Merge branch 'perf/core' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux...
[linux-2.6/x86.git] / kernel / sched.c
blobccacdbdecf452bda8769878ca6e558d13ebb4e74
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>
78 #ifdef CONFIG_PARAVIRT
79 #include <asm/paravirt.h>
80 #endif
82 #include "sched_cpupri.h"
83 #include "workqueue_sched.h"
84 #include "sched_autogroup.h"
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/sched.h>
90 * Convert user-nice values [ -20 ... 0 ... 19 ]
91 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
92 * and back.
94 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
95 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
96 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
99 * 'User priority' is the nice value converted to something we
100 * can work with better when scaling various scheduler parameters,
101 * it's a [ 0 ... 39 ] range.
103 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
104 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
105 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
108 * Helpers for converting nanosecond timing to jiffy resolution
110 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
112 #define NICE_0_LOAD SCHED_LOAD_SCALE
113 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
116 * These are the 'tuning knobs' of the scheduler:
118 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
119 * Timeslices get refilled after they expire.
121 #define DEF_TIMESLICE (100 * HZ / 1000)
124 * single value that denotes runtime == period, ie unlimited time.
126 #define RUNTIME_INF ((u64)~0ULL)
128 static inline int rt_policy(int policy)
130 if (policy == SCHED_FIFO || policy == SCHED_RR)
131 return 1;
132 return 0;
135 static inline int task_has_rt_policy(struct task_struct *p)
137 return rt_policy(p->policy);
141 * This is the priority-queue data structure of the RT scheduling class:
143 struct rt_prio_array {
144 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
145 struct list_head queue[MAX_RT_PRIO];
148 struct rt_bandwidth {
149 /* nests inside the rq lock: */
150 raw_spinlock_t rt_runtime_lock;
151 ktime_t rt_period;
152 u64 rt_runtime;
153 struct hrtimer rt_period_timer;
156 static struct rt_bandwidth def_rt_bandwidth;
158 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
160 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
162 struct rt_bandwidth *rt_b =
163 container_of(timer, struct rt_bandwidth, rt_period_timer);
164 ktime_t now;
165 int overrun;
166 int idle = 0;
168 for (;;) {
169 now = hrtimer_cb_get_time(timer);
170 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
172 if (!overrun)
173 break;
175 idle = do_sched_rt_period_timer(rt_b, overrun);
178 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
181 static
182 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
184 rt_b->rt_period = ns_to_ktime(period);
185 rt_b->rt_runtime = runtime;
187 raw_spin_lock_init(&rt_b->rt_runtime_lock);
189 hrtimer_init(&rt_b->rt_period_timer,
190 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
191 rt_b->rt_period_timer.function = sched_rt_period_timer;
194 static inline int rt_bandwidth_enabled(void)
196 return sysctl_sched_rt_runtime >= 0;
199 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
201 ktime_t now;
203 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
204 return;
206 if (hrtimer_active(&rt_b->rt_period_timer))
207 return;
209 raw_spin_lock(&rt_b->rt_runtime_lock);
210 for (;;) {
211 unsigned long delta;
212 ktime_t soft, hard;
214 if (hrtimer_active(&rt_b->rt_period_timer))
215 break;
217 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
218 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
220 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
221 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
222 delta = ktime_to_ns(ktime_sub(hard, soft));
223 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
224 HRTIMER_MODE_ABS_PINNED, 0);
226 raw_spin_unlock(&rt_b->rt_runtime_lock);
229 #ifdef CONFIG_RT_GROUP_SCHED
230 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
232 hrtimer_cancel(&rt_b->rt_period_timer);
234 #endif
237 * sched_domains_mutex serializes calls to init_sched_domains,
238 * detach_destroy_domains and partition_sched_domains.
240 static DEFINE_MUTEX(sched_domains_mutex);
242 #ifdef CONFIG_CGROUP_SCHED
244 #include <linux/cgroup.h>
246 struct cfs_rq;
248 static LIST_HEAD(task_groups);
250 /* task group related information */
251 struct task_group {
252 struct cgroup_subsys_state css;
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
261 atomic_t load_weight;
262 #endif
264 #ifdef CONFIG_RT_GROUP_SCHED
265 struct sched_rt_entity **rt_se;
266 struct rt_rq **rt_rq;
268 struct rt_bandwidth rt_bandwidth;
269 #endif
271 struct rcu_head rcu;
272 struct list_head list;
274 struct task_group *parent;
275 struct list_head siblings;
276 struct list_head children;
278 #ifdef CONFIG_SCHED_AUTOGROUP
279 struct autogroup *autogroup;
280 #endif
283 /* task_group_lock serializes the addition/removal of task groups */
284 static DEFINE_SPINLOCK(task_group_lock);
286 #ifdef CONFIG_FAIR_GROUP_SCHED
288 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
291 * A weight of 0 or 1 can cause arithmetics problems.
292 * A weight of a cfs_rq is the sum of weights of which entities
293 * are queued on this cfs_rq, so a weight of a entity should not be
294 * too large, so as the shares value of a task group.
295 * (The default weight is 1024 - so there's no practical
296 * limitation from this.)
298 #define MIN_SHARES (1UL << 1)
299 #define MAX_SHARES (1UL << 18)
301 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
302 #endif
304 /* Default task group.
305 * Every task in system belong to this group at bootup.
307 struct task_group root_task_group;
309 #endif /* CONFIG_CGROUP_SCHED */
311 /* CFS-related fields in a runqueue */
312 struct cfs_rq {
313 struct load_weight load;
314 unsigned long nr_running;
316 u64 exec_clock;
317 u64 min_vruntime;
318 #ifndef CONFIG_64BIT
319 u64 min_vruntime_copy;
320 #endif
322 struct rb_root tasks_timeline;
323 struct rb_node *rb_leftmost;
325 struct list_head tasks;
326 struct list_head *balance_iterator;
329 * 'curr' points to currently running entity on this cfs_rq.
330 * It is set to NULL otherwise (i.e when none are currently running).
332 struct sched_entity *curr, *next, *last, *skip;
334 #ifdef CONFIG_SCHED_DEBUG
335 unsigned int nr_spread_over;
336 #endif
338 #ifdef CONFIG_FAIR_GROUP_SCHED
339 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
342 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
343 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
344 * (like users, containers etc.)
346 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
347 * list is used during load balance.
349 int on_list;
350 struct list_head leaf_cfs_rq_list;
351 struct task_group *tg; /* group that "owns" this runqueue */
353 #ifdef CONFIG_SMP
355 * the part of load.weight contributed by tasks
357 unsigned long task_weight;
360 * h_load = weight * f(tg)
362 * Where f(tg) is the recursive weight fraction assigned to
363 * this group.
365 unsigned long h_load;
368 * Maintaining per-cpu shares distribution for group scheduling
370 * load_stamp is the last time we updated the load average
371 * load_last is the last time we updated the load average and saw load
372 * load_unacc_exec_time is currently unaccounted execution time
374 u64 load_avg;
375 u64 load_period;
376 u64 load_stamp, load_last, load_unacc_exec_time;
378 unsigned long load_contribution;
379 #endif
380 #endif
383 /* Real-Time classes' related field in a runqueue: */
384 struct rt_rq {
385 struct rt_prio_array active;
386 unsigned long rt_nr_running;
387 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
388 struct {
389 int curr; /* highest queued rt task prio */
390 #ifdef CONFIG_SMP
391 int next; /* next highest */
392 #endif
393 } highest_prio;
394 #endif
395 #ifdef CONFIG_SMP
396 unsigned long rt_nr_migratory;
397 unsigned long rt_nr_total;
398 int overloaded;
399 struct plist_head pushable_tasks;
400 #endif
401 int rt_throttled;
402 u64 rt_time;
403 u64 rt_runtime;
404 /* Nests inside the rq lock: */
405 raw_spinlock_t rt_runtime_lock;
407 #ifdef CONFIG_RT_GROUP_SCHED
408 unsigned long rt_nr_boosted;
410 struct rq *rq;
411 struct list_head leaf_rt_rq_list;
412 struct task_group *tg;
413 #endif
416 #ifdef CONFIG_SMP
419 * We add the notion of a root-domain which will be used to define per-domain
420 * variables. Each exclusive cpuset essentially defines an island domain by
421 * fully partitioning the member cpus from any other cpuset. Whenever a new
422 * exclusive cpuset is created, we also create and attach a new root-domain
423 * object.
426 struct root_domain {
427 atomic_t refcount;
428 atomic_t rto_count;
429 struct rcu_head rcu;
430 cpumask_var_t span;
431 cpumask_var_t online;
434 * The "RT overload" flag: it gets set if a CPU has more than
435 * one runnable RT task.
437 cpumask_var_t rto_mask;
438 struct cpupri cpupri;
442 * By default the system creates a single root-domain with all cpus as
443 * members (mimicking the global state we have today).
445 static struct root_domain def_root_domain;
447 #endif /* CONFIG_SMP */
450 * This is the main, per-CPU runqueue data structure.
452 * Locking rule: those places that want to lock multiple runqueues
453 * (such as the load balancing or the thread migration code), lock
454 * acquire operations must be ordered by ascending &runqueue.
456 struct rq {
457 /* runqueue lock: */
458 raw_spinlock_t lock;
461 * nr_running and cpu_load should be in the same cacheline because
462 * remote CPUs use both these fields when doing load calculation.
464 unsigned long nr_running;
465 #define CPU_LOAD_IDX_MAX 5
466 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
467 unsigned long last_load_update_tick;
468 #ifdef CONFIG_NO_HZ
469 u64 nohz_stamp;
470 unsigned char nohz_balance_kick;
471 #endif
472 int skip_clock_update;
474 /* capture load from *all* tasks on this cpu: */
475 struct load_weight load;
476 unsigned long nr_load_updates;
477 u64 nr_switches;
479 struct cfs_rq cfs;
480 struct rt_rq rt;
482 #ifdef CONFIG_FAIR_GROUP_SCHED
483 /* list of leaf cfs_rq on this cpu: */
484 struct list_head leaf_cfs_rq_list;
485 #endif
486 #ifdef CONFIG_RT_GROUP_SCHED
487 struct list_head leaf_rt_rq_list;
488 #endif
491 * This is part of a global counter where only the total sum
492 * over all CPUs matters. A task can increase this counter on
493 * one CPU and if it got migrated afterwards it may decrease
494 * it on another CPU. Always updated under the runqueue lock:
496 unsigned long nr_uninterruptible;
498 struct task_struct *curr, *idle, *stop;
499 unsigned long next_balance;
500 struct mm_struct *prev_mm;
502 u64 clock;
503 u64 clock_task;
505 atomic_t nr_iowait;
507 #ifdef CONFIG_SMP
508 struct root_domain *rd;
509 struct sched_domain *sd;
511 unsigned long cpu_power;
513 unsigned char idle_at_tick;
514 /* For active balancing */
515 int post_schedule;
516 int active_balance;
517 int push_cpu;
518 struct cpu_stop_work active_balance_work;
519 /* cpu of this runqueue: */
520 int cpu;
521 int online;
523 unsigned long avg_load_per_task;
525 u64 rt_avg;
526 u64 age_stamp;
527 u64 idle_stamp;
528 u64 avg_idle;
529 #endif
531 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
532 u64 prev_irq_time;
533 #endif
534 #ifdef CONFIG_PARAVIRT
535 u64 prev_steal_time;
536 #endif
537 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
538 u64 prev_steal_time_rq;
539 #endif
541 /* calc_load related fields */
542 unsigned long calc_load_update;
543 long calc_load_active;
545 #ifdef CONFIG_SCHED_HRTICK
546 #ifdef CONFIG_SMP
547 int hrtick_csd_pending;
548 struct call_single_data hrtick_csd;
549 #endif
550 struct hrtimer hrtick_timer;
551 #endif
553 #ifdef CONFIG_SCHEDSTATS
554 /* latency stats */
555 struct sched_info rq_sched_info;
556 unsigned long long rq_cpu_time;
557 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
559 /* sys_sched_yield() stats */
560 unsigned int yld_count;
562 /* schedule() stats */
563 unsigned int sched_switch;
564 unsigned int sched_count;
565 unsigned int sched_goidle;
567 /* try_to_wake_up() stats */
568 unsigned int ttwu_count;
569 unsigned int ttwu_local;
570 #endif
572 #ifdef CONFIG_SMP
573 struct task_struct *wake_list;
574 #endif
577 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
580 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
582 static inline int cpu_of(struct rq *rq)
584 #ifdef CONFIG_SMP
585 return rq->cpu;
586 #else
587 return 0;
588 #endif
591 #define rcu_dereference_check_sched_domain(p) \
592 rcu_dereference_check((p), \
593 lockdep_is_held(&sched_domains_mutex))
596 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
597 * See detach_destroy_domains: synchronize_sched for details.
599 * The domain tree of any CPU may only be accessed from within
600 * preempt-disabled sections.
602 #define for_each_domain(cpu, __sd) \
603 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
605 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
606 #define this_rq() (&__get_cpu_var(runqueues))
607 #define task_rq(p) cpu_rq(task_cpu(p))
608 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
609 #define raw_rq() (&__raw_get_cpu_var(runqueues))
611 #ifdef CONFIG_CGROUP_SCHED
614 * Return the group to which this tasks belongs.
616 * We use task_subsys_state_check() and extend the RCU verification with
617 * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each
618 * task it moves into the cgroup. Therefore by holding either of those locks,
619 * we pin the task to the current cgroup.
621 static inline struct task_group *task_group(struct task_struct *p)
623 struct task_group *tg;
624 struct cgroup_subsys_state *css;
626 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
627 lockdep_is_held(&p->pi_lock) ||
628 lockdep_is_held(&task_rq(p)->lock));
629 tg = container_of(css, struct task_group, css);
631 return autogroup_task_group(p, tg);
634 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
635 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
637 #ifdef CONFIG_FAIR_GROUP_SCHED
638 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
639 p->se.parent = task_group(p)->se[cpu];
640 #endif
642 #ifdef CONFIG_RT_GROUP_SCHED
643 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
644 p->rt.parent = task_group(p)->rt_se[cpu];
645 #endif
648 #else /* CONFIG_CGROUP_SCHED */
650 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
651 static inline struct task_group *task_group(struct task_struct *p)
653 return NULL;
656 #endif /* CONFIG_CGROUP_SCHED */
658 static void update_rq_clock_task(struct rq *rq, s64 delta);
660 static void update_rq_clock(struct rq *rq)
662 s64 delta;
664 if (rq->skip_clock_update > 0)
665 return;
667 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
668 rq->clock += delta;
669 update_rq_clock_task(rq, delta);
673 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
675 #ifdef CONFIG_SCHED_DEBUG
676 # define const_debug __read_mostly
677 #else
678 # define const_debug static const
679 #endif
682 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
683 * @cpu: the processor in question.
685 * This interface allows printk to be called with the runqueue lock
686 * held and know whether or not it is OK to wake up the klogd.
688 int runqueue_is_locked(int cpu)
690 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
694 * Debugging: various feature bits
697 #define SCHED_FEAT(name, enabled) \
698 __SCHED_FEAT_##name ,
700 enum {
701 #include "sched_features.h"
704 #undef SCHED_FEAT
706 #define SCHED_FEAT(name, enabled) \
707 (1UL << __SCHED_FEAT_##name) * enabled |
709 const_debug unsigned int sysctl_sched_features =
710 #include "sched_features.h"
713 #undef SCHED_FEAT
715 #ifdef CONFIG_SCHED_DEBUG
716 #define SCHED_FEAT(name, enabled) \
717 #name ,
719 static __read_mostly char *sched_feat_names[] = {
720 #include "sched_features.h"
721 NULL
724 #undef SCHED_FEAT
726 static int sched_feat_show(struct seq_file *m, void *v)
728 int i;
730 for (i = 0; sched_feat_names[i]; i++) {
731 if (!(sysctl_sched_features & (1UL << i)))
732 seq_puts(m, "NO_");
733 seq_printf(m, "%s ", sched_feat_names[i]);
735 seq_puts(m, "\n");
737 return 0;
740 static ssize_t
741 sched_feat_write(struct file *filp, const char __user *ubuf,
742 size_t cnt, loff_t *ppos)
744 char buf[64];
745 char *cmp;
746 int neg = 0;
747 int i;
749 if (cnt > 63)
750 cnt = 63;
752 if (copy_from_user(&buf, ubuf, cnt))
753 return -EFAULT;
755 buf[cnt] = 0;
756 cmp = strstrip(buf);
758 if (strncmp(cmp, "NO_", 3) == 0) {
759 neg = 1;
760 cmp += 3;
763 for (i = 0; sched_feat_names[i]; i++) {
764 if (strcmp(cmp, sched_feat_names[i]) == 0) {
765 if (neg)
766 sysctl_sched_features &= ~(1UL << i);
767 else
768 sysctl_sched_features |= (1UL << i);
769 break;
773 if (!sched_feat_names[i])
774 return -EINVAL;
776 *ppos += cnt;
778 return cnt;
781 static int sched_feat_open(struct inode *inode, struct file *filp)
783 return single_open(filp, sched_feat_show, NULL);
786 static const struct file_operations sched_feat_fops = {
787 .open = sched_feat_open,
788 .write = sched_feat_write,
789 .read = seq_read,
790 .llseek = seq_lseek,
791 .release = single_release,
794 static __init int sched_init_debug(void)
796 debugfs_create_file("sched_features", 0644, NULL, NULL,
797 &sched_feat_fops);
799 return 0;
801 late_initcall(sched_init_debug);
803 #endif
805 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
808 * Number of tasks to iterate in a single balance run.
809 * Limited because this is done with IRQs disabled.
811 const_debug unsigned int sysctl_sched_nr_migrate = 32;
814 * period over which we average the RT time consumption, measured
815 * in ms.
817 * default: 1s
819 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
822 * period over which we measure -rt task cpu usage in us.
823 * default: 1s
825 unsigned int sysctl_sched_rt_period = 1000000;
827 static __read_mostly int scheduler_running;
830 * part of the period that we allow rt tasks to run in us.
831 * default: 0.95s
833 int sysctl_sched_rt_runtime = 950000;
835 static inline u64 global_rt_period(void)
837 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
840 static inline u64 global_rt_runtime(void)
842 if (sysctl_sched_rt_runtime < 0)
843 return RUNTIME_INF;
845 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
848 #ifndef prepare_arch_switch
849 # define prepare_arch_switch(next) do { } while (0)
850 #endif
851 #ifndef finish_arch_switch
852 # define finish_arch_switch(prev) do { } while (0)
853 #endif
855 static inline int task_current(struct rq *rq, struct task_struct *p)
857 return rq->curr == p;
860 static inline int task_running(struct rq *rq, struct task_struct *p)
862 #ifdef CONFIG_SMP
863 return p->on_cpu;
864 #else
865 return task_current(rq, p);
866 #endif
869 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
870 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
872 #ifdef CONFIG_SMP
874 * We can optimise this out completely for !SMP, because the
875 * SMP rebalancing from interrupt is the only thing that cares
876 * here.
878 next->on_cpu = 1;
879 #endif
882 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
884 #ifdef CONFIG_SMP
886 * After ->on_cpu is cleared, the task can be moved to a different CPU.
887 * We must ensure this doesn't happen until the switch is completely
888 * finished.
890 smp_wmb();
891 prev->on_cpu = 0;
892 #endif
893 #ifdef CONFIG_DEBUG_SPINLOCK
894 /* this is a valid case when another task releases the spinlock */
895 rq->lock.owner = current;
896 #endif
898 * If we are tracking spinlock dependencies then we have to
899 * fix up the runqueue lock - which gets 'carried over' from
900 * prev into current:
902 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
904 raw_spin_unlock_irq(&rq->lock);
907 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
908 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
910 #ifdef CONFIG_SMP
912 * We can optimise this out completely for !SMP, because the
913 * SMP rebalancing from interrupt is the only thing that cares
914 * here.
916 next->on_cpu = 1;
917 #endif
918 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
919 raw_spin_unlock_irq(&rq->lock);
920 #else
921 raw_spin_unlock(&rq->lock);
922 #endif
925 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
927 #ifdef CONFIG_SMP
929 * After ->on_cpu is cleared, the task can be moved to a different CPU.
930 * We must ensure this doesn't happen until the switch is completely
931 * finished.
933 smp_wmb();
934 prev->on_cpu = 0;
935 #endif
936 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
937 local_irq_enable();
938 #endif
940 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
943 * __task_rq_lock - lock the rq @p resides on.
945 static inline struct rq *__task_rq_lock(struct task_struct *p)
946 __acquires(rq->lock)
948 struct rq *rq;
950 lockdep_assert_held(&p->pi_lock);
952 for (;;) {
953 rq = task_rq(p);
954 raw_spin_lock(&rq->lock);
955 if (likely(rq == task_rq(p)))
956 return rq;
957 raw_spin_unlock(&rq->lock);
962 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
964 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
965 __acquires(p->pi_lock)
966 __acquires(rq->lock)
968 struct rq *rq;
970 for (;;) {
971 raw_spin_lock_irqsave(&p->pi_lock, *flags);
972 rq = task_rq(p);
973 raw_spin_lock(&rq->lock);
974 if (likely(rq == task_rq(p)))
975 return rq;
976 raw_spin_unlock(&rq->lock);
977 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
981 static void __task_rq_unlock(struct rq *rq)
982 __releases(rq->lock)
984 raw_spin_unlock(&rq->lock);
987 static inline void
988 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
989 __releases(rq->lock)
990 __releases(p->pi_lock)
992 raw_spin_unlock(&rq->lock);
993 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
997 * this_rq_lock - lock this runqueue and disable interrupts.
999 static struct rq *this_rq_lock(void)
1000 __acquires(rq->lock)
1002 struct rq *rq;
1004 local_irq_disable();
1005 rq = this_rq();
1006 raw_spin_lock(&rq->lock);
1008 return rq;
1011 #ifdef CONFIG_SCHED_HRTICK
1013 * Use HR-timers to deliver accurate preemption points.
1015 * Its all a bit involved since we cannot program an hrt while holding the
1016 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1017 * reschedule event.
1019 * When we get rescheduled we reprogram the hrtick_timer outside of the
1020 * rq->lock.
1024 * Use hrtick when:
1025 * - enabled by features
1026 * - hrtimer is actually high res
1028 static inline int hrtick_enabled(struct rq *rq)
1030 if (!sched_feat(HRTICK))
1031 return 0;
1032 if (!cpu_active(cpu_of(rq)))
1033 return 0;
1034 return hrtimer_is_hres_active(&rq->hrtick_timer);
1037 static void hrtick_clear(struct rq *rq)
1039 if (hrtimer_active(&rq->hrtick_timer))
1040 hrtimer_cancel(&rq->hrtick_timer);
1044 * High-resolution timer tick.
1045 * Runs from hardirq context with interrupts disabled.
1047 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1049 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1051 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1053 raw_spin_lock(&rq->lock);
1054 update_rq_clock(rq);
1055 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1056 raw_spin_unlock(&rq->lock);
1058 return HRTIMER_NORESTART;
1061 #ifdef CONFIG_SMP
1063 * called from hardirq (IPI) context
1065 static void __hrtick_start(void *arg)
1067 struct rq *rq = arg;
1069 raw_spin_lock(&rq->lock);
1070 hrtimer_restart(&rq->hrtick_timer);
1071 rq->hrtick_csd_pending = 0;
1072 raw_spin_unlock(&rq->lock);
1076 * Called to set the hrtick timer state.
1078 * called with rq->lock held and irqs disabled
1080 static void hrtick_start(struct rq *rq, u64 delay)
1082 struct hrtimer *timer = &rq->hrtick_timer;
1083 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1085 hrtimer_set_expires(timer, time);
1087 if (rq == this_rq()) {
1088 hrtimer_restart(timer);
1089 } else if (!rq->hrtick_csd_pending) {
1090 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1091 rq->hrtick_csd_pending = 1;
1095 static int
1096 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1098 int cpu = (int)(long)hcpu;
1100 switch (action) {
1101 case CPU_UP_CANCELED:
1102 case CPU_UP_CANCELED_FROZEN:
1103 case CPU_DOWN_PREPARE:
1104 case CPU_DOWN_PREPARE_FROZEN:
1105 case CPU_DEAD:
1106 case CPU_DEAD_FROZEN:
1107 hrtick_clear(cpu_rq(cpu));
1108 return NOTIFY_OK;
1111 return NOTIFY_DONE;
1114 static __init void init_hrtick(void)
1116 hotcpu_notifier(hotplug_hrtick, 0);
1118 #else
1120 * Called to set the hrtick timer state.
1122 * called with rq->lock held and irqs disabled
1124 static void hrtick_start(struct rq *rq, u64 delay)
1126 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1127 HRTIMER_MODE_REL_PINNED, 0);
1130 static inline void init_hrtick(void)
1133 #endif /* CONFIG_SMP */
1135 static void init_rq_hrtick(struct rq *rq)
1137 #ifdef CONFIG_SMP
1138 rq->hrtick_csd_pending = 0;
1140 rq->hrtick_csd.flags = 0;
1141 rq->hrtick_csd.func = __hrtick_start;
1142 rq->hrtick_csd.info = rq;
1143 #endif
1145 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1146 rq->hrtick_timer.function = hrtick;
1148 #else /* CONFIG_SCHED_HRTICK */
1149 static inline void hrtick_clear(struct rq *rq)
1153 static inline void init_rq_hrtick(struct rq *rq)
1157 static inline void init_hrtick(void)
1160 #endif /* CONFIG_SCHED_HRTICK */
1163 * resched_task - mark a task 'to be rescheduled now'.
1165 * On UP this means the setting of the need_resched flag, on SMP it
1166 * might also involve a cross-CPU call to trigger the scheduler on
1167 * the target CPU.
1169 #ifdef CONFIG_SMP
1171 #ifndef tsk_is_polling
1172 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1173 #endif
1175 static void resched_task(struct task_struct *p)
1177 int cpu;
1179 assert_raw_spin_locked(&task_rq(p)->lock);
1181 if (test_tsk_need_resched(p))
1182 return;
1184 set_tsk_need_resched(p);
1186 cpu = task_cpu(p);
1187 if (cpu == smp_processor_id())
1188 return;
1190 /* NEED_RESCHED must be visible before we test polling */
1191 smp_mb();
1192 if (!tsk_is_polling(p))
1193 smp_send_reschedule(cpu);
1196 static void resched_cpu(int cpu)
1198 struct rq *rq = cpu_rq(cpu);
1199 unsigned long flags;
1201 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1202 return;
1203 resched_task(cpu_curr(cpu));
1204 raw_spin_unlock_irqrestore(&rq->lock, flags);
1207 #ifdef CONFIG_NO_HZ
1209 * In the semi idle case, use the nearest busy cpu for migrating timers
1210 * from an idle cpu. This is good for power-savings.
1212 * We don't do similar optimization for completely idle system, as
1213 * selecting an idle cpu will add more delays to the timers than intended
1214 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1216 int get_nohz_timer_target(void)
1218 int cpu = smp_processor_id();
1219 int i;
1220 struct sched_domain *sd;
1222 rcu_read_lock();
1223 for_each_domain(cpu, sd) {
1224 for_each_cpu(i, sched_domain_span(sd)) {
1225 if (!idle_cpu(i)) {
1226 cpu = i;
1227 goto unlock;
1231 unlock:
1232 rcu_read_unlock();
1233 return cpu;
1236 * When add_timer_on() enqueues a timer into the timer wheel of an
1237 * idle CPU then this timer might expire before the next timer event
1238 * which is scheduled to wake up that CPU. In case of a completely
1239 * idle system the next event might even be infinite time into the
1240 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1241 * leaves the inner idle loop so the newly added timer is taken into
1242 * account when the CPU goes back to idle and evaluates the timer
1243 * wheel for the next timer event.
1245 void wake_up_idle_cpu(int cpu)
1247 struct rq *rq = cpu_rq(cpu);
1249 if (cpu == smp_processor_id())
1250 return;
1253 * This is safe, as this function is called with the timer
1254 * wheel base lock of (cpu) held. When the CPU is on the way
1255 * to idle and has not yet set rq->curr to idle then it will
1256 * be serialized on the timer wheel base lock and take the new
1257 * timer into account automatically.
1259 if (rq->curr != rq->idle)
1260 return;
1263 * We can set TIF_RESCHED on the idle task of the other CPU
1264 * lockless. The worst case is that the other CPU runs the
1265 * idle task through an additional NOOP schedule()
1267 set_tsk_need_resched(rq->idle);
1269 /* NEED_RESCHED must be visible before we test polling */
1270 smp_mb();
1271 if (!tsk_is_polling(rq->idle))
1272 smp_send_reschedule(cpu);
1275 #endif /* CONFIG_NO_HZ */
1277 static u64 sched_avg_period(void)
1279 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1282 static void sched_avg_update(struct rq *rq)
1284 s64 period = sched_avg_period();
1286 while ((s64)(rq->clock - rq->age_stamp) > period) {
1288 * Inline assembly required to prevent the compiler
1289 * optimising this loop into a divmod call.
