thinkpad-acpi: basic ALSA mixer support (v2)
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / kernel / sched.c
blob7d13debbaf9c0d9ac7c56b1d09d7e76b15c481da
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/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
77 #include <asm/tlb.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
85 * and back.
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task);
122 DEFINE_TRACE(sched_wakeup);
123 DEFINE_TRACE(sched_wakeup_new);
124 DEFINE_TRACE(sched_switch);
125 DEFINE_TRACE(sched_migrate_task);
127 #ifdef CONFIG_SMP
129 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
132 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
133 * Since cpu_power is a 'constant', we can use a reciprocal divide.
135 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
137 return reciprocal_divide(load, sg->reciprocal_cpu_power);
141 * Each time a sched group cpu_power is changed,
142 * we must compute its reciprocal value
144 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
146 sg->__cpu_power += val;
147 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
149 #endif
151 static inline int rt_policy(int policy)
153 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
154 return 1;
155 return 0;
158 static inline int task_has_rt_policy(struct task_struct *p)
160 return rt_policy(p->policy);
164 * This is the priority-queue data structure of the RT scheduling class:
166 struct rt_prio_array {
167 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
168 struct list_head queue[MAX_RT_PRIO];
171 struct rt_bandwidth {
172 /* nests inside the rq lock: */
173 spinlock_t rt_runtime_lock;
174 ktime_t rt_period;
175 u64 rt_runtime;
176 struct hrtimer rt_period_timer;
179 static struct rt_bandwidth def_rt_bandwidth;
181 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
183 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
185 struct rt_bandwidth *rt_b =
186 container_of(timer, struct rt_bandwidth, rt_period_timer);
187 ktime_t now;
188 int overrun;
189 int idle = 0;
191 for (;;) {
192 now = hrtimer_cb_get_time(timer);
193 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
195 if (!overrun)
196 break;
198 idle = do_sched_rt_period_timer(rt_b, overrun);
201 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
204 static
205 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
207 rt_b->rt_period = ns_to_ktime(period);
208 rt_b->rt_runtime = runtime;
210 spin_lock_init(&rt_b->rt_runtime_lock);
212 hrtimer_init(&rt_b->rt_period_timer,
213 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
214 rt_b->rt_period_timer.function = sched_rt_period_timer;
217 static inline int rt_bandwidth_enabled(void)
219 return sysctl_sched_rt_runtime >= 0;
222 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
224 ktime_t now;
226 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
227 return;
229 if (hrtimer_active(&rt_b->rt_period_timer))
230 return;
232 spin_lock(&rt_b->rt_runtime_lock);
233 for (;;) {
234 unsigned long delta;
235 ktime_t soft, hard;
237 if (hrtimer_active(&rt_b->rt_period_timer))
238 break;
240 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
241 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
243 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
244 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
245 delta = ktime_to_ns(ktime_sub(hard, soft));
246 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
247 HRTIMER_MODE_ABS, 0);
249 spin_unlock(&rt_b->rt_runtime_lock);
252 #ifdef CONFIG_RT_GROUP_SCHED
253 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
255 hrtimer_cancel(&rt_b->rt_period_timer);
257 #endif
260 * sched_domains_mutex serializes calls to arch_init_sched_domains,
261 * detach_destroy_domains and partition_sched_domains.
263 static DEFINE_MUTEX(sched_domains_mutex);
265 #ifdef CONFIG_GROUP_SCHED
267 #include <linux/cgroup.h>
269 struct cfs_rq;
271 static LIST_HEAD(task_groups);
273 /* task group related information */
274 struct task_group {
275 #ifdef CONFIG_CGROUP_SCHED
276 struct cgroup_subsys_state css;
277 #endif
279 #ifdef CONFIG_USER_SCHED
280 uid_t uid;
281 #endif
283 #ifdef CONFIG_FAIR_GROUP_SCHED
284 /* schedulable entities of this group on each cpu */
285 struct sched_entity **se;
286 /* runqueue "owned" by this group on each cpu */
287 struct cfs_rq **cfs_rq;
288 unsigned long shares;
289 #endif
291 #ifdef CONFIG_RT_GROUP_SCHED
292 struct sched_rt_entity **rt_se;
293 struct rt_rq **rt_rq;
295 struct rt_bandwidth rt_bandwidth;
296 #endif
298 struct rcu_head rcu;
299 struct list_head list;
301 struct task_group *parent;
302 struct list_head siblings;
303 struct list_head children;
306 #ifdef CONFIG_USER_SCHED
308 /* Helper function to pass uid information to create_sched_user() */
309 void set_tg_uid(struct user_struct *user)
311 user->tg->uid = user->uid;
315 * Root task group.
316 * Every UID task group (including init_task_group aka UID-0) will
317 * be a child to this group.
319 struct task_group root_task_group;
321 #ifdef CONFIG_FAIR_GROUP_SCHED
322 /* Default task group's sched entity on each cpu */
323 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
324 /* Default task group's cfs_rq on each cpu */
325 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
326 #endif /* CONFIG_FAIR_GROUP_SCHED */
328 #ifdef CONFIG_RT_GROUP_SCHED
329 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
330 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
331 #endif /* CONFIG_RT_GROUP_SCHED */
332 #else /* !CONFIG_USER_SCHED */
333 #define root_task_group init_task_group
334 #endif /* CONFIG_USER_SCHED */
336 /* task_group_lock serializes add/remove of task groups and also changes to
337 * a task group's cpu shares.
339 static DEFINE_SPINLOCK(task_group_lock);
341 #ifdef CONFIG_FAIR_GROUP_SCHED
342 #ifdef CONFIG_USER_SCHED
343 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
344 #else /* !CONFIG_USER_SCHED */
345 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
346 #endif /* CONFIG_USER_SCHED */
349 * A weight of 0 or 1 can cause arithmetics problems.
350 * A weight of a cfs_rq is the sum of weights of which entities
351 * are queued on this cfs_rq, so a weight of a entity should not be
352 * too large, so as the shares value of a task group.
353 * (The default weight is 1024 - so there's no practical
354 * limitation from this.)
356 #define MIN_SHARES 2
357 #define MAX_SHARES (1UL << 18)
359 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
360 #endif
362 /* Default task group.
363 * Every task in system belong to this group at bootup.
365 struct task_group init_task_group;
367 /* return group to which a task belongs */
368 static inline struct task_group *task_group(struct task_struct *p)
370 struct task_group *tg;
372 #ifdef CONFIG_USER_SCHED
373 rcu_read_lock();
374 tg = __task_cred(p)->user->tg;
375 rcu_read_unlock();
376 #elif defined(CONFIG_CGROUP_SCHED)
377 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
378 struct task_group, css);
379 #else
380 tg = &init_task_group;
381 #endif
382 return tg;
385 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
386 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
388 #ifdef CONFIG_FAIR_GROUP_SCHED
389 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
390 p->se.parent = task_group(p)->se[cpu];
391 #endif
393 #ifdef CONFIG_RT_GROUP_SCHED
394 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
395 p->rt.parent = task_group(p)->rt_se[cpu];
396 #endif
399 #else
401 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
402 static inline struct task_group *task_group(struct task_struct *p)
404 return NULL;
407 #endif /* CONFIG_GROUP_SCHED */
409 /* CFS-related fields in a runqueue */
410 struct cfs_rq {
411 struct load_weight load;
412 unsigned long nr_running;
414 u64 exec_clock;
415 u64 min_vruntime;
417 struct rb_root tasks_timeline;
418 struct rb_node *rb_leftmost;
420 struct list_head tasks;
421 struct list_head *balance_iterator;
424 * 'curr' points to currently running entity on this cfs_rq.
425 * It is set to NULL otherwise (i.e when none are currently running).
427 struct sched_entity *curr, *next, *last;
429 unsigned int nr_spread_over;
431 #ifdef CONFIG_FAIR_GROUP_SCHED
432 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
435 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
436 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
437 * (like users, containers etc.)
439 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
440 * list is used during load balance.
442 struct list_head leaf_cfs_rq_list;
443 struct task_group *tg; /* group that "owns" this runqueue */
445 #ifdef CONFIG_SMP
447 * the part of load.weight contributed by tasks
449 unsigned long task_weight;
452 * h_load = weight * f(tg)
454 * Where f(tg) is the recursive weight fraction assigned to
455 * this group.
457 unsigned long h_load;
460 * this cpu's part of tg->shares
462 unsigned long shares;
465 * load.weight at the time we set shares
467 unsigned long rq_weight;
468 #endif
469 #endif
472 /* Real-Time classes' related field in a runqueue: */
473 struct rt_rq {
474 struct rt_prio_array active;
475 unsigned long rt_nr_running;
476 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
477 int highest_prio; /* highest queued rt task prio */
478 #endif
479 #ifdef CONFIG_SMP
480 unsigned long rt_nr_migratory;
481 int overloaded;
482 #endif
483 int rt_throttled;
484 u64 rt_time;
485 u64 rt_runtime;
486 /* Nests inside the rq lock: */
487 spinlock_t rt_runtime_lock;
489 #ifdef CONFIG_RT_GROUP_SCHED
490 unsigned long rt_nr_boosted;
492 struct rq *rq;
493 struct list_head leaf_rt_rq_list;
494 struct task_group *tg;
495 struct sched_rt_entity *rt_se;
496 #endif
499 #ifdef CONFIG_SMP
502 * We add the notion of a root-domain which will be used to define per-domain
503 * variables. Each exclusive cpuset essentially defines an island domain by
504 * fully partitioning the member cpus from any other cpuset. Whenever a new
505 * exclusive cpuset is created, we also create and attach a new root-domain
506 * object.
509 struct root_domain {
510 atomic_t refcount;
511 cpumask_var_t span;
512 cpumask_var_t online;
515 * The "RT overload" flag: it gets set if a CPU has more than
516 * one runnable RT task.
518 cpumask_var_t rto_mask;
519 atomic_t rto_count;
520 #ifdef CONFIG_SMP
521 struct cpupri cpupri;
522 #endif
523 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
525 * Preferred wake up cpu nominated by sched_mc balance that will be
526 * used when most cpus are idle in the system indicating overall very
527 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
529 unsigned int sched_mc_preferred_wakeup_cpu;
530 #endif
534 * By default the system creates a single root-domain with all cpus as
535 * members (mimicking the global state we have today).
537 static struct root_domain def_root_domain;
539 #endif
542 * This is the main, per-CPU runqueue data structure.
544 * Locking rule: those places that want to lock multiple runqueues
545 * (such as the load balancing or the thread migration code), lock
546 * acquire operations must be ordered by ascending &runqueue.
548 struct rq {
549 /* runqueue lock: */
550 spinlock_t lock;
553 * nr_running and cpu_load should be in the same cacheline because
554 * remote CPUs use both these fields when doing load calculation.
556 unsigned long nr_running;
557 #define CPU_LOAD_IDX_MAX 5
558 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
559 unsigned char idle_at_tick;
560 #ifdef CONFIG_NO_HZ
561 unsigned long last_tick_seen;
562 unsigned char in_nohz_recently;
563 #endif
564 /* capture load from *all* tasks on this cpu: */
565 struct load_weight load;
566 unsigned long nr_load_updates;
567 u64 nr_switches;
569 struct cfs_rq cfs;
570 struct rt_rq rt;
572 #ifdef CONFIG_FAIR_GROUP_SCHED
573 /* list of leaf cfs_rq on this cpu: */
574 struct list_head leaf_cfs_rq_list;
575 #endif
576 #ifdef CONFIG_RT_GROUP_SCHED
577 struct list_head leaf_rt_rq_list;
578 #endif
581 * This is part of a global counter where only the total sum
582 * over all CPUs matters. A task can increase this counter on
583 * one CPU and if it got migrated afterwards it may decrease
584 * it on another CPU. Always updated under the runqueue lock:
586 unsigned long nr_uninterruptible;
588 struct task_struct *curr, *idle;
589 unsigned long next_balance;
590 struct mm_struct *prev_mm;
592 u64 clock;
594 atomic_t nr_iowait;
596 #ifdef CONFIG_SMP
597 struct root_domain *rd;
598 struct sched_domain *sd;
600 /* For active balancing */
601 int active_balance;
602 int push_cpu;
603 /* cpu of this runqueue: */
604 int cpu;
605 int online;
607 unsigned long avg_load_per_task;
609 struct task_struct *migration_thread;
610 struct list_head migration_queue;
611 #endif
613 #ifdef CONFIG_SCHED_HRTICK
614 #ifdef CONFIG_SMP
615 int hrtick_csd_pending;
616 struct call_single_data hrtick_csd;
617 #endif
618 struct hrtimer hrtick_timer;
619 #endif
621 #ifdef CONFIG_SCHEDSTATS
622 /* latency stats */
623 struct sched_info rq_sched_info;
624 unsigned long long rq_cpu_time;
625 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
627 /* sys_sched_yield() stats */
628 unsigned int yld_exp_empty;
629 unsigned int yld_act_empty;
630 unsigned int yld_both_empty;
631 unsigned int yld_count;
633 /* schedule() stats */
634 unsigned int sched_switch;
635 unsigned int sched_count;
636 unsigned int sched_goidle;
638 /* try_to_wake_up() stats */
639 unsigned int ttwu_count;
640 unsigned int ttwu_local;
642 /* BKL stats */
643 unsigned int bkl_count;
644 #endif
647 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
649 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
651 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
654 static inline int cpu_of(struct rq *rq)
656 #ifdef CONFIG_SMP
657 return rq->cpu;
658 #else
659 return 0;
660 #endif
664 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
665 * See detach_destroy_domains: synchronize_sched for details.
667 * The domain tree of any CPU may only be accessed from within
668 * preempt-disabled sections.
670 #define for_each_domain(cpu, __sd) \
671 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
673 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
674 #define this_rq() (&__get_cpu_var(runqueues))
675 #define task_rq(p) cpu_rq(task_cpu(p))
676 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
678 static inline void update_rq_clock(struct rq *rq)
680 rq->clock = sched_clock_cpu(cpu_of(rq));
684 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
686 #ifdef CONFIG_SCHED_DEBUG
687 # define const_debug __read_mostly
688 #else
689 # define const_debug static const
690 #endif
693 * runqueue_is_locked
695 * Returns true if the current cpu runqueue is locked.
696 * This interface allows printk to be called with the runqueue lock
697 * held and know whether or not it is OK to wake up the klogd.
699 int runqueue_is_locked(void)
701 int cpu = get_cpu();
702 struct rq *rq = cpu_rq(cpu);
703 int ret;
705 ret = spin_is_locked(&rq->lock);
706 put_cpu();
707 return ret;
711 * Debugging: various feature bits
714 #define SCHED_FEAT(name, enabled) \
715 __SCHED_FEAT_##name ,
717 enum {
718 #include "sched_features.h"
721 #undef SCHED_FEAT
723 #define SCHED_FEAT(name, enabled) \
724 (1UL << __SCHED_FEAT_##name) * enabled |
726 const_debug unsigned int sysctl_sched_features =
727 #include "sched_features.h"
730 #undef SCHED_FEAT
732 #ifdef CONFIG_SCHED_DEBUG
733 #define SCHED_FEAT(name, enabled) \
734 #name ,
736 static __read_mostly char *sched_feat_names[] = {
737 #include "sched_features.h"
738 NULL
741 #undef SCHED_FEAT
743 static int sched_feat_show(struct seq_file *m, void *v)
745 int i;
747 for (i = 0; sched_feat_names[i]; i++) {
748 if (!(sysctl_sched_features & (1UL << i)))
749 seq_puts(m, "NO_");
750 seq_printf(m, "%s ", sched_feat_names[i]);
752 seq_puts(m, "\n");
754 return 0;
757 static ssize_t
758 sched_feat_write(struct file *filp, const char __user *ubuf,
759 size_t cnt, loff_t *ppos)
761 char buf[64];
762 char *cmp = buf;
763 int neg = 0;
764 int i;
766 if (cnt > 63)
767 cnt = 63;
769 if (copy_from_user(&buf, ubuf, cnt))
770 return -EFAULT;
772 buf[cnt] = 0;
774 if (strncmp(buf, "NO_", 3) == 0) {
775 neg = 1;
776 cmp += 3;
779 for (i = 0; sched_feat_names[i]; i++) {
780 int len = strlen(sched_feat_names[i]);
782 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
783 if (neg)
784 sysctl_sched_features &= ~(1UL << i);
785 else
786 sysctl_sched_features |= (1UL << i);
787 break;
791 if (!sched_feat_names[i])
792 return -EINVAL;
794 filp->f_pos += cnt;
796 return cnt;
799 static int sched_feat_open(struct inode *inode, struct file *filp)
801 return single_open(filp, sched_feat_show, NULL);
804 static struct file_operations sched_feat_fops = {
805 .open = sched_feat_open,
806 .write = sched_feat_write,
807 .read = seq_read,
808 .llseek = seq_lseek,
809 .release = single_release,
812 static __init int sched_init_debug(void)
814 debugfs_create_file("sched_features", 0644, NULL, NULL,
815 &sched_feat_fops);
817 return 0;
819 late_initcall(sched_init_debug);
821 #endif
823 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
826 * Number of tasks to iterate in a single balance run.
827 * Limited because this is done with IRQs disabled.
829 const_debug unsigned int sysctl_sched_nr_migrate = 32;
832 * ratelimit for updating the group shares.
833 * default: 0.25ms
835 unsigned int sysctl_sched_shares_ratelimit = 250000;
838 * Inject some fuzzyness into changing the per-cpu group shares
839 * this avoids remote rq-locks at the expense of fairness.
840 * default: 4
842 unsigned int sysctl_sched_shares_thresh = 4;
845 * period over which we measure -rt task cpu usage in us.
846 * default: 1s
848 unsigned int sysctl_sched_rt_period = 1000000;
850 static __read_mostly int scheduler_running;
853 * part of the period that we allow rt tasks to run in us.
854 * default: 0.95s
856 int sysctl_sched_rt_runtime = 950000;
858 static inline u64 global_rt_period(void)
860 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
863 static inline u64 global_rt_runtime(void)
865 if (sysctl_sched_rt_runtime < 0)
866 return RUNTIME_INF;
868 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
871 #ifndef prepare_arch_switch
872 # define prepare_arch_switch(next) do { } while (0)
873 #endif
874 #ifndef finish_arch_switch
875 # define finish_arch_switch(prev) do { } while (0)
876 #endif
878 static inline int task_current(struct rq *rq, struct task_struct *p)
880 return rq->curr == p;
883 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
884 static inline int task_running(struct rq *rq, struct task_struct *p)
886 return task_current(rq, p);
889 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
893 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
895 #ifdef CONFIG_DEBUG_SPINLOCK
896 /* this is a valid case when another task releases the spinlock */
897 rq->lock.owner = current;
898 #endif
900 * If we are tracking spinlock dependencies then we have to
901 * fix up the runqueue lock - which gets 'carried over' from
902 * prev into current:
904 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
906 spin_unlock_irq(&rq->lock);
909 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
910 static inline int task_running(struct rq *rq, struct task_struct *p)
912 #ifdef CONFIG_SMP
913 return p->oncpu;
914 #else
915 return task_current(rq, p);
916 #endif
919 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
921 #ifdef CONFIG_SMP
923 * We can optimise this out completely for !SMP, because the
924 * SMP rebalancing from interrupt is the only thing that cares
925 * here.
927 next->oncpu = 1;
928 #endif
929 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
930 spin_unlock_irq(&rq->lock);
931 #else
932 spin_unlock(&rq->lock);
933 #endif
936 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
938 #ifdef CONFIG_SMP
940 * After ->oncpu is cleared, the task can be moved to a different CPU.
941 * We must ensure this doesn't happen until the switch is completely
942 * finished.
944 smp_wmb();
945 prev->oncpu = 0;
946 #endif
947 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
948 local_irq_enable();
949 #endif
951 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
954 * __task_rq_lock - lock the runqueue a given task resides on.
955 * Must be called interrupts disabled.
957 static inline struct rq *__task_rq_lock(struct task_struct *p)
958 __acquires(rq->lock)
960 for (;;) {
961 struct rq *rq = task_rq(p);
962 spin_lock(&rq->lock);
963 if (likely(rq == task_rq(p)))
964 return rq;
965 spin_unlock(&rq->lock);
970 * task_rq_lock - lock the runqueue a given task resides on and disable
971 * interrupts. Note the ordering: we can safely lookup the task_rq without
972 * explicitly disabling preemption.
974 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
975 __acquires(rq->lock)
977 struct rq *rq;
979 for (;;) {
980 local_irq_save(*flags);
981 rq = task_rq(p);
982 spin_lock(&rq->lock);
983 if (likely(rq == task_rq(p)))
984 return rq;
985 spin_unlock_irqrestore(&rq->lock, *flags);
989 void task_rq_unlock_wait(struct task_struct *p)
991 struct rq *rq = task_rq(p);
993 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
994 spin_unlock_wait(&rq->lock);
997 static void __task_rq_unlock(struct rq *rq)
998 __releases(rq->lock)
1000 spin_unlock(&rq->lock);
1003 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1004 __releases(rq->lock)
1006 spin_unlock_irqrestore(&rq->lock, *flags);
1010 * this_rq_lock - lock this runqueue and disable interrupts.
1012 static struct rq *this_rq_lock(void)
1013 __acquires(rq->lock)
1015 struct rq *rq;
1017 local_irq_disable();
1018 rq = this_rq();
1019 spin_lock(&rq->lock);
1021 return rq;
1024 #ifdef CONFIG_SCHED_HRTICK
1026 * Use HR-timers to deliver accurate preemption points.
1028 * Its all a bit involved since we cannot program an hrt while holding the
1029 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1030 * reschedule event.
1032 * When we get rescheduled we reprogram the hrtick_timer outside of the
1033 * rq->lock.
1037 * Use hrtick when:
1038 * - enabled by features
1039 * - hrtimer is actually high res
1041 static inline int hrtick_enabled(struct rq *rq)
1043 if (!sched_feat(HRTICK))
1044 return 0;
1045 if (!cpu_active(cpu_of(rq)))
1046 return 0;
1047 return hrtimer_is_hres_active(&rq->hrtick_timer);
1050 static void hrtick_clear(struct rq *rq)
1052 if (hrtimer_active(&rq->hrtick_timer))
1053 hrtimer_cancel(&rq->hrtick_timer);
1057 * High-resolution timer tick.
1058 * Runs from hardirq context with interrupts disabled.
1060 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1062 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1064 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1066 spin_lock(&rq->lock);
1067 update_rq_clock(rq);
1068 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1069 spin_unlock(&rq->lock);
1071 return HRTIMER_NORESTART;
1074 #ifdef CONFIG_SMP
1076 * called from hardirq (IPI) context
1078 static void __hrtick_start(void *arg)
1080 struct rq *rq = arg;
1082 spin_lock(&rq->lock);
1083 hrtimer_restart(&rq->hrtick_timer);
1084 rq->hrtick_csd_pending = 0;
1085 spin_unlock(&rq->lock);
1089 * Called to set the hrtick timer state.
1091 * called with rq->lock held and irqs disabled
1093 static void hrtick_start(struct rq *rq, u64 delay)
1095 struct hrtimer *timer = &rq->hrtick_timer;
1096 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1098 hrtimer_set_expires(timer, time);
1100 if (rq == this_rq()) {
1101 hrtimer_restart(timer);
1102 } else if (!rq->hrtick_csd_pending) {
1103 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1104 rq->hrtick_csd_pending = 1;
1108 static int
1109 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1111 int cpu = (int)(long)hcpu;
1113 switch (action) {
1114 case CPU_UP_CANCELED:
1115 case CPU_UP_CANCELED_FROZEN:
1116 case CPU_DOWN_PREPARE:
1117 case CPU_DOWN_PREPARE_FROZEN:
1118 case CPU_DEAD:
1119 case CPU_DEAD_FROZEN:
1120 hrtick_clear(cpu_rq(cpu));
1121 return NOTIFY_OK;
1124 return NOTIFY_DONE;
1127 static __init void init_hrtick(void)
1129 hotcpu_notifier(hotplug_hrtick, 0);
1131 #else
1133 * Called to set the hrtick timer state.
1135 * called with rq->lock held and irqs disabled
1137 static void hrtick_start(struct rq *rq, u64 delay)
1139 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1140 HRTIMER_MODE_REL, 0);
1143 static inline void init_hrtick(void)
1146 #endif /* CONFIG_SMP */
1148 static void init_rq_hrtick(struct rq *rq)
1150 #ifdef CONFIG_SMP
1151 rq->hrtick_csd_pending = 0;
1153 rq->hrtick_csd.flags = 0;
1154 rq->hrtick_csd.func = __hrtick_start;
1155 rq->hrtick_csd.info = rq;
1156 #endif
1158 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1159 rq->hrtick_timer.function = hrtick;
1161 #else /* CONFIG_SCHED_HRTICK */
1162 static inline void hrtick_clear(struct rq *rq)
1166 static inline void init_rq_hrtick(struct rq *rq)
1170 static inline void init_hrtick(void)
1173 #endif /* CONFIG_SCHED_HRTICK */
1176 * resched_task - mark a task 'to be rescheduled now'.
1178 * On UP this means the setting of the need_resched flag, on SMP it
1179 * might also involve a cross-CPU call to trigger the scheduler on
1180 * the target CPU.
1182 #ifdef CONFIG_SMP
1184 #ifndef tsk_is_polling
1185 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1186 #endif
1188 static void resched_task(struct task_struct *p)
1190 int cpu;
1192 assert_spin_locked(&task_rq(p)->lock);
1194 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1195 return;
1197 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1199 cpu = task_cpu(p);
1200 if (cpu == smp_processor_id())
1201 return;
1203 /* NEED_RESCHED must be visible before we test polling */
1204 smp_mb();
1205 if (!tsk_is_polling(p))
1206 smp_send_reschedule(cpu);
1209 static void resched_cpu(int cpu)
1211 struct rq *rq = cpu_rq(cpu);
1212 unsigned long flags;
1214 if (!spin_trylock_irqsave(&rq->lock, flags))
1215 return;
1216 resched_task(cpu_curr(cpu));
1217 spin_unlock_irqrestore(&rq->lock, flags);
1220 #ifdef CONFIG_NO_HZ
1222 * When add_timer_on() enqueues a timer into the timer wheel of an
1223 * idle CPU then this timer might expire before the next timer event
1224 * which is scheduled to wake up that CPU. In case of a completely
1225 * idle system the next event might even be infinite time into the
1226 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1227 * leaves the inner idle loop so the newly added timer is taken into
1228 * account when the CPU goes back to idle and evaluates the timer
1229 * wheel for the next timer event.
1231 void wake_up_idle_cpu(int cpu)
1233 struct rq *rq = cpu_rq(cpu);
1235 if (cpu == smp_processor_id())
1236 return;
1239 * This is safe, as this function is called with the timer
1240 * wheel base lock of (cpu) held. When the CPU is on the way
1241 * to idle and has not yet set rq->curr to idle then it will
1242 * be serialized on the timer wheel base lock and take the new
1243 * timer into account automatically.
1245 if (rq->curr != rq->idle)
1246 return;
1249 * We can set TIF_RESCHED on the idle task of the other CPU
1250 * lockless. The worst case is that the other CPU runs the
1251 * idle task through an additional NOOP schedule()
1253 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1255 /* NEED_RESCHED must be visible before we test polling */
1256 smp_mb();
1257 if (!tsk_is_polling(rq->idle))
1258 smp_send_reschedule(cpu);
1260 #endif /* CONFIG_NO_HZ */
1262 #else /* !CONFIG_SMP */
1263 static void resched_task(struct task_struct *p)
1265 assert_spin_locked(&task_rq(p)->lock);
1266 set_tsk_need_resched(p);
1268 #endif /* CONFIG_SMP */
1270 #if BITS_PER_LONG == 32
1271 # define WMULT_CONST (~0UL)
1272 #else
1273 # define WMULT_CONST (1UL << 32)
1274 #endif
1276 #define WMULT_SHIFT 32
1279 * Shift right and round:
1281 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1284 * delta *= weight / lw
1286 static unsigned long
1287 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1288 struct load_weight *lw)
1290 u64 tmp;
1292 if (!lw->inv_weight) {
1293 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1294 lw->inv_weight = 1;
1295 else
1296 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1297 / (lw->weight+1);
1300 tmp = (u64)delta_exec * weight;
1302 * Check whether we'd overflow the 64-bit multiplication:
1304 if (unlikely(tmp > WMULT_CONST))
1305 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1306 WMULT_SHIFT/2);
1307 else
1308 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1310 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1313 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1315 lw->weight += inc;
1316 lw->inv_weight = 0;
1319 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1321 lw->weight -= dec;
1322 lw->inv_weight = 0;
1326 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1327 * of tasks with abnormal "nice" values across CPUs the contribution that
1328 * each task makes to its run queue's load is weighted according to its
1329 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1330 * scaled version of the new time slice allocation that they receive on time
1331 * slice expiry etc.
1334 #define WEIGHT_IDLEPRIO 3
1335 #define WMULT_IDLEPRIO 1431655765
1338 * Nice levels are multiplicative, with a gentle 10% change for every
1339 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1340 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1341 * that remained on nice 0.
1343 * The "10% effect" is relative and cumulative: from _any_ nice level,
1344 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1345 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1346 * If a task goes up by ~10% and another task goes down by ~10% then
1347 * the relative distance between them is ~25%.)
1349 static const int prio_to_weight[40] = {
1350 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1351 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1352 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1353 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1354 /* 0 */ 1024, 820, 655, 526, 423,
1355 /* 5 */ 335, 272, 215, 172, 137,
1356 /* 10 */ 110, 87, 70, 56, 45,
1357 /* 15 */ 36, 29, 23, 18, 15,
1361 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1363 * In cases where the weight does not change often, we can use the
1364 * precalculated inverse to speed up arithmetics by turning divisions
1365 * into multiplications:
1367 static const u32 prio_to_wmult[40] = {
1368 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1369 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1370 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1371 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1372 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1373 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1374 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1375 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1378 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1381 * runqueue iterator, to support SMP load-balancing between different
1382 * scheduling classes, without having to expose their internal data
1383 * structures to the load-balancing proper:
1385 struct rq_iterator {
1386 void *arg;
1387 struct task_struct *(*start)(void *);
1388 struct task_struct *(*next)(void *);
1391 #ifdef CONFIG_SMP
1392 static unsigned long
1393 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1394 unsigned long max_load_move, struct sched_domain *sd,
1395 enum cpu_idle_type idle, int *all_pinned,
1396 int *this_best_prio, struct rq_iterator *iterator);
1398 static int
1399 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1400 struct sched_domain *sd, enum cpu_idle_type idle,
1401 struct rq_iterator *iterator);
1402 #endif
1404 #ifdef CONFIG_CGROUP_CPUACCT
1405 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1406 #else
1407 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1408 #endif
1410 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1412 update_load_add(&rq->load, load);
1415 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1417 update_load_sub(&rq->load, load);
1420 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1421 typedef int (*tg_visitor)(struct task_group *, void *);
1424 * Iterate the full tree, calling @down when first entering a node and @up when
1425 * leaving it for the final time.
