projects/linux·blob v3.2 · refs

9785 lines · 233.3 KiB

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 <asm/mmu_context.h>
36#include <linux/interrupt.h>
37#include <linux/capability.h>
38#include <linux/completion.h>
39#include <linux/kernel_stat.h>
40#include <linux/debug_locks.h>
41#include <linux/perf_event.h>
42#include <linux/security.h>
43#include <linux/notifier.h>
44#include <linux/profile.h>
45#include <linux/freezer.h>
46#include <linux/vmalloc.h>
47#include <linux/blkdev.h>
48#include <linux/delay.h>
49#include <linux/pid_namespace.h>
50#include <linux/smp.h>
51#include <linux/threads.h>
52#include <linux/timer.h>
53#include <linux/rcupdate.h>
54#include <linux/cpu.h>
55#include <linux/cpuset.h>
56#include <linux/percpu.h>
57#include <linux/proc_fs.h>
58#include <linux/seq_file.h>
59#include <linux/stop_machine.h>
60#include <linux/sysctl.h>
61#include <linux/syscalls.h>
62#include <linux/times.h>
63#include <linux/tsacct_kern.h>
64#include <linux/kprobes.h>
65#include <linux/delayacct.h>
66#include <linux/unistd.h>
67#include <linux/pagemap.h>
68#include <linux/hrtimer.h>
69#include <linux/tick.h>
70#include <linux/debugfs.h>
71#include <linux/ctype.h>
72#include <linux/ftrace.h>
73#include <linux/slab.h>
74#include <linux/init_task.h>
75
76#include <asm/tlb.h>
77#include <asm/irq_regs.h>
78#include <asm/mutex.h>
79#ifdef CONFIG_PARAVIRT
80#include <asm/paravirt.h>
81#endif
82
83#include "sched_cpupri.h"
84#include "workqueue_sched.h"
85#include "sched_autogroup.h"
86
87#define CREATE_TRACE_POINTS
88#include <trace/events/sched.h>
89
90/*
91 * Convert user-nice values [ -20 ... 0 ... 19 ]
92 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
93 * and back.
94 */
95#define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
96#define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
97#define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
98
99/*
100 * 'User priority' is the nice value converted to something we
101 * can work with better when scaling various scheduler parameters,
102 * it's a [ 0 ... 39 ] range.
103 */
104#define USER_PRIO(p) ((p)-MAX_RT_PRIO)
105#define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
106#define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
107
108/*
109 * Helpers for converting nanosecond timing to jiffy resolution
110 */
111#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
112
113#define NICE_0_LOAD SCHED_LOAD_SCALE
114#define NICE_0_SHIFT SCHED_LOAD_SHIFT
115
116/*
117 * These are the 'tuning knobs' of the scheduler:
118 *
119 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
120 * Timeslices get refilled after they expire.
121 */
122#define DEF_TIMESLICE (100 * HZ / 1000)
123
124/*
125 * single value that denotes runtime == period, ie unlimited time.
126 */
127#define RUNTIME_INF ((u64)~0ULL)
128
129static inline int rt_policy(int policy)
130{
131 if (policy == SCHED_FIFO || policy == SCHED_RR)
132 return 1;
133 return 0;
134}
135
136static inline int task_has_rt_policy(struct task_struct *p)
137{
138 return rt_policy(p->policy);
139}
140
141/*
142 * This is the priority-queue data structure of the RT scheduling class:
143 */
144struct rt_prio_array {
145 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
146 struct list_head queue[MAX_RT_PRIO];
147};
148
149struct rt_bandwidth {
150 /* nests inside the rq lock: */
151 raw_spinlock_t rt_runtime_lock;
152 ktime_t rt_period;
153 u64 rt_runtime;
154 struct hrtimer rt_period_timer;
155};
156
157static struct rt_bandwidth def_rt_bandwidth;
158
159static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
160
161static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
162{
163 struct rt_bandwidth *rt_b =
164 container_of(timer, struct rt_bandwidth, rt_period_timer);
165 ktime_t now;
166 int overrun;
167 int idle = 0;
168
169 for (;;) {
170 now = hrtimer_cb_get_time(timer);
171 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
172
173 if (!overrun)
174 break;
175
176 idle = do_sched_rt_period_timer(rt_b, overrun);
177 }
178
179 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
180}
181
182static
183void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
184{
185 rt_b->rt_period = ns_to_ktime(period);
186 rt_b->rt_runtime = runtime;
187
188 raw_spin_lock_init(&rt_b->rt_runtime_lock);
189
190 hrtimer_init(&rt_b->rt_period_timer,
191 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
192 rt_b->rt_period_timer.function = sched_rt_period_timer;
193}
194
195static inline int rt_bandwidth_enabled(void)
196{
197 return sysctl_sched_rt_runtime >= 0;
198}
199
200static void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
201{
202 unsigned long delta;
203 ktime_t soft, hard, now;
204
205 for (;;) {
206 if (hrtimer_active(period_timer))
207 break;
208
209 now = hrtimer_cb_get_time(period_timer);
210 hrtimer_forward(period_timer, now, period);
211
212 soft = hrtimer_get_softexpires(period_timer);
213 hard = hrtimer_get_expires(period_timer);
214 delta = ktime_to_ns(ktime_sub(hard, soft));
215 __hrtimer_start_range_ns(period_timer, soft, delta,
216 HRTIMER_MODE_ABS_PINNED, 0);
217 }
218}
219
220static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
221{
222 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
223 return;
224
225 if (hrtimer_active(&rt_b->rt_period_timer))
226 return;
227
228 raw_spin_lock(&rt_b->rt_runtime_lock);
229 start_bandwidth_timer(&rt_b->rt_period_timer, rt_b->rt_period);
230 raw_spin_unlock(&rt_b->rt_runtime_lock);
231}
232
233#ifdef CONFIG_RT_GROUP_SCHED
234static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
235{
236 hrtimer_cancel(&rt_b->rt_period_timer);
237}
238#endif
239
240/*
241 * sched_domains_mutex serializes calls to init_sched_domains,
242 * detach_destroy_domains and partition_sched_domains.
243 */
244static DEFINE_MUTEX(sched_domains_mutex);
245
246#ifdef CONFIG_CGROUP_SCHED
247
248#include <linux/cgroup.h>
249
250struct cfs_rq;
251
252static LIST_HEAD(task_groups);
253
254struct cfs_bandwidth {
255#ifdef CONFIG_CFS_BANDWIDTH
256 raw_spinlock_t lock;
257 ktime_t period;
258 u64 quota, runtime;
259 s64 hierarchal_quota;
260 u64 runtime_expires;
261
262 int idle, timer_active;
263 struct hrtimer period_timer, slack_timer;
264 struct list_head throttled_cfs_rq;
265
266 /* statistics */
267 int nr_periods, nr_throttled;
268 u64 throttled_time;
269#endif
270};
271
272/* task group related information */
273struct task_group {
274 struct cgroup_subsys_state css;
275
276#ifdef CONFIG_FAIR_GROUP_SCHED
277 /* schedulable entities of this group on each cpu */
278 struct sched_entity **se;
279 /* runqueue "owned" by this group on each cpu */
280 struct cfs_rq **cfs_rq;
281 unsigned long shares;
282
283 atomic_t load_weight;
284#endif
285
286#ifdef CONFIG_RT_GROUP_SCHED
287 struct sched_rt_entity **rt_se;
288 struct rt_rq **rt_rq;
289
290 struct rt_bandwidth rt_bandwidth;
291#endif
292
293 struct rcu_head rcu;
294 struct list_head list;
295
296 struct task_group *parent;
297 struct list_head siblings;
298 struct list_head children;
299
300#ifdef CONFIG_SCHED_AUTOGROUP
301 struct autogroup *autogroup;
302#endif
303
304 struct cfs_bandwidth cfs_bandwidth;
305};
306
307/* task_group_lock serializes the addition/removal of task groups */
308static DEFINE_SPINLOCK(task_group_lock);
309
310#ifdef CONFIG_FAIR_GROUP_SCHED
311
312# define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
313
314/*
315 * A weight of 0 or 1 can cause arithmetics problems.
316 * A weight of a cfs_rq is the sum of weights of which entities
317 * are queued on this cfs_rq, so a weight of a entity should not be
318 * too large, so as the shares value of a task group.
319 * (The default weight is 1024 - so there's no practical
320 * limitation from this.)
321 */
322#define MIN_SHARES (1UL << 1)
323#define MAX_SHARES (1UL << 18)
324
325static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
326#endif
327
328/* Default task group.
329 * Every task in system belong to this group at bootup.
330 */
331struct task_group root_task_group;
332
333#endif /* CONFIG_CGROUP_SCHED */
334
335/* CFS-related fields in a runqueue */
336struct cfs_rq {
337 struct load_weight load;
338 unsigned long nr_running, h_nr_running;
339
340 u64 exec_clock;
341 u64 min_vruntime;
342#ifndef CONFIG_64BIT
343 u64 min_vruntime_copy;
344#endif
345
346 struct rb_root tasks_timeline;
347 struct rb_node *rb_leftmost;
348
349 struct list_head tasks;
350 struct list_head *balance_iterator;
351
352 /*
353 * 'curr' points to currently running entity on this cfs_rq.
354 * It is set to NULL otherwise (i.e when none are currently running).
355 */
356 struct sched_entity *curr, *next, *last, *skip;
357
358#ifdef CONFIG_SCHED_DEBUG
359 unsigned int nr_spread_over;
360#endif
361
362#ifdef CONFIG_FAIR_GROUP_SCHED
363 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
364
365 /*
366 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
367 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
368 * (like users, containers etc.)
369 *
370 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
371 * list is used during load balance.
372 */
373 int on_list;
374 struct list_head leaf_cfs_rq_list;
375 struct task_group *tg; /* group that "owns" this runqueue */
376
377#ifdef CONFIG_SMP
378 /*
379 * the part of load.weight contributed by tasks
380 */
381 unsigned long task_weight;
382
383 /*
384 * h_load = weight * f(tg)
385 *
386 * Where f(tg) is the recursive weight fraction assigned to
387 * this group.
388 */
389 unsigned long h_load;
390
391 /*
392 * Maintaining per-cpu shares distribution for group scheduling
393 *
394 * load_stamp is the last time we updated the load average
395 * load_last is the last time we updated the load average and saw load
396 * load_unacc_exec_time is currently unaccounted execution time
397 */
398 u64 load_avg;
399 u64 load_period;
400 u64 load_stamp, load_last, load_unacc_exec_time;
401
402 unsigned long load_contribution;
403#endif
404#ifdef CONFIG_CFS_BANDWIDTH
405 int runtime_enabled;
406 u64 runtime_expires;
407 s64 runtime_remaining;
408
409 u64 throttled_timestamp;
410 int throttled, throttle_count;
411 struct list_head throttled_list;
412#endif
413#endif
414};
415
416#ifdef CONFIG_FAIR_GROUP_SCHED
417#ifdef CONFIG_CFS_BANDWIDTH
418static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
419{
420 return &tg->cfs_bandwidth;
421}
422
423static inline u64 default_cfs_period(void);
424static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
425static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
426
427static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
428{
429 struct cfs_bandwidth *cfs_b =
430 container_of(timer, struct cfs_bandwidth, slack_timer);
431 do_sched_cfs_slack_timer(cfs_b);
432
433 return HRTIMER_NORESTART;
434}
435
436static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
437{
438 struct cfs_bandwidth *cfs_b =
439 container_of(timer, struct cfs_bandwidth, period_timer);
440 ktime_t now;
441 int overrun;
442 int idle = 0;
443
444 for (;;) {
445 now = hrtimer_cb_get_time(timer);
446 overrun = hrtimer_forward(timer, now, cfs_b->period);
447
448 if (!overrun)
449 break;
450
451 idle = do_sched_cfs_period_timer(cfs_b, overrun);
452 }
453
454 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
455}
456
457static void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
458{
459 raw_spin_lock_init(&cfs_b->lock);
460 cfs_b->runtime = 0;
461 cfs_b->quota = RUNTIME_INF;
462 cfs_b->period = ns_to_ktime(default_cfs_period());
463
464 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
465 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
466 cfs_b->period_timer.function = sched_cfs_period_timer;
467 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
468 cfs_b->slack_timer.function = sched_cfs_slack_timer;
469}
470
471static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
472{
473 cfs_rq->runtime_enabled = 0;
474 INIT_LIST_HEAD(&cfs_rq->throttled_list);
475}
476
477/* requires cfs_b->lock, may release to reprogram timer */
478static void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
479{
480 /*
481 * The timer may be active because we're trying to set a new bandwidth
482 * period or because we're racing with the tear-down path
483 * (timer_active==0 becomes visible before the hrtimer call-back
484 * terminates). In either case we ensure that it's re-programmed
485 */
486 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
487 raw_spin_unlock(&cfs_b->lock);
488 /* ensure cfs_b->lock is available while we wait */
489 hrtimer_cancel(&cfs_b->period_timer);
490
491 raw_spin_lock(&cfs_b->lock);
492 /* if someone else restarted the timer then we're done */
493 if (cfs_b->timer_active)
494 return;
495 }
496
497 cfs_b->timer_active = 1;
498 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
499}
500
501static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
502{
503 hrtimer_cancel(&cfs_b->period_timer);
504 hrtimer_cancel(&cfs_b->slack_timer);
505}
506#else
507static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
508static void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
509static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
510
511static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
512{
513 return NULL;
514}
515#endif /* CONFIG_CFS_BANDWIDTH */
516#endif /* CONFIG_FAIR_GROUP_SCHED */
517
518/* Real-Time classes' related field in a runqueue: */
519struct rt_rq {
520 struct rt_prio_array active;
521 unsigned long rt_nr_running;
522#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
523 struct {
524 int curr; /* highest queued rt task prio */
525#ifdef CONFIG_SMP
526 int next; /* next highest */
527#endif
528 } highest_prio;
529#endif
530#ifdef CONFIG_SMP
531 unsigned long rt_nr_migratory;
532 unsigned long rt_nr_total;
533 int overloaded;
534 struct plist_head pushable_tasks;
535#endif
536 int rt_throttled;
537 u64 rt_time;
538 u64 rt_runtime;
539 /* Nests inside the rq lock: */
540 raw_spinlock_t rt_runtime_lock;
541
542#ifdef CONFIG_RT_GROUP_SCHED
543 unsigned long rt_nr_boosted;
544
545 struct rq *rq;
546 struct list_head leaf_rt_rq_list;
547 struct task_group *tg;
548#endif
549};
550
551#ifdef CONFIG_SMP
552
553/*
554 * We add the notion of a root-domain which will be used to define per-domain
555 * variables. Each exclusive cpuset essentially defines an island domain by
556 * fully partitioning the member cpus from any other cpuset. Whenever a new
557 * exclusive cpuset is created, we also create and attach a new root-domain
558 * object.
559 *
560 */
561struct root_domain {
562 atomic_t refcount;
563 atomic_t rto_count;
564 struct rcu_head rcu;
565 cpumask_var_t span;
566 cpumask_var_t online;
567
568 /*
569 * The "RT overload" flag: it gets set if a CPU has more than
570 * one runnable RT task.
571 */
572 cpumask_var_t rto_mask;
573 struct cpupri cpupri;
574};
575
576/*
577 * By default the system creates a single root-domain with all cpus as
578 * members (mimicking the global state we have today).
579 */
580static struct root_domain def_root_domain;
581
582#endif /* CONFIG_SMP */
583
584/*
585 * This is the main, per-CPU runqueue data structure.
586 *
587 * Locking rule: those places that want to lock multiple runqueues
588 * (such as the load balancing or the thread migration code), lock
589 * acquire operations must be ordered by ascending &runqueue.
590 */
591struct rq {
592 /* runqueue lock: */
593 raw_spinlock_t lock;
594
595 /*
596 * nr_running and cpu_load should be in the same cacheline because
597 * remote CPUs use both these fields when doing load calculation.
598 */
599 unsigned long nr_running;
600 #define CPU_LOAD_IDX_MAX 5
601 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
602 unsigned long last_load_update_tick;
603#ifdef CONFIG_NO_HZ
604 u64 nohz_stamp;
605 unsigned char nohz_balance_kick;
606#endif
607 int skip_clock_update;
608
609 /* capture load from *all* tasks on this cpu: */
610 struct load_weight load;
611 unsigned long nr_load_updates;
612 u64 nr_switches;
613
614 struct cfs_rq cfs;
615 struct rt_rq rt;
616
617#ifdef CONFIG_FAIR_GROUP_SCHED
618 /* list of leaf cfs_rq on this cpu: */
619 struct list_head leaf_cfs_rq_list;
620#endif
621#ifdef CONFIG_RT_GROUP_SCHED
622 struct list_head leaf_rt_rq_list;
623#endif
624
625 /*
626 * This is part of a global counter where only the total sum
627 * over all CPUs matters. A task can increase this counter on
628 * one CPU and if it got migrated afterwards it may decrease
629 * it on another CPU. Always updated under the runqueue lock:
630 */
631 unsigned long nr_uninterruptible;
632
633 struct task_struct *curr, *idle, *stop;
634 unsigned long next_balance;
635 struct mm_struct *prev_mm;
636
637 u64 clock;
638 u64 clock_task;
639
640 atomic_t nr_iowait;
641
642#ifdef CONFIG_SMP
643 struct root_domain *rd;
644 struct sched_domain *sd;
645
646 unsigned long cpu_power;
647
648 unsigned char idle_balance;
649 /* For active balancing */
650 int post_schedule;
651 int active_balance;
652 int push_cpu;
653 struct cpu_stop_work active_balance_work;
654 /* cpu of this runqueue: */
655 int cpu;
656 int online;
657
658 u64 rt_avg;
659 u64 age_stamp;
660 u64 idle_stamp;
661 u64 avg_idle;
662#endif
663
664#ifdef CONFIG_IRQ_TIME_ACCOUNTING
665 u64 prev_irq_time;
666#endif
667#ifdef CONFIG_PARAVIRT
668 u64 prev_steal_time;
669#endif
670#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
671 u64 prev_steal_time_rq;
672#endif
673
674 /* calc_load related fields */
675 unsigned long calc_load_update;
676 long calc_load_active;
677
678#ifdef CONFIG_SCHED_HRTICK
679#ifdef CONFIG_SMP
680 int hrtick_csd_pending;
681 struct call_single_data hrtick_csd;
682#endif
683 struct hrtimer hrtick_timer;
684#endif
685
686#ifdef CONFIG_SCHEDSTATS
687 /* latency stats */
688 struct sched_info rq_sched_info;
689 unsigned long long rq_cpu_time;
690 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
691
692 /* sys_sched_yield() stats */
693 unsigned int yld_count;
694
695 /* schedule() stats */
696 unsigned int sched_switch;
697 unsigned int sched_count;
698 unsigned int sched_goidle;
699
700 /* try_to_wake_up() stats */
701 unsigned int ttwu_count;
702 unsigned int ttwu_local;
703#endif
704
705#ifdef CONFIG_SMP
706 struct llist_head wake_list;
707#endif
708};
709
710static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
711
712
713static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
714
715static inline int cpu_of(struct rq *rq)
716{
717#ifdef CONFIG_SMP
718 return rq->cpu;
719#else
720 return 0;
721#endif
722}
723
724#define rcu_dereference_check_sched_domain(p) \
725 rcu_dereference_check((p), \
726 lockdep_is_held(&sched_domains_mutex))
727
728/*
729 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
730 * See detach_destroy_domains: synchronize_sched for details.
731 *
732 * The domain tree of any CPU may only be accessed from within
733 * preempt-disabled sections.
734 */
735#define for_each_domain(cpu, __sd) \
736 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
737
738#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
739#define this_rq() (&__get_cpu_var(runqueues))
740#define task_rq(p) cpu_rq(task_cpu(p))
741#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
742#define raw_rq() (&__raw_get_cpu_var(runqueues))
743
744#ifdef CONFIG_CGROUP_SCHED
745
746/*
747 * Return the group to which this tasks belongs.
748 *
749 * We use task_subsys_state_check() and extend the RCU verification with
750 * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each
751 * task it moves into the cgroup. Therefore by holding either of those locks,
752 * we pin the task to the current cgroup.
753 */
754static inline struct task_group *task_group(struct task_struct *p)
755{
756 struct task_group *tg;
757 struct cgroup_subsys_state *css;
758
759 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
760 lockdep_is_held(&p->pi_lock) ||
761 lockdep_is_held(&task_rq(p)->lock));
762 tg = container_of(css, struct task_group, css);
763
764 return autogroup_task_group(p, tg);
765}
766
767/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
768static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
769{
770#ifdef CONFIG_FAIR_GROUP_SCHED
771 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
772 p->se.parent = task_group(p)->se[cpu];
773#endif
774
775#ifdef CONFIG_RT_GROUP_SCHED
776 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
777 p->rt.parent = task_group(p)->rt_se[cpu];
778#endif
779}
780
781#else /* CONFIG_CGROUP_SCHED */
782
783static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
784static inline struct task_group *task_group(struct task_struct *p)
785{
786 return NULL;
787}
788
789#endif /* CONFIG_CGROUP_SCHED */
790
791static void update_rq_clock_task(struct rq *rq, s64 delta);
792
793static void update_rq_clock(struct rq *rq)
794{
795 s64 delta;
796
797 if (rq->skip_clock_update > 0)
798 return;
799
800 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
801 rq->clock += delta;
802 update_rq_clock_task(rq, delta);
803}
804
805/*
806 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
807 */
808#ifdef CONFIG_SCHED_DEBUG
809# define const_debug __read_mostly
810#else
811# define const_debug static const
812#endif
813
814/**
815 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
816 * @cpu: the processor in question.
817 *
818 * This interface allows printk to be called with the runqueue lock
819 * held and know whether or not it is OK to wake up the klogd.
820 */
821int runqueue_is_locked(int cpu)
822{
823 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
824}
825
826/*
827 * Debugging: various feature bits
828 */
829
830#define SCHED_FEAT(name, enabled) \
831 __SCHED_FEAT_##name ,
832
833enum {
834#include "sched_features.h"
835};
836
837#undef SCHED_FEAT
838
839#define SCHED_FEAT(name, enabled) \
840 (1UL << __SCHED_FEAT_##name) * enabled |
841
842const_debug unsigned int sysctl_sched_features =
843#include "sched_features.h"
844 0;
845
846#undef SCHED_FEAT
847
848#ifdef CONFIG_SCHED_DEBUG
849#define SCHED_FEAT(name, enabled) \
850 #name ,
851
852static __read_mostly char *sched_feat_names[] = {
853#include "sched_features.h"
854 NULL
855};
856
857#undef SCHED_FEAT
858
859static int sched_feat_show(struct seq_file *m, void *v)
860{
861 int i;
862
863 for (i = 0; sched_feat_names[i]; i++) {
864 if (!(sysctl_sched_features & (1UL << i)))
865 seq_puts(m, "NO_");
866 seq_printf(m, "%s ", sched_feat_names[i]);
867 }
868 seq_puts(m, "\n");
869
870 return 0;
871}
872
873static ssize_t
874sched_feat_write(struct file *filp, const char __user *ubuf,
875 size_t cnt, loff_t *ppos)
876{
877 char buf[64];
878 char *cmp;
879 int neg = 0;
880 int i;
881
882 if (cnt > 63)
883 cnt = 63;
884
885 if (copy_from_user(&buf, ubuf, cnt))
886 return -EFAULT;
887
888 buf[cnt] = 0;
889 cmp = strstrip(buf);
890
891 if (strncmp(cmp, "NO_", 3) == 0) {
892 neg = 1;
893 cmp += 3;
894 }
895
896 for (i = 0; sched_feat_names[i]; i++) {
897 if (strcmp(cmp, sched_feat_names[i]) == 0) {
898 if (neg)
899 sysctl_sched_features &= ~(1UL << i);
900 else
901 sysctl_sched_features |= (1UL << i);
902 break;
903 }
904 }
905
906 if (!sched_feat_names[i])
907 return -EINVAL;
908
909 *ppos += cnt;
910
911 return cnt;
912}
913
914static int sched_feat_open(struct inode *inode, struct file *filp)
915{
916 return single_open(filp, sched_feat_show, NULL);
917}
918
919static const struct file_operations sched_feat_fops = {
920 .open = sched_feat_open,
921 .write = sched_feat_write,
922 .read = seq_read,
923 .llseek = seq_lseek,
924 .release = single_release,
925};
926
927static __init int sched_init_debug(void)
928{
929 debugfs_create_file("sched_features", 0644, NULL, NULL,
930 &sched_feat_fops);
931
932 return 0;
933}
934late_initcall(sched_init_debug);
935
936#endif
937
938#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
939
940/*
941 * Number of tasks to iterate in a single balance run.
942 * Limited because this is done with IRQs disabled.
943 */
944const_debug unsigned int sysctl_sched_nr_migrate = 32;
945
946/*
947 * period over which we average the RT time consumption, measured
948 * in ms.
949 *
950 * default: 1s
951 */
952const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
953
954/*
955 * period over which we measure -rt task cpu usage in us.
956 * default: 1s
957 */
958unsigned int sysctl_sched_rt_period = 1000000;
959
960static __read_mostly int scheduler_running;
961
962/*
963 * part of the period that we allow rt tasks to run in us.
964 * default: 0.95s
965 */
966int sysctl_sched_rt_runtime = 950000;
967
968static inline u64 global_rt_period(void)
969{
970 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
971}
972
973static inline u64 global_rt_runtime(void)
974{
975 if (sysctl_sched_rt_runtime < 0)
976 return RUNTIME_INF;
977
978 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
979}
980
981#ifndef prepare_arch_switch
982# define prepare_arch_switch(next) do { } while (0)
983#endif
984#ifndef finish_arch_switch
985# define finish_arch_switch(prev) do { } while (0)
986#endif
987
988static inline int task_current(struct rq *rq, struct task_struct *p)
989{
990 return rq->curr == p;
991}
992
993static inline int task_running(struct rq *rq, struct task_struct *p)
994{
995#ifdef CONFIG_SMP
996 return p->on_cpu;
997#else
998 return task_current(rq, p);
999#endif
1000}
1001
1002#ifndef __ARCH_WANT_UNLOCKED_CTXSW
1003static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1004{
1005#ifdef CONFIG_SMP
1006 /*
1007 * We can optimise this out completely for !SMP, because the
1008 * SMP rebalancing from interrupt is the only thing that cares
1009 * here.
1010 */
1011 next->on_cpu = 1;
1012#endif
1013}
1014
1015static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1016{
1017#ifdef CONFIG_SMP
1018 /*
1019 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1020 * We must ensure this doesn't happen until the switch is completely
1021 * finished.
1022 */
1023 smp_wmb();
1024 prev->on_cpu = 0;
1025#endif
1026#ifdef CONFIG_DEBUG_SPINLOCK
1027 /* this is a valid case when another task releases the spinlock */
1028 rq->lock.owner = current;
1029#endif
1030 /*
1031 * If we are tracking spinlock dependencies then we have to
1032 * fix up the runqueue lock - which gets 'carried over' from
1033 * prev into current:
1034 */
1035 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1036
1037 raw_spin_unlock_irq(&rq->lock);
1038}
1039
1040#else /* __ARCH_WANT_UNLOCKED_CTXSW */
1041static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1042{
1043#ifdef CONFIG_SMP
1044 /*
1045 * We can optimise this out completely for !SMP, because the
1046 * SMP rebalancing from interrupt is the only thing that cares
1047 * here.
1048 */
1049 next->on_cpu = 1;
1050#endif
1051#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1052 raw_spin_unlock_irq(&rq->lock);
1053#else
1054 raw_spin_unlock(&rq->lock);
1055#endif
1056}
1057
1058static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1059{
1060#ifdef CONFIG_SMP
1061 /*
1062 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1063 * We must ensure this doesn't happen until the switch is completely
1064 * finished.
1065 */
1066 smp_wmb();
1067 prev->on_cpu = 0;
1068#endif
1069#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1070 local_irq_enable();
1071#endif
1072}
1073#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1074
1075/*
1076 * __task_rq_lock - lock the rq @p resides on.
1077 */
1078static inline struct rq *__task_rq_lock(struct task_struct *p)
1079 __acquires(rq->lock)
1080{
1081 struct rq *rq;
1082
1083 lockdep_assert_held(&p->pi_lock);
1084
1085 for (;;) {
1086 rq = task_rq(p);
1087 raw_spin_lock(&rq->lock);
1088 if (likely(rq == task_rq(p)))
1089 return rq;
1090 raw_spin_unlock(&rq->lock);
1091 }
1092}
1093
1094/*
1095 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
1096 */
1097static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1098 __acquires(p->pi_lock)
1099 __acquires(rq->lock)
1100{
1101 struct rq *rq;
1102
1103 for (;;) {
1104 raw_spin_lock_irqsave(&p->pi_lock, *flags);
1105 rq = task_rq(p);
1106 raw_spin_lock(&rq->lock);
1107 if (likely(rq == task_rq(p)))
1108 return rq;
1109 raw_spin_unlock(&rq->lock);
1110 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1111 }
1112}
1113
1114static void __task_rq_unlock(struct rq *rq)
1115 __releases(rq->lock)
1116{
1117 raw_spin_unlock(&rq->lock);
1118}
1119
1120static inline void
1121task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
1122 __releases(rq->lock)
1123 __releases(p->pi_lock)
1124{
1125 raw_spin_unlock(&rq->lock);
1126 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1127}
1128
1129/*
1130 * this_rq_lock - lock this runqueue and disable interrupts.
1131 */
1132static struct rq *this_rq_lock(void)
1133 __acquires(rq->lock)
1134{
1135 struct rq *rq;
1136
1137 local_irq_disable();
1138 rq = this_rq();
1139 raw_spin_lock(&rq->lock);
1140
1141 return rq;
1142}
1143
1144#ifdef CONFIG_SCHED_HRTICK
1145/*
1146 * Use HR-timers to deliver accurate preemption points.
1147 *
1148 * Its all a bit involved since we cannot program an hrt while holding the
1149 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1150 * reschedule event.
1151 *
1152 * When we get rescheduled we reprogram the hrtick_timer outside of the
1153 * rq->lock.
1154 */
1155
1156/*
1157 * Use hrtick when:
1158 * - enabled by features
1159 * - hrtimer is actually high res
1160 */
1161static inline int hrtick_enabled(struct rq *rq)
1162{
1163 if (!sched_feat(HRTICK))
1164 return 0;
1165 if (!cpu_active(cpu_of(rq)))
1166 return 0;
1167 return hrtimer_is_hres_active(&rq->hrtick_timer);
1168}
1169
1170static void hrtick_clear(struct rq *rq)
1171{
1172 if (hrtimer_active(&rq->hrtick_timer))
1173 hrtimer_cancel(&rq->hrtick_timer);
1174}
1175
1176/*
1177 * High-resolution timer tick.
1178 * Runs from hardirq context with interrupts disabled.
1179 */
1180static enum hrtimer_restart hrtick(struct hrtimer *timer)
1181{
1182 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1183
1184 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1185
1186 raw_spin_lock(&rq->lock);
1187 update_rq_clock(rq);
1188 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1189 raw_spin_unlock(&rq->lock);
1190
1191 return HRTIMER_NORESTART;
1192}
1193
1194#ifdef CONFIG_SMP
1195/*
1196 * called from hardirq (IPI) context
1197 */
1198static void __hrtick_start(void *arg)
1199{
1200 struct rq *rq = arg;
1201
1202 raw_spin_lock(&rq->lock);
1203 hrtimer_restart(&rq->hrtick_timer);
1204 rq->hrtick_csd_pending = 0;
1205 raw_spin_unlock(&rq->lock);
1206}
1207
1208/*
1209 * Called to set the hrtick timer state.
1210 *
1211 * called with rq->lock held and irqs disabled
1212 */
1213static void hrtick_start(struct rq *rq, u64 delay)
1214{
1215 struct hrtimer *timer = &rq->hrtick_timer;
1216 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1217
1218 hrtimer_set_expires(timer, time);
1219
1220 if (rq == this_rq()) {
1221 hrtimer_restart(timer);
1222 } else if (!rq->hrtick_csd_pending) {
1223 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1224 rq->hrtick_csd_pending = 1;
1225 }
1226}
1227
1228static int
1229hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1230{
1231 int cpu = (int)(long)hcpu;
1232
1233 switch (action) {
1234 case CPU_UP_CANCELED:
1235 case CPU_UP_CANCELED_FROZEN:
1236 case CPU_DOWN_PREPARE:
1237 case CPU_DOWN_PREPARE_FROZEN:
1238 case CPU_DEAD:
1239 case CPU_DEAD_FROZEN:
1240 hrtick_clear(cpu_rq(cpu));
1241 return NOTIFY_OK;
1242 }
1243
1244 return NOTIFY_DONE;
1245}
1246
1247static __init void init_hrtick(void)
1248{
1249 hotcpu_notifier(hotplug_hrtick, 0);
1250}
1251#else
1252/*
1253 * Called to set the hrtick timer state.
1254 *
1255 * called with rq->lock held and irqs disabled
1256 */
1257static void hrtick_start(struct rq *rq, u64 delay)
1258{
1259 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1260 HRTIMER_MODE_REL_PINNED, 0);
1261}
1262
1263static inline void init_hrtick(void)
1264{
1265}
1266#endif /* CONFIG_SMP */
1267
1268static void init_rq_hrtick(struct rq *rq)
1269{
1270#ifdef CONFIG_SMP
1271 rq->hrtick_csd_pending = 0;
1272
1273 rq->hrtick_csd.flags = 0;
1274 rq->hrtick_csd.func = __hrtick_start;
1275 rq->hrtick_csd.info = rq;
1276#endif
1277
1278 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1279 rq->hrtick_timer.function = hrtick;
1280}
1281#else /* CONFIG_SCHED_HRTICK */
1282static inline void hrtick_clear(struct rq *rq)
1283{
1284}
1285
1286static inline void init_rq_hrtick(struct rq *rq)
1287{
1288}
1289
1290static inline void init_hrtick(void)
1291{
1292}
1293#endif /* CONFIG_SCHED_HRTICK */
1294
1295/*
1296 * resched_task - mark a task 'to be rescheduled now'.
1297 *
1298 * On UP this means the setting of the need_resched flag, on SMP it
1299 * might also involve a cross-CPU call to trigger the scheduler on
1300 * the target CPU.
