ethtool: Move n-tuple capability check into set_flags
[linux-2.6/libata-dev.git] / kernel / sched.c
blob4508fe7048be5d91ee5a01daef3fe9de8d37b909
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/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
75 #include <asm/tlb.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
86 * and back.
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy)
124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
125 return 1;
126 return 0;
129 static inline int task_has_rt_policy(struct task_struct *p)
131 return rt_policy(p->policy);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
142 struct rt_bandwidth {
143 /* nests inside the rq lock: */
144 raw_spinlock_t rt_runtime_lock;
145 ktime_t rt_period;
146 u64 rt_runtime;
147 struct hrtimer rt_period_timer;
150 static struct rt_bandwidth def_rt_bandwidth;
152 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
154 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
158 ktime_t now;
159 int overrun;
160 int idle = 0;
162 for (;;) {
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
166 if (!overrun)
167 break;
169 idle = do_sched_rt_period_timer(rt_b, overrun);
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
175 static
176 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
181 raw_spin_lock_init(&rt_b->rt_runtime_lock);
183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime >= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
195 ktime_t now;
197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
198 return;
200 if (hrtimer_active(&rt_b->rt_period_timer))
201 return;
203 raw_spin_lock(&rt_b->rt_runtime_lock);
204 for (;;) {
205 unsigned long delta;
206 ktime_t soft, hard;
208 if (hrtimer_active(&rt_b->rt_period_timer))
209 break;
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
218 HRTIMER_MODE_ABS_PINNED, 0);
220 raw_spin_unlock(&rt_b->rt_runtime_lock);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
226 hrtimer_cancel(&rt_b->rt_period_timer);
228 #endif
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex);
236 #ifdef CONFIG_GROUP_SCHED
238 #include <linux/cgroup.h>
240 struct cfs_rq;
242 static LIST_HEAD(task_groups);
244 /* task group related information */
245 struct task_group {
246 #ifdef CONFIG_CGROUP_SCHED
247 struct cgroup_subsys_state css;
248 #endif
250 #ifdef CONFIG_USER_SCHED
251 uid_t uid;
252 #endif
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
260 #endif
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
266 struct rt_bandwidth rt_bandwidth;
267 #endif
269 struct rcu_head rcu;
270 struct list_head list;
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
277 #ifdef CONFIG_USER_SCHED
279 /* Helper function to pass uid information to create_sched_user() */
280 void set_tg_uid(struct user_struct *user)
282 user->tg->uid = user->uid;
286 * Root task group.
287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
290 struct task_group root_task_group;
292 #ifdef CONFIG_FAIR_GROUP_SCHED
293 /* Default task group's sched entity on each cpu */
294 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
295 /* Default task group's cfs_rq on each cpu */
296 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq, init_tg_cfs_rq);
297 #endif /* CONFIG_FAIR_GROUP_SCHED */
299 #ifdef CONFIG_RT_GROUP_SCHED
300 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
301 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq, init_rt_rq_var);
302 #endif /* CONFIG_RT_GROUP_SCHED */
303 #else /* !CONFIG_USER_SCHED */
304 #define root_task_group init_task_group
305 #endif /* CONFIG_USER_SCHED */
307 /* task_group_lock serializes add/remove of task groups and also changes to
308 * a task group's cpu shares.
310 static DEFINE_SPINLOCK(task_group_lock);
312 #ifdef CONFIG_FAIR_GROUP_SCHED
314 #ifdef CONFIG_SMP
315 static int root_task_group_empty(void)
317 return list_empty(&root_task_group.children);
319 #endif
321 #ifdef CONFIG_USER_SCHED
322 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
323 #else /* !CONFIG_USER_SCHED */
324 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
325 #endif /* CONFIG_USER_SCHED */
328 * A weight of 0 or 1 can cause arithmetics problems.
329 * A weight of a cfs_rq is the sum of weights of which entities
330 * are queued on this cfs_rq, so a weight of a entity should not be
331 * too large, so as the shares value of a task group.
332 * (The default weight is 1024 - so there's no practical
333 * limitation from this.)
335 #define MIN_SHARES 2
336 #define MAX_SHARES (1UL << 18)
338 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
339 #endif
341 /* Default task group.
342 * Every task in system belong to this group at bootup.
344 struct task_group init_task_group;
346 /* return group to which a task belongs */
347 static inline struct task_group *task_group(struct task_struct *p)
349 struct task_group *tg;
351 #ifdef CONFIG_USER_SCHED
352 rcu_read_lock();
353 tg = __task_cred(p)->user->tg;
354 rcu_read_unlock();
355 #elif defined(CONFIG_CGROUP_SCHED)
356 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
357 struct task_group, css);
358 #else
359 tg = &init_task_group;
360 #endif
361 return tg;
364 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
365 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
367 #ifdef CONFIG_FAIR_GROUP_SCHED
368 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
369 p->se.parent = task_group(p)->se[cpu];
370 #endif
372 #ifdef CONFIG_RT_GROUP_SCHED
373 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
374 p->rt.parent = task_group(p)->rt_se[cpu];
375 #endif
378 #else
380 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
381 static inline struct task_group *task_group(struct task_struct *p)
383 return NULL;
386 #endif /* CONFIG_GROUP_SCHED */
388 /* CFS-related fields in a runqueue */
389 struct cfs_rq {
390 struct load_weight load;
391 unsigned long nr_running;
393 u64 exec_clock;
394 u64 min_vruntime;
396 struct rb_root tasks_timeline;
397 struct rb_node *rb_leftmost;
399 struct list_head tasks;
400 struct list_head *balance_iterator;
403 * 'curr' points to currently running entity on this cfs_rq.
404 * It is set to NULL otherwise (i.e when none are currently running).
406 struct sched_entity *curr, *next, *last;
408 unsigned int nr_spread_over;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
421 struct list_head leaf_cfs_rq_list;
422 struct task_group *tg; /* group that "owns" this runqueue */
424 #ifdef CONFIG_SMP
426 * the part of load.weight contributed by tasks
428 unsigned long task_weight;
431 * h_load = weight * f(tg)
433 * Where f(tg) is the recursive weight fraction assigned to
434 * this group.
436 unsigned long h_load;
439 * this cpu's part of tg->shares
441 unsigned long shares;
444 * load.weight at the time we set shares
446 unsigned long rq_weight;
447 #endif
448 #endif
451 /* Real-Time classes' related field in a runqueue: */
452 struct rt_rq {
453 struct rt_prio_array active;
454 unsigned long rt_nr_running;
455 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
456 struct {
457 int curr; /* highest queued rt task prio */
458 #ifdef CONFIG_SMP
459 int next; /* next highest */
460 #endif
461 } highest_prio;
462 #endif
463 #ifdef CONFIG_SMP
464 unsigned long rt_nr_migratory;
465 unsigned long rt_nr_total;
466 int overloaded;
467 struct plist_head pushable_tasks;
468 #endif
469 int rt_throttled;
470 u64 rt_time;
471 u64 rt_runtime;
472 /* Nests inside the rq lock: */
473 raw_spinlock_t rt_runtime_lock;
475 #ifdef CONFIG_RT_GROUP_SCHED
476 unsigned long rt_nr_boosted;
478 struct rq *rq;
479 struct list_head leaf_rt_rq_list;
480 struct task_group *tg;
481 struct sched_rt_entity *rt_se;
482 #endif
485 #ifdef CONFIG_SMP
488 * We add the notion of a root-domain which will be used to define per-domain
489 * variables. Each exclusive cpuset essentially defines an island domain by
490 * fully partitioning the member cpus from any other cpuset. Whenever a new
491 * exclusive cpuset is created, we also create and attach a new root-domain
492 * object.
495 struct root_domain {
496 atomic_t refcount;
497 cpumask_var_t span;
498 cpumask_var_t online;
501 * The "RT overload" flag: it gets set if a CPU has more than
502 * one runnable RT task.
504 cpumask_var_t rto_mask;
505 atomic_t rto_count;
506 #ifdef CONFIG_SMP
507 struct cpupri cpupri;
508 #endif
512 * By default the system creates a single root-domain with all cpus as
513 * members (mimicking the global state we have today).
515 static struct root_domain def_root_domain;
517 #endif
520 * This is the main, per-CPU runqueue data structure.
522 * Locking rule: those places that want to lock multiple runqueues
523 * (such as the load balancing or the thread migration code), lock
524 * acquire operations must be ordered by ascending &runqueue.
526 struct rq {
527 /* runqueue lock: */
528 raw_spinlock_t lock;
531 * nr_running and cpu_load should be in the same cacheline because
532 * remote CPUs use both these fields when doing load calculation.
534 unsigned long nr_running;
535 #define CPU_LOAD_IDX_MAX 5
536 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
537 #ifdef CONFIG_NO_HZ
538 unsigned char in_nohz_recently;
539 #endif
540 /* capture load from *all* tasks on this cpu: */
541 struct load_weight load;
542 unsigned long nr_load_updates;
543 u64 nr_switches;
545 struct cfs_rq cfs;
546 struct rt_rq rt;
548 #ifdef CONFIG_FAIR_GROUP_SCHED
549 /* list of leaf cfs_rq on this cpu: */
550 struct list_head leaf_cfs_rq_list;
551 #endif
552 #ifdef CONFIG_RT_GROUP_SCHED
553 struct list_head leaf_rt_rq_list;
554 #endif
557 * This is part of a global counter where only the total sum
558 * over all CPUs matters. A task can increase this counter on
559 * one CPU and if it got migrated afterwards it may decrease
560 * it on another CPU. Always updated under the runqueue lock:
562 unsigned long nr_uninterruptible;
564 struct task_struct *curr, *idle;
565 unsigned long next_balance;
566 struct mm_struct *prev_mm;
568 u64 clock;
570 atomic_t nr_iowait;
572 #ifdef CONFIG_SMP
573 struct root_domain *rd;
574 struct sched_domain *sd;
576 unsigned char idle_at_tick;
577 /* For active balancing */
578 int post_schedule;
579 int active_balance;
580 int push_cpu;
581 /* cpu of this runqueue: */
582 int cpu;
583 int online;
585 unsigned long avg_load_per_task;
587 struct task_struct *migration_thread;
588 struct list_head migration_queue;
590 u64 rt_avg;
591 u64 age_stamp;
592 u64 idle_stamp;
593 u64 avg_idle;
594 #endif
596 /* calc_load related fields */
597 unsigned long calc_load_update;
598 long calc_load_active;
600 #ifdef CONFIG_SCHED_HRTICK
601 #ifdef CONFIG_SMP
602 int hrtick_csd_pending;
603 struct call_single_data hrtick_csd;
604 #endif
605 struct hrtimer hrtick_timer;
606 #endif
608 #ifdef CONFIG_SCHEDSTATS
609 /* latency stats */
610 struct sched_info rq_sched_info;
611 unsigned long long rq_cpu_time;
612 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
614 /* sys_sched_yield() stats */
615 unsigned int yld_count;
617 /* schedule() stats */
618 unsigned int sched_switch;
619 unsigned int sched_count;
620 unsigned int sched_goidle;
622 /* try_to_wake_up() stats */
623 unsigned int ttwu_count;
624 unsigned int ttwu_local;
626 /* BKL stats */
627 unsigned int bkl_count;
628 #endif
631 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
633 static inline
634 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
636 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
639 static inline int cpu_of(struct rq *rq)
641 #ifdef CONFIG_SMP
642 return rq->cpu;
643 #else
644 return 0;
645 #endif
649 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
650 * See detach_destroy_domains: synchronize_sched for details.
652 * The domain tree of any CPU may only be accessed from within
653 * preempt-disabled sections.
655 #define for_each_domain(cpu, __sd) \
656 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
658 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
659 #define this_rq() (&__get_cpu_var(runqueues))
660 #define task_rq(p) cpu_rq(task_cpu(p))
661 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
662 #define raw_rq() (&__raw_get_cpu_var(runqueues))
664 inline void update_rq_clock(struct rq *rq)
666 rq->clock = sched_clock_cpu(cpu_of(rq));
670 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
672 #ifdef CONFIG_SCHED_DEBUG
673 # define const_debug __read_mostly
674 #else
675 # define const_debug static const
676 #endif
679 * runqueue_is_locked
680 * @cpu: the processor in question.
682 * Returns true if the current cpu runqueue is locked.
683 * This interface allows printk to be called with the runqueue lock
684 * held and know whether or not it is OK to wake up the klogd.
686 int runqueue_is_locked(int cpu)
688 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
692 * Debugging: various feature bits
695 #define SCHED_FEAT(name, enabled) \
696 __SCHED_FEAT_##name ,
698 enum {
699 #include "sched_features.h"
702 #undef SCHED_FEAT
704 #define SCHED_FEAT(name, enabled) \
705 (1UL << __SCHED_FEAT_##name) * enabled |
707 const_debug unsigned int sysctl_sched_features =
708 #include "sched_features.h"
711 #undef SCHED_FEAT
713 #ifdef CONFIG_SCHED_DEBUG
714 #define SCHED_FEAT(name, enabled) \
715 #name ,
717 static __read_mostly char *sched_feat_names[] = {
718 #include "sched_features.h"
719 NULL
722 #undef SCHED_FEAT
724 static int sched_feat_show(struct seq_file *m, void *v)
726 int i;
728 for (i = 0; sched_feat_names[i]; i++) {
729 if (!(sysctl_sched_features & (1UL << i)))
730 seq_puts(m, "NO_");
731 seq_printf(m, "%s ", sched_feat_names[i]);
733 seq_puts(m, "\n");
735 return 0;
738 static ssize_t
739 sched_feat_write(struct file *filp, const char __user *ubuf,
740 size_t cnt, loff_t *ppos)
742 char buf[64];
743 char *cmp = buf;
744 int neg = 0;
745 int i;
747 if (cnt > 63)
748 cnt = 63;
750 if (copy_from_user(&buf, ubuf, cnt))
751 return -EFAULT;
753 buf[cnt] = 0;
755 if (strncmp(buf, "NO_", 3) == 0) {
756 neg = 1;
757 cmp += 3;
760 for (i = 0; sched_feat_names[i]; i++) {
761 int len = strlen(sched_feat_names[i]);
763 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
764 if (neg)
765 sysctl_sched_features &= ~(1UL << i);
766 else
767 sysctl_sched_features |= (1UL << i);
768 break;
772 if (!sched_feat_names[i])
773 return -EINVAL;
775 *ppos += cnt;
777 return cnt;
780 static int sched_feat_open(struct inode *inode, struct file *filp)
782 return single_open(filp, sched_feat_show, NULL);
785 static const struct file_operations sched_feat_fops = {
786 .open = sched_feat_open,
787 .write = sched_feat_write,
788 .read = seq_read,
789 .llseek = seq_lseek,
790 .release = single_release,
793 static __init int sched_init_debug(void)
795 debugfs_create_file("sched_features", 0644, NULL, NULL,
796 &sched_feat_fops);
798 return 0;
800 late_initcall(sched_init_debug);
802 #endif
804 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
807 * Number of tasks to iterate in a single balance run.
808 * Limited because this is done with IRQs disabled.
810 const_debug unsigned int sysctl_sched_nr_migrate = 32;
813 * ratelimit for updating the group shares.
814 * default: 0.25ms
816 unsigned int sysctl_sched_shares_ratelimit = 250000;
817 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
820 * Inject some fuzzyness into changing the per-cpu group shares
821 * this avoids remote rq-locks at the expense of fairness.
822 * default: 4
824 unsigned int sysctl_sched_shares_thresh = 4;
827 * period over which we average the RT time consumption, measured
828 * in ms.
830 * default: 1s
832 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
835 * period over which we measure -rt task cpu usage in us.
836 * default: 1s
838 unsigned int sysctl_sched_rt_period = 1000000;
840 static __read_mostly int scheduler_running;
843 * part of the period that we allow rt tasks to run in us.
844 * default: 0.95s
846 int sysctl_sched_rt_runtime = 950000;
848 static inline u64 global_rt_period(void)
850 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
853 static inline u64 global_rt_runtime(void)
855 if (sysctl_sched_rt_runtime < 0)
856 return RUNTIME_INF;
858 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
861 #ifndef prepare_arch_switch
862 # define prepare_arch_switch(next) do { } while (0)
863 #endif
864 #ifndef finish_arch_switch
865 # define finish_arch_switch(prev) do { } while (0)
866 #endif
868 static inline int task_current(struct rq *rq, struct task_struct *p)
870 return rq->curr == p;
873 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
874 static inline int task_running(struct rq *rq, struct task_struct *p)
876 return task_current(rq, p);
879 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
883 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
885 #ifdef CONFIG_DEBUG_SPINLOCK
886 /* this is a valid case when another task releases the spinlock */
887 rq->lock.owner = current;
888 #endif
890 * If we are tracking spinlock dependencies then we have to
891 * fix up the runqueue lock - which gets 'carried over' from
892 * prev into current:
894 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
896 raw_spin_unlock_irq(&rq->lock);
899 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
900 static inline int task_running(struct rq *rq, struct task_struct *p)
902 #ifdef CONFIG_SMP
903 return p->oncpu;
904 #else
905 return task_current(rq, p);
906 #endif
909 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
911 #ifdef CONFIG_SMP
913 * We can optimise this out completely for !SMP, because the
914 * SMP rebalancing from interrupt is the only thing that cares
915 * here.
917 next->oncpu = 1;
918 #endif
919 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
920 raw_spin_unlock_irq(&rq->lock);
921 #else
922 raw_spin_unlock(&rq->lock);
923 #endif
926 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
928 #ifdef CONFIG_SMP
930 * After ->oncpu is cleared, the task can be moved to a different CPU.
931 * We must ensure this doesn't happen until the switch is completely
932 * finished.
934 smp_wmb();
935 prev->oncpu = 0;
936 #endif
937 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
938 local_irq_enable();
939 #endif
941 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
944 * __task_rq_lock - lock the runqueue a given task resides on.
945 * Must be called interrupts disabled.
947 static inline struct rq *__task_rq_lock(struct task_struct *p)
948 __acquires(rq->lock)
950 for (;;) {
951 struct rq *rq = task_rq(p);
952 raw_spin_lock(&rq->lock);
953 if (likely(rq == task_rq(p)))
954 return rq;
955 raw_spin_unlock(&rq->lock);
960 * task_rq_lock - lock the runqueue a given task resides on and disable
961 * interrupts. Note the ordering: we can safely lookup the task_rq without
962 * explicitly disabling preemption.
964 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
965 __acquires(rq->lock)
967 struct rq *rq;
969 for (;;) {
970 local_irq_save(*flags);
971 rq = task_rq(p);
972 raw_spin_lock(&rq->lock);
973 if (likely(rq == task_rq(p)))
974 return rq;
975 raw_spin_unlock_irqrestore(&rq->lock, *flags);
979 void task_rq_unlock_wait(struct task_struct *p)
981 struct rq *rq = task_rq(p);
983 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
984 raw_spin_unlock_wait(&rq->lock);
987 static void __task_rq_unlock(struct rq *rq)
988 __releases(rq->lock)
990 raw_spin_unlock(&rq->lock);
993 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
994 __releases(rq->lock)
996 raw_spin_unlock_irqrestore(&rq->lock, *flags);
1000 * this_rq_lock - lock this runqueue and disable interrupts.
1002 static struct rq *this_rq_lock(void)
1003 __acquires(rq->lock)
1005 struct rq *rq;
1007 local_irq_disable();
1008 rq = this_rq();
1009 raw_spin_lock(&rq->lock);
1011 return rq;
1014 #ifdef CONFIG_SCHED_HRTICK
1016 * Use HR-timers to deliver accurate preemption points.
1018 * Its all a bit involved since we cannot program an hrt while holding the
1019 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1020 * reschedule event.
1022 * When we get rescheduled we reprogram the hrtick_timer outside of the
1023 * rq->lock.
1027 * Use hrtick when:
1028 * - enabled by features
1029 * - hrtimer is actually high res
1031 static inline int hrtick_enabled(struct rq *rq)
1033 if (!sched_feat(HRTICK))
1034 return 0;
1035 if (!cpu_active(cpu_of(rq)))
1036 return 0;
1037 return hrtimer_is_hres_active(&rq->hrtick_timer);
1040 static void hrtick_clear(struct rq *rq)
1042 if (hrtimer_active(&rq->hrtick_timer))
1043 hrtimer_cancel(&rq->hrtick_timer);
1047 * High-resolution timer tick.
1048 * Runs from hardirq context with interrupts disabled.
1050 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1052 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1054 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1056 raw_spin_lock(&rq->lock);
1057 update_rq_clock(rq);
1058 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1059 raw_spin_unlock(&rq->lock);
1061 return HRTIMER_NORESTART;
1064 #ifdef CONFIG_SMP
1066 * called from hardirq (IPI) context
1068 static void __hrtick_start(void *arg)
1070 struct rq *rq = arg;
1072 raw_spin_lock(&rq->lock);
1073 hrtimer_restart(&rq->hrtick_timer);
1074 rq->hrtick_csd_pending = 0;
1075 raw_spin_unlock(&rq->lock);
1079 * Called to set the hrtick timer state.
1081 * called with rq->lock held and irqs disabled
1083 static void hrtick_start(struct rq *rq, u64 delay)
1085 struct hrtimer *timer = &rq->hrtick_timer;
1086 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1088 hrtimer_set_expires(timer, time);
1090 if (rq == this_rq()) {
1091 hrtimer_restart(timer);
1092 } else if (!rq->hrtick_csd_pending) {
1093 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1094 rq->hrtick_csd_pending = 1;
1098 static int
1099 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1101 int cpu = (int)(long)hcpu;
1103 switch (action) {
1104 case CPU_UP_CANCELED:
1105 case CPU_UP_CANCELED_FROZEN:
1106 case CPU_DOWN_PREPARE:
1107 case CPU_DOWN_PREPARE_FROZEN:
1108 case CPU_DEAD:
1109 case CPU_DEAD_FROZEN:
1110 hrtick_clear(cpu_rq(cpu));
1111 return NOTIFY_OK;
1114 return NOTIFY_DONE;
1117 static __init void init_hrtick(void)
1119 hotcpu_notifier(hotplug_hrtick, 0);
1121 #else
1123 * Called to set the hrtick timer state.
1125 * called with rq->lock held and irqs disabled
1127 static void hrtick_start(struct rq *rq, u64 delay)
1129 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1130 HRTIMER_MODE_REL_PINNED, 0);
1133 static inline void init_hrtick(void)
1136 #endif /* CONFIG_SMP */
1138 static void init_rq_hrtick(struct rq *rq)
1140 #ifdef CONFIG_SMP
1141 rq->hrtick_csd_pending = 0;
1143 rq->hrtick_csd.flags = 0;
1144 rq->hrtick_csd.func = __hrtick_start;
1145 rq->hrtick_csd.info = rq;
1146 #endif
1148 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1149 rq->hrtick_timer.function = hrtick;
1151 #else /* CONFIG_SCHED_HRTICK */
1152 static inline void hrtick_clear(struct rq *rq)
1156 static inline void init_rq_hrtick(struct rq *rq)
1160 static inline void init_hrtick(void)
1163 #endif /* CONFIG_SCHED_HRTICK */
1166 * resched_task - mark a task 'to be rescheduled now'.
1168 * On UP this means the setting of the need_resched flag, on SMP it
1169 * might also involve a cross-CPU call to trigger the scheduler on
1170 * the target CPU.
1172 #ifdef CONFIG_SMP
1174 #ifndef tsk_is_polling
1175 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1176 #endif
1178 static void resched_task(struct task_struct *p)
1180 int cpu;
1182 assert_raw_spin_locked(&task_rq(p)->lock);
1184 if (test_tsk_need_resched(p))
1185 return;
1187 set_tsk_need_resched(p);
1189 cpu = task_cpu(p);
1190 if (cpu == smp_processor_id())
1191 return;
1193 /* NEED_RESCHED must be visible before we test polling */
1194 smp_mb();
1195 if (!tsk_is_polling(p))
1196 smp_send_reschedule(cpu);
1199 static void resched_cpu(int cpu)
1201 struct rq *rq = cpu_rq(cpu);
1202 unsigned long flags;
1204 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1205 return;
1206 resched_task(cpu_curr(cpu));
1207 raw_spin_unlock_irqrestore(&rq->lock, flags);
1210 #ifdef CONFIG_NO_HZ
1212 * When add_timer_on() enqueues a timer into the timer wheel of an
1213 * idle CPU then this timer might expire before the next timer event
1214 * which is scheduled to wake up that CPU. In case of a completely
1215 * idle system the next event might even be infinite time into the
1216 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1217 * leaves the inner idle loop so the newly added timer is taken into
1218 * account when the CPU goes back to idle and evaluates the timer
1219 * wheel for the next timer event.
1221 void wake_up_idle_cpu(int cpu)
1223 struct rq *rq = cpu_rq(cpu);
1225 if (cpu == smp_processor_id())
1226 return;
1229 * This is safe, as this function is called with the timer
1230 * wheel base lock of (cpu) held. When the CPU is on the way
1231 * to idle and has not yet set rq->curr to idle then it will
1232 * be serialized on the timer wheel base lock and take the new
1233 * timer into account automatically.
1235 if (rq->curr != rq->idle)
1236 return;
1239 * We can set TIF_RESCHED on the idle task of the other CPU
1240 * lockless. The worst case is that the other CPU runs the
1241 * idle task through an additional NOOP schedule()
1243 set_tsk_need_resched(rq->idle);
1245 /* NEED_RESCHED must be visible before we test polling */
1246 smp_mb();
1247 if (!tsk_is_polling(rq->idle))
1248 smp_send_reschedule(cpu);
1250 #endif /* CONFIG_NO_HZ */
1252 static u64 sched_avg_period(void)
1254 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1257 static void sched_avg_update(struct rq *rq)
1259 s64 period = sched_avg_period();
1261 while ((s64)(rq->clock - rq->age_stamp) > period) {
1262 rq->age_stamp += period;
1263 rq->rt_avg /= 2;
1267 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1269 rq->rt_avg += rt_delta;
1270 sched_avg_update(rq);
1273 #else /* !CONFIG_SMP */
1274 static void resched_task(struct task_struct *p)
1276 assert_raw_spin_locked(&task_rq(p)->lock);
1277 set_tsk_need_resched(p);
1280 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1283 #endif /* CONFIG_SMP */
1285 #if BITS_PER_LONG == 32
1286 # define WMULT_CONST (~0UL)
1287 #else
1288 # define WMULT_CONST (1UL << 32)
1289 #endif
1291 #define WMULT_SHIFT 32
1294 * Shift right and round:
1296 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1299 * delta *= weight / lw
1301 static unsigned long
1302 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1303 struct load_weight *lw)
1305 u64 tmp;
1307 if (!lw->inv_weight) {
1308 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1309 lw->inv_weight = 1;
1310 else
1311 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1312 / (lw->weight+1);
1315 tmp = (u64)delta_exec * weight;
1317 * Check whether we'd overflow the 64-bit multiplication:
1319 if (unlikely(tmp > WMULT_CONST))
1320 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1321 WMULT_SHIFT/2);
1322 else
1323 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1325 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1328 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1330 lw->weight += inc;
1331 lw->inv_weight = 0;
1334 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1336 lw->weight -= dec;
1337 lw->inv_weight = 0;
1341 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1342 * of tasks with abnormal "nice" values across CPUs the contribution that
1343 * each task makes to its run queue's load is weighted according to its
1344 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1345 * scaled version of the new time slice allocation that they receive on time
1346 * slice expiry etc.
1349 #define WEIGHT_IDLEPRIO 3
1350 #define WMULT_IDLEPRIO 1431655765
1353 * Nice levels are multiplicative, with a gentle 10% change for every
1354 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1355 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1356 * that remained on nice 0.
1358 * The "10% effect" is relative and cumulative: from _any_ nice level,
1359 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1360 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1361 * If a task goes up by ~10% and another task goes down by ~10% then
1362 * the relative distance between them is ~25%.)
1364 static const int prio_to_weight[40] = {
1365 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1366 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1367 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1368 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1369 /* 0 */ 1024, 820, 655, 526, 423,
1370 /* 5 */ 335, 272, 215, 172, 137,
1371 /* 10 */ 110, 87, 70, 56, 45,
1372 /* 15 */ 36, 29, 23, 18, 15,
1376 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1378 * In cases where the weight does not change often, we can use the
1379 * precalculated inverse to speed up arithmetics by turning divisions
1380 * into multiplications:
1382 static const u32 prio_to_wmult[40] = {
1383 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1384 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1385 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1386 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1387 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1388 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1389 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1390 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1393 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1396 * runqueue iterator, to support SMP load-balancing between different
1397 * scheduling classes, without having to expose their internal data
1398 * structures to the load-balancing proper:
1400 struct rq_iterator {
1401 void *arg;
1402 struct task_struct *(*start)(void *);
1403 struct task_struct *(*next)(void *);
1406 #ifdef CONFIG_SMP
1407 static unsigned long
1408 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1409 unsigned long max_load_move, struct sched_domain *sd,
1410 enum cpu_idle_type idle, int *all_pinned,
1411 int *this_best_prio, struct rq_iterator *iterator);
1413 static int
1414 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1415 struct sched_domain *sd, enum cpu_idle_type idle,
1416 struct rq_iterator *iterator);
1417 #endif
1419 /* Time spent by the tasks of the cpu accounting group executing in ... */
1420 enum cpuacct_stat_index {
1421 CPUACCT_STAT_USER, /* ... user mode */
1422 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1424 CPUACCT_STAT_NSTATS,
1427 #ifdef CONFIG_CGROUP_CPUACCT
1428 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1429 static void cpuacct_update_stats(struct task_struct *tsk,
1430 enum cpuacct_stat_index idx, cputime_t val);
1431 #else
1432 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1433 static inline void cpuacct_update_stats(struct task_struct *tsk,
1434 enum cpuacct_stat_index idx, cputime_t val) {}
1435 #endif
1437 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1439 update_load_add(&rq->load, load);
1442 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1444 update_load_sub(&rq->load, load);
1447 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1448 typedef int (*tg_visitor)(struct task_group *, void *);
1451 * Iterate the full tree, calling @down when first entering a node and @up when
1452 * leaving it for the final time.
1454 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1456 struct task_group *parent, *child;
1457 int ret;
1459 rcu_read_lock();
1460 parent = &root_task_group;
1461 down:
1462 ret = (*down)(parent, data);
1463 if (ret)
1464 goto out_unlock;
1465 list_for_each_entry_rcu(child, &parent->children, siblings) {
1466 parent = child;
1467 goto down;
1470 continue;
1472 ret = (*up)(parent, data);
1473 if (ret)
1474 goto out_unlock;
1476 child = parent;
1477 parent = parent->parent;
1478 if (parent)
1479 goto up;
1480 out_unlock:
1481 rcu_read_unlock();
1483 return ret;
1486 static int tg_nop(struct task_group *tg, void *data)
1488 return 0;
1490 #endif
1492 #ifdef CONFIG_SMP
1493 /* Used instead of source_load when we know the type == 0 */
1494 static unsigned long weighted_cpuload(const int cpu)
1496 return cpu_rq(cpu)->load.weight;
1500 * Return a low guess at the load of a migration-source cpu weighted
1501 * according to the scheduling class and "nice" value.
1503 * We want to under-estimate the load of migration sources, to
1504 * balance conservatively.
1506 static unsigned long source_load(int cpu, int type)
1508 struct rq *rq = cpu_rq(cpu);
1509 unsigned long total = weighted_cpuload(cpu);
1511 if (type == 0 || !sched_feat(LB_BIAS))
1512 return total;
1514 return min(rq->cpu_load[type-1], total);
1518 * Return a high guess at the load of a migration-target cpu weighted
1519 * according to the scheduling class and "nice" value.
1521 static unsigned long target_load(int cpu, int type)
1523 struct rq *rq = cpu_rq(cpu);
1524 unsigned long total = weighted_cpuload(cpu);
1526 if (type == 0 || !sched_feat(LB_BIAS))
1527 return total;
1529 return max(rq->cpu_load[type-1], total);
1532 static struct sched_group *group_of(int cpu)
1534 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1536 if (!sd)
1537 return NULL;
1539 return sd->groups;
1542 static unsigned long power_of(int cpu)
1544 struct sched_group *group = group_of(cpu);
1546 if (!group)
1547 return SCHED_LOAD_SCALE;
1549 return group->cpu_power;
1552 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1554 static unsigned long cpu_avg_load_per_task(int cpu)
1556 struct rq *rq = cpu_rq(cpu);
1557 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1559 if (nr_running)
1560 rq->avg_load_per_task = rq->load.weight / nr_running;
1561 else
1562 rq->avg_load_per_task = 0;
1564 return rq->avg_load_per_task;
1567 #ifdef CONFIG_FAIR_GROUP_SCHED
1569 static __read_mostly unsigned long *update_shares_data;
1571 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1574 * Calculate and set the cpu's group shares.
1576 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1577 unsigned long sd_shares,
1578 unsigned long sd_rq_weight,
1579 unsigned long *usd_rq_weight)
1581 unsigned long shares, rq_weight;
1582 int boost = 0;
1584 rq_weight = usd_rq_weight[cpu];
1585 if (!rq_weight) {
1586 boost = 1;
1587 rq_weight = NICE_0_LOAD;
1591 * \Sum_j shares_j * rq_weight_i
1592 * shares_i = -----------------------------
1593 * \Sum_j rq_weight_j
1595 shares = (sd_shares * rq_weight) / sd_rq_weight;
1596 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1598 if (abs(shares - tg->se[cpu]->load.weight) >
1599 sysctl_sched_shares_thresh) {
1600 struct rq *rq = cpu_rq(cpu);
1601 unsigned long flags;
1603 raw_spin_lock_irqsave(&rq->lock, flags);
1604 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1605 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1606 __set_se_shares(tg->se[cpu], shares);
1607 raw_spin_unlock_irqrestore(&rq->lock, flags);
1612 * Re-compute the task group their per cpu shares over the given domain.
1613 * This needs to be done in a bottom-up fashion because the rq weight of a
1614 * parent group depends on the shares of its child groups.
1616 static int tg_shares_up(struct task_group *tg, void *data)
1618 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1619 unsigned long *usd_rq_weight;
1620 struct sched_domain *sd = data;
1621 unsigned long flags;
1622 int i;
1624 if (!tg->se[0])
1625 return 0;
1627 local_irq_save(flags);
1628 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1630 for_each_cpu(i, sched_domain_span(sd)) {
1631 weight = tg->cfs_rq[i]->load.weight;
1632 usd_rq_weight[i] = weight;
1634 rq_weight += weight;
1636 * If there are currently no tasks on the cpu pretend there
1637 * is one of average load so that when a new task gets to
1638 * run here it will not get delayed by group starvation.
1640 if (!weight)
1641 weight = NICE_0_LOAD;
1643 sum_weight += weight;
1644 shares += tg->cfs_rq[i]->shares;
1647 if (!rq_weight)
1648 rq_weight = sum_weight;
1650 if ((!shares && rq_weight) || shares > tg->shares)
1651 shares = tg->shares;
1653 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1654 shares = tg->shares;
1656 for_each_cpu(i, sched_domain_span(sd))
1657 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1659 local_irq_restore(flags);
1661 return 0;
1665 * Compute the cpu's hierarchical load factor for each task group.
