Loading...
1/*
2 * kernel/sched/core.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/sysctl.h>
60#include <linux/syscalls.h>
61#include <linux/times.h>
62#include <linux/tsacct_kern.h>
63#include <linux/kprobes.h>
64#include <linux/delayacct.h>
65#include <linux/unistd.h>
66#include <linux/pagemap.h>
67#include <linux/hrtimer.h>
68#include <linux/tick.h>
69#include <linux/debugfs.h>
70#include <linux/ctype.h>
71#include <linux/ftrace.h>
72#include <linux/slab.h>
73#include <linux/init_task.h>
74#include <linux/binfmts.h>
75
76#include <asm/switch_to.h>
77#include <asm/tlb.h>
78#include <asm/irq_regs.h>
79#include <asm/mutex.h>
80#ifdef CONFIG_PARAVIRT
81#include <asm/paravirt.h>
82#endif
83
84#include "sched.h"
85#include "../workqueue_sched.h"
86#include "../smpboot.h"
87
88#define CREATE_TRACE_POINTS
89#include <trace/events/sched.h>
90
91void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
92{
93 unsigned long delta;
94 ktime_t soft, hard, now;
95
96 for (;;) {
97 if (hrtimer_active(period_timer))
98 break;
99
100 now = hrtimer_cb_get_time(period_timer);
101 hrtimer_forward(period_timer, now, period);
102
103 soft = hrtimer_get_softexpires(period_timer);
104 hard = hrtimer_get_expires(period_timer);
105 delta = ktime_to_ns(ktime_sub(hard, soft));
106 __hrtimer_start_range_ns(period_timer, soft, delta,
107 HRTIMER_MODE_ABS_PINNED, 0);
108 }
109}
110
111DEFINE_MUTEX(sched_domains_mutex);
112DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
113
114static void update_rq_clock_task(struct rq *rq, s64 delta);
115
116void update_rq_clock(struct rq *rq)
117{
118 s64 delta;
119
120 if (rq->skip_clock_update > 0)
121 return;
122
123 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
124 rq->clock += delta;
125 update_rq_clock_task(rq, delta);
126}
127
128/*
129 * Debugging: various feature bits
130 */
131
132#define SCHED_FEAT(name, enabled) \
133 (1UL << __SCHED_FEAT_##name) * enabled |
134
135const_debug unsigned int sysctl_sched_features =
136#include "features.h"
137 0;
138
139#undef SCHED_FEAT
140
141#ifdef CONFIG_SCHED_DEBUG
142#define SCHED_FEAT(name, enabled) \
143 #name ,
144
145static const char * const sched_feat_names[] = {
146#include "features.h"
147};
148
149#undef SCHED_FEAT
150
151static int sched_feat_show(struct seq_file *m, void *v)
152{
153 int i;
154
155 for (i = 0; i < __SCHED_FEAT_NR; i++) {
156 if (!(sysctl_sched_features & (1UL << i)))
157 seq_puts(m, "NO_");
158 seq_printf(m, "%s ", sched_feat_names[i]);
159 }
160 seq_puts(m, "\n");
161
162 return 0;
163}
164
165#ifdef HAVE_JUMP_LABEL
166
167#define jump_label_key__true STATIC_KEY_INIT_TRUE
168#define jump_label_key__false STATIC_KEY_INIT_FALSE
169
170#define SCHED_FEAT(name, enabled) \
171 jump_label_key__##enabled ,
172
173struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
174#include "features.h"
175};
176
177#undef SCHED_FEAT
178
179static void sched_feat_disable(int i)
180{
181 if (static_key_enabled(&sched_feat_keys[i]))
182 static_key_slow_dec(&sched_feat_keys[i]);
183}
184
185static void sched_feat_enable(int i)
186{
187 if (!static_key_enabled(&sched_feat_keys[i]))
188 static_key_slow_inc(&sched_feat_keys[i]);
189}
190#else
191static void sched_feat_disable(int i) { };
192static void sched_feat_enable(int i) { };
193#endif /* HAVE_JUMP_LABEL */
194
195static ssize_t
196sched_feat_write(struct file *filp, const char __user *ubuf,
197 size_t cnt, loff_t *ppos)
198{
199 char buf[64];
200 char *cmp;
201 int neg = 0;
202 int i;
203
204 if (cnt > 63)
205 cnt = 63;
206
207 if (copy_from_user(&buf, ubuf, cnt))
208 return -EFAULT;
209
210 buf[cnt] = 0;
211 cmp = strstrip(buf);
212
213 if (strncmp(cmp, "NO_", 3) == 0) {
214 neg = 1;
215 cmp += 3;
216 }
217
218 for (i = 0; i < __SCHED_FEAT_NR; i++) {
219 if (strcmp(cmp, sched_feat_names[i]) == 0) {
220 if (neg) {
221 sysctl_sched_features &= ~(1UL << i);
222 sched_feat_disable(i);
223 } else {
224 sysctl_sched_features |= (1UL << i);
225 sched_feat_enable(i);
226 }
227 break;
228 }
229 }
230
231 if (i == __SCHED_FEAT_NR)
232 return -EINVAL;
233
234 *ppos += cnt;
235
236 return cnt;
237}
238
239static int sched_feat_open(struct inode *inode, struct file *filp)
240{
241 return single_open(filp, sched_feat_show, NULL);
242}
243
244static const struct file_operations sched_feat_fops = {
245 .open = sched_feat_open,
246 .write = sched_feat_write,
247 .read = seq_read,
248 .llseek = seq_lseek,
249 .release = single_release,
250};
251
252static __init int sched_init_debug(void)
253{
254 debugfs_create_file("sched_features", 0644, NULL, NULL,
255 &sched_feat_fops);
256
257 return 0;
258}
259late_initcall(sched_init_debug);
260#endif /* CONFIG_SCHED_DEBUG */
261
262/*
263 * Number of tasks to iterate in a single balance run.
264 * Limited because this is done with IRQs disabled.
265 */
266const_debug unsigned int sysctl_sched_nr_migrate = 32;
267
268/*
269 * period over which we average the RT time consumption, measured
270 * in ms.
271 *
272 * default: 1s
273 */
274const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
275
276/*
277 * period over which we measure -rt task cpu usage in us.
278 * default: 1s
279 */
280unsigned int sysctl_sched_rt_period = 1000000;
281
282__read_mostly int scheduler_running;
283
284/*
285 * part of the period that we allow rt tasks to run in us.
286 * default: 0.95s
287 */
288int sysctl_sched_rt_runtime = 950000;
289
290
291
292/*
293 * __task_rq_lock - lock the rq @p resides on.
294 */
295static inline struct rq *__task_rq_lock(struct task_struct *p)
296 __acquires(rq->lock)
297{
298 struct rq *rq;
299
300 lockdep_assert_held(&p->pi_lock);
301
302 for (;;) {
303 rq = task_rq(p);
304 raw_spin_lock(&rq->lock);
305 if (likely(rq == task_rq(p)))
306 return rq;
307 raw_spin_unlock(&rq->lock);
308 }
309}
310
311/*
312 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
313 */
314static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
315 __acquires(p->pi_lock)
316 __acquires(rq->lock)
317{
318 struct rq *rq;
319
320 for (;;) {
321 raw_spin_lock_irqsave(&p->pi_lock, *flags);
322 rq = task_rq(p);
323 raw_spin_lock(&rq->lock);
324 if (likely(rq == task_rq(p)))
325 return rq;
326 raw_spin_unlock(&rq->lock);
327 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
328 }
329}
330
331static void __task_rq_unlock(struct rq *rq)
332 __releases(rq->lock)
333{
334 raw_spin_unlock(&rq->lock);
335}
336
337static inline void
338task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
339 __releases(rq->lock)
340 __releases(p->pi_lock)
341{
342 raw_spin_unlock(&rq->lock);
343 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
344}
345
346/*
347 * this_rq_lock - lock this runqueue and disable interrupts.
348 */
349static struct rq *this_rq_lock(void)
350 __acquires(rq->lock)
351{
352 struct rq *rq;
353
354 local_irq_disable();
355 rq = this_rq();
356 raw_spin_lock(&rq->lock);
357
358 return rq;
359}
360
361#ifdef CONFIG_SCHED_HRTICK
362/*
363 * Use HR-timers to deliver accurate preemption points.
364 *
365 * Its all a bit involved since we cannot program an hrt while holding the
366 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
367 * reschedule event.
368 *
369 * When we get rescheduled we reprogram the hrtick_timer outside of the
370 * rq->lock.
371 */
372
373static void hrtick_clear(struct rq *rq)
374{
375 if (hrtimer_active(&rq->hrtick_timer))
376 hrtimer_cancel(&rq->hrtick_timer);
377}
378
379/*
380 * High-resolution timer tick.
381 * Runs from hardirq context with interrupts disabled.
382 */
383static enum hrtimer_restart hrtick(struct hrtimer *timer)
384{
385 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
386
387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
388
389 raw_spin_lock(&rq->lock);
390 update_rq_clock(rq);
391 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
392 raw_spin_unlock(&rq->lock);
393
394 return HRTIMER_NORESTART;
395}
396
397#ifdef CONFIG_SMP
398/*
399 * called from hardirq (IPI) context
400 */
401static void __hrtick_start(void *arg)
402{
403 struct rq *rq = arg;
404
405 raw_spin_lock(&rq->lock);
406 hrtimer_restart(&rq->hrtick_timer);
407 rq->hrtick_csd_pending = 0;
408 raw_spin_unlock(&rq->lock);
409}
410
411/*
412 * Called to set the hrtick timer state.
413 *
414 * called with rq->lock held and irqs disabled
415 */
416void hrtick_start(struct rq *rq, u64 delay)
417{
418 struct hrtimer *timer = &rq->hrtick_timer;
419 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
420
421 hrtimer_set_expires(timer, time);
422
423 if (rq == this_rq()) {
424 hrtimer_restart(timer);
425 } else if (!rq->hrtick_csd_pending) {
426 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
427 rq->hrtick_csd_pending = 1;
428 }
429}
430
431static int
432hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
433{
434 int cpu = (int)(long)hcpu;
435
436 switch (action) {
437 case CPU_UP_CANCELED:
438 case CPU_UP_CANCELED_FROZEN:
439 case CPU_DOWN_PREPARE:
440 case CPU_DOWN_PREPARE_FROZEN:
441 case CPU_DEAD:
442 case CPU_DEAD_FROZEN:
443 hrtick_clear(cpu_rq(cpu));
444 return NOTIFY_OK;
445 }
446
447 return NOTIFY_DONE;
448}
449
450static __init void init_hrtick(void)
451{
452 hotcpu_notifier(hotplug_hrtick, 0);
453}
454#else
455/*
456 * Called to set the hrtick timer state.
457 *
458 * called with rq->lock held and irqs disabled
459 */
460void hrtick_start(struct rq *rq, u64 delay)
461{
462 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
463 HRTIMER_MODE_REL_PINNED, 0);
464}
465
466static inline void init_hrtick(void)
467{
468}
469#endif /* CONFIG_SMP */
470
471static void init_rq_hrtick(struct rq *rq)
472{
473#ifdef CONFIG_SMP
474 rq->hrtick_csd_pending = 0;
475
476 rq->hrtick_csd.flags = 0;
477 rq->hrtick_csd.func = __hrtick_start;
478 rq->hrtick_csd.info = rq;
479#endif
480
481 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
482 rq->hrtick_timer.function = hrtick;
483}
484#else /* CONFIG_SCHED_HRTICK */
485static inline void hrtick_clear(struct rq *rq)
486{
487}
488
489static inline void init_rq_hrtick(struct rq *rq)
490{
491}
492
493static inline void init_hrtick(void)
494{
495}
496#endif /* CONFIG_SCHED_HRTICK */
497
498/*
499 * resched_task - mark a task 'to be rescheduled now'.
500 *
501 * On UP this means the setting of the need_resched flag, on SMP it
502 * might also involve a cross-CPU call to trigger the scheduler on
503 * the target CPU.
504 */
505#ifdef CONFIG_SMP
506
507#ifndef tsk_is_polling
508#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
509#endif
510
511void resched_task(struct task_struct *p)
512{
513 int cpu;
514
515 assert_raw_spin_locked(&task_rq(p)->lock);
516
517 if (test_tsk_need_resched(p))
518 return;
519
520 set_tsk_need_resched(p);
521
522 cpu = task_cpu(p);
523 if (cpu == smp_processor_id())
524 return;
525
526 /* NEED_RESCHED must be visible before we test polling */
527 smp_mb();
528 if (!tsk_is_polling(p))
529 smp_send_reschedule(cpu);
530}
531
532void resched_cpu(int cpu)
533{
534 struct rq *rq = cpu_rq(cpu);
535 unsigned long flags;
536
537 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
538 return;
539 resched_task(cpu_curr(cpu));
540 raw_spin_unlock_irqrestore(&rq->lock, flags);
541}
542
543#ifdef CONFIG_NO_HZ
544/*
545 * In the semi idle case, use the nearest busy cpu for migrating timers
546 * from an idle cpu. This is good for power-savings.
547 *
548 * We don't do similar optimization for completely idle system, as
549 * selecting an idle cpu will add more delays to the timers than intended
550 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
551 */
552int get_nohz_timer_target(void)
553{
554 int cpu = smp_processor_id();
555 int i;
556 struct sched_domain *sd;
557
558 rcu_read_lock();
559 for_each_domain(cpu, sd) {
560 for_each_cpu(i, sched_domain_span(sd)) {
561 if (!idle_cpu(i)) {
562 cpu = i;
563 goto unlock;
564 }
565 }
566 }
567unlock:
568 rcu_read_unlock();
569 return cpu;
570}
571/*
572 * When add_timer_on() enqueues a timer into the timer wheel of an
573 * idle CPU then this timer might expire before the next timer event
574 * which is scheduled to wake up that CPU. In case of a completely
575 * idle system the next event might even be infinite time into the
576 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
577 * leaves the inner idle loop so the newly added timer is taken into
578 * account when the CPU goes back to idle and evaluates the timer
579 * wheel for the next timer event.
580 */
581void wake_up_idle_cpu(int cpu)
582{
583 struct rq *rq = cpu_rq(cpu);
584
585 if (cpu == smp_processor_id())
586 return;
587
588 /*
589 * This is safe, as this function is called with the timer
590 * wheel base lock of (cpu) held. When the CPU is on the way
591 * to idle and has not yet set rq->curr to idle then it will
592 * be serialized on the timer wheel base lock and take the new
593 * timer into account automatically.
594 */
595 if (rq->curr != rq->idle)
596 return;
597
598 /*
599 * We can set TIF_RESCHED on the idle task of the other CPU
600 * lockless. The worst case is that the other CPU runs the
601 * idle task through an additional NOOP schedule()
602 */
603 set_tsk_need_resched(rq->idle);
604
605 /* NEED_RESCHED must be visible before we test polling */
606 smp_mb();
607 if (!tsk_is_polling(rq->idle))
608 smp_send_reschedule(cpu);
609}
610
611static inline bool got_nohz_idle_kick(void)
612{
613 int cpu = smp_processor_id();
614 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
615}
616
617#else /* CONFIG_NO_HZ */
618
619static inline bool got_nohz_idle_kick(void)
620{
621 return false;
622}
623
624#endif /* CONFIG_NO_HZ */
625
626void sched_avg_update(struct rq *rq)
627{
628 s64 period = sched_avg_period();
629
630 while ((s64)(rq->clock - rq->age_stamp) > period) {
631 /*
632 * Inline assembly required to prevent the compiler
633 * optimising this loop into a divmod call.
634 * See __iter_div_u64_rem() for another example of this.
635 */
636 asm("" : "+rm" (rq->age_stamp));
637 rq->age_stamp += period;
638 rq->rt_avg /= 2;
639 }
640}
641
642#else /* !CONFIG_SMP */
643void resched_task(struct task_struct *p)
644{
645 assert_raw_spin_locked(&task_rq(p)->lock);
646 set_tsk_need_resched(p);
647}
648#endif /* CONFIG_SMP */
649
650#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
651 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
652/*
653 * Iterate task_group tree rooted at *from, calling @down when first entering a
654 * node and @up when leaving it for the final time.
655 *
656 * Caller must hold rcu_lock or sufficient equivalent.
657 */
658int walk_tg_tree_from(struct task_group *from,
659 tg_visitor down, tg_visitor up, void *data)
660{
661 struct task_group *parent, *child;
662 int ret;
663
664 parent = from;
665
666down:
667 ret = (*down)(parent, data);
668 if (ret)
669 goto out;
670 list_for_each_entry_rcu(child, &parent->children, siblings) {
671 parent = child;
672 goto down;
673
674up:
675 continue;
676 }
677 ret = (*up)(parent, data);
678 if (ret || parent == from)
679 goto out;
680
681 child = parent;
682 parent = parent->parent;
683 if (parent)
684 goto up;
685out:
686 return ret;
687}
688
689int tg_nop(struct task_group *tg, void *data)
690{
691 return 0;
692}
693#endif
694
695static void set_load_weight(struct task_struct *p)
696{
697 int prio = p->static_prio - MAX_RT_PRIO;
698 struct load_weight *load = &p->se.load;
699
700 /*
701 * SCHED_IDLE tasks get minimal weight:
702 */
703 if (p->policy == SCHED_IDLE) {
704 load->weight = scale_load(WEIGHT_IDLEPRIO);
705 load->inv_weight = WMULT_IDLEPRIO;
706 return;
707 }
708
709 load->weight = scale_load(prio_to_weight[prio]);
710 load->inv_weight = prio_to_wmult[prio];
711}
712
713static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
714{
715 update_rq_clock(rq);
716 sched_info_queued(p);
717 p->sched_class->enqueue_task(rq, p, flags);
718}
719
720static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
721{
722 update_rq_clock(rq);
723 sched_info_dequeued(p);
724 p->sched_class->dequeue_task(rq, p, flags);
725}
726
727void activate_task(struct rq *rq, struct task_struct *p, int flags)
728{
729 if (task_contributes_to_load(p))
730 rq->nr_uninterruptible--;
731
732 enqueue_task(rq, p, flags);
733}
734
735void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
736{
737 if (task_contributes_to_load(p))
738 rq->nr_uninterruptible++;
739
740 dequeue_task(rq, p, flags);
741}
742
743#ifdef CONFIG_IRQ_TIME_ACCOUNTING
744
745/*
746 * There are no locks covering percpu hardirq/softirq time.
747 * They are only modified in account_system_vtime, on corresponding CPU
748 * with interrupts disabled. So, writes are safe.
749 * They are read and saved off onto struct rq in update_rq_clock().
750 * This may result in other CPU reading this CPU's irq time and can
751 * race with irq/account_system_vtime on this CPU. We would either get old
752 * or new value with a side effect of accounting a slice of irq time to wrong
753 * task when irq is in progress while we read rq->clock. That is a worthy
754 * compromise in place of having locks on each irq in account_system_time.
755 */
756static DEFINE_PER_CPU(u64, cpu_hardirq_time);
757static DEFINE_PER_CPU(u64, cpu_softirq_time);
758
759static DEFINE_PER_CPU(u64, irq_start_time);
760static int sched_clock_irqtime;
761
762void enable_sched_clock_irqtime(void)
763{
764 sched_clock_irqtime = 1;
765}
766
767void disable_sched_clock_irqtime(void)
768{
769 sched_clock_irqtime = 0;
770}
771
772#ifndef CONFIG_64BIT
773static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
774
775static inline void irq_time_write_begin(void)
776{
777 __this_cpu_inc(irq_time_seq.sequence);
778 smp_wmb();
779}
780
781static inline void irq_time_write_end(void)
782{
783 smp_wmb();
784 __this_cpu_inc(irq_time_seq.sequence);
785}
786
787static inline u64 irq_time_read(int cpu)
788{
789 u64 irq_time;
790 unsigned seq;
791
792 do {
793 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
794 irq_time = per_cpu(cpu_softirq_time, cpu) +
795 per_cpu(cpu_hardirq_time, cpu);
796 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
797
798 return irq_time;
799}
800#else /* CONFIG_64BIT */
801static inline void irq_time_write_begin(void)
802{
803}
804
805static inline void irq_time_write_end(void)
806{
807}
808
809static inline u64 irq_time_read(int cpu)
810{
811 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
812}
813#endif /* CONFIG_64BIT */
814
815/*
816 * Called before incrementing preempt_count on {soft,}irq_enter
817 * and before decrementing preempt_count on {soft,}irq_exit.
818 */
819void account_system_vtime(struct task_struct *curr)
820{
821 unsigned long flags;
822 s64 delta;
823 int cpu;
824
825 if (!sched_clock_irqtime)
826 return;
827
828 local_irq_save(flags);
829
830 cpu = smp_processor_id();
831 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
832 __this_cpu_add(irq_start_time, delta);
833
834 irq_time_write_begin();
835 /*
836 * We do not account for softirq time from ksoftirqd here.
837 * We want to continue accounting softirq time to ksoftirqd thread
838 * in that case, so as not to confuse scheduler with a special task
839 * that do not consume any time, but still wants to run.
840 */
841 if (hardirq_count())
842 __this_cpu_add(cpu_hardirq_time, delta);
843 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
844 __this_cpu_add(cpu_softirq_time, delta);
845
846 irq_time_write_end();
847 local_irq_restore(flags);
848}
849EXPORT_SYMBOL_GPL(account_system_vtime);
850
851#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
852
853#ifdef CONFIG_PARAVIRT
854static inline u64 steal_ticks(u64 steal)
855{
856 if (unlikely(steal > NSEC_PER_SEC))
857 return div_u64(steal, TICK_NSEC);
858
859 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
860}
861#endif
862
863static void update_rq_clock_task(struct rq *rq, s64 delta)
864{
865/*
866 * In theory, the compile should just see 0 here, and optimize out the call
867 * to sched_rt_avg_update. But I don't trust it...
868 */
869#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
870 s64 steal = 0, irq_delta = 0;
871#endif
872#ifdef CONFIG_IRQ_TIME_ACCOUNTING
873 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
874
875 /*
876 * Since irq_time is only updated on {soft,}irq_exit, we might run into
877 * this case when a previous update_rq_clock() happened inside a
878 * {soft,}irq region.
879 *
880 * When this happens, we stop ->clock_task and only update the
881 * prev_irq_time stamp to account for the part that fit, so that a next
882 * update will consume the rest. This ensures ->clock_task is
883 * monotonic.
884 *
885 * It does however cause some slight miss-attribution of {soft,}irq
886 * time, a more accurate solution would be to update the irq_time using
887 * the current rq->clock timestamp, except that would require using
888 * atomic ops.
889 */
890 if (irq_delta > delta)
891 irq_delta = delta;
892
893 rq->prev_irq_time += irq_delta;
894 delta -= irq_delta;
895#endif
896#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
897 if (static_key_false((¶virt_steal_rq_enabled))) {
898 u64 st;
899
900 steal = paravirt_steal_clock(cpu_of(rq));
901 steal -= rq->prev_steal_time_rq;
902
903 if (unlikely(steal > delta))
904 steal = delta;
905
906 st = steal_ticks(steal);
907 steal = st * TICK_NSEC;
908
909 rq->prev_steal_time_rq += steal;
910
911 delta -= steal;
912 }
913#endif
914
915 rq->clock_task += delta;
916
917#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
918 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
919 sched_rt_avg_update(rq, irq_delta + steal);
920#endif
921}
922
923#ifdef CONFIG_IRQ_TIME_ACCOUNTING
924static int irqtime_account_hi_update(void)
925{
926 u64 *cpustat = kcpustat_this_cpu->cpustat;
927 unsigned long flags;
928 u64 latest_ns;
929 int ret = 0;
930
931 local_irq_save(flags);
932 latest_ns = this_cpu_read(cpu_hardirq_time);
933 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
934 ret = 1;
935 local_irq_restore(flags);
936 return ret;
937}
938
939static int irqtime_account_si_update(void)
940{
941 u64 *cpustat = kcpustat_this_cpu->cpustat;
942 unsigned long flags;
943 u64 latest_ns;
944 int ret = 0;
945
946 local_irq_save(flags);
947 latest_ns = this_cpu_read(cpu_softirq_time);
948 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
949 ret = 1;
950 local_irq_restore(flags);
951 return ret;
952}
953
954#else /* CONFIG_IRQ_TIME_ACCOUNTING */
955
956#define sched_clock_irqtime (0)
957
958#endif
959
960void sched_set_stop_task(int cpu, struct task_struct *stop)
961{
962 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
963 struct task_struct *old_stop = cpu_rq(cpu)->stop;
964
965 if (stop) {
966 /*
967 * Make it appear like a SCHED_FIFO task, its something
968 * userspace knows about and won't get confused about.
969 *
970 * Also, it will make PI more or less work without too
971 * much confusion -- but then, stop work should not
972 * rely on PI working anyway.
973 */
974 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
975
976 stop->sched_class = &stop_sched_class;
977 }
978
979 cpu_rq(cpu)->stop = stop;
980
981 if (old_stop) {
982 /*
983 * Reset it back to a normal scheduling class so that
984 * it can die in pieces.
985 */
986 old_stop->sched_class = &rt_sched_class;
987 }
988}
989
990/*
991 * __normal_prio - return the priority that is based on the static prio
992 */
993static inline int __normal_prio(struct task_struct *p)
994{
995 return p->static_prio;
996}
997
998/*
999 * Calculate the expected normal priority: i.e. priority
1000 * without taking RT-inheritance into account. Might be
1001 * boosted by interactivity modifiers. Changes upon fork,
1002 * setprio syscalls, and whenever the interactivity
1003 * estimator recalculates.
1004 */
1005static inline int normal_prio(struct task_struct *p)
1006{
1007 int prio;
1008
1009 if (task_has_rt_policy(p))
1010 prio = MAX_RT_PRIO-1 - p->rt_priority;
1011 else
1012 prio = __normal_prio(p);
1013 return prio;
1014}
1015
1016/*
1017 * Calculate the current priority, i.e. the priority
1018 * taken into account by the scheduler. This value might
1019 * be boosted by RT tasks, or might be boosted by
1020 * interactivity modifiers. Will be RT if the task got
1021 * RT-boosted. If not then it returns p->normal_prio.
1022 */
1023static int effective_prio(struct task_struct *p)
1024{
1025 p->normal_prio = normal_prio(p);
1026 /*
1027 * If we are RT tasks or we were boosted to RT priority,
1028 * keep the priority unchanged. Otherwise, update priority
1029 * to the normal priority:
1030 */
1031 if (!rt_prio(p->prio))
1032 return p->normal_prio;
1033 return p->prio;
1034}
1035
1036/**
1037 * task_curr - is this task currently executing on a CPU?
1038 * @p: the task in question.
1039 */
1040inline int task_curr(const struct task_struct *p)
1041{
1042 return cpu_curr(task_cpu(p)) == p;
1043}
1044
1045static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1046 const struct sched_class *prev_class,
1047 int oldprio)
1048{
1049 if (prev_class != p->sched_class) {
1050 if (prev_class->switched_from)
1051 prev_class->switched_from(rq, p);
1052 p->sched_class->switched_to(rq, p);
1053 } else if (oldprio != p->prio)
1054 p->sched_class->prio_changed(rq, p, oldprio);
1055}
1056
1057void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1058{
1059 const struct sched_class *class;
1060
1061 if (p->sched_class == rq->curr->sched_class) {
1062 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1063 } else {
1064 for_each_class(class) {
1065 if (class == rq->curr->sched_class)
1066 break;
1067 if (class == p->sched_class) {
1068 resched_task(rq->curr);
1069 break;
1070 }
1071 }
1072 }
1073
1074 /*
1075 * A queue event has occurred, and we're going to schedule. In
1076 * this case, we can save a useless back to back clock update.
1077 */
1078 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1079 rq->skip_clock_update = 1;
1080}
1081
1082#ifdef CONFIG_SMP
1083void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1084{
1085#ifdef CONFIG_SCHED_DEBUG
1086 /*
1087 * We should never call set_task_cpu() on a blocked task,
1088 * ttwu() will sort out the placement.
1089 */
1090 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1091 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
1092
1093#ifdef CONFIG_LOCKDEP
1094 /*
1095 * The caller should hold either p->pi_lock or rq->lock, when changing
1096 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1097 *
1098 * sched_move_task() holds both and thus holding either pins the cgroup,
1099 * see task_group().
1100 *
1101 * Furthermore, all task_rq users should acquire both locks, see
1102 * task_rq_lock().
1103 */
1104 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1105 lockdep_is_held(&task_rq(p)->lock)));
1106#endif
1107#endif
1108
1109 trace_sched_migrate_task(p, new_cpu);
1110
1111 if (task_cpu(p) != new_cpu) {
1112 p->se.nr_migrations++;
1113 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1114 }
1115
1116 __set_task_cpu(p, new_cpu);
1117}
1118
1119struct migration_arg {
1120 struct task_struct *task;
1121 int dest_cpu;
1122};
1123
1124static int migration_cpu_stop(void *data);
1125
1126/*
1127 * wait_task_inactive - wait for a thread to unschedule.
1128 *
1129 * If @match_state is nonzero, it's the @p->state value just checked and
1130 * not expected to change. If it changes, i.e. @p might have woken up,
1131 * then return zero. When we succeed in waiting for @p to be off its CPU,
1132 * we return a positive number (its total switch count). If a second call
1133 * a short while later returns the same number, the caller can be sure that
1134 * @p has remained unscheduled the whole time.
1135 *
1136 * The caller must ensure that the task *will* unschedule sometime soon,
1137 * else this function might spin for a *long* time. This function can't
1138 * be called with interrupts off, or it may introduce deadlock with
1139 * smp_call_function() if an IPI is sent by the same process we are
1140 * waiting to become inactive.
1141 */
1142unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1143{
1144 unsigned long flags;
1145 int running, on_rq;
1146 unsigned long ncsw;
1147 struct rq *rq;
1148
1149 for (;;) {
1150 /*
1151 * We do the initial early heuristics without holding
1152 * any task-queue locks at all. We'll only try to get
1153 * the runqueue lock when things look like they will
1154 * work out!
1155 */
1156 rq = task_rq(p);
1157
1158 /*
1159 * If the task is actively running on another CPU
1160 * still, just relax and busy-wait without holding
1161 * any locks.
1162 *
1163 * NOTE! Since we don't hold any locks, it's not
1164 * even sure that "rq" stays as the right runqueue!