1290 * See __iter_div_u64_rem() for another example of this.
1292 asm("" : "+rm" (rq->age_stamp));
1293 rq->age_stamp += period;
1294 rq->rt_avg /= 2;
1298 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1300 rq->rt_avg += rt_delta;
1301 sched_avg_update(rq);
1304 #else /* !CONFIG_SMP */
1305 static void resched_task(struct task_struct *p)
1307 assert_raw_spin_locked(&task_rq(p)->lock);
1308 set_tsk_need_resched(p);
1311 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1315 static void sched_avg_update(struct rq *rq)
1318 #endif /* CONFIG_SMP */
1320 #if BITS_PER_LONG == 32
1321 # define WMULT_CONST (~0UL)
1322 #else
1323 # define WMULT_CONST (1UL << 32)
1324 #endif
1326 #define WMULT_SHIFT 32
1329 * Shift right and round:
1331 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1334 * delta *= weight / lw
1336 static unsigned long
1337 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1338 struct load_weight *lw)
1340 u64 tmp;
1343 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1344 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1345 * 2^SCHED_LOAD_RESOLUTION.
1347 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1348 tmp = (u64)delta_exec * scale_load_down(weight);
1349 else
1350 tmp = (u64)delta_exec;
1352 if (!lw->inv_weight) {
1353 unsigned long w = scale_load_down(lw->weight);
1355 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1356 lw->inv_weight = 1;
1357 else if (unlikely(!w))
1358 lw->inv_weight = WMULT_CONST;
1359 else
1360 lw->inv_weight = WMULT_CONST / w;
1364 * Check whether we'd overflow the 64-bit multiplication:
1366 if (unlikely(tmp > WMULT_CONST))
1367 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1368 WMULT_SHIFT/2);
1369 else
1370 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1372 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1375 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1377 lw->weight += inc;
1378 lw->inv_weight = 0;
1381 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1383 lw->weight -= dec;
1384 lw->inv_weight = 0;
1387 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1389 lw->weight = w;
1390 lw->inv_weight = 0;
1394 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1395 * of tasks with abnormal "nice" values across CPUs the contribution that
1396 * each task makes to its run queue's load is weighted according to its
1397 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1398 * scaled version of the new time slice allocation that they receive on time
1399 * slice expiry etc.
1402 #define WEIGHT_IDLEPRIO 3
1403 #define WMULT_IDLEPRIO 1431655765
1406 * Nice levels are multiplicative, with a gentle 10% change for every
1407 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1408 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1409 * that remained on nice 0.
1411 * The "10% effect" is relative and cumulative: from _any_ nice level,
1412 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1413 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1414 * If a task goes up by ~10% and another task goes down by ~10% then
1415 * the relative distance between them is ~25%.)
1417 static const int prio_to_weight[40] = {
1418 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1419 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1420 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1421 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1422 /* 0 */ 1024, 820, 655, 526, 423,
1423 /* 5 */ 335, 272, 215, 172, 137,
1424 /* 10 */ 110, 87, 70, 56, 45,
1425 /* 15 */ 36, 29, 23, 18, 15,
1429 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1431 * In cases where the weight does not change often, we can use the
1432 * precalculated inverse to speed up arithmetics by turning divisions
1433 * into multiplications:
1435 static const u32 prio_to_wmult[40] = {
1436 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1437 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1438 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1439 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1440 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1441 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1442 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1443 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1446 /* Time spent by the tasks of the cpu accounting group executing in ... */
1447 enum cpuacct_stat_index {
1448 CPUACCT_STAT_USER, /* ... user mode */
1449 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1451 CPUACCT_STAT_NSTATS,
1454 #ifdef CONFIG_CGROUP_CPUACCT
1455 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1456 static void cpuacct_update_stats(struct task_struct *tsk,
1457 enum cpuacct_stat_index idx, cputime_t val);
1458 #else
1459 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1460 static inline void cpuacct_update_stats(struct task_struct *tsk,
1461 enum cpuacct_stat_index idx, cputime_t val) {}
1462 #endif
1464 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1466 update_load_add(&rq->load, load);
1469 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1471 update_load_sub(&rq->load, load);
1474 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1475 typedef int (*tg_visitor)(struct task_group *, void *);
1478 * Iterate the full tree, calling @down when first entering a node and @up when
1479 * leaving it for the final time.
1481 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1483 struct task_group *parent, *child;
1484 int ret;
1486 rcu_read_lock();
1487 parent = &root_task_group;
1488 down:
1489 ret = (*down)(parent, data);
1490 if (ret)
1491 goto out_unlock;
1492 list_for_each_entry_rcu(child, &parent->children, siblings) {
1493 parent = child;
1494 goto down;
1497 continue;
1499 ret = (*up)(parent, data);
1500 if (ret)
1501 goto out_unlock;
1503 child = parent;
1504 parent = parent->parent;
1505 if (parent)
1506 goto up;
1507 out_unlock:
1508 rcu_read_unlock();
1510 return ret;
1513 static int tg_nop(struct task_group *tg, void *data)
1515 return 0;
1517 #endif
1519 #ifdef CONFIG_SMP
1520 /* Used instead of source_load when we know the type == 0 */
1521 static unsigned long weighted_cpuload(const int cpu)
1523 return cpu_rq(cpu)->load.weight;
1527 * Return a low guess at the load of a migration-source cpu weighted
1528 * according to the scheduling class and "nice" value.
1530 * We want to under-estimate the load of migration sources, to
1531 * balance conservatively.
1533 static unsigned long source_load(int cpu, int type)
1535 struct rq *rq = cpu_rq(cpu);
1536 unsigned long total = weighted_cpuload(cpu);
1538 if (type == 0 || !sched_feat(LB_BIAS))
1539 return total;
1541 return min(rq->cpu_load[type-1], total);
1545 * Return a high guess at the load of a migration-target cpu weighted
1546 * according to the scheduling class and "nice" value.
1548 static unsigned long target_load(int cpu, int type)
1550 struct rq *rq = cpu_rq(cpu);
1551 unsigned long total = weighted_cpuload(cpu);
1553 if (type == 0 || !sched_feat(LB_BIAS))
1554 return total;
1556 return max(rq->cpu_load[type-1], total);
1559 static unsigned long power_of(int cpu)
1561 return cpu_rq(cpu)->cpu_power;
1564 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1566 static unsigned long cpu_avg_load_per_task(int cpu)
1568 struct rq *rq = cpu_rq(cpu);
1569 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1571 if (nr_running)
1572 rq->avg_load_per_task = rq->load.weight / nr_running;
1573 else
1574 rq->avg_load_per_task = 0;
1576 return rq->avg_load_per_task;
1579 #ifdef CONFIG_PREEMPT
1581 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1584 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1585 * way at the expense of forcing extra atomic operations in all
1586 * invocations. This assures that the double_lock is acquired using the
1587 * same underlying policy as the spinlock_t on this architecture, which
1588 * reduces latency compared to the unfair variant below. However, it
1589 * also adds more overhead and therefore may reduce throughput.
1591 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1592 __releases(this_rq->lock)
1593 __acquires(busiest->lock)
1594 __acquires(this_rq->lock)
1596 raw_spin_unlock(&this_rq->lock);
1597 double_rq_lock(this_rq, busiest);
1599 return 1;
1602 #else
1604 * Unfair double_lock_balance: Optimizes throughput at the expense of
1605 * latency by eliminating extra atomic operations when the locks are
1606 * already in proper order on entry. This favors lower cpu-ids and will
1607 * grant the double lock to lower cpus over higher ids under contention,
1608 * regardless of entry order into the function.
1610 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1611 __releases(this_rq->lock)
1612 __acquires(busiest->lock)
1613 __acquires(this_rq->lock)
1615 int ret = 0;
1617 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1618 if (busiest < this_rq) {
1619 raw_spin_unlock(&this_rq->lock);
1620 raw_spin_lock(&busiest->lock);
1621 raw_spin_lock_nested(&this_rq->lock,
1622 SINGLE_DEPTH_NESTING);
1623 ret = 1;
1624 } else
1625 raw_spin_lock_nested(&busiest->lock,
1626 SINGLE_DEPTH_NESTING);
1628 return ret;
1631 #endif /* CONFIG_PREEMPT */
1634 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1636 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1638 if (unlikely(!irqs_disabled())) {
1639 /* printk() doesn't work good under rq->lock */
1640 raw_spin_unlock(&this_rq->lock);
1641 BUG_ON(1);
1644 return _double_lock_balance(this_rq, busiest);
1647 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1648 __releases(busiest->lock)
1650 raw_spin_unlock(&busiest->lock);
1651 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1655 * double_rq_lock - safely lock two runqueues
1657 * Note this does not disable interrupts like task_rq_lock,
1658 * you need to do so manually before calling.
1660 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1661 __acquires(rq1->lock)
1662 __acquires(rq2->lock)
1664 BUG_ON(!irqs_disabled());
1665 if (rq1 == rq2) {
1666 raw_spin_lock(&rq1->lock);
1667 __acquire(rq2->lock); /* Fake it out ;) */
1668 } else {
1669 if (rq1 < rq2) {
1670 raw_spin_lock(&rq1->lock);
1671 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1672 } else {
1673 raw_spin_lock(&rq2->lock);
1674 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1680 * double_rq_unlock - safely unlock two runqueues
1682 * Note this does not restore interrupts like task_rq_unlock,
1683 * you need to do so manually after calling.
1685 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1686 __releases(rq1->lock)
1687 __releases(rq2->lock)
1689 raw_spin_unlock(&rq1->lock);
1690 if (rq1 != rq2)
1691 raw_spin_unlock(&rq2->lock);
1692 else
1693 __release(rq2->lock);
1696 #else /* CONFIG_SMP */
1699 * double_rq_lock - safely lock two runqueues
1701 * Note this does not disable interrupts like task_rq_lock,
1702 * you need to do so manually before calling.
1704 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1705 __acquires(rq1->lock)
1706 __acquires(rq2->lock)
1708 BUG_ON(!irqs_disabled());
1709 BUG_ON(rq1 != rq2);
1710 raw_spin_lock(&rq1->lock);
1711 __acquire(rq2->lock); /* Fake it out ;) */
1715 * double_rq_unlock - safely unlock two runqueues
1717 * Note this does not restore interrupts like task_rq_unlock,
1718 * you need to do so manually after calling.
1720 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1721 __releases(rq1->lock)
1722 __releases(rq2->lock)
1724 BUG_ON(rq1 != rq2);
1725 raw_spin_unlock(&rq1->lock);
1726 __release(rq2->lock);
1729 #endif
1731 static void calc_load_account_idle(struct rq *this_rq);
1732 static void update_sysctl(void);
1733 static int get_update_sysctl_factor(void);
1734 static void update_cpu_load(struct rq *this_rq);
1736 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1738 set_task_rq(p, cpu);
1739 #ifdef CONFIG_SMP
1741 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1742 * successfuly executed on another CPU. We must ensure that updates of
1743 * per-task data have been completed by this moment.
1745 smp_wmb();
1746 task_thread_info(p)->cpu = cpu;
1747 #endif
1750 static const struct sched_class rt_sched_class;
1752 #define sched_class_highest (&stop_sched_class)
1753 #define for_each_class(class) \
1754 for (class = sched_class_highest; class; class = class->next)
1756 #include "sched_stats.h"
1758 static void inc_nr_running(struct rq *rq)
1760 rq->nr_running++;
1763 static void dec_nr_running(struct rq *rq)
1765 rq->nr_running--;
1768 static void set_load_weight(struct task_struct *p)
1770 int prio = p->static_prio - MAX_RT_PRIO;
1771 struct load_weight *load = &p->se.load;
1774 * SCHED_IDLE tasks get minimal weight:
1776 if (p->policy == SCHED_IDLE) {
1777 load->weight = scale_load(WEIGHT_IDLEPRIO);
1778 load->inv_weight = WMULT_IDLEPRIO;
1779 return;
1782 load->weight = scale_load(prio_to_weight[prio]);
1783 load->inv_weight = prio_to_wmult[prio];
1786 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1788 update_rq_clock(rq);
1789 sched_info_queued(p);
1790 p->sched_class->enqueue_task(rq, p, flags);
1793 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1795 update_rq_clock(rq);
1796 sched_info_dequeued(p);
1797 p->sched_class->dequeue_task(rq, p, flags);
1801 * activate_task - move a task to the runqueue.
1803 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1805 if (task_contributes_to_load(p))
1806 rq->nr_uninterruptible--;
1808 enqueue_task(rq, p, flags);
1809 inc_nr_running(rq);
1813 * deactivate_task - remove a task from the runqueue.
1815 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1817 if (task_contributes_to_load(p))
1818 rq->nr_uninterruptible++;
1820 dequeue_task(rq, p, flags);
1821 dec_nr_running(rq);
1824 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1827 * There are no locks covering percpu hardirq/softirq time.
1828 * They are only modified in account_system_vtime, on corresponding CPU
1829 * with interrupts disabled. So, writes are safe.
1830 * They are read and saved off onto struct rq in update_rq_clock().
1831 * This may result in other CPU reading this CPU's irq time and can
1832 * race with irq/account_system_vtime on this CPU. We would either get old
1833 * or new value with a side effect of accounting a slice of irq time to wrong
1834 * task when irq is in progress while we read rq->clock. That is a worthy
1835 * compromise in place of having locks on each irq in account_system_time.
1837 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1838 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1840 static DEFINE_PER_CPU(u64, irq_start_time);
1841 static int sched_clock_irqtime;
1843 void enable_sched_clock_irqtime(void)
1845 sched_clock_irqtime = 1;
1848 void disable_sched_clock_irqtime(void)
1850 sched_clock_irqtime = 0;
1853 #ifndef CONFIG_64BIT
1854 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1856 static inline void irq_time_write_begin(void)
1858 __this_cpu_inc(irq_time_seq.sequence);
1859 smp_wmb();
1862 static inline void irq_time_write_end(void)
1864 smp_wmb();
1865 __this_cpu_inc(irq_time_seq.sequence);
1868 static inline u64 irq_time_read(int cpu)
1870 u64 irq_time;
1871 unsigned seq;
1873 do {
1874 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1875 irq_time = per_cpu(cpu_softirq_time, cpu) +
1876 per_cpu(cpu_hardirq_time, cpu);
1877 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1879 return irq_time;
1881 #else /* CONFIG_64BIT */
1882 static inline void irq_time_write_begin(void)
1886 static inline void irq_time_write_end(void)
1890 static inline u64 irq_time_read(int cpu)
1892 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1894 #endif /* CONFIG_64BIT */
1897 * Called before incrementing preempt_count on {soft,}irq_enter
1898 * and before decrementing preempt_count on {soft,}irq_exit.
1900 void account_system_vtime(struct task_struct *curr)
1902 unsigned long flags;
1903 s64 delta;
1904 int cpu;
1906 if (!sched_clock_irqtime)
1907 return;
1909 local_irq_save(flags);
1911 cpu = smp_processor_id();
1912 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1913 __this_cpu_add(irq_start_time, delta);
1915 irq_time_write_begin();
1917 * We do not account for softirq time from ksoftirqd here.
1918 * We want to continue accounting softirq time to ksoftirqd thread
1919 * in that case, so as not to confuse scheduler with a special task
1920 * that do not consume any time, but still wants to run.
1922 if (hardirq_count())
1923 __this_cpu_add(cpu_hardirq_time, delta);
1924 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1925 __this_cpu_add(cpu_softirq_time, delta);
1927 irq_time_write_end();
1928 local_irq_restore(flags);
1930 EXPORT_SYMBOL_GPL(account_system_vtime);
1932 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1934 #ifdef CONFIG_PARAVIRT
1935 static inline u64 steal_ticks(u64 steal)
1937 if (unlikely(steal > NSEC_PER_SEC))
1938 return div_u64(steal, TICK_NSEC);
1940 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
1942 #endif
1944 static void update_rq_clock_task(struct rq *rq, s64 delta)
1947 * In theory, the compile should just see 0 here, and optimize out the call
1948 * to sched_rt_avg_update. But I don't trust it...
1950 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
1951 s64 steal = 0, irq_delta = 0;
1952 #endif
1953 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1954 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1957 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1958 * this case when a previous update_rq_clock() happened inside a
1959 * {soft,}irq region.
1961 * When this happens, we stop ->clock_task and only update the
1962 * prev_irq_time stamp to account for the part that fit, so that a next
1963 * update will consume the rest. This ensures ->clock_task is
1964 * monotonic.
1966 * It does however cause some slight miss-attribution of {soft,}irq
1967 * time, a more accurate solution would be to update the irq_time using
1968 * the current rq->clock timestamp, except that would require using
1969 * atomic ops.
1971 if (irq_delta > delta)
1972 irq_delta = delta;
1974 rq->prev_irq_time += irq_delta;
1975 delta -= irq_delta;
1976 #endif
1977 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
1978 if (static_branch((&paravirt_steal_rq_enabled))) {
1979 u64 st;
1981 steal = paravirt_steal_clock(cpu_of(rq));
1982 steal -= rq->prev_steal_time_rq;
1984 if (unlikely(steal > delta))
1985 steal = delta;
1987 st = steal_ticks(steal);
1988 steal = st * TICK_NSEC;
1990 rq->prev_steal_time_rq += steal;
1992 delta -= steal;
1994 #endif
1996 rq->clock_task += delta;
1998 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
1999 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
2000 sched_rt_avg_update(rq, irq_delta + steal);
2001 #endif
2004 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2005 static int irqtime_account_hi_update(void)
2007 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2008 unsigned long flags;
2009 u64 latest_ns;
2010 int ret = 0;
2012 local_irq_save(flags);
2013 latest_ns = this_cpu_read(cpu_hardirq_time);
2014 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
2015 ret = 1;
2016 local_irq_restore(flags);
2017 return ret;
2020 static int irqtime_account_si_update(void)
2022 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2023 unsigned long flags;
2024 u64 latest_ns;
2025 int ret = 0;
2027 local_irq_save(flags);
2028 latest_ns = this_cpu_read(cpu_softirq_time);
2029 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2030 ret = 1;
2031 local_irq_restore(flags);
2032 return ret;
2035 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2037 #define sched_clock_irqtime (0)
2039 #endif
2041 #include "sched_idletask.c"
2042 #include "sched_fair.c"
2043 #include "sched_rt.c"
2044 #include "sched_autogroup.c"
2045 #include "sched_stoptask.c"
2046 #ifdef CONFIG_SCHED_DEBUG
2047 # include "sched_debug.c"
2048 #endif
2050 void sched_set_stop_task(int cpu, struct task_struct *stop)
2052 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2053 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2055 if (stop) {
2057 * Make it appear like a SCHED_FIFO task, its something
2058 * userspace knows about and won't get confused about.
2060 * Also, it will make PI more or less work without too
2061 * much confusion -- but then, stop work should not
2062 * rely on PI working anyway.
2064 sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
2066 stop->sched_class = &stop_sched_class;
2069 cpu_rq(cpu)->stop = stop;
2071 if (old_stop) {
2073 * Reset it back to a normal scheduling class so that
2074 * it can die in pieces.
2076 old_stop->sched_class = &rt_sched_class;
2081 * __normal_prio - return the priority that is based on the static prio
2083 static inline int __normal_prio(struct task_struct *p)
2085 return p->static_prio;
2089 * Calculate the expected normal priority: i.e. priority
2090 * without taking RT-inheritance into account. Might be
2091 * boosted by interactivity modifiers. Changes upon fork,
2092 * setprio syscalls, and whenever the interactivity
2093 * estimator recalculates.
2095 static inline int normal_prio(struct task_struct *p)
2097 int prio;
2099 if (task_has_rt_policy(p))
2100 prio = MAX_RT_PRIO-1 - p->rt_priority;
2101 else
2102 prio = __normal_prio(p);
2103 return prio;
2107 * Calculate the current priority, i.e. the priority
2108 * taken into account by the scheduler. This value might
2109 * be boosted by RT tasks, or might be boosted by
2110 * interactivity modifiers. Will be RT if the task got
2111 * RT-boosted. If not then it returns p->normal_prio.
2113 static int effective_prio(struct task_struct *p)
2115 p->normal_prio = normal_prio(p);
2117 * If we are RT tasks or we were boosted to RT priority,
2118 * keep the priority unchanged. Otherwise, update priority
2119 * to the normal priority:
2121 if (!rt_prio(p->prio))
2122 return p->normal_prio;
2123 return p->prio;
2127 * task_curr - is this task currently executing on a CPU?
2128 * @p: the task in question.
2130 inline int task_curr(const struct task_struct *p)
2132 return cpu_curr(task_cpu(p)) == p;
2135 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2136 const struct sched_class *prev_class,
2137 int oldprio)
2139 if (prev_class != p->sched_class) {
2140 if (prev_class->switched_from)
2141 prev_class->switched_from(rq, p);
2142 p->sched_class->switched_to(rq, p);
2143 } else if (oldprio != p->prio)
2144 p->sched_class->prio_changed(rq, p, oldprio);
2147 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2149 const struct sched_class *class;
2151 if (p->sched_class == rq->curr->sched_class) {
2152 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2153 } else {
2154 for_each_class(class) {
2155 if (class == rq->curr->sched_class)
2156 break;
2157 if (class == p->sched_class) {
2158 resched_task(rq->curr);
2159 break;
2165 * A queue event has occurred, and we're going to schedule. In
2166 * this case, we can save a useless back to back clock update.
2168 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2169 rq->skip_clock_update = 1;
2172 #ifdef CONFIG_SMP
2174 * Is this task likely cache-hot:
2176 static int
2177 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2179 s64 delta;
2181 if (p->sched_class != &fair_sched_class)
2182 return 0;
2184 if (unlikely(p->policy == SCHED_IDLE))
2185 return 0;
2188 * Buddy candidates are cache hot:
2190 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2191 (&p->se == cfs_rq_of(&p->se)->next ||
2192 &p->se == cfs_rq_of(&p->se)->last))
2193 return 1;
2195 if (sysctl_sched_migration_cost == -1)
2196 return 1;
2197 if (sysctl_sched_migration_cost == 0)
2198 return 0;
2200 delta = now - p->se.exec_start;
2202 return delta < (s64)sysctl_sched_migration_cost;
2205 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2207 #ifdef CONFIG_SCHED_DEBUG
2209 * We should never call set_task_cpu() on a blocked task,
2210 * ttwu() will sort out the placement.
2212 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2213 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2215 #ifdef CONFIG_LOCKDEP
2217 * The caller should hold either p->pi_lock or rq->lock, when changing
2218 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2220 * sched_move_task() holds both and thus holding either pins the cgroup,
2221 * see set_task_rq().
2223 * Furthermore, all task_rq users should acquire both locks, see
2224 * task_rq_lock().
2226 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2227 lockdep_is_held(&task_rq(p)->lock)));
2228 #endif
2229 #endif
2231 trace_sched_migrate_task(p, new_cpu);
2233 if (task_cpu(p) != new_cpu) {
2234 p->se.nr_migrations++;
2235 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
2238 __set_task_cpu(p, new_cpu);
2241 struct migration_arg {
2242 struct task_struct *task;
2243 int dest_cpu;
2246 static int migration_cpu_stop(void *data);
2249 * wait_task_inactive - wait for a thread to unschedule.
2251 * If @match_state is nonzero, it's the @p->state value just checked and
2252 * not expected to change. If it changes, i.e. @p might have woken up,
2253 * then return zero. When we succeed in waiting for @p to be off its CPU,
2254 * we return a positive number (its total switch count). If a second call
2255 * a short while later returns the same number, the caller can be sure that
2256 * @p has remained unscheduled the whole time.
2258 * The caller must ensure that the task *will* unschedule sometime soon,
2259 * else this function might spin for a *long* time. This function can't
2260 * be called with interrupts off, or it may introduce deadlock with
2261 * smp_call_function() if an IPI is sent by the same process we are
2262 * waiting to become inactive.
2264 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2266 unsigned long flags;
2267 int running, on_rq;
2268 unsigned long ncsw;
2269 struct rq *rq;
2271 for (;;) {
2273 * We do the initial early heuristics without holding
2274 * any task-queue locks at all. We'll only try to get
2275 * the runqueue lock when things look like they will
2276 * work out!
2278 rq = task_rq(p);
2281 * If the task is actively running on another CPU
2282 * still, just relax and busy-wait without holding
2283 * any locks.
2285 * NOTE! Since we don't hold any locks, it's not
2286 * even sure that "rq" stays as the right runqueue!
2287 * But we don't care, since "task_running()" will
2288 * return false if the runqueue has changed and p
2289 * is actually now running somewhere else!
2291 while (task_running(rq, p)) {
2292 if (match_state && unlikely(p->state != match_state))
2293 return 0;
2294 cpu_relax();
2298 * Ok, time to look more closely! We need the rq
2299 * lock now, to be *sure*. If we're wrong, we'll
2300 * just go back and repeat.
2302 rq = task_rq_lock(p, &flags);
2303 trace_sched_wait_task(p);
2304 running = task_running(rq, p);
2305 on_rq = p->on_rq;
2306 ncsw = 0;
2307 if (!match_state || p->state == match_state)
2308 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2309 task_rq_unlock(rq, p, &flags);
2312 * If it changed from the expected state, bail out now.
2314 if (unlikely(!ncsw))
2315 break;
2318 * Was it really running after all now that we
2319 * checked with the proper locks actually held?
2321 * Oops. Go back and try again..
2323 if (unlikely(running)) {
2324 cpu_relax();
2325 continue;
2329 * It's not enough that it's not actively running,
2330 * it must be off the runqueue _entirely_, and not
2331 * preempted!
2333 * So if it was still runnable (but just not actively
2334 * running right now), it's preempted, and we should
2335 * yield - it could be a while.
2337 if (unlikely(on_rq)) {
2338 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2340 set_current_state(TASK_UNINTERRUPTIBLE);
2341 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2342 continue;
2346 * Ahh, all good. It wasn't running, and it wasn't
2347 * runnable, which means that it will never become
2348 * running in the future either. We're all done!
2350 break;
2353 return ncsw;
2356 /***
2357 * kick_process - kick a running thread to enter/exit the kernel
2358 * @p: the to-be-kicked thread
2360 * Cause a process which is running on another CPU to enter
2361 * kernel-mode, without any delay. (to get signals handled.)
2363 * NOTE: this function doesn't have to take the runqueue lock,
2364 * because all it wants to ensure is that the remote task enters
2365 * the kernel. If the IPI races and the task has been migrated
2366 * to another CPU then no harm is done and the purpose has been
2367 * achieved as well.
2369 void kick_process(struct task_struct *p)
2371 int cpu;
2373 preempt_disable();
2374 cpu = task_cpu(p);
2375 if ((cpu != smp_processor_id()) && task_curr(p))
2376 smp_send_reschedule(cpu);
2377 preempt_enable();
2379 EXPORT_SYMBOL_GPL(kick_process);
2380 #endif /* CONFIG_SMP */
2382 #ifdef CONFIG_SMP
2384 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2386 static int select_fallback_rq(int cpu, struct task_struct *p)
2388 int dest_cpu;
2389 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2391 /* Look for allowed, online CPU in same node. */
2392 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2393 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2394 return dest_cpu;
2396 /* Any allowed, online CPU? */
2397 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2398 if (dest_cpu < nr_cpu_ids)
2399 return dest_cpu;
2401 /* No more Mr. Nice Guy. */
2402 dest_cpu = cpuset_cpus_allowed_fallback(p);
2404 * Don't tell them about moving exiting tasks or
2405 * kernel threads (both mm NULL), since they never
2406 * leave kernel.
2408 if (p->mm && printk_ratelimit()) {
2409 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2410 task_pid_nr(p), p->comm, cpu);
2413 return dest_cpu;
2417 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2419 static inline
2420 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2422 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2425 * In order not to call set_task_cpu() on a blocking task we need
2426 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2427 * cpu.
2429 * Since this is common to all placement strategies, this lives here.
2431 * [ this allows ->select_task() to simply return task_cpu(p) and
2432 * not worry about this generic constraint ]
2434 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2435 !cpu_online(cpu)))
2436 cpu = select_fallback_rq(task_cpu(p), p);
2438 return cpu;
2441 static void update_avg(u64 *avg, u64 sample)
2443 s64 diff = sample - *avg;
2444 *avg += diff >> 3;
2446 #endif
2448 static void
2449 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2451 #ifdef CONFIG_SCHEDSTATS
2452 struct rq *rq = this_rq();
2454 #ifdef CONFIG_SMP
2455 int this_cpu = smp_processor_id();
2457 if (cpu == this_cpu) {
2458 schedstat_inc(rq, ttwu_local);
2459 schedstat_inc(p, se.statistics.nr_wakeups_local);
2460 } else {
2461 struct sched_domain *sd;
2463 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2464 rcu_read_lock();
2465 for_each_domain(this_cpu, sd) {
2466 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2467 schedstat_inc(sd, ttwu_wake_remote);
2468 break;
2471 rcu_read_unlock();
2474 if (wake_flags & WF_MIGRATED)
2475 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2477 #endif /* CONFIG_SMP */
2479 schedstat_inc(rq, ttwu_count);
2480 schedstat_inc(p, se.statistics.nr_wakeups);
2482 if (wake_flags & WF_SYNC)
2483 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2485 #endif /* CONFIG_SCHEDSTATS */
2488 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2490 activate_task(rq, p, en_flags);
2491 p->on_rq = 1;
2493 /* if a worker is waking up, notify workqueue */
2494 if (p->flags & PF_WQ_WORKER)
2495 wq_worker_waking_up(p, cpu_of(rq));
2499 * Mark the task runnable and perform wakeup-preemption.
