Merge branch 'hwmon-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/groec...
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / kernel / sched.c
blob5e43e9dc65d1c197aa9b085a4ccf546551570177
1 /*
2 * kernel/sched.c
4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
11 * by Andrea Arcangeli
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
22 * by Peter Williams
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
29 #include <linux/mm.h>
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/stop_machine.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
75 #include <asm/tlb.h>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
79 #include "sched_cpupri.h"
80 #include "workqueue_sched.h"
81 #include "sched_autogroup.h"
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/sched.h>
87 * Convert user-nice values [ -20 ... 0 ... 19 ]
88 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 * and back.
91 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
92 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
93 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
96 * 'User priority' is the nice value converted to something we
97 * can work with better when scaling various scheduler parameters,
98 * it's a [ 0 ... 39 ] range.
100 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
101 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
102 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
105 * Helpers for converting nanosecond timing to jiffy resolution
107 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
109 #define NICE_0_LOAD SCHED_LOAD_SCALE
110 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
113 * These are the 'tuning knobs' of the scheduler:
115 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
116 * Timeslices get refilled after they expire.
118 #define DEF_TIMESLICE (100 * HZ / 1000)
121 * single value that denotes runtime == period, ie unlimited time.
123 #define RUNTIME_INF ((u64)~0ULL)
125 static inline int rt_policy(int policy)
127 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
128 return 1;
129 return 0;
132 static inline int task_has_rt_policy(struct task_struct *p)
134 return rt_policy(p->policy);
138 * This is the priority-queue data structure of the RT scheduling class:
140 struct rt_prio_array {
141 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
142 struct list_head queue[MAX_RT_PRIO];
145 struct rt_bandwidth {
146 /* nests inside the rq lock: */
147 raw_spinlock_t rt_runtime_lock;
148 ktime_t rt_period;
149 u64 rt_runtime;
150 struct hrtimer rt_period_timer;
153 static struct rt_bandwidth def_rt_bandwidth;
155 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
157 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
159 struct rt_bandwidth *rt_b =
160 container_of(timer, struct rt_bandwidth, rt_period_timer);
161 ktime_t now;
162 int overrun;
163 int idle = 0;
165 for (;;) {
166 now = hrtimer_cb_get_time(timer);
167 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169 if (!overrun)
170 break;
172 idle = do_sched_rt_period_timer(rt_b, overrun);
175 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
178 static
179 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
181 rt_b->rt_period = ns_to_ktime(period);
182 rt_b->rt_runtime = runtime;
184 raw_spin_lock_init(&rt_b->rt_runtime_lock);
186 hrtimer_init(&rt_b->rt_period_timer,
187 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
188 rt_b->rt_period_timer.function = sched_rt_period_timer;
191 static inline int rt_bandwidth_enabled(void)
193 return sysctl_sched_rt_runtime >= 0;
196 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
198 ktime_t now;
200 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
201 return;
203 if (hrtimer_active(&rt_b->rt_period_timer))
204 return;
206 raw_spin_lock(&rt_b->rt_runtime_lock);
207 for (;;) {
208 unsigned long delta;
209 ktime_t soft, hard;
211 if (hrtimer_active(&rt_b->rt_period_timer))
212 break;
214 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
215 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
217 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
218 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
219 delta = ktime_to_ns(ktime_sub(hard, soft));
220 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
221 HRTIMER_MODE_ABS_PINNED, 0);
223 raw_spin_unlock(&rt_b->rt_runtime_lock);
226 #ifdef CONFIG_RT_GROUP_SCHED
227 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
229 hrtimer_cancel(&rt_b->rt_period_timer);
231 #endif
234 * sched_domains_mutex serializes calls to init_sched_domains,
235 * detach_destroy_domains and partition_sched_domains.
237 static DEFINE_MUTEX(sched_domains_mutex);
239 #ifdef CONFIG_CGROUP_SCHED
241 #include <linux/cgroup.h>
243 struct cfs_rq;
245 static LIST_HEAD(task_groups);
247 /* task group related information */
248 struct task_group {
249 struct cgroup_subsys_state css;
251 #ifdef CONFIG_FAIR_GROUP_SCHED
252 /* schedulable entities of this group on each cpu */
253 struct sched_entity **se;
254 /* runqueue "owned" by this group on each cpu */
255 struct cfs_rq **cfs_rq;
256 unsigned long shares;
258 atomic_t load_weight;
259 #endif
261 #ifdef CONFIG_RT_GROUP_SCHED
262 struct sched_rt_entity **rt_se;
263 struct rt_rq **rt_rq;
265 struct rt_bandwidth rt_bandwidth;
266 #endif
268 struct rcu_head rcu;
269 struct list_head list;
271 struct task_group *parent;
272 struct list_head siblings;
273 struct list_head children;
275 #ifdef CONFIG_SCHED_AUTOGROUP
276 struct autogroup *autogroup;
277 #endif
280 /* task_group_lock serializes the addition/removal of task groups */
281 static DEFINE_SPINLOCK(task_group_lock);
283 #ifdef CONFIG_FAIR_GROUP_SCHED
285 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
288 * A weight of 0 or 1 can cause arithmetics problems.
289 * A weight of a cfs_rq is the sum of weights of which entities
290 * are queued on this cfs_rq, so a weight of a entity should not be
291 * too large, so as the shares value of a task group.
292 * (The default weight is 1024 - so there's no practical
293 * limitation from this.)
295 #define MIN_SHARES 2
296 #define MAX_SHARES (1UL << (18 + SCHED_LOAD_RESOLUTION))
298 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
299 #endif
301 /* Default task group.
302 * Every task in system belong to this group at bootup.
304 struct task_group root_task_group;
306 #endif /* CONFIG_CGROUP_SCHED */
308 /* CFS-related fields in a runqueue */
309 struct cfs_rq {
310 struct load_weight load;
311 unsigned long nr_running;
313 u64 exec_clock;
314 u64 min_vruntime;
315 #ifndef CONFIG_64BIT
316 u64 min_vruntime_copy;
317 #endif
319 struct rb_root tasks_timeline;
320 struct rb_node *rb_leftmost;
322 struct list_head tasks;
323 struct list_head *balance_iterator;
326 * 'curr' points to currently running entity on this cfs_rq.
327 * It is set to NULL otherwise (i.e when none are currently running).
329 struct sched_entity *curr, *next, *last, *skip;
331 #ifdef CONFIG_SCHED_DEBUG
332 unsigned int nr_spread_over;
333 #endif
335 #ifdef CONFIG_FAIR_GROUP_SCHED
336 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
339 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
340 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
341 * (like users, containers etc.)
343 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
344 * list is used during load balance.
346 int on_list;
347 struct list_head leaf_cfs_rq_list;
348 struct task_group *tg; /* group that "owns" this runqueue */
350 #ifdef CONFIG_SMP
352 * the part of load.weight contributed by tasks
354 unsigned long task_weight;
357 * h_load = weight * f(tg)
359 * Where f(tg) is the recursive weight fraction assigned to
360 * this group.
362 unsigned long h_load;
365 * Maintaining per-cpu shares distribution for group scheduling
367 * load_stamp is the last time we updated the load average
368 * load_last is the last time we updated the load average and saw load
369 * load_unacc_exec_time is currently unaccounted execution time
371 u64 load_avg;
372 u64 load_period;
373 u64 load_stamp, load_last, load_unacc_exec_time;
375 unsigned long load_contribution;
376 #endif
377 #endif
380 /* Real-Time classes' related field in a runqueue: */
381 struct rt_rq {
382 struct rt_prio_array active;
383 unsigned long rt_nr_running;
384 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
385 struct {
386 int curr; /* highest queued rt task prio */
387 #ifdef CONFIG_SMP
388 int next; /* next highest */
389 #endif
390 } highest_prio;
391 #endif
392 #ifdef CONFIG_SMP
393 unsigned long rt_nr_migratory;
394 unsigned long rt_nr_total;
395 int overloaded;
396 struct plist_head pushable_tasks;
397 #endif
398 int rt_throttled;
399 u64 rt_time;
400 u64 rt_runtime;
401 /* Nests inside the rq lock: */
402 raw_spinlock_t rt_runtime_lock;
404 #ifdef CONFIG_RT_GROUP_SCHED
405 unsigned long rt_nr_boosted;
407 struct rq *rq;
408 struct list_head leaf_rt_rq_list;
409 struct task_group *tg;
410 #endif
413 #ifdef CONFIG_SMP
416 * We add the notion of a root-domain which will be used to define per-domain
417 * variables. Each exclusive cpuset essentially defines an island domain by
418 * fully partitioning the member cpus from any other cpuset. Whenever a new
419 * exclusive cpuset is created, we also create and attach a new root-domain
420 * object.
423 struct root_domain {
424 atomic_t refcount;
425 struct rcu_head rcu;
426 cpumask_var_t span;
427 cpumask_var_t online;
430 * The "RT overload" flag: it gets set if a CPU has more than
431 * one runnable RT task.
433 cpumask_var_t rto_mask;
434 atomic_t rto_count;
435 struct cpupri cpupri;
439 * By default the system creates a single root-domain with all cpus as
440 * members (mimicking the global state we have today).
442 static struct root_domain def_root_domain;
444 #endif /* CONFIG_SMP */
447 * This is the main, per-CPU runqueue data structure.
449 * Locking rule: those places that want to lock multiple runqueues
450 * (such as the load balancing or the thread migration code), lock
451 * acquire operations must be ordered by ascending &runqueue.
453 struct rq {
454 /* runqueue lock: */
455 raw_spinlock_t lock;
458 * nr_running and cpu_load should be in the same cacheline because
459 * remote CPUs use both these fields when doing load calculation.
461 unsigned long nr_running;
462 #define CPU_LOAD_IDX_MAX 5
463 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
464 unsigned long last_load_update_tick;
465 #ifdef CONFIG_NO_HZ
466 u64 nohz_stamp;
467 unsigned char nohz_balance_kick;
468 #endif
469 int skip_clock_update;
471 /* capture load from *all* tasks on this cpu: */
472 struct load_weight load;
473 unsigned long nr_load_updates;
474 u64 nr_switches;
476 struct cfs_rq cfs;
477 struct rt_rq rt;
479 #ifdef CONFIG_FAIR_GROUP_SCHED
480 /* list of leaf cfs_rq on this cpu: */
481 struct list_head leaf_cfs_rq_list;
482 #endif
483 #ifdef CONFIG_RT_GROUP_SCHED
484 struct list_head leaf_rt_rq_list;
485 #endif
488 * This is part of a global counter where only the total sum
489 * over all CPUs matters. A task can increase this counter on
490 * one CPU and if it got migrated afterwards it may decrease
491 * it on another CPU. Always updated under the runqueue lock:
493 unsigned long nr_uninterruptible;
495 struct task_struct *curr, *idle, *stop;
496 unsigned long next_balance;
497 struct mm_struct *prev_mm;
499 u64 clock;
500 u64 clock_task;
502 atomic_t nr_iowait;
504 #ifdef CONFIG_SMP
505 struct root_domain *rd;
506 struct sched_domain *sd;
508 unsigned long cpu_power;
510 unsigned char idle_at_tick;
511 /* For active balancing */
512 int post_schedule;
513 int active_balance;
514 int push_cpu;
515 struct cpu_stop_work active_balance_work;
516 /* cpu of this runqueue: */
517 int cpu;
518 int online;
520 unsigned long avg_load_per_task;
522 u64 rt_avg;
523 u64 age_stamp;
524 u64 idle_stamp;
525 u64 avg_idle;
526 #endif
528 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
529 u64 prev_irq_time;
530 #endif
532 /* calc_load related fields */
533 unsigned long calc_load_update;
534 long calc_load_active;
536 #ifdef CONFIG_SCHED_HRTICK
537 #ifdef CONFIG_SMP
538 int hrtick_csd_pending;
539 struct call_single_data hrtick_csd;
540 #endif
541 struct hrtimer hrtick_timer;
542 #endif
544 #ifdef CONFIG_SCHEDSTATS
545 /* latency stats */
546 struct sched_info rq_sched_info;
547 unsigned long long rq_cpu_time;
548 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
550 /* sys_sched_yield() stats */
551 unsigned int yld_count;
553 /* schedule() stats */
554 unsigned int sched_switch;
555 unsigned int sched_count;
556 unsigned int sched_goidle;
558 /* try_to_wake_up() stats */
559 unsigned int ttwu_count;
560 unsigned int ttwu_local;
561 #endif
563 #ifdef CONFIG_SMP
564 struct task_struct *wake_list;
565 #endif
568 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
571 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
573 static inline int cpu_of(struct rq *rq)
575 #ifdef CONFIG_SMP
576 return rq->cpu;
577 #else
578 return 0;
579 #endif
582 #define rcu_dereference_check_sched_domain(p) \
583 rcu_dereference_check((p), \
584 rcu_read_lock_held() || \
585 lockdep_is_held(&sched_domains_mutex))
588 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
589 * See detach_destroy_domains: synchronize_sched for details.
591 * The domain tree of any CPU may only be accessed from within
592 * preempt-disabled sections.
594 #define for_each_domain(cpu, __sd) \
595 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
597 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
598 #define this_rq() (&__get_cpu_var(runqueues))
599 #define task_rq(p) cpu_rq(task_cpu(p))
600 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
601 #define raw_rq() (&__raw_get_cpu_var(runqueues))
603 #ifdef CONFIG_CGROUP_SCHED
606 * Return the group to which this tasks belongs.
608 * We use task_subsys_state_check() and extend the RCU verification
609 * with lockdep_is_held(&p->pi_lock) because cpu_cgroup_attach()
610 * holds that lock for each task it moves into the cgroup. Therefore
611 * by holding that lock, we pin the task to the current cgroup.
613 static inline struct task_group *task_group(struct task_struct *p)
615 struct task_group *tg;
616 struct cgroup_subsys_state *css;
618 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
619 lockdep_is_held(&p->pi_lock));
620 tg = container_of(css, struct task_group, css);
622 return autogroup_task_group(p, tg);
625 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
626 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
628 #ifdef CONFIG_FAIR_GROUP_SCHED
629 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
630 p->se.parent = task_group(p)->se[cpu];
631 #endif
633 #ifdef CONFIG_RT_GROUP_SCHED
634 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
635 p->rt.parent = task_group(p)->rt_se[cpu];
636 #endif
639 #else /* CONFIG_CGROUP_SCHED */
641 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
642 static inline struct task_group *task_group(struct task_struct *p)
644 return NULL;
647 #endif /* CONFIG_CGROUP_SCHED */
649 static void update_rq_clock_task(struct rq *rq, s64 delta);
651 static void update_rq_clock(struct rq *rq)
653 s64 delta;
655 if (rq->skip_clock_update > 0)
656 return;
658 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
659 rq->clock += delta;
660 update_rq_clock_task(rq, delta);
664 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
666 #ifdef CONFIG_SCHED_DEBUG
667 # define const_debug __read_mostly
668 #else
669 # define const_debug static const
670 #endif
673 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
674 * @cpu: the processor in question.
676 * This interface allows printk to be called with the runqueue lock
677 * held and know whether or not it is OK to wake up the klogd.
679 int runqueue_is_locked(int cpu)
681 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
685 * Debugging: various feature bits
688 #define SCHED_FEAT(name, enabled) \
689 __SCHED_FEAT_##name ,
691 enum {
692 #include "sched_features.h"
695 #undef SCHED_FEAT
697 #define SCHED_FEAT(name, enabled) \
698 (1UL << __SCHED_FEAT_##name) * enabled |
700 const_debug unsigned int sysctl_sched_features =
701 #include "sched_features.h"
704 #undef SCHED_FEAT
706 #ifdef CONFIG_SCHED_DEBUG
707 #define SCHED_FEAT(name, enabled) \
708 #name ,
710 static __read_mostly char *sched_feat_names[] = {
711 #include "sched_features.h"
712 NULL
715 #undef SCHED_FEAT
717 static int sched_feat_show(struct seq_file *m, void *v)
719 int i;
721 for (i = 0; sched_feat_names[i]; i++) {
722 if (!(sysctl_sched_features & (1UL << i)))
723 seq_puts(m, "NO_");
724 seq_printf(m, "%s ", sched_feat_names[i]);
726 seq_puts(m, "\n");
728 return 0;
731 static ssize_t
732 sched_feat_write(struct file *filp, const char __user *ubuf,
733 size_t cnt, loff_t *ppos)
735 char buf[64];
736 char *cmp;
737 int neg = 0;
738 int i;
740 if (cnt > 63)
741 cnt = 63;
743 if (copy_from_user(&buf, ubuf, cnt))
744 return -EFAULT;
746 buf[cnt] = 0;
747 cmp = strstrip(buf);
749 if (strncmp(cmp, "NO_", 3) == 0) {
750 neg = 1;
751 cmp += 3;
754 for (i = 0; sched_feat_names[i]; i++) {
755 if (strcmp(cmp, sched_feat_names[i]) == 0) {
756 if (neg)
757 sysctl_sched_features &= ~(1UL << i);
758 else
759 sysctl_sched_features |= (1UL << i);
760 break;
764 if (!sched_feat_names[i])
765 return -EINVAL;
767 *ppos += cnt;
769 return cnt;
772 static int sched_feat_open(struct inode *inode, struct file *filp)
774 return single_open(filp, sched_feat_show, NULL);
777 static const struct file_operations sched_feat_fops = {
778 .open = sched_feat_open,
779 .write = sched_feat_write,
780 .read = seq_read,
781 .llseek = seq_lseek,
782 .release = single_release,
785 static __init int sched_init_debug(void)
787 debugfs_create_file("sched_features", 0644, NULL, NULL,
788 &sched_feat_fops);
790 return 0;
792 late_initcall(sched_init_debug);
794 #endif
796 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
799 * Number of tasks to iterate in a single balance run.
800 * Limited because this is done with IRQs disabled.
802 const_debug unsigned int sysctl_sched_nr_migrate = 32;
805 * period over which we average the RT time consumption, measured
806 * in ms.
808 * default: 1s
810 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
813 * period over which we measure -rt task cpu usage in us.
814 * default: 1s
816 unsigned int sysctl_sched_rt_period = 1000000;
818 static __read_mostly int scheduler_running;
821 * part of the period that we allow rt tasks to run in us.
822 * default: 0.95s
824 int sysctl_sched_rt_runtime = 950000;
826 static inline u64 global_rt_period(void)
828 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
831 static inline u64 global_rt_runtime(void)
833 if (sysctl_sched_rt_runtime < 0)
834 return RUNTIME_INF;
836 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
839 #ifndef prepare_arch_switch
840 # define prepare_arch_switch(next) do { } while (0)
841 #endif
842 #ifndef finish_arch_switch
843 # define finish_arch_switch(prev) do { } while (0)
844 #endif
846 static inline int task_current(struct rq *rq, struct task_struct *p)
848 return rq->curr == p;
851 static inline int task_running(struct rq *rq, struct task_struct *p)
853 #ifdef CONFIG_SMP
854 return p->on_cpu;
855 #else
856 return task_current(rq, p);
857 #endif
860 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
861 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
863 #ifdef CONFIG_SMP
865 * We can optimise this out completely for !SMP, because the
866 * SMP rebalancing from interrupt is the only thing that cares
867 * here.
869 next->on_cpu = 1;
870 #endif
873 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
875 #ifdef CONFIG_SMP
877 * After ->on_cpu is cleared, the task can be moved to a different CPU.
878 * We must ensure this doesn't happen until the switch is completely
879 * finished.
881 smp_wmb();
882 prev->on_cpu = 0;
883 #endif
884 #ifdef CONFIG_DEBUG_SPINLOCK
885 /* this is a valid case when another task releases the spinlock */
886 rq->lock.owner = current;
887 #endif
889 * If we are tracking spinlock dependencies then we have to
890 * fix up the runqueue lock - which gets 'carried over' from
891 * prev into current:
893 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
895 raw_spin_unlock_irq(&rq->lock);
898 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
899 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
901 #ifdef CONFIG_SMP
903 * We can optimise this out completely for !SMP, because the
904 * SMP rebalancing from interrupt is the only thing that cares
905 * here.
907 next->on_cpu = 1;
908 #endif
909 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
910 raw_spin_unlock_irq(&rq->lock);
911 #else
912 raw_spin_unlock(&rq->lock);
913 #endif
916 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
918 #ifdef CONFIG_SMP
920 * After ->on_cpu is cleared, the task can be moved to a different CPU.
921 * We must ensure this doesn't happen until the switch is completely
922 * finished.
924 smp_wmb();
925 prev->on_cpu = 0;
926 #endif
927 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
928 local_irq_enable();
929 #endif
931 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
934 * __task_rq_lock - lock the rq @p resides on.
936 static inline struct rq *__task_rq_lock(struct task_struct *p)
937 __acquires(rq->lock)
939 struct rq *rq;
941 lockdep_assert_held(&p->pi_lock);
943 for (;;) {
944 rq = task_rq(p);
945 raw_spin_lock(&rq->lock);
946 if (likely(rq == task_rq(p)))
947 return rq;
948 raw_spin_unlock(&rq->lock);
953 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
955 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
956 __acquires(p->pi_lock)
957 __acquires(rq->lock)
959 struct rq *rq;
961 for (;;) {
962 raw_spin_lock_irqsave(&p->pi_lock, *flags);
963 rq = task_rq(p);
964 raw_spin_lock(&rq->lock);
965 if (likely(rq == task_rq(p)))
966 return rq;
967 raw_spin_unlock(&rq->lock);
968 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
972 static void __task_rq_unlock(struct rq *rq)
973 __releases(rq->lock)
975 raw_spin_unlock(&rq->lock);
978 static inline void
979 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
980 __releases(rq->lock)
981 __releases(p->pi_lock)
983 raw_spin_unlock(&rq->lock);
984 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
988 * this_rq_lock - lock this runqueue and disable interrupts.
990 static struct rq *this_rq_lock(void)
991 __acquires(rq->lock)
993 struct rq *rq;
995 local_irq_disable();
996 rq = this_rq();
997 raw_spin_lock(&rq->lock);
999 return rq;
1002 #ifdef CONFIG_SCHED_HRTICK
1004 * Use HR-timers to deliver accurate preemption points.
1006 * Its all a bit involved since we cannot program an hrt while holding the
1007 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1008 * reschedule event.
1010 * When we get rescheduled we reprogram the hrtick_timer outside of the
1011 * rq->lock.
1015 * Use hrtick when:
1016 * - enabled by features
1017 * - hrtimer is actually high res
1019 static inline int hrtick_enabled(struct rq *rq)
1021 if (!sched_feat(HRTICK))
1022 return 0;
1023 if (!cpu_active(cpu_of(rq)))
1024 return 0;
1025 return hrtimer_is_hres_active(&rq->hrtick_timer);
1028 static void hrtick_clear(struct rq *rq)
1030 if (hrtimer_active(&rq->hrtick_timer))
1031 hrtimer_cancel(&rq->hrtick_timer);
1035 * High-resolution timer tick.
1036 * Runs from hardirq context with interrupts disabled.
1038 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1040 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1042 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1044 raw_spin_lock(&rq->lock);
1045 update_rq_clock(rq);
1046 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1047 raw_spin_unlock(&rq->lock);
1049 return HRTIMER_NORESTART;
1052 #ifdef CONFIG_SMP
1054 * called from hardirq (IPI) context
1056 static void __hrtick_start(void *arg)
1058 struct rq *rq = arg;
1060 raw_spin_lock(&rq->lock);
1061 hrtimer_restart(&rq->hrtick_timer);
1062 rq->hrtick_csd_pending = 0;
1063 raw_spin_unlock(&rq->lock);
1067 * Called to set the hrtick timer state.
1069 * called with rq->lock held and irqs disabled
1071 static void hrtick_start(struct rq *rq, u64 delay)
1073 struct hrtimer *timer = &rq->hrtick_timer;
1074 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1076 hrtimer_set_expires(timer, time);
1078 if (rq == this_rq()) {
1079 hrtimer_restart(timer);
1080 } else if (!rq->hrtick_csd_pending) {
1081 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1082 rq->hrtick_csd_pending = 1;
1086 static int
1087 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1089 int cpu = (int)(long)hcpu;
1091 switch (action) {
1092 case CPU_UP_CANCELED:
1093 case CPU_UP_CANCELED_FROZEN:
1094 case CPU_DOWN_PREPARE:
1095 case CPU_DOWN_PREPARE_FROZEN:
1096 case CPU_DEAD:
1097 case CPU_DEAD_FROZEN:
1098 hrtick_clear(cpu_rq(cpu));
1099 return NOTIFY_OK;
1102 return NOTIFY_DONE;
1105 static __init void init_hrtick(void)
1107 hotcpu_notifier(hotplug_hrtick, 0);
1109 #else
1111 * Called to set the hrtick timer state.
1113 * called with rq->lock held and irqs disabled
1115 static void hrtick_start(struct rq *rq, u64 delay)
1117 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1118 HRTIMER_MODE_REL_PINNED, 0);
1121 static inline void init_hrtick(void)
1124 #endif /* CONFIG_SMP */
1126 static void init_rq_hrtick(struct rq *rq)
1128 #ifdef CONFIG_SMP
1129 rq->hrtick_csd_pending = 0;
1131 rq->hrtick_csd.flags = 0;
1132 rq->hrtick_csd.func = __hrtick_start;
1133 rq->hrtick_csd.info = rq;
1134 #endif
1136 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1137 rq->hrtick_timer.function = hrtick;
1139 #else /* CONFIG_SCHED_HRTICK */
1140 static inline void hrtick_clear(struct rq *rq)
1144 static inline void init_rq_hrtick(struct rq *rq)
1148 static inline void init_hrtick(void)
1151 #endif /* CONFIG_SCHED_HRTICK */
1154 * resched_task - mark a task 'to be rescheduled now'.
1156 * On UP this means the setting of the need_resched flag, on SMP it
1157 * might also involve a cross-CPU call to trigger the scheduler on
1158 * the target CPU.
1160 #ifdef CONFIG_SMP
1162 #ifndef tsk_is_polling
1163 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1164 #endif
1166 static void resched_task(struct task_struct *p)
1168 int cpu;
1170 assert_raw_spin_locked(&task_rq(p)->lock);
1172 if (test_tsk_need_resched(p))
1173 return;
1175 set_tsk_need_resched(p);
1177 cpu = task_cpu(p);
1178 if (cpu == smp_processor_id())
1179 return;
1181 /* NEED_RESCHED must be visible before we test polling */
1182 smp_mb();
1183 if (!tsk_is_polling(p))
1184 smp_send_reschedule(cpu);
1187 static void resched_cpu(int cpu)
1189 struct rq *rq = cpu_rq(cpu);
1190 unsigned long flags;
1192 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1193 return;
1194 resched_task(cpu_curr(cpu));
1195 raw_spin_unlock_irqrestore(&rq->lock, flags);
1198 #ifdef CONFIG_NO_HZ
1200 * In the semi idle case, use the nearest busy cpu for migrating timers
1201 * from an idle cpu. This is good for power-savings.
1203 * We don't do similar optimization for completely idle system, as
1204 * selecting an idle cpu will add more delays to the timers than intended
1205 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1207 int get_nohz_timer_target(void)
1209 int cpu = smp_processor_id();
1210 int i;
1211 struct sched_domain *sd;
1213 rcu_read_lock();
1214 for_each_domain(cpu, sd) {
1215 for_each_cpu(i, sched_domain_span(sd)) {
1216 if (!idle_cpu(i)) {
1217 cpu = i;
1218 goto unlock;
1222 unlock:
1223 rcu_read_unlock();
1224 return cpu;
1227 * When add_timer_on() enqueues a timer into the timer wheel of an
1228 * idle CPU then this timer might expire before the next timer event
1229 * which is scheduled to wake up that CPU. In case of a completely
1230 * idle system the next event might even be infinite time into the
1231 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1232 * leaves the inner idle loop so the newly added timer is taken into
1233 * account when the CPU goes back to idle and evaluates the timer
1234 * wheel for the next timer event.
1236 void wake_up_idle_cpu(int cpu)
1238 struct rq *rq = cpu_rq(cpu);
1240 if (cpu == smp_processor_id())
1241 return;
1244 * This is safe, as this function is called with the timer
1245 * wheel base lock of (cpu) held. When the CPU is on the way
1246 * to idle and has not yet set rq->curr to idle then it will
1247 * be serialized on the timer wheel base lock and take the new
1248 * timer into account automatically.