1165 * But we don't care, since "task_running()" will
1166 * return false if the runqueue has changed and p
1167 * is actually now running somewhere else!
1168 */
1169 while (task_running(rq, p)) {
1170 if (match_state && unlikely(p->state != match_state))
1171 return 0;
1172 cpu_relax();
1173 }
1174
1175 /*
1176 * Ok, time to look more closely! We need the rq
1177 * lock now, to be *sure*. If we're wrong, we'll
1178 * just go back and repeat.
1179 */
1180 rq = task_rq_lock(p, &flags);
1181 trace_sched_wait_task(p);
1182 running = task_running(rq, p);
1183 on_rq = p->on_rq;
1184 ncsw = 0;
1185 if (!match_state || p->state == match_state)
1186 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1187 task_rq_unlock(rq, p, &flags);
1188
1189 /*
1190 * If it changed from the expected state, bail out now.
1191 */
1192 if (unlikely(!ncsw))
1193 break;
1194
1195 /*
1196 * Was it really running after all now that we
1197 * checked with the proper locks actually held?
1198 *
1199 * Oops. Go back and try again..
1200 */
1201 if (unlikely(running)) {
1202 cpu_relax();
1203 continue;
1204 }
1205
1206 /*
1207 * It's not enough that it's not actively running,
1208 * it must be off the runqueue _entirely_, and not
1209 * preempted!
1210 *
1211 * So if it was still runnable (but just not actively
1212 * running right now), it's preempted, and we should
1213 * yield - it could be a while.
1214 */
1215 if (unlikely(on_rq)) {
1216 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1217
1218 set_current_state(TASK_UNINTERRUPTIBLE);
1219 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1220 continue;
1221 }
1222
1223 /*
1224 * Ahh, all good. It wasn't running, and it wasn't
1225 * runnable, which means that it will never become
1226 * running in the future either. We're all done!
1227 */
1228 break;
1229 }
1230
1231 return ncsw;
1232}
1233
1234/***
1235 * kick_process - kick a running thread to enter/exit the kernel
1236 * @p: the to-be-kicked thread
1237 *
1238 * Cause a process which is running on another CPU to enter
1239 * kernel-mode, without any delay. (to get signals handled.)
1240 *
1241 * NOTE: this function doesn't have to take the runqueue lock,
1242 * because all it wants to ensure is that the remote task enters
1243 * the kernel. If the IPI races and the task has been migrated
1244 * to another CPU then no harm is done and the purpose has been
1245 * achieved as well.
1246 */
1247void kick_process(struct task_struct *p)
1248{
1249 int cpu;
1250
1251 preempt_disable();
1252 cpu = task_cpu(p);
1253 if ((cpu != smp_processor_id()) && task_curr(p))
1254 smp_send_reschedule(cpu);
1255 preempt_enable();
1256}
1257EXPORT_SYMBOL_GPL(kick_process);
1258#endif /* CONFIG_SMP */
1259
1260#ifdef CONFIG_SMP
1261/*
1262 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1263 */
1264static int select_fallback_rq(int cpu, struct task_struct *p)
1265{
1266 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1267 enum { cpuset, possible, fail } state = cpuset;
1268 int dest_cpu;
1269
1270 /* Look for allowed, online CPU in same node. */
1271 for_each_cpu(dest_cpu, nodemask) {
1272 if (!cpu_online(dest_cpu))
1273 continue;
1274 if (!cpu_active(dest_cpu))
1275 continue;
1276 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1277 return dest_cpu;
1278 }
1279
1280 for (;;) {
1281 /* Any allowed, online CPU? */
1282 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1283 if (!cpu_online(dest_cpu))
1284 continue;
1285 if (!cpu_active(dest_cpu))
1286 continue;
1287 goto out;
1288 }
1289
1290 switch (state) {
1291 case cpuset:
1292 /* No more Mr. Nice Guy. */
1293 cpuset_cpus_allowed_fallback(p);
1294 state = possible;
1295 break;
1296
1297 case possible:
1298 do_set_cpus_allowed(p, cpu_possible_mask);
1299 state = fail;
1300 break;
1301
1302 case fail:
1303 BUG();
1304 break;
1305 }
1306 }
1307
1308out:
1309 if (state != cpuset) {
1310 /*
1311 * Don't tell them about moving exiting tasks or
1312 * kernel threads (both mm NULL), since they never
1313 * leave kernel.
1314 */
1315 if (p->mm && printk_ratelimit()) {
1316 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1317 task_pid_nr(p), p->comm, cpu);
1318 }
1319 }
1320
1321 return dest_cpu;
1322}
1323
1324/*
1325 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1326 */
1327static inline
1328int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1329{
1330 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1331
1332 /*
1333 * In order not to call set_task_cpu() on a blocking task we need
1334 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1335 * cpu.
1336 *
1337 * Since this is common to all placement strategies, this lives here.
1338 *
1339 * [ this allows ->select_task() to simply return task_cpu(p) and
1340 * not worry about this generic constraint ]
1341 */
1342 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1343 !cpu_online(cpu)))
1344 cpu = select_fallback_rq(task_cpu(p), p);
1345
1346 return cpu;
1347}
1348
1349static void update_avg(u64 *avg, u64 sample)
1350{
1351 s64 diff = sample - *avg;
1352 *avg += diff >> 3;
1353}
1354#endif
1355
1356static void
1357ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1358{
1359#ifdef CONFIG_SCHEDSTATS
1360 struct rq *rq = this_rq();
1361
1362#ifdef CONFIG_SMP
1363 int this_cpu = smp_processor_id();
1364
1365 if (cpu == this_cpu) {
1366 schedstat_inc(rq, ttwu_local);
1367 schedstat_inc(p, se.statistics.nr_wakeups_local);
1368 } else {
1369 struct sched_domain *sd;
1370
1371 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1372 rcu_read_lock();
1373 for_each_domain(this_cpu, sd) {
1374 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1375 schedstat_inc(sd, ttwu_wake_remote);
1376 break;
1377 }
1378 }
1379 rcu_read_unlock();
1380 }
1381
1382 if (wake_flags & WF_MIGRATED)
1383 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1384
1385#endif /* CONFIG_SMP */
1386
1387 schedstat_inc(rq, ttwu_count);
1388 schedstat_inc(p, se.statistics.nr_wakeups);
1389
1390 if (wake_flags & WF_SYNC)
1391 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1392
1393#endif /* CONFIG_SCHEDSTATS */
1394}
1395
1396static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1397{
1398 activate_task(rq, p, en_flags);
1399 p->on_rq = 1;
1400
1401 /* if a worker is waking up, notify workqueue */
1402 if (p->flags & PF_WQ_WORKER)
1403 wq_worker_waking_up(p, cpu_of(rq));
1404}
1405
1406/*
1407 * Mark the task runnable and perform wakeup-preemption.
1408 */
1409static void
1410ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1411{
1412 trace_sched_wakeup(p, true);
1413 check_preempt_curr(rq, p, wake_flags);
1414
1415 p->state = TASK_RUNNING;
1416#ifdef CONFIG_SMP
1417 if (p->sched_class->task_woken)
1418 p->sched_class->task_woken(rq, p);
1419
1420 if (rq->idle_stamp) {
1421 u64 delta = rq->clock - rq->idle_stamp;
1422 u64 max = 2*sysctl_sched_migration_cost;
1423
1424 if (delta > max)
1425 rq->avg_idle = max;
1426 else
1427 update_avg(&rq->avg_idle, delta);
1428 rq->idle_stamp = 0;
1429 }
1430#endif
1431}
1432
1433static void
1434ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1435{
1436#ifdef CONFIG_SMP
1437 if (p->sched_contributes_to_load)
1438 rq->nr_uninterruptible--;
1439#endif
1440
1441 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1442 ttwu_do_wakeup(rq, p, wake_flags);
1443}
1444
1445/*
1446 * Called in case the task @p isn't fully descheduled from its runqueue,
1447 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1448 * since all we need to do is flip p->state to TASK_RUNNING, since
1449 * the task is still ->on_rq.
1450 */
1451static int ttwu_remote(struct task_struct *p, int wake_flags)
1452{
1453 struct rq *rq;
1454 int ret = 0;
1455
1456 rq = __task_rq_lock(p);
1457 if (p->on_rq) {
1458 ttwu_do_wakeup(rq, p, wake_flags);
1459 ret = 1;
1460 }
1461 __task_rq_unlock(rq);
1462
1463 return ret;
1464}
1465
1466#ifdef CONFIG_SMP
1467static void sched_ttwu_pending(void)
1468{
1469 struct rq *rq = this_rq();
1470 struct llist_node *llist = llist_del_all(&rq->wake_list);
1471 struct task_struct *p;
1472
1473 raw_spin_lock(&rq->lock);
1474
1475 while (llist) {
1476 p = llist_entry(llist, struct task_struct, wake_entry);
1477 llist = llist_next(llist);
1478 ttwu_do_activate(rq, p, 0);
1479 }
1480
1481 raw_spin_unlock(&rq->lock);
1482}
1483
1484void scheduler_ipi(void)
1485{
1486 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1487 return;
1488
1489 /*
1490 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1491 * traditionally all their work was done from the interrupt return
1492 * path. Now that we actually do some work, we need to make sure
1493 * we do call them.
1494 *
1495 * Some archs already do call them, luckily irq_enter/exit nest
1496 * properly.
1497 *
1498 * Arguably we should visit all archs and update all handlers,
1499 * however a fair share of IPIs are still resched only so this would
1500 * somewhat pessimize the simple resched case.
1501 */
1502 irq_enter();
1503 sched_ttwu_pending();
1504
1505 /*
1506 * Check if someone kicked us for doing the nohz idle load balance.
1507 */
1508 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1509 this_rq()->idle_balance = 1;
1510 raise_softirq_irqoff(SCHED_SOFTIRQ);
1511 }
1512 irq_exit();
1513}
1514
1515static void ttwu_queue_remote(struct task_struct *p, int cpu)
1516{
1517 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1518 smp_send_reschedule(cpu);
1519}
1520
1521#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1522static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
1523{
1524 struct rq *rq;
1525 int ret = 0;
1526
1527 rq = __task_rq_lock(p);
1528 if (p->on_cpu) {
1529 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1530 ttwu_do_wakeup(rq, p, wake_flags);
1531 ret = 1;
1532 }
1533 __task_rq_unlock(rq);
1534
1535 return ret;
1536
1537}
1538#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1539
1540bool cpus_share_cache(int this_cpu, int that_cpu)
1541{
1542 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1543}
1544#endif /* CONFIG_SMP */
1545
1546static void ttwu_queue(struct task_struct *p, int cpu)
1547{
1548 struct rq *rq = cpu_rq(cpu);
1549
1550#if defined(CONFIG_SMP)
1551 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1552 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1553 ttwu_queue_remote(p, cpu);
1554 return;
1555 }
1556#endif
1557
1558 raw_spin_lock(&rq->lock);
1559 ttwu_do_activate(rq, p, 0);
1560 raw_spin_unlock(&rq->lock);
1561}
1562
1563/**
1564 * try_to_wake_up - wake up a thread
1565 * @p: the thread to be awakened
1566 * @state: the mask of task states that can be woken
1567 * @wake_flags: wake modifier flags (WF_*)
1568 *
1569 * Put it on the run-queue if it's not already there. The "current"
1570 * thread is always on the run-queue (except when the actual
1571 * re-schedule is in progress), and as such you're allowed to do
1572 * the simpler "current->state = TASK_RUNNING" to mark yourself
1573 * runnable without the overhead of this.
1574 *
1575 * Returns %true if @p was woken up, %false if it was already running
1576 * or @state didn't match @p's state.
1577 */
1578static int
1579try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1580{
1581 unsigned long flags;
1582 int cpu, success = 0;
1583
1584 smp_wmb();
1585 raw_spin_lock_irqsave(&p->pi_lock, flags);
1586 if (!(p->state & state))
1587 goto out;
1588
1589 success = 1; /* we're going to change ->state */
1590 cpu = task_cpu(p);
1591
1592 if (p->on_rq && ttwu_remote(p, wake_flags))
1593 goto stat;
1594
1595#ifdef CONFIG_SMP
1596 /*
1597 * If the owning (remote) cpu is still in the middle of schedule() with
1598 * this task as prev, wait until its done referencing the task.
1599 */
1600 while (p->on_cpu) {
1601#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1602 /*
1603 * In case the architecture enables interrupts in
1604 * context_switch(), we cannot busy wait, since that
1605 * would lead to deadlocks when an interrupt hits and
1606 * tries to wake up @prev. So bail and do a complete
1607 * remote wakeup.
1608 */
1609 if (ttwu_activate_remote(p, wake_flags))
1610 goto stat;
1611#else
1612 cpu_relax();
1613#endif
1614 }
1615 /*
1616 * Pairs with the smp_wmb() in finish_lock_switch().
1617 */
1618 smp_rmb();
1619
1620 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1621 p->state = TASK_WAKING;
1622
1623 if (p->sched_class->task_waking)
1624 p->sched_class->task_waking(p);
1625
1626 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1627 if (task_cpu(p) != cpu) {
1628 wake_flags |= WF_MIGRATED;
1629 set_task_cpu(p, cpu);
1630 }
1631#endif /* CONFIG_SMP */
1632
1633 ttwu_queue(p, cpu);
1634stat:
1635 ttwu_stat(p, cpu, wake_flags);
1636out:
1637 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1638
1639 return success;
1640}
1641
1642/**
1643 * try_to_wake_up_local - try to wake up a local task with rq lock held
1644 * @p: the thread to be awakened
1645 *
1646 * Put @p on the run-queue if it's not already there. The caller must
1647 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1648 * the current task.
1649 */
1650static void try_to_wake_up_local(struct task_struct *p)
1651{
1652 struct rq *rq = task_rq(p);
1653
1654 BUG_ON(rq != this_rq());
1655 BUG_ON(p == current);
1656 lockdep_assert_held(&rq->lock);
1657
1658 if (!raw_spin_trylock(&p->pi_lock)) {
1659 raw_spin_unlock(&rq->lock);
1660 raw_spin_lock(&p->pi_lock);
1661 raw_spin_lock(&rq->lock);
1662 }
1663
1664 if (!(p->state & TASK_NORMAL))
1665 goto out;
1666
1667 if (!p->on_rq)
1668 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1669
1670 ttwu_do_wakeup(rq, p, 0);
1671 ttwu_stat(p, smp_processor_id(), 0);
1672out:
1673 raw_spin_unlock(&p->pi_lock);
1674}
1675
1676/**
1677 * wake_up_process - Wake up a specific process
1678 * @p: The process to be woken up.
1679 *
1680 * Attempt to wake up the nominated process and move it to the set of runnable
1681 * processes. Returns 1 if the process was woken up, 0 if it was already
1682 * running.
1683 *
1684 * It may be assumed that this function implies a write memory barrier before
1685 * changing the task state if and only if any tasks are woken up.
1686 */
1687int wake_up_process(struct task_struct *p)
1688{
1689 return try_to_wake_up(p, TASK_ALL, 0);
1690}
1691EXPORT_SYMBOL(wake_up_process);
1692
1693int wake_up_state(struct task_struct *p, unsigned int state)
1694{
1695 return try_to_wake_up(p, state, 0);
1696}
1697
1698/*
1699 * Perform scheduler related setup for a newly forked process p.
1700 * p is forked by current.
1701 *
1702 * __sched_fork() is basic setup used by init_idle() too:
1703 */
1704static void __sched_fork(struct task_struct *p)
1705{
1706 p->on_rq = 0;
1707
1708 p->se.on_rq = 0;
1709 p->se.exec_start = 0;
1710 p->se.sum_exec_runtime = 0;
1711 p->se.prev_sum_exec_runtime = 0;
1712 p->se.nr_migrations = 0;
1713 p->se.vruntime = 0;
1714 INIT_LIST_HEAD(&p->se.group_node);
1715
1716#ifdef CONFIG_SCHEDSTATS
1717 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1718#endif
1719
1720 INIT_LIST_HEAD(&p->rt.run_list);
1721
1722#ifdef CONFIG_PREEMPT_NOTIFIERS
1723 INIT_HLIST_HEAD(&p->preempt_notifiers);
1724#endif
1725}
1726
1727/*
1728 * fork()/clone()-time setup:
1729 */
1730void sched_fork(struct task_struct *p)
1731{
1732 unsigned long flags;
1733 int cpu = get_cpu();
1734
1735 __sched_fork(p);
1736 /*
1737 * We mark the process as running here. This guarantees that
1738 * nobody will actually run it, and a signal or other external
1739 * event cannot wake it up and insert it on the runqueue either.
1740 */
1741 p->state = TASK_RUNNING;
1742
1743 /*
1744 * Make sure we do not leak PI boosting priority to the child.
1745 */
1746 p->prio = current->normal_prio;
1747
1748 /*
1749 * Revert to default priority/policy on fork if requested.
1750 */
1751 if (unlikely(p->sched_reset_on_fork)) {
1752 if (task_has_rt_policy(p)) {
1753 p->policy = SCHED_NORMAL;
1754 p->static_prio = NICE_TO_PRIO(0);
1755 p->rt_priority = 0;
1756 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1757 p->static_prio = NICE_TO_PRIO(0);
1758
1759 p->prio = p->normal_prio = __normal_prio(p);
1760 set_load_weight(p);
1761
1762 /*
1763 * We don't need the reset flag anymore after the fork. It has
1764 * fulfilled its duty:
1765 */
1766 p->sched_reset_on_fork = 0;
1767 }
1768
1769 if (!rt_prio(p->prio))
1770 p->sched_class = &fair_sched_class;
1771
1772 if (p->sched_class->task_fork)
1773 p->sched_class->task_fork(p);
1774
1775 /*
1776 * The child is not yet in the pid-hash so no cgroup attach races,
1777 * and the cgroup is pinned to this child due to cgroup_fork()
1778 * is ran before sched_fork().
1779 *
1780 * Silence PROVE_RCU.
1781 */
1782 raw_spin_lock_irqsave(&p->pi_lock, flags);
1783 set_task_cpu(p, cpu);
1784 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1785
1786#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1787 if (likely(sched_info_on()))
1788 memset(&p->sched_info, 0, sizeof(p->sched_info));
1789#endif
1790#if defined(CONFIG_SMP)
1791 p->on_cpu = 0;
1792#endif
1793#ifdef CONFIG_PREEMPT_COUNT
1794 /* Want to start with kernel preemption disabled. */
1795 task_thread_info(p)->preempt_count = 1;
1796#endif
1797#ifdef CONFIG_SMP
1798 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1799#endif
1800
1801 put_cpu();
1802}
1803
1804/*
1805 * wake_up_new_task - wake up a newly created task for the first time.
1806 *
1807 * This function will do some initial scheduler statistics housekeeping
1808 * that must be done for every newly created context, then puts the task
1809 * on the runqueue and wakes it.
1810 */
1811void wake_up_new_task(struct task_struct *p)
1812{
1813 unsigned long flags;
1814 struct rq *rq;
1815
1816 raw_spin_lock_irqsave(&p->pi_lock, flags);
1817#ifdef CONFIG_SMP
1818 /*
1819 * Fork balancing, do it here and not earlier because:
1820 * - cpus_allowed can change in the fork path
1821 * - any previously selected cpu might disappear through hotplug
1822 */
1823 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1824#endif
1825
1826 rq = __task_rq_lock(p);
1827 activate_task(rq, p, 0);
1828 p->on_rq = 1;
1829 trace_sched_wakeup_new(p, true);
1830 check_preempt_curr(rq, p, WF_FORK);
1831#ifdef CONFIG_SMP
1832 if (p->sched_class->task_woken)
1833 p->sched_class->task_woken(rq, p);
1834#endif
1835 task_rq_unlock(rq, p, &flags);
1836}
1837
1838#ifdef CONFIG_PREEMPT_NOTIFIERS
1839
1840/**
1841 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1842 * @notifier: notifier struct to register
1843 */
1844void preempt_notifier_register(struct preempt_notifier *notifier)
1845{
1846 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1847}
1848EXPORT_SYMBOL_GPL(preempt_notifier_register);
1849
1850/**
1851 * preempt_notifier_unregister - no longer interested in preemption notifications
1852 * @notifier: notifier struct to unregister
1853 *
1854 * This is safe to call from within a preemption notifier.
1855 */
1856void preempt_notifier_unregister(struct preempt_notifier *notifier)
1857{
1858 hlist_del(¬ifier->link);
1859}
1860EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1861
1862static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1863{
1864 struct preempt_notifier *notifier;
1865 struct hlist_node *node;
1866
1867 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1868 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1869}
1870
1871static void
1872fire_sched_out_preempt_notifiers(struct task_struct *curr,
1873 struct task_struct *next)
1874{
1875 struct preempt_notifier *notifier;
1876 struct hlist_node *node;
1877
1878 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1879 notifier->ops->sched_out(notifier, next);
1880}
1881
1882#else /* !CONFIG_PREEMPT_NOTIFIERS */
1883
1884static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1885{
1886}
1887
1888static void
1889fire_sched_out_preempt_notifiers(struct task_struct *curr,
1890 struct task_struct *next)
1891{
1892}
1893
1894#endif /* CONFIG_PREEMPT_NOTIFIERS */
1895
1896/**
1897 * prepare_task_switch - prepare to switch tasks
1898 * @rq: the runqueue preparing to switch
1899 * @prev: the current task that is being switched out
1900 * @next: the task we are going to switch to.
1901 *
1902 * This is called with the rq lock held and interrupts off. It must
1903 * be paired with a subsequent finish_task_switch after the context
1904 * switch.
1905 *
1906 * prepare_task_switch sets up locking and calls architecture specific
1907 * hooks.
1908 */
1909static inline void
1910prepare_task_switch(struct rq *rq, struct task_struct *prev,
1911 struct task_struct *next)
1912{
1913 sched_info_switch(prev, next);
1914 perf_event_task_sched_out(prev, next);
1915 fire_sched_out_preempt_notifiers(prev, next);
1916 prepare_lock_switch(rq, next);
1917 prepare_arch_switch(next);
1918 trace_sched_switch(prev, next);
1919}
1920
1921/**
1922 * finish_task_switch - clean up after a task-switch
1923 * @rq: runqueue associated with task-switch
1924 * @prev: the thread we just switched away from.
1925 *
1926 * finish_task_switch must be called after the context switch, paired
1927 * with a prepare_task_switch call before the context switch.
1928 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1929 * and do any other architecture-specific cleanup actions.
1930 *
1931 * Note that we may have delayed dropping an mm in context_switch(). If
1932 * so, we finish that here outside of the runqueue lock. (Doing it
1933 * with the lock held can cause deadlocks; see schedule() for
1934 * details.)
1935 */
1936static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1937 __releases(rq->lock)
1938{
1939 struct mm_struct *mm = rq->prev_mm;
1940 long prev_state;
1941
1942 rq->prev_mm = NULL;
1943
1944 /*
1945 * A task struct has one reference for the use as "current".
1946 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1947 * schedule one last time. The schedule call will never return, and
1948 * the scheduled task must drop that reference.
1949 * The test for TASK_DEAD must occur while the runqueue locks are
1950 * still held, otherwise prev could be scheduled on another cpu, die
1951 * there before we look at prev->state, and then the reference would
1952 * be dropped twice.
1953 * Manfred Spraul <manfred@colorfullife.com>
1954 */
1955 prev_state = prev->state;
1956 finish_arch_switch(prev);
1957#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1958 local_irq_disable();
1959#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1960 perf_event_task_sched_in(prev, current);
1961#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1962 local_irq_enable();
1963#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1964 finish_lock_switch(rq, prev);
1965 finish_arch_post_lock_switch();
1966
1967 fire_sched_in_preempt_notifiers(current);
1968 if (mm)
1969 mmdrop(mm);
1970 if (unlikely(prev_state == TASK_DEAD)) {
1971 /*
1972 * Remove function-return probe instances associated with this
1973 * task and put them back on the free list.
1974 */
1975 kprobe_flush_task(prev);
1976 put_task_struct(prev);
1977 }
1978}
1979
1980#ifdef CONFIG_SMP
1981
1982/* assumes rq->lock is held */
1983static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1984{
1985 if (prev->sched_class->pre_schedule)
1986 prev->sched_class->pre_schedule(rq, prev);
1987}
1988
1989/* rq->lock is NOT held, but preemption is disabled */
1990static inline void post_schedule(struct rq *rq)
1991{
1992 if (rq->post_schedule) {
1993 unsigned long flags;
1994
1995 raw_spin_lock_irqsave(&rq->lock, flags);
1996 if (rq->curr->sched_class->post_schedule)
1997 rq->curr->sched_class->post_schedule(rq);
1998 raw_spin_unlock_irqrestore(&rq->lock, flags);
1999
2000 rq->post_schedule = 0;
2001 }
2002}
2003
2004#else
2005
2006static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2007{
2008}
2009
2010static inline void post_schedule(struct rq *rq)
2011{
2012}
2013
2014#endif
2015
2016/**
2017 * schedule_tail - first thing a freshly forked thread must call.
2018 * @prev: the thread we just switched away from.
2019 */
2020asmlinkage void schedule_tail(struct task_struct *prev)
2021 __releases(rq->lock)
2022{
2023 struct rq *rq = this_rq();
2024
2025 finish_task_switch(rq, prev);
2026
2027 /*
2028 * FIXME: do we need to worry about rq being invalidated by the
2029 * task_switch?
2030 */
2031 post_schedule(rq);
2032
2033#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2034 /* In this case, finish_task_switch does not reenable preemption */
2035 preempt_enable();
2036#endif
2037 if (current->set_child_tid)
2038 put_user(task_pid_vnr(current), current->set_child_tid);
2039}
2040
2041/*
2042 * context_switch - switch to the new MM and the new
2043 * thread's register state.
2044 */
2045static inline void
2046context_switch(struct rq *rq, struct task_struct *prev,
2047 struct task_struct *next)
2048{
2049 struct mm_struct *mm, *oldmm;
2050
2051 prepare_task_switch(rq, prev, next);
2052
2053 mm = next->mm;
2054 oldmm = prev->active_mm;
2055 /*
2056 * For paravirt, this is coupled with an exit in switch_to to
2057 * combine the page table reload and the switch backend into
2058 * one hypercall.
2059 */
2060 arch_start_context_switch(prev);
2061
2062 if (!mm) {
2063 next->active_mm = oldmm;
2064 atomic_inc(&oldmm->mm_count);
2065 enter_lazy_tlb(oldmm, next);
2066 } else
2067 switch_mm(oldmm, mm, next);
2068
2069 if (!prev->mm) {
2070 prev->active_mm = NULL;
2071 rq->prev_mm = oldmm;
2072 }
2073 /*
2074 * Since the runqueue lock will be released by the next
2075 * task (which is an invalid locking op but in the case
2076 * of the scheduler it's an obvious special-case), so we
2077 * do an early lockdep release here:
2078 */
2079#ifndef __ARCH_WANT_UNLOCKED_CTXSW
2080 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2081#endif
2082
2083 /* Here we just switch the register state and the stack. */
2084 switch_to(prev, next, prev);
2085
2086 barrier();
2087 /*
2088 * this_rq must be evaluated again because prev may have moved
2089 * CPUs since it called schedule(), thus the 'rq' on its stack
2090 * frame will be invalid.
2091 */
2092 finish_task_switch(this_rq(), prev);
2093}
2094
2095/*
2096 * nr_running, nr_uninterruptible and nr_context_switches:
2097 *
2098 * externally visible scheduler statistics: current number of runnable
2099 * threads, current number of uninterruptible-sleeping threads, total
2100 * number of context switches performed since bootup.
2101 */
2102unsigned long nr_running(void)
2103{
2104 unsigned long i, sum = 0;
2105
2106 for_each_online_cpu(i)
2107 sum += cpu_rq(i)->nr_running;
2108
2109 return sum;
2110}
2111
2112unsigned long nr_uninterruptible(void)
2113{
2114 unsigned long i, sum = 0;
2115
2116 for_each_possible_cpu(i)
2117 sum += cpu_rq(i)->nr_uninterruptible;
2118
2119 /*
2120 * Since we read the counters lockless, it might be slightly
2121 * inaccurate. Do not allow it to go below zero though:
2122 */
2123 if (unlikely((long)sum < 0))
2124 sum = 0;
2125
2126 return sum;
2127}
2128
2129unsigned long long nr_context_switches(void)
2130{
2131 int i;
2132 unsigned long long sum = 0;
2133
2134 for_each_possible_cpu(i)
2135 sum += cpu_rq(i)->nr_switches;
2136
2137 return sum;
2138}
2139
2140unsigned long nr_iowait(void)
2141{
2142 unsigned long i, sum = 0;
2143
2144 for_each_possible_cpu(i)
2145 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2146
2147 return sum;
2148}
2149
2150unsigned long nr_iowait_cpu(int cpu)
2151{
2152 struct rq *this = cpu_rq(cpu);
2153 return atomic_read(&this->nr_iowait);
2154}
2155
2156unsigned long this_cpu_load(void)
2157{
2158 struct rq *this = this_rq();
2159 return this->cpu_load[0];
2160}
2161
2162
2163/*
2164 * Global load-average calculations
2165 *
2166 * We take a distributed and async approach to calculating the global load-avg
2167 * in order to minimize overhead.
2168 *
2169 * The global load average is an exponentially decaying average of nr_running +
2170 * nr_uninterruptible.
2171 *
2172 * Once every LOAD_FREQ:
2173 *
2174 * nr_active = 0;
2175 * for_each_possible_cpu(cpu)
2176 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2177 *
2178 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2179 *
2180 * Due to a number of reasons the above turns in the mess below:
2181 *
2182 * - for_each_possible_cpu() is prohibitively expensive on machines with
2183 * serious number of cpus, therefore we need to take a distributed approach
2184 * to calculating nr_active.
2185 *
2186 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2187 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2188 *
2189 * So assuming nr_active := 0 when we start out -- true per definition, we
2190 * can simply take per-cpu deltas and fold those into a global accumulate
2191 * to obtain the same result. See calc_load_fold_active().
2192 *
2193 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2194 * across the machine, we assume 10 ticks is sufficient time for every
2195 * cpu to have completed this task.
2196 *
2197 * This places an upper-bound on the IRQ-off latency of the machine. Then
2198 * again, being late doesn't loose the delta, just wrecks the sample.
