomic. Since do_exit() needs to call into
4082 * schedule() atomically, we ignore that path for now.
4083 * Otherwise, whine if we are scheduling when we should not be.
4085 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4086 __schedule_bug(prev);
4088 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4090 schedstat_inc(this_rq(), sched_count);
4091 #ifdef CONFIG_SCHEDSTATS
4092 if (unlikely(prev->lock_depth >= 0)) {
4093 schedstat_inc(this_rq(), bkl_count);
4094 schedstat_inc(prev, sched_info.bkl_count);
4096 #endif
4100 * Pick up the highest-prio task:
4102 static inline struct task_struct *
4103 pick_next_task(struct rq *rq, struct task_struct *prev)
4105 const struct sched_class *class;
4106 struct task_struct *p;
4109 * Optimization: we know that if all tasks are in
4110 * the fair class we can call that function directly:
4112 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4113 p = fair_sched_class.pick_next_task(rq);
4114 if (likely(p))
4115 return p;
4118 class = sched_class_highest;
4119 for ( ; ; ) {
4120 p = class->pick_next_task(rq);
4121 if (p)
4122 return p;
4124 * Will never be NULL as the idle class always
4125 * returns a non-NULL p:
4127 class = class->next;
4132 * schedule() is the main scheduler function.
4134 asmlinkage void __sched schedule(void)
4136 struct task_struct *prev, *next;
4137 unsigned long *switch_count;
4138 struct rq *rq;
4139 int cpu;
4141 need_resched:
4142 preempt_disable();
4143 cpu = smp_processor_id();
4144 rq = cpu_rq(cpu);
4145 rcu_qsctr_inc(cpu);
4146 prev = rq->curr;
4147 switch_count = &prev->nivcsw;
4149 release_kernel_lock(prev);
4150 need_resched_nonpreemptible:
4152 schedule_debug(prev);
4154 hrtick_clear(rq);
4157 * Do the rq-clock update outside the rq lock:
4159 local_irq_disable();
4160 update_rq_clock(rq);
4161 spin_lock(&rq->lock);
4162 clear_tsk_need_resched(prev);
4164 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4165 if (unlikely(signal_pending_state(prev->state, prev)))
4166 prev->state = TASK_RUNNING;
4167 else
4168 deactivate_task(rq, prev, 1);
4169 switch_count = &prev->nvcsw;
4172 #ifdef CONFIG_SMP
4173 if (prev->sched_class->pre_schedule)
4174 prev->sched_class->pre_schedule(rq, prev);
4175 #endif
4177 if (unlikely(!rq->nr_running))
4178 idle_balance(cpu, rq);
4180 prev->sched_class->put_prev_task(rq, prev);
4181 next = pick_next_task(rq, prev);
4183 if (likely(prev != next)) {
4184 sched_info_switch(prev, next);
4186 rq->nr_switches++;
4187 rq->curr = next;
4188 ++*switch_count;
4190 context_switch(rq, prev, next); /* unlocks the rq */
4192 * the context switch might have flipped the stack from under
4193 * us, hence refresh the local variables.
4195 cpu = smp_processor_id();
4196 rq = cpu_rq(cpu);
4197 } else
4198 spin_unlock_irq(&rq->lock);
4200 hrtick_set(rq);
4202 if (unlikely(reacquire_kernel_lock(current) < 0))
4203 goto need_resched_nonpreemptible;
4205 preempt_enable_no_resched();
4206 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4207 goto need_resched;
4209 EXPORT_SYMBOL(schedule);
4211 #ifdef CONFIG_PREEMPT
4213 * this is the entry point to schedule() from in-kernel preemption
4214 * off of preempt_enable. Kernel preemptions off return from interrupt
4215 * occur there and call schedule directly.
4217 asmlinkage void __sched preempt_schedule(void)
4219 struct thread_info *ti = current_thread_info();
4222 * If there is a non-zero preempt_count or interrupts are disabled,
4223 * we do not want to preempt the current task. Just return..
4225 if (likely(ti->preempt_count || irqs_disabled()))
4226 return;
4228 do {
4229 add_preempt_count(PREEMPT_ACTIVE);
4230 schedule();
4231 sub_preempt_count(PREEMPT_ACTIVE);
4234 * Check again in case we missed a preemption opportunity
4235 * between schedule and now.
4237 barrier();
4238 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4240 EXPORT_SYMBOL(preempt_schedule);
4243 * this is the entry point to schedule() from kernel preemption
4244 * off of irq context.
4245 * Note, that this is called and return with irqs disabled. This will
4246 * protect us against recursive calling from irq.
4248 asmlinkage void __sched preempt_schedule_irq(void)
4250 struct thread_info *ti = current_thread_info();
4252 /* Catch callers which need to be fixed */
4253 BUG_ON(ti->preempt_count || !irqs_disabled());
4255 do {
4256 add_preempt_count(PREEMPT_ACTIVE);
4257 local_irq_enable();
4258 schedule();
4259 local_irq_disable();
4260 sub_preempt_count(PREEMPT_ACTIVE);
4263 * Check again in case we missed a preemption opportunity
4264 * between schedule and now.
4266 barrier();
4267 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4270 #endif /* CONFIG_PREEMPT */
4272 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4273 void *key)
4275 return try_to_wake_up(curr->private, mode, sync);
4277 EXPORT_SYMBOL(default_wake_function);
4280 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4281 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4282 * number) then we wake all the non-exclusive tasks and one exclusive task.
4284 * There are circumstances in which we can try to wake a task which has already
4285 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4286 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4288 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4289 int nr_exclusive, int sync, void *key)
4291 wait_queue_t *curr, *next;
4293 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4294 unsigned flags = curr->flags;
4296 if (curr->func(curr, mode, sync, key) &&
4297 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4298 break;
4303 * __wake_up - wake up threads blocked on a waitqueue.
4304 * @q: the waitqueue
4305 * @mode: which threads
4306 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4307 * @key: is directly passed to the wakeup function
4309 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4310 int nr_exclusive, void *key)
4312 unsigned long flags;
4314 spin_lock_irqsave(&q->lock, flags);
4315 __wake_up_common(q, mode, nr_exclusive, 0, key);
4316 spin_unlock_irqrestore(&q->lock, flags);
4318 EXPORT_SYMBOL(__wake_up);
4321 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4323 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4325 __wake_up_common(q, mode, 1, 0, NULL);
4329 * __wake_up_sync - wake up threads blocked on a waitqueue.
4330 * @q: the waitqueue
4331 * @mode: which threads
4332 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4334 * The sync wakeup differs that the waker knows that it will schedule
4335 * away soon, so while the target thread will be woken up, it will not
4336 * be migrated to another CPU - ie. the two threads are 'synchronized'
4337 * with each other. This can prevent needless bouncing between CPUs.
4339 * On UP it can prevent extra preemption.
4341 void
4342 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4344 unsigned long flags;
4345 int sync = 1;
4347 if (unlikely(!q))
4348 return;
4350 if (unlikely(!nr_exclusive))
4351 sync = 0;
4353 spin_lock_irqsave(&q->lock, flags);
4354 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4355 spin_unlock_irqrestore(&q->lock, flags);
4357 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4359 void complete(struct completion *x)
4361 unsigned long flags;
4363 spin_lock_irqsave(&x->wait.lock, flags);
4364 x->done++;
4365 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4366 spin_unlock_irqrestore(&x->wait.lock, flags);
4368 EXPORT_SYMBOL(complete);
4370 void complete_all(struct completion *x)
4372 unsigned long flags;
4374 spin_lock_irqsave(&x->wait.lock, flags);
4375 x->done += UINT_MAX/2;
4376 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4377 spin_unlock_irqrestore(&x->wait.lock, flags);
4379 EXPORT_SYMBOL(complete_all);
4381 static inline long __sched
4382 do_wait_for_common(struct completion *x, long timeout, int state)
4384 if (!x->done) {
4385 DECLARE_WAITQUEUE(wait, current);
4387 wait.flags |= WQ_FLAG_EXCLUSIVE;
4388 __add_wait_queue_tail(&x->wait, &wait);
4389 do {
4390 if ((state == TASK_INTERRUPTIBLE &&
4391 signal_pending(current)) ||
4392 (state == TASK_KILLABLE &&
4393 fatal_signal_pending(current))) {
4394 __remove_wait_queue(&x->wait, &wait);
4395 return -ERESTARTSYS;
4397 __set_current_state(state);
4398 spin_unlock_irq(&x->wait.lock);
4399 timeout = schedule_timeout(timeout);
4400 spin_lock_irq(&x->wait.lock);
4401 if (!timeout) {
4402 __remove_wait_queue(&x->wait, &wait);
4403 return timeout;
4405 } while (!x->done);
4406 __remove_wait_queue(&x->wait, &wait);
4408 x->done--;
4409 return timeout;
4412 static long __sched
4413 wait_for_common(struct completion *x, long timeout, int state)
4415 might_sleep();
4417 spin_lock_irq(&x->wait.lock);
4418 timeout = do_wait_for_common(x, timeout, state);
4419 spin_unlock_irq(&x->wait.lock);
4420 return timeout;
4423 void __sched wait_for_completion(struct completion *x)
4425 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4427 EXPORT_SYMBOL(wait_for_completion);
4429 unsigned long __sched
4430 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4432 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4434 EXPORT_SYMBOL(wait_for_completion_timeout);
4436 int __sched wait_for_completion_interruptible(struct completion *x)
4438 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4439 if (t == -ERESTARTSYS)
4440 return t;
4441 return 0;
4443 EXPORT_SYMBOL(wait_for_completion_interruptible);
4445 unsigned long __sched
4446 wait_for_completion_interruptible_timeout(struct completion *x,
4447 unsigned long timeout)
4449 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4451 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4453 int __sched wait_for_completion_killable(struct completion *x)
4455 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4456 if (t == -ERESTARTSYS)
4457 return t;
4458 return 0;
4460 EXPORT_SYMBOL(wait_for_completion_killable);
4462 static long __sched
4463 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4465 unsigned long flags;
4466 wait_queue_t wait;
4468 init_waitqueue_entry(&wait, current);
4470 __set_current_state(state);
4472 spin_lock_irqsave(&q->lock, flags);
4473 __add_wait_queue(q, &wait);
4474 spin_unlock(&q->lock);
4475 timeout = schedule_timeout(timeout);
4476 spin_lock_irq(&q->lock);
4477 __remove_wait_queue(q, &wait);
4478 spin_unlock_irqrestore(&q->lock, flags);
4480 return timeout;
4483 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4485 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4487 EXPORT_SYMBOL(interruptible_sleep_on);
4489 long __sched
4490 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4492 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4494 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4496 void __sched sleep_on(wait_queue_head_t *q)
4498 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4500 EXPORT_SYMBOL(sleep_on);
4502 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4504 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4506 EXPORT_SYMBOL(sleep_on_timeout);
4508 #ifdef CONFIG_RT_MUTEXES
4511 * rt_mutex_setprio - set the current priority of a task
4512 * @p: task
4513 * @prio: prio value (kernel-internal form)
4515 * This function changes the 'effective' priority of a task. It does
4516 * not touch ->normal_prio like __setscheduler().
4518 * Used by the rt_mutex code to implement priority inheritance logic.
4520 void rt_mutex_setprio(struct task_struct *p, int prio)
4522 unsigned long flags;
4523 int oldprio, on_rq, running;
4524 struct rq *rq;
4525 const struct sched_class *prev_class = p->sched_class;
4527 BUG_ON(prio < 0 || prio > MAX_PRIO);
4529 rq = task_rq_lock(p, &flags);
4530 update_rq_clock(rq);
4532 oldprio = p->prio;
4533 on_rq = p->se.on_rq;
4534 running = task_current(rq, p);
4535 if (on_rq)
4536 dequeue_task(rq, p, 0);
4537 if (running)
4538 p->sched_class->put_prev_task(rq, p);
4540 if (rt_prio(prio))
4541 p->sched_class = &rt_sched_class;
4542 else
4543 p->sched_class = &fair_sched_class;
4545 p->prio = prio;
4547 if (running)
4548 p->sched_class->set_curr_task(rq);
4549 if (on_rq) {
4550 enqueue_task(rq, p, 0);
4552 check_class_changed(rq, p, prev_class, oldprio, running);
4554 task_rq_unlock(rq, &flags);
4557 #endif
4559 void set_user_nice(struct task_struct *p, long nice)
4561 int old_prio, delta, on_rq;
4562 unsigned long flags;
4563 struct rq *rq;
4565 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4566 return;
4568 * We have to be careful, if called from sys_setpriority(),
4569 * the task might be in the middle of scheduling on another CPU.
4571 rq = task_rq_lock(p, &flags);
4572 update_rq_clock(rq);
4574 * The RT priorities are set via sched_setscheduler(), but we still
4575 * allow the 'normal' nice value to be set - but as expected
4576 * it wont have any effect on scheduling until the task is
4577 * SCHED_FIFO/SCHED_RR:
4579 if (task_has_rt_policy(p)) {
4580 p->static_prio = NICE_TO_PRIO(nice);
4581 goto out_unlock;
4583 on_rq = p->se.on_rq;
4584 if (on_rq) {
4585 dequeue_task(rq, p, 0);
4586 dec_load(rq, p);
4589 p->static_prio = NICE_TO_PRIO(nice);
4590 set_load_weight(p);
4591 old_prio = p->prio;
4592 p->prio = effective_prio(p);
4593 delta = p->prio - old_prio;
4595 if (on_rq) {
4596 enqueue_task(rq, p, 0);
4597 inc_load(rq, p);
4599 * If the task increased its priority or is running and
4600 * lowered its priority, then reschedule its CPU:
4602 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4603 resched_task(rq->curr);
4605 out_unlock:
4606 task_rq_unlock(rq, &flags);
4608 EXPORT_SYMBOL(set_user_nice);
4611 * can_nice - check if a task can reduce its nice value
4612 * @p: task
4613 * @nice: nice value
4615 int can_nice(const struct task_struct *p, const int nice)
4617 /* convert nice value [19,-20] to rlimit style value [1,40] */
4618 int nice_rlim = 20 - nice;
4620 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4621 capable(CAP_SYS_NICE));
4624 #ifdef __ARCH_WANT_SYS_NICE
4627 * sys_nice - change the priority of the current process.
4628 * @increment: priority increment
4630 * sys_setpriority is a more generic, but much slower function that
4631 * does similar things.
4633 asmlinkage long sys_nice(int increment)
4635 long nice, retval;
4638 * Setpriority might change our priority at the same moment.
4639 * We don't have to worry. Conceptually one call occurs first
4640 * and we have a single winner.
4642 if (increment < -40)
4643 increment = -40;
4644 if (increment > 40)
4645 increment = 40;
4647 nice = PRIO_TO_NICE(current->static_prio) + increment;
4648 if (nice < -20)
4649 nice = -20;
4650 if (nice > 19)
4651 nice = 19;
4653 if (increment < 0 && !can_nice(current, nice))
4654 return -EPERM;
4656 retval = security_task_setnice(current, nice);
4657 if (retval)
4658 return retval;
4660 set_user_nice(current, nice);
4661 return 0;
4664 #endif
4667 * task_prio - return the priority value of a given task.
4668 * @p: the task in question.
4670 * This is the priority value as seen by users in /proc.
4671 * RT tasks are offset by -200. Normal tasks are centered
4672 * around 0, value goes from -16 to +15.
4674 int task_prio(const struct task_struct *p)
4676 return p->prio - MAX_RT_PRIO;
4680 * task_nice - return the nice value of a given task.
4681 * @p: the task in question.
4683 int task_nice(const struct task_struct *p)
4685 return TASK_NICE(p);
4687 EXPORT_SYMBOL(task_nice);
4690 * idle_cpu - is a given cpu idle currently?
4691 * @cpu: the processor in question.
4693 int idle_cpu(int cpu)
4695 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4699 * idle_task - return the idle task for a given cpu.
4700 * @cpu: the processor in question.
4702 struct task_struct *idle_task(int cpu)
4704 return cpu_rq(cpu)->idle;
4708 * find_process_by_pid - find a process with a matching PID value.