1250 if (rq->curr != rq->idle)
1251 return;
1254 * We can set TIF_RESCHED on the idle task of the other CPU
1255 * lockless. The worst case is that the other CPU runs the
1256 * idle task through an additional NOOP schedule()
1258 set_tsk_need_resched(rq->idle);
1260 /* NEED_RESCHED must be visible before we test polling */
1261 smp_mb();
1262 if (!tsk_is_polling(rq->idle))
1263 smp_send_reschedule(cpu);
1266 #endif /* CONFIG_NO_HZ */
1268 static u64 sched_avg_period(void)
1270 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1273 static void sched_avg_update(struct rq *rq)
1275 s64 period = sched_avg_period();
1277 while ((s64)(rq->clock - rq->age_stamp) > period) {
1279 * Inline assembly required to prevent the compiler
1280 * optimising this loop into a divmod call.
1281 * See __iter_div_u64_rem() for another example of this.
1283 asm("" : "+rm" (rq->age_stamp));
1284 rq->age_stamp += period;
1285 rq->rt_avg /= 2;
1289 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1291 rq->rt_avg += rt_delta;
1292 sched_avg_update(rq);
1295 #else /* !CONFIG_SMP */
1296 static void resched_task(struct task_struct *p)
1298 assert_raw_spin_locked(&task_rq(p)->lock);
1299 set_tsk_need_resched(p);
1302 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1306 static void sched_avg_update(struct rq *rq)
1309 #endif /* CONFIG_SMP */
1311 #if BITS_PER_LONG == 32
1312 # define WMULT_CONST (~0UL)
1313 #else
1314 # define WMULT_CONST (1UL << 32)
1315 #endif
1317 #define WMULT_SHIFT 32
1320 * Shift right and round:
1322 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1325 * delta *= weight / lw
1327 static unsigned long
1328 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1329 struct load_weight *lw)
1331 u64 tmp;
1334 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1335 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1336 * 2^SCHED_LOAD_RESOLUTION.
1338 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1339 tmp = (u64)delta_exec * scale_load_down(weight);
1340 else
1341 tmp = (u64)delta_exec;
1343 if (!lw->inv_weight) {
1344 unsigned long w = scale_load_down(lw->weight);
1346 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1347 lw->inv_weight = 1;
1348 else if (unlikely(!w))
1349 lw->inv_weight = WMULT_CONST;
1350 else
1351 lw->inv_weight = WMULT_CONST / w;
1355 * Check whether we'd overflow the 64-bit multiplication:
1357 if (unlikely(tmp > WMULT_CONST))
1358 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1359 WMULT_SHIFT/2);
1360 else
1361 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1363 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1366 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1368 lw->weight += inc;
1369 lw->inv_weight = 0;
1372 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1374 lw->weight -= dec;
1375 lw->inv_weight = 0;
1378 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1380 lw->weight = w;
1381 lw->inv_weight = 0;
1385 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1386 * of tasks with abnormal "nice" values across CPUs the contribution that
1387 * each task makes to its run queue's load is weighted according to its
1388 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1389 * scaled version of the new time slice allocation that they receive on time
1390 * slice expiry etc.
1393 #define WEIGHT_IDLEPRIO 3
1394 #define WMULT_IDLEPRIO 1431655765
1397 * Nice levels are multiplicative, with a gentle 10% change for every
1398 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1399 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1400 * that remained on nice 0.
1402 * The "10% effect" is relative and cumulative: from _any_ nice level,
1403 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1404 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1405 * If a task goes up by ~10% and another task goes down by ~10% then
1406 * the relative distance between them is ~25%.)
1408 static const int prio_to_weight[40] = {
1409 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1410 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1411 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1412 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1413 /* 0 */ 1024, 820, 655, 526, 423,
1414 /* 5 */ 335, 272, 215, 172, 137,
1415 /* 10 */ 110, 87, 70, 56, 45,
1416 /* 15 */ 36, 29, 23, 18, 15,
1420 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1422 * In cases where the weight does not change often, we can use the
1423 * precalculated inverse to speed up arithmetics by turning divisions
1424 * into multiplications:
1426 static const u32 prio_to_wmult[40] = {
1427 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1428 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1429 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1430 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1431 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1432 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1433 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1434 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1437 /* Time spent by the tasks of the cpu accounting group executing in ... */
1438 enum cpuacct_stat_index {
1439 CPUACCT_STAT_USER, /* ... user mode */
1440 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1442 CPUACCT_STAT_NSTATS,
1445 #ifdef CONFIG_CGROUP_CPUACCT
1446 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1447 static void cpuacct_update_stats(struct task_struct *tsk,
1448 enum cpuacct_stat_index idx, cputime_t val);
1449 #else
1450 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1451 static inline void cpuacct_update_stats(struct task_struct *tsk,
1452 enum cpuacct_stat_index idx, cputime_t val) {}
1453 #endif
1455 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1457 update_load_add(&rq->load, load);
1460 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1462 update_load_sub(&rq->load, load);
1465 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1466 typedef int (*tg_visitor)(struct task_group *, void *);
1469 * Iterate the full tree, calling @down when first entering a node and @up when
1470 * leaving it for the final time.
1472 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1474 struct task_group *parent, *child;
1475 int ret;
1477 rcu_read_lock();
1478 parent = &root_task_group;
1479 down:
1480 ret = (*down)(parent, data);
1481 if (ret)
1482 goto out_unlock;
1483 list_for_each_entry_rcu(child, &parent->children, siblings) {
1484 parent = child;
1485 goto down;
1488 continue;
1490 ret = (*up)(parent, data);
1491 if (ret)
1492 goto out_unlock;
1494 child = parent;
1495 parent = parent->parent;
1496 if (parent)
1497 goto up;
1498 out_unlock:
1499 rcu_read_unlock();
1501 return ret;
1504 static int tg_nop(struct task_group *tg, void *data)
1506 return 0;
1508 #endif
1510 #ifdef CONFIG_SMP
1511 /* Used instead of source_load when we know the type == 0 */
1512 static unsigned long weighted_cpuload(const int cpu)
1514 return cpu_rq(cpu)->load.weight;
1518 * Return a low guess at the load of a migration-source cpu weighted
1519 * according to the scheduling class and "nice" value.
1521 * We want to under-estimate the load of migration sources, to
1522 * balance conservatively.
1524 static unsigned long source_load(int cpu, int type)
1526 struct rq *rq = cpu_rq(cpu);
1527 unsigned long total = weighted_cpuload(cpu);
1529 if (type == 0 || !sched_feat(LB_BIAS))
1530 return total;
1532 return min(rq->cpu_load[type-1], total);
1536 * Return a high guess at the load of a migration-target cpu weighted
1537 * according to the scheduling class and "nice" value.
1539 static unsigned long target_load(int cpu, int type)
1541 struct rq *rq = cpu_rq(cpu);
1542 unsigned long total = weighted_cpuload(cpu);
1544 if (type == 0 || !sched_feat(LB_BIAS))
1545 return total;
1547 return max(rq->cpu_load[type-1], total);
1550 static unsigned long power_of(int cpu)
1552 return cpu_rq(cpu)->cpu_power;
1555 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1557 static unsigned long cpu_avg_load_per_task(int cpu)
1559 struct rq *rq = cpu_rq(cpu);
1560 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1562 if (nr_running)
1563 rq->avg_load_per_task = rq->load.weight / nr_running;
1564 else
1565 rq->avg_load_per_task = 0;
1567 return rq->avg_load_per_task;
1570 #ifdef CONFIG_FAIR_GROUP_SCHED
1573 * Compute the cpu's hierarchical load factor for each task group.
1574 * This needs to be done in a top-down fashion because the load of a child
1575 * group is a fraction of its parents load.
1577 static int tg_load_down(struct task_group *tg, void *data)
1579 unsigned long load;
1580 long cpu = (long)data;
1582 if (!tg->parent) {
1583 load = cpu_rq(cpu)->load.weight;
1584 } else {
1585 load = tg->parent->cfs_rq[cpu]->h_load;
1586 load *= tg->se[cpu]->load.weight;
1587 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1590 tg->cfs_rq[cpu]->h_load = load;
1592 return 0;
1595 static void update_h_load(long cpu)
1597 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1600 #endif
1602 #ifdef CONFIG_PREEMPT
1604 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1607 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1608 * way at the expense of forcing extra atomic operations in all
1609 * invocations. This assures that the double_lock is acquired using the
1610 * same underlying policy as the spinlock_t on this architecture, which
1611 * reduces latency compared to the unfair variant below. However, it
1612 * also adds more overhead and therefore may reduce throughput.
1614 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1615 __releases(this_rq->lock)
1616 __acquires(busiest->lock)
1617 __acquires(this_rq->lock)
1619 raw_spin_unlock(&this_rq->lock);
1620 double_rq_lock(this_rq, busiest);
1622 return 1;
1625 #else
1627 * Unfair double_lock_balance: Optimizes throughput at the expense of
1628 * latency by eliminating extra atomic operations when the locks are
1629 * already in proper order on entry. This favors lower cpu-ids and will
1630 * grant the double lock to lower cpus over higher ids under contention,
1631 * regardless of entry order into the function.
1633 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1634 __releases(this_rq->lock)
1635 __acquires(busiest->lock)
1636 __acquires(this_rq->lock)
1638 int ret = 0;
1640 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1641 if (busiest < this_rq) {
1642 raw_spin_unlock(&this_rq->lock);
1643 raw_spin_lock(&busiest->lock);
1644 raw_spin_lock_nested(&this_rq->lock,
1645 SINGLE_DEPTH_NESTING);
1646 ret = 1;
1647 } else
1648 raw_spin_lock_nested(&busiest->lock,
1649 SINGLE_DEPTH_NESTING);
1651 return ret;
1654 #endif /* CONFIG_PREEMPT */
1657 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1659 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1661 if (unlikely(!irqs_disabled())) {
1662 /* printk() doesn't work good under rq->lock */
1663 raw_spin_unlock(&this_rq->lock);
1664 BUG_ON(1);
1667 return _double_lock_balance(this_rq, busiest);
1670 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1671 __releases(busiest->lock)
1673 raw_spin_unlock(&busiest->lock);
1674 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1678 * double_rq_lock - safely lock two runqueues
1680 * Note this does not disable interrupts like task_rq_lock,
1681 * you need to do so manually before calling.
1683 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1684 __acquires(rq1->lock)
1685 __acquires(rq2->lock)
1687 BUG_ON(!irqs_disabled());
1688 if (rq1 == rq2) {
1689 raw_spin_lock(&rq1->lock);
1690 __acquire(rq2->lock); /* Fake it out ;) */
1691 } else {
1692 if (rq1 < rq2) {
1693 raw_spin_lock(&rq1->lock);
1694 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1695 } else {
1696 raw_spin_lock(&rq2->lock);
1697 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1703 * double_rq_unlock - safely unlock two runqueues
1705 * Note this does not restore interrupts like task_rq_unlock,
1706 * you need to do so manually after calling.
1708 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1709 __releases(rq1->lock)
1710 __releases(rq2->lock)
1712 raw_spin_unlock(&rq1->lock);
1713 if (rq1 != rq2)
1714 raw_spin_unlock(&rq2->lock);
1715 else
1716 __release(rq2->lock);
1719 #else /* CONFIG_SMP */
1722 * double_rq_lock - safely lock two runqueues
1724 * Note this does not disable interrupts like task_rq_lock,
1725 * you need to do so manually before calling.
1727 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1728 __acquires(rq1->lock)
1729 __acquires(rq2->lock)
1731 BUG_ON(!irqs_disabled());
1732 BUG_ON(rq1 != rq2);
1733 raw_spin_lock(&rq1->lock);
1734 __acquire(rq2->lock); /* Fake it out ;) */
1738 * double_rq_unlock - safely unlock two runqueues
1740 * Note this does not restore interrupts like task_rq_unlock,
1741 * you need to do so manually after calling.
1743 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1744 __releases(rq1->lock)
1745 __releases(rq2->lock)
1747 BUG_ON(rq1 != rq2);
1748 raw_spin_unlock(&rq1->lock);
1749 __release(rq2->lock);
1752 #endif
1754 static void calc_load_account_idle(struct rq *this_rq);
1755 static void update_sysctl(void);
1756 static int get_update_sysctl_factor(void);
1757 static void update_cpu_load(struct rq *this_rq);
1759 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1761 set_task_rq(p, cpu);
1762 #ifdef CONFIG_SMP
1764 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1765 * successfuly executed on another CPU. We must ensure that updates of
1766 * per-task data have been completed by this moment.
1768 smp_wmb();
1769 task_thread_info(p)->cpu = cpu;
1770 #endif
1773 static const struct sched_class rt_sched_class;
1775 #define sched_class_highest (&stop_sched_class)
1776 #define for_each_class(class) \
1777 for (class = sched_class_highest; class; class = class->next)
1779 #include "sched_stats.h"
1781 static void inc_nr_running(struct rq *rq)
1783 rq->nr_running++;
1786 static void dec_nr_running(struct rq *rq)
1788 rq->nr_running--;
1791 static void set_load_weight(struct task_struct *p)
1793 int prio = p->static_prio - MAX_RT_PRIO;
1794 struct load_weight *load = &p->se.load;
1797 * SCHED_IDLE tasks get minimal weight:
1799 if (p->policy == SCHED_IDLE) {
1800 load->weight = scale_load(WEIGHT_IDLEPRIO);
1801 load->inv_weight = WMULT_IDLEPRIO;
1802 return;
1805 load->weight = scale_load(prio_to_weight[prio]);
1806 load->inv_weight = prio_to_wmult[prio];
1809 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1811 update_rq_clock(rq);
1812 sched_info_queued(p);
1813 p->sched_class->enqueue_task(rq, p, flags);
1816 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1818 update_rq_clock(rq);
1819 sched_info_dequeued(p);
1820 p->sched_class->dequeue_task(rq, p, flags);
1824 * activate_task - move a task to the runqueue.
1826 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1828 if (task_contributes_to_load(p))
1829 rq->nr_uninterruptible--;
1831 enqueue_task(rq, p, flags);
1832 inc_nr_running(rq);
1836 * deactivate_task - remove a task from the runqueue.
1838 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1840 if (task_contributes_to_load(p))
1841 rq->nr_uninterruptible++;
1843 dequeue_task(rq, p, flags);
1844 dec_nr_running(rq);
1847 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1850 * There are no locks covering percpu hardirq/softirq time.
1851 * They are only modified in account_system_vtime, on corresponding CPU
1852 * with interrupts disabled. So, writes are safe.
1853 * They are read and saved off onto struct rq in update_rq_clock().
1854 * This may result in other CPU reading this CPU's irq time and can
1855 * race with irq/account_system_vtime on this CPU. We would either get old
1856 * or new value with a side effect of accounting a slice of irq time to wrong
1857 * task when irq is in progress while we read rq->clock. That is a worthy
1858 * compromise in place of having locks on each irq in account_system_time.
1860 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1861 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1863 static DEFINE_PER_CPU(u64, irq_start_time);
1864 static int sched_clock_irqtime;
1866 void enable_sched_clock_irqtime(void)
1868 sched_clock_irqtime = 1;
1871 void disable_sched_clock_irqtime(void)
1873 sched_clock_irqtime = 0;
1876 #ifndef CONFIG_64BIT
1877 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1879 static inline void irq_time_write_begin(void)
1881 __this_cpu_inc(irq_time_seq.sequence);
1882 smp_wmb();
1885 static inline void irq_time_write_end(void)
1887 smp_wmb();
1888 __this_cpu_inc(irq_time_seq.sequence);
1891 static inline u64 irq_time_read(int cpu)
1893 u64 irq_time;
1894 unsigned seq;
1896 do {
1897 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1898 irq_time = per_cpu(cpu_softirq_time, cpu) +
1899 per_cpu(cpu_hardirq_time, cpu);
1900 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1902 return irq_time;
1904 #else /* CONFIG_64BIT */
1905 static inline void irq_time_write_begin(void)
1909 static inline void irq_time_write_end(void)
1913 static inline u64 irq_time_read(int cpu)
1915 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1917 #endif /* CONFIG_64BIT */
1920 * Called before incrementing preempt_count on {soft,}irq_enter
1921 * and before decrementing preempt_count on {soft,}irq_exit.
1923 void account_system_vtime(struct task_struct *curr)
1925 unsigned long flags;
1926 s64 delta;
1927 int cpu;
1929 if (!sched_clock_irqtime)
1930 return;
1932 local_irq_save(flags);
1934 cpu = smp_processor_id();
1935 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1936 __this_cpu_add(irq_start_time, delta);
1938 irq_time_write_begin();
1940 * We do not account for softirq time from ksoftirqd here.
1941 * We want to continue accounting softirq time to ksoftirqd thread
1942 * in that case, so as not to confuse scheduler with a special task
1943 * that do not consume any time, but still wants to run.
1945 if (hardirq_count())
1946 __this_cpu_add(cpu_hardirq_time, delta);
1947 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1948 __this_cpu_add(cpu_softirq_time, delta);
1950 irq_time_write_end();
1951 local_irq_restore(flags);
1953 EXPORT_SYMBOL_GPL(account_system_vtime);
1955 static void update_rq_clock_task(struct rq *rq, s64 delta)
1957 s64 irq_delta;
1959 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1962 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1963 * this case when a previous update_rq_clock() happened inside a
1964 * {soft,}irq region.
1966 * When this happens, we stop ->clock_task and only update the
1967 * prev_irq_time stamp to account for the part that fit, so that a next
1968 * update will consume the rest. This ensures ->clock_task is
1969 * monotonic.
1971 * It does however cause some slight miss-attribution of {soft,}irq
1972 * time, a more accurate solution would be to update the irq_time using
1973 * the current rq->clock timestamp, except that would require using
1974 * atomic ops.
1976 if (irq_delta > delta)
1977 irq_delta = delta;
1979 rq->prev_irq_time += irq_delta;
1980 delta -= irq_delta;
1981 rq->clock_task += delta;
1983 if (irq_delta && sched_feat(NONIRQ_POWER))
1984 sched_rt_avg_update(rq, irq_delta);
1987 static int irqtime_account_hi_update(void)
1989 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1990 unsigned long flags;
1991 u64 latest_ns;
1992 int ret = 0;
1994 local_irq_save(flags);
1995 latest_ns = this_cpu_read(cpu_hardirq_time);
1996 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
1997 ret = 1;
1998 local_irq_restore(flags);
1999 return ret;
2002 static int irqtime_account_si_update(void)
2004 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2005 unsigned long flags;
2006 u64 latest_ns;
2007 int ret = 0;
2009 local_irq_save(flags);
2010 latest_ns = this_cpu_read(cpu_softirq_time);
2011 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2012 ret = 1;
2013 local_irq_restore(flags);
2014 return ret;
2017 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2019 #define sched_clock_irqtime (0)
2021 static void update_rq_clock_task(struct rq *rq, s64 delta)
2023 rq->clock_task += delta;
2026 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2028 #include "sched_idletask.c"
2029 #include "sched_fair.c"
2030 #include "sched_rt.c"
2031 #include "sched_autogroup.c"
2032 #include "sched_stoptask.c"
2033 #ifdef CONFIG_SCHED_DEBUG
2034 # include "sched_debug.c"
2035 #endif
2037 void sched_set_stop_task(int cpu, struct task_struct *stop)
2039 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2040 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2042 if (stop) {
2044 * Make it appear like a SCHED_FIFO task, its something
2045 * userspace knows about and won't get confused about.
2047 * Also, it will make PI more or less work without too
2048 * much confusion -- but then, stop work should not
2049 * rely on PI working anyway.
2051 sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
2053 stop->sched_class = &stop_sched_class;
2056 cpu_rq(cpu)->stop = stop;
2058 if (old_stop) {
2060 * Reset it back to a normal scheduling class so that
2061 * it can die in pieces.
2063 old_stop->sched_class = &rt_sched_class;
2068 * __normal_prio - return the priority that is based on the static prio
2070 static inline int __normal_prio(struct task_struct *p)
2072 return p->static_prio;
2076 * Calculate the expected normal priority: i.e. priority
2077 * without taking RT-inheritance into account. Might be
2078 * boosted by interactivity modifiers. Changes upon fork,
2079 * setprio syscalls, and whenever the interactivity
2080 * estimator recalculates.
2082 static inline int normal_prio(struct task_struct *p)
2084 int prio;
2086 if (task_has_rt_policy(p))
2087 prio = MAX_RT_PRIO-1 - p->rt_priority;
2088 else
2089 prio = __normal_prio(p);
2090 return prio;
2094 * Calculate the current priority, i.e. the priority
2095 * taken into account by the scheduler. This value might
2096 * be boosted by RT tasks, or might be boosted by
2097 * interactivity modifiers. Will be RT if the task got
2098 * RT-boosted. If not then it returns p->normal_prio.
2100 static int effective_prio(struct task_struct *p)
2102 p->normal_prio = normal_prio(p);
2104 * If we are RT tasks or we were boosted to RT priority,
2105 * keep the priority unchanged. Otherwise, update priority
2106 * to the normal priority:
2108 if (!rt_prio(p->prio))
2109 return p->normal_prio;
2110 return p->prio;
2114 * task_curr - is this task currently executing on a CPU?
2115 * @p: the task in question.
2117 inline int task_curr(const struct task_struct *p)
2119 return cpu_curr(task_cpu(p)) == p;
2122 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2123 const struct sched_class *prev_class,
2124 int oldprio)
2126 if (prev_class != p->sched_class) {
2127 if (prev_class->switched_from)
2128 prev_class->switched_from(rq, p);
2129 p->sched_class->switched_to(rq, p);
2130 } else if (oldprio != p->prio)
2131 p->sched_class->prio_changed(rq, p, oldprio);
2134 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2136 const struct sched_class *class;
2138 if (p->sched_class == rq->curr->sched_class) {
2139 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2140 } else {
2141 for_each_class(class) {
2142 if (class == rq->curr->sched_class)
2143 break;
2144 if (class == p->sched_class) {
2145 resched_task(rq->curr);
2146 break;
2152 * A queue event has occurred, and we're going to schedule. In
2153 * this case, we can save a useless back to back clock update.
2155 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2156 rq->skip_clock_update = 1;
2159 #ifdef CONFIG_SMP
2161 * Is this task likely cache-hot:
2163 static int
2164 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2166 s64 delta;
2168 if (p->sched_class != &fair_sched_class)
2169 return 0;
2171 if (unlikely(p->policy == SCHED_IDLE))
2172 return 0;
2175 * Buddy candidates are cache hot:
2177 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2178 (&p->se == cfs_rq_of(&p->se)->next ||
2179 &p->se == cfs_rq_of(&p->se)->last))
2180 return 1;
2182 if (sysctl_sched_migration_cost == -1)
2183 return 1;
2184 if (sysctl_sched_migration_cost == 0)
2185 return 0;
2187 delta = now - p->se.exec_start;
2189 return delta < (s64)sysctl_sched_migration_cost;
2192 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2194 #ifdef CONFIG_SCHED_DEBUG
2196 * We should never call set_task_cpu() on a blocked task,
2197 * ttwu() will sort out the placement.
2199 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2200 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2202 #ifdef CONFIG_LOCKDEP
2203 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2204 lockdep_is_held(&task_rq(p)->lock)));
2205 #endif
2206 #endif
2208 trace_sched_migrate_task(p, new_cpu);
2210 if (task_cpu(p) != new_cpu) {
2211 p->se.nr_migrations++;
2212 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2215 __set_task_cpu(p, new_cpu);
2218 struct migration_arg {
2219 struct task_struct *task;
2220 int dest_cpu;
2223 static int migration_cpu_stop(void *data);
2226 * wait_task_inactive - wait for a thread to unschedule.
2228 * If @match_state is nonzero, it's the @p->state value just checked and
2229 * not expected to change. If it changes, i.e. @p might have woken up,
2230 * then return zero. When we succeed in waiting for @p to be off its CPU,
2231 * we return a positive number (its total switch count). If a second call
2232 * a short while later returns the same number, the caller can be sure that
2233 * @p has remained unscheduled the whole time.
2235 * The caller must ensure that the task *will* unschedule sometime soon,
2236 * else this function might spin for a *long* time. This function can't
2237 * be called with interrupts off, or it may introduce deadlock with
2238 * smp_call_function() if an IPI is sent by the same process we are
2239 * waiting to become inactive.
2241 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2243 unsigned long flags;
2244 int running, on_rq;
2245 unsigned long ncsw;
2246 struct rq *rq;
2248 for (;;) {
2250 * We do the initial early heuristics without holding
2251 * any task-queue locks at all. We'll only try to get
2252 * the runqueue lock when things look like they will
2253 * work out!
2255 rq = task_rq(p);
2258 * If the task is actively running on another CPU
2259 * still, just relax and busy-wait without holding
2260 * any locks.
2262 * NOTE! Since we don't hold any locks, it's not
2263 * even sure that "rq" stays as the right runqueue!
2264 * But we don't care, since "task_running()" will
2265 * return false if the runqueue has changed and p
2266 * is actually now running somewhere else!
2268 while (task_running(rq, p)) {
2269 if (match_state && unlikely(p->state != match_state))
2270 return 0;
2271 cpu_relax();
2275 * Ok, time to look more closely! We need the rq
2276 * lock now, to be *sure*. If we're wrong, we'll
2277 * just go back and repeat.
2279 rq = task_rq_lock(p, &flags);
2280 trace_sched_wait_task(p);
2281 running = task_running(rq, p);
2282 on_rq = p->on_rq;
2283 ncsw = 0;
2284 if (!match_state || p->state == match_state)
2285 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2286 task_rq_unlock(rq, p, &flags);
2289 * If it changed from the expected state, bail out now.
2291 if (unlikely(!ncsw))
2292 break;
2295 * Was it really running after all now that we
2296 * checked with the proper locks actually held?
2298 * Oops. Go back and try again..
2300 if (unlikely(running)) {
2301 cpu_relax();
2302 continue;
2306 * It's not enough that it's not actively running,
2307 * it must be off the runqueue _entirely_, and not
2308 * preempted!
2310 * So if it was still runnable (but just not actively
2311 * running right now), it's preempted, and we should
2312 * yield - it could be a while.
2314 if (unlikely(on_rq)) {
2315 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2317 set_current_state(TASK_UNINTERRUPTIBLE);
2318 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2319 continue;
2323 * Ahh, all good. It wasn't running, and it wasn't
2324 * runnable, which means that it will never become
2325 * running in the future either. We're all done!
2327 break;
2330 return ncsw;
2333 /***
2334 * kick_process - kick a running thread to enter/exit the kernel
2335 * @p: the to-be-kicked thread
2337 * Cause a process which is running on another CPU to enter
2338 * kernel-mode, without any delay. (to get signals handled.)
2340 * NOTE: this function doesn't have to take the runqueue lock,
2341 * because all it wants to ensure is that the remote task enters
2342 * the kernel. If the IPI races and the task has been migrated
2343 * to another CPU then no harm is done and the purpose has been
2344 * achieved as well.
2346 void kick_process(struct task_struct *p)
2348 int cpu;
2350 preempt_disable();
2351 cpu = task_cpu(p);
2352 if ((cpu != smp_processor_id()) && task_curr(p))
2353 smp_send_reschedule(cpu);
2354 preempt_enable();
2356 EXPORT_SYMBOL_GPL(kick_process);
2357 #endif /* CONFIG_SMP */
2359 #ifdef CONFIG_SMP
2361 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2363 static int select_fallback_rq(int cpu, struct task_struct *p)
2365 int dest_cpu;
2366 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2368 /* Look for allowed, online CPU in same node. */
2369 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2370 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2371 return dest_cpu;
2373 /* Any allowed, online CPU? */
2374 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2375 if (dest_cpu < nr_cpu_ids)
2376 return dest_cpu;
2378 /* No more Mr. Nice Guy. */
2379 dest_cpu = cpuset_cpus_allowed_fallback(p);
2381 * Don't tell them about moving exiting tasks or
2382 * kernel threads (both mm NULL), since they never
2383 * leave kernel.
2385 if (p->mm && printk_ratelimit()) {
2386 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2387 task_pid_nr(p), p->comm, cpu);
2390 return dest_cpu;
2394 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2396 static inline
2397 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2399 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2402 * In order not to call set_task_cpu() on a blocking task we need
2403 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2404 * cpu.
2406 * Since this is common to all placement strategies, this lives here.
