projects/linux·blob v3.1 · refs

9347 lines · 222.3 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
75#include <asm/tlb.h>
76#include <asm/irq_regs.h>
77#include <asm/mutex.h>
78#ifdef CONFIG_PARAVIRT
79#include <asm/paravirt.h>
80#endif
81
82#include "sched_cpupri.h"
83#include "workqueue_sched.h"
84#include "sched_autogroup.h"
85
86#define CREATE_TRACE_POINTS
87#include <trace/events/sched.h>
88
89/*
90 * Convert user-nice values [ -20 ... 0 ... 19 ]
91 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
92 * and back.
93 */
94#define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
95#define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
96#define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
97
98/*
99 * 'User priority' is the nice value converted to something we
100 * can work with better when scaling various scheduler parameters,
101 * it's a [ 0 ... 39 ] range.
102 */
103#define USER_PRIO(p) ((p)-MAX_RT_PRIO)
104#define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
105#define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
106
107/*
108 * Helpers for converting nanosecond timing to jiffy resolution
109 */
110#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
111
112#define NICE_0_LOAD SCHED_LOAD_SCALE
113#define NICE_0_SHIFT SCHED_LOAD_SHIFT
114
115/*
116 * These are the 'tuning knobs' of the scheduler:
117 *
118 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
119 * Timeslices get refilled after they expire.
120 */
121#define DEF_TIMESLICE (100 * HZ / 1000)
122
123/*
124 * single value that denotes runtime == period, ie unlimited time.
125 */
126#define RUNTIME_INF ((u64)~0ULL)
127
128static inline int rt_policy(int policy)
129{
130 if (policy == SCHED_FIFO || policy == SCHED_RR)
131 return 1;
132 return 0;
133}
134
135static inline int task_has_rt_policy(struct task_struct *p)
136{
137 return rt_policy(p->policy);
138}
139
140/*
141 * This is the priority-queue data structure of the RT scheduling class:
142 */
143struct rt_prio_array {
144 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
145 struct list_head queue[MAX_RT_PRIO];
146};
147
148struct rt_bandwidth {
149 /* nests inside the rq lock: */
150 raw_spinlock_t rt_runtime_lock;
151 ktime_t rt_period;
152 u64 rt_runtime;
153 struct hrtimer rt_period_timer;
154};
155
156static struct rt_bandwidth def_rt_bandwidth;
157
158static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
159
160static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
161{
162 struct rt_bandwidth *rt_b =
163 container_of(timer, struct rt_bandwidth, rt_period_timer);
164 ktime_t now;
165 int overrun;
166 int idle = 0;
167
168 for (;;) {
169 now = hrtimer_cb_get_time(timer);
170 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
171
172 if (!overrun)
173 break;
174
175 idle = do_sched_rt_period_timer(rt_b, overrun);
176 }
177
178 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
179}
180
181static
182void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
183{
184 rt_b->rt_period = ns_to_ktime(period);
185 rt_b->rt_runtime = runtime;
186
187 raw_spin_lock_init(&rt_b->rt_runtime_lock);
188
189 hrtimer_init(&rt_b->rt_period_timer,
190 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
191 rt_b->rt_period_timer.function = sched_rt_period_timer;
192}
193
194static inline int rt_bandwidth_enabled(void)
195{
196 return sysctl_sched_rt_runtime >= 0;
197}
198
199static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
200{
201 ktime_t now;
202
203 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
204 return;
205
206 if (hrtimer_active(&rt_b->rt_period_timer))
207 return;
208
209 raw_spin_lock(&rt_b->rt_runtime_lock);
210 for (;;) {
211 unsigned long delta;
212 ktime_t soft, hard;
213
214 if (hrtimer_active(&rt_b->rt_period_timer))
215 break;
216
217 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
218 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
219
220 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
221 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
222 delta = ktime_to_ns(ktime_sub(hard, soft));
223 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
224 HRTIMER_MODE_ABS_PINNED, 0);
225 }
226 raw_spin_unlock(&rt_b->rt_runtime_lock);
227}
228
229#ifdef CONFIG_RT_GROUP_SCHED
230static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
231{
232 hrtimer_cancel(&rt_b->rt_period_timer);
233}
234#endif
235
236/*
237 * sched_domains_mutex serializes calls to init_sched_domains,
238 * detach_destroy_domains and partition_sched_domains.
239 */
240static DEFINE_MUTEX(sched_domains_mutex);
241
242#ifdef CONFIG_CGROUP_SCHED
243
244#include <linux/cgroup.h>
245
246struct cfs_rq;
247
248static LIST_HEAD(task_groups);
249
250/* task group related information */
251struct task_group {
252 struct cgroup_subsys_state css;
253
254#ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
260
261 atomic_t load_weight;
262#endif
263
264#ifdef CONFIG_RT_GROUP_SCHED
265 struct sched_rt_entity **rt_se;
266 struct rt_rq **rt_rq;
267
268 struct rt_bandwidth rt_bandwidth;
269#endif
270
271 struct rcu_head rcu;
272 struct list_head list;
273
274 struct task_group *parent;
275 struct list_head siblings;
276 struct list_head children;
277
278#ifdef CONFIG_SCHED_AUTOGROUP
279 struct autogroup *autogroup;
280#endif
281};
282
283/* task_group_lock serializes the addition/removal of task groups */
284static DEFINE_SPINLOCK(task_group_lock);
285
286#ifdef CONFIG_FAIR_GROUP_SCHED
287
288# define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
289
290/*
291 * A weight of 0 or 1 can cause arithmetics problems.
292 * A weight of a cfs_rq is the sum of weights of which entities
293 * are queued on this cfs_rq, so a weight of a entity should not be
294 * too large, so as the shares value of a task group.
295 * (The default weight is 1024 - so there's no practical
296 * limitation from this.)
297 */
298#define MIN_SHARES (1UL << 1)
299#define MAX_SHARES (1UL << 18)
300
301static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
302#endif
303
304/* Default task group.
305 * Every task in system belong to this group at bootup.
306 */
307struct task_group root_task_group;
308
309#endif /* CONFIG_CGROUP_SCHED */
310
311/* CFS-related fields in a runqueue */
312struct cfs_rq {
313 struct load_weight load;
314 unsigned long nr_running;
315
316 u64 exec_clock;
317 u64 min_vruntime;
318#ifndef CONFIG_64BIT
319 u64 min_vruntime_copy;
320#endif
321
322 struct rb_root tasks_timeline;
323 struct rb_node *rb_leftmost;
324
325 struct list_head tasks;
326 struct list_head *balance_iterator;
327
328 /*
329 * 'curr' points to currently running entity on this cfs_rq.
330 * It is set to NULL otherwise (i.e when none are currently running).
331 */
332 struct sched_entity *curr, *next, *last, *skip;
333
334#ifdef CONFIG_SCHED_DEBUG
335 unsigned int nr_spread_over;
336#endif
337
338#ifdef CONFIG_FAIR_GROUP_SCHED
339 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
340
341 /*
342 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
343 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
344 * (like users, containers etc.)
345 *
346 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
347 * list is used during load balance.
348 */
349 int on_list;
350 struct list_head leaf_cfs_rq_list;
351 struct task_group *tg; /* group that "owns" this runqueue */
352
353#ifdef CONFIG_SMP
354 /*
355 * the part of load.weight contributed by tasks
356 */
357 unsigned long task_weight;
358
359 /*
360 * h_load = weight * f(tg)
361 *
362 * Where f(tg) is the recursive weight fraction assigned to
363 * this group.
364 */
365 unsigned long h_load;
366
367 /*
368 * Maintaining per-cpu shares distribution for group scheduling
369 *
370 * load_stamp is the last time we updated the load average
371 * load_last is the last time we updated the load average and saw load
372 * load_unacc_exec_time is currently unaccounted execution time
373 */
374 u64 load_avg;
375 u64 load_period;
376 u64 load_stamp, load_last, load_unacc_exec_time;
377
378 unsigned long load_contribution;
379#endif
380#endif
381};
382
383/* Real-Time classes' related field in a runqueue: */
384struct rt_rq {
385 struct rt_prio_array active;
386 unsigned long rt_nr_running;
387#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
388 struct {
389 int curr; /* highest queued rt task prio */
390#ifdef CONFIG_SMP
391 int next; /* next highest */
392#endif
393 } highest_prio;
394#endif
395#ifdef CONFIG_SMP
396 unsigned long rt_nr_migratory;
397 unsigned long rt_nr_total;
398 int overloaded;
399 struct plist_head pushable_tasks;
400#endif
401 int rt_throttled;
402 u64 rt_time;
403 u64 rt_runtime;
404 /* Nests inside the rq lock: */
405 raw_spinlock_t rt_runtime_lock;
406
407#ifdef CONFIG_RT_GROUP_SCHED
408 unsigned long rt_nr_boosted;
409
410 struct rq *rq;
411 struct list_head leaf_rt_rq_list;
412 struct task_group *tg;
413#endif
414};
415
416#ifdef CONFIG_SMP
417
418/*
419 * We add the notion of a root-domain which will be used to define per-domain
420 * variables. Each exclusive cpuset essentially defines an island domain by
421 * fully partitioning the member cpus from any other cpuset. Whenever a new
422 * exclusive cpuset is created, we also create and attach a new root-domain
423 * object.
424 *
425 */
426struct root_domain {
427 atomic_t refcount;
428 atomic_t rto_count;
429 struct rcu_head rcu;
430 cpumask_var_t span;
431 cpumask_var_t online;
432
433 /*
434 * The "RT overload" flag: it gets set if a CPU has more than
435 * one runnable RT task.
436 */
437 cpumask_var_t rto_mask;
438 struct cpupri cpupri;
439};
440
441/*
442 * By default the system creates a single root-domain with all cpus as
443 * members (mimicking the global state we have today).
444 */
445static struct root_domain def_root_domain;
446
447#endif /* CONFIG_SMP */
448
449/*
450 * This is the main, per-CPU runqueue data structure.
451 *
452 * Locking rule: those places that want to lock multiple runqueues
453 * (such as the load balancing or the thread migration code), lock
454 * acquire operations must be ordered by ascending &runqueue.
455 */
456struct rq {
457 /* runqueue lock: */
458 raw_spinlock_t lock;
459
460 /*
461 * nr_running and cpu_load should be in the same cacheline because
462 * remote CPUs use both these fields when doing load calculation.
463 */
464 unsigned long nr_running;
465 #define CPU_LOAD_IDX_MAX 5
466 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
467 unsigned long last_load_update_tick;
468#ifdef CONFIG_NO_HZ
469 u64 nohz_stamp;
470 unsigned char nohz_balance_kick;
471#endif
472 int skip_clock_update;
473
474 /* capture load from *all* tasks on this cpu: */
475 struct load_weight load;
476 unsigned long nr_load_updates;
477 u64 nr_switches;
478
479 struct cfs_rq cfs;
480 struct rt_rq rt;
481
482#ifdef CONFIG_FAIR_GROUP_SCHED
483 /* list of leaf cfs_rq on this cpu: */
484 struct list_head leaf_cfs_rq_list;
485#endif
486#ifdef CONFIG_RT_GROUP_SCHED
487 struct list_head leaf_rt_rq_list;
488#endif
489
490 /*
491 * This is part of a global counter where only the total sum
492 * over all CPUs matters. A task can increase this counter on
493 * one CPU and if it got migrated afterwards it may decrease
494 * it on another CPU. Always updated under the runqueue lock:
495 */
496 unsigned long nr_uninterruptible;
497
498 struct task_struct *curr, *idle, *stop;
499 unsigned long next_balance;
500 struct mm_struct *prev_mm;
501
502 u64 clock;
503 u64 clock_task;
504
505 atomic_t nr_iowait;
506
507#ifdef CONFIG_SMP
508 struct root_domain *rd;
509 struct sched_domain *sd;
510
511 unsigned long cpu_power;
512
513 unsigned char idle_at_tick;
514 /* For active balancing */
515 int post_schedule;
516 int active_balance;
517 int push_cpu;
518 struct cpu_stop_work active_balance_work;
519 /* cpu of this runqueue: */
520 int cpu;
521 int online;
522
523 unsigned long avg_load_per_task;
524
525 u64 rt_avg;
526 u64 age_stamp;
527 u64 idle_stamp;
528 u64 avg_idle;
529#endif
530
531#ifdef CONFIG_IRQ_TIME_ACCOUNTING
532 u64 prev_irq_time;
533#endif
534#ifdef CONFIG_PARAVIRT
535 u64 prev_steal_time;
536#endif
537#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
538 u64 prev_steal_time_rq;
539#endif
540
541 /* calc_load related fields */
542 unsigned long calc_load_update;
543 long calc_load_active;
544
545#ifdef CONFIG_SCHED_HRTICK
546#ifdef CONFIG_SMP
547 int hrtick_csd_pending;
548 struct call_single_data hrtick_csd;
549#endif
550 struct hrtimer hrtick_timer;
551#endif
552
553#ifdef CONFIG_SCHEDSTATS
554 /* latency stats */
555 struct sched_info rq_sched_info;
556 unsigned long long rq_cpu_time;
557 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
558
559 /* sys_sched_yield() stats */
560 unsigned int yld_count;
561
562 /* schedule() stats */
563 unsigned int sched_switch;
564 unsigned int sched_count;
565 unsigned int sched_goidle;
566
567 /* try_to_wake_up() stats */
568 unsigned int ttwu_count;
569 unsigned int ttwu_local;
570#endif
571
572#ifdef CONFIG_SMP
573 struct task_struct *wake_list;
574#endif
575};
576
577static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
578
579
580static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
581
582static inline int cpu_of(struct rq *rq)
583{
584#ifdef CONFIG_SMP
585 return rq->cpu;
586#else
587 return 0;
588#endif
589}
590
591#define rcu_dereference_check_sched_domain(p) \
592 rcu_dereference_check((p), \
593 lockdep_is_held(&sched_domains_mutex))
594
595/*
596 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
597 * See detach_destroy_domains: synchronize_sched for details.
598 *
599 * The domain tree of any CPU may only be accessed from within
600 * preempt-disabled sections.
601 */
602#define for_each_domain(cpu, __sd) \
603 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
604
605#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
606#define this_rq() (&__get_cpu_var(runqueues))
607#define task_rq(p) cpu_rq(task_cpu(p))
608#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
609#define raw_rq() (&__raw_get_cpu_var(runqueues))
610
611#ifdef CONFIG_CGROUP_SCHED
612
613/*
614 * Return the group to which this tasks belongs.
615 *
616 * We use task_subsys_state_check() and extend the RCU verification with
617 * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each
618 * task it moves into the cgroup. Therefore by holding either of those locks,
619 * we pin the task to the current cgroup.
620 */
621static inline struct task_group *task_group(struct task_struct *p)
622{
623 struct task_group *tg;
624 struct cgroup_subsys_state *css;
625
626 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
627 lockdep_is_held(&p->pi_lock) ||
628 lockdep_is_held(&task_rq(p)->lock));
629 tg = container_of(css, struct task_group, css);
630
631 return autogroup_task_group(p, tg);
632}
633
634/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
635static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
636{
637#ifdef CONFIG_FAIR_GROUP_SCHED
638 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
639 p->se.parent = task_group(p)->se[cpu];
640#endif
641
642#ifdef CONFIG_RT_GROUP_SCHED
643 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
644 p->rt.parent = task_group(p)->rt_se[cpu];
645#endif
646}
647
648#else /* CONFIG_CGROUP_SCHED */
649
650static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
651static inline struct task_group *task_group(struct task_struct *p)
652{
653 return NULL;
654}
655
656#endif /* CONFIG_CGROUP_SCHED */
657
658static void update_rq_clock_task(struct rq *rq, s64 delta);
659
660static void update_rq_clock(struct rq *rq)
661{
662 s64 delta;
663
664 if (rq->skip_clock_update > 0)
665 return;
666
667 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
668 rq->clock += delta;
669 update_rq_clock_task(rq, delta);
670}
671
672/*
673 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
674 */
675#ifdef CONFIG_SCHED_DEBUG
676# define const_debug __read_mostly
677#else
678# define const_debug static const
679#endif
680
681/**
682 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
683 * @cpu: the processor in question.
684 *
685 * This interface allows printk to be called with the runqueue lock
686 * held and know whether or not it is OK to wake up the klogd.
687 */
688int runqueue_is_locked(int cpu)
689{
690 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
691}
692
693/*
694 * Debugging: various feature bits
695 */
696
697#define SCHED_FEAT(name, enabled) \
698 __SCHED_FEAT_##name ,
699
700enum {
701#include "sched_features.h"
702};
703
704#undef SCHED_FEAT
705
706#define SCHED_FEAT(name, enabled) \
707 (1UL << __SCHED_FEAT_##name) * enabled |
708
709const_debug unsigned int sysctl_sched_features =
710#include "sched_features.h"
711 0;
712
713#undef SCHED_FEAT
714
715#ifdef CONFIG_SCHED_DEBUG
716#define SCHED_FEAT(name, enabled) \
717 #name ,
718
719static __read_mostly char *sched_feat_names[] = {
720#include "sched_features.h"
721 NULL
722};
723
724#undef SCHED_FEAT
725
726static int sched_feat_show(struct seq_file *m, void *v)
727{
728 int i;
729
730 for (i = 0; sched_feat_names[i]; i++) {
731 if (!(sysctl_sched_features & (1UL << i)))
732 seq_puts(m, "NO_");
733 seq_printf(m, "%s ", sched_feat_names[i]);
734 }
735 seq_puts(m, "\n");
736
737 return 0;
738}
739
740static ssize_t
741sched_feat_write(struct file *filp, const char __user *ubuf,
742 size_t cnt, loff_t *ppos)
743{
744 char buf[64];
745 char *cmp;
746 int neg = 0;
747 int i;
748
749 if (cnt > 63)
750 cnt = 63;
751
752 if (copy_from_user(&buf, ubuf, cnt))
753 return -EFAULT;
754
755 buf[cnt] = 0;
756 cmp = strstrip(buf);
757
758 if (strncmp(cmp, "NO_", 3) == 0) {
759 neg = 1;
760 cmp += 3;
761 }
762
763 for (i = 0; sched_feat_names[i]; i++) {
764 if (strcmp(cmp, sched_feat_names[i]) == 0) {
765 if (neg)
766 sysctl_sched_features &= ~(1UL << i);
767 else
768 sysctl_sched_features |= (1UL << i);
769 break;
770 }
771 }
772
773 if (!sched_feat_names[i])
774 return -EINVAL;
775
776 *ppos += cnt;
777
778 return cnt;
779}
780
781static int sched_feat_open(struct inode *inode, struct file *filp)
782{
783 return single_open(filp, sched_feat_show, NULL);
784}
785
786static const struct file_operations sched_feat_fops = {
787 .open = sched_feat_open,
788 .write = sched_feat_write,
789 .read = seq_read,
790 .llseek = seq_lseek,
791 .release = single_release,
792};
793
794static __init int sched_init_debug(void)
795{
796 debugfs_create_file("sched_features", 0644, NULL, NULL,
797 &sched_feat_fops);
798
799 return 0;
800}
801late_initcall(sched_init_debug);
802
803#endif
804
805#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
806
807/*
808 * Number of tasks to iterate in a single balance run.
809 * Limited because this is done with IRQs disabled.
810 */
811const_debug unsigned int sysctl_sched_nr_migrate = 32;
812
813/*
814 * period over which we average the RT time consumption, measured
815 * in ms.
816 *
817 * default: 1s
818 */
819const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
820
821/*
822 * period over which we measure -rt task cpu usage in us.
823 * default: 1s
824 */
825unsigned int sysctl_sched_rt_period = 1000000;
826
827static __read_mostly int scheduler_running;
828
829/*
830 * part of the period that we allow rt tasks to run in us.
831 * default: 0.95s
832 */
833int sysctl_sched_rt_runtime = 950000;
834
835static inline u64 global_rt_period(void)
836{
837 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
838}
839
840static inline u64 global_rt_runtime(void)
841{
842 if (sysctl_sched_rt_runtime < 0)
843 return RUNTIME_INF;
844
845 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
846}
847
848#ifndef prepare_arch_switch
849# define prepare_arch_switch(next) do { } while (0)
850#endif
851#ifndef finish_arch_switch
852# define finish_arch_switch(prev) do { } while (0)
853#endif
854
855static inline int task_current(struct rq *rq, struct task_struct *p)
856{
857 return rq->curr == p;
858}
859
860static inline int task_running(struct rq *rq, struct task_struct *p)
861{
862#ifdef CONFIG_SMP
863 return p->on_cpu;
864#else
865 return task_current(rq, p);
866#endif
867}
868
869#ifndef __ARCH_WANT_UNLOCKED_CTXSW
870static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
871{
872#ifdef CONFIG_SMP
873 /*
874 * We can optimise this out completely for !SMP, because the
875 * SMP rebalancing from interrupt is the only thing that cares
876 * here.
877 */
878 next->on_cpu = 1;
879#endif
880}
881
882static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
883{
884#ifdef CONFIG_SMP
885 /*
886 * After ->on_cpu is cleared, the task can be moved to a different CPU.
887 * We must ensure this doesn't happen until the switch is completely
888 * finished.
889 */
890 smp_wmb();
891 prev->on_cpu = 0;
892#endif
893#ifdef CONFIG_DEBUG_SPINLOCK
894 /* this is a valid case when another task releases the spinlock */
895 rq->lock.owner = current;
896#endif
897 /*
898 * If we are tracking spinlock dependencies then we have to
899 * fix up the runqueue lock - which gets 'carried over' from
900 * prev into current:
901 */
902 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
903
904 raw_spin_unlock_irq(&rq->lock);
905}
906
907#else /* __ARCH_WANT_UNLOCKED_CTXSW */
908static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
909{
910#ifdef CONFIG_SMP
911 /*
912 * We can optimise this out completely for !SMP, because the
913 * SMP rebalancing from interrupt is the only thing that cares
914 * here.
915 */
916 next->on_cpu = 1;
917#endif
918#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
919 raw_spin_unlock_irq(&rq->lock);
920#else
921 raw_spin_unlock(&rq->lock);
922#endif
923}
924
925static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
926{
927#ifdef CONFIG_SMP
928 /*
929 * After ->on_cpu is cleared, the task can be moved to a different CPU.
930 * We must ensure this doesn't happen until the switch is completely
931 * finished.
932 */
933 smp_wmb();
934 prev->on_cpu = 0;
935#endif
936#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
937 local_irq_enable();
938#endif
939}
940#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
941
942/*
943 * __task_rq_lock - lock the rq @p resides on.
944 */
945static inline struct rq *__task_rq_lock(struct task_struct *p)
946 __acquires(rq->lock)
947{
948 struct rq *rq;
949
950 lockdep_assert_held(&p->pi_lock);
951
952 for (;;) {
953 rq = task_rq(p);
954 raw_spin_lock(&rq->lock);
955 if (likely(rq == task_rq(p)))
956 return rq;
957 raw_spin_unlock(&rq->lock);
958 }
959}
960
961/*
962 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
963 */
964static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
965 __acquires(p->pi_lock)
966 __acquires(rq->lock)
967{
968 struct rq *rq;
969
970 for (;;) {
971 raw_spin_lock_irqsave(&p->pi_lock, *flags);
972 rq = task_rq(p);
973 raw_spin_lock(&rq->lock);
974 if (likely(rq == task_rq(p)))
975 return rq;
976 raw_spin_unlock(&rq->lock);
977 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
978 }
979}
980
981static void __task_rq_unlock(struct rq *rq)
982 __releases(rq->lock)
983{
984 raw_spin_unlock(&rq->lock);
985}
986
987static inline void
988task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
989 __releases(rq->lock)
990 __releases(p->pi_lock)
991{
992 raw_spin_unlock(&rq->lock);
993 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
994}
995
996/*
997 * this_rq_lock - lock this runqueue and disable interrupts.
998 */
999static struct rq *this_rq_lock(void)
1000 __acquires(rq->lock)
1001{
1002 struct rq *rq;
1003
1004 local_irq_disable();
1005 rq = this_rq();
1006 raw_spin_lock(&rq->lock);
1007
1008 return rq;
1009}
1010
1011#ifdef CONFIG_SCHED_HRTICK
1012/*
1013 * Use HR-timers to deliver accurate preemption points.
1014 *
1015 * Its all a bit involved since we cannot program an hrt while holding the
1016 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1017 * reschedule event.
1018 *
1019 * When we get rescheduled we reprogram the hrtick_timer outside of the
1020 * rq->lock.
1021 */
1022
1023/*
1024 * Use hrtick when:
1025 * - enabled by features
1026 * - hrtimer is actually high res
1027 */
1028static inline int hrtick_enabled(struct rq *rq)
1029{
1030 if (!sched_feat(HRTICK))
1031 return 0;
1032 if (!cpu_active(cpu_of(rq)))
1033 return 0;
1034 return hrtimer_is_hres_active(&rq->hrtick_timer);
1035}
1036
1037static void hrtick_clear(struct rq *rq)
1038{
1039 if (hrtimer_active(&rq->hrtick_timer))
1040 hrtimer_cancel(&rq->hrtick_timer);
1041}
1042
1043/*
1044 * High-resolution timer tick.
1045 * Runs from hardirq context with interrupts disabled.
1046 */
1047static enum hrtimer_restart hrtick(struct hrtimer *timer)
1048{
1049 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1050
1051 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1052
1053 raw_spin_lock(&rq->lock);
1054 update_rq_clock(rq);
1055 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1056 raw_spin_unlock(&rq->lock);
1057
1058 return HRTIMER_NORESTART;
1059}
1060
1061#ifdef CONFIG_SMP
1062/*
1063 * called from hardirq (IPI) context
1064 */
1065static void __hrtick_start(void *arg)
1066{
1067 struct rq *rq = arg;
1068
1069 raw_spin_lock(&rq->lock);
1070 hrtimer_restart(&rq->hrtick_timer);
1071 rq->hrtick_csd_pending = 0;
1072 raw_spin_unlock(&rq->lock);
1073}
1074
1075/*
1076 * Called to set the hrtick timer state.
1077 *
1078 * called with rq->lock held and irqs disabled
1079 */
1080static void hrtick_start(struct rq *rq, u64 delay)
1081{
1082 struct hrtimer *timer = &rq->hrtick_timer;
1083 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1084
1085 hrtimer_set_expires(timer, time);
1086
1087 if (rq == this_rq()) {
1088 hrtimer_restart(timer);
1089 } else if (!rq->hrtick_csd_pending) {
1090 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1091 rq->hrtick_csd_pending = 1;
1092 }
1093}
1094
1095static int
1096hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1097{
1098 int cpu = (int)(long)hcpu;
1099
1100 switch (action) {
1101 case CPU_UP_CANCELED:
1102 case CPU_UP_CANCELED_FROZEN:
1103 case CPU_DOWN_PREPARE:
1104 case CPU_DOWN_PREPARE_FROZEN:
1105 case CPU_DEAD:
1106 case CPU_DEAD_FROZEN:
1107 hrtick_clear(cpu_rq(cpu));
1108 return NOTIFY_OK;
1109 }
1110
1111 return NOTIFY_DONE;
1112}
1113
1114static __init void init_hrtick(void)
1115{
1116 hotcpu_notifier(hotplug_hrtick, 0);
1117}
1118#else
1119/*
1120 * Called to set the hrtick timer state.
1121 *
1122 * called with rq->lock held and irqs disabled
1123 */
1124static void hrtick_start(struct rq *rq, u64 delay)
1125{
1126 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1127 HRTIMER_MODE_REL_PINNED, 0);
1128}
1129
1130static inline void init_hrtick(void)
1131{
1132}
1133#endif /* CONFIG_SMP */
1134
1135static void init_rq_hrtick(struct rq *rq)
1136{
1137#ifdef CONFIG_SMP
1138 rq->hrtick_csd_pending = 0;
1139
1140 rq->hrtick_csd.flags = 0;
1141 rq->hrtick_csd.func = __hrtick_start;
1142 rq->hrtick_csd.info = rq;
1143#endif
1144
1145 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1146 rq->hrtick_timer.function = hrtick;
1147}
1148#else /* CONFIG_SCHED_HRTICK */
1149static inline void hrtick_clear(struct rq *rq)
1150{
1151}
1152
1153static inline void init_rq_hrtick(struct rq *rq)
1154{
1155}
1156
1157static inline void init_hrtick(void)
1158{
1159}
1160#endif /* CONFIG_SCHED_HRTICK */
1161
1162/*
1163 * resched_task - mark a task 'to be rescheduled now'.
1164 *
1165 * On UP this means the setting of the need_resched flag, on SMP it
1166 * might also involve a cross-CPU call to trigger the scheduler on
1167 * the target CPU.
1168 */
1169#ifdef CONFIG_SMP
1170
1171#ifndef tsk_is_polling
1172#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1173#endif
1174
1175static void resched_task(struct task_struct *p)
1176{
1177 int cpu;
1178
1179 assert_raw_spin_locked(&task_rq(p)->lock);
1180
1181 if (test_tsk_need_resched(p))
1182 return;
1183
1184 set_tsk_need_resched(p);
1185
1186 cpu = task_cpu(p);
1187 if (cpu == smp_processor_id())
1188 return;
1189
1190 /* NEED_RESCHED must be visible before we test polling */
1191 smp_mb();
1192 if (!tsk_is_polling(p))
1193 smp_send_reschedule(cpu);
1194}
1195
1196static void resched_cpu(int cpu)
1197{
1198 struct rq *rq = cpu_rq(cpu);
1199 unsigned long flags;
1200
1201 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1202 return;
1203 resched_task(cpu_curr(cpu));
1204 raw_spin_unlock_irqrestore(&rq->lock, flags);
1205}
1206
1207#ifdef CONFIG_NO_HZ
1208/*
1209 * In the semi idle case, use the nearest busy cpu for migrating timers
1210 * from an idle cpu. This is good for power-savings.
1211 *
1212 * We don't do similar optimization for completely idle system, as
1213 * selecting an idle cpu will add more delays to the timers than intended
1214 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1215 */
1216int get_nohz_timer_target(void)
1217{
1218 int cpu = smp_processor_id();
1219 int i;
1220 struct sched_domain *sd;
1221
1222 rcu_read_lock();
1223 for_each_domain(cpu, sd) {
1224 for_each_cpu(i, sched_domain_span(sd)) {
1225 if (!idle_cpu(i)) {
1226 cpu = i;
1227 goto unlock;
1228 }
1229 }
1230 }
1231unlock:
1232 rcu_read_unlock();
1233 return cpu;
1234}
1235/*
1236 * When add_timer_on() enqueues a timer into the timer wheel of an
1237 * idle CPU then this timer might expire before the next timer event
1238 * which is scheduled to wake up that CPU. In case of a completely
1239 * idle system the next event might even be infinite time into the
1240 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1241 * leaves the inner idle loop so the newly added timer is taken into
1242 * account when the CPU goes back to idle and evaluates the timer
1243 * wheel for the next timer event.
1244 */
1245void wake_up_idle_cpu(int cpu)
1246{
1247 struct rq *rq = cpu_rq(cpu);
1248
1249 if (cpu == smp_processor_id())
1250 return;
1251
1252 /*
1253 * This is safe, as this function is called with the timer
1254 * wheel base lock of (cpu) held. When the CPU is on the way
1255 * to idle and has not yet set rq->curr to idle then it will
1256 * be serialized on the timer wheel base lock and take the new
1257 * timer into account automatically.
1258 */
1259 if (rq->curr != rq->idle)
1260 return;
1261
1262 /*
1263 * We can set TIF_RESCHED on the idle task of the other CPU
1264 * lockless. The worst case is that the other CPU runs the
1265 * idle task through an additional NOOP schedule()
1266 */
1267 set_tsk_need_resched(rq->idle);
1268
1269 /* NEED_RESCHED must be visible before we test polling */
1270 smp_mb();
1271 if (!tsk_is_polling(rq->idle))
1272 smp_send_reschedule(cpu);
1273}
1274
1275#endif /* CONFIG_NO_HZ */
1276
1277static u64 sched_avg_period(void)
1278{
1279 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1280}
1281
1282static void sched_avg_update(struct rq *rq)
1283{
1284 s64 period = sched_avg_period();
1285
1286 while ((s64)(rq->clock - rq->age_stamp) > period) {
1287 /*
1288 * Inline assembly required to prevent the compiler
1289 * optimising this loop into a divmod call.
1290 * See __iter_div_u64_rem() for another example of this.