4709 * @pid: the pid in question.
4711 static struct task_struct *find_process_by_pid(pid_t pid)
4713 return pid ? find_task_by_vpid(pid) : current;
4716 /* Actually do priority change: must hold rq lock. */
4717 static void
4718 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4720 BUG_ON(p->se.on_rq);
4722 p->policy = policy;
4723 switch (p->policy) {
4724 case SCHED_NORMAL:
4725 case SCHED_BATCH:
4726 case SCHED_IDLE:
4727 p->sched_class = &fair_sched_class;
4728 break;
4729 case SCHED_FIFO:
4730 case SCHED_RR:
4731 p->sched_class = &rt_sched_class;
4732 break;
4735 p->rt_priority = prio;
4736 p->normal_prio = normal_prio(p);
4737 /* we are holding p->pi_lock already */
4738 p->prio = rt_mutex_getprio(p);
4739 set_load_weight(p);
4743 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4744 * @p: the task in question.
4745 * @policy: new policy.
4746 * @param: structure containing the new RT priority.
4748 * NOTE that the task may be already dead.
4750 int sched_setscheduler(struct task_struct *p, int policy,
4751 struct sched_param *param)
4753 int retval, oldprio, oldpolicy = -1, on_rq, running;
4754 unsigned long flags;
4755 const struct sched_class *prev_class = p->sched_class;
4756 struct rq *rq;
4758 /* may grab non-irq protected spin_locks */
4759 BUG_ON(in_interrupt());
4760 recheck:
4761 /* double check policy once rq lock held */
4762 if (policy < 0)
4763 policy = oldpolicy = p->policy;
4764 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4765 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4766 policy != SCHED_IDLE)
4767 return -EINVAL;
4769 * Valid priorities for SCHED_FIFO and SCHED_RR are
4770 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4771 * SCHED_BATCH and SCHED_IDLE is 0.
4773 if (param->sched_priority < 0 ||
4774 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4775 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4776 return -EINVAL;
4777 if (rt_policy(policy) != (param->sched_priority != 0))
4778 return -EINVAL;
4781 * Allow unprivileged RT tasks to decrease priority:
4783 if (!capable(CAP_SYS_NICE)) {
4784 if (rt_policy(policy)) {
4785 unsigned long rlim_rtprio;
4787 if (!lock_task_sighand(p, &flags))
4788 return -ESRCH;
4789 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4790 unlock_task_sighand(p, &flags);
4792 /* can't set/change the rt policy */
4793 if (policy != p->policy && !rlim_rtprio)
4794 return -EPERM;
4796 /* can't increase priority */
4797 if (param->sched_priority > p->rt_priority &&
4798 param->sched_priority > rlim_rtprio)
4799 return -EPERM;
4802 * Like positive nice levels, dont allow tasks to
4803 * move out of SCHED_IDLE either:
4805 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4806 return -EPERM;
4808 /* can't change other user's priorities */
4809 if ((current->euid != p->euid) &&
4810 (current->euid != p->uid))
4811 return -EPERM;
4814 #ifdef CONFIG_RT_GROUP_SCHED
4816 * Do not allow realtime tasks into groups that have no runtime
4817 * assigned.
4819 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
4820 return -EPERM;
4821 #endif
4823 retval = security_task_setscheduler(p, policy, param);
4824 if (retval)
4825 return retval;
4827 * make sure no PI-waiters arrive (or leave) while we are
4828 * changing the priority of the task:
4830 spin_lock_irqsave(&p->pi_lock, flags);
4832 * To be able to change p->policy safely, the apropriate
4833 * runqueue lock must be held.
4835 rq = __task_rq_lock(p);
4836 /* recheck policy now with rq lock held */
4837 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4838 policy = oldpolicy = -1;
4839 __task_rq_unlock(rq);
4840 spin_unlock_irqrestore(&p->pi_lock, flags);
4841 goto recheck;
4843 update_rq_clock(rq);
4844 on_rq = p->se.on_rq;
4845 running = task_current(rq, p);
4846 if (on_rq)
4847 deactivate_task(rq, p, 0);
4848 if (running)
4849 p->sched_class->put_prev_task(rq, p);
4851 oldprio = p->prio;
4852 __setscheduler(rq, p, policy, param->sched_priority);
4854 if (running)
4855 p->sched_class->set_curr_task(rq);
4856 if (on_rq) {
4857 activate_task(rq, p, 0);
4859 check_class_changed(rq, p, prev_class, oldprio, running);
4861 __task_rq_unlock(rq);
4862 spin_unlock_irqrestore(&p->pi_lock, flags);
4864 rt_mutex_adjust_pi(p);
4866 return 0;
4868 EXPORT_SYMBOL_GPL(sched_setscheduler);
4870 static int
4871 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4873 struct sched_param lparam;
4874 struct task_struct *p;
4875 int retval;
4877 if (!param || pid < 0)
4878 return -EINVAL;
4879 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4880 return -EFAULT;
4882 rcu_read_lock();
4883 retval = -ESRCH;
4884 p = find_process_by_pid(pid);
4885 if (p != NULL)
4886 retval = sched_setscheduler(p, policy, &lparam);
4887 rcu_read_unlock();
4889 return retval;
4893 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4894 * @pid: the pid in question.
4895 * @policy: new policy.
4896 * @param: structure containing the new RT priority.
4898 asmlinkage long
4899 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4901 /* negative values for policy are not valid */
4902 if (policy < 0)
4903 return -EINVAL;
4905 return do_sched_setscheduler(pid, policy, param);
4909 * sys_sched_setparam - set/change the RT priority of a thread
4910 * @pid: the pid in question.
4911 * @param: structure containing the new RT priority.
4913 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4915 return do_sched_setscheduler(pid, -1, param);
4919 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4920 * @pid: the pid in question.
4922 asmlinkage long sys_sched_getscheduler(pid_t pid)
4924 struct task_struct *p;
4925 int retval;
4927 if (pid < 0)
4928 return -EINVAL;
4930 retval = -ESRCH;
4931 read_lock(&tasklist_lock);
4932 p = find_process_by_pid(pid);
4933 if (p) {
4934 retval = security_task_getscheduler(p);
4935 if (!retval)
4936 retval = p->policy;
4938 read_unlock(&tasklist_lock);
4939 return retval;
4943 * sys_sched_getscheduler - get the RT priority of a thread
4944 * @pid: the pid in question.
4945 * @param: structure containing the RT priority.
4947 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4949 struct sched_param lp;
4950 struct task_struct *p;
4951 int retval;
4953 if (!param || pid < 0)
4954 return -EINVAL;
4956 read_lock(&tasklist_lock);
4957 p = find_process_by_pid(pid);
4958 retval = -ESRCH;
4959 if (!p)
4960 goto out_unlock;
4962 retval = security_task_getscheduler(p);
4963 if (retval)
4964 goto out_unlock;
4966 lp.sched_priority = p->rt_priority;
4967 read_unlock(&tasklist_lock);
4970 * This one might sleep, we cannot do it with a spinlock held ...
4972 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4974 return retval;
4976 out_unlock:
4977 read_unlock(&tasklist_lock);
4978 return retval;
4981 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
4983 cpumask_t cpus_allowed;
4984 cpumask_t new_mask = *in_mask;
4985 struct task_struct *p;
4986 int retval;
4988 get_online_cpus();
4989 read_lock(&tasklist_lock);
4991 p = find_process_by_pid(pid);
4992 if (!p) {
4993 read_unlock(&tasklist_lock);
4994 put_online_cpus();
4995 return -ESRCH;
4999 * It is not safe to call set_cpus_allowed with the
5000 * tasklist_lock held. We will bump the task_struct's
5001 * usage count and then drop tasklist_lock.
5003 get_task_struct(p);
5004 read_unlock(&tasklist_lock);
5006 retval = -EPERM;
5007 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5008 !capable(CAP_SYS_NICE))
5009 goto out_unlock;
5011 retval = security_task_setscheduler(p, 0, NULL);
5012 if (retval)
5013 goto out_unlock;
5015 cpuset_cpus_allowed(p, &cpus_allowed);
5016 cpus_and(new_mask, new_mask, cpus_allowed);
5017 again:
5018 retval = set_cpus_allowed_ptr(p, &new_mask);
5020 if (!retval) {
5021 cpuset_cpus_allowed(p, &cpus_allowed);
5022 if (!cpus_subset(new_mask, cpus_allowed)) {
5024 * We must have raced with a concurrent cpuset
5025 * update. Just reset the cpus_allowed to the
5026 * cpuset's cpus_allowed
5028 new_mask = cpus_allowed;
5029 goto again;
5032 out_unlock:
5033 put_task_struct(p);
5034 put_online_cpus();
5035 return retval;
5038 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5039 cpumask_t *new_mask)
5041 if (len < sizeof(cpumask_t)) {
5042 memset(new_mask, 0, sizeof(cpumask_t));
5043 } else if (len > sizeof(cpumask_t)) {
5044 len = sizeof(cpumask_t);
5046 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5050 * sys_sched_setaffinity - set the cpu affinity of a process
5051 * @pid: pid of the process
5052 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5053 * @user_mask_ptr: user-space pointer to the new cpu mask
5055 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5056 unsigned long __user *user_mask_ptr)
5058 cpumask_t new_mask;
5059 int retval;
5061 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5062 if (retval)
5063 return retval;
5065 return sched_setaffinity(pid, &new_mask);
5069 * Represents all cpu's present in the system
5070 * In systems capable of hotplug, this map could dynamically grow
5071 * as new cpu's are detected in the system via any platform specific
5072 * method, such as ACPI for e.g.
5075 cpumask_t cpu_present_map __read_mostly;
5076 EXPORT_SYMBOL(cpu_present_map);
5078 #ifndef CONFIG_SMP
5079 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
5080 EXPORT_SYMBOL(cpu_online_map);
5082 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
5083 EXPORT_SYMBOL(cpu_possible_map);
5084 #endif
5086 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5088 struct task_struct *p;
5089 int retval;
5091 get_online_cpus();
5092 read_lock(&tasklist_lock);
5094 retval = -ESRCH;
5095 p = find_process_by_pid(pid);
5096 if (!p)
5097 goto out_unlock;
5099 retval = security_task_getscheduler(p);
5100 if (retval)
5101 goto out_unlock;
5103 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5105 out_unlock:
5106 read_unlock(&tasklist_lock);
5107 put_online_cpus();
5109 return retval;
5113 * sys_sched_getaffinity - get the cpu affinity of a process
5114 * @pid: pid of the process
5115 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5116 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5118 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5119 unsigned long __user *user_mask_ptr)
5121 int ret;
5122 cpumask_t mask;
5124 if (len < sizeof(cpumask_t))
5125 return -EINVAL;
5127 ret = sched_getaffinity(pid, &mask);
5128 if (ret < 0)
5129 return ret;
5131 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5132 return -EFAULT;
5134 return sizeof(cpumask_t);
5138 * sys_sched_yield - yield the current processor to other threads.
5140 * This function yields the current CPU to other tasks. If there are no
5141 * other threads running on this CPU then this function will return.
5143 asmlinkage long sys_sched_yield(void)
5145 struct rq *rq = this_rq_lock();
5147 schedstat_inc(rq, yld_count);
5148 current->sched_class->yield_task(rq);
5151 * Since we are going to call schedule() anyway, there's
5152 * no need to preempt or enable interrupts:
5154 __release(rq->lock);
5155 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5156 _raw_spin_unlock(&rq->lock);
5157 preempt_enable_no_resched();
5159 schedule();
5161 return 0;
5164 static void __cond_resched(void)
5166 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5167 __might_sleep(__FILE__, __LINE__);
5168 #endif
5170 * The BKS might be reacquired before we have dropped
5171 * PREEMPT_ACTIVE, which could trigger a second
5172 * cond_resched() call.
5174 do {
5175 add_preempt_count(PREEMPT_ACTIVE);
5176 schedule();
5177 sub_preempt_count(PREEMPT_ACTIVE);
5178 } while (need_resched());
5181 int __sched _cond_resched(void)
5183 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5184 system_state == SYSTEM_RUNNING) {
5185 __cond_resched();
5186 return 1;
5188 return 0;
5190 EXPORT_SYMBOL(_cond_resched);
5193 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5194 * call schedule, and on return reacquire the lock.
5196 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5197 * operations here to prevent schedule() from being called twice (once via
5198 * spin_unlock(), once by hand).
5200 int cond_resched_lock(spinlock_t *lock)
5202 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5203 int ret = 0;
5205 if (spin_needbreak(lock) || resched) {
5206 spin_unlock(lock);
5207 if (resched && need_resched())
5208 __cond_resched();
5209 else
5210 cpu_relax();
5211 ret = 1;
5212 spin_lock(lock);
5214 return ret;
5216 EXPORT_SYMBOL(cond_resched_lock);
5218 int __sched cond_resched_softirq(void)
5220 BUG_ON(!in_softirq());
5222 if (need_resched() && system_state == SYSTEM_RUNNING) {
5223 local_bh_enable();
5224 __cond_resched();
5225 local_bh_disable();
5226 return 1;
5228 return 0;
5230 EXPORT_SYMBOL(cond_resched_softirq);
5233 * yield - yield the current processor to other threads.
5235 * This is a shortcut for kernel-space yielding - it marks the
5236 * thread runnable and calls sys_sched_yield().
5238 void __sched yield(void)
5240 set_current_state(TASK_RUNNING);
5241 sys_sched_yield();
5243 EXPORT_SYMBOL(yield);
5246 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5247 * that process accounting knows that this is a task in IO wait state.
5249 * But don't do that if it is a deliberate, throttling IO wait (this task
5250 * has set its backing_dev_info: the queue against which it should throttle)
5252 void __sched io_schedule(void)
5254 struct rq *rq = &__raw_get_cpu_var(runqueues);
5256 delayacct_blkio_start();
5257 atomic_inc(&rq->nr_iowait);
5258 schedule();
5259 atomic_dec(&rq->nr_iowait);
5260 delayacct_blkio_end();
5262 EXPORT_SYMBOL(io_schedule);
5264 long __sched io_schedule_timeout(long timeout)
5266 struct rq *rq = &__raw_get_cpu_var(runqueues);
5267 long ret;
5269 delayacct_blkio_start();
5270 atomic_inc(&rq->nr_iowait);
5271 ret = schedule_timeout(timeout);
5272 atomic_dec(&rq->nr_iowait);
5273 delayacct_blkio_end();
5274 return ret;
5278 * sys_sched_get_priority_max - return maximum RT priority.
5279 * @policy: scheduling class.
5281 * this syscall returns the maximum rt_priority that can be used
5282 * by a given scheduling class.
5284 asmlinkage long sys_sched_get_priority_max(int policy)
5286 int ret = -EINVAL;
5288 switch (policy) {
5289 case SCHED_FIFO:
5290 case SCHED_RR:
5291 ret = MAX_USER_RT_PRIO-1;
5292 break;
5293 case SCHED_NORMAL:
5294 case SCHED_BATCH:
5295 case SCHED_IDLE:
5296 ret = 0;
5297 break;
5299 return ret;
5303 * sys_sched_get_priority_min - return minimum RT priority.
5304 * @policy: scheduling class.
5306 * this syscall returns the minimum rt_priority that can be used
5307 * by a given scheduling class.
5309 asmlinkage long sys_sched_get_priority_min(int policy)
5311 int ret = -EINVAL;
5313 switch (policy) {
5314 case SCHED_FIFO:
5315 case SCHED_RR:
5316 ret = 1;
5317 break;
5318 case SCHED_NORMAL:
5319 case SCHED_BATCH:
5320 case SCHED_IDLE:
5321 ret = 0;
5323 return ret;
5327 * sys_sched_rr_get_interval - return the default timeslice of a process.
5328 * @pid: pid of the process.
5329 * @interval: userspace pointer to the timeslice value.
5331 * this syscall writes the default timeslice value of a given process
5332 * into the user-space timespec buffer. A value of '0' means infinity.