2408 * [ this allows ->select_task() to simply return task_cpu(p) and
2409 * not worry about this generic constraint ]
2411 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2412 !cpu_online(cpu)))
2413 cpu = select_fallback_rq(task_cpu(p), p);
2415 return cpu;
2418 static void update_avg(u64 *avg, u64 sample)
2420 s64 diff = sample - *avg;
2421 *avg += diff >> 3;
2423 #endif
2425 static void
2426 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2428 #ifdef CONFIG_SCHEDSTATS
2429 struct rq *rq = this_rq();
2431 #ifdef CONFIG_SMP
2432 int this_cpu = smp_processor_id();
2434 if (cpu == this_cpu) {
2435 schedstat_inc(rq, ttwu_local);
2436 schedstat_inc(p, se.statistics.nr_wakeups_local);
2437 } else {
2438 struct sched_domain *sd;
2440 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2441 rcu_read_lock();
2442 for_each_domain(this_cpu, sd) {
2443 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2444 schedstat_inc(sd, ttwu_wake_remote);
2445 break;
2448 rcu_read_unlock();
2450 #endif /* CONFIG_SMP */
2452 schedstat_inc(rq, ttwu_count);
2453 schedstat_inc(p, se.statistics.nr_wakeups);
2455 if (wake_flags & WF_SYNC)
2456 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2458 if (cpu != task_cpu(p))
2459 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2461 #endif /* CONFIG_SCHEDSTATS */
2464 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2466 activate_task(rq, p, en_flags);
2467 p->on_rq = 1;
2469 /* if a worker is waking up, notify workqueue */
2470 if (p->flags & PF_WQ_WORKER)
2471 wq_worker_waking_up(p, cpu_of(rq));
2475 * Mark the task runnable and perform wakeup-preemption.
2477 static void
2478 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2480 trace_sched_wakeup(p, true);
2481 check_preempt_curr(rq, p, wake_flags);
2483 p->state = TASK_RUNNING;
2484 #ifdef CONFIG_SMP
2485 if (p->sched_class->task_woken)
2486 p->sched_class->task_woken(rq, p);
2488 if (unlikely(rq->idle_stamp)) {
2489 u64 delta = rq->clock - rq->idle_stamp;
2490 u64 max = 2*sysctl_sched_migration_cost;
2492 if (delta > max)
2493 rq->avg_idle = max;
2494 else
2495 update_avg(&rq->avg_idle, delta);
2496 rq->idle_stamp = 0;
2498 #endif
2501 static void
2502 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2504 #ifdef CONFIG_SMP
2505 if (p->sched_contributes_to_load)
2506 rq->nr_uninterruptible--;
2507 #endif
2509 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2510 ttwu_do_wakeup(rq, p, wake_flags);
2514 * Called in case the task @p isn't fully descheduled from its runqueue,
2515 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2516 * since all we need to do is flip p->state to TASK_RUNNING, since
2517 * the task is still ->on_rq.
2519 static int ttwu_remote(struct task_struct *p, int wake_flags)
2521 struct rq *rq;
2522 int ret = 0;
2524 rq = __task_rq_lock(p);
2525 if (p->on_rq) {
2526 ttwu_do_wakeup(rq, p, wake_flags);
2527 ret = 1;
2529 __task_rq_unlock(rq);
2531 return ret;
2534 #ifdef CONFIG_SMP
2535 static void sched_ttwu_pending(void)
2537 struct rq *rq = this_rq();
2538 struct task_struct *list = xchg(&rq->wake_list, NULL);
2540 if (!list)
2541 return;
2543 raw_spin_lock(&rq->lock);
2545 while (list) {
2546 struct task_struct *p = list;
2547 list = list->wake_entry;
2548 ttwu_do_activate(rq, p, 0);
2551 raw_spin_unlock(&rq->lock);
2554 void scheduler_ipi(void)
2556 sched_ttwu_pending();
2559 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2561 struct rq *rq = cpu_rq(cpu);
2562 struct task_struct *next = rq->wake_list;
2564 for (;;) {
2565 struct task_struct *old = next;
2567 p->wake_entry = next;
2568 next = cmpxchg(&rq->wake_list, old, p);
2569 if (next == old)
2570 break;
2573 if (!next)
2574 smp_send_reschedule(cpu);
2576 #endif
2578 static void ttwu_queue(struct task_struct *p, int cpu)
2580 struct rq *rq = cpu_rq(cpu);
2582 #if defined(CONFIG_SMP)
2583 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2584 ttwu_queue_remote(p, cpu);
2585 return;
2587 #endif
2589 raw_spin_lock(&rq->lock);
2590 ttwu_do_activate(rq, p, 0);
2591 raw_spin_unlock(&rq->lock);
2595 * try_to_wake_up - wake up a thread
2596 * @p: the thread to be awakened
2597 * @state: the mask of task states that can be woken
2598 * @wake_flags: wake modifier flags (WF_*)
2600 * Put it on the run-queue if it's not already there. The "current"
2601 * thread is always on the run-queue (except when the actual
2602 * re-schedule is in progress), and as such you're allowed to do
2603 * the simpler "current->state = TASK_RUNNING" to mark yourself
2604 * runnable without the overhead of this.
2606 * Returns %true if @p was woken up, %false if it was already running
2607 * or @state didn't match @p's state.
2609 static int
2610 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2612 unsigned long flags;
2613 int cpu, success = 0;
2615 smp_wmb();
2616 raw_spin_lock_irqsave(&p->pi_lock, flags);
2617 if (!(p->state & state))
2618 goto out;
2620 success = 1; /* we're going to change ->state */
2621 cpu = task_cpu(p);
2623 if (p->on_rq && ttwu_remote(p, wake_flags))
2624 goto stat;
2626 #ifdef CONFIG_SMP
2628 * If the owning (remote) cpu is still in the middle of schedule() with
2629 * this task as prev, wait until its done referencing the task.
2631 while (p->on_cpu) {
2632 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2634 * If called from interrupt context we could have landed in the
2635 * middle of schedule(), in this case we should take care not
2636 * to spin on ->on_cpu if p is current, since that would
2637 * deadlock.
2639 if (p == current) {
2640 ttwu_queue(p, cpu);
2641 goto stat;
2643 #endif
2644 cpu_relax();
2647 * Pairs with the smp_wmb() in finish_lock_switch().
2649 smp_rmb();
2651 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2652 p->state = TASK_WAKING;
2654 if (p->sched_class->task_waking)
2655 p->sched_class->task_waking(p);
2657 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2658 if (task_cpu(p) != cpu)
2659 set_task_cpu(p, cpu);
2660 #endif /* CONFIG_SMP */
2662 ttwu_queue(p, cpu);
2663 stat:
2664 ttwu_stat(p, cpu, wake_flags);
2665 out:
2666 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2668 return success;
2672 * try_to_wake_up_local - try to wake up a local task with rq lock held
2673 * @p: the thread to be awakened
2675 * Put @p on the run-queue if it's not already there. The caller must
2676 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2677 * the current task.
2679 static void try_to_wake_up_local(struct task_struct *p)
2681 struct rq *rq = task_rq(p);
2683 BUG_ON(rq != this_rq());
2684 BUG_ON(p == current);
2685 lockdep_assert_held(&rq->lock);
2687 if (!raw_spin_trylock(&p->pi_lock)) {
2688 raw_spin_unlock(&rq->lock);
2689 raw_spin_lock(&p->pi_lock);
2690 raw_spin_lock(&rq->lock);
2693 if (!(p->state & TASK_NORMAL))
2694 goto out;
2696 if (!p->on_rq)
2697 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2699 ttwu_do_wakeup(rq, p, 0);
2700 ttwu_stat(p, smp_processor_id(), 0);
2701 out:
2702 raw_spin_unlock(&p->pi_lock);
2706 * wake_up_process - Wake up a specific process
2707 * @p: The process to be woken up.
2709 * Attempt to wake up the nominated process and move it to the set of runnable
2710 * processes. Returns 1 if the process was woken up, 0 if it was already
2711 * running.
2713 * It may be assumed that this function implies a write memory barrier before
2714 * changing the task state if and only if any tasks are woken up.
2716 int wake_up_process(struct task_struct *p)
2718 return try_to_wake_up(p, TASK_ALL, 0);
2720 EXPORT_SYMBOL(wake_up_process);
2722 int wake_up_state(struct task_struct *p, unsigned int state)
2724 return try_to_wake_up(p, state, 0);
2728 * Perform scheduler related setup for a newly forked process p.
2729 * p is forked by current.
2731 * __sched_fork() is basic setup used by init_idle() too:
2733 static void __sched_fork(struct task_struct *p)
2735 p->on_rq = 0;
2737 p->se.on_rq = 0;
2738 p->se.exec_start = 0;
2739 p->se.sum_exec_runtime = 0;
2740 p->se.prev_sum_exec_runtime = 0;
2741 p->se.nr_migrations = 0;
2742 p->se.vruntime = 0;
2743 INIT_LIST_HEAD(&p->se.group_node);
2745 #ifdef CONFIG_SCHEDSTATS
2746 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2747 #endif
2749 INIT_LIST_HEAD(&p->rt.run_list);
2751 #ifdef CONFIG_PREEMPT_NOTIFIERS
2752 INIT_HLIST_HEAD(&p->preempt_notifiers);
2753 #endif
2757 * fork()/clone()-time setup:
2759 void sched_fork(struct task_struct *p)
2761 unsigned long flags;
2762 int cpu = get_cpu();
2764 __sched_fork(p);
2766 * We mark the process as running here. This guarantees that
2767 * nobody will actually run it, and a signal or other external
2768 * event cannot wake it up and insert it on the runqueue either.
2770 p->state = TASK_RUNNING;
2773 * Revert to default priority/policy on fork if requested.
2775 if (unlikely(p->sched_reset_on_fork)) {
2776 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2777 p->policy = SCHED_NORMAL;
2778 p->normal_prio = p->static_prio;
2781 if (PRIO_TO_NICE(p->static_prio) < 0) {
2782 p->static_prio = NICE_TO_PRIO(0);
2783 p->normal_prio = p->static_prio;
2784 set_load_weight(p);
2788 * We don't need the reset flag anymore after the fork. It has
2789 * fulfilled its duty:
2791 p->sched_reset_on_fork = 0;
2795 * Make sure we do not leak PI boosting priority to the child.
2797 p->prio = current->normal_prio;
2799 if (!rt_prio(p->prio))
2800 p->sched_class = &fair_sched_class;
2802 if (p->sched_class->task_fork)
2803 p->sched_class->task_fork(p);
2806 * The child is not yet in the pid-hash so no cgroup attach races,
2807 * and the cgroup is pinned to this child due to cgroup_fork()
2808 * is ran before sched_fork().
2810 * Silence PROVE_RCU.
2812 raw_spin_lock_irqsave(&p->pi_lock, flags);
2813 set_task_cpu(p, cpu);
2814 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2816 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2817 if (likely(sched_info_on()))
2818 memset(&p->sched_info, 0, sizeof(p->sched_info));
2819 #endif
2820 #if defined(CONFIG_SMP)
2821 p->on_cpu = 0;
2822 #endif
2823 #ifdef CONFIG_PREEMPT
2824 /* Want to start with kernel preemption disabled. */
2825 task_thread_info(p)->preempt_count = 1;
2826 #endif
2827 #ifdef CONFIG_SMP
2828 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2829 #endif
2831 put_cpu();
2835 * wake_up_new_task - wake up a newly created task for the first time.
2837 * This function will do some initial scheduler statistics housekeeping
2838 * that must be done for every newly created context, then puts the task
2839 * on the runqueue and wakes it.
2841 void wake_up_new_task(struct task_struct *p)
2843 unsigned long flags;
2844 struct rq *rq;
2846 raw_spin_lock_irqsave(&p->pi_lock, flags);
2847 #ifdef CONFIG_SMP
2849 * Fork balancing, do it here and not earlier because:
2850 * - cpus_allowed can change in the fork path
2851 * - any previously selected cpu might disappear through hotplug
2853 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
2854 #endif
2856 rq = __task_rq_lock(p);
2857 activate_task(rq, p, 0);
2858 p->on_rq = 1;
2859 trace_sched_wakeup_new(p, true);
2860 check_preempt_curr(rq, p, WF_FORK);
2861 #ifdef CONFIG_SMP
2862 if (p->sched_class->task_woken)
2863 p->sched_class->task_woken(rq, p);
2864 #endif
2865 task_rq_unlock(rq, p, &flags);
2868 #ifdef CONFIG_PREEMPT_NOTIFIERS
2871 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2872 * @notifier: notifier struct to register
2874 void preempt_notifier_register(struct preempt_notifier *notifier)
2876 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2878 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2881 * preempt_notifier_unregister - no longer interested in preemption notifications
2882 * @notifier: notifier struct to unregister
2884 * This is safe to call from within a preemption notifier.
2886 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2888 hlist_del(&notifier->link);
2890 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2892 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2894 struct preempt_notifier *notifier;
2895 struct hlist_node *node;
2897 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2898 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2901 static void
2902 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2903 struct task_struct *next)
2905 struct preempt_notifier *notifier;
2906 struct hlist_node *node;
2908 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2909 notifier->ops->sched_out(notifier, next);
2912 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2914 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2918 static void
2919 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2920 struct task_struct *next)
2924 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2927 * prepare_task_switch - prepare to switch tasks
2928 * @rq: the runqueue preparing to switch
2929 * @prev: the current task that is being switched out
2930 * @next: the task we are going to switch to.
2932 * This is called with the rq lock held and interrupts off. It must
2933 * be paired with a subsequent finish_task_switch after the context
2934 * switch.
2936 * prepare_task_switch sets up locking and calls architecture specific
2937 * hooks.
2939 static inline void
2940 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2941 struct task_struct *next)
2943 sched_info_switch(prev, next);
2944 perf_event_task_sched_out(prev, next);
2945 fire_sched_out_preempt_notifiers(prev, next);
2946 prepare_lock_switch(rq, next);
2947 prepare_arch_switch(next);
2948 trace_sched_switch(prev, next);
2952 * finish_task_switch - clean up after a task-switch
2953 * @rq: runqueue associated with task-switch
2954 * @prev: the thread we just switched away from.
2956 * finish_task_switch must be called after the context switch, paired
2957 * with a prepare_task_switch call before the context switch.
2958 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2959 * and do any other architecture-specific cleanup actions.
2961 * Note that we may have delayed dropping an mm in context_switch(). If
2962 * so, we finish that here outside of the runqueue lock. (Doing it
2963 * with the lock held can cause deadlocks; see schedule() for
2964 * details.)
2966 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2967 __releases(rq->lock)
2969 struct mm_struct *mm = rq->prev_mm;
2970 long prev_state;
2972 rq->prev_mm = NULL;
2975 * A task struct has one reference for the use as "current".
2976 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2977 * schedule one last time. The schedule call will never return, and
2978 * the scheduled task must drop that reference.
2979 * The test for TASK_DEAD must occur while the runqueue locks are
2980 * still held, otherwise prev could be scheduled on another cpu, die
2981 * there before we look at prev->state, and then the reference would
2982 * be dropped twice.
2983 * Manfred Spraul <manfred@colorfullife.com>
2985 prev_state = prev->state;
2986 finish_arch_switch(prev);
2987 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2988 local_irq_disable();
2989 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2990 perf_event_task_sched_in(current);
2991 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2992 local_irq_enable();
2993 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2994 finish_lock_switch(rq, prev);
2996 fire_sched_in_preempt_notifiers(current);
2997 if (mm)
2998 mmdrop(mm);
2999 if (unlikely(prev_state == TASK_DEAD)) {
3001 * Remove function-return probe instances associated with this
3002 * task and put them back on the free list.
3004 kprobe_flush_task(prev);
3005 put_task_struct(prev);
3009 #ifdef CONFIG_SMP
3011 /* assumes rq->lock is held */
3012 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3014 if (prev->sched_class->pre_schedule)
3015 prev->sched_class->pre_schedule(rq, prev);
3018 /* rq->lock is NOT held, but preemption is disabled */
3019 static inline void post_schedule(struct rq *rq)
3021 if (rq->post_schedule) {
3022 unsigned long flags;
3024 raw_spin_lock_irqsave(&rq->lock, flags);
3025 if (rq->curr->sched_class->post_schedule)
3026 rq->curr->sched_class->post_schedule(rq);
3027 raw_spin_unlock_irqrestore(&rq->lock, flags);
3029 rq->post_schedule = 0;
3033 #else
3035 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3039 static inline void post_schedule(struct rq *rq)
3043 #endif
3046 * schedule_tail - first thing a freshly forked thread must call.
3047 * @prev: the thread we just switched away from.
3049 asmlinkage void schedule_tail(struct task_struct *prev)
3050 __releases(rq->lock)
3052 struct rq *rq = this_rq();
3054 finish_task_switch(rq, prev);
3057 * FIXME: do we need to worry about rq being invalidated by the
3058 * task_switch?
3060 post_schedule(rq);
3062 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3063 /* In this case, finish_task_switch does not reenable preemption */
3064 preempt_enable();
3065 #endif
3066 if (current->set_child_tid)
3067 put_user(task_pid_vnr(current), current->set_child_tid);
3071 * context_switch - switch to the new MM and the new
3072 * thread's register state.
3074 static inline void
3075 context_switch(struct rq *rq, struct task_struct *prev,
3076 struct task_struct *next)
3078 struct mm_struct *mm, *oldmm;
3080 prepare_task_switch(rq, prev, next);
3082 mm = next->mm;
3083 oldmm = prev->active_mm;
3085 * For paravirt, this is coupled with an exit in switch_to to
3086 * combine the page table reload and the switch backend into
3087 * one hypercall.
3089 arch_start_context_switch(prev);
3091 if (!mm) {
3092 next->active_mm = oldmm;
3093 atomic_inc(&oldmm->mm_count);
3094 enter_lazy_tlb(oldmm, next);
3095 } else
3096 switch_mm(oldmm, mm, next);
3098 if (!prev->mm) {
3099 prev->active_mm = NULL;
3100 rq->prev_mm = oldmm;
3103 * Since the runqueue lock will be released by the next
3104 * task (which is an invalid locking op but in the case
3105 * of the scheduler it's an obvious special-case), so we
3106 * do an early lockdep release here:
3108 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3109 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3110 #endif
3112 /* Here we just switch the register state and the stack. */
3113 switch_to(prev, next, prev);
3115 barrier();
3117 * this_rq must be evaluated again because prev may have moved
3118 * CPUs since it called schedule(), thus the 'rq' on its stack
3119 * frame will be invalid.
3121 finish_task_switch(this_rq(), prev);
3125 * nr_running, nr_uninterruptible and nr_context_switches:
3127 * externally visible scheduler statistics: current number of runnable
3128 * threads, current number of uninterruptible-sleeping threads, total
3129 * number of context switches performed since bootup.
3131 unsigned long nr_running(void)
3133 unsigned long i, sum = 0;
3135 for_each_online_cpu(i)
3136 sum += cpu_rq(i)->nr_running;
3138 return sum;
3141 unsigned long nr_uninterruptible(void)
3143 unsigned long i, sum = 0;
3145 for_each_possible_cpu(i)
3146 sum += cpu_rq(i)->nr_uninterruptible;
3149 * Since we read the counters lockless, it might be slightly
3150 * inaccurate. Do not allow it to go below zero though:
3152 if (unlikely((long)sum < 0))
3153 sum = 0;
3155 return sum;
3158 unsigned long long nr_context_switches(void)
3160 int i;
3161 unsigned long long sum = 0;
3163 for_each_possible_cpu(i)
3164 sum += cpu_rq(i)->nr_switches;
3166 return sum;
3169 unsigned long nr_iowait(void)
3171 unsigned long i, sum = 0;
3173 for_each_possible_cpu(i)
3174 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3176 return sum;
3179 unsigned long nr_iowait_cpu(int cpu)
3181 struct rq *this = cpu_rq(cpu);
3182 return atomic_read(&this->nr_iowait);
3185 unsigned long this_cpu_load(void)
3187 struct rq *this = this_rq();
3188 return this->cpu_load[0];
3192 /* Variables and functions for calc_load */
3193 static atomic_long_t calc_load_tasks;
3194 static unsigned long calc_load_update;
3195 unsigned long avenrun[3];
3196 EXPORT_SYMBOL(avenrun);
3198 static long calc_load_fold_active(struct rq *this_rq)
3200 long nr_active, delta = 0;
3202 nr_active = this_rq->nr_running;
3203 nr_active += (long) this_rq->nr_uninterruptible;
3205 if (nr_active != this_rq->calc_load_active) {
3206 delta = nr_active - this_rq->calc_load_active;
3207 this_rq->calc_load_active = nr_active;
3210 return delta;
3213 static unsigned long
3214 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3216 load *= exp;
3217 load += active * (FIXED_1 - exp);
3218 load += 1UL << (FSHIFT - 1);
3219 return load >> FSHIFT;
3222 #ifdef CONFIG_NO_HZ
3224 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3226 * When making the ILB scale, we should try to pull this in as well.
3228 static atomic_long_t calc_load_tasks_idle;
3230 static void calc_load_account_idle(struct rq *this_rq)
3232 long delta;
3234 delta = calc_load_fold_active(this_rq);
3235 if (delta)
3236 atomic_long_add(delta, &calc_load_tasks_idle);
3239 static long calc_load_fold_idle(void)
3241 long delta = 0;
3244 * Its got a race, we don't care...
3246 if (atomic_long_read(&calc_load_tasks_idle))
3247 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3249 return delta;
3253 * fixed_power_int - compute: x^n, in O(log n) time
3255 * @x: base of the power
3256 * @frac_bits: fractional bits of @x
3257 * @n: power to raise @x to.
3259 * By exploiting the relation between the definition of the natural power
3260 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3261 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3262 * (where: n_i \elem {0, 1}, the binary vector representing n),
3263 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3264 * of course trivially computable in O(log_2 n), the length of our binary
3265 * vector.
3267 static unsigned long
3268 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3270 unsigned long result = 1UL << frac_bits;
3272 if (n) for (;;) {
3273 if (n & 1) {
3274 result *= x;
3275 result += 1UL << (frac_bits - 1);
3276 result >>= frac_bits;
3278 n >>= 1;
3279 if (!n)
3280 break;
3281 x *= x;
3282 x += 1UL << (frac_bits - 1);
3283 x >>= frac_bits;
3286 return result;
3290 * a1 = a0 * e + a * (1 - e)
3292 * a2 = a1 * e + a * (1 - e)
3293 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3294 * = a0 * e^2 + a * (1 - e) * (1 + e)
3296 * a3 = a2 * e + a * (1 - e)
3297 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3298 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3300 * ...
3302 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3303 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3304 * = a0 * e^n + a * (1 - e^n)
3306 * [1] application of the geometric series:
3308 * n 1 - x^(n+1)
3309 * S_n := \Sum x^i = -------------
3310 * i=0 1 - x
3312 static unsigned long
3313 calc_load_n(unsigned long load, unsigned long exp,
3314 unsigned long active, unsigned int n)
3317 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3321 * NO_HZ can leave us missing all per-cpu ticks calling
3322 * calc_load_account_active(), but since an idle CPU folds its delta into
3323 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3324 * in the pending idle delta if our idle period crossed a load cycle boundary.
3326 * Once we've updated the global active value, we need to apply the exponential
3327 * weights adjusted to the number of cycles missed.
3329 static void calc_global_nohz(unsigned long ticks)
3331 long delta, active, n;
3333 if (time_before(jiffies, calc_load_update))
3334 return;
3337 * If we crossed a calc_load_update boundary, make sure to fold
3338 * any pending idle changes, the respective CPUs might have
3339 * missed the tick driven calc_load_account_active() update
3340 * due to NO_HZ.
3342 delta = calc_load_fold_idle();
3343 if (delta)
3344 atomic_long_add(delta, &calc_load_tasks);
3347 * If we were idle for multiple load cycles, apply them.
3349 if (ticks >= LOAD_FREQ) {
3350 n = ticks / LOAD_FREQ;
3352 active = atomic_long_read(&calc_load_tasks);
3353 active = active > 0 ? active * FIXED_1 : 0;
3355 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3356 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3357 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3359 calc_load_update += n * LOAD_FREQ;
3363 * Its possible the remainder of the above division also crosses
3364 * a LOAD_FREQ period, the regular check in calc_global_load()
3365 * which comes after this will take care of that.
3367 * Consider us being 11 ticks before a cycle completion, and us
3368 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3369 * age us 4 cycles, and the test in calc_global_load() will
3370 * pick up the final one.
3373 #else
3374 static void calc_load_account_idle(struct rq *this_rq)
3378 static inline long calc_load_fold_idle(void)
3380 return 0;
3383 static void calc_global_nohz(unsigned long ticks)
3386 #endif
3389 * get_avenrun - get the load average array
3390 * @loads: pointer to dest load array
3391 * @offset: offset to add
3392 * @shift: shift count to shift the result left
3394 * These values are estimates at best, so no need for locking.
3396 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3398 loads[0] = (avenrun[0] + offset) << shift;
3399 loads[1] = (avenrun[1] + offset) << shift;
3400 loads[2] = (avenrun[2] + offset) << shift;
3404 * calc_load - update the avenrun load estimates 10 ticks after the
3405 * CPUs have updated calc_load_tasks.
3407 void calc_global_load(unsigned long ticks)
3409 long active;
3411 calc_global_nohz(ticks);
3413 if (time_before(jiffies, calc_load_update + 10))
3414 return;
3416 active = atomic_long_read(&calc_load_tasks);
3417 active = active > 0 ? active * FIXED_1 : 0;
3419 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3420 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3421 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3423 calc_load_update += LOAD_FREQ;
3427 * Called from update_cpu_load() to periodically update this CPU's
3428 * active count.
3430 static void calc_load_account_active(struct rq *this_rq)
3432 long delta;
3434 if (time_before(jiffies, this_rq->calc_load_update))
3435 return;
3437 delta = calc_load_fold_active(this_rq);
3438 delta += calc_load_fold_idle();
3439 if (delta)
3440 atomic_long_add(delta, &calc_load_tasks);
3442 this_rq->calc_load_update += LOAD_FREQ;
3446 * The exact cpuload at various idx values, calculated at every tick would be
3447 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3449 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3450 * on nth tick when cpu may be busy, then we have:
3451 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3452 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3454 * decay_load_missed() below does efficient calculation of
3455 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3456 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3458 * The calculation is approximated on a 128 point scale.
3459 * degrade_zero_ticks is the number of ticks after which load at any
3460 * particular idx is approximated to be zero.
3461 * degrade_factor is a precomputed table, a row for each load idx.
3462 * Each column corresponds to degradation factor for a power of two ticks,
3463 * based on 128 point scale.
3464 * Example:
3465 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3466 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3468 * With this power of 2 load factors, we can degrade the load n times
3469 * by looking at 1 bits in n and doing as many mult/shift instead of
3470 * n mult/shifts needed by the exact degradation.
3472 #define DEGRADE_SHIFT 7
3473 static const unsigned char
3474 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3475 static const unsigned char
3476 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3477 {0, 0, 0, 0, 0, 0, 0, 0},
3478 {64, 32, 8, 0, 0, 0, 0, 0},
3479 {96, 72, 40, 12, 1, 0, 0},
3480 {112, 98, 75, 43, 15, 1, 0},
3481 {120, 112, 98, 76, 45, 16, 2} };
3484 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3485 * would be when CPU is idle and so we just decay the old load without
3486 * adding any new load.
3488 static unsigned long
3489 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3491 int j = 0;
3493 if (!missed_updates)
3494 return load;
3496 if (missed_updates >= degrade_zero_ticks[idx])
3497 return 0;
3499 if (idx == 1)
3500 return load >> missed_updates;
3502 while (missed_updates) {
3503 if (missed_updates % 2)
3504 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3506 missed_updates >>= 1;
3507 j++;
3509 return load;
3513 * Update rq->cpu_load[] statistics. This function is usually called every
3514 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3515 * every tick. We fix it up based on jiffies.
3517 static void update_cpu_load(struct rq *this_rq)
3519 unsigned long this_load = this_rq->load.weight;
3520 unsigned long curr_jiffies = jiffies;
3521 unsigned long pending_updates;
3522 int i, scale;
3524 this_rq->nr_load_updates++;
3526 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3527 if (curr_jiffies == this_rq->last_load_update_tick)
3528 return;
3530 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3531 this_rq->last_load_update_tick = curr_jiffies;
3533 /* Update our load: */
3534 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3535 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3536 unsigned long old_load, new_load;
3538 /* scale is effectively 1 << i now, and >> i divides by scale */
3540 old_load = this_rq->cpu_load[i];
3541 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3542 new_load = this_load;
3544 * Round up the averaging division if load is increasing. This
3545 * prevents us from getting stuck on 9 if the load is 10, for
3546 * example.