1291 */
1292 asm("" : "+rm" (rq->age_stamp));
1293 rq->age_stamp += period;
1294 rq->rt_avg /= 2;
1295 }
1296}
1297
1298static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1299{
1300 rq->rt_avg += rt_delta;
1301 sched_avg_update(rq);
1302}
1303
1304#else /* !CONFIG_SMP */
1305static void resched_task(struct task_struct *p)
1306{
1307 assert_raw_spin_locked(&task_rq(p)->lock);
1308 set_tsk_need_resched(p);
1309}
1310
1311static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1312{
1313}
1314
1315static void sched_avg_update(struct rq *rq)
1316{
1317}
1318#endif /* CONFIG_SMP */
1319
1320#if BITS_PER_LONG == 32
1321# define WMULT_CONST (~0UL)
1322#else
1323# define WMULT_CONST (1UL << 32)
1324#endif
1325
1326#define WMULT_SHIFT 32
1327
1328/*
1329 * Shift right and round:
1330 */
1331#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1332
1333/*
1334 * delta *= weight / lw
1335 */
1336static unsigned long
1337calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1338 struct load_weight *lw)
1339{
1340 u64 tmp;
1341
1342 /*
1343 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1344 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1345 * 2^SCHED_LOAD_RESOLUTION.
1346 */
1347 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1348 tmp = (u64)delta_exec * scale_load_down(weight);
1349 else
1350 tmp = (u64)delta_exec;
1351
1352 if (!lw->inv_weight) {
1353 unsigned long w = scale_load_down(lw->weight);
1354
1355 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1356 lw->inv_weight = 1;
1357 else if (unlikely(!w))
1358 lw->inv_weight = WMULT_CONST;
1359 else
1360 lw->inv_weight = WMULT_CONST / w;
1361 }
1362
1363 /*
1364 * Check whether we'd overflow the 64-bit multiplication:
1365 */
1366 if (unlikely(tmp > WMULT_CONST))
1367 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1368 WMULT_SHIFT/2);
1369 else
1370 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1371
1372 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1373}
1374
1375static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1376{
1377 lw->weight += inc;
1378 lw->inv_weight = 0;
1379}
1380
1381static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1382{
1383 lw->weight -= dec;
1384 lw->inv_weight = 0;
1385}
1386
1387static inline void update_load_set(struct load_weight *lw, unsigned long w)
1388{
1389 lw->weight = w;
1390 lw->inv_weight = 0;
1391}
1392
1393/*
1394 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1395 * of tasks with abnormal "nice" values across CPUs the contribution that
1396 * each task makes to its run queue's load is weighted according to its
1397 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1398 * scaled version of the new time slice allocation that they receive on time
1399 * slice expiry etc.
1400 */
1401
1402#define WEIGHT_IDLEPRIO 3
1403#define WMULT_IDLEPRIO 1431655765
1404
1405/*
1406 * Nice levels are multiplicative, with a gentle 10% change for every
1407 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1408 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1409 * that remained on nice 0.
1410 *
1411 * The "10% effect" is relative and cumulative: from _any_ nice level,
1412 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1413 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1414 * If a task goes up by ~10% and another task goes down by ~10% then
1415 * the relative distance between them is ~25%.)
1416 */
1417static const int prio_to_weight[40] = {
1418 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1419 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1420 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1421 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1422 /* 0 */ 1024, 820, 655, 526, 423,
1423 /* 5 */ 335, 272, 215, 172, 137,
1424 /* 10 */ 110, 87, 70, 56, 45,
1425 /* 15 */ 36, 29, 23, 18, 15,
1426};
1427
1428/*
1429 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1430 *
1431 * In cases where the weight does not change often, we can use the
1432 * precalculated inverse to speed up arithmetics by turning divisions
1433 * into multiplications:
1434 */
1435static const u32 prio_to_wmult[40] = {
1436 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1437 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1438 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1439 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1440 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1441 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1442 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1443 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1444};
1445
1446/* Time spent by the tasks of the cpu accounting group executing in ... */
1447enum cpuacct_stat_index {
1448 CPUACCT_STAT_USER, /* ... user mode */
1449 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1450
1451 CPUACCT_STAT_NSTATS,
1452};
1453
1454#ifdef CONFIG_CGROUP_CPUACCT
1455static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1456static void cpuacct_update_stats(struct task_struct *tsk,
1457 enum cpuacct_stat_index idx, cputime_t val);
1458#else
1459static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1460static inline void cpuacct_update_stats(struct task_struct *tsk,
1461 enum cpuacct_stat_index idx, cputime_t val) {}
1462#endif
1463
1464static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1465{
1466 update_load_add(&rq->load, load);
1467}
1468
1469static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1470{
1471 update_load_sub(&rq->load, load);
1472}
1473
1474#if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1475typedef int (*tg_visitor)(struct task_group *, void *);
1476
1477/*
1478 * Iterate the full tree, calling @down when first entering a node and @up when
1479 * leaving it for the final time.
1480 */
1481static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1482{
1483 struct task_group *parent, *child;
1484 int ret;
1485
1486 rcu_read_lock();
1487 parent = &root_task_group;
1488down:
1489 ret = (*down)(parent, data);
1490 if (ret)
1491 goto out_unlock;
1492 list_for_each_entry_rcu(child, &parent->children, siblings) {
1493 parent = child;
1494 goto down;
1495
1496up:
1497 continue;
1498 }
1499 ret = (*up)(parent, data);
1500 if (ret)
1501 goto out_unlock;
1502
1503 child = parent;
1504 parent = parent->parent;
1505 if (parent)
1506 goto up;
1507out_unlock:
1508 rcu_read_unlock();
1509
1510 return ret;
1511}
1512
1513static int tg_nop(struct task_group *tg, void *data)
1514{
1515 return 0;
1516}
1517#endif
1518
1519#ifdef CONFIG_SMP
1520/* Used instead of source_load when we know the type == 0 */
1521static unsigned long weighted_cpuload(const int cpu)
1522{
1523 return cpu_rq(cpu)->load.weight;
1524}
1525
1526/*
1527 * Return a low guess at the load of a migration-source cpu weighted
1528 * according to the scheduling class and "nice" value.
1529 *
1530 * We want to under-estimate the load of migration sources, to
1531 * balance conservatively.
1532 */
1533static unsigned long source_load(int cpu, int type)
1534{
1535 struct rq *rq = cpu_rq(cpu);
1536 unsigned long total = weighted_cpuload(cpu);
1537
1538 if (type == 0 || !sched_feat(LB_BIAS))
1539 return total;
1540
1541 return min(rq->cpu_load[type-1], total);
1542}
1543
1544/*
1545 * Return a high guess at the load of a migration-target cpu weighted
1546 * according to the scheduling class and "nice" value.
1547 */
1548static unsigned long target_load(int cpu, int type)
1549{
1550 struct rq *rq = cpu_rq(cpu);
1551 unsigned long total = weighted_cpuload(cpu);
1552
1553 if (type == 0 || !sched_feat(LB_BIAS))
1554 return total;
1555
1556 return max(rq->cpu_load[type-1], total);
1557}
1558
1559static unsigned long power_of(int cpu)
1560{
1561 return cpu_rq(cpu)->cpu_power;
1562}
1563
1564static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1565
1566static unsigned long cpu_avg_load_per_task(int cpu)
1567{
1568 struct rq *rq = cpu_rq(cpu);
1569 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1570
1571 if (nr_running)
1572 rq->avg_load_per_task = rq->load.weight / nr_running;
1573 else
1574 rq->avg_load_per_task = 0;
1575
1576 return rq->avg_load_per_task;
1577}
1578
1579#ifdef CONFIG_PREEMPT
1580
1581static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1582
1583/*
1584 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1585 * way at the expense of forcing extra atomic operations in all
1586 * invocations. This assures that the double_lock is acquired using the
1587 * same underlying policy as the spinlock_t on this architecture, which
1588 * reduces latency compared to the unfair variant below. However, it
1589 * also adds more overhead and therefore may reduce throughput.
1590 */
1591static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1592 __releases(this_rq->lock)
1593 __acquires(busiest->lock)
1594 __acquires(this_rq->lock)
1595{
1596 raw_spin_unlock(&this_rq->lock);
1597 double_rq_lock(this_rq, busiest);
1598
1599 return 1;
1600}
1601
1602#else
1603/*
1604 * Unfair double_lock_balance: Optimizes throughput at the expense of
1605 * latency by eliminating extra atomic operations when the locks are
1606 * already in proper order on entry. This favors lower cpu-ids and will
1607 * grant the double lock to lower cpus over higher ids under contention,
1608 * regardless of entry order into the function.
1609 */
1610static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1611 __releases(this_rq->lock)
1612 __acquires(busiest->lock)
1613 __acquires(this_rq->lock)
1614{
1615 int ret = 0;
1616
1617 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1618 if (busiest < this_rq) {
1619 raw_spin_unlock(&this_rq->lock);
1620 raw_spin_lock(&busiest->lock);
1621 raw_spin_lock_nested(&this_rq->lock,
1622 SINGLE_DEPTH_NESTING);
1623 ret = 1;
1624 } else
1625 raw_spin_lock_nested(&busiest->lock,
1626 SINGLE_DEPTH_NESTING);
1627 }
1628 return ret;
1629}
1630
1631#endif /* CONFIG_PREEMPT */
1632
1633/*
1634 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1635 */
1636static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1637{
1638 if (unlikely(!irqs_disabled())) {
1639 /* printk() doesn't work good under rq->lock */
1640 raw_spin_unlock(&this_rq->lock);
1641 BUG_ON(1);
1642 }
1643
1644 return _double_lock_balance(this_rq, busiest);
1645}
1646
1647static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1648 __releases(busiest->lock)
1649{
1650 raw_spin_unlock(&busiest->lock);
1651 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1652}
1653
1654/*
1655 * double_rq_lock - safely lock two runqueues
1656 *
1657 * Note this does not disable interrupts like task_rq_lock,
1658 * you need to do so manually before calling.
1659 */
1660static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1661 __acquires(rq1->lock)
1662 __acquires(rq2->lock)
1663{
1664 BUG_ON(!irqs_disabled());
1665 if (rq1 == rq2) {
1666 raw_spin_lock(&rq1->lock);
1667 __acquire(rq2->lock); /* Fake it out ;) */
1668 } else {
1669 if (rq1 < rq2) {
1670 raw_spin_lock(&rq1->lock);
1671 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1672 } else {
1673 raw_spin_lock(&rq2->lock);
1674 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1675 }
1676 }
1677}
1678
1679/*
1680 * double_rq_unlock - safely unlock two runqueues
1681 *
1682 * Note this does not restore interrupts like task_rq_unlock,
1683 * you need to do so manually after calling.
1684 */
1685static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1686 __releases(rq1->lock)
1687 __releases(rq2->lock)
1688{
1689 raw_spin_unlock(&rq1->lock);
1690 if (rq1 != rq2)
1691 raw_spin_unlock(&rq2->lock);
1692 else
1693 __release(rq2->lock);
1694}
1695
1696#else /* CONFIG_SMP */
1697
1698/*
1699 * double_rq_lock - safely lock two runqueues
1700 *
1701 * Note this does not disable interrupts like task_rq_lock,
1702 * you need to do so manually before calling.
1703 */
1704static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1705 __acquires(rq1->lock)
1706 __acquires(rq2->lock)
1707{
1708 BUG_ON(!irqs_disabled());
1709 BUG_ON(rq1 != rq2);
1710 raw_spin_lock(&rq1->lock);
1711 __acquire(rq2->lock); /* Fake it out ;) */
1712}
1713
1714/*
1715 * double_rq_unlock - safely unlock two runqueues
1716 *
1717 * Note this does not restore interrupts like task_rq_unlock,
1718 * you need to do so manually after calling.
1719 */
1720static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1721 __releases(rq1->lock)
1722 __releases(rq2->lock)
1723{
1724 BUG_ON(rq1 != rq2);
1725 raw_spin_unlock(&rq1->lock);
1726 __release(rq2->lock);
1727}
1728
1729#endif
1730
1731static void calc_load_account_idle(struct rq *this_rq);
1732static void update_sysctl(void);
1733static int get_update_sysctl_factor(void);
1734static void update_cpu_load(struct rq *this_rq);
1735
1736static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1737{
1738 set_task_rq(p, cpu);
1739#ifdef CONFIG_SMP
1740 /*
1741 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1742 * successfuly executed on another CPU. We must ensure that updates of
1743 * per-task data have been completed by this moment.
1744 */
1745 smp_wmb();
1746 task_thread_info(p)->cpu = cpu;
1747#endif
1748}
1749
1750static const struct sched_class rt_sched_class;
1751
1752#define sched_class_highest (&stop_sched_class)
1753#define for_each_class(class) \
1754 for (class = sched_class_highest; class; class = class->next)
1755
1756#include "sched_stats.h"
1757
1758static void inc_nr_running(struct rq *rq)
1759{
1760 rq->nr_running++;
1761}
1762
1763static void dec_nr_running(struct rq *rq)
1764{
1765 rq->nr_running--;
1766}
1767
1768static void set_load_weight(struct task_struct *p)
1769{
1770 int prio = p->static_prio - MAX_RT_PRIO;
1771 struct load_weight *load = &p->se.load;
1772
1773 /*
1774 * SCHED_IDLE tasks get minimal weight:
1775 */
1776 if (p->policy == SCHED_IDLE) {
1777 load->weight = scale_load(WEIGHT_IDLEPRIO);
1778 load->inv_weight = WMULT_IDLEPRIO;
1779 return;
1780 }
1781
1782 load->weight = scale_load(prio_to_weight[prio]);
1783 load->inv_weight = prio_to_wmult[prio];
1784}
1785
1786static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1787{
1788 update_rq_clock(rq);
1789 sched_info_queued(p);
1790 p->sched_class->enqueue_task(rq, p, flags);
1791}
1792
1793static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1794{
1795 update_rq_clock(rq);
1796 sched_info_dequeued(p);
1797 p->sched_class->dequeue_task(rq, p, flags);
1798}
1799
1800/*
1801 * activate_task - move a task to the runqueue.
1802 */
1803static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1804{
1805 if (task_contributes_to_load(p))
1806 rq->nr_uninterruptible--;
1807
1808 enqueue_task(rq, p, flags);
1809 inc_nr_running(rq);
1810}
1811
1812/*
1813 * deactivate_task - remove a task from the runqueue.
1814 */
1815static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1816{
1817 if (task_contributes_to_load(p))
1818 rq->nr_uninterruptible++;
1819
1820 dequeue_task(rq, p, flags);
1821 dec_nr_running(rq);
1822}
1823
1824#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1825
1826/*
1827 * There are no locks covering percpu hardirq/softirq time.
1828 * They are only modified in account_system_vtime, on corresponding CPU
1829 * with interrupts disabled. So, writes are safe.
1830 * They are read and saved off onto struct rq in update_rq_clock().
1831 * This may result in other CPU reading this CPU's irq time and can
1832 * race with irq/account_system_vtime on this CPU. We would either get old
1833 * or new value with a side effect of accounting a slice of irq time to wrong
1834 * task when irq is in progress while we read rq->clock. That is a worthy
1835 * compromise in place of having locks on each irq in account_system_time.
1836 */
1837static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1838static DEFINE_PER_CPU(u64, cpu_softirq_time);
1839
1840static DEFINE_PER_CPU(u64, irq_start_time);
1841static int sched_clock_irqtime;
1842
1843void enable_sched_clock_irqtime(void)
1844{
1845 sched_clock_irqtime = 1;
1846}
1847
1848void disable_sched_clock_irqtime(void)
1849{
1850 sched_clock_irqtime = 0;
1851}
1852
1853#ifndef CONFIG_64BIT
1854static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1855
1856static inline void irq_time_write_begin(void)
1857{
1858 __this_cpu_inc(irq_time_seq.sequence);
1859 smp_wmb();
1860}
1861
1862static inline void irq_time_write_end(void)
1863{
1864 smp_wmb();
1865 __this_cpu_inc(irq_time_seq.sequence);
1866}
1867
1868static inline u64 irq_time_read(int cpu)
1869{
1870 u64 irq_time;
1871 unsigned seq;
1872
1873 do {
1874 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1875 irq_time = per_cpu(cpu_softirq_time, cpu) +
1876 per_cpu(cpu_hardirq_time, cpu);
1877 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1878
1879 return irq_time;
1880}
1881#else /* CONFIG_64BIT */
1882static inline void irq_time_write_begin(void)
1883{
1884}
1885
1886static inline void irq_time_write_end(void)
1887{
1888}
1889
1890static inline u64 irq_time_read(int cpu)
1891{
1892 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1893}
1894#endif /* CONFIG_64BIT */
1895
1896/*
1897 * Called before incrementing preempt_count on {soft,}irq_enter
1898 * and before decrementing preempt_count on {soft,}irq_exit.
1899 */
1900void account_system_vtime(struct task_struct *curr)
1901{
1902 unsigned long flags;
1903 s64 delta;
1904 int cpu;
1905
1906 if (!sched_clock_irqtime)
1907 return;
1908
1909 local_irq_save(flags);
1910
1911 cpu = smp_processor_id();
1912 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1913 __this_cpu_add(irq_start_time, delta);
1914
1915 irq_time_write_begin();
1916 /*
1917 * We do not account for softirq time from ksoftirqd here.
1918 * We want to continue accounting softirq time to ksoftirqd thread
1919 * in that case, so as not to confuse scheduler with a special task
1920 * that do not consume any time, but still wants to run.
1921 */
1922 if (hardirq_count())
1923 __this_cpu_add(cpu_hardirq_time, delta);
1924 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1925 __this_cpu_add(cpu_softirq_time, delta);
1926
1927 irq_time_write_end();
1928 local_irq_restore(flags);
1929}
1930EXPORT_SYMBOL_GPL(account_system_vtime);
1931
1932#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1933
1934#ifdef CONFIG_PARAVIRT
1935static inline u64 steal_ticks(u64 steal)
1936{
1937 if (unlikely(steal > NSEC_PER_SEC))
1938 return div_u64(steal, TICK_NSEC);
1939
1940 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
1941}
1942#endif
1943
1944static void update_rq_clock_task(struct rq *rq, s64 delta)
1945{
1946/*
1947 * In theory, the compile should just see 0 here, and optimize out the call
1948 * to sched_rt_avg_update. But I don't trust it...
1949 */
1950#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
1951 s64 steal = 0, irq_delta = 0;
1952#endif
1953#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1954 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1955
1956 /*
1957 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1958 * this case when a previous update_rq_clock() happened inside a
1959 * {soft,}irq region.
1960 *
1961 * When this happens, we stop ->clock_task and only update the
1962 * prev_irq_time stamp to account for the part that fit, so that a next
1963 * update will consume the rest. This ensures ->clock_task is
1964 * monotonic.
1965 *
1966 * It does however cause some slight miss-attribution of {soft,}irq
1967 * time, a more accurate solution would be to update the irq_time using
1968 * the current rq->clock timestamp, except that would require using
1969 * atomic ops.
1970 */
1971 if (irq_delta > delta)
1972 irq_delta = delta;
1973
1974 rq->prev_irq_time += irq_delta;
1975 delta -= irq_delta;
1976#endif
1977#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
1978 if (static_branch((&paravirt_steal_rq_enabled))) {
1979 u64 st;
1980
1981 steal = paravirt_steal_clock(cpu_of(rq));
1982 steal -= rq->prev_steal_time_rq;
1983
1984 if (unlikely(steal > delta))
1985 steal = delta;
1986
1987 st = steal_ticks(steal);
1988 steal = st * TICK_NSEC;
1989
1990 rq->prev_steal_time_rq += steal;
1991
1992 delta -= steal;
1993 }
1994#endif
1995
1996 rq->clock_task += delta;
1997
1998#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
1999 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
2000 sched_rt_avg_update(rq, irq_delta + steal);
2001#endif
2002}
2003
2004#ifdef CONFIG_IRQ_TIME_ACCOUNTING
2005static int irqtime_account_hi_update(void)
2006{
2007 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2008 unsigned long flags;
2009 u64 latest_ns;
2010 int ret = 0;
2011
2012 local_irq_save(flags);
2013 latest_ns = this_cpu_read(cpu_hardirq_time);
2014 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
2015 ret = 1;
2016 local_irq_restore(flags);
2017 return ret;
2018}
2019
2020static int irqtime_account_si_update(void)
2021{
2022 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2023 unsigned long flags;
2024 u64 latest_ns;
2025 int ret = 0;
2026
2027 local_irq_save(flags);
2028 latest_ns = this_cpu_read(cpu_softirq_time);
2029 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2030 ret = 1;
2031 local_irq_restore(flags);
2032 return ret;
2033}
2034
2035#else /* CONFIG_IRQ_TIME_ACCOUNTING */
2036
2037#define sched_clock_irqtime (0)
2038
2039#endif
2040
2041#include "sched_idletask.c"
2042#include "sched_fair.c"
2043#include "sched_rt.c"
2044#include "sched_autogroup.c"
2045#include "sched_stoptask.c"
2046#ifdef CONFIG_SCHED_DEBUG
2047# include "sched_debug.c"
2048#endif
2049
2050void sched_set_stop_task(int cpu, struct task_struct *stop)
2051{
2052 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2053 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2054
2055 if (stop) {
2056 /*
2057 * Make it appear like a SCHED_FIFO task, its something
2058 * userspace knows about and won't get confused about.
2059 *
2060 * Also, it will make PI more or less work without too
2061 * much confusion -- but then, stop work should not
2062 * rely on PI working anyway.
2063 */
2064 sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
2065
2066 stop->sched_class = &stop_sched_class;
2067 }
2068
2069 cpu_rq(cpu)->stop = stop;
2070
2071 if (old_stop) {
2072 /*
2073 * Reset it back to a normal scheduling class so that
2074 * it can die in pieces.
2075 */
2076 old_stop->sched_class = &rt_sched_class;
2077 }
2078}
2079
2080/*
2081 * __normal_prio - return the priority that is based on the static prio
2082 */
2083static inline int __normal_prio(struct task_struct *p)
2084{
2085 return p->static_prio;
2086}
2087
2088/*
2089 * Calculate the expected normal priority: i.e. priority
2090 * without taking RT-inheritance into account. Might be
2091 * boosted by interactivity modifiers. Changes upon fork,
2092 * setprio syscalls, and whenever the interactivity
2093 * estimator recalculates.
2094 */
2095static inline int normal_prio(struct task_struct *p)
2096{
2097 int prio;
2098
2099 if (task_has_rt_policy(p))
2100 prio = MAX_RT_PRIO-1 - p->rt_priority;
2101 else
2102 prio = __normal_prio(p);
2103 return prio;
2104}
2105
2106/*
2107 * Calculate the current priority, i.e. the priority
2108 * taken into account by the scheduler. This value might
2109 * be boosted by RT tasks, or might be boosted by
2110 * interactivity modifiers. Will be RT if the task got
2111 * RT-boosted. If not then it returns p->normal_prio.
2112 */
2113static int effective_prio(struct task_struct *p)
2114{
2115 p->normal_prio = normal_prio(p);
2116 /*
2117 * If we are RT tasks or we were boosted to RT priority,
2118 * keep the priority unchanged. Otherwise, update priority
2119 * to the normal priority:
2120 */
2121 if (!rt_prio(p->prio))
2122 return p->normal_prio;
2123 return p->prio;
2124}
2125
2126/**
2127 * task_curr - is this task currently executing on a CPU?
2128 * @p: the task in question.
2129 */
2130inline int task_curr(const struct task_struct *p)
2131{
2132 return cpu_curr(task_cpu(p)) == p;
2133}
2134
2135static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2136 const struct sched_class *prev_class,
2137 int oldprio)
2138{
2139 if (prev_class != p->sched_class) {
2140 if (prev_class->switched_from)
2141 prev_class->switched_from(rq, p);
2142 p->sched_class->switched_to(rq, p);
2143 } else if (oldprio != p->prio)
2144 p->sched_class->prio_changed(rq, p, oldprio);
2145}
2146
2147static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2148{
2149 const struct sched_class *class;
2150
2151 if (p->sched_class == rq->curr->sched_class) {
2152 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2153 } else {
2154 for_each_class(class) {
2155 if (class == rq->curr->sched_class)
2156 break;
2157 if (class == p->sched_class) {
2158 resched_task(rq->curr);
2159 break;
2160 }
2161 }
2162 }
2163
2164 /*
2165 * A queue event has occurred, and we're going to schedule. In
2166 * this case, we can save a useless back to back clock update.
2167 */
2168 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2169 rq->skip_clock_update = 1;
2170}
2171
2172#ifdef CONFIG_SMP
2173/*
2174 * Is this task likely cache-hot:
2175 */
2176static int
2177task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2178{
2179 s64 delta;
2180
2181 if (p->sched_class != &fair_sched_class)
2182 return 0;
2183
2184 if (unlikely(p->policy == SCHED_IDLE))
2185 return 0;
2186
2187 /*
2188 * Buddy candidates are cache hot:
2189 */
2190 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2191 (&p->se == cfs_rq_of(&p->se)->next ||
2192 &p->se == cfs_rq_of(&p->se)->last))
2193 return 1;
2194
2195 if (sysctl_sched_migration_cost == -1)
2196 return 1;
2197 if (sysctl_sched_migration_cost == 0)
2198 return 0;
2199
2200 delta = now - p->se.exec_start;
2201
2202 return delta < (s64)sysctl_sched_migration_cost;
2203}
2204
2205void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2206{
2207#ifdef CONFIG_SCHED_DEBUG
2208 /*
2209 * We should never call set_task_cpu() on a blocked task,
2210 * ttwu() will sort out the placement.
2211 */
2212 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2213 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2214
2215#ifdef CONFIG_LOCKDEP
2216 /*
2217 * The caller should hold either p->pi_lock or rq->lock, when changing
2218 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2219 *
2220 * sched_move_task() holds both and thus holding either pins the cgroup,
2221 * see set_task_rq().
2222 *
2223 * Furthermore, all task_rq users should acquire both locks, see
2224 * task_rq_lock().
2225 */
2226 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2227 lockdep_is_held(&task_rq(p)->lock)));
2228#endif
2229#endif
2230
2231 trace_sched_migrate_task(p, new_cpu);
2232
2233 if (task_cpu(p) != new_cpu) {
2234 p->se.nr_migrations++;
2235 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
2236 }
2237
2238 __set_task_cpu(p, new_cpu);
2239}
2240
2241struct migration_arg {
2242 struct task_struct *task;
2243 int dest_cpu;
2244};
2245
2246static int migration_cpu_stop(void *data);
2247
2248/*
2249 * wait_task_inactive - wait for a thread to unschedule.
2250 *
2251 * If @match_state is nonzero, it's the @p->state value just checked and
2252 * not expected to change. If it changes, i.e. @p might have woken up,
2253 * then return zero. When we succeed in waiting for @p to be off its CPU,
2254 * we return a positive number (its total switch count). If a second call
2255 * a short while later returns the same number, the caller can be sure that
2256 * @p has remained unscheduled the whole time.
2257 *
2258 * The caller must ensure that the task *will* unschedule sometime soon,
2259 * else this function might spin for a *long* time. This function can't
2260 * be called with interrupts off, or it may introduce deadlock with
2261 * smp_call_function() if an IPI is sent by the same process we are
2262 * waiting to become inactive.
2263 */
2264unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2265{
2266 unsigned long flags;
2267 int running, on_rq;
2268 unsigned long ncsw;
2269 struct rq *rq;
2270
2271 for (;;) {
2272 /*
2273 * We do the initial early heuristics without holding
2274 * any task-queue locks at all. We'll only try to get
2275 * the runqueue lock when things look like they will
2276 * work out!
2277 */
2278 rq = task_rq(p);
2279
2280 /*
2281 * If the task is actively running on another CPU
2282 * still, just relax and busy-wait without holding
2283 * any locks.
2284 *
2285 * NOTE! Since we don't hold any locks, it's not
2286 * even sure that "rq" stays as the right runqueue!
2287 * But we don't care, since "task_running()" will
2288 * return false if the runqueue has changed and p
2289 * is actually now running somewhere else!
2290 */
2291 while (task_running(rq, p)) {
2292 if (match_state && unlikely(p->state != match_state))
2293 return 0;
2294 cpu_relax();
2295 }
2296
2297 /*
2298 * Ok, time to look more closely! We need the rq
2299 * lock now, to be *sure*. If we're wrong, we'll
2300 * just go back and repeat.
2301 */
2302 rq = task_rq_lock(p, &flags);
2303 trace_sched_wait_task(p);
2304 running = task_running(rq, p);
2305 on_rq = p->on_rq;
2306 ncsw = 0;
2307 if (!match_state || p->state == match_state)
2308 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2309 task_rq_unlock(rq, p, &flags);
2310
2311 /*
2312 * If it changed from the expected state, bail out now.
2313 */
2314 if (unlikely(!ncsw))
2315 break;
2316
2317 /*
2318 * Was it really running after all now that we
2319 * checked with the proper locks actually held?
2320 *
2321 * Oops. Go back and try again..
2322 */
2323 if (unlikely(running)) {
2324 cpu_relax();
2325 continue;
2326 }
2327
2328 /*
2329 * It's not enough that it's not actively running,
2330 * it must be off the runqueue _entirely_, and not
2331 * preempted!
2332 *
2333 * So if it was still runnable (but just not actively
2334 * running right now), it's preempted, and we should
2335 * yield - it could be a while.
2336 */
2337 if (unlikely(on_rq)) {
2338 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2339
2340 set_current_state(TASK_UNINTERRUPTIBLE);
2341 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2342 continue;
2343 }
2344
2345 /*
2346 * Ahh, all good. It wasn't running, and it wasn't
2347 * runnable, which means that it will never become
2348 * running in the future either. We're all done!
2349 */
2350 break;
2351 }
2352
2353 return ncsw;
2354}
2355
2356/***
2357 * kick_process - kick a running thread to enter/exit the kernel
2358 * @p: the to-be-kicked thread
2359 *
2360 * Cause a process which is running on another CPU to enter
2361 * kernel-mode, without any delay. (to get signals handled.)
2362 *
2363 * NOTE: this function doesn't have to take the runqueue lock,
2364 * because all it wants to ensure is that the remote task enters
2365 * the kernel. If the IPI races and the task has been migrated
2366 * to another CPU then no harm is done and the purpose has been
2367 * achieved as well.
2368 */
2369void kick_process(struct task_struct *p)
2370{
2371 int cpu;
2372
2373 preempt_disable();
2374 cpu = task_cpu(p);
2375 if ((cpu != smp_processor_id()) && task_curr(p))
2376 smp_send_reschedule(cpu);
2377 preempt_enable();
2378}
2379EXPORT_SYMBOL_GPL(kick_process);
2380#endif /* CONFIG_SMP */
2381
2382#ifdef CONFIG_SMP
2383/*
2384 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2385 */
2386static int select_fallback_rq(int cpu, struct task_struct *p)
2387{
2388 int dest_cpu;
2389 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2390
2391 /* Look for allowed, online CPU in same node. */
2392 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2393 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2394 return dest_cpu;
2395
2396 /* Any allowed, online CPU? */
2397 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2398 if (dest_cpu < nr_cpu_ids)
2399 return dest_cpu;
2400
2401 /* No more Mr. Nice Guy. */
2402 dest_cpu = cpuset_cpus_allowed_fallback(p);
2403 /*
2404 * Don't tell them about moving exiting tasks or
2405 * kernel threads (both mm NULL), since they never
2406 * leave kernel.
2407 */
2408 if (p->mm && printk_ratelimit()) {
2409 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2410 task_pid_nr(p), p->comm, cpu);
2411 }
2412
2413 return dest_cpu;
2414}
2415
2416/*
2417 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2418 */
2419static inline
2420int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2421{
2422 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2423
2424 /*
2425 * In order not to call set_task_cpu() on a blocking task we need
2426 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2427 * cpu.
2428 *
2429 * Since this is common to all placement strategies, this lives here.
2430 *
2431 * [ this allows ->select_task() to simply return task_cpu(p) and
2432 * not worry about this generic constraint ]
2433 */
2434 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2435 !cpu_online(cpu)))
2436 cpu = select_fallback_rq(task_cpu(p), p);
2437
2438 return cpu;
2439}
2440
2441static void update_avg(u64 *avg, u64 sample)
2442{
2443 s64 diff = sample - *avg;
2444 *avg += diff >> 3;
2445}
2446#endif
2447
2448static void
2449ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2450{
2451#ifdef CONFIG_SCHEDSTATS
2452 struct rq *rq = this_rq();
2453
2454#ifdef CONFIG_SMP
2455 int this_cpu = smp_processor_id();
2456
2457 if (cpu == this_cpu) {
2458 schedstat_inc(rq, ttwu_local);
2459 schedstat_inc(p, se.statistics.nr_wakeups_local);
2460 } else {
2461 struct sched_domain *sd;
2462
2463 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2464 rcu_read_lock();
2465 for_each_domain(this_cpu, sd) {
2466 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2467 schedstat_inc(sd, ttwu_wake_remote);
2468 break;
2469 }
2470 }
2471 rcu_read_unlock();
2472 }
2473
2474 if (wake_flags & WF_MIGRATED)
2475 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2476
2477#endif /* CONFIG_SMP */
2478
2479 schedstat_inc(rq, ttwu_count);
2480 schedstat_inc(p, se.statistics.nr_wakeups);
2481
2482 if (wake_flags & WF_SYNC)
2483 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2484
2485#endif /* CONFIG_SCHEDSTATS */
2486}
2487
2488static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2489{
2490 activate_task(rq, p, en_flags);
2491 p->on_rq = 1;
2492
2493 /* if a worker is waking up, notify workqueue */
2494 if (p->flags & PF_WQ_WORKER)
2495 wq_worker_waking_up(p, cpu_of(rq));
2496}
2497
2498/*
2499 * Mark the task runnable and perform wakeup-preemption.