1427 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1429 struct task_group *parent, *child;
1430 int ret;
1432 rcu_read_lock();
1433 parent = &root_task_group;
1434 down:
1435 ret = (*down)(parent, data);
1436 if (ret)
1437 goto out_unlock;
1438 list_for_each_entry_rcu(child, &parent->children, siblings) {
1439 parent = child;
1440 goto down;
1443 continue;
1445 ret = (*up)(parent, data);
1446 if (ret)
1447 goto out_unlock;
1449 child = parent;
1450 parent = parent->parent;
1451 if (parent)
1452 goto up;
1453 out_unlock:
1454 rcu_read_unlock();
1456 return ret;
1459 static int tg_nop(struct task_group *tg, void *data)
1461 return 0;
1463 #endif
1465 #ifdef CONFIG_SMP
1466 static unsigned long source_load(int cpu, int type);
1467 static unsigned long target_load(int cpu, int type);
1468 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1470 static unsigned long cpu_avg_load_per_task(int cpu)
1472 struct rq *rq = cpu_rq(cpu);
1473 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1475 if (nr_running)
1476 rq->avg_load_per_task = rq->load.weight / nr_running;
1477 else
1478 rq->avg_load_per_task = 0;
1480 return rq->avg_load_per_task;
1483 #ifdef CONFIG_FAIR_GROUP_SCHED
1485 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1488 * Calculate and set the cpu's group shares.
1490 static void
1491 update_group_shares_cpu(struct task_group *tg, int cpu,
1492 unsigned long sd_shares, unsigned long sd_rq_weight)
1494 unsigned long shares;
1495 unsigned long rq_weight;
1497 if (!tg->se[cpu])
1498 return;
1500 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1503 * \Sum shares * rq_weight
1504 * shares = -----------------------
1505 * \Sum rq_weight
1508 shares = (sd_shares * rq_weight) / sd_rq_weight;
1509 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1511 if (abs(shares - tg->se[cpu]->load.weight) >
1512 sysctl_sched_shares_thresh) {
1513 struct rq *rq = cpu_rq(cpu);
1514 unsigned long flags;
1516 spin_lock_irqsave(&rq->lock, flags);
1517 tg->cfs_rq[cpu]->shares = shares;
1519 __set_se_shares(tg->se[cpu], shares);
1520 spin_unlock_irqrestore(&rq->lock, flags);
1525 * Re-compute the task group their per cpu shares over the given domain.
1526 * This needs to be done in a bottom-up fashion because the rq weight of a
1527 * parent group depends on the shares of its child groups.
1529 static int tg_shares_up(struct task_group *tg, void *data)
1531 unsigned long weight, rq_weight = 0;
1532 unsigned long shares = 0;
1533 struct sched_domain *sd = data;
1534 int i;
1536 for_each_cpu(i, sched_domain_span(sd)) {
1538 * If there are currently no tasks on the cpu pretend there
1539 * is one of average load so that when a new task gets to
1540 * run here it will not get delayed by group starvation.
1542 weight = tg->cfs_rq[i]->load.weight;
1543 if (!weight)
1544 weight = NICE_0_LOAD;
1546 tg->cfs_rq[i]->rq_weight = weight;
1547 rq_weight += weight;
1548 shares += tg->cfs_rq[i]->shares;
1551 if ((!shares && rq_weight) || shares > tg->shares)
1552 shares = tg->shares;
1554 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1555 shares = tg->shares;
1557 for_each_cpu(i, sched_domain_span(sd))
1558 update_group_shares_cpu(tg, i, shares, rq_weight);
1560 return 0;
1564 * Compute the cpu's hierarchical load factor for each task group.
1565 * This needs to be done in a top-down fashion because the load of a child
1566 * group is a fraction of its parents load.
1568 static int tg_load_down(struct task_group *tg, void *data)
1570 unsigned long load;
1571 long cpu = (long)data;
1573 if (!tg->parent) {
1574 load = cpu_rq(cpu)->load.weight;
1575 } else {
1576 load = tg->parent->cfs_rq[cpu]->h_load;
1577 load *= tg->cfs_rq[cpu]->shares;
1578 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1581 tg->cfs_rq[cpu]->h_load = load;
1583 return 0;
1586 static void update_shares(struct sched_domain *sd)
1588 u64 now = cpu_clock(raw_smp_processor_id());
1589 s64 elapsed = now - sd->last_update;
1591 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1592 sd->last_update = now;
1593 walk_tg_tree(tg_nop, tg_shares_up, sd);
1597 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1599 spin_unlock(&rq->lock);
1600 update_shares(sd);
1601 spin_lock(&rq->lock);
1604 static void update_h_load(long cpu)
1606 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1609 #else
1611 static inline void update_shares(struct sched_domain *sd)
1615 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1619 #endif
1622 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1624 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1625 __releases(this_rq->lock)
1626 __acquires(busiest->lock)
1627 __acquires(this_rq->lock)
1629 int ret = 0;
1631 if (unlikely(!irqs_disabled())) {
1632 /* printk() doesn't work good under rq->lock */
1633 spin_unlock(&this_rq->lock);
1634 BUG_ON(1);
1636 if (unlikely(!spin_trylock(&busiest->lock))) {
1637 if (busiest < this_rq) {
1638 spin_unlock(&this_rq->lock);
1639 spin_lock(&busiest->lock);
1640 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1641 ret = 1;
1642 } else
1643 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1645 return ret;
1648 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1649 __releases(busiest->lock)
1651 spin_unlock(&busiest->lock);
1652 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1654 #endif
1656 #ifdef CONFIG_FAIR_GROUP_SCHED
1657 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1659 #ifdef CONFIG_SMP
1660 cfs_rq->shares = shares;
1661 #endif
1663 #endif
1665 #include "sched_stats.h"
1666 #include "sched_idletask.c"
1667 #include "sched_fair.c"
1668 #include "sched_rt.c"
1669 #ifdef CONFIG_SCHED_DEBUG
1670 # include "sched_debug.c"
1671 #endif
1673 #define sched_class_highest (&rt_sched_class)
1674 #define for_each_class(class) \
1675 for (class = sched_class_highest; class; class = class->next)
1677 static void inc_nr_running(struct rq *rq)
1679 rq->nr_running++;
1682 static void dec_nr_running(struct rq *rq)
1684 rq->nr_running--;
1687 static void set_load_weight(struct task_struct *p)
1689 if (task_has_rt_policy(p)) {
1690 p->se.load.weight = prio_to_weight[0] * 2;
1691 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1692 return;
1696 * SCHED_IDLE tasks get minimal weight:
1698 if (p->policy == SCHED_IDLE) {
1699 p->se.load.weight = WEIGHT_IDLEPRIO;
1700 p->se.load.inv_weight = WMULT_IDLEPRIO;
1701 return;
1704 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1705 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1708 static void update_avg(u64 *avg, u64 sample)
1710 s64 diff = sample - *avg;
1711 *avg += diff >> 3;
1714 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1716 sched_info_queued(p);
1717 p->sched_class->enqueue_task(rq, p, wakeup);
1718 p->se.on_rq = 1;
1721 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1723 if (sleep && p->se.last_wakeup) {
1724 update_avg(&p->se.avg_overlap,
1725 p->se.sum_exec_runtime - p->se.last_wakeup);
1726 p->se.last_wakeup = 0;
1729 sched_info_dequeued(p);
1730 p->sched_class->dequeue_task(rq, p, sleep);
1731 p->se.on_rq = 0;
1735 * __normal_prio - return the priority that is based on the static prio
1737 static inline int __normal_prio(struct task_struct *p)
1739 return p->static_prio;
1743 * Calculate the expected normal priority: i.e. priority
1744 * without taking RT-inheritance into account. Might be
1745 * boosted by interactivity modifiers. Changes upon fork,
1746 * setprio syscalls, and whenever the interactivity
1747 * estimator recalculates.
1749 static inline int normal_prio(struct task_struct *p)
1751 int prio;
1753 if (task_has_rt_policy(p))
1754 prio = MAX_RT_PRIO-1 - p->rt_priority;
1755 else
1756 prio = __normal_prio(p);
1757 return prio;
1761 * Calculate the current priority, i.e. the priority
1762 * taken into account by the scheduler. This value might
1763 * be boosted by RT tasks, or might be boosted by
1764 * interactivity modifiers. Will be RT if the task got
1765 * RT-boosted. If not then it returns p->normal_prio.
1767 static int effective_prio(struct task_struct *p)
1769 p->normal_prio = normal_prio(p);
1771 * If we are RT tasks or we were boosted to RT priority,
1772 * keep the priority unchanged. Otherwise, update priority
1773 * to the normal priority:
1775 if (!rt_prio(p->prio))
1776 return p->normal_prio;
1777 return p->prio;
1781 * activate_task - move a task to the runqueue.
1783 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1785 if (task_contributes_to_load(p))
1786 rq->nr_uninterruptible--;
1788 enqueue_task(rq, p, wakeup);
1789 inc_nr_running(rq);
1793 * deactivate_task - remove a task from the runqueue.
1795 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1797 if (task_contributes_to_load(p))
1798 rq->nr_uninterruptible++;
1800 dequeue_task(rq, p, sleep);
1801 dec_nr_running(rq);
1805 * task_curr - is this task currently executing on a CPU?
1806 * @p: the task in question.
1808 inline int task_curr(const struct task_struct *p)
1810 return cpu_curr(task_cpu(p)) == p;
1813 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1815 set_task_rq(p, cpu);
1816 #ifdef CONFIG_SMP
1818 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1819 * successfuly executed on another CPU. We must ensure that updates of
1820 * per-task data have been completed by this moment.
1822 smp_wmb();
1823 task_thread_info(p)->cpu = cpu;
1824 #endif
1827 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1828 const struct sched_class *prev_class,
1829 int oldprio, int running)
1831 if (prev_class != p->sched_class) {
1832 if (prev_class->switched_from)
1833 prev_class->switched_from(rq, p, running);
1834 p->sched_class->switched_to(rq, p, running);
1835 } else
1836 p->sched_class->prio_changed(rq, p, oldprio, running);
1839 #ifdef CONFIG_SMP
1841 /* Used instead of source_load when we know the type == 0 */
1842 static unsigned long weighted_cpuload(const int cpu)
1844 return cpu_rq(cpu)->load.weight;
1848 * Is this task likely cache-hot:
1850 static int
1851 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1853 s64 delta;
1856 * Buddy candidates are cache hot:
1858 if (sched_feat(CACHE_HOT_BUDDY) &&
1859 (&p->se == cfs_rq_of(&p->se)->next ||
1860 &p->se == cfs_rq_of(&p->se)->last))
1861 return 1;
1863 if (p->sched_class != &fair_sched_class)
1864 return 0;
1866 if (sysctl_sched_migration_cost == -1)
1867 return 1;
1868 if (sysctl_sched_migration_cost == 0)
1869 return 0;
1871 delta = now - p->se.exec_start;
1873 return delta < (s64)sysctl_sched_migration_cost;
1877 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1879 int old_cpu = task_cpu(p);
1880 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1881 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1882 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1883 u64 clock_offset;
1885 clock_offset = old_rq->clock - new_rq->clock;
1887 trace_sched_migrate_task(p, task_cpu(p), new_cpu);
1889 #ifdef CONFIG_SCHEDSTATS
1890 if (p->se.wait_start)
1891 p->se.wait_start -= clock_offset;
1892 if (p->se.sleep_start)
1893 p->se.sleep_start -= clock_offset;
1894 if (p->se.block_start)
1895 p->se.block_start -= clock_offset;
1896 if (old_cpu != new_cpu) {
1897 schedstat_inc(p, se.nr_migrations);
1898 if (task_hot(p, old_rq->clock, NULL))
1899 schedstat_inc(p, se.nr_forced2_migrations);
1901 #endif
1902 p->se.vruntime -= old_cfsrq->min_vruntime -
1903 new_cfsrq->min_vruntime;
1905 __set_task_cpu(p, new_cpu);
1908 struct migration_req {
1909 struct list_head list;
1911 struct task_struct *task;
1912 int dest_cpu;
1914 struct completion done;
1918 * The task's runqueue lock must be held.
1919 * Returns true if you have to wait for migration thread.
1921 static int
1922 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1924 struct rq *rq = task_rq(p);
1927 * If the task is not on a runqueue (and not running), then
1928 * it is sufficient to simply update the task's cpu field.
1930 if (!p->se.on_rq && !task_running(rq, p)) {
1931 set_task_cpu(p, dest_cpu);
1932 return 0;
1935 init_completion(&req->done);
1936 req->task = p;
1937 req->dest_cpu = dest_cpu;
1938 list_add(&req->list, &rq->migration_queue);
1940 return 1;
1944 * wait_task_inactive - wait for a thread to unschedule.
1946 * If @match_state is nonzero, it's the @p->state value just checked and
1947 * not expected to change. If it changes, i.e. @p might have woken up,
1948 * then return zero. When we succeed in waiting for @p to be off its CPU,
1949 * we return a positive number (its total switch count). If a second call
1950 * a short while later returns the same number, the caller can be sure that
1951 * @p has remained unscheduled the whole time.
1953 * The caller must ensure that the task *will* unschedule sometime soon,
1954 * else this function might spin for a *long* time. This function can't
1955 * be called with interrupts off, or it may introduce deadlock with
1956 * smp_call_function() if an IPI is sent by the same process we are
1957 * waiting to become inactive.
1959 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1961 unsigned long flags;
1962 int running, on_rq;
1963 unsigned long ncsw;
1964 struct rq *rq;
1966 for (;;) {
1968 * We do the initial early heuristics without holding
1969 * any task-queue locks at all. We'll only try to get
1970 * the runqueue lock when things look like they will
1971 * work out!
1973 rq = task_rq(p);
1976 * If the task is actively running on another CPU
1977 * still, just relax and busy-wait without holding
1978 * any locks.
1980 * NOTE! Since we don't hold any locks, it's not
1981 * even sure that "rq" stays as the right runqueue!
1982 * But we don't care, since "task_running()" will
1983 * return false if the runqueue has changed and p
1984 * is actually now running somewhere else!
1986 while (task_running(rq, p)) {
1987 if (match_state && unlikely(p->state != match_state))
1988 return 0;
1989 cpu_relax();
1993 * Ok, time to look more closely! We need the rq
1994 * lock now, to be *sure*. If we're wrong, we'll
1995 * just go back and repeat.
1997 rq = task_rq_lock(p, &flags);
1998 trace_sched_wait_task(rq, p);
1999 running = task_running(rq, p);
2000 on_rq = p->se.on_rq;
2001 ncsw = 0;
2002 if (!match_state || p->state == match_state)
2003 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2004 task_rq_unlock(rq, &flags);
2007 * If it changed from the expected state, bail out now.
2009 if (unlikely(!ncsw))
2010 break;
2013 * Was it really running after all now that we
2014 * checked with the proper locks actually held?
2016 * Oops. Go back and try again..
2018 if (unlikely(running)) {
2019 cpu_relax();
2020 continue;
2024 * It's not enough that it's not actively running,
2025 * it must be off the runqueue _entirely_, and not
2026 * preempted!
2028 * So if it wa still runnable (but just not actively
2029 * running right now), it's preempted, and we should
2030 * yield - it could be a while.
2032 if (unlikely(on_rq)) {
2033 schedule_timeout_uninterruptible(1);
2034 continue;
2038 * Ahh, all good. It wasn't running, and it wasn't
2039 * runnable, which means that it will never become
2040 * running in the future either. We're all done!
2042 break;
2045 return ncsw;
2048 /***
2049 * kick_process - kick a running thread to enter/exit the kernel
2050 * @p: the to-be-kicked thread
2052 * Cause a process which is running on another CPU to enter
2053 * kernel-mode, without any delay. (to get signals handled.)
2055 * NOTE: this function doesnt have to take the runqueue lock,
2056 * because all it wants to ensure is that the remote task enters
2057 * the kernel. If the IPI races and the task has been migrated
2058 * to another CPU then no harm is done and the purpose has been
2059 * achieved as well.
2061 void kick_process(struct task_struct *p)
2063 int cpu;
2065 preempt_disable();
2066 cpu = task_cpu(p);
2067 if ((cpu != smp_processor_id()) && task_curr(p))
2068 smp_send_reschedule(cpu);
2069 preempt_enable();
2073 * Return a low guess at the load of a migration-source cpu weighted
2074 * according to the scheduling class and "nice" value.
2076 * We want to under-estimate the load of migration sources, to
2077 * balance conservatively.
2079 static unsigned long source_load(int cpu, int type)
2081 struct rq *rq = cpu_rq(cpu);
2082 unsigned long total = weighted_cpuload(cpu);
2084 if (type == 0 || !sched_feat(LB_BIAS))
2085 return total;
2087 return min(rq->cpu_load[type-1], total);
2091 * Return a high guess at the load of a migration-target cpu weighted
2092 * according to the scheduling class and "nice" value.
2094 static unsigned long target_load(int cpu, int type)
2096 struct rq *rq = cpu_rq(cpu);
2097 unsigned long total = weighted_cpuload(cpu);
2099 if (type == 0 || !sched_feat(LB_BIAS))
2100 return total;
2102 return max(rq->cpu_load[type-1], total);
2106 * find_idlest_group finds and returns the least busy CPU group within the
2107 * domain.
2109 static struct sched_group *
2110 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2112 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2113 unsigned long min_load = ULONG_MAX, this_load = 0;
2114 int load_idx = sd->forkexec_idx;
2115 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2117 do {
2118 unsigned long load, avg_load;
2119 int local_group;
2120 int i;
2122 /* Skip over this group if it has no CPUs allowed */
2123 if (!cpumask_intersects(sched_group_cpus(group),
2124 &p->cpus_allowed))
2125 continue;
2127 local_group = cpumask_test_cpu(this_cpu,
2128 sched_group_cpus(group));
2130 /* Tally up the load of all CPUs in the group */
2131 avg_load = 0;
2133 for_each_cpu(i, sched_group_cpus(group)) {
2134 /* Bias balancing toward cpus of our domain */
2135 if (local_group)
2136 load = source_load(i, load_idx);
2137 else
2138 load = target_load(i, load_idx);
2140 avg_load += load;
2143 /* Adjust by relative CPU power of the group */
2144 avg_load = sg_div_cpu_power(group,
2145 avg_load * SCHED_LOAD_SCALE);
2147 if (local_group) {
2148 this_load = avg_load;
2149 this = group;
2150 } else if (avg_load < min_load) {
2151 min_load = avg_load;
2152 idlest = group;
2154 } while (group = group->next, group != sd->groups);
2156 if (!idlest || 100*this_load < imbalance*min_load)
2157 return NULL;
2158 return idlest;
2162 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2164 static int
2165 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2167 unsigned long load, min_load = ULONG_MAX;
2168 int idlest = -1;
2169 int i;
2171 /* Traverse only the allowed CPUs */
2172 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2173 load = weighted_cpuload(i);
2175 if (load < min_load || (load == min_load && i == this_cpu)) {
2176 min_load = load;
2177 idlest = i;
2181 return idlest;
2185 * sched_balance_self: balance the current task (running on cpu) in domains
2186 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2187 * SD_BALANCE_EXEC.
2189 * Balance, ie. select the least loaded group.
2191 * Returns the target CPU number, or the same CPU if no balancing is needed.
2193 * preempt must be disabled.
2195 static int sched_balance_self(int cpu, int flag)
2197 struct task_struct *t = current;
2198 struct sched_domain *tmp, *sd = NULL;
2200 for_each_domain(cpu, tmp) {
2202 * If power savings logic is enabled for a domain, stop there.
2204 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2205 break;
2206 if (tmp->flags & flag)
2207 sd = tmp;
2210 if (sd)
2211 update_shares(sd);
2213 while (sd) {
2214 struct sched_group *group;
2215 int new_cpu, weight;
2217 if (!(sd->flags & flag)) {
2218 sd = sd->child;
2219 continue;
2222 group = find_idlest_group(sd, t, cpu);
2223 if (!group) {
2224 sd = sd->child;
2225 continue;
2228 new_cpu = find_idlest_cpu(group, t, cpu);
2229 if (new_cpu == -1 || new_cpu == cpu) {
2230 /* Now try balancing at a lower domain level of cpu */
2231 sd = sd->child;
2232 continue;
2235 /* Now try balancing at a lower domain level of new_cpu */
2236 cpu = new_cpu;
2237 weight = cpumask_weight(sched_domain_span(sd));
2238 sd = NULL;
2239 for_each_domain(cpu, tmp) {
2240 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2241 break;
2242 if (tmp->flags & flag)
2243 sd = tmp;
2245 /* while loop will break here if sd == NULL */
2248 return cpu;
2251 #endif /* CONFIG_SMP */
2253 /***
2254 * try_to_wake_up - wake up a thread
2255 * @p: the to-be-woken-up thread
2256 * @state: the mask of task states that can be woken
2257 * @sync: do a synchronous wakeup?
2259 * Put it on the run-queue if it's not already there. The "current"
2260 * thread is always on the run-queue (except when the actual
2261 * re-schedule is in progress), and as such you're allowed to do
2262 * the simpler "current->state = TASK_RUNNING" to mark yourself
2263 * runnable without the overhead of this.
2265 * returns failure only if the task is already active.
2267 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2269 int cpu, orig_cpu, this_cpu, success = 0;
2270 unsigned long flags;
2271 long old_state;
2272 struct rq *rq;
2274 if (!sched_feat(SYNC_WAKEUPS))
2275 sync = 0;
2277 #ifdef CONFIG_SMP
2278 if (sched_feat(LB_WAKEUP_UPDATE)) {
2279 struct sched_domain *sd;
2281 this_cpu = raw_smp_processor_id();
2282 cpu = task_cpu(p);
2284 for_each_domain(this_cpu, sd) {
2285 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2286 update_shares(sd);
2287 break;
2291 #endif
2293 smp_wmb();
2294 rq = task_rq_lock(p, &flags);
2295 update_rq_clock(rq);
2296 old_state = p->state;
2297 if (!(old_state & state))
2298 goto out;
2300 if (p->se.on_rq)
2301 goto out_running;
2303 cpu = task_cpu(p);
2304 orig_cpu = cpu;
2305 this_cpu = smp_processor_id();
2307 #ifdef CONFIG_SMP
2308 if (unlikely(task_running(rq, p)))
2309 goto out_activate;
2311 cpu = p->sched_class->select_task_rq(p, sync);
2312 if (cpu != orig_cpu) {
2313 set_task_cpu(p, cpu);
2314 task_rq_unlock(rq, &flags);
2315 /* might preempt at this point */
2316 rq = task_rq_lock(p, &flags);
2317 old_state = p->state;
2318 if (!(old_state & state))
2319 goto out;
2320 if (p->se.on_rq)
2321 goto out_running;
2323 this_cpu = smp_processor_id();
2324 cpu = task_cpu(p);
2327 #ifdef CONFIG_SCHEDSTATS
2328 schedstat_inc(rq, ttwu_count);
2329 if (cpu == this_cpu)
2330 schedstat_inc(rq, ttwu_local);
2331 else {
2332 struct sched_domain *sd;
2333 for_each_domain(this_cpu, sd) {
2334 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2335 schedstat_inc(sd, ttwu_wake_remote);
2336 break;
2340 #endif /* CONFIG_SCHEDSTATS */
2342 out_activate:
2343 #endif /* CONFIG_SMP */
2344 schedstat_inc(p, se.nr_wakeups);
2345 if (sync)
2346 schedstat_inc(p, se.nr_wakeups_sync);
2347 if (orig_cpu != cpu)
2348 schedstat_inc(p, se.nr_wakeups_migrate);
2349 if (cpu == this_cpu)
2350 schedstat_inc(p, se.nr_wakeups_local);
2351 else
2352 schedstat_inc(p, se.nr_wakeups_remote);
2353 activate_task(rq, p, 1);
2354 success = 1;
2356 out_running:
2357 trace_sched_wakeup(rq, p, success);
2358 check_preempt_curr(rq, p, sync);
2360 p->state = TASK_RUNNING;
2361 #ifdef CONFIG_SMP
2362 if (p->sched_class->task_wake_up)
2363 p->sched_class->task_wake_up(rq, p);
2364 #endif
2365 out:
2366 current->se.last_wakeup = current->se.sum_exec_runtime;
2368 task_rq_unlock(rq, &flags);
2370 return success;
2373 int wake_up_process(struct task_struct *p)
2375 return try_to_wake_up(p, TASK_ALL, 0);
2377 EXPORT_SYMBOL(wake_up_process);
2379 int wake_up_state(struct task_struct *p, unsigned int state)
2381 return try_to_wake_up(p, state, 0);
2385 * Perform scheduler related setup for a newly forked process p.
2386 * p is forked by current.
2388 * __sched_fork() is basic setup used by init_idle() too:
2390 static void __sched_fork(struct task_struct *p)
2392 p->se.exec_start = 0;
2393 p->se.sum_exec_runtime = 0;
2394 p->se.prev_sum_exec_runtime = 0;
2395 p->se.last_wakeup = 0;
2396 p->se.avg_overlap = 0;
2398 #ifdef CONFIG_SCHEDSTATS
2399 p->se.wait_start = 0;
2400 p->se.sum_sleep_runtime = 0;
2401 p->se.sleep_start = 0;
2402 p->se.block_start = 0;
2403 p->se.sleep_max = 0;
2404 p->se.block_max = 0;
2405 p->se.exec_max = 0;
2406 p->se.slice_max = 0;
2407 p->se.wait_max = 0;
2408 #endif
2410 INIT_LIST_HEAD(&p->rt.run_list);
2411 p->se.on_rq = 0;
2412 INIT_LIST_HEAD(&p->se.group_node);
2414 #ifdef CONFIG_PREEMPT_NOTIFIERS
2415 INIT_HLIST_HEAD(&p->preempt_notifiers);
2416 #endif
2419 * We mark the process as running here, but have not actually
2420 * inserted it onto the runqueue yet. This guarantees that
2421 * nobody will actually run it, and a signal or other external
2422 * event cannot wake it up and insert it on the runqueue either.
2424 p->state = TASK_RUNNING;
2428 * fork()/clone()-time setup:
2430 void sched_fork(struct task_struct *p, int clone_flags)
2432 int cpu = get_cpu();
2434 __sched_fork(p);
2436 #ifdef CONFIG_SMP
2437 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2438 #endif
2439 set_task_cpu(p, cpu);
2442 * Make sure we do not leak PI boosting priority to the child:
2444 p->prio = current->normal_prio;
2445 if (!rt_prio(p->prio))
2446 p->sched_class = &fair_sched_class;
2448 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2449 if (likely(sched_info_on()))
2450 memset(&p->sched_info, 0, sizeof(p->sched_info));
2451 #endif
2452 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2453 p->oncpu = 0;
2454 #endif
2455 #ifdef CONFIG_PREEMPT
2456 /* Want to start with kernel preemption disabled. */
2457 task_thread_info(p)->preempt_count = 1;
2458 #endif
2459 put_cpu();
2463 * wake_up_new_task - wake up a newly created task for the first time.
2465 * This function will do some initial scheduler statistics housekeeping
2466 * that must be done for every newly created context, then puts the task
2467 * on the runqueue and wakes it.
2469 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2471 unsigned long flags;
2472 struct rq *rq;
2474 rq = task_rq_lock(p, &flags);
2475 BUG_ON(p->state != TASK_RUNNING);
2476 update_rq_clock(rq);
2478 p->prio = effective_prio(p);
2480 if (!p->sched_class->task_new || !current->se.on_rq) {
2481 activate_task(rq, p, 0);
2482 } else {
2484 * Let the scheduling class do new task startup
2485 * management (if any):
2487 p->sched_class->task_new(rq, p);
2488 inc_nr_running(rq);
2490 trace_sched_wakeup_new(rq, p, 1);
2491 check_preempt_curr(rq, p, 0);
2492 #ifdef CONFIG_SMP
2493 if (p->sched_class->task_wake_up)
2494 p->sched_class->task_wake_up(rq, p);
2495 #endif
2496 task_rq_unlock(rq, &flags);
2499 #ifdef CONFIG_PREEMPT_NOTIFIERS
2502 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2503 * @notifier: notifier struct to register
2505 void preempt_notifier_register(struct preempt_notifier *notifier)
2507 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2509 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2512 * preempt_notifier_unregister - no longer interested in preemption notifications
2513 * @notifier: notifier struct to unregister
2515 * This is safe to call from within a preemption notifier.
2517 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2519 hlist_del(&notifier->link);
2521 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2523 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2525 struct preempt_notifier *notifier;
2526 struct hlist_node *node;
2528 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2529 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2532 static void
2533 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2534 struct task_struct *next)
2536 struct preempt_notifier *notifier;
2537 struct hlist_node *node;
2539 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2540 notifier->ops->sched_out(notifier, next);
2543 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2545 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2549 static void
2550 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2551 struct task_struct *next)
2555 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2558 * prepare_task_switch - prepare to switch tasks
2559 * @rq: the runqueue preparing to switch
2560 * @prev: the current task that is being switched out
2561 * @next: the task we are going to switch to.
2563 * This is called with the rq lock held and interrupts off. It must
2564 * be paired with a subsequent finish_task_switch after the context
2565 * switch.
2567 * prepare_task_switch sets up locking and calls architecture specific
2568 * hooks.
2570 static inline void
2571 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2572 struct task_struct *next)
2574 fire_sched_out_preempt_notifiers(prev, next);
2575 prepare_lock_switch(rq, next);
2576 prepare_arch_switch(next);
2580 * finish_task_switch - clean up after a task-switch
2581 * @rq: runqueue associated with task-switch
2582 * @prev: the thread we just switched away from.
2584 * finish_task_switch must be called after the context switch, paired
2585 * with a prepare_task_switch call before the context switch.
2586 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2587 * and do any other architecture-specific cleanup actions.
2589 * Note that we may have delayed dropping an mm in context_switch(). If
2590 * so, we finish that here outside of the runqueue lock. (Doing it
2591 * with the lock held can cause deadlocks; see schedule() for
2592 * details.)
2594 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2595 __releases(rq->lock)
2597 struct mm_struct *mm = rq->prev_mm;
2598 long prev_state;
2600 rq->prev_mm = NULL;
2603 * A task struct has one reference for the use as "current".
2604 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2605 * schedule one last time. The schedule call will never return, and
2606 * the scheduled task must drop that reference.
2607 * The test for TASK_DEAD must occur while the runqueue locks are
2608 * still held, otherwise prev could be scheduled on another cpu, die
2609 * there before we look at prev->state, and then the reference would
2610 * be dropped twice.
2611 * Manfred Spraul <manfred@colorfullife.com>
2613 prev_state = prev->state;
2614 finish_arch_switch(prev);
2615 finish_lock_switch(rq, prev);
2616 #ifdef CONFIG_SMP
2617 if (current->sched_class->post_schedule)
2618 current->sched_class->post_schedule(rq);
2619 #endif
2621 fire_sched_in_preempt_notifiers(current);
2622 if (mm)
2623 mmdrop(mm);
2624 if (unlikely(prev_state == TASK_DEAD)) {
2626 * Remove function-return probe instances associated with this
2627 * task and put them back on the free list.
