]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blob - kernel/sched/core.c
f3f08bf943552e2faae9560e8e8fd73b0b385aaa
[mirror_ubuntu-jammy-kernel.git] / kernel / sched / core.c
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
93 void 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
113 DEFINE_MUTEX(sched_domains_mutex);
114 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115
116 static void update_rq_clock_task(struct rq *rq, s64 delta);
117
118 void 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
137 const_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
147 static const char * const sched_feat_names[] = {
148 #include "features.h"
149 };
150
151 #undef SCHED_FEAT
152
153 static 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
175 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
176 #include "features.h"
177 };
178
179 #undef SCHED_FEAT
180
181 static 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
187 static 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
193 static void sched_feat_disable(int i) { };
194 static void sched_feat_enable(int i) { };
195 #endif /* HAVE_JUMP_LABEL */
196
197 static 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
223 static ssize_t
224 sched_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
249 static int sched_feat_open(struct inode *inode, struct file *filp)
250 {
251 return single_open(filp, sched_feat_show, NULL);
252 }
253
254 static 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
262 static __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 }
269 late_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 */
276 const_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 */
284 const_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 */
290 unsigned 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 */
298 int sysctl_sched_rt_runtime = 950000;
299
300 /*
301 * __task_rq_lock - lock the rq @p resides on.
302 */
303 static 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 */
322 static 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
339 static void __task_rq_unlock(struct rq *rq)
340 __releases(rq->lock)
341 {
342 raw_spin_unlock(&rq->lock);
343 }
344
345 static inline void
346 task_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 */
357 static 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
374 static 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 */
384 static 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
400 static 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 */
411 static 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 */
426 void 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
441 static int
442 hotplug_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
460 static __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 */
470 void 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
476 static inline void init_hrtick(void)
477 {
478 }
479 #endif /* CONFIG_SMP */
480
481 static 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 */
495 static inline void hrtick_clear(struct rq *rq)
496 {
497 }
498
499 static inline void init_rq_hrtick(struct rq *rq)
500 {
501 }
502
503 static 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 */
515 void 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
538 void 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 */
559 int 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 }
577 unlock:
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 */
591 static 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
621 static 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
633 void wake_up_nohz_cpu(int cpu)
634 {
635 if (!wake_up_full_nohz_cpu(cpu))
636 wake_up_idle_cpu(cpu);
637 }
638
639 static 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
659 static 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
667 bool 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
684 void 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 */
710 int 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
718 down:
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
726 up:
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;
737 out:
738 return ret;
739 }
740
741 int tg_nop(struct task_group *tg, void *data)
742 {
743 return 0;
744 }
745 #endif
746
747 static 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
765 static 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
772 static 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
779 void 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
787 void 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
795 static 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((&paravirt_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
849 void 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, &param);
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 */
882 static 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 */
894 static 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 */
914 static 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 */
933 inline int task_curr(const struct task_struct *p)
934 {
935 return cpu_curr(task_cpu(p)) == p;
936 }
937
938 static 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
950 void 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
976 void 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
1014 static 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
1036 struct migration_swap_arg {
1037 struct task_struct *src_task, *dst_task;
1038 int src_cpu, dst_cpu;
1039 };
1040
1041 static 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
1070 unlock:
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 */
1081 int 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
1112 out:
1113 return ret;
1114 }
1115
1116 struct migration_arg {
1117 struct task_struct *task;
1118 int dest_cpu;
1119 };
1120
1121 static 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 */
1139 unsigned 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 */
1244 void 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 }
1254 EXPORT_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 */
1261 static 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
1315 out:
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 */
1334 static inline
1335 int 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
1356 static void update_avg(u64 *avg, u64 sample)
1357 {
1358 s64 diff = sample - *avg;
1359 *avg += diff >> 3;
1360 }
1361 #endif
1362
1363 static void
1364 ttwu_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
1403 static 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 */
1416 static void
1417 ttwu_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
1441 static void
1442 ttwu_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 */
1459 static 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
1477 static 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
1494 void 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
1535 static 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
1541 bool 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
1547 static 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 */
1579 static int
1580 try_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);
1628 stat:
1629 ttwu_stat(p, cpu, wake_flags);
1630 out:
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 */
1644 static 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);
1668 out:
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 */
1684 int 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 }
1689 EXPORT_SYMBOL(wake_up_process);
1690
1691 int 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 */
1702 static 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
1758 void 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
1768 void set_numabalancing_state(bool enabled)
1769 {
1770 numabalancing_enabled = enabled;
1771 }
1772 #endif /* CONFIG_SCHED_DEBUG */
1773
1774 #ifdef CONFIG_PROC_SYSCTL
1775 int 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 */
1800 int 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
1879 unsigned 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
1896 inline struct dl_bw *dl_bw_of(int i)
1897 {
1898 return &cpu_rq(i)->rd->dl_bw;
1899 }
1900
1901 static 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
1912 inline struct dl_bw *dl_bw_of(int i)
1913 {
1914 return &cpu_rq(i)->dl.dl_bw;
1915 }
1916
1917 static inline int dl_bw_cpus(int i)
1918 {
1919 return 1;
1920 }
1921 #endif
1922
1923 static inline
1924 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
1925 {
1926 dl_b->total_bw -= tsk_bw;
1927 }
1928
1929 static inline
1930 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
1931 {
1932 dl_b->total_bw += tsk_bw;
1933 }
1934
1935 static inline
1936 bool __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 */
1950 static 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
1988 extern 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 */
1997 void 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 */
2032 void preempt_notifier_register(struct preempt_notifier *notifier)
2033 {
2034 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2035 }
2036 EXPORT_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 */
2044 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2045 {
2046 hlist_del(&notifier->link);
2047 }
2048 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2049
2050 static 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
2058 static void
2059 fire_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
2070 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2071 {
2072 }
2073
2074 static void
2075 fire_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 */
2095 static inline void
2096 prepare_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 */
2122 static 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 */
2169 static 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
2185 static 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 */
2195 asmlinkage 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 */
2220 static inline void
2221 context_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 */
2277 unsigned 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
2287 unsigned 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
2298 unsigned 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
2308 unsigned 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 */
2320 void 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 }
2338 unlock:
2339 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2340 }
2341
2342 #endif
2343
2344 DEFINE_PER_CPU(struct kernel_stat, kstat);
2345 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2346
2347 EXPORT_PER_CPU_SYMBOL(kstat);
2348 EXPORT_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 */
2356 static 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
2370 unsigned 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 */
2388 unsigned 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 */
2419 void 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 */
2456 u64 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
2470 notrace 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
2483 void __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 }
2508 EXPORT_SYMBOL(preempt_count_add);
2509
2510 void __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 }
2530 EXPORT_SYMBOL(preempt_count_sub);
2531
2532 #endif
2533
2534 /*
2535 * Print scheduling while atomic bug:
2536 */
2537 static 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 */
2563 static 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 */
2582 static inline struct task_struct *
2583 pick_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
2605 again:
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 */
2655 static void __sched __schedule(void)
2656 {
2657 struct task_struct *prev, *next;
2658 unsigned long *switch_count;
2659 struct rq *rq;
2660 int cpu;
2661
2662 need_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
2738 static 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
2750 asmlinkage void __sched schedule(void)
2751 {
2752 struct task_struct *tsk = current;
2753
2754 sched_submit_work(tsk);
2755 __schedule();
2756 }
2757 EXPORT_SYMBOL(schedule);
2758
2759 #ifdef CONFIG_CONTEXT_TRACKING
2760 asmlinkage 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 */
2779 void __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 */
2792 asmlinkage 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 }
2813 EXPORT_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 */
2822 asmlinkage 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
2848 int 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 }
2853 EXPORT_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 */
2868 void 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);
2945 out_unlock:
2946 __task_rq_unlock(rq);
2947 }
2948 #endif
2949
2950 void 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 }
2992 out_unlock:
2993 task_rq_unlock(rq, p, &flags);
2994 }
2995 EXPORT_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 */
3002 int 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 */
3020 SYSCALL_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 */
3061 int 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 */
3072 int 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 */
3096 struct 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 */
3107 static 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 */
3120 static 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
3136 static 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. */
3162 static 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
3181 static 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 (1us); we
3199 * check sched_runtime only since it is always the smaller one.
3200 */
3201 static bool
3202 __checkparam_dl(const struct sched_attr *attr)
3203 {
3204 return attr && attr->sched_deadline != 0 &&
3205 (attr->sched_period == 0 ||
3206 (s64)(attr->sched_period - attr->sched_deadline) >= 0) &&
3207 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 &&
3208 attr->sched_runtime >= (2 << (DL_SCALE - 1));
3209 }
3210
3211 /*
3212 * check the target process has a UID that matches the current process's
3213 */
3214 static bool check_same_owner(struct task_struct *p)
3215 {
3216 const struct cred *cred = current_cred(), *pcred;
3217 bool match;
3218
3219 rcu_read_lock();
3220 pcred = __task_cred(p);
3221 match = (uid_eq(cred->euid, pcred->euid) ||
3222 uid_eq(cred->euid, pcred->uid));
3223 rcu_read_unlock();
3224 return match;
3225 }
3226
3227 static int __sched_setscheduler(struct task_struct *p,
3228 const struct sched_attr *attr,
3229 bool user)
3230 {
3231 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3232 MAX_RT_PRIO - 1 - attr->sched_priority;
3233 int retval, oldprio, oldpolicy = -1, on_rq, running;
3234 int policy = attr->sched_policy;
3235 unsigned long flags;
3236 const struct sched_class *prev_class;
3237 struct rq *rq;
3238 int reset_on_fork;
3239
3240 /* may grab non-irq protected spin_locks */
3241 BUG_ON(in_interrupt());
3242 recheck:
3243 /* double check policy once rq lock held */
3244 if (policy < 0) {
3245 reset_on_fork = p->sched_reset_on_fork;
3246 policy = oldpolicy = p->policy;
3247 } else {
3248 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3249
3250 if (policy != SCHED_DEADLINE &&
3251 policy != SCHED_FIFO && policy != SCHED_RR &&
3252 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3253 policy != SCHED_IDLE)
3254 return -EINVAL;
3255 }
3256
3257 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3258 return -EINVAL;
3259
3260 /*
3261 * Valid priorities for SCHED_FIFO and SCHED_RR are
3262 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3263 * SCHED_BATCH and SCHED_IDLE is 0.
3264 */
3265 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3266 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3267 return -EINVAL;
3268 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3269 (rt_policy(policy) != (attr->sched_priority != 0)))
3270 return -EINVAL;
3271
3272 /*
3273 * Allow unprivileged RT tasks to decrease priority:
3274 */
3275 if (user && !capable(CAP_SYS_NICE)) {
3276 if (fair_policy(policy)) {
3277 if (attr->sched_nice < task_nice(p) &&
3278 !can_nice(p, attr->sched_nice))
3279 return -EPERM;
3280 }
3281
3282 if (rt_policy(policy)) {
3283 unsigned long rlim_rtprio =
3284 task_rlimit(p, RLIMIT_RTPRIO);
3285
3286 /* can't set/change the rt policy */
3287 if (policy != p->policy && !rlim_rtprio)
3288 return -EPERM;
3289
3290 /* can't increase priority */
3291 if (attr->sched_priority > p->rt_priority &&
3292 attr->sched_priority > rlim_rtprio)
3293 return -EPERM;
3294 }
3295
3296 /*
3297 * Can't set/change SCHED_DEADLINE policy at all for now
3298 * (safest behavior); in the future we would like to allow
3299 * unprivileged DL tasks to increase their relative deadline
3300 * or reduce their runtime (both ways reducing utilization)
3301 */
3302 if (dl_policy(policy))
3303 return -EPERM;
3304
3305 /*
3306 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3307 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3308 */
3309 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3310 if (!can_nice(p, task_nice(p)))
3311 return -EPERM;
3312 }
3313
3314 /* can't change other user's priorities */
3315 if (!check_same_owner(p))
3316 return -EPERM;
3317
3318 /* Normal users shall not reset the sched_reset_on_fork flag */
3319 if (p->sched_reset_on_fork && !reset_on_fork)
3320 return -EPERM;
3321 }
3322
3323 if (user) {
3324 retval = security_task_setscheduler(p);
3325 if (retval)
3326 return retval;
3327 }
3328
3329 /*
3330 * make sure no PI-waiters arrive (or leave) while we are
3331 * changing the priority of the task:
3332 *
3333 * To be able to change p->policy safely, the appropriate
3334 * runqueue lock must be held.
3335 */
3336 rq = task_rq_lock(p, &flags);
3337
3338 /*
3339 * Changing the policy of the stop threads its a very bad idea
3340 */
3341 if (p == rq->stop) {
3342 task_rq_unlock(rq, p, &flags);
3343 return -EINVAL;
3344 }
3345
3346 /*
3347 * If not changing anything there's no need to proceed further,
3348 * but store a possible modification of reset_on_fork.
3349 */
3350 if (unlikely(policy == p->policy)) {
3351 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3352 goto change;
3353 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3354 goto change;
3355 if (dl_policy(policy))
3356 goto change;
3357
3358 p->sched_reset_on_fork = reset_on_fork;
3359 task_rq_unlock(rq, p, &flags);
3360 return 0;
3361 }
3362 change:
3363
3364 if (user) {
3365 #ifdef CONFIG_RT_GROUP_SCHED
3366 /*
3367 * Do not allow realtime tasks into groups that have no runtime
3368 * assigned.
3369 */
3370 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3371 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3372 !task_group_is_autogroup(task_group(p))) {
3373 task_rq_unlock(rq, p, &flags);
3374 return -EPERM;
3375 }
3376 #endif
3377 #ifdef CONFIG_SMP
3378 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3379 cpumask_t *span = rq->rd->span;
3380
3381 /*
3382 * Don't allow tasks with an affinity mask smaller than
3383 * the entire root_domain to become SCHED_DEADLINE. We
3384 * will also fail if there's no bandwidth available.
3385 */
3386 if (!cpumask_subset(span, &p->cpus_allowed) ||
3387 rq->rd->dl_bw.bw == 0) {
3388 task_rq_unlock(rq, p, &flags);
3389 return -EPERM;
3390 }
3391 }
3392 #endif
3393 }
3394
3395 /* recheck policy now with rq lock held */
3396 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3397 policy = oldpolicy = -1;
3398 task_rq_unlock(rq, p, &flags);
3399 goto recheck;
3400 }
3401
3402 /*
3403 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3404 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3405 * is available.
3406 */
3407 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3408 task_rq_unlock(rq, p, &flags);
3409 return -EBUSY;
3410 }
3411
3412 p->sched_reset_on_fork = reset_on_fork;
3413 oldprio = p->prio;
3414
3415 /*
3416 * Special case for priority boosted tasks.
3417 *
3418 * If the new priority is lower or equal (user space view)
3419 * than the current (boosted) priority, we just store the new
3420 * normal parameters and do not touch the scheduler class and
3421 * the runqueue. This will be done when the task deboost
3422 * itself.
3423 */
3424 if (rt_mutex_check_prio(p, newprio)) {
3425 __setscheduler_params(p, attr);
3426 task_rq_unlock(rq, p, &flags);
3427 return 0;
3428 }
3429
3430 on_rq = p->on_rq;
3431 running = task_current(rq, p);
3432 if (on_rq)
3433 dequeue_task(rq, p, 0);
3434 if (running)
3435 p->sched_class->put_prev_task(rq, p);
3436
3437 prev_class = p->sched_class;
3438 __setscheduler(rq, p, attr);
3439
3440 if (running)
3441 p->sched_class->set_curr_task(rq);
3442 if (on_rq) {
3443 /*
3444 * We enqueue to tail when the priority of a task is
3445 * increased (user space view).
3446 */
3447 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3448 }
3449
3450 check_class_changed(rq, p, prev_class, oldprio);
3451 task_rq_unlock(rq, p, &flags);
3452
3453 rt_mutex_adjust_pi(p);
3454
3455 return 0;
3456 }
3457
3458 static int _sched_setscheduler(struct task_struct *p, int policy,
3459 const struct sched_param *param, bool check)
3460 {
3461 struct sched_attr attr = {
3462 .sched_policy = policy,
3463 .sched_priority = param->sched_priority,
3464 .sched_nice = PRIO_TO_NICE(p->static_prio),
3465 };
3466
3467 /*
3468 * Fixup the legacy SCHED_RESET_ON_FORK hack
3469 */
3470 if (policy & SCHED_RESET_ON_FORK) {
3471 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3472 policy &= ~SCHED_RESET_ON_FORK;
3473 attr.sched_policy = policy;
3474 }
3475
3476 return __sched_setscheduler(p, &attr, check);
3477 }
3478 /**
3479 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3480 * @p: the task in question.
3481 * @policy: new policy.
3482 * @param: structure containing the new RT priority.
3483 *
3484 * Return: 0 on success. An error code otherwise.
3485 *
3486 * NOTE that the task may be already dead.
3487 */
3488 int sched_setscheduler(struct task_struct *p, int policy,
3489 const struct sched_param *param)
3490 {
3491 return _sched_setscheduler(p, policy, param, true);
3492 }
3493 EXPORT_SYMBOL_GPL(sched_setscheduler);
3494
3495 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3496 {
3497 return __sched_setscheduler(p, attr, true);
3498 }
3499 EXPORT_SYMBOL_GPL(sched_setattr);
3500
3501 /**
3502 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3503 * @p: the task in question.
3504 * @policy: new policy.
3505 * @param: structure containing the new RT priority.
3506 *
3507 * Just like sched_setscheduler, only don't bother checking if the
3508 * current context has permission. For example, this is needed in
3509 * stop_machine(): we create temporary high priority worker threads,
3510 * but our caller might not have that capability.
3511 *
3512 * Return: 0 on success. An error code otherwise.