2199 *
2200 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2201 * this would add another cross-cpu cacheline miss and atomic operation
2202 * to the wakeup path. Instead we increment on whatever cpu the task ran
2203 * when it went into uninterruptible state and decrement on whatever cpu
2204 * did the wakeup. This means that only the sum of nr_uninterruptible over
2205 * all cpus yields the correct result.
2206 *
2207 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2208 */
2209
2210/* Variables and functions for calc_load */
2211static atomic_long_t calc_load_tasks;
2212static unsigned long calc_load_update;
2213unsigned long avenrun[3];
2214EXPORT_SYMBOL(avenrun); /* should be removed */
2215
2216/**
2217 * get_avenrun - get the load average array
2218 * @loads: pointer to dest load array
2219 * @offset: offset to add
2220 * @shift: shift count to shift the result left
2221 *
2222 * These values are estimates at best, so no need for locking.
2223 */
2224void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2225{
2226 loads[0] = (avenrun[0] + offset) << shift;
2227 loads[1] = (avenrun[1] + offset) << shift;
2228 loads[2] = (avenrun[2] + offset) << shift;
2229}
2230
2231static long calc_load_fold_active(struct rq *this_rq)
2232{
2233 long nr_active, delta = 0;
2234
2235 nr_active = this_rq->nr_running;
2236 nr_active += (long) this_rq->nr_uninterruptible;
2237
2238 if (nr_active != this_rq->calc_load_active) {
2239 delta = nr_active - this_rq->calc_load_active;
2240 this_rq->calc_load_active = nr_active;
2241 }
2242
2243 return delta;
2244}
2245
2246/*
2247 * a1 = a0 * e + a * (1 - e)
2248 */
2249static unsigned long
2250calc_load(unsigned long load, unsigned long exp, unsigned long active)
2251{
2252 load *= exp;
2253 load += active * (FIXED_1 - exp);
2254 load += 1UL << (FSHIFT - 1);
2255 return load >> FSHIFT;
2256}
2257
2258#ifdef CONFIG_NO_HZ
2259/*
2260 * Handle NO_HZ for the global load-average.
2261 *
2262 * Since the above described distributed algorithm to compute the global
2263 * load-average relies on per-cpu sampling from the tick, it is affected by
2264 * NO_HZ.
2265 *
2266 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2267 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2268 * when we read the global state.
2269 *
2270 * Obviously reality has to ruin such a delightfully simple scheme:
2271 *
2272 * - When we go NO_HZ idle during the window, we can negate our sample
2273 * contribution, causing under-accounting.
2274 *
2275 * We avoid this by keeping two idle-delta counters and flipping them
2276 * when the window starts, thus separating old and new NO_HZ load.
2277 *
2278 * The only trick is the slight shift in index flip for read vs write.
2279 *
2280 * 0s 5s 10s 15s
2281 * +10 +10 +10 +10
2282 * |-|-----------|-|-----------|-|-----------|-|
2283 * r:0 0 1 1 0 0 1 1 0
2284 * w:0 1 1 0 0 1 1 0 0
2285 *
2286 * This ensures we'll fold the old idle contribution in this window while
2287 * accumlating the new one.
2288 *
2289 * - When we wake up from NO_HZ idle during the window, we push up our
2290 * contribution, since we effectively move our sample point to a known
2291 * busy state.
2292 *
2293 * This is solved by pushing the window forward, and thus skipping the
2294 * sample, for this cpu (effectively using the idle-delta for this cpu which
2295 * was in effect at the time the window opened). This also solves the issue
2296 * of having to deal with a cpu having been in NOHZ idle for multiple
2297 * LOAD_FREQ intervals.
2298 *
2299 * When making the ILB scale, we should try to pull this in as well.
2300 */
2301static atomic_long_t calc_load_idle[2];
2302static int calc_load_idx;
2303
2304static inline int calc_load_write_idx(void)
2305{
2306 int idx = calc_load_idx;
2307
2308 /*
2309 * See calc_global_nohz(), if we observe the new index, we also
2310 * need to observe the new update time.
2311 */
2312 smp_rmb();
2313
2314 /*
2315 * If the folding window started, make sure we start writing in the
2316 * next idle-delta.
2317 */
2318 if (!time_before(jiffies, calc_load_update))
2319 idx++;
2320
2321 return idx & 1;
2322}
2323
2324static inline int calc_load_read_idx(void)
2325{
2326 return calc_load_idx & 1;
2327}
2328
2329void calc_load_enter_idle(void)
2330{
2331 struct rq *this_rq = this_rq();
2332 long delta;
2333
2334 /*
2335 * We're going into NOHZ mode, if there's any pending delta, fold it
2336 * into the pending idle delta.
2337 */
2338 delta = calc_load_fold_active(this_rq);
2339 if (delta) {
2340 int idx = calc_load_write_idx();
2341 atomic_long_add(delta, &calc_load_idle[idx]);
2342 }
2343}
2344
2345void calc_load_exit_idle(void)
2346{
2347 struct rq *this_rq = this_rq();
2348
2349 /*
2350 * If we're still before the sample window, we're done.
2351 */
2352 if (time_before(jiffies, this_rq->calc_load_update))
2353 return;
2354
2355 /*
2356 * We woke inside or after the sample window, this means we're already
2357 * accounted through the nohz accounting, so skip the entire deal and
2358 * sync up for the next window.
2359 */
2360 this_rq->calc_load_update = calc_load_update;
2361 if (time_before(jiffies, this_rq->calc_load_update + 10))
2362 this_rq->calc_load_update += LOAD_FREQ;
2363}
2364
2365static long calc_load_fold_idle(void)
2366{
2367 int idx = calc_load_read_idx();
2368 long delta = 0;
2369
2370 if (atomic_long_read(&calc_load_idle[idx]))
2371 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2372
2373 return delta;
2374}
2375
2376/**
2377 * fixed_power_int - compute: x^n, in O(log n) time
2378 *
2379 * @x: base of the power
2380 * @frac_bits: fractional bits of @x
2381 * @n: power to raise @x to.
2382 *
2383 * By exploiting the relation between the definition of the natural power
2384 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2385 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2386 * (where: n_i \elem {0, 1}, the binary vector representing n),
2387 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2388 * of course trivially computable in O(log_2 n), the length of our binary
2389 * vector.
2390 */
2391static unsigned long
2392fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2393{
2394 unsigned long result = 1UL << frac_bits;
2395
2396 if (n) for (;;) {
2397 if (n & 1) {
2398 result *= x;
2399 result += 1UL << (frac_bits - 1);
2400 result >>= frac_bits;
2401 }
2402 n >>= 1;
2403 if (!n)
2404 break;
2405 x *= x;
2406 x += 1UL << (frac_bits - 1);
2407 x >>= frac_bits;
2408 }
2409
2410 return result;
2411}
2412
2413/*
2414 * a1 = a0 * e + a * (1 - e)
2415 *
2416 * a2 = a1 * e + a * (1 - e)
2417 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2418 * = a0 * e^2 + a * (1 - e) * (1 + e)
2419 *
2420 * a3 = a2 * e + a * (1 - e)
2421 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2422 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2423 *
2424 * ...
2425 *
2426 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2427 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2428 * = a0 * e^n + a * (1 - e^n)
2429 *
2430 * [1] application of the geometric series:
2431 *
2432 * n 1 - x^(n+1)
2433 * S_n := \Sum x^i = -------------
2434 * i=0 1 - x
2435 */
2436static unsigned long
2437calc_load_n(unsigned long load, unsigned long exp,
2438 unsigned long active, unsigned int n)
2439{
2440
2441 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2442}
2443
2444/*
2445 * NO_HZ can leave us missing all per-cpu ticks calling
2446 * calc_load_account_active(), but since an idle CPU folds its delta into
2447 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2448 * in the pending idle delta if our idle period crossed a load cycle boundary.
2449 *
2450 * Once we've updated the global active value, we need to apply the exponential
2451 * weights adjusted to the number of cycles missed.
2452 */
2453static void calc_global_nohz(void)
2454{
2455 long delta, active, n;
2456
2457 if (!time_before(jiffies, calc_load_update + 10)) {
2458 /*
2459 * Catch-up, fold however many we are behind still
2460 */
2461 delta = jiffies - calc_load_update - 10;
2462 n = 1 + (delta / LOAD_FREQ);
2463
2464 active = atomic_long_read(&calc_load_tasks);
2465 active = active > 0 ? active * FIXED_1 : 0;
2466
2467 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2468 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2469 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2470
2471 calc_load_update += n * LOAD_FREQ;
2472 }
2473
2474 /*
2475 * Flip the idle index...
2476 *
2477 * Make sure we first write the new time then flip the index, so that
2478 * calc_load_write_idx() will see the new time when it reads the new
2479 * index, this avoids a double flip messing things up.
2480 */
2481 smp_wmb();
2482 calc_load_idx++;
2483}
2484#else /* !CONFIG_NO_HZ */
2485
2486static inline long calc_load_fold_idle(void) { return 0; }
2487static inline void calc_global_nohz(void) { }
2488
2489#endif /* CONFIG_NO_HZ */
2490
2491/*
2492 * calc_load - update the avenrun load estimates 10 ticks after the
2493 * CPUs have updated calc_load_tasks.
2494 */
2495void calc_global_load(unsigned long ticks)
2496{
2497 long active, delta;
2498
2499 if (time_before(jiffies, calc_load_update + 10))
2500 return;
2501
2502 /*
2503 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2504 */
2505 delta = calc_load_fold_idle();
2506 if (delta)
2507 atomic_long_add(delta, &calc_load_tasks);
2508
2509 active = atomic_long_read(&calc_load_tasks);
2510 active = active > 0 ? active * FIXED_1 : 0;
2511
2512 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2513 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2514 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2515
2516 calc_load_update += LOAD_FREQ;
2517
2518 /*
2519 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2520 */
2521 calc_global_nohz();
2522}
2523
2524/*
2525 * Called from update_cpu_load() to periodically update this CPU's
2526 * active count.
2527 */
2528static void calc_load_account_active(struct rq *this_rq)
2529{
2530 long delta;
2531
2532 if (time_before(jiffies, this_rq->calc_load_update))
2533 return;
2534
2535 delta = calc_load_fold_active(this_rq);
2536 if (delta)
2537 atomic_long_add(delta, &calc_load_tasks);
2538
2539 this_rq->calc_load_update += LOAD_FREQ;
2540}
2541
2542/*
2543 * End of global load-average stuff
2544 */
2545
2546/*
2547 * The exact cpuload at various idx values, calculated at every tick would be
2548 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2549 *
2550 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2551 * on nth tick when cpu may be busy, then we have:
2552 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2553 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2554 *
2555 * decay_load_missed() below does efficient calculation of
2556 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2557 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2558 *
2559 * The calculation is approximated on a 128 point scale.
2560 * degrade_zero_ticks is the number of ticks after which load at any
2561 * particular idx is approximated to be zero.
2562 * degrade_factor is a precomputed table, a row for each load idx.
2563 * Each column corresponds to degradation factor for a power of two ticks,
2564 * based on 128 point scale.
2565 * Example:
2566 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2567 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2568 *
2569 * With this power of 2 load factors, we can degrade the load n times
2570 * by looking at 1 bits in n and doing as many mult/shift instead of
2571 * n mult/shifts needed by the exact degradation.
2572 */
2573#define DEGRADE_SHIFT 7
2574static const unsigned char
2575 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2576static const unsigned char
2577 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2578 {0, 0, 0, 0, 0, 0, 0, 0},
2579 {64, 32, 8, 0, 0, 0, 0, 0},
2580 {96, 72, 40, 12, 1, 0, 0},
2581 {112, 98, 75, 43, 15, 1, 0},
2582 {120, 112, 98, 76, 45, 16, 2} };
2583
2584/*
2585 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2586 * would be when CPU is idle and so we just decay the old load without
2587 * adding any new load.
2588 */
2589static unsigned long
2590decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2591{
2592 int j = 0;
2593
2594 if (!missed_updates)
2595 return load;
2596
2597 if (missed_updates >= degrade_zero_ticks[idx])
2598 return 0;
2599
2600 if (idx == 1)
2601 return load >> missed_updates;
2602
2603 while (missed_updates) {
2604 if (missed_updates % 2)
2605 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2606
2607 missed_updates >>= 1;
2608 j++;
2609 }
2610 return load;
2611}
2612
2613/*
2614 * Update rq->cpu_load[] statistics. This function is usually called every
2615 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2616 * every tick. We fix it up based on jiffies.
2617 */
2618static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2619 unsigned long pending_updates)
2620{
2621 int i, scale;
2622
2623 this_rq->nr_load_updates++;
2624
2625 /* Update our load: */
2626 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2627 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2628 unsigned long old_load, new_load;
2629
2630 /* scale is effectively 1 << i now, and >> i divides by scale */
2631
2632 old_load = this_rq->cpu_load[i];
2633 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2634 new_load = this_load;
2635 /*
2636 * Round up the averaging division if load is increasing. This
2637 * prevents us from getting stuck on 9 if the load is 10, for
2638 * example.
2639 */
2640 if (new_load > old_load)
2641 new_load += scale - 1;
2642
2643 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2644 }
2645
2646 sched_avg_update(this_rq);
2647}
2648
2649#ifdef CONFIG_NO_HZ
2650/*
2651 * There is no sane way to deal with nohz on smp when using jiffies because the
2652 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2653 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2654 *
2655 * Therefore we cannot use the delta approach from the regular tick since that
2656 * would seriously skew the load calculation. However we'll make do for those
2657 * updates happening while idle (nohz_idle_balance) or coming out of idle
2658 * (tick_nohz_idle_exit).
2659 *
2660 * This means we might still be one tick off for nohz periods.
2661 */
2662
2663/*
2664 * Called from nohz_idle_balance() to update the load ratings before doing the
2665 * idle balance.
2666 */
2667void update_idle_cpu_load(struct rq *this_rq)
2668{
2669 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2670 unsigned long load = this_rq->load.weight;
2671 unsigned long pending_updates;
2672
2673 /*
2674 * bail if there's load or we're actually up-to-date.
2675 */
2676 if (load || curr_jiffies == this_rq->last_load_update_tick)
2677 return;
2678
2679 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2680 this_rq->last_load_update_tick = curr_jiffies;
2681
2682 __update_cpu_load(this_rq, load, pending_updates);
2683}
2684
2685/*
2686 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2687 */
2688void update_cpu_load_nohz(void)
2689{
2690 struct rq *this_rq = this_rq();
2691 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2692 unsigned long pending_updates;
2693
2694 if (curr_jiffies == this_rq->last_load_update_tick)
2695 return;
2696
2697 raw_spin_lock(&this_rq->lock);
2698 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2699 if (pending_updates) {
2700 this_rq->last_load_update_tick = curr_jiffies;
2701 /*
2702 * We were idle, this means load 0, the current load might be
2703 * !0 due to remote wakeups and the sort.
2704 */
2705 __update_cpu_load(this_rq, 0, pending_updates);
2706 }
2707 raw_spin_unlock(&this_rq->lock);
2708}
2709#endif /* CONFIG_NO_HZ */
2710
2711/*
2712 * Called from scheduler_tick()
2713 */
2714static void update_cpu_load_active(struct rq *this_rq)
2715{
2716 /*
2717 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2718 */
2719 this_rq->last_load_update_tick = jiffies;
2720 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2721
2722 calc_load_account_active(this_rq);
2723}
2724
2725#ifdef CONFIG_SMP
2726
2727/*
2728 * sched_exec - execve() is a valuable balancing opportunity, because at
2729 * this point the task has the smallest effective memory and cache footprint.
2730 */
2731void sched_exec(void)
2732{
2733 struct task_struct *p = current;
2734 unsigned long flags;
2735 int dest_cpu;
2736
2737 raw_spin_lock_irqsave(&p->pi_lock, flags);
2738 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2739 if (dest_cpu == smp_processor_id())
2740 goto unlock;
2741
2742 if (likely(cpu_active(dest_cpu))) {
2743 struct migration_arg arg = { p, dest_cpu };
2744
2745 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2746 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2747 return;
2748 }
2749unlock:
2750 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2751}
2752
2753#endif
2754
2755DEFINE_PER_CPU(struct kernel_stat, kstat);
2756DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2757
2758EXPORT_PER_CPU_SYMBOL(kstat);
2759EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2760
2761/*
2762 * Return any ns on the sched_clock that have not yet been accounted in
2763 * @p in case that task is currently running.
2764 *
2765 * Called with task_rq_lock() held on @rq.
2766 */
2767static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2768{
2769 u64 ns = 0;
2770
2771 if (task_current(rq, p)) {
2772 update_rq_clock(rq);
2773 ns = rq->clock_task - p->se.exec_start;
2774 if ((s64)ns < 0)
2775 ns = 0;
2776 }
2777
2778 return ns;
2779}
2780
2781unsigned long long task_delta_exec(struct task_struct *p)
2782{
2783 unsigned long flags;
2784 struct rq *rq;
2785 u64 ns = 0;
2786
2787 rq = task_rq_lock(p, &flags);
2788 ns = do_task_delta_exec(p, rq);
2789 task_rq_unlock(rq, p, &flags);
2790
2791 return ns;
2792}
2793
2794/*
2795 * Return accounted runtime for the task.
2796 * In case the task is currently running, return the runtime plus current's
2797 * pending runtime that have not been accounted yet.
2798 */
2799unsigned long long task_sched_runtime(struct task_struct *p)
2800{
2801 unsigned long flags;
2802 struct rq *rq;
2803 u64 ns = 0;
2804
2805 rq = task_rq_lock(p, &flags);
2806 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2807 task_rq_unlock(rq, p, &flags);
2808
2809 return ns;
2810}
2811
2812#ifdef CONFIG_CGROUP_CPUACCT
2813struct cgroup_subsys cpuacct_subsys;
2814struct cpuacct root_cpuacct;
2815#endif
2816
2817static inline void task_group_account_field(struct task_struct *p, int index,
2818 u64 tmp)
2819{
2820#ifdef CONFIG_CGROUP_CPUACCT
2821 struct kernel_cpustat *kcpustat;
2822 struct cpuacct *ca;
2823#endif
2824 /*
2825 * Since all updates are sure to touch the root cgroup, we
2826 * get ourselves ahead and touch it first. If the root cgroup
2827 * is the only cgroup, then nothing else should be necessary.
2828 *
2829 */
2830 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
2831
2832#ifdef CONFIG_CGROUP_CPUACCT
2833 if (unlikely(!cpuacct_subsys.active))
2834 return;
2835
2836 rcu_read_lock();
2837 ca = task_ca(p);
2838 while (ca && (ca != &root_cpuacct)) {
2839 kcpustat = this_cpu_ptr(ca->cpustat);
2840 kcpustat->cpustat[index] += tmp;
2841 ca = parent_ca(ca);
2842 }
2843 rcu_read_unlock();
2844#endif
2845}
2846
2847
2848/*
2849 * Account user cpu time to a process.
2850 * @p: the process that the cpu time gets accounted to
2851 * @cputime: the cpu time spent in user space since the last update
2852 * @cputime_scaled: cputime scaled by cpu frequency
2853 */
2854void account_user_time(struct task_struct *p, cputime_t cputime,
2855 cputime_t cputime_scaled)
2856{
2857 int index;
2858
2859 /* Add user time to process. */
2860 p->utime += cputime;
2861 p->utimescaled += cputime_scaled;
2862 account_group_user_time(p, cputime);
2863
2864 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
2865
2866 /* Add user time to cpustat. */
2867 task_group_account_field(p, index, (__force u64) cputime);
2868
2869 /* Account for user time used */
2870 acct_update_integrals(p);
2871}
2872
2873/*
2874 * Account guest cpu time to a process.
2875 * @p: the process that the cpu time gets accounted to
2876 * @cputime: the cpu time spent in virtual machine since the last update
2877 * @cputime_scaled: cputime scaled by cpu frequency
2878 */
2879static void account_guest_time(struct task_struct *p, cputime_t cputime,
2880 cputime_t cputime_scaled)
2881{
2882 u64 *cpustat = kcpustat_this_cpu->cpustat;
2883
2884 /* Add guest time to process. */
2885 p->utime += cputime;
2886 p->utimescaled += cputime_scaled;
2887 account_group_user_time(p, cputime);
2888 p->gtime += cputime;
2889
2890 /* Add guest time to cpustat. */
2891 if (TASK_NICE(p) > 0) {
2892 cpustat[CPUTIME_NICE] += (__force u64) cputime;
2893 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
2894 } else {
2895 cpustat[CPUTIME_USER] += (__force u64) cputime;
2896 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
2897 }
2898}
2899
2900/*
2901 * Account system cpu time to a process and desired cpustat field
2902 * @p: the process that the cpu time gets accounted to
2903 * @cputime: the cpu time spent in kernel space since the last update
2904 * @cputime_scaled: cputime scaled by cpu frequency
2905 * @target_cputime64: pointer to cpustat field that has to be updated
2906 */
2907static inline
2908void __account_system_time(struct task_struct *p, cputime_t cputime,
2909 cputime_t cputime_scaled, int index)
2910{
2911 /* Add system time to process. */
2912 p->stime += cputime;
2913 p->stimescaled += cputime_scaled;
2914 account_group_system_time(p, cputime);
2915
2916 /* Add system time to cpustat. */
2917 task_group_account_field(p, index, (__force u64) cputime);
2918
2919 /* Account for system time used */
2920 acct_update_integrals(p);
2921}
2922
2923/*
2924 * Account system cpu time to a process.
2925 * @p: the process that the cpu time gets accounted to
2926 * @hardirq_offset: the offset to subtract from hardirq_count()
2927 * @cputime: the cpu time spent in kernel space since the last update
2928 * @cputime_scaled: cputime scaled by cpu frequency
2929 */
2930void account_system_time(struct task_struct *p, int hardirq_offset,
2931 cputime_t cputime, cputime_t cputime_scaled)
2932{
2933 int index;
2934
2935 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2936 account_guest_time(p, cputime, cputime_scaled);
2937 return;
2938 }
2939
2940 if (hardirq_count() - hardirq_offset)
2941 index = CPUTIME_IRQ;
2942 else if (in_serving_softirq())
2943 index = CPUTIME_SOFTIRQ;
2944 else
2945 index = CPUTIME_SYSTEM;
2946
2947 __account_system_time(p, cputime, cputime_scaled, index);
2948}
2949
2950/*
2951 * Account for involuntary wait time.
2952 * @cputime: the cpu time spent in involuntary wait
2953 */
2954void account_steal_time(cputime_t cputime)
2955{
2956 u64 *cpustat = kcpustat_this_cpu->cpustat;
2957
2958 cpustat[CPUTIME_STEAL] += (__force u64) cputime;
2959}
2960
2961/*
2962 * Account for idle time.
2963 * @cputime: the cpu time spent in idle wait
2964 */
2965void account_idle_time(cputime_t cputime)
2966{
2967 u64 *cpustat = kcpustat_this_cpu->cpustat;
2968 struct rq *rq = this_rq();
2969
2970 if (atomic_read(&rq->nr_iowait) > 0)
2971 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
2972 else
2973 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
2974}
2975
2976static __always_inline bool steal_account_process_tick(void)
2977{
2978#ifdef CONFIG_PARAVIRT
2979 if (static_key_false(¶virt_steal_enabled)) {
2980 u64 steal, st = 0;
2981
2982 steal = paravirt_steal_clock(smp_processor_id());
2983 steal -= this_rq()->prev_steal_time;
2984
2985 st = steal_ticks(steal);
2986 this_rq()->prev_steal_time += st * TICK_NSEC;
2987
2988 account_steal_time(st);
2989 return st;
2990 }
2991#endif
2992 return false;
2993}
2994
2995#ifndef CONFIG_VIRT_CPU_ACCOUNTING
2996
2997#ifdef CONFIG_IRQ_TIME_ACCOUNTING
2998/*
2999 * Account a tick to a process and cpustat
3000 * @p: the process that the cpu time gets accounted to
3001 * @user_tick: is the tick from userspace
3002 * @rq: the pointer to rq
3003 *
3004 * Tick demultiplexing follows the order
3005 * - pending hardirq update
3006 * - pending softirq update
3007 * - user_time
3008 * - idle_time
3009 * - system time
3010 * - check for guest_time
3011 * - else account as system_time
3012 *
3013 * Check for hardirq is done both for system and user time as there is
3014 * no timer going off while we are on hardirq and hence we may never get an
3015 * opportunity to update it solely in system time.
3016 * p->stime and friends are only updated on system time and not on irq
3017 * softirq as those do not count in task exec_runtime any more.
3018 */
3019static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3020 struct rq *rq)
3021{
3022 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3023 u64 *cpustat = kcpustat_this_cpu->cpustat;
3024
3025 if (steal_account_process_tick())
3026 return;
3027
3028 if (irqtime_account_hi_update()) {
3029 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
3030 } else if (irqtime_account_si_update()) {
3031 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
3032 } else if (this_cpu_ksoftirqd() == p) {
3033 /*
3034 * ksoftirqd time do not get accounted in cpu_softirq_time.
3035 * So, we have to handle it separately here.
3036 * Also, p->stime needs to be updated for ksoftirqd.
3037 */
3038 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3039 CPUTIME_SOFTIRQ);
3040 } else if (user_tick) {
3041 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3042 } else if (p == rq->idle) {
3043 account_idle_time(cputime_one_jiffy);
3044 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3045 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3046 } else {
3047 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3048 CPUTIME_SYSTEM);
3049 }
3050}
3051
3052static void irqtime_account_idle_ticks(int ticks)
3053{
3054 int i;
3055 struct rq *rq = this_rq();
3056
3057 for (i = 0; i < ticks; i++)
3058 irqtime_account_process_tick(current, 0, rq);
3059}
3060#else /* CONFIG_IRQ_TIME_ACCOUNTING */
3061static void irqtime_account_idle_ticks(int ticks) {}
3062static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3063 struct rq *rq) {}
3064#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3065
3066/*
3067 * Account a single tick of cpu time.
3068 * @p: the process that the cpu time gets accounted to
3069 * @user_tick: indicates if the tick is a user or a system tick
3070 */
3071void account_process_tick(struct task_struct *p, int user_tick)
3072{
3073 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3074 struct rq *rq = this_rq();
3075
3076 if (sched_clock_irqtime) {
3077 irqtime_account_process_tick(p, user_tick, rq);
3078 return;
3079 }
3080
3081 if (steal_account_process_tick())
3082 return;
3083
3084 if (user_tick)
3085 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3086 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3087 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3088 one_jiffy_scaled);
3089 else
3090 account_idle_time(cputime_one_jiffy);
3091}
3092
3093/*
3094 * Account multiple ticks of steal time.
3095 * @p: the process from which the cpu time has been stolen
3096 * @ticks: number of stolen ticks
3097 */
3098void account_steal_ticks(unsigned long ticks)
3099{
3100 account_steal_time(jiffies_to_cputime(ticks));
3101}
3102
3103/*
3104 * Account multiple ticks of idle time.
3105 * @ticks: number of stolen ticks
3106 */
3107void account_idle_ticks(unsigned long ticks)
3108{
3109
3110 if (sched_clock_irqtime) {
3111 irqtime_account_idle_ticks(ticks);
3112 return;
3113 }
3114
3115 account_idle_time(jiffies_to_cputime(ticks));
3116}
3117
3118#endif
3119
3120/*
3121 * Use precise platform statistics if available:
3122 */
3123#ifdef CONFIG_VIRT_CPU_ACCOUNTING
3124void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3125{
3126 *ut = p->utime;
3127 *st = p->stime;
3128}
3129
3130void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3131{
3132 struct task_cputime cputime;
3133
3134 thread_group_cputime(p, &cputime);
3135
3136 *ut = cputime.utime;
3137 *st = cputime.stime;
3138}
3139#else
3140
3141#ifndef nsecs_to_cputime
3142# define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3143#endif
3144
3145static cputime_t scale_utime(cputime_t utime, cputime_t rtime, cputime_t total)
3146{
3147 u64 temp = (__force u64) rtime;
3148
3149 temp *= (__force u64) utime;
3150
3151 if (sizeof(cputime_t) == 4)
3152 temp = div_u64(temp, (__force u32) total);
3153 else
3154 temp = div64_u64(temp, (__force u64) total);
3155
3156 return (__force cputime_t) temp;
3157}
3158
3159void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3160{
3161 cputime_t rtime, utime = p->utime, total = utime + p->stime;
3162
3163 /*
3164 * Use CFS's precise accounting:
3165 */
3166 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3167
3168 if (total)
3169 utime = scale_utime(utime, rtime, total);
3170 else
3171 utime = rtime;
3172
3173 /*
3174 * Compare with previous values, to keep monotonicity:
3175 */
3176 p->prev_utime = max(p->prev_utime, utime);
3177 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime);
3178
3179 *ut = p->prev_utime;
3180 *st = p->prev_stime;
3181}
3182
3183/*
3184 * Must be called with siglock held.
3185 */
3186void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3187{
3188 struct signal_struct *sig = p->signal;
3189 struct task_cputime cputime;
3190 cputime_t rtime, utime, total;
3191
3192 thread_group_cputime(p, &cputime);
3193
3194 total = cputime.utime + cputime.stime;
3195 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3196
3197 if (total)
3198 utime = scale_utime(cputime.utime, rtime, total);
3199 else
3200 utime = rtime;
3201
3202 sig->prev_utime = max(sig->prev_utime, utime);
3203 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime);
3204
3205 *ut = sig->prev_utime;
3206 *st = sig->prev_stime;
3207}
3208#endif
3209
3210/*
3211 * This function gets called by the timer code, with HZ frequency.
3212 * We call it with interrupts disabled.
3213 */
3214void scheduler_tick(void)
3215{
3216 int cpu = smp_processor_id();
3217 struct rq *rq = cpu_rq(cpu);
3218 struct task_struct *curr = rq->curr;
3219
3220 sched_clock_tick();
3221
3222 raw_spin_lock(&rq->lock);
3223 update_rq_clock(rq);
3224 update_cpu_load_active(rq);
3225 curr->sched_class->task_tick(rq, curr, 0);
3226 raw_spin_unlock(&rq->lock);
3227
3228 perf_event_task_tick();
3229
3230#ifdef CONFIG_SMP
3231 rq->idle_balance = idle_cpu(cpu);
3232 trigger_load_balance(rq, cpu);
3233#endif
3234}
3235
3236notrace unsigned long get_parent_ip(unsigned long addr)
3237{
3238 if (in_lock_functions(addr)) {
3239 addr = CALLER_ADDR2;
3240 if (in_lock_functions(addr))
3241 addr = CALLER_ADDR3;
3242 }
3243 return addr;
3244}
3245
3246#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3247 defined(CONFIG_PREEMPT_TRACER))
3248
3249void __kprobes add_preempt_count(int val)
3250{
3251#ifdef CONFIG_DEBUG_PREEMPT
3252 /*
3253 * Underflow?