2501 static void
2502 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2504 trace_sched_wakeup(p, true);
2505 check_preempt_curr(rq, p, wake_flags);
2507 p->state = TASK_RUNNING;
2508 #ifdef CONFIG_SMP
2509 if (p->sched_class->task_woken)
2510 p->sched_class->task_woken(rq, p);
2512 if (rq->idle_stamp) {
2513 u64 delta = rq->clock - rq->idle_stamp;
2514 u64 max = 2*sysctl_sched_migration_cost;
2516 if (delta > max)
2517 rq->avg_idle = max;
2518 else
2519 update_avg(&rq->avg_idle, delta);
2520 rq->idle_stamp = 0;
2522 #endif
2525 static void
2526 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2528 #ifdef CONFIG_SMP
2529 if (p->sched_contributes_to_load)
2530 rq->nr_uninterruptible--;
2531 #endif
2533 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2534 ttwu_do_wakeup(rq, p, wake_flags);
2538 * Called in case the task @p isn't fully descheduled from its runqueue,
2539 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2540 * since all we need to do is flip p->state to TASK_RUNNING, since
2541 * the task is still ->on_rq.
2543 static int ttwu_remote(struct task_struct *p, int wake_flags)
2545 struct rq *rq;
2546 int ret = 0;
2548 rq = __task_rq_lock(p);
2549 if (p->on_rq) {
2550 ttwu_do_wakeup(rq, p, wake_flags);
2551 ret = 1;
2553 __task_rq_unlock(rq);
2555 return ret;
2558 #ifdef CONFIG_SMP
2559 static void sched_ttwu_do_pending(struct task_struct *list)
2561 struct rq *rq = this_rq();
2563 raw_spin_lock(&rq->lock);
2565 while (list) {
2566 struct task_struct *p = list;
2567 list = list->wake_entry;
2568 ttwu_do_activate(rq, p, 0);
2571 raw_spin_unlock(&rq->lock);
2574 #ifdef CONFIG_HOTPLUG_CPU
2576 static void sched_ttwu_pending(void)
2578 struct rq *rq = this_rq();
2579 struct task_struct *list = xchg(&rq->wake_list, NULL);
2581 if (!list)
2582 return;
2584 sched_ttwu_do_pending(list);
2587 #endif /* CONFIG_HOTPLUG_CPU */
2589 void scheduler_ipi(void)
2591 struct rq *rq = this_rq();
2592 struct task_struct *list = xchg(&rq->wake_list, NULL);
2594 if (!list)
2595 return;
2598 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2599 * traditionally all their work was done from the interrupt return
2600 * path. Now that we actually do some work, we need to make sure
2601 * we do call them.
2603 * Some archs already do call them, luckily irq_enter/exit nest
2604 * properly.
2606 * Arguably we should visit all archs and update all handlers,
2607 * however a fair share of IPIs are still resched only so this would
2608 * somewhat pessimize the simple resched case.
2610 irq_enter();
2611 sched_ttwu_do_pending(list);
2612 irq_exit();
2615 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2617 struct rq *rq = cpu_rq(cpu);
2618 struct task_struct *next = rq->wake_list;
2620 for (;;) {
2621 struct task_struct *old = next;
2623 p->wake_entry = next;
2624 next = cmpxchg(&rq->wake_list, old, p);
2625 if (next == old)
2626 break;
2629 if (!next)
2630 smp_send_reschedule(cpu);
2633 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2634 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2636 struct rq *rq;
2637 int ret = 0;
2639 rq = __task_rq_lock(p);
2640 if (p->on_cpu) {
2641 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2642 ttwu_do_wakeup(rq, p, wake_flags);
2643 ret = 1;
2645 __task_rq_unlock(rq);
2647 return ret;
2650 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2651 #endif /* CONFIG_SMP */
2653 static void ttwu_queue(struct task_struct *p, int cpu)
2655 struct rq *rq = cpu_rq(cpu);
2657 #if defined(CONFIG_SMP)
2658 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2659 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2660 ttwu_queue_remote(p, cpu);
2661 return;
2663 #endif
2665 raw_spin_lock(&rq->lock);
2666 ttwu_do_activate(rq, p, 0);
2667 raw_spin_unlock(&rq->lock);
2671 * try_to_wake_up - wake up a thread
2672 * @p: the thread to be awakened
2673 * @state: the mask of task states that can be woken
2674 * @wake_flags: wake modifier flags (WF_*)
2676 * Put it on the run-queue if it's not already there. The "current"
2677 * thread is always on the run-queue (except when the actual
2678 * re-schedule is in progress), and as such you're allowed to do
2679 * the simpler "current->state = TASK_RUNNING" to mark yourself
2680 * runnable without the overhead of this.
2682 * Returns %true if @p was woken up, %false if it was already running
2683 * or @state didn't match @p's state.
2685 static int
2686 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2688 unsigned long flags;
2689 int cpu, success = 0;
2691 smp_wmb();
2692 raw_spin_lock_irqsave(&p->pi_lock, flags);
2693 if (!(p->state & state))
2694 goto out;
2696 success = 1; /* we're going to change ->state */
2697 cpu = task_cpu(p);
2699 if (p->on_rq && ttwu_remote(p, wake_flags))
2700 goto stat;
2702 #ifdef CONFIG_SMP
2704 * If the owning (remote) cpu is still in the middle of schedule() with
2705 * this task as prev, wait until its done referencing the task.
2707 while (p->on_cpu) {
2708 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2710 * In case the architecture enables interrupts in
2711 * context_switch(), we cannot busy wait, since that
2712 * would lead to deadlocks when an interrupt hits and
2713 * tries to wake up @prev. So bail and do a complete
2714 * remote wakeup.
2716 if (ttwu_activate_remote(p, wake_flags))
2717 goto stat;
2718 #else
2719 cpu_relax();
2720 #endif
2723 * Pairs with the smp_wmb() in finish_lock_switch().
2725 smp_rmb();
2727 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2728 p->state = TASK_WAKING;
2730 if (p->sched_class->task_waking)
2731 p->sched_class->task_waking(p);
2733 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2734 if (task_cpu(p) != cpu) {
2735 wake_flags |= WF_MIGRATED;
2736 set_task_cpu(p, cpu);
2738 #endif /* CONFIG_SMP */
2740 ttwu_queue(p, cpu);
2741 stat:
2742 ttwu_stat(p, cpu, wake_flags);
2743 out:
2744 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2746 return success;
2750 * try_to_wake_up_local - try to wake up a local task with rq lock held
2751 * @p: the thread to be awakened
2753 * Put @p on the run-queue if it's not already there. The caller must
2754 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2755 * the current task.
2757 static void try_to_wake_up_local(struct task_struct *p)
2759 struct rq *rq = task_rq(p);
2761 BUG_ON(rq != this_rq());
2762 BUG_ON(p == current);
2763 lockdep_assert_held(&rq->lock);
2765 if (!raw_spin_trylock(&p->pi_lock)) {
2766 raw_spin_unlock(&rq->lock);
2767 raw_spin_lock(&p->pi_lock);
2768 raw_spin_lock(&rq->lock);
2771 if (!(p->state & TASK_NORMAL))
2772 goto out;
2774 if (!p->on_rq)
2775 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2777 ttwu_do_wakeup(rq, p, 0);
2778 ttwu_stat(p, smp_processor_id(), 0);
2779 out:
2780 raw_spin_unlock(&p->pi_lock);
2784 * wake_up_process - Wake up a specific process
2785 * @p: The process to be woken up.
2787 * Attempt to wake up the nominated process and move it to the set of runnable
2788 * processes. Returns 1 if the process was woken up, 0 if it was already
2789 * running.
2791 * It may be assumed that this function implies a write memory barrier before
2792 * changing the task state if and only if any tasks are woken up.
2794 int wake_up_process(struct task_struct *p)
2796 return try_to_wake_up(p, TASK_ALL, 0);
2798 EXPORT_SYMBOL(wake_up_process);
2800 int wake_up_state(struct task_struct *p, unsigned int state)
2802 return try_to_wake_up(p, state, 0);
2806 * Perform scheduler related setup for a newly forked process p.
2807 * p is forked by current.
2809 * __sched_fork() is basic setup used by init_idle() too:
2811 static void __sched_fork(struct task_struct *p)
2813 p->on_rq = 0;
2815 p->se.on_rq = 0;
2816 p->se.exec_start = 0;
2817 p->se.sum_exec_runtime = 0;
2818 p->se.prev_sum_exec_runtime = 0;
2819 p->se.nr_migrations = 0;
2820 p->se.vruntime = 0;
2821 INIT_LIST_HEAD(&p->se.group_node);
2823 #ifdef CONFIG_SCHEDSTATS
2824 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2825 #endif
2827 INIT_LIST_HEAD(&p->rt.run_list);
2829 #ifdef CONFIG_PREEMPT_NOTIFIERS
2830 INIT_HLIST_HEAD(&p->preempt_notifiers);
2831 #endif
2835 * fork()/clone()-time setup:
2837 void sched_fork(struct task_struct *p)
2839 unsigned long flags;
2840 int cpu = get_cpu();
2842 __sched_fork(p);
2844 * We mark the process as running here. This guarantees that
2845 * nobody will actually run it, and a signal or other external
2846 * event cannot wake it up and insert it on the runqueue either.
2848 p->state = TASK_RUNNING;
2851 * Revert to default priority/policy on fork if requested.
2853 if (unlikely(p->sched_reset_on_fork)) {
2854 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2855 p->policy = SCHED_NORMAL;
2856 p->normal_prio = p->static_prio;
2859 if (PRIO_TO_NICE(p->static_prio) < 0) {
2860 p->static_prio = NICE_TO_PRIO(0);
2861 p->normal_prio = p->static_prio;
2862 set_load_weight(p);
2866 * We don't need the reset flag anymore after the fork. It has
2867 * fulfilled its duty:
2869 p->sched_reset_on_fork = 0;
2873 * Make sure we do not leak PI boosting priority to the child.
2875 p->prio = current->normal_prio;
2877 if (!rt_prio(p->prio))
2878 p->sched_class = &fair_sched_class;
2880 if (p->sched_class->task_fork)
2881 p->sched_class->task_fork(p);
2884 * The child is not yet in the pid-hash so no cgroup attach races,
2885 * and the cgroup is pinned to this child due to cgroup_fork()
2886 * is ran before sched_fork().
2888 * Silence PROVE_RCU.
2890 raw_spin_lock_irqsave(&p->pi_lock, flags);
2891 set_task_cpu(p, cpu);
2892 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2894 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2895 if (likely(sched_info_on()))
2896 memset(&p->sched_info, 0, sizeof(p->sched_info));
2897 #endif
2898 #if defined(CONFIG_SMP)
2899 p->on_cpu = 0;
2900 #endif
2901 #ifdef CONFIG_PREEMPT_COUNT
2902 /* Want to start with kernel preemption disabled. */
2903 task_thread_info(p)->preempt_count = 1;
2904 #endif
2905 #ifdef CONFIG_SMP
2906 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2907 #endif
2909 put_cpu();
2913 * wake_up_new_task - wake up a newly created task for the first time.
2915 * This function will do some initial scheduler statistics housekeeping
2916 * that must be done for every newly created context, then puts the task
2917 * on the runqueue and wakes it.
2919 void wake_up_new_task(struct task_struct *p)
2921 unsigned long flags;
2922 struct rq *rq;
2924 raw_spin_lock_irqsave(&p->pi_lock, flags);
2925 #ifdef CONFIG_SMP
2927 * Fork balancing, do it here and not earlier because:
2928 * - cpus_allowed can change in the fork path
2929 * - any previously selected cpu might disappear through hotplug
2931 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
2932 #endif
2934 rq = __task_rq_lock(p);
2935 activate_task(rq, p, 0);
2936 p->on_rq = 1;
2937 trace_sched_wakeup_new(p, true);
2938 check_preempt_curr(rq, p, WF_FORK);
2939 #ifdef CONFIG_SMP
2940 if (p->sched_class->task_woken)
2941 p->sched_class->task_woken(rq, p);
2942 #endif
2943 task_rq_unlock(rq, p, &flags);
2946 #ifdef CONFIG_PREEMPT_NOTIFIERS
2949 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2950 * @notifier: notifier struct to register
2952 void preempt_notifier_register(struct preempt_notifier *notifier)
2954 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2956 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2959 * preempt_notifier_unregister - no longer interested in preemption notifications
2960 * @notifier: notifier struct to unregister
2962 * This is safe to call from within a preemption notifier.
2964 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2966 hlist_del(&notifier->link);
2968 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2970 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2972 struct preempt_notifier *notifier;
2973 struct hlist_node *node;
2975 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2976 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2979 static void
2980 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2981 struct task_struct *next)
2983 struct preempt_notifier *notifier;
2984 struct hlist_node *node;
2986 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2987 notifier->ops->sched_out(notifier, next);
2990 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2992 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2996 static void
2997 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2998 struct task_struct *next)
3002 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3005 * prepare_task_switch - prepare to switch tasks
3006 * @rq: the runqueue preparing to switch
3007 * @prev: the current task that is being switched out
3008 * @next: the task we are going to switch to.
3010 * This is called with the rq lock held and interrupts off. It must
3011 * be paired with a subsequent finish_task_switch after the context
3012 * switch.
3014 * prepare_task_switch sets up locking and calls architecture specific
3015 * hooks.
3017 static inline void
3018 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3019 struct task_struct *next)
3021 sched_info_switch(prev, next);
3022 perf_event_task_sched_out(prev, next);
3023 fire_sched_out_preempt_notifiers(prev, next);
3024 prepare_lock_switch(rq, next);
3025 prepare_arch_switch(next);
3026 trace_sched_switch(prev, next);
3030 * finish_task_switch - clean up after a task-switch
3031 * @rq: runqueue associated with task-switch
3032 * @prev: the thread we just switched away from.
3034 * finish_task_switch must be called after the context switch, paired
3035 * with a prepare_task_switch call before the context switch.
3036 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3037 * and do any other architecture-specific cleanup actions.
3039 * Note that we may have delayed dropping an mm in context_switch(). If
3040 * so, we finish that here outside of the runqueue lock. (Doing it
3041 * with the lock held can cause deadlocks; see schedule() for
3042 * details.)
3044 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3045 __releases(rq->lock)
3047 struct mm_struct *mm = rq->prev_mm;
3048 long prev_state;
3050 rq->prev_mm = NULL;
3053 * A task struct has one reference for the use as "current".
3054 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3055 * schedule one last time. The schedule call will never return, and
3056 * the scheduled task must drop that reference.
3057 * The test for TASK_DEAD must occur while the runqueue locks are
3058 * still held, otherwise prev could be scheduled on another cpu, die
3059 * there before we look at prev->state, and then the reference would
3060 * be dropped twice.
3061 * Manfred Spraul <manfred@colorfullife.com>
3063 prev_state = prev->state;
3064 finish_arch_switch(prev);
3065 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3066 local_irq_disable();
3067 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3068 perf_event_task_sched_in(current);
3069 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3070 local_irq_enable();
3071 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3072 finish_lock_switch(rq, prev);
3074 fire_sched_in_preempt_notifiers(current);
3075 if (mm)
3076 mmdrop(mm);
3077 if (unlikely(prev_state == TASK_DEAD)) {
3079 * Remove function-return probe instances associated with this
3080 * task and put them back on the free list.
3082 kprobe_flush_task(prev);
3083 put_task_struct(prev);
3087 #ifdef CONFIG_SMP
3089 /* assumes rq->lock is held */
3090 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3092 if (prev->sched_class->pre_schedule)
3093 prev->sched_class->pre_schedule(rq, prev);
3096 /* rq->lock is NOT held, but preemption is disabled */
3097 static inline void post_schedule(struct rq *rq)
3099 if (rq->post_schedule) {
3100 unsigned long flags;
3102 raw_spin_lock_irqsave(&rq->lock, flags);
3103 if (rq->curr->sched_class->post_schedule)
3104 rq->curr->sched_class->post_schedule(rq);
3105 raw_spin_unlock_irqrestore(&rq->lock, flags);
3107 rq->post_schedule = 0;
3111 #else
3113 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3117 static inline void post_schedule(struct rq *rq)
3121 #endif
3124 * schedule_tail - first thing a freshly forked thread must call.
3125 * @prev: the thread we just switched away from.
3127 asmlinkage void schedule_tail(struct task_struct *prev)
3128 __releases(rq->lock)
3130 struct rq *rq = this_rq();
3132 finish_task_switch(rq, prev);
3135 * FIXME: do we need to worry about rq being invalidated by the
3136 * task_switch?
3138 post_schedule(rq);
3140 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3141 /* In this case, finish_task_switch does not reenable preemption */
3142 preempt_enable();
3143 #endif
3144 if (current->set_child_tid)
3145 put_user(task_pid_vnr(current), current->set_child_tid);
3149 * context_switch - switch to the new MM and the new
3150 * thread's register state.
3152 static inline void
3153 context_switch(struct rq *rq, struct task_struct *prev,
3154 struct task_struct *next)
3156 struct mm_struct *mm, *oldmm;
3158 prepare_task_switch(rq, prev, next);
3160 mm = next->mm;
3161 oldmm = prev->active_mm;
3163 * For paravirt, this is coupled with an exit in switch_to to
3164 * combine the page table reload and the switch backend into
3165 * one hypercall.
3167 arch_start_context_switch(prev);
3169 if (!mm) {
3170 next->active_mm = oldmm;
3171 atomic_inc(&oldmm->mm_count);
3172 enter_lazy_tlb(oldmm, next);
3173 } else
3174 switch_mm(oldmm, mm, next);
3176 if (!prev->mm) {
3177 prev->active_mm = NULL;
3178 rq->prev_mm = oldmm;
3181 * Since the runqueue lock will be released by the next
3182 * task (which is an invalid locking op but in the case
3183 * of the scheduler it's an obvious special-case), so we
3184 * do an early lockdep release here:
3186 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3187 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3188 #endif
3190 /* Here we just switch the register state and the stack. */
3191 switch_to(prev, next, prev);
3193 barrier();
3195 * this_rq must be evaluated again because prev may have moved
3196 * CPUs since it called schedule(), thus the 'rq' on its stack
3197 * frame will be invalid.
3199 finish_task_switch(this_rq(), prev);
3203 * nr_running, nr_uninterruptible and nr_context_switches:
3205 * externally visible scheduler statistics: current number of runnable
3206 * threads, current number of uninterruptible-sleeping threads, total
3207 * number of context switches performed since bootup.
3209 unsigned long nr_running(void)
3211 unsigned long i, sum = 0;
3213 for_each_online_cpu(i)
3214 sum += cpu_rq(i)->nr_running;
3216 return sum;
3219 unsigned long nr_uninterruptible(void)
3221 unsigned long i, sum = 0;
3223 for_each_possible_cpu(i)
3224 sum += cpu_rq(i)->nr_uninterruptible;
3227 * Since we read the counters lockless, it might be slightly
3228 * inaccurate. Do not allow it to go below zero though:
3230 if (unlikely((long)sum < 0))
3231 sum = 0;
3233 return sum;
3236 unsigned long long nr_context_switches(void)
3238 int i;
3239 unsigned long long sum = 0;
3241 for_each_possible_cpu(i)
3242 sum += cpu_rq(i)->nr_switches;
3244 return sum;
3247 unsigned long nr_iowait(void)
3249 unsigned long i, sum = 0;
3251 for_each_possible_cpu(i)
3252 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3254 return sum;
3257 unsigned long nr_iowait_cpu(int cpu)
3259 struct rq *this = cpu_rq(cpu);
3260 return atomic_read(&this->nr_iowait);
3263 unsigned long this_cpu_load(void)
3265 struct rq *this = this_rq();
3266 return this->cpu_load[0];
3270 /* Variables and functions for calc_load */
3271 static atomic_long_t calc_load_tasks;
3272 static unsigned long calc_load_update;
3273 unsigned long avenrun[3];
3274 EXPORT_SYMBOL(avenrun);
3276 static long calc_load_fold_active(struct rq *this_rq)
3278 long nr_active, delta = 0;
3280 nr_active = this_rq->nr_running;
3281 nr_active += (long) this_rq->nr_uninterruptible;
3283 if (nr_active != this_rq->calc_load_active) {
3284 delta = nr_active - this_rq->calc_load_active;
3285 this_rq->calc_load_active = nr_active;
3288 return delta;
3291 static unsigned long
3292 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3294 load *= exp;
3295 load += active * (FIXED_1 - exp);
3296 load += 1UL << (FSHIFT - 1);
3297 return load >> FSHIFT;
3300 #ifdef CONFIG_NO_HZ
3302 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3304 * When making the ILB scale, we should try to pull this in as well.
3306 static atomic_long_t calc_load_tasks_idle;
3308 static void calc_load_account_idle(struct rq *this_rq)
3310 long delta;
3312 delta = calc_load_fold_active(this_rq);
3313 if (delta)
3314 atomic_long_add(delta, &calc_load_tasks_idle);
3317 static long calc_load_fold_idle(void)
3319 long delta = 0;
3322 * Its got a race, we don't care...
3324 if (atomic_long_read(&calc_load_tasks_idle))
3325 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3327 return delta;
3331 * fixed_power_int - compute: x^n, in O(log n) time
3333 * @x: base of the power
3334 * @frac_bits: fractional bits of @x
3335 * @n: power to raise @x to.
3337 * By exploiting the relation between the definition of the natural power
3338 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3339 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3340 * (where: n_i \elem {0, 1}, the binary vector representing n),
3341 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3342 * of course trivially computable in O(log_2 n), the length of our binary
3343 * vector.
3345 static unsigned long
3346 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3348 unsigned long result = 1UL << frac_bits;
3350 if (n) for (;;) {
3351 if (n & 1) {
3352 result *= x;
3353 result += 1UL << (frac_bits - 1);
3354 result >>= frac_bits;
3356 n >>= 1;
3357 if (!n)
3358 break;
3359 x *= x;
3360 x += 1UL << (frac_bits - 1);
3361 x >>= frac_bits;
3364 return result;
3368 * a1 = a0 * e + a * (1 - e)
3370 * a2 = a1 * e + a * (1 - e)
3371 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3372 * = a0 * e^2 + a * (1 - e) * (1 + e)
3374 * a3 = a2 * e + a * (1 - e)
3375 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3376 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3378 * ...
3380 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3381 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3382 * = a0 * e^n + a * (1 - e^n)
3384 * [1] application of the geometric series:
3386 * n 1 - x^(n+1)
3387 * S_n := \Sum x^i = -------------
3388 * i=0 1 - x
3390 static unsigned long
3391 calc_load_n(unsigned long load, unsigned long exp,
3392 unsigned long active, unsigned int n)
3395 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3399 * NO_HZ can leave us missing all per-cpu ticks calling
3400 * calc_load_account_active(), but since an idle CPU folds its delta into
3401 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3402 * in the pending idle delta if our idle period crossed a load cycle boundary.
3404 * Once we've updated the global active value, we need to apply the exponential
3405 * weights adjusted to the number of cycles missed.
3407 static void calc_global_nohz(unsigned long ticks)
3409 long delta, active, n;
3411 if (time_before(jiffies, calc_load_update))
3412 return;
3415 * If we crossed a calc_load_update boundary, make sure to fold
3416 * any pending idle changes, the respective CPUs might have
3417 * missed the tick driven calc_load_account_active() update
3418 * due to NO_HZ.
3420 delta = calc_load_fold_idle();
3421 if (delta)
3422 atomic_long_add(delta, &calc_load_tasks);
3425 * If we were idle for multiple load cycles, apply them.
3427 if (ticks >= LOAD_FREQ) {
3428 n = ticks / LOAD_FREQ;
3430 active = atomic_long_read(&calc_load_tasks);
3431 active = active > 0 ? active * FIXED_1 : 0;
3433 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3434 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3435 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3437 calc_load_update += n * LOAD_FREQ;
3441 * Its possible the remainder of the above division also crosses
3442 * a LOAD_FREQ period, the regular check in calc_global_load()
3443 * which comes after this will take care of that.
3445 * Consider us being 11 ticks before a cycle completion, and us
3446 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3447 * age us 4 cycles, and the test in calc_global_load() will
3448 * pick up the final one.
3451 #else
3452 static void calc_load_account_idle(struct rq *this_rq)
3456 static inline long calc_load_fold_idle(void)
3458 return 0;
3461 static void calc_global_nohz(unsigned long ticks)
3464 #endif
3467 * get_avenrun - get the load average array
3468 * @loads: pointer to dest load array
3469 * @offset: offset to add
3470 * @shift: shift count to shift the result left
3472 * These values are estimates at best, so no need for locking.
3474 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3476 loads[0] = (avenrun[0] + offset) << shift;
3477 loads[1] = (avenrun[1] + offset) << shift;
3478 loads[2] = (avenrun[2] + offset) << shift;
3482 * calc_load - update the avenrun load estimates 10 ticks after the
3483 * CPUs have updated calc_load_tasks.
3485 void calc_global_load(unsigned long ticks)
3487 long active;
3489 calc_global_nohz(ticks);
3491 if (time_before(jiffies, calc_load_update + 10))
3492 return;
3494 active = atomic_long_read(&calc_load_tasks);
3495 active = active > 0 ? active * FIXED_1 : 0;
3497 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3498 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3499 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3501 calc_load_update += LOAD_FREQ;
3505 * Called from update_cpu_load() to periodically update this CPU's
3506 * active count.
3508 static void calc_load_account_active(struct rq *this_rq)
3510 long delta;
3512 if (time_before(jiffies, this_rq->calc_load_update))
3513 return;
3515 delta = calc_load_fold_active(this_rq);
3516 delta += calc_load_fold_idle();
3517 if (delta)
3518 atomic_long_add(delta, &calc_load_tasks);
3520 this_rq->calc_load_update += LOAD_FREQ;
3524 * The exact cpuload at various idx values, calculated at every tick would be
3525 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3527 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3528 * on nth tick when cpu may be busy, then we have:
3529 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3530 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3532 * decay_load_missed() below does efficient calculation of
3533 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3534 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3536 * The calculation is approximated on a 128 point scale.
3537 * degrade_zero_ticks is the number of ticks after which load at any
3538 * particular idx is approximated to be zero.
3539 * degrade_factor is a precomputed table, a row for each load idx.
3540 * Each column corresponds to degradation factor for a power of two ticks,
3541 * based on 128 point scale.
3542 * Example:
3543 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3544 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3546 * With this power of 2 load factors, we can degrade the load n times
3547 * by looking at 1 bits in n and doing as many mult/shift instead of
3548 * n mult/shifts needed by the exact degradation.
3550 #define DEGRADE_SHIFT 7
3551 static const unsigned char
3552 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3553 static const unsigned char
3554 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3555 {0, 0, 0, 0, 0, 0, 0, 0},
3556 {64, 32, 8, 0, 0, 0, 0, 0},
3557 {96, 72, 40, 12, 1, 0, 0},
3558 {112, 98, 75, 43, 15, 1, 0},
3559 {120, 112, 98, 76, 45, 16, 2} };
3562 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3563 * would be when CPU is idle and so we just decay the old load without
3564 * adding any new load.
3566 static unsigned long
3567 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3569 int j = 0;
3571 if (!missed_updates)
3572 return load;
3574 if (missed_updates >= degrade_zero_ticks[idx])
3575 return 0;
3577 if (idx == 1)
3578 return load >> missed_updates;
3580 while (missed_updates) {
3581 if (missed_updates % 2)
3582 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3584 missed_updates >>= 1;
3585 j++;
3587 return load;
3591 * Update rq->cpu_load[] statistics. This function is usually called every
3592 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3593 * every tick. We fix it up based on jiffies.
3595 static void update_cpu_load(struct rq *this_rq)
3597 unsigned long this_load = this_rq->load.weight;
3598 unsigned long curr_jiffies = jiffies;
3599 unsigned long pending_updates;
3600 int i, scale;
3602 this_rq->nr_load_updates++;
3604 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3605 if (curr_jiffies == this_rq->last_load_update_tick)
3606 return;
3608 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3609 this_rq->last_load_update_tick = curr_jiffies;
3611 /* Update our load: */
3612 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3613 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3614 unsigned long old_load, new_load;
3616 /* scale is effectively 1 << i now, and >> i divides by scale */
3618 old_load = this_rq->cpu_load[i];
3619 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3620 new_load = this_load;
3622 * Round up the averaging division if load is increasing. This
3623 * prevents us from getting stuck on 9 if the load is 10, for
3624 * example.
