cxgb4vf: do vlan cleanup
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / kernel / sched.c
blob3dc716f6d8adfb777ba42dfd92abf00dbafc5988
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 <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/stop_machine.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
75 #include <asm/tlb.h>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
79 #include "sched_cpupri.h"
80 #include "workqueue_sched.h"
81 #include "sched_autogroup.h"
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/sched.h>
87 * Convert user-nice values [ -20 ... 0 ... 19 ]
88 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 * and back.
91 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
92 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
93 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
96 * 'User priority' is the nice value converted to something we
97 * can work with better when scaling various scheduler parameters,
98 * it's a [ 0 ... 39 ] range.
100 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
101 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
102 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
105 * Helpers for converting nanosecond timing to jiffy resolution
107 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
109 #define NICE_0_LOAD SCHED_LOAD_SCALE
110 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
113 * These are the 'tuning knobs' of the scheduler:
115 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
116 * Timeslices get refilled after they expire.
118 #define DEF_TIMESLICE (100 * HZ / 1000)
121 * single value that denotes runtime == period, ie unlimited time.
123 #define RUNTIME_INF ((u64)~0ULL)
125 static inline int rt_policy(int policy)
127 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
128 return 1;
129 return 0;
132 static inline int task_has_rt_policy(struct task_struct *p)
134 return rt_policy(p->policy);
138 * This is the priority-queue data structure of the RT scheduling class:
140 struct rt_prio_array {
141 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
142 struct list_head queue[MAX_RT_PRIO];
145 struct rt_bandwidth {
146 /* nests inside the rq lock: */
147 raw_spinlock_t rt_runtime_lock;
148 ktime_t rt_period;
149 u64 rt_runtime;
150 struct hrtimer rt_period_timer;
153 static struct rt_bandwidth def_rt_bandwidth;
155 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
157 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
159 struct rt_bandwidth *rt_b =
160 container_of(timer, struct rt_bandwidth, rt_period_timer);
161 ktime_t now;
162 int overrun;
163 int idle = 0;
165 for (;;) {
166 now = hrtimer_cb_get_time(timer);
167 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169 if (!overrun)
170 break;
172 idle = do_sched_rt_period_timer(rt_b, overrun);
175 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
178 static
179 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
181 rt_b->rt_period = ns_to_ktime(period);
182 rt_b->rt_runtime = runtime;
184 raw_spin_lock_init(&rt_b->rt_runtime_lock);
186 hrtimer_init(&rt_b->rt_period_timer,
187 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
188 rt_b->rt_period_timer.function = sched_rt_period_timer;
191 static inline int rt_bandwidth_enabled(void)
193 return sysctl_sched_rt_runtime >= 0;
196 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
198 ktime_t now;
200 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
201 return;
203 if (hrtimer_active(&rt_b->rt_period_timer))
204 return;
206 raw_spin_lock(&rt_b->rt_runtime_lock);
207 for (;;) {
208 unsigned long delta;
209 ktime_t soft, hard;
211 if (hrtimer_active(&rt_b->rt_period_timer))
212 break;
214 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
215 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
217 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
218 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
219 delta = ktime_to_ns(ktime_sub(hard, soft));
220 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
221 HRTIMER_MODE_ABS_PINNED, 0);
223 raw_spin_unlock(&rt_b->rt_runtime_lock);
226 #ifdef CONFIG_RT_GROUP_SCHED
227 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
229 hrtimer_cancel(&rt_b->rt_period_timer);
231 #endif
234 * sched_domains_mutex serializes calls to init_sched_domains,
235 * detach_destroy_domains and partition_sched_domains.
237 static DEFINE_MUTEX(sched_domains_mutex);
239 #ifdef CONFIG_CGROUP_SCHED
241 #include <linux/cgroup.h>
243 struct cfs_rq;
245 static LIST_HEAD(task_groups);
247 /* task group related information */
248 struct task_group {
249 struct cgroup_subsys_state css;
251 #ifdef CONFIG_FAIR_GROUP_SCHED
252 /* schedulable entities of this group on each cpu */
253 struct sched_entity **se;
254 /* runqueue "owned" by this group on each cpu */
255 struct cfs_rq **cfs_rq;
256 unsigned long shares;
258 atomic_t load_weight;
259 #endif
261 #ifdef CONFIG_RT_GROUP_SCHED
262 struct sched_rt_entity **rt_se;
263 struct rt_rq **rt_rq;
265 struct rt_bandwidth rt_bandwidth;
266 #endif
268 struct rcu_head rcu;
269 struct list_head list;
271 struct task_group *parent;
272 struct list_head siblings;
273 struct list_head children;
275 #ifdef CONFIG_SCHED_AUTOGROUP
276 struct autogroup *autogroup;
277 #endif
280 /* task_group_lock serializes the addition/removal of task groups */
281 static DEFINE_SPINLOCK(task_group_lock);
283 #ifdef CONFIG_FAIR_GROUP_SCHED
285 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
288 * A weight of 0 or 1 can cause arithmetics problems.
289 * A weight of a cfs_rq is the sum of weights of which entities
290 * are queued on this cfs_rq, so a weight of a entity should not be
291 * too large, so as the shares value of a task group.
292 * (The default weight is 1024 - so there's no practical
293 * limitation from this.)
295 #define MIN_SHARES (1UL << 1)
296 #define MAX_SHARES (1UL << 18)
298 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
299 #endif
301 /* Default task group.
302 * Every task in system belong to this group at bootup.
304 struct task_group root_task_group;
306 #endif /* CONFIG_CGROUP_SCHED */
308 /* CFS-related fields in a runqueue */
309 struct cfs_rq {
310 struct load_weight load;
311 unsigned long nr_running;
313 u64 exec_clock;
314 u64 min_vruntime;
315 #ifndef CONFIG_64BIT
316 u64 min_vruntime_copy;
317 #endif
319 struct rb_root tasks_timeline;
320 struct rb_node *rb_leftmost;
322 struct list_head tasks;
323 struct list_head *balance_iterator;
326 * 'curr' points to currently running entity on this cfs_rq.
327 * It is set to NULL otherwise (i.e when none are currently running).
329 struct sched_entity *curr, *next, *last, *skip;
331 #ifdef CONFIG_SCHED_DEBUG
332 unsigned int nr_spread_over;
333 #endif
335 #ifdef CONFIG_FAIR_GROUP_SCHED
336 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
339 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
340 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
341 * (like users, containers etc.)
343 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
344 * list is used during load balance.
346 int on_list;
347 struct list_head leaf_cfs_rq_list;
348 struct task_group *tg; /* group that "owns" this runqueue */
350 #ifdef CONFIG_SMP
352 * the part of load.weight contributed by tasks
354 unsigned long task_weight;
357 * h_load = weight * f(tg)
359 * Where f(tg) is the recursive weight fraction assigned to
360 * this group.
362 unsigned long h_load;
365 * Maintaining per-cpu shares distribution for group scheduling
367 * load_stamp is the last time we updated the load average
368 * load_last is the last time we updated the load average and saw load
369 * load_unacc_exec_time is currently unaccounted execution time
371 u64 load_avg;
372 u64 load_period;
373 u64 load_stamp, load_last, load_unacc_exec_time;
375 unsigned long load_contribution;
376 #endif
377 #endif
380 /* Real-Time classes' related field in a runqueue: */
381 struct rt_rq {
382 struct rt_prio_array active;
383 unsigned long rt_nr_running;
384 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
385 struct {
386 int curr; /* highest queued rt task prio */
387 #ifdef CONFIG_SMP
388 int next; /* next highest */
389 #endif
390 } highest_prio;
391 #endif
392 #ifdef CONFIG_SMP
393 unsigned long rt_nr_migratory;
394 unsigned long rt_nr_total;
395 int overloaded;
396 struct plist_head pushable_tasks;
397 #endif
398 int rt_throttled;
399 u64 rt_time;
400 u64 rt_runtime;
401 /* Nests inside the rq lock: */
402 raw_spinlock_t rt_runtime_lock;
404 #ifdef CONFIG_RT_GROUP_SCHED
405 unsigned long rt_nr_boosted;
407 struct rq *rq;
408 struct list_head leaf_rt_rq_list;
409 struct task_group *tg;
410 #endif
413 #ifdef CONFIG_SMP
416 * We add the notion of a root-domain which will be used to define per-domain
417 * variables. Each exclusive cpuset essentially defines an island domain by
418 * fully partitioning the member cpus from any other cpuset. Whenever a new
419 * exclusive cpuset is created, we also create and attach a new root-domain
420 * object.
423 struct root_domain {
424 atomic_t refcount;
425 struct rcu_head rcu;
426 cpumask_var_t span;
427 cpumask_var_t online;
430 * The "RT overload" flag: it gets set if a CPU has more than
431 * one runnable RT task.
433 cpumask_var_t rto_mask;
434 atomic_t rto_count;
435 struct cpupri cpupri;
439 * By default the system creates a single root-domain with all cpus as
440 * members (mimicking the global state we have today).
442 static struct root_domain def_root_domain;
444 #endif /* CONFIG_SMP */
447 * This is the main, per-CPU runqueue data structure.
449 * Locking rule: those places that want to lock multiple runqueues
450 * (such as the load balancing or the thread migration code), lock
451 * acquire operations must be ordered by ascending &runqueue.
453 struct rq {
454 /* runqueue lock: */
455 raw_spinlock_t lock;
458 * nr_running and cpu_load should be in the same cacheline because
459 * remote CPUs use both these fields when doing load calculation.
461 unsigned long nr_running;
462 #define CPU_LOAD_IDX_MAX 5
463 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
464 unsigned long last_load_update_tick;
465 #ifdef CONFIG_NO_HZ
466 u64 nohz_stamp;
467 unsigned char nohz_balance_kick;
468 #endif
469 int skip_clock_update;
471 /* capture load from *all* tasks on this cpu: */
472 struct load_weight load;
473 unsigned long nr_load_updates;
474 u64 nr_switches;
476 struct cfs_rq cfs;
477 struct rt_rq rt;
479 #ifdef CONFIG_FAIR_GROUP_SCHED
480 /* list of leaf cfs_rq on this cpu: */
481 struct list_head leaf_cfs_rq_list;
482 #endif
483 #ifdef CONFIG_RT_GROUP_SCHED
484 struct list_head leaf_rt_rq_list;
485 #endif
488 * This is part of a global counter where only the total sum
489 * over all CPUs matters. A task can increase this counter on
490 * one CPU and if it got migrated afterwards it may decrease
491 * it on another CPU. Always updated under the runqueue lock:
493 unsigned long nr_uninterruptible;
495 struct task_struct *curr, *idle, *stop;
496 unsigned long next_balance;
497 struct mm_struct *prev_mm;
499 u64 clock;
500 u64 clock_task;
502 atomic_t nr_iowait;
504 #ifdef CONFIG_SMP
505 struct root_domain *rd;
506 struct sched_domain *sd;
508 unsigned long cpu_power;
510 unsigned char idle_at_tick;
511 /* For active balancing */
512 int post_schedule;
513 int active_balance;
514 int push_cpu;
515 struct cpu_stop_work active_balance_work;
516 /* cpu of this runqueue: */
517 int cpu;
518 int online;
520 unsigned long avg_load_per_task;
522 u64 rt_avg;
523 u64 age_stamp;
524 u64 idle_stamp;
525 u64 avg_idle;
526 #endif
528 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
529 u64 prev_irq_time;
530 #endif
532 /* calc_load related fields */
533 unsigned long calc_load_update;
534 long calc_load_active;
536 #ifdef CONFIG_SCHED_HRTICK
537 #ifdef CONFIG_SMP
538 int hrtick_csd_pending;
539 struct call_single_data hrtick_csd;
540 #endif
541 struct hrtimer hrtick_timer;
542 #endif
544 #ifdef CONFIG_SCHEDSTATS
545 /* latency stats */
546 struct sched_info rq_sched_info;
547 unsigned long long rq_cpu_time;
548 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
550 /* sys_sched_yield() stats */
551 unsigned int yld_count;
553 /* schedule() stats */
554 unsigned int sched_switch;
555 unsigned int sched_count;
556 unsigned int sched_goidle;
558 /* try_to_wake_up() stats */
559 unsigned int ttwu_count;
560 unsigned int ttwu_local;
561 #endif
563 #ifdef CONFIG_SMP
564 struct task_struct *wake_list;
565 #endif
568 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
571 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
573 static inline int cpu_of(struct rq *rq)
575 #ifdef CONFIG_SMP
576 return rq->cpu;
577 #else
578 return 0;
579 #endif
582 #define rcu_dereference_check_sched_domain(p) \
583 rcu_dereference_check((p), \
584 rcu_read_lock_held() || \
585 lockdep_is_held(&sched_domains_mutex))
588 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
589 * See detach_destroy_domains: synchronize_sched for details.
591 * The domain tree of any CPU may only be accessed from within
592 * preempt-disabled sections.
594 #define for_each_domain(cpu, __sd) \
595 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
597 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
598 #define this_rq() (&__get_cpu_var(runqueues))
599 #define task_rq(p) cpu_rq(task_cpu(p))
600 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
601 #define raw_rq() (&__raw_get_cpu_var(runqueues))
603 #ifdef CONFIG_CGROUP_SCHED
606 * Return the group to which this tasks belongs.
608 * We use task_subsys_state_check() and extend the RCU verification with
609 * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each
610 * task it moves into the cgroup. Therefore by holding either of those locks,
611 * we pin the task to the current cgroup.
613 static inline struct task_group *task_group(struct task_struct *p)
615 struct task_group *tg;
616 struct cgroup_subsys_state *css;
618 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
619 lockdep_is_held(&p->pi_lock) ||
620 lockdep_is_held(&task_rq(p)->lock));
621 tg = container_of(css, struct task_group, css);
623 return autogroup_task_group(p, tg);
626 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
627 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
629 #ifdef CONFIG_FAIR_GROUP_SCHED
630 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
631 p->se.parent = task_group(p)->se[cpu];
632 #endif
634 #ifdef CONFIG_RT_GROUP_SCHED
635 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
636 p->rt.parent = task_group(p)->rt_se[cpu];
637 #endif
640 #else /* CONFIG_CGROUP_SCHED */
642 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
643 static inline struct task_group *task_group(struct task_struct *p)
645 return NULL;
648 #endif /* CONFIG_CGROUP_SCHED */
650 static void update_rq_clock_task(struct rq *rq, s64 delta);
652 static void update_rq_clock(struct rq *rq)
654 s64 delta;
656 if (rq->skip_clock_update > 0)
657 return;
659 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
660 rq->clock += delta;
661 update_rq_clock_task(rq, delta);
665 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
667 #ifdef CONFIG_SCHED_DEBUG
668 # define const_debug __read_mostly
669 #else
670 # define const_debug static const
671 #endif
674 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
675 * @cpu: the processor in question.
677 * This interface allows printk to be called with the runqueue lock
678 * held and know whether or not it is OK to wake up the klogd.
680 int runqueue_is_locked(int cpu)
682 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
686 * Debugging: various feature bits
689 #define SCHED_FEAT(name, enabled) \
690 __SCHED_FEAT_##name ,
692 enum {
693 #include "sched_features.h"
696 #undef SCHED_FEAT
698 #define SCHED_FEAT(name, enabled) \
699 (1UL << __SCHED_FEAT_##name) * enabled |
701 const_debug unsigned int sysctl_sched_features =
702 #include "sched_features.h"
705 #undef SCHED_FEAT
707 #ifdef CONFIG_SCHED_DEBUG
708 #define SCHED_FEAT(name, enabled) \
709 #name ,
711 static __read_mostly char *sched_feat_names[] = {
712 #include "sched_features.h"
713 NULL
716 #undef SCHED_FEAT
718 static int sched_feat_show(struct seq_file *m, void *v)
720 int i;
722 for (i = 0; sched_feat_names[i]; i++) {
723 if (!(sysctl_sched_features & (1UL << i)))
724 seq_puts(m, "NO_");
725 seq_printf(m, "%s ", sched_feat_names[i]);
727 seq_puts(m, "\n");
729 return 0;
732 static ssize_t
733 sched_feat_write(struct file *filp, const char __user *ubuf,
734 size_t cnt, loff_t *ppos)
736 char buf[64];
737 char *cmp;
738 int neg = 0;
739 int i;
741 if (cnt > 63)
742 cnt = 63;
744 if (copy_from_user(&buf, ubuf, cnt))
745 return -EFAULT;
747 buf[cnt] = 0;
748 cmp = strstrip(buf);
750 if (strncmp(cmp, "NO_", 3) == 0) {
751 neg = 1;
752 cmp += 3;
755 for (i = 0; sched_feat_names[i]; i++) {
756 if (strcmp(cmp, sched_feat_names[i]) == 0) {
757 if (neg)
758 sysctl_sched_features &= ~(1UL << i);
759 else
760 sysctl_sched_features |= (1UL << i);
761 break;
765 if (!sched_feat_names[i])
766 return -EINVAL;
768 *ppos += cnt;
770 return cnt;
773 static int sched_feat_open(struct inode *inode, struct file *filp)
775 return single_open(filp, sched_feat_show, NULL);
778 static const struct file_operations sched_feat_fops = {
779 .open = sched_feat_open,
780 .write = sched_feat_write,
781 .read = seq_read,
782 .llseek = seq_lseek,
783 .release = single_release,
786 static __init int sched_init_debug(void)
788 debugfs_create_file("sched_features", 0644, NULL, NULL,
789 &sched_feat_fops);
791 return 0;
793 late_initcall(sched_init_debug);
795 #endif
797 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
800 * Number of tasks to iterate in a single balance run.
801 * Limited because this is done with IRQs disabled.
803 const_debug unsigned int sysctl_sched_nr_migrate = 32;
806 * period over which we average the RT time consumption, measured
807 * in ms.
809 * default: 1s
811 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
814 * period over which we measure -rt task cpu usage in us.
815 * default: 1s
817 unsigned int sysctl_sched_rt_period = 1000000;
819 static __read_mostly int scheduler_running;
822 * part of the period that we allow rt tasks to run in us.
823 * default: 0.95s
825 int sysctl_sched_rt_runtime = 950000;
827 static inline u64 global_rt_period(void)
829 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
832 static inline u64 global_rt_runtime(void)
834 if (sysctl_sched_rt_runtime < 0)
835 return RUNTIME_INF;
837 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
840 #ifndef prepare_arch_switch
841 # define prepare_arch_switch(next) do { } while (0)
842 #endif
843 #ifndef finish_arch_switch
844 # define finish_arch_switch(prev) do { } while (0)
845 #endif
847 static inline int task_current(struct rq *rq, struct task_struct *p)
849 return rq->curr == p;
852 static inline int task_running(struct rq *rq, struct task_struct *p)
854 #ifdef CONFIG_SMP
855 return p->on_cpu;
856 #else
857 return task_current(rq, p);
858 #endif
861 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
862 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
864 #ifdef CONFIG_SMP
866 * We can optimise this out completely for !SMP, because the
867 * SMP rebalancing from interrupt is the only thing that cares
868 * here.
870 next->on_cpu = 1;
871 #endif
874 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
876 #ifdef CONFIG_SMP
878 * After ->on_cpu is cleared, the task can be moved to a different CPU.
879 * We must ensure this doesn't happen until the switch is completely
880 * finished.
882 smp_wmb();
883 prev->on_cpu = 0;
884 #endif
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 void prepare_lock_switch(struct rq *rq, struct task_struct *next)
902 #ifdef CONFIG_SMP
904 * We can optimise this out completely for !SMP, because the
905 * SMP rebalancing from interrupt is the only thing that cares
906 * here.
908 next->on_cpu = 1;
909 #endif
910 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
911 raw_spin_unlock_irq(&rq->lock);
912 #else
913 raw_spin_unlock(&rq->lock);
914 #endif
917 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
919 #ifdef CONFIG_SMP
921 * After ->on_cpu is cleared, the task can be moved to a different CPU.
922 * We must ensure this doesn't happen until the switch is completely
923 * finished.
925 smp_wmb();
926 prev->on_cpu = 0;
927 #endif
928 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
929 local_irq_enable();
930 #endif
932 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
935 * __task_rq_lock - lock the rq @p resides on.
937 static inline struct rq *__task_rq_lock(struct task_struct *p)
938 __acquires(rq->lock)
940 struct rq *rq;
942 lockdep_assert_held(&p->pi_lock);
944 for (;;) {
945 rq = task_rq(p);
946 raw_spin_lock(&rq->lock);
947 if (likely(rq == task_rq(p)))
948 return rq;
949 raw_spin_unlock(&rq->lock);
954 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
956 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
957 __acquires(p->pi_lock)
958 __acquires(rq->lock)
960 struct rq *rq;
962 for (;;) {
963 raw_spin_lock_irqsave(&p->pi_lock, *flags);
964 rq = task_rq(p);
965 raw_spin_lock(&rq->lock);
966 if (likely(rq == task_rq(p)))
967 return rq;
968 raw_spin_unlock(&rq->lock);
969 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
973 static void __task_rq_unlock(struct rq *rq)
974 __releases(rq->lock)
976 raw_spin_unlock(&rq->lock);
979 static inline void
980 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
981 __releases(rq->lock)
982 __releases(p->pi_lock)
984 raw_spin_unlock(&rq->lock);
985 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
989 * this_rq_lock - lock this runqueue and disable interrupts.
991 static struct rq *this_rq_lock(void)
992 __acquires(rq->lock)
994 struct rq *rq;
996 local_irq_disable();
997 rq = this_rq();
998 raw_spin_lock(&rq->lock);
1000 return rq;
1003 #ifdef CONFIG_SCHED_HRTICK
1005 * Use HR-timers to deliver accurate preemption points.
1007 * Its all a bit involved since we cannot program an hrt while holding the
1008 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1009 * reschedule event.
1011 * When we get rescheduled we reprogram the hrtick_timer outside of the
1012 * rq->lock.
1016 * Use hrtick when:
1017 * - enabled by features
1018 * - hrtimer is actually high res
1020 static inline int hrtick_enabled(struct rq *rq)
1022 if (!sched_feat(HRTICK))
1023 return 0;
1024 if (!cpu_active(cpu_of(rq)))
1025 return 0;
1026 return hrtimer_is_hres_active(&rq->hrtick_timer);
1029 static void hrtick_clear(struct rq *rq)
1031 if (hrtimer_active(&rq->hrtick_timer))
1032 hrtimer_cancel(&rq->hrtick_timer);
1036 * High-resolution timer tick.
1037 * Runs from hardirq context with interrupts disabled.
1039 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1041 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1043 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1045 raw_spin_lock(&rq->lock);
1046 update_rq_clock(rq);
1047 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1048 raw_spin_unlock(&rq->lock);
1050 return HRTIMER_NORESTART;
1053 #ifdef CONFIG_SMP
1055 * called from hardirq (IPI) context
1057 static void __hrtick_start(void *arg)
1059 struct rq *rq = arg;
1061 raw_spin_lock(&rq->lock);
1062 hrtimer_restart(&rq->hrtick_timer);
1063 rq->hrtick_csd_pending = 0;
1064 raw_spin_unlock(&rq->lock);
1068 * Called to set the hrtick timer state.
1070 * called with rq->lock held and irqs disabled
1072 static void hrtick_start(struct rq *rq, u64 delay)
1074 struct hrtimer *timer = &rq->hrtick_timer;
1075 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1077 hrtimer_set_expires(timer, time);
1079 if (rq == this_rq()) {
1080 hrtimer_restart(timer);
1081 } else if (!rq->hrtick_csd_pending) {
1082 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1083 rq->hrtick_csd_pending = 1;
1087 static int
1088 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1090 int cpu = (int)(long)hcpu;
1092 switch (action) {
1093 case CPU_UP_CANCELED:
1094 case CPU_UP_CANCELED_FROZEN:
1095 case CPU_DOWN_PREPARE:
1096 case CPU_DOWN_PREPARE_FROZEN:
1097 case CPU_DEAD:
1098 case CPU_DEAD_FROZEN:
1099 hrtick_clear(cpu_rq(cpu));
1100 return NOTIFY_OK;
1103 return NOTIFY_DONE;
1106 static __init void init_hrtick(void)
1108 hotcpu_notifier(hotplug_hrtick, 0);
1110 #else
1112 * Called to set the hrtick timer state.
1114 * called with rq->lock held and irqs disabled
1116 static void hrtick_start(struct rq *rq, u64 delay)
1118 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1119 HRTIMER_MODE_REL_PINNED, 0);
1122 static inline void init_hrtick(void)
1125 #endif /* CONFIG_SMP */
1127 static void init_rq_hrtick(struct rq *rq)
1129 #ifdef CONFIG_SMP
1130 rq->hrtick_csd_pending = 0;
1132 rq->hrtick_csd.flags = 0;
1133 rq->hrtick_csd.func = __hrtick_start;
1134 rq->hrtick_csd.info = rq;
1135 #endif
1137 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1138 rq->hrtick_timer.function = hrtick;
1140 #else /* CONFIG_SCHED_HRTICK */
1141 static inline void hrtick_clear(struct rq *rq)
1145 static inline void init_rq_hrtick(struct rq *rq)
1149 static inline void init_hrtick(void)
1152 #endif /* CONFIG_SCHED_HRTICK */
1155 * resched_task - mark a task 'to be rescheduled now'.
1157 * On UP this means the setting of the need_resched flag, on SMP it
1158 * might also involve a cross-CPU call to trigger the scheduler on
1159 * the target CPU.
1161 #ifdef CONFIG_SMP
1163 #ifndef tsk_is_polling
1164 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1165 #endif
1167 static void resched_task(struct task_struct *p)
1169 int cpu;
1171 assert_raw_spin_locked(&task_rq(p)->lock);
1173 if (test_tsk_need_resched(p))
1174 return;
1176 set_tsk_need_resched(p);
1178 cpu = task_cpu(p);
1179 if (cpu == smp_processor_id())
1180 return;
1182 /* NEED_RESCHED must be visible before we test polling */
1183 smp_mb();
1184 if (!tsk_is_polling(p))
1185 smp_send_reschedule(cpu);
1188 static void resched_cpu(int cpu)
1190 struct rq *rq = cpu_rq(cpu);
1191 unsigned long flags;
1193 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1194 return;
1195 resched_task(cpu_curr(cpu));
1196 raw_spin_unlock_irqrestore(&rq->lock, flags);
1199 #ifdef CONFIG_NO_HZ
1201 * In the semi idle case, use the nearest busy cpu for migrating timers
1202 * from an idle cpu. This is good for power-savings.
1204 * We don't do similar optimization for completely idle system, as
1205 * selecting an idle cpu will add more delays to the timers than intended
1206 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1208 int get_nohz_timer_target(void)
1210 int cpu = smp_processor_id();
1211 int i;
1212 struct sched_domain *sd;
1214 rcu_read_lock();
1215 for_each_domain(cpu, sd) {
1216 for_each_cpu(i, sched_domain_span(sd)) {
1217 if (!idle_cpu(i)) {
1218 cpu = i;
1219 goto unlock;
1223 unlock:
1224 rcu_read_unlock();
1225 return cpu;
1228 * When add_timer_on() enqueues a timer into the timer wheel of an
1229 * idle CPU then this timer might expire before the next timer event
1230 * which is scheduled to wake up that CPU. In case of a completely
1231 * idle system the next event might even be infinite time into the
1232 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1233 * leaves the inner idle loop so the newly added timer is taken into
1234 * account when the CPU goes back to idle and evaluates the timer
1235 * wheel for the next timer event.
1237 void wake_up_idle_cpu(int cpu)
1239 struct rq *rq = cpu_rq(cpu);
1241 if (cpu == smp_processor_id())
1242 return;
1245 * This is safe, as this function is called with the timer
1246 * wheel base lock of (cpu) held. When the CPU is on the way
1247 * to idle and has not yet set rq->curr to idle then it will
1248 * be serialized on the timer wheel base lock and take the new
1249 * timer into account automatically.
1251 if (rq->curr != rq->idle)
1252 return;
1255 * We can set TIF_RESCHED on the idle task of the other CPU
1256 * lockless. The worst case is that the other CPU runs the
1257 * idle task through an additional NOOP schedule()
1259 set_tsk_need_resched(rq->idle);
1261 /* NEED_RESCHED must be visible before we test polling */
1262 smp_mb();
1263 if (!tsk_is_polling(rq->idle))
1264 smp_send_reschedule(cpu);
1267 #endif /* CONFIG_NO_HZ */
1269 static u64 sched_avg_period(void)
1271 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1274 static void sched_avg_update(struct rq *rq)
1276 s64 period = sched_avg_period();
1278 while ((s64)(rq->clock - rq->age_stamp) > period) {
1280 * Inline assembly required to prevent the compiler
1281 * optimising this loop into a divmod call.