1666 * This needs to be done in a top-down fashion because the load of a child
1667 * group is a fraction of its parents load.
1669 static int tg_load_down(struct task_group *tg, void *data)
1671 unsigned long load;
1672 long cpu = (long)data;
1674 if (!tg->parent) {
1675 load = cpu_rq(cpu)->load.weight;
1676 } else {
1677 load = tg->parent->cfs_rq[cpu]->h_load;
1678 load *= tg->cfs_rq[cpu]->shares;
1679 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1682 tg->cfs_rq[cpu]->h_load = load;
1684 return 0;
1687 static void update_shares(struct sched_domain *sd)
1689 s64 elapsed;
1690 u64 now;
1692 if (root_task_group_empty())
1693 return;
1695 now = cpu_clock(raw_smp_processor_id());
1696 elapsed = now - sd->last_update;
1698 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1699 sd->last_update = now;
1700 walk_tg_tree(tg_nop, tg_shares_up, sd);
1704 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1706 if (root_task_group_empty())
1707 return;
1709 raw_spin_unlock(&rq->lock);
1710 update_shares(sd);
1711 raw_spin_lock(&rq->lock);
1714 static void update_h_load(long cpu)
1716 if (root_task_group_empty())
1717 return;
1719 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1722 #else
1724 static inline void update_shares(struct sched_domain *sd)
1728 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1732 #endif
1734 #ifdef CONFIG_PREEMPT
1736 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1739 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1740 * way at the expense of forcing extra atomic operations in all
1741 * invocations. This assures that the double_lock is acquired using the
1742 * same underlying policy as the spinlock_t on this architecture, which
1743 * reduces latency compared to the unfair variant below. However, it
1744 * also adds more overhead and therefore may reduce throughput.
1746 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1747 __releases(this_rq->lock)
1748 __acquires(busiest->lock)
1749 __acquires(this_rq->lock)
1751 raw_spin_unlock(&this_rq->lock);
1752 double_rq_lock(this_rq, busiest);
1754 return 1;
1757 #else
1759 * Unfair double_lock_balance: Optimizes throughput at the expense of
1760 * latency by eliminating extra atomic operations when the locks are
1761 * already in proper order on entry. This favors lower cpu-ids and will
1762 * grant the double lock to lower cpus over higher ids under contention,
1763 * regardless of entry order into the function.
1765 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1766 __releases(this_rq->lock)
1767 __acquires(busiest->lock)
1768 __acquires(this_rq->lock)
1770 int ret = 0;
1772 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1773 if (busiest < this_rq) {
1774 raw_spin_unlock(&this_rq->lock);
1775 raw_spin_lock(&busiest->lock);
1776 raw_spin_lock_nested(&this_rq->lock,
1777 SINGLE_DEPTH_NESTING);
1778 ret = 1;
1779 } else
1780 raw_spin_lock_nested(&busiest->lock,
1781 SINGLE_DEPTH_NESTING);
1783 return ret;
1786 #endif /* CONFIG_PREEMPT */
1789 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1791 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1793 if (unlikely(!irqs_disabled())) {
1794 /* printk() doesn't work good under rq->lock */
1795 raw_spin_unlock(&this_rq->lock);
1796 BUG_ON(1);
1799 return _double_lock_balance(this_rq, busiest);
1802 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1803 __releases(busiest->lock)
1805 raw_spin_unlock(&busiest->lock);
1806 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1808 #endif
1810 #ifdef CONFIG_FAIR_GROUP_SCHED
1811 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1813 #ifdef CONFIG_SMP
1814 cfs_rq->shares = shares;
1815 #endif
1817 #endif
1819 static void calc_load_account_active(struct rq *this_rq);
1820 static void update_sysctl(void);
1821 static int get_update_sysctl_factor(void);
1823 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1825 set_task_rq(p, cpu);
1826 #ifdef CONFIG_SMP
1828 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1829 * successfuly executed on another CPU. We must ensure that updates of
1830 * per-task data have been completed by this moment.
1832 smp_wmb();
1833 task_thread_info(p)->cpu = cpu;
1834 #endif
1837 #include "sched_stats.h"
1838 #include "sched_idletask.c"
1839 #include "sched_fair.c"
1840 #include "sched_rt.c"
1841 #ifdef CONFIG_SCHED_DEBUG
1842 # include "sched_debug.c"
1843 #endif
1845 #define sched_class_highest (&rt_sched_class)
1846 #define for_each_class(class) \
1847 for (class = sched_class_highest; class; class = class->next)
1849 static void inc_nr_running(struct rq *rq)
1851 rq->nr_running++;
1854 static void dec_nr_running(struct rq *rq)
1856 rq->nr_running--;
1859 static void set_load_weight(struct task_struct *p)
1861 if (task_has_rt_policy(p)) {
1862 p->se.load.weight = prio_to_weight[0] * 2;
1863 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1864 return;
1868 * SCHED_IDLE tasks get minimal weight:
1870 if (p->policy == SCHED_IDLE) {
1871 p->se.load.weight = WEIGHT_IDLEPRIO;
1872 p->se.load.inv_weight = WMULT_IDLEPRIO;
1873 return;
1876 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1877 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1880 static void update_avg(u64 *avg, u64 sample)
1882 s64 diff = sample - *avg;
1883 *avg += diff >> 3;
1886 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1888 if (wakeup)
1889 p->se.start_runtime = p->se.sum_exec_runtime;
1891 sched_info_queued(p);
1892 p->sched_class->enqueue_task(rq, p, wakeup);
1893 p->se.on_rq = 1;
1896 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1898 if (sleep) {
1899 if (p->se.last_wakeup) {
1900 update_avg(&p->se.avg_overlap,
1901 p->se.sum_exec_runtime - p->se.last_wakeup);
1902 p->se.last_wakeup = 0;
1903 } else {
1904 update_avg(&p->se.avg_wakeup,
1905 sysctl_sched_wakeup_granularity);
1909 sched_info_dequeued(p);
1910 p->sched_class->dequeue_task(rq, p, sleep);
1911 p->se.on_rq = 0;
1915 * __normal_prio - return the priority that is based on the static prio
1917 static inline int __normal_prio(struct task_struct *p)
1919 return p->static_prio;
1923 * Calculate the expected normal priority: i.e. priority
1924 * without taking RT-inheritance into account. Might be
1925 * boosted by interactivity modifiers. Changes upon fork,
1926 * setprio syscalls, and whenever the interactivity
1927 * estimator recalculates.
1929 static inline int normal_prio(struct task_struct *p)
1931 int prio;
1933 if (task_has_rt_policy(p))
1934 prio = MAX_RT_PRIO-1 - p->rt_priority;
1935 else
1936 prio = __normal_prio(p);
1937 return prio;
1941 * Calculate the current priority, i.e. the priority
1942 * taken into account by the scheduler. This value might
1943 * be boosted by RT tasks, or might be boosted by
1944 * interactivity modifiers. Will be RT if the task got
1945 * RT-boosted. If not then it returns p->normal_prio.
1947 static int effective_prio(struct task_struct *p)
1949 p->normal_prio = normal_prio(p);
1951 * If we are RT tasks or we were boosted to RT priority,
1952 * keep the priority unchanged. Otherwise, update priority
1953 * to the normal priority:
1955 if (!rt_prio(p->prio))
1956 return p->normal_prio;
1957 return p->prio;
1961 * activate_task - move a task to the runqueue.
1963 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1965 if (task_contributes_to_load(p))
1966 rq->nr_uninterruptible--;
1968 enqueue_task(rq, p, wakeup);
1969 inc_nr_running(rq);
1973 * deactivate_task - remove a task from the runqueue.
1975 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1977 if (task_contributes_to_load(p))
1978 rq->nr_uninterruptible++;
1980 dequeue_task(rq, p, sleep);
1981 dec_nr_running(rq);
1985 * task_curr - is this task currently executing on a CPU?
1986 * @p: the task in question.
1988 inline int task_curr(const struct task_struct *p)
1990 return cpu_curr(task_cpu(p)) == p;
1993 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1994 const struct sched_class *prev_class,
1995 int oldprio, int running)
1997 if (prev_class != p->sched_class) {
1998 if (prev_class->switched_from)
1999 prev_class->switched_from(rq, p, running);
2000 p->sched_class->switched_to(rq, p, running);
2001 } else
2002 p->sched_class->prio_changed(rq, p, oldprio, running);
2005 #ifdef CONFIG_SMP
2007 * Is this task likely cache-hot:
2009 static int
2010 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2012 s64 delta;
2014 if (p->sched_class != &fair_sched_class)
2015 return 0;
2018 * Buddy candidates are cache hot:
2020 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2021 (&p->se == cfs_rq_of(&p->se)->next ||
2022 &p->se == cfs_rq_of(&p->se)->last))
2023 return 1;
2025 if (sysctl_sched_migration_cost == -1)
2026 return 1;
2027 if (sysctl_sched_migration_cost == 0)
2028 return 0;
2030 delta = now - p->se.exec_start;
2032 return delta < (s64)sysctl_sched_migration_cost;
2035 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2037 #ifdef CONFIG_SCHED_DEBUG
2039 * We should never call set_task_cpu() on a blocked task,
2040 * ttwu() will sort out the placement.
2042 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2043 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2044 #endif
2046 trace_sched_migrate_task(p, new_cpu);
2048 if (task_cpu(p) != new_cpu) {
2049 p->se.nr_migrations++;
2050 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2053 __set_task_cpu(p, new_cpu);
2056 struct migration_req {
2057 struct list_head list;
2059 struct task_struct *task;
2060 int dest_cpu;
2062 struct completion done;
2066 * The task's runqueue lock must be held.
2067 * Returns true if you have to wait for migration thread.
2069 static int
2070 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2072 struct rq *rq = task_rq(p);
2075 * If the task is not on a runqueue (and not running), then
2076 * the next wake-up will properly place the task.
2078 if (!p->se.on_rq && !task_running(rq, p))
2079 return 0;
2081 init_completion(&req->done);
2082 req->task = p;
2083 req->dest_cpu = dest_cpu;
2084 list_add(&req->list, &rq->migration_queue);
2086 return 1;
2090 * wait_task_context_switch - wait for a thread to complete at least one
2091 * context switch.
2093 * @p must not be current.
2095 void wait_task_context_switch(struct task_struct *p)
2097 unsigned long nvcsw, nivcsw, flags;
2098 int running;
2099 struct rq *rq;
2101 nvcsw = p->nvcsw;
2102 nivcsw = p->nivcsw;
2103 for (;;) {
2105 * The runqueue is assigned before the actual context
2106 * switch. We need to take the runqueue lock.
2108 * We could check initially without the lock but it is
2109 * very likely that we need to take the lock in every
2110 * iteration.
2112 rq = task_rq_lock(p, &flags);
2113 running = task_running(rq, p);
2114 task_rq_unlock(rq, &flags);
2116 if (likely(!running))
2117 break;
2119 * The switch count is incremented before the actual
2120 * context switch. We thus wait for two switches to be
2121 * sure at least one completed.
2123 if ((p->nvcsw - nvcsw) > 1)
2124 break;
2125 if ((p->nivcsw - nivcsw) > 1)
2126 break;
2128 cpu_relax();
2133 * wait_task_inactive - wait for a thread to unschedule.
2135 * If @match_state is nonzero, it's the @p->state value just checked and
2136 * not expected to change. If it changes, i.e. @p might have woken up,
2137 * then return zero. When we succeed in waiting for @p to be off its CPU,
2138 * we return a positive number (its total switch count). If a second call
2139 * a short while later returns the same number, the caller can be sure that
2140 * @p has remained unscheduled the whole time.
2142 * The caller must ensure that the task *will* unschedule sometime soon,
2143 * else this function might spin for a *long* time. This function can't
2144 * be called with interrupts off, or it may introduce deadlock with
2145 * smp_call_function() if an IPI is sent by the same process we are
2146 * waiting to become inactive.
2148 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2150 unsigned long flags;
2151 int running, on_rq;
2152 unsigned long ncsw;
2153 struct rq *rq;
2155 for (;;) {
2157 * We do the initial early heuristics without holding
2158 * any task-queue locks at all. We'll only try to get
2159 * the runqueue lock when things look like they will
2160 * work out!
2162 rq = task_rq(p);
2165 * If the task is actively running on another CPU
2166 * still, just relax and busy-wait without holding
2167 * any locks.
2169 * NOTE! Since we don't hold any locks, it's not
2170 * even sure that "rq" stays as the right runqueue!
2171 * But we don't care, since "task_running()" will
2172 * return false if the runqueue has changed and p
2173 * is actually now running somewhere else!
2175 while (task_running(rq, p)) {
2176 if (match_state && unlikely(p->state != match_state))
2177 return 0;
2178 cpu_relax();
2182 * Ok, time to look more closely! We need the rq
2183 * lock now, to be *sure*. If we're wrong, we'll
2184 * just go back and repeat.
2186 rq = task_rq_lock(p, &flags);
2187 trace_sched_wait_task(rq, p);
2188 running = task_running(rq, p);
2189 on_rq = p->se.on_rq;
2190 ncsw = 0;
2191 if (!match_state || p->state == match_state)
2192 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2193 task_rq_unlock(rq, &flags);
2196 * If it changed from the expected state, bail out now.
2198 if (unlikely(!ncsw))
2199 break;
2202 * Was it really running after all now that we
2203 * checked with the proper locks actually held?
2205 * Oops. Go back and try again..
2207 if (unlikely(running)) {
2208 cpu_relax();
2209 continue;
2213 * It's not enough that it's not actively running,
2214 * it must be off the runqueue _entirely_, and not
2215 * preempted!
2217 * So if it was still runnable (but just not actively
2218 * running right now), it's preempted, and we should
2219 * yield - it could be a while.
2221 if (unlikely(on_rq)) {
2222 schedule_timeout_uninterruptible(1);
2223 continue;
2227 * Ahh, all good. It wasn't running, and it wasn't
2228 * runnable, which means that it will never become
2229 * running in the future either. We're all done!
2231 break;
2234 return ncsw;
2237 /***
2238 * kick_process - kick a running thread to enter/exit the kernel
2239 * @p: the to-be-kicked thread
2241 * Cause a process which is running on another CPU to enter
2242 * kernel-mode, without any delay. (to get signals handled.)
2244 * NOTE: this function doesnt have to take the runqueue lock,
2245 * because all it wants to ensure is that the remote task enters
2246 * the kernel. If the IPI races and the task has been migrated
2247 * to another CPU then no harm is done and the purpose has been
2248 * achieved as well.
2250 void kick_process(struct task_struct *p)
2252 int cpu;
2254 preempt_disable();
2255 cpu = task_cpu(p);
2256 if ((cpu != smp_processor_id()) && task_curr(p))
2257 smp_send_reschedule(cpu);
2258 preempt_enable();
2260 EXPORT_SYMBOL_GPL(kick_process);
2261 #endif /* CONFIG_SMP */
2264 * task_oncpu_function_call - call a function on the cpu on which a task runs
2265 * @p: the task to evaluate
2266 * @func: the function to be called
2267 * @info: the function call argument
2269 * Calls the function @func when the task is currently running. This might
2270 * be on the current CPU, which just calls the function directly
2272 void task_oncpu_function_call(struct task_struct *p,
2273 void (*func) (void *info), void *info)
2275 int cpu;
2277 preempt_disable();
2278 cpu = task_cpu(p);
2279 if (task_curr(p))
2280 smp_call_function_single(cpu, func, info, 1);
2281 preempt_enable();
2284 #ifdef CONFIG_SMP
2285 static int select_fallback_rq(int cpu, struct task_struct *p)
2287 int dest_cpu;
2288 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2290 /* Look for allowed, online CPU in same node. */
2291 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2292 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2293 return dest_cpu;
2295 /* Any allowed, online CPU? */
2296 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2297 if (dest_cpu < nr_cpu_ids)
2298 return dest_cpu;
2300 /* No more Mr. Nice Guy. */
2301 if (dest_cpu >= nr_cpu_ids) {
2302 rcu_read_lock();
2303 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2304 rcu_read_unlock();
2305 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2308 * Don't tell them about moving exiting tasks or
2309 * kernel threads (both mm NULL), since they never
2310 * leave kernel.
2312 if (p->mm && printk_ratelimit()) {
2313 printk(KERN_INFO "process %d (%s) no "
2314 "longer affine to cpu%d\n",
2315 task_pid_nr(p), p->comm, cpu);
2319 return dest_cpu;
2323 * Called from:
2325 * - fork, @p is stable because it isn't on the tasklist yet
2327 * - exec, @p is unstable, retry loop
2329 * - wake-up, we serialize ->cpus_allowed against TASK_WAKING so
2330 * we should be good.
2332 static inline
2333 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2335 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2338 * In order not to call set_task_cpu() on a blocking task we need
2339 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2340 * cpu.
2342 * Since this is common to all placement strategies, this lives here.
2344 * [ this allows ->select_task() to simply return task_cpu(p) and
2345 * not worry about this generic constraint ]
2347 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2348 !cpu_online(cpu)))
2349 cpu = select_fallback_rq(task_cpu(p), p);
2351 return cpu;
2353 #endif
2355 /***
2356 * try_to_wake_up - wake up a thread
2357 * @p: the to-be-woken-up thread
2358 * @state: the mask of task states that can be woken
2359 * @sync: do a synchronous wakeup?
2361 * Put it on the run-queue if it's not already there. The "current"
2362 * thread is always on the run-queue (except when the actual
2363 * re-schedule is in progress), and as such you're allowed to do
2364 * the simpler "current->state = TASK_RUNNING" to mark yourself
2365 * runnable without the overhead of this.
2367 * returns failure only if the task is already active.
2369 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2370 int wake_flags)
2372 int cpu, orig_cpu, this_cpu, success = 0;
2373 unsigned long flags;
2374 struct rq *rq, *orig_rq;
2376 if (!sched_feat(SYNC_WAKEUPS))
2377 wake_flags &= ~WF_SYNC;
2379 this_cpu = get_cpu();
2381 smp_wmb();
2382 rq = orig_rq = task_rq_lock(p, &flags);
2383 update_rq_clock(rq);
2384 if (!(p->state & state))
2385 goto out;
2387 if (p->se.on_rq)
2388 goto out_running;
2390 cpu = task_cpu(p);
2391 orig_cpu = cpu;
2393 #ifdef CONFIG_SMP
2394 if (unlikely(task_running(rq, p)))
2395 goto out_activate;
2398 * In order to handle concurrent wakeups and release the rq->lock
2399 * we put the task in TASK_WAKING state.
2401 * First fix up the nr_uninterruptible count:
2403 if (task_contributes_to_load(p))
2404 rq->nr_uninterruptible--;
2405 p->state = TASK_WAKING;
2407 if (p->sched_class->task_waking)
2408 p->sched_class->task_waking(rq, p);
2410 __task_rq_unlock(rq);
2412 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2413 if (cpu != orig_cpu)
2414 set_task_cpu(p, cpu);
2416 rq = __task_rq_lock(p);
2417 update_rq_clock(rq);
2419 WARN_ON(p->state != TASK_WAKING);
2420 cpu = task_cpu(p);
2422 #ifdef CONFIG_SCHEDSTATS
2423 schedstat_inc(rq, ttwu_count);
2424 if (cpu == this_cpu)
2425 schedstat_inc(rq, ttwu_local);
2426 else {
2427 struct sched_domain *sd;
2428 for_each_domain(this_cpu, sd) {
2429 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2430 schedstat_inc(sd, ttwu_wake_remote);
2431 break;
2435 #endif /* CONFIG_SCHEDSTATS */
2437 out_activate:
2438 #endif /* CONFIG_SMP */
2439 schedstat_inc(p, se.nr_wakeups);
2440 if (wake_flags & WF_SYNC)
2441 schedstat_inc(p, se.nr_wakeups_sync);
2442 if (orig_cpu != cpu)
2443 schedstat_inc(p, se.nr_wakeups_migrate);
2444 if (cpu == this_cpu)
2445 schedstat_inc(p, se.nr_wakeups_local);
2446 else
2447 schedstat_inc(p, se.nr_wakeups_remote);
2448 activate_task(rq, p, 1);
2449 success = 1;
2452 * Only attribute actual wakeups done by this task.
2454 if (!in_interrupt()) {
2455 struct sched_entity *se = &current->se;
2456 u64 sample = se->sum_exec_runtime;
2458 if (se->last_wakeup)
2459 sample -= se->last_wakeup;
2460 else
2461 sample -= se->start_runtime;
2462 update_avg(&se->avg_wakeup, sample);
2464 se->last_wakeup = se->sum_exec_runtime;
2467 out_running:
2468 trace_sched_wakeup(rq, p, success);
2469 check_preempt_curr(rq, p, wake_flags);
2471 p->state = TASK_RUNNING;
2472 #ifdef CONFIG_SMP
2473 if (p->sched_class->task_woken)
2474 p->sched_class->task_woken(rq, p);
2476 if (unlikely(rq->idle_stamp)) {
2477 u64 delta = rq->clock - rq->idle_stamp;
2478 u64 max = 2*sysctl_sched_migration_cost;
2480 if (delta > max)
2481 rq->avg_idle = max;
2482 else
2483 update_avg(&rq->avg_idle, delta);
2484 rq->idle_stamp = 0;
2486 #endif
2487 out:
2488 task_rq_unlock(rq, &flags);
2489 put_cpu();
2491 return success;
2495 * wake_up_process - Wake up a specific process
2496 * @p: The process to be woken up.
2498 * Attempt to wake up the nominated process and move it to the set of runnable
2499 * processes. Returns 1 if the process was woken up, 0 if it was already
2500 * running.
2502 * It may be assumed that this function implies a write memory barrier before
2503 * changing the task state if and only if any tasks are woken up.
2505 int wake_up_process(struct task_struct *p)
2507 return try_to_wake_up(p, TASK_ALL, 0);
2509 EXPORT_SYMBOL(wake_up_process);
2511 int wake_up_state(struct task_struct *p, unsigned int state)
2513 return try_to_wake_up(p, state, 0);
2517 * Perform scheduler related setup for a newly forked process p.
2518 * p is forked by current.
2520 * __sched_fork() is basic setup used by init_idle() too:
2522 static void __sched_fork(struct task_struct *p)
2524 p->se.exec_start = 0;
2525 p->se.sum_exec_runtime = 0;
2526 p->se.prev_sum_exec_runtime = 0;
2527 p->se.nr_migrations = 0;
2528 p->se.last_wakeup = 0;
2529 p->se.avg_overlap = 0;
2530 p->se.start_runtime = 0;
2531 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2533 #ifdef CONFIG_SCHEDSTATS
2534 p->se.wait_start = 0;
2535 p->se.wait_max = 0;
2536 p->se.wait_count = 0;
2537 p->se.wait_sum = 0;
2539 p->se.sleep_start = 0;
2540 p->se.sleep_max = 0;
2541 p->se.sum_sleep_runtime = 0;
2543 p->se.block_start = 0;
2544 p->se.block_max = 0;
2545 p->se.exec_max = 0;
2546 p->se.slice_max = 0;
2548 p->se.nr_migrations_cold = 0;
2549 p->se.nr_failed_migrations_affine = 0;
2550 p->se.nr_failed_migrations_running = 0;
2551 p->se.nr_failed_migrations_hot = 0;
2552 p->se.nr_forced_migrations = 0;
2554 p->se.nr_wakeups = 0;
2555 p->se.nr_wakeups_sync = 0;
2556 p->se.nr_wakeups_migrate = 0;
2557 p->se.nr_wakeups_local = 0;
2558 p->se.nr_wakeups_remote = 0;
2559 p->se.nr_wakeups_affine = 0;
2560 p->se.nr_wakeups_affine_attempts = 0;
2561 p->se.nr_wakeups_passive = 0;
2562 p->se.nr_wakeups_idle = 0;
2564 #endif
2566 INIT_LIST_HEAD(&p->rt.run_list);
2567 p->se.on_rq = 0;
2568 INIT_LIST_HEAD(&p->se.group_node);
2570 #ifdef CONFIG_PREEMPT_NOTIFIERS
2571 INIT_HLIST_HEAD(&p->preempt_notifiers);
2572 #endif
2576 * fork()/clone()-time setup:
2578 void sched_fork(struct task_struct *p, int clone_flags)
2580 int cpu = get_cpu();
2582 __sched_fork(p);
2584 * We mark the process as waking here. This guarantees that
2585 * nobody will actually run it, and a signal or other external
2586 * event cannot wake it up and insert it on the runqueue either.
2588 p->state = TASK_WAKING;
2591 * Revert to default priority/policy on fork if requested.
2593 if (unlikely(p->sched_reset_on_fork)) {
2594 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2595 p->policy = SCHED_NORMAL;
2596 p->normal_prio = p->static_prio;
2599 if (PRIO_TO_NICE(p->static_prio) < 0) {
2600 p->static_prio = NICE_TO_PRIO(0);
2601 p->normal_prio = p->static_prio;
2602 set_load_weight(p);
2606 * We don't need the reset flag anymore after the fork. It has
2607 * fulfilled its duty:
2609 p->sched_reset_on_fork = 0;
2613 * Make sure we do not leak PI boosting priority to the child.
2615 p->prio = current->normal_prio;
2617 if (!rt_prio(p->prio))
2618 p->sched_class = &fair_sched_class;
2620 if (p->sched_class->task_fork)
2621 p->sched_class->task_fork(p);
2623 #ifdef CONFIG_SMP
2624 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2625 #endif
2626 set_task_cpu(p, cpu);
2628 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2629 if (likely(sched_info_on()))
2630 memset(&p->sched_info, 0, sizeof(p->sched_info));
2631 #endif
2632 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2633 p->oncpu = 0;
2634 #endif
2635 #ifdef CONFIG_PREEMPT
2636 /* Want to start with kernel preemption disabled. */
2637 task_thread_info(p)->preempt_count = 1;
2638 #endif
2639 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2641 put_cpu();
2645 * wake_up_new_task - wake up a newly created task for the first time.
2647 * This function will do some initial scheduler statistics housekeeping
2648 * that must be done for every newly created context, then puts the task
2649 * on the runqueue and wakes it.
2651 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2653 unsigned long flags;
2654 struct rq *rq;
2656 rq = task_rq_lock(p, &flags);
2657 BUG_ON(p->state != TASK_WAKING);
2658 p->state = TASK_RUNNING;
2659 update_rq_clock(rq);
2660 activate_task(rq, p, 0);
2661 trace_sched_wakeup_new(rq, p, 1);
2662 check_preempt_curr(rq, p, WF_FORK);
2663 #ifdef CONFIG_SMP
2664 if (p->sched_class->task_woken)
2665 p->sched_class->task_woken(rq, p);
2666 #endif
2667 task_rq_unlock(rq, &flags);
2670 #ifdef CONFIG_PREEMPT_NOTIFIERS
2673 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2674 * @notifier: notifier struct to register
2676 void preempt_notifier_register(struct preempt_notifier *notifier)
2678 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2680 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2683 * preempt_notifier_unregister - no longer interested in preemption notifications
2684 * @notifier: notifier struct to unregister
2686 * This is safe to call from within a preemption notifier.
2688 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2690 hlist_del(&notifier->link);
2692 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2694 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2696 struct preempt_notifier *notifier;
2697 struct hlist_node *node;
2699 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2700 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2703 static void
2704 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2705 struct task_struct *next)
2707 struct preempt_notifier *notifier;
2708 struct hlist_node *node;
2710 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2711 notifier->ops->sched_out(notifier, next);
2714 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2716 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2720 static void
2721 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2722 struct task_struct *next)
2726 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2729 * prepare_task_switch - prepare to switch tasks
2730 * @rq: the runqueue preparing to switch
2731 * @prev: the current task that is being switched out
2732 * @next: the task we are going to switch to.
2734 * This is called with the rq lock held and interrupts off. It must
2735 * be paired with a subsequent finish_task_switch after the context
2736 * switch.
2738 * prepare_task_switch sets up locking and calls architecture specific
2739 * hooks.
2741 static inline void
2742 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2743 struct task_struct *next)
2745 fire_sched_out_preempt_notifiers(prev, next);
2746 prepare_lock_switch(rq, next);
2747 prepare_arch_switch(next);
2751 * finish_task_switch - clean up after a task-switch
2752 * @rq: runqueue associated with task-switch
2753 * @prev: the thread we just switched away from.
2755 * finish_task_switch must be called after the context switch, paired
2756 * with a prepare_task_switch call before the context switch.
2757 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2758 * and do any other architecture-specific cleanup actions.
2760 * Note that we may have delayed dropping an mm in context_switch(). If
2761 * so, we finish that here outside of the runqueue lock. (Doing it
2762 * with the lock held can cause deadlocks; see schedule() for
2763 * details.)
2765 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2766 __releases(rq->lock)
2768 struct mm_struct *mm = rq->prev_mm;
2769 long prev_state;
2771 rq->prev_mm = NULL;
2774 * A task struct has one reference for the use as "current".
2775 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2776 * schedule one last time. The schedule call will never return, and
2777 * the scheduled task must drop that reference.
2778 * The test for TASK_DEAD must occur while the runqueue locks are
2779 * still held, otherwise prev could be scheduled on another cpu, die
2780 * there before we look at prev->state, and then the reference would
2781 * be dropped twice.
2782 * Manfred Spraul <manfred@colorfullife.com>
2784 prev_state = prev->state;
2785 finish_arch_switch(prev);
2786 perf_event_task_sched_in(current, cpu_of(rq));
2787 finish_lock_switch(rq, prev);
2789 fire_sched_in_preempt_notifiers(current);
2790 if (mm)
2791 mmdrop(mm);
2792 if (unlikely(prev_state == TASK_DEAD)) {
2794 * Remove function-return probe instances associated with this
2795 * task and put them back on the free list.
2797 kprobe_flush_task(prev);
2798 put_task_struct(prev);
2802 #ifdef CONFIG_SMP
2804 /* assumes rq->lock is held */
2805 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2807 if (prev->sched_class->pre_schedule)
2808 prev->sched_class->pre_schedule(rq, prev);
2811 /* rq->lock is NOT held, but preemption is disabled */
2812 static inline void post_schedule(struct rq *rq)
2814 if (rq->post_schedule) {
2815 unsigned long flags;
2817 raw_spin_lock_irqsave(&rq->lock, flags);
2818 if (rq->curr->sched_class->post_schedule)
2819 rq->curr->sched_class->post_schedule(rq);
2820 raw_spin_unlock_irqrestore(&rq->lock, flags);
2822 rq->post_schedule = 0;
2826 #else
2828 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2832 static inline void post_schedule(struct rq *rq)
2836 #endif
2839 * schedule_tail - first thing a freshly forked thread must call.
2840 * @prev: the thread we just switched away from.
2842 asmlinkage void schedule_tail(struct task_struct *prev)
2843 __releases(rq->lock)
2845 struct rq *rq = this_rq();
2847 finish_task_switch(rq, prev);
2850 * FIXME: do we need to worry about rq being invalidated by the
2851 * task_switch?
2853 post_schedule(rq);
2855 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2856 /* In this case, finish_task_switch does not reenable preemption */
2857 preempt_enable();
2858 #endif
2859 if (current->set_child_tid)
2860 put_user(task_pid_vnr(current), current->set_child_tid);
2864 * context_switch - switch to the new MM and the new
2865 * thread's register state.
2867 static inline void
2868 context_switch(struct rq *rq, struct task_struct *prev,
2869 struct task_struct *next)
2871 struct mm_struct *mm, *oldmm;
2873 prepare_task_switch(rq, prev, next);
2874 trace_sched_switch(rq, prev, next);
2875 mm = next->mm;
2876 oldmm = prev->active_mm;
2878 * For paravirt, this is coupled with an exit in switch_to to
2879 * combine the page table reload and the switch backend into
2880 * one hypercall.
2882 arch_start_context_switch(prev);
2884 if (likely(!mm)) {
2885 next->active_mm = oldmm;
2886 atomic_inc(&oldmm->mm_count);
2887 enter_lazy_tlb(oldmm, next);
2888 } else
2889 switch_mm(oldmm, mm, next);
2891 if (likely(!prev->mm)) {
2892 prev->active_mm = NULL;
2893 rq->prev_mm = oldmm;
2896 * Since the runqueue lock will be released by the next
2897 * task (which is an invalid locking op but in the case
2898 * of the scheduler it's an obvious special-case), so we
2899 * do an early lockdep release here:
2901 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2902 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2903 #endif
2905 /* Here we just switch the register state and the stack. */
2906 switch_to(prev, next, prev);
2908 barrier();
2910 * this_rq must be evaluated again because prev may have moved
2911 * CPUs since it called schedule(), thus the 'rq' on its stack
2912 * frame will be invalid.
2914 finish_task_switch(this_rq(), prev);
2918 * nr_running, nr_uninterruptible and nr_context_switches:
2920 * externally visible scheduler statistics: current number of runnable
2921 * threads, current number of uninterruptible-sleeping threads, total
2922 * number of context switches performed since bootup.
2924 unsigned long nr_running(void)
2926 unsigned long i, sum = 0;
2928 for_each_online_cpu(i)
2929 sum += cpu_rq(i)->nr_running;
2931 return sum;
2934 unsigned long nr_uninterruptible(void)
2936 unsigned long i, sum = 0;
2938 for_each_possible_cpu(i)
2939 sum += cpu_rq(i)->nr_uninterruptible;
2942 * Since we read the counters lockless, it might be slightly
2943 * inaccurate. Do not allow it to go below zero though:
2945 if (unlikely((long)sum < 0))
2946 sum = 0;
2948 return sum;
2951 unsigned long long nr_context_switches(void)
2953 int i;
2954 unsigned long long sum = 0;
2956 for_each_possible_cpu(i)
2957 sum += cpu_rq(i)->nr_switches;
2959 return sum;
2962 unsigned long nr_iowait(void)
2964 unsigned long i, sum = 0;
2966 for_each_possible_cpu(i)
2967 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2969 return sum;
2972 unsigned long nr_iowait_cpu(void)
2974 struct rq *this = this_rq();
2975 return atomic_read(&this->nr_iowait);
2978 unsigned long this_cpu_load(void)
2980 struct rq *this = this_rq();
2981 return this->cpu_load[0];
2985 /* Variables and functions for calc_load */
2986 static atomic_long_t calc_load_tasks;
2987 static unsigned long calc_load_update;
2988 unsigned long avenrun[3];
2989 EXPORT_SYMBOL(avenrun);
2992 * get_avenrun - get the load average array
2993 * @loads: pointer to dest load array
2994 * @offset: offset to add
2995 * @shift: shift count to shift the result left
2997 * These values are estimates at best, so no need for locking.