3254 */
3255 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3256 return;
3257#endif
3258 preempt_count() += val;
3259#ifdef CONFIG_DEBUG_PREEMPT
3260 /*
3261 * Spinlock count overflowing soon?
3262 */
3263 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3264 PREEMPT_MASK - 10);
3265#endif
3266 if (preempt_count() == val)
3267 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3268}
3269EXPORT_SYMBOL(add_preempt_count);
3270
3271void __kprobes sub_preempt_count(int val)
3272{
3273#ifdef CONFIG_DEBUG_PREEMPT
3274 /*
3275 * Underflow?
3276 */
3277 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3278 return;
3279 /*
3280 * Is the spinlock portion underflowing?
3281 */
3282 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3283 !(preempt_count() & PREEMPT_MASK)))
3284 return;
3285#endif
3286
3287 if (preempt_count() == val)
3288 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3289 preempt_count() -= val;
3290}
3291EXPORT_SYMBOL(sub_preempt_count);
3292
3293#endif
3294
3295/*
3296 * Print scheduling while atomic bug:
3297 */
3298static noinline void __schedule_bug(struct task_struct *prev)
3299{
3300 if (oops_in_progress)
3301 return;
3302
3303 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3304 prev->comm, prev->pid, preempt_count());
3305
3306 debug_show_held_locks(prev);
3307 print_modules();
3308 if (irqs_disabled())
3309 print_irqtrace_events(prev);
3310 dump_stack();
3311 add_taint(TAINT_WARN);
3312}
3313
3314/*
3315 * Various schedule()-time debugging checks and statistics:
3316 */
3317static inline void schedule_debug(struct task_struct *prev)
3318{
3319 /*
3320 * Test if we are atomic. Since do_exit() needs to call into
3321 * schedule() atomically, we ignore that path for now.
3322 * Otherwise, whine if we are scheduling when we should not be.
3323 */
3324 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3325 __schedule_bug(prev);
3326 rcu_sleep_check();
3327
3328 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3329
3330 schedstat_inc(this_rq(), sched_count);
3331}
3332
3333static void put_prev_task(struct rq *rq, struct task_struct *prev)
3334{
3335 if (prev->on_rq || rq->skip_clock_update < 0)
3336 update_rq_clock(rq);
3337 prev->sched_class->put_prev_task(rq, prev);
3338}
3339
3340/*
3341 * Pick up the highest-prio task:
3342 */
3343static inline struct task_struct *
3344pick_next_task(struct rq *rq)
3345{
3346 const struct sched_class *class;
3347 struct task_struct *p;
3348
3349 /*
3350 * Optimization: we know that if all tasks are in
3351 * the fair class we can call that function directly:
3352 */
3353 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3354 p = fair_sched_class.pick_next_task(rq);
3355 if (likely(p))
3356 return p;
3357 }
3358
3359 for_each_class(class) {
3360 p = class->pick_next_task(rq);
3361 if (p)
3362 return p;
3363 }
3364
3365 BUG(); /* the idle class will always have a runnable task */
3366}
3367
3368/*
3369 * __schedule() is the main scheduler function.
3370 */
3371static void __sched __schedule(void)
3372{
3373 struct task_struct *prev, *next;
3374 unsigned long *switch_count;
3375 struct rq *rq;
3376 int cpu;
3377
3378need_resched:
3379 preempt_disable();
3380 cpu = smp_processor_id();
3381 rq = cpu_rq(cpu);
3382 rcu_note_context_switch(cpu);
3383 prev = rq->curr;
3384
3385 schedule_debug(prev);
3386
3387 if (sched_feat(HRTICK))
3388 hrtick_clear(rq);
3389
3390 raw_spin_lock_irq(&rq->lock);
3391
3392 switch_count = &prev->nivcsw;
3393 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3394 if (unlikely(signal_pending_state(prev->state, prev))) {
3395 prev->state = TASK_RUNNING;
3396 } else {
3397 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3398 prev->on_rq = 0;
3399
3400 /*
3401 * If a worker went to sleep, notify and ask workqueue
3402 * whether it wants to wake up a task to maintain
3403 * concurrency.
3404 */
3405 if (prev->flags & PF_WQ_WORKER) {
3406 struct task_struct *to_wakeup;
3407
3408 to_wakeup = wq_worker_sleeping(prev, cpu);
3409 if (to_wakeup)
3410 try_to_wake_up_local(to_wakeup);
3411 }
3412 }
3413 switch_count = &prev->nvcsw;
3414 }
3415
3416 pre_schedule(rq, prev);
3417
3418 if (unlikely(!rq->nr_running))
3419 idle_balance(cpu, rq);
3420
3421 put_prev_task(rq, prev);
3422 next = pick_next_task(rq);
3423 clear_tsk_need_resched(prev);
3424 rq->skip_clock_update = 0;
3425
3426 if (likely(prev != next)) {
3427 rq->nr_switches++;
3428 rq->curr = next;
3429 ++*switch_count;
3430
3431 context_switch(rq, prev, next); /* unlocks the rq */
3432 /*
3433 * The context switch have flipped the stack from under us
3434 * and restored the local variables which were saved when
3435 * this task called schedule() in the past. prev == current
3436 * is still correct, but it can be moved to another cpu/rq.
3437 */
3438 cpu = smp_processor_id();
3439 rq = cpu_rq(cpu);
3440 } else
3441 raw_spin_unlock_irq(&rq->lock);
3442
3443 post_schedule(rq);
3444
3445 sched_preempt_enable_no_resched();
3446 if (need_resched())
3447 goto need_resched;
3448}
3449
3450static inline void sched_submit_work(struct task_struct *tsk)
3451{
3452 if (!tsk->state || tsk_is_pi_blocked(tsk))
3453 return;
3454 /*
3455 * If we are going to sleep and we have plugged IO queued,
3456 * make sure to submit it to avoid deadlocks.
3457 */
3458 if (blk_needs_flush_plug(tsk))
3459 blk_schedule_flush_plug(tsk);
3460}
3461
3462asmlinkage void __sched schedule(void)
3463{
3464 struct task_struct *tsk = current;
3465
3466 sched_submit_work(tsk);
3467 __schedule();
3468}
3469EXPORT_SYMBOL(schedule);
3470
3471/**
3472 * schedule_preempt_disabled - called with preemption disabled
3473 *
3474 * Returns with preemption disabled. Note: preempt_count must be 1
3475 */
3476void __sched schedule_preempt_disabled(void)
3477{
3478 sched_preempt_enable_no_resched();
3479 schedule();
3480 preempt_disable();
3481}
3482
3483#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3484
3485static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3486{
3487 if (lock->owner != owner)
3488 return false;
3489
3490 /*
3491 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3492 * lock->owner still matches owner, if that fails, owner might
3493 * point to free()d memory, if it still matches, the rcu_read_lock()
3494 * ensures the memory stays valid.
3495 */
3496 barrier();
3497
3498 return owner->on_cpu;
3499}
3500
3501/*
3502 * Look out! "owner" is an entirely speculative pointer
3503 * access and not reliable.
3504 */
3505int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3506{
3507 if (!sched_feat(OWNER_SPIN))
3508 return 0;
3509
3510 rcu_read_lock();
3511 while (owner_running(lock, owner)) {
3512 if (need_resched())
3513 break;
3514
3515 arch_mutex_cpu_relax();
3516 }
3517 rcu_read_unlock();
3518
3519 /*
3520 * We break out the loop above on need_resched() and when the
3521 * owner changed, which is a sign for heavy contention. Return
3522 * success only when lock->owner is NULL.
3523 */
3524 return lock->owner == NULL;
3525}
3526#endif
3527
3528#ifdef CONFIG_PREEMPT
3529/*
3530 * this is the entry point to schedule() from in-kernel preemption
3531 * off of preempt_enable. Kernel preemptions off return from interrupt
3532 * occur there and call schedule directly.
3533 */
3534asmlinkage void __sched notrace preempt_schedule(void)
3535{
3536 struct thread_info *ti = current_thread_info();
3537
3538 /*
3539 * If there is a non-zero preempt_count or interrupts are disabled,
3540 * we do not want to preempt the current task. Just return..
3541 */
3542 if (likely(ti->preempt_count || irqs_disabled()))
3543 return;
3544
3545 do {
3546 add_preempt_count_notrace(PREEMPT_ACTIVE);
3547 __schedule();
3548 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3549
3550 /*
3551 * Check again in case we missed a preemption opportunity
3552 * between schedule and now.
3553 */
3554 barrier();
3555 } while (need_resched());
3556}
3557EXPORT_SYMBOL(preempt_schedule);
3558
3559/*
3560 * this is the entry point to schedule() from kernel preemption
3561 * off of irq context.
3562 * Note, that this is called and return with irqs disabled. This will
3563 * protect us against recursive calling from irq.
3564 */
3565asmlinkage void __sched preempt_schedule_irq(void)
3566{
3567 struct thread_info *ti = current_thread_info();
3568
3569 /* Catch callers which need to be fixed */
3570 BUG_ON(ti->preempt_count || !irqs_disabled());
3571
3572 do {
3573 add_preempt_count(PREEMPT_ACTIVE);
3574 local_irq_enable();
3575 __schedule();
3576 local_irq_disable();
3577 sub_preempt_count(PREEMPT_ACTIVE);
3578
3579 /*
3580 * Check again in case we missed a preemption opportunity
3581 * between schedule and now.
3582 */
3583 barrier();
3584 } while (need_resched());
3585}
3586
3587#endif /* CONFIG_PREEMPT */
3588
3589int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3590 void *key)
3591{
3592 return try_to_wake_up(curr->private, mode, wake_flags);
3593}
3594EXPORT_SYMBOL(default_wake_function);
3595
3596/*
3597 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3598 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3599 * number) then we wake all the non-exclusive tasks and one exclusive task.
3600 *
3601 * There are circumstances in which we can try to wake a task which has already
3602 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3603 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3604 */
3605static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3606 int nr_exclusive, int wake_flags, void *key)
3607{
3608 wait_queue_t *curr, *next;
3609
3610 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3611 unsigned flags = curr->flags;
3612
3613 if (curr->func(curr, mode, wake_flags, key) &&
3614 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3615 break;
3616 }
3617}
3618
3619/**
3620 * __wake_up - wake up threads blocked on a waitqueue.
3621 * @q: the waitqueue
3622 * @mode: which threads
3623 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3624 * @key: is directly passed to the wakeup function
3625 *
3626 * It may be assumed that this function implies a write memory barrier before
3627 * changing the task state if and only if any tasks are woken up.
3628 */
3629void __wake_up(wait_queue_head_t *q, unsigned int mode,
3630 int nr_exclusive, void *key)
3631{
3632 unsigned long flags;
3633
3634 spin_lock_irqsave(&q->lock, flags);
3635 __wake_up_common(q, mode, nr_exclusive, 0, key);
3636 spin_unlock_irqrestore(&q->lock, flags);
3637}
3638EXPORT_SYMBOL(__wake_up);
3639
3640/*
3641 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3642 */
3643void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3644{
3645 __wake_up_common(q, mode, nr, 0, NULL);
3646}
3647EXPORT_SYMBOL_GPL(__wake_up_locked);
3648
3649void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3650{
3651 __wake_up_common(q, mode, 1, 0, key);
3652}
3653EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3654
3655/**
3656 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3657 * @q: the waitqueue
3658 * @mode: which threads
3659 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3660 * @key: opaque value to be passed to wakeup targets
3661 *
3662 * The sync wakeup differs that the waker knows that it will schedule
3663 * away soon, so while the target thread will be woken up, it will not
3664 * be migrated to another CPU - ie. the two threads are 'synchronized'
3665 * with each other. This can prevent needless bouncing between CPUs.
3666 *
3667 * On UP it can prevent extra preemption.
3668 *
3669 * It may be assumed that this function implies a write memory barrier before
3670 * changing the task state if and only if any tasks are woken up.
3671 */
3672void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3673 int nr_exclusive, void *key)
3674{
3675 unsigned long flags;
3676 int wake_flags = WF_SYNC;
3677
3678 if (unlikely(!q))
3679 return;
3680
3681 if (unlikely(!nr_exclusive))
3682 wake_flags = 0;
3683
3684 spin_lock_irqsave(&q->lock, flags);
3685 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3686 spin_unlock_irqrestore(&q->lock, flags);
3687}
3688EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3689
3690/*
3691 * __wake_up_sync - see __wake_up_sync_key()
3692 */
3693void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3694{
3695 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3696}
3697EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3698
3699/**
3700 * complete: - signals a single thread waiting on this completion
3701 * @x: holds the state of this particular completion
3702 *
3703 * This will wake up a single thread waiting on this completion. Threads will be
3704 * awakened in the same order in which they were queued.
3705 *
3706 * See also complete_all(), wait_for_completion() and related routines.
3707 *
3708 * It may be assumed that this function implies a write memory barrier before
3709 * changing the task state if and only if any tasks are woken up.
3710 */
3711void complete(struct completion *x)
3712{
3713 unsigned long flags;
3714
3715 spin_lock_irqsave(&x->wait.lock, flags);
3716 x->done++;
3717 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3718 spin_unlock_irqrestore(&x->wait.lock, flags);
3719}
3720EXPORT_SYMBOL(complete);
3721
3722/**
3723 * complete_all: - signals all threads waiting on this completion
3724 * @x: holds the state of this particular completion
3725 *
3726 * This will wake up all threads waiting on this particular completion event.
3727 *
3728 * It may be assumed that this function implies a write memory barrier before
3729 * changing the task state if and only if any tasks are woken up.
3730 */
3731void complete_all(struct completion *x)
3732{
3733 unsigned long flags;
3734
3735 spin_lock_irqsave(&x->wait.lock, flags);
3736 x->done += UINT_MAX/2;
3737 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3738 spin_unlock_irqrestore(&x->wait.lock, flags);
3739}
3740EXPORT_SYMBOL(complete_all);
3741
3742static inline long __sched
3743do_wait_for_common(struct completion *x, long timeout, int state)
3744{
3745 if (!x->done) {
3746 DECLARE_WAITQUEUE(wait, current);
3747
3748 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3749 do {
3750 if (signal_pending_state(state, current)) {
3751 timeout = -ERESTARTSYS;
3752 break;
3753 }
3754 __set_current_state(state);
3755 spin_unlock_irq(&x->wait.lock);
3756 timeout = schedule_timeout(timeout);
3757 spin_lock_irq(&x->wait.lock);
3758 } while (!x->done && timeout);
3759 __remove_wait_queue(&x->wait, &wait);
3760 if (!x->done)
3761 return timeout;
3762 }
3763 x->done--;
3764 return timeout ?: 1;
3765}
3766
3767static long __sched
3768wait_for_common(struct completion *x, long timeout, int state)
3769{
3770 might_sleep();
3771
3772 spin_lock_irq(&x->wait.lock);
3773 timeout = do_wait_for_common(x, timeout, state);
3774 spin_unlock_irq(&x->wait.lock);
3775 return timeout;
3776}
3777
3778/**
3779 * wait_for_completion: - waits for completion of a task
3780 * @x: holds the state of this particular completion
3781 *
3782 * This waits to be signaled for completion of a specific task. It is NOT
3783 * interruptible and there is no timeout.
3784 *
3785 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3786 * and interrupt capability. Also see complete().
3787 */
3788void __sched wait_for_completion(struct completion *x)
3789{
3790 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3791}
3792EXPORT_SYMBOL(wait_for_completion);
3793
3794/**
3795 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3796 * @x: holds the state of this particular completion
3797 * @timeout: timeout value in jiffies
3798 *
3799 * This waits for either a completion of a specific task to be signaled or for a
3800 * specified timeout to expire. The timeout is in jiffies. It is not
3801 * interruptible.
3802 *
3803 * The return value is 0 if timed out, and positive (at least 1, or number of
3804 * jiffies left till timeout) if completed.
3805 */
3806unsigned long __sched
3807wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3808{
3809 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3810}
3811EXPORT_SYMBOL(wait_for_completion_timeout);
3812
3813/**
3814 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3815 * @x: holds the state of this particular completion
3816 *
3817 * This waits for completion of a specific task to be signaled. It is
3818 * interruptible.
3819 *
3820 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3821 */
3822int __sched wait_for_completion_interruptible(struct completion *x)
3823{
3824 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3825 if (t == -ERESTARTSYS)
3826 return t;
3827 return 0;
3828}
3829EXPORT_SYMBOL(wait_for_completion_interruptible);
3830
3831/**
3832 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3833 * @x: holds the state of this particular completion
3834 * @timeout: timeout value in jiffies
3835 *
3836 * This waits for either a completion of a specific task to be signaled or for a
3837 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3838 *
3839 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3840 * positive (at least 1, or number of jiffies left till timeout) if completed.
3841 */
3842long __sched
3843wait_for_completion_interruptible_timeout(struct completion *x,
3844 unsigned long timeout)
3845{
3846 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3847}
3848EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3849
3850/**
3851 * wait_for_completion_killable: - waits for completion of a task (killable)
3852 * @x: holds the state of this particular completion
3853 *
3854 * This waits to be signaled for completion of a specific task. It can be
3855 * interrupted by a kill signal.
3856 *
3857 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3858 */
3859int __sched wait_for_completion_killable(struct completion *x)
3860{
3861 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3862 if (t == -ERESTARTSYS)
3863 return t;
3864 return 0;
3865}
3866EXPORT_SYMBOL(wait_for_completion_killable);
3867
3868/**
3869 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3870 * @x: holds the state of this particular completion
3871 * @timeout: timeout value in jiffies
3872 *
3873 * This waits for either a completion of a specific task to be
3874 * signaled or for a specified timeout to expire. It can be
3875 * interrupted by a kill signal. The timeout is in jiffies.
3876 *
3877 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3878 * positive (at least 1, or number of jiffies left till timeout) if completed.
3879 */
3880long __sched
3881wait_for_completion_killable_timeout(struct completion *x,
3882 unsigned long timeout)
3883{
3884 return wait_for_common(x, timeout, TASK_KILLABLE);
3885}
3886EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3887
3888/**
3889 * try_wait_for_completion - try to decrement a completion without blocking
3890 * @x: completion structure
3891 *
3892 * Returns: 0 if a decrement cannot be done without blocking
3893 * 1 if a decrement succeeded.
3894 *
3895 * If a completion is being used as a counting completion,
3896 * attempt to decrement the counter without blocking. This
3897 * enables us to avoid waiting if the resource the completion
3898 * is protecting is not available.
3899 */
3900bool try_wait_for_completion(struct completion *x)
3901{
3902 unsigned long flags;
3903 int ret = 1;
3904
3905 spin_lock_irqsave(&x->wait.lock, flags);
3906 if (!x->done)
3907 ret = 0;
3908 else
3909 x->done--;
3910 spin_unlock_irqrestore(&x->wait.lock, flags);
3911 return ret;
3912}
3913EXPORT_SYMBOL(try_wait_for_completion);
3914
3915/**
3916 * completion_done - Test to see if a completion has any waiters
3917 * @x: completion structure
3918 *
3919 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3920 * 1 if there are no waiters.
3921 *
3922 */
3923bool completion_done(struct completion *x)
3924{
3925 unsigned long flags;
3926 int ret = 1;
3927
3928 spin_lock_irqsave(&x->wait.lock, flags);
3929 if (!x->done)
3930 ret = 0;
3931 spin_unlock_irqrestore(&x->wait.lock, flags);
3932 return ret;
3933}
3934EXPORT_SYMBOL(completion_done);
3935
3936static long __sched
3937sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3938{
3939 unsigned long flags;
3940 wait_queue_t wait;
3941
3942 init_waitqueue_entry(&wait, current);
3943
3944 __set_current_state(state);
3945
3946 spin_lock_irqsave(&q->lock, flags);
3947 __add_wait_queue(q, &wait);
3948 spin_unlock(&q->lock);
3949 timeout = schedule_timeout(timeout);
3950 spin_lock_irq(&q->lock);
3951 __remove_wait_queue(q, &wait);
3952 spin_unlock_irqrestore(&q->lock, flags);
3953
3954 return timeout;
3955}
3956
3957void __sched interruptible_sleep_on(wait_queue_head_t *q)
3958{
3959 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3960}
3961EXPORT_SYMBOL(interruptible_sleep_on);
3962
3963long __sched
3964interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3965{
3966 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3967}
3968EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3969
3970void __sched sleep_on(wait_queue_head_t *q)
3971{
3972 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3973}
3974EXPORT_SYMBOL(sleep_on);
3975
3976long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3977{
3978 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3979}
3980EXPORT_SYMBOL(sleep_on_timeout);
3981
3982#ifdef CONFIG_RT_MUTEXES
3983
3984/*
3985 * rt_mutex_setprio - set the current priority of a task
3986 * @p: task
3987 * @prio: prio value (kernel-internal form)
3988 *
3989 * This function changes the 'effective' priority of a task. It does
3990 * not touch ->normal_prio like __setscheduler().
3991 *
3992 * Used by the rt_mutex code to implement priority inheritance logic.
3993 */
3994void rt_mutex_setprio(struct task_struct *p, int prio)
3995{
3996 int oldprio, on_rq, running;
3997 struct rq *rq;
3998 const struct sched_class *prev_class;
3999
4000 BUG_ON(prio < 0 || prio > MAX_PRIO);
4001
4002 rq = __task_rq_lock(p);
4003
4004 /*
4005 * Idle task boosting is a nono in general. There is one
4006 * exception, when PREEMPT_RT and NOHZ is active:
4007 *
4008 * The idle task calls get_next_timer_interrupt() and holds
4009 * the timer wheel base->lock on the CPU and another CPU wants
4010 * to access the timer (probably to cancel it). We can safely
4011 * ignore the boosting request, as the idle CPU runs this code
4012 * with interrupts disabled and will complete the lock
4013 * protected section without being interrupted. So there is no
4014 * real need to boost.
4015 */
4016 if (unlikely(p == rq->idle)) {
4017 WARN_ON(p != rq->curr);
4018 WARN_ON(p->pi_blocked_on);
4019 goto out_unlock;
4020 }
4021
4022 trace_sched_pi_setprio(p, prio);
4023 oldprio = p->prio;
4024 prev_class = p->sched_class;
4025 on_rq = p->on_rq;
4026 running = task_current(rq, p);
4027 if (on_rq)
4028 dequeue_task(rq, p, 0);
4029 if (running)
4030 p->sched_class->put_prev_task(rq, p);
4031
4032 if (rt_prio(prio))
4033 p->sched_class = &rt_sched_class;
4034 else
4035 p->sched_class = &fair_sched_class;
4036
4037 p->prio = prio;
4038
4039 if (running)
4040 p->sched_class->set_curr_task(rq);
4041 if (on_rq)
4042 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4043
4044 check_class_changed(rq, p, prev_class, oldprio);
4045out_unlock:
4046 __task_rq_unlock(rq);
4047}
4048#endif
4049void set_user_nice(struct task_struct *p, long nice)
4050{
4051 int old_prio, delta, on_rq;
4052 unsigned long flags;
4053 struct rq *rq;
4054
4055 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4056 return;
4057 /*
4058 * We have to be careful, if called from sys_setpriority(),
4059 * the task might be in the middle of scheduling on another CPU.
4060 */
4061 rq = task_rq_lock(p, &flags);
4062 /*
4063 * The RT priorities are set via sched_setscheduler(), but we still
4064 * allow the 'normal' nice value to be set - but as expected
4065 * it wont have any effect on scheduling until the task is
4066 * SCHED_FIFO/SCHED_RR:
4067 */
4068 if (task_has_rt_policy(p)) {
4069 p->static_prio = NICE_TO_PRIO(nice);
4070 goto out_unlock;
4071 }
4072 on_rq = p->on_rq;
4073 if (on_rq)
4074 dequeue_task(rq, p, 0);
4075
4076 p->static_prio = NICE_TO_PRIO(nice);
4077 set_load_weight(p);
4078 old_prio = p->prio;
4079 p->prio = effective_prio(p);
4080 delta = p->prio - old_prio;
4081
4082 if (on_rq) {
4083 enqueue_task(rq, p, 0);
4084 /*
4085 * If the task increased its priority or is running and
4086 * lowered its priority, then reschedule its CPU:
4087 */
4088 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4089 resched_task(rq->curr);
4090 }
4091out_unlock:
4092 task_rq_unlock(rq, p, &flags);
4093}
4094EXPORT_SYMBOL(set_user_nice);
4095
4096/*
4097 * can_nice - check if a task can reduce its nice value
4098 * @p: task
4099 * @nice: nice value
4100 */
4101int can_nice(const struct task_struct *p, const int nice)
4102{
4103 /* convert nice value [19,-20] to rlimit style value [1,40] */
4104 int nice_rlim = 20 - nice;
4105
4106 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4107 capable(CAP_SYS_NICE));
4108}
4109
4110#ifdef __ARCH_WANT_SYS_NICE
4111
4112/*
4113 * sys_nice - change the priority of the current process.
4114 * @increment: priority increment
4115 *
4116 * sys_setpriority is a more generic, but much slower function that
4117 * does similar things.
4118 */
4119SYSCALL_DEFINE1(nice, int, increment)
4120{
4121 long nice, retval;
4122
4123 /*
4124 * Setpriority might change our priority at the same moment.
4125 * We don't have to worry. Conceptually one call occurs first
4126 * and we have a single winner.
4127 */
4128 if (increment < -40)
4129 increment = -40;
4130 if (increment > 40)
4131 increment = 40;
4132
4133 nice = TASK_NICE(current) + increment;
4134 if (nice < -20)
4135 nice = -20;
4136 if (nice > 19)
4137 nice = 19;
4138
4139 if (increment < 0 && !can_nice(current, nice))
4140 return -EPERM;
4141
4142 retval = security_task_setnice(current, nice);
4143 if (retval)
4144 return retval;
4145
4146 set_user_nice(current, nice);
4147 return 0;
4148}
4149
4150#endif
4151
4152/**
4153 * task_prio - return the priority value of a given task.
4154 * @p: the task in question.
4155 *
4156 * This is the priority value as seen by users in /proc.
4157 * RT tasks are offset by -200. Normal tasks are centered
4158 * around 0, value goes from -16 to +15.
4159 */
4160int task_prio(const struct task_struct *p)
4161{
4162 return p->prio - MAX_RT_PRIO;
4163}
4164
4165/**
4166 * task_nice - return the nice value of a given task.
4167 * @p: the task in question.
4168 */
4169int task_nice(const struct task_struct *p)
4170{
4171 return TASK_NICE(p);
4172}
4173EXPORT_SYMBOL(task_nice);
4174
4175/**
4176 * idle_cpu - is a given cpu idle currently?
4177 * @cpu: the processor in question.
4178 */
4179int idle_cpu(int cpu)
4180{
4181 struct rq *rq = cpu_rq(cpu);
4182
4183 if (rq->curr != rq->idle)
4184 return 0;
4185
4186 if (rq->nr_running)
4187 return 0;
4188
4189#ifdef CONFIG_SMP
4190 if (!llist_empty(&rq->wake_list))
4191 return 0;
4192#endif
4193
4194 return 1;
4195}
4196
4197/**
4198 * idle_task - return the idle task for a given cpu.
4199 * @cpu: the processor in question.
4200 */
4201struct task_struct *idle_task(int cpu)
4202{
4203 return cpu_rq(cpu)->idle;
4204}
4205
4206/**
4207 * find_process_by_pid - find a process with a matching PID value.
4208 * @pid: the pid in question.
4209 */
4210static struct task_struct *find_process_by_pid(pid_t pid)
4211{
4212 return pid ? find_task_by_vpid(pid) : current;
4213}
4214
4215/* Actually do priority change: must hold rq lock. */
4216static void
4217__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4218{
4219 p->policy = policy;
4220 p->rt_priority = prio;
4221 p->normal_prio = normal_prio(p);
4222 /* we are holding p->pi_lock already */
4223 p->prio = rt_mutex_getprio(p);
4224 if (rt_prio(p->prio))
4225 p->sched_class = &rt_sched_class;
4226 else
4227 p->sched_class = &fair_sched_class;
4228 set_load_weight(p);
4229}
4230
4231/*
4232 * check the target process has a UID that matches the current process's
4233 */
4234static bool check_same_owner(struct task_struct *p)
4235{
4236 const struct cred *cred = current_cred(), *pcred;
4237 bool match;
4238
4239 rcu_read_lock();
4240 pcred = __task_cred(p);
4241 match = (uid_eq(cred->euid, pcred->euid) ||
4242 uid_eq(cred->euid, pcred->uid));
4243 rcu_read_unlock();
4244 return match;
4245}
4246
4247static int __sched_setscheduler(struct task_struct *p, int policy,
4248 const struct sched_param *param, bool user)
4249{
4250 int retval, oldprio, oldpolicy = -1, on_rq, running;
4251 unsigned long flags;
4252 const struct sched_class *prev_class;
4253 struct rq *rq;
4254 int reset_on_fork;
4255
4256 /* may grab non-irq protected spin_locks */
4257 BUG_ON(in_interrupt());
4258recheck:
4259 /* double check policy once rq lock held */
4260 if (policy < 0) {
4261 reset_on_fork = p->sched_reset_on_fork;
4262 policy = oldpolicy = p->policy;
4263 } else {
4264 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4265 policy &= ~SCHED_RESET_ON_FORK;
4266
4267 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4268 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4269 policy != SCHED_IDLE)
4270 return -EINVAL;
4271 }
4272
4273 /*
4274 * Valid priorities for SCHED_FIFO and SCHED_RR are
4275 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4276 * SCHED_BATCH and SCHED_IDLE is 0.