1301 */
1302#ifdef CONFIG_SMP
1303
1304#ifndef tsk_is_polling
1305#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1306#endif
1307
1308static void resched_task(struct task_struct *p)
1309{
1310 int cpu;
1311
1312 assert_raw_spin_locked(&task_rq(p)->lock);
1313
1314 if (test_tsk_need_resched(p))
1315 return;
1316
1317 set_tsk_need_resched(p);
1318
1319 cpu = task_cpu(p);
1320 if (cpu == smp_processor_id())
1321 return;
1322
1323 /* NEED_RESCHED must be visible before we test polling */
1324 smp_mb();
1325 if (!tsk_is_polling(p))
1326 smp_send_reschedule(cpu);
1327}
1328
1329static void resched_cpu(int cpu)
1330{
1331 struct rq *rq = cpu_rq(cpu);
1332 unsigned long flags;
1333
1334 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1335 return;
1336 resched_task(cpu_curr(cpu));
1337 raw_spin_unlock_irqrestore(&rq->lock, flags);
1338}
1339
1340#ifdef CONFIG_NO_HZ
1341/*
1342 * In the semi idle case, use the nearest busy cpu for migrating timers
1343 * from an idle cpu. This is good for power-savings.
1344 *
1345 * We don't do similar optimization for completely idle system, as
1346 * selecting an idle cpu will add more delays to the timers than intended
1347 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1348 */
1349int get_nohz_timer_target(void)
1350{
1351 int cpu = smp_processor_id();
1352 int i;
1353 struct sched_domain *sd;
1354
1355 rcu_read_lock();
1356 for_each_domain(cpu, sd) {
1357 for_each_cpu(i, sched_domain_span(sd)) {
1358 if (!idle_cpu(i)) {
1359 cpu = i;
1360 goto unlock;
1361 }
1362 }
1363 }
1364unlock:
1365 rcu_read_unlock();
1366 return cpu;
1367}
1368/*
1369 * When add_timer_on() enqueues a timer into the timer wheel of an
1370 * idle CPU then this timer might expire before the next timer event
1371 * which is scheduled to wake up that CPU. In case of a completely
1372 * idle system the next event might even be infinite time into the
1373 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1374 * leaves the inner idle loop so the newly added timer is taken into
1375 * account when the CPU goes back to idle and evaluates the timer
1376 * wheel for the next timer event.
1377 */
1378void wake_up_idle_cpu(int cpu)
1379{
1380 struct rq *rq = cpu_rq(cpu);
1381
1382 if (cpu == smp_processor_id())
1383 return;
1384
1385 /*
1386 * This is safe, as this function is called with the timer
1387 * wheel base lock of (cpu) held. When the CPU is on the way
1388 * to idle and has not yet set rq->curr to idle then it will
1389 * be serialized on the timer wheel base lock and take the new
1390 * timer into account automatically.
1391 */
1392 if (rq->curr != rq->idle)
1393 return;
1394
1395 /*
1396 * We can set TIF_RESCHED on the idle task of the other CPU
1397 * lockless. The worst case is that the other CPU runs the
1398 * idle task through an additional NOOP schedule()
1399 */
1400 set_tsk_need_resched(rq->idle);
1401
1402 /* NEED_RESCHED must be visible before we test polling */
1403 smp_mb();
1404 if (!tsk_is_polling(rq->idle))
1405 smp_send_reschedule(cpu);
1406}
1407
1408static inline bool got_nohz_idle_kick(void)
1409{
1410 return idle_cpu(smp_processor_id()) && this_rq()->nohz_balance_kick;
1411}
1412
1413#else /* CONFIG_NO_HZ */
1414
1415static inline bool got_nohz_idle_kick(void)
1416{
1417 return false;
1418}
1419
1420#endif /* CONFIG_NO_HZ */
1421
1422static u64 sched_avg_period(void)
1423{
1424 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1425}
1426
1427static void sched_avg_update(struct rq *rq)
1428{
1429 s64 period = sched_avg_period();
1430
1431 while ((s64)(rq->clock - rq->age_stamp) > period) {
1432 /*
1433 * Inline assembly required to prevent the compiler
1434 * optimising this loop into a divmod call.
1435 * See __iter_div_u64_rem() for another example of this.
1436 */
1437 asm("" : "+rm" (rq->age_stamp));
1438 rq->age_stamp += period;
1439 rq->rt_avg /= 2;
1440 }
1441}
1442
1443static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1444{
1445 rq->rt_avg += rt_delta;
1446 sched_avg_update(rq);
1447}
1448
1449#else /* !CONFIG_SMP */
1450static void resched_task(struct task_struct *p)
1451{
1452 assert_raw_spin_locked(&task_rq(p)->lock);
1453 set_tsk_need_resched(p);
1454}
1455
1456static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1457{
1458}
1459
1460static void sched_avg_update(struct rq *rq)
1461{
1462}
1463#endif /* CONFIG_SMP */
1464
1465#if BITS_PER_LONG == 32
1466# define WMULT_CONST (~0UL)
1467#else
1468# define WMULT_CONST (1UL << 32)
1469#endif
1470
1471#define WMULT_SHIFT 32
1472
1473/*
1474 * Shift right and round:
1475 */
1476#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1477
1478/*
1479 * delta *= weight / lw
1480 */
1481static unsigned long
1482calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1483 struct load_weight *lw)
1484{
1485 u64 tmp;
1486
1487 /*
1488 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1489 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1490 * 2^SCHED_LOAD_RESOLUTION.
1491 */
1492 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1493 tmp = (u64)delta_exec * scale_load_down(weight);
1494 else
1495 tmp = (u64)delta_exec;
1496
1497 if (!lw->inv_weight) {
1498 unsigned long w = scale_load_down(lw->weight);
1499
1500 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1501 lw->inv_weight = 1;
1502 else if (unlikely(!w))
1503 lw->inv_weight = WMULT_CONST;
1504 else
1505 lw->inv_weight = WMULT_CONST / w;
1506 }
1507
1508 /*
1509 * Check whether we'd overflow the 64-bit multiplication:
1510 */
1511 if (unlikely(tmp > WMULT_CONST))
1512 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1513 WMULT_SHIFT/2);
1514 else
1515 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1516
1517 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1518}
1519
1520static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1521{
1522 lw->weight += inc;
1523 lw->inv_weight = 0;
1524}
1525
1526static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1527{
1528 lw->weight -= dec;
1529 lw->inv_weight = 0;
1530}
1531
1532static inline void update_load_set(struct load_weight *lw, unsigned long w)
1533{
1534 lw->weight = w;
1535 lw->inv_weight = 0;
1536}
1537
1538/*
1539 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1540 * of tasks with abnormal "nice" values across CPUs the contribution that
1541 * each task makes to its run queue's load is weighted according to its
1542 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1543 * scaled version of the new time slice allocation that they receive on time
1544 * slice expiry etc.
1545 */
1546
1547#define WEIGHT_IDLEPRIO 3
1548#define WMULT_IDLEPRIO 1431655765
1549
1550/*
1551 * Nice levels are multiplicative, with a gentle 10% change for every
1552 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1553 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1554 * that remained on nice 0.
1555 *
1556 * The "10% effect" is relative and cumulative: from _any_ nice level,
1557 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1558 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1559 * If a task goes up by ~10% and another task goes down by ~10% then
1560 * the relative distance between them is ~25%.)
1561 */
1562static const int prio_to_weight[40] = {
1563 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1564 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1565 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1566 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1567 /* 0 */ 1024, 820, 655, 526, 423,
1568 /* 5 */ 335, 272, 215, 172, 137,
1569 /* 10 */ 110, 87, 70, 56, 45,
1570 /* 15 */ 36, 29, 23, 18, 15,
1571};
1572
1573/*
1574 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1575 *
1576 * In cases where the weight does not change often, we can use the
1577 * precalculated inverse to speed up arithmetics by turning divisions
1578 * into multiplications:
1579 */
1580static const u32 prio_to_wmult[40] = {
1581 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1582 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1583 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1584 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1585 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1586 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1587 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1588 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1589};
1590
1591/* Time spent by the tasks of the cpu accounting group executing in ... */
1592enum cpuacct_stat_index {
1593 CPUACCT_STAT_USER, /* ... user mode */
1594 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1595
1596 CPUACCT_STAT_NSTATS,
1597};
1598
1599#ifdef CONFIG_CGROUP_CPUACCT
1600static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1601static void cpuacct_update_stats(struct task_struct *tsk,
1602 enum cpuacct_stat_index idx, cputime_t val);
1603#else
1604static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1605static inline void cpuacct_update_stats(struct task_struct *tsk,
1606 enum cpuacct_stat_index idx, cputime_t val) {}
1607#endif
1608
1609static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1610{
1611 update_load_add(&rq->load, load);
1612}
1613
1614static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1615{
1616 update_load_sub(&rq->load, load);
1617}
1618
1619#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1620 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1621typedef int (*tg_visitor)(struct task_group *, void *);
1622
1623/*
1624 * Iterate task_group tree rooted at *from, calling @down when first entering a
1625 * node and @up when leaving it for the final time.
1626 *
1627 * Caller must hold rcu_lock or sufficient equivalent.
1628 */
1629static int walk_tg_tree_from(struct task_group *from,
1630 tg_visitor down, tg_visitor up, void *data)
1631{
1632 struct task_group *parent, *child;
1633 int ret;
1634
1635 parent = from;
1636
1637down:
1638 ret = (*down)(parent, data);
1639 if (ret)
1640 goto out;
1641 list_for_each_entry_rcu(child, &parent->children, siblings) {
1642 parent = child;
1643 goto down;
1644
1645up:
1646 continue;
1647 }
1648 ret = (*up)(parent, data);
1649 if (ret || parent == from)
1650 goto out;
1651
1652 child = parent;
1653 parent = parent->parent;
1654 if (parent)
1655 goto up;
1656out:
1657 return ret;
1658}
1659
1660/*
1661 * Iterate the full tree, calling @down when first entering a node and @up when
1662 * leaving it for the final time.
1663 *
1664 * Caller must hold rcu_lock or sufficient equivalent.
1665 */
1666
1667static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1668{
1669 return walk_tg_tree_from(&root_task_group, down, up, data);
1670}
1671
1672static int tg_nop(struct task_group *tg, void *data)
1673{
1674 return 0;
1675}
1676#endif
1677
1678#ifdef CONFIG_SMP
1679/* Used instead of source_load when we know the type == 0 */
1680static unsigned long weighted_cpuload(const int cpu)
1681{
1682 return cpu_rq(cpu)->load.weight;
1683}
1684
1685/*
1686 * Return a low guess at the load of a migration-source cpu weighted
1687 * according to the scheduling class and "nice" value.
1688 *
1689 * We want to under-estimate the load of migration sources, to
1690 * balance conservatively.
1691 */
1692static unsigned long source_load(int cpu, int type)
1693{
1694 struct rq *rq = cpu_rq(cpu);
1695 unsigned long total = weighted_cpuload(cpu);
1696
1697 if (type == 0 || !sched_feat(LB_BIAS))
1698 return total;
1699
1700 return min(rq->cpu_load[type-1], total);
1701}
1702
1703/*
1704 * Return a high guess at the load of a migration-target cpu weighted
1705 * according to the scheduling class and "nice" value.
1706 */
1707static unsigned long target_load(int cpu, int type)
1708{
1709 struct rq *rq = cpu_rq(cpu);
1710 unsigned long total = weighted_cpuload(cpu);
1711
1712 if (type == 0 || !sched_feat(LB_BIAS))
1713 return total;
1714
1715 return max(rq->cpu_load[type-1], total);
1716}
1717
1718static unsigned long power_of(int cpu)
1719{
1720 return cpu_rq(cpu)->cpu_power;
1721}
1722
1723static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1724
1725static unsigned long cpu_avg_load_per_task(int cpu)
1726{
1727 struct rq *rq = cpu_rq(cpu);
1728 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1729
1730 if (nr_running)
1731 return rq->load.weight / nr_running;
1732
1733 return 0;
1734}
1735
1736#ifdef CONFIG_PREEMPT
1737
1738static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1739
1740/*
1741 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1742 * way at the expense of forcing extra atomic operations in all
1743 * invocations. This assures that the double_lock is acquired using the
1744 * same underlying policy as the spinlock_t on this architecture, which
1745 * reduces latency compared to the unfair variant below. However, it
1746 * also adds more overhead and therefore may reduce throughput.
1747 */
1748static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1749 __releases(this_rq->lock)
1750 __acquires(busiest->lock)
1751 __acquires(this_rq->lock)
1752{
1753 raw_spin_unlock(&this_rq->lock);
1754 double_rq_lock(this_rq, busiest);
1755
1756 return 1;
1757}
1758
1759#else
1760/*
1761 * Unfair double_lock_balance: Optimizes throughput at the expense of
1762 * latency by eliminating extra atomic operations when the locks are
1763 * already in proper order on entry. This favors lower cpu-ids and will
1764 * grant the double lock to lower cpus over higher ids under contention,
1765 * regardless of entry order into the function.
1766 */
1767static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1768 __releases(this_rq->lock)
1769 __acquires(busiest->lock)
1770 __acquires(this_rq->lock)
1771{
1772 int ret = 0;
1773
1774 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1775 if (busiest < this_rq) {
1776 raw_spin_unlock(&this_rq->lock);
1777 raw_spin_lock(&busiest->lock);
1778 raw_spin_lock_nested(&this_rq->lock,
1779 SINGLE_DEPTH_NESTING);
1780 ret = 1;
1781 } else
1782 raw_spin_lock_nested(&busiest->lock,
1783 SINGLE_DEPTH_NESTING);
1784 }
1785 return ret;
1786}
1787
1788#endif /* CONFIG_PREEMPT */
1789
1790/*
1791 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1792 */
1793static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1794{
1795 if (unlikely(!irqs_disabled())) {
1796 /* printk() doesn't work good under rq->lock */
1797 raw_spin_unlock(&this_rq->lock);
1798 BUG_ON(1);
1799 }
1800
1801 return _double_lock_balance(this_rq, busiest);
1802}
1803
1804static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1805 __releases(busiest->lock)
1806{
1807 raw_spin_unlock(&busiest->lock);
1808 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1809}
1810
1811/*
1812 * double_rq_lock - safely lock two runqueues
1813 *
1814 * Note this does not disable interrupts like task_rq_lock,
1815 * you need to do so manually before calling.
1816 */
1817static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1818 __acquires(rq1->lock)
1819 __acquires(rq2->lock)
1820{
1821 BUG_ON(!irqs_disabled());
1822 if (rq1 == rq2) {
1823 raw_spin_lock(&rq1->lock);
1824 __acquire(rq2->lock); /* Fake it out ;) */
1825 } else {
1826 if (rq1 < rq2) {
1827 raw_spin_lock(&rq1->lock);
1828 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1829 } else {
1830 raw_spin_lock(&rq2->lock);
1831 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1832 }
1833 }
1834}
1835
1836/*
1837 * double_rq_unlock - safely unlock two runqueues
1838 *
1839 * Note this does not restore interrupts like task_rq_unlock,
1840 * you need to do so manually after calling.
1841 */
1842static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1843 __releases(rq1->lock)
1844 __releases(rq2->lock)
1845{
1846 raw_spin_unlock(&rq1->lock);
1847 if (rq1 != rq2)
1848 raw_spin_unlock(&rq2->lock);
1849 else
1850 __release(rq2->lock);
1851}
1852
1853#else /* CONFIG_SMP */
1854
1855/*
1856 * double_rq_lock - safely lock two runqueues
1857 *
1858 * Note this does not disable interrupts like task_rq_lock,
1859 * you need to do so manually before calling.
1860 */
1861static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1862 __acquires(rq1->lock)
1863 __acquires(rq2->lock)
1864{
1865 BUG_ON(!irqs_disabled());
1866 BUG_ON(rq1 != rq2);
1867 raw_spin_lock(&rq1->lock);
1868 __acquire(rq2->lock); /* Fake it out ;) */
1869}
1870
1871/*
1872 * double_rq_unlock - safely unlock two runqueues
1873 *
1874 * Note this does not restore interrupts like task_rq_unlock,
1875 * you need to do so manually after calling.
1876 */
1877static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1878 __releases(rq1->lock)
1879 __releases(rq2->lock)
1880{
1881 BUG_ON(rq1 != rq2);
1882 raw_spin_unlock(&rq1->lock);
1883 __release(rq2->lock);
1884}
1885
1886#endif
1887
1888static void calc_load_account_idle(struct rq *this_rq);
1889static void update_sysctl(void);
1890static int get_update_sysctl_factor(void);
1891static void update_cpu_load(struct rq *this_rq);
1892
1893static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1894{
1895 set_task_rq(p, cpu);
1896#ifdef CONFIG_SMP
1897 /*
1898 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1899 * successfully executed on another CPU. We must ensure that updates of
1900 * per-task data have been completed by this moment.
1901 */
1902 smp_wmb();
1903 task_thread_info(p)->cpu = cpu;
1904#endif
1905}
1906
1907static const struct sched_class rt_sched_class;
1908
1909#define sched_class_highest (&stop_sched_class)
1910#define for_each_class(class) \
1911 for (class = sched_class_highest; class; class = class->next)
1912
1913#include "sched_stats.h"
1914
1915static void inc_nr_running(struct rq *rq)
1916{
1917 rq->nr_running++;
1918}
1919
1920static void dec_nr_running(struct rq *rq)
1921{
1922 rq->nr_running--;
1923}
1924
1925static void set_load_weight(struct task_struct *p)
1926{
1927 int prio = p->static_prio - MAX_RT_PRIO;
1928 struct load_weight *load = &p->se.load;
1929
1930 /*
1931 * SCHED_IDLE tasks get minimal weight:
1932 */
1933 if (p->policy == SCHED_IDLE) {
1934 load->weight = scale_load(WEIGHT_IDLEPRIO);
1935 load->inv_weight = WMULT_IDLEPRIO;
1936 return;
1937 }
1938
1939 load->weight = scale_load(prio_to_weight[prio]);
1940 load->inv_weight = prio_to_wmult[prio];
1941}
1942
1943static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1944{
1945 update_rq_clock(rq);
1946 sched_info_queued(p);
1947 p->sched_class->enqueue_task(rq, p, flags);
1948}
1949
1950static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1951{
1952 update_rq_clock(rq);
1953 sched_info_dequeued(p);
1954 p->sched_class->dequeue_task(rq, p, flags);
1955}
1956
1957/*
1958 * activate_task - move a task to the runqueue.
1959 */
1960static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1961{
1962 if (task_contributes_to_load(p))
1963 rq->nr_uninterruptible--;
1964
1965 enqueue_task(rq, p, flags);
1966}
1967
1968/*
1969 * deactivate_task - remove a task from the runqueue.
1970 */
1971static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1972{
1973 if (task_contributes_to_load(p))
1974 rq->nr_uninterruptible++;
1975
1976 dequeue_task(rq, p, flags);
1977}
1978
1979#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1980
1981/*
1982 * There are no locks covering percpu hardirq/softirq time.
1983 * They are only modified in account_system_vtime, on corresponding CPU
1984 * with interrupts disabled. So, writes are safe.
1985 * They are read and saved off onto struct rq in update_rq_clock().
1986 * This may result in other CPU reading this CPU's irq time and can
1987 * race with irq/account_system_vtime on this CPU. We would either get old
1988 * or new value with a side effect of accounting a slice of irq time to wrong
1989 * task when irq is in progress while we read rq->clock. That is a worthy
1990 * compromise in place of having locks on each irq in account_system_time.
1991 */
1992static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1993static DEFINE_PER_CPU(u64, cpu_softirq_time);
1994
1995static DEFINE_PER_CPU(u64, irq_start_time);
1996static int sched_clock_irqtime;
1997
1998void enable_sched_clock_irqtime(void)
1999{
2000 sched_clock_irqtime = 1;
2001}
2002
2003void disable_sched_clock_irqtime(void)
2004{
2005 sched_clock_irqtime = 0;
2006}
2007
2008#ifndef CONFIG_64BIT
2009static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
2010
2011static inline void irq_time_write_begin(void)
2012{
2013 __this_cpu_inc(irq_time_seq.sequence);
2014 smp_wmb();
2015}
2016
2017static inline void irq_time_write_end(void)
2018{
2019 smp_wmb();
2020 __this_cpu_inc(irq_time_seq.sequence);
2021}
2022
2023static inline u64 irq_time_read(int cpu)
2024{
2025 u64 irq_time;
2026 unsigned seq;
2027
2028 do {
2029 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
2030 irq_time = per_cpu(cpu_softirq_time, cpu) +
2031 per_cpu(cpu_hardirq_time, cpu);
2032 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
2033
2034 return irq_time;
2035}
2036#else /* CONFIG_64BIT */
2037static inline void irq_time_write_begin(void)
2038{
2039}
2040
2041static inline void irq_time_write_end(void)
2042{
2043}
2044
2045static inline u64 irq_time_read(int cpu)
2046{
2047 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
2048}
2049#endif /* CONFIG_64BIT */
2050
2051/*
2052 * Called before incrementing preempt_count on {soft,}irq_enter
2053 * and before decrementing preempt_count on {soft,}irq_exit.
2054 */
2055void account_system_vtime(struct task_struct *curr)
2056{
2057 unsigned long flags;
2058 s64 delta;
2059 int cpu;
2060
2061 if (!sched_clock_irqtime)
2062 return;
2063
2064 local_irq_save(flags);
2065
2066 cpu = smp_processor_id();
2067 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
2068 __this_cpu_add(irq_start_time, delta);
2069
2070 irq_time_write_begin();
2071 /*
2072 * We do not account for softirq time from ksoftirqd here.
2073 * We want to continue accounting softirq time to ksoftirqd thread
2074 * in that case, so as not to confuse scheduler with a special task
2075 * that do not consume any time, but still wants to run.
2076 */
2077 if (hardirq_count())
2078 __this_cpu_add(cpu_hardirq_time, delta);
2079 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
2080 __this_cpu_add(cpu_softirq_time, delta);
2081
2082 irq_time_write_end();
2083 local_irq_restore(flags);
2084}
2085EXPORT_SYMBOL_GPL(account_system_vtime);
2086
2087#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2088
2089#ifdef CONFIG_PARAVIRT
2090static inline u64 steal_ticks(u64 steal)
2091{
2092 if (unlikely(steal > NSEC_PER_SEC))
2093 return div_u64(steal, TICK_NSEC);
2094
2095 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
2096}
2097#endif
2098
2099static void update_rq_clock_task(struct rq *rq, s64 delta)
2100{
2101/*
2102 * In theory, the compile should just see 0 here, and optimize out the call
2103 * to sched_rt_avg_update. But I don't trust it...
2104 */
2105#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2106 s64 steal = 0, irq_delta = 0;
2107#endif
2108#ifdef CONFIG_IRQ_TIME_ACCOUNTING
2109 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
2110
2111 /*
2112 * Since irq_time is only updated on {soft,}irq_exit, we might run into
2113 * this case when a previous update_rq_clock() happened inside a
2114 * {soft,}irq region.
2115 *
2116 * When this happens, we stop ->clock_task and only update the
2117 * prev_irq_time stamp to account for the part that fit, so that a next
2118 * update will consume the rest. This ensures ->clock_task is
2119 * monotonic.
2120 *
2121 * It does however cause some slight miss-attribution of {soft,}irq
2122 * time, a more accurate solution would be to update the irq_time using
2123 * the current rq->clock timestamp, except that would require using
2124 * atomic ops.
2125 */
2126 if (irq_delta > delta)
2127 irq_delta = delta;
2128
2129 rq->prev_irq_time += irq_delta;
2130 delta -= irq_delta;
2131#endif
2132#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
2133 if (static_branch((&paravirt_steal_rq_enabled))) {
2134 u64 st;
2135
2136 steal = paravirt_steal_clock(cpu_of(rq));
2137 steal -= rq->prev_steal_time_rq;
2138
2139 if (unlikely(steal > delta))
2140 steal = delta;
2141
2142 st = steal_ticks(steal);
2143 steal = st * TICK_NSEC;
2144
2145 rq->prev_steal_time_rq += steal;
2146
2147 delta -= steal;
2148 }
2149#endif
2150
2151 rq->clock_task += delta;
2152
2153#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2154 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
2155 sched_rt_avg_update(rq, irq_delta + steal);
2156#endif
2157}
2158
2159#ifdef CONFIG_IRQ_TIME_ACCOUNTING
2160static int irqtime_account_hi_update(void)
2161{
2162 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2163 unsigned long flags;
2164 u64 latest_ns;
2165 int ret = 0;
2166
2167 local_irq_save(flags);
2168 latest_ns = this_cpu_read(cpu_hardirq_time);
2169 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
2170 ret = 1;
2171 local_irq_restore(flags);
2172 return ret;
2173}
2174
2175static int irqtime_account_si_update(void)
2176{
2177 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2178 unsigned long flags;
2179 u64 latest_ns;
2180 int ret = 0;
2181
2182 local_irq_save(flags);
2183 latest_ns = this_cpu_read(cpu_softirq_time);
2184 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2185 ret = 1;
2186 local_irq_restore(flags);
2187 return ret;
2188}
2189
2190#else /* CONFIG_IRQ_TIME_ACCOUNTING */
2191
2192#define sched_clock_irqtime (0)
2193
2194#endif
2195
2196#include "sched_idletask.c"
2197#include "sched_fair.c"
2198#include "sched_rt.c"
2199#include "sched_autogroup.c"
2200#include "sched_stoptask.c"
2201#ifdef CONFIG_SCHED_DEBUG
2202# include "sched_debug.c"
2203#endif
2204
2205void sched_set_stop_task(int cpu, struct task_struct *stop)
2206{
2207 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2208 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2209
2210 if (stop) {
2211 /*
2212 * Make it appear like a SCHED_FIFO task, its something
2213 * userspace knows about and won't get confused about.
2214 *
2215 * Also, it will make PI more or less work without too
2216 * much confusion -- but then, stop work should not
2217 * rely on PI working anyway.
2218 */
2219 sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
2220
2221 stop->sched_class = &stop_sched_class;
2222 }
2223
2224 cpu_rq(cpu)->stop = stop;
2225
2226 if (old_stop) {
2227 /*
2228 * Reset it back to a normal scheduling class so that
2229 * it can die in pieces.
2230 */
2231 old_stop->sched_class = &rt_sched_class;
2232 }
2233}
2234
2235/*
2236 * __normal_prio - return the priority that is based on the static prio
2237 */
2238static inline int __normal_prio(struct task_struct *p)
2239{
2240 return p->static_prio;
2241}
2242
2243/*
2244 * Calculate the expected normal priority: i.e. priority
2245 * without taking RT-inheritance into account. Might be
2246 * boosted by interactivity modifiers. Changes upon fork,
2247 * setprio syscalls, and whenever the interactivity
2248 * estimator recalculates.
2249 */
2250static inline int normal_prio(struct task_struct *p)
2251{
2252 int prio;
2253
2254 if (task_has_rt_policy(p))
2255 prio = MAX_RT_PRIO-1 - p->rt_priority;
2256 else
2257 prio = __normal_prio(p);
2258 return prio;
2259}
2260
2261/*
2262 * Calculate the current priority, i.e. the priority
2263 * taken into account by the scheduler. This value might
2264 * be boosted by RT tasks, or might be boosted by
2265 * interactivity modifiers. Will be RT if the task got
2266 * RT-boosted. If not then it returns p->normal_prio.
2267 */
2268static int effective_prio(struct task_struct *p)
2269{
2270 p->normal_prio = normal_prio(p);
2271 /*
2272 * If we are RT tasks or we were boosted to RT priority,
2273 * keep the priority unchanged. Otherwise, update priority
2274 * to the normal priority:
2275 */
2276 if (!rt_prio(p->prio))
2277 return p->normal_prio;
2278 return p->prio;
2279}
2280
2281/**
2282 * task_curr - is this task currently executing on a CPU?
2283 * @p: the task in question.
2284 */
2285inline int task_curr(const struct task_struct *p)
2286{
2287 return cpu_curr(task_cpu(p)) == p;
2288}
2289
2290static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2291 const struct sched_class *prev_class,
2292 int oldprio)
2293{
2294 if (prev_class != p->sched_class) {
2295 if (prev_class->switched_from)
2296 prev_class->switched_from(rq, p);
2297 p->sched_class->switched_to(rq, p);
2298 } else if (oldprio != p->prio)
2299 p->sched_class->prio_changed(rq, p, oldprio);
2300}
2301
2302static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2303{
2304 const struct sched_class *class;
2305
2306 if (p->sched_class == rq->curr->sched_class) {
2307 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2308 } else {
2309 for_each_class(class) {
2310 if (class == rq->curr->sched_class)
2311 break;
2312 if (class == p->sched_class) {
2313 resched_task(rq->curr);
2314 break;
2315 }
2316 }
2317 }
2318
2319 /*
2320 * A queue event has occurred, and we're going to schedule. In
2321 * this case, we can save a useless back to back clock update.
2322 */
2323 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2324 rq->skip_clock_update = 1;
2325}
2326
2327#ifdef CONFIG_SMP
2328/*
2329 * Is this task likely cache-hot:
2330 */
2331static int
2332task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2333{
2334 s64 delta;
2335
2336 if (p->sched_class != &fair_sched_class)
2337 return 0;
2338
2339 if (unlikely(p->policy == SCHED_IDLE))
2340 return 0;
2341
2342 /*
2343 * Buddy candidates are cache hot:
2344 */
2345 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2346 (&p->se == cfs_rq_of(&p->se)->next ||
2347 &p->se == cfs_rq_of(&p->se)->last))
2348 return 1;
2349
2350 if (sysctl_sched_migration_cost == -1)
2351 return 1;
2352 if (sysctl_sched_migration_cost == 0)
2353 return 0;
2354
2355 delta = now - p->se.exec_start;
2356
2357 return delta < (s64)sysctl_sched_migration_cost;
2358}
2359
2360void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2361{
2362#ifdef CONFIG_SCHED_DEBUG
2363 /*
2364 * We should never call set_task_cpu() on a blocked task,
2365 * ttwu() will sort out the placement.
2366 */
2367 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2368 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2369
2370#ifdef CONFIG_LOCKDEP
2371 /*
2372 * The caller should hold either p->pi_lock or rq->lock, when changing
2373 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2374 *
2375 * sched_move_task() holds both and thus holding either pins the cgroup,
2376 * see set_task_rq().
2377 *
2378 * Furthermore, all task_rq users should acquire both locks, see
2379 * task_rq_lock().
2380 */
2381 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2382 lockdep_is_held(&task_rq(p)->lock)));
2383#endif
2384#endif
2385
2386 trace_sched_migrate_task(p, new_cpu);
2387
2388 if (task_cpu(p) != new_cpu) {
2389 p->se.nr_migrations++;
2390 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
2391 }
2392
2393 __set_task_cpu(p, new_cpu);
2394}
2395
2396struct migration_arg {
2397 struct task_struct *task;
2398 int dest_cpu;
2399};
2400
2401static int migration_cpu_stop(void *data);
2402
2403/*
2404 * wait_task_inactive - wait for a thread to unschedule.
2405 *
2406 * If @match_state is nonzero, it's the @p->state value just checked and
2407 * not expected to change. If it changes, i.e. @p might have woken up,
2408 * then return zero. When we succeed in waiting for @p to be off its CPU,
2409 * we return a positive number (its total switch count). If a second call
2410 * a short while later returns the same number, the caller can be sure that
2411 * @p has remained unscheduled the whole time.
2412 *
2413 * The caller must ensure that the task *will* unschedule sometime soon,
2414 * else this function might spin for a *long* time. This function can't
2415 * be called with interrupts off, or it may introduce deadlock with
2416 * smp_call_function() if an IPI is sent by the same process we are
2417 * waiting to become inactive.
2418 */
2419unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2420{
2421 unsigned long flags;
2422 int running, on_rq;
2423 unsigned long ncsw;
2424 struct rq *rq;
2425
2426 for (;;) {
2427 /*
2428 * We do the initial early heuristics without holding
2429 * any task-queue locks at all. We'll only try to get
2430 * the runqueue lock when things look like they will
2431 * work out!
2432 */
2433 rq = task_rq(p);
2434
2435 /*
2436 * If the task is actively running on another CPU
2437 * still, just relax and busy-wait without holding
2438 * any locks.
2439 *
2440 * NOTE! Since we don't hold any locks, it's not
2441 * even sure that "rq" stays as the right runqueue!
2442 * But we don't care, since "task_running()" will
2443 * return false if the runqueue has changed and p
2444 * is actually now running somewhere else!
2445 */
2446 while (task_running(rq, p)) {
2447 if (match_state && unlikely(p->state != match_state))
2448 return 0;
2449 cpu_relax();
2450 }
2451
2452 /*
2453 * Ok, time to look more closely! We need the rq
2454 * lock now, to be *sure*. If we're wrong, we'll
2455 * just go back and repeat.
2456 */
2457 rq = task_rq_lock(p, &flags);
2458 trace_sched_wait_task(p);
2459 running = task_running(rq, p);
2460 on_rq = p->on_rq;
2461 ncsw = 0;
2462 if (!match_state || p->state == match_state)
2463 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2464 task_rq_unlock(rq, p, &flags);
2465
2466 /*
2467 * If it changed from the expected state, bail out now.
2468 */
2469 if (unlikely(!ncsw))
2470 break;
2471
2472 /*
2473 * Was it really running after all now that we
2474 * checked with the proper locks actually held?
2475 *
2476 * Oops. Go back and try again..
2477 */
2478 if (unlikely(running)) {
2479 cpu_relax();
2480 continue;
2481 }
2482
2483 /*
2484 * It's not enough that it's not actively running,
2485 * it must be off the runqueue _entirely_, and not
2486 * preempted!
2487 *
2488 * So if it was still runnable (but just not actively
2489 * running right now), it's preempted, and we should
2490 * yield - it could be a while.
2491 */
2492 if (unlikely(on_rq)) {
2493 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2494
2495 set_current_state(TASK_UNINTERRUPTIBLE);
2496 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2497 continue;
2498 }
2499
2500 /*
2501 * Ahh, all good. It wasn't running, and it wasn't
2502 * runnable, which means that it will never become
2503 * running in the future either. We're all done!
2504 */
2505 break;
2506 }
2507
2508 return ncsw;
2509}
2510
2511/***
2512 * kick_process - kick a running thread to enter/exit the kernel
2513 * @p: the to-be-kicked thread
2514 *
2515 * Cause a process which is running on another CPU to enter
2516 * kernel-mode, without any delay. (to get signals handled.)
2517 *
2518 * NOTE: this function doesn't have to take the runqueue lock,
2519 * because all it wants to ensure is that the remote task enters
2520 * the kernel. If the IPI races and the task has been migrated
2521 * to another CPU then no harm is done and the purpose has been
2522 * achieved as well.