3548 if (new_load > old_load)
3549 new_load += scale - 1;
3551 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3554 sched_avg_update(this_rq);
3557 static void update_cpu_load_active(struct rq *this_rq)
3559 update_cpu_load(this_rq);
3561 calc_load_account_active(this_rq);
3564 #ifdef CONFIG_SMP
3567 * sched_exec - execve() is a valuable balancing opportunity, because at
3568 * this point the task has the smallest effective memory and cache footprint.
3570 void sched_exec(void)
3572 struct task_struct *p = current;
3573 unsigned long flags;
3574 int dest_cpu;
3576 raw_spin_lock_irqsave(&p->pi_lock, flags);
3577 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3578 if (dest_cpu == smp_processor_id())
3579 goto unlock;
3581 if (likely(cpu_active(dest_cpu))) {
3582 struct migration_arg arg = { p, dest_cpu };
3584 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3585 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3586 return;
3588 unlock:
3589 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3592 #endif
3594 DEFINE_PER_CPU(struct kernel_stat, kstat);
3596 EXPORT_PER_CPU_SYMBOL(kstat);
3599 * Return any ns on the sched_clock that have not yet been accounted in
3600 * @p in case that task is currently running.
3602 * Called with task_rq_lock() held on @rq.
3604 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3606 u64 ns = 0;
3608 if (task_current(rq, p)) {
3609 update_rq_clock(rq);
3610 ns = rq->clock_task - p->se.exec_start;
3611 if ((s64)ns < 0)
3612 ns = 0;
3615 return ns;
3618 unsigned long long task_delta_exec(struct task_struct *p)
3620 unsigned long flags;
3621 struct rq *rq;
3622 u64 ns = 0;
3624 rq = task_rq_lock(p, &flags);
3625 ns = do_task_delta_exec(p, rq);
3626 task_rq_unlock(rq, p, &flags);
3628 return ns;
3632 * Return accounted runtime for the task.
3633 * In case the task is currently running, return the runtime plus current's
3634 * pending runtime that have not been accounted yet.
3636 unsigned long long task_sched_runtime(struct task_struct *p)
3638 unsigned long flags;
3639 struct rq *rq;
3640 u64 ns = 0;
3642 rq = task_rq_lock(p, &flags);
3643 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3644 task_rq_unlock(rq, p, &flags);
3646 return ns;
3650 * Return sum_exec_runtime for the thread group.
3651 * In case the task is currently running, return the sum plus current's
3652 * pending runtime that have not been accounted yet.
3654 * Note that the thread group might have other running tasks as well,
3655 * so the return value not includes other pending runtime that other
3656 * running tasks might have.
3658 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3660 struct task_cputime totals;
3661 unsigned long flags;
3662 struct rq *rq;
3663 u64 ns;
3665 rq = task_rq_lock(p, &flags);
3666 thread_group_cputime(p, &totals);
3667 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3668 task_rq_unlock(rq, p, &flags);
3670 return ns;
3674 * Account user cpu time to a process.
3675 * @p: the process that the cpu time gets accounted to
3676 * @cputime: the cpu time spent in user space since the last update
3677 * @cputime_scaled: cputime scaled by cpu frequency
3679 void account_user_time(struct task_struct *p, cputime_t cputime,
3680 cputime_t cputime_scaled)
3682 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3683 cputime64_t tmp;
3685 /* Add user time to process. */
3686 p->utime = cputime_add(p->utime, cputime);
3687 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3688 account_group_user_time(p, cputime);
3690 /* Add user time to cpustat. */
3691 tmp = cputime_to_cputime64(cputime);
3692 if (TASK_NICE(p) > 0)
3693 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3694 else
3695 cpustat->user = cputime64_add(cpustat->user, tmp);
3697 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3698 /* Account for user time used */
3699 acct_update_integrals(p);
3703 * Account guest cpu time to a process.
3704 * @p: the process that the cpu time gets accounted to
3705 * @cputime: the cpu time spent in virtual machine since the last update
3706 * @cputime_scaled: cputime scaled by cpu frequency
3708 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3709 cputime_t cputime_scaled)
3711 cputime64_t tmp;
3712 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3714 tmp = cputime_to_cputime64(cputime);
3716 /* Add guest time to process. */
3717 p->utime = cputime_add(p->utime, cputime);
3718 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3719 account_group_user_time(p, cputime);
3720 p->gtime = cputime_add(p->gtime, cputime);
3722 /* Add guest time to cpustat. */
3723 if (TASK_NICE(p) > 0) {
3724 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3725 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3726 } else {
3727 cpustat->user = cputime64_add(cpustat->user, tmp);
3728 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3733 * Account system cpu time to a process and desired cpustat field
3734 * @p: the process that the cpu time gets accounted to
3735 * @cputime: the cpu time spent in kernel space since the last update
3736 * @cputime_scaled: cputime scaled by cpu frequency
3737 * @target_cputime64: pointer to cpustat field that has to be updated
3739 static inline
3740 void __account_system_time(struct task_struct *p, cputime_t cputime,
3741 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3743 cputime64_t tmp = cputime_to_cputime64(cputime);
3745 /* Add system time to process. */
3746 p->stime = cputime_add(p->stime, cputime);
3747 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3748 account_group_system_time(p, cputime);
3750 /* Add system time to cpustat. */
3751 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3752 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3754 /* Account for system time used */
3755 acct_update_integrals(p);
3759 * Account system cpu time to a process.
3760 * @p: the process that the cpu time gets accounted to
3761 * @hardirq_offset: the offset to subtract from hardirq_count()
3762 * @cputime: the cpu time spent in kernel space since the last update
3763 * @cputime_scaled: cputime scaled by cpu frequency
3765 void account_system_time(struct task_struct *p, int hardirq_offset,
3766 cputime_t cputime, cputime_t cputime_scaled)
3768 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3769 cputime64_t *target_cputime64;
3771 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3772 account_guest_time(p, cputime, cputime_scaled);
3773 return;
3776 if (hardirq_count() - hardirq_offset)
3777 target_cputime64 = &cpustat->irq;
3778 else if (in_serving_softirq())
3779 target_cputime64 = &cpustat->softirq;
3780 else
3781 target_cputime64 = &cpustat->system;
3783 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3787 * Account for involuntary wait time.
3788 * @cputime: the cpu time spent in involuntary wait
3790 void account_steal_time(cputime_t cputime)
3792 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3793 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3795 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3799 * Account for idle time.
3800 * @cputime: the cpu time spent in idle wait
3802 void account_idle_time(cputime_t cputime)
3804 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3805 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3806 struct rq *rq = this_rq();
3808 if (atomic_read(&rq->nr_iowait) > 0)
3809 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3810 else
3811 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3814 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3816 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3818 * Account a tick to a process and cpustat
3819 * @p: the process that the cpu time gets accounted to
3820 * @user_tick: is the tick from userspace
3821 * @rq: the pointer to rq
3823 * Tick demultiplexing follows the order
3824 * - pending hardirq update
3825 * - pending softirq update
3826 * - user_time
3827 * - idle_time
3828 * - system time
3829 * - check for guest_time
3830 * - else account as system_time
3832 * Check for hardirq is done both for system and user time as there is
3833 * no timer going off while we are on hardirq and hence we may never get an
3834 * opportunity to update it solely in system time.
3835 * p->stime and friends are only updated on system time and not on irq
3836 * softirq as those do not count in task exec_runtime any more.
3838 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3839 struct rq *rq)
3841 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3842 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3843 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3845 if (irqtime_account_hi_update()) {
3846 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3847 } else if (irqtime_account_si_update()) {
3848 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3849 } else if (this_cpu_ksoftirqd() == p) {
3851 * ksoftirqd time do not get accounted in cpu_softirq_time.
3852 * So, we have to handle it separately here.
3853 * Also, p->stime needs to be updated for ksoftirqd.
3855 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3856 &cpustat->softirq);
3857 } else if (user_tick) {
3858 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3859 } else if (p == rq->idle) {
3860 account_idle_time(cputime_one_jiffy);
3861 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3862 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3863 } else {
3864 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3865 &cpustat->system);
3869 static void irqtime_account_idle_ticks(int ticks)
3871 int i;
3872 struct rq *rq = this_rq();
3874 for (i = 0; i < ticks; i++)
3875 irqtime_account_process_tick(current, 0, rq);
3877 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3878 static void irqtime_account_idle_ticks(int ticks) {}
3879 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3880 struct rq *rq) {}
3881 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3884 * Account a single tick of cpu time.
3885 * @p: the process that the cpu time gets accounted to
3886 * @user_tick: indicates if the tick is a user or a system tick
3888 void account_process_tick(struct task_struct *p, int user_tick)
3890 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3891 struct rq *rq = this_rq();
3893 if (sched_clock_irqtime) {
3894 irqtime_account_process_tick(p, user_tick, rq);
3895 return;
3898 if (user_tick)
3899 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3900 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3901 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3902 one_jiffy_scaled);
3903 else
3904 account_idle_time(cputime_one_jiffy);
3908 * Account multiple ticks of steal time.
3909 * @p: the process from which the cpu time has been stolen
3910 * @ticks: number of stolen ticks
3912 void account_steal_ticks(unsigned long ticks)
3914 account_steal_time(jiffies_to_cputime(ticks));
3918 * Account multiple ticks of idle time.
3919 * @ticks: number of stolen ticks
3921 void account_idle_ticks(unsigned long ticks)
3924 if (sched_clock_irqtime) {
3925 irqtime_account_idle_ticks(ticks);
3926 return;
3929 account_idle_time(jiffies_to_cputime(ticks));
3932 #endif
3935 * Use precise platform statistics if available:
3937 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3938 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3940 *ut = p->utime;
3941 *st = p->stime;
3944 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3946 struct task_cputime cputime;
3948 thread_group_cputime(p, &cputime);
3950 *ut = cputime.utime;
3951 *st = cputime.stime;
3953 #else
3955 #ifndef nsecs_to_cputime
3956 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3957 #endif
3959 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3961 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3964 * Use CFS's precise accounting:
3966 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3968 if (total) {
3969 u64 temp = rtime;
3971 temp *= utime;
3972 do_div(temp, total);
3973 utime = (cputime_t)temp;
3974 } else
3975 utime = rtime;
3978 * Compare with previous values, to keep monotonicity:
3980 p->prev_utime = max(p->prev_utime, utime);
3981 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3983 *ut = p->prev_utime;
3984 *st = p->prev_stime;
3988 * Must be called with siglock held.
3990 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3992 struct signal_struct *sig = p->signal;
3993 struct task_cputime cputime;
3994 cputime_t rtime, utime, total;
3996 thread_group_cputime(p, &cputime);
3998 total = cputime_add(cputime.utime, cputime.stime);
3999 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4001 if (total) {
4002 u64 temp = rtime;
4004 temp *= cputime.utime;
4005 do_div(temp, total);
4006 utime = (cputime_t)temp;
4007 } else
4008 utime = rtime;
4010 sig->prev_utime = max(sig->prev_utime, utime);
4011 sig->prev_stime = max(sig->prev_stime,
4012 cputime_sub(rtime, sig->prev_utime));
4014 *ut = sig->prev_utime;
4015 *st = sig->prev_stime;
4017 #endif
4020 * This function gets called by the timer code, with HZ frequency.
4021 * We call it with interrupts disabled.
4023 void scheduler_tick(void)
4025 int cpu = smp_processor_id();
4026 struct rq *rq = cpu_rq(cpu);
4027 struct task_struct *curr = rq->curr;
4029 sched_clock_tick();
4031 raw_spin_lock(&rq->lock);
4032 update_rq_clock(rq);
4033 update_cpu_load_active(rq);
4034 curr->sched_class->task_tick(rq, curr, 0);
4035 raw_spin_unlock(&rq->lock);
4037 perf_event_task_tick();
4039 #ifdef CONFIG_SMP
4040 rq->idle_at_tick = idle_cpu(cpu);
4041 trigger_load_balance(rq, cpu);
4042 #endif
4045 notrace unsigned long get_parent_ip(unsigned long addr)
4047 if (in_lock_functions(addr)) {
4048 addr = CALLER_ADDR2;
4049 if (in_lock_functions(addr))
4050 addr = CALLER_ADDR3;
4052 return addr;
4055 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4056 defined(CONFIG_PREEMPT_TRACER))
4058 void __kprobes add_preempt_count(int val)
4060 #ifdef CONFIG_DEBUG_PREEMPT
4062 * Underflow?
4064 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4065 return;
4066 #endif
4067 preempt_count() += val;
4068 #ifdef CONFIG_DEBUG_PREEMPT
4070 * Spinlock count overflowing soon?
4072 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4073 PREEMPT_MASK - 10);
4074 #endif
4075 if (preempt_count() == val)
4076 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4078 EXPORT_SYMBOL(add_preempt_count);
4080 void __kprobes sub_preempt_count(int val)
4082 #ifdef CONFIG_DEBUG_PREEMPT
4084 * Underflow?
4086 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4087 return;
4089 * Is the spinlock portion underflowing?
4091 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4092 !(preempt_count() & PREEMPT_MASK)))
4093 return;
4094 #endif
4096 if (preempt_count() == val)
4097 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4098 preempt_count() -= val;
4100 EXPORT_SYMBOL(sub_preempt_count);
4102 #endif
4105 * Print scheduling while atomic bug:
4107 static noinline void __schedule_bug(struct task_struct *prev)
4109 struct pt_regs *regs = get_irq_regs();
4111 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4112 prev->comm, prev->pid, preempt_count());
4114 debug_show_held_locks(prev);
4115 print_modules();
4116 if (irqs_disabled())
4117 print_irqtrace_events(prev);
4119 if (regs)
4120 show_regs(regs);
4121 else
4122 dump_stack();
4126 * Various schedule()-time debugging checks and statistics:
4128 static inline void schedule_debug(struct task_struct *prev)
4131 * Test if we are atomic. Since do_exit() needs to call into
4132 * schedule() atomically, we ignore that path for now.
4133 * Otherwise, whine if we are scheduling when we should not be.
4135 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4136 __schedule_bug(prev);
4138 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4140 schedstat_inc(this_rq(), sched_count);
4143 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4145 if (prev->on_rq || rq->skip_clock_update < 0)
4146 update_rq_clock(rq);
4147 prev->sched_class->put_prev_task(rq, prev);
4151 * Pick up the highest-prio task:
4153 static inline struct task_struct *
4154 pick_next_task(struct rq *rq)
4156 const struct sched_class *class;
4157 struct task_struct *p;
4160 * Optimization: we know that if all tasks are in
4161 * the fair class we can call that function directly:
4163 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4164 p = fair_sched_class.pick_next_task(rq);
4165 if (likely(p))
4166 return p;
4169 for_each_class(class) {
4170 p = class->pick_next_task(rq);
4171 if (p)
4172 return p;
4175 BUG(); /* the idle class will always have a runnable task */
4179 * schedule() is the main scheduler function.
4181 asmlinkage void __sched schedule(void)
4183 struct task_struct *prev, *next;
4184 unsigned long *switch_count;
4185 struct rq *rq;
4186 int cpu;
4188 need_resched:
4189 preempt_disable();
4190 cpu = smp_processor_id();
4191 rq = cpu_rq(cpu);
4192 rcu_note_context_switch(cpu);
4193 prev = rq->curr;
4195 schedule_debug(prev);
4197 if (sched_feat(HRTICK))
4198 hrtick_clear(rq);
4200 raw_spin_lock_irq(&rq->lock);
4202 switch_count = &prev->nivcsw;
4203 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4204 if (unlikely(signal_pending_state(prev->state, prev))) {
4205 prev->state = TASK_RUNNING;
4206 } else {
4207 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4208 prev->on_rq = 0;
4211 * If a worker went to sleep, notify and ask workqueue
4212 * whether it wants to wake up a task to maintain
4213 * concurrency.
4215 if (prev->flags & PF_WQ_WORKER) {
4216 struct task_struct *to_wakeup;
4218 to_wakeup = wq_worker_sleeping(prev, cpu);
4219 if (to_wakeup)
4220 try_to_wake_up_local(to_wakeup);
4224 * If we are going to sleep and we have plugged IO
4225 * queued, make sure to submit it to avoid deadlocks.
4227 if (blk_needs_flush_plug(prev)) {
4228 raw_spin_unlock(&rq->lock);
4229 blk_schedule_flush_plug(prev);
4230 raw_spin_lock(&rq->lock);
4233 switch_count = &prev->nvcsw;
4236 pre_schedule(rq, prev);
4238 if (unlikely(!rq->nr_running))
4239 idle_balance(cpu, rq);
4241 put_prev_task(rq, prev);
4242 next = pick_next_task(rq);
4243 clear_tsk_need_resched(prev);
4244 rq->skip_clock_update = 0;
4246 if (likely(prev != next)) {
4247 rq->nr_switches++;
4248 rq->curr = next;
4249 ++*switch_count;
4251 context_switch(rq, prev, next); /* unlocks the rq */
4253 * The context switch have flipped the stack from under us
4254 * and restored the local variables which were saved when
4255 * this task called schedule() in the past. prev == current
4256 * is still correct, but it can be moved to another cpu/rq.
4258 cpu = smp_processor_id();
4259 rq = cpu_rq(cpu);
4260 } else
4261 raw_spin_unlock_irq(&rq->lock);
4263 post_schedule(rq);
4265 preempt_enable_no_resched();
4266 if (need_resched())
4267 goto need_resched;
4269 EXPORT_SYMBOL(schedule);
4271 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4273 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4275 bool ret = false;
4277 rcu_read_lock();
4278 if (lock->owner != owner)
4279 goto fail;
4282 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4283 * lock->owner still matches owner, if that fails, owner might
4284 * point to free()d memory, if it still matches, the rcu_read_lock()
4285 * ensures the memory stays valid.
4287 barrier();
4289 ret = owner->on_cpu;
4290 fail:
4291 rcu_read_unlock();
4293 return ret;
4297 * Look out! "owner" is an entirely speculative pointer
4298 * access and not reliable.
4300 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4302 if (!sched_feat(OWNER_SPIN))
4303 return 0;
4305 while (owner_running(lock, owner)) {
4306 if (need_resched())
4307 return 0;
4309 arch_mutex_cpu_relax();
4313 * If the owner changed to another task there is likely
4314 * heavy contention, stop spinning.
4316 if (lock->owner)
4317 return 0;
4319 return 1;
4321 #endif
4323 #ifdef CONFIG_PREEMPT
4325 * this is the entry point to schedule() from in-kernel preemption
4326 * off of preempt_enable. Kernel preemptions off return from interrupt
4327 * occur there and call schedule directly.
4329 asmlinkage void __sched notrace preempt_schedule(void)
4331 struct thread_info *ti = current_thread_info();
4334 * If there is a non-zero preempt_count or interrupts are disabled,
4335 * we do not want to preempt the current task. Just return..
4337 if (likely(ti->preempt_count || irqs_disabled()))
4338 return;
4340 do {
4341 add_preempt_count_notrace(PREEMPT_ACTIVE);
4342 schedule();
4343 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4346 * Check again in case we missed a preemption opportunity
4347 * between schedule and now.
4349 barrier();
4350 } while (need_resched());
4352 EXPORT_SYMBOL(preempt_schedule);
4355 * this is the entry point to schedule() from kernel preemption
4356 * off of irq context.
4357 * Note, that this is called and return with irqs disabled. This will
4358 * protect us against recursive calling from irq.
4360 asmlinkage void __sched preempt_schedule_irq(void)
4362 struct thread_info *ti = current_thread_info();
4364 /* Catch callers which need to be fixed */
4365 BUG_ON(ti->preempt_count || !irqs_disabled());
4367 do {
4368 add_preempt_count(PREEMPT_ACTIVE);
4369 local_irq_enable();
4370 schedule();
4371 local_irq_disable();
4372 sub_preempt_count(PREEMPT_ACTIVE);
4375 * Check again in case we missed a preemption opportunity
4376 * between schedule and now.
4378 barrier();
4379 } while (need_resched());
4382 #endif /* CONFIG_PREEMPT */
4384 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4385 void *key)
4387 return try_to_wake_up(curr->private, mode, wake_flags);
4389 EXPORT_SYMBOL(default_wake_function);
4392 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4393 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4394 * number) then we wake all the non-exclusive tasks and one exclusive task.
4396 * There are circumstances in which we can try to wake a task which has already
4397 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4398 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4400 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4401 int nr_exclusive, int wake_flags, void *key)
4403 wait_queue_t *curr, *next;
4405 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4406 unsigned flags = curr->flags;
4408 if (curr->func(curr, mode, wake_flags, key) &&
4409 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4410 break;
4415 * __wake_up - wake up threads blocked on a waitqueue.
4416 * @q: the waitqueue
4417 * @mode: which threads
4418 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4419 * @key: is directly passed to the wakeup function
4421 * It may be assumed that this function implies a write memory barrier before
4422 * changing the task state if and only if any tasks are woken up.
4424 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4425 int nr_exclusive, void *key)
4427 unsigned long flags;
4429 spin_lock_irqsave(&q->lock, flags);
4430 __wake_up_common(q, mode, nr_exclusive, 0, key);
4431 spin_unlock_irqrestore(&q->lock, flags);
4433 EXPORT_SYMBOL(__wake_up);
4436 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4438 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4440 __wake_up_common(q, mode, 1, 0, NULL);
4442 EXPORT_SYMBOL_GPL(__wake_up_locked);
4444 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4446 __wake_up_common(q, mode, 1, 0, key);
4448 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4451 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4452 * @q: the waitqueue
4453 * @mode: which threads
4454 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4455 * @key: opaque value to be passed to wakeup targets
4457 * The sync wakeup differs that the waker knows that it will schedule
4458 * away soon, so while the target thread will be woken up, it will not
4459 * be migrated to another CPU - ie. the two threads are 'synchronized'
4460 * with each other. This can prevent needless bouncing between CPUs.
4462 * On UP it can prevent extra preemption.
4464 * It may be assumed that this function implies a write memory barrier before
4465 * changing the task state if and only if any tasks are woken up.
4467 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4468 int nr_exclusive, void *key)
4470 unsigned long flags;
4471 int wake_flags = WF_SYNC;
4473 if (unlikely(!q))
4474 return;
4476 if (unlikely(!nr_exclusive))
4477 wake_flags = 0;
4479 spin_lock_irqsave(&q->lock, flags);
4480 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4481 spin_unlock_irqrestore(&q->lock, flags);
4483 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4486 * __wake_up_sync - see __wake_up_sync_key()
4488 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4490 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4492 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4495 * complete: - signals a single thread waiting on this completion
4496 * @x: holds the state of this particular completion
4498 * This will wake up a single thread waiting on this completion. Threads will be
4499 * awakened in the same order in which they were queued.
4501 * See also complete_all(), wait_for_completion() and related routines.
4503 * It may be assumed that this function implies a write memory barrier before
4504 * changing the task state if and only if any tasks are woken up.
4506 void complete(struct completion *x)
4508 unsigned long flags;
4510 spin_lock_irqsave(&x->wait.lock, flags);
4511 x->done++;
4512 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4513 spin_unlock_irqrestore(&x->wait.lock, flags);
4515 EXPORT_SYMBOL(complete);
4518 * complete_all: - signals all threads waiting on this completion
4519 * @x: holds the state of this particular completion
4521 * This will wake up all threads waiting on this particular completion event.
4523 * It may be assumed that this function implies a write memory barrier before
4524 * changing the task state if and only if any tasks are woken up.
4526 void complete_all(struct completion *x)
4528 unsigned long flags;
4530 spin_lock_irqsave(&x->wait.lock, flags);
4531 x->done += UINT_MAX/2;
4532 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4533 spin_unlock_irqrestore(&x->wait.lock, flags);
4535 EXPORT_SYMBOL(complete_all);
4537 static inline long __sched
4538 do_wait_for_common(struct completion *x, long timeout, int state)
4540 if (!x->done) {
4541 DECLARE_WAITQUEUE(wait, current);
4543 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4544 do {
4545 if (signal_pending_state(state, current)) {
4546 timeout = -ERESTARTSYS;
4547 break;
4549 __set_current_state(state);
4550 spin_unlock_irq(&x->wait.lock);
4551 timeout = schedule_timeout(timeout);
4552 spin_lock_irq(&x->wait.lock);
4553 } while (!x->done && timeout);
4554 __remove_wait_queue(&x->wait, &wait);
4555 if (!x->done)
4556 return timeout;
4558 x->done--;
4559 return timeout ?: 1;
4562 static long __sched
4563 wait_for_common(struct completion *x, long timeout, int state)
4565 might_sleep();
4567 spin_lock_irq(&x->wait.lock);
4568 timeout = do_wait_for_common(x, timeout, state);
4569 spin_unlock_irq(&x->wait.lock);
4570 return timeout;
4574 * wait_for_completion: - waits for completion of a task
4575 * @x: holds the state of this particular completion
4577 * This waits to be signaled for completion of a specific task. It is NOT
4578 * interruptible and there is no timeout.
4580 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4581 * and interrupt capability. Also see complete().
4583 void __sched wait_for_completion(struct completion *x)
4585 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4587 EXPORT_SYMBOL(wait_for_completion);
4590 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4591 * @x: holds the state of this particular completion
4592 * @timeout: timeout value in jiffies
4594 * This waits for either a completion of a specific task to be signaled or for a
4595 * specified timeout to expire. The timeout is in jiffies. It is not
4596 * interruptible.
4598 unsigned long __sched
4599 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4601 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4603 EXPORT_SYMBOL(wait_for_completion_timeout);
4606 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4607 * @x: holds the state of this particular completion
4609 * This waits for completion of a specific task to be signaled. It is
4610 * interruptible.
4612 int __sched wait_for_completion_interruptible(struct completion *x)
4614 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4615 if (t == -ERESTARTSYS)
4616 return t;
4617 return 0;
4619 EXPORT_SYMBOL(wait_for_completion_interruptible);
4622 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4623 * @x: holds the state of this particular completion
4624 * @timeout: timeout value in jiffies
4626 * This waits for either a completion of a specific task to be signaled or for a
4627 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4629 long __sched
4630 wait_for_completion_interruptible_timeout(struct completion *x,
4631 unsigned long timeout)
4633 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4635 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4638 * wait_for_completion_killable: - waits for completion of a task (killable)
4639 * @x: holds the state of this particular completion
4641 * This waits to be signaled for completion of a specific task. It can be
4642 * interrupted by a kill signal.
4644 int __sched wait_for_completion_killable(struct completion *x)
4646 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4647 if (t == -ERESTARTSYS)
4648 return t;
4649 return 0;
4651 EXPORT_SYMBOL(wait_for_completion_killable);
4654 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4655 * @x: holds the state of this particular completion
4656 * @timeout: timeout value in jiffies
4658 * This waits for either a completion of a specific task to be
4659 * signaled or for a specified timeout to expire. It can be
4660 * interrupted by a kill signal. The timeout is in jiffies.
4662 long __sched
4663 wait_for_completion_killable_timeout(struct completion *x,
4664 unsigned long timeout)
4666 return wait_for_common(x, timeout, TASK_KILLABLE);
4668 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4671 * try_wait_for_completion - try to decrement a completion without blocking
4672 * @x: completion structure
4674 * Returns: 0 if a decrement cannot be done without blocking
4675 * 1 if a decrement succeeded.
4677 * If a completion is being used as a counting completion,
4678 * attempt to decrement the counter without blocking. This
4679 * enables us to avoid waiting if the resource the completion
4680 * is protecting is not available.
4682 bool try_wait_for_completion(struct completion *x)
4684 unsigned long flags;
4685 int ret = 1;
4687 spin_lock_irqsave(&x->wait.lock, flags);
4688 if (!x->done)
4689 ret = 0;
4690 else
4691 x->done--;
4692 spin_unlock_irqrestore(&x->wait.lock, flags);
4693 return ret;
4695 EXPORT_SYMBOL(try_wait_for_completion);
4698 * completion_done - Test to see if a completion has any waiters
4699 * @x: completion structure
4701 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4702 * 1 if there are no waiters.