4277 */
4278 if (param->sched_priority < 0 ||
4279 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4280 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4281 return -EINVAL;
4282 if (rt_policy(policy) != (param->sched_priority != 0))
4283 return -EINVAL;
4284
4285 /*
4286 * Allow unprivileged RT tasks to decrease priority:
4287 */
4288 if (user && !capable(CAP_SYS_NICE)) {
4289 if (rt_policy(policy)) {
4290 unsigned long rlim_rtprio =
4291 task_rlimit(p, RLIMIT_RTPRIO);
4292
4293 /* can't set/change the rt policy */
4294 if (policy != p->policy && !rlim_rtprio)
4295 return -EPERM;
4296
4297 /* can't increase priority */
4298 if (param->sched_priority > p->rt_priority &&
4299 param->sched_priority > rlim_rtprio)
4300 return -EPERM;
4301 }
4302
4303 /*
4304 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4305 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4306 */
4307 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4308 if (!can_nice(p, TASK_NICE(p)))
4309 return -EPERM;
4310 }
4311
4312 /* can't change other user's priorities */
4313 if (!check_same_owner(p))
4314 return -EPERM;
4315
4316 /* Normal users shall not reset the sched_reset_on_fork flag */
4317 if (p->sched_reset_on_fork && !reset_on_fork)
4318 return -EPERM;
4319 }
4320
4321 if (user) {
4322 retval = security_task_setscheduler(p);
4323 if (retval)
4324 return retval;
4325 }
4326
4327 /*
4328 * make sure no PI-waiters arrive (or leave) while we are
4329 * changing the priority of the task:
4330 *
4331 * To be able to change p->policy safely, the appropriate
4332 * runqueue lock must be held.
4333 */
4334 rq = task_rq_lock(p, &flags);
4335
4336 /*
4337 * Changing the policy of the stop threads its a very bad idea
4338 */
4339 if (p == rq->stop) {
4340 task_rq_unlock(rq, p, &flags);
4341 return -EINVAL;
4342 }
4343
4344 /*
4345 * If not changing anything there's no need to proceed further:
4346 */
4347 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4348 param->sched_priority == p->rt_priority))) {
4349
4350 __task_rq_unlock(rq);
4351 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4352 return 0;
4353 }
4354
4355#ifdef CONFIG_RT_GROUP_SCHED
4356 if (user) {
4357 /*
4358 * Do not allow realtime tasks into groups that have no runtime
4359 * assigned.
4360 */
4361 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4362 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4363 !task_group_is_autogroup(task_group(p))) {
4364 task_rq_unlock(rq, p, &flags);
4365 return -EPERM;
4366 }
4367 }
4368#endif
4369
4370 /* recheck policy now with rq lock held */
4371 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4372 policy = oldpolicy = -1;
4373 task_rq_unlock(rq, p, &flags);
4374 goto recheck;
4375 }
4376 on_rq = p->on_rq;
4377 running = task_current(rq, p);
4378 if (on_rq)
4379 dequeue_task(rq, p, 0);
4380 if (running)
4381 p->sched_class->put_prev_task(rq, p);
4382
4383 p->sched_reset_on_fork = reset_on_fork;
4384
4385 oldprio = p->prio;
4386 prev_class = p->sched_class;
4387 __setscheduler(rq, p, policy, param->sched_priority);
4388
4389 if (running)
4390 p->sched_class->set_curr_task(rq);
4391 if (on_rq)
4392 enqueue_task(rq, p, 0);
4393
4394 check_class_changed(rq, p, prev_class, oldprio);
4395 task_rq_unlock(rq, p, &flags);
4396
4397 rt_mutex_adjust_pi(p);
4398
4399 return 0;
4400}
4401
4402/**
4403 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4404 * @p: the task in question.
4405 * @policy: new policy.
4406 * @param: structure containing the new RT priority.
4407 *
4408 * NOTE that the task may be already dead.
4409 */
4410int sched_setscheduler(struct task_struct *p, int policy,
4411 const struct sched_param *param)
4412{
4413 return __sched_setscheduler(p, policy, param, true);
4414}
4415EXPORT_SYMBOL_GPL(sched_setscheduler);
4416
4417/**
4418 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4419 * @p: the task in question.
4420 * @policy: new policy.
4421 * @param: structure containing the new RT priority.
4422 *
4423 * Just like sched_setscheduler, only don't bother checking if the
4424 * current context has permission. For example, this is needed in
4425 * stop_machine(): we create temporary high priority worker threads,
4426 * but our caller might not have that capability.
4427 */
4428int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4429 const struct sched_param *param)
4430{
4431 return __sched_setscheduler(p, policy, param, false);
4432}
4433
4434static int
4435do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4436{
4437 struct sched_param lparam;
4438 struct task_struct *p;
4439 int retval;
4440
4441 if (!param || pid < 0)
4442 return -EINVAL;
4443 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4444 return -EFAULT;
4445
4446 rcu_read_lock();
4447 retval = -ESRCH;
4448 p = find_process_by_pid(pid);
4449 if (p != NULL)
4450 retval = sched_setscheduler(p, policy, &lparam);
4451 rcu_read_unlock();
4452
4453 return retval;
4454}
4455
4456/**
4457 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4458 * @pid: the pid in question.
4459 * @policy: new policy.
4460 * @param: structure containing the new RT priority.
4461 */
4462SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4463 struct sched_param __user *, param)
4464{
4465 /* negative values for policy are not valid */
4466 if (policy < 0)
4467 return -EINVAL;
4468
4469 return do_sched_setscheduler(pid, policy, param);
4470}
4471
4472/**
4473 * sys_sched_setparam - set/change the RT priority of a thread
4474 * @pid: the pid in question.
4475 * @param: structure containing the new RT priority.
4476 */
4477SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4478{
4479 return do_sched_setscheduler(pid, -1, param);
4480}
4481
4482/**
4483 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4484 * @pid: the pid in question.
4485 */
4486SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4487{
4488 struct task_struct *p;
4489 int retval;
4490
4491 if (pid < 0)
4492 return -EINVAL;
4493
4494 retval = -ESRCH;
4495 rcu_read_lock();
4496 p = find_process_by_pid(pid);
4497 if (p) {
4498 retval = security_task_getscheduler(p);
4499 if (!retval)
4500 retval = p->policy
4501 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4502 }
4503 rcu_read_unlock();
4504 return retval;
4505}
4506
4507/**
4508 * sys_sched_getparam - get the RT priority of a thread
4509 * @pid: the pid in question.
4510 * @param: structure containing the RT priority.
4511 */
4512SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4513{
4514 struct sched_param lp;
4515 struct task_struct *p;
4516 int retval;
4517
4518 if (!param || pid < 0)
4519 return -EINVAL;
4520
4521 rcu_read_lock();
4522 p = find_process_by_pid(pid);
4523 retval = -ESRCH;
4524 if (!p)
4525 goto out_unlock;
4526
4527 retval = security_task_getscheduler(p);
4528 if (retval)
4529 goto out_unlock;
4530
4531 lp.sched_priority = p->rt_priority;
4532 rcu_read_unlock();
4533
4534 /*
4535 * This one might sleep, we cannot do it with a spinlock held ...
4536 */
4537 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4538
4539 return retval;
4540
4541out_unlock:
4542 rcu_read_unlock();
4543 return retval;
4544}
4545
4546long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4547{
4548 cpumask_var_t cpus_allowed, new_mask;
4549 struct task_struct *p;
4550 int retval;
4551
4552 get_online_cpus();
4553 rcu_read_lock();
4554
4555 p = find_process_by_pid(pid);
4556 if (!p) {
4557 rcu_read_unlock();
4558 put_online_cpus();
4559 return -ESRCH;
4560 }
4561
4562 /* Prevent p going away */
4563 get_task_struct(p);
4564 rcu_read_unlock();
4565
4566 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4567 retval = -ENOMEM;
4568 goto out_put_task;
4569 }
4570 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4571 retval = -ENOMEM;
4572 goto out_free_cpus_allowed;
4573 }
4574 retval = -EPERM;
4575 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4576 goto out_unlock;
4577
4578 retval = security_task_setscheduler(p);
4579 if (retval)
4580 goto out_unlock;
4581
4582 cpuset_cpus_allowed(p, cpus_allowed);
4583 cpumask_and(new_mask, in_mask, cpus_allowed);
4584again:
4585 retval = set_cpus_allowed_ptr(p, new_mask);
4586
4587 if (!retval) {
4588 cpuset_cpus_allowed(p, cpus_allowed);
4589 if (!cpumask_subset(new_mask, cpus_allowed)) {
4590 /*
4591 * We must have raced with a concurrent cpuset
4592 * update. Just reset the cpus_allowed to the
4593 * cpuset's cpus_allowed
4594 */
4595 cpumask_copy(new_mask, cpus_allowed);
4596 goto again;
4597 }
4598 }
4599out_unlock:
4600 free_cpumask_var(new_mask);
4601out_free_cpus_allowed:
4602 free_cpumask_var(cpus_allowed);
4603out_put_task:
4604 put_task_struct(p);
4605 put_online_cpus();
4606 return retval;
4607}
4608
4609static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4610 struct cpumask *new_mask)
4611{
4612 if (len < cpumask_size())
4613 cpumask_clear(new_mask);
4614 else if (len > cpumask_size())
4615 len = cpumask_size();
4616
4617 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4618}
4619
4620/**
4621 * sys_sched_setaffinity - set the cpu affinity of a process
4622 * @pid: pid of the process
4623 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4624 * @user_mask_ptr: user-space pointer to the new cpu mask
4625 */
4626SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4627 unsigned long __user *, user_mask_ptr)
4628{
4629 cpumask_var_t new_mask;
4630 int retval;
4631
4632 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4633 return -ENOMEM;
4634
4635 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4636 if (retval == 0)
4637 retval = sched_setaffinity(pid, new_mask);
4638 free_cpumask_var(new_mask);
4639 return retval;
4640}
4641
4642long sched_getaffinity(pid_t pid, struct cpumask *mask)
4643{
4644 struct task_struct *p;
4645 unsigned long flags;
4646 int retval;
4647
4648 get_online_cpus();
4649 rcu_read_lock();
4650
4651 retval = -ESRCH;
4652 p = find_process_by_pid(pid);
4653 if (!p)
4654 goto out_unlock;
4655
4656 retval = security_task_getscheduler(p);
4657 if (retval)
4658 goto out_unlock;
4659
4660 raw_spin_lock_irqsave(&p->pi_lock, flags);
4661 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4662 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4663
4664out_unlock:
4665 rcu_read_unlock();
4666 put_online_cpus();
4667
4668 return retval;
4669}
4670
4671/**
4672 * sys_sched_getaffinity - get the cpu affinity of a process
4673 * @pid: pid of the process
4674 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4675 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4676 */
4677SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4678 unsigned long __user *, user_mask_ptr)
4679{
4680 int ret;
4681 cpumask_var_t mask;
4682
4683 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4684 return -EINVAL;
4685 if (len & (sizeof(unsigned long)-1))
4686 return -EINVAL;
4687
4688 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4689 return -ENOMEM;
4690
4691 ret = sched_getaffinity(pid, mask);
4692 if (ret == 0) {
4693 size_t retlen = min_t(size_t, len, cpumask_size());
4694
4695 if (copy_to_user(user_mask_ptr, mask, retlen))
4696 ret = -EFAULT;
4697 else
4698 ret = retlen;
4699 }
4700 free_cpumask_var(mask);
4701
4702 return ret;
4703}
4704
4705/**
4706 * sys_sched_yield - yield the current processor to other threads.
4707 *
4708 * This function yields the current CPU to other tasks. If there are no
4709 * other threads running on this CPU then this function will return.
4710 */
4711SYSCALL_DEFINE0(sched_yield)
4712{
4713 struct rq *rq = this_rq_lock();
4714
4715 schedstat_inc(rq, yld_count);
4716 current->sched_class->yield_task(rq);
4717
4718 /*
4719 * Since we are going to call schedule() anyway, there's
4720 * no need to preempt or enable interrupts:
4721 */
4722 __release(rq->lock);
4723 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4724 do_raw_spin_unlock(&rq->lock);
4725 sched_preempt_enable_no_resched();
4726
4727 schedule();
4728
4729 return 0;
4730}
4731
4732static inline int should_resched(void)
4733{
4734 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4735}
4736
4737static void __cond_resched(void)
4738{
4739 add_preempt_count(PREEMPT_ACTIVE);
4740 __schedule();
4741 sub_preempt_count(PREEMPT_ACTIVE);
4742}
4743
4744int __sched _cond_resched(void)
4745{
4746 if (should_resched()) {
4747 __cond_resched();
4748 return 1;
4749 }
4750 return 0;
4751}
4752EXPORT_SYMBOL(_cond_resched);
4753
4754/*
4755 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4756 * call schedule, and on return reacquire the lock.
4757 *
4758 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4759 * operations here to prevent schedule() from being called twice (once via
4760 * spin_unlock(), once by hand).
4761 */
4762int __cond_resched_lock(spinlock_t *lock)
4763{
4764 int resched = should_resched();
4765 int ret = 0;
4766
4767 lockdep_assert_held(lock);
4768
4769 if (spin_needbreak(lock) || resched) {
4770 spin_unlock(lock);
4771 if (resched)
4772 __cond_resched();
4773 else
4774 cpu_relax();
4775 ret = 1;
4776 spin_lock(lock);
4777 }
4778 return ret;
4779}
4780EXPORT_SYMBOL(__cond_resched_lock);
4781
4782int __sched __cond_resched_softirq(void)
4783{
4784 BUG_ON(!in_softirq());
4785
4786 if (should_resched()) {
4787 local_bh_enable();
4788 __cond_resched();
4789 local_bh_disable();
4790 return 1;
4791 }
4792 return 0;
4793}
4794EXPORT_SYMBOL(__cond_resched_softirq);
4795
4796/**
4797 * yield - yield the current processor to other threads.
4798 *
4799 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4800 *
4801 * The scheduler is at all times free to pick the calling task as the most
4802 * eligible task to run, if removing the yield() call from your code breaks
4803 * it, its already broken.
4804 *
4805 * Typical broken usage is:
4806 *
4807 * while (!event)
4808 * yield();
4809 *
4810 * where one assumes that yield() will let 'the other' process run that will
4811 * make event true. If the current task is a SCHED_FIFO task that will never
4812 * happen. Never use yield() as a progress guarantee!!
4813 *
4814 * If you want to use yield() to wait for something, use wait_event().
4815 * If you want to use yield() to be 'nice' for others, use cond_resched().
4816 * If you still want to use yield(), do not!
4817 */
4818void __sched yield(void)
4819{
4820 set_current_state(TASK_RUNNING);
4821 sys_sched_yield();
4822}
4823EXPORT_SYMBOL(yield);
4824
4825/**
4826 * yield_to - yield the current processor to another thread in
4827 * your thread group, or accelerate that thread toward the
4828 * processor it's on.
4829 * @p: target task
4830 * @preempt: whether task preemption is allowed or not
4831 *
4832 * It's the caller's job to ensure that the target task struct
4833 * can't go away on us before we can do any checks.
4834 *
4835 * Returns true if we indeed boosted the target task.
4836 */
4837bool __sched yield_to(struct task_struct *p, bool preempt)
4838{
4839 struct task_struct *curr = current;
4840 struct rq *rq, *p_rq;
4841 unsigned long flags;
4842 bool yielded = 0;
4843
4844 local_irq_save(flags);
4845 rq = this_rq();
4846
4847again:
4848 p_rq = task_rq(p);
4849 double_rq_lock(rq, p_rq);
4850 while (task_rq(p) != p_rq) {
4851 double_rq_unlock(rq, p_rq);
4852 goto again;
4853 }
4854
4855 if (!curr->sched_class->yield_to_task)
4856 goto out;
4857
4858 if (curr->sched_class != p->sched_class)
4859 goto out;
4860
4861 if (task_running(p_rq, p) || p->state)
4862 goto out;
4863
4864 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4865 if (yielded) {
4866 schedstat_inc(rq, yld_count);
4867 /*
4868 * Make p's CPU reschedule; pick_next_entity takes care of
4869 * fairness.
4870 */
4871 if (preempt && rq != p_rq)
4872 resched_task(p_rq->curr);
4873 } else {
4874 /*
4875 * We might have set it in task_yield_fair(), but are
4876 * not going to schedule(), so don't want to skip
4877 * the next update.
4878 */
4879 rq->skip_clock_update = 0;
4880 }
4881
4882out:
4883 double_rq_unlock(rq, p_rq);
4884 local_irq_restore(flags);
4885
4886 if (yielded)
4887 schedule();
4888
4889 return yielded;
4890}
4891EXPORT_SYMBOL_GPL(yield_to);
4892
4893/*
4894 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4895 * that process accounting knows that this is a task in IO wait state.
4896 */
4897void __sched io_schedule(void)
4898{
4899 struct rq *rq = raw_rq();
4900
4901 delayacct_blkio_start();
4902 atomic_inc(&rq->nr_iowait);
4903 blk_flush_plug(current);
4904 current->in_iowait = 1;
4905 schedule();
4906 current->in_iowait = 0;
4907 atomic_dec(&rq->nr_iowait);
4908 delayacct_blkio_end();
4909}
4910EXPORT_SYMBOL(io_schedule);
4911
4912long __sched io_schedule_timeout(long timeout)
4913{
4914 struct rq *rq = raw_rq();
4915 long ret;
4916
4917 delayacct_blkio_start();
4918 atomic_inc(&rq->nr_iowait);
4919 blk_flush_plug(current);
4920 current->in_iowait = 1;
4921 ret = schedule_timeout(timeout);
4922 current->in_iowait = 0;
4923 atomic_dec(&rq->nr_iowait);
4924 delayacct_blkio_end();
4925 return ret;
4926}
4927
4928/**
4929 * sys_sched_get_priority_max - return maximum RT priority.
4930 * @policy: scheduling class.
4931 *
4932 * this syscall returns the maximum rt_priority that can be used
4933 * by a given scheduling class.
4934 */
4935SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4936{
4937 int ret = -EINVAL;
4938
4939 switch (policy) {
4940 case SCHED_FIFO:
4941 case SCHED_RR:
4942 ret = MAX_USER_RT_PRIO-1;
4943 break;
4944 case SCHED_NORMAL:
4945 case SCHED_BATCH:
4946 case SCHED_IDLE:
4947 ret = 0;
4948 break;
4949 }
4950 return ret;
4951}
4952
4953/**
4954 * sys_sched_get_priority_min - return minimum RT priority.
4955 * @policy: scheduling class.
4956 *
4957 * this syscall returns the minimum rt_priority that can be used
4958 * by a given scheduling class.
4959 */
4960SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4961{
4962 int ret = -EINVAL;
4963
4964 switch (policy) {
4965 case SCHED_FIFO:
4966 case SCHED_RR:
4967 ret = 1;
4968 break;
4969 case SCHED_NORMAL:
4970 case SCHED_BATCH:
4971 case SCHED_IDLE:
4972 ret = 0;
4973 }
4974 return ret;
4975}
4976
4977/**
4978 * sys_sched_rr_get_interval - return the default timeslice of a process.
4979 * @pid: pid of the process.
4980 * @interval: userspace pointer to the timeslice value.
4981 *
4982 * this syscall writes the default timeslice value of a given process
4983 * into the user-space timespec buffer. A value of '0' means infinity.
4984 */
4985SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4986 struct timespec __user *, interval)
4987{
4988 struct task_struct *p;
4989 unsigned int time_slice;
4990 unsigned long flags;
4991 struct rq *rq;
4992 int retval;
4993 struct timespec t;
4994
4995 if (pid < 0)
4996 return -EINVAL;
4997
4998 retval = -ESRCH;
4999 rcu_read_lock();
5000 p = find_process_by_pid(pid);
5001 if (!p)
5002 goto out_unlock;
5003
5004 retval = security_task_getscheduler(p);
5005 if (retval)
5006 goto out_unlock;
5007
5008 rq = task_rq_lock(p, &flags);
5009 time_slice = p->sched_class->get_rr_interval(rq, p);
5010 task_rq_unlock(rq, p, &flags);
5011
5012 rcu_read_unlock();
5013 jiffies_to_timespec(time_slice, &t);
5014 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5015 return retval;
5016
5017out_unlock:
5018 rcu_read_unlock();
5019 return retval;
5020}
5021
5022static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5023
5024void sched_show_task(struct task_struct *p)
5025{
5026 unsigned long free = 0;
5027 unsigned state;
5028
5029 state = p->state ? __ffs(p->state) + 1 : 0;
5030 printk(KERN_INFO "%-15.15s %c", p->comm,
5031 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5032#if BITS_PER_LONG == 32
5033 if (state == TASK_RUNNING)
5034 printk(KERN_CONT " running ");
5035 else
5036 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5037#else
5038 if (state == TASK_RUNNING)
5039 printk(KERN_CONT " running task ");
5040 else
5041 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5042#endif
5043#ifdef CONFIG_DEBUG_STACK_USAGE
5044 free = stack_not_used(p);
5045#endif
5046 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5047 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
5048 (unsigned long)task_thread_info(p)->flags);
5049
5050 show_stack(p, NULL);
5051}
5052
5053void show_state_filter(unsigned long state_filter)
5054{
5055 struct task_struct *g, *p;
5056
5057#if BITS_PER_LONG == 32
5058 printk(KERN_INFO
5059 " task PC stack pid father\n");
5060#else
5061 printk(KERN_INFO
5062 " task PC stack pid father\n");
5063#endif
5064 rcu_read_lock();
5065 do_each_thread(g, p) {
5066 /*
5067 * reset the NMI-timeout, listing all files on a slow
5068 * console might take a lot of time:
5069 */
5070 touch_nmi_watchdog();
5071 if (!state_filter || (p->state & state_filter))
5072 sched_show_task(p);
5073 } while_each_thread(g, p);
5074
5075 touch_all_softlockup_watchdogs();
5076
5077#ifdef CONFIG_SCHED_DEBUG
5078 sysrq_sched_debug_show();
5079#endif
5080 rcu_read_unlock();
5081 /*
5082 * Only show locks if all tasks are dumped:
5083 */
5084 if (!state_filter)
5085 debug_show_all_locks();
5086}
5087
5088void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5089{
5090 idle->sched_class = &idle_sched_class;
5091}
5092
5093/**
5094 * init_idle - set up an idle thread for a given CPU
5095 * @idle: task in question
5096 * @cpu: cpu the idle task belongs to
5097 *
5098 * NOTE: this function does not set the idle thread's NEED_RESCHED
5099 * flag, to make booting more robust.
5100 */
5101void __cpuinit init_idle(struct task_struct *idle, int cpu)
5102{
5103 struct rq *rq = cpu_rq(cpu);
5104 unsigned long flags;
5105
5106 raw_spin_lock_irqsave(&rq->lock, flags);
5107
5108 __sched_fork(idle);
5109 idle->state = TASK_RUNNING;
5110 idle->se.exec_start = sched_clock();
5111
5112 do_set_cpus_allowed(idle, cpumask_of(cpu));
5113 /*
5114 * We're having a chicken and egg problem, even though we are
5115 * holding rq->lock, the cpu isn't yet set to this cpu so the
5116 * lockdep check in task_group() will fail.
5117 *
5118 * Similar case to sched_fork(). / Alternatively we could
5119 * use task_rq_lock() here and obtain the other rq->lock.
5120 *
5121 * Silence PROVE_RCU
5122 */
5123 rcu_read_lock();
5124 __set_task_cpu(idle, cpu);
5125 rcu_read_unlock();
5126
5127 rq->curr = rq->idle = idle;
5128#if defined(CONFIG_SMP)
5129 idle->on_cpu = 1;
5130#endif
5131 raw_spin_unlock_irqrestore(&rq->lock, flags);
5132
5133 /* Set the preempt count _outside_ the spinlocks! */
5134 task_thread_info(idle)->preempt_count = 0;
5135
5136 /*
5137 * The idle tasks have their own, simple scheduling class:
5138 */
5139 idle->sched_class = &idle_sched_class;
5140 ftrace_graph_init_idle_task(idle, cpu);
5141#if defined(CONFIG_SMP)
5142 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5143#endif
5144}
5145
5146#ifdef CONFIG_SMP
5147void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
5148{
5149 if (p->sched_class && p->sched_class->set_cpus_allowed)
5150 p->sched_class->set_cpus_allowed(p, new_mask);
5151
5152 cpumask_copy(&p->cpus_allowed, new_mask);
5153 p->nr_cpus_allowed = cpumask_weight(new_mask);
5154}
5155
5156/*
5157 * This is how migration works:
5158 *
5159 * 1) we invoke migration_cpu_stop() on the target CPU using
5160 * stop_one_cpu().
5161 * 2) stopper starts to run (implicitly forcing the migrated thread
5162 * off the CPU)
5163 * 3) it checks whether the migrated task is still in the wrong runqueue.
5164 * 4) if it's in the wrong runqueue then the migration thread removes
5165 * it and puts it into the right queue.
5166 * 5) stopper completes and stop_one_cpu() returns and the migration
5167 * is done.
5168 */
5169
5170/*
5171 * Change a given task's CPU affinity. Migrate the thread to a
5172 * proper CPU and schedule it away if the CPU it's executing on
5173 * is removed from the allowed bitmask.
5174 *
5175 * NOTE: the caller must have a valid reference to the task, the
5176 * task must not exit() & deallocate itself prematurely. The
5177 * call is not atomic; no spinlocks may be held.
5178 */
5179int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5180{
5181 unsigned long flags;
5182 struct rq *rq;
5183 unsigned int dest_cpu;
5184 int ret = 0;
5185
5186 rq = task_rq_lock(p, &flags);
5187
5188 if (cpumask_equal(&p->cpus_allowed, new_mask))
5189 goto out;
5190
5191 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5192 ret = -EINVAL;
5193 goto out;
5194 }
5195
5196 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
5197 ret = -EINVAL;
5198 goto out;
5199 }
5200
5201 do_set_cpus_allowed(p, new_mask);
5202
5203 /* Can the task run on the task's current CPU? If so, we're done */
5204 if (cpumask_test_cpu(task_cpu(p), new_mask))
5205 goto out;
5206
5207 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5208 if (p->on_rq) {
5209 struct migration_arg arg = { p, dest_cpu };
5210 /* Need help from migration thread: drop lock and wait. */
5211 task_rq_unlock(rq, p, &flags);
5212 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5213 tlb_migrate_finish(p->mm);
5214 return 0;
5215 }
5216out:
5217 task_rq_unlock(rq, p, &flags);
5218
5219 return ret;
5220}
5221EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5222
5223/*
5224 * Move (not current) task off this cpu, onto dest cpu. We're doing
5225 * this because either it can't run here any more (set_cpus_allowed()
5226 * away from this CPU, or CPU going down), or because we're
5227 * attempting to rebalance this task on exec (sched_exec).
5228 *
5229 * So we race with normal scheduler movements, but that's OK, as long
5230 * as the task is no longer on this CPU.
5231 *
5232 * Returns non-zero if task was successfully migrated.
5233 */
5234static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5235{
5236 struct rq *rq_dest, *rq_src;
5237 int ret = 0;
5238
5239 if (unlikely(!cpu_active(dest_cpu)))
5240 return ret;
5241
5242 rq_src = cpu_rq(src_cpu);
5243 rq_dest = cpu_rq(dest_cpu);
5244
5245 raw_spin_lock(&p->pi_lock);
5246 double_rq_lock(rq_src, rq_dest);
5247 /* Already moved. */
5248 if (task_cpu(p) != src_cpu)
5249 goto done;
5250 /* Affinity changed (again). */
5251 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
5252 goto fail;
5253
5254 /*
5255 * If we're not on a rq, the next wake-up will ensure we're
5256 * placed properly.
5257 */
5258 if (p->on_rq) {
5259 dequeue_task(rq_src, p, 0);
5260 set_task_cpu(p, dest_cpu);
5261 enqueue_task(rq_dest, p, 0);
5262 check_preempt_curr(rq_dest, p, 0);
5263 }
5264done:
5265 ret = 1;
5266fail:
5267 double_rq_unlock(rq_src, rq_dest);
5268 raw_spin_unlock(&p->pi_lock);
5269 return ret;
5270}
5271
5272/*
5273 * migration_cpu_stop - this will be executed by a highprio stopper thread
5274 * and performs thread migration by bumping thread off CPU then
5275 * 'pushing' onto another runqueue.
5276 */
5277static int migration_cpu_stop(void *data)
5278{
5279 struct migration_arg *arg = data;
5280
5281 /*
5282 * The original target cpu might have gone down and we might
5283 * be on another cpu but it doesn't matter.
5284 */
5285 local_irq_disable();
5286 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5287 local_irq_enable();
5288 return 0;
5289}
5290
5291#ifdef CONFIG_HOTPLUG_CPU
5292
5293/*
5294 * Ensures that the idle task is using init_mm right before its cpu goes
5295 * offline.
5296 */
5297void idle_task_exit(void)
5298{
5299 struct mm_struct *mm = current->active_mm;
5300
5301 BUG_ON(cpu_online(smp_processor_id()));
5302
5303 if (mm != &init_mm)
5304 switch_mm(mm, &init_mm, current);
5305 mmdrop(mm);
5306}
5307
5308/*
5309 * While a dead CPU has no uninterruptible tasks queued at this point,
5310 * it might still have a nonzero ->nr_uninterruptible counter, because
5311 * for performance reasons the counter is not stricly tracking tasks to
5312 * their home CPUs. So we just add the counter to another CPU's counter,
5313 * to keep the global sum constant after CPU-down:
5314 */
5315static void migrate_nr_uninterruptible(struct rq *rq_src)
5316{
5317 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5318
5319 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5320 rq_src->nr_uninterruptible = 0;
5321}
5322
5323/*
5324 * remove the tasks which were accounted by rq from calc_load_tasks.
5325 */
5326static void calc_global_load_remove(struct rq *rq)
5327{
5328 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5329 rq->calc_load_active = 0;
5330}
5331
5332/*
5333 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5334 * try_to_wake_up()->select_task_rq().
5335 *
5336 * Called with rq->lock held even though we'er in stop_machine() and
5337 * there's no concurrency possible, we hold the required locks anyway
5338 * because of lock validation efforts.
5339 */
5340static void migrate_tasks(unsigned int dead_cpu)
5341{
5342 struct rq *rq = cpu_rq(dead_cpu);
5343 struct task_struct *next, *stop = rq->stop;
5344 int dest_cpu;
5345
5346 /*
5347 * Fudge the rq selection such that the below task selection loop
5348 * doesn't get stuck on the currently eligible stop task.