1282 * See __iter_div_u64_rem() for another example of this.
1284 asm("" : "+rm" (rq->age_stamp));
1285 rq->age_stamp += period;
1286 rq->rt_avg /= 2;
1290 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1292 rq->rt_avg += rt_delta;
1293 sched_avg_update(rq);
1296 #else /* !CONFIG_SMP */
1297 static void resched_task(struct task_struct *p)
1299 assert_raw_spin_locked(&task_rq(p)->lock);
1300 set_tsk_need_resched(p);
1303 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1307 static void sched_avg_update(struct rq *rq)
1310 #endif /* CONFIG_SMP */
1312 #if BITS_PER_LONG == 32
1313 # define WMULT_CONST (~0UL)
1314 #else
1315 # define WMULT_CONST (1UL << 32)
1316 #endif
1318 #define WMULT_SHIFT 32
1321 * Shift right and round:
1323 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1326 * delta *= weight / lw
1328 static unsigned long
1329 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1330 struct load_weight *lw)
1332 u64 tmp;
1335 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1336 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1337 * 2^SCHED_LOAD_RESOLUTION.
1339 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1340 tmp = (u64)delta_exec * scale_load_down(weight);
1341 else
1342 tmp = (u64)delta_exec;
1344 if (!lw->inv_weight) {
1345 unsigned long w = scale_load_down(lw->weight);
1347 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1348 lw->inv_weight = 1;
1349 else if (unlikely(!w))
1350 lw->inv_weight = WMULT_CONST;
1351 else
1352 lw->inv_weight = WMULT_CONST / w;
1356 * Check whether we'd overflow the 64-bit multiplication:
1358 if (unlikely(tmp > WMULT_CONST))
1359 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1360 WMULT_SHIFT/2);
1361 else
1362 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1364 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1367 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1369 lw->weight += inc;
1370 lw->inv_weight = 0;
1373 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1375 lw->weight -= dec;
1376 lw->inv_weight = 0;
1379 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1381 lw->weight = w;
1382 lw->inv_weight = 0;
1386 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1387 * of tasks with abnormal "nice" values across CPUs the contribution that
1388 * each task makes to its run queue's load is weighted according to its
1389 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1390 * scaled version of the new time slice allocation that they receive on time
1391 * slice expiry etc.
1394 #define WEIGHT_IDLEPRIO 3
1395 #define WMULT_IDLEPRIO 1431655765
1398 * Nice levels are multiplicative, with a gentle 10% change for every
1399 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1400 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1401 * that remained on nice 0.
1403 * The "10% effect" is relative and cumulative: from _any_ nice level,
1404 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1405 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1406 * If a task goes up by ~10% and another task goes down by ~10% then
1407 * the relative distance between them is ~25%.)
1409 static const int prio_to_weight[40] = {
1410 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1411 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1412 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1413 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1414 /* 0 */ 1024, 820, 655, 526, 423,
1415 /* 5 */ 335, 272, 215, 172, 137,
1416 /* 10 */ 110, 87, 70, 56, 45,
1417 /* 15 */ 36, 29, 23, 18, 15,
1421 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1423 * In cases where the weight does not change often, we can use the
1424 * precalculated inverse to speed up arithmetics by turning divisions
1425 * into multiplications:
1427 static const u32 prio_to_wmult[40] = {
1428 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1429 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1430 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1431 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1432 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1433 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1434 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1435 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1438 /* Time spent by the tasks of the cpu accounting group executing in ... */
1439 enum cpuacct_stat_index {
1440 CPUACCT_STAT_USER, /* ... user mode */
1441 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1443 CPUACCT_STAT_NSTATS,
1446 #ifdef CONFIG_CGROUP_CPUACCT
1447 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1448 static void cpuacct_update_stats(struct task_struct *tsk,
1449 enum cpuacct_stat_index idx, cputime_t val);
1450 #else
1451 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1452 static inline void cpuacct_update_stats(struct task_struct *tsk,
1453 enum cpuacct_stat_index idx, cputime_t val) {}
1454 #endif
1456 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1458 update_load_add(&rq->load, load);
1461 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1463 update_load_sub(&rq->load, load);
1466 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1467 typedef int (*tg_visitor)(struct task_group *, void *);
1470 * Iterate the full tree, calling @down when first entering a node and @up when
1471 * leaving it for the final time.
1473 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1475 struct task_group *parent, *child;
1476 int ret;
1478 rcu_read_lock();
1479 parent = &root_task_group;
1480 down:
1481 ret = (*down)(parent, data);
1482 if (ret)
1483 goto out_unlock;
1484 list_for_each_entry_rcu(child, &parent->children, siblings) {
1485 parent = child;
1486 goto down;
1489 continue;
1491 ret = (*up)(parent, data);
1492 if (ret)
1493 goto out_unlock;
1495 child = parent;
1496 parent = parent->parent;
1497 if (parent)
1498 goto up;
1499 out_unlock:
1500 rcu_read_unlock();
1502 return ret;
1505 static int tg_nop(struct task_group *tg, void *data)
1507 return 0;
1509 #endif
1511 #ifdef CONFIG_SMP
1512 /* Used instead of source_load when we know the type == 0 */
1513 static unsigned long weighted_cpuload(const int cpu)
1515 return cpu_rq(cpu)->load.weight;
1519 * Return a low guess at the load of a migration-source cpu weighted
1520 * according to the scheduling class and "nice" value.
1522 * We want to under-estimate the load of migration sources, to
1523 * balance conservatively.
1525 static unsigned long source_load(int cpu, int type)
1527 struct rq *rq = cpu_rq(cpu);
1528 unsigned long total = weighted_cpuload(cpu);
1530 if (type == 0 || !sched_feat(LB_BIAS))
1531 return total;
1533 return min(rq->cpu_load[type-1], total);
1537 * Return a high guess at the load of a migration-target cpu weighted
1538 * according to the scheduling class and "nice" value.
1540 static unsigned long target_load(int cpu, int type)
1542 struct rq *rq = cpu_rq(cpu);
1543 unsigned long total = weighted_cpuload(cpu);
1545 if (type == 0 || !sched_feat(LB_BIAS))
1546 return total;
1548 return max(rq->cpu_load[type-1], total);
1551 static unsigned long power_of(int cpu)
1553 return cpu_rq(cpu)->cpu_power;
1556 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1558 static unsigned long cpu_avg_load_per_task(int cpu)
1560 struct rq *rq = cpu_rq(cpu);
1561 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1563 if (nr_running)
1564 rq->avg_load_per_task = rq->load.weight / nr_running;
1565 else
1566 rq->avg_load_per_task = 0;
1568 return rq->avg_load_per_task;
1571 #ifdef CONFIG_FAIR_GROUP_SCHED
1574 * Compute the cpu's hierarchical load factor for each task group.
1575 * This needs to be done in a top-down fashion because the load of a child
1576 * group is a fraction of its parents load.
1578 static int tg_load_down(struct task_group *tg, void *data)
1580 unsigned long load;
1581 long cpu = (long)data;
1583 if (!tg->parent) {
1584 load = cpu_rq(cpu)->load.weight;
1585 } else {
1586 load = tg->parent->cfs_rq[cpu]->h_load;
1587 load *= tg->se[cpu]->load.weight;
1588 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1591 tg->cfs_rq[cpu]->h_load = load;
1593 return 0;
1596 static void update_h_load(long cpu)
1598 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1601 #endif
1603 #ifdef CONFIG_PREEMPT
1605 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1608 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1609 * way at the expense of forcing extra atomic operations in all
1610 * invocations. This assures that the double_lock is acquired using the
1611 * same underlying policy as the spinlock_t on this architecture, which
1612 * reduces latency compared to the unfair variant below. However, it
1613 * also adds more overhead and therefore may reduce throughput.
1615 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1616 __releases(this_rq->lock)
1617 __acquires(busiest->lock)
1618 __acquires(this_rq->lock)
1620 raw_spin_unlock(&this_rq->lock);
1621 double_rq_lock(this_rq, busiest);
1623 return 1;
1626 #else
1628 * Unfair double_lock_balance: Optimizes throughput at the expense of
1629 * latency by eliminating extra atomic operations when the locks are
1630 * already in proper order on entry. This favors lower cpu-ids and will
1631 * grant the double lock to lower cpus over higher ids under contention,
1632 * regardless of entry order into the function.
1634 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1635 __releases(this_rq->lock)
1636 __acquires(busiest->lock)
1637 __acquires(this_rq->lock)
1639 int ret = 0;
1641 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1642 if (busiest < this_rq) {
1643 raw_spin_unlock(&this_rq->lock);
1644 raw_spin_lock(&busiest->lock);
1645 raw_spin_lock_nested(&this_rq->lock,
1646 SINGLE_DEPTH_NESTING);
1647 ret = 1;
1648 } else
1649 raw_spin_lock_nested(&busiest->lock,
1650 SINGLE_DEPTH_NESTING);
1652 return ret;
1655 #endif /* CONFIG_PREEMPT */
1658 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1660 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1662 if (unlikely(!irqs_disabled())) {
1663 /* printk() doesn't work good under rq->lock */
1664 raw_spin_unlock(&this_rq->lock);
1665 BUG_ON(1);
1668 return _double_lock_balance(this_rq, busiest);
1671 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1672 __releases(busiest->lock)
1674 raw_spin_unlock(&busiest->lock);
1675 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1679 * double_rq_lock - safely lock two runqueues
1681 * Note this does not disable interrupts like task_rq_lock,
1682 * you need to do so manually before calling.
1684 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1685 __acquires(rq1->lock)
1686 __acquires(rq2->lock)
1688 BUG_ON(!irqs_disabled());
1689 if (rq1 == rq2) {
1690 raw_spin_lock(&rq1->lock);
1691 __acquire(rq2->lock); /* Fake it out ;) */
1692 } else {
1693 if (rq1 < rq2) {
1694 raw_spin_lock(&rq1->lock);
1695 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1696 } else {
1697 raw_spin_lock(&rq2->lock);
1698 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1704 * double_rq_unlock - safely unlock two runqueues
1706 * Note this does not restore interrupts like task_rq_unlock,
1707 * you need to do so manually after calling.
1709 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1710 __releases(rq1->lock)
1711 __releases(rq2->lock)
1713 raw_spin_unlock(&rq1->lock);
1714 if (rq1 != rq2)
1715 raw_spin_unlock(&rq2->lock);
1716 else
1717 __release(rq2->lock);
1720 #else /* CONFIG_SMP */
1723 * double_rq_lock - safely lock two runqueues
1725 * Note this does not disable interrupts like task_rq_lock,
1726 * you need to do so manually before calling.
1728 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1729 __acquires(rq1->lock)
1730 __acquires(rq2->lock)
1732 BUG_ON(!irqs_disabled());
1733 BUG_ON(rq1 != rq2);
1734 raw_spin_lock(&rq1->lock);
1735 __acquire(rq2->lock); /* Fake it out ;) */
1739 * double_rq_unlock - safely unlock two runqueues
1741 * Note this does not restore interrupts like task_rq_unlock,
1742 * you need to do so manually after calling.
1744 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1745 __releases(rq1->lock)
1746 __releases(rq2->lock)
1748 BUG_ON(rq1 != rq2);
1749 raw_spin_unlock(&rq1->lock);
1750 __release(rq2->lock);
1753 #endif
1755 static void calc_load_account_idle(struct rq *this_rq);
1756 static void update_sysctl(void);
1757 static int get_update_sysctl_factor(void);
1758 static void update_cpu_load(struct rq *this_rq);
1760 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1762 set_task_rq(p, cpu);
1763 #ifdef CONFIG_SMP
1765 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1766 * successfuly executed on another CPU. We must ensure that updates of
1767 * per-task data have been completed by this moment.
1769 smp_wmb();
1770 task_thread_info(p)->cpu = cpu;
1771 #endif
1774 static const struct sched_class rt_sched_class;
1776 #define sched_class_highest (&stop_sched_class)
1777 #define for_each_class(class) \
1778 for (class = sched_class_highest; class; class = class->next)
1780 #include "sched_stats.h"
1782 static void inc_nr_running(struct rq *rq)
1784 rq->nr_running++;
1787 static void dec_nr_running(struct rq *rq)
1789 rq->nr_running--;
1792 static void set_load_weight(struct task_struct *p)
1794 int prio = p->static_prio - MAX_RT_PRIO;
1795 struct load_weight *load = &p->se.load;
1798 * SCHED_IDLE tasks get minimal weight:
1800 if (p->policy == SCHED_IDLE) {
1801 load->weight = scale_load(WEIGHT_IDLEPRIO);
1802 load->inv_weight = WMULT_IDLEPRIO;
1803 return;
1806 load->weight = scale_load(prio_to_weight[prio]);
1807 load->inv_weight = prio_to_wmult[prio];
1810 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1812 update_rq_clock(rq);
1813 sched_info_queued(p);
1814 p->sched_class->enqueue_task(rq, p, flags);
1817 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1819 update_rq_clock(rq);
1820 sched_info_dequeued(p);
1821 p->sched_class->dequeue_task(rq, p, flags);
1825 * activate_task - move a task to the runqueue.
1827 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1829 if (task_contributes_to_load(p))
1830 rq->nr_uninterruptible--;
1832 enqueue_task(rq, p, flags);
1833 inc_nr_running(rq);
1837 * deactivate_task - remove a task from the runqueue.
1839 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1841 if (task_contributes_to_load(p))
1842 rq->nr_uninterruptible++;
1844 dequeue_task(rq, p, flags);
1845 dec_nr_running(rq);
1848 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1851 * There are no locks covering percpu hardirq/softirq time.
1852 * They are only modified in account_system_vtime, on corresponding CPU
1853 * with interrupts disabled. So, writes are safe.
1854 * They are read and saved off onto struct rq in update_rq_clock().
1855 * This may result in other CPU reading this CPU's irq time and can
1856 * race with irq/account_system_vtime on this CPU. We would either get old
1857 * or new value with a side effect of accounting a slice of irq time to wrong
1858 * task when irq is in progress while we read rq->clock. That is a worthy
1859 * compromise in place of having locks on each irq in account_system_time.
1861 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1862 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1864 static DEFINE_PER_CPU(u64, irq_start_time);
1865 static int sched_clock_irqtime;
1867 void enable_sched_clock_irqtime(void)
1869 sched_clock_irqtime = 1;
1872 void disable_sched_clock_irqtime(void)
1874 sched_clock_irqtime = 0;
1877 #ifndef CONFIG_64BIT
1878 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1880 static inline void irq_time_write_begin(void)
1882 __this_cpu_inc(irq_time_seq.sequence);
1883 smp_wmb();
1886 static inline void irq_time_write_end(void)
1888 smp_wmb();
1889 __this_cpu_inc(irq_time_seq.sequence);
1892 static inline u64 irq_time_read(int cpu)
1894 u64 irq_time;
1895 unsigned seq;
1897 do {
1898 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1899 irq_time = per_cpu(cpu_softirq_time, cpu) +
1900 per_cpu(cpu_hardirq_time, cpu);
1901 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1903 return irq_time;
1905 #else /* CONFIG_64BIT */
1906 static inline void irq_time_write_begin(void)
1910 static inline void irq_time_write_end(void)
1914 static inline u64 irq_time_read(int cpu)
1916 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1918 #endif /* CONFIG_64BIT */
1921 * Called before incrementing preempt_count on {soft,}irq_enter
1922 * and before decrementing preempt_count on {soft,}irq_exit.
1924 void account_system_vtime(struct task_struct *curr)
1926 unsigned long flags;
1927 s64 delta;
1928 int cpu;
1930 if (!sched_clock_irqtime)
1931 return;
1933 local_irq_save(flags);
1935 cpu = smp_processor_id();
1936 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1937 __this_cpu_add(irq_start_time, delta);
1939 irq_time_write_begin();
1941 * We do not account for softirq time from ksoftirqd here.
1942 * We want to continue accounting softirq time to ksoftirqd thread
1943 * in that case, so as not to confuse scheduler with a special task
1944 * that do not consume any time, but still wants to run.
1946 if (hardirq_count())
1947 __this_cpu_add(cpu_hardirq_time, delta);
1948 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1949 __this_cpu_add(cpu_softirq_time, delta);
1951 irq_time_write_end();
1952 local_irq_restore(flags);
1954 EXPORT_SYMBOL_GPL(account_system_vtime);
1956 static void update_rq_clock_task(struct rq *rq, s64 delta)
1958 s64 irq_delta;
1960 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1963 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1964 * this case when a previous update_rq_clock() happened inside a
1965 * {soft,}irq region.
1967 * When this happens, we stop ->clock_task and only update the
1968 * prev_irq_time stamp to account for the part that fit, so that a next
1969 * update will consume the rest. This ensures ->clock_task is
1970 * monotonic.
1972 * It does however cause some slight miss-attribution of {soft,}irq
1973 * time, a more accurate solution would be to update the irq_time using
1974 * the current rq->clock timestamp, except that would require using
1975 * atomic ops.
1977 if (irq_delta > delta)
1978 irq_delta = delta;
1980 rq->prev_irq_time += irq_delta;
1981 delta -= irq_delta;
1982 rq->clock_task += delta;
1984 if (irq_delta && sched_feat(NONIRQ_POWER))
1985 sched_rt_avg_update(rq, irq_delta);
1988 static int irqtime_account_hi_update(void)
1990 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1991 unsigned long flags;
1992 u64 latest_ns;
1993 int ret = 0;
1995 local_irq_save(flags);
1996 latest_ns = this_cpu_read(cpu_hardirq_time);
1997 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
1998 ret = 1;
1999 local_irq_restore(flags);
2000 return ret;
2003 static int irqtime_account_si_update(void)
2005 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2006 unsigned long flags;
2007 u64 latest_ns;
2008 int ret = 0;
2010 local_irq_save(flags);
2011 latest_ns = this_cpu_read(cpu_softirq_time);
2012 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2013 ret = 1;
2014 local_irq_restore(flags);
2015 return ret;
2018 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2020 #define sched_clock_irqtime (0)
2022 static void update_rq_clock_task(struct rq *rq, s64 delta)
2024 rq->clock_task += delta;
2027 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2029 #include "sched_idletask.c"
2030 #include "sched_fair.c"
2031 #include "sched_rt.c"
2032 #include "sched_autogroup.c"
2033 #include "sched_stoptask.c"
2034 #ifdef CONFIG_SCHED_DEBUG
2035 # include "sched_debug.c"
2036 #endif
2038 void sched_set_stop_task(int cpu, struct task_struct *stop)
2040 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2041 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2043 if (stop) {
2045 * Make it appear like a SCHED_FIFO task, its something
2046 * userspace knows about and won't get confused about.
2048 * Also, it will make PI more or less work without too
2049 * much confusion -- but then, stop work should not
2050 * rely on PI working anyway.
2052 sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
2054 stop->sched_class = &stop_sched_class;
2057 cpu_rq(cpu)->stop = stop;
2059 if (old_stop) {
2061 * Reset it back to a normal scheduling class so that
2062 * it can die in pieces.
2064 old_stop->sched_class = &rt_sched_class;
2069 * __normal_prio - return the priority that is based on the static prio
2071 static inline int __normal_prio(struct task_struct *p)
2073 return p->static_prio;
2077 * Calculate the expected normal priority: i.e. priority
2078 * without taking RT-inheritance into account. Might be
2079 * boosted by interactivity modifiers. Changes upon fork,
2080 * setprio syscalls, and whenever the interactivity
2081 * estimator recalculates.
2083 static inline int normal_prio(struct task_struct *p)
2085 int prio;
2087 if (task_has_rt_policy(p))
2088 prio = MAX_RT_PRIO-1 - p->rt_priority;
2089 else
2090 prio = __normal_prio(p);
2091 return prio;
2095 * Calculate the current priority, i.e. the priority
2096 * taken into account by the scheduler. This value might
2097 * be boosted by RT tasks, or might be boosted by
2098 * interactivity modifiers. Will be RT if the task got
2099 * RT-boosted. If not then it returns p->normal_prio.
2101 static int effective_prio(struct task_struct *p)
2103 p->normal_prio = normal_prio(p);
2105 * If we are RT tasks or we were boosted to RT priority,
2106 * keep the priority unchanged. Otherwise, update priority
2107 * to the normal priority:
2109 if (!rt_prio(p->prio))
2110 return p->normal_prio;
2111 return p->prio;
2115 * task_curr - is this task currently executing on a CPU?
2116 * @p: the task in question.
2118 inline int task_curr(const struct task_struct *p)
2120 return cpu_curr(task_cpu(p)) == p;
2123 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2124 const struct sched_class *prev_class,
2125 int oldprio)
2127 if (prev_class != p->sched_class) {
2128 if (prev_class->switched_from)
2129 prev_class->switched_from(rq, p);
2130 p->sched_class->switched_to(rq, p);
2131 } else if (oldprio != p->prio)
2132 p->sched_class->prio_changed(rq, p, oldprio);
2135 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2137 const struct sched_class *class;
2139 if (p->sched_class == rq->curr->sched_class) {
2140 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2141 } else {
2142 for_each_class(class) {
2143 if (class == rq->curr->sched_class)
2144 break;
2145 if (class == p->sched_class) {
2146 resched_task(rq->curr);
2147 break;
2153 * A queue event has occurred, and we're going to schedule. In
2154 * this case, we can save a useless back to back clock update.
2156 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2157 rq->skip_clock_update = 1;
2160 #ifdef CONFIG_SMP
2162 * Is this task likely cache-hot:
2164 static int
2165 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2167 s64 delta;
2169 if (p->sched_class != &fair_sched_class)
2170 return 0;
2172 if (unlikely(p->policy == SCHED_IDLE))
2173 return 0;
2176 * Buddy candidates are cache hot:
2178 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2179 (&p->se == cfs_rq_of(&p->se)->next ||
2180 &p->se == cfs_rq_of(&p->se)->last))
2181 return 1;
2183 if (sysctl_sched_migration_cost == -1)
2184 return 1;
2185 if (sysctl_sched_migration_cost == 0)
2186 return 0;
2188 delta = now - p->se.exec_start;
2190 return delta < (s64)sysctl_sched_migration_cost;
2193 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2195 #ifdef CONFIG_SCHED_DEBUG
2197 * We should never call set_task_cpu() on a blocked task,
2198 * ttwu() will sort out the placement.
2200 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2201 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2203 #ifdef CONFIG_LOCKDEP
2205 * The caller should hold either p->pi_lock or rq->lock, when changing
2206 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2208 * sched_move_task() holds both and thus holding either pins the cgroup,
2209 * see set_task_rq().
2211 * Furthermore, all task_rq users should acquire both locks, see
2212 * task_rq_lock().
2214 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2215 lockdep_is_held(&task_rq(p)->lock)));
2216 #endif
2217 #endif
2219 trace_sched_migrate_task(p, new_cpu);
2221 if (task_cpu(p) != new_cpu) {
2222 p->se.nr_migrations++;
2223 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2226 __set_task_cpu(p, new_cpu);
2229 struct migration_arg {
2230 struct task_struct *task;
2231 int dest_cpu;
2234 static int migration_cpu_stop(void *data);
2237 * wait_task_inactive - wait for a thread to unschedule.
2239 * If @match_state is nonzero, it's the @p->state value just checked and
2240 * not expected to change. If it changes, i.e. @p might have woken up,
2241 * then return zero. When we succeed in waiting for @p to be off its CPU,
2242 * we return a positive number (its total switch count). If a second call
2243 * a short while later returns the same number, the caller can be sure that
2244 * @p has remained unscheduled the whole time.
2246 * The caller must ensure that the task *will* unschedule sometime soon,
2247 * else this function might spin for a *long* time. This function can't
2248 * be called with interrupts off, or it may introduce deadlock with
2249 * smp_call_function() if an IPI is sent by the same process we are
2250 * waiting to become inactive.
2252 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2254 unsigned long flags;
2255 int running, on_rq;
2256 unsigned long ncsw;
2257 struct rq *rq;
2259 for (;;) {
2261 * We do the initial early heuristics without holding
2262 * any task-queue locks at all. We'll only try to get
2263 * the runqueue lock when things look like they will
2264 * work out!
2266 rq = task_rq(p);
2269 * If the task is actively running on another CPU
2270 * still, just relax and busy-wait without holding
2271 * any locks.
2273 * NOTE! Since we don't hold any locks, it's not
2274 * even sure that "rq" stays as the right runqueue!
2275 * But we don't care, since "task_running()" will
2276 * return false if the runqueue has changed and p
2277 * is actually now running somewhere else!
2279 while (task_running(rq, p)) {
2280 if (match_state && unlikely(p->state != match_state))
2281 return 0;
2282 cpu_relax();
2286 * Ok, time to look more closely! We need the rq
2287 * lock now, to be *sure*. If we're wrong, we'll
2288 * just go back and repeat.
2290 rq = task_rq_lock(p, &flags);
2291 trace_sched_wait_task(p);
2292 running = task_running(rq, p);
2293 on_rq = p->on_rq;
2294 ncsw = 0;
2295 if (!match_state || p->state == match_state)
2296 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2297 task_rq_unlock(rq, p, &flags);
2300 * If it changed from the expected state, bail out now.
2302 if (unlikely(!ncsw))
2303 break;
2306 * Was it really running after all now that we
2307 * checked with the proper locks actually held?
2309 * Oops. Go back and try again..
2311 if (unlikely(running)) {
2312 cpu_relax();
2313 continue;
2317 * It's not enough that it's not actively running,
2318 * it must be off the runqueue _entirely_, and not
2319 * preempted!
2321 * So if it was still runnable (but just not actively
2322 * running right now), it's preempted, and we should
2323 * yield - it could be a while.
2325 if (unlikely(on_rq)) {
2326 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2328 set_current_state(TASK_UNINTERRUPTIBLE);
2329 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2330 continue;
2334 * Ahh, all good. It wasn't running, and it wasn't
2335 * runnable, which means that it will never become
2336 * running in the future either. We're all done!
2338 break;
2341 return ncsw;
2344 /***
2345 * kick_process - kick a running thread to enter/exit the kernel
2346 * @p: the to-be-kicked thread
2348 * Cause a process which is running on another CPU to enter
2349 * kernel-mode, without any delay. (to get signals handled.)
2351 * NOTE: this function doesn't have to take the runqueue lock,
2352 * because all it wants to ensure is that the remote task enters
2353 * the kernel. If the IPI races and the task has been migrated
2354 * to another CPU then no harm is done and the purpose has been
2355 * achieved as well.
2357 void kick_process(struct task_struct *p)
2359 int cpu;
2361 preempt_disable();
2362 cpu = task_cpu(p);
2363 if ((cpu != smp_processor_id()) && task_curr(p))
2364 smp_send_reschedule(cpu);
2365 preempt_enable();
2367 EXPORT_SYMBOL_GPL(kick_process);
2368 #endif /* CONFIG_SMP */
2370 #ifdef CONFIG_SMP
2372 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2374 static int select_fallback_rq(int cpu, struct task_struct *p)
2376 int dest_cpu;
2377 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2379 /* Look for allowed, online CPU in same node. */
2380 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2381 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2382 return dest_cpu;
2384 /* Any allowed, online CPU? */
2385 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2386 if (dest_cpu < nr_cpu_ids)
2387 return dest_cpu;
2389 /* No more Mr. Nice Guy. */
2390 dest_cpu = cpuset_cpus_allowed_fallback(p);
2392 * Don't tell them about moving exiting tasks or
2393 * kernel threads (both mm NULL), since they never
2394 * leave kernel.
2396 if (p->mm && printk_ratelimit()) {
2397 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2398 task_pid_nr(p), p->comm, cpu);
2401 return dest_cpu;
2405 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2407 static inline
2408 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2410 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2413 * In order not to call set_task_cpu() on a blocking task we need
2414 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2415 * cpu.
2417 * Since this is common to all placement strategies, this lives here.