3513 */
3514 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3515 const struct sched_param *param)
3516 {
3517 return _sched_setscheduler(p, policy, param, false);
3518 }
3519
3520 static int
3521 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3522 {
3523 struct sched_param lparam;
3524 struct task_struct *p;
3525 int retval;
3526
3527 if (!param || pid < 0)
3528 return -EINVAL;
3529 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3530 return -EFAULT;
3531
3532 rcu_read_lock();
3533 retval = -ESRCH;
3534 p = find_process_by_pid(pid);
3535 if (p != NULL)
3536 retval = sched_setscheduler(p, policy, &lparam);
3537 rcu_read_unlock();
3538
3539 return retval;
3540 }
3541
3542 /*
3543 * Mimics kernel/events/core.c perf_copy_attr().
3544 */
3545 static int sched_copy_attr(struct sched_attr __user *uattr,
3546 struct sched_attr *attr)
3547 {
3548 u32 size;
3549 int ret;
3550
3551 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3552 return -EFAULT;
3553
3554 /*
3555 * zero the full structure, so that a short copy will be nice.
3556 */
3557 memset(attr, 0, sizeof(*attr));
3558
3559 ret = get_user(size, &uattr->size);
3560 if (ret)
3561 return ret;
3562
3563 if (size > PAGE_SIZE) /* silly large */
3564 goto err_size;
3565
3566 if (!size) /* abi compat */
3567 size = SCHED_ATTR_SIZE_VER0;
3568
3569 if (size < SCHED_ATTR_SIZE_VER0)
3570 goto err_size;
3571
3572 /*
3573 * If we're handed a bigger struct than we know of,
3574 * ensure all the unknown bits are 0 - i.e. new
3575 * user-space does not rely on any kernel feature
3576 * extensions we dont know about yet.
3577 */
3578 if (size > sizeof(*attr)) {
3579 unsigned char __user *addr;
3580 unsigned char __user *end;
3581 unsigned char val;
3582
3583 addr = (void __user *)uattr + sizeof(*attr);
3584 end = (void __user *)uattr + size;
3585
3586 for (; addr < end; addr++) {
3587 ret = get_user(val, addr);
3588 if (ret)
3589 return ret;
3590 if (val)
3591 goto err_size;
3592 }
3593 size = sizeof(*attr);
3594 }
3595
3596 ret = copy_from_user(attr, uattr, size);
3597 if (ret)
3598 return -EFAULT;
3599
3600 /*
3601 * XXX: do we want to be lenient like existing syscalls; or do we want
3602 * to be strict and return an error on out-of-bounds values?
3603 */
3604 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3605
3606 out:
3607 return ret;
3608
3609 err_size:
3610 put_user(sizeof(*attr), &uattr->size);
3611 ret = -E2BIG;
3612 goto out;
3613 }
3614
3615 /**
3616 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3617 * @pid: the pid in question.
3618 * @policy: new policy.
3619 * @param: structure containing the new RT priority.
3620 *
3621 * Return: 0 on success. An error code otherwise.
3622 */
3623 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3624 struct sched_param __user *, param)
3625 {
3626 /* negative values for policy are not valid */
3627 if (policy < 0)
3628 return -EINVAL;
3629
3630 return do_sched_setscheduler(pid, policy, param);
3631 }
3632
3633 /**
3634 * sys_sched_setparam - set/change the RT priority of a thread
3635 * @pid: the pid in question.
3636 * @param: structure containing the new RT priority.
3637 *
3638 * Return: 0 on success. An error code otherwise.
3639 */
3640 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3641 {
3642 return do_sched_setscheduler(pid, -1, param);
3643 }
3644
3645 /**
3646 * sys_sched_setattr - same as above, but with extended sched_attr
3647 * @pid: the pid in question.
3648 * @uattr: structure containing the extended parameters.
3649 * @flags: for future extension.
3650 */
3651 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3652 unsigned int, flags)
3653 {
3654 struct sched_attr attr;
3655 struct task_struct *p;
3656 int retval;
3657
3658 if (!uattr || pid < 0 || flags)
3659 return -EINVAL;
3660
3661 retval = sched_copy_attr(uattr, &attr);
3662 if (retval)
3663 return retval;
3664
3665 if (attr.sched_policy < 0)
3666 return -EINVAL;
3667
3668 rcu_read_lock();
3669 retval = -ESRCH;
3670 p = find_process_by_pid(pid);
3671 if (p != NULL)
3672 retval = sched_setattr(p, &attr);
3673 rcu_read_unlock();
3674
3675 return retval;
3676 }
3677
3678 /**
3679 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3680 * @pid: the pid in question.
3681 *
3682 * Return: On success, the policy of the thread. Otherwise, a negative error
3683 * code.
3684 */
3685 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3686 {
3687 struct task_struct *p;
3688 int retval;
3689
3690 if (pid < 0)
3691 return -EINVAL;
3692
3693 retval = -ESRCH;
3694 rcu_read_lock();
3695 p = find_process_by_pid(pid);
3696 if (p) {
3697 retval = security_task_getscheduler(p);
3698 if (!retval)
3699 retval = p->policy
3700 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3701 }
3702 rcu_read_unlock();
3703 return retval;
3704 }
3705
3706 /**
3707 * sys_sched_getparam - get the RT priority of a thread
3708 * @pid: the pid in question.
3709 * @param: structure containing the RT priority.
3710 *
3711 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3712 * code.
3713 */
3714 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3715 {
3716 struct sched_param lp = { .sched_priority = 0 };
3717 struct task_struct *p;
3718 int retval;
3719
3720 if (!param || pid < 0)
3721 return -EINVAL;
3722
3723 rcu_read_lock();
3724 p = find_process_by_pid(pid);
3725 retval = -ESRCH;
3726 if (!p)
3727 goto out_unlock;
3728
3729 retval = security_task_getscheduler(p);
3730 if (retval)
3731 goto out_unlock;
3732
3733 if (task_has_rt_policy(p))
3734 lp.sched_priority = p->rt_priority;
3735 rcu_read_unlock();
3736
3737 /*
3738 * This one might sleep, we cannot do it with a spinlock held ...
3739 */
3740 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3741
3742 return retval;
3743
3744 out_unlock:
3745 rcu_read_unlock();
3746 return retval;
3747 }
3748
3749 static int sched_read_attr(struct sched_attr __user *uattr,
3750 struct sched_attr *attr,
3751 unsigned int usize)
3752 {
3753 int ret;
3754
3755 if (!access_ok(VERIFY_WRITE, uattr, usize))
3756 return -EFAULT;
3757
3758 /*
3759 * If we're handed a smaller struct than we know of,
3760 * ensure all the unknown bits are 0 - i.e. old
3761 * user-space does not get uncomplete information.
3762 */
3763 if (usize < sizeof(*attr)) {
3764 unsigned char *addr;
3765 unsigned char *end;
3766
3767 addr = (void *)attr + usize;
3768 end = (void *)attr + sizeof(*attr);
3769
3770 for (; addr < end; addr++) {
3771 if (*addr)
3772 goto err_size;
3773 }
3774
3775 attr->size = usize;
3776 }
3777
3778 ret = copy_to_user(uattr, attr, attr->size);
3779 if (ret)
3780 return -EFAULT;
3781
3782 out:
3783 return ret;
3784
3785 err_size:
3786 ret = -E2BIG;
3787 goto out;
3788 }
3789
3790 /**
3791 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3792 * @pid: the pid in question.
3793 * @uattr: structure containing the extended parameters.
3794 * @size: sizeof(attr) for fwd/bwd comp.
3795 * @flags: for future extension.
3796 */
3797 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3798 unsigned int, size, unsigned int, flags)
3799 {
3800 struct sched_attr attr = {
3801 .size = sizeof(struct sched_attr),
3802 };
3803 struct task_struct *p;
3804 int retval;
3805
3806 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3807 size < SCHED_ATTR_SIZE_VER0 || flags)
3808 return -EINVAL;
3809
3810 rcu_read_lock();
3811 p = find_process_by_pid(pid);
3812 retval = -ESRCH;
3813 if (!p)
3814 goto out_unlock;
3815
3816 retval = security_task_getscheduler(p);
3817 if (retval)
3818 goto out_unlock;
3819
3820 attr.sched_policy = p->policy;
3821 if (p->sched_reset_on_fork)
3822 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3823 if (task_has_dl_policy(p))
3824 __getparam_dl(p, &attr);
3825 else if (task_has_rt_policy(p))
3826 attr.sched_priority = p->rt_priority;
3827 else
3828 attr.sched_nice = task_nice(p);
3829
3830 rcu_read_unlock();
3831
3832 retval = sched_read_attr(uattr, &attr, size);
3833 return retval;
3834
3835 out_unlock:
3836 rcu_read_unlock();
3837 return retval;
3838 }
3839
3840 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3841 {
3842 cpumask_var_t cpus_allowed, new_mask;
3843 struct task_struct *p;
3844 int retval;
3845
3846 rcu_read_lock();
3847
3848 p = find_process_by_pid(pid);
3849 if (!p) {
3850 rcu_read_unlock();
3851 return -ESRCH;
3852 }
3853
3854 /* Prevent p going away */
3855 get_task_struct(p);
3856 rcu_read_unlock();
3857
3858 if (p->flags & PF_NO_SETAFFINITY) {
3859 retval = -EINVAL;
3860 goto out_put_task;
3861 }
3862 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3863 retval = -ENOMEM;
3864 goto out_put_task;
3865 }
3866 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3867 retval = -ENOMEM;
3868 goto out_free_cpus_allowed;
3869 }
3870 retval = -EPERM;
3871 if (!check_same_owner(p)) {
3872 rcu_read_lock();
3873 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3874 rcu_read_unlock();
3875 goto out_unlock;
3876 }
3877 rcu_read_unlock();
3878 }
3879
3880 retval = security_task_setscheduler(p);
3881 if (retval)
3882 goto out_unlock;
3883
3884
3885 cpuset_cpus_allowed(p, cpus_allowed);
3886 cpumask_and(new_mask, in_mask, cpus_allowed);
3887
3888 /*
3889 * Since bandwidth control happens on root_domain basis,
3890 * if admission test is enabled, we only admit -deadline
3891 * tasks allowed to run on all the CPUs in the task's
3892 * root_domain.
3893 */
3894 #ifdef CONFIG_SMP
3895 if (task_has_dl_policy(p)) {
3896 const struct cpumask *span = task_rq(p)->rd->span;
3897
3898 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
3899 retval = -EBUSY;
3900 goto out_unlock;
3901 }
3902 }
3903 #endif
3904 again:
3905 retval = set_cpus_allowed_ptr(p, new_mask);
3906
3907 if (!retval) {
3908 cpuset_cpus_allowed(p, cpus_allowed);
3909 if (!cpumask_subset(new_mask, cpus_allowed)) {
3910 /*
3911 * We must have raced with a concurrent cpuset
3912 * update. Just reset the cpus_allowed to the
3913 * cpuset's cpus_allowed
3914 */
3915 cpumask_copy(new_mask, cpus_allowed);
3916 goto again;
3917 }
3918 }
3919 out_unlock:
3920 free_cpumask_var(new_mask);
3921 out_free_cpus_allowed:
3922 free_cpumask_var(cpus_allowed);
3923 out_put_task:
3924 put_task_struct(p);
3925 return retval;
3926 }
3927
3928 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3929 struct cpumask *new_mask)
3930 {
3931 if (len < cpumask_size())
3932 cpumask_clear(new_mask);
3933 else if (len > cpumask_size())
3934 len = cpumask_size();
3935
3936 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3937 }
3938
3939 /**
3940 * sys_sched_setaffinity - set the cpu affinity of a process
3941 * @pid: pid of the process
3942 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3943 * @user_mask_ptr: user-space pointer to the new cpu mask
3944 *
3945 * Return: 0 on success. An error code otherwise.
3946 */
3947 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3948 unsigned long __user *, user_mask_ptr)
3949 {
3950 cpumask_var_t new_mask;
3951 int retval;
3952
3953 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3954 return -ENOMEM;
3955
3956 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3957 if (retval == 0)
3958 retval = sched_setaffinity(pid, new_mask);
3959 free_cpumask_var(new_mask);
3960 return retval;
3961 }
3962
3963 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3964 {
3965 struct task_struct *p;
3966 unsigned long flags;
3967 int retval;
3968
3969 rcu_read_lock();
3970
3971 retval = -ESRCH;
3972 p = find_process_by_pid(pid);
3973 if (!p)
3974 goto out_unlock;
3975
3976 retval = security_task_getscheduler(p);
3977 if (retval)
3978 goto out_unlock;
3979
3980 raw_spin_lock_irqsave(&p->pi_lock, flags);
3981 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
3982 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3983
3984 out_unlock:
3985 rcu_read_unlock();
3986
3987 return retval;
3988 }
3989
3990 /**
3991 * sys_sched_getaffinity - get the cpu affinity of a process
3992 * @pid: pid of the process
3993 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3994 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3995 *
3996 * Return: 0 on success. An error code otherwise.
3997 */
3998 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
3999 unsigned long __user *, user_mask_ptr)
4000 {
4001 int ret;
4002 cpumask_var_t mask;
4003
4004 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4005 return -EINVAL;
4006 if (len & (sizeof(unsigned long)-1))
4007 return -EINVAL;
4008
4009 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4010 return -ENOMEM;
4011
4012 ret = sched_getaffinity(pid, mask);
4013 if (ret == 0) {
4014 size_t retlen = min_t(size_t, len, cpumask_size());
4015
4016 if (copy_to_user(user_mask_ptr, mask, retlen))
4017 ret = -EFAULT;
4018 else
4019 ret = retlen;
4020 }
4021 free_cpumask_var(mask);
4022
4023 return ret;
4024 }
4025
4026 /**
4027 * sys_sched_yield - yield the current processor to other threads.
4028 *
4029 * This function yields the current CPU to other tasks. If there are no
4030 * other threads running on this CPU then this function will return.
4031 *
4032 * Return: 0.
4033 */
4034 SYSCALL_DEFINE0(sched_yield)
4035 {
4036 struct rq *rq = this_rq_lock();
4037
4038 schedstat_inc(rq, yld_count);
4039 current->sched_class->yield_task(rq);
4040
4041 /*
4042 * Since we are going to call schedule() anyway, there's
4043 * no need to preempt or enable interrupts:
4044 */
4045 __release(rq->lock);
4046 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4047 do_raw_spin_unlock(&rq->lock);
4048 sched_preempt_enable_no_resched();
4049
4050 schedule();
4051
4052 return 0;
4053 }
4054
4055 static void __cond_resched(void)
4056 {
4057 __preempt_count_add(PREEMPT_ACTIVE);
4058 __schedule();
4059 __preempt_count_sub(PREEMPT_ACTIVE);
4060 }
4061
4062 int __sched _cond_resched(void)
4063 {
4064 if (should_resched()) {
4065 __cond_resched();
4066 return 1;
4067 }
4068 return 0;
4069 }
4070 EXPORT_SYMBOL(_cond_resched);
4071
4072 /*
4073 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4074 * call schedule, and on return reacquire the lock.
4075 *
4076 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4077 * operations here to prevent schedule() from being called twice (once via
4078 * spin_unlock(), once by hand).
4079 */
4080 int __cond_resched_lock(spinlock_t *lock)
4081 {
4082 int resched = should_resched();
4083 int ret = 0;
4084
4085 lockdep_assert_held(lock);
4086
4087 if (spin_needbreak(lock) || resched) {
4088 spin_unlock(lock);
4089 if (resched)
4090 __cond_resched();
4091 else
4092 cpu_relax();
4093 ret = 1;
4094 spin_lock(lock);
4095 }
4096 return ret;
4097 }
4098 EXPORT_SYMBOL(__cond_resched_lock);
4099
4100 int __sched __cond_resched_softirq(void)
4101 {
4102 BUG_ON(!in_softirq());
4103
4104 if (should_resched()) {
4105 local_bh_enable();
4106 __cond_resched();
4107 local_bh_disable();
4108 return 1;
4109 }
4110 return 0;
4111 }
4112 EXPORT_SYMBOL(__cond_resched_softirq);
4113
4114 /**
4115 * yield - yield the current processor to other threads.
4116 *
4117 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4118 *
4119 * The scheduler is at all times free to pick the calling task as the most
4120 * eligible task to run, if removing the yield() call from your code breaks
4121 * it, its already broken.
4122 *
4123 * Typical broken usage is:
4124 *
4125 * while (!event)
4126 * yield();
4127 *
4128 * where one assumes that yield() will let 'the other' process run that will
4129 * make event true. If the current task is a SCHED_FIFO task that will never
4130 * happen. Never use yield() as a progress guarantee!!
4131 *
4132 * If you want to use yield() to wait for something, use wait_event().
4133 * If you want to use yield() to be 'nice' for others, use cond_resched().
4134 * If you still want to use yield(), do not!
4135 */
4136 void __sched yield(void)
4137 {
4138 set_current_state(TASK_RUNNING);
4139 sys_sched_yield();
4140 }
4141 EXPORT_SYMBOL(yield);
4142
4143 /**
4144 * yield_to - yield the current processor to another thread in
4145 * your thread group, or accelerate that thread toward the
4146 * processor it's on.
4147 * @p: target task
4148 * @preempt: whether task preemption is allowed or not
4149 *
4150 * It's the caller's job to ensure that the target task struct
4151 * can't go away on us before we can do any checks.
4152 *
4153 * Return:
4154 * true (>0) if we indeed boosted the target task.
4155 * false (0) if we failed to boost the target.
4156 * -ESRCH if there's no task to yield to.
4157 */
4158 bool __sched yield_to(struct task_struct *p, bool preempt)
4159 {
4160 struct task_struct *curr = current;
4161 struct rq *rq, *p_rq;
4162 unsigned long flags;
4163 int yielded = 0;
4164
4165 local_irq_save(flags);
4166 rq = this_rq();
4167
4168 again:
4169 p_rq = task_rq(p);
4170 /*
4171 * If we're the only runnable task on the rq and target rq also
4172 * has only one task, there's absolutely no point in yielding.