2500 */
2501static void
2502ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2503{
2504 trace_sched_wakeup(p, true);
2505 check_preempt_curr(rq, p, wake_flags);
2506
2507 p->state = TASK_RUNNING;
2508#ifdef CONFIG_SMP
2509 if (p->sched_class->task_woken)
2510 p->sched_class->task_woken(rq, p);
2511
2512 if (rq->idle_stamp) {
2513 u64 delta = rq->clock - rq->idle_stamp;
2514 u64 max = 2*sysctl_sched_migration_cost;
2515
2516 if (delta > max)
2517 rq->avg_idle = max;
2518 else
2519 update_avg(&rq->avg_idle, delta);
2520 rq->idle_stamp = 0;
2521 }
2522#endif
2523}
2524
2525static void
2526ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2527{
2528#ifdef CONFIG_SMP
2529 if (p->sched_contributes_to_load)
2530 rq->nr_uninterruptible--;
2531#endif
2532
2533 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2534 ttwu_do_wakeup(rq, p, wake_flags);
2535}
2536
2537/*
2538 * Called in case the task @p isn't fully descheduled from its runqueue,
2539 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2540 * since all we need to do is flip p->state to TASK_RUNNING, since
2541 * the task is still ->on_rq.
2542 */
2543static int ttwu_remote(struct task_struct *p, int wake_flags)
2544{
2545 struct rq *rq;
2546 int ret = 0;
2547
2548 rq = __task_rq_lock(p);
2549 if (p->on_rq) {
2550 ttwu_do_wakeup(rq, p, wake_flags);
2551 ret = 1;
2552 }
2553 __task_rq_unlock(rq);
2554
2555 return ret;
2556}
2557
2558#ifdef CONFIG_SMP
2559static void sched_ttwu_do_pending(struct task_struct *list)
2560{
2561 struct rq *rq = this_rq();
2562
2563 raw_spin_lock(&rq->lock);
2564
2565 while (list) {
2566 struct task_struct *p = list;
2567 list = list->wake_entry;
2568 ttwu_do_activate(rq, p, 0);
2569 }
2570
2571 raw_spin_unlock(&rq->lock);
2572}
2573
2574#ifdef CONFIG_HOTPLUG_CPU
2575
2576static void sched_ttwu_pending(void)
2577{
2578 struct rq *rq = this_rq();
2579 struct task_struct *list = xchg(&rq->wake_list, NULL);
2580
2581 if (!list)
2582 return;
2583
2584 sched_ttwu_do_pending(list);
2585}
2586
2587#endif /* CONFIG_HOTPLUG_CPU */
2588
2589void scheduler_ipi(void)
2590{
2591 struct rq *rq = this_rq();
2592 struct task_struct *list = xchg(&rq->wake_list, NULL);
2593
2594 if (!list)
2595 return;
2596
2597 /*
2598 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2599 * traditionally all their work was done from the interrupt return
2600 * path. Now that we actually do some work, we need to make sure
2601 * we do call them.
2602 *
2603 * Some archs already do call them, luckily irq_enter/exit nest
2604 * properly.
2605 *
2606 * Arguably we should visit all archs and update all handlers,
2607 * however a fair share of IPIs are still resched only so this would
2608 * somewhat pessimize the simple resched case.
2609 */
2610 irq_enter();
2611 sched_ttwu_do_pending(list);
2612 irq_exit();
2613}
2614
2615static void ttwu_queue_remote(struct task_struct *p, int cpu)
2616{
2617 struct rq *rq = cpu_rq(cpu);
2618 struct task_struct *next = rq->wake_list;
2619
2620 for (;;) {
2621 struct task_struct *old = next;
2622
2623 p->wake_entry = next;
2624 next = cmpxchg(&rq->wake_list, old, p);
2625 if (next == old)
2626 break;
2627 }
2628
2629 if (!next)
2630 smp_send_reschedule(cpu);
2631}
2632
2633#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2634static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2635{
2636 struct rq *rq;
2637 int ret = 0;
2638
2639 rq = __task_rq_lock(p);
2640 if (p->on_cpu) {
2641 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2642 ttwu_do_wakeup(rq, p, wake_flags);
2643 ret = 1;
2644 }
2645 __task_rq_unlock(rq);
2646
2647 return ret;
2648
2649}
2650#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2651#endif /* CONFIG_SMP */
2652
2653static void ttwu_queue(struct task_struct *p, int cpu)
2654{
2655 struct rq *rq = cpu_rq(cpu);
2656
2657#if defined(CONFIG_SMP)
2658 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2659 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2660 ttwu_queue_remote(p, cpu);
2661 return;
2662 }
2663#endif
2664
2665 raw_spin_lock(&rq->lock);
2666 ttwu_do_activate(rq, p, 0);
2667 raw_spin_unlock(&rq->lock);
2668}
2669
2670/**
2671 * try_to_wake_up - wake up a thread
2672 * @p: the thread to be awakened
2673 * @state: the mask of task states that can be woken
2674 * @wake_flags: wake modifier flags (WF_*)
2675 *
2676 * Put it on the run-queue if it's not already there. The "current"
2677 * thread is always on the run-queue (except when the actual
2678 * re-schedule is in progress), and as such you're allowed to do
2679 * the simpler "current->state = TASK_RUNNING" to mark yourself
2680 * runnable without the overhead of this.
2681 *
2682 * Returns %true if @p was woken up, %false if it was already running
2683 * or @state didn't match @p's state.
2684 */
2685static int
2686try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2687{
2688 unsigned long flags;
2689 int cpu, success = 0;
2690
2691 smp_wmb();
2692 raw_spin_lock_irqsave(&p->pi_lock, flags);
2693 if (!(p->state & state))
2694 goto out;
2695
2696 success = 1; /* we're going to change ->state */
2697 cpu = task_cpu(p);
2698
2699 if (p->on_rq && ttwu_remote(p, wake_flags))
2700 goto stat;
2701
2702#ifdef CONFIG_SMP
2703 /*
2704 * If the owning (remote) cpu is still in the middle of schedule() with
2705 * this task as prev, wait until its done referencing the task.
2706 */
2707 while (p->on_cpu) {
2708#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2709 /*
2710 * In case the architecture enables interrupts in
2711 * context_switch(), we cannot busy wait, since that
2712 * would lead to deadlocks when an interrupt hits and
2713 * tries to wake up @prev. So bail and do a complete
2714 * remote wakeup.
2715 */
2716 if (ttwu_activate_remote(p, wake_flags))
2717 goto stat;
2718#else
2719 cpu_relax();
2720#endif
2721 }
2722 /*
2723 * Pairs with the smp_wmb() in finish_lock_switch().
2724 */
2725 smp_rmb();
2726
2727 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2728 p->state = TASK_WAKING;
2729
2730 if (p->sched_class->task_waking)
2731 p->sched_class->task_waking(p);
2732
2733 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2734 if (task_cpu(p) != cpu) {
2735 wake_flags |= WF_MIGRATED;
2736 set_task_cpu(p, cpu);
2737 }
2738#endif /* CONFIG_SMP */
2739
2740 ttwu_queue(p, cpu);
2741stat:
2742 ttwu_stat(p, cpu, wake_flags);
2743out:
2744 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2745
2746 return success;
2747}
2748
2749/**
2750 * try_to_wake_up_local - try to wake up a local task with rq lock held
2751 * @p: the thread to be awakened
2752 *
2753 * Put @p on the run-queue if it's not already there. The caller must
2754 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2755 * the current task.
2756 */
2757static void try_to_wake_up_local(struct task_struct *p)
2758{
2759 struct rq *rq = task_rq(p);
2760
2761 BUG_ON(rq != this_rq());
2762 BUG_ON(p == current);
2763 lockdep_assert_held(&rq->lock);
2764
2765 if (!raw_spin_trylock(&p->pi_lock)) {
2766 raw_spin_unlock(&rq->lock);
2767 raw_spin_lock(&p->pi_lock);
2768 raw_spin_lock(&rq->lock);
2769 }
2770
2771 if (!(p->state & TASK_NORMAL))
2772 goto out;
2773
2774 if (!p->on_rq)
2775 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2776
2777 ttwu_do_wakeup(rq, p, 0);
2778 ttwu_stat(p, smp_processor_id(), 0);
2779out:
2780 raw_spin_unlock(&p->pi_lock);
2781}
2782
2783/**
2784 * wake_up_process - Wake up a specific process
2785 * @p: The process to be woken up.
2786 *
2787 * Attempt to wake up the nominated process and move it to the set of runnable
2788 * processes. Returns 1 if the process was woken up, 0 if it was already
2789 * running.
2790 *
2791 * It may be assumed that this function implies a write memory barrier before
2792 * changing the task state if and only if any tasks are woken up.
2793 */
2794int wake_up_process(struct task_struct *p)
2795{
2796 return try_to_wake_up(p, TASK_ALL, 0);
2797}
2798EXPORT_SYMBOL(wake_up_process);
2799
2800int wake_up_state(struct task_struct *p, unsigned int state)
2801{
2802 return try_to_wake_up(p, state, 0);
2803}
2804
2805/*
2806 * Perform scheduler related setup for a newly forked process p.
2807 * p is forked by current.
2808 *
2809 * __sched_fork() is basic setup used by init_idle() too:
2810 */
2811static void __sched_fork(struct task_struct *p)
2812{
2813 p->on_rq = 0;
2814
2815 p->se.on_rq = 0;
2816 p->se.exec_start = 0;
2817 p->se.sum_exec_runtime = 0;
2818 p->se.prev_sum_exec_runtime = 0;
2819 p->se.nr_migrations = 0;
2820 p->se.vruntime = 0;
2821 INIT_LIST_HEAD(&p->se.group_node);
2822
2823#ifdef CONFIG_SCHEDSTATS
2824 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2825#endif
2826
2827 INIT_LIST_HEAD(&p->rt.run_list);
2828
2829#ifdef CONFIG_PREEMPT_NOTIFIERS
2830 INIT_HLIST_HEAD(&p->preempt_notifiers);
2831#endif
2832}
2833
2834/*
2835 * fork()/clone()-time setup:
2836 */
2837void sched_fork(struct task_struct *p)
2838{
2839 unsigned long flags;
2840 int cpu = get_cpu();
2841
2842 __sched_fork(p);
2843 /*
2844 * We mark the process as running here. This guarantees that
2845 * nobody will actually run it, and a signal or other external
2846 * event cannot wake it up and insert it on the runqueue either.
2847 */
2848 p->state = TASK_RUNNING;
2849
2850 /*
2851 * Revert to default priority/policy on fork if requested.
2852 */
2853 if (unlikely(p->sched_reset_on_fork)) {
2854 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2855 p->policy = SCHED_NORMAL;
2856 p->normal_prio = p->static_prio;
2857 }
2858
2859 if (PRIO_TO_NICE(p->static_prio) < 0) {
2860 p->static_prio = NICE_TO_PRIO(0);
2861 p->normal_prio = p->static_prio;
2862 set_load_weight(p);
2863 }
2864
2865 /*
2866 * We don't need the reset flag anymore after the fork. It has
2867 * fulfilled its duty:
2868 */
2869 p->sched_reset_on_fork = 0;
2870 }
2871
2872 /*
2873 * Make sure we do not leak PI boosting priority to the child.
2874 */
2875 p->prio = current->normal_prio;
2876
2877 if (!rt_prio(p->prio))
2878 p->sched_class = &fair_sched_class;
2879
2880 if (p->sched_class->task_fork)
2881 p->sched_class->task_fork(p);
2882
2883 /*
2884 * The child is not yet in the pid-hash so no cgroup attach races,
2885 * and the cgroup is pinned to this child due to cgroup_fork()
2886 * is ran before sched_fork().
2887 *
2888 * Silence PROVE_RCU.
2889 */
2890 raw_spin_lock_irqsave(&p->pi_lock, flags);
2891 set_task_cpu(p, cpu);
2892 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2893
2894#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2895 if (likely(sched_info_on()))
2896 memset(&p->sched_info, 0, sizeof(p->sched_info));
2897#endif
2898#if defined(CONFIG_SMP)
2899 p->on_cpu = 0;
2900#endif
2901#ifdef CONFIG_PREEMPT_COUNT
2902 /* Want to start with kernel preemption disabled. */
2903 task_thread_info(p)->preempt_count = 1;
2904#endif
2905#ifdef CONFIG_SMP
2906 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2907#endif
2908
2909 put_cpu();
2910}
2911
2912/*
2913 * wake_up_new_task - wake up a newly created task for the first time.
2914 *
2915 * This function will do some initial scheduler statistics housekeeping
2916 * that must be done for every newly created context, then puts the task
2917 * on the runqueue and wakes it.
2918 */
2919void wake_up_new_task(struct task_struct *p)
2920{
2921 unsigned long flags;
2922 struct rq *rq;
2923
2924 raw_spin_lock_irqsave(&p->pi_lock, flags);
2925#ifdef CONFIG_SMP
2926 /*
2927 * Fork balancing, do it here and not earlier because:
2928 * - cpus_allowed can change in the fork path
2929 * - any previously selected cpu might disappear through hotplug
2930 */
2931 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
2932#endif
2933
2934 rq = __task_rq_lock(p);
2935 activate_task(rq, p, 0);
2936 p->on_rq = 1;
2937 trace_sched_wakeup_new(p, true);
2938 check_preempt_curr(rq, p, WF_FORK);
2939#ifdef CONFIG_SMP
2940 if (p->sched_class->task_woken)
2941 p->sched_class->task_woken(rq, p);
2942#endif
2943 task_rq_unlock(rq, p, &flags);
2944}
2945
2946#ifdef CONFIG_PREEMPT_NOTIFIERS
2947
2948/**
2949 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2950 * @notifier: notifier struct to register
2951 */
2952void preempt_notifier_register(struct preempt_notifier *notifier)
2953{
2954 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2955}
2956EXPORT_SYMBOL_GPL(preempt_notifier_register);
2957
2958/**
2959 * preempt_notifier_unregister - no longer interested in preemption notifications
2960 * @notifier: notifier struct to unregister
2961 *
2962 * This is safe to call from within a preemption notifier.
2963 */
2964void preempt_notifier_unregister(struct preempt_notifier *notifier)
2965{
2966 hlist_del(&notifier->link);
2967}
2968EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2969
2970static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2971{
2972 struct preempt_notifier *notifier;
2973 struct hlist_node *node;
2974
2975 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2976 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2977}
2978
2979static void
2980fire_sched_out_preempt_notifiers(struct task_struct *curr,
2981 struct task_struct *next)
2982{
2983 struct preempt_notifier *notifier;
2984 struct hlist_node *node;
2985
2986 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2987 notifier->ops->sched_out(notifier, next);
2988}
2989
2990#else /* !CONFIG_PREEMPT_NOTIFIERS */
2991
2992static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2993{
2994}
2995
2996static void
2997fire_sched_out_preempt_notifiers(struct task_struct *curr,
2998 struct task_struct *next)
2999{
3000}
3001
3002#endif /* CONFIG_PREEMPT_NOTIFIERS */
3003
3004/**
3005 * prepare_task_switch - prepare to switch tasks
3006 * @rq: the runqueue preparing to switch
3007 * @prev: the current task that is being switched out
3008 * @next: the task we are going to switch to.
3009 *
3010 * This is called with the rq lock held and interrupts off. It must
3011 * be paired with a subsequent finish_task_switch after the context
3012 * switch.
3013 *
3014 * prepare_task_switch sets up locking and calls architecture specific
3015 * hooks.
3016 */
3017static inline void
3018prepare_task_switch(struct rq *rq, struct task_struct *prev,
3019 struct task_struct *next)
3020{
3021 sched_info_switch(prev, next);
3022 perf_event_task_sched_out(prev, next);
3023 fire_sched_out_preempt_notifiers(prev, next);
3024 prepare_lock_switch(rq, next);
3025 prepare_arch_switch(next);
3026 trace_sched_switch(prev, next);
3027}
3028
3029/**
3030 * finish_task_switch - clean up after a task-switch
3031 * @rq: runqueue associated with task-switch
3032 * @prev: the thread we just switched away from.
3033 *
3034 * finish_task_switch must be called after the context switch, paired
3035 * with a prepare_task_switch call before the context switch.
3036 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3037 * and do any other architecture-specific cleanup actions.
3038 *
3039 * Note that we may have delayed dropping an mm in context_switch(). If
3040 * so, we finish that here outside of the runqueue lock. (Doing it
3041 * with the lock held can cause deadlocks; see schedule() for
3042 * details.)
3043 */
3044static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3045 __releases(rq->lock)
3046{
3047 struct mm_struct *mm = rq->prev_mm;
3048 long prev_state;
3049
3050 rq->prev_mm = NULL;
3051
3052 /*
3053 * A task struct has one reference for the use as "current".
3054 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3055 * schedule one last time. The schedule call will never return, and
3056 * the scheduled task must drop that reference.
3057 * The test for TASK_DEAD must occur while the runqueue locks are
3058 * still held, otherwise prev could be scheduled on another cpu, die
3059 * there before we look at prev->state, and then the reference would
3060 * be dropped twice.
3061 * Manfred Spraul <[email protected]>
3062 */
3063 prev_state = prev->state;
3064 finish_arch_switch(prev);
3065#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3066 local_irq_disable();
3067#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3068 perf_event_task_sched_in(prev, current);
3069#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3070 local_irq_enable();
3071#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3072 finish_lock_switch(rq, prev);
3073
3074 fire_sched_in_preempt_notifiers(current);
3075 if (mm)
3076 mmdrop(mm);
3077 if (unlikely(prev_state == TASK_DEAD)) {
3078 /*
3079 * Remove function-return probe instances associated with this
3080 * task and put them back on the free list.
3081 */
3082 kprobe_flush_task(prev);
3083 put_task_struct(prev);
3084 }
3085}
3086
3087#ifdef CONFIG_SMP
3088
3089/* assumes rq->lock is held */
3090static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3091{
3092 if (prev->sched_class->pre_schedule)
3093 prev->sched_class->pre_schedule(rq, prev);
3094}
3095
3096/* rq->lock is NOT held, but preemption is disabled */
3097static inline void post_schedule(struct rq *rq)
3098{
3099 if (rq->post_schedule) {
3100 unsigned long flags;
3101
3102 raw_spin_lock_irqsave(&rq->lock, flags);
3103 if (rq->curr->sched_class->post_schedule)
3104 rq->curr->sched_class->post_schedule(rq);
3105 raw_spin_unlock_irqrestore(&rq->lock, flags);
3106
3107 rq->post_schedule = 0;
3108 }
3109}
3110
3111#else
3112
3113static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3114{
3115}
3116
3117static inline void post_schedule(struct rq *rq)
3118{
3119}
3120
3121#endif
3122
3123/**
3124 * schedule_tail - first thing a freshly forked thread must call.
3125 * @prev: the thread we just switched away from.
3126 */
3127asmlinkage void schedule_tail(struct task_struct *prev)
3128 __releases(rq->lock)
3129{
3130 struct rq *rq = this_rq();
3131
3132 finish_task_switch(rq, prev);
3133
3134 /*
3135 * FIXME: do we need to worry about rq being invalidated by the
3136 * task_switch?
3137 */
3138 post_schedule(rq);
3139
3140#ifdef __ARCH_WANT_UNLOCKED_CTXSW
3141 /* In this case, finish_task_switch does not reenable preemption */
3142 preempt_enable();
3143#endif
3144 if (current->set_child_tid)
3145 put_user(task_pid_vnr(current), current->set_child_tid);
3146}
3147
3148/*
3149 * context_switch - switch to the new MM and the new
3150 * thread's register state.
3151 */
3152static inline void
3153context_switch(struct rq *rq, struct task_struct *prev,
3154 struct task_struct *next)
3155{
3156 struct mm_struct *mm, *oldmm;
3157
3158 prepare_task_switch(rq, prev, next);
3159
3160 mm = next->mm;
3161 oldmm = prev->active_mm;
3162 /*
3163 * For paravirt, this is coupled with an exit in switch_to to
3164 * combine the page table reload and the switch backend into
3165 * one hypercall.
3166 */
3167 arch_start_context_switch(prev);
3168
3169 if (!mm) {
3170 next->active_mm = oldmm;
3171 atomic_inc(&oldmm->mm_count);
3172 enter_lazy_tlb(oldmm, next);
3173 } else
3174 switch_mm(oldmm, mm, next);
3175
3176 if (!prev->mm) {
3177 prev->active_mm = NULL;
3178 rq->prev_mm = oldmm;
3179 }
3180 /*
3181 * Since the runqueue lock will be released by the next
3182 * task (which is an invalid locking op but in the case
3183 * of the scheduler it's an obvious special-case), so we
3184 * do an early lockdep release here:
3185 */
3186#ifndef __ARCH_WANT_UNLOCKED_CTXSW
3187 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3188#endif
3189
3190 /* Here we just switch the register state and the stack. */
3191 switch_to(prev, next, prev);
3192
3193 barrier();
3194 /*
3195 * this_rq must be evaluated again because prev may have moved
3196 * CPUs since it called schedule(), thus the 'rq' on its stack
3197 * frame will be invalid.
3198 */
3199 finish_task_switch(this_rq(), prev);
3200}
3201
3202/*
3203 * nr_running, nr_uninterruptible and nr_context_switches:
3204 *
3205 * externally visible scheduler statistics: current number of runnable
3206 * threads, current number of uninterruptible-sleeping threads, total
3207 * number of context switches performed since bootup.
3208 */
3209unsigned long nr_running(void)
3210{
3211 unsigned long i, sum = 0;
3212
3213 for_each_online_cpu(i)
3214 sum += cpu_rq(i)->nr_running;
3215
3216 return sum;
3217}
3218
3219unsigned long nr_uninterruptible(void)
3220{
3221 unsigned long i, sum = 0;
3222
3223 for_each_possible_cpu(i)
3224 sum += cpu_rq(i)->nr_uninterruptible;
3225
3226 /*
3227 * Since we read the counters lockless, it might be slightly
3228 * inaccurate. Do not allow it to go below zero though:
3229 */
3230 if (unlikely((long)sum < 0))
3231 sum = 0;
3232
3233 return sum;
3234}
3235
3236unsigned long long nr_context_switches(void)
3237{
3238 int i;
3239 unsigned long long sum = 0;
3240
3241 for_each_possible_cpu(i)
3242 sum += cpu_rq(i)->nr_switches;
3243
3244 return sum;
3245}
3246
3247unsigned long nr_iowait(void)
3248{
3249 unsigned long i, sum = 0;
3250
3251 for_each_possible_cpu(i)
3252 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3253
3254 return sum;
3255}
3256
3257unsigned long nr_iowait_cpu(int cpu)
3258{
3259 struct rq *this = cpu_rq(cpu);
3260 return atomic_read(&this->nr_iowait);
3261}
3262
3263unsigned long this_cpu_load(void)
3264{
3265 struct rq *this = this_rq();
3266 return this->cpu_load[0];
3267}
3268
3269
3270/* Variables and functions for calc_load */
3271static atomic_long_t calc_load_tasks;
3272static unsigned long calc_load_update;
3273unsigned long avenrun[3];
3274EXPORT_SYMBOL(avenrun);
3275
3276static long calc_load_fold_active(struct rq *this_rq)
3277{
3278 long nr_active, delta = 0;
3279
3280 nr_active = this_rq->nr_running;
3281 nr_active += (long) this_rq->nr_uninterruptible;
3282
3283 if (nr_active != this_rq->calc_load_active) {
3284 delta = nr_active - this_rq->calc_load_active;
3285 this_rq->calc_load_active = nr_active;
3286 }
3287
3288 return delta;
3289}
3290
3291static unsigned long
3292calc_load(unsigned long load, unsigned long exp, unsigned long active)
3293{
3294 load *= exp;
3295 load += active * (FIXED_1 - exp);
3296 load += 1UL << (FSHIFT - 1);
3297 return load >> FSHIFT;
3298}
3299
3300#ifdef CONFIG_NO_HZ
3301/*
3302 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3303 *
3304 * When making the ILB scale, we should try to pull this in as well.
3305 */
3306static atomic_long_t calc_load_tasks_idle;
3307
3308static void calc_load_account_idle(struct rq *this_rq)
3309{
3310 long delta;
3311
3312 delta = calc_load_fold_active(this_rq);
3313 if (delta)
3314 atomic_long_add(delta, &calc_load_tasks_idle);
3315}
3316
3317static long calc_load_fold_idle(void)
3318{
3319 long delta = 0;
3320
3321 /*
3322 * Its got a race, we don't care...
3323 */
3324 if (atomic_long_read(&calc_load_tasks_idle))
3325 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3326
3327 return delta;
3328}
3329
3330/**
3331 * fixed_power_int - compute: x^n, in O(log n) time
3332 *
3333 * @x: base of the power
3334 * @frac_bits: fractional bits of @x
3335 * @n: power to raise @x to.
3336 *
3337 * By exploiting the relation between the definition of the natural power
3338 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3339 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3340 * (where: n_i \elem {0, 1}, the binary vector representing n),
3341 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3342 * of course trivially computable in O(log_2 n), the length of our binary
3343 * vector.
3344 */
3345static unsigned long
3346fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3347{
3348 unsigned long result = 1UL << frac_bits;
3349
3350 if (n) for (;;) {
3351 if (n & 1) {
3352 result *= x;
3353 result += 1UL << (frac_bits - 1);
3354 result >>= frac_bits;
3355 }
3356 n >>= 1;
3357 if (!n)
3358 break;
3359 x *= x;
3360 x += 1UL << (frac_bits - 1);
3361 x >>= frac_bits;
3362 }
3363
3364 return result;
3365}
3366
3367/*
3368 * a1 = a0 * e + a * (1 - e)
3369 *
3370 * a2 = a1 * e + a * (1 - e)
3371 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3372 * = a0 * e^2 + a * (1 - e) * (1 + e)
3373 *
3374 * a3 = a2 * e + a * (1 - e)
3375 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3376 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3377 *
3378 * ...
3379 *
3380 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3381 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3382 * = a0 * e^n + a * (1 - e^n)
3383 *
3384 * [1] application of the geometric series:
3385 *
3386 * n 1 - x^(n+1)
3387 * S_n := \Sum x^i = -------------
3388 * i=0 1 - x
3389 */
3390static unsigned long
3391calc_load_n(unsigned long load, unsigned long exp,
3392 unsigned long active, unsigned int n)
3393{
3394
3395 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3396}
3397
3398/*
3399 * NO_HZ can leave us missing all per-cpu ticks calling
3400 * calc_load_account_active(), but since an idle CPU folds its delta into
3401 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3402 * in the pending idle delta if our idle period crossed a load cycle boundary.
3403 *
3404 * Once we've updated the global active value, we need to apply the exponential
3405 * weights adjusted to the number of cycles missed.
3406 */
3407static void calc_global_nohz(unsigned long ticks)
3408{
3409 long delta, active, n;
3410
3411 if (time_before(jiffies, calc_load_update))
3412 return;
3413
3414 /*
3415 * If we crossed a calc_load_update boundary, make sure to fold
3416 * any pending idle changes, the respective CPUs might have
3417 * missed the tick driven calc_load_account_active() update
3418 * due to NO_HZ.
3419 */
3420 delta = calc_load_fold_idle();
3421 if (delta)
3422 atomic_long_add(delta, &calc_load_tasks);
3423
3424 /*
3425 * If we were idle for multiple load cycles, apply them.
3426 */
3427 if (ticks >= LOAD_FREQ) {
3428 n = ticks / LOAD_FREQ;
3429
3430 active = atomic_long_read(&calc_load_tasks);
3431 active = active > 0 ? active * FIXED_1 : 0;
3432
3433 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3434 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3435 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3436
3437 calc_load_update += n * LOAD_FREQ;
3438 }
3439
3440 /*
3441 * Its possible the remainder of the above division also crosses
3442 * a LOAD_FREQ period, the regular check in calc_global_load()
3443 * which comes after this will take care of that.
3444 *
3445 * Consider us being 11 ticks before a cycle completion, and us
3446 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3447 * age us 4 cycles, and the test in calc_global_load() will
3448 * pick up the final one.
3449 */
3450}
3451#else
3452static void calc_load_account_idle(struct rq *this_rq)
3453{
3454}
3455
3456static inline long calc_load_fold_idle(void)
3457{
3458 return 0;
3459}
3460
3461static void calc_global_nohz(unsigned long ticks)
3462{
3463}
3464#endif
3465
3466/**
3467 * get_avenrun - get the load average array
3468 * @loads: pointer to dest load array
3469 * @offset: offset to add
3470 * @shift: shift count to shift the result left
3471 *
3472 * These values are estimates at best, so no need for locking.
3473 */
3474void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3475{
3476 loads[0] = (avenrun[0] + offset) << shift;
3477 loads[1] = (avenrun[1] + offset) << shift;
3478 loads[2] = (avenrun[2] + offset) << shift;
3479}
3480
3481/*
3482 * calc_load - update the avenrun load estimates 10 ticks after the
3483 * CPUs have updated calc_load_tasks.
3484 */
3485void calc_global_load(unsigned long ticks)
3486{
3487 long active;
3488
3489 calc_global_nohz(ticks);
3490
3491 if (time_before(jiffies, calc_load_update + 10))
3492 return;
3493
3494 active = atomic_long_read(&calc_load_tasks);
3495 active = active > 0 ? active * FIXED_1 : 0;
3496
3497 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3498 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3499 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3500
3501 calc_load_update += LOAD_FREQ;
3502}
3503
3504/*
3505 * Called from update_cpu_load() to periodically update this CPU's
3506 * active count.
3507 */
3508static void calc_load_account_active(struct rq *this_rq)
3509{
3510 long delta;
3511
3512 if (time_before(jiffies, this_rq->calc_load_update))
3513 return;
3514
3515 delta = calc_load_fold_active(this_rq);
3516 delta += calc_load_fold_idle();
3517 if (delta)
3518 atomic_long_add(delta, &calc_load_tasks);
3519
3520 this_rq->calc_load_update += LOAD_FREQ;
3521}
3522
3523/*
3524 * The exact cpuload at various idx values, calculated at every tick would be
3525 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3526 *
3527 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3528 * on nth tick when cpu may be busy, then we have:
3529 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3530 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3531 *
3532 * decay_load_missed() below does efficient calculation of
3533 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3534 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3535 *
3536 * The calculation is approximated on a 128 point scale.
3537 * degrade_zero_ticks is the number of ticks after which load at any
3538 * particular idx is approximated to be zero.
3539 * degrade_factor is a precomputed table, a row for each load idx.
3540 * Each column corresponds to degradation factor for a power of two ticks,
3541 * based on 128 point scale.
3542 * Example:
3543 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3544 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3545 *
3546 * With this power of 2 load factors, we can degrade the load n times
3547 * by looking at 1 bits in n and doing as many mult/shift instead of
3548 * n mult/shifts needed by the exact degradation.
3549 */
3550#define DEGRADE_SHIFT 7
3551static const unsigned char
3552 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3553static const unsigned char
3554 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3555 {0, 0, 0, 0, 0, 0, 0, 0},
3556 {64, 32, 8, 0, 0, 0, 0, 0},
3557 {96, 72, 40, 12, 1, 0, 0},
3558 {112, 98, 75, 43, 15, 1, 0},
3559 {120, 112, 98, 76, 45, 16, 2} };
3560
3561/*
3562 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3563 * would be when CPU is idle and so we just decay the old load without
3564 * adding any new load.
3565 */
3566static unsigned long
3567decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3568{
3569 int j = 0;
3570
3571 if (!missed_updates)
3572 return load;
3573
3574 if (missed_updates >= degrade_zero_ticks[idx])
3575 return 0;
3576
3577 if (idx == 1)
3578 return load >> missed_updates;
3579
3580 while (missed_updates) {
3581 if (missed_updates % 2)
3582 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3583
3584 missed_updates >>= 1;
3585 j++;
3586 }
3587 return load;
3588}
3589
3590/*
3591 * Update rq->cpu_load[] statistics. This function is usually called every
3592 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3593 * every tick. We fix it up based on jiffies.