2523 */
2524void kick_process(struct task_struct *p)
2525{
2526 int cpu;
2527
2528 preempt_disable();
2529 cpu = task_cpu(p);
2530 if ((cpu != smp_processor_id()) && task_curr(p))
2531 smp_send_reschedule(cpu);
2532 preempt_enable();
2533}
2534EXPORT_SYMBOL_GPL(kick_process);
2535#endif /* CONFIG_SMP */
2536
2537#ifdef CONFIG_SMP
2538/*
2539 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2540 */
2541static int select_fallback_rq(int cpu, struct task_struct *p)
2542{
2543 int dest_cpu;
2544 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2545
2546 /* Look for allowed, online CPU in same node. */
2547 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2548 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
2549 return dest_cpu;
2550
2551 /* Any allowed, online CPU? */
2552 dest_cpu = cpumask_any_and(tsk_cpus_allowed(p), cpu_active_mask);
2553 if (dest_cpu < nr_cpu_ids)
2554 return dest_cpu;
2555
2556 /* No more Mr. Nice Guy. */
2557 dest_cpu = cpuset_cpus_allowed_fallback(p);
2558 /*
2559 * Don't tell them about moving exiting tasks or
2560 * kernel threads (both mm NULL), since they never
2561 * leave kernel.
2562 */
2563 if (p->mm && printk_ratelimit()) {
2564 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2565 task_pid_nr(p), p->comm, cpu);
2566 }
2567
2568 return dest_cpu;
2569}
2570
2571/*
2572 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2573 */
2574static inline
2575int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2576{
2577 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2578
2579 /*
2580 * In order not to call set_task_cpu() on a blocking task we need
2581 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2582 * cpu.
2583 *
2584 * Since this is common to all placement strategies, this lives here.
2585 *
2586 * [ this allows ->select_task() to simply return task_cpu(p) and
2587 * not worry about this generic constraint ]
2588 */
2589 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
2590 !cpu_online(cpu)))
2591 cpu = select_fallback_rq(task_cpu(p), p);
2592
2593 return cpu;
2594}
2595
2596static void update_avg(u64 *avg, u64 sample)
2597{
2598 s64 diff = sample - *avg;
2599 *avg += diff >> 3;
2600}
2601#endif
2602
2603static void
2604ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2605{
2606#ifdef CONFIG_SCHEDSTATS
2607 struct rq *rq = this_rq();
2608
2609#ifdef CONFIG_SMP
2610 int this_cpu = smp_processor_id();
2611
2612 if (cpu == this_cpu) {
2613 schedstat_inc(rq, ttwu_local);
2614 schedstat_inc(p, se.statistics.nr_wakeups_local);
2615 } else {
2616 struct sched_domain *sd;
2617
2618 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2619 rcu_read_lock();
2620 for_each_domain(this_cpu, sd) {
2621 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2622 schedstat_inc(sd, ttwu_wake_remote);
2623 break;
2624 }
2625 }
2626 rcu_read_unlock();
2627 }
2628
2629 if (wake_flags & WF_MIGRATED)
2630 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2631
2632#endif /* CONFIG_SMP */
2633
2634 schedstat_inc(rq, ttwu_count);
2635 schedstat_inc(p, se.statistics.nr_wakeups);
2636
2637 if (wake_flags & WF_SYNC)
2638 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2639
2640#endif /* CONFIG_SCHEDSTATS */
2641}
2642
2643static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2644{
2645 activate_task(rq, p, en_flags);
2646 p->on_rq = 1;
2647
2648 /* if a worker is waking up, notify workqueue */
2649 if (p->flags & PF_WQ_WORKER)
2650 wq_worker_waking_up(p, cpu_of(rq));
2651}
2652
2653/*
2654 * Mark the task runnable and perform wakeup-preemption.
2655 */
2656static void
2657ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2658{
2659 trace_sched_wakeup(p, true);
2660 check_preempt_curr(rq, p, wake_flags);
2661
2662 p->state = TASK_RUNNING;
2663#ifdef CONFIG_SMP
2664 if (p->sched_class->task_woken)
2665 p->sched_class->task_woken(rq, p);
2666
2667 if (rq->idle_stamp) {
2668 u64 delta = rq->clock - rq->idle_stamp;
2669 u64 max = 2*sysctl_sched_migration_cost;
2670
2671 if (delta > max)
2672 rq->avg_idle = max;
2673 else
2674 update_avg(&rq->avg_idle, delta);
2675 rq->idle_stamp = 0;
2676 }
2677#endif
2678}
2679
2680static void
2681ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2682{
2683#ifdef CONFIG_SMP
2684 if (p->sched_contributes_to_load)
2685 rq->nr_uninterruptible--;
2686#endif
2687
2688 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2689 ttwu_do_wakeup(rq, p, wake_flags);
2690}
2691
2692/*
2693 * Called in case the task @p isn't fully descheduled from its runqueue,
2694 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2695 * since all we need to do is flip p->state to TASK_RUNNING, since
2696 * the task is still ->on_rq.
2697 */
2698static int ttwu_remote(struct task_struct *p, int wake_flags)
2699{
2700 struct rq *rq;
2701 int ret = 0;
2702
2703 rq = __task_rq_lock(p);
2704 if (p->on_rq) {
2705 ttwu_do_wakeup(rq, p, wake_flags);
2706 ret = 1;
2707 }
2708 __task_rq_unlock(rq);
2709
2710 return ret;
2711}
2712
2713#ifdef CONFIG_SMP
2714static void sched_ttwu_pending(void)
2715{
2716 struct rq *rq = this_rq();
2717 struct llist_node *llist = llist_del_all(&rq->wake_list);
2718 struct task_struct *p;
2719
2720 raw_spin_lock(&rq->lock);
2721
2722 while (llist) {
2723 p = llist_entry(llist, struct task_struct, wake_entry);
2724 llist = llist_next(llist);
2725 ttwu_do_activate(rq, p, 0);
2726 }
2727
2728 raw_spin_unlock(&rq->lock);
2729}
2730
2731void scheduler_ipi(void)
2732{
2733 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
2734 return;
2735
2736 /*
2737 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2738 * traditionally all their work was done from the interrupt return
2739 * path. Now that we actually do some work, we need to make sure
2740 * we do call them.
2741 *
2742 * Some archs already do call them, luckily irq_enter/exit nest
2743 * properly.
2744 *
2745 * Arguably we should visit all archs and update all handlers,
2746 * however a fair share of IPIs are still resched only so this would
2747 * somewhat pessimize the simple resched case.
2748 */
2749 irq_enter();
2750 sched_ttwu_pending();
2751
2752 /*
2753 * Check if someone kicked us for doing the nohz idle load balance.
2754 */
2755 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
2756 this_rq()->idle_balance = 1;
2757 raise_softirq_irqoff(SCHED_SOFTIRQ);
2758 }
2759 irq_exit();
2760}
2761
2762static void ttwu_queue_remote(struct task_struct *p, int cpu)
2763{
2764 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
2765 smp_send_reschedule(cpu);
2766}
2767
2768#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2769static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2770{
2771 struct rq *rq;
2772 int ret = 0;
2773
2774 rq = __task_rq_lock(p);
2775 if (p->on_cpu) {
2776 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2777 ttwu_do_wakeup(rq, p, wake_flags);
2778 ret = 1;
2779 }
2780 __task_rq_unlock(rq);
2781
2782 return ret;
2783
2784}
2785#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2786#endif /* CONFIG_SMP */
2787
2788static void ttwu_queue(struct task_struct *p, int cpu)
2789{
2790 struct rq *rq = cpu_rq(cpu);
2791
2792#if defined(CONFIG_SMP)
2793 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2794 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2795 ttwu_queue_remote(p, cpu);
2796 return;
2797 }
2798#endif
2799
2800 raw_spin_lock(&rq->lock);
2801 ttwu_do_activate(rq, p, 0);
2802 raw_spin_unlock(&rq->lock);
2803}
2804
2805/**
2806 * try_to_wake_up - wake up a thread
2807 * @p: the thread to be awakened
2808 * @state: the mask of task states that can be woken
2809 * @wake_flags: wake modifier flags (WF_*)
2810 *
2811 * Put it on the run-queue if it's not already there. The "current"
2812 * thread is always on the run-queue (except when the actual
2813 * re-schedule is in progress), and as such you're allowed to do
2814 * the simpler "current->state = TASK_RUNNING" to mark yourself
2815 * runnable without the overhead of this.
2816 *
2817 * Returns %true if @p was woken up, %false if it was already running
2818 * or @state didn't match @p's state.
2819 */
2820static int
2821try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2822{
2823 unsigned long flags;
2824 int cpu, success = 0;
2825
2826 smp_wmb();
2827 raw_spin_lock_irqsave(&p->pi_lock, flags);
2828 if (!(p->state & state))
2829 goto out;
2830
2831 success = 1; /* we're going to change ->state */
2832 cpu = task_cpu(p);
2833
2834 if (p->on_rq && ttwu_remote(p, wake_flags))
2835 goto stat;
2836
2837#ifdef CONFIG_SMP
2838 /*
2839 * If the owning (remote) cpu is still in the middle of schedule() with
2840 * this task as prev, wait until its done referencing the task.
2841 */
2842 while (p->on_cpu) {
2843#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2844 /*
2845 * In case the architecture enables interrupts in
2846 * context_switch(), we cannot busy wait, since that
2847 * would lead to deadlocks when an interrupt hits and
2848 * tries to wake up @prev. So bail and do a complete
2849 * remote wakeup.
2850 */
2851 if (ttwu_activate_remote(p, wake_flags))
2852 goto stat;
2853#else
2854 cpu_relax();
2855#endif
2856 }
2857 /*
2858 * Pairs with the smp_wmb() in finish_lock_switch().
2859 */
2860 smp_rmb();
2861
2862 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2863 p->state = TASK_WAKING;
2864
2865 if (p->sched_class->task_waking)
2866 p->sched_class->task_waking(p);
2867
2868 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2869 if (task_cpu(p) != cpu) {
2870 wake_flags |= WF_MIGRATED;
2871 set_task_cpu(p, cpu);
2872 }
2873#endif /* CONFIG_SMP */
2874
2875 ttwu_queue(p, cpu);
2876stat:
2877 ttwu_stat(p, cpu, wake_flags);
2878out:
2879 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2880
2881 return success;
2882}
2883
2884/**
2885 * try_to_wake_up_local - try to wake up a local task with rq lock held
2886 * @p: the thread to be awakened
2887 *
2888 * Put @p on the run-queue if it's not already there. The caller must
2889 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2890 * the current task.
2891 */
2892static void try_to_wake_up_local(struct task_struct *p)
2893{
2894 struct rq *rq = task_rq(p);
2895
2896 BUG_ON(rq != this_rq());
2897 BUG_ON(p == current);
2898 lockdep_assert_held(&rq->lock);
2899
2900 if (!raw_spin_trylock(&p->pi_lock)) {
2901 raw_spin_unlock(&rq->lock);
2902 raw_spin_lock(&p->pi_lock);
2903 raw_spin_lock(&rq->lock);
2904 }
2905
2906 if (!(p->state & TASK_NORMAL))
2907 goto out;
2908
2909 if (!p->on_rq)
2910 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2911
2912 ttwu_do_wakeup(rq, p, 0);
2913 ttwu_stat(p, smp_processor_id(), 0);
2914out:
2915 raw_spin_unlock(&p->pi_lock);
2916}
2917
2918/**
2919 * wake_up_process - Wake up a specific process
2920 * @p: The process to be woken up.
2921 *
2922 * Attempt to wake up the nominated process and move it to the set of runnable
2923 * processes. Returns 1 if the process was woken up, 0 if it was already
2924 * running.
2925 *
2926 * It may be assumed that this function implies a write memory barrier before
2927 * changing the task state if and only if any tasks are woken up.
2928 */
2929int wake_up_process(struct task_struct *p)
2930{
2931 return try_to_wake_up(p, TASK_ALL, 0);
2932}
2933EXPORT_SYMBOL(wake_up_process);
2934
2935int wake_up_state(struct task_struct *p, unsigned int state)
2936{
2937 return try_to_wake_up(p, state, 0);
2938}
2939
2940/*
2941 * Perform scheduler related setup for a newly forked process p.
2942 * p is forked by current.
2943 *
2944 * __sched_fork() is basic setup used by init_idle() too:
2945 */
2946static void __sched_fork(struct task_struct *p)
2947{
2948 p->on_rq = 0;
2949
2950 p->se.on_rq = 0;
2951 p->se.exec_start = 0;
2952 p->se.sum_exec_runtime = 0;
2953 p->se.prev_sum_exec_runtime = 0;
2954 p->se.nr_migrations = 0;
2955 p->se.vruntime = 0;
2956 INIT_LIST_HEAD(&p->se.group_node);
2957
2958#ifdef CONFIG_SCHEDSTATS
2959 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2960#endif
2961
2962 INIT_LIST_HEAD(&p->rt.run_list);
2963
2964#ifdef CONFIG_PREEMPT_NOTIFIERS
2965 INIT_HLIST_HEAD(&p->preempt_notifiers);
2966#endif
2967}
2968
2969/*
2970 * fork()/clone()-time setup:
2971 */
2972void sched_fork(struct task_struct *p)
2973{
2974 unsigned long flags;
2975 int cpu = get_cpu();
2976
2977 __sched_fork(p);
2978 /*
2979 * We mark the process as running here. This guarantees that
2980 * nobody will actually run it, and a signal or other external
2981 * event cannot wake it up and insert it on the runqueue either.
2982 */
2983 p->state = TASK_RUNNING;
2984
2985 /*
2986 * Make sure we do not leak PI boosting priority to the child.
2987 */
2988 p->prio = current->normal_prio;
2989
2990 /*
2991 * Revert to default priority/policy on fork if requested.
2992 */
2993 if (unlikely(p->sched_reset_on_fork)) {
2994 if (task_has_rt_policy(p)) {
2995 p->policy = SCHED_NORMAL;
2996 p->static_prio = NICE_TO_PRIO(0);
2997 p->rt_priority = 0;
2998 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2999 p->static_prio = NICE_TO_PRIO(0);
3000
3001 p->prio = p->normal_prio = __normal_prio(p);
3002 set_load_weight(p);
3003
3004 /*
3005 * We don't need the reset flag anymore after the fork. It has
3006 * fulfilled its duty:
3007 */
3008 p->sched_reset_on_fork = 0;
3009 }
3010
3011 if (!rt_prio(p->prio))
3012 p->sched_class = &fair_sched_class;
3013
3014 if (p->sched_class->task_fork)
3015 p->sched_class->task_fork(p);
3016
3017 /*
3018 * The child is not yet in the pid-hash so no cgroup attach races,
3019 * and the cgroup is pinned to this child due to cgroup_fork()
3020 * is ran before sched_fork().
3021 *
3022 * Silence PROVE_RCU.
3023 */
3024 raw_spin_lock_irqsave(&p->pi_lock, flags);
3025 set_task_cpu(p, cpu);
3026 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3027
3028#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
3029 if (likely(sched_info_on()))
3030 memset(&p->sched_info, 0, sizeof(p->sched_info));
3031#endif
3032#if defined(CONFIG_SMP)
3033 p->on_cpu = 0;
3034#endif
3035#ifdef CONFIG_PREEMPT_COUNT
3036 /* Want to start with kernel preemption disabled. */
3037 task_thread_info(p)->preempt_count = 1;
3038#endif
3039#ifdef CONFIG_SMP
3040 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3041#endif
3042
3043 put_cpu();
3044}
3045
3046/*
3047 * wake_up_new_task - wake up a newly created task for the first time.
3048 *
3049 * This function will do some initial scheduler statistics housekeeping
3050 * that must be done for every newly created context, then puts the task
3051 * on the runqueue and wakes it.
3052 */
3053void wake_up_new_task(struct task_struct *p)
3054{
3055 unsigned long flags;
3056 struct rq *rq;
3057
3058 raw_spin_lock_irqsave(&p->pi_lock, flags);
3059#ifdef CONFIG_SMP
3060 /*
3061 * Fork balancing, do it here and not earlier because:
3062 * - cpus_allowed can change in the fork path
3063 * - any previously selected cpu might disappear through hotplug
3064 */
3065 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
3066#endif
3067
3068 rq = __task_rq_lock(p);
3069 activate_task(rq, p, 0);
3070 p->on_rq = 1;
3071 trace_sched_wakeup_new(p, true);
3072 check_preempt_curr(rq, p, WF_FORK);
3073#ifdef CONFIG_SMP
3074 if (p->sched_class->task_woken)
3075 p->sched_class->task_woken(rq, p);
3076#endif
3077 task_rq_unlock(rq, p, &flags);
3078}
3079
3080#ifdef CONFIG_PREEMPT_NOTIFIERS
3081
3082/**
3083 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3084 * @notifier: notifier struct to register
3085 */
3086void preempt_notifier_register(struct preempt_notifier *notifier)
3087{
3088 hlist_add_head(&notifier->link, &current->preempt_notifiers);
3089}
3090EXPORT_SYMBOL_GPL(preempt_notifier_register);
3091
3092/**
3093 * preempt_notifier_unregister - no longer interested in preemption notifications
3094 * @notifier: notifier struct to unregister
3095 *
3096 * This is safe to call from within a preemption notifier.
3097 */
3098void preempt_notifier_unregister(struct preempt_notifier *notifier)
3099{
3100 hlist_del(&notifier->link);
3101}
3102EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3103
3104static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3105{
3106 struct preempt_notifier *notifier;
3107 struct hlist_node *node;
3108
3109 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3110 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3111}
3112
3113static void
3114fire_sched_out_preempt_notifiers(struct task_struct *curr,
3115 struct task_struct *next)
3116{
3117 struct preempt_notifier *notifier;
3118 struct hlist_node *node;
3119
3120 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3121 notifier->ops->sched_out(notifier, next);
3122}
3123
3124#else /* !CONFIG_PREEMPT_NOTIFIERS */
3125
3126static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3127{
3128}
3129
3130static void
3131fire_sched_out_preempt_notifiers(struct task_struct *curr,
3132 struct task_struct *next)
3133{
3134}
3135
3136#endif /* CONFIG_PREEMPT_NOTIFIERS */
3137
3138/**
3139 * prepare_task_switch - prepare to switch tasks
3140 * @rq: the runqueue preparing to switch
3141 * @prev: the current task that is being switched out
3142 * @next: the task we are going to switch to.
3143 *
3144 * This is called with the rq lock held and interrupts off. It must
3145 * be paired with a subsequent finish_task_switch after the context
3146 * switch.
3147 *
3148 * prepare_task_switch sets up locking and calls architecture specific
3149 * hooks.
3150 */
3151static inline void
3152prepare_task_switch(struct rq *rq, struct task_struct *prev,
3153 struct task_struct *next)
3154{
3155 sched_info_switch(prev, next);
3156 perf_event_task_sched_out(prev, next);
3157 fire_sched_out_preempt_notifiers(prev, next);
3158 prepare_lock_switch(rq, next);
3159 prepare_arch_switch(next);
3160 trace_sched_switch(prev, next);
3161}
3162
3163/**
3164 * finish_task_switch - clean up after a task-switch
3165 * @rq: runqueue associated with task-switch
3166 * @prev: the thread we just switched away from.
3167 *
3168 * finish_task_switch must be called after the context switch, paired
3169 * with a prepare_task_switch call before the context switch.
3170 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3171 * and do any other architecture-specific cleanup actions.
3172 *
3173 * Note that we may have delayed dropping an mm in context_switch(). If
3174 * so, we finish that here outside of the runqueue lock. (Doing it
3175 * with the lock held can cause deadlocks; see schedule() for
3176 * details.)
3177 */
3178static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3179 __releases(rq->lock)
3180{
3181 struct mm_struct *mm = rq->prev_mm;
3182 long prev_state;
3183
3184 rq->prev_mm = NULL;
3185
3186 /*
3187 * A task struct has one reference for the use as "current".
3188 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3189 * schedule one last time. The schedule call will never return, and
3190 * the scheduled task must drop that reference.
3191 * The test for TASK_DEAD must occur while the runqueue locks are
3192 * still held, otherwise prev could be scheduled on another cpu, die
3193 * there before we look at prev->state, and then the reference would
3194 * be dropped twice.
3195 * Manfred Spraul <[email protected]>
3196 */
3197 prev_state = prev->state;
3198 finish_arch_switch(prev);
3199#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3200 local_irq_disable();
3201#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3202 perf_event_task_sched_in(prev, current);
3203#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3204 local_irq_enable();
3205#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3206 finish_lock_switch(rq, prev);
3207
3208 fire_sched_in_preempt_notifiers(current);
3209 if (mm)
3210 mmdrop(mm);
3211 if (unlikely(prev_state == TASK_DEAD)) {
3212 /*
3213 * Remove function-return probe instances associated with this
3214 * task and put them back on the free list.
3215 */
3216 kprobe_flush_task(prev);
3217 put_task_struct(prev);
3218 }
3219}
3220
3221#ifdef CONFIG_SMP
3222
3223/* assumes rq->lock is held */
3224static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3225{
3226 if (prev->sched_class->pre_schedule)
3227 prev->sched_class->pre_schedule(rq, prev);
3228}
3229
3230/* rq->lock is NOT held, but preemption is disabled */
3231static inline void post_schedule(struct rq *rq)
3232{
3233 if (rq->post_schedule) {
3234 unsigned long flags;
3235
3236 raw_spin_lock_irqsave(&rq->lock, flags);
3237 if (rq->curr->sched_class->post_schedule)
3238 rq->curr->sched_class->post_schedule(rq);
3239 raw_spin_unlock_irqrestore(&rq->lock, flags);
3240
3241 rq->post_schedule = 0;
3242 }
3243}
3244
3245#else
3246
3247static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3248{
3249}
3250
3251static inline void post_schedule(struct rq *rq)
3252{
3253}
3254
3255#endif
3256
3257/**
3258 * schedule_tail - first thing a freshly forked thread must call.
3259 * @prev: the thread we just switched away from.
3260 */
3261asmlinkage void schedule_tail(struct task_struct *prev)
3262 __releases(rq->lock)
3263{
3264 struct rq *rq = this_rq();
3265
3266 finish_task_switch(rq, prev);
3267
3268 /*
3269 * FIXME: do we need to worry about rq being invalidated by the
3270 * task_switch?
3271 */
3272 post_schedule(rq);
3273
3274#ifdef __ARCH_WANT_UNLOCKED_CTXSW
3275 /* In this case, finish_task_switch does not reenable preemption */
3276 preempt_enable();
3277#endif
3278 if (current->set_child_tid)
3279 put_user(task_pid_vnr(current), current->set_child_tid);
3280}
3281
3282/*
3283 * context_switch - switch to the new MM and the new
3284 * thread's register state.
3285 */
3286static inline void
3287context_switch(struct rq *rq, struct task_struct *prev,
3288 struct task_struct *next)
3289{
3290 struct mm_struct *mm, *oldmm;
3291
3292 prepare_task_switch(rq, prev, next);
3293
3294 mm = next->mm;
3295 oldmm = prev->active_mm;
3296 /*
3297 * For paravirt, this is coupled with an exit in switch_to to
3298 * combine the page table reload and the switch backend into
3299 * one hypercall.
3300 */
3301 arch_start_context_switch(prev);
3302
3303 if (!mm) {
3304 next->active_mm = oldmm;
3305 atomic_inc(&oldmm->mm_count);
3306 enter_lazy_tlb(oldmm, next);
3307 } else
3308 switch_mm(oldmm, mm, next);
3309
3310 if (!prev->mm) {
3311 prev->active_mm = NULL;
3312 rq->prev_mm = oldmm;
3313 }
3314 /*
3315 * Since the runqueue lock will be released by the next
3316 * task (which is an invalid locking op but in the case
3317 * of the scheduler it's an obvious special-case), so we
3318 * do an early lockdep release here:
3319 */
3320#ifndef __ARCH_WANT_UNLOCKED_CTXSW
3321 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3322#endif
3323
3324 /* Here we just switch the register state and the stack. */
3325 switch_to(prev, next, prev);
3326
3327 barrier();
3328 /*
3329 * this_rq must be evaluated again because prev may have moved
3330 * CPUs since it called schedule(), thus the 'rq' on its stack
3331 * frame will be invalid.
3332 */
3333 finish_task_switch(this_rq(), prev);
3334}
3335
3336/*
3337 * nr_running, nr_uninterruptible and nr_context_switches:
3338 *
3339 * externally visible scheduler statistics: current number of runnable
3340 * threads, current number of uninterruptible-sleeping threads, total
3341 * number of context switches performed since bootup.
3342 */
3343unsigned long nr_running(void)
3344{
3345 unsigned long i, sum = 0;
3346
3347 for_each_online_cpu(i)
3348 sum += cpu_rq(i)->nr_running;
3349
3350 return sum;
3351}
3352
3353unsigned long nr_uninterruptible(void)
3354{
3355 unsigned long i, sum = 0;
3356
3357 for_each_possible_cpu(i)
3358 sum += cpu_rq(i)->nr_uninterruptible;
3359
3360 /*
3361 * Since we read the counters lockless, it might be slightly
3362 * inaccurate. Do not allow it to go below zero though:
3363 */
3364 if (unlikely((long)sum < 0))
3365 sum = 0;
3366
3367 return sum;
3368}
3369
3370unsigned long long nr_context_switches(void)
3371{
3372 int i;
3373 unsigned long long sum = 0;
3374
3375 for_each_possible_cpu(i)
3376 sum += cpu_rq(i)->nr_switches;
3377
3378 return sum;
3379}
3380
3381unsigned long nr_iowait(void)
3382{
3383 unsigned long i, sum = 0;
3384
3385 for_each_possible_cpu(i)
3386 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3387
3388 return sum;
3389}
3390
3391unsigned long nr_iowait_cpu(int cpu)
3392{
3393 struct rq *this = cpu_rq(cpu);
3394 return atomic_read(&this->nr_iowait);
3395}
3396
3397unsigned long this_cpu_load(void)
3398{
3399 struct rq *this = this_rq();
3400 return this->cpu_load[0];
3401}
3402
3403
3404/* Variables and functions for calc_load */
3405static atomic_long_t calc_load_tasks;
3406static unsigned long calc_load_update;
3407unsigned long avenrun[3];
3408EXPORT_SYMBOL(avenrun);
3409
3410static long calc_load_fold_active(struct rq *this_rq)
3411{
3412 long nr_active, delta = 0;
3413
3414 nr_active = this_rq->nr_running;
3415 nr_active += (long) this_rq->nr_uninterruptible;
3416
3417 if (nr_active != this_rq->calc_load_active) {
3418 delta = nr_active - this_rq->calc_load_active;
3419 this_rq->calc_load_active = nr_active;
3420 }
3421
3422 return delta;
3423}
3424
3425static unsigned long
3426calc_load(unsigned long load, unsigned long exp, unsigned long active)
3427{
3428 load *= exp;
3429 load += active * (FIXED_1 - exp);
3430 load += 1UL << (FSHIFT - 1);
3431 return load >> FSHIFT;
3432}
3433
3434#ifdef CONFIG_NO_HZ
3435/*
3436 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3437 *
3438 * When making the ILB scale, we should try to pull this in as well.
3439 */
3440static atomic_long_t calc_load_tasks_idle;
3441
3442static void calc_load_account_idle(struct rq *this_rq)
3443{
3444 long delta;
3445
3446 delta = calc_load_fold_active(this_rq);
3447 if (delta)
3448 atomic_long_add(delta, &calc_load_tasks_idle);
3449}
3450
3451static long calc_load_fold_idle(void)
3452{
3453 long delta = 0;
3454
3455 /*
3456 * Its got a race, we don't care...
3457 */
3458 if (atomic_long_read(&calc_load_tasks_idle))
3459 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3460
3461 return delta;
3462}
3463
3464/**
3465 * fixed_power_int - compute: x^n, in O(log n) time
3466 *
3467 * @x: base of the power
3468 * @frac_bits: fractional bits of @x
3469 * @n: power to raise @x to.
3470 *
3471 * By exploiting the relation between the definition of the natural power
3472 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3473 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3474 * (where: n_i \elem {0, 1}, the binary vector representing n),
3475 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3476 * of course trivially computable in O(log_2 n), the length of our binary
3477 * vector.
3478 */
3479static unsigned long
3480fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3481{
3482 unsigned long result = 1UL << frac_bits;
3483
3484 if (n) for (;;) {
3485 if (n & 1) {
3486 result *= x;
3487 result += 1UL << (frac_bits - 1);
3488 result >>= frac_bits;
3489 }
3490 n >>= 1;
3491 if (!n)
3492 break;
3493 x *= x;
3494 x += 1UL << (frac_bits - 1);
3495 x >>= frac_bits;
3496 }
3497
3498 return result;
3499}
3500
3501/*
3502 * a1 = a0 * e + a * (1 - e)
3503 *
3504 * a2 = a1 * e + a * (1 - e)
3505 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3506 * = a0 * e^2 + a * (1 - e) * (1 + e)
3507 *
3508 * a3 = a2 * e + a * (1 - e)
3509 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3510 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3511 *
3512 * ...
3513 *
3514 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3515 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3516 * = a0 * e^n + a * (1 - e^n)
3517 *
3518 * [1] application of the geometric series:
3519 *
3520 * n 1 - x^(n+1)
3521 * S_n := \Sum x^i = -------------
3522 * i=0 1 - x
3523 */
3524static unsigned long
3525calc_load_n(unsigned long load, unsigned long exp,
3526 unsigned long active, unsigned int n)
3527{
3528
3529 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3530}
3531
3532/*
3533 * NO_HZ can leave us missing all per-cpu ticks calling
3534 * calc_load_account_active(), but since an idle CPU folds its delta into
3535 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3536 * in the pending idle delta if our idle period crossed a load cycle boundary.
3537 *
3538 * Once we've updated the global active value, we need to apply the exponential
3539 * weights adjusted to the number of cycles missed.
3540 */
3541static void calc_global_nohz(unsigned long ticks)
3542{
3543 long delta, active, n;
3544
3545 if (time_before(jiffies, calc_load_update))
3546 return;
3547
3548 /*
3549 * If we crossed a calc_load_update boundary, make sure to fold
3550 * any pending idle changes, the respective CPUs might have
3551 * missed the tick driven calc_load_account_active() update
3552 * due to NO_HZ.
3553 */
3554 delta = calc_load_fold_idle();
3555 if (delta)
3556 atomic_long_add(delta, &calc_load_tasks);
3557
3558 /*
3559 * If we were idle for multiple load cycles, apply them.
3560 */
3561 if (ticks >= LOAD_FREQ) {
3562 n = ticks / LOAD_FREQ;
3563
3564 active = atomic_long_read(&calc_load_tasks);
3565 active = active > 0 ? active * FIXED_1 : 0;
3566
3567 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3568 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3569 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3570
3571 calc_load_update += n * LOAD_FREQ;
3572 }
3573
3574 /*
3575 * Its possible the remainder of the above division also crosses
3576 * a LOAD_FREQ period, the regular check in calc_global_load()
3577 * which comes after this will take care of that.
3578 *
3579 * Consider us being 11 ticks before a cycle completion, and us
3580 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3581 * age us 4 cycles, and the test in calc_global_load() will
3582 * pick up the final one.
3583 */
3584}
3585#else
3586static void calc_load_account_idle(struct rq *this_rq)
3587{
3588}
3589
3590static inline long calc_load_fold_idle(void)
3591{
3592 return 0;
3593}
3594
3595static void calc_global_nohz(unsigned long ticks)
3596{
3597}
3598#endif
3599
3600/**
3601 * get_avenrun - get the load average array
3602 * @loads: pointer to dest load array
3603 * @offset: offset to add
3604 * @shift: shift count to shift the result left
3605 *
3606 * These values are estimates at best, so no need for locking.
3607 */
3608void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3609{
3610 loads[0] = (avenrun[0] + offset) << shift;
3611 loads[1] = (avenrun[1] + offset) << shift;
3612 loads[2] = (avenrun[2] + offset) << shift;
3613}
3614
3615/*
3616 * calc_load - update the avenrun load estimates 10 ticks after the
3617 * CPUs have updated calc_load_tasks.
3618 */
3619void calc_global_load(unsigned long ticks)
3620{
3621 long active;
3622
3623 calc_global_nohz(ticks);
3624
3625 if (time_before(jiffies, calc_load_update + 10))
3626 return;
3627
3628 active = atomic_long_read(&calc_load_tasks);
3629 active = active > 0 ? active * FIXED_1 : 0;
3630
3631 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3632 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3633 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3634
3635 calc_load_update += LOAD_FREQ;
3636}
3637
3638/*
3639 * Called from update_cpu_load() to periodically update this CPU's
3640 * active count.
3641 */
3642static void calc_load_account_active(struct rq *this_rq)
3643{
3644 long delta;
3645
3646 if (time_before(jiffies, this_rq->calc_load_update))
3647 return;
3648
3649 delta = calc_load_fold_active(this_rq);
3650 delta += calc_load_fold_idle();
3651 if (delta)
3652 atomic_long_add(delta, &calc_load_tasks);
3653
3654 this_rq->calc_load_update += LOAD_FREQ;
3655}
3656
3657/*
3658 * The exact cpuload at various idx values, calculated at every tick would be
3659 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3660 *
3661 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3662 * on nth tick when cpu may be busy, then we have:
3663 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3664 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3665 *
3666 * decay_load_missed() below does efficient calculation of
3667 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3668 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3669 *
3670 * The calculation is approximated on a 128 point scale.
3671 * degrade_zero_ticks is the number of ticks after which load at any
3672 * particular idx is approximated to be zero.
3673 * degrade_factor is a precomputed table, a row for each load idx.
3674 * Each column corresponds to degradation factor for a power of two ticks,
3675 * based on 128 point scale.
3676 * Example:
3677 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3678 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3679 *
3680 * With this power of 2 load factors, we can degrade the load n times
3681 * by looking at 1 bits in n and doing as many mult/shift instead of
3682 * n mult/shifts needed by the exact degradation.
3683 */
3684#define DEGRADE_SHIFT 7
3685static const unsigned char
3686 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3687static const unsigned char
3688 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3689 {0, 0, 0, 0, 0, 0, 0, 0},
3690 {64, 32, 8, 0, 0, 0, 0, 0},
3691 {96, 72, 40, 12, 1, 0, 0},
3692 {112, 98, 75, 43, 15, 1, 0},
3693 {120, 112, 98, 76, 45, 16, 2} };
3694
3695/*
3696 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3697 * would be when CPU is idle and so we just decay the old load without
3698 * adding any new load.