5334 asmlinkage
5335 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5337 struct task_struct *p;
5338 unsigned int time_slice;
5339 int retval;
5340 struct timespec t;
5342 if (pid < 0)
5343 return -EINVAL;
5345 retval = -ESRCH;
5346 read_lock(&tasklist_lock);
5347 p = find_process_by_pid(pid);
5348 if (!p)
5349 goto out_unlock;
5351 retval = security_task_getscheduler(p);
5352 if (retval)
5353 goto out_unlock;
5356 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5357 * tasks that are on an otherwise idle runqueue:
5359 time_slice = 0;
5360 if (p->policy == SCHED_RR) {
5361 time_slice = DEF_TIMESLICE;
5362 } else if (p->policy != SCHED_FIFO) {
5363 struct sched_entity *se = &p->se;
5364 unsigned long flags;
5365 struct rq *rq;
5367 rq = task_rq_lock(p, &flags);
5368 if (rq->cfs.load.weight)
5369 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5370 task_rq_unlock(rq, &flags);
5372 read_unlock(&tasklist_lock);
5373 jiffies_to_timespec(time_slice, &t);
5374 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5375 return retval;
5377 out_unlock:
5378 read_unlock(&tasklist_lock);
5379 return retval;
5382 static const char stat_nam[] = "RSDTtZX";
5384 void sched_show_task(struct task_struct *p)
5386 unsigned long free = 0;
5387 unsigned state;
5389 state = p->state ? __ffs(p->state) + 1 : 0;
5390 printk(KERN_INFO "%-13.13s %c", p->comm,
5391 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5392 #if BITS_PER_LONG == 32
5393 if (state == TASK_RUNNING)
5394 printk(KERN_CONT " running ");
5395 else
5396 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5397 #else
5398 if (state == TASK_RUNNING)
5399 printk(KERN_CONT " running task ");
5400 else
5401 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5402 #endif
5403 #ifdef CONFIG_DEBUG_STACK_USAGE
5405 unsigned long *n = end_of_stack(p);
5406 while (!*n)
5407 n++;
5408 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5410 #endif
5411 printk(KERN_CONT "%5lu %5d %6d\n", free,
5412 task_pid_nr(p), task_pid_nr(p->real_parent));
5414 show_stack(p, NULL);
5417 void show_state_filter(unsigned long state_filter)
5419 struct task_struct *g, *p;
5421 #if BITS_PER_LONG == 32
5422 printk(KERN_INFO
5423 " task PC stack pid father\n");
5424 #else
5425 printk(KERN_INFO
5426 " task PC stack pid father\n");
5427 #endif
5428 read_lock(&tasklist_lock);
5429 do_each_thread(g, p) {
5431 * reset the NMI-timeout, listing all files on a slow
5432 * console might take alot of time:
5434 touch_nmi_watchdog();
5435 if (!state_filter || (p->state & state_filter))
5436 sched_show_task(p);
5437 } while_each_thread(g, p);
5439 touch_all_softlockup_watchdogs();
5441 #ifdef CONFIG_SCHED_DEBUG
5442 sysrq_sched_debug_show();
5443 #endif
5444 read_unlock(&tasklist_lock);
5446 * Only show locks if all tasks are dumped:
5448 if (state_filter == -1)
5449 debug_show_all_locks();
5452 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5454 idle->sched_class = &idle_sched_class;
5458 * init_idle - set up an idle thread for a given CPU
5459 * @idle: task in question
5460 * @cpu: cpu the idle task belongs to
5462 * NOTE: this function does not set the idle thread's NEED_RESCHED
5463 * flag, to make booting more robust.
5465 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5467 struct rq *rq = cpu_rq(cpu);
5468 unsigned long flags;
5470 __sched_fork(idle);
5471 idle->se.exec_start = sched_clock();
5473 idle->prio = idle->normal_prio = MAX_PRIO;
5474 idle->cpus_allowed = cpumask_of_cpu(cpu);
5475 __set_task_cpu(idle, cpu);
5477 spin_lock_irqsave(&rq->lock, flags);
5478 rq->curr = rq->idle = idle;
5479 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5480 idle->oncpu = 1;
5481 #endif
5482 spin_unlock_irqrestore(&rq->lock, flags);
5484 /* Set the preempt count _outside_ the spinlocks! */
5485 #if defined(CONFIG_PREEMPT)
5486 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5487 #else
5488 task_thread_info(idle)->preempt_count = 0;
5489 #endif
5491 * The idle tasks have their own, simple scheduling class:
5493 idle->sched_class = &idle_sched_class;
5497 * In a system that switches off the HZ timer nohz_cpu_mask
5498 * indicates which cpus entered this state. This is used
5499 * in the rcu update to wait only for active cpus. For system
5500 * which do not switch off the HZ timer nohz_cpu_mask should
5501 * always be CPU_MASK_NONE.
5503 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5506 * Increase the granularity value when there are more CPUs,
5507 * because with more CPUs the 'effective latency' as visible
5508 * to users decreases. But the relationship is not linear,
5509 * so pick a second-best guess by going with the log2 of the
5510 * number of CPUs.
5512 * This idea comes from the SD scheduler of Con Kolivas:
5514 static inline void sched_init_granularity(void)
5516 unsigned int factor = 1 + ilog2(num_online_cpus());
5517 const unsigned long limit = 200000000;
5519 sysctl_sched_min_granularity *= factor;
5520 if (sysctl_sched_min_granularity > limit)
5521 sysctl_sched_min_granularity = limit;
5523 sysctl_sched_latency *= factor;
5524 if (sysctl_sched_latency > limit)
5525 sysctl_sched_latency = limit;
5527 sysctl_sched_wakeup_granularity *= factor;
5530 #ifdef CONFIG_SMP
5532 * This is how migration works:
5534 * 1) we queue a struct migration_req structure in the source CPU's
5535 * runqueue and wake up that CPU's migration thread.
5536 * 2) we down() the locked semaphore => thread blocks.
5537 * 3) migration thread wakes up (implicitly it forces the migrated
5538 * thread off the CPU)
5539 * 4) it gets the migration request and checks whether the migrated
5540 * task is still in the wrong runqueue.
5541 * 5) if it's in the wrong runqueue then the migration thread removes
5542 * it and puts it into the right queue.
5543 * 6) migration thread up()s the semaphore.
5544 * 7) we wake up and the migration is done.
5548 * Change a given task's CPU affinity. Migrate the thread to a
5549 * proper CPU and schedule it away if the CPU it's executing on
5550 * is removed from the allowed bitmask.
5552 * NOTE: the caller must have a valid reference to the task, the
5553 * task must not exit() & deallocate itself prematurely. The
5554 * call is not atomic; no spinlocks may be held.
5556 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5558 struct migration_req req;
5559 unsigned long flags;
5560 struct rq *rq;
5561 int ret = 0;
5563 rq = task_rq_lock(p, &flags);
5564 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5565 ret = -EINVAL;
5566 goto out;
5569 if (p->sched_class->set_cpus_allowed)
5570 p->sched_class->set_cpus_allowed(p, new_mask);
5571 else {
5572 p->cpus_allowed = *new_mask;
5573 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5576 /* Can the task run on the task's current CPU? If so, we're done */
5577 if (cpu_isset(task_cpu(p), *new_mask))
5578 goto out;
5580 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5581 /* Need help from migration thread: drop lock and wait. */
5582 task_rq_unlock(rq, &flags);
5583 wake_up_process(rq->migration_thread);
5584 wait_for_completion(&req.done);
5585 tlb_migrate_finish(p->mm);
5586 return 0;
5588 out:
5589 task_rq_unlock(rq, &flags);
5591 return ret;
5593 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5596 * Move (not current) task off this cpu, onto dest cpu. We're doing
5597 * this because either it can't run here any more (set_cpus_allowed()
5598 * away from this CPU, or CPU going down), or because we're
5599 * attempting to rebalance this task on exec (sched_exec).
5601 * So we race with normal scheduler movements, but that's OK, as long
5602 * as the task is no longer on this CPU.
5604 * Returns non-zero if task was successfully migrated.
5606 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5608 struct rq *rq_dest, *rq_src;
5609 int ret = 0, on_rq;
5611 if (unlikely(cpu_is_offline(dest_cpu)))
5612 return ret;
5614 rq_src = cpu_rq(src_cpu);
5615 rq_dest = cpu_rq(dest_cpu);
5617 double_rq_lock(rq_src, rq_dest);
5618 /* Already moved. */
5619 if (task_cpu(p) != src_cpu)
5620 goto out;
5621 /* Affinity changed (again). */
5622 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5623 goto out;
5625 on_rq = p->se.on_rq;
5626 if (on_rq)
5627 deactivate_task(rq_src, p, 0);
5629 set_task_cpu(p, dest_cpu);
5630 if (on_rq) {
5631 activate_task(rq_dest, p, 0);
5632 check_preempt_curr(rq_dest, p);
5634 ret = 1;
5635 out:
5636 double_rq_unlock(rq_src, rq_dest);
5637 return ret;
5641 * migration_thread - this is a highprio system thread that performs
5642 * thread migration by bumping thread off CPU then 'pushing' onto
5643 * another runqueue.
5645 static int migration_thread(void *data)
5647 int cpu = (long)data;
5648 struct rq *rq;
5650 rq = cpu_rq(cpu);
5651 BUG_ON(rq->migration_thread != current);
5653 set_current_state(TASK_INTERRUPTIBLE);
5654 while (!kthread_should_stop()) {
5655 struct migration_req *req;
5656 struct list_head *head;
5658 spin_lock_irq(&rq->lock);
5660 if (cpu_is_offline(cpu)) {
5661 spin_unlock_irq(&rq->lock);
5662 goto wait_to_die;
5665 if (rq->active_balance) {
5666 active_load_balance(rq, cpu);
5667 rq->active_balance = 0;
5670 head = &rq->migration_queue;
5672 if (list_empty(head)) {
5673 spin_unlock_irq(&rq->lock);
5674 schedule();
5675 set_current_state(TASK_INTERRUPTIBLE);
5676 continue;
5678 req = list_entry(head->next, struct migration_req, list);
5679 list_del_init(head->next);
5681 spin_unlock(&rq->lock);
5682 __migrate_task(req->task, cpu, req->dest_cpu);
5683 local_irq_enable();
5685 complete(&req->done);
5687 __set_current_state(TASK_RUNNING);
5688 return 0;
5690 wait_to_die:
5691 /* Wait for kthread_stop */
5692 set_current_state(TASK_INTERRUPTIBLE);
5693 while (!kthread_should_stop()) {
5694 schedule();
5695 set_current_state(TASK_INTERRUPTIBLE);
5697 __set_current_state(TASK_RUNNING);
5698 return 0;
5701 #ifdef CONFIG_HOTPLUG_CPU
5703 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5705 int ret;
5707 local_irq_disable();
5708 ret = __migrate_task(p, src_cpu, dest_cpu);
5709 local_irq_enable();
5710 return ret;
5714 * Figure out where task on dead CPU should go, use force if necessary.
5715 * NOTE: interrupts should be disabled by the caller
5717 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5719 unsigned long flags;
5720 cpumask_t mask;
5721 struct rq *rq;
5722 int dest_cpu;
5724 do {
5725 /* On same node? */
5726 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5727 cpus_and(mask, mask, p->cpus_allowed);
5728 dest_cpu = any_online_cpu(mask);
5730 /* On any allowed CPU? */
5731 if (dest_cpu >= nr_cpu_ids)
5732 dest_cpu = any_online_cpu(p->cpus_allowed);
5734 /* No more Mr. Nice Guy. */
5735 if (dest_cpu >= nr_cpu_ids) {
5736 cpumask_t cpus_allowed;
5738 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5740 * Try to stay on the same cpuset, where the
5741 * current cpuset may be a subset of all cpus.
5742 * The cpuset_cpus_allowed_locked() variant of
5743 * cpuset_cpus_allowed() will not block. It must be
5744 * called within calls to cpuset_lock/cpuset_unlock.
5746 rq = task_rq_lock(p, &flags);
5747 p->cpus_allowed = cpus_allowed;
5748 dest_cpu = any_online_cpu(p->cpus_allowed);
5749 task_rq_unlock(rq, &flags);
5752 * Don't tell them about moving exiting tasks or
5753 * kernel threads (both mm NULL), since they never
5754 * leave kernel.
5756 if (p->mm && printk_ratelimit()) {
5757 printk(KERN_INFO "process %d (%s) no "
5758 "longer affine to cpu%d\n",
5759 task_pid_nr(p), p->comm, dead_cpu);
5762 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5766 * While a dead CPU has no uninterruptible tasks queued at this point,
5767 * it might still have a nonzero ->nr_uninterruptible counter, because
5768 * for performance reasons the counter is not stricly tracking tasks to
5769 * their home CPUs. So we just add the counter to another CPU's counter,
5770 * to keep the global sum constant after CPU-down:
5772 static void migrate_nr_uninterruptible(struct rq *rq_src)
5774 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5775 unsigned long flags;
5777 local_irq_save(flags);
5778 double_rq_lock(rq_src, rq_dest);
5779 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5780 rq_src->nr_uninterruptible = 0;
5781 double_rq_unlock(rq_src, rq_dest);
5782 local_irq_restore(flags);
5785 /* Run through task list and migrate tasks from the dead cpu. */
5786 static void migrate_live_tasks(int src_cpu)
5788 struct task_struct *p, *t;
5790 read_lock(&tasklist_lock);
5792 do_each_thread(t, p) {
5793 if (p == current)
5794 continue;
5796 if (task_cpu(p) == src_cpu)
5797 move_task_off_dead_cpu(src_cpu, p);
5798 } while_each_thread(t, p);
5800 read_unlock(&tasklist_lock);
5804 * Schedules idle task to be the next runnable task on current CPU.
5805 * It does so by boosting its priority to highest possible.
5806 * Used by CPU offline code.
5808 void sched_idle_next(void)
5810 int this_cpu = smp_processor_id();
5811 struct rq *rq = cpu_rq(this_cpu);
5812 struct task_struct *p = rq->idle;
5813 unsigned long flags;
5815 /* cpu has to be offline */
5816 BUG_ON(cpu_online(this_cpu));
5819 * Strictly not necessary since rest of the CPUs are stopped by now
5820 * and interrupts disabled on the current cpu.
5822 spin_lock_irqsave(&rq->lock, flags);
5824 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5826 update_rq_clock(rq);
5827 activate_task(rq, p, 0);
5829 spin_unlock_irqrestore(&rq->lock, flags);
5833 * Ensures that the idle task is using init_mm right before its cpu goes
5834 * offline.
5836 void idle_task_exit(void)
5838 struct mm_struct *mm = current->active_mm;
5840 BUG_ON(cpu_online(smp_processor_id()));
5842 if (mm != &init_mm)
5843 switch_mm(mm, &init_mm, current);
5844 mmdrop(mm);
5847 /* called under rq->lock with disabled interrupts */
5848 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5850 struct rq *rq = cpu_rq(dead_cpu);
5852 /* Must be exiting, otherwise would be on tasklist. */
5853 BUG_ON(!p->exit_state);
5855 /* Cannot have done final schedule yet: would have vanished. */
5856 BUG_ON(p->state == TASK_DEAD);
5858 get_task_struct(p);
5861 * Drop lock around migration; if someone else moves it,
5862 * that's OK. No task can be added to this CPU, so iteration is
5863 * fine.