4705 bool completion_done(struct completion *x)
4707 unsigned long flags;
4708 int ret = 1;
4710 spin_lock_irqsave(&x->wait.lock, flags);
4711 if (!x->done)
4712 ret = 0;
4713 spin_unlock_irqrestore(&x->wait.lock, flags);
4714 return ret;
4716 EXPORT_SYMBOL(completion_done);
4718 static long __sched
4719 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4721 unsigned long flags;
4722 wait_queue_t wait;
4724 init_waitqueue_entry(&wait, current);
4726 __set_current_state(state);
4728 spin_lock_irqsave(&q->lock, flags);
4729 __add_wait_queue(q, &wait);
4730 spin_unlock(&q->lock);
4731 timeout = schedule_timeout(timeout);
4732 spin_lock_irq(&q->lock);
4733 __remove_wait_queue(q, &wait);
4734 spin_unlock_irqrestore(&q->lock, flags);
4736 return timeout;
4739 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4741 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4743 EXPORT_SYMBOL(interruptible_sleep_on);
4745 long __sched
4746 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4748 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4750 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4752 void __sched sleep_on(wait_queue_head_t *q)
4754 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4756 EXPORT_SYMBOL(sleep_on);
4758 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4760 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4762 EXPORT_SYMBOL(sleep_on_timeout);
4764 #ifdef CONFIG_RT_MUTEXES
4767 * rt_mutex_setprio - set the current priority of a task
4768 * @p: task
4769 * @prio: prio value (kernel-internal form)
4771 * This function changes the 'effective' priority of a task. It does
4772 * not touch ->normal_prio like __setscheduler().
4774 * Used by the rt_mutex code to implement priority inheritance logic.
4776 void rt_mutex_setprio(struct task_struct *p, int prio)
4778 int oldprio, on_rq, running;
4779 struct rq *rq;
4780 const struct sched_class *prev_class;
4782 BUG_ON(prio < 0 || prio > MAX_PRIO);
4784 rq = __task_rq_lock(p);
4786 trace_sched_pi_setprio(p, prio);
4787 oldprio = p->prio;
4788 prev_class = p->sched_class;
4789 on_rq = p->on_rq;
4790 running = task_current(rq, p);
4791 if (on_rq)
4792 dequeue_task(rq, p, 0);
4793 if (running)
4794 p->sched_class->put_prev_task(rq, p);
4796 if (rt_prio(prio))
4797 p->sched_class = &rt_sched_class;
4798 else
4799 p->sched_class = &fair_sched_class;
4801 p->prio = prio;
4803 if (running)
4804 p->sched_class->set_curr_task(rq);
4805 if (on_rq)
4806 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4808 check_class_changed(rq, p, prev_class, oldprio);
4809 __task_rq_unlock(rq);
4812 #endif
4814 void set_user_nice(struct task_struct *p, long nice)
4816 int old_prio, delta, on_rq;
4817 unsigned long flags;
4818 struct rq *rq;
4820 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4821 return;
4823 * We have to be careful, if called from sys_setpriority(),
4824 * the task might be in the middle of scheduling on another CPU.
4826 rq = task_rq_lock(p, &flags);
4828 * The RT priorities are set via sched_setscheduler(), but we still
4829 * allow the 'normal' nice value to be set - but as expected
4830 * it wont have any effect on scheduling until the task is
4831 * SCHED_FIFO/SCHED_RR:
4833 if (task_has_rt_policy(p)) {
4834 p->static_prio = NICE_TO_PRIO(nice);
4835 goto out_unlock;
4837 on_rq = p->on_rq;
4838 if (on_rq)
4839 dequeue_task(rq, p, 0);
4841 p->static_prio = NICE_TO_PRIO(nice);
4842 set_load_weight(p);
4843 old_prio = p->prio;
4844 p->prio = effective_prio(p);
4845 delta = p->prio - old_prio;
4847 if (on_rq) {
4848 enqueue_task(rq, p, 0);
4850 * If the task increased its priority or is running and
4851 * lowered its priority, then reschedule its CPU:
4853 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4854 resched_task(rq->curr);
4856 out_unlock:
4857 task_rq_unlock(rq, p, &flags);
4859 EXPORT_SYMBOL(set_user_nice);
4862 * can_nice - check if a task can reduce its nice value
4863 * @p: task
4864 * @nice: nice value
4866 int can_nice(const struct task_struct *p, const int nice)
4868 /* convert nice value [19,-20] to rlimit style value [1,40] */
4869 int nice_rlim = 20 - nice;
4871 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4872 capable(CAP_SYS_NICE));
4875 #ifdef __ARCH_WANT_SYS_NICE
4878 * sys_nice - change the priority of the current process.
4879 * @increment: priority increment
4881 * sys_setpriority is a more generic, but much slower function that
4882 * does similar things.
4884 SYSCALL_DEFINE1(nice, int, increment)
4886 long nice, retval;
4889 * Setpriority might change our priority at the same moment.
4890 * We don't have to worry. Conceptually one call occurs first
4891 * and we have a single winner.
4893 if (increment < -40)
4894 increment = -40;
4895 if (increment > 40)
4896 increment = 40;
4898 nice = TASK_NICE(current) + increment;
4899 if (nice < -20)
4900 nice = -20;
4901 if (nice > 19)
4902 nice = 19;
4904 if (increment < 0 && !can_nice(current, nice))
4905 return -EPERM;
4907 retval = security_task_setnice(current, nice);
4908 if (retval)
4909 return retval;
4911 set_user_nice(current, nice);
4912 return 0;
4915 #endif
4918 * task_prio - return the priority value of a given task.
4919 * @p: the task in question.
4921 * This is the priority value as seen by users in /proc.
4922 * RT tasks are offset by -200. Normal tasks are centered
4923 * around 0, value goes from -16 to +15.
4925 int task_prio(const struct task_struct *p)
4927 return p->prio - MAX_RT_PRIO;
4931 * task_nice - return the nice value of a given task.
4932 * @p: the task in question.
4934 int task_nice(const struct task_struct *p)
4936 return TASK_NICE(p);
4938 EXPORT_SYMBOL(task_nice);
4941 * idle_cpu - is a given cpu idle currently?
4942 * @cpu: the processor in question.
4944 int idle_cpu(int cpu)
4946 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4950 * idle_task - return the idle task for a given cpu.
4951 * @cpu: the processor in question.
4953 struct task_struct *idle_task(int cpu)
4955 return cpu_rq(cpu)->idle;
4959 * find_process_by_pid - find a process with a matching PID value.
4960 * @pid: the pid in question.
4962 static struct task_struct *find_process_by_pid(pid_t pid)
4964 return pid ? find_task_by_vpid(pid) : current;
4967 /* Actually do priority change: must hold rq lock. */
4968 static void
4969 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4971 p->policy = policy;
4972 p->rt_priority = prio;
4973 p->normal_prio = normal_prio(p);
4974 /* we are holding p->pi_lock already */
4975 p->prio = rt_mutex_getprio(p);
4976 if (rt_prio(p->prio))
4977 p->sched_class = &rt_sched_class;
4978 else
4979 p->sched_class = &fair_sched_class;
4980 set_load_weight(p);
4984 * check the target process has a UID that matches the current process's
4986 static bool check_same_owner(struct task_struct *p)
4988 const struct cred *cred = current_cred(), *pcred;
4989 bool match;
4991 rcu_read_lock();
4992 pcred = __task_cred(p);
4993 if (cred->user->user_ns == pcred->user->user_ns)
4994 match = (cred->euid == pcred->euid ||
4995 cred->euid == pcred->uid);
4996 else
4997 match = false;
4998 rcu_read_unlock();
4999 return match;
5002 static int __sched_setscheduler(struct task_struct *p, int policy,
5003 const struct sched_param *param, bool user)
5005 int retval, oldprio, oldpolicy = -1, on_rq, running;
5006 unsigned long flags;
5007 const struct sched_class *prev_class;
5008 struct rq *rq;
5009 int reset_on_fork;
5011 /* may grab non-irq protected spin_locks */
5012 BUG_ON(in_interrupt());
5013 recheck:
5014 /* double check policy once rq lock held */
5015 if (policy < 0) {
5016 reset_on_fork = p->sched_reset_on_fork;
5017 policy = oldpolicy = p->policy;
5018 } else {
5019 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5020 policy &= ~SCHED_RESET_ON_FORK;
5022 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5023 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5024 policy != SCHED_IDLE)
5025 return -EINVAL;
5029 * Valid priorities for SCHED_FIFO and SCHED_RR are
5030 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5031 * SCHED_BATCH and SCHED_IDLE is 0.
5033 if (param->sched_priority < 0 ||
5034 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5035 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5036 return -EINVAL;
5037 if (rt_policy(policy) != (param->sched_priority != 0))
5038 return -EINVAL;
5041 * Allow unprivileged RT tasks to decrease priority:
5043 if (user && !capable(CAP_SYS_NICE)) {
5044 if (rt_policy(policy)) {
5045 unsigned long rlim_rtprio =
5046 task_rlimit(p, RLIMIT_RTPRIO);
5048 /* can't set/change the rt policy */
5049 if (policy != p->policy && !rlim_rtprio)
5050 return -EPERM;
5052 /* can't increase priority */
5053 if (param->sched_priority > p->rt_priority &&
5054 param->sched_priority > rlim_rtprio)
5055 return -EPERM;
5059 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5060 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5062 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5063 if (!can_nice(p, TASK_NICE(p)))
5064 return -EPERM;
5067 /* can't change other user's priorities */
5068 if (!check_same_owner(p))
5069 return -EPERM;
5071 /* Normal users shall not reset the sched_reset_on_fork flag */
5072 if (p->sched_reset_on_fork && !reset_on_fork)
5073 return -EPERM;
5076 if (user) {
5077 retval = security_task_setscheduler(p);
5078 if (retval)
5079 return retval;
5083 * make sure no PI-waiters arrive (or leave) while we are
5084 * changing the priority of the task:
5086 * To be able to change p->policy safely, the appropriate
5087 * runqueue lock must be held.
5089 rq = task_rq_lock(p, &flags);
5092 * Changing the policy of the stop threads its a very bad idea
5094 if (p == rq->stop) {
5095 task_rq_unlock(rq, p, &flags);
5096 return -EINVAL;
5100 * If not changing anything there's no need to proceed further:
5102 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5103 param->sched_priority == p->rt_priority))) {
5105 __task_rq_unlock(rq);
5106 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5107 return 0;
5110 #ifdef CONFIG_RT_GROUP_SCHED
5111 if (user) {
5113 * Do not allow realtime tasks into groups that have no runtime
5114 * assigned.
5116 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5117 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5118 !task_group_is_autogroup(task_group(p))) {
5119 task_rq_unlock(rq, p, &flags);
5120 return -EPERM;
5123 #endif
5125 /* recheck policy now with rq lock held */
5126 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5127 policy = oldpolicy = -1;
5128 task_rq_unlock(rq, p, &flags);
5129 goto recheck;
5131 on_rq = p->on_rq;
5132 running = task_current(rq, p);
5133 if (on_rq)
5134 deactivate_task(rq, p, 0);
5135 if (running)
5136 p->sched_class->put_prev_task(rq, p);
5138 p->sched_reset_on_fork = reset_on_fork;
5140 oldprio = p->prio;
5141 prev_class = p->sched_class;
5142 __setscheduler(rq, p, policy, param->sched_priority);
5144 if (running)
5145 p->sched_class->set_curr_task(rq);
5146 if (on_rq)
5147 activate_task(rq, p, 0);
5149 check_class_changed(rq, p, prev_class, oldprio);
5150 task_rq_unlock(rq, p, &flags);
5152 rt_mutex_adjust_pi(p);
5154 return 0;
5158 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5159 * @p: the task in question.
5160 * @policy: new policy.
5161 * @param: structure containing the new RT priority.
5163 * NOTE that the task may be already dead.
5165 int sched_setscheduler(struct task_struct *p, int policy,
5166 const struct sched_param *param)
5168 return __sched_setscheduler(p, policy, param, true);
5170 EXPORT_SYMBOL_GPL(sched_setscheduler);
5173 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5174 * @p: the task in question.
5175 * @policy: new policy.
5176 * @param: structure containing the new RT priority.
5178 * Just like sched_setscheduler, only don't bother checking if the
5179 * current context has permission. For example, this is needed in
5180 * stop_machine(): we create temporary high priority worker threads,
5181 * but our caller might not have that capability.
5183 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5184 const struct sched_param *param)
5186 return __sched_setscheduler(p, policy, param, false);
5189 static int
5190 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5192 struct sched_param lparam;
5193 struct task_struct *p;
5194 int retval;
5196 if (!param || pid < 0)
5197 return -EINVAL;
5198 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5199 return -EFAULT;
5201 rcu_read_lock();
5202 retval = -ESRCH;
5203 p = find_process_by_pid(pid);
5204 if (p != NULL)
5205 retval = sched_setscheduler(p, policy, &lparam);
5206 rcu_read_unlock();
5208 return retval;
5212 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5213 * @pid: the pid in question.
5214 * @policy: new policy.
5215 * @param: structure containing the new RT priority.
5217 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5218 struct sched_param __user *, param)
5220 /* negative values for policy are not valid */
5221 if (policy < 0)
5222 return -EINVAL;
5224 return do_sched_setscheduler(pid, policy, param);
5228 * sys_sched_setparam - set/change the RT priority of a thread
5229 * @pid: the pid in question.
5230 * @param: structure containing the new RT priority.
5232 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5234 return do_sched_setscheduler(pid, -1, param);
5238 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5239 * @pid: the pid in question.
5241 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5243 struct task_struct *p;
5244 int retval;
5246 if (pid < 0)
5247 return -EINVAL;
5249 retval = -ESRCH;
5250 rcu_read_lock();
5251 p = find_process_by_pid(pid);
5252 if (p) {
5253 retval = security_task_getscheduler(p);
5254 if (!retval)
5255 retval = p->policy
5256 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5258 rcu_read_unlock();
5259 return retval;
5263 * sys_sched_getparam - get the RT priority of a thread
5264 * @pid: the pid in question.
5265 * @param: structure containing the RT priority.
5267 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5269 struct sched_param lp;
5270 struct task_struct *p;
5271 int retval;
5273 if (!param || pid < 0)
5274 return -EINVAL;
5276 rcu_read_lock();
5277 p = find_process_by_pid(pid);
5278 retval = -ESRCH;
5279 if (!p)
5280 goto out_unlock;
5282 retval = security_task_getscheduler(p);
5283 if (retval)
5284 goto out_unlock;
5286 lp.sched_priority = p->rt_priority;
5287 rcu_read_unlock();
5290 * This one might sleep, we cannot do it with a spinlock held ...
5292 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5294 return retval;
5296 out_unlock:
5297 rcu_read_unlock();
5298 return retval;
5301 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5303 cpumask_var_t cpus_allowed, new_mask;
5304 struct task_struct *p;
5305 int retval;
5307 get_online_cpus();
5308 rcu_read_lock();
5310 p = find_process_by_pid(pid);
5311 if (!p) {
5312 rcu_read_unlock();
5313 put_online_cpus();
5314 return -ESRCH;
5317 /* Prevent p going away */
5318 get_task_struct(p);
5319 rcu_read_unlock();
5321 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5322 retval = -ENOMEM;
5323 goto out_put_task;
5325 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5326 retval = -ENOMEM;
5327 goto out_free_cpus_allowed;
5329 retval = -EPERM;
5330 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5331 goto out_unlock;
5333 retval = security_task_setscheduler(p);
5334 if (retval)
5335 goto out_unlock;
5337 cpuset_cpus_allowed(p, cpus_allowed);
5338 cpumask_and(new_mask, in_mask, cpus_allowed);
5339 again:
5340 retval = set_cpus_allowed_ptr(p, new_mask);
5342 if (!retval) {
5343 cpuset_cpus_allowed(p, cpus_allowed);
5344 if (!cpumask_subset(new_mask, cpus_allowed)) {
5346 * We must have raced with a concurrent cpuset
5347 * update. Just reset the cpus_allowed to the
5348 * cpuset's cpus_allowed
5350 cpumask_copy(new_mask, cpus_allowed);
5351 goto again;
5354 out_unlock:
5355 free_cpumask_var(new_mask);
5356 out_free_cpus_allowed:
5357 free_cpumask_var(cpus_allowed);
5358 out_put_task:
5359 put_task_struct(p);
5360 put_online_cpus();
5361 return retval;
5364 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5365 struct cpumask *new_mask)
5367 if (len < cpumask_size())
5368 cpumask_clear(new_mask);
5369 else if (len > cpumask_size())
5370 len = cpumask_size();
5372 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5376 * sys_sched_setaffinity - set the cpu affinity of a process
5377 * @pid: pid of the process
5378 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5379 * @user_mask_ptr: user-space pointer to the new cpu mask
5381 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5382 unsigned long __user *, user_mask_ptr)
5384 cpumask_var_t new_mask;
5385 int retval;
5387 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5388 return -ENOMEM;
5390 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5391 if (retval == 0)
5392 retval = sched_setaffinity(pid, new_mask);
5393 free_cpumask_var(new_mask);
5394 return retval;
5397 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5399 struct task_struct *p;
5400 unsigned long flags;
5401 int retval;
5403 get_online_cpus();
5404 rcu_read_lock();
5406 retval = -ESRCH;
5407 p = find_process_by_pid(pid);
5408 if (!p)
5409 goto out_unlock;
5411 retval = security_task_getscheduler(p);
5412 if (retval)
5413 goto out_unlock;
5415 raw_spin_lock_irqsave(&p->pi_lock, flags);
5416 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5417 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5419 out_unlock:
5420 rcu_read_unlock();
5421 put_online_cpus();
5423 return retval;
5427 * sys_sched_getaffinity - get the cpu affinity of a process
5428 * @pid: pid of the process
5429 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5430 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5432 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5433 unsigned long __user *, user_mask_ptr)
5435 int ret;
5436 cpumask_var_t mask;
5438 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5439 return -EINVAL;
5440 if (len & (sizeof(unsigned long)-1))
5441 return -EINVAL;
5443 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5444 return -ENOMEM;
5446 ret = sched_getaffinity(pid, mask);
5447 if (ret == 0) {
5448 size_t retlen = min_t(size_t, len, cpumask_size());
5450 if (copy_to_user(user_mask_ptr, mask, retlen))
5451 ret = -EFAULT;
5452 else
5453 ret = retlen;
5455 free_cpumask_var(mask);
5457 return ret;
5461 * sys_sched_yield - yield the current processor to other threads.
5463 * This function yields the current CPU to other tasks. If there are no
5464 * other threads running on this CPU then this function will return.
5466 SYSCALL_DEFINE0(sched_yield)
5468 struct rq *rq = this_rq_lock();
5470 schedstat_inc(rq, yld_count);
5471 current->sched_class->yield_task(rq);
5474 * Since we are going to call schedule() anyway, there's
5475 * no need to preempt or enable interrupts:
5477 __release(rq->lock);
5478 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5479 do_raw_spin_unlock(&rq->lock);
5480 preempt_enable_no_resched();
5482 schedule();
5484 return 0;
5487 static inline int should_resched(void)
5489 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5492 static void __cond_resched(void)
5494 add_preempt_count(PREEMPT_ACTIVE);
5495 schedule();
5496 sub_preempt_count(PREEMPT_ACTIVE);
5499 int __sched _cond_resched(void)
5501 if (should_resched()) {
5502 __cond_resched();
5503 return 1;
5505 return 0;
5507 EXPORT_SYMBOL(_cond_resched);
5510 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5511 * call schedule, and on return reacquire the lock.
5513 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5514 * operations here to prevent schedule() from being called twice (once via
5515 * spin_unlock(), once by hand).
5517 int __cond_resched_lock(spinlock_t *lock)
5519 int resched = should_resched();
5520 int ret = 0;
5522 lockdep_assert_held(lock);
5524 if (spin_needbreak(lock) || resched) {
5525 spin_unlock(lock);
5526 if (resched)
5527 __cond_resched();
5528 else
5529 cpu_relax();
5530 ret = 1;
5531 spin_lock(lock);
5533 return ret;
5535 EXPORT_SYMBOL(__cond_resched_lock);
5537 int __sched __cond_resched_softirq(void)
5539 BUG_ON(!in_softirq());
5541 if (should_resched()) {
5542 local_bh_enable();
5543 __cond_resched();
5544 local_bh_disable();
5545 return 1;
5547 return 0;
5549 EXPORT_SYMBOL(__cond_resched_softirq);
5552 * yield - yield the current processor to other threads.
5554 * This is a shortcut for kernel-space yielding - it marks the
5555 * thread runnable and calls sys_sched_yield().
5557 void __sched yield(void)
5559 set_current_state(TASK_RUNNING);
5560 sys_sched_yield();
5562 EXPORT_SYMBOL(yield);
5565 * yield_to - yield the current processor to another thread in
5566 * your thread group, or accelerate that thread toward the
5567 * processor it's on.
5568 * @p: target task
5569 * @preempt: whether task preemption is allowed or not
5571 * It's the caller's job to ensure that the target task struct
5572 * can't go away on us before we can do any checks.
5574 * Returns true if we indeed boosted the target task.
5576 bool __sched yield_to(struct task_struct *p, bool preempt)
5578 struct task_struct *curr = current;
5579 struct rq *rq, *p_rq;
5580 unsigned long flags;
5581 bool yielded = 0;
5583 local_irq_save(flags);
5584 rq = this_rq();
5586 again:
5587 p_rq = task_rq(p);
5588 double_rq_lock(rq, p_rq);
5589 while (task_rq(p) != p_rq) {
5590 double_rq_unlock(rq, p_rq);
5591 goto again;
5594 if (!curr->sched_class->yield_to_task)
5595 goto out;
5597 if (curr->sched_class != p->sched_class)
5598 goto out;
5600 if (task_running(p_rq, p) || p->state)
5601 goto out;
5603 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5604 if (yielded) {
5605 schedstat_inc(rq, yld_count);
5607 * Make p's CPU reschedule; pick_next_entity takes care of
5608 * fairness.
5610 if (preempt && rq != p_rq)
5611 resched_task(p_rq->curr);
5614 out:
5615 double_rq_unlock(rq, p_rq);
5616 local_irq_restore(flags);
5618 if (yielded)
5619 schedule();
5621 return yielded;
5623 EXPORT_SYMBOL_GPL(yield_to);
5626 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5627 * that process accounting knows that this is a task in IO wait state.
5629 void __sched io_schedule(void)
5631 struct rq *rq = raw_rq();
5633 delayacct_blkio_start();
5634 atomic_inc(&rq->nr_iowait);
5635 blk_flush_plug(current);
5636 current->in_iowait = 1;
5637 schedule();
5638 current->in_iowait = 0;
5639 atomic_dec(&rq->nr_iowait);
5640 delayacct_blkio_end();
5642 EXPORT_SYMBOL(io_schedule);
5644 long __sched io_schedule_timeout(long timeout)
5646 struct rq *rq = raw_rq();
5647 long ret;
5649 delayacct_blkio_start();
5650 atomic_inc(&rq->nr_iowait);
5651 blk_flush_plug(current);
5652 current->in_iowait = 1;
5653 ret = schedule_timeout(timeout);
5654 current->in_iowait = 0;
5655 atomic_dec(&rq->nr_iowait);
5656 delayacct_blkio_end();
5657 return ret;
5661 * sys_sched_get_priority_max - return maximum RT priority.
5662 * @policy: scheduling class.
5664 * this syscall returns the maximum rt_priority that can be used
5665 * by a given scheduling class.
5667 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5669 int ret = -EINVAL;
5671 switch (policy) {
5672 case SCHED_FIFO:
5673 case SCHED_RR:
5674 ret = MAX_USER_RT_PRIO-1;
5675 break;
5676 case SCHED_NORMAL:
5677 case SCHED_BATCH:
5678 case SCHED_IDLE:
5679 ret = 0;
5680 break;
5682 return ret;
5686 * sys_sched_get_priority_min - return minimum RT priority.
5687 * @policy: scheduling class.
5689 * this syscall returns the minimum rt_priority that can be used
5690 * by a given scheduling class.
5692 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5694 int ret = -EINVAL;
5696 switch (policy) {
5697 case SCHED_FIFO:
5698 case SCHED_RR:
5699 ret = 1;
5700 break;
5701 case SCHED_NORMAL:
5702 case SCHED_BATCH:
5703 case SCHED_IDLE:
5704 ret = 0;
5706 return ret;
5710 * sys_sched_rr_get_interval - return the default timeslice of a process.
5711 * @pid: pid of the process.
5712 * @interval: userspace pointer to the timeslice value.
5714 * this syscall writes the default timeslice value of a given process
5715 * into the user-space timespec buffer. A value of '0' means infinity.
5717 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5718 struct timespec __user *, interval)
5720 struct task_struct *p;
5721 unsigned int time_slice;
5722 unsigned long flags;
5723 struct rq *rq;
5724 int retval;
5725 struct timespec t;
5727 if (pid < 0)
5728 return -EINVAL;
5730 retval = -ESRCH;
5731 rcu_read_lock();
5732 p = find_process_by_pid(pid);
5733 if (!p)
5734 goto out_unlock;
5736 retval = security_task_getscheduler(p);
5737 if (retval)
5738 goto out_unlock;
5740 rq = task_rq_lock(p, &flags);
5741 time_slice = p->sched_class->get_rr_interval(rq, p);
5742 task_rq_unlock(rq, p, &flags);
5744 rcu_read_unlock();
5745 jiffies_to_timespec(time_slice, &t);
5746 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5747 return retval;
5749 out_unlock:
5750 rcu_read_unlock();
5751 return retval;
5754 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5756 void sched_show_task(struct task_struct *p)
5758 unsigned long free = 0;
5759 unsigned state;
5761 state = p->state ? __ffs(p->state) + 1 : 0;
5762 printk(KERN_INFO "%-15.15s %c", p->comm,
5763 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5764 #if BITS_PER_LONG == 32
5765 if (state == TASK_RUNNING)
5766 printk(KERN_CONT " running ");
5767 else
5768 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5769 #else
5770 if (state == TASK_RUNNING)
5771 printk(KERN_CONT " running task ");
5772 else
5773 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5774 #endif
5775 #ifdef CONFIG_DEBUG_STACK_USAGE
5776 free = stack_not_used(p);
5777 #endif
5778 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5779 task_pid_nr(p), task_pid_nr(p->real_parent),
5780 (unsigned long)task_thread_info(p)->flags);
5782 show_stack(p, NULL);
5785 void show_state_filter(unsigned long state_filter)
5787 struct task_struct *g, *p;
5789 #if BITS_PER_LONG == 32
5790 printk(KERN_INFO
5791 " task PC stack pid father\n");
5792 #else
5793 printk(KERN_INFO
5794 " task PC stack pid father\n");
5795 #endif
5796 read_lock(&tasklist_lock);
5797 do_each_thread(g, p) {
5799 * reset the NMI-timeout, listing all files on a slow
5800 * console might take a lot of time:
5802 touch_nmi_watchdog();
5803 if (!state_filter || (p->state & state_filter))
5804 sched_show_task(p);
5805 } while_each_thread(g, p);
5807 touch_all_softlockup_watchdogs();
5809 #ifdef CONFIG_SCHED_DEBUG
5810 sysrq_sched_debug_show();
5811 #endif
5812 read_unlock(&tasklist_lock);
5814 * Only show locks if all tasks are dumped:
5816 if (!state_filter)
5817 debug_show_all_locks();
5820 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5822 idle->sched_class = &idle_sched_class;
5826 * init_idle - set up an idle thread for a given CPU
5827 * @idle: task in question
5828 * @cpu: cpu the idle task belongs to
5830 * NOTE: this function does not set the idle thread's NEED_RESCHED
5831 * flag, to make booting more robust.
5833 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5835 struct rq *rq = cpu_rq(cpu);
5836 unsigned long flags;
5838 raw_spin_lock_irqsave(&rq->lock, flags);
5840 __sched_fork(idle);
5841 idle->state = TASK_RUNNING;
5842 idle->se.exec_start = sched_clock();
5844 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5846 * We're having a chicken and egg problem, even though we are
5847 * holding rq->lock, the cpu isn't yet set to this cpu so the
5848 * lockdep check in task_group() will fail.
5850 * Similar case to sched_fork(). / Alternatively we could
5851 * use task_rq_lock() here and obtain the other rq->lock.
