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#include <linux/context_tracking.h>
76#include <linux/compiler.h>
77
78#include <asm/switch_to.h>
79#include <asm/tlb.h>
80#include <asm/irq_regs.h>
81#include <asm/mutex.h>
82#ifdef CONFIG_PARAVIRT
83#include <asm/paravirt.h>
84#endif
85
86#include "sched.h"
87#include "../workqueue_internal.h"
88#include "../smpboot.h"
89
90#define CREATE_TRACE_POINTS
91#include <trace/events/sched.h>
92
93void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
94{
95 unsigned long delta;
96 ktime_t soft, hard, now;
97
98 for (;;) {
99 if (hrtimer_active(period_timer))
100 break;
101
102 now = hrtimer_cb_get_time(period_timer);
103 hrtimer_forward(period_timer, now, period);
104
105 soft = hrtimer_get_softexpires(period_timer);
106 hard = hrtimer_get_expires(period_timer);
107 delta = ktime_to_ns(ktime_sub(hard, soft));
108 __hrtimer_start_range_ns(period_timer, soft, delta,
109 HRTIMER_MODE_ABS_PINNED, 0);
110 }
111}
112
113DEFINE_MUTEX(sched_domains_mutex);
114DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115
116static void update_rq_clock_task(struct rq *rq, s64 delta);
117
118void update_rq_clock(struct rq *rq)
119{
120 s64 delta;
121
122 if (rq->skip_clock_update > 0)
123 return;
124
125 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
126 rq->clock += delta;
127 update_rq_clock_task(rq, delta);
128}
129
130/*
131 * Debugging: various feature bits
132 */
133
134#define SCHED_FEAT(name, enabled) \
135 (1UL << __SCHED_FEAT_##name) * enabled |
136
137const_debug unsigned int sysctl_sched_features =
138#include "features.h"
139 0;
140
141#undef SCHED_FEAT
142
143#ifdef CONFIG_SCHED_DEBUG
144#define SCHED_FEAT(name, enabled) \
145 #name ,
146
147static const char * const sched_feat_names[] = {
148#include "features.h"
149};
150
151#undef SCHED_FEAT
152
153static int sched_feat_show(struct seq_file *m, void *v)
154{
155 int i;
156
157 for (i = 0; i < __SCHED_FEAT_NR; i++) {
158 if (!(sysctl_sched_features & (1UL << i)))
159 seq_puts(m, "NO_");
160 seq_printf(m, "%s ", sched_feat_names[i]);
161 }
162 seq_puts(m, "\n");
163
164 return 0;
165}
166
167#ifdef HAVE_JUMP_LABEL
168
169#define jump_label_key__true STATIC_KEY_INIT_TRUE
170#define jump_label_key__false STATIC_KEY_INIT_FALSE
171
172#define SCHED_FEAT(name, enabled) \
173 jump_label_key__##enabled ,
174
175struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
176#include "features.h"
177};
178
179#undef SCHED_FEAT
180
181static void sched_feat_disable(int i)
182{
183 if (static_key_enabled(&sched_feat_keys[i]))
184 static_key_slow_dec(&sched_feat_keys[i]);
185}
186
187static void sched_feat_enable(int i)
188{
189 if (!static_key_enabled(&sched_feat_keys[i]))
190 static_key_slow_inc(&sched_feat_keys[i]);
191}
192#else
193static void sched_feat_disable(int i) { };
194static void sched_feat_enable(int i) { };
195#endif /* HAVE_JUMP_LABEL */
196
197static int sched_feat_set(char *cmp)
198{
199 int i;
200 int neg = 0;
201
202 if (strncmp(cmp, "NO_", 3) == 0) {
203 neg = 1;
204 cmp += 3;
205 }
206
207 for (i = 0; i < __SCHED_FEAT_NR; i++) {
208 if (strcmp(cmp, sched_feat_names[i]) == 0) {
209 if (neg) {
210 sysctl_sched_features &= ~(1UL << i);
211 sched_feat_disable(i);
212 } else {
213 sysctl_sched_features |= (1UL << i);
214 sched_feat_enable(i);
215 }
216 break;
217 }
218 }
219
220 return i;
221}
222
223static ssize_t
224sched_feat_write(struct file *filp, const char __user *ubuf,
225 size_t cnt, loff_t *ppos)
226{
227 char buf[64];
228 char *cmp;
229 int i;
230
231 if (cnt > 63)
232 cnt = 63;
233
234 if (copy_from_user(&buf, ubuf, cnt))
235 return -EFAULT;
236
237 buf[cnt] = 0;
238 cmp = strstrip(buf);
239
240 i = sched_feat_set(cmp);
241 if (i == __SCHED_FEAT_NR)
242 return -EINVAL;
243
244 *ppos += cnt;
245
246 return cnt;
247}
248
249static int sched_feat_open(struct inode *inode, struct file *filp)
250{
251 return single_open(filp, sched_feat_show, NULL);
252}
253
254static const struct file_operations sched_feat_fops = {
255 .open = sched_feat_open,
256 .write = sched_feat_write,
257 .read = seq_read,
258 .llseek = seq_lseek,
259 .release = single_release,
260};
261
262static __init int sched_init_debug(void)
263{
264 debugfs_create_file("sched_features", 0644, NULL, NULL,
265 &sched_feat_fops);
266
267 return 0;
268}
269late_initcall(sched_init_debug);
270#endif /* CONFIG_SCHED_DEBUG */
271
272/*
273 * Number of tasks to iterate in a single balance run.
274 * Limited because this is done with IRQs disabled.
275 */
276const_debug unsigned int sysctl_sched_nr_migrate = 32;
277
278/*
279 * period over which we average the RT time consumption, measured
280 * in ms.
281 *
282 * default: 1s
283 */
284const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
285
286/*
287 * period over which we measure -rt task cpu usage in us.
288 * default: 1s
289 */
290unsigned int sysctl_sched_rt_period = 1000000;
291
292__read_mostly int scheduler_running;
293
294/*
295 * part of the period that we allow rt tasks to run in us.
296 * default: 0.95s
297 */
298int sysctl_sched_rt_runtime = 950000;
299
300/*
301 * __task_rq_lock - lock the rq @p resides on.
302 */
303static inline struct rq *__task_rq_lock(struct task_struct *p)
304 __acquires(rq->lock)
305{
306 struct rq *rq;
307
308 lockdep_assert_held(&p->pi_lock);
309
310 for (;;) {
311 rq = task_rq(p);
312 raw_spin_lock(&rq->lock);
313 if (likely(rq == task_rq(p)))
314 return rq;
315 raw_spin_unlock(&rq->lock);
316 }
317}
318
319/*
320 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
321 */
322static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
323 __acquires(p->pi_lock)
324 __acquires(rq->lock)
325{
326 struct rq *rq;
327
328 for (;;) {
329 raw_spin_lock_irqsave(&p->pi_lock, *flags);
330 rq = task_rq(p);
331 raw_spin_lock(&rq->lock);
332 if (likely(rq == task_rq(p)))
333 return rq;
334 raw_spin_unlock(&rq->lock);
335 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
336 }
337}
338
339static void __task_rq_unlock(struct rq *rq)
340 __releases(rq->lock)
341{
342 raw_spin_unlock(&rq->lock);
343}
344
345static inline void
346task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
347 __releases(rq->lock)
348 __releases(p->pi_lock)
349{
350 raw_spin_unlock(&rq->lock);
351 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
352}
353
354/*
355 * this_rq_lock - lock this runqueue and disable interrupts.
356 */
357static struct rq *this_rq_lock(void)
358 __acquires(rq->lock)
359{
360 struct rq *rq;
361
362 local_irq_disable();
363 rq = this_rq();
364 raw_spin_lock(&rq->lock);
365
366 return rq;
367}
368
369#ifdef CONFIG_SCHED_HRTICK
370/*
371 * Use HR-timers to deliver accurate preemption points.
372 */
373
374static void hrtick_clear(struct rq *rq)
375{
376 if (hrtimer_active(&rq->hrtick_timer))
377 hrtimer_cancel(&rq->hrtick_timer);
378}
379
380/*
381 * High-resolution timer tick.
382 * Runs from hardirq context with interrupts disabled.
383 */
384static enum hrtimer_restart hrtick(struct hrtimer *timer)
385{
386 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
387
388 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
389
390 raw_spin_lock(&rq->lock);
391 update_rq_clock(rq);
392 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
393 raw_spin_unlock(&rq->lock);
394
395 return HRTIMER_NORESTART;
396}
397
398#ifdef CONFIG_SMP
399
400static int __hrtick_restart(struct rq *rq)
401{
402 struct hrtimer *timer = &rq->hrtick_timer;
403 ktime_t time = hrtimer_get_softexpires(timer);
404
405 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
406}
407
408/*
409 * called from hardirq (IPI) context
410 */
411static void __hrtick_start(void *arg)
412{
413 struct rq *rq = arg;
414
415 raw_spin_lock(&rq->lock);
416 __hrtick_restart(rq);
417 rq->hrtick_csd_pending = 0;
418 raw_spin_unlock(&rq->lock);
419}
420
421/*
422 * Called to set the hrtick timer state.
423 *
424 * called with rq->lock held and irqs disabled
425 */
426void hrtick_start(struct rq *rq, u64 delay)
427{
428 struct hrtimer *timer = &rq->hrtick_timer;
429 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
430
431 hrtimer_set_expires(timer, time);
432
433 if (rq == this_rq()) {
434 __hrtick_restart(rq);
435 } else if (!rq->hrtick_csd_pending) {
436 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
437 rq->hrtick_csd_pending = 1;
438 }
439}
440
441static int
442hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
443{
444 int cpu = (int)(long)hcpu;
445
446 switch (action) {
447 case CPU_UP_CANCELED:
448 case CPU_UP_CANCELED_FROZEN:
449 case CPU_DOWN_PREPARE:
450 case CPU_DOWN_PREPARE_FROZEN:
451 case CPU_DEAD:
452 case CPU_DEAD_FROZEN:
453 hrtick_clear(cpu_rq(cpu));
454 return NOTIFY_OK;
455 }
456
457 return NOTIFY_DONE;
458}
459
460static __init void init_hrtick(void)
461{
462 hotcpu_notifier(hotplug_hrtick, 0);
463}
464#else
465/*
466 * Called to set the hrtick timer state.
467 *
468 * called with rq->lock held and irqs disabled
469 */
470void hrtick_start(struct rq *rq, u64 delay)
471{
472 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
473 HRTIMER_MODE_REL_PINNED, 0);
474}
475
476static inline void init_hrtick(void)
477{
478}
479#endif /* CONFIG_SMP */
480
481static void init_rq_hrtick(struct rq *rq)
482{
483#ifdef CONFIG_SMP
484 rq->hrtick_csd_pending = 0;
485
486 rq->hrtick_csd.flags = 0;
487 rq->hrtick_csd.func = __hrtick_start;
488 rq->hrtick_csd.info = rq;
489#endif
490
491 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
492 rq->hrtick_timer.function = hrtick;
493}
494#else /* CONFIG_SCHED_HRTICK */
495static inline void hrtick_clear(struct rq *rq)
496{
497}
498
499static inline void init_rq_hrtick(struct rq *rq)
500{
501}
502
503static inline void init_hrtick(void)
504{
505}
506#endif /* CONFIG_SCHED_HRTICK */
507
508/*
509 * resched_task - mark a task 'to be rescheduled now'.
510 *
511 * On UP this means the setting of the need_resched flag, on SMP it
512 * might also involve a cross-CPU call to trigger the scheduler on
513 * the target CPU.
514 */
515void resched_task(struct task_struct *p)
516{
517 int cpu;
518
519 lockdep_assert_held(&task_rq(p)->lock);
520
521 if (test_tsk_need_resched(p))
522 return;
523
524 set_tsk_need_resched(p);
525
526 cpu = task_cpu(p);
527 if (cpu == smp_processor_id()) {
528 set_preempt_need_resched();
529 return;
530 }
531
532 /* NEED_RESCHED must be visible before we test polling */
533 smp_mb();
534 if (!tsk_is_polling(p))
535 smp_send_reschedule(cpu);
536}
537
538void resched_cpu(int cpu)
539{
540 struct rq *rq = cpu_rq(cpu);
541 unsigned long flags;
542
543 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
544 return;
545 resched_task(cpu_curr(cpu));
546 raw_spin_unlock_irqrestore(&rq->lock, flags);
547}
548
549#ifdef CONFIG_SMP
550#ifdef CONFIG_NO_HZ_COMMON
551/*
552 * In the semi idle case, use the nearest busy cpu for migrating timers
553 * from an idle cpu. This is good for power-savings.
554 *
555 * We don't do similar optimization for completely idle system, as
556 * selecting an idle cpu will add more delays to the timers than intended
557 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
558 */
559int get_nohz_timer_target(int pinned)
560{
561 int cpu = smp_processor_id();
562 int i;
563 struct sched_domain *sd;
564
565 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
566 return cpu;
567
568 rcu_read_lock();
569 for_each_domain(cpu, sd) {
570 for_each_cpu(i, sched_domain_span(sd)) {
571 if (!idle_cpu(i)) {
572 cpu = i;
573 goto unlock;
574 }
575 }
576 }
577unlock:
578 rcu_read_unlock();
579 return cpu;
580}
581/*
582 * When add_timer_on() enqueues a timer into the timer wheel of an
583 * idle CPU then this timer might expire before the next timer event
584 * which is scheduled to wake up that CPU. In case of a completely
585 * idle system the next event might even be infinite time into the
586 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
587 * leaves the inner idle loop so the newly added timer is taken into
588 * account when the CPU goes back to idle and evaluates the timer
589 * wheel for the next timer event.
590 */
591static void wake_up_idle_cpu(int cpu)
592{
593 struct rq *rq = cpu_rq(cpu);
594
595 if (cpu == smp_processor_id())
596 return;
597
598 /*
599 * This is safe, as this function is called with the timer
600 * wheel base lock of (cpu) held. When the CPU is on the way
601 * to idle and has not yet set rq->curr to idle then it will
602 * be serialized on the timer wheel base lock and take the new
603 * timer into account automatically.
604 */
605 if (rq->curr != rq->idle)
606 return;
607
608 /*
609 * We can set TIF_RESCHED on the idle task of the other CPU
610 * lockless. The worst case is that the other CPU runs the
611 * idle task through an additional NOOP schedule()
612 */
613 set_tsk_need_resched(rq->idle);
614
615 /* NEED_RESCHED must be visible before we test polling */
616 smp_mb();
617 if (!tsk_is_polling(rq->idle))
618 smp_send_reschedule(cpu);
619}
620
621static bool wake_up_full_nohz_cpu(int cpu)
622{
623 if (tick_nohz_full_cpu(cpu)) {
624 if (cpu != smp_processor_id() ||
625 tick_nohz_tick_stopped())
626 smp_send_reschedule(cpu);
627 return true;
628 }
629
630 return false;
631}
632
633void wake_up_nohz_cpu(int cpu)
634{
635 if (!wake_up_full_nohz_cpu(cpu))
636 wake_up_idle_cpu(cpu);
637}
638
639static inline bool got_nohz_idle_kick(void)
640{
641 int cpu = smp_processor_id();
642
643 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
644 return false;
645
646 if (idle_cpu(cpu) && !need_resched())
647 return true;
648
649 /*
650 * We can't run Idle Load Balance on this CPU for this time so we
651 * cancel it and clear NOHZ_BALANCE_KICK
652 */
653 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
654 return false;
655}
656
657#else /* CONFIG_NO_HZ_COMMON */
658
659static inline bool got_nohz_idle_kick(void)
660{
661 return false;
662}
663
664#endif /* CONFIG_NO_HZ_COMMON */
665
666#ifdef CONFIG_NO_HZ_FULL
667bool sched_can_stop_tick(void)
668{
669 struct rq *rq;
670
671 rq = this_rq();
672
673 /* Make sure rq->nr_running update is visible after the IPI */
674 smp_rmb();
675
676 /* More than one running task need preemption */
677 if (rq->nr_running > 1)
678 return false;
679
680 return true;
681}
682#endif /* CONFIG_NO_HZ_FULL */
683
684void sched_avg_update(struct rq *rq)
685{
686 s64 period = sched_avg_period();
687
688 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
689 /*
690 * Inline assembly required to prevent the compiler
691 * optimising this loop into a divmod call.
692 * See __iter_div_u64_rem() for another example of this.
693 */
694 asm("" : "+rm" (rq->age_stamp));
695 rq->age_stamp += period;
696 rq->rt_avg /= 2;
697 }
698}
699
700#endif /* CONFIG_SMP */
701
702#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
703 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
704/*
705 * Iterate task_group tree rooted at *from, calling @down when first entering a
706 * node and @up when leaving it for the final time.
707 *
708 * Caller must hold rcu_lock or sufficient equivalent.
709 */
710int walk_tg_tree_from(struct task_group *from,
711 tg_visitor down, tg_visitor up, void *data)
712{
713 struct task_group *parent, *child;
714 int ret;
715
716 parent = from;
717
718down:
719 ret = (*down)(parent, data);
720 if (ret)
721 goto out;
722 list_for_each_entry_rcu(child, &parent->children, siblings) {
723 parent = child;
724 goto down;
725
726up:
727 continue;
728 }
729 ret = (*up)(parent, data);
730 if (ret || parent == from)
731 goto out;
732
733 child = parent;
734 parent = parent->parent;
735 if (parent)
736 goto up;
737out:
738 return ret;
739}
740
741int tg_nop(struct task_group *tg, void *data)
742{
743 return 0;
744}
745#endif
746
747static void set_load_weight(struct task_struct *p)
748{
749 int prio = p->static_prio - MAX_RT_PRIO;
750 struct load_weight *load = &p->se.load;
751
752 /*
753 * SCHED_IDLE tasks get minimal weight:
754 */
755 if (p->policy == SCHED_IDLE) {
756 load->weight = scale_load(WEIGHT_IDLEPRIO);
757 load->inv_weight = WMULT_IDLEPRIO;
758 return;
759 }
760
761 load->weight = scale_load(prio_to_weight[prio]);
762 load->inv_weight = prio_to_wmult[prio];
763}
764
765static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
766{
767 update_rq_clock(rq);
768 sched_info_queued(rq, p);
769 p->sched_class->enqueue_task(rq, p, flags);
770}
771
772static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
773{
774 update_rq_clock(rq);
775 sched_info_dequeued(rq, p);
776 p->sched_class->dequeue_task(rq, p, flags);
777}
778
779void activate_task(struct rq *rq, struct task_struct *p, int flags)
780{
781 if (task_contributes_to_load(p))
782 rq->nr_uninterruptible--;
783
784 enqueue_task(rq, p, flags);
785}
786
787void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
788{
789 if (task_contributes_to_load(p))
790 rq->nr_uninterruptible++;
791
792 dequeue_task(rq, p, flags);
793}
794
795static void update_rq_clock_task(struct rq *rq, s64 delta)
796{
797/*
798 * In theory, the compile should just see 0 here, and optimize out the call
799 * to sched_rt_avg_update. But I don't trust it...
800 */
801#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
802 s64 steal = 0, irq_delta = 0;
803#endif
804#ifdef CONFIG_IRQ_TIME_ACCOUNTING
805 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
806
807 /*
808 * Since irq_time is only updated on {soft,}irq_exit, we might run into
809 * this case when a previous update_rq_clock() happened inside a
810 * {soft,}irq region.
811 *
812 * When this happens, we stop ->clock_task and only update the
813 * prev_irq_time stamp to account for the part that fit, so that a next
814 * update will consume the rest. This ensures ->clock_task is
815 * monotonic.
816 *
817 * It does however cause some slight miss-attribution of {soft,}irq
818 * time, a more accurate solution would be to update the irq_time using
819 * the current rq->clock timestamp, except that would require using
820 * atomic ops.
821 */
822 if (irq_delta > delta)
823 irq_delta = delta;
824
825 rq->prev_irq_time += irq_delta;
826 delta -= irq_delta;
827#endif
828#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
829 if (static_key_false((¶virt_steal_rq_enabled))) {
830 steal = paravirt_steal_clock(cpu_of(rq));
831 steal -= rq->prev_steal_time_rq;
832
833 if (unlikely(steal > delta))
834 steal = delta;
835
836 rq->prev_steal_time_rq += steal;
837 delta -= steal;
838 }
839#endif
840
841 rq->clock_task += delta;
842
843#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
844 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
845 sched_rt_avg_update(rq, irq_delta + steal);
846#endif
847}
848
849void sched_set_stop_task(int cpu, struct task_struct *stop)
850{
851 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
852 struct task_struct *old_stop = cpu_rq(cpu)->stop;
853
854 if (stop) {
855 /*
856 * Make it appear like a SCHED_FIFO task, its something
857 * userspace knows about and won't get confused about.
858 *
859 * Also, it will make PI more or less work without too
860 * much confusion -- but then, stop work should not
861 * rely on PI working anyway.
862 */
863 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
864
865 stop->sched_class = &stop_sched_class;
866 }
867
868 cpu_rq(cpu)->stop = stop;
869
870 if (old_stop) {
871 /*
872 * Reset it back to a normal scheduling class so that
873 * it can die in pieces.
874 */
875 old_stop->sched_class = &rt_sched_class;
876 }
877}
878
879/*
880 * __normal_prio - return the priority that is based on the static prio
881 */
882static inline int __normal_prio(struct task_struct *p)
883{
884 return p->static_prio;
885}
886
887/*
888 * Calculate the expected normal priority: i.e. priority
889 * without taking RT-inheritance into account. Might be
890 * boosted by interactivity modifiers. Changes upon fork,
891 * setprio syscalls, and whenever the interactivity
892 * estimator recalculates.
893 */
894static inline int normal_prio(struct task_struct *p)
895{
896 int prio;
897
898 if (task_has_dl_policy(p))
899 prio = MAX_DL_PRIO-1;
900 else if (task_has_rt_policy(p))
901 prio = MAX_RT_PRIO-1 - p->rt_priority;
902 else
903 prio = __normal_prio(p);
904 return prio;
905}
906
907/*
908 * Calculate the current priority, i.e. the priority
909 * taken into account by the scheduler. This value might
910 * be boosted by RT tasks, or might be boosted by
911 * interactivity modifiers. Will be RT if the task got
912 * RT-boosted. If not then it returns p->normal_prio.
913 */
914static int effective_prio(struct task_struct *p)
915{
916 p->normal_prio = normal_prio(p);
917 /*
918 * If we are RT tasks or we were boosted to RT priority,
919 * keep the priority unchanged. Otherwise, update priority
920 * to the normal priority:
921 */
922 if (!rt_prio(p->prio))
923 return p->normal_prio;
924 return p->prio;
925}
926
927/**
928 * task_curr - is this task currently executing on a CPU?
929 * @p: the task in question.
930 *
931 * Return: 1 if the task is currently executing. 0 otherwise.
932 */
933inline int task_curr(const struct task_struct *p)
934{
935 return cpu_curr(task_cpu(p)) == p;
936}
937
938static inline void check_class_changed(struct rq *rq, struct task_struct *p,
939 const struct sched_class *prev_class,
940 int oldprio)
941{
942 if (prev_class != p->sched_class) {
943 if (prev_class->switched_from)
944 prev_class->switched_from(rq, p);
945 p->sched_class->switched_to(rq, p);
946 } else if (oldprio != p->prio || dl_task(p))
947 p->sched_class->prio_changed(rq, p, oldprio);
948}
949
950void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
951{
952 const struct sched_class *class;
953
954 if (p->sched_class == rq->curr->sched_class) {
955 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
956 } else {
957 for_each_class(class) {
958 if (class == rq->curr->sched_class)
959 break;
960 if (class == p->sched_class) {
961 resched_task(rq->curr);
962 break;
963 }
964 }
965 }
966
967 /*
968 * A queue event has occurred, and we're going to schedule. In
969 * this case, we can save a useless back to back clock update.
970 */
971 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
972 rq->skip_clock_update = 1;
973}
974
975#ifdef CONFIG_SMP
976void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
977{
978#ifdef CONFIG_SCHED_DEBUG
979 /*
980 * We should never call set_task_cpu() on a blocked task,
981 * ttwu() will sort out the placement.
982 */
983 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
984 !(task_preempt_count(p) & PREEMPT_ACTIVE));
985
986#ifdef CONFIG_LOCKDEP
987 /*
988 * The caller should hold either p->pi_lock or rq->lock, when changing
989 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
990 *
991 * sched_move_task() holds both and thus holding either pins the cgroup,
992 * see task_group().
993 *
994 * Furthermore, all task_rq users should acquire both locks, see
995 * task_rq_lock().
996 */
997 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
998 lockdep_is_held(&task_rq(p)->lock)));
999#endif
1000#endif
1001
1002 trace_sched_migrate_task(p, new_cpu);
1003
1004 if (task_cpu(p) != new_cpu) {
1005 if (p->sched_class->migrate_task_rq)
1006 p->sched_class->migrate_task_rq(p, new_cpu);
1007 p->se.nr_migrations++;
1008 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1009 }
1010
1011 __set_task_cpu(p, new_cpu);
1012}
1013
1014static void __migrate_swap_task(struct task_struct *p, int cpu)
1015{
1016 if (p->on_rq) {
1017 struct rq *src_rq, *dst_rq;
1018
1019 src_rq = task_rq(p);
1020 dst_rq = cpu_rq(cpu);
1021
1022 deactivate_task(src_rq, p, 0);
1023 set_task_cpu(p, cpu);
1024 activate_task(dst_rq, p, 0);
1025 check_preempt_curr(dst_rq, p, 0);
1026 } else {
1027 /*
1028 * Task isn't running anymore; make it appear like we migrated
1029 * it before it went to sleep. This means on wakeup we make the
1030 * previous cpu our targer instead of where it really is.
1031 */
1032 p->wake_cpu = cpu;
1033 }
1034}
1035
1036struct migration_swap_arg {
1037 struct task_struct *src_task, *dst_task;
1038 int src_cpu, dst_cpu;
1039};
1040
1041static int migrate_swap_stop(void *data)
1042{
1043 struct migration_swap_arg *arg = data;
1044 struct rq *src_rq, *dst_rq;
1045 int ret = -EAGAIN;
1046
1047 src_rq = cpu_rq(arg->src_cpu);
1048 dst_rq = cpu_rq(arg->dst_cpu);
1049
1050 double_raw_lock(&arg->src_task->pi_lock,
1051 &arg->dst_task->pi_lock);
1052 double_rq_lock(src_rq, dst_rq);
1053 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1054 goto unlock;
1055
1056 if (task_cpu(arg->src_task) != arg->src_cpu)
1057 goto unlock;
1058
1059 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1060 goto unlock;
1061
1062 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1063 goto unlock;
1064
1065 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1066 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1067
1068 ret = 0;
1069
1070unlock:
1071 double_rq_unlock(src_rq, dst_rq);
1072 raw_spin_unlock(&arg->dst_task->pi_lock);
1073 raw_spin_unlock(&arg->src_task->pi_lock);
1074
1075 return ret;
1076}
1077
1078/*
1079 * Cross migrate two tasks
1080 */
1081int migrate_swap(struct task_struct *cur, struct task_struct *p)
1082{
1083 struct migration_swap_arg arg;
1084 int ret = -EINVAL;
1085
1086 arg = (struct migration_swap_arg){
1087 .src_task = cur,
1088 .src_cpu = task_cpu(cur),
1089 .dst_task = p,
1090 .dst_cpu = task_cpu(p),
1091 };
1092
1093 if (arg.src_cpu == arg.dst_cpu)
1094 goto out;
1095
1096 /*
1097 * These three tests are all lockless; this is OK since all of them
1098 * will be re-checked with proper locks held further down the line.
1099 */
1100 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1101 goto out;
1102
1103 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1104 goto out;
1105
1106 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1107 goto out;
1108
1109 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1110 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1111
1112out:
1113 return ret;
1114}
1115
1116struct migration_arg {
1117 struct task_struct *task;
1118 int dest_cpu;
1119};
1120
1121static int migration_cpu_stop(void *data);
1122
1123/*
1124 * wait_task_inactive - wait for a thread to unschedule.
1125 *
1126 * If @match_state is nonzero, it's the @p->state value just checked and
1127 * not expected to change. If it changes, i.e. @p might have woken up,
1128 * then return zero. When we succeed in waiting for @p to be off its CPU,
1129 * we return a positive number (its total switch count). If a second call
1130 * a short while later returns the same number, the caller can be sure that
1131 * @p has remained unscheduled the whole time.
1132 *
1133 * The caller must ensure that the task *will* unschedule sometime soon,
1134 * else this function might spin for a *long* time. This function can't
1135 * be called with interrupts off, or it may introduce deadlock with
1136 * smp_call_function() if an IPI is sent by the same process we are
1137 * waiting to become inactive.
1138 */
1139unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1140{
1141 unsigned long flags;
1142 int running, on_rq;
1143 unsigned long ncsw;
1144 struct rq *rq;
1145
1146 for (;;) {
1147 /*
1148 * We do the initial early heuristics without holding
1149 * any task-queue locks at all. We'll only try to get
1150 * the runqueue lock when things look like they will
1151 * work out!
1152 */
1153 rq = task_rq(p);
1154
1155 /*
1156 * If the task is actively running on another CPU
1157 * still, just relax and busy-wait without holding
1158 * any locks.
1159 *
1160 * NOTE! Since we don't hold any locks, it's not
1161 * even sure that "rq" stays as the right runqueue!
1162 * But we don't care, since "task_running()" will
1163 * return false if the runqueue has changed and p
1164 * is actually now running somewhere else!
1165 */
1166 while (task_running(rq, p)) {
1167 if (match_state && unlikely(p->state != match_state))
1168 return 0;
1169 cpu_relax();
1170 }
1171
1172 /*
1173 * Ok, time to look more closely! We need the rq
1174 * lock now, to be *sure*. If we're wrong, we'll
1175 * just go back and repeat.
1176 */
1177 rq = task_rq_lock(p, &flags);
1178 trace_sched_wait_task(p);
1179 running = task_running(rq, p);
1180 on_rq = p->on_rq;
1181 ncsw = 0;
1182 if (!match_state || p->state == match_state)
1183 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1184 task_rq_unlock(rq, p, &flags);
1185
1186 /*
1187 * If it changed from the expected state, bail out now.
1188 */
1189 if (unlikely(!ncsw))
1190 break;
1191
1192 /*
1193 * Was it really running after all now that we
1194 * checked with the proper locks actually held?
1195 *
1196 * Oops. Go back and try again..
1197 */
1198 if (unlikely(running)) {
1199 cpu_relax();
1200 continue;
1201 }
1202
1203 /*
1204 * It's not enough that it's not actively running,
1205 * it must be off the runqueue _entirely_, and not
1206 * preempted!
1207 *
1208 * So if it was still runnable (but just not actively
1209 * running right now), it's preempted, and we should
1210 * yield - it could be a while.
1211 */
1212 if (unlikely(on_rq)) {
1213 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1214
1215 set_current_state(TASK_UNINTERRUPTIBLE);
1216 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1217 continue;
1218 }
1219
1220 /*
1221 * Ahh, all good. It wasn't running, and it wasn't
1222 * runnable, which means that it will never become
1223 * running in the future either. We're all done!
1224 */
1225 break;
1226 }
1227
1228 return ncsw;
1229}
1230
1231/***
1232 * kick_process - kick a running thread to enter/exit the kernel
1233 * @p: the to-be-kicked thread
1234 *
1235 * Cause a process which is running on another CPU to enter
1236 * kernel-mode, without any delay. (to get signals handled.)
1237 *
1238 * NOTE: this function doesn't have to take the runqueue lock,
1239 * because all it wants to ensure is that the remote task enters
1240 * the kernel. If the IPI races and the task has been migrated
1241 * to another CPU then no harm is done and the purpose has been
1242 * achieved as well.
1243 */
1244void kick_process(struct task_struct *p)
1245{
1246 int cpu;
1247
1248 preempt_disable();
1249 cpu = task_cpu(p);
1250 if ((cpu != smp_processor_id()) && task_curr(p))
1251 smp_send_reschedule(cpu);
1252 preempt_enable();
1253}
1254EXPORT_SYMBOL_GPL(kick_process);
1255#endif /* CONFIG_SMP */
1256
1257#ifdef CONFIG_SMP
1258/*
1259 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1260 */
1261static int select_fallback_rq(int cpu, struct task_struct *p)
1262{
1263 int nid = cpu_to_node(cpu);
1264 const struct cpumask *nodemask = NULL;
1265 enum { cpuset, possible, fail } state = cpuset;
1266 int dest_cpu;
1267
1268 /*
1269 * If the node that the cpu is on has been offlined, cpu_to_node()
1270 * will return -1. There is no cpu on the node, and we should
1271 * select the cpu on the other node.
1272 */
1273 if (nid != -1) {
1274 nodemask = cpumask_of_node(nid);
1275
1276 /* Look for allowed, online CPU in same node. */
1277 for_each_cpu(dest_cpu, nodemask) {
1278 if (!cpu_online(dest_cpu))
1279 continue;
1280 if (!cpu_active(dest_cpu))
1281 continue;
1282 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1283 return dest_cpu;
1284 }
1285 }
1286
1287 for (;;) {
1288 /* Any allowed, online CPU? */
1289 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1290 if (!cpu_online(dest_cpu))
1291 continue;
1292 if (!cpu_active(dest_cpu))
1293 continue;
1294 goto out;
1295 }
1296
1297 switch (state) {
1298 case cpuset:
1299 /* No more Mr. Nice Guy. */
1300 cpuset_cpus_allowed_fallback(p);
1301 state = possible;
1302 break;
1303
1304 case possible:
1305 do_set_cpus_allowed(p, cpu_possible_mask);
1306 state = fail;
1307 break;
1308
1309 case fail:
1310 BUG();
1311 break;
1312 }
1313 }
1314
1315out:
1316 if (state != cpuset) {
1317 /*
1318 * Don't tell them about moving exiting tasks or
1319 * kernel threads (both mm NULL), since they never
1320 * leave kernel.
1321 */
1322 if (p->mm && printk_ratelimit()) {
1323 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1324 task_pid_nr(p), p->comm, cpu);
1325 }
1326 }
1327
1328 return dest_cpu;
1329}
1330
1331/*
1332 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1333 */
1334static inline
1335int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1336{
1337 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1338
1339 /*
1340 * In order not to call set_task_cpu() on a blocking task we need
1341 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1342 * cpu.
1343 *
1344 * Since this is common to all placement strategies, this lives here.
1345 *
1346 * [ this allows ->select_task() to simply return task_cpu(p) and
1347 * not worry about this generic constraint ]
1348 */
1349 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1350 !cpu_online(cpu)))
1351 cpu = select_fallback_rq(task_cpu(p), p);
1352
1353 return cpu;
1354}
1355
1356static void update_avg(u64 *avg, u64 sample)
1357{
1358 s64 diff = sample - *avg;
1359 *avg += diff >> 3;
1360}
1361#endif
1362
1363static void
1364ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1365{
1366#ifdef CONFIG_SCHEDSTATS
1367 struct rq *rq = this_rq();
1368
1369#ifdef CONFIG_SMP
1370 int this_cpu = smp_processor_id();
1371
1372 if (cpu == this_cpu) {
1373 schedstat_inc(rq, ttwu_local);
1374 schedstat_inc(p, se.statistics.nr_wakeups_local);
1375 } else {
1376 struct sched_domain *sd;
1377
1378 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1379 rcu_read_lock();
1380 for_each_domain(this_cpu, sd) {
1381 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1382 schedstat_inc(sd, ttwu_wake_remote);
1383 break;
1384 }
1385 }
1386 rcu_read_unlock();
1387 }
1388
1389 if (wake_flags & WF_MIGRATED)
1390 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1391
1392#endif /* CONFIG_SMP */
1393
1394 schedstat_inc(rq, ttwu_count);
1395 schedstat_inc(p, se.statistics.nr_wakeups);
1396
1397 if (wake_flags & WF_SYNC)
1398 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1399
1400#endif /* CONFIG_SCHEDSTATS */
1401}
1402
1403static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1404{
1405 activate_task(rq, p, en_flags);
1406 p->on_rq = 1;
1407
1408 /* if a worker is waking up, notify workqueue */
1409 if (p->flags & PF_WQ_WORKER)
1410 wq_worker_waking_up(p, cpu_of(rq));
1411}
1412
1413/*
1414 * Mark the task runnable and perform wakeup-preemption.
1415 */
1416static void
1417ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1418{
1419 check_preempt_curr(rq, p, wake_flags);
1420 trace_sched_wakeup(p, true);
1421
1422 p->state = TASK_RUNNING;
1423#ifdef CONFIG_SMP
1424 if (p->sched_class->task_woken)
1425 p->sched_class->task_woken(rq, p);
1426
1427 if (rq->idle_stamp) {
1428 u64 delta = rq_clock(rq) - rq->idle_stamp;
1429 u64 max = 2*rq->max_idle_balance_cost;
1430
1431 update_avg(&rq->avg_idle, delta);
1432
1433 if (rq->avg_idle > max)
1434 rq->avg_idle = max;
1435
1436 rq->idle_stamp = 0;
1437 }
1438#endif
1439}
1440
1441static void
1442ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1443{
1444#ifdef CONFIG_SMP
1445 if (p->sched_contributes_to_load)
1446 rq->nr_uninterruptible--;
1447#endif
1448
1449 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1450 ttwu_do_wakeup(rq, p, wake_flags);
1451}
1452
1453/*
1454 * Called in case the task @p isn't fully descheduled from its runqueue,
1455 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1456 * since all we need to do is flip p->state to TASK_RUNNING, since
1457 * the task is still ->on_rq.
1458 */
1459static int ttwu_remote(struct task_struct *p, int wake_flags)
1460{
1461 struct rq *rq;
1462 int ret = 0;
1463
1464 rq = __task_rq_lock(p);
1465 if (p->on_rq) {
1466 /* check_preempt_curr() may use rq clock */
1467 update_rq_clock(rq);
1468 ttwu_do_wakeup(rq, p, wake_flags);
1469 ret = 1;
1470 }
1471 __task_rq_unlock(rq);
1472
1473 return ret;
1474}
1475
1476#ifdef CONFIG_SMP
1477static void sched_ttwu_pending(void)
1478{
1479 struct rq *rq = this_rq();
1480 struct llist_node *llist = llist_del_all(&rq->wake_list);
1481 struct task_struct *p;
1482
1483 raw_spin_lock(&rq->lock);
1484
1485 while (llist) {
1486 p = llist_entry(llist, struct task_struct, wake_entry);
1487 llist = llist_next(llist);
1488 ttwu_do_activate(rq, p, 0);
1489 }
1490
1491 raw_spin_unlock(&rq->lock);
1492}
1493
1494void scheduler_ipi(void)
1495{
1496 /*
1497 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1498 * TIF_NEED_RESCHED remotely (for the first time) will also send
1499 * this IPI.
1500 */
1501 preempt_fold_need_resched();
1502
1503 if (llist_empty(&this_rq()->wake_list)
1504 && !tick_nohz_full_cpu(smp_processor_id())
1505 && !got_nohz_idle_kick())
1506 return;
1507
1508 /*
1509 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1510 * traditionally all their work was done from the interrupt return
1511 * path. Now that we actually do some work, we need to make sure
1512 * we do call them.
1513 *
1514 * Some archs already do call them, luckily irq_enter/exit nest
1515 * properly.
1516 *
1517 * Arguably we should visit all archs and update all handlers,
1518 * however a fair share of IPIs are still resched only so this would
1519 * somewhat pessimize the simple resched case.
1520 */
1521 irq_enter();
1522 tick_nohz_full_check();
1523 sched_ttwu_pending();
1524
1525 /*
1526 * Check if someone kicked us for doing the nohz idle load balance.
1527 */
1528 if (unlikely(got_nohz_idle_kick())) {
1529 this_rq()->idle_balance = 1;
1530 raise_softirq_irqoff(SCHED_SOFTIRQ);
1531 }
1532 irq_exit();
1533}
1534
1535static void ttwu_queue_remote(struct task_struct *p, int cpu)
1536{
1537 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1538 smp_send_reschedule(cpu);
1539}
1540
1541bool cpus_share_cache(int this_cpu, int that_cpu)
1542{
1543 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1544}
1545#endif /* CONFIG_SMP */
1546
1547static void ttwu_queue(struct task_struct *p, int cpu)
1548{
1549 struct rq *rq = cpu_rq(cpu);
1550
1551#if defined(CONFIG_SMP)
1552 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1553 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1554 ttwu_queue_remote(p, cpu);
1555 return;
1556 }
1557#endif
1558
1559 raw_spin_lock(&rq->lock);
1560 ttwu_do_activate(rq, p, 0);
1561 raw_spin_unlock(&rq->lock);
1562}
1563
1564/**
1565 * try_to_wake_up - wake up a thread
1566 * @p: the thread to be awakened
1567 * @state: the mask of task states that can be woken
1568 * @wake_flags: wake modifier flags (WF_*)
1569 *
1570 * Put it on the run-queue if it's not already there. The "current"
1571 * thread is always on the run-queue (except when the actual
1572 * re-schedule is in progress), and as such you're allowed to do
1573 * the simpler "current->state = TASK_RUNNING" to mark yourself
1574 * runnable without the overhead of this.
1575 *
1576 * Return: %true if @p was woken up, %false if it was already running.
1577 * or @state didn't match @p's state.
1578 */
1579static int
1580try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1581{
1582 unsigned long flags;
1583 int cpu, success = 0;
1584
1585 /*
1586 * If we are going to wake up a thread waiting for CONDITION we
1587 * need to ensure that CONDITION=1 done by the caller can not be
1588 * reordered with p->state check below. This pairs with mb() in
1589 * set_current_state() the waiting thread does.
1590 */
1591 smp_mb__before_spinlock();
1592 raw_spin_lock_irqsave(&p->pi_lock, flags);
1593 if (!(p->state & state))
1594 goto out;
1595
1596 success = 1; /* we're going to change ->state */
1597 cpu = task_cpu(p);
1598
1599 if (p->on_rq && ttwu_remote(p, wake_flags))
1600 goto stat;
1601
1602#ifdef CONFIG_SMP
1603 /*
1604 * If the owning (remote) cpu is still in the middle of schedule() with
1605 * this task as prev, wait until its done referencing the task.
1606 */
1607 while (p->on_cpu)
1608 cpu_relax();
1609 /*
1610 * Pairs with the smp_wmb() in finish_lock_switch().
1611 */
1612 smp_rmb();
1613
1614 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1615 p->state = TASK_WAKING;
1616
1617 if (p->sched_class->task_waking)
1618 p->sched_class->task_waking(p);
1619
1620 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1621 if (task_cpu(p) != cpu) {
1622 wake_flags |= WF_MIGRATED;
1623 set_task_cpu(p, cpu);
1624 }
1625#endif /* CONFIG_SMP */
1626
1627 ttwu_queue(p, cpu);
1628stat:
1629 ttwu_stat(p, cpu, wake_flags);
1630out:
1631 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1632
1633 return success;
1634}
1635
1636/**
1637 * try_to_wake_up_local - try to wake up a local task with rq lock held
1638 * @p: the thread to be awakened
1639 *
1640 * Put @p on the run-queue if it's not already there. The caller must
1641 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1642 * the current task.
1643 */
1644static void try_to_wake_up_local(struct task_struct *p)
1645{
1646 struct rq *rq = task_rq(p);
1647
1648 if (WARN_ON_ONCE(rq != this_rq()) ||
1649 WARN_ON_ONCE(p == current))
1650 return;
1651
1652 lockdep_assert_held(&rq->lock);
1653
1654 if (!raw_spin_trylock(&p->pi_lock)) {
1655 raw_spin_unlock(&rq->lock);
1656 raw_spin_lock(&p->pi_lock);
1657 raw_spin_lock(&rq->lock);
1658 }
1659
1660 if (!(p->state & TASK_NORMAL))
1661 goto out;
1662
1663 if (!p->on_rq)
1664 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1665
1666 ttwu_do_wakeup(rq, p, 0);
1667 ttwu_stat(p, smp_processor_id(), 0);
1668out:
1669 raw_spin_unlock(&p->pi_lock);
1670}
1671
1672/**
1673 * wake_up_process - Wake up a specific process
1674 * @p: The process to be woken up.
1675 *
1676 * Attempt to wake up the nominated process and move it to the set of runnable
1677 * processes.
1678 *
1679 * Return: 1 if the process was woken up, 0 if it was already running.
1680 *
1681 * It may be assumed that this function implies a write memory barrier before
1682 * changing the task state if and only if any tasks are woken up.
1683 */
1684int wake_up_process(struct task_struct *p)
1685{
1686 WARN_ON(task_is_stopped_or_traced(p));
1687 return try_to_wake_up(p, TASK_NORMAL, 0);
1688}
1689EXPORT_SYMBOL(wake_up_process);
1690
1691int wake_up_state(struct task_struct *p, unsigned int state)
1692{
1693 return try_to_wake_up(p, state, 0);
1694}
1695
1696/*
1697 * Perform scheduler related setup for a newly forked process p.
1698 * p is forked by current.
1699 *
1700 * __sched_fork() is basic setup used by init_idle() too:
1701 */
1702static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1703{
1704 p->on_rq = 0;
1705
1706 p->se.on_rq = 0;
1707 p->se.exec_start = 0;
1708 p->se.sum_exec_runtime = 0;
1709 p->se.prev_sum_exec_runtime = 0;
1710 p->se.nr_migrations = 0;
1711 p->se.vruntime = 0;
1712 INIT_LIST_HEAD(&p->se.group_node);
1713
1714#ifdef CONFIG_SCHEDSTATS
1715 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1716#endif
1717
1718 RB_CLEAR_NODE(&p->dl.rb_node);
1719 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1720 p->dl.dl_runtime = p->dl.runtime = 0;
1721 p->dl.dl_deadline = p->dl.deadline = 0;
1722 p->dl.dl_period = 0;
1723 p->dl.flags = 0;
1724
1725 INIT_LIST_HEAD(&p->rt.run_list);
1726
1727#ifdef CONFIG_PREEMPT_NOTIFIERS
1728 INIT_HLIST_HEAD(&p->preempt_notifiers);
1729#endif
1730
1731#ifdef CONFIG_NUMA_BALANCING
1732 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1733 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1734 p->mm->numa_scan_seq = 0;
1735 }
1736
1737 if (clone_flags & CLONE_VM)
1738 p->numa_preferred_nid = current->numa_preferred_nid;
1739 else
1740 p->numa_preferred_nid = -1;
1741
1742 p->node_stamp = 0ULL;
1743 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1744 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1745 p->numa_work.next = &p->numa_work;
1746 p->numa_faults_memory = NULL;
1747 p->numa_faults_buffer_memory = NULL;
1748 p->last_task_numa_placement = 0;
1749 p->last_sum_exec_runtime = 0;
1750
1751 INIT_LIST_HEAD(&p->numa_entry);
1752 p->numa_group = NULL;
1753#endif /* CONFIG_NUMA_BALANCING */
1754}
1755
1756#ifdef CONFIG_NUMA_BALANCING
1757#ifdef CONFIG_SCHED_DEBUG
1758void set_numabalancing_state(bool enabled)
1759{
1760 if (enabled)
1761 sched_feat_set("NUMA");
1762 else
1763 sched_feat_set("NO_NUMA");
1764}
1765#else
1766__read_mostly bool numabalancing_enabled;
1767
1768void set_numabalancing_state(bool enabled)
1769{
1770 numabalancing_enabled = enabled;
1771}
1772#endif /* CONFIG_SCHED_DEBUG */
1773
1774#ifdef CONFIG_PROC_SYSCTL
1775int sysctl_numa_balancing(struct ctl_table *table, int write,
1776 void __user *buffer, size_t *lenp, loff_t *ppos)
1777{
1778 struct ctl_table t;
1779 int err;
1780 int state = numabalancing_enabled;
1781
1782 if (write && !capable(CAP_SYS_ADMIN))
1783 return -EPERM;
1784
1785 t = *table;
1786 t.data = &state;
1787 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1788 if (err < 0)
1789 return err;
1790 if (write)
1791 set_numabalancing_state(state);
1792 return err;
1793}
1794#endif
1795#endif
1796
1797/*
1798 * fork()/clone()-time setup:
1799 */
1800int sched_fork(unsigned long clone_flags, struct task_struct *p)
1801{
1802 unsigned long flags;
1803 int cpu = get_cpu();
1804
1805 __sched_fork(clone_flags, p);
1806 /*
1807 * We mark the process as running here. This guarantees that
1808 * nobody will actually run it, and a signal or other external
1809 * event cannot wake it up and insert it on the runqueue either.
1810 */
1811 p->state = TASK_RUNNING;
1812
1813 /*
1814 * Make sure we do not leak PI boosting priority to the child.
1815 */
1816 p->prio = current->normal_prio;
1817
1818 /*
1819 * Revert to default priority/policy on fork if requested.
1820 */
1821 if (unlikely(p->sched_reset_on_fork)) {
1822 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1823 p->policy = SCHED_NORMAL;
1824 p->static_prio = NICE_TO_PRIO(0);
1825 p->rt_priority = 0;
1826 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1827 p->static_prio = NICE_TO_PRIO(0);
1828
1829 p->prio = p->normal_prio = __normal_prio(p);
1830 set_load_weight(p);
1831
1832 /*
1833 * We don't need the reset flag anymore after the fork. It has
1834 * fulfilled its duty:
1835 */
1836 p->sched_reset_on_fork = 0;
1837 }
1838
1839 if (dl_prio(p->prio)) {
1840 put_cpu();
1841 return -EAGAIN;
1842 } else if (rt_prio(p->prio)) {
1843 p->sched_class = &rt_sched_class;
1844 } else {
1845 p->sched_class = &fair_sched_class;
1846 }
1847
1848 if (p->sched_class->task_fork)
1849 p->sched_class->task_fork(p);
1850
1851 /*
1852 * The child is not yet in the pid-hash so no cgroup attach races,
1853 * and the cgroup is pinned to this child due to cgroup_fork()
1854 * is ran before sched_fork().
1855 *
1856 * Silence PROVE_RCU.
1857 */
1858 raw_spin_lock_irqsave(&p->pi_lock, flags);
1859 set_task_cpu(p, cpu);
1860 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1861
1862#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1863 if (likely(sched_info_on()))
1864 memset(&p->sched_info, 0, sizeof(p->sched_info));
1865#endif
1866#if defined(CONFIG_SMP)
1867 p->on_cpu = 0;
1868#endif
1869 init_task_preempt_count(p);
1870#ifdef CONFIG_SMP
1871 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1872 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1873#endif
1874
1875 put_cpu();
1876 return 0;
1877}
1878
1879unsigned long to_ratio(u64 period, u64 runtime)
1880{
1881 if (runtime == RUNTIME_INF)
1882 return 1ULL << 20;
1883
1884 /*
1885 * Doing this here saves a lot of checks in all
1886 * the calling paths, and returning zero seems
1887 * safe for them anyway.
1888 */
1889 if (period == 0)
1890 return 0;
1891
1892 return div64_u64(runtime << 20, period);
1893}
1894
1895#ifdef CONFIG_SMP
1896inline struct dl_bw *dl_bw_of(int i)
1897{
1898 return &cpu_rq(i)->rd->dl_bw;
1899}
1900
1901static inline int dl_bw_cpus(int i)
1902{
1903 struct root_domain *rd = cpu_rq(i)->rd;
1904 int cpus = 0;
1905
1906 for_each_cpu_and(i, rd->span, cpu_active_mask)
1907 cpus++;
1908
1909 return cpus;
1910}
1911#else
1912inline struct dl_bw *dl_bw_of(int i)
1913{
1914 return &cpu_rq(i)->dl.dl_bw;
1915}
1916
1917static inline int dl_bw_cpus(int i)
1918{
1919 return 1;
1920}
1921#endif
1922
1923static inline
1924void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
1925{
1926 dl_b->total_bw -= tsk_bw;
1927}
1928
1929static inline
1930void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
1931{
1932 dl_b->total_bw += tsk_bw;
1933}
1934
1935static inline
1936bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
1937{
1938 return dl_b->bw != -1 &&
1939 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
1940}
1941
1942/*
1943 * We must be sure that accepting a new task (or allowing changing the
1944 * parameters of an existing one) is consistent with the bandwidth
1945 * constraints. If yes, this function also accordingly updates the currently
1946 * allocated bandwidth to reflect the new situation.
1947 *
1948 * This function is called while holding p's rq->lock.
1949 */
1950static int dl_overflow(struct task_struct *p, int policy,
1951 const struct sched_attr *attr)
1952{
1953
1954 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1955 u64 period = attr->sched_period ?: attr->sched_deadline;
1956 u64 runtime = attr->sched_runtime;
1957 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
1958 int cpus, err = -1;
1959
1960 if (new_bw == p->dl.dl_bw)
1961 return 0;
1962
1963 /*
1964 * Either if a task, enters, leave, or stays -deadline but changes
1965 * its parameters, we may need to update accordingly the total
1966 * allocated bandwidth of the container.
1967 */
1968 raw_spin_lock(&dl_b->lock);
1969 cpus = dl_bw_cpus(task_cpu(p));
1970 if (dl_policy(policy) && !task_has_dl_policy(p) &&
1971 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
1972 __dl_add(dl_b, new_bw);
1973 err = 0;
1974 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
1975 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
1976 __dl_clear(dl_b, p->dl.dl_bw);
1977 __dl_add(dl_b, new_bw);
1978 err = 0;
1979 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
1980 __dl_clear(dl_b, p->dl.dl_bw);
1981 err = 0;
1982 }
1983 raw_spin_unlock(&dl_b->lock);
1984
1985 return err;
1986}
1987
1988extern void init_dl_bw(struct dl_bw *dl_b);
1989
1990/*
1991 * wake_up_new_task - wake up a newly created task for the first time.
1992 *
1993 * This function will do some initial scheduler statistics housekeeping
1994 * that must be done for every newly created context, then puts the task
1995 * on the runqueue and wakes it.
1996 */
1997void wake_up_new_task(struct task_struct *p)
1998{
1999 unsigned long flags;
2000 struct rq *rq;
2001
2002 raw_spin_lock_irqsave(&p->pi_lock, flags);
2003#ifdef CONFIG_SMP
2004 /*
2005 * Fork balancing, do it here and not earlier because:
2006 * - cpus_allowed can change in the fork path
2007 * - any previously selected cpu might disappear through hotplug
2008 */
2009 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2010#endif
2011
2012 /* Initialize new task's runnable average */
2013 init_task_runnable_average(p);
2014 rq = __task_rq_lock(p);
2015 activate_task(rq, p, 0);
2016 p->on_rq = 1;
2017 trace_sched_wakeup_new(p, true);
2018 check_preempt_curr(rq, p, WF_FORK);
2019#ifdef CONFIG_SMP
2020 if (p->sched_class->task_woken)
2021 p->sched_class->task_woken(rq, p);
2022#endif
2023 task_rq_unlock(rq, p, &flags);
2024}
2025
2026#ifdef CONFIG_PREEMPT_NOTIFIERS
2027
2028/**
2029 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2030 * @notifier: notifier struct to register
2031 */
2032void preempt_notifier_register(struct preempt_notifier *notifier)
2033{
2034 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2035}
2036EXPORT_SYMBOL_GPL(preempt_notifier_register);
2037
2038/**
2039 * preempt_notifier_unregister - no longer interested in preemption notifications
2040 * @notifier: notifier struct to unregister
2041 *
2042 * This is safe to call from within a preemption notifier.
2043 */
2044void preempt_notifier_unregister(struct preempt_notifier *notifier)
2045{
2046 hlist_del(¬ifier->link);
2047}
2048EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2049
2050static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2051{
2052 struct preempt_notifier *notifier;
2053
2054 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2055 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2056}
2057
2058static void
2059fire_sched_out_preempt_notifiers(struct task_struct *curr,
2060 struct task_struct *next)
2061{
2062 struct preempt_notifier *notifier;
2063
2064 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2065 notifier->ops->sched_out(notifier, next);
2066}
2067
2068#else /* !CONFIG_PREEMPT_NOTIFIERS */
2069
2070static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2071{
2072}
2073
2074static void
2075fire_sched_out_preempt_notifiers(struct task_struct *curr,
2076 struct task_struct *next)
2077{
2078}
2079
2080#endif /* CONFIG_PREEMPT_NOTIFIERS */
2081
2082/**
2083 * prepare_task_switch - prepare to switch tasks
2084 * @rq: the runqueue preparing to switch
2085 * @prev: the current task that is being switched out
2086 * @next: the task we are going to switch to.
2087 *
2088 * This is called with the rq lock held and interrupts off. It must
2089 * be paired with a subsequent finish_task_switch after the context
2090 * switch.
2091 *
2092 * prepare_task_switch sets up locking and calls architecture specific
2093 * hooks.
2094 */
2095static inline void
2096prepare_task_switch(struct rq *rq, struct task_struct *prev,
2097 struct task_struct *next)
2098{
2099 trace_sched_switch(prev, next);
2100 sched_info_switch(rq, prev, next);
2101 perf_event_task_sched_out(prev, next);
2102 fire_sched_out_preempt_notifiers(prev, next);
2103 prepare_lock_switch(rq, next);
2104 prepare_arch_switch(next);
2105}
2106
2107/**
2108 * finish_task_switch - clean up after a task-switch
2109 * @rq: runqueue associated with task-switch
2110 * @prev: the thread we just switched away from.
2111 *
2112 * finish_task_switch must be called after the context switch, paired
2113 * with a prepare_task_switch call before the context switch.
2114 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2115 * and do any other architecture-specific cleanup actions.
2116 *
2117 * Note that we may have delayed dropping an mm in context_switch(). If
2118 * so, we finish that here outside of the runqueue lock. (Doing it
2119 * with the lock held can cause deadlocks; see schedule() for
2120 * details.)
2121 */
2122static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2123 __releases(rq->lock)
2124{
2125 struct mm_struct *mm = rq->prev_mm;
2126 long prev_state;
2127
2128 rq->prev_mm = NULL;
2129
2130 /*
2131 * A task struct has one reference for the use as "current".
2132 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2133 * schedule one last time. The schedule call will never return, and
2134 * the scheduled task must drop that reference.
2135 * The test for TASK_DEAD must occur while the runqueue locks are
2136 * still held, otherwise prev could be scheduled on another cpu, die
2137 * there before we look at prev->state, and then the reference would
2138 * be dropped twice.
2139 * Manfred Spraul <manfred@colorfullife.com>
2140 */
2141 prev_state = prev->state;
2142 vtime_task_switch(prev);
2143 finish_arch_switch(prev);
2144 perf_event_task_sched_in(prev, current);
2145 finish_lock_switch(rq, prev);
2146 finish_arch_post_lock_switch();
2147
2148 fire_sched_in_preempt_notifiers(current);
2149 if (mm)
2150 mmdrop(mm);
2151 if (unlikely(prev_state == TASK_DEAD)) {
2152 if (prev->sched_class->task_dead)
2153 prev->sched_class->task_dead(prev);
2154
2155 /*
2156 * Remove function-return probe instances associated with this
2157 * task and put them back on the free list.