4173 */
4174 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4175 yielded = -ESRCH;
4176 goto out_irq;
4177 }
4178
4179 double_rq_lock(rq, p_rq);
4180 if (task_rq(p) != p_rq) {
4181 double_rq_unlock(rq, p_rq);
4182 goto again;
4183 }
4184
4185 if (!curr->sched_class->yield_to_task)
4186 goto out_unlock;
4187
4188 if (curr->sched_class != p->sched_class)
4189 goto out_unlock;
4190
4191 if (task_running(p_rq, p) || p->state)
4192 goto out_unlock;
4193
4194 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4195 if (yielded) {
4196 schedstat_inc(rq, yld_count);
4197 /*
4198 * Make p's CPU reschedule; pick_next_entity takes care of
4199 * fairness.
4200 */
4201 if (preempt && rq != p_rq)
4202 resched_task(p_rq->curr);
4203 }
4204
4205 out_unlock:
4206 double_rq_unlock(rq, p_rq);
4207 out_irq:
4208 local_irq_restore(flags);
4209
4210 if (yielded > 0)
4211 schedule();
4212
4213 return yielded;
4214 }
4215 EXPORT_SYMBOL_GPL(yield_to);
4216
4217 /*
4218 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4219 * that process accounting knows that this is a task in IO wait state.
4220 */
4221 void __sched io_schedule(void)
4222 {
4223 struct rq *rq = raw_rq();
4224
4225 delayacct_blkio_start();
4226 atomic_inc(&rq->nr_iowait);
4227 blk_flush_plug(current);
4228 current->in_iowait = 1;
4229 schedule();
4230 current->in_iowait = 0;
4231 atomic_dec(&rq->nr_iowait);
4232 delayacct_blkio_end();
4233 }
4234 EXPORT_SYMBOL(io_schedule);
4235
4236 long __sched io_schedule_timeout(long timeout)
4237 {
4238 struct rq *rq = raw_rq();
4239 long ret;
4240
4241 delayacct_blkio_start();
4242 atomic_inc(&rq->nr_iowait);
4243 blk_flush_plug(current);
4244 current->in_iowait = 1;
4245 ret = schedule_timeout(timeout);
4246 current->in_iowait = 0;
4247 atomic_dec(&rq->nr_iowait);
4248 delayacct_blkio_end();
4249 return ret;
4250 }
4251
4252 /**
4253 * sys_sched_get_priority_max - return maximum RT priority.
4254 * @policy: scheduling class.
4255 *
4256 * Return: On success, this syscall returns the maximum
4257 * rt_priority that can be used by a given scheduling class.
4258 * On failure, a negative error code is returned.
4259 */
4260 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4261 {
4262 int ret = -EINVAL;
4263
4264 switch (policy) {
4265 case SCHED_FIFO:
4266 case SCHED_RR:
4267 ret = MAX_USER_RT_PRIO-1;
4268 break;
4269 case SCHED_DEADLINE:
4270 case SCHED_NORMAL:
4271 case SCHED_BATCH:
4272 case SCHED_IDLE:
4273 ret = 0;
4274 break;
4275 }
4276 return ret;
4277 }
4278
4279 /**
4280 * sys_sched_get_priority_min - return minimum RT priority.
4281 * @policy: scheduling class.
4282 *
4283 * Return: On success, this syscall returns the minimum
4284 * rt_priority that can be used by a given scheduling class.
4285 * On failure, a negative error code is returned.
4286 */
4287 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4288 {
4289 int ret = -EINVAL;
4290
4291 switch (policy) {
4292 case SCHED_FIFO:
4293 case SCHED_RR:
4294 ret = 1;
4295 break;
4296 case SCHED_DEADLINE:
4297 case SCHED_NORMAL:
4298 case SCHED_BATCH:
4299 case SCHED_IDLE:
4300 ret = 0;
4301 }
4302 return ret;
4303 }
4304
4305 /**
4306 * sys_sched_rr_get_interval - return the default timeslice of a process.
4307 * @pid: pid of the process.
4308 * @interval: userspace pointer to the timeslice value.
4309 *
4310 * this syscall writes the default timeslice value of a given process
4311 * into the user-space timespec buffer. A value of '0' means infinity.
4312 *
4313 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4314 * an error code.
4315 */
4316 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4317 struct timespec __user *, interval)
4318 {
4319 struct task_struct *p;
4320 unsigned int time_slice;
4321 unsigned long flags;
4322 struct rq *rq;
4323 int retval;
4324 struct timespec t;
4325
4326 if (pid < 0)
4327 return -EINVAL;
4328
4329 retval = -ESRCH;
4330 rcu_read_lock();
4331 p = find_process_by_pid(pid);
4332 if (!p)
4333 goto out_unlock;
4334
4335 retval = security_task_getscheduler(p);
4336 if (retval)
4337 goto out_unlock;
4338
4339 rq = task_rq_lock(p, &flags);
4340 time_slice = 0;
4341 if (p->sched_class->get_rr_interval)
4342 time_slice = p->sched_class->get_rr_interval(rq, p);
4343 task_rq_unlock(rq, p, &flags);
4344
4345 rcu_read_unlock();
4346 jiffies_to_timespec(time_slice, &t);
4347 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4348 return retval;
4349
4350 out_unlock:
4351 rcu_read_unlock();
4352 return retval;
4353 }
4354
4355 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4356
4357 void sched_show_task(struct task_struct *p)
4358 {
4359 unsigned long free = 0;
4360 int ppid;
4361 unsigned state;
4362
4363 state = p->state ? __ffs(p->state) + 1 : 0;
4364 printk(KERN_INFO "%-15.15s %c", p->comm,
4365 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4366 #if BITS_PER_LONG == 32
4367 if (state == TASK_RUNNING)
4368 printk(KERN_CONT " running ");
4369 else
4370 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4371 #else
4372 if (state == TASK_RUNNING)
4373 printk(KERN_CONT " running task ");
4374 else
4375 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4376 #endif
4377 #ifdef CONFIG_DEBUG_STACK_USAGE
4378 free = stack_not_used(p);
4379 #endif
4380 rcu_read_lock();
4381 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4382 rcu_read_unlock();
4383 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4384 task_pid_nr(p), ppid,
4385 (unsigned long)task_thread_info(p)->flags);
4386
4387 print_worker_info(KERN_INFO, p);
4388 show_stack(p, NULL);
4389 }
4390
4391 void show_state_filter(unsigned long state_filter)
4392 {
4393 struct task_struct *g, *p;
4394
4395 #if BITS_PER_LONG == 32
4396 printk(KERN_INFO
4397 " task PC stack pid father\n");
4398 #else
4399 printk(KERN_INFO
4400 " task PC stack pid father\n");
4401 #endif
4402 rcu_read_lock();
4403 do_each_thread(g, p) {
4404 /*
4405 * reset the NMI-timeout, listing all files on a slow
4406 * console might take a lot of time:
4407 */
4408 touch_nmi_watchdog();
4409 if (!state_filter || (p->state & state_filter))
4410 sched_show_task(p);
4411 } while_each_thread(g, p);
4412
4413 touch_all_softlockup_watchdogs();
4414
4415 #ifdef CONFIG_SCHED_DEBUG
4416 sysrq_sched_debug_show();
4417 #endif
4418 rcu_read_unlock();
4419 /*
4420 * Only show locks if all tasks are dumped:
4421 */
4422 if (!state_filter)
4423 debug_show_all_locks();
4424 }
4425
4426 void init_idle_bootup_task(struct task_struct *idle)
4427 {
4428 idle->sched_class = &idle_sched_class;
4429 }
4430
4431 /**
4432 * init_idle - set up an idle thread for a given CPU
4433 * @idle: task in question
4434 * @cpu: cpu the idle task belongs to
4435 *
4436 * NOTE: this function does not set the idle thread's NEED_RESCHED
4437 * flag, to make booting more robust.
4438 */
4439 void init_idle(struct task_struct *idle, int cpu)
4440 {
4441 struct rq *rq = cpu_rq(cpu);
4442 unsigned long flags;
4443
4444 raw_spin_lock_irqsave(&rq->lock, flags);
4445
4446 __sched_fork(0, idle);
4447 idle->state = TASK_RUNNING;
4448 idle->se.exec_start = sched_clock();
4449
4450 do_set_cpus_allowed(idle, cpumask_of(cpu));
4451 /*
4452 * We're having a chicken and egg problem, even though we are
4453 * holding rq->lock, the cpu isn't yet set to this cpu so the
4454 * lockdep check in task_group() will fail.
4455 *
4456 * Similar case to sched_fork(). / Alternatively we could
4457 * use task_rq_lock() here and obtain the other rq->lock.
4458 *
4459 * Silence PROVE_RCU
4460 */
4461 rcu_read_lock();
4462 __set_task_cpu(idle, cpu);
4463 rcu_read_unlock();
4464
4465 rq->curr = rq->idle = idle;
4466 idle->on_rq = 1;
4467 #if defined(CONFIG_SMP)
4468 idle->on_cpu = 1;
4469 #endif
4470 raw_spin_unlock_irqrestore(&rq->lock, flags);
4471
4472 /* Set the preempt count _outside_ the spinlocks! */
4473 init_idle_preempt_count(idle, cpu);
4474
4475 /*
4476 * The idle tasks have their own, simple scheduling class:
4477 */
4478 idle->sched_class = &idle_sched_class;
4479 ftrace_graph_init_idle_task(idle, cpu);
4480 vtime_init_idle(idle, cpu);
4481 #if defined(CONFIG_SMP)
4482 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4483 #endif
4484 }
4485
4486 #ifdef CONFIG_SMP
4487 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4488 {
4489 if (p->sched_class && p->sched_class->set_cpus_allowed)
4490 p->sched_class->set_cpus_allowed(p, new_mask);
4491
4492 cpumask_copy(&p->cpus_allowed, new_mask);
4493 p->nr_cpus_allowed = cpumask_weight(new_mask);
4494 }
4495
4496 /*
4497 * This is how migration works:
4498 *
4499 * 1) we invoke migration_cpu_stop() on the target CPU using
4500 * stop_one_cpu().
4501 * 2) stopper starts to run (implicitly forcing the migrated thread
4502 * off the CPU)
4503 * 3) it checks whether the migrated task is still in the wrong runqueue.
4504 * 4) if it's in the wrong runqueue then the migration thread removes
4505 * it and puts it into the right queue.
4506 * 5) stopper completes and stop_one_cpu() returns and the migration
4507 * is done.
4508 */
4509
4510 /*
4511 * Change a given task's CPU affinity. Migrate the thread to a
4512 * proper CPU and schedule it away if the CPU it's executing on
4513 * is removed from the allowed bitmask.
4514 *
4515 * NOTE: the caller must have a valid reference to the task, the
4516 * task must not exit() & deallocate itself prematurely. The
4517 * call is not atomic; no spinlocks may be held.
4518 */
4519 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4520 {
4521 unsigned long flags;
4522 struct rq *rq;
4523 unsigned int dest_cpu;
4524 int ret = 0;
4525
4526 rq = task_rq_lock(p, &flags);
4527
4528 if (cpumask_equal(&p->cpus_allowed, new_mask))
4529 goto out;
4530
4531 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4532 ret = -EINVAL;
4533 goto out;
4534 }
4535
4536 do_set_cpus_allowed(p, new_mask);
4537
4538 /* Can the task run on the task's current CPU? If so, we're done */
4539 if (cpumask_test_cpu(task_cpu(p), new_mask))
4540 goto out;
4541
4542 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4543 if (p->on_rq) {
4544 struct migration_arg arg = { p, dest_cpu };
4545 /* Need help from migration thread: drop lock and wait. */
4546 task_rq_unlock(rq, p, &flags);
4547 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4548 tlb_migrate_finish(p->mm);
4549 return 0;
4550 }
4551 out:
4552 task_rq_unlock(rq, p, &flags);
4553
4554 return ret;
4555 }
4556 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4557
4558 /*
4559 * Move (not current) task off this cpu, onto dest cpu. We're doing
4560 * this because either it can't run here any more (set_cpus_allowed()
4561 * away from this CPU, or CPU going down), or because we're
4562 * attempting to rebalance this task on exec (sched_exec).
4563 *
4564 * So we race with normal scheduler movements, but that's OK, as long
4565 * as the task is no longer on this CPU.
4566 *
4567 * Returns non-zero if task was successfully migrated.
4568 */
4569 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4570 {
4571 struct rq *rq_dest, *rq_src;
4572 int ret = 0;
4573
4574 if (unlikely(!cpu_active(dest_cpu)))
4575 return ret;
4576
4577 rq_src = cpu_rq(src_cpu);
4578 rq_dest = cpu_rq(dest_cpu);
4579
4580 raw_spin_lock(&p->pi_lock);
4581 double_rq_lock(rq_src, rq_dest);
4582 /* Already moved. */
4583 if (task_cpu(p) != src_cpu)
4584 goto done;
4585 /* Affinity changed (again). */
4586 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4587 goto fail;
4588
4589 /*
4590 * If we're not on a rq, the next wake-up will ensure we're
4591 * placed properly.
4592 */
4593 if (p->on_rq) {
4594 dequeue_task(rq_src, p, 0);
4595 set_task_cpu(p, dest_cpu);
4596 enqueue_task(rq_dest, p, 0);
4597 check_preempt_curr(rq_dest, p, 0);
4598 }
4599 done:
4600 ret = 1;
4601 fail:
4602 double_rq_unlock(rq_src, rq_dest);
4603 raw_spin_unlock(&p->pi_lock);
4604 return ret;
4605 }
4606
4607 #ifdef CONFIG_NUMA_BALANCING
4608 /* Migrate current task p to target_cpu */
4609 int migrate_task_to(struct task_struct *p, int target_cpu)
4610 {
4611 struct migration_arg arg = { p, target_cpu };
4612 int curr_cpu = task_cpu(p);
4613
4614 if (curr_cpu == target_cpu)
4615 return 0;
4616
4617 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4618 return -EINVAL;
4619
4620 /* TODO: This is not properly updating schedstats */
4621
4622 trace_sched_move_numa(p, curr_cpu, target_cpu);
4623 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4624 }
4625
4626 /*
4627 * Requeue a task on a given node and accurately track the number of NUMA
4628 * tasks on the runqueues
4629 */
4630 void sched_setnuma(struct task_struct *p, int nid)
4631 {
4632 struct rq *rq;
4633 unsigned long flags;
4634 bool on_rq, running;
4635
4636 rq = task_rq_lock(p, &flags);
4637 on_rq = p->on_rq;
4638 running = task_current(rq, p);
4639
4640 if (on_rq)
4641 dequeue_task(rq, p, 0);
4642 if (running)
4643 p->sched_class->put_prev_task(rq, p);
4644
4645 p->numa_preferred_nid = nid;
4646
4647 if (running)
4648 p->sched_class->set_curr_task(rq);
4649 if (on_rq)
4650 enqueue_task(rq, p, 0);
4651 task_rq_unlock(rq, p, &flags);
4652 }
4653 #endif
4654
4655 /*
4656 * migration_cpu_stop - this will be executed by a highprio stopper thread
4657 * and performs thread migration by bumping thread off CPU then
4658 * 'pushing' onto another runqueue.
4659 */
4660 static int migration_cpu_stop(void *data)
4661 {
4662 struct migration_arg *arg = data;
4663
4664 /*
4665 * The original target cpu might have gone down and we might
4666 * be on another cpu but it doesn't matter.
4667 */
4668 local_irq_disable();
4669 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4670 local_irq_enable();
4671 return 0;
4672 }
4673
4674 #ifdef CONFIG_HOTPLUG_CPU
4675
4676 /*
4677 * Ensures that the idle task is using init_mm right before its cpu goes
4678 * offline.
4679 */
4680 void idle_task_exit(void)
4681 {
4682 struct mm_struct *mm = current->active_mm;
4683
4684 BUG_ON(cpu_online(smp_processor_id()));
4685
4686 if (mm != &init_mm) {
4687 switch_mm(mm, &init_mm, current);
4688 finish_arch_post_lock_switch();
4689 }
4690 mmdrop(mm);
4691 }
4692
4693 /*
4694 * Since this CPU is going 'away' for a while, fold any nr_active delta
4695 * we might have. Assumes we're called after migrate_tasks() so that the
4696 * nr_active count is stable.
4697 *
4698 * Also see the comment "Global load-average calculations".
4699 */
4700 static void calc_load_migrate(struct rq *rq)
4701 {
4702 long delta = calc_load_fold_active(rq);
4703 if (delta)
4704 atomic_long_add(delta, &calc_load_tasks);
4705 }
4706
4707 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4708 {
4709 }
4710
4711 static const struct sched_class fake_sched_class = {
4712 .put_prev_task = put_prev_task_fake,
4713 };
4714
4715 static struct task_struct fake_task = {
4716 /*
4717 * Avoid pull_{rt,dl}_task()
4718 */
4719 .prio = MAX_PRIO + 1,
4720 .sched_class = &fake_sched_class,
4721 };
4722
4723 /*
4724 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4725 * try_to_wake_up()->select_task_rq().
4726 *
4727 * Called with rq->lock held even though we'er in stop_machine() and
4728 * there's no concurrency possible, we hold the required locks anyway
4729 * because of lock validation efforts.
4730 */
4731 static void migrate_tasks(unsigned int dead_cpu)
4732 {
4733 struct rq *rq = cpu_rq(dead_cpu);
4734 struct task_struct *next, *stop = rq->stop;
4735 int dest_cpu;
4736
4737 /*
4738 * Fudge the rq selection such that the below task selection loop
4739 * doesn't get stuck on the currently eligible stop task.
4740 *
4741 * We're currently inside stop_machine() and the rq is either stuck
4742 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4743 * either way we should never end up calling schedule() until we're
4744 * done here.
4745 */
4746 rq->stop = NULL;
4747
4748 /*
4749 * put_prev_task() and pick_next_task() sched
4750 * class method both need to have an up-to-date
4751 * value of rq->clock[_task]
4752 */
4753 update_rq_clock(rq);
4754
4755 for ( ; ; ) {
4756 /*
4757 * There's this thread running, bail when that's the only
4758 * remaining thread.