5853 * Silence PROVE_RCU
5855 rcu_read_lock();
5856 __set_task_cpu(idle, cpu);
5857 rcu_read_unlock();
5859 rq->curr = rq->idle = idle;
5860 #if defined(CONFIG_SMP)
5861 idle->on_cpu = 1;
5862 #endif
5863 raw_spin_unlock_irqrestore(&rq->lock, flags);
5865 /* Set the preempt count _outside_ the spinlocks! */
5866 task_thread_info(idle)->preempt_count = 0;
5869 * The idle tasks have their own, simple scheduling class:
5871 idle->sched_class = &idle_sched_class;
5872 ftrace_graph_init_idle_task(idle, cpu);
5876 * In a system that switches off the HZ timer nohz_cpu_mask
5877 * indicates which cpus entered this state. This is used
5878 * in the rcu update to wait only for active cpus. For system
5879 * which do not switch off the HZ timer nohz_cpu_mask should
5880 * always be CPU_BITS_NONE.
5882 cpumask_var_t nohz_cpu_mask;
5885 * Increase the granularity value when there are more CPUs,
5886 * because with more CPUs the 'effective latency' as visible
5887 * to users decreases. But the relationship is not linear,
5888 * so pick a second-best guess by going with the log2 of the
5889 * number of CPUs.
5891 * This idea comes from the SD scheduler of Con Kolivas:
5893 static int get_update_sysctl_factor(void)
5895 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5896 unsigned int factor;
5898 switch (sysctl_sched_tunable_scaling) {
5899 case SCHED_TUNABLESCALING_NONE:
5900 factor = 1;
5901 break;
5902 case SCHED_TUNABLESCALING_LINEAR:
5903 factor = cpus;
5904 break;
5905 case SCHED_TUNABLESCALING_LOG:
5906 default:
5907 factor = 1 + ilog2(cpus);
5908 break;
5911 return factor;
5914 static void update_sysctl(void)
5916 unsigned int factor = get_update_sysctl_factor();
5918 #define SET_SYSCTL(name) \
5919 (sysctl_##name = (factor) * normalized_sysctl_##name)
5920 SET_SYSCTL(sched_min_granularity);
5921 SET_SYSCTL(sched_latency);
5922 SET_SYSCTL(sched_wakeup_granularity);
5923 #undef SET_SYSCTL
5926 static inline void sched_init_granularity(void)
5928 update_sysctl();
5931 #ifdef CONFIG_SMP
5933 * This is how migration works:
5935 * 1) we invoke migration_cpu_stop() on the target CPU using
5936 * stop_one_cpu().
5937 * 2) stopper starts to run (implicitly forcing the migrated thread
5938 * off the CPU)
5939 * 3) it checks whether the migrated task is still in the wrong runqueue.
5940 * 4) if it's in the wrong runqueue then the migration thread removes
5941 * it and puts it into the right queue.
5942 * 5) stopper completes and stop_one_cpu() returns and the migration
5943 * is done.
5947 * Change a given task's CPU affinity. Migrate the thread to a
5948 * proper CPU and schedule it away if the CPU it's executing on
5949 * is removed from the allowed bitmask.
5951 * NOTE: the caller must have a valid reference to the task, the
5952 * task must not exit() & deallocate itself prematurely. The
5953 * call is not atomic; no spinlocks may be held.
5955 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5957 unsigned long flags;
5958 struct rq *rq;
5959 unsigned int dest_cpu;
5960 int ret = 0;
5962 rq = task_rq_lock(p, &flags);
5964 if (cpumask_equal(&p->cpus_allowed, new_mask))
5965 goto out;
5967 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5968 ret = -EINVAL;
5969 goto out;
5972 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
5973 ret = -EINVAL;
5974 goto out;
5977 if (p->sched_class->set_cpus_allowed)
5978 p->sched_class->set_cpus_allowed(p, new_mask);
5979 else {
5980 cpumask_copy(&p->cpus_allowed, new_mask);
5981 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5984 /* Can the task run on the task's current CPU? If so, we're done */
5985 if (cpumask_test_cpu(task_cpu(p), new_mask))
5986 goto out;
5988 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5989 if (p->on_rq) {
5990 struct migration_arg arg = { p, dest_cpu };
5991 /* Need help from migration thread: drop lock and wait. */
5992 task_rq_unlock(rq, p, &flags);
5993 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5994 tlb_migrate_finish(p->mm);
5995 return 0;
5997 out:
5998 task_rq_unlock(rq, p, &flags);
6000 return ret;
6002 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6005 * Move (not current) task off this cpu, onto dest cpu. We're doing
6006 * this because either it can't run here any more (set_cpus_allowed()
6007 * away from this CPU, or CPU going down), or because we're
6008 * attempting to rebalance this task on exec (sched_exec).
6010 * So we race with normal scheduler movements, but that's OK, as long
6011 * as the task is no longer on this CPU.
6013 * Returns non-zero if task was successfully migrated.
6015 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6017 struct rq *rq_dest, *rq_src;
6018 int ret = 0;
6020 if (unlikely(!cpu_active(dest_cpu)))
6021 return ret;
6023 rq_src = cpu_rq(src_cpu);
6024 rq_dest = cpu_rq(dest_cpu);
6026 raw_spin_lock(&p->pi_lock);
6027 double_rq_lock(rq_src, rq_dest);
6028 /* Already moved. */
6029 if (task_cpu(p) != src_cpu)
6030 goto done;
6031 /* Affinity changed (again). */
6032 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6033 goto fail;
6036 * If we're not on a rq, the next wake-up will ensure we're
6037 * placed properly.
6039 if (p->on_rq) {
6040 deactivate_task(rq_src, p, 0);
6041 set_task_cpu(p, dest_cpu);
6042 activate_task(rq_dest, p, 0);
6043 check_preempt_curr(rq_dest, p, 0);
6045 done:
6046 ret = 1;
6047 fail:
6048 double_rq_unlock(rq_src, rq_dest);
6049 raw_spin_unlock(&p->pi_lock);
6050 return ret;
6054 * migration_cpu_stop - this will be executed by a highprio stopper thread
6055 * and performs thread migration by bumping thread off CPU then
6056 * 'pushing' onto another runqueue.
6058 static int migration_cpu_stop(void *data)
6060 struct migration_arg *arg = data;
6063 * The original target cpu might have gone down and we might
6064 * be on another cpu but it doesn't matter.
6066 local_irq_disable();
6067 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6068 local_irq_enable();
6069 return 0;
6072 #ifdef CONFIG_HOTPLUG_CPU
6075 * Ensures that the idle task is using init_mm right before its cpu goes
6076 * offline.
6078 void idle_task_exit(void)
6080 struct mm_struct *mm = current->active_mm;
6082 BUG_ON(cpu_online(smp_processor_id()));
6084 if (mm != &init_mm)
6085 switch_mm(mm, &init_mm, current);
6086 mmdrop(mm);
6090 * While a dead CPU has no uninterruptible tasks queued at this point,
6091 * it might still have a nonzero ->nr_uninterruptible counter, because
6092 * for performance reasons the counter is not stricly tracking tasks to
6093 * their home CPUs. So we just add the counter to another CPU's counter,
6094 * to keep the global sum constant after CPU-down:
6096 static void migrate_nr_uninterruptible(struct rq *rq_src)
6098 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6100 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6101 rq_src->nr_uninterruptible = 0;
6105 * remove the tasks which were accounted by rq from calc_load_tasks.
6107 static void calc_global_load_remove(struct rq *rq)
6109 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6110 rq->calc_load_active = 0;
6114 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6115 * try_to_wake_up()->select_task_rq().
6117 * Called with rq->lock held even though we'er in stop_machine() and
6118 * there's no concurrency possible, we hold the required locks anyway
6119 * because of lock validation efforts.
6121 static void migrate_tasks(unsigned int dead_cpu)
6123 struct rq *rq = cpu_rq(dead_cpu);
6124 struct task_struct *next, *stop = rq->stop;
6125 int dest_cpu;
6128 * Fudge the rq selection such that the below task selection loop
6129 * doesn't get stuck on the currently eligible stop task.
6131 * We're currently inside stop_machine() and the rq is either stuck
6132 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6133 * either way we should never end up calling schedule() until we're
6134 * done here.
6136 rq->stop = NULL;
6138 for ( ; ; ) {
6140 * There's this thread running, bail when that's the only
6141 * remaining thread.
6143 if (rq->nr_running == 1)
6144 break;
6146 next = pick_next_task(rq);
6147 BUG_ON(!next);
6148 next->sched_class->put_prev_task(rq, next);
6150 /* Find suitable destination for @next, with force if needed. */
6151 dest_cpu = select_fallback_rq(dead_cpu, next);
6152 raw_spin_unlock(&rq->lock);
6154 __migrate_task(next, dead_cpu, dest_cpu);
6156 raw_spin_lock(&rq->lock);
6159 rq->stop = stop;
6162 #endif /* CONFIG_HOTPLUG_CPU */
6164 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6166 static struct ctl_table sd_ctl_dir[] = {
6168 .procname = "sched_domain",
6169 .mode = 0555,
6174 static struct ctl_table sd_ctl_root[] = {
6176 .procname = "kernel",
6177 .mode = 0555,
6178 .child = sd_ctl_dir,
6183 static struct ctl_table *sd_alloc_ctl_entry(int n)
6185 struct ctl_table *entry =
6186 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6188 return entry;
6191 static void sd_free_ctl_entry(struct ctl_table **tablep)
6193 struct ctl_table *entry;
6196 * In the intermediate directories, both the child directory and
6197 * procname are dynamically allocated and could fail but the mode
6198 * will always be set. In the lowest directory the names are
6199 * static strings and all have proc handlers.
6201 for (entry = *tablep; entry->mode; entry++) {
6202 if (entry->child)
6203 sd_free_ctl_entry(&entry->child);
6204 if (entry->proc_handler == NULL)
6205 kfree(entry->procname);
6208 kfree(*tablep);
6209 *tablep = NULL;
6212 static void
6213 set_table_entry(struct ctl_table *entry,
6214 const char *procname, void *data, int maxlen,
6215 mode_t mode, proc_handler *proc_handler)
6217 entry->procname = procname;
6218 entry->data = data;
6219 entry->maxlen = maxlen;
6220 entry->mode = mode;
6221 entry->proc_handler = proc_handler;
6224 static struct ctl_table *
6225 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6227 struct ctl_table *table = sd_alloc_ctl_entry(13);
6229 if (table == NULL)
6230 return NULL;
6232 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6233 sizeof(long), 0644, proc_doulongvec_minmax);
6234 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6235 sizeof(long), 0644, proc_doulongvec_minmax);
6236 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6237 sizeof(int), 0644, proc_dointvec_minmax);
6238 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6239 sizeof(int), 0644, proc_dointvec_minmax);
6240 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6241 sizeof(int), 0644, proc_dointvec_minmax);
6242 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6243 sizeof(int), 0644, proc_dointvec_minmax);
6244 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6245 sizeof(int), 0644, proc_dointvec_minmax);
6246 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6247 sizeof(int), 0644, proc_dointvec_minmax);
6248 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6249 sizeof(int), 0644, proc_dointvec_minmax);
6250 set_table_entry(&table[9], "cache_nice_tries",
6251 &sd->cache_nice_tries,
6252 sizeof(int), 0644, proc_dointvec_minmax);
6253 set_table_entry(&table[10], "flags", &sd->flags,
6254 sizeof(int), 0644, proc_dointvec_minmax);
6255 set_table_entry(&table[11], "name", sd->name,
6256 CORENAME_MAX_SIZE, 0444, proc_dostring);
6257 /* &table[12] is terminator */
6259 return table;
6262 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6264 struct ctl_table *entry, *table;
6265 struct sched_domain *sd;
6266 int domain_num = 0, i;
6267 char buf[32];
6269 for_each_domain(cpu, sd)
6270 domain_num++;
6271 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6272 if (table == NULL)
6273 return NULL;
6275 i = 0;
6276 for_each_domain(cpu, sd) {
6277 snprintf(buf, 32, "domain%d", i);
6278 entry->procname = kstrdup(buf, GFP_KERNEL);
6279 entry->mode = 0555;
6280 entry->child = sd_alloc_ctl_domain_table(sd);
6281 entry++;
6282 i++;
6284 return table;
6287 static struct ctl_table_header *sd_sysctl_header;
6288 static void register_sched_domain_sysctl(void)
6290 int i, cpu_num = num_possible_cpus();
6291 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6292 char buf[32];
6294 WARN_ON(sd_ctl_dir[0].child);
6295 sd_ctl_dir[0].child = entry;
6297 if (entry == NULL)
6298 return;
6300 for_each_possible_cpu(i) {
6301 snprintf(buf, 32, "cpu%d", i);
6302 entry->procname = kstrdup(buf, GFP_KERNEL);
6303 entry->mode = 0555;
6304 entry->child = sd_alloc_ctl_cpu_table(i);
6305 entry++;
6308 WARN_ON(sd_sysctl_header);
6309 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6312 /* may be called multiple times per register */
6313 static void unregister_sched_domain_sysctl(void)
6315 if (sd_sysctl_header)
6316 unregister_sysctl_table(sd_sysctl_header);
6317 sd_sysctl_header = NULL;
6318 if (sd_ctl_dir[0].child)
6319 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6321 #else
6322 static void register_sched_domain_sysctl(void)
6325 static void unregister_sched_domain_sysctl(void)
6328 #endif
6330 static void set_rq_online(struct rq *rq)
6332 if (!rq->online) {
6333 const struct sched_class *class;
6335 cpumask_set_cpu(rq->cpu, rq->rd->online);
6336 rq->online = 1;
6338 for_each_class(class) {
6339 if (class->rq_online)
6340 class->rq_online(rq);
6345 static void set_rq_offline(struct rq *rq)
6347 if (rq->online) {
6348 const struct sched_class *class;
6350 for_each_class(class) {
6351 if (class->rq_offline)
6352 class->rq_offline(rq);
6355 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6356 rq->online = 0;
6361 * migration_call - callback that gets triggered when a CPU is added.
6362 * Here we can start up the necessary migration thread for the new CPU.
6364 static int __cpuinit
6365 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6367 int cpu = (long)hcpu;
6368 unsigned long flags;
6369 struct rq *rq = cpu_rq(cpu);
6371 switch (action & ~CPU_TASKS_FROZEN) {
6373 case CPU_UP_PREPARE:
6374 rq->calc_load_update = calc_load_update;
6375 break;
6377 case CPU_ONLINE:
6378 /* Update our root-domain */
6379 raw_spin_lock_irqsave(&rq->lock, flags);
6380 if (rq->rd) {
6381 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6383 set_rq_online(rq);
6385 raw_spin_unlock_irqrestore(&rq->lock, flags);
6386 break;
6388 #ifdef CONFIG_HOTPLUG_CPU
6389 case CPU_DYING:
6390 sched_ttwu_pending();
6391 /* Update our root-domain */
6392 raw_spin_lock_irqsave(&rq->lock, flags);
6393 if (rq->rd) {
6394 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6395 set_rq_offline(rq);
6397 migrate_tasks(cpu);
6398 BUG_ON(rq->nr_running != 1); /* the migration thread */
6399 raw_spin_unlock_irqrestore(&rq->lock, flags);
6401 migrate_nr_uninterruptible(rq);
6402 calc_global_load_remove(rq);
6403 break;
6404 #endif
6407 update_max_interval();
6409 return NOTIFY_OK;
6413 * Register at high priority so that task migration (migrate_all_tasks)
6414 * happens before everything else. This has to be lower priority than
6415 * the notifier in the perf_event subsystem, though.
6417 static struct notifier_block __cpuinitdata migration_notifier = {
6418 .notifier_call = migration_call,
6419 .priority = CPU_PRI_MIGRATION,
6422 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6423 unsigned long action, void *hcpu)
6425 switch (action & ~CPU_TASKS_FROZEN) {
6426 case CPU_ONLINE:
6427 case CPU_DOWN_FAILED:
6428 set_cpu_active((long)hcpu, true);
6429 return NOTIFY_OK;
6430 default:
6431 return NOTIFY_DONE;
6435 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6436 unsigned long action, void *hcpu)
6438 switch (action & ~CPU_TASKS_FROZEN) {
6439 case CPU_DOWN_PREPARE:
6440 set_cpu_active((long)hcpu, false);
6441 return NOTIFY_OK;
6442 default:
6443 return NOTIFY_DONE;
6447 static int __init migration_init(void)
6449 void *cpu = (void *)(long)smp_processor_id();
6450 int err;
6452 /* Initialize migration for the boot CPU */
6453 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6454 BUG_ON(err == NOTIFY_BAD);
6455 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6456 register_cpu_notifier(&migration_notifier);
6458 /* Register cpu active notifiers */
6459 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6460 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6462 return 0;
6464 early_initcall(migration_init);
6465 #endif
6467 #ifdef CONFIG_SMP
6469 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6471 #ifdef CONFIG_SCHED_DEBUG
6473 static __read_mostly int sched_domain_debug_enabled;
6475 static int __init sched_domain_debug_setup(char *str)
6477 sched_domain_debug_enabled = 1;
6479 return 0;
6481 early_param("sched_debug", sched_domain_debug_setup);
6483 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6484 struct cpumask *groupmask)
6486 struct sched_group *group = sd->groups;
6487 char str[256];
6489 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6490 cpumask_clear(groupmask);
6492 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6494 if (!(sd->flags & SD_LOAD_BALANCE)) {
6495 printk("does not load-balance\n");
6496 if (sd->parent)
6497 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6498 " has parent");
6499 return -1;
6502 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6504 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6505 printk(KERN_ERR "ERROR: domain->span does not contain "
6506 "CPU%d\n", cpu);
6508 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6509 printk(KERN_ERR "ERROR: domain->groups does not contain"
6510 " CPU%d\n", cpu);
6513 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6514 do {
6515 if (!group) {
6516 printk("\n");
6517 printk(KERN_ERR "ERROR: group is NULL\n");
6518 break;
6521 if (!group->cpu_power) {
6522 printk(KERN_CONT "\n");
6523 printk(KERN_ERR "ERROR: domain->cpu_power not "
6524 "set\n");
6525 break;
6528 if (!cpumask_weight(sched_group_cpus(group))) {
6529 printk(KERN_CONT "\n");
6530 printk(KERN_ERR "ERROR: empty group\n");
6531 break;
6534 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6535 printk(KERN_CONT "\n");
6536 printk(KERN_ERR "ERROR: repeated CPUs\n");
6537 break;
6540 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6542 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6544 printk(KERN_CONT " %s", str);
6545 if (group->cpu_power != SCHED_POWER_SCALE) {
6546 printk(KERN_CONT " (cpu_power = %d)",
6547 group->cpu_power);
6550 group = group->next;
6551 } while (group != sd->groups);
6552 printk(KERN_CONT "\n");
6554 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6555 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6557 if (sd->parent &&
6558 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6559 printk(KERN_ERR "ERROR: parent span is not a superset "
6560 "of domain->span\n");
6561 return 0;
6564 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6566 int level = 0;
6568 if (!sched_domain_debug_enabled)
6569 return;
6571 if (!sd) {
6572 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6573 return;
6576 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6578 for (;;) {
6579 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6580 break;
6581 level++;
6582 sd = sd->parent;
6583 if (!sd)
6584 break;
6587 #else /* !CONFIG_SCHED_DEBUG */
6588 # define sched_domain_debug(sd, cpu) do { } while (0)
6589 #endif /* CONFIG_SCHED_DEBUG */
6591 static int sd_degenerate(struct sched_domain *sd)
6593 if (cpumask_weight(sched_domain_span(sd)) == 1)
6594 return 1;
6596 /* Following flags need at least 2 groups */
6597 if (sd->flags & (SD_LOAD_BALANCE |
6598 SD_BALANCE_NEWIDLE |
6599 SD_BALANCE_FORK |
6600 SD_BALANCE_EXEC |
6601 SD_SHARE_CPUPOWER |
6602 SD_SHARE_PKG_RESOURCES)) {
6603 if (sd->groups != sd->groups->next)
6604 return 0;
6607 /* Following flags don't use groups */
6608 if (sd->flags & (SD_WAKE_AFFINE))
6609 return 0;
6611 return 1;
6614 static int
6615 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6617 unsigned long cflags = sd->flags, pflags = parent->flags;
6619 if (sd_degenerate(parent))
6620 return 1;
6622 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6623 return 0;
6625 /* Flags needing groups don't count if only 1 group in parent */
6626 if (parent->groups == parent->groups->next) {
6627 pflags &= ~(SD_LOAD_BALANCE |
6628 SD_BALANCE_NEWIDLE |
6629 SD_BALANCE_FORK |
6630 SD_BALANCE_EXEC |
6631 SD_SHARE_CPUPOWER |
6632 SD_SHARE_PKG_RESOURCES);
6633 if (nr_node_ids == 1)
6634 pflags &= ~SD_SERIALIZE;
6636 if (~cflags & pflags)
6637 return 0;
6639 return 1;
6642 static void free_rootdomain(struct rcu_head *rcu)
6644 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6646 cpupri_cleanup(&rd->cpupri);
6647 free_cpumask_var(rd->rto_mask);
6648 free_cpumask_var(rd->online);
6649 free_cpumask_var(rd->span);
6650 kfree(rd);
6653 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6655 struct root_domain *old_rd = NULL;
6656 unsigned long flags;
6658 raw_spin_lock_irqsave(&rq->lock, flags);
6660 if (rq->rd) {
6661 old_rd = rq->rd;
6663 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6664 set_rq_offline(rq);
6666 cpumask_clear_cpu(rq->cpu, old_rd->span);
6669 * If we dont want to free the old_rt yet then
6670 * set old_rd to NULL to skip the freeing later
6671 * in this function:
6673 if (!atomic_dec_and_test(&old_rd->refcount))
6674 old_rd = NULL;
6677 atomic_inc(&rd->refcount);
6678 rq->rd = rd;
6680 cpumask_set_cpu(rq->cpu, rd->span);
6681 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6682 set_rq_online(rq);
6684 raw_spin_unlock_irqrestore(&rq->lock, flags);
6686 if (old_rd)
6687 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6690 static int init_rootdomain(struct root_domain *rd)
6692 memset(rd, 0, sizeof(*rd));
6694 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6695 goto out;
6696 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6697 goto free_span;
6698 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6699 goto free_online;
6701 if (cpupri_init(&rd->cpupri) != 0)
6702 goto free_rto_mask;
6703 return 0;
6705 free_rto_mask:
6706 free_cpumask_var(rd->rto_mask);
6707 free_online:
6708 free_cpumask_var(rd->online);
6709 free_span:
6710 free_cpumask_var(rd->span);
6711 out:
6712 return -ENOMEM;
6715 static void init_defrootdomain(void)
6717 init_rootdomain(&def_root_domain);
6719 atomic_set(&def_root_domain.refcount, 1);
6722 static struct root_domain *alloc_rootdomain(void)
6724 struct root_domain *rd;
6726 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6727 if (!rd)
6728 return NULL;
6730 if (init_rootdomain(rd) != 0) {
6731 kfree(rd);
6732 return NULL;
6735 return rd;
6738 static void free_sched_domain(struct rcu_head *rcu)
6740 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6741 if (atomic_dec_and_test(&sd->groups->ref))
6742 kfree(sd->groups);
6743 kfree(sd);
6746 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6748 call_rcu(&sd->rcu, free_sched_domain);
6751 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6753 for (; sd; sd = sd->parent)
6754 destroy_sched_domain(sd, cpu);
6758 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6759 * hold the hotplug lock.
6761 static void
6762 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6764 struct rq *rq = cpu_rq(cpu);
6765 struct sched_domain *tmp;
6767 /* Remove the sched domains which do not contribute to scheduling. */
6768 for (tmp = sd; tmp; ) {
6769 struct sched_domain *parent = tmp->parent;
6770 if (!parent)
6771 break;
6773 if (sd_parent_degenerate(tmp, parent)) {
6774 tmp->parent = parent->parent;
6775 if (parent->parent)
6776 parent->parent->child = tmp;
6777 destroy_sched_domain(parent, cpu);
6778 } else
6779 tmp = tmp->parent;
6782 if (sd && sd_degenerate(sd)) {
6783 tmp = sd;
6784 sd = sd->parent;
6785 destroy_sched_domain(tmp, cpu);
6786 if (sd)
6787 sd->child = NULL;
6790 sched_domain_debug(sd, cpu);
6792 rq_attach_root(rq, rd);
6793 tmp = rq->sd;
6794 rcu_assign_pointer(rq->sd, sd);
6795 destroy_sched_domains(tmp, cpu);
6798 /* cpus with isolated domains */
6799 static cpumask_var_t cpu_isolated_map;
6801 /* Setup the mask of cpus configured for isolated domains */
6802 static int __init isolated_cpu_setup(char *str)
6804 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6805 cpulist_parse(str, cpu_isolated_map);
6806 return 1;
6809 __setup("isolcpus=", isolated_cpu_setup);
6811 #define SD_NODES_PER_DOMAIN 16
6813 #ifdef CONFIG_NUMA
6816 * find_next_best_node - find the next node to include in a sched_domain
6817 * @node: node whose sched_domain we're building
6818 * @used_nodes: nodes already in the sched_domain
6820 * Find the next node to include in a given scheduling domain. Simply
6821 * finds the closest node not already in the @used_nodes map.
6823 * Should use nodemask_t.
6825 static int find_next_best_node(int node, nodemask_t *used_nodes)
6827 int i, n, val, min_val, best_node = -1;
6829 min_val = INT_MAX;
6831 for (i = 0; i < nr_node_ids; i++) {
6832 /* Start at @node */
6833 n = (node + i) % nr_node_ids;
6835 if (!nr_cpus_node(n))
6836 continue;
6838 /* Skip already used nodes */
6839 if (node_isset(n, *used_nodes))
6840 continue;
6842 /* Simple min distance search */
6843 val = node_distance(node, n);
6845 if (val < min_val) {
6846 min_val = val;
6847 best_node = n;
6851 if (best_node != -1)
6852 node_set(best_node, *used_nodes);
6853 return best_node;
6857 * sched_domain_node_span - get a cpumask for a node's sched_domain
6858 * @node: node whose cpumask we're constructing
6859 * @span: resulting cpumask
6861 * Given a node, construct a good cpumask for its sched_domain to span. It
6862 * should be one that prevents unnecessary balancing, but also spreads tasks
6863 * out optimally.
6865 static void sched_domain_node_span(int node, struct cpumask *span)
6867 nodemask_t used_nodes;
6868 int i;
6870 cpumask_clear(span);
6871 nodes_clear(used_nodes);
6873 cpumask_or(span, span, cpumask_of_node(node));
6874 node_set(node, used_nodes);
6876 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6877 int next_node = find_next_best_node(node, &used_nodes);
6878 if (next_node < 0)
6879 break;
6880 cpumask_or(span, span, cpumask_of_node(next_node));
6884 static const struct cpumask *cpu_node_mask(int cpu)
6886 lockdep_assert_held(&sched_domains_mutex);
6888 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
6890 return sched_domains_tmpmask;
6893 static const struct cpumask *cpu_allnodes_mask(int cpu)
6895 return cpu_possible_mask;
6897 #endif /* CONFIG_NUMA */
6899 static const struct cpumask *cpu_cpu_mask(int cpu)
6901 return cpumask_of_node(cpu_to_node(cpu));
6904 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6906 struct sd_data {
6907 struct sched_domain **__percpu sd;
6908 struct sched_group **__percpu sg;
6911 struct s_data {
6912 struct sched_domain ** __percpu sd;
6913 struct root_domain *rd;
6916 enum s_alloc {
6917 sa_rootdomain,
6918 sa_sd,
6919 sa_sd_storage,
6920 sa_none,
6923 struct sched_domain_topology_level;
6925 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
6926 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
6928 struct sched_domain_topology_level {
6929 sched_domain_init_f init;
6930 sched_domain_mask_f mask;
6931 struct sd_data data;
6935 * Assumes the sched_domain tree is fully constructed
6937 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6939 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6940 struct sched_domain *child = sd->child;
6942 if (child)
6943 cpu = cpumask_first(sched_domain_span(child));
6945 if (sg)
6946 *sg = *per_cpu_ptr(sdd->sg, cpu);
6948 return cpu;
6952 * build_sched_groups takes the cpumask we wish to span, and a pointer
6953 * to a function which identifies what group(along with sched group) a CPU
6954 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6955 * (due to the fact that we keep track of groups covered with a struct cpumask).
