Merge branches 'stable/ia64', 'stable/blkfront-cleanup' and 'stable/cleanup' of git...
[deliverable/linux.git] / kernel / sched.c
1 /*
2 * kernel/sched.c
3 *
4 * Kernel scheduler and related syscalls
5 *
6 * Copyright (C) 1991-2002 Linus Torvalds
7 *
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
27 */
28
29 #include <linux/mm.h>
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/stop_machine.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
75
76 #include <asm/tlb.h>
77 #include <asm/irq_regs.h>
78 #include <asm/mutex.h>
79
80 #include "sched_cpupri.h"
81 #include "workqueue_sched.h"
82 #include "sched_autogroup.h"
83
84 #define CREATE_TRACE_POINTS
85 #include <trace/events/sched.h>
86
87 /*
88 * Convert user-nice values [ -20 ... 0 ... 19 ]
89 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
90 * and back.
91 */
92 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
93 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
94 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
95
96 /*
97 * 'User priority' is the nice value converted to something we
98 * can work with better when scaling various scheduler parameters,
99 * it's a [ 0 ... 39 ] range.
100 */
101 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
102 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
103 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
104
105 /*
106 * Helpers for converting nanosecond timing to jiffy resolution
107 */
108 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
109
110 #define NICE_0_LOAD SCHED_LOAD_SCALE
111 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
112
113 /*
114 * These are the 'tuning knobs' of the scheduler:
115 *
116 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
117 * Timeslices get refilled after they expire.
118 */
119 #define DEF_TIMESLICE (100 * HZ / 1000)
120
121 /*
122 * single value that denotes runtime == period, ie unlimited time.
123 */
124 #define RUNTIME_INF ((u64)~0ULL)
125
126 static inline int rt_policy(int policy)
127 {
128 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
129 return 1;
130 return 0;
131 }
132
133 static inline int task_has_rt_policy(struct task_struct *p)
134 {
135 return rt_policy(p->policy);
136 }
137
138 /*
139 * This is the priority-queue data structure of the RT scheduling class:
140 */
141 struct rt_prio_array {
142 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
143 struct list_head queue[MAX_RT_PRIO];
144 };
145
146 struct rt_bandwidth {
147 /* nests inside the rq lock: */
148 raw_spinlock_t rt_runtime_lock;
149 ktime_t rt_period;
150 u64 rt_runtime;
151 struct hrtimer rt_period_timer;
152 };
153
154 static struct rt_bandwidth def_rt_bandwidth;
155
156 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
157
158 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
159 {
160 struct rt_bandwidth *rt_b =
161 container_of(timer, struct rt_bandwidth, rt_period_timer);
162 ktime_t now;
163 int overrun;
164 int idle = 0;
165
166 for (;;) {
167 now = hrtimer_cb_get_time(timer);
168 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169
170 if (!overrun)
171 break;
172
173 idle = do_sched_rt_period_timer(rt_b, overrun);
174 }
175
176 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
177 }
178
179 static
180 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
181 {
182 rt_b->rt_period = ns_to_ktime(period);
183 rt_b->rt_runtime = runtime;
184
185 raw_spin_lock_init(&rt_b->rt_runtime_lock);
186
187 hrtimer_init(&rt_b->rt_period_timer,
188 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
189 rt_b->rt_period_timer.function = sched_rt_period_timer;
190 }
191
192 static inline int rt_bandwidth_enabled(void)
193 {
194 return sysctl_sched_rt_runtime >= 0;
195 }
196
197 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
198 {
199 ktime_t now;
200
201 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
202 return;
203
204 if (hrtimer_active(&rt_b->rt_period_timer))
205 return;
206
207 raw_spin_lock(&rt_b->rt_runtime_lock);
208 for (;;) {
209 unsigned long delta;
210 ktime_t soft, hard;
211
212 if (hrtimer_active(&rt_b->rt_period_timer))
213 break;
214
215 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
216 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
217
218 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
219 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
220 delta = ktime_to_ns(ktime_sub(hard, soft));
221 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
222 HRTIMER_MODE_ABS_PINNED, 0);
223 }
224 raw_spin_unlock(&rt_b->rt_runtime_lock);
225 }
226
227 #ifdef CONFIG_RT_GROUP_SCHED
228 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
229 {
230 hrtimer_cancel(&rt_b->rt_period_timer);
231 }
232 #endif
233
234 /*
235 * sched_domains_mutex serializes calls to arch_init_sched_domains,
236 * detach_destroy_domains and partition_sched_domains.
237 */
238 static DEFINE_MUTEX(sched_domains_mutex);
239
240 #ifdef CONFIG_CGROUP_SCHED
241
242 #include <linux/cgroup.h>
243
244 struct cfs_rq;
245
246 static LIST_HEAD(task_groups);
247
248 /* task group related information */
249 struct task_group {
250 struct cgroup_subsys_state css;
251
252 #ifdef CONFIG_FAIR_GROUP_SCHED
253 /* schedulable entities of this group on each cpu */
254 struct sched_entity **se;
255 /* runqueue "owned" by this group on each cpu */
256 struct cfs_rq **cfs_rq;
257 unsigned long shares;
258
259 atomic_t load_weight;
260 #endif
261
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
265
266 struct rt_bandwidth rt_bandwidth;
267 #endif
268
269 struct rcu_head rcu;
270 struct list_head list;
271
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
275
276 #ifdef CONFIG_SCHED_AUTOGROUP
277 struct autogroup *autogroup;
278 #endif
279 };
280
281 /* task_group_lock serializes the addition/removal of task groups */
282 static DEFINE_SPINLOCK(task_group_lock);
283
284 #ifdef CONFIG_FAIR_GROUP_SCHED
285
286 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
287
288 /*
289 * A weight of 0 or 1 can cause arithmetics problems.
290 * A weight of a cfs_rq is the sum of weights of which entities
291 * are queued on this cfs_rq, so a weight of a entity should not be
292 * too large, so as the shares value of a task group.
293 * (The default weight is 1024 - so there's no practical
294 * limitation from this.)
295 */
296 #define MIN_SHARES 2
297 #define MAX_SHARES (1UL << 18)
298
299 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
300 #endif
301
302 /* Default task group.
303 * Every task in system belong to this group at bootup.
304 */
305 struct task_group root_task_group;
306
307 #endif /* CONFIG_CGROUP_SCHED */
308
309 /* CFS-related fields in a runqueue */
310 struct cfs_rq {
311 struct load_weight load;
312 unsigned long nr_running;
313
314 u64 exec_clock;
315 u64 min_vruntime;
316
317 struct rb_root tasks_timeline;
318 struct rb_node *rb_leftmost;
319
320 struct list_head tasks;
321 struct list_head *balance_iterator;
322
323 /*
324 * 'curr' points to currently running entity on this cfs_rq.
325 * It is set to NULL otherwise (i.e when none are currently running).
326 */
327 struct sched_entity *curr, *next, *last;
328
329 unsigned int nr_spread_over;
330
331 #ifdef CONFIG_FAIR_GROUP_SCHED
332 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
333
334 /*
335 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
336 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
337 * (like users, containers etc.)
338 *
339 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
340 * list is used during load balance.
341 */
342 int on_list;
343 struct list_head leaf_cfs_rq_list;
344 struct task_group *tg; /* group that "owns" this runqueue */
345
346 #ifdef CONFIG_SMP
347 /*
348 * the part of load.weight contributed by tasks
349 */
350 unsigned long task_weight;
351
352 /*
353 * h_load = weight * f(tg)
354 *
355 * Where f(tg) is the recursive weight fraction assigned to
356 * this group.
357 */
358 unsigned long h_load;
359
360 /*
361 * Maintaining per-cpu shares distribution for group scheduling
362 *
363 * load_stamp is the last time we updated the load average
364 * load_last is the last time we updated the load average and saw load
365 * load_unacc_exec_time is currently unaccounted execution time
366 */
367 u64 load_avg;
368 u64 load_period;
369 u64 load_stamp, load_last, load_unacc_exec_time;
370
371 unsigned long load_contribution;
372 #endif
373 #endif
374 };
375
376 /* Real-Time classes' related field in a runqueue: */
377 struct rt_rq {
378 struct rt_prio_array active;
379 unsigned long rt_nr_running;
380 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
381 struct {
382 int curr; /* highest queued rt task prio */
383 #ifdef CONFIG_SMP
384 int next; /* next highest */
385 #endif
386 } highest_prio;
387 #endif
388 #ifdef CONFIG_SMP
389 unsigned long rt_nr_migratory;
390 unsigned long rt_nr_total;
391 int overloaded;
392 struct plist_head pushable_tasks;
393 #endif
394 int rt_throttled;
395 u64 rt_time;
396 u64 rt_runtime;
397 /* Nests inside the rq lock: */
398 raw_spinlock_t rt_runtime_lock;
399
400 #ifdef CONFIG_RT_GROUP_SCHED
401 unsigned long rt_nr_boosted;
402
403 struct rq *rq;
404 struct list_head leaf_rt_rq_list;
405 struct task_group *tg;
406 #endif
407 };
408
409 #ifdef CONFIG_SMP
410
411 /*
412 * We add the notion of a root-domain which will be used to define per-domain
413 * variables. Each exclusive cpuset essentially defines an island domain by
414 * fully partitioning the member cpus from any other cpuset. Whenever a new
415 * exclusive cpuset is created, we also create and attach a new root-domain
416 * object.
417 *
418 */
419 struct root_domain {
420 atomic_t refcount;
421 cpumask_var_t span;
422 cpumask_var_t online;
423
424 /*
425 * The "RT overload" flag: it gets set if a CPU has more than
426 * one runnable RT task.
427 */
428 cpumask_var_t rto_mask;
429 atomic_t rto_count;
430 struct cpupri cpupri;
431 };
432
433 /*
434 * By default the system creates a single root-domain with all cpus as
435 * members (mimicking the global state we have today).
436 */
437 static struct root_domain def_root_domain;
438
439 #endif /* CONFIG_SMP */
440
441 /*
442 * This is the main, per-CPU runqueue data structure.
443 *
444 * Locking rule: those places that want to lock multiple runqueues
445 * (such as the load balancing or the thread migration code), lock
446 * acquire operations must be ordered by ascending &runqueue.
447 */
448 struct rq {
449 /* runqueue lock: */
450 raw_spinlock_t lock;
451
452 /*
453 * nr_running and cpu_load should be in the same cacheline because
454 * remote CPUs use both these fields when doing load calculation.
455 */
456 unsigned long nr_running;
457 #define CPU_LOAD_IDX_MAX 5
458 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
459 unsigned long last_load_update_tick;
460 #ifdef CONFIG_NO_HZ
461 u64 nohz_stamp;
462 unsigned char nohz_balance_kick;
463 #endif
464 unsigned int skip_clock_update;
465
466 /* capture load from *all* tasks on this cpu: */
467 struct load_weight load;
468 unsigned long nr_load_updates;
469 u64 nr_switches;
470
471 struct cfs_rq cfs;
472 struct rt_rq rt;
473
474 #ifdef CONFIG_FAIR_GROUP_SCHED
475 /* list of leaf cfs_rq on this cpu: */
476 struct list_head leaf_cfs_rq_list;
477 #endif
478 #ifdef CONFIG_RT_GROUP_SCHED
479 struct list_head leaf_rt_rq_list;
480 #endif
481
482 /*
483 * This is part of a global counter where only the total sum
484 * over all CPUs matters. A task can increase this counter on
485 * one CPU and if it got migrated afterwards it may decrease
486 * it on another CPU. Always updated under the runqueue lock:
487 */
488 unsigned long nr_uninterruptible;
489
490 struct task_struct *curr, *idle, *stop;
491 unsigned long next_balance;
492 struct mm_struct *prev_mm;
493
494 u64 clock;
495 u64 clock_task;
496
497 atomic_t nr_iowait;
498
499 #ifdef CONFIG_SMP
500 struct root_domain *rd;
501 struct sched_domain *sd;
502
503 unsigned long cpu_power;
504
505 unsigned char idle_at_tick;
506 /* For active balancing */
507 int post_schedule;
508 int active_balance;
509 int push_cpu;
510 struct cpu_stop_work active_balance_work;
511 /* cpu of this runqueue: */
512 int cpu;
513 int online;
514
515 unsigned long avg_load_per_task;
516
517 u64 rt_avg;
518 u64 age_stamp;
519 u64 idle_stamp;
520 u64 avg_idle;
521 #endif
522
523 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
524 u64 prev_irq_time;
525 #endif
526
527 /* calc_load related fields */
528 unsigned long calc_load_update;
529 long calc_load_active;
530
531 #ifdef CONFIG_SCHED_HRTICK
532 #ifdef CONFIG_SMP
533 int hrtick_csd_pending;
534 struct call_single_data hrtick_csd;
535 #endif
536 struct hrtimer hrtick_timer;
537 #endif
538
539 #ifdef CONFIG_SCHEDSTATS
540 /* latency stats */
541 struct sched_info rq_sched_info;
542 unsigned long long rq_cpu_time;
543 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
544
545 /* sys_sched_yield() stats */
546 unsigned int yld_count;
547
548 /* schedule() stats */
549 unsigned int sched_switch;
550 unsigned int sched_count;
551 unsigned int sched_goidle;
552
553 /* try_to_wake_up() stats */
554 unsigned int ttwu_count;
555 unsigned int ttwu_local;
556 #endif
557 };
558
559 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
560
561
562 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
563
564 static inline int cpu_of(struct rq *rq)
565 {
566 #ifdef CONFIG_SMP
567 return rq->cpu;
568 #else
569 return 0;
570 #endif
571 }
572
573 #define rcu_dereference_check_sched_domain(p) \
574 rcu_dereference_check((p), \
575 rcu_read_lock_sched_held() || \
576 lockdep_is_held(&sched_domains_mutex))
577
578 /*
579 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
580 * See detach_destroy_domains: synchronize_sched for details.
581 *
582 * The domain tree of any CPU may only be accessed from within
583 * preempt-disabled sections.
584 */
585 #define for_each_domain(cpu, __sd) \
586 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
587
588 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
589 #define this_rq() (&__get_cpu_var(runqueues))
590 #define task_rq(p) cpu_rq(task_cpu(p))
591 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
592 #define raw_rq() (&__raw_get_cpu_var(runqueues))
593
594 #ifdef CONFIG_CGROUP_SCHED
595
596 /*
597 * Return the group to which this tasks belongs.
598 *
599 * We use task_subsys_state_check() and extend the RCU verification
600 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
601 * holds that lock for each task it moves into the cgroup. Therefore
602 * by holding that lock, we pin the task to the current cgroup.
603 */
604 static inline struct task_group *task_group(struct task_struct *p)
605 {
606 struct task_group *tg;
607 struct cgroup_subsys_state *css;
608
609 if (p->flags & PF_EXITING)
610 return &root_task_group;
611
612 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
613 lockdep_is_held(&task_rq(p)->lock));
614 tg = container_of(css, struct task_group, css);
615
616 return autogroup_task_group(p, tg);
617 }
618
619 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
620 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
621 {
622 #ifdef CONFIG_FAIR_GROUP_SCHED
623 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
624 p->se.parent = task_group(p)->se[cpu];
625 #endif
626
627 #ifdef CONFIG_RT_GROUP_SCHED
628 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
629 p->rt.parent = task_group(p)->rt_se[cpu];
630 #endif
631 }
632
633 #else /* CONFIG_CGROUP_SCHED */
634
635 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
636 static inline struct task_group *task_group(struct task_struct *p)
637 {
638 return NULL;
639 }
640
641 #endif /* CONFIG_CGROUP_SCHED */
642
643 static void update_rq_clock_task(struct rq *rq, s64 delta);
644
645 static void update_rq_clock(struct rq *rq)
646 {
647 s64 delta;
648
649 if (rq->skip_clock_update)
650 return;
651
652 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
653 rq->clock += delta;
654 update_rq_clock_task(rq, delta);
655 }
656
657 /*
658 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
659 */
660 #ifdef CONFIG_SCHED_DEBUG
661 # define const_debug __read_mostly
662 #else
663 # define const_debug static const
664 #endif
665
666 /**
667 * runqueue_is_locked
668 * @cpu: the processor in question.
669 *
670 * Returns true if the current cpu runqueue is locked.
671 * This interface allows printk to be called with the runqueue lock
672 * held and know whether or not it is OK to wake up the klogd.
673 */
674 int runqueue_is_locked(int cpu)
675 {
676 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
677 }
678
679 /*
680 * Debugging: various feature bits
681 */
682
683 #define SCHED_FEAT(name, enabled) \
684 __SCHED_FEAT_##name ,
685
686 enum {
687 #include "sched_features.h"
688 };
689
690 #undef SCHED_FEAT
691
692 #define SCHED_FEAT(name, enabled) \
693 (1UL << __SCHED_FEAT_##name) * enabled |
694
695 const_debug unsigned int sysctl_sched_features =
696 #include "sched_features.h"
697 0;
698
699 #undef SCHED_FEAT
700
701 #ifdef CONFIG_SCHED_DEBUG
702 #define SCHED_FEAT(name, enabled) \
703 #name ,
704
705 static __read_mostly char *sched_feat_names[] = {
706 #include "sched_features.h"
707 NULL
708 };
709
710 #undef SCHED_FEAT
711
712 static int sched_feat_show(struct seq_file *m, void *v)
713 {
714 int i;
715
716 for (i = 0; sched_feat_names[i]; i++) {
717 if (!(sysctl_sched_features & (1UL << i)))
718 seq_puts(m, "NO_");
719 seq_printf(m, "%s ", sched_feat_names[i]);
720 }
721 seq_puts(m, "\n");
722
723 return 0;
724 }
725
726 static ssize_t
727 sched_feat_write(struct file *filp, const char __user *ubuf,
728 size_t cnt, loff_t *ppos)
729 {
730 char buf[64];
731 char *cmp;
732 int neg = 0;
733 int i;
734
735 if (cnt > 63)
736 cnt = 63;
737
738 if (copy_from_user(&buf, ubuf, cnt))
739 return -EFAULT;
740
741 buf[cnt] = 0;
742 cmp = strstrip(buf);
743
744 if (strncmp(cmp, "NO_", 3) == 0) {
745 neg = 1;
746 cmp += 3;
747 }
748
749 for (i = 0; sched_feat_names[i]; i++) {
750 if (strcmp(cmp, sched_feat_names[i]) == 0) {
751 if (neg)
752 sysctl_sched_features &= ~(1UL << i);
753 else
754 sysctl_sched_features |= (1UL << i);
755 break;
756 }
757 }
758
759 if (!sched_feat_names[i])
760 return -EINVAL;
761
762 *ppos += cnt;
763
764 return cnt;
765 }
766
767 static int sched_feat_open(struct inode *inode, struct file *filp)
768 {
769 return single_open(filp, sched_feat_show, NULL);
770 }
771
772 static const struct file_operations sched_feat_fops = {
773 .open = sched_feat_open,
774 .write = sched_feat_write,
775 .read = seq_read,
776 .llseek = seq_lseek,
777 .release = single_release,
778 };
779
780 static __init int sched_init_debug(void)
781 {
782 debugfs_create_file("sched_features", 0644, NULL, NULL,
783 &sched_feat_fops);
784
785 return 0;
786 }
787 late_initcall(sched_init_debug);
788
789 #endif
790
791 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
792
793 /*
794 * Number of tasks to iterate in a single balance run.
795 * Limited because this is done with IRQs disabled.
796 */
797 const_debug unsigned int sysctl_sched_nr_migrate = 32;
798
799 /*
800 * period over which we average the RT time consumption, measured
801 * in ms.
802 *
803 * default: 1s
804 */
805 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
806
807 /*
808 * period over which we measure -rt task cpu usage in us.
809 * default: 1s
810 */
811 unsigned int sysctl_sched_rt_period = 1000000;
812
813 static __read_mostly int scheduler_running;
814
815 /*
816 * part of the period that we allow rt tasks to run in us.
817 * default: 0.95s
818 */
819 int sysctl_sched_rt_runtime = 950000;
820
821 static inline u64 global_rt_period(void)
822 {
823 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
824 }
825
826 static inline u64 global_rt_runtime(void)
827 {
828 if (sysctl_sched_rt_runtime < 0)
829 return RUNTIME_INF;
830
831 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
832 }
833
834 #ifndef prepare_arch_switch
835 # define prepare_arch_switch(next) do { } while (0)
836 #endif
837 #ifndef finish_arch_switch
838 # define finish_arch_switch(prev) do { } while (0)
839 #endif
840
841 static inline int task_current(struct rq *rq, struct task_struct *p)
842 {
843 return rq->curr == p;
844 }
845
846 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
847 static inline int task_running(struct rq *rq, struct task_struct *p)
848 {
849 return task_current(rq, p);
850 }
851
852 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
853 {
854 }
855
856 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
857 {
858 #ifdef CONFIG_DEBUG_SPINLOCK
859 /* this is a valid case when another task releases the spinlock */
860 rq->lock.owner = current;
861 #endif
862 /*
863 * If we are tracking spinlock dependencies then we have to
864 * fix up the runqueue lock - which gets 'carried over' from
865 * prev into current:
866 */
867 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
868
869 raw_spin_unlock_irq(&rq->lock);
870 }
871
872 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
873 static inline int task_running(struct rq *rq, struct task_struct *p)
874 {
875 #ifdef CONFIG_SMP
876 return p->oncpu;
877 #else
878 return task_current(rq, p);
879 #endif
880 }
881
882 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
883 {
884 #ifdef CONFIG_SMP
885 /*
886 * We can optimise this out completely for !SMP, because the
887 * SMP rebalancing from interrupt is the only thing that cares
888 * here.
889 */
890 next->oncpu = 1;
891 #endif
892 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
893 raw_spin_unlock_irq(&rq->lock);
894 #else
895 raw_spin_unlock(&rq->lock);
896 #endif
897 }
898
899 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
900 {
901 #ifdef CONFIG_SMP
902 /*
903 * After ->oncpu is cleared, the task can be moved to a different CPU.
904 * We must ensure this doesn't happen until the switch is completely
905 * finished.
906 */
907 smp_wmb();
908 prev->oncpu = 0;
909 #endif
910 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
911 local_irq_enable();
912 #endif
913 }
914 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
915
916 /*
917 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
918 * against ttwu().
919 */
920 static inline int task_is_waking(struct task_struct *p)
921 {
922 return unlikely(p->state == TASK_WAKING);
923 }
924
925 /*
926 * __task_rq_lock - lock the runqueue a given task resides on.
927 * Must be called interrupts disabled.
928 */
929 static inline struct rq *__task_rq_lock(struct task_struct *p)
930 __acquires(rq->lock)
931 {
932 struct rq *rq;
933
934 for (;;) {
935 rq = task_rq(p);
936 raw_spin_lock(&rq->lock);
937 if (likely(rq == task_rq(p)))
938 return rq;
939 raw_spin_unlock(&rq->lock);
940 }
941 }
942
943 /*
944 * task_rq_lock - lock the runqueue a given task resides on and disable
945 * interrupts. Note the ordering: we can safely lookup the task_rq without
946 * explicitly disabling preemption.
947 */
948 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
949 __acquires(rq->lock)
950 {
951 struct rq *rq;
952
953 for (;;) {
954 local_irq_save(*flags);
955 rq = task_rq(p);
956 raw_spin_lock(&rq->lock);
957 if (likely(rq == task_rq(p)))
958 return rq;
959 raw_spin_unlock_irqrestore(&rq->lock, *flags);
960 }
961 }
962
963 static void __task_rq_unlock(struct rq *rq)
964 __releases(rq->lock)
965 {
966 raw_spin_unlock(&rq->lock);
967 }
968
969 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
970 __releases(rq->lock)
971 {
972 raw_spin_unlock_irqrestore(&rq->lock, *flags);
973 }
974
975 /*
976 * this_rq_lock - lock this runqueue and disable interrupts.
977 */
978 static struct rq *this_rq_lock(void)
979 __acquires(rq->lock)
980 {
981 struct rq *rq;
982
983 local_irq_disable();
984 rq = this_rq();
985 raw_spin_lock(&rq->lock);
986
987 return rq;
988 }
989
990 #ifdef CONFIG_SCHED_HRTICK
991 /*
992 * Use HR-timers to deliver accurate preemption points.
993 *
994 * Its all a bit involved since we cannot program an hrt while holding the
995 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
996 * reschedule event.
997 *
998 * When we get rescheduled we reprogram the hrtick_timer outside of the
999 * rq->lock.
1000 */
1001
1002 /*
1003 * Use hrtick when:
1004 * - enabled by features
1005 * - hrtimer is actually high res
1006 */
1007 static inline int hrtick_enabled(struct rq *rq)
1008 {
1009 if (!sched_feat(HRTICK))
1010 return 0;
1011 if (!cpu_active(cpu_of(rq)))
1012 return 0;
1013 return hrtimer_is_hres_active(&rq->hrtick_timer);
1014 }
1015
1016 static void hrtick_clear(struct rq *rq)
1017 {
1018 if (hrtimer_active(&rq->hrtick_timer))
1019 hrtimer_cancel(&rq->hrtick_timer);
1020 }
1021
1022 /*
1023 * High-resolution timer tick.
1024 * Runs from hardirq context with interrupts disabled.
1025 */
1026 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1027 {
1028 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1029
1030 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1031
1032 raw_spin_lock(&rq->lock);
1033 update_rq_clock(rq);
1034 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1035 raw_spin_unlock(&rq->lock);
1036
1037 return HRTIMER_NORESTART;
1038 }
1039
1040 #ifdef CONFIG_SMP
1041 /*
1042 * called from hardirq (IPI) context
1043 */
1044 static void __hrtick_start(void *arg)
1045 {
1046 struct rq *rq = arg;
1047
1048 raw_spin_lock(&rq->lock);
1049 hrtimer_restart(&rq->hrtick_timer);
1050 rq->hrtick_csd_pending = 0;
1051 raw_spin_unlock(&rq->lock);
1052 }
1053
1054 /*
1055 * Called to set the hrtick timer state.
1056 *
1057 * called with rq->lock held and irqs disabled
1058 */
1059 static void hrtick_start(struct rq *rq, u64 delay)
1060 {
1061 struct hrtimer *timer = &rq->hrtick_timer;
1062 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1063
1064 hrtimer_set_expires(timer, time);
1065
1066 if (rq == this_rq()) {
1067 hrtimer_restart(timer);
1068 } else if (!rq->hrtick_csd_pending) {
1069 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1070 rq->hrtick_csd_pending = 1;
1071 }
1072 }
1073
1074 static int
1075 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1076 {
1077 int cpu = (int)(long)hcpu;
1078
1079 switch (action) {
1080 case CPU_UP_CANCELED:
1081 case CPU_UP_CANCELED_FROZEN:
1082 case CPU_DOWN_PREPARE:
1083 case CPU_DOWN_PREPARE_FROZEN:
1084 case CPU_DEAD:
1085 case CPU_DEAD_FROZEN:
1086 hrtick_clear(cpu_rq(cpu));
1087 return NOTIFY_OK;
1088 }
1089
1090 return NOTIFY_DONE;
1091 }
1092
1093 static __init void init_hrtick(void)
1094 {
1095 hotcpu_notifier(hotplug_hrtick, 0);
1096 }
1097 #else
1098 /*
1099 * Called to set the hrtick timer state.
1100 *
1101 * called with rq->lock held and irqs disabled
1102 */
1103 static void hrtick_start(struct rq *rq, u64 delay)
1104 {
1105 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1106 HRTIMER_MODE_REL_PINNED, 0);
1107 }
1108
1109 static inline void init_hrtick(void)
1110 {
1111 }
1112 #endif /* CONFIG_SMP */
1113
1114 static void init_rq_hrtick(struct rq *rq)
1115 {
1116 #ifdef CONFIG_SMP
1117 rq->hrtick_csd_pending = 0;
1118
1119 rq->hrtick_csd.flags = 0;
1120 rq->hrtick_csd.func = __hrtick_start;
1121 rq->hrtick_csd.info = rq;
1122 #endif
1123
1124 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1125 rq->hrtick_timer.function = hrtick;
1126 }
1127 #else /* CONFIG_SCHED_HRTICK */
1128 static inline void hrtick_clear(struct rq *rq)
1129 {
1130 }
1131
1132 static inline void init_rq_hrtick(struct rq *rq)
1133 {
1134 }
1135
1136 static inline void init_hrtick(void)
1137 {
1138 }
1139 #endif /* CONFIG_SCHED_HRTICK */
1140
1141 /*
1142 * resched_task - mark a task 'to be rescheduled now'.
1143 *
1144 * On UP this means the setting of the need_resched flag, on SMP it
1145 * might also involve a cross-CPU call to trigger the scheduler on
1146 * the target CPU.
1147 */
1148 #ifdef CONFIG_SMP
1149
1150 #ifndef tsk_is_polling
1151 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1152 #endif
1153
1154 static void resched_task(struct task_struct *p)
1155 {
1156 int cpu;
1157
1158 assert_raw_spin_locked(&task_rq(p)->lock);
1159
1160 if (test_tsk_need_resched(p))
1161 return;
1162
1163 set_tsk_need_resched(p);
1164
1165 cpu = task_cpu(p);
1166 if (cpu == smp_processor_id())
1167 return;
1168
1169 /* NEED_RESCHED must be visible before we test polling */
1170 smp_mb();
1171 if (!tsk_is_polling(p))
1172 smp_send_reschedule(cpu);
1173 }
1174
1175 static void resched_cpu(int cpu)
1176 {
1177 struct rq *rq = cpu_rq(cpu);
1178 unsigned long flags;
1179
1180 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1181 return;
1182 resched_task(cpu_curr(cpu));
1183 raw_spin_unlock_irqrestore(&rq->lock, flags);
1184 }
1185
1186 #ifdef CONFIG_NO_HZ
1187 /*
1188 * In the semi idle case, use the nearest busy cpu for migrating timers
1189 * from an idle cpu. This is good for power-savings.
1190 *
1191 * We don't do similar optimization for completely idle system, as
1192 * selecting an idle cpu will add more delays to the timers than intended
1193 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1194 */
1195 int get_nohz_timer_target(void)
1196 {
1197 int cpu = smp_processor_id();
1198 int i;
1199 struct sched_domain *sd;
1200
1201 for_each_domain(cpu, sd) {
1202 for_each_cpu(i, sched_domain_span(sd))
1203 if (!idle_cpu(i))
1204 return i;
1205 }
1206 return cpu;
1207 }
1208 /*
1209 * When add_timer_on() enqueues a timer into the timer wheel of an
1210 * idle CPU then this timer might expire before the next timer event
1211 * which is scheduled to wake up that CPU. In case of a completely
1212 * idle system the next event might even be infinite time into the
1213 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1214 * leaves the inner idle loop so the newly added timer is taken into
1215 * account when the CPU goes back to idle and evaluates the timer
1216 * wheel for the next timer event.
1217 */
1218 void wake_up_idle_cpu(int cpu)
1219 {
1220 struct rq *rq = cpu_rq(cpu);
1221
1222 if (cpu == smp_processor_id())
1223 return;
1224
1225 /*
1226 * This is safe, as this function is called with the timer
1227 * wheel base lock of (cpu) held. When the CPU is on the way
1228 * to idle and has not yet set rq->curr to idle then it will
1229 * be serialized on the timer wheel base lock and take the new
1230 * timer into account automatically.
1231 */
1232 if (rq->curr != rq->idle)
1233 return;
1234
1235 /*
1236 * We can set TIF_RESCHED on the idle task of the other CPU
1237 * lockless. The worst case is that the other CPU runs the
1238 * idle task through an additional NOOP schedule()
1239 */
1240 set_tsk_need_resched(rq->idle);
1241
1242 /* NEED_RESCHED must be visible before we test polling */
1243 smp_mb();
1244 if (!tsk_is_polling(rq->idle))
1245 smp_send_reschedule(cpu);
1246 }
1247
1248 #endif /* CONFIG_NO_HZ */
1249
1250 static u64 sched_avg_period(void)
1251 {
1252 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1253 }
1254
1255 static void sched_avg_update(struct rq *rq)
1256 {
1257 s64 period = sched_avg_period();
1258
1259 while ((s64)(rq->clock - rq->age_stamp) > period) {
1260 /*
1261 * Inline assembly required to prevent the compiler
1262 * optimising this loop into a divmod call.
1263 * See __iter_div_u64_rem() for another example of this.