5349 *
5350 * We're currently inside stop_machine() and the rq is either stuck
5351 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5352 * either way we should never end up calling schedule() until we're
5353 * done here.
5354 */
5355 rq->stop = NULL;
5356
5357 /* Ensure any throttled groups are reachable by pick_next_task */
5358 unthrottle_offline_cfs_rqs(rq);
5359
5360 for ( ; ; ) {
5361 /*
5362 * There's this thread running, bail when that's the only
5363 * remaining thread.
5364 */
5365 if (rq->nr_running == 1)
5366 break;
5367
5368 next = pick_next_task(rq);
5369 BUG_ON(!next);
5370 next->sched_class->put_prev_task(rq, next);
5371
5372 /* Find suitable destination for @next, with force if needed. */
5373 dest_cpu = select_fallback_rq(dead_cpu, next);
5374 raw_spin_unlock(&rq->lock);
5375
5376 __migrate_task(next, dead_cpu, dest_cpu);
5377
5378 raw_spin_lock(&rq->lock);
5379 }
5380
5381 rq->stop = stop;
5382}
5383
5384#endif /* CONFIG_HOTPLUG_CPU */
5385
5386#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5387
5388static struct ctl_table sd_ctl_dir[] = {
5389 {
5390 .procname = "sched_domain",
5391 .mode = 0555,
5392 },
5393 {}
5394};
5395
5396static struct ctl_table sd_ctl_root[] = {
5397 {
5398 .procname = "kernel",
5399 .mode = 0555,
5400 .child = sd_ctl_dir,
5401 },
5402 {}
5403};
5404
5405static struct ctl_table *sd_alloc_ctl_entry(int n)
5406{
5407 struct ctl_table *entry =
5408 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5409
5410 return entry;
5411}
5412
5413static void sd_free_ctl_entry(struct ctl_table **tablep)
5414{
5415 struct ctl_table *entry;
5416
5417 /*
5418 * In the intermediate directories, both the child directory and
5419 * procname are dynamically allocated and could fail but the mode
5420 * will always be set. In the lowest directory the names are
5421 * static strings and all have proc handlers.
5422 */
5423 for (entry = *tablep; entry->mode; entry++) {
5424 if (entry->child)
5425 sd_free_ctl_entry(&entry->child);
5426 if (entry->proc_handler == NULL)
5427 kfree(entry->procname);
5428 }
5429
5430 kfree(*tablep);
5431 *tablep = NULL;
5432}
5433
5434static void
5435set_table_entry(struct ctl_table *entry,
5436 const char *procname, void *data, int maxlen,
5437 umode_t mode, proc_handler *proc_handler)
5438{
5439 entry->procname = procname;
5440 entry->data = data;
5441 entry->maxlen = maxlen;
5442 entry->mode = mode;
5443 entry->proc_handler = proc_handler;
5444}
5445
5446static struct ctl_table *
5447sd_alloc_ctl_domain_table(struct sched_domain *sd)
5448{
5449 struct ctl_table *table = sd_alloc_ctl_entry(13);
5450
5451 if (table == NULL)
5452 return NULL;
5453
5454 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5455 sizeof(long), 0644, proc_doulongvec_minmax);
5456 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5457 sizeof(long), 0644, proc_doulongvec_minmax);
5458 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5459 sizeof(int), 0644, proc_dointvec_minmax);
5460 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5461 sizeof(int), 0644, proc_dointvec_minmax);
5462 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5463 sizeof(int), 0644, proc_dointvec_minmax);
5464 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5465 sizeof(int), 0644, proc_dointvec_minmax);
5466 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5467 sizeof(int), 0644, proc_dointvec_minmax);
5468 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5469 sizeof(int), 0644, proc_dointvec_minmax);
5470 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5471 sizeof(int), 0644, proc_dointvec_minmax);
5472 set_table_entry(&table[9], "cache_nice_tries",
5473 &sd->cache_nice_tries,
5474 sizeof(int), 0644, proc_dointvec_minmax);
5475 set_table_entry(&table[10], "flags", &sd->flags,
5476 sizeof(int), 0644, proc_dointvec_minmax);
5477 set_table_entry(&table[11], "name", sd->name,
5478 CORENAME_MAX_SIZE, 0444, proc_dostring);
5479 /* &table[12] is terminator */
5480
5481 return table;
5482}
5483
5484static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5485{
5486 struct ctl_table *entry, *table;
5487 struct sched_domain *sd;
5488 int domain_num = 0, i;
5489 char buf[32];
5490
5491 for_each_domain(cpu, sd)
5492 domain_num++;
5493 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5494 if (table == NULL)
5495 return NULL;
5496
5497 i = 0;
5498 for_each_domain(cpu, sd) {
5499 snprintf(buf, 32, "domain%d", i);
5500 entry->procname = kstrdup(buf, GFP_KERNEL);
5501 entry->mode = 0555;
5502 entry->child = sd_alloc_ctl_domain_table(sd);
5503 entry++;
5504 i++;
5505 }
5506 return table;
5507}
5508
5509static struct ctl_table_header *sd_sysctl_header;
5510static void register_sched_domain_sysctl(void)
5511{
5512 int i, cpu_num = num_possible_cpus();
5513 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5514 char buf[32];
5515
5516 WARN_ON(sd_ctl_dir[0].child);
5517 sd_ctl_dir[0].child = entry;
5518
5519 if (entry == NULL)
5520 return;
5521
5522 for_each_possible_cpu(i) {
5523 snprintf(buf, 32, "cpu%d", i);
5524 entry->procname = kstrdup(buf, GFP_KERNEL);
5525 entry->mode = 0555;
5526 entry->child = sd_alloc_ctl_cpu_table(i);
5527 entry++;
5528 }
5529
5530 WARN_ON(sd_sysctl_header);
5531 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5532}
5533
5534/* may be called multiple times per register */
5535static void unregister_sched_domain_sysctl(void)
5536{
5537 if (sd_sysctl_header)
5538 unregister_sysctl_table(sd_sysctl_header);
5539 sd_sysctl_header = NULL;
5540 if (sd_ctl_dir[0].child)
5541 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5542}
5543#else
5544static void register_sched_domain_sysctl(void)
5545{
5546}
5547static void unregister_sched_domain_sysctl(void)
5548{
5549}
5550#endif
5551
5552static void set_rq_online(struct rq *rq)
5553{
5554 if (!rq->online) {
5555 const struct sched_class *class;
5556
5557 cpumask_set_cpu(rq->cpu, rq->rd->online);
5558 rq->online = 1;
5559
5560 for_each_class(class) {
5561 if (class->rq_online)
5562 class->rq_online(rq);
5563 }
5564 }
5565}
5566
5567static void set_rq_offline(struct rq *rq)
5568{
5569 if (rq->online) {
5570 const struct sched_class *class;
5571
5572 for_each_class(class) {
5573 if (class->rq_offline)
5574 class->rq_offline(rq);
5575 }
5576
5577 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5578 rq->online = 0;
5579 }
5580}
5581
5582/*
5583 * migration_call - callback that gets triggered when a CPU is added.
5584 * Here we can start up the necessary migration thread for the new CPU.
5585 */
5586static int __cpuinit
5587migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5588{
5589 int cpu = (long)hcpu;
5590 unsigned long flags;
5591 struct rq *rq = cpu_rq(cpu);
5592
5593 switch (action & ~CPU_TASKS_FROZEN) {
5594
5595 case CPU_UP_PREPARE:
5596 rq->calc_load_update = calc_load_update;
5597 break;
5598
5599 case CPU_ONLINE:
5600 /* Update our root-domain */
5601 raw_spin_lock_irqsave(&rq->lock, flags);
5602 if (rq->rd) {
5603 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5604
5605 set_rq_online(rq);
5606 }
5607 raw_spin_unlock_irqrestore(&rq->lock, flags);
5608 break;
5609
5610#ifdef CONFIG_HOTPLUG_CPU
5611 case CPU_DYING:
5612 sched_ttwu_pending();
5613 /* Update our root-domain */
5614 raw_spin_lock_irqsave(&rq->lock, flags);
5615 if (rq->rd) {
5616 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5617 set_rq_offline(rq);
5618 }
5619 migrate_tasks(cpu);
5620 BUG_ON(rq->nr_running != 1); /* the migration thread */
5621 raw_spin_unlock_irqrestore(&rq->lock, flags);
5622
5623 migrate_nr_uninterruptible(rq);
5624 calc_global_load_remove(rq);
5625 break;
5626#endif
5627 }
5628
5629 update_max_interval();
5630
5631 return NOTIFY_OK;
5632}
5633
5634/*
5635 * Register at high priority so that task migration (migrate_all_tasks)
5636 * happens before everything else. This has to be lower priority than
5637 * the notifier in the perf_event subsystem, though.
5638 */
5639static struct notifier_block __cpuinitdata migration_notifier = {
5640 .notifier_call = migration_call,
5641 .priority = CPU_PRI_MIGRATION,
5642};
5643
5644static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5645 unsigned long action, void *hcpu)
5646{
5647 switch (action & ~CPU_TASKS_FROZEN) {
5648 case CPU_STARTING:
5649 case CPU_DOWN_FAILED:
5650 set_cpu_active((long)hcpu, true);
5651 return NOTIFY_OK;
5652 default:
5653 return NOTIFY_DONE;
5654 }
5655}
5656
5657static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5658 unsigned long action, void *hcpu)
5659{
5660 switch (action & ~CPU_TASKS_FROZEN) {
5661 case CPU_DOWN_PREPARE:
5662 set_cpu_active((long)hcpu, false);
5663 return NOTIFY_OK;
5664 default:
5665 return NOTIFY_DONE;
5666 }
5667}
5668
5669static int __init migration_init(void)
5670{
5671 void *cpu = (void *)(long)smp_processor_id();
5672 int err;
5673
5674 /* Initialize migration for the boot CPU */
5675 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5676 BUG_ON(err == NOTIFY_BAD);
5677 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5678 register_cpu_notifier(&migration_notifier);
5679
5680 /* Register cpu active notifiers */
5681 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5682 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5683
5684 return 0;
5685}
5686early_initcall(migration_init);
5687#endif
5688
5689#ifdef CONFIG_SMP
5690
5691static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5692
5693#ifdef CONFIG_SCHED_DEBUG
5694
5695static __read_mostly int sched_debug_enabled;
5696
5697static int __init sched_debug_setup(char *str)
5698{
5699 sched_debug_enabled = 1;
5700
5701 return 0;
5702}
5703early_param("sched_debug", sched_debug_setup);
5704
5705static inline bool sched_debug(void)
5706{
5707 return sched_debug_enabled;
5708}
5709
5710static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5711 struct cpumask *groupmask)
5712{
5713 struct sched_group *group = sd->groups;
5714 char str[256];
5715
5716 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5717 cpumask_clear(groupmask);
5718
5719 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5720
5721 if (!(sd->flags & SD_LOAD_BALANCE)) {
5722 printk("does not load-balance\n");
5723 if (sd->parent)
5724 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5725 " has parent");
5726 return -1;
5727 }
5728
5729 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5730
5731 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5732 printk(KERN_ERR "ERROR: domain->span does not contain "
5733 "CPU%d\n", cpu);
5734 }
5735 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5736 printk(KERN_ERR "ERROR: domain->groups does not contain"
5737 " CPU%d\n", cpu);
5738 }
5739
5740 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5741 do {
5742 if (!group) {
5743 printk("\n");
5744 printk(KERN_ERR "ERROR: group is NULL\n");
5745 break;
5746 }
5747
5748 /*
5749 * Even though we initialize ->power to something semi-sane,
5750 * we leave power_orig unset. This allows us to detect if
5751 * domain iteration is still funny without causing /0 traps.
5752 */
5753 if (!group->sgp->power_orig) {
5754 printk(KERN_CONT "\n");
5755 printk(KERN_ERR "ERROR: domain->cpu_power not "
5756 "set\n");
5757 break;
5758 }
5759
5760 if (!cpumask_weight(sched_group_cpus(group))) {
5761 printk(KERN_CONT "\n");
5762 printk(KERN_ERR "ERROR: empty group\n");
5763 break;
5764 }
5765
5766 if (!(sd->flags & SD_OVERLAP) &&
5767 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5768 printk(KERN_CONT "\n");
5769 printk(KERN_ERR "ERROR: repeated CPUs\n");
5770 break;
5771 }
5772
5773 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5774
5775 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5776
5777 printk(KERN_CONT " %s", str);
5778 if (group->sgp->power != SCHED_POWER_SCALE) {
5779 printk(KERN_CONT " (cpu_power = %d)",
5780 group->sgp->power);
5781 }
5782
5783 group = group->next;
5784 } while (group != sd->groups);
5785 printk(KERN_CONT "\n");
5786
5787 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5788 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5789
5790 if (sd->parent &&
5791 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5792 printk(KERN_ERR "ERROR: parent span is not a superset "
5793 "of domain->span\n");
5794 return 0;
5795}
5796
5797static void sched_domain_debug(struct sched_domain *sd, int cpu)
5798{
5799 int level = 0;
5800
5801 if (!sched_debug_enabled)
5802 return;
5803
5804 if (!sd) {
5805 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5806 return;
5807 }
5808
5809 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5810
5811 for (;;) {
5812 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5813 break;
5814 level++;
5815 sd = sd->parent;
5816 if (!sd)
5817 break;
5818 }
5819}
5820#else /* !CONFIG_SCHED_DEBUG */
5821# define sched_domain_debug(sd, cpu) do { } while (0)
5822static inline bool sched_debug(void)
5823{
5824 return false;
5825}
5826#endif /* CONFIG_SCHED_DEBUG */
5827
5828static int sd_degenerate(struct sched_domain *sd)
5829{
5830 if (cpumask_weight(sched_domain_span(sd)) == 1)
5831 return 1;
5832
5833 /* Following flags need at least 2 groups */
5834 if (sd->flags & (SD_LOAD_BALANCE |
5835 SD_BALANCE_NEWIDLE |
5836 SD_BALANCE_FORK |
5837 SD_BALANCE_EXEC |
5838 SD_SHARE_CPUPOWER |
5839 SD_SHARE_PKG_RESOURCES)) {
5840 if (sd->groups != sd->groups->next)
5841 return 0;
5842 }
5843
5844 /* Following flags don't use groups */
5845 if (sd->flags & (SD_WAKE_AFFINE))
5846 return 0;
5847
5848 return 1;
5849}
5850
5851static int
5852sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5853{
5854 unsigned long cflags = sd->flags, pflags = parent->flags;
5855
5856 if (sd_degenerate(parent))
5857 return 1;
5858
5859 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5860 return 0;
5861
5862 /* Flags needing groups don't count if only 1 group in parent */
5863 if (parent->groups == parent->groups->next) {
5864 pflags &= ~(SD_LOAD_BALANCE |
5865 SD_BALANCE_NEWIDLE |
5866 SD_BALANCE_FORK |
5867 SD_BALANCE_EXEC |
5868 SD_SHARE_CPUPOWER |
5869 SD_SHARE_PKG_RESOURCES);
5870 if (nr_node_ids == 1)
5871 pflags &= ~SD_SERIALIZE;
5872 }
5873 if (~cflags & pflags)
5874 return 0;
5875
5876 return 1;
5877}
5878
5879static void free_rootdomain(struct rcu_head *rcu)
5880{
5881 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5882
5883 cpupri_cleanup(&rd->cpupri);
5884 free_cpumask_var(rd->rto_mask);
5885 free_cpumask_var(rd->online);
5886 free_cpumask_var(rd->span);
5887 kfree(rd);
5888}
5889
5890static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5891{
5892 struct root_domain *old_rd = NULL;
5893 unsigned long flags;
5894
5895 raw_spin_lock_irqsave(&rq->lock, flags);
5896
5897 if (rq->rd) {
5898 old_rd = rq->rd;
5899
5900 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5901 set_rq_offline(rq);
5902
5903 cpumask_clear_cpu(rq->cpu, old_rd->span);
5904
5905 /*
5906 * If we dont want to free the old_rt yet then
5907 * set old_rd to NULL to skip the freeing later
5908 * in this function:
5909 */
5910 if (!atomic_dec_and_test(&old_rd->refcount))
5911 old_rd = NULL;
5912 }
5913
5914 atomic_inc(&rd->refcount);
5915 rq->rd = rd;
5916
5917 cpumask_set_cpu(rq->cpu, rd->span);
5918 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5919 set_rq_online(rq);
5920
5921 raw_spin_unlock_irqrestore(&rq->lock, flags);
5922
5923 if (old_rd)
5924 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5925}
5926
5927static int init_rootdomain(struct root_domain *rd)
5928{
5929 memset(rd, 0, sizeof(*rd));
5930
5931 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5932 goto out;
5933 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5934 goto free_span;
5935 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5936 goto free_online;
5937
5938 if (cpupri_init(&rd->cpupri) != 0)
5939 goto free_rto_mask;
5940 return 0;
5941
5942free_rto_mask:
5943 free_cpumask_var(rd->rto_mask);
5944free_online:
5945 free_cpumask_var(rd->online);
5946free_span:
5947 free_cpumask_var(rd->span);
5948out:
5949 return -ENOMEM;
5950}
5951
5952/*
5953 * By default the system creates a single root-domain with all cpus as
5954 * members (mimicking the global state we have today).
5955 */
5956struct root_domain def_root_domain;
5957
5958static void init_defrootdomain(void)
5959{
5960 init_rootdomain(&def_root_domain);
5961
5962 atomic_set(&def_root_domain.refcount, 1);
5963}
5964
5965static struct root_domain *alloc_rootdomain(void)
5966{
5967 struct root_domain *rd;
5968
5969 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5970 if (!rd)
5971 return NULL;
5972
5973 if (init_rootdomain(rd) != 0) {
5974 kfree(rd);
5975 return NULL;
5976 }
5977
5978 return rd;
5979}
5980
5981static void free_sched_groups(struct sched_group *sg, int free_sgp)
5982{
5983 struct sched_group *tmp, *first;
5984
5985 if (!sg)
5986 return;
5987
5988 first = sg;
5989 do {
5990 tmp = sg->next;
5991
5992 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5993 kfree(sg->sgp);
5994
5995 kfree(sg);
5996 sg = tmp;
5997 } while (sg != first);
5998}
5999
6000static void free_sched_domain(struct rcu_head *rcu)
6001{
6002 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6003
6004 /*
6005 * If its an overlapping domain it has private groups, iterate and
6006 * nuke them all.
6007 */
6008 if (sd->flags & SD_OVERLAP) {
6009 free_sched_groups(sd->groups, 1);
6010 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6011 kfree(sd->groups->sgp);
6012 kfree(sd->groups);
6013 }
6014 kfree(sd);
6015}
6016
6017static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6018{
6019 call_rcu(&sd->rcu, free_sched_domain);
6020}
6021
6022static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6023{
6024 for (; sd; sd = sd->parent)
6025 destroy_sched_domain(sd, cpu);
6026}
6027
6028/*
6029 * Keep a special pointer to the highest sched_domain that has
6030 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6031 * allows us to avoid some pointer chasing select_idle_sibling().
6032 *
6033 * Also keep a unique ID per domain (we use the first cpu number in
6034 * the cpumask of the domain), this allows us to quickly tell if
6035 * two cpus are in the same cache domain, see cpus_share_cache().
6036 */
6037DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6038DEFINE_PER_CPU(int, sd_llc_id);
6039
6040static void update_top_cache_domain(int cpu)
6041{
6042 struct sched_domain *sd;
6043 int id = cpu;
6044
6045 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6046 if (sd)
6047 id = cpumask_first(sched_domain_span(sd));
6048
6049 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6050 per_cpu(sd_llc_id, cpu) = id;
6051}
6052
6053/*
6054 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6055 * hold the hotplug lock.
6056 */
6057static void
6058cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6059{
6060 struct rq *rq = cpu_rq(cpu);
6061 struct sched_domain *tmp;
6062
6063 /* Remove the sched domains which do not contribute to scheduling. */
6064 for (tmp = sd; tmp; ) {
6065 struct sched_domain *parent = tmp->parent;
6066 if (!parent)
6067 break;
6068
6069 if (sd_parent_degenerate(tmp, parent)) {
6070 tmp->parent = parent->parent;
6071 if (parent->parent)
6072 parent->parent->child = tmp;
6073 destroy_sched_domain(parent, cpu);
6074 } else
6075 tmp = tmp->parent;
6076 }
6077
6078 if (sd && sd_degenerate(sd)) {
6079 tmp = sd;
6080 sd = sd->parent;
6081 destroy_sched_domain(tmp, cpu);
6082 if (sd)
6083 sd->child = NULL;
6084 }
6085
6086 sched_domain_debug(sd, cpu);
6087
6088 rq_attach_root(rq, rd);
6089 tmp = rq->sd;
6090 rcu_assign_pointer(rq->sd, sd);
6091 destroy_sched_domains(tmp, cpu);
6092
6093 update_top_cache_domain(cpu);
6094}
6095
6096/* cpus with isolated domains */
6097static cpumask_var_t cpu_isolated_map;
6098
6099/* Setup the mask of cpus configured for isolated domains */
6100static int __init isolated_cpu_setup(char *str)
6101{
6102 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6103 cpulist_parse(str, cpu_isolated_map);
6104 return 1;
6105}
6106
6107__setup("isolcpus=", isolated_cpu_setup);
6108
6109static const struct cpumask *cpu_cpu_mask(int cpu)
6110{
6111 return cpumask_of_node(cpu_to_node(cpu));
6112}
6113
6114struct sd_data {
6115 struct sched_domain **__percpu sd;
6116 struct sched_group **__percpu sg;
6117 struct sched_group_power **__percpu sgp;
6118};
6119
6120struct s_data {
6121 struct sched_domain ** __percpu sd;
6122 struct root_domain *rd;
6123};
6124
6125enum s_alloc {
6126 sa_rootdomain,
6127 sa_sd,
6128 sa_sd_storage,
6129 sa_none,
6130};
6131
6132struct sched_domain_topology_level;
6133
6134typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
6135typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
6136
6137#define SDTL_OVERLAP 0x01
6138
6139struct sched_domain_topology_level {
6140 sched_domain_init_f init;
6141 sched_domain_mask_f mask;
6142 int flags;
6143 int numa_level;
6144 struct sd_data data;
6145};
6146
6147/*
6148 * Build an iteration mask that can exclude certain CPUs from the upwards
6149 * domain traversal.
6150 *
6151 * Asymmetric node setups can result in situations where the domain tree is of
6152 * unequal depth, make sure to skip domains that already cover the entire
6153 * range.
6154 *
6155 * In that case build_sched_domains() will have terminated the iteration early
6156 * and our sibling sd spans will be empty. Domains should always include the
6157 * cpu they're built on, so check that.
6158 *
6159 */
6160static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6161{
6162 const struct cpumask *span = sched_domain_span(sd);
6163 struct sd_data *sdd = sd->private;
6164 struct sched_domain *sibling;
6165 int i;
6166
6167 for_each_cpu(i, span) {
6168 sibling = *per_cpu_ptr(sdd->sd, i);
6169 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6170 continue;
6171
6172 cpumask_set_cpu(i, sched_group_mask(sg));
6173 }
6174}
6175
6176/*
6177 * Return the canonical balance cpu for this group, this is the first cpu
6178 * of this group that's also in the iteration mask.
6179 */
6180int group_balance_cpu(struct sched_group *sg)
6181{
6182 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6183}
6184
6185static int
6186build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6187{
6188 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6189 const struct cpumask *span = sched_domain_span(sd);
6190 struct cpumask *covered = sched_domains_tmpmask;
6191 struct sd_data *sdd = sd->private;
6192 struct sched_domain *child;
6193 int i;
6194
6195 cpumask_clear(covered);
6196
6197 for_each_cpu(i, span) {
6198 struct cpumask *sg_span;
6199
6200 if (cpumask_test_cpu(i, covered))
6201 continue;
6202
6203 child = *per_cpu_ptr(sdd->sd, i);
6204
6205 /* See the comment near build_group_mask(). */
6206 if (!cpumask_test_cpu(i, sched_domain_span(child)))
6207 continue;
6208
6209 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6210 GFP_KERNEL, cpu_to_node(cpu));
6211
6212 if (!sg)
6213 goto fail;
6214
6215 sg_span = sched_group_cpus(sg);
6216 if (child->child) {
6217 child = child->child;
6218 cpumask_copy(sg_span, sched_domain_span(child));
6219 } else
6220 cpumask_set_cpu(i, sg_span);
6221
6222 cpumask_or(covered, covered, sg_span);
6223
6224 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
6225 if (atomic_inc_return(&sg->sgp->ref) == 1)
6226 build_group_mask(sd, sg);
6227
6228 /*
6229 * Initialize sgp->power such that even if we mess up the
6230 * domains and no possible iteration will get us here, we won't
6231 * die on a /0 trap.
6232 */
6233 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
6234
6235 /*
6236 * Make sure the first group of this domain contains the
6237 * canonical balance cpu. Otherwise the sched_domain iteration
6238 * breaks. See update_sg_lb_stats().
6239 */
6240 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6241 group_balance_cpu(sg) == cpu)
6242 groups = sg;
6243
6244 if (!first)
6245 first = sg;
6246 if (last)
6247 last->next = sg;
6248 last = sg;
6249 last->next = first;
6250 }
6251 sd->groups = groups;
6252
6253 return 0;
6254
6255fail:
6256 free_sched_groups(first, 0);
6257
6258 return -ENOMEM;
6259}
6260
6261static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6262{
6263 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6264 struct sched_domain *child = sd->child;
6265
6266 if (child)
6267 cpu = cpumask_first(sched_domain_span(child));
6268
6269 if (sg) {
6270 *sg = *per_cpu_ptr(sdd->sg, cpu);
6271 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
6272 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
6273 }
6274
6275 return cpu;
6276}
6277
6278/*
6279 * build_sched_groups will build a circular linked list of the groups
6280 * covered by the given span, and will set each group's ->cpumask correctly,
6281 * and ->cpu_power to 0.
6282 *
6283 * Assumes the sched_domain tree is fully constructed
6284 */
6285static int
6286build_sched_groups(struct sched_domain *sd, int cpu)
6287{
6288 struct sched_group *first = NULL, *last = NULL;
6289 struct sd_data *sdd = sd->private;
6290 const struct cpumask *span = sched_domain_span(sd);
6291 struct cpumask *covered;
6292 int i;
6293
6294 get_group(cpu, sdd, &sd->groups);
6295 atomic_inc(&sd->groups->ref);
6296
6297 if (cpu != cpumask_first(sched_domain_span(sd)))
6298 return 0;
6299
6300 lockdep_assert_held(&sched_domains_mutex);
6301 covered = sched_domains_tmpmask;
6302
6303 cpumask_clear(covered);
6304
6305 for_each_cpu(i, span) {
6306 struct sched_group *sg;
6307 int group = get_group(i, sdd, &sg);
6308 int j;
6309
6310 if (cpumask_test_cpu(i, covered))
6311 continue;
6312
6313 cpumask_clear(sched_group_cpus(sg));
6314 sg->sgp->power = 0;
6315 cpumask_setall(sched_group_mask(sg));
6316
6317 for_each_cpu(j, span) {
6318 if (get_group(j, sdd, NULL) != group)
6319 continue;
6320
6321 cpumask_set_cpu(j, covered);
6322 cpumask_set_cpu(j, sched_group_cpus(sg));
6323 }
6324
6325 if (!first)
6326 first = sg;
6327 if (last)
6328 last->next = sg;
6329 last = sg;
6330 }
6331 last->next = first;
6332
6333 return 0;
6334}
6335
6336/*
6337 * Initialize sched groups cpu_power.
6338 *
6339 * cpu_power indicates the capacity of sched group, which is used while
6340 * distributing the load between different sched groups in a sched domain.
6341 * Typically cpu_power for all the groups in a sched domain will be same unless
6342 * there are asymmetries in the topology. If there are asymmetries, group
6343 * having more cpu_power will pickup more load compared to the group having
6344 * less cpu_power.
6345 */
6346static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6347{
6348 struct sched_group *sg = sd->groups;
6349
6350 WARN_ON(!sd || !sg);
6351
6352 do {
6353 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6354 sg = sg->next;
6355 } while (sg != sd->groups);
6356
6357 if (cpu != group_balance_cpu(sg))
6358 return;
6359
6360 update_group_power(sd, cpu);
6361 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6362}
6363
6364int __weak arch_sd_sibling_asym_packing(void)
6365{
6366 return 0*SD_ASYM_PACKING;
6367}
6368
6369/*
6370 * Initializers for schedule domains
6371 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6372 */
6373
6374#ifdef CONFIG_SCHED_DEBUG
6375# define SD_INIT_NAME(sd, type) sd->name = #type
6376#else
6377# define SD_INIT_NAME(sd, type) do { } while (0)
6378#endif
6379
6380#define SD_INIT_FUNC(type) \
6381static noinline struct sched_domain * \
6382sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6383{ \
6384 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6385 *sd = SD_##type##_INIT; \
6386 SD_INIT_NAME(sd, type); \
6387 sd->private = &tl->data; \
6388 return sd; \
6389}
6390
6391SD_INIT_FUNC(CPU)
6392#ifdef CONFIG_SCHED_SMT
6393 SD_INIT_FUNC(SIBLING)
6394#endif
6395#ifdef CONFIG_SCHED_MC
6396 SD_INIT_FUNC(MC)
6397#endif
6398#ifdef CONFIG_SCHED_BOOK
6399 SD_INIT_FUNC(BOOK)
6400#endif
6401
6402static int default_relax_domain_level = -1;
6403int sched_domain_level_max;
6404
6405static int __init setup_relax_domain_level(char *str)
6406{
6407 if (kstrtoint(str, 0, &default_relax_domain_level))
6408 pr_warn("Unable to set relax_domain_level\n");
6409
6410 return 1;
6411}
6412__setup("relax_domain_level=", setup_relax_domain_level);
6413
6414static void set_domain_attribute(struct sched_domain *sd,
6415 struct sched_domain_attr *attr)
6416{
6417 int request;
6418
6419 if (!attr || attr->relax_domain_level < 0) {
6420 if (default_relax_domain_level < 0)
6421 return;
6422 else
6423 request = default_relax_domain_level;
6424 } else
6425 request = attr->relax_domain_level;
6426 if (request < sd->level) {
6427 /* turn off idle balance on this domain */
6428 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6429 } else {
6430 /* turn on idle balance on this domain */
6431 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6432 }
6433}
6434
6435static void __sdt_free(const struct cpumask *cpu_map);
6436static int __sdt_alloc(const struct cpumask *cpu_map);
6437
6438static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6439 const struct cpumask *cpu_map)
6440{
6441 switch (what) {
6442 case sa_rootdomain:
6443 if (!atomic_read(&d->rd->refcount))
6444 free_rootdomain(&d->rd->rcu); /* fall through */
6445 case sa_sd:
6446 free_percpu(d->sd); /* fall through */
6447 case sa_sd_storage:
6448 __sdt_free(cpu_map); /* fall through */
6449 case sa_none:
6450 break;
6451 }
6452}
6453
6454static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6455 const struct cpumask *cpu_map)
6456{
6457 memset(d, 0, sizeof(*d));
6458
6459 if (__sdt_alloc(cpu_map))
6460 return sa_sd_storage;
6461 d->sd = alloc_percpu(struct sched_domain *);
6462 if (!d->sd)
6463 return sa_sd_storage;
6464 d->rd = alloc_rootdomain();
6465 if (!d->rd)
6466 return sa_sd;
6467 return sa_rootdomain;
6468}
6469
6470/*
6471 * NULL the sd_data elements we've used to build the sched_domain and
6472 * sched_group structure so that the subsequent __free_domain_allocs()
6473 * will not free the data we're using.