3594 */
3595static void update_cpu_load(struct rq *this_rq)
3596{
3597 unsigned long this_load = this_rq->load.weight;
3598 unsigned long curr_jiffies = jiffies;
3599 unsigned long pending_updates;
3600 int i, scale;
3601
3602 this_rq->nr_load_updates++;
3603
3604 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3605 if (curr_jiffies == this_rq->last_load_update_tick)
3606 return;
3607
3608 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3609 this_rq->last_load_update_tick = curr_jiffies;
3610
3611 /* Update our load: */
3612 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3613 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3614 unsigned long old_load, new_load;
3615
3616 /* scale is effectively 1 << i now, and >> i divides by scale */
3617
3618 old_load = this_rq->cpu_load[i];
3619 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3620 new_load = this_load;
3621 /*
3622 * Round up the averaging division if load is increasing. This
3623 * prevents us from getting stuck on 9 if the load is 10, for
3624 * example.
3625 */
3626 if (new_load > old_load)
3627 new_load += scale - 1;
3628
3629 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3630 }
3631
3632 sched_avg_update(this_rq);
3633}
3634
3635static void update_cpu_load_active(struct rq *this_rq)
3636{
3637 update_cpu_load(this_rq);
3638
3639 calc_load_account_active(this_rq);
3640}
3641
3642#ifdef CONFIG_SMP
3643
3644/*
3645 * sched_exec - execve() is a valuable balancing opportunity, because at
3646 * this point the task has the smallest effective memory and cache footprint.
3647 */
3648void sched_exec(void)
3649{
3650 struct task_struct *p = current;
3651 unsigned long flags;
3652 int dest_cpu;
3653
3654 raw_spin_lock_irqsave(&p->pi_lock, flags);
3655 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3656 if (dest_cpu == smp_processor_id())
3657 goto unlock;
3658
3659 if (likely(cpu_active(dest_cpu))) {
3660 struct migration_arg arg = { p, dest_cpu };
3661
3662 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3663 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3664 return;
3665 }
3666unlock:
3667 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3668}
3669
3670#endif
3671
3672DEFINE_PER_CPU(struct kernel_stat, kstat);
3673
3674EXPORT_PER_CPU_SYMBOL(kstat);
3675
3676/*
3677 * Return any ns on the sched_clock that have not yet been accounted in
3678 * @p in case that task is currently running.
3679 *
3680 * Called with task_rq_lock() held on @rq.
3681 */
3682static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3683{
3684 u64 ns = 0;
3685
3686 if (task_current(rq, p)) {
3687 update_rq_clock(rq);
3688 ns = rq->clock_task - p->se.exec_start;
3689 if ((s64)ns < 0)
3690 ns = 0;
3691 }
3692
3693 return ns;
3694}
3695
3696unsigned long long task_delta_exec(struct task_struct *p)
3697{
3698 unsigned long flags;
3699 struct rq *rq;
3700 u64 ns = 0;
3701
3702 rq = task_rq_lock(p, &flags);
3703 ns = do_task_delta_exec(p, rq);
3704 task_rq_unlock(rq, p, &flags);
3705
3706 return ns;
3707}
3708
3709/*
3710 * Return accounted runtime for the task.
3711 * In case the task is currently running, return the runtime plus current's
3712 * pending runtime that have not been accounted yet.
3713 */
3714unsigned long long task_sched_runtime(struct task_struct *p)
3715{
3716 unsigned long flags;
3717 struct rq *rq;
3718 u64 ns = 0;
3719
3720 rq = task_rq_lock(p, &flags);
3721 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3722 task_rq_unlock(rq, p, &flags);
3723
3724 return ns;
3725}
3726
3727/*
3728 * Account user cpu time to a process.
3729 * @p: the process that the cpu time gets accounted to
3730 * @cputime: the cpu time spent in user space since the last update
3731 * @cputime_scaled: cputime scaled by cpu frequency
3732 */
3733void account_user_time(struct task_struct *p, cputime_t cputime,
3734 cputime_t cputime_scaled)
3735{
3736 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3737 cputime64_t tmp;
3738
3739 /* Add user time to process. */
3740 p->utime = cputime_add(p->utime, cputime);
3741 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3742 account_group_user_time(p, cputime);
3743
3744 /* Add user time to cpustat. */
3745 tmp = cputime_to_cputime64(cputime);
3746 if (TASK_NICE(p) > 0)
3747 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3748 else
3749 cpustat->user = cputime64_add(cpustat->user, tmp);
3750
3751 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3752 /* Account for user time used */
3753 acct_update_integrals(p);
3754}
3755
3756/*
3757 * Account guest cpu time to a process.
3758 * @p: the process that the cpu time gets accounted to
3759 * @cputime: the cpu time spent in virtual machine since the last update
3760 * @cputime_scaled: cputime scaled by cpu frequency
3761 */
3762static void account_guest_time(struct task_struct *p, cputime_t cputime,
3763 cputime_t cputime_scaled)
3764{
3765 cputime64_t tmp;
3766 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3767
3768 tmp = cputime_to_cputime64(cputime);
3769
3770 /* Add guest time to process. */
3771 p->utime = cputime_add(p->utime, cputime);
3772 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3773 account_group_user_time(p, cputime);
3774 p->gtime = cputime_add(p->gtime, cputime);
3775
3776 /* Add guest time to cpustat. */
3777 if (TASK_NICE(p) > 0) {
3778 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3779 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3780 } else {
3781 cpustat->user = cputime64_add(cpustat->user, tmp);
3782 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3783 }
3784}
3785
3786/*
3787 * Account system cpu time to a process and desired cpustat field
3788 * @p: the process that the cpu time gets accounted to
3789 * @cputime: the cpu time spent in kernel space since the last update
3790 * @cputime_scaled: cputime scaled by cpu frequency
3791 * @target_cputime64: pointer to cpustat field that has to be updated
3792 */
3793static inline
3794void __account_system_time(struct task_struct *p, cputime_t cputime,
3795 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3796{
3797 cputime64_t tmp = cputime_to_cputime64(cputime);
3798
3799 /* Add system time to process. */
3800 p->stime = cputime_add(p->stime, cputime);
3801 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3802 account_group_system_time(p, cputime);
3803
3804 /* Add system time to cpustat. */
3805 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3806 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3807
3808 /* Account for system time used */
3809 acct_update_integrals(p);
3810}
3811
3812/*
3813 * Account system cpu time to a process.
3814 * @p: the process that the cpu time gets accounted to
3815 * @hardirq_offset: the offset to subtract from hardirq_count()
3816 * @cputime: the cpu time spent in kernel space since the last update
3817 * @cputime_scaled: cputime scaled by cpu frequency
3818 */
3819void account_system_time(struct task_struct *p, int hardirq_offset,
3820 cputime_t cputime, cputime_t cputime_scaled)
3821{
3822 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3823 cputime64_t *target_cputime64;
3824
3825 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3826 account_guest_time(p, cputime, cputime_scaled);
3827 return;
3828 }
3829
3830 if (hardirq_count() - hardirq_offset)
3831 target_cputime64 = &cpustat->irq;
3832 else if (in_serving_softirq())
3833 target_cputime64 = &cpustat->softirq;
3834 else
3835 target_cputime64 = &cpustat->system;
3836
3837 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3838}
3839
3840/*
3841 * Account for involuntary wait time.
3842 * @cputime: the cpu time spent in involuntary wait
3843 */
3844void account_steal_time(cputime_t cputime)
3845{
3846 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3847 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3848
3849 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3850}
3851
3852/*
3853 * Account for idle time.
3854 * @cputime: the cpu time spent in idle wait
3855 */
3856void account_idle_time(cputime_t cputime)
3857{
3858 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3859 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3860 struct rq *rq = this_rq();
3861
3862 if (atomic_read(&rq->nr_iowait) > 0)
3863 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3864 else
3865 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3866}
3867
3868static __always_inline bool steal_account_process_tick(void)
3869{
3870#ifdef CONFIG_PARAVIRT
3871 if (static_branch(&paravirt_steal_enabled)) {
3872 u64 steal, st = 0;
3873
3874 steal = paravirt_steal_clock(smp_processor_id());
3875 steal -= this_rq()->prev_steal_time;
3876
3877 st = steal_ticks(steal);
3878 this_rq()->prev_steal_time += st * TICK_NSEC;
3879
3880 account_steal_time(st);
3881 return st;
3882 }
3883#endif
3884 return false;
3885}
3886
3887#ifndef CONFIG_VIRT_CPU_ACCOUNTING
3888
3889#ifdef CONFIG_IRQ_TIME_ACCOUNTING
3890/*
3891 * Account a tick to a process and cpustat
3892 * @p: the process that the cpu time gets accounted to
3893 * @user_tick: is the tick from userspace
3894 * @rq: the pointer to rq
3895 *
3896 * Tick demultiplexing follows the order
3897 * - pending hardirq update
3898 * - pending softirq update
3899 * - user_time
3900 * - idle_time
3901 * - system time
3902 * - check for guest_time
3903 * - else account as system_time
3904 *
3905 * Check for hardirq is done both for system and user time as there is
3906 * no timer going off while we are on hardirq and hence we may never get an
3907 * opportunity to update it solely in system time.
3908 * p->stime and friends are only updated on system time and not on irq
3909 * softirq as those do not count in task exec_runtime any more.
3910 */
3911static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3912 struct rq *rq)
3913{
3914 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3915 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3916 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3917
3918 if (steal_account_process_tick())
3919 return;
3920
3921 if (irqtime_account_hi_update()) {
3922 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3923 } else if (irqtime_account_si_update()) {
3924 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3925 } else if (this_cpu_ksoftirqd() == p) {
3926 /*
3927 * ksoftirqd time do not get accounted in cpu_softirq_time.
3928 * So, we have to handle it separately here.
3929 * Also, p->stime needs to be updated for ksoftirqd.
3930 */
3931 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3932 &cpustat->softirq);
3933 } else if (user_tick) {
3934 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3935 } else if (p == rq->idle) {
3936 account_idle_time(cputime_one_jiffy);
3937 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3938 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3939 } else {
3940 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3941 &cpustat->system);
3942 }
3943}
3944
3945static void irqtime_account_idle_ticks(int ticks)
3946{
3947 int i;
3948 struct rq *rq = this_rq();
3949
3950 for (i = 0; i < ticks; i++)
3951 irqtime_account_process_tick(current, 0, rq);
3952}
3953#else /* CONFIG_IRQ_TIME_ACCOUNTING */
3954static void irqtime_account_idle_ticks(int ticks) {}
3955static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3956 struct rq *rq) {}
3957#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3958
3959/*
3960 * Account a single tick of cpu time.
3961 * @p: the process that the cpu time gets accounted to
3962 * @user_tick: indicates if the tick is a user or a system tick
3963 */
3964void account_process_tick(struct task_struct *p, int user_tick)
3965{
3966 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3967 struct rq *rq = this_rq();
3968
3969 if (sched_clock_irqtime) {
3970 irqtime_account_process_tick(p, user_tick, rq);
3971 return;
3972 }
3973
3974 if (steal_account_process_tick())
3975 return;
3976
3977 if (user_tick)
3978 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3979 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3980 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3981 one_jiffy_scaled);
3982 else
3983 account_idle_time(cputime_one_jiffy);
3984}
3985
3986/*
3987 * Account multiple ticks of steal time.
3988 * @p: the process from which the cpu time has been stolen
3989 * @ticks: number of stolen ticks
3990 */
3991void account_steal_ticks(unsigned long ticks)
3992{
3993 account_steal_time(jiffies_to_cputime(ticks));
3994}
3995
3996/*
3997 * Account multiple ticks of idle time.
3998 * @ticks: number of stolen ticks
3999 */
4000void account_idle_ticks(unsigned long ticks)
4001{
4002
4003 if (sched_clock_irqtime) {
4004 irqtime_account_idle_ticks(ticks);
4005 return;
4006 }
4007
4008 account_idle_time(jiffies_to_cputime(ticks));
4009}
4010
4011#endif
4012
4013/*
4014 * Use precise platform statistics if available:
4015 */
4016#ifdef CONFIG_VIRT_CPU_ACCOUNTING
4017void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4018{
4019 *ut = p->utime;
4020 *st = p->stime;
4021}
4022
4023void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4024{
4025 struct task_cputime cputime;
4026
4027 thread_group_cputime(p, &cputime);
4028
4029 *ut = cputime.utime;
4030 *st = cputime.stime;
4031}
4032#else
4033
4034#ifndef nsecs_to_cputime
4035# define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4036#endif
4037
4038void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4039{
4040 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4041
4042 /*
4043 * Use CFS's precise accounting:
4044 */
4045 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4046
4047 if (total) {
4048 u64 temp = rtime;
4049
4050 temp *= utime;
4051 do_div(temp, total);
4052 utime = (cputime_t)temp;
4053 } else
4054 utime = rtime;
4055
4056 /*
4057 * Compare with previous values, to keep monotonicity:
4058 */
4059 p->prev_utime = max(p->prev_utime, utime);
4060 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4061
4062 *ut = p->prev_utime;
4063 *st = p->prev_stime;
4064}
4065
4066/*
4067 * Must be called with siglock held.
4068 */
4069void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4070{
4071 struct signal_struct *sig = p->signal;
4072 struct task_cputime cputime;
4073 cputime_t rtime, utime, total;
4074
4075 thread_group_cputime(p, &cputime);
4076
4077 total = cputime_add(cputime.utime, cputime.stime);
4078 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4079
4080 if (total) {
4081 u64 temp = rtime;
4082
4083 temp *= cputime.utime;
4084 do_div(temp, total);
4085 utime = (cputime_t)temp;
4086 } else
4087 utime = rtime;
4088
4089 sig->prev_utime = max(sig->prev_utime, utime);
4090 sig->prev_stime = max(sig->prev_stime,
4091 cputime_sub(rtime, sig->prev_utime));
4092
4093 *ut = sig->prev_utime;
4094 *st = sig->prev_stime;
4095}
4096#endif
4097
4098/*
4099 * This function gets called by the timer code, with HZ frequency.
4100 * We call it with interrupts disabled.
4101 */
4102void scheduler_tick(void)
4103{
4104 int cpu = smp_processor_id();
4105 struct rq *rq = cpu_rq(cpu);
4106 struct task_struct *curr = rq->curr;
4107
4108 sched_clock_tick();
4109
4110 raw_spin_lock(&rq->lock);
4111 update_rq_clock(rq);
4112 update_cpu_load_active(rq);
4113 curr->sched_class->task_tick(rq, curr, 0);
4114 raw_spin_unlock(&rq->lock);
4115
4116 perf_event_task_tick();
4117
4118#ifdef CONFIG_SMP
4119 rq->idle_at_tick = idle_cpu(cpu);
4120 trigger_load_balance(rq, cpu);
4121#endif
4122}
4123
4124notrace unsigned long get_parent_ip(unsigned long addr)
4125{
4126 if (in_lock_functions(addr)) {
4127 addr = CALLER_ADDR2;
4128 if (in_lock_functions(addr))
4129 addr = CALLER_ADDR3;
4130 }
4131 return addr;
4132}
4133
4134#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4135 defined(CONFIG_PREEMPT_TRACER))
4136
4137void __kprobes add_preempt_count(int val)
4138{
4139#ifdef CONFIG_DEBUG_PREEMPT
4140 /*
4141 * Underflow?
4142 */
4143 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4144 return;
4145#endif
4146 preempt_count() += val;
4147#ifdef CONFIG_DEBUG_PREEMPT
4148 /*
4149 * Spinlock count overflowing soon?
4150 */
4151 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4152 PREEMPT_MASK - 10);
4153#endif
4154 if (preempt_count() == val)
4155 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4156}
4157EXPORT_SYMBOL(add_preempt_count);
4158
4159void __kprobes sub_preempt_count(int val)
4160{
4161#ifdef CONFIG_DEBUG_PREEMPT
4162 /*
4163 * Underflow?
4164 */
4165 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4166 return;
4167 /*
4168 * Is the spinlock portion underflowing?
4169 */
4170 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4171 !(preempt_count() & PREEMPT_MASK)))
4172 return;
4173#endif
4174
4175 if (preempt_count() == val)
4176 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4177 preempt_count() -= val;
4178}
4179EXPORT_SYMBOL(sub_preempt_count);
4180
4181#endif
4182
4183/*
4184 * Print scheduling while atomic bug:
4185 */
4186static noinline void __schedule_bug(struct task_struct *prev)
4187{
4188 struct pt_regs *regs = get_irq_regs();
4189
4190 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4191 prev->comm, prev->pid, preempt_count());
4192
4193 debug_show_held_locks(prev);
4194 print_modules();
4195 if (irqs_disabled())
4196 print_irqtrace_events(prev);
4197
4198 if (regs)
4199 show_regs(regs);
4200 else
4201 dump_stack();
4202}
4203
4204/*
4205 * Various schedule()-time debugging checks and statistics:
4206 */
4207static inline void schedule_debug(struct task_struct *prev)
4208{
4209 /*
4210 * Test if we are atomic. Since do_exit() needs to call into
4211 * schedule() atomically, we ignore that path for now.
4212 * Otherwise, whine if we are scheduling when we should not be.
4213 */
4214 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4215 __schedule_bug(prev);
4216
4217 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4218
4219 schedstat_inc(this_rq(), sched_count);
4220}
4221
4222static void put_prev_task(struct rq *rq, struct task_struct *prev)
4223{
4224 if (prev->on_rq || rq->skip_clock_update < 0)
4225 update_rq_clock(rq);
4226 prev->sched_class->put_prev_task(rq, prev);
4227}
4228
4229/*
4230 * Pick up the highest-prio task:
4231 */
4232static inline struct task_struct *
4233pick_next_task(struct rq *rq)
4234{
4235 const struct sched_class *class;
4236 struct task_struct *p;
4237
4238 /*
4239 * Optimization: we know that if all tasks are in
4240 * the fair class we can call that function directly:
4241 */
4242 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4243 p = fair_sched_class.pick_next_task(rq);
4244 if (likely(p))
4245 return p;
4246 }
4247
4248 for_each_class(class) {
4249 p = class->pick_next_task(rq);
4250 if (p)
4251 return p;
4252 }
4253
4254 BUG(); /* the idle class will always have a runnable task */
4255}
4256
4257/*
4258 * __schedule() is the main scheduler function.
4259 */
4260static void __sched __schedule(void)
4261{
4262 struct task_struct *prev, *next;
4263 unsigned long *switch_count;
4264 struct rq *rq;
4265 int cpu;
4266
4267need_resched:
4268 preempt_disable();
4269 cpu = smp_processor_id();
4270 rq = cpu_rq(cpu);
4271 rcu_note_context_switch(cpu);
4272 prev = rq->curr;
4273
4274 schedule_debug(prev);
4275
4276 if (sched_feat(HRTICK))
4277 hrtick_clear(rq);
4278
4279 raw_spin_lock_irq(&rq->lock);
4280
4281 switch_count = &prev->nivcsw;
4282 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4283 if (unlikely(signal_pending_state(prev->state, prev))) {
4284 prev->state = TASK_RUNNING;
4285 } else {
4286 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4287 prev->on_rq = 0;
4288
4289 /*
4290 * If a worker went to sleep, notify and ask workqueue
4291 * whether it wants to wake up a task to maintain
4292 * concurrency.
4293 */
4294 if (prev->flags & PF_WQ_WORKER) {
4295 struct task_struct *to_wakeup;
4296
4297 to_wakeup = wq_worker_sleeping(prev, cpu);
4298 if (to_wakeup)
4299 try_to_wake_up_local(to_wakeup);
4300 }
4301 }
4302 switch_count = &prev->nvcsw;
4303 }
4304
4305 pre_schedule(rq, prev);
4306
4307 if (unlikely(!rq->nr_running))
4308 idle_balance(cpu, rq);
4309
4310 put_prev_task(rq, prev);
4311 next = pick_next_task(rq);
4312 clear_tsk_need_resched(prev);
4313 rq->skip_clock_update = 0;
4314
4315 if (likely(prev != next)) {
4316 rq->nr_switches++;
4317 rq->curr = next;
4318 ++*switch_count;
4319
4320 context_switch(rq, prev, next); /* unlocks the rq */
4321 /*
4322 * The context switch have flipped the stack from under us
4323 * and restored the local variables which were saved when
4324 * this task called schedule() in the past. prev == current
4325 * is still correct, but it can be moved to another cpu/rq.
4326 */
4327 cpu = smp_processor_id();
4328 rq = cpu_rq(cpu);
4329 } else
4330 raw_spin_unlock_irq(&rq->lock);
4331
4332 post_schedule(rq);
4333
4334 preempt_enable_no_resched();
4335 if (need_resched())
4336 goto need_resched;
4337}
4338
4339static inline void sched_submit_work(struct task_struct *tsk)
4340{
4341 if (!tsk->state)
4342 return;
4343 /*
4344 * If we are going to sleep and we have plugged IO queued,
4345 * make sure to submit it to avoid deadlocks.
4346 */
4347 if (blk_needs_flush_plug(tsk))
4348 blk_schedule_flush_plug(tsk);
4349}
4350
4351asmlinkage void __sched schedule(void)
4352{
4353 struct task_struct *tsk = current;
4354
4355 sched_submit_work(tsk);
4356 __schedule();
4357}
4358EXPORT_SYMBOL(schedule);
4359
4360#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4361
4362static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4363{
4364 if (lock->owner != owner)
4365 return false;
4366
4367 /*
4368 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4369 * lock->owner still matches owner, if that fails, owner might
4370 * point to free()d memory, if it still matches, the rcu_read_lock()
4371 * ensures the memory stays valid.
4372 */
4373 barrier();
4374
4375 return owner->on_cpu;
4376}
4377
4378/*
4379 * Look out! "owner" is an entirely speculative pointer
4380 * access and not reliable.
4381 */
4382int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4383{
4384 if (!sched_feat(OWNER_SPIN))
4385 return 0;
4386
4387 rcu_read_lock();
4388 while (owner_running(lock, owner)) {
4389 if (need_resched())
4390 break;
4391
4392 arch_mutex_cpu_relax();
4393 }
4394 rcu_read_unlock();
4395
4396 /*
4397 * We break out the loop above on need_resched() and when the
4398 * owner changed, which is a sign for heavy contention. Return
4399 * success only when lock->owner is NULL.
4400 */
4401 return lock->owner == NULL;
4402}
4403#endif
4404
4405#ifdef CONFIG_PREEMPT
4406/*
4407 * this is the entry point to schedule() from in-kernel preemption
4408 * off of preempt_enable. Kernel preemptions off return from interrupt
4409 * occur there and call schedule directly.
4410 */
4411asmlinkage void __sched notrace preempt_schedule(void)
4412{
4413 struct thread_info *ti = current_thread_info();
4414
4415 /*
4416 * If there is a non-zero preempt_count or interrupts are disabled,
4417 * we do not want to preempt the current task. Just return..
4418 */
4419 if (likely(ti->preempt_count || irqs_disabled()))
4420 return;
4421
4422 do {
4423 add_preempt_count_notrace(PREEMPT_ACTIVE);
4424 __schedule();
4425 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4426
4427 /*
4428 * Check again in case we missed a preemption opportunity
4429 * between schedule and now.
4430 */
4431 barrier();
4432 } while (need_resched());
4433}
4434EXPORT_SYMBOL(preempt_schedule);
4435
4436/*
4437 * this is the entry point to schedule() from kernel preemption
4438 * off of irq context.
4439 * Note, that this is called and return with irqs disabled. This will
4440 * protect us against recursive calling from irq.
4441 */
4442asmlinkage void __sched preempt_schedule_irq(void)
4443{
4444 struct thread_info *ti = current_thread_info();
4445
4446 /* Catch callers which need to be fixed */
4447 BUG_ON(ti->preempt_count || !irqs_disabled());
4448
4449 do {
4450 add_preempt_count(PREEMPT_ACTIVE);
4451 local_irq_enable();
4452 __schedule();
4453 local_irq_disable();
4454 sub_preempt_count(PREEMPT_ACTIVE);
4455
4456 /*
4457 * Check again in case we missed a preemption opportunity
4458 * between schedule and now.
4459 */
4460 barrier();
4461 } while (need_resched());
4462}
4463
4464#endif /* CONFIG_PREEMPT */
4465
4466int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4467 void *key)
4468{
4469 return try_to_wake_up(curr->private, mode, wake_flags);
4470}
4471EXPORT_SYMBOL(default_wake_function);
4472
4473/*
4474 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4475 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4476 * number) then we wake all the non-exclusive tasks and one exclusive task.
4477 *
4478 * There are circumstances in which we can try to wake a task which has already
4479 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4480 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4481 */
4482static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4483 int nr_exclusive, int wake_flags, void *key)
4484{
4485 wait_queue_t *curr, *next;
4486
4487 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4488 unsigned flags = curr->flags;
4489
4490 if (curr->func(curr, mode, wake_flags, key) &&
4491 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4492 break;
4493 }
4494}
4495
4496/**
4497 * __wake_up - wake up threads blocked on a waitqueue.
4498 * @q: the waitqueue
4499 * @mode: which threads
4500 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4501 * @key: is directly passed to the wakeup function
4502 *
4503 * It may be assumed that this function implies a write memory barrier before
4504 * changing the task state if and only if any tasks are woken up.
4505 */
4506void __wake_up(wait_queue_head_t *q, unsigned int mode,
4507 int nr_exclusive, void *key)
4508{
4509 unsigned long flags;
4510
4511 spin_lock_irqsave(&q->lock, flags);
4512 __wake_up_common(q, mode, nr_exclusive, 0, key);
4513 spin_unlock_irqrestore(&q->lock, flags);
4514}
4515EXPORT_SYMBOL(__wake_up);
4516
4517/*
4518 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4519 */
4520void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4521{
4522 __wake_up_common(q, mode, 1, 0, NULL);
4523}
4524EXPORT_SYMBOL_GPL(__wake_up_locked);
4525
4526void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4527{
4528 __wake_up_common(q, mode, 1, 0, key);
4529}
4530EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4531
4532/**
4533 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4534 * @q: the waitqueue
4535 * @mode: which threads
4536 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4537 * @key: opaque value to be passed to wakeup targets
4538 *
4539 * The sync wakeup differs that the waker knows that it will schedule
4540 * away soon, so while the target thread will be woken up, it will not
4541 * be migrated to another CPU - ie. the two threads are 'synchronized'
4542 * with each other. This can prevent needless bouncing between CPUs.
4543 *
4544 * On UP it can prevent extra preemption.
4545 *
4546 * It may be assumed that this function implies a write memory barrier before
4547 * changing the task state if and only if any tasks are woken up.
4548 */
4549void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4550 int nr_exclusive, void *key)
4551{
4552 unsigned long flags;
4553 int wake_flags = WF_SYNC;
4554
4555 if (unlikely(!q))
4556 return;
4557
4558 if (unlikely(!nr_exclusive))
4559 wake_flags = 0;
4560
4561 spin_lock_irqsave(&q->lock, flags);
4562 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4563 spin_unlock_irqrestore(&q->lock, flags);
4564}
4565EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4566
4567/*
4568 * __wake_up_sync - see __wake_up_sync_key()
4569 */
4570void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4571{
4572 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4573}
4574EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4575
4576/**
4577 * complete: - signals a single thread waiting on this completion
4578 * @x: holds the state of this particular completion
4579 *
4580 * This will wake up a single thread waiting on this completion. Threads will be
4581 * awakened in the same order in which they were queued.
4582 *
4583 * See also complete_all(), wait_for_completion() and related routines.
4584 *
4585 * It may be assumed that this function implies a write memory barrier before
4586 * changing the task state if and only if any tasks are woken up.
4587 */
4588void complete(struct completion *x)
4589{
4590 unsigned long flags;
4591
4592 spin_lock_irqsave(&x->wait.lock, flags);
4593 x->done++;
4594 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4595 spin_unlock_irqrestore(&x->wait.lock, flags);
4596}
4597EXPORT_SYMBOL(complete);
4598
4599/**
4600 * complete_all: - signals all threads waiting on this completion
4601 * @x: holds the state of this particular completion
4602 *
4603 * This will wake up all threads waiting on this particular completion event.
4604 *
4605 * It may be assumed that this function implies a write memory barrier before
4606 * changing the task state if and only if any tasks are woken up.
4607 */
4608void complete_all(struct completion *x)
4609{
4610 unsigned long flags;
4611
4612 spin_lock_irqsave(&x->wait.lock, flags);
4613 x->done += UINT_MAX/2;
4614 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4615 spin_unlock_irqrestore(&x->wait.lock, flags);
4616}
4617EXPORT_SYMBOL(complete_all);
4618
4619static inline long __sched
4620do_wait_for_common(struct completion *x, long timeout, int state)
4621{
4622 if (!x->done) {
4623 DECLARE_WAITQUEUE(wait, current);
4624
4625 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4626 do {
4627 if (signal_pending_state(state, current)) {
4628 timeout = -ERESTARTSYS;
4629 break;
4630 }
4631 __set_current_state(state);
4632 spin_unlock_irq(&x->wait.lock);
4633 timeout = schedule_timeout(timeout);
4634 spin_lock_irq(&x->wait.lock);
4635 } while (!x->done && timeout);
4636 __remove_wait_queue(&x->wait, &wait);
4637 if (!x->done)
4638 return timeout;
4639 }
4640 x->done--;
4641 return timeout ?: 1;
4642}
4643
4644static long __sched
4645wait_for_common(struct completion *x, long timeout, int state)
4646{
4647 might_sleep();
4648
4649 spin_lock_irq(&x->wait.lock);
4650 timeout = do_wait_for_common(x, timeout, state);
4651 spin_unlock_irq(&x->wait.lock);
4652 return timeout;
4653}
4654
4655/**
4656 * wait_for_completion: - waits for completion of a task
4657 * @x: holds the state of this particular completion
4658 *
4659 * This waits to be signaled for completion of a specific task. It is NOT
4660 * interruptible and there is no timeout.
4661 *
4662 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4663 * and interrupt capability. Also see complete().
4664 */
4665void __sched wait_for_completion(struct completion *x)
4666{
4667 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4668}
4669EXPORT_SYMBOL(wait_for_completion);
4670
4671/**
4672 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4673 * @x: holds the state of this particular completion
4674 * @timeout: timeout value in jiffies
4675 *
4676 * This waits for either a completion of a specific task to be signaled or for a
4677 * specified timeout to expire. The timeout is in jiffies. It is not
4678 * interruptible.
4679 */
4680unsigned long __sched
4681wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4682{
4683 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4684}
4685EXPORT_SYMBOL(wait_for_completion_timeout);
4686
4687/**
4688 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4689 * @x: holds the state of this particular completion
4690 *
4691 * This waits for completion of a specific task to be signaled. It is
4692 * interruptible.
4693 */
4694int __sched wait_for_completion_interruptible(struct completion *x)
4695{
4696 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4697 if (t == -ERESTARTSYS)
4698 return t;
4699 return 0;
4700}
4701EXPORT_SYMBOL(wait_for_completion_interruptible);
4702
4703/**
4704 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4705 * @x: holds the state of this particular completion
4706 * @timeout: timeout value in jiffies
4707 *
4708 * This waits for either a completion of a specific task to be signaled or for a
4709 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4710 */
4711long __sched
4712wait_for_completion_interruptible_timeout(struct completion *x,
4713 unsigned long timeout)
4714{
4715 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4716}
4717EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4718
4719/**
4720 * wait_for_completion_killable: - waits for completion of a task (killable)
4721 * @x: holds the state of this particular completion
4722 *
4723 * This waits to be signaled for completion of a specific task. It can be
4724 * interrupted by a kill signal.
4725 */
4726int __sched wait_for_completion_killable(struct completion *x)
4727{
4728 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4729 if (t == -ERESTARTSYS)
4730 return t;
4731 return 0;
4732}
4733EXPORT_SYMBOL(wait_for_completion_killable);
4734
4735/**
4736 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4737 * @x: holds the state of this particular completion
4738 * @timeout: timeout value in jiffies
4739 *
4740 * This waits for either a completion of a specific task to be
4741 * signaled or for a specified timeout to expire. It can be
4742 * interrupted by a kill signal. The timeout is in jiffies.
4743 */
4744long __sched
4745wait_for_completion_killable_timeout(struct completion *x,
4746 unsigned long timeout)
4747{
4748 return wait_for_common(x, timeout, TASK_KILLABLE);
4749}
4750EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4751
4752/**
4753 * try_wait_for_completion - try to decrement a completion without blocking
4754 * @x: completion structure
4755 *
4756 * Returns: 0 if a decrement cannot be done without blocking
4757 * 1 if a decrement succeeded.