2158 */
2159 kprobe_flush_task(prev);
2160 put_task_struct(prev);
2161 }
2162
2163 tick_nohz_task_switch(current);
2164}
2165
2166#ifdef CONFIG_SMP
2167
2168/* rq->lock is NOT held, but preemption is disabled */
2169static inline void post_schedule(struct rq *rq)
2170{
2171 if (rq->post_schedule) {
2172 unsigned long flags;
2173
2174 raw_spin_lock_irqsave(&rq->lock, flags);
2175 if (rq->curr->sched_class->post_schedule)
2176 rq->curr->sched_class->post_schedule(rq);
2177 raw_spin_unlock_irqrestore(&rq->lock, flags);
2178
2179 rq->post_schedule = 0;
2180 }
2181}
2182
2183#else
2184
2185static inline void post_schedule(struct rq *rq)
2186{
2187}
2188
2189#endif
2190
2191/**
2192 * schedule_tail - first thing a freshly forked thread must call.
2193 * @prev: the thread we just switched away from.
2194 */
2195asmlinkage __visible void schedule_tail(struct task_struct *prev)
2196 __releases(rq->lock)
2197{
2198 struct rq *rq = this_rq();
2199
2200 finish_task_switch(rq, prev);
2201
2202 /*
2203 * FIXME: do we need to worry about rq being invalidated by the
2204 * task_switch?
2205 */
2206 post_schedule(rq);
2207
2208#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2209 /* In this case, finish_task_switch does not reenable preemption */
2210 preempt_enable();
2211#endif
2212 if (current->set_child_tid)
2213 put_user(task_pid_vnr(current), current->set_child_tid);
2214}
2215
2216/*
2217 * context_switch - switch to the new MM and the new
2218 * thread's register state.
2219 */
2220static inline void
2221context_switch(struct rq *rq, struct task_struct *prev,
2222 struct task_struct *next)
2223{
2224 struct mm_struct *mm, *oldmm;
2225
2226 prepare_task_switch(rq, prev, next);
2227
2228 mm = next->mm;
2229 oldmm = prev->active_mm;
2230 /*
2231 * For paravirt, this is coupled with an exit in switch_to to
2232 * combine the page table reload and the switch backend into
2233 * one hypercall.
2234 */
2235 arch_start_context_switch(prev);
2236
2237 if (!mm) {
2238 next->active_mm = oldmm;
2239 atomic_inc(&oldmm->mm_count);
2240 enter_lazy_tlb(oldmm, next);
2241 } else
2242 switch_mm(oldmm, mm, next);
2243
2244 if (!prev->mm) {
2245 prev->active_mm = NULL;
2246 rq->prev_mm = oldmm;
2247 }
2248 /*
2249 * Since the runqueue lock will be released by the next
2250 * task (which is an invalid locking op but in the case
2251 * of the scheduler it's an obvious special-case), so we
2252 * do an early lockdep release here:
2253 */
2254#ifndef __ARCH_WANT_UNLOCKED_CTXSW
2255 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2256#endif
2257
2258 context_tracking_task_switch(prev, next);
2259 /* Here we just switch the register state and the stack. */
2260 switch_to(prev, next, prev);
2261
2262 barrier();
2263 /*
2264 * this_rq must be evaluated again because prev may have moved
2265 * CPUs since it called schedule(), thus the 'rq' on its stack
2266 * frame will be invalid.
2267 */
2268 finish_task_switch(this_rq(), prev);
2269}
2270
2271/*
2272 * nr_running and nr_context_switches:
2273 *
2274 * externally visible scheduler statistics: current number of runnable
2275 * threads, total number of context switches performed since bootup.
2276 */
2277unsigned long nr_running(void)
2278{
2279 unsigned long i, sum = 0;
2280
2281 for_each_online_cpu(i)
2282 sum += cpu_rq(i)->nr_running;
2283
2284 return sum;
2285}
2286
2287unsigned long long nr_context_switches(void)
2288{
2289 int i;
2290 unsigned long long sum = 0;
2291
2292 for_each_possible_cpu(i)
2293 sum += cpu_rq(i)->nr_switches;
2294
2295 return sum;
2296}
2297
2298unsigned long nr_iowait(void)
2299{
2300 unsigned long i, sum = 0;
2301
2302 for_each_possible_cpu(i)
2303 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2304
2305 return sum;
2306}
2307
2308unsigned long nr_iowait_cpu(int cpu)
2309{
2310 struct rq *this = cpu_rq(cpu);
2311 return atomic_read(&this->nr_iowait);
2312}
2313
2314#ifdef CONFIG_SMP
2315
2316/*
2317 * sched_exec - execve() is a valuable balancing opportunity, because at
2318 * this point the task has the smallest effective memory and cache footprint.
2319 */
2320void sched_exec(void)
2321{
2322 struct task_struct *p = current;
2323 unsigned long flags;
2324 int dest_cpu;
2325
2326 raw_spin_lock_irqsave(&p->pi_lock, flags);
2327 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2328 if (dest_cpu == smp_processor_id())
2329 goto unlock;
2330
2331 if (likely(cpu_active(dest_cpu))) {
2332 struct migration_arg arg = { p, dest_cpu };
2333
2334 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2335 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2336 return;
2337 }
2338unlock:
2339 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2340}
2341
2342#endif
2343
2344DEFINE_PER_CPU(struct kernel_stat, kstat);
2345DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2346
2347EXPORT_PER_CPU_SYMBOL(kstat);
2348EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2349
2350/*
2351 * Return any ns on the sched_clock that have not yet been accounted in
2352 * @p in case that task is currently running.
2353 *
2354 * Called with task_rq_lock() held on @rq.
2355 */
2356static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2357{
2358 u64 ns = 0;
2359
2360 if (task_current(rq, p)) {
2361 update_rq_clock(rq);
2362 ns = rq_clock_task(rq) - p->se.exec_start;
2363 if ((s64)ns < 0)
2364 ns = 0;
2365 }
2366
2367 return ns;
2368}
2369
2370unsigned long long task_delta_exec(struct task_struct *p)
2371{
2372 unsigned long flags;
2373 struct rq *rq;
2374 u64 ns = 0;
2375
2376 rq = task_rq_lock(p, &flags);
2377 ns = do_task_delta_exec(p, rq);
2378 task_rq_unlock(rq, p, &flags);
2379
2380 return ns;
2381}
2382
2383/*
2384 * Return accounted runtime for the task.
2385 * In case the task is currently running, return the runtime plus current's
2386 * pending runtime that have not been accounted yet.
2387 */
2388unsigned long long task_sched_runtime(struct task_struct *p)
2389{
2390 unsigned long flags;
2391 struct rq *rq;
2392 u64 ns = 0;
2393
2394#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2395 /*
2396 * 64-bit doesn't need locks to atomically read a 64bit value.
2397 * So we have a optimization chance when the task's delta_exec is 0.
2398 * Reading ->on_cpu is racy, but this is ok.
2399 *
2400 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2401 * If we race with it entering cpu, unaccounted time is 0. This is
2402 * indistinguishable from the read occurring a few cycles earlier.
2403 */
2404 if (!p->on_cpu)
2405 return p->se.sum_exec_runtime;
2406#endif
2407
2408 rq = task_rq_lock(p, &flags);
2409 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2410 task_rq_unlock(rq, p, &flags);
2411
2412 return ns;
2413}
2414
2415/*
2416 * This function gets called by the timer code, with HZ frequency.
2417 * We call it with interrupts disabled.
2418 */
2419void scheduler_tick(void)
2420{
2421 int cpu = smp_processor_id();
2422 struct rq *rq = cpu_rq(cpu);
2423 struct task_struct *curr = rq->curr;
2424
2425 sched_clock_tick();
2426
2427 raw_spin_lock(&rq->lock);
2428 update_rq_clock(rq);
2429 curr->sched_class->task_tick(rq, curr, 0);
2430 update_cpu_load_active(rq);
2431 raw_spin_unlock(&rq->lock);
2432
2433 perf_event_task_tick();
2434
2435#ifdef CONFIG_SMP
2436 rq->idle_balance = idle_cpu(cpu);
2437 trigger_load_balance(rq);
2438#endif
2439 rq_last_tick_reset(rq);
2440}
2441
2442#ifdef CONFIG_NO_HZ_FULL
2443/**
2444 * scheduler_tick_max_deferment
2445 *
2446 * Keep at least one tick per second when a single
2447 * active task is running because the scheduler doesn't
2448 * yet completely support full dynticks environment.
2449 *
2450 * This makes sure that uptime, CFS vruntime, load
2451 * balancing, etc... continue to move forward, even
2452 * with a very low granularity.
2453 *
2454 * Return: Maximum deferment in nanoseconds.
2455 */
2456u64 scheduler_tick_max_deferment(void)
2457{
2458 struct rq *rq = this_rq();
2459 unsigned long next, now = ACCESS_ONCE(jiffies);
2460
2461 next = rq->last_sched_tick + HZ;
2462
2463 if (time_before_eq(next, now))
2464 return 0;
2465
2466 return jiffies_to_nsecs(next - now);
2467}
2468#endif
2469
2470notrace unsigned long get_parent_ip(unsigned long addr)
2471{
2472 if (in_lock_functions(addr)) {
2473 addr = CALLER_ADDR2;
2474 if (in_lock_functions(addr))
2475 addr = CALLER_ADDR3;
2476 }
2477 return addr;
2478}
2479
2480#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2481 defined(CONFIG_PREEMPT_TRACER))
2482
2483void __kprobes preempt_count_add(int val)
2484{
2485#ifdef CONFIG_DEBUG_PREEMPT
2486 /*
2487 * Underflow?
2488 */
2489 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2490 return;
2491#endif
2492 __preempt_count_add(val);
2493#ifdef CONFIG_DEBUG_PREEMPT
2494 /*
2495 * Spinlock count overflowing soon?
2496 */
2497 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2498 PREEMPT_MASK - 10);
2499#endif
2500 if (preempt_count() == val) {
2501 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2502#ifdef CONFIG_DEBUG_PREEMPT
2503 current->preempt_disable_ip = ip;
2504#endif
2505 trace_preempt_off(CALLER_ADDR0, ip);
2506 }
2507}
2508EXPORT_SYMBOL(preempt_count_add);
2509
2510void __kprobes preempt_count_sub(int val)
2511{
2512#ifdef CONFIG_DEBUG_PREEMPT
2513 /*
2514 * Underflow?
2515 */
2516 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2517 return;
2518 /*
2519 * Is the spinlock portion underflowing?
2520 */
2521 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2522 !(preempt_count() & PREEMPT_MASK)))
2523 return;
2524#endif
2525
2526 if (preempt_count() == val)
2527 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2528 __preempt_count_sub(val);
2529}
2530EXPORT_SYMBOL(preempt_count_sub);
2531
2532#endif
2533
2534/*
2535 * Print scheduling while atomic bug:
2536 */
2537static noinline void __schedule_bug(struct task_struct *prev)
2538{
2539 if (oops_in_progress)
2540 return;
2541
2542 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2543 prev->comm, prev->pid, preempt_count());
2544
2545 debug_show_held_locks(prev);
2546 print_modules();
2547 if (irqs_disabled())
2548 print_irqtrace_events(prev);
2549#ifdef CONFIG_DEBUG_PREEMPT
2550 if (in_atomic_preempt_off()) {
2551 pr_err("Preemption disabled at:");
2552 print_ip_sym(current->preempt_disable_ip);
2553 pr_cont("\n");
2554 }
2555#endif
2556 dump_stack();
2557 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2558}
2559
2560/*
2561 * Various schedule()-time debugging checks and statistics:
2562 */
2563static inline void schedule_debug(struct task_struct *prev)
2564{
2565 /*
2566 * Test if we are atomic. Since do_exit() needs to call into
2567 * schedule() atomically, we ignore that path. Otherwise whine
2568 * if we are scheduling when we should not.
2569 */
2570 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2571 __schedule_bug(prev);
2572 rcu_sleep_check();
2573
2574 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2575
2576 schedstat_inc(this_rq(), sched_count);
2577}
2578
2579/*
2580 * Pick up the highest-prio task:
2581 */
2582static inline struct task_struct *
2583pick_next_task(struct rq *rq, struct task_struct *prev)
2584{
2585 const struct sched_class *class = &fair_sched_class;
2586 struct task_struct *p;
2587
2588 /*
2589 * Optimization: we know that if all tasks are in
2590 * the fair class we can call that function directly:
2591 */
2592 if (likely(prev->sched_class == class &&
2593 rq->nr_running == rq->cfs.h_nr_running)) {
2594 p = fair_sched_class.pick_next_task(rq, prev);
2595 if (unlikely(p == RETRY_TASK))
2596 goto again;
2597
2598 /* assumes fair_sched_class->next == idle_sched_class */
2599 if (unlikely(!p))
2600 p = idle_sched_class.pick_next_task(rq, prev);
2601
2602 return p;
2603 }
2604
2605again:
2606 for_each_class(class) {
2607 p = class->pick_next_task(rq, prev);
2608 if (p) {
2609 if (unlikely(p == RETRY_TASK))
2610 goto again;
2611 return p;
2612 }
2613 }
2614
2615 BUG(); /* the idle class will always have a runnable task */
2616}
2617
2618/*
2619 * __schedule() is the main scheduler function.
2620 *
2621 * The main means of driving the scheduler and thus entering this function are:
2622 *
2623 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2624 *
2625 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2626 * paths. For example, see arch/x86/entry_64.S.
2627 *
2628 * To drive preemption between tasks, the scheduler sets the flag in timer
2629 * interrupt handler scheduler_tick().
2630 *
2631 * 3. Wakeups don't really cause entry into schedule(). They add a
2632 * task to the run-queue and that's it.
2633 *
2634 * Now, if the new task added to the run-queue preempts the current
2635 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2636 * called on the nearest possible occasion:
2637 *
2638 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2639 *
2640 * - in syscall or exception context, at the next outmost
2641 * preempt_enable(). (this might be as soon as the wake_up()'s
2642 * spin_unlock()!)
2643 *
2644 * - in IRQ context, return from interrupt-handler to
2645 * preemptible context
2646 *
2647 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2648 * then at the next:
2649 *
2650 * - cond_resched() call
2651 * - explicit schedule() call
2652 * - return from syscall or exception to user-space
2653 * - return from interrupt-handler to user-space
2654 */
2655static void __sched __schedule(void)
2656{
2657 struct task_struct *prev, *next;
2658 unsigned long *switch_count;
2659 struct rq *rq;
2660 int cpu;
2661
2662need_resched:
2663 preempt_disable();
2664 cpu = smp_processor_id();
2665 rq = cpu_rq(cpu);
2666 rcu_note_context_switch(cpu);
2667 prev = rq->curr;
2668
2669 schedule_debug(prev);
2670
2671 if (sched_feat(HRTICK))
2672 hrtick_clear(rq);
2673
2674 /*
2675 * Make sure that signal_pending_state()->signal_pending() below
2676 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2677 * done by the caller to avoid the race with signal_wake_up().
2678 */
2679 smp_mb__before_spinlock();
2680 raw_spin_lock_irq(&rq->lock);
2681
2682 switch_count = &prev->nivcsw;
2683 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2684 if (unlikely(signal_pending_state(prev->state, prev))) {
2685 prev->state = TASK_RUNNING;
2686 } else {
2687 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2688 prev->on_rq = 0;
2689
2690 /*
2691 * If a worker went to sleep, notify and ask workqueue
2692 * whether it wants to wake up a task to maintain
2693 * concurrency.
2694 */
2695 if (prev->flags & PF_WQ_WORKER) {
2696 struct task_struct *to_wakeup;
2697
2698 to_wakeup = wq_worker_sleeping(prev, cpu);
2699 if (to_wakeup)
2700 try_to_wake_up_local(to_wakeup);
2701 }
2702 }
2703 switch_count = &prev->nvcsw;
2704 }
2705
2706 if (prev->on_rq || rq->skip_clock_update < 0)
2707 update_rq_clock(rq);
2708
2709 next = pick_next_task(rq, prev);
2710 clear_tsk_need_resched(prev);
2711 clear_preempt_need_resched();
2712 rq->skip_clock_update = 0;
2713
2714 if (likely(prev != next)) {
2715 rq->nr_switches++;
2716 rq->curr = next;
2717 ++*switch_count;
2718
2719 context_switch(rq, prev, next); /* unlocks the rq */
2720 /*
2721 * The context switch have flipped the stack from under us
2722 * and restored the local variables which were saved when
2723 * this task called schedule() in the past. prev == current
2724 * is still correct, but it can be moved to another cpu/rq.
2725 */
2726 cpu = smp_processor_id();
2727 rq = cpu_rq(cpu);
2728 } else
2729 raw_spin_unlock_irq(&rq->lock);
2730
2731 post_schedule(rq);
2732
2733 sched_preempt_enable_no_resched();
2734 if (need_resched())
2735 goto need_resched;
2736}
2737
2738static inline void sched_submit_work(struct task_struct *tsk)
2739{
2740 if (!tsk->state || tsk_is_pi_blocked(tsk))
2741 return;
2742 /*
2743 * If we are going to sleep and we have plugged IO queued,
2744 * make sure to submit it to avoid deadlocks.
2745 */
2746 if (blk_needs_flush_plug(tsk))
2747 blk_schedule_flush_plug(tsk);
2748}
2749
2750asmlinkage __visible void __sched schedule(void)
2751{
2752 struct task_struct *tsk = current;
2753
2754 sched_submit_work(tsk);
2755 __schedule();
2756}
2757EXPORT_SYMBOL(schedule);
2758
2759#ifdef CONFIG_CONTEXT_TRACKING
2760asmlinkage __visible void __sched schedule_user(void)
2761{
2762 /*
2763 * If we come here after a random call to set_need_resched(),
2764 * or we have been woken up remotely but the IPI has not yet arrived,
2765 * we haven't yet exited the RCU idle mode. Do it here manually until
2766 * we find a better solution.
2767 */
2768 user_exit();
2769 schedule();
2770 user_enter();
2771}
2772#endif
2773
2774/**
2775 * schedule_preempt_disabled - called with preemption disabled
2776 *
2777 * Returns with preemption disabled. Note: preempt_count must be 1
2778 */
2779void __sched schedule_preempt_disabled(void)
2780{
2781 sched_preempt_enable_no_resched();
2782 schedule();
2783 preempt_disable();
2784}
2785
2786#ifdef CONFIG_PREEMPT
2787/*
2788 * this is the entry point to schedule() from in-kernel preemption
2789 * off of preempt_enable. Kernel preemptions off return from interrupt
2790 * occur there and call schedule directly.
2791 */
2792asmlinkage __visible void __sched notrace preempt_schedule(void)
2793{
2794 /*
2795 * If there is a non-zero preempt_count or interrupts are disabled,
2796 * we do not want to preempt the current task. Just return..
2797 */
2798 if (likely(!preemptible()))
2799 return;
2800
2801 do {
2802 __preempt_count_add(PREEMPT_ACTIVE);
2803 __schedule();
2804 __preempt_count_sub(PREEMPT_ACTIVE);
2805
2806 /*
2807 * Check again in case we missed a preemption opportunity
2808 * between schedule and now.
2809 */
2810 barrier();
2811 } while (need_resched());
2812}
2813EXPORT_SYMBOL(preempt_schedule);
2814#endif /* CONFIG_PREEMPT */
2815
2816/*
2817 * this is the entry point to schedule() from kernel preemption
2818 * off of irq context.
2819 * Note, that this is called and return with irqs disabled. This will
2820 * protect us against recursive calling from irq.
2821 */
2822asmlinkage __visible void __sched preempt_schedule_irq(void)
2823{
2824 enum ctx_state prev_state;
2825
2826 /* Catch callers which need to be fixed */
2827 BUG_ON(preempt_count() || !irqs_disabled());
2828
2829 prev_state = exception_enter();
2830
2831 do {
2832 __preempt_count_add(PREEMPT_ACTIVE);
2833 local_irq_enable();
2834 __schedule();
2835 local_irq_disable();
2836 __preempt_count_sub(PREEMPT_ACTIVE);
2837
2838 /*
2839 * Check again in case we missed a preemption opportunity
2840 * between schedule and now.
2841 */
2842 barrier();
2843 } while (need_resched());
2844
2845 exception_exit(prev_state);
2846}
2847
2848int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2849 void *key)
2850{
2851 return try_to_wake_up(curr->private, mode, wake_flags);
2852}
2853EXPORT_SYMBOL(default_wake_function);
2854
2855#ifdef CONFIG_RT_MUTEXES
2856
2857/*
2858 * rt_mutex_setprio - set the current priority of a task
2859 * @p: task
2860 * @prio: prio value (kernel-internal form)
2861 *
2862 * This function changes the 'effective' priority of a task. It does
2863 * not touch ->normal_prio like __setscheduler().
2864 *
2865 * Used by the rt_mutex code to implement priority inheritance
2866 * logic. Call site only calls if the priority of the task changed.
2867 */
2868void rt_mutex_setprio(struct task_struct *p, int prio)
2869{
2870 int oldprio, on_rq, running, enqueue_flag = 0;
2871 struct rq *rq;
2872 const struct sched_class *prev_class;
2873
2874 BUG_ON(prio > MAX_PRIO);
2875
2876 rq = __task_rq_lock(p);
2877
2878 /*
2879 * Idle task boosting is a nono in general. There is one
2880 * exception, when PREEMPT_RT and NOHZ is active:
2881 *
2882 * The idle task calls get_next_timer_interrupt() and holds
2883 * the timer wheel base->lock on the CPU and another CPU wants
2884 * to access the timer (probably to cancel it). We can safely
2885 * ignore the boosting request, as the idle CPU runs this code
2886 * with interrupts disabled and will complete the lock
2887 * protected section without being interrupted. So there is no
2888 * real need to boost.
2889 */
2890 if (unlikely(p == rq->idle)) {
2891 WARN_ON(p != rq->curr);
2892 WARN_ON(p->pi_blocked_on);
2893 goto out_unlock;
2894 }
2895
2896 trace_sched_pi_setprio(p, prio);
2897 p->pi_top_task = rt_mutex_get_top_task(p);
2898 oldprio = p->prio;
2899 prev_class = p->sched_class;
2900 on_rq = p->on_rq;
2901 running = task_current(rq, p);
2902 if (on_rq)
2903 dequeue_task(rq, p, 0);
2904 if (running)
2905 p->sched_class->put_prev_task(rq, p);
2906
2907 /*
2908 * Boosting condition are:
2909 * 1. -rt task is running and holds mutex A
2910 * --> -dl task blocks on mutex A
2911 *
2912 * 2. -dl task is running and holds mutex A
2913 * --> -dl task blocks on mutex A and could preempt the
2914 * running task
2915 */
2916 if (dl_prio(prio)) {
2917 if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
2918 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
2919 p->dl.dl_boosted = 1;
2920 p->dl.dl_throttled = 0;
2921 enqueue_flag = ENQUEUE_REPLENISH;
2922 } else
2923 p->dl.dl_boosted = 0;
2924 p->sched_class = &dl_sched_class;
2925 } else if (rt_prio(prio)) {
2926 if (dl_prio(oldprio))
2927 p->dl.dl_boosted = 0;
2928 if (oldprio < prio)
2929 enqueue_flag = ENQUEUE_HEAD;
2930 p->sched_class = &rt_sched_class;
2931 } else {
2932 if (dl_prio(oldprio))
2933 p->dl.dl_boosted = 0;
2934 p->sched_class = &fair_sched_class;
2935 }
2936
2937 p->prio = prio;
2938
2939 if (running)
2940 p->sched_class->set_curr_task(rq);
2941 if (on_rq)
2942 enqueue_task(rq, p, enqueue_flag);
2943
2944 check_class_changed(rq, p, prev_class, oldprio);
2945out_unlock:
2946 __task_rq_unlock(rq);
2947}
2948#endif
2949
2950void set_user_nice(struct task_struct *p, long nice)
2951{
2952 int old_prio, delta, on_rq;
2953 unsigned long flags;
2954 struct rq *rq;
2955
2956 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
2957 return;
2958 /*
2959 * We have to be careful, if called from sys_setpriority(),
2960 * the task might be in the middle of scheduling on another CPU.
2961 */
2962 rq = task_rq_lock(p, &flags);
2963 /*
2964 * The RT priorities are set via sched_setscheduler(), but we still
2965 * allow the 'normal' nice value to be set - but as expected
2966 * it wont have any effect on scheduling until the task is
2967 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
2968 */
2969 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2970 p->static_prio = NICE_TO_PRIO(nice);
2971 goto out_unlock;
2972 }
2973 on_rq = p->on_rq;
2974 if (on_rq)
2975 dequeue_task(rq, p, 0);
2976
2977 p->static_prio = NICE_TO_PRIO(nice);
2978 set_load_weight(p);
2979 old_prio = p->prio;
2980 p->prio = effective_prio(p);
2981 delta = p->prio - old_prio;
2982
2983 if (on_rq) {
2984 enqueue_task(rq, p, 0);
2985 /*
2986 * If the task increased its priority or is running and
2987 * lowered its priority, then reschedule its CPU:
2988 */
2989 if (delta < 0 || (delta > 0 && task_running(rq, p)))
2990 resched_task(rq->curr);
2991 }
2992out_unlock:
2993 task_rq_unlock(rq, p, &flags);
2994}
2995EXPORT_SYMBOL(set_user_nice);
2996
2997/*
2998 * can_nice - check if a task can reduce its nice value
2999 * @p: task
3000 * @nice: nice value
3001 */
3002int can_nice(const struct task_struct *p, const int nice)
3003{
3004 /* convert nice value [19,-20] to rlimit style value [1,40] */
3005 int nice_rlim = 20 - nice;
3006
3007 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3008 capable(CAP_SYS_NICE));
3009}
3010
3011#ifdef __ARCH_WANT_SYS_NICE
3012
3013/*
3014 * sys_nice - change the priority of the current process.
3015 * @increment: priority increment
3016 *
3017 * sys_setpriority is a more generic, but much slower function that
3018 * does similar things.
3019 */
3020SYSCALL_DEFINE1(nice, int, increment)
3021{
3022 long nice, retval;
3023
3024 /*
3025 * Setpriority might change our priority at the same moment.
3026 * We don't have to worry. Conceptually one call occurs first
3027 * and we have a single winner.
3028 */
3029 if (increment < -40)
3030 increment = -40;
3031 if (increment > 40)
3032 increment = 40;
3033
3034 nice = task_nice(current) + increment;
3035 if (nice < MIN_NICE)
3036 nice = MIN_NICE;
3037 if (nice > MAX_NICE)
3038 nice = MAX_NICE;
3039
3040 if (increment < 0 && !can_nice(current, nice))
3041 return -EPERM;
3042
3043 retval = security_task_setnice(current, nice);
3044 if (retval)
3045 return retval;
3046
3047 set_user_nice(current, nice);
3048 return 0;
3049}
3050
3051#endif
3052
3053/**
3054 * task_prio - return the priority value of a given task.
3055 * @p: the task in question.
3056 *
3057 * Return: The priority value as seen by users in /proc.
3058 * RT tasks are offset by -200. Normal tasks are centered
3059 * around 0, value goes from -16 to +15.
3060 */
3061int task_prio(const struct task_struct *p)
3062{
3063 return p->prio - MAX_RT_PRIO;
3064}
3065
3066/**
3067 * idle_cpu - is a given cpu idle currently?
3068 * @cpu: the processor in question.
3069 *
3070 * Return: 1 if the CPU is currently idle. 0 otherwise.
3071 */
3072int idle_cpu(int cpu)
3073{
3074 struct rq *rq = cpu_rq(cpu);
3075
3076 if (rq->curr != rq->idle)
3077 return 0;
3078
3079 if (rq->nr_running)
3080 return 0;
3081
3082#ifdef CONFIG_SMP
3083 if (!llist_empty(&rq->wake_list))
3084 return 0;
3085#endif
3086
3087 return 1;
3088}
3089
3090/**
3091 * idle_task - return the idle task for a given cpu.
3092 * @cpu: the processor in question.
3093 *
3094 * Return: The idle task for the cpu @cpu.
3095 */
3096struct task_struct *idle_task(int cpu)
3097{
3098 return cpu_rq(cpu)->idle;
3099}
3100
3101/**
3102 * find_process_by_pid - find a process with a matching PID value.
3103 * @pid: the pid in question.
3104 *
3105 * The task of @pid, if found. %NULL otherwise.
3106 */
3107static struct task_struct *find_process_by_pid(pid_t pid)
3108{
3109 return pid ? find_task_by_vpid(pid) : current;
3110}
3111
3112/*
3113 * This function initializes the sched_dl_entity of a newly becoming
3114 * SCHED_DEADLINE task.
3115 *
3116 * Only the static values are considered here, the actual runtime and the
3117 * absolute deadline will be properly calculated when the task is enqueued
3118 * for the first time with its new policy.
3119 */
3120static void
3121__setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3122{
3123 struct sched_dl_entity *dl_se = &p->dl;
3124
3125 init_dl_task_timer(dl_se);
3126 dl_se->dl_runtime = attr->sched_runtime;
3127 dl_se->dl_deadline = attr->sched_deadline;
3128 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3129 dl_se->flags = attr->sched_flags;
3130 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3131 dl_se->dl_throttled = 0;
3132 dl_se->dl_new = 1;
3133 dl_se->dl_yielded = 0;
3134}
3135
3136static void __setscheduler_params(struct task_struct *p,
3137 const struct sched_attr *attr)
3138{
3139 int policy = attr->sched_policy;
3140
3141 if (policy == -1) /* setparam */
3142 policy = p->policy;
3143
3144 p->policy = policy;
3145
3146 if (dl_policy(policy))
3147 __setparam_dl(p, attr);
3148 else if (fair_policy(policy))
3149 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3150
3151 /*
3152 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3153 * !rt_policy. Always setting this ensures that things like
3154 * getparam()/getattr() don't report silly values for !rt tasks.
3155 */
3156 p->rt_priority = attr->sched_priority;
3157 p->normal_prio = normal_prio(p);
3158 set_load_weight(p);
3159}
3160
3161/* Actually do priority change: must hold pi & rq lock. */
3162static void __setscheduler(struct rq *rq, struct task_struct *p,
3163 const struct sched_attr *attr)
3164{
3165 __setscheduler_params(p, attr);
3166
3167 /*
3168 * If we get here, there was no pi waiters boosting the
3169 * task. It is safe to use the normal prio.
3170 */
3171 p->prio = normal_prio(p);
3172
3173 if (dl_prio(p->prio))
3174 p->sched_class = &dl_sched_class;
3175 else if (rt_prio(p->prio))
3176 p->sched_class = &rt_sched_class;
3177 else
3178 p->sched_class = &fair_sched_class;
3179}
3180
3181static void
3182__getparam_dl(struct task_struct *p, struct sched_attr *attr)
3183{
3184 struct sched_dl_entity *dl_se = &p->dl;
3185
3186 attr->sched_priority = p->rt_priority;
3187 attr->sched_runtime = dl_se->dl_runtime;
3188 attr->sched_deadline = dl_se->dl_deadline;
3189 attr->sched_period = dl_se->dl_period;
3190 attr->sched_flags = dl_se->flags;
3191}
3192
3193/*
3194 * This function validates the new parameters of a -deadline task.
3195 * We ask for the deadline not being zero, and greater or equal
3196 * than the runtime, as well as the period of being zero or
3197 * greater than deadline. Furthermore, we have to be sure that
3198 * user parameters are above the internal resolution of 1us (we
3199 * check sched_runtime only since it is always the smaller one) and
3200 * below 2^63 ns (we have to check both sched_deadline and
3201 * sched_period, as the latter can be zero).
3202 */
3203static bool
3204__checkparam_dl(const struct sched_attr *attr)
3205{
3206 /* deadline != 0 */
3207 if (attr->sched_deadline == 0)
3208 return false;
3209
3210 /*
3211 * Since we truncate DL_SCALE bits, make sure we're at least
3212 * that big.
3213 */
3214 if (attr->sched_runtime < (1ULL << DL_SCALE))
3215 return false;
3216
3217 /*
3218 * Since we use the MSB for wrap-around and sign issues, make
3219 * sure it's not set (mind that period can be equal to zero).
3220 */
3221 if (attr->sched_deadline & (1ULL << 63) ||
3222 attr->sched_period & (1ULL << 63))
3223 return false;
3224
3225 /* runtime <= deadline <= period (if period != 0) */
3226 if ((attr->sched_period != 0 &&
3227 attr->sched_period < attr->sched_deadline) ||
3228 attr->sched_deadline < attr->sched_runtime)
3229 return false;
3230
3231 return true;
3232}
3233
3234/*
3235 * check the target process has a UID that matches the current process's
3236 */
3237static bool check_same_owner(struct task_struct *p)
3238{
3239 const struct cred *cred = current_cred(), *pcred;
3240 bool match;
3241
3242 rcu_read_lock();
3243 pcred = __task_cred(p);
3244 match = (uid_eq(cred->euid, pcred->euid) ||
3245 uid_eq(cred->euid, pcred->uid));
3246 rcu_read_unlock();
3247 return match;
3248}
3249
3250static int __sched_setscheduler(struct task_struct *p,
3251 const struct sched_attr *attr,
3252 bool user)
3253{
3254 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3255 MAX_RT_PRIO - 1 - attr->sched_priority;
3256 int retval, oldprio, oldpolicy = -1, on_rq, running;
3257 int policy = attr->sched_policy;
3258 unsigned long flags;
3259 const struct sched_class *prev_class;
3260 struct rq *rq;
3261 int reset_on_fork;
3262
3263 /* may grab non-irq protected spin_locks */
3264 BUG_ON(in_interrupt());
3265recheck:
3266 /* double check policy once rq lock held */
3267 if (policy < 0) {
3268 reset_on_fork = p->sched_reset_on_fork;
3269 policy = oldpolicy = p->policy;
3270 } else {
3271 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3272
3273 if (policy != SCHED_DEADLINE &&
3274 policy != SCHED_FIFO && policy != SCHED_RR &&
3275 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3276 policy != SCHED_IDLE)
3277 return -EINVAL;
3278 }
3279
3280 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3281 return -EINVAL;
3282
3283 /*
3284 * Valid priorities for SCHED_FIFO and SCHED_RR are
3285 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3286 * SCHED_BATCH and SCHED_IDLE is 0.
3287 */
3288 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3289 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3290 return -EINVAL;
3291 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3292 (rt_policy(policy) != (attr->sched_priority != 0)))
3293 return -EINVAL;
3294
3295 /*
3296 * Allow unprivileged RT tasks to decrease priority:
3297 */
3298 if (user && !capable(CAP_SYS_NICE)) {
3299 if (fair_policy(policy)) {
3300 if (attr->sched_nice < task_nice(p) &&
3301 !can_nice(p, attr->sched_nice))
3302 return -EPERM;
3303 }
3304
3305 if (rt_policy(policy)) {
3306 unsigned long rlim_rtprio =
3307 task_rlimit(p, RLIMIT_RTPRIO);
3308
3309 /* can't set/change the rt policy */
3310 if (policy != p->policy && !rlim_rtprio)
3311 return -EPERM;
3312
3313 /* can't increase priority */
3314 if (attr->sched_priority > p->rt_priority &&
3315 attr->sched_priority > rlim_rtprio)
3316 return -EPERM;
3317 }
3318
3319 /*
3320 * Can't set/change SCHED_DEADLINE policy at all for now
3321 * (safest behavior); in the future we would like to allow
3322 * unprivileged DL tasks to increase their relative deadline
3323 * or reduce their runtime (both ways reducing utilization)
3324 */
3325 if (dl_policy(policy))
3326 return -EPERM;
3327
3328 /*
3329 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3330 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3331 */
3332 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3333 if (!can_nice(p, task_nice(p)))
3334 return -EPERM;
3335 }
3336
3337 /* can't change other user's priorities */
3338 if (!check_same_owner(p))
3339 return -EPERM;
3340
3341 /* Normal users shall not reset the sched_reset_on_fork flag */
3342 if (p->sched_reset_on_fork && !reset_on_fork)
3343 return -EPERM;
3344 }
3345
3346 if (user) {
3347 retval = security_task_setscheduler(p);
3348 if (retval)
3349 return retval;
3350 }
3351
3352 /*
3353 * make sure no PI-waiters arrive (or leave) while we are
3354 * changing the priority of the task:
3355 *
3356 * To be able to change p->policy safely, the appropriate
3357 * runqueue lock must be held.
3358 */
3359 rq = task_rq_lock(p, &flags);
3360
3361 /*
3362 * Changing the policy of the stop threads its a very bad idea
3363 */
3364 if (p == rq->stop) {
3365 task_rq_unlock(rq, p, &flags);
3366 return -EINVAL;
3367 }
3368
3369 /*
3370 * If not changing anything there's no need to proceed further,
3371 * but store a possible modification of reset_on_fork.
3372 */
3373 if (unlikely(policy == p->policy)) {
3374 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3375 goto change;
3376 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3377 goto change;
3378 if (dl_policy(policy))
3379 goto change;
3380
3381 p->sched_reset_on_fork = reset_on_fork;
3382 task_rq_unlock(rq, p, &flags);
3383 return 0;
3384 }
3385change:
3386
3387 if (user) {
3388#ifdef CONFIG_RT_GROUP_SCHED
3389 /*
3390 * Do not allow realtime tasks into groups that have no runtime
3391 * assigned.
3392 */
3393 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3394 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3395 !task_group_is_autogroup(task_group(p))) {
3396 task_rq_unlock(rq, p, &flags);
3397 return -EPERM;
3398 }
3399#endif
3400#ifdef CONFIG_SMP
3401 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3402 cpumask_t *span = rq->rd->span;
3403
3404 /*
3405 * Don't allow tasks with an affinity mask smaller than
3406 * the entire root_domain to become SCHED_DEADLINE. We
3407 * will also fail if there's no bandwidth available.
3408 */
3409 if (!cpumask_subset(span, &p->cpus_allowed) ||
3410 rq->rd->dl_bw.bw == 0) {
3411 task_rq_unlock(rq, p, &flags);
3412 return -EPERM;
3413 }
3414 }
3415#endif
3416 }
3417
3418 /* recheck policy now with rq lock held */
3419 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3420 policy = oldpolicy = -1;
3421 task_rq_unlock(rq, p, &flags);
3422 goto recheck;
3423 }
3424
3425 /*
3426 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3427 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3428 * is available.
3429 */
3430 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3431 task_rq_unlock(rq, p, &flags);
3432 return -EBUSY;
3433 }
3434
3435 p->sched_reset_on_fork = reset_on_fork;
3436 oldprio = p->prio;
3437
3438 /*
3439 * Special case for priority boosted tasks.
3440 *
3441 * If the new priority is lower or equal (user space view)
3442 * than the current (boosted) priority, we just store the new
3443 * normal parameters and do not touch the scheduler class and
3444 * the runqueue. This will be done when the task deboost
3445 * itself.
3446 */
3447 if (rt_mutex_check_prio(p, newprio)) {
3448 __setscheduler_params(p, attr);
3449 task_rq_unlock(rq, p, &flags);
3450 return 0;
3451 }
3452
3453 on_rq = p->on_rq;
3454 running = task_current(rq, p);
3455 if (on_rq)
3456 dequeue_task(rq, p, 0);
3457 if (running)
3458 p->sched_class->put_prev_task(rq, p);
3459
3460 prev_class = p->sched_class;
3461 __setscheduler(rq, p, attr);
3462
3463 if (running)
3464 p->sched_class->set_curr_task(rq);
3465 if (on_rq) {
3466 /*
3467 * We enqueue to tail when the priority of a task is
3468 * increased (user space view).
3469 */
3470 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3471 }
3472
3473 check_class_changed(rq, p, prev_class, oldprio);
3474 task_rq_unlock(rq, p, &flags);
3475
3476 rt_mutex_adjust_pi(p);
3477
3478 return 0;
3479}
3480
3481static int _sched_setscheduler(struct task_struct *p, int policy,
3482 const struct sched_param *param, bool check)
3483{
3484 struct sched_attr attr = {
3485 .sched_policy = policy,
3486 .sched_priority = param->sched_priority,
3487 .sched_nice = PRIO_TO_NICE(p->static_prio),
3488 };
3489
3490 /*
3491 * Fixup the legacy SCHED_RESET_ON_FORK hack
3492 */
3493 if (policy & SCHED_RESET_ON_FORK) {
3494 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3495 policy &= ~SCHED_RESET_ON_FORK;
3496 attr.sched_policy = policy;
3497 }
3498
3499 return __sched_setscheduler(p, &attr, check);
3500}
3501/**
3502 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3503 * @p: the task in question.
3504 * @policy: new policy.
3505 * @param: structure containing the new RT priority.
3506 *
3507 * Return: 0 on success. An error code otherwise.
3508 *
3509 * NOTE that the task may be already dead.
3510 */
3511int sched_setscheduler(struct task_struct *p, int policy,
3512 const struct sched_param *param)
3513{
3514 return _sched_setscheduler(p, policy, param, true);
3515}
3516EXPORT_SYMBOL_GPL(sched_setscheduler);
3517
3518int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3519{
3520 return __sched_setscheduler(p, attr, true);
3521}
3522EXPORT_SYMBOL_GPL(sched_setattr);
3523
3524/**
3525 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3526 * @p: the task in question.
3527 * @policy: new policy.
3528 * @param: structure containing the new RT priority.
3529 *
3530 * Just like sched_setscheduler, only don't bother checking if the
3531 * current context has permission. For example, this is needed in
3532 * stop_machine(): we create temporary high priority worker threads,
3533 * but our caller might not have that capability.
3534 *
3535 * Return: 0 on success. An error code otherwise.
3536 */
3537int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3538 const struct sched_param *param)
3539{
3540 return _sched_setscheduler(p, policy, param, false);
3541}
3542
3543static int
3544do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3545{
3546 struct sched_param lparam;
3547 struct task_struct *p;
3548 int retval;
3549
3550 if (!param || pid < 0)
3551 return -EINVAL;
3552 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3553 return -EFAULT;
3554
3555 rcu_read_lock();
3556 retval = -ESRCH;
3557 p = find_process_by_pid(pid);
3558 if (p != NULL)
3559 retval = sched_setscheduler(p, policy, &lparam);
3560 rcu_read_unlock();
3561
3562 return retval;
3563}
3564
3565/*
3566 * Mimics kernel/events/core.c perf_copy_attr().
3567 */
3568static int sched_copy_attr(struct sched_attr __user *uattr,
3569 struct sched_attr *attr)
3570{
3571 u32 size;
3572 int ret;
3573
3574 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3575 return -EFAULT;
3576
3577 /*
3578 * zero the full structure, so that a short copy will be nice.
3579 */
3580 memset(attr, 0, sizeof(*attr));
3581
3582 ret = get_user(size, &uattr->size);
3583 if (ret)
3584 return ret;
3585
3586 if (size > PAGE_SIZE) /* silly large */
3587 goto err_size;
3588
3589 if (!size) /* abi compat */
3590 size = SCHED_ATTR_SIZE_VER0;
3591
3592 if (size < SCHED_ATTR_SIZE_VER0)
3593 goto err_size;
3594
3595 /*
3596 * If we're handed a bigger struct than we know of,
3597 * ensure all the unknown bits are 0 - i.e. new
3598 * user-space does not rely on any kernel feature
3599 * extensions we dont know about yet.
3600 */
3601 if (size > sizeof(*attr)) {
3602 unsigned char __user *addr;
3603 unsigned char __user *end;
3604 unsigned char val;
3605
3606 addr = (void __user *)uattr + sizeof(*attr);
3607 end = (void __user *)uattr + size;
3608
3609 for (; addr < end; addr++) {
3610 ret = get_user(val, addr);
3611 if (ret)
3612 return ret;
3613 if (val)
3614 goto err_size;
3615 }
3616 size = sizeof(*attr);
3617 }
3618
3619 ret = copy_from_user(attr, uattr, size);
3620 if (ret)
3621 return -EFAULT;
3622
3623 /*
3624 * XXX: do we want to be lenient like existing syscalls; or do we want
3625 * to be strict and return an error on out-of-bounds values?
3626 */
3627 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3628
3629out:
3630 return ret;
3631
3632err_size:
3633 put_user(sizeof(*attr), &uattr->size);
3634 ret = -E2BIG;
3635 goto out;
3636}
3637
3638/**
3639 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3640 * @pid: the pid in question.
3641 * @policy: new policy.
3642 * @param: structure containing the new RT priority.
3643 *
3644 * Return: 0 on success. An error code otherwise.
3645 */
3646SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3647 struct sched_param __user *, param)
3648{
3649 /* negative values for policy are not valid */
3650 if (policy < 0)
3651 return -EINVAL;
3652
3653 return do_sched_setscheduler(pid, policy, param);
3654}
3655
3656/**
3657 * sys_sched_setparam - set/change the RT priority of a thread
3658 * @pid: the pid in question.
3659 * @param: structure containing the new RT priority.
3660 *
3661 * Return: 0 on success. An error code otherwise.
3662 */
3663SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3664{
3665 return do_sched_setscheduler(pid, -1, param);
3666}
3667
3668/**
3669 * sys_sched_setattr - same as above, but with extended sched_attr
3670 * @pid: the pid in question.
3671 * @uattr: structure containing the extended parameters.
3672 * @flags: for future extension.
3673 */
3674SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3675 unsigned int, flags)
3676{
3677 struct sched_attr attr;
3678 struct task_struct *p;
3679 int retval;
3680
3681 if (!uattr || pid < 0 || flags)
3682 return -EINVAL;
3683
3684 retval = sched_copy_attr(uattr, &attr);
3685 if (retval)
3686 return retval;
3687
3688 if ((int)attr.sched_policy < 0)
3689 return -EINVAL;
3690
3691 rcu_read_lock();
3692 retval = -ESRCH;
3693 p = find_process_by_pid(pid);
3694 if (p != NULL)
3695 retval = sched_setattr(p, &attr);
3696 rcu_read_unlock();
3697
3698 return retval;
3699}
3700
3701/**
3702 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3703 * @pid: the pid in question.
3704 *
3705 * Return: On success, the policy of the thread. Otherwise, a negative error
3706 * code.
3707 */
3708SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3709{
3710 struct task_struct *p;
3711 int retval;
3712
3713 if (pid < 0)
3714 return -EINVAL;
3715
3716 retval = -ESRCH;
3717 rcu_read_lock();
3718 p = find_process_by_pid(pid);
3719 if (p) {
3720 retval = security_task_getscheduler(p);
3721 if (!retval)
3722 retval = p->policy
3723 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3724 }
3725 rcu_read_unlock();
3726 return retval;
3727}
3728
3729/**
3730 * sys_sched_getparam - get the RT priority of a thread
3731 * @pid: the pid in question.
3732 * @param: structure containing the RT priority.
3733 *
3734 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3735 * code.
3736 */
3737SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3738{
3739 struct sched_param lp = { .sched_priority = 0 };
3740 struct task_struct *p;
3741 int retval;
3742
3743 if (!param || pid < 0)
3744 return -EINVAL;
3745
3746 rcu_read_lock();
3747 p = find_process_by_pid(pid);
3748 retval = -ESRCH;
3749 if (!p)
3750 goto out_unlock;
3751
3752 retval = security_task_getscheduler(p);
3753 if (retval)
3754 goto out_unlock;
3755
3756 if (task_has_rt_policy(p))
3757 lp.sched_priority = p->rt_priority;
3758 rcu_read_unlock();
3759
3760 /*
3761 * This one might sleep, we cannot do it with a spinlock held ...
3762 */
3763 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3764
3765 return retval;
3766
3767out_unlock:
3768 rcu_read_unlock();
3769 return retval;
3770}
3771
3772static int sched_read_attr(struct sched_attr __user *uattr,
3773 struct sched_attr *attr,
3774 unsigned int usize)
3775{
3776 int ret;
3777
3778 if (!access_ok(VERIFY_WRITE, uattr, usize))
3779 return -EFAULT;
3780
3781 /*
3782 * If we're handed a smaller struct than we know of,
3783 * ensure all the unknown bits are 0 - i.e. old
3784 * user-space does not get uncomplete information.
3785 */
3786 if (usize < sizeof(*attr)) {
3787 unsigned char *addr;
3788 unsigned char *end;
3789
3790 addr = (void *)attr + usize;
3791 end = (void *)attr + sizeof(*attr);
3792
3793 for (; addr < end; addr++) {
3794 if (*addr)
3795 goto err_size;
3796 }
3797
3798 attr->size = usize;
3799 }
3800
3801 ret = copy_to_user(uattr, attr, attr->size);
3802 if (ret)
3803 return -EFAULT;
3804
3805out:
3806 return ret;
3807
3808err_size:
3809 ret = -E2BIG;
3810 goto out;
3811}
3812
3813/**
3814 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3815 * @pid: the pid in question.
3816 * @uattr: structure containing the extended parameters.
3817 * @size: sizeof(attr) for fwd/bwd comp.
3818 * @flags: for future extension.
3819 */
3820SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3821 unsigned int, size, unsigned int, flags)
3822{
3823 struct sched_attr attr = {
3824 .size = sizeof(struct sched_attr),
3825 };
3826 struct task_struct *p;
3827 int retval;
3828
3829 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3830 size < SCHED_ATTR_SIZE_VER0 || flags)
3831 return -EINVAL;
3832
3833 rcu_read_lock();
3834 p = find_process_by_pid(pid);
3835 retval = -ESRCH;
3836 if (!p)
3837 goto out_unlock;
3838
3839 retval = security_task_getscheduler(p);
3840 if (retval)
3841 goto out_unlock;
3842
3843 attr.sched_policy = p->policy;
3844 if (p->sched_reset_on_fork)
3845 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3846 if (task_has_dl_policy(p))
3847 __getparam_dl(p, &attr);
3848 else if (task_has_rt_policy(p))
3849 attr.sched_priority = p->rt_priority;
3850 else
3851 attr.sched_nice = task_nice(p);
3852
3853 rcu_read_unlock();
3854
3855 retval = sched_read_attr(uattr, &attr, size);
3856 return retval;
3857
3858out_unlock:
3859 rcu_read_unlock();
3860 return retval;
3861}
3862
3863long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3864{
3865 cpumask_var_t cpus_allowed, new_mask;
3866 struct task_struct *p;
3867 int retval;
3868
3869 rcu_read_lock();
3870
3871 p = find_process_by_pid(pid);
3872 if (!p) {
3873 rcu_read_unlock();
3874 return -ESRCH;
3875 }
3876
3877 /* Prevent p going away */
3878 get_task_struct(p);
3879 rcu_read_unlock();
3880
3881 if (p->flags & PF_NO_SETAFFINITY) {
3882 retval = -EINVAL;
3883 goto out_put_task;
3884 }
3885 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3886 retval = -ENOMEM;
3887 goto out_put_task;
3888 }
3889 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3890 retval = -ENOMEM;
3891 goto out_free_cpus_allowed;
3892 }
3893 retval = -EPERM;
3894 if (!check_same_owner(p)) {
3895 rcu_read_lock();
3896 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3897 rcu_read_unlock();
3898 goto out_unlock;
3899 }
3900 rcu_read_unlock();
3901 }
3902
3903 retval = security_task_setscheduler(p);
3904 if (retval)
3905 goto out_unlock;
3906
3907
3908 cpuset_cpus_allowed(p, cpus_allowed);
3909 cpumask_and(new_mask, in_mask, cpus_allowed);
3910
3911 /*
3912 * Since bandwidth control happens on root_domain basis,
3913 * if admission test is enabled, we only admit -deadline
3914 * tasks allowed to run on all the CPUs in the task's
3915 * root_domain.
3916 */
3917#ifdef CONFIG_SMP
3918 if (task_has_dl_policy(p)) {
3919 const struct cpumask *span = task_rq(p)->rd->span;
3920
3921 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
3922 retval = -EBUSY;
3923 goto out_unlock;
3924 }
3925 }
3926#endif
3927again:
3928 retval = set_cpus_allowed_ptr(p, new_mask);
3929
3930 if (!retval) {
3931 cpuset_cpus_allowed(p, cpus_allowed);
3932 if (!cpumask_subset(new_mask, cpus_allowed)) {
3933 /*
3934 * We must have raced with a concurrent cpuset
3935 * update. Just reset the cpus_allowed to the
3936 * cpuset's cpus_allowed
3937 */
3938 cpumask_copy(new_mask, cpus_allowed);
3939 goto again;
3940 }
3941 }
3942out_unlock:
3943 free_cpumask_var(new_mask);
3944out_free_cpus_allowed:
3945 free_cpumask_var(cpus_allowed);
3946out_put_task:
3947 put_task_struct(p);
3948 return retval;
3949}
3950
3951static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3952 struct cpumask *new_mask)
3953{
3954 if (len < cpumask_size())
3955 cpumask_clear(new_mask);
3956 else if (len > cpumask_size())
3957 len = cpumask_size();
3958
3959 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3960}
3961
3962/**
3963 * sys_sched_setaffinity - set the cpu affinity of a process
3964 * @pid: pid of the process
3965 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3966 * @user_mask_ptr: user-space pointer to the new cpu mask
3967 *
3968 * Return: 0 on success. An error code otherwise.
3969 */
3970SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3971 unsigned long __user *, user_mask_ptr)
3972{
3973 cpumask_var_t new_mask;
3974 int retval;
3975
3976 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3977 return -ENOMEM;
3978
3979 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3980 if (retval == 0)
3981 retval = sched_setaffinity(pid, new_mask);
3982 free_cpumask_var(new_mask);
3983 return retval;
3984}
3985
3986long sched_getaffinity(pid_t pid, struct cpumask *mask)
3987{
3988 struct task_struct *p;
3989 unsigned long flags;
3990 int retval;
3991
3992 rcu_read_lock();
3993
3994 retval = -ESRCH;
3995 p = find_process_by_pid(pid);
3996 if (!p)
3997 goto out_unlock;
3998
3999 retval = security_task_getscheduler(p);
4000 if (retval)
4001 goto out_unlock;
4002
4003 raw_spin_lock_irqsave(&p->pi_lock, flags);
4004 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4005 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4006
4007out_unlock:
4008 rcu_read_unlock();
4009
4010 return retval;
4011}
4012
4013/**
4014 * sys_sched_getaffinity - get the cpu affinity of a process
4015 * @pid: pid of the process
4016 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4017 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4018 *
4019 * Return: 0 on success. An error code otherwise.
4020 */
4021SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4022 unsigned long __user *, user_mask_ptr)
4023{
4024 int ret;
4025 cpumask_var_t mask;
4026
4027 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4028 return -EINVAL;
4029 if (len & (sizeof(unsigned long)-1))
4030 return -EINVAL;
4031
4032 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4033 return -ENOMEM;
4034
4035 ret = sched_getaffinity(pid, mask);
4036 if (ret == 0) {
4037 size_t retlen = min_t(size_t, len, cpumask_size());
4038
4039 if (copy_to_user(user_mask_ptr, mask, retlen))
4040 ret = -EFAULT;
4041 else
4042 ret = retlen;
4043 }
4044 free_cpumask_var(mask);
4045
4046 return ret;
4047}
4048
4049/**
4050 * sys_sched_yield - yield the current processor to other threads.
4051 *
4052 * This function yields the current CPU to other tasks. If there are no
4053 * other threads running on this CPU then this function will return.
4054 *
4055 * Return: 0.
4056 */
4057SYSCALL_DEFINE0(sched_yield)
4058{
4059 struct rq *rq = this_rq_lock();
4060
4061 schedstat_inc(rq, yld_count);
4062 current->sched_class->yield_task(rq);
4063
4064 /*
4065 * Since we are going to call schedule() anyway, there's
4066 * no need to preempt or enable interrupts:
4067 */
4068 __release(rq->lock);
4069 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4070 do_raw_spin_unlock(&rq->lock);
4071 sched_preempt_enable_no_resched();
4072
4073 schedule();
4074
4075 return 0;
4076}
4077
4078static void __cond_resched(void)
4079{
4080 __preempt_count_add(PREEMPT_ACTIVE);
4081 __schedule();
4082 __preempt_count_sub(PREEMPT_ACTIVE);
4083}
4084
4085int __sched _cond_resched(void)
4086{
4087 if (should_resched()) {
4088 __cond_resched();
4089 return 1;
4090 }
4091 return 0;
4092}
4093EXPORT_SYMBOL(_cond_resched);
4094
4095/*
4096 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4097 * call schedule, and on return reacquire the lock.
4098 *
4099 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4100 * operations here to prevent schedule() from being called twice (once via
4101 * spin_unlock(), once by hand).
4102 */
4103int __cond_resched_lock(spinlock_t *lock)
4104{
4105 int resched = should_resched();
4106 int ret = 0;
4107
4108 lockdep_assert_held(lock);
4109
4110 if (spin_needbreak(lock) || resched) {
4111 spin_unlock(lock);
4112 if (resched)
4113 __cond_resched();
4114 else
4115 cpu_relax();
4116 ret = 1;
4117 spin_lock(lock);
4118 }
4119 return ret;
4120}
4121EXPORT_SYMBOL(__cond_resched_lock);
4122
4123int __sched __cond_resched_softirq(void)
4124{
4125 BUG_ON(!in_softirq());
4126
4127 if (should_resched()) {
4128 local_bh_enable();
4129 __cond_resched();
4130 local_bh_disable();
4131 return 1;
4132 }
4133 return 0;
4134}
4135EXPORT_SYMBOL(__cond_resched_softirq);
4136
4137/**
4138 * yield - yield the current processor to other threads.
4139 *
4140 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4141 *
4142 * The scheduler is at all times free to pick the calling task as the most
4143 * eligible task to run, if removing the yield() call from your code breaks
4144 * it, its already broken.
4145 *
4146 * Typical broken usage is:
4147 *
4148 * while (!event)
4149 * yield();
4150 *
4151 * where one assumes that yield() will let 'the other' process run that will
4152 * make event true. If the current task is a SCHED_FIFO task that will never
4153 * happen. Never use yield() as a progress guarantee!!
4154 *
4155 * If you want to use yield() to wait for something, use wait_event().
4156 * If you want to use yield() to be 'nice' for others, use cond_resched().
4157 * If you still want to use yield(), do not!
4158 */
4159void __sched yield(void)
4160{
4161 set_current_state(TASK_RUNNING);
4162 sys_sched_yield();
4163}
4164EXPORT_SYMBOL(yield);
4165
4166/**
4167 * yield_to - yield the current processor to another thread in
4168 * your thread group, or accelerate that thread toward the
4169 * processor it's on.
4170 * @p: target task
4171 * @preempt: whether task preemption is allowed or not
4172 *
4173 * It's the caller's job to ensure that the target task struct
4174 * can't go away on us before we can do any checks.
4175 *
4176 * Return:
4177 * true (>0) if we indeed boosted the target task.
4178 * false (0) if we failed to boost the target.
4179 * -ESRCH if there's no task to yield to.
4180 */
4181bool __sched yield_to(struct task_struct *p, bool preempt)
4182{
4183 struct task_struct *curr = current;
4184 struct rq *rq, *p_rq;
4185 unsigned long flags;
4186 int yielded = 0;
4187
4188 local_irq_save(flags);
4189 rq = this_rq();
4190
4191again:
4192 p_rq = task_rq(p);
4193 /*
4194 * If we're the only runnable task on the rq and target rq also
4195 * has only one task, there's absolutely no point in yielding.
4196 */
4197 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4198 yielded = -ESRCH;
4199 goto out_irq;
4200 }
4201
4202 double_rq_lock(rq, p_rq);
4203 if (task_rq(p) != p_rq) {
4204 double_rq_unlock(rq, p_rq);
4205 goto again;
4206 }
4207
4208 if (!curr->sched_class->yield_to_task)
4209 goto out_unlock;
4210
4211 if (curr->sched_class != p->sched_class)
4212 goto out_unlock;
4213
4214 if (task_running(p_rq, p) || p->state)
4215 goto out_unlock;
4216
4217 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4218 if (yielded) {
4219 schedstat_inc(rq, yld_count);
4220 /*
4221 * Make p's CPU reschedule; pick_next_entity takes care of
4222 * fairness.