6474 */
6475static void claim_allocations(int cpu, struct sched_domain *sd)
6476{
6477 struct sd_data *sdd = sd->private;
6478
6479 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6480 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6481
6482 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6483 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6484
6485 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6486 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6487}
6488
6489#ifdef CONFIG_SCHED_SMT
6490static const struct cpumask *cpu_smt_mask(int cpu)
6491{
6492 return topology_thread_cpumask(cpu);
6493}
6494#endif
6495
6496/*
6497 * Topology list, bottom-up.
6498 */
6499static struct sched_domain_topology_level default_topology[] = {
6500#ifdef CONFIG_SCHED_SMT
6501 { sd_init_SIBLING, cpu_smt_mask, },
6502#endif
6503#ifdef CONFIG_SCHED_MC
6504 { sd_init_MC, cpu_coregroup_mask, },
6505#endif
6506#ifdef CONFIG_SCHED_BOOK
6507 { sd_init_BOOK, cpu_book_mask, },
6508#endif
6509 { sd_init_CPU, cpu_cpu_mask, },
6510 { NULL, },
6511};
6512
6513static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6514
6515#ifdef CONFIG_NUMA
6516
6517static int sched_domains_numa_levels;
6518static int *sched_domains_numa_distance;
6519static struct cpumask ***sched_domains_numa_masks;
6520static int sched_domains_curr_level;
6521
6522static inline int sd_local_flags(int level)
6523{
6524 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6525 return 0;
6526
6527 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6528}
6529
6530static struct sched_domain *
6531sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6532{
6533 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6534 int level = tl->numa_level;
6535 int sd_weight = cpumask_weight(
6536 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6537
6538 *sd = (struct sched_domain){
6539 .min_interval = sd_weight,
6540 .max_interval = 2*sd_weight,
6541 .busy_factor = 32,
6542 .imbalance_pct = 125,
6543 .cache_nice_tries = 2,
6544 .busy_idx = 3,
6545 .idle_idx = 2,
6546 .newidle_idx = 0,
6547 .wake_idx = 0,
6548 .forkexec_idx = 0,
6549
6550 .flags = 1*SD_LOAD_BALANCE
6551 | 1*SD_BALANCE_NEWIDLE
6552 | 0*SD_BALANCE_EXEC
6553 | 0*SD_BALANCE_FORK
6554 | 0*SD_BALANCE_WAKE
6555 | 0*SD_WAKE_AFFINE
6556 | 0*SD_PREFER_LOCAL
6557 | 0*SD_SHARE_CPUPOWER
6558 | 0*SD_SHARE_PKG_RESOURCES
6559 | 1*SD_SERIALIZE
6560 | 0*SD_PREFER_SIBLING
6561 | sd_local_flags(level)
6562 ,
6563 .last_balance = jiffies,
6564 .balance_interval = sd_weight,
6565 };
6566 SD_INIT_NAME(sd, NUMA);
6567 sd->private = &tl->data;
6568
6569 /*
6570 * Ugly hack to pass state to sd_numa_mask()...
6571 */
6572 sched_domains_curr_level = tl->numa_level;
6573
6574 return sd;
6575}
6576
6577static const struct cpumask *sd_numa_mask(int cpu)
6578{
6579 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6580}
6581
6582static void sched_numa_warn(const char *str)
6583{
6584 static int done = false;
6585 int i,j;
6586
6587 if (done)
6588 return;
6589
6590 done = true;
6591
6592 printk(KERN_WARNING "ERROR: %s\n\n", str);
6593
6594 for (i = 0; i < nr_node_ids; i++) {
6595 printk(KERN_WARNING " ");
6596 for (j = 0; j < nr_node_ids; j++)
6597 printk(KERN_CONT "%02d ", node_distance(i,j));
6598 printk(KERN_CONT "\n");
6599 }
6600 printk(KERN_WARNING "\n");
6601}
6602
6603static bool find_numa_distance(int distance)
6604{
6605 int i;
6606
6607 if (distance == node_distance(0, 0))
6608 return true;
6609
6610 for (i = 0; i < sched_domains_numa_levels; i++) {
6611 if (sched_domains_numa_distance[i] == distance)
6612 return true;
6613 }
6614
6615 return false;
6616}
6617
6618static void sched_init_numa(void)
6619{
6620 int next_distance, curr_distance = node_distance(0, 0);
6621 struct sched_domain_topology_level *tl;
6622 int level = 0;
6623 int i, j, k;
6624
6625 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6626 if (!sched_domains_numa_distance)
6627 return;
6628
6629 /*
6630 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6631 * unique distances in the node_distance() table.
6632 *
6633 * Assumes node_distance(0,j) includes all distances in
6634 * node_distance(i,j) in order to avoid cubic time.
6635 */
6636 next_distance = curr_distance;
6637 for (i = 0; i < nr_node_ids; i++) {
6638 for (j = 0; j < nr_node_ids; j++) {
6639 for (k = 0; k < nr_node_ids; k++) {
6640 int distance = node_distance(i, k);
6641
6642 if (distance > curr_distance &&
6643 (distance < next_distance ||
6644 next_distance == curr_distance))
6645 next_distance = distance;
6646
6647 /*
6648 * While not a strong assumption it would be nice to know
6649 * about cases where if node A is connected to B, B is not
6650 * equally connected to A.
6651 */
6652 if (sched_debug() && node_distance(k, i) != distance)
6653 sched_numa_warn("Node-distance not symmetric");
6654
6655 if (sched_debug() && i && !find_numa_distance(distance))
6656 sched_numa_warn("Node-0 not representative");
6657 }
6658 if (next_distance != curr_distance) {
6659 sched_domains_numa_distance[level++] = next_distance;
6660 sched_domains_numa_levels = level;
6661 curr_distance = next_distance;
6662 } else break;
6663 }
6664
6665 /*
6666 * In case of sched_debug() we verify the above assumption.
6667 */
6668 if (!sched_debug())
6669 break;
6670 }
6671 /*
6672 * 'level' contains the number of unique distances, excluding the
6673 * identity distance node_distance(i,i).
6674 *
6675 * The sched_domains_nume_distance[] array includes the actual distance
6676 * numbers.
6677 */
6678
6679 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6680 if (!sched_domains_numa_masks)
6681 return;
6682
6683 /*
6684 * Now for each level, construct a mask per node which contains all
6685 * cpus of nodes that are that many hops away from us.
6686 */
6687 for (i = 0; i < level; i++) {
6688 sched_domains_numa_masks[i] =
6689 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6690 if (!sched_domains_numa_masks[i])
6691 return;
6692
6693 for (j = 0; j < nr_node_ids; j++) {
6694 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6695 if (!mask)
6696 return;
6697
6698 sched_domains_numa_masks[i][j] = mask;
6699
6700 for (k = 0; k < nr_node_ids; k++) {
6701 if (node_distance(j, k) > sched_domains_numa_distance[i])
6702 continue;
6703
6704 cpumask_or(mask, mask, cpumask_of_node(k));
6705 }
6706 }
6707 }
6708
6709 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6710 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6711 if (!tl)
6712 return;
6713
6714 /*
6715 * Copy the default topology bits..
6716 */
6717 for (i = 0; default_topology[i].init; i++)
6718 tl[i] = default_topology[i];
6719
6720 /*
6721 * .. and append 'j' levels of NUMA goodness.
6722 */
6723 for (j = 0; j < level; i++, j++) {
6724 tl[i] = (struct sched_domain_topology_level){
6725 .init = sd_numa_init,
6726 .mask = sd_numa_mask,
6727 .flags = SDTL_OVERLAP,
6728 .numa_level = j,
6729 };
6730 }
6731
6732 sched_domain_topology = tl;
6733}
6734#else
6735static inline void sched_init_numa(void)
6736{
6737}
6738#endif /* CONFIG_NUMA */
6739
6740static int __sdt_alloc(const struct cpumask *cpu_map)
6741{
6742 struct sched_domain_topology_level *tl;
6743 int j;
6744
6745 for (tl = sched_domain_topology; tl->init; tl++) {
6746 struct sd_data *sdd = &tl->data;
6747
6748 sdd->sd = alloc_percpu(struct sched_domain *);
6749 if (!sdd->sd)
6750 return -ENOMEM;
6751
6752 sdd->sg = alloc_percpu(struct sched_group *);
6753 if (!sdd->sg)
6754 return -ENOMEM;
6755
6756 sdd->sgp = alloc_percpu(struct sched_group_power *);
6757 if (!sdd->sgp)
6758 return -ENOMEM;
6759
6760 for_each_cpu(j, cpu_map) {
6761 struct sched_domain *sd;
6762 struct sched_group *sg;
6763 struct sched_group_power *sgp;
6764
6765 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6766 GFP_KERNEL, cpu_to_node(j));
6767 if (!sd)
6768 return -ENOMEM;
6769
6770 *per_cpu_ptr(sdd->sd, j) = sd;
6771
6772 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6773 GFP_KERNEL, cpu_to_node(j));
6774 if (!sg)
6775 return -ENOMEM;
6776
6777 sg->next = sg;
6778
6779 *per_cpu_ptr(sdd->sg, j) = sg;
6780
6781 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6782 GFP_KERNEL, cpu_to_node(j));
6783 if (!sgp)
6784 return -ENOMEM;
6785
6786 *per_cpu_ptr(sdd->sgp, j) = sgp;
6787 }
6788 }
6789
6790 return 0;
6791}
6792
6793static void __sdt_free(const struct cpumask *cpu_map)
6794{
6795 struct sched_domain_topology_level *tl;
6796 int j;
6797
6798 for (tl = sched_domain_topology; tl->init; tl++) {
6799 struct sd_data *sdd = &tl->data;
6800
6801 for_each_cpu(j, cpu_map) {
6802 struct sched_domain *sd;
6803
6804 if (sdd->sd) {
6805 sd = *per_cpu_ptr(sdd->sd, j);
6806 if (sd && (sd->flags & SD_OVERLAP))
6807 free_sched_groups(sd->groups, 0);
6808 kfree(*per_cpu_ptr(sdd->sd, j));
6809 }
6810
6811 if (sdd->sg)
6812 kfree(*per_cpu_ptr(sdd->sg, j));
6813 if (sdd->sgp)
6814 kfree(*per_cpu_ptr(sdd->sgp, j));
6815 }
6816 free_percpu(sdd->sd);
6817 sdd->sd = NULL;
6818 free_percpu(sdd->sg);
6819 sdd->sg = NULL;
6820 free_percpu(sdd->sgp);
6821 sdd->sgp = NULL;
6822 }
6823}
6824
6825struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6826 struct s_data *d, const struct cpumask *cpu_map,
6827 struct sched_domain_attr *attr, struct sched_domain *child,
6828 int cpu)
6829{
6830 struct sched_domain *sd = tl->init(tl, cpu);
6831 if (!sd)
6832 return child;
6833
6834 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6835 if (child) {
6836 sd->level = child->level + 1;
6837 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6838 child->parent = sd;
6839 }
6840 sd->child = child;
6841 set_domain_attribute(sd, attr);
6842
6843 return sd;
6844}
6845
6846/*
6847 * Build sched domains for a given set of cpus and attach the sched domains
6848 * to the individual cpus
6849 */
6850static int build_sched_domains(const struct cpumask *cpu_map,
6851 struct sched_domain_attr *attr)
6852{
6853 enum s_alloc alloc_state = sa_none;
6854 struct sched_domain *sd;
6855 struct s_data d;
6856 int i, ret = -ENOMEM;
6857
6858 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6859 if (alloc_state != sa_rootdomain)
6860 goto error;
6861
6862 /* Set up domains for cpus specified by the cpu_map. */
6863 for_each_cpu(i, cpu_map) {
6864 struct sched_domain_topology_level *tl;
6865
6866 sd = NULL;
6867 for (tl = sched_domain_topology; tl->init; tl++) {
6868 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6869 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6870 sd->flags |= SD_OVERLAP;
6871 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6872 break;
6873 }
6874
6875 while (sd->child)
6876 sd = sd->child;
6877
6878 *per_cpu_ptr(d.sd, i) = sd;
6879 }
6880
6881 /* Build the groups for the domains */
6882 for_each_cpu(i, cpu_map) {
6883 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6884 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6885 if (sd->flags & SD_OVERLAP) {
6886 if (build_overlap_sched_groups(sd, i))
6887 goto error;
6888 } else {
6889 if (build_sched_groups(sd, i))
6890 goto error;
6891 }
6892 }
6893 }
6894
6895 /* Calculate CPU power for physical packages and nodes */
6896 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6897 if (!cpumask_test_cpu(i, cpu_map))
6898 continue;
6899
6900 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6901 claim_allocations(i, sd);
6902 init_sched_groups_power(i, sd);
6903 }
6904 }
6905
6906 /* Attach the domains */
6907 rcu_read_lock();
6908 for_each_cpu(i, cpu_map) {
6909 sd = *per_cpu_ptr(d.sd, i);
6910 cpu_attach_domain(sd, d.rd, i);
6911 }
6912 rcu_read_unlock();
6913
6914 ret = 0;
6915error:
6916 __free_domain_allocs(&d, alloc_state, cpu_map);
6917 return ret;
6918}
6919
6920static cpumask_var_t *doms_cur; /* current sched domains */
6921static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6922static struct sched_domain_attr *dattr_cur;
6923 /* attribues of custom domains in 'doms_cur' */
6924
6925/*
6926 * Special case: If a kmalloc of a doms_cur partition (array of
6927 * cpumask) fails, then fallback to a single sched domain,
6928 * as determined by the single cpumask fallback_doms.
6929 */
6930static cpumask_var_t fallback_doms;
6931
6932/*
6933 * arch_update_cpu_topology lets virtualized architectures update the
6934 * cpu core maps. It is supposed to return 1 if the topology changed
6935 * or 0 if it stayed the same.
6936 */
6937int __attribute__((weak)) arch_update_cpu_topology(void)
6938{
6939 return 0;
6940}
6941
6942cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6943{
6944 int i;
6945 cpumask_var_t *doms;
6946
6947 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6948 if (!doms)
6949 return NULL;
6950 for (i = 0; i < ndoms; i++) {
6951 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6952 free_sched_domains(doms, i);
6953 return NULL;
6954 }
6955 }
6956 return doms;
6957}
6958
6959void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6960{
6961 unsigned int i;
6962 for (i = 0; i < ndoms; i++)
6963 free_cpumask_var(doms[i]);
6964 kfree(doms);
6965}
6966
6967/*
6968 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6969 * For now this just excludes isolated cpus, but could be used to
6970 * exclude other special cases in the future.
6971 */
6972static int init_sched_domains(const struct cpumask *cpu_map)
6973{
6974 int err;
6975
6976 arch_update_cpu_topology();
6977 ndoms_cur = 1;
6978 doms_cur = alloc_sched_domains(ndoms_cur);
6979 if (!doms_cur)
6980 doms_cur = &fallback_doms;
6981 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6982 err = build_sched_domains(doms_cur[0], NULL);
6983 register_sched_domain_sysctl();
6984
6985 return err;
6986}
6987
6988/*
6989 * Detach sched domains from a group of cpus specified in cpu_map
6990 * These cpus will now be attached to the NULL domain
6991 */
6992static void detach_destroy_domains(const struct cpumask *cpu_map)
6993{
6994 int i;
6995
6996 rcu_read_lock();
6997 for_each_cpu(i, cpu_map)
6998 cpu_attach_domain(NULL, &def_root_domain, i);
6999 rcu_read_unlock();
7000}
7001
7002/* handle null as "default" */
7003static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7004 struct sched_domain_attr *new, int idx_new)
7005{
7006 struct sched_domain_attr tmp;
7007
7008 /* fast path */
7009 if (!new && !cur)
7010 return 1;
7011
7012 tmp = SD_ATTR_INIT;
7013 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7014 new ? (new + idx_new) : &tmp,
7015 sizeof(struct sched_domain_attr));
7016}
7017
7018/*
7019 * Partition sched domains as specified by the 'ndoms_new'
7020 * cpumasks in the array doms_new[] of cpumasks. This compares
7021 * doms_new[] to the current sched domain partitioning, doms_cur[].
7022 * It destroys each deleted domain and builds each new domain.
7023 *
7024 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7025 * The masks don't intersect (don't overlap.) We should setup one
7026 * sched domain for each mask. CPUs not in any of the cpumasks will
7027 * not be load balanced. If the same cpumask appears both in the
7028 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7029 * it as it is.
7030 *
7031 * The passed in 'doms_new' should be allocated using
7032 * alloc_sched_domains. This routine takes ownership of it and will
7033 * free_sched_domains it when done with it. If the caller failed the
7034 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7035 * and partition_sched_domains() will fallback to the single partition
7036 * 'fallback_doms', it also forces the domains to be rebuilt.
7037 *
7038 * If doms_new == NULL it will be replaced with cpu_online_mask.
7039 * ndoms_new == 0 is a special case for destroying existing domains,
7040 * and it will not create the default domain.
7041 *
7042 * Call with hotplug lock held
7043 */
7044void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7045 struct sched_domain_attr *dattr_new)
7046{
7047 int i, j, n;
7048 int new_topology;
7049
7050 mutex_lock(&sched_domains_mutex);
7051
7052 /* always unregister in case we don't destroy any domains */
7053 unregister_sched_domain_sysctl();
7054
7055 /* Let architecture update cpu core mappings. */
7056 new_topology = arch_update_cpu_topology();
7057
7058 n = doms_new ? ndoms_new : 0;
7059
7060 /* Destroy deleted domains */
7061 for (i = 0; i < ndoms_cur; i++) {
7062 for (j = 0; j < n && !new_topology; j++) {
7063 if (cpumask_equal(doms_cur[i], doms_new[j])
7064 && dattrs_equal(dattr_cur, i, dattr_new, j))
7065 goto match1;
7066 }
7067 /* no match - a current sched domain not in new doms_new[] */
7068 detach_destroy_domains(doms_cur[i]);
7069match1:
7070 ;
7071 }
7072
7073 if (doms_new == NULL) {
7074 ndoms_cur = 0;
7075 doms_new = &fallback_doms;
7076 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7077 WARN_ON_ONCE(dattr_new);
7078 }
7079
7080 /* Build new domains */
7081 for (i = 0; i < ndoms_new; i++) {
7082 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7083 if (cpumask_equal(doms_new[i], doms_cur[j])
7084 && dattrs_equal(dattr_new, i, dattr_cur, j))
7085 goto match2;
7086 }
7087 /* no match - add a new doms_new */
7088 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7089match2:
7090 ;
7091 }
7092
7093 /* Remember the new sched domains */
7094 if (doms_cur != &fallback_doms)
7095 free_sched_domains(doms_cur, ndoms_cur);
7096 kfree(dattr_cur); /* kfree(NULL) is safe */
7097 doms_cur = doms_new;
7098 dattr_cur = dattr_new;
7099 ndoms_cur = ndoms_new;
7100
7101 register_sched_domain_sysctl();
7102
7103 mutex_unlock(&sched_domains_mutex);
7104}
7105
7106/*
7107 * Update cpusets according to cpu_active mask. If cpusets are
7108 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7109 * around partition_sched_domains().
7110 */
7111static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7112 void *hcpu)
7113{
7114 switch (action & ~CPU_TASKS_FROZEN) {
7115 case CPU_ONLINE:
7116 case CPU_DOWN_FAILED:
7117 cpuset_update_active_cpus();
7118 return NOTIFY_OK;
7119 default:
7120 return NOTIFY_DONE;
7121 }
7122}
7123
7124static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7125 void *hcpu)
7126{
7127 switch (action & ~CPU_TASKS_FROZEN) {
7128 case CPU_DOWN_PREPARE:
7129 cpuset_update_active_cpus();
7130 return NOTIFY_OK;
7131 default:
7132 return NOTIFY_DONE;
7133 }
7134}
7135
7136void __init sched_init_smp(void)
7137{
7138 cpumask_var_t non_isolated_cpus;
7139
7140 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7141 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7142
7143 sched_init_numa();
7144
7145 get_online_cpus();
7146 mutex_lock(&sched_domains_mutex);
7147 init_sched_domains(cpu_active_mask);
7148 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7149 if (cpumask_empty(non_isolated_cpus))
7150 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7151 mutex_unlock(&sched_domains_mutex);
7152 put_online_cpus();
7153
7154 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7155 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7156
7157 /* RT runtime code needs to handle some hotplug events */
7158 hotcpu_notifier(update_runtime, 0);
7159
7160 init_hrtick();
7161
7162 /* Move init over to a non-isolated CPU */
7163 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7164 BUG();
7165 sched_init_granularity();
7166 free_cpumask_var(non_isolated_cpus);
7167
7168 init_sched_rt_class();
7169}
7170#else
7171void __init sched_init_smp(void)
7172{
7173 sched_init_granularity();
7174}
7175#endif /* CONFIG_SMP */
7176
7177const_debug unsigned int sysctl_timer_migration = 1;
7178
7179int in_sched_functions(unsigned long addr)
7180{
7181 return in_lock_functions(addr) ||
7182 (addr >= (unsigned long)__sched_text_start
7183 && addr < (unsigned long)__sched_text_end);
7184}
7185
7186#ifdef CONFIG_CGROUP_SCHED
7187struct task_group root_task_group;
7188LIST_HEAD(task_groups);
7189#endif
7190
7191DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
7192
7193void __init sched_init(void)
7194{
7195 int i, j;
7196 unsigned long alloc_size = 0, ptr;
7197
7198#ifdef CONFIG_FAIR_GROUP_SCHED
7199 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7200#endif
7201#ifdef CONFIG_RT_GROUP_SCHED
7202 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7203#endif
7204#ifdef CONFIG_CPUMASK_OFFSTACK
7205 alloc_size += num_possible_cpus() * cpumask_size();
7206#endif
7207 if (alloc_size) {
7208 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7209
7210#ifdef CONFIG_FAIR_GROUP_SCHED
7211 root_task_group.se = (struct sched_entity **)ptr;
7212 ptr += nr_cpu_ids * sizeof(void **);
7213
7214 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7215 ptr += nr_cpu_ids * sizeof(void **);
7216
7217#endif /* CONFIG_FAIR_GROUP_SCHED */
7218#ifdef CONFIG_RT_GROUP_SCHED
7219 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7220 ptr += nr_cpu_ids * sizeof(void **);
7221
7222 root_task_group.rt_rq = (struct rt_rq **)ptr;
7223 ptr += nr_cpu_ids * sizeof(void **);
7224
7225#endif /* CONFIG_RT_GROUP_SCHED */
7226#ifdef CONFIG_CPUMASK_OFFSTACK
7227 for_each_possible_cpu(i) {
7228 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7229 ptr += cpumask_size();
7230 }
7231#endif /* CONFIG_CPUMASK_OFFSTACK */
7232 }
7233
7234#ifdef CONFIG_SMP
7235 init_defrootdomain();
7236#endif
7237
7238 init_rt_bandwidth(&def_rt_bandwidth,
7239 global_rt_period(), global_rt_runtime());
7240
7241#ifdef CONFIG_RT_GROUP_SCHED
7242 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7243 global_rt_period(), global_rt_runtime());
7244#endif /* CONFIG_RT_GROUP_SCHED */
7245
7246#ifdef CONFIG_CGROUP_SCHED
7247 list_add(&root_task_group.list, &task_groups);
7248 INIT_LIST_HEAD(&root_task_group.children);
7249 INIT_LIST_HEAD(&root_task_group.siblings);
7250 autogroup_init(&init_task);
7251
7252#endif /* CONFIG_CGROUP_SCHED */
7253
7254#ifdef CONFIG_CGROUP_CPUACCT
7255 root_cpuacct.cpustat = &kernel_cpustat;
7256 root_cpuacct.cpuusage = alloc_percpu(u64);
7257 /* Too early, not expected to fail */
7258 BUG_ON(!root_cpuacct.cpuusage);
7259#endif
7260 for_each_possible_cpu(i) {
7261 struct rq *rq;
7262
7263 rq = cpu_rq(i);
7264 raw_spin_lock_init(&rq->lock);
7265 rq->nr_running = 0;
7266 rq->calc_load_active = 0;
7267 rq->calc_load_update = jiffies + LOAD_FREQ;
7268 init_cfs_rq(&rq->cfs);
7269 init_rt_rq(&rq->rt, rq);
7270#ifdef CONFIG_FAIR_GROUP_SCHED
7271 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7272 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7273 /*
7274 * How much cpu bandwidth does root_task_group get?
7275 *
7276 * In case of task-groups formed thr' the cgroup filesystem, it
7277 * gets 100% of the cpu resources in the system. This overall
7278 * system cpu resource is divided among the tasks of
7279 * root_task_group and its child task-groups in a fair manner,
7280 * based on each entity's (task or task-group's) weight
7281 * (se->load.weight).
7282 *
7283 * In other words, if root_task_group has 10 tasks of weight
7284 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7285 * then A0's share of the cpu resource is:
7286 *
7287 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7288 *
7289 * We achieve this by letting root_task_group's tasks sit
7290 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7291 */
7292 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7293 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7294#endif /* CONFIG_FAIR_GROUP_SCHED */
7295
7296 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7297#ifdef CONFIG_RT_GROUP_SCHED
7298 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7299 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7300#endif
7301
7302 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7303 rq->cpu_load[j] = 0;
7304
7305 rq->last_load_update_tick = jiffies;
7306
7307#ifdef CONFIG_SMP
7308 rq->sd = NULL;
7309 rq->rd = NULL;
7310 rq->cpu_power = SCHED_POWER_SCALE;
7311 rq->post_schedule = 0;
7312 rq->active_balance = 0;
7313 rq->next_balance = jiffies;
7314 rq->push_cpu = 0;
7315 rq->cpu = i;
7316 rq->online = 0;
7317 rq->idle_stamp = 0;
7318 rq->avg_idle = 2*sysctl_sched_migration_cost;
7319
7320 INIT_LIST_HEAD(&rq->cfs_tasks);
7321
7322 rq_attach_root(rq, &def_root_domain);
7323#ifdef CONFIG_NO_HZ
7324 rq->nohz_flags = 0;
7325#endif
7326#endif
7327 init_rq_hrtick(rq);
7328 atomic_set(&rq->nr_iowait, 0);
7329 }
7330
7331 set_load_weight(&init_task);
7332
7333#ifdef CONFIG_PREEMPT_NOTIFIERS
7334 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7335#endif
7336
7337#ifdef CONFIG_RT_MUTEXES
7338 plist_head_init(&init_task.pi_waiters);
7339#endif
7340
7341 /*
7342 * The boot idle thread does lazy MMU switching as well:
7343 */
7344 atomic_inc(&init_mm.mm_count);
7345 enter_lazy_tlb(&init_mm, current);
7346
7347 /*
7348 * Make us the idle thread. Technically, schedule() should not be
7349 * called from this thread, however somewhere below it might be,
7350 * but because we are the idle thread, we just pick up running again
7351 * when this runqueue becomes "idle".
7352 */
7353 init_idle(current, smp_processor_id());
7354
7355 calc_load_update = jiffies + LOAD_FREQ;
7356
7357 /*
7358 * During early bootup we pretend to be a normal task:
7359 */
7360 current->sched_class = &fair_sched_class;
7361
7362#ifdef CONFIG_SMP
7363 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7364 /* May be allocated at isolcpus cmdline parse time */
7365 if (cpu_isolated_map == NULL)
7366 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7367 idle_thread_set_boot_cpu();
7368#endif
7369 init_sched_fair_class();
7370
7371 scheduler_running = 1;
7372}
7373
7374#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7375static inline int preempt_count_equals(int preempt_offset)
7376{
7377 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7378
7379 return (nested == preempt_offset);
7380}
7381
7382void __might_sleep(const char *file, int line, int preempt_offset)
7383{
7384 static unsigned long prev_jiffy; /* ratelimiting */
7385
7386 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7387 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7388 system_state != SYSTEM_RUNNING || oops_in_progress)
7389 return;
7390 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7391 return;
7392 prev_jiffy = jiffies;
7393
7394 printk(KERN_ERR
7395 "BUG: sleeping function called from invalid context at %s:%d\n",
7396 file, line);
7397 printk(KERN_ERR
7398 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7399 in_atomic(), irqs_disabled(),
7400 current->pid, current->comm);
7401
7402 debug_show_held_locks(current);
7403 if (irqs_disabled())
7404 print_irqtrace_events(current);
7405 dump_stack();
7406}
7407EXPORT_SYMBOL(__might_sleep);
7408#endif
7409
7410#ifdef CONFIG_MAGIC_SYSRQ
7411static void normalize_task(struct rq *rq, struct task_struct *p)
7412{
7413 const struct sched_class *prev_class = p->sched_class;
7414 int old_prio = p->prio;
7415 int on_rq;
7416
7417 on_rq = p->on_rq;
7418 if (on_rq)
7419 dequeue_task(rq, p, 0);
7420 __setscheduler(rq, p, SCHED_NORMAL, 0);
7421 if (on_rq) {
7422 enqueue_task(rq, p, 0);
7423 resched_task(rq->curr);
7424 }
7425
7426 check_class_changed(rq, p, prev_class, old_prio);
7427}
7428
7429void normalize_rt_tasks(void)
7430{
7431 struct task_struct *g, *p;
7432 unsigned long flags;
7433 struct rq *rq;
7434
7435 read_lock_irqsave(&tasklist_lock, flags);
7436 do_each_thread(g, p) {
7437 /*
7438 * Only normalize user tasks:
7439 */
7440 if (!p->mm)
7441 continue;
7442
7443 p->se.exec_start = 0;
7444#ifdef CONFIG_SCHEDSTATS
7445 p->se.statistics.wait_start = 0;
7446 p->se.statistics.sleep_start = 0;
7447 p->se.statistics.block_start = 0;
7448#endif
7449
7450 if (!rt_task(p)) {
7451 /*
7452 * Renice negative nice level userspace
7453 * tasks back to 0:
7454 */
7455 if (TASK_NICE(p) < 0 && p->mm)
7456 set_user_nice(p, 0);
7457 continue;
7458 }
7459
7460 raw_spin_lock(&p->pi_lock);
7461 rq = __task_rq_lock(p);
7462
7463 normalize_task(rq, p);
7464
7465 __task_rq_unlock(rq);
7466 raw_spin_unlock(&p->pi_lock);
7467 } while_each_thread(g, p);
7468
7469 read_unlock_irqrestore(&tasklist_lock, flags);
7470}
7471
7472#endif /* CONFIG_MAGIC_SYSRQ */
7473
7474#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7475/*
7476 * These functions are only useful for the IA64 MCA handling, or kdb.