4758 *
4759 * If a completion is being used as a counting completion,
4760 * attempt to decrement the counter without blocking. This
4761 * enables us to avoid waiting if the resource the completion
4762 * is protecting is not available.
4763 */
4764bool try_wait_for_completion(struct completion *x)
4765{
4766 unsigned long flags;
4767 int ret = 1;
4768
4769 spin_lock_irqsave(&x->wait.lock, flags);
4770 if (!x->done)
4771 ret = 0;
4772 else
4773 x->done--;
4774 spin_unlock_irqrestore(&x->wait.lock, flags);
4775 return ret;
4776}
4777EXPORT_SYMBOL(try_wait_for_completion);
4778
4779/**
4780 * completion_done - Test to see if a completion has any waiters
4781 * @x: completion structure
4782 *
4783 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4784 * 1 if there are no waiters.
4785 *
4786 */
4787bool completion_done(struct completion *x)
4788{
4789 unsigned long flags;
4790 int ret = 1;
4791
4792 spin_lock_irqsave(&x->wait.lock, flags);
4793 if (!x->done)
4794 ret = 0;
4795 spin_unlock_irqrestore(&x->wait.lock, flags);
4796 return ret;
4797}
4798EXPORT_SYMBOL(completion_done);
4799
4800static long __sched
4801sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4802{
4803 unsigned long flags;
4804 wait_queue_t wait;
4805
4806 init_waitqueue_entry(&wait, current);
4807
4808 __set_current_state(state);
4809
4810 spin_lock_irqsave(&q->lock, flags);
4811 __add_wait_queue(q, &wait);
4812 spin_unlock(&q->lock);
4813 timeout = schedule_timeout(timeout);
4814 spin_lock_irq(&q->lock);
4815 __remove_wait_queue(q, &wait);
4816 spin_unlock_irqrestore(&q->lock, flags);
4817
4818 return timeout;
4819}
4820
4821void __sched interruptible_sleep_on(wait_queue_head_t *q)
4822{
4823 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4824}
4825EXPORT_SYMBOL(interruptible_sleep_on);
4826
4827long __sched
4828interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4829{
4830 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4831}
4832EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4833
4834void __sched sleep_on(wait_queue_head_t *q)
4835{
4836 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4837}
4838EXPORT_SYMBOL(sleep_on);
4839
4840long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4841{
4842 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4843}
4844EXPORT_SYMBOL(sleep_on_timeout);
4845
4846#ifdef CONFIG_RT_MUTEXES
4847
4848/*
4849 * rt_mutex_setprio - set the current priority of a task
4850 * @p: task
4851 * @prio: prio value (kernel-internal form)
4852 *
4853 * This function changes the 'effective' priority of a task. It does
4854 * not touch ->normal_prio like __setscheduler().
4855 *
4856 * Used by the rt_mutex code to implement priority inheritance logic.
4857 */
4858void rt_mutex_setprio(struct task_struct *p, int prio)
4859{
4860 int oldprio, on_rq, running;
4861 struct rq *rq;
4862 const struct sched_class *prev_class;
4863
4864 BUG_ON(prio < 0 || prio > MAX_PRIO);
4865
4866 rq = __task_rq_lock(p);
4867
4868 trace_sched_pi_setprio(p, prio);
4869 oldprio = p->prio;
4870 prev_class = p->sched_class;
4871 on_rq = p->on_rq;
4872 running = task_current(rq, p);
4873 if (on_rq)
4874 dequeue_task(rq, p, 0);
4875 if (running)
4876 p->sched_class->put_prev_task(rq, p);
4877
4878 if (rt_prio(prio))
4879 p->sched_class = &rt_sched_class;
4880 else
4881 p->sched_class = &fair_sched_class;
4882
4883 p->prio = prio;
4884
4885 if (running)
4886 p->sched_class->set_curr_task(rq);
4887 if (on_rq)
4888 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4889
4890 check_class_changed(rq, p, prev_class, oldprio);
4891 __task_rq_unlock(rq);
4892}
4893
4894#endif
4895
4896void set_user_nice(struct task_struct *p, long nice)
4897{
4898 int old_prio, delta, on_rq;
4899 unsigned long flags;
4900 struct rq *rq;
4901
4902 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4903 return;
4904 /*
4905 * We have to be careful, if called from sys_setpriority(),
4906 * the task might be in the middle of scheduling on another CPU.
4907 */
4908 rq = task_rq_lock(p, &flags);
4909 /*
4910 * The RT priorities are set via sched_setscheduler(), but we still
4911 * allow the 'normal' nice value to be set - but as expected
4912 * it wont have any effect on scheduling until the task is
4913 * SCHED_FIFO/SCHED_RR:
4914 */
4915 if (task_has_rt_policy(p)) {
4916 p->static_prio = NICE_TO_PRIO(nice);
4917 goto out_unlock;
4918 }
4919 on_rq = p->on_rq;
4920 if (on_rq)
4921 dequeue_task(rq, p, 0);
4922
4923 p->static_prio = NICE_TO_PRIO(nice);
4924 set_load_weight(p);
4925 old_prio = p->prio;
4926 p->prio = effective_prio(p);
4927 delta = p->prio - old_prio;
4928
4929 if (on_rq) {
4930 enqueue_task(rq, p, 0);
4931 /*
4932 * If the task increased its priority or is running and
4933 * lowered its priority, then reschedule its CPU:
4934 */
4935 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4936 resched_task(rq->curr);
4937 }
4938out_unlock:
4939 task_rq_unlock(rq, p, &flags);
4940}
4941EXPORT_SYMBOL(set_user_nice);
4942
4943/*
4944 * can_nice - check if a task can reduce its nice value
4945 * @p: task
4946 * @nice: nice value
4947 */
4948int can_nice(const struct task_struct *p, const int nice)
4949{
4950 /* convert nice value [19,-20] to rlimit style value [1,40] */
4951 int nice_rlim = 20 - nice;
4952
4953 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4954 capable(CAP_SYS_NICE));
4955}
4956
4957#ifdef __ARCH_WANT_SYS_NICE
4958
4959/*
4960 * sys_nice - change the priority of the current process.
4961 * @increment: priority increment
4962 *
4963 * sys_setpriority is a more generic, but much slower function that
4964 * does similar things.
4965 */
4966SYSCALL_DEFINE1(nice, int, increment)
4967{
4968 long nice, retval;
4969
4970 /*
4971 * Setpriority might change our priority at the same moment.
4972 * We don't have to worry. Conceptually one call occurs first
4973 * and we have a single winner.
4974 */
4975 if (increment < -40)
4976 increment = -40;
4977 if (increment > 40)
4978 increment = 40;
4979
4980 nice = TASK_NICE(current) + increment;
4981 if (nice < -20)
4982 nice = -20;
4983 if (nice > 19)
4984 nice = 19;
4985
4986 if (increment < 0 && !can_nice(current, nice))
4987 return -EPERM;
4988
4989 retval = security_task_setnice(current, nice);
4990 if (retval)
4991 return retval;
4992
4993 set_user_nice(current, nice);
4994 return 0;
4995}
4996
4997#endif
4998
4999/**
5000 * task_prio - return the priority value of a given task.
5001 * @p: the task in question.
5002 *
5003 * This is the priority value as seen by users in /proc.
5004 * RT tasks are offset by -200. Normal tasks are centered
5005 * around 0, value goes from -16 to +15.
5006 */
5007int task_prio(const struct task_struct *p)
5008{
5009 return p->prio - MAX_RT_PRIO;
5010}
5011
5012/**
5013 * task_nice - return the nice value of a given task.
5014 * @p: the task in question.
5015 */
5016int task_nice(const struct task_struct *p)
5017{
5018 return TASK_NICE(p);
5019}
5020EXPORT_SYMBOL(task_nice);
5021
5022/**
5023 * idle_cpu - is a given cpu idle currently?
5024 * @cpu: the processor in question.
5025 */
5026int idle_cpu(int cpu)
5027{
5028 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5029}
5030
5031/**
5032 * idle_task - return the idle task for a given cpu.
5033 * @cpu: the processor in question.
5034 */
5035struct task_struct *idle_task(int cpu)
5036{
5037 return cpu_rq(cpu)->idle;
5038}
5039
5040/**
5041 * find_process_by_pid - find a process with a matching PID value.
5042 * @pid: the pid in question.
5043 */
5044static struct task_struct *find_process_by_pid(pid_t pid)
5045{
5046 return pid ? find_task_by_vpid(pid) : current;
5047}
5048
5049/* Actually do priority change: must hold rq lock. */
5050static void
5051__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5052{
5053 p->policy = policy;
5054 p->rt_priority = prio;
5055 p->normal_prio = normal_prio(p);
5056 /* we are holding p->pi_lock already */
5057 p->prio = rt_mutex_getprio(p);
5058 if (rt_prio(p->prio))
5059 p->sched_class = &rt_sched_class;
5060 else
5061 p->sched_class = &fair_sched_class;
5062 set_load_weight(p);
5063}
5064
5065/*
5066 * check the target process has a UID that matches the current process's
5067 */
5068static bool check_same_owner(struct task_struct *p)
5069{
5070 const struct cred *cred = current_cred(), *pcred;
5071 bool match;
5072
5073 rcu_read_lock();
5074 pcred = __task_cred(p);
5075 if (cred->user->user_ns == pcred->user->user_ns)
5076 match = (cred->euid == pcred->euid ||
5077 cred->euid == pcred->uid);
5078 else
5079 match = false;
5080 rcu_read_unlock();
5081 return match;
5082}
5083
5084static int __sched_setscheduler(struct task_struct *p, int policy,
5085 const struct sched_param *param, bool user)
5086{
5087 int retval, oldprio, oldpolicy = -1, on_rq, running;
5088 unsigned long flags;
5089 const struct sched_class *prev_class;
5090 struct rq *rq;
5091 int reset_on_fork;
5092
5093 /* may grab non-irq protected spin_locks */
5094 BUG_ON(in_interrupt());
5095recheck:
5096 /* double check policy once rq lock held */
5097 if (policy < 0) {
5098 reset_on_fork = p->sched_reset_on_fork;
5099 policy = oldpolicy = p->policy;
5100 } else {
5101 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5102 policy &= ~SCHED_RESET_ON_FORK;
5103
5104 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5105 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5106 policy != SCHED_IDLE)
5107 return -EINVAL;
5108 }
5109
5110 /*
5111 * Valid priorities for SCHED_FIFO and SCHED_RR are
5112 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5113 * SCHED_BATCH and SCHED_IDLE is 0.
5114 */
5115 if (param->sched_priority < 0 ||
5116 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5117 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5118 return -EINVAL;
5119 if (rt_policy(policy) != (param->sched_priority != 0))
5120 return -EINVAL;
5121
5122 /*
5123 * Allow unprivileged RT tasks to decrease priority:
5124 */
5125 if (user && !capable(CAP_SYS_NICE)) {
5126 if (rt_policy(policy)) {
5127 unsigned long rlim_rtprio =
5128 task_rlimit(p, RLIMIT_RTPRIO);
5129
5130 /* can't set/change the rt policy */
5131 if (policy != p->policy && !rlim_rtprio)
5132 return -EPERM;
5133
5134 /* can't increase priority */
5135 if (param->sched_priority > p->rt_priority &&
5136 param->sched_priority > rlim_rtprio)
5137 return -EPERM;
5138 }
5139
5140 /*
5141 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5142 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5143 */
5144 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5145 if (!can_nice(p, TASK_NICE(p)))
5146 return -EPERM;
5147 }
5148
5149 /* can't change other user's priorities */
5150 if (!check_same_owner(p))
5151 return -EPERM;
5152
5153 /* Normal users shall not reset the sched_reset_on_fork flag */
5154 if (p->sched_reset_on_fork && !reset_on_fork)
5155 return -EPERM;
5156 }
5157
5158 if (user) {
5159 retval = security_task_setscheduler(p);
5160 if (retval)
5161 return retval;
5162 }
5163
5164 /*
5165 * make sure no PI-waiters arrive (or leave) while we are
5166 * changing the priority of the task:
5167 *
5168 * To be able to change p->policy safely, the appropriate
5169 * runqueue lock must be held.
5170 */
5171 rq = task_rq_lock(p, &flags);
5172
5173 /*
5174 * Changing the policy of the stop threads its a very bad idea
5175 */
5176 if (p == rq->stop) {
5177 task_rq_unlock(rq, p, &flags);
5178 return -EINVAL;
5179 }
5180
5181 /*
5182 * If not changing anything there's no need to proceed further:
5183 */
5184 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5185 param->sched_priority == p->rt_priority))) {
5186
5187 __task_rq_unlock(rq);
5188 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5189 return 0;
5190 }
5191
5192#ifdef CONFIG_RT_GROUP_SCHED
5193 if (user) {
5194 /*
5195 * Do not allow realtime tasks into groups that have no runtime
5196 * assigned.
5197 */
5198 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5199 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5200 !task_group_is_autogroup(task_group(p))) {
5201 task_rq_unlock(rq, p, &flags);
5202 return -EPERM;
5203 }
5204 }
5205#endif
5206
5207 /* recheck policy now with rq lock held */
5208 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5209 policy = oldpolicy = -1;
5210 task_rq_unlock(rq, p, &flags);
5211 goto recheck;
5212 }
5213 on_rq = p->on_rq;
5214 running = task_current(rq, p);
5215 if (on_rq)
5216 deactivate_task(rq, p, 0);
5217 if (running)
5218 p->sched_class->put_prev_task(rq, p);
5219
5220 p->sched_reset_on_fork = reset_on_fork;
5221
5222 oldprio = p->prio;
5223 prev_class = p->sched_class;
5224 __setscheduler(rq, p, policy, param->sched_priority);
5225
5226 if (running)
5227 p->sched_class->set_curr_task(rq);
5228 if (on_rq)
5229 activate_task(rq, p, 0);
5230
5231 check_class_changed(rq, p, prev_class, oldprio);
5232 task_rq_unlock(rq, p, &flags);
5233
5234 rt_mutex_adjust_pi(p);
5235
5236 return 0;
5237}
5238
5239/**
5240 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5241 * @p: the task in question.
5242 * @policy: new policy.
5243 * @param: structure containing the new RT priority.
5244 *
5245 * NOTE that the task may be already dead.
5246 */
5247int sched_setscheduler(struct task_struct *p, int policy,
5248 const struct sched_param *param)
5249{
5250 return __sched_setscheduler(p, policy, param, true);
5251}
5252EXPORT_SYMBOL_GPL(sched_setscheduler);
5253
5254/**
5255 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5256 * @p: the task in question.
5257 * @policy: new policy.
5258 * @param: structure containing the new RT priority.
5259 *
5260 * Just like sched_setscheduler, only don't bother checking if the
5261 * current context has permission. For example, this is needed in
5262 * stop_machine(): we create temporary high priority worker threads,
5263 * but our caller might not have that capability.
5264 */
5265int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5266 const struct sched_param *param)
5267{
5268 return __sched_setscheduler(p, policy, param, false);
5269}
5270
5271static int
5272do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5273{
5274 struct sched_param lparam;
5275 struct task_struct *p;
5276 int retval;
5277
5278 if (!param || pid < 0)
5279 return -EINVAL;
5280 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5281 return -EFAULT;
5282
5283 rcu_read_lock();
5284 retval = -ESRCH;
5285 p = find_process_by_pid(pid);
5286 if (p != NULL)
5287 retval = sched_setscheduler(p, policy, &lparam);
5288 rcu_read_unlock();
5289
5290 return retval;
5291}
5292
5293/**
5294 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5295 * @pid: the pid in question.
5296 * @policy: new policy.
5297 * @param: structure containing the new RT priority.
5298 */
5299SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5300 struct sched_param __user *, param)
5301{
5302 /* negative values for policy are not valid */
5303 if (policy < 0)
5304 return -EINVAL;
5305
5306 return do_sched_setscheduler(pid, policy, param);
5307}
5308
5309/**
5310 * sys_sched_setparam - set/change the RT priority of a thread
5311 * @pid: the pid in question.
5312 * @param: structure containing the new RT priority.
5313 */
5314SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5315{
5316 return do_sched_setscheduler(pid, -1, param);
5317}
5318
5319/**
5320 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5321 * @pid: the pid in question.
5322 */
5323SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5324{
5325 struct task_struct *p;
5326 int retval;
5327
5328 if (pid < 0)
5329 return -EINVAL;
5330
5331 retval = -ESRCH;
5332 rcu_read_lock();
5333 p = find_process_by_pid(pid);
5334 if (p) {
5335 retval = security_task_getscheduler(p);
5336 if (!retval)
5337 retval = p->policy
5338 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5339 }
5340 rcu_read_unlock();
5341 return retval;
5342}
5343
5344/**
5345 * sys_sched_getparam - get the RT priority of a thread
5346 * @pid: the pid in question.
5347 * @param: structure containing the RT priority.
5348 */
5349SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5350{
5351 struct sched_param lp;
5352 struct task_struct *p;
5353 int retval;
5354
5355 if (!param || pid < 0)
5356 return -EINVAL;
5357
5358 rcu_read_lock();
5359 p = find_process_by_pid(pid);
5360 retval = -ESRCH;
5361 if (!p)
5362 goto out_unlock;
5363
5364 retval = security_task_getscheduler(p);
5365 if (retval)
5366 goto out_unlock;
5367
5368 lp.sched_priority = p->rt_priority;
5369 rcu_read_unlock();
5370
5371 /*
5372 * This one might sleep, we cannot do it with a spinlock held ...
5373 */
5374 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5375
5376 return retval;
5377
5378out_unlock:
5379 rcu_read_unlock();
5380 return retval;
5381}
5382
5383long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5384{
5385 cpumask_var_t cpus_allowed, new_mask;
5386 struct task_struct *p;
5387 int retval;
5388
5389 get_online_cpus();
5390 rcu_read_lock();
5391
5392 p = find_process_by_pid(pid);
5393 if (!p) {
5394 rcu_read_unlock();
5395 put_online_cpus();
5396 return -ESRCH;
5397 }
5398
5399 /* Prevent p going away */
5400 get_task_struct(p);
5401 rcu_read_unlock();
5402
5403 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5404 retval = -ENOMEM;
5405 goto out_put_task;
5406 }
5407 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5408 retval = -ENOMEM;
5409 goto out_free_cpus_allowed;
5410 }
5411 retval = -EPERM;
5412 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5413 goto out_unlock;
5414
5415 retval = security_task_setscheduler(p);
5416 if (retval)
5417 goto out_unlock;
5418
5419 cpuset_cpus_allowed(p, cpus_allowed);
5420 cpumask_and(new_mask, in_mask, cpus_allowed);
5421again:
5422 retval = set_cpus_allowed_ptr(p, new_mask);
5423
5424 if (!retval) {
5425 cpuset_cpus_allowed(p, cpus_allowed);
5426 if (!cpumask_subset(new_mask, cpus_allowed)) {
5427 /*
5428 * We must have raced with a concurrent cpuset
5429 * update. Just reset the cpus_allowed to the
5430 * cpuset's cpus_allowed
5431 */
5432 cpumask_copy(new_mask, cpus_allowed);
5433 goto again;
5434 }
5435 }
5436out_unlock:
5437 free_cpumask_var(new_mask);
5438out_free_cpus_allowed:
5439 free_cpumask_var(cpus_allowed);
5440out_put_task:
5441 put_task_struct(p);
5442 put_online_cpus();
5443 return retval;
5444}
5445
5446static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5447 struct cpumask *new_mask)
5448{
5449 if (len < cpumask_size())
5450 cpumask_clear(new_mask);
5451 else if (len > cpumask_size())
5452 len = cpumask_size();
5453
5454 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5455}
5456
5457/**
5458 * sys_sched_setaffinity - set the cpu affinity of a process
5459 * @pid: pid of the process
5460 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5461 * @user_mask_ptr: user-space pointer to the new cpu mask
5462 */
5463SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5464 unsigned long __user *, user_mask_ptr)
5465{
5466 cpumask_var_t new_mask;
5467 int retval;
5468
5469 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5470 return -ENOMEM;
5471
5472 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5473 if (retval == 0)
5474 retval = sched_setaffinity(pid, new_mask);
5475 free_cpumask_var(new_mask);
5476 return retval;
5477}
5478
5479long sched_getaffinity(pid_t pid, struct cpumask *mask)
5480{
5481 struct task_struct *p;
5482 unsigned long flags;
5483 int retval;
5484
5485 get_online_cpus();
5486 rcu_read_lock();
5487
5488 retval = -ESRCH;
5489 p = find_process_by_pid(pid);
5490 if (!p)
5491 goto out_unlock;
5492
5493 retval = security_task_getscheduler(p);
5494 if (retval)
5495 goto out_unlock;
5496
5497 raw_spin_lock_irqsave(&p->pi_lock, flags);
5498 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5499 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5500
5501out_unlock:
5502 rcu_read_unlock();
5503 put_online_cpus();
5504
5505 return retval;
5506}
5507
5508/**
5509 * sys_sched_getaffinity - get the cpu affinity of a process
5510 * @pid: pid of the process
5511 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5512 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5513 */
5514SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5515 unsigned long __user *, user_mask_ptr)
5516{
5517 int ret;
5518 cpumask_var_t mask;
5519
5520 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5521 return -EINVAL;
5522 if (len & (sizeof(unsigned long)-1))
5523 return -EINVAL;
5524
5525 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5526 return -ENOMEM;
5527
5528 ret = sched_getaffinity(pid, mask);
5529 if (ret == 0) {
5530 size_t retlen = min_t(size_t, len, cpumask_size());
5531
5532 if (copy_to_user(user_mask_ptr, mask, retlen))
5533 ret = -EFAULT;
5534 else
5535 ret = retlen;
5536 }
5537 free_cpumask_var(mask);
5538
5539 return ret;
5540}
5541
5542/**
5543 * sys_sched_yield - yield the current processor to other threads.
5544 *
5545 * This function yields the current CPU to other tasks. If there are no
5546 * other threads running on this CPU then this function will return.
5547 */
5548SYSCALL_DEFINE0(sched_yield)
5549{
5550 struct rq *rq = this_rq_lock();
5551
5552 schedstat_inc(rq, yld_count);
5553 current->sched_class->yield_task(rq);
5554
5555 /*
5556 * Since we are going to call schedule() anyway, there's
5557 * no need to preempt or enable interrupts:
5558 */
5559 __release(rq->lock);
5560 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5561 do_raw_spin_unlock(&rq->lock);
5562 preempt_enable_no_resched();
5563
5564 schedule();
5565
5566 return 0;
5567}
5568
5569static inline int should_resched(void)
5570{
5571 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5572}
5573
5574static void __cond_resched(void)
5575{
5576 add_preempt_count(PREEMPT_ACTIVE);
5577 __schedule();
5578 sub_preempt_count(PREEMPT_ACTIVE);
5579}
5580
5581int __sched _cond_resched(void)
5582{
5583 if (should_resched()) {
5584 __cond_resched();
5585 return 1;
5586 }
5587 return 0;
5588}
5589EXPORT_SYMBOL(_cond_resched);
5590
5591/*
5592 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5593 * call schedule, and on return reacquire the lock.
5594 *
5595 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5596 * operations here to prevent schedule() from being called twice (once via
5597 * spin_unlock(), once by hand).
5598 */
5599int __cond_resched_lock(spinlock_t *lock)
5600{
5601 int resched = should_resched();
5602 int ret = 0;
5603
5604 lockdep_assert_held(lock);
5605
5606 if (spin_needbreak(lock) || resched) {
5607 spin_unlock(lock);
5608 if (resched)
5609 __cond_resched();
5610 else
5611 cpu_relax();
5612 ret = 1;
5613 spin_lock(lock);
5614 }
5615 return ret;
5616}
5617EXPORT_SYMBOL(__cond_resched_lock);
5618
5619int __sched __cond_resched_softirq(void)
5620{
5621 BUG_ON(!in_softirq());
5622
5623 if (should_resched()) {
5624 local_bh_enable();
5625 __cond_resched();
5626 local_bh_disable();
5627 return 1;
5628 }
5629 return 0;
5630}
5631EXPORT_SYMBOL(__cond_resched_softirq);
5632
5633/**
5634 * yield - yield the current processor to other threads.
5635 *
5636 * This is a shortcut for kernel-space yielding - it marks the
5637 * thread runnable and calls sys_sched_yield().
5638 */
5639void __sched yield(void)
5640{
5641 set_current_state(TASK_RUNNING);
5642 sys_sched_yield();
5643}
5644EXPORT_SYMBOL(yield);
5645
5646/**
5647 * yield_to - yield the current processor to another thread in
5648 * your thread group, or accelerate that thread toward the
5649 * processor it's on.
5650 * @p: target task
5651 * @preempt: whether task preemption is allowed or not
5652 *
5653 * It's the caller's job to ensure that the target task struct
5654 * can't go away on us before we can do any checks.
5655 *
5656 * Returns true if we indeed boosted the target task.
5657 */
5658bool __sched yield_to(struct task_struct *p, bool preempt)
5659{
5660 struct task_struct *curr = current;
5661 struct rq *rq, *p_rq;
5662 unsigned long flags;
5663 bool yielded = 0;
5664
5665 local_irq_save(flags);
5666 rq = this_rq();
5667
5668again:
5669 p_rq = task_rq(p);
5670 double_rq_lock(rq, p_rq);
5671 while (task_rq(p) != p_rq) {
5672 double_rq_unlock(rq, p_rq);
5673 goto again;
5674 }
5675
5676 if (!curr->sched_class->yield_to_task)
5677 goto out;
5678
5679 if (curr->sched_class != p->sched_class)
5680 goto out;
5681
5682 if (task_running(p_rq, p) || p->state)
5683 goto out;
5684
5685 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5686 if (yielded) {
5687 schedstat_inc(rq, yld_count);
5688 /*
5689 * Make p's CPU reschedule; pick_next_entity takes care of
5690 * fairness.
5691 */
5692 if (preempt && rq != p_rq)
5693 resched_task(p_rq->curr);
5694 }
5695
5696out:
5697 double_rq_unlock(rq, p_rq);
5698 local_irq_restore(flags);
5699
5700 if (yielded)
5701 schedule();
5702
5703 return yielded;
5704}
5705EXPORT_SYMBOL_GPL(yield_to);
5706
5707/*
5708 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5709 * that process accounting knows that this is a task in IO wait state.
5710 */
5711void __sched io_schedule(void)
5712{
5713 struct rq *rq = raw_rq();
5714
5715 delayacct_blkio_start();
5716 atomic_inc(&rq->nr_iowait);
5717 blk_flush_plug(current);
5718 current->in_iowait = 1;
5719 schedule();
5720 current->in_iowait = 0;
5721 atomic_dec(&rq->nr_iowait);
5722 delayacct_blkio_end();
5723}
5724EXPORT_SYMBOL(io_schedule);
5725
5726long __sched io_schedule_timeout(long timeout)
5727{
5728 struct rq *rq = raw_rq();
5729 long ret;
5730
5731 delayacct_blkio_start();
5732 atomic_inc(&rq->nr_iowait);
5733 blk_flush_plug(current);
5734 current->in_iowait = 1;
5735 ret = schedule_timeout(timeout);
5736 current->in_iowait = 0;
5737 atomic_dec(&rq->nr_iowait);
5738 delayacct_blkio_end();
5739 return ret;
5740}
5741
5742/**
5743 * sys_sched_get_priority_max - return maximum RT priority.
5744 * @policy: scheduling class.
5745 *
5746 * this syscall returns the maximum rt_priority that can be used
5747 * by a given scheduling class.
5748 */
5749SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5750{
5751 int ret = -EINVAL;
5752
5753 switch (policy) {
5754 case SCHED_FIFO:
5755 case SCHED_RR:
5756 ret = MAX_USER_RT_PRIO-1;
5757 break;
5758 case SCHED_NORMAL:
5759 case SCHED_BATCH:
5760 case SCHED_IDLE:
5761 ret = 0;
5762 break;
5763 }
5764 return ret;
5765}
5766
5767/**
5768 * sys_sched_get_priority_min - return minimum RT priority.
5769 * @policy: scheduling class.
5770 *
5771 * this syscall returns the minimum rt_priority that can be used
5772 * by a given scheduling class.
5773 */
5774SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5775{
5776 int ret = -EINVAL;
5777
5778 switch (policy) {
5779 case SCHED_FIFO:
5780 case SCHED_RR:
5781 ret = 1;
5782 break;
5783 case SCHED_NORMAL:
5784 case SCHED_BATCH:
5785 case SCHED_IDLE:
5786 ret = 0;
5787 }
5788 return ret;
5789}
5790
5791/**
5792 * sys_sched_rr_get_interval - return the default timeslice of a process.
5793 * @pid: pid of the process.
5794 * @interval: userspace pointer to the timeslice value.
5795 *
5796 * this syscall writes the default timeslice value of a given process
5797 * into the user-space timespec buffer. A value of '0' means infinity.
5798 */
5799SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5800 struct timespec __user *, interval)
5801{
5802 struct task_struct *p;
5803 unsigned int time_slice;
5804 unsigned long flags;
5805 struct rq *rq;
5806 int retval;
5807 struct timespec t;
5808
5809 if (pid < 0)
5810 return -EINVAL;
5811
5812 retval = -ESRCH;
5813 rcu_read_lock();
5814 p = find_process_by_pid(pid);
5815 if (!p)
5816 goto out_unlock;
5817
5818 retval = security_task_getscheduler(p);
5819 if (retval)
5820 goto out_unlock;
5821
5822 rq = task_rq_lock(p, &flags);
5823 time_slice = p->sched_class->get_rr_interval(rq, p);
5824 task_rq_unlock(rq, p, &flags);
5825
5826 rcu_read_unlock();
5827 jiffies_to_timespec(time_slice, &t);
5828 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5829 return retval;
5830
5831out_unlock:
5832 rcu_read_unlock();
5833 return retval;
5834}
5835
5836static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5837
5838void sched_show_task(struct task_struct *p)
5839{
5840 unsigned long free = 0;
5841 unsigned state;
5842
5843 state = p->state ? __ffs(p->state) + 1 : 0;
5844 printk(KERN_INFO "%-15.15s %c", p->comm,
5845 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5846#if BITS_PER_LONG == 32
5847 if (state == TASK_RUNNING)
5848 printk(KERN_CONT " running ");
5849 else
5850 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5851#else
5852 if (state == TASK_RUNNING)
5853 printk(KERN_CONT " running task ");
5854 else
5855 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5856#endif
5857#ifdef CONFIG_DEBUG_STACK_USAGE
5858 free = stack_not_used(p);
5859#endif
5860 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5861 task_pid_nr(p), task_pid_nr(p->real_parent),
5862 (unsigned long)task_thread_info(p)->flags);
5863
5864 show_stack(p, NULL);
5865}
5866
5867void show_state_filter(unsigned long state_filter)
5868{
5869 struct task_struct *g, *p;
5870
5871#if BITS_PER_LONG == 32
5872 printk(KERN_INFO
5873 " task PC stack pid father\n");
5874#else
5875 printk(KERN_INFO
5876 " task PC stack pid father\n");
5877#endif
5878 read_lock(&tasklist_lock);
5879 do_each_thread(g, p) {
5880 /*
5881 * reset the NMI-timeout, listing all files on a slow
5882 * console might take a lot of time:
5883 */
5884 touch_nmi_watchdog();
5885 if (!state_filter || (p->state & state_filter))
5886 sched_show_task(p);
5887 } while_each_thread(g, p);
5888
5889 touch_all_softlockup_watchdogs();
5890
5891#ifdef CONFIG_SCHED_DEBUG
5892 sysrq_sched_debug_show();
5893#endif
5894 read_unlock(&tasklist_lock);
5895 /*
5896 * Only show locks if all tasks are dumped:
5897 */
5898 if (!state_filter)
5899 debug_show_all_locks();
5900}
5901
5902void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5903{
5904 idle->sched_class = &idle_sched_class;
5905}
5906
5907/**
5908 * init_idle - set up an idle thread for a given CPU
5909 * @idle: task in question
5910 * @cpu: cpu the idle task belongs to
5911 *
5912 * NOTE: this function does not set the idle thread's NEED_RESCHED
5913 * flag, to make booting more robust.
5914 */
5915void __cpuinit init_idle(struct task_struct *idle, int cpu)
5916{
5917 struct rq *rq = cpu_rq(cpu);
5918 unsigned long flags;
5919
5920 raw_spin_lock_irqsave(&rq->lock, flags);
5921
5922 __sched_fork(idle);
5923 idle->state = TASK_RUNNING;
5924 idle->se.exec_start = sched_clock();
5925
5926 do_set_cpus_allowed(idle, cpumask_of(cpu));
5927 /*
5928 * We're having a chicken and egg problem, even though we are
5929 * holding rq->lock, the cpu isn't yet set to this cpu so the
5930 * lockdep check in task_group() will fail.