2419 * [ this allows ->select_task() to simply return task_cpu(p) and
2420 * not worry about this generic constraint ]
2422 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2423 !cpu_online(cpu)))
2424 cpu = select_fallback_rq(task_cpu(p), p);
2426 return cpu;
2429 static void update_avg(u64 *avg, u64 sample)
2431 s64 diff = sample - *avg;
2432 *avg += diff >> 3;
2434 #endif
2436 static void
2437 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2439 #ifdef CONFIG_SCHEDSTATS
2440 struct rq *rq = this_rq();
2442 #ifdef CONFIG_SMP
2443 int this_cpu = smp_processor_id();
2445 if (cpu == this_cpu) {
2446 schedstat_inc(rq, ttwu_local);
2447 schedstat_inc(p, se.statistics.nr_wakeups_local);
2448 } else {
2449 struct sched_domain *sd;
2451 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2452 rcu_read_lock();
2453 for_each_domain(this_cpu, sd) {
2454 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2455 schedstat_inc(sd, ttwu_wake_remote);
2456 break;
2459 rcu_read_unlock();
2462 if (wake_flags & WF_MIGRATED)
2463 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2465 #endif /* CONFIG_SMP */
2467 schedstat_inc(rq, ttwu_count);
2468 schedstat_inc(p, se.statistics.nr_wakeups);
2470 if (wake_flags & WF_SYNC)
2471 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2473 #endif /* CONFIG_SCHEDSTATS */
2476 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2478 activate_task(rq, p, en_flags);
2479 p->on_rq = 1;
2481 /* if a worker is waking up, notify workqueue */
2482 if (p->flags & PF_WQ_WORKER)
2483 wq_worker_waking_up(p, cpu_of(rq));
2487 * Mark the task runnable and perform wakeup-preemption.
2489 static void
2490 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2492 trace_sched_wakeup(p, true);
2493 check_preempt_curr(rq, p, wake_flags);
2495 p->state = TASK_RUNNING;
2496 #ifdef CONFIG_SMP
2497 if (p->sched_class->task_woken)
2498 p->sched_class->task_woken(rq, p);
2500 if (unlikely(rq->idle_stamp)) {
2501 u64 delta = rq->clock - rq->idle_stamp;
2502 u64 max = 2*sysctl_sched_migration_cost;
2504 if (delta > max)
2505 rq->avg_idle = max;
2506 else
2507 update_avg(&rq->avg_idle, delta);
2508 rq->idle_stamp = 0;
2510 #endif
2513 static void
2514 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2516 #ifdef CONFIG_SMP
2517 if (p->sched_contributes_to_load)
2518 rq->nr_uninterruptible--;
2519 #endif
2521 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2522 ttwu_do_wakeup(rq, p, wake_flags);
2526 * Called in case the task @p isn't fully descheduled from its runqueue,
2527 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2528 * since all we need to do is flip p->state to TASK_RUNNING, since
2529 * the task is still ->on_rq.
2531 static int ttwu_remote(struct task_struct *p, int wake_flags)
2533 struct rq *rq;
2534 int ret = 0;
2536 rq = __task_rq_lock(p);
2537 if (p->on_rq) {
2538 ttwu_do_wakeup(rq, p, wake_flags);
2539 ret = 1;
2541 __task_rq_unlock(rq);
2543 return ret;
2546 #ifdef CONFIG_SMP
2547 static void sched_ttwu_pending(void)
2549 struct rq *rq = this_rq();
2550 struct task_struct *list = xchg(&rq->wake_list, NULL);
2552 if (!list)
2553 return;
2555 raw_spin_lock(&rq->lock);
2557 while (list) {
2558 struct task_struct *p = list;
2559 list = list->wake_entry;
2560 ttwu_do_activate(rq, p, 0);
2563 raw_spin_unlock(&rq->lock);
2566 void scheduler_ipi(void)
2568 sched_ttwu_pending();
2571 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2573 struct rq *rq = cpu_rq(cpu);
2574 struct task_struct *next = rq->wake_list;
2576 for (;;) {
2577 struct task_struct *old = next;
2579 p->wake_entry = next;
2580 next = cmpxchg(&rq->wake_list, old, p);
2581 if (next == old)
2582 break;
2585 if (!next)
2586 smp_send_reschedule(cpu);
2589 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2590 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2592 struct rq *rq;
2593 int ret = 0;
2595 rq = __task_rq_lock(p);
2596 if (p->on_cpu) {
2597 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2598 ttwu_do_wakeup(rq, p, wake_flags);
2599 ret = 1;
2601 __task_rq_unlock(rq);
2603 return ret;
2606 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2607 #endif /* CONFIG_SMP */
2609 static void ttwu_queue(struct task_struct *p, int cpu)
2611 struct rq *rq = cpu_rq(cpu);
2613 #if defined(CONFIG_SMP)
2614 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2615 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2616 ttwu_queue_remote(p, cpu);
2617 return;
2619 #endif
2621 raw_spin_lock(&rq->lock);
2622 ttwu_do_activate(rq, p, 0);
2623 raw_spin_unlock(&rq->lock);
2627 * try_to_wake_up - wake up a thread
2628 * @p: the thread to be awakened
2629 * @state: the mask of task states that can be woken
2630 * @wake_flags: wake modifier flags (WF_*)
2632 * Put it on the run-queue if it's not already there. The "current"
2633 * thread is always on the run-queue (except when the actual
2634 * re-schedule is in progress), and as such you're allowed to do
2635 * the simpler "current->state = TASK_RUNNING" to mark yourself
2636 * runnable without the overhead of this.
2638 * Returns %true if @p was woken up, %false if it was already running
2639 * or @state didn't match @p's state.
2641 static int
2642 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2644 unsigned long flags;
2645 int cpu, success = 0;
2647 smp_wmb();
2648 raw_spin_lock_irqsave(&p->pi_lock, flags);
2649 if (!(p->state & state))
2650 goto out;
2652 success = 1; /* we're going to change ->state */
2653 cpu = task_cpu(p);
2655 if (p->on_rq && ttwu_remote(p, wake_flags))
2656 goto stat;
2658 #ifdef CONFIG_SMP
2660 * If the owning (remote) cpu is still in the middle of schedule() with
2661 * this task as prev, wait until its done referencing the task.
2663 while (p->on_cpu) {
2664 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2666 * In case the architecture enables interrupts in
2667 * context_switch(), we cannot busy wait, since that
2668 * would lead to deadlocks when an interrupt hits and
2669 * tries to wake up @prev. So bail and do a complete
2670 * remote wakeup.
2672 if (ttwu_activate_remote(p, wake_flags))
2673 goto stat;
2674 #else
2675 cpu_relax();
2676 #endif
2679 * Pairs with the smp_wmb() in finish_lock_switch().
2681 smp_rmb();
2683 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2684 p->state = TASK_WAKING;
2686 if (p->sched_class->task_waking)
2687 p->sched_class->task_waking(p);
2689 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2690 if (task_cpu(p) != cpu) {
2691 wake_flags |= WF_MIGRATED;
2692 set_task_cpu(p, cpu);
2694 #endif /* CONFIG_SMP */
2696 ttwu_queue(p, cpu);
2697 stat:
2698 ttwu_stat(p, cpu, wake_flags);
2699 out:
2700 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2702 return success;
2706 * try_to_wake_up_local - try to wake up a local task with rq lock held
2707 * @p: the thread to be awakened
2709 * Put @p on the run-queue if it's not already there. The caller must
2710 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2711 * the current task.
2713 static void try_to_wake_up_local(struct task_struct *p)
2715 struct rq *rq = task_rq(p);
2717 BUG_ON(rq != this_rq());
2718 BUG_ON(p == current);
2719 lockdep_assert_held(&rq->lock);
2721 if (!raw_spin_trylock(&p->pi_lock)) {
2722 raw_spin_unlock(&rq->lock);
2723 raw_spin_lock(&p->pi_lock);
2724 raw_spin_lock(&rq->lock);
2727 if (!(p->state & TASK_NORMAL))
2728 goto out;
2730 if (!p->on_rq)
2731 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2733 ttwu_do_wakeup(rq, p, 0);
2734 ttwu_stat(p, smp_processor_id(), 0);
2735 out:
2736 raw_spin_unlock(&p->pi_lock);
2740 * wake_up_process - Wake up a specific process
2741 * @p: The process to be woken up.
2743 * Attempt to wake up the nominated process and move it to the set of runnable
2744 * processes. Returns 1 if the process was woken up, 0 if it was already
2745 * running.
2747 * It may be assumed that this function implies a write memory barrier before
2748 * changing the task state if and only if any tasks are woken up.
2750 int wake_up_process(struct task_struct *p)
2752 return try_to_wake_up(p, TASK_ALL, 0);
2754 EXPORT_SYMBOL(wake_up_process);
2756 int wake_up_state(struct task_struct *p, unsigned int state)
2758 return try_to_wake_up(p, state, 0);
2762 * Perform scheduler related setup for a newly forked process p.
2763 * p is forked by current.
2765 * __sched_fork() is basic setup used by init_idle() too:
2767 static void __sched_fork(struct task_struct *p)
2769 p->on_rq = 0;
2771 p->se.on_rq = 0;
2772 p->se.exec_start = 0;
2773 p->se.sum_exec_runtime = 0;
2774 p->se.prev_sum_exec_runtime = 0;
2775 p->se.nr_migrations = 0;
2776 p->se.vruntime = 0;
2777 INIT_LIST_HEAD(&p->se.group_node);
2779 #ifdef CONFIG_SCHEDSTATS
2780 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2781 #endif
2783 INIT_LIST_HEAD(&p->rt.run_list);
2785 #ifdef CONFIG_PREEMPT_NOTIFIERS
2786 INIT_HLIST_HEAD(&p->preempt_notifiers);
2787 #endif
2791 * fork()/clone()-time setup:
2793 void sched_fork(struct task_struct *p)
2795 unsigned long flags;
2796 int cpu = get_cpu();
2798 __sched_fork(p);
2800 * We mark the process as running here. This guarantees that
2801 * nobody will actually run it, and a signal or other external
2802 * event cannot wake it up and insert it on the runqueue either.
2804 p->state = TASK_RUNNING;
2807 * Revert to default priority/policy on fork if requested.
2809 if (unlikely(p->sched_reset_on_fork)) {
2810 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2811 p->policy = SCHED_NORMAL;
2812 p->normal_prio = p->static_prio;
2815 if (PRIO_TO_NICE(p->static_prio) < 0) {
2816 p->static_prio = NICE_TO_PRIO(0);
2817 p->normal_prio = p->static_prio;
2818 set_load_weight(p);
2822 * We don't need the reset flag anymore after the fork. It has
2823 * fulfilled its duty:
2825 p->sched_reset_on_fork = 0;
2829 * Make sure we do not leak PI boosting priority to the child.
2831 p->prio = current->normal_prio;
2833 if (!rt_prio(p->prio))
2834 p->sched_class = &fair_sched_class;
2836 if (p->sched_class->task_fork)
2837 p->sched_class->task_fork(p);
2840 * The child is not yet in the pid-hash so no cgroup attach races,
2841 * and the cgroup is pinned to this child due to cgroup_fork()
2842 * is ran before sched_fork().
2844 * Silence PROVE_RCU.
2846 raw_spin_lock_irqsave(&p->pi_lock, flags);
2847 set_task_cpu(p, cpu);
2848 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2850 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2851 if (likely(sched_info_on()))
2852 memset(&p->sched_info, 0, sizeof(p->sched_info));
2853 #endif
2854 #if defined(CONFIG_SMP)
2855 p->on_cpu = 0;
2856 #endif
2857 #ifdef CONFIG_PREEMPT
2858 /* Want to start with kernel preemption disabled. */
2859 task_thread_info(p)->preempt_count = 1;
2860 #endif
2861 #ifdef CONFIG_SMP
2862 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2863 #endif
2865 put_cpu();
2869 * wake_up_new_task - wake up a newly created task for the first time.
2871 * This function will do some initial scheduler statistics housekeeping
2872 * that must be done for every newly created context, then puts the task
2873 * on the runqueue and wakes it.
2875 void wake_up_new_task(struct task_struct *p)
2877 unsigned long flags;
2878 struct rq *rq;
2880 raw_spin_lock_irqsave(&p->pi_lock, flags);
2881 #ifdef CONFIG_SMP
2883 * Fork balancing, do it here and not earlier because:
2884 * - cpus_allowed can change in the fork path
2885 * - any previously selected cpu might disappear through hotplug
2887 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
2888 #endif
2890 rq = __task_rq_lock(p);
2891 activate_task(rq, p, 0);
2892 p->on_rq = 1;
2893 trace_sched_wakeup_new(p, true);
2894 check_preempt_curr(rq, p, WF_FORK);
2895 #ifdef CONFIG_SMP
2896 if (p->sched_class->task_woken)
2897 p->sched_class->task_woken(rq, p);
2898 #endif
2899 task_rq_unlock(rq, p, &flags);
2902 #ifdef CONFIG_PREEMPT_NOTIFIERS
2905 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2906 * @notifier: notifier struct to register
2908 void preempt_notifier_register(struct preempt_notifier *notifier)
2910 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2912 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2915 * preempt_notifier_unregister - no longer interested in preemption notifications
2916 * @notifier: notifier struct to unregister
2918 * This is safe to call from within a preemption notifier.
2920 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2922 hlist_del(&notifier->link);
2924 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2926 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2928 struct preempt_notifier *notifier;
2929 struct hlist_node *node;
2931 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2932 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2935 static void
2936 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2937 struct task_struct *next)
2939 struct preempt_notifier *notifier;
2940 struct hlist_node *node;
2942 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2943 notifier->ops->sched_out(notifier, next);
2946 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2948 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2952 static void
2953 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2954 struct task_struct *next)
2958 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2961 * prepare_task_switch - prepare to switch tasks
2962 * @rq: the runqueue preparing to switch
2963 * @prev: the current task that is being switched out
2964 * @next: the task we are going to switch to.
2966 * This is called with the rq lock held and interrupts off. It must
2967 * be paired with a subsequent finish_task_switch after the context
2968 * switch.
2970 * prepare_task_switch sets up locking and calls architecture specific
2971 * hooks.
2973 static inline void
2974 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2975 struct task_struct *next)
2977 sched_info_switch(prev, next);
2978 perf_event_task_sched_out(prev, next);
2979 fire_sched_out_preempt_notifiers(prev, next);
2980 prepare_lock_switch(rq, next);
2981 prepare_arch_switch(next);
2982 trace_sched_switch(prev, next);
2986 * finish_task_switch - clean up after a task-switch
2987 * @rq: runqueue associated with task-switch
2988 * @prev: the thread we just switched away from.
2990 * finish_task_switch must be called after the context switch, paired
2991 * with a prepare_task_switch call before the context switch.
2992 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2993 * and do any other architecture-specific cleanup actions.
2995 * Note that we may have delayed dropping an mm in context_switch(). If
2996 * so, we finish that here outside of the runqueue lock. (Doing it
2997 * with the lock held can cause deadlocks; see schedule() for
2998 * details.)
3000 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3001 __releases(rq->lock)
3003 struct mm_struct *mm = rq->prev_mm;
3004 long prev_state;
3006 rq->prev_mm = NULL;
3009 * A task struct has one reference for the use as "current".
3010 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3011 * schedule one last time. The schedule call will never return, and
3012 * the scheduled task must drop that reference.
3013 * The test for TASK_DEAD must occur while the runqueue locks are
3014 * still held, otherwise prev could be scheduled on another cpu, die
3015 * there before we look at prev->state, and then the reference would
3016 * be dropped twice.
3017 * Manfred Spraul <manfred@colorfullife.com>
3019 prev_state = prev->state;
3020 finish_arch_switch(prev);
3021 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3022 local_irq_disable();
3023 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3024 perf_event_task_sched_in(current);
3025 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3026 local_irq_enable();
3027 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3028 finish_lock_switch(rq, prev);
3030 fire_sched_in_preempt_notifiers(current);
3031 if (mm)
3032 mmdrop(mm);
3033 if (unlikely(prev_state == TASK_DEAD)) {
3035 * Remove function-return probe instances associated with this
3036 * task and put them back on the free list.
3038 kprobe_flush_task(prev);
3039 put_task_struct(prev);
3043 #ifdef CONFIG_SMP
3045 /* assumes rq->lock is held */
3046 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3048 if (prev->sched_class->pre_schedule)
3049 prev->sched_class->pre_schedule(rq, prev);
3052 /* rq->lock is NOT held, but preemption is disabled */
3053 static inline void post_schedule(struct rq *rq)
3055 if (rq->post_schedule) {
3056 unsigned long flags;
3058 raw_spin_lock_irqsave(&rq->lock, flags);
3059 if (rq->curr->sched_class->post_schedule)
3060 rq->curr->sched_class->post_schedule(rq);
3061 raw_spin_unlock_irqrestore(&rq->lock, flags);
3063 rq->post_schedule = 0;
3067 #else
3069 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3073 static inline void post_schedule(struct rq *rq)
3077 #endif
3080 * schedule_tail - first thing a freshly forked thread must call.
3081 * @prev: the thread we just switched away from.
3083 asmlinkage void schedule_tail(struct task_struct *prev)
3084 __releases(rq->lock)
3086 struct rq *rq = this_rq();
3088 finish_task_switch(rq, prev);
3091 * FIXME: do we need to worry about rq being invalidated by the
3092 * task_switch?
3094 post_schedule(rq);
3096 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3097 /* In this case, finish_task_switch does not reenable preemption */
3098 preempt_enable();
3099 #endif
3100 if (current->set_child_tid)
3101 put_user(task_pid_vnr(current), current->set_child_tid);
3105 * context_switch - switch to the new MM and the new
3106 * thread's register state.
3108 static inline void
3109 context_switch(struct rq *rq, struct task_struct *prev,
3110 struct task_struct *next)
3112 struct mm_struct *mm, *oldmm;
3114 prepare_task_switch(rq, prev, next);
3116 mm = next->mm;
3117 oldmm = prev->active_mm;
3119 * For paravirt, this is coupled with an exit in switch_to to
3120 * combine the page table reload and the switch backend into
3121 * one hypercall.
3123 arch_start_context_switch(prev);
3125 if (!mm) {
3126 next->active_mm = oldmm;
3127 atomic_inc(&oldmm->mm_count);
3128 enter_lazy_tlb(oldmm, next);
3129 } else
3130 switch_mm(oldmm, mm, next);
3132 if (!prev->mm) {
3133 prev->active_mm = NULL;
3134 rq->prev_mm = oldmm;
3137 * Since the runqueue lock will be released by the next
3138 * task (which is an invalid locking op but in the case
3139 * of the scheduler it's an obvious special-case), so we
3140 * do an early lockdep release here:
3142 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3143 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3144 #endif
3146 /* Here we just switch the register state and the stack. */
3147 switch_to(prev, next, prev);
3149 barrier();
3151 * this_rq must be evaluated again because prev may have moved
3152 * CPUs since it called schedule(), thus the 'rq' on its stack
3153 * frame will be invalid.
3155 finish_task_switch(this_rq(), prev);
3159 * nr_running, nr_uninterruptible and nr_context_switches:
3161 * externally visible scheduler statistics: current number of runnable
3162 * threads, current number of uninterruptible-sleeping threads, total
3163 * number of context switches performed since bootup.
3165 unsigned long nr_running(void)
3167 unsigned long i, sum = 0;
3169 for_each_online_cpu(i)
3170 sum += cpu_rq(i)->nr_running;
3172 return sum;
3175 unsigned long nr_uninterruptible(void)
3177 unsigned long i, sum = 0;
3179 for_each_possible_cpu(i)
3180 sum += cpu_rq(i)->nr_uninterruptible;
3183 * Since we read the counters lockless, it might be slightly
3184 * inaccurate. Do not allow it to go below zero though:
3186 if (unlikely((long)sum < 0))
3187 sum = 0;
3189 return sum;
3192 unsigned long long nr_context_switches(void)
3194 int i;
3195 unsigned long long sum = 0;
3197 for_each_possible_cpu(i)
3198 sum += cpu_rq(i)->nr_switches;
3200 return sum;
3203 unsigned long nr_iowait(void)
3205 unsigned long i, sum = 0;
3207 for_each_possible_cpu(i)
3208 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3210 return sum;
3213 unsigned long nr_iowait_cpu(int cpu)
3215 struct rq *this = cpu_rq(cpu);
3216 return atomic_read(&this->nr_iowait);
3219 unsigned long this_cpu_load(void)
3221 struct rq *this = this_rq();
3222 return this->cpu_load[0];
3226 /* Variables and functions for calc_load */
3227 static atomic_long_t calc_load_tasks;
3228 static unsigned long calc_load_update;
3229 unsigned long avenrun[3];
3230 EXPORT_SYMBOL(avenrun);
3232 static long calc_load_fold_active(struct rq *this_rq)
3234 long nr_active, delta = 0;
3236 nr_active = this_rq->nr_running;
3237 nr_active += (long) this_rq->nr_uninterruptible;
3239 if (nr_active != this_rq->calc_load_active) {
3240 delta = nr_active - this_rq->calc_load_active;
3241 this_rq->calc_load_active = nr_active;
3244 return delta;
3247 static unsigned long
3248 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3250 load *= exp;
3251 load += active * (FIXED_1 - exp);
3252 load += 1UL << (FSHIFT - 1);
3253 return load >> FSHIFT;
3256 #ifdef CONFIG_NO_HZ
3258 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3260 * When making the ILB scale, we should try to pull this in as well.
3262 static atomic_long_t calc_load_tasks_idle;
3264 static void calc_load_account_idle(struct rq *this_rq)
3266 long delta;
3268 delta = calc_load_fold_active(this_rq);
3269 if (delta)
3270 atomic_long_add(delta, &calc_load_tasks_idle);
3273 static long calc_load_fold_idle(void)
3275 long delta = 0;
3278 * Its got a race, we don't care...
3280 if (atomic_long_read(&calc_load_tasks_idle))
3281 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3283 return delta;
3287 * fixed_power_int - compute: x^n, in O(log n) time
3289 * @x: base of the power
3290 * @frac_bits: fractional bits of @x
3291 * @n: power to raise @x to.
3293 * By exploiting the relation between the definition of the natural power
3294 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3295 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3296 * (where: n_i \elem {0, 1}, the binary vector representing n),
3297 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3298 * of course trivially computable in O(log_2 n), the length of our binary
3299 * vector.
3301 static unsigned long
3302 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3304 unsigned long result = 1UL << frac_bits;
3306 if (n) for (;;) {
3307 if (n & 1) {
3308 result *= x;
3309 result += 1UL << (frac_bits - 1);
3310 result >>= frac_bits;
3312 n >>= 1;
3313 if (!n)
3314 break;
3315 x *= x;
3316 x += 1UL << (frac_bits - 1);
3317 x >>= frac_bits;
3320 return result;
3324 * a1 = a0 * e + a * (1 - e)
3326 * a2 = a1 * e + a * (1 - e)
3327 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3328 * = a0 * e^2 + a * (1 - e) * (1 + e)
3330 * a3 = a2 * e + a * (1 - e)
3331 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3332 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3334 * ...
3336 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3337 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3338 * = a0 * e^n + a * (1 - e^n)
3340 * [1] application of the geometric series:
3342 * n 1 - x^(n+1)
3343 * S_n := \Sum x^i = -------------
3344 * i=0 1 - x
3346 static unsigned long
3347 calc_load_n(unsigned long load, unsigned long exp,
3348 unsigned long active, unsigned int n)
3351 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3355 * NO_HZ can leave us missing all per-cpu ticks calling
3356 * calc_load_account_active(), but since an idle CPU folds its delta into
3357 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3358 * in the pending idle delta if our idle period crossed a load cycle boundary.
3360 * Once we've updated the global active value, we need to apply the exponential
3361 * weights adjusted to the number of cycles missed.
3363 static void calc_global_nohz(unsigned long ticks)
3365 long delta, active, n;
3367 if (time_before(jiffies, calc_load_update))
3368 return;
3371 * If we crossed a calc_load_update boundary, make sure to fold
3372 * any pending idle changes, the respective CPUs might have
3373 * missed the tick driven calc_load_account_active() update
3374 * due to NO_HZ.
3376 delta = calc_load_fold_idle();
3377 if (delta)
3378 atomic_long_add(delta, &calc_load_tasks);
3381 * If we were idle for multiple load cycles, apply them.
3383 if (ticks >= LOAD_FREQ) {
3384 n = ticks / LOAD_FREQ;
3386 active = atomic_long_read(&calc_load_tasks);
3387 active = active > 0 ? active * FIXED_1 : 0;
3389 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3390 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3391 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3393 calc_load_update += n * LOAD_FREQ;
3397 * Its possible the remainder of the above division also crosses
3398 * a LOAD_FREQ period, the regular check in calc_global_load()
3399 * which comes after this will take care of that.
3401 * Consider us being 11 ticks before a cycle completion, and us
3402 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3403 * age us 4 cycles, and the test in calc_global_load() will
3404 * pick up the final one.
3407 #else
3408 static void calc_load_account_idle(struct rq *this_rq)
3412 static inline long calc_load_fold_idle(void)
3414 return 0;
3417 static void calc_global_nohz(unsigned long ticks)
3420 #endif
3423 * get_avenrun - get the load average array
3424 * @loads: pointer to dest load array
3425 * @offset: offset to add
3426 * @shift: shift count to shift the result left
3428 * These values are estimates at best, so no need for locking.
3430 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3432 loads[0] = (avenrun[0] + offset) << shift;
3433 loads[1] = (avenrun[1] + offset) << shift;
3434 loads[2] = (avenrun[2] + offset) << shift;
3438 * calc_load - update the avenrun load estimates 10 ticks after the
3439 * CPUs have updated calc_load_tasks.
3441 void calc_global_load(unsigned long ticks)
3443 long active;
3445 calc_global_nohz(ticks);
3447 if (time_before(jiffies, calc_load_update + 10))
3448 return;
3450 active = atomic_long_read(&calc_load_tasks);
3451 active = active > 0 ? active * FIXED_1 : 0;
3453 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3454 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3455 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3457 calc_load_update += LOAD_FREQ;
3461 * Called from update_cpu_load() to periodically update this CPU's
3462 * active count.
3464 static void calc_load_account_active(struct rq *this_rq)
3466 long delta;
3468 if (time_before(jiffies, this_rq->calc_load_update))
3469 return;
3471 delta = calc_load_fold_active(this_rq);
3472 delta += calc_load_fold_idle();
3473 if (delta)
3474 atomic_long_add(delta, &calc_load_tasks);
3476 this_rq->calc_load_update += LOAD_FREQ;
3480 * The exact cpuload at various idx values, calculated at every tick would be
3481 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3483 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3484 * on nth tick when cpu may be busy, then we have:
3485 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3486 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3488 * decay_load_missed() below does efficient calculation of
3489 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3490 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3492 * The calculation is approximated on a 128 point scale.
3493 * degrade_zero_ticks is the number of ticks after which load at any
3494 * particular idx is approximated to be zero.
3495 * degrade_factor is a precomputed table, a row for each load idx.
3496 * Each column corresponds to degradation factor for a power of two ticks,
3497 * based on 128 point scale.
3498 * Example:
3499 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3500 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3502 * With this power of 2 load factors, we can degrade the load n times
3503 * by looking at 1 bits in n and doing as many mult/shift instead of
3504 * n mult/shifts needed by the exact degradation.
3506 #define DEGRADE_SHIFT 7
3507 static const unsigned char
3508 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3509 static const unsigned char
3510 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3511 {0, 0, 0, 0, 0, 0, 0, 0},
3512 {64, 32, 8, 0, 0, 0, 0, 0},
3513 {96, 72, 40, 12, 1, 0, 0},
3514 {112, 98, 75, 43, 15, 1, 0},
3515 {120, 112, 98, 76, 45, 16, 2} };
3518 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3519 * would be when CPU is idle and so we just decay the old load without
3520 * adding any new load.
3522 static unsigned long
3523 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3525 int j = 0;
3527 if (!missed_updates)
3528 return load;
3530 if (missed_updates >= degrade_zero_ticks[idx])
3531 return 0;
3533 if (idx == 1)
3534 return load >> missed_updates;
3536 while (missed_updates) {
3537 if (missed_updates % 2)
3538 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3540 missed_updates >>= 1;
3541 j++;
3543 return load;
3547 * Update rq->cpu_load[] statistics. This function is usually called every
3548 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3549 * every tick. We fix it up based on jiffies.