4759 */
4760 if (rq->nr_running == 1)
4761 break;
4762
4763 next = pick_next_task(rq, &fake_task);
4764 BUG_ON(!next);
4765 next->sched_class->put_prev_task(rq, next);
4766
4767 /* Find suitable destination for @next, with force if needed. */
4768 dest_cpu = select_fallback_rq(dead_cpu, next);
4769 raw_spin_unlock(&rq->lock);
4770
4771 __migrate_task(next, dead_cpu, dest_cpu);
4772
4773 raw_spin_lock(&rq->lock);
4774 }
4775
4776 rq->stop = stop;
4777 }
4778
4779 #endif /* CONFIG_HOTPLUG_CPU */
4780
4781 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4782
4783 static struct ctl_table sd_ctl_dir[] = {
4784 {
4785 .procname = "sched_domain",
4786 .mode = 0555,
4787 },
4788 {}
4789 };
4790
4791 static struct ctl_table sd_ctl_root[] = {
4792 {
4793 .procname = "kernel",
4794 .mode = 0555,
4795 .child = sd_ctl_dir,
4796 },
4797 {}
4798 };
4799
4800 static struct ctl_table *sd_alloc_ctl_entry(int n)
4801 {
4802 struct ctl_table *entry =
4803 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4804
4805 return entry;
4806 }
4807
4808 static void sd_free_ctl_entry(struct ctl_table **tablep)
4809 {
4810 struct ctl_table *entry;
4811
4812 /*
4813 * In the intermediate directories, both the child directory and
4814 * procname are dynamically allocated and could fail but the mode
4815 * will always be set. In the lowest directory the names are
4816 * static strings and all have proc handlers.
4817 */
4818 for (entry = *tablep; entry->mode; entry++) {
4819 if (entry->child)
4820 sd_free_ctl_entry(&entry->child);
4821 if (entry->proc_handler == NULL)
4822 kfree(entry->procname);
4823 }
4824
4825 kfree(*tablep);
4826 *tablep = NULL;
4827 }
4828
4829 static int min_load_idx = 0;
4830 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4831
4832 static void
4833 set_table_entry(struct ctl_table *entry,
4834 const char *procname, void *data, int maxlen,
4835 umode_t mode, proc_handler *proc_handler,
4836 bool load_idx)
4837 {
4838 entry->procname = procname;
4839 entry->data = data;
4840 entry->maxlen = maxlen;
4841 entry->mode = mode;
4842 entry->proc_handler = proc_handler;
4843
4844 if (load_idx) {
4845 entry->extra1 = &min_load_idx;
4846 entry->extra2 = &max_load_idx;
4847 }
4848 }
4849
4850 static struct ctl_table *
4851 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4852 {
4853 struct ctl_table *table = sd_alloc_ctl_entry(14);
4854
4855 if (table == NULL)
4856 return NULL;
4857
4858 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4859 sizeof(long), 0644, proc_doulongvec_minmax, false);
4860 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4861 sizeof(long), 0644, proc_doulongvec_minmax, false);
4862 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4863 sizeof(int), 0644, proc_dointvec_minmax, true);
4864 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4865 sizeof(int), 0644, proc_dointvec_minmax, true);
4866 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4867 sizeof(int), 0644, proc_dointvec_minmax, true);
4868 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4869 sizeof(int), 0644, proc_dointvec_minmax, true);
4870 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4871 sizeof(int), 0644, proc_dointvec_minmax, true);
4872 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4873 sizeof(int), 0644, proc_dointvec_minmax, false);
4874 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4875 sizeof(int), 0644, proc_dointvec_minmax, false);
4876 set_table_entry(&table[9], "cache_nice_tries",
4877 &sd->cache_nice_tries,
4878 sizeof(int), 0644, proc_dointvec_minmax, false);
4879 set_table_entry(&table[10], "flags", &sd->flags,
4880 sizeof(int), 0644, proc_dointvec_minmax, false);
4881 set_table_entry(&table[11], "max_newidle_lb_cost",
4882 &sd->max_newidle_lb_cost,
4883 sizeof(long), 0644, proc_doulongvec_minmax, false);
4884 set_table_entry(&table[12], "name", sd->name,
4885 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4886 /* &table[13] is terminator */
4887
4888 return table;
4889 }
4890
4891 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4892 {
4893 struct ctl_table *entry, *table;
4894 struct sched_domain *sd;
4895 int domain_num = 0, i;
4896 char buf[32];
4897
4898 for_each_domain(cpu, sd)
4899 domain_num++;
4900 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4901 if (table == NULL)
4902 return NULL;
4903
4904 i = 0;
4905 for_each_domain(cpu, sd) {
4906 snprintf(buf, 32, "domain%d", i);
4907 entry->procname = kstrdup(buf, GFP_KERNEL);
4908 entry->mode = 0555;
4909 entry->child = sd_alloc_ctl_domain_table(sd);
4910 entry++;
4911 i++;
4912 }
4913 return table;
4914 }
4915
4916 static struct ctl_table_header *sd_sysctl_header;
4917 static void register_sched_domain_sysctl(void)
4918 {
4919 int i, cpu_num = num_possible_cpus();
4920 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4921 char buf[32];
4922
4923 WARN_ON(sd_ctl_dir[0].child);
4924 sd_ctl_dir[0].child = entry;
4925
4926 if (entry == NULL)
4927 return;
4928
4929 for_each_possible_cpu(i) {
4930 snprintf(buf, 32, "cpu%d", i);
4931 entry->procname = kstrdup(buf, GFP_KERNEL);
4932 entry->mode = 0555;
4933 entry->child = sd_alloc_ctl_cpu_table(i);
4934 entry++;
4935 }
4936
4937 WARN_ON(sd_sysctl_header);
4938 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4939 }
4940
4941 /* may be called multiple times per register */
4942 static void unregister_sched_domain_sysctl(void)
4943 {
4944 if (sd_sysctl_header)
4945 unregister_sysctl_table(sd_sysctl_header);
4946 sd_sysctl_header = NULL;
4947 if (sd_ctl_dir[0].child)
4948 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4949 }
4950 #else
4951 static void register_sched_domain_sysctl(void)
4952 {
4953 }
4954 static void unregister_sched_domain_sysctl(void)
4955 {
4956 }
4957 #endif
4958
4959 static void set_rq_online(struct rq *rq)
4960 {
4961 if (!rq->online) {
4962 const struct sched_class *class;
4963
4964 cpumask_set_cpu(rq->cpu, rq->rd->online);
4965 rq->online = 1;
4966
4967 for_each_class(class) {
4968 if (class->rq_online)
4969 class->rq_online(rq);
4970 }
4971 }
4972 }
4973
4974 static void set_rq_offline(struct rq *rq)
4975 {
4976 if (rq->online) {
4977 const struct sched_class *class;
4978
4979 for_each_class(class) {
4980 if (class->rq_offline)
4981 class->rq_offline(rq);
4982 }
4983
4984 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4985 rq->online = 0;
4986 }
4987 }
4988
4989 /*
4990 * migration_call - callback that gets triggered when a CPU is added.
4991 * Here we can start up the necessary migration thread for the new CPU.
4992 */
4993 static int
4994 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4995 {
4996 int cpu = (long)hcpu;
4997 unsigned long flags;
4998 struct rq *rq = cpu_rq(cpu);
4999
5000 switch (action & ~CPU_TASKS_FROZEN) {
5001
5002 case CPU_UP_PREPARE:
5003 rq->calc_load_update = calc_load_update;
5004 break;
5005
5006 case CPU_ONLINE:
5007 /* Update our root-domain */
5008 raw_spin_lock_irqsave(&rq->lock, flags);
5009 if (rq->rd) {
5010 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5011
5012 set_rq_online(rq);
5013 }
5014 raw_spin_unlock_irqrestore(&rq->lock, flags);
5015 break;
5016
5017 #ifdef CONFIG_HOTPLUG_CPU
5018 case CPU_DYING:
5019 sched_ttwu_pending();
5020 /* Update our root-domain */
5021 raw_spin_lock_irqsave(&rq->lock, flags);
5022 if (rq->rd) {
5023 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5024 set_rq_offline(rq);
5025 }
5026 migrate_tasks(cpu);
5027 BUG_ON(rq->nr_running != 1); /* the migration thread */
5028 raw_spin_unlock_irqrestore(&rq->lock, flags);
5029 break;
5030
5031 case CPU_DEAD:
5032 calc_load_migrate(rq);
5033 break;
5034 #endif
5035 }
5036
5037 update_max_interval();
5038
5039 return NOTIFY_OK;
5040 }
5041
5042 /*
5043 * Register at high priority so that task migration (migrate_all_tasks)
5044 * happens before everything else. This has to be lower priority than
5045 * the notifier in the perf_event subsystem, though.
5046 */
5047 static struct notifier_block migration_notifier = {
5048 .notifier_call = migration_call,
5049 .priority = CPU_PRI_MIGRATION,
5050 };
5051
5052 static int sched_cpu_active(struct notifier_block *nfb,
5053 unsigned long action, void *hcpu)
5054 {
5055 switch (action & ~CPU_TASKS_FROZEN) {
5056 case CPU_STARTING:
5057 case CPU_DOWN_FAILED:
5058 set_cpu_active((long)hcpu, true);
5059 return NOTIFY_OK;
5060 default:
5061 return NOTIFY_DONE;
5062 }
5063 }
5064
5065 static int sched_cpu_inactive(struct notifier_block *nfb,
5066 unsigned long action, void *hcpu)
5067 {
5068 unsigned long flags;
5069 long cpu = (long)hcpu;
5070
5071 switch (action & ~CPU_TASKS_FROZEN) {
5072 case CPU_DOWN_PREPARE:
5073 set_cpu_active(cpu, false);
5074
5075 /* explicitly allow suspend */
5076 if (!(action & CPU_TASKS_FROZEN)) {
5077 struct dl_bw *dl_b = dl_bw_of(cpu);
5078 bool overflow;
5079 int cpus;
5080
5081 raw_spin_lock_irqsave(&dl_b->lock, flags);
5082 cpus = dl_bw_cpus(cpu);
5083 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5084 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5085
5086 if (overflow)
5087 return notifier_from_errno(-EBUSY);
5088 }
5089 return NOTIFY_OK;
5090 }
5091
5092 return NOTIFY_DONE;
5093 }
5094
5095 static int __init migration_init(void)
5096 {
5097 void *cpu = (void *)(long)smp_processor_id();
5098 int err;
5099
5100 /* Initialize migration for the boot CPU */
5101 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5102 BUG_ON(err == NOTIFY_BAD);
5103 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5104 register_cpu_notifier(&migration_notifier);
5105
5106 /* Register cpu active notifiers */
5107 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5108 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5109
5110 return 0;
5111 }
5112 early_initcall(migration_init);
5113 #endif
5114
5115 #ifdef CONFIG_SMP
5116
5117 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5118
5119 #ifdef CONFIG_SCHED_DEBUG
5120
5121 static __read_mostly int sched_debug_enabled;
5122
5123 static int __init sched_debug_setup(char *str)
5124 {
5125 sched_debug_enabled = 1;
5126
5127 return 0;
5128 }
5129 early_param("sched_debug", sched_debug_setup);
5130
5131 static inline bool sched_debug(void)
5132 {
5133 return sched_debug_enabled;
5134 }
5135
5136 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5137 struct cpumask *groupmask)
5138 {
5139 struct sched_group *group = sd->groups;
5140 char str[256];
5141
5142 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5143 cpumask_clear(groupmask);
5144
5145 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5146
5147 if (!(sd->flags & SD_LOAD_BALANCE)) {
5148 printk("does not load-balance\n");
5149 if (sd->parent)
5150 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5151 " has parent");
5152 return -1;
5153 }
5154
5155 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5156
5157 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5158 printk(KERN_ERR "ERROR: domain->span does not contain "
5159 "CPU%d\n", cpu);
5160 }
5161 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5162 printk(KERN_ERR "ERROR: domain->groups does not contain"
5163 " CPU%d\n", cpu);
5164 }
5165
5166 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5167 do {
5168 if (!group) {
5169 printk("\n");
5170 printk(KERN_ERR "ERROR: group is NULL\n");
5171 break;
5172 }
5173
5174 /*
5175 * Even though we initialize ->power to something semi-sane,
5176 * we leave power_orig unset. This allows us to detect if
5177 * domain iteration is still funny without causing /0 traps.
5178 */
5179 if (!group->sgp->power_orig) {
5180 printk(KERN_CONT "\n");
5181 printk(KERN_ERR "ERROR: domain->cpu_power not "
5182 "set\n");
5183 break;
5184 }
5185
5186 if (!cpumask_weight(sched_group_cpus(group))) {
5187 printk(KERN_CONT "\n");
5188 printk(KERN_ERR "ERROR: empty group\n");
5189 break;
5190 }
5191
5192 if (!(sd->flags & SD_OVERLAP) &&
5193 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5194 printk(KERN_CONT "\n");
5195 printk(KERN_ERR "ERROR: repeated CPUs\n");
5196 break;
5197 }
5198
5199 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5200
5201 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5202
5203 printk(KERN_CONT " %s", str);
5204 if (group->sgp->power != SCHED_POWER_SCALE) {
5205 printk(KERN_CONT " (cpu_power = %d)",
5206 group->sgp->power);
5207 }
5208
5209 group = group->next;
5210 } while (group != sd->groups);
5211 printk(KERN_CONT "\n");
5212
5213 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5214 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5215
5216 if (sd->parent &&
5217 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5218 printk(KERN_ERR "ERROR: parent span is not a superset "
5219 "of domain->span\n");
5220 return 0;
5221 }
5222
5223 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5224 {
5225 int level = 0;
5226
5227 if (!sched_debug_enabled)
5228 return;
5229
5230 if (!sd) {
5231 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5232 return;
5233 }
5234
5235 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5236
5237 for (;;) {
5238 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5239 break;
5240 level++;
5241 sd = sd->parent;
5242 if (!sd)
5243 break;
5244 }
5245 }
5246 #else /* !CONFIG_SCHED_DEBUG */
5247 # define sched_domain_debug(sd, cpu) do { } while (0)
5248 static inline bool sched_debug(void)
5249 {
5250 return false;
5251 }
5252 #endif /* CONFIG_SCHED_DEBUG */
5253
5254 static int sd_degenerate(struct sched_domain *sd)
5255 {
5256 if (cpumask_weight(sched_domain_span(sd)) == 1)
5257 return 1;
5258
5259 /* Following flags need at least 2 groups */
5260 if (sd->flags & (SD_LOAD_BALANCE |
5261 SD_BALANCE_NEWIDLE |
5262 SD_BALANCE_FORK |
5263 SD_BALANCE_EXEC |
5264 SD_SHARE_CPUPOWER |
5265 SD_SHARE_PKG_RESOURCES)) {
5266 if (sd->groups != sd->groups->next)
5267 return 0;
5268 }
5269
5270 /* Following flags don't use groups */
5271 if (sd->flags & (SD_WAKE_AFFINE))
5272 return 0;
5273
5274 return 1;
5275 }
5276
5277 static int
5278 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5279 {
5280 unsigned long cflags = sd->flags, pflags = parent->flags;
5281
5282 if (sd_degenerate(parent))
5283 return 1;
5284
5285 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5286 return 0;
5287
5288 /* Flags needing groups don't count if only 1 group in parent */
5289 if (parent->groups == parent->groups->next) {
5290 pflags &= ~(SD_LOAD_BALANCE |
5291 SD_BALANCE_NEWIDLE |
5292 SD_BALANCE_FORK |
5293 SD_BALANCE_EXEC |
5294 SD_SHARE_CPUPOWER |
5295 SD_SHARE_PKG_RESOURCES |
5296 SD_PREFER_SIBLING);
5297 if (nr_node_ids == 1)
5298 pflags &= ~SD_SERIALIZE;
5299 }
5300 if (~cflags & pflags)
5301 return 0;
5302
5303 return 1;
5304 }
5305
5306 static void free_rootdomain(struct rcu_head *rcu)
5307 {
5308 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5309
5310 cpupri_cleanup(&rd->cpupri);
5311 cpudl_cleanup(&rd->cpudl);
5312 free_cpumask_var(rd->dlo_mask);
5313 free_cpumask_var(rd->rto_mask);
5314 free_cpumask_var(rd->online);
5315 free_cpumask_var(rd->span);
5316 kfree(rd);
5317 }
5318
5319 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5320 {
5321 struct root_domain *old_rd = NULL;
5322 unsigned long flags;
5323
5324 raw_spin_lock_irqsave(&rq->lock, flags);
5325
5326 if (rq->rd) {
5327 old_rd = rq->rd;
5328
5329 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5330 set_rq_offline(rq);
5331
5332 cpumask_clear_cpu(rq->cpu, old_rd->span);
5333
5334 /*
5335 * If we dont want to free the old_rd yet then
5336 * set old_rd to NULL to skip the freeing later
5337 * in this function:
5338 */
5339 if (!atomic_dec_and_test(&old_rd->refcount))
5340 old_rd = NULL;
5341 }
5342
5343 atomic_inc(&rd->refcount);
5344 rq->rd = rd;
5345
5346 cpumask_set_cpu(rq->cpu, rd->span);
5347 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5348 set_rq_online(rq);
5349
5350 raw_spin_unlock_irqrestore(&rq->lock, flags);
5351
5352 if (old_rd)
5353 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5354 }
5355
5356 static int init_rootdomain(struct root_domain *rd)
5357 {
5358 memset(rd, 0, sizeof(*rd));
5359
5360 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5361 goto out;
5362 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5363 goto free_span;
5364 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5365 goto free_online;
5366 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5367 goto free_dlo_mask;
5368
5369 init_dl_bw(&rd->dl_bw);
5370 if (cpudl_init(&rd->cpudl) != 0)
5371 goto free_dlo_mask;
5372
5373 if (cpupri_init(&rd->cpupri) != 0)
5374 goto free_rto_mask;
5375 return 0;
5376
5377 free_rto_mask:
5378 free_cpumask_var(rd->rto_mask);
5379 free_dlo_mask:
5380 free_cpumask_var(rd->dlo_mask);
5381 free_online:
5382 free_cpumask_var(rd->online);
5383 free_span:
5384 free_cpumask_var(rd->span);
5385 out:
5386 return -ENOMEM;
5387 }
5388
5389 /*
5390 * By default the system creates a single root-domain with all cpus as
5391 * members (mimicking the global state we have today).