1264 */
1265 asm("" : "+rm" (rq->age_stamp));
1266 rq->age_stamp += period;
1267 rq->rt_avg /= 2;
1268 }
1269 }
1270
1271 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1272 {
1273 rq->rt_avg += rt_delta;
1274 sched_avg_update(rq);
1275 }
1276
1277 #else /* !CONFIG_SMP */
1278 static void resched_task(struct task_struct *p)
1279 {
1280 assert_raw_spin_locked(&task_rq(p)->lock);
1281 set_tsk_need_resched(p);
1282 }
1283
1284 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1285 {
1286 }
1287
1288 static void sched_avg_update(struct rq *rq)
1289 {
1290 }
1291 #endif /* CONFIG_SMP */
1292
1293 #if BITS_PER_LONG == 32
1294 # define WMULT_CONST (~0UL)
1295 #else
1296 # define WMULT_CONST (1UL << 32)
1297 #endif
1298
1299 #define WMULT_SHIFT 32
1300
1301 /*
1302 * Shift right and round:
1303 */
1304 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1305
1306 /*
1307 * delta *= weight / lw
1308 */
1309 static unsigned long
1310 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1311 struct load_weight *lw)
1312 {
1313 u64 tmp;
1314
1315 if (!lw->inv_weight) {
1316 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1317 lw->inv_weight = 1;
1318 else
1319 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1320 / (lw->weight+1);
1321 }
1322
1323 tmp = (u64)delta_exec * weight;
1324 /*
1325 * Check whether we'd overflow the 64-bit multiplication:
1326 */
1327 if (unlikely(tmp > WMULT_CONST))
1328 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1329 WMULT_SHIFT/2);
1330 else
1331 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1332
1333 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1334 }
1335
1336 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1337 {
1338 lw->weight += inc;
1339 lw->inv_weight = 0;
1340 }
1341
1342 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1343 {
1344 lw->weight -= dec;
1345 lw->inv_weight = 0;
1346 }
1347
1348 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1349 {
1350 lw->weight = w;
1351 lw->inv_weight = 0;
1352 }
1353
1354 /*
1355 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1356 * of tasks with abnormal "nice" values across CPUs the contribution that
1357 * each task makes to its run queue's load is weighted according to its
1358 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1359 * scaled version of the new time slice allocation that they receive on time
1360 * slice expiry etc.
1361 */
1362
1363 #define WEIGHT_IDLEPRIO 3
1364 #define WMULT_IDLEPRIO 1431655765
1365
1366 /*
1367 * Nice levels are multiplicative, with a gentle 10% change for every
1368 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1369 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1370 * that remained on nice 0.
1371 *
1372 * The "10% effect" is relative and cumulative: from _any_ nice level,
1373 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1374 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1375 * If a task goes up by ~10% and another task goes down by ~10% then
1376 * the relative distance between them is ~25%.)
1377 */
1378 static const int prio_to_weight[40] = {
1379 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1380 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1381 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1382 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1383 /* 0 */ 1024, 820, 655, 526, 423,
1384 /* 5 */ 335, 272, 215, 172, 137,
1385 /* 10 */ 110, 87, 70, 56, 45,
1386 /* 15 */ 36, 29, 23, 18, 15,
1387 };
1388
1389 /*
1390 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1391 *
1392 * In cases where the weight does not change often, we can use the
1393 * precalculated inverse to speed up arithmetics by turning divisions
1394 * into multiplications:
1395 */
1396 static const u32 prio_to_wmult[40] = {
1397 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1398 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1399 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1400 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1401 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1402 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1403 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1404 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1405 };
1406
1407 /* Time spent by the tasks of the cpu accounting group executing in ... */
1408 enum cpuacct_stat_index {
1409 CPUACCT_STAT_USER, /* ... user mode */
1410 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1411
1412 CPUACCT_STAT_NSTATS,
1413 };
1414
1415 #ifdef CONFIG_CGROUP_CPUACCT
1416 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1417 static void cpuacct_update_stats(struct task_struct *tsk,
1418 enum cpuacct_stat_index idx, cputime_t val);
1419 #else
1420 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1421 static inline void cpuacct_update_stats(struct task_struct *tsk,
1422 enum cpuacct_stat_index idx, cputime_t val) {}
1423 #endif
1424
1425 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1426 {
1427 update_load_add(&rq->load, load);
1428 }
1429
1430 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1431 {
1432 update_load_sub(&rq->load, load);
1433 }
1434
1435 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1436 typedef int (*tg_visitor)(struct task_group *, void *);
1437
1438 /*
1439 * Iterate the full tree, calling @down when first entering a node and @up when
1440 * leaving it for the final time.
1441 */
1442 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1443 {
1444 struct task_group *parent, *child;
1445 int ret;
1446
1447 rcu_read_lock();
1448 parent = &root_task_group;
1449 down:
1450 ret = (*down)(parent, data);
1451 if (ret)
1452 goto out_unlock;
1453 list_for_each_entry_rcu(child, &parent->children, siblings) {
1454 parent = child;
1455 goto down;
1456
1457 up:
1458 continue;
1459 }
1460 ret = (*up)(parent, data);
1461 if (ret)
1462 goto out_unlock;
1463
1464 child = parent;
1465 parent = parent->parent;
1466 if (parent)
1467 goto up;
1468 out_unlock:
1469 rcu_read_unlock();
1470
1471 return ret;
1472 }
1473
1474 static int tg_nop(struct task_group *tg, void *data)
1475 {
1476 return 0;
1477 }
1478 #endif
1479
1480 #ifdef CONFIG_SMP
1481 /* Used instead of source_load when we know the type == 0 */
1482 static unsigned long weighted_cpuload(const int cpu)
1483 {
1484 return cpu_rq(cpu)->load.weight;
1485 }
1486
1487 /*
1488 * Return a low guess at the load of a migration-source cpu weighted
1489 * according to the scheduling class and "nice" value.
1490 *
1491 * We want to under-estimate the load of migration sources, to
1492 * balance conservatively.
1493 */
1494 static unsigned long source_load(int cpu, int type)
1495 {
1496 struct rq *rq = cpu_rq(cpu);
1497 unsigned long total = weighted_cpuload(cpu);
1498
1499 if (type == 0 || !sched_feat(LB_BIAS))
1500 return total;
1501
1502 return min(rq->cpu_load[type-1], total);
1503 }
1504
1505 /*
1506 * Return a high guess at the load of a migration-target cpu weighted
1507 * according to the scheduling class and "nice" value.
1508 */
1509 static unsigned long target_load(int cpu, int type)
1510 {
1511 struct rq *rq = cpu_rq(cpu);
1512 unsigned long total = weighted_cpuload(cpu);
1513
1514 if (type == 0 || !sched_feat(LB_BIAS))
1515 return total;
1516
1517 return max(rq->cpu_load[type-1], total);
1518 }
1519
1520 static unsigned long power_of(int cpu)
1521 {
1522 return cpu_rq(cpu)->cpu_power;
1523 }
1524
1525 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1526
1527 static unsigned long cpu_avg_load_per_task(int cpu)
1528 {
1529 struct rq *rq = cpu_rq(cpu);
1530 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1531
1532 if (nr_running)
1533 rq->avg_load_per_task = rq->load.weight / nr_running;
1534 else
1535 rq->avg_load_per_task = 0;
1536
1537 return rq->avg_load_per_task;
1538 }
1539
1540 #ifdef CONFIG_FAIR_GROUP_SCHED
1541
1542 /*
1543 * Compute the cpu's hierarchical load factor for each task group.
1544 * This needs to be done in a top-down fashion because the load of a child
1545 * group is a fraction of its parents load.
1546 */
1547 static int tg_load_down(struct task_group *tg, void *data)
1548 {
1549 unsigned long load;
1550 long cpu = (long)data;
1551
1552 if (!tg->parent) {
1553 load = cpu_rq(cpu)->load.weight;
1554 } else {
1555 load = tg->parent->cfs_rq[cpu]->h_load;
1556 load *= tg->se[cpu]->load.weight;
1557 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1558 }
1559
1560 tg->cfs_rq[cpu]->h_load = load;
1561
1562 return 0;
1563 }
1564
1565 static void update_h_load(long cpu)
1566 {
1567 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1568 }
1569
1570 #endif
1571
1572 #ifdef CONFIG_PREEMPT
1573
1574 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1575
1576 /*
1577 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1578 * way at the expense of forcing extra atomic operations in all
1579 * invocations. This assures that the double_lock is acquired using the
1580 * same underlying policy as the spinlock_t on this architecture, which
1581 * reduces latency compared to the unfair variant below. However, it
1582 * also adds more overhead and therefore may reduce throughput.
1583 */
1584 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1585 __releases(this_rq->lock)
1586 __acquires(busiest->lock)
1587 __acquires(this_rq->lock)
1588 {
1589 raw_spin_unlock(&this_rq->lock);
1590 double_rq_lock(this_rq, busiest);
1591
1592 return 1;
1593 }
1594
1595 #else
1596 /*
1597 * Unfair double_lock_balance: Optimizes throughput at the expense of
1598 * latency by eliminating extra atomic operations when the locks are
1599 * already in proper order on entry. This favors lower cpu-ids and will
1600 * grant the double lock to lower cpus over higher ids under contention,
1601 * regardless of entry order into the function.
1602 */
1603 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1604 __releases(this_rq->lock)
1605 __acquires(busiest->lock)
1606 __acquires(this_rq->lock)
1607 {
1608 int ret = 0;
1609
1610 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1611 if (busiest < this_rq) {
1612 raw_spin_unlock(&this_rq->lock);
1613 raw_spin_lock(&busiest->lock);
1614 raw_spin_lock_nested(&this_rq->lock,
1615 SINGLE_DEPTH_NESTING);
1616 ret = 1;
1617 } else
1618 raw_spin_lock_nested(&busiest->lock,
1619 SINGLE_DEPTH_NESTING);
1620 }
1621 return ret;
1622 }
1623
1624 #endif /* CONFIG_PREEMPT */
1625
1626 /*
1627 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1628 */
1629 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1630 {
1631 if (unlikely(!irqs_disabled())) {
1632 /* printk() doesn't work good under rq->lock */
1633 raw_spin_unlock(&this_rq->lock);
1634 BUG_ON(1);
1635 }
1636
1637 return _double_lock_balance(this_rq, busiest);
1638 }
1639
1640 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1641 __releases(busiest->lock)
1642 {
1643 raw_spin_unlock(&busiest->lock);
1644 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1645 }
1646
1647 /*
1648 * double_rq_lock - safely lock two runqueues
1649 *
1650 * Note this does not disable interrupts like task_rq_lock,
1651 * you need to do so manually before calling.
1652 */
1653 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1654 __acquires(rq1->lock)
1655 __acquires(rq2->lock)
1656 {
1657 BUG_ON(!irqs_disabled());
1658 if (rq1 == rq2) {
1659 raw_spin_lock(&rq1->lock);
1660 __acquire(rq2->lock); /* Fake it out ;) */
1661 } else {
1662 if (rq1 < rq2) {
1663 raw_spin_lock(&rq1->lock);
1664 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1665 } else {
1666 raw_spin_lock(&rq2->lock);
1667 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1668 }
1669 }
1670 }
1671
1672 /*
1673 * double_rq_unlock - safely unlock two runqueues
1674 *
1675 * Note this does not restore interrupts like task_rq_unlock,
1676 * you need to do so manually after calling.
1677 */
1678 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1679 __releases(rq1->lock)
1680 __releases(rq2->lock)
1681 {
1682 raw_spin_unlock(&rq1->lock);
1683 if (rq1 != rq2)
1684 raw_spin_unlock(&rq2->lock);
1685 else
1686 __release(rq2->lock);
1687 }
1688
1689 #endif
1690
1691 static void calc_load_account_idle(struct rq *this_rq);
1692 static void update_sysctl(void);
1693 static int get_update_sysctl_factor(void);
1694 static void update_cpu_load(struct rq *this_rq);
1695
1696 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1697 {
1698 set_task_rq(p, cpu);
1699 #ifdef CONFIG_SMP
1700 /*
1701 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1702 * successfuly executed on another CPU. We must ensure that updates of
1703 * per-task data have been completed by this moment.
1704 */
1705 smp_wmb();
1706 task_thread_info(p)->cpu = cpu;
1707 #endif
1708 }
1709
1710 static const struct sched_class rt_sched_class;
1711
1712 #define sched_class_highest (&stop_sched_class)
1713 #define for_each_class(class) \
1714 for (class = sched_class_highest; class; class = class->next)
1715
1716 #include "sched_stats.h"
1717
1718 static void inc_nr_running(struct rq *rq)
1719 {
1720 rq->nr_running++;
1721 }
1722
1723 static void dec_nr_running(struct rq *rq)
1724 {
1725 rq->nr_running--;
1726 }
1727
1728 static void set_load_weight(struct task_struct *p)
1729 {
1730 /*
1731 * SCHED_IDLE tasks get minimal weight:
1732 */
1733 if (p->policy == SCHED_IDLE) {
1734 p->se.load.weight = WEIGHT_IDLEPRIO;
1735 p->se.load.inv_weight = WMULT_IDLEPRIO;
1736 return;
1737 }
1738
1739 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1740 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1741 }
1742
1743 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1744 {
1745 update_rq_clock(rq);
1746 sched_info_queued(p);
1747 p->sched_class->enqueue_task(rq, p, flags);
1748 p->se.on_rq = 1;
1749 }
1750
1751 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1752 {
1753 update_rq_clock(rq);
1754 sched_info_dequeued(p);
1755 p->sched_class->dequeue_task(rq, p, flags);
1756 p->se.on_rq = 0;
1757 }
1758
1759 /*
1760 * activate_task - move a task to the runqueue.
1761 */
1762 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1763 {
1764 if (task_contributes_to_load(p))
1765 rq->nr_uninterruptible--;
1766
1767 enqueue_task(rq, p, flags);
1768 inc_nr_running(rq);
1769 }
1770
1771 /*
1772 * deactivate_task - remove a task from the runqueue.
1773 */
1774 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1775 {
1776 if (task_contributes_to_load(p))
1777 rq->nr_uninterruptible++;
1778
1779 dequeue_task(rq, p, flags);
1780 dec_nr_running(rq);
1781 }
1782
1783 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1784
1785 /*
1786 * There are no locks covering percpu hardirq/softirq time.
1787 * They are only modified in account_system_vtime, on corresponding CPU
1788 * with interrupts disabled. So, writes are safe.
1789 * They are read and saved off onto struct rq in update_rq_clock().
1790 * This may result in other CPU reading this CPU's irq time and can
1791 * race with irq/account_system_vtime on this CPU. We would either get old
1792 * or new value with a side effect of accounting a slice of irq time to wrong
1793 * task when irq is in progress while we read rq->clock. That is a worthy
1794 * compromise in place of having locks on each irq in account_system_time.
1795 */
1796 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1797 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1798
1799 static DEFINE_PER_CPU(u64, irq_start_time);
1800 static int sched_clock_irqtime;
1801
1802 void enable_sched_clock_irqtime(void)
1803 {
1804 sched_clock_irqtime = 1;
1805 }
1806
1807 void disable_sched_clock_irqtime(void)
1808 {
1809 sched_clock_irqtime = 0;
1810 }
1811
1812 #ifndef CONFIG_64BIT
1813 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1814
1815 static inline void irq_time_write_begin(void)
1816 {
1817 __this_cpu_inc(irq_time_seq.sequence);
1818 smp_wmb();
1819 }
1820
1821 static inline void irq_time_write_end(void)
1822 {
1823 smp_wmb();
1824 __this_cpu_inc(irq_time_seq.sequence);
1825 }
1826
1827 static inline u64 irq_time_read(int cpu)
1828 {
1829 u64 irq_time;
1830 unsigned seq;
1831
1832 do {
1833 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1834 irq_time = per_cpu(cpu_softirq_time, cpu) +
1835 per_cpu(cpu_hardirq_time, cpu);
1836 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1837
1838 return irq_time;
1839 }
1840 #else /* CONFIG_64BIT */
1841 static inline void irq_time_write_begin(void)
1842 {
1843 }
1844
1845 static inline void irq_time_write_end(void)
1846 {
1847 }
1848
1849 static inline u64 irq_time_read(int cpu)
1850 {
1851 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1852 }
1853 #endif /* CONFIG_64BIT */
1854
1855 /*
1856 * Called before incrementing preempt_count on {soft,}irq_enter
1857 * and before decrementing preempt_count on {soft,}irq_exit.
1858 */
1859 void account_system_vtime(struct task_struct *curr)
1860 {
1861 unsigned long flags;
1862 s64 delta;
1863 int cpu;
1864
1865 if (!sched_clock_irqtime)
1866 return;
1867
1868 local_irq_save(flags);
1869
1870 cpu = smp_processor_id();
1871 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1872 __this_cpu_add(irq_start_time, delta);
1873
1874 irq_time_write_begin();
1875 /*
1876 * We do not account for softirq time from ksoftirqd here.
1877 * We want to continue accounting softirq time to ksoftirqd thread
1878 * in that case, so as not to confuse scheduler with a special task
1879 * that do not consume any time, but still wants to run.
1880 */
1881 if (hardirq_count())
1882 __this_cpu_add(cpu_hardirq_time, delta);
1883 else if (in_serving_softirq() && !(curr->flags & PF_KSOFTIRQD))
1884 __this_cpu_add(cpu_softirq_time, delta);
1885
1886 irq_time_write_end();
1887 local_irq_restore(flags);
1888 }
1889 EXPORT_SYMBOL_GPL(account_system_vtime);
1890
1891 static void update_rq_clock_task(struct rq *rq, s64 delta)
1892 {
1893 s64 irq_delta;
1894
1895 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1896
1897 /*
1898 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1899 * this case when a previous update_rq_clock() happened inside a
1900 * {soft,}irq region.
1901 *
1902 * When this happens, we stop ->clock_task and only update the
1903 * prev_irq_time stamp to account for the part that fit, so that a next
1904 * update will consume the rest. This ensures ->clock_task is
1905 * monotonic.
1906 *
1907 * It does however cause some slight miss-attribution of {soft,}irq
1908 * time, a more accurate solution would be to update the irq_time using
1909 * the current rq->clock timestamp, except that would require using
1910 * atomic ops.
1911 */
1912 if (irq_delta > delta)
1913 irq_delta = delta;
1914
1915 rq->prev_irq_time += irq_delta;
1916 delta -= irq_delta;
1917 rq->clock_task += delta;
1918
1919 if (irq_delta && sched_feat(NONIRQ_POWER))
1920 sched_rt_avg_update(rq, irq_delta);
1921 }
1922
1923 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
1924
1925 static void update_rq_clock_task(struct rq *rq, s64 delta)
1926 {
1927 rq->clock_task += delta;
1928 }
1929
1930 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1931
1932 #include "sched_idletask.c"
1933 #include "sched_fair.c"
1934 #include "sched_rt.c"
1935 #include "sched_autogroup.c"
1936 #include "sched_stoptask.c"
1937 #ifdef CONFIG_SCHED_DEBUG
1938 # include "sched_debug.c"
1939 #endif
1940
1941 void sched_set_stop_task(int cpu, struct task_struct *stop)
1942 {
1943 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1944 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1945
1946 if (stop) {
1947 /*
1948 * Make it appear like a SCHED_FIFO task, its something
1949 * userspace knows about and won't get confused about.
1950 *
1951 * Also, it will make PI more or less work without too
1952 * much confusion -- but then, stop work should not
1953 * rely on PI working anyway.
1954 */
1955 sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
1956
1957 stop->sched_class = &stop_sched_class;
1958 }
1959
1960 cpu_rq(cpu)->stop = stop;
1961
1962 if (old_stop) {
1963 /*
1964 * Reset it back to a normal scheduling class so that
1965 * it can die in pieces.
1966 */
1967 old_stop->sched_class = &rt_sched_class;
1968 }
1969 }
1970
1971 /*
1972 * __normal_prio - return the priority that is based on the static prio
1973 */
1974 static inline int __normal_prio(struct task_struct *p)
1975 {
1976 return p->static_prio;
1977 }
1978
1979 /*
1980 * Calculate the expected normal priority: i.e. priority
1981 * without taking RT-inheritance into account. Might be
1982 * boosted by interactivity modifiers. Changes upon fork,
1983 * setprio syscalls, and whenever the interactivity
1984 * estimator recalculates.
1985 */
1986 static inline int normal_prio(struct task_struct *p)
1987 {
1988 int prio;
1989
1990 if (task_has_rt_policy(p))
1991 prio = MAX_RT_PRIO-1 - p->rt_priority;
1992 else
1993 prio = __normal_prio(p);
1994 return prio;
1995 }
1996
1997 /*
1998 * Calculate the current priority, i.e. the priority
1999 * taken into account by the scheduler. This value might
2000 * be boosted by RT tasks, or might be boosted by
2001 * interactivity modifiers. Will be RT if the task got
2002 * RT-boosted. If not then it returns p->normal_prio.
2003 */
2004 static int effective_prio(struct task_struct *p)
2005 {
2006 p->normal_prio = normal_prio(p);
2007 /*
2008 * If we are RT tasks or we were boosted to RT priority,
2009 * keep the priority unchanged. Otherwise, update priority
2010 * to the normal priority:
2011 */
2012 if (!rt_prio(p->prio))
2013 return p->normal_prio;
2014 return p->prio;
2015 }
2016
2017 /**
2018 * task_curr - is this task currently executing on a CPU?
2019 * @p: the task in question.
2020 */
2021 inline int task_curr(const struct task_struct *p)
2022 {
2023 return cpu_curr(task_cpu(p)) == p;
2024 }
2025
2026 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2027 const struct sched_class *prev_class,
2028 int oldprio, int running)
2029 {
2030 if (prev_class != p->sched_class) {
2031 if (prev_class->switched_from)
2032 prev_class->switched_from(rq, p, running);
2033 p->sched_class->switched_to(rq, p, running);
2034 } else
2035 p->sched_class->prio_changed(rq, p, oldprio, running);
2036 }
2037
2038 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2039 {
2040 const struct sched_class *class;
2041
2042 if (p->sched_class == rq->curr->sched_class) {
2043 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2044 } else {
2045 for_each_class(class) {
2046 if (class == rq->curr->sched_class)
2047 break;
2048 if (class == p->sched_class) {
2049 resched_task(rq->curr);
2050 break;
2051 }
2052 }
2053 }
2054
2055 /*
2056 * A queue event has occurred, and we're going to schedule. In
2057 * this case, we can save a useless back to back clock update.
2058 */
2059 if (rq->curr->se.on_rq && test_tsk_need_resched(rq->curr))
2060 rq->skip_clock_update = 1;
2061 }
2062
2063 #ifdef CONFIG_SMP
2064 /*
2065 * Is this task likely cache-hot:
2066 */
2067 static int
2068 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2069 {
2070 s64 delta;
2071
2072 if (p->sched_class != &fair_sched_class)
2073 return 0;
2074
2075 if (unlikely(p->policy == SCHED_IDLE))
2076 return 0;
2077
2078 /*
2079 * Buddy candidates are cache hot:
2080 */
2081 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2082 (&p->se == cfs_rq_of(&p->se)->next ||
2083 &p->se == cfs_rq_of(&p->se)->last))
2084 return 1;
2085
2086 if (sysctl_sched_migration_cost == -1)
2087 return 1;
2088 if (sysctl_sched_migration_cost == 0)
2089 return 0;
2090
2091 delta = now - p->se.exec_start;
2092
2093 return delta < (s64)sysctl_sched_migration_cost;
2094 }
2095
2096 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2097 {
2098 #ifdef CONFIG_SCHED_DEBUG
2099 /*
2100 * We should never call set_task_cpu() on a blocked task,
2101 * ttwu() will sort out the placement.
2102 */
2103 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2104 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2105 #endif
2106
2107 trace_sched_migrate_task(p, new_cpu);
2108
2109 if (task_cpu(p) != new_cpu) {
2110 p->se.nr_migrations++;
2111 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2112 }
2113
2114 __set_task_cpu(p, new_cpu);
2115 }
2116
2117 struct migration_arg {
2118 struct task_struct *task;
2119 int dest_cpu;
2120 };
2121
2122 static int migration_cpu_stop(void *data);
2123
2124 /*
2125 * The task's runqueue lock must be held.
2126 * Returns true if you have to wait for migration thread.
2127 */
2128 static bool migrate_task(struct task_struct *p, struct rq *rq)
2129 {
2130 /*
2131 * If the task is not on a runqueue (and not running), then
2132 * the next wake-up will properly place the task.
2133 */
2134 return p->se.on_rq || task_running(rq, p);
2135 }
2136
2137 /*
2138 * wait_task_inactive - wait for a thread to unschedule.
2139 *
2140 * If @match_state is nonzero, it's the @p->state value just checked and
2141 * not expected to change. If it changes, i.e. @p might have woken up,
2142 * then return zero. When we succeed in waiting for @p to be off its CPU,
2143 * we return a positive number (its total switch count). If a second call
2144 * a short while later returns the same number, the caller can be sure that
2145 * @p has remained unscheduled the whole time.
2146 *
2147 * The caller must ensure that the task *will* unschedule sometime soon,
2148 * else this function might spin for a *long* time. This function can't
2149 * be called with interrupts off, or it may introduce deadlock with
2150 * smp_call_function() if an IPI is sent by the same process we are
2151 * waiting to become inactive.
2152 */
2153 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2154 {
2155 unsigned long flags;
2156 int running, on_rq;
2157 unsigned long ncsw;
2158 struct rq *rq;
2159
2160 for (;;) {
2161 /*
2162 * We do the initial early heuristics without holding
2163 * any task-queue locks at all. We'll only try to get
2164 * the runqueue lock when things look like they will
2165 * work out!
2166 */
2167 rq = task_rq(p);
2168
2169 /*
2170 * If the task is actively running on another CPU
2171 * still, just relax and busy-wait without holding
2172 * any locks.
2173 *
2174 * NOTE! Since we don't hold any locks, it's not
2175 * even sure that "rq" stays as the right runqueue!
2176 * But we don't care, since "task_running()" will
2177 * return false if the runqueue has changed and p
2178 * is actually now running somewhere else!
2179 */
2180 while (task_running(rq, p)) {
2181 if (match_state && unlikely(p->state != match_state))
2182 return 0;
2183 cpu_relax();
2184 }
2185
2186 /*
2187 * Ok, time to look more closely! We need the rq
2188 * lock now, to be *sure*. If we're wrong, we'll
2189 * just go back and repeat.
2190 */
2191 rq = task_rq_lock(p, &flags);
2192 trace_sched_wait_task(p);
2193 running = task_running(rq, p);
2194 on_rq = p->se.on_rq;
2195 ncsw = 0;
2196 if (!match_state || p->state == match_state)
2197 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2198 task_rq_unlock(rq, &flags);
2199
2200 /*
2201 * If it changed from the expected state, bail out now.
2202 */
2203 if (unlikely(!ncsw))
2204 break;
2205
2206 /*
2207 * Was it really running after all now that we
2208 * checked with the proper locks actually held?
2209 *
2210 * Oops. Go back and try again..
2211 */
2212 if (unlikely(running)) {
2213 cpu_relax();
2214 continue;
2215 }
2216
2217 /*
2218 * It's not enough that it's not actively running,
2219 * it must be off the runqueue _entirely_, and not
2220 * preempted!
2221 *
2222 * So if it was still runnable (but just not actively
2223 * running right now), it's preempted, and we should
2224 * yield - it could be a while.
2225 */
2226 if (unlikely(on_rq)) {
2227 schedule_timeout_uninterruptible(1);
2228 continue;
2229 }
2230
2231 /*
2232 * Ahh, all good. It wasn't running, and it wasn't
2233 * runnable, which means that it will never become
2234 * running in the future either. We're all done!
2235 */
2236 break;
2237 }
2238
2239 return ncsw;
2240 }
2241
2242 /***
2243 * kick_process - kick a running thread to enter/exit the kernel
2244 * @p: the to-be-kicked thread
2245 *
2246 * Cause a process which is running on another CPU to enter
2247 * kernel-mode, without any delay. (to get signals handled.)
2248 *
2249 * NOTE: this function doesnt have to take the runqueue lock,
2250 * because all it wants to ensure is that the remote task enters
2251 * the kernel. If the IPI races and the task has been migrated
2252 * to another CPU then no harm is done and the purpose has been
2253 * achieved as well.
2254 */
2255 void kick_process(struct task_struct *p)
2256 {
2257 int cpu;
2258
2259 preempt_disable();
2260 cpu = task_cpu(p);
2261 if ((cpu != smp_processor_id()) && task_curr(p))
2262 smp_send_reschedule(cpu);
2263 preempt_enable();
2264 }
2265 EXPORT_SYMBOL_GPL(kick_process);
2266 #endif /* CONFIG_SMP */
2267
2268 /**
2269 * task_oncpu_function_call - call a function on the cpu on which a task runs
2270 * @p: the task to evaluate
2271 * @func: the function to be called
2272 * @info: the function call argument
2273 *
2274 * Calls the function @func when the task is currently running. This might
2275 * be on the current CPU, which just calls the function directly
2276 */
2277 void task_oncpu_function_call(struct task_struct *p,
2278 void (*func) (void *info), void *info)
2279 {
2280 int cpu;
2281
2282 preempt_disable();
2283 cpu = task_cpu(p);
2284 if (task_curr(p))
2285 smp_call_function_single(cpu, func, info, 1);
2286 preempt_enable();
2287 }
2288
2289 #ifdef CONFIG_SMP
2290 /*
2291 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2292 */
2293 static int select_fallback_rq(int cpu, struct task_struct *p)
2294 {
2295 int dest_cpu;
2296 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2297
2298 /* Look for allowed, online CPU in same node. */
2299 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2300 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2301 return dest_cpu;
2302
2303 /* Any allowed, online CPU? */
2304 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2305 if (dest_cpu < nr_cpu_ids)
2306 return dest_cpu;
2307
2308 /* No more Mr. Nice Guy. */
2309 dest_cpu = cpuset_cpus_allowed_fallback(p);
2310 /*
2311 * Don't tell them about moving exiting tasks or
2312 * kernel threads (both mm NULL), since they never
2313 * leave kernel.
2314 */
2315 if (p->mm && printk_ratelimit()) {
2316 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2317 task_pid_nr(p), p->comm, cpu);
2318 }
2319
2320 return dest_cpu;
2321 }
2322
2323 /*
2324 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2325 */
2326 static inline
2327 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2328 {
2329 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2330
2331 /*
2332 * In order not to call set_task_cpu() on a blocking task we need
2333 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2334 * cpu.
2335 *
2336 * Since this is common to all placement strategies, this lives here.
2337 *
2338 * [ this allows ->select_task() to simply return task_cpu(p) and
2339 * not worry about this generic constraint ]
2340 */
2341 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2342 !cpu_online(cpu)))
2343 cpu = select_fallback_rq(task_cpu(p), p);
2344
2345 return cpu;
2346 }
2347
2348 static void update_avg(u64 *avg, u64 sample)
2349 {
2350 s64 diff = sample - *avg;
2351 *avg += diff >> 3;
2352 }
2353 #endif
2354
2355 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2356 bool is_sync, bool is_migrate, bool is_local,
2357 unsigned long en_flags)
2358 {
2359 schedstat_inc(p, se.statistics.nr_wakeups);
2360 if (is_sync)
2361 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2362 if (is_migrate)
2363 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2364 if (is_local)
2365 schedstat_inc(p, se.statistics.nr_wakeups_local);
2366 else
2367 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2368
2369 activate_task(rq, p, en_flags);
2370 }
2371
2372 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2373 int wake_flags, bool success)
2374 {
2375 trace_sched_wakeup(p, success);
2376 check_preempt_curr(rq, p, wake_flags);
2377
2378 p->state = TASK_RUNNING;
2379 #ifdef CONFIG_SMP
2380 if (p->sched_class->task_woken)
2381 p->sched_class->task_woken(rq, p);
2382
2383 if (unlikely(rq->idle_stamp)) {
2384 u64 delta = rq->clock - rq->idle_stamp;
2385 u64 max = 2*sysctl_sched_migration_cost;
2386
2387 if (delta > max)
2388 rq->avg_idle = max;
2389 else
2390 update_avg(&rq->avg_idle, delta);
2391 rq->idle_stamp = 0;
2392 }
2393 #endif
2394 /* if a worker is waking up, notify workqueue */
2395 if ((p->flags & PF_WQ_WORKER) && success)
2396 wq_worker_waking_up(p, cpu_of(rq));
2397 }
2398
2399 /**
2400 * try_to_wake_up - wake up a thread
2401 * @p: the thread to be awakened
2402 * @state: the mask of task states that can be woken
2403 * @wake_flags: wake modifier flags (WF_*)
2404 *
2405 * Put it on the run-queue if it's not already there. The "current"
2406 * thread is always on the run-queue (except when the actual
2407 * re-schedule is in progress), and as such you're allowed to do
2408 * the simpler "current->state = TASK_RUNNING" to mark yourself
2409 * runnable without the overhead of this.
2410 *
2411 * Returns %true if @p was woken up, %false if it was already running
2412 * or @state didn't match @p's state.