5931 *
5932 * Similar case to sched_fork(). / Alternatively we could
5933 * use task_rq_lock() here and obtain the other rq->lock.
5934 *
5935 * Silence PROVE_RCU
5936 */
5937 rcu_read_lock();
5938 __set_task_cpu(idle, cpu);
5939 rcu_read_unlock();
5940
5941 rq->curr = rq->idle = idle;
5942#if defined(CONFIG_SMP)
5943 idle->on_cpu = 1;
5944#endif
5945 raw_spin_unlock_irqrestore(&rq->lock, flags);
5946
5947 /* Set the preempt count _outside_ the spinlocks! */
5948 task_thread_info(idle)->preempt_count = 0;
5949
5950 /*
5951 * The idle tasks have their own, simple scheduling class:
5952 */
5953 idle->sched_class = &idle_sched_class;
5954 ftrace_graph_init_idle_task(idle, cpu);
5955}
5956
5957/*
5958 * In a system that switches off the HZ timer nohz_cpu_mask
5959 * indicates which cpus entered this state. This is used
5960 * in the rcu update to wait only for active cpus. For system
5961 * which do not switch off the HZ timer nohz_cpu_mask should
5962 * always be CPU_BITS_NONE.
5963 */
5964cpumask_var_t nohz_cpu_mask;
5965
5966/*
5967 * Increase the granularity value when there are more CPUs,
5968 * because with more CPUs the 'effective latency' as visible
5969 * to users decreases. But the relationship is not linear,
5970 * so pick a second-best guess by going with the log2 of the
5971 * number of CPUs.
5972 *
5973 * This idea comes from the SD scheduler of Con Kolivas:
5974 */
5975static int get_update_sysctl_factor(void)
5976{
5977 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5978 unsigned int factor;
5979
5980 switch (sysctl_sched_tunable_scaling) {
5981 case SCHED_TUNABLESCALING_NONE:
5982 factor = 1;
5983 break;
5984 case SCHED_TUNABLESCALING_LINEAR:
5985 factor = cpus;
5986 break;
5987 case SCHED_TUNABLESCALING_LOG:
5988 default:
5989 factor = 1 + ilog2(cpus);
5990 break;
5991 }
5992
5993 return factor;
5994}
5995
5996static void update_sysctl(void)
5997{
5998 unsigned int factor = get_update_sysctl_factor();
5999
6000#define SET_SYSCTL(name) \
6001 (sysctl_##name = (factor) * normalized_sysctl_##name)
6002 SET_SYSCTL(sched_min_granularity);
6003 SET_SYSCTL(sched_latency);
6004 SET_SYSCTL(sched_wakeup_granularity);
6005#undef SET_SYSCTL
6006}
6007
6008static inline void sched_init_granularity(void)
6009{
6010 update_sysctl();
6011}
6012
6013#ifdef CONFIG_SMP
6014void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
6015{
6016 if (p->sched_class && p->sched_class->set_cpus_allowed)
6017 p->sched_class->set_cpus_allowed(p, new_mask);
6018 else {
6019 cpumask_copy(&p->cpus_allowed, new_mask);
6020 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6021 }
6022}
6023
6024/*
6025 * This is how migration works:
6026 *
6027 * 1) we invoke migration_cpu_stop() on the target CPU using
6028 * stop_one_cpu().
6029 * 2) stopper starts to run (implicitly forcing the migrated thread
6030 * off the CPU)
6031 * 3) it checks whether the migrated task is still in the wrong runqueue.
6032 * 4) if it's in the wrong runqueue then the migration thread removes
6033 * it and puts it into the right queue.
6034 * 5) stopper completes and stop_one_cpu() returns and the migration
6035 * is done.
6036 */
6037
6038/*
6039 * Change a given task's CPU affinity. Migrate the thread to a
6040 * proper CPU and schedule it away if the CPU it's executing on
6041 * is removed from the allowed bitmask.
6042 *
6043 * NOTE: the caller must have a valid reference to the task, the
6044 * task must not exit() & deallocate itself prematurely. The
6045 * call is not atomic; no spinlocks may be held.
6046 */
6047int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6048{
6049 unsigned long flags;
6050 struct rq *rq;
6051 unsigned int dest_cpu;
6052 int ret = 0;
6053
6054 rq = task_rq_lock(p, &flags);
6055
6056 if (cpumask_equal(&p->cpus_allowed, new_mask))
6057 goto out;
6058
6059 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6060 ret = -EINVAL;
6061 goto out;
6062 }
6063
6064 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6065 ret = -EINVAL;
6066 goto out;
6067 }
6068
6069 do_set_cpus_allowed(p, new_mask);
6070
6071 /* Can the task run on the task's current CPU? If so, we're done */
6072 if (cpumask_test_cpu(task_cpu(p), new_mask))
6073 goto out;
6074
6075 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6076 if (p->on_rq) {
6077 struct migration_arg arg = { p, dest_cpu };
6078 /* Need help from migration thread: drop lock and wait. */
6079 task_rq_unlock(rq, p, &flags);
6080 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6081 tlb_migrate_finish(p->mm);
6082 return 0;
6083 }
6084out:
6085 task_rq_unlock(rq, p, &flags);
6086
6087 return ret;
6088}
6089EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6090
6091/*
6092 * Move (not current) task off this cpu, onto dest cpu. We're doing
6093 * this because either it can't run here any more (set_cpus_allowed()
6094 * away from this CPU, or CPU going down), or because we're
6095 * attempting to rebalance this task on exec (sched_exec).
6096 *
6097 * So we race with normal scheduler movements, but that's OK, as long
6098 * as the task is no longer on this CPU.
6099 *
6100 * Returns non-zero if task was successfully migrated.
6101 */
6102static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6103{
6104 struct rq *rq_dest, *rq_src;
6105 int ret = 0;
6106
6107 if (unlikely(!cpu_active(dest_cpu)))
6108 return ret;
6109
6110 rq_src = cpu_rq(src_cpu);
6111 rq_dest = cpu_rq(dest_cpu);
6112
6113 raw_spin_lock(&p->pi_lock);
6114 double_rq_lock(rq_src, rq_dest);
6115 /* Already moved. */
6116 if (task_cpu(p) != src_cpu)
6117 goto done;
6118 /* Affinity changed (again). */
6119 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6120 goto fail;
6121
6122 /*
6123 * If we're not on a rq, the next wake-up will ensure we're
6124 * placed properly.
6125 */
6126 if (p->on_rq) {
6127 deactivate_task(rq_src, p, 0);
6128 set_task_cpu(p, dest_cpu);
6129 activate_task(rq_dest, p, 0);
6130 check_preempt_curr(rq_dest, p, 0);
6131 }
6132done:
6133 ret = 1;
6134fail:
6135 double_rq_unlock(rq_src, rq_dest);
6136 raw_spin_unlock(&p->pi_lock);
6137 return ret;
6138}
6139
6140/*
6141 * migration_cpu_stop - this will be executed by a highprio stopper thread
6142 * and performs thread migration by bumping thread off CPU then
6143 * 'pushing' onto another runqueue.
6144 */
6145static int migration_cpu_stop(void *data)
6146{
6147 struct migration_arg *arg = data;
6148
6149 /*
6150 * The original target cpu might have gone down and we might
6151 * be on another cpu but it doesn't matter.
6152 */
6153 local_irq_disable();
6154 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6155 local_irq_enable();
6156 return 0;
6157}
6158
6159#ifdef CONFIG_HOTPLUG_CPU
6160
6161/*
6162 * Ensures that the idle task is using init_mm right before its cpu goes
6163 * offline.
6164 */
6165void idle_task_exit(void)
6166{
6167 struct mm_struct *mm = current->active_mm;
6168
6169 BUG_ON(cpu_online(smp_processor_id()));
6170
6171 if (mm != &init_mm)
6172 switch_mm(mm, &init_mm, current);
6173 mmdrop(mm);
6174}
6175
6176/*
6177 * While a dead CPU has no uninterruptible tasks queued at this point,
6178 * it might still have a nonzero ->nr_uninterruptible counter, because
6179 * for performance reasons the counter is not stricly tracking tasks to
6180 * their home CPUs. So we just add the counter to another CPU's counter,
6181 * to keep the global sum constant after CPU-down:
6182 */
6183static void migrate_nr_uninterruptible(struct rq *rq_src)
6184{
6185 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6186
6187 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6188 rq_src->nr_uninterruptible = 0;
6189}
6190
6191/*
6192 * remove the tasks which were accounted by rq from calc_load_tasks.
6193 */
6194static void calc_global_load_remove(struct rq *rq)
6195{
6196 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6197 rq->calc_load_active = 0;
6198}
6199
6200/*
6201 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6202 * try_to_wake_up()->select_task_rq().
6203 *
6204 * Called with rq->lock held even though we'er in stop_machine() and
6205 * there's no concurrency possible, we hold the required locks anyway
6206 * because of lock validation efforts.
6207 */
6208static void migrate_tasks(unsigned int dead_cpu)
6209{
6210 struct rq *rq = cpu_rq(dead_cpu);
6211 struct task_struct *next, *stop = rq->stop;
6212 int dest_cpu;
6213
6214 /*
6215 * Fudge the rq selection such that the below task selection loop
6216 * doesn't get stuck on the currently eligible stop task.
6217 *
6218 * We're currently inside stop_machine() and the rq is either stuck
6219 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6220 * either way we should never end up calling schedule() until we're
6221 * done here.
6222 */
6223 rq->stop = NULL;
6224
6225 for ( ; ; ) {
6226 /*
6227 * There's this thread running, bail when that's the only
6228 * remaining thread.
6229 */
6230 if (rq->nr_running == 1)
6231 break;
6232
6233 next = pick_next_task(rq);
6234 BUG_ON(!next);
6235 next->sched_class->put_prev_task(rq, next);
6236
6237 /* Find suitable destination for @next, with force if needed. */
6238 dest_cpu = select_fallback_rq(dead_cpu, next);
6239 raw_spin_unlock(&rq->lock);
6240
6241 __migrate_task(next, dead_cpu, dest_cpu);
6242
6243 raw_spin_lock(&rq->lock);
6244 }
6245
6246 rq->stop = stop;
6247}
6248
6249#endif /* CONFIG_HOTPLUG_CPU */
6250
6251#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6252
6253static struct ctl_table sd_ctl_dir[] = {
6254 {
6255 .procname = "sched_domain",
6256 .mode = 0555,
6257 },
6258 {}
6259};
6260
6261static struct ctl_table sd_ctl_root[] = {
6262 {
6263 .procname = "kernel",
6264 .mode = 0555,
6265 .child = sd_ctl_dir,
6266 },
6267 {}
6268};
6269
6270static struct ctl_table *sd_alloc_ctl_entry(int n)
6271{
6272 struct ctl_table *entry =
6273 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6274
6275 return entry;
6276}
6277
6278static void sd_free_ctl_entry(struct ctl_table **tablep)
6279{
6280 struct ctl_table *entry;
6281
6282 /*
6283 * In the intermediate directories, both the child directory and
6284 * procname are dynamically allocated and could fail but the mode
6285 * will always be set. In the lowest directory the names are
6286 * static strings and all have proc handlers.
6287 */
6288 for (entry = *tablep; entry->mode; entry++) {
6289 if (entry->child)
6290 sd_free_ctl_entry(&entry->child);
6291 if (entry->proc_handler == NULL)
6292 kfree(entry->procname);
6293 }
6294
6295 kfree(*tablep);
6296 *tablep = NULL;
6297}
6298
6299static void
6300set_table_entry(struct ctl_table *entry,
6301 const char *procname, void *data, int maxlen,
6302 mode_t mode, proc_handler *proc_handler)
6303{
6304 entry->procname = procname;
6305 entry->data = data;
6306 entry->maxlen = maxlen;
6307 entry->mode = mode;
6308 entry->proc_handler = proc_handler;
6309}
6310
6311static struct ctl_table *
6312sd_alloc_ctl_domain_table(struct sched_domain *sd)
6313{
6314 struct ctl_table *table = sd_alloc_ctl_entry(13);
6315
6316 if (table == NULL)
6317 return NULL;
6318
6319 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6320 sizeof(long), 0644, proc_doulongvec_minmax);
6321 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6322 sizeof(long), 0644, proc_doulongvec_minmax);
6323 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6324 sizeof(int), 0644, proc_dointvec_minmax);
6325 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6326 sizeof(int), 0644, proc_dointvec_minmax);
6327 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6328 sizeof(int), 0644, proc_dointvec_minmax);
6329 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6330 sizeof(int), 0644, proc_dointvec_minmax);
6331 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6332 sizeof(int), 0644, proc_dointvec_minmax);
6333 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6334 sizeof(int), 0644, proc_dointvec_minmax);
6335 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6336 sizeof(int), 0644, proc_dointvec_minmax);
6337 set_table_entry(&table[9], "cache_nice_tries",
6338 &sd->cache_nice_tries,
6339 sizeof(int), 0644, proc_dointvec_minmax);
6340 set_table_entry(&table[10], "flags", &sd->flags,
6341 sizeof(int), 0644, proc_dointvec_minmax);
6342 set_table_entry(&table[11], "name", sd->name,
6343 CORENAME_MAX_SIZE, 0444, proc_dostring);
6344 /* &table[12] is terminator */
6345
6346 return table;
6347}
6348
6349static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6350{
6351 struct ctl_table *entry, *table;
6352 struct sched_domain *sd;
6353 int domain_num = 0, i;
6354 char buf[32];
6355
6356 for_each_domain(cpu, sd)
6357 domain_num++;
6358 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6359 if (table == NULL)
6360 return NULL;
6361
6362 i = 0;
6363 for_each_domain(cpu, sd) {
6364 snprintf(buf, 32, "domain%d", i);
6365 entry->procname = kstrdup(buf, GFP_KERNEL);
6366 entry->mode = 0555;
6367 entry->child = sd_alloc_ctl_domain_table(sd);
6368 entry++;
6369 i++;
6370 }
6371 return table;
6372}
6373
6374static struct ctl_table_header *sd_sysctl_header;
6375static void register_sched_domain_sysctl(void)
6376{
6377 int i, cpu_num = num_possible_cpus();
6378 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6379 char buf[32];
6380
6381 WARN_ON(sd_ctl_dir[0].child);
6382 sd_ctl_dir[0].child = entry;
6383
6384 if (entry == NULL)
6385 return;
6386
6387 for_each_possible_cpu(i) {
6388 snprintf(buf, 32, "cpu%d", i);
6389 entry->procname = kstrdup(buf, GFP_KERNEL);
6390 entry->mode = 0555;
6391 entry->child = sd_alloc_ctl_cpu_table(i);
6392 entry++;
6393 }
6394
6395 WARN_ON(sd_sysctl_header);
6396 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6397}
6398
6399/* may be called multiple times per register */
6400static void unregister_sched_domain_sysctl(void)
6401{
6402 if (sd_sysctl_header)
6403 unregister_sysctl_table(sd_sysctl_header);
6404 sd_sysctl_header = NULL;
6405 if (sd_ctl_dir[0].child)
6406 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6407}
6408#else
6409static void register_sched_domain_sysctl(void)
6410{
6411}
6412static void unregister_sched_domain_sysctl(void)
6413{
6414}
6415#endif
6416
6417static void set_rq_online(struct rq *rq)
6418{
6419 if (!rq->online) {
6420 const struct sched_class *class;
6421
6422 cpumask_set_cpu(rq->cpu, rq->rd->online);
6423 rq->online = 1;
6424
6425 for_each_class(class) {
6426 if (class->rq_online)
6427 class->rq_online(rq);
6428 }
6429 }
6430}
6431
6432static void set_rq_offline(struct rq *rq)
6433{
6434 if (rq->online) {
6435 const struct sched_class *class;
6436
6437 for_each_class(class) {
6438 if (class->rq_offline)
6439 class->rq_offline(rq);
6440 }
6441
6442 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6443 rq->online = 0;
6444 }
6445}
6446
6447/*
6448 * migration_call - callback that gets triggered when a CPU is added.
6449 * Here we can start up the necessary migration thread for the new CPU.
6450 */
6451static int __cpuinit
6452migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6453{
6454 int cpu = (long)hcpu;
6455 unsigned long flags;
6456 struct rq *rq = cpu_rq(cpu);
6457
6458 switch (action & ~CPU_TASKS_FROZEN) {
6459
6460 case CPU_UP_PREPARE:
6461 rq->calc_load_update = calc_load_update;
6462 break;
6463
6464 case CPU_ONLINE:
6465 /* Update our root-domain */
6466 raw_spin_lock_irqsave(&rq->lock, flags);
6467 if (rq->rd) {
6468 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6469
6470 set_rq_online(rq);
6471 }
6472 raw_spin_unlock_irqrestore(&rq->lock, flags);
6473 break;
6474
6475#ifdef CONFIG_HOTPLUG_CPU
6476 case CPU_DYING:
6477 sched_ttwu_pending();
6478 /* Update our root-domain */
6479 raw_spin_lock_irqsave(&rq->lock, flags);
6480 if (rq->rd) {
6481 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6482 set_rq_offline(rq);
6483 }
6484 migrate_tasks(cpu);
6485 BUG_ON(rq->nr_running != 1); /* the migration thread */
6486 raw_spin_unlock_irqrestore(&rq->lock, flags);
6487
6488 migrate_nr_uninterruptible(rq);
6489 calc_global_load_remove(rq);
6490 break;
6491#endif
6492 }
6493
6494 update_max_interval();
6495
6496 return NOTIFY_OK;
6497}
6498
6499/*
6500 * Register at high priority so that task migration (migrate_all_tasks)
6501 * happens before everything else. This has to be lower priority than
6502 * the notifier in the perf_event subsystem, though.
6503 */
6504static struct notifier_block __cpuinitdata migration_notifier = {
6505 .notifier_call = migration_call,
6506 .priority = CPU_PRI_MIGRATION,
6507};
6508
6509static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6510 unsigned long action, void *hcpu)
6511{
6512 switch (action & ~CPU_TASKS_FROZEN) {
6513 case CPU_ONLINE:
6514 case CPU_DOWN_FAILED:
6515 set_cpu_active((long)hcpu, true);
6516 return NOTIFY_OK;
6517 default:
6518 return NOTIFY_DONE;
6519 }
6520}
6521
6522static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6523 unsigned long action, void *hcpu)
6524{
6525 switch (action & ~CPU_TASKS_FROZEN) {
6526 case CPU_DOWN_PREPARE:
6527 set_cpu_active((long)hcpu, false);
6528 return NOTIFY_OK;
6529 default:
6530 return NOTIFY_DONE;
6531 }
6532}
6533
6534static int __init migration_init(void)
6535{
6536 void *cpu = (void *)(long)smp_processor_id();
6537 int err;
6538
6539 /* Initialize migration for the boot CPU */
6540 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6541 BUG_ON(err == NOTIFY_BAD);
6542 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6543 register_cpu_notifier(&migration_notifier);
6544
6545 /* Register cpu active notifiers */
6546 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6547 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6548
6549 return 0;
6550}
6551early_initcall(migration_init);
6552#endif
6553
6554#ifdef CONFIG_SMP
6555
6556static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6557
6558#ifdef CONFIG_SCHED_DEBUG
6559
6560static __read_mostly int sched_domain_debug_enabled;
6561
6562static int __init sched_domain_debug_setup(char *str)
6563{
6564 sched_domain_debug_enabled = 1;
6565
6566 return 0;
6567}
6568early_param("sched_debug", sched_domain_debug_setup);
6569
6570static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6571 struct cpumask *groupmask)
6572{
6573 struct sched_group *group = sd->groups;
6574 char str[256];
6575
6576 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6577 cpumask_clear(groupmask);
6578
6579 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6580
6581 if (!(sd->flags & SD_LOAD_BALANCE)) {
6582 printk("does not load-balance\n");
6583 if (sd->parent)
6584 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6585 " has parent");
6586 return -1;
6587 }
6588
6589 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6590
6591 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6592 printk(KERN_ERR "ERROR: domain->span does not contain "
6593 "CPU%d\n", cpu);
6594 }
6595 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6596 printk(KERN_ERR "ERROR: domain->groups does not contain"
6597 " CPU%d\n", cpu);
6598 }
6599
6600 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6601 do {
6602 if (!group) {
6603 printk("\n");
6604 printk(KERN_ERR "ERROR: group is NULL\n");
6605 break;
6606 }
6607
6608 if (!group->sgp->power) {
6609 printk(KERN_CONT "\n");
6610 printk(KERN_ERR "ERROR: domain->cpu_power not "
6611 "set\n");
6612 break;
6613 }
6614
6615 if (!cpumask_weight(sched_group_cpus(group))) {
6616 printk(KERN_CONT "\n");
6617 printk(KERN_ERR "ERROR: empty group\n");
6618 break;
6619 }
6620
6621 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6622 printk(KERN_CONT "\n");
6623 printk(KERN_ERR "ERROR: repeated CPUs\n");
6624 break;
6625 }
6626
6627 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6628
6629 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6630
6631 printk(KERN_CONT " %s", str);
6632 if (group->sgp->power != SCHED_POWER_SCALE) {
6633 printk(KERN_CONT " (cpu_power = %d)",
6634 group->sgp->power);
6635 }
6636
6637 group = group->next;
6638 } while (group != sd->groups);
6639 printk(KERN_CONT "\n");
6640
6641 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6642 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6643
6644 if (sd->parent &&
6645 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6646 printk(KERN_ERR "ERROR: parent span is not a superset "
6647 "of domain->span\n");
6648 return 0;
6649}
6650
6651static void sched_domain_debug(struct sched_domain *sd, int cpu)
6652{
6653 int level = 0;
6654
6655 if (!sched_domain_debug_enabled)
6656 return;
6657
6658 if (!sd) {
6659 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6660 return;
6661 }
6662
6663 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6664
6665 for (;;) {
6666 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6667 break;
6668 level++;
6669 sd = sd->parent;
6670 if (!sd)
6671 break;
6672 }
6673}
6674#else /* !CONFIG_SCHED_DEBUG */
6675# define sched_domain_debug(sd, cpu) do { } while (0)
6676#endif /* CONFIG_SCHED_DEBUG */
6677
6678static int sd_degenerate(struct sched_domain *sd)
6679{
6680 if (cpumask_weight(sched_domain_span(sd)) == 1)
6681 return 1;
6682
6683 /* Following flags need at least 2 groups */
6684 if (sd->flags & (SD_LOAD_BALANCE |
6685 SD_BALANCE_NEWIDLE |
6686 SD_BALANCE_FORK |
6687 SD_BALANCE_EXEC |
6688 SD_SHARE_CPUPOWER |
6689 SD_SHARE_PKG_RESOURCES)) {
6690 if (sd->groups != sd->groups->next)
6691 return 0;
6692 }
6693
6694 /* Following flags don't use groups */
6695 if (sd->flags & (SD_WAKE_AFFINE))
6696 return 0;
6697
6698 return 1;
6699}
6700
6701static int
6702sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6703{
6704 unsigned long cflags = sd->flags, pflags = parent->flags;
6705
6706 if (sd_degenerate(parent))
6707 return 1;
6708
6709 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6710 return 0;
6711
6712 /* Flags needing groups don't count if only 1 group in parent */
6713 if (parent->groups == parent->groups->next) {
6714 pflags &= ~(SD_LOAD_BALANCE |
6715 SD_BALANCE_NEWIDLE |
6716 SD_BALANCE_FORK |
6717 SD_BALANCE_EXEC |
6718 SD_SHARE_CPUPOWER |
6719 SD_SHARE_PKG_RESOURCES);
6720 if (nr_node_ids == 1)
6721 pflags &= ~SD_SERIALIZE;
6722 }
6723 if (~cflags & pflags)
6724 return 0;
6725
6726 return 1;
6727}
6728
6729static void free_rootdomain(struct rcu_head *rcu)
6730{
6731 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6732
6733 cpupri_cleanup(&rd->cpupri);
6734 free_cpumask_var(rd->rto_mask);
6735 free_cpumask_var(rd->online);
6736 free_cpumask_var(rd->span);
6737 kfree(rd);
6738}
6739
6740static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6741{
6742 struct root_domain *old_rd = NULL;
6743 unsigned long flags;
6744
6745 raw_spin_lock_irqsave(&rq->lock, flags);
6746
6747 if (rq->rd) {
6748 old_rd = rq->rd;
6749
6750 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6751 set_rq_offline(rq);
6752
6753 cpumask_clear_cpu(rq->cpu, old_rd->span);
6754
6755 /*
6756 * If we dont want to free the old_rt yet then
6757 * set old_rd to NULL to skip the freeing later
6758 * in this function:
6759 */
6760 if (!atomic_dec_and_test(&old_rd->refcount))
6761 old_rd = NULL;
6762 }
6763
6764 atomic_inc(&rd->refcount);
6765 rq->rd = rd;
6766
6767 cpumask_set_cpu(rq->cpu, rd->span);
6768 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6769 set_rq_online(rq);
6770
6771 raw_spin_unlock_irqrestore(&rq->lock, flags);
6772
6773 if (old_rd)
6774 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6775}
6776
6777static int init_rootdomain(struct root_domain *rd)
6778{
6779 memset(rd, 0, sizeof(*rd));
6780
6781 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6782 goto out;
6783 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6784 goto free_span;
6785 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6786 goto free_online;
6787
6788 if (cpupri_init(&rd->cpupri) != 0)
6789 goto free_rto_mask;
6790 return 0;
6791
6792free_rto_mask:
6793 free_cpumask_var(rd->rto_mask);
6794free_online:
6795 free_cpumask_var(rd->online);
6796free_span:
6797 free_cpumask_var(rd->span);
6798out:
6799 return -ENOMEM;
6800}
6801
6802static void init_defrootdomain(void)
6803{
6804 init_rootdomain(&def_root_domain);
6805
6806 atomic_set(&def_root_domain.refcount, 1);
6807}
6808
6809static struct root_domain *alloc_rootdomain(void)
6810{
6811 struct root_domain *rd;
6812
6813 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6814 if (!rd)
6815 return NULL;
6816
6817 if (init_rootdomain(rd) != 0) {
6818 kfree(rd);
6819 return NULL;
6820 }
6821
6822 return rd;
6823}
6824
6825static void free_sched_groups(struct sched_group *sg, int free_sgp)
6826{
6827 struct sched_group *tmp, *first;
6828
6829 if (!sg)
6830 return;
6831
6832 first = sg;
6833 do {
6834 tmp = sg->next;
6835
6836 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
6837 kfree(sg->sgp);
6838
6839 kfree(sg);
6840 sg = tmp;
6841 } while (sg != first);
6842}
6843
6844static void free_sched_domain(struct rcu_head *rcu)
6845{
6846 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6847
6848 /*
6849 * If its an overlapping domain it has private groups, iterate and
6850 * nuke them all.
6851 */
6852 if (sd->flags & SD_OVERLAP) {
6853 free_sched_groups(sd->groups, 1);
6854 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6855 kfree(sd->groups->sgp);
6856 kfree(sd->groups);
6857 }
6858 kfree(sd);
6859}
6860
6861static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6862{
6863 call_rcu(&sd->rcu, free_sched_domain);
6864}
6865
6866static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6867{
6868 for (; sd; sd = sd->parent)
6869 destroy_sched_domain(sd, cpu);
6870}
6871
6872/*
6873 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6874 * hold the hotplug lock.
6875 */
6876static void
6877cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6878{
6879 struct rq *rq = cpu_rq(cpu);
6880 struct sched_domain *tmp;
6881
6882 /* Remove the sched domains which do not contribute to scheduling. */
6883 for (tmp = sd; tmp; ) {
6884 struct sched_domain *parent = tmp->parent;
6885 if (!parent)
6886 break;
6887
6888 if (sd_parent_degenerate(tmp, parent)) {
6889 tmp->parent = parent->parent;
6890 if (parent->parent)
6891 parent->parent->child = tmp;
6892 destroy_sched_domain(parent, cpu);
6893 } else
6894 tmp = tmp->parent;
6895 }
6896
6897 if (sd && sd_degenerate(sd)) {
6898 tmp = sd;
6899 sd = sd->parent;
6900 destroy_sched_domain(tmp, cpu);
6901 if (sd)
6902 sd->child = NULL;
6903 }
6904
6905 sched_domain_debug(sd, cpu);
6906
6907 rq_attach_root(rq, rd);
6908 tmp = rq->sd;
6909 rcu_assign_pointer(rq->sd, sd);
6910 destroy_sched_domains(tmp, cpu);
6911}
6912
6913/* cpus with isolated domains */
6914static cpumask_var_t cpu_isolated_map;
6915
6916/* Setup the mask of cpus configured for isolated domains */
6917static int __init isolated_cpu_setup(char *str)
6918{
6919 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6920 cpulist_parse(str, cpu_isolated_map);
6921 return 1;
6922}
6923
6924__setup("isolcpus=", isolated_cpu_setup);
6925
6926#define SD_NODES_PER_DOMAIN 16
6927
6928#ifdef CONFIG_NUMA
6929
6930/**
6931 * find_next_best_node - find the next node to include in a sched_domain
6932 * @node: node whose sched_domain we're building
6933 * @used_nodes: nodes already in the sched_domain
6934 *
6935 * Find the next node to include in a given scheduling domain. Simply
6936 * finds the closest node not already in the @used_nodes map.
6937 *
6938 * Should use nodemask_t.
6939 */
6940static int find_next_best_node(int node, nodemask_t *used_nodes)
6941{
6942 int i, n, val, min_val, best_node = -1;
6943
6944 min_val = INT_MAX;
6945
6946 for (i = 0; i < nr_node_ids; i++) {
6947 /* Start at @node */
6948 n = (node + i) % nr_node_ids;
6949
6950 if (!nr_cpus_node(n))
6951 continue;
6952
6953 /* Skip already used nodes */
6954 if (node_isset(n, *used_nodes))
6955 continue;
6956
6957 /* Simple min distance search */
6958 val = node_distance(node, n);
6959
6960 if (val < min_val) {
6961 min_val = val;
6962 best_node = n;
6963 }
6964 }
6965
6966 if (best_node != -1)
6967 node_set(best_node, *used_nodes);
6968 return best_node;
6969}
6970
6971/**
6972 * sched_domain_node_span - get a cpumask for a node's sched_domain
6973 * @node: node whose cpumask we're constructing
6974 * @span: resulting cpumask
6975 *
6976 * Given a node, construct a good cpumask for its sched_domain to span. It
6977 * should be one that prevents unnecessary balancing, but also spreads tasks
6978 * out optimally.
6979 */
6980static void sched_domain_node_span(int node, struct cpumask *span)
6981{
6982 nodemask_t used_nodes;
6983 int i;
6984
6985 cpumask_clear(span);
6986 nodes_clear(used_nodes);
6987
6988 cpumask_or(span, span, cpumask_of_node(node));
6989 node_set(node, used_nodes);
6990
6991 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6992 int next_node = find_next_best_node(node, &used_nodes);
6993 if (next_node < 0)
6994 break;
6995 cpumask_or(span, span, cpumask_of_node(next_node));
6996 }
6997}
6998
6999static const struct cpumask *cpu_node_mask(int cpu)
7000{
7001 lockdep_assert_held(&sched_domains_mutex);
7002
7003 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
7004
7005 return sched_domains_tmpmask;
7006}
7007
7008static const struct cpumask *cpu_allnodes_mask(int cpu)
7009{
7010 return cpu_possible_mask;
7011}
7012#endif /* CONFIG_NUMA */
7013
7014static const struct cpumask *cpu_cpu_mask(int cpu)
7015{
7016 return cpumask_of_node(cpu_to_node(cpu));
7017}
7018
7019int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7020
7021struct sd_data {
7022 struct sched_domain **__percpu sd;
7023 struct sched_group **__percpu sg;
7024 struct sched_group_power **__percpu sgp;
7025};
7026
7027struct s_data {
7028 struct sched_domain ** __percpu sd;
7029 struct root_domain *rd;
7030};
7031
7032enum s_alloc {
7033 sa_rootdomain,
7034 sa_sd,
7035 sa_sd_storage,
7036 sa_none,
7037};
7038
7039struct sched_domain_topology_level;
7040
7041typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7042typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7043
7044#define SDTL_OVERLAP 0x01
7045
7046struct sched_domain_topology_level {
7047 sched_domain_init_f init;
7048 sched_domain_mask_f mask;
7049 int flags;
7050 struct sd_data data;
7051};
7052
7053static int
7054build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7055{
7056 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7057 const struct cpumask *span = sched_domain_span(sd);
7058 struct cpumask *covered = sched_domains_tmpmask;
7059 struct sd_data *sdd = sd->private;
7060 struct sched_domain *child;
7061 int i;
7062
7063 cpumask_clear(covered);
7064
7065 for_each_cpu(i, span) {
7066 struct cpumask *sg_span;
7067
7068 if (cpumask_test_cpu(i, covered))
7069 continue;
7070
7071 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7072 GFP_KERNEL, cpu_to_node(i));
7073
7074 if (!sg)
7075 goto fail;
7076
7077 sg_span = sched_group_cpus(sg);
7078
7079 child = *per_cpu_ptr(sdd->sd, i);
7080 if (child->child) {
7081 child = child->child;
7082 cpumask_copy(sg_span, sched_domain_span(child));
7083 } else
7084 cpumask_set_cpu(i, sg_span);
7085
7086 cpumask_or(covered, covered, sg_span);
7087
7088 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7089 atomic_inc(&sg->sgp->ref);
7090
7091 if (cpumask_test_cpu(cpu, sg_span))
7092 groups = sg;
7093
7094 if (!first)
7095 first = sg;
7096 if (last)
7097 last->next = sg;
7098 last = sg;
7099 last->next = first;
7100 }
7101 sd->groups = groups;
7102
7103 return 0;
7104
7105fail:
7106 free_sched_groups(first, 0);
7107
7108 return -ENOMEM;
7109}
7110
7111static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7112{
7113 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7114 struct sched_domain *child = sd->child;
7115
7116 if (child)
7117 cpu = cpumask_first(sched_domain_span(child));
7118
7119 if (sg) {
7120 *sg = *per_cpu_ptr(sdd->sg, cpu);
7121 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7122 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7123 }
7124
7125 return cpu;
7126}
7127
7128/*
7129 * build_sched_groups will build a circular linked list of the groups
7130 * covered by the given span, and will set each group's ->cpumask correctly,
7131 * and ->cpu_power to 0.