2999 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3001 loads[0] = (avenrun[0] + offset) << shift;
3002 loads[1] = (avenrun[1] + offset) << shift;
3003 loads[2] = (avenrun[2] + offset) << shift;
3006 static unsigned long
3007 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3009 load *= exp;
3010 load += active * (FIXED_1 - exp);
3011 return load >> FSHIFT;
3015 * calc_load - update the avenrun load estimates 10 ticks after the
3016 * CPUs have updated calc_load_tasks.
3018 void calc_global_load(void)
3020 unsigned long upd = calc_load_update + 10;
3021 long active;
3023 if (time_before(jiffies, upd))
3024 return;
3026 active = atomic_long_read(&calc_load_tasks);
3027 active = active > 0 ? active * FIXED_1 : 0;
3029 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3030 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3031 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3033 calc_load_update += LOAD_FREQ;
3037 * Either called from update_cpu_load() or from a cpu going idle
3039 static void calc_load_account_active(struct rq *this_rq)
3041 long nr_active, delta;
3043 nr_active = this_rq->nr_running;
3044 nr_active += (long) this_rq->nr_uninterruptible;
3046 if (nr_active != this_rq->calc_load_active) {
3047 delta = nr_active - this_rq->calc_load_active;
3048 this_rq->calc_load_active = nr_active;
3049 atomic_long_add(delta, &calc_load_tasks);
3054 * Update rq->cpu_load[] statistics. This function is usually called every
3055 * scheduler tick (TICK_NSEC).
3057 static void update_cpu_load(struct rq *this_rq)
3059 unsigned long this_load = this_rq->load.weight;
3060 int i, scale;
3062 this_rq->nr_load_updates++;
3064 /* Update our load: */
3065 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3066 unsigned long old_load, new_load;
3068 /* scale is effectively 1 << i now, and >> i divides by scale */
3070 old_load = this_rq->cpu_load[i];
3071 new_load = this_load;
3073 * Round up the averaging division if load is increasing. This
3074 * prevents us from getting stuck on 9 if the load is 10, for
3075 * example.
3077 if (new_load > old_load)
3078 new_load += scale-1;
3079 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3082 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3083 this_rq->calc_load_update += LOAD_FREQ;
3084 calc_load_account_active(this_rq);
3088 #ifdef CONFIG_SMP
3091 * double_rq_lock - safely lock two runqueues
3093 * Note this does not disable interrupts like task_rq_lock,
3094 * you need to do so manually before calling.
3096 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3097 __acquires(rq1->lock)
3098 __acquires(rq2->lock)
3100 BUG_ON(!irqs_disabled());
3101 if (rq1 == rq2) {
3102 raw_spin_lock(&rq1->lock);
3103 __acquire(rq2->lock); /* Fake it out ;) */
3104 } else {
3105 if (rq1 < rq2) {
3106 raw_spin_lock(&rq1->lock);
3107 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3108 } else {
3109 raw_spin_lock(&rq2->lock);
3110 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3113 update_rq_clock(rq1);
3114 update_rq_clock(rq2);
3118 * double_rq_unlock - safely unlock two runqueues
3120 * Note this does not restore interrupts like task_rq_unlock,
3121 * you need to do so manually after calling.
3123 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3124 __releases(rq1->lock)
3125 __releases(rq2->lock)
3127 raw_spin_unlock(&rq1->lock);
3128 if (rq1 != rq2)
3129 raw_spin_unlock(&rq2->lock);
3130 else
3131 __release(rq2->lock);
3135 * sched_exec - execve() is a valuable balancing opportunity, because at
3136 * this point the task has the smallest effective memory and cache footprint.
3138 void sched_exec(void)
3140 struct task_struct *p = current;
3141 struct migration_req req;
3142 int dest_cpu, this_cpu;
3143 unsigned long flags;
3144 struct rq *rq;
3146 again:
3147 this_cpu = get_cpu();
3148 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3149 if (dest_cpu == this_cpu) {
3150 put_cpu();
3151 return;
3154 rq = task_rq_lock(p, &flags);
3155 put_cpu();
3158 * select_task_rq() can race against ->cpus_allowed
3160 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3161 || unlikely(!cpu_active(dest_cpu))) {
3162 task_rq_unlock(rq, &flags);
3163 goto again;
3166 /* force the process onto the specified CPU */
3167 if (migrate_task(p, dest_cpu, &req)) {
3168 /* Need to wait for migration thread (might exit: take ref). */
3169 struct task_struct *mt = rq->migration_thread;
3171 get_task_struct(mt);
3172 task_rq_unlock(rq, &flags);
3173 wake_up_process(mt);
3174 put_task_struct(mt);
3175 wait_for_completion(&req.done);
3177 return;
3179 task_rq_unlock(rq, &flags);
3183 * pull_task - move a task from a remote runqueue to the local runqueue.
3184 * Both runqueues must be locked.
3186 static void pull_task(struct rq *src_rq, struct task_struct *p,
3187 struct rq *this_rq, int this_cpu)
3189 deactivate_task(src_rq, p, 0);
3190 set_task_cpu(p, this_cpu);
3191 activate_task(this_rq, p, 0);
3192 check_preempt_curr(this_rq, p, 0);
3196 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3198 static
3199 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3200 struct sched_domain *sd, enum cpu_idle_type idle,
3201 int *all_pinned)
3203 int tsk_cache_hot = 0;
3205 * We do not migrate tasks that are:
3206 * 1) running (obviously), or
3207 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3208 * 3) are cache-hot on their current CPU.
3210 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3211 schedstat_inc(p, se.nr_failed_migrations_affine);
3212 return 0;
3214 *all_pinned = 0;
3216 if (task_running(rq, p)) {
3217 schedstat_inc(p, se.nr_failed_migrations_running);
3218 return 0;
3222 * Aggressive migration if:
3223 * 1) task is cache cold, or
3224 * 2) too many balance attempts have failed.
3227 tsk_cache_hot = task_hot(p, rq->clock, sd);
3228 if (!tsk_cache_hot ||
3229 sd->nr_balance_failed > sd->cache_nice_tries) {
3230 #ifdef CONFIG_SCHEDSTATS
3231 if (tsk_cache_hot) {
3232 schedstat_inc(sd, lb_hot_gained[idle]);
3233 schedstat_inc(p, se.nr_forced_migrations);
3235 #endif
3236 return 1;
3239 if (tsk_cache_hot) {
3240 schedstat_inc(p, se.nr_failed_migrations_hot);
3241 return 0;
3243 return 1;
3246 static unsigned long
3247 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3248 unsigned long max_load_move, struct sched_domain *sd,
3249 enum cpu_idle_type idle, int *all_pinned,
3250 int *this_best_prio, struct rq_iterator *iterator)
3252 int loops = 0, pulled = 0, pinned = 0;
3253 struct task_struct *p;
3254 long rem_load_move = max_load_move;
3256 if (max_load_move == 0)
3257 goto out;
3259 pinned = 1;
3262 * Start the load-balancing iterator:
3264 p = iterator->start(iterator->arg);
3265 next:
3266 if (!p || loops++ > sysctl_sched_nr_migrate)
3267 goto out;
3269 if ((p->se.load.weight >> 1) > rem_load_move ||
3270 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3271 p = iterator->next(iterator->arg);
3272 goto next;
3275 pull_task(busiest, p, this_rq, this_cpu);
3276 pulled++;
3277 rem_load_move -= p->se.load.weight;
3279 #ifdef CONFIG_PREEMPT
3281 * NEWIDLE balancing is a source of latency, so preemptible kernels
3282 * will stop after the first task is pulled to minimize the critical
3283 * section.
3285 if (idle == CPU_NEWLY_IDLE)
3286 goto out;
3287 #endif
3290 * We only want to steal up to the prescribed amount of weighted load.
3292 if (rem_load_move > 0) {
3293 if (p->prio < *this_best_prio)
3294 *this_best_prio = p->prio;
3295 p = iterator->next(iterator->arg);
3296 goto next;
3298 out:
3300 * Right now, this is one of only two places pull_task() is called,
3301 * so we can safely collect pull_task() stats here rather than
3302 * inside pull_task().
3304 schedstat_add(sd, lb_gained[idle], pulled);
3306 if (all_pinned)
3307 *all_pinned = pinned;
3309 return max_load_move - rem_load_move;
3313 * move_tasks tries to move up to max_load_move weighted load from busiest to
3314 * this_rq, as part of a balancing operation within domain "sd".
3315 * Returns 1 if successful and 0 otherwise.
3317 * Called with both runqueues locked.
3319 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3320 unsigned long max_load_move,
3321 struct sched_domain *sd, enum cpu_idle_type idle,
3322 int *all_pinned)
3324 const struct sched_class *class = sched_class_highest;
3325 unsigned long total_load_moved = 0;
3326 int this_best_prio = this_rq->curr->prio;
3328 do {
3329 total_load_moved +=
3330 class->load_balance(this_rq, this_cpu, busiest,
3331 max_load_move - total_load_moved,
3332 sd, idle, all_pinned, &this_best_prio);
3333 class = class->next;
3335 #ifdef CONFIG_PREEMPT
3337 * NEWIDLE balancing is a source of latency, so preemptible
3338 * kernels will stop after the first task is pulled to minimize
3339 * the critical section.
3341 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3342 break;
3343 #endif
3344 } while (class && max_load_move > total_load_moved);
3346 return total_load_moved > 0;
3349 static int
3350 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3351 struct sched_domain *sd, enum cpu_idle_type idle,
3352 struct rq_iterator *iterator)
3354 struct task_struct *p = iterator->start(iterator->arg);
3355 int pinned = 0;
3357 while (p) {
3358 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3359 pull_task(busiest, p, this_rq, this_cpu);
3361 * Right now, this is only the second place pull_task()
3362 * is called, so we can safely collect pull_task()
3363 * stats here rather than inside pull_task().
3365 schedstat_inc(sd, lb_gained[idle]);
3367 return 1;
3369 p = iterator->next(iterator->arg);
3372 return 0;
3376 * move_one_task tries to move exactly one task from busiest to this_rq, as
3377 * part of active balancing operations within "domain".
3378 * Returns 1 if successful and 0 otherwise.
3380 * Called with both runqueues locked.
3382 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3383 struct sched_domain *sd, enum cpu_idle_type idle)
3385 const struct sched_class *class;
3387 for_each_class(class) {
3388 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3389 return 1;
3392 return 0;
3394 /********** Helpers for find_busiest_group ************************/
3396 * sd_lb_stats - Structure to store the statistics of a sched_domain
3397 * during load balancing.
3399 struct sd_lb_stats {
3400 struct sched_group *busiest; /* Busiest group in this sd */
3401 struct sched_group *this; /* Local group in this sd */
3402 unsigned long total_load; /* Total load of all groups in sd */
3403 unsigned long total_pwr; /* Total power of all groups in sd */
3404 unsigned long avg_load; /* Average load across all groups in sd */
3406 /** Statistics of this group */
3407 unsigned long this_load;
3408 unsigned long this_load_per_task;
3409 unsigned long this_nr_running;
3411 /* Statistics of the busiest group */
3412 unsigned long max_load;
3413 unsigned long busiest_load_per_task;
3414 unsigned long busiest_nr_running;
3416 int group_imb; /* Is there imbalance in this sd */
3417 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3418 int power_savings_balance; /* Is powersave balance needed for this sd */
3419 struct sched_group *group_min; /* Least loaded group in sd */
3420 struct sched_group *group_leader; /* Group which relieves group_min */
3421 unsigned long min_load_per_task; /* load_per_task in group_min */
3422 unsigned long leader_nr_running; /* Nr running of group_leader */
3423 unsigned long min_nr_running; /* Nr running of group_min */
3424 #endif
3428 * sg_lb_stats - stats of a sched_group required for load_balancing
3430 struct sg_lb_stats {
3431 unsigned long avg_load; /*Avg load across the CPUs of the group */
3432 unsigned long group_load; /* Total load over the CPUs of the group */
3433 unsigned long sum_nr_running; /* Nr tasks running in the group */
3434 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3435 unsigned long group_capacity;
3436 int group_imb; /* Is there an imbalance in the group ? */
3440 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3441 * @group: The group whose first cpu is to be returned.
3443 static inline unsigned int group_first_cpu(struct sched_group *group)
3445 return cpumask_first(sched_group_cpus(group));
3449 * get_sd_load_idx - Obtain the load index for a given sched domain.
3450 * @sd: The sched_domain whose load_idx is to be obtained.
3451 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3453 static inline int get_sd_load_idx(struct sched_domain *sd,
3454 enum cpu_idle_type idle)
3456 int load_idx;
3458 switch (idle) {
3459 case CPU_NOT_IDLE:
3460 load_idx = sd->busy_idx;
3461 break;
3463 case CPU_NEWLY_IDLE:
3464 load_idx = sd->newidle_idx;
3465 break;
3466 default:
3467 load_idx = sd->idle_idx;
3468 break;
3471 return load_idx;
3475 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3477 * init_sd_power_savings_stats - Initialize power savings statistics for
3478 * the given sched_domain, during load balancing.
3480 * @sd: Sched domain whose power-savings statistics are to be initialized.
3481 * @sds: Variable containing the statistics for sd.
3482 * @idle: Idle status of the CPU at which we're performing load-balancing.
3484 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3485 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3488 * Busy processors will not participate in power savings
3489 * balance.
3491 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3492 sds->power_savings_balance = 0;
3493 else {
3494 sds->power_savings_balance = 1;
3495 sds->min_nr_running = ULONG_MAX;
3496 sds->leader_nr_running = 0;
3501 * update_sd_power_savings_stats - Update the power saving stats for a
3502 * sched_domain while performing load balancing.
3504 * @group: sched_group belonging to the sched_domain under consideration.
3505 * @sds: Variable containing the statistics of the sched_domain
3506 * @local_group: Does group contain the CPU for which we're performing
3507 * load balancing ?
3508 * @sgs: Variable containing the statistics of the group.
3510 static inline void update_sd_power_savings_stats(struct sched_group *group,
3511 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3514 if (!sds->power_savings_balance)
3515 return;
3518 * If the local group is idle or completely loaded
3519 * no need to do power savings balance at this domain
3521 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3522 !sds->this_nr_running))
3523 sds->power_savings_balance = 0;
3526 * If a group is already running at full capacity or idle,
3527 * don't include that group in power savings calculations
3529 if (!sds->power_savings_balance ||
3530 sgs->sum_nr_running >= sgs->group_capacity ||
3531 !sgs->sum_nr_running)
3532 return;
3535 * Calculate the group which has the least non-idle load.
3536 * This is the group from where we need to pick up the load
3537 * for saving power
3539 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3540 (sgs->sum_nr_running == sds->min_nr_running &&
3541 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3542 sds->group_min = group;
3543 sds->min_nr_running = sgs->sum_nr_running;
3544 sds->min_load_per_task = sgs->sum_weighted_load /
3545 sgs->sum_nr_running;
3549 * Calculate the group which is almost near its
3550 * capacity but still has some space to pick up some load
3551 * from other group and save more power
3553 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3554 return;
3556 if (sgs->sum_nr_running > sds->leader_nr_running ||
3557 (sgs->sum_nr_running == sds->leader_nr_running &&
3558 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3559 sds->group_leader = group;
3560 sds->leader_nr_running = sgs->sum_nr_running;
3565 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3566 * @sds: Variable containing the statistics of the sched_domain
3567 * under consideration.
3568 * @this_cpu: Cpu at which we're currently performing load-balancing.
3569 * @imbalance: Variable to store the imbalance.
3571 * Description:
3572 * Check if we have potential to perform some power-savings balance.
3573 * If yes, set the busiest group to be the least loaded group in the
3574 * sched_domain, so that it's CPUs can be put to idle.
3576 * Returns 1 if there is potential to perform power-savings balance.
3577 * Else returns 0.
3579 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3580 int this_cpu, unsigned long *imbalance)
3582 if (!sds->power_savings_balance)
3583 return 0;
3585 if (sds->this != sds->group_leader ||
3586 sds->group_leader == sds->group_min)
3587 return 0;
3589 *imbalance = sds->min_load_per_task;
3590 sds->busiest = sds->group_min;
3592 return 1;
3595 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3596 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3597 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3599 return;
3602 static inline void update_sd_power_savings_stats(struct sched_group *group,
3603 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3605 return;
3608 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3609 int this_cpu, unsigned long *imbalance)
3611 return 0;
3613 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3616 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3618 return SCHED_LOAD_SCALE;
3621 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3623 return default_scale_freq_power(sd, cpu);
3626 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3628 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3629 unsigned long smt_gain = sd->smt_gain;
3631 smt_gain /= weight;
3633 return smt_gain;
3636 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3638 return default_scale_smt_power(sd, cpu);
3641 unsigned long scale_rt_power(int cpu)
3643 struct rq *rq = cpu_rq(cpu);
3644 u64 total, available;
3646 sched_avg_update(rq);
3648 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3649 available = total - rq->rt_avg;
3651 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3652 total = SCHED_LOAD_SCALE;
3654 total >>= SCHED_LOAD_SHIFT;
3656 return div_u64(available, total);
3659 static void update_cpu_power(struct sched_domain *sd, int cpu)
3661 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3662 unsigned long power = SCHED_LOAD_SCALE;
3663 struct sched_group *sdg = sd->groups;
3665 if (sched_feat(ARCH_POWER))
3666 power *= arch_scale_freq_power(sd, cpu);
3667 else
3668 power *= default_scale_freq_power(sd, cpu);
3670 power >>= SCHED_LOAD_SHIFT;
3672 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3673 if (sched_feat(ARCH_POWER))
3674 power *= arch_scale_smt_power(sd, cpu);
3675 else
3676 power *= default_scale_smt_power(sd, cpu);
3678 power >>= SCHED_LOAD_SHIFT;
3681 power *= scale_rt_power(cpu);
3682 power >>= SCHED_LOAD_SHIFT;
3684 if (!power)
3685 power = 1;
3687 sdg->cpu_power = power;
3690 static void update_group_power(struct sched_domain *sd, int cpu)
3692 struct sched_domain *child = sd->child;
3693 struct sched_group *group, *sdg = sd->groups;
3694 unsigned long power;
3696 if (!child) {
3697 update_cpu_power(sd, cpu);
3698 return;
3701 power = 0;
3703 group = child->groups;
3704 do {
3705 power += group->cpu_power;
3706 group = group->next;
3707 } while (group != child->groups);
3709 sdg->cpu_power = power;
3713 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3714 * @sd: The sched_domain whose statistics are to be updated.
3715 * @group: sched_group whose statistics are to be updated.
3716 * @this_cpu: Cpu for which load balance is currently performed.
3717 * @idle: Idle status of this_cpu
3718 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3719 * @sd_idle: Idle status of the sched_domain containing group.
3720 * @local_group: Does group contain this_cpu.
3721 * @cpus: Set of cpus considered for load balancing.
3722 * @balance: Should we balance.
3723 * @sgs: variable to hold the statistics for this group.
3725 static inline void update_sg_lb_stats(struct sched_domain *sd,
3726 struct sched_group *group, int this_cpu,
3727 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3728 int local_group, const struct cpumask *cpus,
3729 int *balance, struct sg_lb_stats *sgs)
3731 unsigned long load, max_cpu_load, min_cpu_load;
3732 int i;
3733 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3734 unsigned long sum_avg_load_per_task;
3735 unsigned long avg_load_per_task;
3737 if (local_group) {
3738 balance_cpu = group_first_cpu(group);
3739 if (balance_cpu == this_cpu)
3740 update_group_power(sd, this_cpu);
3743 /* Tally up the load of all CPUs in the group */
3744 sum_avg_load_per_task = avg_load_per_task = 0;
3745 max_cpu_load = 0;
3746 min_cpu_load = ~0UL;
3748 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3749 struct rq *rq = cpu_rq(i);
3751 if (*sd_idle && rq->nr_running)
3752 *sd_idle = 0;
3754 /* Bias balancing toward cpus of our domain */
3755 if (local_group) {
3756 if (idle_cpu(i) && !first_idle_cpu) {
3757 first_idle_cpu = 1;
3758 balance_cpu = i;
3761 load = target_load(i, load_idx);
3762 } else {
3763 load = source_load(i, load_idx);
3764 if (load > max_cpu_load)
3765 max_cpu_load = load;
3766 if (min_cpu_load > load)
3767 min_cpu_load = load;
3770 sgs->group_load += load;
3771 sgs->sum_nr_running += rq->nr_running;
3772 sgs->sum_weighted_load += weighted_cpuload(i);
3774 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3778 * First idle cpu or the first cpu(busiest) in this sched group
3779 * is eligible for doing load balancing at this and above
3780 * domains. In the newly idle case, we will allow all the cpu's
3781 * to do the newly idle load balance.
3783 if (idle != CPU_NEWLY_IDLE && local_group &&
3784 balance_cpu != this_cpu && balance) {
3785 *balance = 0;
3786 return;
3789 /* Adjust by relative CPU power of the group */
3790 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3794 * Consider the group unbalanced when the imbalance is larger
3795 * than the average weight of two tasks.
3797 * APZ: with cgroup the avg task weight can vary wildly and
3798 * might not be a suitable number - should we keep a
3799 * normalized nr_running number somewhere that negates
3800 * the hierarchy?
3802 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3803 group->cpu_power;
3805 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3806 sgs->group_imb = 1;
3808 sgs->group_capacity =
3809 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3813 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3814 * @sd: sched_domain whose statistics are to be updated.
3815 * @this_cpu: Cpu for which load balance is currently performed.
3816 * @idle: Idle status of this_cpu
3817 * @sd_idle: Idle status of the sched_domain containing group.
3818 * @cpus: Set of cpus considered for load balancing.
3819 * @balance: Should we balance.
3820 * @sds: variable to hold the statistics for this sched_domain.
3822 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3823 enum cpu_idle_type idle, int *sd_idle,
3824 const struct cpumask *cpus, int *balance,
3825 struct sd_lb_stats *sds)
3827 struct sched_domain *child = sd->child;
3828 struct sched_group *group = sd->groups;
3829 struct sg_lb_stats sgs;
3830 int load_idx, prefer_sibling = 0;
3832 if (child && child->flags & SD_PREFER_SIBLING)
3833 prefer_sibling = 1;
3835 init_sd_power_savings_stats(sd, sds, idle);
3836 load_idx = get_sd_load_idx(sd, idle);
3838 do {
3839 int local_group;
3841 local_group = cpumask_test_cpu(this_cpu,
3842 sched_group_cpus(group));
3843 memset(&sgs, 0, sizeof(sgs));
3844 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3845 local_group, cpus, balance, &sgs);
3847 if (local_group && balance && !(*balance))
3848 return;
3850 sds->total_load += sgs.group_load;
3851 sds->total_pwr += group->cpu_power;
3854 * In case the child domain prefers tasks go to siblings
3855 * first, lower the group capacity to one so that we'll try
3856 * and move all the excess tasks away.
3858 if (prefer_sibling)
3859 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3861 if (local_group) {
3862 sds->this_load = sgs.avg_load;
3863 sds->this = group;
3864 sds->this_nr_running = sgs.sum_nr_running;
3865 sds->this_load_per_task = sgs.sum_weighted_load;
3866 } else if (sgs.avg_load > sds->max_load &&
3867 (sgs.sum_nr_running > sgs.group_capacity ||
3868 sgs.group_imb)) {
3869 sds->max_load = sgs.avg_load;
3870 sds->busiest = group;
3871 sds->busiest_nr_running = sgs.sum_nr_running;
3872 sds->busiest_load_per_task = sgs.sum_weighted_load;
3873 sds->group_imb = sgs.group_imb;
3876 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3877 group = group->next;
3878 } while (group != sd->groups);
3882 * fix_small_imbalance - Calculate the minor imbalance that exists
3883 * amongst the groups of a sched_domain, during
3884 * load balancing.
3885 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3886 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3887 * @imbalance: Variable to store the imbalance.
3889 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3890 int this_cpu, unsigned long *imbalance)
3892 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3893 unsigned int imbn = 2;
3895 if (sds->this_nr_running) {
3896 sds->this_load_per_task /= sds->this_nr_running;
3897 if (sds->busiest_load_per_task >
3898 sds->this_load_per_task)
3899 imbn = 1;
3900 } else
3901 sds->this_load_per_task =
3902 cpu_avg_load_per_task(this_cpu);
3904 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3905 sds->busiest_load_per_task * imbn) {
3906 *imbalance = sds->busiest_load_per_task;
3907 return;
3911 * OK, we don't have enough imbalance to justify moving tasks,
3912 * however we may be able to increase total CPU power used by
3913 * moving them.
3916 pwr_now += sds->busiest->cpu_power *
3917 min(sds->busiest_load_per_task, sds->max_load);
3918 pwr_now += sds->this->cpu_power *
3919 min(sds->this_load_per_task, sds->this_load);
3920 pwr_now /= SCHED_LOAD_SCALE;
3922 /* Amount of load we'd subtract */
3923 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3924 sds->busiest->cpu_power;
3925 if (sds->max_load > tmp)
3926 pwr_move += sds->busiest->cpu_power *
3927 min(sds->busiest_load_per_task, sds->max_load - tmp);
3929 /* Amount of load we'd add */
3930 if (sds->max_load * sds->busiest->cpu_power <
3931 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3932 tmp = (sds->max_load * sds->busiest->cpu_power) /
3933 sds->this->cpu_power;
3934 else
3935 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3936 sds->this->cpu_power;
3937 pwr_move += sds->this->cpu_power *
3938 min(sds->this_load_per_task, sds->this_load + tmp);
3939 pwr_move /= SCHED_LOAD_SCALE;
3941 /* Move if we gain throughput */
3942 if (pwr_move > pwr_now)
3943 *imbalance = sds->busiest_load_per_task;
3947 * calculate_imbalance - Calculate the amount of imbalance present within the
3948 * groups of a given sched_domain during load balance.
3949 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3950 * @this_cpu: Cpu for which currently load balance is being performed.
3951 * @imbalance: The variable to store the imbalance.
3953 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3954 unsigned long *imbalance)
3956 unsigned long max_pull;
3958 * In the presence of smp nice balancing, certain scenarios can have
3959 * max load less than avg load(as we skip the groups at or below
3960 * its cpu_power, while calculating max_load..)
3962 if (sds->max_load < sds->avg_load) {
3963 *imbalance = 0;
3964 return fix_small_imbalance(sds, this_cpu, imbalance);
3967 /* Don't want to pull so many tasks that a group would go idle */
3968 max_pull = min(sds->max_load - sds->avg_load,
3969 sds->max_load - sds->busiest_load_per_task);
3971 /* How much load to actually move to equalise the imbalance */
3972 *imbalance = min(max_pull * sds->busiest->cpu_power,
3973 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3974 / SCHED_LOAD_SCALE;
3977 * if *imbalance is less than the average load per runnable task
3978 * there is no gaurantee that any tasks will be moved so we'll have
3979 * a think about bumping its value to force at least one task to be
3980 * moved
3982 if (*imbalance < sds->busiest_load_per_task)
3983 return fix_small_imbalance(sds, this_cpu, imbalance);
3986 /******* find_busiest_group() helpers end here *********************/
3989 * find_busiest_group - Returns the busiest group within the sched_domain
3990 * if there is an imbalance. If there isn't an imbalance, and
3991 * the user has opted for power-savings, it returns a group whose
3992 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3993 * such a group exists.
3995 * Also calculates the amount of weighted load which should be moved
3996 * to restore balance.
3998 * @sd: The sched_domain whose busiest group is to be returned.
3999 * @this_cpu: The cpu for which load balancing is currently being performed.
4000 * @imbalance: Variable which stores amount of weighted load which should
4001 * be moved to restore balance/put a group to idle.
4002 * @idle: The idle status of this_cpu.
4003 * @sd_idle: The idleness of sd
4004 * @cpus: The set of CPUs under consideration for load-balancing.
4005 * @balance: Pointer to a variable indicating if this_cpu
4006 * is the appropriate cpu to perform load balancing at this_level.
4008 * Returns: - the busiest group if imbalance exists.
4009 * - If no imbalance and user has opted for power-savings balance,
4010 * return the least loaded group whose CPUs can be
4011 * put to idle by rebalancing its tasks onto our group.
4013 static struct sched_group *
4014 find_busiest_group(struct sched_domain *sd, int this_cpu,
4015 unsigned long *imbalance, enum cpu_idle_type idle,
4016 int *sd_idle, const struct cpumask *cpus, int *balance)
4018 struct sd_lb_stats sds;
4020 memset(&sds, 0, sizeof(sds));
4023 * Compute the various statistics relavent for load balancing at
4024 * this level.
4026 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4027 balance, &sds);
4029 /* Cases where imbalance does not exist from POV of this_cpu */
4030 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4031 * at this level.
4032 * 2) There is no busy sibling group to pull from.
4033 * 3) This group is the busiest group.
4034 * 4) This group is more busy than the avg busieness at this
4035 * sched_domain.
4036 * 5) The imbalance is within the specified limit.
4037 * 6) Any rebalance would lead to ping-pong
4039 if (balance && !(*balance))
4040 goto ret;
4042 if (!sds.busiest || sds.busiest_nr_running == 0)
4043 goto out_balanced;
4045 if (sds.this_load >= sds.max_load)
4046 goto out_balanced;
4048 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4050 if (sds.this_load >= sds.avg_load)
4051 goto out_balanced;
4053 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4054 goto out_balanced;
4056 sds.busiest_load_per_task /= sds.busiest_nr_running;
4057 if (sds.group_imb)
4058 sds.busiest_load_per_task =
4059 min(sds.busiest_load_per_task, sds.avg_load);
4062 * We're trying to get all the cpus to the average_load, so we don't
4063 * want to push ourselves above the average load, nor do we wish to
4064 * reduce the max loaded cpu below the average load, as either of these
4065 * actions would just result in more rebalancing later, and ping-pong
4066 * tasks around. Thus we look for the minimum possible imbalance.
4067 * Negative imbalances (*we* are more loaded than anyone else) will
4068 * be counted as no imbalance for these purposes -- we can't fix that
4069 * by pulling tasks to us. Be careful of negative numbers as they'll
4070 * appear as very large values with unsigned longs.
4072 if (sds.max_load <= sds.busiest_load_per_task)
4073 goto out_balanced;
4075 /* Looks like there is an imbalance. Compute it */
4076 calculate_imbalance(&sds, this_cpu, imbalance);
4077 return sds.busiest;
4079 out_balanced:
4081 * There is no obvious imbalance. But check if we can do some balancing
4082 * to save power.
4084 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4085 return sds.busiest;
4086 ret:
4087 *imbalance = 0;
4088 return NULL;
4092 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4094 static struct rq *
4095 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4096 unsigned long imbalance, const struct cpumask *cpus)
4098 struct rq *busiest = NULL, *rq;
4099 unsigned long max_load = 0;
4100 int i;
4102 for_each_cpu(i, sched_group_cpus(group)) {
4103 unsigned long power = power_of(i);
4104 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4105 unsigned long wl;
4107 if (!cpumask_test_cpu(i, cpus))
4108 continue;
4110 rq = cpu_rq(i);
4111 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4112 wl /= power;
4114 if (capacity && rq->nr_running == 1 && wl > imbalance)
4115 continue;
4117 if (wl > max_load) {
4118 max_load = wl;
4119 busiest = rq;
4123 return busiest;
4127 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4128 * so long as it is large enough.
4130 #define MAX_PINNED_INTERVAL 512
4132 /* Working cpumask for load_balance and load_balance_newidle. */
4133 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4136 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4137 * tasks if there is an imbalance.
4139 static int load_balance(int this_cpu, struct rq *this_rq,
4140 struct sched_domain *sd, enum cpu_idle_type idle,
4141 int *balance)
4143 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4144 struct sched_group *group;
4145 unsigned long imbalance;
4146 struct rq *busiest;
4147 unsigned long flags;
4148 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4150 cpumask_copy(cpus, cpu_active_mask);
4153 * When power savings policy is enabled for the parent domain, idle
4154 * sibling can pick up load irrespective of busy siblings. In this case,
4155 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4156 * portraying it as CPU_NOT_IDLE.
4158 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4159 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4160 sd_idle = 1;
4162 schedstat_inc(sd, lb_count[idle]);
4164 redo:
4165 update_shares(sd);
4166 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4167 cpus, balance);
4169 if (*balance == 0)
4170 goto out_balanced;
4172 if (!group) {
4173 schedstat_inc(sd, lb_nobusyg[idle]);
4174 goto out_balanced;
4177 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4178 if (!busiest) {
4179 schedstat_inc(sd, lb_nobusyq[idle]);
4180 goto out_balanced;
4183 BUG_ON(busiest == this_rq);
4185 schedstat_add(sd, lb_imbalance[idle], imbalance);
4187 ld_moved = 0;
4188 if (busiest->nr_running > 1) {
4190 * Attempt to move tasks. If find_busiest_group has found
4191 * an imbalance but busiest->nr_running <= 1, the group is
4192 * still unbalanced. ld_moved simply stays zero, so it is
4193 * correctly treated as an imbalance.
4195 local_irq_save(flags);
4196 double_rq_lock(this_rq, busiest);
4197 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4198 imbalance, sd, idle, &all_pinned);
4199 double_rq_unlock(this_rq, busiest);
4200 local_irq_restore(flags);
4203 * some other cpu did the load balance for us.
4205 if (ld_moved && this_cpu != smp_processor_id())
4206 resched_cpu(this_cpu);
4208 /* All tasks on this runqueue were pinned by CPU affinity */
4209 if (unlikely(all_pinned)) {
4210 cpumask_clear_cpu(cpu_of(busiest), cpus);
4211 if (!cpumask_empty(cpus))
4212 goto redo;
4213 goto out_balanced;
4217 if (!ld_moved) {
4218 schedstat_inc(sd, lb_failed[idle]);
4219 sd->nr_balance_failed++;
4221 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4223 raw_spin_lock_irqsave(&busiest->lock, flags);
4225 /* don't kick the migration_thread, if the curr
4226 * task on busiest cpu can't be moved to this_cpu
4228 if (!cpumask_test_cpu(this_cpu,
4229 &busiest->curr->cpus_allowed)) {
4230 raw_spin_unlock_irqrestore(&busiest->lock,
4231 flags);
4232 all_pinned = 1;
4233 goto out_one_pinned;
4236 if (!busiest->active_balance) {
4237 busiest->active_balance = 1;
4238 busiest->push_cpu = this_cpu;
4239 active_balance = 1;
4241 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4242 if (active_balance)
4243 wake_up_process(busiest->migration_thread);
4246 * We've kicked active balancing, reset the failure
4247 * counter.
4249 sd->nr_balance_failed = sd->cache_nice_tries+1;
4251 } else
4252 sd->nr_balance_failed = 0;
4254 if (likely(!active_balance)) {
4255 /* We were unbalanced, so reset the balancing interval */
4256 sd->balance_interval = sd->min_interval;
4257 } else {
4259 * If we've begun active balancing, start to back off. This
4260 * case may not be covered by the all_pinned logic if there
4261 * is only 1 task on the busy runqueue (because we don't call
4262 * move_tasks).