6957 * build_sched_groups will build a circular linked list of the groups
6958 * covered by the given span, and will set each group's ->cpumask correctly,
6959 * and ->cpu_power to 0.
6961 static void
6962 build_sched_groups(struct sched_domain *sd)
6964 struct sched_group *first = NULL, *last = NULL;
6965 struct sd_data *sdd = sd->private;
6966 const struct cpumask *span = sched_domain_span(sd);
6967 struct cpumask *covered;
6968 int i;
6970 lockdep_assert_held(&sched_domains_mutex);
6971 covered = sched_domains_tmpmask;
6973 cpumask_clear(covered);
6975 for_each_cpu(i, span) {
6976 struct sched_group *sg;
6977 int group = get_group(i, sdd, &sg);
6978 int j;
6980 if (cpumask_test_cpu(i, covered))
6981 continue;
6983 cpumask_clear(sched_group_cpus(sg));
6984 sg->cpu_power = 0;
6986 for_each_cpu(j, span) {
6987 if (get_group(j, sdd, NULL) != group)
6988 continue;
6990 cpumask_set_cpu(j, covered);
6991 cpumask_set_cpu(j, sched_group_cpus(sg));
6994 if (!first)
6995 first = sg;
6996 if (last)
6997 last->next = sg;
6998 last = sg;
7000 last->next = first;
7004 * Initialize sched groups cpu_power.
7006 * cpu_power indicates the capacity of sched group, which is used while
7007 * distributing the load between different sched groups in a sched domain.
7008 * Typically cpu_power for all the groups in a sched domain will be same unless
7009 * there are asymmetries in the topology. If there are asymmetries, group
7010 * having more cpu_power will pickup more load compared to the group having
7011 * less cpu_power.
7013 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7015 WARN_ON(!sd || !sd->groups);
7017 if (cpu != group_first_cpu(sd->groups))
7018 return;
7020 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
7022 update_group_power(sd, cpu);
7026 * Initializers for schedule domains
7027 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7030 #ifdef CONFIG_SCHED_DEBUG
7031 # define SD_INIT_NAME(sd, type) sd->name = #type
7032 #else
7033 # define SD_INIT_NAME(sd, type) do { } while (0)
7034 #endif
7036 #define SD_INIT_FUNC(type) \
7037 static noinline struct sched_domain * \
7038 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7040 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7041 *sd = SD_##type##_INIT; \
7042 SD_INIT_NAME(sd, type); \
7043 sd->private = &tl->data; \
7044 return sd; \
7047 SD_INIT_FUNC(CPU)
7048 #ifdef CONFIG_NUMA
7049 SD_INIT_FUNC(ALLNODES)
7050 SD_INIT_FUNC(NODE)
7051 #endif
7052 #ifdef CONFIG_SCHED_SMT
7053 SD_INIT_FUNC(SIBLING)
7054 #endif
7055 #ifdef CONFIG_SCHED_MC
7056 SD_INIT_FUNC(MC)
7057 #endif
7058 #ifdef CONFIG_SCHED_BOOK
7059 SD_INIT_FUNC(BOOK)
7060 #endif
7062 static int default_relax_domain_level = -1;
7063 int sched_domain_level_max;
7065 static int __init setup_relax_domain_level(char *str)
7067 unsigned long val;
7069 val = simple_strtoul(str, NULL, 0);
7070 if (val < sched_domain_level_max)
7071 default_relax_domain_level = val;
7073 return 1;
7075 __setup("relax_domain_level=", setup_relax_domain_level);
7077 static void set_domain_attribute(struct sched_domain *sd,
7078 struct sched_domain_attr *attr)
7080 int request;
7082 if (!attr || attr->relax_domain_level < 0) {
7083 if (default_relax_domain_level < 0)
7084 return;
7085 else
7086 request = default_relax_domain_level;
7087 } else
7088 request = attr->relax_domain_level;
7089 if (request < sd->level) {
7090 /* turn off idle balance on this domain */
7091 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7092 } else {
7093 /* turn on idle balance on this domain */
7094 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7098 static void __sdt_free(const struct cpumask *cpu_map);
7099 static int __sdt_alloc(const struct cpumask *cpu_map);
7101 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7102 const struct cpumask *cpu_map)
7104 switch (what) {
7105 case sa_rootdomain:
7106 if (!atomic_read(&d->rd->refcount))
7107 free_rootdomain(&d->rd->rcu); /* fall through */
7108 case sa_sd:
7109 free_percpu(d->sd); /* fall through */
7110 case sa_sd_storage:
7111 __sdt_free(cpu_map); /* fall through */
7112 case sa_none:
7113 break;
7117 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7118 const struct cpumask *cpu_map)
7120 memset(d, 0, sizeof(*d));
7122 if (__sdt_alloc(cpu_map))
7123 return sa_sd_storage;
7124 d->sd = alloc_percpu(struct sched_domain *);
7125 if (!d->sd)
7126 return sa_sd_storage;
7127 d->rd = alloc_rootdomain();
7128 if (!d->rd)
7129 return sa_sd;
7130 return sa_rootdomain;
7134 * NULL the sd_data elements we've used to build the sched_domain and
7135 * sched_group structure so that the subsequent __free_domain_allocs()
7136 * will not free the data we're using.
7138 static void claim_allocations(int cpu, struct sched_domain *sd)
7140 struct sd_data *sdd = sd->private;
7141 struct sched_group *sg = sd->groups;
7143 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7144 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7146 if (cpu == cpumask_first(sched_group_cpus(sg))) {
7147 WARN_ON_ONCE(*per_cpu_ptr(sdd->sg, cpu) != sg);
7148 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7152 #ifdef CONFIG_SCHED_SMT
7153 static const struct cpumask *cpu_smt_mask(int cpu)
7155 return topology_thread_cpumask(cpu);
7157 #endif
7160 * Topology list, bottom-up.
7162 static struct sched_domain_topology_level default_topology[] = {
7163 #ifdef CONFIG_SCHED_SMT
7164 { sd_init_SIBLING, cpu_smt_mask, },
7165 #endif
7166 #ifdef CONFIG_SCHED_MC
7167 { sd_init_MC, cpu_coregroup_mask, },
7168 #endif
7169 #ifdef CONFIG_SCHED_BOOK
7170 { sd_init_BOOK, cpu_book_mask, },
7171 #endif
7172 { sd_init_CPU, cpu_cpu_mask, },
7173 #ifdef CONFIG_NUMA
7174 { sd_init_NODE, cpu_node_mask, },
7175 { sd_init_ALLNODES, cpu_allnodes_mask, },
7176 #endif
7177 { NULL, },
7180 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7182 static int __sdt_alloc(const struct cpumask *cpu_map)
7184 struct sched_domain_topology_level *tl;
7185 int j;
7187 for (tl = sched_domain_topology; tl->init; tl++) {
7188 struct sd_data *sdd = &tl->data;
7190 sdd->sd = alloc_percpu(struct sched_domain *);
7191 if (!sdd->sd)
7192 return -ENOMEM;
7194 sdd->sg = alloc_percpu(struct sched_group *);
7195 if (!sdd->sg)
7196 return -ENOMEM;
7198 for_each_cpu(j, cpu_map) {
7199 struct sched_domain *sd;
7200 struct sched_group *sg;
7202 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7203 GFP_KERNEL, cpu_to_node(j));
7204 if (!sd)
7205 return -ENOMEM;
7207 *per_cpu_ptr(sdd->sd, j) = sd;
7209 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7210 GFP_KERNEL, cpu_to_node(j));
7211 if (!sg)
7212 return -ENOMEM;
7214 *per_cpu_ptr(sdd->sg, j) = sg;
7218 return 0;
7221 static void __sdt_free(const struct cpumask *cpu_map)
7223 struct sched_domain_topology_level *tl;
7224 int j;
7226 for (tl = sched_domain_topology; tl->init; tl++) {
7227 struct sd_data *sdd = &tl->data;
7229 for_each_cpu(j, cpu_map) {
7230 kfree(*per_cpu_ptr(sdd->sd, j));
7231 kfree(*per_cpu_ptr(sdd->sg, j));
7233 free_percpu(sdd->sd);
7234 free_percpu(sdd->sg);
7238 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7239 struct s_data *d, const struct cpumask *cpu_map,
7240 struct sched_domain_attr *attr, struct sched_domain *child,
7241 int cpu)
7243 struct sched_domain *sd = tl->init(tl, cpu);
7244 if (!sd)
7245 return child;
7247 set_domain_attribute(sd, attr);
7248 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7249 if (child) {
7250 sd->level = child->level + 1;
7251 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7252 child->parent = sd;
7254 sd->child = child;
7256 return sd;
7260 * Build sched domains for a given set of cpus and attach the sched domains
7261 * to the individual cpus
7263 static int build_sched_domains(const struct cpumask *cpu_map,
7264 struct sched_domain_attr *attr)
7266 enum s_alloc alloc_state = sa_none;
7267 struct sched_domain *sd;
7268 struct s_data d;
7269 int i, ret = -ENOMEM;
7271 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7272 if (alloc_state != sa_rootdomain)
7273 goto error;
7275 /* Set up domains for cpus specified by the cpu_map. */
7276 for_each_cpu(i, cpu_map) {
7277 struct sched_domain_topology_level *tl;
7279 sd = NULL;
7280 for (tl = sched_domain_topology; tl->init; tl++)
7281 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7283 while (sd->child)
7284 sd = sd->child;
7286 *per_cpu_ptr(d.sd, i) = sd;
7289 /* Build the groups for the domains */
7290 for_each_cpu(i, cpu_map) {
7291 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7292 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7293 get_group(i, sd->private, &sd->groups);
7294 atomic_inc(&sd->groups->ref);
7296 if (i != cpumask_first(sched_domain_span(sd)))
7297 continue;
7299 build_sched_groups(sd);
7303 /* Calculate CPU power for physical packages and nodes */
7304 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7305 if (!cpumask_test_cpu(i, cpu_map))
7306 continue;
7308 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7309 claim_allocations(i, sd);
7310 init_sched_groups_power(i, sd);
7314 /* Attach the domains */
7315 rcu_read_lock();
7316 for_each_cpu(i, cpu_map) {
7317 sd = *per_cpu_ptr(d.sd, i);
7318 cpu_attach_domain(sd, d.rd, i);
7320 rcu_read_unlock();
7322 ret = 0;
7323 error:
7324 __free_domain_allocs(&d, alloc_state, cpu_map);
7325 return ret;
7328 static cpumask_var_t *doms_cur; /* current sched domains */
7329 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7330 static struct sched_domain_attr *dattr_cur;
7331 /* attribues of custom domains in 'doms_cur' */
7334 * Special case: If a kmalloc of a doms_cur partition (array of
7335 * cpumask) fails, then fallback to a single sched domain,
7336 * as determined by the single cpumask fallback_doms.
7338 static cpumask_var_t fallback_doms;
7341 * arch_update_cpu_topology lets virtualized architectures update the
7342 * cpu core maps. It is supposed to return 1 if the topology changed
7343 * or 0 if it stayed the same.
7345 int __attribute__((weak)) arch_update_cpu_topology(void)
7347 return 0;
7350 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7352 int i;
7353 cpumask_var_t *doms;
7355 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7356 if (!doms)
7357 return NULL;
7358 for (i = 0; i < ndoms; i++) {
7359 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7360 free_sched_domains(doms, i);
7361 return NULL;
7364 return doms;
7367 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7369 unsigned int i;
7370 for (i = 0; i < ndoms; i++)
7371 free_cpumask_var(doms[i]);
7372 kfree(doms);
7376 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7377 * For now this just excludes isolated cpus, but could be used to
7378 * exclude other special cases in the future.
7380 static int init_sched_domains(const struct cpumask *cpu_map)
7382 int err;
7384 arch_update_cpu_topology();
7385 ndoms_cur = 1;
7386 doms_cur = alloc_sched_domains(ndoms_cur);
7387 if (!doms_cur)
7388 doms_cur = &fallback_doms;
7389 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7390 dattr_cur = NULL;
7391 err = build_sched_domains(doms_cur[0], NULL);
7392 register_sched_domain_sysctl();
7394 return err;
7398 * Detach sched domains from a group of cpus specified in cpu_map
7399 * These cpus will now be attached to the NULL domain
7401 static void detach_destroy_domains(const struct cpumask *cpu_map)
7403 int i;
7405 rcu_read_lock();
7406 for_each_cpu(i, cpu_map)
7407 cpu_attach_domain(NULL, &def_root_domain, i);
7408 rcu_read_unlock();
7411 /* handle null as "default" */
7412 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7413 struct sched_domain_attr *new, int idx_new)
7415 struct sched_domain_attr tmp;
7417 /* fast path */
7418 if (!new && !cur)
7419 return 1;
7421 tmp = SD_ATTR_INIT;
7422 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7423 new ? (new + idx_new) : &tmp,
7424 sizeof(struct sched_domain_attr));
7428 * Partition sched domains as specified by the 'ndoms_new'
7429 * cpumasks in the array doms_new[] of cpumasks. This compares
7430 * doms_new[] to the current sched domain partitioning, doms_cur[].
7431 * It destroys each deleted domain and builds each new domain.
7433 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7434 * The masks don't intersect (don't overlap.) We should setup one
7435 * sched domain for each mask. CPUs not in any of the cpumasks will
7436 * not be load balanced. If the same cpumask appears both in the
7437 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7438 * it as it is.
7440 * The passed in 'doms_new' should be allocated using
7441 * alloc_sched_domains. This routine takes ownership of it and will
7442 * free_sched_domains it when done with it. If the caller failed the
7443 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7444 * and partition_sched_domains() will fallback to the single partition
7445 * 'fallback_doms', it also forces the domains to be rebuilt.
7447 * If doms_new == NULL it will be replaced with cpu_online_mask.
7448 * ndoms_new == 0 is a special case for destroying existing domains,
7449 * and it will not create the default domain.
7451 * Call with hotplug lock held
7453 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7454 struct sched_domain_attr *dattr_new)
7456 int i, j, n;
7457 int new_topology;
7459 mutex_lock(&sched_domains_mutex);
7461 /* always unregister in case we don't destroy any domains */
7462 unregister_sched_domain_sysctl();
7464 /* Let architecture update cpu core mappings. */
7465 new_topology = arch_update_cpu_topology();
7467 n = doms_new ? ndoms_new : 0;
7469 /* Destroy deleted domains */
7470 for (i = 0; i < ndoms_cur; i++) {
7471 for (j = 0; j < n && !new_topology; j++) {
7472 if (cpumask_equal(doms_cur[i], doms_new[j])
7473 && dattrs_equal(dattr_cur, i, dattr_new, j))
7474 goto match1;
7476 /* no match - a current sched domain not in new doms_new[] */
7477 detach_destroy_domains(doms_cur[i]);
7478 match1:
7482 if (doms_new == NULL) {
7483 ndoms_cur = 0;
7484 doms_new = &fallback_doms;
7485 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7486 WARN_ON_ONCE(dattr_new);
7489 /* Build new domains */
7490 for (i = 0; i < ndoms_new; i++) {
7491 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7492 if (cpumask_equal(doms_new[i], doms_cur[j])
7493 && dattrs_equal(dattr_new, i, dattr_cur, j))
7494 goto match2;
7496 /* no match - add a new doms_new */
7497 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7498 match2:
7502 /* Remember the new sched domains */
7503 if (doms_cur != &fallback_doms)
7504 free_sched_domains(doms_cur, ndoms_cur);
7505 kfree(dattr_cur); /* kfree(NULL) is safe */
7506 doms_cur = doms_new;
7507 dattr_cur = dattr_new;
7508 ndoms_cur = ndoms_new;
7510 register_sched_domain_sysctl();
7512 mutex_unlock(&sched_domains_mutex);
7515 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7516 static void reinit_sched_domains(void)
7518 get_online_cpus();
7520 /* Destroy domains first to force the rebuild */
7521 partition_sched_domains(0, NULL, NULL);
7523 rebuild_sched_domains();
7524 put_online_cpus();
7527 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7529 unsigned int level = 0;
7531 if (sscanf(buf, "%u", &level) != 1)
7532 return -EINVAL;
7535 * level is always be positive so don't check for
7536 * level < POWERSAVINGS_BALANCE_NONE which is 0
7537 * What happens on 0 or 1 byte write,
7538 * need to check for count as well?
7541 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7542 return -EINVAL;
7544 if (smt)
7545 sched_smt_power_savings = level;
7546 else
7547 sched_mc_power_savings = level;
7549 reinit_sched_domains();
7551 return count;
7554 #ifdef CONFIG_SCHED_MC
7555 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7556 struct sysdev_class_attribute *attr,
7557 char *page)
7559 return sprintf(page, "%u\n", sched_mc_power_savings);
7561 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7562 struct sysdev_class_attribute *attr,
7563 const char *buf, size_t count)
7565 return sched_power_savings_store(buf, count, 0);
7567 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7568 sched_mc_power_savings_show,
7569 sched_mc_power_savings_store);
7570 #endif
7572 #ifdef CONFIG_SCHED_SMT
7573 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7574 struct sysdev_class_attribute *attr,
7575 char *page)
7577 return sprintf(page, "%u\n", sched_smt_power_savings);
7579 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7580 struct sysdev_class_attribute *attr,
7581 const char *buf, size_t count)
7583 return sched_power_savings_store(buf, count, 1);
7585 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7586 sched_smt_power_savings_show,
7587 sched_smt_power_savings_store);
7588 #endif
7590 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7592 int err = 0;
7594 #ifdef CONFIG_SCHED_SMT
7595 if (smt_capable())
7596 err = sysfs_create_file(&cls->kset.kobj,
7597 &attr_sched_smt_power_savings.attr);
7598 #endif
7599 #ifdef CONFIG_SCHED_MC
7600 if (!err && mc_capable())
7601 err = sysfs_create_file(&cls->kset.kobj,
7602 &attr_sched_mc_power_savings.attr);
7603 #endif
7604 return err;
7606 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7609 * Update cpusets according to cpu_active mask. If cpusets are
7610 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7611 * around partition_sched_domains().
7613 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7614 void *hcpu)
7616 switch (action & ~CPU_TASKS_FROZEN) {
7617 case CPU_ONLINE:
7618 case CPU_DOWN_FAILED:
7619 cpuset_update_active_cpus();
7620 return NOTIFY_OK;
7621 default:
7622 return NOTIFY_DONE;
7626 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7627 void *hcpu)
7629 switch (action & ~CPU_TASKS_FROZEN) {
7630 case CPU_DOWN_PREPARE:
7631 cpuset_update_active_cpus();
7632 return NOTIFY_OK;
7633 default:
7634 return NOTIFY_DONE;
7638 static int update_runtime(struct notifier_block *nfb,
7639 unsigned long action, void *hcpu)
7641 int cpu = (int)(long)hcpu;
7643 switch (action) {
7644 case CPU_DOWN_PREPARE:
7645 case CPU_DOWN_PREPARE_FROZEN:
7646 disable_runtime(cpu_rq(cpu));
7647 return NOTIFY_OK;
7649 case CPU_DOWN_FAILED:
7650 case CPU_DOWN_FAILED_FROZEN:
7651 case CPU_ONLINE:
7652 case CPU_ONLINE_FROZEN:
7653 enable_runtime(cpu_rq(cpu));
7654 return NOTIFY_OK;
7656 default:
7657 return NOTIFY_DONE;
7661 void __init sched_init_smp(void)
7663 cpumask_var_t non_isolated_cpus;
7665 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7666 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7668 get_online_cpus();
7669 mutex_lock(&sched_domains_mutex);
7670 init_sched_domains(cpu_active_mask);
7671 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7672 if (cpumask_empty(non_isolated_cpus))
7673 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7674 mutex_unlock(&sched_domains_mutex);
7675 put_online_cpus();
7677 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7678 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7680 /* RT runtime code needs to handle some hotplug events */
7681 hotcpu_notifier(update_runtime, 0);
7683 init_hrtick();
7685 /* Move init over to a non-isolated CPU */
7686 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7687 BUG();
7688 sched_init_granularity();
7689 free_cpumask_var(non_isolated_cpus);
7691 init_sched_rt_class();
7693 #else
7694 void __init sched_init_smp(void)
7696 sched_init_granularity();
7698 #endif /* CONFIG_SMP */
7700 const_debug unsigned int sysctl_timer_migration = 1;
7702 int in_sched_functions(unsigned long addr)
7704 return in_lock_functions(addr) ||
7705 (addr >= (unsigned long)__sched_text_start
7706 && addr < (unsigned long)__sched_text_end);
7709 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7711 cfs_rq->tasks_timeline = RB_ROOT;
7712 INIT_LIST_HEAD(&cfs_rq->tasks);
7713 #ifdef CONFIG_FAIR_GROUP_SCHED
7714 cfs_rq->rq = rq;
7715 /* allow initial update_cfs_load() to truncate */
7716 #ifdef CONFIG_SMP
7717 cfs_rq->load_stamp = 1;
7718 #endif
7719 #endif
7720 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7723 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7725 struct rt_prio_array *array;
7726 int i;
7728 array = &rt_rq->active;
7729 for (i = 0; i < MAX_RT_PRIO; i++) {
7730 INIT_LIST_HEAD(array->queue + i);
7731 __clear_bit(i, array->bitmap);
7733 /* delimiter for bitsearch: */
7734 __set_bit(MAX_RT_PRIO, array->bitmap);
7736 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7737 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7738 #ifdef CONFIG_SMP
7739 rt_rq->highest_prio.next = MAX_RT_PRIO;
7740 #endif
7741 #endif
7742 #ifdef CONFIG_SMP
7743 rt_rq->rt_nr_migratory = 0;
7744 rt_rq->overloaded = 0;
7745 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7746 #endif
7748 rt_rq->rt_time = 0;
7749 rt_rq->rt_throttled = 0;
7750 rt_rq->rt_runtime = 0;
7751 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7753 #ifdef CONFIG_RT_GROUP_SCHED
7754 rt_rq->rt_nr_boosted = 0;
7755 rt_rq->rq = rq;
7756 #endif
7759 #ifdef CONFIG_FAIR_GROUP_SCHED
7760 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7761 struct sched_entity *se, int cpu,
7762 struct sched_entity *parent)
7764 struct rq *rq = cpu_rq(cpu);
7765 tg->cfs_rq[cpu] = cfs_rq;
7766 init_cfs_rq(cfs_rq, rq);
7767 cfs_rq->tg = tg;
7769 tg->se[cpu] = se;
7770 /* se could be NULL for root_task_group */
7771 if (!se)
7772 return;
7774 if (!parent)
7775 se->cfs_rq = &rq->cfs;
7776 else
7777 se->cfs_rq = parent->my_q;
7779 se->my_q = cfs_rq;
7780 update_load_set(&se->load, 0);
7781 se->parent = parent;
7783 #endif
7785 #ifdef CONFIG_RT_GROUP_SCHED
7786 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7787 struct sched_rt_entity *rt_se, int cpu,
7788 struct sched_rt_entity *parent)
7790 struct rq *rq = cpu_rq(cpu);
7792 tg->rt_rq[cpu] = rt_rq;
7793 init_rt_rq(rt_rq, rq);
7794 rt_rq->tg = tg;
7795 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7797 tg->rt_se[cpu] = rt_se;
7798 if (!rt_se)
7799 return;
7801 if (!parent)
7802 rt_se->rt_rq = &rq->rt;
7803 else
7804 rt_se->rt_rq = parent->my_q;
7806 rt_se->my_q = rt_rq;
7807 rt_se->parent = parent;
7808 INIT_LIST_HEAD(&rt_se->run_list);
7810 #endif
7812 void __init sched_init(void)
7814 int i, j;
7815 unsigned long alloc_size = 0, ptr;
7817 #ifdef CONFIG_FAIR_GROUP_SCHED
7818 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7819 #endif
7820 #ifdef CONFIG_RT_GROUP_SCHED
7821 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7822 #endif
7823 #ifdef CONFIG_CPUMASK_OFFSTACK
7824 alloc_size += num_possible_cpus() * cpumask_size();
7825 #endif
7826 if (alloc_size) {
7827 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7829 #ifdef CONFIG_FAIR_GROUP_SCHED
7830 root_task_group.se = (struct sched_entity **)ptr;
7831 ptr += nr_cpu_ids * sizeof(void **);
7833 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7834 ptr += nr_cpu_ids * sizeof(void **);
7836 #endif /* CONFIG_FAIR_GROUP_SCHED */
7837 #ifdef CONFIG_RT_GROUP_SCHED
7838 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7839 ptr += nr_cpu_ids * sizeof(void **);
7841 root_task_group.rt_rq = (struct rt_rq **)ptr;
7842 ptr += nr_cpu_ids * sizeof(void **);
7844 #endif /* CONFIG_RT_GROUP_SCHED */
7845 #ifdef CONFIG_CPUMASK_OFFSTACK
7846 for_each_possible_cpu(i) {
7847 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7848 ptr += cpumask_size();
7850 #endif /* CONFIG_CPUMASK_OFFSTACK */
7853 #ifdef CONFIG_SMP
7854 init_defrootdomain();
7855 #endif
7857 init_rt_bandwidth(&def_rt_bandwidth,
7858 global_rt_period(), global_rt_runtime());
7860 #ifdef CONFIG_RT_GROUP_SCHED
7861 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7862 global_rt_period(), global_rt_runtime());
7863 #endif /* CONFIG_RT_GROUP_SCHED */
7865 #ifdef CONFIG_CGROUP_SCHED
7866 list_add(&root_task_group.list, &task_groups);
7867 INIT_LIST_HEAD(&root_task_group.children);
7868 autogroup_init(&init_task);
7869 #endif /* CONFIG_CGROUP_SCHED */
7871 for_each_possible_cpu(i) {
7872 struct rq *rq;
7874 rq = cpu_rq(i);
7875 raw_spin_lock_init(&rq->lock);
7876 rq->nr_running = 0;
7877 rq->calc_load_active = 0;
7878 rq->calc_load_update = jiffies + LOAD_FREQ;
7879 init_cfs_rq(&rq->cfs, rq);
7880 init_rt_rq(&rq->rt, rq);
7881 #ifdef CONFIG_FAIR_GROUP_SCHED
7882 root_task_group.shares = root_task_group_load;
7883 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7885 * How much cpu bandwidth does root_task_group get?
7887 * In case of task-groups formed thr' the cgroup filesystem, it
7888 * gets 100% of the cpu resources in the system. This overall
7889 * system cpu resource is divided among the tasks of
7890 * root_task_group and its child task-groups in a fair manner,
7891 * based on each entity's (task or task-group's) weight
7892 * (se->load.weight).
7894 * In other words, if root_task_group has 10 tasks of weight
7895 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7896 * then A0's share of the cpu resource is:
7898 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7900 * We achieve this by letting root_task_group's tasks sit
7901 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7903 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7904 #endif /* CONFIG_FAIR_GROUP_SCHED */
7906 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7907 #ifdef CONFIG_RT_GROUP_SCHED
7908 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7909 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7910 #endif
7912 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7913 rq->cpu_load[j] = 0;
7915 rq->last_load_update_tick = jiffies;
7917 #ifdef CONFIG_SMP
7918 rq->sd = NULL;
7919 rq->rd = NULL;
7920 rq->cpu_power = SCHED_POWER_SCALE;
7921 rq->post_schedule = 0;
7922 rq->active_balance = 0;
7923 rq->next_balance = jiffies;
7924 rq->push_cpu = 0;
7925 rq->cpu = i;
7926 rq->online = 0;
7927 rq->idle_stamp = 0;
7928 rq->avg_idle = 2*sysctl_sched_migration_cost;
7929 rq_attach_root(rq, &def_root_domain);
7930 #ifdef CONFIG_NO_HZ
7931 rq->nohz_balance_kick = 0;
7932 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
7933 #endif
7934 #endif
7935 init_rq_hrtick(rq);
7936 atomic_set(&rq->nr_iowait, 0);
7939 set_load_weight(&init_task);
7941 #ifdef CONFIG_PREEMPT_NOTIFIERS
7942 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7943 #endif
7945 #ifdef CONFIG_SMP
7946 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7947 #endif
7949 #ifdef CONFIG_RT_MUTEXES
7950 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7951 #endif
7954 * The boot idle thread does lazy MMU switching as well:
7956 atomic_inc(&init_mm.mm_count);
7957 enter_lazy_tlb(&init_mm, current);
7960 * Make us the idle thread. Technically, schedule() should not be
7961 * called from this thread, however somewhere below it might be,
7962 * but because we are the idle thread, we just pick up running again
7963 * when this runqueue becomes "idle".