3626 if (new_load > old_load)
3627 new_load += scale - 1;
3629 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3632 sched_avg_update(this_rq);
3635 static void update_cpu_load_active(struct rq *this_rq)
3637 update_cpu_load(this_rq);
3639 calc_load_account_active(this_rq);
3642 #ifdef CONFIG_SMP
3645 * sched_exec - execve() is a valuable balancing opportunity, because at
3646 * this point the task has the smallest effective memory and cache footprint.
3648 void sched_exec(void)
3650 struct task_struct *p = current;
3651 unsigned long flags;
3652 int dest_cpu;
3654 raw_spin_lock_irqsave(&p->pi_lock, flags);
3655 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3656 if (dest_cpu == smp_processor_id())
3657 goto unlock;
3659 if (likely(cpu_active(dest_cpu))) {
3660 struct migration_arg arg = { p, dest_cpu };
3662 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3663 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3664 return;
3666 unlock:
3667 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3670 #endif
3672 DEFINE_PER_CPU(struct kernel_stat, kstat);
3674 EXPORT_PER_CPU_SYMBOL(kstat);
3677 * Return any ns on the sched_clock that have not yet been accounted in
3678 * @p in case that task is currently running.
3680 * Called with task_rq_lock() held on @rq.
3682 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3684 u64 ns = 0;
3686 if (task_current(rq, p)) {
3687 update_rq_clock(rq);
3688 ns = rq->clock_task - p->se.exec_start;
3689 if ((s64)ns < 0)
3690 ns = 0;
3693 return ns;
3696 unsigned long long task_delta_exec(struct task_struct *p)
3698 unsigned long flags;
3699 struct rq *rq;
3700 u64 ns = 0;
3702 rq = task_rq_lock(p, &flags);
3703 ns = do_task_delta_exec(p, rq);
3704 task_rq_unlock(rq, p, &flags);
3706 return ns;
3710 * Return accounted runtime for the task.
3711 * In case the task is currently running, return the runtime plus current's
3712 * pending runtime that have not been accounted yet.
3714 unsigned long long task_sched_runtime(struct task_struct *p)
3716 unsigned long flags;
3717 struct rq *rq;
3718 u64 ns = 0;
3720 rq = task_rq_lock(p, &flags);
3721 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3722 task_rq_unlock(rq, p, &flags);
3724 return ns;
3728 * Return sum_exec_runtime for the thread group.
3729 * In case the task is currently running, return the sum plus current's
3730 * pending runtime that have not been accounted yet.
3732 * Note that the thread group might have other running tasks as well,
3733 * so the return value not includes other pending runtime that other
3734 * running tasks might have.
3736 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3738 struct task_cputime totals;
3739 unsigned long flags;
3740 struct rq *rq;
3741 u64 ns;
3743 rq = task_rq_lock(p, &flags);
3744 thread_group_cputime(p, &totals);
3745 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3746 task_rq_unlock(rq, p, &flags);
3748 return ns;
3752 * Account user cpu time to a process.
3753 * @p: the process that the cpu time gets accounted to
3754 * @cputime: the cpu time spent in user space since the last update
3755 * @cputime_scaled: cputime scaled by cpu frequency
3757 void account_user_time(struct task_struct *p, cputime_t cputime,
3758 cputime_t cputime_scaled)
3760 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3761 cputime64_t tmp;
3763 /* Add user time to process. */
3764 p->utime = cputime_add(p->utime, cputime);
3765 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3766 account_group_user_time(p, cputime);
3768 /* Add user time to cpustat. */
3769 tmp = cputime_to_cputime64(cputime);
3770 if (TASK_NICE(p) > 0)
3771 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3772 else
3773 cpustat->user = cputime64_add(cpustat->user, tmp);
3775 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3776 /* Account for user time used */
3777 acct_update_integrals(p);
3781 * Account guest cpu time to a process.
3782 * @p: the process that the cpu time gets accounted to
3783 * @cputime: the cpu time spent in virtual machine since the last update
3784 * @cputime_scaled: cputime scaled by cpu frequency
3786 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3787 cputime_t cputime_scaled)
3789 cputime64_t tmp;
3790 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3792 tmp = cputime_to_cputime64(cputime);
3794 /* Add guest time to process. */
3795 p->utime = cputime_add(p->utime, cputime);
3796 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3797 account_group_user_time(p, cputime);
3798 p->gtime = cputime_add(p->gtime, cputime);
3800 /* Add guest time to cpustat. */
3801 if (TASK_NICE(p) > 0) {
3802 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3803 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3804 } else {
3805 cpustat->user = cputime64_add(cpustat->user, tmp);
3806 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3811 * Account system cpu time to a process and desired cpustat field
3812 * @p: the process that the cpu time gets accounted to
3813 * @cputime: the cpu time spent in kernel space since the last update
3814 * @cputime_scaled: cputime scaled by cpu frequency
3815 * @target_cputime64: pointer to cpustat field that has to be updated
3817 static inline
3818 void __account_system_time(struct task_struct *p, cputime_t cputime,
3819 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3821 cputime64_t tmp = cputime_to_cputime64(cputime);
3823 /* Add system time to process. */
3824 p->stime = cputime_add(p->stime, cputime);
3825 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3826 account_group_system_time(p, cputime);
3828 /* Add system time to cpustat. */
3829 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3830 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3832 /* Account for system time used */
3833 acct_update_integrals(p);
3837 * Account system cpu time to a process.
3838 * @p: the process that the cpu time gets accounted to
3839 * @hardirq_offset: the offset to subtract from hardirq_count()
3840 * @cputime: the cpu time spent in kernel space since the last update
3841 * @cputime_scaled: cputime scaled by cpu frequency
3843 void account_system_time(struct task_struct *p, int hardirq_offset,
3844 cputime_t cputime, cputime_t cputime_scaled)
3846 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3847 cputime64_t *target_cputime64;
3849 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3850 account_guest_time(p, cputime, cputime_scaled);
3851 return;
3854 if (hardirq_count() - hardirq_offset)
3855 target_cputime64 = &cpustat->irq;
3856 else if (in_serving_softirq())
3857 target_cputime64 = &cpustat->softirq;
3858 else
3859 target_cputime64 = &cpustat->system;
3861 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3865 * Account for involuntary wait time.
3866 * @cputime: the cpu time spent in involuntary wait
3868 void account_steal_time(cputime_t cputime)
3870 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3871 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3873 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3877 * Account for idle time.
3878 * @cputime: the cpu time spent in idle wait
3880 void account_idle_time(cputime_t cputime)
3882 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3883 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3884 struct rq *rq = this_rq();
3886 if (atomic_read(&rq->nr_iowait) > 0)
3887 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3888 else
3889 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3892 static __always_inline bool steal_account_process_tick(void)
3894 #ifdef CONFIG_PARAVIRT
3895 if (static_branch(&paravirt_steal_enabled)) {
3896 u64 steal, st = 0;
3898 steal = paravirt_steal_clock(smp_processor_id());
3899 steal -= this_rq()->prev_steal_time;
3901 st = steal_ticks(steal);
3902 this_rq()->prev_steal_time += st * TICK_NSEC;
3904 account_steal_time(st);
3905 return st;
3907 #endif
3908 return false;
3911 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3913 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3915 * Account a tick to a process and cpustat
3916 * @p: the process that the cpu time gets accounted to
3917 * @user_tick: is the tick from userspace
3918 * @rq: the pointer to rq
3920 * Tick demultiplexing follows the order
3921 * - pending hardirq update
3922 * - pending softirq update
3923 * - user_time
3924 * - idle_time
3925 * - system time
3926 * - check for guest_time
3927 * - else account as system_time
3929 * Check for hardirq is done both for system and user time as there is
3930 * no timer going off while we are on hardirq and hence we may never get an
3931 * opportunity to update it solely in system time.
3932 * p->stime and friends are only updated on system time and not on irq
3933 * softirq as those do not count in task exec_runtime any more.
3935 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3936 struct rq *rq)
3938 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3939 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3940 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3942 if (steal_account_process_tick())
3943 return;
3945 if (irqtime_account_hi_update()) {
3946 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3947 } else if (irqtime_account_si_update()) {
3948 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3949 } else if (this_cpu_ksoftirqd() == p) {
3951 * ksoftirqd time do not get accounted in cpu_softirq_time.
3952 * So, we have to handle it separately here.
3953 * Also, p->stime needs to be updated for ksoftirqd.
3955 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3956 &cpustat->softirq);
3957 } else if (user_tick) {
3958 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3959 } else if (p == rq->idle) {
3960 account_idle_time(cputime_one_jiffy);
3961 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3962 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3963 } else {
3964 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3965 &cpustat->system);
3969 static void irqtime_account_idle_ticks(int ticks)
3971 int i;
3972 struct rq *rq = this_rq();
3974 for (i = 0; i < ticks; i++)
3975 irqtime_account_process_tick(current, 0, rq);
3977 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3978 static void irqtime_account_idle_ticks(int ticks) {}
3979 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3980 struct rq *rq) {}
3981 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3984 * Account a single tick of cpu time.
3985 * @p: the process that the cpu time gets accounted to
3986 * @user_tick: indicates if the tick is a user or a system tick
3988 void account_process_tick(struct task_struct *p, int user_tick)
3990 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3991 struct rq *rq = this_rq();
3993 if (sched_clock_irqtime) {
3994 irqtime_account_process_tick(p, user_tick, rq);
3995 return;
3998 if (steal_account_process_tick())
3999 return;
4001 if (user_tick)
4002 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4003 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
4004 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
4005 one_jiffy_scaled);
4006 else
4007 account_idle_time(cputime_one_jiffy);
4011 * Account multiple ticks of steal time.
4012 * @p: the process from which the cpu time has been stolen
4013 * @ticks: number of stolen ticks
4015 void account_steal_ticks(unsigned long ticks)
4017 account_steal_time(jiffies_to_cputime(ticks));
4021 * Account multiple ticks of idle time.
4022 * @ticks: number of stolen ticks
4024 void account_idle_ticks(unsigned long ticks)
4027 if (sched_clock_irqtime) {
4028 irqtime_account_idle_ticks(ticks);
4029 return;
4032 account_idle_time(jiffies_to_cputime(ticks));
4035 #endif
4038 * Use precise platform statistics if available:
4040 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4041 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4043 *ut = p->utime;
4044 *st = p->stime;
4047 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4049 struct task_cputime cputime;
4051 thread_group_cputime(p, &cputime);
4053 *ut = cputime.utime;
4054 *st = cputime.stime;
4056 #else
4058 #ifndef nsecs_to_cputime
4059 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4060 #endif
4062 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4064 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4067 * Use CFS's precise accounting:
4069 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4071 if (total) {
4072 u64 temp = rtime;
4074 temp *= utime;
4075 do_div(temp, total);
4076 utime = (cputime_t)temp;
4077 } else
4078 utime = rtime;
4081 * Compare with previous values, to keep monotonicity:
4083 p->prev_utime = max(p->prev_utime, utime);
4084 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4086 *ut = p->prev_utime;
4087 *st = p->prev_stime;
4091 * Must be called with siglock held.
4093 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4095 struct signal_struct *sig = p->signal;
4096 struct task_cputime cputime;
4097 cputime_t rtime, utime, total;
4099 thread_group_cputime(p, &cputime);
4101 total = cputime_add(cputime.utime, cputime.stime);
4102 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4104 if (total) {
4105 u64 temp = rtime;
4107 temp *= cputime.utime;
4108 do_div(temp, total);
4109 utime = (cputime_t)temp;
4110 } else
4111 utime = rtime;
4113 sig->prev_utime = max(sig->prev_utime, utime);
4114 sig->prev_stime = max(sig->prev_stime,
4115 cputime_sub(rtime, sig->prev_utime));
4117 *ut = sig->prev_utime;
4118 *st = sig->prev_stime;
4120 #endif
4123 * This function gets called by the timer code, with HZ frequency.
4124 * We call it with interrupts disabled.
4126 void scheduler_tick(void)
4128 int cpu = smp_processor_id();
4129 struct rq *rq = cpu_rq(cpu);
4130 struct task_struct *curr = rq->curr;
4132 sched_clock_tick();
4134 raw_spin_lock(&rq->lock);
4135 update_rq_clock(rq);
4136 update_cpu_load_active(rq);
4137 curr->sched_class->task_tick(rq, curr, 0);
4138 raw_spin_unlock(&rq->lock);
4140 perf_event_task_tick();
4142 #ifdef CONFIG_SMP
4143 rq->idle_at_tick = idle_cpu(cpu);
4144 trigger_load_balance(rq, cpu);
4145 #endif
4148 notrace unsigned long get_parent_ip(unsigned long addr)
4150 if (in_lock_functions(addr)) {
4151 addr = CALLER_ADDR2;
4152 if (in_lock_functions(addr))
4153 addr = CALLER_ADDR3;
4155 return addr;
4158 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4159 defined(CONFIG_PREEMPT_TRACER))
4161 void __kprobes add_preempt_count(int val)
4163 #ifdef CONFIG_DEBUG_PREEMPT
4165 * Underflow?
4167 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4168 return;
4169 #endif
4170 preempt_count() += val;
4171 #ifdef CONFIG_DEBUG_PREEMPT
4173 * Spinlock count overflowing soon?
4175 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4176 PREEMPT_MASK - 10);
4177 #endif
4178 if (preempt_count() == val)
4179 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4181 EXPORT_SYMBOL(add_preempt_count);
4183 void __kprobes sub_preempt_count(int val)
4185 #ifdef CONFIG_DEBUG_PREEMPT
4187 * Underflow?
4189 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4190 return;
4192 * Is the spinlock portion underflowing?
4194 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4195 !(preempt_count() & PREEMPT_MASK)))
4196 return;
4197 #endif
4199 if (preempt_count() == val)
4200 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4201 preempt_count() -= val;
4203 EXPORT_SYMBOL(sub_preempt_count);
4205 #endif
4208 * Print scheduling while atomic bug:
4210 static noinline void __schedule_bug(struct task_struct *prev)
4212 struct pt_regs *regs = get_irq_regs();
4214 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4215 prev->comm, prev->pid, preempt_count());
4217 debug_show_held_locks(prev);
4218 print_modules();
4219 if (irqs_disabled())
4220 print_irqtrace_events(prev);
4222 if (regs)
4223 show_regs(regs);
4224 else
4225 dump_stack();
4229 * Various schedule()-time debugging checks and statistics:
4231 static inline void schedule_debug(struct task_struct *prev)
4234 * Test if we are atomic. Since do_exit() needs to call into
4235 * schedule() atomically, we ignore that path for now.
4236 * Otherwise, whine if we are scheduling when we should not be.
4238 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4239 __schedule_bug(prev);
4241 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4243 schedstat_inc(this_rq(), sched_count);
4246 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4248 if (prev->on_rq || rq->skip_clock_update < 0)
4249 update_rq_clock(rq);
4250 prev->sched_class->put_prev_task(rq, prev);
4254 * Pick up the highest-prio task:
4256 static inline struct task_struct *
4257 pick_next_task(struct rq *rq)
4259 const struct sched_class *class;
4260 struct task_struct *p;
4263 * Optimization: we know that if all tasks are in
4264 * the fair class we can call that function directly:
4266 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4267 p = fair_sched_class.pick_next_task(rq);
4268 if (likely(p))
4269 return p;
4272 for_each_class(class) {
4273 p = class->pick_next_task(rq);
4274 if (p)
4275 return p;
4278 BUG(); /* the idle class will always have a runnable task */
4282 * schedule() is the main scheduler function.
4284 asmlinkage void __sched schedule(void)
4286 struct task_struct *prev, *next;
4287 unsigned long *switch_count;
4288 struct rq *rq;
4289 int cpu;
4291 need_resched:
4292 preempt_disable();
4293 cpu = smp_processor_id();
4294 rq = cpu_rq(cpu);
4295 rcu_note_context_switch(cpu);
4296 prev = rq->curr;
4298 schedule_debug(prev);
4300 if (sched_feat(HRTICK))
4301 hrtick_clear(rq);
4303 raw_spin_lock_irq(&rq->lock);
4305 switch_count = &prev->nivcsw;
4306 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4307 if (unlikely(signal_pending_state(prev->state, prev))) {
4308 prev->state = TASK_RUNNING;
4309 } else {
4310 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4311 prev->on_rq = 0;
4314 * If a worker went to sleep, notify and ask workqueue
4315 * whether it wants to wake up a task to maintain
4316 * concurrency.
4318 if (prev->flags & PF_WQ_WORKER) {
4319 struct task_struct *to_wakeup;
4321 to_wakeup = wq_worker_sleeping(prev, cpu);
4322 if (to_wakeup)
4323 try_to_wake_up_local(to_wakeup);
4327 * If we are going to sleep and we have plugged IO
4328 * queued, make sure to submit it to avoid deadlocks.
4330 if (blk_needs_flush_plug(prev)) {
4331 raw_spin_unlock(&rq->lock);
4332 blk_schedule_flush_plug(prev);
4333 raw_spin_lock(&rq->lock);
4336 switch_count = &prev->nvcsw;
4339 pre_schedule(rq, prev);
4341 if (unlikely(!rq->nr_running))
4342 idle_balance(cpu, rq);
4344 put_prev_task(rq, prev);
4345 next = pick_next_task(rq);
4346 clear_tsk_need_resched(prev);
4347 rq->skip_clock_update = 0;
4349 if (likely(prev != next)) {
4350 rq->nr_switches++;
4351 rq->curr = next;
4352 ++*switch_count;
4354 context_switch(rq, prev, next); /* unlocks the rq */
4356 * The context switch have flipped the stack from under us
4357 * and restored the local variables which were saved when
4358 * this task called schedule() in the past. prev == current
4359 * is still correct, but it can be moved to another cpu/rq.
4361 cpu = smp_processor_id();
4362 rq = cpu_rq(cpu);
4363 } else
4364 raw_spin_unlock_irq(&rq->lock);
4366 post_schedule(rq);
4368 preempt_enable_no_resched();
4369 if (need_resched())
4370 goto need_resched;
4372 EXPORT_SYMBOL(schedule);
4374 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4376 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4378 if (lock->owner != owner)
4379 return false;
4382 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4383 * lock->owner still matches owner, if that fails, owner might
4384 * point to free()d memory, if it still matches, the rcu_read_lock()
4385 * ensures the memory stays valid.
4387 barrier();
4389 return owner->on_cpu;
4393 * Look out! "owner" is an entirely speculative pointer
4394 * access and not reliable.
4396 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4398 if (!sched_feat(OWNER_SPIN))
4399 return 0;
4401 rcu_read_lock();
4402 while (owner_running(lock, owner)) {
4403 if (need_resched())
4404 break;
4406 arch_mutex_cpu_relax();
4408 rcu_read_unlock();
4411 * We break out the loop above on need_resched() and when the
4412 * owner changed, which is a sign for heavy contention. Return
4413 * success only when lock->owner is NULL.
4415 return lock->owner == NULL;
4417 #endif
4419 #ifdef CONFIG_PREEMPT
4421 * this is the entry point to schedule() from in-kernel preemption
4422 * off of preempt_enable. Kernel preemptions off return from interrupt
4423 * occur there and call schedule directly.
4425 asmlinkage void __sched notrace preempt_schedule(void)
4427 struct thread_info *ti = current_thread_info();
4430 * If there is a non-zero preempt_count or interrupts are disabled,
4431 * we do not want to preempt the current task. Just return..
4433 if (likely(ti->preempt_count || irqs_disabled()))
4434 return;
4436 do {
4437 add_preempt_count_notrace(PREEMPT_ACTIVE);
4438 schedule();
4439 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4442 * Check again in case we missed a preemption opportunity
4443 * between schedule and now.
4445 barrier();
4446 } while (need_resched());
4448 EXPORT_SYMBOL(preempt_schedule);
4451 * this is the entry point to schedule() from kernel preemption
4452 * off of irq context.
4453 * Note, that this is called and return with irqs disabled. This will
4454 * protect us against recursive calling from irq.
4456 asmlinkage void __sched preempt_schedule_irq(void)
4458 struct thread_info *ti = current_thread_info();
4460 /* Catch callers which need to be fixed */
4461 BUG_ON(ti->preempt_count || !irqs_disabled());
4463 do {
4464 add_preempt_count(PREEMPT_ACTIVE);
4465 local_irq_enable();
4466 schedule();
4467 local_irq_disable();
4468 sub_preempt_count(PREEMPT_ACTIVE);
4471 * Check again in case we missed a preemption opportunity
4472 * between schedule and now.
4474 barrier();
4475 } while (need_resched());
4478 #endif /* CONFIG_PREEMPT */
4480 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4481 void *key)
4483 return try_to_wake_up(curr->private, mode, wake_flags);
4485 EXPORT_SYMBOL(default_wake_function);
4488 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4489 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4490 * number) then we wake all the non-exclusive tasks and one exclusive task.
4492 * There are circumstances in which we can try to wake a task which has already
4493 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4494 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4496 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4497 int nr_exclusive, int wake_flags, void *key)
4499 wait_queue_t *curr, *next;
4501 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4502 unsigned flags = curr->flags;
4504 if (curr->func(curr, mode, wake_flags, key) &&
4505 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4506 break;
4511 * __wake_up - wake up threads blocked on a waitqueue.
4512 * @q: the waitqueue
4513 * @mode: which threads
4514 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4515 * @key: is directly passed to the wakeup function
4517 * It may be assumed that this function implies a write memory barrier before
4518 * changing the task state if and only if any tasks are woken up.
4520 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4521 int nr_exclusive, void *key)
4523 unsigned long flags;
4525 spin_lock_irqsave(&q->lock, flags);
4526 __wake_up_common(q, mode, nr_exclusive, 0, key);
4527 spin_unlock_irqrestore(&q->lock, flags);
4529 EXPORT_SYMBOL(__wake_up);
4532 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4534 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4536 __wake_up_common(q, mode, 1, 0, NULL);
4538 EXPORT_SYMBOL_GPL(__wake_up_locked);
4540 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4542 __wake_up_common(q, mode, 1, 0, key);
4544 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4547 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4548 * @q: the waitqueue
4549 * @mode: which threads
4550 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4551 * @key: opaque value to be passed to wakeup targets
4553 * The sync wakeup differs that the waker knows that it will schedule
4554 * away soon, so while the target thread will be woken up, it will not
4555 * be migrated to another CPU - ie. the two threads are 'synchronized'
4556 * with each other. This can prevent needless bouncing between CPUs.
4558 * On UP it can prevent extra preemption.
4560 * It may be assumed that this function implies a write memory barrier before
4561 * changing the task state if and only if any tasks are woken up.
4563 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4564 int nr_exclusive, void *key)
4566 unsigned long flags;
4567 int wake_flags = WF_SYNC;
4569 if (unlikely(!q))
4570 return;
4572 if (unlikely(!nr_exclusive))
4573 wake_flags = 0;
4575 spin_lock_irqsave(&q->lock, flags);
4576 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4577 spin_unlock_irqrestore(&q->lock, flags);
4579 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4582 * __wake_up_sync - see __wake_up_sync_key()
4584 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4586 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4588 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4591 * complete: - signals a single thread waiting on this completion
4592 * @x: holds the state of this particular completion
4594 * This will wake up a single thread waiting on this completion. Threads will be
4595 * awakened in the same order in which they were queued.
4597 * See also complete_all(), wait_for_completion() and related routines.
4599 * It may be assumed that this function implies a write memory barrier before
4600 * changing the task state if and only if any tasks are woken up.
4602 void complete(struct completion *x)
4604 unsigned long flags;
4606 spin_lock_irqsave(&x->wait.lock, flags);
4607 x->done++;
4608 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4609 spin_unlock_irqrestore(&x->wait.lock, flags);
4611 EXPORT_SYMBOL(complete);
4614 * complete_all: - signals all threads waiting on this completion
4615 * @x: holds the state of this particular completion
4617 * This will wake up all threads waiting on this particular completion event.
4619 * It may be assumed that this function implies a write memory barrier before
4620 * changing the task state if and only if any tasks are woken up.
4622 void complete_all(struct completion *x)
4624 unsigned long flags;
4626 spin_lock_irqsave(&x->wait.lock, flags);
4627 x->done += UINT_MAX/2;
4628 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4629 spin_unlock_irqrestore(&x->wait.lock, flags);
4631 EXPORT_SYMBOL(complete_all);
4633 static inline long __sched
4634 do_wait_for_common(struct completion *x, long timeout, int state)
4636 if (!x->done) {
4637 DECLARE_WAITQUEUE(wait, current);
4639 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4640 do {
4641 if (signal_pending_state(state, current)) {
4642 timeout = -ERESTARTSYS;
4643 break;
4645 __set_current_state(state);
4646 spin_unlock_irq(&x->wait.lock);
4647 timeout = schedule_timeout(timeout);
4648 spin_lock_irq(&x->wait.lock);
4649 } while (!x->done && timeout);
4650 __remove_wait_queue(&x->wait, &wait);
4651 if (!x->done)
4652 return timeout;
4654 x->done--;
4655 return timeout ?: 1;
4658 static long __sched
4659 wait_for_common(struct completion *x, long timeout, int state)
4661 might_sleep();
4663 spin_lock_irq(&x->wait.lock);
4664 timeout = do_wait_for_common(x, timeout, state);
4665 spin_unlock_irq(&x->wait.lock);
4666 return timeout;
4670 * wait_for_completion: - waits for completion of a task
4671 * @x: holds the state of this particular completion
4673 * This waits to be signaled for completion of a specific task. It is NOT
4674 * interruptible and there is no timeout.
4676 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4677 * and interrupt capability. Also see complete().
4679 void __sched wait_for_completion(struct completion *x)
4681 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4683 EXPORT_SYMBOL(wait_for_completion);
4686 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4687 * @x: holds the state of this particular completion
4688 * @timeout: timeout value in jiffies
4690 * This waits for either a completion of a specific task to be signaled or for a
4691 * specified timeout to expire. The timeout is in jiffies. It is not
4692 * interruptible.
4694 unsigned long __sched
4695 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4697 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4699 EXPORT_SYMBOL(wait_for_completion_timeout);
4702 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4703 * @x: holds the state of this particular completion
4705 * This waits for completion of a specific task to be signaled. It is
4706 * interruptible.
4708 int __sched wait_for_completion_interruptible(struct completion *x)
4710 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4711 if (t == -ERESTARTSYS)
4712 return t;
4713 return 0;
4715 EXPORT_SYMBOL(wait_for_completion_interruptible);
4718 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4719 * @x: holds the state of this particular completion
4720 * @timeout: timeout value in jiffies
4722 * This waits for either a completion of a specific task to be signaled or for a
4723 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4725 long __sched
4726 wait_for_completion_interruptible_timeout(struct completion *x,
4727 unsigned long timeout)
4729 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4731 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4734 * wait_for_completion_killable: - waits for completion of a task (killable)
4735 * @x: holds the state of this particular completion
4737 * This waits to be signaled for completion of a specific task. It can be
4738 * interrupted by a kill signal.
4740 int __sched wait_for_completion_killable(struct completion *x)
4742 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4743 if (t == -ERESTARTSYS)
4744 return t;
4745 return 0;
4747 EXPORT_SYMBOL(wait_for_completion_killable);
4750 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4751 * @x: holds the state of this particular completion
4752 * @timeout: timeout value in jiffies
4754 * This waits for either a completion of a specific task to be
4755 * signaled or for a specified timeout to expire. It can be
4756 * interrupted by a kill signal. The timeout is in jiffies.
4758 long __sched
4759 wait_for_completion_killable_timeout(struct completion *x,
4760 unsigned long timeout)
4762 return wait_for_common(x, timeout, TASK_KILLABLE);
4764 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4767 * try_wait_for_completion - try to decrement a completion without blocking
4768 * @x: completion structure
4770 * Returns: 0 if a decrement cannot be done without blocking
4771 * 1 if a decrement succeeded.
4773 * If a completion is being used as a counting completion,
4774 * attempt to decrement the counter without blocking. This
4775 * enables us to avoid waiting if the resource the completion
4776 * is protecting is not available.
4778 bool try_wait_for_completion(struct completion *x)
4780 unsigned long flags;
4781 int ret = 1;
4783 spin_lock_irqsave(&x->wait.lock, flags);
4784 if (!x->done)
4785 ret = 0;
4786 else
4787 x->done--;
4788 spin_unlock_irqrestore(&x->wait.lock, flags);
4789 return ret;
4791 EXPORT_SYMBOL(try_wait_for_completion);
4794 * completion_done - Test to see if a completion has any waiters
4795 * @x: completion structure
4797 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4798 * 1 if there are no waiters.