4223 */
4224 if (preempt && rq != p_rq)
4225 resched_task(p_rq->curr);
4226 }
4227
4228out_unlock:
4229 double_rq_unlock(rq, p_rq);
4230out_irq:
4231 local_irq_restore(flags);
4232
4233 if (yielded > 0)
4234 schedule();
4235
4236 return yielded;
4237}
4238EXPORT_SYMBOL_GPL(yield_to);
4239
4240/*
4241 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4242 * that process accounting knows that this is a task in IO wait state.
4243 */
4244void __sched io_schedule(void)
4245{
4246 struct rq *rq = raw_rq();
4247
4248 delayacct_blkio_start();
4249 atomic_inc(&rq->nr_iowait);
4250 blk_flush_plug(current);
4251 current->in_iowait = 1;
4252 schedule();
4253 current->in_iowait = 0;
4254 atomic_dec(&rq->nr_iowait);
4255 delayacct_blkio_end();
4256}
4257EXPORT_SYMBOL(io_schedule);
4258
4259long __sched io_schedule_timeout(long timeout)
4260{
4261 struct rq *rq = raw_rq();
4262 long ret;
4263
4264 delayacct_blkio_start();
4265 atomic_inc(&rq->nr_iowait);
4266 blk_flush_plug(current);
4267 current->in_iowait = 1;
4268 ret = schedule_timeout(timeout);
4269 current->in_iowait = 0;
4270 atomic_dec(&rq->nr_iowait);
4271 delayacct_blkio_end();
4272 return ret;
4273}
4274
4275/**
4276 * sys_sched_get_priority_max - return maximum RT priority.
4277 * @policy: scheduling class.
4278 *
4279 * Return: On success, this syscall returns the maximum
4280 * rt_priority that can be used by a given scheduling class.
4281 * On failure, a negative error code is returned.
4282 */
4283SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4284{
4285 int ret = -EINVAL;
4286
4287 switch (policy) {
4288 case SCHED_FIFO:
4289 case SCHED_RR:
4290 ret = MAX_USER_RT_PRIO-1;
4291 break;
4292 case SCHED_DEADLINE:
4293 case SCHED_NORMAL:
4294 case SCHED_BATCH:
4295 case SCHED_IDLE:
4296 ret = 0;
4297 break;
4298 }
4299 return ret;
4300}
4301
4302/**
4303 * sys_sched_get_priority_min - return minimum RT priority.
4304 * @policy: scheduling class.
4305 *
4306 * Return: On success, this syscall returns the minimum
4307 * rt_priority that can be used by a given scheduling class.
4308 * On failure, a negative error code is returned.
4309 */
4310SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4311{
4312 int ret = -EINVAL;
4313
4314 switch (policy) {
4315 case SCHED_FIFO:
4316 case SCHED_RR:
4317 ret = 1;
4318 break;
4319 case SCHED_DEADLINE:
4320 case SCHED_NORMAL:
4321 case SCHED_BATCH:
4322 case SCHED_IDLE:
4323 ret = 0;
4324 }
4325 return ret;
4326}
4327
4328/**
4329 * sys_sched_rr_get_interval - return the default timeslice of a process.
4330 * @pid: pid of the process.
4331 * @interval: userspace pointer to the timeslice value.
4332 *
4333 * this syscall writes the default timeslice value of a given process
4334 * into the user-space timespec buffer. A value of '0' means infinity.
4335 *
4336 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4337 * an error code.
4338 */
4339SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4340 struct timespec __user *, interval)
4341{
4342 struct task_struct *p;
4343 unsigned int time_slice;
4344 unsigned long flags;
4345 struct rq *rq;
4346 int retval;
4347 struct timespec t;
4348
4349 if (pid < 0)
4350 return -EINVAL;
4351
4352 retval = -ESRCH;
4353 rcu_read_lock();
4354 p = find_process_by_pid(pid);
4355 if (!p)
4356 goto out_unlock;
4357
4358 retval = security_task_getscheduler(p);
4359 if (retval)
4360 goto out_unlock;
4361
4362 rq = task_rq_lock(p, &flags);
4363 time_slice = 0;
4364 if (p->sched_class->get_rr_interval)
4365 time_slice = p->sched_class->get_rr_interval(rq, p);
4366 task_rq_unlock(rq, p, &flags);
4367
4368 rcu_read_unlock();
4369 jiffies_to_timespec(time_slice, &t);
4370 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4371 return retval;
4372
4373out_unlock:
4374 rcu_read_unlock();
4375 return retval;
4376}
4377
4378static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4379
4380void sched_show_task(struct task_struct *p)
4381{
4382 unsigned long free = 0;
4383 int ppid;
4384 unsigned state;
4385
4386 state = p->state ? __ffs(p->state) + 1 : 0;
4387 printk(KERN_INFO "%-15.15s %c", p->comm,
4388 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4389#if BITS_PER_LONG == 32
4390 if (state == TASK_RUNNING)
4391 printk(KERN_CONT " running ");
4392 else
4393 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4394#else
4395 if (state == TASK_RUNNING)
4396 printk(KERN_CONT " running task ");
4397 else
4398 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4399#endif
4400#ifdef CONFIG_DEBUG_STACK_USAGE
4401 free = stack_not_used(p);
4402#endif
4403 rcu_read_lock();
4404 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4405 rcu_read_unlock();
4406 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4407 task_pid_nr(p), ppid,
4408 (unsigned long)task_thread_info(p)->flags);
4409
4410 print_worker_info(KERN_INFO, p);
4411 show_stack(p, NULL);
4412}
4413
4414void show_state_filter(unsigned long state_filter)
4415{
4416 struct task_struct *g, *p;
4417
4418#if BITS_PER_LONG == 32
4419 printk(KERN_INFO
4420 " task PC stack pid father\n");
4421#else
4422 printk(KERN_INFO
4423 " task PC stack pid father\n");
4424#endif
4425 rcu_read_lock();
4426 do_each_thread(g, p) {
4427 /*
4428 * reset the NMI-timeout, listing all files on a slow
4429 * console might take a lot of time:
4430 */
4431 touch_nmi_watchdog();
4432 if (!state_filter || (p->state & state_filter))
4433 sched_show_task(p);
4434 } while_each_thread(g, p);
4435
4436 touch_all_softlockup_watchdogs();
4437
4438#ifdef CONFIG_SCHED_DEBUG
4439 sysrq_sched_debug_show();
4440#endif
4441 rcu_read_unlock();
4442 /*
4443 * Only show locks if all tasks are dumped:
4444 */
4445 if (!state_filter)
4446 debug_show_all_locks();
4447}
4448
4449void init_idle_bootup_task(struct task_struct *idle)
4450{
4451 idle->sched_class = &idle_sched_class;
4452}
4453
4454/**
4455 * init_idle - set up an idle thread for a given CPU
4456 * @idle: task in question
4457 * @cpu: cpu the idle task belongs to
4458 *
4459 * NOTE: this function does not set the idle thread's NEED_RESCHED
4460 * flag, to make booting more robust.
4461 */
4462void init_idle(struct task_struct *idle, int cpu)
4463{
4464 struct rq *rq = cpu_rq(cpu);
4465 unsigned long flags;
4466
4467 raw_spin_lock_irqsave(&rq->lock, flags);
4468
4469 __sched_fork(0, idle);
4470 idle->state = TASK_RUNNING;
4471 idle->se.exec_start = sched_clock();
4472
4473 do_set_cpus_allowed(idle, cpumask_of(cpu));
4474 /*
4475 * We're having a chicken and egg problem, even though we are
4476 * holding rq->lock, the cpu isn't yet set to this cpu so the
4477 * lockdep check in task_group() will fail.
4478 *
4479 * Similar case to sched_fork(). / Alternatively we could
4480 * use task_rq_lock() here and obtain the other rq->lock.
4481 *
4482 * Silence PROVE_RCU
4483 */
4484 rcu_read_lock();
4485 __set_task_cpu(idle, cpu);
4486 rcu_read_unlock();
4487
4488 rq->curr = rq->idle = idle;
4489 idle->on_rq = 1;
4490#if defined(CONFIG_SMP)
4491 idle->on_cpu = 1;
4492#endif
4493 raw_spin_unlock_irqrestore(&rq->lock, flags);
4494
4495 /* Set the preempt count _outside_ the spinlocks! */
4496 init_idle_preempt_count(idle, cpu);
4497
4498 /*
4499 * The idle tasks have their own, simple scheduling class:
4500 */
4501 idle->sched_class = &idle_sched_class;
4502 ftrace_graph_init_idle_task(idle, cpu);
4503 vtime_init_idle(idle, cpu);
4504#if defined(CONFIG_SMP)
4505 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4506#endif
4507}
4508
4509#ifdef CONFIG_SMP
4510void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4511{
4512 if (p->sched_class && p->sched_class->set_cpus_allowed)
4513 p->sched_class->set_cpus_allowed(p, new_mask);
4514
4515 cpumask_copy(&p->cpus_allowed, new_mask);
4516 p->nr_cpus_allowed = cpumask_weight(new_mask);
4517}
4518
4519/*
4520 * This is how migration works:
4521 *
4522 * 1) we invoke migration_cpu_stop() on the target CPU using
4523 * stop_one_cpu().
4524 * 2) stopper starts to run (implicitly forcing the migrated thread
4525 * off the CPU)
4526 * 3) it checks whether the migrated task is still in the wrong runqueue.
4527 * 4) if it's in the wrong runqueue then the migration thread removes
4528 * it and puts it into the right queue.
4529 * 5) stopper completes and stop_one_cpu() returns and the migration
4530 * is done.
4531 */
4532
4533/*
4534 * Change a given task's CPU affinity. Migrate the thread to a
4535 * proper CPU and schedule it away if the CPU it's executing on
4536 * is removed from the allowed bitmask.
4537 *
4538 * NOTE: the caller must have a valid reference to the task, the
4539 * task must not exit() & deallocate itself prematurely. The
4540 * call is not atomic; no spinlocks may be held.
4541 */
4542int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4543{
4544 unsigned long flags;
4545 struct rq *rq;
4546 unsigned int dest_cpu;
4547 int ret = 0;
4548
4549 rq = task_rq_lock(p, &flags);
4550
4551 if (cpumask_equal(&p->cpus_allowed, new_mask))
4552 goto out;
4553
4554 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4555 ret = -EINVAL;
4556 goto out;
4557 }
4558
4559 do_set_cpus_allowed(p, new_mask);
4560
4561 /* Can the task run on the task's current CPU? If so, we're done */
4562 if (cpumask_test_cpu(task_cpu(p), new_mask))
4563 goto out;
4564
4565 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4566 if (p->on_rq) {
4567 struct migration_arg arg = { p, dest_cpu };
4568 /* Need help from migration thread: drop lock and wait. */
4569 task_rq_unlock(rq, p, &flags);
4570 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4571 tlb_migrate_finish(p->mm);
4572 return 0;
4573 }
4574out:
4575 task_rq_unlock(rq, p, &flags);
4576
4577 return ret;
4578}
4579EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4580
4581/*
4582 * Move (not current) task off this cpu, onto dest cpu. We're doing
4583 * this because either it can't run here any more (set_cpus_allowed()
4584 * away from this CPU, or CPU going down), or because we're
4585 * attempting to rebalance this task on exec (sched_exec).
4586 *
4587 * So we race with normal scheduler movements, but that's OK, as long
4588 * as the task is no longer on this CPU.
4589 *
4590 * Returns non-zero if task was successfully migrated.
4591 */
4592static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4593{
4594 struct rq *rq_dest, *rq_src;
4595 int ret = 0;
4596
4597 if (unlikely(!cpu_active(dest_cpu)))
4598 return ret;
4599
4600 rq_src = cpu_rq(src_cpu);
4601 rq_dest = cpu_rq(dest_cpu);
4602
4603 raw_spin_lock(&p->pi_lock);
4604 double_rq_lock(rq_src, rq_dest);
4605 /* Already moved. */
4606 if (task_cpu(p) != src_cpu)
4607 goto done;
4608 /* Affinity changed (again). */
4609 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4610 goto fail;
4611
4612 /*
4613 * If we're not on a rq, the next wake-up will ensure we're
4614 * placed properly.
4615 */
4616 if (p->on_rq) {
4617 dequeue_task(rq_src, p, 0);
4618 set_task_cpu(p, dest_cpu);
4619 enqueue_task(rq_dest, p, 0);
4620 check_preempt_curr(rq_dest, p, 0);
4621 }
4622done:
4623 ret = 1;
4624fail:
4625 double_rq_unlock(rq_src, rq_dest);
4626 raw_spin_unlock(&p->pi_lock);
4627 return ret;
4628}
4629
4630#ifdef CONFIG_NUMA_BALANCING
4631/* Migrate current task p to target_cpu */
4632int migrate_task_to(struct task_struct *p, int target_cpu)
4633{
4634 struct migration_arg arg = { p, target_cpu };
4635 int curr_cpu = task_cpu(p);
4636
4637 if (curr_cpu == target_cpu)
4638 return 0;
4639
4640 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4641 return -EINVAL;
4642
4643 /* TODO: This is not properly updating schedstats */
4644
4645 trace_sched_move_numa(p, curr_cpu, target_cpu);
4646 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4647}
4648
4649/*
4650 * Requeue a task on a given node and accurately track the number of NUMA
4651 * tasks on the runqueues
4652 */
4653void sched_setnuma(struct task_struct *p, int nid)
4654{
4655 struct rq *rq;
4656 unsigned long flags;
4657 bool on_rq, running;
4658
4659 rq = task_rq_lock(p, &flags);
4660 on_rq = p->on_rq;
4661 running = task_current(rq, p);
4662
4663 if (on_rq)
4664 dequeue_task(rq, p, 0);
4665 if (running)
4666 p->sched_class->put_prev_task(rq, p);
4667
4668 p->numa_preferred_nid = nid;
4669
4670 if (running)
4671 p->sched_class->set_curr_task(rq);
4672 if (on_rq)
4673 enqueue_task(rq, p, 0);
4674 task_rq_unlock(rq, p, &flags);
4675}
4676#endif
4677
4678/*
4679 * migration_cpu_stop - this will be executed by a highprio stopper thread
4680 * and performs thread migration by bumping thread off CPU then
4681 * 'pushing' onto another runqueue.
4682 */
4683static int migration_cpu_stop(void *data)
4684{
4685 struct migration_arg *arg = data;
4686
4687 /*
4688 * The original target cpu might have gone down and we might
4689 * be on another cpu but it doesn't matter.
4690 */
4691 local_irq_disable();
4692 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4693 local_irq_enable();
4694 return 0;
4695}
4696
4697#ifdef CONFIG_HOTPLUG_CPU
4698
4699/*
4700 * Ensures that the idle task is using init_mm right before its cpu goes
4701 * offline.
4702 */
4703void idle_task_exit(void)
4704{
4705 struct mm_struct *mm = current->active_mm;
4706
4707 BUG_ON(cpu_online(smp_processor_id()));
4708
4709 if (mm != &init_mm) {
4710 switch_mm(mm, &init_mm, current);
4711 finish_arch_post_lock_switch();
4712 }
4713 mmdrop(mm);
4714}
4715
4716/*
4717 * Since this CPU is going 'away' for a while, fold any nr_active delta
4718 * we might have. Assumes we're called after migrate_tasks() so that the
4719 * nr_active count is stable.
4720 *
4721 * Also see the comment "Global load-average calculations".
4722 */
4723static void calc_load_migrate(struct rq *rq)
4724{
4725 long delta = calc_load_fold_active(rq);
4726 if (delta)
4727 atomic_long_add(delta, &calc_load_tasks);
4728}
4729
4730static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4731{
4732}
4733
4734static const struct sched_class fake_sched_class = {
4735 .put_prev_task = put_prev_task_fake,
4736};
4737
4738static struct task_struct fake_task = {
4739 /*
4740 * Avoid pull_{rt,dl}_task()
4741 */
4742 .prio = MAX_PRIO + 1,
4743 .sched_class = &fake_sched_class,
4744};
4745
4746/*
4747 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4748 * try_to_wake_up()->select_task_rq().
4749 *
4750 * Called with rq->lock held even though we'er in stop_machine() and
4751 * there's no concurrency possible, we hold the required locks anyway
4752 * because of lock validation efforts.
4753 */
4754static void migrate_tasks(unsigned int dead_cpu)
4755{
4756 struct rq *rq = cpu_rq(dead_cpu);
4757 struct task_struct *next, *stop = rq->stop;
4758 int dest_cpu;
4759
4760 /*
4761 * Fudge the rq selection such that the below task selection loop
4762 * doesn't get stuck on the currently eligible stop task.
4763 *
4764 * We're currently inside stop_machine() and the rq is either stuck
4765 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4766 * either way we should never end up calling schedule() until we're
4767 * done here.
4768 */
4769 rq->stop = NULL;
4770
4771 /*
4772 * put_prev_task() and pick_next_task() sched
4773 * class method both need to have an up-to-date
4774 * value of rq->clock[_task]
4775 */
4776 update_rq_clock(rq);
4777
4778 for ( ; ; ) {
4779 /*
4780 * There's this thread running, bail when that's the only
4781 * remaining thread.
4782 */
4783 if (rq->nr_running == 1)
4784 break;
4785
4786 next = pick_next_task(rq, &fake_task);
4787 BUG_ON(!next);
4788 next->sched_class->put_prev_task(rq, next);
4789
4790 /* Find suitable destination for @next, with force if needed. */
4791 dest_cpu = select_fallback_rq(dead_cpu, next);
4792 raw_spin_unlock(&rq->lock);
4793
4794 __migrate_task(next, dead_cpu, dest_cpu);
4795
4796 raw_spin_lock(&rq->lock);
4797 }
4798
4799 rq->stop = stop;
4800}
4801
4802#endif /* CONFIG_HOTPLUG_CPU */
4803
4804#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4805
4806static struct ctl_table sd_ctl_dir[] = {
4807 {
4808 .procname = "sched_domain",
4809 .mode = 0555,
4810 },
4811 {}
4812};
4813
4814static struct ctl_table sd_ctl_root[] = {
4815 {
4816 .procname = "kernel",
4817 .mode = 0555,
4818 .child = sd_ctl_dir,
4819 },
4820 {}
4821};
4822
4823static struct ctl_table *sd_alloc_ctl_entry(int n)
4824{
4825 struct ctl_table *entry =
4826 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4827
4828 return entry;
4829}
4830
4831static void sd_free_ctl_entry(struct ctl_table **tablep)
4832{
4833 struct ctl_table *entry;
4834
4835 /*
4836 * In the intermediate directories, both the child directory and
4837 * procname are dynamically allocated and could fail but the mode
4838 * will always be set. In the lowest directory the names are
4839 * static strings and all have proc handlers.
4840 */
4841 for (entry = *tablep; entry->mode; entry++) {
4842 if (entry->child)
4843 sd_free_ctl_entry(&entry->child);
4844 if (entry->proc_handler == NULL)
4845 kfree(entry->procname);
4846 }
4847
4848 kfree(*tablep);
4849 *tablep = NULL;
4850}
4851
4852static int min_load_idx = 0;
4853static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4854
4855static void
4856set_table_entry(struct ctl_table *entry,
4857 const char *procname, void *data, int maxlen,
4858 umode_t mode, proc_handler *proc_handler,
4859 bool load_idx)
4860{
4861 entry->procname = procname;
4862 entry->data = data;
4863 entry->maxlen = maxlen;
4864 entry->mode = mode;
4865 entry->proc_handler = proc_handler;
4866
4867 if (load_idx) {
4868 entry->extra1 = &min_load_idx;
4869 entry->extra2 = &max_load_idx;
4870 }
4871}
4872
4873static struct ctl_table *
4874sd_alloc_ctl_domain_table(struct sched_domain *sd)
4875{
4876 struct ctl_table *table = sd_alloc_ctl_entry(14);
4877
4878 if (table == NULL)
4879 return NULL;
4880
4881 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4882 sizeof(long), 0644, proc_doulongvec_minmax, false);
4883 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4884 sizeof(long), 0644, proc_doulongvec_minmax, false);
4885 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4886 sizeof(int), 0644, proc_dointvec_minmax, true);
4887 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4888 sizeof(int), 0644, proc_dointvec_minmax, true);
4889 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4890 sizeof(int), 0644, proc_dointvec_minmax, true);
4891 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4892 sizeof(int), 0644, proc_dointvec_minmax, true);
4893 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4894 sizeof(int), 0644, proc_dointvec_minmax, true);
4895 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4896 sizeof(int), 0644, proc_dointvec_minmax, false);
4897 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4898 sizeof(int), 0644, proc_dointvec_minmax, false);
4899 set_table_entry(&table[9], "cache_nice_tries",
4900 &sd->cache_nice_tries,
4901 sizeof(int), 0644, proc_dointvec_minmax, false);
4902 set_table_entry(&table[10], "flags", &sd->flags,
4903 sizeof(int), 0644, proc_dointvec_minmax, false);
4904 set_table_entry(&table[11], "max_newidle_lb_cost",
4905 &sd->max_newidle_lb_cost,
4906 sizeof(long), 0644, proc_doulongvec_minmax, false);
4907 set_table_entry(&table[12], "name", sd->name,
4908 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4909 /* &table[13] is terminator */
4910
4911 return table;
4912}
4913
4914static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4915{
4916 struct ctl_table *entry, *table;
4917 struct sched_domain *sd;
4918 int domain_num = 0, i;
4919 char buf[32];
4920
4921 for_each_domain(cpu, sd)
4922 domain_num++;
4923 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4924 if (table == NULL)
4925 return NULL;
4926
4927 i = 0;
4928 for_each_domain(cpu, sd) {
4929 snprintf(buf, 32, "domain%d", i);
4930 entry->procname = kstrdup(buf, GFP_KERNEL);
4931 entry->mode = 0555;
4932 entry->child = sd_alloc_ctl_domain_table(sd);
4933 entry++;
4934 i++;
4935 }
4936 return table;
4937}
4938
4939static struct ctl_table_header *sd_sysctl_header;
4940static void register_sched_domain_sysctl(void)
4941{
4942 int i, cpu_num = num_possible_cpus();
4943 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4944 char buf[32];
4945
4946 WARN_ON(sd_ctl_dir[0].child);
4947 sd_ctl_dir[0].child = entry;
4948
4949 if (entry == NULL)
4950 return;
4951
4952 for_each_possible_cpu(i) {
4953 snprintf(buf, 32, "cpu%d", i);
4954 entry->procname = kstrdup(buf, GFP_KERNEL);
4955 entry->mode = 0555;
4956 entry->child = sd_alloc_ctl_cpu_table(i);
4957 entry++;
4958 }
4959
4960 WARN_ON(sd_sysctl_header);
4961 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4962}
4963
4964/* may be called multiple times per register */
4965static void unregister_sched_domain_sysctl(void)
4966{
4967 if (sd_sysctl_header)
4968 unregister_sysctl_table(sd_sysctl_header);
4969 sd_sysctl_header = NULL;
4970 if (sd_ctl_dir[0].child)
4971 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4972}
4973#else
4974static void register_sched_domain_sysctl(void)
4975{
4976}
4977static void unregister_sched_domain_sysctl(void)
4978{
4979}
4980#endif
4981
4982static void set_rq_online(struct rq *rq)
4983{
4984 if (!rq->online) {
4985 const struct sched_class *class;
4986
4987 cpumask_set_cpu(rq->cpu, rq->rd->online);
4988 rq->online = 1;
4989
4990 for_each_class(class) {
4991 if (class->rq_online)
4992 class->rq_online(rq);
4993 }
4994 }
4995}
4996
4997static void set_rq_offline(struct rq *rq)
4998{
4999 if (rq->online) {
5000 const struct sched_class *class;
5001
5002 for_each_class(class) {
5003 if (class->rq_offline)
5004 class->rq_offline(rq);
5005 }
5006
5007 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5008 rq->online = 0;
5009 }
5010}
5011
5012/*
5013 * migration_call - callback that gets triggered when a CPU is added.
5014 * Here we can start up the necessary migration thread for the new CPU.
5015 */
5016static int
5017migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5018{
5019 int cpu = (long)hcpu;
5020 unsigned long flags;
5021 struct rq *rq = cpu_rq(cpu);
5022
5023 switch (action & ~CPU_TASKS_FROZEN) {
5024
5025 case CPU_UP_PREPARE:
5026 rq->calc_load_update = calc_load_update;
5027 break;
5028
5029 case CPU_ONLINE:
5030 /* Update our root-domain */
5031 raw_spin_lock_irqsave(&rq->lock, flags);
5032 if (rq->rd) {
5033 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5034
5035 set_rq_online(rq);
5036 }
5037 raw_spin_unlock_irqrestore(&rq->lock, flags);
5038 break;
5039
5040#ifdef CONFIG_HOTPLUG_CPU
5041 case CPU_DYING:
5042 sched_ttwu_pending();
5043 /* Update our root-domain */
5044 raw_spin_lock_irqsave(&rq->lock, flags);
5045 if (rq->rd) {
5046 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5047 set_rq_offline(rq);
5048 }
5049 migrate_tasks(cpu);
5050 BUG_ON(rq->nr_running != 1); /* the migration thread */
5051 raw_spin_unlock_irqrestore(&rq->lock, flags);
5052 break;
5053
5054 case CPU_DEAD:
5055 calc_load_migrate(rq);
5056 break;
5057#endif
5058 }
5059
5060 update_max_interval();
5061
5062 return NOTIFY_OK;
5063}
5064
5065/*
5066 * Register at high priority so that task migration (migrate_all_tasks)
5067 * happens before everything else. This has to be lower priority than
5068 * the notifier in the perf_event subsystem, though.
5069 */
5070static struct notifier_block migration_notifier = {
5071 .notifier_call = migration_call,
5072 .priority = CPU_PRI_MIGRATION,
5073};
5074
5075static int sched_cpu_active(struct notifier_block *nfb,
5076 unsigned long action, void *hcpu)
5077{
5078 switch (action & ~CPU_TASKS_FROZEN) {
5079 case CPU_DOWN_FAILED:
5080 set_cpu_active((long)hcpu, true);
5081 return NOTIFY_OK;
5082 default:
5083 return NOTIFY_DONE;
5084 }
5085}
5086
5087static int sched_cpu_inactive(struct notifier_block *nfb,
5088 unsigned long action, void *hcpu)
5089{
5090 unsigned long flags;
5091 long cpu = (long)hcpu;
5092
5093 switch (action & ~CPU_TASKS_FROZEN) {
5094 case CPU_DOWN_PREPARE:
5095 set_cpu_active(cpu, false);
5096
5097 /* explicitly allow suspend */
5098 if (!(action & CPU_TASKS_FROZEN)) {
5099 struct dl_bw *dl_b = dl_bw_of(cpu);
5100 bool overflow;
5101 int cpus;
5102
5103 raw_spin_lock_irqsave(&dl_b->lock, flags);
5104 cpus = dl_bw_cpus(cpu);
5105 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5106 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5107
5108 if (overflow)
5109 return notifier_from_errno(-EBUSY);
5110 }
5111 return NOTIFY_OK;
5112 }
5113
5114 return NOTIFY_DONE;
5115}
5116
5117static int __init migration_init(void)
5118{
5119 void *cpu = (void *)(long)smp_processor_id();
5120 int err;
5121
5122 /* Initialize migration for the boot CPU */
5123 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5124 BUG_ON(err == NOTIFY_BAD);
5125 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5126 register_cpu_notifier(&migration_notifier);
5127
5128 /* Register cpu active notifiers */
5129 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5130 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5131
5132 return 0;
5133}
5134early_initcall(migration_init);
5135#endif
5136
5137#ifdef CONFIG_SMP
5138
5139static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5140
5141#ifdef CONFIG_SCHED_DEBUG
5142
5143static __read_mostly int sched_debug_enabled;
5144
5145static int __init sched_debug_setup(char *str)
5146{
5147 sched_debug_enabled = 1;
5148
5149 return 0;
5150}
5151early_param("sched_debug", sched_debug_setup);
5152
5153static inline bool sched_debug(void)
5154{
5155 return sched_debug_enabled;
5156}
5157
5158static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5159 struct cpumask *groupmask)
5160{
5161 struct sched_group *group = sd->groups;
5162 char str[256];
5163
5164 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5165 cpumask_clear(groupmask);
5166
5167 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5168
5169 if (!(sd->flags & SD_LOAD_BALANCE)) {
5170 printk("does not load-balance\n");
5171 if (sd->parent)
5172 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5173 " has parent");
5174 return -1;
5175 }
5176
5177 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5178
5179 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5180 printk(KERN_ERR "ERROR: domain->span does not contain "
5181 "CPU%d\n", cpu);
5182 }
5183 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5184 printk(KERN_ERR "ERROR: domain->groups does not contain"
5185 " CPU%d\n", cpu);
5186 }
5187
5188 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5189 do {
5190 if (!group) {
5191 printk("\n");
5192 printk(KERN_ERR "ERROR: group is NULL\n");
5193 break;
5194 }
5195
5196 /*
5197 * Even though we initialize ->power to something semi-sane,
5198 * we leave power_orig unset. This allows us to detect if
5199 * domain iteration is still funny without causing /0 traps.
5200 */
5201 if (!group->sgp->power_orig) {
5202 printk(KERN_CONT "\n");
5203 printk(KERN_ERR "ERROR: domain->cpu_power not "
5204 "set\n");
5205 break;
5206 }
5207
5208 if (!cpumask_weight(sched_group_cpus(group))) {
5209 printk(KERN_CONT "\n");
5210 printk(KERN_ERR "ERROR: empty group\n");
5211 break;
5212 }
5213
5214 if (!(sd->flags & SD_OVERLAP) &&
5215 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5216 printk(KERN_CONT "\n");
5217 printk(KERN_ERR "ERROR: repeated CPUs\n");
5218 break;
5219 }
5220
5221 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5222
5223 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5224
5225 printk(KERN_CONT " %s", str);
5226 if (group->sgp->power != SCHED_POWER_SCALE) {
5227 printk(KERN_CONT " (cpu_power = %d)",
5228 group->sgp->power);
5229 }
5230
5231 group = group->next;
5232 } while (group != sd->groups);
5233 printk(KERN_CONT "\n");
5234
5235 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5236 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5237
5238 if (sd->parent &&
5239 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5240 printk(KERN_ERR "ERROR: parent span is not a superset "
5241 "of domain->span\n");
5242 return 0;
5243}
5244
5245static void sched_domain_debug(struct sched_domain *sd, int cpu)
5246{
5247 int level = 0;
5248
5249 if (!sched_debug_enabled)
5250 return;
5251
5252 if (!sd) {
5253 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5254 return;
5255 }
5256
5257 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5258
5259 for (;;) {
5260 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5261 break;
5262 level++;
5263 sd = sd->parent;
5264 if (!sd)
5265 break;
5266 }
5267}
5268#else /* !CONFIG_SCHED_DEBUG */
5269# define sched_domain_debug(sd, cpu) do { } while (0)
5270static inline bool sched_debug(void)
5271{
5272 return false;
5273}
5274#endif /* CONFIG_SCHED_DEBUG */
5275
5276static int sd_degenerate(struct sched_domain *sd)
5277{
5278 if (cpumask_weight(sched_domain_span(sd)) == 1)
5279 return 1;
5280
5281 /* Following flags need at least 2 groups */
5282 if (sd->flags & (SD_LOAD_BALANCE |
5283 SD_BALANCE_NEWIDLE |
5284 SD_BALANCE_FORK |
5285 SD_BALANCE_EXEC |
5286 SD_SHARE_CPUPOWER |
5287 SD_SHARE_PKG_RESOURCES)) {
5288 if (sd->groups != sd->groups->next)
5289 return 0;
5290 }
5291
5292 /* Following flags don't use groups */
5293 if (sd->flags & (SD_WAKE_AFFINE))
5294 return 0;
5295
5296 return 1;
5297}
5298
5299static int
5300sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5301{
5302 unsigned long cflags = sd->flags, pflags = parent->flags;
5303
5304 if (sd_degenerate(parent))
5305 return 1;
5306
5307 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5308 return 0;
5309
5310 /* Flags needing groups don't count if only 1 group in parent */
5311 if (parent->groups == parent->groups->next) {
5312 pflags &= ~(SD_LOAD_BALANCE |
5313 SD_BALANCE_NEWIDLE |
5314 SD_BALANCE_FORK |
5315 SD_BALANCE_EXEC |
5316 SD_SHARE_CPUPOWER |
5317 SD_SHARE_PKG_RESOURCES |
5318 SD_PREFER_SIBLING);
5319 if (nr_node_ids == 1)
5320 pflags &= ~SD_SERIALIZE;
5321 }
5322 if (~cflags & pflags)
5323 return 0;
5324
5325 return 1;
5326}
5327
5328static void free_rootdomain(struct rcu_head *rcu)
5329{
5330 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5331
5332 cpupri_cleanup(&rd->cpupri);
5333 cpudl_cleanup(&rd->cpudl);
5334 free_cpumask_var(rd->dlo_mask);
5335 free_cpumask_var(rd->rto_mask);
5336 free_cpumask_var(rd->online);
5337 free_cpumask_var(rd->span);
5338 kfree(rd);
5339}
5340
5341static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5342{
5343 struct root_domain *old_rd = NULL;
5344 unsigned long flags;
5345
5346 raw_spin_lock_irqsave(&rq->lock, flags);
5347
5348 if (rq->rd) {
5349 old_rd = rq->rd;
5350
5351 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5352 set_rq_offline(rq);
5353
5354 cpumask_clear_cpu(rq->cpu, old_rd->span);
5355
5356 /*
5357 * If we dont want to free the old_rd yet then
5358 * set old_rd to NULL to skip the freeing later
5359 * in this function:
5360 */
5361 if (!atomic_dec_and_test(&old_rd->refcount))
5362 old_rd = NULL;
5363 }
5364
5365 atomic_inc(&rd->refcount);
5366 rq->rd = rd;
5367
5368 cpumask_set_cpu(rq->cpu, rd->span);
5369 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5370 set_rq_online(rq);
5371
5372 raw_spin_unlock_irqrestore(&rq->lock, flags);
5373
5374 if (old_rd)
5375 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5376}
5377
5378static int init_rootdomain(struct root_domain *rd)
5379{
5380 memset(rd, 0, sizeof(*rd));
5381
5382 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5383 goto out;
5384 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5385 goto free_span;
5386 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5387 goto free_online;
5388 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5389 goto free_dlo_mask;
5390
5391 init_dl_bw(&rd->dl_bw);
5392 if (cpudl_init(&rd->cpudl) != 0)
5393 goto free_dlo_mask;
5394
5395 if (cpupri_init(&rd->cpupri) != 0)
5396 goto free_rto_mask;
5397 return 0;
5398
5399free_rto_mask:
5400 free_cpumask_var(rd->rto_mask);
5401free_dlo_mask:
5402 free_cpumask_var(rd->dlo_mask);
5403free_online:
5404 free_cpumask_var(rd->online);
5405free_span:
5406 free_cpumask_var(rd->span);
5407out:
5408 return -ENOMEM;
5409}
5410
5411/*
5412 * By default the system creates a single root-domain with all cpus as
5413 * members (mimicking the global state we have today).
5414 */
5415struct root_domain def_root_domain;
5416
5417static void init_defrootdomain(void)
5418{
5419 init_rootdomain(&def_root_domain);
5420
5421 atomic_set(&def_root_domain.refcount, 1);
5422}
5423
5424static struct root_domain *alloc_rootdomain(void)
5425{
5426 struct root_domain *rd;
5427
5428 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5429 if (!rd)
5430 return NULL;
5431
5432 if (init_rootdomain(rd) != 0) {
5433 kfree(rd);
5434 return NULL;
5435 }
5436
5437 return rd;
5438}
5439
5440static void free_sched_groups(struct sched_group *sg, int free_sgp)
5441{
5442 struct sched_group *tmp, *first;
5443
5444 if (!sg)
5445 return;
5446
5447 first = sg;
5448 do {
5449 tmp = sg->next;
5450
5451 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5452 kfree(sg->sgp);
5453
5454 kfree(sg);
5455 sg = tmp;
5456 } while (sg != first);
5457}
5458
5459static void free_sched_domain(struct rcu_head *rcu)
5460{
5461 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5462
5463 /*
5464 * If its an overlapping domain it has private groups, iterate and
5465 * nuke them all.
5466 */
5467 if (sd->flags & SD_OVERLAP) {
5468 free_sched_groups(sd->groups, 1);
5469 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5470 kfree(sd->groups->sgp);
5471 kfree(sd->groups);
5472 }
5473 kfree(sd);
5474}
5475
5476static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5477{
5478 call_rcu(&sd->rcu, free_sched_domain);
5479}
5480
5481static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5482{
5483 for (; sd; sd = sd->parent)
5484 destroy_sched_domain(sd, cpu);
5485}
5486
5487/*
5488 * Keep a special pointer to the highest sched_domain that has
5489 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5490 * allows us to avoid some pointer chasing select_idle_sibling().
5491 *
5492 * Also keep a unique ID per domain (we use the first cpu number in
5493 * the cpumask of the domain), this allows us to quickly tell if
5494 * two cpus are in the same cache domain, see cpus_share_cache().
5495 */
5496DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5497DEFINE_PER_CPU(int, sd_llc_size);
5498DEFINE_PER_CPU(int, sd_llc_id);
5499DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5500DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5501DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5502
5503static void update_top_cache_domain(int cpu)
5504{
5505 struct sched_domain *sd;
5506 struct sched_domain *busy_sd = NULL;
5507 int id = cpu;
5508 int size = 1;
5509
5510 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5511 if (sd) {
5512 id = cpumask_first(sched_domain_span(sd));
5513 size = cpumask_weight(sched_domain_span(sd));
5514 busy_sd = sd->parent; /* sd_busy */
5515 }
5516 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5517
5518 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5519 per_cpu(sd_llc_size, cpu) = size;
5520 per_cpu(sd_llc_id, cpu) = id;
5521
5522 sd = lowest_flag_domain(cpu, SD_NUMA);
5523 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5524
5525 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5526 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5527}
5528
5529/*
5530 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5531 * hold the hotplug lock.
5532 */
5533static void
5534cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5535{
5536 struct rq *rq = cpu_rq(cpu);
5537 struct sched_domain *tmp;
5538
5539 /* Remove the sched domains which do not contribute to scheduling. */
5540 for (tmp = sd; tmp; ) {
5541 struct sched_domain *parent = tmp->parent;
5542 if (!parent)
5543 break;
5544
5545 if (sd_parent_degenerate(tmp, parent)) {
5546 tmp->parent = parent->parent;
5547 if (parent->parent)
5548 parent->parent->child = tmp;
5549 /*
5550 * Transfer SD_PREFER_SIBLING down in case of a
5551 * degenerate parent; the spans match for this
5552 * so the property transfers.
5553 */
5554 if (parent->flags & SD_PREFER_SIBLING)
5555 tmp->flags |= SD_PREFER_SIBLING;
5556 destroy_sched_domain(parent, cpu);
5557 } else
5558 tmp = tmp->parent;
5559 }
5560
5561 if (sd && sd_degenerate(sd)) {
5562 tmp = sd;
5563 sd = sd->parent;
5564 destroy_sched_domain(tmp, cpu);
5565 if (sd)
5566 sd->child = NULL;
5567 }
5568
5569 sched_domain_debug(sd, cpu);
5570
5571 rq_attach_root(rq, rd);
5572 tmp = rq->sd;
5573 rcu_assign_pointer(rq->sd, sd);
5574 destroy_sched_domains(tmp, cpu);
5575
5576 update_top_cache_domain(cpu);
5577}
5578
5579/* cpus with isolated domains */
5580static cpumask_var_t cpu_isolated_map;
5581
5582/* Setup the mask of cpus configured for isolated domains */
5583static int __init isolated_cpu_setup(char *str)
5584{
5585 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5586 cpulist_parse(str, cpu_isolated_map);
5587 return 1;
5588}
5589
5590__setup("isolcpus=", isolated_cpu_setup);
5591
5592static const struct cpumask *cpu_cpu_mask(int cpu)
5593{
5594 return cpumask_of_node(cpu_to_node(cpu));
5595}
5596
5597struct sd_data {
5598 struct sched_domain **__percpu sd;
5599 struct sched_group **__percpu sg;
5600 struct sched_group_power **__percpu sgp;
5601};
5602
5603struct s_data {
5604 struct sched_domain ** __percpu sd;
5605 struct root_domain *rd;
5606};
5607
5608enum s_alloc {
5609 sa_rootdomain,
5610 sa_sd,
5611 sa_sd_storage,
5612 sa_none,
5613};
5614
5615struct sched_domain_topology_level;
5616
5617typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5618typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5619
5620#define SDTL_OVERLAP 0x01
5621
5622struct sched_domain_topology_level {
5623 sched_domain_init_f init;
5624 sched_domain_mask_f mask;
5625 int flags;
5626 int numa_level;
5627 struct sd_data data;
5628};
5629
5630/*
5631 * Build an iteration mask that can exclude certain CPUs from the upwards
5632 * domain traversal.
5633 *
5634 * Asymmetric node setups can result in situations where the domain tree is of
5635 * unequal depth, make sure to skip domains that already cover the entire
5636 * range.
5637 *
5638 * In that case build_sched_domains() will have terminated the iteration early
5639 * and our sibling sd spans will be empty. Domains should always include the
5640 * cpu they're built on, so check that.
5641 *
5642 */
5643static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5644{
5645 const struct cpumask *span = sched_domain_span(sd);
5646 struct sd_data *sdd = sd->private;
5647 struct sched_domain *sibling;
5648 int i;
5649
5650 for_each_cpu(i, span) {
5651 sibling = *per_cpu_ptr(sdd->sd, i);
5652 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5653 continue;
5654
5655 cpumask_set_cpu(i, sched_group_mask(sg));
5656 }
5657}
5658
5659/*
5660 * Return the canonical balance cpu for this group, this is the first cpu
5661 * of this group that's also in the iteration mask.
5662 */
5663int group_balance_cpu(struct sched_group *sg)
5664{
5665 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5666}
5667
5668static int
5669build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5670{
5671 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5672 const struct cpumask *span = sched_domain_span(sd);
5673 struct cpumask *covered = sched_domains_tmpmask;
5674 struct sd_data *sdd = sd->private;
5675 struct sched_domain *child;
5676 int i;
5677
5678 cpumask_clear(covered);
5679
5680 for_each_cpu(i, span) {
5681 struct cpumask *sg_span;
5682
5683 if (cpumask_test_cpu(i, covered))
5684 continue;
5685
5686 child = *per_cpu_ptr(sdd->sd, i);
5687
5688 /* See the comment near build_group_mask(). */
5689 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5690 continue;
5691
5692 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5693 GFP_KERNEL, cpu_to_node(cpu));
5694
5695 if (!sg)
5696 goto fail;
5697
5698 sg_span = sched_group_cpus(sg);
5699 if (child->child) {
5700 child = child->child;
5701 cpumask_copy(sg_span, sched_domain_span(child));
5702 } else
5703 cpumask_set_cpu(i, sg_span);
5704
5705 cpumask_or(covered, covered, sg_span);
5706
5707 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5708 if (atomic_inc_return(&sg->sgp->ref) == 1)
5709 build_group_mask(sd, sg);
5710
5711 /*
5712 * Initialize sgp->power such that even if we mess up the
5713 * domains and no possible iteration will get us here, we won't
5714 * die on a /0 trap.
5715 */
5716 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5717 sg->sgp->power_orig = sg->sgp->power;
5718
5719 /*
5720 * Make sure the first group of this domain contains the
5721 * canonical balance cpu. Otherwise the sched_domain iteration
5722 * breaks. See update_sg_lb_stats().
5723 */
5724 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5725 group_balance_cpu(sg) == cpu)
5726 groups = sg;
5727
5728 if (!first)
5729 first = sg;
5730 if (last)
5731 last->next = sg;
5732 last = sg;
5733 last->next = first;
5734 }
5735 sd->groups = groups;
5736
5737 return 0;
5738
5739fail:
5740 free_sched_groups(first, 0);
5741
5742 return -ENOMEM;
5743}
5744
5745static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5746{
5747 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5748 struct sched_domain *child = sd->child;
5749
5750 if (child)
5751 cpu = cpumask_first(sched_domain_span(child));
5752
5753 if (sg) {
5754 *sg = *per_cpu_ptr(sdd->sg, cpu);
5755 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5756 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5757 }
5758
5759 return cpu;
5760}
5761
5762/*
5763 * build_sched_groups will build a circular linked list of the groups
5764 * covered by the given span, and will set each group's ->cpumask correctly,
5765 * and ->cpu_power to 0.
5766 *
5767 * Assumes the sched_domain tree is fully constructed
5768 */
5769static int
5770build_sched_groups(struct sched_domain *sd, int cpu)
5771{
5772 struct sched_group *first = NULL, *last = NULL;
5773 struct sd_data *sdd = sd->private;
5774 const struct cpumask *span = sched_domain_span(sd);
5775 struct cpumask *covered;
5776 int i;
5777
5778 get_group(cpu, sdd, &sd->groups);
5779 atomic_inc(&sd->groups->ref);
5780
5781 if (cpu != cpumask_first(span))
5782 return 0;
5783
5784 lockdep_assert_held(&sched_domains_mutex);
5785 covered = sched_domains_tmpmask;
5786
5787 cpumask_clear(covered);
5788
5789 for_each_cpu(i, span) {
5790 struct sched_group *sg;
5791 int group, j;
5792
5793 if (cpumask_test_cpu(i, covered))
5794 continue;
5795
5796 group = get_group(i, sdd, &sg);
5797 cpumask_clear(sched_group_cpus(sg));
5798 sg->sgp->power = 0;
5799 cpumask_setall(sched_group_mask(sg));
5800
5801 for_each_cpu(j, span) {
5802 if (get_group(j, sdd, NULL) != group)
5803 continue;
5804
5805 cpumask_set_cpu(j, covered);
5806 cpumask_set_cpu(j, sched_group_cpus(sg));
5807 }
5808
5809 if (!first)
5810 first = sg;
5811 if (last)
5812 last->next = sg;
5813 last = sg;
5814 }
5815 last->next = first;
5816
5817 return 0;
5818}
5819
5820/*
5821 * Initialize sched groups cpu_power.
5822 *
5823 * cpu_power indicates the capacity of sched group, which is used while
5824 * distributing the load between different sched groups in a sched domain.
5825 * Typically cpu_power for all the groups in a sched domain will be same unless
5826 * there are asymmetries in the topology. If there are asymmetries, group
5827 * having more cpu_power will pickup more load compared to the group having
5828 * less cpu_power.
5829 */
5830static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5831{
5832 struct sched_group *sg = sd->groups;
5833
5834 WARN_ON(!sg);
5835
5836 do {
5837 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5838 sg = sg->next;
5839 } while (sg != sd->groups);
5840
5841 if (cpu != group_balance_cpu(sg))
5842 return;
5843
5844 update_group_power(sd, cpu);
5845 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5846}
5847
5848int __weak arch_sd_sibling_asym_packing(void)
5849{
5850 return 0*SD_ASYM_PACKING;
5851}
5852
5853/*
5854 * Initializers for schedule domains
5855 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5856 */
5857
5858#ifdef CONFIG_SCHED_DEBUG
5859# define SD_INIT_NAME(sd, type) sd->name = #type
5860#else
5861# define SD_INIT_NAME(sd, type) do { } while (0)
5862#endif
5863
5864#define SD_INIT_FUNC(type) \
5865static noinline struct sched_domain * \
5866sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5867{ \
5868 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5869 *sd = SD_##type##_INIT; \
5870 SD_INIT_NAME(sd, type); \
5871 sd->private = &tl->data; \
5872 return sd; \
5873}
5874
5875SD_INIT_FUNC(CPU)
5876#ifdef CONFIG_SCHED_SMT
5877 SD_INIT_FUNC(SIBLING)
5878#endif
5879#ifdef CONFIG_SCHED_MC
5880 SD_INIT_FUNC(MC)
5881#endif
5882#ifdef CONFIG_SCHED_BOOK
5883 SD_INIT_FUNC(BOOK)
5884#endif
5885
5886static int default_relax_domain_level = -1;
5887int sched_domain_level_max;
5888
5889static int __init setup_relax_domain_level(char *str)
5890{
5891 if (kstrtoint(str, 0, &default_relax_domain_level))
5892 pr_warn("Unable to set relax_domain_level\n");
5893
5894 return 1;
5895}
5896__setup("relax_domain_level=", setup_relax_domain_level);
5897
5898static void set_domain_attribute(struct sched_domain *sd,
5899 struct sched_domain_attr *attr)
5900{
5901 int request;
5902
5903 if (!attr || attr->relax_domain_level < 0) {
5904 if (default_relax_domain_level < 0)
5905 return;
5906 else
5907 request = default_relax_domain_level;
5908 } else
5909 request = attr->relax_domain_level;
5910 if (request < sd->level) {
5911 /* turn off idle balance on this domain */
5912 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5913 } else {
5914 /* turn on idle balance on this domain */
5915 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5916 }
5917}
5918
5919static void __sdt_free(const struct cpumask *cpu_map);
5920static int __sdt_alloc(const struct cpumask *cpu_map);
5921
5922static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5923 const struct cpumask *cpu_map)
5924{
5925 switch (what) {
5926 case sa_rootdomain:
5927 if (!atomic_read(&d->rd->refcount))
5928 free_rootdomain(&d->rd->rcu); /* fall through */
5929 case sa_sd:
5930 free_percpu(d->sd); /* fall through */
5931 case sa_sd_storage:
5932 __sdt_free(cpu_map); /* fall through */
5933 case sa_none:
5934 break;
5935 }
5936}
5937
5938static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5939 const struct cpumask *cpu_map)
5940{
5941 memset(d, 0, sizeof(*d));
5942
5943 if (__sdt_alloc(cpu_map))
5944 return sa_sd_storage;
5945 d->sd = alloc_percpu(struct sched_domain *);
5946 if (!d->sd)
5947 return sa_sd_storage;
5948 d->rd = alloc_rootdomain();
5949 if (!d->rd)
5950 return sa_sd;
5951 return sa_rootdomain;
5952}
5953
5954/*
5955 * NULL the sd_data elements we've used to build the sched_domain and
5956 * sched_group structure so that the subsequent __free_domain_allocs()
5957 * will not free the data we're using.
5958 */
5959static void claim_allocations(int cpu, struct sched_domain *sd)
5960{
5961 struct sd_data *sdd = sd->private;
5962
5963 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5964 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5965
5966 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5967 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5968
5969 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5970 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5971}
5972
5973#ifdef CONFIG_SCHED_SMT
5974static const struct cpumask *cpu_smt_mask(int cpu)
5975{
5976 return topology_thread_cpumask(cpu);
5977}
5978#endif
5979
5980/*
5981 * Topology list, bottom-up.
5982 */
5983static struct sched_domain_topology_level default_topology[] = {
5984#ifdef CONFIG_SCHED_SMT
5985 { sd_init_SIBLING, cpu_smt_mask, },
5986#endif
5987#ifdef CONFIG_SCHED_MC
5988 { sd_init_MC, cpu_coregroup_mask, },
5989#endif
5990#ifdef CONFIG_SCHED_BOOK
5991 { sd_init_BOOK, cpu_book_mask, },
5992#endif
5993 { sd_init_CPU, cpu_cpu_mask, },
5994 { NULL, },
5995};
5996
5997static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5998
5999#define for_each_sd_topology(tl) \
6000 for (tl = sched_domain_topology; tl->init; tl++)
6001
6002#ifdef CONFIG_NUMA
6003
6004static int sched_domains_numa_levels;
6005static int *sched_domains_numa_distance;
6006static struct cpumask ***sched_domains_numa_masks;
6007static int sched_domains_curr_level;
6008
6009static inline int sd_local_flags(int level)
6010{
6011 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6012 return 0;
6013
6014 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6015}
6016
6017static struct sched_domain *
6018sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6019{
6020 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6021 int level = tl->numa_level;
6022 int sd_weight = cpumask_weight(
6023 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6024
6025 *sd = (struct sched_domain){
6026 .min_interval = sd_weight,
6027 .max_interval = 2*sd_weight,
6028 .busy_factor = 32,
6029 .imbalance_pct = 125,
6030 .cache_nice_tries = 2,
6031 .busy_idx = 3,
6032 .idle_idx = 2,
6033 .newidle_idx = 0,
6034 .wake_idx = 0,
6035 .forkexec_idx = 0,
6036
6037 .flags = 1*SD_LOAD_BALANCE
6038 | 1*SD_BALANCE_NEWIDLE
6039 | 0*SD_BALANCE_EXEC
6040 | 0*SD_BALANCE_FORK
6041 | 0*SD_BALANCE_WAKE
6042 | 0*SD_WAKE_AFFINE
6043 | 0*SD_SHARE_CPUPOWER
6044 | 0*SD_SHARE_PKG_RESOURCES
6045 | 1*SD_SERIALIZE
6046 | 0*SD_PREFER_SIBLING
6047 | 1*SD_NUMA
6048 | sd_local_flags(level)
6049 ,
6050 .last_balance = jiffies,
6051 .balance_interval = sd_weight,
6052 .max_newidle_lb_cost = 0,
6053 .next_decay_max_lb_cost = jiffies,
6054 };
6055 SD_INIT_NAME(sd, NUMA);
6056 sd->private = &tl->data;
6057
6058 /*
6059 * Ugly hack to pass state to sd_numa_mask()...
6060 */
6061 sched_domains_curr_level = tl->numa_level;
6062
6063 return sd;
6064}
6065
6066static const struct cpumask *sd_numa_mask(int cpu)
6067{
6068 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6069}
6070
6071static void sched_numa_warn(const char *str)
6072{
6073 static int done = false;
6074 int i,j;
6075
6076 if (done)
6077 return;
6078
6079 done = true;
6080
6081 printk(KERN_WARNING "ERROR: %s\n\n", str);
6082
6083 for (i = 0; i < nr_node_ids; i++) {
6084 printk(KERN_WARNING " ");
6085 for (j = 0; j < nr_node_ids; j++)
6086 printk(KERN_CONT "%02d ", node_distance(i,j));
6087 printk(KERN_CONT "\n");
6088 }
6089 printk(KERN_WARNING "\n");
6090}
6091
6092static bool find_numa_distance(int distance)
6093{
6094 int i;
6095
6096 if (distance == node_distance(0, 0))
6097 return true;
6098
6099 for (i = 0; i < sched_domains_numa_levels; i++) {
6100 if (sched_domains_numa_distance[i] == distance)
6101 return true;
6102 }
6103
6104 return false;
6105}
6106
6107static void sched_init_numa(void)
6108{
6109 int next_distance, curr_distance = node_distance(0, 0);
6110 struct sched_domain_topology_level *tl;
6111 int level = 0;
6112 int i, j, k;
6113
6114 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6115 if (!sched_domains_numa_distance)
6116 return;
6117
6118 /*
6119 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6120 * unique distances in the node_distance() table.
6121 *
6122 * Assumes node_distance(0,j) includes all distances in
6123 * node_distance(i,j) in order to avoid cubic time.
6124 */
6125 next_distance = curr_distance;
6126 for (i = 0; i < nr_node_ids; i++) {
6127 for (j = 0; j < nr_node_ids; j++) {
6128 for (k = 0; k < nr_node_ids; k++) {
6129 int distance = node_distance(i, k);
6130
6131 if (distance > curr_distance &&
6132 (distance < next_distance ||
6133 next_distance == curr_distance))
6134 next_distance = distance;
6135
6136 /*
6137 * While not a strong assumption it would be nice to know
6138 * about cases where if node A is connected to B, B is not
6139 * equally connected to A.
6140 */
6141 if (sched_debug() && node_distance(k, i) != distance)
6142 sched_numa_warn("Node-distance not symmetric");
6143
6144 if (sched_debug() && i && !find_numa_distance(distance))
6145 sched_numa_warn("Node-0 not representative");
6146 }
6147 if (next_distance != curr_distance) {
6148 sched_domains_numa_distance[level++] = next_distance;
6149 sched_domains_numa_levels = level;
6150 curr_distance = next_distance;
6151 } else break;
6152 }
6153
6154 /*
6155 * In case of sched_debug() we verify the above assumption.
6156 */
6157 if (!sched_debug())
6158 break;
6159 }
6160 /*
6161 * 'level' contains the number of unique distances, excluding the
6162 * identity distance node_distance(i,i).
6163 *
6164 * The sched_domains_numa_distance[] array includes the actual distance
6165 * numbers.
6166 */
6167
6168 /*
6169 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6170 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6171 * the array will contain less then 'level' members. This could be
6172 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6173 * in other functions.
6174 *
6175 * We reset it to 'level' at the end of this function.
6176 */
6177 sched_domains_numa_levels = 0;
6178
6179 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6180 if (!sched_domains_numa_masks)
6181 return;
6182
6183 /*
6184 * Now for each level, construct a mask per node which contains all
6185 * cpus of nodes that are that many hops away from us.
6186 */
6187 for (i = 0; i < level; i++) {
6188 sched_domains_numa_masks[i] =
6189 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6190 if (!sched_domains_numa_masks[i])
6191 return;
6192
6193 for (j = 0; j < nr_node_ids; j++) {
6194 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6195 if (!mask)
6196 return;
6197
6198 sched_domains_numa_masks[i][j] = mask;
6199
6200 for (k = 0; k < nr_node_ids; k++) {
6201 if (node_distance(j, k) > sched_domains_numa_distance[i])
6202 continue;
6203
6204 cpumask_or(mask, mask, cpumask_of_node(k));
6205 }
6206 }
6207 }
6208
6209 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6210 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6211 if (!tl)
6212 return;
6213
6214 /*
6215 * Copy the default topology bits..
6216 */
6217 for (i = 0; default_topology[i].init; i++)
6218 tl[i] = default_topology[i];
6219
6220 /*
6221 * .. and append 'j' levels of NUMA goodness.
6222 */
6223 for (j = 0; j < level; i++, j++) {
6224 tl[i] = (struct sched_domain_topology_level){
6225 .init = sd_numa_init,
6226 .mask = sd_numa_mask,
6227 .flags = SDTL_OVERLAP,
6228 .numa_level = j,
6229 };
6230 }
6231
6232 sched_domain_topology = tl;
6233
6234 sched_domains_numa_levels = level;
6235}
6236
6237static void sched_domains_numa_masks_set(int cpu)
6238{
6239 int i, j;
6240 int node = cpu_to_node(cpu);
6241
6242 for (i = 0; i < sched_domains_numa_levels; i++) {
6243 for (j = 0; j < nr_node_ids; j++) {
6244 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6245 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6246 }
6247 }
6248}
6249
6250static void sched_domains_numa_masks_clear(int cpu)
6251{
6252 int i, j;
6253 for (i = 0; i < sched_domains_numa_levels; i++) {
6254 for (j = 0; j < nr_node_ids; j++)
6255 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6256 }
6257}
6258
6259/*
6260 * Update sched_domains_numa_masks[level][node] array when new cpus
6261 * are onlined.