7965 init_idle(current, smp_processor_id());
7967 calc_load_update = jiffies + LOAD_FREQ;
7970 * During early bootup we pretend to be a normal task:
7972 current->sched_class = &fair_sched_class;
7974 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7975 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7976 #ifdef CONFIG_SMP
7977 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7978 #ifdef CONFIG_NO_HZ
7979 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7980 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
7981 atomic_set(&nohz.load_balancer, nr_cpu_ids);
7982 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
7983 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
7984 #endif
7985 /* May be allocated at isolcpus cmdline parse time */
7986 if (cpu_isolated_map == NULL)
7987 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7988 #endif /* SMP */
7990 scheduler_running = 1;
7993 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7994 static inline int preempt_count_equals(int preempt_offset)
7996 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7998 return (nested == preempt_offset);
8001 void __might_sleep(const char *file, int line, int preempt_offset)
8003 #ifdef in_atomic
8004 static unsigned long prev_jiffy; /* ratelimiting */
8006 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8007 system_state != SYSTEM_RUNNING || oops_in_progress)
8008 return;
8009 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8010 return;
8011 prev_jiffy = jiffies;
8013 printk(KERN_ERR
8014 "BUG: sleeping function called from invalid context at %s:%d\n",
8015 file, line);
8016 printk(KERN_ERR
8017 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8018 in_atomic(), irqs_disabled(),
8019 current->pid, current->comm);
8021 debug_show_held_locks(current);
8022 if (irqs_disabled())
8023 print_irqtrace_events(current);
8024 dump_stack();
8025 #endif
8027 EXPORT_SYMBOL(__might_sleep);
8028 #endif
8030 #ifdef CONFIG_MAGIC_SYSRQ
8031 static void normalize_task(struct rq *rq, struct task_struct *p)
8033 const struct sched_class *prev_class = p->sched_class;
8034 int old_prio = p->prio;
8035 int on_rq;
8037 on_rq = p->on_rq;
8038 if (on_rq)
8039 deactivate_task(rq, p, 0);
8040 __setscheduler(rq, p, SCHED_NORMAL, 0);
8041 if (on_rq) {
8042 activate_task(rq, p, 0);
8043 resched_task(rq->curr);
8046 check_class_changed(rq, p, prev_class, old_prio);
8049 void normalize_rt_tasks(void)
8051 struct task_struct *g, *p;
8052 unsigned long flags;
8053 struct rq *rq;
8055 read_lock_irqsave(&tasklist_lock, flags);
8056 do_each_thread(g, p) {
8058 * Only normalize user tasks:
8060 if (!p->mm)
8061 continue;
8063 p->se.exec_start = 0;
8064 #ifdef CONFIG_SCHEDSTATS
8065 p->se.statistics.wait_start = 0;
8066 p->se.statistics.sleep_start = 0;
8067 p->se.statistics.block_start = 0;
8068 #endif
8070 if (!rt_task(p)) {
8072 * Renice negative nice level userspace
8073 * tasks back to 0:
8075 if (TASK_NICE(p) < 0 && p->mm)
8076 set_user_nice(p, 0);
8077 continue;
8080 raw_spin_lock(&p->pi_lock);
8081 rq = __task_rq_lock(p);
8083 normalize_task(rq, p);
8085 __task_rq_unlock(rq);
8086 raw_spin_unlock(&p->pi_lock);
8087 } while_each_thread(g, p);
8089 read_unlock_irqrestore(&tasklist_lock, flags);
8092 #endif /* CONFIG_MAGIC_SYSRQ */
8094 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8096 * These functions are only useful for the IA64 MCA handling, or kdb.
8098 * They can only be called when the whole system has been
8099 * stopped - every CPU needs to be quiescent, and no scheduling
8100 * activity can take place. Using them for anything else would
8101 * be a serious bug, and as a result, they aren't even visible
8102 * under any other configuration.
8106 * curr_task - return the current task for a given cpu.
8107 * @cpu: the processor in question.
8109 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8111 struct task_struct *curr_task(int cpu)
8113 return cpu_curr(cpu);
8116 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8118 #ifdef CONFIG_IA64
8120 * set_curr_task - set the current task for a given cpu.
8121 * @cpu: the processor in question.
8122 * @p: the task pointer to set.
8124 * Description: This function must only be used when non-maskable interrupts
8125 * are serviced on a separate stack. It allows the architecture to switch the
8126 * notion of the current task on a cpu in a non-blocking manner. This function
8127 * must be called with all CPU's synchronized, and interrupts disabled, the
8128 * and caller must save the original value of the current task (see
8129 * curr_task() above) and restore that value before reenabling interrupts and
8130 * re-starting the system.
8132 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8134 void set_curr_task(int cpu, struct task_struct *p)
8136 cpu_curr(cpu) = p;
8139 #endif
8141 #ifdef CONFIG_FAIR_GROUP_SCHED
8142 static void free_fair_sched_group(struct task_group *tg)
8144 int i;
8146 for_each_possible_cpu(i) {
8147 if (tg->cfs_rq)
8148 kfree(tg->cfs_rq[i]);
8149 if (tg->se)
8150 kfree(tg->se[i]);
8153 kfree(tg->cfs_rq);
8154 kfree(tg->se);
8157 static
8158 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8160 struct cfs_rq *cfs_rq;
8161 struct sched_entity *se;
8162 int i;
8164 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8165 if (!tg->cfs_rq)
8166 goto err;
8167 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8168 if (!tg->se)
8169 goto err;
8171 tg->shares = NICE_0_LOAD;
8173 for_each_possible_cpu(i) {
8174 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8175 GFP_KERNEL, cpu_to_node(i));
8176 if (!cfs_rq)
8177 goto err;
8179 se = kzalloc_node(sizeof(struct sched_entity),
8180 GFP_KERNEL, cpu_to_node(i));
8181 if (!se)
8182 goto err_free_rq;
8184 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8187 return 1;
8189 err_free_rq:
8190 kfree(cfs_rq);
8191 err:
8192 return 0;
8195 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8197 struct rq *rq = cpu_rq(cpu);
8198 unsigned long flags;
8201 * Only empty task groups can be destroyed; so we can speculatively
8202 * check on_list without danger of it being re-added.
8204 if (!tg->cfs_rq[cpu]->on_list)
8205 return;
8207 raw_spin_lock_irqsave(&rq->lock, flags);
8208 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8209 raw_spin_unlock_irqrestore(&rq->lock, flags);
8211 #else /* !CONFG_FAIR_GROUP_SCHED */
8212 static inline void free_fair_sched_group(struct task_group *tg)
8216 static inline
8217 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8219 return 1;
8222 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8225 #endif /* CONFIG_FAIR_GROUP_SCHED */
8227 #ifdef CONFIG_RT_GROUP_SCHED
8228 static void free_rt_sched_group(struct task_group *tg)
8230 int i;
8232 destroy_rt_bandwidth(&tg->rt_bandwidth);
8234 for_each_possible_cpu(i) {
8235 if (tg->rt_rq)
8236 kfree(tg->rt_rq[i]);
8237 if (tg->rt_se)
8238 kfree(tg->rt_se[i]);
8241 kfree(tg->rt_rq);
8242 kfree(tg->rt_se);
8245 static
8246 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8248 struct rt_rq *rt_rq;
8249 struct sched_rt_entity *rt_se;
8250 int i;
8252 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8253 if (!tg->rt_rq)
8254 goto err;
8255 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8256 if (!tg->rt_se)
8257 goto err;
8259 init_rt_bandwidth(&tg->rt_bandwidth,
8260 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8262 for_each_possible_cpu(i) {
8263 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8264 GFP_KERNEL, cpu_to_node(i));
8265 if (!rt_rq)
8266 goto err;
8268 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8269 GFP_KERNEL, cpu_to_node(i));
8270 if (!rt_se)
8271 goto err_free_rq;
8273 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8276 return 1;
8278 err_free_rq:
8279 kfree(rt_rq);
8280 err:
8281 return 0;
8283 #else /* !CONFIG_RT_GROUP_SCHED */
8284 static inline void free_rt_sched_group(struct task_group *tg)
8288 static inline
8289 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8291 return 1;
8293 #endif /* CONFIG_RT_GROUP_SCHED */
8295 #ifdef CONFIG_CGROUP_SCHED
8296 static void free_sched_group(struct task_group *tg)
8298 free_fair_sched_group(tg);
8299 free_rt_sched_group(tg);
8300 autogroup_free(tg);
8301 kfree(tg);
8304 /* allocate runqueue etc for a new task group */
8305 struct task_group *sched_create_group(struct task_group *parent)
8307 struct task_group *tg;
8308 unsigned long flags;
8310 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8311 if (!tg)
8312 return ERR_PTR(-ENOMEM);
8314 if (!alloc_fair_sched_group(tg, parent))
8315 goto err;
8317 if (!alloc_rt_sched_group(tg, parent))
8318 goto err;
8320 spin_lock_irqsave(&task_group_lock, flags);
8321 list_add_rcu(&tg->list, &task_groups);
8323 WARN_ON(!parent); /* root should already exist */
8325 tg->parent = parent;
8326 INIT_LIST_HEAD(&tg->children);
8327 list_add_rcu(&tg->siblings, &parent->children);
8328 spin_unlock_irqrestore(&task_group_lock, flags);
8330 return tg;
8332 err:
8333 free_sched_group(tg);
8334 return ERR_PTR(-ENOMEM);
8337 /* rcu callback to free various structures associated with a task group */
8338 static void free_sched_group_rcu(struct rcu_head *rhp)
8340 /* now it should be safe to free those cfs_rqs */
8341 free_sched_group(container_of(rhp, struct task_group, rcu));
8344 /* Destroy runqueue etc associated with a task group */
8345 void sched_destroy_group(struct task_group *tg)
8347 unsigned long flags;
8348 int i;
8350 /* end participation in shares distribution */
8351 for_each_possible_cpu(i)
8352 unregister_fair_sched_group(tg, i);
8354 spin_lock_irqsave(&task_group_lock, flags);
8355 list_del_rcu(&tg->list);
8356 list_del_rcu(&tg->siblings);
8357 spin_unlock_irqrestore(&task_group_lock, flags);
8359 /* wait for possible concurrent references to cfs_rqs complete */
8360 call_rcu(&tg->rcu, free_sched_group_rcu);
8363 /* change task's runqueue when it moves between groups.
8364 * The caller of this function should have put the task in its new group
8365 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8366 * reflect its new group.
8368 void sched_move_task(struct task_struct *tsk)
8370 int on_rq, running;
8371 unsigned long flags;
8372 struct rq *rq;
8374 rq = task_rq_lock(tsk, &flags);
8376 running = task_current(rq, tsk);
8377 on_rq = tsk->on_rq;
8379 if (on_rq)
8380 dequeue_task(rq, tsk, 0);
8381 if (unlikely(running))
8382 tsk->sched_class->put_prev_task(rq, tsk);
8384 #ifdef CONFIG_FAIR_GROUP_SCHED
8385 if (tsk->sched_class->task_move_group)
8386 tsk->sched_class->task_move_group(tsk, on_rq);
8387 else
8388 #endif
8389 set_task_rq(tsk, task_cpu(tsk));
8391 if (unlikely(running))
8392 tsk->sched_class->set_curr_task(rq);
8393 if (on_rq)
8394 enqueue_task(rq, tsk, 0);
8396 task_rq_unlock(rq, tsk, &flags);
8398 #endif /* CONFIG_CGROUP_SCHED */
8400 #ifdef CONFIG_FAIR_GROUP_SCHED
8401 static DEFINE_MUTEX(shares_mutex);
8403 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8405 int i;
8406 unsigned long flags;
8409 * We can't change the weight of the root cgroup.
8411 if (!tg->se[0])
8412 return -EINVAL;
8414 if (shares < MIN_SHARES)
8415 shares = MIN_SHARES;
8416 else if (shares > MAX_SHARES)
8417 shares = MAX_SHARES;
8419 mutex_lock(&shares_mutex);
8420 if (tg->shares == shares)
8421 goto done;
8423 tg->shares = shares;
8424 for_each_possible_cpu(i) {
8425 struct rq *rq = cpu_rq(i);
8426 struct sched_entity *se;
8428 se = tg->se[i];
8429 /* Propagate contribution to hierarchy */
8430 raw_spin_lock_irqsave(&rq->lock, flags);
8431 for_each_sched_entity(se)
8432 update_cfs_shares(group_cfs_rq(se));
8433 raw_spin_unlock_irqrestore(&rq->lock, flags);
8436 done:
8437 mutex_unlock(&shares_mutex);
8438 return 0;
8441 unsigned long sched_group_shares(struct task_group *tg)
8443 return tg->shares;
8445 #endif
8447 #ifdef CONFIG_RT_GROUP_SCHED
8449 * Ensure that the real time constraints are schedulable.
8451 static DEFINE_MUTEX(rt_constraints_mutex);
8453 static unsigned long to_ratio(u64 period, u64 runtime)
8455 if (runtime == RUNTIME_INF)
8456 return 1ULL << 20;
8458 return div64_u64(runtime << 20, period);
8461 /* Must be called with tasklist_lock held */
8462 static inline int tg_has_rt_tasks(struct task_group *tg)
8464 struct task_struct *g, *p;
8466 do_each_thread(g, p) {
8467 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8468 return 1;
8469 } while_each_thread(g, p);
8471 return 0;
8474 struct rt_schedulable_data {
8475 struct task_group *tg;
8476 u64 rt_period;
8477 u64 rt_runtime;
8480 static int tg_schedulable(struct task_group *tg, void *data)
8482 struct rt_schedulable_data *d = data;
8483 struct task_group *child;
8484 unsigned long total, sum = 0;
8485 u64 period, runtime;
8487 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8488 runtime = tg->rt_bandwidth.rt_runtime;
8490 if (tg == d->tg) {
8491 period = d->rt_period;
8492 runtime = d->rt_runtime;
8496 * Cannot have more runtime than the period.
8498 if (runtime > period && runtime != RUNTIME_INF)
8499 return -EINVAL;
8502 * Ensure we don't starve existing RT tasks.
8504 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8505 return -EBUSY;
8507 total = to_ratio(period, runtime);
8510 * Nobody can have more than the global setting allows.
8512 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8513 return -EINVAL;
8516 * The sum of our children's runtime should not exceed our own.
8518 list_for_each_entry_rcu(child, &tg->children, siblings) {
8519 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8520 runtime = child->rt_bandwidth.rt_runtime;
8522 if (child == d->tg) {
8523 period = d->rt_period;
8524 runtime = d->rt_runtime;
8527 sum += to_ratio(period, runtime);
8530 if (sum > total)
8531 return -EINVAL;
8533 return 0;
8536 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8538 struct rt_schedulable_data data = {
8539 .tg = tg,
8540 .rt_period = period,
8541 .rt_runtime = runtime,
8544 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8547 static int tg_set_bandwidth(struct task_group *tg,
8548 u64 rt_period, u64 rt_runtime)
8550 int i, err = 0;
8552 mutex_lock(&rt_constraints_mutex);
8553 read_lock(&tasklist_lock);
8554 err = __rt_schedulable(tg, rt_period, rt_runtime);
8555 if (err)
8556 goto unlock;
8558 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8559 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8560 tg->rt_bandwidth.rt_runtime = rt_runtime;
8562 for_each_possible_cpu(i) {
8563 struct rt_rq *rt_rq = tg->rt_rq[i];
8565 raw_spin_lock(&rt_rq->rt_runtime_lock);
8566 rt_rq->rt_runtime = rt_runtime;
8567 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8569 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8570 unlock:
8571 read_unlock(&tasklist_lock);
8572 mutex_unlock(&rt_constraints_mutex);
8574 return err;
8577 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8579 u64 rt_runtime, rt_period;
8581 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8582 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8583 if (rt_runtime_us < 0)
8584 rt_runtime = RUNTIME_INF;
8586 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8589 long sched_group_rt_runtime(struct task_group *tg)
8591 u64 rt_runtime_us;
8593 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8594 return -1;
8596 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8597 do_div(rt_runtime_us, NSEC_PER_USEC);
8598 return rt_runtime_us;
8601 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8603 u64 rt_runtime, rt_period;
8605 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8606 rt_runtime = tg->rt_bandwidth.rt_runtime;
8608 if (rt_period == 0)
8609 return -EINVAL;
8611 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8614 long sched_group_rt_period(struct task_group *tg)
8616 u64 rt_period_us;
8618 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8619 do_div(rt_period_us, NSEC_PER_USEC);
8620 return rt_period_us;
8623 static int sched_rt_global_constraints(void)
8625 u64 runtime, period;
8626 int ret = 0;
8628 if (sysctl_sched_rt_period <= 0)
8629 return -EINVAL;
8631 runtime = global_rt_runtime();
8632 period = global_rt_period();
8635 * Sanity check on the sysctl variables.
8637 if (runtime > period && runtime != RUNTIME_INF)
8638 return -EINVAL;
8640 mutex_lock(&rt_constraints_mutex);
8641 read_lock(&tasklist_lock);
8642 ret = __rt_schedulable(NULL, 0, 0);
8643 read_unlock(&tasklist_lock);
8644 mutex_unlock(&rt_constraints_mutex);
8646 return ret;
8649 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8651 /* Don't accept realtime tasks when there is no way for them to run */
8652 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8653 return 0;
8655 return 1;
8658 #else /* !CONFIG_RT_GROUP_SCHED */
8659 static int sched_rt_global_constraints(void)
8661 unsigned long flags;
8662 int i;
8664 if (sysctl_sched_rt_period <= 0)
8665 return -EINVAL;
8668 * There's always some RT tasks in the root group
8669 * -- migration, kstopmachine etc..
8671 if (sysctl_sched_rt_runtime == 0)
8672 return -EBUSY;
8674 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8675 for_each_possible_cpu(i) {
8676 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8678 raw_spin_lock(&rt_rq->rt_runtime_lock);
8679 rt_rq->rt_runtime = global_rt_runtime();
8680 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8682 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8684 return 0;
8686 #endif /* CONFIG_RT_GROUP_SCHED */
8688 int sched_rt_handler(struct ctl_table *table, int write,
8689 void __user *buffer, size_t *lenp,
8690 loff_t *ppos)
8692 int ret;
8693 int old_period, old_runtime;
8694 static DEFINE_MUTEX(mutex);
8696 mutex_lock(&mutex);
8697 old_period = sysctl_sched_rt_period;
8698 old_runtime = sysctl_sched_rt_runtime;
8700 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8702 if (!ret && write) {
8703 ret = sched_rt_global_constraints();
8704 if (ret) {
8705 sysctl_sched_rt_period = old_period;
8706 sysctl_sched_rt_runtime = old_runtime;
8707 } else {
8708 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8709 def_rt_bandwidth.rt_period =
8710 ns_to_ktime(global_rt_period());
8713 mutex_unlock(&mutex);
8715 return ret;
8718 #ifdef CONFIG_CGROUP_SCHED
8720 /* return corresponding task_group object of a cgroup */
8721 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8723 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8724 struct task_group, css);
8727 static struct cgroup_subsys_state *
8728 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8730 struct task_group *tg, *parent;
8732 if (!cgrp->parent) {
8733 /* This is early initialization for the top cgroup */
8734 return &root_task_group.css;
8737 parent = cgroup_tg(cgrp->parent);
8738 tg = sched_create_group(parent);
8739 if (IS_ERR(tg))
8740 return ERR_PTR(-ENOMEM);
8742 return &tg->css;
8745 static void
8746 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8748 struct task_group *tg = cgroup_tg(cgrp);
8750 sched_destroy_group(tg);
8753 static int
8754 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8756 #ifdef CONFIG_RT_GROUP_SCHED
8757 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8758 return -EINVAL;
8759 #else
8760 /* We don't support RT-tasks being in separate groups */
8761 if (tsk->sched_class != &fair_sched_class)
8762 return -EINVAL;
8763 #endif
8764 return 0;
8767 static void
8768 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8770 sched_move_task(tsk);
8773 static void
8774 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
8775 struct cgroup *old_cgrp, struct task_struct *task)
8778 * cgroup_exit() is called in the copy_process() failure path.
8779 * Ignore this case since the task hasn't ran yet, this avoids
8780 * trying to poke a half freed task state from generic code.
8782 if (!(task->flags & PF_EXITING))
8783 return;
8785 sched_move_task(task);
8788 #ifdef CONFIG_FAIR_GROUP_SCHED
8789 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8790 u64 shareval)
8792 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
8795 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8797 struct task_group *tg = cgroup_tg(cgrp);
8799 return (u64) scale_load_down(tg->shares);
8801 #endif /* CONFIG_FAIR_GROUP_SCHED */
8803 #ifdef CONFIG_RT_GROUP_SCHED
8804 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8805 s64 val)
8807 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8810 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8812 return sched_group_rt_runtime(cgroup_tg(cgrp));
8815 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8816 u64 rt_period_us)
8818 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8821 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8823 return sched_group_rt_period(cgroup_tg(cgrp));
8825 #endif /* CONFIG_RT_GROUP_SCHED */
8827 static struct cftype cpu_files[] = {
8828 #ifdef CONFIG_FAIR_GROUP_SCHED
8830 .name = "shares",
8831 .read_u64 = cpu_shares_read_u64,
8832 .write_u64 = cpu_shares_write_u64,
8834 #endif
8835 #ifdef CONFIG_RT_GROUP_SCHED
8837 .name = "rt_runtime_us",
8838 .read_s64 = cpu_rt_runtime_read,
8839 .write_s64 = cpu_rt_runtime_write,
8842 .name = "rt_period_us",
8843 .read_u64 = cpu_rt_period_read_uint,
8844 .write_u64 = cpu_rt_period_write_uint,
8846 #endif
8849 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8851 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8854 struct cgroup_subsys cpu_cgroup_subsys = {
8855 .name = "cpu",
8856 .create = cpu_cgroup_create,
8857 .destroy = cpu_cgroup_destroy,
8858 .can_attach_task = cpu_cgroup_can_attach_task,
8859 .attach_task = cpu_cgroup_attach_task,
8860 .exit = cpu_cgroup_exit,
8861 .populate = cpu_cgroup_populate,
8862 .subsys_id = cpu_cgroup_subsys_id,
8863 .early_init = 1,
8866 #endif /* CONFIG_CGROUP_SCHED */
8868 #ifdef CONFIG_CGROUP_CPUACCT
8871 * CPU accounting code for task groups.
8873 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8874 * (balbir@in.ibm.com).
8877 /* track cpu usage of a group of tasks and its child groups */
8878 struct cpuacct {
8879 struct cgroup_subsys_state css;
8880 /* cpuusage holds pointer to a u64-type object on every cpu */
8881 u64 __percpu *cpuusage;
8882 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8883 struct cpuacct *parent;
8886 struct cgroup_subsys cpuacct_subsys;
8888 /* return cpu accounting group corresponding to this container */
8889 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8891 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8892 struct cpuacct, css);
8895 /* return cpu accounting group to which this task belongs */
8896 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8898 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8899 struct cpuacct, css);
8902 /* create a new cpu accounting group */
8903 static struct cgroup_subsys_state *cpuacct_create(
8904 struct cgroup_subsys *ss, struct cgroup *cgrp)
8906 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8907 int i;
8909 if (!ca)
8910 goto out;
8912 ca->cpuusage = alloc_percpu(u64);
8913 if (!ca->cpuusage)
8914 goto out_free_ca;
8916 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8917 if (percpu_counter_init(&ca->cpustat[i], 0))
8918 goto out_free_counters;
8920 if (cgrp->parent)
8921 ca->parent = cgroup_ca(cgrp->parent);
8923 return &ca->css;
8925 out_free_counters:
8926 while (--i >= 0)
8927 percpu_counter_destroy(&ca->cpustat[i]);
8928 free_percpu(ca->cpuusage);
8929 out_free_ca:
8930 kfree(ca);
8931 out:
8932 return ERR_PTR(-ENOMEM);
8935 /* destroy an existing cpu accounting group */
8936 static void
8937 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8939 struct cpuacct *ca = cgroup_ca(cgrp);
8940 int i;
8942 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8943 percpu_counter_destroy(&ca->cpustat[i]);
8944 free_percpu(ca->cpuusage);
8945 kfree(ca);
8948 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8950 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8951 u64 data;
8953 #ifndef CONFIG_64BIT
8955 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8957 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8958 data = *cpuusage;
8959 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8960 #else
8961 data = *cpuusage;
8962 #endif
8964 return data;
8967 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8969 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8971 #ifndef CONFIG_64BIT
8973 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8975 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8976 *cpuusage = val;
8977 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8978 #else
8979 *cpuusage = val;
8980 #endif
8983 /* return total cpu usage (in nanoseconds) of a group */
8984 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8986 struct cpuacct *ca = cgroup_ca(cgrp);
8987 u64 totalcpuusage = 0;
8988 int i;
8990 for_each_present_cpu(i)
8991 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8993 return totalcpuusage;
8996 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8997 u64 reset)
8999 struct cpuacct *ca = cgroup_ca(cgrp);
9000 int err = 0;
9001 int i;
9003 if (reset) {
9004 err = -EINVAL;
9005 goto out;
9008 for_each_present_cpu(i)
9009 cpuacct_cpuusage_write(ca, i, 0);
9011 out:
9012 return err;
9015 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9016 struct seq_file *m)
9018 struct cpuacct *ca = cgroup_ca(cgroup);
9019 u64 percpu;
9020 int i;
9022 for_each_present_cpu(i) {
9023 percpu = cpuacct_cpuusage_read(ca, i);
9024 seq_printf(m, "%llu ", (unsigned long long) percpu);
9026 seq_printf(m, "\n");
9027 return 0;
9030 static const char *cpuacct_stat_desc[] = {
9031 [CPUACCT_STAT_USER] = "user",
9032 [CPUACCT_STAT_SYSTEM] = "system",
9035 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9036 struct cgroup_map_cb *cb)
9038 struct cpuacct *ca = cgroup_ca(cgrp);
9039 int i;
9041 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9042 s64 val = percpu_counter_read(&ca->cpustat[i]);
9043 val = cputime64_to_clock_t(val);
9044 cb->fill(cb, cpuacct_stat_desc[i], val);
9046 return 0;
9049 static struct cftype files[] = {
9051 .name = "usage",
9052 .read_u64 = cpuusage_read,
9053 .write_u64 = cpuusage_write,
9056 .name = "usage_percpu",
9057 .read_seq_string = cpuacct_percpu_seq_read,
9060 .name = "stat",
9061 .read_map = cpuacct_stats_show,
9065 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9067 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9071 * charge this task's execution time to its accounting group.
9073 * called with rq->lock held.
9075 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9077 struct cpuacct *ca;
9078 int cpu;
9080 if (unlikely(!cpuacct_subsys.active))
9081 return;
9083 cpu = task_cpu(tsk);
9085 rcu_read_lock();
9087 ca = task_ca(tsk);
9089 for (; ca; ca = ca->parent) {
9090 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9091 *cpuusage += cputime;
9094 rcu_read_unlock();
9098 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9099 * in cputime_t units. As a result, cpuacct_update_stats calls
9100 * percpu_counter_add with values large enough to always overflow the
9101 * per cpu batch limit causing bad SMP scalability.
9103 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9104 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9105 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9107 #ifdef CONFIG_SMP
9108 #define CPUACCT_BATCH \
9109 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9110 #else
9111 #define CPUACCT_BATCH 0
9112 #endif
9115 * Charge the system/user time to the task's accounting group.
9117 static void cpuacct_update_stats(struct task_struct *tsk,
9118 enum cpuacct_stat_index idx, cputime_t val)
9120 struct cpuacct *ca;
9121 int batch = CPUACCT_BATCH;
9123 if (unlikely(!cpuacct_subsys.active))
9124 return;
9126 rcu_read_lock();
9127 ca = task_ca(tsk);
9129 do {
9130 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9131 ca = ca->parent;
9132 } while (ca);
9133 rcu_read_unlock();
9136 struct cgroup_subsys cpuacct_subsys = {
9137 .name = "cpuacct",
9138 .create = cpuacct_create,
9139 .destroy = cpuacct_destroy,
9140 .populate = cpuacct_populate,
9141 .subsys_id = cpuacct_subsys_id,
9143 #endif /* CONFIG_CGROUP_CPUACCT */