2413 */
2414 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2415 int wake_flags)
2416 {
2417 int cpu, orig_cpu, this_cpu, success = 0;
2418 unsigned long flags;
2419 unsigned long en_flags = ENQUEUE_WAKEUP;
2420 struct rq *rq;
2421
2422 this_cpu = get_cpu();
2423
2424 smp_wmb();
2425 rq = task_rq_lock(p, &flags);
2426 if (!(p->state & state))
2427 goto out;
2428
2429 if (p->se.on_rq)
2430 goto out_running;
2431
2432 cpu = task_cpu(p);
2433 orig_cpu = cpu;
2434
2435 #ifdef CONFIG_SMP
2436 if (unlikely(task_running(rq, p)))
2437 goto out_activate;
2438
2439 /*
2440 * In order to handle concurrent wakeups and release the rq->lock
2441 * we put the task in TASK_WAKING state.
2442 *
2443 * First fix up the nr_uninterruptible count:
2444 */
2445 if (task_contributes_to_load(p)) {
2446 if (likely(cpu_online(orig_cpu)))
2447 rq->nr_uninterruptible--;
2448 else
2449 this_rq()->nr_uninterruptible--;
2450 }
2451 p->state = TASK_WAKING;
2452
2453 if (p->sched_class->task_waking) {
2454 p->sched_class->task_waking(rq, p);
2455 en_flags |= ENQUEUE_WAKING;
2456 }
2457
2458 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2459 if (cpu != orig_cpu)
2460 set_task_cpu(p, cpu);
2461 __task_rq_unlock(rq);
2462
2463 rq = cpu_rq(cpu);
2464 raw_spin_lock(&rq->lock);
2465
2466 /*
2467 * We migrated the task without holding either rq->lock, however
2468 * since the task is not on the task list itself, nobody else
2469 * will try and migrate the task, hence the rq should match the
2470 * cpu we just moved it to.
2471 */
2472 WARN_ON(task_cpu(p) != cpu);
2473 WARN_ON(p->state != TASK_WAKING);
2474
2475 #ifdef CONFIG_SCHEDSTATS
2476 schedstat_inc(rq, ttwu_count);
2477 if (cpu == this_cpu)
2478 schedstat_inc(rq, ttwu_local);
2479 else {
2480 struct sched_domain *sd;
2481 for_each_domain(this_cpu, sd) {
2482 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2483 schedstat_inc(sd, ttwu_wake_remote);
2484 break;
2485 }
2486 }
2487 }
2488 #endif /* CONFIG_SCHEDSTATS */
2489
2490 out_activate:
2491 #endif /* CONFIG_SMP */
2492 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2493 cpu == this_cpu, en_flags);
2494 success = 1;
2495 out_running:
2496 ttwu_post_activation(p, rq, wake_flags, success);
2497 out:
2498 task_rq_unlock(rq, &flags);
2499 put_cpu();
2500
2501 return success;
2502 }
2503
2504 /**
2505 * try_to_wake_up_local - try to wake up a local task with rq lock held
2506 * @p: the thread to be awakened
2507 *
2508 * Put @p on the run-queue if it's not already there. The caller must
2509 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2510 * the current task. this_rq() stays locked over invocation.
2511 */
2512 static void try_to_wake_up_local(struct task_struct *p)
2513 {
2514 struct rq *rq = task_rq(p);
2515 bool success = false;
2516
2517 BUG_ON(rq != this_rq());
2518 BUG_ON(p == current);
2519 lockdep_assert_held(&rq->lock);
2520
2521 if (!(p->state & TASK_NORMAL))
2522 return;
2523
2524 if (!p->se.on_rq) {
2525 if (likely(!task_running(rq, p))) {
2526 schedstat_inc(rq, ttwu_count);
2527 schedstat_inc(rq, ttwu_local);
2528 }
2529 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2530 success = true;
2531 }
2532 ttwu_post_activation(p, rq, 0, success);
2533 }
2534
2535 /**
2536 * wake_up_process - Wake up a specific process
2537 * @p: The process to be woken up.
2538 *
2539 * Attempt to wake up the nominated process and move it to the set of runnable
2540 * processes. Returns 1 if the process was woken up, 0 if it was already
2541 * running.
2542 *
2543 * It may be assumed that this function implies a write memory barrier before
2544 * changing the task state if and only if any tasks are woken up.
2545 */
2546 int wake_up_process(struct task_struct *p)
2547 {
2548 return try_to_wake_up(p, TASK_ALL, 0);
2549 }
2550 EXPORT_SYMBOL(wake_up_process);
2551
2552 int wake_up_state(struct task_struct *p, unsigned int state)
2553 {
2554 return try_to_wake_up(p, state, 0);
2555 }
2556
2557 /*
2558 * Perform scheduler related setup for a newly forked process p.
2559 * p is forked by current.
2560 *
2561 * __sched_fork() is basic setup used by init_idle() too:
2562 */
2563 static void __sched_fork(struct task_struct *p)
2564 {
2565 p->se.exec_start = 0;
2566 p->se.sum_exec_runtime = 0;
2567 p->se.prev_sum_exec_runtime = 0;
2568 p->se.nr_migrations = 0;
2569
2570 #ifdef CONFIG_SCHEDSTATS
2571 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2572 #endif
2573
2574 INIT_LIST_HEAD(&p->rt.run_list);
2575 p->se.on_rq = 0;
2576 INIT_LIST_HEAD(&p->se.group_node);
2577
2578 #ifdef CONFIG_PREEMPT_NOTIFIERS
2579 INIT_HLIST_HEAD(&p->preempt_notifiers);
2580 #endif
2581 }
2582
2583 /*
2584 * fork()/clone()-time setup:
2585 */
2586 void sched_fork(struct task_struct *p, int clone_flags)
2587 {
2588 int cpu = get_cpu();
2589
2590 __sched_fork(p);
2591 /*
2592 * We mark the process as running here. This guarantees that
2593 * nobody will actually run it, and a signal or other external
2594 * event cannot wake it up and insert it on the runqueue either.
2595 */
2596 p->state = TASK_RUNNING;
2597
2598 /*
2599 * Revert to default priority/policy on fork if requested.
2600 */
2601 if (unlikely(p->sched_reset_on_fork)) {
2602 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2603 p->policy = SCHED_NORMAL;
2604 p->normal_prio = p->static_prio;
2605 }
2606
2607 if (PRIO_TO_NICE(p->static_prio) < 0) {
2608 p->static_prio = NICE_TO_PRIO(0);
2609 p->normal_prio = p->static_prio;
2610 set_load_weight(p);
2611 }
2612
2613 /*
2614 * We don't need the reset flag anymore after the fork. It has
2615 * fulfilled its duty:
2616 */
2617 p->sched_reset_on_fork = 0;
2618 }
2619
2620 /*
2621 * Make sure we do not leak PI boosting priority to the child.
2622 */
2623 p->prio = current->normal_prio;
2624
2625 if (!rt_prio(p->prio))
2626 p->sched_class = &fair_sched_class;
2627
2628 if (p->sched_class->task_fork)
2629 p->sched_class->task_fork(p);
2630
2631 /*
2632 * The child is not yet in the pid-hash so no cgroup attach races,
2633 * and the cgroup is pinned to this child due to cgroup_fork()
2634 * is ran before sched_fork().
2635 *
2636 * Silence PROVE_RCU.
2637 */
2638 rcu_read_lock();
2639 set_task_cpu(p, cpu);
2640 rcu_read_unlock();
2641
2642 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2643 if (likely(sched_info_on()))
2644 memset(&p->sched_info, 0, sizeof(p->sched_info));
2645 #endif
2646 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2647 p->oncpu = 0;
2648 #endif
2649 #ifdef CONFIG_PREEMPT
2650 /* Want to start with kernel preemption disabled. */
2651 task_thread_info(p)->preempt_count = 1;
2652 #endif
2653 #ifdef CONFIG_SMP
2654 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2655 #endif
2656
2657 put_cpu();
2658 }
2659
2660 /*
2661 * wake_up_new_task - wake up a newly created task for the first time.
2662 *
2663 * This function will do some initial scheduler statistics housekeeping
2664 * that must be done for every newly created context, then puts the task
2665 * on the runqueue and wakes it.
2666 */
2667 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2668 {
2669 unsigned long flags;
2670 struct rq *rq;
2671 int cpu __maybe_unused = get_cpu();
2672
2673 #ifdef CONFIG_SMP
2674 rq = task_rq_lock(p, &flags);
2675 p->state = TASK_WAKING;
2676
2677 /*
2678 * Fork balancing, do it here and not earlier because:
2679 * - cpus_allowed can change in the fork path
2680 * - any previously selected cpu might disappear through hotplug
2681 *
2682 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2683 * without people poking at ->cpus_allowed.
2684 */
2685 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2686 set_task_cpu(p, cpu);
2687
2688 p->state = TASK_RUNNING;
2689 task_rq_unlock(rq, &flags);
2690 #endif
2691
2692 rq = task_rq_lock(p, &flags);
2693 activate_task(rq, p, 0);
2694 trace_sched_wakeup_new(p, 1);
2695 check_preempt_curr(rq, p, WF_FORK);
2696 #ifdef CONFIG_SMP
2697 if (p->sched_class->task_woken)
2698 p->sched_class->task_woken(rq, p);
2699 #endif
2700 task_rq_unlock(rq, &flags);
2701 put_cpu();
2702 }
2703
2704 #ifdef CONFIG_PREEMPT_NOTIFIERS
2705
2706 /**
2707 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2708 * @notifier: notifier struct to register
2709 */
2710 void preempt_notifier_register(struct preempt_notifier *notifier)
2711 {
2712 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2713 }
2714 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2715
2716 /**
2717 * preempt_notifier_unregister - no longer interested in preemption notifications
2718 * @notifier: notifier struct to unregister
2719 *
2720 * This is safe to call from within a preemption notifier.
2721 */
2722 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2723 {
2724 hlist_del(&notifier->link);
2725 }
2726 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2727
2728 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2729 {
2730 struct preempt_notifier *notifier;
2731 struct hlist_node *node;
2732
2733 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2734 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2735 }
2736
2737 static void
2738 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2739 struct task_struct *next)
2740 {
2741 struct preempt_notifier *notifier;
2742 struct hlist_node *node;
2743
2744 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2745 notifier->ops->sched_out(notifier, next);
2746 }
2747
2748 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2749
2750 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2751 {
2752 }
2753
2754 static void
2755 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2756 struct task_struct *next)
2757 {
2758 }
2759
2760 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2761
2762 /**
2763 * prepare_task_switch - prepare to switch tasks
2764 * @rq: the runqueue preparing to switch
2765 * @prev: the current task that is being switched out
2766 * @next: the task we are going to switch to.
2767 *
2768 * This is called with the rq lock held and interrupts off. It must
2769 * be paired with a subsequent finish_task_switch after the context
2770 * switch.
2771 *
2772 * prepare_task_switch sets up locking and calls architecture specific
2773 * hooks.
2774 */
2775 static inline void
2776 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2777 struct task_struct *next)
2778 {
2779 fire_sched_out_preempt_notifiers(prev, next);
2780 prepare_lock_switch(rq, next);
2781 prepare_arch_switch(next);
2782 }
2783
2784 /**
2785 * finish_task_switch - clean up after a task-switch
2786 * @rq: runqueue associated with task-switch
2787 * @prev: the thread we just switched away from.
2788 *
2789 * finish_task_switch must be called after the context switch, paired
2790 * with a prepare_task_switch call before the context switch.
2791 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2792 * and do any other architecture-specific cleanup actions.
2793 *
2794 * Note that we may have delayed dropping an mm in context_switch(). If
2795 * so, we finish that here outside of the runqueue lock. (Doing it
2796 * with the lock held can cause deadlocks; see schedule() for
2797 * details.)
2798 */
2799 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2800 __releases(rq->lock)
2801 {
2802 struct mm_struct *mm = rq->prev_mm;
2803 long prev_state;
2804
2805 rq->prev_mm = NULL;
2806
2807 /*
2808 * A task struct has one reference for the use as "current".
2809 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2810 * schedule one last time. The schedule call will never return, and
2811 * the scheduled task must drop that reference.
2812 * The test for TASK_DEAD must occur while the runqueue locks are
2813 * still held, otherwise prev could be scheduled on another cpu, die
2814 * there before we look at prev->state, and then the reference would
2815 * be dropped twice.
2816 * Manfred Spraul <manfred@colorfullife.com>
2817 */
2818 prev_state = prev->state;
2819 finish_arch_switch(prev);
2820 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2821 local_irq_disable();
2822 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2823 perf_event_task_sched_in(current);
2824 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2825 local_irq_enable();
2826 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2827 finish_lock_switch(rq, prev);
2828
2829 fire_sched_in_preempt_notifiers(current);
2830 if (mm)
2831 mmdrop(mm);
2832 if (unlikely(prev_state == TASK_DEAD)) {
2833 /*
2834 * Remove function-return probe instances associated with this
2835 * task and put them back on the free list.
2836 */
2837 kprobe_flush_task(prev);
2838 put_task_struct(prev);
2839 }
2840 }
2841
2842 #ifdef CONFIG_SMP
2843
2844 /* assumes rq->lock is held */
2845 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2846 {
2847 if (prev->sched_class->pre_schedule)
2848 prev->sched_class->pre_schedule(rq, prev);
2849 }
2850
2851 /* rq->lock is NOT held, but preemption is disabled */
2852 static inline void post_schedule(struct rq *rq)
2853 {
2854 if (rq->post_schedule) {
2855 unsigned long flags;
2856
2857 raw_spin_lock_irqsave(&rq->lock, flags);
2858 if (rq->curr->sched_class->post_schedule)
2859 rq->curr->sched_class->post_schedule(rq);
2860 raw_spin_unlock_irqrestore(&rq->lock, flags);
2861
2862 rq->post_schedule = 0;
2863 }
2864 }
2865
2866 #else
2867
2868 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2869 {
2870 }
2871
2872 static inline void post_schedule(struct rq *rq)
2873 {
2874 }
2875
2876 #endif
2877
2878 /**
2879 * schedule_tail - first thing a freshly forked thread must call.
2880 * @prev: the thread we just switched away from.
2881 */
2882 asmlinkage void schedule_tail(struct task_struct *prev)
2883 __releases(rq->lock)
2884 {
2885 struct rq *rq = this_rq();
2886
2887 finish_task_switch(rq, prev);
2888
2889 /*
2890 * FIXME: do we need to worry about rq being invalidated by the
2891 * task_switch?
2892 */
2893 post_schedule(rq);
2894
2895 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2896 /* In this case, finish_task_switch does not reenable preemption */
2897 preempt_enable();
2898 #endif
2899 if (current->set_child_tid)
2900 put_user(task_pid_vnr(current), current->set_child_tid);
2901 }
2902
2903 /*
2904 * context_switch - switch to the new MM and the new
2905 * thread's register state.
2906 */
2907 static inline void
2908 context_switch(struct rq *rq, struct task_struct *prev,
2909 struct task_struct *next)
2910 {
2911 struct mm_struct *mm, *oldmm;
2912
2913 prepare_task_switch(rq, prev, next);
2914 trace_sched_switch(prev, next);
2915 mm = next->mm;
2916 oldmm = prev->active_mm;
2917 /*
2918 * For paravirt, this is coupled with an exit in switch_to to
2919 * combine the page table reload and the switch backend into
2920 * one hypercall.
2921 */
2922 arch_start_context_switch(prev);
2923
2924 if (!mm) {
2925 next->active_mm = oldmm;
2926 atomic_inc(&oldmm->mm_count);
2927 enter_lazy_tlb(oldmm, next);
2928 } else
2929 switch_mm(oldmm, mm, next);
2930
2931 if (!prev->mm) {
2932 prev->active_mm = NULL;
2933 rq->prev_mm = oldmm;
2934 }
2935 /*
2936 * Since the runqueue lock will be released by the next
2937 * task (which is an invalid locking op but in the case
2938 * of the scheduler it's an obvious special-case), so we
2939 * do an early lockdep release here:
2940 */
2941 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2942 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2943 #endif
2944
2945 /* Here we just switch the register state and the stack. */
2946 switch_to(prev, next, prev);
2947
2948 barrier();
2949 /*
2950 * this_rq must be evaluated again because prev may have moved
2951 * CPUs since it called schedule(), thus the 'rq' on its stack
2952 * frame will be invalid.
2953 */
2954 finish_task_switch(this_rq(), prev);
2955 }
2956
2957 /*
2958 * nr_running, nr_uninterruptible and nr_context_switches:
2959 *
2960 * externally visible scheduler statistics: current number of runnable
2961 * threads, current number of uninterruptible-sleeping threads, total
2962 * number of context switches performed since bootup.
2963 */
2964 unsigned long nr_running(void)
2965 {
2966 unsigned long i, sum = 0;
2967
2968 for_each_online_cpu(i)
2969 sum += cpu_rq(i)->nr_running;
2970
2971 return sum;
2972 }
2973
2974 unsigned long nr_uninterruptible(void)
2975 {
2976 unsigned long i, sum = 0;
2977
2978 for_each_possible_cpu(i)
2979 sum += cpu_rq(i)->nr_uninterruptible;
2980
2981 /*
2982 * Since we read the counters lockless, it might be slightly
2983 * inaccurate. Do not allow it to go below zero though:
2984 */
2985 if (unlikely((long)sum < 0))
2986 sum = 0;
2987
2988 return sum;
2989 }
2990
2991 unsigned long long nr_context_switches(void)
2992 {
2993 int i;
2994 unsigned long long sum = 0;
2995
2996 for_each_possible_cpu(i)
2997 sum += cpu_rq(i)->nr_switches;
2998
2999 return sum;
3000 }
3001
3002 unsigned long nr_iowait(void)
3003 {
3004 unsigned long i, sum = 0;
3005
3006 for_each_possible_cpu(i)
3007 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3008
3009 return sum;
3010 }
3011
3012 unsigned long nr_iowait_cpu(int cpu)
3013 {
3014 struct rq *this = cpu_rq(cpu);
3015 return atomic_read(&this->nr_iowait);
3016 }
3017
3018 unsigned long this_cpu_load(void)
3019 {
3020 struct rq *this = this_rq();
3021 return this->cpu_load[0];
3022 }
3023
3024
3025 /* Variables and functions for calc_load */
3026 static atomic_long_t calc_load_tasks;
3027 static unsigned long calc_load_update;
3028 unsigned long avenrun[3];
3029 EXPORT_SYMBOL(avenrun);
3030
3031 static long calc_load_fold_active(struct rq *this_rq)
3032 {
3033 long nr_active, delta = 0;
3034
3035 nr_active = this_rq->nr_running;
3036 nr_active += (long) this_rq->nr_uninterruptible;
3037
3038 if (nr_active != this_rq->calc_load_active) {
3039 delta = nr_active - this_rq->calc_load_active;
3040 this_rq->calc_load_active = nr_active;
3041 }
3042
3043 return delta;
3044 }
3045
3046 static unsigned long
3047 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3048 {
3049 load *= exp;
3050 load += active * (FIXED_1 - exp);
3051 load += 1UL << (FSHIFT - 1);
3052 return load >> FSHIFT;
3053 }
3054
3055 #ifdef CONFIG_NO_HZ
3056 /*
3057 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3058 *
3059 * When making the ILB scale, we should try to pull this in as well.
3060 */
3061 static atomic_long_t calc_load_tasks_idle;
3062
3063 static void calc_load_account_idle(struct rq *this_rq)
3064 {
3065 long delta;
3066
3067 delta = calc_load_fold_active(this_rq);
3068 if (delta)
3069 atomic_long_add(delta, &calc_load_tasks_idle);
3070 }
3071
3072 static long calc_load_fold_idle(void)
3073 {
3074 long delta = 0;
3075
3076 /*
3077 * Its got a race, we don't care...
3078 */
3079 if (atomic_long_read(&calc_load_tasks_idle))
3080 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3081
3082 return delta;
3083 }
3084
3085 /**
3086 * fixed_power_int - compute: x^n, in O(log n) time
3087 *
3088 * @x: base of the power
3089 * @frac_bits: fractional bits of @x
3090 * @n: power to raise @x to.
3091 *
3092 * By exploiting the relation between the definition of the natural power
3093 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3094 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3095 * (where: n_i \elem {0, 1}, the binary vector representing n),
3096 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3097 * of course trivially computable in O(log_2 n), the length of our binary
3098 * vector.
3099 */
3100 static unsigned long
3101 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3102 {
3103 unsigned long result = 1UL << frac_bits;
3104
3105 if (n) for (;;) {
3106 if (n & 1) {
3107 result *= x;
3108 result += 1UL << (frac_bits - 1);
3109 result >>= frac_bits;
3110 }
3111 n >>= 1;
3112 if (!n)
3113 break;
3114 x *= x;
3115 x += 1UL << (frac_bits - 1);
3116 x >>= frac_bits;
3117 }
3118
3119 return result;
3120 }
3121
3122 /*
3123 * a1 = a0 * e + a * (1 - e)
3124 *
3125 * a2 = a1 * e + a * (1 - e)
3126 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3127 * = a0 * e^2 + a * (1 - e) * (1 + e)
3128 *
3129 * a3 = a2 * e + a * (1 - e)
3130 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3131 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3132 *
3133 * ...
3134 *
3135 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3136 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3137 * = a0 * e^n + a * (1 - e^n)
3138 *
3139 * [1] application of the geometric series:
3140 *
3141 * n 1 - x^(n+1)
3142 * S_n := \Sum x^i = -------------
3143 * i=0 1 - x
3144 */
3145 static unsigned long
3146 calc_load_n(unsigned long load, unsigned long exp,
3147 unsigned long active, unsigned int n)
3148 {
3149
3150 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3151 }
3152
3153 /*
3154 * NO_HZ can leave us missing all per-cpu ticks calling
3155 * calc_load_account_active(), but since an idle CPU folds its delta into
3156 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3157 * in the pending idle delta if our idle period crossed a load cycle boundary.
3158 *
3159 * Once we've updated the global active value, we need to apply the exponential
3160 * weights adjusted to the number of cycles missed.
3161 */
3162 static void calc_global_nohz(unsigned long ticks)
3163 {
3164 long delta, active, n;
3165
3166 if (time_before(jiffies, calc_load_update))
3167 return;
3168
3169 /*
3170 * If we crossed a calc_load_update boundary, make sure to fold
3171 * any pending idle changes, the respective CPUs might have
3172 * missed the tick driven calc_load_account_active() update
3173 * due to NO_HZ.
3174 */
3175 delta = calc_load_fold_idle();
3176 if (delta)
3177 atomic_long_add(delta, &calc_load_tasks);
3178
3179 /*
3180 * If we were idle for multiple load cycles, apply them.
3181 */
3182 if (ticks >= LOAD_FREQ) {
3183 n = ticks / LOAD_FREQ;
3184
3185 active = atomic_long_read(&calc_load_tasks);
3186 active = active > 0 ? active * FIXED_1 : 0;
3187
3188 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3189 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3190 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3191
3192 calc_load_update += n * LOAD_FREQ;
3193 }
3194
3195 /*
3196 * Its possible the remainder of the above division also crosses
3197 * a LOAD_FREQ period, the regular check in calc_global_load()
3198 * which comes after this will take care of that.
3199 *
3200 * Consider us being 11 ticks before a cycle completion, and us
3201 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3202 * age us 4 cycles, and the test in calc_global_load() will
3203 * pick up the final one.
3204 */
3205 }
3206 #else
3207 static void calc_load_account_idle(struct rq *this_rq)
3208 {
3209 }
3210
3211 static inline long calc_load_fold_idle(void)
3212 {
3213 return 0;
3214 }
3215
3216 static void calc_global_nohz(unsigned long ticks)
3217 {
3218 }
3219 #endif
3220
3221 /**
3222 * get_avenrun - get the load average array
3223 * @loads: pointer to dest load array
3224 * @offset: offset to add
3225 * @shift: shift count to shift the result left
3226 *
3227 * These values are estimates at best, so no need for locking.
3228 */
3229 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3230 {
3231 loads[0] = (avenrun[0] + offset) << shift;
3232 loads[1] = (avenrun[1] + offset) << shift;
3233 loads[2] = (avenrun[2] + offset) << shift;
3234 }
3235
3236 /*
3237 * calc_load - update the avenrun load estimates 10 ticks after the
3238 * CPUs have updated calc_load_tasks.
3239 */
3240 void calc_global_load(unsigned long ticks)
3241 {
3242 long active;
3243
3244 calc_global_nohz(ticks);
3245
3246 if (time_before(jiffies, calc_load_update + 10))
3247 return;
3248
3249 active = atomic_long_read(&calc_load_tasks);
3250 active = active > 0 ? active * FIXED_1 : 0;
3251
3252 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3253 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3254 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3255
3256 calc_load_update += LOAD_FREQ;
3257 }
3258
3259 /*
3260 * Called from update_cpu_load() to periodically update this CPU's
3261 * active count.
3262 */
3263 static void calc_load_account_active(struct rq *this_rq)
3264 {
3265 long delta;
3266
3267 if (time_before(jiffies, this_rq->calc_load_update))
3268 return;
3269
3270 delta = calc_load_fold_active(this_rq);
3271 delta += calc_load_fold_idle();
3272 if (delta)
3273 atomic_long_add(delta, &calc_load_tasks);
3274
3275 this_rq->calc_load_update += LOAD_FREQ;
3276 }
3277
3278 /*
3279 * The exact cpuload at various idx values, calculated at every tick would be
3280 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3281 *
3282 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3283 * on nth tick when cpu may be busy, then we have:
3284 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3285 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3286 *
3287 * decay_load_missed() below does efficient calculation of
3288 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3289 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3290 *
3291 * The calculation is approximated on a 128 point scale.
3292 * degrade_zero_ticks is the number of ticks after which load at any
3293 * particular idx is approximated to be zero.
3294 * degrade_factor is a precomputed table, a row for each load idx.
3295 * Each column corresponds to degradation factor for a power of two ticks,
3296 * based on 128 point scale.
3297 * Example:
3298 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3299 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3300 *
3301 * With this power of 2 load factors, we can degrade the load n times
3302 * by looking at 1 bits in n and doing as many mult/shift instead of
3303 * n mult/shifts needed by the exact degradation.
3304 */
3305 #define DEGRADE_SHIFT 7
3306 static const unsigned char
3307 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3308 static const unsigned char
3309 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3310 {0, 0, 0, 0, 0, 0, 0, 0},
3311 {64, 32, 8, 0, 0, 0, 0, 0},
3312 {96, 72, 40, 12, 1, 0, 0},
3313 {112, 98, 75, 43, 15, 1, 0},
3314 {120, 112, 98, 76, 45, 16, 2} };
3315
3316 /*
3317 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3318 * would be when CPU is idle and so we just decay the old load without
3319 * adding any new load.
3320 */
3321 static unsigned long
3322 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3323 {
3324 int j = 0;
3325
3326 if (!missed_updates)
3327 return load;
3328
3329 if (missed_updates >= degrade_zero_ticks[idx])
3330 return 0;
3331
3332 if (idx == 1)
3333 return load >> missed_updates;
3334
3335 while (missed_updates) {
3336 if (missed_updates % 2)
3337 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3338
3339 missed_updates >>= 1;
3340 j++;
3341 }
3342 return load;
3343 }
3344
3345 /*
3346 * Update rq->cpu_load[] statistics. This function is usually called every
3347 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3348 * every tick. We fix it up based on jiffies.
3349 */
3350 static void update_cpu_load(struct rq *this_rq)
3351 {
3352 unsigned long this_load = this_rq->load.weight;
3353 unsigned long curr_jiffies = jiffies;
3354 unsigned long pending_updates;
3355 int i, scale;
3356
3357 this_rq->nr_load_updates++;
3358
3359 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3360 if (curr_jiffies == this_rq->last_load_update_tick)
3361 return;
3362
3363 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3364 this_rq->last_load_update_tick = curr_jiffies;
3365
3366 /* Update our load: */
3367 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3368 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3369 unsigned long old_load, new_load;
3370
3371 /* scale is effectively 1 << i now, and >> i divides by scale */
3372
3373 old_load = this_rq->cpu_load[i];
3374 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3375 new_load = this_load;
3376 /*
3377 * Round up the averaging division if load is increasing. This
3378 * prevents us from getting stuck on 9 if the load is 10, for
3379 * example.
3380 */
3381 if (new_load > old_load)
3382 new_load += scale - 1;
3383
3384 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3385 }
3386
3387 sched_avg_update(this_rq);
3388 }
3389
3390 static void update_cpu_load_active(struct rq *this_rq)
3391 {
3392 update_cpu_load(this_rq);
3393
3394 calc_load_account_active(this_rq);
3395 }
3396
3397 #ifdef CONFIG_SMP
3398
3399 /*
3400 * sched_exec - execve() is a valuable balancing opportunity, because at
3401 * this point the task has the smallest effective memory and cache footprint.
3402 */
3403 void sched_exec(void)
3404 {
3405 struct task_struct *p = current;
3406 unsigned long flags;
3407 struct rq *rq;
3408 int dest_cpu;
3409
3410 rq = task_rq_lock(p, &flags);
3411 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3412 if (dest_cpu == smp_processor_id())
3413 goto unlock;
3414
3415 /*
3416 * select_task_rq() can race against ->cpus_allowed
3417 */
3418 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3419 likely(cpu_active(dest_cpu)) && migrate_task(p, rq)) {
3420 struct migration_arg arg = { p, dest_cpu };
3421
3422 task_rq_unlock(rq, &flags);
3423 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3424 return;
3425 }
3426 unlock:
3427 task_rq_unlock(rq, &flags);
3428 }
3429
3430 #endif
3431
3432 DEFINE_PER_CPU(struct kernel_stat, kstat);
3433
3434 EXPORT_PER_CPU_SYMBOL(kstat);
3435
3436 /*
3437 * Return any ns on the sched_clock that have not yet been accounted in
3438 * @p in case that task is currently running.
3439 *
3440 * Called with task_rq_lock() held on @rq.
3441 */
3442 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3443 {
3444 u64 ns = 0;
3445
3446 if (task_current(rq, p)) {
3447 update_rq_clock(rq);
3448 ns = rq->clock_task - p->se.exec_start;
3449 if ((s64)ns < 0)
3450 ns = 0;
3451 }
3452
3453 return ns;
3454 }
3455
3456 unsigned long long task_delta_exec(struct task_struct *p)
3457 {
3458 unsigned long flags;
3459 struct rq *rq;
3460 u64 ns = 0;
3461
3462 rq = task_rq_lock(p, &flags);
3463 ns = do_task_delta_exec(p, rq);
3464 task_rq_unlock(rq, &flags);
3465
3466 return ns;
3467 }
3468
3469 /*
3470 * Return accounted runtime for the task.
3471 * In case the task is currently running, return the runtime plus current's
3472 * pending runtime that have not been accounted yet.
3473 */
3474 unsigned long long task_sched_runtime(struct task_struct *p)
3475 {
3476 unsigned long flags;
3477 struct rq *rq;
3478 u64 ns = 0;
3479
3480 rq = task_rq_lock(p, &flags);
3481 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3482 task_rq_unlock(rq, &flags);
3483
3484 return ns;
3485 }
3486
3487 /*
3488 * Return sum_exec_runtime for the thread group.
3489 * In case the task is currently running, return the sum plus current's
3490 * pending runtime that have not been accounted yet.
3491 *
3492 * Note that the thread group might have other running tasks as well,
3493 * so the return value not includes other pending runtime that other
3494 * running tasks might have.
3495 */
3496 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3497 {
3498 struct task_cputime totals;
3499 unsigned long flags;
3500 struct rq *rq;
3501 u64 ns;
3502
3503 rq = task_rq_lock(p, &flags);
3504 thread_group_cputime(p, &totals);
3505 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3506 task_rq_unlock(rq, &flags);
3507
3508 return ns;
3509 }
3510
3511 /*
3512 * Account user cpu time to a process.
3513 * @p: the process that the cpu time gets accounted to
3514 * @cputime: the cpu time spent in user space since the last update
3515 * @cputime_scaled: cputime scaled by cpu frequency
3516 */
3517 void account_user_time(struct task_struct *p, cputime_t cputime,
3518 cputime_t cputime_scaled)
3519 {
3520 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3521 cputime64_t tmp;
3522
3523 /* Add user time to process. */
3524 p->utime = cputime_add(p->utime, cputime);
3525 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3526 account_group_user_time(p, cputime);
3527
3528 /* Add user time to cpustat. */
3529 tmp = cputime_to_cputime64(cputime);
3530 if (TASK_NICE(p) > 0)
3531 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3532 else
3533 cpustat->user = cputime64_add(cpustat->user, tmp);
3534
3535 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3536 /* Account for user time used */
3537 acct_update_integrals(p);
3538 }
3539
3540 /*
3541 * Account guest cpu time to a process.