2629 kprobe_flush_task(prev);
2630 put_task_struct(prev);
2635 * schedule_tail - first thing a freshly forked thread must call.
2636 * @prev: the thread we just switched away from.
2638 asmlinkage void schedule_tail(struct task_struct *prev)
2639 __releases(rq->lock)
2641 struct rq *rq = this_rq();
2643 finish_task_switch(rq, prev);
2644 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2645 /* In this case, finish_task_switch does not reenable preemption */
2646 preempt_enable();
2647 #endif
2648 if (current->set_child_tid)
2649 put_user(task_pid_vnr(current), current->set_child_tid);
2653 * context_switch - switch to the new MM and the new
2654 * thread's register state.
2656 static inline void
2657 context_switch(struct rq *rq, struct task_struct *prev,
2658 struct task_struct *next)
2660 struct mm_struct *mm, *oldmm;
2662 prepare_task_switch(rq, prev, next);
2663 trace_sched_switch(rq, prev, next);
2664 mm = next->mm;
2665 oldmm = prev->active_mm;
2667 * For paravirt, this is coupled with an exit in switch_to to
2668 * combine the page table reload and the switch backend into
2669 * one hypercall.
2671 arch_enter_lazy_cpu_mode();
2673 if (unlikely(!mm)) {
2674 next->active_mm = oldmm;
2675 atomic_inc(&oldmm->mm_count);
2676 enter_lazy_tlb(oldmm, next);
2677 } else
2678 switch_mm(oldmm, mm, next);
2680 if (unlikely(!prev->mm)) {
2681 prev->active_mm = NULL;
2682 rq->prev_mm = oldmm;
2685 * Since the runqueue lock will be released by the next
2686 * task (which is an invalid locking op but in the case
2687 * of the scheduler it's an obvious special-case), so we
2688 * do an early lockdep release here:
2690 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2691 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2692 #endif
2694 /* Here we just switch the register state and the stack. */
2695 switch_to(prev, next, prev);
2697 barrier();
2699 * this_rq must be evaluated again because prev may have moved
2700 * CPUs since it called schedule(), thus the 'rq' on its stack
2701 * frame will be invalid.
2703 finish_task_switch(this_rq(), prev);
2707 * nr_running, nr_uninterruptible and nr_context_switches:
2709 * externally visible scheduler statistics: current number of runnable
2710 * threads, current number of uninterruptible-sleeping threads, total
2711 * number of context switches performed since bootup.
2713 unsigned long nr_running(void)
2715 unsigned long i, sum = 0;
2717 for_each_online_cpu(i)
2718 sum += cpu_rq(i)->nr_running;
2720 return sum;
2723 unsigned long nr_uninterruptible(void)
2725 unsigned long i, sum = 0;
2727 for_each_possible_cpu(i)
2728 sum += cpu_rq(i)->nr_uninterruptible;
2731 * Since we read the counters lockless, it might be slightly
2732 * inaccurate. Do not allow it to go below zero though:
2734 if (unlikely((long)sum < 0))
2735 sum = 0;
2737 return sum;
2740 unsigned long long nr_context_switches(void)
2742 int i;
2743 unsigned long long sum = 0;
2745 for_each_possible_cpu(i)
2746 sum += cpu_rq(i)->nr_switches;
2748 return sum;
2751 unsigned long nr_iowait(void)
2753 unsigned long i, sum = 0;
2755 for_each_possible_cpu(i)
2756 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2758 return sum;
2761 unsigned long nr_active(void)
2763 unsigned long i, running = 0, uninterruptible = 0;
2765 for_each_online_cpu(i) {
2766 running += cpu_rq(i)->nr_running;
2767 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2770 if (unlikely((long)uninterruptible < 0))
2771 uninterruptible = 0;
2773 return running + uninterruptible;
2777 * Update rq->cpu_load[] statistics. This function is usually called every
2778 * scheduler tick (TICK_NSEC).
2780 static void update_cpu_load(struct rq *this_rq)
2782 unsigned long this_load = this_rq->load.weight;
2783 int i, scale;
2785 this_rq->nr_load_updates++;
2787 /* Update our load: */
2788 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2789 unsigned long old_load, new_load;
2791 /* scale is effectively 1 << i now, and >> i divides by scale */
2793 old_load = this_rq->cpu_load[i];
2794 new_load = this_load;
2796 * Round up the averaging division if load is increasing. This
2797 * prevents us from getting stuck on 9 if the load is 10, for
2798 * example.
2800 if (new_load > old_load)
2801 new_load += scale-1;
2802 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2806 #ifdef CONFIG_SMP
2809 * double_rq_lock - safely lock two runqueues
2811 * Note this does not disable interrupts like task_rq_lock,
2812 * you need to do so manually before calling.
2814 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2815 __acquires(rq1->lock)
2816 __acquires(rq2->lock)
2818 BUG_ON(!irqs_disabled());
2819 if (rq1 == rq2) {
2820 spin_lock(&rq1->lock);
2821 __acquire(rq2->lock); /* Fake it out ;) */
2822 } else {
2823 if (rq1 < rq2) {
2824 spin_lock(&rq1->lock);
2825 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2826 } else {
2827 spin_lock(&rq2->lock);
2828 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2831 update_rq_clock(rq1);
2832 update_rq_clock(rq2);
2836 * double_rq_unlock - safely unlock two runqueues
2838 * Note this does not restore interrupts like task_rq_unlock,
2839 * you need to do so manually after calling.
2841 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2842 __releases(rq1->lock)
2843 __releases(rq2->lock)
2845 spin_unlock(&rq1->lock);
2846 if (rq1 != rq2)
2847 spin_unlock(&rq2->lock);
2848 else
2849 __release(rq2->lock);
2853 * If dest_cpu is allowed for this process, migrate the task to it.
2854 * This is accomplished by forcing the cpu_allowed mask to only
2855 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2856 * the cpu_allowed mask is restored.
2858 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2860 struct migration_req req;
2861 unsigned long flags;
2862 struct rq *rq;
2864 rq = task_rq_lock(p, &flags);
2865 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
2866 || unlikely(!cpu_active(dest_cpu)))
2867 goto out;
2869 /* force the process onto the specified CPU */
2870 if (migrate_task(p, dest_cpu, &req)) {
2871 /* Need to wait for migration thread (might exit: take ref). */
2872 struct task_struct *mt = rq->migration_thread;
2874 get_task_struct(mt);
2875 task_rq_unlock(rq, &flags);
2876 wake_up_process(mt);
2877 put_task_struct(mt);
2878 wait_for_completion(&req.done);
2880 return;
2882 out:
2883 task_rq_unlock(rq, &flags);
2887 * sched_exec - execve() is a valuable balancing opportunity, because at
2888 * this point the task has the smallest effective memory and cache footprint.
2890 void sched_exec(void)
2892 int new_cpu, this_cpu = get_cpu();
2893 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2894 put_cpu();
2895 if (new_cpu != this_cpu)
2896 sched_migrate_task(current, new_cpu);
2900 * pull_task - move a task from a remote runqueue to the local runqueue.
2901 * Both runqueues must be locked.
2903 static void pull_task(struct rq *src_rq, struct task_struct *p,
2904 struct rq *this_rq, int this_cpu)
2906 deactivate_task(src_rq, p, 0);
2907 set_task_cpu(p, this_cpu);
2908 activate_task(this_rq, p, 0);
2910 * Note that idle threads have a prio of MAX_PRIO, for this test
2911 * to be always true for them.
2913 check_preempt_curr(this_rq, p, 0);
2917 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2919 static
2920 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2921 struct sched_domain *sd, enum cpu_idle_type idle,
2922 int *all_pinned)
2925 * We do not migrate tasks that are:
2926 * 1) running (obviously), or
2927 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2928 * 3) are cache-hot on their current CPU.
2930 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
2931 schedstat_inc(p, se.nr_failed_migrations_affine);
2932 return 0;
2934 *all_pinned = 0;
2936 if (task_running(rq, p)) {
2937 schedstat_inc(p, se.nr_failed_migrations_running);
2938 return 0;
2942 * Aggressive migration if:
2943 * 1) task is cache cold, or
2944 * 2) too many balance attempts have failed.
2947 if (!task_hot(p, rq->clock, sd) ||
2948 sd->nr_balance_failed > sd->cache_nice_tries) {
2949 #ifdef CONFIG_SCHEDSTATS
2950 if (task_hot(p, rq->clock, sd)) {
2951 schedstat_inc(sd, lb_hot_gained[idle]);
2952 schedstat_inc(p, se.nr_forced_migrations);
2954 #endif
2955 return 1;
2958 if (task_hot(p, rq->clock, sd)) {
2959 schedstat_inc(p, se.nr_failed_migrations_hot);
2960 return 0;
2962 return 1;
2965 static unsigned long
2966 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2967 unsigned long max_load_move, struct sched_domain *sd,
2968 enum cpu_idle_type idle, int *all_pinned,
2969 int *this_best_prio, struct rq_iterator *iterator)
2971 int loops = 0, pulled = 0, pinned = 0;
2972 struct task_struct *p;
2973 long rem_load_move = max_load_move;
2975 if (max_load_move == 0)
2976 goto out;
2978 pinned = 1;
2981 * Start the load-balancing iterator:
2983 p = iterator->start(iterator->arg);
2984 next:
2985 if (!p || loops++ > sysctl_sched_nr_migrate)
2986 goto out;
2988 if ((p->se.load.weight >> 1) > rem_load_move ||
2989 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2990 p = iterator->next(iterator->arg);
2991 goto next;
2994 pull_task(busiest, p, this_rq, this_cpu);
2995 pulled++;
2996 rem_load_move -= p->se.load.weight;
2999 * We only want to steal up to the prescribed amount of weighted load.
3001 if (rem_load_move > 0) {
3002 if (p->prio < *this_best_prio)
3003 *this_best_prio = p->prio;
3004 p = iterator->next(iterator->arg);
3005 goto next;
3007 out:
3009 * Right now, this is one of only two places pull_task() is called,
3010 * so we can safely collect pull_task() stats here rather than
3011 * inside pull_task().
3013 schedstat_add(sd, lb_gained[idle], pulled);
3015 if (all_pinned)
3016 *all_pinned = pinned;
3018 return max_load_move - rem_load_move;
3022 * move_tasks tries to move up to max_load_move weighted load from busiest to
3023 * this_rq, as part of a balancing operation within domain "sd".
3024 * Returns 1 if successful and 0 otherwise.
3026 * Called with both runqueues locked.
3028 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3029 unsigned long max_load_move,
3030 struct sched_domain *sd, enum cpu_idle_type idle,
3031 int *all_pinned)
3033 const struct sched_class *class = sched_class_highest;
3034 unsigned long total_load_moved = 0;
3035 int this_best_prio = this_rq->curr->prio;
3037 do {
3038 total_load_moved +=
3039 class->load_balance(this_rq, this_cpu, busiest,
3040 max_load_move - total_load_moved,
3041 sd, idle, all_pinned, &this_best_prio);
3042 class = class->next;
3044 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3045 break;
3047 } while (class && max_load_move > total_load_moved);
3049 return total_load_moved > 0;
3052 static int
3053 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3054 struct sched_domain *sd, enum cpu_idle_type idle,
3055 struct rq_iterator *iterator)
3057 struct task_struct *p = iterator->start(iterator->arg);
3058 int pinned = 0;
3060 while (p) {
3061 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3062 pull_task(busiest, p, this_rq, this_cpu);
3064 * Right now, this is only the second place pull_task()
3065 * is called, so we can safely collect pull_task()
3066 * stats here rather than inside pull_task().
3068 schedstat_inc(sd, lb_gained[idle]);
3070 return 1;
3072 p = iterator->next(iterator->arg);
3075 return 0;
3079 * move_one_task tries to move exactly one task from busiest to this_rq, as
3080 * part of active balancing operations within "domain".
3081 * Returns 1 if successful and 0 otherwise.
3083 * Called with both runqueues locked.
3085 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3086 struct sched_domain *sd, enum cpu_idle_type idle)
3088 const struct sched_class *class;
3090 for (class = sched_class_highest; class; class = class->next)
3091 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3092 return 1;
3094 return 0;
3098 * find_busiest_group finds and returns the busiest CPU group within the
3099 * domain. It calculates and returns the amount of weighted load which
3100 * should be moved to restore balance via the imbalance parameter.
3102 static struct sched_group *
3103 find_busiest_group(struct sched_domain *sd, int this_cpu,
3104 unsigned long *imbalance, enum cpu_idle_type idle,
3105 int *sd_idle, const struct cpumask *cpus, int *balance)
3107 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3108 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3109 unsigned long max_pull;
3110 unsigned long busiest_load_per_task, busiest_nr_running;
3111 unsigned long this_load_per_task, this_nr_running;
3112 int load_idx, group_imb = 0;
3113 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3114 int power_savings_balance = 1;
3115 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3116 unsigned long min_nr_running = ULONG_MAX;
3117 struct sched_group *group_min = NULL, *group_leader = NULL;
3118 #endif
3120 max_load = this_load = total_load = total_pwr = 0;
3121 busiest_load_per_task = busiest_nr_running = 0;
3122 this_load_per_task = this_nr_running = 0;
3124 if (idle == CPU_NOT_IDLE)
3125 load_idx = sd->busy_idx;
3126 else if (idle == CPU_NEWLY_IDLE)
3127 load_idx = sd->newidle_idx;
3128 else
3129 load_idx = sd->idle_idx;
3131 do {
3132 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3133 int local_group;
3134 int i;
3135 int __group_imb = 0;
3136 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3137 unsigned long sum_nr_running, sum_weighted_load;
3138 unsigned long sum_avg_load_per_task;
3139 unsigned long avg_load_per_task;
3141 local_group = cpumask_test_cpu(this_cpu,
3142 sched_group_cpus(group));
3144 if (local_group)
3145 balance_cpu = cpumask_first(sched_group_cpus(group));
3147 /* Tally up the load of all CPUs in the group */
3148 sum_weighted_load = sum_nr_running = avg_load = 0;
3149 sum_avg_load_per_task = avg_load_per_task = 0;
3151 max_cpu_load = 0;
3152 min_cpu_load = ~0UL;
3154 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3155 struct rq *rq = cpu_rq(i);
3157 if (*sd_idle && rq->nr_running)
3158 *sd_idle = 0;
3160 /* Bias balancing toward cpus of our domain */
3161 if (local_group) {
3162 if (idle_cpu(i) && !first_idle_cpu) {
3163 first_idle_cpu = 1;
3164 balance_cpu = i;
3167 load = target_load(i, load_idx);
3168 } else {
3169 load = source_load(i, load_idx);
3170 if (load > max_cpu_load)
3171 max_cpu_load = load;
3172 if (min_cpu_load > load)
3173 min_cpu_load = load;
3176 avg_load += load;
3177 sum_nr_running += rq->nr_running;
3178 sum_weighted_load += weighted_cpuload(i);
3180 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3184 * First idle cpu or the first cpu(busiest) in this sched group
3185 * is eligible for doing load balancing at this and above
3186 * domains. In the newly idle case, we will allow all the cpu's
3187 * to do the newly idle load balance.
3189 if (idle != CPU_NEWLY_IDLE && local_group &&
3190 balance_cpu != this_cpu && balance) {
3191 *balance = 0;
3192 goto ret;
3195 total_load += avg_load;
3196 total_pwr += group->__cpu_power;
3198 /* Adjust by relative CPU power of the group */
3199 avg_load = sg_div_cpu_power(group,
3200 avg_load * SCHED_LOAD_SCALE);
3204 * Consider the group unbalanced when the imbalance is larger
3205 * than the average weight of two tasks.
3207 * APZ: with cgroup the avg task weight can vary wildly and
3208 * might not be a suitable number - should we keep a
3209 * normalized nr_running number somewhere that negates
3210 * the hierarchy?
3212 avg_load_per_task = sg_div_cpu_power(group,
3213 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3215 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3216 __group_imb = 1;
3218 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3220 if (local_group) {
3221 this_load = avg_load;
3222 this = group;
3223 this_nr_running = sum_nr_running;
3224 this_load_per_task = sum_weighted_load;
3225 } else if (avg_load > max_load &&
3226 (sum_nr_running > group_capacity || __group_imb)) {
3227 max_load = avg_load;
3228 busiest = group;
3229 busiest_nr_running = sum_nr_running;
3230 busiest_load_per_task = sum_weighted_load;
3231 group_imb = __group_imb;
3234 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3236 * Busy processors will not participate in power savings
3237 * balance.
3239 if (idle == CPU_NOT_IDLE ||
3240 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3241 goto group_next;
3244 * If the local group is idle or completely loaded
3245 * no need to do power savings balance at this domain
3247 if (local_group && (this_nr_running >= group_capacity ||
3248 !this_nr_running))
3249 power_savings_balance = 0;
3252 * If a group is already running at full capacity or idle,
3253 * don't include that group in power savings calculations
3255 if (!power_savings_balance || sum_nr_running >= group_capacity
3256 || !sum_nr_running)
3257 goto group_next;
3260 * Calculate the group which has the least non-idle load.
3261 * This is the group from where we need to pick up the load
3262 * for saving power
3264 if ((sum_nr_running < min_nr_running) ||
3265 (sum_nr_running == min_nr_running &&
3266 cpumask_first(sched_group_cpus(group)) >
3267 cpumask_first(sched_group_cpus(group_min)))) {
3268 group_min = group;
3269 min_nr_running = sum_nr_running;
3270 min_load_per_task = sum_weighted_load /
3271 sum_nr_running;
3275 * Calculate the group which is almost near its
3276 * capacity but still has some space to pick up some load
3277 * from other group and save more power
3279 if (sum_nr_running <= group_capacity - 1) {
3280 if (sum_nr_running > leader_nr_running ||
3281 (sum_nr_running == leader_nr_running &&
3282 cpumask_first(sched_group_cpus(group)) <
3283 cpumask_first(sched_group_cpus(group_leader)))) {
3284 group_leader = group;
3285 leader_nr_running = sum_nr_running;
3288 group_next:
3289 #endif
3290 group = group->next;
3291 } while (group != sd->groups);
3293 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3294 goto out_balanced;
3296 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3298 if (this_load >= avg_load ||
3299 100*max_load <= sd->imbalance_pct*this_load)
3300 goto out_balanced;
3302 busiest_load_per_task /= busiest_nr_running;
3303 if (group_imb)
3304 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3307 * We're trying to get all the cpus to the average_load, so we don't
3308 * want to push ourselves above the average load, nor do we wish to
3309 * reduce the max loaded cpu below the average load, as either of these
3310 * actions would just result in more rebalancing later, and ping-pong
3311 * tasks around. Thus we look for the minimum possible imbalance.
3312 * Negative imbalances (*we* are more loaded than anyone else) will
3313 * be counted as no imbalance for these purposes -- we can't fix that
3314 * by pulling tasks to us. Be careful of negative numbers as they'll
3315 * appear as very large values with unsigned longs.
3317 if (max_load <= busiest_load_per_task)
3318 goto out_balanced;
3321 * In the presence of smp nice balancing, certain scenarios can have
3322 * max load less than avg load(as we skip the groups at or below
3323 * its cpu_power, while calculating max_load..)
3325 if (max_load < avg_load) {
3326 *imbalance = 0;
3327 goto small_imbalance;
3330 /* Don't want to pull so many tasks that a group would go idle */
3331 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3333 /* How much load to actually move to equalise the imbalance */
3334 *imbalance = min(max_pull * busiest->__cpu_power,
3335 (avg_load - this_load) * this->__cpu_power)
3336 / SCHED_LOAD_SCALE;
3339 * if *imbalance is less than the average load per runnable task
3340 * there is no gaurantee that any tasks will be moved so we'll have
3341 * a think about bumping its value to force at least one task to be
3342 * moved
3344 if (*imbalance < busiest_load_per_task) {
3345 unsigned long tmp, pwr_now, pwr_move;
3346 unsigned int imbn;
3348 small_imbalance:
3349 pwr_move = pwr_now = 0;
3350 imbn = 2;
3351 if (this_nr_running) {
3352 this_load_per_task /= this_nr_running;
3353 if (busiest_load_per_task > this_load_per_task)
3354 imbn = 1;
3355 } else
3356 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3358 if (max_load - this_load + busiest_load_per_task >=
3359 busiest_load_per_task * imbn) {
3360 *imbalance = busiest_load_per_task;
3361 return busiest;
3365 * OK, we don't have enough imbalance to justify moving tasks,
3366 * however we may be able to increase total CPU power used by
3367 * moving them.
3370 pwr_now += busiest->__cpu_power *
3371 min(busiest_load_per_task, max_load);
3372 pwr_now += this->__cpu_power *
3373 min(this_load_per_task, this_load);
3374 pwr_now /= SCHED_LOAD_SCALE;
3376 /* Amount of load we'd subtract */
3377 tmp = sg_div_cpu_power(busiest,
3378 busiest_load_per_task * SCHED_LOAD_SCALE);
3379 if (max_load > tmp)
3380 pwr_move += busiest->__cpu_power *
3381 min(busiest_load_per_task, max_load - tmp);
3383 /* Amount of load we'd add */
3384 if (max_load * busiest->__cpu_power <
3385 busiest_load_per_task * SCHED_LOAD_SCALE)
3386 tmp = sg_div_cpu_power(this,
3387 max_load * busiest->__cpu_power);
3388 else
3389 tmp = sg_div_cpu_power(this,
3390 busiest_load_per_task * SCHED_LOAD_SCALE);
3391 pwr_move += this->__cpu_power *
3392 min(this_load_per_task, this_load + tmp);
3393 pwr_move /= SCHED_LOAD_SCALE;
3395 /* Move if we gain throughput */
3396 if (pwr_move > pwr_now)
3397 *imbalance = busiest_load_per_task;
3400 return busiest;
3402 out_balanced:
3403 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3404 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3405 goto ret;
3407 if (this == group_leader && group_leader != group_min) {
3408 *imbalance = min_load_per_task;
3409 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3410 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3411 cpumask_first(sched_group_cpus(group_leader));
3413 return group_min;
3415 #endif
3416 ret:
3417 *imbalance = 0;
3418 return NULL;
3422 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3424 static struct rq *
3425 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3426 unsigned long imbalance, const struct cpumask *cpus)
3428 struct rq *busiest = NULL, *rq;
3429 unsigned long max_load = 0;
3430 int i;
3432 for_each_cpu(i, sched_group_cpus(group)) {
3433 unsigned long wl;
3435 if (!cpumask_test_cpu(i, cpus))
3436 continue;
3438 rq = cpu_rq(i);
3439 wl = weighted_cpuload(i);
3441 if (rq->nr_running == 1 && wl > imbalance)
3442 continue;
3444 if (wl > max_load) {
3445 max_load = wl;
3446 busiest = rq;
3450 return busiest;
3454 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3455 * so long as it is large enough.
3457 #define MAX_PINNED_INTERVAL 512
3460 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3461 * tasks if there is an imbalance.
3463 static int load_balance(int this_cpu, struct rq *this_rq,
3464 struct sched_domain *sd, enum cpu_idle_type idle,
3465 int *balance, struct cpumask *cpus)
3467 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3468 struct sched_group *group;
3469 unsigned long imbalance;
3470 struct rq *busiest;
3471 unsigned long flags;
3473 cpumask_setall(cpus);
3476 * When power savings policy is enabled for the parent domain, idle
3477 * sibling can pick up load irrespective of busy siblings. In this case,
3478 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3479 * portraying it as CPU_NOT_IDLE.
3481 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3482 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3483 sd_idle = 1;
3485 schedstat_inc(sd, lb_count[idle]);
3487 redo:
3488 update_shares(sd);
3489 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3490 cpus, balance);
3492 if (*balance == 0)
3493 goto out_balanced;
3495 if (!group) {
3496 schedstat_inc(sd, lb_nobusyg[idle]);
3497 goto out_balanced;
3500 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3501 if (!busiest) {
3502 schedstat_inc(sd, lb_nobusyq[idle]);
3503 goto out_balanced;
3506 BUG_ON(busiest == this_rq);
3508 schedstat_add(sd, lb_imbalance[idle], imbalance);
3510 ld_moved = 0;
3511 if (busiest->nr_running > 1) {
3513 * Attempt to move tasks. If find_busiest_group has found
3514 * an imbalance but busiest->nr_running <= 1, the group is
3515 * still unbalanced. ld_moved simply stays zero, so it is
3516 * correctly treated as an imbalance.
3518 local_irq_save(flags);
3519 double_rq_lock(this_rq, busiest);
3520 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3521 imbalance, sd, idle, &all_pinned);
3522 double_rq_unlock(this_rq, busiest);
3523 local_irq_restore(flags);
3526 * some other cpu did the load balance for us.
3528 if (ld_moved && this_cpu != smp_processor_id())
3529 resched_cpu(this_cpu);
3531 /* All tasks on this runqueue were pinned by CPU affinity */
3532 if (unlikely(all_pinned)) {
3533 cpumask_clear_cpu(cpu_of(busiest), cpus);
3534 if (!cpumask_empty(cpus))
3535 goto redo;
3536 goto out_balanced;
3540 if (!ld_moved) {
3541 schedstat_inc(sd, lb_failed[idle]);
3542 sd->nr_balance_failed++;
3544 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3546 spin_lock_irqsave(&busiest->lock, flags);
3548 /* don't kick the migration_thread, if the curr
3549 * task on busiest cpu can't be moved to this_cpu
3551 if (!cpumask_test_cpu(this_cpu,
3552 &busiest->curr->cpus_allowed)) {
3553 spin_unlock_irqrestore(&busiest->lock, flags);
3554 all_pinned = 1;
3555 goto out_one_pinned;
3558 if (!busiest->active_balance) {
3559 busiest->active_balance = 1;
3560 busiest->push_cpu = this_cpu;
3561 active_balance = 1;
3563 spin_unlock_irqrestore(&busiest->lock, flags);
3564 if (active_balance)
3565 wake_up_process(busiest->migration_thread);
3568 * We've kicked active balancing, reset the failure
3569 * counter.
3571 sd->nr_balance_failed = sd->cache_nice_tries+1;
3573 } else
3574 sd->nr_balance_failed = 0;
3576 if (likely(!active_balance)) {
3577 /* We were unbalanced, so reset the balancing interval */
3578 sd->balance_interval = sd->min_interval;
3579 } else {
3581 * If we've begun active balancing, start to back off. This
3582 * case may not be covered by the all_pinned logic if there
3583 * is only 1 task on the busy runqueue (because we don't call
3584 * move_tasks).
3586 if (sd->balance_interval < sd->max_interval)
3587 sd->balance_interval *= 2;
3590 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3591 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3592 ld_moved = -1;
3594 goto out;
3596 out_balanced:
3597 schedstat_inc(sd, lb_balanced[idle]);
3599 sd->nr_balance_failed = 0;
3601 out_one_pinned:
3602 /* tune up the balancing interval */
3603 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3604 (sd->balance_interval < sd->max_interval))
3605 sd->balance_interval *= 2;
3607 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3608 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3609 ld_moved = -1;
3610 else
3611 ld_moved = 0;
3612 out:
3613 if (ld_moved)
3614 update_shares(sd);
3615 return ld_moved;
3619 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3620 * tasks if there is an imbalance.
3622 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3623 * this_rq is locked.
3625 static int
3626 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3627 struct cpumask *cpus)
3629 struct sched_group *group;
3630 struct rq *busiest = NULL;
3631 unsigned long imbalance;
3632 int ld_moved = 0;
3633 int sd_idle = 0;
3634 int all_pinned = 0;
3636 cpumask_setall(cpus);
3639 * When power savings policy is enabled for the parent domain, idle
3640 * sibling can pick up load irrespective of busy siblings. In this case,
3641 * let the state of idle sibling percolate up as IDLE, instead of
3642 * portraying it as CPU_NOT_IDLE.
3644 if (sd->flags & SD_SHARE_CPUPOWER &&
3645 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3646 sd_idle = 1;
3648 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3649 redo:
3650 update_shares_locked(this_rq, sd);
3651 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3652 &sd_idle, cpus, NULL);
3653 if (!group) {
3654 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3655 goto out_balanced;
3658 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3659 if (!busiest) {
3660 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3661 goto out_balanced;
3664 BUG_ON(busiest == this_rq);
3666 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3668 ld_moved = 0;
3669 if (busiest->nr_running > 1) {
3670 /* Attempt to move tasks */
3671 double_lock_balance(this_rq, busiest);
3672 /* this_rq->clock is already updated */
3673 update_rq_clock(busiest);
3674 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3675 imbalance, sd, CPU_NEWLY_IDLE,
3676 &all_pinned);
3677 double_unlock_balance(this_rq, busiest);
3679 if (unlikely(all_pinned)) {
3680 cpumask_clear_cpu(cpu_of(busiest), cpus);
3681 if (!cpumask_empty(cpus))
3682 goto redo;
3686 if (!ld_moved) {
3687 int active_balance = 0;
3689 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3690 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3691 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3692 return -1;
3694 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3695 return -1;
3697 if (sd->nr_balance_failed++ < 2)
3698 return -1;
3701 * The only task running in a non-idle cpu can be moved to this
3702 * cpu in an attempt to completely freeup the other CPU
3703 * package. The same method used to move task in load_balance()
3704 * have been extended for load_balance_newidle() to speedup
3705 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
3707 * The package power saving logic comes from
3708 * find_busiest_group(). If there are no imbalance, then
3709 * f_b_g() will return NULL. However when sched_mc={1,2} then
3710 * f_b_g() will select a group from which a running task may be
3711 * pulled to this cpu in order to make the other package idle.
3712 * If there is no opportunity to make a package idle and if
3713 * there are no imbalance, then f_b_g() will return NULL and no
3714 * action will be taken in load_balance_newidle().
3716 * Under normal task pull operation due to imbalance, there
3717 * will be more than one task in the source run queue and
3718 * move_tasks() will succeed. ld_moved will be true and this
3719 * active balance code will not be triggered.
3722 /* Lock busiest in correct order while this_rq is held */
3723 double_lock_balance(this_rq, busiest);
3726 * don't kick the migration_thread, if the curr
3727 * task on busiest cpu can't be moved to this_cpu
3729 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
3730 double_unlock_balance(this_rq, busiest);
3731 all_pinned = 1;
3732 return ld_moved;
3735 if (!busiest->active_balance) {
3736 busiest->active_balance = 1;
3737 busiest->push_cpu = this_cpu;
3738 active_balance = 1;
3741 double_unlock_balance(this_rq, busiest);
3743 * Should not call ttwu while holding a rq->lock
3745 spin_unlock(&this_rq->lock);
3746 if (active_balance)
3747 wake_up_process(busiest->migration_thread);
3748 spin_lock(&this_rq->lock);
3750 } else
3751 sd->nr_balance_failed = 0;
3753 update_shares_locked(this_rq, sd);
3754 return ld_moved;
3756 out_balanced:
3757 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3758 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3759 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3760 return -1;
3761 sd->nr_balance_failed = 0;
3763 return 0;
3767 * idle_balance is called by schedule() if this_cpu is about to become
3768 * idle. Attempts to pull tasks from other CPUs.