3699 */
3700static unsigned long
3701decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3702{
3703 int j = 0;
3704
3705 if (!missed_updates)
3706 return load;
3707
3708 if (missed_updates >= degrade_zero_ticks[idx])
3709 return 0;
3710
3711 if (idx == 1)
3712 return load >> missed_updates;
3713
3714 while (missed_updates) {
3715 if (missed_updates % 2)
3716 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3717
3718 missed_updates >>= 1;
3719 j++;
3720 }
3721 return load;
3722}
3723
3724/*
3725 * Update rq->cpu_load[] statistics. This function is usually called every
3726 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3727 * every tick. We fix it up based on jiffies.
3728 */
3729static void update_cpu_load(struct rq *this_rq)
3730{
3731 unsigned long this_load = this_rq->load.weight;
3732 unsigned long curr_jiffies = jiffies;
3733 unsigned long pending_updates;
3734 int i, scale;
3735
3736 this_rq->nr_load_updates++;
3737
3738 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3739 if (curr_jiffies == this_rq->last_load_update_tick)
3740 return;
3741
3742 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3743 this_rq->last_load_update_tick = curr_jiffies;
3744
3745 /* Update our load: */
3746 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3747 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3748 unsigned long old_load, new_load;
3749
3750 /* scale is effectively 1 << i now, and >> i divides by scale */
3751
3752 old_load = this_rq->cpu_load[i];
3753 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3754 new_load = this_load;
3755 /*
3756 * Round up the averaging division if load is increasing. This
3757 * prevents us from getting stuck on 9 if the load is 10, for
3758 * example.
3759 */
3760 if (new_load > old_load)
3761 new_load += scale - 1;
3762
3763 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3764 }
3765
3766 sched_avg_update(this_rq);
3767}
3768
3769static void update_cpu_load_active(struct rq *this_rq)
3770{
3771 update_cpu_load(this_rq);
3772
3773 calc_load_account_active(this_rq);
3774}
3775
3776#ifdef CONFIG_SMP
3777
3778/*
3779 * sched_exec - execve() is a valuable balancing opportunity, because at
3780 * this point the task has the smallest effective memory and cache footprint.
3781 */
3782void sched_exec(void)
3783{
3784 struct task_struct *p = current;
3785 unsigned long flags;
3786 int dest_cpu;
3787
3788 raw_spin_lock_irqsave(&p->pi_lock, flags);
3789 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3790 if (dest_cpu == smp_processor_id())
3791 goto unlock;
3792
3793 if (likely(cpu_active(dest_cpu))) {
3794 struct migration_arg arg = { p, dest_cpu };
3795
3796 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3797 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3798 return;
3799 }
3800unlock:
3801 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3802}
3803
3804#endif
3805
3806DEFINE_PER_CPU(struct kernel_stat, kstat);
3807
3808EXPORT_PER_CPU_SYMBOL(kstat);
3809
3810/*
3811 * Return any ns on the sched_clock that have not yet been accounted in
3812 * @p in case that task is currently running.
3813 *
3814 * Called with task_rq_lock() held on @rq.
3815 */
3816static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3817{
3818 u64 ns = 0;
3819
3820 if (task_current(rq, p)) {
3821 update_rq_clock(rq);
3822 ns = rq->clock_task - p->se.exec_start;
3823 if ((s64)ns < 0)
3824 ns = 0;
3825 }
3826
3827 return ns;
3828}
3829
3830unsigned long long task_delta_exec(struct task_struct *p)
3831{
3832 unsigned long flags;
3833 struct rq *rq;
3834 u64 ns = 0;
3835
3836 rq = task_rq_lock(p, &flags);
3837 ns = do_task_delta_exec(p, rq);
3838 task_rq_unlock(rq, p, &flags);
3839
3840 return ns;
3841}
3842
3843/*
3844 * Return accounted runtime for the task.
3845 * In case the task is currently running, return the runtime plus current's
3846 * pending runtime that have not been accounted yet.
3847 */
3848unsigned long long task_sched_runtime(struct task_struct *p)
3849{
3850 unsigned long flags;
3851 struct rq *rq;
3852 u64 ns = 0;
3853
3854 rq = task_rq_lock(p, &flags);
3855 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3856 task_rq_unlock(rq, p, &flags);
3857
3858 return ns;
3859}
3860
3861/*
3862 * Account user cpu time to a process.
3863 * @p: the process that the cpu time gets accounted to
3864 * @cputime: the cpu time spent in user space since the last update
3865 * @cputime_scaled: cputime scaled by cpu frequency
3866 */
3867void account_user_time(struct task_struct *p, cputime_t cputime,
3868 cputime_t cputime_scaled)
3869{
3870 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3871 cputime64_t tmp;
3872
3873 /* Add user time to process. */
3874 p->utime = cputime_add(p->utime, cputime);
3875 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3876 account_group_user_time(p, cputime);
3877
3878 /* Add user time to cpustat. */
3879 tmp = cputime_to_cputime64(cputime);
3880 if (TASK_NICE(p) > 0)
3881 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3882 else
3883 cpustat->user = cputime64_add(cpustat->user, tmp);
3884
3885 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3886 /* Account for user time used */
3887 acct_update_integrals(p);
3888}
3889
3890/*
3891 * Account guest cpu time to a process.
3892 * @p: the process that the cpu time gets accounted to
3893 * @cputime: the cpu time spent in virtual machine since the last update
3894 * @cputime_scaled: cputime scaled by cpu frequency
3895 */
3896static void account_guest_time(struct task_struct *p, cputime_t cputime,
3897 cputime_t cputime_scaled)
3898{
3899 cputime64_t tmp;
3900 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3901
3902 tmp = cputime_to_cputime64(cputime);
3903
3904 /* Add guest time to process. */
3905 p->utime = cputime_add(p->utime, cputime);
3906 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3907 account_group_user_time(p, cputime);
3908 p->gtime = cputime_add(p->gtime, cputime);
3909
3910 /* Add guest time to cpustat. */
3911 if (TASK_NICE(p) > 0) {
3912 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3913 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3914 } else {
3915 cpustat->user = cputime64_add(cpustat->user, tmp);
3916 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3917 }
3918}
3919
3920/*
3921 * Account system cpu time to a process and desired cpustat field
3922 * @p: the process that the cpu time gets accounted to
3923 * @cputime: the cpu time spent in kernel space since the last update
3924 * @cputime_scaled: cputime scaled by cpu frequency
3925 * @target_cputime64: pointer to cpustat field that has to be updated
3926 */
3927static inline
3928void __account_system_time(struct task_struct *p, cputime_t cputime,
3929 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3930{
3931 cputime64_t tmp = cputime_to_cputime64(cputime);
3932
3933 /* Add system time to process. */
3934 p->stime = cputime_add(p->stime, cputime);
3935 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3936 account_group_system_time(p, cputime);
3937
3938 /* Add system time to cpustat. */
3939 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3940 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3941
3942 /* Account for system time used */
3943 acct_update_integrals(p);
3944}
3945
3946/*
3947 * Account system cpu time to a process.
3948 * @p: the process that the cpu time gets accounted to
3949 * @hardirq_offset: the offset to subtract from hardirq_count()
3950 * @cputime: the cpu time spent in kernel space since the last update
3951 * @cputime_scaled: cputime scaled by cpu frequency
3952 */
3953void account_system_time(struct task_struct *p, int hardirq_offset,
3954 cputime_t cputime, cputime_t cputime_scaled)
3955{
3956 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3957 cputime64_t *target_cputime64;
3958
3959 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3960 account_guest_time(p, cputime, cputime_scaled);
3961 return;
3962 }
3963
3964 if (hardirq_count() - hardirq_offset)
3965 target_cputime64 = &cpustat->irq;
3966 else if (in_serving_softirq())
3967 target_cputime64 = &cpustat->softirq;
3968 else
3969 target_cputime64 = &cpustat->system;
3970
3971 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3972}
3973
3974/*
3975 * Account for involuntary wait time.
3976 * @cputime: the cpu time spent in involuntary wait
3977 */
3978void account_steal_time(cputime_t cputime)
3979{
3980 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3981 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3982
3983 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3984}
3985
3986/*
3987 * Account for idle time.
3988 * @cputime: the cpu time spent in idle wait
3989 */
3990void account_idle_time(cputime_t cputime)
3991{
3992 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3993 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3994 struct rq *rq = this_rq();
3995
3996 if (atomic_read(&rq->nr_iowait) > 0)
3997 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3998 else
3999 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4000}
4001
4002static __always_inline bool steal_account_process_tick(void)
4003{
4004#ifdef CONFIG_PARAVIRT
4005 if (static_branch(&paravirt_steal_enabled)) {
4006 u64 steal, st = 0;
4007
4008 steal = paravirt_steal_clock(smp_processor_id());
4009 steal -= this_rq()->prev_steal_time;
4010
4011 st = steal_ticks(steal);
4012 this_rq()->prev_steal_time += st * TICK_NSEC;
4013
4014 account_steal_time(st);
4015 return st;
4016 }
4017#endif
4018 return false;
4019}
4020
4021#ifndef CONFIG_VIRT_CPU_ACCOUNTING
4022
4023#ifdef CONFIG_IRQ_TIME_ACCOUNTING
4024/*
4025 * Account a tick to a process and cpustat
4026 * @p: the process that the cpu time gets accounted to
4027 * @user_tick: is the tick from userspace
4028 * @rq: the pointer to rq
4029 *
4030 * Tick demultiplexing follows the order
4031 * - pending hardirq update
4032 * - pending softirq update
4033 * - user_time
4034 * - idle_time
4035 * - system time
4036 * - check for guest_time
4037 * - else account as system_time
4038 *
4039 * Check for hardirq is done both for system and user time as there is
4040 * no timer going off while we are on hardirq and hence we may never get an
4041 * opportunity to update it solely in system time.
4042 * p->stime and friends are only updated on system time and not on irq
4043 * softirq as those do not count in task exec_runtime any more.
4044 */
4045static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4046 struct rq *rq)
4047{
4048 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4049 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
4050 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4051
4052 if (steal_account_process_tick())
4053 return;
4054
4055 if (irqtime_account_hi_update()) {
4056 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4057 } else if (irqtime_account_si_update()) {
4058 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4059 } else if (this_cpu_ksoftirqd() == p) {
4060 /*
4061 * ksoftirqd time do not get accounted in cpu_softirq_time.
4062 * So, we have to handle it separately here.
4063 * Also, p->stime needs to be updated for ksoftirqd.
4064 */
4065 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4066 &cpustat->softirq);
4067 } else if (user_tick) {
4068 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4069 } else if (p == rq->idle) {
4070 account_idle_time(cputime_one_jiffy);
4071 } else if (p->flags & PF_VCPU) { /* System time or guest time */
4072 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
4073 } else {
4074 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4075 &cpustat->system);
4076 }
4077}
4078
4079static void irqtime_account_idle_ticks(int ticks)
4080{
4081 int i;
4082 struct rq *rq = this_rq();
4083
4084 for (i = 0; i < ticks; i++)
4085 irqtime_account_process_tick(current, 0, rq);
4086}
4087#else /* CONFIG_IRQ_TIME_ACCOUNTING */
4088static void irqtime_account_idle_ticks(int ticks) {}
4089static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4090 struct rq *rq) {}
4091#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
4092
4093/*
4094 * Account a single tick of cpu time.
4095 * @p: the process that the cpu time gets accounted to
4096 * @user_tick: indicates if the tick is a user or a system tick
4097 */
4098void account_process_tick(struct task_struct *p, int user_tick)
4099{
4100 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4101 struct rq *rq = this_rq();
4102
4103 if (sched_clock_irqtime) {
4104 irqtime_account_process_tick(p, user_tick, rq);
4105 return;
4106 }
4107
4108 if (steal_account_process_tick())
4109 return;
4110
4111 if (user_tick)
4112 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4113 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
4114 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
4115 one_jiffy_scaled);
4116 else
4117 account_idle_time(cputime_one_jiffy);
4118}
4119
4120/*
4121 * Account multiple ticks of steal time.
4122 * @p: the process from which the cpu time has been stolen
4123 * @ticks: number of stolen ticks
4124 */
4125void account_steal_ticks(unsigned long ticks)
4126{
4127 account_steal_time(jiffies_to_cputime(ticks));
4128}
4129
4130/*
4131 * Account multiple ticks of idle time.
4132 * @ticks: number of stolen ticks
4133 */
4134void account_idle_ticks(unsigned long ticks)
4135{
4136
4137 if (sched_clock_irqtime) {
4138 irqtime_account_idle_ticks(ticks);
4139 return;
4140 }
4141
4142 account_idle_time(jiffies_to_cputime(ticks));
4143}
4144
4145#endif
4146
4147/*
4148 * Use precise platform statistics if available:
4149 */
4150#ifdef CONFIG_VIRT_CPU_ACCOUNTING
4151void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4152{
4153 *ut = p->utime;
4154 *st = p->stime;
4155}
4156
4157void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4158{
4159 struct task_cputime cputime;
4160
4161 thread_group_cputime(p, &cputime);
4162
4163 *ut = cputime.utime;
4164 *st = cputime.stime;
4165}
4166#else
4167
4168#ifndef nsecs_to_cputime
4169# define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4170#endif
4171
4172void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4173{
4174 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4175
4176 /*
4177 * Use CFS's precise accounting:
4178 */
4179 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4180
4181 if (total) {
4182 u64 temp = rtime;
4183
4184 temp *= utime;
4185 do_div(temp, total);
4186 utime = (cputime_t)temp;
4187 } else
4188 utime = rtime;
4189
4190 /*
4191 * Compare with previous values, to keep monotonicity:
4192 */
4193 p->prev_utime = max(p->prev_utime, utime);
4194 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4195
4196 *ut = p->prev_utime;
4197 *st = p->prev_stime;
4198}
4199
4200/*
4201 * Must be called with siglock held.
4202 */
4203void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4204{
4205 struct signal_struct *sig = p->signal;
4206 struct task_cputime cputime;
4207 cputime_t rtime, utime, total;
4208
4209 thread_group_cputime(p, &cputime);
4210
4211 total = cputime_add(cputime.utime, cputime.stime);
4212 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4213
4214 if (total) {
4215 u64 temp = rtime;
4216
4217 temp *= cputime.utime;
4218 do_div(temp, total);
4219 utime = (cputime_t)temp;
4220 } else
4221 utime = rtime;
4222
4223 sig->prev_utime = max(sig->prev_utime, utime);
4224 sig->prev_stime = max(sig->prev_stime,
4225 cputime_sub(rtime, sig->prev_utime));
4226
4227 *ut = sig->prev_utime;
4228 *st = sig->prev_stime;
4229}
4230#endif
4231
4232/*
4233 * This function gets called by the timer code, with HZ frequency.
4234 * We call it with interrupts disabled.
4235 */
4236void scheduler_tick(void)
4237{
4238 int cpu = smp_processor_id();
4239 struct rq *rq = cpu_rq(cpu);
4240 struct task_struct *curr = rq->curr;
4241
4242 sched_clock_tick();
4243
4244 raw_spin_lock(&rq->lock);
4245 update_rq_clock(rq);
4246 update_cpu_load_active(rq);
4247 curr->sched_class->task_tick(rq, curr, 0);
4248 raw_spin_unlock(&rq->lock);
4249
4250 perf_event_task_tick();
4251
4252#ifdef CONFIG_SMP
4253 rq->idle_balance = idle_cpu(cpu);
4254 trigger_load_balance(rq, cpu);
4255#endif
4256}
4257
4258notrace unsigned long get_parent_ip(unsigned long addr)
4259{
4260 if (in_lock_functions(addr)) {
4261 addr = CALLER_ADDR2;
4262 if (in_lock_functions(addr))
4263 addr = CALLER_ADDR3;
4264 }
4265 return addr;
4266}
4267
4268#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4269 defined(CONFIG_PREEMPT_TRACER))
4270
4271void __kprobes add_preempt_count(int val)
4272{
4273#ifdef CONFIG_DEBUG_PREEMPT
4274 /*
4275 * Underflow?
4276 */
4277 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4278 return;
4279#endif
4280 preempt_count() += val;
4281#ifdef CONFIG_DEBUG_PREEMPT
4282 /*
4283 * Spinlock count overflowing soon?
4284 */
4285 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4286 PREEMPT_MASK - 10);
4287#endif
4288 if (preempt_count() == val)
4289 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4290}
4291EXPORT_SYMBOL(add_preempt_count);
4292
4293void __kprobes sub_preempt_count(int val)
4294{
4295#ifdef CONFIG_DEBUG_PREEMPT
4296 /*
4297 * Underflow?
4298 */
4299 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4300 return;
4301 /*
4302 * Is the spinlock portion underflowing?
4303 */
4304 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4305 !(preempt_count() & PREEMPT_MASK)))
4306 return;
4307#endif
4308
4309 if (preempt_count() == val)
4310 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4311 preempt_count() -= val;
4312}
4313EXPORT_SYMBOL(sub_preempt_count);
4314
4315#endif
4316
4317/*
4318 * Print scheduling while atomic bug:
4319 */
4320static noinline void __schedule_bug(struct task_struct *prev)
4321{
4322 struct pt_regs *regs = get_irq_regs();
4323
4324 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4325 prev->comm, prev->pid, preempt_count());
4326
4327 debug_show_held_locks(prev);
4328 print_modules();
4329 if (irqs_disabled())
4330 print_irqtrace_events(prev);
4331
4332 if (regs)
4333 show_regs(regs);
4334 else
4335 dump_stack();
4336}
4337
4338/*
4339 * Various schedule()-time debugging checks and statistics:
4340 */
4341static inline void schedule_debug(struct task_struct *prev)
4342{
4343 /*
4344 * Test if we are atomic. Since do_exit() needs to call into
4345 * schedule() atomically, we ignore that path for now.
4346 * Otherwise, whine if we are scheduling when we should not be.
4347 */
4348 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4349 __schedule_bug(prev);
4350 rcu_sleep_check();
4351
4352 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4353
4354 schedstat_inc(this_rq(), sched_count);
4355}
4356
4357static void put_prev_task(struct rq *rq, struct task_struct *prev)
4358{
4359 if (prev->on_rq || rq->skip_clock_update < 0)
4360 update_rq_clock(rq);
4361 prev->sched_class->put_prev_task(rq, prev);
4362}
4363
4364/*
4365 * Pick up the highest-prio task:
4366 */
4367static inline struct task_struct *
4368pick_next_task(struct rq *rq)
4369{
4370 const struct sched_class *class;
4371 struct task_struct *p;
4372
4373 /*
4374 * Optimization: we know that if all tasks are in
4375 * the fair class we can call that function directly:
4376 */
4377 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
4378 p = fair_sched_class.pick_next_task(rq);
4379 if (likely(p))
4380 return p;
4381 }
4382
4383 for_each_class(class) {
4384 p = class->pick_next_task(rq);
4385 if (p)
4386 return p;
4387 }
4388
4389 BUG(); /* the idle class will always have a runnable task */
4390}
4391
4392/*
4393 * __schedule() is the main scheduler function.
4394 */
4395static void __sched __schedule(void)
4396{
4397 struct task_struct *prev, *next;
4398 unsigned long *switch_count;
4399 struct rq *rq;
4400 int cpu;
4401
4402need_resched:
4403 preempt_disable();
4404 cpu = smp_processor_id();
4405 rq = cpu_rq(cpu);
4406 rcu_note_context_switch(cpu);
4407 prev = rq->curr;
4408
4409 schedule_debug(prev);
4410
4411 if (sched_feat(HRTICK))
4412 hrtick_clear(rq);
4413
4414 raw_spin_lock_irq(&rq->lock);
4415
4416 switch_count = &prev->nivcsw;
4417 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4418 if (unlikely(signal_pending_state(prev->state, prev))) {
4419 prev->state = TASK_RUNNING;
4420 } else {
4421 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4422 prev->on_rq = 0;
4423
4424 /*
4425 * If a worker went to sleep, notify and ask workqueue
4426 * whether it wants to wake up a task to maintain
4427 * concurrency.
4428 */
4429 if (prev->flags & PF_WQ_WORKER) {
4430 struct task_struct *to_wakeup;
4431
4432 to_wakeup = wq_worker_sleeping(prev, cpu);
4433 if (to_wakeup)
4434 try_to_wake_up_local(to_wakeup);
4435 }
4436 }
4437 switch_count = &prev->nvcsw;
4438 }
4439
4440 pre_schedule(rq, prev);
4441
4442 if (unlikely(!rq->nr_running))
4443 idle_balance(cpu, rq);
4444
4445 put_prev_task(rq, prev);
4446 next = pick_next_task(rq);
4447 clear_tsk_need_resched(prev);
4448 rq->skip_clock_update = 0;
4449
4450 if (likely(prev != next)) {
4451 rq->nr_switches++;
4452 rq->curr = next;
4453 ++*switch_count;
4454
4455 context_switch(rq, prev, next); /* unlocks the rq */
4456 /*
4457 * The context switch have flipped the stack from under us
4458 * and restored the local variables which were saved when
4459 * this task called schedule() in the past. prev == current
4460 * is still correct, but it can be moved to another cpu/rq.
4461 */
4462 cpu = smp_processor_id();
4463 rq = cpu_rq(cpu);
4464 } else
4465 raw_spin_unlock_irq(&rq->lock);
4466
4467 post_schedule(rq);
4468
4469 preempt_enable_no_resched();
4470 if (need_resched())
4471 goto need_resched;
4472}
4473
4474static inline void sched_submit_work(struct task_struct *tsk)
4475{
4476 if (!tsk->state)
4477 return;
4478 /*
4479 * If we are going to sleep and we have plugged IO queued,
4480 * make sure to submit it to avoid deadlocks.
4481 */
4482 if (blk_needs_flush_plug(tsk))
4483 blk_schedule_flush_plug(tsk);
4484}
4485
4486asmlinkage void __sched schedule(void)
4487{
4488 struct task_struct *tsk = current;
4489
4490 sched_submit_work(tsk);
4491 __schedule();
4492}
4493EXPORT_SYMBOL(schedule);
4494
4495#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4496
4497static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4498{
4499 if (lock->owner != owner)
4500 return false;
4501
4502 /*
4503 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4504 * lock->owner still matches owner, if that fails, owner might
4505 * point to free()d memory, if it still matches, the rcu_read_lock()
4506 * ensures the memory stays valid.
4507 */
4508 barrier();
4509
4510 return owner->on_cpu;
4511}
4512
4513/*
4514 * Look out! "owner" is an entirely speculative pointer
4515 * access and not reliable.
4516 */
4517int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4518{
4519 if (!sched_feat(OWNER_SPIN))
4520 return 0;
4521
4522 rcu_read_lock();
4523 while (owner_running(lock, owner)) {
4524 if (need_resched())
4525 break;
4526
4527 arch_mutex_cpu_relax();
4528 }
4529 rcu_read_unlock();
4530
4531 /*
4532 * We break out the loop above on need_resched() and when the
4533 * owner changed, which is a sign for heavy contention. Return
4534 * success only when lock->owner is NULL.
4535 */
4536 return lock->owner == NULL;
4537}
4538#endif
4539
4540#ifdef CONFIG_PREEMPT
4541/*
4542 * this is the entry point to schedule() from in-kernel preemption
4543 * off of preempt_enable. Kernel preemptions off return from interrupt
4544 * occur there and call schedule directly.
4545 */
4546asmlinkage void __sched notrace preempt_schedule(void)
4547{
4548 struct thread_info *ti = current_thread_info();
4549
4550 /*
4551 * If there is a non-zero preempt_count or interrupts are disabled,
4552 * we do not want to preempt the current task. Just return..
4553 */
4554 if (likely(ti->preempt_count || irqs_disabled()))
4555 return;
4556
4557 do {
4558 add_preempt_count_notrace(PREEMPT_ACTIVE);
4559 __schedule();
4560 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4561
4562 /*
4563 * Check again in case we missed a preemption opportunity
4564 * between schedule and now.
4565 */
4566 barrier();
4567 } while (need_resched());
4568}
4569EXPORT_SYMBOL(preempt_schedule);
4570
4571/*
4572 * this is the entry point to schedule() from kernel preemption
4573 * off of irq context.
4574 * Note, that this is called and return with irqs disabled. This will
4575 * protect us against recursive calling from irq.
4576 */
4577asmlinkage void __sched preempt_schedule_irq(void)
4578{
4579 struct thread_info *ti = current_thread_info();
4580
4581 /* Catch callers which need to be fixed */
4582 BUG_ON(ti->preempt_count || !irqs_disabled());
4583
4584 do {
4585 add_preempt_count(PREEMPT_ACTIVE);
4586 local_irq_enable();
4587 __schedule();
4588 local_irq_disable();
4589 sub_preempt_count(PREEMPT_ACTIVE);
4590
4591 /*
4592 * Check again in case we missed a preemption opportunity
4593 * between schedule and now.
4594 */
4595 barrier();
4596 } while (need_resched());
4597}
4598
4599#endif /* CONFIG_PREEMPT */
4600
4601int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4602 void *key)
4603{
4604 return try_to_wake_up(curr->private, mode, wake_flags);
4605}
4606EXPORT_SYMBOL(default_wake_function);
4607
4608/*
4609 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4610 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4611 * number) then we wake all the non-exclusive tasks and one exclusive task.
4612 *
4613 * There are circumstances in which we can try to wake a task which has already
4614 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4615 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4616 */
4617static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4618 int nr_exclusive, int wake_flags, void *key)
4619{
4620 wait_queue_t *curr, *next;
4621
4622 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4623 unsigned flags = curr->flags;
4624
4625 if (curr->func(curr, mode, wake_flags, key) &&
4626 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4627 break;
4628 }
4629}
4630
4631/**
4632 * __wake_up - wake up threads blocked on a waitqueue.
4633 * @q: the waitqueue
4634 * @mode: which threads
4635 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4636 * @key: is directly passed to the wakeup function
4637 *
4638 * It may be assumed that this function implies a write memory barrier before
4639 * changing the task state if and only if any tasks are woken up.
4640 */
4641void __wake_up(wait_queue_head_t *q, unsigned int mode,
4642 int nr_exclusive, void *key)
4643{
4644 unsigned long flags;
4645
4646 spin_lock_irqsave(&q->lock, flags);
4647 __wake_up_common(q, mode, nr_exclusive, 0, key);
4648 spin_unlock_irqrestore(&q->lock, flags);
4649}
4650EXPORT_SYMBOL(__wake_up);
4651
4652/*
4653 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4654 */
4655void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4656{
4657 __wake_up_common(q, mode, 1, 0, NULL);
4658}
4659EXPORT_SYMBOL_GPL(__wake_up_locked);
4660
4661void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4662{
4663 __wake_up_common(q, mode, 1, 0, key);
4664}
4665EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4666
4667/**
4668 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4669 * @q: the waitqueue
4670 * @mode: which threads
4671 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4672 * @key: opaque value to be passed to wakeup targets
4673 *
4674 * The sync wakeup differs that the waker knows that it will schedule
4675 * away soon, so while the target thread will be woken up, it will not
4676 * be migrated to another CPU - ie. the two threads are 'synchronized'
4677 * with each other. This can prevent needless bouncing between CPUs.
4678 *
4679 * On UP it can prevent extra preemption.
4680 *
4681 * It may be assumed that this function implies a write memory barrier before
4682 * changing the task state if and only if any tasks are woken up.
4683 */
4684void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4685 int nr_exclusive, void *key)
4686{
4687 unsigned long flags;
4688 int wake_flags = WF_SYNC;
4689
4690 if (unlikely(!q))
4691 return;
4692
4693 if (unlikely(!nr_exclusive))
4694 wake_flags = 0;
4695
4696 spin_lock_irqsave(&q->lock, flags);
4697 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4698 spin_unlock_irqrestore(&q->lock, flags);
4699}
4700EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4701
4702/*
4703 * __wake_up_sync - see __wake_up_sync_key()
4704 */
4705void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4706{
4707 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4708}
4709EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4710
4711/**
4712 * complete: - signals a single thread waiting on this completion
4713 * @x: holds the state of this particular completion
4714 *
4715 * This will wake up a single thread waiting on this completion. Threads will be
4716 * awakened in the same order in which they were queued.
4717 *
4718 * See also complete_all(), wait_for_completion() and related routines.
4719 *
4720 * It may be assumed that this function implies a write memory barrier before
4721 * changing the task state if and only if any tasks are woken up.
4722 */
4723void complete(struct completion *x)
4724{
4725 unsigned long flags;
4726
4727 spin_lock_irqsave(&x->wait.lock, flags);
4728 x->done++;
4729 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4730 spin_unlock_irqrestore(&x->wait.lock, flags);
4731}
4732EXPORT_SYMBOL(complete);
4733
4734/**
4735 * complete_all: - signals all threads waiting on this completion
4736 * @x: holds the state of this particular completion
4737 *
4738 * This will wake up all threads waiting on this particular completion event.
4739 *
4740 * It may be assumed that this function implies a write memory barrier before
4741 * changing the task state if and only if any tasks are woken up.
4742 */
4743void complete_all(struct completion *x)
4744{
4745 unsigned long flags;
4746
4747 spin_lock_irqsave(&x->wait.lock, flags);
4748 x->done += UINT_MAX/2;
4749 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4750 spin_unlock_irqrestore(&x->wait.lock, flags);
4751}
4752EXPORT_SYMBOL(complete_all);
4753
4754static inline long __sched
4755do_wait_for_common(struct completion *x, long timeout, int state)
4756{
4757 if (!x->done) {
4758 DECLARE_WAITQUEUE(wait, current);
4759
4760 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4761 do {
4762 if (signal_pending_state(state, current)) {
4763 timeout = -ERESTARTSYS;
4764 break;
4765 }
4766 __set_current_state(state);
4767 spin_unlock_irq(&x->wait.lock);
4768 timeout = schedule_timeout(timeout);
4769 spin_lock_irq(&x->wait.lock);
4770 } while (!x->done && timeout);
4771 __remove_wait_queue(&x->wait, &wait);
4772 if (!x->done)
4773 return timeout;
4774 }
4775 x->done--;
4776 return timeout ?: 1;
4777}
4778
4779static long __sched
4780wait_for_common(struct completion *x, long timeout, int state)
4781{
4782 might_sleep();
4783
4784 spin_lock_irq(&x->wait.lock);
4785 timeout = do_wait_for_common(x, timeout, state);
4786 spin_unlock_irq(&x->wait.lock);
4787 return timeout;
4788}
4789
4790/**
4791 * wait_for_completion: - waits for completion of a task
4792 * @x: holds the state of this particular completion
4793 *
4794 * This waits to be signaled for completion of a specific task. It is NOT
4795 * interruptible and there is no timeout.
4796 *
4797 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4798 * and interrupt capability. Also see complete().
4799 */
4800void __sched wait_for_completion(struct completion *x)
4801{
4802 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4803}
4804EXPORT_SYMBOL(wait_for_completion);
4805
4806/**
4807 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4808 * @x: holds the state of this particular completion
4809 * @timeout: timeout value in jiffies
4810 *
4811 * This waits for either a completion of a specific task to be signaled or for a
4812 * specified timeout to expire. The timeout is in jiffies. It is not
4813 * interruptible.
4814 *
4815 * The return value is 0 if timed out, and positive (at least 1, or number of
4816 * jiffies left till timeout) if completed.
4817 */
4818unsigned long __sched
4819wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4820{
4821 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4822}
4823EXPORT_SYMBOL(wait_for_completion_timeout);
4824
4825/**
4826 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4827 * @x: holds the state of this particular completion
4828 *
4829 * This waits for completion of a specific task to be signaled. It is
4830 * interruptible.
4831 *
4832 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
4833 */
4834int __sched wait_for_completion_interruptible(struct completion *x)
4835{
4836 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4837 if (t == -ERESTARTSYS)
4838 return t;
4839 return 0;
4840}
4841EXPORT_SYMBOL(wait_for_completion_interruptible);
4842
4843/**
4844 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4845 * @x: holds the state of this particular completion
4846 * @timeout: timeout value in jiffies
4847 *
4848 * This waits for either a completion of a specific task to be signaled or for a
4849 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4850 *
4851 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
4852 * positive (at least 1, or number of jiffies left till timeout) if completed.
4853 */
4854long __sched
4855wait_for_completion_interruptible_timeout(struct completion *x,
4856 unsigned long timeout)
4857{
4858 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4859}
4860EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4861
4862/**
4863 * wait_for_completion_killable: - waits for completion of a task (killable)
4864 * @x: holds the state of this particular completion
4865 *
4866 * This waits to be signaled for completion of a specific task. It can be
4867 * interrupted by a kill signal.
4868 *
4869 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
4870 */
4871int __sched wait_for_completion_killable(struct completion *x)
4872{
4873 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4874 if (t == -ERESTARTSYS)
4875 return t;
4876 return 0;
4877}
4878EXPORT_SYMBOL(wait_for_completion_killable);
4879
4880/**
4881 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4882 * @x: holds the state of this particular completion
4883 * @timeout: timeout value in jiffies
4884 *
4885 * This waits for either a completion of a specific task to be
4886 * signaled or for a specified timeout to expire. It can be
4887 * interrupted by a kill signal. The timeout is in jiffies.
4888 *
4889 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
4890 * positive (at least 1, or number of jiffies left till timeout) if completed.
4891 */
4892long __sched
4893wait_for_completion_killable_timeout(struct completion *x,
4894 unsigned long timeout)
4895{
4896 return wait_for_common(x, timeout, TASK_KILLABLE);
4897}
4898EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4899
4900/**
4901 * try_wait_for_completion - try to decrement a completion without blocking
4902 * @x: completion structure
4903 *
4904 * Returns: 0 if a decrement cannot be done without blocking
4905 * 1 if a decrement succeeded.
4906 *
4907 * If a completion is being used as a counting completion,
4908 * attempt to decrement the counter without blocking. This
4909 * enables us to avoid waiting if the resource the completion
4910 * is protecting is not available.
4911 */
4912bool try_wait_for_completion(struct completion *x)
4913{
4914 unsigned long flags;
4915 int ret = 1;
4916
4917 spin_lock_irqsave(&x->wait.lock, flags);
4918 if (!x->done)
4919 ret = 0;
4920 else
4921 x->done--;
4922 spin_unlock_irqrestore(&x->wait.lock, flags);
4923 return ret;
4924}
4925EXPORT_SYMBOL(try_wait_for_completion);
4926
4927/**
4928 * completion_done - Test to see if a completion has any waiters
4929 * @x: completion structure
4930 *
4931 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4932 * 1 if there are no waiters.