4801 bool completion_done(struct completion *x)
4803 unsigned long flags;
4804 int ret = 1;
4806 spin_lock_irqsave(&x->wait.lock, flags);
4807 if (!x->done)
4808 ret = 0;
4809 spin_unlock_irqrestore(&x->wait.lock, flags);
4810 return ret;
4812 EXPORT_SYMBOL(completion_done);
4814 static long __sched
4815 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4817 unsigned long flags;
4818 wait_queue_t wait;
4820 init_waitqueue_entry(&wait, current);
4822 __set_current_state(state);
4824 spin_lock_irqsave(&q->lock, flags);
4825 __add_wait_queue(q, &wait);
4826 spin_unlock(&q->lock);
4827 timeout = schedule_timeout(timeout);
4828 spin_lock_irq(&q->lock);
4829 __remove_wait_queue(q, &wait);
4830 spin_unlock_irqrestore(&q->lock, flags);
4832 return timeout;
4835 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4837 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4839 EXPORT_SYMBOL(interruptible_sleep_on);
4841 long __sched
4842 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4844 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4846 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4848 void __sched sleep_on(wait_queue_head_t *q)
4850 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4852 EXPORT_SYMBOL(sleep_on);
4854 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4856 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4858 EXPORT_SYMBOL(sleep_on_timeout);
4860 #ifdef CONFIG_RT_MUTEXES
4863 * rt_mutex_setprio - set the current priority of a task
4864 * @p: task
4865 * @prio: prio value (kernel-internal form)
4867 * This function changes the 'effective' priority of a task. It does
4868 * not touch ->normal_prio like __setscheduler().
4870 * Used by the rt_mutex code to implement priority inheritance logic.
4872 void rt_mutex_setprio(struct task_struct *p, int prio)
4874 int oldprio, on_rq, running;
4875 struct rq *rq;
4876 const struct sched_class *prev_class;
4878 BUG_ON(prio < 0 || prio > MAX_PRIO);
4880 rq = __task_rq_lock(p);
4882 trace_sched_pi_setprio(p, prio);
4883 oldprio = p->prio;
4884 prev_class = p->sched_class;
4885 on_rq = p->on_rq;
4886 running = task_current(rq, p);
4887 if (on_rq)
4888 dequeue_task(rq, p, 0);
4889 if (running)
4890 p->sched_class->put_prev_task(rq, p);
4892 if (rt_prio(prio))
4893 p->sched_class = &rt_sched_class;
4894 else
4895 p->sched_class = &fair_sched_class;
4897 p->prio = prio;
4899 if (running)
4900 p->sched_class->set_curr_task(rq);
4901 if (on_rq)
4902 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4904 check_class_changed(rq, p, prev_class, oldprio);
4905 __task_rq_unlock(rq);
4908 #endif
4910 void set_user_nice(struct task_struct *p, long nice)
4912 int old_prio, delta, on_rq;
4913 unsigned long flags;
4914 struct rq *rq;
4916 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4917 return;
4919 * We have to be careful, if called from sys_setpriority(),
4920 * the task might be in the middle of scheduling on another CPU.
4922 rq = task_rq_lock(p, &flags);
4924 * The RT priorities are set via sched_setscheduler(), but we still
4925 * allow the 'normal' nice value to be set - but as expected
4926 * it wont have any effect on scheduling until the task is
4927 * SCHED_FIFO/SCHED_RR:
4929 if (task_has_rt_policy(p)) {
4930 p->static_prio = NICE_TO_PRIO(nice);
4931 goto out_unlock;
4933 on_rq = p->on_rq;
4934 if (on_rq)
4935 dequeue_task(rq, p, 0);
4937 p->static_prio = NICE_TO_PRIO(nice);
4938 set_load_weight(p);
4939 old_prio = p->prio;
4940 p->prio = effective_prio(p);
4941 delta = p->prio - old_prio;
4943 if (on_rq) {
4944 enqueue_task(rq, p, 0);
4946 * If the task increased its priority or is running and
4947 * lowered its priority, then reschedule its CPU:
4949 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4950 resched_task(rq->curr);
4952 out_unlock:
4953 task_rq_unlock(rq, p, &flags);
4955 EXPORT_SYMBOL(set_user_nice);
4958 * can_nice - check if a task can reduce its nice value
4959 * @p: task
4960 * @nice: nice value
4962 int can_nice(const struct task_struct *p, const int nice)
4964 /* convert nice value [19,-20] to rlimit style value [1,40] */
4965 int nice_rlim = 20 - nice;
4967 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4968 capable(CAP_SYS_NICE));
4971 #ifdef __ARCH_WANT_SYS_NICE
4974 * sys_nice - change the priority of the current process.
4975 * @increment: priority increment
4977 * sys_setpriority is a more generic, but much slower function that
4978 * does similar things.
4980 SYSCALL_DEFINE1(nice, int, increment)
4982 long nice, retval;
4985 * Setpriority might change our priority at the same moment.
4986 * We don't have to worry. Conceptually one call occurs first
4987 * and we have a single winner.
4989 if (increment < -40)
4990 increment = -40;
4991 if (increment > 40)
4992 increment = 40;
4994 nice = TASK_NICE(current) + increment;
4995 if (nice < -20)
4996 nice = -20;
4997 if (nice > 19)
4998 nice = 19;
5000 if (increment < 0 && !can_nice(current, nice))
5001 return -EPERM;
5003 retval = security_task_setnice(current, nice);
5004 if (retval)
5005 return retval;
5007 set_user_nice(current, nice);
5008 return 0;
5011 #endif
5014 * task_prio - return the priority value of a given task.
5015 * @p: the task in question.
5017 * This is the priority value as seen by users in /proc.
5018 * RT tasks are offset by -200. Normal tasks are centered
5019 * around 0, value goes from -16 to +15.
5021 int task_prio(const struct task_struct *p)
5023 return p->prio - MAX_RT_PRIO;
5027 * task_nice - return the nice value of a given task.
5028 * @p: the task in question.
5030 int task_nice(const struct task_struct *p)
5032 return TASK_NICE(p);
5034 EXPORT_SYMBOL(task_nice);
5037 * idle_cpu - is a given cpu idle currently?
5038 * @cpu: the processor in question.
5040 int idle_cpu(int cpu)
5042 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5046 * idle_task - return the idle task for a given cpu.
5047 * @cpu: the processor in question.
5049 struct task_struct *idle_task(int cpu)
5051 return cpu_rq(cpu)->idle;
5055 * find_process_by_pid - find a process with a matching PID value.
5056 * @pid: the pid in question.
5058 static struct task_struct *find_process_by_pid(pid_t pid)
5060 return pid ? find_task_by_vpid(pid) : current;
5063 /* Actually do priority change: must hold rq lock. */
5064 static void
5065 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5067 p->policy = policy;
5068 p->rt_priority = prio;
5069 p->normal_prio = normal_prio(p);
5070 /* we are holding p->pi_lock already */
5071 p->prio = rt_mutex_getprio(p);
5072 if (rt_prio(p->prio))
5073 p->sched_class = &rt_sched_class;
5074 else
5075 p->sched_class = &fair_sched_class;
5076 set_load_weight(p);
5080 * check the target process has a UID that matches the current process's
5082 static bool check_same_owner(struct task_struct *p)
5084 const struct cred *cred = current_cred(), *pcred;
5085 bool match;
5087 rcu_read_lock();
5088 pcred = __task_cred(p);
5089 if (cred->user->user_ns == pcred->user->user_ns)
5090 match = (cred->euid == pcred->euid ||
5091 cred->euid == pcred->uid);
5092 else
5093 match = false;
5094 rcu_read_unlock();
5095 return match;
5098 static int __sched_setscheduler(struct task_struct *p, int policy,
5099 const struct sched_param *param, bool user)
5101 int retval, oldprio, oldpolicy = -1, on_rq, running;
5102 unsigned long flags;
5103 const struct sched_class *prev_class;
5104 struct rq *rq;
5105 int reset_on_fork;
5107 /* may grab non-irq protected spin_locks */
5108 BUG_ON(in_interrupt());
5109 recheck:
5110 /* double check policy once rq lock held */
5111 if (policy < 0) {
5112 reset_on_fork = p->sched_reset_on_fork;
5113 policy = oldpolicy = p->policy;
5114 } else {
5115 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5116 policy &= ~SCHED_RESET_ON_FORK;
5118 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5119 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5120 policy != SCHED_IDLE)
5121 return -EINVAL;
5125 * Valid priorities for SCHED_FIFO and SCHED_RR are
5126 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5127 * SCHED_BATCH and SCHED_IDLE is 0.
5129 if (param->sched_priority < 0 ||
5130 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5131 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5132 return -EINVAL;
5133 if (rt_policy(policy) != (param->sched_priority != 0))
5134 return -EINVAL;
5137 * Allow unprivileged RT tasks to decrease priority:
5139 if (user && !capable(CAP_SYS_NICE)) {
5140 if (rt_policy(policy)) {
5141 unsigned long rlim_rtprio =
5142 task_rlimit(p, RLIMIT_RTPRIO);
5144 /* can't set/change the rt policy */
5145 if (policy != p->policy && !rlim_rtprio)
5146 return -EPERM;
5148 /* can't increase priority */
5149 if (param->sched_priority > p->rt_priority &&
5150 param->sched_priority > rlim_rtprio)
5151 return -EPERM;
5155 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5156 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5158 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5159 if (!can_nice(p, TASK_NICE(p)))
5160 return -EPERM;
5163 /* can't change other user's priorities */
5164 if (!check_same_owner(p))
5165 return -EPERM;
5167 /* Normal users shall not reset the sched_reset_on_fork flag */
5168 if (p->sched_reset_on_fork && !reset_on_fork)
5169 return -EPERM;
5172 if (user) {
5173 retval = security_task_setscheduler(p);
5174 if (retval)
5175 return retval;
5179 * make sure no PI-waiters arrive (or leave) while we are
5180 * changing the priority of the task:
5182 * To be able to change p->policy safely, the appropriate
5183 * runqueue lock must be held.
5185 rq = task_rq_lock(p, &flags);
5188 * Changing the policy of the stop threads its a very bad idea
5190 if (p == rq->stop) {
5191 task_rq_unlock(rq, p, &flags);
5192 return -EINVAL;
5196 * If not changing anything there's no need to proceed further:
5198 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5199 param->sched_priority == p->rt_priority))) {
5201 __task_rq_unlock(rq);
5202 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5203 return 0;
5206 #ifdef CONFIG_RT_GROUP_SCHED
5207 if (user) {
5209 * Do not allow realtime tasks into groups that have no runtime
5210 * assigned.
5212 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5213 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5214 !task_group_is_autogroup(task_group(p))) {
5215 task_rq_unlock(rq, p, &flags);
5216 return -EPERM;
5219 #endif
5221 /* recheck policy now with rq lock held */
5222 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5223 policy = oldpolicy = -1;
5224 task_rq_unlock(rq, p, &flags);
5225 goto recheck;
5227 on_rq = p->on_rq;
5228 running = task_current(rq, p);
5229 if (on_rq)
5230 deactivate_task(rq, p, 0);
5231 if (running)
5232 p->sched_class->put_prev_task(rq, p);
5234 p->sched_reset_on_fork = reset_on_fork;
5236 oldprio = p->prio;
5237 prev_class = p->sched_class;
5238 __setscheduler(rq, p, policy, param->sched_priority);
5240 if (running)
5241 p->sched_class->set_curr_task(rq);
5242 if (on_rq)
5243 activate_task(rq, p, 0);
5245 check_class_changed(rq, p, prev_class, oldprio);
5246 task_rq_unlock(rq, p, &flags);
5248 rt_mutex_adjust_pi(p);
5250 return 0;
5254 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5255 * @p: the task in question.
5256 * @policy: new policy.
5257 * @param: structure containing the new RT priority.
5259 * NOTE that the task may be already dead.
5261 int sched_setscheduler(struct task_struct *p, int policy,
5262 const struct sched_param *param)
5264 return __sched_setscheduler(p, policy, param, true);
5266 EXPORT_SYMBOL_GPL(sched_setscheduler);
5269 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5270 * @p: the task in question.
5271 * @policy: new policy.
5272 * @param: structure containing the new RT priority.
5274 * Just like sched_setscheduler, only don't bother checking if the
5275 * current context has permission. For example, this is needed in
5276 * stop_machine(): we create temporary high priority worker threads,
5277 * but our caller might not have that capability.
5279 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5280 const struct sched_param *param)
5282 return __sched_setscheduler(p, policy, param, false);
5285 static int
5286 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5288 struct sched_param lparam;
5289 struct task_struct *p;
5290 int retval;
5292 if (!param || pid < 0)
5293 return -EINVAL;
5294 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5295 return -EFAULT;
5297 rcu_read_lock();
5298 retval = -ESRCH;
5299 p = find_process_by_pid(pid);
5300 if (p != NULL)
5301 retval = sched_setscheduler(p, policy, &lparam);
5302 rcu_read_unlock();
5304 return retval;
5308 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5309 * @pid: the pid in question.
5310 * @policy: new policy.
5311 * @param: structure containing the new RT priority.
5313 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5314 struct sched_param __user *, param)
5316 /* negative values for policy are not valid */
5317 if (policy < 0)
5318 return -EINVAL;
5320 return do_sched_setscheduler(pid, policy, param);
5324 * sys_sched_setparam - set/change the RT priority of a thread
5325 * @pid: the pid in question.
5326 * @param: structure containing the new RT priority.
5328 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5330 return do_sched_setscheduler(pid, -1, param);
5334 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5335 * @pid: the pid in question.
5337 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5339 struct task_struct *p;
5340 int retval;
5342 if (pid < 0)
5343 return -EINVAL;
5345 retval = -ESRCH;
5346 rcu_read_lock();
5347 p = find_process_by_pid(pid);
5348 if (p) {
5349 retval = security_task_getscheduler(p);
5350 if (!retval)
5351 retval = p->policy
5352 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5354 rcu_read_unlock();
5355 return retval;
5359 * sys_sched_getparam - get the RT priority of a thread
5360 * @pid: the pid in question.
5361 * @param: structure containing the RT priority.
5363 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5365 struct sched_param lp;
5366 struct task_struct *p;
5367 int retval;
5369 if (!param || pid < 0)
5370 return -EINVAL;
5372 rcu_read_lock();
5373 p = find_process_by_pid(pid);
5374 retval = -ESRCH;
5375 if (!p)
5376 goto out_unlock;
5378 retval = security_task_getscheduler(p);
5379 if (retval)
5380 goto out_unlock;
5382 lp.sched_priority = p->rt_priority;
5383 rcu_read_unlock();
5386 * This one might sleep, we cannot do it with a spinlock held ...
5388 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5390 return retval;
5392 out_unlock:
5393 rcu_read_unlock();
5394 return retval;
5397 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5399 cpumask_var_t cpus_allowed, new_mask;
5400 struct task_struct *p;
5401 int retval;
5403 get_online_cpus();
5404 rcu_read_lock();
5406 p = find_process_by_pid(pid);
5407 if (!p) {
5408 rcu_read_unlock();
5409 put_online_cpus();
5410 return -ESRCH;
5413 /* Prevent p going away */
5414 get_task_struct(p);
5415 rcu_read_unlock();
5417 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5418 retval = -ENOMEM;
5419 goto out_put_task;
5421 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5422 retval = -ENOMEM;
5423 goto out_free_cpus_allowed;
5425 retval = -EPERM;
5426 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5427 goto out_unlock;
5429 retval = security_task_setscheduler(p);
5430 if (retval)
5431 goto out_unlock;
5433 cpuset_cpus_allowed(p, cpus_allowed);
5434 cpumask_and(new_mask, in_mask, cpus_allowed);
5435 again:
5436 retval = set_cpus_allowed_ptr(p, new_mask);
5438 if (!retval) {
5439 cpuset_cpus_allowed(p, cpus_allowed);
5440 if (!cpumask_subset(new_mask, cpus_allowed)) {
5442 * We must have raced with a concurrent cpuset
5443 * update. Just reset the cpus_allowed to the
5444 * cpuset's cpus_allowed
5446 cpumask_copy(new_mask, cpus_allowed);
5447 goto again;
5450 out_unlock:
5451 free_cpumask_var(new_mask);
5452 out_free_cpus_allowed:
5453 free_cpumask_var(cpus_allowed);
5454 out_put_task:
5455 put_task_struct(p);
5456 put_online_cpus();
5457 return retval;
5460 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5461 struct cpumask *new_mask)
5463 if (len < cpumask_size())
5464 cpumask_clear(new_mask);
5465 else if (len > cpumask_size())
5466 len = cpumask_size();
5468 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5472 * sys_sched_setaffinity - set the cpu affinity of a process
5473 * @pid: pid of the process
5474 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5475 * @user_mask_ptr: user-space pointer to the new cpu mask
5477 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5478 unsigned long __user *, user_mask_ptr)
5480 cpumask_var_t new_mask;
5481 int retval;
5483 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5484 return -ENOMEM;
5486 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5487 if (retval == 0)
5488 retval = sched_setaffinity(pid, new_mask);
5489 free_cpumask_var(new_mask);
5490 return retval;
5493 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5495 struct task_struct *p;
5496 unsigned long flags;
5497 int retval;
5499 get_online_cpus();
5500 rcu_read_lock();
5502 retval = -ESRCH;
5503 p = find_process_by_pid(pid);
5504 if (!p)
5505 goto out_unlock;
5507 retval = security_task_getscheduler(p);
5508 if (retval)
5509 goto out_unlock;
5511 raw_spin_lock_irqsave(&p->pi_lock, flags);
5512 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5513 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5515 out_unlock:
5516 rcu_read_unlock();
5517 put_online_cpus();
5519 return retval;
5523 * sys_sched_getaffinity - get the cpu affinity of a process
5524 * @pid: pid of the process
5525 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5526 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5528 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5529 unsigned long __user *, user_mask_ptr)
5531 int ret;
5532 cpumask_var_t mask;
5534 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5535 return -EINVAL;
5536 if (len & (sizeof(unsigned long)-1))
5537 return -EINVAL;
5539 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5540 return -ENOMEM;
5542 ret = sched_getaffinity(pid, mask);
5543 if (ret == 0) {
5544 size_t retlen = min_t(size_t, len, cpumask_size());
5546 if (copy_to_user(user_mask_ptr, mask, retlen))
5547 ret = -EFAULT;
5548 else
5549 ret = retlen;
5551 free_cpumask_var(mask);
5553 return ret;
5557 * sys_sched_yield - yield the current processor to other threads.
5559 * This function yields the current CPU to other tasks. If there are no
5560 * other threads running on this CPU then this function will return.
5562 SYSCALL_DEFINE0(sched_yield)
5564 struct rq *rq = this_rq_lock();
5566 schedstat_inc(rq, yld_count);
5567 current->sched_class->yield_task(rq);
5570 * Since we are going to call schedule() anyway, there's
5571 * no need to preempt or enable interrupts:
5573 __release(rq->lock);
5574 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5575 do_raw_spin_unlock(&rq->lock);
5576 preempt_enable_no_resched();
5578 schedule();
5580 return 0;
5583 static inline int should_resched(void)
5585 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5588 static void __cond_resched(void)
5590 add_preempt_count(PREEMPT_ACTIVE);
5591 schedule();
5592 sub_preempt_count(PREEMPT_ACTIVE);
5595 int __sched _cond_resched(void)
5597 if (should_resched()) {
5598 __cond_resched();
5599 return 1;
5601 return 0;
5603 EXPORT_SYMBOL(_cond_resched);
5606 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5607 * call schedule, and on return reacquire the lock.
5609 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5610 * operations here to prevent schedule() from being called twice (once via
5611 * spin_unlock(), once by hand).
5613 int __cond_resched_lock(spinlock_t *lock)
5615 int resched = should_resched();
5616 int ret = 0;
5618 lockdep_assert_held(lock);
5620 if (spin_needbreak(lock) || resched) {
5621 spin_unlock(lock);
5622 if (resched)
5623 __cond_resched();
5624 else
5625 cpu_relax();
5626 ret = 1;
5627 spin_lock(lock);
5629 return ret;
5631 EXPORT_SYMBOL(__cond_resched_lock);
5633 int __sched __cond_resched_softirq(void)
5635 BUG_ON(!in_softirq());
5637 if (should_resched()) {
5638 local_bh_enable();
5639 __cond_resched();
5640 local_bh_disable();
5641 return 1;
5643 return 0;
5645 EXPORT_SYMBOL(__cond_resched_softirq);
5648 * yield - yield the current processor to other threads.
5650 * This is a shortcut for kernel-space yielding - it marks the
5651 * thread runnable and calls sys_sched_yield().
5653 void __sched yield(void)
5655 set_current_state(TASK_RUNNING);
5656 sys_sched_yield();
5658 EXPORT_SYMBOL(yield);
5661 * yield_to - yield the current processor to another thread in
5662 * your thread group, or accelerate that thread toward the
5663 * processor it's on.
5664 * @p: target task
5665 * @preempt: whether task preemption is allowed or not
5667 * It's the caller's job to ensure that the target task struct
5668 * can't go away on us before we can do any checks.
5670 * Returns true if we indeed boosted the target task.
5672 bool __sched yield_to(struct task_struct *p, bool preempt)
5674 struct task_struct *curr = current;
5675 struct rq *rq, *p_rq;
5676 unsigned long flags;
5677 bool yielded = 0;
5679 local_irq_save(flags);
5680 rq = this_rq();
5682 again:
5683 p_rq = task_rq(p);
5684 double_rq_lock(rq, p_rq);
5685 while (task_rq(p) != p_rq) {
5686 double_rq_unlock(rq, p_rq);
5687 goto again;
5690 if (!curr->sched_class->yield_to_task)
5691 goto out;
5693 if (curr->sched_class != p->sched_class)
5694 goto out;
5696 if (task_running(p_rq, p) || p->state)
5697 goto out;
5699 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5700 if (yielded) {
5701 schedstat_inc(rq, yld_count);
5703 * Make p's CPU reschedule; pick_next_entity takes care of
5704 * fairness.
5706 if (preempt && rq != p_rq)
5707 resched_task(p_rq->curr);
5710 out:
5711 double_rq_unlock(rq, p_rq);
5712 local_irq_restore(flags);
5714 if (yielded)
5715 schedule();
5717 return yielded;
5719 EXPORT_SYMBOL_GPL(yield_to);
5722 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5723 * that process accounting knows that this is a task in IO wait state.
5725 void __sched io_schedule(void)
5727 struct rq *rq = raw_rq();
5729 delayacct_blkio_start();
5730 atomic_inc(&rq->nr_iowait);
5731 blk_flush_plug(current);
5732 current->in_iowait = 1;
5733 schedule();
5734 current->in_iowait = 0;
5735 atomic_dec(&rq->nr_iowait);
5736 delayacct_blkio_end();
5738 EXPORT_SYMBOL(io_schedule);
5740 long __sched io_schedule_timeout(long timeout)
5742 struct rq *rq = raw_rq();
5743 long ret;
5745 delayacct_blkio_start();
5746 atomic_inc(&rq->nr_iowait);
5747 blk_flush_plug(current);
5748 current->in_iowait = 1;
5749 ret = schedule_timeout(timeout);
5750 current->in_iowait = 0;
5751 atomic_dec(&rq->nr_iowait);
5752 delayacct_blkio_end();
5753 return ret;
5757 * sys_sched_get_priority_max - return maximum RT priority.
5758 * @policy: scheduling class.
5760 * this syscall returns the maximum rt_priority that can be used
5761 * by a given scheduling class.
5763 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5765 int ret = -EINVAL;
5767 switch (policy) {
5768 case SCHED_FIFO:
5769 case SCHED_RR:
5770 ret = MAX_USER_RT_PRIO-1;
5771 break;
5772 case SCHED_NORMAL:
5773 case SCHED_BATCH:
5774 case SCHED_IDLE:
5775 ret = 0;
5776 break;
5778 return ret;
5782 * sys_sched_get_priority_min - return minimum RT priority.
5783 * @policy: scheduling class.
5785 * this syscall returns the minimum rt_priority that can be used
5786 * by a given scheduling class.
5788 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5790 int ret = -EINVAL;
5792 switch (policy) {
5793 case SCHED_FIFO:
5794 case SCHED_RR:
5795 ret = 1;
5796 break;
5797 case SCHED_NORMAL:
5798 case SCHED_BATCH:
5799 case SCHED_IDLE:
5800 ret = 0;
5802 return ret;
5806 * sys_sched_rr_get_interval - return the default timeslice of a process.
5807 * @pid: pid of the process.
5808 * @interval: userspace pointer to the timeslice value.
5810 * this syscall writes the default timeslice value of a given process
5811 * into the user-space timespec buffer. A value of '0' means infinity.
5813 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5814 struct timespec __user *, interval)
5816 struct task_struct *p;
5817 unsigned int time_slice;
5818 unsigned long flags;
5819 struct rq *rq;
5820 int retval;
5821 struct timespec t;
5823 if (pid < 0)
5824 return -EINVAL;
5826 retval = -ESRCH;
5827 rcu_read_lock();
5828 p = find_process_by_pid(pid);
5829 if (!p)
5830 goto out_unlock;
5832 retval = security_task_getscheduler(p);
5833 if (retval)
5834 goto out_unlock;
5836 rq = task_rq_lock(p, &flags);
5837 time_slice = p->sched_class->get_rr_interval(rq, p);
5838 task_rq_unlock(rq, p, &flags);
5840 rcu_read_unlock();
5841 jiffies_to_timespec(time_slice, &t);
5842 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5843 return retval;
5845 out_unlock:
5846 rcu_read_unlock();
5847 return retval;
5850 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5852 void sched_show_task(struct task_struct *p)
5854 unsigned long free = 0;
5855 unsigned state;
5857 state = p->state ? __ffs(p->state) + 1 : 0;
5858 printk(KERN_INFO "%-15.15s %c", p->comm,
5859 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5860 #if BITS_PER_LONG == 32
5861 if (state == TASK_RUNNING)
5862 printk(KERN_CONT " running ");
5863 else
5864 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5865 #else
5866 if (state == TASK_RUNNING)
5867 printk(KERN_CONT " running task ");
5868 else
5869 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5870 #endif
5871 #ifdef CONFIG_DEBUG_STACK_USAGE
5872 free = stack_not_used(p);
5873 #endif
5874 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5875 task_pid_nr(p), task_pid_nr(p->real_parent),
5876 (unsigned long)task_thread_info(p)->flags);
5878 show_stack(p, NULL);
5881 void show_state_filter(unsigned long state_filter)
5883 struct task_struct *g, *p;
5885 #if BITS_PER_LONG == 32
5886 printk(KERN_INFO
5887 " task PC stack pid father\n");
5888 #else
5889 printk(KERN_INFO
5890 " task PC stack pid father\n");
5891 #endif
5892 read_lock(&tasklist_lock);
5893 do_each_thread(g, p) {
5895 * reset the NMI-timeout, listing all files on a slow
5896 * console might take a lot of time:
5898 touch_nmi_watchdog();
5899 if (!state_filter || (p->state & state_filter))
5900 sched_show_task(p);
5901 } while_each_thread(g, p);
5903 touch_all_softlockup_watchdogs();
5905 #ifdef CONFIG_SCHED_DEBUG
5906 sysrq_sched_debug_show();
5907 #endif
5908 read_unlock(&tasklist_lock);
5910 * Only show locks if all tasks are dumped:
5912 if (!state_filter)
5913 debug_show_all_locks();
5916 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5918 idle->sched_class = &idle_sched_class;
5922 * init_idle - set up an idle thread for a given CPU
5923 * @idle: task in question
5924 * @cpu: cpu the idle task belongs to
5926 * NOTE: this function does not set the idle thread's NEED_RESCHED
5927 * flag, to make booting more robust.
5929 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5931 struct rq *rq = cpu_rq(cpu);
5932 unsigned long flags;
5934 raw_spin_lock_irqsave(&rq->lock, flags);
5936 __sched_fork(idle);
5937 idle->state = TASK_RUNNING;
5938 idle->se.exec_start = sched_clock();
5940 do_set_cpus_allowed(idle, cpumask_of(cpu));
5942 * We're having a chicken and egg problem, even though we are
5943 * holding rq->lock, the cpu isn't yet set to this cpu so the
5944 * lockdep check in task_group() will fail.
5946 * Similar case to sched_fork(). / Alternatively we could
5947 * use task_rq_lock() here and obtain the other rq->lock.
5949 * Silence PROVE_RCU
5951 rcu_read_lock();
5952 __set_task_cpu(idle, cpu);
5953 rcu_read_unlock();
5955 rq->curr = rq->idle = idle;
5956 #if defined(CONFIG_SMP)
5957 idle->on_cpu = 1;
5958 #endif
5959 raw_spin_unlock_irqrestore(&rq->lock, flags);
5961 /* Set the preempt count _outside_ the spinlocks! */
5962 task_thread_info(idle)->preempt_count = 0;
5965 * The idle tasks have their own, simple scheduling class:
5967 idle->sched_class = &idle_sched_class;
5968 ftrace_graph_init_idle_task(idle, cpu);
5972 * In a system that switches off the HZ timer nohz_cpu_mask
5973 * indicates which cpus entered this state. This is used
5974 * in the rcu update to wait only for active cpus. For system
5975 * which do not switch off the HZ timer nohz_cpu_mask should
5976 * always be CPU_BITS_NONE.