3770 static void idle_balance(int this_cpu, struct rq *this_rq)
3772 struct sched_domain *sd;
3773 int pulled_task = 0;
3774 unsigned long next_balance = jiffies + HZ;
3775 cpumask_var_t tmpmask;
3777 if (!alloc_cpumask_var(&tmpmask, GFP_ATOMIC))
3778 return;
3780 for_each_domain(this_cpu, sd) {
3781 unsigned long interval;
3783 if (!(sd->flags & SD_LOAD_BALANCE))
3784 continue;
3786 if (sd->flags & SD_BALANCE_NEWIDLE)
3787 /* If we've pulled tasks over stop searching: */
3788 pulled_task = load_balance_newidle(this_cpu, this_rq,
3789 sd, tmpmask);
3791 interval = msecs_to_jiffies(sd->balance_interval);
3792 if (time_after(next_balance, sd->last_balance + interval))
3793 next_balance = sd->last_balance + interval;
3794 if (pulled_task)
3795 break;
3797 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3799 * We are going idle. next_balance may be set based on
3800 * a busy processor. So reset next_balance.
3802 this_rq->next_balance = next_balance;
3804 free_cpumask_var(tmpmask);
3808 * active_load_balance is run by migration threads. It pushes running tasks
3809 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3810 * running on each physical CPU where possible, and avoids physical /
3811 * logical imbalances.
3813 * Called with busiest_rq locked.
3815 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3817 int target_cpu = busiest_rq->push_cpu;
3818 struct sched_domain *sd;
3819 struct rq *target_rq;
3821 /* Is there any task to move? */
3822 if (busiest_rq->nr_running <= 1)
3823 return;
3825 target_rq = cpu_rq(target_cpu);
3828 * This condition is "impossible", if it occurs
3829 * we need to fix it. Originally reported by
3830 * Bjorn Helgaas on a 128-cpu setup.
3832 BUG_ON(busiest_rq == target_rq);
3834 /* move a task from busiest_rq to target_rq */
3835 double_lock_balance(busiest_rq, target_rq);
3836 update_rq_clock(busiest_rq);
3837 update_rq_clock(target_rq);
3839 /* Search for an sd spanning us and the target CPU. */
3840 for_each_domain(target_cpu, sd) {
3841 if ((sd->flags & SD_LOAD_BALANCE) &&
3842 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3843 break;
3846 if (likely(sd)) {
3847 schedstat_inc(sd, alb_count);
3849 if (move_one_task(target_rq, target_cpu, busiest_rq,
3850 sd, CPU_IDLE))
3851 schedstat_inc(sd, alb_pushed);
3852 else
3853 schedstat_inc(sd, alb_failed);
3855 double_unlock_balance(busiest_rq, target_rq);
3858 #ifdef CONFIG_NO_HZ
3859 static struct {
3860 atomic_t load_balancer;
3861 cpumask_var_t cpu_mask;
3862 } nohz ____cacheline_aligned = {
3863 .load_balancer = ATOMIC_INIT(-1),
3867 * This routine will try to nominate the ilb (idle load balancing)
3868 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3869 * load balancing on behalf of all those cpus. If all the cpus in the system
3870 * go into this tickless mode, then there will be no ilb owner (as there is
3871 * no need for one) and all the cpus will sleep till the next wakeup event
3872 * arrives...
3874 * For the ilb owner, tick is not stopped. And this tick will be used
3875 * for idle load balancing. ilb owner will still be part of
3876 * nohz.cpu_mask..
3878 * While stopping the tick, this cpu will become the ilb owner if there
3879 * is no other owner. And will be the owner till that cpu becomes busy
3880 * or if all cpus in the system stop their ticks at which point
3881 * there is no need for ilb owner.
3883 * When the ilb owner becomes busy, it nominates another owner, during the
3884 * next busy scheduler_tick()
3886 int select_nohz_load_balancer(int stop_tick)
3888 int cpu = smp_processor_id();
3890 if (stop_tick) {
3891 cpu_rq(cpu)->in_nohz_recently = 1;
3893 if (!cpu_active(cpu)) {
3894 if (atomic_read(&nohz.load_balancer) != cpu)
3895 return 0;
3898 * If we are going offline and still the leader,
3899 * give up!
3901 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3902 BUG();
3904 return 0;
3907 cpumask_set_cpu(cpu, nohz.cpu_mask);
3909 /* time for ilb owner also to sleep */
3910 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
3911 if (atomic_read(&nohz.load_balancer) == cpu)
3912 atomic_set(&nohz.load_balancer, -1);
3913 return 0;
3916 if (atomic_read(&nohz.load_balancer) == -1) {
3917 /* make me the ilb owner */
3918 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3919 return 1;
3920 } else if (atomic_read(&nohz.load_balancer) == cpu)
3921 return 1;
3922 } else {
3923 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
3924 return 0;
3926 cpumask_clear_cpu(cpu, nohz.cpu_mask);
3928 if (atomic_read(&nohz.load_balancer) == cpu)
3929 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3930 BUG();
3932 return 0;
3934 #endif
3936 static DEFINE_SPINLOCK(balancing);
3939 * It checks each scheduling domain to see if it is due to be balanced,
3940 * and initiates a balancing operation if so.
3942 * Balancing parameters are set up in arch_init_sched_domains.
3944 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3946 int balance = 1;
3947 struct rq *rq = cpu_rq(cpu);
3948 unsigned long interval;
3949 struct sched_domain *sd;
3950 /* Earliest time when we have to do rebalance again */
3951 unsigned long next_balance = jiffies + 60*HZ;
3952 int update_next_balance = 0;
3953 int need_serialize;
3954 cpumask_var_t tmp;
3956 /* Fails alloc? Rebalancing probably not a priority right now. */
3957 if (!alloc_cpumask_var(&tmp, GFP_ATOMIC))
3958 return;
3960 for_each_domain(cpu, sd) {
3961 if (!(sd->flags & SD_LOAD_BALANCE))
3962 continue;
3964 interval = sd->balance_interval;
3965 if (idle != CPU_IDLE)
3966 interval *= sd->busy_factor;
3968 /* scale ms to jiffies */
3969 interval = msecs_to_jiffies(interval);
3970 if (unlikely(!interval))
3971 interval = 1;
3972 if (interval > HZ*NR_CPUS/10)
3973 interval = HZ*NR_CPUS/10;
3975 need_serialize = sd->flags & SD_SERIALIZE;
3977 if (need_serialize) {
3978 if (!spin_trylock(&balancing))
3979 goto out;
3982 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3983 if (load_balance(cpu, rq, sd, idle, &balance, tmp)) {
3985 * We've pulled tasks over so either we're no
3986 * longer idle, or one of our SMT siblings is
3987 * not idle.
3989 idle = CPU_NOT_IDLE;
3991 sd->last_balance = jiffies;
3993 if (need_serialize)
3994 spin_unlock(&balancing);
3995 out:
3996 if (time_after(next_balance, sd->last_balance + interval)) {
3997 next_balance = sd->last_balance + interval;
3998 update_next_balance = 1;
4002 * Stop the load balance at this level. There is another
4003 * CPU in our sched group which is doing load balancing more
4004 * actively.
4006 if (!balance)
4007 break;
4011 * next_balance will be updated only when there is a need.
4012 * When the cpu is attached to null domain for ex, it will not be
4013 * updated.
4015 if (likely(update_next_balance))
4016 rq->next_balance = next_balance;
4018 free_cpumask_var(tmp);
4022 * run_rebalance_domains is triggered when needed from the scheduler tick.
4023 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4024 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4026 static void run_rebalance_domains(struct softirq_action *h)
4028 int this_cpu = smp_processor_id();
4029 struct rq *this_rq = cpu_rq(this_cpu);
4030 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4031 CPU_IDLE : CPU_NOT_IDLE;
4033 rebalance_domains(this_cpu, idle);
4035 #ifdef CONFIG_NO_HZ
4037 * If this cpu is the owner for idle load balancing, then do the
4038 * balancing on behalf of the other idle cpus whose ticks are
4039 * stopped.
4041 if (this_rq->idle_at_tick &&
4042 atomic_read(&nohz.load_balancer) == this_cpu) {
4043 struct rq *rq;
4044 int balance_cpu;
4046 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4047 if (balance_cpu == this_cpu)
4048 continue;
4051 * If this cpu gets work to do, stop the load balancing
4052 * work being done for other cpus. Next load
4053 * balancing owner will pick it up.
4055 if (need_resched())
4056 break;
4058 rebalance_domains(balance_cpu, CPU_IDLE);
4060 rq = cpu_rq(balance_cpu);
4061 if (time_after(this_rq->next_balance, rq->next_balance))
4062 this_rq->next_balance = rq->next_balance;
4065 #endif
4069 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4071 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4072 * idle load balancing owner or decide to stop the periodic load balancing,
4073 * if the whole system is idle.
4075 static inline void trigger_load_balance(struct rq *rq, int cpu)
4077 #ifdef CONFIG_NO_HZ
4079 * If we were in the nohz mode recently and busy at the current
4080 * scheduler tick, then check if we need to nominate new idle
4081 * load balancer.
4083 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4084 rq->in_nohz_recently = 0;
4086 if (atomic_read(&nohz.load_balancer) == cpu) {
4087 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4088 atomic_set(&nohz.load_balancer, -1);
4091 if (atomic_read(&nohz.load_balancer) == -1) {
4093 * simple selection for now: Nominate the
4094 * first cpu in the nohz list to be the next
4095 * ilb owner.
4097 * TBD: Traverse the sched domains and nominate
4098 * the nearest cpu in the nohz.cpu_mask.
4100 int ilb = cpumask_first(nohz.cpu_mask);
4102 if (ilb < nr_cpu_ids)
4103 resched_cpu(ilb);
4108 * If this cpu is idle and doing idle load balancing for all the
4109 * cpus with ticks stopped, is it time for that to stop?
4111 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4112 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4113 resched_cpu(cpu);
4114 return;
4118 * If this cpu is idle and the idle load balancing is done by
4119 * someone else, then no need raise the SCHED_SOFTIRQ
4121 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4122 cpumask_test_cpu(cpu, nohz.cpu_mask))
4123 return;
4124 #endif
4125 if (time_after_eq(jiffies, rq->next_balance))
4126 raise_softirq(SCHED_SOFTIRQ);
4129 #else /* CONFIG_SMP */
4132 * on UP we do not need to balance between CPUs:
4134 static inline void idle_balance(int cpu, struct rq *rq)
4138 #endif
4140 DEFINE_PER_CPU(struct kernel_stat, kstat);
4142 EXPORT_PER_CPU_SYMBOL(kstat);
4145 * Return any ns on the sched_clock that have not yet been accounted in
4146 * @p in case that task is currently running.
4148 * Called with task_rq_lock() held on @rq.
4150 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4152 u64 ns = 0;
4154 if (task_current(rq, p)) {
4155 update_rq_clock(rq);
4156 ns = rq->clock - p->se.exec_start;
4157 if ((s64)ns < 0)
4158 ns = 0;
4161 return ns;
4164 unsigned long long task_delta_exec(struct task_struct *p)
4166 unsigned long flags;
4167 struct rq *rq;
4168 u64 ns = 0;
4170 rq = task_rq_lock(p, &flags);
4171 ns = do_task_delta_exec(p, rq);
4172 task_rq_unlock(rq, &flags);
4174 return ns;
4178 * Return accounted runtime for the task.
4179 * In case the task is currently running, return the runtime plus current's
4180 * pending runtime that have not been accounted yet.
4182 unsigned long long task_sched_runtime(struct task_struct *p)
4184 unsigned long flags;
4185 struct rq *rq;
4186 u64 ns = 0;
4188 rq = task_rq_lock(p, &flags);
4189 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4190 task_rq_unlock(rq, &flags);
4192 return ns;
4196 * Return sum_exec_runtime for the thread group.
4197 * In case the task is currently running, return the sum plus current's
4198 * pending runtime that have not been accounted yet.
4200 * Note that the thread group might have other running tasks as well,
4201 * so the return value not includes other pending runtime that other
4202 * running tasks might have.
4204 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4206 struct task_cputime totals;
4207 unsigned long flags;
4208 struct rq *rq;
4209 u64 ns;
4211 rq = task_rq_lock(p, &flags);
4212 thread_group_cputime(p, &totals);
4213 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4214 task_rq_unlock(rq, &flags);
4216 return ns;
4220 * Account user cpu time to a process.
4221 * @p: the process that the cpu time gets accounted to
4222 * @cputime: the cpu time spent in user space since the last update
4223 * @cputime_scaled: cputime scaled by cpu frequency
4225 void account_user_time(struct task_struct *p, cputime_t cputime,
4226 cputime_t cputime_scaled)
4228 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4229 cputime64_t tmp;
4231 /* Add user time to process. */
4232 p->utime = cputime_add(p->utime, cputime);
4233 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4234 account_group_user_time(p, cputime);
4236 /* Add user time to cpustat. */
4237 tmp = cputime_to_cputime64(cputime);
4238 if (TASK_NICE(p) > 0)
4239 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4240 else
4241 cpustat->user = cputime64_add(cpustat->user, tmp);
4242 /* Account for user time used */
4243 acct_update_integrals(p);
4247 * Account guest cpu time to a process.
4248 * @p: the process that the cpu time gets accounted to
4249 * @cputime: the cpu time spent in virtual machine since the last update
4250 * @cputime_scaled: cputime scaled by cpu frequency
4252 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4253 cputime_t cputime_scaled)
4255 cputime64_t tmp;
4256 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4258 tmp = cputime_to_cputime64(cputime);
4260 /* Add guest time to process. */
4261 p->utime = cputime_add(p->utime, cputime);
4262 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4263 account_group_user_time(p, cputime);
4264 p->gtime = cputime_add(p->gtime, cputime);
4266 /* Add guest time to cpustat. */
4267 cpustat->user = cputime64_add(cpustat->user, tmp);
4268 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4272 * Account system cpu time to a process.
4273 * @p: the process that the cpu time gets accounted to
4274 * @hardirq_offset: the offset to subtract from hardirq_count()
4275 * @cputime: the cpu time spent in kernel space since the last update
4276 * @cputime_scaled: cputime scaled by cpu frequency
4278 void account_system_time(struct task_struct *p, int hardirq_offset,
4279 cputime_t cputime, cputime_t cputime_scaled)
4281 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4282 cputime64_t tmp;
4284 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4285 account_guest_time(p, cputime, cputime_scaled);
4286 return;
4289 /* Add system time to process. */
4290 p->stime = cputime_add(p->stime, cputime);
4291 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4292 account_group_system_time(p, cputime);
4294 /* Add system time to cpustat. */
4295 tmp = cputime_to_cputime64(cputime);
4296 if (hardirq_count() - hardirq_offset)
4297 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4298 else if (softirq_count())
4299 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4300 else
4301 cpustat->system = cputime64_add(cpustat->system, tmp);
4303 /* Account for system time used */
4304 acct_update_integrals(p);
4308 * Account for involuntary wait time.
4309 * @steal: the cpu time spent in involuntary wait
4311 void account_steal_time(cputime_t cputime)
4313 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4314 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4316 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4320 * Account for idle time.
4321 * @cputime: the cpu time spent in idle wait
4323 void account_idle_time(cputime_t cputime)
4325 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4326 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4327 struct rq *rq = this_rq();
4329 if (atomic_read(&rq->nr_iowait) > 0)
4330 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4331 else
4332 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4335 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4338 * Account a single tick of cpu time.
4339 * @p: the process that the cpu time gets accounted to
4340 * @user_tick: indicates if the tick is a user or a system tick
4342 void account_process_tick(struct task_struct *p, int user_tick)
4344 cputime_t one_jiffy = jiffies_to_cputime(1);
4345 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
4346 struct rq *rq = this_rq();
4348 if (user_tick)
4349 account_user_time(p, one_jiffy, one_jiffy_scaled);
4350 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
4351 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
4352 one_jiffy_scaled);
4353 else
4354 account_idle_time(one_jiffy);
4358 * Account multiple ticks of steal time.
4359 * @p: the process from which the cpu time has been stolen
4360 * @ticks: number of stolen ticks
4362 void account_steal_ticks(unsigned long ticks)
4364 account_steal_time(jiffies_to_cputime(ticks));
4368 * Account multiple ticks of idle time.
4369 * @ticks: number of stolen ticks
4371 void account_idle_ticks(unsigned long ticks)
4373 account_idle_time(jiffies_to_cputime(ticks));
4376 #endif
4379 * Use precise platform statistics if available:
4381 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4382 cputime_t task_utime(struct task_struct *p)
4384 return p->utime;
4387 cputime_t task_stime(struct task_struct *p)
4389 return p->stime;
4391 #else
4392 cputime_t task_utime(struct task_struct *p)
4394 clock_t utime = cputime_to_clock_t(p->utime),
4395 total = utime + cputime_to_clock_t(p->stime);
4396 u64 temp;
4399 * Use CFS's precise accounting:
4401 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4403 if (total) {
4404 temp *= utime;
4405 do_div(temp, total);
4407 utime = (clock_t)temp;
4409 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4410 return p->prev_utime;
4413 cputime_t task_stime(struct task_struct *p)
4415 clock_t stime;
4418 * Use CFS's precise accounting. (we subtract utime from
4419 * the total, to make sure the total observed by userspace
4420 * grows monotonically - apps rely on that):
4422 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4423 cputime_to_clock_t(task_utime(p));
4425 if (stime >= 0)
4426 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4428 return p->prev_stime;
4430 #endif
4432 inline cputime_t task_gtime(struct task_struct *p)
4434 return p->gtime;
4438 * This function gets called by the timer code, with HZ frequency.
4439 * We call it with interrupts disabled.
4441 * It also gets called by the fork code, when changing the parent's
4442 * timeslices.
4444 void scheduler_tick(void)
4446 int cpu = smp_processor_id();
4447 struct rq *rq = cpu_rq(cpu);
4448 struct task_struct *curr = rq->curr;
4450 sched_clock_tick();
4452 spin_lock(&rq->lock);
4453 update_rq_clock(rq);
4454 update_cpu_load(rq);
4455 curr->sched_class->task_tick(rq, curr, 0);
4456 spin_unlock(&rq->lock);
4458 #ifdef CONFIG_SMP
4459 rq->idle_at_tick = idle_cpu(cpu);
4460 trigger_load_balance(rq, cpu);
4461 #endif
4464 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4465 defined(CONFIG_PREEMPT_TRACER))
4467 static inline unsigned long get_parent_ip(unsigned long addr)
4469 if (in_lock_functions(addr)) {
4470 addr = CALLER_ADDR2;
4471 if (in_lock_functions(addr))
4472 addr = CALLER_ADDR3;
4474 return addr;
4477 void __kprobes add_preempt_count(int val)
4479 #ifdef CONFIG_DEBUG_PREEMPT
4481 * Underflow?
4483 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4484 return;
4485 #endif
4486 preempt_count() += val;
4487 #ifdef CONFIG_DEBUG_PREEMPT
4489 * Spinlock count overflowing soon?
4491 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4492 PREEMPT_MASK - 10);
4493 #endif
4494 if (preempt_count() == val)
4495 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4497 EXPORT_SYMBOL(add_preempt_count);
4499 void __kprobes sub_preempt_count(int val)
4501 #ifdef CONFIG_DEBUG_PREEMPT
4503 * Underflow?
4505 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4506 return;
4508 * Is the spinlock portion underflowing?
4510 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4511 !(preempt_count() & PREEMPT_MASK)))
4512 return;
4513 #endif
4515 if (preempt_count() == val)
4516 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4517 preempt_count() -= val;
4519 EXPORT_SYMBOL(sub_preempt_count);
4521 #endif
4524 * Print scheduling while atomic bug:
4526 static noinline void __schedule_bug(struct task_struct *prev)
4528 struct pt_regs *regs = get_irq_regs();
4530 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4531 prev->comm, prev->pid, preempt_count());
4533 debug_show_held_locks(prev);
4534 print_modules();
4535 if (irqs_disabled())
4536 print_irqtrace_events(prev);
4538 if (regs)
4539 show_regs(regs);
4540 else
4541 dump_stack();
4545 * Various schedule()-time debugging checks and statistics:
4547 static inline void schedule_debug(struct task_struct *prev)
4550 * Test if we are atomic. Since do_exit() needs to call into
4551 * schedule() atomically, we ignore that path for now.
4552 * Otherwise, whine if we are scheduling when we should not be.
4554 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4555 __schedule_bug(prev);
4557 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4559 schedstat_inc(this_rq(), sched_count);
4560 #ifdef CONFIG_SCHEDSTATS
4561 if (unlikely(prev->lock_depth >= 0)) {
4562 schedstat_inc(this_rq(), bkl_count);
4563 schedstat_inc(prev, sched_info.bkl_count);
4565 #endif
4569 * Pick up the highest-prio task:
4571 static inline struct task_struct *
4572 pick_next_task(struct rq *rq, struct task_struct *prev)
4574 const struct sched_class *class;
4575 struct task_struct *p;
4578 * Optimization: we know that if all tasks are in
4579 * the fair class we can call that function directly:
4581 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4582 p = fair_sched_class.pick_next_task(rq);
4583 if (likely(p))
4584 return p;
4587 class = sched_class_highest;
4588 for ( ; ; ) {
4589 p = class->pick_next_task(rq);
4590 if (p)
4591 return p;
4593 * Will never be NULL as the idle class always
4594 * returns a non-NULL p:
4596 class = class->next;
4601 * schedule() is the main scheduler function.
4603 asmlinkage void __sched schedule(void)
4605 struct task_struct *prev, *next;
4606 unsigned long *switch_count;
4607 struct rq *rq;
4608 int cpu;
4610 need_resched:
4611 preempt_disable();
4612 cpu = smp_processor_id();
4613 rq = cpu_rq(cpu);
4614 rcu_qsctr_inc(cpu);
4615 prev = rq->curr;
4616 switch_count = &prev->nivcsw;
4618 release_kernel_lock(prev);
4619 need_resched_nonpreemptible:
4621 schedule_debug(prev);
4623 if (sched_feat(HRTICK))
4624 hrtick_clear(rq);
4626 spin_lock_irq(&rq->lock);
4627 update_rq_clock(rq);
4628 clear_tsk_need_resched(prev);
4630 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4631 if (unlikely(signal_pending_state(prev->state, prev)))
4632 prev->state = TASK_RUNNING;
4633 else
4634 deactivate_task(rq, prev, 1);
4635 switch_count = &prev->nvcsw;
4638 #ifdef CONFIG_SMP
4639 if (prev->sched_class->pre_schedule)
4640 prev->sched_class->pre_schedule(rq, prev);
4641 #endif
4643 if (unlikely(!rq->nr_running))
4644 idle_balance(cpu, rq);
4646 prev->sched_class->put_prev_task(rq, prev);
4647 next = pick_next_task(rq, prev);
4649 if (likely(prev != next)) {
4650 sched_info_switch(prev, next);
4652 rq->nr_switches++;
4653 rq->curr = next;
4654 ++*switch_count;
4656 context_switch(rq, prev, next); /* unlocks the rq */
4658 * the context switch might have flipped the stack from under
4659 * us, hence refresh the local variables.
4661 cpu = smp_processor_id();
4662 rq = cpu_rq(cpu);
4663 } else
4664 spin_unlock_irq(&rq->lock);
4666 if (unlikely(reacquire_kernel_lock(current) < 0))
4667 goto need_resched_nonpreemptible;
4669 preempt_enable_no_resched();
4670 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4671 goto need_resched;
4673 EXPORT_SYMBOL(schedule);
4675 #ifdef CONFIG_PREEMPT
4677 * this is the entry point to schedule() from in-kernel preemption
4678 * off of preempt_enable. Kernel preemptions off return from interrupt
4679 * occur there and call schedule directly.
4681 asmlinkage void __sched preempt_schedule(void)
4683 struct thread_info *ti = current_thread_info();
4686 * If there is a non-zero preempt_count or interrupts are disabled,
4687 * we do not want to preempt the current task. Just return..
4689 if (likely(ti->preempt_count || irqs_disabled()))
4690 return;
4692 do {
4693 add_preempt_count(PREEMPT_ACTIVE);
4694 schedule();
4695 sub_preempt_count(PREEMPT_ACTIVE);
4698 * Check again in case we missed a preemption opportunity
4699 * between schedule and now.
4701 barrier();
4702 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4704 EXPORT_SYMBOL(preempt_schedule);
4707 * this is the entry point to schedule() from kernel preemption
4708 * off of irq context.
4709 * Note, that this is called and return with irqs disabled. This will
4710 * protect us against recursive calling from irq.
4712 asmlinkage void __sched preempt_schedule_irq(void)
4714 struct thread_info *ti = current_thread_info();
4716 /* Catch callers which need to be fixed */
4717 BUG_ON(ti->preempt_count || !irqs_disabled());
4719 do {
4720 add_preempt_count(PREEMPT_ACTIVE);
4721 local_irq_enable();
4722 schedule();
4723 local_irq_disable();
4724 sub_preempt_count(PREEMPT_ACTIVE);
4727 * Check again in case we missed a preemption opportunity
4728 * between schedule and now.
4730 barrier();
4731 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4734 #endif /* CONFIG_PREEMPT */
4736 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4737 void *key)
4739 return try_to_wake_up(curr->private, mode, sync);
4741 EXPORT_SYMBOL(default_wake_function);
4744 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4745 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4746 * number) then we wake all the non-exclusive tasks and one exclusive task.
4748 * There are circumstances in which we can try to wake a task which has already
4749 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4750 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4752 void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4753 int nr_exclusive, int sync, void *key)
4755 wait_queue_t *curr, *next;
4757 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4758 unsigned flags = curr->flags;
4760 if (curr->func(curr, mode, sync, key) &&
4761 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4762 break;
4767 * __wake_up - wake up threads blocked on a waitqueue.
4768 * @q: the waitqueue
4769 * @mode: which threads
4770 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4771 * @key: is directly passed to the wakeup function
4773 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4774 int nr_exclusive, void *key)
4776 unsigned long flags;
4778 spin_lock_irqsave(&q->lock, flags);
4779 __wake_up_common(q, mode, nr_exclusive, 0, key);
4780 spin_unlock_irqrestore(&q->lock, flags);
4782 EXPORT_SYMBOL(__wake_up);
4785 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4787 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4789 __wake_up_common(q, mode, 1, 0, NULL);
4793 * __wake_up_sync - wake up threads blocked on a waitqueue.
4794 * @q: the waitqueue
4795 * @mode: which threads
4796 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4798 * The sync wakeup differs that the waker knows that it will schedule
4799 * away soon, so while the target thread will be woken up, it will not
4800 * be migrated to another CPU - ie. the two threads are 'synchronized'
4801 * with each other. This can prevent needless bouncing between CPUs.
4803 * On UP it can prevent extra preemption.
4805 void
4806 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4808 unsigned long flags;
4809 int sync = 1;
4811 if (unlikely(!q))
4812 return;
4814 if (unlikely(!nr_exclusive))
4815 sync = 0;
4817 spin_lock_irqsave(&q->lock, flags);
4818 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4819 spin_unlock_irqrestore(&q->lock, flags);
4821 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4824 * complete: - signals a single thread waiting on this completion
4825 * @x: holds the state of this particular completion
4827 * This will wake up a single thread waiting on this completion. Threads will be
4828 * awakened in the same order in which they were queued.
4830 * See also complete_all(), wait_for_completion() and related routines.
4832 void complete(struct completion *x)
4834 unsigned long flags;
4836 spin_lock_irqsave(&x->wait.lock, flags);
4837 x->done++;
4838 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4839 spin_unlock_irqrestore(&x->wait.lock, flags);
4841 EXPORT_SYMBOL(complete);
4844 * complete_all: - signals all threads waiting on this completion
4845 * @x: holds the state of this particular completion
4847 * This will wake up all threads waiting on this particular completion event.
4849 void complete_all(struct completion *x)
4851 unsigned long flags;
4853 spin_lock_irqsave(&x->wait.lock, flags);
4854 x->done += UINT_MAX/2;
4855 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4856 spin_unlock_irqrestore(&x->wait.lock, flags);
4858 EXPORT_SYMBOL(complete_all);
4860 static inline long __sched
4861 do_wait_for_common(struct completion *x, long timeout, int state)
4863 if (!x->done) {
4864 DECLARE_WAITQUEUE(wait, current);
4866 wait.flags |= WQ_FLAG_EXCLUSIVE;
4867 __add_wait_queue_tail(&x->wait, &wait);
4868 do {
4869 if (signal_pending_state(state, current)) {
4870 timeout = -ERESTARTSYS;
4871 break;
4873 __set_current_state(state);
4874 spin_unlock_irq(&x->wait.lock);
4875 timeout = schedule_timeout(timeout);
4876 spin_lock_irq(&x->wait.lock);
4877 } while (!x->done && timeout);
4878 __remove_wait_queue(&x->wait, &wait);
4879 if (!x->done)
4880 return timeout;
4882 x->done--;
4883 return timeout ?: 1;
4886 static long __sched
4887 wait_for_common(struct completion *x, long timeout, int state)
4889 might_sleep();
4891 spin_lock_irq(&x->wait.lock);
4892 timeout = do_wait_for_common(x, timeout, state);
4893 spin_unlock_irq(&x->wait.lock);
4894 return timeout;
4898 * wait_for_completion: - waits for completion of a task
4899 * @x: holds the state of this particular completion
4901 * This waits to be signaled for completion of a specific task. It is NOT
4902 * interruptible and there is no timeout.
4904 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4905 * and interrupt capability. Also see complete().
4907 void __sched wait_for_completion(struct completion *x)
4909 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4911 EXPORT_SYMBOL(wait_for_completion);
4914 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4915 * @x: holds the state of this particular completion
4916 * @timeout: timeout value in jiffies
4918 * This waits for either a completion of a specific task to be signaled or for a
4919 * specified timeout to expire. The timeout is in jiffies. It is not
4920 * interruptible.
4922 unsigned long __sched
4923 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4925 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4927 EXPORT_SYMBOL(wait_for_completion_timeout);
4930 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4931 * @x: holds the state of this particular completion
4933 * This waits for completion of a specific task to be signaled. It is
4934 * interruptible.
4936 int __sched wait_for_completion_interruptible(struct completion *x)
4938 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4939 if (t == -ERESTARTSYS)
4940 return t;
4941 return 0;
4943 EXPORT_SYMBOL(wait_for_completion_interruptible);
4946 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4947 * @x: holds the state of this particular completion
4948 * @timeout: timeout value in jiffies
4950 * This waits for either a completion of a specific task to be signaled or for a
4951 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4953 unsigned long __sched
4954 wait_for_completion_interruptible_timeout(struct completion *x,
4955 unsigned long timeout)
4957 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4959 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4962 * wait_for_completion_killable: - waits for completion of a task (killable)
4963 * @x: holds the state of this particular completion
4965 * This waits to be signaled for completion of a specific task. It can be
4966 * interrupted by a kill signal.