4933 *
4934 */
4935bool completion_done(struct completion *x)
4936{
4937 unsigned long flags;
4938 int ret = 1;
4939
4940 spin_lock_irqsave(&x->wait.lock, flags);
4941 if (!x->done)
4942 ret = 0;
4943 spin_unlock_irqrestore(&x->wait.lock, flags);
4944 return ret;
4945}
4946EXPORT_SYMBOL(completion_done);
4947
4948static long __sched
4949sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4950{
4951 unsigned long flags;
4952 wait_queue_t wait;
4953
4954 init_waitqueue_entry(&wait, current);
4955
4956 __set_current_state(state);
4957
4958 spin_lock_irqsave(&q->lock, flags);
4959 __add_wait_queue(q, &wait);
4960 spin_unlock(&q->lock);
4961 timeout = schedule_timeout(timeout);
4962 spin_lock_irq(&q->lock);
4963 __remove_wait_queue(q, &wait);
4964 spin_unlock_irqrestore(&q->lock, flags);
4965
4966 return timeout;
4967}
4968
4969void __sched interruptible_sleep_on(wait_queue_head_t *q)
4970{
4971 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4972}
4973EXPORT_SYMBOL(interruptible_sleep_on);
4974
4975long __sched
4976interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4977{
4978 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4979}
4980EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4981
4982void __sched sleep_on(wait_queue_head_t *q)
4983{
4984 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4985}
4986EXPORT_SYMBOL(sleep_on);
4987
4988long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4989{
4990 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4991}
4992EXPORT_SYMBOL(sleep_on_timeout);
4993
4994#ifdef CONFIG_RT_MUTEXES
4995
4996/*
4997 * rt_mutex_setprio - set the current priority of a task
4998 * @p: task
4999 * @prio: prio value (kernel-internal form)
5000 *
5001 * This function changes the 'effective' priority of a task. It does
5002 * not touch ->normal_prio like __setscheduler().
5003 *
5004 * Used by the rt_mutex code to implement priority inheritance logic.
5005 */
5006void rt_mutex_setprio(struct task_struct *p, int prio)
5007{
5008 int oldprio, on_rq, running;
5009 struct rq *rq;
5010 const struct sched_class *prev_class;
5011
5012 BUG_ON(prio < 0 || prio > MAX_PRIO);
5013
5014 rq = __task_rq_lock(p);
5015
5016 trace_sched_pi_setprio(p, prio);
5017 oldprio = p->prio;
5018 prev_class = p->sched_class;
5019 on_rq = p->on_rq;
5020 running = task_current(rq, p);
5021 if (on_rq)
5022 dequeue_task(rq, p, 0);
5023 if (running)
5024 p->sched_class->put_prev_task(rq, p);
5025
5026 if (rt_prio(prio))
5027 p->sched_class = &rt_sched_class;
5028 else
5029 p->sched_class = &fair_sched_class;
5030
5031 p->prio = prio;
5032
5033 if (running)
5034 p->sched_class->set_curr_task(rq);
5035 if (on_rq)
5036 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
5037
5038 check_class_changed(rq, p, prev_class, oldprio);
5039 __task_rq_unlock(rq);
5040}
5041
5042#endif
5043
5044void set_user_nice(struct task_struct *p, long nice)
5045{
5046 int old_prio, delta, on_rq;
5047 unsigned long flags;
5048 struct rq *rq;
5049
5050 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5051 return;
5052 /*
5053 * We have to be careful, if called from sys_setpriority(),
5054 * the task might be in the middle of scheduling on another CPU.
5055 */
5056 rq = task_rq_lock(p, &flags);
5057 /*
5058 * The RT priorities are set via sched_setscheduler(), but we still
5059 * allow the 'normal' nice value to be set - but as expected
5060 * it wont have any effect on scheduling until the task is
5061 * SCHED_FIFO/SCHED_RR:
5062 */
5063 if (task_has_rt_policy(p)) {
5064 p->static_prio = NICE_TO_PRIO(nice);
5065 goto out_unlock;
5066 }
5067 on_rq = p->on_rq;
5068 if (on_rq)
5069 dequeue_task(rq, p, 0);
5070
5071 p->static_prio = NICE_TO_PRIO(nice);
5072 set_load_weight(p);
5073 old_prio = p->prio;
5074 p->prio = effective_prio(p);
5075 delta = p->prio - old_prio;
5076
5077 if (on_rq) {
5078 enqueue_task(rq, p, 0);
5079 /*
5080 * If the task increased its priority or is running and
5081 * lowered its priority, then reschedule its CPU:
5082 */
5083 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5084 resched_task(rq->curr);
5085 }
5086out_unlock:
5087 task_rq_unlock(rq, p, &flags);
5088}
5089EXPORT_SYMBOL(set_user_nice);
5090
5091/*
5092 * can_nice - check if a task can reduce its nice value
5093 * @p: task
5094 * @nice: nice value
5095 */
5096int can_nice(const struct task_struct *p, const int nice)
5097{
5098 /* convert nice value [19,-20] to rlimit style value [1,40] */
5099 int nice_rlim = 20 - nice;
5100
5101 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
5102 capable(CAP_SYS_NICE));
5103}
5104
5105#ifdef __ARCH_WANT_SYS_NICE
5106
5107/*
5108 * sys_nice - change the priority of the current process.
5109 * @increment: priority increment
5110 *
5111 * sys_setpriority is a more generic, but much slower function that
5112 * does similar things.
5113 */
5114SYSCALL_DEFINE1(nice, int, increment)
5115{
5116 long nice, retval;
5117
5118 /*
5119 * Setpriority might change our priority at the same moment.
5120 * We don't have to worry. Conceptually one call occurs first
5121 * and we have a single winner.
5122 */
5123 if (increment < -40)
5124 increment = -40;
5125 if (increment > 40)
5126 increment = 40;
5127
5128 nice = TASK_NICE(current) + increment;
5129 if (nice < -20)
5130 nice = -20;
5131 if (nice > 19)
5132 nice = 19;
5133
5134 if (increment < 0 && !can_nice(current, nice))
5135 return -EPERM;
5136
5137 retval = security_task_setnice(current, nice);
5138 if (retval)
5139 return retval;
5140
5141 set_user_nice(current, nice);
5142 return 0;
5143}
5144
5145#endif
5146
5147/**
5148 * task_prio - return the priority value of a given task.
5149 * @p: the task in question.
5150 *
5151 * This is the priority value as seen by users in /proc.
5152 * RT tasks are offset by -200. Normal tasks are centered
5153 * around 0, value goes from -16 to +15.
5154 */
5155int task_prio(const struct task_struct *p)
5156{
5157 return p->prio - MAX_RT_PRIO;
5158}
5159
5160/**
5161 * task_nice - return the nice value of a given task.
5162 * @p: the task in question.
5163 */
5164int task_nice(const struct task_struct *p)
5165{
5166 return TASK_NICE(p);
5167}
5168EXPORT_SYMBOL(task_nice);
5169
5170/**
5171 * idle_cpu - is a given cpu idle currently?
5172 * @cpu: the processor in question.
5173 */
5174int idle_cpu(int cpu)
5175{
5176 struct rq *rq = cpu_rq(cpu);
5177
5178 if (rq->curr != rq->idle)
5179 return 0;
5180
5181 if (rq->nr_running)
5182 return 0;
5183
5184#ifdef CONFIG_SMP
5185 if (!llist_empty(&rq->wake_list))
5186 return 0;
5187#endif
5188
5189 return 1;
5190}
5191
5192/**
5193 * idle_task - return the idle task for a given cpu.
5194 * @cpu: the processor in question.
5195 */
5196struct task_struct *idle_task(int cpu)
5197{
5198 return cpu_rq(cpu)->idle;
5199}
5200
5201/**
5202 * find_process_by_pid - find a process with a matching PID value.
5203 * @pid: the pid in question.
5204 */
5205static struct task_struct *find_process_by_pid(pid_t pid)
5206{
5207 return pid ? find_task_by_vpid(pid) : current;
5208}
5209
5210/* Actually do priority change: must hold rq lock. */
5211static void
5212__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5213{
5214 p->policy = policy;
5215 p->rt_priority = prio;
5216 p->normal_prio = normal_prio(p);
5217 /* we are holding p->pi_lock already */
5218 p->prio = rt_mutex_getprio(p);
5219 if (rt_prio(p->prio))
5220 p->sched_class = &rt_sched_class;
5221 else
5222 p->sched_class = &fair_sched_class;
5223 set_load_weight(p);
5224}
5225
5226/*
5227 * check the target process has a UID that matches the current process's
5228 */
5229static bool check_same_owner(struct task_struct *p)
5230{
5231 const struct cred *cred = current_cred(), *pcred;
5232 bool match;
5233
5234 rcu_read_lock();
5235 pcred = __task_cred(p);
5236 if (cred->user->user_ns == pcred->user->user_ns)
5237 match = (cred->euid == pcred->euid ||
5238 cred->euid == pcred->uid);
5239 else
5240 match = false;
5241 rcu_read_unlock();
5242 return match;
5243}
5244
5245static int __sched_setscheduler(struct task_struct *p, int policy,
5246 const struct sched_param *param, bool user)
5247{
5248 int retval, oldprio, oldpolicy = -1, on_rq, running;
5249 unsigned long flags;
5250 const struct sched_class *prev_class;
5251 struct rq *rq;
5252 int reset_on_fork;
5253
5254 /* may grab non-irq protected spin_locks */
5255 BUG_ON(in_interrupt());
5256recheck:
5257 /* double check policy once rq lock held */
5258 if (policy < 0) {
5259 reset_on_fork = p->sched_reset_on_fork;
5260 policy = oldpolicy = p->policy;
5261 } else {
5262 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5263 policy &= ~SCHED_RESET_ON_FORK;
5264
5265 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5266 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5267 policy != SCHED_IDLE)
5268 return -EINVAL;
5269 }
5270
5271 /*
5272 * Valid priorities for SCHED_FIFO and SCHED_RR are
5273 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5274 * SCHED_BATCH and SCHED_IDLE is 0.
5275 */
5276 if (param->sched_priority < 0 ||
5277 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5278 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5279 return -EINVAL;
5280 if (rt_policy(policy) != (param->sched_priority != 0))
5281 return -EINVAL;
5282
5283 /*
5284 * Allow unprivileged RT tasks to decrease priority:
5285 */
5286 if (user && !capable(CAP_SYS_NICE)) {
5287 if (rt_policy(policy)) {
5288 unsigned long rlim_rtprio =
5289 task_rlimit(p, RLIMIT_RTPRIO);
5290
5291 /* can't set/change the rt policy */
5292 if (policy != p->policy && !rlim_rtprio)
5293 return -EPERM;
5294
5295 /* can't increase priority */
5296 if (param->sched_priority > p->rt_priority &&
5297 param->sched_priority > rlim_rtprio)
5298 return -EPERM;
5299 }
5300
5301 /*
5302 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5303 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5304 */
5305 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5306 if (!can_nice(p, TASK_NICE(p)))
5307 return -EPERM;
5308 }
5309
5310 /* can't change other user's priorities */
5311 if (!check_same_owner(p))
5312 return -EPERM;
5313
5314 /* Normal users shall not reset the sched_reset_on_fork flag */
5315 if (p->sched_reset_on_fork && !reset_on_fork)
5316 return -EPERM;
5317 }
5318
5319 if (user) {
5320 retval = security_task_setscheduler(p);
5321 if (retval)
5322 return retval;
5323 }
5324
5325 /*
5326 * make sure no PI-waiters arrive (or leave) while we are
5327 * changing the priority of the task:
5328 *
5329 * To be able to change p->policy safely, the appropriate
5330 * runqueue lock must be held.
5331 */
5332 rq = task_rq_lock(p, &flags);
5333
5334 /*
5335 * Changing the policy of the stop threads its a very bad idea
5336 */
5337 if (p == rq->stop) {
5338 task_rq_unlock(rq, p, &flags);
5339 return -EINVAL;
5340 }
5341
5342 /*
5343 * If not changing anything there's no need to proceed further:
5344 */
5345 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5346 param->sched_priority == p->rt_priority))) {
5347
5348 __task_rq_unlock(rq);
5349 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5350 return 0;
5351 }
5352
5353#ifdef CONFIG_RT_GROUP_SCHED
5354 if (user) {
5355 /*
5356 * Do not allow realtime tasks into groups that have no runtime
5357 * assigned.
5358 */
5359 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5360 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5361 !task_group_is_autogroup(task_group(p))) {
5362 task_rq_unlock(rq, p, &flags);
5363 return -EPERM;
5364 }
5365 }
5366#endif
5367
5368 /* recheck policy now with rq lock held */
5369 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5370 policy = oldpolicy = -1;
5371 task_rq_unlock(rq, p, &flags);
5372 goto recheck;
5373 }
5374 on_rq = p->on_rq;
5375 running = task_current(rq, p);
5376 if (on_rq)
5377 deactivate_task(rq, p, 0);
5378 if (running)
5379 p->sched_class->put_prev_task(rq, p);
5380
5381 p->sched_reset_on_fork = reset_on_fork;
5382
5383 oldprio = p->prio;
5384 prev_class = p->sched_class;
5385 __setscheduler(rq, p, policy, param->sched_priority);
5386
5387 if (running)
5388 p->sched_class->set_curr_task(rq);
5389 if (on_rq)
5390 activate_task(rq, p, 0);
5391
5392 check_class_changed(rq, p, prev_class, oldprio);
5393 task_rq_unlock(rq, p, &flags);
5394
5395 rt_mutex_adjust_pi(p);
5396
5397 return 0;
5398}
5399
5400/**
5401 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5402 * @p: the task in question.
5403 * @policy: new policy.
5404 * @param: structure containing the new RT priority.
5405 *
5406 * NOTE that the task may be already dead.
5407 */
5408int sched_setscheduler(struct task_struct *p, int policy,
5409 const struct sched_param *param)
5410{
5411 return __sched_setscheduler(p, policy, param, true);
5412}
5413EXPORT_SYMBOL_GPL(sched_setscheduler);
5414
5415/**
5416 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5417 * @p: the task in question.
5418 * @policy: new policy.
5419 * @param: structure containing the new RT priority.
5420 *
5421 * Just like sched_setscheduler, only don't bother checking if the
5422 * current context has permission. For example, this is needed in
5423 * stop_machine(): we create temporary high priority worker threads,
5424 * but our caller might not have that capability.
5425 */
5426int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5427 const struct sched_param *param)
5428{
5429 return __sched_setscheduler(p, policy, param, false);
5430}
5431
5432static int
5433do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5434{
5435 struct sched_param lparam;
5436 struct task_struct *p;
5437 int retval;
5438
5439 if (!param || pid < 0)
5440 return -EINVAL;
5441 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5442 return -EFAULT;
5443
5444 rcu_read_lock();
5445 retval = -ESRCH;
5446 p = find_process_by_pid(pid);
5447 if (p != NULL)
5448 retval = sched_setscheduler(p, policy, &lparam);
5449 rcu_read_unlock();
5450
5451 return retval;
5452}
5453
5454/**
5455 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5456 * @pid: the pid in question.
5457 * @policy: new policy.
5458 * @param: structure containing the new RT priority.
5459 */
5460SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5461 struct sched_param __user *, param)
5462{
5463 /* negative values for policy are not valid */
5464 if (policy < 0)
5465 return -EINVAL;
5466
5467 return do_sched_setscheduler(pid, policy, param);
5468}
5469
5470/**
5471 * sys_sched_setparam - set/change the RT priority of a thread
5472 * @pid: the pid in question.
5473 * @param: structure containing the new RT priority.
5474 */
5475SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5476{
5477 return do_sched_setscheduler(pid, -1, param);
5478}
5479
5480/**
5481 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5482 * @pid: the pid in question.
5483 */
5484SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5485{
5486 struct task_struct *p;
5487 int retval;
5488
5489 if (pid < 0)
5490 return -EINVAL;
5491
5492 retval = -ESRCH;
5493 rcu_read_lock();
5494 p = find_process_by_pid(pid);
5495 if (p) {
5496 retval = security_task_getscheduler(p);
5497 if (!retval)
5498 retval = p->policy
5499 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5500 }
5501 rcu_read_unlock();
5502 return retval;
5503}
5504
5505/**
5506 * sys_sched_getparam - get the RT priority of a thread
5507 * @pid: the pid in question.
5508 * @param: structure containing the RT priority.
5509 */
5510SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5511{
5512 struct sched_param lp;
5513 struct task_struct *p;
5514 int retval;
5515
5516 if (!param || pid < 0)
5517 return -EINVAL;
5518
5519 rcu_read_lock();
5520 p = find_process_by_pid(pid);
5521 retval = -ESRCH;
5522 if (!p)
5523 goto out_unlock;
5524
5525 retval = security_task_getscheduler(p);
5526 if (retval)
5527 goto out_unlock;
5528
5529 lp.sched_priority = p->rt_priority;
5530 rcu_read_unlock();
5531
5532 /*
5533 * This one might sleep, we cannot do it with a spinlock held ...
5534 */
5535 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5536
5537 return retval;
5538
5539out_unlock:
5540 rcu_read_unlock();
5541 return retval;
5542}
5543
5544long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5545{
5546 cpumask_var_t cpus_allowed, new_mask;
5547 struct task_struct *p;
5548 int retval;
5549
5550 get_online_cpus();
5551 rcu_read_lock();
5552
5553 p = find_process_by_pid(pid);
5554 if (!p) {
5555 rcu_read_unlock();
5556 put_online_cpus();
5557 return -ESRCH;
5558 }
5559
5560 /* Prevent p going away */
5561 get_task_struct(p);
5562 rcu_read_unlock();
5563
5564 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5565 retval = -ENOMEM;
5566 goto out_put_task;
5567 }
5568 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5569 retval = -ENOMEM;
5570 goto out_free_cpus_allowed;
5571 }
5572 retval = -EPERM;
5573 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5574 goto out_unlock;
5575
5576 retval = security_task_setscheduler(p);
5577 if (retval)
5578 goto out_unlock;
5579
5580 cpuset_cpus_allowed(p, cpus_allowed);
5581 cpumask_and(new_mask, in_mask, cpus_allowed);
5582again:
5583 retval = set_cpus_allowed_ptr(p, new_mask);
5584
5585 if (!retval) {
5586 cpuset_cpus_allowed(p, cpus_allowed);
5587 if (!cpumask_subset(new_mask, cpus_allowed)) {
5588 /*
5589 * We must have raced with a concurrent cpuset
5590 * update. Just reset the cpus_allowed to the
5591 * cpuset's cpus_allowed
5592 */
5593 cpumask_copy(new_mask, cpus_allowed);
5594 goto again;
5595 }
5596 }
5597out_unlock:
5598 free_cpumask_var(new_mask);
5599out_free_cpus_allowed:
5600 free_cpumask_var(cpus_allowed);
5601out_put_task:
5602 put_task_struct(p);
5603 put_online_cpus();
5604 return retval;
5605}
5606
5607static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5608 struct cpumask *new_mask)
5609{
5610 if (len < cpumask_size())
5611 cpumask_clear(new_mask);
5612 else if (len > cpumask_size())
5613 len = cpumask_size();
5614
5615 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5616}
5617
5618/**
5619 * sys_sched_setaffinity - set the cpu affinity of a process
5620 * @pid: pid of the process
5621 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5622 * @user_mask_ptr: user-space pointer to the new cpu mask
5623 */
5624SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5625 unsigned long __user *, user_mask_ptr)
5626{
5627 cpumask_var_t new_mask;
5628 int retval;
5629
5630 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5631 return -ENOMEM;
5632
5633 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5634 if (retval == 0)
5635 retval = sched_setaffinity(pid, new_mask);
5636 free_cpumask_var(new_mask);
5637 return retval;
5638}
5639
5640long sched_getaffinity(pid_t pid, struct cpumask *mask)
5641{
5642 struct task_struct *p;
5643 unsigned long flags;
5644 int retval;
5645
5646 get_online_cpus();
5647 rcu_read_lock();
5648
5649 retval = -ESRCH;
5650 p = find_process_by_pid(pid);
5651 if (!p)
5652 goto out_unlock;
5653
5654 retval = security_task_getscheduler(p);
5655 if (retval)
5656 goto out_unlock;
5657
5658 raw_spin_lock_irqsave(&p->pi_lock, flags);
5659 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5660 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5661
5662out_unlock:
5663 rcu_read_unlock();
5664 put_online_cpus();
5665
5666 return retval;
5667}
5668
5669/**
5670 * sys_sched_getaffinity - get the cpu affinity of a process
5671 * @pid: pid of the process
5672 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5673 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5674 */
5675SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5676 unsigned long __user *, user_mask_ptr)
5677{
5678 int ret;
5679 cpumask_var_t mask;
5680
5681 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5682 return -EINVAL;
5683 if (len & (sizeof(unsigned long)-1))
5684 return -EINVAL;
5685
5686 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5687 return -ENOMEM;
5688
5689 ret = sched_getaffinity(pid, mask);
5690 if (ret == 0) {
5691 size_t retlen = min_t(size_t, len, cpumask_size());
5692
5693 if (copy_to_user(user_mask_ptr, mask, retlen))
5694 ret = -EFAULT;
5695 else
5696 ret = retlen;
5697 }
5698 free_cpumask_var(mask);
5699
5700 return ret;
5701}
5702
5703/**
5704 * sys_sched_yield - yield the current processor to other threads.
5705 *
5706 * This function yields the current CPU to other tasks. If there are no
5707 * other threads running on this CPU then this function will return.
5708 */
5709SYSCALL_DEFINE0(sched_yield)
5710{
5711 struct rq *rq = this_rq_lock();
5712
5713 schedstat_inc(rq, yld_count);
5714 current->sched_class->yield_task(rq);
5715
5716 /*
5717 * Since we are going to call schedule() anyway, there's
5718 * no need to preempt or enable interrupts:
5719 */
5720 __release(rq->lock);
5721 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5722 do_raw_spin_unlock(&rq->lock);
5723 preempt_enable_no_resched();
5724
5725 schedule();
5726
5727 return 0;
5728}
5729
5730static inline int should_resched(void)
5731{
5732 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5733}
5734
5735static void __cond_resched(void)
5736{
5737 add_preempt_count(PREEMPT_ACTIVE);
5738 __schedule();
5739 sub_preempt_count(PREEMPT_ACTIVE);
5740}
5741
5742int __sched _cond_resched(void)
5743{
5744 if (should_resched()) {
5745 __cond_resched();
5746 return 1;
5747 }
5748 return 0;
5749}
5750EXPORT_SYMBOL(_cond_resched);
5751
5752/*
5753 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5754 * call schedule, and on return reacquire the lock.
5755 *
5756 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5757 * operations here to prevent schedule() from being called twice (once via
5758 * spin_unlock(), once by hand).
5759 */
5760int __cond_resched_lock(spinlock_t *lock)
5761{
5762 int resched = should_resched();
5763 int ret = 0;
5764
5765 lockdep_assert_held(lock);
5766
5767 if (spin_needbreak(lock) || resched) {
5768 spin_unlock(lock);
5769 if (resched)
5770 __cond_resched();
5771 else
5772 cpu_relax();
5773 ret = 1;
5774 spin_lock(lock);
5775 }
5776 return ret;
5777}
5778EXPORT_SYMBOL(__cond_resched_lock);
5779
5780int __sched __cond_resched_softirq(void)
5781{
5782 BUG_ON(!in_softirq());
5783
5784 if (should_resched()) {
5785 local_bh_enable();
5786 __cond_resched();
5787 local_bh_disable();
5788 return 1;
5789 }
5790 return 0;
5791}
5792EXPORT_SYMBOL(__cond_resched_softirq);
5793
5794/**
5795 * yield - yield the current processor to other threads.
5796 *
5797 * This is a shortcut for kernel-space yielding - it marks the
5798 * thread runnable and calls sys_sched_yield().
5799 */
5800void __sched yield(void)
5801{
5802 set_current_state(TASK_RUNNING);
5803 sys_sched_yield();
5804}
5805EXPORT_SYMBOL(yield);
5806
5807/**
5808 * yield_to - yield the current processor to another thread in
5809 * your thread group, or accelerate that thread toward the
5810 * processor it's on.
5811 * @p: target task
5812 * @preempt: whether task preemption is allowed or not
5813 *
5814 * It's the caller's job to ensure that the target task struct
5815 * can't go away on us before we can do any checks.
5816 *
5817 * Returns true if we indeed boosted the target task.
5818 */
5819bool __sched yield_to(struct task_struct *p, bool preempt)
5820{
5821 struct task_struct *curr = current;
5822 struct rq *rq, *p_rq;
5823 unsigned long flags;
5824 bool yielded = 0;
5825
5826 local_irq_save(flags);
5827 rq = this_rq();
5828
5829again:
5830 p_rq = task_rq(p);
5831 double_rq_lock(rq, p_rq);
5832 while (task_rq(p) != p_rq) {
5833 double_rq_unlock(rq, p_rq);
5834 goto again;
5835 }
5836
5837 if (!curr->sched_class->yield_to_task)
5838 goto out;
5839
5840 if (curr->sched_class != p->sched_class)
5841 goto out;
5842
5843 if (task_running(p_rq, p) || p->state)
5844 goto out;
5845
5846 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5847 if (yielded) {
5848 schedstat_inc(rq, yld_count);
5849 /*
5850 * Make p's CPU reschedule; pick_next_entity takes care of
5851 * fairness.
5852 */
5853 if (preempt && rq != p_rq)
5854 resched_task(p_rq->curr);
5855 }
5856
5857out:
5858 double_rq_unlock(rq, p_rq);
5859 local_irq_restore(flags);
5860
5861 if (yielded)
5862 schedule();
5863
5864 return yielded;
5865}
5866EXPORT_SYMBOL_GPL(yield_to);
5867
5868/*
5869 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5870 * that process accounting knows that this is a task in IO wait state.
5871 */
5872void __sched io_schedule(void)
5873{
5874 struct rq *rq = raw_rq();
5875
5876 delayacct_blkio_start();
5877 atomic_inc(&rq->nr_iowait);
5878 blk_flush_plug(current);
5879 current->in_iowait = 1;
5880 schedule();
5881 current->in_iowait = 0;
5882 atomic_dec(&rq->nr_iowait);
5883 delayacct_blkio_end();
5884}
5885EXPORT_SYMBOL(io_schedule);
5886
5887long __sched io_schedule_timeout(long timeout)
5888{
5889 struct rq *rq = raw_rq();
5890 long ret;
5891
5892 delayacct_blkio_start();
5893 atomic_inc(&rq->nr_iowait);
5894 blk_flush_plug(current);
5895 current->in_iowait = 1;
5896 ret = schedule_timeout(timeout);
5897 current->in_iowait = 0;
5898 atomic_dec(&rq->nr_iowait);
5899 delayacct_blkio_end();
5900 return ret;
5901}
5902
5903/**
5904 * sys_sched_get_priority_max - return maximum RT priority.
5905 * @policy: scheduling class.
5906 *
5907 * this syscall returns the maximum rt_priority that can be used
5908 * by a given scheduling class.
5909 */
5910SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5911{
5912 int ret = -EINVAL;
5913
5914 switch (policy) {
5915 case SCHED_FIFO:
5916 case SCHED_RR:
5917 ret = MAX_USER_RT_PRIO-1;
5918 break;
5919 case SCHED_NORMAL:
5920 case SCHED_BATCH:
5921 case SCHED_IDLE:
5922 ret = 0;
5923 break;
5924 }
5925 return ret;
5926}
5927
5928/**
5929 * sys_sched_get_priority_min - return minimum RT priority.
5930 * @policy: scheduling class.
5931 *
5932 * this syscall returns the minimum rt_priority that can be used
5933 * by a given scheduling class.
5934 */
5935SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5936{
5937 int ret = -EINVAL;
5938
5939 switch (policy) {
5940 case SCHED_FIFO:
5941 case SCHED_RR:
5942 ret = 1;
5943 break;
5944 case SCHED_NORMAL:
5945 case SCHED_BATCH:
5946 case SCHED_IDLE:
5947 ret = 0;
5948 }
5949 return ret;
5950}
5951
5952/**
5953 * sys_sched_rr_get_interval - return the default timeslice of a process.
5954 * @pid: pid of the process.
5955 * @interval: userspace pointer to the timeslice value.
5956 *
5957 * this syscall writes the default timeslice value of a given process
5958 * into the user-space timespec buffer. A value of '0' means infinity.
5959 */
5960SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5961 struct timespec __user *, interval)
5962{
5963 struct task_struct *p;
5964 unsigned int time_slice;
5965 unsigned long flags;
5966 struct rq *rq;
5967 int retval;
5968 struct timespec t;
5969
5970 if (pid < 0)
5971 return -EINVAL;
5972
5973 retval = -ESRCH;
5974 rcu_read_lock();
5975 p = find_process_by_pid(pid);
5976 if (!p)
5977 goto out_unlock;
5978
5979 retval = security_task_getscheduler(p);
5980 if (retval)
5981 goto out_unlock;
5982
5983 rq = task_rq_lock(p, &flags);
5984 time_slice = p->sched_class->get_rr_interval(rq, p);
5985 task_rq_unlock(rq, p, &flags);
5986
5987 rcu_read_unlock();
5988 jiffies_to_timespec(time_slice, &t);
5989 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5990 return retval;
5991
5992out_unlock:
5993 rcu_read_unlock();
5994 return retval;
5995}
5996
5997static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5998
5999void sched_show_task(struct task_struct *p)
6000{
6001 unsigned long free = 0;
6002 unsigned state;
6003
6004 state = p->state ? __ffs(p->state) + 1 : 0;
6005 printk(KERN_INFO "%-15.15s %c", p->comm,
6006 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6007#if BITS_PER_LONG == 32
6008 if (state == TASK_RUNNING)
6009 printk(KERN_CONT " running ");
6010 else
6011 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6012#else
6013 if (state == TASK_RUNNING)
6014 printk(KERN_CONT " running task ");
6015 else
6016 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6017#endif
6018#ifdef CONFIG_DEBUG_STACK_USAGE
6019 free = stack_not_used(p);
6020#endif
6021 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6022 task_pid_nr(p), task_pid_nr(p->real_parent),
6023 (unsigned long)task_thread_info(p)->flags);
6024
6025 show_stack(p, NULL);
6026}
6027
6028void show_state_filter(unsigned long state_filter)
6029{
6030 struct task_struct *g, *p;
6031
6032#if BITS_PER_LONG == 32
6033 printk(KERN_INFO
6034 " task PC stack pid father\n");
6035#else
6036 printk(KERN_INFO
6037 " task PC stack pid father\n");
6038#endif
6039 rcu_read_lock();
6040 do_each_thread(g, p) {
6041 /*
6042 * reset the NMI-timeout, listing all files on a slow
6043 * console might take a lot of time:
6044 */
6045 touch_nmi_watchdog();
6046 if (!state_filter || (p->state & state_filter))
6047 sched_show_task(p);
6048 } while_each_thread(g, p);
6049
6050 touch_all_softlockup_watchdogs();
6051
6052#ifdef CONFIG_SCHED_DEBUG
6053 sysrq_sched_debug_show();
6054#endif
6055 rcu_read_unlock();
6056 /*
6057 * Only show locks if all tasks are dumped:
6058 */
6059 if (!state_filter)
6060 debug_show_all_locks();
6061}
6062
6063void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6064{
6065 idle->sched_class = &idle_sched_class;
6066}
6067
6068/**
6069 * init_idle - set up an idle thread for a given CPU
6070 * @idle: task in question
6071 * @cpu: cpu the idle task belongs to
6072 *
6073 * NOTE: this function does not set the idle thread's NEED_RESCHED
6074 * flag, to make booting more robust.
6075 */
6076void __cpuinit init_idle(struct task_struct *idle, int cpu)
6077{
6078 struct rq *rq = cpu_rq(cpu);
6079 unsigned long flags;
6080
6081 raw_spin_lock_irqsave(&rq->lock, flags);
6082
6083 __sched_fork(idle);
6084 idle->state = TASK_RUNNING;
6085 idle->se.exec_start = sched_clock();
6086
6087 do_set_cpus_allowed(idle, cpumask_of(cpu));
6088 /*
6089 * We're having a chicken and egg problem, even though we are
6090 * holding rq->lock, the cpu isn't yet set to this cpu so the
6091 * lockdep check in task_group() will fail.
6092 *
6093 * Similar case to sched_fork(). / Alternatively we could
6094 * use task_rq_lock() here and obtain the other rq->lock.
6095 *
6096 * Silence PROVE_RCU
6097 */
6098 rcu_read_lock();
6099 __set_task_cpu(idle, cpu);
6100 rcu_read_unlock();
6101
6102 rq->curr = rq->idle = idle;
6103#if defined(CONFIG_SMP)
6104 idle->on_cpu = 1;
6105#endif
6106 raw_spin_unlock_irqrestore(&rq->lock, flags);
6107
6108 /* Set the preempt count _outside_ the spinlocks! */
6109 task_thread_info(idle)->preempt_count = 0;
6110
6111 /*
6112 * The idle tasks have their own, simple scheduling class:
6113 */
6114 idle->sched_class = &idle_sched_class;
6115 ftrace_graph_init_idle_task(idle, cpu);
6116#if defined(CONFIG_SMP)
6117 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
6118#endif
6119}
6120
6121/*
6122 * Increase the granularity value when there are more CPUs,
6123 * because with more CPUs the 'effective latency' as visible
6124 * to users decreases. But the relationship is not linear,
6125 * so pick a second-best guess by going with the log2 of the
6126 * number of CPUs.
6127 *
6128 * This idea comes from the SD scheduler of Con Kolivas:
6129 */
6130static int get_update_sysctl_factor(void)
6131{
6132 unsigned int cpus = min_t(int, num_online_cpus(), 8);
6133 unsigned int factor;
6134
6135 switch (sysctl_sched_tunable_scaling) {
6136 case SCHED_TUNABLESCALING_NONE:
6137 factor = 1;
6138 break;
6139 case SCHED_TUNABLESCALING_LINEAR:
6140 factor = cpus;
6141 break;
6142 case SCHED_TUNABLESCALING_LOG:
6143 default:
6144 factor = 1 + ilog2(cpus);
6145 break;
6146 }
6147
6148 return factor;
6149}
6150
6151static void update_sysctl(void)
6152{
6153 unsigned int factor = get_update_sysctl_factor();
6154
6155#define SET_SYSCTL(name) \
6156 (sysctl_##name = (factor) * normalized_sysctl_##name)
6157 SET_SYSCTL(sched_min_granularity);
6158 SET_SYSCTL(sched_latency);
6159 SET_SYSCTL(sched_wakeup_granularity);
6160#undef SET_SYSCTL
6161}
6162
6163static inline void sched_init_granularity(void)
6164{
6165 update_sysctl();
6166}
6167
6168#ifdef CONFIG_SMP
6169void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
6170{
6171 if (p->sched_class && p->sched_class->set_cpus_allowed)
6172 p->sched_class->set_cpus_allowed(p, new_mask);
6173
6174 cpumask_copy(&p->cpus_allowed, new_mask);
6175 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6176}
6177
6178/*
6179 * This is how migration works:
6180 *
6181 * 1) we invoke migration_cpu_stop() on the target CPU using
6182 * stop_one_cpu().