3542 * @p: the process that the cpu time gets accounted to
3543 * @cputime: the cpu time spent in virtual machine since the last update
3544 * @cputime_scaled: cputime scaled by cpu frequency
3545 */
3546 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3547 cputime_t cputime_scaled)
3548 {
3549 cputime64_t tmp;
3550 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3551
3552 tmp = cputime_to_cputime64(cputime);
3553
3554 /* Add guest time to process. */
3555 p->utime = cputime_add(p->utime, cputime);
3556 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3557 account_group_user_time(p, cputime);
3558 p->gtime = cputime_add(p->gtime, cputime);
3559
3560 /* Add guest time to cpustat. */
3561 if (TASK_NICE(p) > 0) {
3562 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3563 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3564 } else {
3565 cpustat->user = cputime64_add(cpustat->user, tmp);
3566 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3567 }
3568 }
3569
3570 /*
3571 * Account system cpu time to a process.
3572 * @p: the process that the cpu time gets accounted to
3573 * @hardirq_offset: the offset to subtract from hardirq_count()
3574 * @cputime: the cpu time spent in kernel space since the last update
3575 * @cputime_scaled: cputime scaled by cpu frequency
3576 */
3577 void account_system_time(struct task_struct *p, int hardirq_offset,
3578 cputime_t cputime, cputime_t cputime_scaled)
3579 {
3580 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3581 cputime64_t tmp;
3582
3583 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3584 account_guest_time(p, cputime, cputime_scaled);
3585 return;
3586 }
3587
3588 /* Add system time to process. */
3589 p->stime = cputime_add(p->stime, cputime);
3590 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3591 account_group_system_time(p, cputime);
3592
3593 /* Add system time to cpustat. */
3594 tmp = cputime_to_cputime64(cputime);
3595 if (hardirq_count() - hardirq_offset)
3596 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3597 else if (in_serving_softirq())
3598 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3599 else
3600 cpustat->system = cputime64_add(cpustat->system, tmp);
3601
3602 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3603
3604 /* Account for system time used */
3605 acct_update_integrals(p);
3606 }
3607
3608 /*
3609 * Account for involuntary wait time.
3610 * @steal: the cpu time spent in involuntary wait
3611 */
3612 void account_steal_time(cputime_t cputime)
3613 {
3614 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3615 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3616
3617 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3618 }
3619
3620 /*
3621 * Account for idle time.
3622 * @cputime: the cpu time spent in idle wait
3623 */
3624 void account_idle_time(cputime_t cputime)
3625 {
3626 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3627 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3628 struct rq *rq = this_rq();
3629
3630 if (atomic_read(&rq->nr_iowait) > 0)
3631 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3632 else
3633 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3634 }
3635
3636 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3637
3638 /*
3639 * Account a single tick of cpu time.
3640 * @p: the process that the cpu time gets accounted to
3641 * @user_tick: indicates if the tick is a user or a system tick
3642 */
3643 void account_process_tick(struct task_struct *p, int user_tick)
3644 {
3645 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3646 struct rq *rq = this_rq();
3647
3648 if (user_tick)
3649 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3650 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3651 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3652 one_jiffy_scaled);
3653 else
3654 account_idle_time(cputime_one_jiffy);
3655 }
3656
3657 /*
3658 * Account multiple ticks of steal time.
3659 * @p: the process from which the cpu time has been stolen
3660 * @ticks: number of stolen ticks
3661 */
3662 void account_steal_ticks(unsigned long ticks)
3663 {
3664 account_steal_time(jiffies_to_cputime(ticks));
3665 }
3666
3667 /*
3668 * Account multiple ticks of idle time.
3669 * @ticks: number of stolen ticks
3670 */
3671 void account_idle_ticks(unsigned long ticks)
3672 {
3673 account_idle_time(jiffies_to_cputime(ticks));
3674 }
3675
3676 #endif
3677
3678 /*
3679 * Use precise platform statistics if available:
3680 */
3681 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3682 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3683 {
3684 *ut = p->utime;
3685 *st = p->stime;
3686 }
3687
3688 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3689 {
3690 struct task_cputime cputime;
3691
3692 thread_group_cputime(p, &cputime);
3693
3694 *ut = cputime.utime;
3695 *st = cputime.stime;
3696 }
3697 #else
3698
3699 #ifndef nsecs_to_cputime
3700 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3701 #endif
3702
3703 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3704 {
3705 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3706
3707 /*
3708 * Use CFS's precise accounting:
3709 */
3710 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3711
3712 if (total) {
3713 u64 temp = rtime;
3714
3715 temp *= utime;
3716 do_div(temp, total);
3717 utime = (cputime_t)temp;
3718 } else
3719 utime = rtime;
3720
3721 /*
3722 * Compare with previous values, to keep monotonicity:
3723 */
3724 p->prev_utime = max(p->prev_utime, utime);
3725 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3726
3727 *ut = p->prev_utime;
3728 *st = p->prev_stime;
3729 }
3730
3731 /*
3732 * Must be called with siglock held.
3733 */
3734 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3735 {
3736 struct signal_struct *sig = p->signal;
3737 struct task_cputime cputime;
3738 cputime_t rtime, utime, total;
3739
3740 thread_group_cputime(p, &cputime);
3741
3742 total = cputime_add(cputime.utime, cputime.stime);
3743 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3744
3745 if (total) {
3746 u64 temp = rtime;
3747
3748 temp *= cputime.utime;
3749 do_div(temp, total);
3750 utime = (cputime_t)temp;
3751 } else
3752 utime = rtime;
3753
3754 sig->prev_utime = max(sig->prev_utime, utime);
3755 sig->prev_stime = max(sig->prev_stime,
3756 cputime_sub(rtime, sig->prev_utime));
3757
3758 *ut = sig->prev_utime;
3759 *st = sig->prev_stime;
3760 }
3761 #endif
3762
3763 /*
3764 * This function gets called by the timer code, with HZ frequency.
3765 * We call it with interrupts disabled.
3766 *
3767 * It also gets called by the fork code, when changing the parent's
3768 * timeslices.
3769 */
3770 void scheduler_tick(void)
3771 {
3772 int cpu = smp_processor_id();
3773 struct rq *rq = cpu_rq(cpu);
3774 struct task_struct *curr = rq->curr;
3775
3776 sched_clock_tick();
3777
3778 raw_spin_lock(&rq->lock);
3779 update_rq_clock(rq);
3780 update_cpu_load_active(rq);
3781 curr->sched_class->task_tick(rq, curr, 0);
3782 raw_spin_unlock(&rq->lock);
3783
3784 perf_event_task_tick();
3785
3786 #ifdef CONFIG_SMP
3787 rq->idle_at_tick = idle_cpu(cpu);
3788 trigger_load_balance(rq, cpu);
3789 #endif
3790 }
3791
3792 notrace unsigned long get_parent_ip(unsigned long addr)
3793 {
3794 if (in_lock_functions(addr)) {
3795 addr = CALLER_ADDR2;
3796 if (in_lock_functions(addr))
3797 addr = CALLER_ADDR3;
3798 }
3799 return addr;
3800 }
3801
3802 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3803 defined(CONFIG_PREEMPT_TRACER))
3804
3805 void __kprobes add_preempt_count(int val)
3806 {
3807 #ifdef CONFIG_DEBUG_PREEMPT
3808 /*
3809 * Underflow?
3810 */
3811 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3812 return;
3813 #endif
3814 preempt_count() += val;
3815 #ifdef CONFIG_DEBUG_PREEMPT
3816 /*
3817 * Spinlock count overflowing soon?
3818 */
3819 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3820 PREEMPT_MASK - 10);
3821 #endif
3822 if (preempt_count() == val)
3823 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3824 }
3825 EXPORT_SYMBOL(add_preempt_count);
3826
3827 void __kprobes sub_preempt_count(int val)
3828 {
3829 #ifdef CONFIG_DEBUG_PREEMPT
3830 /*
3831 * Underflow?
3832 */
3833 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3834 return;
3835 /*
3836 * Is the spinlock portion underflowing?
3837 */
3838 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3839 !(preempt_count() & PREEMPT_MASK)))
3840 return;
3841 #endif
3842
3843 if (preempt_count() == val)
3844 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3845 preempt_count() -= val;
3846 }
3847 EXPORT_SYMBOL(sub_preempt_count);
3848
3849 #endif
3850
3851 /*
3852 * Print scheduling while atomic bug:
3853 */
3854 static noinline void __schedule_bug(struct task_struct *prev)
3855 {
3856 struct pt_regs *regs = get_irq_regs();
3857
3858 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3859 prev->comm, prev->pid, preempt_count());
3860
3861 debug_show_held_locks(prev);
3862 print_modules();
3863 if (irqs_disabled())
3864 print_irqtrace_events(prev);
3865
3866 if (regs)
3867 show_regs(regs);
3868 else
3869 dump_stack();
3870 }
3871
3872 /*
3873 * Various schedule()-time debugging checks and statistics:
3874 */
3875 static inline void schedule_debug(struct task_struct *prev)
3876 {
3877 /*
3878 * Test if we are atomic. Since do_exit() needs to call into
3879 * schedule() atomically, we ignore that path for now.
3880 * Otherwise, whine if we are scheduling when we should not be.
3881 */
3882 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3883 __schedule_bug(prev);
3884
3885 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3886
3887 schedstat_inc(this_rq(), sched_count);
3888 #ifdef CONFIG_SCHEDSTATS
3889 if (unlikely(prev->lock_depth >= 0)) {
3890 schedstat_inc(this_rq(), rq_sched_info.bkl_count);
3891 schedstat_inc(prev, sched_info.bkl_count);
3892 }
3893 #endif
3894 }
3895
3896 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3897 {
3898 if (prev->se.on_rq)
3899 update_rq_clock(rq);
3900 prev->sched_class->put_prev_task(rq, prev);
3901 }
3902
3903 /*
3904 * Pick up the highest-prio task:
3905 */
3906 static inline struct task_struct *
3907 pick_next_task(struct rq *rq)
3908 {
3909 const struct sched_class *class;
3910 struct task_struct *p;
3911
3912 /*
3913 * Optimization: we know that if all tasks are in
3914 * the fair class we can call that function directly:
3915 */
3916 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3917 p = fair_sched_class.pick_next_task(rq);
3918 if (likely(p))
3919 return p;
3920 }
3921
3922 for_each_class(class) {
3923 p = class->pick_next_task(rq);
3924 if (p)
3925 return p;
3926 }
3927
3928 BUG(); /* the idle class will always have a runnable task */
3929 }
3930
3931 /*
3932 * schedule() is the main scheduler function.
3933 */
3934 asmlinkage void __sched schedule(void)
3935 {
3936 struct task_struct *prev, *next;
3937 unsigned long *switch_count;
3938 struct rq *rq;
3939 int cpu;
3940
3941 need_resched:
3942 preempt_disable();
3943 cpu = smp_processor_id();
3944 rq = cpu_rq(cpu);
3945 rcu_note_context_switch(cpu);
3946 prev = rq->curr;
3947
3948 release_kernel_lock(prev);
3949 need_resched_nonpreemptible:
3950
3951 schedule_debug(prev);
3952
3953 if (sched_feat(HRTICK))
3954 hrtick_clear(rq);
3955
3956 raw_spin_lock_irq(&rq->lock);
3957
3958 switch_count = &prev->nivcsw;
3959 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3960 if (unlikely(signal_pending_state(prev->state, prev))) {
3961 prev->state = TASK_RUNNING;
3962 } else {
3963 /*
3964 * If a worker is going to sleep, notify and
3965 * ask workqueue whether it wants to wake up a
3966 * task to maintain concurrency. If so, wake
3967 * up the task.
3968 */
3969 if (prev->flags & PF_WQ_WORKER) {
3970 struct task_struct *to_wakeup;
3971
3972 to_wakeup = wq_worker_sleeping(prev, cpu);
3973 if (to_wakeup)
3974 try_to_wake_up_local(to_wakeup);
3975 }
3976 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3977 }
3978 switch_count = &prev->nvcsw;
3979 }
3980
3981 pre_schedule(rq, prev);
3982
3983 if (unlikely(!rq->nr_running))
3984 idle_balance(cpu, rq);
3985
3986 put_prev_task(rq, prev);
3987 next = pick_next_task(rq);
3988 clear_tsk_need_resched(prev);
3989 rq->skip_clock_update = 0;
3990
3991 if (likely(prev != next)) {
3992 sched_info_switch(prev, next);
3993 perf_event_task_sched_out(prev, next);
3994
3995 rq->nr_switches++;
3996 rq->curr = next;
3997 ++*switch_count;
3998
3999 context_switch(rq, prev, next); /* unlocks the rq */
4000 /*
4001 * The context switch have flipped the stack from under us
4002 * and restored the local variables which were saved when
4003 * this task called schedule() in the past. prev == current
4004 * is still correct, but it can be moved to another cpu/rq.
4005 */
4006 cpu = smp_processor_id();
4007 rq = cpu_rq(cpu);
4008 } else
4009 raw_spin_unlock_irq(&rq->lock);
4010
4011 post_schedule(rq);
4012
4013 if (unlikely(reacquire_kernel_lock(prev)))
4014 goto need_resched_nonpreemptible;
4015
4016 preempt_enable_no_resched();
4017 if (need_resched())
4018 goto need_resched;
4019 }
4020 EXPORT_SYMBOL(schedule);
4021
4022 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4023 /*
4024 * Look out! "owner" is an entirely speculative pointer
4025 * access and not reliable.
4026 */
4027 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
4028 {
4029 unsigned int cpu;
4030 struct rq *rq;
4031
4032 if (!sched_feat(OWNER_SPIN))
4033 return 0;
4034
4035 #ifdef CONFIG_DEBUG_PAGEALLOC
4036 /*
4037 * Need to access the cpu field knowing that
4038 * DEBUG_PAGEALLOC could have unmapped it if
4039 * the mutex owner just released it and exited.
4040 */
4041 if (probe_kernel_address(&owner->cpu, cpu))
4042 return 0;
4043 #else
4044 cpu = owner->cpu;
4045 #endif
4046
4047 /*
4048 * Even if the access succeeded (likely case),
4049 * the cpu field may no longer be valid.
4050 */
4051 if (cpu >= nr_cpumask_bits)
4052 return 0;
4053
4054 /*
4055 * We need to validate that we can do a
4056 * get_cpu() and that we have the percpu area.
4057 */
4058 if (!cpu_online(cpu))
4059 return 0;
4060
4061 rq = cpu_rq(cpu);
4062
4063 for (;;) {
4064 /*
4065 * Owner changed, break to re-assess state.
4066 */
4067 if (lock->owner != owner) {
4068 /*
4069 * If the lock has switched to a different owner,
4070 * we likely have heavy contention. Return 0 to quit
4071 * optimistic spinning and not contend further:
4072 */
4073 if (lock->owner)
4074 return 0;
4075 break;
4076 }
4077
4078 /*
4079 * Is that owner really running on that cpu?
4080 */
4081 if (task_thread_info(rq->curr) != owner || need_resched())
4082 return 0;
4083
4084 arch_mutex_cpu_relax();
4085 }
4086
4087 return 1;
4088 }
4089 #endif
4090
4091 #ifdef CONFIG_PREEMPT
4092 /*
4093 * this is the entry point to schedule() from in-kernel preemption
4094 * off of preempt_enable. Kernel preemptions off return from interrupt
4095 * occur there and call schedule directly.
4096 */
4097 asmlinkage void __sched notrace preempt_schedule(void)
4098 {
4099 struct thread_info *ti = current_thread_info();
4100
4101 /*
4102 * If there is a non-zero preempt_count or interrupts are disabled,
4103 * we do not want to preempt the current task. Just return..
4104 */
4105 if (likely(ti->preempt_count || irqs_disabled()))
4106 return;
4107
4108 do {
4109 add_preempt_count_notrace(PREEMPT_ACTIVE);
4110 schedule();
4111 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4112
4113 /*
4114 * Check again in case we missed a preemption opportunity
4115 * between schedule and now.
4116 */
4117 barrier();
4118 } while (need_resched());
4119 }
4120 EXPORT_SYMBOL(preempt_schedule);
4121
4122 /*
4123 * this is the entry point to schedule() from kernel preemption
4124 * off of irq context.
4125 * Note, that this is called and return with irqs disabled. This will
4126 * protect us against recursive calling from irq.
4127 */
4128 asmlinkage void __sched preempt_schedule_irq(void)
4129 {
4130 struct thread_info *ti = current_thread_info();
4131
4132 /* Catch callers which need to be fixed */
4133 BUG_ON(ti->preempt_count || !irqs_disabled());
4134
4135 do {
4136 add_preempt_count(PREEMPT_ACTIVE);
4137 local_irq_enable();
4138 schedule();
4139 local_irq_disable();
4140 sub_preempt_count(PREEMPT_ACTIVE);
4141
4142 /*
4143 * Check again in case we missed a preemption opportunity
4144 * between schedule and now.
4145 */
4146 barrier();
4147 } while (need_resched());
4148 }
4149
4150 #endif /* CONFIG_PREEMPT */
4151
4152 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4153 void *key)
4154 {
4155 return try_to_wake_up(curr->private, mode, wake_flags);
4156 }
4157 EXPORT_SYMBOL(default_wake_function);
4158
4159 /*
4160 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4161 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4162 * number) then we wake all the non-exclusive tasks and one exclusive task.
4163 *
4164 * There are circumstances in which we can try to wake a task which has already
4165 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4166 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4167 */
4168 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4169 int nr_exclusive, int wake_flags, void *key)
4170 {
4171 wait_queue_t *curr, *next;
4172
4173 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4174 unsigned flags = curr->flags;
4175
4176 if (curr->func(curr, mode, wake_flags, key) &&
4177 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4178 break;
4179 }
4180 }
4181
4182 /**
4183 * __wake_up - wake up threads blocked on a waitqueue.
4184 * @q: the waitqueue
4185 * @mode: which threads
4186 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4187 * @key: is directly passed to the wakeup function
4188 *
4189 * It may be assumed that this function implies a write memory barrier before
4190 * changing the task state if and only if any tasks are woken up.
4191 */
4192 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4193 int nr_exclusive, void *key)
4194 {
4195 unsigned long flags;
4196
4197 spin_lock_irqsave(&q->lock, flags);
4198 __wake_up_common(q, mode, nr_exclusive, 0, key);
4199 spin_unlock_irqrestore(&q->lock, flags);
4200 }
4201 EXPORT_SYMBOL(__wake_up);
4202
4203 /*
4204 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4205 */
4206 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4207 {
4208 __wake_up_common(q, mode, 1, 0, NULL);
4209 }
4210 EXPORT_SYMBOL_GPL(__wake_up_locked);
4211
4212 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4213 {
4214 __wake_up_common(q, mode, 1, 0, key);
4215 }
4216 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4217
4218 /**
4219 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4220 * @q: the waitqueue
4221 * @mode: which threads
4222 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4223 * @key: opaque value to be passed to wakeup targets
4224 *
4225 * The sync wakeup differs that the waker knows that it will schedule
4226 * away soon, so while the target thread will be woken up, it will not
4227 * be migrated to another CPU - ie. the two threads are 'synchronized'
4228 * with each other. This can prevent needless bouncing between CPUs.
4229 *
4230 * On UP it can prevent extra preemption.
4231 *
4232 * It may be assumed that this function implies a write memory barrier before
4233 * changing the task state if and only if any tasks are woken up.
4234 */
4235 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4236 int nr_exclusive, void *key)
4237 {
4238 unsigned long flags;
4239 int wake_flags = WF_SYNC;
4240
4241 if (unlikely(!q))
4242 return;
4243
4244 if (unlikely(!nr_exclusive))
4245 wake_flags = 0;
4246
4247 spin_lock_irqsave(&q->lock, flags);
4248 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4249 spin_unlock_irqrestore(&q->lock, flags);
4250 }
4251 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4252
4253 /*
4254 * __wake_up_sync - see __wake_up_sync_key()
4255 */
4256 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4257 {
4258 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4259 }
4260 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4261
4262 /**
4263 * complete: - signals a single thread waiting on this completion
4264 * @x: holds the state of this particular completion
4265 *
4266 * This will wake up a single thread waiting on this completion. Threads will be
4267 * awakened in the same order in which they were queued.
4268 *
4269 * See also complete_all(), wait_for_completion() and related routines.
4270 *
4271 * It may be assumed that this function implies a write memory barrier before
4272 * changing the task state if and only if any tasks are woken up.
4273 */
4274 void complete(struct completion *x)
4275 {
4276 unsigned long flags;
4277
4278 spin_lock_irqsave(&x->wait.lock, flags);
4279 x->done++;
4280 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4281 spin_unlock_irqrestore(&x->wait.lock, flags);
4282 }
4283 EXPORT_SYMBOL(complete);
4284
4285 /**
4286 * complete_all: - signals all threads waiting on this completion
4287 * @x: holds the state of this particular completion
4288 *
4289 * This will wake up all threads waiting on this particular completion event.
4290 *
4291 * It may be assumed that this function implies a write memory barrier before
4292 * changing the task state if and only if any tasks are woken up.
4293 */
4294 void complete_all(struct completion *x)
4295 {
4296 unsigned long flags;
4297
4298 spin_lock_irqsave(&x->wait.lock, flags);
4299 x->done += UINT_MAX/2;
4300 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4301 spin_unlock_irqrestore(&x->wait.lock, flags);
4302 }
4303 EXPORT_SYMBOL(complete_all);
4304
4305 static inline long __sched
4306 do_wait_for_common(struct completion *x, long timeout, int state)
4307 {
4308 if (!x->done) {
4309 DECLARE_WAITQUEUE(wait, current);
4310
4311 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4312 do {
4313 if (signal_pending_state(state, current)) {
4314 timeout = -ERESTARTSYS;
4315 break;
4316 }
4317 __set_current_state(state);
4318 spin_unlock_irq(&x->wait.lock);
4319 timeout = schedule_timeout(timeout);
4320 spin_lock_irq(&x->wait.lock);
4321 } while (!x->done && timeout);
4322 __remove_wait_queue(&x->wait, &wait);
4323 if (!x->done)
4324 return timeout;
4325 }
4326 x->done--;
4327 return timeout ?: 1;
4328 }
4329
4330 static long __sched
4331 wait_for_common(struct completion *x, long timeout, int state)
4332 {
4333 might_sleep();
4334
4335 spin_lock_irq(&x->wait.lock);
4336 timeout = do_wait_for_common(x, timeout, state);
4337 spin_unlock_irq(&x->wait.lock);
4338 return timeout;
4339 }
4340
4341 /**
4342 * wait_for_completion: - waits for completion of a task
4343 * @x: holds the state of this particular completion
4344 *
4345 * This waits to be signaled for completion of a specific task. It is NOT
4346 * interruptible and there is no timeout.
4347 *
4348 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4349 * and interrupt capability. Also see complete().
4350 */
4351 void __sched wait_for_completion(struct completion *x)
4352 {
4353 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4354 }
4355 EXPORT_SYMBOL(wait_for_completion);
4356
4357 /**
4358 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4359 * @x: holds the state of this particular completion
4360 * @timeout: timeout value in jiffies
4361 *
4362 * This waits for either a completion of a specific task to be signaled or for a
4363 * specified timeout to expire. The timeout is in jiffies. It is not
4364 * interruptible.
4365 */
4366 unsigned long __sched
4367 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4368 {
4369 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4370 }
4371 EXPORT_SYMBOL(wait_for_completion_timeout);
4372
4373 /**
4374 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4375 * @x: holds the state of this particular completion
4376 *
4377 * This waits for completion of a specific task to be signaled. It is
4378 * interruptible.
4379 */
4380 int __sched wait_for_completion_interruptible(struct completion *x)
4381 {
4382 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4383 if (t == -ERESTARTSYS)
4384 return t;
4385 return 0;
4386 }
4387 EXPORT_SYMBOL(wait_for_completion_interruptible);
4388
4389 /**
4390 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4391 * @x: holds the state of this particular completion
4392 * @timeout: timeout value in jiffies
4393 *
4394 * This waits for either a completion of a specific task to be signaled or for a
4395 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4396 */
4397 long __sched
4398 wait_for_completion_interruptible_timeout(struct completion *x,
4399 unsigned long timeout)
4400 {
4401 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4402 }
4403 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4404
4405 /**
4406 * wait_for_completion_killable: - waits for completion of a task (killable)
4407 * @x: holds the state of this particular completion
4408 *
4409 * This waits to be signaled for completion of a specific task. It can be
4410 * interrupted by a kill signal.
4411 */
4412 int __sched wait_for_completion_killable(struct completion *x)
4413 {
4414 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4415 if (t == -ERESTARTSYS)
4416 return t;
4417 return 0;
4418 }
4419 EXPORT_SYMBOL(wait_for_completion_killable);
4420
4421 /**
4422 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4423 * @x: holds the state of this particular completion
4424 * @timeout: timeout value in jiffies
4425 *
4426 * This waits for either a completion of a specific task to be
4427 * signaled or for a specified timeout to expire. It can be
4428 * interrupted by a kill signal. The timeout is in jiffies.
4429 */
4430 long __sched
4431 wait_for_completion_killable_timeout(struct completion *x,
4432 unsigned long timeout)
4433 {
4434 return wait_for_common(x, timeout, TASK_KILLABLE);
4435 }
4436 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4437
4438 /**
4439 * try_wait_for_completion - try to decrement a completion without blocking
4440 * @x: completion structure
4441 *
4442 * Returns: 0 if a decrement cannot be done without blocking
4443 * 1 if a decrement succeeded.
4444 *
4445 * If a completion is being used as a counting completion,
4446 * attempt to decrement the counter without blocking. This
4447 * enables us to avoid waiting if the resource the completion
4448 * is protecting is not available.
4449 */
4450 bool try_wait_for_completion(struct completion *x)
4451 {
4452 unsigned long flags;
4453 int ret = 1;
4454
4455 spin_lock_irqsave(&x->wait.lock, flags);
4456 if (!x->done)
4457 ret = 0;
4458 else
4459 x->done--;
4460 spin_unlock_irqrestore(&x->wait.lock, flags);
4461 return ret;
4462 }
4463 EXPORT_SYMBOL(try_wait_for_completion);
4464
4465 /**
4466 * completion_done - Test to see if a completion has any waiters
4467 * @x: completion structure
4468 *
4469 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4470 * 1 if there are no waiters.
4471 *
4472 */
4473 bool completion_done(struct completion *x)
4474 {
4475 unsigned long flags;
4476 int ret = 1;
4477
4478 spin_lock_irqsave(&x->wait.lock, flags);
4479 if (!x->done)
4480 ret = 0;
4481 spin_unlock_irqrestore(&x->wait.lock, flags);
4482 return ret;
4483 }
4484 EXPORT_SYMBOL(completion_done);
4485
4486 static long __sched
4487 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4488 {
4489 unsigned long flags;
4490 wait_queue_t wait;
4491
4492 init_waitqueue_entry(&wait, current);
4493
4494 __set_current_state(state);
4495
4496 spin_lock_irqsave(&q->lock, flags);
4497 __add_wait_queue(q, &wait);
4498 spin_unlock(&q->lock);
4499 timeout = schedule_timeout(timeout);
4500 spin_lock_irq(&q->lock);
4501 __remove_wait_queue(q, &wait);
4502 spin_unlock_irqrestore(&q->lock, flags);
4503
4504 return timeout;
4505 }
4506
4507 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4508 {
4509 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4510 }
4511 EXPORT_SYMBOL(interruptible_sleep_on);
4512
4513 long __sched
4514 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4515 {
4516 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4517 }
4518 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4519
4520 void __sched sleep_on(wait_queue_head_t *q)
4521 {
4522 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4523 }
4524 EXPORT_SYMBOL(sleep_on);
4525
4526 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4527 {
4528 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4529 }
4530 EXPORT_SYMBOL(sleep_on_timeout);
4531
4532 #ifdef CONFIG_RT_MUTEXES
4533
4534 /*
4535 * rt_mutex_setprio - set the current priority of a task
4536 * @p: task
4537 * @prio: prio value (kernel-internal form)
4538 *
4539 * This function changes the 'effective' priority of a task. It does
4540 * not touch ->normal_prio like __setscheduler().
4541 *
4542 * Used by the rt_mutex code to implement priority inheritance logic.
4543 */
4544 void rt_mutex_setprio(struct task_struct *p, int prio)
4545 {
4546 unsigned long flags;
4547 int oldprio, on_rq, running;
4548 struct rq *rq;
4549 const struct sched_class *prev_class;
4550
4551 BUG_ON(prio < 0 || prio > MAX_PRIO);
4552
4553 rq = task_rq_lock(p, &flags);
4554
4555 trace_sched_pi_setprio(p, prio);
4556 oldprio = p->prio;
4557 prev_class = p->sched_class;
4558 on_rq = p->se.on_rq;
4559 running = task_current(rq, p);
4560 if (on_rq)
4561 dequeue_task(rq, p, 0);
4562 if (running)
4563 p->sched_class->put_prev_task(rq, p);
4564
4565 if (rt_prio(prio))
4566 p->sched_class = &rt_sched_class;
4567 else
4568 p->sched_class = &fair_sched_class;
4569
4570 p->prio = prio;
4571
4572 if (running)
4573 p->sched_class->set_curr_task(rq);
4574 if (on_rq) {
4575 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4576
4577 check_class_changed(rq, p, prev_class, oldprio, running);
4578 }
4579 task_rq_unlock(rq, &flags);
4580 }
4581
4582 #endif
4583
4584 void set_user_nice(struct task_struct *p, long nice)
4585 {
4586 int old_prio, delta, on_rq;
4587 unsigned long flags;
4588 struct rq *rq;
4589
4590 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4591 return;
4592 /*
4593 * We have to be careful, if called from sys_setpriority(),
4594 * the task might be in the middle of scheduling on another CPU.
4595 */
4596 rq = task_rq_lock(p, &flags);
4597 /*
4598 * The RT priorities are set via sched_setscheduler(), but we still
4599 * allow the 'normal' nice value to be set - but as expected
4600 * it wont have any effect on scheduling until the task is
4601 * SCHED_FIFO/SCHED_RR:
4602 */
4603 if (task_has_rt_policy(p)) {
4604 p->static_prio = NICE_TO_PRIO(nice);
4605 goto out_unlock;
4606 }
4607 on_rq = p->se.on_rq;
4608 if (on_rq)
4609 dequeue_task(rq, p, 0);
4610
4611 p->static_prio = NICE_TO_PRIO(nice);
4612 set_load_weight(p);
4613 old_prio = p->prio;
4614 p->prio = effective_prio(p);
4615 delta = p->prio - old_prio;
4616
4617 if (on_rq) {
4618 enqueue_task(rq, p, 0);
4619 /*
4620 * If the task increased its priority or is running and
4621 * lowered its priority, then reschedule its CPU:
4622 */
4623 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4624 resched_task(rq->curr);
4625 }
4626 out_unlock:
4627 task_rq_unlock(rq, &flags);
4628 }
4629 EXPORT_SYMBOL(set_user_nice);
4630
4631 /*
4632 * can_nice - check if a task can reduce its nice value
4633 * @p: task
4634 * @nice: nice value
4635 */
4636 int can_nice(const struct task_struct *p, const int nice)
4637 {
4638 /* convert nice value [19,-20] to rlimit style value [1,40] */
4639 int nice_rlim = 20 - nice;
4640
4641 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4642 capable(CAP_SYS_NICE));
4643 }
4644
4645 #ifdef __ARCH_WANT_SYS_NICE
4646
4647 /*
4648 * sys_nice - change the priority of the current process.
4649 * @increment: priority increment
4650 *
4651 * sys_setpriority is a more generic, but much slower function that
4652 * does similar things.
4653 */
4654 SYSCALL_DEFINE1(nice, int, increment)
4655 {
4656 long nice, retval;
4657
4658 /*
4659 * Setpriority might change our priority at the same moment.
4660 * We don't have to worry. Conceptually one call occurs first
4661 * and we have a single winner.
4662 */
4663 if (increment < -40)
4664 increment = -40;
4665 if (increment > 40)
4666 increment = 40;
4667
4668 nice = TASK_NICE(current) + increment;
4669 if (nice < -20)
4670 nice = -20;
4671 if (nice > 19)
4672 nice = 19;
4673
4674 if (increment < 0 && !can_nice(current, nice))
4675 return -EPERM;
4676
4677 retval = security_task_setnice(current, nice);
4678 if (retval)
4679 return retval;
4680
4681 set_user_nice(current, nice);
4682 return 0;
4683 }
4684
4685 #endif
4686
4687 /**
4688 * task_prio - return the priority value of a given task.
4689 * @p: the task in question.
4690 *
4691 * This is the priority value as seen by users in /proc.
4692 * RT tasks are offset by -200. Normal tasks are centered
4693 * around 0, value goes from -16 to +15.
4694 */
4695 int task_prio(const struct task_struct *p)
4696 {
4697 return p->prio - MAX_RT_PRIO;
4698 }
4699
4700 /**
4701 * task_nice - return the nice value of a given task.
4702 * @p: the task in question.
4703 */
4704 int task_nice(const struct task_struct *p)
4705 {
4706 return TASK_NICE(p);
4707 }
4708 EXPORT_SYMBOL(task_nice);
4709
4710 /**
4711 * idle_cpu - is a given cpu idle currently?