3551 static void update_cpu_load(struct rq *this_rq)
3553 unsigned long this_load = this_rq->load.weight;
3554 unsigned long curr_jiffies = jiffies;
3555 unsigned long pending_updates;
3556 int i, scale;
3558 this_rq->nr_load_updates++;
3560 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3561 if (curr_jiffies == this_rq->last_load_update_tick)
3562 return;
3564 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3565 this_rq->last_load_update_tick = curr_jiffies;
3567 /* Update our load: */
3568 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3569 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3570 unsigned long old_load, new_load;
3572 /* scale is effectively 1 << i now, and >> i divides by scale */
3574 old_load = this_rq->cpu_load[i];
3575 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3576 new_load = this_load;
3578 * Round up the averaging division if load is increasing. This
3579 * prevents us from getting stuck on 9 if the load is 10, for
3580 * example.
3582 if (new_load > old_load)
3583 new_load += scale - 1;
3585 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3588 sched_avg_update(this_rq);
3591 static void update_cpu_load_active(struct rq *this_rq)
3593 update_cpu_load(this_rq);
3595 calc_load_account_active(this_rq);
3598 #ifdef CONFIG_SMP
3601 * sched_exec - execve() is a valuable balancing opportunity, because at
3602 * this point the task has the smallest effective memory and cache footprint.
3604 void sched_exec(void)
3606 struct task_struct *p = current;
3607 unsigned long flags;
3608 int dest_cpu;
3610 raw_spin_lock_irqsave(&p->pi_lock, flags);
3611 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3612 if (dest_cpu == smp_processor_id())
3613 goto unlock;
3615 if (likely(cpu_active(dest_cpu))) {
3616 struct migration_arg arg = { p, dest_cpu };
3618 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3619 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3620 return;
3622 unlock:
3623 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3626 #endif
3628 DEFINE_PER_CPU(struct kernel_stat, kstat);
3630 EXPORT_PER_CPU_SYMBOL(kstat);
3633 * Return any ns on the sched_clock that have not yet been accounted in
3634 * @p in case that task is currently running.
3636 * Called with task_rq_lock() held on @rq.
3638 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3640 u64 ns = 0;
3642 if (task_current(rq, p)) {
3643 update_rq_clock(rq);
3644 ns = rq->clock_task - p->se.exec_start;
3645 if ((s64)ns < 0)
3646 ns = 0;
3649 return ns;
3652 unsigned long long task_delta_exec(struct task_struct *p)
3654 unsigned long flags;
3655 struct rq *rq;
3656 u64 ns = 0;
3658 rq = task_rq_lock(p, &flags);
3659 ns = do_task_delta_exec(p, rq);
3660 task_rq_unlock(rq, p, &flags);
3662 return ns;
3666 * Return accounted runtime for the task.
3667 * In case the task is currently running, return the runtime plus current's
3668 * pending runtime that have not been accounted yet.
3670 unsigned long long task_sched_runtime(struct task_struct *p)
3672 unsigned long flags;
3673 struct rq *rq;
3674 u64 ns = 0;
3676 rq = task_rq_lock(p, &flags);
3677 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3678 task_rq_unlock(rq, p, &flags);
3680 return ns;
3684 * Return sum_exec_runtime for the thread group.
3685 * In case the task is currently running, return the sum plus current's
3686 * pending runtime that have not been accounted yet.
3688 * Note that the thread group might have other running tasks as well,
3689 * so the return value not includes other pending runtime that other
3690 * running tasks might have.
3692 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3694 struct task_cputime totals;
3695 unsigned long flags;
3696 struct rq *rq;
3697 u64 ns;
3699 rq = task_rq_lock(p, &flags);
3700 thread_group_cputime(p, &totals);
3701 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3702 task_rq_unlock(rq, p, &flags);
3704 return ns;
3708 * Account user cpu time to a process.
3709 * @p: the process that the cpu time gets accounted to
3710 * @cputime: the cpu time spent in user space since the last update
3711 * @cputime_scaled: cputime scaled by cpu frequency
3713 void account_user_time(struct task_struct *p, cputime_t cputime,
3714 cputime_t cputime_scaled)
3716 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3717 cputime64_t tmp;
3719 /* Add user time to process. */
3720 p->utime = cputime_add(p->utime, cputime);
3721 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3722 account_group_user_time(p, cputime);
3724 /* Add user time to cpustat. */
3725 tmp = cputime_to_cputime64(cputime);
3726 if (TASK_NICE(p) > 0)
3727 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3728 else
3729 cpustat->user = cputime64_add(cpustat->user, tmp);
3731 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3732 /* Account for user time used */
3733 acct_update_integrals(p);
3737 * Account guest cpu time to a process.
3738 * @p: the process that the cpu time gets accounted to
3739 * @cputime: the cpu time spent in virtual machine since the last update
3740 * @cputime_scaled: cputime scaled by cpu frequency
3742 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3743 cputime_t cputime_scaled)
3745 cputime64_t tmp;
3746 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3748 tmp = cputime_to_cputime64(cputime);
3750 /* Add guest time to process. */
3751 p->utime = cputime_add(p->utime, cputime);
3752 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3753 account_group_user_time(p, cputime);
3754 p->gtime = cputime_add(p->gtime, cputime);
3756 /* Add guest time to cpustat. */
3757 if (TASK_NICE(p) > 0) {
3758 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3759 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3760 } else {
3761 cpustat->user = cputime64_add(cpustat->user, tmp);
3762 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3767 * Account system cpu time to a process and desired cpustat field
3768 * @p: the process that the cpu time gets accounted to
3769 * @cputime: the cpu time spent in kernel space since the last update
3770 * @cputime_scaled: cputime scaled by cpu frequency
3771 * @target_cputime64: pointer to cpustat field that has to be updated
3773 static inline
3774 void __account_system_time(struct task_struct *p, cputime_t cputime,
3775 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3777 cputime64_t tmp = cputime_to_cputime64(cputime);
3779 /* Add system time to process. */
3780 p->stime = cputime_add(p->stime, cputime);
3781 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3782 account_group_system_time(p, cputime);
3784 /* Add system time to cpustat. */
3785 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3786 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3788 /* Account for system time used */
3789 acct_update_integrals(p);
3793 * Account system cpu time to a process.
3794 * @p: the process that the cpu time gets accounted to
3795 * @hardirq_offset: the offset to subtract from hardirq_count()
3796 * @cputime: the cpu time spent in kernel space since the last update
3797 * @cputime_scaled: cputime scaled by cpu frequency
3799 void account_system_time(struct task_struct *p, int hardirq_offset,
3800 cputime_t cputime, cputime_t cputime_scaled)
3802 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3803 cputime64_t *target_cputime64;
3805 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3806 account_guest_time(p, cputime, cputime_scaled);
3807 return;
3810 if (hardirq_count() - hardirq_offset)
3811 target_cputime64 = &cpustat->irq;
3812 else if (in_serving_softirq())
3813 target_cputime64 = &cpustat->softirq;
3814 else
3815 target_cputime64 = &cpustat->system;
3817 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3821 * Account for involuntary wait time.
3822 * @cputime: the cpu time spent in involuntary wait
3824 void account_steal_time(cputime_t cputime)
3826 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3827 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3829 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3833 * Account for idle time.
3834 * @cputime: the cpu time spent in idle wait
3836 void account_idle_time(cputime_t cputime)
3838 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3839 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3840 struct rq *rq = this_rq();
3842 if (atomic_read(&rq->nr_iowait) > 0)
3843 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3844 else
3845 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3848 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3850 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3852 * Account a tick to a process and cpustat
3853 * @p: the process that the cpu time gets accounted to
3854 * @user_tick: is the tick from userspace
3855 * @rq: the pointer to rq
3857 * Tick demultiplexing follows the order
3858 * - pending hardirq update
3859 * - pending softirq update
3860 * - user_time
3861 * - idle_time
3862 * - system time
3863 * - check for guest_time
3864 * - else account as system_time
3866 * Check for hardirq is done both for system and user time as there is
3867 * no timer going off while we are on hardirq and hence we may never get an
3868 * opportunity to update it solely in system time.
3869 * p->stime and friends are only updated on system time and not on irq
3870 * softirq as those do not count in task exec_runtime any more.
3872 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3873 struct rq *rq)
3875 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3876 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3877 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3879 if (irqtime_account_hi_update()) {
3880 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3881 } else if (irqtime_account_si_update()) {
3882 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3883 } else if (this_cpu_ksoftirqd() == p) {
3885 * ksoftirqd time do not get accounted in cpu_softirq_time.
3886 * So, we have to handle it separately here.
3887 * Also, p->stime needs to be updated for ksoftirqd.
3889 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3890 &cpustat->softirq);
3891 } else if (user_tick) {
3892 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3893 } else if (p == rq->idle) {
3894 account_idle_time(cputime_one_jiffy);
3895 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3896 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3897 } else {
3898 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3899 &cpustat->system);
3903 static void irqtime_account_idle_ticks(int ticks)
3905 int i;
3906 struct rq *rq = this_rq();
3908 for (i = 0; i < ticks; i++)
3909 irqtime_account_process_tick(current, 0, rq);
3911 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3912 static void irqtime_account_idle_ticks(int ticks) {}
3913 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3914 struct rq *rq) {}
3915 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3918 * Account a single tick of cpu time.
3919 * @p: the process that the cpu time gets accounted to
3920 * @user_tick: indicates if the tick is a user or a system tick
3922 void account_process_tick(struct task_struct *p, int user_tick)
3924 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3925 struct rq *rq = this_rq();
3927 if (sched_clock_irqtime) {
3928 irqtime_account_process_tick(p, user_tick, rq);
3929 return;
3932 if (user_tick)
3933 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3934 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3935 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3936 one_jiffy_scaled);
3937 else
3938 account_idle_time(cputime_one_jiffy);
3942 * Account multiple ticks of steal time.
3943 * @p: the process from which the cpu time has been stolen
3944 * @ticks: number of stolen ticks
3946 void account_steal_ticks(unsigned long ticks)
3948 account_steal_time(jiffies_to_cputime(ticks));
3952 * Account multiple ticks of idle time.
3953 * @ticks: number of stolen ticks
3955 void account_idle_ticks(unsigned long ticks)
3958 if (sched_clock_irqtime) {
3959 irqtime_account_idle_ticks(ticks);
3960 return;
3963 account_idle_time(jiffies_to_cputime(ticks));
3966 #endif
3969 * Use precise platform statistics if available:
3971 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3972 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3974 *ut = p->utime;
3975 *st = p->stime;
3978 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3980 struct task_cputime cputime;
3982 thread_group_cputime(p, &cputime);
3984 *ut = cputime.utime;
3985 *st = cputime.stime;
3987 #else
3989 #ifndef nsecs_to_cputime
3990 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3991 #endif
3993 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3995 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3998 * Use CFS's precise accounting:
4000 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4002 if (total) {
4003 u64 temp = rtime;
4005 temp *= utime;
4006 do_div(temp, total);
4007 utime = (cputime_t)temp;
4008 } else
4009 utime = rtime;
4012 * Compare with previous values, to keep monotonicity:
4014 p->prev_utime = max(p->prev_utime, utime);
4015 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4017 *ut = p->prev_utime;
4018 *st = p->prev_stime;
4022 * Must be called with siglock held.
4024 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4026 struct signal_struct *sig = p->signal;
4027 struct task_cputime cputime;
4028 cputime_t rtime, utime, total;
4030 thread_group_cputime(p, &cputime);
4032 total = cputime_add(cputime.utime, cputime.stime);
4033 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4035 if (total) {
4036 u64 temp = rtime;
4038 temp *= cputime.utime;
4039 do_div(temp, total);
4040 utime = (cputime_t)temp;
4041 } else
4042 utime = rtime;
4044 sig->prev_utime = max(sig->prev_utime, utime);
4045 sig->prev_stime = max(sig->prev_stime,
4046 cputime_sub(rtime, sig->prev_utime));
4048 *ut = sig->prev_utime;
4049 *st = sig->prev_stime;
4051 #endif
4054 * This function gets called by the timer code, with HZ frequency.
4055 * We call it with interrupts disabled.
4057 void scheduler_tick(void)
4059 int cpu = smp_processor_id();
4060 struct rq *rq = cpu_rq(cpu);
4061 struct task_struct *curr = rq->curr;
4063 sched_clock_tick();
4065 raw_spin_lock(&rq->lock);
4066 update_rq_clock(rq);
4067 update_cpu_load_active(rq);
4068 curr->sched_class->task_tick(rq, curr, 0);
4069 raw_spin_unlock(&rq->lock);
4071 perf_event_task_tick();
4073 #ifdef CONFIG_SMP
4074 rq->idle_at_tick = idle_cpu(cpu);
4075 trigger_load_balance(rq, cpu);
4076 #endif
4079 notrace unsigned long get_parent_ip(unsigned long addr)
4081 if (in_lock_functions(addr)) {
4082 addr = CALLER_ADDR2;
4083 if (in_lock_functions(addr))
4084 addr = CALLER_ADDR3;
4086 return addr;
4089 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4090 defined(CONFIG_PREEMPT_TRACER))
4092 void __kprobes add_preempt_count(int val)
4094 #ifdef CONFIG_DEBUG_PREEMPT
4096 * Underflow?
4098 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4099 return;
4100 #endif
4101 preempt_count() += val;
4102 #ifdef CONFIG_DEBUG_PREEMPT
4104 * Spinlock count overflowing soon?
4106 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4107 PREEMPT_MASK - 10);
4108 #endif
4109 if (preempt_count() == val)
4110 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4112 EXPORT_SYMBOL(add_preempt_count);
4114 void __kprobes sub_preempt_count(int val)
4116 #ifdef CONFIG_DEBUG_PREEMPT
4118 * Underflow?
4120 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4121 return;
4123 * Is the spinlock portion underflowing?
4125 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4126 !(preempt_count() & PREEMPT_MASK)))
4127 return;
4128 #endif
4130 if (preempt_count() == val)
4131 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4132 preempt_count() -= val;
4134 EXPORT_SYMBOL(sub_preempt_count);
4136 #endif
4139 * Print scheduling while atomic bug:
4141 static noinline void __schedule_bug(struct task_struct *prev)
4143 struct pt_regs *regs = get_irq_regs();
4145 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4146 prev->comm, prev->pid, preempt_count());
4148 debug_show_held_locks(prev);
4149 print_modules();
4150 if (irqs_disabled())
4151 print_irqtrace_events(prev);
4153 if (regs)
4154 show_regs(regs);
4155 else
4156 dump_stack();
4160 * Various schedule()-time debugging checks and statistics:
4162 static inline void schedule_debug(struct task_struct *prev)
4165 * Test if we are atomic. Since do_exit() needs to call into
4166 * schedule() atomically, we ignore that path for now.
4167 * Otherwise, whine if we are scheduling when we should not be.
4169 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4170 __schedule_bug(prev);
4172 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4174 schedstat_inc(this_rq(), sched_count);
4177 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4179 if (prev->on_rq || rq->skip_clock_update < 0)
4180 update_rq_clock(rq);
4181 prev->sched_class->put_prev_task(rq, prev);
4185 * Pick up the highest-prio task:
4187 static inline struct task_struct *
4188 pick_next_task(struct rq *rq)
4190 const struct sched_class *class;
4191 struct task_struct *p;
4194 * Optimization: we know that if all tasks are in
4195 * the fair class we can call that function directly:
4197 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4198 p = fair_sched_class.pick_next_task(rq);
4199 if (likely(p))
4200 return p;
4203 for_each_class(class) {
4204 p = class->pick_next_task(rq);
4205 if (p)
4206 return p;
4209 BUG(); /* the idle class will always have a runnable task */
4213 * schedule() is the main scheduler function.
4215 asmlinkage void __sched schedule(void)
4217 struct task_struct *prev, *next;
4218 unsigned long *switch_count;
4219 struct rq *rq;
4220 int cpu;
4222 need_resched:
4223 preempt_disable();
4224 cpu = smp_processor_id();
4225 rq = cpu_rq(cpu);
4226 rcu_note_context_switch(cpu);
4227 prev = rq->curr;
4229 schedule_debug(prev);
4231 if (sched_feat(HRTICK))
4232 hrtick_clear(rq);
4234 raw_spin_lock_irq(&rq->lock);
4236 switch_count = &prev->nivcsw;
4237 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4238 if (unlikely(signal_pending_state(prev->state, prev))) {
4239 prev->state = TASK_RUNNING;
4240 } else {
4241 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4242 prev->on_rq = 0;
4245 * If a worker went to sleep, notify and ask workqueue
4246 * whether it wants to wake up a task to maintain
4247 * concurrency.
4249 if (prev->flags & PF_WQ_WORKER) {
4250 struct task_struct *to_wakeup;
4252 to_wakeup = wq_worker_sleeping(prev, cpu);
4253 if (to_wakeup)
4254 try_to_wake_up_local(to_wakeup);
4258 * If we are going to sleep and we have plugged IO
4259 * queued, make sure to submit it to avoid deadlocks.
4261 if (blk_needs_flush_plug(prev)) {
4262 raw_spin_unlock(&rq->lock);
4263 blk_schedule_flush_plug(prev);
4264 raw_spin_lock(&rq->lock);
4267 switch_count = &prev->nvcsw;
4270 pre_schedule(rq, prev);
4272 if (unlikely(!rq->nr_running))
4273 idle_balance(cpu, rq);
4275 put_prev_task(rq, prev);
4276 next = pick_next_task(rq);
4277 clear_tsk_need_resched(prev);
4278 rq->skip_clock_update = 0;
4280 if (likely(prev != next)) {
4281 rq->nr_switches++;
4282 rq->curr = next;
4283 ++*switch_count;
4285 context_switch(rq, prev, next); /* unlocks the rq */
4287 * The context switch have flipped the stack from under us
4288 * and restored the local variables which were saved when
4289 * this task called schedule() in the past. prev == current
4290 * is still correct, but it can be moved to another cpu/rq.
4292 cpu = smp_processor_id();
4293 rq = cpu_rq(cpu);
4294 } else
4295 raw_spin_unlock_irq(&rq->lock);
4297 post_schedule(rq);
4299 preempt_enable_no_resched();
4300 if (need_resched())
4301 goto need_resched;
4303 EXPORT_SYMBOL(schedule);
4305 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4307 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4309 bool ret = false;
4311 rcu_read_lock();
4312 if (lock->owner != owner)
4313 goto fail;
4316 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4317 * lock->owner still matches owner, if that fails, owner might
4318 * point to free()d memory, if it still matches, the rcu_read_lock()
4319 * ensures the memory stays valid.
4321 barrier();
4323 ret = owner->on_cpu;
4324 fail:
4325 rcu_read_unlock();
4327 return ret;
4331 * Look out! "owner" is an entirely speculative pointer
4332 * access and not reliable.
4334 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4336 if (!sched_feat(OWNER_SPIN))
4337 return 0;
4339 while (owner_running(lock, owner)) {
4340 if (need_resched())
4341 return 0;
4343 arch_mutex_cpu_relax();
4347 * If the owner changed to another task there is likely
4348 * heavy contention, stop spinning.
4350 if (lock->owner)
4351 return 0;
4353 return 1;
4355 #endif
4357 #ifdef CONFIG_PREEMPT
4359 * this is the entry point to schedule() from in-kernel preemption
4360 * off of preempt_enable. Kernel preemptions off return from interrupt
4361 * occur there and call schedule directly.
4363 asmlinkage void __sched notrace preempt_schedule(void)
4365 struct thread_info *ti = current_thread_info();
4368 * If there is a non-zero preempt_count or interrupts are disabled,
4369 * we do not want to preempt the current task. Just return..
4371 if (likely(ti->preempt_count || irqs_disabled()))
4372 return;
4374 do {
4375 add_preempt_count_notrace(PREEMPT_ACTIVE);
4376 schedule();
4377 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4380 * Check again in case we missed a preemption opportunity
4381 * between schedule and now.
4383 barrier();
4384 } while (need_resched());
4386 EXPORT_SYMBOL(preempt_schedule);
4389 * this is the entry point to schedule() from kernel preemption
4390 * off of irq context.
4391 * Note, that this is called and return with irqs disabled. This will
4392 * protect us against recursive calling from irq.
4394 asmlinkage void __sched preempt_schedule_irq(void)
4396 struct thread_info *ti = current_thread_info();
4398 /* Catch callers which need to be fixed */
4399 BUG_ON(ti->preempt_count || !irqs_disabled());
4401 do {
4402 add_preempt_count(PREEMPT_ACTIVE);
4403 local_irq_enable();
4404 schedule();
4405 local_irq_disable();
4406 sub_preempt_count(PREEMPT_ACTIVE);
4409 * Check again in case we missed a preemption opportunity
4410 * between schedule and now.
4412 barrier();
4413 } while (need_resched());
4416 #endif /* CONFIG_PREEMPT */
4418 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4419 void *key)
4421 return try_to_wake_up(curr->private, mode, wake_flags);
4423 EXPORT_SYMBOL(default_wake_function);
4426 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4427 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4428 * number) then we wake all the non-exclusive tasks and one exclusive task.
4430 * There are circumstances in which we can try to wake a task which has already
4431 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4432 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4434 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4435 int nr_exclusive, int wake_flags, void *key)
4437 wait_queue_t *curr, *next;
4439 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4440 unsigned flags = curr->flags;
4442 if (curr->func(curr, mode, wake_flags, key) &&
4443 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4444 break;
4449 * __wake_up - wake up threads blocked on a waitqueue.
4450 * @q: the waitqueue
4451 * @mode: which threads
4452 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4453 * @key: is directly passed to the wakeup function
4455 * It may be assumed that this function implies a write memory barrier before
4456 * changing the task state if and only if any tasks are woken up.
4458 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4459 int nr_exclusive, void *key)
4461 unsigned long flags;
4463 spin_lock_irqsave(&q->lock, flags);
4464 __wake_up_common(q, mode, nr_exclusive, 0, key);
4465 spin_unlock_irqrestore(&q->lock, flags);
4467 EXPORT_SYMBOL(__wake_up);
4470 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4472 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4474 __wake_up_common(q, mode, 1, 0, NULL);
4476 EXPORT_SYMBOL_GPL(__wake_up_locked);
4478 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4480 __wake_up_common(q, mode, 1, 0, key);
4482 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4485 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4486 * @q: the waitqueue
4487 * @mode: which threads
4488 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4489 * @key: opaque value to be passed to wakeup targets
4491 * The sync wakeup differs that the waker knows that it will schedule
4492 * away soon, so while the target thread will be woken up, it will not
4493 * be migrated to another CPU - ie. the two threads are 'synchronized'
4494 * with each other. This can prevent needless bouncing between CPUs.
4496 * On UP it can prevent extra preemption.
4498 * It may be assumed that this function implies a write memory barrier before
4499 * changing the task state if and only if any tasks are woken up.
4501 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4502 int nr_exclusive, void *key)
4504 unsigned long flags;
4505 int wake_flags = WF_SYNC;
4507 if (unlikely(!q))
4508 return;
4510 if (unlikely(!nr_exclusive))
4511 wake_flags = 0;
4513 spin_lock_irqsave(&q->lock, flags);
4514 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4515 spin_unlock_irqrestore(&q->lock, flags);
4517 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4520 * __wake_up_sync - see __wake_up_sync_key()
4522 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4524 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4526 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4529 * complete: - signals a single thread waiting on this completion
4530 * @x: holds the state of this particular completion
4532 * This will wake up a single thread waiting on this completion. Threads will be
4533 * awakened in the same order in which they were queued.
4535 * See also complete_all(), wait_for_completion() and related routines.
4537 * It may be assumed that this function implies a write memory barrier before
4538 * changing the task state if and only if any tasks are woken up.
4540 void complete(struct completion *x)
4542 unsigned long flags;
4544 spin_lock_irqsave(&x->wait.lock, flags);
4545 x->done++;
4546 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4547 spin_unlock_irqrestore(&x->wait.lock, flags);
4549 EXPORT_SYMBOL(complete);
4552 * complete_all: - signals all threads waiting on this completion
4553 * @x: holds the state of this particular completion
4555 * This will wake up all threads waiting on this particular completion event.
4557 * It may be assumed that this function implies a write memory barrier before
4558 * changing the task state if and only if any tasks are woken up.
4560 void complete_all(struct completion *x)
4562 unsigned long flags;
4564 spin_lock_irqsave(&x->wait.lock, flags);
4565 x->done += UINT_MAX/2;
4566 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4567 spin_unlock_irqrestore(&x->wait.lock, flags);
4569 EXPORT_SYMBOL(complete_all);
4571 static inline long __sched
4572 do_wait_for_common(struct completion *x, long timeout, int state)
4574 if (!x->done) {
4575 DECLARE_WAITQUEUE(wait, current);
4577 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4578 do {
4579 if (signal_pending_state(state, current)) {
4580 timeout = -ERESTARTSYS;
4581 break;
4583 __set_current_state(state);
4584 spin_unlock_irq(&x->wait.lock);
4585 timeout = schedule_timeout(timeout);
4586 spin_lock_irq(&x->wait.lock);
4587 } while (!x->done && timeout);
4588 __remove_wait_queue(&x->wait, &wait);
4589 if (!x->done)
4590 return timeout;
4592 x->done--;
4593 return timeout ?: 1;
4596 static long __sched
4597 wait_for_common(struct completion *x, long timeout, int state)
4599 might_sleep();
4601 spin_lock_irq(&x->wait.lock);
4602 timeout = do_wait_for_common(x, timeout, state);
4603 spin_unlock_irq(&x->wait.lock);
4604 return timeout;
4608 * wait_for_completion: - waits for completion of a task
4609 * @x: holds the state of this particular completion
4611 * This waits to be signaled for completion of a specific task. It is NOT
4612 * interruptible and there is no timeout.
4614 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4615 * and interrupt capability. Also see complete().
4617 void __sched wait_for_completion(struct completion *x)
4619 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4621 EXPORT_SYMBOL(wait_for_completion);
4624 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4625 * @x: holds the state of this particular completion
4626 * @timeout: timeout value in jiffies
4628 * This waits for either a completion of a specific task to be signaled or for a
4629 * specified timeout to expire. The timeout is in jiffies. It is not
4630 * interruptible.
4632 unsigned long __sched
4633 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4635 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4637 EXPORT_SYMBOL(wait_for_completion_timeout);
4640 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4641 * @x: holds the state of this particular completion
4643 * This waits for completion of a specific task to be signaled. It is
4644 * interruptible.
4646 int __sched wait_for_completion_interruptible(struct completion *x)
4648 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4649 if (t == -ERESTARTSYS)
4650 return t;
4651 return 0;
4653 EXPORT_SYMBOL(wait_for_completion_interruptible);
4656 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4657 * @x: holds the state of this particular completion
4658 * @timeout: timeout value in jiffies
4660 * This waits for either a completion of a specific task to be signaled or for a
4661 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4663 long __sched
4664 wait_for_completion_interruptible_timeout(struct completion *x,
4665 unsigned long timeout)
4667 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4669 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4672 * wait_for_completion_killable: - waits for completion of a task (killable)
4673 * @x: holds the state of this particular completion
4675 * This waits to be signaled for completion of a specific task. It can be
4676 * interrupted by a kill signal.
4678 int __sched wait_for_completion_killable(struct completion *x)
4680 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4681 if (t == -ERESTARTSYS)
4682 return t;
4683 return 0;
4685 EXPORT_SYMBOL(wait_for_completion_killable);
4688 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4689 * @x: holds the state of this particular completion
4690 * @timeout: timeout value in jiffies
4692 * This waits for either a completion of a specific task to be
4693 * signaled or for a specified timeout to expire. It can be
4694 * interrupted by a kill signal. The timeout is in jiffies.
4696 long __sched
4697 wait_for_completion_killable_timeout(struct completion *x,
4698 unsigned long timeout)
4700 return wait_for_common(x, timeout, TASK_KILLABLE);
4702 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4705 * try_wait_for_completion - try to decrement a completion without blocking
4706 * @x: completion structure
4708 * Returns: 0 if a decrement cannot be done without blocking
4709 * 1 if a decrement succeeded.
4711 * If a completion is being used as a counting completion,
4712 * attempt to decrement the counter without blocking. This
4713 * enables us to avoid waiting if the resource the completion
4714 * is protecting is not available.
4716 bool try_wait_for_completion(struct completion *x)
4718 unsigned long flags;
4719 int ret = 1;
4721 spin_lock_irqsave(&x->wait.lock, flags);
4722 if (!x->done)
4723 ret = 0;
4724 else
4725 x->done--;
4726 spin_unlock_irqrestore(&x->wait.lock, flags);
4727 return ret;
4729 EXPORT_SYMBOL(try_wait_for_completion);
4732 * completion_done - Test to see if a completion has any waiters
4733 * @x: completion structure
4735 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4736 * 1 if there are no waiters.