6262 */
6263static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6264 unsigned long action,
6265 void *hcpu)
6266{
6267 int cpu = (long)hcpu;
6268
6269 switch (action & ~CPU_TASKS_FROZEN) {
6270 case CPU_ONLINE:
6271 sched_domains_numa_masks_set(cpu);
6272 break;
6273
6274 case CPU_DEAD:
6275 sched_domains_numa_masks_clear(cpu);
6276 break;
6277
6278 default:
6279 return NOTIFY_DONE;
6280 }
6281
6282 return NOTIFY_OK;
6283}
6284#else
6285static inline void sched_init_numa(void)
6286{
6287}
6288
6289static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6290 unsigned long action,
6291 void *hcpu)
6292{
6293 return 0;
6294}
6295#endif /* CONFIG_NUMA */
6296
6297static int __sdt_alloc(const struct cpumask *cpu_map)
6298{
6299 struct sched_domain_topology_level *tl;
6300 int j;
6301
6302 for_each_sd_topology(tl) {
6303 struct sd_data *sdd = &tl->data;
6304
6305 sdd->sd = alloc_percpu(struct sched_domain *);
6306 if (!sdd->sd)
6307 return -ENOMEM;
6308
6309 sdd->sg = alloc_percpu(struct sched_group *);
6310 if (!sdd->sg)
6311 return -ENOMEM;
6312
6313 sdd->sgp = alloc_percpu(struct sched_group_power *);
6314 if (!sdd->sgp)
6315 return -ENOMEM;
6316
6317 for_each_cpu(j, cpu_map) {
6318 struct sched_domain *sd;
6319 struct sched_group *sg;
6320 struct sched_group_power *sgp;
6321
6322 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6323 GFP_KERNEL, cpu_to_node(j));
6324 if (!sd)
6325 return -ENOMEM;
6326
6327 *per_cpu_ptr(sdd->sd, j) = sd;
6328
6329 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6330 GFP_KERNEL, cpu_to_node(j));
6331 if (!sg)
6332 return -ENOMEM;
6333
6334 sg->next = sg;
6335
6336 *per_cpu_ptr(sdd->sg, j) = sg;
6337
6338 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6339 GFP_KERNEL, cpu_to_node(j));
6340 if (!sgp)
6341 return -ENOMEM;
6342
6343 *per_cpu_ptr(sdd->sgp, j) = sgp;
6344 }
6345 }
6346
6347 return 0;
6348}
6349
6350static void __sdt_free(const struct cpumask *cpu_map)
6351{
6352 struct sched_domain_topology_level *tl;
6353 int j;
6354
6355 for_each_sd_topology(tl) {
6356 struct sd_data *sdd = &tl->data;
6357
6358 for_each_cpu(j, cpu_map) {
6359 struct sched_domain *sd;
6360
6361 if (sdd->sd) {
6362 sd = *per_cpu_ptr(sdd->sd, j);
6363 if (sd && (sd->flags & SD_OVERLAP))
6364 free_sched_groups(sd->groups, 0);
6365 kfree(*per_cpu_ptr(sdd->sd, j));
6366 }
6367
6368 if (sdd->sg)
6369 kfree(*per_cpu_ptr(sdd->sg, j));
6370 if (sdd->sgp)
6371 kfree(*per_cpu_ptr(sdd->sgp, j));
6372 }
6373 free_percpu(sdd->sd);
6374 sdd->sd = NULL;
6375 free_percpu(sdd->sg);
6376 sdd->sg = NULL;
6377 free_percpu(sdd->sgp);
6378 sdd->sgp = NULL;
6379 }
6380}
6381
6382struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6383 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6384 struct sched_domain *child, int cpu)
6385{
6386 struct sched_domain *sd = tl->init(tl, cpu);
6387 if (!sd)
6388 return child;
6389
6390 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6391 if (child) {
6392 sd->level = child->level + 1;
6393 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6394 child->parent = sd;
6395 sd->child = child;
6396 }
6397 set_domain_attribute(sd, attr);
6398
6399 return sd;
6400}
6401
6402/*
6403 * Build sched domains for a given set of cpus and attach the sched domains
6404 * to the individual cpus
6405 */
6406static int build_sched_domains(const struct cpumask *cpu_map,
6407 struct sched_domain_attr *attr)
6408{
6409 enum s_alloc alloc_state;
6410 struct sched_domain *sd;
6411 struct s_data d;
6412 int i, ret = -ENOMEM;
6413
6414 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6415 if (alloc_state != sa_rootdomain)
6416 goto error;
6417
6418 /* Set up domains for cpus specified by the cpu_map. */
6419 for_each_cpu(i, cpu_map) {
6420 struct sched_domain_topology_level *tl;
6421
6422 sd = NULL;
6423 for_each_sd_topology(tl) {
6424 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6425 if (tl == sched_domain_topology)
6426 *per_cpu_ptr(d.sd, i) = sd;
6427 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6428 sd->flags |= SD_OVERLAP;
6429 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6430 break;
6431 }
6432 }
6433
6434 /* Build the groups for the domains */
6435 for_each_cpu(i, cpu_map) {
6436 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6437 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6438 if (sd->flags & SD_OVERLAP) {
6439 if (build_overlap_sched_groups(sd, i))
6440 goto error;
6441 } else {
6442 if (build_sched_groups(sd, i))
6443 goto error;
6444 }
6445 }
6446 }
6447
6448 /* Calculate CPU power for physical packages and nodes */
6449 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6450 if (!cpumask_test_cpu(i, cpu_map))
6451 continue;
6452
6453 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6454 claim_allocations(i, sd);
6455 init_sched_groups_power(i, sd);
6456 }
6457 }
6458
6459 /* Attach the domains */
6460 rcu_read_lock();
6461 for_each_cpu(i, cpu_map) {
6462 sd = *per_cpu_ptr(d.sd, i);
6463 cpu_attach_domain(sd, d.rd, i);
6464 }
6465 rcu_read_unlock();
6466
6467 ret = 0;
6468error:
6469 __free_domain_allocs(&d, alloc_state, cpu_map);
6470 return ret;
6471}
6472
6473static cpumask_var_t *doms_cur; /* current sched domains */
6474static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6475static struct sched_domain_attr *dattr_cur;
6476 /* attribues of custom domains in 'doms_cur' */
6477
6478/*
6479 * Special case: If a kmalloc of a doms_cur partition (array of
6480 * cpumask) fails, then fallback to a single sched domain,
6481 * as determined by the single cpumask fallback_doms.
6482 */
6483static cpumask_var_t fallback_doms;
6484
6485/*
6486 * arch_update_cpu_topology lets virtualized architectures update the
6487 * cpu core maps. It is supposed to return 1 if the topology changed
6488 * or 0 if it stayed the same.
6489 */
6490int __weak arch_update_cpu_topology(void)
6491{
6492 return 0;
6493}
6494
6495cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6496{
6497 int i;
6498 cpumask_var_t *doms;
6499
6500 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6501 if (!doms)
6502 return NULL;
6503 for (i = 0; i < ndoms; i++) {
6504 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6505 free_sched_domains(doms, i);
6506 return NULL;
6507 }
6508 }
6509 return doms;
6510}
6511
6512void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6513{
6514 unsigned int i;
6515 for (i = 0; i < ndoms; i++)
6516 free_cpumask_var(doms[i]);
6517 kfree(doms);
6518}
6519
6520/*
6521 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6522 * For now this just excludes isolated cpus, but could be used to
6523 * exclude other special cases in the future.
6524 */
6525static int init_sched_domains(const struct cpumask *cpu_map)
6526{
6527 int err;
6528
6529 arch_update_cpu_topology();
6530 ndoms_cur = 1;
6531 doms_cur = alloc_sched_domains(ndoms_cur);
6532 if (!doms_cur)
6533 doms_cur = &fallback_doms;
6534 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6535 err = build_sched_domains(doms_cur[0], NULL);
6536 register_sched_domain_sysctl();
6537
6538 return err;
6539}
6540
6541/*
6542 * Detach sched domains from a group of cpus specified in cpu_map
6543 * These cpus will now be attached to the NULL domain
6544 */
6545static void detach_destroy_domains(const struct cpumask *cpu_map)
6546{
6547 int i;
6548
6549 rcu_read_lock();
6550 for_each_cpu(i, cpu_map)
6551 cpu_attach_domain(NULL, &def_root_domain, i);
6552 rcu_read_unlock();
6553}
6554
6555/* handle null as "default" */
6556static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6557 struct sched_domain_attr *new, int idx_new)
6558{
6559 struct sched_domain_attr tmp;
6560
6561 /* fast path */
6562 if (!new && !cur)
6563 return 1;
6564
6565 tmp = SD_ATTR_INIT;
6566 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6567 new ? (new + idx_new) : &tmp,
6568 sizeof(struct sched_domain_attr));
6569}
6570
6571/*
6572 * Partition sched domains as specified by the 'ndoms_new'
6573 * cpumasks in the array doms_new[] of cpumasks. This compares
6574 * doms_new[] to the current sched domain partitioning, doms_cur[].
6575 * It destroys each deleted domain and builds each new domain.
6576 *
6577 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6578 * The masks don't intersect (don't overlap.) We should setup one
6579 * sched domain for each mask. CPUs not in any of the cpumasks will
6580 * not be load balanced. If the same cpumask appears both in the
6581 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6582 * it as it is.
6583 *
6584 * The passed in 'doms_new' should be allocated using
6585 * alloc_sched_domains. This routine takes ownership of it and will
6586 * free_sched_domains it when done with it. If the caller failed the
6587 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6588 * and partition_sched_domains() will fallback to the single partition
6589 * 'fallback_doms', it also forces the domains to be rebuilt.
6590 *
6591 * If doms_new == NULL it will be replaced with cpu_online_mask.
6592 * ndoms_new == 0 is a special case for destroying existing domains,
6593 * and it will not create the default domain.
6594 *
6595 * Call with hotplug lock held
6596 */
6597void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6598 struct sched_domain_attr *dattr_new)
6599{
6600 int i, j, n;
6601 int new_topology;
6602
6603 mutex_lock(&sched_domains_mutex);
6604
6605 /* always unregister in case we don't destroy any domains */
6606 unregister_sched_domain_sysctl();
6607
6608 /* Let architecture update cpu core mappings. */
6609 new_topology = arch_update_cpu_topology();
6610
6611 n = doms_new ? ndoms_new : 0;
6612
6613 /* Destroy deleted domains */
6614 for (i = 0; i < ndoms_cur; i++) {
6615 for (j = 0; j < n && !new_topology; j++) {
6616 if (cpumask_equal(doms_cur[i], doms_new[j])
6617 && dattrs_equal(dattr_cur, i, dattr_new, j))
6618 goto match1;
6619 }
6620 /* no match - a current sched domain not in new doms_new[] */
6621 detach_destroy_domains(doms_cur[i]);
6622match1:
6623 ;
6624 }
6625
6626 n = ndoms_cur;
6627 if (doms_new == NULL) {
6628 n = 0;
6629 doms_new = &fallback_doms;
6630 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6631 WARN_ON_ONCE(dattr_new);
6632 }
6633
6634 /* Build new domains */
6635 for (i = 0; i < ndoms_new; i++) {
6636 for (j = 0; j < n && !new_topology; j++) {
6637 if (cpumask_equal(doms_new[i], doms_cur[j])
6638 && dattrs_equal(dattr_new, i, dattr_cur, j))
6639 goto match2;
6640 }
6641 /* no match - add a new doms_new */
6642 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6643match2:
6644 ;
6645 }
6646
6647 /* Remember the new sched domains */
6648 if (doms_cur != &fallback_doms)
6649 free_sched_domains(doms_cur, ndoms_cur);
6650 kfree(dattr_cur); /* kfree(NULL) is safe */
6651 doms_cur = doms_new;
6652 dattr_cur = dattr_new;
6653 ndoms_cur = ndoms_new;
6654
6655 register_sched_domain_sysctl();
6656
6657 mutex_unlock(&sched_domains_mutex);
6658}
6659
6660static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6661
6662/*
6663 * Update cpusets according to cpu_active mask. If cpusets are
6664 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6665 * around partition_sched_domains().
6666 *
6667 * If we come here as part of a suspend/resume, don't touch cpusets because we
6668 * want to restore it back to its original state upon resume anyway.
6669 */
6670static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6671 void *hcpu)
6672{
6673 switch (action) {
6674 case CPU_ONLINE_FROZEN:
6675 case CPU_DOWN_FAILED_FROZEN:
6676
6677 /*
6678 * num_cpus_frozen tracks how many CPUs are involved in suspend
6679 * resume sequence. As long as this is not the last online
6680 * operation in the resume sequence, just build a single sched
6681 * domain, ignoring cpusets.
6682 */
6683 num_cpus_frozen--;
6684 if (likely(num_cpus_frozen)) {
6685 partition_sched_domains(1, NULL, NULL);
6686 break;
6687 }
6688
6689 /*
6690 * This is the last CPU online operation. So fall through and
6691 * restore the original sched domains by considering the
6692 * cpuset configurations.
6693 */
6694
6695 case CPU_ONLINE:
6696 case CPU_DOWN_FAILED:
6697 cpuset_update_active_cpus(true);
6698 break;
6699 default:
6700 return NOTIFY_DONE;
6701 }
6702 return NOTIFY_OK;
6703}
6704
6705static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6706 void *hcpu)
6707{
6708 switch (action) {
6709 case CPU_DOWN_PREPARE:
6710 cpuset_update_active_cpus(false);
6711 break;
6712 case CPU_DOWN_PREPARE_FROZEN:
6713 num_cpus_frozen++;
6714 partition_sched_domains(1, NULL, NULL);
6715 break;
6716 default:
6717 return NOTIFY_DONE;
6718 }
6719 return NOTIFY_OK;
6720}
6721
6722void __init sched_init_smp(void)
6723{
6724 cpumask_var_t non_isolated_cpus;
6725
6726 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6727 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6728
6729 sched_init_numa();
6730
6731 /*
6732 * There's no userspace yet to cause hotplug operations; hence all the
6733 * cpu masks are stable and all blatant races in the below code cannot
6734 * happen.
6735 */
6736 mutex_lock(&sched_domains_mutex);
6737 init_sched_domains(cpu_active_mask);
6738 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6739 if (cpumask_empty(non_isolated_cpus))
6740 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6741 mutex_unlock(&sched_domains_mutex);
6742
6743 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6744 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6745 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6746
6747 init_hrtick();
6748
6749 /* Move init over to a non-isolated CPU */
6750 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6751 BUG();
6752 sched_init_granularity();
6753 free_cpumask_var(non_isolated_cpus);
6754
6755 init_sched_rt_class();
6756 init_sched_dl_class();
6757}
6758#else
6759void __init sched_init_smp(void)
6760{
6761 sched_init_granularity();
6762}
6763#endif /* CONFIG_SMP */
6764
6765const_debug unsigned int sysctl_timer_migration = 1;
6766
6767int in_sched_functions(unsigned long addr)
6768{
6769 return in_lock_functions(addr) ||
6770 (addr >= (unsigned long)__sched_text_start
6771 && addr < (unsigned long)__sched_text_end);
6772}
6773
6774#ifdef CONFIG_CGROUP_SCHED
6775/*
6776 * Default task group.
6777 * Every task in system belongs to this group at bootup.
6778 */
6779struct task_group root_task_group;
6780LIST_HEAD(task_groups);
6781#endif
6782
6783DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6784
6785void __init sched_init(void)
6786{
6787 int i, j;
6788 unsigned long alloc_size = 0, ptr;
6789
6790#ifdef CONFIG_FAIR_GROUP_SCHED
6791 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6792#endif
6793#ifdef CONFIG_RT_GROUP_SCHED
6794 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6795#endif
6796#ifdef CONFIG_CPUMASK_OFFSTACK
6797 alloc_size += num_possible_cpus() * cpumask_size();
6798#endif
6799 if (alloc_size) {
6800 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6801
6802#ifdef CONFIG_FAIR_GROUP_SCHED
6803 root_task_group.se = (struct sched_entity **)ptr;
6804 ptr += nr_cpu_ids * sizeof(void **);
6805
6806 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6807 ptr += nr_cpu_ids * sizeof(void **);
6808
6809#endif /* CONFIG_FAIR_GROUP_SCHED */
6810#ifdef CONFIG_RT_GROUP_SCHED
6811 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6812 ptr += nr_cpu_ids * sizeof(void **);
6813
6814 root_task_group.rt_rq = (struct rt_rq **)ptr;
6815 ptr += nr_cpu_ids * sizeof(void **);
6816
6817#endif /* CONFIG_RT_GROUP_SCHED */
6818#ifdef CONFIG_CPUMASK_OFFSTACK
6819 for_each_possible_cpu(i) {
6820 per_cpu(load_balance_mask, i) = (void *)ptr;
6821 ptr += cpumask_size();
6822 }
6823#endif /* CONFIG_CPUMASK_OFFSTACK */
6824 }
6825
6826 init_rt_bandwidth(&def_rt_bandwidth,
6827 global_rt_period(), global_rt_runtime());
6828 init_dl_bandwidth(&def_dl_bandwidth,
6829 global_rt_period(), global_rt_runtime());
6830
6831#ifdef CONFIG_SMP
6832 init_defrootdomain();
6833#endif
6834
6835#ifdef CONFIG_RT_GROUP_SCHED
6836 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6837 global_rt_period(), global_rt_runtime());
6838#endif /* CONFIG_RT_GROUP_SCHED */
6839
6840#ifdef CONFIG_CGROUP_SCHED
6841 list_add(&root_task_group.list, &task_groups);
6842 INIT_LIST_HEAD(&root_task_group.children);
6843 INIT_LIST_HEAD(&root_task_group.siblings);
6844 autogroup_init(&init_task);
6845
6846#endif /* CONFIG_CGROUP_SCHED */
6847
6848 for_each_possible_cpu(i) {
6849 struct rq *rq;
6850
6851 rq = cpu_rq(i);
6852 raw_spin_lock_init(&rq->lock);
6853 rq->nr_running = 0;
6854 rq->calc_load_active = 0;
6855 rq->calc_load_update = jiffies + LOAD_FREQ;
6856 init_cfs_rq(&rq->cfs);
6857 init_rt_rq(&rq->rt, rq);
6858 init_dl_rq(&rq->dl, rq);
6859#ifdef CONFIG_FAIR_GROUP_SCHED
6860 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6861 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6862 /*
6863 * How much cpu bandwidth does root_task_group get?
6864 *
6865 * In case of task-groups formed thr' the cgroup filesystem, it
6866 * gets 100% of the cpu resources in the system. This overall
6867 * system cpu resource is divided among the tasks of
6868 * root_task_group and its child task-groups in a fair manner,
6869 * based on each entity's (task or task-group's) weight
6870 * (se->load.weight).
6871 *
6872 * In other words, if root_task_group has 10 tasks of weight
6873 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6874 * then A0's share of the cpu resource is:
6875 *
6876 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6877 *
6878 * We achieve this by letting root_task_group's tasks sit
6879 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6880 */
6881 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6882 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6883#endif /* CONFIG_FAIR_GROUP_SCHED */
6884
6885 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6886#ifdef CONFIG_RT_GROUP_SCHED
6887 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6888#endif
6889
6890 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6891 rq->cpu_load[j] = 0;
6892
6893 rq->last_load_update_tick = jiffies;
6894
6895#ifdef CONFIG_SMP
6896 rq->sd = NULL;
6897 rq->rd = NULL;
6898 rq->cpu_power = SCHED_POWER_SCALE;
6899 rq->post_schedule = 0;
6900 rq->active_balance = 0;
6901 rq->next_balance = jiffies;
6902 rq->push_cpu = 0;
6903 rq->cpu = i;
6904 rq->online = 0;
6905 rq->idle_stamp = 0;
6906 rq->avg_idle = 2*sysctl_sched_migration_cost;
6907 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6908
6909 INIT_LIST_HEAD(&rq->cfs_tasks);
6910
6911 rq_attach_root(rq, &def_root_domain);
6912#ifdef CONFIG_NO_HZ_COMMON
6913 rq->nohz_flags = 0;
6914#endif
6915#ifdef CONFIG_NO_HZ_FULL
6916 rq->last_sched_tick = 0;
6917#endif
6918#endif
6919 init_rq_hrtick(rq);
6920 atomic_set(&rq->nr_iowait, 0);
6921 }
6922
6923 set_load_weight(&init_task);
6924
6925#ifdef CONFIG_PREEMPT_NOTIFIERS
6926 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6927#endif
6928
6929 /*
6930 * The boot idle thread does lazy MMU switching as well:
6931 */
6932 atomic_inc(&init_mm.mm_count);
6933 enter_lazy_tlb(&init_mm, current);
6934
6935 /*
6936 * Make us the idle thread. Technically, schedule() should not be
6937 * called from this thread, however somewhere below it might be,
6938 * but because we are the idle thread, we just pick up running again
6939 * when this runqueue becomes "idle".
6940 */
6941 init_idle(current, smp_processor_id());
6942
6943 calc_load_update = jiffies + LOAD_FREQ;
6944
6945 /*
6946 * During early bootup we pretend to be a normal task:
6947 */
6948 current->sched_class = &fair_sched_class;
6949
6950#ifdef CONFIG_SMP
6951 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6952 /* May be allocated at isolcpus cmdline parse time */
6953 if (cpu_isolated_map == NULL)
6954 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6955 idle_thread_set_boot_cpu();
6956#endif
6957 init_sched_fair_class();
6958
6959 scheduler_running = 1;
6960}
6961
6962#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6963static inline int preempt_count_equals(int preempt_offset)
6964{
6965 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6966
6967 return (nested == preempt_offset);
6968}
6969
6970void __might_sleep(const char *file, int line, int preempt_offset)
6971{
6972 static unsigned long prev_jiffy; /* ratelimiting */
6973
6974 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6975 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6976 !is_idle_task(current)) ||
6977 system_state != SYSTEM_RUNNING || oops_in_progress)
6978 return;
6979 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6980 return;
6981 prev_jiffy = jiffies;
6982
6983 printk(KERN_ERR
6984 "BUG: sleeping function called from invalid context at %s:%d\n",
6985 file, line);
6986 printk(KERN_ERR
6987 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6988 in_atomic(), irqs_disabled(),
6989 current->pid, current->comm);
6990
6991 debug_show_held_locks(current);
6992 if (irqs_disabled())
6993 print_irqtrace_events(current);
6994#ifdef CONFIG_DEBUG_PREEMPT
6995 if (!preempt_count_equals(preempt_offset)) {
6996 pr_err("Preemption disabled at:");
6997 print_ip_sym(current->preempt_disable_ip);
6998 pr_cont("\n");
6999 }
7000#endif
7001 dump_stack();
7002}
7003EXPORT_SYMBOL(__might_sleep);
7004#endif
7005
7006#ifdef CONFIG_MAGIC_SYSRQ
7007static void normalize_task(struct rq *rq, struct task_struct *p)
7008{
7009 const struct sched_class *prev_class = p->sched_class;
7010 struct sched_attr attr = {
7011 .sched_policy = SCHED_NORMAL,
7012 };
7013 int old_prio = p->prio;
7014 int on_rq;
7015
7016 on_rq = p->on_rq;
7017 if (on_rq)
7018 dequeue_task(rq, p, 0);
7019 __setscheduler(rq, p, &attr);
7020 if (on_rq) {
7021 enqueue_task(rq, p, 0);
7022 resched_task(rq->curr);
7023 }
7024
7025 check_class_changed(rq, p, prev_class, old_prio);
7026}
7027
7028void normalize_rt_tasks(void)
7029{
7030 struct task_struct *g, *p;
7031 unsigned long flags;
7032 struct rq *rq;
7033
7034 read_lock_irqsave(&tasklist_lock, flags);
7035 do_each_thread(g, p) {
7036 /*
7037 * Only normalize user tasks:
7038 */
7039 if (!p->mm)
7040 continue;
7041
7042 p->se.exec_start = 0;
7043#ifdef CONFIG_SCHEDSTATS
7044 p->se.statistics.wait_start = 0;
7045 p->se.statistics.sleep_start = 0;
7046 p->se.statistics.block_start = 0;
7047#endif
7048
7049 if (!dl_task(p) && !rt_task(p)) {
7050 /*
7051 * Renice negative nice level userspace
7052 * tasks back to 0:
7053 */
7054 if (task_nice(p) < 0 && p->mm)
7055 set_user_nice(p, 0);
7056 continue;
7057 }
7058
7059 raw_spin_lock(&p->pi_lock);
7060 rq = __task_rq_lock(p);
7061
7062 normalize_task(rq, p);
7063
7064 __task_rq_unlock(rq);
7065 raw_spin_unlock(&p->pi_lock);
7066 } while_each_thread(g, p);
7067
7068 read_unlock_irqrestore(&tasklist_lock, flags);
7069}
7070
7071#endif /* CONFIG_MAGIC_SYSRQ */
7072
7073#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7074/*
7075 * These functions are only useful for the IA64 MCA handling, or kdb.
7076 *
7077 * They can only be called when the whole system has been
7078 * stopped - every CPU needs to be quiescent, and no scheduling
7079 * activity can take place. Using them for anything else would
7080 * be a serious bug, and as a result, they aren't even visible
7081 * under any other configuration.
7082 */
7083
7084/**
7085 * curr_task - return the current task for a given cpu.
7086 * @cpu: the processor in question.
7087 *
7088 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7089 *
7090 * Return: The current task for @cpu.
7091 */
7092struct task_struct *curr_task(int cpu)
7093{
7094 return cpu_curr(cpu);
7095}
7096
7097#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7098
7099#ifdef CONFIG_IA64
7100/**
7101 * set_curr_task - set the current task for a given cpu.
7102 * @cpu: the processor in question.
7103 * @p: the task pointer to set.
7104 *
7105 * Description: This function must only be used when non-maskable interrupts
7106 * are serviced on a separate stack. It allows the architecture to switch the
7107 * notion of the current task on a cpu in a non-blocking manner. This function
7108 * must be called with all CPU's synchronized, and interrupts disabled, the
7109 * and caller must save the original value of the current task (see
7110 * curr_task() above) and restore that value before reenabling interrupts and
7111 * re-starting the system.
7112 *
7113 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7114 */
7115void set_curr_task(int cpu, struct task_struct *p)
7116{
7117 cpu_curr(cpu) = p;
7118}
7119
7120#endif
7121
7122#ifdef CONFIG_CGROUP_SCHED
7123/* task_group_lock serializes the addition/removal of task groups */
7124static DEFINE_SPINLOCK(task_group_lock);
7125
7126static void free_sched_group(struct task_group *tg)
7127{
7128 free_fair_sched_group(tg);
7129 free_rt_sched_group(tg);
7130 autogroup_free(tg);
7131 kfree(tg);
7132}
7133
7134/* allocate runqueue etc for a new task group */
7135struct task_group *sched_create_group(struct task_group *parent)
7136{
7137 struct task_group *tg;
7138
7139 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7140 if (!tg)
7141 return ERR_PTR(-ENOMEM);
7142
7143 if (!alloc_fair_sched_group(tg, parent))
7144 goto err;
7145
7146 if (!alloc_rt_sched_group(tg, parent))
7147 goto err;
7148
7149 return tg;
7150
7151err:
7152 free_sched_group(tg);
7153 return ERR_PTR(-ENOMEM);
7154}
7155
7156void sched_online_group(struct task_group *tg, struct task_group *parent)
7157{
7158 unsigned long flags;
7159
7160 spin_lock_irqsave(&task_group_lock, flags);
7161 list_add_rcu(&tg->list, &task_groups);
7162
7163 WARN_ON(!parent); /* root should already exist */
7164
7165 tg->parent = parent;
7166 INIT_LIST_HEAD(&tg->children);
7167 list_add_rcu(&tg->siblings, &parent->children);
7168 spin_unlock_irqrestore(&task_group_lock, flags);
7169}
7170
7171/* rcu callback to free various structures associated with a task group */
7172static void free_sched_group_rcu(struct rcu_head *rhp)
7173{
7174 /* now it should be safe to free those cfs_rqs */
7175 free_sched_group(container_of(rhp, struct task_group, rcu));
7176}
7177
7178/* Destroy runqueue etc associated with a task group */
7179void sched_destroy_group(struct task_group *tg)
7180{
7181 /* wait for possible concurrent references to cfs_rqs complete */
7182 call_rcu(&tg->rcu, free_sched_group_rcu);
7183}
7184
7185void sched_offline_group(struct task_group *tg)
7186{
7187 unsigned long flags;
7188 int i;
7189
7190 /* end participation in shares distribution */
7191 for_each_possible_cpu(i)
7192 unregister_fair_sched_group(tg, i);
7193
7194 spin_lock_irqsave(&task_group_lock, flags);
7195 list_del_rcu(&tg->list);
7196 list_del_rcu(&tg->siblings);
7197 spin_unlock_irqrestore(&task_group_lock, flags);
7198}
7199
7200/* change task's runqueue when it moves between groups.
7201 * The caller of this function should have put the task in its new group
7202 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7203 * reflect its new group.
7204 */
7205void sched_move_task(struct task_struct *tsk)
7206{
7207 struct task_group *tg;
7208 int on_rq, running;
7209 unsigned long flags;
7210 struct rq *rq;
7211
7212 rq = task_rq_lock(tsk, &flags);
7213
7214 running = task_current(rq, tsk);
7215 on_rq = tsk->on_rq;
7216
7217 if (on_rq)
7218 dequeue_task(rq, tsk, 0);
7219 if (unlikely(running))
7220 tsk->sched_class->put_prev_task(rq, tsk);
7221
7222 tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7223 lockdep_is_held(&tsk->sighand->siglock)),
7224 struct task_group, css);
7225 tg = autogroup_task_group(tsk, tg);
7226 tsk->sched_task_group = tg;
7227
7228#ifdef CONFIG_FAIR_GROUP_SCHED
7229 if (tsk->sched_class->task_move_group)
7230 tsk->sched_class->task_move_group(tsk, on_rq);
7231 else
7232#endif
7233 set_task_rq(tsk, task_cpu(tsk));
7234
7235 if (unlikely(running))
7236 tsk->sched_class->set_curr_task(rq);
7237 if (on_rq)
7238 enqueue_task(rq, tsk, 0);
7239
7240 task_rq_unlock(rq, tsk, &flags);
7241}
7242#endif /* CONFIG_CGROUP_SCHED */
7243
7244#ifdef CONFIG_RT_GROUP_SCHED
7245/*
7246 * Ensure that the real time constraints are schedulable.
7247 */
7248static DEFINE_MUTEX(rt_constraints_mutex);
7249
7250/* Must be called with tasklist_lock held */
7251static inline int tg_has_rt_tasks(struct task_group *tg)
7252{
7253 struct task_struct *g, *p;
7254
7255 do_each_thread(g, p) {
7256 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7257 return 1;
7258 } while_each_thread(g, p);
7259
7260 return 0;
7261}
7262
7263struct rt_schedulable_data {
7264 struct task_group *tg;
7265 u64 rt_period;
7266 u64 rt_runtime;
7267};
7268
7269static int tg_rt_schedulable(struct task_group *tg, void *data)
7270{
7271 struct rt_schedulable_data *d = data;
7272 struct task_group *child;
7273 unsigned long total, sum = 0;
7274 u64 period, runtime;
7275
7276 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7277 runtime = tg->rt_bandwidth.rt_runtime;
7278
7279 if (tg == d->tg) {
7280 period = d->rt_period;
7281 runtime = d->rt_runtime;
7282 }
7283
7284 /*
7285 * Cannot have more runtime than the period.
7286 */
7287 if (runtime > period && runtime != RUNTIME_INF)
7288 return -EINVAL;
7289
7290 /*
7291 * Ensure we don't starve existing RT tasks.
7292 */
7293 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7294 return -EBUSY;
7295
7296 total = to_ratio(period, runtime);
7297
7298 /*
7299 * Nobody can have more than the global setting allows.
7300 */
7301 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7302 return -EINVAL;
7303
7304 /*
7305 * The sum of our children's runtime should not exceed our own.
7306 */
7307 list_for_each_entry_rcu(child, &tg->children, siblings) {
7308 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7309 runtime = child->rt_bandwidth.rt_runtime;
7310
7311 if (child == d->tg) {
7312 period = d->rt_period;
7313 runtime = d->rt_runtime;
7314 }
7315
7316 sum += to_ratio(period, runtime);
7317 }
7318
7319 if (sum > total)
7320 return -EINVAL;
7321
7322 return 0;
7323}
7324
7325static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7326{
7327 int ret;
7328
7329 struct rt_schedulable_data data = {
7330 .tg = tg,
7331 .rt_period = period,
7332 .rt_runtime = runtime,
7333 };
7334
7335 rcu_read_lock();
7336 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7337 rcu_read_unlock();
7338
7339 return ret;
7340}
7341
7342static int tg_set_rt_bandwidth(struct task_group *tg,
7343 u64 rt_period, u64 rt_runtime)
7344{
7345 int i, err = 0;
7346
7347 mutex_lock(&rt_constraints_mutex);
7348 read_lock(&tasklist_lock);
7349 err = __rt_schedulable(tg, rt_period, rt_runtime);
7350 if (err)
7351 goto unlock;
7352
7353 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7354 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7355 tg->rt_bandwidth.rt_runtime = rt_runtime;
7356
7357 for_each_possible_cpu(i) {
7358 struct rt_rq *rt_rq = tg->rt_rq[i];
7359
7360 raw_spin_lock(&rt_rq->rt_runtime_lock);
7361 rt_rq->rt_runtime = rt_runtime;
7362 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7363 }
7364 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7365unlock:
7366 read_unlock(&tasklist_lock);
7367 mutex_unlock(&rt_constraints_mutex);
7368
7369 return err;
7370}
7371
7372static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7373{
7374 u64 rt_runtime, rt_period;
7375
7376 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7377 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7378 if (rt_runtime_us < 0)
7379 rt_runtime = RUNTIME_INF;
7380
7381 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7382}
7383
7384static long sched_group_rt_runtime(struct task_group *tg)
7385{
7386 u64 rt_runtime_us;
7387
7388 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7389 return -1;
7390
7391 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7392 do_div(rt_runtime_us, NSEC_PER_USEC);
7393 return rt_runtime_us;
7394}
7395
7396static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7397{
7398 u64 rt_runtime, rt_period;
7399
7400 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7401 rt_runtime = tg->rt_bandwidth.rt_runtime;
7402
7403 if (rt_period == 0)
7404 return -EINVAL;
7405
7406 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7407}
7408
7409static long sched_group_rt_period(struct task_group *tg)
7410{
7411 u64 rt_period_us;
7412
7413 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7414 do_div(rt_period_us, NSEC_PER_USEC);
7415 return rt_period_us;
7416}
7417#endif /* CONFIG_RT_GROUP_SCHED */
7418
7419#ifdef CONFIG_RT_GROUP_SCHED
7420static int sched_rt_global_constraints(void)
7421{
7422 int ret = 0;
7423
7424 mutex_lock(&rt_constraints_mutex);
7425 read_lock(&tasklist_lock);
7426 ret = __rt_schedulable(NULL, 0, 0);
7427 read_unlock(&tasklist_lock);
7428 mutex_unlock(&rt_constraints_mutex);
7429
7430 return ret;
7431}
7432
7433static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7434{
7435 /* Don't accept realtime tasks when there is no way for them to run */
7436 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7437 return 0;
7438
7439 return 1;
7440}
7441
7442#else /* !CONFIG_RT_GROUP_SCHED */
7443static int sched_rt_global_constraints(void)
7444{
7445 unsigned long flags;
7446 int i, ret = 0;
7447
7448 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7449 for_each_possible_cpu(i) {
7450 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7451
7452 raw_spin_lock(&rt_rq->rt_runtime_lock);
7453 rt_rq->rt_runtime = global_rt_runtime();
7454 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7455 }
7456 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7457
7458 return ret;
7459}
7460#endif /* CONFIG_RT_GROUP_SCHED */
7461
7462static int sched_dl_global_constraints(void)
7463{
7464 u64 runtime = global_rt_runtime();
7465 u64 period = global_rt_period();
7466 u64 new_bw = to_ratio(period, runtime);
7467 int cpu, ret = 0;
7468 unsigned long flags;
7469
7470 /*
7471 * Here we want to check the bandwidth not being set to some
7472 * value smaller than the currently allocated bandwidth in
7473 * any of the root_domains.
7474 *
7475 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7476 * cycling on root_domains... Discussion on different/better
7477 * solutions is welcome!
7478 */
7479 for_each_possible_cpu(cpu) {
7480 struct dl_bw *dl_b = dl_bw_of(cpu);
7481
7482 raw_spin_lock_irqsave(&dl_b->lock, flags);
7483 if (new_bw < dl_b->total_bw)
7484 ret = -EBUSY;
7485 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7486
7487 if (ret)
7488 break;
7489 }
7490
7491 return ret;
7492}
7493
7494static void sched_dl_do_global(void)
7495{
7496 u64 new_bw = -1;
7497 int cpu;
7498 unsigned long flags;
7499
7500 def_dl_bandwidth.dl_period = global_rt_period();
7501 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7502
7503 if (global_rt_runtime() != RUNTIME_INF)
7504 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7505
7506 /*
7507 * FIXME: As above...
7508 */
7509 for_each_possible_cpu(cpu) {
7510 struct dl_bw *dl_b = dl_bw_of(cpu);
7511
7512 raw_spin_lock_irqsave(&dl_b->lock, flags);
7513 dl_b->bw = new_bw;
7514 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7515 }
7516}
7517
7518static int sched_rt_global_validate(void)
7519{
7520 if (sysctl_sched_rt_period <= 0)
7521 return -EINVAL;
7522
7523 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7524 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7525 return -EINVAL;
7526
7527 return 0;
7528}
7529
7530static void sched_rt_do_global(void)
7531{
7532 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7533 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7534}
7535
7536int sched_rt_handler(struct ctl_table *table, int write,
7537 void __user *buffer, size_t *lenp,
7538 loff_t *ppos)
7539{
7540 int old_period, old_runtime;
7541 static DEFINE_MUTEX(mutex);
7542 int ret;
7543
7544 mutex_lock(&mutex);
7545 old_period = sysctl_sched_rt_period;
7546 old_runtime = sysctl_sched_rt_runtime;
7547
7548 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7549
7550 if (!ret && write) {
7551 ret = sched_rt_global_validate();
7552 if (ret)
7553 goto undo;
7554
7555 ret = sched_rt_global_constraints();
7556 if (ret)
7557 goto undo;
7558
7559 ret = sched_dl_global_constraints();
7560 if (ret)
7561 goto undo;
7562
7563 sched_rt_do_global();
7564 sched_dl_do_global();
7565 }
7566 if (0) {
7567undo:
7568 sysctl_sched_rt_period = old_period;
7569 sysctl_sched_rt_runtime = old_runtime;
7570 }
7571 mutex_unlock(&mutex);
7572
7573 return ret;
7574}
7575
7576int sched_rr_handler(struct ctl_table *table, int write,
7577 void __user *buffer, size_t *lenp,
7578 loff_t *ppos)
7579{
7580 int ret;
7581 static DEFINE_MUTEX(mutex);
7582
7583 mutex_lock(&mutex);
7584 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7585 /* make sure that internally we keep jiffies */
7586 /* also, writing zero resets timeslice to default */
7587 if (!ret && write) {
7588 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7589 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7590 }
7591 mutex_unlock(&mutex);
7592 return ret;
7593}
7594
7595#ifdef CONFIG_CGROUP_SCHED
7596
7597static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7598{
7599 return css ? container_of(css, struct task_group, css) : NULL;
7600}
7601
7602static struct cgroup_subsys_state *
7603cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7604{
7605 struct task_group *parent = css_tg(parent_css);
7606 struct task_group *tg;
7607
7608 if (!parent) {
7609 /* This is early initialization for the top cgroup */
7610 return &root_task_group.css;
7611 }
7612
7613 tg = sched_create_group(parent);
7614 if (IS_ERR(tg))
7615 return ERR_PTR(-ENOMEM);
7616
7617 return &tg->css;
7618}
7619
7620static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7621{
7622 struct task_group *tg = css_tg(css);
7623 struct task_group *parent = css_tg(css_parent(css));
7624
7625 if (parent)
7626 sched_online_group(tg, parent);
7627 return 0;
7628}
7629
7630static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7631{
7632 struct task_group *tg = css_tg(css);
7633
7634 sched_destroy_group(tg);
7635}
7636
7637static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7638{
7639 struct task_group *tg = css_tg(css);
7640
7641 sched_offline_group(tg);
7642}
7643
7644static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7645 struct cgroup_taskset *tset)
7646{
7647 struct task_struct *task;
7648
7649 cgroup_taskset_for_each(task, tset) {
7650#ifdef CONFIG_RT_GROUP_SCHED
7651 if (!sched_rt_can_attach(css_tg(css), task))
7652 return -EINVAL;
7653#else
7654 /* We don't support RT-tasks being in separate groups */
7655 if (task->sched_class != &fair_sched_class)
7656 return -EINVAL;
7657#endif
7658 }
7659 return 0;
7660}
7661
7662static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7663 struct cgroup_taskset *tset)
7664{
7665 struct task_struct *task;
7666
7667 cgroup_taskset_for_each(task, tset)
7668 sched_move_task(task);
7669}
7670
7671static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7672 struct cgroup_subsys_state *old_css,
7673 struct task_struct *task)
7674{
7675 /*
7676 * cgroup_exit() is called in the copy_process() failure path.
7677 * Ignore this case since the task hasn't ran yet, this avoids
7678 * trying to poke a half freed task state from generic code.
7679 */
7680 if (!(task->flags & PF_EXITING))
7681 return;
7682
7683 sched_move_task(task);
7684}
7685
7686#ifdef CONFIG_FAIR_GROUP_SCHED
7687static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7688 struct cftype *cftype, u64 shareval)
7689{
7690 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7691}
7692
7693static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7694 struct cftype *cft)
7695{
7696 struct task_group *tg = css_tg(css);
7697
7698 return (u64) scale_load_down(tg->shares);
7699}
7700
7701#ifdef CONFIG_CFS_BANDWIDTH
7702static DEFINE_MUTEX(cfs_constraints_mutex);
7703
7704const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7705const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7706
7707static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7708
7709static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7710{
7711 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7712 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7713
7714 if (tg == &root_task_group)
7715 return -EINVAL;
7716
7717 /*
7718 * Ensure we have at some amount of bandwidth every period. This is
7719 * to prevent reaching a state of large arrears when throttled via
7720 * entity_tick() resulting in prolonged exit starvation.
7721 */
7722 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7723 return -EINVAL;
7724
7725 /*
7726 * Likewise, bound things on the otherside by preventing insane quota
7727 * periods. This also allows us to normalize in computing quota
7728 * feasibility.
7729 */
7730 if (period > max_cfs_quota_period)
7731 return -EINVAL;
7732
7733 mutex_lock(&cfs_constraints_mutex);
7734 ret = __cfs_schedulable(tg, period, quota);
7735 if (ret)
7736 goto out_unlock;
7737
7738 runtime_enabled = quota != RUNTIME_INF;
7739 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7740 /*
7741 * If we need to toggle cfs_bandwidth_used, off->on must occur
7742 * before making related changes, and on->off must occur afterwards
7743 */
7744 if (runtime_enabled && !runtime_was_enabled)
7745 cfs_bandwidth_usage_inc();
7746 raw_spin_lock_irq(&cfs_b->lock);
7747 cfs_b->period = ns_to_ktime(period);
7748 cfs_b->quota = quota;
7749
7750 __refill_cfs_bandwidth_runtime(cfs_b);
7751 /* restart the period timer (if active) to handle new period expiry */
7752 if (runtime_enabled && cfs_b->timer_active) {
7753 /* force a reprogram */
7754 __start_cfs_bandwidth(cfs_b, true);
7755 }
7756 raw_spin_unlock_irq(&cfs_b->lock);
7757
7758 for_each_possible_cpu(i) {
7759 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7760 struct rq *rq = cfs_rq->rq;
7761
7762 raw_spin_lock_irq(&rq->lock);
7763 cfs_rq->runtime_enabled = runtime_enabled;
7764 cfs_rq->runtime_remaining = 0;
7765
7766 if (cfs_rq->throttled)
7767 unthrottle_cfs_rq(cfs_rq);
7768 raw_spin_unlock_irq(&rq->lock);
7769 }
7770 if (runtime_was_enabled && !runtime_enabled)
7771 cfs_bandwidth_usage_dec();
7772out_unlock:
7773 mutex_unlock(&cfs_constraints_mutex);
7774
7775 return ret;
7776}
7777
7778int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7779{
7780 u64 quota, period;
7781
7782 period = ktime_to_ns(tg->cfs_bandwidth.period);
7783 if (cfs_quota_us < 0)
7784 quota = RUNTIME_INF;
7785 else
7786 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7787
7788 return tg_set_cfs_bandwidth(tg, period, quota);
7789}
7790
7791long tg_get_cfs_quota(struct task_group *tg)
7792{
7793 u64 quota_us;
7794
7795 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7796 return -1;
7797
7798 quota_us = tg->cfs_bandwidth.quota;
7799 do_div(quota_us, NSEC_PER_USEC);
7800
7801 return quota_us;
7802}
7803
7804int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7805{
7806 u64 quota, period;
7807
7808 period = (u64)cfs_period_us * NSEC_PER_USEC;
7809 quota = tg->cfs_bandwidth.quota;
7810
7811 return tg_set_cfs_bandwidth(tg, period, quota);
7812}
7813
7814long tg_get_cfs_period(struct task_group *tg)
7815{
7816 u64 cfs_period_us;
7817
7818 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7819 do_div(cfs_period_us, NSEC_PER_USEC);
7820
7821 return cfs_period_us;
7822}
7823
7824static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7825 struct cftype *cft)
7826{
7827 return tg_get_cfs_quota(css_tg(css));
7828}
7829
7830static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7831 struct cftype *cftype, s64 cfs_quota_us)
7832{
7833 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7834}
7835
7836static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7837 struct cftype *cft)
7838{
7839 return tg_get_cfs_period(css_tg(css));
7840}
7841
7842static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7843 struct cftype *cftype, u64 cfs_period_us)
7844{
7845 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7846}
7847
7848struct cfs_schedulable_data {
7849 struct task_group *tg;
7850 u64 period, quota;
7851};
7852
7853/*
7854 * normalize group quota/period to be quota/max_period
7855 * note: units are usecs
7856 */
7857static u64 normalize_cfs_quota(struct task_group *tg,
7858 struct cfs_schedulable_data *d)
7859{
7860 u64 quota, period;
7861
7862 if (tg == d->tg) {
7863 period = d->period;
7864 quota = d->quota;
7865 } else {
7866 period = tg_get_cfs_period(tg);
7867 quota = tg_get_cfs_quota(tg);
7868 }
7869
7870 /* note: these should typically be equivalent */
7871 if (quota == RUNTIME_INF || quota == -1)
7872 return RUNTIME_INF;
7873
7874 return to_ratio(period, quota);
7875}
7876
7877static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7878{
7879 struct cfs_schedulable_data *d = data;
7880 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7881 s64 quota = 0, parent_quota = -1;
7882
7883 if (!tg->parent) {
7884 quota = RUNTIME_INF;
7885 } else {
7886 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7887
7888 quota = normalize_cfs_quota(tg, d);
7889 parent_quota = parent_b->hierarchal_quota;
7890
7891 /*
7892 * ensure max(child_quota) <= parent_quota, inherit when no
7893 * limit is set
7894 */
7895 if (quota == RUNTIME_INF)
7896 quota = parent_quota;
7897 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7898 return -EINVAL;
7899 }
7900 cfs_b->hierarchal_quota = quota;
7901
7902 return 0;
7903}
7904
7905static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7906{
7907 int ret;
7908 struct cfs_schedulable_data data = {
7909 .tg = tg,
7910 .period = period,
7911 .quota = quota,
7912 };
7913
7914 if (quota != RUNTIME_INF) {
7915 do_div(data.period, NSEC_PER_USEC);
7916 do_div(data.quota, NSEC_PER_USEC);
7917 }
7918
7919 rcu_read_lock();
7920 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7921 rcu_read_unlock();
7922
7923 return ret;
7924}
7925
7926static int cpu_stats_show(struct seq_file *sf, void *v)
7927{
7928 struct task_group *tg = css_tg(seq_css(sf));
7929 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7930
7931 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7932 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7933 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7934
7935 return 0;
7936}
7937#endif /* CONFIG_CFS_BANDWIDTH */
7938#endif /* CONFIG_FAIR_GROUP_SCHED */
7939
7940#ifdef CONFIG_RT_GROUP_SCHED
7941static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7942 struct cftype *cft, s64 val)
7943{
7944 return sched_group_set_rt_runtime(css_tg(css), val);
7945}
7946
7947static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7948 struct cftype *cft)
7949{
7950 return sched_group_rt_runtime(css_tg(css));
7951}
7952
7953static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7954 struct cftype *cftype, u64 rt_period_us)
7955{
7956 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7957}
7958
7959static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7960 struct cftype *cft)
7961{
7962 return sched_group_rt_period(css_tg(css));
7963}
7964#endif /* CONFIG_RT_GROUP_SCHED */
7965
7966static struct cftype cpu_files[] = {
7967#ifdef CONFIG_FAIR_GROUP_SCHED
7968 {
7969 .name = "shares",
7970 .read_u64 = cpu_shares_read_u64,
7971 .write_u64 = cpu_shares_write_u64,
7972 },
7973#endif
7974#ifdef CONFIG_CFS_BANDWIDTH
7975 {
7976 .name = "cfs_quota_us",
7977 .read_s64 = cpu_cfs_quota_read_s64,
7978 .write_s64 = cpu_cfs_quota_write_s64,
7979 },
7980 {
7981 .name = "cfs_period_us",
7982 .read_u64 = cpu_cfs_period_read_u64,
7983 .write_u64 = cpu_cfs_period_write_u64,
7984 },
7985 {
7986 .name = "stat",
7987 .seq_show = cpu_stats_show,
7988 },
7989#endif
7990#ifdef CONFIG_RT_GROUP_SCHED
7991 {
7992 .name = "rt_runtime_us",
7993 .read_s64 = cpu_rt_runtime_read,
7994 .write_s64 = cpu_rt_runtime_write,
7995 },
7996 {
7997 .name = "rt_period_us",
7998 .read_u64 = cpu_rt_period_read_uint,
7999 .write_u64 = cpu_rt_period_write_uint,
8000 },
8001#endif
8002 { } /* terminate */
8003};
8004
8005struct cgroup_subsys cpu_cgrp_subsys = {
8006 .css_alloc = cpu_cgroup_css_alloc,
8007 .css_free = cpu_cgroup_css_free,
8008 .css_online = cpu_cgroup_css_online,
8009 .css_offline = cpu_cgroup_css_offline,
8010 .can_attach = cpu_cgroup_can_attach,
8011 .attach = cpu_cgroup_attach,
8012 .exit = cpu_cgroup_exit,
8013 .base_cftypes = cpu_files,
8014 .early_init = 1,
8015};
8016
8017#endif /* CONFIG_CGROUP_SCHED */
8018
8019void dump_cpu_task(int cpu)
8020{
8021 pr_info("Task dump for CPU %d:\n", cpu);
8022 sched_show_task(cpu_curr(cpu));
8023}
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/kasan.h>
30#include <linux/mm.h>
31#include <linux/module.h>
32#include <linux/nmi.h>
33#include <linux/init.h>
34#include <linux/uaccess.h>
35#include <linux/highmem.h>
36#include <asm/mmu_context.h>
37#include <linux/interrupt.h>
38#include <linux/capability.h>
39#include <linux/completion.h>
40#include <linux/kernel_stat.h>
41#include <linux/debug_locks.h>
42#include <linux/perf_event.h>
43#include <linux/security.h>
44#include <linux/notifier.h>
45#include <linux/profile.h>
46#include <linux/freezer.h>
47#include <linux/vmalloc.h>
48#include <linux/blkdev.h>
49#include <linux/delay.h>
50#include <linux/pid_namespace.h>
51#include <linux/smp.h>
52#include <linux/threads.h>
53#include <linux/timer.h>
54#include <linux/rcupdate.h>
55#include <linux/cpu.h>
56#include <linux/cpuset.h>
57#include <linux/percpu.h>
58#include <linux/proc_fs.h>
59#include <linux/seq_file.h>
60#include <linux/sysctl.h>
61#include <linux/syscalls.h>
62#include <linux/times.h>
63#include <linux/tsacct_kern.h>
64#include <linux/kprobes.h>
65#include <linux/delayacct.h>
66#include <linux/unistd.h>
67#include <linux/pagemap.h>
68#include <linux/hrtimer.h>
69#include <linux/tick.h>
70#include <linux/ctype.h>
71#include <linux/ftrace.h>
72#include <linux/slab.h>
73#include <linux/init_task.h>
74#include <linux/context_tracking.h>
75#include <linux/compiler.h>
76#include <linux/frame.h>
77
78#include <asm/switch_to.h>
79#include <asm/tlb.h>
80#include <asm/irq_regs.h>
81#include <asm/mutex.h>
82#ifdef CONFIG_PARAVIRT
83#include <asm/paravirt.h>
84#endif
85
86#include "sched.h"
87#include "../workqueue_internal.h"
88#include "../smpboot.h"
89
90#define CREATE_TRACE_POINTS
91#include <trace/events/sched.h>
92
93DEFINE_MUTEX(sched_domains_mutex);
94DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
95
96static void update_rq_clock_task(struct rq *rq, s64 delta);
97
98void update_rq_clock(struct rq *rq)
99{
100 s64 delta;
101
102 lockdep_assert_held(&rq->lock);
103
104 if (rq->clock_skip_update & RQCF_ACT_SKIP)
105 return;
106
107 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
108 if (delta < 0)
109 return;
110 rq->clock += delta;
111 update_rq_clock_task(rq, delta);
112}
113
114/*
115 * Debugging: various feature bits
116 */
117
118#define SCHED_FEAT(name, enabled) \
119 (1UL << __SCHED_FEAT_##name) * enabled |
120
121const_debug unsigned int sysctl_sched_features =
122#include "features.h"
123 0;
124
125#undef SCHED_FEAT
126
127/*
128 * Number of tasks to iterate in a single balance run.
129 * Limited because this is done with IRQs disabled.
130 */
131const_debug unsigned int sysctl_sched_nr_migrate = 32;
132
133/*
134 * period over which we average the RT time consumption, measured
135 * in ms.
136 *
137 * default: 1s
138 */
139const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
140
141/*
142 * period over which we measure -rt task cpu usage in us.
143 * default: 1s
144 */
145unsigned int sysctl_sched_rt_period = 1000000;
146
147__read_mostly int scheduler_running;
148
149/*
150 * part of the period that we allow rt tasks to run in us.
151 * default: 0.95s
152 */
153int sysctl_sched_rt_runtime = 950000;
154
155/* cpus with isolated domains */
156cpumask_var_t cpu_isolated_map;
157
158/*
159 * this_rq_lock - lock this runqueue and disable interrupts.
160 */
161static struct rq *this_rq_lock(void)
162 __acquires(rq->lock)
163{
164 struct rq *rq;
165
166 local_irq_disable();
167 rq = this_rq();
168 raw_spin_lock(&rq->lock);
169
170 return rq;
171}
172
173#ifdef CONFIG_SCHED_HRTICK
174/*
175 * Use HR-timers to deliver accurate preemption points.
176 */
177
178static void hrtick_clear(struct rq *rq)
179{
180 if (hrtimer_active(&rq->hrtick_timer))
181 hrtimer_cancel(&rq->hrtick_timer);
182}
183
184/*
185 * High-resolution timer tick.
186 * Runs from hardirq context with interrupts disabled.
187 */
188static enum hrtimer_restart hrtick(struct hrtimer *timer)
189{
190 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
191
192 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
193
194 raw_spin_lock(&rq->lock);
195 update_rq_clock(rq);
196 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
197 raw_spin_unlock(&rq->lock);
198
199 return HRTIMER_NORESTART;
200}
201
202#ifdef CONFIG_SMP
203
204static void __hrtick_restart(struct rq *rq)
205{
206 struct hrtimer *timer = &rq->hrtick_timer;
207
208 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
209}
210
211/*
212 * called from hardirq (IPI) context
213 */
214static void __hrtick_start(void *arg)
215{
216 struct rq *rq = arg;
217
218 raw_spin_lock(&rq->lock);
219 __hrtick_restart(rq);
220 rq->hrtick_csd_pending = 0;
221 raw_spin_unlock(&rq->lock);
222}
223
224/*
225 * Called to set the hrtick timer state.
226 *
227 * called with rq->lock held and irqs disabled
228 */
229void hrtick_start(struct rq *rq, u64 delay)
230{
231 struct hrtimer *timer = &rq->hrtick_timer;
232 ktime_t time;
233 s64 delta;
234
235 /*
236 * Don't schedule slices shorter than 10000ns, that just
237 * doesn't make sense and can cause timer DoS.
238 */
239 delta = max_t(s64, delay, 10000LL);
240 time = ktime_add_ns(timer->base->get_time(), delta);
241
242 hrtimer_set_expires(timer, time);
243
244 if (rq == this_rq()) {
245 __hrtick_restart(rq);
246 } else if (!rq->hrtick_csd_pending) {
247 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
248 rq->hrtick_csd_pending = 1;
249 }
250}
251
252static int
253hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
254{
255 int cpu = (int)(long)hcpu;
256
257 switch (action) {
258 case CPU_UP_CANCELED:
259 case CPU_UP_CANCELED_FROZEN:
260 case CPU_DOWN_PREPARE:
261 case CPU_DOWN_PREPARE_FROZEN:
262 case CPU_DEAD:
263 case CPU_DEAD_FROZEN:
264 hrtick_clear(cpu_rq(cpu));
265 return NOTIFY_OK;
266 }
267
268 return NOTIFY_DONE;
269}
270
271static __init void init_hrtick(void)
272{
273 hotcpu_notifier(hotplug_hrtick, 0);
274}
275#else
276/*
277 * Called to set the hrtick timer state.
278 *
279 * called with rq->lock held and irqs disabled
280 */
281void hrtick_start(struct rq *rq, u64 delay)
282{
283 /*
284 * Don't schedule slices shorter than 10000ns, that just
285 * doesn't make sense. Rely on vruntime for fairness.
286 */
287 delay = max_t(u64, delay, 10000LL);
288 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
289 HRTIMER_MODE_REL_PINNED);
290}
291
292static inline void init_hrtick(void)
293{
294}
295#endif /* CONFIG_SMP */
296
297static void init_rq_hrtick(struct rq *rq)
298{
299#ifdef CONFIG_SMP
300 rq->hrtick_csd_pending = 0;
301
302 rq->hrtick_csd.flags = 0;
303 rq->hrtick_csd.func = __hrtick_start;
304 rq->hrtick_csd.info = rq;
305#endif
306
307 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
308 rq->hrtick_timer.function = hrtick;
309}
310#else /* CONFIG_SCHED_HRTICK */
311static inline void hrtick_clear(struct rq *rq)
312{
313}
314
315static inline void init_rq_hrtick(struct rq *rq)
316{
317}
318
319static inline void init_hrtick(void)
320{
321}
322#endif /* CONFIG_SCHED_HRTICK */
323
324/*
325 * cmpxchg based fetch_or, macro so it works for different integer types
326 */
327#define fetch_or(ptr, mask) \
328 ({ \
329 typeof(ptr) _ptr = (ptr); \
330 typeof(mask) _mask = (mask); \
331 typeof(*_ptr) _old, _val = *_ptr; \
332 \
333 for (;;) { \
334 _old = cmpxchg(_ptr, _val, _val | _mask); \
335 if (_old == _val) \
336 break; \
337 _val = _old; \
338 } \
339 _old; \
340})
341
342#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
343/*
344 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
345 * this avoids any races wrt polling state changes and thereby avoids
346 * spurious IPIs.