5392 */
5393 struct root_domain def_root_domain;
5394
5395 static void init_defrootdomain(void)
5396 {
5397 init_rootdomain(&def_root_domain);
5398
5399 atomic_set(&def_root_domain.refcount, 1);
5400 }
5401
5402 static struct root_domain *alloc_rootdomain(void)
5403 {
5404 struct root_domain *rd;
5405
5406 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5407 if (!rd)
5408 return NULL;
5409
5410 if (init_rootdomain(rd) != 0) {
5411 kfree(rd);
5412 return NULL;
5413 }
5414
5415 return rd;
5416 }
5417
5418 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5419 {
5420 struct sched_group *tmp, *first;
5421
5422 if (!sg)
5423 return;
5424
5425 first = sg;
5426 do {
5427 tmp = sg->next;
5428
5429 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5430 kfree(sg->sgp);
5431
5432 kfree(sg);
5433 sg = tmp;
5434 } while (sg != first);
5435 }
5436
5437 static void free_sched_domain(struct rcu_head *rcu)
5438 {
5439 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5440
5441 /*
5442 * If its an overlapping domain it has private groups, iterate and
5443 * nuke them all.
5444 */
5445 if (sd->flags & SD_OVERLAP) {
5446 free_sched_groups(sd->groups, 1);
5447 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5448 kfree(sd->groups->sgp);
5449 kfree(sd->groups);
5450 }
5451 kfree(sd);
5452 }
5453
5454 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5455 {
5456 call_rcu(&sd->rcu, free_sched_domain);
5457 }
5458
5459 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5460 {
5461 for (; sd; sd = sd->parent)
5462 destroy_sched_domain(sd, cpu);
5463 }
5464
5465 /*
5466 * Keep a special pointer to the highest sched_domain that has
5467 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5468 * allows us to avoid some pointer chasing select_idle_sibling().
5469 *
5470 * Also keep a unique ID per domain (we use the first cpu number in
5471 * the cpumask of the domain), this allows us to quickly tell if
5472 * two cpus are in the same cache domain, see cpus_share_cache().
5473 */
5474 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5475 DEFINE_PER_CPU(int, sd_llc_size);
5476 DEFINE_PER_CPU(int, sd_llc_id);
5477 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5478 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5479 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5480
5481 static void update_top_cache_domain(int cpu)
5482 {
5483 struct sched_domain *sd;
5484 struct sched_domain *busy_sd = NULL;
5485 int id = cpu;
5486 int size = 1;
5487
5488 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5489 if (sd) {
5490 id = cpumask_first(sched_domain_span(sd));
5491 size = cpumask_weight(sched_domain_span(sd));
5492 busy_sd = sd->parent; /* sd_busy */
5493 }
5494 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5495
5496 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5497 per_cpu(sd_llc_size, cpu) = size;
5498 per_cpu(sd_llc_id, cpu) = id;
5499
5500 sd = lowest_flag_domain(cpu, SD_NUMA);
5501 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5502
5503 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5504 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5505 }
5506
5507 /*
5508 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5509 * hold the hotplug lock.
5510 */
5511 static void
5512 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5513 {
5514 struct rq *rq = cpu_rq(cpu);
5515 struct sched_domain *tmp;
5516
5517 /* Remove the sched domains which do not contribute to scheduling. */
5518 for (tmp = sd; tmp; ) {
5519 struct sched_domain *parent = tmp->parent;
5520 if (!parent)
5521 break;
5522
5523 if (sd_parent_degenerate(tmp, parent)) {
5524 tmp->parent = parent->parent;
5525 if (parent->parent)
5526 parent->parent->child = tmp;
5527 /*
5528 * Transfer SD_PREFER_SIBLING down in case of a
5529 * degenerate parent; the spans match for this
5530 * so the property transfers.
5531 */
5532 if (parent->flags & SD_PREFER_SIBLING)
5533 tmp->flags |= SD_PREFER_SIBLING;
5534 destroy_sched_domain(parent, cpu);
5535 } else
5536 tmp = tmp->parent;
5537 }
5538
5539 if (sd && sd_degenerate(sd)) {
5540 tmp = sd;
5541 sd = sd->parent;
5542 destroy_sched_domain(tmp, cpu);
5543 if (sd)
5544 sd->child = NULL;
5545 }
5546
5547 sched_domain_debug(sd, cpu);
5548
5549 rq_attach_root(rq, rd);
5550 tmp = rq->sd;
5551 rcu_assign_pointer(rq->sd, sd);
5552 destroy_sched_domains(tmp, cpu);
5553
5554 update_top_cache_domain(cpu);
5555 }
5556
5557 /* cpus with isolated domains */
5558 static cpumask_var_t cpu_isolated_map;
5559
5560 /* Setup the mask of cpus configured for isolated domains */
5561 static int __init isolated_cpu_setup(char *str)
5562 {
5563 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5564 cpulist_parse(str, cpu_isolated_map);
5565 return 1;
5566 }
5567
5568 __setup("isolcpus=", isolated_cpu_setup);
5569
5570 static const struct cpumask *cpu_cpu_mask(int cpu)
5571 {
5572 return cpumask_of_node(cpu_to_node(cpu));
5573 }
5574
5575 struct sd_data {
5576 struct sched_domain **__percpu sd;
5577 struct sched_group **__percpu sg;
5578 struct sched_group_power **__percpu sgp;
5579 };
5580
5581 struct s_data {
5582 struct sched_domain ** __percpu sd;
5583 struct root_domain *rd;
5584 };
5585
5586 enum s_alloc {
5587 sa_rootdomain,
5588 sa_sd,
5589 sa_sd_storage,
5590 sa_none,
5591 };
5592
5593 struct sched_domain_topology_level;
5594
5595 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5596 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5597
5598 #define SDTL_OVERLAP 0x01
5599
5600 struct sched_domain_topology_level {
5601 sched_domain_init_f init;
5602 sched_domain_mask_f mask;
5603 int flags;
5604 int numa_level;
5605 struct sd_data data;
5606 };
5607
5608 /*
5609 * Build an iteration mask that can exclude certain CPUs from the upwards
5610 * domain traversal.
5611 *
5612 * Asymmetric node setups can result in situations where the domain tree is of
5613 * unequal depth, make sure to skip domains that already cover the entire
5614 * range.
5615 *
5616 * In that case build_sched_domains() will have terminated the iteration early
5617 * and our sibling sd spans will be empty. Domains should always include the
5618 * cpu they're built on, so check that.
5619 *
5620 */
5621 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5622 {
5623 const struct cpumask *span = sched_domain_span(sd);
5624 struct sd_data *sdd = sd->private;
5625 struct sched_domain *sibling;
5626 int i;
5627
5628 for_each_cpu(i, span) {
5629 sibling = *per_cpu_ptr(sdd->sd, i);
5630 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5631 continue;
5632
5633 cpumask_set_cpu(i, sched_group_mask(sg));
5634 }
5635 }
5636
5637 /*
5638 * Return the canonical balance cpu for this group, this is the first cpu
5639 * of this group that's also in the iteration mask.
5640 */
5641 int group_balance_cpu(struct sched_group *sg)
5642 {
5643 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5644 }
5645
5646 static int
5647 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5648 {
5649 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5650 const struct cpumask *span = sched_domain_span(sd);
5651 struct cpumask *covered = sched_domains_tmpmask;
5652 struct sd_data *sdd = sd->private;
5653 struct sched_domain *child;
5654 int i;
5655
5656 cpumask_clear(covered);
5657
5658 for_each_cpu(i, span) {
5659 struct cpumask *sg_span;
5660
5661 if (cpumask_test_cpu(i, covered))
5662 continue;
5663
5664 child = *per_cpu_ptr(sdd->sd, i);
5665
5666 /* See the comment near build_group_mask(). */
5667 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5668 continue;
5669
5670 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5671 GFP_KERNEL, cpu_to_node(cpu));
5672
5673 if (!sg)
5674 goto fail;
5675
5676 sg_span = sched_group_cpus(sg);
5677 if (child->child) {
5678 child = child->child;
5679 cpumask_copy(sg_span, sched_domain_span(child));
5680 } else
5681 cpumask_set_cpu(i, sg_span);
5682
5683 cpumask_or(covered, covered, sg_span);
5684
5685 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5686 if (atomic_inc_return(&sg->sgp->ref) == 1)
5687 build_group_mask(sd, sg);
5688
5689 /*
5690 * Initialize sgp->power such that even if we mess up the
5691 * domains and no possible iteration will get us here, we won't
5692 * die on a /0 trap.
5693 */
5694 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5695 sg->sgp->power_orig = sg->sgp->power;
5696
5697 /*
5698 * Make sure the first group of this domain contains the
5699 * canonical balance cpu. Otherwise the sched_domain iteration
5700 * breaks. See update_sg_lb_stats().
5701 */
5702 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5703 group_balance_cpu(sg) == cpu)
5704 groups = sg;
5705
5706 if (!first)
5707 first = sg;
5708 if (last)
5709 last->next = sg;
5710 last = sg;
5711 last->next = first;
5712 }
5713 sd->groups = groups;
5714
5715 return 0;
5716
5717 fail:
5718 free_sched_groups(first, 0);
5719
5720 return -ENOMEM;
5721 }
5722
5723 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5724 {
5725 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5726 struct sched_domain *child = sd->child;
5727
5728 if (child)
5729 cpu = cpumask_first(sched_domain_span(child));
5730
5731 if (sg) {
5732 *sg = *per_cpu_ptr(sdd->sg, cpu);
5733 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5734 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5735 }
5736
5737 return cpu;
5738 }
5739
5740 /*
5741 * build_sched_groups will build a circular linked list of the groups
5742 * covered by the given span, and will set each group's ->cpumask correctly,
5743 * and ->cpu_power to 0.
5744 *
5745 * Assumes the sched_domain tree is fully constructed
5746 */
5747 static int
5748 build_sched_groups(struct sched_domain *sd, int cpu)
5749 {
5750 struct sched_group *first = NULL, *last = NULL;
5751 struct sd_data *sdd = sd->private;
5752 const struct cpumask *span = sched_domain_span(sd);
5753 struct cpumask *covered;
5754 int i;
5755
5756 get_group(cpu, sdd, &sd->groups);
5757 atomic_inc(&sd->groups->ref);
5758
5759 if (cpu != cpumask_first(span))
5760 return 0;
5761
5762 lockdep_assert_held(&sched_domains_mutex);
5763 covered = sched_domains_tmpmask;
5764
5765 cpumask_clear(covered);
5766
5767 for_each_cpu(i, span) {
5768 struct sched_group *sg;
5769 int group, j;
5770
5771 if (cpumask_test_cpu(i, covered))
5772 continue;
5773
5774 group = get_group(i, sdd, &sg);
5775 cpumask_clear(sched_group_cpus(sg));
5776 sg->sgp->power = 0;
5777 cpumask_setall(sched_group_mask(sg));
5778
5779 for_each_cpu(j, span) {
5780 if (get_group(j, sdd, NULL) != group)
5781 continue;
5782
5783 cpumask_set_cpu(j, covered);
5784 cpumask_set_cpu(j, sched_group_cpus(sg));
5785 }
5786
5787 if (!first)
5788 first = sg;
5789 if (last)
5790 last->next = sg;
5791 last = sg;
5792 }
5793 last->next = first;
5794
5795 return 0;
5796 }
5797
5798 /*
5799 * Initialize sched groups cpu_power.
5800 *
5801 * cpu_power indicates the capacity of sched group, which is used while
5802 * distributing the load between different sched groups in a sched domain.
5803 * Typically cpu_power for all the groups in a sched domain will be same unless
5804 * there are asymmetries in the topology. If there are asymmetries, group
5805 * having more cpu_power will pickup more load compared to the group having
5806 * less cpu_power.
5807 */
5808 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5809 {
5810 struct sched_group *sg = sd->groups;
5811
5812 WARN_ON(!sg);
5813
5814 do {
5815 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5816 sg = sg->next;
5817 } while (sg != sd->groups);
5818
5819 if (cpu != group_balance_cpu(sg))
5820 return;
5821
5822 update_group_power(sd, cpu);
5823 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5824 }
5825
5826 int __weak arch_sd_sibling_asym_packing(void)
5827 {
5828 return 0*SD_ASYM_PACKING;
5829 }
5830
5831 /*
5832 * Initializers for schedule domains
5833 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5834 */
5835
5836 #ifdef CONFIG_SCHED_DEBUG
5837 # define SD_INIT_NAME(sd, type) sd->name = #type
5838 #else
5839 # define SD_INIT_NAME(sd, type) do { } while (0)
5840 #endif
5841
5842 #define SD_INIT_FUNC(type) \
5843 static noinline struct sched_domain * \
5844 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5845 { \
5846 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5847 *sd = SD_##type##_INIT; \
5848 SD_INIT_NAME(sd, type); \
5849 sd->private = &tl->data; \
5850 return sd; \
5851 }
5852
5853 SD_INIT_FUNC(CPU)
5854 #ifdef CONFIG_SCHED_SMT
5855 SD_INIT_FUNC(SIBLING)
5856 #endif
5857 #ifdef CONFIG_SCHED_MC
5858 SD_INIT_FUNC(MC)
5859 #endif
5860 #ifdef CONFIG_SCHED_BOOK
5861 SD_INIT_FUNC(BOOK)
5862 #endif
5863
5864 static int default_relax_domain_level = -1;
5865 int sched_domain_level_max;
5866
5867 static int __init setup_relax_domain_level(char *str)
5868 {
5869 if (kstrtoint(str, 0, &default_relax_domain_level))
5870 pr_warn("Unable to set relax_domain_level\n");
5871
5872 return 1;
5873 }
5874 __setup("relax_domain_level=", setup_relax_domain_level);
5875
5876 static void set_domain_attribute(struct sched_domain *sd,
5877 struct sched_domain_attr *attr)
5878 {
5879 int request;
5880
5881 if (!attr || attr->relax_domain_level < 0) {
5882 if (default_relax_domain_level < 0)
5883 return;
5884 else
5885 request = default_relax_domain_level;
5886 } else
5887 request = attr->relax_domain_level;
5888 if (request < sd->level) {
5889 /* turn off idle balance on this domain */
5890 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5891 } else {
5892 /* turn on idle balance on this domain */
5893 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5894 }
5895 }
5896
5897 static void __sdt_free(const struct cpumask *cpu_map);
5898 static int __sdt_alloc(const struct cpumask *cpu_map);
5899
5900 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5901 const struct cpumask *cpu_map)
5902 {
5903 switch (what) {
5904 case sa_rootdomain:
5905 if (!atomic_read(&d->rd->refcount))
5906 free_rootdomain(&d->rd->rcu); /* fall through */
5907 case sa_sd:
5908 free_percpu(d->sd); /* fall through */
5909 case sa_sd_storage:
5910 __sdt_free(cpu_map); /* fall through */
5911 case sa_none:
5912 break;
5913 }
5914 }
5915
5916 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5917 const struct cpumask *cpu_map)
5918 {
5919 memset(d, 0, sizeof(*d));
5920
5921 if (__sdt_alloc(cpu_map))
5922 return sa_sd_storage;
5923 d->sd = alloc_percpu(struct sched_domain *);
5924 if (!d->sd)
5925 return sa_sd_storage;
5926 d->rd = alloc_rootdomain();
5927 if (!d->rd)
5928 return sa_sd;
5929 return sa_rootdomain;
5930 }
5931
5932 /*
5933 * NULL the sd_data elements we've used to build the sched_domain and
5934 * sched_group structure so that the subsequent __free_domain_allocs()
5935 * will not free the data we're using.
5936 */
5937 static void claim_allocations(int cpu, struct sched_domain *sd)
5938 {
5939 struct sd_data *sdd = sd->private;
5940
5941 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5942 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5943
5944 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5945 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5946
5947 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5948 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5949 }
5950
5951 #ifdef CONFIG_SCHED_SMT
5952 static const struct cpumask *cpu_smt_mask(int cpu)
5953 {
5954 return topology_thread_cpumask(cpu);
5955 }
5956 #endif
5957
5958 /*
5959 * Topology list, bottom-up.
5960 */
5961 static struct sched_domain_topology_level default_topology[] = {
5962 #ifdef CONFIG_SCHED_SMT
5963 { sd_init_SIBLING, cpu_smt_mask, },
5964 #endif
5965 #ifdef CONFIG_SCHED_MC
5966 { sd_init_MC, cpu_coregroup_mask, },
5967 #endif
5968 #ifdef CONFIG_SCHED_BOOK
5969 { sd_init_BOOK, cpu_book_mask, },
5970 #endif
5971 { sd_init_CPU, cpu_cpu_mask, },
5972 { NULL, },
5973 };
5974
5975 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5976
5977 #define for_each_sd_topology(tl) \
5978 for (tl = sched_domain_topology; tl->init; tl++)
5979
5980 #ifdef CONFIG_NUMA
5981
5982 static int sched_domains_numa_levels;
5983 static int *sched_domains_numa_distance;
5984 static struct cpumask ***sched_domains_numa_masks;
5985 static int sched_domains_curr_level;
5986
5987 static inline int sd_local_flags(int level)
5988 {
5989 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5990 return 0;
5991
5992 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5993 }
5994
5995 static struct sched_domain *
5996 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5997 {
5998 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5999 int level = tl->numa_level;
6000 int sd_weight = cpumask_weight(
6001 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6002
6003 *sd = (struct sched_domain){
6004 .min_interval = sd_weight,
6005 .max_interval = 2*sd_weight,
6006 .busy_factor = 32,
6007 .imbalance_pct = 125,
6008 .cache_nice_tries = 2,
6009 .busy_idx = 3,
6010 .idle_idx = 2,
6011 .newidle_idx = 0,
6012 .wake_idx = 0,
6013 .forkexec_idx = 0,
6014
6015 .flags = 1*SD_LOAD_BALANCE
6016 | 1*SD_BALANCE_NEWIDLE
6017 | 0*SD_BALANCE_EXEC
6018 | 0*SD_BALANCE_FORK
6019 | 0*SD_BALANCE_WAKE
6020 | 0*SD_WAKE_AFFINE
6021 | 0*SD_SHARE_CPUPOWER
6022 | 0*SD_SHARE_PKG_RESOURCES
6023 | 1*SD_SERIALIZE
6024 | 0*SD_PREFER_SIBLING
6025 | 1*SD_NUMA
6026 | sd_local_flags(level)
6027 ,
6028 .last_balance = jiffies,
6029 .balance_interval = sd_weight,
6030 .max_newidle_lb_cost = 0,
6031 .next_decay_max_lb_cost = jiffies,
6032 };
6033 SD_INIT_NAME(sd, NUMA);
6034 sd->private = &tl->data;
6035
6036 /*
6037 * Ugly hack to pass state to sd_numa_mask()...