4739 bool completion_done(struct completion *x)
4741 unsigned long flags;
4742 int ret = 1;
4744 spin_lock_irqsave(&x->wait.lock, flags);
4745 if (!x->done)
4746 ret = 0;
4747 spin_unlock_irqrestore(&x->wait.lock, flags);
4748 return ret;
4750 EXPORT_SYMBOL(completion_done);
4752 static long __sched
4753 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4755 unsigned long flags;
4756 wait_queue_t wait;
4758 init_waitqueue_entry(&wait, current);
4760 __set_current_state(state);
4762 spin_lock_irqsave(&q->lock, flags);
4763 __add_wait_queue(q, &wait);
4764 spin_unlock(&q->lock);
4765 timeout = schedule_timeout(timeout);
4766 spin_lock_irq(&q->lock);
4767 __remove_wait_queue(q, &wait);
4768 spin_unlock_irqrestore(&q->lock, flags);
4770 return timeout;
4773 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4775 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4777 EXPORT_SYMBOL(interruptible_sleep_on);
4779 long __sched
4780 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4782 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4784 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4786 void __sched sleep_on(wait_queue_head_t *q)
4788 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4790 EXPORT_SYMBOL(sleep_on);
4792 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4794 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4796 EXPORT_SYMBOL(sleep_on_timeout);
4798 #ifdef CONFIG_RT_MUTEXES
4801 * rt_mutex_setprio - set the current priority of a task
4802 * @p: task
4803 * @prio: prio value (kernel-internal form)
4805 * This function changes the 'effective' priority of a task. It does
4806 * not touch ->normal_prio like __setscheduler().
4808 * Used by the rt_mutex code to implement priority inheritance logic.
4810 void rt_mutex_setprio(struct task_struct *p, int prio)
4812 int oldprio, on_rq, running;
4813 struct rq *rq;
4814 const struct sched_class *prev_class;
4816 BUG_ON(prio < 0 || prio > MAX_PRIO);
4818 rq = __task_rq_lock(p);
4820 trace_sched_pi_setprio(p, prio);
4821 oldprio = p->prio;
4822 prev_class = p->sched_class;
4823 on_rq = p->on_rq;
4824 running = task_current(rq, p);
4825 if (on_rq)
4826 dequeue_task(rq, p, 0);
4827 if (running)
4828 p->sched_class->put_prev_task(rq, p);
4830 if (rt_prio(prio))
4831 p->sched_class = &rt_sched_class;
4832 else
4833 p->sched_class = &fair_sched_class;
4835 p->prio = prio;
4837 if (running)
4838 p->sched_class->set_curr_task(rq);
4839 if (on_rq)
4840 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4842 check_class_changed(rq, p, prev_class, oldprio);
4843 __task_rq_unlock(rq);
4846 #endif
4848 void set_user_nice(struct task_struct *p, long nice)
4850 int old_prio, delta, on_rq;
4851 unsigned long flags;
4852 struct rq *rq;
4854 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4855 return;
4857 * We have to be careful, if called from sys_setpriority(),
4858 * the task might be in the middle of scheduling on another CPU.
4860 rq = task_rq_lock(p, &flags);
4862 * The RT priorities are set via sched_setscheduler(), but we still
4863 * allow the 'normal' nice value to be set - but as expected
4864 * it wont have any effect on scheduling until the task is
4865 * SCHED_FIFO/SCHED_RR:
4867 if (task_has_rt_policy(p)) {
4868 p->static_prio = NICE_TO_PRIO(nice);
4869 goto out_unlock;
4871 on_rq = p->on_rq;
4872 if (on_rq)
4873 dequeue_task(rq, p, 0);
4875 p->static_prio = NICE_TO_PRIO(nice);
4876 set_load_weight(p);
4877 old_prio = p->prio;
4878 p->prio = effective_prio(p);
4879 delta = p->prio - old_prio;
4881 if (on_rq) {
4882 enqueue_task(rq, p, 0);
4884 * If the task increased its priority or is running and
4885 * lowered its priority, then reschedule its CPU:
4887 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4888 resched_task(rq->curr);
4890 out_unlock:
4891 task_rq_unlock(rq, p, &flags);
4893 EXPORT_SYMBOL(set_user_nice);
4896 * can_nice - check if a task can reduce its nice value
4897 * @p: task
4898 * @nice: nice value
4900 int can_nice(const struct task_struct *p, const int nice)
4902 /* convert nice value [19,-20] to rlimit style value [1,40] */
4903 int nice_rlim = 20 - nice;
4905 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4906 capable(CAP_SYS_NICE));
4909 #ifdef __ARCH_WANT_SYS_NICE
4912 * sys_nice - change the priority of the current process.
4913 * @increment: priority increment
4915 * sys_setpriority is a more generic, but much slower function that
4916 * does similar things.
4918 SYSCALL_DEFINE1(nice, int, increment)
4920 long nice, retval;
4923 * Setpriority might change our priority at the same moment.
4924 * We don't have to worry. Conceptually one call occurs first
4925 * and we have a single winner.
4927 if (increment < -40)
4928 increment = -40;
4929 if (increment > 40)
4930 increment = 40;
4932 nice = TASK_NICE(current) + increment;
4933 if (nice < -20)
4934 nice = -20;
4935 if (nice > 19)
4936 nice = 19;
4938 if (increment < 0 && !can_nice(current, nice))
4939 return -EPERM;
4941 retval = security_task_setnice(current, nice);
4942 if (retval)
4943 return retval;
4945 set_user_nice(current, nice);
4946 return 0;
4949 #endif
4952 * task_prio - return the priority value of a given task.
4953 * @p: the task in question.
4955 * This is the priority value as seen by users in /proc.
4956 * RT tasks are offset by -200. Normal tasks are centered
4957 * around 0, value goes from -16 to +15.
4959 int task_prio(const struct task_struct *p)
4961 return p->prio - MAX_RT_PRIO;
4965 * task_nice - return the nice value of a given task.
4966 * @p: the task in question.
4968 int task_nice(const struct task_struct *p)
4970 return TASK_NICE(p);
4972 EXPORT_SYMBOL(task_nice);
4975 * idle_cpu - is a given cpu idle currently?
4976 * @cpu: the processor in question.
4978 int idle_cpu(int cpu)
4980 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4984 * idle_task - return the idle task for a given cpu.
4985 * @cpu: the processor in question.
4987 struct task_struct *idle_task(int cpu)
4989 return cpu_rq(cpu)->idle;
4993 * find_process_by_pid - find a process with a matching PID value.
4994 * @pid: the pid in question.
4996 static struct task_struct *find_process_by_pid(pid_t pid)
4998 return pid ? find_task_by_vpid(pid) : current;
5001 /* Actually do priority change: must hold rq lock. */
5002 static void
5003 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5005 p->policy = policy;
5006 p->rt_priority = prio;
5007 p->normal_prio = normal_prio(p);
5008 /* we are holding p->pi_lock already */
5009 p->prio = rt_mutex_getprio(p);
5010 if (rt_prio(p->prio))
5011 p->sched_class = &rt_sched_class;
5012 else
5013 p->sched_class = &fair_sched_class;
5014 set_load_weight(p);
5018 * check the target process has a UID that matches the current process's
5020 static bool check_same_owner(struct task_struct *p)
5022 const struct cred *cred = current_cred(), *pcred;
5023 bool match;
5025 rcu_read_lock();
5026 pcred = __task_cred(p);
5027 if (cred->user->user_ns == pcred->user->user_ns)
5028 match = (cred->euid == pcred->euid ||
5029 cred->euid == pcred->uid);
5030 else
5031 match = false;
5032 rcu_read_unlock();
5033 return match;
5036 static int __sched_setscheduler(struct task_struct *p, int policy,
5037 const struct sched_param *param, bool user)
5039 int retval, oldprio, oldpolicy = -1, on_rq, running;
5040 unsigned long flags;
5041 const struct sched_class *prev_class;
5042 struct rq *rq;
5043 int reset_on_fork;
5045 /* may grab non-irq protected spin_locks */
5046 BUG_ON(in_interrupt());
5047 recheck:
5048 /* double check policy once rq lock held */
5049 if (policy < 0) {
5050 reset_on_fork = p->sched_reset_on_fork;
5051 policy = oldpolicy = p->policy;
5052 } else {
5053 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5054 policy &= ~SCHED_RESET_ON_FORK;
5056 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5057 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5058 policy != SCHED_IDLE)
5059 return -EINVAL;
5063 * Valid priorities for SCHED_FIFO and SCHED_RR are
5064 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5065 * SCHED_BATCH and SCHED_IDLE is 0.
5067 if (param->sched_priority < 0 ||
5068 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5069 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5070 return -EINVAL;
5071 if (rt_policy(policy) != (param->sched_priority != 0))
5072 return -EINVAL;
5075 * Allow unprivileged RT tasks to decrease priority:
5077 if (user && !capable(CAP_SYS_NICE)) {
5078 if (rt_policy(policy)) {
5079 unsigned long rlim_rtprio =
5080 task_rlimit(p, RLIMIT_RTPRIO);
5082 /* can't set/change the rt policy */
5083 if (policy != p->policy && !rlim_rtprio)
5084 return -EPERM;
5086 /* can't increase priority */
5087 if (param->sched_priority > p->rt_priority &&
5088 param->sched_priority > rlim_rtprio)
5089 return -EPERM;
5093 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5094 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5096 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5097 if (!can_nice(p, TASK_NICE(p)))
5098 return -EPERM;
5101 /* can't change other user's priorities */
5102 if (!check_same_owner(p))
5103 return -EPERM;
5105 /* Normal users shall not reset the sched_reset_on_fork flag */
5106 if (p->sched_reset_on_fork && !reset_on_fork)
5107 return -EPERM;
5110 if (user) {
5111 retval = security_task_setscheduler(p);
5112 if (retval)
5113 return retval;
5117 * make sure no PI-waiters arrive (or leave) while we are
5118 * changing the priority of the task:
5120 * To be able to change p->policy safely, the appropriate
5121 * runqueue lock must be held.
5123 rq = task_rq_lock(p, &flags);
5126 * Changing the policy of the stop threads its a very bad idea
5128 if (p == rq->stop) {
5129 task_rq_unlock(rq, p, &flags);
5130 return -EINVAL;
5134 * If not changing anything there's no need to proceed further:
5136 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5137 param->sched_priority == p->rt_priority))) {
5139 __task_rq_unlock(rq);
5140 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5141 return 0;
5144 #ifdef CONFIG_RT_GROUP_SCHED
5145 if (user) {
5147 * Do not allow realtime tasks into groups that have no runtime
5148 * assigned.
5150 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5151 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5152 !task_group_is_autogroup(task_group(p))) {
5153 task_rq_unlock(rq, p, &flags);
5154 return -EPERM;
5157 #endif
5159 /* recheck policy now with rq lock held */
5160 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5161 policy = oldpolicy = -1;
5162 task_rq_unlock(rq, p, &flags);
5163 goto recheck;
5165 on_rq = p->on_rq;
5166 running = task_current(rq, p);
5167 if (on_rq)
5168 deactivate_task(rq, p, 0);
5169 if (running)
5170 p->sched_class->put_prev_task(rq, p);
5172 p->sched_reset_on_fork = reset_on_fork;
5174 oldprio = p->prio;
5175 prev_class = p->sched_class;
5176 __setscheduler(rq, p, policy, param->sched_priority);
5178 if (running)
5179 p->sched_class->set_curr_task(rq);
5180 if (on_rq)
5181 activate_task(rq, p, 0);
5183 check_class_changed(rq, p, prev_class, oldprio);
5184 task_rq_unlock(rq, p, &flags);
5186 rt_mutex_adjust_pi(p);
5188 return 0;
5192 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5193 * @p: the task in question.
5194 * @policy: new policy.
5195 * @param: structure containing the new RT priority.
5197 * NOTE that the task may be already dead.
5199 int sched_setscheduler(struct task_struct *p, int policy,
5200 const struct sched_param *param)
5202 return __sched_setscheduler(p, policy, param, true);
5204 EXPORT_SYMBOL_GPL(sched_setscheduler);
5207 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5208 * @p: the task in question.
5209 * @policy: new policy.
5210 * @param: structure containing the new RT priority.
5212 * Just like sched_setscheduler, only don't bother checking if the
5213 * current context has permission. For example, this is needed in
5214 * stop_machine(): we create temporary high priority worker threads,
5215 * but our caller might not have that capability.
5217 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5218 const struct sched_param *param)
5220 return __sched_setscheduler(p, policy, param, false);
5223 static int
5224 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5226 struct sched_param lparam;
5227 struct task_struct *p;
5228 int retval;
5230 if (!param || pid < 0)
5231 return -EINVAL;
5232 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5233 return -EFAULT;
5235 rcu_read_lock();
5236 retval = -ESRCH;
5237 p = find_process_by_pid(pid);
5238 if (p != NULL)
5239 retval = sched_setscheduler(p, policy, &lparam);
5240 rcu_read_unlock();
5242 return retval;
5246 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5247 * @pid: the pid in question.
5248 * @policy: new policy.
5249 * @param: structure containing the new RT priority.
5251 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5252 struct sched_param __user *, param)
5254 /* negative values for policy are not valid */
5255 if (policy < 0)
5256 return -EINVAL;
5258 return do_sched_setscheduler(pid, policy, param);
5262 * sys_sched_setparam - set/change the RT priority of a thread
5263 * @pid: the pid in question.
5264 * @param: structure containing the new RT priority.
5266 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5268 return do_sched_setscheduler(pid, -1, param);
5272 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5273 * @pid: the pid in question.
5275 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5277 struct task_struct *p;
5278 int retval;
5280 if (pid < 0)
5281 return -EINVAL;
5283 retval = -ESRCH;
5284 rcu_read_lock();
5285 p = find_process_by_pid(pid);
5286 if (p) {
5287 retval = security_task_getscheduler(p);
5288 if (!retval)
5289 retval = p->policy
5290 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5292 rcu_read_unlock();
5293 return retval;
5297 * sys_sched_getparam - get the RT priority of a thread
5298 * @pid: the pid in question.
5299 * @param: structure containing the RT priority.
5301 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5303 struct sched_param lp;
5304 struct task_struct *p;
5305 int retval;
5307 if (!param || pid < 0)
5308 return -EINVAL;
5310 rcu_read_lock();
5311 p = find_process_by_pid(pid);
5312 retval = -ESRCH;
5313 if (!p)
5314 goto out_unlock;
5316 retval = security_task_getscheduler(p);
5317 if (retval)
5318 goto out_unlock;
5320 lp.sched_priority = p->rt_priority;
5321 rcu_read_unlock();
5324 * This one might sleep, we cannot do it with a spinlock held ...
5326 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5328 return retval;
5330 out_unlock:
5331 rcu_read_unlock();
5332 return retval;
5335 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5337 cpumask_var_t cpus_allowed, new_mask;
5338 struct task_struct *p;
5339 int retval;
5341 get_online_cpus();
5342 rcu_read_lock();
5344 p = find_process_by_pid(pid);
5345 if (!p) {
5346 rcu_read_unlock();
5347 put_online_cpus();
5348 return -ESRCH;
5351 /* Prevent p going away */
5352 get_task_struct(p);
5353 rcu_read_unlock();
5355 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5356 retval = -ENOMEM;
5357 goto out_put_task;
5359 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5360 retval = -ENOMEM;
5361 goto out_free_cpus_allowed;
5363 retval = -EPERM;
5364 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5365 goto out_unlock;
5367 retval = security_task_setscheduler(p);
5368 if (retval)
5369 goto out_unlock;
5371 cpuset_cpus_allowed(p, cpus_allowed);
5372 cpumask_and(new_mask, in_mask, cpus_allowed);
5373 again:
5374 retval = set_cpus_allowed_ptr(p, new_mask);
5376 if (!retval) {
5377 cpuset_cpus_allowed(p, cpus_allowed);
5378 if (!cpumask_subset(new_mask, cpus_allowed)) {
5380 * We must have raced with a concurrent cpuset
5381 * update. Just reset the cpus_allowed to the
5382 * cpuset's cpus_allowed
5384 cpumask_copy(new_mask, cpus_allowed);
5385 goto again;
5388 out_unlock:
5389 free_cpumask_var(new_mask);
5390 out_free_cpus_allowed:
5391 free_cpumask_var(cpus_allowed);
5392 out_put_task:
5393 put_task_struct(p);
5394 put_online_cpus();
5395 return retval;
5398 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5399 struct cpumask *new_mask)
5401 if (len < cpumask_size())
5402 cpumask_clear(new_mask);
5403 else if (len > cpumask_size())
5404 len = cpumask_size();
5406 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5410 * sys_sched_setaffinity - set the cpu affinity of a process
5411 * @pid: pid of the process
5412 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5413 * @user_mask_ptr: user-space pointer to the new cpu mask
5415 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5416 unsigned long __user *, user_mask_ptr)
5418 cpumask_var_t new_mask;
5419 int retval;
5421 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5422 return -ENOMEM;
5424 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5425 if (retval == 0)
5426 retval = sched_setaffinity(pid, new_mask);
5427 free_cpumask_var(new_mask);
5428 return retval;
5431 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5433 struct task_struct *p;
5434 unsigned long flags;
5435 int retval;
5437 get_online_cpus();
5438 rcu_read_lock();
5440 retval = -ESRCH;
5441 p = find_process_by_pid(pid);
5442 if (!p)
5443 goto out_unlock;
5445 retval = security_task_getscheduler(p);
5446 if (retval)
5447 goto out_unlock;
5449 raw_spin_lock_irqsave(&p->pi_lock, flags);
5450 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5451 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5453 out_unlock:
5454 rcu_read_unlock();
5455 put_online_cpus();
5457 return retval;
5461 * sys_sched_getaffinity - get the cpu affinity of a process
5462 * @pid: pid of the process
5463 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5464 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5466 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5467 unsigned long __user *, user_mask_ptr)
5469 int ret;
5470 cpumask_var_t mask;
5472 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5473 return -EINVAL;
5474 if (len & (sizeof(unsigned long)-1))
5475 return -EINVAL;
5477 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5478 return -ENOMEM;
5480 ret = sched_getaffinity(pid, mask);
5481 if (ret == 0) {
5482 size_t retlen = min_t(size_t, len, cpumask_size());
5484 if (copy_to_user(user_mask_ptr, mask, retlen))
5485 ret = -EFAULT;
5486 else
5487 ret = retlen;
5489 free_cpumask_var(mask);
5491 return ret;
5495 * sys_sched_yield - yield the current processor to other threads.
5497 * This function yields the current CPU to other tasks. If there are no
5498 * other threads running on this CPU then this function will return.
5500 SYSCALL_DEFINE0(sched_yield)
5502 struct rq *rq = this_rq_lock();
5504 schedstat_inc(rq, yld_count);
5505 current->sched_class->yield_task(rq);
5508 * Since we are going to call schedule() anyway, there's
5509 * no need to preempt or enable interrupts:
5511 __release(rq->lock);
5512 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5513 do_raw_spin_unlock(&rq->lock);
5514 preempt_enable_no_resched();
5516 schedule();
5518 return 0;
5521 static inline int should_resched(void)
5523 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5526 static void __cond_resched(void)
5528 add_preempt_count(PREEMPT_ACTIVE);
5529 schedule();
5530 sub_preempt_count(PREEMPT_ACTIVE);
5533 int __sched _cond_resched(void)
5535 if (should_resched()) {
5536 __cond_resched();
5537 return 1;
5539 return 0;
5541 EXPORT_SYMBOL(_cond_resched);
5544 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5545 * call schedule, and on return reacquire the lock.
5547 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5548 * operations here to prevent schedule() from being called twice (once via
5549 * spin_unlock(), once by hand).
5551 int __cond_resched_lock(spinlock_t *lock)
5553 int resched = should_resched();
5554 int ret = 0;
5556 lockdep_assert_held(lock);
5558 if (spin_needbreak(lock) || resched) {
5559 spin_unlock(lock);
5560 if (resched)
5561 __cond_resched();
5562 else
5563 cpu_relax();
5564 ret = 1;
5565 spin_lock(lock);
5567 return ret;
5569 EXPORT_SYMBOL(__cond_resched_lock);
5571 int __sched __cond_resched_softirq(void)
5573 BUG_ON(!in_softirq());
5575 if (should_resched()) {
5576 local_bh_enable();
5577 __cond_resched();
5578 local_bh_disable();
5579 return 1;
5581 return 0;
5583 EXPORT_SYMBOL(__cond_resched_softirq);
5586 * yield - yield the current processor to other threads.
5588 * This is a shortcut for kernel-space yielding - it marks the
5589 * thread runnable and calls sys_sched_yield().
5591 void __sched yield(void)
5593 set_current_state(TASK_RUNNING);
5594 sys_sched_yield();
5596 EXPORT_SYMBOL(yield);
5599 * yield_to - yield the current processor to another thread in
5600 * your thread group, or accelerate that thread toward the
5601 * processor it's on.
5602 * @p: target task
5603 * @preempt: whether task preemption is allowed or not
5605 * It's the caller's job to ensure that the target task struct
5606 * can't go away on us before we can do any checks.
5608 * Returns true if we indeed boosted the target task.
5610 bool __sched yield_to(struct task_struct *p, bool preempt)
5612 struct task_struct *curr = current;
5613 struct rq *rq, *p_rq;
5614 unsigned long flags;
5615 bool yielded = 0;
5617 local_irq_save(flags);
5618 rq = this_rq();
5620 again:
5621 p_rq = task_rq(p);
5622 double_rq_lock(rq, p_rq);
5623 while (task_rq(p) != p_rq) {
5624 double_rq_unlock(rq, p_rq);
5625 goto again;
5628 if (!curr->sched_class->yield_to_task)
5629 goto out;
5631 if (curr->sched_class != p->sched_class)
5632 goto out;
5634 if (task_running(p_rq, p) || p->state)
5635 goto out;
5637 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5638 if (yielded) {
5639 schedstat_inc(rq, yld_count);
5641 * Make p's CPU reschedule; pick_next_entity takes care of
5642 * fairness.
5644 if (preempt && rq != p_rq)
5645 resched_task(p_rq->curr);
5648 out:
5649 double_rq_unlock(rq, p_rq);
5650 local_irq_restore(flags);
5652 if (yielded)
5653 schedule();
5655 return yielded;
5657 EXPORT_SYMBOL_GPL(yield_to);
5660 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5661 * that process accounting knows that this is a task in IO wait state.
5663 void __sched io_schedule(void)
5665 struct rq *rq = raw_rq();
5667 delayacct_blkio_start();
5668 atomic_inc(&rq->nr_iowait);
5669 blk_flush_plug(current);
5670 current->in_iowait = 1;
5671 schedule();
5672 current->in_iowait = 0;
5673 atomic_dec(&rq->nr_iowait);
5674 delayacct_blkio_end();
5676 EXPORT_SYMBOL(io_schedule);
5678 long __sched io_schedule_timeout(long timeout)
5680 struct rq *rq = raw_rq();
5681 long ret;
5683 delayacct_blkio_start();
5684 atomic_inc(&rq->nr_iowait);
5685 blk_flush_plug(current);
5686 current->in_iowait = 1;
5687 ret = schedule_timeout(timeout);
5688 current->in_iowait = 0;
5689 atomic_dec(&rq->nr_iowait);
5690 delayacct_blkio_end();
5691 return ret;
5695 * sys_sched_get_priority_max - return maximum RT priority.
5696 * @policy: scheduling class.
5698 * this syscall returns the maximum rt_priority that can be used
5699 * by a given scheduling class.
5701 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5703 int ret = -EINVAL;
5705 switch (policy) {
5706 case SCHED_FIFO:
5707 case SCHED_RR:
5708 ret = MAX_USER_RT_PRIO-1;
5709 break;
5710 case SCHED_NORMAL:
5711 case SCHED_BATCH:
5712 case SCHED_IDLE:
5713 ret = 0;
5714 break;
5716 return ret;
5720 * sys_sched_get_priority_min - return minimum RT priority.
5721 * @policy: scheduling class.
5723 * this syscall returns the minimum rt_priority that can be used
5724 * by a given scheduling class.
5726 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5728 int ret = -EINVAL;
5730 switch (policy) {
5731 case SCHED_FIFO:
5732 case SCHED_RR:
5733 ret = 1;
5734 break;
5735 case SCHED_NORMAL:
5736 case SCHED_BATCH:
5737 case SCHED_IDLE:
5738 ret = 0;
5740 return ret;
5744 * sys_sched_rr_get_interval - return the default timeslice of a process.
5745 * @pid: pid of the process.
5746 * @interval: userspace pointer to the timeslice value.
5748 * this syscall writes the default timeslice value of a given process
5749 * into the user-space timespec buffer. A value of '0' means infinity.
5751 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5752 struct timespec __user *, interval)
5754 struct task_struct *p;
5755 unsigned int time_slice;
5756 unsigned long flags;
5757 struct rq *rq;
5758 int retval;
5759 struct timespec t;
5761 if (pid < 0)
5762 return -EINVAL;
5764 retval = -ESRCH;
5765 rcu_read_lock();
5766 p = find_process_by_pid(pid);
5767 if (!p)
5768 goto out_unlock;
5770 retval = security_task_getscheduler(p);
5771 if (retval)
5772 goto out_unlock;
5774 rq = task_rq_lock(p, &flags);
5775 time_slice = p->sched_class->get_rr_interval(rq, p);
5776 task_rq_unlock(rq, p, &flags);
5778 rcu_read_unlock();
5779 jiffies_to_timespec(time_slice, &t);
5780 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5781 return retval;
5783 out_unlock:
5784 rcu_read_unlock();
5785 return retval;
5788 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5790 void sched_show_task(struct task_struct *p)
5792 unsigned long free = 0;
5793 unsigned state;
5795 state = p->state ? __ffs(p->state) + 1 : 0;
5796 printk(KERN_INFO "%-15.15s %c", p->comm,
5797 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5798 #if BITS_PER_LONG == 32
5799 if (state == TASK_RUNNING)
5800 printk(KERN_CONT " running ");
5801 else
5802 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5803 #else
5804 if (state == TASK_RUNNING)
5805 printk(KERN_CONT " running task ");
5806 else
5807 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5808 #endif
5809 #ifdef CONFIG_DEBUG_STACK_USAGE
5810 free = stack_not_used(p);
5811 #endif
5812 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5813 task_pid_nr(p), task_pid_nr(p->real_parent),
5814 (unsigned long)task_thread_info(p)->flags);
5816 show_stack(p, NULL);
5819 void show_state_filter(unsigned long state_filter)
5821 struct task_struct *g, *p;
5823 #if BITS_PER_LONG == 32
5824 printk(KERN_INFO
5825 " task PC stack pid father\n");
5826 #else
5827 printk(KERN_INFO
5828 " task PC stack pid father\n");
5829 #endif
5830 read_lock(&tasklist_lock);
5831 do_each_thread(g, p) {
5833 * reset the NMI-timeout, listing all files on a slow
5834 * console might take a lot of time:
5836 touch_nmi_watchdog();
5837 if (!state_filter || (p->state & state_filter))
5838 sched_show_task(p);
5839 } while_each_thread(g, p);
5841 touch_all_softlockup_watchdogs();
5843 #ifdef CONFIG_SCHED_DEBUG
5844 sysrq_sched_debug_show();
5845 #endif
5846 read_unlock(&tasklist_lock);
5848 * Only show locks if all tasks are dumped:
5850 if (!state_filter)
5851 debug_show_all_locks();
5854 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5856 idle->sched_class = &idle_sched_class;
5860 * init_idle - set up an idle thread for a given CPU
5861 * @idle: task in question
5862 * @cpu: cpu the idle task belongs to
5864 * NOTE: this function does not set the idle thread's NEED_RESCHED
5865 * flag, to make booting more robust.
5867 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5869 struct rq *rq = cpu_rq(cpu);
5870 unsigned long flags;
5872 raw_spin_lock_irqsave(&rq->lock, flags);
5874 __sched_fork(idle);
5875 idle->state = TASK_RUNNING;
5876 idle->se.exec_start = sched_clock();
5878 do_set_cpus_allowed(idle, cpumask_of(cpu));
5880 * We're having a chicken and egg problem, even though we are
5881 * holding rq->lock, the cpu isn't yet set to this cpu so the
5882 * lockdep check in task_group() will fail.