4968 int __sched wait_for_completion_killable(struct completion *x)
4970 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4971 if (t == -ERESTARTSYS)
4972 return t;
4973 return 0;
4975 EXPORT_SYMBOL(wait_for_completion_killable);
4978 * try_wait_for_completion - try to decrement a completion without blocking
4979 * @x: completion structure
4981 * Returns: 0 if a decrement cannot be done without blocking
4982 * 1 if a decrement succeeded.
4984 * If a completion is being used as a counting completion,
4985 * attempt to decrement the counter without blocking. This
4986 * enables us to avoid waiting if the resource the completion
4987 * is protecting is not available.
4989 bool try_wait_for_completion(struct completion *x)
4991 int ret = 1;
4993 spin_lock_irq(&x->wait.lock);
4994 if (!x->done)
4995 ret = 0;
4996 else
4997 x->done--;
4998 spin_unlock_irq(&x->wait.lock);
4999 return ret;
5001 EXPORT_SYMBOL(try_wait_for_completion);
5004 * completion_done - Test to see if a completion has any waiters
5005 * @x: completion structure
5007 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5008 * 1 if there are no waiters.
5011 bool completion_done(struct completion *x)
5013 int ret = 1;
5015 spin_lock_irq(&x->wait.lock);
5016 if (!x->done)
5017 ret = 0;
5018 spin_unlock_irq(&x->wait.lock);
5019 return ret;
5021 EXPORT_SYMBOL(completion_done);
5023 static long __sched
5024 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5026 unsigned long flags;
5027 wait_queue_t wait;
5029 init_waitqueue_entry(&wait, current);
5031 __set_current_state(state);
5033 spin_lock_irqsave(&q->lock, flags);
5034 __add_wait_queue(q, &wait);
5035 spin_unlock(&q->lock);
5036 timeout = schedule_timeout(timeout);
5037 spin_lock_irq(&q->lock);
5038 __remove_wait_queue(q, &wait);
5039 spin_unlock_irqrestore(&q->lock, flags);
5041 return timeout;
5044 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5046 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5048 EXPORT_SYMBOL(interruptible_sleep_on);
5050 long __sched
5051 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5053 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5055 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5057 void __sched sleep_on(wait_queue_head_t *q)
5059 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5061 EXPORT_SYMBOL(sleep_on);
5063 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5065 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5067 EXPORT_SYMBOL(sleep_on_timeout);
5069 #ifdef CONFIG_RT_MUTEXES
5072 * rt_mutex_setprio - set the current priority of a task
5073 * @p: task
5074 * @prio: prio value (kernel-internal form)
5076 * This function changes the 'effective' priority of a task. It does
5077 * not touch ->normal_prio like __setscheduler().
5079 * Used by the rt_mutex code to implement priority inheritance logic.
5081 void rt_mutex_setprio(struct task_struct *p, int prio)
5083 unsigned long flags;
5084 int oldprio, on_rq, running;
5085 struct rq *rq;
5086 const struct sched_class *prev_class = p->sched_class;
5088 BUG_ON(prio < 0 || prio > MAX_PRIO);
5090 rq = task_rq_lock(p, &flags);
5091 update_rq_clock(rq);
5093 oldprio = p->prio;
5094 on_rq = p->se.on_rq;
5095 running = task_current(rq, p);
5096 if (on_rq)
5097 dequeue_task(rq, p, 0);
5098 if (running)
5099 p->sched_class->put_prev_task(rq, p);
5101 if (rt_prio(prio))
5102 p->sched_class = &rt_sched_class;
5103 else
5104 p->sched_class = &fair_sched_class;
5106 p->prio = prio;
5108 if (running)
5109 p->sched_class->set_curr_task(rq);
5110 if (on_rq) {
5111 enqueue_task(rq, p, 0);
5113 check_class_changed(rq, p, prev_class, oldprio, running);
5115 task_rq_unlock(rq, &flags);
5118 #endif
5120 void set_user_nice(struct task_struct *p, long nice)
5122 int old_prio, delta, on_rq;
5123 unsigned long flags;
5124 struct rq *rq;
5126 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5127 return;
5129 * We have to be careful, if called from sys_setpriority(),
5130 * the task might be in the middle of scheduling on another CPU.
5132 rq = task_rq_lock(p, &flags);
5133 update_rq_clock(rq);
5135 * The RT priorities are set via sched_setscheduler(), but we still
5136 * allow the 'normal' nice value to be set - but as expected
5137 * it wont have any effect on scheduling until the task is
5138 * SCHED_FIFO/SCHED_RR:
5140 if (task_has_rt_policy(p)) {
5141 p->static_prio = NICE_TO_PRIO(nice);
5142 goto out_unlock;
5144 on_rq = p->se.on_rq;
5145 if (on_rq)
5146 dequeue_task(rq, p, 0);
5148 p->static_prio = NICE_TO_PRIO(nice);
5149 set_load_weight(p);
5150 old_prio = p->prio;
5151 p->prio = effective_prio(p);
5152 delta = p->prio - old_prio;
5154 if (on_rq) {
5155 enqueue_task(rq, p, 0);
5157 * If the task increased its priority or is running and
5158 * lowered its priority, then reschedule its CPU:
5160 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5161 resched_task(rq->curr);
5163 out_unlock:
5164 task_rq_unlock(rq, &flags);
5166 EXPORT_SYMBOL(set_user_nice);
5169 * can_nice - check if a task can reduce its nice value
5170 * @p: task
5171 * @nice: nice value
5173 int can_nice(const struct task_struct *p, const int nice)
5175 /* convert nice value [19,-20] to rlimit style value [1,40] */
5176 int nice_rlim = 20 - nice;
5178 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5179 capable(CAP_SYS_NICE));
5182 #ifdef __ARCH_WANT_SYS_NICE
5185 * sys_nice - change the priority of the current process.
5186 * @increment: priority increment
5188 * sys_setpriority is a more generic, but much slower function that
5189 * does similar things.
5191 SYSCALL_DEFINE1(nice, int, increment)
5193 long nice, retval;
5196 * Setpriority might change our priority at the same moment.
5197 * We don't have to worry. Conceptually one call occurs first
5198 * and we have a single winner.
5200 if (increment < -40)
5201 increment = -40;
5202 if (increment > 40)
5203 increment = 40;
5205 nice = PRIO_TO_NICE(current->static_prio) + increment;
5206 if (nice < -20)
5207 nice = -20;
5208 if (nice > 19)
5209 nice = 19;
5211 if (increment < 0 && !can_nice(current, nice))
5212 return -EPERM;
5214 retval = security_task_setnice(current, nice);
5215 if (retval)
5216 return retval;
5218 set_user_nice(current, nice);
5219 return 0;
5222 #endif
5225 * task_prio - return the priority value of a given task.
5226 * @p: the task in question.
5228 * This is the priority value as seen by users in /proc.
5229 * RT tasks are offset by -200. Normal tasks are centered
5230 * around 0, value goes from -16 to +15.
5232 int task_prio(const struct task_struct *p)
5234 return p->prio - MAX_RT_PRIO;
5238 * task_nice - return the nice value of a given task.
5239 * @p: the task in question.
5241 int task_nice(const struct task_struct *p)
5243 return TASK_NICE(p);
5245 EXPORT_SYMBOL(task_nice);
5248 * idle_cpu - is a given cpu idle currently?
5249 * @cpu: the processor in question.
5251 int idle_cpu(int cpu)
5253 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5257 * idle_task - return the idle task for a given cpu.
5258 * @cpu: the processor in question.
5260 struct task_struct *idle_task(int cpu)
5262 return cpu_rq(cpu)->idle;
5266 * find_process_by_pid - find a process with a matching PID value.
5267 * @pid: the pid in question.
5269 static struct task_struct *find_process_by_pid(pid_t pid)
5271 return pid ? find_task_by_vpid(pid) : current;
5274 /* Actually do priority change: must hold rq lock. */
5275 static void
5276 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5278 BUG_ON(p->se.on_rq);
5280 p->policy = policy;
5281 switch (p->policy) {
5282 case SCHED_NORMAL:
5283 case SCHED_BATCH:
5284 case SCHED_IDLE:
5285 p->sched_class = &fair_sched_class;
5286 break;
5287 case SCHED_FIFO:
5288 case SCHED_RR:
5289 p->sched_class = &rt_sched_class;
5290 break;
5293 p->rt_priority = prio;
5294 p->normal_prio = normal_prio(p);
5295 /* we are holding p->pi_lock already */
5296 p->prio = rt_mutex_getprio(p);
5297 set_load_weight(p);
5301 * check the target process has a UID that matches the current process's
5303 static bool check_same_owner(struct task_struct *p)
5305 const struct cred *cred = current_cred(), *pcred;
5306 bool match;
5308 rcu_read_lock();
5309 pcred = __task_cred(p);
5310 match = (cred->euid == pcred->euid ||
5311 cred->euid == pcred->uid);
5312 rcu_read_unlock();
5313 return match;
5316 static int __sched_setscheduler(struct task_struct *p, int policy,
5317 struct sched_param *param, bool user)
5319 int retval, oldprio, oldpolicy = -1, on_rq, running;
5320 unsigned long flags;
5321 const struct sched_class *prev_class = p->sched_class;
5322 struct rq *rq;
5324 /* may grab non-irq protected spin_locks */
5325 BUG_ON(in_interrupt());
5326 recheck:
5327 /* double check policy once rq lock held */
5328 if (policy < 0)
5329 policy = oldpolicy = p->policy;
5330 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5331 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5332 policy != SCHED_IDLE)
5333 return -EINVAL;
5335 * Valid priorities for SCHED_FIFO and SCHED_RR are
5336 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5337 * SCHED_BATCH and SCHED_IDLE is 0.
5339 if (param->sched_priority < 0 ||
5340 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5341 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5342 return -EINVAL;
5343 if (rt_policy(policy) != (param->sched_priority != 0))
5344 return -EINVAL;
5347 * Allow unprivileged RT tasks to decrease priority:
5349 if (user && !capable(CAP_SYS_NICE)) {
5350 if (rt_policy(policy)) {
5351 unsigned long rlim_rtprio;
5353 if (!lock_task_sighand(p, &flags))
5354 return -ESRCH;
5355 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5356 unlock_task_sighand(p, &flags);
5358 /* can't set/change the rt policy */
5359 if (policy != p->policy && !rlim_rtprio)
5360 return -EPERM;
5362 /* can't increase priority */
5363 if (param->sched_priority > p->rt_priority &&
5364 param->sched_priority > rlim_rtprio)
5365 return -EPERM;
5368 * Like positive nice levels, dont allow tasks to
5369 * move out of SCHED_IDLE either:
5371 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5372 return -EPERM;
5374 /* can't change other user's priorities */
5375 if (!check_same_owner(p))
5376 return -EPERM;
5379 if (user) {
5380 #ifdef CONFIG_RT_GROUP_SCHED
5382 * Do not allow realtime tasks into groups that have no runtime
5383 * assigned.
5385 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5386 task_group(p)->rt_bandwidth.rt_runtime == 0)
5387 return -EPERM;
5388 #endif
5390 retval = security_task_setscheduler(p, policy, param);
5391 if (retval)
5392 return retval;
5396 * make sure no PI-waiters arrive (or leave) while we are
5397 * changing the priority of the task:
5399 spin_lock_irqsave(&p->pi_lock, flags);
5401 * To be able to change p->policy safely, the apropriate
5402 * runqueue lock must be held.
5404 rq = __task_rq_lock(p);
5405 /* recheck policy now with rq lock held */
5406 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5407 policy = oldpolicy = -1;
5408 __task_rq_unlock(rq);
5409 spin_unlock_irqrestore(&p->pi_lock, flags);
5410 goto recheck;
5412 update_rq_clock(rq);
5413 on_rq = p->se.on_rq;
5414 running = task_current(rq, p);
5415 if (on_rq)
5416 deactivate_task(rq, p, 0);
5417 if (running)
5418 p->sched_class->put_prev_task(rq, p);
5420 oldprio = p->prio;
5421 __setscheduler(rq, p, policy, param->sched_priority);
5423 if (running)
5424 p->sched_class->set_curr_task(rq);
5425 if (on_rq) {
5426 activate_task(rq, p, 0);
5428 check_class_changed(rq, p, prev_class, oldprio, running);
5430 __task_rq_unlock(rq);
5431 spin_unlock_irqrestore(&p->pi_lock, flags);
5433 rt_mutex_adjust_pi(p);
5435 return 0;
5439 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5440 * @p: the task in question.
5441 * @policy: new policy.
5442 * @param: structure containing the new RT priority.
5444 * NOTE that the task may be already dead.
5446 int sched_setscheduler(struct task_struct *p, int policy,
5447 struct sched_param *param)
5449 return __sched_setscheduler(p, policy, param, true);
5451 EXPORT_SYMBOL_GPL(sched_setscheduler);
5454 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5455 * @p: the task in question.
5456 * @policy: new policy.
5457 * @param: structure containing the new RT priority.
5459 * Just like sched_setscheduler, only don't bother checking if the
5460 * current context has permission. For example, this is needed in
5461 * stop_machine(): we create temporary high priority worker threads,
5462 * but our caller might not have that capability.
5464 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5465 struct sched_param *param)
5467 return __sched_setscheduler(p, policy, param, false);
5470 static int
5471 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5473 struct sched_param lparam;
5474 struct task_struct *p;
5475 int retval;
5477 if (!param || pid < 0)
5478 return -EINVAL;
5479 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5480 return -EFAULT;
5482 rcu_read_lock();
5483 retval = -ESRCH;
5484 p = find_process_by_pid(pid);
5485 if (p != NULL)
5486 retval = sched_setscheduler(p, policy, &lparam);
5487 rcu_read_unlock();
5489 return retval;
5493 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5494 * @pid: the pid in question.
5495 * @policy: new policy.
5496 * @param: structure containing the new RT priority.
5498 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5499 struct sched_param __user *, param)
5501 /* negative values for policy are not valid */
5502 if (policy < 0)
5503 return -EINVAL;
5505 return do_sched_setscheduler(pid, policy, param);
5509 * sys_sched_setparam - set/change the RT priority of a thread
5510 * @pid: the pid in question.
5511 * @param: structure containing the new RT priority.
5513 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5515 return do_sched_setscheduler(pid, -1, param);
5519 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5520 * @pid: the pid in question.
5522 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5524 struct task_struct *p;
5525 int retval;
5527 if (pid < 0)
5528 return -EINVAL;
5530 retval = -ESRCH;
5531 read_lock(&tasklist_lock);
5532 p = find_process_by_pid(pid);
5533 if (p) {
5534 retval = security_task_getscheduler(p);
5535 if (!retval)
5536 retval = p->policy;
5538 read_unlock(&tasklist_lock);
5539 return retval;
5543 * sys_sched_getscheduler - get the RT priority of a thread
5544 * @pid: the pid in question.
5545 * @param: structure containing the RT priority.
5547 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5549 struct sched_param lp;
5550 struct task_struct *p;
5551 int retval;
5553 if (!param || pid < 0)
5554 return -EINVAL;
5556 read_lock(&tasklist_lock);
5557 p = find_process_by_pid(pid);
5558 retval = -ESRCH;
5559 if (!p)
5560 goto out_unlock;
5562 retval = security_task_getscheduler(p);
5563 if (retval)
5564 goto out_unlock;
5566 lp.sched_priority = p->rt_priority;
5567 read_unlock(&tasklist_lock);
5570 * This one might sleep, we cannot do it with a spinlock held ...
5572 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5574 return retval;
5576 out_unlock:
5577 read_unlock(&tasklist_lock);
5578 return retval;
5581 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5583 cpumask_var_t cpus_allowed, new_mask;
5584 struct task_struct *p;
5585 int retval;
5587 get_online_cpus();
5588 read_lock(&tasklist_lock);
5590 p = find_process_by_pid(pid);
5591 if (!p) {
5592 read_unlock(&tasklist_lock);
5593 put_online_cpus();
5594 return -ESRCH;
5598 * It is not safe to call set_cpus_allowed with the
5599 * tasklist_lock held. We will bump the task_struct's
5600 * usage count and then drop tasklist_lock.
5602 get_task_struct(p);
5603 read_unlock(&tasklist_lock);
5605 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5606 retval = -ENOMEM;
5607 goto out_put_task;
5609 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5610 retval = -ENOMEM;
5611 goto out_free_cpus_allowed;
5613 retval = -EPERM;
5614 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5615 goto out_unlock;
5617 retval = security_task_setscheduler(p, 0, NULL);
5618 if (retval)
5619 goto out_unlock;
5621 cpuset_cpus_allowed(p, cpus_allowed);
5622 cpumask_and(new_mask, in_mask, cpus_allowed);
5623 again:
5624 retval = set_cpus_allowed_ptr(p, new_mask);
5626 if (!retval) {
5627 cpuset_cpus_allowed(p, cpus_allowed);
5628 if (!cpumask_subset(new_mask, cpus_allowed)) {
5630 * We must have raced with a concurrent cpuset
5631 * update. Just reset the cpus_allowed to the
5632 * cpuset's cpus_allowed
5634 cpumask_copy(new_mask, cpus_allowed);
5635 goto again;
5638 out_unlock:
5639 free_cpumask_var(new_mask);
5640 out_free_cpus_allowed:
5641 free_cpumask_var(cpus_allowed);
5642 out_put_task:
5643 put_task_struct(p);
5644 put_online_cpus();
5645 return retval;
5648 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5649 struct cpumask *new_mask)
5651 if (len < cpumask_size())
5652 cpumask_clear(new_mask);
5653 else if (len > cpumask_size())
5654 len = cpumask_size();
5656 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5660 * sys_sched_setaffinity - set the cpu affinity of a process
5661 * @pid: pid of the process
5662 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5663 * @user_mask_ptr: user-space pointer to the new cpu mask
5665 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5666 unsigned long __user *, user_mask_ptr)
5668 cpumask_var_t new_mask;
5669 int retval;
5671 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5672 return -ENOMEM;
5674 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5675 if (retval == 0)
5676 retval = sched_setaffinity(pid, new_mask);
5677 free_cpumask_var(new_mask);
5678 return retval;
5681 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5683 struct task_struct *p;
5684 int retval;
5686 get_online_cpus();
5687 read_lock(&tasklist_lock);
5689 retval = -ESRCH;
5690 p = find_process_by_pid(pid);
5691 if (!p)
5692 goto out_unlock;
5694 retval = security_task_getscheduler(p);
5695 if (retval)
5696 goto out_unlock;
5698 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5700 out_unlock:
5701 read_unlock(&tasklist_lock);
5702 put_online_cpus();
5704 return retval;
5708 * sys_sched_getaffinity - get the cpu affinity of a process
5709 * @pid: pid of the process
5710 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5711 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5713 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5714 unsigned long __user *, user_mask_ptr)
5716 int ret;
5717 cpumask_var_t mask;
5719 if (len < cpumask_size())
5720 return -EINVAL;
5722 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5723 return -ENOMEM;
5725 ret = sched_getaffinity(pid, mask);
5726 if (ret == 0) {
5727 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
5728 ret = -EFAULT;
5729 else
5730 ret = cpumask_size();
5732 free_cpumask_var(mask);
5734 return ret;
5738 * sys_sched_yield - yield the current processor to other threads.
5740 * This function yields the current CPU to other tasks. If there are no
5741 * other threads running on this CPU then this function will return.
5743 SYSCALL_DEFINE0(sched_yield)
5745 struct rq *rq = this_rq_lock();
5747 schedstat_inc(rq, yld_count);
5748 current->sched_class->yield_task(rq);
5751 * Since we are going to call schedule() anyway, there's
5752 * no need to preempt or enable interrupts:
5754 __release(rq->lock);
5755 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5756 _raw_spin_unlock(&rq->lock);
5757 preempt_enable_no_resched();
5759 schedule();
5761 return 0;
5764 static void __cond_resched(void)
5766 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5767 __might_sleep(__FILE__, __LINE__);
5768 #endif
5770 * The BKS might be reacquired before we have dropped
5771 * PREEMPT_ACTIVE, which could trigger a second
5772 * cond_resched() call.
5774 do {
5775 add_preempt_count(PREEMPT_ACTIVE);
5776 schedule();
5777 sub_preempt_count(PREEMPT_ACTIVE);
5778 } while (need_resched());
5781 int __sched _cond_resched(void)
5783 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5784 system_state == SYSTEM_RUNNING) {
5785 __cond_resched();
5786 return 1;
5788 return 0;
5790 EXPORT_SYMBOL(_cond_resched);
5793 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5794 * call schedule, and on return reacquire the lock.
5796 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5797 * operations here to prevent schedule() from being called twice (once via
5798 * spin_unlock(), once by hand).
5800 int cond_resched_lock(spinlock_t *lock)
5802 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5803 int ret = 0;
5805 if (spin_needbreak(lock) || resched) {
5806 spin_unlock(lock);
5807 if (resched && need_resched())
5808 __cond_resched();
5809 else
5810 cpu_relax();
5811 ret = 1;
5812 spin_lock(lock);
5814 return ret;
5816 EXPORT_SYMBOL(cond_resched_lock);
5818 int __sched cond_resched_softirq(void)
5820 BUG_ON(!in_softirq());
5822 if (need_resched() && system_state == SYSTEM_RUNNING) {
5823 local_bh_enable();
5824 __cond_resched();
5825 local_bh_disable();
5826 return 1;
5828 return 0;
5830 EXPORT_SYMBOL(cond_resched_softirq);
5833 * yield - yield the current processor to other threads.
5835 * This is a shortcut for kernel-space yielding - it marks the
5836 * thread runnable and calls sys_sched_yield().
5838 void __sched yield(void)
5840 set_current_state(TASK_RUNNING);
5841 sys_sched_yield();
5843 EXPORT_SYMBOL(yield);
5846 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5847 * that process accounting knows that this is a task in IO wait state.
5849 * But don't do that if it is a deliberate, throttling IO wait (this task
5850 * has set its backing_dev_info: the queue against which it should throttle)
5852 void __sched io_schedule(void)
5854 struct rq *rq = &__raw_get_cpu_var(runqueues);
5856 delayacct_blkio_start();
5857 atomic_inc(&rq->nr_iowait);
5858 schedule();
5859 atomic_dec(&rq->nr_iowait);
5860 delayacct_blkio_end();
5862 EXPORT_SYMBOL(io_schedule);
5864 long __sched io_schedule_timeout(long timeout)
5866 struct rq *rq = &__raw_get_cpu_var(runqueues);
5867 long ret;
5869 delayacct_blkio_start();
5870 atomic_inc(&rq->nr_iowait);
5871 ret = schedule_timeout(timeout);
5872 atomic_dec(&rq->nr_iowait);
5873 delayacct_blkio_end();
5874 return ret;
5878 * sys_sched_get_priority_max - return maximum RT priority.
5879 * @policy: scheduling class.
5881 * this syscall returns the maximum rt_priority that can be used
5882 * by a given scheduling class.
5884 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5886 int ret = -EINVAL;
5888 switch (policy) {
5889 case SCHED_FIFO:
5890 case SCHED_RR:
5891 ret = MAX_USER_RT_PRIO-1;
5892 break;
5893 case SCHED_NORMAL:
5894 case SCHED_BATCH:
5895 case SCHED_IDLE:
5896 ret = 0;
5897 break;
5899 return ret;
5903 * sys_sched_get_priority_min - return minimum RT priority.
5904 * @policy: scheduling class.
5906 * this syscall returns the minimum rt_priority that can be used
5907 * by a given scheduling class.
5909 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5911 int ret = -EINVAL;
5913 switch (policy) {
5914 case SCHED_FIFO:
5915 case SCHED_RR:
5916 ret = 1;
5917 break;
5918 case SCHED_NORMAL:
5919 case SCHED_BATCH:
5920 case SCHED_IDLE:
5921 ret = 0;
5923 return ret;
5927 * sys_sched_rr_get_interval - return the default timeslice of a process.
5928 * @pid: pid of the process.
5929 * @interval: userspace pointer to the timeslice value.
5931 * this syscall writes the default timeslice value of a given process
5932 * into the user-space timespec buffer. A value of '0' means infinity.
5934 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5935 struct timespec __user *, interval)
5937 struct task_struct *p;
5938 unsigned int time_slice;
5939 int retval;
5940 struct timespec t;
5942 if (pid < 0)
5943 return -EINVAL;
5945 retval = -ESRCH;
5946 read_lock(&tasklist_lock);
5947 p = find_process_by_pid(pid);
5948 if (!p)
5949 goto out_unlock;
5951 retval = security_task_getscheduler(p);
5952 if (retval)
5953 goto out_unlock;
5956 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5957 * tasks that are on an otherwise idle runqueue:
5959 time_slice = 0;
5960 if (p->policy == SCHED_RR) {
5961 time_slice = DEF_TIMESLICE;
5962 } else if (p->policy != SCHED_FIFO) {
5963 struct sched_entity *se = &p->se;
5964 unsigned long flags;
5965 struct rq *rq;
5967 rq = task_rq_lock(p, &flags);
5968 if (rq->cfs.load.weight)
5969 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5970 task_rq_unlock(rq, &flags);
5972 read_unlock(&tasklist_lock);
5973 jiffies_to_timespec(time_slice, &t);
5974 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5975 return retval;
5977 out_unlock:
5978 read_unlock(&tasklist_lock);
5979 return retval;
5982 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5984 void sched_show_task(struct task_struct *p)
5986 unsigned long free = 0;
5987 unsigned state;
5989 state = p->state ? __ffs(p->state) + 1 : 0;
5990 printk(KERN_INFO "%-13.13s %c", p->comm,
5991 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5992 #if BITS_PER_LONG == 32
5993 if (state == TASK_RUNNING)
5994 printk(KERN_CONT " running ");
5995 else
5996 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5997 #else
5998 if (state == TASK_RUNNING)
5999 printk(KERN_CONT " running task ");
6000 else
6001 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6002 #endif
6003 #ifdef CONFIG_DEBUG_STACK_USAGE
6005 unsigned long *n = end_of_stack(p);
6006 while (!*n)
6007 n++;
6008 free = (unsigned long)n - (unsigned long)end_of_stack(p);
6010 #endif
6011 printk(KERN_CONT "%5lu %5d %6d\n", free,
6012 task_pid_nr(p), task_pid_nr(p->real_parent));
6014 show_stack(p, NULL);
6017 void show_state_filter(unsigned long state_filter)
6019 struct task_struct *g, *p;
6021 #if BITS_PER_LONG == 32
6022 printk(KERN_INFO
6023 " task PC stack pid father\n");
6024 #else
6025 printk(KERN_INFO
6026 " task PC stack pid father\n");
6027 #endif
6028 read_lock(&tasklist_lock);
6029 do_each_thread(g, p) {
6031 * reset the NMI-timeout, listing all files on a slow
6032 * console might take alot of time:
6034 touch_nmi_watchdog();
6035 if (!state_filter || (p->state & state_filter))
6036 sched_show_task(p);
6037 } while_each_thread(g, p);
6039 touch_all_softlockup_watchdogs();
6041 #ifdef CONFIG_SCHED_DEBUG
6042 sysrq_sched_debug_show();
6043 #endif
6044 read_unlock(&tasklist_lock);
6046 * Only show locks if all tasks are dumped:
6048 if (state_filter == -1)
6049 debug_show_all_locks();
6052 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6054 idle->sched_class = &idle_sched_class;
6058 * init_idle - set up an idle thread for a given CPU
6059 * @idle: task in question
6060 * @cpu: cpu the idle task belongs to
6062 * NOTE: this function does not set the idle thread's NEED_RESCHED
6063 * flag, to make booting more robust.
6065 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6067 struct rq *rq = cpu_rq(cpu);
6068 unsigned long flags;
6070 spin_lock_irqsave(&rq->lock, flags);
6072 __sched_fork(idle);
6073 idle->se.exec_start = sched_clock();
6075 idle->prio = idle->normal_prio = MAX_PRIO;
6076 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6077 __set_task_cpu(idle, cpu);
6079 rq->curr = rq->idle = idle;
6080 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6081 idle->oncpu = 1;
6082 #endif
6083 spin_unlock_irqrestore(&rq->lock, flags);
6085 /* Set the preempt count _outside_ the spinlocks! */
6086 #if defined(CONFIG_PREEMPT)
6087 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6088 #else
6089 task_thread_info(idle)->preempt_count = 0;
6090 #endif
6092 * The idle tasks have their own, simple scheduling class:
6094 idle->sched_class = &idle_sched_class;
6095 ftrace_graph_init_task(idle);
6099 * In a system that switches off the HZ timer nohz_cpu_mask
6100 * indicates which cpus entered this state. This is used
6101 * in the rcu update to wait only for active cpus. For system
6102 * which do not switch off the HZ timer nohz_cpu_mask should
6103 * always be CPU_BITS_NONE.
6105 cpumask_var_t nohz_cpu_mask;
6108 * Increase the granularity value when there are more CPUs,
6109 * because with more CPUs the 'effective latency' as visible
6110 * to users decreases. But the relationship is not linear,
6111 * so pick a second-best guess by going with the log2 of the
6112 * number of CPUs.
6114 * This idea comes from the SD scheduler of Con Kolivas:
6116 static inline void sched_init_granularity(void)
6118 unsigned int factor = 1 + ilog2(num_online_cpus());
6119 const unsigned long limit = 200000000;
6121 sysctl_sched_min_granularity *= factor;
6122 if (sysctl_sched_min_granularity > limit)
6123 sysctl_sched_min_granularity = limit;
6125 sysctl_sched_latency *= factor;
6126 if (sysctl_sched_latency > limit)
6127 sysctl_sched_latency = limit;
6129 sysctl_sched_wakeup_granularity *= factor;
6131 sysctl_sched_shares_ratelimit *= factor;
6134 #ifdef CONFIG_SMP
6136 * This is how migration works:
6138 * 1) we queue a struct migration_req structure in the source CPU's
6139 * runqueue and wake up that CPU's migration thread.
6140 * 2) we down() the locked semaphore => thread blocks.
6141 * 3) migration thread wakes up (implicitly it forces the migrated
6142 * thread off the CPU)
6143 * 4) it gets the migration request and checks whether the migrated
6144 * task is still in the wrong runqueue.
6145 * 5) if it's in the wrong runqueue then the migration thread removes
6146 * it and puts it into the right queue.
6147 * 6) migration thread up()s the semaphore.
6148 * 7) we wake up and the migration is done.
6152 * Change a given task's CPU affinity. Migrate the thread to a
6153 * proper CPU and schedule it away if the CPU it's executing on
6154 * is removed from the allowed bitmask.
6156 * NOTE: the caller must have a valid reference to the task, the
6157 * task must not exit() & deallocate itself prematurely. The
6158 * call is not atomic; no spinlocks may be held.