4712 * @cpu: the processor in question.
4713 */
4714 int idle_cpu(int cpu)
4715 {
4716 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4717 }
4718
4719 /**
4720 * idle_task - return the idle task for a given cpu.
4721 * @cpu: the processor in question.
4722 */
4723 struct task_struct *idle_task(int cpu)
4724 {
4725 return cpu_rq(cpu)->idle;
4726 }
4727
4728 /**
4729 * find_process_by_pid - find a process with a matching PID value.
4730 * @pid: the pid in question.
4731 */
4732 static struct task_struct *find_process_by_pid(pid_t pid)
4733 {
4734 return pid ? find_task_by_vpid(pid) : current;
4735 }
4736
4737 /* Actually do priority change: must hold rq lock. */
4738 static void
4739 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4740 {
4741 BUG_ON(p->se.on_rq);
4742
4743 p->policy = policy;
4744 p->rt_priority = prio;
4745 p->normal_prio = normal_prio(p);
4746 /* we are holding p->pi_lock already */
4747 p->prio = rt_mutex_getprio(p);
4748 if (rt_prio(p->prio))
4749 p->sched_class = &rt_sched_class;
4750 else
4751 p->sched_class = &fair_sched_class;
4752 set_load_weight(p);
4753 }
4754
4755 /*
4756 * check the target process has a UID that matches the current process's
4757 */
4758 static bool check_same_owner(struct task_struct *p)
4759 {
4760 const struct cred *cred = current_cred(), *pcred;
4761 bool match;
4762
4763 rcu_read_lock();
4764 pcred = __task_cred(p);
4765 match = (cred->euid == pcred->euid ||
4766 cred->euid == pcred->uid);
4767 rcu_read_unlock();
4768 return match;
4769 }
4770
4771 static int __sched_setscheduler(struct task_struct *p, int policy,
4772 const struct sched_param *param, bool user)
4773 {
4774 int retval, oldprio, oldpolicy = -1, on_rq, running;
4775 unsigned long flags;
4776 const struct sched_class *prev_class;
4777 struct rq *rq;
4778 int reset_on_fork;
4779
4780 /* may grab non-irq protected spin_locks */
4781 BUG_ON(in_interrupt());
4782 recheck:
4783 /* double check policy once rq lock held */
4784 if (policy < 0) {
4785 reset_on_fork = p->sched_reset_on_fork;
4786 policy = oldpolicy = p->policy;
4787 } else {
4788 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4789 policy &= ~SCHED_RESET_ON_FORK;
4790
4791 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4792 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4793 policy != SCHED_IDLE)
4794 return -EINVAL;
4795 }
4796
4797 /*
4798 * Valid priorities for SCHED_FIFO and SCHED_RR are
4799 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4800 * SCHED_BATCH and SCHED_IDLE is 0.
4801 */
4802 if (param->sched_priority < 0 ||
4803 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4804 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4805 return -EINVAL;
4806 if (rt_policy(policy) != (param->sched_priority != 0))
4807 return -EINVAL;
4808
4809 /*
4810 * Allow unprivileged RT tasks to decrease priority:
4811 */
4812 if (user && !capable(CAP_SYS_NICE)) {
4813 if (rt_policy(policy)) {
4814 unsigned long rlim_rtprio =
4815 task_rlimit(p, RLIMIT_RTPRIO);
4816
4817 /* can't set/change the rt policy */
4818 if (policy != p->policy && !rlim_rtprio)
4819 return -EPERM;
4820
4821 /* can't increase priority */
4822 if (param->sched_priority > p->rt_priority &&
4823 param->sched_priority > rlim_rtprio)
4824 return -EPERM;
4825 }
4826 /*
4827 * Like positive nice levels, dont allow tasks to
4828 * move out of SCHED_IDLE either:
4829 */
4830 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4831 return -EPERM;
4832
4833 /* can't change other user's priorities */
4834 if (!check_same_owner(p))
4835 return -EPERM;
4836
4837 /* Normal users shall not reset the sched_reset_on_fork flag */
4838 if (p->sched_reset_on_fork && !reset_on_fork)
4839 return -EPERM;
4840 }
4841
4842 if (user) {
4843 retval = security_task_setscheduler(p);
4844 if (retval)
4845 return retval;
4846 }
4847
4848 /*
4849 * make sure no PI-waiters arrive (or leave) while we are
4850 * changing the priority of the task:
4851 */
4852 raw_spin_lock_irqsave(&p->pi_lock, flags);
4853 /*
4854 * To be able to change p->policy safely, the apropriate
4855 * runqueue lock must be held.
4856 */
4857 rq = __task_rq_lock(p);
4858
4859 /*
4860 * Changing the policy of the stop threads its a very bad idea
4861 */
4862 if (p == rq->stop) {
4863 __task_rq_unlock(rq);
4864 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4865 return -EINVAL;
4866 }
4867
4868 #ifdef CONFIG_RT_GROUP_SCHED
4869 if (user) {
4870 /*
4871 * Do not allow realtime tasks into groups that have no runtime
4872 * assigned.
4873 */
4874 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4875 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4876 !task_group_is_autogroup(task_group(p))) {
4877 __task_rq_unlock(rq);
4878 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4879 return -EPERM;
4880 }
4881 }
4882 #endif
4883
4884 /* recheck policy now with rq lock held */
4885 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4886 policy = oldpolicy = -1;
4887 __task_rq_unlock(rq);
4888 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4889 goto recheck;
4890 }
4891 on_rq = p->se.on_rq;
4892 running = task_current(rq, p);
4893 if (on_rq)
4894 deactivate_task(rq, p, 0);
4895 if (running)
4896 p->sched_class->put_prev_task(rq, p);
4897
4898 p->sched_reset_on_fork = reset_on_fork;
4899
4900 oldprio = p->prio;
4901 prev_class = p->sched_class;
4902 __setscheduler(rq, p, policy, param->sched_priority);
4903
4904 if (running)
4905 p->sched_class->set_curr_task(rq);
4906 if (on_rq) {
4907 activate_task(rq, p, 0);
4908
4909 check_class_changed(rq, p, prev_class, oldprio, running);
4910 }
4911 __task_rq_unlock(rq);
4912 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4913
4914 rt_mutex_adjust_pi(p);
4915
4916 return 0;
4917 }
4918
4919 /**
4920 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4921 * @p: the task in question.
4922 * @policy: new policy.
4923 * @param: structure containing the new RT priority.
4924 *
4925 * NOTE that the task may be already dead.
4926 */
4927 int sched_setscheduler(struct task_struct *p, int policy,
4928 const struct sched_param *param)
4929 {
4930 return __sched_setscheduler(p, policy, param, true);
4931 }
4932 EXPORT_SYMBOL_GPL(sched_setscheduler);
4933
4934 /**
4935 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4936 * @p: the task in question.
4937 * @policy: new policy.
4938 * @param: structure containing the new RT priority.
4939 *
4940 * Just like sched_setscheduler, only don't bother checking if the
4941 * current context has permission. For example, this is needed in
4942 * stop_machine(): we create temporary high priority worker threads,
4943 * but our caller might not have that capability.
4944 */
4945 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4946 const struct sched_param *param)
4947 {
4948 return __sched_setscheduler(p, policy, param, false);
4949 }
4950
4951 static int
4952 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4953 {
4954 struct sched_param lparam;
4955 struct task_struct *p;
4956 int retval;
4957
4958 if (!param || pid < 0)
4959 return -EINVAL;
4960 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4961 return -EFAULT;
4962
4963 rcu_read_lock();
4964 retval = -ESRCH;
4965 p = find_process_by_pid(pid);
4966 if (p != NULL)
4967 retval = sched_setscheduler(p, policy, &lparam);
4968 rcu_read_unlock();
4969
4970 return retval;
4971 }
4972
4973 /**
4974 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4975 * @pid: the pid in question.
4976 * @policy: new policy.
4977 * @param: structure containing the new RT priority.
4978 */
4979 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4980 struct sched_param __user *, param)
4981 {
4982 /* negative values for policy are not valid */
4983 if (policy < 0)
4984 return -EINVAL;
4985
4986 return do_sched_setscheduler(pid, policy, param);
4987 }
4988
4989 /**
4990 * sys_sched_setparam - set/change the RT priority of a thread
4991 * @pid: the pid in question.
4992 * @param: structure containing the new RT priority.
4993 */
4994 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4995 {
4996 return do_sched_setscheduler(pid, -1, param);
4997 }
4998
4999 /**
5000 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5001 * @pid: the pid in question.
5002 */
5003 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5004 {
5005 struct task_struct *p;
5006 int retval;
5007
5008 if (pid < 0)
5009 return -EINVAL;
5010
5011 retval = -ESRCH;
5012 rcu_read_lock();
5013 p = find_process_by_pid(pid);
5014 if (p) {
5015 retval = security_task_getscheduler(p);
5016 if (!retval)
5017 retval = p->policy
5018 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5019 }
5020 rcu_read_unlock();
5021 return retval;
5022 }
5023
5024 /**
5025 * sys_sched_getparam - get the RT priority of a thread
5026 * @pid: the pid in question.
5027 * @param: structure containing the RT priority.
5028 */
5029 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5030 {
5031 struct sched_param lp;
5032 struct task_struct *p;
5033 int retval;
5034
5035 if (!param || pid < 0)
5036 return -EINVAL;
5037
5038 rcu_read_lock();
5039 p = find_process_by_pid(pid);
5040 retval = -ESRCH;
5041 if (!p)
5042 goto out_unlock;
5043
5044 retval = security_task_getscheduler(p);
5045 if (retval)
5046 goto out_unlock;
5047
5048 lp.sched_priority = p->rt_priority;
5049 rcu_read_unlock();
5050
5051 /*
5052 * This one might sleep, we cannot do it with a spinlock held ...
5053 */
5054 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5055
5056 return retval;
5057
5058 out_unlock:
5059 rcu_read_unlock();
5060 return retval;
5061 }
5062
5063 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5064 {
5065 cpumask_var_t cpus_allowed, new_mask;
5066 struct task_struct *p;
5067 int retval;
5068
5069 get_online_cpus();
5070 rcu_read_lock();
5071
5072 p = find_process_by_pid(pid);
5073 if (!p) {
5074 rcu_read_unlock();
5075 put_online_cpus();
5076 return -ESRCH;
5077 }
5078
5079 /* Prevent p going away */
5080 get_task_struct(p);
5081 rcu_read_unlock();
5082
5083 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5084 retval = -ENOMEM;
5085 goto out_put_task;
5086 }
5087 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5088 retval = -ENOMEM;
5089 goto out_free_cpus_allowed;
5090 }
5091 retval = -EPERM;
5092 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5093 goto out_unlock;
5094
5095 retval = security_task_setscheduler(p);
5096 if (retval)
5097 goto out_unlock;
5098
5099 cpuset_cpus_allowed(p, cpus_allowed);
5100 cpumask_and(new_mask, in_mask, cpus_allowed);
5101 again:
5102 retval = set_cpus_allowed_ptr(p, new_mask);
5103
5104 if (!retval) {
5105 cpuset_cpus_allowed(p, cpus_allowed);
5106 if (!cpumask_subset(new_mask, cpus_allowed)) {
5107 /*
5108 * We must have raced with a concurrent cpuset
5109 * update. Just reset the cpus_allowed to the
5110 * cpuset's cpus_allowed
5111 */
5112 cpumask_copy(new_mask, cpus_allowed);
5113 goto again;
5114 }
5115 }
5116 out_unlock:
5117 free_cpumask_var(new_mask);
5118 out_free_cpus_allowed:
5119 free_cpumask_var(cpus_allowed);
5120 out_put_task:
5121 put_task_struct(p);
5122 put_online_cpus();
5123 return retval;
5124 }
5125
5126 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5127 struct cpumask *new_mask)
5128 {
5129 if (len < cpumask_size())
5130 cpumask_clear(new_mask);
5131 else if (len > cpumask_size())
5132 len = cpumask_size();
5133
5134 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5135 }
5136
5137 /**
5138 * sys_sched_setaffinity - set the cpu affinity of a process
5139 * @pid: pid of the process
5140 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5141 * @user_mask_ptr: user-space pointer to the new cpu mask
5142 */
5143 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5144 unsigned long __user *, user_mask_ptr)
5145 {
5146 cpumask_var_t new_mask;
5147 int retval;
5148
5149 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5150 return -ENOMEM;
5151
5152 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5153 if (retval == 0)
5154 retval = sched_setaffinity(pid, new_mask);
5155 free_cpumask_var(new_mask);
5156 return retval;
5157 }
5158
5159 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5160 {
5161 struct task_struct *p;
5162 unsigned long flags;
5163 struct rq *rq;
5164 int retval;
5165
5166 get_online_cpus();
5167 rcu_read_lock();
5168
5169 retval = -ESRCH;
5170 p = find_process_by_pid(pid);
5171 if (!p)
5172 goto out_unlock;
5173
5174 retval = security_task_getscheduler(p);
5175 if (retval)
5176 goto out_unlock;
5177
5178 rq = task_rq_lock(p, &flags);
5179 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5180 task_rq_unlock(rq, &flags);
5181
5182 out_unlock:
5183 rcu_read_unlock();
5184 put_online_cpus();
5185
5186 return retval;
5187 }
5188
5189 /**
5190 * sys_sched_getaffinity - get the cpu affinity of a process
5191 * @pid: pid of the process
5192 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5193 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5194 */
5195 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5196 unsigned long __user *, user_mask_ptr)
5197 {
5198 int ret;
5199 cpumask_var_t mask;
5200
5201 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5202 return -EINVAL;
5203 if (len & (sizeof(unsigned long)-1))
5204 return -EINVAL;
5205
5206 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5207 return -ENOMEM;
5208
5209 ret = sched_getaffinity(pid, mask);
5210 if (ret == 0) {
5211 size_t retlen = min_t(size_t, len, cpumask_size());
5212
5213 if (copy_to_user(user_mask_ptr, mask, retlen))
5214 ret = -EFAULT;
5215 else
5216 ret = retlen;
5217 }
5218 free_cpumask_var(mask);
5219
5220 return ret;
5221 }
5222
5223 /**
5224 * sys_sched_yield - yield the current processor to other threads.
5225 *
5226 * This function yields the current CPU to other tasks. If there are no
5227 * other threads running on this CPU then this function will return.
5228 */
5229 SYSCALL_DEFINE0(sched_yield)
5230 {
5231 struct rq *rq = this_rq_lock();
5232
5233 schedstat_inc(rq, yld_count);
5234 current->sched_class->yield_task(rq);
5235
5236 /*
5237 * Since we are going to call schedule() anyway, there's
5238 * no need to preempt or enable interrupts:
5239 */
5240 __release(rq->lock);
5241 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5242 do_raw_spin_unlock(&rq->lock);
5243 preempt_enable_no_resched();
5244
5245 schedule();
5246
5247 return 0;
5248 }
5249
5250 static inline int should_resched(void)
5251 {
5252 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5253 }
5254
5255 static void __cond_resched(void)
5256 {
5257 add_preempt_count(PREEMPT_ACTIVE);
5258 schedule();
5259 sub_preempt_count(PREEMPT_ACTIVE);
5260 }
5261
5262 int __sched _cond_resched(void)
5263 {
5264 if (should_resched()) {
5265 __cond_resched();
5266 return 1;
5267 }
5268 return 0;
5269 }
5270 EXPORT_SYMBOL(_cond_resched);
5271
5272 /*
5273 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5274 * call schedule, and on return reacquire the lock.
5275 *
5276 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5277 * operations here to prevent schedule() from being called twice (once via
5278 * spin_unlock(), once by hand).
5279 */
5280 int __cond_resched_lock(spinlock_t *lock)
5281 {
5282 int resched = should_resched();
5283 int ret = 0;
5284
5285 lockdep_assert_held(lock);
5286
5287 if (spin_needbreak(lock) || resched) {
5288 spin_unlock(lock);
5289 if (resched)
5290 __cond_resched();
5291 else
5292 cpu_relax();
5293 ret = 1;
5294 spin_lock(lock);
5295 }
5296 return ret;
5297 }
5298 EXPORT_SYMBOL(__cond_resched_lock);
5299
5300 int __sched __cond_resched_softirq(void)
5301 {
5302 BUG_ON(!in_softirq());
5303
5304 if (should_resched()) {
5305 local_bh_enable();
5306 __cond_resched();
5307 local_bh_disable();
5308 return 1;
5309 }
5310 return 0;
5311 }
5312 EXPORT_SYMBOL(__cond_resched_softirq);
5313
5314 /**
5315 * yield - yield the current processor to other threads.
5316 *
5317 * This is a shortcut for kernel-space yielding - it marks the
5318 * thread runnable and calls sys_sched_yield().
5319 */
5320 void __sched yield(void)
5321 {
5322 set_current_state(TASK_RUNNING);
5323 sys_sched_yield();
5324 }
5325 EXPORT_SYMBOL(yield);
5326
5327 /*
5328 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5329 * that process accounting knows that this is a task in IO wait state.
5330 */
5331 void __sched io_schedule(void)
5332 {
5333 struct rq *rq = raw_rq();
5334
5335 delayacct_blkio_start();
5336 atomic_inc(&rq->nr_iowait);
5337 current->in_iowait = 1;
5338 schedule();
5339 current->in_iowait = 0;
5340 atomic_dec(&rq->nr_iowait);
5341 delayacct_blkio_end();
5342 }
5343 EXPORT_SYMBOL(io_schedule);
5344
5345 long __sched io_schedule_timeout(long timeout)
5346 {
5347 struct rq *rq = raw_rq();
5348 long ret;
5349
5350 delayacct_blkio_start();
5351 atomic_inc(&rq->nr_iowait);
5352 current->in_iowait = 1;
5353 ret = schedule_timeout(timeout);
5354 current->in_iowait = 0;
5355 atomic_dec(&rq->nr_iowait);
5356 delayacct_blkio_end();
5357 return ret;
5358 }
5359
5360 /**
5361 * sys_sched_get_priority_max - return maximum RT priority.
5362 * @policy: scheduling class.
5363 *
5364 * this syscall returns the maximum rt_priority that can be used
5365 * by a given scheduling class.
5366 */
5367 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5368 {
5369 int ret = -EINVAL;
5370
5371 switch (policy) {
5372 case SCHED_FIFO:
5373 case SCHED_RR:
5374 ret = MAX_USER_RT_PRIO-1;
5375 break;
5376 case SCHED_NORMAL:
5377 case SCHED_BATCH:
5378 case SCHED_IDLE:
5379 ret = 0;
5380 break;
5381 }
5382 return ret;
5383 }
5384
5385 /**
5386 * sys_sched_get_priority_min - return minimum RT priority.
5387 * @policy: scheduling class.
5388 *
5389 * this syscall returns the minimum rt_priority that can be used
5390 * by a given scheduling class.
5391 */
5392 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5393 {
5394 int ret = -EINVAL;
5395
5396 switch (policy) {
5397 case SCHED_FIFO:
5398 case SCHED_RR:
5399 ret = 1;
5400 break;
5401 case SCHED_NORMAL:
5402 case SCHED_BATCH:
5403 case SCHED_IDLE:
5404 ret = 0;
5405 }
5406 return ret;
5407 }
5408
5409 /**
5410 * sys_sched_rr_get_interval - return the default timeslice of a process.
5411 * @pid: pid of the process.
5412 * @interval: userspace pointer to the timeslice value.
5413 *
5414 * this syscall writes the default timeslice value of a given process
5415 * into the user-space timespec buffer. A value of '0' means infinity.
5416 */
5417 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5418 struct timespec __user *, interval)
5419 {
5420 struct task_struct *p;
5421 unsigned int time_slice;
5422 unsigned long flags;
5423 struct rq *rq;
5424 int retval;
5425 struct timespec t;
5426
5427 if (pid < 0)
5428 return -EINVAL;
5429
5430 retval = -ESRCH;
5431 rcu_read_lock();
5432 p = find_process_by_pid(pid);
5433 if (!p)
5434 goto out_unlock;
5435
5436 retval = security_task_getscheduler(p);
5437 if (retval)
5438 goto out_unlock;
5439
5440 rq = task_rq_lock(p, &flags);
5441 time_slice = p->sched_class->get_rr_interval(rq, p);
5442 task_rq_unlock(rq, &flags);
5443
5444 rcu_read_unlock();
5445 jiffies_to_timespec(time_slice, &t);
5446 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5447 return retval;
5448
5449 out_unlock:
5450 rcu_read_unlock();
5451 return retval;
5452 }
5453
5454 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5455
5456 void sched_show_task(struct task_struct *p)
5457 {
5458 unsigned long free = 0;
5459 unsigned state;
5460
5461 state = p->state ? __ffs(p->state) + 1 : 0;
5462 printk(KERN_INFO "%-15.15s %c", p->comm,
5463 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5464 #if BITS_PER_LONG == 32
5465 if (state == TASK_RUNNING)
5466 printk(KERN_CONT " running ");
5467 else
5468 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5469 #else
5470 if (state == TASK_RUNNING)
5471 printk(KERN_CONT " running task ");
5472 else
5473 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5474 #endif
5475 #ifdef CONFIG_DEBUG_STACK_USAGE
5476 free = stack_not_used(p);
5477 #endif
5478 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5479 task_pid_nr(p), task_pid_nr(p->real_parent),
5480 (unsigned long)task_thread_info(p)->flags);
5481
5482 show_stack(p, NULL);
5483 }
5484
5485 void show_state_filter(unsigned long state_filter)
5486 {
5487 struct task_struct *g, *p;
5488
5489 #if BITS_PER_LONG == 32
5490 printk(KERN_INFO
5491 " task PC stack pid father\n");
5492 #else
5493 printk(KERN_INFO
5494 " task PC stack pid father\n");
5495 #endif
5496 read_lock(&tasklist_lock);
5497 do_each_thread(g, p) {
5498 /*
5499 * reset the NMI-timeout, listing all files on a slow
5500 * console might take alot of time:
5501 */
5502 touch_nmi_watchdog();
5503 if (!state_filter || (p->state & state_filter))
5504 sched_show_task(p);
5505 } while_each_thread(g, p);
5506
5507 touch_all_softlockup_watchdogs();
5508
5509 #ifdef CONFIG_SCHED_DEBUG
5510 sysrq_sched_debug_show();
5511 #endif
5512 read_unlock(&tasklist_lock);
5513 /*
5514 * Only show locks if all tasks are dumped:
5515 */
5516 if (!state_filter)
5517 debug_show_all_locks();
5518 }
5519
5520 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5521 {
5522 idle->sched_class = &idle_sched_class;
5523 }
5524
5525 /**
5526 * init_idle - set up an idle thread for a given CPU
5527 * @idle: task in question
5528 * @cpu: cpu the idle task belongs to
5529 *
5530 * NOTE: this function does not set the idle thread's NEED_RESCHED
5531 * flag, to make booting more robust.
5532 */
5533 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5534 {
5535 struct rq *rq = cpu_rq(cpu);
5536 unsigned long flags;
5537
5538 raw_spin_lock_irqsave(&rq->lock, flags);
5539
5540 __sched_fork(idle);
5541 idle->state = TASK_RUNNING;
5542 idle->se.exec_start = sched_clock();
5543
5544 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5545 /*
5546 * We're having a chicken and egg problem, even though we are
5547 * holding rq->lock, the cpu isn't yet set to this cpu so the
5548 * lockdep check in task_group() will fail.
5549 *
5550 * Similar case to sched_fork(). / Alternatively we could
5551 * use task_rq_lock() here and obtain the other rq->lock.
5552 *
5553 * Silence PROVE_RCU
5554 */
5555 rcu_read_lock();
5556 __set_task_cpu(idle, cpu);
5557 rcu_read_unlock();
5558
5559 rq->curr = rq->idle = idle;
5560 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5561 idle->oncpu = 1;
5562 #endif
5563 raw_spin_unlock_irqrestore(&rq->lock, flags);
5564
5565 /* Set the preempt count _outside_ the spinlocks! */
5566 #if defined(CONFIG_PREEMPT)
5567 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5568 #else
5569 task_thread_info(idle)->preempt_count = 0;
5570 #endif
5571 /*
5572 * The idle tasks have their own, simple scheduling class:
5573 */
5574 idle->sched_class = &idle_sched_class;
5575 ftrace_graph_init_task(idle);
5576 }
5577
5578 /*
5579 * In a system that switches off the HZ timer nohz_cpu_mask
5580 * indicates which cpus entered this state. This is used
5581 * in the rcu update to wait only for active cpus. For system
5582 * which do not switch off the HZ timer nohz_cpu_mask should
5583 * always be CPU_BITS_NONE.
5584 */
5585 cpumask_var_t nohz_cpu_mask;
5586
5587 /*
5588 * Increase the granularity value when there are more CPUs,
5589 * because with more CPUs the 'effective latency' as visible
5590 * to users decreases. But the relationship is not linear,
5591 * so pick a second-best guess by going with the log2 of the
5592 * number of CPUs.
5593 *
5594 * This idea comes from the SD scheduler of Con Kolivas:
5595 */
5596 static int get_update_sysctl_factor(void)
5597 {
5598 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5599 unsigned int factor;
5600
5601 switch (sysctl_sched_tunable_scaling) {
5602 case SCHED_TUNABLESCALING_NONE:
5603 factor = 1;
5604 break;
5605 case SCHED_TUNABLESCALING_LINEAR:
5606 factor = cpus;
5607 break;
5608 case SCHED_TUNABLESCALING_LOG:
5609 default:
5610 factor = 1 + ilog2(cpus);
5611 break;
5612 }
5613
5614 return factor;
5615 }
5616
5617 static void update_sysctl(void)
5618 {
5619 unsigned int factor = get_update_sysctl_factor();
5620
5621 #define SET_SYSCTL(name) \
5622 (sysctl_##name = (factor) * normalized_sysctl_##name)
5623 SET_SYSCTL(sched_min_granularity);
5624 SET_SYSCTL(sched_latency);
5625 SET_SYSCTL(sched_wakeup_granularity);
5626 #undef SET_SYSCTL
5627 }
5628
5629 static inline void sched_init_granularity(void)
5630 {
5631 update_sysctl();
5632 }
5633
5634 #ifdef CONFIG_SMP
5635 /*
5636 * This is how migration works:
5637 *
5638 * 1) we invoke migration_cpu_stop() on the target CPU using
5639 * stop_one_cpu().
5640 * 2) stopper starts to run (implicitly forcing the migrated thread
5641 * off the CPU)
5642 * 3) it checks whether the migrated task is still in the wrong runqueue.
5643 * 4) if it's in the wrong runqueue then the migration thread removes
5644 * it and puts it into the right queue.
5645 * 5) stopper completes and stop_one_cpu() returns and the migration
5646 * is done.
5647 */
5648
5649 /*
5650 * Change a given task's CPU affinity. Migrate the thread to a
5651 * proper CPU and schedule it away if the CPU it's executing on
5652 * is removed from the allowed bitmask.
5653 *
5654 * NOTE: the caller must have a valid reference to the task, the
5655 * task must not exit() & deallocate itself prematurely. The
5656 * call is not atomic; no spinlocks may be held.
5657 */
5658 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5659 {
5660 unsigned long flags;
5661 struct rq *rq;
5662 unsigned int dest_cpu;
5663 int ret = 0;
5664
5665 /*
5666 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5667 * drop the rq->lock and still rely on ->cpus_allowed.
5668 */
5669 again:
5670 while (task_is_waking(p))
5671 cpu_relax();
5672 rq = task_rq_lock(p, &flags);
5673 if (task_is_waking(p)) {
5674 task_rq_unlock(rq, &flags);
5675 goto again;
5676 }
5677
5678 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5679 ret = -EINVAL;
5680 goto out;
5681 }
5682
5683 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5684 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5685 ret = -EINVAL;
5686 goto out;
5687 }
5688
5689 if (p->sched_class->set_cpus_allowed)
5690 p->sched_class->set_cpus_allowed(p, new_mask);
5691 else {
5692 cpumask_copy(&p->cpus_allowed, new_mask);
5693 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5694 }
5695
5696 /* Can the task run on the task's current CPU? If so, we're done */
5697 if (cpumask_test_cpu(task_cpu(p), new_mask))
5698 goto out;
5699
5700 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5701 if (migrate_task(p, rq)) {
5702 struct migration_arg arg = { p, dest_cpu };
5703 /* Need help from migration thread: drop lock and wait. */
5704 task_rq_unlock(rq, &flags);
5705 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5706 tlb_migrate_finish(p->mm);
5707 return 0;
5708 }
5709 out:
5710 task_rq_unlock(rq, &flags);
5711
5712 return ret;
5713 }
5714 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5715
5716 /*
5717 * Move (not current) task off this cpu, onto dest cpu. We're doing
5718 * this because either it can't run here any more (set_cpus_allowed()
5719 * away from this CPU, or CPU going down), or because we're
5720 * attempting to rebalance this task on exec (sched_exec).
5721 *
5722 * So we race with normal scheduler movements, but that's OK, as long
5723 * as the task is no longer on this CPU.
5724 *
5725 * Returns non-zero if task was successfully migrated.
5726 */
5727 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5728 {
5729 struct rq *rq_dest, *rq_src;
5730 int ret = 0;
5731
5732 if (unlikely(!cpu_active(dest_cpu)))
5733 return ret;
5734
5735 rq_src = cpu_rq(src_cpu);
5736 rq_dest = cpu_rq(dest_cpu);
5737
5738 double_rq_lock(rq_src, rq_dest);
5739 /* Already moved. */
5740 if (task_cpu(p) != src_cpu)
5741 goto done;
5742 /* Affinity changed (again). */
5743 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5744 goto fail;
5745
5746 /*
5747 * If we're not on a rq, the next wake-up will ensure we're
5748 * placed properly.
5749 */
5750 if (p->se.on_rq) {
5751 deactivate_task(rq_src, p, 0);
5752 set_task_cpu(p, dest_cpu);
5753 activate_task(rq_dest, p, 0);
5754 check_preempt_curr(rq_dest, p, 0);
5755 }
5756 done:
5757 ret = 1;
5758 fail:
5759 double_rq_unlock(rq_src, rq_dest);
5760 return ret;
5761 }
5762
5763 /*
5764 * migration_cpu_stop - this will be executed by a highprio stopper thread
5765 * and performs thread migration by bumping thread off CPU then
5766 * 'pushing' onto another runqueue.
5767 */
5768 static int migration_cpu_stop(void *data)
5769 {
5770 struct migration_arg *arg = data;
5771
5772 /*
5773 * The original target cpu might have gone down and we might
5774 * be on another cpu but it doesn't matter.
5775 */
5776 local_irq_disable();
5777 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5778 local_irq_enable();
5779 return 0;
5780 }
5781
5782 #ifdef CONFIG_HOTPLUG_CPU
5783
5784 /*
5785 * Ensures that the idle task is using init_mm right before its cpu goes
5786 * offline.
5787 */
5788 void idle_task_exit(void)
5789 {
5790 struct mm_struct *mm = current->active_mm;
5791
5792 BUG_ON(cpu_online(smp_processor_id()));
5793
5794 if (mm != &init_mm)
5795 switch_mm(mm, &init_mm, current);
5796 mmdrop(mm);
5797 }
5798
5799 /*
5800 * While a dead CPU has no uninterruptible tasks queued at this point,
5801 * it might still have a nonzero ->nr_uninterruptible counter, because
5802 * for performance reasons the counter is not stricly tracking tasks to
5803 * their home CPUs. So we just add the counter to another CPU's counter,
5804 * to keep the global sum constant after CPU-down:
5805 */
5806 static void migrate_nr_uninterruptible(struct rq *rq_src)
5807 {
5808 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5809
5810 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5811 rq_src->nr_uninterruptible = 0;
5812 }
5813
5814 /*
5815 * remove the tasks which were accounted by rq from calc_load_tasks.
5816 */
5817 static void calc_global_load_remove(struct rq *rq)
5818 {
5819 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5820 rq->calc_load_active = 0;
5821 }
5822
5823 /*
5824 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5825 * try_to_wake_up()->select_task_rq().
5826 *
5827 * Called with rq->lock held even though we'er in stop_machine() and
5828 * there's no concurrency possible, we hold the required locks anyway
5829 * because of lock validation efforts.
5830 */
5831 static void migrate_tasks(unsigned int dead_cpu)
5832 {
5833 struct rq *rq = cpu_rq(dead_cpu);
5834 struct task_struct *next, *stop = rq->stop;
5835 int dest_cpu;
5836
5837 /*
5838 * Fudge the rq selection such that the below task selection loop
5839 * doesn't get stuck on the currently eligible stop task.
5840 *
5841 * We're currently inside stop_machine() and the rq is either stuck
5842 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5843 * either way we should never end up calling schedule() until we're
5844 * done here.
5845 */
5846 rq->stop = NULL;
5847
5848 for ( ; ; ) {
5849 /*
5850 * There's this thread running, bail when that's the only
5851 * remaining thread.