6038 */
6039 sched_domains_curr_level = tl->numa_level;
6040
6041 return sd;
6042 }
6043
6044 static const struct cpumask *sd_numa_mask(int cpu)
6045 {
6046 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6047 }
6048
6049 static void sched_numa_warn(const char *str)
6050 {
6051 static int done = false;
6052 int i,j;
6053
6054 if (done)
6055 return;
6056
6057 done = true;
6058
6059 printk(KERN_WARNING "ERROR: %s\n\n", str);
6060
6061 for (i = 0; i < nr_node_ids; i++) {
6062 printk(KERN_WARNING " ");
6063 for (j = 0; j < nr_node_ids; j++)
6064 printk(KERN_CONT "%02d ", node_distance(i,j));
6065 printk(KERN_CONT "\n");
6066 }
6067 printk(KERN_WARNING "\n");
6068 }
6069
6070 static bool find_numa_distance(int distance)
6071 {
6072 int i;
6073
6074 if (distance == node_distance(0, 0))
6075 return true;
6076
6077 for (i = 0; i < sched_domains_numa_levels; i++) {
6078 if (sched_domains_numa_distance[i] == distance)
6079 return true;
6080 }
6081
6082 return false;
6083 }
6084
6085 static void sched_init_numa(void)
6086 {
6087 int next_distance, curr_distance = node_distance(0, 0);
6088 struct sched_domain_topology_level *tl;
6089 int level = 0;
6090 int i, j, k;
6091
6092 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6093 if (!sched_domains_numa_distance)
6094 return;
6095
6096 /*
6097 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6098 * unique distances in the node_distance() table.
6099 *
6100 * Assumes node_distance(0,j) includes all distances in
6101 * node_distance(i,j) in order to avoid cubic time.
6102 */
6103 next_distance = curr_distance;
6104 for (i = 0; i < nr_node_ids; i++) {
6105 for (j = 0; j < nr_node_ids; j++) {
6106 for (k = 0; k < nr_node_ids; k++) {
6107 int distance = node_distance(i, k);
6108
6109 if (distance > curr_distance &&
6110 (distance < next_distance ||
6111 next_distance == curr_distance))
6112 next_distance = distance;
6113
6114 /*
6115 * While not a strong assumption it would be nice to know
6116 * about cases where if node A is connected to B, B is not
6117 * equally connected to A.
6118 */
6119 if (sched_debug() && node_distance(k, i) != distance)
6120 sched_numa_warn("Node-distance not symmetric");
6121
6122 if (sched_debug() && i && !find_numa_distance(distance))
6123 sched_numa_warn("Node-0 not representative");
6124 }
6125 if (next_distance != curr_distance) {
6126 sched_domains_numa_distance[level++] = next_distance;
6127 sched_domains_numa_levels = level;
6128 curr_distance = next_distance;
6129 } else break;
6130 }
6131
6132 /*
6133 * In case of sched_debug() we verify the above assumption.
6134 */
6135 if (!sched_debug())
6136 break;
6137 }
6138 /*
6139 * 'level' contains the number of unique distances, excluding the
6140 * identity distance node_distance(i,i).
6141 *
6142 * The sched_domains_numa_distance[] array includes the actual distance
6143 * numbers.
6144 */
6145
6146 /*
6147 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6148 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6149 * the array will contain less then 'level' members. This could be
6150 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6151 * in other functions.
6152 *
6153 * We reset it to 'level' at the end of this function.
6154 */
6155 sched_domains_numa_levels = 0;
6156
6157 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6158 if (!sched_domains_numa_masks)
6159 return;
6160
6161 /*
6162 * Now for each level, construct a mask per node which contains all
6163 * cpus of nodes that are that many hops away from us.
6164 */
6165 for (i = 0; i < level; i++) {
6166 sched_domains_numa_masks[i] =
6167 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6168 if (!sched_domains_numa_masks[i])
6169 return;
6170
6171 for (j = 0; j < nr_node_ids; j++) {
6172 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6173 if (!mask)
6174 return;
6175
6176 sched_domains_numa_masks[i][j] = mask;
6177
6178 for (k = 0; k < nr_node_ids; k++) {
6179 if (node_distance(j, k) > sched_domains_numa_distance[i])
6180 continue;
6181
6182 cpumask_or(mask, mask, cpumask_of_node(k));
6183 }
6184 }
6185 }
6186
6187 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6188 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6189 if (!tl)
6190 return;
6191
6192 /*
6193 * Copy the default topology bits..
6194 */
6195 for (i = 0; default_topology[i].init; i++)
6196 tl[i] = default_topology[i];
6197
6198 /*
6199 * .. and append 'j' levels of NUMA goodness.
6200 */
6201 for (j = 0; j < level; i++, j++) {
6202 tl[i] = (struct sched_domain_topology_level){
6203 .init = sd_numa_init,
6204 .mask = sd_numa_mask,
6205 .flags = SDTL_OVERLAP,
6206 .numa_level = j,
6207 };
6208 }
6209
6210 sched_domain_topology = tl;
6211
6212 sched_domains_numa_levels = level;
6213 }
6214
6215 static void sched_domains_numa_masks_set(int cpu)
6216 {
6217 int i, j;
6218 int node = cpu_to_node(cpu);
6219
6220 for (i = 0; i < sched_domains_numa_levels; i++) {
6221 for (j = 0; j < nr_node_ids; j++) {
6222 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6223 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6224 }
6225 }
6226 }
6227
6228 static void sched_domains_numa_masks_clear(int cpu)
6229 {
6230 int i, j;
6231 for (i = 0; i < sched_domains_numa_levels; i++) {
6232 for (j = 0; j < nr_node_ids; j++)
6233 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6234 }
6235 }
6236
6237 /*
6238 * Update sched_domains_numa_masks[level][node] array when new cpus
6239 * are onlined.
6240 */
6241 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6242 unsigned long action,
6243 void *hcpu)
6244 {
6245 int cpu = (long)hcpu;
6246
6247 switch (action & ~CPU_TASKS_FROZEN) {
6248 case CPU_ONLINE:
6249 sched_domains_numa_masks_set(cpu);
6250 break;
6251
6252 case CPU_DEAD:
6253 sched_domains_numa_masks_clear(cpu);
6254 break;
6255
6256 default:
6257 return NOTIFY_DONE;
6258 }
6259
6260 return NOTIFY_OK;
6261 }
6262 #else
6263 static inline void sched_init_numa(void)
6264 {
6265 }
6266
6267 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6268 unsigned long action,
6269 void *hcpu)
6270 {
6271 return 0;
6272 }
6273 #endif /* CONFIG_NUMA */
6274
6275 static int __sdt_alloc(const struct cpumask *cpu_map)
6276 {
6277 struct sched_domain_topology_level *tl;
6278 int j;
6279
6280 for_each_sd_topology(tl) {
6281 struct sd_data *sdd = &tl->data;
6282
6283 sdd->sd = alloc_percpu(struct sched_domain *);
6284 if (!sdd->sd)
6285 return -ENOMEM;
6286
6287 sdd->sg = alloc_percpu(struct sched_group *);
6288 if (!sdd->sg)
6289 return -ENOMEM;
6290
6291 sdd->sgp = alloc_percpu(struct sched_group_power *);
6292 if (!sdd->sgp)
6293 return -ENOMEM;
6294
6295 for_each_cpu(j, cpu_map) {
6296 struct sched_domain *sd;
6297 struct sched_group *sg;
6298 struct sched_group_power *sgp;
6299
6300 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6301 GFP_KERNEL, cpu_to_node(j));
6302 if (!sd)
6303 return -ENOMEM;
6304
6305 *per_cpu_ptr(sdd->sd, j) = sd;
6306
6307 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6308 GFP_KERNEL, cpu_to_node(j));
6309 if (!sg)
6310 return -ENOMEM;
6311
6312 sg->next = sg;
6313
6314 *per_cpu_ptr(sdd->sg, j) = sg;
6315
6316 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6317 GFP_KERNEL, cpu_to_node(j));
6318 if (!sgp)
6319 return -ENOMEM;
6320
6321 *per_cpu_ptr(sdd->sgp, j) = sgp;
6322 }
6323 }
6324
6325 return 0;
6326 }
6327
6328 static void __sdt_free(const struct cpumask *cpu_map)
6329 {
6330 struct sched_domain_topology_level *tl;
6331 int j;
6332
6333 for_each_sd_topology(tl) {
6334 struct sd_data *sdd = &tl->data;
6335
6336 for_each_cpu(j, cpu_map) {
6337 struct sched_domain *sd;
6338
6339 if (sdd->sd) {
6340 sd = *per_cpu_ptr(sdd->sd, j);
6341 if (sd && (sd->flags & SD_OVERLAP))
6342 free_sched_groups(sd->groups, 0);
6343 kfree(*per_cpu_ptr(sdd->sd, j));
6344 }
6345
6346 if (sdd->sg)
6347 kfree(*per_cpu_ptr(sdd->sg, j));
6348 if (sdd->sgp)
6349 kfree(*per_cpu_ptr(sdd->sgp, j));
6350 }
6351 free_percpu(sdd->sd);
6352 sdd->sd = NULL;
6353 free_percpu(sdd->sg);
6354 sdd->sg = NULL;
6355 free_percpu(sdd->sgp);
6356 sdd->sgp = NULL;
6357 }
6358 }
6359
6360 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6361 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6362 struct sched_domain *child, int cpu)
6363 {
6364 struct sched_domain *sd = tl->init(tl, cpu);
6365 if (!sd)
6366 return child;
6367
6368 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6369 if (child) {
6370 sd->level = child->level + 1;
6371 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6372 child->parent = sd;
6373 sd->child = child;
6374 }
6375 set_domain_attribute(sd, attr);
6376
6377 return sd;
6378 }
6379
6380 /*
6381 * Build sched domains for a given set of cpus and attach the sched domains
6382 * to the individual cpus
6383 */
6384 static int build_sched_domains(const struct cpumask *cpu_map,
6385 struct sched_domain_attr *attr)
6386 {
6387 enum s_alloc alloc_state;
6388 struct sched_domain *sd;
6389 struct s_data d;
6390 int i, ret = -ENOMEM;
6391
6392 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6393 if (alloc_state != sa_rootdomain)
6394 goto error;
6395
6396 /* Set up domains for cpus specified by the cpu_map. */
6397 for_each_cpu(i, cpu_map) {
6398 struct sched_domain_topology_level *tl;
6399
6400 sd = NULL;
6401 for_each_sd_topology(tl) {
6402 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6403 if (tl == sched_domain_topology)
6404 *per_cpu_ptr(d.sd, i) = sd;
6405 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6406 sd->flags |= SD_OVERLAP;
6407 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6408 break;
6409 }
6410 }
6411
6412 /* Build the groups for the domains */
6413 for_each_cpu(i, cpu_map) {
6414 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6415 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6416 if (sd->flags & SD_OVERLAP) {
6417 if (build_overlap_sched_groups(sd, i))
6418 goto error;
6419 } else {
6420 if (build_sched_groups(sd, i))
6421 goto error;
6422 }
6423 }
6424 }
6425
6426 /* Calculate CPU power for physical packages and nodes */
6427 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6428 if (!cpumask_test_cpu(i, cpu_map))
6429 continue;
6430
6431 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6432 claim_allocations(i, sd);
6433 init_sched_groups_power(i, sd);
6434 }
6435 }
6436
6437 /* Attach the domains */
6438 rcu_read_lock();
6439 for_each_cpu(i, cpu_map) {
6440 sd = *per_cpu_ptr(d.sd, i);
6441 cpu_attach_domain(sd, d.rd, i);
6442 }
6443 rcu_read_unlock();
6444
6445 ret = 0;
6446 error:
6447 __free_domain_allocs(&d, alloc_state, cpu_map);
6448 return ret;
6449 }
6450
6451 static cpumask_var_t *doms_cur; /* current sched domains */
6452 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6453 static struct sched_domain_attr *dattr_cur;
6454 /* attribues of custom domains in 'doms_cur' */
6455
6456 /*
6457 * Special case: If a kmalloc of a doms_cur partition (array of
6458 * cpumask) fails, then fallback to a single sched domain,
6459 * as determined by the single cpumask fallback_doms.
6460 */
6461 static cpumask_var_t fallback_doms;
6462
6463 /*
6464 * arch_update_cpu_topology lets virtualized architectures update the
6465 * cpu core maps. It is supposed to return 1 if the topology changed
6466 * or 0 if it stayed the same.
6467 */
6468 int __weak arch_update_cpu_topology(void)
6469 {
6470 return 0;
6471 }
6472
6473 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6474 {
6475 int i;
6476 cpumask_var_t *doms;
6477
6478 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6479 if (!doms)
6480 return NULL;
6481 for (i = 0; i < ndoms; i++) {
6482 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6483 free_sched_domains(doms, i);
6484 return NULL;
6485 }
6486 }
6487 return doms;
6488 }
6489
6490 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6491 {
6492 unsigned int i;
6493 for (i = 0; i < ndoms; i++)
6494 free_cpumask_var(doms[i]);
6495 kfree(doms);
6496 }
6497
6498 /*
6499 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6500 * For now this just excludes isolated cpus, but could be used to
6501 * exclude other special cases in the future.
6502 */
6503 static int init_sched_domains(const struct cpumask *cpu_map)
6504 {
6505 int err;
6506
6507 arch_update_cpu_topology();
6508 ndoms_cur = 1;
6509 doms_cur = alloc_sched_domains(ndoms_cur);
6510 if (!doms_cur)
6511 doms_cur = &fallback_doms;
6512 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6513 err = build_sched_domains(doms_cur[0], NULL);
6514 register_sched_domain_sysctl();
6515
6516 return err;
6517 }
6518
6519 /*
6520 * Detach sched domains from a group of cpus specified in cpu_map
6521 * These cpus will now be attached to the NULL domain
6522 */
6523 static void detach_destroy_domains(const struct cpumask *cpu_map)
6524 {
6525 int i;
6526
6527 rcu_read_lock();
6528 for_each_cpu(i, cpu_map)
6529 cpu_attach_domain(NULL, &def_root_domain, i);
6530 rcu_read_unlock();
6531 }
6532
6533 /* handle null as "default" */
6534 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6535 struct sched_domain_attr *new, int idx_new)
6536 {
6537 struct sched_domain_attr tmp;
6538
6539 /* fast path */
6540 if (!new && !cur)
6541 return 1;
6542
6543 tmp = SD_ATTR_INIT;
6544 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6545 new ? (new + idx_new) : &tmp,
6546 sizeof(struct sched_domain_attr));
6547 }
6548
6549 /*
6550 * Partition sched domains as specified by the 'ndoms_new'
6551 * cpumasks in the array doms_new[] of cpumasks. This compares
6552 * doms_new[] to the current sched domain partitioning, doms_cur[].
6553 * It destroys each deleted domain and builds each new domain.
6554 *
6555 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6556 * The masks don't intersect (don't overlap.) We should setup one
6557 * sched domain for each mask. CPUs not in any of the cpumasks will
6558 * not be load balanced. If the same cpumask appears both in the
6559 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6560 * it as it is.
6561 *
6562 * The passed in 'doms_new' should be allocated using
6563 * alloc_sched_domains. This routine takes ownership of it and will
6564 * free_sched_domains it when done with it. If the caller failed the
6565 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6566 * and partition_sched_domains() will fallback to the single partition
6567 * 'fallback_doms', it also forces the domains to be rebuilt.
6568 *
6569 * If doms_new == NULL it will be replaced with cpu_online_mask.
6570 * ndoms_new == 0 is a special case for destroying existing domains,
6571 * and it will not create the default domain.
6572 *
6573 * Call with hotplug lock held
6574 */
6575 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6576 struct sched_domain_attr *dattr_new)
6577 {
6578 int i, j, n;
6579 int new_topology;
6580
6581 mutex_lock(&sched_domains_mutex);
6582
6583 /* always unregister in case we don't destroy any domains */
6584 unregister_sched_domain_sysctl();
6585
6586 /* Let architecture update cpu core mappings. */
6587 new_topology = arch_update_cpu_topology();
6588
6589 n = doms_new ? ndoms_new : 0;
6590
6591 /* Destroy deleted domains */
6592 for (i = 0; i < ndoms_cur; i++) {
6593 for (j = 0; j < n && !new_topology; j++) {
6594 if (cpumask_equal(doms_cur[i], doms_new[j])
6595 && dattrs_equal(dattr_cur, i, dattr_new, j))
6596 goto match1;
6597 }
6598 /* no match - a current sched domain not in new doms_new[] */
6599 detach_destroy_domains(doms_cur[i]);
6600 match1:
6601 ;
6602 }
6603
6604 n = ndoms_cur;
6605 if (doms_new == NULL) {
6606 n = 0;
6607 doms_new = &fallback_doms;
6608 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6609 WARN_ON_ONCE(dattr_new);
6610 }
6611
6612 /* Build new domains */
6613 for (i = 0; i < ndoms_new; i++) {
6614 for (j = 0; j < n && !new_topology; j++) {
6615 if (cpumask_equal(doms_new[i], doms_cur[j])
6616 && dattrs_equal(dattr_new, i, dattr_cur, j))
6617 goto match2;
6618 }
6619 /* no match - add a new doms_new */
6620 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6621 match2:
6622 ;
6623 }
6624
6625 /* Remember the new sched domains */
6626 if (doms_cur != &fallback_doms)
6627 free_sched_domains(doms_cur, ndoms_cur);
6628 kfree(dattr_cur); /* kfree(NULL) is safe */
6629 doms_cur = doms_new;
6630 dattr_cur = dattr_new;
6631 ndoms_cur = ndoms_new;
6632
6633 register_sched_domain_sysctl();
6634
6635 mutex_unlock(&sched_domains_mutex);
6636 }
6637
6638 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6639
6640 /*
6641 * Update cpusets according to cpu_active mask. If cpusets are
6642 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6643 * around partition_sched_domains().