4264 if (sd->balance_interval < sd->max_interval)
4265 sd->balance_interval *= 2;
4268 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4269 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4270 ld_moved = -1;
4272 goto out;
4274 out_balanced:
4275 schedstat_inc(sd, lb_balanced[idle]);
4277 sd->nr_balance_failed = 0;
4279 out_one_pinned:
4280 /* tune up the balancing interval */
4281 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4282 (sd->balance_interval < sd->max_interval))
4283 sd->balance_interval *= 2;
4285 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4286 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4287 ld_moved = -1;
4288 else
4289 ld_moved = 0;
4290 out:
4291 if (ld_moved)
4292 update_shares(sd);
4293 return ld_moved;
4297 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4298 * tasks if there is an imbalance.
4300 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4301 * this_rq is locked.
4303 static int
4304 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4306 struct sched_group *group;
4307 struct rq *busiest = NULL;
4308 unsigned long imbalance;
4309 int ld_moved = 0;
4310 int sd_idle = 0;
4311 int all_pinned = 0;
4312 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4314 cpumask_copy(cpus, cpu_active_mask);
4317 * When power savings policy is enabled for the parent domain, idle
4318 * sibling can pick up load irrespective of busy siblings. In this case,
4319 * let the state of idle sibling percolate up as IDLE, instead of
4320 * portraying it as CPU_NOT_IDLE.
4322 if (sd->flags & SD_SHARE_CPUPOWER &&
4323 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4324 sd_idle = 1;
4326 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4327 redo:
4328 update_shares_locked(this_rq, sd);
4329 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4330 &sd_idle, cpus, NULL);
4331 if (!group) {
4332 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4333 goto out_balanced;
4336 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4337 if (!busiest) {
4338 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4339 goto out_balanced;
4342 BUG_ON(busiest == this_rq);
4344 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4346 ld_moved = 0;
4347 if (busiest->nr_running > 1) {
4348 /* Attempt to move tasks */
4349 double_lock_balance(this_rq, busiest);
4350 /* this_rq->clock is already updated */
4351 update_rq_clock(busiest);
4352 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4353 imbalance, sd, CPU_NEWLY_IDLE,
4354 &all_pinned);
4355 double_unlock_balance(this_rq, busiest);
4357 if (unlikely(all_pinned)) {
4358 cpumask_clear_cpu(cpu_of(busiest), cpus);
4359 if (!cpumask_empty(cpus))
4360 goto redo;
4364 if (!ld_moved) {
4365 int active_balance = 0;
4367 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4368 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4369 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4370 return -1;
4372 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4373 return -1;
4375 if (sd->nr_balance_failed++ < 2)
4376 return -1;
4379 * The only task running in a non-idle cpu can be moved to this
4380 * cpu in an attempt to completely freeup the other CPU
4381 * package. The same method used to move task in load_balance()
4382 * have been extended for load_balance_newidle() to speedup
4383 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4385 * The package power saving logic comes from
4386 * find_busiest_group(). If there are no imbalance, then
4387 * f_b_g() will return NULL. However when sched_mc={1,2} then
4388 * f_b_g() will select a group from which a running task may be
4389 * pulled to this cpu in order to make the other package idle.
4390 * If there is no opportunity to make a package idle and if
4391 * there are no imbalance, then f_b_g() will return NULL and no
4392 * action will be taken in load_balance_newidle().
4394 * Under normal task pull operation due to imbalance, there
4395 * will be more than one task in the source run queue and
4396 * move_tasks() will succeed. ld_moved will be true and this
4397 * active balance code will not be triggered.
4400 /* Lock busiest in correct order while this_rq is held */
4401 double_lock_balance(this_rq, busiest);
4404 * don't kick the migration_thread, if the curr
4405 * task on busiest cpu can't be moved to this_cpu
4407 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4408 double_unlock_balance(this_rq, busiest);
4409 all_pinned = 1;
4410 return ld_moved;
4413 if (!busiest->active_balance) {
4414 busiest->active_balance = 1;
4415 busiest->push_cpu = this_cpu;
4416 active_balance = 1;
4419 double_unlock_balance(this_rq, busiest);
4421 * Should not call ttwu while holding a rq->lock
4423 raw_spin_unlock(&this_rq->lock);
4424 if (active_balance)
4425 wake_up_process(busiest->migration_thread);
4426 raw_spin_lock(&this_rq->lock);
4428 } else
4429 sd->nr_balance_failed = 0;
4431 update_shares_locked(this_rq, sd);
4432 return ld_moved;
4434 out_balanced:
4435 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4436 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4437 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4438 return -1;
4439 sd->nr_balance_failed = 0;
4441 return 0;
4445 * idle_balance is called by schedule() if this_cpu is about to become
4446 * idle. Attempts to pull tasks from other CPUs.
4448 static void idle_balance(int this_cpu, struct rq *this_rq)
4450 struct sched_domain *sd;
4451 int pulled_task = 0;
4452 unsigned long next_balance = jiffies + HZ;
4454 this_rq->idle_stamp = this_rq->clock;
4456 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4457 return;
4459 for_each_domain(this_cpu, sd) {
4460 unsigned long interval;
4462 if (!(sd->flags & SD_LOAD_BALANCE))
4463 continue;
4465 if (sd->flags & SD_BALANCE_NEWIDLE)
4466 /* If we've pulled tasks over stop searching: */
4467 pulled_task = load_balance_newidle(this_cpu, this_rq,
4468 sd);
4470 interval = msecs_to_jiffies(sd->balance_interval);
4471 if (time_after(next_balance, sd->last_balance + interval))
4472 next_balance = sd->last_balance + interval;
4473 if (pulled_task) {
4474 this_rq->idle_stamp = 0;
4475 break;
4478 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4480 * We are going idle. next_balance may be set based on
4481 * a busy processor. So reset next_balance.
4483 this_rq->next_balance = next_balance;
4488 * active_load_balance is run by migration threads. It pushes running tasks
4489 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4490 * running on each physical CPU where possible, and avoids physical /
4491 * logical imbalances.
4493 * Called with busiest_rq locked.
4495 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4497 int target_cpu = busiest_rq->push_cpu;
4498 struct sched_domain *sd;
4499 struct rq *target_rq;
4501 /* Is there any task to move? */
4502 if (busiest_rq->nr_running <= 1)
4503 return;
4505 target_rq = cpu_rq(target_cpu);
4508 * This condition is "impossible", if it occurs
4509 * we need to fix it. Originally reported by
4510 * Bjorn Helgaas on a 128-cpu setup.
4512 BUG_ON(busiest_rq == target_rq);
4514 /* move a task from busiest_rq to target_rq */
4515 double_lock_balance(busiest_rq, target_rq);
4516 update_rq_clock(busiest_rq);
4517 update_rq_clock(target_rq);
4519 /* Search for an sd spanning us and the target CPU. */
4520 for_each_domain(target_cpu, sd) {
4521 if ((sd->flags & SD_LOAD_BALANCE) &&
4522 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4523 break;
4526 if (likely(sd)) {
4527 schedstat_inc(sd, alb_count);
4529 if (move_one_task(target_rq, target_cpu, busiest_rq,
4530 sd, CPU_IDLE))
4531 schedstat_inc(sd, alb_pushed);
4532 else
4533 schedstat_inc(sd, alb_failed);
4535 double_unlock_balance(busiest_rq, target_rq);
4538 #ifdef CONFIG_NO_HZ
4539 static struct {
4540 atomic_t load_balancer;
4541 cpumask_var_t cpu_mask;
4542 cpumask_var_t ilb_grp_nohz_mask;
4543 } nohz ____cacheline_aligned = {
4544 .load_balancer = ATOMIC_INIT(-1),
4547 int get_nohz_load_balancer(void)
4549 return atomic_read(&nohz.load_balancer);
4552 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4554 * lowest_flag_domain - Return lowest sched_domain containing flag.
4555 * @cpu: The cpu whose lowest level of sched domain is to
4556 * be returned.
4557 * @flag: The flag to check for the lowest sched_domain
4558 * for the given cpu.
4560 * Returns the lowest sched_domain of a cpu which contains the given flag.
4562 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4564 struct sched_domain *sd;
4566 for_each_domain(cpu, sd)
4567 if (sd && (sd->flags & flag))
4568 break;
4570 return sd;
4574 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4575 * @cpu: The cpu whose domains we're iterating over.
4576 * @sd: variable holding the value of the power_savings_sd
4577 * for cpu.
4578 * @flag: The flag to filter the sched_domains to be iterated.
4580 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4581 * set, starting from the lowest sched_domain to the highest.
4583 #define for_each_flag_domain(cpu, sd, flag) \
4584 for (sd = lowest_flag_domain(cpu, flag); \
4585 (sd && (sd->flags & flag)); sd = sd->parent)
4588 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4589 * @ilb_group: group to be checked for semi-idleness
4591 * Returns: 1 if the group is semi-idle. 0 otherwise.
4593 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4594 * and atleast one non-idle CPU. This helper function checks if the given
4595 * sched_group is semi-idle or not.
4597 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4599 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4600 sched_group_cpus(ilb_group));
4603 * A sched_group is semi-idle when it has atleast one busy cpu
4604 * and atleast one idle cpu.
4606 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4607 return 0;
4609 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4610 return 0;
4612 return 1;
4615 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4616 * @cpu: The cpu which is nominating a new idle_load_balancer.
4618 * Returns: Returns the id of the idle load balancer if it exists,
4619 * Else, returns >= nr_cpu_ids.
4621 * This algorithm picks the idle load balancer such that it belongs to a
4622 * semi-idle powersavings sched_domain. The idea is to try and avoid
4623 * completely idle packages/cores just for the purpose of idle load balancing
4624 * when there are other idle cpu's which are better suited for that job.
4626 static int find_new_ilb(int cpu)
4628 struct sched_domain *sd;
4629 struct sched_group *ilb_group;
4632 * Have idle load balancer selection from semi-idle packages only
4633 * when power-aware load balancing is enabled
4635 if (!(sched_smt_power_savings || sched_mc_power_savings))
4636 goto out_done;
4639 * Optimize for the case when we have no idle CPUs or only one
4640 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4642 if (cpumask_weight(nohz.cpu_mask) < 2)
4643 goto out_done;
4645 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4646 ilb_group = sd->groups;
4648 do {
4649 if (is_semi_idle_group(ilb_group))
4650 return cpumask_first(nohz.ilb_grp_nohz_mask);
4652 ilb_group = ilb_group->next;
4654 } while (ilb_group != sd->groups);
4657 out_done:
4658 return cpumask_first(nohz.cpu_mask);
4660 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4661 static inline int find_new_ilb(int call_cpu)
4663 return cpumask_first(nohz.cpu_mask);
4665 #endif
4668 * This routine will try to nominate the ilb (idle load balancing)
4669 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4670 * load balancing on behalf of all those cpus. If all the cpus in the system
4671 * go into this tickless mode, then there will be no ilb owner (as there is
4672 * no need for one) and all the cpus will sleep till the next wakeup event
4673 * arrives...
4675 * For the ilb owner, tick is not stopped. And this tick will be used
4676 * for idle load balancing. ilb owner will still be part of
4677 * nohz.cpu_mask..
4679 * While stopping the tick, this cpu will become the ilb owner if there
4680 * is no other owner. And will be the owner till that cpu becomes busy
4681 * or if all cpus in the system stop their ticks at which point
4682 * there is no need for ilb owner.
4684 * When the ilb owner becomes busy, it nominates another owner, during the
4685 * next busy scheduler_tick()
4687 int select_nohz_load_balancer(int stop_tick)
4689 int cpu = smp_processor_id();
4691 if (stop_tick) {
4692 cpu_rq(cpu)->in_nohz_recently = 1;
4694 if (!cpu_active(cpu)) {
4695 if (atomic_read(&nohz.load_balancer) != cpu)
4696 return 0;
4699 * If we are going offline and still the leader,
4700 * give up!
4702 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4703 BUG();
4705 return 0;
4708 cpumask_set_cpu(cpu, nohz.cpu_mask);
4710 /* time for ilb owner also to sleep */
4711 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4712 if (atomic_read(&nohz.load_balancer) == cpu)
4713 atomic_set(&nohz.load_balancer, -1);
4714 return 0;
4717 if (atomic_read(&nohz.load_balancer) == -1) {
4718 /* make me the ilb owner */
4719 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4720 return 1;
4721 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4722 int new_ilb;
4724 if (!(sched_smt_power_savings ||
4725 sched_mc_power_savings))
4726 return 1;
4728 * Check to see if there is a more power-efficient
4729 * ilb.
4731 new_ilb = find_new_ilb(cpu);
4732 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4733 atomic_set(&nohz.load_balancer, -1);
4734 resched_cpu(new_ilb);
4735 return 0;
4737 return 1;
4739 } else {
4740 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4741 return 0;
4743 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4745 if (atomic_read(&nohz.load_balancer) == cpu)
4746 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4747 BUG();
4749 return 0;
4751 #endif
4753 static DEFINE_SPINLOCK(balancing);
4756 * It checks each scheduling domain to see if it is due to be balanced,
4757 * and initiates a balancing operation if so.
4759 * Balancing parameters are set up in arch_init_sched_domains.
4761 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4763 int balance = 1;
4764 struct rq *rq = cpu_rq(cpu);
4765 unsigned long interval;
4766 struct sched_domain *sd;
4767 /* Earliest time when we have to do rebalance again */
4768 unsigned long next_balance = jiffies + 60*HZ;
4769 int update_next_balance = 0;
4770 int need_serialize;
4772 for_each_domain(cpu, sd) {
4773 if (!(sd->flags & SD_LOAD_BALANCE))
4774 continue;
4776 interval = sd->balance_interval;
4777 if (idle != CPU_IDLE)
4778 interval *= sd->busy_factor;
4780 /* scale ms to jiffies */
4781 interval = msecs_to_jiffies(interval);
4782 if (unlikely(!interval))
4783 interval = 1;
4784 if (interval > HZ*NR_CPUS/10)
4785 interval = HZ*NR_CPUS/10;
4787 need_serialize = sd->flags & SD_SERIALIZE;
4789 if (need_serialize) {
4790 if (!spin_trylock(&balancing))
4791 goto out;
4794 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4795 if (load_balance(cpu, rq, sd, idle, &balance)) {
4797 * We've pulled tasks over so either we're no
4798 * longer idle, or one of our SMT siblings is
4799 * not idle.
4801 idle = CPU_NOT_IDLE;
4803 sd->last_balance = jiffies;
4805 if (need_serialize)
4806 spin_unlock(&balancing);
4807 out:
4808 if (time_after(next_balance, sd->last_balance + interval)) {
4809 next_balance = sd->last_balance + interval;
4810 update_next_balance = 1;
4814 * Stop the load balance at this level. There is another
4815 * CPU in our sched group which is doing load balancing more
4816 * actively.
4818 if (!balance)
4819 break;
4823 * next_balance will be updated only when there is a need.
4824 * When the cpu is attached to null domain for ex, it will not be
4825 * updated.
4827 if (likely(update_next_balance))
4828 rq->next_balance = next_balance;
4832 * run_rebalance_domains is triggered when needed from the scheduler tick.
4833 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4834 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4836 static void run_rebalance_domains(struct softirq_action *h)
4838 int this_cpu = smp_processor_id();
4839 struct rq *this_rq = cpu_rq(this_cpu);
4840 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4841 CPU_IDLE : CPU_NOT_IDLE;
4843 rebalance_domains(this_cpu, idle);
4845 #ifdef CONFIG_NO_HZ
4847 * If this cpu is the owner for idle load balancing, then do the
4848 * balancing on behalf of the other idle cpus whose ticks are
4849 * stopped.
4851 if (this_rq->idle_at_tick &&
4852 atomic_read(&nohz.load_balancer) == this_cpu) {
4853 struct rq *rq;
4854 int balance_cpu;
4856 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4857 if (balance_cpu == this_cpu)
4858 continue;
4861 * If this cpu gets work to do, stop the load balancing
4862 * work being done for other cpus. Next load
4863 * balancing owner will pick it up.
4865 if (need_resched())
4866 break;
4868 rebalance_domains(balance_cpu, CPU_IDLE);
4870 rq = cpu_rq(balance_cpu);
4871 if (time_after(this_rq->next_balance, rq->next_balance))
4872 this_rq->next_balance = rq->next_balance;
4875 #endif
4878 static inline int on_null_domain(int cpu)
4880 return !rcu_dereference(cpu_rq(cpu)->sd);
4884 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4886 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4887 * idle load balancing owner or decide to stop the periodic load balancing,
4888 * if the whole system is idle.
4890 static inline void trigger_load_balance(struct rq *rq, int cpu)
4892 #ifdef CONFIG_NO_HZ
4894 * If we were in the nohz mode recently and busy at the current
4895 * scheduler tick, then check if we need to nominate new idle
4896 * load balancer.
4898 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4899 rq->in_nohz_recently = 0;
4901 if (atomic_read(&nohz.load_balancer) == cpu) {
4902 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4903 atomic_set(&nohz.load_balancer, -1);
4906 if (atomic_read(&nohz.load_balancer) == -1) {
4907 int ilb = find_new_ilb(cpu);
4909 if (ilb < nr_cpu_ids)
4910 resched_cpu(ilb);
4915 * If this cpu is idle and doing idle load balancing for all the
4916 * cpus with ticks stopped, is it time for that to stop?
4918 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4919 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4920 resched_cpu(cpu);
4921 return;
4925 * If this cpu is idle and the idle load balancing is done by
4926 * someone else, then no need raise the SCHED_SOFTIRQ
4928 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4929 cpumask_test_cpu(cpu, nohz.cpu_mask))
4930 return;
4931 #endif
4932 /* Don't need to rebalance while attached to NULL domain */
4933 if (time_after_eq(jiffies, rq->next_balance) &&
4934 likely(!on_null_domain(cpu)))
4935 raise_softirq(SCHED_SOFTIRQ);
4938 #else /* CONFIG_SMP */
4941 * on UP we do not need to balance between CPUs:
4943 static inline void idle_balance(int cpu, struct rq *rq)
4947 #endif
4949 DEFINE_PER_CPU(struct kernel_stat, kstat);
4951 EXPORT_PER_CPU_SYMBOL(kstat);
4954 * Return any ns on the sched_clock that have not yet been accounted in
4955 * @p in case that task is currently running.
4957 * Called with task_rq_lock() held on @rq.
4959 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4961 u64 ns = 0;
4963 if (task_current(rq, p)) {
4964 update_rq_clock(rq);
4965 ns = rq->clock - p->se.exec_start;
4966 if ((s64)ns < 0)
4967 ns = 0;
4970 return ns;
4973 unsigned long long task_delta_exec(struct task_struct *p)
4975 unsigned long flags;
4976 struct rq *rq;
4977 u64 ns = 0;
4979 rq = task_rq_lock(p, &flags);
4980 ns = do_task_delta_exec(p, rq);
4981 task_rq_unlock(rq, &flags);
4983 return ns;
4987 * Return accounted runtime for the task.
4988 * In case the task is currently running, return the runtime plus current's
4989 * pending runtime that have not been accounted yet.
4991 unsigned long long task_sched_runtime(struct task_struct *p)
4993 unsigned long flags;
4994 struct rq *rq;
4995 u64 ns = 0;
4997 rq = task_rq_lock(p, &flags);
4998 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4999 task_rq_unlock(rq, &flags);
5001 return ns;
5005 * Return sum_exec_runtime for the thread group.
5006 * In case the task is currently running, return the sum plus current's
5007 * pending runtime that have not been accounted yet.
5009 * Note that the thread group might have other running tasks as well,
5010 * so the return value not includes other pending runtime that other
5011 * running tasks might have.
5013 unsigned long long thread_group_sched_runtime(struct task_struct *p)
5015 struct task_cputime totals;
5016 unsigned long flags;
5017 struct rq *rq;
5018 u64 ns;
5020 rq = task_rq_lock(p, &flags);
5021 thread_group_cputime(p, &totals);
5022 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
5023 task_rq_unlock(rq, &flags);
5025 return ns;
5029 * Account user cpu time to a process.
5030 * @p: the process that the cpu time gets accounted to
5031 * @cputime: the cpu time spent in user space since the last update
5032 * @cputime_scaled: cputime scaled by cpu frequency
5034 void account_user_time(struct task_struct *p, cputime_t cputime,
5035 cputime_t cputime_scaled)
5037 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5038 cputime64_t tmp;
5040 /* Add user time to process. */
5041 p->utime = cputime_add(p->utime, cputime);
5042 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5043 account_group_user_time(p, cputime);
5045 /* Add user time to cpustat. */
5046 tmp = cputime_to_cputime64(cputime);
5047 if (TASK_NICE(p) > 0)
5048 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5049 else
5050 cpustat->user = cputime64_add(cpustat->user, tmp);
5052 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5053 /* Account for user time used */
5054 acct_update_integrals(p);
5058 * Account guest cpu time to a process.
5059 * @p: the process that the cpu time gets accounted to
5060 * @cputime: the cpu time spent in virtual machine since the last update
5061 * @cputime_scaled: cputime scaled by cpu frequency
5063 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5064 cputime_t cputime_scaled)
5066 cputime64_t tmp;
5067 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5069 tmp = cputime_to_cputime64(cputime);
5071 /* Add guest time to process. */
5072 p->utime = cputime_add(p->utime, cputime);
5073 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5074 account_group_user_time(p, cputime);
5075 p->gtime = cputime_add(p->gtime, cputime);
5077 /* Add guest time to cpustat. */
5078 if (TASK_NICE(p) > 0) {
5079 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5080 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
5081 } else {
5082 cpustat->user = cputime64_add(cpustat->user, tmp);
5083 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5088 * Account system cpu time to a process.
5089 * @p: the process that the cpu time gets accounted to
5090 * @hardirq_offset: the offset to subtract from hardirq_count()
5091 * @cputime: the cpu time spent in kernel space since the last update
5092 * @cputime_scaled: cputime scaled by cpu frequency
5094 void account_system_time(struct task_struct *p, int hardirq_offset,
5095 cputime_t cputime, cputime_t cputime_scaled)
5097 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5098 cputime64_t tmp;
5100 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5101 account_guest_time(p, cputime, cputime_scaled);
5102 return;
5105 /* Add system time to process. */
5106 p->stime = cputime_add(p->stime, cputime);
5107 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5108 account_group_system_time(p, cputime);
5110 /* Add system time to cpustat. */
5111 tmp = cputime_to_cputime64(cputime);
5112 if (hardirq_count() - hardirq_offset)
5113 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5114 else if (softirq_count())
5115 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5116 else
5117 cpustat->system = cputime64_add(cpustat->system, tmp);
5119 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5121 /* Account for system time used */
5122 acct_update_integrals(p);
5126 * Account for involuntary wait time.
5127 * @steal: the cpu time spent in involuntary wait
5129 void account_steal_time(cputime_t cputime)
5131 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5132 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5134 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5138 * Account for idle time.
5139 * @cputime: the cpu time spent in idle wait
5141 void account_idle_time(cputime_t cputime)
5143 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5144 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5145 struct rq *rq = this_rq();
5147 if (atomic_read(&rq->nr_iowait) > 0)
5148 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5149 else
5150 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5153 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5156 * Account a single tick of cpu time.
5157 * @p: the process that the cpu time gets accounted to
5158 * @user_tick: indicates if the tick is a user or a system tick
5160 void account_process_tick(struct task_struct *p, int user_tick)
5162 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5163 struct rq *rq = this_rq();
5165 if (user_tick)
5166 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5167 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5168 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5169 one_jiffy_scaled);
5170 else
5171 account_idle_time(cputime_one_jiffy);
5175 * Account multiple ticks of steal time.
5176 * @p: the process from which the cpu time has been stolen
5177 * @ticks: number of stolen ticks
5179 void account_steal_ticks(unsigned long ticks)
5181 account_steal_time(jiffies_to_cputime(ticks));
5185 * Account multiple ticks of idle time.
5186 * @ticks: number of stolen ticks
5188 void account_idle_ticks(unsigned long ticks)
5190 account_idle_time(jiffies_to_cputime(ticks));
5193 #endif
5196 * Use precise platform statistics if available:
5198 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5199 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5201 *ut = p->utime;
5202 *st = p->stime;
5205 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5207 struct task_cputime cputime;
5209 thread_group_cputime(p, &cputime);
5211 *ut = cputime.utime;
5212 *st = cputime.stime;
5214 #else
5216 #ifndef nsecs_to_cputime
5217 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
5218 #endif
5220 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5222 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
5225 * Use CFS's precise accounting:
5227 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
5229 if (total) {
5230 u64 temp;
5232 temp = (u64)(rtime * utime);
5233 do_div(temp, total);
5234 utime = (cputime_t)temp;
5235 } else
5236 utime = rtime;
5239 * Compare with previous values, to keep monotonicity:
5241 p->prev_utime = max(p->prev_utime, utime);
5242 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
5244 *ut = p->prev_utime;
5245 *st = p->prev_stime;
5249 * Must be called with siglock held.
5251 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5253 struct signal_struct *sig = p->signal;
5254 struct task_cputime cputime;
5255 cputime_t rtime, utime, total;
5257 thread_group_cputime(p, &cputime);
5259 total = cputime_add(cputime.utime, cputime.stime);
5260 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5262 if (total) {
5263 u64 temp;
5265 temp = (u64)(rtime * cputime.utime);
5266 do_div(temp, total);
5267 utime = (cputime_t)temp;
5268 } else
5269 utime = rtime;
5271 sig->prev_utime = max(sig->prev_utime, utime);
5272 sig->prev_stime = max(sig->prev_stime,
5273 cputime_sub(rtime, sig->prev_utime));
5275 *ut = sig->prev_utime;
5276 *st = sig->prev_stime;
5278 #endif
5281 * This function gets called by the timer code, with HZ frequency.
5282 * We call it with interrupts disabled.
5284 * It also gets called by the fork code, when changing the parent's
5285 * timeslices.
5287 void scheduler_tick(void)
5289 int cpu = smp_processor_id();
5290 struct rq *rq = cpu_rq(cpu);
5291 struct task_struct *curr = rq->curr;
5293 sched_clock_tick();
5295 raw_spin_lock(&rq->lock);
5296 update_rq_clock(rq);
5297 update_cpu_load(rq);
5298 curr->sched_class->task_tick(rq, curr, 0);
5299 raw_spin_unlock(&rq->lock);
5301 perf_event_task_tick(curr, cpu);
5303 #ifdef CONFIG_SMP
5304 rq->idle_at_tick = idle_cpu(cpu);
5305 trigger_load_balance(rq, cpu);
5306 #endif
5309 notrace unsigned long get_parent_ip(unsigned long addr)
5311 if (in_lock_functions(addr)) {
5312 addr = CALLER_ADDR2;
5313 if (in_lock_functions(addr))
5314 addr = CALLER_ADDR3;
5316 return addr;
5319 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5320 defined(CONFIG_PREEMPT_TRACER))
5322 void __kprobes add_preempt_count(int val)
5324 #ifdef CONFIG_DEBUG_PREEMPT
5326 * Underflow?
5328 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5329 return;
5330 #endif
5331 preempt_count() += val;
5332 #ifdef CONFIG_DEBUG_PREEMPT
5334 * Spinlock count overflowing soon?
5336 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5337 PREEMPT_MASK - 10);
5338 #endif
5339 if (preempt_count() == val)
5340 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5342 EXPORT_SYMBOL(add_preempt_count);
5344 void __kprobes sub_preempt_count(int val)
5346 #ifdef CONFIG_DEBUG_PREEMPT
5348 * Underflow?
5350 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5351 return;
5353 * Is the spinlock portion underflowing?
5355 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5356 !(preempt_count() & PREEMPT_MASK)))
5357 return;
5358 #endif
5360 if (preempt_count() == val)
5361 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5362 preempt_count() -= val;
5364 EXPORT_SYMBOL(sub_preempt_count);
5366 #endif
5369 * Print scheduling while atomic bug:
5371 static noinline void __schedule_bug(struct task_struct *prev)
5373 struct pt_regs *regs = get_irq_regs();
5375 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5376 prev->comm, prev->pid, preempt_count());
5378 debug_show_held_locks(prev);
5379 print_modules();
5380 if (irqs_disabled())
5381 print_irqtrace_events(prev);
5383 if (regs)
5384 show_regs(regs);
5385 else
5386 dump_stack();
5390 * Various schedule()-time debugging checks and statistics:
5392 static inline void schedule_debug(struct task_struct *prev)
5395 * Test if we are atomic. Since do_exit() needs to call into
5396 * schedule() atomically, we ignore that path for now.
5397 * Otherwise, whine if we are scheduling when we should not be.
5399 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5400 __schedule_bug(prev);
5402 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5404 schedstat_inc(this_rq(), sched_count);
5405 #ifdef CONFIG_SCHEDSTATS
5406 if (unlikely(prev->lock_depth >= 0)) {
5407 schedstat_inc(this_rq(), bkl_count);
5408 schedstat_inc(prev, sched_info.bkl_count);
5410 #endif
5413 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5415 if (prev->state == TASK_RUNNING) {
5416 u64 runtime = prev->se.sum_exec_runtime;
5418 runtime -= prev->se.prev_sum_exec_runtime;
5419 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5422 * In order to avoid avg_overlap growing stale when we are
5423 * indeed overlapping and hence not getting put to sleep, grow
5424 * the avg_overlap on preemption.
5426 * We use the average preemption runtime because that
5427 * correlates to the amount of cache footprint a task can
5428 * build up.
5430 update_avg(&prev->se.avg_overlap, runtime);
5432 prev->sched_class->put_prev_task(rq, prev);
5436 * Pick up the highest-prio task:
5438 static inline struct task_struct *
5439 pick_next_task(struct rq *rq)
5441 const struct sched_class *class;
5442 struct task_struct *p;
5445 * Optimization: we know that if all tasks are in
5446 * the fair class we can call that function directly:
5448 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5449 p = fair_sched_class.pick_next_task(rq);
5450 if (likely(p))
5451 return p;
5454 class = sched_class_highest;
5455 for ( ; ; ) {
5456 p = class->pick_next_task(rq);
5457 if (p)
5458 return p;
5460 * Will never be NULL as the idle class always
5461 * returns a non-NULL p:
5463 class = class->next;
5468 * schedule() is the main scheduler function.
5470 asmlinkage void __sched schedule(void)
5472 struct task_struct *prev, *next;
5473 unsigned long *switch_count;
5474 struct rq *rq;
5475 int cpu;
5477 need_resched:
5478 preempt_disable();
5479 cpu = smp_processor_id();
5480 rq = cpu_rq(cpu);
5481 rcu_sched_qs(cpu);
5482 prev = rq->curr;
5483 switch_count = &prev->nivcsw;
5485 release_kernel_lock(prev);
5486 need_resched_nonpreemptible:
5488 schedule_debug(prev);
5490 if (sched_feat(HRTICK))
5491 hrtick_clear(rq);
5493 raw_spin_lock_irq(&rq->lock);
5494 update_rq_clock(rq);
5495 clear_tsk_need_resched(prev);
5497 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5498 if (unlikely(signal_pending_state(prev->state, prev)))
5499 prev->state = TASK_RUNNING;
5500 else
5501 deactivate_task(rq, prev, 1);
5502 switch_count = &prev->nvcsw;
5505 pre_schedule(rq, prev);
5507 if (unlikely(!rq->nr_running))
5508 idle_balance(cpu, rq);
5510 put_prev_task(rq, prev);
5511 next = pick_next_task(rq);
5513 if (likely(prev != next)) {
5514 sched_info_switch(prev, next);
5515 perf_event_task_sched_out(prev, next, cpu);
5517 rq->nr_switches++;
5518 rq->curr = next;
5519 ++*switch_count;
5521 context_switch(rq, prev, next); /* unlocks the rq */
5523 * the context switch might have flipped the stack from under
5524 * us, hence refresh the local variables.
5526 cpu = smp_processor_id();
5527 rq = cpu_rq(cpu);
5528 } else
5529 raw_spin_unlock_irq(&rq->lock);
5531 post_schedule(rq);
5533 if (unlikely(reacquire_kernel_lock(current) < 0)) {
5534 prev = rq->curr;
5535 switch_count = &prev->nivcsw;
5536 goto need_resched_nonpreemptible;
5539 preempt_enable_no_resched();
5540 if (need_resched())
5541 goto need_resched;
5543 EXPORT_SYMBOL(schedule);
5545 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5547 * Look out! "owner" is an entirely speculative pointer
5548 * access and not reliable.
5550 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5552 unsigned int cpu;
5553 struct rq *rq;
5555 if (!sched_feat(OWNER_SPIN))
5556 return 0;
5558 #ifdef CONFIG_DEBUG_PAGEALLOC
5560 * Need to access the cpu field knowing that
5561 * DEBUG_PAGEALLOC could have unmapped it if
5562 * the mutex owner just released it and exited.
5564 if (probe_kernel_address(&owner->cpu, cpu))
5565 goto out;
5566 #else
5567 cpu = owner->cpu;
5568 #endif
5571 * Even if the access succeeded (likely case),
5572 * the cpu field may no longer be valid.
5574 if (cpu >= nr_cpumask_bits)
5575 goto out;
5578 * We need to validate that we can do a
5579 * get_cpu() and that we have the percpu area.
5581 if (!cpu_online(cpu))
5582 goto out;
5584 rq = cpu_rq(cpu);
5586 for (;;) {
5588 * Owner changed, break to re-assess state.
5590 if (lock->owner != owner)
5591 break;
5594 * Is that owner really running on that cpu?
5596 if (task_thread_info(rq->curr) != owner || need_resched())
5597 return 0;
5599 cpu_relax();
5601 out:
5602 return 1;
5604 #endif
5606 #ifdef CONFIG_PREEMPT
5608 * this is the entry point to schedule() from in-kernel preemption
5609 * off of preempt_enable. Kernel preemptions off return from interrupt
5610 * occur there and call schedule directly.
5612 asmlinkage void __sched preempt_schedule(void)
5614 struct thread_info *ti = current_thread_info();
5617 * If there is a non-zero preempt_count or interrupts are disabled,
5618 * we do not want to preempt the current task. Just return..
5620 if (likely(ti->preempt_count || irqs_disabled()))
5621 return;
5623 do {
5624 add_preempt_count(PREEMPT_ACTIVE);
5625 schedule();
5626 sub_preempt_count(PREEMPT_ACTIVE);
5629 * Check again in case we missed a preemption opportunity
5630 * between schedule and now.
5632 barrier();
5633 } while (need_resched());
5635 EXPORT_SYMBOL(preempt_schedule);
5638 * this is the entry point to schedule() from kernel preemption
5639 * off of irq context.
5640 * Note, that this is called and return with irqs disabled. This will
5641 * protect us against recursive calling from irq.
5643 asmlinkage void __sched preempt_schedule_irq(void)
5645 struct thread_info *ti = current_thread_info();
5647 /* Catch callers which need to be fixed */
5648 BUG_ON(ti->preempt_count || !irqs_disabled());
5650 do {
5651 add_preempt_count(PREEMPT_ACTIVE);
5652 local_irq_enable();
5653 schedule();
5654 local_irq_disable();
5655 sub_preempt_count(PREEMPT_ACTIVE);
5658 * Check again in case we missed a preemption opportunity
5659 * between schedule and now.