5852 */
5853 if (rq->nr_running == 1)
5854 break;
5855
5856 next = pick_next_task(rq);
5857 BUG_ON(!next);
5858 next->sched_class->put_prev_task(rq, next);
5859
5860 /* Find suitable destination for @next, with force if needed. */
5861 dest_cpu = select_fallback_rq(dead_cpu, next);
5862 raw_spin_unlock(&rq->lock);
5863
5864 __migrate_task(next, dead_cpu, dest_cpu);
5865
5866 raw_spin_lock(&rq->lock);
5867 }
5868
5869 rq->stop = stop;
5870 }
5871
5872 #endif /* CONFIG_HOTPLUG_CPU */
5873
5874 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5875
5876 static struct ctl_table sd_ctl_dir[] = {
5877 {
5878 .procname = "sched_domain",
5879 .mode = 0555,
5880 },
5881 {}
5882 };
5883
5884 static struct ctl_table sd_ctl_root[] = {
5885 {
5886 .procname = "kernel",
5887 .mode = 0555,
5888 .child = sd_ctl_dir,
5889 },
5890 {}
5891 };
5892
5893 static struct ctl_table *sd_alloc_ctl_entry(int n)
5894 {
5895 struct ctl_table *entry =
5896 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5897
5898 return entry;
5899 }
5900
5901 static void sd_free_ctl_entry(struct ctl_table **tablep)
5902 {
5903 struct ctl_table *entry;
5904
5905 /*
5906 * In the intermediate directories, both the child directory and
5907 * procname are dynamically allocated and could fail but the mode
5908 * will always be set. In the lowest directory the names are
5909 * static strings and all have proc handlers.
5910 */
5911 for (entry = *tablep; entry->mode; entry++) {
5912 if (entry->child)
5913 sd_free_ctl_entry(&entry->child);
5914 if (entry->proc_handler == NULL)
5915 kfree(entry->procname);
5916 }
5917
5918 kfree(*tablep);
5919 *tablep = NULL;
5920 }
5921
5922 static void
5923 set_table_entry(struct ctl_table *entry,
5924 const char *procname, void *data, int maxlen,
5925 mode_t mode, proc_handler *proc_handler)
5926 {
5927 entry->procname = procname;
5928 entry->data = data;
5929 entry->maxlen = maxlen;
5930 entry->mode = mode;
5931 entry->proc_handler = proc_handler;
5932 }
5933
5934 static struct ctl_table *
5935 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5936 {
5937 struct ctl_table *table = sd_alloc_ctl_entry(13);
5938
5939 if (table == NULL)
5940 return NULL;
5941
5942 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5943 sizeof(long), 0644, proc_doulongvec_minmax);
5944 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5945 sizeof(long), 0644, proc_doulongvec_minmax);
5946 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5947 sizeof(int), 0644, proc_dointvec_minmax);
5948 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5949 sizeof(int), 0644, proc_dointvec_minmax);
5950 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5951 sizeof(int), 0644, proc_dointvec_minmax);
5952 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5953 sizeof(int), 0644, proc_dointvec_minmax);
5954 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5955 sizeof(int), 0644, proc_dointvec_minmax);
5956 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5957 sizeof(int), 0644, proc_dointvec_minmax);
5958 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5959 sizeof(int), 0644, proc_dointvec_minmax);
5960 set_table_entry(&table[9], "cache_nice_tries",
5961 &sd->cache_nice_tries,
5962 sizeof(int), 0644, proc_dointvec_minmax);
5963 set_table_entry(&table[10], "flags", &sd->flags,
5964 sizeof(int), 0644, proc_dointvec_minmax);
5965 set_table_entry(&table[11], "name", sd->name,
5966 CORENAME_MAX_SIZE, 0444, proc_dostring);
5967 /* &table[12] is terminator */
5968
5969 return table;
5970 }
5971
5972 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5973 {
5974 struct ctl_table *entry, *table;
5975 struct sched_domain *sd;
5976 int domain_num = 0, i;
5977 char buf[32];
5978
5979 for_each_domain(cpu, sd)
5980 domain_num++;
5981 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5982 if (table == NULL)
5983 return NULL;
5984
5985 i = 0;
5986 for_each_domain(cpu, sd) {
5987 snprintf(buf, 32, "domain%d", i);
5988 entry->procname = kstrdup(buf, GFP_KERNEL);
5989 entry->mode = 0555;
5990 entry->child = sd_alloc_ctl_domain_table(sd);
5991 entry++;
5992 i++;
5993 }
5994 return table;
5995 }
5996
5997 static struct ctl_table_header *sd_sysctl_header;
5998 static void register_sched_domain_sysctl(void)
5999 {
6000 int i, cpu_num = num_possible_cpus();
6001 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6002 char buf[32];
6003
6004 WARN_ON(sd_ctl_dir[0].child);
6005 sd_ctl_dir[0].child = entry;
6006
6007 if (entry == NULL)
6008 return;
6009
6010 for_each_possible_cpu(i) {
6011 snprintf(buf, 32, "cpu%d", i);
6012 entry->procname = kstrdup(buf, GFP_KERNEL);
6013 entry->mode = 0555;
6014 entry->child = sd_alloc_ctl_cpu_table(i);
6015 entry++;
6016 }
6017
6018 WARN_ON(sd_sysctl_header);
6019 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6020 }
6021
6022 /* may be called multiple times per register */
6023 static void unregister_sched_domain_sysctl(void)
6024 {
6025 if (sd_sysctl_header)
6026 unregister_sysctl_table(sd_sysctl_header);
6027 sd_sysctl_header = NULL;
6028 if (sd_ctl_dir[0].child)
6029 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6030 }
6031 #else
6032 static void register_sched_domain_sysctl(void)
6033 {
6034 }
6035 static void unregister_sched_domain_sysctl(void)
6036 {
6037 }
6038 #endif
6039
6040 static void set_rq_online(struct rq *rq)
6041 {
6042 if (!rq->online) {
6043 const struct sched_class *class;
6044
6045 cpumask_set_cpu(rq->cpu, rq->rd->online);
6046 rq->online = 1;
6047
6048 for_each_class(class) {
6049 if (class->rq_online)
6050 class->rq_online(rq);
6051 }
6052 }
6053 }
6054
6055 static void set_rq_offline(struct rq *rq)
6056 {
6057 if (rq->online) {
6058 const struct sched_class *class;
6059
6060 for_each_class(class) {
6061 if (class->rq_offline)
6062 class->rq_offline(rq);
6063 }
6064
6065 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6066 rq->online = 0;
6067 }
6068 }
6069
6070 /*
6071 * migration_call - callback that gets triggered when a CPU is added.
6072 * Here we can start up the necessary migration thread for the new CPU.
6073 */
6074 static int __cpuinit
6075 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6076 {
6077 int cpu = (long)hcpu;
6078 unsigned long flags;
6079 struct rq *rq = cpu_rq(cpu);
6080
6081 switch (action & ~CPU_TASKS_FROZEN) {
6082
6083 case CPU_UP_PREPARE:
6084 rq->calc_load_update = calc_load_update;
6085 break;
6086
6087 case CPU_ONLINE:
6088 /* Update our root-domain */
6089 raw_spin_lock_irqsave(&rq->lock, flags);
6090 if (rq->rd) {
6091 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6092
6093 set_rq_online(rq);
6094 }
6095 raw_spin_unlock_irqrestore(&rq->lock, flags);
6096 break;
6097
6098 #ifdef CONFIG_HOTPLUG_CPU
6099 case CPU_DYING:
6100 /* Update our root-domain */
6101 raw_spin_lock_irqsave(&rq->lock, flags);
6102 if (rq->rd) {
6103 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6104 set_rq_offline(rq);
6105 }
6106 migrate_tasks(cpu);
6107 BUG_ON(rq->nr_running != 1); /* the migration thread */
6108 raw_spin_unlock_irqrestore(&rq->lock, flags);
6109
6110 migrate_nr_uninterruptible(rq);
6111 calc_global_load_remove(rq);
6112 break;
6113 #endif
6114 }
6115 return NOTIFY_OK;
6116 }
6117
6118 /*
6119 * Register at high priority so that task migration (migrate_all_tasks)
6120 * happens before everything else. This has to be lower priority than
6121 * the notifier in the perf_event subsystem, though.
6122 */
6123 static struct notifier_block __cpuinitdata migration_notifier = {
6124 .notifier_call = migration_call,
6125 .priority = CPU_PRI_MIGRATION,
6126 };
6127
6128 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6129 unsigned long action, void *hcpu)
6130 {
6131 switch (action & ~CPU_TASKS_FROZEN) {
6132 case CPU_ONLINE:
6133 case CPU_DOWN_FAILED:
6134 set_cpu_active((long)hcpu, true);
6135 return NOTIFY_OK;
6136 default:
6137 return NOTIFY_DONE;
6138 }
6139 }
6140
6141 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6142 unsigned long action, void *hcpu)
6143 {
6144 switch (action & ~CPU_TASKS_FROZEN) {
6145 case CPU_DOWN_PREPARE:
6146 set_cpu_active((long)hcpu, false);
6147 return NOTIFY_OK;
6148 default:
6149 return NOTIFY_DONE;
6150 }
6151 }
6152
6153 static int __init migration_init(void)
6154 {
6155 void *cpu = (void *)(long)smp_processor_id();
6156 int err;
6157
6158 /* Initialize migration for the boot CPU */
6159 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6160 BUG_ON(err == NOTIFY_BAD);
6161 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6162 register_cpu_notifier(&migration_notifier);
6163
6164 /* Register cpu active notifiers */
6165 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6166 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6167
6168 return 0;
6169 }
6170 early_initcall(migration_init);
6171 #endif
6172
6173 #ifdef CONFIG_SMP
6174
6175 #ifdef CONFIG_SCHED_DEBUG
6176
6177 static __read_mostly int sched_domain_debug_enabled;
6178
6179 static int __init sched_domain_debug_setup(char *str)
6180 {
6181 sched_domain_debug_enabled = 1;
6182
6183 return 0;
6184 }
6185 early_param("sched_debug", sched_domain_debug_setup);
6186
6187 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6188 struct cpumask *groupmask)
6189 {
6190 struct sched_group *group = sd->groups;
6191 char str[256];
6192
6193 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6194 cpumask_clear(groupmask);
6195
6196 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6197
6198 if (!(sd->flags & SD_LOAD_BALANCE)) {
6199 printk("does not load-balance\n");
6200 if (sd->parent)
6201 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6202 " has parent");
6203 return -1;
6204 }
6205
6206 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6207
6208 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6209 printk(KERN_ERR "ERROR: domain->span does not contain "
6210 "CPU%d\n", cpu);
6211 }
6212 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6213 printk(KERN_ERR "ERROR: domain->groups does not contain"
6214 " CPU%d\n", cpu);
6215 }
6216
6217 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6218 do {
6219 if (!group) {
6220 printk("\n");
6221 printk(KERN_ERR "ERROR: group is NULL\n");
6222 break;
6223 }
6224
6225 if (!group->cpu_power) {
6226 printk(KERN_CONT "\n");
6227 printk(KERN_ERR "ERROR: domain->cpu_power not "
6228 "set\n");
6229 break;
6230 }
6231
6232 if (!cpumask_weight(sched_group_cpus(group))) {
6233 printk(KERN_CONT "\n");
6234 printk(KERN_ERR "ERROR: empty group\n");
6235 break;
6236 }
6237
6238 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6239 printk(KERN_CONT "\n");
6240 printk(KERN_ERR "ERROR: repeated CPUs\n");
6241 break;
6242 }
6243
6244 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6245
6246 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6247
6248 printk(KERN_CONT " %s", str);
6249 if (group->cpu_power != SCHED_LOAD_SCALE) {
6250 printk(KERN_CONT " (cpu_power = %d)",
6251 group->cpu_power);
6252 }
6253
6254 group = group->next;
6255 } while (group != sd->groups);
6256 printk(KERN_CONT "\n");
6257
6258 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6259 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6260
6261 if (sd->parent &&
6262 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6263 printk(KERN_ERR "ERROR: parent span is not a superset "
6264 "of domain->span\n");
6265 return 0;
6266 }
6267
6268 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6269 {
6270 cpumask_var_t groupmask;
6271 int level = 0;
6272
6273 if (!sched_domain_debug_enabled)
6274 return;
6275
6276 if (!sd) {
6277 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6278 return;
6279 }
6280
6281 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6282
6283 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6284 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6285 return;
6286 }
6287
6288 for (;;) {
6289 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6290 break;
6291 level++;
6292 sd = sd->parent;
6293 if (!sd)
6294 break;
6295 }
6296 free_cpumask_var(groupmask);
6297 }
6298 #else /* !CONFIG_SCHED_DEBUG */
6299 # define sched_domain_debug(sd, cpu) do { } while (0)
6300 #endif /* CONFIG_SCHED_DEBUG */
6301
6302 static int sd_degenerate(struct sched_domain *sd)
6303 {
6304 if (cpumask_weight(sched_domain_span(sd)) == 1)
6305 return 1;
6306
6307 /* Following flags need at least 2 groups */
6308 if (sd->flags & (SD_LOAD_BALANCE |
6309 SD_BALANCE_NEWIDLE |
6310 SD_BALANCE_FORK |
6311 SD_BALANCE_EXEC |
6312 SD_SHARE_CPUPOWER |
6313 SD_SHARE_PKG_RESOURCES)) {
6314 if (sd->groups != sd->groups->next)
6315 return 0;
6316 }
6317
6318 /* Following flags don't use groups */
6319 if (sd->flags & (SD_WAKE_AFFINE))
6320 return 0;
6321
6322 return 1;
6323 }
6324
6325 static int
6326 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6327 {
6328 unsigned long cflags = sd->flags, pflags = parent->flags;
6329
6330 if (sd_degenerate(parent))
6331 return 1;
6332
6333 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6334 return 0;
6335
6336 /* Flags needing groups don't count if only 1 group in parent */
6337 if (parent->groups == parent->groups->next) {
6338 pflags &= ~(SD_LOAD_BALANCE |
6339 SD_BALANCE_NEWIDLE |
6340 SD_BALANCE_FORK |
6341 SD_BALANCE_EXEC |
6342 SD_SHARE_CPUPOWER |
6343 SD_SHARE_PKG_RESOURCES);
6344 if (nr_node_ids == 1)
6345 pflags &= ~SD_SERIALIZE;
6346 }
6347 if (~cflags & pflags)
6348 return 0;
6349
6350 return 1;
6351 }
6352
6353 static void free_rootdomain(struct root_domain *rd)
6354 {
6355 synchronize_sched();
6356
6357 cpupri_cleanup(&rd->cpupri);
6358
6359 free_cpumask_var(rd->rto_mask);
6360 free_cpumask_var(rd->online);
6361 free_cpumask_var(rd->span);
6362 kfree(rd);
6363 }
6364
6365 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6366 {
6367 struct root_domain *old_rd = NULL;
6368 unsigned long flags;
6369
6370 raw_spin_lock_irqsave(&rq->lock, flags);
6371
6372 if (rq->rd) {
6373 old_rd = rq->rd;
6374
6375 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6376 set_rq_offline(rq);
6377
6378 cpumask_clear_cpu(rq->cpu, old_rd->span);
6379
6380 /*
6381 * If we dont want to free the old_rt yet then
6382 * set old_rd to NULL to skip the freeing later
6383 * in this function:
6384 */
6385 if (!atomic_dec_and_test(&old_rd->refcount))
6386 old_rd = NULL;
6387 }
6388
6389 atomic_inc(&rd->refcount);
6390 rq->rd = rd;
6391
6392 cpumask_set_cpu(rq->cpu, rd->span);
6393 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6394 set_rq_online(rq);
6395
6396 raw_spin_unlock_irqrestore(&rq->lock, flags);
6397
6398 if (old_rd)
6399 free_rootdomain(old_rd);
6400 }
6401
6402 static int init_rootdomain(struct root_domain *rd)
6403 {
6404 memset(rd, 0, sizeof(*rd));
6405
6406 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6407 goto out;
6408 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6409 goto free_span;
6410 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6411 goto free_online;
6412
6413 if (cpupri_init(&rd->cpupri) != 0)
6414 goto free_rto_mask;
6415 return 0;
6416
6417 free_rto_mask:
6418 free_cpumask_var(rd->rto_mask);
6419 free_online:
6420 free_cpumask_var(rd->online);
6421 free_span:
6422 free_cpumask_var(rd->span);
6423 out:
6424 return -ENOMEM;
6425 }
6426
6427 static void init_defrootdomain(void)
6428 {
6429 init_rootdomain(&def_root_domain);
6430
6431 atomic_set(&def_root_domain.refcount, 1);
6432 }
6433
6434 static struct root_domain *alloc_rootdomain(void)
6435 {
6436 struct root_domain *rd;
6437
6438 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6439 if (!rd)
6440 return NULL;
6441
6442 if (init_rootdomain(rd) != 0) {
6443 kfree(rd);
6444 return NULL;
6445 }
6446
6447 return rd;
6448 }
6449
6450 /*
6451 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6452 * hold the hotplug lock.
6453 */
6454 static void
6455 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6456 {
6457 struct rq *rq = cpu_rq(cpu);
6458 struct sched_domain *tmp;
6459
6460 for (tmp = sd; tmp; tmp = tmp->parent)
6461 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6462
6463 /* Remove the sched domains which do not contribute to scheduling. */
6464 for (tmp = sd; tmp; ) {
6465 struct sched_domain *parent = tmp->parent;
6466 if (!parent)
6467 break;
6468
6469 if (sd_parent_degenerate(tmp, parent)) {
6470 tmp->parent = parent->parent;
6471 if (parent->parent)
6472 parent->parent->child = tmp;
6473 } else
6474 tmp = tmp->parent;
6475 }
6476
6477 if (sd && sd_degenerate(sd)) {
6478 sd = sd->parent;
6479 if (sd)
6480 sd->child = NULL;
6481 }
6482
6483 sched_domain_debug(sd, cpu);
6484
6485 rq_attach_root(rq, rd);
6486 rcu_assign_pointer(rq->sd, sd);
6487 }
6488
6489 /* cpus with isolated domains */
6490 static cpumask_var_t cpu_isolated_map;
6491
6492 /* Setup the mask of cpus configured for isolated domains */
6493 static int __init isolated_cpu_setup(char *str)
6494 {
6495 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6496 cpulist_parse(str, cpu_isolated_map);
6497 return 1;
6498 }
6499
6500 __setup("isolcpus=", isolated_cpu_setup);
6501
6502 /*
6503 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6504 * to a function which identifies what group(along with sched group) a CPU
6505 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6506 * (due to the fact that we keep track of groups covered with a struct cpumask).
6507 *
6508 * init_sched_build_groups will build a circular linked list of the groups
6509 * covered by the given span, and will set each group's ->cpumask correctly,
6510 * and ->cpu_power to 0.
6511 */
6512 static void
6513 init_sched_build_groups(const struct cpumask *span,
6514 const struct cpumask *cpu_map,
6515 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6516 struct sched_group **sg,
6517 struct cpumask *tmpmask),
6518 struct cpumask *covered, struct cpumask *tmpmask)
6519 {
6520 struct sched_group *first = NULL, *last = NULL;
6521 int i;
6522
6523 cpumask_clear(covered);
6524
6525 for_each_cpu(i, span) {
6526 struct sched_group *sg;
6527 int group = group_fn(i, cpu_map, &sg, tmpmask);
6528 int j;
6529
6530 if (cpumask_test_cpu(i, covered))
6531 continue;
6532
6533 cpumask_clear(sched_group_cpus(sg));
6534 sg->cpu_power = 0;
6535
6536 for_each_cpu(j, span) {
6537 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6538 continue;
6539
6540 cpumask_set_cpu(j, covered);
6541 cpumask_set_cpu(j, sched_group_cpus(sg));
6542 }
6543 if (!first)
6544 first = sg;
6545 if (last)
6546 last->next = sg;
6547 last = sg;
6548 }
6549 last->next = first;
6550 }
6551
6552 #define SD_NODES_PER_DOMAIN 16
6553
6554 #ifdef CONFIG_NUMA
6555
6556 /**
6557 * find_next_best_node - find the next node to include in a sched_domain
6558 * @node: node whose sched_domain we're building
6559 * @used_nodes: nodes already in the sched_domain
6560 *
6561 * Find the next node to include in a given scheduling domain. Simply
6562 * finds the closest node not already in the @used_nodes map.
6563 *
6564 * Should use nodemask_t.
6565 */
6566 static int find_next_best_node(int node, nodemask_t *used_nodes)
6567 {
6568 int i, n, val, min_val, best_node = 0;
6569
6570 min_val = INT_MAX;
6571
6572 for (i = 0; i < nr_node_ids; i++) {
6573 /* Start at @node */
6574 n = (node + i) % nr_node_ids;
6575
6576 if (!nr_cpus_node(n))
6577 continue;
6578
6579 /* Skip already used nodes */
6580 if (node_isset(n, *used_nodes))
6581 continue;
6582
6583 /* Simple min distance search */
6584 val = node_distance(node, n);
6585
6586 if (val < min_val) {
6587 min_val = val;
6588 best_node = n;
6589 }
6590 }
6591
6592 node_set(best_node, *used_nodes);
6593 return best_node;
6594 }
6595
6596 /**
6597 * sched_domain_node_span - get a cpumask for a node's sched_domain
6598 * @node: node whose cpumask we're constructing
6599 * @span: resulting cpumask
6600 *
6601 * Given a node, construct a good cpumask for its sched_domain to span. It
6602 * should be one that prevents unnecessary balancing, but also spreads tasks
6603 * out optimally.
6604 */
6605 static void sched_domain_node_span(int node, struct cpumask *span)
6606 {
6607 nodemask_t used_nodes;
6608 int i;
6609
6610 cpumask_clear(span);
6611 nodes_clear(used_nodes);
6612
6613 cpumask_or(span, span, cpumask_of_node(node));
6614 node_set(node, used_nodes);
6615
6616 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6617 int next_node = find_next_best_node(node, &used_nodes);
6618
6619 cpumask_or(span, span, cpumask_of_node(next_node));
6620 }
6621 }
6622 #endif /* CONFIG_NUMA */
6623
6624 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6625
6626 /*
6627 * The cpus mask in sched_group and sched_domain hangs off the end.
6628 *
6629 * ( See the the comments in include/linux/sched.h:struct sched_group
6630 * and struct sched_domain. )
6631 */
6632 struct static_sched_group {
6633 struct sched_group sg;
6634 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6635 };
6636
6637 struct static_sched_domain {
6638 struct sched_domain sd;
6639 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6640 };
6641
6642 struct s_data {
6643 #ifdef CONFIG_NUMA
6644 int sd_allnodes;
6645 cpumask_var_t domainspan;
6646 cpumask_var_t covered;
6647 cpumask_var_t notcovered;
6648 #endif
6649 cpumask_var_t nodemask;
6650 cpumask_var_t this_sibling_map;
6651 cpumask_var_t this_core_map;
6652 cpumask_var_t this_book_map;
6653 cpumask_var_t send_covered;
6654 cpumask_var_t tmpmask;
6655 struct sched_group **sched_group_nodes;
6656 struct root_domain *rd;
6657 };
6658
6659 enum s_alloc {
6660 sa_sched_groups = 0,
6661 sa_rootdomain,
6662 sa_tmpmask,
6663 sa_send_covered,
6664 sa_this_book_map,
6665 sa_this_core_map,
6666 sa_this_sibling_map,
6667 sa_nodemask,
6668 sa_sched_group_nodes,
6669 #ifdef CONFIG_NUMA
6670 sa_notcovered,
6671 sa_covered,
6672 sa_domainspan,
6673 #endif
6674 sa_none,
6675 };
6676
6677 /*
6678 * SMT sched-domains:
6679 */
6680 #ifdef CONFIG_SCHED_SMT
6681 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6682 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6683
6684 static int
6685 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6686 struct sched_group **sg, struct cpumask *unused)
6687 {
6688 if (sg)
6689 *sg = &per_cpu(sched_groups, cpu).sg;
6690 return cpu;
6691 }
6692 #endif /* CONFIG_SCHED_SMT */
6693
6694 /*
6695 * multi-core sched-domains:
6696 */
6697 #ifdef CONFIG_SCHED_MC
6698 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6699 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6700
6701 static int
6702 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6703 struct sched_group **sg, struct cpumask *mask)
6704 {
6705 int group;
6706 #ifdef CONFIG_SCHED_SMT
6707 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6708 group = cpumask_first(mask);
6709 #else
6710 group = cpu;
6711 #endif
6712 if (sg)
6713 *sg = &per_cpu(sched_group_core, group).sg;
6714 return group;
6715 }
6716 #endif /* CONFIG_SCHED_MC */
6717
6718 /*
6719 * book sched-domains:
6720 */
6721 #ifdef CONFIG_SCHED_BOOK
6722 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6723 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6724
6725 static int
6726 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6727 struct sched_group **sg, struct cpumask *mask)
6728 {
6729 int group = cpu;
6730 #ifdef CONFIG_SCHED_MC
6731 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6732 group = cpumask_first(mask);
6733 #elif defined(CONFIG_SCHED_SMT)
6734 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6735 group = cpumask_first(mask);
6736 #endif
6737 if (sg)
6738 *sg = &per_cpu(sched_group_book, group).sg;
6739 return group;
6740 }
6741 #endif /* CONFIG_SCHED_BOOK */
6742
6743 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6744 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6745
6746 static int
6747 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6748 struct sched_group **sg, struct cpumask *mask)
6749 {
6750 int group;
6751 #ifdef CONFIG_SCHED_BOOK
6752 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6753 group = cpumask_first(mask);
6754 #elif defined(CONFIG_SCHED_MC)
6755 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6756 group = cpumask_first(mask);
6757 #elif defined(CONFIG_SCHED_SMT)
6758 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6759 group = cpumask_first(mask);
6760 #else
6761 group = cpu;
6762 #endif
6763 if (sg)
6764 *sg = &per_cpu(sched_group_phys, group).sg;
6765 return group;
6766 }
6767
6768 #ifdef CONFIG_NUMA
6769 /*
6770 * The init_sched_build_groups can't handle what we want to do with node
6771 * groups, so roll our own. Now each node has its own list of groups which
6772 * gets dynamically allocated.
6773 */
6774 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6775 static struct sched_group ***sched_group_nodes_bycpu;
6776
6777 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6778 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6779
6780 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6781 struct sched_group **sg,
6782 struct cpumask *nodemask)
6783 {
6784 int group;
6785
6786 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6787 group = cpumask_first(nodemask);
6788
6789 if (sg)
6790 *sg = &per_cpu(sched_group_allnodes, group).sg;
6791 return group;
6792 }
6793
6794 static void init_numa_sched_groups_power(struct sched_group *group_head)
6795 {
6796 struct sched_group *sg = group_head;
6797 int j;
6798
6799 if (!sg)
6800 return;
6801 do {
6802 for_each_cpu(j, sched_group_cpus(sg)) {
6803 struct sched_domain *sd;
6804
6805 sd = &per_cpu(phys_domains, j).sd;
6806 if (j != group_first_cpu(sd->groups)) {
6807 /*
6808 * Only add "power" once for each
6809 * physical package.
6810 */
6811 continue;
6812 }
6813
6814 sg->cpu_power += sd->groups->cpu_power;
6815 }
6816 sg = sg->next;
6817 } while (sg != group_head);
6818 }
6819
6820 static int build_numa_sched_groups(struct s_data *d,
6821 const struct cpumask *cpu_map, int num)
6822 {
6823 struct sched_domain *sd;
6824 struct sched_group *sg, *prev;
6825 int n, j;
6826
6827 cpumask_clear(d->covered);
6828 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6829 if (cpumask_empty(d->nodemask)) {
6830 d->sched_group_nodes[num] = NULL;
6831 goto out;
6832 }
6833
6834 sched_domain_node_span(num, d->domainspan);
6835 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6836
6837 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6838 GFP_KERNEL, num);
6839 if (!sg) {
6840 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6841 num);
6842 return -ENOMEM;
6843 }
6844 d->sched_group_nodes[num] = sg;
6845
6846 for_each_cpu(j, d->nodemask) {
6847 sd = &per_cpu(node_domains, j).sd;
6848 sd->groups = sg;
6849 }
6850
6851 sg->cpu_power = 0;
6852 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6853 sg->next = sg;
6854 cpumask_or(d->covered, d->covered, d->nodemask);
6855
6856 prev = sg;
6857 for (j = 0; j < nr_node_ids; j++) {
6858 n = (num + j) % nr_node_ids;
6859 cpumask_complement(d->notcovered, d->covered);
6860 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6861 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6862 if (cpumask_empty(d->tmpmask))
6863 break;
6864 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6865 if (cpumask_empty(d->tmpmask))
6866 continue;
6867 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6868 GFP_KERNEL, num);
6869 if (!sg) {
6870 printk(KERN_WARNING
6871 "Can not alloc domain group for node %d\n", j);
6872 return -ENOMEM;
6873 }
6874 sg->cpu_power = 0;
6875 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6876 sg->next = prev->next;
6877 cpumask_or(d->covered, d->covered, d->tmpmask);
6878 prev->next = sg;
6879 prev = sg;
6880 }
6881 out:
6882 return 0;
6883 }
6884 #endif /* CONFIG_NUMA */
6885
6886 #ifdef CONFIG_NUMA
6887 /* Free memory allocated for various sched_group structures */
6888 static void free_sched_groups(const struct cpumask *cpu_map,
6889 struct cpumask *nodemask)
6890 {
6891 int cpu, i;
6892
6893 for_each_cpu(cpu, cpu_map) {
6894 struct sched_group **sched_group_nodes
6895 = sched_group_nodes_bycpu[cpu];
6896
6897 if (!sched_group_nodes)
6898 continue;
6899
6900 for (i = 0; i < nr_node_ids; i++) {
6901 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6902
6903 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6904 if (cpumask_empty(nodemask))
6905 continue;
6906
6907 if (sg == NULL)
6908 continue;
6909 sg = sg->next;
6910 next_sg:
6911 oldsg = sg;
6912 sg = sg->next;
6913 kfree(oldsg);
6914 if (oldsg != sched_group_nodes[i])
6915 goto next_sg;
6916 }
6917 kfree(sched_group_nodes);
6918 sched_group_nodes_bycpu[cpu] = NULL;
6919 }
6920 }
6921 #else /* !CONFIG_NUMA */
6922 static void free_sched_groups(const struct cpumask *cpu_map,
6923 struct cpumask *nodemask)
6924 {
6925 }
6926 #endif /* CONFIG_NUMA */
6927
6928 /*
6929 * Initialize sched groups cpu_power.
6930 *
6931 * cpu_power indicates the capacity of sched group, which is used while
6932 * distributing the load between different sched groups in a sched domain.
6933 * Typically cpu_power for all the groups in a sched domain will be same unless
6934 * there are asymmetries in the topology. If there are asymmetries, group
6935 * having more cpu_power will pickup more load compared to the group having
6936 * less cpu_power.
6937 */
6938 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6939 {
6940 struct sched_domain *child;
6941 struct sched_group *group;
6942 long power;
6943 int weight;
6944
6945 WARN_ON(!sd || !sd->groups);
6946
6947 if (cpu != group_first_cpu(sd->groups))
6948 return;
6949
6950 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
6951
6952 child = sd->child;
6953
6954 sd->groups->cpu_power = 0;
6955
6956 if (!child) {
6957 power = SCHED_LOAD_SCALE;
6958 weight = cpumask_weight(sched_domain_span(sd));
6959 /*
6960 * SMT siblings share the power of a single core.
6961 * Usually multiple threads get a better yield out of
6962 * that one core than a single thread would have,
6963 * reflect that in sd->smt_gain.
6964 */
6965 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6966 power *= sd->smt_gain;
6967 power /= weight;
6968 power >>= SCHED_LOAD_SHIFT;
6969 }
6970 sd->groups->cpu_power += power;
6971 return;
6972 }
6973
6974 /*
6975 * Add cpu_power of each child group to this groups cpu_power.