7132 *
7133 * Assumes the sched_domain tree is fully constructed
7134 */
7135static int
7136build_sched_groups(struct sched_domain *sd, int cpu)
7137{
7138 struct sched_group *first = NULL, *last = NULL;
7139 struct sd_data *sdd = sd->private;
7140 const struct cpumask *span = sched_domain_span(sd);
7141 struct cpumask *covered;
7142 int i;
7143
7144 get_group(cpu, sdd, &sd->groups);
7145 atomic_inc(&sd->groups->ref);
7146
7147 if (cpu != cpumask_first(sched_domain_span(sd)))
7148 return 0;
7149
7150 lockdep_assert_held(&sched_domains_mutex);
7151 covered = sched_domains_tmpmask;
7152
7153 cpumask_clear(covered);
7154
7155 for_each_cpu(i, span) {
7156 struct sched_group *sg;
7157 int group = get_group(i, sdd, &sg);
7158 int j;
7159
7160 if (cpumask_test_cpu(i, covered))
7161 continue;
7162
7163 cpumask_clear(sched_group_cpus(sg));
7164 sg->sgp->power = 0;
7165
7166 for_each_cpu(j, span) {
7167 if (get_group(j, sdd, NULL) != group)
7168 continue;
7169
7170 cpumask_set_cpu(j, covered);
7171 cpumask_set_cpu(j, sched_group_cpus(sg));
7172 }
7173
7174 if (!first)
7175 first = sg;
7176 if (last)
7177 last->next = sg;
7178 last = sg;
7179 }
7180 last->next = first;
7181
7182 return 0;
7183}
7184
7185/*
7186 * Initialize sched groups cpu_power.
7187 *
7188 * cpu_power indicates the capacity of sched group, which is used while
7189 * distributing the load between different sched groups in a sched domain.
7190 * Typically cpu_power for all the groups in a sched domain will be same unless
7191 * there are asymmetries in the topology. If there are asymmetries, group
7192 * having more cpu_power will pickup more load compared to the group having
7193 * less cpu_power.
7194 */
7195static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7196{
7197 struct sched_group *sg = sd->groups;
7198
7199 WARN_ON(!sd || !sg);
7200
7201 do {
7202 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7203 sg = sg->next;
7204 } while (sg != sd->groups);
7205
7206 if (cpu != group_first_cpu(sg))
7207 return;
7208
7209 update_group_power(sd, cpu);
7210}
7211
7212/*
7213 * Initializers for schedule domains
7214 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7215 */
7216
7217#ifdef CONFIG_SCHED_DEBUG
7218# define SD_INIT_NAME(sd, type) sd->name = #type
7219#else
7220# define SD_INIT_NAME(sd, type) do { } while (0)
7221#endif
7222
7223#define SD_INIT_FUNC(type) \
7224static noinline struct sched_domain * \
7225sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7226{ \
7227 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7228 *sd = SD_##type##_INIT; \
7229 SD_INIT_NAME(sd, type); \
7230 sd->private = &tl->data; \
7231 return sd; \
7232}
7233
7234SD_INIT_FUNC(CPU)
7235#ifdef CONFIG_NUMA
7236 SD_INIT_FUNC(ALLNODES)
7237 SD_INIT_FUNC(NODE)
7238#endif
7239#ifdef CONFIG_SCHED_SMT
7240 SD_INIT_FUNC(SIBLING)
7241#endif
7242#ifdef CONFIG_SCHED_MC
7243 SD_INIT_FUNC(MC)
7244#endif
7245#ifdef CONFIG_SCHED_BOOK
7246 SD_INIT_FUNC(BOOK)
7247#endif
7248
7249static int default_relax_domain_level = -1;
7250int sched_domain_level_max;
7251
7252static int __init setup_relax_domain_level(char *str)
7253{
7254 unsigned long val;
7255
7256 val = simple_strtoul(str, NULL, 0);
7257 if (val < sched_domain_level_max)
7258 default_relax_domain_level = val;
7259
7260 return 1;
7261}
7262__setup("relax_domain_level=", setup_relax_domain_level);
7263
7264static void set_domain_attribute(struct sched_domain *sd,
7265 struct sched_domain_attr *attr)
7266{
7267 int request;
7268
7269 if (!attr || attr->relax_domain_level < 0) {
7270 if (default_relax_domain_level < 0)
7271 return;
7272 else
7273 request = default_relax_domain_level;
7274 } else
7275 request = attr->relax_domain_level;
7276 if (request < sd->level) {
7277 /* turn off idle balance on this domain */
7278 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7279 } else {
7280 /* turn on idle balance on this domain */
7281 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7282 }
7283}
7284
7285static void __sdt_free(const struct cpumask *cpu_map);
7286static int __sdt_alloc(const struct cpumask *cpu_map);
7287
7288static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7289 const struct cpumask *cpu_map)
7290{
7291 switch (what) {
7292 case sa_rootdomain:
7293 if (!atomic_read(&d->rd->refcount))
7294 free_rootdomain(&d->rd->rcu); /* fall through */
7295 case sa_sd:
7296 free_percpu(d->sd); /* fall through */
7297 case sa_sd_storage:
7298 __sdt_free(cpu_map); /* fall through */
7299 case sa_none:
7300 break;
7301 }
7302}
7303
7304static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7305 const struct cpumask *cpu_map)
7306{
7307 memset(d, 0, sizeof(*d));
7308
7309 if (__sdt_alloc(cpu_map))
7310 return sa_sd_storage;
7311 d->sd = alloc_percpu(struct sched_domain *);
7312 if (!d->sd)
7313 return sa_sd_storage;
7314 d->rd = alloc_rootdomain();
7315 if (!d->rd)
7316 return sa_sd;
7317 return sa_rootdomain;
7318}
7319
7320/*
7321 * NULL the sd_data elements we've used to build the sched_domain and
7322 * sched_group structure so that the subsequent __free_domain_allocs()
7323 * will not free the data we're using.
7324 */
7325static void claim_allocations(int cpu, struct sched_domain *sd)
7326{
7327 struct sd_data *sdd = sd->private;
7328
7329 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7330 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7331
7332 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7333 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7334
7335 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7336 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7337}
7338
7339#ifdef CONFIG_SCHED_SMT
7340static const struct cpumask *cpu_smt_mask(int cpu)
7341{
7342 return topology_thread_cpumask(cpu);
7343}
7344#endif
7345
7346/*
7347 * Topology list, bottom-up.
7348 */
7349static struct sched_domain_topology_level default_topology[] = {
7350#ifdef CONFIG_SCHED_SMT
7351 { sd_init_SIBLING, cpu_smt_mask, },
7352#endif
7353#ifdef CONFIG_SCHED_MC
7354 { sd_init_MC, cpu_coregroup_mask, },
7355#endif
7356#ifdef CONFIG_SCHED_BOOK
7357 { sd_init_BOOK, cpu_book_mask, },
7358#endif
7359 { sd_init_CPU, cpu_cpu_mask, },
7360#ifdef CONFIG_NUMA
7361 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7362 { sd_init_ALLNODES, cpu_allnodes_mask, },
7363#endif
7364 { NULL, },
7365};
7366
7367static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7368
7369static int __sdt_alloc(const struct cpumask *cpu_map)
7370{
7371 struct sched_domain_topology_level *tl;
7372 int j;
7373
7374 for (tl = sched_domain_topology; tl->init; tl++) {
7375 struct sd_data *sdd = &tl->data;
7376
7377 sdd->sd = alloc_percpu(struct sched_domain *);
7378 if (!sdd->sd)
7379 return -ENOMEM;
7380
7381 sdd->sg = alloc_percpu(struct sched_group *);
7382 if (!sdd->sg)
7383 return -ENOMEM;
7384
7385 sdd->sgp = alloc_percpu(struct sched_group_power *);
7386 if (!sdd->sgp)
7387 return -ENOMEM;
7388
7389 for_each_cpu(j, cpu_map) {
7390 struct sched_domain *sd;
7391 struct sched_group *sg;
7392 struct sched_group_power *sgp;
7393
7394 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7395 GFP_KERNEL, cpu_to_node(j));
7396 if (!sd)
7397 return -ENOMEM;
7398
7399 *per_cpu_ptr(sdd->sd, j) = sd;
7400
7401 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7402 GFP_KERNEL, cpu_to_node(j));
7403 if (!sg)
7404 return -ENOMEM;
7405
7406 *per_cpu_ptr(sdd->sg, j) = sg;
7407
7408 sgp = kzalloc_node(sizeof(struct sched_group_power),
7409 GFP_KERNEL, cpu_to_node(j));
7410 if (!sgp)
7411 return -ENOMEM;
7412
7413 *per_cpu_ptr(sdd->sgp, j) = sgp;
7414 }
7415 }
7416
7417 return 0;
7418}
7419
7420static void __sdt_free(const struct cpumask *cpu_map)
7421{
7422 struct sched_domain_topology_level *tl;
7423 int j;
7424
7425 for (tl = sched_domain_topology; tl->init; tl++) {
7426 struct sd_data *sdd = &tl->data;
7427
7428 for_each_cpu(j, cpu_map) {
7429 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
7430 if (sd && (sd->flags & SD_OVERLAP))
7431 free_sched_groups(sd->groups, 0);
7432 kfree(*per_cpu_ptr(sdd->sd, j));
7433 kfree(*per_cpu_ptr(sdd->sg, j));
7434 kfree(*per_cpu_ptr(sdd->sgp, j));
7435 }
7436 free_percpu(sdd->sd);
7437 free_percpu(sdd->sg);
7438 free_percpu(sdd->sgp);
7439 }
7440}
7441
7442struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7443 struct s_data *d, const struct cpumask *cpu_map,
7444 struct sched_domain_attr *attr, struct sched_domain *child,
7445 int cpu)
7446{
7447 struct sched_domain *sd = tl->init(tl, cpu);
7448 if (!sd)
7449 return child;
7450
7451 set_domain_attribute(sd, attr);
7452 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7453 if (child) {
7454 sd->level = child->level + 1;
7455 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7456 child->parent = sd;
7457 }
7458 sd->child = child;
7459
7460 return sd;
7461}
7462
7463/*
7464 * Build sched domains for a given set of cpus and attach the sched domains
7465 * to the individual cpus
7466 */
7467static int build_sched_domains(const struct cpumask *cpu_map,
7468 struct sched_domain_attr *attr)
7469{
7470 enum s_alloc alloc_state = sa_none;
7471 struct sched_domain *sd;
7472 struct s_data d;
7473 int i, ret = -ENOMEM;
7474
7475 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7476 if (alloc_state != sa_rootdomain)
7477 goto error;
7478
7479 /* Set up domains for cpus specified by the cpu_map. */
7480 for_each_cpu(i, cpu_map) {
7481 struct sched_domain_topology_level *tl;
7482
7483 sd = NULL;
7484 for (tl = sched_domain_topology; tl->init; tl++) {
7485 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7486 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7487 sd->flags |= SD_OVERLAP;
7488 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7489 break;
7490 }
7491
7492 while (sd->child)
7493 sd = sd->child;
7494
7495 *per_cpu_ptr(d.sd, i) = sd;
7496 }
7497
7498 /* Build the groups for the domains */
7499 for_each_cpu(i, cpu_map) {
7500 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7501 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7502 if (sd->flags & SD_OVERLAP) {
7503 if (build_overlap_sched_groups(sd, i))
7504 goto error;
7505 } else {
7506 if (build_sched_groups(sd, i))
7507 goto error;
7508 }
7509 }
7510 }
7511
7512 /* Calculate CPU power for physical packages and nodes */
7513 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7514 if (!cpumask_test_cpu(i, cpu_map))
7515 continue;
7516
7517 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7518 claim_allocations(i, sd);
7519 init_sched_groups_power(i, sd);
7520 }
7521 }
7522
7523 /* Attach the domains */
7524 rcu_read_lock();
7525 for_each_cpu(i, cpu_map) {
7526 sd = *per_cpu_ptr(d.sd, i);
7527 cpu_attach_domain(sd, d.rd, i);
7528 }
7529 rcu_read_unlock();
7530
7531 ret = 0;
7532error:
7533 __free_domain_allocs(&d, alloc_state, cpu_map);
7534 return ret;
7535}
7536
7537static cpumask_var_t *doms_cur; /* current sched domains */
7538static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7539static struct sched_domain_attr *dattr_cur;
7540 /* attribues of custom domains in 'doms_cur' */
7541
7542/*
7543 * Special case: If a kmalloc of a doms_cur partition (array of
7544 * cpumask) fails, then fallback to a single sched domain,
7545 * as determined by the single cpumask fallback_doms.
7546 */
7547static cpumask_var_t fallback_doms;
7548
7549/*
7550 * arch_update_cpu_topology lets virtualized architectures update the
7551 * cpu core maps. It is supposed to return 1 if the topology changed
7552 * or 0 if it stayed the same.
7553 */
7554int __attribute__((weak)) arch_update_cpu_topology(void)
7555{
7556 return 0;
7557}
7558
7559cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7560{
7561 int i;
7562 cpumask_var_t *doms;
7563
7564 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7565 if (!doms)
7566 return NULL;
7567 for (i = 0; i < ndoms; i++) {
7568 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7569 free_sched_domains(doms, i);
7570 return NULL;
7571 }
7572 }
7573 return doms;
7574}
7575
7576void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7577{
7578 unsigned int i;
7579 for (i = 0; i < ndoms; i++)
7580 free_cpumask_var(doms[i]);
7581 kfree(doms);
7582}
7583
7584/*
7585 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7586 * For now this just excludes isolated cpus, but could be used to
7587 * exclude other special cases in the future.
7588 */
7589static int init_sched_domains(const struct cpumask *cpu_map)
7590{
7591 int err;
7592
7593 arch_update_cpu_topology();
7594 ndoms_cur = 1;
7595 doms_cur = alloc_sched_domains(ndoms_cur);
7596 if (!doms_cur)
7597 doms_cur = &fallback_doms;
7598 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7599 dattr_cur = NULL;
7600 err = build_sched_domains(doms_cur[0], NULL);
7601 register_sched_domain_sysctl();
7602
7603 return err;
7604}
7605
7606/*
7607 * Detach sched domains from a group of cpus specified in cpu_map
7608 * These cpus will now be attached to the NULL domain
7609 */
7610static void detach_destroy_domains(const struct cpumask *cpu_map)
7611{
7612 int i;
7613
7614 rcu_read_lock();
7615 for_each_cpu(i, cpu_map)
7616 cpu_attach_domain(NULL, &def_root_domain, i);
7617 rcu_read_unlock();
7618}
7619
7620/* handle null as "default" */
7621static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7622 struct sched_domain_attr *new, int idx_new)
7623{
7624 struct sched_domain_attr tmp;
7625
7626 /* fast path */
7627 if (!new && !cur)
7628 return 1;
7629
7630 tmp = SD_ATTR_INIT;
7631 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7632 new ? (new + idx_new) : &tmp,
7633 sizeof(struct sched_domain_attr));
7634}
7635
7636/*
7637 * Partition sched domains as specified by the 'ndoms_new'
7638 * cpumasks in the array doms_new[] of cpumasks. This compares
7639 * doms_new[] to the current sched domain partitioning, doms_cur[].
7640 * It destroys each deleted domain and builds each new domain.
7641 *
7642 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7643 * The masks don't intersect (don't overlap.) We should setup one
7644 * sched domain for each mask. CPUs not in any of the cpumasks will
7645 * not be load balanced. If the same cpumask appears both in the
7646 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7647 * it as it is.
7648 *
7649 * The passed in 'doms_new' should be allocated using
7650 * alloc_sched_domains. This routine takes ownership of it and will
7651 * free_sched_domains it when done with it. If the caller failed the
7652 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7653 * and partition_sched_domains() will fallback to the single partition
7654 * 'fallback_doms', it also forces the domains to be rebuilt.
7655 *
7656 * If doms_new == NULL it will be replaced with cpu_online_mask.
7657 * ndoms_new == 0 is a special case for destroying existing domains,
7658 * and it will not create the default domain.
7659 *
7660 * Call with hotplug lock held
7661 */
7662void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7663 struct sched_domain_attr *dattr_new)
7664{
7665 int i, j, n;
7666 int new_topology;
7667
7668 mutex_lock(&sched_domains_mutex);
7669
7670 /* always unregister in case we don't destroy any domains */
7671 unregister_sched_domain_sysctl();
7672
7673 /* Let architecture update cpu core mappings. */
7674 new_topology = arch_update_cpu_topology();
7675
7676 n = doms_new ? ndoms_new : 0;
7677
7678 /* Destroy deleted domains */
7679 for (i = 0; i < ndoms_cur; i++) {
7680 for (j = 0; j < n && !new_topology; j++) {
7681 if (cpumask_equal(doms_cur[i], doms_new[j])
7682 && dattrs_equal(dattr_cur, i, dattr_new, j))
7683 goto match1;
7684 }
7685 /* no match - a current sched domain not in new doms_new[] */
7686 detach_destroy_domains(doms_cur[i]);
7687match1:
7688 ;
7689 }
7690
7691 if (doms_new == NULL) {
7692 ndoms_cur = 0;
7693 doms_new = &fallback_doms;
7694 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7695 WARN_ON_ONCE(dattr_new);
7696 }
7697
7698 /* Build new domains */
7699 for (i = 0; i < ndoms_new; i++) {
7700 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7701 if (cpumask_equal(doms_new[i], doms_cur[j])
7702 && dattrs_equal(dattr_new, i, dattr_cur, j))
7703 goto match2;
7704 }
7705 /* no match - add a new doms_new */
7706 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7707match2:
7708 ;
7709 }
7710
7711 /* Remember the new sched domains */
7712 if (doms_cur != &fallback_doms)
7713 free_sched_domains(doms_cur, ndoms_cur);
7714 kfree(dattr_cur); /* kfree(NULL) is safe */
7715 doms_cur = doms_new;
7716 dattr_cur = dattr_new;
7717 ndoms_cur = ndoms_new;
7718
7719 register_sched_domain_sysctl();
7720
7721 mutex_unlock(&sched_domains_mutex);
7722}
7723
7724#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7725static void reinit_sched_domains(void)
7726{
7727 get_online_cpus();
7728
7729 /* Destroy domains first to force the rebuild */
7730 partition_sched_domains(0, NULL, NULL);
7731
7732 rebuild_sched_domains();
7733 put_online_cpus();
7734}
7735
7736static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7737{
7738 unsigned int level = 0;
7739
7740 if (sscanf(buf, "%u", &level) != 1)
7741 return -EINVAL;
7742
7743 /*
7744 * level is always be positive so don't check for
7745 * level < POWERSAVINGS_BALANCE_NONE which is 0
7746 * What happens on 0 or 1 byte write,
7747 * need to check for count as well?
7748 */
7749
7750 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7751 return -EINVAL;
7752
7753 if (smt)
7754 sched_smt_power_savings = level;
7755 else
7756 sched_mc_power_savings = level;
7757
7758 reinit_sched_domains();
7759
7760 return count;
7761}
7762
7763#ifdef CONFIG_SCHED_MC
7764static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7765 struct sysdev_class_attribute *attr,
7766 char *page)
7767{
7768 return sprintf(page, "%u\n", sched_mc_power_savings);
7769}
7770static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7771 struct sysdev_class_attribute *attr,
7772 const char *buf, size_t count)
7773{
7774 return sched_power_savings_store(buf, count, 0);
7775}
7776static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7777 sched_mc_power_savings_show,
7778 sched_mc_power_savings_store);
7779#endif
7780
7781#ifdef CONFIG_SCHED_SMT
7782static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7783 struct sysdev_class_attribute *attr,
7784 char *page)
7785{
7786 return sprintf(page, "%u\n", sched_smt_power_savings);
7787}
7788static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7789 struct sysdev_class_attribute *attr,
7790 const char *buf, size_t count)
7791{
7792 return sched_power_savings_store(buf, count, 1);
7793}
7794static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7795 sched_smt_power_savings_show,
7796 sched_smt_power_savings_store);
7797#endif
7798
7799int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7800{
7801 int err = 0;
7802
7803#ifdef CONFIG_SCHED_SMT
7804 if (smt_capable())
7805 err = sysfs_create_file(&cls->kset.kobj,
7806 &attr_sched_smt_power_savings.attr);
7807#endif
7808#ifdef CONFIG_SCHED_MC
7809 if (!err && mc_capable())
7810 err = sysfs_create_file(&cls->kset.kobj,
7811 &attr_sched_mc_power_savings.attr);
7812#endif
7813 return err;
7814}
7815#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7816
7817/*
7818 * Update cpusets according to cpu_active mask. If cpusets are
7819 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7820 * around partition_sched_domains().
7821 */
7822static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7823 void *hcpu)
7824{
7825 switch (action & ~CPU_TASKS_FROZEN) {
7826 case CPU_ONLINE:
7827 case CPU_DOWN_FAILED:
7828 cpuset_update_active_cpus();
7829 return NOTIFY_OK;
7830 default:
7831 return NOTIFY_DONE;
7832 }
7833}
7834
7835static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7836 void *hcpu)
7837{
7838 switch (action & ~CPU_TASKS_FROZEN) {
7839 case CPU_DOWN_PREPARE:
7840 cpuset_update_active_cpus();
7841 return NOTIFY_OK;
7842 default:
7843 return NOTIFY_DONE;
7844 }
7845}
7846
7847static int update_runtime(struct notifier_block *nfb,
7848 unsigned long action, void *hcpu)
7849{
7850 int cpu = (int)(long)hcpu;
7851
7852 switch (action) {
7853 case CPU_DOWN_PREPARE:
7854 case CPU_DOWN_PREPARE_FROZEN:
7855 disable_runtime(cpu_rq(cpu));
7856 return NOTIFY_OK;
7857
7858 case CPU_DOWN_FAILED:
7859 case CPU_DOWN_FAILED_FROZEN:
7860 case CPU_ONLINE:
7861 case CPU_ONLINE_FROZEN:
7862 enable_runtime(cpu_rq(cpu));
7863 return NOTIFY_OK;
7864
7865 default:
7866 return NOTIFY_DONE;
7867 }
7868}
7869
7870void __init sched_init_smp(void)
7871{
7872 cpumask_var_t non_isolated_cpus;
7873
7874 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7875 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7876
7877 get_online_cpus();
7878 mutex_lock(&sched_domains_mutex);
7879 init_sched_domains(cpu_active_mask);
7880 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7881 if (cpumask_empty(non_isolated_cpus))
7882 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7883 mutex_unlock(&sched_domains_mutex);
7884 put_online_cpus();
7885
7886 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7887 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7888
7889 /* RT runtime code needs to handle some hotplug events */
7890 hotcpu_notifier(update_runtime, 0);
7891
7892 init_hrtick();
7893
7894 /* Move init over to a non-isolated CPU */
7895 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7896 BUG();
7897 sched_init_granularity();
7898 free_cpumask_var(non_isolated_cpus);
7899
7900 init_sched_rt_class();
7901}
7902#else
7903void __init sched_init_smp(void)
7904{
7905 sched_init_granularity();
7906}
7907#endif /* CONFIG_SMP */
7908
7909const_debug unsigned int sysctl_timer_migration = 1;
7910
7911int in_sched_functions(unsigned long addr)
7912{
7913 return in_lock_functions(addr) ||
7914 (addr >= (unsigned long)__sched_text_start
7915 && addr < (unsigned long)__sched_text_end);
7916}
7917
7918static void init_cfs_rq(struct cfs_rq *cfs_rq)
7919{
7920 cfs_rq->tasks_timeline = RB_ROOT;
7921 INIT_LIST_HEAD(&cfs_rq->tasks);
7922 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7923#ifndef CONFIG_64BIT
7924 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7925#endif
7926}
7927
7928static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7929{
7930 struct rt_prio_array *array;
7931 int i;
7932
7933 array = &rt_rq->active;
7934 for (i = 0; i < MAX_RT_PRIO; i++) {
7935 INIT_LIST_HEAD(array->queue + i);
7936 __clear_bit(i, array->bitmap);
7937 }
7938 /* delimiter for bitsearch: */
7939 __set_bit(MAX_RT_PRIO, array->bitmap);
7940
7941#if defined CONFIG_SMP
7942 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7943 rt_rq->highest_prio.next = MAX_RT_PRIO;
7944 rt_rq->rt_nr_migratory = 0;
7945 rt_rq->overloaded = 0;
7946 plist_head_init(&rt_rq->pushable_tasks);
7947#endif
7948
7949 rt_rq->rt_time = 0;
7950 rt_rq->rt_throttled = 0;
7951 rt_rq->rt_runtime = 0;
7952 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7953}
7954
7955#ifdef CONFIG_FAIR_GROUP_SCHED
7956static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7957 struct sched_entity *se, int cpu,
7958 struct sched_entity *parent)
7959{
7960 struct rq *rq = cpu_rq(cpu);
7961
7962 cfs_rq->tg = tg;
7963 cfs_rq->rq = rq;
7964#ifdef CONFIG_SMP
7965 /* allow initial update_cfs_load() to truncate */
7966 cfs_rq->load_stamp = 1;
7967#endif
7968
7969 tg->cfs_rq[cpu] = cfs_rq;
7970 tg->se[cpu] = se;
7971
7972 /* se could be NULL for root_task_group */
7973 if (!se)
7974 return;
7975
7976 if (!parent)
7977 se->cfs_rq = &rq->cfs;
7978 else
7979 se->cfs_rq = parent->my_q;
7980
7981 se->my_q = cfs_rq;
7982 update_load_set(&se->load, 0);
7983 se->parent = parent;
7984}
7985#endif
7986
7987#ifdef CONFIG_RT_GROUP_SCHED
7988static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7989 struct sched_rt_entity *rt_se, int cpu,
7990 struct sched_rt_entity *parent)
7991{
7992 struct rq *rq = cpu_rq(cpu);
7993
7994 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7995 rt_rq->rt_nr_boosted = 0;
7996 rt_rq->rq = rq;
7997 rt_rq->tg = tg;
7998
7999 tg->rt_rq[cpu] = rt_rq;
8000 tg->rt_se[cpu] = rt_se;
8001
8002 if (!rt_se)
8003 return;
8004
8005 if (!parent)
8006 rt_se->rt_rq = &rq->rt;
8007 else
8008 rt_se->rt_rq = parent->my_q;
8009
8010 rt_se->my_q = rt_rq;
8011 rt_se->parent = parent;
8012 INIT_LIST_HEAD(&rt_se->run_list);
8013}
8014#endif
8015
8016void __init sched_init(void)
8017{
8018 int i, j;
8019 unsigned long alloc_size = 0, ptr;
8020
8021#ifdef CONFIG_FAIR_GROUP_SCHED
8022 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8023#endif
8024#ifdef CONFIG_RT_GROUP_SCHED
8025 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8026#endif
8027#ifdef CONFIG_CPUMASK_OFFSTACK
8028 alloc_size += num_possible_cpus() * cpumask_size();
8029#endif
8030 if (alloc_size) {
8031 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8032
8033#ifdef CONFIG_FAIR_GROUP_SCHED
8034 root_task_group.se = (struct sched_entity **)ptr;
8035 ptr += nr_cpu_ids * sizeof(void **);
8036
8037 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8038 ptr += nr_cpu_ids * sizeof(void **);
8039
8040#endif /* CONFIG_FAIR_GROUP_SCHED */
8041#ifdef CONFIG_RT_GROUP_SCHED
8042 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8043 ptr += nr_cpu_ids * sizeof(void **);
8044
8045 root_task_group.rt_rq = (struct rt_rq **)ptr;
8046 ptr += nr_cpu_ids * sizeof(void **);
8047
8048#endif /* CONFIG_RT_GROUP_SCHED */
8049#ifdef CONFIG_CPUMASK_OFFSTACK
8050 for_each_possible_cpu(i) {
8051 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8052 ptr += cpumask_size();
8053 }
8054#endif /* CONFIG_CPUMASK_OFFSTACK */
8055 }
8056
8057#ifdef CONFIG_SMP
8058 init_defrootdomain();
8059#endif
8060
8061 init_rt_bandwidth(&def_rt_bandwidth,
8062 global_rt_period(), global_rt_runtime());
8063
8064#ifdef CONFIG_RT_GROUP_SCHED
8065 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8066 global_rt_period(), global_rt_runtime());
8067#endif /* CONFIG_RT_GROUP_SCHED */
8068
8069#ifdef CONFIG_CGROUP_SCHED
8070 list_add(&root_task_group.list, &task_groups);
8071 INIT_LIST_HEAD(&root_task_group.children);
8072 autogroup_init(&init_task);
8073#endif /* CONFIG_CGROUP_SCHED */
8074
8075 for_each_possible_cpu(i) {
8076 struct rq *rq;
8077
8078 rq = cpu_rq(i);
8079 raw_spin_lock_init(&rq->lock);
8080 rq->nr_running = 0;
8081 rq->calc_load_active = 0;
8082 rq->calc_load_update = jiffies + LOAD_FREQ;
8083 init_cfs_rq(&rq->cfs);
8084 init_rt_rq(&rq->rt, rq);
8085#ifdef CONFIG_FAIR_GROUP_SCHED
8086 root_task_group.shares = root_task_group_load;
8087 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8088 /*
8089 * How much cpu bandwidth does root_task_group get?
8090 *
8091 * In case of task-groups formed thr' the cgroup filesystem, it
8092 * gets 100% of the cpu resources in the system. This overall
8093 * system cpu resource is divided among the tasks of
8094 * root_task_group and its child task-groups in a fair manner,
8095 * based on each entity's (task or task-group's) weight
8096 * (se->load.weight).
8097 *
8098 * In other words, if root_task_group has 10 tasks of weight
8099 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8100 * then A0's share of the cpu resource is:
8101 *
8102 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8103 *
8104 * We achieve this by letting root_task_group's tasks sit
8105 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8106 */
8107 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8108#endif /* CONFIG_FAIR_GROUP_SCHED */
8109
8110 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8111#ifdef CONFIG_RT_GROUP_SCHED
8112 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8113 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8114#endif
8115
8116 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8117 rq->cpu_load[j] = 0;
8118
8119 rq->last_load_update_tick = jiffies;
8120
8121#ifdef CONFIG_SMP
8122 rq->sd = NULL;
8123 rq->rd = NULL;
8124 rq->cpu_power = SCHED_POWER_SCALE;
8125 rq->post_schedule = 0;
8126 rq->active_balance = 0;
8127 rq->next_balance = jiffies;
8128 rq->push_cpu = 0;
8129 rq->cpu = i;
8130 rq->online = 0;
8131 rq->idle_stamp = 0;
8132 rq->avg_idle = 2*sysctl_sched_migration_cost;
8133 rq_attach_root(rq, &def_root_domain);
8134#ifdef CONFIG_NO_HZ
8135 rq->nohz_balance_kick = 0;
8136 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8137#endif
8138#endif
8139 init_rq_hrtick(rq);
8140 atomic_set(&rq->nr_iowait, 0);
8141 }
8142
8143 set_load_weight(&init_task);
8144
8145#ifdef CONFIG_PREEMPT_NOTIFIERS
8146 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8147#endif
8148
8149#ifdef CONFIG_SMP
8150 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8151#endif
8152
8153#ifdef CONFIG_RT_MUTEXES
8154 plist_head_init(&init_task.pi_waiters);
8155#endif
8156
8157 /*
8158 * The boot idle thread does lazy MMU switching as well:
8159 */
8160 atomic_inc(&init_mm.mm_count);
8161 enter_lazy_tlb(&init_mm, current);
8162
8163 /*
8164 * Make us the idle thread. Technically, schedule() should not be
8165 * called from this thread, however somewhere below it might be,
8166 * but because we are the idle thread, we just pick up running again
8167 * when this runqueue becomes "idle".