5661 barrier();
5662 } while (need_resched());
5665 #endif /* CONFIG_PREEMPT */
5667 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5668 void *key)
5670 return try_to_wake_up(curr->private, mode, wake_flags);
5672 EXPORT_SYMBOL(default_wake_function);
5675 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5676 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5677 * number) then we wake all the non-exclusive tasks and one exclusive task.
5679 * There are circumstances in which we can try to wake a task which has already
5680 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5681 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5683 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5684 int nr_exclusive, int wake_flags, void *key)
5686 wait_queue_t *curr, *next;
5688 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5689 unsigned flags = curr->flags;
5691 if (curr->func(curr, mode, wake_flags, key) &&
5692 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5693 break;
5698 * __wake_up - wake up threads blocked on a waitqueue.
5699 * @q: the waitqueue
5700 * @mode: which threads
5701 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5702 * @key: is directly passed to the wakeup function
5704 * It may be assumed that this function implies a write memory barrier before
5705 * changing the task state if and only if any tasks are woken up.
5707 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5708 int nr_exclusive, void *key)
5710 unsigned long flags;
5712 spin_lock_irqsave(&q->lock, flags);
5713 __wake_up_common(q, mode, nr_exclusive, 0, key);
5714 spin_unlock_irqrestore(&q->lock, flags);
5716 EXPORT_SYMBOL(__wake_up);
5719 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5721 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5723 __wake_up_common(q, mode, 1, 0, NULL);
5726 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5728 __wake_up_common(q, mode, 1, 0, key);
5732 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5733 * @q: the waitqueue
5734 * @mode: which threads
5735 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5736 * @key: opaque value to be passed to wakeup targets
5738 * The sync wakeup differs that the waker knows that it will schedule
5739 * away soon, so while the target thread will be woken up, it will not
5740 * be migrated to another CPU - ie. the two threads are 'synchronized'
5741 * with each other. This can prevent needless bouncing between CPUs.
5743 * On UP it can prevent extra preemption.
5745 * It may be assumed that this function implies a write memory barrier before
5746 * changing the task state if and only if any tasks are woken up.
5748 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5749 int nr_exclusive, void *key)
5751 unsigned long flags;
5752 int wake_flags = WF_SYNC;
5754 if (unlikely(!q))
5755 return;
5757 if (unlikely(!nr_exclusive))
5758 wake_flags = 0;
5760 spin_lock_irqsave(&q->lock, flags);
5761 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5762 spin_unlock_irqrestore(&q->lock, flags);
5764 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5767 * __wake_up_sync - see __wake_up_sync_key()
5769 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5771 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5773 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5776 * complete: - signals a single thread waiting on this completion
5777 * @x: holds the state of this particular completion
5779 * This will wake up a single thread waiting on this completion. Threads will be
5780 * awakened in the same order in which they were queued.
5782 * See also complete_all(), wait_for_completion() and related routines.
5784 * It may be assumed that this function implies a write memory barrier before
5785 * changing the task state if and only if any tasks are woken up.
5787 void complete(struct completion *x)
5789 unsigned long flags;
5791 spin_lock_irqsave(&x->wait.lock, flags);
5792 x->done++;
5793 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5794 spin_unlock_irqrestore(&x->wait.lock, flags);
5796 EXPORT_SYMBOL(complete);
5799 * complete_all: - signals all threads waiting on this completion
5800 * @x: holds the state of this particular completion
5802 * This will wake up all threads waiting on this particular completion event.
5804 * It may be assumed that this function implies a write memory barrier before
5805 * changing the task state if and only if any tasks are woken up.
5807 void complete_all(struct completion *x)
5809 unsigned long flags;
5811 spin_lock_irqsave(&x->wait.lock, flags);
5812 x->done += UINT_MAX/2;
5813 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5814 spin_unlock_irqrestore(&x->wait.lock, flags);
5816 EXPORT_SYMBOL(complete_all);
5818 static inline long __sched
5819 do_wait_for_common(struct completion *x, long timeout, int state)
5821 if (!x->done) {
5822 DECLARE_WAITQUEUE(wait, current);
5824 wait.flags |= WQ_FLAG_EXCLUSIVE;
5825 __add_wait_queue_tail(&x->wait, &wait);
5826 do {
5827 if (signal_pending_state(state, current)) {
5828 timeout = -ERESTARTSYS;
5829 break;
5831 __set_current_state(state);
5832 spin_unlock_irq(&x->wait.lock);
5833 timeout = schedule_timeout(timeout);
5834 spin_lock_irq(&x->wait.lock);
5835 } while (!x->done && timeout);
5836 __remove_wait_queue(&x->wait, &wait);
5837 if (!x->done)
5838 return timeout;
5840 x->done--;
5841 return timeout ?: 1;
5844 static long __sched
5845 wait_for_common(struct completion *x, long timeout, int state)
5847 might_sleep();
5849 spin_lock_irq(&x->wait.lock);
5850 timeout = do_wait_for_common(x, timeout, state);
5851 spin_unlock_irq(&x->wait.lock);
5852 return timeout;
5856 * wait_for_completion: - waits for completion of a task
5857 * @x: holds the state of this particular completion
5859 * This waits to be signaled for completion of a specific task. It is NOT
5860 * interruptible and there is no timeout.
5862 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5863 * and interrupt capability. Also see complete().
5865 void __sched wait_for_completion(struct completion *x)
5867 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5869 EXPORT_SYMBOL(wait_for_completion);
5872 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5873 * @x: holds the state of this particular completion
5874 * @timeout: timeout value in jiffies
5876 * This waits for either a completion of a specific task to be signaled or for a
5877 * specified timeout to expire. The timeout is in jiffies. It is not
5878 * interruptible.
5880 unsigned long __sched
5881 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5883 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5885 EXPORT_SYMBOL(wait_for_completion_timeout);
5888 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5889 * @x: holds the state of this particular completion
5891 * This waits for completion of a specific task to be signaled. It is
5892 * interruptible.
5894 int __sched wait_for_completion_interruptible(struct completion *x)
5896 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5897 if (t == -ERESTARTSYS)
5898 return t;
5899 return 0;
5901 EXPORT_SYMBOL(wait_for_completion_interruptible);
5904 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5905 * @x: holds the state of this particular completion
5906 * @timeout: timeout value in jiffies
5908 * This waits for either a completion of a specific task to be signaled or for a
5909 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5911 unsigned long __sched
5912 wait_for_completion_interruptible_timeout(struct completion *x,
5913 unsigned long timeout)
5915 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5917 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5920 * wait_for_completion_killable: - waits for completion of a task (killable)
5921 * @x: holds the state of this particular completion
5923 * This waits to be signaled for completion of a specific task. It can be
5924 * interrupted by a kill signal.
5926 int __sched wait_for_completion_killable(struct completion *x)
5928 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5929 if (t == -ERESTARTSYS)
5930 return t;
5931 return 0;
5933 EXPORT_SYMBOL(wait_for_completion_killable);
5936 * try_wait_for_completion - try to decrement a completion without blocking
5937 * @x: completion structure
5939 * Returns: 0 if a decrement cannot be done without blocking
5940 * 1 if a decrement succeeded.
5942 * If a completion is being used as a counting completion,
5943 * attempt to decrement the counter without blocking. This
5944 * enables us to avoid waiting if the resource the completion
5945 * is protecting is not available.
5947 bool try_wait_for_completion(struct completion *x)
5949 unsigned long flags;
5950 int ret = 1;
5952 spin_lock_irqsave(&x->wait.lock, flags);
5953 if (!x->done)
5954 ret = 0;
5955 else
5956 x->done--;
5957 spin_unlock_irqrestore(&x->wait.lock, flags);
5958 return ret;
5960 EXPORT_SYMBOL(try_wait_for_completion);
5963 * completion_done - Test to see if a completion has any waiters
5964 * @x: completion structure
5966 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5967 * 1 if there are no waiters.
5970 bool completion_done(struct completion *x)
5972 unsigned long flags;
5973 int ret = 1;
5975 spin_lock_irqsave(&x->wait.lock, flags);
5976 if (!x->done)
5977 ret = 0;
5978 spin_unlock_irqrestore(&x->wait.lock, flags);
5979 return ret;
5981 EXPORT_SYMBOL(completion_done);
5983 static long __sched
5984 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5986 unsigned long flags;
5987 wait_queue_t wait;
5989 init_waitqueue_entry(&wait, current);
5991 __set_current_state(state);
5993 spin_lock_irqsave(&q->lock, flags);
5994 __add_wait_queue(q, &wait);
5995 spin_unlock(&q->lock);
5996 timeout = schedule_timeout(timeout);
5997 spin_lock_irq(&q->lock);
5998 __remove_wait_queue(q, &wait);
5999 spin_unlock_irqrestore(&q->lock, flags);
6001 return timeout;
6004 void __sched interruptible_sleep_on(wait_queue_head_t *q)
6006 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6008 EXPORT_SYMBOL(interruptible_sleep_on);
6010 long __sched
6011 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
6013 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
6015 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
6017 void __sched sleep_on(wait_queue_head_t *q)
6019 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6021 EXPORT_SYMBOL(sleep_on);
6023 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
6025 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
6027 EXPORT_SYMBOL(sleep_on_timeout);
6029 #ifdef CONFIG_RT_MUTEXES
6032 * rt_mutex_setprio - set the current priority of a task
6033 * @p: task
6034 * @prio: prio value (kernel-internal form)
6036 * This function changes the 'effective' priority of a task. It does
6037 * not touch ->normal_prio like __setscheduler().
6039 * Used by the rt_mutex code to implement priority inheritance logic.
6041 void rt_mutex_setprio(struct task_struct *p, int prio)
6043 unsigned long flags;
6044 int oldprio, on_rq, running;
6045 struct rq *rq;
6046 const struct sched_class *prev_class = p->sched_class;
6048 BUG_ON(prio < 0 || prio > MAX_PRIO);
6050 rq = task_rq_lock(p, &flags);
6051 update_rq_clock(rq);
6053 oldprio = p->prio;
6054 on_rq = p->se.on_rq;
6055 running = task_current(rq, p);
6056 if (on_rq)
6057 dequeue_task(rq, p, 0);
6058 if (running)
6059 p->sched_class->put_prev_task(rq, p);
6061 if (rt_prio(prio))
6062 p->sched_class = &rt_sched_class;
6063 else
6064 p->sched_class = &fair_sched_class;
6066 p->prio = prio;
6068 if (running)
6069 p->sched_class->set_curr_task(rq);
6070 if (on_rq) {
6071 enqueue_task(rq, p, 0);
6073 check_class_changed(rq, p, prev_class, oldprio, running);
6075 task_rq_unlock(rq, &flags);
6078 #endif
6080 void set_user_nice(struct task_struct *p, long nice)
6082 int old_prio, delta, on_rq;
6083 unsigned long flags;
6084 struct rq *rq;
6086 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6087 return;
6089 * We have to be careful, if called from sys_setpriority(),
6090 * the task might be in the middle of scheduling on another CPU.
6092 rq = task_rq_lock(p, &flags);
6093 update_rq_clock(rq);
6095 * The RT priorities are set via sched_setscheduler(), but we still
6096 * allow the 'normal' nice value to be set - but as expected
6097 * it wont have any effect on scheduling until the task is
6098 * SCHED_FIFO/SCHED_RR:
6100 if (task_has_rt_policy(p)) {
6101 p->static_prio = NICE_TO_PRIO(nice);
6102 goto out_unlock;
6104 on_rq = p->se.on_rq;
6105 if (on_rq)
6106 dequeue_task(rq, p, 0);
6108 p->static_prio = NICE_TO_PRIO(nice);
6109 set_load_weight(p);
6110 old_prio = p->prio;
6111 p->prio = effective_prio(p);
6112 delta = p->prio - old_prio;
6114 if (on_rq) {
6115 enqueue_task(rq, p, 0);
6117 * If the task increased its priority or is running and
6118 * lowered its priority, then reschedule its CPU:
6120 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6121 resched_task(rq->curr);
6123 out_unlock:
6124 task_rq_unlock(rq, &flags);
6126 EXPORT_SYMBOL(set_user_nice);
6129 * can_nice - check if a task can reduce its nice value
6130 * @p: task
6131 * @nice: nice value
6133 int can_nice(const struct task_struct *p, const int nice)
6135 /* convert nice value [19,-20] to rlimit style value [1,40] */
6136 int nice_rlim = 20 - nice;
6138 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6139 capable(CAP_SYS_NICE));
6142 #ifdef __ARCH_WANT_SYS_NICE
6145 * sys_nice - change the priority of the current process.
6146 * @increment: priority increment
6148 * sys_setpriority is a more generic, but much slower function that
6149 * does similar things.
6151 SYSCALL_DEFINE1(nice, int, increment)
6153 long nice, retval;
6156 * Setpriority might change our priority at the same moment.
6157 * We don't have to worry. Conceptually one call occurs first
6158 * and we have a single winner.
6160 if (increment < -40)
6161 increment = -40;
6162 if (increment > 40)
6163 increment = 40;
6165 nice = TASK_NICE(current) + increment;
6166 if (nice < -20)
6167 nice = -20;
6168 if (nice > 19)
6169 nice = 19;
6171 if (increment < 0 && !can_nice(current, nice))
6172 return -EPERM;
6174 retval = security_task_setnice(current, nice);
6175 if (retval)
6176 return retval;
6178 set_user_nice(current, nice);
6179 return 0;
6182 #endif
6185 * task_prio - return the priority value of a given task.
6186 * @p: the task in question.
6188 * This is the priority value as seen by users in /proc.
6189 * RT tasks are offset by -200. Normal tasks are centered
6190 * around 0, value goes from -16 to +15.
6192 int task_prio(const struct task_struct *p)
6194 return p->prio - MAX_RT_PRIO;
6198 * task_nice - return the nice value of a given task.
6199 * @p: the task in question.
6201 int task_nice(const struct task_struct *p)
6203 return TASK_NICE(p);
6205 EXPORT_SYMBOL(task_nice);
6208 * idle_cpu - is a given cpu idle currently?
6209 * @cpu: the processor in question.
6211 int idle_cpu(int cpu)
6213 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6217 * idle_task - return the idle task for a given cpu.
6218 * @cpu: the processor in question.
6220 struct task_struct *idle_task(int cpu)
6222 return cpu_rq(cpu)->idle;
6226 * find_process_by_pid - find a process with a matching PID value.
6227 * @pid: the pid in question.
6229 static struct task_struct *find_process_by_pid(pid_t pid)
6231 return pid ? find_task_by_vpid(pid) : current;
6234 /* Actually do priority change: must hold rq lock. */
6235 static void
6236 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6238 BUG_ON(p->se.on_rq);
6240 p->policy = policy;
6241 p->rt_priority = prio;
6242 p->normal_prio = normal_prio(p);
6243 /* we are holding p->pi_lock already */
6244 p->prio = rt_mutex_getprio(p);
6245 if (rt_prio(p->prio))
6246 p->sched_class = &rt_sched_class;
6247 else
6248 p->sched_class = &fair_sched_class;
6249 set_load_weight(p);
6253 * check the target process has a UID that matches the current process's
6255 static bool check_same_owner(struct task_struct *p)
6257 const struct cred *cred = current_cred(), *pcred;
6258 bool match;
6260 rcu_read_lock();
6261 pcred = __task_cred(p);
6262 match = (cred->euid == pcred->euid ||
6263 cred->euid == pcred->uid);
6264 rcu_read_unlock();
6265 return match;
6268 static int __sched_setscheduler(struct task_struct *p, int policy,
6269 struct sched_param *param, bool user)
6271 int retval, oldprio, oldpolicy = -1, on_rq, running;
6272 unsigned long flags;
6273 const struct sched_class *prev_class = p->sched_class;
6274 struct rq *rq;
6275 int reset_on_fork;
6277 /* may grab non-irq protected spin_locks */
6278 BUG_ON(in_interrupt());
6279 recheck:
6280 /* double check policy once rq lock held */
6281 if (policy < 0) {
6282 reset_on_fork = p->sched_reset_on_fork;
6283 policy = oldpolicy = p->policy;
6284 } else {
6285 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6286 policy &= ~SCHED_RESET_ON_FORK;
6288 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6289 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6290 policy != SCHED_IDLE)
6291 return -EINVAL;
6295 * Valid priorities for SCHED_FIFO and SCHED_RR are
6296 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6297 * SCHED_BATCH and SCHED_IDLE is 0.
6299 if (param->sched_priority < 0 ||
6300 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6301 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6302 return -EINVAL;
6303 if (rt_policy(policy) != (param->sched_priority != 0))
6304 return -EINVAL;
6307 * Allow unprivileged RT tasks to decrease priority:
6309 if (user && !capable(CAP_SYS_NICE)) {
6310 if (rt_policy(policy)) {
6311 unsigned long rlim_rtprio;
6313 if (!lock_task_sighand(p, &flags))
6314 return -ESRCH;
6315 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6316 unlock_task_sighand(p, &flags);
6318 /* can't set/change the rt policy */
6319 if (policy != p->policy && !rlim_rtprio)
6320 return -EPERM;
6322 /* can't increase priority */
6323 if (param->sched_priority > p->rt_priority &&
6324 param->sched_priority > rlim_rtprio)
6325 return -EPERM;
6328 * Like positive nice levels, dont allow tasks to
6329 * move out of SCHED_IDLE either:
6331 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6332 return -EPERM;
6334 /* can't change other user's priorities */
6335 if (!check_same_owner(p))
6336 return -EPERM;
6338 /* Normal users shall not reset the sched_reset_on_fork flag */
6339 if (p->sched_reset_on_fork && !reset_on_fork)
6340 return -EPERM;
6343 if (user) {
6344 #ifdef CONFIG_RT_GROUP_SCHED
6346 * Do not allow realtime tasks into groups that have no runtime
6347 * assigned.
6349 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6350 task_group(p)->rt_bandwidth.rt_runtime == 0)
6351 return -EPERM;
6352 #endif
6354 retval = security_task_setscheduler(p, policy, param);
6355 if (retval)
6356 return retval;
6360 * make sure no PI-waiters arrive (or leave) while we are
6361 * changing the priority of the task:
6363 raw_spin_lock_irqsave(&p->pi_lock, flags);
6365 * To be able to change p->policy safely, the apropriate
6366 * runqueue lock must be held.
6368 rq = __task_rq_lock(p);
6369 /* recheck policy now with rq lock held */
6370 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6371 policy = oldpolicy = -1;
6372 __task_rq_unlock(rq);
6373 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6374 goto recheck;
6376 update_rq_clock(rq);
6377 on_rq = p->se.on_rq;
6378 running = task_current(rq, p);
6379 if (on_rq)
6380 deactivate_task(rq, p, 0);
6381 if (running)
6382 p->sched_class->put_prev_task(rq, p);
6384 p->sched_reset_on_fork = reset_on_fork;
6386 oldprio = p->prio;
6387 __setscheduler(rq, p, policy, param->sched_priority);
6389 if (running)
6390 p->sched_class->set_curr_task(rq);
6391 if (on_rq) {
6392 activate_task(rq, p, 0);
6394 check_class_changed(rq, p, prev_class, oldprio, running);
6396 __task_rq_unlock(rq);
6397 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6399 rt_mutex_adjust_pi(p);
6401 return 0;
6405 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6406 * @p: the task in question.
6407 * @policy: new policy.
6408 * @param: structure containing the new RT priority.
6410 * NOTE that the task may be already dead.
6412 int sched_setscheduler(struct task_struct *p, int policy,
6413 struct sched_param *param)
6415 return __sched_setscheduler(p, policy, param, true);
6417 EXPORT_SYMBOL_GPL(sched_setscheduler);
6420 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6421 * @p: the task in question.
6422 * @policy: new policy.
6423 * @param: structure containing the new RT priority.
6425 * Just like sched_setscheduler, only don't bother checking if the
6426 * current context has permission. For example, this is needed in
6427 * stop_machine(): we create temporary high priority worker threads,
6428 * but our caller might not have that capability.
6430 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6431 struct sched_param *param)
6433 return __sched_setscheduler(p, policy, param, false);
6436 static int
6437 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6439 struct sched_param lparam;
6440 struct task_struct *p;
6441 int retval;
6443 if (!param || pid < 0)
6444 return -EINVAL;
6445 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6446 return -EFAULT;
6448 rcu_read_lock();
6449 retval = -ESRCH;
6450 p = find_process_by_pid(pid);
6451 if (p != NULL)
6452 retval = sched_setscheduler(p, policy, &lparam);
6453 rcu_read_unlock();
6455 return retval;
6459 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6460 * @pid: the pid in question.
6461 * @policy: new policy.
6462 * @param: structure containing the new RT priority.
6464 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6465 struct sched_param __user *, param)
6467 /* negative values for policy are not valid */
6468 if (policy < 0)
6469 return -EINVAL;
6471 return do_sched_setscheduler(pid, policy, param);
6475 * sys_sched_setparam - set/change the RT priority of a thread
6476 * @pid: the pid in question.
6477 * @param: structure containing the new RT priority.
6479 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6481 return do_sched_setscheduler(pid, -1, param);
6485 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6486 * @pid: the pid in question.
6488 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6490 struct task_struct *p;
6491 int retval;
6493 if (pid < 0)
6494 return -EINVAL;
6496 retval = -ESRCH;
6497 rcu_read_lock();
6498 p = find_process_by_pid(pid);
6499 if (p) {
6500 retval = security_task_getscheduler(p);
6501 if (!retval)
6502 retval = p->policy
6503 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6505 rcu_read_unlock();
6506 return retval;
6510 * sys_sched_getparam - get the RT priority of a thread
6511 * @pid: the pid in question.
6512 * @param: structure containing the RT priority.
6514 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6516 struct sched_param lp;
6517 struct task_struct *p;
6518 int retval;
6520 if (!param || pid < 0)
6521 return -EINVAL;
6523 rcu_read_lock();
6524 p = find_process_by_pid(pid);
6525 retval = -ESRCH;
6526 if (!p)
6527 goto out_unlock;
6529 retval = security_task_getscheduler(p);
6530 if (retval)
6531 goto out_unlock;
6533 lp.sched_priority = p->rt_priority;
6534 rcu_read_unlock();
6537 * This one might sleep, we cannot do it with a spinlock held ...
6539 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6541 return retval;
6543 out_unlock:
6544 rcu_read_unlock();
6545 return retval;
6548 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6550 cpumask_var_t cpus_allowed, new_mask;
6551 struct task_struct *p;
6552 int retval;
6554 get_online_cpus();
6555 rcu_read_lock();
6557 p = find_process_by_pid(pid);
6558 if (!p) {
6559 rcu_read_unlock();
6560 put_online_cpus();
6561 return -ESRCH;
6564 /* Prevent p going away */
6565 get_task_struct(p);
6566 rcu_read_unlock();
6568 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6569 retval = -ENOMEM;
6570 goto out_put_task;
6572 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6573 retval = -ENOMEM;
6574 goto out_free_cpus_allowed;
6576 retval = -EPERM;
6577 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6578 goto out_unlock;
6580 retval = security_task_setscheduler(p, 0, NULL);
6581 if (retval)
6582 goto out_unlock;
6584 cpuset_cpus_allowed(p, cpus_allowed);
6585 cpumask_and(new_mask, in_mask, cpus_allowed);
6586 again:
6587 retval = set_cpus_allowed_ptr(p, new_mask);
6589 if (!retval) {
6590 cpuset_cpus_allowed(p, cpus_allowed);
6591 if (!cpumask_subset(new_mask, cpus_allowed)) {
6593 * We must have raced with a concurrent cpuset
6594 * update. Just reset the cpus_allowed to the
6595 * cpuset's cpus_allowed
6597 cpumask_copy(new_mask, cpus_allowed);
6598 goto again;
6601 out_unlock:
6602 free_cpumask_var(new_mask);
6603 out_free_cpus_allowed:
6604 free_cpumask_var(cpus_allowed);
6605 out_put_task:
6606 put_task_struct(p);
6607 put_online_cpus();
6608 return retval;
6611 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6612 struct cpumask *new_mask)
6614 if (len < cpumask_size())
6615 cpumask_clear(new_mask);
6616 else if (len > cpumask_size())
6617 len = cpumask_size();
6619 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6623 * sys_sched_setaffinity - set the cpu affinity of a process
6624 * @pid: pid of the process
6625 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6626 * @user_mask_ptr: user-space pointer to the new cpu mask
6628 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6629 unsigned long __user *, user_mask_ptr)
6631 cpumask_var_t new_mask;
6632 int retval;
6634 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6635 return -ENOMEM;
6637 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6638 if (retval == 0)
6639 retval = sched_setaffinity(pid, new_mask);
6640 free_cpumask_var(new_mask);
6641 return retval;
6644 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6646 struct task_struct *p;
6647 unsigned long flags;
6648 struct rq *rq;
6649 int retval;
6651 get_online_cpus();
6652 rcu_read_lock();
6654 retval = -ESRCH;
6655 p = find_process_by_pid(pid);
6656 if (!p)
6657 goto out_unlock;
6659 retval = security_task_getscheduler(p);
6660 if (retval)
6661 goto out_unlock;
6663 rq = task_rq_lock(p, &flags);
6664 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6665 task_rq_unlock(rq, &flags);
6667 out_unlock:
6668 rcu_read_unlock();
6669 put_online_cpus();
6671 return retval;
6675 * sys_sched_getaffinity - get the cpu affinity of a process
6676 * @pid: pid of the process
6677 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6678 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6680 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6681 unsigned long __user *, user_mask_ptr)
6683 int ret;
6684 cpumask_var_t mask;
6686 if (len < cpumask_size())
6687 return -EINVAL;
6689 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6690 return -ENOMEM;
6692 ret = sched_getaffinity(pid, mask);
6693 if (ret == 0) {
6694 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6695 ret = -EFAULT;
6696 else
6697 ret = cpumask_size();
6699 free_cpumask_var(mask);
6701 return ret;
6705 * sys_sched_yield - yield the current processor to other threads.
6707 * This function yields the current CPU to other tasks. If there are no
6708 * other threads running on this CPU then this function will return.
6710 SYSCALL_DEFINE0(sched_yield)
6712 struct rq *rq = this_rq_lock();
6714 schedstat_inc(rq, yld_count);
6715 current->sched_class->yield_task(rq);
6718 * Since we are going to call schedule() anyway, there's
6719 * no need to preempt or enable interrupts:
6721 __release(rq->lock);
6722 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6723 do_raw_spin_unlock(&rq->lock);
6724 preempt_enable_no_resched();
6726 schedule();
6728 return 0;
6731 static inline int should_resched(void)
6733 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6736 static void __cond_resched(void)
6738 add_preempt_count(PREEMPT_ACTIVE);
6739 schedule();
6740 sub_preempt_count(PREEMPT_ACTIVE);
6743 int __sched _cond_resched(void)
6745 if (should_resched()) {
6746 __cond_resched();
6747 return 1;
6749 return 0;
6751 EXPORT_SYMBOL(_cond_resched);
6754 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6755 * call schedule, and on return reacquire the lock.
6757 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6758 * operations here to prevent schedule() from being called twice (once via
6759 * spin_unlock(), once by hand).
6761 int __cond_resched_lock(spinlock_t *lock)
6763 int resched = should_resched();
6764 int ret = 0;
6766 lockdep_assert_held(lock);
6768 if (spin_needbreak(lock) || resched) {
6769 spin_unlock(lock);
6770 if (resched)
6771 __cond_resched();
6772 else
6773 cpu_relax();
6774 ret = 1;
6775 spin_lock(lock);
6777 return ret;
6779 EXPORT_SYMBOL(__cond_resched_lock);
6781 int __sched __cond_resched_softirq(void)
6783 BUG_ON(!in_softirq());
6785 if (should_resched()) {
6786 local_bh_enable();
6787 __cond_resched();
6788 local_bh_disable();
6789 return 1;
6791 return 0;
6793 EXPORT_SYMBOL(__cond_resched_softirq);
6796 * yield - yield the current processor to other threads.
6798 * This is a shortcut for kernel-space yielding - it marks the
6799 * thread runnable and calls sys_sched_yield().
6801 void __sched yield(void)
6803 set_current_state(TASK_RUNNING);
6804 sys_sched_yield();
6806 EXPORT_SYMBOL(yield);
6809 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6810 * that process accounting knows that this is a task in IO wait state.
6812 void __sched io_schedule(void)
6814 struct rq *rq = raw_rq();
6816 delayacct_blkio_start();
6817 atomic_inc(&rq->nr_iowait);
6818 current->in_iowait = 1;
6819 schedule();
6820 current->in_iowait = 0;
6821 atomic_dec(&rq->nr_iowait);
6822 delayacct_blkio_end();
6824 EXPORT_SYMBOL(io_schedule);
6826 long __sched io_schedule_timeout(long timeout)
6828 struct rq *rq = raw_rq();
6829 long ret;
6831 delayacct_blkio_start();
6832 atomic_inc(&rq->nr_iowait);
6833 current->in_iowait = 1;
6834 ret = schedule_timeout(timeout);
6835 current->in_iowait = 0;
6836 atomic_dec(&rq->nr_iowait);
6837 delayacct_blkio_end();
6838 return ret;
6842 * sys_sched_get_priority_max - return maximum RT priority.
6843 * @policy: scheduling class.
6845 * this syscall returns the maximum rt_priority that can be used
6846 * by a given scheduling class.
6848 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6850 int ret = -EINVAL;
6852 switch (policy) {
6853 case SCHED_FIFO:
6854 case SCHED_RR:
6855 ret = MAX_USER_RT_PRIO-1;
6856 break;
6857 case SCHED_NORMAL:
6858 case SCHED_BATCH:
6859 case SCHED_IDLE:
6860 ret = 0;
6861 break;
6863 return ret;
6867 * sys_sched_get_priority_min - return minimum RT priority.
6868 * @policy: scheduling class.
6870 * this syscall returns the minimum rt_priority that can be used
6871 * by a given scheduling class.
6873 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6875 int ret = -EINVAL;
6877 switch (policy) {
6878 case SCHED_FIFO:
6879 case SCHED_RR:
6880 ret = 1;
6881 break;
6882 case SCHED_NORMAL:
6883 case SCHED_BATCH:
6884 case SCHED_IDLE:
6885 ret = 0;
6887 return ret;
6891 * sys_sched_rr_get_interval - return the default timeslice of a process.
6892 * @pid: pid of the process.
6893 * @interval: userspace pointer to the timeslice value.
6895 * this syscall writes the default timeslice value of a given process
6896 * into the user-space timespec buffer. A value of '0' means infinity.
6898 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6899 struct timespec __user *, interval)
6901 struct task_struct *p;
6902 unsigned int time_slice;
6903 unsigned long flags;
6904 struct rq *rq;
6905 int retval;
6906 struct timespec t;
6908 if (pid < 0)
6909 return -EINVAL;
6911 retval = -ESRCH;
6912 rcu_read_lock();
6913 p = find_process_by_pid(pid);
6914 if (!p)
6915 goto out_unlock;
6917 retval = security_task_getscheduler(p);
6918 if (retval)
6919 goto out_unlock;
6921 rq = task_rq_lock(p, &flags);
6922 time_slice = p->sched_class->get_rr_interval(rq, p);
6923 task_rq_unlock(rq, &flags);
6925 rcu_read_unlock();
6926 jiffies_to_timespec(time_slice, &t);
6927 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6928 return retval;
6930 out_unlock:
6931 rcu_read_unlock();
6932 return retval;
6935 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6937 void sched_show_task(struct task_struct *p)
6939 unsigned long free = 0;
6940 unsigned state;
6942 state = p->state ? __ffs(p->state) + 1 : 0;
6943 printk(KERN_INFO "%-13.13s %c", p->comm,
6944 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6945 #if BITS_PER_LONG == 32
6946 if (state == TASK_RUNNING)
6947 printk(KERN_CONT " running ");
6948 else
6949 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6950 #else
6951 if (state == TASK_RUNNING)
6952 printk(KERN_CONT " running task ");
6953 else
6954 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6955 #endif
6956 #ifdef CONFIG_DEBUG_STACK_USAGE
6957 free = stack_not_used(p);
6958 #endif
6959 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6960 task_pid_nr(p), task_pid_nr(p->real_parent),
6961 (unsigned long)task_thread_info(p)->flags);
6963 show_stack(p, NULL);
6966 void show_state_filter(unsigned long state_filter)
6968 struct task_struct *g, *p;
6970 #if BITS_PER_LONG == 32
6971 printk(KERN_INFO
6972 " task PC stack pid father\n");
6973 #else
6974 printk(KERN_INFO
6975 " task PC stack pid father\n");
6976 #endif
6977 read_lock(&tasklist_lock);
6978 do_each_thread(g, p) {
6980 * reset the NMI-timeout, listing all files on a slow
6981 * console might take alot of time:
6983 touch_nmi_watchdog();
6984 if (!state_filter || (p->state & state_filter))
6985 sched_show_task(p);
6986 } while_each_thread(g, p);
6988 touch_all_softlockup_watchdogs();
6990 #ifdef CONFIG_SCHED_DEBUG
6991 sysrq_sched_debug_show();
6992 #endif
6993 read_unlock(&tasklist_lock);
6995 * Only show locks if all tasks are dumped:
6997 if (!state_filter)
6998 debug_show_all_locks();
7001 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
7003 idle->sched_class = &idle_sched_class;
7007 * init_idle - set up an idle thread for a given CPU
7008 * @idle: task in question
7009 * @cpu: cpu the idle task belongs to
7011 * NOTE: this function does not set the idle thread's NEED_RESCHED
7012 * flag, to make booting more robust.
7014 void __cpuinit init_idle(struct task_struct *idle, int cpu)
7016 struct rq *rq = cpu_rq(cpu);
7017 unsigned long flags;
7019 raw_spin_lock_irqsave(&rq->lock, flags);
7021 __sched_fork(idle);
7022 idle->state = TASK_RUNNING;
7023 idle->se.exec_start = sched_clock();
7025 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
7026 __set_task_cpu(idle, cpu);
7028 rq->curr = rq->idle = idle;
7029 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7030 idle->oncpu = 1;
7031 #endif
7032 raw_spin_unlock_irqrestore(&rq->lock, flags);
7034 /* Set the preempt count _outside_ the spinlocks! */
7035 #if defined(CONFIG_PREEMPT)
7036 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7037 #else
7038 task_thread_info(idle)->preempt_count = 0;
7039 #endif
7041 * The idle tasks have their own, simple scheduling class:
7043 idle->sched_class = &idle_sched_class;
7044 ftrace_graph_init_task(idle);
7048 * In a system that switches off the HZ timer nohz_cpu_mask
7049 * indicates which cpus entered this state. This is used
7050 * in the rcu update to wait only for active cpus. For system
7051 * which do not switch off the HZ timer nohz_cpu_mask should
7052 * always be CPU_BITS_NONE.
7054 cpumask_var_t nohz_cpu_mask;
7057 * Increase the granularity value when there are more CPUs,
7058 * because with more CPUs the 'effective latency' as visible
7059 * to users decreases. But the relationship is not linear,
7060 * so pick a second-best guess by going with the log2 of the
7061 * number of CPUs.