5884 * Similar case to sched_fork(). / Alternatively we could
5885 * use task_rq_lock() here and obtain the other rq->lock.
5887 * Silence PROVE_RCU
5889 rcu_read_lock();
5890 __set_task_cpu(idle, cpu);
5891 rcu_read_unlock();
5893 rq->curr = rq->idle = idle;
5894 #if defined(CONFIG_SMP)
5895 idle->on_cpu = 1;
5896 #endif
5897 raw_spin_unlock_irqrestore(&rq->lock, flags);
5899 /* Set the preempt count _outside_ the spinlocks! */
5900 task_thread_info(idle)->preempt_count = 0;
5903 * The idle tasks have their own, simple scheduling class:
5905 idle->sched_class = &idle_sched_class;
5906 ftrace_graph_init_idle_task(idle, cpu);
5910 * In a system that switches off the HZ timer nohz_cpu_mask
5911 * indicates which cpus entered this state. This is used
5912 * in the rcu update to wait only for active cpus. For system
5913 * which do not switch off the HZ timer nohz_cpu_mask should
5914 * always be CPU_BITS_NONE.
5916 cpumask_var_t nohz_cpu_mask;
5919 * Increase the granularity value when there are more CPUs,
5920 * because with more CPUs the 'effective latency' as visible
5921 * to users decreases. But the relationship is not linear,
5922 * so pick a second-best guess by going with the log2 of the
5923 * number of CPUs.
5925 * This idea comes from the SD scheduler of Con Kolivas:
5927 static int get_update_sysctl_factor(void)
5929 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5930 unsigned int factor;
5932 switch (sysctl_sched_tunable_scaling) {
5933 case SCHED_TUNABLESCALING_NONE:
5934 factor = 1;
5935 break;
5936 case SCHED_TUNABLESCALING_LINEAR:
5937 factor = cpus;
5938 break;
5939 case SCHED_TUNABLESCALING_LOG:
5940 default:
5941 factor = 1 + ilog2(cpus);
5942 break;
5945 return factor;
5948 static void update_sysctl(void)
5950 unsigned int factor = get_update_sysctl_factor();
5952 #define SET_SYSCTL(name) \
5953 (sysctl_##name = (factor) * normalized_sysctl_##name)
5954 SET_SYSCTL(sched_min_granularity);
5955 SET_SYSCTL(sched_latency);
5956 SET_SYSCTL(sched_wakeup_granularity);
5957 #undef SET_SYSCTL
5960 static inline void sched_init_granularity(void)
5962 update_sysctl();
5965 #ifdef CONFIG_SMP
5966 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
5968 if (p->sched_class && p->sched_class->set_cpus_allowed)
5969 p->sched_class->set_cpus_allowed(p, new_mask);
5970 else {
5971 cpumask_copy(&p->cpus_allowed, new_mask);
5972 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5977 * This is how migration works:
5979 * 1) we invoke migration_cpu_stop() on the target CPU using
5980 * stop_one_cpu().
5981 * 2) stopper starts to run (implicitly forcing the migrated thread
5982 * off the CPU)
5983 * 3) it checks whether the migrated task is still in the wrong runqueue.
5984 * 4) if it's in the wrong runqueue then the migration thread removes
5985 * it and puts it into the right queue.
5986 * 5) stopper completes and stop_one_cpu() returns and the migration
5987 * is done.
5991 * Change a given task's CPU affinity. Migrate the thread to a
5992 * proper CPU and schedule it away if the CPU it's executing on
5993 * is removed from the allowed bitmask.
5995 * NOTE: the caller must have a valid reference to the task, the
5996 * task must not exit() & deallocate itself prematurely. The
5997 * call is not atomic; no spinlocks may be held.
5999 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6001 unsigned long flags;
6002 struct rq *rq;
6003 unsigned int dest_cpu;
6004 int ret = 0;
6006 rq = task_rq_lock(p, &flags);
6008 if (cpumask_equal(&p->cpus_allowed, new_mask))
6009 goto out;
6011 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6012 ret = -EINVAL;
6013 goto out;
6016 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6017 ret = -EINVAL;
6018 goto out;
6021 do_set_cpus_allowed(p, new_mask);
6023 /* Can the task run on the task's current CPU? If so, we're done */
6024 if (cpumask_test_cpu(task_cpu(p), new_mask))
6025 goto out;
6027 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6028 if (p->on_rq) {
6029 struct migration_arg arg = { p, dest_cpu };
6030 /* Need help from migration thread: drop lock and wait. */
6031 task_rq_unlock(rq, p, &flags);
6032 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6033 tlb_migrate_finish(p->mm);
6034 return 0;
6036 out:
6037 task_rq_unlock(rq, p, &flags);
6039 return ret;
6041 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6044 * Move (not current) task off this cpu, onto dest cpu. We're doing
6045 * this because either it can't run here any more (set_cpus_allowed()
6046 * away from this CPU, or CPU going down), or because we're
6047 * attempting to rebalance this task on exec (sched_exec).
6049 * So we race with normal scheduler movements, but that's OK, as long
6050 * as the task is no longer on this CPU.
6052 * Returns non-zero if task was successfully migrated.
6054 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6056 struct rq *rq_dest, *rq_src;
6057 int ret = 0;
6059 if (unlikely(!cpu_active(dest_cpu)))
6060 return ret;
6062 rq_src = cpu_rq(src_cpu);
6063 rq_dest = cpu_rq(dest_cpu);
6065 raw_spin_lock(&p->pi_lock);
6066 double_rq_lock(rq_src, rq_dest);
6067 /* Already moved. */
6068 if (task_cpu(p) != src_cpu)
6069 goto done;
6070 /* Affinity changed (again). */
6071 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6072 goto fail;
6075 * If we're not on a rq, the next wake-up will ensure we're
6076 * placed properly.
6078 if (p->on_rq) {
6079 deactivate_task(rq_src, p, 0);
6080 set_task_cpu(p, dest_cpu);
6081 activate_task(rq_dest, p, 0);
6082 check_preempt_curr(rq_dest, p, 0);
6084 done:
6085 ret = 1;
6086 fail:
6087 double_rq_unlock(rq_src, rq_dest);
6088 raw_spin_unlock(&p->pi_lock);
6089 return ret;
6093 * migration_cpu_stop - this will be executed by a highprio stopper thread
6094 * and performs thread migration by bumping thread off CPU then
6095 * 'pushing' onto another runqueue.
6097 static int migration_cpu_stop(void *data)
6099 struct migration_arg *arg = data;
6102 * The original target cpu might have gone down and we might
6103 * be on another cpu but it doesn't matter.
6105 local_irq_disable();
6106 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6107 local_irq_enable();
6108 return 0;
6111 #ifdef CONFIG_HOTPLUG_CPU
6114 * Ensures that the idle task is using init_mm right before its cpu goes
6115 * offline.
6117 void idle_task_exit(void)
6119 struct mm_struct *mm = current->active_mm;
6121 BUG_ON(cpu_online(smp_processor_id()));
6123 if (mm != &init_mm)
6124 switch_mm(mm, &init_mm, current);
6125 mmdrop(mm);
6129 * While a dead CPU has no uninterruptible tasks queued at this point,
6130 * it might still have a nonzero ->nr_uninterruptible counter, because
6131 * for performance reasons the counter is not stricly tracking tasks to
6132 * their home CPUs. So we just add the counter to another CPU's counter,
6133 * to keep the global sum constant after CPU-down:
6135 static void migrate_nr_uninterruptible(struct rq *rq_src)
6137 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6139 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6140 rq_src->nr_uninterruptible = 0;
6144 * remove the tasks which were accounted by rq from calc_load_tasks.
6146 static void calc_global_load_remove(struct rq *rq)
6148 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6149 rq->calc_load_active = 0;
6153 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6154 * try_to_wake_up()->select_task_rq().
6156 * Called with rq->lock held even though we'er in stop_machine() and
6157 * there's no concurrency possible, we hold the required locks anyway
6158 * because of lock validation efforts.
6160 static void migrate_tasks(unsigned int dead_cpu)
6162 struct rq *rq = cpu_rq(dead_cpu);
6163 struct task_struct *next, *stop = rq->stop;
6164 int dest_cpu;
6167 * Fudge the rq selection such that the below task selection loop
6168 * doesn't get stuck on the currently eligible stop task.
6170 * We're currently inside stop_machine() and the rq is either stuck
6171 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6172 * either way we should never end up calling schedule() until we're
6173 * done here.
6175 rq->stop = NULL;
6177 for ( ; ; ) {
6179 * There's this thread running, bail when that's the only
6180 * remaining thread.
6182 if (rq->nr_running == 1)
6183 break;
6185 next = pick_next_task(rq);
6186 BUG_ON(!next);
6187 next->sched_class->put_prev_task(rq, next);
6189 /* Find suitable destination for @next, with force if needed. */
6190 dest_cpu = select_fallback_rq(dead_cpu, next);
6191 raw_spin_unlock(&rq->lock);
6193 __migrate_task(next, dead_cpu, dest_cpu);
6195 raw_spin_lock(&rq->lock);
6198 rq->stop = stop;
6201 #endif /* CONFIG_HOTPLUG_CPU */
6203 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6205 static struct ctl_table sd_ctl_dir[] = {
6207 .procname = "sched_domain",
6208 .mode = 0555,
6213 static struct ctl_table sd_ctl_root[] = {
6215 .procname = "kernel",
6216 .mode = 0555,
6217 .child = sd_ctl_dir,
6222 static struct ctl_table *sd_alloc_ctl_entry(int n)
6224 struct ctl_table *entry =
6225 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6227 return entry;
6230 static void sd_free_ctl_entry(struct ctl_table **tablep)
6232 struct ctl_table *entry;
6235 * In the intermediate directories, both the child directory and
6236 * procname are dynamically allocated and could fail but the mode
6237 * will always be set. In the lowest directory the names are
6238 * static strings and all have proc handlers.
6240 for (entry = *tablep; entry->mode; entry++) {
6241 if (entry->child)
6242 sd_free_ctl_entry(&entry->child);
6243 if (entry->proc_handler == NULL)
6244 kfree(entry->procname);
6247 kfree(*tablep);
6248 *tablep = NULL;
6251 static void
6252 set_table_entry(struct ctl_table *entry,
6253 const char *procname, void *data, int maxlen,
6254 mode_t mode, proc_handler *proc_handler)
6256 entry->procname = procname;
6257 entry->data = data;
6258 entry->maxlen = maxlen;
6259 entry->mode = mode;
6260 entry->proc_handler = proc_handler;
6263 static struct ctl_table *
6264 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6266 struct ctl_table *table = sd_alloc_ctl_entry(13);
6268 if (table == NULL)
6269 return NULL;
6271 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6272 sizeof(long), 0644, proc_doulongvec_minmax);
6273 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6274 sizeof(long), 0644, proc_doulongvec_minmax);
6275 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6276 sizeof(int), 0644, proc_dointvec_minmax);
6277 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6278 sizeof(int), 0644, proc_dointvec_minmax);
6279 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6280 sizeof(int), 0644, proc_dointvec_minmax);
6281 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6282 sizeof(int), 0644, proc_dointvec_minmax);
6283 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6284 sizeof(int), 0644, proc_dointvec_minmax);
6285 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6286 sizeof(int), 0644, proc_dointvec_minmax);
6287 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6288 sizeof(int), 0644, proc_dointvec_minmax);
6289 set_table_entry(&table[9], "cache_nice_tries",
6290 &sd->cache_nice_tries,
6291 sizeof(int), 0644, proc_dointvec_minmax);
6292 set_table_entry(&table[10], "flags", &sd->flags,
6293 sizeof(int), 0644, proc_dointvec_minmax);
6294 set_table_entry(&table[11], "name", sd->name,
6295 CORENAME_MAX_SIZE, 0444, proc_dostring);
6296 /* &table[12] is terminator */
6298 return table;
6301 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6303 struct ctl_table *entry, *table;
6304 struct sched_domain *sd;
6305 int domain_num = 0, i;
6306 char buf[32];
6308 for_each_domain(cpu, sd)
6309 domain_num++;
6310 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6311 if (table == NULL)
6312 return NULL;
6314 i = 0;
6315 for_each_domain(cpu, sd) {
6316 snprintf(buf, 32, "domain%d", i);
6317 entry->procname = kstrdup(buf, GFP_KERNEL);
6318 entry->mode = 0555;
6319 entry->child = sd_alloc_ctl_domain_table(sd);
6320 entry++;
6321 i++;
6323 return table;
6326 static struct ctl_table_header *sd_sysctl_header;
6327 static void register_sched_domain_sysctl(void)
6329 int i, cpu_num = num_possible_cpus();
6330 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6331 char buf[32];
6333 WARN_ON(sd_ctl_dir[0].child);
6334 sd_ctl_dir[0].child = entry;
6336 if (entry == NULL)
6337 return;
6339 for_each_possible_cpu(i) {
6340 snprintf(buf, 32, "cpu%d", i);
6341 entry->procname = kstrdup(buf, GFP_KERNEL);
6342 entry->mode = 0555;
6343 entry->child = sd_alloc_ctl_cpu_table(i);
6344 entry++;
6347 WARN_ON(sd_sysctl_header);
6348 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6351 /* may be called multiple times per register */
6352 static void unregister_sched_domain_sysctl(void)
6354 if (sd_sysctl_header)
6355 unregister_sysctl_table(sd_sysctl_header);
6356 sd_sysctl_header = NULL;
6357 if (sd_ctl_dir[0].child)
6358 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6360 #else
6361 static void register_sched_domain_sysctl(void)
6364 static void unregister_sched_domain_sysctl(void)
6367 #endif
6369 static void set_rq_online(struct rq *rq)
6371 if (!rq->online) {
6372 const struct sched_class *class;
6374 cpumask_set_cpu(rq->cpu, rq->rd->online);
6375 rq->online = 1;
6377 for_each_class(class) {
6378 if (class->rq_online)
6379 class->rq_online(rq);
6384 static void set_rq_offline(struct rq *rq)
6386 if (rq->online) {
6387 const struct sched_class *class;
6389 for_each_class(class) {
6390 if (class->rq_offline)
6391 class->rq_offline(rq);
6394 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6395 rq->online = 0;
6400 * migration_call - callback that gets triggered when a CPU is added.
6401 * Here we can start up the necessary migration thread for the new CPU.
6403 static int __cpuinit
6404 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6406 int cpu = (long)hcpu;
6407 unsigned long flags;
6408 struct rq *rq = cpu_rq(cpu);
6410 switch (action & ~CPU_TASKS_FROZEN) {
6412 case CPU_UP_PREPARE:
6413 rq->calc_load_update = calc_load_update;
6414 break;
6416 case CPU_ONLINE:
6417 /* Update our root-domain */
6418 raw_spin_lock_irqsave(&rq->lock, flags);
6419 if (rq->rd) {
6420 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6422 set_rq_online(rq);
6424 raw_spin_unlock_irqrestore(&rq->lock, flags);
6425 break;
6427 #ifdef CONFIG_HOTPLUG_CPU
6428 case CPU_DYING:
6429 sched_ttwu_pending();
6430 /* Update our root-domain */
6431 raw_spin_lock_irqsave(&rq->lock, flags);
6432 if (rq->rd) {
6433 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6434 set_rq_offline(rq);
6436 migrate_tasks(cpu);
6437 BUG_ON(rq->nr_running != 1); /* the migration thread */
6438 raw_spin_unlock_irqrestore(&rq->lock, flags);
6440 migrate_nr_uninterruptible(rq);
6441 calc_global_load_remove(rq);
6442 break;
6443 #endif
6446 update_max_interval();
6448 return NOTIFY_OK;
6452 * Register at high priority so that task migration (migrate_all_tasks)
6453 * happens before everything else. This has to be lower priority than
6454 * the notifier in the perf_event subsystem, though.
6456 static struct notifier_block __cpuinitdata migration_notifier = {
6457 .notifier_call = migration_call,
6458 .priority = CPU_PRI_MIGRATION,
6461 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6462 unsigned long action, void *hcpu)
6464 switch (action & ~CPU_TASKS_FROZEN) {
6465 case CPU_ONLINE:
6466 case CPU_DOWN_FAILED:
6467 set_cpu_active((long)hcpu, true);
6468 return NOTIFY_OK;
6469 default:
6470 return NOTIFY_DONE;
6474 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6475 unsigned long action, void *hcpu)
6477 switch (action & ~CPU_TASKS_FROZEN) {
6478 case CPU_DOWN_PREPARE:
6479 set_cpu_active((long)hcpu, false);
6480 return NOTIFY_OK;
6481 default:
6482 return NOTIFY_DONE;
6486 static int __init migration_init(void)
6488 void *cpu = (void *)(long)smp_processor_id();
6489 int err;
6491 /* Initialize migration for the boot CPU */
6492 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6493 BUG_ON(err == NOTIFY_BAD);
6494 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6495 register_cpu_notifier(&migration_notifier);
6497 /* Register cpu active notifiers */
6498 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6499 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6501 return 0;
6503 early_initcall(migration_init);
6504 #endif
6506 #ifdef CONFIG_SMP
6508 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6510 #ifdef CONFIG_SCHED_DEBUG
6512 static __read_mostly int sched_domain_debug_enabled;
6514 static int __init sched_domain_debug_setup(char *str)
6516 sched_domain_debug_enabled = 1;
6518 return 0;
6520 early_param("sched_debug", sched_domain_debug_setup);
6522 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6523 struct cpumask *groupmask)
6525 struct sched_group *group = sd->groups;
6526 char str[256];
6528 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6529 cpumask_clear(groupmask);
6531 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6533 if (!(sd->flags & SD_LOAD_BALANCE)) {
6534 printk("does not load-balance\n");
6535 if (sd->parent)
6536 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6537 " has parent");
6538 return -1;
6541 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6543 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6544 printk(KERN_ERR "ERROR: domain->span does not contain "
6545 "CPU%d\n", cpu);
6547 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6548 printk(KERN_ERR "ERROR: domain->groups does not contain"
6549 " CPU%d\n", cpu);
6552 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6553 do {
6554 if (!group) {
6555 printk("\n");
6556 printk(KERN_ERR "ERROR: group is NULL\n");
6557 break;
6560 if (!group->cpu_power) {
6561 printk(KERN_CONT "\n");
6562 printk(KERN_ERR "ERROR: domain->cpu_power not "
6563 "set\n");
6564 break;
6567 if (!cpumask_weight(sched_group_cpus(group))) {
6568 printk(KERN_CONT "\n");
6569 printk(KERN_ERR "ERROR: empty group\n");
6570 break;
6573 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6574 printk(KERN_CONT "\n");
6575 printk(KERN_ERR "ERROR: repeated CPUs\n");
6576 break;
6579 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6581 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6583 printk(KERN_CONT " %s", str);
6584 if (group->cpu_power != SCHED_POWER_SCALE) {
6585 printk(KERN_CONT " (cpu_power = %d)",
6586 group->cpu_power);
6589 group = group->next;
6590 } while (group != sd->groups);
6591 printk(KERN_CONT "\n");
6593 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6594 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6596 if (sd->parent &&
6597 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6598 printk(KERN_ERR "ERROR: parent span is not a superset "
6599 "of domain->span\n");
6600 return 0;
6603 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6605 int level = 0;
6607 if (!sched_domain_debug_enabled)
6608 return;
6610 if (!sd) {
6611 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6612 return;
6615 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6617 for (;;) {
6618 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6619 break;
6620 level++;
6621 sd = sd->parent;
6622 if (!sd)
6623 break;
6626 #else /* !CONFIG_SCHED_DEBUG */
6627 # define sched_domain_debug(sd, cpu) do { } while (0)
6628 #endif /* CONFIG_SCHED_DEBUG */
6630 static int sd_degenerate(struct sched_domain *sd)
6632 if (cpumask_weight(sched_domain_span(sd)) == 1)
6633 return 1;
6635 /* Following flags need at least 2 groups */
6636 if (sd->flags & (SD_LOAD_BALANCE |
6637 SD_BALANCE_NEWIDLE |
6638 SD_BALANCE_FORK |
6639 SD_BALANCE_EXEC |
6640 SD_SHARE_CPUPOWER |
6641 SD_SHARE_PKG_RESOURCES)) {
6642 if (sd->groups != sd->groups->next)
6643 return 0;
6646 /* Following flags don't use groups */
6647 if (sd->flags & (SD_WAKE_AFFINE))
6648 return 0;
6650 return 1;
6653 static int
6654 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6656 unsigned long cflags = sd->flags, pflags = parent->flags;
6658 if (sd_degenerate(parent))
6659 return 1;
6661 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6662 return 0;
6664 /* Flags needing groups don't count if only 1 group in parent */
6665 if (parent->groups == parent->groups->next) {
6666 pflags &= ~(SD_LOAD_BALANCE |
6667 SD_BALANCE_NEWIDLE |
6668 SD_BALANCE_FORK |
6669 SD_BALANCE_EXEC |
6670 SD_SHARE_CPUPOWER |
6671 SD_SHARE_PKG_RESOURCES);
6672 if (nr_node_ids == 1)
6673 pflags &= ~SD_SERIALIZE;
6675 if (~cflags & pflags)
6676 return 0;
6678 return 1;
6681 static void free_rootdomain(struct rcu_head *rcu)
6683 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6685 cpupri_cleanup(&rd->cpupri);
6686 free_cpumask_var(rd->rto_mask);
6687 free_cpumask_var(rd->online);
6688 free_cpumask_var(rd->span);
6689 kfree(rd);
6692 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6694 struct root_domain *old_rd = NULL;
6695 unsigned long flags;
6697 raw_spin_lock_irqsave(&rq->lock, flags);
6699 if (rq->rd) {
6700 old_rd = rq->rd;
6702 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6703 set_rq_offline(rq);
6705 cpumask_clear_cpu(rq->cpu, old_rd->span);
6708 * If we dont want to free the old_rt yet then
6709 * set old_rd to NULL to skip the freeing later
6710 * in this function:
6712 if (!atomic_dec_and_test(&old_rd->refcount))
6713 old_rd = NULL;
6716 atomic_inc(&rd->refcount);
6717 rq->rd = rd;
6719 cpumask_set_cpu(rq->cpu, rd->span);
6720 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6721 set_rq_online(rq);
6723 raw_spin_unlock_irqrestore(&rq->lock, flags);
6725 if (old_rd)
6726 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6729 static int init_rootdomain(struct root_domain *rd)
6731 memset(rd, 0, sizeof(*rd));
6733 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6734 goto out;
6735 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6736 goto free_span;
6737 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6738 goto free_online;
6740 if (cpupri_init(&rd->cpupri) != 0)
6741 goto free_rto_mask;
6742 return 0;
6744 free_rto_mask:
6745 free_cpumask_var(rd->rto_mask);
6746 free_online:
6747 free_cpumask_var(rd->online);
6748 free_span:
6749 free_cpumask_var(rd->span);
6750 out:
6751 return -ENOMEM;
6754 static void init_defrootdomain(void)
6756 init_rootdomain(&def_root_domain);
6758 atomic_set(&def_root_domain.refcount, 1);
6761 static struct root_domain *alloc_rootdomain(void)
6763 struct root_domain *rd;
6765 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6766 if (!rd)
6767 return NULL;
6769 if (init_rootdomain(rd) != 0) {
6770 kfree(rd);
6771 return NULL;
6774 return rd;
6777 static void free_sched_domain(struct rcu_head *rcu)
6779 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6780 if (atomic_dec_and_test(&sd->groups->ref))
6781 kfree(sd->groups);
6782 kfree(sd);
6785 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6787 call_rcu(&sd->rcu, free_sched_domain);
6790 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6792 for (; sd; sd = sd->parent)
6793 destroy_sched_domain(sd, cpu);
6797 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6798 * hold the hotplug lock.
6800 static void
6801 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6803 struct rq *rq = cpu_rq(cpu);
6804 struct sched_domain *tmp;
6806 /* Remove the sched domains which do not contribute to scheduling. */
6807 for (tmp = sd; tmp; ) {
6808 struct sched_domain *parent = tmp->parent;
6809 if (!parent)
6810 break;
6812 if (sd_parent_degenerate(tmp, parent)) {
6813 tmp->parent = parent->parent;
6814 if (parent->parent)
6815 parent->parent->child = tmp;
6816 destroy_sched_domain(parent, cpu);
6817 } else
6818 tmp = tmp->parent;
6821 if (sd && sd_degenerate(sd)) {
6822 tmp = sd;
6823 sd = sd->parent;
6824 destroy_sched_domain(tmp, cpu);
6825 if (sd)
6826 sd->child = NULL;
6829 sched_domain_debug(sd, cpu);
6831 rq_attach_root(rq, rd);
6832 tmp = rq->sd;
6833 rcu_assign_pointer(rq->sd, sd);
6834 destroy_sched_domains(tmp, cpu);
6837 /* cpus with isolated domains */
6838 static cpumask_var_t cpu_isolated_map;
6840 /* Setup the mask of cpus configured for isolated domains */
6841 static int __init isolated_cpu_setup(char *str)
6843 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6844 cpulist_parse(str, cpu_isolated_map);
6845 return 1;
6848 __setup("isolcpus=", isolated_cpu_setup);
6850 #define SD_NODES_PER_DOMAIN 16
6852 #ifdef CONFIG_NUMA
6855 * find_next_best_node - find the next node to include in a sched_domain
6856 * @node: node whose sched_domain we're building
6857 * @used_nodes: nodes already in the sched_domain
6859 * Find the next node to include in a given scheduling domain. Simply
6860 * finds the closest node not already in the @used_nodes map.
6862 * Should use nodemask_t.
6864 static int find_next_best_node(int node, nodemask_t *used_nodes)
6866 int i, n, val, min_val, best_node = -1;
6868 min_val = INT_MAX;
6870 for (i = 0; i < nr_node_ids; i++) {
6871 /* Start at @node */
6872 n = (node + i) % nr_node_ids;
6874 if (!nr_cpus_node(n))
6875 continue;
6877 /* Skip already used nodes */
6878 if (node_isset(n, *used_nodes))
6879 continue;
6881 /* Simple min distance search */
6882 val = node_distance(node, n);
6884 if (val < min_val) {
6885 min_val = val;
6886 best_node = n;
6890 if (best_node != -1)
6891 node_set(best_node, *used_nodes);
6892 return best_node;
6896 * sched_domain_node_span - get a cpumask for a node's sched_domain
6897 * @node: node whose cpumask we're constructing
6898 * @span: resulting cpumask
6900 * Given a node, construct a good cpumask for its sched_domain to span. It
6901 * should be one that prevents unnecessary balancing, but also spreads tasks
6902 * out optimally.
6904 static void sched_domain_node_span(int node, struct cpumask *span)
6906 nodemask_t used_nodes;
6907 int i;
6909 cpumask_clear(span);
6910 nodes_clear(used_nodes);
6912 cpumask_or(span, span, cpumask_of_node(node));
6913 node_set(node, used_nodes);
6915 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6916 int next_node = find_next_best_node(node, &used_nodes);
6917 if (next_node < 0)
6918 break;
6919 cpumask_or(span, span, cpumask_of_node(next_node));
6923 static const struct cpumask *cpu_node_mask(int cpu)
6925 lockdep_assert_held(&sched_domains_mutex);
6927 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
6929 return sched_domains_tmpmask;
6932 static const struct cpumask *cpu_allnodes_mask(int cpu)
6934 return cpu_possible_mask;
6936 #endif /* CONFIG_NUMA */
6938 static const struct cpumask *cpu_cpu_mask(int cpu)
6940 return cpumask_of_node(cpu_to_node(cpu));
6943 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6945 struct sd_data {
6946 struct sched_domain **__percpu sd;
6947 struct sched_group **__percpu sg;
6950 struct s_data {
6951 struct sched_domain ** __percpu sd;
6952 struct root_domain *rd;
6955 enum s_alloc {
6956 sa_rootdomain,
6957 sa_sd,
6958 sa_sd_storage,
6959 sa_none,
6962 struct sched_domain_topology_level;
6964 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
6965 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
6967 struct sched_domain_topology_level {
6968 sched_domain_init_f init;
6969 sched_domain_mask_f mask;
6970 struct sd_data data;
6974 * Assumes the sched_domain tree is fully constructed
6976 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6978 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6979 struct sched_domain *child = sd->child;
6981 if (child)
6982 cpu = cpumask_first(sched_domain_span(child));
6984 if (sg)
6985 *sg = *per_cpu_ptr(sdd->sg, cpu);
6987 return cpu;
6991 * build_sched_groups takes the cpumask we wish to span, and a pointer
6992 * to a function which identifies what group(along with sched group) a CPU
6993 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6994 * (due to the fact that we keep track of groups covered with a struct cpumask).