5865 spin_unlock_irq(&rq->lock);
5866 move_task_off_dead_cpu(dead_cpu, p);
5867 spin_lock_irq(&rq->lock);
5869 put_task_struct(p);
5872 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5873 static void migrate_dead_tasks(unsigned int dead_cpu)
5875 struct rq *rq = cpu_rq(dead_cpu);
5876 struct task_struct *next;
5878 for ( ; ; ) {
5879 if (!rq->nr_running)
5880 break;
5881 update_rq_clock(rq);
5882 next = pick_next_task(rq, rq->curr);
5883 if (!next)
5884 break;
5885 migrate_dead(dead_cpu, next);
5889 #endif /* CONFIG_HOTPLUG_CPU */
5891 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5893 static struct ctl_table sd_ctl_dir[] = {
5895 .procname = "sched_domain",
5896 .mode = 0555,
5898 {0, },
5901 static struct ctl_table sd_ctl_root[] = {
5903 .ctl_name = CTL_KERN,
5904 .procname = "kernel",
5905 .mode = 0555,
5906 .child = sd_ctl_dir,
5908 {0, },
5911 static struct ctl_table *sd_alloc_ctl_entry(int n)
5913 struct ctl_table *entry =
5914 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5916 return entry;
5919 static void sd_free_ctl_entry(struct ctl_table **tablep)
5921 struct ctl_table *entry;
5924 * In the intermediate directories, both the child directory and
5925 * procname are dynamically allocated and could fail but the mode
5926 * will always be set. In the lowest directory the names are
5927 * static strings and all have proc handlers.
5929 for (entry = *tablep; entry->mode; entry++) {
5930 if (entry->child)
5931 sd_free_ctl_entry(&entry->child);
5932 if (entry->proc_handler == NULL)
5933 kfree(entry->procname);
5936 kfree(*tablep);
5937 *tablep = NULL;
5940 static void
5941 set_table_entry(struct ctl_table *entry,
5942 const char *procname, void *data, int maxlen,
5943 mode_t mode, proc_handler *proc_handler)
5945 entry->procname = procname;
5946 entry->data = data;
5947 entry->maxlen = maxlen;
5948 entry->mode = mode;
5949 entry->proc_handler = proc_handler;
5952 static struct ctl_table *
5953 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5955 struct ctl_table *table = sd_alloc_ctl_entry(12);
5957 if (table == NULL)
5958 return NULL;
5960 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5961 sizeof(long), 0644, proc_doulongvec_minmax);
5962 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5963 sizeof(long), 0644, proc_doulongvec_minmax);
5964 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5965 sizeof(int), 0644, proc_dointvec_minmax);
5966 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5967 sizeof(int), 0644, proc_dointvec_minmax);
5968 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5969 sizeof(int), 0644, proc_dointvec_minmax);
5970 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5971 sizeof(int), 0644, proc_dointvec_minmax);
5972 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5973 sizeof(int), 0644, proc_dointvec_minmax);
5974 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5975 sizeof(int), 0644, proc_dointvec_minmax);
5976 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5977 sizeof(int), 0644, proc_dointvec_minmax);
5978 set_table_entry(&table[9], "cache_nice_tries",
5979 &sd->cache_nice_tries,
5980 sizeof(int), 0644, proc_dointvec_minmax);
5981 set_table_entry(&table[10], "flags", &sd->flags,
5982 sizeof(int), 0644, proc_dointvec_minmax);
5983 /* &table[11] is terminator */
5985 return table;
5988 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5990 struct ctl_table *entry, *table;
5991 struct sched_domain *sd;
5992 int domain_num = 0, i;
5993 char buf[32];
5995 for_each_domain(cpu, sd)
5996 domain_num++;
5997 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5998 if (table == NULL)
5999 return NULL;
6001 i = 0;
6002 for_each_domain(cpu, sd) {
6003 snprintf(buf, 32, "domain%d", i);
6004 entry->procname = kstrdup(buf, GFP_KERNEL);
6005 entry->mode = 0555;
6006 entry->child = sd_alloc_ctl_domain_table(sd);
6007 entry++;
6008 i++;
6010 return table;
6013 static struct ctl_table_header *sd_sysctl_header;
6014 static void register_sched_domain_sysctl(void)
6016 int i, cpu_num = num_online_cpus();
6017 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6018 char buf[32];
6020 WARN_ON(sd_ctl_dir[0].child);
6021 sd_ctl_dir[0].child = entry;
6023 if (entry == NULL)
6024 return;
6026 for_each_online_cpu(i) {
6027 snprintf(buf, 32, "cpu%d", i);
6028 entry->procname = kstrdup(buf, GFP_KERNEL);
6029 entry->mode = 0555;
6030 entry->child = sd_alloc_ctl_cpu_table(i);
6031 entry++;
6034 WARN_ON(sd_sysctl_header);
6035 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6038 /* may be called multiple times per register */
6039 static void unregister_sched_domain_sysctl(void)
6041 if (sd_sysctl_header)
6042 unregister_sysctl_table(sd_sysctl_header);
6043 sd_sysctl_header = NULL;
6044 if (sd_ctl_dir[0].child)
6045 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6047 #else
6048 static void register_sched_domain_sysctl(void)
6051 static void unregister_sched_domain_sysctl(void)
6054 #endif
6057 * migration_call - callback that gets triggered when a CPU is added.
6058 * Here we can start up the necessary migration thread for the new CPU.
6060 static int __cpuinit
6061 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6063 struct task_struct *p;
6064 int cpu = (long)hcpu;
6065 unsigned long flags;
6066 struct rq *rq;
6068 switch (action) {
6070 case CPU_UP_PREPARE:
6071 case CPU_UP_PREPARE_FROZEN:
6072 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6073 if (IS_ERR(p))
6074 return NOTIFY_BAD;
6075 kthread_bind(p, cpu);
6076 /* Must be high prio: stop_machine expects to yield to it. */
6077 rq = task_rq_lock(p, &flags);
6078 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6079 task_rq_unlock(rq, &flags);
6080 cpu_rq(cpu)->migration_thread = p;
6081 break;
6083 case CPU_ONLINE:
6084 case CPU_ONLINE_FROZEN:
6085 /* Strictly unnecessary, as first user will wake it. */
6086 wake_up_process(cpu_rq(cpu)->migration_thread);
6088 /* Update our root-domain */
6089 rq = cpu_rq(cpu);
6090 spin_lock_irqsave(&rq->lock, flags);
6091 if (rq->rd) {
6092 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6093 cpu_set(cpu, rq->rd->online);
6095 spin_unlock_irqrestore(&rq->lock, flags);
6096 break;
6098 #ifdef CONFIG_HOTPLUG_CPU
6099 case CPU_UP_CANCELED:
6100 case CPU_UP_CANCELED_FROZEN:
6101 if (!cpu_rq(cpu)->migration_thread)
6102 break;
6103 /* Unbind it from offline cpu so it can run. Fall thru. */
6104 kthread_bind(cpu_rq(cpu)->migration_thread,
6105 any_online_cpu(cpu_online_map));
6106 kthread_stop(cpu_rq(cpu)->migration_thread);
6107 cpu_rq(cpu)->migration_thread = NULL;
6108 break;
6110 case CPU_DEAD:
6111 case CPU_DEAD_FROZEN:
6112 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6113 migrate_live_tasks(cpu);
6114 rq = cpu_rq(cpu);
6115 kthread_stop(rq->migration_thread);
6116 rq->migration_thread = NULL;
6117 /* Idle task back to normal (off runqueue, low prio) */
6118 spin_lock_irq(&rq->lock);
6119 update_rq_clock(rq);
6120 deactivate_task(rq, rq->idle, 0);
6121 rq->idle->static_prio = MAX_PRIO;
6122 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6123 rq->idle->sched_class = &idle_sched_class;
6124 migrate_dead_tasks(cpu);
6125 spin_unlock_irq(&rq->lock);
6126 cpuset_unlock();
6127 migrate_nr_uninterruptible(rq);
6128 BUG_ON(rq->nr_running != 0);
6131 * No need to migrate the tasks: it was best-effort if
6132 * they didn't take sched_hotcpu_mutex. Just wake up
6133 * the requestors.
6135 spin_lock_irq(&rq->lock);
6136 while (!list_empty(&rq->migration_queue)) {
6137 struct migration_req *req;
6139 req = list_entry(rq->migration_queue.next,
6140 struct migration_req, list);
6141 list_del_init(&req->list);
6142 complete(&req->done);
6144 spin_unlock_irq(&rq->lock);
6145 break;
6147 case CPU_DYING:
6148 case CPU_DYING_FROZEN:
6149 /* Update our root-domain */
6150 rq = cpu_rq(cpu);
6151 spin_lock_irqsave(&rq->lock, flags);
6152 if (rq->rd) {
6153 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6154 cpu_clear(cpu, rq->rd->online);
6156 spin_unlock_irqrestore(&rq->lock, flags);
6157 break;
6158 #endif
6160 return NOTIFY_OK;
6163 /* Register at highest priority so that task migration (migrate_all_tasks)
6164 * happens before everything else.
6166 static struct notifier_block __cpuinitdata migration_notifier = {
6167 .notifier_call = migration_call,
6168 .priority = 10
6171 void __init migration_init(void)
6173 void *cpu = (void *)(long)smp_processor_id();
6174 int err;
6176 /* Start one for the boot CPU: */
6177 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6178 BUG_ON(err == NOTIFY_BAD);
6179 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6180 register_cpu_notifier(&migration_notifier);
6182 #endif
6184 #ifdef CONFIG_SMP
6186 #ifdef CONFIG_SCHED_DEBUG
6188 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6189 cpumask_t *groupmask)
6191 struct sched_group *group = sd->groups;
6192 char str[256];
6194 cpulist_scnprintf(str, sizeof(str), sd->span);
6195 cpus_clear(*groupmask);
6197 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6199 if (!(sd->flags & SD_LOAD_BALANCE)) {
6200 printk("does not load-balance\n");
6201 if (sd->parent)
6202 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6203 " has parent");
6204 return -1;
6207 printk(KERN_CONT "span %s\n", str);
6209 if (!cpu_isset(cpu, sd->span)) {
6210 printk(KERN_ERR "ERROR: domain->span does not contain "
6211 "CPU%d\n", cpu);
6213 if (!cpu_isset(cpu, group->cpumask)) {
6214 printk(KERN_ERR "ERROR: domain->groups does not contain"
6215 " CPU%d\n", cpu);
6218 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6219 do {
6220 if (!group) {
6221 printk("\n");
6222 printk(KERN_ERR "ERROR: group is NULL\n");
6223 break;
6226 if (!group->__cpu_power) {
6227 printk(KERN_CONT "\n");
6228 printk(KERN_ERR "ERROR: domain->cpu_power not "
6229 "set\n");
6230 break;
6233 if (!cpus_weight(group->cpumask)) {
6234 printk(KERN_CONT "\n");
6235 printk(KERN_ERR "ERROR: empty group\n");
6236 break;
6239 if (cpus_intersects(*groupmask, group->cpumask)) {
6240 printk(KERN_CONT "\n");
6241 printk(KERN_ERR "ERROR: repeated CPUs\n");
6242 break;
6245 cpus_or(*groupmask, *groupmask, group->cpumask);
6247 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6248 printk(KERN_CONT " %s", str);
6250 group = group->next;
6251 } while (group != sd->groups);
6252 printk(KERN_CONT "\n");
6254 if (!cpus_equal(sd->span, *groupmask))
6255 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6257 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6258 printk(KERN_ERR "ERROR: parent span is not a superset "
6259 "of domain->span\n");
6260 return 0;
6263 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6265 cpumask_t *groupmask;
6266 int level = 0;
6268 if (!sd) {
6269 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6270 return;
6273 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6275 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6276 if (!groupmask) {
6277 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6278 return;
6281 for (;;) {
6282 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6283 break;
6284 level++;
6285 sd = sd->parent;
6286 if (!sd)
6287 break;
6289 kfree(groupmask);
6291 #else
6292 # define sched_domain_debug(sd, cpu) do { } while (0)
6293 #endif
6295 static int sd_degenerate(struct sched_domain *sd)
6297 if (cpus_weight(sd->span) == 1)
6298 return 1;
6300 /* Following flags need at least 2 groups */
6301 if (sd->flags & (SD_LOAD_BALANCE |
6302 SD_BALANCE_NEWIDLE |
6303 SD_BALANCE_FORK |
6304 SD_BALANCE_EXEC |
6305 SD_SHARE_CPUPOWER |
6306 SD_SHARE_PKG_RESOURCES)) {
6307 if (sd->groups != sd->groups->next)
6308 return 0;
6311 /* Following flags don't use groups */
6312 if (sd->flags & (SD_WAKE_IDLE |
6313 SD_WAKE_AFFINE |
6314 SD_WAKE_BALANCE))
6315 return 0;
6317 return 1;
6320 static int
6321 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6323 unsigned long cflags = sd->flags, pflags = parent->flags;
6325 if (sd_degenerate(parent))
6326 return 1;
6328 if (!cpus_equal(sd->span, parent->span))
6329 return 0;
6331 /* Does parent contain flags not in child? */
6332 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6333 if (cflags & SD_WAKE_AFFINE)
6334 pflags &= ~SD_WAKE_BALANCE;
6335 /* Flags needing groups don't count if only 1 group in parent */
6336 if (parent->groups == parent->groups->next) {
6337 pflags &= ~(SD_LOAD_BALANCE |
6338 SD_BALANCE_NEWIDLE |
6339 SD_BALANCE_FORK |
6340 SD_BALANCE_EXEC |
6341 SD_SHARE_CPUPOWER |
6342 SD_SHARE_PKG_RESOURCES);
6344 if (~cflags & pflags)
6345 return 0;
6347 return 1;
6350 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6352 unsigned long flags;
6353 const struct sched_class *class;
6355 spin_lock_irqsave(&rq->lock, flags);
6357 if (rq->rd) {
6358 struct root_domain *old_rd = rq->rd;
6360 for (class = sched_class_highest; class; class = class->next) {
6361 if (class->leave_domain)
6362 class->leave_domain(rq);
6365 cpu_clear(rq->cpu, old_rd->span);
6366 cpu_clear(rq->cpu, old_rd->online);
6368 if (atomic_dec_and_test(&old_rd->refcount))
6369 kfree(old_rd);
6372 atomic_inc(&rd->refcount);
6373 rq->rd = rd;
6375 cpu_set(rq->cpu, rd->span);
6376 if (cpu_isset(rq->cpu, cpu_online_map))
6377 cpu_set(rq->cpu, rd->online);
6379 for (class = sched_class_highest; class; class = class->next) {
6380 if (class->join_domain)
6381 class->join_domain(rq);
6384 spin_unlock_irqrestore(&rq->lock, flags);
6387 static void init_rootdomain(struct root_domain *rd)
6389 memset(rd, 0, sizeof(*rd));
6391 cpus_clear(rd->span);
6392 cpus_clear(rd->online);
6395 static void init_defrootdomain(void)
6397 init_rootdomain(&def_root_domain);
6398 atomic_set(&def_root_domain.refcount, 1);
6401 static struct root_domain *alloc_rootdomain(void)
6403 struct root_domain *rd;
6405 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6406 if (!rd)
6407 return NULL;
6409 init_rootdomain(rd);
6411 return rd;
6415 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6416 * hold the hotplug lock.
6418 static void
6419 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6421 struct rq *rq = cpu_rq(cpu);
6422 struct sched_domain *tmp;
6424 /* Remove the sched domains which do not contribute to scheduling. */
6425 for (tmp = sd; tmp; tmp = tmp->parent) {
6426 struct sched_domain *parent = tmp->parent;
6427 if (!parent)
6428 break;
6429 if (sd_parent_degenerate(tmp, parent)) {
6430 tmp->parent = parent->parent;
6431 if (parent->parent)
6432 parent->parent->child = tmp;
6436 if (sd && sd_degenerate(sd)) {
6437 sd = sd->parent;
6438 if (sd)
6439 sd->child = NULL;
6442 sched_domain_debug(sd, cpu);
6444 rq_attach_root(rq, rd);
6445 rcu_assign_pointer(rq->sd, sd);
6448 /* cpus with isolated domains */
6449 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6451 /* Setup the mask of cpus configured for isolated domains */
6452 static int __init isolated_cpu_setup(char *str)
6454 int ints[NR_CPUS], i;
6456 str = get_options(str, ARRAY_SIZE(ints), ints);
6457 cpus_clear(cpu_isolated_map);
6458 for (i = 1; i <= ints[0]; i++)
6459 if (ints[i] < NR_CPUS)
6460 cpu_set(ints[i], cpu_isolated_map);
6461 return 1;
6464 __setup("isolcpus=", isolated_cpu_setup);
6467 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6468 * to a function which identifies what group(along with sched group) a CPU
6469 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6470 * (due to the fact that we keep track of groups covered with a cpumask_t).