7063 * This idea comes from the SD scheduler of Con Kolivas:
7065 static int get_update_sysctl_factor(void)
7067 unsigned int cpus = min_t(int, num_online_cpus(), 8);
7068 unsigned int factor;
7070 switch (sysctl_sched_tunable_scaling) {
7071 case SCHED_TUNABLESCALING_NONE:
7072 factor = 1;
7073 break;
7074 case SCHED_TUNABLESCALING_LINEAR:
7075 factor = cpus;
7076 break;
7077 case SCHED_TUNABLESCALING_LOG:
7078 default:
7079 factor = 1 + ilog2(cpus);
7080 break;
7083 return factor;
7086 static void update_sysctl(void)
7088 unsigned int factor = get_update_sysctl_factor();
7090 #define SET_SYSCTL(name) \
7091 (sysctl_##name = (factor) * normalized_sysctl_##name)
7092 SET_SYSCTL(sched_min_granularity);
7093 SET_SYSCTL(sched_latency);
7094 SET_SYSCTL(sched_wakeup_granularity);
7095 SET_SYSCTL(sched_shares_ratelimit);
7096 #undef SET_SYSCTL
7099 static inline void sched_init_granularity(void)
7101 update_sysctl();
7104 #ifdef CONFIG_SMP
7106 * This is how migration works:
7108 * 1) we queue a struct migration_req structure in the source CPU's
7109 * runqueue and wake up that CPU's migration thread.
7110 * 2) we down() the locked semaphore => thread blocks.
7111 * 3) migration thread wakes up (implicitly it forces the migrated
7112 * thread off the CPU)
7113 * 4) it gets the migration request and checks whether the migrated
7114 * task is still in the wrong runqueue.
7115 * 5) if it's in the wrong runqueue then the migration thread removes
7116 * it and puts it into the right queue.
7117 * 6) migration thread up()s the semaphore.
7118 * 7) we wake up and the migration is done.
7122 * Change a given task's CPU affinity. Migrate the thread to a
7123 * proper CPU and schedule it away if the CPU it's executing on
7124 * is removed from the allowed bitmask.
7126 * NOTE: the caller must have a valid reference to the task, the
7127 * task must not exit() & deallocate itself prematurely. The
7128 * call is not atomic; no spinlocks may be held.
7130 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7132 struct migration_req req;
7133 unsigned long flags;
7134 struct rq *rq;
7135 int ret = 0;
7138 * Since we rely on wake-ups to migrate sleeping tasks, don't change
7139 * the ->cpus_allowed mask from under waking tasks, which would be
7140 * possible when we change rq->lock in ttwu(), so synchronize against
7141 * TASK_WAKING to avoid that.
7143 again:
7144 while (p->state == TASK_WAKING)
7145 cpu_relax();
7147 rq = task_rq_lock(p, &flags);
7149 if (p->state == TASK_WAKING) {
7150 task_rq_unlock(rq, &flags);
7151 goto again;
7154 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7155 ret = -EINVAL;
7156 goto out;
7159 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7160 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7161 ret = -EINVAL;
7162 goto out;
7165 if (p->sched_class->set_cpus_allowed)
7166 p->sched_class->set_cpus_allowed(p, new_mask);
7167 else {
7168 cpumask_copy(&p->cpus_allowed, new_mask);
7169 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7172 /* Can the task run on the task's current CPU? If so, we're done */
7173 if (cpumask_test_cpu(task_cpu(p), new_mask))
7174 goto out;
7176 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7177 /* Need help from migration thread: drop lock and wait. */
7178 struct task_struct *mt = rq->migration_thread;
7180 get_task_struct(mt);
7181 task_rq_unlock(rq, &flags);
7182 wake_up_process(rq->migration_thread);
7183 put_task_struct(mt);
7184 wait_for_completion(&req.done);
7185 tlb_migrate_finish(p->mm);
7186 return 0;
7188 out:
7189 task_rq_unlock(rq, &flags);
7191 return ret;
7193 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7196 * Move (not current) task off this cpu, onto dest cpu. We're doing
7197 * this because either it can't run here any more (set_cpus_allowed()
7198 * away from this CPU, or CPU going down), or because we're
7199 * attempting to rebalance this task on exec (sched_exec).
7201 * So we race with normal scheduler movements, but that's OK, as long
7202 * as the task is no longer on this CPU.
7204 * Returns non-zero if task was successfully migrated.
7206 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7208 struct rq *rq_dest, *rq_src;
7209 int ret = 0;
7211 if (unlikely(!cpu_active(dest_cpu)))
7212 return ret;
7214 rq_src = cpu_rq(src_cpu);
7215 rq_dest = cpu_rq(dest_cpu);
7217 double_rq_lock(rq_src, rq_dest);
7218 /* Already moved. */
7219 if (task_cpu(p) != src_cpu)
7220 goto done;
7221 /* Affinity changed (again). */
7222 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7223 goto fail;
7226 * If we're not on a rq, the next wake-up will ensure we're
7227 * placed properly.
7229 if (p->se.on_rq) {
7230 deactivate_task(rq_src, p, 0);
7231 set_task_cpu(p, dest_cpu);
7232 activate_task(rq_dest, p, 0);
7233 check_preempt_curr(rq_dest, p, 0);
7235 done:
7236 ret = 1;
7237 fail:
7238 double_rq_unlock(rq_src, rq_dest);
7239 return ret;
7242 #define RCU_MIGRATION_IDLE 0
7243 #define RCU_MIGRATION_NEED_QS 1
7244 #define RCU_MIGRATION_GOT_QS 2
7245 #define RCU_MIGRATION_MUST_SYNC 3
7248 * migration_thread - this is a highprio system thread that performs
7249 * thread migration by bumping thread off CPU then 'pushing' onto
7250 * another runqueue.
7252 static int migration_thread(void *data)
7254 int badcpu;
7255 int cpu = (long)data;
7256 struct rq *rq;
7258 rq = cpu_rq(cpu);
7259 BUG_ON(rq->migration_thread != current);
7261 set_current_state(TASK_INTERRUPTIBLE);
7262 while (!kthread_should_stop()) {
7263 struct migration_req *req;
7264 struct list_head *head;
7266 raw_spin_lock_irq(&rq->lock);
7268 if (cpu_is_offline(cpu)) {
7269 raw_spin_unlock_irq(&rq->lock);
7270 break;
7273 if (rq->active_balance) {
7274 active_load_balance(rq, cpu);
7275 rq->active_balance = 0;
7278 head = &rq->migration_queue;
7280 if (list_empty(head)) {
7281 raw_spin_unlock_irq(&rq->lock);
7282 schedule();
7283 set_current_state(TASK_INTERRUPTIBLE);
7284 continue;
7286 req = list_entry(head->next, struct migration_req, list);
7287 list_del_init(head->next);
7289 if (req->task != NULL) {
7290 raw_spin_unlock(&rq->lock);
7291 __migrate_task(req->task, cpu, req->dest_cpu);
7292 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7293 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7294 raw_spin_unlock(&rq->lock);
7295 } else {
7296 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7297 raw_spin_unlock(&rq->lock);
7298 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7300 local_irq_enable();
7302 complete(&req->done);
7304 __set_current_state(TASK_RUNNING);
7306 return 0;
7309 #ifdef CONFIG_HOTPLUG_CPU
7311 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7313 int ret;
7315 local_irq_disable();
7316 ret = __migrate_task(p, src_cpu, dest_cpu);
7317 local_irq_enable();
7318 return ret;
7322 * Figure out where task on dead CPU should go, use force if necessary.
7324 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7326 int dest_cpu;
7328 again:
7329 dest_cpu = select_fallback_rq(dead_cpu, p);
7331 /* It can have affinity changed while we were choosing. */
7332 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7333 goto again;
7337 * While a dead CPU has no uninterruptible tasks queued at this point,
7338 * it might still have a nonzero ->nr_uninterruptible counter, because
7339 * for performance reasons the counter is not stricly tracking tasks to
7340 * their home CPUs. So we just add the counter to another CPU's counter,
7341 * to keep the global sum constant after CPU-down:
7343 static void migrate_nr_uninterruptible(struct rq *rq_src)
7345 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7346 unsigned long flags;
7348 local_irq_save(flags);
7349 double_rq_lock(rq_src, rq_dest);
7350 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7351 rq_src->nr_uninterruptible = 0;
7352 double_rq_unlock(rq_src, rq_dest);
7353 local_irq_restore(flags);
7356 /* Run through task list and migrate tasks from the dead cpu. */
7357 static void migrate_live_tasks(int src_cpu)
7359 struct task_struct *p, *t;
7361 read_lock(&tasklist_lock);
7363 do_each_thread(t, p) {
7364 if (p == current)
7365 continue;
7367 if (task_cpu(p) == src_cpu)
7368 move_task_off_dead_cpu(src_cpu, p);
7369 } while_each_thread(t, p);
7371 read_unlock(&tasklist_lock);
7375 * Schedules idle task to be the next runnable task on current CPU.
7376 * It does so by boosting its priority to highest possible.
7377 * Used by CPU offline code.
7379 void sched_idle_next(void)
7381 int this_cpu = smp_processor_id();
7382 struct rq *rq = cpu_rq(this_cpu);
7383 struct task_struct *p = rq->idle;
7384 unsigned long flags;
7386 /* cpu has to be offline */
7387 BUG_ON(cpu_online(this_cpu));
7390 * Strictly not necessary since rest of the CPUs are stopped by now
7391 * and interrupts disabled on the current cpu.
7393 raw_spin_lock_irqsave(&rq->lock, flags);
7395 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7397 update_rq_clock(rq);
7398 activate_task(rq, p, 0);
7400 raw_spin_unlock_irqrestore(&rq->lock, flags);
7404 * Ensures that the idle task is using init_mm right before its cpu goes
7405 * offline.
7407 void idle_task_exit(void)
7409 struct mm_struct *mm = current->active_mm;
7411 BUG_ON(cpu_online(smp_processor_id()));
7413 if (mm != &init_mm)
7414 switch_mm(mm, &init_mm, current);
7415 mmdrop(mm);
7418 /* called under rq->lock with disabled interrupts */
7419 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7421 struct rq *rq = cpu_rq(dead_cpu);
7423 /* Must be exiting, otherwise would be on tasklist. */
7424 BUG_ON(!p->exit_state);
7426 /* Cannot have done final schedule yet: would have vanished. */
7427 BUG_ON(p->state == TASK_DEAD);
7429 get_task_struct(p);
7432 * Drop lock around migration; if someone else moves it,
7433 * that's OK. No task can be added to this CPU, so iteration is
7434 * fine.
7436 raw_spin_unlock_irq(&rq->lock);
7437 move_task_off_dead_cpu(dead_cpu, p);
7438 raw_spin_lock_irq(&rq->lock);
7440 put_task_struct(p);
7443 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7444 static void migrate_dead_tasks(unsigned int dead_cpu)
7446 struct rq *rq = cpu_rq(dead_cpu);
7447 struct task_struct *next;
7449 for ( ; ; ) {
7450 if (!rq->nr_running)
7451 break;
7452 update_rq_clock(rq);
7453 next = pick_next_task(rq);
7454 if (!next)
7455 break;
7456 next->sched_class->put_prev_task(rq, next);
7457 migrate_dead(dead_cpu, next);
7463 * remove the tasks which were accounted by rq from calc_load_tasks.
7465 static void calc_global_load_remove(struct rq *rq)
7467 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7468 rq->calc_load_active = 0;
7470 #endif /* CONFIG_HOTPLUG_CPU */
7472 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7474 static struct ctl_table sd_ctl_dir[] = {
7476 .procname = "sched_domain",
7477 .mode = 0555,
7482 static struct ctl_table sd_ctl_root[] = {
7484 .procname = "kernel",
7485 .mode = 0555,
7486 .child = sd_ctl_dir,
7491 static struct ctl_table *sd_alloc_ctl_entry(int n)
7493 struct ctl_table *entry =
7494 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7496 return entry;
7499 static void sd_free_ctl_entry(struct ctl_table **tablep)
7501 struct ctl_table *entry;
7504 * In the intermediate directories, both the child directory and
7505 * procname are dynamically allocated and could fail but the mode
7506 * will always be set. In the lowest directory the names are
7507 * static strings and all have proc handlers.
7509 for (entry = *tablep; entry->mode; entry++) {
7510 if (entry->child)
7511 sd_free_ctl_entry(&entry->child);
7512 if (entry->proc_handler == NULL)
7513 kfree(entry->procname);
7516 kfree(*tablep);
7517 *tablep = NULL;
7520 static void
7521 set_table_entry(struct ctl_table *entry,
7522 const char *procname, void *data, int maxlen,
7523 mode_t mode, proc_handler *proc_handler)
7525 entry->procname = procname;
7526 entry->data = data;
7527 entry->maxlen = maxlen;
7528 entry->mode = mode;
7529 entry->proc_handler = proc_handler;
7532 static struct ctl_table *
7533 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7535 struct ctl_table *table = sd_alloc_ctl_entry(13);
7537 if (table == NULL)
7538 return NULL;
7540 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7541 sizeof(long), 0644, proc_doulongvec_minmax);
7542 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7543 sizeof(long), 0644, proc_doulongvec_minmax);
7544 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7545 sizeof(int), 0644, proc_dointvec_minmax);
7546 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7547 sizeof(int), 0644, proc_dointvec_minmax);
7548 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7549 sizeof(int), 0644, proc_dointvec_minmax);
7550 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7551 sizeof(int), 0644, proc_dointvec_minmax);
7552 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7553 sizeof(int), 0644, proc_dointvec_minmax);
7554 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7555 sizeof(int), 0644, proc_dointvec_minmax);
7556 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7557 sizeof(int), 0644, proc_dointvec_minmax);
7558 set_table_entry(&table[9], "cache_nice_tries",
7559 &sd->cache_nice_tries,
7560 sizeof(int), 0644, proc_dointvec_minmax);
7561 set_table_entry(&table[10], "flags", &sd->flags,
7562 sizeof(int), 0644, proc_dointvec_minmax);
7563 set_table_entry(&table[11], "name", sd->name,
7564 CORENAME_MAX_SIZE, 0444, proc_dostring);
7565 /* &table[12] is terminator */
7567 return table;
7570 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7572 struct ctl_table *entry, *table;
7573 struct sched_domain *sd;
7574 int domain_num = 0, i;
7575 char buf[32];
7577 for_each_domain(cpu, sd)
7578 domain_num++;
7579 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7580 if (table == NULL)
7581 return NULL;
7583 i = 0;
7584 for_each_domain(cpu, sd) {
7585 snprintf(buf, 32, "domain%d", i);
7586 entry->procname = kstrdup(buf, GFP_KERNEL);
7587 entry->mode = 0555;
7588 entry->child = sd_alloc_ctl_domain_table(sd);
7589 entry++;
7590 i++;
7592 return table;
7595 static struct ctl_table_header *sd_sysctl_header;
7596 static void register_sched_domain_sysctl(void)
7598 int i, cpu_num = num_possible_cpus();
7599 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7600 char buf[32];
7602 WARN_ON(sd_ctl_dir[0].child);
7603 sd_ctl_dir[0].child = entry;
7605 if (entry == NULL)
7606 return;
7608 for_each_possible_cpu(i) {
7609 snprintf(buf, 32, "cpu%d", i);
7610 entry->procname = kstrdup(buf, GFP_KERNEL);
7611 entry->mode = 0555;
7612 entry->child = sd_alloc_ctl_cpu_table(i);
7613 entry++;
7616 WARN_ON(sd_sysctl_header);
7617 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7620 /* may be called multiple times per register */
7621 static void unregister_sched_domain_sysctl(void)
7623 if (sd_sysctl_header)
7624 unregister_sysctl_table(sd_sysctl_header);
7625 sd_sysctl_header = NULL;
7626 if (sd_ctl_dir[0].child)
7627 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7629 #else
7630 static void register_sched_domain_sysctl(void)
7633 static void unregister_sched_domain_sysctl(void)
7636 #endif
7638 static void set_rq_online(struct rq *rq)
7640 if (!rq->online) {
7641 const struct sched_class *class;
7643 cpumask_set_cpu(rq->cpu, rq->rd->online);
7644 rq->online = 1;
7646 for_each_class(class) {
7647 if (class->rq_online)
7648 class->rq_online(rq);
7653 static void set_rq_offline(struct rq *rq)
7655 if (rq->online) {
7656 const struct sched_class *class;
7658 for_each_class(class) {
7659 if (class->rq_offline)
7660 class->rq_offline(rq);
7663 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7664 rq->online = 0;
7669 * migration_call - callback that gets triggered when a CPU is added.
7670 * Here we can start up the necessary migration thread for the new CPU.
7672 static int __cpuinit
7673 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7675 struct task_struct *p;
7676 int cpu = (long)hcpu;
7677 unsigned long flags;
7678 struct rq *rq;
7680 switch (action) {
7682 case CPU_UP_PREPARE:
7683 case CPU_UP_PREPARE_FROZEN:
7684 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7685 if (IS_ERR(p))
7686 return NOTIFY_BAD;
7687 kthread_bind(p, cpu);
7688 /* Must be high prio: stop_machine expects to yield to it. */
7689 rq = task_rq_lock(p, &flags);
7690 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7691 task_rq_unlock(rq, &flags);
7692 get_task_struct(p);
7693 cpu_rq(cpu)->migration_thread = p;
7694 rq->calc_load_update = calc_load_update;
7695 break;
7697 case CPU_ONLINE:
7698 case CPU_ONLINE_FROZEN:
7699 /* Strictly unnecessary, as first user will wake it. */
7700 wake_up_process(cpu_rq(cpu)->migration_thread);
7702 /* Update our root-domain */
7703 rq = cpu_rq(cpu);
7704 raw_spin_lock_irqsave(&rq->lock, flags);
7705 if (rq->rd) {
7706 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7708 set_rq_online(rq);
7710 raw_spin_unlock_irqrestore(&rq->lock, flags);
7711 break;
7713 #ifdef CONFIG_HOTPLUG_CPU
7714 case CPU_UP_CANCELED:
7715 case CPU_UP_CANCELED_FROZEN:
7716 if (!cpu_rq(cpu)->migration_thread)
7717 break;
7718 /* Unbind it from offline cpu so it can run. Fall thru. */
7719 kthread_bind(cpu_rq(cpu)->migration_thread,
7720 cpumask_any(cpu_online_mask));
7721 kthread_stop(cpu_rq(cpu)->migration_thread);
7722 put_task_struct(cpu_rq(cpu)->migration_thread);
7723 cpu_rq(cpu)->migration_thread = NULL;
7724 break;
7726 case CPU_DEAD:
7727 case CPU_DEAD_FROZEN:
7728 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7729 migrate_live_tasks(cpu);
7730 rq = cpu_rq(cpu);
7731 kthread_stop(rq->migration_thread);
7732 put_task_struct(rq->migration_thread);
7733 rq->migration_thread = NULL;
7734 /* Idle task back to normal (off runqueue, low prio) */
7735 raw_spin_lock_irq(&rq->lock);
7736 update_rq_clock(rq);
7737 deactivate_task(rq, rq->idle, 0);
7738 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7739 rq->idle->sched_class = &idle_sched_class;
7740 migrate_dead_tasks(cpu);
7741 raw_spin_unlock_irq(&rq->lock);
7742 cpuset_unlock();
7743 migrate_nr_uninterruptible(rq);
7744 BUG_ON(rq->nr_running != 0);
7745 calc_global_load_remove(rq);
7747 * No need to migrate the tasks: it was best-effort if
7748 * they didn't take sched_hotcpu_mutex. Just wake up
7749 * the requestors.
7751 raw_spin_lock_irq(&rq->lock);
7752 while (!list_empty(&rq->migration_queue)) {
7753 struct migration_req *req;
7755 req = list_entry(rq->migration_queue.next,
7756 struct migration_req, list);
7757 list_del_init(&req->list);
7758 raw_spin_unlock_irq(&rq->lock);
7759 complete(&req->done);
7760 raw_spin_lock_irq(&rq->lock);
7762 raw_spin_unlock_irq(&rq->lock);
7763 break;
7765 case CPU_DYING:
7766 case CPU_DYING_FROZEN:
7767 /* Update our root-domain */
7768 rq = cpu_rq(cpu);
7769 raw_spin_lock_irqsave(&rq->lock, flags);
7770 if (rq->rd) {
7771 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7772 set_rq_offline(rq);
7774 raw_spin_unlock_irqrestore(&rq->lock, flags);
7775 break;
7776 #endif
7778 return NOTIFY_OK;
7782 * Register at high priority so that task migration (migrate_all_tasks)
7783 * happens before everything else. This has to be lower priority than
7784 * the notifier in the perf_event subsystem, though.
7786 static struct notifier_block __cpuinitdata migration_notifier = {
7787 .notifier_call = migration_call,
7788 .priority = 10
7791 static int __init migration_init(void)
7793 void *cpu = (void *)(long)smp_processor_id();
7794 int err;
7796 /* Start one for the boot CPU: */
7797 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7798 BUG_ON(err == NOTIFY_BAD);
7799 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7800 register_cpu_notifier(&migration_notifier);
7802 return 0;
7804 early_initcall(migration_init);
7805 #endif
7807 #ifdef CONFIG_SMP
7809 #ifdef CONFIG_SCHED_DEBUG
7811 static __read_mostly int sched_domain_debug_enabled;
7813 static int __init sched_domain_debug_setup(char *str)
7815 sched_domain_debug_enabled = 1;
7817 return 0;
7819 early_param("sched_debug", sched_domain_debug_setup);
7821 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7822 struct cpumask *groupmask)
7824 struct sched_group *group = sd->groups;
7825 char str[256];
7827 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7828 cpumask_clear(groupmask);
7830 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7832 if (!(sd->flags & SD_LOAD_BALANCE)) {
7833 printk("does not load-balance\n");
7834 if (sd->parent)
7835 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7836 " has parent");
7837 return -1;
7840 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7842 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7843 printk(KERN_ERR "ERROR: domain->span does not contain "
7844 "CPU%d\n", cpu);
7846 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7847 printk(KERN_ERR "ERROR: domain->groups does not contain"
7848 " CPU%d\n", cpu);
7851 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7852 do {
7853 if (!group) {
7854 printk("\n");
7855 printk(KERN_ERR "ERROR: group is NULL\n");
7856 break;
7859 if (!group->cpu_power) {
7860 printk(KERN_CONT "\n");
7861 printk(KERN_ERR "ERROR: domain->cpu_power not "
7862 "set\n");
7863 break;
7866 if (!cpumask_weight(sched_group_cpus(group))) {
7867 printk(KERN_CONT "\n");
7868 printk(KERN_ERR "ERROR: empty group\n");
7869 break;
7872 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7873 printk(KERN_CONT "\n");
7874 printk(KERN_ERR "ERROR: repeated CPUs\n");
7875 break;
7878 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7880 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7882 printk(KERN_CONT " %s", str);
7883 if (group->cpu_power != SCHED_LOAD_SCALE) {
7884 printk(KERN_CONT " (cpu_power = %d)",
7885 group->cpu_power);
7888 group = group->next;
7889 } while (group != sd->groups);
7890 printk(KERN_CONT "\n");
7892 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7893 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7895 if (sd->parent &&
7896 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7897 printk(KERN_ERR "ERROR: parent span is not a superset "
7898 "of domain->span\n");
7899 return 0;
7902 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7904 cpumask_var_t groupmask;
7905 int level = 0;
7907 if (!sched_domain_debug_enabled)
7908 return;
7910 if (!sd) {
7911 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7912 return;
7915 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7917 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7918 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7919 return;
7922 for (;;) {
7923 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7924 break;
7925 level++;
7926 sd = sd->parent;
7927 if (!sd)
7928 break;
7930 free_cpumask_var(groupmask);
7932 #else /* !CONFIG_SCHED_DEBUG */
7933 # define sched_domain_debug(sd, cpu) do { } while (0)
7934 #endif /* CONFIG_SCHED_DEBUG */
7936 static int sd_degenerate(struct sched_domain *sd)
7938 if (cpumask_weight(sched_domain_span(sd)) == 1)
7939 return 1;
7941 /* Following flags need at least 2 groups */
7942 if (sd->flags & (SD_LOAD_BALANCE |
7943 SD_BALANCE_NEWIDLE |
7944 SD_BALANCE_FORK |
7945 SD_BALANCE_EXEC |
7946 SD_SHARE_CPUPOWER |
7947 SD_SHARE_PKG_RESOURCES)) {
7948 if (sd->groups != sd->groups->next)
7949 return 0;
7952 /* Following flags don't use groups */
7953 if (sd->flags & (SD_WAKE_AFFINE))
7954 return 0;
7956 return 1;
7959 static int
7960 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7962 unsigned long cflags = sd->flags, pflags = parent->flags;
7964 if (sd_degenerate(parent))
7965 return 1;
7967 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7968 return 0;
7970 /* Flags needing groups don't count if only 1 group in parent */
7971 if (parent->groups == parent->groups->next) {
7972 pflags &= ~(SD_LOAD_BALANCE |
7973 SD_BALANCE_NEWIDLE |
7974 SD_BALANCE_FORK |
7975 SD_BALANCE_EXEC |
7976 SD_SHARE_CPUPOWER |
7977 SD_SHARE_PKG_RESOURCES);
7978 if (nr_node_ids == 1)
7979 pflags &= ~SD_SERIALIZE;
7981 if (~cflags & pflags)
7982 return 0;
7984 return 1;
7987 static void free_rootdomain(struct root_domain *rd)
7989 synchronize_sched();
7991 cpupri_cleanup(&rd->cpupri);
7993 free_cpumask_var(rd->rto_mask);
7994 free_cpumask_var(rd->online);
7995 free_cpumask_var(rd->span);
7996 kfree(rd);
7999 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
8001 struct root_domain *old_rd = NULL;
8002 unsigned long flags;
8004 raw_spin_lock_irqsave(&rq->lock, flags);
8006 if (rq->rd) {
8007 old_rd = rq->rd;
8009 if (cpumask_test_cpu(rq->cpu, old_rd->online))
8010 set_rq_offline(rq);
8012 cpumask_clear_cpu(rq->cpu, old_rd->span);
8015 * If we dont want to free the old_rt yet then
8016 * set old_rd to NULL to skip the freeing later
8017 * in this function:
8019 if (!atomic_dec_and_test(&old_rd->refcount))
8020 old_rd = NULL;
8023 atomic_inc(&rd->refcount);
8024 rq->rd = rd;
8026 cpumask_set_cpu(rq->cpu, rd->span);
8027 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
8028 set_rq_online(rq);
8030 raw_spin_unlock_irqrestore(&rq->lock, flags);
8032 if (old_rd)
8033 free_rootdomain(old_rd);
8036 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8038 gfp_t gfp = GFP_KERNEL;
8040 memset(rd, 0, sizeof(*rd));
8042 if (bootmem)
8043 gfp = GFP_NOWAIT;
8045 if (!alloc_cpumask_var(&rd->span, gfp))
8046 goto out;
8047 if (!alloc_cpumask_var(&rd->online, gfp))
8048 goto free_span;
8049 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8050 goto free_online;
8052 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8053 goto free_rto_mask;
8054 return 0;
8056 free_rto_mask:
8057 free_cpumask_var(rd->rto_mask);
8058 free_online:
8059 free_cpumask_var(rd->online);
8060 free_span:
8061 free_cpumask_var(rd->span);
8062 out:
8063 return -ENOMEM;
8066 static void init_defrootdomain(void)
8068 init_rootdomain(&def_root_domain, true);
8070 atomic_set(&def_root_domain.refcount, 1);
8073 static struct root_domain *alloc_rootdomain(void)
8075 struct root_domain *rd;
8077 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8078 if (!rd)
8079 return NULL;
8081 if (init_rootdomain(rd, false) != 0) {
8082 kfree(rd);
8083 return NULL;
8086 return rd;
8090 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8091 * hold the hotplug lock.
8093 static void
8094 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8096 struct rq *rq = cpu_rq(cpu);
8097 struct sched_domain *tmp;
8099 /* Remove the sched domains which do not contribute to scheduling. */
8100 for (tmp = sd; tmp; ) {
8101 struct sched_domain *parent = tmp->parent;
8102 if (!parent)
8103 break;
8105 if (sd_parent_degenerate(tmp, parent)) {
8106 tmp->parent = parent->parent;
8107 if (parent->parent)
8108 parent->parent->child = tmp;
8109 } else
8110 tmp = tmp->parent;
8113 if (sd && sd_degenerate(sd)) {
8114 sd = sd->parent;
8115 if (sd)
8116 sd->child = NULL;
8119 sched_domain_debug(sd, cpu);
8121 rq_attach_root(rq, rd);
8122 rcu_assign_pointer(rq->sd, sd);
8125 /* cpus with isolated domains */
8126 static cpumask_var_t cpu_isolated_map;
8128 /* Setup the mask of cpus configured for isolated domains */
8129 static int __init isolated_cpu_setup(char *str)
8131 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8132 cpulist_parse(str, cpu_isolated_map);
8133 return 1;
8136 __setup("isolcpus=", isolated_cpu_setup);
8139 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8140 * to a function which identifies what group(along with sched group) a CPU
8141 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8142 * (due to the fact that we keep track of groups covered with a struct cpumask).
8144 * init_sched_build_groups will build a circular linked list of the groups
8145 * covered by the given span, and will set each group's ->cpumask correctly,
8146 * and ->cpu_power to 0.
8148 static void
8149 init_sched_build_groups(const struct cpumask *span,
8150 const struct cpumask *cpu_map,
8151 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8152 struct sched_group **sg,
8153 struct cpumask *tmpmask),
8154 struct cpumask *covered, struct cpumask *tmpmask)
8156 struct sched_group *first = NULL, *last = NULL;
8157 int i;
8159 cpumask_clear(covered);
8161 for_each_cpu(i, span) {
8162 struct sched_group *sg;
8163 int group = group_fn(i, cpu_map, &sg, tmpmask);
8164 int j;
8166 if (cpumask_test_cpu(i, covered))
8167 continue;
8169 cpumask_clear(sched_group_cpus(sg));
8170 sg->cpu_power = 0;
8172 for_each_cpu(j, span) {
8173 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8174 continue;
8176 cpumask_set_cpu(j, covered);
8177 cpumask_set_cpu(j, sched_group_cpus(sg));
8179 if (!first)
8180 first = sg;
8181 if (last)
8182 last->next = sg;
8183 last = sg;
8185 last->next = first;
8188 #define SD_NODES_PER_DOMAIN 16
8190 #ifdef CONFIG_NUMA
8193 * find_next_best_node - find the next node to include in a sched_domain
8194 * @node: node whose sched_domain we're building
8195 * @used_nodes: nodes already in the sched_domain
8197 * Find the next node to include in a given scheduling domain. Simply
8198 * finds the closest node not already in the @used_nodes map.
8200 * Should use nodemask_t.
8202 static int find_next_best_node(int node, nodemask_t *used_nodes)
8204 int i, n, val, min_val, best_node = 0;
8206 min_val = INT_MAX;
8208 for (i = 0; i < nr_node_ids; i++) {
8209 /* Start at @node */
8210 n = (node + i) % nr_node_ids;
8212 if (!nr_cpus_node(n))
8213 continue;
8215 /* Skip already used nodes */
8216 if (node_isset(n, *used_nodes))
8217 continue;
8219 /* Simple min distance search */
8220 val = node_distance(node, n);
8222 if (val < min_val) {
8223 min_val = val;
8224 best_node = n;
8228 node_set(best_node, *used_nodes);
8229 return best_node;
8233 * sched_domain_node_span - get a cpumask for a node's sched_domain
8234 * @node: node whose cpumask we're constructing
8235 * @span: resulting cpumask
8237 * Given a node, construct a good cpumask for its sched_domain to span. It
8238 * should be one that prevents unnecessary balancing, but also spreads tasks
8239 * out optimally.
8241 static void sched_domain_node_span(int node, struct cpumask *span)
8243 nodemask_t used_nodes;
8244 int i;
8246 cpumask_clear(span);
8247 nodes_clear(used_nodes);
8249 cpumask_or(span, span, cpumask_of_node(node));
8250 node_set(node, used_nodes);
8252 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8253 int next_node = find_next_best_node(node, &used_nodes);
8255 cpumask_or(span, span, cpumask_of_node(next_node));
8258 #endif /* CONFIG_NUMA */
8260 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8263 * The cpus mask in sched_group and sched_domain hangs off the end.
8265 * ( See the the comments in include/linux/sched.h:struct sched_group
8266 * and struct sched_domain. )
8268 struct static_sched_group {
8269 struct sched_group sg;
8270 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8273 struct static_sched_domain {
8274 struct sched_domain sd;
8275 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8278 struct s_data {
8279 #ifdef CONFIG_NUMA
8280 int sd_allnodes;
8281 cpumask_var_t domainspan;
8282 cpumask_var_t covered;
8283 cpumask_var_t notcovered;
8284 #endif
8285 cpumask_var_t nodemask;
8286 cpumask_var_t this_sibling_map;
8287 cpumask_var_t this_core_map;
8288 cpumask_var_t send_covered;
8289 cpumask_var_t tmpmask;
8290 struct sched_group **sched_group_nodes;
8291 struct root_domain *rd;
8294 enum s_alloc {
8295 sa_sched_groups = 0,
8296 sa_rootdomain,
8297 sa_tmpmask,
8298 sa_send_covered,
8299 sa_this_core_map,
8300 sa_this_sibling_map,
8301 sa_nodemask,
8302 sa_sched_group_nodes,
8303 #ifdef CONFIG_NUMA
8304 sa_notcovered,
8305 sa_covered,
8306 sa_domainspan,
8307 #endif
8308 sa_none,
8312 * SMT sched-domains:
8314 #ifdef CONFIG_SCHED_SMT
8315 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8316 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
8318 static int
8319 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8320 struct sched_group **sg, struct cpumask *unused)
8322 if (sg)
8323 *sg = &per_cpu(sched_groups, cpu).sg;
8324 return cpu;
8326 #endif /* CONFIG_SCHED_SMT */
8329 * multi-core sched-domains:
8331 #ifdef CONFIG_SCHED_MC
8332 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8333 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8334 #endif /* CONFIG_SCHED_MC */
8336 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8337 static int
8338 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8339 struct sched_group **sg, struct cpumask *mask)
8341 int group;
8343 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8344 group = cpumask_first(mask);
8345 if (sg)
8346 *sg = &per_cpu(sched_group_core, group).sg;
8347 return group;
8349 #elif defined(CONFIG_SCHED_MC)
8350 static int
8351 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8352 struct sched_group **sg, struct cpumask *unused)
8354 if (sg)
8355 *sg = &per_cpu(sched_group_core, cpu).sg;
8356 return cpu;
8358 #endif
8360 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8361 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8363 static int
8364 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8365 struct sched_group **sg, struct cpumask *mask)
8367 int group;
8368 #ifdef CONFIG_SCHED_MC
8369 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8370 group = cpumask_first(mask);
8371 #elif defined(CONFIG_SCHED_SMT)
8372 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8373 group = cpumask_first(mask);
8374 #else
8375 group = cpu;
8376 #endif
8377 if (sg)
8378 *sg = &per_cpu(sched_group_phys, group).sg;
8379 return group;
8382 #ifdef CONFIG_NUMA
8384 * The init_sched_build_groups can't handle what we want to do with node
8385 * groups, so roll our own. Now each node has its own list of groups which
8386 * gets dynamically allocated.