6996 * build_sched_groups will build a circular linked list of the groups
6997 * covered by the given span, and will set each group's ->cpumask correctly,
6998 * and ->cpu_power to 0.
7000 static void
7001 build_sched_groups(struct sched_domain *sd)
7003 struct sched_group *first = NULL, *last = NULL;
7004 struct sd_data *sdd = sd->private;
7005 const struct cpumask *span = sched_domain_span(sd);
7006 struct cpumask *covered;
7007 int i;
7009 lockdep_assert_held(&sched_domains_mutex);
7010 covered = sched_domains_tmpmask;
7012 cpumask_clear(covered);
7014 for_each_cpu(i, span) {
7015 struct sched_group *sg;
7016 int group = get_group(i, sdd, &sg);
7017 int j;
7019 if (cpumask_test_cpu(i, covered))
7020 continue;
7022 cpumask_clear(sched_group_cpus(sg));
7023 sg->cpu_power = 0;
7025 for_each_cpu(j, span) {
7026 if (get_group(j, sdd, NULL) != group)
7027 continue;
7029 cpumask_set_cpu(j, covered);
7030 cpumask_set_cpu(j, sched_group_cpus(sg));
7033 if (!first)
7034 first = sg;
7035 if (last)
7036 last->next = sg;
7037 last = sg;
7039 last->next = first;
7043 * Initialize sched groups cpu_power.
7045 * cpu_power indicates the capacity of sched group, which is used while
7046 * distributing the load between different sched groups in a sched domain.
7047 * Typically cpu_power for all the groups in a sched domain will be same unless
7048 * there are asymmetries in the topology. If there are asymmetries, group
7049 * having more cpu_power will pickup more load compared to the group having
7050 * less cpu_power.
7052 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7054 WARN_ON(!sd || !sd->groups);
7056 if (cpu != group_first_cpu(sd->groups))
7057 return;
7059 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
7061 update_group_power(sd, cpu);
7065 * Initializers for schedule domains
7066 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7069 #ifdef CONFIG_SCHED_DEBUG
7070 # define SD_INIT_NAME(sd, type) sd->name = #type
7071 #else
7072 # define SD_INIT_NAME(sd, type) do { } while (0)
7073 #endif
7075 #define SD_INIT_FUNC(type) \
7076 static noinline struct sched_domain * \
7077 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7079 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7080 *sd = SD_##type##_INIT; \
7081 SD_INIT_NAME(sd, type); \
7082 sd->private = &tl->data; \
7083 return sd; \
7086 SD_INIT_FUNC(CPU)
7087 #ifdef CONFIG_NUMA
7088 SD_INIT_FUNC(ALLNODES)
7089 SD_INIT_FUNC(NODE)
7090 #endif
7091 #ifdef CONFIG_SCHED_SMT
7092 SD_INIT_FUNC(SIBLING)
7093 #endif
7094 #ifdef CONFIG_SCHED_MC
7095 SD_INIT_FUNC(MC)
7096 #endif
7097 #ifdef CONFIG_SCHED_BOOK
7098 SD_INIT_FUNC(BOOK)
7099 #endif
7101 static int default_relax_domain_level = -1;
7102 int sched_domain_level_max;
7104 static int __init setup_relax_domain_level(char *str)
7106 unsigned long val;
7108 val = simple_strtoul(str, NULL, 0);
7109 if (val < sched_domain_level_max)
7110 default_relax_domain_level = val;
7112 return 1;
7114 __setup("relax_domain_level=", setup_relax_domain_level);
7116 static void set_domain_attribute(struct sched_domain *sd,
7117 struct sched_domain_attr *attr)
7119 int request;
7121 if (!attr || attr->relax_domain_level < 0) {
7122 if (default_relax_domain_level < 0)
7123 return;
7124 else
7125 request = default_relax_domain_level;
7126 } else
7127 request = attr->relax_domain_level;
7128 if (request < sd->level) {
7129 /* turn off idle balance on this domain */
7130 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7131 } else {
7132 /* turn on idle balance on this domain */
7133 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7137 static void __sdt_free(const struct cpumask *cpu_map);
7138 static int __sdt_alloc(const struct cpumask *cpu_map);
7140 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7141 const struct cpumask *cpu_map)
7143 switch (what) {
7144 case sa_rootdomain:
7145 if (!atomic_read(&d->rd->refcount))
7146 free_rootdomain(&d->rd->rcu); /* fall through */
7147 case sa_sd:
7148 free_percpu(d->sd); /* fall through */
7149 case sa_sd_storage:
7150 __sdt_free(cpu_map); /* fall through */
7151 case sa_none:
7152 break;
7156 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7157 const struct cpumask *cpu_map)
7159 memset(d, 0, sizeof(*d));
7161 if (__sdt_alloc(cpu_map))
7162 return sa_sd_storage;
7163 d->sd = alloc_percpu(struct sched_domain *);
7164 if (!d->sd)
7165 return sa_sd_storage;
7166 d->rd = alloc_rootdomain();
7167 if (!d->rd)
7168 return sa_sd;
7169 return sa_rootdomain;
7173 * NULL the sd_data elements we've used to build the sched_domain and
7174 * sched_group structure so that the subsequent __free_domain_allocs()
7175 * will not free the data we're using.
7177 static void claim_allocations(int cpu, struct sched_domain *sd)
7179 struct sd_data *sdd = sd->private;
7180 struct sched_group *sg = sd->groups;
7182 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7183 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7185 if (cpu == cpumask_first(sched_group_cpus(sg))) {
7186 WARN_ON_ONCE(*per_cpu_ptr(sdd->sg, cpu) != sg);
7187 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7191 #ifdef CONFIG_SCHED_SMT
7192 static const struct cpumask *cpu_smt_mask(int cpu)
7194 return topology_thread_cpumask(cpu);
7196 #endif
7199 * Topology list, bottom-up.
7201 static struct sched_domain_topology_level default_topology[] = {
7202 #ifdef CONFIG_SCHED_SMT
7203 { sd_init_SIBLING, cpu_smt_mask, },
7204 #endif
7205 #ifdef CONFIG_SCHED_MC
7206 { sd_init_MC, cpu_coregroup_mask, },
7207 #endif
7208 #ifdef CONFIG_SCHED_BOOK
7209 { sd_init_BOOK, cpu_book_mask, },
7210 #endif
7211 { sd_init_CPU, cpu_cpu_mask, },
7212 #ifdef CONFIG_NUMA
7213 { sd_init_NODE, cpu_node_mask, },
7214 { sd_init_ALLNODES, cpu_allnodes_mask, },
7215 #endif
7216 { NULL, },
7219 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7221 static int __sdt_alloc(const struct cpumask *cpu_map)
7223 struct sched_domain_topology_level *tl;
7224 int j;
7226 for (tl = sched_domain_topology; tl->init; tl++) {
7227 struct sd_data *sdd = &tl->data;
7229 sdd->sd = alloc_percpu(struct sched_domain *);
7230 if (!sdd->sd)
7231 return -ENOMEM;
7233 sdd->sg = alloc_percpu(struct sched_group *);
7234 if (!sdd->sg)
7235 return -ENOMEM;
7237 for_each_cpu(j, cpu_map) {
7238 struct sched_domain *sd;
7239 struct sched_group *sg;
7241 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7242 GFP_KERNEL, cpu_to_node(j));
7243 if (!sd)
7244 return -ENOMEM;
7246 *per_cpu_ptr(sdd->sd, j) = sd;
7248 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7249 GFP_KERNEL, cpu_to_node(j));
7250 if (!sg)
7251 return -ENOMEM;
7253 *per_cpu_ptr(sdd->sg, j) = sg;
7257 return 0;
7260 static void __sdt_free(const struct cpumask *cpu_map)
7262 struct sched_domain_topology_level *tl;
7263 int j;
7265 for (tl = sched_domain_topology; tl->init; tl++) {
7266 struct sd_data *sdd = &tl->data;
7268 for_each_cpu(j, cpu_map) {
7269 kfree(*per_cpu_ptr(sdd->sd, j));
7270 kfree(*per_cpu_ptr(sdd->sg, j));
7272 free_percpu(sdd->sd);
7273 free_percpu(sdd->sg);
7277 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7278 struct s_data *d, const struct cpumask *cpu_map,
7279 struct sched_domain_attr *attr, struct sched_domain *child,
7280 int cpu)
7282 struct sched_domain *sd = tl->init(tl, cpu);
7283 if (!sd)
7284 return child;
7286 set_domain_attribute(sd, attr);
7287 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7288 if (child) {
7289 sd->level = child->level + 1;
7290 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7291 child->parent = sd;
7293 sd->child = child;
7295 return sd;
7299 * Build sched domains for a given set of cpus and attach the sched domains
7300 * to the individual cpus
7302 static int build_sched_domains(const struct cpumask *cpu_map,
7303 struct sched_domain_attr *attr)
7305 enum s_alloc alloc_state = sa_none;
7306 struct sched_domain *sd;
7307 struct s_data d;
7308 int i, ret = -ENOMEM;
7310 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7311 if (alloc_state != sa_rootdomain)
7312 goto error;
7314 /* Set up domains for cpus specified by the cpu_map. */
7315 for_each_cpu(i, cpu_map) {
7316 struct sched_domain_topology_level *tl;
7318 sd = NULL;
7319 for (tl = sched_domain_topology; tl->init; tl++)
7320 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7322 while (sd->child)
7323 sd = sd->child;
7325 *per_cpu_ptr(d.sd, i) = sd;
7328 /* Build the groups for the domains */
7329 for_each_cpu(i, cpu_map) {
7330 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7331 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7332 get_group(i, sd->private, &sd->groups);
7333 atomic_inc(&sd->groups->ref);
7335 if (i != cpumask_first(sched_domain_span(sd)))
7336 continue;
7338 build_sched_groups(sd);
7342 /* Calculate CPU power for physical packages and nodes */
7343 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7344 if (!cpumask_test_cpu(i, cpu_map))
7345 continue;
7347 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7348 claim_allocations(i, sd);
7349 init_sched_groups_power(i, sd);
7353 /* Attach the domains */
7354 rcu_read_lock();
7355 for_each_cpu(i, cpu_map) {
7356 sd = *per_cpu_ptr(d.sd, i);
7357 cpu_attach_domain(sd, d.rd, i);
7359 rcu_read_unlock();
7361 ret = 0;
7362 error:
7363 __free_domain_allocs(&d, alloc_state, cpu_map);
7364 return ret;
7367 static cpumask_var_t *doms_cur; /* current sched domains */
7368 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7369 static struct sched_domain_attr *dattr_cur;
7370 /* attribues of custom domains in 'doms_cur' */
7373 * Special case: If a kmalloc of a doms_cur partition (array of
7374 * cpumask) fails, then fallback to a single sched domain,
7375 * as determined by the single cpumask fallback_doms.
7377 static cpumask_var_t fallback_doms;
7380 * arch_update_cpu_topology lets virtualized architectures update the
7381 * cpu core maps. It is supposed to return 1 if the topology changed
7382 * or 0 if it stayed the same.
7384 int __attribute__((weak)) arch_update_cpu_topology(void)
7386 return 0;
7389 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7391 int i;
7392 cpumask_var_t *doms;
7394 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7395 if (!doms)
7396 return NULL;
7397 for (i = 0; i < ndoms; i++) {
7398 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7399 free_sched_domains(doms, i);
7400 return NULL;
7403 return doms;
7406 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7408 unsigned int i;
7409 for (i = 0; i < ndoms; i++)
7410 free_cpumask_var(doms[i]);
7411 kfree(doms);
7415 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7416 * For now this just excludes isolated cpus, but could be used to
7417 * exclude other special cases in the future.
7419 static int init_sched_domains(const struct cpumask *cpu_map)
7421 int err;
7423 arch_update_cpu_topology();
7424 ndoms_cur = 1;
7425 doms_cur = alloc_sched_domains(ndoms_cur);
7426 if (!doms_cur)
7427 doms_cur = &fallback_doms;
7428 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7429 dattr_cur = NULL;
7430 err = build_sched_domains(doms_cur[0], NULL);
7431 register_sched_domain_sysctl();
7433 return err;
7437 * Detach sched domains from a group of cpus specified in cpu_map
7438 * These cpus will now be attached to the NULL domain
7440 static void detach_destroy_domains(const struct cpumask *cpu_map)
7442 int i;
7444 rcu_read_lock();
7445 for_each_cpu(i, cpu_map)
7446 cpu_attach_domain(NULL, &def_root_domain, i);
7447 rcu_read_unlock();
7450 /* handle null as "default" */
7451 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7452 struct sched_domain_attr *new, int idx_new)
7454 struct sched_domain_attr tmp;
7456 /* fast path */
7457 if (!new && !cur)
7458 return 1;
7460 tmp = SD_ATTR_INIT;
7461 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7462 new ? (new + idx_new) : &tmp,
7463 sizeof(struct sched_domain_attr));
7467 * Partition sched domains as specified by the 'ndoms_new'
7468 * cpumasks in the array doms_new[] of cpumasks. This compares
7469 * doms_new[] to the current sched domain partitioning, doms_cur[].
7470 * It destroys each deleted domain and builds each new domain.
7472 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7473 * The masks don't intersect (don't overlap.) We should setup one
7474 * sched domain for each mask. CPUs not in any of the cpumasks will
7475 * not be load balanced. If the same cpumask appears both in the
7476 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7477 * it as it is.
7479 * The passed in 'doms_new' should be allocated using
7480 * alloc_sched_domains. This routine takes ownership of it and will
7481 * free_sched_domains it when done with it. If the caller failed the
7482 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7483 * and partition_sched_domains() will fallback to the single partition
7484 * 'fallback_doms', it also forces the domains to be rebuilt.
7486 * If doms_new == NULL it will be replaced with cpu_online_mask.
7487 * ndoms_new == 0 is a special case for destroying existing domains,
7488 * and it will not create the default domain.
7490 * Call with hotplug lock held
7492 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7493 struct sched_domain_attr *dattr_new)
7495 int i, j, n;
7496 int new_topology;
7498 mutex_lock(&sched_domains_mutex);
7500 /* always unregister in case we don't destroy any domains */
7501 unregister_sched_domain_sysctl();
7503 /* Let architecture update cpu core mappings. */
7504 new_topology = arch_update_cpu_topology();
7506 n = doms_new ? ndoms_new : 0;
7508 /* Destroy deleted domains */
7509 for (i = 0; i < ndoms_cur; i++) {
7510 for (j = 0; j < n && !new_topology; j++) {
7511 if (cpumask_equal(doms_cur[i], doms_new[j])
7512 && dattrs_equal(dattr_cur, i, dattr_new, j))
7513 goto match1;
7515 /* no match - a current sched domain not in new doms_new[] */
7516 detach_destroy_domains(doms_cur[i]);
7517 match1:
7521 if (doms_new == NULL) {
7522 ndoms_cur = 0;
7523 doms_new = &fallback_doms;
7524 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7525 WARN_ON_ONCE(dattr_new);
7528 /* Build new domains */
7529 for (i = 0; i < ndoms_new; i++) {
7530 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7531 if (cpumask_equal(doms_new[i], doms_cur[j])
7532 && dattrs_equal(dattr_new, i, dattr_cur, j))
7533 goto match2;
7535 /* no match - add a new doms_new */
7536 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7537 match2:
7541 /* Remember the new sched domains */
7542 if (doms_cur != &fallback_doms)
7543 free_sched_domains(doms_cur, ndoms_cur);
7544 kfree(dattr_cur); /* kfree(NULL) is safe */
7545 doms_cur = doms_new;
7546 dattr_cur = dattr_new;
7547 ndoms_cur = ndoms_new;
7549 register_sched_domain_sysctl();
7551 mutex_unlock(&sched_domains_mutex);
7554 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7555 static void reinit_sched_domains(void)
7557 get_online_cpus();
7559 /* Destroy domains first to force the rebuild */
7560 partition_sched_domains(0, NULL, NULL);
7562 rebuild_sched_domains();
7563 put_online_cpus();
7566 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7568 unsigned int level = 0;
7570 if (sscanf(buf, "%u", &level) != 1)
7571 return -EINVAL;
7574 * level is always be positive so don't check for
7575 * level < POWERSAVINGS_BALANCE_NONE which is 0
7576 * What happens on 0 or 1 byte write,
7577 * need to check for count as well?
7580 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7581 return -EINVAL;
7583 if (smt)
7584 sched_smt_power_savings = level;
7585 else
7586 sched_mc_power_savings = level;
7588 reinit_sched_domains();
7590 return count;
7593 #ifdef CONFIG_SCHED_MC
7594 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7595 struct sysdev_class_attribute *attr,
7596 char *page)
7598 return sprintf(page, "%u\n", sched_mc_power_savings);
7600 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7601 struct sysdev_class_attribute *attr,
7602 const char *buf, size_t count)
7604 return sched_power_savings_store(buf, count, 0);
7606 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7607 sched_mc_power_savings_show,
7608 sched_mc_power_savings_store);
7609 #endif
7611 #ifdef CONFIG_SCHED_SMT
7612 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7613 struct sysdev_class_attribute *attr,
7614 char *page)
7616 return sprintf(page, "%u\n", sched_smt_power_savings);
7618 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7619 struct sysdev_class_attribute *attr,
7620 const char *buf, size_t count)
7622 return sched_power_savings_store(buf, count, 1);
7624 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7625 sched_smt_power_savings_show,
7626 sched_smt_power_savings_store);
7627 #endif
7629 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7631 int err = 0;
7633 #ifdef CONFIG_SCHED_SMT
7634 if (smt_capable())
7635 err = sysfs_create_file(&cls->kset.kobj,
7636 &attr_sched_smt_power_savings.attr);
7637 #endif
7638 #ifdef CONFIG_SCHED_MC
7639 if (!err && mc_capable())
7640 err = sysfs_create_file(&cls->kset.kobj,
7641 &attr_sched_mc_power_savings.attr);
7642 #endif
7643 return err;
7645 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7648 * Update cpusets according to cpu_active mask. If cpusets are
7649 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7650 * around partition_sched_domains().
7652 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7653 void *hcpu)
7655 switch (action & ~CPU_TASKS_FROZEN) {
7656 case CPU_ONLINE:
7657 case CPU_DOWN_FAILED:
7658 cpuset_update_active_cpus();
7659 return NOTIFY_OK;
7660 default:
7661 return NOTIFY_DONE;
7665 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7666 void *hcpu)
7668 switch (action & ~CPU_TASKS_FROZEN) {
7669 case CPU_DOWN_PREPARE:
7670 cpuset_update_active_cpus();
7671 return NOTIFY_OK;
7672 default:
7673 return NOTIFY_DONE;
7677 static int update_runtime(struct notifier_block *nfb,
7678 unsigned long action, void *hcpu)
7680 int cpu = (int)(long)hcpu;
7682 switch (action) {
7683 case CPU_DOWN_PREPARE:
7684 case CPU_DOWN_PREPARE_FROZEN:
7685 disable_runtime(cpu_rq(cpu));
7686 return NOTIFY_OK;
7688 case CPU_DOWN_FAILED:
7689 case CPU_DOWN_FAILED_FROZEN:
7690 case CPU_ONLINE:
7691 case CPU_ONLINE_FROZEN:
7692 enable_runtime(cpu_rq(cpu));
7693 return NOTIFY_OK;
7695 default:
7696 return NOTIFY_DONE;
7700 void __init sched_init_smp(void)
7702 cpumask_var_t non_isolated_cpus;
7704 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7705 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7707 get_online_cpus();
7708 mutex_lock(&sched_domains_mutex);
7709 init_sched_domains(cpu_active_mask);
7710 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7711 if (cpumask_empty(non_isolated_cpus))
7712 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7713 mutex_unlock(&sched_domains_mutex);
7714 put_online_cpus();
7716 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7717 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7719 /* RT runtime code needs to handle some hotplug events */
7720 hotcpu_notifier(update_runtime, 0);
7722 init_hrtick();
7724 /* Move init over to a non-isolated CPU */
7725 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7726 BUG();
7727 sched_init_granularity();
7728 free_cpumask_var(non_isolated_cpus);
7730 init_sched_rt_class();
7732 #else
7733 void __init sched_init_smp(void)
7735 sched_init_granularity();
7737 #endif /* CONFIG_SMP */
7739 const_debug unsigned int sysctl_timer_migration = 1;
7741 int in_sched_functions(unsigned long addr)
7743 return in_lock_functions(addr) ||
7744 (addr >= (unsigned long)__sched_text_start
7745 && addr < (unsigned long)__sched_text_end);
7748 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7750 cfs_rq->tasks_timeline = RB_ROOT;
7751 INIT_LIST_HEAD(&cfs_rq->tasks);
7752 #ifdef CONFIG_FAIR_GROUP_SCHED
7753 cfs_rq->rq = rq;
7754 /* allow initial update_cfs_load() to truncate */
7755 #ifdef CONFIG_SMP
7756 cfs_rq->load_stamp = 1;
7757 #endif
7758 #endif
7759 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7760 #ifndef CONFIG_64BIT
7761 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7762 #endif
7765 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7767 struct rt_prio_array *array;
7768 int i;
7770 array = &rt_rq->active;
7771 for (i = 0; i < MAX_RT_PRIO; i++) {
7772 INIT_LIST_HEAD(array->queue + i);
7773 __clear_bit(i, array->bitmap);
7775 /* delimiter for bitsearch: */
7776 __set_bit(MAX_RT_PRIO, array->bitmap);
7778 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7779 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7780 #ifdef CONFIG_SMP
7781 rt_rq->highest_prio.next = MAX_RT_PRIO;
7782 #endif
7783 #endif
7784 #ifdef CONFIG_SMP
7785 rt_rq->rt_nr_migratory = 0;
7786 rt_rq->overloaded = 0;
7787 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7788 #endif
7790 rt_rq->rt_time = 0;
7791 rt_rq->rt_throttled = 0;
7792 rt_rq->rt_runtime = 0;
7793 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7795 #ifdef CONFIG_RT_GROUP_SCHED
7796 rt_rq->rt_nr_boosted = 0;
7797 rt_rq->rq = rq;
7798 #endif
7801 #ifdef CONFIG_FAIR_GROUP_SCHED
7802 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7803 struct sched_entity *se, int cpu,
7804 struct sched_entity *parent)
7806 struct rq *rq = cpu_rq(cpu);
7807 tg->cfs_rq[cpu] = cfs_rq;
7808 init_cfs_rq(cfs_rq, rq);
7809 cfs_rq->tg = tg;
7811 tg->se[cpu] = se;
7812 /* se could be NULL for root_task_group */
7813 if (!se)
7814 return;
7816 if (!parent)
7817 se->cfs_rq = &rq->cfs;
7818 else
7819 se->cfs_rq = parent->my_q;
7821 se->my_q = cfs_rq;
7822 update_load_set(&se->load, 0);
7823 se->parent = parent;
7825 #endif
7827 #ifdef CONFIG_RT_GROUP_SCHED
7828 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7829 struct sched_rt_entity *rt_se, int cpu,
7830 struct sched_rt_entity *parent)
7832 struct rq *rq = cpu_rq(cpu);
7834 tg->rt_rq[cpu] = rt_rq;
7835 init_rt_rq(rt_rq, rq);
7836 rt_rq->tg = tg;
7837 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7839 tg->rt_se[cpu] = rt_se;
7840 if (!rt_se)
7841 return;
7843 if (!parent)
7844 rt_se->rt_rq = &rq->rt;
7845 else
7846 rt_se->rt_rq = parent->my_q;
7848 rt_se->my_q = rt_rq;
7849 rt_se->parent = parent;
7850 INIT_LIST_HEAD(&rt_se->run_list);
7852 #endif
7854 void __init sched_init(void)
7856 int i, j;
7857 unsigned long alloc_size = 0, ptr;
7859 #ifdef CONFIG_FAIR_GROUP_SCHED
7860 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7861 #endif
7862 #ifdef CONFIG_RT_GROUP_SCHED
7863 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7864 #endif
7865 #ifdef CONFIG_CPUMASK_OFFSTACK
7866 alloc_size += num_possible_cpus() * cpumask_size();
7867 #endif
7868 if (alloc_size) {
7869 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7871 #ifdef CONFIG_FAIR_GROUP_SCHED
7872 root_task_group.se = (struct sched_entity **)ptr;
7873 ptr += nr_cpu_ids * sizeof(void **);
7875 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7876 ptr += nr_cpu_ids * sizeof(void **);
7878 #endif /* CONFIG_FAIR_GROUP_SCHED */
7879 #ifdef CONFIG_RT_GROUP_SCHED
7880 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7881 ptr += nr_cpu_ids * sizeof(void **);
7883 root_task_group.rt_rq = (struct rt_rq **)ptr;
7884 ptr += nr_cpu_ids * sizeof(void **);
7886 #endif /* CONFIG_RT_GROUP_SCHED */
7887 #ifdef CONFIG_CPUMASK_OFFSTACK
7888 for_each_possible_cpu(i) {
7889 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7890 ptr += cpumask_size();
7892 #endif /* CONFIG_CPUMASK_OFFSTACK */
7895 #ifdef CONFIG_SMP
7896 init_defrootdomain();
7897 #endif
7899 init_rt_bandwidth(&def_rt_bandwidth,
7900 global_rt_period(), global_rt_runtime());
7902 #ifdef CONFIG_RT_GROUP_SCHED
7903 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7904 global_rt_period(), global_rt_runtime());
7905 #endif /* CONFIG_RT_GROUP_SCHED */
7907 #ifdef CONFIG_CGROUP_SCHED
7908 list_add(&root_task_group.list, &task_groups);
7909 INIT_LIST_HEAD(&root_task_group.children);
7910 autogroup_init(&init_task);
7911 #endif /* CONFIG_CGROUP_SCHED */
7913 for_each_possible_cpu(i) {
7914 struct rq *rq;
7916 rq = cpu_rq(i);
7917 raw_spin_lock_init(&rq->lock);
7918 rq->nr_running = 0;
7919 rq->calc_load_active = 0;
7920 rq->calc_load_update = jiffies + LOAD_FREQ;
7921 init_cfs_rq(&rq->cfs, rq);
7922 init_rt_rq(&rq->rt, rq);
7923 #ifdef CONFIG_FAIR_GROUP_SCHED
7924 root_task_group.shares = root_task_group_load;
7925 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7927 * How much cpu bandwidth does root_task_group get?
7929 * In case of task-groups formed thr' the cgroup filesystem, it
7930 * gets 100% of the cpu resources in the system. This overall
7931 * system cpu resource is divided among the tasks of
7932 * root_task_group and its child task-groups in a fair manner,
7933 * based on each entity's (task or task-group's) weight
7934 * (se->load.weight).
7936 * In other words, if root_task_group has 10 tasks of weight
7937 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7938 * then A0's share of the cpu resource is:
7940 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7942 * We achieve this by letting root_task_group's tasks sit
7943 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7945 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7946 #endif /* CONFIG_FAIR_GROUP_SCHED */
7948 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7949 #ifdef CONFIG_RT_GROUP_SCHED
7950 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7951 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7952 #endif
7954 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7955 rq->cpu_load[j] = 0;
7957 rq->last_load_update_tick = jiffies;
7959 #ifdef CONFIG_SMP
7960 rq->sd = NULL;
7961 rq->rd = NULL;
7962 rq->cpu_power = SCHED_POWER_SCALE;
7963 rq->post_schedule = 0;
7964 rq->active_balance = 0;
7965 rq->next_balance = jiffies;
7966 rq->push_cpu = 0;
7967 rq->cpu = i;
7968 rq->online = 0;
7969 rq->idle_stamp = 0;
7970 rq->avg_idle = 2*sysctl_sched_migration_cost;
7971 rq_attach_root(rq, &def_root_domain);
7972 #ifdef CONFIG_NO_HZ
7973 rq->nohz_balance_kick = 0;
7974 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
7975 #endif
7976 #endif
7977 init_rq_hrtick(rq);
7978 atomic_set(&rq->nr_iowait, 0);
7981 set_load_weight(&init_task);
7983 #ifdef CONFIG_PREEMPT_NOTIFIERS
7984 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7985 #endif
7987 #ifdef CONFIG_SMP
7988 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7989 #endif
7991 #ifdef CONFIG_RT_MUTEXES
7992 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7993 #endif
7996 * The boot idle thread does lazy MMU switching as well:
7998 atomic_inc(&init_mm.mm_count);
7999 enter_lazy_tlb(&init_mm, current);
8002 * Make us the idle thread. Technically, schedule() should not be
8003 * called from this thread, however somewhere below it might be,
8004 * but because we are the idle thread, we just pick up running again
8005 * when this runqueue becomes "idle".