6976 */
6977 group = child->groups;
6978 do {
6979 sd->groups->cpu_power += group->cpu_power;
6980 group = group->next;
6981 } while (group != child->groups);
6982 }
6983
6984 /*
6985 * Initializers for schedule domains
6986 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6987 */
6988
6989 #ifdef CONFIG_SCHED_DEBUG
6990 # define SD_INIT_NAME(sd, type) sd->name = #type
6991 #else
6992 # define SD_INIT_NAME(sd, type) do { } while (0)
6993 #endif
6994
6995 #define SD_INIT(sd, type) sd_init_##type(sd)
6996
6997 #define SD_INIT_FUNC(type) \
6998 static noinline void sd_init_##type(struct sched_domain *sd) \
6999 { \
7000 memset(sd, 0, sizeof(*sd)); \
7001 *sd = SD_##type##_INIT; \
7002 sd->level = SD_LV_##type; \
7003 SD_INIT_NAME(sd, type); \
7004 }
7005
7006 SD_INIT_FUNC(CPU)
7007 #ifdef CONFIG_NUMA
7008 SD_INIT_FUNC(ALLNODES)
7009 SD_INIT_FUNC(NODE)
7010 #endif
7011 #ifdef CONFIG_SCHED_SMT
7012 SD_INIT_FUNC(SIBLING)
7013 #endif
7014 #ifdef CONFIG_SCHED_MC
7015 SD_INIT_FUNC(MC)
7016 #endif
7017 #ifdef CONFIG_SCHED_BOOK
7018 SD_INIT_FUNC(BOOK)
7019 #endif
7020
7021 static int default_relax_domain_level = -1;
7022
7023 static int __init setup_relax_domain_level(char *str)
7024 {
7025 unsigned long val;
7026
7027 val = simple_strtoul(str, NULL, 0);
7028 if (val < SD_LV_MAX)
7029 default_relax_domain_level = val;
7030
7031 return 1;
7032 }
7033 __setup("relax_domain_level=", setup_relax_domain_level);
7034
7035 static void set_domain_attribute(struct sched_domain *sd,
7036 struct sched_domain_attr *attr)
7037 {
7038 int request;
7039
7040 if (!attr || attr->relax_domain_level < 0) {
7041 if (default_relax_domain_level < 0)
7042 return;
7043 else
7044 request = default_relax_domain_level;
7045 } else
7046 request = attr->relax_domain_level;
7047 if (request < sd->level) {
7048 /* turn off idle balance on this domain */
7049 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7050 } else {
7051 /* turn on idle balance on this domain */
7052 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7053 }
7054 }
7055
7056 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7057 const struct cpumask *cpu_map)
7058 {
7059 switch (what) {
7060 case sa_sched_groups:
7061 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
7062 d->sched_group_nodes = NULL;
7063 case sa_rootdomain:
7064 free_rootdomain(d->rd); /* fall through */
7065 case sa_tmpmask:
7066 free_cpumask_var(d->tmpmask); /* fall through */
7067 case sa_send_covered:
7068 free_cpumask_var(d->send_covered); /* fall through */
7069 case sa_this_book_map:
7070 free_cpumask_var(d->this_book_map); /* fall through */
7071 case sa_this_core_map:
7072 free_cpumask_var(d->this_core_map); /* fall through */
7073 case sa_this_sibling_map:
7074 free_cpumask_var(d->this_sibling_map); /* fall through */
7075 case sa_nodemask:
7076 free_cpumask_var(d->nodemask); /* fall through */
7077 case sa_sched_group_nodes:
7078 #ifdef CONFIG_NUMA
7079 kfree(d->sched_group_nodes); /* fall through */
7080 case sa_notcovered:
7081 free_cpumask_var(d->notcovered); /* fall through */
7082 case sa_covered:
7083 free_cpumask_var(d->covered); /* fall through */
7084 case sa_domainspan:
7085 free_cpumask_var(d->domainspan); /* fall through */
7086 #endif
7087 case sa_none:
7088 break;
7089 }
7090 }
7091
7092 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7093 const struct cpumask *cpu_map)
7094 {
7095 #ifdef CONFIG_NUMA
7096 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
7097 return sa_none;
7098 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
7099 return sa_domainspan;
7100 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
7101 return sa_covered;
7102 /* Allocate the per-node list of sched groups */
7103 d->sched_group_nodes = kcalloc(nr_node_ids,
7104 sizeof(struct sched_group *), GFP_KERNEL);
7105 if (!d->sched_group_nodes) {
7106 printk(KERN_WARNING "Can not alloc sched group node list\n");
7107 return sa_notcovered;
7108 }
7109 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
7110 #endif
7111 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
7112 return sa_sched_group_nodes;
7113 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
7114 return sa_nodemask;
7115 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
7116 return sa_this_sibling_map;
7117 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
7118 return sa_this_core_map;
7119 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
7120 return sa_this_book_map;
7121 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
7122 return sa_send_covered;
7123 d->rd = alloc_rootdomain();
7124 if (!d->rd) {
7125 printk(KERN_WARNING "Cannot alloc root domain\n");
7126 return sa_tmpmask;
7127 }
7128 return sa_rootdomain;
7129 }
7130
7131 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
7132 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
7133 {
7134 struct sched_domain *sd = NULL;
7135 #ifdef CONFIG_NUMA
7136 struct sched_domain *parent;
7137
7138 d->sd_allnodes = 0;
7139 if (cpumask_weight(cpu_map) >
7140 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
7141 sd = &per_cpu(allnodes_domains, i).sd;
7142 SD_INIT(sd, ALLNODES);
7143 set_domain_attribute(sd, attr);
7144 cpumask_copy(sched_domain_span(sd), cpu_map);
7145 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7146 d->sd_allnodes = 1;
7147 }
7148 parent = sd;
7149
7150 sd = &per_cpu(node_domains, i).sd;
7151 SD_INIT(sd, NODE);
7152 set_domain_attribute(sd, attr);
7153 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7154 sd->parent = parent;
7155 if (parent)
7156 parent->child = sd;
7157 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7158 #endif
7159 return sd;
7160 }
7161
7162 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7163 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7164 struct sched_domain *parent, int i)
7165 {
7166 struct sched_domain *sd;
7167 sd = &per_cpu(phys_domains, i).sd;
7168 SD_INIT(sd, CPU);
7169 set_domain_attribute(sd, attr);
7170 cpumask_copy(sched_domain_span(sd), d->nodemask);
7171 sd->parent = parent;
7172 if (parent)
7173 parent->child = sd;
7174 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7175 return sd;
7176 }
7177
7178 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
7179 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7180 struct sched_domain *parent, int i)
7181 {
7182 struct sched_domain *sd = parent;
7183 #ifdef CONFIG_SCHED_BOOK
7184 sd = &per_cpu(book_domains, i).sd;
7185 SD_INIT(sd, BOOK);
7186 set_domain_attribute(sd, attr);
7187 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7188 sd->parent = parent;
7189 parent->child = sd;
7190 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7191 #endif
7192 return sd;
7193 }
7194
7195 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7196 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7197 struct sched_domain *parent, int i)
7198 {
7199 struct sched_domain *sd = parent;
7200 #ifdef CONFIG_SCHED_MC
7201 sd = &per_cpu(core_domains, i).sd;
7202 SD_INIT(sd, MC);
7203 set_domain_attribute(sd, attr);
7204 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7205 sd->parent = parent;
7206 parent->child = sd;
7207 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7208 #endif
7209 return sd;
7210 }
7211
7212 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7213 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7214 struct sched_domain *parent, int i)
7215 {
7216 struct sched_domain *sd = parent;
7217 #ifdef CONFIG_SCHED_SMT
7218 sd = &per_cpu(cpu_domains, i).sd;
7219 SD_INIT(sd, SIBLING);
7220 set_domain_attribute(sd, attr);
7221 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7222 sd->parent = parent;
7223 parent->child = sd;
7224 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7225 #endif
7226 return sd;
7227 }
7228
7229 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7230 const struct cpumask *cpu_map, int cpu)
7231 {
7232 switch (l) {
7233 #ifdef CONFIG_SCHED_SMT
7234 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7235 cpumask_and(d->this_sibling_map, cpu_map,
7236 topology_thread_cpumask(cpu));
7237 if (cpu == cpumask_first(d->this_sibling_map))
7238 init_sched_build_groups(d->this_sibling_map, cpu_map,
7239 &cpu_to_cpu_group,
7240 d->send_covered, d->tmpmask);
7241 break;
7242 #endif
7243 #ifdef CONFIG_SCHED_MC
7244 case SD_LV_MC: /* set up multi-core groups */
7245 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7246 if (cpu == cpumask_first(d->this_core_map))
7247 init_sched_build_groups(d->this_core_map, cpu_map,
7248 &cpu_to_core_group,
7249 d->send_covered, d->tmpmask);
7250 break;
7251 #endif
7252 #ifdef CONFIG_SCHED_BOOK
7253 case SD_LV_BOOK: /* set up book groups */
7254 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7255 if (cpu == cpumask_first(d->this_book_map))
7256 init_sched_build_groups(d->this_book_map, cpu_map,
7257 &cpu_to_book_group,
7258 d->send_covered, d->tmpmask);
7259 break;
7260 #endif
7261 case SD_LV_CPU: /* set up physical groups */
7262 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7263 if (!cpumask_empty(d->nodemask))
7264 init_sched_build_groups(d->nodemask, cpu_map,
7265 &cpu_to_phys_group,
7266 d->send_covered, d->tmpmask);
7267 break;
7268 #ifdef CONFIG_NUMA
7269 case SD_LV_ALLNODES:
7270 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7271 d->send_covered, d->tmpmask);
7272 break;
7273 #endif
7274 default:
7275 break;
7276 }
7277 }
7278
7279 /*
7280 * Build sched domains for a given set of cpus and attach the sched domains
7281 * to the individual cpus
7282 */
7283 static int __build_sched_domains(const struct cpumask *cpu_map,
7284 struct sched_domain_attr *attr)
7285 {
7286 enum s_alloc alloc_state = sa_none;
7287 struct s_data d;
7288 struct sched_domain *sd;
7289 int i;
7290 #ifdef CONFIG_NUMA
7291 d.sd_allnodes = 0;
7292 #endif
7293
7294 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7295 if (alloc_state != sa_rootdomain)
7296 goto error;
7297 alloc_state = sa_sched_groups;
7298
7299 /*
7300 * Set up domains for cpus specified by the cpu_map.
7301 */
7302 for_each_cpu(i, cpu_map) {
7303 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7304 cpu_map);
7305
7306 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7307 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7308 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7309 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7310 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7311 }
7312
7313 for_each_cpu(i, cpu_map) {
7314 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7315 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7316 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7317 }
7318
7319 /* Set up physical groups */
7320 for (i = 0; i < nr_node_ids; i++)
7321 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7322
7323 #ifdef CONFIG_NUMA
7324 /* Set up node groups */
7325 if (d.sd_allnodes)
7326 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7327
7328 for (i = 0; i < nr_node_ids; i++)
7329 if (build_numa_sched_groups(&d, cpu_map, i))
7330 goto error;
7331 #endif
7332
7333 /* Calculate CPU power for physical packages and nodes */
7334 #ifdef CONFIG_SCHED_SMT
7335 for_each_cpu(i, cpu_map) {
7336 sd = &per_cpu(cpu_domains, i).sd;
7337 init_sched_groups_power(i, sd);
7338 }
7339 #endif
7340 #ifdef CONFIG_SCHED_MC
7341 for_each_cpu(i, cpu_map) {
7342 sd = &per_cpu(core_domains, i).sd;
7343 init_sched_groups_power(i, sd);
7344 }
7345 #endif
7346 #ifdef CONFIG_SCHED_BOOK
7347 for_each_cpu(i, cpu_map) {
7348 sd = &per_cpu(book_domains, i).sd;
7349 init_sched_groups_power(i, sd);
7350 }
7351 #endif
7352
7353 for_each_cpu(i, cpu_map) {
7354 sd = &per_cpu(phys_domains, i).sd;
7355 init_sched_groups_power(i, sd);
7356 }
7357
7358 #ifdef CONFIG_NUMA
7359 for (i = 0; i < nr_node_ids; i++)
7360 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7361
7362 if (d.sd_allnodes) {
7363 struct sched_group *sg;
7364
7365 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7366 d.tmpmask);
7367 init_numa_sched_groups_power(sg);
7368 }
7369 #endif
7370
7371 /* Attach the domains */
7372 for_each_cpu(i, cpu_map) {
7373 #ifdef CONFIG_SCHED_SMT
7374 sd = &per_cpu(cpu_domains, i).sd;
7375 #elif defined(CONFIG_SCHED_MC)
7376 sd = &per_cpu(core_domains, i).sd;
7377 #elif defined(CONFIG_SCHED_BOOK)
7378 sd = &per_cpu(book_domains, i).sd;
7379 #else
7380 sd = &per_cpu(phys_domains, i).sd;
7381 #endif
7382 cpu_attach_domain(sd, d.rd, i);
7383 }
7384
7385 d.sched_group_nodes = NULL; /* don't free this we still need it */
7386 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7387 return 0;
7388
7389 error:
7390 __free_domain_allocs(&d, alloc_state, cpu_map);
7391 return -ENOMEM;
7392 }
7393
7394 static int build_sched_domains(const struct cpumask *cpu_map)
7395 {
7396 return __build_sched_domains(cpu_map, NULL);
7397 }
7398
7399 static cpumask_var_t *doms_cur; /* current sched domains */
7400 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7401 static struct sched_domain_attr *dattr_cur;
7402 /* attribues of custom domains in 'doms_cur' */
7403
7404 /*
7405 * Special case: If a kmalloc of a doms_cur partition (array of
7406 * cpumask) fails, then fallback to a single sched domain,
7407 * as determined by the single cpumask fallback_doms.
7408 */
7409 static cpumask_var_t fallback_doms;
7410
7411 /*
7412 * arch_update_cpu_topology lets virtualized architectures update the
7413 * cpu core maps. It is supposed to return 1 if the topology changed
7414 * or 0 if it stayed the same.
7415 */
7416 int __attribute__((weak)) arch_update_cpu_topology(void)
7417 {
7418 return 0;
7419 }
7420
7421 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7422 {
7423 int i;
7424 cpumask_var_t *doms;
7425
7426 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7427 if (!doms)
7428 return NULL;
7429 for (i = 0; i < ndoms; i++) {
7430 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7431 free_sched_domains(doms, i);
7432 return NULL;
7433 }
7434 }
7435 return doms;
7436 }
7437
7438 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7439 {
7440 unsigned int i;
7441 for (i = 0; i < ndoms; i++)
7442 free_cpumask_var(doms[i]);
7443 kfree(doms);
7444 }
7445
7446 /*
7447 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7448 * For now this just excludes isolated cpus, but could be used to
7449 * exclude other special cases in the future.
7450 */
7451 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7452 {
7453 int err;
7454
7455 arch_update_cpu_topology();
7456 ndoms_cur = 1;
7457 doms_cur = alloc_sched_domains(ndoms_cur);
7458 if (!doms_cur)
7459 doms_cur = &fallback_doms;
7460 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7461 dattr_cur = NULL;
7462 err = build_sched_domains(doms_cur[0]);
7463 register_sched_domain_sysctl();
7464
7465 return err;
7466 }
7467
7468 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7469 struct cpumask *tmpmask)
7470 {
7471 free_sched_groups(cpu_map, tmpmask);
7472 }
7473
7474 /*
7475 * Detach sched domains from a group of cpus specified in cpu_map
7476 * These cpus will now be attached to the NULL domain
7477 */
7478 static void detach_destroy_domains(const struct cpumask *cpu_map)
7479 {
7480 /* Save because hotplug lock held. */
7481 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7482 int i;
7483
7484 for_each_cpu(i, cpu_map)
7485 cpu_attach_domain(NULL, &def_root_domain, i);
7486 synchronize_sched();
7487 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7488 }
7489
7490 /* handle null as "default" */
7491 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7492 struct sched_domain_attr *new, int idx_new)
7493 {
7494 struct sched_domain_attr tmp;
7495
7496 /* fast path */
7497 if (!new && !cur)
7498 return 1;
7499
7500 tmp = SD_ATTR_INIT;
7501 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7502 new ? (new + idx_new) : &tmp,
7503 sizeof(struct sched_domain_attr));
7504 }
7505
7506 /*
7507 * Partition sched domains as specified by the 'ndoms_new'
7508 * cpumasks in the array doms_new[] of cpumasks. This compares
7509 * doms_new[] to the current sched domain partitioning, doms_cur[].
7510 * It destroys each deleted domain and builds each new domain.
7511 *
7512 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7513 * The masks don't intersect (don't overlap.) We should setup one
7514 * sched domain for each mask. CPUs not in any of the cpumasks will
7515 * not be load balanced. If the same cpumask appears both in the
7516 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7517 * it as it is.
7518 *
7519 * The passed in 'doms_new' should be allocated using
7520 * alloc_sched_domains. This routine takes ownership of it and will
7521 * free_sched_domains it when done with it. If the caller failed the
7522 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7523 * and partition_sched_domains() will fallback to the single partition
7524 * 'fallback_doms', it also forces the domains to be rebuilt.
7525 *
7526 * If doms_new == NULL it will be replaced with cpu_online_mask.
7527 * ndoms_new == 0 is a special case for destroying existing domains,
7528 * and it will not create the default domain.
7529 *
7530 * Call with hotplug lock held
7531 */
7532 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7533 struct sched_domain_attr *dattr_new)
7534 {
7535 int i, j, n;
7536 int new_topology;
7537
7538 mutex_lock(&sched_domains_mutex);
7539
7540 /* always unregister in case we don't destroy any domains */
7541 unregister_sched_domain_sysctl();
7542
7543 /* Let architecture update cpu core mappings. */
7544 new_topology = arch_update_cpu_topology();
7545
7546 n = doms_new ? ndoms_new : 0;
7547
7548 /* Destroy deleted domains */
7549 for (i = 0; i < ndoms_cur; i++) {
7550 for (j = 0; j < n && !new_topology; j++) {
7551 if (cpumask_equal(doms_cur[i], doms_new[j])
7552 && dattrs_equal(dattr_cur, i, dattr_new, j))
7553 goto match1;
7554 }
7555 /* no match - a current sched domain not in new doms_new[] */
7556 detach_destroy_domains(doms_cur[i]);
7557 match1:
7558 ;
7559 }
7560
7561 if (doms_new == NULL) {
7562 ndoms_cur = 0;
7563 doms_new = &fallback_doms;
7564 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7565 WARN_ON_ONCE(dattr_new);
7566 }
7567
7568 /* Build new domains */
7569 for (i = 0; i < ndoms_new; i++) {
7570 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7571 if (cpumask_equal(doms_new[i], doms_cur[j])
7572 && dattrs_equal(dattr_new, i, dattr_cur, j))
7573 goto match2;
7574 }
7575 /* no match - add a new doms_new */
7576 __build_sched_domains(doms_new[i],
7577 dattr_new ? dattr_new + i : NULL);
7578 match2:
7579 ;
7580 }
7581
7582 /* Remember the new sched domains */
7583 if (doms_cur != &fallback_doms)
7584 free_sched_domains(doms_cur, ndoms_cur);
7585 kfree(dattr_cur); /* kfree(NULL) is safe */
7586 doms_cur = doms_new;
7587 dattr_cur = dattr_new;
7588 ndoms_cur = ndoms_new;
7589
7590 register_sched_domain_sysctl();
7591
7592 mutex_unlock(&sched_domains_mutex);
7593 }
7594
7595 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7596 static void arch_reinit_sched_domains(void)
7597 {
7598 get_online_cpus();
7599
7600 /* Destroy domains first to force the rebuild */
7601 partition_sched_domains(0, NULL, NULL);
7602
7603 rebuild_sched_domains();
7604 put_online_cpus();
7605 }
7606
7607 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7608 {
7609 unsigned int level = 0;
7610
7611 if (sscanf(buf, "%u", &level) != 1)
7612 return -EINVAL;
7613
7614 /*
7615 * level is always be positive so don't check for
7616 * level < POWERSAVINGS_BALANCE_NONE which is 0
7617 * What happens on 0 or 1 byte write,
7618 * need to check for count as well?
7619 */
7620
7621 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7622 return -EINVAL;
7623
7624 if (smt)
7625 sched_smt_power_savings = level;
7626 else
7627 sched_mc_power_savings = level;
7628
7629 arch_reinit_sched_domains();
7630
7631 return count;
7632 }
7633
7634 #ifdef CONFIG_SCHED_MC
7635 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7636 struct sysdev_class_attribute *attr,
7637 char *page)
7638 {
7639 return sprintf(page, "%u\n", sched_mc_power_savings);
7640 }
7641 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7642 struct sysdev_class_attribute *attr,
7643 const char *buf, size_t count)
7644 {
7645 return sched_power_savings_store(buf, count, 0);
7646 }
7647 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7648 sched_mc_power_savings_show,
7649 sched_mc_power_savings_store);
7650 #endif
7651
7652 #ifdef CONFIG_SCHED_SMT
7653 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7654 struct sysdev_class_attribute *attr,
7655 char *page)
7656 {
7657 return sprintf(page, "%u\n", sched_smt_power_savings);
7658 }
7659 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7660 struct sysdev_class_attribute *attr,
7661 const char *buf, size_t count)
7662 {
7663 return sched_power_savings_store(buf, count, 1);
7664 }
7665 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7666 sched_smt_power_savings_show,
7667 sched_smt_power_savings_store);
7668 #endif
7669
7670 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7671 {
7672 int err = 0;
7673
7674 #ifdef CONFIG_SCHED_SMT
7675 if (smt_capable())
7676 err = sysfs_create_file(&cls->kset.kobj,
7677 &attr_sched_smt_power_savings.attr);
7678 #endif
7679 #ifdef CONFIG_SCHED_MC
7680 if (!err && mc_capable())
7681 err = sysfs_create_file(&cls->kset.kobj,
7682 &attr_sched_mc_power_savings.attr);
7683 #endif
7684 return err;
7685 }
7686 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7687
7688 /*
7689 * Update cpusets according to cpu_active mask. If cpusets are
7690 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7691 * around partition_sched_domains().
7692 */
7693 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7694 void *hcpu)
7695 {
7696 switch (action & ~CPU_TASKS_FROZEN) {
7697 case CPU_ONLINE:
7698 case CPU_DOWN_FAILED:
7699 cpuset_update_active_cpus();
7700 return NOTIFY_OK;
7701 default:
7702 return NOTIFY_DONE;
7703 }
7704 }
7705
7706 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7707 void *hcpu)
7708 {
7709 switch (action & ~CPU_TASKS_FROZEN) {
7710 case CPU_DOWN_PREPARE:
7711 cpuset_update_active_cpus();
7712 return NOTIFY_OK;
7713 default:
7714 return NOTIFY_DONE;
7715 }
7716 }
7717
7718 static int update_runtime(struct notifier_block *nfb,
7719 unsigned long action, void *hcpu)
7720 {
7721 int cpu = (int)(long)hcpu;
7722
7723 switch (action) {
7724 case CPU_DOWN_PREPARE:
7725 case CPU_DOWN_PREPARE_FROZEN:
7726 disable_runtime(cpu_rq(cpu));
7727 return NOTIFY_OK;
7728
7729 case CPU_DOWN_FAILED:
7730 case CPU_DOWN_FAILED_FROZEN:
7731 case CPU_ONLINE:
7732 case CPU_ONLINE_FROZEN:
7733 enable_runtime(cpu_rq(cpu));
7734 return NOTIFY_OK;
7735
7736 default:
7737 return NOTIFY_DONE;
7738 }
7739 }
7740
7741 void __init sched_init_smp(void)
7742 {
7743 cpumask_var_t non_isolated_cpus;
7744
7745 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7746 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7747
7748 #if defined(CONFIG_NUMA)
7749 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7750 GFP_KERNEL);
7751 BUG_ON(sched_group_nodes_bycpu == NULL);
7752 #endif
7753 get_online_cpus();
7754 mutex_lock(&sched_domains_mutex);
7755 arch_init_sched_domains(cpu_active_mask);
7756 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7757 if (cpumask_empty(non_isolated_cpus))
7758 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7759 mutex_unlock(&sched_domains_mutex);
7760 put_online_cpus();
7761
7762 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7763 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7764
7765 /* RT runtime code needs to handle some hotplug events */
7766 hotcpu_notifier(update_runtime, 0);
7767
7768 init_hrtick();
7769
7770 /* Move init over to a non-isolated CPU */
7771 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7772 BUG();
7773 sched_init_granularity();
7774 free_cpumask_var(non_isolated_cpus);
7775
7776 init_sched_rt_class();
7777 }
7778 #else
7779 void __init sched_init_smp(void)
7780 {
7781 sched_init_granularity();
7782 }
7783 #endif /* CONFIG_SMP */
7784
7785 const_debug unsigned int sysctl_timer_migration = 1;
7786
7787 int in_sched_functions(unsigned long addr)
7788 {
7789 return in_lock_functions(addr) ||
7790 (addr >= (unsigned long)__sched_text_start
7791 && addr < (unsigned long)__sched_text_end);
7792 }
7793
7794 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7795 {
7796 cfs_rq->tasks_timeline = RB_ROOT;
7797 INIT_LIST_HEAD(&cfs_rq->tasks);
7798 #ifdef CONFIG_FAIR_GROUP_SCHED
7799 cfs_rq->rq = rq;
7800 #endif
7801 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7802 }
7803
7804 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7805 {
7806 struct rt_prio_array *array;
7807 int i;
7808
7809 array = &rt_rq->active;
7810 for (i = 0; i < MAX_RT_PRIO; i++) {
7811 INIT_LIST_HEAD(array->queue + i);
7812 __clear_bit(i, array->bitmap);
7813 }
7814 /* delimiter for bitsearch: */
7815 __set_bit(MAX_RT_PRIO, array->bitmap);
7816
7817 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7818 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7819 #ifdef CONFIG_SMP
7820 rt_rq->highest_prio.next = MAX_RT_PRIO;
7821 #endif
7822 #endif
7823 #ifdef CONFIG_SMP
7824 rt_rq->rt_nr_migratory = 0;
7825 rt_rq->overloaded = 0;
7826 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7827 #endif
7828
7829 rt_rq->rt_time = 0;
7830 rt_rq->rt_throttled = 0;
7831 rt_rq->rt_runtime = 0;
7832 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7833
7834 #ifdef CONFIG_RT_GROUP_SCHED
7835 rt_rq->rt_nr_boosted = 0;
7836 rt_rq->rq = rq;
7837 #endif
7838 }
7839
7840 #ifdef CONFIG_FAIR_GROUP_SCHED
7841 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7842 struct sched_entity *se, int cpu,
7843 struct sched_entity *parent)
7844 {
7845 struct rq *rq = cpu_rq(cpu);
7846 tg->cfs_rq[cpu] = cfs_rq;
7847 init_cfs_rq(cfs_rq, rq);
7848 cfs_rq->tg = tg;
7849
7850 tg->se[cpu] = se;
7851 /* se could be NULL for root_task_group */
7852 if (!se)
7853 return;
7854
7855 if (!parent)
7856 se->cfs_rq = &rq->cfs;
7857 else
7858 se->cfs_rq = parent->my_q;
7859
7860 se->my_q = cfs_rq;
7861 update_load_set(&se->load, 0);
7862 se->parent = parent;
7863 }
7864 #endif
7865
7866 #ifdef CONFIG_RT_GROUP_SCHED
7867 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7868 struct sched_rt_entity *rt_se, int cpu,
7869 struct sched_rt_entity *parent)
7870 {
7871 struct rq *rq = cpu_rq(cpu);
7872
7873 tg->rt_rq[cpu] = rt_rq;
7874 init_rt_rq(rt_rq, rq);
7875 rt_rq->tg = tg;
7876 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7877
7878 tg->rt_se[cpu] = rt_se;
7879 if (!rt_se)
7880 return;
7881
7882 if (!parent)
7883 rt_se->rt_rq = &rq->rt;
7884 else
7885 rt_se->rt_rq = parent->my_q;
7886
7887 rt_se->my_q = rt_rq;
7888 rt_se->parent = parent;
7889 INIT_LIST_HEAD(&rt_se->run_list);
7890 }
7891 #endif
7892
7893 void __init sched_init(void)
7894 {
7895 int i, j;
7896 unsigned long alloc_size = 0, ptr;
7897
7898 #ifdef CONFIG_FAIR_GROUP_SCHED
7899 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7900 #endif
7901 #ifdef CONFIG_RT_GROUP_SCHED
7902 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7903 #endif
7904 #ifdef CONFIG_CPUMASK_OFFSTACK
7905 alloc_size += num_possible_cpus() * cpumask_size();
7906 #endif
7907 if (alloc_size) {
7908 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7909
7910 #ifdef CONFIG_FAIR_GROUP_SCHED
7911 root_task_group.se = (struct sched_entity **)ptr;
7912 ptr += nr_cpu_ids * sizeof(void **);
7913
7914 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7915 ptr += nr_cpu_ids * sizeof(void **);
7916
7917 #endif /* CONFIG_FAIR_GROUP_SCHED */
7918 #ifdef CONFIG_RT_GROUP_SCHED
7919 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7920 ptr += nr_cpu_ids * sizeof(void **);
7921
7922 root_task_group.rt_rq = (struct rt_rq **)ptr;
7923 ptr += nr_cpu_ids * sizeof(void **);
7924
7925 #endif /* CONFIG_RT_GROUP_SCHED */
7926 #ifdef CONFIG_CPUMASK_OFFSTACK
7927 for_each_possible_cpu(i) {
7928 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7929 ptr += cpumask_size();
7930 }
7931 #endif /* CONFIG_CPUMASK_OFFSTACK */
7932 }
7933
7934 #ifdef CONFIG_SMP
7935 init_defrootdomain();
7936 #endif
7937
7938 init_rt_bandwidth(&def_rt_bandwidth,
7939 global_rt_period(), global_rt_runtime());
7940
7941 #ifdef CONFIG_RT_GROUP_SCHED
7942 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7943 global_rt_period(), global_rt_runtime());
7944 #endif /* CONFIG_RT_GROUP_SCHED */
7945
7946 #ifdef CONFIG_CGROUP_SCHED
7947 list_add(&root_task_group.list, &task_groups);
7948 INIT_LIST_HEAD(&root_task_group.children);
7949 autogroup_init(&init_task);
7950 #endif /* CONFIG_CGROUP_SCHED */
7951
7952 for_each_possible_cpu(i) {
7953 struct rq *rq;
7954
7955 rq = cpu_rq(i);
7956 raw_spin_lock_init(&rq->lock);
7957 rq->nr_running = 0;
7958 rq->calc_load_active = 0;
7959 rq->calc_load_update = jiffies + LOAD_FREQ;
7960 init_cfs_rq(&rq->cfs, rq);
7961 init_rt_rq(&rq->rt, rq);
7962 #ifdef CONFIG_FAIR_GROUP_SCHED
7963 root_task_group.shares = root_task_group_load;
7964 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7965 /*
7966 * How much cpu bandwidth does root_task_group get?
7967 *
7968 * In case of task-groups formed thr' the cgroup filesystem, it
7969 * gets 100% of the cpu resources in the system. This overall
7970 * system cpu resource is divided among the tasks of
7971 * root_task_group and its child task-groups in a fair manner,
7972 * based on each entity's (task or task-group's) weight
7973 * (se->load.weight).
7974 *
7975 * In other words, if root_task_group has 10 tasks of weight
7976 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7977 * then A0's share of the cpu resource is:
7978 *
7979 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7980 *
7981 * We achieve this by letting root_task_group's tasks sit
7982 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7983 */
7984 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7985 #endif /* CONFIG_FAIR_GROUP_SCHED */
7986
7987 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7988 #ifdef CONFIG_RT_GROUP_SCHED
7989 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7990 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7991 #endif
7992
7993 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7994 rq->cpu_load[j] = 0;
7995
7996 rq->last_load_update_tick = jiffies;
7997
7998 #ifdef CONFIG_SMP
7999 rq->sd = NULL;
8000 rq->rd = NULL;
8001 rq->cpu_power = SCHED_LOAD_SCALE;
8002 rq->post_schedule = 0;
8003 rq->active_balance = 0;
8004 rq->next_balance = jiffies;
8005 rq->push_cpu = 0;
8006 rq->cpu = i;
8007 rq->online = 0;
8008 rq->idle_stamp = 0;
8009 rq->avg_idle = 2*sysctl_sched_migration_cost;
8010 rq_attach_root(rq, &def_root_domain);
8011 #ifdef CONFIG_NO_HZ
8012 rq->nohz_balance_kick = 0;
8013 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8014 #endif
8015 #endif
8016 init_rq_hrtick(rq);
8017 atomic_set(&rq->nr_iowait, 0);
8018 }
8019
8020 set_load_weight(&init_task);
8021
8022 #ifdef CONFIG_PREEMPT_NOTIFIERS
8023 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8024 #endif
8025
8026 #ifdef CONFIG_SMP
8027 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8028 #endif
8029
8030 #ifdef CONFIG_RT_MUTEXES
8031 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
8032 #endif
8033
8034 /*
8035 * The boot idle thread does lazy MMU switching as well:
8036 */
8037 atomic_inc(&init_mm.mm_count);
8038 enter_lazy_tlb(&init_mm, current);
8039
8040 /*
8041 * Make us the idle thread. Technically, schedule() should not be
8042 * called from this thread, however somewhere below it might be,
8043 * but because we are the idle thread, we just pick up running again
8044 * when this runqueue becomes "idle".