5978 cpumask_var_t nohz_cpu_mask;
5981 * Increase the granularity value when there are more CPUs,
5982 * because with more CPUs the 'effective latency' as visible
5983 * to users decreases. But the relationship is not linear,
5984 * so pick a second-best guess by going with the log2 of the
5985 * number of CPUs.
5987 * This idea comes from the SD scheduler of Con Kolivas:
5989 static int get_update_sysctl_factor(void)
5991 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5992 unsigned int factor;
5994 switch (sysctl_sched_tunable_scaling) {
5995 case SCHED_TUNABLESCALING_NONE:
5996 factor = 1;
5997 break;
5998 case SCHED_TUNABLESCALING_LINEAR:
5999 factor = cpus;
6000 break;
6001 case SCHED_TUNABLESCALING_LOG:
6002 default:
6003 factor = 1 + ilog2(cpus);
6004 break;
6007 return factor;
6010 static void update_sysctl(void)
6012 unsigned int factor = get_update_sysctl_factor();
6014 #define SET_SYSCTL(name) \
6015 (sysctl_##name = (factor) * normalized_sysctl_##name)
6016 SET_SYSCTL(sched_min_granularity);
6017 SET_SYSCTL(sched_latency);
6018 SET_SYSCTL(sched_wakeup_granularity);
6019 #undef SET_SYSCTL
6022 static inline void sched_init_granularity(void)
6024 update_sysctl();
6027 #ifdef CONFIG_SMP
6028 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
6030 if (p->sched_class && p->sched_class->set_cpus_allowed)
6031 p->sched_class->set_cpus_allowed(p, new_mask);
6032 else {
6033 cpumask_copy(&p->cpus_allowed, new_mask);
6034 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6039 * This is how migration works:
6041 * 1) we invoke migration_cpu_stop() on the target CPU using
6042 * stop_one_cpu().
6043 * 2) stopper starts to run (implicitly forcing the migrated thread
6044 * off the CPU)
6045 * 3) it checks whether the migrated task is still in the wrong runqueue.
6046 * 4) if it's in the wrong runqueue then the migration thread removes
6047 * it and puts it into the right queue.
6048 * 5) stopper completes and stop_one_cpu() returns and the migration
6049 * is done.
6053 * Change a given task's CPU affinity. Migrate the thread to a
6054 * proper CPU and schedule it away if the CPU it's executing on
6055 * is removed from the allowed bitmask.
6057 * NOTE: the caller must have a valid reference to the task, the
6058 * task must not exit() & deallocate itself prematurely. The
6059 * call is not atomic; no spinlocks may be held.
6061 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6063 unsigned long flags;
6064 struct rq *rq;
6065 unsigned int dest_cpu;
6066 int ret = 0;
6068 rq = task_rq_lock(p, &flags);
6070 if (cpumask_equal(&p->cpus_allowed, new_mask))
6071 goto out;
6073 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6074 ret = -EINVAL;
6075 goto out;
6078 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6079 ret = -EINVAL;
6080 goto out;
6083 do_set_cpus_allowed(p, new_mask);
6085 /* Can the task run on the task's current CPU? If so, we're done */
6086 if (cpumask_test_cpu(task_cpu(p), new_mask))
6087 goto out;
6089 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6090 if (p->on_rq) {
6091 struct migration_arg arg = { p, dest_cpu };
6092 /* Need help from migration thread: drop lock and wait. */
6093 task_rq_unlock(rq, p, &flags);
6094 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6095 tlb_migrate_finish(p->mm);
6096 return 0;
6098 out:
6099 task_rq_unlock(rq, p, &flags);
6101 return ret;
6103 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6106 * Move (not current) task off this cpu, onto dest cpu. We're doing
6107 * this because either it can't run here any more (set_cpus_allowed()
6108 * away from this CPU, or CPU going down), or because we're
6109 * attempting to rebalance this task on exec (sched_exec).
6111 * So we race with normal scheduler movements, but that's OK, as long
6112 * as the task is no longer on this CPU.
6114 * Returns non-zero if task was successfully migrated.
6116 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6118 struct rq *rq_dest, *rq_src;
6119 int ret = 0;
6121 if (unlikely(!cpu_active(dest_cpu)))
6122 return ret;
6124 rq_src = cpu_rq(src_cpu);
6125 rq_dest = cpu_rq(dest_cpu);
6127 raw_spin_lock(&p->pi_lock);
6128 double_rq_lock(rq_src, rq_dest);
6129 /* Already moved. */
6130 if (task_cpu(p) != src_cpu)
6131 goto done;
6132 /* Affinity changed (again). */
6133 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6134 goto fail;
6137 * If we're not on a rq, the next wake-up will ensure we're
6138 * placed properly.
6140 if (p->on_rq) {
6141 deactivate_task(rq_src, p, 0);
6142 set_task_cpu(p, dest_cpu);
6143 activate_task(rq_dest, p, 0);
6144 check_preempt_curr(rq_dest, p, 0);
6146 done:
6147 ret = 1;
6148 fail:
6149 double_rq_unlock(rq_src, rq_dest);
6150 raw_spin_unlock(&p->pi_lock);
6151 return ret;
6155 * migration_cpu_stop - this will be executed by a highprio stopper thread
6156 * and performs thread migration by bumping thread off CPU then
6157 * 'pushing' onto another runqueue.
6159 static int migration_cpu_stop(void *data)
6161 struct migration_arg *arg = data;
6164 * The original target cpu might have gone down and we might
6165 * be on another cpu but it doesn't matter.
6167 local_irq_disable();
6168 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6169 local_irq_enable();
6170 return 0;
6173 #ifdef CONFIG_HOTPLUG_CPU
6176 * Ensures that the idle task is using init_mm right before its cpu goes
6177 * offline.
6179 void idle_task_exit(void)
6181 struct mm_struct *mm = current->active_mm;
6183 BUG_ON(cpu_online(smp_processor_id()));
6185 if (mm != &init_mm)
6186 switch_mm(mm, &init_mm, current);
6187 mmdrop(mm);
6191 * While a dead CPU has no uninterruptible tasks queued at this point,
6192 * it might still have a nonzero ->nr_uninterruptible counter, because
6193 * for performance reasons the counter is not stricly tracking tasks to
6194 * their home CPUs. So we just add the counter to another CPU's counter,
6195 * to keep the global sum constant after CPU-down:
6197 static void migrate_nr_uninterruptible(struct rq *rq_src)
6199 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6201 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6202 rq_src->nr_uninterruptible = 0;
6206 * remove the tasks which were accounted by rq from calc_load_tasks.
6208 static void calc_global_load_remove(struct rq *rq)
6210 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6211 rq->calc_load_active = 0;
6215 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6216 * try_to_wake_up()->select_task_rq().
6218 * Called with rq->lock held even though we'er in stop_machine() and
6219 * there's no concurrency possible, we hold the required locks anyway
6220 * because of lock validation efforts.
6222 static void migrate_tasks(unsigned int dead_cpu)
6224 struct rq *rq = cpu_rq(dead_cpu);
6225 struct task_struct *next, *stop = rq->stop;
6226 int dest_cpu;
6229 * Fudge the rq selection such that the below task selection loop
6230 * doesn't get stuck on the currently eligible stop task.
6232 * We're currently inside stop_machine() and the rq is either stuck
6233 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6234 * either way we should never end up calling schedule() until we're
6235 * done here.
6237 rq->stop = NULL;
6239 for ( ; ; ) {
6241 * There's this thread running, bail when that's the only
6242 * remaining thread.
6244 if (rq->nr_running == 1)
6245 break;
6247 next = pick_next_task(rq);
6248 BUG_ON(!next);
6249 next->sched_class->put_prev_task(rq, next);
6251 /* Find suitable destination for @next, with force if needed. */
6252 dest_cpu = select_fallback_rq(dead_cpu, next);
6253 raw_spin_unlock(&rq->lock);
6255 __migrate_task(next, dead_cpu, dest_cpu);
6257 raw_spin_lock(&rq->lock);
6260 rq->stop = stop;
6263 #endif /* CONFIG_HOTPLUG_CPU */
6265 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6267 static struct ctl_table sd_ctl_dir[] = {
6269 .procname = "sched_domain",
6270 .mode = 0555,
6275 static struct ctl_table sd_ctl_root[] = {
6277 .procname = "kernel",
6278 .mode = 0555,
6279 .child = sd_ctl_dir,
6284 static struct ctl_table *sd_alloc_ctl_entry(int n)
6286 struct ctl_table *entry =
6287 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6289 return entry;
6292 static void sd_free_ctl_entry(struct ctl_table **tablep)
6294 struct ctl_table *entry;
6297 * In the intermediate directories, both the child directory and
6298 * procname are dynamically allocated and could fail but the mode
6299 * will always be set. In the lowest directory the names are
6300 * static strings and all have proc handlers.
6302 for (entry = *tablep; entry->mode; entry++) {
6303 if (entry->child)
6304 sd_free_ctl_entry(&entry->child);
6305 if (entry->proc_handler == NULL)
6306 kfree(entry->procname);
6309 kfree(*tablep);
6310 *tablep = NULL;
6313 static void
6314 set_table_entry(struct ctl_table *entry,
6315 const char *procname, void *data, int maxlen,
6316 mode_t mode, proc_handler *proc_handler)
6318 entry->procname = procname;
6319 entry->data = data;
6320 entry->maxlen = maxlen;
6321 entry->mode = mode;
6322 entry->proc_handler = proc_handler;
6325 static struct ctl_table *
6326 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6328 struct ctl_table *table = sd_alloc_ctl_entry(13);
6330 if (table == NULL)
6331 return NULL;
6333 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6334 sizeof(long), 0644, proc_doulongvec_minmax);
6335 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6336 sizeof(long), 0644, proc_doulongvec_minmax);
6337 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6338 sizeof(int), 0644, proc_dointvec_minmax);
6339 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6340 sizeof(int), 0644, proc_dointvec_minmax);
6341 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6342 sizeof(int), 0644, proc_dointvec_minmax);
6343 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6344 sizeof(int), 0644, proc_dointvec_minmax);
6345 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6346 sizeof(int), 0644, proc_dointvec_minmax);
6347 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6348 sizeof(int), 0644, proc_dointvec_minmax);
6349 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6350 sizeof(int), 0644, proc_dointvec_minmax);
6351 set_table_entry(&table[9], "cache_nice_tries",
6352 &sd->cache_nice_tries,
6353 sizeof(int), 0644, proc_dointvec_minmax);
6354 set_table_entry(&table[10], "flags", &sd->flags,
6355 sizeof(int), 0644, proc_dointvec_minmax);
6356 set_table_entry(&table[11], "name", sd->name,
6357 CORENAME_MAX_SIZE, 0444, proc_dostring);
6358 /* &table[12] is terminator */
6360 return table;
6363 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6365 struct ctl_table *entry, *table;
6366 struct sched_domain *sd;
6367 int domain_num = 0, i;
6368 char buf[32];
6370 for_each_domain(cpu, sd)
6371 domain_num++;
6372 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6373 if (table == NULL)
6374 return NULL;
6376 i = 0;
6377 for_each_domain(cpu, sd) {
6378 snprintf(buf, 32, "domain%d", i);
6379 entry->procname = kstrdup(buf, GFP_KERNEL);
6380 entry->mode = 0555;
6381 entry->child = sd_alloc_ctl_domain_table(sd);
6382 entry++;
6383 i++;
6385 return table;
6388 static struct ctl_table_header *sd_sysctl_header;
6389 static void register_sched_domain_sysctl(void)
6391 int i, cpu_num = num_possible_cpus();
6392 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6393 char buf[32];
6395 WARN_ON(sd_ctl_dir[0].child);
6396 sd_ctl_dir[0].child = entry;
6398 if (entry == NULL)
6399 return;
6401 for_each_possible_cpu(i) {
6402 snprintf(buf, 32, "cpu%d", i);
6403 entry->procname = kstrdup(buf, GFP_KERNEL);
6404 entry->mode = 0555;
6405 entry->child = sd_alloc_ctl_cpu_table(i);
6406 entry++;
6409 WARN_ON(sd_sysctl_header);
6410 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6413 /* may be called multiple times per register */
6414 static void unregister_sched_domain_sysctl(void)
6416 if (sd_sysctl_header)
6417 unregister_sysctl_table(sd_sysctl_header);
6418 sd_sysctl_header = NULL;
6419 if (sd_ctl_dir[0].child)
6420 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6422 #else
6423 static void register_sched_domain_sysctl(void)
6426 static void unregister_sched_domain_sysctl(void)
6429 #endif
6431 static void set_rq_online(struct rq *rq)
6433 if (!rq->online) {
6434 const struct sched_class *class;
6436 cpumask_set_cpu(rq->cpu, rq->rd->online);
6437 rq->online = 1;
6439 for_each_class(class) {
6440 if (class->rq_online)
6441 class->rq_online(rq);
6446 static void set_rq_offline(struct rq *rq)
6448 if (rq->online) {
6449 const struct sched_class *class;
6451 for_each_class(class) {
6452 if (class->rq_offline)
6453 class->rq_offline(rq);
6456 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6457 rq->online = 0;
6462 * migration_call - callback that gets triggered when a CPU is added.
6463 * Here we can start up the necessary migration thread for the new CPU.
6465 static int __cpuinit
6466 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6468 int cpu = (long)hcpu;
6469 unsigned long flags;
6470 struct rq *rq = cpu_rq(cpu);
6472 switch (action & ~CPU_TASKS_FROZEN) {
6474 case CPU_UP_PREPARE:
6475 rq->calc_load_update = calc_load_update;
6476 break;
6478 case CPU_ONLINE:
6479 /* Update our root-domain */
6480 raw_spin_lock_irqsave(&rq->lock, flags);
6481 if (rq->rd) {
6482 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6484 set_rq_online(rq);
6486 raw_spin_unlock_irqrestore(&rq->lock, flags);
6487 break;
6489 #ifdef CONFIG_HOTPLUG_CPU
6490 case CPU_DYING:
6491 sched_ttwu_pending();
6492 /* Update our root-domain */
6493 raw_spin_lock_irqsave(&rq->lock, flags);
6494 if (rq->rd) {
6495 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6496 set_rq_offline(rq);
6498 migrate_tasks(cpu);
6499 BUG_ON(rq->nr_running != 1); /* the migration thread */
6500 raw_spin_unlock_irqrestore(&rq->lock, flags);
6502 migrate_nr_uninterruptible(rq);
6503 calc_global_load_remove(rq);
6504 break;
6505 #endif
6508 update_max_interval();
6510 return NOTIFY_OK;
6514 * Register at high priority so that task migration (migrate_all_tasks)
6515 * happens before everything else. This has to be lower priority than
6516 * the notifier in the perf_event subsystem, though.
6518 static struct notifier_block __cpuinitdata migration_notifier = {
6519 .notifier_call = migration_call,
6520 .priority = CPU_PRI_MIGRATION,
6523 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6524 unsigned long action, void *hcpu)
6526 switch (action & ~CPU_TASKS_FROZEN) {
6527 case CPU_ONLINE:
6528 case CPU_DOWN_FAILED:
6529 set_cpu_active((long)hcpu, true);
6530 return NOTIFY_OK;
6531 default:
6532 return NOTIFY_DONE;
6536 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6537 unsigned long action, void *hcpu)
6539 switch (action & ~CPU_TASKS_FROZEN) {
6540 case CPU_DOWN_PREPARE:
6541 set_cpu_active((long)hcpu, false);
6542 return NOTIFY_OK;
6543 default:
6544 return NOTIFY_DONE;
6548 static int __init migration_init(void)
6550 void *cpu = (void *)(long)smp_processor_id();
6551 int err;
6553 /* Initialize migration for the boot CPU */
6554 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6555 BUG_ON(err == NOTIFY_BAD);
6556 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6557 register_cpu_notifier(&migration_notifier);
6559 /* Register cpu active notifiers */
6560 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6561 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6563 return 0;
6565 early_initcall(migration_init);
6566 #endif
6568 #ifdef CONFIG_SMP
6570 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6572 #ifdef CONFIG_SCHED_DEBUG
6574 static __read_mostly int sched_domain_debug_enabled;
6576 static int __init sched_domain_debug_setup(char *str)
6578 sched_domain_debug_enabled = 1;
6580 return 0;
6582 early_param("sched_debug", sched_domain_debug_setup);
6584 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6585 struct cpumask *groupmask)
6587 struct sched_group *group = sd->groups;
6588 char str[256];
6590 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6591 cpumask_clear(groupmask);
6593 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6595 if (!(sd->flags & SD_LOAD_BALANCE)) {
6596 printk("does not load-balance\n");
6597 if (sd->parent)
6598 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6599 " has parent");
6600 return -1;
6603 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6605 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6606 printk(KERN_ERR "ERROR: domain->span does not contain "
6607 "CPU%d\n", cpu);
6609 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6610 printk(KERN_ERR "ERROR: domain->groups does not contain"
6611 " CPU%d\n", cpu);
6614 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6615 do {
6616 if (!group) {
6617 printk("\n");
6618 printk(KERN_ERR "ERROR: group is NULL\n");
6619 break;
6622 if (!group->sgp->power) {
6623 printk(KERN_CONT "\n");
6624 printk(KERN_ERR "ERROR: domain->cpu_power not "
6625 "set\n");
6626 break;
6629 if (!cpumask_weight(sched_group_cpus(group))) {
6630 printk(KERN_CONT "\n");
6631 printk(KERN_ERR "ERROR: empty group\n");
6632 break;
6635 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6636 printk(KERN_CONT "\n");
6637 printk(KERN_ERR "ERROR: repeated CPUs\n");
6638 break;
6641 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6643 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6645 printk(KERN_CONT " %s", str);
6646 if (group->sgp->power != SCHED_POWER_SCALE) {
6647 printk(KERN_CONT " (cpu_power = %d)",
6648 group->sgp->power);
6651 group = group->next;
6652 } while (group != sd->groups);
6653 printk(KERN_CONT "\n");
6655 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6656 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6658 if (sd->parent &&
6659 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6660 printk(KERN_ERR "ERROR: parent span is not a superset "
6661 "of domain->span\n");
6662 return 0;
6665 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6667 int level = 0;
6669 if (!sched_domain_debug_enabled)
6670 return;
6672 if (!sd) {
6673 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6674 return;
6677 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6679 for (;;) {
6680 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6681 break;
6682 level++;
6683 sd = sd->parent;
6684 if (!sd)
6685 break;
6688 #else /* !CONFIG_SCHED_DEBUG */
6689 # define sched_domain_debug(sd, cpu) do { } while (0)
6690 #endif /* CONFIG_SCHED_DEBUG */
6692 static int sd_degenerate(struct sched_domain *sd)
6694 if (cpumask_weight(sched_domain_span(sd)) == 1)
6695 return 1;
6697 /* Following flags need at least 2 groups */
6698 if (sd->flags & (SD_LOAD_BALANCE |
6699 SD_BALANCE_NEWIDLE |
6700 SD_BALANCE_FORK |
6701 SD_BALANCE_EXEC |
6702 SD_SHARE_CPUPOWER |
6703 SD_SHARE_PKG_RESOURCES)) {
6704 if (sd->groups != sd->groups->next)
6705 return 0;
6708 /* Following flags don't use groups */
6709 if (sd->flags & (SD_WAKE_AFFINE))
6710 return 0;
6712 return 1;
6715 static int
6716 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6718 unsigned long cflags = sd->flags, pflags = parent->flags;
6720 if (sd_degenerate(parent))
6721 return 1;
6723 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6724 return 0;
6726 /* Flags needing groups don't count if only 1 group in parent */
6727 if (parent->groups == parent->groups->next) {
6728 pflags &= ~(SD_LOAD_BALANCE |
6729 SD_BALANCE_NEWIDLE |
6730 SD_BALANCE_FORK |
6731 SD_BALANCE_EXEC |
6732 SD_SHARE_CPUPOWER |
6733 SD_SHARE_PKG_RESOURCES);
6734 if (nr_node_ids == 1)
6735 pflags &= ~SD_SERIALIZE;
6737 if (~cflags & pflags)
6738 return 0;
6740 return 1;
6743 static void free_rootdomain(struct rcu_head *rcu)
6745 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6747 cpupri_cleanup(&rd->cpupri);
6748 free_cpumask_var(rd->rto_mask);
6749 free_cpumask_var(rd->online);
6750 free_cpumask_var(rd->span);
6751 kfree(rd);
6754 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6756 struct root_domain *old_rd = NULL;
6757 unsigned long flags;
6759 raw_spin_lock_irqsave(&rq->lock, flags);
6761 if (rq->rd) {
6762 old_rd = rq->rd;
6764 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6765 set_rq_offline(rq);
6767 cpumask_clear_cpu(rq->cpu, old_rd->span);
6770 * If we dont want to free the old_rt yet then
6771 * set old_rd to NULL to skip the freeing later
6772 * in this function:
6774 if (!atomic_dec_and_test(&old_rd->refcount))
6775 old_rd = NULL;
6778 atomic_inc(&rd->refcount);
6779 rq->rd = rd;
6781 cpumask_set_cpu(rq->cpu, rd->span);
6782 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6783 set_rq_online(rq);
6785 raw_spin_unlock_irqrestore(&rq->lock, flags);
6787 if (old_rd)
6788 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6791 static int init_rootdomain(struct root_domain *rd)
6793 memset(rd, 0, sizeof(*rd));
6795 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6796 goto out;
6797 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6798 goto free_span;
6799 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6800 goto free_online;
6802 if (cpupri_init(&rd->cpupri) != 0)
6803 goto free_rto_mask;
6804 return 0;
6806 free_rto_mask:
6807 free_cpumask_var(rd->rto_mask);
6808 free_online:
6809 free_cpumask_var(rd->online);
6810 free_span:
6811 free_cpumask_var(rd->span);
6812 out:
6813 return -ENOMEM;
6816 static void init_defrootdomain(void)
6818 init_rootdomain(&def_root_domain);
6820 atomic_set(&def_root_domain.refcount, 1);
6823 static struct root_domain *alloc_rootdomain(void)
6825 struct root_domain *rd;
6827 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6828 if (!rd)
6829 return NULL;
6831 if (init_rootdomain(rd) != 0) {
6832 kfree(rd);
6833 return NULL;
6836 return rd;
6839 static void free_sched_groups(struct sched_group *sg, int free_sgp)
6841 struct sched_group *tmp, *first;
6843 if (!sg)
6844 return;
6846 first = sg;
6847 do {
6848 tmp = sg->next;
6850 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
6851 kfree(sg->sgp);
6853 kfree(sg);
6854 sg = tmp;
6855 } while (sg != first);
6858 static void free_sched_domain(struct rcu_head *rcu)
6860 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6863 * If its an overlapping domain it has private groups, iterate and
6864 * nuke them all.
6866 if (sd->flags & SD_OVERLAP) {
6867 free_sched_groups(sd->groups, 1);
6868 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6869 kfree(sd->groups->sgp);
6870 kfree(sd->groups);
6872 kfree(sd);
6875 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6877 call_rcu(&sd->rcu, free_sched_domain);
6880 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6882 for (; sd; sd = sd->parent)
6883 destroy_sched_domain(sd, cpu);
6887 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6888 * hold the hotplug lock.
6890 static void
6891 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6893 struct rq *rq = cpu_rq(cpu);
6894 struct sched_domain *tmp;
6896 /* Remove the sched domains which do not contribute to scheduling. */
6897 for (tmp = sd; tmp; ) {
6898 struct sched_domain *parent = tmp->parent;
6899 if (!parent)
6900 break;
6902 if (sd_parent_degenerate(tmp, parent)) {
6903 tmp->parent = parent->parent;
6904 if (parent->parent)
6905 parent->parent->child = tmp;
6906 destroy_sched_domain(parent, cpu);
6907 } else
6908 tmp = tmp->parent;
6911 if (sd && sd_degenerate(sd)) {
6912 tmp = sd;
6913 sd = sd->parent;
6914 destroy_sched_domain(tmp, cpu);
6915 if (sd)
6916 sd->child = NULL;
6919 sched_domain_debug(sd, cpu);
6921 rq_attach_root(rq, rd);
6922 tmp = rq->sd;
6923 rcu_assign_pointer(rq->sd, sd);
6924 destroy_sched_domains(tmp, cpu);
6927 /* cpus with isolated domains */
6928 static cpumask_var_t cpu_isolated_map;
6930 /* Setup the mask of cpus configured for isolated domains */
6931 static int __init isolated_cpu_setup(char *str)
6933 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6934 cpulist_parse(str, cpu_isolated_map);
6935 return 1;
6938 __setup("isolcpus=", isolated_cpu_setup);
6940 #define SD_NODES_PER_DOMAIN 16
6942 #ifdef CONFIG_NUMA
6945 * find_next_best_node - find the next node to include in a sched_domain
6946 * @node: node whose sched_domain we're building
6947 * @used_nodes: nodes already in the sched_domain
6949 * Find the next node to include in a given scheduling domain. Simply
6950 * finds the closest node not already in the @used_nodes map.
6952 * Should use nodemask_t.
6954 static int find_next_best_node(int node, nodemask_t *used_nodes)
6956 int i, n, val, min_val, best_node = -1;
6958 min_val = INT_MAX;
6960 for (i = 0; i < nr_node_ids; i++) {
6961 /* Start at @node */
6962 n = (node + i) % nr_node_ids;
6964 if (!nr_cpus_node(n))
6965 continue;
6967 /* Skip already used nodes */
6968 if (node_isset(n, *used_nodes))
6969 continue;
6971 /* Simple min distance search */
6972 val = node_distance(node, n);
6974 if (val < min_val) {
6975 min_val = val;
6976 best_node = n;
6980 if (best_node != -1)
6981 node_set(best_node, *used_nodes);
6982 return best_node;
6986 * sched_domain_node_span - get a cpumask for a node's sched_domain
6987 * @node: node whose cpumask we're constructing
6988 * @span: resulting cpumask
6990 * Given a node, construct a good cpumask for its sched_domain to span. It
6991 * should be one that prevents unnecessary balancing, but also spreads tasks
6992 * out optimally.
6994 static void sched_domain_node_span(int node, struct cpumask *span)
6996 nodemask_t used_nodes;
6997 int i;
6999 cpumask_clear(span);
7000 nodes_clear(used_nodes);
7002 cpumask_or(span, span, cpumask_of_node(node));
7003 node_set(node, used_nodes);
7005 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7006 int next_node = find_next_best_node(node, &used_nodes);
7007 if (next_node < 0)
7008 break;
7009 cpumask_or(span, span, cpumask_of_node(next_node));
7013 static const struct cpumask *cpu_node_mask(int cpu)
7015 lockdep_assert_held(&sched_domains_mutex);
7017 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
7019 return sched_domains_tmpmask;
7022 static const struct cpumask *cpu_allnodes_mask(int cpu)
7024 return cpu_possible_mask;
7026 #endif /* CONFIG_NUMA */
7028 static const struct cpumask *cpu_cpu_mask(int cpu)
7030 return cpumask_of_node(cpu_to_node(cpu));
7033 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7035 struct sd_data {
7036 struct sched_domain **__percpu sd;
7037 struct sched_group **__percpu sg;
7038 struct sched_group_power **__percpu sgp;
7041 struct s_data {
7042 struct sched_domain ** __percpu sd;
7043 struct root_domain *rd;
7046 enum s_alloc {
7047 sa_rootdomain,
7048 sa_sd,
7049 sa_sd_storage,
7050 sa_none,
7053 struct sched_domain_topology_level;
7055 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7056 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7058 #define SDTL_OVERLAP 0x01
7060 struct sched_domain_topology_level {
7061 sched_domain_init_f init;
7062 sched_domain_mask_f mask;
7063 int flags;
7064 struct sd_data data;
7067 static int
7068 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7070 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7071 const struct cpumask *span = sched_domain_span(sd);
7072 struct cpumask *covered = sched_domains_tmpmask;
7073 struct sd_data *sdd = sd->private;
7074 struct sched_domain *child;
7075 int i;
7077 cpumask_clear(covered);
7079 for_each_cpu(i, span) {
7080 struct cpumask *sg_span;
7082 if (cpumask_test_cpu(i, covered))
7083 continue;
7085 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7086 GFP_KERNEL, cpu_to_node(i));
7088 if (!sg)
7089 goto fail;
7091 sg_span = sched_group_cpus(sg);
7093 child = *per_cpu_ptr(sdd->sd, i);
7094 if (child->child) {
7095 child = child->child;
7096 cpumask_copy(sg_span, sched_domain_span(child));
7097 } else
7098 cpumask_set_cpu(i, sg_span);
7100 cpumask_or(covered, covered, sg_span);
7102 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7103 atomic_inc(&sg->sgp->ref);
7105 if (cpumask_test_cpu(cpu, sg_span))
7106 groups = sg;
7108 if (!first)
7109 first = sg;
7110 if (last)
7111 last->next = sg;
7112 last = sg;
7113 last->next = first;
7115 sd->groups = groups;
7117 return 0;
7119 fail:
7120 free_sched_groups(first, 0);
7122 return -ENOMEM;
7125 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7127 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7128 struct sched_domain *child = sd->child;
7130 if (child)
7131 cpu = cpumask_first(sched_domain_span(child));
7133 if (sg) {
7134 *sg = *per_cpu_ptr(sdd->sg, cpu);
7135 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7136 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7139 return cpu;
7143 * build_sched_groups will build a circular linked list of the groups
7144 * covered by the given span, and will set each group's ->cpumask correctly,
7145 * and ->cpu_power to 0.