6160 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6162 struct migration_req req;
6163 unsigned long flags;
6164 struct rq *rq;
6165 int ret = 0;
6167 rq = task_rq_lock(p, &flags);
6168 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6169 ret = -EINVAL;
6170 goto out;
6173 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6174 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6175 ret = -EINVAL;
6176 goto out;
6179 if (p->sched_class->set_cpus_allowed)
6180 p->sched_class->set_cpus_allowed(p, new_mask);
6181 else {
6182 cpumask_copy(&p->cpus_allowed, new_mask);
6183 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6186 /* Can the task run on the task's current CPU? If so, we're done */
6187 if (cpumask_test_cpu(task_cpu(p), new_mask))
6188 goto out;
6190 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6191 /* Need help from migration thread: drop lock and wait. */
6192 task_rq_unlock(rq, &flags);
6193 wake_up_process(rq->migration_thread);
6194 wait_for_completion(&req.done);
6195 tlb_migrate_finish(p->mm);
6196 return 0;
6198 out:
6199 task_rq_unlock(rq, &flags);
6201 return ret;
6203 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6206 * Move (not current) task off this cpu, onto dest cpu. We're doing
6207 * this because either it can't run here any more (set_cpus_allowed()
6208 * away from this CPU, or CPU going down), or because we're
6209 * attempting to rebalance this task on exec (sched_exec).
6211 * So we race with normal scheduler movements, but that's OK, as long
6212 * as the task is no longer on this CPU.
6214 * Returns non-zero if task was successfully migrated.
6216 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6218 struct rq *rq_dest, *rq_src;
6219 int ret = 0, on_rq;
6221 if (unlikely(!cpu_active(dest_cpu)))
6222 return ret;
6224 rq_src = cpu_rq(src_cpu);
6225 rq_dest = cpu_rq(dest_cpu);
6227 double_rq_lock(rq_src, rq_dest);
6228 /* Already moved. */
6229 if (task_cpu(p) != src_cpu)
6230 goto done;
6231 /* Affinity changed (again). */
6232 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6233 goto fail;
6235 on_rq = p->se.on_rq;
6236 if (on_rq)
6237 deactivate_task(rq_src, p, 0);
6239 set_task_cpu(p, dest_cpu);
6240 if (on_rq) {
6241 activate_task(rq_dest, p, 0);
6242 check_preempt_curr(rq_dest, p, 0);
6244 done:
6245 ret = 1;
6246 fail:
6247 double_rq_unlock(rq_src, rq_dest);
6248 return ret;
6252 * migration_thread - this is a highprio system thread that performs
6253 * thread migration by bumping thread off CPU then 'pushing' onto
6254 * another runqueue.
6256 static int migration_thread(void *data)
6258 int cpu = (long)data;
6259 struct rq *rq;
6261 rq = cpu_rq(cpu);
6262 BUG_ON(rq->migration_thread != current);
6264 set_current_state(TASK_INTERRUPTIBLE);
6265 while (!kthread_should_stop()) {
6266 struct migration_req *req;
6267 struct list_head *head;
6269 spin_lock_irq(&rq->lock);
6271 if (cpu_is_offline(cpu)) {
6272 spin_unlock_irq(&rq->lock);
6273 goto wait_to_die;
6276 if (rq->active_balance) {
6277 active_load_balance(rq, cpu);
6278 rq->active_balance = 0;
6281 head = &rq->migration_queue;
6283 if (list_empty(head)) {
6284 spin_unlock_irq(&rq->lock);
6285 schedule();
6286 set_current_state(TASK_INTERRUPTIBLE);
6287 continue;
6289 req = list_entry(head->next, struct migration_req, list);
6290 list_del_init(head->next);
6292 spin_unlock(&rq->lock);
6293 __migrate_task(req->task, cpu, req->dest_cpu);
6294 local_irq_enable();
6296 complete(&req->done);
6298 __set_current_state(TASK_RUNNING);
6299 return 0;
6301 wait_to_die:
6302 /* Wait for kthread_stop */
6303 set_current_state(TASK_INTERRUPTIBLE);
6304 while (!kthread_should_stop()) {
6305 schedule();
6306 set_current_state(TASK_INTERRUPTIBLE);
6308 __set_current_state(TASK_RUNNING);
6309 return 0;
6312 #ifdef CONFIG_HOTPLUG_CPU
6314 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6316 int ret;
6318 local_irq_disable();
6319 ret = __migrate_task(p, src_cpu, dest_cpu);
6320 local_irq_enable();
6321 return ret;
6325 * Figure out where task on dead CPU should go, use force if necessary.
6327 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6329 int dest_cpu;
6330 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
6332 again:
6333 /* Look for allowed, online CPU in same node. */
6334 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
6335 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6336 goto move;
6338 /* Any allowed, online CPU? */
6339 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
6340 if (dest_cpu < nr_cpu_ids)
6341 goto move;
6343 /* No more Mr. Nice Guy. */
6344 if (dest_cpu >= nr_cpu_ids) {
6345 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
6346 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
6349 * Don't tell them about moving exiting tasks or
6350 * kernel threads (both mm NULL), since they never
6351 * leave kernel.
6353 if (p->mm && printk_ratelimit()) {
6354 printk(KERN_INFO "process %d (%s) no "
6355 "longer affine to cpu%d\n",
6356 task_pid_nr(p), p->comm, dead_cpu);
6360 move:
6361 /* It can have affinity changed while we were choosing. */
6362 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
6363 goto again;
6367 * While a dead CPU has no uninterruptible tasks queued at this point,
6368 * it might still have a nonzero ->nr_uninterruptible counter, because
6369 * for performance reasons the counter is not stricly tracking tasks to
6370 * their home CPUs. So we just add the counter to another CPU's counter,
6371 * to keep the global sum constant after CPU-down:
6373 static void migrate_nr_uninterruptible(struct rq *rq_src)
6375 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
6376 unsigned long flags;
6378 local_irq_save(flags);
6379 double_rq_lock(rq_src, rq_dest);
6380 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6381 rq_src->nr_uninterruptible = 0;
6382 double_rq_unlock(rq_src, rq_dest);
6383 local_irq_restore(flags);
6386 /* Run through task list and migrate tasks from the dead cpu. */
6387 static void migrate_live_tasks(int src_cpu)
6389 struct task_struct *p, *t;
6391 read_lock(&tasklist_lock);
6393 do_each_thread(t, p) {
6394 if (p == current)
6395 continue;
6397 if (task_cpu(p) == src_cpu)
6398 move_task_off_dead_cpu(src_cpu, p);
6399 } while_each_thread(t, p);
6401 read_unlock(&tasklist_lock);
6405 * Schedules idle task to be the next runnable task on current CPU.
6406 * It does so by boosting its priority to highest possible.
6407 * Used by CPU offline code.
6409 void sched_idle_next(void)
6411 int this_cpu = smp_processor_id();
6412 struct rq *rq = cpu_rq(this_cpu);
6413 struct task_struct *p = rq->idle;
6414 unsigned long flags;
6416 /* cpu has to be offline */
6417 BUG_ON(cpu_online(this_cpu));
6420 * Strictly not necessary since rest of the CPUs are stopped by now
6421 * and interrupts disabled on the current cpu.
6423 spin_lock_irqsave(&rq->lock, flags);
6425 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6427 update_rq_clock(rq);
6428 activate_task(rq, p, 0);
6430 spin_unlock_irqrestore(&rq->lock, flags);
6434 * Ensures that the idle task is using init_mm right before its cpu goes
6435 * offline.
6437 void idle_task_exit(void)
6439 struct mm_struct *mm = current->active_mm;
6441 BUG_ON(cpu_online(smp_processor_id()));
6443 if (mm != &init_mm)
6444 switch_mm(mm, &init_mm, current);
6445 mmdrop(mm);
6448 /* called under rq->lock with disabled interrupts */
6449 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6451 struct rq *rq = cpu_rq(dead_cpu);
6453 /* Must be exiting, otherwise would be on tasklist. */
6454 BUG_ON(!p->exit_state);
6456 /* Cannot have done final schedule yet: would have vanished. */
6457 BUG_ON(p->state == TASK_DEAD);
6459 get_task_struct(p);
6462 * Drop lock around migration; if someone else moves it,
6463 * that's OK. No task can be added to this CPU, so iteration is
6464 * fine.
6466 spin_unlock_irq(&rq->lock);
6467 move_task_off_dead_cpu(dead_cpu, p);
6468 spin_lock_irq(&rq->lock);
6470 put_task_struct(p);
6473 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6474 static void migrate_dead_tasks(unsigned int dead_cpu)
6476 struct rq *rq = cpu_rq(dead_cpu);
6477 struct task_struct *next;
6479 for ( ; ; ) {
6480 if (!rq->nr_running)
6481 break;
6482 update_rq_clock(rq);
6483 next = pick_next_task(rq, rq->curr);
6484 if (!next)
6485 break;
6486 next->sched_class->put_prev_task(rq, next);
6487 migrate_dead(dead_cpu, next);
6491 #endif /* CONFIG_HOTPLUG_CPU */
6493 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6495 static struct ctl_table sd_ctl_dir[] = {
6497 .procname = "sched_domain",
6498 .mode = 0555,
6500 {0, },
6503 static struct ctl_table sd_ctl_root[] = {
6505 .ctl_name = CTL_KERN,
6506 .procname = "kernel",
6507 .mode = 0555,
6508 .child = sd_ctl_dir,
6510 {0, },
6513 static struct ctl_table *sd_alloc_ctl_entry(int n)
6515 struct ctl_table *entry =
6516 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6518 return entry;
6521 static void sd_free_ctl_entry(struct ctl_table **tablep)
6523 struct ctl_table *entry;
6526 * In the intermediate directories, both the child directory and
6527 * procname are dynamically allocated and could fail but the mode
6528 * will always be set. In the lowest directory the names are
6529 * static strings and all have proc handlers.
6531 for (entry = *tablep; entry->mode; entry++) {
6532 if (entry->child)
6533 sd_free_ctl_entry(&entry->child);
6534 if (entry->proc_handler == NULL)
6535 kfree(entry->procname);
6538 kfree(*tablep);
6539 *tablep = NULL;
6542 static void
6543 set_table_entry(struct ctl_table *entry,
6544 const char *procname, void *data, int maxlen,
6545 mode_t mode, proc_handler *proc_handler)
6547 entry->procname = procname;
6548 entry->data = data;
6549 entry->maxlen = maxlen;
6550 entry->mode = mode;
6551 entry->proc_handler = proc_handler;
6554 static struct ctl_table *
6555 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6557 struct ctl_table *table = sd_alloc_ctl_entry(13);
6559 if (table == NULL)
6560 return NULL;
6562 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6563 sizeof(long), 0644, proc_doulongvec_minmax);
6564 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6565 sizeof(long), 0644, proc_doulongvec_minmax);
6566 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6567 sizeof(int), 0644, proc_dointvec_minmax);
6568 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6569 sizeof(int), 0644, proc_dointvec_minmax);
6570 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6571 sizeof(int), 0644, proc_dointvec_minmax);
6572 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6573 sizeof(int), 0644, proc_dointvec_minmax);
6574 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6575 sizeof(int), 0644, proc_dointvec_minmax);
6576 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6577 sizeof(int), 0644, proc_dointvec_minmax);
6578 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6579 sizeof(int), 0644, proc_dointvec_minmax);
6580 set_table_entry(&table[9], "cache_nice_tries",
6581 &sd->cache_nice_tries,
6582 sizeof(int), 0644, proc_dointvec_minmax);
6583 set_table_entry(&table[10], "flags", &sd->flags,
6584 sizeof(int), 0644, proc_dointvec_minmax);
6585 set_table_entry(&table[11], "name", sd->name,
6586 CORENAME_MAX_SIZE, 0444, proc_dostring);
6587 /* &table[12] is terminator */
6589 return table;
6592 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6594 struct ctl_table *entry, *table;
6595 struct sched_domain *sd;
6596 int domain_num = 0, i;
6597 char buf[32];
6599 for_each_domain(cpu, sd)
6600 domain_num++;
6601 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6602 if (table == NULL)
6603 return NULL;
6605 i = 0;
6606 for_each_domain(cpu, sd) {
6607 snprintf(buf, 32, "domain%d", i);
6608 entry->procname = kstrdup(buf, GFP_KERNEL);
6609 entry->mode = 0555;
6610 entry->child = sd_alloc_ctl_domain_table(sd);
6611 entry++;
6612 i++;
6614 return table;
6617 static struct ctl_table_header *sd_sysctl_header;
6618 static void register_sched_domain_sysctl(void)
6620 int i, cpu_num = num_online_cpus();
6621 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6622 char buf[32];
6624 WARN_ON(sd_ctl_dir[0].child);
6625 sd_ctl_dir[0].child = entry;
6627 if (entry == NULL)
6628 return;
6630 for_each_online_cpu(i) {
6631 snprintf(buf, 32, "cpu%d", i);
6632 entry->procname = kstrdup(buf, GFP_KERNEL);
6633 entry->mode = 0555;
6634 entry->child = sd_alloc_ctl_cpu_table(i);
6635 entry++;
6638 WARN_ON(sd_sysctl_header);
6639 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6642 /* may be called multiple times per register */
6643 static void unregister_sched_domain_sysctl(void)
6645 if (sd_sysctl_header)
6646 unregister_sysctl_table(sd_sysctl_header);
6647 sd_sysctl_header = NULL;
6648 if (sd_ctl_dir[0].child)
6649 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6651 #else
6652 static void register_sched_domain_sysctl(void)
6655 static void unregister_sched_domain_sysctl(void)
6658 #endif
6660 static void set_rq_online(struct rq *rq)
6662 if (!rq->online) {
6663 const struct sched_class *class;
6665 cpumask_set_cpu(rq->cpu, rq->rd->online);
6666 rq->online = 1;
6668 for_each_class(class) {
6669 if (class->rq_online)
6670 class->rq_online(rq);
6675 static void set_rq_offline(struct rq *rq)
6677 if (rq->online) {
6678 const struct sched_class *class;
6680 for_each_class(class) {
6681 if (class->rq_offline)
6682 class->rq_offline(rq);
6685 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6686 rq->online = 0;
6691 * migration_call - callback that gets triggered when a CPU is added.
6692 * Here we can start up the necessary migration thread for the new CPU.
6694 static int __cpuinit
6695 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6697 struct task_struct *p;
6698 int cpu = (long)hcpu;
6699 unsigned long flags;
6700 struct rq *rq;
6702 switch (action) {
6704 case CPU_UP_PREPARE:
6705 case CPU_UP_PREPARE_FROZEN:
6706 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6707 if (IS_ERR(p))
6708 return NOTIFY_BAD;
6709 kthread_bind(p, cpu);
6710 /* Must be high prio: stop_machine expects to yield to it. */
6711 rq = task_rq_lock(p, &flags);
6712 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6713 task_rq_unlock(rq, &flags);
6714 cpu_rq(cpu)->migration_thread = p;
6715 break;
6717 case CPU_ONLINE:
6718 case CPU_ONLINE_FROZEN:
6719 /* Strictly unnecessary, as first user will wake it. */
6720 wake_up_process(cpu_rq(cpu)->migration_thread);
6722 /* Update our root-domain */
6723 rq = cpu_rq(cpu);
6724 spin_lock_irqsave(&rq->lock, flags);
6725 if (rq->rd) {
6726 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6728 set_rq_online(rq);
6730 spin_unlock_irqrestore(&rq->lock, flags);
6731 break;
6733 #ifdef CONFIG_HOTPLUG_CPU
6734 case CPU_UP_CANCELED:
6735 case CPU_UP_CANCELED_FROZEN:
6736 if (!cpu_rq(cpu)->migration_thread)
6737 break;
6738 /* Unbind it from offline cpu so it can run. Fall thru. */
6739 kthread_bind(cpu_rq(cpu)->migration_thread,
6740 cpumask_any(cpu_online_mask));
6741 kthread_stop(cpu_rq(cpu)->migration_thread);
6742 cpu_rq(cpu)->migration_thread = NULL;
6743 break;
6745 case CPU_DEAD:
6746 case CPU_DEAD_FROZEN:
6747 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6748 migrate_live_tasks(cpu);
6749 rq = cpu_rq(cpu);
6750 kthread_stop(rq->migration_thread);
6751 rq->migration_thread = NULL;
6752 /* Idle task back to normal (off runqueue, low prio) */
6753 spin_lock_irq(&rq->lock);
6754 update_rq_clock(rq);
6755 deactivate_task(rq, rq->idle, 0);
6756 rq->idle->static_prio = MAX_PRIO;
6757 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6758 rq->idle->sched_class = &idle_sched_class;
6759 migrate_dead_tasks(cpu);
6760 spin_unlock_irq(&rq->lock);
6761 cpuset_unlock();
6762 migrate_nr_uninterruptible(rq);
6763 BUG_ON(rq->nr_running != 0);
6766 * No need to migrate the tasks: it was best-effort if
6767 * they didn't take sched_hotcpu_mutex. Just wake up
6768 * the requestors.
6770 spin_lock_irq(&rq->lock);
6771 while (!list_empty(&rq->migration_queue)) {
6772 struct migration_req *req;
6774 req = list_entry(rq->migration_queue.next,
6775 struct migration_req, list);
6776 list_del_init(&req->list);
6777 spin_unlock_irq(&rq->lock);
6778 complete(&req->done);
6779 spin_lock_irq(&rq->lock);
6781 spin_unlock_irq(&rq->lock);
6782 break;
6784 case CPU_DYING:
6785 case CPU_DYING_FROZEN:
6786 /* Update our root-domain */
6787 rq = cpu_rq(cpu);
6788 spin_lock_irqsave(&rq->lock, flags);
6789 if (rq->rd) {
6790 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6791 set_rq_offline(rq);
6793 spin_unlock_irqrestore(&rq->lock, flags);
6794 break;
6795 #endif
6797 return NOTIFY_OK;
6800 /* Register at highest priority so that task migration (migrate_all_tasks)
6801 * happens before everything else.
6803 static struct notifier_block __cpuinitdata migration_notifier = {
6804 .notifier_call = migration_call,
6805 .priority = 10
6808 static int __init migration_init(void)
6810 void *cpu = (void *)(long)smp_processor_id();
6811 int err;
6813 /* Start one for the boot CPU: */
6814 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6815 BUG_ON(err == NOTIFY_BAD);
6816 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6817 register_cpu_notifier(&migration_notifier);
6819 return err;
6821 early_initcall(migration_init);
6822 #endif
6824 #ifdef CONFIG_SMP
6826 #ifdef CONFIG_SCHED_DEBUG
6828 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6829 struct cpumask *groupmask)
6831 struct sched_group *group = sd->groups;
6832 char str[256];
6834 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6835 cpumask_clear(groupmask);
6837 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6839 if (!(sd->flags & SD_LOAD_BALANCE)) {
6840 printk("does not load-balance\n");
6841 if (sd->parent)
6842 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6843 " has parent");
6844 return -1;
6847 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6849 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6850 printk(KERN_ERR "ERROR: domain->span does not contain "
6851 "CPU%d\n", cpu);
6853 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6854 printk(KERN_ERR "ERROR: domain->groups does not contain"
6855 " CPU%d\n", cpu);
6858 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6859 do {
6860 if (!group) {
6861 printk("\n");
6862 printk(KERN_ERR "ERROR: group is NULL\n");
6863 break;
6866 if (!group->__cpu_power) {
6867 printk(KERN_CONT "\n");
6868 printk(KERN_ERR "ERROR: domain->cpu_power not "
6869 "set\n");
6870 break;
6873 if (!cpumask_weight(sched_group_cpus(group))) {
6874 printk(KERN_CONT "\n");
6875 printk(KERN_ERR "ERROR: empty group\n");
6876 break;
6879 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6880 printk(KERN_CONT "\n");
6881 printk(KERN_ERR "ERROR: repeated CPUs\n");
6882 break;
6885 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6887 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6888 printk(KERN_CONT " %s", str);
6890 group = group->next;
6891 } while (group != sd->groups);
6892 printk(KERN_CONT "\n");
6894 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6895 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6897 if (sd->parent &&
6898 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6899 printk(KERN_ERR "ERROR: parent span is not a superset "
6900 "of domain->span\n");
6901 return 0;
6904 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6906 cpumask_var_t groupmask;
6907 int level = 0;
6909 if (!sd) {
6910 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6911 return;
6914 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6916 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6917 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6918 return;
6921 for (;;) {
6922 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6923 break;
6924 level++;
6925 sd = sd->parent;
6926 if (!sd)
6927 break;
6929 free_cpumask_var(groupmask);
6931 #else /* !CONFIG_SCHED_DEBUG */
6932 # define sched_domain_debug(sd, cpu) do { } while (0)
6933 #endif /* CONFIG_SCHED_DEBUG */
6935 static int sd_degenerate(struct sched_domain *sd)
6937 if (cpumask_weight(sched_domain_span(sd)) == 1)
6938 return 1;
6940 /* Following flags need at least 2 groups */
6941 if (sd->flags & (SD_LOAD_BALANCE |
6942 SD_BALANCE_NEWIDLE |
6943 SD_BALANCE_FORK |
6944 SD_BALANCE_EXEC |
6945 SD_SHARE_CPUPOWER |
6946 SD_SHARE_PKG_RESOURCES)) {
6947 if (sd->groups != sd->groups->next)
6948 return 0;
6951 /* Following flags don't use groups */
6952 if (sd->flags & (SD_WAKE_IDLE |
6953 SD_WAKE_AFFINE |
6954 SD_WAKE_BALANCE))
6955 return 0;
6957 return 1;
6960 static int
6961 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6963 unsigned long cflags = sd->flags, pflags = parent->flags;
6965 if (sd_degenerate(parent))
6966 return 1;
6968 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6969 return 0;
6971 /* Does parent contain flags not in child? */
6972 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6973 if (cflags & SD_WAKE_AFFINE)
6974 pflags &= ~SD_WAKE_BALANCE;
6975 /* Flags needing groups don't count if only 1 group in parent */
6976 if (parent->groups == parent->groups->next) {
6977 pflags &= ~(SD_LOAD_BALANCE |
6978 SD_BALANCE_NEWIDLE |
6979 SD_BALANCE_FORK |
6980 SD_BALANCE_EXEC |
6981 SD_SHARE_CPUPOWER |
6982 SD_SHARE_PKG_RESOURCES);
6983 if (nr_node_ids == 1)
6984 pflags &= ~SD_SERIALIZE;
6986 if (~cflags & pflags)
6987 return 0;
6989 return 1;
6992 static void free_rootdomain(struct root_domain *rd)
6994 cpupri_cleanup(&rd->cpupri);
6996 free_cpumask_var(rd->rto_mask);
6997 free_cpumask_var(rd->online);
6998 free_cpumask_var(rd->span);
6999 kfree(rd);
7002 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7004 struct root_domain *old_rd = NULL;
7005 unsigned long flags;
7007 spin_lock_irqsave(&rq->lock, flags);
7009 if (rq->rd) {
7010 old_rd = rq->rd;
7012 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7013 set_rq_offline(rq);
7015 cpumask_clear_cpu(rq->cpu, old_rd->span);
7018 * If we dont want to free the old_rt yet then
7019 * set old_rd to NULL to skip the freeing later
7020 * in this function:
7022 if (!atomic_dec_and_test(&old_rd->refcount))
7023 old_rd = NULL;
7026 atomic_inc(&rd->refcount);
7027 rq->rd = rd;
7029 cpumask_set_cpu(rq->cpu, rd->span);
7030 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7031 set_rq_online(rq);
7033 spin_unlock_irqrestore(&rq->lock, flags);
7035 if (old_rd)
7036 free_rootdomain(old_rd);
7039 static int __init_refok init_rootdomain(struct root_domain *rd, bool bootmem)
7041 memset(rd, 0, sizeof(*rd));
7043 if (bootmem) {
7044 alloc_bootmem_cpumask_var(&def_root_domain.span);
7045 alloc_bootmem_cpumask_var(&def_root_domain.online);
7046 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
7047 cpupri_init(&rd->cpupri, true);
7048 return 0;
7051 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7052 goto out;
7053 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7054 goto free_span;
7055 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7056 goto free_online;
7058 if (cpupri_init(&rd->cpupri, false) != 0)
7059 goto free_rto_mask;
7060 return 0;
7062 free_rto_mask:
7063 free_cpumask_var(rd->rto_mask);
7064 free_online:
7065 free_cpumask_var(rd->online);
7066 free_span:
7067 free_cpumask_var(rd->span);
7068 out:
7069 return -ENOMEM;
7072 static void init_defrootdomain(void)
7074 init_rootdomain(&def_root_domain, true);
7076 atomic_set(&def_root_domain.refcount, 1);
7079 static struct root_domain *alloc_rootdomain(void)
7081 struct root_domain *rd;
7083 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7084 if (!rd)
7085 return NULL;
7087 if (init_rootdomain(rd, false) != 0) {
7088 kfree(rd);
7089 return NULL;
7092 return rd;
7096 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7097 * hold the hotplug lock.
7099 static void
7100 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7102 struct rq *rq = cpu_rq(cpu);
7103 struct sched_domain *tmp;
7105 /* Remove the sched domains which do not contribute to scheduling. */
7106 for (tmp = sd; tmp; ) {
7107 struct sched_domain *parent = tmp->parent;
7108 if (!parent)
7109 break;
7111 if (sd_parent_degenerate(tmp, parent)) {
7112 tmp->parent = parent->parent;
7113 if (parent->parent)
7114 parent->parent->child = tmp;
7115 } else
7116 tmp = tmp->parent;
7119 if (sd && sd_degenerate(sd)) {
7120 sd = sd->parent;
7121 if (sd)
7122 sd->child = NULL;
7125 sched_domain_debug(sd, cpu);
7127 rq_attach_root(rq, rd);
7128 rcu_assign_pointer(rq->sd, sd);
7131 /* cpus with isolated domains */
7132 static cpumask_var_t cpu_isolated_map;
7134 /* Setup the mask of cpus configured for isolated domains */
7135 static int __init isolated_cpu_setup(char *str)
7137 cpulist_parse(str, cpu_isolated_map);
7138 return 1;
7141 __setup("isolcpus=", isolated_cpu_setup);
7144 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7145 * to a function which identifies what group(along with sched group) a CPU
7146 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7147 * (due to the fact that we keep track of groups covered with a struct cpumask).
7149 * init_sched_build_groups will build a circular linked list of the groups
7150 * covered by the given span, and will set each group's ->cpumask correctly,
7151 * and ->cpu_power to 0.
7153 static void
7154 init_sched_build_groups(const struct cpumask *span,
7155 const struct cpumask *cpu_map,
7156 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7157 struct sched_group **sg,
7158 struct cpumask *tmpmask),
7159 struct cpumask *covered, struct cpumask *tmpmask)
7161 struct sched_group *first = NULL, *last = NULL;
7162 int i;
7164 cpumask_clear(covered);
7166 for_each_cpu(i, span) {
7167 struct sched_group *sg;
7168 int group = group_fn(i, cpu_map, &sg, tmpmask);
7169 int j;
7171 if (cpumask_test_cpu(i, covered))
7172 continue;
7174 cpumask_clear(sched_group_cpus(sg));
7175 sg->__cpu_power = 0;
7177 for_each_cpu(j, span) {
7178 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7179 continue;
7181 cpumask_set_cpu(j, covered);
7182 cpumask_set_cpu(j, sched_group_cpus(sg));
7184 if (!first)
7185 first = sg;
7186 if (last)
7187 last->next = sg;
7188 last = sg;
7190 last->next = first;
7193 #define SD_NODES_PER_DOMAIN 16
7195 #ifdef CONFIG_NUMA
7198 * find_next_best_node - find the next node to include in a sched_domain
7199 * @node: node whose sched_domain we're building
7200 * @used_nodes: nodes already in the sched_domain
7202 * Find the next node to include in a given scheduling domain. Simply
7203 * finds the closest node not already in the @used_nodes map.
7205 * Should use nodemask_t.
7207 static int find_next_best_node(int node, nodemask_t *used_nodes)
7209 int i, n, val, min_val, best_node = 0;
7211 min_val = INT_MAX;
7213 for (i = 0; i < nr_node_ids; i++) {
7214 /* Start at @node */
7215 n = (node + i) % nr_node_ids;
7217 if (!nr_cpus_node(n))
7218 continue;
7220 /* Skip already used nodes */
7221 if (node_isset(n, *used_nodes))
7222 continue;
7224 /* Simple min distance search */
7225 val = node_distance(node, n);
7227 if (val < min_val) {
7228 min_val = val;
7229 best_node = n;
7233 node_set(best_node, *used_nodes);
7234 return best_node;
7238 * sched_domain_node_span - get a cpumask for a node's sched_domain
7239 * @node: node whose cpumask we're constructing
7240 * @span: resulting cpumask
7242 * Given a node, construct a good cpumask for its sched_domain to span. It
7243 * should be one that prevents unnecessary balancing, but also spreads tasks
7244 * out optimally.
7246 static void sched_domain_node_span(int node, struct cpumask *span)
7248 nodemask_t used_nodes;
7249 int i;
7251 cpumask_clear(span);
7252 nodes_clear(used_nodes);
7254 cpumask_or(span, span, cpumask_of_node(node));
7255 node_set(node, used_nodes);
7257 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7258 int next_node = find_next_best_node(node, &used_nodes);
7260 cpumask_or(span, span, cpumask_of_node(next_node));
7263 #endif /* CONFIG_NUMA */
7265 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7268 * The cpus mask in sched_group and sched_domain hangs off the end.
7269 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7270 * for nr_cpu_ids < CONFIG_NR_CPUS.
7272 struct static_sched_group {
7273 struct sched_group sg;
7274 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7277 struct static_sched_domain {
7278 struct sched_domain sd;
7279 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7283 * SMT sched-domains:
7285 #ifdef CONFIG_SCHED_SMT
7286 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7287 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7289 static int
7290 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
7291 struct sched_group **sg, struct cpumask *unused)
7293 if (sg)
7294 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7295 return cpu;
7297 #endif /* CONFIG_SCHED_SMT */
7300 * multi-core sched-domains:
7302 #ifdef CONFIG_SCHED_MC
7303 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7304 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7305 #endif /* CONFIG_SCHED_MC */
7307 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7308 static int
7309 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7310 struct sched_group **sg, struct cpumask *mask)
7312 int group;
7314 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7315 group = cpumask_first(mask);
7316 if (sg)
7317 *sg = &per_cpu(sched_group_core, group).sg;
7318 return group;
7320 #elif defined(CONFIG_SCHED_MC)
7321 static int
7322 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7323 struct sched_group **sg, struct cpumask *unused)
7325 if (sg)
7326 *sg = &per_cpu(sched_group_core, cpu).sg;
7327 return cpu;
7329 #endif
7331 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
7332 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
7334 static int
7335 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
7336 struct sched_group **sg, struct cpumask *mask)
7338 int group;
7339 #ifdef CONFIG_SCHED_MC
7340 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
7341 group = cpumask_first(mask);
7342 #elif defined(CONFIG_SCHED_SMT)
7343 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7344 group = cpumask_first(mask);
7345 #else
7346 group = cpu;
7347 #endif
7348 if (sg)
7349 *sg = &per_cpu(sched_group_phys, group).sg;
7350 return group;
7353 #ifdef CONFIG_NUMA
7355 * The init_sched_build_groups can't handle what we want to do with node
7356 * groups, so roll our own. Now each node has its own list of groups which
7357 * gets dynamically allocated.