347 */
348static bool set_nr_and_not_polling(struct task_struct *p)
349{
350 struct thread_info *ti = task_thread_info(p);
351 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
352}
353
354/*
355 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
356 *
357 * If this returns true, then the idle task promises to call
358 * sched_ttwu_pending() and reschedule soon.
359 */
360static bool set_nr_if_polling(struct task_struct *p)
361{
362 struct thread_info *ti = task_thread_info(p);
363 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
364
365 for (;;) {
366 if (!(val & _TIF_POLLING_NRFLAG))
367 return false;
368 if (val & _TIF_NEED_RESCHED)
369 return true;
370 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
371 if (old == val)
372 break;
373 val = old;
374 }
375 return true;
376}
377
378#else
379static bool set_nr_and_not_polling(struct task_struct *p)
380{
381 set_tsk_need_resched(p);
382 return true;
383}
384
385#ifdef CONFIG_SMP
386static bool set_nr_if_polling(struct task_struct *p)
387{
388 return false;
389}
390#endif
391#endif
392
393void wake_q_add(struct wake_q_head *head, struct task_struct *task)
394{
395 struct wake_q_node *node = &task->wake_q;
396
397 /*
398 * Atomically grab the task, if ->wake_q is !nil already it means
399 * its already queued (either by us or someone else) and will get the
400 * wakeup due to that.
401 *
402 * This cmpxchg() implies a full barrier, which pairs with the write
403 * barrier implied by the wakeup in wake_up_list().
404 */
405 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
406 return;
407
408 get_task_struct(task);
409
410 /*
411 * The head is context local, there can be no concurrency.
412 */
413 *head->lastp = node;
414 head->lastp = &node->next;
415}
416
417void wake_up_q(struct wake_q_head *head)
418{
419 struct wake_q_node *node = head->first;
420
421 while (node != WAKE_Q_TAIL) {
422 struct task_struct *task;
423
424 task = container_of(node, struct task_struct, wake_q);
425 BUG_ON(!task);
426 /* task can safely be re-inserted now */
427 node = node->next;
428 task->wake_q.next = NULL;
429
430 /*
431 * wake_up_process() implies a wmb() to pair with the queueing
432 * in wake_q_add() so as not to miss wakeups.
433 */
434 wake_up_process(task);
435 put_task_struct(task);
436 }
437}
438
439/*
440 * resched_curr - mark rq's current task 'to be rescheduled now'.
441 *
442 * On UP this means the setting of the need_resched flag, on SMP it
443 * might also involve a cross-CPU call to trigger the scheduler on
444 * the target CPU.
445 */
446void resched_curr(struct rq *rq)
447{
448 struct task_struct *curr = rq->curr;
449 int cpu;
450
451 lockdep_assert_held(&rq->lock);
452
453 if (test_tsk_need_resched(curr))
454 return;
455
456 cpu = cpu_of(rq);
457
458 if (cpu == smp_processor_id()) {
459 set_tsk_need_resched(curr);
460 set_preempt_need_resched();
461 return;
462 }
463
464 if (set_nr_and_not_polling(curr))
465 smp_send_reschedule(cpu);
466 else
467 trace_sched_wake_idle_without_ipi(cpu);
468}
469
470void resched_cpu(int cpu)
471{
472 struct rq *rq = cpu_rq(cpu);
473 unsigned long flags;
474
475 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
476 return;
477 resched_curr(rq);
478 raw_spin_unlock_irqrestore(&rq->lock, flags);
479}
480
481#ifdef CONFIG_SMP
482#ifdef CONFIG_NO_HZ_COMMON
483/*
484 * In the semi idle case, use the nearest busy cpu for migrating timers
485 * from an idle cpu. This is good for power-savings.
486 *
487 * We don't do similar optimization for completely idle system, as
488 * selecting an idle cpu will add more delays to the timers than intended
489 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
490 */
491int get_nohz_timer_target(void)
492{
493 int i, cpu = smp_processor_id();
494 struct sched_domain *sd;
495
496 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
497 return cpu;
498
499 rcu_read_lock();
500 for_each_domain(cpu, sd) {
501 for_each_cpu(i, sched_domain_span(sd)) {
502 if (!idle_cpu(i) && is_housekeeping_cpu(cpu)) {
503 cpu = i;
504 goto unlock;
505 }
506 }
507 }
508
509 if (!is_housekeeping_cpu(cpu))
510 cpu = housekeeping_any_cpu();
511unlock:
512 rcu_read_unlock();
513 return cpu;
514}
515/*
516 * When add_timer_on() enqueues a timer into the timer wheel of an
517 * idle CPU then this timer might expire before the next timer event
518 * which is scheduled to wake up that CPU. In case of a completely
519 * idle system the next event might even be infinite time into the
520 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
521 * leaves the inner idle loop so the newly added timer is taken into
522 * account when the CPU goes back to idle and evaluates the timer
523 * wheel for the next timer event.
524 */
525static void wake_up_idle_cpu(int cpu)
526{
527 struct rq *rq = cpu_rq(cpu);
528
529 if (cpu == smp_processor_id())
530 return;
531
532 if (set_nr_and_not_polling(rq->idle))
533 smp_send_reschedule(cpu);
534 else
535 trace_sched_wake_idle_without_ipi(cpu);
536}
537
538static bool wake_up_full_nohz_cpu(int cpu)
539{
540 /*
541 * We just need the target to call irq_exit() and re-evaluate
542 * the next tick. The nohz full kick at least implies that.
543 * If needed we can still optimize that later with an
544 * empty IRQ.
545 */
546 if (tick_nohz_full_cpu(cpu)) {
547 if (cpu != smp_processor_id() ||
548 tick_nohz_tick_stopped())
549 tick_nohz_full_kick_cpu(cpu);
550 return true;
551 }
552
553 return false;
554}
555
556void wake_up_nohz_cpu(int cpu)
557{
558 if (!wake_up_full_nohz_cpu(cpu))
559 wake_up_idle_cpu(cpu);
560}
561
562static inline bool got_nohz_idle_kick(void)
563{
564 int cpu = smp_processor_id();
565
566 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
567 return false;
568
569 if (idle_cpu(cpu) && !need_resched())
570 return true;
571
572 /*
573 * We can't run Idle Load Balance on this CPU for this time so we
574 * cancel it and clear NOHZ_BALANCE_KICK
575 */
576 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
577 return false;
578}
579
580#else /* CONFIG_NO_HZ_COMMON */
581
582static inline bool got_nohz_idle_kick(void)
583{
584 return false;
585}
586
587#endif /* CONFIG_NO_HZ_COMMON */
588
589#ifdef CONFIG_NO_HZ_FULL
590bool sched_can_stop_tick(struct rq *rq)
591{
592 int fifo_nr_running;
593
594 /* Deadline tasks, even if single, need the tick */
595 if (rq->dl.dl_nr_running)
596 return false;
597
598 /*
599 * If there are more than one RR tasks, we need the tick to effect the
600 * actual RR behaviour.
601 */
602 if (rq->rt.rr_nr_running) {
603 if (rq->rt.rr_nr_running == 1)
604 return true;
605 else
606 return false;
607 }
608
609 /*
610 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
611 * forced preemption between FIFO tasks.
612 */
613 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
614 if (fifo_nr_running)
615 return true;
616
617 /*
618 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
619 * if there's more than one we need the tick for involuntary
620 * preemption.
621 */
622 if (rq->nr_running > 1)
623 return false;
624
625 return true;
626}
627#endif /* CONFIG_NO_HZ_FULL */
628
629void sched_avg_update(struct rq *rq)
630{
631 s64 period = sched_avg_period();
632
633 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
634 /*
635 * Inline assembly required to prevent the compiler
636 * optimising this loop into a divmod call.
637 * See __iter_div_u64_rem() for another example of this.
638 */
639 asm("" : "+rm" (rq->age_stamp));
640 rq->age_stamp += period;
641 rq->rt_avg /= 2;
642 }
643}
644
645#endif /* CONFIG_SMP */
646
647#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
648 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
649/*
650 * Iterate task_group tree rooted at *from, calling @down when first entering a
651 * node and @up when leaving it for the final time.
652 *
653 * Caller must hold rcu_lock or sufficient equivalent.
654 */
655int walk_tg_tree_from(struct task_group *from,
656 tg_visitor down, tg_visitor up, void *data)
657{
658 struct task_group *parent, *child;
659 int ret;
660
661 parent = from;
662
663down:
664 ret = (*down)(parent, data);
665 if (ret)
666 goto out;
667 list_for_each_entry_rcu(child, &parent->children, siblings) {
668 parent = child;
669 goto down;
670
671up:
672 continue;
673 }
674 ret = (*up)(parent, data);
675 if (ret || parent == from)
676 goto out;
677
678 child = parent;
679 parent = parent->parent;
680 if (parent)
681 goto up;
682out:
683 return ret;
684}
685
686int tg_nop(struct task_group *tg, void *data)
687{
688 return 0;
689}
690#endif
691
692static void set_load_weight(struct task_struct *p)
693{
694 int prio = p->static_prio - MAX_RT_PRIO;
695 struct load_weight *load = &p->se.load;
696
697 /*
698 * SCHED_IDLE tasks get minimal weight:
699 */
700 if (idle_policy(p->policy)) {
701 load->weight = scale_load(WEIGHT_IDLEPRIO);
702 load->inv_weight = WMULT_IDLEPRIO;
703 return;
704 }
705
706 load->weight = scale_load(sched_prio_to_weight[prio]);
707 load->inv_weight = sched_prio_to_wmult[prio];
708}
709
710static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
711{
712 update_rq_clock(rq);
713 if (!(flags & ENQUEUE_RESTORE))
714 sched_info_queued(rq, p);
715 p->sched_class->enqueue_task(rq, p, flags);
716}
717
718static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
719{
720 update_rq_clock(rq);
721 if (!(flags & DEQUEUE_SAVE))
722 sched_info_dequeued(rq, p);
723 p->sched_class->dequeue_task(rq, p, flags);
724}
725
726void activate_task(struct rq *rq, struct task_struct *p, int flags)
727{
728 if (task_contributes_to_load(p))
729 rq->nr_uninterruptible--;
730
731 enqueue_task(rq, p, flags);
732}
733
734void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
735{
736 if (task_contributes_to_load(p))
737 rq->nr_uninterruptible++;
738
739 dequeue_task(rq, p, flags);
740}
741
742static void update_rq_clock_task(struct rq *rq, s64 delta)
743{
744/*
745 * In theory, the compile should just see 0 here, and optimize out the call
746 * to sched_rt_avg_update. But I don't trust it...
747 */
748#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
749 s64 steal = 0, irq_delta = 0;
750#endif
751#ifdef CONFIG_IRQ_TIME_ACCOUNTING
752 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
753
754 /*
755 * Since irq_time is only updated on {soft,}irq_exit, we might run into
756 * this case when a previous update_rq_clock() happened inside a
757 * {soft,}irq region.
758 *
759 * When this happens, we stop ->clock_task and only update the
760 * prev_irq_time stamp to account for the part that fit, so that a next
761 * update will consume the rest. This ensures ->clock_task is
762 * monotonic.
763 *
764 * It does however cause some slight miss-attribution of {soft,}irq
765 * time, a more accurate solution would be to update the irq_time using
766 * the current rq->clock timestamp, except that would require using
767 * atomic ops.
768 */
769 if (irq_delta > delta)
770 irq_delta = delta;
771
772 rq->prev_irq_time += irq_delta;
773 delta -= irq_delta;
774#endif
775#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
776 if (static_key_false((¶virt_steal_rq_enabled))) {
777 steal = paravirt_steal_clock(cpu_of(rq));
778 steal -= rq->prev_steal_time_rq;
779
780 if (unlikely(steal > delta))
781 steal = delta;
782
783 rq->prev_steal_time_rq += steal;
784 delta -= steal;
785 }
786#endif
787
788 rq->clock_task += delta;
789
790#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
791 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
792 sched_rt_avg_update(rq, irq_delta + steal);
793#endif
794}
795
796void sched_set_stop_task(int cpu, struct task_struct *stop)
797{
798 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
799 struct task_struct *old_stop = cpu_rq(cpu)->stop;
800
801 if (stop) {
802 /*
803 * Make it appear like a SCHED_FIFO task, its something
804 * userspace knows about and won't get confused about.
805 *
806 * Also, it will make PI more or less work without too
807 * much confusion -- but then, stop work should not
808 * rely on PI working anyway.
809 */
810 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
811
812 stop->sched_class = &stop_sched_class;
813 }
814
815 cpu_rq(cpu)->stop = stop;
816
817 if (old_stop) {
818 /*
819 * Reset it back to a normal scheduling class so that
820 * it can die in pieces.
821 */
822 old_stop->sched_class = &rt_sched_class;
823 }
824}
825
826/*
827 * __normal_prio - return the priority that is based on the static prio
828 */
829static inline int __normal_prio(struct task_struct *p)
830{
831 return p->static_prio;
832}
833
834/*
835 * Calculate the expected normal priority: i.e. priority
836 * without taking RT-inheritance into account. Might be
837 * boosted by interactivity modifiers. Changes upon fork,
838 * setprio syscalls, and whenever the interactivity
839 * estimator recalculates.
840 */
841static inline int normal_prio(struct task_struct *p)
842{
843 int prio;
844
845 if (task_has_dl_policy(p))
846 prio = MAX_DL_PRIO-1;
847 else if (task_has_rt_policy(p))
848 prio = MAX_RT_PRIO-1 - p->rt_priority;
849 else
850 prio = __normal_prio(p);
851 return prio;
852}
853
854/*
855 * Calculate the current priority, i.e. the priority
856 * taken into account by the scheduler. This value might
857 * be boosted by RT tasks, or might be boosted by
858 * interactivity modifiers. Will be RT if the task got
859 * RT-boosted. If not then it returns p->normal_prio.
860 */
861static int effective_prio(struct task_struct *p)
862{
863 p->normal_prio = normal_prio(p);
864 /*
865 * If we are RT tasks or we were boosted to RT priority,
866 * keep the priority unchanged. Otherwise, update priority
867 * to the normal priority:
868 */
869 if (!rt_prio(p->prio))
870 return p->normal_prio;
871 return p->prio;
872}
873
874/**
875 * task_curr - is this task currently executing on a CPU?
876 * @p: the task in question.
877 *
878 * Return: 1 if the task is currently executing. 0 otherwise.
879 */
880inline int task_curr(const struct task_struct *p)
881{
882 return cpu_curr(task_cpu(p)) == p;
883}
884
885/*
886 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
887 * use the balance_callback list if you want balancing.
888 *
889 * this means any call to check_class_changed() must be followed by a call to
890 * balance_callback().
891 */
892static inline void check_class_changed(struct rq *rq, struct task_struct *p,
893 const struct sched_class *prev_class,
894 int oldprio)
895{
896 if (prev_class != p->sched_class) {
897 if (prev_class->switched_from)
898 prev_class->switched_from(rq, p);
899
900 p->sched_class->switched_to(rq, p);
901 } else if (oldprio != p->prio || dl_task(p))
902 p->sched_class->prio_changed(rq, p, oldprio);
903}
904
905void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
906{
907 const struct sched_class *class;
908
909 if (p->sched_class == rq->curr->sched_class) {
910 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
911 } else {
912 for_each_class(class) {
913 if (class == rq->curr->sched_class)
914 break;
915 if (class == p->sched_class) {
916 resched_curr(rq);
917 break;
918 }
919 }
920 }
921
922 /*
923 * A queue event has occurred, and we're going to schedule. In
924 * this case, we can save a useless back to back clock update.
925 */
926 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
927 rq_clock_skip_update(rq, true);
928}
929
930#ifdef CONFIG_SMP
931/*
932 * This is how migration works:
933 *
934 * 1) we invoke migration_cpu_stop() on the target CPU using
935 * stop_one_cpu().
936 * 2) stopper starts to run (implicitly forcing the migrated thread
937 * off the CPU)
938 * 3) it checks whether the migrated task is still in the wrong runqueue.
939 * 4) if it's in the wrong runqueue then the migration thread removes
940 * it and puts it into the right queue.
941 * 5) stopper completes and stop_one_cpu() returns and the migration
942 * is done.
943 */
944
945/*
946 * move_queued_task - move a queued task to new rq.
947 *
948 * Returns (locked) new rq. Old rq's lock is released.
949 */
950static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
951{
952 lockdep_assert_held(&rq->lock);
953
954 p->on_rq = TASK_ON_RQ_MIGRATING;
955 dequeue_task(rq, p, 0);
956 set_task_cpu(p, new_cpu);
957 raw_spin_unlock(&rq->lock);
958
959 rq = cpu_rq(new_cpu);
960
961 raw_spin_lock(&rq->lock);
962 BUG_ON(task_cpu(p) != new_cpu);
963 enqueue_task(rq, p, 0);
964 p->on_rq = TASK_ON_RQ_QUEUED;
965 check_preempt_curr(rq, p, 0);
966
967 return rq;
968}
969
970struct migration_arg {
971 struct task_struct *task;
972 int dest_cpu;
973};
974
975/*
976 * Move (not current) task off this cpu, onto dest cpu. We're doing
977 * this because either it can't run here any more (set_cpus_allowed()
978 * away from this CPU, or CPU going down), or because we're
979 * attempting to rebalance this task on exec (sched_exec).
980 *
981 * So we race with normal scheduler movements, but that's OK, as long
982 * as the task is no longer on this CPU.
983 */
984static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
985{
986 if (unlikely(!cpu_active(dest_cpu)))
987 return rq;
988
989 /* Affinity changed (again). */
990 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
991 return rq;
992
993 rq = move_queued_task(rq, p, dest_cpu);
994
995 return rq;
996}
997
998/*
999 * migration_cpu_stop - this will be executed by a highprio stopper thread
1000 * and performs thread migration by bumping thread off CPU then
1001 * 'pushing' onto another runqueue.
1002 */
1003static int migration_cpu_stop(void *data)
1004{
1005 struct migration_arg *arg = data;
1006 struct task_struct *p = arg->task;
1007 struct rq *rq = this_rq();
1008
1009 /*
1010 * The original target cpu might have gone down and we might
1011 * be on another cpu but it doesn't matter.
1012 */
1013 local_irq_disable();
1014 /*
1015 * We need to explicitly wake pending tasks before running
1016 * __migrate_task() such that we will not miss enforcing cpus_allowed
1017 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1018 */
1019 sched_ttwu_pending();
1020
1021 raw_spin_lock(&p->pi_lock);
1022 raw_spin_lock(&rq->lock);
1023 /*
1024 * If task_rq(p) != rq, it cannot be migrated here, because we're
1025 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1026 * we're holding p->pi_lock.
1027 */
1028 if (task_rq(p) == rq && task_on_rq_queued(p))
1029 rq = __migrate_task(rq, p, arg->dest_cpu);
1030 raw_spin_unlock(&rq->lock);
1031 raw_spin_unlock(&p->pi_lock);
1032
1033 local_irq_enable();
1034 return 0;
1035}
1036
1037/*
1038 * sched_class::set_cpus_allowed must do the below, but is not required to
1039 * actually call this function.
1040 */
1041void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1042{
1043 cpumask_copy(&p->cpus_allowed, new_mask);
1044 p->nr_cpus_allowed = cpumask_weight(new_mask);
1045}
1046
1047void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1048{
1049 struct rq *rq = task_rq(p);
1050 bool queued, running;
1051
1052 lockdep_assert_held(&p->pi_lock);
1053
1054 queued = task_on_rq_queued(p);
1055 running = task_current(rq, p);
1056
1057 if (queued) {
1058 /*
1059 * Because __kthread_bind() calls this on blocked tasks without
1060 * holding rq->lock.
1061 */
1062 lockdep_assert_held(&rq->lock);
1063 dequeue_task(rq, p, DEQUEUE_SAVE);
1064 }
1065 if (running)
1066 put_prev_task(rq, p);
1067
1068 p->sched_class->set_cpus_allowed(p, new_mask);
1069
1070 if (running)
1071 p->sched_class->set_curr_task(rq);
1072 if (queued)
1073 enqueue_task(rq, p, ENQUEUE_RESTORE);
1074}
1075
1076/*
1077 * Change a given task's CPU affinity. Migrate the thread to a
1078 * proper CPU and schedule it away if the CPU it's executing on
1079 * is removed from the allowed bitmask.
1080 *
1081 * NOTE: the caller must have a valid reference to the task, the
1082 * task must not exit() & deallocate itself prematurely. The
1083 * call is not atomic; no spinlocks may be held.
1084 */
1085static int __set_cpus_allowed_ptr(struct task_struct *p,
1086 const struct cpumask *new_mask, bool check)
1087{
1088 unsigned long flags;
1089 struct rq *rq;
1090 unsigned int dest_cpu;
1091 int ret = 0;
1092
1093 rq = task_rq_lock(p, &flags);
1094
1095 /*
1096 * Must re-check here, to close a race against __kthread_bind(),
1097 * sched_setaffinity() is not guaranteed to observe the flag.
1098 */
1099 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1100 ret = -EINVAL;
1101 goto out;
1102 }
1103
1104 if (cpumask_equal(&p->cpus_allowed, new_mask))
1105 goto out;
1106
1107 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1108 ret = -EINVAL;
1109 goto out;
1110 }
1111
1112 do_set_cpus_allowed(p, new_mask);
1113
1114 /* Can the task run on the task's current CPU? If so, we're done */
1115 if (cpumask_test_cpu(task_cpu(p), new_mask))
1116 goto out;
1117
1118 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1119 if (task_running(rq, p) || p->state == TASK_WAKING) {
1120 struct migration_arg arg = { p, dest_cpu };
1121 /* Need help from migration thread: drop lock and wait. */
1122 task_rq_unlock(rq, p, &flags);
1123 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1124 tlb_migrate_finish(p->mm);
1125 return 0;
1126 } else if (task_on_rq_queued(p)) {
1127 /*
1128 * OK, since we're going to drop the lock immediately
1129 * afterwards anyway.
1130 */
1131 lockdep_unpin_lock(&rq->lock);
1132 rq = move_queued_task(rq, p, dest_cpu);
1133 lockdep_pin_lock(&rq->lock);
1134 }
1135out:
1136 task_rq_unlock(rq, p, &flags);
1137
1138 return ret;
1139}
1140
1141int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1142{
1143 return __set_cpus_allowed_ptr(p, new_mask, false);
1144}
1145EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1146
1147void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1148{
1149#ifdef CONFIG_SCHED_DEBUG
1150 /*
1151 * We should never call set_task_cpu() on a blocked task,
1152 * ttwu() will sort out the placement.
1153 */
1154 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1155 !p->on_rq);
1156
1157 /*
1158 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1159 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1160 * time relying on p->on_rq.
1161 */
1162 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1163 p->sched_class == &fair_sched_class &&
1164 (p->on_rq && !task_on_rq_migrating(p)));
1165
1166#ifdef CONFIG_LOCKDEP
1167 /*
1168 * The caller should hold either p->pi_lock or rq->lock, when changing
1169 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1170 *
1171 * sched_move_task() holds both and thus holding either pins the cgroup,
1172 * see task_group().
1173 *
1174 * Furthermore, all task_rq users should acquire both locks, see
1175 * task_rq_lock().
1176 */
1177 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1178 lockdep_is_held(&task_rq(p)->lock)));
1179#endif
1180#endif
1181
1182 trace_sched_migrate_task(p, new_cpu);
1183
1184 if (task_cpu(p) != new_cpu) {
1185 if (p->sched_class->migrate_task_rq)
1186 p->sched_class->migrate_task_rq(p);
1187 p->se.nr_migrations++;
1188 perf_event_task_migrate(p);
1189 }
1190
1191 __set_task_cpu(p, new_cpu);
1192}
1193
1194static void __migrate_swap_task(struct task_struct *p, int cpu)
1195{
1196 if (task_on_rq_queued(p)) {
1197 struct rq *src_rq, *dst_rq;
1198
1199 src_rq = task_rq(p);
1200 dst_rq = cpu_rq(cpu);
1201
1202 p->on_rq = TASK_ON_RQ_MIGRATING;
1203 deactivate_task(src_rq, p, 0);
1204 set_task_cpu(p, cpu);
1205 activate_task(dst_rq, p, 0);
1206 p->on_rq = TASK_ON_RQ_QUEUED;
1207 check_preempt_curr(dst_rq, p, 0);
1208 } else {
1209 /*
1210 * Task isn't running anymore; make it appear like we migrated
1211 * it before it went to sleep. This means on wakeup we make the
1212 * previous cpu our targer instead of where it really is.
1213 */
1214 p->wake_cpu = cpu;
1215 }
1216}
1217
1218struct migration_swap_arg {
1219 struct task_struct *src_task, *dst_task;
1220 int src_cpu, dst_cpu;
1221};
1222
1223static int migrate_swap_stop(void *data)
1224{
1225 struct migration_swap_arg *arg = data;
1226 struct rq *src_rq, *dst_rq;
1227 int ret = -EAGAIN;
1228
1229 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1230 return -EAGAIN;
1231
1232 src_rq = cpu_rq(arg->src_cpu);
1233 dst_rq = cpu_rq(arg->dst_cpu);
1234
1235 double_raw_lock(&arg->src_task->pi_lock,
1236 &arg->dst_task->pi_lock);
1237 double_rq_lock(src_rq, dst_rq);
1238
1239 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1240 goto unlock;
1241
1242 if (task_cpu(arg->src_task) != arg->src_cpu)
1243 goto unlock;
1244
1245 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1246 goto unlock;
1247
1248 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1249 goto unlock;
1250
1251 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1252 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1253
1254 ret = 0;
1255
1256unlock:
1257 double_rq_unlock(src_rq, dst_rq);
1258 raw_spin_unlock(&arg->dst_task->pi_lock);
1259 raw_spin_unlock(&arg->src_task->pi_lock);
1260
1261 return ret;
1262}
1263
1264/*
1265 * Cross migrate two tasks
1266 */
1267int migrate_swap(struct task_struct *cur, struct task_struct *p)
1268{
1269 struct migration_swap_arg arg;
1270 int ret = -EINVAL;
1271
1272 arg = (struct migration_swap_arg){
1273 .src_task = cur,
1274 .src_cpu = task_cpu(cur),
1275 .dst_task = p,
1276 .dst_cpu = task_cpu(p),
1277 };
1278
1279 if (arg.src_cpu == arg.dst_cpu)
1280 goto out;
1281
1282 /*
1283 * These three tests are all lockless; this is OK since all of them
1284 * will be re-checked with proper locks held further down the line.
1285 */
1286 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1287 goto out;
1288
1289 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1290 goto out;
1291
1292 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1293 goto out;
1294
1295 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1296 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1297
1298out:
1299 return ret;
1300}
1301
1302/*
1303 * wait_task_inactive - wait for a thread to unschedule.
1304 *
1305 * If @match_state is nonzero, it's the @p->state value just checked and
1306 * not expected to change. If it changes, i.e. @p might have woken up,
1307 * then return zero. When we succeed in waiting for @p to be off its CPU,
1308 * we return a positive number (its total switch count). If a second call
1309 * a short while later returns the same number, the caller can be sure that
1310 * @p has remained unscheduled the whole time.
1311 *
1312 * The caller must ensure that the task *will* unschedule sometime soon,
1313 * else this function might spin for a *long* time. This function can't
1314 * be called with interrupts off, or it may introduce deadlock with
1315 * smp_call_function() if an IPI is sent by the same process we are
1316 * waiting to become inactive.
1317 */
1318unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1319{
1320 unsigned long flags;
1321 int running, queued;
1322 unsigned long ncsw;
1323 struct rq *rq;
1324
1325 for (;;) {
1326 /*
1327 * We do the initial early heuristics without holding
1328 * any task-queue locks at all. We'll only try to get
1329 * the runqueue lock when things look like they will
1330 * work out!
1331 */
1332 rq = task_rq(p);
1333
1334 /*
1335 * If the task is actively running on another CPU
1336 * still, just relax and busy-wait without holding
1337 * any locks.
1338 *
1339 * NOTE! Since we don't hold any locks, it's not
1340 * even sure that "rq" stays as the right runqueue!
1341 * But we don't care, since "task_running()" will
1342 * return false if the runqueue has changed and p
1343 * is actually now running somewhere else!
1344 */
1345 while (task_running(rq, p)) {
1346 if (match_state && unlikely(p->state != match_state))
1347 return 0;
1348 cpu_relax();
1349 }
1350
1351 /*
1352 * Ok, time to look more closely! We need the rq
1353 * lock now, to be *sure*. If we're wrong, we'll
1354 * just go back and repeat.
1355 */
1356 rq = task_rq_lock(p, &flags);
1357 trace_sched_wait_task(p);
1358 running = task_running(rq, p);
1359 queued = task_on_rq_queued(p);
1360 ncsw = 0;
1361 if (!match_state || p->state == match_state)
1362 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1363 task_rq_unlock(rq, p, &flags);
1364
1365 /*
1366 * If it changed from the expected state, bail out now.
1367 */
1368 if (unlikely(!ncsw))
1369 break;
1370
1371 /*
1372 * Was it really running after all now that we
1373 * checked with the proper locks actually held?
1374 *
1375 * Oops. Go back and try again..
1376 */
1377 if (unlikely(running)) {
1378 cpu_relax();
1379 continue;
1380 }
1381
1382 /*
1383 * It's not enough that it's not actively running,
1384 * it must be off the runqueue _entirely_, and not
1385 * preempted!
1386 *
1387 * So if it was still runnable (but just not actively
1388 * running right now), it's preempted, and we should
1389 * yield - it could be a while.
1390 */
1391 if (unlikely(queued)) {
1392 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1393
1394 set_current_state(TASK_UNINTERRUPTIBLE);
1395 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1396 continue;
1397 }
1398
1399 /*
1400 * Ahh, all good. It wasn't running, and it wasn't
1401 * runnable, which means that it will never become
1402 * running in the future either. We're all done!
1403 */
1404 break;
1405 }
1406
1407 return ncsw;
1408}
1409
1410/***
1411 * kick_process - kick a running thread to enter/exit the kernel
1412 * @p: the to-be-kicked thread
1413 *
1414 * Cause a process which is running on another CPU to enter
1415 * kernel-mode, without any delay. (to get signals handled.)
1416 *
1417 * NOTE: this function doesn't have to take the runqueue lock,
1418 * because all it wants to ensure is that the remote task enters
1419 * the kernel. If the IPI races and the task has been migrated
1420 * to another CPU then no harm is done and the purpose has been
1421 * achieved as well.
1422 */
1423void kick_process(struct task_struct *p)
1424{
1425 int cpu;
1426
1427 preempt_disable();
1428 cpu = task_cpu(p);
1429 if ((cpu != smp_processor_id()) && task_curr(p))
1430 smp_send_reschedule(cpu);
1431 preempt_enable();
1432}
1433EXPORT_SYMBOL_GPL(kick_process);
1434
1435/*
1436 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1437 */
1438static int select_fallback_rq(int cpu, struct task_struct *p)
1439{
1440 int nid = cpu_to_node(cpu);
1441 const struct cpumask *nodemask = NULL;
1442 enum { cpuset, possible, fail } state = cpuset;
1443 int dest_cpu;
1444
1445 /*
1446 * If the node that the cpu is on has been offlined, cpu_to_node()
1447 * will return -1. There is no cpu on the node, and we should
1448 * select the cpu on the other node.
1449 */
1450 if (nid != -1) {
1451 nodemask = cpumask_of_node(nid);
1452
1453 /* Look for allowed, online CPU in same node. */
1454 for_each_cpu(dest_cpu, nodemask) {
1455 if (!cpu_online(dest_cpu))
1456 continue;
1457 if (!cpu_active(dest_cpu))
1458 continue;
1459 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1460 return dest_cpu;
1461 }
1462 }
1463
1464 for (;;) {
1465 /* Any allowed, online CPU? */
1466 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1467 if (!cpu_online(dest_cpu))
1468 continue;
1469 if (!cpu_active(dest_cpu))
1470 continue;
1471 goto out;
1472 }
1473
1474 /* No more Mr. Nice Guy. */
1475 switch (state) {
1476 case cpuset:
1477 if (IS_ENABLED(CONFIG_CPUSETS)) {
1478 cpuset_cpus_allowed_fallback(p);
1479 state = possible;
1480 break;
1481 }
1482 /* fall-through */
1483 case possible:
1484 do_set_cpus_allowed(p, cpu_possible_mask);
1485 state = fail;
1486 break;
1487
1488 case fail:
1489 BUG();
1490 break;
1491 }
1492 }
1493
1494out:
1495 if (state != cpuset) {
1496 /*
1497 * Don't tell them about moving exiting tasks or
1498 * kernel threads (both mm NULL), since they never
1499 * leave kernel.
1500 */
1501 if (p->mm && printk_ratelimit()) {
1502 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1503 task_pid_nr(p), p->comm, cpu);
1504 }
1505 }
1506
1507 return dest_cpu;
1508}
1509
1510/*
1511 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1512 */
1513static inline
1514int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1515{
1516 lockdep_assert_held(&p->pi_lock);
1517
1518 if (p->nr_cpus_allowed > 1)
1519 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1520
1521 /*
1522 * In order not to call set_task_cpu() on a blocking task we need
1523 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1524 * cpu.
1525 *
1526 * Since this is common to all placement strategies, this lives here.
1527 *
1528 * [ this allows ->select_task() to simply return task_cpu(p) and
1529 * not worry about this generic constraint ]
1530 */
1531 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1532 !cpu_online(cpu)))
1533 cpu = select_fallback_rq(task_cpu(p), p);
1534
1535 return cpu;
1536}
1537
1538static void update_avg(u64 *avg, u64 sample)
1539{
1540 s64 diff = sample - *avg;
1541 *avg += diff >> 3;
1542}
1543
1544#else
1545
1546static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1547 const struct cpumask *new_mask, bool check)
1548{
1549 return set_cpus_allowed_ptr(p, new_mask);
1550}
1551
1552#endif /* CONFIG_SMP */
1553
1554static void
1555ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1556{
1557#ifdef CONFIG_SCHEDSTATS
1558 struct rq *rq = this_rq();
1559
1560#ifdef CONFIG_SMP
1561 int this_cpu = smp_processor_id();
1562
1563 if (cpu == this_cpu) {
1564 schedstat_inc(rq, ttwu_local);
1565 schedstat_inc(p, se.statistics.nr_wakeups_local);
1566 } else {
1567 struct sched_domain *sd;
1568
1569 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1570 rcu_read_lock();
1571 for_each_domain(this_cpu, sd) {
1572 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1573 schedstat_inc(sd, ttwu_wake_remote);
1574 break;
1575 }
1576 }
1577 rcu_read_unlock();
1578 }
1579
1580 if (wake_flags & WF_MIGRATED)
1581 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1582
1583#endif /* CONFIG_SMP */
1584
1585 schedstat_inc(rq, ttwu_count);
1586 schedstat_inc(p, se.statistics.nr_wakeups);
1587
1588 if (wake_flags & WF_SYNC)
1589 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1590
1591#endif /* CONFIG_SCHEDSTATS */
1592}
1593
1594static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1595{
1596 activate_task(rq, p, en_flags);
1597 p->on_rq = TASK_ON_RQ_QUEUED;
1598
1599 /* if a worker is waking up, notify workqueue */
1600 if (p->flags & PF_WQ_WORKER)
1601 wq_worker_waking_up(p, cpu_of(rq));
1602}
1603
1604/*
1605 * Mark the task runnable and perform wakeup-preemption.
1606 */
1607static void
1608ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1609{
1610 check_preempt_curr(rq, p, wake_flags);
1611 p->state = TASK_RUNNING;
1612 trace_sched_wakeup(p);
1613
1614#ifdef CONFIG_SMP
1615 if (p->sched_class->task_woken) {
1616 /*
1617 * Our task @p is fully woken up and running; so its safe to
1618 * drop the rq->lock, hereafter rq is only used for statistics.
1619 */
1620 lockdep_unpin_lock(&rq->lock);
1621 p->sched_class->task_woken(rq, p);
1622 lockdep_pin_lock(&rq->lock);
1623 }
1624
1625 if (rq->idle_stamp) {
1626 u64 delta = rq_clock(rq) - rq->idle_stamp;
1627 u64 max = 2*rq->max_idle_balance_cost;
1628
1629 update_avg(&rq->avg_idle, delta);
1630
1631 if (rq->avg_idle > max)
1632 rq->avg_idle = max;
1633
1634 rq->idle_stamp = 0;
1635 }
1636#endif
1637}
1638
1639static void
1640ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1641{
1642 lockdep_assert_held(&rq->lock);
1643
1644#ifdef CONFIG_SMP
1645 if (p->sched_contributes_to_load)
1646 rq->nr_uninterruptible--;
1647#endif
1648
1649 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1650 ttwu_do_wakeup(rq, p, wake_flags);
1651}
1652
1653/*
1654 * Called in case the task @p isn't fully descheduled from its runqueue,
1655 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1656 * since all we need to do is flip p->state to TASK_RUNNING, since
1657 * the task is still ->on_rq.
1658 */
1659static int ttwu_remote(struct task_struct *p, int wake_flags)
1660{
1661 struct rq *rq;
1662 int ret = 0;
1663
1664 rq = __task_rq_lock(p);
1665 if (task_on_rq_queued(p)) {
1666 /* check_preempt_curr() may use rq clock */
1667 update_rq_clock(rq);
1668 ttwu_do_wakeup(rq, p, wake_flags);
1669 ret = 1;
1670 }
1671 __task_rq_unlock(rq);
1672
1673 return ret;
1674}
1675
1676#ifdef CONFIG_SMP
1677void sched_ttwu_pending(void)
1678{
1679 struct rq *rq = this_rq();
1680 struct llist_node *llist = llist_del_all(&rq->wake_list);
1681 struct task_struct *p;
1682 unsigned long flags;
1683
1684 if (!llist)
1685 return;
1686
1687 raw_spin_lock_irqsave(&rq->lock, flags);
1688 lockdep_pin_lock(&rq->lock);
1689
1690 while (llist) {
1691 p = llist_entry(llist, struct task_struct, wake_entry);
1692 llist = llist_next(llist);
1693 ttwu_do_activate(rq, p, 0);
1694 }
1695
1696 lockdep_unpin_lock(&rq->lock);
1697 raw_spin_unlock_irqrestore(&rq->lock, flags);
1698}
1699
1700void scheduler_ipi(void)
1701{
1702 /*
1703 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1704 * TIF_NEED_RESCHED remotely (for the first time) will also send
1705 * this IPI.
1706 */
1707 preempt_fold_need_resched();
1708
1709 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1710 return;
1711
1712 /*
1713 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1714 * traditionally all their work was done from the interrupt return
1715 * path. Now that we actually do some work, we need to make sure
1716 * we do call them.
1717 *
1718 * Some archs already do call them, luckily irq_enter/exit nest
1719 * properly.
1720 *
1721 * Arguably we should visit all archs and update all handlers,
1722 * however a fair share of IPIs are still resched only so this would
1723 * somewhat pessimize the simple resched case.
1724 */
1725 irq_enter();
1726 sched_ttwu_pending();
1727
1728 /*
1729 * Check if someone kicked us for doing the nohz idle load balance.
1730 */
1731 if (unlikely(got_nohz_idle_kick())) {
1732 this_rq()->idle_balance = 1;
1733 raise_softirq_irqoff(SCHED_SOFTIRQ);
1734 }
1735 irq_exit();
1736}
1737
1738static void ttwu_queue_remote(struct task_struct *p, int cpu)
1739{
1740 struct rq *rq = cpu_rq(cpu);
1741
1742 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1743 if (!set_nr_if_polling(rq->idle))
1744 smp_send_reschedule(cpu);
1745 else
1746 trace_sched_wake_idle_without_ipi(cpu);
1747 }
1748}
1749
1750void wake_up_if_idle(int cpu)
1751{
1752 struct rq *rq = cpu_rq(cpu);
1753 unsigned long flags;
1754
1755 rcu_read_lock();
1756
1757 if (!is_idle_task(rcu_dereference(rq->curr)))
1758 goto out;
1759
1760 if (set_nr_if_polling(rq->idle)) {
1761 trace_sched_wake_idle_without_ipi(cpu);
1762 } else {
1763 raw_spin_lock_irqsave(&rq->lock, flags);
1764 if (is_idle_task(rq->curr))
1765 smp_send_reschedule(cpu);
1766 /* Else cpu is not in idle, do nothing here */
1767 raw_spin_unlock_irqrestore(&rq->lock, flags);
1768 }
1769
1770out:
1771 rcu_read_unlock();
1772}
1773
1774bool cpus_share_cache(int this_cpu, int that_cpu)
1775{
1776 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1777}
1778#endif /* CONFIG_SMP */
1779
1780static void ttwu_queue(struct task_struct *p, int cpu)
1781{
1782 struct rq *rq = cpu_rq(cpu);
1783
1784#if defined(CONFIG_SMP)
1785 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1786 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1787 ttwu_queue_remote(p, cpu);
1788 return;
1789 }
1790#endif
1791
1792 raw_spin_lock(&rq->lock);
1793 lockdep_pin_lock(&rq->lock);
1794 ttwu_do_activate(rq, p, 0);
1795 lockdep_unpin_lock(&rq->lock);
1796 raw_spin_unlock(&rq->lock);
1797}
1798
1799/*
1800 * Notes on Program-Order guarantees on SMP systems.
1801 *
1802 * MIGRATION
1803 *
1804 * The basic program-order guarantee on SMP systems is that when a task [t]
1805 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1806 * execution on its new cpu [c1].
1807 *
1808 * For migration (of runnable tasks) this is provided by the following means:
1809 *
1810 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1811 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1812 * rq(c1)->lock (if not at the same time, then in that order).
1813 * C) LOCK of the rq(c1)->lock scheduling in task
1814 *
1815 * Transitivity guarantees that B happens after A and C after B.
1816 * Note: we only require RCpc transitivity.
1817 * Note: the cpu doing B need not be c0 or c1
1818 *
1819 * Example:
1820 *
1821 * CPU0 CPU1 CPU2
1822 *
1823 * LOCK rq(0)->lock
1824 * sched-out X
1825 * sched-in Y
1826 * UNLOCK rq(0)->lock
1827 *
1828 * LOCK rq(0)->lock // orders against CPU0
1829 * dequeue X
1830 * UNLOCK rq(0)->lock
1831 *
1832 * LOCK rq(1)->lock
1833 * enqueue X
1834 * UNLOCK rq(1)->lock
1835 *
1836 * LOCK rq(1)->lock // orders against CPU2
1837 * sched-out Z
1838 * sched-in X
1839 * UNLOCK rq(1)->lock
1840 *
1841 *
1842 * BLOCKING -- aka. SLEEP + WAKEUP
1843 *
1844 * For blocking we (obviously) need to provide the same guarantee as for
1845 * migration. However the means are completely different as there is no lock
1846 * chain to provide order. Instead we do:
1847 *
1848 * 1) smp_store_release(X->on_cpu, 0)
1849 * 2) smp_cond_acquire(!X->on_cpu)
1850 *
1851 * Example:
1852 *
1853 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1854 *
1855 * LOCK rq(0)->lock LOCK X->pi_lock
1856 * dequeue X
1857 * sched-out X
1858 * smp_store_release(X->on_cpu, 0);
1859 *
1860 * smp_cond_acquire(!X->on_cpu);
1861 * X->state = WAKING
1862 * set_task_cpu(X,2)
1863 *
1864 * LOCK rq(2)->lock
1865 * enqueue X
1866 * X->state = RUNNING
1867 * UNLOCK rq(2)->lock
1868 *
1869 * LOCK rq(2)->lock // orders against CPU1
1870 * sched-out Z
1871 * sched-in X
1872 * UNLOCK rq(2)->lock
1873 *
1874 * UNLOCK X->pi_lock
1875 * UNLOCK rq(0)->lock
1876 *
1877 *
1878 * However; for wakeups there is a second guarantee we must provide, namely we
1879 * must observe the state that lead to our wakeup. That is, not only must our
1880 * task observe its own prior state, it must also observe the stores prior to
1881 * its wakeup.
1882 *
1883 * This means that any means of doing remote wakeups must order the CPU doing
1884 * the wakeup against the CPU the task is going to end up running on. This,
1885 * however, is already required for the regular Program-Order guarantee above,
1886 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
1887 *
1888 */
1889
1890/**
1891 * try_to_wake_up - wake up a thread
1892 * @p: the thread to be awakened
1893 * @state: the mask of task states that can be woken
1894 * @wake_flags: wake modifier flags (WF_*)
1895 *
1896 * Put it on the run-queue if it's not already there. The "current"
1897 * thread is always on the run-queue (except when the actual
1898 * re-schedule is in progress), and as such you're allowed to do
1899 * the simpler "current->state = TASK_RUNNING" to mark yourself
1900 * runnable without the overhead of this.
1901 *
1902 * Return: %true if @p was woken up, %false if it was already running.
1903 * or @state didn't match @p's state.
1904 */
1905static int
1906try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1907{
1908 unsigned long flags;
1909 int cpu, success = 0;
1910
1911 /*
1912 * If we are going to wake up a thread waiting for CONDITION we
1913 * need to ensure that CONDITION=1 done by the caller can not be
1914 * reordered with p->state check below. This pairs with mb() in
1915 * set_current_state() the waiting thread does.
1916 */
1917 smp_mb__before_spinlock();
1918 raw_spin_lock_irqsave(&p->pi_lock, flags);
1919 if (!(p->state & state))
1920 goto out;
1921
1922 trace_sched_waking(p);
1923
1924 success = 1; /* we're going to change ->state */
1925 cpu = task_cpu(p);
1926
1927 if (p->on_rq && ttwu_remote(p, wake_flags))
1928 goto stat;
1929
1930#ifdef CONFIG_SMP
1931 /*
1932 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1933 * possible to, falsely, observe p->on_cpu == 0.
1934 *
1935 * One must be running (->on_cpu == 1) in order to remove oneself
1936 * from the runqueue.
1937 *
1938 * [S] ->on_cpu = 1; [L] ->on_rq
1939 * UNLOCK rq->lock
1940 * RMB
1941 * LOCK rq->lock
1942 * [S] ->on_rq = 0; [L] ->on_cpu
1943 *
1944 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
1945 * from the consecutive calls to schedule(); the first switching to our
1946 * task, the second putting it to sleep.
1947 */
1948 smp_rmb();
1949
1950 /*
1951 * If the owning (remote) cpu is still in the middle of schedule() with
1952 * this task as prev, wait until its done referencing the task.
1953 *
1954 * Pairs with the smp_store_release() in finish_lock_switch().
1955 *
1956 * This ensures that tasks getting woken will be fully ordered against
1957 * their previous state and preserve Program Order.
1958 */
1959 smp_cond_acquire(!p->on_cpu);
1960
1961 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1962 p->state = TASK_WAKING;
1963
1964 if (p->sched_class->task_waking)
1965 p->sched_class->task_waking(p);
1966
1967 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1968 if (task_cpu(p) != cpu) {
1969 wake_flags |= WF_MIGRATED;
1970 set_task_cpu(p, cpu);
1971 }
1972#endif /* CONFIG_SMP */
1973
1974 ttwu_queue(p, cpu);
1975stat:
1976 if (schedstat_enabled())
1977 ttwu_stat(p, cpu, wake_flags);
1978out:
1979 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1980
1981 return success;
1982}
1983
1984/**
1985 * try_to_wake_up_local - try to wake up a local task with rq lock held
1986 * @p: the thread to be awakened
1987 *
1988 * Put @p on the run-queue if it's not already there. The caller must
1989 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1990 * the current task.
1991 */
1992static void try_to_wake_up_local(struct task_struct *p)
1993{
1994 struct rq *rq = task_rq(p);
1995
1996 if (WARN_ON_ONCE(rq != this_rq()) ||
1997 WARN_ON_ONCE(p == current))
1998 return;
1999
2000 lockdep_assert_held(&rq->lock);
2001
2002 if (!raw_spin_trylock(&p->pi_lock)) {
2003 /*
2004 * This is OK, because current is on_cpu, which avoids it being
2005 * picked for load-balance and preemption/IRQs are still
2006 * disabled avoiding further scheduler activity on it and we've
2007 * not yet picked a replacement task.
2008 */
2009 lockdep_unpin_lock(&rq->lock);
2010 raw_spin_unlock(&rq->lock);
2011 raw_spin_lock(&p->pi_lock);
2012 raw_spin_lock(&rq->lock);
2013 lockdep_pin_lock(&rq->lock);
2014 }
2015
2016 if (!(p->state & TASK_NORMAL))
2017 goto out;
2018
2019 trace_sched_waking(p);
2020
2021 if (!task_on_rq_queued(p))
2022 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2023
2024 ttwu_do_wakeup(rq, p, 0);
2025 if (schedstat_enabled())
2026 ttwu_stat(p, smp_processor_id(), 0);
2027out:
2028 raw_spin_unlock(&p->pi_lock);
2029}
2030
2031/**
2032 * wake_up_process - Wake up a specific process
2033 * @p: The process to be woken up.
2034 *
2035 * Attempt to wake up the nominated process and move it to the set of runnable
2036 * processes.
2037 *
2038 * Return: 1 if the process was woken up, 0 if it was already running.
2039 *
2040 * It may be assumed that this function implies a write memory barrier before
2041 * changing the task state if and only if any tasks are woken up.
2042 */
2043int wake_up_process(struct task_struct *p)
2044{
2045 return try_to_wake_up(p, TASK_NORMAL, 0);
2046}
2047EXPORT_SYMBOL(wake_up_process);
2048
2049int wake_up_state(struct task_struct *p, unsigned int state)
2050{
2051 return try_to_wake_up(p, state, 0);
2052}
2053
2054/*
2055 * This function clears the sched_dl_entity static params.
2056 */
2057void __dl_clear_params(struct task_struct *p)
2058{
2059 struct sched_dl_entity *dl_se = &p->dl;
2060
2061 dl_se->dl_runtime = 0;
2062 dl_se->dl_deadline = 0;
2063 dl_se->dl_period = 0;
2064 dl_se->flags = 0;
2065 dl_se->dl_bw = 0;
2066
2067 dl_se->dl_throttled = 0;
2068 dl_se->dl_yielded = 0;
2069}
2070
2071/*
2072 * Perform scheduler related setup for a newly forked process p.
2073 * p is forked by current.
2074 *
2075 * __sched_fork() is basic setup used by init_idle() too:
2076 */
2077static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2078{
2079 p->on_rq = 0;
2080
2081 p->se.on_rq = 0;
2082 p->se.exec_start = 0;
2083 p->se.sum_exec_runtime = 0;
2084 p->se.prev_sum_exec_runtime = 0;
2085 p->se.nr_migrations = 0;
2086 p->se.vruntime = 0;
2087 INIT_LIST_HEAD(&p->se.group_node);
2088
2089#ifdef CONFIG_FAIR_GROUP_SCHED
2090 p->se.cfs_rq = NULL;
2091#endif
2092
2093#ifdef CONFIG_SCHEDSTATS
2094 /* Even if schedstat is disabled, there should not be garbage */
2095 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2096#endif
2097
2098 RB_CLEAR_NODE(&p->dl.rb_node);
2099 init_dl_task_timer(&p->dl);
2100 __dl_clear_params(p);
2101
2102 INIT_LIST_HEAD(&p->rt.run_list);
2103 p->rt.timeout = 0;
2104 p->rt.time_slice = sched_rr_timeslice;
2105 p->rt.on_rq = 0;
2106 p->rt.on_list = 0;
2107
2108#ifdef CONFIG_PREEMPT_NOTIFIERS
2109 INIT_HLIST_HEAD(&p->preempt_notifiers);
2110#endif
2111
2112#ifdef CONFIG_NUMA_BALANCING
2113 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2114 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2115 p->mm->numa_scan_seq = 0;
2116 }
2117
2118 if (clone_flags & CLONE_VM)
2119 p->numa_preferred_nid = current->numa_preferred_nid;
2120 else
2121 p->numa_preferred_nid = -1;
2122
2123 p->node_stamp = 0ULL;
2124 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2125 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2126 p->numa_work.next = &p->numa_work;
2127 p->numa_faults = NULL;
2128 p->last_task_numa_placement = 0;
2129 p->last_sum_exec_runtime = 0;
2130
2131 p->numa_group = NULL;
2132#endif /* CONFIG_NUMA_BALANCING */
2133}
2134
2135DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2136
2137#ifdef CONFIG_NUMA_BALANCING
2138
2139void set_numabalancing_state(bool enabled)
2140{
2141 if (enabled)
2142 static_branch_enable(&sched_numa_balancing);
2143 else
2144 static_branch_disable(&sched_numa_balancing);
2145}
2146
2147#ifdef CONFIG_PROC_SYSCTL
2148int sysctl_numa_balancing(struct ctl_table *table, int write,
2149 void __user *buffer, size_t *lenp, loff_t *ppos)
2150{
2151 struct ctl_table t;
2152 int err;
2153 int state = static_branch_likely(&sched_numa_balancing);
2154
2155 if (write && !capable(CAP_SYS_ADMIN))
2156 return -EPERM;
2157
2158 t = *table;
2159 t.data = &state;
2160 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2161 if (err < 0)
2162 return err;
2163 if (write)
2164 set_numabalancing_state(state);
2165 return err;
2166}
2167#endif
2168#endif
2169
2170DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2171
2172#ifdef CONFIG_SCHEDSTATS
2173static void set_schedstats(bool enabled)
2174{
2175 if (enabled)
2176 static_branch_enable(&sched_schedstats);
2177 else
2178 static_branch_disable(&sched_schedstats);
2179}
2180
2181void force_schedstat_enabled(void)
2182{
2183 if (!schedstat_enabled()) {
2184 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2185 static_branch_enable(&sched_schedstats);
2186 }
2187}
2188
2189static int __init setup_schedstats(char *str)
2190{
2191 int ret = 0;
2192 if (!str)
2193 goto out;
2194
2195 if (!strcmp(str, "enable")) {
2196 set_schedstats(true);
2197 ret = 1;
2198 } else if (!strcmp(str, "disable")) {
2199 set_schedstats(false);
2200 ret = 1;
2201 }
2202out:
2203 if (!ret)
2204 pr_warn("Unable to parse schedstats=\n");
2205
2206 return ret;
2207}
2208__setup("schedstats=", setup_schedstats);
2209
2210#ifdef CONFIG_PROC_SYSCTL
2211int sysctl_schedstats(struct ctl_table *table, int write,
2212 void __user *buffer, size_t *lenp, loff_t *ppos)
2213{
2214 struct ctl_table t;
2215 int err;
2216 int state = static_branch_likely(&sched_schedstats);
2217
2218 if (write && !capable(CAP_SYS_ADMIN))
2219 return -EPERM;
2220
2221 t = *table;
2222 t.data = &state;
2223 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2224 if (err < 0)
2225 return err;
2226 if (write)
2227 set_schedstats(state);
2228 return err;
2229}
2230#endif
2231#endif
2232
2233/*
2234 * fork()/clone()-time setup:
2235 */
2236int sched_fork(unsigned long clone_flags, struct task_struct *p)
2237{
2238 unsigned long flags;
2239 int cpu = get_cpu();
2240
2241 __sched_fork(clone_flags, p);
2242 /*
2243 * We mark the process as running here. This guarantees that
2244 * nobody will actually run it, and a signal or other external
2245 * event cannot wake it up and insert it on the runqueue either.
2246 */
2247 p->state = TASK_RUNNING;
2248
2249 /*
2250 * Make sure we do not leak PI boosting priority to the child.
2251 */
2252 p->prio = current->normal_prio;
2253
2254 /*
2255 * Revert to default priority/policy on fork if requested.
2256 */
2257 if (unlikely(p->sched_reset_on_fork)) {
2258 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2259 p->policy = SCHED_NORMAL;
2260 p->static_prio = NICE_TO_PRIO(0);
2261 p->rt_priority = 0;
2262 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2263 p->static_prio = NICE_TO_PRIO(0);
2264
2265 p->prio = p->normal_prio = __normal_prio(p);
2266 set_load_weight(p);
2267
2268 /*
2269 * We don't need the reset flag anymore after the fork. It has
2270 * fulfilled its duty:
2271 */
2272 p->sched_reset_on_fork = 0;
2273 }
2274
2275 if (dl_prio(p->prio)) {
2276 put_cpu();
2277 return -EAGAIN;
2278 } else if (rt_prio(p->prio)) {
2279 p->sched_class = &rt_sched_class;
2280 } else {
2281 p->sched_class = &fair_sched_class;
2282 }
2283
2284 if (p->sched_class->task_fork)
2285 p->sched_class->task_fork(p);
2286
2287 /*
2288 * The child is not yet in the pid-hash so no cgroup attach races,
2289 * and the cgroup is pinned to this child due to cgroup_fork()
2290 * is ran before sched_fork().
2291 *
2292 * Silence PROVE_RCU.
2293 */
2294 raw_spin_lock_irqsave(&p->pi_lock, flags);
2295 set_task_cpu(p, cpu);
2296 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2297
2298#ifdef CONFIG_SCHED_INFO
2299 if (likely(sched_info_on()))
2300 memset(&p->sched_info, 0, sizeof(p->sched_info));
2301#endif
2302#if defined(CONFIG_SMP)
2303 p->on_cpu = 0;
2304#endif
2305 init_task_preempt_count(p);
2306#ifdef CONFIG_SMP
2307 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2308 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2309#endif
2310
2311 put_cpu();
2312 return 0;
2313}
2314
2315unsigned long to_ratio(u64 period, u64 runtime)
2316{
2317 if (runtime == RUNTIME_INF)
2318 return 1ULL << 20;
2319
2320 /*
2321 * Doing this here saves a lot of checks in all
2322 * the calling paths, and returning zero seems
2323 * safe for them anyway.
2324 */
2325 if (period == 0)
2326 return 0;
2327
2328 return div64_u64(runtime << 20, period);
2329}
2330
2331#ifdef CONFIG_SMP
2332inline struct dl_bw *dl_bw_of(int i)
2333{
2334 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2335 "sched RCU must be held");
2336 return &cpu_rq(i)->rd->dl_bw;
2337}
2338
2339static inline int dl_bw_cpus(int i)
2340{
2341 struct root_domain *rd = cpu_rq(i)->rd;
2342 int cpus = 0;
2343
2344 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2345 "sched RCU must be held");
2346 for_each_cpu_and(i, rd->span, cpu_active_mask)
2347 cpus++;
2348
2349 return cpus;
2350}
2351#else
2352inline struct dl_bw *dl_bw_of(int i)
2353{
2354 return &cpu_rq(i)->dl.dl_bw;
2355}
2356
2357static inline int dl_bw_cpus(int i)
2358{
2359 return 1;
2360}
2361#endif
2362
2363/*
2364 * We must be sure that accepting a new task (or allowing changing the
2365 * parameters of an existing one) is consistent with the bandwidth
2366 * constraints. If yes, this function also accordingly updates the currently
2367 * allocated bandwidth to reflect the new situation.
2368 *
2369 * This function is called while holding p's rq->lock.
2370 *
2371 * XXX we should delay bw change until the task's 0-lag point, see
2372 * __setparam_dl().
2373 */
2374static int dl_overflow(struct task_struct *p, int policy,
2375 const struct sched_attr *attr)
2376{
2377
2378 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2379 u64 period = attr->sched_period ?: attr->sched_deadline;
2380 u64 runtime = attr->sched_runtime;
2381 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2382 int cpus, err = -1;
2383
2384 if (new_bw == p->dl.dl_bw)
2385 return 0;
2386
2387 /*
2388 * Either if a task, enters, leave, or stays -deadline but changes
2389 * its parameters, we may need to update accordingly the total
2390 * allocated bandwidth of the container.