6183 * 2) stopper starts to run (implicitly forcing the migrated thread
6184 * off the CPU)
6185 * 3) it checks whether the migrated task is still in the wrong runqueue.
6186 * 4) if it's in the wrong runqueue then the migration thread removes
6187 * it and puts it into the right queue.
6188 * 5) stopper completes and stop_one_cpu() returns and the migration
6189 * is done.
6190 */
6191
6192/*
6193 * Change a given task's CPU affinity. Migrate the thread to a
6194 * proper CPU and schedule it away if the CPU it's executing on
6195 * is removed from the allowed bitmask.
6196 *
6197 * NOTE: the caller must have a valid reference to the task, the
6198 * task must not exit() & deallocate itself prematurely. The
6199 * call is not atomic; no spinlocks may be held.
6200 */
6201int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6202{
6203 unsigned long flags;
6204 struct rq *rq;
6205 unsigned int dest_cpu;
6206 int ret = 0;
6207
6208 rq = task_rq_lock(p, &flags);
6209
6210 if (cpumask_equal(&p->cpus_allowed, new_mask))
6211 goto out;
6212
6213 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6214 ret = -EINVAL;
6215 goto out;
6216 }
6217
6218 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6219 ret = -EINVAL;
6220 goto out;
6221 }
6222
6223 do_set_cpus_allowed(p, new_mask);
6224
6225 /* Can the task run on the task's current CPU? If so, we're done */
6226 if (cpumask_test_cpu(task_cpu(p), new_mask))
6227 goto out;
6228
6229 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6230 if (p->on_rq) {
6231 struct migration_arg arg = { p, dest_cpu };
6232 /* Need help from migration thread: drop lock and wait. */
6233 task_rq_unlock(rq, p, &flags);
6234 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6235 tlb_migrate_finish(p->mm);
6236 return 0;
6237 }
6238out:
6239 task_rq_unlock(rq, p, &flags);
6240
6241 return ret;
6242}
6243EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6244
6245/*
6246 * Move (not current) task off this cpu, onto dest cpu. We're doing
6247 * this because either it can't run here any more (set_cpus_allowed()
6248 * away from this CPU, or CPU going down), or because we're
6249 * attempting to rebalance this task on exec (sched_exec).
6250 *
6251 * So we race with normal scheduler movements, but that's OK, as long
6252 * as the task is no longer on this CPU.
6253 *
6254 * Returns non-zero if task was successfully migrated.
6255 */
6256static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6257{
6258 struct rq *rq_dest, *rq_src;
6259 int ret = 0;
6260
6261 if (unlikely(!cpu_active(dest_cpu)))
6262 return ret;
6263
6264 rq_src = cpu_rq(src_cpu);
6265 rq_dest = cpu_rq(dest_cpu);
6266
6267 raw_spin_lock(&p->pi_lock);
6268 double_rq_lock(rq_src, rq_dest);
6269 /* Already moved. */
6270 if (task_cpu(p) != src_cpu)
6271 goto done;
6272 /* Affinity changed (again). */
6273 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
6274 goto fail;
6275
6276 /*
6277 * If we're not on a rq, the next wake-up will ensure we're
6278 * placed properly.
6279 */
6280 if (p->on_rq) {
6281 deactivate_task(rq_src, p, 0);
6282 set_task_cpu(p, dest_cpu);
6283 activate_task(rq_dest, p, 0);
6284 check_preempt_curr(rq_dest, p, 0);
6285 }
6286done:
6287 ret = 1;
6288fail:
6289 double_rq_unlock(rq_src, rq_dest);
6290 raw_spin_unlock(&p->pi_lock);
6291 return ret;
6292}
6293
6294/*
6295 * migration_cpu_stop - this will be executed by a highprio stopper thread
6296 * and performs thread migration by bumping thread off CPU then
6297 * 'pushing' onto another runqueue.
6298 */
6299static int migration_cpu_stop(void *data)
6300{
6301 struct migration_arg *arg = data;
6302
6303 /*
6304 * The original target cpu might have gone down and we might
6305 * be on another cpu but it doesn't matter.
6306 */
6307 local_irq_disable();
6308 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6309 local_irq_enable();
6310 return 0;
6311}
6312
6313#ifdef CONFIG_HOTPLUG_CPU
6314
6315/*
6316 * Ensures that the idle task is using init_mm right before its cpu goes
6317 * offline.
6318 */
6319void idle_task_exit(void)
6320{
6321 struct mm_struct *mm = current->active_mm;
6322
6323 BUG_ON(cpu_online(smp_processor_id()));
6324
6325 if (mm != &init_mm)
6326 switch_mm(mm, &init_mm, current);
6327 mmdrop(mm);
6328}
6329
6330/*
6331 * While a dead CPU has no uninterruptible tasks queued at this point,
6332 * it might still have a nonzero ->nr_uninterruptible counter, because
6333 * for performance reasons the counter is not stricly tracking tasks to
6334 * their home CPUs. So we just add the counter to another CPU's counter,
6335 * to keep the global sum constant after CPU-down:
6336 */
6337static void migrate_nr_uninterruptible(struct rq *rq_src)
6338{
6339 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6340
6341 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6342 rq_src->nr_uninterruptible = 0;
6343}
6344
6345/*
6346 * remove the tasks which were accounted by rq from calc_load_tasks.
6347 */
6348static void calc_global_load_remove(struct rq *rq)
6349{
6350 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6351 rq->calc_load_active = 0;
6352}
6353
6354#ifdef CONFIG_CFS_BANDWIDTH
6355static void unthrottle_offline_cfs_rqs(struct rq *rq)
6356{
6357 struct cfs_rq *cfs_rq;
6358
6359 for_each_leaf_cfs_rq(rq, cfs_rq) {
6360 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6361
6362 if (!cfs_rq->runtime_enabled)
6363 continue;
6364
6365 /*
6366 * clock_task is not advancing so we just need to make sure
6367 * there's some valid quota amount
6368 */
6369 cfs_rq->runtime_remaining = cfs_b->quota;
6370 if (cfs_rq_throttled(cfs_rq))
6371 unthrottle_cfs_rq(cfs_rq);
6372 }
6373}
6374#else
6375static void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6376#endif
6377
6378/*
6379 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6380 * try_to_wake_up()->select_task_rq().
6381 *
6382 * Called with rq->lock held even though we'er in stop_machine() and
6383 * there's no concurrency possible, we hold the required locks anyway
6384 * because of lock validation efforts.
6385 */
6386static void migrate_tasks(unsigned int dead_cpu)
6387{
6388 struct rq *rq = cpu_rq(dead_cpu);
6389 struct task_struct *next, *stop = rq->stop;
6390 int dest_cpu;
6391
6392 /*
6393 * Fudge the rq selection such that the below task selection loop
6394 * doesn't get stuck on the currently eligible stop task.
6395 *
6396 * We're currently inside stop_machine() and the rq is either stuck
6397 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6398 * either way we should never end up calling schedule() until we're
6399 * done here.
6400 */
6401 rq->stop = NULL;
6402
6403 /* Ensure any throttled groups are reachable by pick_next_task */
6404 unthrottle_offline_cfs_rqs(rq);
6405
6406 for ( ; ; ) {
6407 /*
6408 * There's this thread running, bail when that's the only
6409 * remaining thread.
6410 */
6411 if (rq->nr_running == 1)
6412 break;
6413
6414 next = pick_next_task(rq);
6415 BUG_ON(!next);
6416 next->sched_class->put_prev_task(rq, next);
6417
6418 /* Find suitable destination for @next, with force if needed. */
6419 dest_cpu = select_fallback_rq(dead_cpu, next);
6420 raw_spin_unlock(&rq->lock);
6421
6422 __migrate_task(next, dead_cpu, dest_cpu);
6423
6424 raw_spin_lock(&rq->lock);
6425 }
6426
6427 rq->stop = stop;
6428}
6429
6430#endif /* CONFIG_HOTPLUG_CPU */
6431
6432#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6433
6434static struct ctl_table sd_ctl_dir[] = {
6435 {
6436 .procname = "sched_domain",
6437 .mode = 0555,
6438 },
6439 {}
6440};
6441
6442static struct ctl_table sd_ctl_root[] = {
6443 {
6444 .procname = "kernel",
6445 .mode = 0555,
6446 .child = sd_ctl_dir,
6447 },
6448 {}
6449};
6450
6451static struct ctl_table *sd_alloc_ctl_entry(int n)
6452{
6453 struct ctl_table *entry =
6454 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6455
6456 return entry;
6457}
6458
6459static void sd_free_ctl_entry(struct ctl_table **tablep)
6460{
6461 struct ctl_table *entry;
6462
6463 /*
6464 * In the intermediate directories, both the child directory and
6465 * procname are dynamically allocated and could fail but the mode
6466 * will always be set. In the lowest directory the names are
6467 * static strings and all have proc handlers.
6468 */
6469 for (entry = *tablep; entry->mode; entry++) {
6470 if (entry->child)
6471 sd_free_ctl_entry(&entry->child);
6472 if (entry->proc_handler == NULL)
6473 kfree(entry->procname);
6474 }
6475
6476 kfree(*tablep);
6477 *tablep = NULL;
6478}
6479
6480static void
6481set_table_entry(struct ctl_table *entry,
6482 const char *procname, void *data, int maxlen,
6483 mode_t mode, proc_handler *proc_handler)
6484{
6485 entry->procname = procname;
6486 entry->data = data;
6487 entry->maxlen = maxlen;
6488 entry->mode = mode;
6489 entry->proc_handler = proc_handler;
6490}
6491
6492static struct ctl_table *
6493sd_alloc_ctl_domain_table(struct sched_domain *sd)
6494{
6495 struct ctl_table *table = sd_alloc_ctl_entry(13);
6496
6497 if (table == NULL)
6498 return NULL;
6499
6500 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6501 sizeof(long), 0644, proc_doulongvec_minmax);
6502 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6503 sizeof(long), 0644, proc_doulongvec_minmax);
6504 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6505 sizeof(int), 0644, proc_dointvec_minmax);
6506 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6507 sizeof(int), 0644, proc_dointvec_minmax);
6508 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6509 sizeof(int), 0644, proc_dointvec_minmax);
6510 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6511 sizeof(int), 0644, proc_dointvec_minmax);
6512 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6513 sizeof(int), 0644, proc_dointvec_minmax);
6514 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6515 sizeof(int), 0644, proc_dointvec_minmax);
6516 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6517 sizeof(int), 0644, proc_dointvec_minmax);
6518 set_table_entry(&table[9], "cache_nice_tries",
6519 &sd->cache_nice_tries,
6520 sizeof(int), 0644, proc_dointvec_minmax);
6521 set_table_entry(&table[10], "flags", &sd->flags,
6522 sizeof(int), 0644, proc_dointvec_minmax);
6523 set_table_entry(&table[11], "name", sd->name,
6524 CORENAME_MAX_SIZE, 0444, proc_dostring);
6525 /* &table[12] is terminator */
6526
6527 return table;
6528}
6529
6530static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6531{
6532 struct ctl_table *entry, *table;
6533 struct sched_domain *sd;
6534 int domain_num = 0, i;
6535 char buf[32];
6536
6537 for_each_domain(cpu, sd)
6538 domain_num++;
6539 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6540 if (table == NULL)
6541 return NULL;
6542
6543 i = 0;
6544 for_each_domain(cpu, sd) {
6545 snprintf(buf, 32, "domain%d", i);
6546 entry->procname = kstrdup(buf, GFP_KERNEL);
6547 entry->mode = 0555;
6548 entry->child = sd_alloc_ctl_domain_table(sd);
6549 entry++;
6550 i++;
6551 }
6552 return table;
6553}
6554
6555static struct ctl_table_header *sd_sysctl_header;
6556static void register_sched_domain_sysctl(void)
6557{
6558 int i, cpu_num = num_possible_cpus();
6559 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6560 char buf[32];
6561
6562 WARN_ON(sd_ctl_dir[0].child);
6563 sd_ctl_dir[0].child = entry;
6564
6565 if (entry == NULL)
6566 return;
6567
6568 for_each_possible_cpu(i) {
6569 snprintf(buf, 32, "cpu%d", i);
6570 entry->procname = kstrdup(buf, GFP_KERNEL);
6571 entry->mode = 0555;
6572 entry->child = sd_alloc_ctl_cpu_table(i);
6573 entry++;
6574 }
6575
6576 WARN_ON(sd_sysctl_header);
6577 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6578}
6579
6580/* may be called multiple times per register */
6581static void unregister_sched_domain_sysctl(void)
6582{
6583 if (sd_sysctl_header)
6584 unregister_sysctl_table(sd_sysctl_header);
6585 sd_sysctl_header = NULL;
6586 if (sd_ctl_dir[0].child)
6587 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6588}
6589#else
6590static void register_sched_domain_sysctl(void)
6591{
6592}
6593static void unregister_sched_domain_sysctl(void)
6594{
6595}
6596#endif
6597
6598static void set_rq_online(struct rq *rq)
6599{
6600 if (!rq->online) {
6601 const struct sched_class *class;
6602
6603 cpumask_set_cpu(rq->cpu, rq->rd->online);
6604 rq->online = 1;
6605
6606 for_each_class(class) {
6607 if (class->rq_online)
6608 class->rq_online(rq);
6609 }
6610 }
6611}
6612
6613static void set_rq_offline(struct rq *rq)
6614{
6615 if (rq->online) {
6616 const struct sched_class *class;
6617
6618 for_each_class(class) {
6619 if (class->rq_offline)
6620 class->rq_offline(rq);
6621 }
6622
6623 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6624 rq->online = 0;
6625 }
6626}
6627
6628/*
6629 * migration_call - callback that gets triggered when a CPU is added.
6630 * Here we can start up the necessary migration thread for the new CPU.
6631 */
6632static int __cpuinit
6633migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6634{
6635 int cpu = (long)hcpu;
6636 unsigned long flags;
6637 struct rq *rq = cpu_rq(cpu);
6638
6639 switch (action & ~CPU_TASKS_FROZEN) {
6640
6641 case CPU_UP_PREPARE:
6642 rq->calc_load_update = calc_load_update;
6643 break;
6644
6645 case CPU_ONLINE:
6646 /* Update our root-domain */
6647 raw_spin_lock_irqsave(&rq->lock, flags);
6648 if (rq->rd) {
6649 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6650
6651 set_rq_online(rq);
6652 }
6653 raw_spin_unlock_irqrestore(&rq->lock, flags);
6654 break;
6655
6656#ifdef CONFIG_HOTPLUG_CPU
6657 case CPU_DYING:
6658 sched_ttwu_pending();
6659 /* Update our root-domain */
6660 raw_spin_lock_irqsave(&rq->lock, flags);
6661 if (rq->rd) {
6662 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6663 set_rq_offline(rq);
6664 }
6665 migrate_tasks(cpu);
6666 BUG_ON(rq->nr_running != 1); /* the migration thread */
6667 raw_spin_unlock_irqrestore(&rq->lock, flags);
6668
6669 migrate_nr_uninterruptible(rq);
6670 calc_global_load_remove(rq);
6671 break;
6672#endif
6673 }
6674
6675 update_max_interval();
6676
6677 return NOTIFY_OK;
6678}
6679
6680/*
6681 * Register at high priority so that task migration (migrate_all_tasks)
6682 * happens before everything else. This has to be lower priority than
6683 * the notifier in the perf_event subsystem, though.
6684 */
6685static struct notifier_block __cpuinitdata migration_notifier = {
6686 .notifier_call = migration_call,
6687 .priority = CPU_PRI_MIGRATION,
6688};
6689
6690static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6691 unsigned long action, void *hcpu)
6692{
6693 switch (action & ~CPU_TASKS_FROZEN) {
6694 case CPU_ONLINE:
6695 case CPU_DOWN_FAILED:
6696 set_cpu_active((long)hcpu, true);
6697 return NOTIFY_OK;
6698 default:
6699 return NOTIFY_DONE;
6700 }
6701}
6702
6703static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6704 unsigned long action, void *hcpu)
6705{
6706 switch (action & ~CPU_TASKS_FROZEN) {
6707 case CPU_DOWN_PREPARE:
6708 set_cpu_active((long)hcpu, false);
6709 return NOTIFY_OK;
6710 default:
6711 return NOTIFY_DONE;
6712 }
6713}
6714
6715static int __init migration_init(void)
6716{
6717 void *cpu = (void *)(long)smp_processor_id();
6718 int err;
6719
6720 /* Initialize migration for the boot CPU */
6721 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6722 BUG_ON(err == NOTIFY_BAD);
6723 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6724 register_cpu_notifier(&migration_notifier);
6725
6726 /* Register cpu active notifiers */
6727 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6728 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6729
6730 return 0;
6731}
6732early_initcall(migration_init);
6733#endif
6734
6735#ifdef CONFIG_SMP
6736
6737static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6738
6739#ifdef CONFIG_SCHED_DEBUG
6740
6741static __read_mostly int sched_domain_debug_enabled;
6742
6743static int __init sched_domain_debug_setup(char *str)
6744{
6745 sched_domain_debug_enabled = 1;
6746
6747 return 0;
6748}
6749early_param("sched_debug", sched_domain_debug_setup);
6750
6751static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6752 struct cpumask *groupmask)
6753{
6754 struct sched_group *group = sd->groups;
6755 char str[256];
6756
6757 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6758 cpumask_clear(groupmask);
6759
6760 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6761
6762 if (!(sd->flags & SD_LOAD_BALANCE)) {
6763 printk("does not load-balance\n");
6764 if (sd->parent)
6765 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6766 " has parent");
6767 return -1;
6768 }
6769
6770 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6771
6772 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6773 printk(KERN_ERR "ERROR: domain->span does not contain "
6774 "CPU%d\n", cpu);
6775 }
6776 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6777 printk(KERN_ERR "ERROR: domain->groups does not contain"
6778 " CPU%d\n", cpu);
6779 }
6780
6781 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6782 do {
6783 if (!group) {
6784 printk("\n");
6785 printk(KERN_ERR "ERROR: group is NULL\n");
6786 break;
6787 }
6788
6789 if (!group->sgp->power) {
6790 printk(KERN_CONT "\n");
6791 printk(KERN_ERR "ERROR: domain->cpu_power not "
6792 "set\n");
6793 break;
6794 }
6795
6796 if (!cpumask_weight(sched_group_cpus(group))) {
6797 printk(KERN_CONT "\n");
6798 printk(KERN_ERR "ERROR: empty group\n");
6799 break;
6800 }
6801
6802 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6803 printk(KERN_CONT "\n");
6804 printk(KERN_ERR "ERROR: repeated CPUs\n");
6805 break;
6806 }
6807
6808 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6809
6810 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6811
6812 printk(KERN_CONT " %s", str);
6813 if (group->sgp->power != SCHED_POWER_SCALE) {
6814 printk(KERN_CONT " (cpu_power = %d)",
6815 group->sgp->power);
6816 }
6817
6818 group = group->next;
6819 } while (group != sd->groups);
6820 printk(KERN_CONT "\n");
6821
6822 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6823 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6824
6825 if (sd->parent &&
6826 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6827 printk(KERN_ERR "ERROR: parent span is not a superset "
6828 "of domain->span\n");
6829 return 0;
6830}
6831
6832static void sched_domain_debug(struct sched_domain *sd, int cpu)
6833{
6834 int level = 0;
6835
6836 if (!sched_domain_debug_enabled)
6837 return;
6838
6839 if (!sd) {
6840 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6841 return;
6842 }
6843
6844 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6845
6846 for (;;) {
6847 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6848 break;
6849 level++;
6850 sd = sd->parent;
6851 if (!sd)
6852 break;
6853 }
6854}
6855#else /* !CONFIG_SCHED_DEBUG */
6856# define sched_domain_debug(sd, cpu) do { } while (0)
6857#endif /* CONFIG_SCHED_DEBUG */
6858
6859static int sd_degenerate(struct sched_domain *sd)
6860{
6861 if (cpumask_weight(sched_domain_span(sd)) == 1)
6862 return 1;
6863
6864 /* Following flags need at least 2 groups */
6865 if (sd->flags & (SD_LOAD_BALANCE |
6866 SD_BALANCE_NEWIDLE |
6867 SD_BALANCE_FORK |
6868 SD_BALANCE_EXEC |
6869 SD_SHARE_CPUPOWER |
6870 SD_SHARE_PKG_RESOURCES)) {
6871 if (sd->groups != sd->groups->next)
6872 return 0;
6873 }
6874
6875 /* Following flags don't use groups */
6876 if (sd->flags & (SD_WAKE_AFFINE))
6877 return 0;
6878
6879 return 1;
6880}
6881
6882static int
6883sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6884{
6885 unsigned long cflags = sd->flags, pflags = parent->flags;
6886
6887 if (sd_degenerate(parent))
6888 return 1;
6889
6890 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6891 return 0;
6892
6893 /* Flags needing groups don't count if only 1 group in parent */
6894 if (parent->groups == parent->groups->next) {
6895 pflags &= ~(SD_LOAD_BALANCE |
6896 SD_BALANCE_NEWIDLE |
6897 SD_BALANCE_FORK |
6898 SD_BALANCE_EXEC |
6899 SD_SHARE_CPUPOWER |
6900 SD_SHARE_PKG_RESOURCES);
6901 if (nr_node_ids == 1)
6902 pflags &= ~SD_SERIALIZE;
6903 }
6904 if (~cflags & pflags)
6905 return 0;
6906
6907 return 1;
6908}
6909
6910static void free_rootdomain(struct rcu_head *rcu)
6911{
6912 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6913
6914 cpupri_cleanup(&rd->cpupri);
6915 free_cpumask_var(rd->rto_mask);
6916 free_cpumask_var(rd->online);
6917 free_cpumask_var(rd->span);
6918 kfree(rd);
6919}
6920
6921static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6922{
6923 struct root_domain *old_rd = NULL;
6924 unsigned long flags;
6925
6926 raw_spin_lock_irqsave(&rq->lock, flags);
6927
6928 if (rq->rd) {
6929 old_rd = rq->rd;
6930
6931 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6932 set_rq_offline(rq);
6933
6934 cpumask_clear_cpu(rq->cpu, old_rd->span);
6935
6936 /*
6937 * If we dont want to free the old_rt yet then
6938 * set old_rd to NULL to skip the freeing later
6939 * in this function:
6940 */
6941 if (!atomic_dec_and_test(&old_rd->refcount))
6942 old_rd = NULL;
6943 }
6944
6945 atomic_inc(&rd->refcount);
6946 rq->rd = rd;
6947
6948 cpumask_set_cpu(rq->cpu, rd->span);
6949 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6950 set_rq_online(rq);
6951
6952 raw_spin_unlock_irqrestore(&rq->lock, flags);
6953
6954 if (old_rd)
6955 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6956}
6957
6958static int init_rootdomain(struct root_domain *rd)
6959{
6960 memset(rd, 0, sizeof(*rd));
6961
6962 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6963 goto out;
6964 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6965 goto free_span;
6966 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6967 goto free_online;
6968
6969 if (cpupri_init(&rd->cpupri) != 0)
6970 goto free_rto_mask;
6971 return 0;
6972
6973free_rto_mask:
6974 free_cpumask_var(rd->rto_mask);
6975free_online:
6976 free_cpumask_var(rd->online);
6977free_span:
6978 free_cpumask_var(rd->span);
6979out:
6980 return -ENOMEM;
6981}
6982
6983static void init_defrootdomain(void)
6984{
6985 init_rootdomain(&def_root_domain);
6986
6987 atomic_set(&def_root_domain.refcount, 1);
6988}
6989
6990static struct root_domain *alloc_rootdomain(void)
6991{
6992 struct root_domain *rd;
6993
6994 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6995 if (!rd)
6996 return NULL;
6997
6998 if (init_rootdomain(rd) != 0) {
6999 kfree(rd);
7000 return NULL;
7001 }
7002
7003 return rd;
7004}
7005
7006static void free_sched_groups(struct sched_group *sg, int free_sgp)
7007{
7008 struct sched_group *tmp, *first;
7009
7010 if (!sg)
7011 return;
7012
7013 first = sg;
7014 do {
7015 tmp = sg->next;
7016
7017 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
7018 kfree(sg->sgp);
7019
7020 kfree(sg);
7021 sg = tmp;
7022 } while (sg != first);
7023}
7024
7025static void free_sched_domain(struct rcu_head *rcu)
7026{
7027 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
7028
7029 /*
7030 * If its an overlapping domain it has private groups, iterate and
7031 * nuke them all.
7032 */
7033 if (sd->flags & SD_OVERLAP) {
7034 free_sched_groups(sd->groups, 1);
7035 } else if (atomic_dec_and_test(&sd->groups->ref)) {
7036 kfree(sd->groups->sgp);
7037 kfree(sd->groups);
7038 }
7039 kfree(sd);
7040}
7041
7042static void destroy_sched_domain(struct sched_domain *sd, int cpu)
7043{
7044 call_rcu(&sd->rcu, free_sched_domain);
7045}
7046
7047static void destroy_sched_domains(struct sched_domain *sd, int cpu)
7048{
7049 for (; sd; sd = sd->parent)
7050 destroy_sched_domain(sd, cpu);
7051}
7052
7053/*
7054 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7055 * hold the hotplug lock.
7056 */
7057static void
7058cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7059{
7060 struct rq *rq = cpu_rq(cpu);
7061 struct sched_domain *tmp;
7062
7063 /* Remove the sched domains which do not contribute to scheduling. */
7064 for (tmp = sd; tmp; ) {
7065 struct sched_domain *parent = tmp->parent;
7066 if (!parent)
7067 break;
7068
7069 if (sd_parent_degenerate(tmp, parent)) {
7070 tmp->parent = parent->parent;
7071 if (parent->parent)
7072 parent->parent->child = tmp;
7073 destroy_sched_domain(parent, cpu);
7074 } else
7075 tmp = tmp->parent;
7076 }
7077
7078 if (sd && sd_degenerate(sd)) {
7079 tmp = sd;
7080 sd = sd->parent;
7081 destroy_sched_domain(tmp, cpu);
7082 if (sd)
7083 sd->child = NULL;
7084 }
7085
7086 sched_domain_debug(sd, cpu);
7087
7088 rq_attach_root(rq, rd);
7089 tmp = rq->sd;
7090 rcu_assign_pointer(rq->sd, sd);
7091 destroy_sched_domains(tmp, cpu);
7092}
7093
7094/* cpus with isolated domains */
7095static cpumask_var_t cpu_isolated_map;
7096
7097/* Setup the mask of cpus configured for isolated domains */
7098static int __init isolated_cpu_setup(char *str)
7099{
7100 alloc_bootmem_cpumask_var(&cpu_isolated_map);
7101 cpulist_parse(str, cpu_isolated_map);
7102 return 1;
7103}
7104
7105__setup("isolcpus=", isolated_cpu_setup);
7106
7107#ifdef CONFIG_NUMA
7108
7109/**
7110 * find_next_best_node - find the next node to include in a sched_domain
7111 * @node: node whose sched_domain we're building
7112 * @used_nodes: nodes already in the sched_domain
7113 *
7114 * Find the next node to include in a given scheduling domain. Simply
7115 * finds the closest node not already in the @used_nodes map.
7116 *
7117 * Should use nodemask_t.
7118 */
7119static int find_next_best_node(int node, nodemask_t *used_nodes)
7120{
7121 int i, n, val, min_val, best_node = -1;
7122
7123 min_val = INT_MAX;
7124
7125 for (i = 0; i < nr_node_ids; i++) {
7126 /* Start at @node */
7127 n = (node + i) % nr_node_ids;
7128
7129 if (!nr_cpus_node(n))
7130 continue;
7131
7132 /* Skip already used nodes */
7133 if (node_isset(n, *used_nodes))
7134 continue;
7135
7136 /* Simple min distance search */
7137 val = node_distance(node, n);
7138
7139 if (val < min_val) {
7140 min_val = val;
7141 best_node = n;
7142 }
7143 }
7144
7145 if (best_node != -1)
7146 node_set(best_node, *used_nodes);
7147 return best_node;
7148}
7149
7150/**
7151 * sched_domain_node_span - get a cpumask for a node's sched_domain
7152 * @node: node whose cpumask we're constructing
7153 * @span: resulting cpumask
7154 *
7155 * Given a node, construct a good cpumask for its sched_domain to span. It
7156 * should be one that prevents unnecessary balancing, but also spreads tasks
7157 * out optimally.
7158 */
7159static void sched_domain_node_span(int node, struct cpumask *span)
7160{
7161 nodemask_t used_nodes;
7162 int i;
7163
7164 cpumask_clear(span);
7165 nodes_clear(used_nodes);
7166
7167 cpumask_or(span, span, cpumask_of_node(node));
7168 node_set(node, used_nodes);
7169
7170 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7171 int next_node = find_next_best_node(node, &used_nodes);
7172 if (next_node < 0)
7173 break;
7174 cpumask_or(span, span, cpumask_of_node(next_node));
7175 }
7176}
7177
7178static const struct cpumask *cpu_node_mask(int cpu)
7179{
7180 lockdep_assert_held(&sched_domains_mutex);
7181
7182 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
7183
7184 return sched_domains_tmpmask;
7185}
7186
7187static const struct cpumask *cpu_allnodes_mask(int cpu)
7188{
7189 return cpu_possible_mask;
7190}
7191#endif /* CONFIG_NUMA */
7192
7193static const struct cpumask *cpu_cpu_mask(int cpu)
7194{
7195 return cpumask_of_node(cpu_to_node(cpu));
7196}
7197
7198int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7199
7200struct sd_data {
7201 struct sched_domain **__percpu sd;
7202 struct sched_group **__percpu sg;
7203 struct sched_group_power **__percpu sgp;
7204};
7205
7206struct s_data {
7207 struct sched_domain ** __percpu sd;
7208 struct root_domain *rd;
7209};
7210
7211enum s_alloc {
7212 sa_rootdomain,
7213 sa_sd,
7214 sa_sd_storage,
7215 sa_none,
7216};
7217
7218struct sched_domain_topology_level;
7219
7220typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7221typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7222
7223#define SDTL_OVERLAP 0x01
7224
7225struct sched_domain_topology_level {
7226 sched_domain_init_f init;
7227 sched_domain_mask_f mask;
7228 int flags;
7229 struct sd_data data;
7230};
7231
7232static int
7233build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7234{
7235 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7236 const struct cpumask *span = sched_domain_span(sd);
7237 struct cpumask *covered = sched_domains_tmpmask;
7238 struct sd_data *sdd = sd->private;
7239 struct sched_domain *child;
7240 int i;
7241
7242 cpumask_clear(covered);
7243
7244 for_each_cpu(i, span) {
7245 struct cpumask *sg_span;
7246
7247 if (cpumask_test_cpu(i, covered))
7248 continue;
7249
7250 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7251 GFP_KERNEL, cpu_to_node(i));
7252
7253 if (!sg)
7254 goto fail;
7255
7256 sg_span = sched_group_cpus(sg);
7257
7258 child = *per_cpu_ptr(sdd->sd, i);
7259 if (child->child) {
7260 child = child->child;
7261 cpumask_copy(sg_span, sched_domain_span(child));
7262 } else
7263 cpumask_set_cpu(i, sg_span);
7264
7265 cpumask_or(covered, covered, sg_span);
7266
7267 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7268 atomic_inc(&sg->sgp->ref);
7269
7270 if (cpumask_test_cpu(cpu, sg_span))
7271 groups = sg;
7272
7273 if (!first)
7274 first = sg;
7275 if (last)
7276 last->next = sg;
7277 last = sg;
7278 last->next = first;
7279 }
7280 sd->groups = groups;
7281
7282 return 0;
7283
7284fail:
7285 free_sched_groups(first, 0);
7286
7287 return -ENOMEM;
7288}
7289
7290static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7291{
7292 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7293 struct sched_domain *child = sd->child;
7294
7295 if (child)
7296 cpu = cpumask_first(sched_domain_span(child));
7297
7298 if (sg) {
7299 *sg = *per_cpu_ptr(sdd->sg, cpu);
7300 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7301 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7302 }
7303
7304 return cpu;
7305}
7306
7307/*
7308 * build_sched_groups will build a circular linked list of the groups
7309 * covered by the given span, and will set each group's ->cpumask correctly,
7310 * and ->cpu_power to 0.
7311 *
7312 * Assumes the sched_domain tree is fully constructed
7313 */
7314static int
7315build_sched_groups(struct sched_domain *sd, int cpu)
7316{
7317 struct sched_group *first = NULL, *last = NULL;
7318 struct sd_data *sdd = sd->private;
7319 const struct cpumask *span = sched_domain_span(sd);
7320 struct cpumask *covered;
7321 int i;
7322
7323 get_group(cpu, sdd, &sd->groups);
7324 atomic_inc(&sd->groups->ref);
7325
7326 if (cpu != cpumask_first(sched_domain_span(sd)))
7327 return 0;
7328
7329 lockdep_assert_held(&sched_domains_mutex);
7330 covered = sched_domains_tmpmask;
7331
7332 cpumask_clear(covered);
7333
7334 for_each_cpu(i, span) {
7335 struct sched_group *sg;
7336 int group = get_group(i, sdd, &sg);
7337 int j;
7338
7339 if (cpumask_test_cpu(i, covered))
7340 continue;
7341
7342 cpumask_clear(sched_group_cpus(sg));
7343 sg->sgp->power = 0;
7344
7345 for_each_cpu(j, span) {
7346 if (get_group(j, sdd, NULL) != group)
7347 continue;
7348
7349 cpumask_set_cpu(j, covered);
7350 cpumask_set_cpu(j, sched_group_cpus(sg));
7351 }
7352
7353 if (!first)
7354 first = sg;
7355 if (last)
7356 last->next = sg;
7357 last = sg;
7358 }
7359 last->next = first;
7360
7361 return 0;
7362}
7363
7364/*
7365 * Initialize sched groups cpu_power.