8388 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8389 static struct sched_group ***sched_group_nodes_bycpu;
8391 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8392 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8394 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8395 struct sched_group **sg,
8396 struct cpumask *nodemask)
8398 int group;
8400 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8401 group = cpumask_first(nodemask);
8403 if (sg)
8404 *sg = &per_cpu(sched_group_allnodes, group).sg;
8405 return group;
8408 static void init_numa_sched_groups_power(struct sched_group *group_head)
8410 struct sched_group *sg = group_head;
8411 int j;
8413 if (!sg)
8414 return;
8415 do {
8416 for_each_cpu(j, sched_group_cpus(sg)) {
8417 struct sched_domain *sd;
8419 sd = &per_cpu(phys_domains, j).sd;
8420 if (j != group_first_cpu(sd->groups)) {
8422 * Only add "power" once for each
8423 * physical package.
8425 continue;
8428 sg->cpu_power += sd->groups->cpu_power;
8430 sg = sg->next;
8431 } while (sg != group_head);
8434 static int build_numa_sched_groups(struct s_data *d,
8435 const struct cpumask *cpu_map, int num)
8437 struct sched_domain *sd;
8438 struct sched_group *sg, *prev;
8439 int n, j;
8441 cpumask_clear(d->covered);
8442 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8443 if (cpumask_empty(d->nodemask)) {
8444 d->sched_group_nodes[num] = NULL;
8445 goto out;
8448 sched_domain_node_span(num, d->domainspan);
8449 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8451 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8452 GFP_KERNEL, num);
8453 if (!sg) {
8454 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8455 num);
8456 return -ENOMEM;
8458 d->sched_group_nodes[num] = sg;
8460 for_each_cpu(j, d->nodemask) {
8461 sd = &per_cpu(node_domains, j).sd;
8462 sd->groups = sg;
8465 sg->cpu_power = 0;
8466 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8467 sg->next = sg;
8468 cpumask_or(d->covered, d->covered, d->nodemask);
8470 prev = sg;
8471 for (j = 0; j < nr_node_ids; j++) {
8472 n = (num + j) % nr_node_ids;
8473 cpumask_complement(d->notcovered, d->covered);
8474 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8475 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8476 if (cpumask_empty(d->tmpmask))
8477 break;
8478 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8479 if (cpumask_empty(d->tmpmask))
8480 continue;
8481 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8482 GFP_KERNEL, num);
8483 if (!sg) {
8484 printk(KERN_WARNING
8485 "Can not alloc domain group for node %d\n", j);
8486 return -ENOMEM;
8488 sg->cpu_power = 0;
8489 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8490 sg->next = prev->next;
8491 cpumask_or(d->covered, d->covered, d->tmpmask);
8492 prev->next = sg;
8493 prev = sg;
8495 out:
8496 return 0;
8498 #endif /* CONFIG_NUMA */
8500 #ifdef CONFIG_NUMA
8501 /* Free memory allocated for various sched_group structures */
8502 static void free_sched_groups(const struct cpumask *cpu_map,
8503 struct cpumask *nodemask)
8505 int cpu, i;
8507 for_each_cpu(cpu, cpu_map) {
8508 struct sched_group **sched_group_nodes
8509 = sched_group_nodes_bycpu[cpu];
8511 if (!sched_group_nodes)
8512 continue;
8514 for (i = 0; i < nr_node_ids; i++) {
8515 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8517 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8518 if (cpumask_empty(nodemask))
8519 continue;
8521 if (sg == NULL)
8522 continue;
8523 sg = sg->next;
8524 next_sg:
8525 oldsg = sg;
8526 sg = sg->next;
8527 kfree(oldsg);
8528 if (oldsg != sched_group_nodes[i])
8529 goto next_sg;
8531 kfree(sched_group_nodes);
8532 sched_group_nodes_bycpu[cpu] = NULL;
8535 #else /* !CONFIG_NUMA */
8536 static void free_sched_groups(const struct cpumask *cpu_map,
8537 struct cpumask *nodemask)
8540 #endif /* CONFIG_NUMA */
8543 * Initialize sched groups cpu_power.
8545 * cpu_power indicates the capacity of sched group, which is used while
8546 * distributing the load between different sched groups in a sched domain.
8547 * Typically cpu_power for all the groups in a sched domain will be same unless
8548 * there are asymmetries in the topology. If there are asymmetries, group
8549 * having more cpu_power will pickup more load compared to the group having
8550 * less cpu_power.
8552 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8554 struct sched_domain *child;
8555 struct sched_group *group;
8556 long power;
8557 int weight;
8559 WARN_ON(!sd || !sd->groups);
8561 if (cpu != group_first_cpu(sd->groups))
8562 return;
8564 child = sd->child;
8566 sd->groups->cpu_power = 0;
8568 if (!child) {
8569 power = SCHED_LOAD_SCALE;
8570 weight = cpumask_weight(sched_domain_span(sd));
8572 * SMT siblings share the power of a single core.
8573 * Usually multiple threads get a better yield out of
8574 * that one core than a single thread would have,
8575 * reflect that in sd->smt_gain.
8577 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8578 power *= sd->smt_gain;
8579 power /= weight;
8580 power >>= SCHED_LOAD_SHIFT;
8582 sd->groups->cpu_power += power;
8583 return;
8587 * Add cpu_power of each child group to this groups cpu_power.
8589 group = child->groups;
8590 do {
8591 sd->groups->cpu_power += group->cpu_power;
8592 group = group->next;
8593 } while (group != child->groups);
8597 * Initializers for schedule domains
8598 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8601 #ifdef CONFIG_SCHED_DEBUG
8602 # define SD_INIT_NAME(sd, type) sd->name = #type
8603 #else
8604 # define SD_INIT_NAME(sd, type) do { } while (0)
8605 #endif
8607 #define SD_INIT(sd, type) sd_init_##type(sd)
8609 #define SD_INIT_FUNC(type) \
8610 static noinline void sd_init_##type(struct sched_domain *sd) \
8612 memset(sd, 0, sizeof(*sd)); \
8613 *sd = SD_##type##_INIT; \
8614 sd->level = SD_LV_##type; \
8615 SD_INIT_NAME(sd, type); \
8618 SD_INIT_FUNC(CPU)
8619 #ifdef CONFIG_NUMA
8620 SD_INIT_FUNC(ALLNODES)
8621 SD_INIT_FUNC(NODE)
8622 #endif
8623 #ifdef CONFIG_SCHED_SMT
8624 SD_INIT_FUNC(SIBLING)
8625 #endif
8626 #ifdef CONFIG_SCHED_MC
8627 SD_INIT_FUNC(MC)
8628 #endif
8630 static int default_relax_domain_level = -1;
8632 static int __init setup_relax_domain_level(char *str)
8634 unsigned long val;
8636 val = simple_strtoul(str, NULL, 0);
8637 if (val < SD_LV_MAX)
8638 default_relax_domain_level = val;
8640 return 1;
8642 __setup("relax_domain_level=", setup_relax_domain_level);
8644 static void set_domain_attribute(struct sched_domain *sd,
8645 struct sched_domain_attr *attr)
8647 int request;
8649 if (!attr || attr->relax_domain_level < 0) {
8650 if (default_relax_domain_level < 0)
8651 return;
8652 else
8653 request = default_relax_domain_level;
8654 } else
8655 request = attr->relax_domain_level;
8656 if (request < sd->level) {
8657 /* turn off idle balance on this domain */
8658 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8659 } else {
8660 /* turn on idle balance on this domain */
8661 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8665 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8666 const struct cpumask *cpu_map)
8668 switch (what) {
8669 case sa_sched_groups:
8670 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8671 d->sched_group_nodes = NULL;
8672 case sa_rootdomain:
8673 free_rootdomain(d->rd); /* fall through */
8674 case sa_tmpmask:
8675 free_cpumask_var(d->tmpmask); /* fall through */
8676 case sa_send_covered:
8677 free_cpumask_var(d->send_covered); /* fall through */
8678 case sa_this_core_map:
8679 free_cpumask_var(d->this_core_map); /* fall through */
8680 case sa_this_sibling_map:
8681 free_cpumask_var(d->this_sibling_map); /* fall through */
8682 case sa_nodemask:
8683 free_cpumask_var(d->nodemask); /* fall through */
8684 case sa_sched_group_nodes:
8685 #ifdef CONFIG_NUMA
8686 kfree(d->sched_group_nodes); /* fall through */
8687 case sa_notcovered:
8688 free_cpumask_var(d->notcovered); /* fall through */
8689 case sa_covered:
8690 free_cpumask_var(d->covered); /* fall through */
8691 case sa_domainspan:
8692 free_cpumask_var(d->domainspan); /* fall through */
8693 #endif
8694 case sa_none:
8695 break;
8699 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8700 const struct cpumask *cpu_map)
8702 #ifdef CONFIG_NUMA
8703 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8704 return sa_none;
8705 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8706 return sa_domainspan;
8707 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8708 return sa_covered;
8709 /* Allocate the per-node list of sched groups */
8710 d->sched_group_nodes = kcalloc(nr_node_ids,
8711 sizeof(struct sched_group *), GFP_KERNEL);
8712 if (!d->sched_group_nodes) {
8713 printk(KERN_WARNING "Can not alloc sched group node list\n");
8714 return sa_notcovered;
8716 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8717 #endif
8718 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8719 return sa_sched_group_nodes;
8720 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8721 return sa_nodemask;
8722 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8723 return sa_this_sibling_map;
8724 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8725 return sa_this_core_map;
8726 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8727 return sa_send_covered;
8728 d->rd = alloc_rootdomain();
8729 if (!d->rd) {
8730 printk(KERN_WARNING "Cannot alloc root domain\n");
8731 return sa_tmpmask;
8733 return sa_rootdomain;
8736 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8737 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8739 struct sched_domain *sd = NULL;
8740 #ifdef CONFIG_NUMA
8741 struct sched_domain *parent;
8743 d->sd_allnodes = 0;
8744 if (cpumask_weight(cpu_map) >
8745 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8746 sd = &per_cpu(allnodes_domains, i).sd;
8747 SD_INIT(sd, ALLNODES);
8748 set_domain_attribute(sd, attr);
8749 cpumask_copy(sched_domain_span(sd), cpu_map);
8750 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8751 d->sd_allnodes = 1;
8753 parent = sd;
8755 sd = &per_cpu(node_domains, i).sd;
8756 SD_INIT(sd, NODE);
8757 set_domain_attribute(sd, attr);
8758 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8759 sd->parent = parent;
8760 if (parent)
8761 parent->child = sd;
8762 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8763 #endif
8764 return sd;
8767 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8768 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8769 struct sched_domain *parent, int i)
8771 struct sched_domain *sd;
8772 sd = &per_cpu(phys_domains, i).sd;
8773 SD_INIT(sd, CPU);
8774 set_domain_attribute(sd, attr);
8775 cpumask_copy(sched_domain_span(sd), d->nodemask);
8776 sd->parent = parent;
8777 if (parent)
8778 parent->child = sd;
8779 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8780 return sd;
8783 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8784 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8785 struct sched_domain *parent, int i)
8787 struct sched_domain *sd = parent;
8788 #ifdef CONFIG_SCHED_MC
8789 sd = &per_cpu(core_domains, i).sd;
8790 SD_INIT(sd, MC);
8791 set_domain_attribute(sd, attr);
8792 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8793 sd->parent = parent;
8794 parent->child = sd;
8795 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8796 #endif
8797 return sd;
8800 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8801 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8802 struct sched_domain *parent, int i)
8804 struct sched_domain *sd = parent;
8805 #ifdef CONFIG_SCHED_SMT
8806 sd = &per_cpu(cpu_domains, i).sd;
8807 SD_INIT(sd, SIBLING);
8808 set_domain_attribute(sd, attr);
8809 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8810 sd->parent = parent;
8811 parent->child = sd;
8812 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8813 #endif
8814 return sd;
8817 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8818 const struct cpumask *cpu_map, int cpu)
8820 switch (l) {
8821 #ifdef CONFIG_SCHED_SMT
8822 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8823 cpumask_and(d->this_sibling_map, cpu_map,
8824 topology_thread_cpumask(cpu));
8825 if (cpu == cpumask_first(d->this_sibling_map))
8826 init_sched_build_groups(d->this_sibling_map, cpu_map,
8827 &cpu_to_cpu_group,
8828 d->send_covered, d->tmpmask);
8829 break;
8830 #endif
8831 #ifdef CONFIG_SCHED_MC
8832 case SD_LV_MC: /* set up multi-core groups */
8833 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8834 if (cpu == cpumask_first(d->this_core_map))
8835 init_sched_build_groups(d->this_core_map, cpu_map,
8836 &cpu_to_core_group,
8837 d->send_covered, d->tmpmask);
8838 break;
8839 #endif
8840 case SD_LV_CPU: /* set up physical groups */
8841 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8842 if (!cpumask_empty(d->nodemask))
8843 init_sched_build_groups(d->nodemask, cpu_map,
8844 &cpu_to_phys_group,
8845 d->send_covered, d->tmpmask);
8846 break;
8847 #ifdef CONFIG_NUMA
8848 case SD_LV_ALLNODES:
8849 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8850 d->send_covered, d->tmpmask);
8851 break;
8852 #endif
8853 default:
8854 break;
8859 * Build sched domains for a given set of cpus and attach the sched domains
8860 * to the individual cpus
8862 static int __build_sched_domains(const struct cpumask *cpu_map,
8863 struct sched_domain_attr *attr)
8865 enum s_alloc alloc_state = sa_none;
8866 struct s_data d;
8867 struct sched_domain *sd;
8868 int i;
8869 #ifdef CONFIG_NUMA
8870 d.sd_allnodes = 0;
8871 #endif
8873 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8874 if (alloc_state != sa_rootdomain)
8875 goto error;
8876 alloc_state = sa_sched_groups;
8879 * Set up domains for cpus specified by the cpu_map.
8881 for_each_cpu(i, cpu_map) {
8882 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8883 cpu_map);
8885 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8886 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8887 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8888 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8891 for_each_cpu(i, cpu_map) {
8892 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8893 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8896 /* Set up physical groups */
8897 for (i = 0; i < nr_node_ids; i++)
8898 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8900 #ifdef CONFIG_NUMA
8901 /* Set up node groups */
8902 if (d.sd_allnodes)
8903 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8905 for (i = 0; i < nr_node_ids; i++)
8906 if (build_numa_sched_groups(&d, cpu_map, i))
8907 goto error;
8908 #endif
8910 /* Calculate CPU power for physical packages and nodes */
8911 #ifdef CONFIG_SCHED_SMT
8912 for_each_cpu(i, cpu_map) {
8913 sd = &per_cpu(cpu_domains, i).sd;
8914 init_sched_groups_power(i, sd);
8916 #endif
8917 #ifdef CONFIG_SCHED_MC
8918 for_each_cpu(i, cpu_map) {
8919 sd = &per_cpu(core_domains, i).sd;
8920 init_sched_groups_power(i, sd);
8922 #endif
8924 for_each_cpu(i, cpu_map) {
8925 sd = &per_cpu(phys_domains, i).sd;
8926 init_sched_groups_power(i, sd);
8929 #ifdef CONFIG_NUMA
8930 for (i = 0; i < nr_node_ids; i++)
8931 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8933 if (d.sd_allnodes) {
8934 struct sched_group *sg;
8936 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8937 d.tmpmask);
8938 init_numa_sched_groups_power(sg);
8940 #endif
8942 /* Attach the domains */
8943 for_each_cpu(i, cpu_map) {
8944 #ifdef CONFIG_SCHED_SMT
8945 sd = &per_cpu(cpu_domains, i).sd;
8946 #elif defined(CONFIG_SCHED_MC)
8947 sd = &per_cpu(core_domains, i).sd;
8948 #else
8949 sd = &per_cpu(phys_domains, i).sd;
8950 #endif
8951 cpu_attach_domain(sd, d.rd, i);
8954 d.sched_group_nodes = NULL; /* don't free this we still need it */
8955 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8956 return 0;
8958 error:
8959 __free_domain_allocs(&d, alloc_state, cpu_map);
8960 return -ENOMEM;
8963 static int build_sched_domains(const struct cpumask *cpu_map)
8965 return __build_sched_domains(cpu_map, NULL);
8968 static cpumask_var_t *doms_cur; /* current sched domains */
8969 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8970 static struct sched_domain_attr *dattr_cur;
8971 /* attribues of custom domains in 'doms_cur' */
8974 * Special case: If a kmalloc of a doms_cur partition (array of
8975 * cpumask) fails, then fallback to a single sched domain,
8976 * as determined by the single cpumask fallback_doms.
8978 static cpumask_var_t fallback_doms;
8981 * arch_update_cpu_topology lets virtualized architectures update the
8982 * cpu core maps. It is supposed to return 1 if the topology changed
8983 * or 0 if it stayed the same.
8985 int __attribute__((weak)) arch_update_cpu_topology(void)
8987 return 0;
8990 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
8992 int i;
8993 cpumask_var_t *doms;
8995 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
8996 if (!doms)
8997 return NULL;
8998 for (i = 0; i < ndoms; i++) {
8999 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
9000 free_sched_domains(doms, i);
9001 return NULL;
9004 return doms;
9007 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
9009 unsigned int i;
9010 for (i = 0; i < ndoms; i++)
9011 free_cpumask_var(doms[i]);
9012 kfree(doms);
9016 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9017 * For now this just excludes isolated cpus, but could be used to
9018 * exclude other special cases in the future.
9020 static int arch_init_sched_domains(const struct cpumask *cpu_map)
9022 int err;
9024 arch_update_cpu_topology();
9025 ndoms_cur = 1;
9026 doms_cur = alloc_sched_domains(ndoms_cur);
9027 if (!doms_cur)
9028 doms_cur = &fallback_doms;
9029 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
9030 dattr_cur = NULL;
9031 err = build_sched_domains(doms_cur[0]);
9032 register_sched_domain_sysctl();
9034 return err;
9037 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9038 struct cpumask *tmpmask)
9040 free_sched_groups(cpu_map, tmpmask);
9044 * Detach sched domains from a group of cpus specified in cpu_map
9045 * These cpus will now be attached to the NULL domain
9047 static void detach_destroy_domains(const struct cpumask *cpu_map)
9049 /* Save because hotplug lock held. */
9050 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9051 int i;
9053 for_each_cpu(i, cpu_map)
9054 cpu_attach_domain(NULL, &def_root_domain, i);
9055 synchronize_sched();
9056 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9059 /* handle null as "default" */
9060 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9061 struct sched_domain_attr *new, int idx_new)
9063 struct sched_domain_attr tmp;
9065 /* fast path */
9066 if (!new && !cur)
9067 return 1;
9069 tmp = SD_ATTR_INIT;
9070 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9071 new ? (new + idx_new) : &tmp,
9072 sizeof(struct sched_domain_attr));
9076 * Partition sched domains as specified by the 'ndoms_new'
9077 * cpumasks in the array doms_new[] of cpumasks. This compares
9078 * doms_new[] to the current sched domain partitioning, doms_cur[].
9079 * It destroys each deleted domain and builds each new domain.
9081 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
9082 * The masks don't intersect (don't overlap.) We should setup one
9083 * sched domain for each mask. CPUs not in any of the cpumasks will
9084 * not be load balanced. If the same cpumask appears both in the
9085 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9086 * it as it is.
9088 * The passed in 'doms_new' should be allocated using
9089 * alloc_sched_domains. This routine takes ownership of it and will
9090 * free_sched_domains it when done with it. If the caller failed the
9091 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9092 * and partition_sched_domains() will fallback to the single partition
9093 * 'fallback_doms', it also forces the domains to be rebuilt.
9095 * If doms_new == NULL it will be replaced with cpu_online_mask.
9096 * ndoms_new == 0 is a special case for destroying existing domains,
9097 * and it will not create the default domain.
9099 * Call with hotplug lock held
9101 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
9102 struct sched_domain_attr *dattr_new)
9104 int i, j, n;
9105 int new_topology;
9107 mutex_lock(&sched_domains_mutex);
9109 /* always unregister in case we don't destroy any domains */
9110 unregister_sched_domain_sysctl();
9112 /* Let architecture update cpu core mappings. */
9113 new_topology = arch_update_cpu_topology();
9115 n = doms_new ? ndoms_new : 0;
9117 /* Destroy deleted domains */
9118 for (i = 0; i < ndoms_cur; i++) {
9119 for (j = 0; j < n && !new_topology; j++) {
9120 if (cpumask_equal(doms_cur[i], doms_new[j])
9121 && dattrs_equal(dattr_cur, i, dattr_new, j))
9122 goto match1;
9124 /* no match - a current sched domain not in new doms_new[] */
9125 detach_destroy_domains(doms_cur[i]);
9126 match1:
9130 if (doms_new == NULL) {
9131 ndoms_cur = 0;
9132 doms_new = &fallback_doms;
9133 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
9134 WARN_ON_ONCE(dattr_new);
9137 /* Build new domains */
9138 for (i = 0; i < ndoms_new; i++) {
9139 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9140 if (cpumask_equal(doms_new[i], doms_cur[j])
9141 && dattrs_equal(dattr_new, i, dattr_cur, j))
9142 goto match2;
9144 /* no match - add a new doms_new */
9145 __build_sched_domains(doms_new[i],
9146 dattr_new ? dattr_new + i : NULL);
9147 match2:
9151 /* Remember the new sched domains */
9152 if (doms_cur != &fallback_doms)
9153 free_sched_domains(doms_cur, ndoms_cur);
9154 kfree(dattr_cur); /* kfree(NULL) is safe */
9155 doms_cur = doms_new;
9156 dattr_cur = dattr_new;
9157 ndoms_cur = ndoms_new;
9159 register_sched_domain_sysctl();
9161 mutex_unlock(&sched_domains_mutex);
9164 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9165 static void arch_reinit_sched_domains(void)
9167 get_online_cpus();
9169 /* Destroy domains first to force the rebuild */
9170 partition_sched_domains(0, NULL, NULL);
9172 rebuild_sched_domains();
9173 put_online_cpus();
9176 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9178 unsigned int level = 0;
9180 if (sscanf(buf, "%u", &level) != 1)
9181 return -EINVAL;
9184 * level is always be positive so don't check for
9185 * level < POWERSAVINGS_BALANCE_NONE which is 0
9186 * What happens on 0 or 1 byte write,
9187 * need to check for count as well?
9190 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9191 return -EINVAL;
9193 if (smt)
9194 sched_smt_power_savings = level;
9195 else
9196 sched_mc_power_savings = level;
9198 arch_reinit_sched_domains();
9200 return count;
9203 #ifdef CONFIG_SCHED_MC
9204 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9205 char *page)
9207 return sprintf(page, "%u\n", sched_mc_power_savings);
9209 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9210 const char *buf, size_t count)
9212 return sched_power_savings_store(buf, count, 0);
9214 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9215 sched_mc_power_savings_show,
9216 sched_mc_power_savings_store);
9217 #endif
9219 #ifdef CONFIG_SCHED_SMT
9220 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9221 char *page)
9223 return sprintf(page, "%u\n", sched_smt_power_savings);
9225 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9226 const char *buf, size_t count)
9228 return sched_power_savings_store(buf, count, 1);
9230 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9231 sched_smt_power_savings_show,
9232 sched_smt_power_savings_store);
9233 #endif
9235 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9237 int err = 0;
9239 #ifdef CONFIG_SCHED_SMT
9240 if (smt_capable())
9241 err = sysfs_create_file(&cls->kset.kobj,
9242 &attr_sched_smt_power_savings.attr);
9243 #endif
9244 #ifdef CONFIG_SCHED_MC
9245 if (!err && mc_capable())
9246 err = sysfs_create_file(&cls->kset.kobj,
9247 &attr_sched_mc_power_savings.attr);
9248 #endif
9249 return err;
9251 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9253 #ifndef CONFIG_CPUSETS
9255 * Add online and remove offline CPUs from the scheduler domains.
9256 * When cpusets are enabled they take over this function.
9258 static int update_sched_domains(struct notifier_block *nfb,
9259 unsigned long action, void *hcpu)
9261 switch (action) {
9262 case CPU_ONLINE:
9263 case CPU_ONLINE_FROZEN:
9264 case CPU_DOWN_PREPARE:
9265 case CPU_DOWN_PREPARE_FROZEN:
9266 case CPU_DOWN_FAILED:
9267 case CPU_DOWN_FAILED_FROZEN:
9268 partition_sched_domains(1, NULL, NULL);
9269 return NOTIFY_OK;
9271 default:
9272 return NOTIFY_DONE;
9275 #endif
9277 static int update_runtime(struct notifier_block *nfb,
9278 unsigned long action, void *hcpu)
9280 int cpu = (int)(long)hcpu;
9282 switch (action) {
9283 case CPU_DOWN_PREPARE:
9284 case CPU_DOWN_PREPARE_FROZEN:
9285 disable_runtime(cpu_rq(cpu));
9286 return NOTIFY_OK;
9288 case CPU_DOWN_FAILED:
9289 case CPU_DOWN_FAILED_FROZEN:
9290 case CPU_ONLINE:
9291 case CPU_ONLINE_FROZEN:
9292 enable_runtime(cpu_rq(cpu));
9293 return NOTIFY_OK;
9295 default:
9296 return NOTIFY_DONE;
9300 void __init sched_init_smp(void)
9302 cpumask_var_t non_isolated_cpus;
9304 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9305 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9307 #if defined(CONFIG_NUMA)
9308 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9309 GFP_KERNEL);
9310 BUG_ON(sched_group_nodes_bycpu == NULL);
9311 #endif
9312 get_online_cpus();
9313 mutex_lock(&sched_domains_mutex);
9314 arch_init_sched_domains(cpu_active_mask);
9315 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9316 if (cpumask_empty(non_isolated_cpus))
9317 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9318 mutex_unlock(&sched_domains_mutex);
9319 put_online_cpus();
9321 #ifndef CONFIG_CPUSETS
9322 /* XXX: Theoretical race here - CPU may be hotplugged now */
9323 hotcpu_notifier(update_sched_domains, 0);
9324 #endif
9326 /* RT runtime code needs to handle some hotplug events */
9327 hotcpu_notifier(update_runtime, 0);
9329 init_hrtick();
9331 /* Move init over to a non-isolated CPU */
9332 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9333 BUG();
9334 sched_init_granularity();
9335 free_cpumask_var(non_isolated_cpus);
9337 init_sched_rt_class();
9339 #else
9340 void __init sched_init_smp(void)
9342 sched_init_granularity();
9344 #endif /* CONFIG_SMP */
9346 const_debug unsigned int sysctl_timer_migration = 1;
9348 int in_sched_functions(unsigned long addr)
9350 return in_lock_functions(addr) ||
9351 (addr >= (unsigned long)__sched_text_start
9352 && addr < (unsigned long)__sched_text_end);
9355 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9357 cfs_rq->tasks_timeline = RB_ROOT;
9358 INIT_LIST_HEAD(&cfs_rq->tasks);
9359 #ifdef CONFIG_FAIR_GROUP_SCHED
9360 cfs_rq->rq = rq;
9361 #endif
9362 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9365 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9367 struct rt_prio_array *array;
9368 int i;
9370 array = &rt_rq->active;
9371 for (i = 0; i < MAX_RT_PRIO; i++) {
9372 INIT_LIST_HEAD(array->queue + i);
9373 __clear_bit(i, array->bitmap);
9375 /* delimiter for bitsearch: */
9376 __set_bit(MAX_RT_PRIO, array->bitmap);
9378 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9379 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9380 #ifdef CONFIG_SMP
9381 rt_rq->highest_prio.next = MAX_RT_PRIO;
9382 #endif
9383 #endif
9384 #ifdef CONFIG_SMP
9385 rt_rq->rt_nr_migratory = 0;
9386 rt_rq->overloaded = 0;
9387 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
9388 #endif
9390 rt_rq->rt_time = 0;
9391 rt_rq->rt_throttled = 0;
9392 rt_rq->rt_runtime = 0;
9393 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
9395 #ifdef CONFIG_RT_GROUP_SCHED
9396 rt_rq->rt_nr_boosted = 0;
9397 rt_rq->rq = rq;
9398 #endif
9401 #ifdef CONFIG_FAIR_GROUP_SCHED
9402 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9403 struct sched_entity *se, int cpu, int add,
9404 struct sched_entity *parent)
9406 struct rq *rq = cpu_rq(cpu);
9407 tg->cfs_rq[cpu] = cfs_rq;
9408 init_cfs_rq(cfs_rq, rq);
9409 cfs_rq->tg = tg;
9410 if (add)
9411 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9413 tg->se[cpu] = se;
9414 /* se could be NULL for init_task_group */
9415 if (!se)
9416 return;
9418 if (!parent)
9419 se->cfs_rq = &rq->cfs;
9420 else
9421 se->cfs_rq = parent->my_q;
9423 se->my_q = cfs_rq;
9424 se->load.weight = tg->shares;
9425 se->load.inv_weight = 0;
9426 se->parent = parent;
9428 #endif
9430 #ifdef CONFIG_RT_GROUP_SCHED
9431 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9432 struct sched_rt_entity *rt_se, int cpu, int add,
9433 struct sched_rt_entity *parent)
9435 struct rq *rq = cpu_rq(cpu);
9437 tg->rt_rq[cpu] = rt_rq;
9438 init_rt_rq(rt_rq, rq);
9439 rt_rq->tg = tg;
9440 rt_rq->rt_se = rt_se;
9441 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9442 if (add)
9443 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9445 tg->rt_se[cpu] = rt_se;
9446 if (!rt_se)
9447 return;
9449 if (!parent)
9450 rt_se->rt_rq = &rq->rt;
9451 else
9452 rt_se->rt_rq = parent->my_q;
9454 rt_se->my_q = rt_rq;
9455 rt_se->parent = parent;
9456 INIT_LIST_HEAD(&rt_se->run_list);
9458 #endif
9460 void __init sched_init(void)
9462 int i, j;
9463 unsigned long alloc_size = 0, ptr;
9465 #ifdef CONFIG_FAIR_GROUP_SCHED
9466 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9467 #endif
9468 #ifdef CONFIG_RT_GROUP_SCHED
9469 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9470 #endif
9471 #ifdef CONFIG_USER_SCHED
9472 alloc_size *= 2;
9473 #endif
9474 #ifdef CONFIG_CPUMASK_OFFSTACK
9475 alloc_size += num_possible_cpus() * cpumask_size();
9476 #endif
9477 if (alloc_size) {
9478 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9480 #ifdef CONFIG_FAIR_GROUP_SCHED
9481 init_task_group.se = (struct sched_entity **)ptr;
9482 ptr += nr_cpu_ids * sizeof(void **);
9484 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9485 ptr += nr_cpu_ids * sizeof(void **);
9487 #ifdef CONFIG_USER_SCHED
9488 root_task_group.se = (struct sched_entity **)ptr;
9489 ptr += nr_cpu_ids * sizeof(void **);
9491 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9492 ptr += nr_cpu_ids * sizeof(void **);
9493 #endif /* CONFIG_USER_SCHED */
9494 #endif /* CONFIG_FAIR_GROUP_SCHED */
9495 #ifdef CONFIG_RT_GROUP_SCHED
9496 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9497 ptr += nr_cpu_ids * sizeof(void **);
9499 init_task_group.rt_rq = (struct rt_rq **)ptr;
9500 ptr += nr_cpu_ids * sizeof(void **);
9502 #ifdef CONFIG_USER_SCHED
9503 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9504 ptr += nr_cpu_ids * sizeof(void **);
9506 root_task_group.rt_rq = (struct rt_rq **)ptr;
9507 ptr += nr_cpu_ids * sizeof(void **);
9508 #endif /* CONFIG_USER_SCHED */
9509 #endif /* CONFIG_RT_GROUP_SCHED */
9510 #ifdef CONFIG_CPUMASK_OFFSTACK
9511 for_each_possible_cpu(i) {
9512 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9513 ptr += cpumask_size();
9515 #endif /* CONFIG_CPUMASK_OFFSTACK */
9518 #ifdef CONFIG_SMP
9519 init_defrootdomain();
9520 #endif
9522 init_rt_bandwidth(&def_rt_bandwidth,
9523 global_rt_period(), global_rt_runtime());
9525 #ifdef CONFIG_RT_GROUP_SCHED
9526 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9527 global_rt_period(), global_rt_runtime());
9528 #ifdef CONFIG_USER_SCHED
9529 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9530 global_rt_period(), RUNTIME_INF);
9531 #endif /* CONFIG_USER_SCHED */
9532 #endif /* CONFIG_RT_GROUP_SCHED */
9534 #ifdef CONFIG_GROUP_SCHED
9535 list_add(&init_task_group.list, &task_groups);
9536 INIT_LIST_HEAD(&init_task_group.children);
9538 #ifdef CONFIG_USER_SCHED
9539 INIT_LIST_HEAD(&root_task_group.children);
9540 init_task_group.parent = &root_task_group;
9541 list_add(&init_task_group.siblings, &root_task_group.children);
9542 #endif /* CONFIG_USER_SCHED */
9543 #endif /* CONFIG_GROUP_SCHED */
9545 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9546 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9547 __alignof__(unsigned long));
9548 #endif
9549 for_each_possible_cpu(i) {
9550 struct rq *rq;
9552 rq = cpu_rq(i);
9553 raw_spin_lock_init(&rq->lock);
9554 rq->nr_running = 0;
9555 rq->calc_load_active = 0;
9556 rq->calc_load_update = jiffies + LOAD_FREQ;
9557 init_cfs_rq(&rq->cfs, rq);
9558 init_rt_rq(&rq->rt, rq);
9559 #ifdef CONFIG_FAIR_GROUP_SCHED
9560 init_task_group.shares = init_task_group_load;
9561 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9562 #ifdef CONFIG_CGROUP_SCHED
9564 * How much cpu bandwidth does init_task_group get?
9566 * In case of task-groups formed thr' the cgroup filesystem, it
9567 * gets 100% of the cpu resources in the system. This overall
9568 * system cpu resource is divided among the tasks of
9569 * init_task_group and its child task-groups in a fair manner,
9570 * based on each entity's (task or task-group's) weight
9571 * (se->load.weight).
9573 * In other words, if init_task_group has 10 tasks of weight
9574 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9575 * then A0's share of the cpu resource is:
9577 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9579 * We achieve this by letting init_task_group's tasks sit
9580 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9582 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9583 #elif defined CONFIG_USER_SCHED
9584 root_task_group.shares = NICE_0_LOAD;
9585 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9587 * In case of task-groups formed thr' the user id of tasks,
9588 * init_task_group represents tasks belonging to root user.