8045 */
8046 init_idle(current, smp_processor_id());
8047
8048 calc_load_update = jiffies + LOAD_FREQ;
8049
8050 /*
8051 * During early bootup we pretend to be a normal task:
8052 */
8053 current->sched_class = &fair_sched_class;
8054
8055 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8056 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8057 #ifdef CONFIG_SMP
8058 #ifdef CONFIG_NO_HZ
8059 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8060 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8061 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8062 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8063 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8064 #endif
8065 /* May be allocated at isolcpus cmdline parse time */
8066 if (cpu_isolated_map == NULL)
8067 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8068 #endif /* SMP */
8069
8070 scheduler_running = 1;
8071 }
8072
8073 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8074 static inline int preempt_count_equals(int preempt_offset)
8075 {
8076 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8077
8078 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
8079 }
8080
8081 void __might_sleep(const char *file, int line, int preempt_offset)
8082 {
8083 #ifdef in_atomic
8084 static unsigned long prev_jiffy; /* ratelimiting */
8085
8086 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8087 system_state != SYSTEM_RUNNING || oops_in_progress)
8088 return;
8089 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8090 return;
8091 prev_jiffy = jiffies;
8092
8093 printk(KERN_ERR
8094 "BUG: sleeping function called from invalid context at %s:%d\n",
8095 file, line);
8096 printk(KERN_ERR
8097 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8098 in_atomic(), irqs_disabled(),
8099 current->pid, current->comm);
8100
8101 debug_show_held_locks(current);
8102 if (irqs_disabled())
8103 print_irqtrace_events(current);
8104 dump_stack();
8105 #endif
8106 }
8107 EXPORT_SYMBOL(__might_sleep);
8108 #endif
8109
8110 #ifdef CONFIG_MAGIC_SYSRQ
8111 static void normalize_task(struct rq *rq, struct task_struct *p)
8112 {
8113 int on_rq;
8114
8115 on_rq = p->se.on_rq;
8116 if (on_rq)
8117 deactivate_task(rq, p, 0);
8118 __setscheduler(rq, p, SCHED_NORMAL, 0);
8119 if (on_rq) {
8120 activate_task(rq, p, 0);
8121 resched_task(rq->curr);
8122 }
8123 }
8124
8125 void normalize_rt_tasks(void)
8126 {
8127 struct task_struct *g, *p;
8128 unsigned long flags;
8129 struct rq *rq;
8130
8131 read_lock_irqsave(&tasklist_lock, flags);
8132 do_each_thread(g, p) {
8133 /*
8134 * Only normalize user tasks:
8135 */
8136 if (!p->mm)
8137 continue;
8138
8139 p->se.exec_start = 0;
8140 #ifdef CONFIG_SCHEDSTATS
8141 p->se.statistics.wait_start = 0;
8142 p->se.statistics.sleep_start = 0;
8143 p->se.statistics.block_start = 0;
8144 #endif
8145
8146 if (!rt_task(p)) {
8147 /*
8148 * Renice negative nice level userspace
8149 * tasks back to 0:
8150 */
8151 if (TASK_NICE(p) < 0 && p->mm)
8152 set_user_nice(p, 0);
8153 continue;
8154 }
8155
8156 raw_spin_lock(&p->pi_lock);
8157 rq = __task_rq_lock(p);
8158
8159 normalize_task(rq, p);
8160
8161 __task_rq_unlock(rq);
8162 raw_spin_unlock(&p->pi_lock);
8163 } while_each_thread(g, p);
8164
8165 read_unlock_irqrestore(&tasklist_lock, flags);
8166 }
8167
8168 #endif /* CONFIG_MAGIC_SYSRQ */
8169
8170 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8171 /*
8172 * These functions are only useful for the IA64 MCA handling, or kdb.
8173 *
8174 * They can only be called when the whole system has been
8175 * stopped - every CPU needs to be quiescent, and no scheduling
8176 * activity can take place. Using them for anything else would
8177 * be a serious bug, and as a result, they aren't even visible
8178 * under any other configuration.
8179 */
8180
8181 /**
8182 * curr_task - return the current task for a given cpu.
8183 * @cpu: the processor in question.
8184 *
8185 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8186 */
8187 struct task_struct *curr_task(int cpu)
8188 {
8189 return cpu_curr(cpu);
8190 }
8191
8192 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8193
8194 #ifdef CONFIG_IA64
8195 /**
8196 * set_curr_task - set the current task for a given cpu.
8197 * @cpu: the processor in question.
8198 * @p: the task pointer to set.
8199 *
8200 * Description: This function must only be used when non-maskable interrupts
8201 * are serviced on a separate stack. It allows the architecture to switch the
8202 * notion of the current task on a cpu in a non-blocking manner. This function
8203 * must be called with all CPU's synchronized, and interrupts disabled, the
8204 * and caller must save the original value of the current task (see
8205 * curr_task() above) and restore that value before reenabling interrupts and
8206 * re-starting the system.
8207 *
8208 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8209 */
8210 void set_curr_task(int cpu, struct task_struct *p)
8211 {
8212 cpu_curr(cpu) = p;
8213 }
8214
8215 #endif
8216
8217 #ifdef CONFIG_FAIR_GROUP_SCHED
8218 static void free_fair_sched_group(struct task_group *tg)
8219 {
8220 int i;
8221
8222 for_each_possible_cpu(i) {
8223 if (tg->cfs_rq)
8224 kfree(tg->cfs_rq[i]);
8225 if (tg->se)
8226 kfree(tg->se[i]);
8227 }
8228
8229 kfree(tg->cfs_rq);
8230 kfree(tg->se);
8231 }
8232
8233 static
8234 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8235 {
8236 struct cfs_rq *cfs_rq;
8237 struct sched_entity *se;
8238 struct rq *rq;
8239 int i;
8240
8241 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8242 if (!tg->cfs_rq)
8243 goto err;
8244 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8245 if (!tg->se)
8246 goto err;
8247
8248 tg->shares = NICE_0_LOAD;
8249
8250 for_each_possible_cpu(i) {
8251 rq = cpu_rq(i);
8252
8253 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8254 GFP_KERNEL, cpu_to_node(i));
8255 if (!cfs_rq)
8256 goto err;
8257
8258 se = kzalloc_node(sizeof(struct sched_entity),
8259 GFP_KERNEL, cpu_to_node(i));
8260 if (!se)
8261 goto err_free_rq;
8262
8263 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8264 }
8265
8266 return 1;
8267
8268 err_free_rq:
8269 kfree(cfs_rq);
8270 err:
8271 return 0;
8272 }
8273
8274 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8275 {
8276 struct rq *rq = cpu_rq(cpu);
8277 unsigned long flags;
8278
8279 /*
8280 * Only empty task groups can be destroyed; so we can speculatively
8281 * check on_list without danger of it being re-added.
8282 */
8283 if (!tg->cfs_rq[cpu]->on_list)
8284 return;
8285
8286 raw_spin_lock_irqsave(&rq->lock, flags);
8287 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8288 raw_spin_unlock_irqrestore(&rq->lock, flags);
8289 }
8290 #else /* !CONFG_FAIR_GROUP_SCHED */
8291 static inline void free_fair_sched_group(struct task_group *tg)
8292 {
8293 }
8294
8295 static inline
8296 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8297 {
8298 return 1;
8299 }
8300
8301 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8302 {
8303 }
8304 #endif /* CONFIG_FAIR_GROUP_SCHED */
8305
8306 #ifdef CONFIG_RT_GROUP_SCHED
8307 static void free_rt_sched_group(struct task_group *tg)
8308 {
8309 int i;
8310
8311 destroy_rt_bandwidth(&tg->rt_bandwidth);
8312
8313 for_each_possible_cpu(i) {
8314 if (tg->rt_rq)
8315 kfree(tg->rt_rq[i]);
8316 if (tg->rt_se)
8317 kfree(tg->rt_se[i]);
8318 }
8319
8320 kfree(tg->rt_rq);
8321 kfree(tg->rt_se);
8322 }
8323
8324 static
8325 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8326 {
8327 struct rt_rq *rt_rq;
8328 struct sched_rt_entity *rt_se;
8329 struct rq *rq;
8330 int i;
8331
8332 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8333 if (!tg->rt_rq)
8334 goto err;
8335 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8336 if (!tg->rt_se)
8337 goto err;
8338
8339 init_rt_bandwidth(&tg->rt_bandwidth,
8340 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8341
8342 for_each_possible_cpu(i) {
8343 rq = cpu_rq(i);
8344
8345 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8346 GFP_KERNEL, cpu_to_node(i));
8347 if (!rt_rq)
8348 goto err;
8349
8350 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8351 GFP_KERNEL, cpu_to_node(i));
8352 if (!rt_se)
8353 goto err_free_rq;
8354
8355 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8356 }
8357
8358 return 1;
8359
8360 err_free_rq:
8361 kfree(rt_rq);
8362 err:
8363 return 0;
8364 }
8365 #else /* !CONFIG_RT_GROUP_SCHED */
8366 static inline void free_rt_sched_group(struct task_group *tg)
8367 {
8368 }
8369
8370 static inline
8371 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8372 {
8373 return 1;
8374 }
8375 #endif /* CONFIG_RT_GROUP_SCHED */
8376
8377 #ifdef CONFIG_CGROUP_SCHED
8378 static void free_sched_group(struct task_group *tg)
8379 {
8380 free_fair_sched_group(tg);
8381 free_rt_sched_group(tg);
8382 autogroup_free(tg);
8383 kfree(tg);
8384 }
8385
8386 /* allocate runqueue etc for a new task group */
8387 struct task_group *sched_create_group(struct task_group *parent)
8388 {
8389 struct task_group *tg;
8390 unsigned long flags;
8391
8392 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8393 if (!tg)
8394 return ERR_PTR(-ENOMEM);
8395
8396 if (!alloc_fair_sched_group(tg, parent))
8397 goto err;
8398
8399 if (!alloc_rt_sched_group(tg, parent))
8400 goto err;
8401
8402 spin_lock_irqsave(&task_group_lock, flags);
8403 list_add_rcu(&tg->list, &task_groups);
8404
8405 WARN_ON(!parent); /* root should already exist */
8406
8407 tg->parent = parent;
8408 INIT_LIST_HEAD(&tg->children);
8409 list_add_rcu(&tg->siblings, &parent->children);
8410 spin_unlock_irqrestore(&task_group_lock, flags);
8411
8412 return tg;
8413
8414 err:
8415 free_sched_group(tg);
8416 return ERR_PTR(-ENOMEM);
8417 }
8418
8419 /* rcu callback to free various structures associated with a task group */
8420 static void free_sched_group_rcu(struct rcu_head *rhp)
8421 {
8422 /* now it should be safe to free those cfs_rqs */
8423 free_sched_group(container_of(rhp, struct task_group, rcu));
8424 }
8425
8426 /* Destroy runqueue etc associated with a task group */
8427 void sched_destroy_group(struct task_group *tg)
8428 {
8429 unsigned long flags;
8430 int i;
8431
8432 /* end participation in shares distribution */
8433 for_each_possible_cpu(i)
8434 unregister_fair_sched_group(tg, i);
8435
8436 spin_lock_irqsave(&task_group_lock, flags);
8437 list_del_rcu(&tg->list);
8438 list_del_rcu(&tg->siblings);
8439 spin_unlock_irqrestore(&task_group_lock, flags);
8440
8441 /* wait for possible concurrent references to cfs_rqs complete */
8442 call_rcu(&tg->rcu, free_sched_group_rcu);
8443 }
8444
8445 /* change task's runqueue when it moves between groups.
8446 * The caller of this function should have put the task in its new group
8447 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8448 * reflect its new group.
8449 */
8450 void sched_move_task(struct task_struct *tsk)
8451 {
8452 int on_rq, running;
8453 unsigned long flags;
8454 struct rq *rq;
8455
8456 rq = task_rq_lock(tsk, &flags);
8457
8458 running = task_current(rq, tsk);
8459 on_rq = tsk->se.on_rq;
8460
8461 if (on_rq)
8462 dequeue_task(rq, tsk, 0);
8463 if (unlikely(running))
8464 tsk->sched_class->put_prev_task(rq, tsk);
8465
8466 #ifdef CONFIG_FAIR_GROUP_SCHED
8467 if (tsk->sched_class->task_move_group)
8468 tsk->sched_class->task_move_group(tsk, on_rq);
8469 else
8470 #endif
8471 set_task_rq(tsk, task_cpu(tsk));
8472
8473 if (unlikely(running))
8474 tsk->sched_class->set_curr_task(rq);
8475 if (on_rq)
8476 enqueue_task(rq, tsk, 0);
8477
8478 task_rq_unlock(rq, &flags);
8479 }
8480 #endif /* CONFIG_CGROUP_SCHED */
8481
8482 #ifdef CONFIG_FAIR_GROUP_SCHED
8483 static DEFINE_MUTEX(shares_mutex);
8484
8485 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8486 {
8487 int i;
8488 unsigned long flags;
8489
8490 /*
8491 * We can't change the weight of the root cgroup.
8492 */
8493 if (!tg->se[0])
8494 return -EINVAL;
8495
8496 if (shares < MIN_SHARES)
8497 shares = MIN_SHARES;
8498 else if (shares > MAX_SHARES)
8499 shares = MAX_SHARES;
8500
8501 mutex_lock(&shares_mutex);
8502 if (tg->shares == shares)
8503 goto done;
8504
8505 tg->shares = shares;
8506 for_each_possible_cpu(i) {
8507 struct rq *rq = cpu_rq(i);
8508 struct sched_entity *se;
8509
8510 se = tg->se[i];
8511 /* Propagate contribution to hierarchy */
8512 raw_spin_lock_irqsave(&rq->lock, flags);
8513 for_each_sched_entity(se)
8514 update_cfs_shares(group_cfs_rq(se), 0);
8515 raw_spin_unlock_irqrestore(&rq->lock, flags);
8516 }
8517
8518 done:
8519 mutex_unlock(&shares_mutex);
8520 return 0;
8521 }
8522
8523 unsigned long sched_group_shares(struct task_group *tg)
8524 {
8525 return tg->shares;
8526 }
8527 #endif
8528
8529 #ifdef CONFIG_RT_GROUP_SCHED
8530 /*
8531 * Ensure that the real time constraints are schedulable.
8532 */
8533 static DEFINE_MUTEX(rt_constraints_mutex);
8534
8535 static unsigned long to_ratio(u64 period, u64 runtime)
8536 {
8537 if (runtime == RUNTIME_INF)
8538 return 1ULL << 20;
8539
8540 return div64_u64(runtime << 20, period);
8541 }
8542
8543 /* Must be called with tasklist_lock held */
8544 static inline int tg_has_rt_tasks(struct task_group *tg)
8545 {
8546 struct task_struct *g, *p;
8547
8548 do_each_thread(g, p) {
8549 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8550 return 1;
8551 } while_each_thread(g, p);
8552
8553 return 0;
8554 }
8555
8556 struct rt_schedulable_data {
8557 struct task_group *tg;
8558 u64 rt_period;
8559 u64 rt_runtime;
8560 };
8561
8562 static int tg_schedulable(struct task_group *tg, void *data)
8563 {
8564 struct rt_schedulable_data *d = data;
8565 struct task_group *child;
8566 unsigned long total, sum = 0;
8567 u64 period, runtime;
8568
8569 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8570 runtime = tg->rt_bandwidth.rt_runtime;
8571
8572 if (tg == d->tg) {
8573 period = d->rt_period;
8574 runtime = d->rt_runtime;
8575 }
8576
8577 /*
8578 * Cannot have more runtime than the period.
8579 */
8580 if (runtime > period && runtime != RUNTIME_INF)
8581 return -EINVAL;
8582
8583 /*
8584 * Ensure we don't starve existing RT tasks.
8585 */
8586 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8587 return -EBUSY;
8588
8589 total = to_ratio(period, runtime);
8590
8591 /*
8592 * Nobody can have more than the global setting allows.
8593 */
8594 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8595 return -EINVAL;
8596
8597 /*
8598 * The sum of our children's runtime should not exceed our own.
8599 */
8600 list_for_each_entry_rcu(child, &tg->children, siblings) {
8601 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8602 runtime = child->rt_bandwidth.rt_runtime;
8603
8604 if (child == d->tg) {
8605 period = d->rt_period;
8606 runtime = d->rt_runtime;
8607 }
8608
8609 sum += to_ratio(period, runtime);
8610 }
8611
8612 if (sum > total)
8613 return -EINVAL;
8614
8615 return 0;
8616 }
8617
8618 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8619 {
8620 struct rt_schedulable_data data = {
8621 .tg = tg,
8622 .rt_period = period,
8623 .rt_runtime = runtime,
8624 };
8625
8626 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8627 }
8628
8629 static int tg_set_bandwidth(struct task_group *tg,
8630 u64 rt_period, u64 rt_runtime)
8631 {
8632 int i, err = 0;
8633
8634 mutex_lock(&rt_constraints_mutex);
8635 read_lock(&tasklist_lock);
8636 err = __rt_schedulable(tg, rt_period, rt_runtime);
8637 if (err)
8638 goto unlock;
8639
8640 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8641 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8642 tg->rt_bandwidth.rt_runtime = rt_runtime;
8643
8644 for_each_possible_cpu(i) {
8645 struct rt_rq *rt_rq = tg->rt_rq[i];
8646
8647 raw_spin_lock(&rt_rq->rt_runtime_lock);
8648 rt_rq->rt_runtime = rt_runtime;
8649 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8650 }
8651 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8652 unlock:
8653 read_unlock(&tasklist_lock);
8654 mutex_unlock(&rt_constraints_mutex);
8655
8656 return err;
8657 }
8658
8659 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8660 {
8661 u64 rt_runtime, rt_period;
8662
8663 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8664 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8665 if (rt_runtime_us < 0)
8666 rt_runtime = RUNTIME_INF;
8667
8668 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8669 }
8670
8671 long sched_group_rt_runtime(struct task_group *tg)
8672 {
8673 u64 rt_runtime_us;
8674
8675 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8676 return -1;
8677
8678 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8679 do_div(rt_runtime_us, NSEC_PER_USEC);
8680 return rt_runtime_us;
8681 }
8682
8683 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8684 {
8685 u64 rt_runtime, rt_period;
8686
8687 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8688 rt_runtime = tg->rt_bandwidth.rt_runtime;
8689
8690 if (rt_period == 0)
8691 return -EINVAL;
8692
8693 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8694 }
8695
8696 long sched_group_rt_period(struct task_group *tg)
8697 {
8698 u64 rt_period_us;
8699
8700 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8701 do_div(rt_period_us, NSEC_PER_USEC);
8702 return rt_period_us;
8703 }
8704
8705 static int sched_rt_global_constraints(void)
8706 {
8707 u64 runtime, period;
8708 int ret = 0;
8709
8710 if (sysctl_sched_rt_period <= 0)
8711 return -EINVAL;
8712
8713 runtime = global_rt_runtime();
8714 period = global_rt_period();
8715
8716 /*
8717 * Sanity check on the sysctl variables.
8718 */
8719 if (runtime > period && runtime != RUNTIME_INF)
8720 return -EINVAL;
8721
8722 mutex_lock(&rt_constraints_mutex);
8723 read_lock(&tasklist_lock);
8724 ret = __rt_schedulable(NULL, 0, 0);
8725 read_unlock(&tasklist_lock);
8726 mutex_unlock(&rt_constraints_mutex);
8727
8728 return ret;
8729 }
8730
8731 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8732 {
8733 /* Don't accept realtime tasks when there is no way for them to run */
8734 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8735 return 0;
8736
8737 return 1;
8738 }
8739
8740 #else /* !CONFIG_RT_GROUP_SCHED */
8741 static int sched_rt_global_constraints(void)
8742 {
8743 unsigned long flags;
8744 int i;
8745
8746 if (sysctl_sched_rt_period <= 0)
8747 return -EINVAL;
8748
8749 /*
8750 * There's always some RT tasks in the root group
8751 * -- migration, kstopmachine etc..
8752 */
8753 if (sysctl_sched_rt_runtime == 0)
8754 return -EBUSY;
8755
8756 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8757 for_each_possible_cpu(i) {
8758 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8759
8760 raw_spin_lock(&rt_rq->rt_runtime_lock);
8761 rt_rq->rt_runtime = global_rt_runtime();
8762 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8763 }
8764 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8765
8766 return 0;
8767 }
8768 #endif /* CONFIG_RT_GROUP_SCHED */
8769
8770 int sched_rt_handler(struct ctl_table *table, int write,
8771 void __user *buffer, size_t *lenp,
8772 loff_t *ppos)
8773 {
8774 int ret;
8775 int old_period, old_runtime;
8776 static DEFINE_MUTEX(mutex);
8777
8778 mutex_lock(&mutex);
8779 old_period = sysctl_sched_rt_period;
8780 old_runtime = sysctl_sched_rt_runtime;
8781
8782 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8783
8784 if (!ret && write) {
8785 ret = sched_rt_global_constraints();
8786 if (ret) {
8787 sysctl_sched_rt_period = old_period;
8788 sysctl_sched_rt_runtime = old_runtime;
8789 } else {
8790 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8791 def_rt_bandwidth.rt_period =
8792 ns_to_ktime(global_rt_period());
8793 }
8794 }
8795 mutex_unlock(&mutex);
8796
8797 return ret;
8798 }
8799
8800 #ifdef CONFIG_CGROUP_SCHED
8801
8802 /* return corresponding task_group object of a cgroup */
8803 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8804 {
8805 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8806 struct task_group, css);
8807 }
8808
8809 static struct cgroup_subsys_state *
8810 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8811 {
8812 struct task_group *tg, *parent;
8813
8814 if (!cgrp->parent) {
8815 /* This is early initialization for the top cgroup */
8816 return &root_task_group.css;
8817 }
8818
8819 parent = cgroup_tg(cgrp->parent);
8820 tg = sched_create_group(parent);
8821 if (IS_ERR(tg))
8822 return ERR_PTR(-ENOMEM);
8823
8824 return &tg->css;
8825 }
8826
8827 static void
8828 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8829 {
8830 struct task_group *tg = cgroup_tg(cgrp);
8831
8832 sched_destroy_group(tg);
8833 }
8834
8835 static int
8836 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8837 {
8838 #ifdef CONFIG_RT_GROUP_SCHED
8839 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8840 return -EINVAL;
8841 #else
8842 /* We don't support RT-tasks being in separate groups */
8843 if (tsk->sched_class != &fair_sched_class)
8844 return -EINVAL;
8845 #endif
8846 return 0;
8847 }
8848
8849 static int
8850 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8851 struct task_struct *tsk, bool threadgroup)
8852 {
8853 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8854 if (retval)
8855 return retval;
8856 if (threadgroup) {
8857 struct task_struct *c;
8858 rcu_read_lock();
8859 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8860 retval = cpu_cgroup_can_attach_task(cgrp, c);
8861 if (retval) {
8862 rcu_read_unlock();
8863 return retval;
8864 }
8865 }
8866 rcu_read_unlock();
8867 }
8868 return 0;
8869 }
8870
8871 static void
8872 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8873 struct cgroup *old_cont, struct task_struct *tsk,
8874 bool threadgroup)
8875 {
8876 sched_move_task(tsk);
8877 if (threadgroup) {
8878 struct task_struct *c;
8879 rcu_read_lock();
8880 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8881 sched_move_task(c);
8882 }
8883 rcu_read_unlock();
8884 }
8885 }
8886
8887 static void
8888 cpu_cgroup_exit(struct cgroup_subsys *ss, struct task_struct *task)
8889 {
8890 /*
8891 * cgroup_exit() is called in the copy_process() failure path.
8892 * Ignore this case since the task hasn't ran yet, this avoids
8893 * trying to poke a half freed task state from generic code.
8894 */
8895 if (!(task->flags & PF_EXITING))
8896 return;
8897
8898 sched_move_task(task);
8899 }
8900
8901 #ifdef CONFIG_FAIR_GROUP_SCHED
8902 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8903 u64 shareval)
8904 {
8905 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8906 }
8907
8908 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8909 {
8910 struct task_group *tg = cgroup_tg(cgrp);
8911
8912 return (u64) tg->shares;
8913 }
8914 #endif /* CONFIG_FAIR_GROUP_SCHED */
8915
8916 #ifdef CONFIG_RT_GROUP_SCHED
8917 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8918 s64 val)
8919 {
8920 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8921 }
8922
8923 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8924 {
8925 return sched_group_rt_runtime(cgroup_tg(cgrp));
8926 }
8927
8928 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8929 u64 rt_period_us)
8930 {
8931 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8932 }
8933
8934 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8935 {
8936 return sched_group_rt_period(cgroup_tg(cgrp));
8937 }
8938 #endif /* CONFIG_RT_GROUP_SCHED */
8939
8940 static struct cftype cpu_files[] = {
8941 #ifdef CONFIG_FAIR_GROUP_SCHED
8942 {
8943 .name = "shares",
8944 .read_u64 = cpu_shares_read_u64,
8945 .write_u64 = cpu_shares_write_u64,
8946 },
8947 #endif
8948 #ifdef CONFIG_RT_GROUP_SCHED
8949 {
8950 .name = "rt_runtime_us",
8951 .read_s64 = cpu_rt_runtime_read,
8952 .write_s64 = cpu_rt_runtime_write,
8953 },
8954 {
8955 .name = "rt_period_us",
8956 .read_u64 = cpu_rt_period_read_uint,
8957 .write_u64 = cpu_rt_period_write_uint,
8958 },
8959 #endif
8960 };
8961
8962 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8963 {
8964 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8965 }
8966
8967 struct cgroup_subsys cpu_cgroup_subsys = {
8968 .name = "cpu",
8969 .create = cpu_cgroup_create,
8970 .destroy = cpu_cgroup_destroy,
8971 .can_attach = cpu_cgroup_can_attach,
8972 .attach = cpu_cgroup_attach,
8973 .exit = cpu_cgroup_exit,
8974 .populate = cpu_cgroup_populate,
8975 .subsys_id = cpu_cgroup_subsys_id,
8976 .early_init = 1,
8977 };
8978
8979 #endif /* CONFIG_CGROUP_SCHED */
8980
8981 #ifdef CONFIG_CGROUP_CPUACCT
8982
8983 /*
8984 * CPU accounting code for task groups.
8985 *
8986 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8987 * (balbir@in.ibm.com).
8988 */
8989
8990 /* track cpu usage of a group of tasks and its child groups */
8991 struct cpuacct {
8992 struct cgroup_subsys_state css;
8993 /* cpuusage holds pointer to a u64-type object on every cpu */
8994 u64 __percpu *cpuusage;
8995 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8996 struct cpuacct *parent;
8997 };
8998
8999 struct cgroup_subsys cpuacct_subsys;
9000
9001 /* return cpu accounting group corresponding to this container */
9002 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9003 {
9004 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9005 struct cpuacct, css);
9006 }
9007
9008 /* return cpu accounting group to which this task belongs */
9009 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9010 {
9011 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9012 struct cpuacct, css);
9013 }
9014
9015 /* create a new cpu accounting group */
9016 static struct cgroup_subsys_state *cpuacct_create(
9017 struct cgroup_subsys *ss, struct cgroup *cgrp)
9018 {
9019 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9020 int i;
9021
9022 if (!ca)
9023 goto out;
9024
9025 ca->cpuusage = alloc_percpu(u64);
9026 if (!ca->cpuusage)
9027 goto out_free_ca;
9028
9029 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9030 if (percpu_counter_init(&ca->cpustat[i], 0))
9031 goto out_free_counters;
9032
9033 if (cgrp->parent)
9034 ca->parent = cgroup_ca(cgrp->parent);
9035
9036 return &ca->css;
9037
9038 out_free_counters:
9039 while (--i >= 0)
9040 percpu_counter_destroy(&ca->cpustat[i]);
9041 free_percpu(ca->cpuusage);
9042 out_free_ca:
9043 kfree(ca);
9044 out:
9045 return ERR_PTR(-ENOMEM);
9046 }
9047
9048 /* destroy an existing cpu accounting group */
9049 static void
9050 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9051 {
9052 struct cpuacct *ca = cgroup_ca(cgrp);
9053 int i;
9054
9055 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9056 percpu_counter_destroy(&ca->cpustat[i]);
9057 free_percpu(ca->cpuusage);
9058 kfree(ca);
9059 }
9060
9061 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9062 {
9063 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9064 u64 data;
9065
9066 #ifndef CONFIG_64BIT
9067 /*
9068 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9069 */
9070 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9071 data = *cpuusage;
9072 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9073 #else
9074 data = *cpuusage;
9075 #endif
9076
9077 return data;
9078 }
9079
9080 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9081 {
9082 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9083
9084 #ifndef CONFIG_64BIT
9085 /*
9086 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9087 */
9088 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9089 *cpuusage = val;
9090 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9091 #else
9092 *cpuusage = val;
9093 #endif
9094 }
9095
9096 /* return total cpu usage (in nanoseconds) of a group */
9097 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9098 {
9099 struct cpuacct *ca = cgroup_ca(cgrp);
9100 u64 totalcpuusage = 0;
9101 int i;
9102
9103 for_each_present_cpu(i)
9104 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9105
9106 return totalcpuusage;
9107 }
9108
9109 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9110 u64 reset)
9111 {
9112 struct cpuacct *ca = cgroup_ca(cgrp);
9113 int err = 0;
9114 int i;
9115
9116 if (reset) {
9117 err = -EINVAL;
9118 goto out;
9119 }
9120
9121 for_each_present_cpu(i)
9122 cpuacct_cpuusage_write(ca, i, 0);
9123
9124 out:
9125 return err;
9126 }
9127
9128 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9129 struct seq_file *m)
9130 {
9131 struct cpuacct *ca = cgroup_ca(cgroup);
9132 u64 percpu;
9133 int i;
9134
9135 for_each_present_cpu(i) {
9136 percpu = cpuacct_cpuusage_read(ca, i);
9137 seq_printf(m, "%llu ", (unsigned long long) percpu);
9138 }
9139 seq_printf(m, "\n");
9140 return 0;
9141 }
9142
9143 static const char *cpuacct_stat_desc[] = {
9144 [CPUACCT_STAT_USER] = "user",
9145 [CPUACCT_STAT_SYSTEM] = "system",
9146 };
9147
9148 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9149 struct cgroup_map_cb *cb)
9150 {
9151 struct cpuacct *ca = cgroup_ca(cgrp);
9152 int i;
9153
9154 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9155 s64 val = percpu_counter_read(&ca->cpustat[i]);
9156 val = cputime64_to_clock_t(val);
9157 cb->fill(cb, cpuacct_stat_desc[i], val);
9158 }
9159 return 0;
9160 }
9161
9162 static struct cftype files[] = {
9163 {
9164 .name = "usage",
9165 .read_u64 = cpuusage_read,
9166 .write_u64 = cpuusage_write,
9167 },
9168 {
9169 .name = "usage_percpu",
9170 .read_seq_string = cpuacct_percpu_seq_read,
9171 },
9172 {
9173 .name = "stat",
9174 .read_map = cpuacct_stats_show,
9175 },
9176 };
9177
9178 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9179 {
9180 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9181 }
9182
9183 /*
9184 * charge this task's execution time to its accounting group.
9185 *
9186 * called with rq->lock held.
9187 */
9188 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9189 {
9190 struct cpuacct *ca;
9191 int cpu;
9192
9193 if (unlikely(!cpuacct_subsys.active))
9194 return;
9195
9196 cpu = task_cpu(tsk);
9197
9198 rcu_read_lock();
9199
9200 ca = task_ca(tsk);
9201
9202 for (; ca; ca = ca->parent) {
9203 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9204 *cpuusage += cputime;
9205 }
9206
9207 rcu_read_unlock();
9208 }
9209
9210 /*
9211 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9212 * in cputime_t units. As a result, cpuacct_update_stats calls
9213 * percpu_counter_add with values large enough to always overflow the
9214 * per cpu batch limit causing bad SMP scalability.
9215 *
9216 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9217 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9218 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9219 */
9220 #ifdef CONFIG_SMP
9221 #define CPUACCT_BATCH \
9222 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9223 #else
9224 #define CPUACCT_BATCH 0
9225 #endif
9226
9227 /*
9228 * Charge the system/user time to the task's accounting group.
9229 */
9230 static void cpuacct_update_stats(struct task_struct *tsk,
9231 enum cpuacct_stat_index idx, cputime_t val)
9232 {
9233 struct cpuacct *ca;
9234 int batch = CPUACCT_BATCH;
9235
9236 if (unlikely(!cpuacct_subsys.active))
9237 return;
9238
9239 rcu_read_lock();
9240 ca = task_ca(tsk);
9241
9242 do {
9243 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9244 ca = ca->parent;
9245 } while (ca);
9246 rcu_read_unlock();
9247 }
9248
9249 struct cgroup_subsys cpuacct_subsys = {
9250 .name = "cpuacct",
9251 .create = cpuacct_create,
9252 .destroy = cpuacct_destroy,
9253 .populate = cpuacct_populate,
9254 .subsys_id = cpuacct_subsys_id,
9255 };
9256 #endif /* CONFIG_CGROUP_CPUACCT */
9257
This page took 0.221512 seconds and 5 git commands to generate.