2391 */
2392 raw_spin_lock(&dl_b->lock);
2393 cpus = dl_bw_cpus(task_cpu(p));
2394 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2395 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2396 __dl_add(dl_b, new_bw);
2397 err = 0;
2398 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2399 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2400 __dl_clear(dl_b, p->dl.dl_bw);
2401 __dl_add(dl_b, new_bw);
2402 err = 0;
2403 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2404 __dl_clear(dl_b, p->dl.dl_bw);
2405 err = 0;
2406 }
2407 raw_spin_unlock(&dl_b->lock);
2408
2409 return err;
2410}
2411
2412extern void init_dl_bw(struct dl_bw *dl_b);
2413
2414/*
2415 * wake_up_new_task - wake up a newly created task for the first time.
2416 *
2417 * This function will do some initial scheduler statistics housekeeping
2418 * that must be done for every newly created context, then puts the task
2419 * on the runqueue and wakes it.
2420 */
2421void wake_up_new_task(struct task_struct *p)
2422{
2423 unsigned long flags;
2424 struct rq *rq;
2425
2426 raw_spin_lock_irqsave(&p->pi_lock, flags);
2427 /* Initialize new task's runnable average */
2428 init_entity_runnable_average(&p->se);
2429#ifdef CONFIG_SMP
2430 /*
2431 * Fork balancing, do it here and not earlier because:
2432 * - cpus_allowed can change in the fork path
2433 * - any previously selected cpu might disappear through hotplug
2434 */
2435 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2436#endif
2437
2438 rq = __task_rq_lock(p);
2439 activate_task(rq, p, 0);
2440 p->on_rq = TASK_ON_RQ_QUEUED;
2441 trace_sched_wakeup_new(p);
2442 check_preempt_curr(rq, p, WF_FORK);
2443#ifdef CONFIG_SMP
2444 if (p->sched_class->task_woken) {
2445 /*
2446 * Nothing relies on rq->lock after this, so its fine to
2447 * drop it.
2448 */
2449 lockdep_unpin_lock(&rq->lock);
2450 p->sched_class->task_woken(rq, p);
2451 lockdep_pin_lock(&rq->lock);
2452 }
2453#endif
2454 task_rq_unlock(rq, p, &flags);
2455}
2456
2457#ifdef CONFIG_PREEMPT_NOTIFIERS
2458
2459static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2460
2461void preempt_notifier_inc(void)
2462{
2463 static_key_slow_inc(&preempt_notifier_key);
2464}
2465EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2466
2467void preempt_notifier_dec(void)
2468{
2469 static_key_slow_dec(&preempt_notifier_key);
2470}
2471EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2472
2473/**
2474 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2475 * @notifier: notifier struct to register
2476 */
2477void preempt_notifier_register(struct preempt_notifier *notifier)
2478{
2479 if (!static_key_false(&preempt_notifier_key))
2480 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2481
2482 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2483}
2484EXPORT_SYMBOL_GPL(preempt_notifier_register);
2485
2486/**
2487 * preempt_notifier_unregister - no longer interested in preemption notifications
2488 * @notifier: notifier struct to unregister
2489 *
2490 * This is *not* safe to call from within a preemption notifier.
2491 */
2492void preempt_notifier_unregister(struct preempt_notifier *notifier)
2493{
2494 hlist_del(¬ifier->link);
2495}
2496EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2497
2498static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2499{
2500 struct preempt_notifier *notifier;
2501
2502 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2503 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2504}
2505
2506static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2507{
2508 if (static_key_false(&preempt_notifier_key))
2509 __fire_sched_in_preempt_notifiers(curr);
2510}
2511
2512static void
2513__fire_sched_out_preempt_notifiers(struct task_struct *curr,
2514 struct task_struct *next)
2515{
2516 struct preempt_notifier *notifier;
2517
2518 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2519 notifier->ops->sched_out(notifier, next);
2520}
2521
2522static __always_inline void
2523fire_sched_out_preempt_notifiers(struct task_struct *curr,
2524 struct task_struct *next)
2525{
2526 if (static_key_false(&preempt_notifier_key))
2527 __fire_sched_out_preempt_notifiers(curr, next);
2528}
2529
2530#else /* !CONFIG_PREEMPT_NOTIFIERS */
2531
2532static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2533{
2534}
2535
2536static inline void
2537fire_sched_out_preempt_notifiers(struct task_struct *curr,
2538 struct task_struct *next)
2539{
2540}
2541
2542#endif /* CONFIG_PREEMPT_NOTIFIERS */
2543
2544/**
2545 * prepare_task_switch - prepare to switch tasks
2546 * @rq: the runqueue preparing to switch
2547 * @prev: the current task that is being switched out
2548 * @next: the task we are going to switch to.
2549 *
2550 * This is called with the rq lock held and interrupts off. It must
2551 * be paired with a subsequent finish_task_switch after the context
2552 * switch.
2553 *
2554 * prepare_task_switch sets up locking and calls architecture specific
2555 * hooks.
2556 */
2557static inline void
2558prepare_task_switch(struct rq *rq, struct task_struct *prev,
2559 struct task_struct *next)
2560{
2561 sched_info_switch(rq, prev, next);
2562 perf_event_task_sched_out(prev, next);
2563 fire_sched_out_preempt_notifiers(prev, next);
2564 prepare_lock_switch(rq, next);
2565 prepare_arch_switch(next);
2566}
2567
2568/**
2569 * finish_task_switch - clean up after a task-switch
2570 * @prev: the thread we just switched away from.
2571 *
2572 * finish_task_switch must be called after the context switch, paired
2573 * with a prepare_task_switch call before the context switch.
2574 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2575 * and do any other architecture-specific cleanup actions.
2576 *
2577 * Note that we may have delayed dropping an mm in context_switch(). If
2578 * so, we finish that here outside of the runqueue lock. (Doing it
2579 * with the lock held can cause deadlocks; see schedule() for
2580 * details.)
2581 *
2582 * The context switch have flipped the stack from under us and restored the
2583 * local variables which were saved when this task called schedule() in the
2584 * past. prev == current is still correct but we need to recalculate this_rq
2585 * because prev may have moved to another CPU.
2586 */
2587static struct rq *finish_task_switch(struct task_struct *prev)
2588 __releases(rq->lock)
2589{
2590 struct rq *rq = this_rq();
2591 struct mm_struct *mm = rq->prev_mm;
2592 long prev_state;
2593
2594 /*
2595 * The previous task will have left us with a preempt_count of 2
2596 * because it left us after:
2597 *
2598 * schedule()
2599 * preempt_disable(); // 1
2600 * __schedule()
2601 * raw_spin_lock_irq(&rq->lock) // 2
2602 *
2603 * Also, see FORK_PREEMPT_COUNT.
2604 */
2605 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2606 "corrupted preempt_count: %s/%d/0x%x\n",
2607 current->comm, current->pid, preempt_count()))
2608 preempt_count_set(FORK_PREEMPT_COUNT);
2609
2610 rq->prev_mm = NULL;
2611
2612 /*
2613 * A task struct has one reference for the use as "current".
2614 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2615 * schedule one last time. The schedule call will never return, and
2616 * the scheduled task must drop that reference.
2617 *
2618 * We must observe prev->state before clearing prev->on_cpu (in
2619 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2620 * running on another CPU and we could rave with its RUNNING -> DEAD
2621 * transition, resulting in a double drop.
2622 */
2623 prev_state = prev->state;
2624 vtime_task_switch(prev);
2625 perf_event_task_sched_in(prev, current);
2626 finish_lock_switch(rq, prev);
2627 finish_arch_post_lock_switch();
2628
2629 fire_sched_in_preempt_notifiers(current);
2630 if (mm)
2631 mmdrop(mm);
2632 if (unlikely(prev_state == TASK_DEAD)) {
2633 if (prev->sched_class->task_dead)
2634 prev->sched_class->task_dead(prev);
2635
2636 /*
2637 * Remove function-return probe instances associated with this
2638 * task and put them back on the free list.
2639 */
2640 kprobe_flush_task(prev);
2641 put_task_struct(prev);
2642 }
2643
2644 tick_nohz_task_switch();
2645 return rq;
2646}
2647
2648#ifdef CONFIG_SMP
2649
2650/* rq->lock is NOT held, but preemption is disabled */
2651static void __balance_callback(struct rq *rq)
2652{
2653 struct callback_head *head, *next;
2654 void (*func)(struct rq *rq);
2655 unsigned long flags;
2656
2657 raw_spin_lock_irqsave(&rq->lock, flags);
2658 head = rq->balance_callback;
2659 rq->balance_callback = NULL;
2660 while (head) {
2661 func = (void (*)(struct rq *))head->func;
2662 next = head->next;
2663 head->next = NULL;
2664 head = next;
2665
2666 func(rq);
2667 }
2668 raw_spin_unlock_irqrestore(&rq->lock, flags);
2669}
2670
2671static inline void balance_callback(struct rq *rq)
2672{
2673 if (unlikely(rq->balance_callback))
2674 __balance_callback(rq);
2675}
2676
2677#else
2678
2679static inline void balance_callback(struct rq *rq)
2680{
2681}
2682
2683#endif
2684
2685/**
2686 * schedule_tail - first thing a freshly forked thread must call.
2687 * @prev: the thread we just switched away from.
2688 */
2689asmlinkage __visible void schedule_tail(struct task_struct *prev)
2690 __releases(rq->lock)
2691{
2692 struct rq *rq;
2693
2694 /*
2695 * New tasks start with FORK_PREEMPT_COUNT, see there and
2696 * finish_task_switch() for details.
2697 *
2698 * finish_task_switch() will drop rq->lock() and lower preempt_count
2699 * and the preempt_enable() will end up enabling preemption (on
2700 * PREEMPT_COUNT kernels).
2701 */
2702
2703 rq = finish_task_switch(prev);
2704 balance_callback(rq);
2705 preempt_enable();
2706
2707 if (current->set_child_tid)
2708 put_user(task_pid_vnr(current), current->set_child_tid);
2709}
2710
2711/*
2712 * context_switch - switch to the new MM and the new thread's register state.
2713 */
2714static __always_inline struct rq *
2715context_switch(struct rq *rq, struct task_struct *prev,
2716 struct task_struct *next)
2717{
2718 struct mm_struct *mm, *oldmm;
2719
2720 prepare_task_switch(rq, prev, next);
2721
2722 mm = next->mm;
2723 oldmm = prev->active_mm;
2724 /*
2725 * For paravirt, this is coupled with an exit in switch_to to
2726 * combine the page table reload and the switch backend into
2727 * one hypercall.
2728 */
2729 arch_start_context_switch(prev);
2730
2731 if (!mm) {
2732 next->active_mm = oldmm;
2733 atomic_inc(&oldmm->mm_count);
2734 enter_lazy_tlb(oldmm, next);
2735 } else
2736 switch_mm(oldmm, mm, next);
2737
2738 if (!prev->mm) {
2739 prev->active_mm = NULL;
2740 rq->prev_mm = oldmm;
2741 }
2742 /*
2743 * Since the runqueue lock will be released by the next
2744 * task (which is an invalid locking op but in the case
2745 * of the scheduler it's an obvious special-case), so we
2746 * do an early lockdep release here:
2747 */
2748 lockdep_unpin_lock(&rq->lock);
2749 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2750
2751 /* Here we just switch the register state and the stack. */
2752 switch_to(prev, next, prev);
2753 barrier();
2754
2755 return finish_task_switch(prev);
2756}
2757
2758/*
2759 * nr_running and nr_context_switches:
2760 *
2761 * externally visible scheduler statistics: current number of runnable
2762 * threads, total number of context switches performed since bootup.
2763 */
2764unsigned long nr_running(void)
2765{
2766 unsigned long i, sum = 0;
2767
2768 for_each_online_cpu(i)
2769 sum += cpu_rq(i)->nr_running;
2770
2771 return sum;
2772}
2773
2774/*
2775 * Check if only the current task is running on the cpu.
2776 *
2777 * Caution: this function does not check that the caller has disabled
2778 * preemption, thus the result might have a time-of-check-to-time-of-use
2779 * race. The caller is responsible to use it correctly, for example:
2780 *
2781 * - from a non-preemptable section (of course)
2782 *
2783 * - from a thread that is bound to a single CPU
2784 *
2785 * - in a loop with very short iterations (e.g. a polling loop)
2786 */
2787bool single_task_running(void)
2788{
2789 return raw_rq()->nr_running == 1;
2790}
2791EXPORT_SYMBOL(single_task_running);
2792
2793unsigned long long nr_context_switches(void)
2794{
2795 int i;
2796 unsigned long long sum = 0;
2797
2798 for_each_possible_cpu(i)
2799 sum += cpu_rq(i)->nr_switches;
2800
2801 return sum;
2802}
2803
2804unsigned long nr_iowait(void)
2805{
2806 unsigned long i, sum = 0;
2807
2808 for_each_possible_cpu(i)
2809 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2810
2811 return sum;
2812}
2813
2814unsigned long nr_iowait_cpu(int cpu)
2815{
2816 struct rq *this = cpu_rq(cpu);
2817 return atomic_read(&this->nr_iowait);
2818}
2819
2820void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2821{
2822 struct rq *rq = this_rq();
2823 *nr_waiters = atomic_read(&rq->nr_iowait);
2824 *load = rq->load.weight;
2825}
2826
2827#ifdef CONFIG_SMP
2828
2829/*
2830 * sched_exec - execve() is a valuable balancing opportunity, because at
2831 * this point the task has the smallest effective memory and cache footprint.
2832 */
2833void sched_exec(void)
2834{
2835 struct task_struct *p = current;
2836 unsigned long flags;
2837 int dest_cpu;
2838
2839 raw_spin_lock_irqsave(&p->pi_lock, flags);
2840 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2841 if (dest_cpu == smp_processor_id())
2842 goto unlock;
2843
2844 if (likely(cpu_active(dest_cpu))) {
2845 struct migration_arg arg = { p, dest_cpu };
2846
2847 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2848 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2849 return;
2850 }
2851unlock:
2852 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2853}
2854
2855#endif
2856
2857DEFINE_PER_CPU(struct kernel_stat, kstat);
2858DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2859
2860EXPORT_PER_CPU_SYMBOL(kstat);
2861EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2862
2863/*
2864 * Return accounted runtime for the task.
2865 * In case the task is currently running, return the runtime plus current's
2866 * pending runtime that have not been accounted yet.
2867 */
2868unsigned long long task_sched_runtime(struct task_struct *p)
2869{
2870 unsigned long flags;
2871 struct rq *rq;
2872 u64 ns;
2873
2874#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2875 /*
2876 * 64-bit doesn't need locks to atomically read a 64bit value.
2877 * So we have a optimization chance when the task's delta_exec is 0.
2878 * Reading ->on_cpu is racy, but this is ok.
2879 *
2880 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2881 * If we race with it entering cpu, unaccounted time is 0. This is
2882 * indistinguishable from the read occurring a few cycles earlier.
2883 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2884 * been accounted, so we're correct here as well.
2885 */
2886 if (!p->on_cpu || !task_on_rq_queued(p))
2887 return p->se.sum_exec_runtime;
2888#endif
2889
2890 rq = task_rq_lock(p, &flags);
2891 /*
2892 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2893 * project cycles that may never be accounted to this
2894 * thread, breaking clock_gettime().
2895 */
2896 if (task_current(rq, p) && task_on_rq_queued(p)) {
2897 update_rq_clock(rq);
2898 p->sched_class->update_curr(rq);
2899 }
2900 ns = p->se.sum_exec_runtime;
2901 task_rq_unlock(rq, p, &flags);
2902
2903 return ns;
2904}
2905
2906/*
2907 * This function gets called by the timer code, with HZ frequency.
2908 * We call it with interrupts disabled.
2909 */
2910void scheduler_tick(void)
2911{
2912 int cpu = smp_processor_id();
2913 struct rq *rq = cpu_rq(cpu);
2914 struct task_struct *curr = rq->curr;
2915
2916 sched_clock_tick();
2917
2918 raw_spin_lock(&rq->lock);
2919 update_rq_clock(rq);
2920 curr->sched_class->task_tick(rq, curr, 0);
2921 update_cpu_load_active(rq);
2922 calc_global_load_tick(rq);
2923 raw_spin_unlock(&rq->lock);
2924
2925 perf_event_task_tick();
2926
2927#ifdef CONFIG_SMP
2928 rq->idle_balance = idle_cpu(cpu);
2929 trigger_load_balance(rq);
2930#endif
2931 rq_last_tick_reset(rq);
2932}
2933
2934#ifdef CONFIG_NO_HZ_FULL
2935/**
2936 * scheduler_tick_max_deferment
2937 *
2938 * Keep at least one tick per second when a single
2939 * active task is running because the scheduler doesn't
2940 * yet completely support full dynticks environment.
2941 *
2942 * This makes sure that uptime, CFS vruntime, load
2943 * balancing, etc... continue to move forward, even
2944 * with a very low granularity.
2945 *
2946 * Return: Maximum deferment in nanoseconds.
2947 */
2948u64 scheduler_tick_max_deferment(void)
2949{
2950 struct rq *rq = this_rq();
2951 unsigned long next, now = READ_ONCE(jiffies);
2952
2953 next = rq->last_sched_tick + HZ;
2954
2955 if (time_before_eq(next, now))
2956 return 0;
2957
2958 return jiffies_to_nsecs(next - now);
2959}
2960#endif
2961
2962#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2963 defined(CONFIG_PREEMPT_TRACER))
2964
2965void preempt_count_add(int val)
2966{
2967#ifdef CONFIG_DEBUG_PREEMPT
2968 /*
2969 * Underflow?
2970 */
2971 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2972 return;
2973#endif
2974 __preempt_count_add(val);
2975#ifdef CONFIG_DEBUG_PREEMPT
2976 /*
2977 * Spinlock count overflowing soon?
2978 */
2979 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2980 PREEMPT_MASK - 10);
2981#endif
2982 if (preempt_count() == val) {
2983 unsigned long ip = get_lock_parent_ip();
2984#ifdef CONFIG_DEBUG_PREEMPT
2985 current->preempt_disable_ip = ip;
2986#endif
2987 trace_preempt_off(CALLER_ADDR0, ip);
2988 }
2989}
2990EXPORT_SYMBOL(preempt_count_add);
2991NOKPROBE_SYMBOL(preempt_count_add);
2992
2993void preempt_count_sub(int val)
2994{
2995#ifdef CONFIG_DEBUG_PREEMPT
2996 /*
2997 * Underflow?
2998 */
2999 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3000 return;
3001 /*
3002 * Is the spinlock portion underflowing?
3003 */
3004 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3005 !(preempt_count() & PREEMPT_MASK)))
3006 return;
3007#endif
3008
3009 if (preempt_count() == val)
3010 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3011 __preempt_count_sub(val);
3012}
3013EXPORT_SYMBOL(preempt_count_sub);
3014NOKPROBE_SYMBOL(preempt_count_sub);
3015
3016#endif
3017
3018/*
3019 * Print scheduling while atomic bug:
3020 */
3021static noinline void __schedule_bug(struct task_struct *prev)
3022{
3023 if (oops_in_progress)
3024 return;
3025
3026 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3027 prev->comm, prev->pid, preempt_count());
3028
3029 debug_show_held_locks(prev);
3030 print_modules();
3031 if (irqs_disabled())
3032 print_irqtrace_events(prev);
3033#ifdef CONFIG_DEBUG_PREEMPT
3034 if (in_atomic_preempt_off()) {
3035 pr_err("Preemption disabled at:");
3036 print_ip_sym(current->preempt_disable_ip);
3037 pr_cont("\n");
3038 }
3039#endif
3040 dump_stack();
3041 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3042}
3043
3044/*
3045 * Various schedule()-time debugging checks and statistics:
3046 */
3047static inline void schedule_debug(struct task_struct *prev)
3048{
3049#ifdef CONFIG_SCHED_STACK_END_CHECK
3050 BUG_ON(task_stack_end_corrupted(prev));
3051#endif
3052
3053 if (unlikely(in_atomic_preempt_off())) {
3054 __schedule_bug(prev);
3055 preempt_count_set(PREEMPT_DISABLED);
3056 }
3057 rcu_sleep_check();
3058
3059 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3060
3061 schedstat_inc(this_rq(), sched_count);
3062}
3063
3064/*
3065 * Pick up the highest-prio task:
3066 */
3067static inline struct task_struct *
3068pick_next_task(struct rq *rq, struct task_struct *prev)
3069{
3070 const struct sched_class *class = &fair_sched_class;
3071 struct task_struct *p;
3072
3073 /*
3074 * Optimization: we know that if all tasks are in
3075 * the fair class we can call that function directly:
3076 */
3077 if (likely(prev->sched_class == class &&
3078 rq->nr_running == rq->cfs.h_nr_running)) {
3079 p = fair_sched_class.pick_next_task(rq, prev);
3080 if (unlikely(p == RETRY_TASK))
3081 goto again;
3082
3083 /* assumes fair_sched_class->next == idle_sched_class */
3084 if (unlikely(!p))
3085 p = idle_sched_class.pick_next_task(rq, prev);
3086
3087 return p;
3088 }
3089
3090again:
3091 for_each_class(class) {
3092 p = class->pick_next_task(rq, prev);
3093 if (p) {
3094 if (unlikely(p == RETRY_TASK))
3095 goto again;
3096 return p;
3097 }
3098 }
3099
3100 BUG(); /* the idle class will always have a runnable task */
3101}
3102
3103/*
3104 * __schedule() is the main scheduler function.
3105 *
3106 * The main means of driving the scheduler and thus entering this function are:
3107 *
3108 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3109 *
3110 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3111 * paths. For example, see arch/x86/entry_64.S.
3112 *
3113 * To drive preemption between tasks, the scheduler sets the flag in timer
3114 * interrupt handler scheduler_tick().
3115 *
3116 * 3. Wakeups don't really cause entry into schedule(). They add a
3117 * task to the run-queue and that's it.
3118 *
3119 * Now, if the new task added to the run-queue preempts the current
3120 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3121 * called on the nearest possible occasion:
3122 *
3123 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3124 *
3125 * - in syscall or exception context, at the next outmost
3126 * preempt_enable(). (this might be as soon as the wake_up()'s
3127 * spin_unlock()!)
3128 *
3129 * - in IRQ context, return from interrupt-handler to
3130 * preemptible context
3131 *
3132 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3133 * then at the next:
3134 *
3135 * - cond_resched() call
3136 * - explicit schedule() call
3137 * - return from syscall or exception to user-space
3138 * - return from interrupt-handler to user-space
3139 *
3140 * WARNING: must be called with preemption disabled!
3141 */
3142static void __sched notrace __schedule(bool preempt)
3143{
3144 struct task_struct *prev, *next;
3145 unsigned long *switch_count;
3146 struct rq *rq;
3147 int cpu;
3148
3149 cpu = smp_processor_id();
3150 rq = cpu_rq(cpu);
3151 prev = rq->curr;
3152
3153 /*
3154 * do_exit() calls schedule() with preemption disabled as an exception;
3155 * however we must fix that up, otherwise the next task will see an
3156 * inconsistent (higher) preempt count.
3157 *
3158 * It also avoids the below schedule_debug() test from complaining
3159 * about this.
3160 */
3161 if (unlikely(prev->state == TASK_DEAD))
3162 preempt_enable_no_resched_notrace();
3163
3164 schedule_debug(prev);
3165
3166 if (sched_feat(HRTICK))
3167 hrtick_clear(rq);
3168
3169 local_irq_disable();
3170 rcu_note_context_switch();
3171
3172 /*
3173 * Make sure that signal_pending_state()->signal_pending() below
3174 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3175 * done by the caller to avoid the race with signal_wake_up().
3176 */
3177 smp_mb__before_spinlock();
3178 raw_spin_lock(&rq->lock);
3179 lockdep_pin_lock(&rq->lock);
3180
3181 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3182
3183 switch_count = &prev->nivcsw;
3184 if (!preempt && prev->state) {
3185 if (unlikely(signal_pending_state(prev->state, prev))) {
3186 prev->state = TASK_RUNNING;
3187 } else {
3188 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3189 prev->on_rq = 0;
3190
3191 /*
3192 * If a worker went to sleep, notify and ask workqueue
3193 * whether it wants to wake up a task to maintain
3194 * concurrency.
3195 */
3196 if (prev->flags & PF_WQ_WORKER) {
3197 struct task_struct *to_wakeup;
3198
3199 to_wakeup = wq_worker_sleeping(prev);
3200 if (to_wakeup)
3201 try_to_wake_up_local(to_wakeup);
3202 }
3203 }
3204 switch_count = &prev->nvcsw;
3205 }
3206
3207 if (task_on_rq_queued(prev))
3208 update_rq_clock(rq);
3209
3210 next = pick_next_task(rq, prev);
3211 clear_tsk_need_resched(prev);
3212 clear_preempt_need_resched();
3213 rq->clock_skip_update = 0;
3214
3215 if (likely(prev != next)) {
3216 rq->nr_switches++;
3217 rq->curr = next;
3218 ++*switch_count;
3219
3220 trace_sched_switch(preempt, prev, next);
3221 rq = context_switch(rq, prev, next); /* unlocks the rq */
3222 } else {
3223 lockdep_unpin_lock(&rq->lock);
3224 raw_spin_unlock_irq(&rq->lock);
3225 }
3226
3227 balance_callback(rq);
3228}
3229STACK_FRAME_NON_STANDARD(__schedule); /* switch_to() */
3230
3231static inline void sched_submit_work(struct task_struct *tsk)
3232{
3233 if (!tsk->state || tsk_is_pi_blocked(tsk))
3234 return;
3235 /*
3236 * If we are going to sleep and we have plugged IO queued,
3237 * make sure to submit it to avoid deadlocks.
3238 */
3239 if (blk_needs_flush_plug(tsk))
3240 blk_schedule_flush_plug(tsk);
3241}
3242
3243asmlinkage __visible void __sched schedule(void)
3244{
3245 struct task_struct *tsk = current;
3246
3247 sched_submit_work(tsk);
3248 do {
3249 preempt_disable();
3250 __schedule(false);
3251 sched_preempt_enable_no_resched();
3252 } while (need_resched());
3253}
3254EXPORT_SYMBOL(schedule);
3255
3256#ifdef CONFIG_CONTEXT_TRACKING
3257asmlinkage __visible void __sched schedule_user(void)
3258{
3259 /*
3260 * If we come here after a random call to set_need_resched(),
3261 * or we have been woken up remotely but the IPI has not yet arrived,
3262 * we haven't yet exited the RCU idle mode. Do it here manually until
3263 * we find a better solution.
3264 *
3265 * NB: There are buggy callers of this function. Ideally we
3266 * should warn if prev_state != CONTEXT_USER, but that will trigger
3267 * too frequently to make sense yet.
3268 */
3269 enum ctx_state prev_state = exception_enter();
3270 schedule();
3271 exception_exit(prev_state);
3272}
3273#endif
3274
3275/**
3276 * schedule_preempt_disabled - called with preemption disabled
3277 *
3278 * Returns with preemption disabled. Note: preempt_count must be 1
3279 */
3280void __sched schedule_preempt_disabled(void)
3281{
3282 sched_preempt_enable_no_resched();
3283 schedule();
3284 preempt_disable();
3285}
3286
3287static void __sched notrace preempt_schedule_common(void)
3288{
3289 do {
3290 preempt_disable_notrace();
3291 __schedule(true);
3292 preempt_enable_no_resched_notrace();
3293
3294 /*
3295 * Check again in case we missed a preemption opportunity
3296 * between schedule and now.
3297 */
3298 } while (need_resched());
3299}
3300
3301#ifdef CONFIG_PREEMPT
3302/*
3303 * this is the entry point to schedule() from in-kernel preemption
3304 * off of preempt_enable. Kernel preemptions off return from interrupt
3305 * occur there and call schedule directly.
3306 */
3307asmlinkage __visible void __sched notrace preempt_schedule(void)
3308{
3309 /*
3310 * If there is a non-zero preempt_count or interrupts are disabled,
3311 * we do not want to preempt the current task. Just return..
3312 */
3313 if (likely(!preemptible()))
3314 return;
3315
3316 preempt_schedule_common();
3317}
3318NOKPROBE_SYMBOL(preempt_schedule);
3319EXPORT_SYMBOL(preempt_schedule);
3320
3321/**
3322 * preempt_schedule_notrace - preempt_schedule called by tracing
3323 *
3324 * The tracing infrastructure uses preempt_enable_notrace to prevent
3325 * recursion and tracing preempt enabling caused by the tracing
3326 * infrastructure itself. But as tracing can happen in areas coming
3327 * from userspace or just about to enter userspace, a preempt enable
3328 * can occur before user_exit() is called. This will cause the scheduler
3329 * to be called when the system is still in usermode.
3330 *
3331 * To prevent this, the preempt_enable_notrace will use this function
3332 * instead of preempt_schedule() to exit user context if needed before
3333 * calling the scheduler.
3334 */
3335asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3336{
3337 enum ctx_state prev_ctx;
3338
3339 if (likely(!preemptible()))
3340 return;
3341
3342 do {
3343 preempt_disable_notrace();
3344 /*
3345 * Needs preempt disabled in case user_exit() is traced
3346 * and the tracer calls preempt_enable_notrace() causing
3347 * an infinite recursion.
3348 */
3349 prev_ctx = exception_enter();
3350 __schedule(true);
3351 exception_exit(prev_ctx);
3352
3353 preempt_enable_no_resched_notrace();
3354 } while (need_resched());
3355}
3356EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3357
3358#endif /* CONFIG_PREEMPT */
3359
3360/*
3361 * this is the entry point to schedule() from kernel preemption
3362 * off of irq context.
3363 * Note, that this is called and return with irqs disabled. This will
3364 * protect us against recursive calling from irq.
3365 */
3366asmlinkage __visible void __sched preempt_schedule_irq(void)
3367{
3368 enum ctx_state prev_state;
3369
3370 /* Catch callers which need to be fixed */
3371 BUG_ON(preempt_count() || !irqs_disabled());
3372
3373 prev_state = exception_enter();
3374
3375 do {
3376 preempt_disable();
3377 local_irq_enable();
3378 __schedule(true);
3379 local_irq_disable();
3380 sched_preempt_enable_no_resched();
3381 } while (need_resched());
3382
3383 exception_exit(prev_state);
3384}
3385
3386int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3387 void *key)
3388{
3389 return try_to_wake_up(curr->private, mode, wake_flags);
3390}
3391EXPORT_SYMBOL(default_wake_function);
3392
3393#ifdef CONFIG_RT_MUTEXES
3394
3395/*
3396 * rt_mutex_setprio - set the current priority of a task
3397 * @p: task
3398 * @prio: prio value (kernel-internal form)
3399 *
3400 * This function changes the 'effective' priority of a task. It does
3401 * not touch ->normal_prio like __setscheduler().
3402 *
3403 * Used by the rt_mutex code to implement priority inheritance
3404 * logic. Call site only calls if the priority of the task changed.
3405 */
3406void rt_mutex_setprio(struct task_struct *p, int prio)
3407{
3408 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3409 struct rq *rq;
3410 const struct sched_class *prev_class;
3411
3412 BUG_ON(prio > MAX_PRIO);
3413
3414 rq = __task_rq_lock(p);
3415
3416 /*
3417 * Idle task boosting is a nono in general. There is one
3418 * exception, when PREEMPT_RT and NOHZ is active:
3419 *
3420 * The idle task calls get_next_timer_interrupt() and holds
3421 * the timer wheel base->lock on the CPU and another CPU wants
3422 * to access the timer (probably to cancel it). We can safely
3423 * ignore the boosting request, as the idle CPU runs this code
3424 * with interrupts disabled and will complete the lock
3425 * protected section without being interrupted. So there is no
3426 * real need to boost.
3427 */
3428 if (unlikely(p == rq->idle)) {
3429 WARN_ON(p != rq->curr);
3430 WARN_ON(p->pi_blocked_on);
3431 goto out_unlock;
3432 }
3433
3434 trace_sched_pi_setprio(p, prio);
3435 oldprio = p->prio;
3436
3437 if (oldprio == prio)
3438 queue_flag &= ~DEQUEUE_MOVE;
3439
3440 prev_class = p->sched_class;
3441 queued = task_on_rq_queued(p);
3442 running = task_current(rq, p);
3443 if (queued)
3444 dequeue_task(rq, p, queue_flag);
3445 if (running)
3446 put_prev_task(rq, p);
3447
3448 /*
3449 * Boosting condition are:
3450 * 1. -rt task is running and holds mutex A
3451 * --> -dl task blocks on mutex A
3452 *
3453 * 2. -dl task is running and holds mutex A
3454 * --> -dl task blocks on mutex A and could preempt the
3455 * running task
3456 */
3457 if (dl_prio(prio)) {
3458 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3459 if (!dl_prio(p->normal_prio) ||
3460 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3461 p->dl.dl_boosted = 1;
3462 queue_flag |= ENQUEUE_REPLENISH;
3463 } else
3464 p->dl.dl_boosted = 0;
3465 p->sched_class = &dl_sched_class;
3466 } else if (rt_prio(prio)) {
3467 if (dl_prio(oldprio))
3468 p->dl.dl_boosted = 0;
3469 if (oldprio < prio)
3470 queue_flag |= ENQUEUE_HEAD;
3471 p->sched_class = &rt_sched_class;
3472 } else {
3473 if (dl_prio(oldprio))
3474 p->dl.dl_boosted = 0;
3475 if (rt_prio(oldprio))
3476 p->rt.timeout = 0;
3477 p->sched_class = &fair_sched_class;
3478 }
3479
3480 p->prio = prio;
3481
3482 if (running)
3483 p->sched_class->set_curr_task(rq);
3484 if (queued)
3485 enqueue_task(rq, p, queue_flag);
3486
3487 check_class_changed(rq, p, prev_class, oldprio);
3488out_unlock:
3489 preempt_disable(); /* avoid rq from going away on us */
3490 __task_rq_unlock(rq);
3491
3492 balance_callback(rq);
3493 preempt_enable();
3494}
3495#endif
3496
3497void set_user_nice(struct task_struct *p, long nice)
3498{
3499 int old_prio, delta, queued;
3500 unsigned long flags;
3501 struct rq *rq;
3502
3503 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3504 return;
3505 /*
3506 * We have to be careful, if called from sys_setpriority(),
3507 * the task might be in the middle of scheduling on another CPU.
3508 */
3509 rq = task_rq_lock(p, &flags);
3510 /*
3511 * The RT priorities are set via sched_setscheduler(), but we still
3512 * allow the 'normal' nice value to be set - but as expected
3513 * it wont have any effect on scheduling until the task is
3514 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3515 */
3516 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3517 p->static_prio = NICE_TO_PRIO(nice);
3518 goto out_unlock;
3519 }
3520 queued = task_on_rq_queued(p);
3521 if (queued)
3522 dequeue_task(rq, p, DEQUEUE_SAVE);
3523
3524 p->static_prio = NICE_TO_PRIO(nice);
3525 set_load_weight(p);
3526 old_prio = p->prio;
3527 p->prio = effective_prio(p);
3528 delta = p->prio - old_prio;
3529
3530 if (queued) {
3531 enqueue_task(rq, p, ENQUEUE_RESTORE);
3532 /*
3533 * If the task increased its priority or is running and
3534 * lowered its priority, then reschedule its CPU:
3535 */
3536 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3537 resched_curr(rq);
3538 }
3539out_unlock:
3540 task_rq_unlock(rq, p, &flags);
3541}
3542EXPORT_SYMBOL(set_user_nice);
3543
3544/*
3545 * can_nice - check if a task can reduce its nice value
3546 * @p: task
3547 * @nice: nice value
3548 */
3549int can_nice(const struct task_struct *p, const int nice)
3550{
3551 /* convert nice value [19,-20] to rlimit style value [1,40] */
3552 int nice_rlim = nice_to_rlimit(nice);
3553
3554 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3555 capable(CAP_SYS_NICE));
3556}
3557
3558#ifdef __ARCH_WANT_SYS_NICE
3559
3560/*
3561 * sys_nice - change the priority of the current process.
3562 * @increment: priority increment
3563 *
3564 * sys_setpriority is a more generic, but much slower function that
3565 * does similar things.
3566 */
3567SYSCALL_DEFINE1(nice, int, increment)
3568{
3569 long nice, retval;
3570
3571 /*
3572 * Setpriority might change our priority at the same moment.
3573 * We don't have to worry. Conceptually one call occurs first
3574 * and we have a single winner.
3575 */
3576 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3577 nice = task_nice(current) + increment;
3578
3579 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3580 if (increment < 0 && !can_nice(current, nice))
3581 return -EPERM;
3582
3583 retval = security_task_setnice(current, nice);
3584 if (retval)
3585 return retval;
3586
3587 set_user_nice(current, nice);
3588 return 0;
3589}
3590
3591#endif
3592
3593/**
3594 * task_prio - return the priority value of a given task.
3595 * @p: the task in question.
3596 *
3597 * Return: The priority value as seen by users in /proc.
3598 * RT tasks are offset by -200. Normal tasks are centered
3599 * around 0, value goes from -16 to +15.
3600 */
3601int task_prio(const struct task_struct *p)
3602{
3603 return p->prio - MAX_RT_PRIO;
3604}
3605
3606/**
3607 * idle_cpu - is a given cpu idle currently?
3608 * @cpu: the processor in question.
3609 *
3610 * Return: 1 if the CPU is currently idle. 0 otherwise.
3611 */
3612int idle_cpu(int cpu)
3613{
3614 struct rq *rq = cpu_rq(cpu);
3615
3616 if (rq->curr != rq->idle)
3617 return 0;
3618
3619 if (rq->nr_running)
3620 return 0;
3621
3622#ifdef CONFIG_SMP
3623 if (!llist_empty(&rq->wake_list))
3624 return 0;
3625#endif
3626
3627 return 1;
3628}
3629
3630/**
3631 * idle_task - return the idle task for a given cpu.
3632 * @cpu: the processor in question.
3633 *
3634 * Return: The idle task for the cpu @cpu.
3635 */
3636struct task_struct *idle_task(int cpu)
3637{
3638 return cpu_rq(cpu)->idle;
3639}
3640
3641/**
3642 * find_process_by_pid - find a process with a matching PID value.
3643 * @pid: the pid in question.
3644 *
3645 * The task of @pid, if found. %NULL otherwise.
3646 */
3647static struct task_struct *find_process_by_pid(pid_t pid)
3648{
3649 return pid ? find_task_by_vpid(pid) : current;
3650}
3651
3652/*
3653 * This function initializes the sched_dl_entity of a newly becoming
3654 * SCHED_DEADLINE task.
3655 *
3656 * Only the static values are considered here, the actual runtime and the
3657 * absolute deadline will be properly calculated when the task is enqueued
3658 * for the first time with its new policy.
3659 */
3660static void
3661__setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3662{
3663 struct sched_dl_entity *dl_se = &p->dl;
3664
3665 dl_se->dl_runtime = attr->sched_runtime;
3666 dl_se->dl_deadline = attr->sched_deadline;
3667 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3668 dl_se->flags = attr->sched_flags;
3669 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3670
3671 /*
3672 * Changing the parameters of a task is 'tricky' and we're not doing
3673 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3674 *
3675 * What we SHOULD do is delay the bandwidth release until the 0-lag
3676 * point. This would include retaining the task_struct until that time
3677 * and change dl_overflow() to not immediately decrement the current
3678 * amount.
3679 *
3680 * Instead we retain the current runtime/deadline and let the new
3681 * parameters take effect after the current reservation period lapses.
3682 * This is safe (albeit pessimistic) because the 0-lag point is always
3683 * before the current scheduling deadline.
3684 *
3685 * We can still have temporary overloads because we do not delay the
3686 * change in bandwidth until that time; so admission control is
3687 * not on the safe side. It does however guarantee tasks will never
3688 * consume more than promised.
3689 */
3690}
3691
3692/*
3693 * sched_setparam() passes in -1 for its policy, to let the functions
3694 * it calls know not to change it.
3695 */
3696#define SETPARAM_POLICY -1
3697
3698static void __setscheduler_params(struct task_struct *p,
3699 const struct sched_attr *attr)
3700{
3701 int policy = attr->sched_policy;
3702
3703 if (policy == SETPARAM_POLICY)
3704 policy = p->policy;
3705
3706 p->policy = policy;
3707
3708 if (dl_policy(policy))
3709 __setparam_dl(p, attr);
3710 else if (fair_policy(policy))
3711 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3712
3713 /*
3714 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3715 * !rt_policy. Always setting this ensures that things like
3716 * getparam()/getattr() don't report silly values for !rt tasks.
3717 */
3718 p->rt_priority = attr->sched_priority;
3719 p->normal_prio = normal_prio(p);
3720 set_load_weight(p);
3721}
3722
3723/* Actually do priority change: must hold pi & rq lock. */
3724static void __setscheduler(struct rq *rq, struct task_struct *p,
3725 const struct sched_attr *attr, bool keep_boost)
3726{
3727 __setscheduler_params(p, attr);
3728
3729 /*
3730 * Keep a potential priority boosting if called from
3731 * sched_setscheduler().
3732 */
3733 if (keep_boost)
3734 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3735 else
3736 p->prio = normal_prio(p);
3737
3738 if (dl_prio(p->prio))
3739 p->sched_class = &dl_sched_class;
3740 else if (rt_prio(p->prio))
3741 p->sched_class = &rt_sched_class;
3742 else
3743 p->sched_class = &fair_sched_class;
3744}
3745
3746static void
3747__getparam_dl(struct task_struct *p, struct sched_attr *attr)
3748{
3749 struct sched_dl_entity *dl_se = &p->dl;
3750
3751 attr->sched_priority = p->rt_priority;
3752 attr->sched_runtime = dl_se->dl_runtime;
3753 attr->sched_deadline = dl_se->dl_deadline;
3754 attr->sched_period = dl_se->dl_period;
3755 attr->sched_flags = dl_se->flags;
3756}
3757
3758/*
3759 * This function validates the new parameters of a -deadline task.
3760 * We ask for the deadline not being zero, and greater or equal
3761 * than the runtime, as well as the period of being zero or
3762 * greater than deadline. Furthermore, we have to be sure that
3763 * user parameters are above the internal resolution of 1us (we
3764 * check sched_runtime only since it is always the smaller one) and
3765 * below 2^63 ns (we have to check both sched_deadline and
3766 * sched_period, as the latter can be zero).
3767 */
3768static bool
3769__checkparam_dl(const struct sched_attr *attr)
3770{
3771 /* deadline != 0 */
3772 if (attr->sched_deadline == 0)
3773 return false;
3774
3775 /*
3776 * Since we truncate DL_SCALE bits, make sure we're at least
3777 * that big.
3778 */
3779 if (attr->sched_runtime < (1ULL << DL_SCALE))
3780 return false;
3781
3782 /*
3783 * Since we use the MSB for wrap-around and sign issues, make
3784 * sure it's not set (mind that period can be equal to zero).
3785 */
3786 if (attr->sched_deadline & (1ULL << 63) ||
3787 attr->sched_period & (1ULL << 63))
3788 return false;
3789
3790 /* runtime <= deadline <= period (if period != 0) */
3791 if ((attr->sched_period != 0 &&
3792 attr->sched_period < attr->sched_deadline) ||
3793 attr->sched_deadline < attr->sched_runtime)
3794 return false;
3795
3796 return true;
3797}
3798
3799/*
3800 * check the target process has a UID that matches the current process's
3801 */
3802static bool check_same_owner(struct task_struct *p)
3803{
3804 const struct cred *cred = current_cred(), *pcred;
3805 bool match;
3806
3807 rcu_read_lock();
3808 pcred = __task_cred(p);
3809 match = (uid_eq(cred->euid, pcred->euid) ||
3810 uid_eq(cred->euid, pcred->uid));
3811 rcu_read_unlock();
3812 return match;
3813}
3814
3815static bool dl_param_changed(struct task_struct *p,
3816 const struct sched_attr *attr)
3817{
3818 struct sched_dl_entity *dl_se = &p->dl;
3819
3820 if (dl_se->dl_runtime != attr->sched_runtime ||
3821 dl_se->dl_deadline != attr->sched_deadline ||
3822 dl_se->dl_period != attr->sched_period ||
3823 dl_se->flags != attr->sched_flags)
3824 return true;
3825
3826 return false;
3827}
3828
3829static int __sched_setscheduler(struct task_struct *p,
3830 const struct sched_attr *attr,
3831 bool user, bool pi)
3832{
3833 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3834 MAX_RT_PRIO - 1 - attr->sched_priority;
3835 int retval, oldprio, oldpolicy = -1, queued, running;
3836 int new_effective_prio, policy = attr->sched_policy;
3837 unsigned long flags;
3838 const struct sched_class *prev_class;
3839 struct rq *rq;
3840 int reset_on_fork;
3841 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
3842
3843 /* may grab non-irq protected spin_locks */
3844 BUG_ON(in_interrupt());
3845recheck:
3846 /* double check policy once rq lock held */
3847 if (policy < 0) {
3848 reset_on_fork = p->sched_reset_on_fork;
3849 policy = oldpolicy = p->policy;
3850 } else {
3851 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3852
3853 if (!valid_policy(policy))
3854 return -EINVAL;
3855 }
3856
3857 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3858 return -EINVAL;
3859
3860 /*
3861 * Valid priorities for SCHED_FIFO and SCHED_RR are
3862 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3863 * SCHED_BATCH and SCHED_IDLE is 0.
3864 */
3865 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3866 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3867 return -EINVAL;
3868 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3869 (rt_policy(policy) != (attr->sched_priority != 0)))
3870 return -EINVAL;
3871
3872 /*
3873 * Allow unprivileged RT tasks to decrease priority:
3874 */
3875 if (user && !capable(CAP_SYS_NICE)) {
3876 if (fair_policy(policy)) {
3877 if (attr->sched_nice < task_nice(p) &&
3878 !can_nice(p, attr->sched_nice))
3879 return -EPERM;
3880 }
3881
3882 if (rt_policy(policy)) {
3883 unsigned long rlim_rtprio =
3884 task_rlimit(p, RLIMIT_RTPRIO);
3885
3886 /* can't set/change the rt policy */
3887 if (policy != p->policy && !rlim_rtprio)
3888 return -EPERM;
3889
3890 /* can't increase priority */
3891 if (attr->sched_priority > p->rt_priority &&
3892 attr->sched_priority > rlim_rtprio)
3893 return -EPERM;
3894 }
3895
3896 /*
3897 * Can't set/change SCHED_DEADLINE policy at all for now
3898 * (safest behavior); in the future we would like to allow
3899 * unprivileged DL tasks to increase their relative deadline
3900 * or reduce their runtime (both ways reducing utilization)
3901 */
3902 if (dl_policy(policy))
3903 return -EPERM;
3904
3905 /*
3906 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3907 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3908 */
3909 if (idle_policy(p->policy) && !idle_policy(policy)) {
3910 if (!can_nice(p, task_nice(p)))
3911 return -EPERM;
3912 }
3913
3914 /* can't change other user's priorities */
3915 if (!check_same_owner(p))
3916 return -EPERM;
3917
3918 /* Normal users shall not reset the sched_reset_on_fork flag */
3919 if (p->sched_reset_on_fork && !reset_on_fork)
3920 return -EPERM;
3921 }
3922
3923 if (user) {
3924 retval = security_task_setscheduler(p);
3925 if (retval)
3926 return retval;
3927 }
3928
3929 /*
3930 * make sure no PI-waiters arrive (or leave) while we are
3931 * changing the priority of the task:
3932 *
3933 * To be able to change p->policy safely, the appropriate
3934 * runqueue lock must be held.
3935 */
3936 rq = task_rq_lock(p, &flags);
3937
3938 /*
3939 * Changing the policy of the stop threads its a very bad idea
3940 */
3941 if (p == rq->stop) {
3942 task_rq_unlock(rq, p, &flags);
3943 return -EINVAL;
3944 }
3945
3946 /*
3947 * If not changing anything there's no need to proceed further,
3948 * but store a possible modification of reset_on_fork.
3949 */
3950 if (unlikely(policy == p->policy)) {
3951 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3952 goto change;
3953 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3954 goto change;
3955 if (dl_policy(policy) && dl_param_changed(p, attr))
3956 goto change;
3957
3958 p->sched_reset_on_fork = reset_on_fork;
3959 task_rq_unlock(rq, p, &flags);
3960 return 0;
3961 }
3962change:
3963
3964 if (user) {
3965#ifdef CONFIG_RT_GROUP_SCHED
3966 /*
3967 * Do not allow realtime tasks into groups that have no runtime
3968 * assigned.
3969 */
3970 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3971 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3972 !task_group_is_autogroup(task_group(p))) {
3973 task_rq_unlock(rq, p, &flags);
3974 return -EPERM;
3975 }
3976#endif
3977#ifdef CONFIG_SMP
3978 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3979 cpumask_t *span = rq->rd->span;
3980
3981 /*
3982 * Don't allow tasks with an affinity mask smaller than
3983 * the entire root_domain to become SCHED_DEADLINE. We
3984 * will also fail if there's no bandwidth available.
3985 */
3986 if (!cpumask_subset(span, &p->cpus_allowed) ||
3987 rq->rd->dl_bw.bw == 0) {
3988 task_rq_unlock(rq, p, &flags);
3989 return -EPERM;
3990 }
3991 }
3992#endif
3993 }
3994
3995 /* recheck policy now with rq lock held */
3996 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3997 policy = oldpolicy = -1;
3998 task_rq_unlock(rq, p, &flags);
3999 goto recheck;
4000 }
4001
4002 /*
4003 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4004 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4005 * is available.
4006 */
4007 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4008 task_rq_unlock(rq, p, &flags);
4009 return -EBUSY;
4010 }
4011
4012 p->sched_reset_on_fork = reset_on_fork;
4013 oldprio = p->prio;
4014
4015 if (pi) {
4016 /*
4017 * Take priority boosted tasks into account. If the new
4018 * effective priority is unchanged, we just store the new
4019 * normal parameters and do not touch the scheduler class and
4020 * the runqueue. This will be done when the task deboost
4021 * itself.
4022 */
4023 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4024 if (new_effective_prio == oldprio)
4025 queue_flags &= ~DEQUEUE_MOVE;
4026 }
4027
4028 queued = task_on_rq_queued(p);
4029 running = task_current(rq, p);
4030 if (queued)
4031 dequeue_task(rq, p, queue_flags);
4032 if (running)
4033 put_prev_task(rq, p);
4034
4035 prev_class = p->sched_class;
4036 __setscheduler(rq, p, attr, pi);
4037
4038 if (running)
4039 p->sched_class->set_curr_task(rq);
4040 if (queued) {
4041 /*
4042 * We enqueue to tail when the priority of a task is
4043 * increased (user space view).
4044 */
4045 if (oldprio < p->prio)
4046 queue_flags |= ENQUEUE_HEAD;
4047
4048 enqueue_task(rq, p, queue_flags);
4049 }
4050
4051 check_class_changed(rq, p, prev_class, oldprio);
4052 preempt_disable(); /* avoid rq from going away on us */
4053 task_rq_unlock(rq, p, &flags);
4054
4055 if (pi)
4056 rt_mutex_adjust_pi(p);
4057
4058 /*
4059 * Run balance callbacks after we've adjusted the PI chain.
4060 */
4061 balance_callback(rq);
4062 preempt_enable();
4063
4064 return 0;
4065}
4066
4067static int _sched_setscheduler(struct task_struct *p, int policy,
4068 const struct sched_param *param, bool check)
4069{
4070 struct sched_attr attr = {
4071 .sched_policy = policy,
4072 .sched_priority = param->sched_priority,
4073 .sched_nice = PRIO_TO_NICE(p->static_prio),
4074 };
4075
4076 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4077 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4078 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4079 policy &= ~SCHED_RESET_ON_FORK;
4080 attr.sched_policy = policy;
4081 }
4082
4083 return __sched_setscheduler(p, &attr, check, true);
4084}
4085/**
4086 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4087 * @p: the task in question.
4088 * @policy: new policy.
4089 * @param: structure containing the new RT priority.
4090 *
4091 * Return: 0 on success. An error code otherwise.
4092 *
4093 * NOTE that the task may be already dead.
4094 */
4095int sched_setscheduler(struct task_struct *p, int policy,
4096 const struct sched_param *param)
4097{
4098 return _sched_setscheduler(p, policy, param, true);
4099}
4100EXPORT_SYMBOL_GPL(sched_setscheduler);
4101
4102int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4103{
4104 return __sched_setscheduler(p, attr, true, true);
4105}
4106EXPORT_SYMBOL_GPL(sched_setattr);
4107
4108/**
4109 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4110 * @p: the task in question.
4111 * @policy: new policy.
4112 * @param: structure containing the new RT priority.
4113 *
4114 * Just like sched_setscheduler, only don't bother checking if the
4115 * current context has permission. For example, this is needed in
4116 * stop_machine(): we create temporary high priority worker threads,
4117 * but our caller might not have that capability.
4118 *
4119 * Return: 0 on success. An error code otherwise.
4120 */
4121int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4122 const struct sched_param *param)
4123{
4124 return _sched_setscheduler(p, policy, param, false);
4125}
4126EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4127
4128static int
4129do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4130{
4131 struct sched_param lparam;
4132 struct task_struct *p;
4133 int retval;
4134
4135 if (!param || pid < 0)
4136 return -EINVAL;
4137 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4138 return -EFAULT;
4139
4140 rcu_read_lock();
4141 retval = -ESRCH;
4142 p = find_process_by_pid(pid);
4143 if (p != NULL)
4144 retval = sched_setscheduler(p, policy, &lparam);
4145 rcu_read_unlock();
4146
4147 return retval;
4148}
4149
4150/*
4151 * Mimics kernel/events/core.c perf_copy_attr().
4152 */
4153static int sched_copy_attr(struct sched_attr __user *uattr,
4154 struct sched_attr *attr)
4155{
4156 u32 size;
4157 int ret;
4158
4159 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4160 return -EFAULT;
4161
4162 /*
4163 * zero the full structure, so that a short copy will be nice.
4164 */
4165 memset(attr, 0, sizeof(*attr));
4166
4167 ret = get_user(size, &uattr->size);
4168 if (ret)
4169 return ret;
4170
4171 if (size > PAGE_SIZE) /* silly large */
4172 goto err_size;
4173
4174 if (!size) /* abi compat */
4175 size = SCHED_ATTR_SIZE_VER0;
4176
4177 if (size < SCHED_ATTR_SIZE_VER0)
4178 goto err_size;
4179
4180 /*
4181 * If we're handed a bigger struct than we know of,
4182 * ensure all the unknown bits are 0 - i.e. new
4183 * user-space does not rely on any kernel feature
4184 * extensions we dont know about yet.
4185 */
4186 if (size > sizeof(*attr)) {
4187 unsigned char __user *addr;
4188 unsigned char __user *end;
4189 unsigned char val;
4190
4191 addr = (void __user *)uattr + sizeof(*attr);
4192 end = (void __user *)uattr + size;
4193
4194 for (; addr < end; addr++) {
4195 ret = get_user(val, addr);
4196 if (ret)
4197 return ret;
4198 if (val)
4199 goto err_size;
4200 }
4201 size = sizeof(*attr);
4202 }
4203
4204 ret = copy_from_user(attr, uattr, size);
4205 if (ret)
4206 return -EFAULT;
4207
4208 /*
4209 * XXX: do we want to be lenient like existing syscalls; or do we want
4210 * to be strict and return an error on out-of-bounds values?
4211 */
4212 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4213
4214 return 0;
4215
4216err_size:
4217 put_user(sizeof(*attr), &uattr->size);
4218 return -E2BIG;
4219}
4220
4221/**
4222 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4223 * @pid: the pid in question.
4224 * @policy: new policy.
4225 * @param: structure containing the new RT priority.
4226 *
4227 * Return: 0 on success. An error code otherwise.
4228 */
4229SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4230 struct sched_param __user *, param)
4231{
4232 /* negative values for policy are not valid */
4233 if (policy < 0)
4234 return -EINVAL;
4235
4236 return do_sched_setscheduler(pid, policy, param);
4237}
4238
4239/**
4240 * sys_sched_setparam - set/change the RT priority of a thread
4241 * @pid: the pid in question.
4242 * @param: structure containing the new RT priority.
4243 *
4244 * Return: 0 on success. An error code otherwise.
4245 */
4246SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4247{
4248 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4249}
4250
4251/**
4252 * sys_sched_setattr - same as above, but with extended sched_attr
4253 * @pid: the pid in question.
4254 * @uattr: structure containing the extended parameters.
4255 * @flags: for future extension.
4256 */
4257SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4258 unsigned int, flags)
4259{
4260 struct sched_attr attr;
4261 struct task_struct *p;
4262 int retval;
4263
4264 if (!uattr || pid < 0 || flags)
4265 return -EINVAL;
4266
4267 retval = sched_copy_attr(uattr, &attr);
4268 if (retval)
4269 return retval;
4270
4271 if ((int)attr.sched_policy < 0)
4272 return -EINVAL;
4273
4274 rcu_read_lock();
4275 retval = -ESRCH;
4276 p = find_process_by_pid(pid);
4277 if (p != NULL)
4278 retval = sched_setattr(p, &attr);
4279 rcu_read_unlock();
4280
4281 return retval;
4282}
4283
4284/**
4285 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4286 * @pid: the pid in question.
4287 *
4288 * Return: On success, the policy of the thread. Otherwise, a negative error
4289 * code.
4290 */
4291SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4292{
4293 struct task_struct *p;
4294 int retval;
4295
4296 if (pid < 0)
4297 return -EINVAL;
4298
4299 retval = -ESRCH;
4300 rcu_read_lock();
4301 p = find_process_by_pid(pid);
4302 if (p) {
4303 retval = security_task_getscheduler(p);
4304 if (!retval)
4305 retval = p->policy
4306 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4307 }
4308 rcu_read_unlock();
4309 return retval;
4310}
4311
4312/**
4313 * sys_sched_getparam - get the RT priority of a thread
4314 * @pid: the pid in question.
4315 * @param: structure containing the RT priority.
4316 *
4317 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4318 * code.
4319 */
4320SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4321{
4322 struct sched_param lp = { .sched_priority = 0 };
4323 struct task_struct *p;
4324 int retval;
4325
4326 if (!param || pid < 0)
4327 return -EINVAL;
4328
4329 rcu_read_lock();
4330 p = find_process_by_pid(pid);
4331 retval = -ESRCH;
4332 if (!p)
4333 goto out_unlock;
4334
4335 retval = security_task_getscheduler(p);
4336 if (retval)
4337 goto out_unlock;
4338
4339 if (task_has_rt_policy(p))
4340 lp.sched_priority = p->rt_priority;
4341 rcu_read_unlock();
4342
4343 /*
4344 * This one might sleep, we cannot do it with a spinlock held ...
4345 */
4346 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4347
4348 return retval;
4349
4350out_unlock:
4351 rcu_read_unlock();
4352 return retval;
4353}
4354
4355static int sched_read_attr(struct sched_attr __user *uattr,
4356 struct sched_attr *attr,
4357 unsigned int usize)
4358{
4359 int ret;
4360
4361 if (!access_ok(VERIFY_WRITE, uattr, usize))
4362 return -EFAULT;
4363
4364 /*
4365 * If we're handed a smaller struct than we know of,
4366 * ensure all the unknown bits are 0 - i.e. old
4367 * user-space does not get uncomplete information.