6644 *
6645 * If we come here as part of a suspend/resume, don't touch cpusets because we
6646 * want to restore it back to its original state upon resume anyway.
6647 */
6648 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6649 void *hcpu)
6650 {
6651 switch (action) {
6652 case CPU_ONLINE_FROZEN:
6653 case CPU_DOWN_FAILED_FROZEN:
6654
6655 /*
6656 * num_cpus_frozen tracks how many CPUs are involved in suspend
6657 * resume sequence. As long as this is not the last online
6658 * operation in the resume sequence, just build a single sched
6659 * domain, ignoring cpusets.
6660 */
6661 num_cpus_frozen--;
6662 if (likely(num_cpus_frozen)) {
6663 partition_sched_domains(1, NULL, NULL);
6664 break;
6665 }
6666
6667 /*
6668 * This is the last CPU online operation. So fall through and
6669 * restore the original sched domains by considering the
6670 * cpuset configurations.
6671 */
6672
6673 case CPU_ONLINE:
6674 case CPU_DOWN_FAILED:
6675 cpuset_update_active_cpus(true);
6676 break;
6677 default:
6678 return NOTIFY_DONE;
6679 }
6680 return NOTIFY_OK;
6681 }
6682
6683 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6684 void *hcpu)
6685 {
6686 switch (action) {
6687 case CPU_DOWN_PREPARE:
6688 cpuset_update_active_cpus(false);
6689 break;
6690 case CPU_DOWN_PREPARE_FROZEN:
6691 num_cpus_frozen++;
6692 partition_sched_domains(1, NULL, NULL);
6693 break;
6694 default:
6695 return NOTIFY_DONE;
6696 }
6697 return NOTIFY_OK;
6698 }
6699
6700 void __init sched_init_smp(void)
6701 {
6702 cpumask_var_t non_isolated_cpus;
6703
6704 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6705 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6706
6707 sched_init_numa();
6708
6709 /*
6710 * There's no userspace yet to cause hotplug operations; hence all the
6711 * cpu masks are stable and all blatant races in the below code cannot
6712 * happen.
6713 */
6714 mutex_lock(&sched_domains_mutex);
6715 init_sched_domains(cpu_active_mask);
6716 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6717 if (cpumask_empty(non_isolated_cpus))
6718 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6719 mutex_unlock(&sched_domains_mutex);
6720
6721 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6722 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6723 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6724
6725 init_hrtick();
6726
6727 /* Move init over to a non-isolated CPU */
6728 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6729 BUG();
6730 sched_init_granularity();
6731 free_cpumask_var(non_isolated_cpus);
6732
6733 init_sched_rt_class();
6734 init_sched_dl_class();
6735 }
6736 #else
6737 void __init sched_init_smp(void)
6738 {
6739 sched_init_granularity();
6740 }
6741 #endif /* CONFIG_SMP */
6742
6743 const_debug unsigned int sysctl_timer_migration = 1;
6744
6745 int in_sched_functions(unsigned long addr)
6746 {
6747 return in_lock_functions(addr) ||
6748 (addr >= (unsigned long)__sched_text_start
6749 && addr < (unsigned long)__sched_text_end);
6750 }
6751
6752 #ifdef CONFIG_CGROUP_SCHED
6753 /*
6754 * Default task group.
6755 * Every task in system belongs to this group at bootup.
6756 */
6757 struct task_group root_task_group;
6758 LIST_HEAD(task_groups);
6759 #endif
6760
6761 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6762
6763 void __init sched_init(void)
6764 {
6765 int i, j;
6766 unsigned long alloc_size = 0, ptr;
6767
6768 #ifdef CONFIG_FAIR_GROUP_SCHED
6769 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6770 #endif
6771 #ifdef CONFIG_RT_GROUP_SCHED
6772 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6773 #endif
6774 #ifdef CONFIG_CPUMASK_OFFSTACK
6775 alloc_size += num_possible_cpus() * cpumask_size();
6776 #endif
6777 if (alloc_size) {
6778 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6779
6780 #ifdef CONFIG_FAIR_GROUP_SCHED
6781 root_task_group.se = (struct sched_entity **)ptr;
6782 ptr += nr_cpu_ids * sizeof(void **);
6783
6784 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6785 ptr += nr_cpu_ids * sizeof(void **);
6786
6787 #endif /* CONFIG_FAIR_GROUP_SCHED */
6788 #ifdef CONFIG_RT_GROUP_SCHED
6789 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6790 ptr += nr_cpu_ids * sizeof(void **);
6791
6792 root_task_group.rt_rq = (struct rt_rq **)ptr;
6793 ptr += nr_cpu_ids * sizeof(void **);
6794
6795 #endif /* CONFIG_RT_GROUP_SCHED */
6796 #ifdef CONFIG_CPUMASK_OFFSTACK
6797 for_each_possible_cpu(i) {
6798 per_cpu(load_balance_mask, i) = (void *)ptr;
6799 ptr += cpumask_size();
6800 }
6801 #endif /* CONFIG_CPUMASK_OFFSTACK */
6802 }
6803
6804 init_rt_bandwidth(&def_rt_bandwidth,
6805 global_rt_period(), global_rt_runtime());
6806 init_dl_bandwidth(&def_dl_bandwidth,
6807 global_rt_period(), global_rt_runtime());
6808
6809 #ifdef CONFIG_SMP
6810 init_defrootdomain();
6811 #endif
6812
6813 #ifdef CONFIG_RT_GROUP_SCHED
6814 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6815 global_rt_period(), global_rt_runtime());
6816 #endif /* CONFIG_RT_GROUP_SCHED */
6817
6818 #ifdef CONFIG_CGROUP_SCHED
6819 list_add(&root_task_group.list, &task_groups);
6820 INIT_LIST_HEAD(&root_task_group.children);
6821 INIT_LIST_HEAD(&root_task_group.siblings);
6822 autogroup_init(&init_task);
6823
6824 #endif /* CONFIG_CGROUP_SCHED */
6825
6826 for_each_possible_cpu(i) {
6827 struct rq *rq;
6828
6829 rq = cpu_rq(i);
6830 raw_spin_lock_init(&rq->lock);
6831 rq->nr_running = 0;
6832 rq->calc_load_active = 0;
6833 rq->calc_load_update = jiffies + LOAD_FREQ;
6834 init_cfs_rq(&rq->cfs);
6835 init_rt_rq(&rq->rt, rq);
6836 init_dl_rq(&rq->dl, rq);
6837 #ifdef CONFIG_FAIR_GROUP_SCHED
6838 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6839 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6840 /*
6841 * How much cpu bandwidth does root_task_group get?
6842 *
6843 * In case of task-groups formed thr' the cgroup filesystem, it
6844 * gets 100% of the cpu resources in the system. This overall
6845 * system cpu resource is divided among the tasks of
6846 * root_task_group and its child task-groups in a fair manner,
6847 * based on each entity's (task or task-group's) weight
6848 * (se->load.weight).
6849 *
6850 * In other words, if root_task_group has 10 tasks of weight
6851 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6852 * then A0's share of the cpu resource is:
6853 *
6854 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6855 *
6856 * We achieve this by letting root_task_group's tasks sit
6857 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6858 */
6859 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6860 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6861 #endif /* CONFIG_FAIR_GROUP_SCHED */
6862
6863 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6864 #ifdef CONFIG_RT_GROUP_SCHED
6865 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6866 #endif
6867
6868 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6869 rq->cpu_load[j] = 0;
6870
6871 rq->last_load_update_tick = jiffies;
6872
6873 #ifdef CONFIG_SMP
6874 rq->sd = NULL;
6875 rq->rd = NULL;
6876 rq->cpu_power = SCHED_POWER_SCALE;
6877 rq->post_schedule = 0;
6878 rq->active_balance = 0;
6879 rq->next_balance = jiffies;
6880 rq->push_cpu = 0;
6881 rq->cpu = i;
6882 rq->online = 0;
6883 rq->idle_stamp = 0;
6884 rq->avg_idle = 2*sysctl_sched_migration_cost;
6885 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6886
6887 INIT_LIST_HEAD(&rq->cfs_tasks);
6888
6889 rq_attach_root(rq, &def_root_domain);
6890 #ifdef CONFIG_NO_HZ_COMMON
6891 rq->nohz_flags = 0;
6892 #endif
6893 #ifdef CONFIG_NO_HZ_FULL
6894 rq->last_sched_tick = 0;
6895 #endif
6896 #endif
6897 init_rq_hrtick(rq);
6898 atomic_set(&rq->nr_iowait, 0);
6899 }
6900
6901 set_load_weight(&init_task);
6902
6903 #ifdef CONFIG_PREEMPT_NOTIFIERS
6904 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6905 #endif
6906
6907 /*
6908 * The boot idle thread does lazy MMU switching as well:
6909 */
6910 atomic_inc(&init_mm.mm_count);
6911 enter_lazy_tlb(&init_mm, current);
6912
6913 /*
6914 * Make us the idle thread. Technically, schedule() should not be
6915 * called from this thread, however somewhere below it might be,
6916 * but because we are the idle thread, we just pick up running again
6917 * when this runqueue becomes "idle".
6918 */
6919 init_idle(current, smp_processor_id());
6920
6921 calc_load_update = jiffies + LOAD_FREQ;
6922
6923 /*
6924 * During early bootup we pretend to be a normal task:
6925 */
6926 current->sched_class = &fair_sched_class;
6927
6928 #ifdef CONFIG_SMP
6929 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6930 /* May be allocated at isolcpus cmdline parse time */
6931 if (cpu_isolated_map == NULL)
6932 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6933 idle_thread_set_boot_cpu();
6934 #endif
6935 init_sched_fair_class();
6936
6937 scheduler_running = 1;
6938 }
6939
6940 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6941 static inline int preempt_count_equals(int preempt_offset)
6942 {
6943 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6944
6945 return (nested == preempt_offset);
6946 }
6947
6948 void __might_sleep(const char *file, int line, int preempt_offset)
6949 {
6950 static unsigned long prev_jiffy; /* ratelimiting */
6951
6952 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6953 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6954 !is_idle_task(current)) ||
6955 system_state != SYSTEM_RUNNING || oops_in_progress)
6956 return;
6957 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6958 return;
6959 prev_jiffy = jiffies;
6960
6961 printk(KERN_ERR
6962 "BUG: sleeping function called from invalid context at %s:%d\n",
6963 file, line);
6964 printk(KERN_ERR
6965 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6966 in_atomic(), irqs_disabled(),
6967 current->pid, current->comm);
6968
6969 debug_show_held_locks(current);
6970 if (irqs_disabled())
6971 print_irqtrace_events(current);
6972 #ifdef CONFIG_DEBUG_PREEMPT
6973 if (!preempt_count_equals(preempt_offset)) {
6974 pr_err("Preemption disabled at:");
6975 print_ip_sym(current->preempt_disable_ip);
6976 pr_cont("\n");
6977 }
6978 #endif
6979 dump_stack();
6980 }
6981 EXPORT_SYMBOL(__might_sleep);
6982 #endif
6983
6984 #ifdef CONFIG_MAGIC_SYSRQ
6985 static void normalize_task(struct rq *rq, struct task_struct *p)
6986 {
6987 const struct sched_class *prev_class = p->sched_class;
6988 struct sched_attr attr = {
6989 .sched_policy = SCHED_NORMAL,
6990 };
6991 int old_prio = p->prio;
6992 int on_rq;
6993
6994 on_rq = p->on_rq;
6995 if (on_rq)
6996 dequeue_task(rq, p, 0);
6997 __setscheduler(rq, p, &attr);
6998 if (on_rq) {
6999 enqueue_task(rq, p, 0);
7000 resched_task(rq->curr);
7001 }
7002
7003 check_class_changed(rq, p, prev_class, old_prio);
7004 }
7005
7006 void normalize_rt_tasks(void)
7007 {
7008 struct task_struct *g, *p;
7009 unsigned long flags;
7010 struct rq *rq;
7011
7012 read_lock_irqsave(&tasklist_lock, flags);
7013 do_each_thread(g, p) {
7014 /*
7015 * Only normalize user tasks:
7016 */
7017 if (!p->mm)
7018 continue;
7019
7020 p->se.exec_start = 0;
7021 #ifdef CONFIG_SCHEDSTATS
7022 p->se.statistics.wait_start = 0;
7023 p->se.statistics.sleep_start = 0;
7024 p->se.statistics.block_start = 0;
7025 #endif
7026
7027 if (!dl_task(p) && !rt_task(p)) {
7028 /*
7029 * Renice negative nice level userspace
7030 * tasks back to 0:
7031 */
7032 if (task_nice(p) < 0 && p->mm)
7033 set_user_nice(p, 0);
7034 continue;
7035 }
7036
7037 raw_spin_lock(&p->pi_lock);
7038 rq = __task_rq_lock(p);
7039
7040 normalize_task(rq, p);
7041
7042 __task_rq_unlock(rq);
7043 raw_spin_unlock(&p->pi_lock);
7044 } while_each_thread(g, p);
7045
7046 read_unlock_irqrestore(&tasklist_lock, flags);
7047 }
7048
7049 #endif /* CONFIG_MAGIC_SYSRQ */
7050
7051 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7052 /*
7053 * These functions are only useful for the IA64 MCA handling, or kdb.
7054 *
7055 * They can only be called when the whole system has been
7056 * stopped - every CPU needs to be quiescent, and no scheduling
7057 * activity can take place. Using them for anything else would
7058 * be a serious bug, and as a result, they aren't even visible
7059 * under any other configuration.
7060 */
7061
7062 /**
7063 * curr_task - return the current task for a given cpu.
7064 * @cpu: the processor in question.
7065 *
7066 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7067 *
7068 * Return: The current task for @cpu.
7069 */
7070 struct task_struct *curr_task(int cpu)
7071 {
7072 return cpu_curr(cpu);
7073 }
7074
7075 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7076
7077 #ifdef CONFIG_IA64
7078 /**
7079 * set_curr_task - set the current task for a given cpu.
7080 * @cpu: the processor in question.
7081 * @p: the task pointer to set.
7082 *
7083 * Description: This function must only be used when non-maskable interrupts
7084 * are serviced on a separate stack. It allows the architecture to switch the
7085 * notion of the current task on a cpu in a non-blocking manner. This function
7086 * must be called with all CPU's synchronized, and interrupts disabled, the
7087 * and caller must save the original value of the current task (see
7088 * curr_task() above) and restore that value before reenabling interrupts and
7089 * re-starting the system.
7090 *
7091 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7092 */
7093 void set_curr_task(int cpu, struct task_struct *p)
7094 {
7095 cpu_curr(cpu) = p;
7096 }
7097
7098 #endif
7099
7100 #ifdef CONFIG_CGROUP_SCHED
7101 /* task_group_lock serializes the addition/removal of task groups */
7102 static DEFINE_SPINLOCK(task_group_lock);
7103
7104 static void free_sched_group(struct task_group *tg)
7105 {
7106 free_fair_sched_group(tg);
7107 free_rt_sched_group(tg);
7108 autogroup_free(tg);
7109 kfree(tg);
7110 }
7111
7112 /* allocate runqueue etc for a new task group */
7113 struct task_group *sched_create_group(struct task_group *parent)
7114 {
7115 struct task_group *tg;
7116
7117 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7118 if (!tg)
7119 return ERR_PTR(-ENOMEM);
7120
7121 if (!alloc_fair_sched_group(tg, parent))
7122 goto err;
7123
7124 if (!alloc_rt_sched_group(tg, parent))
7125 goto err;
7126
7127 return tg;
7128
7129 err:
7130 free_sched_group(tg);
7131 return ERR_PTR(-ENOMEM);
7132 }
7133
7134 void sched_online_group(struct task_group *tg, struct task_group *parent)
7135 {
7136 unsigned long flags;
7137
7138 spin_lock_irqsave(&task_group_lock, flags);
7139 list_add_rcu(&tg->list, &task_groups);
7140
7141 WARN_ON(!parent); /* root should already exist */
7142
7143 tg->parent = parent;
7144 INIT_LIST_HEAD(&tg->children);
7145 list_add_rcu(&tg->siblings, &parent->children);
7146 spin_unlock_irqrestore(&task_group_lock, flags);
7147 }
7148
7149 /* rcu callback to free various structures associated with a task group */
7150 static void free_sched_group_rcu(struct rcu_head *rhp)
7151 {
7152 /* now it should be safe to free those cfs_rqs */
7153 free_sched_group(container_of(rhp, struct task_group, rcu));
7154 }
7155
7156 /* Destroy runqueue etc associated with a task group */
7157 void sched_destroy_group(struct task_group *tg)
7158 {
7159 /* wait for possible concurrent references to cfs_rqs complete */
7160 call_rcu(&tg->rcu, free_sched_group_rcu);
7161 }
7162
7163 void sched_offline_group(struct task_group *tg)
7164 {
7165 unsigned long flags;
7166 int i;
7167
7168 /* end participation in shares distribution */
7169 for_each_possible_cpu(i)
7170 unregister_fair_sched_group(tg, i);
7171
7172 spin_lock_irqsave(&task_group_lock, flags);
7173 list_del_rcu(&tg->list);
7174 list_del_rcu(&tg->siblings);
7175 spin_unlock_irqrestore(&task_group_lock, flags);
7176 }
7177
7178 /* change task's runqueue when it moves between groups.