6472 * init_sched_build_groups will build a circular linked list of the groups
6473 * covered by the given span, and will set each group's ->cpumask correctly,
6474 * and ->cpu_power to 0.
6476 static void
6477 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6478 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6479 struct sched_group **sg,
6480 cpumask_t *tmpmask),
6481 cpumask_t *covered, cpumask_t *tmpmask)
6483 struct sched_group *first = NULL, *last = NULL;
6484 int i;
6486 cpus_clear(*covered);
6488 for_each_cpu_mask(i, *span) {
6489 struct sched_group *sg;
6490 int group = group_fn(i, cpu_map, &sg, tmpmask);
6491 int j;
6493 if (cpu_isset(i, *covered))
6494 continue;
6496 cpus_clear(sg->cpumask);
6497 sg->__cpu_power = 0;
6499 for_each_cpu_mask(j, *span) {
6500 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6501 continue;
6503 cpu_set(j, *covered);
6504 cpu_set(j, sg->cpumask);
6506 if (!first)
6507 first = sg;
6508 if (last)
6509 last->next = sg;
6510 last = sg;
6512 last->next = first;
6515 #define SD_NODES_PER_DOMAIN 16
6517 #ifdef CONFIG_NUMA
6520 * find_next_best_node - find the next node to include in a sched_domain
6521 * @node: node whose sched_domain we're building
6522 * @used_nodes: nodes already in the sched_domain
6524 * Find the next node to include in a given scheduling domain. Simply
6525 * finds the closest node not already in the @used_nodes map.
6527 * Should use nodemask_t.
6529 static int find_next_best_node(int node, nodemask_t *used_nodes)
6531 int i, n, val, min_val, best_node = 0;
6533 min_val = INT_MAX;
6535 for (i = 0; i < MAX_NUMNODES; i++) {
6536 /* Start at @node */
6537 n = (node + i) % MAX_NUMNODES;
6539 if (!nr_cpus_node(n))
6540 continue;
6542 /* Skip already used nodes */
6543 if (node_isset(n, *used_nodes))
6544 continue;
6546 /* Simple min distance search */
6547 val = node_distance(node, n);
6549 if (val < min_val) {
6550 min_val = val;
6551 best_node = n;
6555 node_set(best_node, *used_nodes);
6556 return best_node;
6560 * sched_domain_node_span - get a cpumask for a node's sched_domain
6561 * @node: node whose cpumask we're constructing
6562 * @span: resulting cpumask
6564 * Given a node, construct a good cpumask for its sched_domain to span. It
6565 * should be one that prevents unnecessary balancing, but also spreads tasks
6566 * out optimally.
6568 static void sched_domain_node_span(int node, cpumask_t *span)
6570 nodemask_t used_nodes;
6571 node_to_cpumask_ptr(nodemask, node);
6572 int i;
6574 cpus_clear(*span);
6575 nodes_clear(used_nodes);
6577 cpus_or(*span, *span, *nodemask);
6578 node_set(node, used_nodes);
6580 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6581 int next_node = find_next_best_node(node, &used_nodes);
6583 node_to_cpumask_ptr_next(nodemask, next_node);
6584 cpus_or(*span, *span, *nodemask);
6587 #endif
6589 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6592 * SMT sched-domains:
6594 #ifdef CONFIG_SCHED_SMT
6595 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6596 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6598 static int
6599 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6600 cpumask_t *unused)
6602 if (sg)
6603 *sg = &per_cpu(sched_group_cpus, cpu);
6604 return cpu;
6606 #endif
6609 * multi-core sched-domains:
6611 #ifdef CONFIG_SCHED_MC
6612 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6613 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6614 #endif
6616 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6617 static int
6618 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6619 cpumask_t *mask)
6621 int group;
6623 *mask = per_cpu(cpu_sibling_map, cpu);
6624 cpus_and(*mask, *mask, *cpu_map);
6625 group = first_cpu(*mask);
6626 if (sg)
6627 *sg = &per_cpu(sched_group_core, group);
6628 return group;
6630 #elif defined(CONFIG_SCHED_MC)
6631 static int
6632 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6633 cpumask_t *unused)
6635 if (sg)
6636 *sg = &per_cpu(sched_group_core, cpu);
6637 return cpu;
6639 #endif
6641 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6642 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6644 static int
6645 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6646 cpumask_t *mask)
6648 int group;
6649 #ifdef CONFIG_SCHED_MC
6650 *mask = cpu_coregroup_map(cpu);
6651 cpus_and(*mask, *mask, *cpu_map);
6652 group = first_cpu(*mask);
6653 #elif defined(CONFIG_SCHED_SMT)
6654 *mask = per_cpu(cpu_sibling_map, cpu);
6655 cpus_and(*mask, *mask, *cpu_map);
6656 group = first_cpu(*mask);
6657 #else
6658 group = cpu;
6659 #endif
6660 if (sg)
6661 *sg = &per_cpu(sched_group_phys, group);
6662 return group;
6665 #ifdef CONFIG_NUMA
6667 * The init_sched_build_groups can't handle what we want to do with node
6668 * groups, so roll our own. Now each node has its own list of groups which
6669 * gets dynamically allocated.
6671 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6672 static struct sched_group ***sched_group_nodes_bycpu;
6674 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6675 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6677 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6678 struct sched_group **sg, cpumask_t *nodemask)
6680 int group;
6682 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6683 cpus_and(*nodemask, *nodemask, *cpu_map);
6684 group = first_cpu(*nodemask);
6686 if (sg)
6687 *sg = &per_cpu(sched_group_allnodes, group);
6688 return group;
6691 static void init_numa_sched_groups_power(struct sched_group *group_head)
6693 struct sched_group *sg = group_head;
6694 int j;
6696 if (!sg)
6697 return;
6698 do {
6699 for_each_cpu_mask(j, sg->cpumask) {
6700 struct sched_domain *sd;
6702 sd = &per_cpu(phys_domains, j);
6703 if (j != first_cpu(sd->groups->cpumask)) {
6705 * Only add "power" once for each
6706 * physical package.
6708 continue;
6711 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6713 sg = sg->next;
6714 } while (sg != group_head);
6716 #endif
6718 #ifdef CONFIG_NUMA
6719 /* Free memory allocated for various sched_group structures */
6720 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6722 int cpu, i;
6724 for_each_cpu_mask(cpu, *cpu_map) {
6725 struct sched_group **sched_group_nodes
6726 = sched_group_nodes_bycpu[cpu];
6728 if (!sched_group_nodes)
6729 continue;
6731 for (i = 0; i < MAX_NUMNODES; i++) {
6732 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6734 *nodemask = node_to_cpumask(i);
6735 cpus_and(*nodemask, *nodemask, *cpu_map);
6736 if (cpus_empty(*nodemask))
6737 continue;
6739 if (sg == NULL)
6740 continue;
6741 sg = sg->next;
6742 next_sg:
6743 oldsg = sg;
6744 sg = sg->next;
6745 kfree(oldsg);
6746 if (oldsg != sched_group_nodes[i])
6747 goto next_sg;
6749 kfree(sched_group_nodes);
6750 sched_group_nodes_bycpu[cpu] = NULL;
6753 #else
6754 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6757 #endif
6760 * Initialize sched groups cpu_power.
6762 * cpu_power indicates the capacity of sched group, which is used while
6763 * distributing the load between different sched groups in a sched domain.
6764 * Typically cpu_power for all the groups in a sched domain will be same unless
6765 * there are asymmetries in the topology. If there are asymmetries, group
6766 * having more cpu_power will pickup more load compared to the group having
6767 * less cpu_power.
6769 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6770 * the maximum number of tasks a group can handle in the presence of other idle
6771 * or lightly loaded groups in the same sched domain.
6773 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6775 struct sched_domain *child;
6776 struct sched_group *group;
6778 WARN_ON(!sd || !sd->groups);
6780 if (cpu != first_cpu(sd->groups->cpumask))
6781 return;
6783 child = sd->child;
6785 sd->groups->__cpu_power = 0;
6788 * For perf policy, if the groups in child domain share resources
6789 * (for example cores sharing some portions of the cache hierarchy
6790 * or SMT), then set this domain groups cpu_power such that each group
6791 * can handle only one task, when there are other idle groups in the
6792 * same sched domain.
6794 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6795 (child->flags &
6796 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6797 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6798 return;
6802 * add cpu_power of each child group to this groups cpu_power
6804 group = child->groups;
6805 do {
6806 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6807 group = group->next;
6808 } while (group != child->groups);
6812 * Initializers for schedule domains
6813 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6816 #define SD_INIT(sd, type) sd_init_##type(sd)
6817 #define SD_INIT_FUNC(type) \
6818 static noinline void sd_init_##type(struct sched_domain *sd) \
6820 memset(sd, 0, sizeof(*sd)); \
6821 *sd = SD_##type##_INIT; \
6822 sd->level = SD_LV_##type; \
6825 SD_INIT_FUNC(CPU)
6826 #ifdef CONFIG_NUMA
6827 SD_INIT_FUNC(ALLNODES)
6828 SD_INIT_FUNC(NODE)
6829 #endif
6830 #ifdef CONFIG_SCHED_SMT
6831 SD_INIT_FUNC(SIBLING)
6832 #endif
6833 #ifdef CONFIG_SCHED_MC
6834 SD_INIT_FUNC(MC)
6835 #endif
6838 * To minimize stack usage kmalloc room for cpumasks and share the
6839 * space as the usage in build_sched_domains() dictates. Used only
6840 * if the amount of space is significant.
6842 struct allmasks {
6843 cpumask_t tmpmask; /* make this one first */
6844 union {
6845 cpumask_t nodemask;
6846 cpumask_t this_sibling_map;
6847 cpumask_t this_core_map;
6849 cpumask_t send_covered;
6851 #ifdef CONFIG_NUMA
6852 cpumask_t domainspan;
6853 cpumask_t covered;
6854 cpumask_t notcovered;
6855 #endif
6858 #if NR_CPUS > 128
6859 #define SCHED_CPUMASK_ALLOC 1
6860 #define SCHED_CPUMASK_FREE(v) kfree(v)
6861 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
6862 #else
6863 #define SCHED_CPUMASK_ALLOC 0
6864 #define SCHED_CPUMASK_FREE(v)
6865 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
6866 #endif
6868 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
6869 ((unsigned long)(a) + offsetof(struct allmasks, v))
6871 static int default_relax_domain_level = -1;
6873 static int __init setup_relax_domain_level(char *str)
6875 default_relax_domain_level = simple_strtoul(str, NULL, 0);
6876 return 1;
6878 __setup("relax_domain_level=", setup_relax_domain_level);
6880 static void set_domain_attribute(struct sched_domain *sd,
6881 struct sched_domain_attr *attr)
6883 int request;
6885 if (!attr || attr->relax_domain_level < 0) {
6886 if (default_relax_domain_level < 0)
6887 return;
6888 else
6889 request = default_relax_domain_level;
6890 } else
6891 request = attr->relax_domain_level;
6892 if (request < sd->level) {
6893 /* turn off idle balance on this domain */
6894 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
6895 } else {
6896 /* turn on idle balance on this domain */
6897 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
6902 * Build sched domains for a given set of cpus and attach the sched domains
6903 * to the individual cpus
6905 static int __build_sched_domains(const cpumask_t *cpu_map,
6906 struct sched_domain_attr *attr)
6908 int i;
6909 struct root_domain *rd;
6910 SCHED_CPUMASK_DECLARE(allmasks);
6911 cpumask_t *tmpmask;
6912 #ifdef CONFIG_NUMA
6913 struct sched_group **sched_group_nodes = NULL;
6914 int sd_allnodes = 0;
6917 * Allocate the per-node list of sched groups
6919 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6920 GFP_KERNEL);
6921 if (!sched_group_nodes) {
6922 printk(KERN_WARNING "Can not alloc sched group node list\n");
6923 return -ENOMEM;
6925 #endif
6927 rd = alloc_rootdomain();
6928 if (!rd) {
6929 printk(KERN_WARNING "Cannot alloc root domain\n");
6930 #ifdef CONFIG_NUMA
6931 kfree(sched_group_nodes);
6932 #endif
6933 return -ENOMEM;
6936 #if SCHED_CPUMASK_ALLOC
6937 /* get space for all scratch cpumask variables */
6938 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
6939 if (!allmasks) {
6940 printk(KERN_WARNING "Cannot alloc cpumask array\n");
6941 kfree(rd);
6942 #ifdef CONFIG_NUMA
6943 kfree(sched_group_nodes);
6944 #endif
6945 return -ENOMEM;
6947 #endif
6948 tmpmask = (cpumask_t *)allmasks;
6951 #ifdef CONFIG_NUMA
6952 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6953 #endif
6956 * Set up domains for cpus specified by the cpu_map.