7359 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
7360 static struct sched_group ***sched_group_nodes_bycpu;
7362 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
7363 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7365 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
7366 struct sched_group **sg,
7367 struct cpumask *nodemask)
7369 int group;
7371 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7372 group = cpumask_first(nodemask);
7374 if (sg)
7375 *sg = &per_cpu(sched_group_allnodes, group).sg;
7376 return group;
7379 static void init_numa_sched_groups_power(struct sched_group *group_head)
7381 struct sched_group *sg = group_head;
7382 int j;
7384 if (!sg)
7385 return;
7386 do {
7387 for_each_cpu(j, sched_group_cpus(sg)) {
7388 struct sched_domain *sd;
7390 sd = &per_cpu(phys_domains, j).sd;
7391 if (j != cpumask_first(sched_group_cpus(sd->groups))) {
7393 * Only add "power" once for each
7394 * physical package.
7396 continue;
7399 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7401 sg = sg->next;
7402 } while (sg != group_head);
7404 #endif /* CONFIG_NUMA */
7406 #ifdef CONFIG_NUMA
7407 /* Free memory allocated for various sched_group structures */
7408 static void free_sched_groups(const struct cpumask *cpu_map,
7409 struct cpumask *nodemask)
7411 int cpu, i;
7413 for_each_cpu(cpu, cpu_map) {
7414 struct sched_group **sched_group_nodes
7415 = sched_group_nodes_bycpu[cpu];
7417 if (!sched_group_nodes)
7418 continue;
7420 for (i = 0; i < nr_node_ids; i++) {
7421 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7423 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7424 if (cpumask_empty(nodemask))
7425 continue;
7427 if (sg == NULL)
7428 continue;
7429 sg = sg->next;
7430 next_sg:
7431 oldsg = sg;
7432 sg = sg->next;
7433 kfree(oldsg);
7434 if (oldsg != sched_group_nodes[i])
7435 goto next_sg;
7437 kfree(sched_group_nodes);
7438 sched_group_nodes_bycpu[cpu] = NULL;
7441 #else /* !CONFIG_NUMA */
7442 static void free_sched_groups(const struct cpumask *cpu_map,
7443 struct cpumask *nodemask)
7446 #endif /* CONFIG_NUMA */
7449 * Initialize sched groups cpu_power.
7451 * cpu_power indicates the capacity of sched group, which is used while
7452 * distributing the load between different sched groups in a sched domain.
7453 * Typically cpu_power for all the groups in a sched domain will be same unless
7454 * there are asymmetries in the topology. If there are asymmetries, group
7455 * having more cpu_power will pickup more load compared to the group having
7456 * less cpu_power.
7458 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7459 * the maximum number of tasks a group can handle in the presence of other idle
7460 * or lightly loaded groups in the same sched domain.
7462 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7464 struct sched_domain *child;
7465 struct sched_group *group;
7467 WARN_ON(!sd || !sd->groups);
7469 if (cpu != cpumask_first(sched_group_cpus(sd->groups)))
7470 return;
7472 child = sd->child;
7474 sd->groups->__cpu_power = 0;
7477 * For perf policy, if the groups in child domain share resources
7478 * (for example cores sharing some portions of the cache hierarchy
7479 * or SMT), then set this domain groups cpu_power such that each group
7480 * can handle only one task, when there are other idle groups in the
7481 * same sched domain.
7483 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7484 (child->flags &
7485 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7486 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7487 return;
7491 * add cpu_power of each child group to this groups cpu_power
7493 group = child->groups;
7494 do {
7495 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7496 group = group->next;
7497 } while (group != child->groups);
7501 * Initializers for schedule domains
7502 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7505 #ifdef CONFIG_SCHED_DEBUG
7506 # define SD_INIT_NAME(sd, type) sd->name = #type
7507 #else
7508 # define SD_INIT_NAME(sd, type) do { } while (0)
7509 #endif
7511 #define SD_INIT(sd, type) sd_init_##type(sd)
7513 #define SD_INIT_FUNC(type) \
7514 static noinline void sd_init_##type(struct sched_domain *sd) \
7516 memset(sd, 0, sizeof(*sd)); \
7517 *sd = SD_##type##_INIT; \
7518 sd->level = SD_LV_##type; \
7519 SD_INIT_NAME(sd, type); \
7522 SD_INIT_FUNC(CPU)
7523 #ifdef CONFIG_NUMA
7524 SD_INIT_FUNC(ALLNODES)
7525 SD_INIT_FUNC(NODE)
7526 #endif
7527 #ifdef CONFIG_SCHED_SMT
7528 SD_INIT_FUNC(SIBLING)
7529 #endif
7530 #ifdef CONFIG_SCHED_MC
7531 SD_INIT_FUNC(MC)
7532 #endif
7534 static int default_relax_domain_level = -1;
7536 static int __init setup_relax_domain_level(char *str)
7538 unsigned long val;
7540 val = simple_strtoul(str, NULL, 0);
7541 if (val < SD_LV_MAX)
7542 default_relax_domain_level = val;
7544 return 1;
7546 __setup("relax_domain_level=", setup_relax_domain_level);
7548 static void set_domain_attribute(struct sched_domain *sd,
7549 struct sched_domain_attr *attr)
7551 int request;
7553 if (!attr || attr->relax_domain_level < 0) {
7554 if (default_relax_domain_level < 0)
7555 return;
7556 else
7557 request = default_relax_domain_level;
7558 } else
7559 request = attr->relax_domain_level;
7560 if (request < sd->level) {
7561 /* turn off idle balance on this domain */
7562 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7563 } else {
7564 /* turn on idle balance on this domain */
7565 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7570 * Build sched domains for a given set of cpus and attach the sched domains
7571 * to the individual cpus
7573 static int __build_sched_domains(const struct cpumask *cpu_map,
7574 struct sched_domain_attr *attr)
7576 int i, err = -ENOMEM;
7577 struct root_domain *rd;
7578 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
7579 tmpmask;
7580 #ifdef CONFIG_NUMA
7581 cpumask_var_t domainspan, covered, notcovered;
7582 struct sched_group **sched_group_nodes = NULL;
7583 int sd_allnodes = 0;
7585 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
7586 goto out;
7587 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
7588 goto free_domainspan;
7589 if (!alloc_cpumask_var(&notcovered, GFP_KERNEL))
7590 goto free_covered;
7591 #endif
7593 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
7594 goto free_notcovered;
7595 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
7596 goto free_nodemask;
7597 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
7598 goto free_this_sibling_map;
7599 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
7600 goto free_this_core_map;
7601 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
7602 goto free_send_covered;
7604 #ifdef CONFIG_NUMA
7606 * Allocate the per-node list of sched groups
7608 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7609 GFP_KERNEL);
7610 if (!sched_group_nodes) {
7611 printk(KERN_WARNING "Can not alloc sched group node list\n");
7612 goto free_tmpmask;
7614 #endif
7616 rd = alloc_rootdomain();
7617 if (!rd) {
7618 printk(KERN_WARNING "Cannot alloc root domain\n");
7619 goto free_sched_groups;
7622 #ifdef CONFIG_NUMA
7623 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
7624 #endif
7627 * Set up domains for cpus specified by the cpu_map.
7629 for_each_cpu(i, cpu_map) {
7630 struct sched_domain *sd = NULL, *p;
7632 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
7634 #ifdef CONFIG_NUMA
7635 if (cpumask_weight(cpu_map) >
7636 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
7637 sd = &per_cpu(allnodes_domains, i).sd;
7638 SD_INIT(sd, ALLNODES);
7639 set_domain_attribute(sd, attr);
7640 cpumask_copy(sched_domain_span(sd), cpu_map);
7641 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7642 p = sd;
7643 sd_allnodes = 1;
7644 } else
7645 p = NULL;
7647 sd = &per_cpu(node_domains, i).sd;
7648 SD_INIT(sd, NODE);
7649 set_domain_attribute(sd, attr);
7650 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7651 sd->parent = p;
7652 if (p)
7653 p->child = sd;
7654 cpumask_and(sched_domain_span(sd),
7655 sched_domain_span(sd), cpu_map);
7656 #endif
7658 p = sd;
7659 sd = &per_cpu(phys_domains, i).sd;
7660 SD_INIT(sd, CPU);
7661 set_domain_attribute(sd, attr);
7662 cpumask_copy(sched_domain_span(sd), nodemask);
7663 sd->parent = p;
7664 if (p)
7665 p->child = sd;
7666 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7668 #ifdef CONFIG_SCHED_MC
7669 p = sd;
7670 sd = &per_cpu(core_domains, i).sd;
7671 SD_INIT(sd, MC);
7672 set_domain_attribute(sd, attr);
7673 cpumask_and(sched_domain_span(sd), cpu_map,
7674 cpu_coregroup_mask(i));
7675 sd->parent = p;
7676 p->child = sd;
7677 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7678 #endif
7680 #ifdef CONFIG_SCHED_SMT
7681 p = sd;
7682 sd = &per_cpu(cpu_domains, i).sd;
7683 SD_INIT(sd, SIBLING);
7684 set_domain_attribute(sd, attr);
7685 cpumask_and(sched_domain_span(sd),
7686 &per_cpu(cpu_sibling_map, i), cpu_map);
7687 sd->parent = p;
7688 p->child = sd;
7689 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7690 #endif
7693 #ifdef CONFIG_SCHED_SMT
7694 /* Set up CPU (sibling) groups */
7695 for_each_cpu(i, cpu_map) {
7696 cpumask_and(this_sibling_map,
7697 &per_cpu(cpu_sibling_map, i), cpu_map);
7698 if (i != cpumask_first(this_sibling_map))
7699 continue;
7701 init_sched_build_groups(this_sibling_map, cpu_map,
7702 &cpu_to_cpu_group,
7703 send_covered, tmpmask);
7705 #endif
7707 #ifdef CONFIG_SCHED_MC
7708 /* Set up multi-core groups */
7709 for_each_cpu(i, cpu_map) {
7710 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
7711 if (i != cpumask_first(this_core_map))
7712 continue;
7714 init_sched_build_groups(this_core_map, cpu_map,
7715 &cpu_to_core_group,
7716 send_covered, tmpmask);
7718 #endif
7720 /* Set up physical groups */
7721 for (i = 0; i < nr_node_ids; i++) {
7722 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7723 if (cpumask_empty(nodemask))
7724 continue;
7726 init_sched_build_groups(nodemask, cpu_map,
7727 &cpu_to_phys_group,
7728 send_covered, tmpmask);
7731 #ifdef CONFIG_NUMA
7732 /* Set up node groups */
7733 if (sd_allnodes) {
7734 init_sched_build_groups(cpu_map, cpu_map,
7735 &cpu_to_allnodes_group,
7736 send_covered, tmpmask);
7739 for (i = 0; i < nr_node_ids; i++) {
7740 /* Set up node groups */
7741 struct sched_group *sg, *prev;
7742 int j;
7744 cpumask_clear(covered);
7745 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7746 if (cpumask_empty(nodemask)) {
7747 sched_group_nodes[i] = NULL;
7748 continue;
7751 sched_domain_node_span(i, domainspan);
7752 cpumask_and(domainspan, domainspan, cpu_map);
7754 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7755 GFP_KERNEL, i);
7756 if (!sg) {
7757 printk(KERN_WARNING "Can not alloc domain group for "
7758 "node %d\n", i);
7759 goto error;
7761 sched_group_nodes[i] = sg;
7762 for_each_cpu(j, nodemask) {
7763 struct sched_domain *sd;
7765 sd = &per_cpu(node_domains, j).sd;
7766 sd->groups = sg;
7768 sg->__cpu_power = 0;
7769 cpumask_copy(sched_group_cpus(sg), nodemask);
7770 sg->next = sg;
7771 cpumask_or(covered, covered, nodemask);
7772 prev = sg;
7774 for (j = 0; j < nr_node_ids; j++) {
7775 int n = (i + j) % nr_node_ids;
7777 cpumask_complement(notcovered, covered);
7778 cpumask_and(tmpmask, notcovered, cpu_map);
7779 cpumask_and(tmpmask, tmpmask, domainspan);
7780 if (cpumask_empty(tmpmask))
7781 break;
7783 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
7784 if (cpumask_empty(tmpmask))
7785 continue;
7787 sg = kmalloc_node(sizeof(struct sched_group) +
7788 cpumask_size(),
7789 GFP_KERNEL, i);
7790 if (!sg) {
7791 printk(KERN_WARNING
7792 "Can not alloc domain group for node %d\n", j);
7793 goto error;
7795 sg->__cpu_power = 0;
7796 cpumask_copy(sched_group_cpus(sg), tmpmask);
7797 sg->next = prev->next;
7798 cpumask_or(covered, covered, tmpmask);
7799 prev->next = sg;
7800 prev = sg;
7803 #endif
7805 /* Calculate CPU power for physical packages and nodes */
7806 #ifdef CONFIG_SCHED_SMT
7807 for_each_cpu(i, cpu_map) {
7808 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
7810 init_sched_groups_power(i, sd);
7812 #endif
7813 #ifdef CONFIG_SCHED_MC
7814 for_each_cpu(i, cpu_map) {
7815 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
7817 init_sched_groups_power(i, sd);
7819 #endif
7821 for_each_cpu(i, cpu_map) {
7822 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
7824 init_sched_groups_power(i, sd);
7827 #ifdef CONFIG_NUMA
7828 for (i = 0; i < nr_node_ids; i++)
7829 init_numa_sched_groups_power(sched_group_nodes[i]);
7831 if (sd_allnodes) {
7832 struct sched_group *sg;
7834 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7835 tmpmask);
7836 init_numa_sched_groups_power(sg);
7838 #endif
7840 /* Attach the domains */
7841 for_each_cpu(i, cpu_map) {
7842 struct sched_domain *sd;
7843 #ifdef CONFIG_SCHED_SMT
7844 sd = &per_cpu(cpu_domains, i).sd;
7845 #elif defined(CONFIG_SCHED_MC)
7846 sd = &per_cpu(core_domains, i).sd;
7847 #else
7848 sd = &per_cpu(phys_domains, i).sd;
7849 #endif
7850 cpu_attach_domain(sd, rd, i);
7853 err = 0;
7855 free_tmpmask:
7856 free_cpumask_var(tmpmask);
7857 free_send_covered:
7858 free_cpumask_var(send_covered);
7859 free_this_core_map:
7860 free_cpumask_var(this_core_map);
7861 free_this_sibling_map:
7862 free_cpumask_var(this_sibling_map);
7863 free_nodemask:
7864 free_cpumask_var(nodemask);
7865 free_notcovered:
7866 #ifdef CONFIG_NUMA
7867 free_cpumask_var(notcovered);
7868 free_covered:
7869 free_cpumask_var(covered);
7870 free_domainspan:
7871 free_cpumask_var(domainspan);
7872 out:
7873 #endif
7874 return err;
7876 free_sched_groups:
7877 #ifdef CONFIG_NUMA
7878 kfree(sched_group_nodes);
7879 #endif
7880 goto free_tmpmask;
7882 #ifdef CONFIG_NUMA
7883 error:
7884 free_sched_groups(cpu_map, tmpmask);
7885 free_rootdomain(rd);
7886 goto free_tmpmask;
7887 #endif
7890 static int build_sched_domains(const struct cpumask *cpu_map)
7892 return __build_sched_domains(cpu_map, NULL);
7895 static struct cpumask *doms_cur; /* current sched domains */
7896 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7897 static struct sched_domain_attr *dattr_cur;
7898 /* attribues of custom domains in 'doms_cur' */
7901 * Special case: If a kmalloc of a doms_cur partition (array of
7902 * cpumask) fails, then fallback to a single sched domain,
7903 * as determined by the single cpumask fallback_doms.
7905 static cpumask_var_t fallback_doms;
7908 * arch_update_cpu_topology lets virtualized architectures update the
7909 * cpu core maps. It is supposed to return 1 if the topology changed
7910 * or 0 if it stayed the same.
7912 int __attribute__((weak)) arch_update_cpu_topology(void)
7914 return 0;
7918 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7919 * For now this just excludes isolated cpus, but could be used to
7920 * exclude other special cases in the future.
7922 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7924 int err;
7926 arch_update_cpu_topology();
7927 ndoms_cur = 1;
7928 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
7929 if (!doms_cur)
7930 doms_cur = fallback_doms;
7931 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
7932 dattr_cur = NULL;
7933 err = build_sched_domains(doms_cur);
7934 register_sched_domain_sysctl();
7936 return err;
7939 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7940 struct cpumask *tmpmask)
7942 free_sched_groups(cpu_map, tmpmask);
7946 * Detach sched domains from a group of cpus specified in cpu_map
7947 * These cpus will now be attached to the NULL domain
7949 static void detach_destroy_domains(const struct cpumask *cpu_map)
7951 /* Save because hotplug lock held. */
7952 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7953 int i;
7955 for_each_cpu(i, cpu_map)
7956 cpu_attach_domain(NULL, &def_root_domain, i);
7957 synchronize_sched();
7958 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7961 /* handle null as "default" */
7962 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7963 struct sched_domain_attr *new, int idx_new)
7965 struct sched_domain_attr tmp;
7967 /* fast path */
7968 if (!new && !cur)
7969 return 1;
7971 tmp = SD_ATTR_INIT;
7972 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7973 new ? (new + idx_new) : &tmp,
7974 sizeof(struct sched_domain_attr));
7978 * Partition sched domains as specified by the 'ndoms_new'
7979 * cpumasks in the array doms_new[] of cpumasks. This compares
7980 * doms_new[] to the current sched domain partitioning, doms_cur[].
7981 * It destroys each deleted domain and builds each new domain.
7983 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
7984 * The masks don't intersect (don't overlap.) We should setup one
7985 * sched domain for each mask. CPUs not in any of the cpumasks will
7986 * not be load balanced. If the same cpumask appears both in the
7987 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7988 * it as it is.
7990 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7991 * ownership of it and will kfree it when done with it. If the caller
7992 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7993 * ndoms_new == 1, and partition_sched_domains() will fallback to
7994 * the single partition 'fallback_doms', it also forces the domains
7995 * to be rebuilt.
7997 * If doms_new == NULL it will be replaced with cpu_online_mask.
7998 * ndoms_new == 0 is a special case for destroying existing domains,
7999 * and it will not create the default domain.
8001 * Call with hotplug lock held
8003 /* FIXME: Change to struct cpumask *doms_new[] */
8004 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8005 struct sched_domain_attr *dattr_new)
8007 int i, j, n;
8008 int new_topology;
8010 mutex_lock(&sched_domains_mutex);
8012 /* always unregister in case we don't destroy any domains */
8013 unregister_sched_domain_sysctl();
8015 /* Let architecture update cpu core mappings. */
8016 new_topology = arch_update_cpu_topology();
8018 n = doms_new ? ndoms_new : 0;
8020 /* Destroy deleted domains */
8021 for (i = 0; i < ndoms_cur; i++) {
8022 for (j = 0; j < n && !new_topology; j++) {
8023 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8024 && dattrs_equal(dattr_cur, i, dattr_new, j))
8025 goto match1;
8027 /* no match - a current sched domain not in new doms_new[] */
8028 detach_destroy_domains(doms_cur + i);
8029 match1:
8033 if (doms_new == NULL) {
8034 ndoms_cur = 0;
8035 doms_new = fallback_doms;
8036 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8037 WARN_ON_ONCE(dattr_new);
8040 /* Build new domains */
8041 for (i = 0; i < ndoms_new; i++) {
8042 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8043 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8044 && dattrs_equal(dattr_new, i, dattr_cur, j))
8045 goto match2;
8047 /* no match - add a new doms_new */
8048 __build_sched_domains(doms_new + i,
8049 dattr_new ? dattr_new + i : NULL);
8050 match2:
8054 /* Remember the new sched domains */
8055 if (doms_cur != fallback_doms)
8056 kfree(doms_cur);
8057 kfree(dattr_cur); /* kfree(NULL) is safe */
8058 doms_cur = doms_new;
8059 dattr_cur = dattr_new;
8060 ndoms_cur = ndoms_new;
8062 register_sched_domain_sysctl();
8064 mutex_unlock(&sched_domains_mutex);
8067 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8068 static void arch_reinit_sched_domains(void)
8070 get_online_cpus();
8072 /* Destroy domains first to force the rebuild */
8073 partition_sched_domains(0, NULL, NULL);
8075 rebuild_sched_domains();
8076 put_online_cpus();
8079 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8081 unsigned int level = 0;
8083 if (sscanf(buf, "%u", &level) != 1)
8084 return -EINVAL;
8087 * level is always be positive so don't check for
8088 * level < POWERSAVINGS_BALANCE_NONE which is 0
8089 * What happens on 0 or 1 byte write,
8090 * need to check for count as well?
8093 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8094 return -EINVAL;
8096 if (smt)
8097 sched_smt_power_savings = level;
8098 else
8099 sched_mc_power_savings = level;
8101 arch_reinit_sched_domains();
8103 return count;
8106 #ifdef CONFIG_SCHED_MC
8107 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8108 char *page)
8110 return sprintf(page, "%u\n", sched_mc_power_savings);
8112 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8113 const char *buf, size_t count)
8115 return sched_power_savings_store(buf, count, 0);
8117 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8118 sched_mc_power_savings_show,
8119 sched_mc_power_savings_store);
8120 #endif
8122 #ifdef CONFIG_SCHED_SMT
8123 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8124 char *page)
8126 return sprintf(page, "%u\n", sched_smt_power_savings);
8128 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8129 const char *buf, size_t count)
8131 return sched_power_savings_store(buf, count, 1);
8133 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8134 sched_smt_power_savings_show,
8135 sched_smt_power_savings_store);
8136 #endif
8138 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8140 int err = 0;
8142 #ifdef CONFIG_SCHED_SMT
8143 if (smt_capable())
8144 err = sysfs_create_file(&cls->kset.kobj,
8145 &attr_sched_smt_power_savings.attr);
8146 #endif
8147 #ifdef CONFIG_SCHED_MC
8148 if (!err && mc_capable())
8149 err = sysfs_create_file(&cls->kset.kobj,
8150 &attr_sched_mc_power_savings.attr);
8151 #endif
8152 return err;
8154 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8156 #ifndef CONFIG_CPUSETS
8158 * Add online and remove offline CPUs from the scheduler domains.
8159 * When cpusets are enabled they take over this function.
8161 static int update_sched_domains(struct notifier_block *nfb,
8162 unsigned long action, void *hcpu)
8164 switch (action) {
8165 case CPU_ONLINE:
8166 case CPU_ONLINE_FROZEN:
8167 case CPU_DEAD:
8168 case CPU_DEAD_FROZEN:
8169 partition_sched_domains(1, NULL, NULL);
8170 return NOTIFY_OK;
8172 default:
8173 return NOTIFY_DONE;
8176 #endif
8178 static int update_runtime(struct notifier_block *nfb,
8179 unsigned long action, void *hcpu)
8181 int cpu = (int)(long)hcpu;
8183 switch (action) {
8184 case CPU_DOWN_PREPARE:
8185 case CPU_DOWN_PREPARE_FROZEN:
8186 disable_runtime(cpu_rq(cpu));
8187 return NOTIFY_OK;
8189 case CPU_DOWN_FAILED:
8190 case CPU_DOWN_FAILED_FROZEN:
8191 case CPU_ONLINE:
8192 case CPU_ONLINE_FROZEN:
8193 enable_runtime(cpu_rq(cpu));
8194 return NOTIFY_OK;
8196 default:
8197 return NOTIFY_DONE;
8201 void __init sched_init_smp(void)
8203 cpumask_var_t non_isolated_cpus;
8205 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8207 #if defined(CONFIG_NUMA)
8208 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8209 GFP_KERNEL);
8210 BUG_ON(sched_group_nodes_bycpu == NULL);
8211 #endif
8212 get_online_cpus();
8213 mutex_lock(&sched_domains_mutex);
8214 arch_init_sched_domains(cpu_online_mask);
8215 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8216 if (cpumask_empty(non_isolated_cpus))
8217 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8218 mutex_unlock(&sched_domains_mutex);
8219 put_online_cpus();
8221 #ifndef CONFIG_CPUSETS
8222 /* XXX: Theoretical race here - CPU may be hotplugged now */
8223 hotcpu_notifier(update_sched_domains, 0);
8224 #endif
8226 /* RT runtime code needs to handle some hotplug events */
8227 hotcpu_notifier(update_runtime, 0);
8229 init_hrtick();
8231 /* Move init over to a non-isolated CPU */
8232 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8233 BUG();
8234 sched_init_granularity();
8235 free_cpumask_var(non_isolated_cpus);
8237 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8238 init_sched_rt_class();
8240 #else
8241 void __init sched_init_smp(void)
8243 sched_init_granularity();
8245 #endif /* CONFIG_SMP */
8247 int in_sched_functions(unsigned long addr)
8249 return in_lock_functions(addr) ||
8250 (addr >= (unsigned long)__sched_text_start
8251 && addr < (unsigned long)__sched_text_end);
8254 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8256 cfs_rq->tasks_timeline = RB_ROOT;
8257 INIT_LIST_HEAD(&cfs_rq->tasks);
8258 #ifdef CONFIG_FAIR_GROUP_SCHED
8259 cfs_rq->rq = rq;
8260 #endif
8261 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8264 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8266 struct rt_prio_array *array;
8267 int i;
8269 array = &rt_rq->active;
8270 for (i = 0; i < MAX_RT_PRIO; i++) {
8271 INIT_LIST_HEAD(array->queue + i);
8272 __clear_bit(i, array->bitmap);
8274 /* delimiter for bitsearch: */
8275 __set_bit(MAX_RT_PRIO, array->bitmap);
8277 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8278 rt_rq->highest_prio = MAX_RT_PRIO;
8279 #endif
8280 #ifdef CONFIG_SMP
8281 rt_rq->rt_nr_migratory = 0;
8282 rt_rq->overloaded = 0;
8283 #endif
8285 rt_rq->rt_time = 0;
8286 rt_rq->rt_throttled = 0;
8287 rt_rq->rt_runtime = 0;
8288 spin_lock_init(&rt_rq->rt_runtime_lock);
8290 #ifdef CONFIG_RT_GROUP_SCHED
8291 rt_rq->rt_nr_boosted = 0;
8292 rt_rq->rq = rq;
8293 #endif
8296 #ifdef CONFIG_FAIR_GROUP_SCHED
8297 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8298 struct sched_entity *se, int cpu, int add,
8299 struct sched_entity *parent)
8301 struct rq *rq = cpu_rq(cpu);
8302 tg->cfs_rq[cpu] = cfs_rq;
8303 init_cfs_rq(cfs_rq, rq);
8304 cfs_rq->tg = tg;
8305 if (add)
8306 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8308 tg->se[cpu] = se;
8309 /* se could be NULL for init_task_group */
8310 if (!se)
8311 return;
8313 if (!parent)
8314 se->cfs_rq = &rq->cfs;
8315 else
8316 se->cfs_rq = parent->my_q;
8318 se->my_q = cfs_rq;
8319 se->load.weight = tg->shares;
8320 se->load.inv_weight = 0;
8321 se->parent = parent;
8323 #endif
8325 #ifdef CONFIG_RT_GROUP_SCHED
8326 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8327 struct sched_rt_entity *rt_se, int cpu, int add,
8328 struct sched_rt_entity *parent)
8330 struct rq *rq = cpu_rq(cpu);
8332 tg->rt_rq[cpu] = rt_rq;
8333 init_rt_rq(rt_rq, rq);
8334 rt_rq->tg = tg;
8335 rt_rq->rt_se = rt_se;
8336 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8337 if (add)
8338 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8340 tg->rt_se[cpu] = rt_se;
8341 if (!rt_se)
8342 return;
8344 if (!parent)
8345 rt_se->rt_rq = &rq->rt;
8346 else
8347 rt_se->rt_rq = parent->my_q;
8349 rt_se->my_q = rt_rq;
8350 rt_se->parent = parent;
8351 INIT_LIST_HEAD(&rt_se->run_list);
8353 #endif
8355 void __init sched_init(void)
8357 int i, j;
8358 unsigned long alloc_size = 0, ptr;
8360 #ifdef CONFIG_FAIR_GROUP_SCHED
8361 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8362 #endif
8363 #ifdef CONFIG_RT_GROUP_SCHED
8364 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8365 #endif
8366 #ifdef CONFIG_USER_SCHED
8367 alloc_size *= 2;
8368 #endif
8370 * As sched_init() is called before page_alloc is setup,
8371 * we use alloc_bootmem().
8373 if (alloc_size) {
8374 ptr = (unsigned long)alloc_bootmem(alloc_size);
8376 #ifdef CONFIG_FAIR_GROUP_SCHED
8377 init_task_group.se = (struct sched_entity **)ptr;
8378 ptr += nr_cpu_ids * sizeof(void **);
8380 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8381 ptr += nr_cpu_ids * sizeof(void **);
8383 #ifdef CONFIG_USER_SCHED
8384 root_task_group.se = (struct sched_entity **)ptr;
8385 ptr += nr_cpu_ids * sizeof(void **);
8387 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8388 ptr += nr_cpu_ids * sizeof(void **);
8389 #endif /* CONFIG_USER_SCHED */
8390 #endif /* CONFIG_FAIR_GROUP_SCHED */
8391 #ifdef CONFIG_RT_GROUP_SCHED
8392 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8393 ptr += nr_cpu_ids * sizeof(void **);
8395 init_task_group.rt_rq = (struct rt_rq **)ptr;
8396 ptr += nr_cpu_ids * sizeof(void **);
8398 #ifdef CONFIG_USER_SCHED
8399 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8400 ptr += nr_cpu_ids * sizeof(void **);
8402 root_task_group.rt_rq = (struct rt_rq **)ptr;
8403 ptr += nr_cpu_ids * sizeof(void **);
8404 #endif /* CONFIG_USER_SCHED */
8405 #endif /* CONFIG_RT_GROUP_SCHED */
8408 #ifdef CONFIG_SMP
8409 init_defrootdomain();
8410 #endif
8412 init_rt_bandwidth(&def_rt_bandwidth,
8413 global_rt_period(), global_rt_runtime());
8415 #ifdef CONFIG_RT_GROUP_SCHED
8416 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8417 global_rt_period(), global_rt_runtime());
8418 #ifdef CONFIG_USER_SCHED
8419 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8420 global_rt_period(), RUNTIME_INF);
8421 #endif /* CONFIG_USER_SCHED */
8422 #endif /* CONFIG_RT_GROUP_SCHED */
8424 #ifdef CONFIG_GROUP_SCHED
8425 list_add(&init_task_group.list, &task_groups);
8426 INIT_LIST_HEAD(&init_task_group.children);
8428 #ifdef CONFIG_USER_SCHED
8429 INIT_LIST_HEAD(&root_task_group.children);
8430 init_task_group.parent = &root_task_group;
8431 list_add(&init_task_group.siblings, &root_task_group.children);
8432 #endif /* CONFIG_USER_SCHED */
8433 #endif /* CONFIG_GROUP_SCHED */
8435 for_each_possible_cpu(i) {
8436 struct rq *rq;
8438 rq = cpu_rq(i);
8439 spin_lock_init(&rq->lock);
8440 rq->nr_running = 0;
8441 init_cfs_rq(&rq->cfs, rq);
8442 init_rt_rq(&rq->rt, rq);
8443 #ifdef CONFIG_FAIR_GROUP_SCHED
8444 init_task_group.shares = init_task_group_load;
8445 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8446 #ifdef CONFIG_CGROUP_SCHED
8448 * How much cpu bandwidth does init_task_group get?