7147 * Assumes the sched_domain tree is fully constructed
7149 static int
7150 build_sched_groups(struct sched_domain *sd, int cpu)
7152 struct sched_group *first = NULL, *last = NULL;
7153 struct sd_data *sdd = sd->private;
7154 const struct cpumask *span = sched_domain_span(sd);
7155 struct cpumask *covered;
7156 int i;
7158 get_group(cpu, sdd, &sd->groups);
7159 atomic_inc(&sd->groups->ref);
7161 if (cpu != cpumask_first(sched_domain_span(sd)))
7162 return 0;
7164 lockdep_assert_held(&sched_domains_mutex);
7165 covered = sched_domains_tmpmask;
7167 cpumask_clear(covered);
7169 for_each_cpu(i, span) {
7170 struct sched_group *sg;
7171 int group = get_group(i, sdd, &sg);
7172 int j;
7174 if (cpumask_test_cpu(i, covered))
7175 continue;
7177 cpumask_clear(sched_group_cpus(sg));
7178 sg->sgp->power = 0;
7180 for_each_cpu(j, span) {
7181 if (get_group(j, sdd, NULL) != group)
7182 continue;
7184 cpumask_set_cpu(j, covered);
7185 cpumask_set_cpu(j, sched_group_cpus(sg));
7188 if (!first)
7189 first = sg;
7190 if (last)
7191 last->next = sg;
7192 last = sg;
7194 last->next = first;
7196 return 0;
7200 * Initialize sched groups cpu_power.
7202 * cpu_power indicates the capacity of sched group, which is used while
7203 * distributing the load between different sched groups in a sched domain.
7204 * Typically cpu_power for all the groups in a sched domain will be same unless
7205 * there are asymmetries in the topology. If there are asymmetries, group
7206 * having more cpu_power will pickup more load compared to the group having
7207 * less cpu_power.
7209 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7211 struct sched_group *sg = sd->groups;
7213 WARN_ON(!sd || !sg);
7215 do {
7216 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7217 sg = sg->next;
7218 } while (sg != sd->groups);
7220 if (cpu != group_first_cpu(sg))
7221 return;
7223 update_group_power(sd, cpu);
7227 * Initializers for schedule domains
7228 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7231 #ifdef CONFIG_SCHED_DEBUG
7232 # define SD_INIT_NAME(sd, type) sd->name = #type
7233 #else
7234 # define SD_INIT_NAME(sd, type) do { } while (0)
7235 #endif
7237 #define SD_INIT_FUNC(type) \
7238 static noinline struct sched_domain * \
7239 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7241 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7242 *sd = SD_##type##_INIT; \
7243 SD_INIT_NAME(sd, type); \
7244 sd->private = &tl->data; \
7245 return sd; \
7248 SD_INIT_FUNC(CPU)
7249 #ifdef CONFIG_NUMA
7250 SD_INIT_FUNC(ALLNODES)
7251 SD_INIT_FUNC(NODE)
7252 #endif
7253 #ifdef CONFIG_SCHED_SMT
7254 SD_INIT_FUNC(SIBLING)
7255 #endif
7256 #ifdef CONFIG_SCHED_MC
7257 SD_INIT_FUNC(MC)
7258 #endif
7259 #ifdef CONFIG_SCHED_BOOK
7260 SD_INIT_FUNC(BOOK)
7261 #endif
7263 static int default_relax_domain_level = -1;
7264 int sched_domain_level_max;
7266 static int __init setup_relax_domain_level(char *str)
7268 unsigned long val;
7270 val = simple_strtoul(str, NULL, 0);
7271 if (val < sched_domain_level_max)
7272 default_relax_domain_level = val;
7274 return 1;
7276 __setup("relax_domain_level=", setup_relax_domain_level);
7278 static void set_domain_attribute(struct sched_domain *sd,
7279 struct sched_domain_attr *attr)
7281 int request;
7283 if (!attr || attr->relax_domain_level < 0) {
7284 if (default_relax_domain_level < 0)
7285 return;
7286 else
7287 request = default_relax_domain_level;
7288 } else
7289 request = attr->relax_domain_level;
7290 if (request < sd->level) {
7291 /* turn off idle balance on this domain */
7292 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7293 } else {
7294 /* turn on idle balance on this domain */
7295 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7299 static void __sdt_free(const struct cpumask *cpu_map);
7300 static int __sdt_alloc(const struct cpumask *cpu_map);
7302 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7303 const struct cpumask *cpu_map)
7305 switch (what) {
7306 case sa_rootdomain:
7307 if (!atomic_read(&d->rd->refcount))
7308 free_rootdomain(&d->rd->rcu); /* fall through */
7309 case sa_sd:
7310 free_percpu(d->sd); /* fall through */
7311 case sa_sd_storage:
7312 __sdt_free(cpu_map); /* fall through */
7313 case sa_none:
7314 break;
7318 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7319 const struct cpumask *cpu_map)
7321 memset(d, 0, sizeof(*d));
7323 if (__sdt_alloc(cpu_map))
7324 return sa_sd_storage;
7325 d->sd = alloc_percpu(struct sched_domain *);
7326 if (!d->sd)
7327 return sa_sd_storage;
7328 d->rd = alloc_rootdomain();
7329 if (!d->rd)
7330 return sa_sd;
7331 return sa_rootdomain;
7335 * NULL the sd_data elements we've used to build the sched_domain and
7336 * sched_group structure so that the subsequent __free_domain_allocs()
7337 * will not free the data we're using.
7339 static void claim_allocations(int cpu, struct sched_domain *sd)
7341 struct sd_data *sdd = sd->private;
7343 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7344 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7346 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7347 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7349 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7350 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7353 #ifdef CONFIG_SCHED_SMT
7354 static const struct cpumask *cpu_smt_mask(int cpu)
7356 return topology_thread_cpumask(cpu);
7358 #endif
7361 * Topology list, bottom-up.
7363 static struct sched_domain_topology_level default_topology[] = {
7364 #ifdef CONFIG_SCHED_SMT
7365 { sd_init_SIBLING, cpu_smt_mask, },
7366 #endif
7367 #ifdef CONFIG_SCHED_MC
7368 { sd_init_MC, cpu_coregroup_mask, },
7369 #endif
7370 #ifdef CONFIG_SCHED_BOOK
7371 { sd_init_BOOK, cpu_book_mask, },
7372 #endif
7373 { sd_init_CPU, cpu_cpu_mask, },
7374 #ifdef CONFIG_NUMA
7375 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7376 { sd_init_ALLNODES, cpu_allnodes_mask, },
7377 #endif
7378 { NULL, },
7381 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7383 static int __sdt_alloc(const struct cpumask *cpu_map)
7385 struct sched_domain_topology_level *tl;
7386 int j;
7388 for (tl = sched_domain_topology; tl->init; tl++) {
7389 struct sd_data *sdd = &tl->data;
7391 sdd->sd = alloc_percpu(struct sched_domain *);
7392 if (!sdd->sd)
7393 return -ENOMEM;
7395 sdd->sg = alloc_percpu(struct sched_group *);
7396 if (!sdd->sg)
7397 return -ENOMEM;
7399 sdd->sgp = alloc_percpu(struct sched_group_power *);
7400 if (!sdd->sgp)
7401 return -ENOMEM;
7403 for_each_cpu(j, cpu_map) {
7404 struct sched_domain *sd;
7405 struct sched_group *sg;
7406 struct sched_group_power *sgp;
7408 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7409 GFP_KERNEL, cpu_to_node(j));
7410 if (!sd)
7411 return -ENOMEM;
7413 *per_cpu_ptr(sdd->sd, j) = sd;
7415 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7416 GFP_KERNEL, cpu_to_node(j));
7417 if (!sg)
7418 return -ENOMEM;
7420 *per_cpu_ptr(sdd->sg, j) = sg;
7422 sgp = kzalloc_node(sizeof(struct sched_group_power),
7423 GFP_KERNEL, cpu_to_node(j));
7424 if (!sgp)
7425 return -ENOMEM;
7427 *per_cpu_ptr(sdd->sgp, j) = sgp;
7431 return 0;
7434 static void __sdt_free(const struct cpumask *cpu_map)
7436 struct sched_domain_topology_level *tl;
7437 int j;
7439 for (tl = sched_domain_topology; tl->init; tl++) {
7440 struct sd_data *sdd = &tl->data;
7442 for_each_cpu(j, cpu_map) {
7443 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
7444 if (sd && (sd->flags & SD_OVERLAP))
7445 free_sched_groups(sd->groups, 0);
7446 kfree(*per_cpu_ptr(sdd->sg, j));
7447 kfree(*per_cpu_ptr(sdd->sgp, j));
7449 free_percpu(sdd->sd);
7450 free_percpu(sdd->sg);
7451 free_percpu(sdd->sgp);
7455 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7456 struct s_data *d, const struct cpumask *cpu_map,
7457 struct sched_domain_attr *attr, struct sched_domain *child,
7458 int cpu)
7460 struct sched_domain *sd = tl->init(tl, cpu);
7461 if (!sd)
7462 return child;
7464 set_domain_attribute(sd, attr);
7465 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7466 if (child) {
7467 sd->level = child->level + 1;
7468 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7469 child->parent = sd;
7471 sd->child = child;
7473 return sd;
7477 * Build sched domains for a given set of cpus and attach the sched domains
7478 * to the individual cpus
7480 static int build_sched_domains(const struct cpumask *cpu_map,
7481 struct sched_domain_attr *attr)
7483 enum s_alloc alloc_state = sa_none;
7484 struct sched_domain *sd;
7485 struct s_data d;
7486 int i, ret = -ENOMEM;
7488 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7489 if (alloc_state != sa_rootdomain)
7490 goto error;
7492 /* Set up domains for cpus specified by the cpu_map. */
7493 for_each_cpu(i, cpu_map) {
7494 struct sched_domain_topology_level *tl;
7496 sd = NULL;
7497 for (tl = sched_domain_topology; tl->init; tl++) {
7498 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7499 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7500 sd->flags |= SD_OVERLAP;
7501 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7502 break;
7505 while (sd->child)
7506 sd = sd->child;
7508 *per_cpu_ptr(d.sd, i) = sd;
7511 /* Build the groups for the domains */
7512 for_each_cpu(i, cpu_map) {
7513 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7514 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7515 if (sd->flags & SD_OVERLAP) {
7516 if (build_overlap_sched_groups(sd, i))
7517 goto error;
7518 } else {
7519 if (build_sched_groups(sd, i))
7520 goto error;
7525 /* Calculate CPU power for physical packages and nodes */
7526 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7527 if (!cpumask_test_cpu(i, cpu_map))
7528 continue;
7530 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7531 claim_allocations(i, sd);
7532 init_sched_groups_power(i, sd);
7536 /* Attach the domains */
7537 rcu_read_lock();
7538 for_each_cpu(i, cpu_map) {
7539 sd = *per_cpu_ptr(d.sd, i);
7540 cpu_attach_domain(sd, d.rd, i);
7542 rcu_read_unlock();
7544 ret = 0;
7545 error:
7546 __free_domain_allocs(&d, alloc_state, cpu_map);
7547 return ret;
7550 static cpumask_var_t *doms_cur; /* current sched domains */
7551 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7552 static struct sched_domain_attr *dattr_cur;
7553 /* attribues of custom domains in 'doms_cur' */
7556 * Special case: If a kmalloc of a doms_cur partition (array of
7557 * cpumask) fails, then fallback to a single sched domain,
7558 * as determined by the single cpumask fallback_doms.
7560 static cpumask_var_t fallback_doms;
7563 * arch_update_cpu_topology lets virtualized architectures update the
7564 * cpu core maps. It is supposed to return 1 if the topology changed
7565 * or 0 if it stayed the same.
7567 int __attribute__((weak)) arch_update_cpu_topology(void)
7569 return 0;
7572 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7574 int i;
7575 cpumask_var_t *doms;
7577 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7578 if (!doms)
7579 return NULL;
7580 for (i = 0; i < ndoms; i++) {
7581 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7582 free_sched_domains(doms, i);
7583 return NULL;
7586 return doms;
7589 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7591 unsigned int i;
7592 for (i = 0; i < ndoms; i++)
7593 free_cpumask_var(doms[i]);
7594 kfree(doms);
7598 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7599 * For now this just excludes isolated cpus, but could be used to
7600 * exclude other special cases in the future.
7602 static int init_sched_domains(const struct cpumask *cpu_map)
7604 int err;
7606 arch_update_cpu_topology();
7607 ndoms_cur = 1;
7608 doms_cur = alloc_sched_domains(ndoms_cur);
7609 if (!doms_cur)
7610 doms_cur = &fallback_doms;
7611 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7612 dattr_cur = NULL;
7613 err = build_sched_domains(doms_cur[0], NULL);
7614 register_sched_domain_sysctl();
7616 return err;
7620 * Detach sched domains from a group of cpus specified in cpu_map
7621 * These cpus will now be attached to the NULL domain
7623 static void detach_destroy_domains(const struct cpumask *cpu_map)
7625 int i;
7627 rcu_read_lock();
7628 for_each_cpu(i, cpu_map)
7629 cpu_attach_domain(NULL, &def_root_domain, i);
7630 rcu_read_unlock();
7633 /* handle null as "default" */
7634 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7635 struct sched_domain_attr *new, int idx_new)
7637 struct sched_domain_attr tmp;
7639 /* fast path */
7640 if (!new && !cur)
7641 return 1;
7643 tmp = SD_ATTR_INIT;
7644 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7645 new ? (new + idx_new) : &tmp,
7646 sizeof(struct sched_domain_attr));
7650 * Partition sched domains as specified by the 'ndoms_new'
7651 * cpumasks in the array doms_new[] of cpumasks. This compares
7652 * doms_new[] to the current sched domain partitioning, doms_cur[].
7653 * It destroys each deleted domain and builds each new domain.
7655 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7656 * The masks don't intersect (don't overlap.) We should setup one
7657 * sched domain for each mask. CPUs not in any of the cpumasks will
7658 * not be load balanced. If the same cpumask appears both in the
7659 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7660 * it as it is.
7662 * The passed in 'doms_new' should be allocated using
7663 * alloc_sched_domains. This routine takes ownership of it and will
7664 * free_sched_domains it when done with it. If the caller failed the
7665 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7666 * and partition_sched_domains() will fallback to the single partition
7667 * 'fallback_doms', it also forces the domains to be rebuilt.
7669 * If doms_new == NULL it will be replaced with cpu_online_mask.
7670 * ndoms_new == 0 is a special case for destroying existing domains,
7671 * and it will not create the default domain.
7673 * Call with hotplug lock held
7675 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7676 struct sched_domain_attr *dattr_new)
7678 int i, j, n;
7679 int new_topology;
7681 mutex_lock(&sched_domains_mutex);
7683 /* always unregister in case we don't destroy any domains */
7684 unregister_sched_domain_sysctl();
7686 /* Let architecture update cpu core mappings. */
7687 new_topology = arch_update_cpu_topology();
7689 n = doms_new ? ndoms_new : 0;
7691 /* Destroy deleted domains */
7692 for (i = 0; i < ndoms_cur; i++) {
7693 for (j = 0; j < n && !new_topology; j++) {
7694 if (cpumask_equal(doms_cur[i], doms_new[j])
7695 && dattrs_equal(dattr_cur, i, dattr_new, j))
7696 goto match1;
7698 /* no match - a current sched domain not in new doms_new[] */
7699 detach_destroy_domains(doms_cur[i]);
7700 match1:
7704 if (doms_new == NULL) {
7705 ndoms_cur = 0;
7706 doms_new = &fallback_doms;
7707 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7708 WARN_ON_ONCE(dattr_new);
7711 /* Build new domains */
7712 for (i = 0; i < ndoms_new; i++) {
7713 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7714 if (cpumask_equal(doms_new[i], doms_cur[j])
7715 && dattrs_equal(dattr_new, i, dattr_cur, j))
7716 goto match2;
7718 /* no match - add a new doms_new */
7719 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7720 match2:
7724 /* Remember the new sched domains */
7725 if (doms_cur != &fallback_doms)
7726 free_sched_domains(doms_cur, ndoms_cur);
7727 kfree(dattr_cur); /* kfree(NULL) is safe */
7728 doms_cur = doms_new;
7729 dattr_cur = dattr_new;
7730 ndoms_cur = ndoms_new;
7732 register_sched_domain_sysctl();
7734 mutex_unlock(&sched_domains_mutex);
7737 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7738 static void reinit_sched_domains(void)
7740 get_online_cpus();
7742 /* Destroy domains first to force the rebuild */
7743 partition_sched_domains(0, NULL, NULL);
7745 rebuild_sched_domains();
7746 put_online_cpus();
7749 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7751 unsigned int level = 0;
7753 if (sscanf(buf, "%u", &level) != 1)
7754 return -EINVAL;
7757 * level is always be positive so don't check for
7758 * level < POWERSAVINGS_BALANCE_NONE which is 0
7759 * What happens on 0 or 1 byte write,
7760 * need to check for count as well?
7763 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7764 return -EINVAL;
7766 if (smt)
7767 sched_smt_power_savings = level;
7768 else
7769 sched_mc_power_savings = level;
7771 reinit_sched_domains();
7773 return count;
7776 #ifdef CONFIG_SCHED_MC
7777 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7778 struct sysdev_class_attribute *attr,
7779 char *page)
7781 return sprintf(page, "%u\n", sched_mc_power_savings);
7783 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7784 struct sysdev_class_attribute *attr,
7785 const char *buf, size_t count)
7787 return sched_power_savings_store(buf, count, 0);
7789 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7790 sched_mc_power_savings_show,
7791 sched_mc_power_savings_store);
7792 #endif
7794 #ifdef CONFIG_SCHED_SMT
7795 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7796 struct sysdev_class_attribute *attr,
7797 char *page)
7799 return sprintf(page, "%u\n", sched_smt_power_savings);
7801 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7802 struct sysdev_class_attribute *attr,
7803 const char *buf, size_t count)
7805 return sched_power_savings_store(buf, count, 1);
7807 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7808 sched_smt_power_savings_show,
7809 sched_smt_power_savings_store);
7810 #endif
7812 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7814 int err = 0;
7816 #ifdef CONFIG_SCHED_SMT
7817 if (smt_capable())
7818 err = sysfs_create_file(&cls->kset.kobj,
7819 &attr_sched_smt_power_savings.attr);
7820 #endif
7821 #ifdef CONFIG_SCHED_MC
7822 if (!err && mc_capable())
7823 err = sysfs_create_file(&cls->kset.kobj,
7824 &attr_sched_mc_power_savings.attr);
7825 #endif
7826 return err;
7828 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7831 * Update cpusets according to cpu_active mask. If cpusets are
7832 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7833 * around partition_sched_domains().
7835 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7836 void *hcpu)
7838 switch (action & ~CPU_TASKS_FROZEN) {
7839 case CPU_ONLINE:
7840 case CPU_DOWN_FAILED:
7841 cpuset_update_active_cpus();
7842 return NOTIFY_OK;
7843 default:
7844 return NOTIFY_DONE;
7848 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7849 void *hcpu)
7851 switch (action & ~CPU_TASKS_FROZEN) {
7852 case CPU_DOWN_PREPARE:
7853 cpuset_update_active_cpus();
7854 return NOTIFY_OK;
7855 default:
7856 return NOTIFY_DONE;
7860 static int update_runtime(struct notifier_block *nfb,
7861 unsigned long action, void *hcpu)
7863 int cpu = (int)(long)hcpu;
7865 switch (action) {
7866 case CPU_DOWN_PREPARE:
7867 case CPU_DOWN_PREPARE_FROZEN:
7868 disable_runtime(cpu_rq(cpu));
7869 return NOTIFY_OK;
7871 case CPU_DOWN_FAILED:
7872 case CPU_DOWN_FAILED_FROZEN:
7873 case CPU_ONLINE:
7874 case CPU_ONLINE_FROZEN:
7875 enable_runtime(cpu_rq(cpu));
7876 return NOTIFY_OK;
7878 default:
7879 return NOTIFY_DONE;
7883 void __init sched_init_smp(void)
7885 cpumask_var_t non_isolated_cpus;
7887 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7888 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7890 get_online_cpus();
7891 mutex_lock(&sched_domains_mutex);
7892 init_sched_domains(cpu_active_mask);
7893 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7894 if (cpumask_empty(non_isolated_cpus))
7895 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7896 mutex_unlock(&sched_domains_mutex);
7897 put_online_cpus();
7899 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7900 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7902 /* RT runtime code needs to handle some hotplug events */
7903 hotcpu_notifier(update_runtime, 0);
7905 init_hrtick();
7907 /* Move init over to a non-isolated CPU */
7908 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7909 BUG();
7910 sched_init_granularity();
7911 free_cpumask_var(non_isolated_cpus);
7913 init_sched_rt_class();
7915 #else
7916 void __init sched_init_smp(void)
7918 sched_init_granularity();
7920 #endif /* CONFIG_SMP */
7922 const_debug unsigned int sysctl_timer_migration = 1;
7924 int in_sched_functions(unsigned long addr)
7926 return in_lock_functions(addr) ||
7927 (addr >= (unsigned long)__sched_text_start
7928 && addr < (unsigned long)__sched_text_end);
7931 static void init_cfs_rq(struct cfs_rq *cfs_rq)
7933 cfs_rq->tasks_timeline = RB_ROOT;
7934 INIT_LIST_HEAD(&cfs_rq->tasks);
7935 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7936 #ifndef CONFIG_64BIT
7937 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7938 #endif
7941 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7943 struct rt_prio_array *array;
7944 int i;
7946 array = &rt_rq->active;
7947 for (i = 0; i < MAX_RT_PRIO; i++) {
7948 INIT_LIST_HEAD(array->queue + i);
7949 __clear_bit(i, array->bitmap);
7951 /* delimiter for bitsearch: */
7952 __set_bit(MAX_RT_PRIO, array->bitmap);
7954 #if defined CONFIG_SMP
7955 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7956 rt_rq->highest_prio.next = MAX_RT_PRIO;
7957 rt_rq->rt_nr_migratory = 0;
7958 rt_rq->overloaded = 0;
7959 plist_head_init(&rt_rq->pushable_tasks);
7960 #endif
7962 rt_rq->rt_time = 0;
7963 rt_rq->rt_throttled = 0;
7964 rt_rq->rt_runtime = 0;
7965 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7968 #ifdef CONFIG_FAIR_GROUP_SCHED
7969 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7970 struct sched_entity *se, int cpu,
7971 struct sched_entity *parent)
7973 struct rq *rq = cpu_rq(cpu);
7975 cfs_rq->tg = tg;
7976 cfs_rq->rq = rq;
7977 #ifdef CONFIG_SMP
7978 /* allow initial update_cfs_load() to truncate */
7979 cfs_rq->load_stamp = 1;
7980 #endif
7982 tg->cfs_rq[cpu] = cfs_rq;
7983 tg->se[cpu] = se;
7985 /* se could be NULL for root_task_group */
7986 if (!se)
7987 return;
7989 if (!parent)
7990 se->cfs_rq = &rq->cfs;
7991 else
7992 se->cfs_rq = parent->my_q;
7994 se->my_q = cfs_rq;
7995 update_load_set(&se->load, 0);
7996 se->parent = parent;
7998 #endif
8000 #ifdef CONFIG_RT_GROUP_SCHED
8001 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8002 struct sched_rt_entity *rt_se, int cpu,
8003 struct sched_rt_entity *parent)
8005 struct rq *rq = cpu_rq(cpu);
8007 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8008 rt_rq->rt_nr_boosted = 0;
8009 rt_rq->rq = rq;
8010 rt_rq->tg = tg;
8012 tg->rt_rq[cpu] = rt_rq;
8013 tg->rt_se[cpu] = rt_se;
8015 if (!rt_se)
8016 return;
8018 if (!parent)
8019 rt_se->rt_rq = &rq->rt;
8020 else
8021 rt_se->rt_rq = parent->my_q;
8023 rt_se->my_q = rt_rq;
8024 rt_se->parent = parent;
8025 INIT_LIST_HEAD(&rt_se->run_list);
8027 #endif
8029 void __init sched_init(void)
8031 int i, j;
8032 unsigned long alloc_size = 0, ptr;
8034 #ifdef CONFIG_FAIR_GROUP_SCHED
8035 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8036 #endif
8037 #ifdef CONFIG_RT_GROUP_SCHED
8038 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8039 #endif
8040 #ifdef CONFIG_CPUMASK_OFFSTACK
8041 alloc_size += num_possible_cpus() * cpumask_size();
8042 #endif
8043 if (alloc_size) {
8044 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8046 #ifdef CONFIG_FAIR_GROUP_SCHED
8047 root_task_group.se = (struct sched_entity **)ptr;
8048 ptr += nr_cpu_ids * sizeof(void **);
8050 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8051 ptr += nr_cpu_ids * sizeof(void **);
8053 #endif /* CONFIG_FAIR_GROUP_SCHED */
8054 #ifdef CONFIG_RT_GROUP_SCHED
8055 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8056 ptr += nr_cpu_ids * sizeof(void **);
8058 root_task_group.rt_rq = (struct rt_rq **)ptr;
8059 ptr += nr_cpu_ids * sizeof(void **);
8061 #endif /* CONFIG_RT_GROUP_SCHED */
8062 #ifdef CONFIG_CPUMASK_OFFSTACK
8063 for_each_possible_cpu(i) {
8064 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8065 ptr += cpumask_size();
8067 #endif /* CONFIG_CPUMASK_OFFSTACK */
8070 #ifdef CONFIG_SMP
8071 init_defrootdomain();
8072 #endif
8074 init_rt_bandwidth(&def_rt_bandwidth,
8075 global_rt_period(), global_rt_runtime());
8077 #ifdef CONFIG_RT_GROUP_SCHED
8078 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8079 global_rt_period(), global_rt_runtime());
8080 #endif /* CONFIG_RT_GROUP_SCHED */
8082 #ifdef CONFIG_CGROUP_SCHED
8083 list_add(&root_task_group.list, &task_groups);
8084 INIT_LIST_HEAD(&root_task_group.children);
8085 autogroup_init(&init_task);
8086 #endif /* CONFIG_CGROUP_SCHED */
8088 for_each_possible_cpu(i) {
8089 struct rq *rq;
8091 rq = cpu_rq(i);
8092 raw_spin_lock_init(&rq->lock);
8093 rq->nr_running = 0;
8094 rq->calc_load_active = 0;
8095 rq->calc_load_update = jiffies + LOAD_FREQ;
8096 init_cfs_rq(&rq->cfs);
8097 init_rt_rq(&rq->rt, rq);
8098 #ifdef CONFIG_FAIR_GROUP_SCHED
8099 root_task_group.shares = root_task_group_load;
8100 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8102 * How much cpu bandwidth does root_task_group get?
8104 * In case of task-groups formed thr' the cgroup filesystem, it
8105 * gets 100% of the cpu resources in the system. This overall
8106 * system cpu resource is divided among the tasks of
8107 * root_task_group and its child task-groups in a fair manner,
8108 * based on each entity's (task or task-group's) weight
8109 * (se->load.weight).
8111 * In other words, if root_task_group has 10 tasks of weight
8112 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8113 * then A0's share of the cpu resource is:
8115 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8117 * We achieve this by letting root_task_group's tasks sit
8118 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8120 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8121 #endif /* CONFIG_FAIR_GROUP_SCHED */
8123 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8124 #ifdef CONFIG_RT_GROUP_SCHED
8125 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8126 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8127 #endif
8129 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8130 rq->cpu_load[j] = 0;
8132 rq->last_load_update_tick = jiffies;
8134 #ifdef CONFIG_SMP
8135 rq->sd = NULL;
8136 rq->rd = NULL;
8137 rq->cpu_power = SCHED_POWER_SCALE;
8138 rq->post_schedule = 0;
8139 rq->active_balance = 0;
8140 rq->next_balance = jiffies;
8141 rq->push_cpu = 0;
8142 rq->cpu = i;
8143 rq->online = 0;
8144 rq->idle_stamp = 0;
8145 rq->avg_idle = 2*sysctl_sched_migration_cost;
8146 rq_attach_root(rq, &def_root_domain);
8147 #ifdef CONFIG_NO_HZ
8148 rq->nohz_balance_kick = 0;
8149 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8150 #endif
8151 #endif
8152 init_rq_hrtick(rq);
8153 atomic_set(&rq->nr_iowait, 0);
8156 set_load_weight(&init_task);
8158 #ifdef CONFIG_PREEMPT_NOTIFIERS
8159 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8160 #endif
8162 #ifdef CONFIG_SMP
8163 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8164 #endif
8166 #ifdef CONFIG_RT_MUTEXES
8167 plist_head_init(&init_task.pi_waiters);
8168 #endif
8171 * The boot idle thread does lazy MMU switching as well:
8173 atomic_inc(&init_mm.mm_count);
8174 enter_lazy_tlb(&init_mm, current);
8177 * Make us the idle thread. Technically, schedule() should not be
8178 * called from this thread, however somewhere below it might be,
8179 * but because we are the idle thread, we just pick up running again
8180 * when this runqueue becomes "idle".