6958 for_each_cpu_mask(i, *cpu_map) {
6959 struct sched_domain *sd = NULL, *p;
6960 SCHED_CPUMASK_VAR(nodemask, allmasks);
6962 *nodemask = node_to_cpumask(cpu_to_node(i));
6963 cpus_and(*nodemask, *nodemask, *cpu_map);
6965 #ifdef CONFIG_NUMA
6966 if (cpus_weight(*cpu_map) >
6967 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
6968 sd = &per_cpu(allnodes_domains, i);
6969 SD_INIT(sd, ALLNODES);
6970 set_domain_attribute(sd, attr);
6971 sd->span = *cpu_map;
6972 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
6973 p = sd;
6974 sd_allnodes = 1;
6975 } else
6976 p = NULL;
6978 sd = &per_cpu(node_domains, i);
6979 SD_INIT(sd, NODE);
6980 set_domain_attribute(sd, attr);
6981 sched_domain_node_span(cpu_to_node(i), &sd->span);
6982 sd->parent = p;
6983 if (p)
6984 p->child = sd;
6985 cpus_and(sd->span, sd->span, *cpu_map);
6986 #endif
6988 p = sd;
6989 sd = &per_cpu(phys_domains, i);
6990 SD_INIT(sd, CPU);
6991 set_domain_attribute(sd, attr);
6992 sd->span = *nodemask;
6993 sd->parent = p;
6994 if (p)
6995 p->child = sd;
6996 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
6998 #ifdef CONFIG_SCHED_MC
6999 p = sd;
7000 sd = &per_cpu(core_domains, i);
7001 SD_INIT(sd, MC);
7002 set_domain_attribute(sd, attr);
7003 sd->span = cpu_coregroup_map(i);
7004 cpus_and(sd->span, sd->span, *cpu_map);
7005 sd->parent = p;
7006 p->child = sd;
7007 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7008 #endif
7010 #ifdef CONFIG_SCHED_SMT
7011 p = sd;
7012 sd = &per_cpu(cpu_domains, i);
7013 SD_INIT(sd, SIBLING);
7014 set_domain_attribute(sd, attr);
7015 sd->span = per_cpu(cpu_sibling_map, i);
7016 cpus_and(sd->span, sd->span, *cpu_map);
7017 sd->parent = p;
7018 p->child = sd;
7019 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7020 #endif
7023 #ifdef CONFIG_SCHED_SMT
7024 /* Set up CPU (sibling) groups */
7025 for_each_cpu_mask(i, *cpu_map) {
7026 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7027 SCHED_CPUMASK_VAR(send_covered, allmasks);
7029 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7030 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7031 if (i != first_cpu(*this_sibling_map))
7032 continue;
7034 init_sched_build_groups(this_sibling_map, cpu_map,
7035 &cpu_to_cpu_group,
7036 send_covered, tmpmask);
7038 #endif
7040 #ifdef CONFIG_SCHED_MC
7041 /* Set up multi-core groups */
7042 for_each_cpu_mask(i, *cpu_map) {
7043 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7044 SCHED_CPUMASK_VAR(send_covered, allmasks);
7046 *this_core_map = cpu_coregroup_map(i);
7047 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7048 if (i != first_cpu(*this_core_map))
7049 continue;
7051 init_sched_build_groups(this_core_map, cpu_map,
7052 &cpu_to_core_group,
7053 send_covered, tmpmask);
7055 #endif
7057 /* Set up physical groups */
7058 for (i = 0; i < MAX_NUMNODES; i++) {
7059 SCHED_CPUMASK_VAR(nodemask, allmasks);
7060 SCHED_CPUMASK_VAR(send_covered, allmasks);
7062 *nodemask = node_to_cpumask(i);
7063 cpus_and(*nodemask, *nodemask, *cpu_map);
7064 if (cpus_empty(*nodemask))
7065 continue;
7067 init_sched_build_groups(nodemask, cpu_map,
7068 &cpu_to_phys_group,
7069 send_covered, tmpmask);
7072 #ifdef CONFIG_NUMA
7073 /* Set up node groups */
7074 if (sd_allnodes) {
7075 SCHED_CPUMASK_VAR(send_covered, allmasks);
7077 init_sched_build_groups(cpu_map, cpu_map,
7078 &cpu_to_allnodes_group,
7079 send_covered, tmpmask);
7082 for (i = 0; i < MAX_NUMNODES; i++) {
7083 /* Set up node groups */
7084 struct sched_group *sg, *prev;
7085 SCHED_CPUMASK_VAR(nodemask, allmasks);
7086 SCHED_CPUMASK_VAR(domainspan, allmasks);
7087 SCHED_CPUMASK_VAR(covered, allmasks);
7088 int j;
7090 *nodemask = node_to_cpumask(i);
7091 cpus_clear(*covered);
7093 cpus_and(*nodemask, *nodemask, *cpu_map);
7094 if (cpus_empty(*nodemask)) {
7095 sched_group_nodes[i] = NULL;
7096 continue;
7099 sched_domain_node_span(i, domainspan);
7100 cpus_and(*domainspan, *domainspan, *cpu_map);
7102 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7103 if (!sg) {
7104 printk(KERN_WARNING "Can not alloc domain group for "
7105 "node %d\n", i);
7106 goto error;
7108 sched_group_nodes[i] = sg;
7109 for_each_cpu_mask(j, *nodemask) {
7110 struct sched_domain *sd;
7112 sd = &per_cpu(node_domains, j);
7113 sd->groups = sg;
7115 sg->__cpu_power = 0;
7116 sg->cpumask = *nodemask;
7117 sg->next = sg;
7118 cpus_or(*covered, *covered, *nodemask);
7119 prev = sg;
7121 for (j = 0; j < MAX_NUMNODES; j++) {
7122 SCHED_CPUMASK_VAR(notcovered, allmasks);
7123 int n = (i + j) % MAX_NUMNODES;
7124 node_to_cpumask_ptr(pnodemask, n);
7126 cpus_complement(*notcovered, *covered);
7127 cpus_and(*tmpmask, *notcovered, *cpu_map);
7128 cpus_and(*tmpmask, *tmpmask, *domainspan);
7129 if (cpus_empty(*tmpmask))
7130 break;
7132 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7133 if (cpus_empty(*tmpmask))
7134 continue;
7136 sg = kmalloc_node(sizeof(struct sched_group),
7137 GFP_KERNEL, i);
7138 if (!sg) {
7139 printk(KERN_WARNING
7140 "Can not alloc domain group for node %d\n", j);
7141 goto error;
7143 sg->__cpu_power = 0;
7144 sg->cpumask = *tmpmask;
7145 sg->next = prev->next;
7146 cpus_or(*covered, *covered, *tmpmask);
7147 prev->next = sg;
7148 prev = sg;
7151 #endif
7153 /* Calculate CPU power for physical packages and nodes */
7154 #ifdef CONFIG_SCHED_SMT
7155 for_each_cpu_mask(i, *cpu_map) {
7156 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7158 init_sched_groups_power(i, sd);
7160 #endif
7161 #ifdef CONFIG_SCHED_MC
7162 for_each_cpu_mask(i, *cpu_map) {
7163 struct sched_domain *sd = &per_cpu(core_domains, i);
7165 init_sched_groups_power(i, sd);
7167 #endif
7169 for_each_cpu_mask(i, *cpu_map) {
7170 struct sched_domain *sd = &per_cpu(phys_domains, i);
7172 init_sched_groups_power(i, sd);
7175 #ifdef CONFIG_NUMA
7176 for (i = 0; i < MAX_NUMNODES; i++)
7177 init_numa_sched_groups_power(sched_group_nodes[i]);
7179 if (sd_allnodes) {
7180 struct sched_group *sg;
7182 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7183 tmpmask);
7184 init_numa_sched_groups_power(sg);
7186 #endif
7188 /* Attach the domains */
7189 for_each_cpu_mask(i, *cpu_map) {
7190 struct sched_domain *sd;
7191 #ifdef CONFIG_SCHED_SMT
7192 sd = &per_cpu(cpu_domains, i);
7193 #elif defined(CONFIG_SCHED_MC)
7194 sd = &per_cpu(core_domains, i);
7195 #else
7196 sd = &per_cpu(phys_domains, i);
7197 #endif
7198 cpu_attach_domain(sd, rd, i);
7201 SCHED_CPUMASK_FREE((void *)allmasks);
7202 return 0;
7204 #ifdef CONFIG_NUMA
7205 error:
7206 free_sched_groups(cpu_map, tmpmask);
7207 SCHED_CPUMASK_FREE((void *)allmasks);
7208 return -ENOMEM;
7209 #endif
7212 static int build_sched_domains(const cpumask_t *cpu_map)
7214 return __build_sched_domains(cpu_map, NULL);
7217 static cpumask_t *doms_cur; /* current sched domains */
7218 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7219 static struct sched_domain_attr *dattr_cur;
7220 /* attribues of custom domains in 'doms_cur' */
7223 * Special case: If a kmalloc of a doms_cur partition (array of
7224 * cpumask_t) fails, then fallback to a single sched domain,
7225 * as determined by the single cpumask_t fallback_doms.
7227 static cpumask_t fallback_doms;
7229 void __attribute__((weak)) arch_update_cpu_topology(void)
7234 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7235 * For now this just excludes isolated cpus, but could be used to
7236 * exclude other special cases in the future.
7238 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7240 int err;
7242 arch_update_cpu_topology();
7243 ndoms_cur = 1;
7244 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7245 if (!doms_cur)
7246 doms_cur = &fallback_doms;
7247 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7248 dattr_cur = NULL;
7249 err = build_sched_domains(doms_cur);
7250 register_sched_domain_sysctl();
7252 return err;
7255 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7256 cpumask_t *tmpmask)
7258 free_sched_groups(cpu_map, tmpmask);
7262 * Detach sched domains from a group of cpus specified in cpu_map
7263 * These cpus will now be attached to the NULL domain
7265 static void detach_destroy_domains(const cpumask_t *cpu_map)
7267 cpumask_t tmpmask;
7268 int i;
7270 unregister_sched_domain_sysctl();
7272 for_each_cpu_mask(i, *cpu_map)
7273 cpu_attach_domain(NULL, &def_root_domain, i);
7274 synchronize_sched();
7275 arch_destroy_sched_domains(cpu_map, &tmpmask);
7278 /* handle null as "default" */
7279 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7280 struct sched_domain_attr *new, int idx_new)
7282 struct sched_domain_attr tmp;
7284 /* fast path */
7285 if (!new && !cur)
7286 return 1;
7288 tmp = SD_ATTR_INIT;
7289 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7290 new ? (new + idx_new) : &tmp,
7291 sizeof(struct sched_domain_attr));
7295 * Partition sched domains as specified by the 'ndoms_new'
7296 * cpumasks in the array doms_new[] of cpumasks. This compares
7297 * doms_new[] to the current sched domain partitioning, doms_cur[].
7298 * It destroys each deleted domain and builds each new domain.
7300 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7301 * The masks don't intersect (don't overlap.) We should setup one
7302 * sched domain for each mask. CPUs not in any of the cpumasks will
7303 * not be load balanced. If the same cpumask appears both in the
7304 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7305 * it as it is.
7307 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7308 * ownership of it and will kfree it when done with it. If the caller
7309 * failed the kmalloc call, then it can pass in doms_new == NULL,
7310 * and partition_sched_domains() will fallback to the single partition
7311 * 'fallback_doms'.
7313 * Call with hotplug lock held
7315 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7316 struct sched_domain_attr *dattr_new)
7318 int i, j;
7320 mutex_lock(&sched_domains_mutex);
7322 /* always unregister in case we don't destroy any domains */
7323 unregister_sched_domain_sysctl();
7325 if (doms_new == NULL) {
7326 ndoms_new = 1;
7327 doms_new = &fallback_doms;
7328 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7329 dattr_new = NULL;
7332 /* Destroy deleted domains */
7333 for (i = 0; i < ndoms_cur; i++) {
7334 for (j = 0; j < ndoms_new; j++) {
7335 if (cpus_equal(doms_cur[i], doms_new[j])
7336 && dattrs_equal(dattr_cur, i, dattr_new, j))
7337 goto match1;
7339 /* no match - a current sched domain not in new doms_new[] */
7340 detach_destroy_domains(doms_cur + i);
7341 match1:
7345 /* Build new domains */
7346 for (i = 0; i < ndoms_new; i++) {
7347 for (j = 0; j < ndoms_cur; j++) {
7348 if (cpus_equal(doms_new[i], doms_cur[j])
7349 && dattrs_equal(dattr_new, i, dattr_cur, j))
7350 goto match2;
7352 /* no match - add a new doms_new */
7353 __build_sched_domains(doms_new + i,
7354 dattr_new ? dattr_new + i : NULL);
7355 match2:
7359 /* Remember the new sched domains */
7360 if (doms_cur != &fallback_doms)
7361 kfree(doms_cur);
7362 kfree(dattr_cur); /* kfree(NULL) is safe */
7363 doms_cur = doms_new;
7364 dattr_cur = dattr_new;
7365 ndoms_cur = ndoms_new;
7367 register_sched_domain_sysctl();
7369 mutex_unlock(&sched_domains_mutex);
7372 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7373 int arch_reinit_sched_domains(void)
7375 int err;
7377 get_online_cpus();
7378 mutex_lock(&sched_domains_mutex);
7379 detach_destroy_domains(&cpu_online_map);
7380 err = arch_init_sched_domains(&cpu_online_map);
7381 mutex_unlock(&sched_domains_mutex);
7382 put_online_cpus();
7384 return err;
7387 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7389 int ret;
7391 if (buf[0] != '0' && buf[0] != '1')
7392 return -EINVAL;
7394 if (smt)
7395 sched_smt_power_savings = (buf[0] == '1');
7396 else
7397 sched_mc_power_savings = (buf[0] == '1');
7399 ret = arch_reinit_sched_domains();
7401 return ret ? ret : count;
7404 #ifdef CONFIG_SCHED_MC
7405 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7407 return sprintf(page, "%u\n", sched_mc_power_savings);
7409 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7410 const char *buf, size_t count)
7412 return sched_power_savings_store(buf, count, 0);
7414 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7415 sched_mc_power_savings_store);
7416 #endif
7418 #ifdef CONFIG_SCHED_SMT
7419 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7421 return sprintf(page, "%u\n", sched_smt_power_savings);
7423 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7424 const char *buf, size_t count)
7426 return sched_power_savings_store(buf, count, 1);
7428 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7429 sched_smt_power_savings_store);
7430 #endif
7432 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7434 int err = 0;
7436 #ifdef CONFIG_SCHED_SMT
7437 if (smt_capable())
7438 err = sysfs_create_file(&cls->kset.kobj,
7439 &attr_sched_smt_power_savings.attr);
7440 #endif
7441 #ifdef CONFIG_SCHED_MC
7442 if (!err && mc_capable())
7443 err = sysfs_create_file(&cls->kset.kobj,
7444 &attr_sched_mc_power_savings.attr);
7445 #endif
7446 return err;
7448 #endif
7451 * Force a reinitialization of the sched domains hierarchy. The domains
7452 * and groups cannot be updated in place without racing with the balancing
7453 * code, so we temporarily attach all running cpus to the NULL domain
7454 * which will prevent rebalancing while the sched domains are recalculated.
7456 static int update_sched_domains(struct notifier_block *nfb,
7457 unsigned long action, void *hcpu)
7459 switch (action) {
7460 case CPU_UP_PREPARE:
7461 case CPU_UP_PREPARE_FROZEN:
7462 case CPU_DOWN_PREPARE:
7463 case CPU_DOWN_PREPARE_FROZEN:
7464 detach_destroy_domains(&cpu_online_map);
7465 return NOTIFY_OK;
7467 case CPU_UP_CANCELED:
7468 case CPU_UP_CANCELED_FROZEN:
7469 case CPU_DOWN_FAILED:
7470 case CPU_DOWN_FAILED_FROZEN:
7471 case CPU_ONLINE:
7472 case CPU_ONLINE_FROZEN:
7473 case CPU_DEAD:
7474 case CPU_DEAD_FROZEN:
7476 * Fall through and re-initialise the domains.
7478 break;
7479 default:
7480 return NOTIFY_DONE;
7483 /* The hotplug lock is already held by cpu_up/cpu_down */
7484 arch_init_sched_domains(&cpu_online_map);
7486 return NOTIFY_OK;
7489 void __init sched_init_smp(void)
7491 cpumask_t non_isolated_cpus;
7493 #if defined(CONFIG_NUMA)
7494 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7495 GFP_KERNEL);
7496 BUG_ON(sched_group_nodes_bycpu == NULL);
7497 #endif
7498 get_online_cpus();
7499 mutex_lock(&sched_domains_mutex);
7500 arch_init_sched_domains(&cpu_online_map);
7501 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7502 if (cpus_empty(non_isolated_cpus))
7503 cpu_set(smp_processor_id(), non_isolated_cpus);
7504 mutex_unlock(&sched_domains_mutex);
7505 put_online_cpus();
7506 /* XXX: Theoretical race here - CPU may be hotplugged now */
7507 hotcpu_notifier(update_sched_domains, 0);
7508 init_hrtick();
7510 /* Move init over to a non-isolated CPU */
7511 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7512 BUG();
7513 sched_init_granularity();
7515 #else
7516 void __init sched_init_smp(void)
7518 sched_init_granularity();
7520 #endif /* CONFIG_SMP */
7522 int in_sched_functions(unsigned long addr)
7524 return in_lock_functions(addr) ||
7525 (addr >= (unsigned long)__sched_text_start
7526 && addr < (unsigned long)__sched_text_end);
7529 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7531 cfs_rq->tasks_timeline = RB_ROOT;
7532 INIT_LIST_HEAD(&cfs_rq->tasks);
7533 #ifdef CONFIG_FAIR_GROUP_SCHED
7534 cfs_rq->rq = rq;
7535 #endif
7536 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7539 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7541 struct rt_prio_array *array;
7542 int i;
7544 array = &rt_rq->active;
7545 for (i = 0; i < MAX_RT_PRIO; i++) {
7546 INIT_LIST_HEAD(array->queue + i);
7547 __clear_bit(i, array->bitmap);
7549 /* delimiter for bitsearch: */
7550 __set_bit(MAX_RT_PRIO, array->bitmap);
7552 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7553 rt_rq->highest_prio = MAX_RT_PRIO;
7554 #endif
7555 #ifdef CONFIG_SMP
7556 rt_rq->rt_nr_migratory = 0;
7557 rt_rq->overloaded = 0;
7558 #endif
7560 rt_rq->rt_time = 0;
7561 rt_rq->rt_throttled = 0;
7562 rt_rq->rt_runtime = 0;
7563 spin_lock_init(&rt_rq->rt_runtime_lock);
7565 #ifdef CONFIG_RT_GROUP_SCHED
7566 rt_rq->rt_nr_boosted = 0;
7567 rt_rq->rq = rq;
7568 #endif
7571 #ifdef CONFIG_FAIR_GROUP_SCHED
7572 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7573 struct sched_entity *se, int cpu, int add,
7574 struct sched_entity *parent)
7576 struct rq *rq = cpu_rq(cpu);
7577 tg->cfs_rq[cpu] = cfs_rq;
7578 init_cfs_rq(cfs_rq, rq);
7579 cfs_rq->tg = tg;
7580 if (add)
7581 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7583 tg->se[cpu] = se;
7584 /* se could be NULL for init_task_group */
7585 if (!se)
7586 return;
7588 if (!parent)
7589 se->cfs_rq = &rq->cfs;
7590 else
7591 se->cfs_rq = parent->my_q;
7593 se->my_q = cfs_rq;
7594 se->load.weight = tg->shares;
7595 se->load.inv_weight = 0;
7596 se->parent = parent;
7598 #endif
7600 #ifdef CONFIG_RT_GROUP_SCHED
7601 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7602 struct sched_rt_entity *rt_se, int cpu, int add,
7603 struct sched_rt_entity *parent)
7605 struct rq *rq = cpu_rq(cpu);
7607 tg->rt_rq[cpu] = rt_rq;
7608 init_rt_rq(rt_rq, rq);
7609 rt_rq->tg = tg;
7610 rt_rq->rt_se = rt_se;
7611 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7612 if (add)
7613 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7615 tg->rt_se[cpu] = rt_se;
7616 if (!rt_se)
7617 return;
7619 if (!parent)
7620 rt_se->rt_rq = &rq->rt;
7621 else
7622 rt_se->rt_rq = parent->my_q;
7624 rt_se->rt_rq = &rq->rt;
7625 rt_se->my_q = rt_rq;
7626 rt_se->parent = parent;
7627 INIT_LIST_HEAD(&rt_se->run_list);
7629 #endif
7631 void __init sched_init(void)
7633 int i, j;
7634 unsigned long alloc_size = 0, ptr;
7636 #ifdef CONFIG_FAIR_GROUP_SCHED
7637 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7638 #endif
7639 #ifdef CONFIG_RT_GROUP_SCHED
7640 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7641 #endif
7642 #ifdef CONFIG_USER_SCHED
7643 alloc_size *= 2;
7644 #endif
7646 * As sched_init() is called before page_alloc is setup,
7647 * we use alloc_bootmem().