8450 * In case of task-groups formed thr' the cgroup filesystem, it
8451 * gets 100% of the cpu resources in the system. This overall
8452 * system cpu resource is divided among the tasks of
8453 * init_task_group and its child task-groups in a fair manner,
8454 * based on each entity's (task or task-group's) weight
8455 * (se->load.weight).
8457 * In other words, if init_task_group has 10 tasks of weight
8458 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8459 * then A0's share of the cpu resource is:
8461 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8463 * We achieve this by letting init_task_group's tasks sit
8464 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8466 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8467 #elif defined CONFIG_USER_SCHED
8468 root_task_group.shares = NICE_0_LOAD;
8469 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8471 * In case of task-groups formed thr' the user id of tasks,
8472 * init_task_group represents tasks belonging to root user.
8473 * Hence it forms a sibling of all subsequent groups formed.
8474 * In this case, init_task_group gets only a fraction of overall
8475 * system cpu resource, based on the weight assigned to root
8476 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8477 * by letting tasks of init_task_group sit in a separate cfs_rq
8478 * (init_cfs_rq) and having one entity represent this group of
8479 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8481 init_tg_cfs_entry(&init_task_group,
8482 &per_cpu(init_cfs_rq, i),
8483 &per_cpu(init_sched_entity, i), i, 1,
8484 root_task_group.se[i]);
8486 #endif
8487 #endif /* CONFIG_FAIR_GROUP_SCHED */
8489 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8490 #ifdef CONFIG_RT_GROUP_SCHED
8491 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8492 #ifdef CONFIG_CGROUP_SCHED
8493 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8494 #elif defined CONFIG_USER_SCHED
8495 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8496 init_tg_rt_entry(&init_task_group,
8497 &per_cpu(init_rt_rq, i),
8498 &per_cpu(init_sched_rt_entity, i), i, 1,
8499 root_task_group.rt_se[i]);
8500 #endif
8501 #endif
8503 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8504 rq->cpu_load[j] = 0;
8505 #ifdef CONFIG_SMP
8506 rq->sd = NULL;
8507 rq->rd = NULL;
8508 rq->active_balance = 0;
8509 rq->next_balance = jiffies;
8510 rq->push_cpu = 0;
8511 rq->cpu = i;
8512 rq->online = 0;
8513 rq->migration_thread = NULL;
8514 INIT_LIST_HEAD(&rq->migration_queue);
8515 rq_attach_root(rq, &def_root_domain);
8516 #endif
8517 init_rq_hrtick(rq);
8518 atomic_set(&rq->nr_iowait, 0);
8521 set_load_weight(&init_task);
8523 #ifdef CONFIG_PREEMPT_NOTIFIERS
8524 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8525 #endif
8527 #ifdef CONFIG_SMP
8528 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8529 #endif
8531 #ifdef CONFIG_RT_MUTEXES
8532 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8533 #endif
8536 * The boot idle thread does lazy MMU switching as well:
8538 atomic_inc(&init_mm.mm_count);
8539 enter_lazy_tlb(&init_mm, current);
8542 * Make us the idle thread. Technically, schedule() should not be
8543 * called from this thread, however somewhere below it might be,
8544 * but because we are the idle thread, we just pick up running again
8545 * when this runqueue becomes "idle".
8547 init_idle(current, smp_processor_id());
8549 * During early bootup we pretend to be a normal task:
8551 current->sched_class = &fair_sched_class;
8553 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8554 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
8555 #ifdef CONFIG_SMP
8556 #ifdef CONFIG_NO_HZ
8557 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
8558 #endif
8559 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8560 #endif /* SMP */
8562 scheduler_running = 1;
8565 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8566 void __might_sleep(char *file, int line)
8568 #ifdef in_atomic
8569 static unsigned long prev_jiffy; /* ratelimiting */
8571 if ((!in_atomic() && !irqs_disabled()) ||
8572 system_state != SYSTEM_RUNNING || oops_in_progress)
8573 return;
8574 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8575 return;
8576 prev_jiffy = jiffies;
8578 printk(KERN_ERR
8579 "BUG: sleeping function called from invalid context at %s:%d\n",
8580 file, line);
8581 printk(KERN_ERR
8582 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8583 in_atomic(), irqs_disabled(),
8584 current->pid, current->comm);
8586 debug_show_held_locks(current);
8587 if (irqs_disabled())
8588 print_irqtrace_events(current);
8589 dump_stack();
8590 #endif
8592 EXPORT_SYMBOL(__might_sleep);
8593 #endif
8595 #ifdef CONFIG_MAGIC_SYSRQ
8596 static void normalize_task(struct rq *rq, struct task_struct *p)
8598 int on_rq;
8600 update_rq_clock(rq);
8601 on_rq = p->se.on_rq;
8602 if (on_rq)
8603 deactivate_task(rq, p, 0);
8604 __setscheduler(rq, p, SCHED_NORMAL, 0);
8605 if (on_rq) {
8606 activate_task(rq, p, 0);
8607 resched_task(rq->curr);
8611 void normalize_rt_tasks(void)
8613 struct task_struct *g, *p;
8614 unsigned long flags;
8615 struct rq *rq;
8617 read_lock_irqsave(&tasklist_lock, flags);
8618 do_each_thread(g, p) {
8620 * Only normalize user tasks:
8622 if (!p->mm)
8623 continue;
8625 p->se.exec_start = 0;
8626 #ifdef CONFIG_SCHEDSTATS
8627 p->se.wait_start = 0;
8628 p->se.sleep_start = 0;
8629 p->se.block_start = 0;
8630 #endif
8632 if (!rt_task(p)) {
8634 * Renice negative nice level userspace
8635 * tasks back to 0:
8637 if (TASK_NICE(p) < 0 && p->mm)
8638 set_user_nice(p, 0);
8639 continue;
8642 spin_lock(&p->pi_lock);
8643 rq = __task_rq_lock(p);
8645 normalize_task(rq, p);
8647 __task_rq_unlock(rq);
8648 spin_unlock(&p->pi_lock);
8649 } while_each_thread(g, p);
8651 read_unlock_irqrestore(&tasklist_lock, flags);
8654 #endif /* CONFIG_MAGIC_SYSRQ */
8656 #ifdef CONFIG_IA64
8658 * These functions are only useful for the IA64 MCA handling.
8660 * They can only be called when the whole system has been
8661 * stopped - every CPU needs to be quiescent, and no scheduling
8662 * activity can take place. Using them for anything else would
8663 * be a serious bug, and as a result, they aren't even visible
8664 * under any other configuration.
8668 * curr_task - return the current task for a given cpu.
8669 * @cpu: the processor in question.
8671 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8673 struct task_struct *curr_task(int cpu)
8675 return cpu_curr(cpu);
8679 * set_curr_task - set the current task for a given cpu.
8680 * @cpu: the processor in question.
8681 * @p: the task pointer to set.
8683 * Description: This function must only be used when non-maskable interrupts
8684 * are serviced on a separate stack. It allows the architecture to switch the
8685 * notion of the current task on a cpu in a non-blocking manner. This function
8686 * must be called with all CPU's synchronized, and interrupts disabled, the
8687 * and caller must save the original value of the current task (see
8688 * curr_task() above) and restore that value before reenabling interrupts and
8689 * re-starting the system.
8691 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8693 void set_curr_task(int cpu, struct task_struct *p)
8695 cpu_curr(cpu) = p;
8698 #endif
8700 #ifdef CONFIG_FAIR_GROUP_SCHED
8701 static void free_fair_sched_group(struct task_group *tg)
8703 int i;
8705 for_each_possible_cpu(i) {
8706 if (tg->cfs_rq)
8707 kfree(tg->cfs_rq[i]);
8708 if (tg->se)
8709 kfree(tg->se[i]);
8712 kfree(tg->cfs_rq);
8713 kfree(tg->se);
8716 static
8717 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8719 struct cfs_rq *cfs_rq;
8720 struct sched_entity *se;
8721 struct rq *rq;
8722 int i;
8724 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8725 if (!tg->cfs_rq)
8726 goto err;
8727 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8728 if (!tg->se)
8729 goto err;
8731 tg->shares = NICE_0_LOAD;
8733 for_each_possible_cpu(i) {
8734 rq = cpu_rq(i);
8736 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8737 GFP_KERNEL, cpu_to_node(i));
8738 if (!cfs_rq)
8739 goto err;
8741 se = kzalloc_node(sizeof(struct sched_entity),
8742 GFP_KERNEL, cpu_to_node(i));
8743 if (!se)
8744 goto err;
8746 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8749 return 1;
8751 err:
8752 return 0;
8755 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8757 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8758 &cpu_rq(cpu)->leaf_cfs_rq_list);
8761 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8763 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8765 #else /* !CONFG_FAIR_GROUP_SCHED */
8766 static inline void free_fair_sched_group(struct task_group *tg)
8770 static inline
8771 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8773 return 1;
8776 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8780 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8783 #endif /* CONFIG_FAIR_GROUP_SCHED */
8785 #ifdef CONFIG_RT_GROUP_SCHED
8786 static void free_rt_sched_group(struct task_group *tg)
8788 int i;
8790 destroy_rt_bandwidth(&tg->rt_bandwidth);
8792 for_each_possible_cpu(i) {
8793 if (tg->rt_rq)
8794 kfree(tg->rt_rq[i]);
8795 if (tg->rt_se)
8796 kfree(tg->rt_se[i]);
8799 kfree(tg->rt_rq);
8800 kfree(tg->rt_se);
8803 static
8804 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8806 struct rt_rq *rt_rq;
8807 struct sched_rt_entity *rt_se;
8808 struct rq *rq;
8809 int i;
8811 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8812 if (!tg->rt_rq)
8813 goto err;
8814 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8815 if (!tg->rt_se)
8816 goto err;
8818 init_rt_bandwidth(&tg->rt_bandwidth,
8819 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8821 for_each_possible_cpu(i) {
8822 rq = cpu_rq(i);
8824 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8825 GFP_KERNEL, cpu_to_node(i));
8826 if (!rt_rq)
8827 goto err;
8829 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8830 GFP_KERNEL, cpu_to_node(i));
8831 if (!rt_se)
8832 goto err;
8834 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8837 return 1;
8839 err:
8840 return 0;
8843 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8845 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8846 &cpu_rq(cpu)->leaf_rt_rq_list);
8849 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8851 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8853 #else /* !CONFIG_RT_GROUP_SCHED */
8854 static inline void free_rt_sched_group(struct task_group *tg)
8858 static inline
8859 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8861 return 1;
8864 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8868 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8871 #endif /* CONFIG_RT_GROUP_SCHED */
8873 #ifdef CONFIG_GROUP_SCHED
8874 static void free_sched_group(struct task_group *tg)
8876 free_fair_sched_group(tg);
8877 free_rt_sched_group(tg);
8878 kfree(tg);
8881 /* allocate runqueue etc for a new task group */
8882 struct task_group *sched_create_group(struct task_group *parent)
8884 struct task_group *tg;
8885 unsigned long flags;
8886 int i;
8888 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8889 if (!tg)
8890 return ERR_PTR(-ENOMEM);
8892 if (!alloc_fair_sched_group(tg, parent))
8893 goto err;
8895 if (!alloc_rt_sched_group(tg, parent))
8896 goto err;
8898 spin_lock_irqsave(&task_group_lock, flags);
8899 for_each_possible_cpu(i) {
8900 register_fair_sched_group(tg, i);
8901 register_rt_sched_group(tg, i);
8903 list_add_rcu(&tg->list, &task_groups);
8905 WARN_ON(!parent); /* root should already exist */
8907 tg->parent = parent;
8908 INIT_LIST_HEAD(&tg->children);
8909 list_add_rcu(&tg->siblings, &parent->children);
8910 spin_unlock_irqrestore(&task_group_lock, flags);
8912 return tg;
8914 err:
8915 free_sched_group(tg);
8916 return ERR_PTR(-ENOMEM);
8919 /* rcu callback to free various structures associated with a task group */
8920 static void free_sched_group_rcu(struct rcu_head *rhp)
8922 /* now it should be safe to free those cfs_rqs */
8923 free_sched_group(container_of(rhp, struct task_group, rcu));
8926 /* Destroy runqueue etc associated with a task group */
8927 void sched_destroy_group(struct task_group *tg)
8929 unsigned long flags;
8930 int i;
8932 spin_lock_irqsave(&task_group_lock, flags);
8933 for_each_possible_cpu(i) {
8934 unregister_fair_sched_group(tg, i);
8935 unregister_rt_sched_group(tg, i);
8937 list_del_rcu(&tg->list);
8938 list_del_rcu(&tg->siblings);
8939 spin_unlock_irqrestore(&task_group_lock, flags);
8941 /* wait for possible concurrent references to cfs_rqs complete */
8942 call_rcu(&tg->rcu, free_sched_group_rcu);
8945 /* change task's runqueue when it moves between groups.
8946 * The caller of this function should have put the task in its new group
8947 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8948 * reflect its new group.
8950 void sched_move_task(struct task_struct *tsk)
8952 int on_rq, running;
8953 unsigned long flags;
8954 struct rq *rq;
8956 rq = task_rq_lock(tsk, &flags);
8958 update_rq_clock(rq);
8960 running = task_current(rq, tsk);
8961 on_rq = tsk->se.on_rq;
8963 if (on_rq)
8964 dequeue_task(rq, tsk, 0);
8965 if (unlikely(running))
8966 tsk->sched_class->put_prev_task(rq, tsk);
8968 set_task_rq(tsk, task_cpu(tsk));
8970 #ifdef CONFIG_FAIR_GROUP_SCHED
8971 if (tsk->sched_class->moved_group)
8972 tsk->sched_class->moved_group(tsk);
8973 #endif
8975 if (unlikely(running))
8976 tsk->sched_class->set_curr_task(rq);
8977 if (on_rq)
8978 enqueue_task(rq, tsk, 0);
8980 task_rq_unlock(rq, &flags);
8982 #endif /* CONFIG_GROUP_SCHED */
8984 #ifdef CONFIG_FAIR_GROUP_SCHED
8985 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8987 struct cfs_rq *cfs_rq = se->cfs_rq;
8988 int on_rq;
8990 on_rq = se->on_rq;
8991 if (on_rq)
8992 dequeue_entity(cfs_rq, se, 0);
8994 se->load.weight = shares;
8995 se->load.inv_weight = 0;
8997 if (on_rq)
8998 enqueue_entity(cfs_rq, se, 0);
9001 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9003 struct cfs_rq *cfs_rq = se->cfs_rq;
9004 struct rq *rq = cfs_rq->rq;
9005 unsigned long flags;
9007 spin_lock_irqsave(&rq->lock, flags);
9008 __set_se_shares(se, shares);
9009 spin_unlock_irqrestore(&rq->lock, flags);
9012 static DEFINE_MUTEX(shares_mutex);
9014 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9016 int i;
9017 unsigned long flags;
9020 * We can't change the weight of the root cgroup.
9022 if (!tg->se[0])
9023 return -EINVAL;
9025 if (shares < MIN_SHARES)
9026 shares = MIN_SHARES;
9027 else if (shares > MAX_SHARES)
9028 shares = MAX_SHARES;
9030 mutex_lock(&shares_mutex);
9031 if (tg->shares == shares)
9032 goto done;
9034 spin_lock_irqsave(&task_group_lock, flags);
9035 for_each_possible_cpu(i)
9036 unregister_fair_sched_group(tg, i);
9037 list_del_rcu(&tg->siblings);
9038 spin_unlock_irqrestore(&task_group_lock, flags);
9040 /* wait for any ongoing reference to this group to finish */
9041 synchronize_sched();
9044 * Now we are free to modify the group's share on each cpu
9045 * w/o tripping rebalance_share or load_balance_fair.
9047 tg->shares = shares;
9048 for_each_possible_cpu(i) {
9050 * force a rebalance
9052 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9053 set_se_shares(tg->se[i], shares);
9057 * Enable load balance activity on this group, by inserting it back on
9058 * each cpu's rq->leaf_cfs_rq_list.
9060 spin_lock_irqsave(&task_group_lock, flags);
9061 for_each_possible_cpu(i)
9062 register_fair_sched_group(tg, i);
9063 list_add_rcu(&tg->siblings, &tg->parent->children);
9064 spin_unlock_irqrestore(&task_group_lock, flags);
9065 done:
9066 mutex_unlock(&shares_mutex);
9067 return 0;
9070 unsigned long sched_group_shares(struct task_group *tg)
9072 return tg->shares;
9074 #endif
9076 #ifdef CONFIG_RT_GROUP_SCHED
9078 * Ensure that the real time constraints are schedulable.
9080 static DEFINE_MUTEX(rt_constraints_mutex);
9082 static unsigned long to_ratio(u64 period, u64 runtime)
9084 if (runtime == RUNTIME_INF)
9085 return 1ULL << 20;
9087 return div64_u64(runtime << 20, period);
9090 /* Must be called with tasklist_lock held */
9091 static inline int tg_has_rt_tasks(struct task_group *tg)
9093 struct task_struct *g, *p;
9095 do_each_thread(g, p) {
9096 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9097 return 1;
9098 } while_each_thread(g, p);
9100 return 0;
9103 struct rt_schedulable_data {
9104 struct task_group *tg;
9105 u64 rt_period;
9106 u64 rt_runtime;
9109 static int tg_schedulable(struct task_group *tg, void *data)
9111 struct rt_schedulable_data *d = data;
9112 struct task_group *child;
9113 unsigned long total, sum = 0;
9114 u64 period, runtime;
9116 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9117 runtime = tg->rt_bandwidth.rt_runtime;
9119 if (tg == d->tg) {
9120 period = d->rt_period;
9121 runtime = d->rt_runtime;
9124 #ifdef CONFIG_USER_SCHED
9125 if (tg == &root_task_group) {
9126 period = global_rt_period();
9127 runtime = global_rt_runtime();
9129 #endif
9132 * Cannot have more runtime than the period.
9134 if (runtime > period && runtime != RUNTIME_INF)
9135 return -EINVAL;
9138 * Ensure we don't starve existing RT tasks.
9140 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9141 return -EBUSY;
9143 total = to_ratio(period, runtime);
9146 * Nobody can have more than the global setting allows.
9148 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9149 return -EINVAL;
9152 * The sum of our children's runtime should not exceed our own.
9154 list_for_each_entry_rcu(child, &tg->children, siblings) {
9155 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9156 runtime = child->rt_bandwidth.rt_runtime;
9158 if (child == d->tg) {
9159 period = d->rt_period;
9160 runtime = d->rt_runtime;
9163 sum += to_ratio(period, runtime);
9166 if (sum > total)
9167 return -EINVAL;
9169 return 0;
9172 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9174 struct rt_schedulable_data data = {
9175 .tg = tg,
9176 .rt_period = period,
9177 .rt_runtime = runtime,
9180 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9183 static int tg_set_bandwidth(struct task_group *tg,
9184 u64 rt_period, u64 rt_runtime)
9186 int i, err = 0;
9188 mutex_lock(&rt_constraints_mutex);
9189 read_lock(&tasklist_lock);
9190 err = __rt_schedulable(tg, rt_period, rt_runtime);
9191 if (err)
9192 goto unlock;
9194 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9195 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9196 tg->rt_bandwidth.rt_runtime = rt_runtime;
9198 for_each_possible_cpu(i) {
9199 struct rt_rq *rt_rq = tg->rt_rq[i];
9201 spin_lock(&rt_rq->rt_runtime_lock);
9202 rt_rq->rt_runtime = rt_runtime;
9203 spin_unlock(&rt_rq->rt_runtime_lock);
9205 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9206 unlock:
9207 read_unlock(&tasklist_lock);
9208 mutex_unlock(&rt_constraints_mutex);
9210 return err;
9213 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9215 u64 rt_runtime, rt_period;
9217 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9218 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9219 if (rt_runtime_us < 0)
9220 rt_runtime = RUNTIME_INF;
9222 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9225 long sched_group_rt_runtime(struct task_group *tg)
9227 u64 rt_runtime_us;
9229 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9230 return -1;
9232 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9233 do_div(rt_runtime_us, NSEC_PER_USEC);
9234 return rt_runtime_us;
9237 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9239 u64 rt_runtime, rt_period;
9241 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9242 rt_runtime = tg->rt_bandwidth.rt_runtime;
9244 if (rt_period == 0)
9245 return -EINVAL;
9247 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9250 long sched_group_rt_period(struct task_group *tg)
9252 u64 rt_period_us;
9254 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9255 do_div(rt_period_us, NSEC_PER_USEC);
9256 return rt_period_us;
9259 static int sched_rt_global_constraints(void)
9261 u64 runtime, period;
9262 int ret = 0;
9264 if (sysctl_sched_rt_period <= 0)
9265 return -EINVAL;
9267 runtime = global_rt_runtime();
9268 period = global_rt_period();
9271 * Sanity check on the sysctl variables.
9273 if (runtime > period && runtime != RUNTIME_INF)
9274 return -EINVAL;
9276 mutex_lock(&rt_constraints_mutex);
9277 read_lock(&tasklist_lock);
9278 ret = __rt_schedulable(NULL, 0, 0);
9279 read_unlock(&tasklist_lock);
9280 mutex_unlock(&rt_constraints_mutex);
9282 return ret;
9285 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9287 /* Don't accept realtime tasks when there is no way for them to run */
9288 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9289 return 0;
9291 return 1;
9294 #else /* !CONFIG_RT_GROUP_SCHED */
9295 static int sched_rt_global_constraints(void)
9297 unsigned long flags;
9298 int i;
9300 if (sysctl_sched_rt_period <= 0)
9301 return -EINVAL;
9303 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9304 for_each_possible_cpu(i) {
9305 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9307 spin_lock(&rt_rq->rt_runtime_lock);
9308 rt_rq->rt_runtime = global_rt_runtime();
9309 spin_unlock(&rt_rq->rt_runtime_lock);
9311 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9313 return 0;
9315 #endif /* CONFIG_RT_GROUP_SCHED */
9317 int sched_rt_handler(struct ctl_table *table, int write,
9318 struct file *filp, void __user *buffer, size_t *lenp,
9319 loff_t *ppos)
9321 int ret;
9322 int old_period, old_runtime;
9323 static DEFINE_MUTEX(mutex);
9325 mutex_lock(&mutex);
9326 old_period = sysctl_sched_rt_period;
9327 old_runtime = sysctl_sched_rt_runtime;
9329 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9331 if (!ret && write) {
9332 ret = sched_rt_global_constraints();
9333 if (ret) {
9334 sysctl_sched_rt_period = old_period;
9335 sysctl_sched_rt_runtime = old_runtime;
9336 } else {
9337 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9338 def_rt_bandwidth.rt_period =
9339 ns_to_ktime(global_rt_period());
9342 mutex_unlock(&mutex);
9344 return ret;
9347 #ifdef CONFIG_CGROUP_SCHED
9349 /* return corresponding task_group object of a cgroup */
9350 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9352 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9353 struct task_group, css);
9356 static struct cgroup_subsys_state *
9357 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9359 struct task_group *tg, *parent;
9361 if (!cgrp->parent) {
9362 /* This is early initialization for the top cgroup */
9363 return &init_task_group.css;
9366 parent = cgroup_tg(cgrp->parent);
9367 tg = sched_create_group(parent);
9368 if (IS_ERR(tg))
9369 return ERR_PTR(-ENOMEM);
9371 return &tg->css;
9374 static void
9375 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9377 struct task_group *tg = cgroup_tg(cgrp);
9379 sched_destroy_group(tg);
9382 static int
9383 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9384 struct task_struct *tsk)
9386 #ifdef CONFIG_RT_GROUP_SCHED
9387 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9388 return -EINVAL;
9389 #else
9390 /* We don't support RT-tasks being in separate groups */
9391 if (tsk->sched_class != &fair_sched_class)
9392 return -EINVAL;
9393 #endif
9395 return 0;
9398 static void
9399 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9400 struct cgroup *old_cont, struct task_struct *tsk)
9402 sched_move_task(tsk);
9405 #ifdef CONFIG_FAIR_GROUP_SCHED
9406 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9407 u64 shareval)
9409 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9412 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9414 struct task_group *tg = cgroup_tg(cgrp);
9416 return (u64) tg->shares;
9418 #endif /* CONFIG_FAIR_GROUP_SCHED */
9420 #ifdef CONFIG_RT_GROUP_SCHED
9421 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9422 s64 val)
9424 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9427 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9429 return sched_group_rt_runtime(cgroup_tg(cgrp));
9432 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9433 u64 rt_period_us)
9435 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9438 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9440 return sched_group_rt_period(cgroup_tg(cgrp));
9442 #endif /* CONFIG_RT_GROUP_SCHED */
9444 static struct cftype cpu_files[] = {
9445 #ifdef CONFIG_FAIR_GROUP_SCHED
9447 .name = "shares",
9448 .read_u64 = cpu_shares_read_u64,
9449 .write_u64 = cpu_shares_write_u64,
9451 #endif
9452 #ifdef CONFIG_RT_GROUP_SCHED
9454 .name = "rt_runtime_us",
9455 .read_s64 = cpu_rt_runtime_read,
9456 .write_s64 = cpu_rt_runtime_write,
9459 .name = "rt_period_us",
9460 .read_u64 = cpu_rt_period_read_uint,
9461 .write_u64 = cpu_rt_period_write_uint,
9463 #endif
9466 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9468 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9471 struct cgroup_subsys cpu_cgroup_subsys = {
9472 .name = "cpu",
9473 .create = cpu_cgroup_create,
9474 .destroy = cpu_cgroup_destroy,
9475 .can_attach = cpu_cgroup_can_attach,
9476 .attach = cpu_cgroup_attach,
9477 .populate = cpu_cgroup_populate,
9478 .subsys_id = cpu_cgroup_subsys_id,
9479 .early_init = 1,
9482 #endif /* CONFIG_CGROUP_SCHED */
9484 #ifdef CONFIG_CGROUP_CPUACCT
9487 * CPU accounting code for task groups.
9489 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9490 * (balbir@in.ibm.com).
9493 /* track cpu usage of a group of tasks and its child groups */
9494 struct cpuacct {
9495 struct cgroup_subsys_state css;
9496 /* cpuusage holds pointer to a u64-type object on every cpu */
9497 u64 *cpuusage;
9498 struct cpuacct *parent;
9501 struct cgroup_subsys cpuacct_subsys;
9503 /* return cpu accounting group corresponding to this container */
9504 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9506 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9507 struct cpuacct, css);
9510 /* return cpu accounting group to which this task belongs */
9511 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9513 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9514 struct cpuacct, css);
9517 /* create a new cpu accounting group */
9518 static struct cgroup_subsys_state *cpuacct_create(
9519 struct cgroup_subsys *ss, struct cgroup *cgrp)
9521 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9523 if (!ca)
9524 return ERR_PTR(-ENOMEM);
9526 ca->cpuusage = alloc_percpu(u64);
9527 if (!ca->cpuusage) {
9528 kfree(ca);
9529 return ERR_PTR(-ENOMEM);
9532 if (cgrp->parent)
9533 ca->parent = cgroup_ca(cgrp->parent);
9535 return &ca->css;
9538 /* destroy an existing cpu accounting group */
9539 static void
9540 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9542 struct cpuacct *ca = cgroup_ca(cgrp);
9544 free_percpu(ca->cpuusage);
9545 kfree(ca);
9548 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9550 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9551 u64 data;
9553 #ifndef CONFIG_64BIT
9555 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9557 spin_lock_irq(&cpu_rq(cpu)->lock);
9558 data = *cpuusage;
9559 spin_unlock_irq(&cpu_rq(cpu)->lock);
9560 #else
9561 data = *cpuusage;
9562 #endif
9564 return data;
9567 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9569 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9571 #ifndef CONFIG_64BIT
9573 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9575 spin_lock_irq(&cpu_rq(cpu)->lock);
9576 *cpuusage = val;
9577 spin_unlock_irq(&cpu_rq(cpu)->lock);
9578 #else
9579 *cpuusage = val;
9580 #endif
9583 /* return total cpu usage (in nanoseconds) of a group */
9584 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9586 struct cpuacct *ca = cgroup_ca(cgrp);
9587 u64 totalcpuusage = 0;
9588 int i;
9590 for_each_present_cpu(i)
9591 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9593 return totalcpuusage;
9596 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9597 u64 reset)
9599 struct cpuacct *ca = cgroup_ca(cgrp);
9600 int err = 0;
9601 int i;
9603 if (reset) {
9604 err = -EINVAL;
9605 goto out;
9608 for_each_present_cpu(i)
9609 cpuacct_cpuusage_write(ca, i, 0);
9611 out:
9612 return err;
9615 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9616 struct seq_file *m)
9618 struct cpuacct *ca = cgroup_ca(cgroup);
9619 u64 percpu;
9620 int i;
9622 for_each_present_cpu(i) {
9623 percpu = cpuacct_cpuusage_read(ca, i);
9624 seq_printf(m, "%llu ", (unsigned long long) percpu);
9626 seq_printf(m, "\n");
9627 return 0;
9630 static struct cftype files[] = {
9632 .name = "usage",
9633 .read_u64 = cpuusage_read,
9634 .write_u64 = cpuusage_write,
9637 .name = "usage_percpu",
9638 .read_seq_string = cpuacct_percpu_seq_read,
9643 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9645 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9649 * charge this task's execution time to its accounting group.
9651 * called with rq->lock held.
9653 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9655 struct cpuacct *ca;
9656 int cpu;
9658 if (!cpuacct_subsys.active)
9659 return;
9661 cpu = task_cpu(tsk);
9662 ca = task_ca(tsk);
9664 for (; ca; ca = ca->parent) {
9665 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9666 *cpuusage += cputime;
9670 struct cgroup_subsys cpuacct_subsys = {
9671 .name = "cpuacct",
9672 .create = cpuacct_create,
9673 .destroy = cpuacct_destroy,
9674 .populate = cpuacct_populate,
9675 .subsys_id = cpuacct_subsys_id,
9677 #endif /* CONFIG_CGROUP_CPUACCT */