8182 init_idle(current, smp_processor_id());
8184 calc_load_update = jiffies + LOAD_FREQ;
8187 * During early bootup we pretend to be a normal task:
8189 current->sched_class = &fair_sched_class;
8191 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8192 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8193 #ifdef CONFIG_SMP
8194 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8195 #ifdef CONFIG_NO_HZ
8196 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8197 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8198 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8199 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8200 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8201 #endif
8202 /* May be allocated at isolcpus cmdline parse time */
8203 if (cpu_isolated_map == NULL)
8204 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8205 #endif /* SMP */
8207 scheduler_running = 1;
8210 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8211 static inline int preempt_count_equals(int preempt_offset)
8213 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8215 return (nested == preempt_offset);
8218 void __might_sleep(const char *file, int line, int preempt_offset)
8220 static unsigned long prev_jiffy; /* ratelimiting */
8222 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8223 system_state != SYSTEM_RUNNING || oops_in_progress)
8224 return;
8225 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8226 return;
8227 prev_jiffy = jiffies;
8229 printk(KERN_ERR
8230 "BUG: sleeping function called from invalid context at %s:%d\n",
8231 file, line);
8232 printk(KERN_ERR
8233 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8234 in_atomic(), irqs_disabled(),
8235 current->pid, current->comm);
8237 debug_show_held_locks(current);
8238 if (irqs_disabled())
8239 print_irqtrace_events(current);
8240 dump_stack();
8242 EXPORT_SYMBOL(__might_sleep);
8243 #endif
8245 #ifdef CONFIG_MAGIC_SYSRQ
8246 static void normalize_task(struct rq *rq, struct task_struct *p)
8248 const struct sched_class *prev_class = p->sched_class;
8249 int old_prio = p->prio;
8250 int on_rq;
8252 on_rq = p->on_rq;
8253 if (on_rq)
8254 deactivate_task(rq, p, 0);
8255 __setscheduler(rq, p, SCHED_NORMAL, 0);
8256 if (on_rq) {
8257 activate_task(rq, p, 0);
8258 resched_task(rq->curr);
8261 check_class_changed(rq, p, prev_class, old_prio);
8264 void normalize_rt_tasks(void)
8266 struct task_struct *g, *p;
8267 unsigned long flags;
8268 struct rq *rq;
8270 read_lock_irqsave(&tasklist_lock, flags);
8271 do_each_thread(g, p) {
8273 * Only normalize user tasks:
8275 if (!p->mm)
8276 continue;
8278 p->se.exec_start = 0;
8279 #ifdef CONFIG_SCHEDSTATS
8280 p->se.statistics.wait_start = 0;
8281 p->se.statistics.sleep_start = 0;
8282 p->se.statistics.block_start = 0;
8283 #endif
8285 if (!rt_task(p)) {
8287 * Renice negative nice level userspace
8288 * tasks back to 0:
8290 if (TASK_NICE(p) < 0 && p->mm)
8291 set_user_nice(p, 0);
8292 continue;
8295 raw_spin_lock(&p->pi_lock);
8296 rq = __task_rq_lock(p);
8298 normalize_task(rq, p);
8300 __task_rq_unlock(rq);
8301 raw_spin_unlock(&p->pi_lock);
8302 } while_each_thread(g, p);
8304 read_unlock_irqrestore(&tasklist_lock, flags);
8307 #endif /* CONFIG_MAGIC_SYSRQ */
8309 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8311 * These functions are only useful for the IA64 MCA handling, or kdb.
8313 * They can only be called when the whole system has been
8314 * stopped - every CPU needs to be quiescent, and no scheduling
8315 * activity can take place. Using them for anything else would
8316 * be a serious bug, and as a result, they aren't even visible
8317 * under any other configuration.
8321 * curr_task - return the current task for a given cpu.
8322 * @cpu: the processor in question.
8324 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8326 struct task_struct *curr_task(int cpu)
8328 return cpu_curr(cpu);
8331 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8333 #ifdef CONFIG_IA64
8335 * set_curr_task - set the current task for a given cpu.
8336 * @cpu: the processor in question.
8337 * @p: the task pointer to set.
8339 * Description: This function must only be used when non-maskable interrupts
8340 * are serviced on a separate stack. It allows the architecture to switch the
8341 * notion of the current task on a cpu in a non-blocking manner. This function
8342 * must be called with all CPU's synchronized, and interrupts disabled, the
8343 * and caller must save the original value of the current task (see
8344 * curr_task() above) and restore that value before reenabling interrupts and
8345 * re-starting the system.
8347 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8349 void set_curr_task(int cpu, struct task_struct *p)
8351 cpu_curr(cpu) = p;
8354 #endif
8356 #ifdef CONFIG_FAIR_GROUP_SCHED
8357 static void free_fair_sched_group(struct task_group *tg)
8359 int i;
8361 for_each_possible_cpu(i) {
8362 if (tg->cfs_rq)
8363 kfree(tg->cfs_rq[i]);
8364 if (tg->se)
8365 kfree(tg->se[i]);
8368 kfree(tg->cfs_rq);
8369 kfree(tg->se);
8372 static
8373 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8375 struct cfs_rq *cfs_rq;
8376 struct sched_entity *se;
8377 int i;
8379 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8380 if (!tg->cfs_rq)
8381 goto err;
8382 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8383 if (!tg->se)
8384 goto err;
8386 tg->shares = NICE_0_LOAD;
8388 for_each_possible_cpu(i) {
8389 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8390 GFP_KERNEL, cpu_to_node(i));
8391 if (!cfs_rq)
8392 goto err;
8394 se = kzalloc_node(sizeof(struct sched_entity),
8395 GFP_KERNEL, cpu_to_node(i));
8396 if (!se)
8397 goto err_free_rq;
8399 init_cfs_rq(cfs_rq);
8400 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8403 return 1;
8405 err_free_rq:
8406 kfree(cfs_rq);
8407 err:
8408 return 0;
8411 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8413 struct rq *rq = cpu_rq(cpu);
8414 unsigned long flags;
8417 * Only empty task groups can be destroyed; so we can speculatively
8418 * check on_list without danger of it being re-added.
8420 if (!tg->cfs_rq[cpu]->on_list)
8421 return;
8423 raw_spin_lock_irqsave(&rq->lock, flags);
8424 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8425 raw_spin_unlock_irqrestore(&rq->lock, flags);
8427 #else /* !CONFIG_FAIR_GROUP_SCHED */
8428 static inline void free_fair_sched_group(struct task_group *tg)
8432 static inline
8433 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8435 return 1;
8438 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8441 #endif /* CONFIG_FAIR_GROUP_SCHED */
8443 #ifdef CONFIG_RT_GROUP_SCHED
8444 static void free_rt_sched_group(struct task_group *tg)
8446 int i;
8448 if (tg->rt_se)
8449 destroy_rt_bandwidth(&tg->rt_bandwidth);
8451 for_each_possible_cpu(i) {
8452 if (tg->rt_rq)
8453 kfree(tg->rt_rq[i]);
8454 if (tg->rt_se)
8455 kfree(tg->rt_se[i]);
8458 kfree(tg->rt_rq);
8459 kfree(tg->rt_se);
8462 static
8463 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8465 struct rt_rq *rt_rq;
8466 struct sched_rt_entity *rt_se;
8467 int i;
8469 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8470 if (!tg->rt_rq)
8471 goto err;
8472 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8473 if (!tg->rt_se)
8474 goto err;
8476 init_rt_bandwidth(&tg->rt_bandwidth,
8477 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8479 for_each_possible_cpu(i) {
8480 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8481 GFP_KERNEL, cpu_to_node(i));
8482 if (!rt_rq)
8483 goto err;
8485 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8486 GFP_KERNEL, cpu_to_node(i));
8487 if (!rt_se)
8488 goto err_free_rq;
8490 init_rt_rq(rt_rq, cpu_rq(i));
8491 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8492 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8495 return 1;
8497 err_free_rq:
8498 kfree(rt_rq);
8499 err:
8500 return 0;
8502 #else /* !CONFIG_RT_GROUP_SCHED */
8503 static inline void free_rt_sched_group(struct task_group *tg)
8507 static inline
8508 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8510 return 1;
8512 #endif /* CONFIG_RT_GROUP_SCHED */
8514 #ifdef CONFIG_CGROUP_SCHED
8515 static void free_sched_group(struct task_group *tg)
8517 free_fair_sched_group(tg);
8518 free_rt_sched_group(tg);
8519 autogroup_free(tg);
8520 kfree(tg);
8523 /* allocate runqueue etc for a new task group */
8524 struct task_group *sched_create_group(struct task_group *parent)
8526 struct task_group *tg;
8527 unsigned long flags;
8529 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8530 if (!tg)
8531 return ERR_PTR(-ENOMEM);
8533 if (!alloc_fair_sched_group(tg, parent))
8534 goto err;
8536 if (!alloc_rt_sched_group(tg, parent))
8537 goto err;
8539 spin_lock_irqsave(&task_group_lock, flags);
8540 list_add_rcu(&tg->list, &task_groups);
8542 WARN_ON(!parent); /* root should already exist */
8544 tg->parent = parent;
8545 INIT_LIST_HEAD(&tg->children);
8546 list_add_rcu(&tg->siblings, &parent->children);
8547 spin_unlock_irqrestore(&task_group_lock, flags);
8549 return tg;
8551 err:
8552 free_sched_group(tg);
8553 return ERR_PTR(-ENOMEM);
8556 /* rcu callback to free various structures associated with a task group */
8557 static void free_sched_group_rcu(struct rcu_head *rhp)
8559 /* now it should be safe to free those cfs_rqs */
8560 free_sched_group(container_of(rhp, struct task_group, rcu));
8563 /* Destroy runqueue etc associated with a task group */
8564 void sched_destroy_group(struct task_group *tg)
8566 unsigned long flags;
8567 int i;
8569 /* end participation in shares distribution */
8570 for_each_possible_cpu(i)
8571 unregister_fair_sched_group(tg, i);
8573 spin_lock_irqsave(&task_group_lock, flags);
8574 list_del_rcu(&tg->list);
8575 list_del_rcu(&tg->siblings);
8576 spin_unlock_irqrestore(&task_group_lock, flags);
8578 /* wait for possible concurrent references to cfs_rqs complete */
8579 call_rcu(&tg->rcu, free_sched_group_rcu);
8582 /* change task's runqueue when it moves between groups.
8583 * The caller of this function should have put the task in its new group
8584 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8585 * reflect its new group.
8587 void sched_move_task(struct task_struct *tsk)
8589 int on_rq, running;
8590 unsigned long flags;
8591 struct rq *rq;
8593 rq = task_rq_lock(tsk, &flags);
8595 running = task_current(rq, tsk);
8596 on_rq = tsk->on_rq;
8598 if (on_rq)
8599 dequeue_task(rq, tsk, 0);
8600 if (unlikely(running))
8601 tsk->sched_class->put_prev_task(rq, tsk);
8603 #ifdef CONFIG_FAIR_GROUP_SCHED
8604 if (tsk->sched_class->task_move_group)
8605 tsk->sched_class->task_move_group(tsk, on_rq);
8606 else
8607 #endif
8608 set_task_rq(tsk, task_cpu(tsk));
8610 if (unlikely(running))
8611 tsk->sched_class->set_curr_task(rq);
8612 if (on_rq)
8613 enqueue_task(rq, tsk, 0);
8615 task_rq_unlock(rq, tsk, &flags);
8617 #endif /* CONFIG_CGROUP_SCHED */
8619 #ifdef CONFIG_FAIR_GROUP_SCHED
8620 static DEFINE_MUTEX(shares_mutex);
8622 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8624 int i;
8625 unsigned long flags;
8628 * We can't change the weight of the root cgroup.
8630 if (!tg->se[0])
8631 return -EINVAL;
8633 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8635 mutex_lock(&shares_mutex);
8636 if (tg->shares == shares)
8637 goto done;
8639 tg->shares = shares;
8640 for_each_possible_cpu(i) {
8641 struct rq *rq = cpu_rq(i);
8642 struct sched_entity *se;
8644 se = tg->se[i];
8645 /* Propagate contribution to hierarchy */
8646 raw_spin_lock_irqsave(&rq->lock, flags);
8647 for_each_sched_entity(se)
8648 update_cfs_shares(group_cfs_rq(se));
8649 raw_spin_unlock_irqrestore(&rq->lock, flags);
8652 done:
8653 mutex_unlock(&shares_mutex);
8654 return 0;
8657 unsigned long sched_group_shares(struct task_group *tg)
8659 return tg->shares;
8661 #endif
8663 #ifdef CONFIG_RT_GROUP_SCHED
8665 * Ensure that the real time constraints are schedulable.
8667 static DEFINE_MUTEX(rt_constraints_mutex);
8669 static unsigned long to_ratio(u64 period, u64 runtime)
8671 if (runtime == RUNTIME_INF)
8672 return 1ULL << 20;
8674 return div64_u64(runtime << 20, period);
8677 /* Must be called with tasklist_lock held */
8678 static inline int tg_has_rt_tasks(struct task_group *tg)
8680 struct task_struct *g, *p;
8682 do_each_thread(g, p) {
8683 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8684 return 1;
8685 } while_each_thread(g, p);
8687 return 0;
8690 struct rt_schedulable_data {
8691 struct task_group *tg;
8692 u64 rt_period;
8693 u64 rt_runtime;
8696 static int tg_schedulable(struct task_group *tg, void *data)
8698 struct rt_schedulable_data *d = data;
8699 struct task_group *child;
8700 unsigned long total, sum = 0;
8701 u64 period, runtime;
8703 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8704 runtime = tg->rt_bandwidth.rt_runtime;
8706 if (tg == d->tg) {
8707 period = d->rt_period;
8708 runtime = d->rt_runtime;
8712 * Cannot have more runtime than the period.
8714 if (runtime > period && runtime != RUNTIME_INF)
8715 return -EINVAL;
8718 * Ensure we don't starve existing RT tasks.
8720 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8721 return -EBUSY;
8723 total = to_ratio(period, runtime);
8726 * Nobody can have more than the global setting allows.
8728 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8729 return -EINVAL;
8732 * The sum of our children's runtime should not exceed our own.
8734 list_for_each_entry_rcu(child, &tg->children, siblings) {
8735 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8736 runtime = child->rt_bandwidth.rt_runtime;
8738 if (child == d->tg) {
8739 period = d->rt_period;
8740 runtime = d->rt_runtime;
8743 sum += to_ratio(period, runtime);
8746 if (sum > total)
8747 return -EINVAL;
8749 return 0;
8752 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8754 struct rt_schedulable_data data = {
8755 .tg = tg,
8756 .rt_period = period,
8757 .rt_runtime = runtime,
8760 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8763 static int tg_set_bandwidth(struct task_group *tg,
8764 u64 rt_period, u64 rt_runtime)
8766 int i, err = 0;
8768 mutex_lock(&rt_constraints_mutex);
8769 read_lock(&tasklist_lock);
8770 err = __rt_schedulable(tg, rt_period, rt_runtime);
8771 if (err)
8772 goto unlock;
8774 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8775 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8776 tg->rt_bandwidth.rt_runtime = rt_runtime;
8778 for_each_possible_cpu(i) {
8779 struct rt_rq *rt_rq = tg->rt_rq[i];
8781 raw_spin_lock(&rt_rq->rt_runtime_lock);
8782 rt_rq->rt_runtime = rt_runtime;
8783 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8785 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8786 unlock:
8787 read_unlock(&tasklist_lock);
8788 mutex_unlock(&rt_constraints_mutex);
8790 return err;
8793 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8795 u64 rt_runtime, rt_period;
8797 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8798 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8799 if (rt_runtime_us < 0)
8800 rt_runtime = RUNTIME_INF;
8802 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8805 long sched_group_rt_runtime(struct task_group *tg)
8807 u64 rt_runtime_us;
8809 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8810 return -1;
8812 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8813 do_div(rt_runtime_us, NSEC_PER_USEC);
8814 return rt_runtime_us;
8817 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8819 u64 rt_runtime, rt_period;
8821 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8822 rt_runtime = tg->rt_bandwidth.rt_runtime;
8824 if (rt_period == 0)
8825 return -EINVAL;
8827 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8830 long sched_group_rt_period(struct task_group *tg)
8832 u64 rt_period_us;
8834 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8835 do_div(rt_period_us, NSEC_PER_USEC);
8836 return rt_period_us;
8839 static int sched_rt_global_constraints(void)
8841 u64 runtime, period;
8842 int ret = 0;
8844 if (sysctl_sched_rt_period <= 0)
8845 return -EINVAL;
8847 runtime = global_rt_runtime();
8848 period = global_rt_period();
8851 * Sanity check on the sysctl variables.
8853 if (runtime > period && runtime != RUNTIME_INF)
8854 return -EINVAL;
8856 mutex_lock(&rt_constraints_mutex);
8857 read_lock(&tasklist_lock);
8858 ret = __rt_schedulable(NULL, 0, 0);
8859 read_unlock(&tasklist_lock);
8860 mutex_unlock(&rt_constraints_mutex);
8862 return ret;
8865 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8867 /* Don't accept realtime tasks when there is no way for them to run */
8868 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8869 return 0;
8871 return 1;
8874 #else /* !CONFIG_RT_GROUP_SCHED */
8875 static int sched_rt_global_constraints(void)
8877 unsigned long flags;
8878 int i;
8880 if (sysctl_sched_rt_period <= 0)
8881 return -EINVAL;
8884 * There's always some RT tasks in the root group
8885 * -- migration, kstopmachine etc..
8887 if (sysctl_sched_rt_runtime == 0)
8888 return -EBUSY;
8890 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8891 for_each_possible_cpu(i) {
8892 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8894 raw_spin_lock(&rt_rq->rt_runtime_lock);
8895 rt_rq->rt_runtime = global_rt_runtime();
8896 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8898 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8900 return 0;
8902 #endif /* CONFIG_RT_GROUP_SCHED */
8904 int sched_rt_handler(struct ctl_table *table, int write,
8905 void __user *buffer, size_t *lenp,
8906 loff_t *ppos)
8908 int ret;
8909 int old_period, old_runtime;
8910 static DEFINE_MUTEX(mutex);
8912 mutex_lock(&mutex);
8913 old_period = sysctl_sched_rt_period;
8914 old_runtime = sysctl_sched_rt_runtime;
8916 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8918 if (!ret && write) {
8919 ret = sched_rt_global_constraints();
8920 if (ret) {
8921 sysctl_sched_rt_period = old_period;
8922 sysctl_sched_rt_runtime = old_runtime;
8923 } else {
8924 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8925 def_rt_bandwidth.rt_period =
8926 ns_to_ktime(global_rt_period());
8929 mutex_unlock(&mutex);
8931 return ret;
8934 #ifdef CONFIG_CGROUP_SCHED
8936 /* return corresponding task_group object of a cgroup */
8937 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8939 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8940 struct task_group, css);
8943 static struct cgroup_subsys_state *
8944 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8946 struct task_group *tg, *parent;
8948 if (!cgrp->parent) {
8949 /* This is early initialization for the top cgroup */
8950 return &root_task_group.css;
8953 parent = cgroup_tg(cgrp->parent);
8954 tg = sched_create_group(parent);
8955 if (IS_ERR(tg))
8956 return ERR_PTR(-ENOMEM);
8958 return &tg->css;
8961 static void
8962 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8964 struct task_group *tg = cgroup_tg(cgrp);
8966 sched_destroy_group(tg);
8969 static int
8970 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8972 #ifdef CONFIG_RT_GROUP_SCHED
8973 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8974 return -EINVAL;
8975 #else
8976 /* We don't support RT-tasks being in separate groups */
8977 if (tsk->sched_class != &fair_sched_class)
8978 return -EINVAL;
8979 #endif
8980 return 0;
8983 static void
8984 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8986 sched_move_task(tsk);
8989 static void
8990 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
8991 struct cgroup *old_cgrp, struct task_struct *task)
8994 * cgroup_exit() is called in the copy_process() failure path.
8995 * Ignore this case since the task hasn't ran yet, this avoids
8996 * trying to poke a half freed task state from generic code.
8998 if (!(task->flags & PF_EXITING))
8999 return;
9001 sched_move_task(task);
9004 #ifdef CONFIG_FAIR_GROUP_SCHED
9005 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9006 u64 shareval)
9008 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
9011 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9013 struct task_group *tg = cgroup_tg(cgrp);
9015 return (u64) scale_load_down(tg->shares);
9017 #endif /* CONFIG_FAIR_GROUP_SCHED */
9019 #ifdef CONFIG_RT_GROUP_SCHED
9020 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9021 s64 val)
9023 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9026 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9028 return sched_group_rt_runtime(cgroup_tg(cgrp));
9031 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9032 u64 rt_period_us)
9034 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9037 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9039 return sched_group_rt_period(cgroup_tg(cgrp));
9041 #endif /* CONFIG_RT_GROUP_SCHED */
9043 static struct cftype cpu_files[] = {
9044 #ifdef CONFIG_FAIR_GROUP_SCHED
9046 .name = "shares",
9047 .read_u64 = cpu_shares_read_u64,
9048 .write_u64 = cpu_shares_write_u64,
9050 #endif
9051 #ifdef CONFIG_RT_GROUP_SCHED
9053 .name = "rt_runtime_us",
9054 .read_s64 = cpu_rt_runtime_read,
9055 .write_s64 = cpu_rt_runtime_write,
9058 .name = "rt_period_us",
9059 .read_u64 = cpu_rt_period_read_uint,
9060 .write_u64 = cpu_rt_period_write_uint,
9062 #endif
9065 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9067 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9070 struct cgroup_subsys cpu_cgroup_subsys = {
9071 .name = "cpu",
9072 .create = cpu_cgroup_create,
9073 .destroy = cpu_cgroup_destroy,
9074 .can_attach_task = cpu_cgroup_can_attach_task,
9075 .attach_task = cpu_cgroup_attach_task,
9076 .exit = cpu_cgroup_exit,
9077 .populate = cpu_cgroup_populate,
9078 .subsys_id = cpu_cgroup_subsys_id,
9079 .early_init = 1,
9082 #endif /* CONFIG_CGROUP_SCHED */
9084 #ifdef CONFIG_CGROUP_CPUACCT
9087 * CPU accounting code for task groups.
9089 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9090 * (balbir@in.ibm.com).
9093 /* track cpu usage of a group of tasks and its child groups */
9094 struct cpuacct {
9095 struct cgroup_subsys_state css;
9096 /* cpuusage holds pointer to a u64-type object on every cpu */
9097 u64 __percpu *cpuusage;
9098 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9099 struct cpuacct *parent;
9102 struct cgroup_subsys cpuacct_subsys;
9104 /* return cpu accounting group corresponding to this container */
9105 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9107 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9108 struct cpuacct, css);
9111 /* return cpu accounting group to which this task belongs */
9112 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9114 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9115 struct cpuacct, css);
9118 /* create a new cpu accounting group */
9119 static struct cgroup_subsys_state *cpuacct_create(
9120 struct cgroup_subsys *ss, struct cgroup *cgrp)
9122 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9123 int i;
9125 if (!ca)
9126 goto out;
9128 ca->cpuusage = alloc_percpu(u64);
9129 if (!ca->cpuusage)
9130 goto out_free_ca;
9132 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9133 if (percpu_counter_init(&ca->cpustat[i], 0))
9134 goto out_free_counters;
9136 if (cgrp->parent)
9137 ca->parent = cgroup_ca(cgrp->parent);
9139 return &ca->css;
9141 out_free_counters:
9142 while (--i >= 0)
9143 percpu_counter_destroy(&ca->cpustat[i]);
9144 free_percpu(ca->cpuusage);
9145 out_free_ca:
9146 kfree(ca);
9147 out:
9148 return ERR_PTR(-ENOMEM);
9151 /* destroy an existing cpu accounting group */
9152 static void
9153 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9155 struct cpuacct *ca = cgroup_ca(cgrp);
9156 int i;
9158 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9159 percpu_counter_destroy(&ca->cpustat[i]);
9160 free_percpu(ca->cpuusage);
9161 kfree(ca);
9164 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9166 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9167 u64 data;
9169 #ifndef CONFIG_64BIT
9171 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9173 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9174 data = *cpuusage;
9175 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9176 #else
9177 data = *cpuusage;
9178 #endif
9180 return data;
9183 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9185 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9187 #ifndef CONFIG_64BIT
9189 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9191 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9192 *cpuusage = val;
9193 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9194 #else
9195 *cpuusage = val;
9196 #endif
9199 /* return total cpu usage (in nanoseconds) of a group */
9200 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9202 struct cpuacct *ca = cgroup_ca(cgrp);
9203 u64 totalcpuusage = 0;
9204 int i;
9206 for_each_present_cpu(i)
9207 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9209 return totalcpuusage;
9212 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9213 u64 reset)
9215 struct cpuacct *ca = cgroup_ca(cgrp);
9216 int err = 0;
9217 int i;
9219 if (reset) {
9220 err = -EINVAL;
9221 goto out;
9224 for_each_present_cpu(i)
9225 cpuacct_cpuusage_write(ca, i, 0);
9227 out:
9228 return err;
9231 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9232 struct seq_file *m)
9234 struct cpuacct *ca = cgroup_ca(cgroup);
9235 u64 percpu;
9236 int i;
9238 for_each_present_cpu(i) {
9239 percpu = cpuacct_cpuusage_read(ca, i);
9240 seq_printf(m, "%llu ", (unsigned long long) percpu);
9242 seq_printf(m, "\n");
9243 return 0;
9246 static const char *cpuacct_stat_desc[] = {
9247 [CPUACCT_STAT_USER] = "user",
9248 [CPUACCT_STAT_SYSTEM] = "system",
9251 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9252 struct cgroup_map_cb *cb)
9254 struct cpuacct *ca = cgroup_ca(cgrp);
9255 int i;
9257 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9258 s64 val = percpu_counter_read(&ca->cpustat[i]);
9259 val = cputime64_to_clock_t(val);
9260 cb->fill(cb, cpuacct_stat_desc[i], val);
9262 return 0;
9265 static struct cftype files[] = {
9267 .name = "usage",
9268 .read_u64 = cpuusage_read,
9269 .write_u64 = cpuusage_write,
9272 .name = "usage_percpu",
9273 .read_seq_string = cpuacct_percpu_seq_read,
9276 .name = "stat",
9277 .read_map = cpuacct_stats_show,
9281 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9283 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9287 * charge this task's execution time to its accounting group.
9289 * called with rq->lock held.
9291 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9293 struct cpuacct *ca;
9294 int cpu;
9296 if (unlikely(!cpuacct_subsys.active))
9297 return;
9299 cpu = task_cpu(tsk);
9301 rcu_read_lock();
9303 ca = task_ca(tsk);
9305 for (; ca; ca = ca->parent) {
9306 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9307 *cpuusage += cputime;
9310 rcu_read_unlock();
9314 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9315 * in cputime_t units. As a result, cpuacct_update_stats calls
9316 * percpu_counter_add with values large enough to always overflow the
9317 * per cpu batch limit causing bad SMP scalability.
9319 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9320 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9321 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9323 #ifdef CONFIG_SMP
9324 #define CPUACCT_BATCH \
9325 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9326 #else
9327 #define CPUACCT_BATCH 0
9328 #endif
9331 * Charge the system/user time to the task's accounting group.
9333 static void cpuacct_update_stats(struct task_struct *tsk,
9334 enum cpuacct_stat_index idx, cputime_t val)
9336 struct cpuacct *ca;
9337 int batch = CPUACCT_BATCH;
9339 if (unlikely(!cpuacct_subsys.active))
9340 return;
9342 rcu_read_lock();
9343 ca = task_ca(tsk);
9345 do {
9346 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9347 ca = ca->parent;
9348 } while (ca);
9349 rcu_read_unlock();
9352 struct cgroup_subsys cpuacct_subsys = {
9353 .name = "cpuacct",
9354 .create = cpuacct_create,
9355 .destroy = cpuacct_destroy,
9356 .populate = cpuacct_populate,
9357 .subsys_id = cpuacct_subsys_id,
9359 #endif /* CONFIG_CGROUP_CPUACCT */