7649 if (alloc_size) {
7650 ptr = (unsigned long)alloc_bootmem(alloc_size);
7652 #ifdef CONFIG_FAIR_GROUP_SCHED
7653 init_task_group.se = (struct sched_entity **)ptr;
7654 ptr += nr_cpu_ids * sizeof(void **);
7656 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7657 ptr += nr_cpu_ids * sizeof(void **);
7659 #ifdef CONFIG_USER_SCHED
7660 root_task_group.se = (struct sched_entity **)ptr;
7661 ptr += nr_cpu_ids * sizeof(void **);
7663 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7664 ptr += nr_cpu_ids * sizeof(void **);
7665 #endif
7666 #endif
7667 #ifdef CONFIG_RT_GROUP_SCHED
7668 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7669 ptr += nr_cpu_ids * sizeof(void **);
7671 init_task_group.rt_rq = (struct rt_rq **)ptr;
7672 ptr += nr_cpu_ids * sizeof(void **);
7674 #ifdef CONFIG_USER_SCHED
7675 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7676 ptr += nr_cpu_ids * sizeof(void **);
7678 root_task_group.rt_rq = (struct rt_rq **)ptr;
7679 ptr += nr_cpu_ids * sizeof(void **);
7680 #endif
7681 #endif
7684 #ifdef CONFIG_SMP
7685 init_defrootdomain();
7686 #endif
7688 init_rt_bandwidth(&def_rt_bandwidth,
7689 global_rt_period(), global_rt_runtime());
7691 #ifdef CONFIG_RT_GROUP_SCHED
7692 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7693 global_rt_period(), global_rt_runtime());
7694 #ifdef CONFIG_USER_SCHED
7695 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7696 global_rt_period(), RUNTIME_INF);
7697 #endif
7698 #endif
7700 #ifdef CONFIG_GROUP_SCHED
7701 list_add(&init_task_group.list, &task_groups);
7702 INIT_LIST_HEAD(&init_task_group.children);
7704 #ifdef CONFIG_USER_SCHED
7705 INIT_LIST_HEAD(&root_task_group.children);
7706 init_task_group.parent = &root_task_group;
7707 list_add(&init_task_group.siblings, &root_task_group.children);
7708 #endif
7709 #endif
7711 for_each_possible_cpu(i) {
7712 struct rq *rq;
7714 rq = cpu_rq(i);
7715 spin_lock_init(&rq->lock);
7716 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7717 rq->nr_running = 0;
7718 init_cfs_rq(&rq->cfs, rq);
7719 init_rt_rq(&rq->rt, rq);
7720 #ifdef CONFIG_FAIR_GROUP_SCHED
7721 init_task_group.shares = init_task_group_load;
7722 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7723 #ifdef CONFIG_CGROUP_SCHED
7725 * How much cpu bandwidth does init_task_group get?
7727 * In case of task-groups formed thr' the cgroup filesystem, it
7728 * gets 100% of the cpu resources in the system. This overall
7729 * system cpu resource is divided among the tasks of
7730 * init_task_group and its child task-groups in a fair manner,
7731 * based on each entity's (task or task-group's) weight
7732 * (se->load.weight).
7734 * In other words, if init_task_group has 10 tasks of weight
7735 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7736 * then A0's share of the cpu resource is:
7738 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7740 * We achieve this by letting init_task_group's tasks sit
7741 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7743 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7744 #elif defined CONFIG_USER_SCHED
7745 root_task_group.shares = NICE_0_LOAD;
7746 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
7748 * In case of task-groups formed thr' the user id of tasks,
7749 * init_task_group represents tasks belonging to root user.
7750 * Hence it forms a sibling of all subsequent groups formed.
7751 * In this case, init_task_group gets only a fraction of overall
7752 * system cpu resource, based on the weight assigned to root
7753 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7754 * by letting tasks of init_task_group sit in a separate cfs_rq
7755 * (init_cfs_rq) and having one entity represent this group of
7756 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7758 init_tg_cfs_entry(&init_task_group,
7759 &per_cpu(init_cfs_rq, i),
7760 &per_cpu(init_sched_entity, i), i, 1,
7761 root_task_group.se[i]);
7763 #endif
7764 #endif /* CONFIG_FAIR_GROUP_SCHED */
7766 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7767 #ifdef CONFIG_RT_GROUP_SCHED
7768 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7769 #ifdef CONFIG_CGROUP_SCHED
7770 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7771 #elif defined CONFIG_USER_SCHED
7772 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
7773 init_tg_rt_entry(&init_task_group,
7774 &per_cpu(init_rt_rq, i),
7775 &per_cpu(init_sched_rt_entity, i), i, 1,
7776 root_task_group.rt_se[i]);
7777 #endif
7778 #endif
7780 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7781 rq->cpu_load[j] = 0;
7782 #ifdef CONFIG_SMP
7783 rq->sd = NULL;
7784 rq->rd = NULL;
7785 rq->active_balance = 0;
7786 rq->next_balance = jiffies;
7787 rq->push_cpu = 0;
7788 rq->cpu = i;
7789 rq->migration_thread = NULL;
7790 INIT_LIST_HEAD(&rq->migration_queue);
7791 rq_attach_root(rq, &def_root_domain);
7792 #endif
7793 init_rq_hrtick(rq);
7794 atomic_set(&rq->nr_iowait, 0);
7797 set_load_weight(&init_task);
7799 #ifdef CONFIG_PREEMPT_NOTIFIERS
7800 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7801 #endif
7803 #ifdef CONFIG_SMP
7804 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7805 #endif
7807 #ifdef CONFIG_RT_MUTEXES
7808 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7809 #endif
7812 * The boot idle thread does lazy MMU switching as well:
7814 atomic_inc(&init_mm.mm_count);
7815 enter_lazy_tlb(&init_mm, current);
7818 * Make us the idle thread. Technically, schedule() should not be
7819 * called from this thread, however somewhere below it might be,
7820 * but because we are the idle thread, we just pick up running again
7821 * when this runqueue becomes "idle".
7823 init_idle(current, smp_processor_id());
7825 * During early bootup we pretend to be a normal task:
7827 current->sched_class = &fair_sched_class;
7829 scheduler_running = 1;
7832 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7833 void __might_sleep(char *file, int line)
7835 #ifdef in_atomic
7836 static unsigned long prev_jiffy; /* ratelimiting */
7838 if ((in_atomic() || irqs_disabled()) &&
7839 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7840 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7841 return;
7842 prev_jiffy = jiffies;
7843 printk(KERN_ERR "BUG: sleeping function called from invalid"
7844 " context at %s:%d\n", file, line);
7845 printk("in_atomic():%d, irqs_disabled():%d\n",
7846 in_atomic(), irqs_disabled());
7847 debug_show_held_locks(current);
7848 if (irqs_disabled())
7849 print_irqtrace_events(current);
7850 dump_stack();
7852 #endif
7854 EXPORT_SYMBOL(__might_sleep);
7855 #endif
7857 #ifdef CONFIG_MAGIC_SYSRQ
7858 static void normalize_task(struct rq *rq, struct task_struct *p)
7860 int on_rq;
7862 update_rq_clock(rq);
7863 on_rq = p->se.on_rq;
7864 if (on_rq)
7865 deactivate_task(rq, p, 0);
7866 __setscheduler(rq, p, SCHED_NORMAL, 0);
7867 if (on_rq) {
7868 activate_task(rq, p, 0);
7869 resched_task(rq->curr);
7873 void normalize_rt_tasks(void)
7875 struct task_struct *g, *p;
7876 unsigned long flags;
7877 struct rq *rq;
7879 read_lock_irqsave(&tasklist_lock, flags);
7880 do_each_thread(g, p) {
7882 * Only normalize user tasks:
7884 if (!p->mm)
7885 continue;
7887 p->se.exec_start = 0;
7888 #ifdef CONFIG_SCHEDSTATS
7889 p->se.wait_start = 0;
7890 p->se.sleep_start = 0;
7891 p->se.block_start = 0;
7892 #endif
7894 if (!rt_task(p)) {
7896 * Renice negative nice level userspace
7897 * tasks back to 0:
7899 if (TASK_NICE(p) < 0 && p->mm)
7900 set_user_nice(p, 0);
7901 continue;
7904 spin_lock(&p->pi_lock);
7905 rq = __task_rq_lock(p);
7907 normalize_task(rq, p);
7909 __task_rq_unlock(rq);
7910 spin_unlock(&p->pi_lock);
7911 } while_each_thread(g, p);
7913 read_unlock_irqrestore(&tasklist_lock, flags);
7916 #endif /* CONFIG_MAGIC_SYSRQ */
7918 #ifdef CONFIG_IA64
7920 * These functions are only useful for the IA64 MCA handling.
7922 * They can only be called when the whole system has been
7923 * stopped - every CPU needs to be quiescent, and no scheduling
7924 * activity can take place. Using them for anything else would
7925 * be a serious bug, and as a result, they aren't even visible
7926 * under any other configuration.
7930 * curr_task - return the current task for a given cpu.
7931 * @cpu: the processor in question.
7933 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7935 struct task_struct *curr_task(int cpu)
7937 return cpu_curr(cpu);
7941 * set_curr_task - set the current task for a given cpu.
7942 * @cpu: the processor in question.
7943 * @p: the task pointer to set.
7945 * Description: This function must only be used when non-maskable interrupts
7946 * are serviced on a separate stack. It allows the architecture to switch the
7947 * notion of the current task on a cpu in a non-blocking manner. This function
7948 * must be called with all CPU's synchronized, and interrupts disabled, the
7949 * and caller must save the original value of the current task (see
7950 * curr_task() above) and restore that value before reenabling interrupts and
7951 * re-starting the system.
7953 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7955 void set_curr_task(int cpu, struct task_struct *p)
7957 cpu_curr(cpu) = p;
7960 #endif
7962 #ifdef CONFIG_FAIR_GROUP_SCHED
7963 static void free_fair_sched_group(struct task_group *tg)
7965 int i;
7967 for_each_possible_cpu(i) {
7968 if (tg->cfs_rq)
7969 kfree(tg->cfs_rq[i]);
7970 if (tg->se)
7971 kfree(tg->se[i]);
7974 kfree(tg->cfs_rq);
7975 kfree(tg->se);
7978 static
7979 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7981 struct cfs_rq *cfs_rq;
7982 struct sched_entity *se, *parent_se;
7983 struct rq *rq;
7984 int i;
7986 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7987 if (!tg->cfs_rq)
7988 goto err;
7989 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7990 if (!tg->se)
7991 goto err;
7993 tg->shares = NICE_0_LOAD;
7995 for_each_possible_cpu(i) {
7996 rq = cpu_rq(i);
7998 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
7999 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8000 if (!cfs_rq)
8001 goto err;
8003 se = kmalloc_node(sizeof(struct sched_entity),
8004 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8005 if (!se)
8006 goto err;
8008 parent_se = parent ? parent->se[i] : NULL;
8009 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8012 return 1;
8014 err:
8015 return 0;
8018 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8020 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8021 &cpu_rq(cpu)->leaf_cfs_rq_list);
8024 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8026 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8028 #else
8029 static inline void free_fair_sched_group(struct task_group *tg)
8033 static inline
8034 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8036 return 1;
8039 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8043 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8046 #endif
8048 #ifdef CONFIG_RT_GROUP_SCHED
8049 static void free_rt_sched_group(struct task_group *tg)
8051 int i;
8053 destroy_rt_bandwidth(&tg->rt_bandwidth);
8055 for_each_possible_cpu(i) {
8056 if (tg->rt_rq)
8057 kfree(tg->rt_rq[i]);
8058 if (tg->rt_se)
8059 kfree(tg->rt_se[i]);
8062 kfree(tg->rt_rq);
8063 kfree(tg->rt_se);
8066 static
8067 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8069 struct rt_rq *rt_rq;
8070 struct sched_rt_entity *rt_se, *parent_se;
8071 struct rq *rq;
8072 int i;
8074 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8075 if (!tg->rt_rq)
8076 goto err;
8077 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8078 if (!tg->rt_se)
8079 goto err;
8081 init_rt_bandwidth(&tg->rt_bandwidth,
8082 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8084 for_each_possible_cpu(i) {
8085 rq = cpu_rq(i);
8087 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8088 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8089 if (!rt_rq)
8090 goto err;
8092 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8093 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8094 if (!rt_se)
8095 goto err;
8097 parent_se = parent ? parent->rt_se[i] : NULL;
8098 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8101 return 1;
8103 err:
8104 return 0;
8107 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8109 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8110 &cpu_rq(cpu)->leaf_rt_rq_list);
8113 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8115 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8117 #else
8118 static inline void free_rt_sched_group(struct task_group *tg)
8122 static inline
8123 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8125 return 1;
8128 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8132 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8135 #endif
8137 #ifdef CONFIG_GROUP_SCHED
8138 static void free_sched_group(struct task_group *tg)
8140 free_fair_sched_group(tg);
8141 free_rt_sched_group(tg);
8142 kfree(tg);
8145 /* allocate runqueue etc for a new task group */
8146 struct task_group *sched_create_group(struct task_group *parent)
8148 struct task_group *tg;
8149 unsigned long flags;
8150 int i;
8152 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8153 if (!tg)
8154 return ERR_PTR(-ENOMEM);
8156 if (!alloc_fair_sched_group(tg, parent))
8157 goto err;
8159 if (!alloc_rt_sched_group(tg, parent))
8160 goto err;
8162 spin_lock_irqsave(&task_group_lock, flags);
8163 for_each_possible_cpu(i) {
8164 register_fair_sched_group(tg, i);
8165 register_rt_sched_group(tg, i);
8167 list_add_rcu(&tg->list, &task_groups);
8169 WARN_ON(!parent); /* root should already exist */
8171 tg->parent = parent;
8172 list_add_rcu(&tg->siblings, &parent->children);
8173 INIT_LIST_HEAD(&tg->children);
8174 spin_unlock_irqrestore(&task_group_lock, flags);
8176 return tg;
8178 err:
8179 free_sched_group(tg);
8180 return ERR_PTR(-ENOMEM);
8183 /* rcu callback to free various structures associated with a task group */
8184 static void free_sched_group_rcu(struct rcu_head *rhp)
8186 /* now it should be safe to free those cfs_rqs */
8187 free_sched_group(container_of(rhp, struct task_group, rcu));
8190 /* Destroy runqueue etc associated with a task group */
8191 void sched_destroy_group(struct task_group *tg)
8193 unsigned long flags;
8194 int i;
8196 spin_lock_irqsave(&task_group_lock, flags);
8197 for_each_possible_cpu(i) {
8198 unregister_fair_sched_group(tg, i);
8199 unregister_rt_sched_group(tg, i);
8201 list_del_rcu(&tg->list);
8202 list_del_rcu(&tg->siblings);
8203 spin_unlock_irqrestore(&task_group_lock, flags);
8205 /* wait for possible concurrent references to cfs_rqs complete */
8206 call_rcu(&tg->rcu, free_sched_group_rcu);
8209 /* change task's runqueue when it moves between groups.