8168 */
8169 init_idle(current, smp_processor_id());
8170
8171 calc_load_update = jiffies + LOAD_FREQ;
8172
8173 /*
8174 * During early bootup we pretend to be a normal task:
8175 */
8176 current->sched_class = &fair_sched_class;
8177
8178 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8179 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8180#ifdef CONFIG_SMP
8181 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8182#ifdef CONFIG_NO_HZ
8183 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8184 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8185 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8186 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8187 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8188#endif
8189 /* May be allocated at isolcpus cmdline parse time */
8190 if (cpu_isolated_map == NULL)
8191 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8192#endif /* SMP */
8193
8194 scheduler_running = 1;
8195}
8196
8197#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8198static inline int preempt_count_equals(int preempt_offset)
8199{
8200 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8201
8202 return (nested == preempt_offset);
8203}
8204
8205void __might_sleep(const char *file, int line, int preempt_offset)
8206{
8207 static unsigned long prev_jiffy; /* ratelimiting */
8208
8209 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8210 system_state != SYSTEM_RUNNING || oops_in_progress)
8211 return;
8212 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8213 return;
8214 prev_jiffy = jiffies;
8215
8216 printk(KERN_ERR
8217 "BUG: sleeping function called from invalid context at %s:%d\n",
8218 file, line);
8219 printk(KERN_ERR
8220 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8221 in_atomic(), irqs_disabled(),
8222 current->pid, current->comm);
8223
8224 debug_show_held_locks(current);
8225 if (irqs_disabled())
8226 print_irqtrace_events(current);
8227 dump_stack();
8228}
8229EXPORT_SYMBOL(__might_sleep);
8230#endif
8231
8232#ifdef CONFIG_MAGIC_SYSRQ
8233static void normalize_task(struct rq *rq, struct task_struct *p)
8234{
8235 const struct sched_class *prev_class = p->sched_class;
8236 int old_prio = p->prio;
8237 int on_rq;
8238
8239 on_rq = p->on_rq;
8240 if (on_rq)
8241 deactivate_task(rq, p, 0);
8242 __setscheduler(rq, p, SCHED_NORMAL, 0);
8243 if (on_rq) {
8244 activate_task(rq, p, 0);
8245 resched_task(rq->curr);
8246 }
8247
8248 check_class_changed(rq, p, prev_class, old_prio);
8249}
8250
8251void normalize_rt_tasks(void)
8252{
8253 struct task_struct *g, *p;
8254 unsigned long flags;
8255 struct rq *rq;
8256
8257 read_lock_irqsave(&tasklist_lock, flags);
8258 do_each_thread(g, p) {
8259 /*
8260 * Only normalize user tasks:
8261 */
8262 if (!p->mm)
8263 continue;
8264
8265 p->se.exec_start = 0;
8266#ifdef CONFIG_SCHEDSTATS
8267 p->se.statistics.wait_start = 0;
8268 p->se.statistics.sleep_start = 0;
8269 p->se.statistics.block_start = 0;
8270#endif
8271
8272 if (!rt_task(p)) {
8273 /*
8274 * Renice negative nice level userspace
8275 * tasks back to 0:
8276 */
8277 if (TASK_NICE(p) < 0 && p->mm)
8278 set_user_nice(p, 0);
8279 continue;
8280 }
8281
8282 raw_spin_lock(&p->pi_lock);
8283 rq = __task_rq_lock(p);
8284
8285 normalize_task(rq, p);
8286
8287 __task_rq_unlock(rq);
8288 raw_spin_unlock(&p->pi_lock);
8289 } while_each_thread(g, p);
8290
8291 read_unlock_irqrestore(&tasklist_lock, flags);
8292}
8293
8294#endif /* CONFIG_MAGIC_SYSRQ */
8295
8296#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8297/*
8298 * These functions are only useful for the IA64 MCA handling, or kdb.
8299 *
8300 * They can only be called when the whole system has been
8301 * stopped - every CPU needs to be quiescent, and no scheduling
8302 * activity can take place. Using them for anything else would
8303 * be a serious bug, and as a result, they aren't even visible
8304 * under any other configuration.
8305 */
8306
8307/**
8308 * curr_task - return the current task for a given cpu.
8309 * @cpu: the processor in question.
8310 *
8311 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8312 */
8313struct task_struct *curr_task(int cpu)
8314{
8315 return cpu_curr(cpu);
8316}
8317
8318#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8319
8320#ifdef CONFIG_IA64
8321/**
8322 * set_curr_task - set the current task for a given cpu.
8323 * @cpu: the processor in question.
8324 * @p: the task pointer to set.
8325 *
8326 * Description: This function must only be used when non-maskable interrupts
8327 * are serviced on a separate stack. It allows the architecture to switch the
8328 * notion of the current task on a cpu in a non-blocking manner. This function
8329 * must be called with all CPU's synchronized, and interrupts disabled, the
8330 * and caller must save the original value of the current task (see
8331 * curr_task() above) and restore that value before reenabling interrupts and
8332 * re-starting the system.
8333 *
8334 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8335 */
8336void set_curr_task(int cpu, struct task_struct *p)
8337{
8338 cpu_curr(cpu) = p;
8339}
8340
8341#endif
8342
8343#ifdef CONFIG_FAIR_GROUP_SCHED
8344static void free_fair_sched_group(struct task_group *tg)
8345{
8346 int i;
8347
8348 for_each_possible_cpu(i) {
8349 if (tg->cfs_rq)
8350 kfree(tg->cfs_rq[i]);
8351 if (tg->se)
8352 kfree(tg->se[i]);
8353 }
8354
8355 kfree(tg->cfs_rq);
8356 kfree(tg->se);
8357}
8358
8359static
8360int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8361{
8362 struct cfs_rq *cfs_rq;
8363 struct sched_entity *se;
8364 int i;
8365
8366 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8367 if (!tg->cfs_rq)
8368 goto err;
8369 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8370 if (!tg->se)
8371 goto err;
8372
8373 tg->shares = NICE_0_LOAD;
8374
8375 for_each_possible_cpu(i) {
8376 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8377 GFP_KERNEL, cpu_to_node(i));
8378 if (!cfs_rq)
8379 goto err;
8380
8381 se = kzalloc_node(sizeof(struct sched_entity),
8382 GFP_KERNEL, cpu_to_node(i));
8383 if (!se)
8384 goto err_free_rq;
8385
8386 init_cfs_rq(cfs_rq);
8387 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8388 }
8389
8390 return 1;
8391
8392err_free_rq:
8393 kfree(cfs_rq);
8394err:
8395 return 0;
8396}
8397
8398static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8399{
8400 struct rq *rq = cpu_rq(cpu);
8401 unsigned long flags;
8402
8403 /*
8404 * Only empty task groups can be destroyed; so we can speculatively
8405 * check on_list without danger of it being re-added.
8406 */
8407 if (!tg->cfs_rq[cpu]->on_list)
8408 return;
8409
8410 raw_spin_lock_irqsave(&rq->lock, flags);
8411 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8412 raw_spin_unlock_irqrestore(&rq->lock, flags);
8413}
8414#else /* !CONFIG_FAIR_GROUP_SCHED */
8415static inline void free_fair_sched_group(struct task_group *tg)
8416{
8417}
8418
8419static inline
8420int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8421{
8422 return 1;
8423}
8424
8425static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8426{
8427}
8428#endif /* CONFIG_FAIR_GROUP_SCHED */
8429
8430#ifdef CONFIG_RT_GROUP_SCHED
8431static void free_rt_sched_group(struct task_group *tg)
8432{
8433 int i;
8434
8435 if (tg->rt_se)
8436 destroy_rt_bandwidth(&tg->rt_bandwidth);
8437
8438 for_each_possible_cpu(i) {
8439 if (tg->rt_rq)
8440 kfree(tg->rt_rq[i]);
8441 if (tg->rt_se)
8442 kfree(tg->rt_se[i]);
8443 }
8444
8445 kfree(tg->rt_rq);
8446 kfree(tg->rt_se);
8447}
8448
8449static
8450int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8451{
8452 struct rt_rq *rt_rq;
8453 struct sched_rt_entity *rt_se;
8454 int i;
8455
8456 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8457 if (!tg->rt_rq)
8458 goto err;
8459 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8460 if (!tg->rt_se)
8461 goto err;
8462
8463 init_rt_bandwidth(&tg->rt_bandwidth,
8464 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8465
8466 for_each_possible_cpu(i) {
8467 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8468 GFP_KERNEL, cpu_to_node(i));
8469 if (!rt_rq)
8470 goto err;
8471
8472 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8473 GFP_KERNEL, cpu_to_node(i));
8474 if (!rt_se)
8475 goto err_free_rq;
8476
8477 init_rt_rq(rt_rq, cpu_rq(i));
8478 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8479 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8480 }
8481
8482 return 1;
8483
8484err_free_rq:
8485 kfree(rt_rq);
8486err:
8487 return 0;
8488}
8489#else /* !CONFIG_RT_GROUP_SCHED */
8490static inline void free_rt_sched_group(struct task_group *tg)
8491{
8492}
8493
8494static inline
8495int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8496{
8497 return 1;
8498}
8499#endif /* CONFIG_RT_GROUP_SCHED */
8500
8501#ifdef CONFIG_CGROUP_SCHED
8502static void free_sched_group(struct task_group *tg)
8503{
8504 free_fair_sched_group(tg);
8505 free_rt_sched_group(tg);
8506 autogroup_free(tg);
8507 kfree(tg);
8508}
8509
8510/* allocate runqueue etc for a new task group */
8511struct task_group *sched_create_group(struct task_group *parent)
8512{
8513 struct task_group *tg;
8514 unsigned long flags;
8515
8516 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8517 if (!tg)
8518 return ERR_PTR(-ENOMEM);
8519
8520 if (!alloc_fair_sched_group(tg, parent))
8521 goto err;
8522
8523 if (!alloc_rt_sched_group(tg, parent))
8524 goto err;
8525
8526 spin_lock_irqsave(&task_group_lock, flags);
8527 list_add_rcu(&tg->list, &task_groups);
8528
8529 WARN_ON(!parent); /* root should already exist */
8530
8531 tg->parent = parent;
8532 INIT_LIST_HEAD(&tg->children);
8533 list_add_rcu(&tg->siblings, &parent->children);
8534 spin_unlock_irqrestore(&task_group_lock, flags);
8535
8536 return tg;
8537
8538err:
8539 free_sched_group(tg);
8540 return ERR_PTR(-ENOMEM);
8541}
8542
8543/* rcu callback to free various structures associated with a task group */
8544static void free_sched_group_rcu(struct rcu_head *rhp)
8545{
8546 /* now it should be safe to free those cfs_rqs */
8547 free_sched_group(container_of(rhp, struct task_group, rcu));
8548}
8549
8550/* Destroy runqueue etc associated with a task group */
8551void sched_destroy_group(struct task_group *tg)
8552{
8553 unsigned long flags;
8554 int i;
8555
8556 /* end participation in shares distribution */
8557 for_each_possible_cpu(i)
8558 unregister_fair_sched_group(tg, i);
8559
8560 spin_lock_irqsave(&task_group_lock, flags);
8561 list_del_rcu(&tg->list);
8562 list_del_rcu(&tg->siblings);
8563 spin_unlock_irqrestore(&task_group_lock, flags);
8564
8565 /* wait for possible concurrent references to cfs_rqs complete */
8566 call_rcu(&tg->rcu, free_sched_group_rcu);
8567}
8568
8569/* change task's runqueue when it moves between groups.
8570 * The caller of this function should have put the task in its new group
8571 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8572 * reflect its new group.
8573 */
8574void sched_move_task(struct task_struct *tsk)
8575{
8576 int on_rq, running;
8577 unsigned long flags;
8578 struct rq *rq;
8579
8580 rq = task_rq_lock(tsk, &flags);
8581
8582 running = task_current(rq, tsk);
8583 on_rq = tsk->on_rq;
8584
8585 if (on_rq)
8586 dequeue_task(rq, tsk, 0);
8587 if (unlikely(running))
8588 tsk->sched_class->put_prev_task(rq, tsk);
8589
8590#ifdef CONFIG_FAIR_GROUP_SCHED
8591 if (tsk->sched_class->task_move_group)
8592 tsk->sched_class->task_move_group(tsk, on_rq);
8593 else
8594#endif
8595 set_task_rq(tsk, task_cpu(tsk));
8596
8597 if (unlikely(running))
8598 tsk->sched_class->set_curr_task(rq);
8599 if (on_rq)
8600 enqueue_task(rq, tsk, 0);
8601
8602 task_rq_unlock(rq, tsk, &flags);
8603}
8604#endif /* CONFIG_CGROUP_SCHED */
8605
8606#ifdef CONFIG_FAIR_GROUP_SCHED
8607static DEFINE_MUTEX(shares_mutex);
8608
8609int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8610{
8611 int i;
8612 unsigned long flags;
8613
8614 /*
8615 * We can't change the weight of the root cgroup.
8616 */
8617 if (!tg->se[0])
8618 return -EINVAL;
8619
8620 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8621
8622 mutex_lock(&shares_mutex);
8623 if (tg->shares == shares)
8624 goto done;
8625
8626 tg->shares = shares;
8627 for_each_possible_cpu(i) {
8628 struct rq *rq = cpu_rq(i);
8629 struct sched_entity *se;
8630
8631 se = tg->se[i];
8632 /* Propagate contribution to hierarchy */
8633 raw_spin_lock_irqsave(&rq->lock, flags);
8634 for_each_sched_entity(se)
8635 update_cfs_shares(group_cfs_rq(se));
8636 raw_spin_unlock_irqrestore(&rq->lock, flags);
8637 }
8638
8639done:
8640 mutex_unlock(&shares_mutex);
8641 return 0;
8642}
8643
8644unsigned long sched_group_shares(struct task_group *tg)
8645{
8646 return tg->shares;
8647}
8648#endif
8649
8650#ifdef CONFIG_RT_GROUP_SCHED
8651/*
8652 * Ensure that the real time constraints are schedulable.
8653 */
8654static DEFINE_MUTEX(rt_constraints_mutex);
8655
8656static unsigned long to_ratio(u64 period, u64 runtime)
8657{
8658 if (runtime == RUNTIME_INF)
8659 return 1ULL << 20;
8660
8661 return div64_u64(runtime << 20, period);
8662}
8663
8664/* Must be called with tasklist_lock held */
8665static inline int tg_has_rt_tasks(struct task_group *tg)
8666{
8667 struct task_struct *g, *p;
8668
8669 do_each_thread(g, p) {
8670 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8671 return 1;
8672 } while_each_thread(g, p);
8673
8674 return 0;
8675}
8676
8677struct rt_schedulable_data {
8678 struct task_group *tg;
8679 u64 rt_period;
8680 u64 rt_runtime;
8681};
8682
8683static int tg_schedulable(struct task_group *tg, void *data)
8684{
8685 struct rt_schedulable_data *d = data;
8686 struct task_group *child;
8687 unsigned long total, sum = 0;
8688 u64 period, runtime;
8689
8690 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8691 runtime = tg->rt_bandwidth.rt_runtime;
8692
8693 if (tg == d->tg) {
8694 period = d->rt_period;
8695 runtime = d->rt_runtime;
8696 }
8697
8698 /*
8699 * Cannot have more runtime than the period.
8700 */
8701 if (runtime > period && runtime != RUNTIME_INF)
8702 return -EINVAL;
8703
8704 /*
8705 * Ensure we don't starve existing RT tasks.
8706 */
8707 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8708 return -EBUSY;
8709
8710 total = to_ratio(period, runtime);
8711
8712 /*
8713 * Nobody can have more than the global setting allows.
8714 */
8715 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8716 return -EINVAL;
8717
8718 /*
8719 * The sum of our children's runtime should not exceed our own.
8720 */
8721 list_for_each_entry_rcu(child, &tg->children, siblings) {
8722 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8723 runtime = child->rt_bandwidth.rt_runtime;
8724
8725 if (child == d->tg) {
8726 period = d->rt_period;
8727 runtime = d->rt_runtime;
8728 }
8729
8730 sum += to_ratio(period, runtime);
8731 }
8732
8733 if (sum > total)
8734 return -EINVAL;
8735
8736 return 0;
8737}
8738
8739static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8740{
8741 struct rt_schedulable_data data = {
8742 .tg = tg,
8743 .rt_period = period,
8744 .rt_runtime = runtime,
8745 };
8746
8747 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8748}
8749
8750static int tg_set_bandwidth(struct task_group *tg,
8751 u64 rt_period, u64 rt_runtime)
8752{
8753 int i, err = 0;
8754
8755 mutex_lock(&rt_constraints_mutex);
8756 read_lock(&tasklist_lock);
8757 err = __rt_schedulable(tg, rt_period, rt_runtime);
8758 if (err)
8759 goto unlock;
8760
8761 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8762 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8763 tg->rt_bandwidth.rt_runtime = rt_runtime;
8764
8765 for_each_possible_cpu(i) {
8766 struct rt_rq *rt_rq = tg->rt_rq[i];
8767
8768 raw_spin_lock(&rt_rq->rt_runtime_lock);
8769 rt_rq->rt_runtime = rt_runtime;
8770 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8771 }
8772 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8773unlock:
8774 read_unlock(&tasklist_lock);
8775 mutex_unlock(&rt_constraints_mutex);
8776
8777 return err;
8778}
8779
8780int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8781{
8782 u64 rt_runtime, rt_period;
8783
8784 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8785 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8786 if (rt_runtime_us < 0)
8787 rt_runtime = RUNTIME_INF;
8788
8789 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8790}
8791
8792long sched_group_rt_runtime(struct task_group *tg)
8793{
8794 u64 rt_runtime_us;
8795
8796 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8797 return -1;
8798
8799 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8800 do_div(rt_runtime_us, NSEC_PER_USEC);
8801 return rt_runtime_us;
8802}
8803
8804int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8805{
8806 u64 rt_runtime, rt_period;
8807
8808 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8809 rt_runtime = tg->rt_bandwidth.rt_runtime;
8810
8811 if (rt_period == 0)
8812 return -EINVAL;
8813
8814 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8815}
8816
8817long sched_group_rt_period(struct task_group *tg)
8818{
8819 u64 rt_period_us;
8820
8821 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8822 do_div(rt_period_us, NSEC_PER_USEC);
8823 return rt_period_us;
8824}
8825
8826static int sched_rt_global_constraints(void)
8827{
8828 u64 runtime, period;
8829 int ret = 0;
8830
8831 if (sysctl_sched_rt_period <= 0)
8832 return -EINVAL;
8833
8834 runtime = global_rt_runtime();
8835 period = global_rt_period();
8836
8837 /*
8838 * Sanity check on the sysctl variables.
8839 */
8840 if (runtime > period && runtime != RUNTIME_INF)
8841 return -EINVAL;
8842
8843 mutex_lock(&rt_constraints_mutex);
8844 read_lock(&tasklist_lock);
8845 ret = __rt_schedulable(NULL, 0, 0);
8846 read_unlock(&tasklist_lock);
8847 mutex_unlock(&rt_constraints_mutex);
8848
8849 return ret;
8850}
8851
8852int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8853{
8854 /* Don't accept realtime tasks when there is no way for them to run */
8855 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8856 return 0;
8857
8858 return 1;
8859}
8860
8861#else /* !CONFIG_RT_GROUP_SCHED */
8862static int sched_rt_global_constraints(void)
8863{
8864 unsigned long flags;
8865 int i;
8866
8867 if (sysctl_sched_rt_period <= 0)
8868 return -EINVAL;
8869
8870 /*
8871 * There's always some RT tasks in the root group
8872 * -- migration, kstopmachine etc..
8873 */
8874 if (sysctl_sched_rt_runtime == 0)
8875 return -EBUSY;
8876
8877 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8878 for_each_possible_cpu(i) {
8879 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8880
8881 raw_spin_lock(&rt_rq->rt_runtime_lock);
8882 rt_rq->rt_runtime = global_rt_runtime();
8883 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8884 }
8885 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8886
8887 return 0;
8888}
8889#endif /* CONFIG_RT_GROUP_SCHED */
8890
8891int sched_rt_handler(struct ctl_table *table, int write,
8892 void __user *buffer, size_t *lenp,
8893 loff_t *ppos)
8894{
8895 int ret;
8896 int old_period, old_runtime;
8897 static DEFINE_MUTEX(mutex);
8898
8899 mutex_lock(&mutex);
8900 old_period = sysctl_sched_rt_period;
8901 old_runtime = sysctl_sched_rt_runtime;
8902
8903 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8904
8905 if (!ret && write) {
8906 ret = sched_rt_global_constraints();
8907 if (ret) {
8908 sysctl_sched_rt_period = old_period;
8909 sysctl_sched_rt_runtime = old_runtime;
8910 } else {
8911 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8912 def_rt_bandwidth.rt_period =
8913 ns_to_ktime(global_rt_period());
8914 }
8915 }
8916 mutex_unlock(&mutex);
8917
8918 return ret;
8919}
8920
8921#ifdef CONFIG_CGROUP_SCHED
8922
8923/* return corresponding task_group object of a cgroup */
8924static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8925{
8926 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8927 struct task_group, css);
8928}
8929
8930static struct cgroup_subsys_state *
8931cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8932{
8933 struct task_group *tg, *parent;
8934
8935 if (!cgrp->parent) {
8936 /* This is early initialization for the top cgroup */
8937 return &root_task_group.css;
8938 }
8939
8940 parent = cgroup_tg(cgrp->parent);
8941 tg = sched_create_group(parent);
8942 if (IS_ERR(tg))
8943 return ERR_PTR(-ENOMEM);
8944
8945 return &tg->css;
8946}
8947
8948static void
8949cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8950{
8951 struct task_group *tg = cgroup_tg(cgrp);
8952
8953 sched_destroy_group(tg);
8954}
8955
8956static int
8957cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8958{
8959#ifdef CONFIG_RT_GROUP_SCHED
8960 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8961 return -EINVAL;
8962#else
8963 /* We don't support RT-tasks being in separate groups */
8964 if (tsk->sched_class != &fair_sched_class)
8965 return -EINVAL;
8966#endif
8967 return 0;
8968}
8969
8970static void
8971cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8972{
8973 sched_move_task(tsk);
8974}
8975
8976static void
8977cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
8978 struct cgroup *old_cgrp, struct task_struct *task)
8979{
8980 /*
8981 * cgroup_exit() is called in the copy_process() failure path.
8982 * Ignore this case since the task hasn't ran yet, this avoids
8983 * trying to poke a half freed task state from generic code.
8984 */
8985 if (!(task->flags & PF_EXITING))
8986 return;
8987
8988 sched_move_task(task);
8989}
8990
8991#ifdef CONFIG_FAIR_GROUP_SCHED
8992static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8993 u64 shareval)
8994{
8995 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
8996}
8997
8998static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8999{
9000 struct task_group *tg = cgroup_tg(cgrp);
9001
9002 return (u64) scale_load_down(tg->shares);
9003}
9004#endif /* CONFIG_FAIR_GROUP_SCHED */
9005
9006#ifdef CONFIG_RT_GROUP_SCHED
9007static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9008 s64 val)
9009{
9010 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9011}
9012
9013static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9014{
9015 return sched_group_rt_runtime(cgroup_tg(cgrp));
9016}
9017
9018static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9019 u64 rt_period_us)
9020{
9021 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9022}
9023
9024static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9025{
9026 return sched_group_rt_period(cgroup_tg(cgrp));
9027}
9028#endif /* CONFIG_RT_GROUP_SCHED */
9029
9030static struct cftype cpu_files[] = {
9031#ifdef CONFIG_FAIR_GROUP_SCHED
9032 {
9033 .name = "shares",
9034 .read_u64 = cpu_shares_read_u64,
9035 .write_u64 = cpu_shares_write_u64,
9036 },
9037#endif
9038#ifdef CONFIG_RT_GROUP_SCHED
9039 {
9040 .name = "rt_runtime_us",
9041 .read_s64 = cpu_rt_runtime_read,
9042 .write_s64 = cpu_rt_runtime_write,
9043 },
9044 {
9045 .name = "rt_period_us",
9046 .read_u64 = cpu_rt_period_read_uint,
9047 .write_u64 = cpu_rt_period_write_uint,
9048 },
9049#endif
9050};
9051
9052static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9053{
9054 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9055}
9056
9057struct cgroup_subsys cpu_cgroup_subsys = {
9058 .name = "cpu",
9059 .create = cpu_cgroup_create,
9060 .destroy = cpu_cgroup_destroy,
9061 .can_attach_task = cpu_cgroup_can_attach_task,
9062 .attach_task = cpu_cgroup_attach_task,
9063 .exit = cpu_cgroup_exit,
9064 .populate = cpu_cgroup_populate,
9065 .subsys_id = cpu_cgroup_subsys_id,
9066 .early_init = 1,
9067};
9068
9069#endif /* CONFIG_CGROUP_SCHED */
9070
9071#ifdef CONFIG_CGROUP_CPUACCT
9072
9073/*
9074 * CPU accounting code for task groups.
9075 *
9076 * Based on the work by Paul Menage ([email protected]) and Balbir Singh
9077 * ([email protected]).
9078 */
9079
9080/* track cpu usage of a group of tasks and its child groups */
9081struct cpuacct {
9082 struct cgroup_subsys_state css;
9083 /* cpuusage holds pointer to a u64-type object on every cpu */
9084 u64 __percpu *cpuusage;
9085 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9086 struct cpuacct *parent;
9087};
9088
9089struct cgroup_subsys cpuacct_subsys;
9090
9091/* return cpu accounting group corresponding to this container */
9092static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9093{
9094 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9095 struct cpuacct, css);
9096}
9097
9098/* return cpu accounting group to which this task belongs */
9099static inline struct cpuacct *task_ca(struct task_struct *tsk)
9100{
9101 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9102 struct cpuacct, css);
9103}
9104
9105/* create a new cpu accounting group */
9106static struct cgroup_subsys_state *cpuacct_create(
9107 struct cgroup_subsys *ss, struct cgroup *cgrp)
9108{
9109 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9110 int i;
9111
9112 if (!ca)
9113 goto out;
9114
9115 ca->cpuusage = alloc_percpu(u64);
9116 if (!ca->cpuusage)
9117 goto out_free_ca;
9118
9119 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9120 if (percpu_counter_init(&ca->cpustat[i], 0))
9121 goto out_free_counters;
9122
9123 if (cgrp->parent)
9124 ca->parent = cgroup_ca(cgrp->parent);
9125
9126 return &ca->css;
9127
9128out_free_counters:
9129 while (--i >= 0)
9130 percpu_counter_destroy(&ca->cpustat[i]);
9131 free_percpu(ca->cpuusage);
9132out_free_ca:
9133 kfree(ca);
9134out:
9135 return ERR_PTR(-ENOMEM);
9136}
9137
9138/* destroy an existing cpu accounting group */
9139static void
9140cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9141{
9142 struct cpuacct *ca = cgroup_ca(cgrp);
9143 int i;
9144
9145 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9146 percpu_counter_destroy(&ca->cpustat[i]);
9147 free_percpu(ca->cpuusage);
9148 kfree(ca);
9149}
9150
9151static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9152{
9153 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9154 u64 data;
9155
9156#ifndef CONFIG_64BIT
9157 /*
9158 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9159 */
9160 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9161 data = *cpuusage;
9162 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9163#else
9164 data = *cpuusage;
9165#endif
9166
9167 return data;
9168}
9169
9170static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9171{
9172 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9173
9174#ifndef CONFIG_64BIT
9175 /*
9176 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9177 */
9178 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9179 *cpuusage = val;
9180 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9181#else
9182 *cpuusage = val;
9183#endif
9184}
9185
9186/* return total cpu usage (in nanoseconds) of a group */
9187static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9188{
9189 struct cpuacct *ca = cgroup_ca(cgrp);
9190 u64 totalcpuusage = 0;
9191 int i;
9192
9193 for_each_present_cpu(i)
9194 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9195
9196 return totalcpuusage;
9197}
9198
9199static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9200 u64 reset)
9201{
9202 struct cpuacct *ca = cgroup_ca(cgrp);
9203 int err = 0;
9204 int i;
9205
9206 if (reset) {
9207 err = -EINVAL;
9208 goto out;
9209 }
9210
9211 for_each_present_cpu(i)
9212 cpuacct_cpuusage_write(ca, i, 0);
9213
9214out:
9215 return err;
9216}
9217
9218static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9219 struct seq_file *m)
9220{
9221 struct cpuacct *ca = cgroup_ca(cgroup);
9222 u64 percpu;
9223 int i;
9224
9225 for_each_present_cpu(i) {
9226 percpu = cpuacct_cpuusage_read(ca, i);
9227 seq_printf(m, "%llu ", (unsigned long long) percpu);
9228 }
9229 seq_printf(m, "\n");
9230 return 0;
9231}
9232
9233static const char *cpuacct_stat_desc[] = {
9234 [CPUACCT_STAT_USER] = "user",
9235 [CPUACCT_STAT_SYSTEM] = "system",
9236};
9237
9238static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9239 struct cgroup_map_cb *cb)
9240{
9241 struct cpuacct *ca = cgroup_ca(cgrp);
9242 int i;
9243
9244 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9245 s64 val = percpu_counter_read(&ca->cpustat[i]);
9246 val = cputime64_to_clock_t(val);
9247 cb->fill(cb, cpuacct_stat_desc[i], val);
9248 }
9249 return 0;
9250}
9251
9252static struct cftype files[] = {
9253 {
9254 .name = "usage",
9255 .read_u64 = cpuusage_read,
9256 .write_u64 = cpuusage_write,
9257 },
9258 {
9259 .name = "usage_percpu",
9260 .read_seq_string = cpuacct_percpu_seq_read,
9261 },
9262 {
9263 .name = "stat",
9264 .read_map = cpuacct_stats_show,
9265 },
9266};
9267
9268static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9269{
9270 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9271}
9272
9273/*
9274 * charge this task's execution time to its accounting group.
9275 *
9276 * called with rq->lock held.
9277 */
9278static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9279{
9280 struct cpuacct *ca;
9281 int cpu;
9282
9283 if (unlikely(!cpuacct_subsys.active))
9284 return;
9285
9286 cpu = task_cpu(tsk);
9287
9288 rcu_read_lock();
9289
9290 ca = task_ca(tsk);
9291
9292 for (; ca; ca = ca->parent) {
9293 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9294 *cpuusage += cputime;
9295 }
9296
9297 rcu_read_unlock();
9298}
9299
9300/*
9301 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9302 * in cputime_t units. As a result, cpuacct_update_stats calls
9303 * percpu_counter_add with values large enough to always overflow the
9304 * per cpu batch limit causing bad SMP scalability.
9305 *
9306 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9307 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9308 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9309 */
9310#ifdef CONFIG_SMP
9311#define CPUACCT_BATCH \
9312 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9313#else
9314#define CPUACCT_BATCH 0
9315#endif
9316
9317/*
9318 * Charge the system/user time to the task's accounting group.
9319 */
9320static void cpuacct_update_stats(struct task_struct *tsk,
9321 enum cpuacct_stat_index idx, cputime_t val)
9322{
9323 struct cpuacct *ca;
9324 int batch = CPUACCT_BATCH;
9325
9326 if (unlikely(!cpuacct_subsys.active))
9327 return;
9328
9329 rcu_read_lock();
9330 ca = task_ca(tsk);
9331
9332 do {
9333 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9334 ca = ca->parent;
9335 } while (ca);
9336 rcu_read_unlock();
9337}
9338
9339struct cgroup_subsys cpuacct_subsys = {
9340 .name = "cpuacct",
9341 .create = cpuacct_create,
9342 .destroy = cpuacct_destroy,
9343 .populate = cpuacct_populate,
9344 .subsys_id = cpuacct_subsys_id,
9345};
9346#endif /* CONFIG_CGROUP_CPUACCT */
9347