9589 * Hence it forms a sibling of all subsequent groups formed.
9590 * In this case, init_task_group gets only a fraction of overall
9591 * system cpu resource, based on the weight assigned to root
9592 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9593 * by letting tasks of init_task_group sit in a separate cfs_rq
9594 * (init_tg_cfs_rq) and having one entity represent this group of
9595 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9597 init_tg_cfs_entry(&init_task_group,
9598 &per_cpu(init_tg_cfs_rq, i),
9599 &per_cpu(init_sched_entity, i), i, 1,
9600 root_task_group.se[i]);
9602 #endif
9603 #endif /* CONFIG_FAIR_GROUP_SCHED */
9605 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9606 #ifdef CONFIG_RT_GROUP_SCHED
9607 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9608 #ifdef CONFIG_CGROUP_SCHED
9609 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9610 #elif defined CONFIG_USER_SCHED
9611 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9612 init_tg_rt_entry(&init_task_group,
9613 &per_cpu(init_rt_rq_var, i),
9614 &per_cpu(init_sched_rt_entity, i), i, 1,
9615 root_task_group.rt_se[i]);
9616 #endif
9617 #endif
9619 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9620 rq->cpu_load[j] = 0;
9621 #ifdef CONFIG_SMP
9622 rq->sd = NULL;
9623 rq->rd = NULL;
9624 rq->post_schedule = 0;
9625 rq->active_balance = 0;
9626 rq->next_balance = jiffies;
9627 rq->push_cpu = 0;
9628 rq->cpu = i;
9629 rq->online = 0;
9630 rq->migration_thread = NULL;
9631 rq->idle_stamp = 0;
9632 rq->avg_idle = 2*sysctl_sched_migration_cost;
9633 INIT_LIST_HEAD(&rq->migration_queue);
9634 rq_attach_root(rq, &def_root_domain);
9635 #endif
9636 init_rq_hrtick(rq);
9637 atomic_set(&rq->nr_iowait, 0);
9640 set_load_weight(&init_task);
9642 #ifdef CONFIG_PREEMPT_NOTIFIERS
9643 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9644 #endif
9646 #ifdef CONFIG_SMP
9647 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9648 #endif
9650 #ifdef CONFIG_RT_MUTEXES
9651 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
9652 #endif
9655 * The boot idle thread does lazy MMU switching as well:
9657 atomic_inc(&init_mm.mm_count);
9658 enter_lazy_tlb(&init_mm, current);
9661 * Make us the idle thread. Technically, schedule() should not be
9662 * called from this thread, however somewhere below it might be,
9663 * but because we are the idle thread, we just pick up running again
9664 * when this runqueue becomes "idle".
9666 init_idle(current, smp_processor_id());
9668 calc_load_update = jiffies + LOAD_FREQ;
9671 * During early bootup we pretend to be a normal task:
9673 current->sched_class = &fair_sched_class;
9675 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9676 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9677 #ifdef CONFIG_SMP
9678 #ifdef CONFIG_NO_HZ
9679 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9680 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9681 #endif
9682 /* May be allocated at isolcpus cmdline parse time */
9683 if (cpu_isolated_map == NULL)
9684 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9685 #endif /* SMP */
9687 perf_event_init();
9689 scheduler_running = 1;
9692 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9693 static inline int preempt_count_equals(int preempt_offset)
9695 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
9697 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9700 void __might_sleep(char *file, int line, int preempt_offset)
9702 #ifdef in_atomic
9703 static unsigned long prev_jiffy; /* ratelimiting */
9705 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9706 system_state != SYSTEM_RUNNING || oops_in_progress)
9707 return;
9708 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9709 return;
9710 prev_jiffy = jiffies;
9712 printk(KERN_ERR
9713 "BUG: sleeping function called from invalid context at %s:%d\n",
9714 file, line);
9715 printk(KERN_ERR
9716 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9717 in_atomic(), irqs_disabled(),
9718 current->pid, current->comm);
9720 debug_show_held_locks(current);
9721 if (irqs_disabled())
9722 print_irqtrace_events(current);
9723 dump_stack();
9724 #endif
9726 EXPORT_SYMBOL(__might_sleep);
9727 #endif
9729 #ifdef CONFIG_MAGIC_SYSRQ
9730 static void normalize_task(struct rq *rq, struct task_struct *p)
9732 int on_rq;
9734 update_rq_clock(rq);
9735 on_rq = p->se.on_rq;
9736 if (on_rq)
9737 deactivate_task(rq, p, 0);
9738 __setscheduler(rq, p, SCHED_NORMAL, 0);
9739 if (on_rq) {
9740 activate_task(rq, p, 0);
9741 resched_task(rq->curr);
9745 void normalize_rt_tasks(void)
9747 struct task_struct *g, *p;
9748 unsigned long flags;
9749 struct rq *rq;
9751 read_lock_irqsave(&tasklist_lock, flags);
9752 do_each_thread(g, p) {
9754 * Only normalize user tasks:
9756 if (!p->mm)
9757 continue;
9759 p->se.exec_start = 0;
9760 #ifdef CONFIG_SCHEDSTATS
9761 p->se.wait_start = 0;
9762 p->se.sleep_start = 0;
9763 p->se.block_start = 0;
9764 #endif
9766 if (!rt_task(p)) {
9768 * Renice negative nice level userspace
9769 * tasks back to 0:
9771 if (TASK_NICE(p) < 0 && p->mm)
9772 set_user_nice(p, 0);
9773 continue;
9776 raw_spin_lock(&p->pi_lock);
9777 rq = __task_rq_lock(p);
9779 normalize_task(rq, p);
9781 __task_rq_unlock(rq);
9782 raw_spin_unlock(&p->pi_lock);
9783 } while_each_thread(g, p);
9785 read_unlock_irqrestore(&tasklist_lock, flags);
9788 #endif /* CONFIG_MAGIC_SYSRQ */
9790 #ifdef CONFIG_IA64
9792 * These functions are only useful for the IA64 MCA handling.
9794 * They can only be called when the whole system has been
9795 * stopped - every CPU needs to be quiescent, and no scheduling
9796 * activity can take place. Using them for anything else would
9797 * be a serious bug, and as a result, they aren't even visible
9798 * under any other configuration.
9802 * curr_task - return the current task for a given cpu.
9803 * @cpu: the processor in question.
9805 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9807 struct task_struct *curr_task(int cpu)
9809 return cpu_curr(cpu);
9813 * set_curr_task - set the current task for a given cpu.
9814 * @cpu: the processor in question.
9815 * @p: the task pointer to set.
9817 * Description: This function must only be used when non-maskable interrupts
9818 * are serviced on a separate stack. It allows the architecture to switch the
9819 * notion of the current task on a cpu in a non-blocking manner. This function
9820 * must be called with all CPU's synchronized, and interrupts disabled, the
9821 * and caller must save the original value of the current task (see
9822 * curr_task() above) and restore that value before reenabling interrupts and
9823 * re-starting the system.
9825 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9827 void set_curr_task(int cpu, struct task_struct *p)
9829 cpu_curr(cpu) = p;
9832 #endif
9834 #ifdef CONFIG_FAIR_GROUP_SCHED
9835 static void free_fair_sched_group(struct task_group *tg)
9837 int i;
9839 for_each_possible_cpu(i) {
9840 if (tg->cfs_rq)
9841 kfree(tg->cfs_rq[i]);
9842 if (tg->se)
9843 kfree(tg->se[i]);
9846 kfree(tg->cfs_rq);
9847 kfree(tg->se);
9850 static
9851 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9853 struct cfs_rq *cfs_rq;
9854 struct sched_entity *se;
9855 struct rq *rq;
9856 int i;
9858 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9859 if (!tg->cfs_rq)
9860 goto err;
9861 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9862 if (!tg->se)
9863 goto err;
9865 tg->shares = NICE_0_LOAD;
9867 for_each_possible_cpu(i) {
9868 rq = cpu_rq(i);
9870 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9871 GFP_KERNEL, cpu_to_node(i));
9872 if (!cfs_rq)
9873 goto err;
9875 se = kzalloc_node(sizeof(struct sched_entity),
9876 GFP_KERNEL, cpu_to_node(i));
9877 if (!se)
9878 goto err_free_rq;
9880 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9883 return 1;
9885 err_free_rq:
9886 kfree(cfs_rq);
9887 err:
9888 return 0;
9891 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9893 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9894 &cpu_rq(cpu)->leaf_cfs_rq_list);
9897 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9899 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9901 #else /* !CONFG_FAIR_GROUP_SCHED */
9902 static inline void free_fair_sched_group(struct task_group *tg)
9906 static inline
9907 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9909 return 1;
9912 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9916 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9919 #endif /* CONFIG_FAIR_GROUP_SCHED */
9921 #ifdef CONFIG_RT_GROUP_SCHED
9922 static void free_rt_sched_group(struct task_group *tg)
9924 int i;
9926 destroy_rt_bandwidth(&tg->rt_bandwidth);
9928 for_each_possible_cpu(i) {
9929 if (tg->rt_rq)
9930 kfree(tg->rt_rq[i]);
9931 if (tg->rt_se)
9932 kfree(tg->rt_se[i]);
9935 kfree(tg->rt_rq);
9936 kfree(tg->rt_se);
9939 static
9940 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9942 struct rt_rq *rt_rq;
9943 struct sched_rt_entity *rt_se;
9944 struct rq *rq;
9945 int i;
9947 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9948 if (!tg->rt_rq)
9949 goto err;
9950 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9951 if (!tg->rt_se)
9952 goto err;
9954 init_rt_bandwidth(&tg->rt_bandwidth,
9955 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9957 for_each_possible_cpu(i) {
9958 rq = cpu_rq(i);
9960 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9961 GFP_KERNEL, cpu_to_node(i));
9962 if (!rt_rq)
9963 goto err;
9965 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9966 GFP_KERNEL, cpu_to_node(i));
9967 if (!rt_se)
9968 goto err_free_rq;
9970 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9973 return 1;
9975 err_free_rq:
9976 kfree(rt_rq);
9977 err:
9978 return 0;
9981 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9983 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9984 &cpu_rq(cpu)->leaf_rt_rq_list);
9987 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9989 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9991 #else /* !CONFIG_RT_GROUP_SCHED */
9992 static inline void free_rt_sched_group(struct task_group *tg)
9996 static inline
9997 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9999 return 1;
10002 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10006 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10009 #endif /* CONFIG_RT_GROUP_SCHED */
10011 #ifdef CONFIG_GROUP_SCHED
10012 static void free_sched_group(struct task_group *tg)
10014 free_fair_sched_group(tg);
10015 free_rt_sched_group(tg);
10016 kfree(tg);
10019 /* allocate runqueue etc for a new task group */
10020 struct task_group *sched_create_group(struct task_group *parent)
10022 struct task_group *tg;
10023 unsigned long flags;
10024 int i;
10026 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
10027 if (!tg)
10028 return ERR_PTR(-ENOMEM);
10030 if (!alloc_fair_sched_group(tg, parent))
10031 goto err;
10033 if (!alloc_rt_sched_group(tg, parent))
10034 goto err;
10036 spin_lock_irqsave(&task_group_lock, flags);
10037 for_each_possible_cpu(i) {
10038 register_fair_sched_group(tg, i);
10039 register_rt_sched_group(tg, i);
10041 list_add_rcu(&tg->list, &task_groups);
10043 WARN_ON(!parent); /* root should already exist */
10045 tg->parent = parent;
10046 INIT_LIST_HEAD(&tg->children);
10047 list_add_rcu(&tg->siblings, &parent->children);
10048 spin_unlock_irqrestore(&task_group_lock, flags);
10050 return tg;
10052 err:
10053 free_sched_group(tg);
10054 return ERR_PTR(-ENOMEM);
10057 /* rcu callback to free various structures associated with a task group */
10058 static void free_sched_group_rcu(struct rcu_head *rhp)
10060 /* now it should be safe to free those cfs_rqs */
10061 free_sched_group(container_of(rhp, struct task_group, rcu));
10064 /* Destroy runqueue etc associated with a task group */
10065 void sched_destroy_group(struct task_group *tg)
10067 unsigned long flags;
10068 int i;
10070 spin_lock_irqsave(&task_group_lock, flags);
10071 for_each_possible_cpu(i) {
10072 unregister_fair_sched_group(tg, i);
10073 unregister_rt_sched_group(tg, i);
10075 list_del_rcu(&tg->list);
10076 list_del_rcu(&tg->siblings);
10077 spin_unlock_irqrestore(&task_group_lock, flags);
10079 /* wait for possible concurrent references to cfs_rqs complete */
10080 call_rcu(&tg->rcu, free_sched_group_rcu);
10083 /* change task's runqueue when it moves between groups.
10084 * The caller of this function should have put the task in its new group
10085 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10086 * reflect its new group.
10088 void sched_move_task(struct task_struct *tsk)
10090 int on_rq, running;
10091 unsigned long flags;
10092 struct rq *rq;
10094 rq = task_rq_lock(tsk, &flags);
10096 update_rq_clock(rq);
10098 running = task_current(rq, tsk);
10099 on_rq = tsk->se.on_rq;
10101 if (on_rq)
10102 dequeue_task(rq, tsk, 0);
10103 if (unlikely(running))
10104 tsk->sched_class->put_prev_task(rq, tsk);
10106 set_task_rq(tsk, task_cpu(tsk));
10108 #ifdef CONFIG_FAIR_GROUP_SCHED
10109 if (tsk->sched_class->moved_group)
10110 tsk->sched_class->moved_group(tsk, on_rq);
10111 #endif
10113 if (unlikely(running))
10114 tsk->sched_class->set_curr_task(rq);
10115 if (on_rq)
10116 enqueue_task(rq, tsk, 0);
10118 task_rq_unlock(rq, &flags);
10120 #endif /* CONFIG_GROUP_SCHED */
10122 #ifdef CONFIG_FAIR_GROUP_SCHED
10123 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10125 struct cfs_rq *cfs_rq = se->cfs_rq;
10126 int on_rq;
10128 on_rq = se->on_rq;
10129 if (on_rq)
10130 dequeue_entity(cfs_rq, se, 0);
10132 se->load.weight = shares;
10133 se->load.inv_weight = 0;
10135 if (on_rq)
10136 enqueue_entity(cfs_rq, se, 0);
10139 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10141 struct cfs_rq *cfs_rq = se->cfs_rq;
10142 struct rq *rq = cfs_rq->rq;
10143 unsigned long flags;
10145 raw_spin_lock_irqsave(&rq->lock, flags);
10146 __set_se_shares(se, shares);
10147 raw_spin_unlock_irqrestore(&rq->lock, flags);
10150 static DEFINE_MUTEX(shares_mutex);
10152 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10154 int i;
10155 unsigned long flags;
10158 * We can't change the weight of the root cgroup.
10160 if (!tg->se[0])
10161 return -EINVAL;
10163 if (shares < MIN_SHARES)
10164 shares = MIN_SHARES;
10165 else if (shares > MAX_SHARES)
10166 shares = MAX_SHARES;
10168 mutex_lock(&shares_mutex);
10169 if (tg->shares == shares)
10170 goto done;
10172 spin_lock_irqsave(&task_group_lock, flags);
10173 for_each_possible_cpu(i)
10174 unregister_fair_sched_group(tg, i);
10175 list_del_rcu(&tg->siblings);
10176 spin_unlock_irqrestore(&task_group_lock, flags);
10178 /* wait for any ongoing reference to this group to finish */
10179 synchronize_sched();
10182 * Now we are free to modify the group's share on each cpu
10183 * w/o tripping rebalance_share or load_balance_fair.
10185 tg->shares = shares;
10186 for_each_possible_cpu(i) {
10188 * force a rebalance
10190 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10191 set_se_shares(tg->se[i], shares);
10195 * Enable load balance activity on this group, by inserting it back on
10196 * each cpu's rq->leaf_cfs_rq_list.
10198 spin_lock_irqsave(&task_group_lock, flags);
10199 for_each_possible_cpu(i)
10200 register_fair_sched_group(tg, i);
10201 list_add_rcu(&tg->siblings, &tg->parent->children);
10202 spin_unlock_irqrestore(&task_group_lock, flags);
10203 done:
10204 mutex_unlock(&shares_mutex);
10205 return 0;
10208 unsigned long sched_group_shares(struct task_group *tg)
10210 return tg->shares;
10212 #endif
10214 #ifdef CONFIG_RT_GROUP_SCHED
10216 * Ensure that the real time constraints are schedulable.
10218 static DEFINE_MUTEX(rt_constraints_mutex);
10220 static unsigned long to_ratio(u64 period, u64 runtime)
10222 if (runtime == RUNTIME_INF)
10223 return 1ULL << 20;
10225 return div64_u64(runtime << 20, period);
10228 /* Must be called with tasklist_lock held */
10229 static inline int tg_has_rt_tasks(struct task_group *tg)
10231 struct task_struct *g, *p;
10233 do_each_thread(g, p) {
10234 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10235 return 1;
10236 } while_each_thread(g, p);
10238 return 0;
10241 struct rt_schedulable_data {
10242 struct task_group *tg;
10243 u64 rt_period;
10244 u64 rt_runtime;
10247 static int tg_schedulable(struct task_group *tg, void *data)
10249 struct rt_schedulable_data *d = data;
10250 struct task_group *child;
10251 unsigned long total, sum = 0;
10252 u64 period, runtime;
10254 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10255 runtime = tg->rt_bandwidth.rt_runtime;
10257 if (tg == d->tg) {
10258 period = d->rt_period;
10259 runtime = d->rt_runtime;
10262 #ifdef CONFIG_USER_SCHED
10263 if (tg == &root_task_group) {
10264 period = global_rt_period();
10265 runtime = global_rt_runtime();
10267 #endif
10270 * Cannot have more runtime than the period.
10272 if (runtime > period && runtime != RUNTIME_INF)
10273 return -EINVAL;
10276 * Ensure we don't starve existing RT tasks.
10278 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10279 return -EBUSY;
10281 total = to_ratio(period, runtime);
10284 * Nobody can have more than the global setting allows.
10286 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10287 return -EINVAL;
10290 * The sum of our children's runtime should not exceed our own.
10292 list_for_each_entry_rcu(child, &tg->children, siblings) {
10293 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10294 runtime = child->rt_bandwidth.rt_runtime;
10296 if (child == d->tg) {
10297 period = d->rt_period;
10298 runtime = d->rt_runtime;
10301 sum += to_ratio(period, runtime);
10304 if (sum > total)
10305 return -EINVAL;
10307 return 0;
10310 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10312 struct rt_schedulable_data data = {
10313 .tg = tg,
10314 .rt_period = period,
10315 .rt_runtime = runtime,
10318 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10321 static int tg_set_bandwidth(struct task_group *tg,
10322 u64 rt_period, u64 rt_runtime)
10324 int i, err = 0;
10326 mutex_lock(&rt_constraints_mutex);
10327 read_lock(&tasklist_lock);
10328 err = __rt_schedulable(tg, rt_period, rt_runtime);
10329 if (err)
10330 goto unlock;
10332 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10333 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10334 tg->rt_bandwidth.rt_runtime = rt_runtime;
10336 for_each_possible_cpu(i) {
10337 struct rt_rq *rt_rq = tg->rt_rq[i];
10339 raw_spin_lock(&rt_rq->rt_runtime_lock);
10340 rt_rq->rt_runtime = rt_runtime;
10341 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10343 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10344 unlock:
10345 read_unlock(&tasklist_lock);
10346 mutex_unlock(&rt_constraints_mutex);
10348 return err;
10351 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10353 u64 rt_runtime, rt_period;
10355 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10356 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10357 if (rt_runtime_us < 0)
10358 rt_runtime = RUNTIME_INF;
10360 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10363 long sched_group_rt_runtime(struct task_group *tg)
10365 u64 rt_runtime_us;
10367 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10368 return -1;
10370 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10371 do_div(rt_runtime_us, NSEC_PER_USEC);
10372 return rt_runtime_us;
10375 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10377 u64 rt_runtime, rt_period;
10379 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10380 rt_runtime = tg->rt_bandwidth.rt_runtime;
10382 if (rt_period == 0)
10383 return -EINVAL;
10385 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10388 long sched_group_rt_period(struct task_group *tg)
10390 u64 rt_period_us;
10392 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10393 do_div(rt_period_us, NSEC_PER_USEC);
10394 return rt_period_us;
10397 static int sched_rt_global_constraints(void)
10399 u64 runtime, period;
10400 int ret = 0;
10402 if (sysctl_sched_rt_period <= 0)
10403 return -EINVAL;
10405 runtime = global_rt_runtime();
10406 period = global_rt_period();
10409 * Sanity check on the sysctl variables.
10411 if (runtime > period && runtime != RUNTIME_INF)
10412 return -EINVAL;
10414 mutex_lock(&rt_constraints_mutex);
10415 read_lock(&tasklist_lock);
10416 ret = __rt_schedulable(NULL, 0, 0);
10417 read_unlock(&tasklist_lock);
10418 mutex_unlock(&rt_constraints_mutex);
10420 return ret;
10423 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10425 /* Don't accept realtime tasks when there is no way for them to run */
10426 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10427 return 0;
10429 return 1;
10432 #else /* !CONFIG_RT_GROUP_SCHED */
10433 static int sched_rt_global_constraints(void)
10435 unsigned long flags;
10436 int i;
10438 if (sysctl_sched_rt_period <= 0)
10439 return -EINVAL;
10442 * There's always some RT tasks in the root group
10443 * -- migration, kstopmachine etc..
10445 if (sysctl_sched_rt_runtime == 0)
10446 return -EBUSY;
10448 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10449 for_each_possible_cpu(i) {
10450 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10452 raw_spin_lock(&rt_rq->rt_runtime_lock);
10453 rt_rq->rt_runtime = global_rt_runtime();
10454 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10456 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10458 return 0;
10460 #endif /* CONFIG_RT_GROUP_SCHED */
10462 int sched_rt_handler(struct ctl_table *table, int write,
10463 void __user *buffer, size_t *lenp,
10464 loff_t *ppos)
10466 int ret;
10467 int old_period, old_runtime;
10468 static DEFINE_MUTEX(mutex);
10470 mutex_lock(&mutex);
10471 old_period = sysctl_sched_rt_period;
10472 old_runtime = sysctl_sched_rt_runtime;
10474 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10476 if (!ret && write) {
10477 ret = sched_rt_global_constraints();
10478 if (ret) {
10479 sysctl_sched_rt_period = old_period;
10480 sysctl_sched_rt_runtime = old_runtime;
10481 } else {
10482 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10483 def_rt_bandwidth.rt_period =
10484 ns_to_ktime(global_rt_period());
10487 mutex_unlock(&mutex);
10489 return ret;
10492 #ifdef CONFIG_CGROUP_SCHED
10494 /* return corresponding task_group object of a cgroup */
10495 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10497 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10498 struct task_group, css);
10501 static struct cgroup_subsys_state *
10502 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10504 struct task_group *tg, *parent;
10506 if (!cgrp->parent) {
10507 /* This is early initialization for the top cgroup */
10508 return &init_task_group.css;
10511 parent = cgroup_tg(cgrp->parent);
10512 tg = sched_create_group(parent);
10513 if (IS_ERR(tg))
10514 return ERR_PTR(-ENOMEM);
10516 return &tg->css;
10519 static void
10520 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10522 struct task_group *tg = cgroup_tg(cgrp);
10524 sched_destroy_group(tg);
10527 static int
10528 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10530 #ifdef CONFIG_RT_GROUP_SCHED
10531 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10532 return -EINVAL;
10533 #else
10534 /* We don't support RT-tasks being in separate groups */
10535 if (tsk->sched_class != &fair_sched_class)
10536 return -EINVAL;
10537 #endif
10538 return 0;
10541 static int
10542 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10543 struct task_struct *tsk, bool threadgroup)
10545 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10546 if (retval)
10547 return retval;
10548 if (threadgroup) {
10549 struct task_struct *c;
10550 rcu_read_lock();
10551 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10552 retval = cpu_cgroup_can_attach_task(cgrp, c);
10553 if (retval) {
10554 rcu_read_unlock();
10555 return retval;
10558 rcu_read_unlock();
10560 return 0;
10563 static void
10564 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10565 struct cgroup *old_cont, struct task_struct *tsk,
10566 bool threadgroup)
10568 sched_move_task(tsk);
10569 if (threadgroup) {
10570 struct task_struct *c;
10571 rcu_read_lock();
10572 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10573 sched_move_task(c);
10575 rcu_read_unlock();
10579 #ifdef CONFIG_FAIR_GROUP_SCHED
10580 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10581 u64 shareval)
10583 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10586 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10588 struct task_group *tg = cgroup_tg(cgrp);
10590 return (u64) tg->shares;
10592 #endif /* CONFIG_FAIR_GROUP_SCHED */
10594 #ifdef CONFIG_RT_GROUP_SCHED
10595 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10596 s64 val)
10598 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10601 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10603 return sched_group_rt_runtime(cgroup_tg(cgrp));
10606 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10607 u64 rt_period_us)
10609 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10612 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10614 return sched_group_rt_period(cgroup_tg(cgrp));
10616 #endif /* CONFIG_RT_GROUP_SCHED */
10618 static struct cftype cpu_files[] = {
10619 #ifdef CONFIG_FAIR_GROUP_SCHED
10621 .name = "shares",
10622 .read_u64 = cpu_shares_read_u64,
10623 .write_u64 = cpu_shares_write_u64,
10625 #endif
10626 #ifdef CONFIG_RT_GROUP_SCHED
10628 .name = "rt_runtime_us",
10629 .read_s64 = cpu_rt_runtime_read,
10630 .write_s64 = cpu_rt_runtime_write,
10633 .name = "rt_period_us",
10634 .read_u64 = cpu_rt_period_read_uint,
10635 .write_u64 = cpu_rt_period_write_uint,
10637 #endif
10640 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10642 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10645 struct cgroup_subsys cpu_cgroup_subsys = {
10646 .name = "cpu",
10647 .create = cpu_cgroup_create,
10648 .destroy = cpu_cgroup_destroy,
10649 .can_attach = cpu_cgroup_can_attach,
10650 .attach = cpu_cgroup_attach,
10651 .populate = cpu_cgroup_populate,
10652 .subsys_id = cpu_cgroup_subsys_id,
10653 .early_init = 1,
10656 #endif /* CONFIG_CGROUP_SCHED */
10658 #ifdef CONFIG_CGROUP_CPUACCT
10661 * CPU accounting code for task groups.
10663 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10664 * (balbir@in.ibm.com).
10667 /* track cpu usage of a group of tasks and its child groups */
10668 struct cpuacct {
10669 struct cgroup_subsys_state css;
10670 /* cpuusage holds pointer to a u64-type object on every cpu */
10671 u64 *cpuusage;
10672 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10673 struct cpuacct *parent;
10676 struct cgroup_subsys cpuacct_subsys;
10678 /* return cpu accounting group corresponding to this container */
10679 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10681 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10682 struct cpuacct, css);
10685 /* return cpu accounting group to which this task belongs */
10686 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10688 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10689 struct cpuacct, css);
10692 /* create a new cpu accounting group */
10693 static struct cgroup_subsys_state *cpuacct_create(
10694 struct cgroup_subsys *ss, struct cgroup *cgrp)
10696 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10697 int i;
10699 if (!ca)
10700 goto out;
10702 ca->cpuusage = alloc_percpu(u64);
10703 if (!ca->cpuusage)
10704 goto out_free_ca;
10706 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10707 if (percpu_counter_init(&ca->cpustat[i], 0))
10708 goto out_free_counters;
10710 if (cgrp->parent)
10711 ca->parent = cgroup_ca(cgrp->parent);
10713 return &ca->css;
10715 out_free_counters:
10716 while (--i >= 0)
10717 percpu_counter_destroy(&ca->cpustat[i]);
10718 free_percpu(ca->cpuusage);
10719 out_free_ca:
10720 kfree(ca);
10721 out:
10722 return ERR_PTR(-ENOMEM);
10725 /* destroy an existing cpu accounting group */
10726 static void
10727 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10729 struct cpuacct *ca = cgroup_ca(cgrp);
10730 int i;
10732 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10733 percpu_counter_destroy(&ca->cpustat[i]);
10734 free_percpu(ca->cpuusage);
10735 kfree(ca);
10738 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10740 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10741 u64 data;
10743 #ifndef CONFIG_64BIT
10745 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10747 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10748 data = *cpuusage;
10749 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10750 #else
10751 data = *cpuusage;
10752 #endif
10754 return data;
10757 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10759 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10761 #ifndef CONFIG_64BIT
10763 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10765 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10766 *cpuusage = val;
10767 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10768 #else
10769 *cpuusage = val;
10770 #endif
10773 /* return total cpu usage (in nanoseconds) of a group */
10774 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10776 struct cpuacct *ca = cgroup_ca(cgrp);
10777 u64 totalcpuusage = 0;
10778 int i;
10780 for_each_present_cpu(i)
10781 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10783 return totalcpuusage;
10786 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10787 u64 reset)
10789 struct cpuacct *ca = cgroup_ca(cgrp);
10790 int err = 0;
10791 int i;
10793 if (reset) {
10794 err = -EINVAL;
10795 goto out;
10798 for_each_present_cpu(i)
10799 cpuacct_cpuusage_write(ca, i, 0);
10801 out:
10802 return err;
10805 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10806 struct seq_file *m)
10808 struct cpuacct *ca = cgroup_ca(cgroup);
10809 u64 percpu;
10810 int i;
10812 for_each_present_cpu(i) {
10813 percpu = cpuacct_cpuusage_read(ca, i);
10814 seq_printf(m, "%llu ", (unsigned long long) percpu);
10816 seq_printf(m, "\n");
10817 return 0;
10820 static const char *cpuacct_stat_desc[] = {
10821 [CPUACCT_STAT_USER] = "user",
10822 [CPUACCT_STAT_SYSTEM] = "system",
10825 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10826 struct cgroup_map_cb *cb)
10828 struct cpuacct *ca = cgroup_ca(cgrp);
10829 int i;
10831 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10832 s64 val = percpu_counter_read(&ca->cpustat[i]);
10833 val = cputime64_to_clock_t(val);
10834 cb->fill(cb, cpuacct_stat_desc[i], val);
10836 return 0;
10839 static struct cftype files[] = {
10841 .name = "usage",
10842 .read_u64 = cpuusage_read,
10843 .write_u64 = cpuusage_write,
10846 .name = "usage_percpu",
10847 .read_seq_string = cpuacct_percpu_seq_read,
10850 .name = "stat",
10851 .read_map = cpuacct_stats_show,
10855 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10857 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10861 * charge this task's execution time to its accounting group.
10863 * called with rq->lock held.
10865 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10867 struct cpuacct *ca;
10868 int cpu;
10870 if (unlikely(!cpuacct_subsys.active))
10871 return;
10873 cpu = task_cpu(tsk);
10875 rcu_read_lock();
10877 ca = task_ca(tsk);
10879 for (; ca; ca = ca->parent) {
10880 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10881 *cpuusage += cputime;
10884 rcu_read_unlock();
10888 * Charge the system/user time to the task's accounting group.
10890 static void cpuacct_update_stats(struct task_struct *tsk,
10891 enum cpuacct_stat_index idx, cputime_t val)
10893 struct cpuacct *ca;
10895 if (unlikely(!cpuacct_subsys.active))
10896 return;
10898 rcu_read_lock();
10899 ca = task_ca(tsk);
10901 do {
10902 percpu_counter_add(&ca->cpustat[idx], val);
10903 ca = ca->parent;
10904 } while (ca);
10905 rcu_read_unlock();
10908 struct cgroup_subsys cpuacct_subsys = {
10909 .name = "cpuacct",
10910 .create = cpuacct_create,
10911 .destroy = cpuacct_destroy,
10912 .populate = cpuacct_populate,
10913 .subsys_id = cpuacct_subsys_id,
10915 #endif /* CONFIG_CGROUP_CPUACCT */
10917 #ifndef CONFIG_SMP
10919 int rcu_expedited_torture_stats(char *page)
10921 return 0;
10923 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10925 void synchronize_sched_expedited(void)
10928 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10930 #else /* #ifndef CONFIG_SMP */
10932 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10933 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10935 #define RCU_EXPEDITED_STATE_POST -2
10936 #define RCU_EXPEDITED_STATE_IDLE -1
10938 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10940 int rcu_expedited_torture_stats(char *page)
10942 int cnt = 0;
10943 int cpu;
10945 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10946 for_each_online_cpu(cpu) {
10947 cnt += sprintf(&page[cnt], " %d:%d",
10948 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10950 cnt += sprintf(&page[cnt], "\n");
10951 return cnt;
10953 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10955 static long synchronize_sched_expedited_count;
10958 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10959 * approach to force grace period to end quickly. This consumes
10960 * significant time on all CPUs, and is thus not recommended for
10961 * any sort of common-case code.
10963 * Note that it is illegal to call this function while holding any
10964 * lock that is acquired by a CPU-hotplug notifier. Failing to
10965 * observe this restriction will result in deadlock.
10967 void synchronize_sched_expedited(void)
10969 int cpu;
10970 unsigned long flags;
10971 bool need_full_sync = 0;
10972 struct rq *rq;
10973 struct migration_req *req;
10974 long snap;
10975 int trycount = 0;
10977 smp_mb(); /* ensure prior mod happens before capturing snap. */
10978 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10979 get_online_cpus();
10980 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10981 put_online_cpus();
10982 if (trycount++ < 10)
10983 udelay(trycount * num_online_cpus());
10984 else {
10985 synchronize_sched();
10986 return;
10988 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10989 smp_mb(); /* ensure test happens before caller kfree */
10990 return;
10992 get_online_cpus();
10994 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10995 for_each_online_cpu(cpu) {
10996 rq = cpu_rq(cpu);
10997 req = &per_cpu(rcu_migration_req, cpu);
10998 init_completion(&req->done);
10999 req->task = NULL;
11000 req->dest_cpu = RCU_MIGRATION_NEED_QS;
11001 raw_spin_lock_irqsave(&rq->lock, flags);
11002 list_add(&req->list, &rq->migration_queue);
11003 raw_spin_unlock_irqrestore(&rq->lock, flags);
11004 wake_up_process(rq->migration_thread);
11006 for_each_online_cpu(cpu) {
11007 rcu_expedited_state = cpu;
11008 req = &per_cpu(rcu_migration_req, cpu);
11009 rq = cpu_rq(cpu);
11010 wait_for_completion(&req->done);
11011 raw_spin_lock_irqsave(&rq->lock, flags);
11012 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
11013 need_full_sync = 1;
11014 req->dest_cpu = RCU_MIGRATION_IDLE;
11015 raw_spin_unlock_irqrestore(&rq->lock, flags);
11017 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
11018 synchronize_sched_expedited_count++;
11019 mutex_unlock(&rcu_sched_expedited_mutex);
11020 put_online_cpus();
11021 if (need_full_sync)
11022 synchronize_sched();
11024 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
11026 #endif /* #else #ifndef CONFIG_SMP */