8210 * The caller of this function should have put the task in its new group
8211 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8212 * reflect its new group.
8214 void sched_move_task(struct task_struct *tsk)
8216 int on_rq, running;
8217 unsigned long flags;
8218 struct rq *rq;
8220 rq = task_rq_lock(tsk, &flags);
8222 update_rq_clock(rq);
8224 running = task_current(rq, tsk);
8225 on_rq = tsk->se.on_rq;
8227 if (on_rq)
8228 dequeue_task(rq, tsk, 0);
8229 if (unlikely(running))
8230 tsk->sched_class->put_prev_task(rq, tsk);
8232 set_task_rq(tsk, task_cpu(tsk));
8234 #ifdef CONFIG_FAIR_GROUP_SCHED
8235 if (tsk->sched_class->moved_group)
8236 tsk->sched_class->moved_group(tsk);
8237 #endif
8239 if (unlikely(running))
8240 tsk->sched_class->set_curr_task(rq);
8241 if (on_rq)
8242 enqueue_task(rq, tsk, 0);
8244 task_rq_unlock(rq, &flags);
8246 #endif
8248 #ifdef CONFIG_FAIR_GROUP_SCHED
8249 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8251 struct cfs_rq *cfs_rq = se->cfs_rq;
8252 struct rq *rq = cfs_rq->rq;
8253 int on_rq;
8255 spin_lock_irq(&rq->lock);
8257 on_rq = se->on_rq;
8258 if (on_rq)
8259 dequeue_entity(cfs_rq, se, 0);
8261 se->load.weight = shares;
8262 se->load.inv_weight = 0;
8264 if (on_rq)
8265 enqueue_entity(cfs_rq, se, 0);
8267 spin_unlock_irq(&rq->lock);
8270 static DEFINE_MUTEX(shares_mutex);
8272 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8274 int i;
8275 unsigned long flags;
8278 * We can't change the weight of the root cgroup.
8280 if (!tg->se[0])
8281 return -EINVAL;
8283 if (shares < MIN_SHARES)
8284 shares = MIN_SHARES;
8285 else if (shares > MAX_SHARES)
8286 shares = MAX_SHARES;
8288 mutex_lock(&shares_mutex);
8289 if (tg->shares == shares)
8290 goto done;
8292 spin_lock_irqsave(&task_group_lock, flags);
8293 for_each_possible_cpu(i)
8294 unregister_fair_sched_group(tg, i);
8295 list_del_rcu(&tg->siblings);
8296 spin_unlock_irqrestore(&task_group_lock, flags);
8298 /* wait for any ongoing reference to this group to finish */
8299 synchronize_sched();
8302 * Now we are free to modify the group's share on each cpu
8303 * w/o tripping rebalance_share or load_balance_fair.
8305 tg->shares = shares;
8306 for_each_possible_cpu(i)
8307 set_se_shares(tg->se[i], shares);
8310 * Enable load balance activity on this group, by inserting it back on
8311 * each cpu's rq->leaf_cfs_rq_list.
8313 spin_lock_irqsave(&task_group_lock, flags);
8314 for_each_possible_cpu(i)
8315 register_fair_sched_group(tg, i);
8316 list_add_rcu(&tg->siblings, &tg->parent->children);
8317 spin_unlock_irqrestore(&task_group_lock, flags);
8318 done:
8319 mutex_unlock(&shares_mutex);
8320 return 0;
8323 unsigned long sched_group_shares(struct task_group *tg)
8325 return tg->shares;
8327 #endif
8329 #ifdef CONFIG_RT_GROUP_SCHED
8331 * Ensure that the real time constraints are schedulable.
8333 static DEFINE_MUTEX(rt_constraints_mutex);
8335 static unsigned long to_ratio(u64 period, u64 runtime)
8337 if (runtime == RUNTIME_INF)
8338 return 1ULL << 16;
8340 return div64_u64(runtime << 16, period);
8343 #ifdef CONFIG_CGROUP_SCHED
8344 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8346 struct task_group *tgi, *parent = tg->parent;
8347 unsigned long total = 0;
8349 if (!parent) {
8350 if (global_rt_period() < period)
8351 return 0;
8353 return to_ratio(period, runtime) <
8354 to_ratio(global_rt_period(), global_rt_runtime());
8357 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8358 return 0;
8360 rcu_read_lock();
8361 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8362 if (tgi == tg)
8363 continue;
8365 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8366 tgi->rt_bandwidth.rt_runtime);
8368 rcu_read_unlock();
8370 return total + to_ratio(period, runtime) <
8371 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8372 parent->rt_bandwidth.rt_runtime);
8374 #elif defined CONFIG_USER_SCHED
8375 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8377 struct task_group *tgi;
8378 unsigned long total = 0;
8379 unsigned long global_ratio =
8380 to_ratio(global_rt_period(), global_rt_runtime());
8382 rcu_read_lock();
8383 list_for_each_entry_rcu(tgi, &task_groups, list) {
8384 if (tgi == tg)
8385 continue;
8387 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8388 tgi->rt_bandwidth.rt_runtime);
8390 rcu_read_unlock();
8392 return total + to_ratio(period, runtime) < global_ratio;
8394 #endif
8396 /* Must be called with tasklist_lock held */
8397 static inline int tg_has_rt_tasks(struct task_group *tg)
8399 struct task_struct *g, *p;
8400 do_each_thread(g, p) {
8401 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8402 return 1;
8403 } while_each_thread(g, p);
8404 return 0;
8407 static int tg_set_bandwidth(struct task_group *tg,
8408 u64 rt_period, u64 rt_runtime)
8410 int i, err = 0;
8412 mutex_lock(&rt_constraints_mutex);
8413 read_lock(&tasklist_lock);
8414 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8415 err = -EBUSY;
8416 goto unlock;
8418 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8419 err = -EINVAL;
8420 goto unlock;
8423 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8424 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8425 tg->rt_bandwidth.rt_runtime = rt_runtime;
8427 for_each_possible_cpu(i) {
8428 struct rt_rq *rt_rq = tg->rt_rq[i];
8430 spin_lock(&rt_rq->rt_runtime_lock);
8431 rt_rq->rt_runtime = rt_runtime;
8432 spin_unlock(&rt_rq->rt_runtime_lock);
8434 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8435 unlock:
8436 read_unlock(&tasklist_lock);
8437 mutex_unlock(&rt_constraints_mutex);
8439 return err;
8442 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8444 u64 rt_runtime, rt_period;
8446 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8447 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8448 if (rt_runtime_us < 0)
8449 rt_runtime = RUNTIME_INF;
8451 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8454 long sched_group_rt_runtime(struct task_group *tg)
8456 u64 rt_runtime_us;
8458 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8459 return -1;
8461 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8462 do_div(rt_runtime_us, NSEC_PER_USEC);
8463 return rt_runtime_us;
8466 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8468 u64 rt_runtime, rt_period;
8470 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8471 rt_runtime = tg->rt_bandwidth.rt_runtime;
8473 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8476 long sched_group_rt_period(struct task_group *tg)
8478 u64 rt_period_us;
8480 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8481 do_div(rt_period_us, NSEC_PER_USEC);
8482 return rt_period_us;
8485 static int sched_rt_global_constraints(void)
8487 int ret = 0;
8489 mutex_lock(&rt_constraints_mutex);
8490 if (!__rt_schedulable(NULL, 1, 0))
8491 ret = -EINVAL;
8492 mutex_unlock(&rt_constraints_mutex);
8494 return ret;
8496 #else
8497 static int sched_rt_global_constraints(void)
8499 unsigned long flags;
8500 int i;
8502 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8503 for_each_possible_cpu(i) {
8504 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8506 spin_lock(&rt_rq->rt_runtime_lock);
8507 rt_rq->rt_runtime = global_rt_runtime();
8508 spin_unlock(&rt_rq->rt_runtime_lock);
8510 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8512 return 0;
8514 #endif
8516 int sched_rt_handler(struct ctl_table *table, int write,
8517 struct file *filp, void __user *buffer, size_t *lenp,
8518 loff_t *ppos)
8520 int ret;
8521 int old_period, old_runtime;
8522 static DEFINE_MUTEX(mutex);
8524 mutex_lock(&mutex);
8525 old_period = sysctl_sched_rt_period;
8526 old_runtime = sysctl_sched_rt_runtime;
8528 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8530 if (!ret && write) {
8531 ret = sched_rt_global_constraints();
8532 if (ret) {
8533 sysctl_sched_rt_period = old_period;
8534 sysctl_sched_rt_runtime = old_runtime;
8535 } else {
8536 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8537 def_rt_bandwidth.rt_period =
8538 ns_to_ktime(global_rt_period());
8541 mutex_unlock(&mutex);
8543 return ret;
8546 #ifdef CONFIG_CGROUP_SCHED
8548 /* return corresponding task_group object of a cgroup */
8549 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8551 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8552 struct task_group, css);
8555 static struct cgroup_subsys_state *
8556 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8558 struct task_group *tg, *parent;
8560 if (!cgrp->parent) {
8561 /* This is early initialization for the top cgroup */
8562 init_task_group.css.cgroup = cgrp;
8563 return &init_task_group.css;
8566 parent = cgroup_tg(cgrp->parent);
8567 tg = sched_create_group(parent);
8568 if (IS_ERR(tg))
8569 return ERR_PTR(-ENOMEM);
8571 /* Bind the cgroup to task_group object we just created */
8572 tg->css.cgroup = cgrp;
8574 return &tg->css;
8577 static void
8578 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8580 struct task_group *tg = cgroup_tg(cgrp);
8582 sched_destroy_group(tg);
8585 static int
8586 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8587 struct task_struct *tsk)
8589 #ifdef CONFIG_RT_GROUP_SCHED
8590 /* Don't accept realtime tasks when there is no way for them to run */
8591 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8592 return -EINVAL;
8593 #else
8594 /* We don't support RT-tasks being in separate groups */
8595 if (tsk->sched_class != &fair_sched_class)
8596 return -EINVAL;
8597 #endif
8599 return 0;
8602 static void
8603 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8604 struct cgroup *old_cont, struct task_struct *tsk)
8606 sched_move_task(tsk);
8609 #ifdef CONFIG_FAIR_GROUP_SCHED
8610 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8611 u64 shareval)
8613 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8616 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8618 struct task_group *tg = cgroup_tg(cgrp);
8620 return (u64) tg->shares;
8622 #endif
8624 #ifdef CONFIG_RT_GROUP_SCHED
8625 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8626 s64 val)
8628 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8631 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8633 return sched_group_rt_runtime(cgroup_tg(cgrp));
8636 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8637 u64 rt_period_us)
8639 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8642 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8644 return sched_group_rt_period(cgroup_tg(cgrp));
8646 #endif
8648 static struct cftype cpu_files[] = {
8649 #ifdef CONFIG_FAIR_GROUP_SCHED
8651 .name = "shares",
8652 .read_u64 = cpu_shares_read_u64,
8653 .write_u64 = cpu_shares_write_u64,
8655 #endif
8656 #ifdef CONFIG_RT_GROUP_SCHED
8658 .name = "rt_runtime_us",
8659 .read_s64 = cpu_rt_runtime_read,
8660 .write_s64 = cpu_rt_runtime_write,
8663 .name = "rt_period_us",
8664 .read_u64 = cpu_rt_period_read_uint,
8665 .write_u64 = cpu_rt_period_write_uint,
8667 #endif
8670 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8672 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8675 struct cgroup_subsys cpu_cgroup_subsys = {
8676 .name = "cpu",
8677 .create = cpu_cgroup_create,
8678 .destroy = cpu_cgroup_destroy,
8679 .can_attach = cpu_cgroup_can_attach,
8680 .attach = cpu_cgroup_attach,
8681 .populate = cpu_cgroup_populate,
8682 .subsys_id = cpu_cgroup_subsys_id,
8683 .early_init = 1,
8686 #endif /* CONFIG_CGROUP_SCHED */
8688 #ifdef CONFIG_CGROUP_CPUACCT
8691 * CPU accounting code for task groups.
8693 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8694 * (balbir@in.ibm.com).
8697 /* track cpu usage of a group of tasks */
8698 struct cpuacct {
8699 struct cgroup_subsys_state css;
8700 /* cpuusage holds pointer to a u64-type object on every cpu */
8701 u64 *cpuusage;
8704 struct cgroup_subsys cpuacct_subsys;
8706 /* return cpu accounting group corresponding to this container */
8707 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8709 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8710 struct cpuacct, css);
8713 /* return cpu accounting group to which this task belongs */
8714 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8716 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8717 struct cpuacct, css);
8720 /* create a new cpu accounting group */
8721 static struct cgroup_subsys_state *cpuacct_create(
8722 struct cgroup_subsys *ss, struct cgroup *cgrp)
8724 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8726 if (!ca)
8727 return ERR_PTR(-ENOMEM);
8729 ca->cpuusage = alloc_percpu(u64);
8730 if (!ca->cpuusage) {
8731 kfree(ca);
8732 return ERR_PTR(-ENOMEM);
8735 return &ca->css;
8738 /* destroy an existing cpu accounting group */
8739 static void
8740 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8742 struct cpuacct *ca = cgroup_ca(cgrp);
8744 free_percpu(ca->cpuusage);
8745 kfree(ca);
8748 /* return total cpu usage (in nanoseconds) of a group */
8749 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8751 struct cpuacct *ca = cgroup_ca(cgrp);
8752 u64 totalcpuusage = 0;
8753 int i;
8755 for_each_possible_cpu(i) {
8756 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8759 * Take rq->lock to make 64-bit addition safe on 32-bit
8760 * platforms.
8762 spin_lock_irq(&cpu_rq(i)->lock);
8763 totalcpuusage += *cpuusage;
8764 spin_unlock_irq(&cpu_rq(i)->lock);
8767 return totalcpuusage;
8770 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8771 u64 reset)
8773 struct cpuacct *ca = cgroup_ca(cgrp);
8774 int err = 0;
8775 int i;
8777 if (reset) {
8778 err = -EINVAL;
8779 goto out;
8782 for_each_possible_cpu(i) {
8783 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8785 spin_lock_irq(&cpu_rq(i)->lock);
8786 *cpuusage = 0;
8787 spin_unlock_irq(&cpu_rq(i)->lock);
8789 out:
8790 return err;
8793 static struct cftype files[] = {
8795 .name = "usage",
8796 .read_u64 = cpuusage_read,
8797 .write_u64 = cpuusage_write,
8801 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8803 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8807 * charge this task's execution time to its accounting group.
8809 * called with rq->lock held.
8811 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8813 struct cpuacct *ca;
8815 if (!cpuacct_subsys.active)
8816 return;
8818 ca = task_ca(tsk);
8819 if (ca) {
8820 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8822 *cpuusage += cputime;
8826 struct cgroup_subsys cpuacct_subsys = {
8827 .name = "cpuacct",
8828 .create = cpuacct_create,
8829 .destroy = cpuacct_destroy,
8830 .populate = cpuacct_populate,
8831 .subsys_id = cpuacct_subsys_id,
8833 #endif /* CONFIG_CGROUP_CPUACCT */