8007 init_idle(current, smp_processor_id());
8009 calc_load_update = jiffies + LOAD_FREQ;
8012 * During early bootup we pretend to be a normal task:
8014 current->sched_class = &fair_sched_class;
8016 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8017 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8018 #ifdef CONFIG_SMP
8019 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8020 #ifdef CONFIG_NO_HZ
8021 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8022 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8023 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8024 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8025 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8026 #endif
8027 /* May be allocated at isolcpus cmdline parse time */
8028 if (cpu_isolated_map == NULL)
8029 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8030 #endif /* SMP */
8032 scheduler_running = 1;
8035 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8036 static inline int preempt_count_equals(int preempt_offset)
8038 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8040 return (nested == preempt_offset);
8043 void __might_sleep(const char *file, int line, int preempt_offset)
8045 #ifdef in_atomic
8046 static unsigned long prev_jiffy; /* ratelimiting */
8048 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8049 system_state != SYSTEM_RUNNING || oops_in_progress)
8050 return;
8051 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8052 return;
8053 prev_jiffy = jiffies;
8055 printk(KERN_ERR
8056 "BUG: sleeping function called from invalid context at %s:%d\n",
8057 file, line);
8058 printk(KERN_ERR
8059 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8060 in_atomic(), irqs_disabled(),
8061 current->pid, current->comm);
8063 debug_show_held_locks(current);
8064 if (irqs_disabled())
8065 print_irqtrace_events(current);
8066 dump_stack();
8067 #endif
8069 EXPORT_SYMBOL(__might_sleep);
8070 #endif
8072 #ifdef CONFIG_MAGIC_SYSRQ
8073 static void normalize_task(struct rq *rq, struct task_struct *p)
8075 const struct sched_class *prev_class = p->sched_class;
8076 int old_prio = p->prio;
8077 int on_rq;
8079 on_rq = p->on_rq;
8080 if (on_rq)
8081 deactivate_task(rq, p, 0);
8082 __setscheduler(rq, p, SCHED_NORMAL, 0);
8083 if (on_rq) {
8084 activate_task(rq, p, 0);
8085 resched_task(rq->curr);
8088 check_class_changed(rq, p, prev_class, old_prio);
8091 void normalize_rt_tasks(void)
8093 struct task_struct *g, *p;
8094 unsigned long flags;
8095 struct rq *rq;
8097 read_lock_irqsave(&tasklist_lock, flags);
8098 do_each_thread(g, p) {
8100 * Only normalize user tasks:
8102 if (!p->mm)
8103 continue;
8105 p->se.exec_start = 0;
8106 #ifdef CONFIG_SCHEDSTATS
8107 p->se.statistics.wait_start = 0;
8108 p->se.statistics.sleep_start = 0;
8109 p->se.statistics.block_start = 0;
8110 #endif
8112 if (!rt_task(p)) {
8114 * Renice negative nice level userspace
8115 * tasks back to 0:
8117 if (TASK_NICE(p) < 0 && p->mm)
8118 set_user_nice(p, 0);
8119 continue;
8122 raw_spin_lock(&p->pi_lock);
8123 rq = __task_rq_lock(p);
8125 normalize_task(rq, p);
8127 __task_rq_unlock(rq);
8128 raw_spin_unlock(&p->pi_lock);
8129 } while_each_thread(g, p);
8131 read_unlock_irqrestore(&tasklist_lock, flags);
8134 #endif /* CONFIG_MAGIC_SYSRQ */
8136 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8138 * These functions are only useful for the IA64 MCA handling, or kdb.
8140 * They can only be called when the whole system has been
8141 * stopped - every CPU needs to be quiescent, and no scheduling
8142 * activity can take place. Using them for anything else would
8143 * be a serious bug, and as a result, they aren't even visible
8144 * under any other configuration.
8148 * curr_task - return the current task for a given cpu.
8149 * @cpu: the processor in question.
8151 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8153 struct task_struct *curr_task(int cpu)
8155 return cpu_curr(cpu);
8158 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8160 #ifdef CONFIG_IA64
8162 * set_curr_task - set the current task for a given cpu.
8163 * @cpu: the processor in question.
8164 * @p: the task pointer to set.
8166 * Description: This function must only be used when non-maskable interrupts
8167 * are serviced on a separate stack. It allows the architecture to switch the
8168 * notion of the current task on a cpu in a non-blocking manner. This function
8169 * must be called with all CPU's synchronized, and interrupts disabled, the
8170 * and caller must save the original value of the current task (see
8171 * curr_task() above) and restore that value before reenabling interrupts and
8172 * re-starting the system.
8174 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8176 void set_curr_task(int cpu, struct task_struct *p)
8178 cpu_curr(cpu) = p;
8181 #endif
8183 #ifdef CONFIG_FAIR_GROUP_SCHED
8184 static void free_fair_sched_group(struct task_group *tg)
8186 int i;
8188 for_each_possible_cpu(i) {
8189 if (tg->cfs_rq)
8190 kfree(tg->cfs_rq[i]);
8191 if (tg->se)
8192 kfree(tg->se[i]);
8195 kfree(tg->cfs_rq);
8196 kfree(tg->se);
8199 static
8200 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8202 struct cfs_rq *cfs_rq;
8203 struct sched_entity *se;
8204 int i;
8206 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8207 if (!tg->cfs_rq)
8208 goto err;
8209 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8210 if (!tg->se)
8211 goto err;
8213 tg->shares = NICE_0_LOAD;
8215 for_each_possible_cpu(i) {
8216 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8217 GFP_KERNEL, cpu_to_node(i));
8218 if (!cfs_rq)
8219 goto err;
8221 se = kzalloc_node(sizeof(struct sched_entity),
8222 GFP_KERNEL, cpu_to_node(i));
8223 if (!se)
8224 goto err_free_rq;
8226 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8229 return 1;
8231 err_free_rq:
8232 kfree(cfs_rq);
8233 err:
8234 return 0;
8237 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8239 struct rq *rq = cpu_rq(cpu);
8240 unsigned long flags;
8243 * Only empty task groups can be destroyed; so we can speculatively
8244 * check on_list without danger of it being re-added.
8246 if (!tg->cfs_rq[cpu]->on_list)
8247 return;
8249 raw_spin_lock_irqsave(&rq->lock, flags);
8250 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8251 raw_spin_unlock_irqrestore(&rq->lock, flags);
8253 #else /* !CONFG_FAIR_GROUP_SCHED */
8254 static inline void free_fair_sched_group(struct task_group *tg)
8258 static inline
8259 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8261 return 1;
8264 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8267 #endif /* CONFIG_FAIR_GROUP_SCHED */
8269 #ifdef CONFIG_RT_GROUP_SCHED
8270 static void free_rt_sched_group(struct task_group *tg)
8272 int i;
8274 destroy_rt_bandwidth(&tg->rt_bandwidth);
8276 for_each_possible_cpu(i) {
8277 if (tg->rt_rq)
8278 kfree(tg->rt_rq[i]);
8279 if (tg->rt_se)
8280 kfree(tg->rt_se[i]);
8283 kfree(tg->rt_rq);
8284 kfree(tg->rt_se);
8287 static
8288 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8290 struct rt_rq *rt_rq;
8291 struct sched_rt_entity *rt_se;
8292 int i;
8294 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8295 if (!tg->rt_rq)
8296 goto err;
8297 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8298 if (!tg->rt_se)
8299 goto err;
8301 init_rt_bandwidth(&tg->rt_bandwidth,
8302 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8304 for_each_possible_cpu(i) {
8305 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8306 GFP_KERNEL, cpu_to_node(i));
8307 if (!rt_rq)
8308 goto err;
8310 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8311 GFP_KERNEL, cpu_to_node(i));
8312 if (!rt_se)
8313 goto err_free_rq;
8315 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8318 return 1;
8320 err_free_rq:
8321 kfree(rt_rq);
8322 err:
8323 return 0;
8325 #else /* !CONFIG_RT_GROUP_SCHED */
8326 static inline void free_rt_sched_group(struct task_group *tg)
8330 static inline
8331 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8333 return 1;
8335 #endif /* CONFIG_RT_GROUP_SCHED */
8337 #ifdef CONFIG_CGROUP_SCHED
8338 static void free_sched_group(struct task_group *tg)
8340 free_fair_sched_group(tg);
8341 free_rt_sched_group(tg);
8342 autogroup_free(tg);
8343 kfree(tg);
8346 /* allocate runqueue etc for a new task group */
8347 struct task_group *sched_create_group(struct task_group *parent)
8349 struct task_group *tg;
8350 unsigned long flags;
8352 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8353 if (!tg)
8354 return ERR_PTR(-ENOMEM);
8356 if (!alloc_fair_sched_group(tg, parent))
8357 goto err;
8359 if (!alloc_rt_sched_group(tg, parent))
8360 goto err;
8362 spin_lock_irqsave(&task_group_lock, flags);
8363 list_add_rcu(&tg->list, &task_groups);
8365 WARN_ON(!parent); /* root should already exist */
8367 tg->parent = parent;
8368 INIT_LIST_HEAD(&tg->children);
8369 list_add_rcu(&tg->siblings, &parent->children);
8370 spin_unlock_irqrestore(&task_group_lock, flags);
8372 return tg;
8374 err:
8375 free_sched_group(tg);
8376 return ERR_PTR(-ENOMEM);
8379 /* rcu callback to free various structures associated with a task group */
8380 static void free_sched_group_rcu(struct rcu_head *rhp)
8382 /* now it should be safe to free those cfs_rqs */
8383 free_sched_group(container_of(rhp, struct task_group, rcu));
8386 /* Destroy runqueue etc associated with a task group */
8387 void sched_destroy_group(struct task_group *tg)
8389 unsigned long flags;
8390 int i;
8392 /* end participation in shares distribution */
8393 for_each_possible_cpu(i)
8394 unregister_fair_sched_group(tg, i);
8396 spin_lock_irqsave(&task_group_lock, flags);
8397 list_del_rcu(&tg->list);
8398 list_del_rcu(&tg->siblings);
8399 spin_unlock_irqrestore(&task_group_lock, flags);
8401 /* wait for possible concurrent references to cfs_rqs complete */
8402 call_rcu(&tg->rcu, free_sched_group_rcu);
8405 /* change task's runqueue when it moves between groups.
8406 * The caller of this function should have put the task in its new group
8407 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8408 * reflect its new group.
8410 void sched_move_task(struct task_struct *tsk)
8412 int on_rq, running;
8413 unsigned long flags;
8414 struct rq *rq;
8416 rq = task_rq_lock(tsk, &flags);
8418 running = task_current(rq, tsk);
8419 on_rq = tsk->on_rq;
8421 if (on_rq)
8422 dequeue_task(rq, tsk, 0);
8423 if (unlikely(running))
8424 tsk->sched_class->put_prev_task(rq, tsk);
8426 #ifdef CONFIG_FAIR_GROUP_SCHED
8427 if (tsk->sched_class->task_move_group)
8428 tsk->sched_class->task_move_group(tsk, on_rq);
8429 else
8430 #endif
8431 set_task_rq(tsk, task_cpu(tsk));
8433 if (unlikely(running))
8434 tsk->sched_class->set_curr_task(rq);
8435 if (on_rq)
8436 enqueue_task(rq, tsk, 0);
8438 task_rq_unlock(rq, tsk, &flags);
8440 #endif /* CONFIG_CGROUP_SCHED */
8442 #ifdef CONFIG_FAIR_GROUP_SCHED
8443 static DEFINE_MUTEX(shares_mutex);
8445 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8447 int i;
8448 unsigned long flags;
8451 * We can't change the weight of the root cgroup.
8453 if (!tg->se[0])
8454 return -EINVAL;
8456 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8458 mutex_lock(&shares_mutex);
8459 if (tg->shares == shares)
8460 goto done;
8462 tg->shares = shares;
8463 for_each_possible_cpu(i) {
8464 struct rq *rq = cpu_rq(i);
8465 struct sched_entity *se;
8467 se = tg->se[i];
8468 /* Propagate contribution to hierarchy */
8469 raw_spin_lock_irqsave(&rq->lock, flags);
8470 for_each_sched_entity(se)
8471 update_cfs_shares(group_cfs_rq(se));
8472 raw_spin_unlock_irqrestore(&rq->lock, flags);
8475 done:
8476 mutex_unlock(&shares_mutex);
8477 return 0;
8480 unsigned long sched_group_shares(struct task_group *tg)
8482 return tg->shares;
8484 #endif
8486 #ifdef CONFIG_RT_GROUP_SCHED
8488 * Ensure that the real time constraints are schedulable.
8490 static DEFINE_MUTEX(rt_constraints_mutex);
8492 static unsigned long to_ratio(u64 period, u64 runtime)
8494 if (runtime == RUNTIME_INF)
8495 return 1ULL << 20;
8497 return div64_u64(runtime << 20, period);
8500 /* Must be called with tasklist_lock held */
8501 static inline int tg_has_rt_tasks(struct task_group *tg)
8503 struct task_struct *g, *p;
8505 do_each_thread(g, p) {
8506 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8507 return 1;
8508 } while_each_thread(g, p);
8510 return 0;
8513 struct rt_schedulable_data {
8514 struct task_group *tg;
8515 u64 rt_period;
8516 u64 rt_runtime;
8519 static int tg_schedulable(struct task_group *tg, void *data)
8521 struct rt_schedulable_data *d = data;
8522 struct task_group *child;
8523 unsigned long total, sum = 0;
8524 u64 period, runtime;
8526 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8527 runtime = tg->rt_bandwidth.rt_runtime;
8529 if (tg == d->tg) {
8530 period = d->rt_period;
8531 runtime = d->rt_runtime;
8535 * Cannot have more runtime than the period.
8537 if (runtime > period && runtime != RUNTIME_INF)
8538 return -EINVAL;
8541 * Ensure we don't starve existing RT tasks.
8543 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8544 return -EBUSY;
8546 total = to_ratio(period, runtime);
8549 * Nobody can have more than the global setting allows.
8551 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8552 return -EINVAL;
8555 * The sum of our children's runtime should not exceed our own.
8557 list_for_each_entry_rcu(child, &tg->children, siblings) {
8558 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8559 runtime = child->rt_bandwidth.rt_runtime;
8561 if (child == d->tg) {
8562 period = d->rt_period;
8563 runtime = d->rt_runtime;
8566 sum += to_ratio(period, runtime);
8569 if (sum > total)
8570 return -EINVAL;
8572 return 0;
8575 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8577 struct rt_schedulable_data data = {
8578 .tg = tg,
8579 .rt_period = period,
8580 .rt_runtime = runtime,
8583 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8586 static int tg_set_bandwidth(struct task_group *tg,
8587 u64 rt_period, u64 rt_runtime)
8589 int i, err = 0;
8591 mutex_lock(&rt_constraints_mutex);
8592 read_lock(&tasklist_lock);
8593 err = __rt_schedulable(tg, rt_period, rt_runtime);
8594 if (err)
8595 goto unlock;
8597 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8598 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8599 tg->rt_bandwidth.rt_runtime = rt_runtime;
8601 for_each_possible_cpu(i) {
8602 struct rt_rq *rt_rq = tg->rt_rq[i];
8604 raw_spin_lock(&rt_rq->rt_runtime_lock);
8605 rt_rq->rt_runtime = rt_runtime;
8606 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8608 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8609 unlock:
8610 read_unlock(&tasklist_lock);
8611 mutex_unlock(&rt_constraints_mutex);
8613 return err;
8616 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8618 u64 rt_runtime, rt_period;
8620 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8621 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8622 if (rt_runtime_us < 0)
8623 rt_runtime = RUNTIME_INF;
8625 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8628 long sched_group_rt_runtime(struct task_group *tg)
8630 u64 rt_runtime_us;
8632 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8633 return -1;
8635 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8636 do_div(rt_runtime_us, NSEC_PER_USEC);
8637 return rt_runtime_us;
8640 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8642 u64 rt_runtime, rt_period;
8644 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8645 rt_runtime = tg->rt_bandwidth.rt_runtime;
8647 if (rt_period == 0)
8648 return -EINVAL;
8650 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8653 long sched_group_rt_period(struct task_group *tg)
8655 u64 rt_period_us;
8657 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8658 do_div(rt_period_us, NSEC_PER_USEC);
8659 return rt_period_us;
8662 static int sched_rt_global_constraints(void)
8664 u64 runtime, period;
8665 int ret = 0;
8667 if (sysctl_sched_rt_period <= 0)
8668 return -EINVAL;
8670 runtime = global_rt_runtime();
8671 period = global_rt_period();
8674 * Sanity check on the sysctl variables.
8676 if (runtime > period && runtime != RUNTIME_INF)
8677 return -EINVAL;
8679 mutex_lock(&rt_constraints_mutex);
8680 read_lock(&tasklist_lock);
8681 ret = __rt_schedulable(NULL, 0, 0);
8682 read_unlock(&tasklist_lock);
8683 mutex_unlock(&rt_constraints_mutex);
8685 return ret;
8688 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8690 /* Don't accept realtime tasks when there is no way for them to run */
8691 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8692 return 0;
8694 return 1;
8697 #else /* !CONFIG_RT_GROUP_SCHED */
8698 static int sched_rt_global_constraints(void)
8700 unsigned long flags;
8701 int i;
8703 if (sysctl_sched_rt_period <= 0)
8704 return -EINVAL;
8707 * There's always some RT tasks in the root group
8708 * -- migration, kstopmachine etc..
8710 if (sysctl_sched_rt_runtime == 0)
8711 return -EBUSY;
8713 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8714 for_each_possible_cpu(i) {
8715 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8717 raw_spin_lock(&rt_rq->rt_runtime_lock);
8718 rt_rq->rt_runtime = global_rt_runtime();
8719 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8721 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8723 return 0;
8725 #endif /* CONFIG_RT_GROUP_SCHED */
8727 int sched_rt_handler(struct ctl_table *table, int write,
8728 void __user *buffer, size_t *lenp,
8729 loff_t *ppos)
8731 int ret;
8732 int old_period, old_runtime;
8733 static DEFINE_MUTEX(mutex);
8735 mutex_lock(&mutex);
8736 old_period = sysctl_sched_rt_period;
8737 old_runtime = sysctl_sched_rt_runtime;
8739 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8741 if (!ret && write) {
8742 ret = sched_rt_global_constraints();
8743 if (ret) {
8744 sysctl_sched_rt_period = old_period;
8745 sysctl_sched_rt_runtime = old_runtime;
8746 } else {
8747 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8748 def_rt_bandwidth.rt_period =
8749 ns_to_ktime(global_rt_period());
8752 mutex_unlock(&mutex);
8754 return ret;
8757 #ifdef CONFIG_CGROUP_SCHED
8759 /* return corresponding task_group object of a cgroup */
8760 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8762 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8763 struct task_group, css);
8766 static struct cgroup_subsys_state *
8767 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8769 struct task_group *tg, *parent;
8771 if (!cgrp->parent) {
8772 /* This is early initialization for the top cgroup */
8773 return &root_task_group.css;
8776 parent = cgroup_tg(cgrp->parent);
8777 tg = sched_create_group(parent);
8778 if (IS_ERR(tg))
8779 return ERR_PTR(-ENOMEM);
8781 return &tg->css;
8784 static void
8785 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8787 struct task_group *tg = cgroup_tg(cgrp);
8789 sched_destroy_group(tg);
8792 static int
8793 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8795 #ifdef CONFIG_RT_GROUP_SCHED
8796 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8797 return -EINVAL;
8798 #else
8799 /* We don't support RT-tasks being in separate groups */
8800 if (tsk->sched_class != &fair_sched_class)
8801 return -EINVAL;
8802 #endif
8803 return 0;
8806 static void
8807 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8809 sched_move_task(tsk);
8812 static void
8813 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
8814 struct cgroup *old_cgrp, struct task_struct *task)
8817 * cgroup_exit() is called in the copy_process() failure path.
8818 * Ignore this case since the task hasn't ran yet, this avoids
8819 * trying to poke a half freed task state from generic code.
8821 if (!(task->flags & PF_EXITING))
8822 return;
8824 sched_move_task(task);
8827 #ifdef CONFIG_FAIR_GROUP_SCHED
8828 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8829 u64 shareval)
8831 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
8834 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8836 struct task_group *tg = cgroup_tg(cgrp);
8838 return (u64) scale_load_down(tg->shares);
8840 #endif /* CONFIG_FAIR_GROUP_SCHED */
8842 #ifdef CONFIG_RT_GROUP_SCHED
8843 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8844 s64 val)
8846 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8849 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8851 return sched_group_rt_runtime(cgroup_tg(cgrp));
8854 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8855 u64 rt_period_us)
8857 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8860 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8862 return sched_group_rt_period(cgroup_tg(cgrp));
8864 #endif /* CONFIG_RT_GROUP_SCHED */
8866 static struct cftype cpu_files[] = {
8867 #ifdef CONFIG_FAIR_GROUP_SCHED
8869 .name = "shares",
8870 .read_u64 = cpu_shares_read_u64,
8871 .write_u64 = cpu_shares_write_u64,
8873 #endif
8874 #ifdef CONFIG_RT_GROUP_SCHED
8876 .name = "rt_runtime_us",
8877 .read_s64 = cpu_rt_runtime_read,
8878 .write_s64 = cpu_rt_runtime_write,
8881 .name = "rt_period_us",
8882 .read_u64 = cpu_rt_period_read_uint,
8883 .write_u64 = cpu_rt_period_write_uint,
8885 #endif
8888 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8890 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8893 struct cgroup_subsys cpu_cgroup_subsys = {
8894 .name = "cpu",
8895 .create = cpu_cgroup_create,
8896 .destroy = cpu_cgroup_destroy,
8897 .can_attach_task = cpu_cgroup_can_attach_task,
8898 .attach_task = cpu_cgroup_attach_task,
8899 .exit = cpu_cgroup_exit,
8900 .populate = cpu_cgroup_populate,
8901 .subsys_id = cpu_cgroup_subsys_id,
8902 .early_init = 1,
8905 #endif /* CONFIG_CGROUP_SCHED */
8907 #ifdef CONFIG_CGROUP_CPUACCT
8910 * CPU accounting code for task groups.
8912 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8913 * (balbir@in.ibm.com).
8916 /* track cpu usage of a group of tasks and its child groups */
8917 struct cpuacct {
8918 struct cgroup_subsys_state css;
8919 /* cpuusage holds pointer to a u64-type object on every cpu */
8920 u64 __percpu *cpuusage;
8921 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8922 struct cpuacct *parent;
8925 struct cgroup_subsys cpuacct_subsys;
8927 /* return cpu accounting group corresponding to this container */
8928 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8930 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8931 struct cpuacct, css);
8934 /* return cpu accounting group to which this task belongs */
8935 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8937 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8938 struct cpuacct, css);
8941 /* create a new cpu accounting group */
8942 static struct cgroup_subsys_state *cpuacct_create(
8943 struct cgroup_subsys *ss, struct cgroup *cgrp)
8945 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8946 int i;
8948 if (!ca)
8949 goto out;
8951 ca->cpuusage = alloc_percpu(u64);
8952 if (!ca->cpuusage)
8953 goto out_free_ca;
8955 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8956 if (percpu_counter_init(&ca->cpustat[i], 0))
8957 goto out_free_counters;
8959 if (cgrp->parent)
8960 ca->parent = cgroup_ca(cgrp->parent);
8962 return &ca->css;
8964 out_free_counters:
8965 while (--i >= 0)
8966 percpu_counter_destroy(&ca->cpustat[i]);
8967 free_percpu(ca->cpuusage);
8968 out_free_ca:
8969 kfree(ca);
8970 out:
8971 return ERR_PTR(-ENOMEM);
8974 /* destroy an existing cpu accounting group */
8975 static void
8976 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8978 struct cpuacct *ca = cgroup_ca(cgrp);
8979 int i;
8981 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8982 percpu_counter_destroy(&ca->cpustat[i]);
8983 free_percpu(ca->cpuusage);
8984 kfree(ca);
8987 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8989 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8990 u64 data;
8992 #ifndef CONFIG_64BIT
8994 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8996 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8997 data = *cpuusage;
8998 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8999 #else
9000 data = *cpuusage;
9001 #endif
9003 return data;
9006 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9008 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9010 #ifndef CONFIG_64BIT
9012 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9014 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9015 *cpuusage = val;
9016 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9017 #else
9018 *cpuusage = val;
9019 #endif
9022 /* return total cpu usage (in nanoseconds) of a group */
9023 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9025 struct cpuacct *ca = cgroup_ca(cgrp);
9026 u64 totalcpuusage = 0;
9027 int i;
9029 for_each_present_cpu(i)
9030 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9032 return totalcpuusage;
9035 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9036 u64 reset)
9038 struct cpuacct *ca = cgroup_ca(cgrp);
9039 int err = 0;
9040 int i;
9042 if (reset) {
9043 err = -EINVAL;
9044 goto out;
9047 for_each_present_cpu(i)
9048 cpuacct_cpuusage_write(ca, i, 0);
9050 out:
9051 return err;
9054 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9055 struct seq_file *m)
9057 struct cpuacct *ca = cgroup_ca(cgroup);
9058 u64 percpu;
9059 int i;
9061 for_each_present_cpu(i) {
9062 percpu = cpuacct_cpuusage_read(ca, i);
9063 seq_printf(m, "%llu ", (unsigned long long) percpu);
9065 seq_printf(m, "\n");
9066 return 0;
9069 static const char *cpuacct_stat_desc[] = {
9070 [CPUACCT_STAT_USER] = "user",
9071 [CPUACCT_STAT_SYSTEM] = "system",
9074 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9075 struct cgroup_map_cb *cb)
9077 struct cpuacct *ca = cgroup_ca(cgrp);
9078 int i;
9080 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9081 s64 val = percpu_counter_read(&ca->cpustat[i]);
9082 val = cputime64_to_clock_t(val);
9083 cb->fill(cb, cpuacct_stat_desc[i], val);
9085 return 0;
9088 static struct cftype files[] = {
9090 .name = "usage",
9091 .read_u64 = cpuusage_read,
9092 .write_u64 = cpuusage_write,
9095 .name = "usage_percpu",
9096 .read_seq_string = cpuacct_percpu_seq_read,
9099 .name = "stat",
9100 .read_map = cpuacct_stats_show,
9104 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9106 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9110 * charge this task's execution time to its accounting group.
9112 * called with rq->lock held.
9114 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9116 struct cpuacct *ca;
9117 int cpu;
9119 if (unlikely(!cpuacct_subsys.active))
9120 return;
9122 cpu = task_cpu(tsk);
9124 rcu_read_lock();
9126 ca = task_ca(tsk);
9128 for (; ca; ca = ca->parent) {
9129 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9130 *cpuusage += cputime;
9133 rcu_read_unlock();
9137 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9138 * in cputime_t units. As a result, cpuacct_update_stats calls
9139 * percpu_counter_add with values large enough to always overflow the
9140 * per cpu batch limit causing bad SMP scalability.
9142 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9143 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9144 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9146 #ifdef CONFIG_SMP
9147 #define CPUACCT_BATCH \
9148 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9149 #else
9150 #define CPUACCT_BATCH 0
9151 #endif
9154 * Charge the system/user time to the task's accounting group.
9156 static void cpuacct_update_stats(struct task_struct *tsk,
9157 enum cpuacct_stat_index idx, cputime_t val)
9159 struct cpuacct *ca;
9160 int batch = CPUACCT_BATCH;
9162 if (unlikely(!cpuacct_subsys.active))
9163 return;
9165 rcu_read_lock();
9166 ca = task_ca(tsk);
9168 do {
9169 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9170 ca = ca->parent;
9171 } while (ca);
9172 rcu_read_unlock();
9175 struct cgroup_subsys cpuacct_subsys = {
9176 .name = "cpuacct",
9177 .create = cpuacct_create,
9178 .destroy = cpuacct_destroy,
9179 .populate = cpuacct_populate,
9180 .subsys_id = cpuacct_subsys_id,
9182 #endif /* CONFIG_CGROUP_CPUACCT */