4368 */
4369 if (usize < sizeof(*attr)) {
4370 unsigned char *addr;
4371 unsigned char *end;
4372
4373 addr = (void *)attr + usize;
4374 end = (void *)attr + sizeof(*attr);
4375
4376 for (; addr < end; addr++) {
4377 if (*addr)
4378 return -EFBIG;
4379 }
4380
4381 attr->size = usize;
4382 }
4383
4384 ret = copy_to_user(uattr, attr, attr->size);
4385 if (ret)
4386 return -EFAULT;
4387
4388 return 0;
4389}
4390
4391/**
4392 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4393 * @pid: the pid in question.
4394 * @uattr: structure containing the extended parameters.
4395 * @size: sizeof(attr) for fwd/bwd comp.
4396 * @flags: for future extension.
4397 */
4398SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4399 unsigned int, size, unsigned int, flags)
4400{
4401 struct sched_attr attr = {
4402 .size = sizeof(struct sched_attr),
4403 };
4404 struct task_struct *p;
4405 int retval;
4406
4407 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4408 size < SCHED_ATTR_SIZE_VER0 || flags)
4409 return -EINVAL;
4410
4411 rcu_read_lock();
4412 p = find_process_by_pid(pid);
4413 retval = -ESRCH;
4414 if (!p)
4415 goto out_unlock;
4416
4417 retval = security_task_getscheduler(p);
4418 if (retval)
4419 goto out_unlock;
4420
4421 attr.sched_policy = p->policy;
4422 if (p->sched_reset_on_fork)
4423 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4424 if (task_has_dl_policy(p))
4425 __getparam_dl(p, &attr);
4426 else if (task_has_rt_policy(p))
4427 attr.sched_priority = p->rt_priority;
4428 else
4429 attr.sched_nice = task_nice(p);
4430
4431 rcu_read_unlock();
4432
4433 retval = sched_read_attr(uattr, &attr, size);
4434 return retval;
4435
4436out_unlock:
4437 rcu_read_unlock();
4438 return retval;
4439}
4440
4441long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4442{
4443 cpumask_var_t cpus_allowed, new_mask;
4444 struct task_struct *p;
4445 int retval;
4446
4447 rcu_read_lock();
4448
4449 p = find_process_by_pid(pid);
4450 if (!p) {
4451 rcu_read_unlock();
4452 return -ESRCH;
4453 }
4454
4455 /* Prevent p going away */
4456 get_task_struct(p);
4457 rcu_read_unlock();
4458
4459 if (p->flags & PF_NO_SETAFFINITY) {
4460 retval = -EINVAL;
4461 goto out_put_task;
4462 }
4463 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4464 retval = -ENOMEM;
4465 goto out_put_task;
4466 }
4467 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4468 retval = -ENOMEM;
4469 goto out_free_cpus_allowed;
4470 }
4471 retval = -EPERM;
4472 if (!check_same_owner(p)) {
4473 rcu_read_lock();
4474 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4475 rcu_read_unlock();
4476 goto out_free_new_mask;
4477 }
4478 rcu_read_unlock();
4479 }
4480
4481 retval = security_task_setscheduler(p);
4482 if (retval)
4483 goto out_free_new_mask;
4484
4485
4486 cpuset_cpus_allowed(p, cpus_allowed);
4487 cpumask_and(new_mask, in_mask, cpus_allowed);
4488
4489 /*
4490 * Since bandwidth control happens on root_domain basis,
4491 * if admission test is enabled, we only admit -deadline
4492 * tasks allowed to run on all the CPUs in the task's
4493 * root_domain.
4494 */
4495#ifdef CONFIG_SMP
4496 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4497 rcu_read_lock();
4498 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4499 retval = -EBUSY;
4500 rcu_read_unlock();
4501 goto out_free_new_mask;
4502 }
4503 rcu_read_unlock();
4504 }
4505#endif
4506again:
4507 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4508
4509 if (!retval) {
4510 cpuset_cpus_allowed(p, cpus_allowed);
4511 if (!cpumask_subset(new_mask, cpus_allowed)) {
4512 /*
4513 * We must have raced with a concurrent cpuset
4514 * update. Just reset the cpus_allowed to the
4515 * cpuset's cpus_allowed
4516 */
4517 cpumask_copy(new_mask, cpus_allowed);
4518 goto again;
4519 }
4520 }
4521out_free_new_mask:
4522 free_cpumask_var(new_mask);
4523out_free_cpus_allowed:
4524 free_cpumask_var(cpus_allowed);
4525out_put_task:
4526 put_task_struct(p);
4527 return retval;
4528}
4529
4530static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4531 struct cpumask *new_mask)
4532{
4533 if (len < cpumask_size())
4534 cpumask_clear(new_mask);
4535 else if (len > cpumask_size())
4536 len = cpumask_size();
4537
4538 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4539}
4540
4541/**
4542 * sys_sched_setaffinity - set the cpu affinity of a process
4543 * @pid: pid of the process
4544 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4545 * @user_mask_ptr: user-space pointer to the new cpu mask
4546 *
4547 * Return: 0 on success. An error code otherwise.
4548 */
4549SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4550 unsigned long __user *, user_mask_ptr)
4551{
4552 cpumask_var_t new_mask;
4553 int retval;
4554
4555 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4556 return -ENOMEM;
4557
4558 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4559 if (retval == 0)
4560 retval = sched_setaffinity(pid, new_mask);
4561 free_cpumask_var(new_mask);
4562 return retval;
4563}
4564
4565long sched_getaffinity(pid_t pid, struct cpumask *mask)
4566{
4567 struct task_struct *p;
4568 unsigned long flags;
4569 int retval;
4570
4571 rcu_read_lock();
4572
4573 retval = -ESRCH;
4574 p = find_process_by_pid(pid);
4575 if (!p)
4576 goto out_unlock;
4577
4578 retval = security_task_getscheduler(p);
4579 if (retval)
4580 goto out_unlock;
4581
4582 raw_spin_lock_irqsave(&p->pi_lock, flags);
4583 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4584 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4585
4586out_unlock:
4587 rcu_read_unlock();
4588
4589 return retval;
4590}
4591
4592/**
4593 * sys_sched_getaffinity - get the cpu affinity of a process
4594 * @pid: pid of the process
4595 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4596 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4597 *
4598 * Return: 0 on success. An error code otherwise.
4599 */
4600SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4601 unsigned long __user *, user_mask_ptr)
4602{
4603 int ret;
4604 cpumask_var_t mask;
4605
4606 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4607 return -EINVAL;
4608 if (len & (sizeof(unsigned long)-1))
4609 return -EINVAL;
4610
4611 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4612 return -ENOMEM;
4613
4614 ret = sched_getaffinity(pid, mask);
4615 if (ret == 0) {
4616 size_t retlen = min_t(size_t, len, cpumask_size());
4617
4618 if (copy_to_user(user_mask_ptr, mask, retlen))
4619 ret = -EFAULT;
4620 else
4621 ret = retlen;
4622 }
4623 free_cpumask_var(mask);
4624
4625 return ret;
4626}
4627
4628/**
4629 * sys_sched_yield - yield the current processor to other threads.
4630 *
4631 * This function yields the current CPU to other tasks. If there are no
4632 * other threads running on this CPU then this function will return.
4633 *
4634 * Return: 0.
4635 */
4636SYSCALL_DEFINE0(sched_yield)
4637{
4638 struct rq *rq = this_rq_lock();
4639
4640 schedstat_inc(rq, yld_count);
4641 current->sched_class->yield_task(rq);
4642
4643 /*
4644 * Since we are going to call schedule() anyway, there's
4645 * no need to preempt or enable interrupts:
4646 */
4647 __release(rq->lock);
4648 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4649 do_raw_spin_unlock(&rq->lock);
4650 sched_preempt_enable_no_resched();
4651
4652 schedule();
4653
4654 return 0;
4655}
4656
4657int __sched _cond_resched(void)
4658{
4659 if (should_resched(0)) {
4660 preempt_schedule_common();
4661 return 1;
4662 }
4663 return 0;
4664}
4665EXPORT_SYMBOL(_cond_resched);
4666
4667/*
4668 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4669 * call schedule, and on return reacquire the lock.
4670 *
4671 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4672 * operations here to prevent schedule() from being called twice (once via
4673 * spin_unlock(), once by hand).
4674 */
4675int __cond_resched_lock(spinlock_t *lock)
4676{
4677 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4678 int ret = 0;
4679
4680 lockdep_assert_held(lock);
4681
4682 if (spin_needbreak(lock) || resched) {
4683 spin_unlock(lock);
4684 if (resched)
4685 preempt_schedule_common();
4686 else
4687 cpu_relax();
4688 ret = 1;
4689 spin_lock(lock);
4690 }
4691 return ret;
4692}
4693EXPORT_SYMBOL(__cond_resched_lock);
4694
4695int __sched __cond_resched_softirq(void)
4696{
4697 BUG_ON(!in_softirq());
4698
4699 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4700 local_bh_enable();
4701 preempt_schedule_common();
4702 local_bh_disable();
4703 return 1;
4704 }
4705 return 0;
4706}
4707EXPORT_SYMBOL(__cond_resched_softirq);
4708
4709/**
4710 * yield - yield the current processor to other threads.
4711 *
4712 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4713 *
4714 * The scheduler is at all times free to pick the calling task as the most
4715 * eligible task to run, if removing the yield() call from your code breaks
4716 * it, its already broken.
4717 *
4718 * Typical broken usage is:
4719 *
4720 * while (!event)
4721 * yield();
4722 *
4723 * where one assumes that yield() will let 'the other' process run that will
4724 * make event true. If the current task is a SCHED_FIFO task that will never
4725 * happen. Never use yield() as a progress guarantee!!
4726 *
4727 * If you want to use yield() to wait for something, use wait_event().
4728 * If you want to use yield() to be 'nice' for others, use cond_resched().
4729 * If you still want to use yield(), do not!
4730 */
4731void __sched yield(void)
4732{
4733 set_current_state(TASK_RUNNING);
4734 sys_sched_yield();
4735}
4736EXPORT_SYMBOL(yield);
4737
4738/**
4739 * yield_to - yield the current processor to another thread in
4740 * your thread group, or accelerate that thread toward the
4741 * processor it's on.
4742 * @p: target task
4743 * @preempt: whether task preemption is allowed or not
4744 *
4745 * It's the caller's job to ensure that the target task struct
4746 * can't go away on us before we can do any checks.
4747 *
4748 * Return:
4749 * true (>0) if we indeed boosted the target task.
4750 * false (0) if we failed to boost the target.
4751 * -ESRCH if there's no task to yield to.
4752 */
4753int __sched yield_to(struct task_struct *p, bool preempt)
4754{
4755 struct task_struct *curr = current;
4756 struct rq *rq, *p_rq;
4757 unsigned long flags;
4758 int yielded = 0;
4759
4760 local_irq_save(flags);
4761 rq = this_rq();
4762
4763again:
4764 p_rq = task_rq(p);
4765 /*
4766 * If we're the only runnable task on the rq and target rq also
4767 * has only one task, there's absolutely no point in yielding.
4768 */
4769 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4770 yielded = -ESRCH;
4771 goto out_irq;
4772 }
4773
4774 double_rq_lock(rq, p_rq);
4775 if (task_rq(p) != p_rq) {
4776 double_rq_unlock(rq, p_rq);
4777 goto again;
4778 }
4779
4780 if (!curr->sched_class->yield_to_task)
4781 goto out_unlock;
4782
4783 if (curr->sched_class != p->sched_class)
4784 goto out_unlock;
4785
4786 if (task_running(p_rq, p) || p->state)
4787 goto out_unlock;
4788
4789 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4790 if (yielded) {
4791 schedstat_inc(rq, yld_count);
4792 /*
4793 * Make p's CPU reschedule; pick_next_entity takes care of
4794 * fairness.
4795 */
4796 if (preempt && rq != p_rq)
4797 resched_curr(p_rq);
4798 }
4799
4800out_unlock:
4801 double_rq_unlock(rq, p_rq);
4802out_irq:
4803 local_irq_restore(flags);
4804
4805 if (yielded > 0)
4806 schedule();
4807
4808 return yielded;
4809}
4810EXPORT_SYMBOL_GPL(yield_to);
4811
4812/*
4813 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4814 * that process accounting knows that this is a task in IO wait state.
4815 */
4816long __sched io_schedule_timeout(long timeout)
4817{
4818 int old_iowait = current->in_iowait;
4819 struct rq *rq;
4820 long ret;
4821
4822 current->in_iowait = 1;
4823 blk_schedule_flush_plug(current);
4824
4825 delayacct_blkio_start();
4826 rq = raw_rq();
4827 atomic_inc(&rq->nr_iowait);
4828 ret = schedule_timeout(timeout);
4829 current->in_iowait = old_iowait;
4830 atomic_dec(&rq->nr_iowait);
4831 delayacct_blkio_end();
4832
4833 return ret;
4834}
4835EXPORT_SYMBOL(io_schedule_timeout);
4836
4837/**
4838 * sys_sched_get_priority_max - return maximum RT priority.
4839 * @policy: scheduling class.
4840 *
4841 * Return: On success, this syscall returns the maximum
4842 * rt_priority that can be used by a given scheduling class.
4843 * On failure, a negative error code is returned.
4844 */
4845SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4846{
4847 int ret = -EINVAL;
4848
4849 switch (policy) {
4850 case SCHED_FIFO:
4851 case SCHED_RR:
4852 ret = MAX_USER_RT_PRIO-1;
4853 break;
4854 case SCHED_DEADLINE:
4855 case SCHED_NORMAL:
4856 case SCHED_BATCH:
4857 case SCHED_IDLE:
4858 ret = 0;
4859 break;
4860 }
4861 return ret;
4862}
4863
4864/**
4865 * sys_sched_get_priority_min - return minimum RT priority.
4866 * @policy: scheduling class.
4867 *
4868 * Return: On success, this syscall returns the minimum
4869 * rt_priority that can be used by a given scheduling class.
4870 * On failure, a negative error code is returned.
4871 */
4872SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4873{
4874 int ret = -EINVAL;
4875
4876 switch (policy) {
4877 case SCHED_FIFO:
4878 case SCHED_RR:
4879 ret = 1;
4880 break;
4881 case SCHED_DEADLINE:
4882 case SCHED_NORMAL:
4883 case SCHED_BATCH:
4884 case SCHED_IDLE:
4885 ret = 0;
4886 }
4887 return ret;
4888}
4889
4890/**
4891 * sys_sched_rr_get_interval - return the default timeslice of a process.
4892 * @pid: pid of the process.
4893 * @interval: userspace pointer to the timeslice value.
4894 *
4895 * this syscall writes the default timeslice value of a given process
4896 * into the user-space timespec buffer. A value of '0' means infinity.
4897 *
4898 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4899 * an error code.
4900 */
4901SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4902 struct timespec __user *, interval)
4903{
4904 struct task_struct *p;
4905 unsigned int time_slice;
4906 unsigned long flags;
4907 struct rq *rq;
4908 int retval;
4909 struct timespec t;
4910
4911 if (pid < 0)
4912 return -EINVAL;
4913
4914 retval = -ESRCH;
4915 rcu_read_lock();
4916 p = find_process_by_pid(pid);
4917 if (!p)
4918 goto out_unlock;
4919
4920 retval = security_task_getscheduler(p);
4921 if (retval)
4922 goto out_unlock;
4923
4924 rq = task_rq_lock(p, &flags);
4925 time_slice = 0;
4926 if (p->sched_class->get_rr_interval)
4927 time_slice = p->sched_class->get_rr_interval(rq, p);
4928 task_rq_unlock(rq, p, &flags);
4929
4930 rcu_read_unlock();
4931 jiffies_to_timespec(time_slice, &t);
4932 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4933 return retval;
4934
4935out_unlock:
4936 rcu_read_unlock();
4937 return retval;
4938}
4939
4940static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4941
4942void sched_show_task(struct task_struct *p)
4943{
4944 unsigned long free = 0;
4945 int ppid;
4946 unsigned long state = p->state;
4947
4948 if (state)
4949 state = __ffs(state) + 1;
4950 printk(KERN_INFO "%-15.15s %c", p->comm,
4951 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4952#if BITS_PER_LONG == 32
4953 if (state == TASK_RUNNING)
4954 printk(KERN_CONT " running ");
4955 else
4956 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4957#else
4958 if (state == TASK_RUNNING)
4959 printk(KERN_CONT " running task ");
4960 else
4961 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4962#endif
4963#ifdef CONFIG_DEBUG_STACK_USAGE
4964 free = stack_not_used(p);
4965#endif
4966 ppid = 0;
4967 rcu_read_lock();
4968 if (pid_alive(p))
4969 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4970 rcu_read_unlock();
4971 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4972 task_pid_nr(p), ppid,
4973 (unsigned long)task_thread_info(p)->flags);
4974
4975 print_worker_info(KERN_INFO, p);
4976 show_stack(p, NULL);
4977}
4978
4979void show_state_filter(unsigned long state_filter)
4980{
4981 struct task_struct *g, *p;
4982
4983#if BITS_PER_LONG == 32
4984 printk(KERN_INFO
4985 " task PC stack pid father\n");
4986#else
4987 printk(KERN_INFO
4988 " task PC stack pid father\n");
4989#endif
4990 rcu_read_lock();
4991 for_each_process_thread(g, p) {
4992 /*
4993 * reset the NMI-timeout, listing all files on a slow
4994 * console might take a lot of time:
4995 */
4996 touch_nmi_watchdog();
4997 if (!state_filter || (p->state & state_filter))
4998 sched_show_task(p);
4999 }
5000
5001 touch_all_softlockup_watchdogs();
5002
5003#ifdef CONFIG_SCHED_DEBUG
5004 sysrq_sched_debug_show();
5005#endif
5006 rcu_read_unlock();
5007 /*
5008 * Only show locks if all tasks are dumped:
5009 */
5010 if (!state_filter)
5011 debug_show_all_locks();
5012}
5013
5014void init_idle_bootup_task(struct task_struct *idle)
5015{
5016 idle->sched_class = &idle_sched_class;
5017}
5018
5019/**
5020 * init_idle - set up an idle thread for a given CPU
5021 * @idle: task in question
5022 * @cpu: cpu the idle task belongs to
5023 *
5024 * NOTE: this function does not set the idle thread's NEED_RESCHED
5025 * flag, to make booting more robust.
5026 */
5027void init_idle(struct task_struct *idle, int cpu)
5028{
5029 struct rq *rq = cpu_rq(cpu);
5030 unsigned long flags;
5031
5032 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5033 raw_spin_lock(&rq->lock);
5034
5035 __sched_fork(0, idle);
5036 idle->state = TASK_RUNNING;
5037 idle->se.exec_start = sched_clock();
5038
5039 kasan_unpoison_task_stack(idle);
5040
5041#ifdef CONFIG_SMP
5042 /*
5043 * Its possible that init_idle() gets called multiple times on a task,
5044 * in that case do_set_cpus_allowed() will not do the right thing.
5045 *
5046 * And since this is boot we can forgo the serialization.
5047 */
5048 set_cpus_allowed_common(idle, cpumask_of(cpu));
5049#endif
5050 /*
5051 * We're having a chicken and egg problem, even though we are
5052 * holding rq->lock, the cpu isn't yet set to this cpu so the
5053 * lockdep check in task_group() will fail.
5054 *
5055 * Similar case to sched_fork(). / Alternatively we could
5056 * use task_rq_lock() here and obtain the other rq->lock.
5057 *
5058 * Silence PROVE_RCU
5059 */
5060 rcu_read_lock();
5061 __set_task_cpu(idle, cpu);
5062 rcu_read_unlock();
5063
5064 rq->curr = rq->idle = idle;
5065 idle->on_rq = TASK_ON_RQ_QUEUED;
5066#ifdef CONFIG_SMP
5067 idle->on_cpu = 1;
5068#endif
5069 raw_spin_unlock(&rq->lock);
5070 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5071
5072 /* Set the preempt count _outside_ the spinlocks! */
5073 init_idle_preempt_count(idle, cpu);
5074
5075 /*
5076 * The idle tasks have their own, simple scheduling class:
5077 */
5078 idle->sched_class = &idle_sched_class;
5079 ftrace_graph_init_idle_task(idle, cpu);
5080 vtime_init_idle(idle, cpu);
5081#ifdef CONFIG_SMP
5082 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5083#endif
5084}
5085
5086int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5087 const struct cpumask *trial)
5088{
5089 int ret = 1, trial_cpus;
5090 struct dl_bw *cur_dl_b;
5091 unsigned long flags;
5092
5093 if (!cpumask_weight(cur))
5094 return ret;
5095
5096 rcu_read_lock_sched();
5097 cur_dl_b = dl_bw_of(cpumask_any(cur));
5098 trial_cpus = cpumask_weight(trial);
5099
5100 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5101 if (cur_dl_b->bw != -1 &&
5102 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5103 ret = 0;
5104 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5105 rcu_read_unlock_sched();
5106
5107 return ret;
5108}
5109
5110int task_can_attach(struct task_struct *p,
5111 const struct cpumask *cs_cpus_allowed)
5112{
5113 int ret = 0;
5114
5115 /*
5116 * Kthreads which disallow setaffinity shouldn't be moved
5117 * to a new cpuset; we don't want to change their cpu
5118 * affinity and isolating such threads by their set of
5119 * allowed nodes is unnecessary. Thus, cpusets are not
5120 * applicable for such threads. This prevents checking for
5121 * success of set_cpus_allowed_ptr() on all attached tasks
5122 * before cpus_allowed may be changed.
5123 */
5124 if (p->flags & PF_NO_SETAFFINITY) {
5125 ret = -EINVAL;
5126 goto out;
5127 }
5128
5129#ifdef CONFIG_SMP
5130 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5131 cs_cpus_allowed)) {
5132 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5133 cs_cpus_allowed);
5134 struct dl_bw *dl_b;
5135 bool overflow;
5136 int cpus;
5137 unsigned long flags;
5138
5139 rcu_read_lock_sched();
5140 dl_b = dl_bw_of(dest_cpu);
5141 raw_spin_lock_irqsave(&dl_b->lock, flags);
5142 cpus = dl_bw_cpus(dest_cpu);
5143 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5144 if (overflow)
5145 ret = -EBUSY;
5146 else {
5147 /*
5148 * We reserve space for this task in the destination
5149 * root_domain, as we can't fail after this point.
5150 * We will free resources in the source root_domain
5151 * later on (see set_cpus_allowed_dl()).
5152 */
5153 __dl_add(dl_b, p->dl.dl_bw);
5154 }
5155 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5156 rcu_read_unlock_sched();
5157
5158 }
5159#endif
5160out:
5161 return ret;
5162}
5163
5164#ifdef CONFIG_SMP
5165
5166#ifdef CONFIG_NUMA_BALANCING
5167/* Migrate current task p to target_cpu */
5168int migrate_task_to(struct task_struct *p, int target_cpu)
5169{
5170 struct migration_arg arg = { p, target_cpu };
5171 int curr_cpu = task_cpu(p);
5172
5173 if (curr_cpu == target_cpu)
5174 return 0;
5175
5176 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5177 return -EINVAL;
5178
5179 /* TODO: This is not properly updating schedstats */
5180
5181 trace_sched_move_numa(p, curr_cpu, target_cpu);
5182 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5183}
5184
5185/*
5186 * Requeue a task on a given node and accurately track the number of NUMA
5187 * tasks on the runqueues
5188 */
5189void sched_setnuma(struct task_struct *p, int nid)
5190{
5191 struct rq *rq;
5192 unsigned long flags;
5193 bool queued, running;
5194
5195 rq = task_rq_lock(p, &flags);
5196 queued = task_on_rq_queued(p);
5197 running = task_current(rq, p);
5198
5199 if (queued)
5200 dequeue_task(rq, p, DEQUEUE_SAVE);
5201 if (running)
5202 put_prev_task(rq, p);
5203
5204 p->numa_preferred_nid = nid;
5205
5206 if (running)
5207 p->sched_class->set_curr_task(rq);
5208 if (queued)
5209 enqueue_task(rq, p, ENQUEUE_RESTORE);
5210 task_rq_unlock(rq, p, &flags);
5211}
5212#endif /* CONFIG_NUMA_BALANCING */
5213
5214#ifdef CONFIG_HOTPLUG_CPU
5215/*
5216 * Ensures that the idle task is using init_mm right before its cpu goes
5217 * offline.
5218 */
5219void idle_task_exit(void)
5220{
5221 struct mm_struct *mm = current->active_mm;
5222
5223 BUG_ON(cpu_online(smp_processor_id()));
5224
5225 if (mm != &init_mm) {
5226 switch_mm(mm, &init_mm, current);
5227 finish_arch_post_lock_switch();
5228 }
5229 mmdrop(mm);
5230}
5231
5232/*
5233 * Since this CPU is going 'away' for a while, fold any nr_active delta
5234 * we might have. Assumes we're called after migrate_tasks() so that the
5235 * nr_active count is stable.
5236 *
5237 * Also see the comment "Global load-average calculations".
5238 */
5239static void calc_load_migrate(struct rq *rq)
5240{
5241 long delta = calc_load_fold_active(rq);
5242 if (delta)
5243 atomic_long_add(delta, &calc_load_tasks);
5244}
5245
5246static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5247{
5248}
5249
5250static const struct sched_class fake_sched_class = {
5251 .put_prev_task = put_prev_task_fake,
5252};
5253
5254static struct task_struct fake_task = {
5255 /*
5256 * Avoid pull_{rt,dl}_task()
5257 */
5258 .prio = MAX_PRIO + 1,
5259 .sched_class = &fake_sched_class,
5260};
5261
5262/*
5263 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5264 * try_to_wake_up()->select_task_rq().
5265 *
5266 * Called with rq->lock held even though we'er in stop_machine() and
5267 * there's no concurrency possible, we hold the required locks anyway
5268 * because of lock validation efforts.
5269 */
5270static void migrate_tasks(struct rq *dead_rq)
5271{
5272 struct rq *rq = dead_rq;
5273 struct task_struct *next, *stop = rq->stop;
5274 int dest_cpu;
5275
5276 /*
5277 * Fudge the rq selection such that the below task selection loop
5278 * doesn't get stuck on the currently eligible stop task.
5279 *
5280 * We're currently inside stop_machine() and the rq is either stuck
5281 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5282 * either way we should never end up calling schedule() until we're
5283 * done here.
5284 */
5285 rq->stop = NULL;
5286
5287 /*
5288 * put_prev_task() and pick_next_task() sched
5289 * class method both need to have an up-to-date
5290 * value of rq->clock[_task]
5291 */
5292 update_rq_clock(rq);
5293
5294 for (;;) {
5295 /*
5296 * There's this thread running, bail when that's the only
5297 * remaining thread.
5298 */
5299 if (rq->nr_running == 1)
5300 break;
5301
5302 /*
5303 * pick_next_task assumes pinned rq->lock.
5304 */
5305 lockdep_pin_lock(&rq->lock);
5306 next = pick_next_task(rq, &fake_task);
5307 BUG_ON(!next);
5308 next->sched_class->put_prev_task(rq, next);
5309
5310 /*
5311 * Rules for changing task_struct::cpus_allowed are holding
5312 * both pi_lock and rq->lock, such that holding either
5313 * stabilizes the mask.
5314 *
5315 * Drop rq->lock is not quite as disastrous as it usually is
5316 * because !cpu_active at this point, which means load-balance
5317 * will not interfere. Also, stop-machine.
5318 */
5319 lockdep_unpin_lock(&rq->lock);
5320 raw_spin_unlock(&rq->lock);
5321 raw_spin_lock(&next->pi_lock);
5322 raw_spin_lock(&rq->lock);
5323
5324 /*
5325 * Since we're inside stop-machine, _nothing_ should have
5326 * changed the task, WARN if weird stuff happened, because in
5327 * that case the above rq->lock drop is a fail too.
5328 */
5329 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5330 raw_spin_unlock(&next->pi_lock);
5331 continue;
5332 }
5333
5334 /* Find suitable destination for @next, with force if needed. */
5335 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5336
5337 rq = __migrate_task(rq, next, dest_cpu);
5338 if (rq != dead_rq) {
5339 raw_spin_unlock(&rq->lock);
5340 rq = dead_rq;
5341 raw_spin_lock(&rq->lock);
5342 }
5343 raw_spin_unlock(&next->pi_lock);
5344 }
5345
5346 rq->stop = stop;
5347}
5348#endif /* CONFIG_HOTPLUG_CPU */
5349
5350static void set_rq_online(struct rq *rq)
5351{
5352 if (!rq->online) {
5353 const struct sched_class *class;
5354
5355 cpumask_set_cpu(rq->cpu, rq->rd->online);
5356 rq->online = 1;
5357
5358 for_each_class(class) {
5359 if (class->rq_online)
5360 class->rq_online(rq);
5361 }
5362 }
5363}
5364
5365static void set_rq_offline(struct rq *rq)
5366{
5367 if (rq->online) {
5368 const struct sched_class *class;
5369
5370 for_each_class(class) {
5371 if (class->rq_offline)
5372 class->rq_offline(rq);
5373 }
5374
5375 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5376 rq->online = 0;
5377 }
5378}
5379
5380/*
5381 * migration_call - callback that gets triggered when a CPU is added.
5382 * Here we can start up the necessary migration thread for the new CPU.
5383 */
5384static int
5385migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5386{
5387 int cpu = (long)hcpu;
5388 unsigned long flags;
5389 struct rq *rq = cpu_rq(cpu);
5390
5391 switch (action & ~CPU_TASKS_FROZEN) {
5392
5393 case CPU_UP_PREPARE:
5394 rq->calc_load_update = calc_load_update;
5395 account_reset_rq(rq);
5396 break;
5397
5398 case CPU_ONLINE:
5399 /* Update our root-domain */
5400 raw_spin_lock_irqsave(&rq->lock, flags);
5401 if (rq->rd) {
5402 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5403
5404 set_rq_online(rq);
5405 }
5406 raw_spin_unlock_irqrestore(&rq->lock, flags);
5407 break;
5408
5409#ifdef CONFIG_HOTPLUG_CPU
5410 case CPU_DYING:
5411 sched_ttwu_pending();
5412 /* Update our root-domain */
5413 raw_spin_lock_irqsave(&rq->lock, flags);
5414 if (rq->rd) {
5415 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5416 set_rq_offline(rq);
5417 }
5418 migrate_tasks(rq);
5419 BUG_ON(rq->nr_running != 1); /* the migration thread */
5420 raw_spin_unlock_irqrestore(&rq->lock, flags);
5421 break;
5422
5423 case CPU_DEAD:
5424 calc_load_migrate(rq);
5425 break;
5426#endif
5427 }
5428
5429 update_max_interval();
5430
5431 return NOTIFY_OK;
5432}
5433
5434/*
5435 * Register at high priority so that task migration (migrate_all_tasks)
5436 * happens before everything else. This has to be lower priority than
5437 * the notifier in the perf_event subsystem, though.
5438 */
5439static struct notifier_block migration_notifier = {
5440 .notifier_call = migration_call,
5441 .priority = CPU_PRI_MIGRATION,
5442};
5443
5444static void set_cpu_rq_start_time(void)
5445{
5446 int cpu = smp_processor_id();
5447 struct rq *rq = cpu_rq(cpu);
5448 rq->age_stamp = sched_clock_cpu(cpu);
5449}
5450
5451static int sched_cpu_active(struct notifier_block *nfb,
5452 unsigned long action, void *hcpu)
5453{
5454 int cpu = (long)hcpu;
5455
5456 switch (action & ~CPU_TASKS_FROZEN) {
5457 case CPU_STARTING:
5458 set_cpu_rq_start_time();
5459 return NOTIFY_OK;
5460
5461 case CPU_DOWN_FAILED:
5462 set_cpu_active(cpu, true);
5463 return NOTIFY_OK;
5464
5465 default:
5466 return NOTIFY_DONE;
5467 }
5468}
5469
5470static int sched_cpu_inactive(struct notifier_block *nfb,
5471 unsigned long action, void *hcpu)
5472{
5473 switch (action & ~CPU_TASKS_FROZEN) {
5474 case CPU_DOWN_PREPARE:
5475 set_cpu_active((long)hcpu, false);
5476 return NOTIFY_OK;
5477 default:
5478 return NOTIFY_DONE;
5479 }
5480}
5481
5482static int __init migration_init(void)
5483{
5484 void *cpu = (void *)(long)smp_processor_id();
5485 int err;
5486
5487 /* Initialize migration for the boot CPU */
5488 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5489 BUG_ON(err == NOTIFY_BAD);
5490 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5491 register_cpu_notifier(&migration_notifier);
5492
5493 /* Register cpu active notifiers */
5494 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5495 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5496
5497 return 0;
5498}
5499early_initcall(migration_init);
5500
5501static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5502
5503#ifdef CONFIG_SCHED_DEBUG
5504
5505static __read_mostly int sched_debug_enabled;
5506
5507static int __init sched_debug_setup(char *str)
5508{
5509 sched_debug_enabled = 1;
5510
5511 return 0;
5512}
5513early_param("sched_debug", sched_debug_setup);
5514
5515static inline bool sched_debug(void)
5516{
5517 return sched_debug_enabled;
5518}
5519
5520static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5521 struct cpumask *groupmask)
5522{
5523 struct sched_group *group = sd->groups;
5524
5525 cpumask_clear(groupmask);
5526
5527 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5528
5529 if (!(sd->flags & SD_LOAD_BALANCE)) {
5530 printk("does not load-balance\n");
5531 if (sd->parent)
5532 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5533 " has parent");
5534 return -1;
5535 }
5536
5537 printk(KERN_CONT "span %*pbl level %s\n",
5538 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5539
5540 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5541 printk(KERN_ERR "ERROR: domain->span does not contain "
5542 "CPU%d\n", cpu);
5543 }
5544 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5545 printk(KERN_ERR "ERROR: domain->groups does not contain"
5546 " CPU%d\n", cpu);
5547 }
5548
5549 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5550 do {
5551 if (!group) {
5552 printk("\n");
5553 printk(KERN_ERR "ERROR: group is NULL\n");
5554 break;
5555 }
5556
5557 if (!cpumask_weight(sched_group_cpus(group))) {
5558 printk(KERN_CONT "\n");
5559 printk(KERN_ERR "ERROR: empty group\n");
5560 break;
5561 }
5562
5563 if (!(sd->flags & SD_OVERLAP) &&
5564 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5565 printk(KERN_CONT "\n");
5566 printk(KERN_ERR "ERROR: repeated CPUs\n");
5567 break;
5568 }
5569
5570 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5571
5572 printk(KERN_CONT " %*pbl",
5573 cpumask_pr_args(sched_group_cpus(group)));
5574 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5575 printk(KERN_CONT " (cpu_capacity = %d)",
5576 group->sgc->capacity);
5577 }
5578
5579 group = group->next;
5580 } while (group != sd->groups);
5581 printk(KERN_CONT "\n");
5582
5583 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5584 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5585
5586 if (sd->parent &&
5587 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5588 printk(KERN_ERR "ERROR: parent span is not a superset "
5589 "of domain->span\n");
5590 return 0;
5591}
5592
5593static void sched_domain_debug(struct sched_domain *sd, int cpu)
5594{
5595 int level = 0;
5596
5597 if (!sched_debug_enabled)
5598 return;
5599
5600 if (!sd) {
5601 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5602 return;
5603 }
5604
5605 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5606
5607 for (;;) {
5608 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5609 break;
5610 level++;
5611 sd = sd->parent;
5612 if (!sd)
5613 break;
5614 }
5615}
5616#else /* !CONFIG_SCHED_DEBUG */
5617# define sched_domain_debug(sd, cpu) do { } while (0)
5618static inline bool sched_debug(void)
5619{
5620 return false;
5621}
5622#endif /* CONFIG_SCHED_DEBUG */
5623
5624static int sd_degenerate(struct sched_domain *sd)
5625{
5626 if (cpumask_weight(sched_domain_span(sd)) == 1)
5627 return 1;
5628
5629 /* Following flags need at least 2 groups */
5630 if (sd->flags & (SD_LOAD_BALANCE |
5631 SD_BALANCE_NEWIDLE |
5632 SD_BALANCE_FORK |
5633 SD_BALANCE_EXEC |
5634 SD_SHARE_CPUCAPACITY |
5635 SD_SHARE_PKG_RESOURCES |
5636 SD_SHARE_POWERDOMAIN)) {
5637 if (sd->groups != sd->groups->next)
5638 return 0;
5639 }
5640
5641 /* Following flags don't use groups */
5642 if (sd->flags & (SD_WAKE_AFFINE))
5643 return 0;
5644
5645 return 1;
5646}
5647
5648static int
5649sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5650{
5651 unsigned long cflags = sd->flags, pflags = parent->flags;
5652
5653 if (sd_degenerate(parent))
5654 return 1;
5655
5656 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5657 return 0;
5658
5659 /* Flags needing groups don't count if only 1 group in parent */
5660 if (parent->groups == parent->groups->next) {
5661 pflags &= ~(SD_LOAD_BALANCE |
5662 SD_BALANCE_NEWIDLE |
5663 SD_BALANCE_FORK |
5664 SD_BALANCE_EXEC |
5665 SD_SHARE_CPUCAPACITY |
5666 SD_SHARE_PKG_RESOURCES |
5667 SD_PREFER_SIBLING |
5668 SD_SHARE_POWERDOMAIN);
5669 if (nr_node_ids == 1)
5670 pflags &= ~SD_SERIALIZE;
5671 }
5672 if (~cflags & pflags)
5673 return 0;
5674
5675 return 1;
5676}
5677
5678static void free_rootdomain(struct rcu_head *rcu)
5679{
5680 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5681
5682 cpupri_cleanup(&rd->cpupri);
5683 cpudl_cleanup(&rd->cpudl);
5684 free_cpumask_var(rd->dlo_mask);
5685 free_cpumask_var(rd->rto_mask);
5686 free_cpumask_var(rd->online);
5687 free_cpumask_var(rd->span);
5688 kfree(rd);
5689}
5690
5691static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5692{
5693 struct root_domain *old_rd = NULL;
5694 unsigned long flags;
5695
5696 raw_spin_lock_irqsave(&rq->lock, flags);
5697
5698 if (rq->rd) {
5699 old_rd = rq->rd;
5700
5701 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5702 set_rq_offline(rq);
5703
5704 cpumask_clear_cpu(rq->cpu, old_rd->span);
5705
5706 /*
5707 * If we dont want to free the old_rd yet then
5708 * set old_rd to NULL to skip the freeing later
5709 * in this function:
5710 */
5711 if (!atomic_dec_and_test(&old_rd->refcount))
5712 old_rd = NULL;
5713 }
5714
5715 atomic_inc(&rd->refcount);
5716 rq->rd = rd;
5717
5718 cpumask_set_cpu(rq->cpu, rd->span);
5719 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5720 set_rq_online(rq);
5721
5722 raw_spin_unlock_irqrestore(&rq->lock, flags);
5723
5724 if (old_rd)
5725 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5726}
5727
5728static int init_rootdomain(struct root_domain *rd)
5729{
5730 memset(rd, 0, sizeof(*rd));
5731
5732 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5733 goto out;
5734 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5735 goto free_span;
5736 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5737 goto free_online;
5738 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5739 goto free_dlo_mask;
5740
5741 init_dl_bw(&rd->dl_bw);
5742 if (cpudl_init(&rd->cpudl) != 0)
5743 goto free_dlo_mask;
5744
5745 if (cpupri_init(&rd->cpupri) != 0)
5746 goto free_rto_mask;
5747 return 0;
5748
5749free_rto_mask:
5750 free_cpumask_var(rd->rto_mask);
5751free_dlo_mask:
5752 free_cpumask_var(rd->dlo_mask);
5753free_online:
5754 free_cpumask_var(rd->online);
5755free_span:
5756 free_cpumask_var(rd->span);
5757out:
5758 return -ENOMEM;
5759}
5760
5761/*
5762 * By default the system creates a single root-domain with all cpus as
5763 * members (mimicking the global state we have today).
5764 */
5765struct root_domain def_root_domain;
5766
5767static void init_defrootdomain(void)
5768{
5769 init_rootdomain(&def_root_domain);
5770
5771 atomic_set(&def_root_domain.refcount, 1);
5772}
5773
5774static struct root_domain *alloc_rootdomain(void)
5775{
5776 struct root_domain *rd;
5777
5778 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5779 if (!rd)
5780 return NULL;
5781
5782 if (init_rootdomain(rd) != 0) {
5783 kfree(rd);
5784 return NULL;
5785 }
5786
5787 return rd;
5788}
5789
5790static void free_sched_groups(struct sched_group *sg, int free_sgc)
5791{
5792 struct sched_group *tmp, *first;
5793
5794 if (!sg)
5795 return;
5796
5797 first = sg;
5798 do {
5799 tmp = sg->next;
5800
5801 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5802 kfree(sg->sgc);
5803
5804 kfree(sg);
5805 sg = tmp;
5806 } while (sg != first);
5807}
5808
5809static void free_sched_domain(struct rcu_head *rcu)
5810{
5811 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5812
5813 /*
5814 * If its an overlapping domain it has private groups, iterate and
5815 * nuke them all.
5816 */
5817 if (sd->flags & SD_OVERLAP) {
5818 free_sched_groups(sd->groups, 1);
5819 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5820 kfree(sd->groups->sgc);
5821 kfree(sd->groups);
5822 }
5823 kfree(sd);
5824}
5825
5826static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5827{
5828 call_rcu(&sd->rcu, free_sched_domain);
5829}
5830
5831static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5832{
5833 for (; sd; sd = sd->parent)
5834 destroy_sched_domain(sd, cpu);
5835}
5836
5837/*
5838 * Keep a special pointer to the highest sched_domain that has
5839 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5840 * allows us to avoid some pointer chasing select_idle_sibling().
5841 *
5842 * Also keep a unique ID per domain (we use the first cpu number in
5843 * the cpumask of the domain), this allows us to quickly tell if
5844 * two cpus are in the same cache domain, see cpus_share_cache().
5845 */
5846DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5847DEFINE_PER_CPU(int, sd_llc_size);
5848DEFINE_PER_CPU(int, sd_llc_id);
5849DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5850DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5851DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5852
5853static void update_top_cache_domain(int cpu)
5854{
5855 struct sched_domain *sd;
5856 struct sched_domain *busy_sd = NULL;
5857 int id = cpu;
5858 int size = 1;
5859
5860 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5861 if (sd) {
5862 id = cpumask_first(sched_domain_span(sd));
5863 size = cpumask_weight(sched_domain_span(sd));
5864 busy_sd = sd->parent; /* sd_busy */
5865 }
5866 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5867
5868 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5869 per_cpu(sd_llc_size, cpu) = size;
5870 per_cpu(sd_llc_id, cpu) = id;
5871
5872 sd = lowest_flag_domain(cpu, SD_NUMA);
5873 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5874
5875 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5876 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5877}
5878
5879/*
5880 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5881 * hold the hotplug lock.
5882 */
5883static void
5884cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5885{
5886 struct rq *rq = cpu_rq(cpu);
5887 struct sched_domain *tmp;
5888
5889 /* Remove the sched domains which do not contribute to scheduling. */
5890 for (tmp = sd; tmp; ) {
5891 struct sched_domain *parent = tmp->parent;
5892 if (!parent)
5893 break;
5894
5895 if (sd_parent_degenerate(tmp, parent)) {
5896 tmp->parent = parent->parent;
5897 if (parent->parent)
5898 parent->parent->child = tmp;
5899 /*
5900 * Transfer SD_PREFER_SIBLING down in case of a
5901 * degenerate parent; the spans match for this
5902 * so the property transfers.
5903 */
5904 if (parent->flags & SD_PREFER_SIBLING)
5905 tmp->flags |= SD_PREFER_SIBLING;
5906 destroy_sched_domain(parent, cpu);
5907 } else
5908 tmp = tmp->parent;
5909 }
5910
5911 if (sd && sd_degenerate(sd)) {
5912 tmp = sd;
5913 sd = sd->parent;
5914 destroy_sched_domain(tmp, cpu);
5915 if (sd)
5916 sd->child = NULL;
5917 }
5918
5919 sched_domain_debug(sd, cpu);
5920
5921 rq_attach_root(rq, rd);
5922 tmp = rq->sd;
5923 rcu_assign_pointer(rq->sd, sd);
5924 destroy_sched_domains(tmp, cpu);
5925
5926 update_top_cache_domain(cpu);
5927}
5928
5929/* Setup the mask of cpus configured for isolated domains */
5930static int __init isolated_cpu_setup(char *str)
5931{
5932 int ret;
5933
5934 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5935 ret = cpulist_parse(str, cpu_isolated_map);
5936 if (ret) {
5937 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids);
5938 return 0;
5939 }
5940 return 1;
5941}
5942__setup("isolcpus=", isolated_cpu_setup);
5943
5944struct s_data {
5945 struct sched_domain ** __percpu sd;
5946 struct root_domain *rd;
5947};
5948
5949enum s_alloc {
5950 sa_rootdomain,
5951 sa_sd,
5952 sa_sd_storage,
5953 sa_none,
5954};
5955
5956/*
5957 * Build an iteration mask that can exclude certain CPUs from the upwards
5958 * domain traversal.
5959 *
5960 * Asymmetric node setups can result in situations where the domain tree is of
5961 * unequal depth, make sure to skip domains that already cover the entire
5962 * range.
5963 *
5964 * In that case build_sched_domains() will have terminated the iteration early
5965 * and our sibling sd spans will be empty. Domains should always include the
5966 * cpu they're built on, so check that.
5967 *
5968 */
5969static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5970{
5971 const struct cpumask *span = sched_domain_span(sd);
5972 struct sd_data *sdd = sd->private;
5973 struct sched_domain *sibling;
5974 int i;
5975
5976 for_each_cpu(i, span) {
5977 sibling = *per_cpu_ptr(sdd->sd, i);
5978 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5979 continue;
5980
5981 cpumask_set_cpu(i, sched_group_mask(sg));
5982 }
5983}
5984
5985/*
5986 * Return the canonical balance cpu for this group, this is the first cpu
5987 * of this group that's also in the iteration mask.
5988 */
5989int group_balance_cpu(struct sched_group *sg)
5990{
5991 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5992}
5993
5994static int
5995build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5996{
5997 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5998 const struct cpumask *span = sched_domain_span(sd);
5999 struct cpumask *covered = sched_domains_tmpmask;
6000 struct sd_data *sdd = sd->private;
6001 struct sched_domain *sibling;
6002 int i;
6003
6004 cpumask_clear(covered);
6005
6006 for_each_cpu(i, span) {
6007 struct cpumask *sg_span;
6008
6009 if (cpumask_test_cpu(i, covered))
6010 continue;
6011
6012 sibling = *per_cpu_ptr(sdd->sd, i);
6013
6014 /* See the comment near build_group_mask(). */
6015 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6016 continue;
6017
6018 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6019 GFP_KERNEL, cpu_to_node(cpu));
6020
6021 if (!sg)
6022 goto fail;
6023
6024 sg_span = sched_group_cpus(sg);
6025 if (sibling->child)
6026 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6027 else
6028 cpumask_set_cpu(i, sg_span);
6029
6030 cpumask_or(covered, covered, sg_span);
6031
6032 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6033 if (atomic_inc_return(&sg->sgc->ref) == 1)
6034 build_group_mask(sd, sg);
6035
6036 /*
6037 * Initialize sgc->capacity such that even if we mess up the
6038 * domains and no possible iteration will get us here, we won't
6039 * die on a /0 trap.
6040 */
6041 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6042
6043 /*
6044 * Make sure the first group of this domain contains the
6045 * canonical balance cpu. Otherwise the sched_domain iteration
6046 * breaks. See update_sg_lb_stats().
6047 */
6048 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6049 group_balance_cpu(sg) == cpu)
6050 groups = sg;
6051
6052 if (!first)
6053 first = sg;
6054 if (last)
6055 last->next = sg;
6056 last = sg;
6057 last->next = first;
6058 }
6059 sd->groups = groups;
6060
6061 return 0;
6062
6063fail:
6064 free_sched_groups(first, 0);
6065
6066 return -ENOMEM;
6067}
6068
6069static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6070{
6071 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6072 struct sched_domain *child = sd->child;
6073
6074 if (child)
6075 cpu = cpumask_first(sched_domain_span(child));
6076
6077 if (sg) {
6078 *sg = *per_cpu_ptr(sdd->sg, cpu);
6079 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6080 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6081 }
6082
6083 return cpu;
6084}
6085
6086/*
6087 * build_sched_groups will build a circular linked list of the groups
6088 * covered by the given span, and will set each group's ->cpumask correctly,
6089 * and ->cpu_capacity to 0.
6090 *
6091 * Assumes the sched_domain tree is fully constructed
6092 */
6093static int
6094build_sched_groups(struct sched_domain *sd, int cpu)
6095{
6096 struct sched_group *first = NULL, *last = NULL;
6097 struct sd_data *sdd = sd->private;
6098 const struct cpumask *span = sched_domain_span(sd);
6099 struct cpumask *covered;
6100 int i;
6101
6102 get_group(cpu, sdd, &sd->groups);
6103 atomic_inc(&sd->groups->ref);
6104
6105 if (cpu != cpumask_first(span))
6106 return 0;
6107
6108 lockdep_assert_held(&sched_domains_mutex);
6109 covered = sched_domains_tmpmask;
6110
6111 cpumask_clear(covered);
6112
6113 for_each_cpu(i, span) {
6114 struct sched_group *sg;
6115 int group, j;
6116
6117 if (cpumask_test_cpu(i, covered))
6118 continue;
6119
6120 group = get_group(i, sdd, &sg);
6121 cpumask_setall(sched_group_mask(sg));
6122
6123 for_each_cpu(j, span) {
6124 if (get_group(j, sdd, NULL) != group)
6125 continue;
6126
6127 cpumask_set_cpu(j, covered);
6128 cpumask_set_cpu(j, sched_group_cpus(sg));
6129 }
6130
6131 if (!first)
6132 first = sg;
6133 if (last)
6134 last->next = sg;
6135 last = sg;
6136 }
6137 last->next = first;
6138
6139 return 0;
6140}
6141
6142/*
6143 * Initialize sched groups cpu_capacity.
6144 *
6145 * cpu_capacity indicates the capacity of sched group, which is used while
6146 * distributing the load between different sched groups in a sched domain.
6147 * Typically cpu_capacity for all the groups in a sched domain will be same
6148 * unless there are asymmetries in the topology. If there are asymmetries,
6149 * group having more cpu_capacity will pickup more load compared to the
6150 * group having less cpu_capacity.
6151 */
6152static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6153{
6154 struct sched_group *sg = sd->groups;
6155
6156 WARN_ON(!sg);
6157
6158 do {
6159 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6160 sg = sg->next;
6161 } while (sg != sd->groups);
6162
6163 if (cpu != group_balance_cpu(sg))
6164 return;
6165
6166 update_group_capacity(sd, cpu);
6167 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6168}
6169
6170/*
6171 * Initializers for schedule domains
6172 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6173 */
6174
6175static int default_relax_domain_level = -1;
6176int sched_domain_level_max;
6177
6178static int __init setup_relax_domain_level(char *str)
6179{
6180 if (kstrtoint(str, 0, &default_relax_domain_level))
6181 pr_warn("Unable to set relax_domain_level\n");
6182
6183 return 1;
6184}
6185__setup("relax_domain_level=", setup_relax_domain_level);
6186
6187static void set_domain_attribute(struct sched_domain *sd,
6188 struct sched_domain_attr *attr)
6189{
6190 int request;
6191
6192 if (!attr || attr->relax_domain_level < 0) {
6193 if (default_relax_domain_level < 0)
6194 return;
6195 else
6196 request = default_relax_domain_level;
6197 } else
6198 request = attr->relax_domain_level;
6199 if (request < sd->level) {
6200 /* turn off idle balance on this domain */
6201 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6202 } else {
6203 /* turn on idle balance on this domain */
6204 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6205 }
6206}
6207
6208static void __sdt_free(const struct cpumask *cpu_map);
6209static int __sdt_alloc(const struct cpumask *cpu_map);
6210
6211static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6212 const struct cpumask *cpu_map)
6213{
6214 switch (what) {
6215 case sa_rootdomain:
6216 if (!atomic_read(&d->rd->refcount))
6217 free_rootdomain(&d->rd->rcu); /* fall through */
6218 case sa_sd:
6219 free_percpu(d->sd); /* fall through */
6220 case sa_sd_storage:
6221 __sdt_free(cpu_map); /* fall through */
6222 case sa_none:
6223 break;
6224 }
6225}
6226
6227static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6228 const struct cpumask *cpu_map)
6229{
6230 memset(d, 0, sizeof(*d));
6231
6232 if (__sdt_alloc(cpu_map))
6233 return sa_sd_storage;
6234 d->sd = alloc_percpu(struct sched_domain *);
6235 if (!d->sd)
6236 return sa_sd_storage;
6237 d->rd = alloc_rootdomain();
6238 if (!d->rd)
6239 return sa_sd;
6240 return sa_rootdomain;
6241}
6242
6243/*
6244 * NULL the sd_data elements we've used to build the sched_domain and
6245 * sched_group structure so that the subsequent __free_domain_allocs()
6246 * will not free the data we're using.
6247 */
6248static void claim_allocations(int cpu, struct sched_domain *sd)
6249{
6250 struct sd_data *sdd = sd->private;
6251
6252 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6253 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6254
6255 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6256 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6257
6258 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6259 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6260}
6261
6262#ifdef CONFIG_NUMA
6263static int sched_domains_numa_levels;
6264enum numa_topology_type sched_numa_topology_type;
6265static int *sched_domains_numa_distance;
6266int sched_max_numa_distance;
6267static struct cpumask ***sched_domains_numa_masks;
6268static int sched_domains_curr_level;
6269#endif
6270
6271/*
6272 * SD_flags allowed in topology descriptions.
6273 *
6274 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6275 * SD_SHARE_PKG_RESOURCES - describes shared caches
6276 * SD_NUMA - describes NUMA topologies
6277 * SD_SHARE_POWERDOMAIN - describes shared power domain
6278 *
6279 * Odd one out:
6280 * SD_ASYM_PACKING - describes SMT quirks
6281 */
6282#define TOPOLOGY_SD_FLAGS \
6283 (SD_SHARE_CPUCAPACITY | \
6284 SD_SHARE_PKG_RESOURCES | \
6285 SD_NUMA | \
6286 SD_ASYM_PACKING | \
6287 SD_SHARE_POWERDOMAIN)
6288
6289static struct sched_domain *
6290sd_init(struct sched_domain_topology_level *tl, int cpu)
6291{
6292 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6293 int sd_weight, sd_flags = 0;
6294
6295#ifdef CONFIG_NUMA
6296 /*
6297 * Ugly hack to pass state to sd_numa_mask()...
6298 */
6299 sched_domains_curr_level = tl->numa_level;
6300#endif
6301
6302 sd_weight = cpumask_weight(tl->mask(cpu));
6303
6304 if (tl->sd_flags)
6305 sd_flags = (*tl->sd_flags)();
6306 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6307 "wrong sd_flags in topology description\n"))
6308 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6309
6310 *sd = (struct sched_domain){
6311 .min_interval = sd_weight,
6312 .max_interval = 2*sd_weight,
6313 .busy_factor = 32,
6314 .imbalance_pct = 125,
6315
6316 .cache_nice_tries = 0,
6317 .busy_idx = 0,
6318 .idle_idx = 0,
6319 .newidle_idx = 0,
6320 .wake_idx = 0,
6321 .forkexec_idx = 0,
6322
6323 .flags = 1*SD_LOAD_BALANCE
6324 | 1*SD_BALANCE_NEWIDLE
6325 | 1*SD_BALANCE_EXEC
6326 | 1*SD_BALANCE_FORK
6327 | 0*SD_BALANCE_WAKE
6328 | 1*SD_WAKE_AFFINE
6329 | 0*SD_SHARE_CPUCAPACITY
6330 | 0*SD_SHARE_PKG_RESOURCES
6331 | 0*SD_SERIALIZE
6332 | 0*SD_PREFER_SIBLING
6333 | 0*SD_NUMA
6334 | sd_flags
6335 ,
6336
6337 .last_balance = jiffies,
6338 .balance_interval = sd_weight,
6339 .smt_gain = 0,
6340 .max_newidle_lb_cost = 0,
6341 .next_decay_max_lb_cost = jiffies,
6342#ifdef CONFIG_SCHED_DEBUG
6343 .name = tl->name,
6344#endif
6345 };
6346
6347 /*
6348 * Convert topological properties into behaviour.
6349 */
6350
6351 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6352 sd->flags |= SD_PREFER_SIBLING;
6353 sd->imbalance_pct = 110;
6354 sd->smt_gain = 1178; /* ~15% */
6355
6356 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6357 sd->imbalance_pct = 117;
6358 sd->cache_nice_tries = 1;
6359 sd->busy_idx = 2;
6360
6361#ifdef CONFIG_NUMA
6362 } else if (sd->flags & SD_NUMA) {
6363 sd->cache_nice_tries = 2;
6364 sd->busy_idx = 3;
6365 sd->idle_idx = 2;
6366
6367 sd->flags |= SD_SERIALIZE;
6368 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6369 sd->flags &= ~(SD_BALANCE_EXEC |
6370 SD_BALANCE_FORK |
6371 SD_WAKE_AFFINE);
6372 }
6373
6374#endif
6375 } else {
6376 sd->flags |= SD_PREFER_SIBLING;
6377 sd->cache_nice_tries = 1;
6378 sd->busy_idx = 2;
6379 sd->idle_idx = 1;
6380 }
6381
6382 sd->private = &tl->data;
6383
6384 return sd;
6385}
6386
6387/*
6388 * Topology list, bottom-up.
6389 */
6390static struct sched_domain_topology_level default_topology[] = {
6391#ifdef CONFIG_SCHED_SMT
6392 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6393#endif
6394#ifdef CONFIG_SCHED_MC
6395 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6396#endif
6397 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6398 { NULL, },
6399};
6400
6401static struct sched_domain_topology_level *sched_domain_topology =
6402 default_topology;
6403
6404#define for_each_sd_topology(tl) \
6405 for (tl = sched_domain_topology; tl->mask; tl++)
6406
6407void set_sched_topology(struct sched_domain_topology_level *tl)
6408{
6409 sched_domain_topology = tl;
6410}
6411
6412#ifdef CONFIG_NUMA
6413
6414static const struct cpumask *sd_numa_mask(int cpu)
6415{
6416 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6417}
6418
6419static void sched_numa_warn(const char *str)
6420{
6421 static int done = false;
6422 int i,j;
6423
6424 if (done)
6425 return;
6426
6427 done = true;
6428
6429 printk(KERN_WARNING "ERROR: %s\n\n", str);
6430
6431 for (i = 0; i < nr_node_ids; i++) {
6432 printk(KERN_WARNING " ");
6433 for (j = 0; j < nr_node_ids; j++)
6434 printk(KERN_CONT "%02d ", node_distance(i,j));
6435 printk(KERN_CONT "\n");
6436 }
6437 printk(KERN_WARNING "\n");
6438}
6439
6440bool find_numa_distance(int distance)
6441{
6442 int i;
6443
6444 if (distance == node_distance(0, 0))
6445 return true;
6446
6447 for (i = 0; i < sched_domains_numa_levels; i++) {
6448 if (sched_domains_numa_distance[i] == distance)
6449 return true;
6450 }
6451
6452 return false;
6453}
6454
6455/*
6456 * A system can have three types of NUMA topology:
6457 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6458 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6459 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6460 *
6461 * The difference between a glueless mesh topology and a backplane
6462 * topology lies in whether communication between not directly
6463 * connected nodes goes through intermediary nodes (where programs
6464 * could run), or through backplane controllers. This affects
6465 * placement of programs.