7366 *
7367 * cpu_power indicates the capacity of sched group, which is used while
7368 * distributing the load between different sched groups in a sched domain.
7369 * Typically cpu_power for all the groups in a sched domain will be same unless
7370 * there are asymmetries in the topology. If there are asymmetries, group
7371 * having more cpu_power will pickup more load compared to the group having
7372 * less cpu_power.
7373 */
7374static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7375{
7376 struct sched_group *sg = sd->groups;
7377
7378 WARN_ON(!sd || !sg);
7379
7380 do {
7381 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7382 sg = sg->next;
7383 } while (sg != sd->groups);
7384
7385 if (cpu != group_first_cpu(sg))
7386 return;
7387
7388 update_group_power(sd, cpu);
7389}
7390
7391/*
7392 * Initializers for schedule domains
7393 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7394 */
7395
7396#ifdef CONFIG_SCHED_DEBUG
7397# define SD_INIT_NAME(sd, type) sd->name = #type
7398#else
7399# define SD_INIT_NAME(sd, type) do { } while (0)
7400#endif
7401
7402#define SD_INIT_FUNC(type) \
7403static noinline struct sched_domain * \
7404sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7405{ \
7406 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7407 *sd = SD_##type##_INIT; \
7408 SD_INIT_NAME(sd, type); \
7409 sd->private = &tl->data; \
7410 return sd; \
7411}
7412
7413SD_INIT_FUNC(CPU)
7414#ifdef CONFIG_NUMA
7415 SD_INIT_FUNC(ALLNODES)
7416 SD_INIT_FUNC(NODE)
7417#endif
7418#ifdef CONFIG_SCHED_SMT
7419 SD_INIT_FUNC(SIBLING)
7420#endif
7421#ifdef CONFIG_SCHED_MC
7422 SD_INIT_FUNC(MC)
7423#endif
7424#ifdef CONFIG_SCHED_BOOK
7425 SD_INIT_FUNC(BOOK)
7426#endif
7427
7428static int default_relax_domain_level = -1;
7429int sched_domain_level_max;
7430
7431static int __init setup_relax_domain_level(char *str)
7432{
7433 unsigned long val;
7434
7435 val = simple_strtoul(str, NULL, 0);
7436 if (val < sched_domain_level_max)
7437 default_relax_domain_level = val;
7438
7439 return 1;
7440}
7441__setup("relax_domain_level=", setup_relax_domain_level);
7442
7443static void set_domain_attribute(struct sched_domain *sd,
7444 struct sched_domain_attr *attr)
7445{
7446 int request;
7447
7448 if (!attr || attr->relax_domain_level < 0) {
7449 if (default_relax_domain_level < 0)
7450 return;
7451 else
7452 request = default_relax_domain_level;
7453 } else
7454 request = attr->relax_domain_level;
7455 if (request < sd->level) {
7456 /* turn off idle balance on this domain */
7457 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7458 } else {
7459 /* turn on idle balance on this domain */
7460 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7461 }
7462}
7463
7464static void __sdt_free(const struct cpumask *cpu_map);
7465static int __sdt_alloc(const struct cpumask *cpu_map);
7466
7467static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7468 const struct cpumask *cpu_map)
7469{
7470 switch (what) {
7471 case sa_rootdomain:
7472 if (!atomic_read(&d->rd->refcount))
7473 free_rootdomain(&d->rd->rcu); /* fall through */
7474 case sa_sd:
7475 free_percpu(d->sd); /* fall through */
7476 case sa_sd_storage:
7477 __sdt_free(cpu_map); /* fall through */
7478 case sa_none:
7479 break;
7480 }
7481}
7482
7483static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7484 const struct cpumask *cpu_map)
7485{
7486 memset(d, 0, sizeof(*d));
7487
7488 if (__sdt_alloc(cpu_map))
7489 return sa_sd_storage;
7490 d->sd = alloc_percpu(struct sched_domain *);
7491 if (!d->sd)
7492 return sa_sd_storage;
7493 d->rd = alloc_rootdomain();
7494 if (!d->rd)
7495 return sa_sd;
7496 return sa_rootdomain;
7497}
7498
7499/*
7500 * NULL the sd_data elements we've used to build the sched_domain and
7501 * sched_group structure so that the subsequent __free_domain_allocs()
7502 * will not free the data we're using.
7503 */
7504static void claim_allocations(int cpu, struct sched_domain *sd)
7505{
7506 struct sd_data *sdd = sd->private;
7507
7508 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7509 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7510
7511 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7512 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7513
7514 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7515 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7516}
7517
7518#ifdef CONFIG_SCHED_SMT
7519static const struct cpumask *cpu_smt_mask(int cpu)
7520{
7521 return topology_thread_cpumask(cpu);
7522}
7523#endif
7524
7525/*
7526 * Topology list, bottom-up.
7527 */
7528static struct sched_domain_topology_level default_topology[] = {
7529#ifdef CONFIG_SCHED_SMT
7530 { sd_init_SIBLING, cpu_smt_mask, },
7531#endif
7532#ifdef CONFIG_SCHED_MC
7533 { sd_init_MC, cpu_coregroup_mask, },
7534#endif
7535#ifdef CONFIG_SCHED_BOOK
7536 { sd_init_BOOK, cpu_book_mask, },
7537#endif
7538 { sd_init_CPU, cpu_cpu_mask, },
7539#ifdef CONFIG_NUMA
7540 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7541 { sd_init_ALLNODES, cpu_allnodes_mask, },
7542#endif
7543 { NULL, },
7544};
7545
7546static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7547
7548static int __sdt_alloc(const struct cpumask *cpu_map)
7549{
7550 struct sched_domain_topology_level *tl;
7551 int j;
7552
7553 for (tl = sched_domain_topology; tl->init; tl++) {
7554 struct sd_data *sdd = &tl->data;
7555
7556 sdd->sd = alloc_percpu(struct sched_domain *);
7557 if (!sdd->sd)
7558 return -ENOMEM;
7559
7560 sdd->sg = alloc_percpu(struct sched_group *);
7561 if (!sdd->sg)
7562 return -ENOMEM;
7563
7564 sdd->sgp = alloc_percpu(struct sched_group_power *);
7565 if (!sdd->sgp)
7566 return -ENOMEM;
7567
7568 for_each_cpu(j, cpu_map) {
7569 struct sched_domain *sd;
7570 struct sched_group *sg;
7571 struct sched_group_power *sgp;
7572
7573 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7574 GFP_KERNEL, cpu_to_node(j));
7575 if (!sd)
7576 return -ENOMEM;
7577
7578 *per_cpu_ptr(sdd->sd, j) = sd;
7579
7580 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7581 GFP_KERNEL, cpu_to_node(j));
7582 if (!sg)
7583 return -ENOMEM;
7584
7585 *per_cpu_ptr(sdd->sg, j) = sg;
7586
7587 sgp = kzalloc_node(sizeof(struct sched_group_power),
7588 GFP_KERNEL, cpu_to_node(j));
7589 if (!sgp)
7590 return -ENOMEM;
7591
7592 *per_cpu_ptr(sdd->sgp, j) = sgp;
7593 }
7594 }
7595
7596 return 0;
7597}
7598
7599static void __sdt_free(const struct cpumask *cpu_map)
7600{
7601 struct sched_domain_topology_level *tl;
7602 int j;
7603
7604 for (tl = sched_domain_topology; tl->init; tl++) {
7605 struct sd_data *sdd = &tl->data;
7606
7607 for_each_cpu(j, cpu_map) {
7608 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
7609 if (sd && (sd->flags & SD_OVERLAP))
7610 free_sched_groups(sd->groups, 0);
7611 kfree(*per_cpu_ptr(sdd->sd, j));
7612 kfree(*per_cpu_ptr(sdd->sg, j));
7613 kfree(*per_cpu_ptr(sdd->sgp, j));
7614 }
7615 free_percpu(sdd->sd);
7616 free_percpu(sdd->sg);
7617 free_percpu(sdd->sgp);
7618 }
7619}
7620
7621struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7622 struct s_data *d, const struct cpumask *cpu_map,
7623 struct sched_domain_attr *attr, struct sched_domain *child,
7624 int cpu)
7625{
7626 struct sched_domain *sd = tl->init(tl, cpu);
7627 if (!sd)
7628 return child;
7629
7630 set_domain_attribute(sd, attr);
7631 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7632 if (child) {
7633 sd->level = child->level + 1;
7634 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7635 child->parent = sd;
7636 }
7637 sd->child = child;
7638
7639 return sd;
7640}
7641
7642/*
7643 * Build sched domains for a given set of cpus and attach the sched domains
7644 * to the individual cpus
7645 */
7646static int build_sched_domains(const struct cpumask *cpu_map,
7647 struct sched_domain_attr *attr)
7648{
7649 enum s_alloc alloc_state = sa_none;
7650 struct sched_domain *sd;
7651 struct s_data d;
7652 int i, ret = -ENOMEM;
7653
7654 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7655 if (alloc_state != sa_rootdomain)
7656 goto error;
7657
7658 /* Set up domains for cpus specified by the cpu_map. */
7659 for_each_cpu(i, cpu_map) {
7660 struct sched_domain_topology_level *tl;
7661
7662 sd = NULL;
7663 for (tl = sched_domain_topology; tl->init; tl++) {
7664 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7665 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7666 sd->flags |= SD_OVERLAP;
7667 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7668 break;
7669 }
7670
7671 while (sd->child)
7672 sd = sd->child;
7673
7674 *per_cpu_ptr(d.sd, i) = sd;
7675 }
7676
7677 /* Build the groups for the domains */
7678 for_each_cpu(i, cpu_map) {
7679 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7680 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7681 if (sd->flags & SD_OVERLAP) {
7682 if (build_overlap_sched_groups(sd, i))
7683 goto error;
7684 } else {
7685 if (build_sched_groups(sd, i))
7686 goto error;
7687 }
7688 }
7689 }
7690
7691 /* Calculate CPU power for physical packages and nodes */
7692 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7693 if (!cpumask_test_cpu(i, cpu_map))
7694 continue;
7695
7696 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7697 claim_allocations(i, sd);
7698 init_sched_groups_power(i, sd);
7699 }
7700 }
7701
7702 /* Attach the domains */
7703 rcu_read_lock();
7704 for_each_cpu(i, cpu_map) {
7705 sd = *per_cpu_ptr(d.sd, i);
7706 cpu_attach_domain(sd, d.rd, i);
7707 }
7708 rcu_read_unlock();
7709
7710 ret = 0;
7711error:
7712 __free_domain_allocs(&d, alloc_state, cpu_map);
7713 return ret;
7714}
7715
7716static cpumask_var_t *doms_cur; /* current sched domains */
7717static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7718static struct sched_domain_attr *dattr_cur;
7719 /* attribues of custom domains in 'doms_cur' */
7720
7721/*
7722 * Special case: If a kmalloc of a doms_cur partition (array of
7723 * cpumask) fails, then fallback to a single sched domain,
7724 * as determined by the single cpumask fallback_doms.
7725 */
7726static cpumask_var_t fallback_doms;
7727
7728/*
7729 * arch_update_cpu_topology lets virtualized architectures update the
7730 * cpu core maps. It is supposed to return 1 if the topology changed
7731 * or 0 if it stayed the same.
7732 */
7733int __attribute__((weak)) arch_update_cpu_topology(void)
7734{
7735 return 0;
7736}
7737
7738cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7739{
7740 int i;
7741 cpumask_var_t *doms;
7742
7743 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7744 if (!doms)
7745 return NULL;
7746 for (i = 0; i < ndoms; i++) {
7747 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7748 free_sched_domains(doms, i);
7749 return NULL;
7750 }
7751 }
7752 return doms;
7753}
7754
7755void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7756{
7757 unsigned int i;
7758 for (i = 0; i < ndoms; i++)
7759 free_cpumask_var(doms[i]);
7760 kfree(doms);
7761}
7762
7763/*
7764 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7765 * For now this just excludes isolated cpus, but could be used to
7766 * exclude other special cases in the future.
7767 */
7768static int init_sched_domains(const struct cpumask *cpu_map)
7769{
7770 int err;
7771
7772 arch_update_cpu_topology();
7773 ndoms_cur = 1;
7774 doms_cur = alloc_sched_domains(ndoms_cur);
7775 if (!doms_cur)
7776 doms_cur = &fallback_doms;
7777 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7778 dattr_cur = NULL;
7779 err = build_sched_domains(doms_cur[0], NULL);
7780 register_sched_domain_sysctl();
7781
7782 return err;
7783}
7784
7785/*
7786 * Detach sched domains from a group of cpus specified in cpu_map
7787 * These cpus will now be attached to the NULL domain
7788 */
7789static void detach_destroy_domains(const struct cpumask *cpu_map)
7790{
7791 int i;
7792
7793 rcu_read_lock();
7794 for_each_cpu(i, cpu_map)
7795 cpu_attach_domain(NULL, &def_root_domain, i);
7796 rcu_read_unlock();
7797}
7798
7799/* handle null as "default" */
7800static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7801 struct sched_domain_attr *new, int idx_new)
7802{
7803 struct sched_domain_attr tmp;
7804
7805 /* fast path */
7806 if (!new && !cur)
7807 return 1;
7808
7809 tmp = SD_ATTR_INIT;
7810 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7811 new ? (new + idx_new) : &tmp,
7812 sizeof(struct sched_domain_attr));
7813}
7814
7815/*
7816 * Partition sched domains as specified by the 'ndoms_new'
7817 * cpumasks in the array doms_new[] of cpumasks. This compares
7818 * doms_new[] to the current sched domain partitioning, doms_cur[].
7819 * It destroys each deleted domain and builds each new domain.
7820 *
7821 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7822 * The masks don't intersect (don't overlap.) We should setup one
7823 * sched domain for each mask. CPUs not in any of the cpumasks will
7824 * not be load balanced. If the same cpumask appears both in the
7825 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7826 * it as it is.
7827 *
7828 * The passed in 'doms_new' should be allocated using
7829 * alloc_sched_domains. This routine takes ownership of it and will
7830 * free_sched_domains it when done with it. If the caller failed the
7831 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7832 * and partition_sched_domains() will fallback to the single partition
7833 * 'fallback_doms', it also forces the domains to be rebuilt.
7834 *
7835 * If doms_new == NULL it will be replaced with cpu_online_mask.
7836 * ndoms_new == 0 is a special case for destroying existing domains,
7837 * and it will not create the default domain.
7838 *
7839 * Call with hotplug lock held
7840 */
7841void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7842 struct sched_domain_attr *dattr_new)
7843{
7844 int i, j, n;
7845 int new_topology;
7846
7847 mutex_lock(&sched_domains_mutex);
7848
7849 /* always unregister in case we don't destroy any domains */
7850 unregister_sched_domain_sysctl();
7851
7852 /* Let architecture update cpu core mappings. */
7853 new_topology = arch_update_cpu_topology();
7854
7855 n = doms_new ? ndoms_new : 0;
7856
7857 /* Destroy deleted domains */
7858 for (i = 0; i < ndoms_cur; i++) {
7859 for (j = 0; j < n && !new_topology; j++) {
7860 if (cpumask_equal(doms_cur[i], doms_new[j])
7861 && dattrs_equal(dattr_cur, i, dattr_new, j))
7862 goto match1;
7863 }
7864 /* no match - a current sched domain not in new doms_new[] */
7865 detach_destroy_domains(doms_cur[i]);
7866match1:
7867 ;
7868 }
7869
7870 if (doms_new == NULL) {
7871 ndoms_cur = 0;
7872 doms_new = &fallback_doms;
7873 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7874 WARN_ON_ONCE(dattr_new);
7875 }
7876
7877 /* Build new domains */
7878 for (i = 0; i < ndoms_new; i++) {
7879 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7880 if (cpumask_equal(doms_new[i], doms_cur[j])
7881 && dattrs_equal(dattr_new, i, dattr_cur, j))
7882 goto match2;
7883 }
7884 /* no match - add a new doms_new */
7885 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7886match2:
7887 ;
7888 }
7889
7890 /* Remember the new sched domains */
7891 if (doms_cur != &fallback_doms)
7892 free_sched_domains(doms_cur, ndoms_cur);
7893 kfree(dattr_cur); /* kfree(NULL) is safe */
7894 doms_cur = doms_new;
7895 dattr_cur = dattr_new;
7896 ndoms_cur = ndoms_new;
7897
7898 register_sched_domain_sysctl();
7899
7900 mutex_unlock(&sched_domains_mutex);
7901}
7902
7903#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7904static void reinit_sched_domains(void)
7905{
7906 get_online_cpus();
7907
7908 /* Destroy domains first to force the rebuild */
7909 partition_sched_domains(0, NULL, NULL);
7910
7911 rebuild_sched_domains();
7912 put_online_cpus();
7913}
7914
7915static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7916{
7917 unsigned int level = 0;
7918
7919 if (sscanf(buf, "%u", &level) != 1)
7920 return -EINVAL;
7921
7922 /*
7923 * level is always be positive so don't check for
7924 * level < POWERSAVINGS_BALANCE_NONE which is 0
7925 * What happens on 0 or 1 byte write,
7926 * need to check for count as well?
7927 */
7928
7929 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7930 return -EINVAL;
7931
7932 if (smt)
7933 sched_smt_power_savings = level;
7934 else
7935 sched_mc_power_savings = level;
7936
7937 reinit_sched_domains();
7938
7939 return count;
7940}
7941
7942#ifdef CONFIG_SCHED_MC
7943static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7944 struct sysdev_class_attribute *attr,
7945 char *page)
7946{
7947 return sprintf(page, "%u\n", sched_mc_power_savings);
7948}
7949static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7950 struct sysdev_class_attribute *attr,
7951 const char *buf, size_t count)
7952{
7953 return sched_power_savings_store(buf, count, 0);
7954}
7955static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7956 sched_mc_power_savings_show,
7957 sched_mc_power_savings_store);
7958#endif
7959
7960#ifdef CONFIG_SCHED_SMT
7961static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7962 struct sysdev_class_attribute *attr,
7963 char *page)
7964{
7965 return sprintf(page, "%u\n", sched_smt_power_savings);
7966}
7967static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7968 struct sysdev_class_attribute *attr,
7969 const char *buf, size_t count)
7970{
7971 return sched_power_savings_store(buf, count, 1);
7972}
7973static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7974 sched_smt_power_savings_show,
7975 sched_smt_power_savings_store);
7976#endif
7977
7978int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7979{
7980 int err = 0;
7981
7982#ifdef CONFIG_SCHED_SMT
7983 if (smt_capable())
7984 err = sysfs_create_file(&cls->kset.kobj,
7985 &attr_sched_smt_power_savings.attr);
7986#endif
7987#ifdef CONFIG_SCHED_MC
7988 if (!err && mc_capable())
7989 err = sysfs_create_file(&cls->kset.kobj,
7990 &attr_sched_mc_power_savings.attr);
7991#endif
7992 return err;
7993}
7994#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7995
7996/*
7997 * Update cpusets according to cpu_active mask. If cpusets are
7998 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7999 * around partition_sched_domains().
8000 */
8001static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
8002 void *hcpu)
8003{
8004 switch (action & ~CPU_TASKS_FROZEN) {
8005 case CPU_ONLINE:
8006 case CPU_DOWN_FAILED:
8007 cpuset_update_active_cpus();
8008 return NOTIFY_OK;
8009 default:
8010 return NOTIFY_DONE;
8011 }
8012}
8013
8014static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
8015 void *hcpu)
8016{
8017 switch (action & ~CPU_TASKS_FROZEN) {
8018 case CPU_DOWN_PREPARE:
8019 cpuset_update_active_cpus();
8020 return NOTIFY_OK;
8021 default:
8022 return NOTIFY_DONE;
8023 }
8024}
8025
8026static int update_runtime(struct notifier_block *nfb,
8027 unsigned long action, void *hcpu)
8028{
8029 int cpu = (int)(long)hcpu;
8030
8031 switch (action) {
8032 case CPU_DOWN_PREPARE:
8033 case CPU_DOWN_PREPARE_FROZEN:
8034 disable_runtime(cpu_rq(cpu));
8035 return NOTIFY_OK;
8036
8037 case CPU_DOWN_FAILED:
8038 case CPU_DOWN_FAILED_FROZEN:
8039 case CPU_ONLINE:
8040 case CPU_ONLINE_FROZEN:
8041 enable_runtime(cpu_rq(cpu));
8042 return NOTIFY_OK;
8043
8044 default:
8045 return NOTIFY_DONE;
8046 }
8047}
8048
8049void __init sched_init_smp(void)
8050{
8051 cpumask_var_t non_isolated_cpus;
8052
8053 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8054 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8055
8056 get_online_cpus();
8057 mutex_lock(&sched_domains_mutex);
8058 init_sched_domains(cpu_active_mask);
8059 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8060 if (cpumask_empty(non_isolated_cpus))
8061 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8062 mutex_unlock(&sched_domains_mutex);
8063 put_online_cpus();
8064
8065 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
8066 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
8067
8068 /* RT runtime code needs to handle some hotplug events */
8069 hotcpu_notifier(update_runtime, 0);
8070
8071 init_hrtick();
8072
8073 /* Move init over to a non-isolated CPU */
8074 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8075 BUG();
8076 sched_init_granularity();
8077 free_cpumask_var(non_isolated_cpus);
8078
8079 init_sched_rt_class();
8080}
8081#else
8082void __init sched_init_smp(void)
8083{
8084 sched_init_granularity();
8085}
8086#endif /* CONFIG_SMP */
8087
8088const_debug unsigned int sysctl_timer_migration = 1;
8089
8090int in_sched_functions(unsigned long addr)
8091{
8092 return in_lock_functions(addr) ||
8093 (addr >= (unsigned long)__sched_text_start
8094 && addr < (unsigned long)__sched_text_end);
8095}
8096
8097static void init_cfs_rq(struct cfs_rq *cfs_rq)
8098{
8099 cfs_rq->tasks_timeline = RB_ROOT;
8100 INIT_LIST_HEAD(&cfs_rq->tasks);
8101 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8102#ifndef CONFIG_64BIT
8103 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8104#endif
8105}
8106
8107static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8108{
8109 struct rt_prio_array *array;
8110 int i;
8111
8112 array = &rt_rq->active;
8113 for (i = 0; i < MAX_RT_PRIO; i++) {
8114 INIT_LIST_HEAD(array->queue + i);
8115 __clear_bit(i, array->bitmap);
8116 }
8117 /* delimiter for bitsearch: */
8118 __set_bit(MAX_RT_PRIO, array->bitmap);
8119
8120#if defined CONFIG_SMP
8121 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8122 rt_rq->highest_prio.next = MAX_RT_PRIO;
8123 rt_rq->rt_nr_migratory = 0;
8124 rt_rq->overloaded = 0;
8125 plist_head_init(&rt_rq->pushable_tasks);
8126#endif
8127
8128 rt_rq->rt_time = 0;
8129 rt_rq->rt_throttled = 0;
8130 rt_rq->rt_runtime = 0;
8131 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8132}
8133
8134#ifdef CONFIG_FAIR_GROUP_SCHED
8135static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8136 struct sched_entity *se, int cpu,
8137 struct sched_entity *parent)
8138{
8139 struct rq *rq = cpu_rq(cpu);
8140
8141 cfs_rq->tg = tg;
8142 cfs_rq->rq = rq;
8143#ifdef CONFIG_SMP
8144 /* allow initial update_cfs_load() to truncate */
8145 cfs_rq->load_stamp = 1;
8146#endif
8147 init_cfs_rq_runtime(cfs_rq);
8148
8149 tg->cfs_rq[cpu] = cfs_rq;
8150 tg->se[cpu] = se;
8151
8152 /* se could be NULL for root_task_group */
8153 if (!se)
8154 return;
8155
8156 if (!parent)
8157 se->cfs_rq = &rq->cfs;
8158 else
8159 se->cfs_rq = parent->my_q;
8160
8161 se->my_q = cfs_rq;
8162 update_load_set(&se->load, 0);
8163 se->parent = parent;
8164}
8165#endif
8166
8167#ifdef CONFIG_RT_GROUP_SCHED
8168static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8169 struct sched_rt_entity *rt_se, int cpu,
8170 struct sched_rt_entity *parent)
8171{
8172 struct rq *rq = cpu_rq(cpu);
8173
8174 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8175 rt_rq->rt_nr_boosted = 0;
8176 rt_rq->rq = rq;
8177 rt_rq->tg = tg;
8178
8179 tg->rt_rq[cpu] = rt_rq;
8180 tg->rt_se[cpu] = rt_se;
8181
8182 if (!rt_se)
8183 return;
8184
8185 if (!parent)
8186 rt_se->rt_rq = &rq->rt;
8187 else
8188 rt_se->rt_rq = parent->my_q;
8189
8190 rt_se->my_q = rt_rq;
8191 rt_se->parent = parent;
8192 INIT_LIST_HEAD(&rt_se->run_list);
8193}
8194#endif
8195
8196void __init sched_init(void)
8197{
8198 int i, j;
8199 unsigned long alloc_size = 0, ptr;
8200
8201#ifdef CONFIG_FAIR_GROUP_SCHED
8202 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8203#endif
8204#ifdef CONFIG_RT_GROUP_SCHED
8205 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8206#endif
8207#ifdef CONFIG_CPUMASK_OFFSTACK
8208 alloc_size += num_possible_cpus() * cpumask_size();
8209#endif
8210 if (alloc_size) {
8211 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8212
8213#ifdef CONFIG_FAIR_GROUP_SCHED
8214 root_task_group.se = (struct sched_entity **)ptr;
8215 ptr += nr_cpu_ids * sizeof(void **);
8216
8217 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8218 ptr += nr_cpu_ids * sizeof(void **);
8219
8220#endif /* CONFIG_FAIR_GROUP_SCHED */
8221#ifdef CONFIG_RT_GROUP_SCHED
8222 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8223 ptr += nr_cpu_ids * sizeof(void **);
8224
8225 root_task_group.rt_rq = (struct rt_rq **)ptr;
8226 ptr += nr_cpu_ids * sizeof(void **);
8227
8228#endif /* CONFIG_RT_GROUP_SCHED */
8229#ifdef CONFIG_CPUMASK_OFFSTACK
8230 for_each_possible_cpu(i) {
8231 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8232 ptr += cpumask_size();
8233 }
8234#endif /* CONFIG_CPUMASK_OFFSTACK */
8235 }
8236
8237#ifdef CONFIG_SMP
8238 init_defrootdomain();
8239#endif
8240
8241 init_rt_bandwidth(&def_rt_bandwidth,
8242 global_rt_period(), global_rt_runtime());
8243
8244#ifdef CONFIG_RT_GROUP_SCHED
8245 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8246 global_rt_period(), global_rt_runtime());
8247#endif /* CONFIG_RT_GROUP_SCHED */
8248
8249#ifdef CONFIG_CGROUP_SCHED
8250 list_add(&root_task_group.list, &task_groups);
8251 INIT_LIST_HEAD(&root_task_group.children);
8252 autogroup_init(&init_task);
8253#endif /* CONFIG_CGROUP_SCHED */
8254
8255 for_each_possible_cpu(i) {
8256 struct rq *rq;
8257
8258 rq = cpu_rq(i);
8259 raw_spin_lock_init(&rq->lock);
8260 rq->nr_running = 0;
8261 rq->calc_load_active = 0;
8262 rq->calc_load_update = jiffies + LOAD_FREQ;
8263 init_cfs_rq(&rq->cfs);
8264 init_rt_rq(&rq->rt, rq);
8265#ifdef CONFIG_FAIR_GROUP_SCHED
8266 root_task_group.shares = root_task_group_load;
8267 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8268 /*
8269 * How much cpu bandwidth does root_task_group get?
8270 *
8271 * In case of task-groups formed thr' the cgroup filesystem, it
8272 * gets 100% of the cpu resources in the system. This overall
8273 * system cpu resource is divided among the tasks of
8274 * root_task_group and its child task-groups in a fair manner,
8275 * based on each entity's (task or task-group's) weight
8276 * (se->load.weight).
8277 *
8278 * In other words, if root_task_group has 10 tasks of weight
8279 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8280 * then A0's share of the cpu resource is:
8281 *
8282 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8283 *
8284 * We achieve this by letting root_task_group's tasks sit
8285 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8286 */
8287 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
8288 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8289#endif /* CONFIG_FAIR_GROUP_SCHED */
8290
8291 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8292#ifdef CONFIG_RT_GROUP_SCHED
8293 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8294 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8295#endif
8296
8297 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8298 rq->cpu_load[j] = 0;
8299
8300 rq->last_load_update_tick = jiffies;
8301
8302#ifdef CONFIG_SMP
8303 rq->sd = NULL;
8304 rq->rd = NULL;
8305 rq->cpu_power = SCHED_POWER_SCALE;
8306 rq->post_schedule = 0;
8307 rq->active_balance = 0;
8308 rq->next_balance = jiffies;
8309 rq->push_cpu = 0;
8310 rq->cpu = i;
8311 rq->online = 0;
8312 rq->idle_stamp = 0;
8313 rq->avg_idle = 2*sysctl_sched_migration_cost;
8314 rq_attach_root(rq, &def_root_domain);
8315#ifdef CONFIG_NO_HZ
8316 rq->nohz_balance_kick = 0;
8317#endif
8318#endif
8319 init_rq_hrtick(rq);
8320 atomic_set(&rq->nr_iowait, 0);
8321 }
8322
8323 set_load_weight(&init_task);
8324
8325#ifdef CONFIG_PREEMPT_NOTIFIERS
8326 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8327#endif
8328
8329#ifdef CONFIG_SMP
8330 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8331#endif
8332
8333#ifdef CONFIG_RT_MUTEXES
8334 plist_head_init(&init_task.pi_waiters);
8335#endif
8336
8337 /*
8338 * The boot idle thread does lazy MMU switching as well:
8339 */
8340 atomic_inc(&init_mm.mm_count);
8341 enter_lazy_tlb(&init_mm, current);
8342
8343 /*
8344 * Make us the idle thread. Technically, schedule() should not be
8345 * called from this thread, however somewhere below it might be,
8346 * but because we are the idle thread, we just pick up running again
8347 * when this runqueue becomes "idle".
8348 */
8349 init_idle(current, smp_processor_id());
8350
8351 calc_load_update = jiffies + LOAD_FREQ;
8352
8353 /*
8354 * During early bootup we pretend to be a normal task:
8355 */
8356 current->sched_class = &fair_sched_class;
8357
8358#ifdef CONFIG_SMP
8359 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8360#ifdef CONFIG_NO_HZ
8361 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8362 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8363 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8364 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8365 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8366#endif
8367 /* May be allocated at isolcpus cmdline parse time */
8368 if (cpu_isolated_map == NULL)
8369 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8370#endif /* SMP */
8371
8372 scheduler_running = 1;
8373}
8374
8375#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8376static inline int preempt_count_equals(int preempt_offset)
8377{
8378 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8379
8380 return (nested == preempt_offset);
8381}
8382
8383void __might_sleep(const char *file, int line, int preempt_offset)
8384{
8385 static unsigned long prev_jiffy; /* ratelimiting */
8386
8387 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
8388 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8389 system_state != SYSTEM_RUNNING || oops_in_progress)
8390 return;
8391 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8392 return;
8393 prev_jiffy = jiffies;
8394
8395 printk(KERN_ERR
8396 "BUG: sleeping function called from invalid context at %s:%d\n",
8397 file, line);
8398 printk(KERN_ERR
8399 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8400 in_atomic(), irqs_disabled(),
8401 current->pid, current->comm);
8402
8403 debug_show_held_locks(current);
8404 if (irqs_disabled())
8405 print_irqtrace_events(current);
8406 dump_stack();
8407}
8408EXPORT_SYMBOL(__might_sleep);
8409#endif
8410
8411#ifdef CONFIG_MAGIC_SYSRQ
8412static void normalize_task(struct rq *rq, struct task_struct *p)
8413{
8414 const struct sched_class *prev_class = p->sched_class;
8415 int old_prio = p->prio;
8416 int on_rq;
8417
8418 on_rq = p->on_rq;
8419 if (on_rq)
8420 deactivate_task(rq, p, 0);
8421 __setscheduler(rq, p, SCHED_NORMAL, 0);
8422 if (on_rq) {
8423 activate_task(rq, p, 0);
8424 resched_task(rq->curr);
8425 }
8426
8427 check_class_changed(rq, p, prev_class, old_prio);
8428}
8429
8430void normalize_rt_tasks(void)
8431{
8432 struct task_struct *g, *p;
8433 unsigned long flags;
8434 struct rq *rq;
8435
8436 read_lock_irqsave(&tasklist_lock, flags);
8437 do_each_thread(g, p) {
8438 /*
8439 * Only normalize user tasks:
8440 */
8441 if (!p->mm)
8442 continue;
8443
8444 p->se.exec_start = 0;
8445#ifdef CONFIG_SCHEDSTATS
8446 p->se.statistics.wait_start = 0;
8447 p->se.statistics.sleep_start = 0;
8448 p->se.statistics.block_start = 0;
8449#endif
8450
8451 if (!rt_task(p)) {
8452 /*
8453 * Renice negative nice level userspace
8454 * tasks back to 0:
8455 */
8456 if (TASK_NICE(p) < 0 && p->mm)
8457 set_user_nice(p, 0);
8458 continue;
8459 }
8460
8461 raw_spin_lock(&p->pi_lock);
8462 rq = __task_rq_lock(p);
8463
8464 normalize_task(rq, p);
8465
8466 __task_rq_unlock(rq);
8467 raw_spin_unlock(&p->pi_lock);
8468 } while_each_thread(g, p);
8469
8470 read_unlock_irqrestore(&tasklist_lock, flags);
8471}
8472
8473#endif /* CONFIG_MAGIC_SYSRQ */
8474
8475#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8476/*
8477 * These functions are only useful for the IA64 MCA handling, or kdb.
8478 *
8479 * They can only be called when the whole system has been
8480 * stopped - every CPU needs to be quiescent, and no scheduling
8481 * activity can take place. Using them for anything else would
8482 * be a serious bug, and as a result, they aren't even visible
8483 * under any other configuration.
8484 */
8485
8486/**
8487 * curr_task - return the current task for a given cpu.
8488 * @cpu: the processor in question.
8489 *
8490 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8491 */
8492struct task_struct *curr_task(int cpu)
8493{
8494 return cpu_curr(cpu);
8495}
8496
8497#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8498
8499#ifdef CONFIG_IA64
8500/**
8501 * set_curr_task - set the current task for a given cpu.
8502 * @cpu: the processor in question.
8503 * @p: the task pointer to set.