7477 *
7478 * They can only be called when the whole system has been
7479 * stopped - every CPU needs to be quiescent, and no scheduling
7480 * activity can take place. Using them for anything else would
7481 * be a serious bug, and as a result, they aren't even visible
7482 * under any other configuration.
7483 */
7484
7485/**
7486 * curr_task - return the current task for a given cpu.
7487 * @cpu: the processor in question.
7488 *
7489 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7490 */
7491struct task_struct *curr_task(int cpu)
7492{
7493 return cpu_curr(cpu);
7494}
7495
7496#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7497
7498#ifdef CONFIG_IA64
7499/**
7500 * set_curr_task - set the current task for a given cpu.
7501 * @cpu: the processor in question.
7502 * @p: the task pointer to set.
7503 *
7504 * Description: This function must only be used when non-maskable interrupts
7505 * are serviced on a separate stack. It allows the architecture to switch the
7506 * notion of the current task on a cpu in a non-blocking manner. This function
7507 * must be called with all CPU's synchronized, and interrupts disabled, the
7508 * and caller must save the original value of the current task (see
7509 * curr_task() above) and restore that value before reenabling interrupts and
7510 * re-starting the system.
7511 *
7512 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7513 */
7514void set_curr_task(int cpu, struct task_struct *p)
7515{
7516 cpu_curr(cpu) = p;
7517}
7518
7519#endif
7520
7521#ifdef CONFIG_CGROUP_SCHED
7522/* task_group_lock serializes the addition/removal of task groups */
7523static DEFINE_SPINLOCK(task_group_lock);
7524
7525static void free_sched_group(struct task_group *tg)
7526{
7527 free_fair_sched_group(tg);
7528 free_rt_sched_group(tg);
7529 autogroup_free(tg);
7530 kfree(tg);
7531}
7532
7533/* allocate runqueue etc for a new task group */
7534struct task_group *sched_create_group(struct task_group *parent)
7535{
7536 struct task_group *tg;
7537 unsigned long flags;
7538
7539 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7540 if (!tg)
7541 return ERR_PTR(-ENOMEM);
7542
7543 if (!alloc_fair_sched_group(tg, parent))
7544 goto err;
7545
7546 if (!alloc_rt_sched_group(tg, parent))
7547 goto err;
7548
7549 spin_lock_irqsave(&task_group_lock, flags);
7550 list_add_rcu(&tg->list, &task_groups);
7551
7552 WARN_ON(!parent); /* root should already exist */
7553
7554 tg->parent = parent;
7555 INIT_LIST_HEAD(&tg->children);
7556 list_add_rcu(&tg->siblings, &parent->children);
7557 spin_unlock_irqrestore(&task_group_lock, flags);
7558
7559 return tg;
7560
7561err:
7562 free_sched_group(tg);
7563 return ERR_PTR(-ENOMEM);
7564}
7565
7566/* rcu callback to free various structures associated with a task group */
7567static void free_sched_group_rcu(struct rcu_head *rhp)
7568{
7569 /* now it should be safe to free those cfs_rqs */
7570 free_sched_group(container_of(rhp, struct task_group, rcu));
7571}
7572
7573/* Destroy runqueue etc associated with a task group */
7574void sched_destroy_group(struct task_group *tg)
7575{
7576 unsigned long flags;
7577 int i;
7578
7579 /* end participation in shares distribution */
7580 for_each_possible_cpu(i)
7581 unregister_fair_sched_group(tg, i);
7582
7583 spin_lock_irqsave(&task_group_lock, flags);
7584 list_del_rcu(&tg->list);
7585 list_del_rcu(&tg->siblings);
7586 spin_unlock_irqrestore(&task_group_lock, flags);
7587
7588 /* wait for possible concurrent references to cfs_rqs complete */
7589 call_rcu(&tg->rcu, free_sched_group_rcu);
7590}
7591
7592/* change task's runqueue when it moves between groups.
7593 * The caller of this function should have put the task in its new group
7594 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7595 * reflect its new group.
7596 */
7597void sched_move_task(struct task_struct *tsk)
7598{
7599 struct task_group *tg;
7600 int on_rq, running;
7601 unsigned long flags;
7602 struct rq *rq;
7603
7604 rq = task_rq_lock(tsk, &flags);
7605
7606 running = task_current(rq, tsk);
7607 on_rq = tsk->on_rq;
7608
7609 if (on_rq)
7610 dequeue_task(rq, tsk, 0);
7611 if (unlikely(running))
7612 tsk->sched_class->put_prev_task(rq, tsk);
7613
7614 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7615 lockdep_is_held(&tsk->sighand->siglock)),
7616 struct task_group, css);
7617 tg = autogroup_task_group(tsk, tg);
7618 tsk->sched_task_group = tg;
7619
7620#ifdef CONFIG_FAIR_GROUP_SCHED
7621 if (tsk->sched_class->task_move_group)
7622 tsk->sched_class->task_move_group(tsk, on_rq);
7623 else
7624#endif
7625 set_task_rq(tsk, task_cpu(tsk));
7626
7627 if (unlikely(running))
7628 tsk->sched_class->set_curr_task(rq);
7629 if (on_rq)
7630 enqueue_task(rq, tsk, 0);
7631
7632 task_rq_unlock(rq, tsk, &flags);
7633}
7634#endif /* CONFIG_CGROUP_SCHED */
7635
7636#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7637static unsigned long to_ratio(u64 period, u64 runtime)
7638{
7639 if (runtime == RUNTIME_INF)
7640 return 1ULL << 20;
7641
7642 return div64_u64(runtime << 20, period);
7643}
7644#endif
7645
7646#ifdef CONFIG_RT_GROUP_SCHED
7647/*
7648 * Ensure that the real time constraints are schedulable.
7649 */
7650static DEFINE_MUTEX(rt_constraints_mutex);
7651
7652/* Must be called with tasklist_lock held */
7653static inline int tg_has_rt_tasks(struct task_group *tg)
7654{
7655 struct task_struct *g, *p;
7656
7657 do_each_thread(g, p) {
7658 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7659 return 1;
7660 } while_each_thread(g, p);
7661
7662 return 0;
7663}
7664
7665struct rt_schedulable_data {
7666 struct task_group *tg;
7667 u64 rt_period;
7668 u64 rt_runtime;
7669};
7670
7671static int tg_rt_schedulable(struct task_group *tg, void *data)
7672{
7673 struct rt_schedulable_data *d = data;
7674 struct task_group *child;
7675 unsigned long total, sum = 0;
7676 u64 period, runtime;
7677
7678 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7679 runtime = tg->rt_bandwidth.rt_runtime;
7680
7681 if (tg == d->tg) {
7682 period = d->rt_period;
7683 runtime = d->rt_runtime;
7684 }
7685
7686 /*
7687 * Cannot have more runtime than the period.
7688 */
7689 if (runtime > period && runtime != RUNTIME_INF)
7690 return -EINVAL;
7691
7692 /*
7693 * Ensure we don't starve existing RT tasks.
7694 */
7695 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7696 return -EBUSY;
7697
7698 total = to_ratio(period, runtime);
7699
7700 /*
7701 * Nobody can have more than the global setting allows.
7702 */
7703 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7704 return -EINVAL;
7705
7706 /*
7707 * The sum of our children's runtime should not exceed our own.
7708 */
7709 list_for_each_entry_rcu(child, &tg->children, siblings) {
7710 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7711 runtime = child->rt_bandwidth.rt_runtime;
7712
7713 if (child == d->tg) {
7714 period = d->rt_period;
7715 runtime = d->rt_runtime;
7716 }
7717
7718 sum += to_ratio(period, runtime);
7719 }
7720
7721 if (sum > total)
7722 return -EINVAL;
7723
7724 return 0;
7725}
7726
7727static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7728{
7729 int ret;
7730
7731 struct rt_schedulable_data data = {
7732 .tg = tg,
7733 .rt_period = period,
7734 .rt_runtime = runtime,
7735 };
7736
7737 rcu_read_lock();
7738 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7739 rcu_read_unlock();
7740
7741 return ret;
7742}
7743
7744static int tg_set_rt_bandwidth(struct task_group *tg,
7745 u64 rt_period, u64 rt_runtime)
7746{
7747 int i, err = 0;
7748
7749 mutex_lock(&rt_constraints_mutex);
7750 read_lock(&tasklist_lock);
7751 err = __rt_schedulable(tg, rt_period, rt_runtime);
7752 if (err)
7753 goto unlock;
7754
7755 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7756 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7757 tg->rt_bandwidth.rt_runtime = rt_runtime;
7758
7759 for_each_possible_cpu(i) {
7760 struct rt_rq *rt_rq = tg->rt_rq[i];
7761
7762 raw_spin_lock(&rt_rq->rt_runtime_lock);
7763 rt_rq->rt_runtime = rt_runtime;
7764 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7765 }
7766 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7767unlock:
7768 read_unlock(&tasklist_lock);
7769 mutex_unlock(&rt_constraints_mutex);
7770
7771 return err;
7772}
7773
7774int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7775{
7776 u64 rt_runtime, rt_period;
7777
7778 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7779 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7780 if (rt_runtime_us < 0)
7781 rt_runtime = RUNTIME_INF;
7782
7783 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7784}
7785
7786long sched_group_rt_runtime(struct task_group *tg)
7787{
7788 u64 rt_runtime_us;
7789
7790 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7791 return -1;
7792
7793 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7794 do_div(rt_runtime_us, NSEC_PER_USEC);
7795 return rt_runtime_us;
7796}
7797
7798int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7799{
7800 u64 rt_runtime, rt_period;
7801
7802 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7803 rt_runtime = tg->rt_bandwidth.rt_runtime;
7804
7805 if (rt_period == 0)
7806 return -EINVAL;
7807
7808 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7809}
7810
7811long sched_group_rt_period(struct task_group *tg)
7812{
7813 u64 rt_period_us;
7814
7815 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7816 do_div(rt_period_us, NSEC_PER_USEC);
7817 return rt_period_us;
7818}
7819
7820static int sched_rt_global_constraints(void)
7821{
7822 u64 runtime, period;
7823 int ret = 0;
7824
7825 if (sysctl_sched_rt_period <= 0)
7826 return -EINVAL;
7827
7828 runtime = global_rt_runtime();
7829 period = global_rt_period();
7830
7831 /*
7832 * Sanity check on the sysctl variables.
7833 */
7834 if (runtime > period && runtime != RUNTIME_INF)
7835 return -EINVAL;
7836
7837 mutex_lock(&rt_constraints_mutex);
7838 read_lock(&tasklist_lock);
7839 ret = __rt_schedulable(NULL, 0, 0);
7840 read_unlock(&tasklist_lock);
7841 mutex_unlock(&rt_constraints_mutex);
7842
7843 return ret;
7844}
7845
7846int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7847{
7848 /* Don't accept realtime tasks when there is no way for them to run */
7849 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7850 return 0;
7851
7852 return 1;
7853}
7854
7855#else /* !CONFIG_RT_GROUP_SCHED */
7856static int sched_rt_global_constraints(void)
7857{
7858 unsigned long flags;
7859 int i;
7860
7861 if (sysctl_sched_rt_period <= 0)
7862 return -EINVAL;
7863
7864 /*
7865 * There's always some RT tasks in the root group
7866 * -- migration, kstopmachine etc..
7867 */
7868 if (sysctl_sched_rt_runtime == 0)
7869 return -EBUSY;
7870
7871 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7872 for_each_possible_cpu(i) {
7873 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7874
7875 raw_spin_lock(&rt_rq->rt_runtime_lock);
7876 rt_rq->rt_runtime = global_rt_runtime();
7877 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7878 }
7879 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7880
7881 return 0;
7882}
7883#endif /* CONFIG_RT_GROUP_SCHED */
7884
7885int sched_rt_handler(struct ctl_table *table, int write,
7886 void __user *buffer, size_t *lenp,
7887 loff_t *ppos)
7888{
7889 int ret;
7890 int old_period, old_runtime;
7891 static DEFINE_MUTEX(mutex);
7892
7893 mutex_lock(&mutex);
7894 old_period = sysctl_sched_rt_period;
7895 old_runtime = sysctl_sched_rt_runtime;
7896
7897 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7898
7899 if (!ret && write) {
7900 ret = sched_rt_global_constraints();
7901 if (ret) {
7902 sysctl_sched_rt_period = old_period;
7903 sysctl_sched_rt_runtime = old_runtime;
7904 } else {
7905 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7906 def_rt_bandwidth.rt_period =
7907 ns_to_ktime(global_rt_period());
7908 }
7909 }
7910 mutex_unlock(&mutex);
7911
7912 return ret;
7913}
7914
7915#ifdef CONFIG_CGROUP_SCHED
7916
7917/* return corresponding task_group object of a cgroup */
7918static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7919{
7920 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7921 struct task_group, css);
7922}
7923
7924static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7925{
7926 struct task_group *tg, *parent;
7927
7928 if (!cgrp->parent) {
7929 /* This is early initialization for the top cgroup */
7930 return &root_task_group.css;
7931 }
7932
7933 parent = cgroup_tg(cgrp->parent);
7934 tg = sched_create_group(parent);
7935 if (IS_ERR(tg))
7936 return ERR_PTR(-ENOMEM);
7937
7938 return &tg->css;
7939}
7940
7941static void cpu_cgroup_destroy(struct cgroup *cgrp)
7942{
7943 struct task_group *tg = cgroup_tg(cgrp);
7944
7945 sched_destroy_group(tg);
7946}
7947
7948static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7949 struct cgroup_taskset *tset)
7950{
7951 struct task_struct *task;
7952
7953 cgroup_taskset_for_each(task, cgrp, tset) {
7954#ifdef CONFIG_RT_GROUP_SCHED
7955 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7956 return -EINVAL;
7957#else
7958 /* We don't support RT-tasks being in separate groups */
7959 if (task->sched_class != &fair_sched_class)
7960 return -EINVAL;
7961#endif
7962 }
7963 return 0;
7964}
7965
7966static void cpu_cgroup_attach(struct cgroup *cgrp,
7967 struct cgroup_taskset *tset)
7968{
7969 struct task_struct *task;
7970
7971 cgroup_taskset_for_each(task, cgrp, tset)
7972 sched_move_task(task);
7973}
7974
7975static void
7976cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7977 struct task_struct *task)
7978{
7979 /*
7980 * cgroup_exit() is called in the copy_process() failure path.
7981 * Ignore this case since the task hasn't ran yet, this avoids
7982 * trying to poke a half freed task state from generic code.
7983 */
7984 if (!(task->flags & PF_EXITING))
7985 return;
7986
7987 sched_move_task(task);
7988}
7989
7990#ifdef CONFIG_FAIR_GROUP_SCHED
7991static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7992 u64 shareval)
7993{
7994 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7995}
7996
7997static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7998{
7999 struct task_group *tg = cgroup_tg(cgrp);
8000
8001 return (u64) scale_load_down(tg->shares);
8002}
8003
8004#ifdef CONFIG_CFS_BANDWIDTH
8005static DEFINE_MUTEX(cfs_constraints_mutex);
8006
8007const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8008const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8009
8010static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8011
8012static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8013{
8014 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8015 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8016
8017 if (tg == &root_task_group)
8018 return -EINVAL;
8019
8020 /*
8021 * Ensure we have at some amount of bandwidth every period. This is
8022 * to prevent reaching a state of large arrears when throttled via
8023 * entity_tick() resulting in prolonged exit starvation.
8024 */
8025 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8026 return -EINVAL;
8027
8028 /*
8029 * Likewise, bound things on the otherside by preventing insane quota
8030 * periods. This also allows us to normalize in computing quota
8031 * feasibility.
8032 */
8033 if (period > max_cfs_quota_period)
8034 return -EINVAL;
8035
8036 mutex_lock(&cfs_constraints_mutex);
8037 ret = __cfs_schedulable(tg, period, quota);
8038 if (ret)
8039 goto out_unlock;
8040
8041 runtime_enabled = quota != RUNTIME_INF;
8042 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8043 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
8044 raw_spin_lock_irq(&cfs_b->lock);
8045 cfs_b->period = ns_to_ktime(period);
8046 cfs_b->quota = quota;
8047
8048 __refill_cfs_bandwidth_runtime(cfs_b);
8049 /* restart the period timer (if active) to handle new period expiry */
8050 if (runtime_enabled && cfs_b->timer_active) {
8051 /* force a reprogram */
8052 cfs_b->timer_active = 0;
8053 __start_cfs_bandwidth(cfs_b);
8054 }
8055 raw_spin_unlock_irq(&cfs_b->lock);
8056
8057 for_each_possible_cpu(i) {
8058 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8059 struct rq *rq = cfs_rq->rq;
8060
8061 raw_spin_lock_irq(&rq->lock);
8062 cfs_rq->runtime_enabled = runtime_enabled;
8063 cfs_rq->runtime_remaining = 0;
8064
8065 if (cfs_rq->throttled)
8066 unthrottle_cfs_rq(cfs_rq);
8067 raw_spin_unlock_irq(&rq->lock);
8068 }
8069out_unlock:
8070 mutex_unlock(&cfs_constraints_mutex);
8071
8072 return ret;
8073}
8074
8075int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8076{
8077 u64 quota, period;
8078
8079 period = ktime_to_ns(tg->cfs_bandwidth.period);
8080 if (cfs_quota_us < 0)
8081 quota = RUNTIME_INF;
8082 else
8083 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8084
8085 return tg_set_cfs_bandwidth(tg, period, quota);
8086}
8087
8088long tg_get_cfs_quota(struct task_group *tg)
8089{
8090 u64 quota_us;
8091
8092 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8093 return -1;
8094
8095 quota_us = tg->cfs_bandwidth.quota;
8096 do_div(quota_us, NSEC_PER_USEC);
8097
8098 return quota_us;
8099}
8100
8101int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8102{
8103 u64 quota, period;
8104
8105 period = (u64)cfs_period_us * NSEC_PER_USEC;
8106 quota = tg->cfs_bandwidth.quota;
8107
8108 return tg_set_cfs_bandwidth(tg, period, quota);
8109}
8110
8111long tg_get_cfs_period(struct task_group *tg)
8112{
8113 u64 cfs_period_us;
8114
8115 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8116 do_div(cfs_period_us, NSEC_PER_USEC);
8117
8118 return cfs_period_us;
8119}
8120
8121static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
8122{
8123 return tg_get_cfs_quota(cgroup_tg(cgrp));
8124}
8125
8126static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
8127 s64 cfs_quota_us)
8128{
8129 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
8130}
8131
8132static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
8133{
8134 return tg_get_cfs_period(cgroup_tg(cgrp));
8135}
8136
8137static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8138 u64 cfs_period_us)
8139{
8140 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
8141}
8142
8143struct cfs_schedulable_data {
8144 struct task_group *tg;
8145 u64 period, quota;
8146};
8147
8148/*
8149 * normalize group quota/period to be quota/max_period
8150 * note: units are usecs
8151 */
8152static u64 normalize_cfs_quota(struct task_group *tg,
8153 struct cfs_schedulable_data *d)
8154{
8155 u64 quota, period;
8156
8157 if (tg == d->tg) {
8158 period = d->period;
8159 quota = d->quota;
8160 } else {
8161 period = tg_get_cfs_period(tg);
8162 quota = tg_get_cfs_quota(tg);
8163 }
8164
8165 /* note: these should typically be equivalent */
8166 if (quota == RUNTIME_INF || quota == -1)
8167 return RUNTIME_INF;
8168
8169 return to_ratio(period, quota);
8170}
8171
8172static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8173{
8174 struct cfs_schedulable_data *d = data;
8175 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8176 s64 quota = 0, parent_quota = -1;
8177
8178 if (!tg->parent) {
8179 quota = RUNTIME_INF;
8180 } else {
8181 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8182
8183 quota = normalize_cfs_quota(tg, d);
8184 parent_quota = parent_b->hierarchal_quota;
8185
8186 /*
8187 * ensure max(child_quota) <= parent_quota, inherit when no
8188 * limit is set
8189 */
8190 if (quota == RUNTIME_INF)
8191 quota = parent_quota;
8192 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8193 return -EINVAL;
8194 }
8195 cfs_b->hierarchal_quota = quota;
8196
8197 return 0;
8198}
8199
8200static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8201{
8202 int ret;
8203 struct cfs_schedulable_data data = {
8204 .tg = tg,
8205 .period = period,
8206 .quota = quota,
8207 };
8208
8209 if (quota != RUNTIME_INF) {
8210 do_div(data.period, NSEC_PER_USEC);
8211 do_div(data.quota, NSEC_PER_USEC);
8212 }
8213
8214 rcu_read_lock();
8215 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8216 rcu_read_unlock();
8217
8218 return ret;
8219}
8220
8221static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
8222 struct cgroup_map_cb *cb)
8223{
8224 struct task_group *tg = cgroup_tg(cgrp);
8225 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8226
8227 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
8228 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
8229 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
8230
8231 return 0;
8232}
8233#endif /* CONFIG_CFS_BANDWIDTH */
8234#endif /* CONFIG_FAIR_GROUP_SCHED */
8235
8236#ifdef CONFIG_RT_GROUP_SCHED
8237static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8238 s64 val)
8239{
8240 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8241}
8242
8243static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8244{
8245 return sched_group_rt_runtime(cgroup_tg(cgrp));
8246}
8247
8248static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8249 u64 rt_period_us)
8250{
8251 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8252}
8253
8254static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8255{
8256 return sched_group_rt_period(cgroup_tg(cgrp));
8257}
8258#endif /* CONFIG_RT_GROUP_SCHED */
8259
8260static struct cftype cpu_files[] = {
8261#ifdef CONFIG_FAIR_GROUP_SCHED
8262 {
8263 .name = "shares",
8264 .read_u64 = cpu_shares_read_u64,
8265 .write_u64 = cpu_shares_write_u64,
8266 },
8267#endif
8268#ifdef CONFIG_CFS_BANDWIDTH
8269 {
8270 .name = "cfs_quota_us",
8271 .read_s64 = cpu_cfs_quota_read_s64,
8272 .write_s64 = cpu_cfs_quota_write_s64,
8273 },
8274 {
8275 .name = "cfs_period_us",
8276 .read_u64 = cpu_cfs_period_read_u64,
8277 .write_u64 = cpu_cfs_period_write_u64,
8278 },
8279 {
8280 .name = "stat",
8281 .read_map = cpu_stats_show,
8282 },
8283#endif
8284#ifdef CONFIG_RT_GROUP_SCHED
8285 {
8286 .name = "rt_runtime_us",
8287 .read_s64 = cpu_rt_runtime_read,
8288 .write_s64 = cpu_rt_runtime_write,
8289 },
8290 {
8291 .name = "rt_period_us",
8292 .read_u64 = cpu_rt_period_read_uint,
8293 .write_u64 = cpu_rt_period_write_uint,
8294 },
8295#endif
8296 { } /* terminate */
8297};
8298
8299struct cgroup_subsys cpu_cgroup_subsys = {
8300 .name = "cpu",
8301 .create = cpu_cgroup_create,
8302 .destroy = cpu_cgroup_destroy,
8303 .can_attach = cpu_cgroup_can_attach,
8304 .attach = cpu_cgroup_attach,
8305 .exit = cpu_cgroup_exit,
8306 .subsys_id = cpu_cgroup_subsys_id,
8307 .base_cftypes = cpu_files,
8308 .early_init = 1,
8309};
8310
8311#endif /* CONFIG_CGROUP_SCHED */
8312
8313#ifdef CONFIG_CGROUP_CPUACCT
8314
8315/*
8316 * CPU accounting code for task groups.
8317 *
8318 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8319 * (balbir@in.ibm.com).
8320 */
8321
8322/* create a new cpu accounting group */
8323static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
8324{
8325 struct cpuacct *ca;
8326
8327 if (!cgrp->parent)
8328 return &root_cpuacct.css;
8329
8330 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8331 if (!ca)
8332 goto out;
8333
8334 ca->cpuusage = alloc_percpu(u64);
8335 if (!ca->cpuusage)
8336 goto out_free_ca;
8337
8338 ca->cpustat = alloc_percpu(struct kernel_cpustat);
8339 if (!ca->cpustat)
8340 goto out_free_cpuusage;
8341
8342 return &ca->css;
8343
8344out_free_cpuusage:
8345 free_percpu(ca->cpuusage);
8346out_free_ca:
8347 kfree(ca);
8348out:
8349 return ERR_PTR(-ENOMEM);
8350}
8351
8352/* destroy an existing cpu accounting group */
8353static void cpuacct_destroy(struct cgroup *cgrp)
8354{
8355 struct cpuacct *ca = cgroup_ca(cgrp);
8356
8357 free_percpu(ca->cpustat);
8358 free_percpu(ca->cpuusage);
8359 kfree(ca);
8360}
8361
8362static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8363{
8364 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8365 u64 data;
8366
8367#ifndef CONFIG_64BIT
8368 /*
8369 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8370 */
8371 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8372 data = *cpuusage;
8373 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8374#else
8375 data = *cpuusage;
8376#endif
8377
8378 return data;
8379}
8380
8381static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8382{
8383 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8384
8385#ifndef CONFIG_64BIT
8386 /*
8387 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8388 */
8389 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8390 *cpuusage = val;
8391 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8392#else
8393 *cpuusage = val;
8394#endif
8395}
8396
8397/* return total cpu usage (in nanoseconds) of a group */
8398static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8399{
8400 struct cpuacct *ca = cgroup_ca(cgrp);
8401 u64 totalcpuusage = 0;
8402 int i;
8403
8404 for_each_present_cpu(i)
8405 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8406
8407 return totalcpuusage;
8408}
8409
8410static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8411 u64 reset)
8412{
8413 struct cpuacct *ca = cgroup_ca(cgrp);
8414 int err = 0;
8415 int i;
8416
8417 if (reset) {
8418 err = -EINVAL;
8419 goto out;
8420 }
8421
8422 for_each_present_cpu(i)
8423 cpuacct_cpuusage_write(ca, i, 0);
8424
8425out:
8426 return err;
8427}
8428
8429static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8430 struct seq_file *m)
8431{
8432 struct cpuacct *ca = cgroup_ca(cgroup);
8433 u64 percpu;
8434 int i;
8435
8436 for_each_present_cpu(i) {
8437 percpu = cpuacct_cpuusage_read(ca, i);
8438 seq_printf(m, "%llu ", (unsigned long long) percpu);
8439 }
8440 seq_printf(m, "\n");
8441 return 0;
8442}
8443
8444static const char *cpuacct_stat_desc[] = {
8445 [CPUACCT_STAT_USER] = "user",
8446 [CPUACCT_STAT_SYSTEM] = "system",
8447};
8448
8449static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8450 struct cgroup_map_cb *cb)
8451{
8452 struct cpuacct *ca = cgroup_ca(cgrp);
8453 int cpu;
8454 s64 val = 0;
8455
8456 for_each_online_cpu(cpu) {
8457 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8458 val += kcpustat->cpustat[CPUTIME_USER];
8459 val += kcpustat->cpustat[CPUTIME_NICE];
8460 }
8461 val = cputime64_to_clock_t(val);
8462 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8463
8464 val = 0;
8465 for_each_online_cpu(cpu) {
8466 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8467 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8468 val += kcpustat->cpustat[CPUTIME_IRQ];
8469 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8470 }
8471
8472 val = cputime64_to_clock_t(val);
8473 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8474
8475 return 0;
8476}
8477
8478static struct cftype files[] = {
8479 {
8480 .name = "usage",
8481 .read_u64 = cpuusage_read,
8482 .write_u64 = cpuusage_write,
8483 },
8484 {
8485 .name = "usage_percpu",
8486 .read_seq_string = cpuacct_percpu_seq_read,
8487 },
8488 {
8489 .name = "stat",
8490 .read_map = cpuacct_stats_show,
8491 },
8492 { } /* terminate */
8493};
8494
8495/*
8496 * charge this task's execution time to its accounting group.
8497 *
8498 * called with rq->lock held.
8499 */
8500void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8501{
8502 struct cpuacct *ca;
8503 int cpu;
8504
8505 if (unlikely(!cpuacct_subsys.active))
8506 return;
8507
8508 cpu = task_cpu(tsk);
8509
8510 rcu_read_lock();
8511
8512 ca = task_ca(tsk);
8513
8514 for (; ca; ca = parent_ca(ca)) {
8515 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8516 *cpuusage += cputime;
8517 }
8518
8519 rcu_read_unlock();
8520}
8521
8522struct cgroup_subsys cpuacct_subsys = {
8523 .name = "cpuacct",
8524 .create = cpuacct_create,
8525 .destroy = cpuacct_destroy,
8526 .subsys_id = cpuacct_subsys_id,
8527 .base_cftypes = files,
8528};
8529#endif /* CONFIG_CGROUP_CPUACCT */