6466 *
6467 * The type of topology can be discerned with the following tests:
6468 * - If the maximum distance between any nodes is 1 hop, the system
6469 * is directly connected.
6470 * - If for two nodes A and B, located N > 1 hops away from each other,
6471 * there is an intermediary node C, which is < N hops away from both
6472 * nodes A and B, the system is a glueless mesh.
6473 */
6474static void init_numa_topology_type(void)
6475{
6476 int a, b, c, n;
6477
6478 n = sched_max_numa_distance;
6479
6480 if (sched_domains_numa_levels <= 1) {
6481 sched_numa_topology_type = NUMA_DIRECT;
6482 return;
6483 }
6484
6485 for_each_online_node(a) {
6486 for_each_online_node(b) {
6487 /* Find two nodes furthest removed from each other. */
6488 if (node_distance(a, b) < n)
6489 continue;
6490
6491 /* Is there an intermediary node between a and b? */
6492 for_each_online_node(c) {
6493 if (node_distance(a, c) < n &&
6494 node_distance(b, c) < n) {
6495 sched_numa_topology_type =
6496 NUMA_GLUELESS_MESH;
6497 return;
6498 }
6499 }
6500
6501 sched_numa_topology_type = NUMA_BACKPLANE;
6502 return;
6503 }
6504 }
6505}
6506
6507static void sched_init_numa(void)
6508{
6509 int next_distance, curr_distance = node_distance(0, 0);
6510 struct sched_domain_topology_level *tl;
6511 int level = 0;
6512 int i, j, k;
6513
6514 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6515 if (!sched_domains_numa_distance)
6516 return;
6517
6518 /*
6519 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6520 * unique distances in the node_distance() table.
6521 *
6522 * Assumes node_distance(0,j) includes all distances in
6523 * node_distance(i,j) in order to avoid cubic time.
6524 */
6525 next_distance = curr_distance;
6526 for (i = 0; i < nr_node_ids; i++) {
6527 for (j = 0; j < nr_node_ids; j++) {
6528 for (k = 0; k < nr_node_ids; k++) {
6529 int distance = node_distance(i, k);
6530
6531 if (distance > curr_distance &&
6532 (distance < next_distance ||
6533 next_distance == curr_distance))
6534 next_distance = distance;
6535
6536 /*
6537 * While not a strong assumption it would be nice to know
6538 * about cases where if node A is connected to B, B is not
6539 * equally connected to A.
6540 */
6541 if (sched_debug() && node_distance(k, i) != distance)
6542 sched_numa_warn("Node-distance not symmetric");
6543
6544 if (sched_debug() && i && !find_numa_distance(distance))
6545 sched_numa_warn("Node-0 not representative");
6546 }
6547 if (next_distance != curr_distance) {
6548 sched_domains_numa_distance[level++] = next_distance;
6549 sched_domains_numa_levels = level;
6550 curr_distance = next_distance;
6551 } else break;
6552 }
6553
6554 /*
6555 * In case of sched_debug() we verify the above assumption.
6556 */
6557 if (!sched_debug())
6558 break;
6559 }
6560
6561 if (!level)
6562 return;
6563
6564 /*
6565 * 'level' contains the number of unique distances, excluding the
6566 * identity distance node_distance(i,i).
6567 *
6568 * The sched_domains_numa_distance[] array includes the actual distance
6569 * numbers.
6570 */
6571
6572 /*
6573 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6574 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6575 * the array will contain less then 'level' members. This could be
6576 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6577 * in other functions.
6578 *
6579 * We reset it to 'level' at the end of this function.
6580 */
6581 sched_domains_numa_levels = 0;
6582
6583 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6584 if (!sched_domains_numa_masks)
6585 return;
6586
6587 /*
6588 * Now for each level, construct a mask per node which contains all
6589 * cpus of nodes that are that many hops away from us.
6590 */
6591 for (i = 0; i < level; i++) {
6592 sched_domains_numa_masks[i] =
6593 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6594 if (!sched_domains_numa_masks[i])
6595 return;
6596
6597 for (j = 0; j < nr_node_ids; j++) {
6598 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6599 if (!mask)
6600 return;
6601
6602 sched_domains_numa_masks[i][j] = mask;
6603
6604 for_each_node(k) {
6605 if (node_distance(j, k) > sched_domains_numa_distance[i])
6606 continue;
6607
6608 cpumask_or(mask, mask, cpumask_of_node(k));
6609 }
6610 }
6611 }
6612
6613 /* Compute default topology size */
6614 for (i = 0; sched_domain_topology[i].mask; i++);
6615
6616 tl = kzalloc((i + level + 1) *
6617 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6618 if (!tl)
6619 return;
6620
6621 /*
6622 * Copy the default topology bits..
6623 */
6624 for (i = 0; sched_domain_topology[i].mask; i++)
6625 tl[i] = sched_domain_topology[i];
6626
6627 /*
6628 * .. and append 'j' levels of NUMA goodness.
6629 */
6630 for (j = 0; j < level; i++, j++) {
6631 tl[i] = (struct sched_domain_topology_level){
6632 .mask = sd_numa_mask,
6633 .sd_flags = cpu_numa_flags,
6634 .flags = SDTL_OVERLAP,
6635 .numa_level = j,
6636 SD_INIT_NAME(NUMA)
6637 };
6638 }
6639
6640 sched_domain_topology = tl;
6641
6642 sched_domains_numa_levels = level;
6643 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6644
6645 init_numa_topology_type();
6646}
6647
6648static void sched_domains_numa_masks_set(int cpu)
6649{
6650 int i, j;
6651 int node = cpu_to_node(cpu);
6652
6653 for (i = 0; i < sched_domains_numa_levels; i++) {
6654 for (j = 0; j < nr_node_ids; j++) {
6655 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6656 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6657 }
6658 }
6659}
6660
6661static void sched_domains_numa_masks_clear(int cpu)
6662{
6663 int i, j;
6664 for (i = 0; i < sched_domains_numa_levels; i++) {
6665 for (j = 0; j < nr_node_ids; j++)
6666 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6667 }
6668}
6669
6670/*
6671 * Update sched_domains_numa_masks[level][node] array when new cpus
6672 * are onlined.
6673 */
6674static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6675 unsigned long action,
6676 void *hcpu)
6677{
6678 int cpu = (long)hcpu;
6679
6680 switch (action & ~CPU_TASKS_FROZEN) {
6681 case CPU_ONLINE:
6682 sched_domains_numa_masks_set(cpu);
6683 break;
6684
6685 case CPU_DEAD:
6686 sched_domains_numa_masks_clear(cpu);
6687 break;
6688
6689 default:
6690 return NOTIFY_DONE;
6691 }
6692
6693 return NOTIFY_OK;
6694}
6695#else
6696static inline void sched_init_numa(void)
6697{
6698}
6699
6700static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6701 unsigned long action,
6702 void *hcpu)
6703{
6704 return 0;
6705}
6706#endif /* CONFIG_NUMA */
6707
6708static int __sdt_alloc(const struct cpumask *cpu_map)
6709{
6710 struct sched_domain_topology_level *tl;
6711 int j;
6712
6713 for_each_sd_topology(tl) {
6714 struct sd_data *sdd = &tl->data;
6715
6716 sdd->sd = alloc_percpu(struct sched_domain *);
6717 if (!sdd->sd)
6718 return -ENOMEM;
6719
6720 sdd->sg = alloc_percpu(struct sched_group *);
6721 if (!sdd->sg)
6722 return -ENOMEM;
6723
6724 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6725 if (!sdd->sgc)
6726 return -ENOMEM;
6727
6728 for_each_cpu(j, cpu_map) {
6729 struct sched_domain *sd;
6730 struct sched_group *sg;
6731 struct sched_group_capacity *sgc;
6732
6733 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6734 GFP_KERNEL, cpu_to_node(j));
6735 if (!sd)
6736 return -ENOMEM;
6737
6738 *per_cpu_ptr(sdd->sd, j) = sd;
6739
6740 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6741 GFP_KERNEL, cpu_to_node(j));
6742 if (!sg)
6743 return -ENOMEM;
6744
6745 sg->next = sg;
6746
6747 *per_cpu_ptr(sdd->sg, j) = sg;
6748
6749 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6750 GFP_KERNEL, cpu_to_node(j));
6751 if (!sgc)
6752 return -ENOMEM;
6753
6754 *per_cpu_ptr(sdd->sgc, j) = sgc;
6755 }
6756 }
6757
6758 return 0;
6759}
6760
6761static void __sdt_free(const struct cpumask *cpu_map)
6762{
6763 struct sched_domain_topology_level *tl;
6764 int j;
6765
6766 for_each_sd_topology(tl) {
6767 struct sd_data *sdd = &tl->data;
6768
6769 for_each_cpu(j, cpu_map) {
6770 struct sched_domain *sd;
6771
6772 if (sdd->sd) {
6773 sd = *per_cpu_ptr(sdd->sd, j);
6774 if (sd && (sd->flags & SD_OVERLAP))
6775 free_sched_groups(sd->groups, 0);
6776 kfree(*per_cpu_ptr(sdd->sd, j));
6777 }
6778
6779 if (sdd->sg)
6780 kfree(*per_cpu_ptr(sdd->sg, j));
6781 if (sdd->sgc)
6782 kfree(*per_cpu_ptr(sdd->sgc, j));
6783 }
6784 free_percpu(sdd->sd);
6785 sdd->sd = NULL;
6786 free_percpu(sdd->sg);
6787 sdd->sg = NULL;
6788 free_percpu(sdd->sgc);
6789 sdd->sgc = NULL;
6790 }
6791}
6792
6793struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6794 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6795 struct sched_domain *child, int cpu)
6796{
6797 struct sched_domain *sd = sd_init(tl, cpu);
6798 if (!sd)
6799 return child;
6800
6801 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6802 if (child) {
6803 sd->level = child->level + 1;
6804 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6805 child->parent = sd;
6806 sd->child = child;
6807
6808 if (!cpumask_subset(sched_domain_span(child),
6809 sched_domain_span(sd))) {
6810 pr_err("BUG: arch topology borken\n");
6811#ifdef CONFIG_SCHED_DEBUG
6812 pr_err(" the %s domain not a subset of the %s domain\n",
6813 child->name, sd->name);
6814#endif
6815 /* Fixup, ensure @sd has at least @child cpus. */
6816 cpumask_or(sched_domain_span(sd),
6817 sched_domain_span(sd),
6818 sched_domain_span(child));
6819 }
6820
6821 }
6822 set_domain_attribute(sd, attr);
6823
6824 return sd;
6825}
6826
6827/*
6828 * Build sched domains for a given set of cpus and attach the sched domains
6829 * to the individual cpus
6830 */
6831static int build_sched_domains(const struct cpumask *cpu_map,
6832 struct sched_domain_attr *attr)
6833{
6834 enum s_alloc alloc_state;
6835 struct sched_domain *sd;
6836 struct s_data d;
6837 int i, ret = -ENOMEM;
6838
6839 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6840 if (alloc_state != sa_rootdomain)
6841 goto error;
6842
6843 /* Set up domains for cpus specified by the cpu_map. */
6844 for_each_cpu(i, cpu_map) {
6845 struct sched_domain_topology_level *tl;
6846
6847 sd = NULL;
6848 for_each_sd_topology(tl) {
6849 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6850 if (tl == sched_domain_topology)
6851 *per_cpu_ptr(d.sd, i) = sd;
6852 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6853 sd->flags |= SD_OVERLAP;
6854 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6855 break;
6856 }
6857 }
6858
6859 /* Build the groups for the domains */
6860 for_each_cpu(i, cpu_map) {
6861 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6862 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6863 if (sd->flags & SD_OVERLAP) {
6864 if (build_overlap_sched_groups(sd, i))
6865 goto error;
6866 } else {
6867 if (build_sched_groups(sd, i))
6868 goto error;
6869 }
6870 }
6871 }
6872
6873 /* Calculate CPU capacity for physical packages and nodes */
6874 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6875 if (!cpumask_test_cpu(i, cpu_map))
6876 continue;
6877
6878 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6879 claim_allocations(i, sd);
6880 init_sched_groups_capacity(i, sd);
6881 }
6882 }
6883
6884 /* Attach the domains */
6885 rcu_read_lock();
6886 for_each_cpu(i, cpu_map) {
6887 sd = *per_cpu_ptr(d.sd, i);
6888 cpu_attach_domain(sd, d.rd, i);
6889 }
6890 rcu_read_unlock();
6891
6892 ret = 0;
6893error:
6894 __free_domain_allocs(&d, alloc_state, cpu_map);
6895 return ret;
6896}
6897
6898static cpumask_var_t *doms_cur; /* current sched domains */
6899static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6900static struct sched_domain_attr *dattr_cur;
6901 /* attribues of custom domains in 'doms_cur' */
6902
6903/*
6904 * Special case: If a kmalloc of a doms_cur partition (array of
6905 * cpumask) fails, then fallback to a single sched domain,
6906 * as determined by the single cpumask fallback_doms.
6907 */
6908static cpumask_var_t fallback_doms;
6909
6910/*
6911 * arch_update_cpu_topology lets virtualized architectures update the
6912 * cpu core maps. It is supposed to return 1 if the topology changed
6913 * or 0 if it stayed the same.
6914 */
6915int __weak arch_update_cpu_topology(void)
6916{
6917 return 0;
6918}
6919
6920cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6921{
6922 int i;
6923 cpumask_var_t *doms;
6924
6925 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6926 if (!doms)
6927 return NULL;
6928 for (i = 0; i < ndoms; i++) {
6929 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6930 free_sched_domains(doms, i);
6931 return NULL;
6932 }
6933 }
6934 return doms;
6935}
6936
6937void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6938{
6939 unsigned int i;
6940 for (i = 0; i < ndoms; i++)
6941 free_cpumask_var(doms[i]);
6942 kfree(doms);
6943}
6944
6945/*
6946 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6947 * For now this just excludes isolated cpus, but could be used to
6948 * exclude other special cases in the future.
6949 */
6950static int init_sched_domains(const struct cpumask *cpu_map)
6951{
6952 int err;
6953
6954 arch_update_cpu_topology();
6955 ndoms_cur = 1;
6956 doms_cur = alloc_sched_domains(ndoms_cur);
6957 if (!doms_cur)
6958 doms_cur = &fallback_doms;
6959 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6960 err = build_sched_domains(doms_cur[0], NULL);
6961 register_sched_domain_sysctl();
6962
6963 return err;
6964}
6965
6966/*
6967 * Detach sched domains from a group of cpus specified in cpu_map
6968 * These cpus will now be attached to the NULL domain
6969 */
6970static void detach_destroy_domains(const struct cpumask *cpu_map)
6971{
6972 int i;
6973
6974 rcu_read_lock();
6975 for_each_cpu(i, cpu_map)
6976 cpu_attach_domain(NULL, &def_root_domain, i);
6977 rcu_read_unlock();
6978}
6979
6980/* handle null as "default" */
6981static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6982 struct sched_domain_attr *new, int idx_new)
6983{
6984 struct sched_domain_attr tmp;
6985
6986 /* fast path */
6987 if (!new && !cur)
6988 return 1;
6989
6990 tmp = SD_ATTR_INIT;
6991 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6992 new ? (new + idx_new) : &tmp,
6993 sizeof(struct sched_domain_attr));
6994}
6995
6996/*
6997 * Partition sched domains as specified by the 'ndoms_new'
6998 * cpumasks in the array doms_new[] of cpumasks. This compares
6999 * doms_new[] to the current sched domain partitioning, doms_cur[].
7000 * It destroys each deleted domain and builds each new domain.
7001 *
7002 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7003 * The masks don't intersect (don't overlap.) We should setup one
7004 * sched domain for each mask. CPUs not in any of the cpumasks will
7005 * not be load balanced. If the same cpumask appears both in the
7006 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7007 * it as it is.
7008 *
7009 * The passed in 'doms_new' should be allocated using
7010 * alloc_sched_domains. This routine takes ownership of it and will
7011 * free_sched_domains it when done with it. If the caller failed the
7012 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7013 * and partition_sched_domains() will fallback to the single partition
7014 * 'fallback_doms', it also forces the domains to be rebuilt.
7015 *
7016 * If doms_new == NULL it will be replaced with cpu_online_mask.
7017 * ndoms_new == 0 is a special case for destroying existing domains,
7018 * and it will not create the default domain.
7019 *
7020 * Call with hotplug lock held
7021 */
7022void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7023 struct sched_domain_attr *dattr_new)
7024{
7025 int i, j, n;
7026 int new_topology;
7027
7028 mutex_lock(&sched_domains_mutex);
7029
7030 /* always unregister in case we don't destroy any domains */
7031 unregister_sched_domain_sysctl();
7032
7033 /* Let architecture update cpu core mappings. */
7034 new_topology = arch_update_cpu_topology();
7035
7036 n = doms_new ? ndoms_new : 0;
7037
7038 /* Destroy deleted domains */
7039 for (i = 0; i < ndoms_cur; i++) {
7040 for (j = 0; j < n && !new_topology; j++) {
7041 if (cpumask_equal(doms_cur[i], doms_new[j])
7042 && dattrs_equal(dattr_cur, i, dattr_new, j))
7043 goto match1;
7044 }
7045 /* no match - a current sched domain not in new doms_new[] */
7046 detach_destroy_domains(doms_cur[i]);
7047match1:
7048 ;
7049 }
7050
7051 n = ndoms_cur;
7052 if (doms_new == NULL) {
7053 n = 0;
7054 doms_new = &fallback_doms;
7055 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7056 WARN_ON_ONCE(dattr_new);
7057 }
7058
7059 /* Build new domains */
7060 for (i = 0; i < ndoms_new; i++) {
7061 for (j = 0; j < n && !new_topology; j++) {
7062 if (cpumask_equal(doms_new[i], doms_cur[j])
7063 && dattrs_equal(dattr_new, i, dattr_cur, j))
7064 goto match2;
7065 }
7066 /* no match - add a new doms_new */
7067 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7068match2:
7069 ;
7070 }
7071
7072 /* Remember the new sched domains */
7073 if (doms_cur != &fallback_doms)
7074 free_sched_domains(doms_cur, ndoms_cur);
7075 kfree(dattr_cur); /* kfree(NULL) is safe */
7076 doms_cur = doms_new;
7077 dattr_cur = dattr_new;
7078 ndoms_cur = ndoms_new;
7079
7080 register_sched_domain_sysctl();
7081
7082 mutex_unlock(&sched_domains_mutex);
7083}
7084
7085static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7086
7087/*
7088 * Update cpusets according to cpu_active mask. If cpusets are
7089 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7090 * around partition_sched_domains().
7091 *
7092 * If we come here as part of a suspend/resume, don't touch cpusets because we
7093 * want to restore it back to its original state upon resume anyway.
7094 */
7095static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7096 void *hcpu)
7097{
7098 switch (action) {
7099 case CPU_ONLINE_FROZEN:
7100 case CPU_DOWN_FAILED_FROZEN:
7101
7102 /*
7103 * num_cpus_frozen tracks how many CPUs are involved in suspend
7104 * resume sequence. As long as this is not the last online
7105 * operation in the resume sequence, just build a single sched
7106 * domain, ignoring cpusets.
7107 */
7108 num_cpus_frozen--;
7109 if (likely(num_cpus_frozen)) {
7110 partition_sched_domains(1, NULL, NULL);
7111 break;
7112 }
7113
7114 /*
7115 * This is the last CPU online operation. So fall through and
7116 * restore the original sched domains by considering the
7117 * cpuset configurations.
7118 */
7119
7120 case CPU_ONLINE:
7121 cpuset_update_active_cpus(true);
7122 break;
7123 default:
7124 return NOTIFY_DONE;
7125 }
7126 return NOTIFY_OK;
7127}
7128
7129static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7130 void *hcpu)
7131{
7132 unsigned long flags;
7133 long cpu = (long)hcpu;
7134 struct dl_bw *dl_b;
7135 bool overflow;
7136 int cpus;
7137
7138 switch (action) {
7139 case CPU_DOWN_PREPARE:
7140 rcu_read_lock_sched();
7141 dl_b = dl_bw_of(cpu);
7142
7143 raw_spin_lock_irqsave(&dl_b->lock, flags);
7144 cpus = dl_bw_cpus(cpu);
7145 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7146 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7147
7148 rcu_read_unlock_sched();
7149
7150 if (overflow)
7151 return notifier_from_errno(-EBUSY);
7152 cpuset_update_active_cpus(false);
7153 break;
7154 case CPU_DOWN_PREPARE_FROZEN:
7155 num_cpus_frozen++;
7156 partition_sched_domains(1, NULL, NULL);
7157 break;
7158 default:
7159 return NOTIFY_DONE;
7160 }
7161 return NOTIFY_OK;
7162}
7163
7164void __init sched_init_smp(void)
7165{
7166 cpumask_var_t non_isolated_cpus;
7167
7168 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7169 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7170
7171 sched_init_numa();
7172
7173 /*
7174 * There's no userspace yet to cause hotplug operations; hence all the
7175 * cpu masks are stable and all blatant races in the below code cannot
7176 * happen.
7177 */
7178 mutex_lock(&sched_domains_mutex);
7179 init_sched_domains(cpu_active_mask);
7180 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7181 if (cpumask_empty(non_isolated_cpus))
7182 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7183 mutex_unlock(&sched_domains_mutex);
7184
7185 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7186 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7187 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7188
7189 init_hrtick();
7190
7191 /* Move init over to a non-isolated CPU */
7192 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7193 BUG();
7194 sched_init_granularity();
7195 free_cpumask_var(non_isolated_cpus);
7196
7197 init_sched_rt_class();
7198 init_sched_dl_class();
7199}
7200#else
7201void __init sched_init_smp(void)
7202{
7203 sched_init_granularity();
7204}
7205#endif /* CONFIG_SMP */
7206
7207int in_sched_functions(unsigned long addr)
7208{
7209 return in_lock_functions(addr) ||
7210 (addr >= (unsigned long)__sched_text_start
7211 && addr < (unsigned long)__sched_text_end);
7212}
7213
7214#ifdef CONFIG_CGROUP_SCHED
7215/*
7216 * Default task group.
7217 * Every task in system belongs to this group at bootup.
7218 */
7219struct task_group root_task_group;
7220LIST_HEAD(task_groups);
7221
7222/* Cacheline aligned slab cache for task_group */
7223static struct kmem_cache *task_group_cache __read_mostly;
7224#endif
7225
7226DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7227
7228void __init sched_init(void)
7229{
7230 int i, j;
7231 unsigned long alloc_size = 0, ptr;
7232
7233#ifdef CONFIG_FAIR_GROUP_SCHED
7234 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7235#endif
7236#ifdef CONFIG_RT_GROUP_SCHED
7237 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7238#endif
7239 if (alloc_size) {
7240 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7241
7242#ifdef CONFIG_FAIR_GROUP_SCHED
7243 root_task_group.se = (struct sched_entity **)ptr;
7244 ptr += nr_cpu_ids * sizeof(void **);
7245
7246 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7247 ptr += nr_cpu_ids * sizeof(void **);
7248
7249#endif /* CONFIG_FAIR_GROUP_SCHED */
7250#ifdef CONFIG_RT_GROUP_SCHED
7251 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7252 ptr += nr_cpu_ids * sizeof(void **);
7253
7254 root_task_group.rt_rq = (struct rt_rq **)ptr;
7255 ptr += nr_cpu_ids * sizeof(void **);
7256
7257#endif /* CONFIG_RT_GROUP_SCHED */
7258 }
7259#ifdef CONFIG_CPUMASK_OFFSTACK
7260 for_each_possible_cpu(i) {
7261 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7262 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7263 }
7264#endif /* CONFIG_CPUMASK_OFFSTACK */
7265
7266 init_rt_bandwidth(&def_rt_bandwidth,
7267 global_rt_period(), global_rt_runtime());
7268 init_dl_bandwidth(&def_dl_bandwidth,
7269 global_rt_period(), global_rt_runtime());
7270
7271#ifdef CONFIG_SMP
7272 init_defrootdomain();
7273#endif
7274
7275#ifdef CONFIG_RT_GROUP_SCHED
7276 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7277 global_rt_period(), global_rt_runtime());
7278#endif /* CONFIG_RT_GROUP_SCHED */
7279
7280#ifdef CONFIG_CGROUP_SCHED
7281 task_group_cache = KMEM_CACHE(task_group, 0);
7282
7283 list_add(&root_task_group.list, &task_groups);
7284 INIT_LIST_HEAD(&root_task_group.children);
7285 INIT_LIST_HEAD(&root_task_group.siblings);
7286 autogroup_init(&init_task);
7287#endif /* CONFIG_CGROUP_SCHED */
7288
7289 for_each_possible_cpu(i) {
7290 struct rq *rq;
7291
7292 rq = cpu_rq(i);
7293 raw_spin_lock_init(&rq->lock);
7294 rq->nr_running = 0;
7295 rq->calc_load_active = 0;
7296 rq->calc_load_update = jiffies + LOAD_FREQ;
7297 init_cfs_rq(&rq->cfs);
7298 init_rt_rq(&rq->rt);
7299 init_dl_rq(&rq->dl);
7300#ifdef CONFIG_FAIR_GROUP_SCHED
7301 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7302 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7303 /*
7304 * How much cpu bandwidth does root_task_group get?
7305 *
7306 * In case of task-groups formed thr' the cgroup filesystem, it
7307 * gets 100% of the cpu resources in the system. This overall
7308 * system cpu resource is divided among the tasks of
7309 * root_task_group and its child task-groups in a fair manner,
7310 * based on each entity's (task or task-group's) weight
7311 * (se->load.weight).
7312 *
7313 * In other words, if root_task_group has 10 tasks of weight
7314 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7315 * then A0's share of the cpu resource is:
7316 *
7317 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7318 *
7319 * We achieve this by letting root_task_group's tasks sit
7320 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7321 */
7322 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7323 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7324#endif /* CONFIG_FAIR_GROUP_SCHED */
7325
7326 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7327#ifdef CONFIG_RT_GROUP_SCHED
7328 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7329#endif
7330
7331 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7332 rq->cpu_load[j] = 0;
7333
7334 rq->last_load_update_tick = jiffies;
7335
7336#ifdef CONFIG_SMP
7337 rq->sd = NULL;
7338 rq->rd = NULL;
7339 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7340 rq->balance_callback = NULL;
7341 rq->active_balance = 0;
7342 rq->next_balance = jiffies;
7343 rq->push_cpu = 0;
7344 rq->cpu = i;
7345 rq->online = 0;
7346 rq->idle_stamp = 0;
7347 rq->avg_idle = 2*sysctl_sched_migration_cost;
7348 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7349
7350 INIT_LIST_HEAD(&rq->cfs_tasks);
7351
7352 rq_attach_root(rq, &def_root_domain);
7353#ifdef CONFIG_NO_HZ_COMMON
7354 rq->nohz_flags = 0;
7355#endif
7356#ifdef CONFIG_NO_HZ_FULL
7357 rq->last_sched_tick = 0;
7358#endif
7359#endif
7360 init_rq_hrtick(rq);
7361 atomic_set(&rq->nr_iowait, 0);
7362 }
7363
7364 set_load_weight(&init_task);
7365
7366#ifdef CONFIG_PREEMPT_NOTIFIERS
7367 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7368#endif
7369
7370 /*
7371 * The boot idle thread does lazy MMU switching as well:
7372 */
7373 atomic_inc(&init_mm.mm_count);
7374 enter_lazy_tlb(&init_mm, current);
7375
7376 /*
7377 * During early bootup we pretend to be a normal task:
7378 */
7379 current->sched_class = &fair_sched_class;
7380
7381 /*
7382 * Make us the idle thread. Technically, schedule() should not be
7383 * called from this thread, however somewhere below it might be,
7384 * but because we are the idle thread, we just pick up running again
7385 * when this runqueue becomes "idle".
7386 */
7387 init_idle(current, smp_processor_id());
7388
7389 calc_load_update = jiffies + LOAD_FREQ;
7390
7391#ifdef CONFIG_SMP
7392 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7393 /* May be allocated at isolcpus cmdline parse time */
7394 if (cpu_isolated_map == NULL)
7395 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7396 idle_thread_set_boot_cpu();
7397 set_cpu_rq_start_time();
7398#endif
7399 init_sched_fair_class();
7400
7401 scheduler_running = 1;
7402}
7403
7404#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7405static inline int preempt_count_equals(int preempt_offset)
7406{
7407 int nested = preempt_count() + rcu_preempt_depth();
7408
7409 return (nested == preempt_offset);
7410}
7411
7412void __might_sleep(const char *file, int line, int preempt_offset)
7413{
7414 /*
7415 * Blocking primitives will set (and therefore destroy) current->state,
7416 * since we will exit with TASK_RUNNING make sure we enter with it,
7417 * otherwise we will destroy state.
7418 */
7419 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7420 "do not call blocking ops when !TASK_RUNNING; "
7421 "state=%lx set at [<%p>] %pS\n",
7422 current->state,
7423 (void *)current->task_state_change,
7424 (void *)current->task_state_change);
7425
7426 ___might_sleep(file, line, preempt_offset);
7427}
7428EXPORT_SYMBOL(__might_sleep);
7429
7430void ___might_sleep(const char *file, int line, int preempt_offset)
7431{
7432 static unsigned long prev_jiffy; /* ratelimiting */
7433
7434 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7435 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7436 !is_idle_task(current)) ||
7437 system_state != SYSTEM_RUNNING || oops_in_progress)
7438 return;
7439 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7440 return;
7441 prev_jiffy = jiffies;
7442
7443 printk(KERN_ERR
7444 "BUG: sleeping function called from invalid context at %s:%d\n",
7445 file, line);
7446 printk(KERN_ERR
7447 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7448 in_atomic(), irqs_disabled(),
7449 current->pid, current->comm);
7450
7451 if (task_stack_end_corrupted(current))
7452 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7453
7454 debug_show_held_locks(current);
7455 if (irqs_disabled())
7456 print_irqtrace_events(current);
7457#ifdef CONFIG_DEBUG_PREEMPT
7458 if (!preempt_count_equals(preempt_offset)) {
7459 pr_err("Preemption disabled at:");
7460 print_ip_sym(current->preempt_disable_ip);
7461 pr_cont("\n");
7462 }
7463#endif
7464 dump_stack();
7465}
7466EXPORT_SYMBOL(___might_sleep);
7467#endif
7468
7469#ifdef CONFIG_MAGIC_SYSRQ
7470void normalize_rt_tasks(void)
7471{
7472 struct task_struct *g, *p;
7473 struct sched_attr attr = {
7474 .sched_policy = SCHED_NORMAL,
7475 };
7476
7477 read_lock(&tasklist_lock);
7478 for_each_process_thread(g, p) {
7479 /*
7480 * Only normalize user tasks:
7481 */
7482 if (p->flags & PF_KTHREAD)
7483 continue;
7484
7485 p->se.exec_start = 0;
7486#ifdef CONFIG_SCHEDSTATS
7487 p->se.statistics.wait_start = 0;
7488 p->se.statistics.sleep_start = 0;
7489 p->se.statistics.block_start = 0;
7490#endif
7491
7492 if (!dl_task(p) && !rt_task(p)) {
7493 /*
7494 * Renice negative nice level userspace
7495 * tasks back to 0:
7496 */
7497 if (task_nice(p) < 0)
7498 set_user_nice(p, 0);
7499 continue;
7500 }
7501
7502 __sched_setscheduler(p, &attr, false, false);
7503 }
7504 read_unlock(&tasklist_lock);
7505}
7506
7507#endif /* CONFIG_MAGIC_SYSRQ */
7508
7509#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7510/*
7511 * These functions are only useful for the IA64 MCA handling, or kdb.
7512 *
7513 * They can only be called when the whole system has been
7514 * stopped - every CPU needs to be quiescent, and no scheduling
7515 * activity can take place. Using them for anything else would
7516 * be a serious bug, and as a result, they aren't even visible
7517 * under any other configuration.
7518 */
7519
7520/**
7521 * curr_task - return the current task for a given cpu.
7522 * @cpu: the processor in question.
7523 *
7524 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7525 *
7526 * Return: The current task for @cpu.
7527 */
7528struct task_struct *curr_task(int cpu)
7529{
7530 return cpu_curr(cpu);
7531}
7532
7533#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7534
7535#ifdef CONFIG_IA64
7536/**
7537 * set_curr_task - set the current task for a given cpu.
7538 * @cpu: the processor in question.
7539 * @p: the task pointer to set.
7540 *
7541 * Description: This function must only be used when non-maskable interrupts
7542 * are serviced on a separate stack. It allows the architecture to switch the
7543 * notion of the current task on a cpu in a non-blocking manner. This function
7544 * must be called with all CPU's synchronized, and interrupts disabled, the
7545 * and caller must save the original value of the current task (see
7546 * curr_task() above) and restore that value before reenabling interrupts and
7547 * re-starting the system.
7548 *
7549 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7550 */
7551void set_curr_task(int cpu, struct task_struct *p)
7552{
7553 cpu_curr(cpu) = p;
7554}
7555
7556#endif
7557
7558#ifdef CONFIG_CGROUP_SCHED
7559/* task_group_lock serializes the addition/removal of task groups */
7560static DEFINE_SPINLOCK(task_group_lock);
7561
7562static void sched_free_group(struct task_group *tg)
7563{
7564 free_fair_sched_group(tg);
7565 free_rt_sched_group(tg);
7566 autogroup_free(tg);
7567 kmem_cache_free(task_group_cache, tg);
7568}
7569
7570/* allocate runqueue etc for a new task group */
7571struct task_group *sched_create_group(struct task_group *parent)
7572{
7573 struct task_group *tg;
7574
7575 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7576 if (!tg)
7577 return ERR_PTR(-ENOMEM);
7578
7579 if (!alloc_fair_sched_group(tg, parent))
7580 goto err;
7581
7582 if (!alloc_rt_sched_group(tg, parent))
7583 goto err;
7584
7585 return tg;
7586
7587err:
7588 sched_free_group(tg);
7589 return ERR_PTR(-ENOMEM);
7590}
7591
7592void sched_online_group(struct task_group *tg, struct task_group *parent)
7593{
7594 unsigned long flags;
7595
7596 spin_lock_irqsave(&task_group_lock, flags);
7597 list_add_rcu(&tg->list, &task_groups);
7598
7599 WARN_ON(!parent); /* root should already exist */
7600
7601 tg->parent = parent;
7602 INIT_LIST_HEAD(&tg->children);
7603 list_add_rcu(&tg->siblings, &parent->children);
7604 spin_unlock_irqrestore(&task_group_lock, flags);
7605}
7606
7607/* rcu callback to free various structures associated with a task group */
7608static void sched_free_group_rcu(struct rcu_head *rhp)
7609{
7610 /* now it should be safe to free those cfs_rqs */
7611 sched_free_group(container_of(rhp, struct task_group, rcu));
7612}
7613
7614void sched_destroy_group(struct task_group *tg)
7615{
7616 /* wait for possible concurrent references to cfs_rqs complete */
7617 call_rcu(&tg->rcu, sched_free_group_rcu);
7618}
7619
7620void sched_offline_group(struct task_group *tg)
7621{
7622 unsigned long flags;
7623
7624 /* end participation in shares distribution */
7625 unregister_fair_sched_group(tg);
7626
7627 spin_lock_irqsave(&task_group_lock, flags);
7628 list_del_rcu(&tg->list);
7629 list_del_rcu(&tg->siblings);
7630 spin_unlock_irqrestore(&task_group_lock, flags);
7631}
7632
7633/* change task's runqueue when it moves between groups.
7634 * The caller of this function should have put the task in its new group
7635 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7636 * reflect its new group.
7637 */
7638void sched_move_task(struct task_struct *tsk)
7639{
7640 struct task_group *tg;
7641 int queued, running;
7642 unsigned long flags;
7643 struct rq *rq;
7644
7645 rq = task_rq_lock(tsk, &flags);
7646
7647 running = task_current(rq, tsk);
7648 queued = task_on_rq_queued(tsk);
7649
7650 if (queued)
7651 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
7652 if (unlikely(running))
7653 put_prev_task(rq, tsk);
7654
7655 /*
7656 * All callers are synchronized by task_rq_lock(); we do not use RCU
7657 * which is pointless here. Thus, we pass "true" to task_css_check()
7658 * to prevent lockdep warnings.
7659 */
7660 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7661 struct task_group, css);
7662 tg = autogroup_task_group(tsk, tg);
7663 tsk->sched_task_group = tg;
7664
7665#ifdef CONFIG_FAIR_GROUP_SCHED
7666 if (tsk->sched_class->task_move_group)
7667 tsk->sched_class->task_move_group(tsk);
7668 else
7669#endif
7670 set_task_rq(tsk, task_cpu(tsk));
7671
7672 if (unlikely(running))
7673 tsk->sched_class->set_curr_task(rq);
7674 if (queued)
7675 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
7676
7677 task_rq_unlock(rq, tsk, &flags);
7678}
7679#endif /* CONFIG_CGROUP_SCHED */
7680
7681#ifdef CONFIG_RT_GROUP_SCHED
7682/*
7683 * Ensure that the real time constraints are schedulable.
7684 */
7685static DEFINE_MUTEX(rt_constraints_mutex);
7686
7687/* Must be called with tasklist_lock held */
7688static inline int tg_has_rt_tasks(struct task_group *tg)
7689{
7690 struct task_struct *g, *p;
7691
7692 /*
7693 * Autogroups do not have RT tasks; see autogroup_create().
7694 */
7695 if (task_group_is_autogroup(tg))
7696 return 0;
7697
7698 for_each_process_thread(g, p) {
7699 if (rt_task(p) && task_group(p) == tg)
7700 return 1;
7701 }
7702
7703 return 0;
7704}
7705
7706struct rt_schedulable_data {
7707 struct task_group *tg;
7708 u64 rt_period;
7709 u64 rt_runtime;
7710};
7711
7712static int tg_rt_schedulable(struct task_group *tg, void *data)
7713{
7714 struct rt_schedulable_data *d = data;
7715 struct task_group *child;
7716 unsigned long total, sum = 0;
7717 u64 period, runtime;
7718
7719 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7720 runtime = tg->rt_bandwidth.rt_runtime;
7721
7722 if (tg == d->tg) {
7723 period = d->rt_period;
7724 runtime = d->rt_runtime;
7725 }
7726
7727 /*
7728 * Cannot have more runtime than the period.
7729 */
7730 if (runtime > period && runtime != RUNTIME_INF)
7731 return -EINVAL;
7732
7733 /*
7734 * Ensure we don't starve existing RT tasks.
7735 */
7736 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7737 return -EBUSY;
7738
7739 total = to_ratio(period, runtime);
7740
7741 /*
7742 * Nobody can have more than the global setting allows.
7743 */
7744 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7745 return -EINVAL;
7746
7747 /*
7748 * The sum of our children's runtime should not exceed our own.
7749 */
7750 list_for_each_entry_rcu(child, &tg->children, siblings) {
7751 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7752 runtime = child->rt_bandwidth.rt_runtime;
7753
7754 if (child == d->tg) {
7755 period = d->rt_period;
7756 runtime = d->rt_runtime;
7757 }
7758
7759 sum += to_ratio(period, runtime);
7760 }
7761
7762 if (sum > total)
7763 return -EINVAL;
7764
7765 return 0;
7766}
7767
7768static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7769{
7770 int ret;
7771
7772 struct rt_schedulable_data data = {
7773 .tg = tg,
7774 .rt_period = period,
7775 .rt_runtime = runtime,
7776 };
7777
7778 rcu_read_lock();
7779 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7780 rcu_read_unlock();
7781
7782 return ret;
7783}
7784
7785static int tg_set_rt_bandwidth(struct task_group *tg,
7786 u64 rt_period, u64 rt_runtime)
7787{
7788 int i, err = 0;
7789
7790 /*
7791 * Disallowing the root group RT runtime is BAD, it would disallow the
7792 * kernel creating (and or operating) RT threads.
7793 */
7794 if (tg == &root_task_group && rt_runtime == 0)
7795 return -EINVAL;
7796
7797 /* No period doesn't make any sense. */
7798 if (rt_period == 0)
7799 return -EINVAL;
7800
7801 mutex_lock(&rt_constraints_mutex);
7802 read_lock(&tasklist_lock);
7803 err = __rt_schedulable(tg, rt_period, rt_runtime);
7804 if (err)
7805 goto unlock;
7806
7807 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7808 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7809 tg->rt_bandwidth.rt_runtime = rt_runtime;
7810
7811 for_each_possible_cpu(i) {
7812 struct rt_rq *rt_rq = tg->rt_rq[i];
7813
7814 raw_spin_lock(&rt_rq->rt_runtime_lock);
7815 rt_rq->rt_runtime = rt_runtime;
7816 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7817 }
7818 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7819unlock:
7820 read_unlock(&tasklist_lock);
7821 mutex_unlock(&rt_constraints_mutex);
7822
7823 return err;
7824}
7825
7826static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7827{
7828 u64 rt_runtime, rt_period;
7829
7830 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7831 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7832 if (rt_runtime_us < 0)
7833 rt_runtime = RUNTIME_INF;
7834
7835 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7836}
7837
7838static long sched_group_rt_runtime(struct task_group *tg)
7839{
7840 u64 rt_runtime_us;
7841
7842 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7843 return -1;
7844
7845 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7846 do_div(rt_runtime_us, NSEC_PER_USEC);
7847 return rt_runtime_us;
7848}
7849
7850static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
7851{
7852 u64 rt_runtime, rt_period;
7853
7854 rt_period = rt_period_us * NSEC_PER_USEC;
7855 rt_runtime = tg->rt_bandwidth.rt_runtime;
7856
7857 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7858}
7859
7860static long sched_group_rt_period(struct task_group *tg)
7861{
7862 u64 rt_period_us;
7863
7864 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7865 do_div(rt_period_us, NSEC_PER_USEC);
7866 return rt_period_us;
7867}
7868#endif /* CONFIG_RT_GROUP_SCHED */
7869
7870#ifdef CONFIG_RT_GROUP_SCHED
7871static int sched_rt_global_constraints(void)
7872{
7873 int ret = 0;
7874
7875 mutex_lock(&rt_constraints_mutex);
7876 read_lock(&tasklist_lock);
7877 ret = __rt_schedulable(NULL, 0, 0);
7878 read_unlock(&tasklist_lock);
7879 mutex_unlock(&rt_constraints_mutex);
7880
7881 return ret;
7882}
7883
7884static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7885{
7886 /* Don't accept realtime tasks when there is no way for them to run */
7887 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7888 return 0;
7889
7890 return 1;
7891}
7892
7893#else /* !CONFIG_RT_GROUP_SCHED */
7894static int sched_rt_global_constraints(void)
7895{
7896 unsigned long flags;
7897 int i, ret = 0;
7898
7899 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7900 for_each_possible_cpu(i) {
7901 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7902
7903 raw_spin_lock(&rt_rq->rt_runtime_lock);
7904 rt_rq->rt_runtime = global_rt_runtime();
7905 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7906 }
7907 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7908
7909 return ret;
7910}
7911#endif /* CONFIG_RT_GROUP_SCHED */
7912
7913static int sched_dl_global_validate(void)
7914{
7915 u64 runtime = global_rt_runtime();
7916 u64 period = global_rt_period();
7917 u64 new_bw = to_ratio(period, runtime);
7918 struct dl_bw *dl_b;
7919 int cpu, ret = 0;
7920 unsigned long flags;
7921
7922 /*
7923 * Here we want to check the bandwidth not being set to some
7924 * value smaller than the currently allocated bandwidth in
7925 * any of the root_domains.
7926 *
7927 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7928 * cycling on root_domains... Discussion on different/better
7929 * solutions is welcome!
7930 */
7931 for_each_possible_cpu(cpu) {
7932 rcu_read_lock_sched();
7933 dl_b = dl_bw_of(cpu);
7934
7935 raw_spin_lock_irqsave(&dl_b->lock, flags);
7936 if (new_bw < dl_b->total_bw)
7937 ret = -EBUSY;
7938 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7939
7940 rcu_read_unlock_sched();
7941
7942 if (ret)
7943 break;
7944 }
7945
7946 return ret;
7947}
7948
7949static void sched_dl_do_global(void)
7950{
7951 u64 new_bw = -1;
7952 struct dl_bw *dl_b;
7953 int cpu;
7954 unsigned long flags;
7955
7956 def_dl_bandwidth.dl_period = global_rt_period();
7957 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7958
7959 if (global_rt_runtime() != RUNTIME_INF)
7960 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7961
7962 /*
7963 * FIXME: As above...
7964 */
7965 for_each_possible_cpu(cpu) {
7966 rcu_read_lock_sched();
7967 dl_b = dl_bw_of(cpu);
7968
7969 raw_spin_lock_irqsave(&dl_b->lock, flags);
7970 dl_b->bw = new_bw;
7971 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7972
7973 rcu_read_unlock_sched();
7974 }
7975}
7976
7977static int sched_rt_global_validate(void)
7978{
7979 if (sysctl_sched_rt_period <= 0)
7980 return -EINVAL;
7981
7982 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7983 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7984 return -EINVAL;
7985
7986 return 0;
7987}
7988
7989static void sched_rt_do_global(void)
7990{
7991 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7992 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7993}
7994
7995int sched_rt_handler(struct ctl_table *table, int write,
7996 void __user *buffer, size_t *lenp,
7997 loff_t *ppos)
7998{
7999 int old_period, old_runtime;
8000 static DEFINE_MUTEX(mutex);
8001 int ret;
8002
8003 mutex_lock(&mutex);
8004 old_period = sysctl_sched_rt_period;
8005 old_runtime = sysctl_sched_rt_runtime;
8006
8007 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8008
8009 if (!ret && write) {
8010 ret = sched_rt_global_validate();
8011 if (ret)
8012 goto undo;
8013
8014 ret = sched_dl_global_validate();
8015 if (ret)
8016 goto undo;
8017
8018 ret = sched_rt_global_constraints();
8019 if (ret)
8020 goto undo;
8021
8022 sched_rt_do_global();
8023 sched_dl_do_global();
8024 }
8025 if (0) {
8026undo:
8027 sysctl_sched_rt_period = old_period;
8028 sysctl_sched_rt_runtime = old_runtime;
8029 }
8030 mutex_unlock(&mutex);
8031
8032 return ret;
8033}
8034
8035int sched_rr_handler(struct ctl_table *table, int write,
8036 void __user *buffer, size_t *lenp,
8037 loff_t *ppos)
8038{
8039 int ret;
8040 static DEFINE_MUTEX(mutex);
8041
8042 mutex_lock(&mutex);
8043 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8044 /* make sure that internally we keep jiffies */
8045 /* also, writing zero resets timeslice to default */
8046 if (!ret && write) {
8047 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8048 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8049 }
8050 mutex_unlock(&mutex);
8051 return ret;
8052}
8053
8054#ifdef CONFIG_CGROUP_SCHED
8055
8056static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8057{
8058 return css ? container_of(css, struct task_group, css) : NULL;
8059}
8060
8061static struct cgroup_subsys_state *
8062cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8063{
8064 struct task_group *parent = css_tg(parent_css);
8065 struct task_group *tg;
8066
8067 if (!parent) {
8068 /* This is early initialization for the top cgroup */
8069 return &root_task_group.css;
8070 }
8071
8072 tg = sched_create_group(parent);
8073 if (IS_ERR(tg))
8074 return ERR_PTR(-ENOMEM);
8075
8076 sched_online_group(tg, parent);
8077
8078 return &tg->css;
8079}
8080
8081static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8082{
8083 struct task_group *tg = css_tg(css);
8084
8085 sched_offline_group(tg);
8086}
8087
8088static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8089{
8090 struct task_group *tg = css_tg(css);
8091
8092 /*
8093 * Relies on the RCU grace period between css_released() and this.
8094 */
8095 sched_free_group(tg);
8096}
8097
8098static void cpu_cgroup_fork(struct task_struct *task)
8099{
8100 sched_move_task(task);
8101}
8102
8103static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8104{
8105 struct task_struct *task;
8106 struct cgroup_subsys_state *css;
8107
8108 cgroup_taskset_for_each(task, css, tset) {
8109#ifdef CONFIG_RT_GROUP_SCHED
8110 if (!sched_rt_can_attach(css_tg(css), task))
8111 return -EINVAL;
8112#else
8113 /* We don't support RT-tasks being in separate groups */
8114 if (task->sched_class != &fair_sched_class)
8115 return -EINVAL;
8116#endif
8117 }
8118 return 0;
8119}
8120
8121static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8122{
8123 struct task_struct *task;
8124 struct cgroup_subsys_state *css;
8125
8126 cgroup_taskset_for_each(task, css, tset)
8127 sched_move_task(task);
8128}
8129
8130#ifdef CONFIG_FAIR_GROUP_SCHED
8131static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8132 struct cftype *cftype, u64 shareval)
8133{
8134 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8135}
8136
8137static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8138 struct cftype *cft)
8139{
8140 struct task_group *tg = css_tg(css);
8141
8142 return (u64) scale_load_down(tg->shares);
8143}
8144
8145#ifdef CONFIG_CFS_BANDWIDTH
8146static DEFINE_MUTEX(cfs_constraints_mutex);
8147
8148const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8149const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8150
8151static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8152
8153static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8154{
8155 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8156 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8157
8158 if (tg == &root_task_group)
8159 return -EINVAL;
8160
8161 /*
8162 * Ensure we have at some amount of bandwidth every period. This is
8163 * to prevent reaching a state of large arrears when throttled via
8164 * entity_tick() resulting in prolonged exit starvation.
8165 */
8166 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8167 return -EINVAL;
8168
8169 /*
8170 * Likewise, bound things on the otherside by preventing insane quota
8171 * periods. This also allows us to normalize in computing quota
8172 * feasibility.
8173 */
8174 if (period > max_cfs_quota_period)
8175 return -EINVAL;
8176
8177 /*
8178 * Prevent race between setting of cfs_rq->runtime_enabled and
8179 * unthrottle_offline_cfs_rqs().
8180 */
8181 get_online_cpus();
8182 mutex_lock(&cfs_constraints_mutex);
8183 ret = __cfs_schedulable(tg, period, quota);
8184 if (ret)
8185 goto out_unlock;
8186
8187 runtime_enabled = quota != RUNTIME_INF;
8188 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8189 /*
8190 * If we need to toggle cfs_bandwidth_used, off->on must occur
8191 * before making related changes, and on->off must occur afterwards
8192 */
8193 if (runtime_enabled && !runtime_was_enabled)
8194 cfs_bandwidth_usage_inc();
8195 raw_spin_lock_irq(&cfs_b->lock);
8196 cfs_b->period = ns_to_ktime(period);
8197 cfs_b->quota = quota;
8198
8199 __refill_cfs_bandwidth_runtime(cfs_b);
8200 /* restart the period timer (if active) to handle new period expiry */
8201 if (runtime_enabled)
8202 start_cfs_bandwidth(cfs_b);
8203 raw_spin_unlock_irq(&cfs_b->lock);
8204
8205 for_each_online_cpu(i) {
8206 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8207 struct rq *rq = cfs_rq->rq;
8208
8209 raw_spin_lock_irq(&rq->lock);
8210 cfs_rq->runtime_enabled = runtime_enabled;
8211 cfs_rq->runtime_remaining = 0;
8212
8213 if (cfs_rq->throttled)
8214 unthrottle_cfs_rq(cfs_rq);
8215 raw_spin_unlock_irq(&rq->lock);
8216 }
8217 if (runtime_was_enabled && !runtime_enabled)
8218 cfs_bandwidth_usage_dec();
8219out_unlock:
8220 mutex_unlock(&cfs_constraints_mutex);
8221 put_online_cpus();
8222
8223 return ret;
8224}
8225
8226int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8227{
8228 u64 quota, period;
8229
8230 period = ktime_to_ns(tg->cfs_bandwidth.period);
8231 if (cfs_quota_us < 0)
8232 quota = RUNTIME_INF;
8233 else
8234 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8235
8236 return tg_set_cfs_bandwidth(tg, period, quota);
8237}
8238
8239long tg_get_cfs_quota(struct task_group *tg)
8240{
8241 u64 quota_us;
8242
8243 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8244 return -1;
8245
8246 quota_us = tg->cfs_bandwidth.quota;
8247 do_div(quota_us, NSEC_PER_USEC);
8248
8249 return quota_us;
8250}
8251
8252int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8253{
8254 u64 quota, period;
8255
8256 period = (u64)cfs_period_us * NSEC_PER_USEC;
8257 quota = tg->cfs_bandwidth.quota;
8258
8259 return tg_set_cfs_bandwidth(tg, period, quota);
8260}
8261
8262long tg_get_cfs_period(struct task_group *tg)
8263{
8264 u64 cfs_period_us;
8265
8266 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8267 do_div(cfs_period_us, NSEC_PER_USEC);
8268
8269 return cfs_period_us;
8270}
8271
8272static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8273 struct cftype *cft)
8274{
8275 return tg_get_cfs_quota(css_tg(css));
8276}
8277
8278static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8279 struct cftype *cftype, s64 cfs_quota_us)
8280{
8281 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8282}
8283
8284static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8285 struct cftype *cft)
8286{
8287 return tg_get_cfs_period(css_tg(css));
8288}
8289
8290static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8291 struct cftype *cftype, u64 cfs_period_us)
8292{
8293 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8294}
8295
8296struct cfs_schedulable_data {
8297 struct task_group *tg;
8298 u64 period, quota;
8299};
8300
8301/*
8302 * normalize group quota/period to be quota/max_period
8303 * note: units are usecs
8304 */
8305static u64 normalize_cfs_quota(struct task_group *tg,
8306 struct cfs_schedulable_data *d)
8307{
8308 u64 quota, period;
8309
8310 if (tg == d->tg) {
8311 period = d->period;
8312 quota = d->quota;
8313 } else {
8314 period = tg_get_cfs_period(tg);
8315 quota = tg_get_cfs_quota(tg);
8316 }
8317
8318 /* note: these should typically be equivalent */
8319 if (quota == RUNTIME_INF || quota == -1)
8320 return RUNTIME_INF;
8321
8322 return to_ratio(period, quota);
8323}
8324
8325static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8326{
8327 struct cfs_schedulable_data *d = data;
8328 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8329 s64 quota = 0, parent_quota = -1;
8330
8331 if (!tg->parent) {
8332 quota = RUNTIME_INF;
8333 } else {
8334 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8335
8336 quota = normalize_cfs_quota(tg, d);
8337 parent_quota = parent_b->hierarchical_quota;
8338
8339 /*
8340 * ensure max(child_quota) <= parent_quota, inherit when no
8341 * limit is set
8342 */
8343 if (quota == RUNTIME_INF)
8344 quota = parent_quota;
8345 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8346 return -EINVAL;
8347 }
8348 cfs_b->hierarchical_quota = quota;
8349
8350 return 0;
8351}
8352
8353static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8354{
8355 int ret;
8356 struct cfs_schedulable_data data = {
8357 .tg = tg,
8358 .period = period,
8359 .quota = quota,
8360 };
8361
8362 if (quota != RUNTIME_INF) {
8363 do_div(data.period, NSEC_PER_USEC);
8364 do_div(data.quota, NSEC_PER_USEC);
8365 }
8366
8367 rcu_read_lock();
8368 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8369 rcu_read_unlock();
8370
8371 return ret;
8372}
8373
8374static int cpu_stats_show(struct seq_file *sf, void *v)
8375{
8376 struct task_group *tg = css_tg(seq_css(sf));
8377 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8378
8379 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8380 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8381 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8382
8383 return 0;
8384}
8385#endif /* CONFIG_CFS_BANDWIDTH */
8386#endif /* CONFIG_FAIR_GROUP_SCHED */
8387
8388#ifdef CONFIG_RT_GROUP_SCHED
8389static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8390 struct cftype *cft, s64 val)
8391{
8392 return sched_group_set_rt_runtime(css_tg(css), val);
8393}
8394
8395static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8396 struct cftype *cft)
8397{
8398 return sched_group_rt_runtime(css_tg(css));
8399}
8400
8401static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8402 struct cftype *cftype, u64 rt_period_us)
8403{
8404 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8405}
8406
8407static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8408 struct cftype *cft)
8409{
8410 return sched_group_rt_period(css_tg(css));
8411}
8412#endif /* CONFIG_RT_GROUP_SCHED */
8413
8414static struct cftype cpu_files[] = {
8415#ifdef CONFIG_FAIR_GROUP_SCHED
8416 {
8417 .name = "shares",
8418 .read_u64 = cpu_shares_read_u64,
8419 .write_u64 = cpu_shares_write_u64,
8420 },
8421#endif
8422#ifdef CONFIG_CFS_BANDWIDTH
8423 {
8424 .name = "cfs_quota_us",
8425 .read_s64 = cpu_cfs_quota_read_s64,
8426 .write_s64 = cpu_cfs_quota_write_s64,
8427 },
8428 {
8429 .name = "cfs_period_us",
8430 .read_u64 = cpu_cfs_period_read_u64,
8431 .write_u64 = cpu_cfs_period_write_u64,
8432 },
8433 {
8434 .name = "stat",
8435 .seq_show = cpu_stats_show,
8436 },
8437#endif
8438#ifdef CONFIG_RT_GROUP_SCHED
8439 {
8440 .name = "rt_runtime_us",
8441 .read_s64 = cpu_rt_runtime_read,
8442 .write_s64 = cpu_rt_runtime_write,
8443 },
8444 {
8445 .name = "rt_period_us",
8446 .read_u64 = cpu_rt_period_read_uint,
8447 .write_u64 = cpu_rt_period_write_uint,
8448 },
8449#endif
8450 { } /* terminate */
8451};
8452
8453struct cgroup_subsys cpu_cgrp_subsys = {
8454 .css_alloc = cpu_cgroup_css_alloc,
8455 .css_released = cpu_cgroup_css_released,
8456 .css_free = cpu_cgroup_css_free,
8457 .fork = cpu_cgroup_fork,
8458 .can_attach = cpu_cgroup_can_attach,
8459 .attach = cpu_cgroup_attach,
8460 .legacy_cftypes = cpu_files,
8461 .early_init = true,
8462};
8463
8464#endif /* CONFIG_CGROUP_SCHED */
8465
8466void dump_cpu_task(int cpu)
8467{
8468 pr_info("Task dump for CPU %d:\n", cpu);
8469 sched_show_task(cpu_curr(cpu));
8470}
8471
8472/*
8473 * Nice levels are multiplicative, with a gentle 10% change for every
8474 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8475 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8476 * that remained on nice 0.
8477 *
8478 * The "10% effect" is relative and cumulative: from _any_ nice level,
8479 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8480 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8481 * If a task goes up by ~10% and another task goes down by ~10% then
8482 * the relative distance between them is ~25%.)
8483 */
8484const int sched_prio_to_weight[40] = {
8485 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8486 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8487 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8488 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8489 /* 0 */ 1024, 820, 655, 526, 423,
8490 /* 5 */ 335, 272, 215, 172, 137,
8491 /* 10 */ 110, 87, 70, 56, 45,
8492 /* 15 */ 36, 29, 23, 18, 15,
8493};
8494
8495/*
8496 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8497 *
8498 * In cases where the weight does not change often, we can use the
8499 * precalculated inverse to speed up arithmetics by turning divisions
8500 * into multiplications:
8501 */
8502const u32 sched_prio_to_wmult[40] = {
8503 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8504 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8505 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8506 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8507 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8508 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8509 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8510 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
8511};