8504 *
8505 * Description: This function must only be used when non-maskable interrupts
8506 * are serviced on a separate stack. It allows the architecture to switch the
8507 * notion of the current task on a cpu in a non-blocking manner. This function
8508 * must be called with all CPU's synchronized, and interrupts disabled, the
8509 * and caller must save the original value of the current task (see
8510 * curr_task() above) and restore that value before reenabling interrupts and
8511 * re-starting the system.
8512 *
8513 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8514 */
8515void set_curr_task(int cpu, struct task_struct *p)
8516{
8517 cpu_curr(cpu) = p;
8518}
8519
8520#endif
8521
8522#ifdef CONFIG_FAIR_GROUP_SCHED
8523static void free_fair_sched_group(struct task_group *tg)
8524{
8525 int i;
8526
8527 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8528
8529 for_each_possible_cpu(i) {
8530 if (tg->cfs_rq)
8531 kfree(tg->cfs_rq[i]);
8532 if (tg->se)
8533 kfree(tg->se[i]);
8534 }
8535
8536 kfree(tg->cfs_rq);
8537 kfree(tg->se);
8538}
8539
8540static
8541int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8542{
8543 struct cfs_rq *cfs_rq;
8544 struct sched_entity *se;
8545 int i;
8546
8547 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8548 if (!tg->cfs_rq)
8549 goto err;
8550 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8551 if (!tg->se)
8552 goto err;
8553
8554 tg->shares = NICE_0_LOAD;
8555
8556 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8557
8558 for_each_possible_cpu(i) {
8559 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8560 GFP_KERNEL, cpu_to_node(i));
8561 if (!cfs_rq)
8562 goto err;
8563
8564 se = kzalloc_node(sizeof(struct sched_entity),
8565 GFP_KERNEL, cpu_to_node(i));
8566 if (!se)
8567 goto err_free_rq;
8568
8569 init_cfs_rq(cfs_rq);
8570 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8571 }
8572
8573 return 1;
8574
8575err_free_rq:
8576 kfree(cfs_rq);
8577err:
8578 return 0;
8579}
8580
8581static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8582{
8583 struct rq *rq = cpu_rq(cpu);
8584 unsigned long flags;
8585
8586 /*
8587 * Only empty task groups can be destroyed; so we can speculatively
8588 * check on_list without danger of it being re-added.
8589 */
8590 if (!tg->cfs_rq[cpu]->on_list)
8591 return;
8592
8593 raw_spin_lock_irqsave(&rq->lock, flags);
8594 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8595 raw_spin_unlock_irqrestore(&rq->lock, flags);
8596}
8597#else /* !CONFIG_FAIR_GROUP_SCHED */
8598static inline void free_fair_sched_group(struct task_group *tg)
8599{
8600}
8601
8602static inline
8603int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8604{
8605 return 1;
8606}
8607
8608static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8609{
8610}
8611#endif /* CONFIG_FAIR_GROUP_SCHED */
8612
8613#ifdef CONFIG_RT_GROUP_SCHED
8614static void free_rt_sched_group(struct task_group *tg)
8615{
8616 int i;
8617
8618 if (tg->rt_se)
8619 destroy_rt_bandwidth(&tg->rt_bandwidth);
8620
8621 for_each_possible_cpu(i) {
8622 if (tg->rt_rq)
8623 kfree(tg->rt_rq[i]);
8624 if (tg->rt_se)
8625 kfree(tg->rt_se[i]);
8626 }
8627
8628 kfree(tg->rt_rq);
8629 kfree(tg->rt_se);
8630}
8631
8632static
8633int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8634{
8635 struct rt_rq *rt_rq;
8636 struct sched_rt_entity *rt_se;
8637 int i;
8638
8639 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8640 if (!tg->rt_rq)
8641 goto err;
8642 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8643 if (!tg->rt_se)
8644 goto err;
8645
8646 init_rt_bandwidth(&tg->rt_bandwidth,
8647 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8648
8649 for_each_possible_cpu(i) {
8650 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8651 GFP_KERNEL, cpu_to_node(i));
8652 if (!rt_rq)
8653 goto err;
8654
8655 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8656 GFP_KERNEL, cpu_to_node(i));
8657 if (!rt_se)
8658 goto err_free_rq;
8659
8660 init_rt_rq(rt_rq, cpu_rq(i));
8661 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8662 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8663 }
8664
8665 return 1;
8666
8667err_free_rq:
8668 kfree(rt_rq);
8669err:
8670 return 0;
8671}
8672#else /* !CONFIG_RT_GROUP_SCHED */
8673static inline void free_rt_sched_group(struct task_group *tg)
8674{
8675}
8676
8677static inline
8678int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8679{
8680 return 1;
8681}
8682#endif /* CONFIG_RT_GROUP_SCHED */
8683
8684#ifdef CONFIG_CGROUP_SCHED
8685static void free_sched_group(struct task_group *tg)
8686{
8687 free_fair_sched_group(tg);
8688 free_rt_sched_group(tg);
8689 autogroup_free(tg);
8690 kfree(tg);
8691}
8692
8693/* allocate runqueue etc for a new task group */
8694struct task_group *sched_create_group(struct task_group *parent)
8695{
8696 struct task_group *tg;
8697 unsigned long flags;
8698
8699 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8700 if (!tg)
8701 return ERR_PTR(-ENOMEM);
8702
8703 if (!alloc_fair_sched_group(tg, parent))
8704 goto err;
8705
8706 if (!alloc_rt_sched_group(tg, parent))
8707 goto err;
8708
8709 spin_lock_irqsave(&task_group_lock, flags);
8710 list_add_rcu(&tg->list, &task_groups);
8711
8712 WARN_ON(!parent); /* root should already exist */
8713
8714 tg->parent = parent;
8715 INIT_LIST_HEAD(&tg->children);
8716 list_add_rcu(&tg->siblings, &parent->children);
8717 spin_unlock_irqrestore(&task_group_lock, flags);
8718
8719 return tg;
8720
8721err:
8722 free_sched_group(tg);
8723 return ERR_PTR(-ENOMEM);
8724}
8725
8726/* rcu callback to free various structures associated with a task group */
8727static void free_sched_group_rcu(struct rcu_head *rhp)
8728{
8729 /* now it should be safe to free those cfs_rqs */
8730 free_sched_group(container_of(rhp, struct task_group, rcu));
8731}
8732
8733/* Destroy runqueue etc associated with a task group */
8734void sched_destroy_group(struct task_group *tg)
8735{
8736 unsigned long flags;
8737 int i;
8738
8739 /* end participation in shares distribution */
8740 for_each_possible_cpu(i)
8741 unregister_fair_sched_group(tg, i);
8742
8743 spin_lock_irqsave(&task_group_lock, flags);
8744 list_del_rcu(&tg->list);
8745 list_del_rcu(&tg->siblings);
8746 spin_unlock_irqrestore(&task_group_lock, flags);
8747
8748 /* wait for possible concurrent references to cfs_rqs complete */
8749 call_rcu(&tg->rcu, free_sched_group_rcu);
8750}
8751
8752/* change task's runqueue when it moves between groups.
8753 * The caller of this function should have put the task in its new group
8754 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8755 * reflect its new group.
8756 */
8757void sched_move_task(struct task_struct *tsk)
8758{
8759 int on_rq, running;
8760 unsigned long flags;
8761 struct rq *rq;
8762
8763 rq = task_rq_lock(tsk, &flags);
8764
8765 running = task_current(rq, tsk);
8766 on_rq = tsk->on_rq;
8767
8768 if (on_rq)
8769 dequeue_task(rq, tsk, 0);
8770 if (unlikely(running))
8771 tsk->sched_class->put_prev_task(rq, tsk);
8772
8773#ifdef CONFIG_FAIR_GROUP_SCHED
8774 if (tsk->sched_class->task_move_group)
8775 tsk->sched_class->task_move_group(tsk, on_rq);
8776 else
8777#endif
8778 set_task_rq(tsk, task_cpu(tsk));
8779
8780 if (unlikely(running))
8781 tsk->sched_class->set_curr_task(rq);
8782 if (on_rq)
8783 enqueue_task(rq, tsk, 0);
8784
8785 task_rq_unlock(rq, tsk, &flags);
8786}
8787#endif /* CONFIG_CGROUP_SCHED */
8788
8789#ifdef CONFIG_FAIR_GROUP_SCHED
8790static DEFINE_MUTEX(shares_mutex);
8791
8792int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8793{
8794 int i;
8795 unsigned long flags;
8796
8797 /*
8798 * We can't change the weight of the root cgroup.
8799 */
8800 if (!tg->se[0])
8801 return -EINVAL;
8802
8803 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8804
8805 mutex_lock(&shares_mutex);
8806 if (tg->shares == shares)
8807 goto done;
8808
8809 tg->shares = shares;
8810 for_each_possible_cpu(i) {
8811 struct rq *rq = cpu_rq(i);
8812 struct sched_entity *se;
8813
8814 se = tg->se[i];
8815 /* Propagate contribution to hierarchy */
8816 raw_spin_lock_irqsave(&rq->lock, flags);
8817 for_each_sched_entity(se)
8818 update_cfs_shares(group_cfs_rq(se));
8819 raw_spin_unlock_irqrestore(&rq->lock, flags);
8820 }
8821
8822done:
8823 mutex_unlock(&shares_mutex);
8824 return 0;
8825}
8826
8827unsigned long sched_group_shares(struct task_group *tg)
8828{
8829 return tg->shares;
8830}
8831#endif
8832
8833#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
8834static unsigned long to_ratio(u64 period, u64 runtime)
8835{
8836 if (runtime == RUNTIME_INF)
8837 return 1ULL << 20;
8838
8839 return div64_u64(runtime << 20, period);
8840}
8841#endif
8842
8843#ifdef CONFIG_RT_GROUP_SCHED
8844/*
8845 * Ensure that the real time constraints are schedulable.
8846 */
8847static DEFINE_MUTEX(rt_constraints_mutex);
8848
8849/* Must be called with tasklist_lock held */
8850static inline int tg_has_rt_tasks(struct task_group *tg)
8851{
8852 struct task_struct *g, *p;
8853
8854 do_each_thread(g, p) {
8855 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8856 return 1;
8857 } while_each_thread(g, p);
8858
8859 return 0;
8860}
8861
8862struct rt_schedulable_data {
8863 struct task_group *tg;
8864 u64 rt_period;
8865 u64 rt_runtime;
8866};
8867
8868static int tg_rt_schedulable(struct task_group *tg, void *data)
8869{
8870 struct rt_schedulable_data *d = data;
8871 struct task_group *child;
8872 unsigned long total, sum = 0;
8873 u64 period, runtime;
8874
8875 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8876 runtime = tg->rt_bandwidth.rt_runtime;
8877
8878 if (tg == d->tg) {
8879 period = d->rt_period;
8880 runtime = d->rt_runtime;
8881 }
8882
8883 /*
8884 * Cannot have more runtime than the period.
8885 */
8886 if (runtime > period && runtime != RUNTIME_INF)
8887 return -EINVAL;
8888
8889 /*
8890 * Ensure we don't starve existing RT tasks.
8891 */
8892 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8893 return -EBUSY;
8894
8895 total = to_ratio(period, runtime);
8896
8897 /*
8898 * Nobody can have more than the global setting allows.
8899 */
8900 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8901 return -EINVAL;
8902
8903 /*
8904 * The sum of our children's runtime should not exceed our own.
8905 */
8906 list_for_each_entry_rcu(child, &tg->children, siblings) {
8907 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8908 runtime = child->rt_bandwidth.rt_runtime;
8909
8910 if (child == d->tg) {
8911 period = d->rt_period;
8912 runtime = d->rt_runtime;
8913 }
8914
8915 sum += to_ratio(period, runtime);
8916 }
8917
8918 if (sum > total)
8919 return -EINVAL;
8920
8921 return 0;
8922}
8923
8924static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8925{
8926 int ret;
8927
8928 struct rt_schedulable_data data = {
8929 .tg = tg,
8930 .rt_period = period,
8931 .rt_runtime = runtime,
8932 };
8933
8934 rcu_read_lock();
8935 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8936 rcu_read_unlock();
8937
8938 return ret;
8939}
8940
8941static int tg_set_rt_bandwidth(struct task_group *tg,
8942 u64 rt_period, u64 rt_runtime)
8943{
8944 int i, err = 0;
8945
8946 mutex_lock(&rt_constraints_mutex);
8947 read_lock(&tasklist_lock);
8948 err = __rt_schedulable(tg, rt_period, rt_runtime);
8949 if (err)
8950 goto unlock;
8951
8952 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8953 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8954 tg->rt_bandwidth.rt_runtime = rt_runtime;
8955
8956 for_each_possible_cpu(i) {
8957 struct rt_rq *rt_rq = tg->rt_rq[i];
8958
8959 raw_spin_lock(&rt_rq->rt_runtime_lock);
8960 rt_rq->rt_runtime = rt_runtime;
8961 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8962 }
8963 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8964unlock:
8965 read_unlock(&tasklist_lock);
8966 mutex_unlock(&rt_constraints_mutex);
8967
8968 return err;
8969}
8970
8971int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8972{
8973 u64 rt_runtime, rt_period;
8974
8975 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8976 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8977 if (rt_runtime_us < 0)
8978 rt_runtime = RUNTIME_INF;
8979
8980 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8981}
8982
8983long sched_group_rt_runtime(struct task_group *tg)
8984{
8985 u64 rt_runtime_us;
8986
8987 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8988 return -1;
8989
8990 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8991 do_div(rt_runtime_us, NSEC_PER_USEC);
8992 return rt_runtime_us;
8993}
8994
8995int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8996{
8997 u64 rt_runtime, rt_period;
8998
8999 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9000 rt_runtime = tg->rt_bandwidth.rt_runtime;
9001
9002 if (rt_period == 0)
9003 return -EINVAL;
9004
9005 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
9006}
9007
9008long sched_group_rt_period(struct task_group *tg)
9009{
9010 u64 rt_period_us;
9011
9012 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9013 do_div(rt_period_us, NSEC_PER_USEC);
9014 return rt_period_us;
9015}
9016
9017static int sched_rt_global_constraints(void)
9018{
9019 u64 runtime, period;
9020 int ret = 0;
9021
9022 if (sysctl_sched_rt_period <= 0)
9023 return -EINVAL;
9024
9025 runtime = global_rt_runtime();
9026 period = global_rt_period();
9027
9028 /*
9029 * Sanity check on the sysctl variables.
9030 */
9031 if (runtime > period && runtime != RUNTIME_INF)
9032 return -EINVAL;
9033
9034 mutex_lock(&rt_constraints_mutex);
9035 read_lock(&tasklist_lock);
9036 ret = __rt_schedulable(NULL, 0, 0);
9037 read_unlock(&tasklist_lock);
9038 mutex_unlock(&rt_constraints_mutex);
9039
9040 return ret;
9041}
9042
9043int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9044{
9045 /* Don't accept realtime tasks when there is no way for them to run */
9046 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9047 return 0;
9048
9049 return 1;
9050}
9051
9052#else /* !CONFIG_RT_GROUP_SCHED */
9053static int sched_rt_global_constraints(void)
9054{
9055 unsigned long flags;
9056 int i;
9057
9058 if (sysctl_sched_rt_period <= 0)
9059 return -EINVAL;
9060
9061 /*
9062 * There's always some RT tasks in the root group
9063 * -- migration, kstopmachine etc..
9064 */
9065 if (sysctl_sched_rt_runtime == 0)
9066 return -EBUSY;
9067
9068 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9069 for_each_possible_cpu(i) {
9070 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9071
9072 raw_spin_lock(&rt_rq->rt_runtime_lock);
9073 rt_rq->rt_runtime = global_rt_runtime();
9074 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9075 }
9076 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9077
9078 return 0;
9079}
9080#endif /* CONFIG_RT_GROUP_SCHED */
9081
9082int sched_rt_handler(struct ctl_table *table, int write,
9083 void __user *buffer, size_t *lenp,
9084 loff_t *ppos)
9085{
9086 int ret;
9087 int old_period, old_runtime;
9088 static DEFINE_MUTEX(mutex);
9089
9090 mutex_lock(&mutex);
9091 old_period = sysctl_sched_rt_period;
9092 old_runtime = sysctl_sched_rt_runtime;
9093
9094 ret = proc_dointvec(table, write, buffer, lenp, ppos);
9095
9096 if (!ret && write) {
9097 ret = sched_rt_global_constraints();
9098 if (ret) {
9099 sysctl_sched_rt_period = old_period;
9100 sysctl_sched_rt_runtime = old_runtime;
9101 } else {
9102 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9103 def_rt_bandwidth.rt_period =
9104 ns_to_ktime(global_rt_period());
9105 }
9106 }
9107 mutex_unlock(&mutex);
9108
9109 return ret;
9110}
9111
9112#ifdef CONFIG_CGROUP_SCHED
9113
9114/* return corresponding task_group object of a cgroup */
9115static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9116{
9117 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9118 struct task_group, css);
9119}
9120
9121static struct cgroup_subsys_state *
9122cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9123{
9124 struct task_group *tg, *parent;
9125
9126 if (!cgrp->parent) {
9127 /* This is early initialization for the top cgroup */
9128 return &root_task_group.css;
9129 }
9130
9131 parent = cgroup_tg(cgrp->parent);
9132 tg = sched_create_group(parent);
9133 if (IS_ERR(tg))
9134 return ERR_PTR(-ENOMEM);
9135
9136 return &tg->css;
9137}
9138
9139static void
9140cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9141{
9142 struct task_group *tg = cgroup_tg(cgrp);
9143
9144 sched_destroy_group(tg);
9145}
9146
9147static int
9148cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9149{
9150#ifdef CONFIG_RT_GROUP_SCHED
9151 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9152 return -EINVAL;
9153#else
9154 /* We don't support RT-tasks being in separate groups */
9155 if (tsk->sched_class != &fair_sched_class)
9156 return -EINVAL;
9157#endif
9158 return 0;
9159}
9160
9161static void
9162cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9163{
9164 sched_move_task(tsk);
9165}
9166
9167static void
9168cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9169 struct cgroup *old_cgrp, struct task_struct *task)
9170{
9171 /*
9172 * cgroup_exit() is called in the copy_process() failure path.
9173 * Ignore this case since the task hasn't ran yet, this avoids
9174 * trying to poke a half freed task state from generic code.
9175 */
9176 if (!(task->flags & PF_EXITING))
9177 return;
9178
9179 sched_move_task(task);
9180}
9181
9182#ifdef CONFIG_FAIR_GROUP_SCHED
9183static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9184 u64 shareval)
9185{
9186 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
9187}
9188
9189static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9190{
9191 struct task_group *tg = cgroup_tg(cgrp);
9192
9193 return (u64) scale_load_down(tg->shares);
9194}
9195
9196#ifdef CONFIG_CFS_BANDWIDTH
9197static DEFINE_MUTEX(cfs_constraints_mutex);
9198
9199const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
9200const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
9201
9202static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
9203
9204static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
9205{
9206 int i, ret = 0, runtime_enabled;
9207 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9208
9209 if (tg == &root_task_group)
9210 return -EINVAL;
9211
9212 /*
9213 * Ensure we have at some amount of bandwidth every period. This is
9214 * to prevent reaching a state of large arrears when throttled via
9215 * entity_tick() resulting in prolonged exit starvation.
9216 */
9217 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
9218 return -EINVAL;
9219
9220 /*
9221 * Likewise, bound things on the otherside by preventing insane quota
9222 * periods. This also allows us to normalize in computing quota
9223 * feasibility.
9224 */
9225 if (period > max_cfs_quota_period)
9226 return -EINVAL;
9227
9228 mutex_lock(&cfs_constraints_mutex);
9229 ret = __cfs_schedulable(tg, period, quota);
9230 if (ret)
9231 goto out_unlock;
9232
9233 runtime_enabled = quota != RUNTIME_INF;
9234 raw_spin_lock_irq(&cfs_b->lock);
9235 cfs_b->period = ns_to_ktime(period);
9236 cfs_b->quota = quota;
9237
9238 __refill_cfs_bandwidth_runtime(cfs_b);
9239 /* restart the period timer (if active) to handle new period expiry */
9240 if (runtime_enabled && cfs_b->timer_active) {
9241 /* force a reprogram */
9242 cfs_b->timer_active = 0;
9243 __start_cfs_bandwidth(cfs_b);
9244 }
9245 raw_spin_unlock_irq(&cfs_b->lock);
9246
9247 for_each_possible_cpu(i) {
9248 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
9249 struct rq *rq = rq_of(cfs_rq);
9250
9251 raw_spin_lock_irq(&rq->lock);
9252 cfs_rq->runtime_enabled = runtime_enabled;
9253 cfs_rq->runtime_remaining = 0;
9254
9255 if (cfs_rq_throttled(cfs_rq))
9256 unthrottle_cfs_rq(cfs_rq);
9257 raw_spin_unlock_irq(&rq->lock);
9258 }
9259out_unlock:
9260 mutex_unlock(&cfs_constraints_mutex);
9261
9262 return ret;
9263}
9264
9265int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
9266{
9267 u64 quota, period;
9268
9269 period = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9270 if (cfs_quota_us < 0)
9271 quota = RUNTIME_INF;
9272 else
9273 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
9274
9275 return tg_set_cfs_bandwidth(tg, period, quota);
9276}
9277
9278long tg_get_cfs_quota(struct task_group *tg)
9279{
9280 u64 quota_us;
9281
9282 if (tg_cfs_bandwidth(tg)->quota == RUNTIME_INF)
9283 return -1;
9284
9285 quota_us = tg_cfs_bandwidth(tg)->quota;
9286 do_div(quota_us, NSEC_PER_USEC);
9287
9288 return quota_us;
9289}
9290
9291int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
9292{
9293 u64 quota, period;
9294
9295 period = (u64)cfs_period_us * NSEC_PER_USEC;
9296 quota = tg_cfs_bandwidth(tg)->quota;
9297
9298 if (period <= 0)
9299 return -EINVAL;
9300
9301 return tg_set_cfs_bandwidth(tg, period, quota);
9302}
9303
9304long tg_get_cfs_period(struct task_group *tg)
9305{
9306 u64 cfs_period_us;
9307
9308 cfs_period_us = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9309 do_div(cfs_period_us, NSEC_PER_USEC);
9310
9311 return cfs_period_us;
9312}
9313
9314static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
9315{
9316 return tg_get_cfs_quota(cgroup_tg(cgrp));
9317}
9318
9319static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
9320 s64 cfs_quota_us)
9321{
9322 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
9323}
9324
9325static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
9326{
9327 return tg_get_cfs_period(cgroup_tg(cgrp));
9328}
9329
9330static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9331 u64 cfs_period_us)
9332{
9333 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
9334}
9335
9336struct cfs_schedulable_data {
9337 struct task_group *tg;
9338 u64 period, quota;
9339};
9340
9341/*
9342 * normalize group quota/period to be quota/max_period
9343 * note: units are usecs
9344 */
9345static u64 normalize_cfs_quota(struct task_group *tg,
9346 struct cfs_schedulable_data *d)
9347{
9348 u64 quota, period;
9349
9350 if (tg == d->tg) {
9351 period = d->period;
9352 quota = d->quota;
9353 } else {
9354 period = tg_get_cfs_period(tg);
9355 quota = tg_get_cfs_quota(tg);
9356 }
9357
9358 /* note: these should typically be equivalent */
9359 if (quota == RUNTIME_INF || quota == -1)
9360 return RUNTIME_INF;
9361
9362 return to_ratio(period, quota);
9363}
9364
9365static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
9366{
9367 struct cfs_schedulable_data *d = data;
9368 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9369 s64 quota = 0, parent_quota = -1;
9370
9371 if (!tg->parent) {
9372 quota = RUNTIME_INF;
9373 } else {
9374 struct cfs_bandwidth *parent_b = tg_cfs_bandwidth(tg->parent);
9375
9376 quota = normalize_cfs_quota(tg, d);
9377 parent_quota = parent_b->hierarchal_quota;
9378
9379 /*
9380 * ensure max(child_quota) <= parent_quota, inherit when no
9381 * limit is set
9382 */
9383 if (quota == RUNTIME_INF)
9384 quota = parent_quota;
9385 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
9386 return -EINVAL;
9387 }
9388 cfs_b->hierarchal_quota = quota;
9389
9390 return 0;
9391}
9392
9393static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
9394{
9395 int ret;
9396 struct cfs_schedulable_data data = {
9397 .tg = tg,
9398 .period = period,
9399 .quota = quota,
9400 };
9401
9402 if (quota != RUNTIME_INF) {
9403 do_div(data.period, NSEC_PER_USEC);
9404 do_div(data.quota, NSEC_PER_USEC);
9405 }
9406
9407 rcu_read_lock();
9408 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
9409 rcu_read_unlock();
9410
9411 return ret;
9412}
9413
9414static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
9415 struct cgroup_map_cb *cb)
9416{
9417 struct task_group *tg = cgroup_tg(cgrp);
9418 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9419
9420 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
9421 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
9422 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
9423
9424 return 0;
9425}
9426#endif /* CONFIG_CFS_BANDWIDTH */
9427#endif /* CONFIG_FAIR_GROUP_SCHED */
9428
9429#ifdef CONFIG_RT_GROUP_SCHED
9430static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9431 s64 val)
9432{
9433 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9434}
9435
9436static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9437{
9438 return sched_group_rt_runtime(cgroup_tg(cgrp));
9439}
9440
9441static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9442 u64 rt_period_us)
9443{
9444 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9445}
9446
9447static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9448{
9449 return sched_group_rt_period(cgroup_tg(cgrp));
9450}
9451#endif /* CONFIG_RT_GROUP_SCHED */
9452
9453static struct cftype cpu_files[] = {
9454#ifdef CONFIG_FAIR_GROUP_SCHED
9455 {
9456 .name = "shares",
9457 .read_u64 = cpu_shares_read_u64,
9458 .write_u64 = cpu_shares_write_u64,
9459 },
9460#endif
9461#ifdef CONFIG_CFS_BANDWIDTH
9462 {
9463 .name = "cfs_quota_us",
9464 .read_s64 = cpu_cfs_quota_read_s64,
9465 .write_s64 = cpu_cfs_quota_write_s64,
9466 },
9467 {
9468 .name = "cfs_period_us",
9469 .read_u64 = cpu_cfs_period_read_u64,
9470 .write_u64 = cpu_cfs_period_write_u64,
9471 },
9472 {
9473 .name = "stat",
9474 .read_map = cpu_stats_show,
9475 },
9476#endif
9477#ifdef CONFIG_RT_GROUP_SCHED
9478 {
9479 .name = "rt_runtime_us",
9480 .read_s64 = cpu_rt_runtime_read,
9481 .write_s64 = cpu_rt_runtime_write,
9482 },
9483 {
9484 .name = "rt_period_us",
9485 .read_u64 = cpu_rt_period_read_uint,
9486 .write_u64 = cpu_rt_period_write_uint,
9487 },
9488#endif
9489};
9490
9491static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9492{
9493 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9494}
9495
9496struct cgroup_subsys cpu_cgroup_subsys = {
9497 .name = "cpu",
9498 .create = cpu_cgroup_create,
9499 .destroy = cpu_cgroup_destroy,
9500 .can_attach_task = cpu_cgroup_can_attach_task,
9501 .attach_task = cpu_cgroup_attach_task,
9502 .exit = cpu_cgroup_exit,
9503 .populate = cpu_cgroup_populate,
9504 .subsys_id = cpu_cgroup_subsys_id,
9505 .early_init = 1,
9506};
9507
9508#endif /* CONFIG_CGROUP_SCHED */
9509
9510#ifdef CONFIG_CGROUP_CPUACCT
9511
9512/*
9513 * CPU accounting code for task groups.
9514 *
9515 * Based on the work by Paul Menage ([email protected]) and Balbir Singh
9516 * ([email protected]).
9517 */
9518
9519/* track cpu usage of a group of tasks and its child groups */
9520struct cpuacct {
9521 struct cgroup_subsys_state css;
9522 /* cpuusage holds pointer to a u64-type object on every cpu */
9523 u64 __percpu *cpuusage;
9524 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9525 struct cpuacct *parent;
9526};
9527
9528struct cgroup_subsys cpuacct_subsys;
9529
9530/* return cpu accounting group corresponding to this container */
9531static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9532{
9533 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9534 struct cpuacct, css);
9535}
9536
9537/* return cpu accounting group to which this task belongs */
9538static inline struct cpuacct *task_ca(struct task_struct *tsk)
9539{
9540 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9541 struct cpuacct, css);
9542}
9543
9544/* create a new cpu accounting group */
9545static struct cgroup_subsys_state *cpuacct_create(
9546 struct cgroup_subsys *ss, struct cgroup *cgrp)
9547{
9548 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9549 int i;
9550
9551 if (!ca)
9552 goto out;
9553
9554 ca->cpuusage = alloc_percpu(u64);
9555 if (!ca->cpuusage)
9556 goto out_free_ca;
9557
9558 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9559 if (percpu_counter_init(&ca->cpustat[i], 0))
9560 goto out_free_counters;
9561
9562 if (cgrp->parent)
9563 ca->parent = cgroup_ca(cgrp->parent);
9564
9565 return &ca->css;
9566
9567out_free_counters:
9568 while (--i >= 0)
9569 percpu_counter_destroy(&ca->cpustat[i]);
9570 free_percpu(ca->cpuusage);
9571out_free_ca:
9572 kfree(ca);
9573out:
9574 return ERR_PTR(-ENOMEM);
9575}
9576
9577/* destroy an existing cpu accounting group */
9578static void
9579cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9580{
9581 struct cpuacct *ca = cgroup_ca(cgrp);
9582 int i;
9583
9584 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9585 percpu_counter_destroy(&ca->cpustat[i]);
9586 free_percpu(ca->cpuusage);
9587 kfree(ca);
9588}
9589
9590static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9591{
9592 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9593 u64 data;
9594
9595#ifndef CONFIG_64BIT
9596 /*
9597 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9598 */
9599 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9600 data = *cpuusage;
9601 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9602#else
9603 data = *cpuusage;
9604#endif
9605
9606 return data;
9607}
9608
9609static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9610{
9611 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9612
9613#ifndef CONFIG_64BIT
9614 /*
9615 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9616 */
9617 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9618 *cpuusage = val;
9619 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9620#else
9621 *cpuusage = val;
9622#endif
9623}
9624
9625/* return total cpu usage (in nanoseconds) of a group */
9626static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9627{
9628 struct cpuacct *ca = cgroup_ca(cgrp);
9629 u64 totalcpuusage = 0;
9630 int i;
9631
9632 for_each_present_cpu(i)
9633 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9634
9635 return totalcpuusage;
9636}
9637
9638static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9639 u64 reset)
9640{
9641 struct cpuacct *ca = cgroup_ca(cgrp);
9642 int err = 0;
9643 int i;
9644
9645 if (reset) {
9646 err = -EINVAL;
9647 goto out;
9648 }
9649
9650 for_each_present_cpu(i)
9651 cpuacct_cpuusage_write(ca, i, 0);
9652
9653out:
9654 return err;
9655}
9656
9657static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9658 struct seq_file *m)
9659{
9660 struct cpuacct *ca = cgroup_ca(cgroup);
9661 u64 percpu;
9662 int i;
9663
9664 for_each_present_cpu(i) {
9665 percpu = cpuacct_cpuusage_read(ca, i);
9666 seq_printf(m, "%llu ", (unsigned long long) percpu);
9667 }
9668 seq_printf(m, "\n");
9669 return 0;
9670}
9671
9672static const char *cpuacct_stat_desc[] = {
9673 [CPUACCT_STAT_USER] = "user",
9674 [CPUACCT_STAT_SYSTEM] = "system",
9675};
9676
9677static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9678 struct cgroup_map_cb *cb)
9679{
9680 struct cpuacct *ca = cgroup_ca(cgrp);
9681 int i;
9682
9683 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9684 s64 val = percpu_counter_read(&ca->cpustat[i]);
9685 val = cputime64_to_clock_t(val);
9686 cb->fill(cb, cpuacct_stat_desc[i], val);
9687 }
9688 return 0;
9689}
9690
9691static struct cftype files[] = {
9692 {
9693 .name = "usage",
9694 .read_u64 = cpuusage_read,
9695 .write_u64 = cpuusage_write,
9696 },
9697 {
9698 .name = "usage_percpu",
9699 .read_seq_string = cpuacct_percpu_seq_read,
9700 },
9701 {
9702 .name = "stat",
9703 .read_map = cpuacct_stats_show,
9704 },
9705};
9706
9707static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9708{
9709 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9710}
9711
9712/*
9713 * charge this task's execution time to its accounting group.
9714 *
9715 * called with rq->lock held.
9716 */
9717static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9718{
9719 struct cpuacct *ca;
9720 int cpu;
9721
9722 if (unlikely(!cpuacct_subsys.active))
9723 return;
9724
9725 cpu = task_cpu(tsk);
9726
9727 rcu_read_lock();
9728
9729 ca = task_ca(tsk);
9730
9731 for (; ca; ca = ca->parent) {
9732 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9733 *cpuusage += cputime;
9734 }
9735
9736 rcu_read_unlock();
9737}
9738
9739/*
9740 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9741 * in cputime_t units. As a result, cpuacct_update_stats calls
9742 * percpu_counter_add with values large enough to always overflow the
9743 * per cpu batch limit causing bad SMP scalability.
9744 *
9745 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9746 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9747 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9748 */
9749#ifdef CONFIG_SMP
9750#define CPUACCT_BATCH \
9751 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9752#else
9753#define CPUACCT_BATCH 0
9754#endif
9755
9756/*
9757 * Charge the system/user time to the task's accounting group.
9758 */
9759static void cpuacct_update_stats(struct task_struct *tsk,
9760 enum cpuacct_stat_index idx, cputime_t val)
9761{
9762 struct cpuacct *ca;
9763 int batch = CPUACCT_BATCH;
9764
9765 if (unlikely(!cpuacct_subsys.active))
9766 return;
9767
9768 rcu_read_lock();
9769 ca = task_ca(tsk);
9770
9771 do {
9772 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9773 ca = ca->parent;
9774 } while (ca);
9775 rcu_read_unlock();
9776}
9777
9778struct cgroup_subsys cpuacct_subsys = {
9779 .name = "cpuacct",
9780 .create = cpuacct_create,
9781 .destroy = cpuacct_destroy,
9782 .populate = cpuacct_populate,
9783 .subsys_id = cpuacct_subsys_id,
9784};
9785#endif /* CONFIG_CGROUP_CPUACCT */