7179 * The caller of this function should have put the task in its new group
7180 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7181 * reflect its new group.
7182 */
7183 void sched_move_task(struct task_struct *tsk)
7184 {
7185 struct task_group *tg;
7186 int on_rq, running;
7187 unsigned long flags;
7188 struct rq *rq;
7189
7190 rq = task_rq_lock(tsk, &flags);
7191
7192 running = task_current(rq, tsk);
7193 on_rq = tsk->on_rq;
7194
7195 if (on_rq)
7196 dequeue_task(rq, tsk, 0);
7197 if (unlikely(running))
7198 tsk->sched_class->put_prev_task(rq, tsk);
7199
7200 tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7201 lockdep_is_held(&tsk->sighand->siglock)),
7202 struct task_group, css);
7203 tg = autogroup_task_group(tsk, tg);
7204 tsk->sched_task_group = tg;
7205
7206 #ifdef CONFIG_FAIR_GROUP_SCHED
7207 if (tsk->sched_class->task_move_group)
7208 tsk->sched_class->task_move_group(tsk, on_rq);
7209 else
7210 #endif
7211 set_task_rq(tsk, task_cpu(tsk));
7212
7213 if (unlikely(running))
7214 tsk->sched_class->set_curr_task(rq);
7215 if (on_rq)
7216 enqueue_task(rq, tsk, 0);
7217
7218 task_rq_unlock(rq, tsk, &flags);
7219 }
7220 #endif /* CONFIG_CGROUP_SCHED */
7221
7222 #ifdef CONFIG_RT_GROUP_SCHED
7223 /*
7224 * Ensure that the real time constraints are schedulable.
7225 */
7226 static DEFINE_MUTEX(rt_constraints_mutex);
7227
7228 /* Must be called with tasklist_lock held */
7229 static inline int tg_has_rt_tasks(struct task_group *tg)
7230 {
7231 struct task_struct *g, *p;
7232
7233 do_each_thread(g, p) {
7234 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7235 return 1;
7236 } while_each_thread(g, p);
7237
7238 return 0;
7239 }
7240
7241 struct rt_schedulable_data {
7242 struct task_group *tg;
7243 u64 rt_period;
7244 u64 rt_runtime;
7245 };
7246
7247 static int tg_rt_schedulable(struct task_group *tg, void *data)
7248 {
7249 struct rt_schedulable_data *d = data;
7250 struct task_group *child;
7251 unsigned long total, sum = 0;
7252 u64 period, runtime;
7253
7254 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7255 runtime = tg->rt_bandwidth.rt_runtime;
7256
7257 if (tg == d->tg) {
7258 period = d->rt_period;
7259 runtime = d->rt_runtime;
7260 }
7261
7262 /*
7263 * Cannot have more runtime than the period.
7264 */
7265 if (runtime > period && runtime != RUNTIME_INF)
7266 return -EINVAL;
7267
7268 /*
7269 * Ensure we don't starve existing RT tasks.
7270 */
7271 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7272 return -EBUSY;
7273
7274 total = to_ratio(period, runtime);
7275
7276 /*
7277 * Nobody can have more than the global setting allows.
7278 */
7279 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7280 return -EINVAL;
7281
7282 /*
7283 * The sum of our children's runtime should not exceed our own.
7284 */
7285 list_for_each_entry_rcu(child, &tg->children, siblings) {
7286 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7287 runtime = child->rt_bandwidth.rt_runtime;
7288
7289 if (child == d->tg) {
7290 period = d->rt_period;
7291 runtime = d->rt_runtime;
7292 }
7293
7294 sum += to_ratio(period, runtime);
7295 }
7296
7297 if (sum > total)
7298 return -EINVAL;
7299
7300 return 0;
7301 }
7302
7303 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7304 {
7305 int ret;
7306
7307 struct rt_schedulable_data data = {
7308 .tg = tg,
7309 .rt_period = period,
7310 .rt_runtime = runtime,
7311 };
7312
7313 rcu_read_lock();
7314 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7315 rcu_read_unlock();
7316
7317 return ret;
7318 }
7319
7320 static int tg_set_rt_bandwidth(struct task_group *tg,
7321 u64 rt_period, u64 rt_runtime)
7322 {
7323 int i, err = 0;
7324
7325 mutex_lock(&rt_constraints_mutex);
7326 read_lock(&tasklist_lock);
7327 err = __rt_schedulable(tg, rt_period, rt_runtime);
7328 if (err)
7329 goto unlock;
7330
7331 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7332 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7333 tg->rt_bandwidth.rt_runtime = rt_runtime;
7334
7335 for_each_possible_cpu(i) {
7336 struct rt_rq *rt_rq = tg->rt_rq[i];
7337
7338 raw_spin_lock(&rt_rq->rt_runtime_lock);
7339 rt_rq->rt_runtime = rt_runtime;
7340 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7341 }
7342 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7343 unlock:
7344 read_unlock(&tasklist_lock);
7345 mutex_unlock(&rt_constraints_mutex);
7346
7347 return err;
7348 }
7349
7350 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7351 {
7352 u64 rt_runtime, rt_period;
7353
7354 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7355 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7356 if (rt_runtime_us < 0)
7357 rt_runtime = RUNTIME_INF;
7358
7359 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7360 }
7361
7362 static long sched_group_rt_runtime(struct task_group *tg)
7363 {
7364 u64 rt_runtime_us;
7365
7366 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7367 return -1;
7368
7369 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7370 do_div(rt_runtime_us, NSEC_PER_USEC);
7371 return rt_runtime_us;
7372 }
7373
7374 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7375 {
7376 u64 rt_runtime, rt_period;
7377
7378 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7379 rt_runtime = tg->rt_bandwidth.rt_runtime;
7380
7381 if (rt_period == 0)
7382 return -EINVAL;
7383
7384 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7385 }
7386
7387 static long sched_group_rt_period(struct task_group *tg)
7388 {
7389 u64 rt_period_us;
7390
7391 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7392 do_div(rt_period_us, NSEC_PER_USEC);
7393 return rt_period_us;
7394 }
7395 #endif /* CONFIG_RT_GROUP_SCHED */
7396
7397 #ifdef CONFIG_RT_GROUP_SCHED
7398 static int sched_rt_global_constraints(void)
7399 {
7400 int ret = 0;
7401
7402 mutex_lock(&rt_constraints_mutex);
7403 read_lock(&tasklist_lock);
7404 ret = __rt_schedulable(NULL, 0, 0);
7405 read_unlock(&tasklist_lock);
7406 mutex_unlock(&rt_constraints_mutex);
7407
7408 return ret;
7409 }
7410
7411 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7412 {
7413 /* Don't accept realtime tasks when there is no way for them to run */
7414 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7415 return 0;
7416
7417 return 1;
7418 }
7419
7420 #else /* !CONFIG_RT_GROUP_SCHED */
7421 static int sched_rt_global_constraints(void)
7422 {
7423 unsigned long flags;
7424 int i, ret = 0;
7425
7426 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7427 for_each_possible_cpu(i) {
7428 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7429
7430 raw_spin_lock(&rt_rq->rt_runtime_lock);
7431 rt_rq->rt_runtime = global_rt_runtime();
7432 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7433 }
7434 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7435
7436 return ret;
7437 }
7438 #endif /* CONFIG_RT_GROUP_SCHED */
7439
7440 static int sched_dl_global_constraints(void)
7441 {
7442 u64 runtime = global_rt_runtime();
7443 u64 period = global_rt_period();
7444 u64 new_bw = to_ratio(period, runtime);
7445 int cpu, ret = 0;
7446 unsigned long flags;
7447
7448 /*
7449 * Here we want to check the bandwidth not being set to some
7450 * value smaller than the currently allocated bandwidth in
7451 * any of the root_domains.
7452 *
7453 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7454 * cycling on root_domains... Discussion on different/better
7455 * solutions is welcome!
7456 */
7457 for_each_possible_cpu(cpu) {
7458 struct dl_bw *dl_b = dl_bw_of(cpu);
7459
7460 raw_spin_lock_irqsave(&dl_b->lock, flags);
7461 if (new_bw < dl_b->total_bw)
7462 ret = -EBUSY;
7463 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7464
7465 if (ret)
7466 break;
7467 }
7468
7469 return ret;
7470 }
7471
7472 static void sched_dl_do_global(void)
7473 {
7474 u64 new_bw = -1;
7475 int cpu;
7476 unsigned long flags;
7477
7478 def_dl_bandwidth.dl_period = global_rt_period();
7479 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7480
7481 if (global_rt_runtime() != RUNTIME_INF)
7482 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7483
7484 /*
7485 * FIXME: As above...
7486 */
7487 for_each_possible_cpu(cpu) {
7488 struct dl_bw *dl_b = dl_bw_of(cpu);
7489
7490 raw_spin_lock_irqsave(&dl_b->lock, flags);
7491 dl_b->bw = new_bw;
7492 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7493 }
7494 }
7495
7496 static int sched_rt_global_validate(void)
7497 {
7498 if (sysctl_sched_rt_period <= 0)
7499 return -EINVAL;
7500
7501 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7502 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7503 return -EINVAL;
7504
7505 return 0;
7506 }
7507
7508 static void sched_rt_do_global(void)
7509 {
7510 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7511 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7512 }
7513
7514 int sched_rt_handler(struct ctl_table *table, int write,
7515 void __user *buffer, size_t *lenp,
7516 loff_t *ppos)
7517 {
7518 int old_period, old_runtime;
7519 static DEFINE_MUTEX(mutex);
7520 int ret;
7521
7522 mutex_lock(&mutex);
7523 old_period = sysctl_sched_rt_period;
7524 old_runtime = sysctl_sched_rt_runtime;
7525
7526 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7527
7528 if (!ret && write) {
7529 ret = sched_rt_global_validate();
7530 if (ret)
7531 goto undo;
7532
7533 ret = sched_rt_global_constraints();
7534 if (ret)
7535 goto undo;
7536
7537 ret = sched_dl_global_constraints();
7538 if (ret)
7539 goto undo;
7540
7541 sched_rt_do_global();
7542 sched_dl_do_global();
7543 }
7544 if (0) {
7545 undo:
7546 sysctl_sched_rt_period = old_period;
7547 sysctl_sched_rt_runtime = old_runtime;
7548 }
7549 mutex_unlock(&mutex);
7550
7551 return ret;
7552 }
7553
7554 int sched_rr_handler(struct ctl_table *table, int write,
7555 void __user *buffer, size_t *lenp,
7556 loff_t *ppos)
7557 {
7558 int ret;
7559 static DEFINE_MUTEX(mutex);
7560
7561 mutex_lock(&mutex);
7562 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7563 /* make sure that internally we keep jiffies */
7564 /* also, writing zero resets timeslice to default */
7565 if (!ret && write) {
7566 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7567 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7568 }
7569 mutex_unlock(&mutex);
7570 return ret;
7571 }
7572
7573 #ifdef CONFIG_CGROUP_SCHED
7574
7575 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7576 {
7577 return css ? container_of(css, struct task_group, css) : NULL;
7578 }
7579
7580 static struct cgroup_subsys_state *
7581 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7582 {
7583 struct task_group *parent = css_tg(parent_css);
7584 struct task_group *tg;
7585
7586 if (!parent) {
7587 /* This is early initialization for the top cgroup */
7588 return &root_task_group.css;
7589 }
7590
7591 tg = sched_create_group(parent);
7592 if (IS_ERR(tg))
7593 return ERR_PTR(-ENOMEM);
7594
7595 return &tg->css;
7596 }
7597
7598 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7599 {
7600 struct task_group *tg = css_tg(css);
7601 struct task_group *parent = css_tg(css_parent(css));
7602
7603 if (parent)
7604 sched_online_group(tg, parent);
7605 return 0;
7606 }
7607
7608 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7609 {
7610 struct task_group *tg = css_tg(css);
7611
7612 sched_destroy_group(tg);
7613 }
7614
7615 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7616 {
7617 struct task_group *tg = css_tg(css);
7618
7619 sched_offline_group(tg);
7620 }
7621
7622 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7623 struct cgroup_taskset *tset)
7624 {
7625 struct task_struct *task;
7626
7627 cgroup_taskset_for_each(task, tset) {
7628 #ifdef CONFIG_RT_GROUP_SCHED
7629 if (!sched_rt_can_attach(css_tg(css), task))
7630 return -EINVAL;
7631 #else
7632 /* We don't support RT-tasks being in separate groups */
7633 if (task->sched_class != &fair_sched_class)
7634 return -EINVAL;
7635 #endif
7636 }
7637 return 0;
7638 }
7639
7640 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7641 struct cgroup_taskset *tset)
7642 {
7643 struct task_struct *task;
7644
7645 cgroup_taskset_for_each(task, tset)
7646 sched_move_task(task);
7647 }
7648
7649 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7650 struct cgroup_subsys_state *old_css,
7651 struct task_struct *task)
7652 {
7653 /*
7654 * cgroup_exit() is called in the copy_process() failure path.
7655 * Ignore this case since the task hasn't ran yet, this avoids
7656 * trying to poke a half freed task state from generic code.
7657 */
7658 if (!(task->flags & PF_EXITING))
7659 return;
7660
7661 sched_move_task(task);
7662 }
7663
7664 #ifdef CONFIG_FAIR_GROUP_SCHED
7665 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7666 struct cftype *cftype, u64 shareval)
7667 {
7668 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7669 }
7670
7671 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7672 struct cftype *cft)
7673 {
7674 struct task_group *tg = css_tg(css);
7675
7676 return (u64) scale_load_down(tg->shares);
7677 }
7678
7679 #ifdef CONFIG_CFS_BANDWIDTH
7680 static DEFINE_MUTEX(cfs_constraints_mutex);
7681
7682 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7683 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7684
7685 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7686
7687 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7688 {
7689 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7690 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7691
7692 if (tg == &root_task_group)
7693 return -EINVAL;
7694
7695 /*
7696 * Ensure we have at some amount of bandwidth every period. This is
7697 * to prevent reaching a state of large arrears when throttled via
7698 * entity_tick() resulting in prolonged exit starvation.
7699 */
7700 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7701 return -EINVAL;
7702
7703 /*
7704 * Likewise, bound things on the otherside by preventing insane quota
7705 * periods. This also allows us to normalize in computing quota
7706 * feasibility.
7707 */
7708 if (period > max_cfs_quota_period)
7709 return -EINVAL;
7710
7711 mutex_lock(&cfs_constraints_mutex);
7712 ret = __cfs_schedulable(tg, period, quota);
7713 if (ret)
7714 goto out_unlock;
7715
7716 runtime_enabled = quota != RUNTIME_INF;
7717 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7718 /*
7719 * If we need to toggle cfs_bandwidth_used, off->on must occur
7720 * before making related changes, and on->off must occur afterwards
7721 */
7722 if (runtime_enabled && !runtime_was_enabled)
7723 cfs_bandwidth_usage_inc();
7724 raw_spin_lock_irq(&cfs_b->lock);
7725 cfs_b->period = ns_to_ktime(period);
7726 cfs_b->quota = quota;
7727
7728 __refill_cfs_bandwidth_runtime(cfs_b);
7729 /* restart the period timer (if active) to handle new period expiry */
7730 if (runtime_enabled && cfs_b->timer_active) {
7731 /* force a reprogram */
7732 cfs_b->timer_active = 0;
7733 __start_cfs_bandwidth(cfs_b);
7734 }
7735 raw_spin_unlock_irq(&cfs_b->lock);
7736
7737 for_each_possible_cpu(i) {
7738 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7739 struct rq *rq = cfs_rq->rq;
7740
7741 raw_spin_lock_irq(&rq->lock);
7742 cfs_rq->runtime_enabled = runtime_enabled;
7743 cfs_rq->runtime_remaining = 0;
7744
7745 if (cfs_rq->throttled)
7746 unthrottle_cfs_rq(cfs_rq);
7747 raw_spin_unlock_irq(&rq->lock);
7748 }
7749 if (runtime_was_enabled && !runtime_enabled)
7750 cfs_bandwidth_usage_dec();
7751 out_unlock:
7752 mutex_unlock(&cfs_constraints_mutex);
7753
7754 return ret;
7755 }
7756
7757 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7758 {
7759 u64 quota, period;
7760
7761 period = ktime_to_ns(tg->cfs_bandwidth.period);
7762 if (cfs_quota_us < 0)
7763 quota = RUNTIME_INF;
7764 else
7765 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7766
7767 return tg_set_cfs_bandwidth(tg, period, quota);
7768 }
7769
7770 long tg_get_cfs_quota(struct task_group *tg)
7771 {
7772 u64 quota_us;
7773
7774 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7775 return -1;
7776
7777 quota_us = tg->cfs_bandwidth.quota;
7778 do_div(quota_us, NSEC_PER_USEC);
7779
7780 return quota_us;
7781 }
7782
7783 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7784 {
7785 u64 quota, period;
7786
7787 period = (u64)cfs_period_us * NSEC_PER_USEC;
7788 quota = tg->cfs_bandwidth.quota;
7789
7790 return tg_set_cfs_bandwidth(tg, period, quota);
7791 }
7792
7793 long tg_get_cfs_period(struct task_group *tg)
7794 {
7795 u64 cfs_period_us;
7796
7797 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7798 do_div(cfs_period_us, NSEC_PER_USEC);
7799
7800 return cfs_period_us;
7801 }
7802
7803 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7804 struct cftype *cft)
7805 {
7806 return tg_get_cfs_quota(css_tg(css));
7807 }
7808
7809 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7810 struct cftype *cftype, s64 cfs_quota_us)
7811 {
7812 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7813 }
7814
7815 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7816 struct cftype *cft)
7817 {
7818 return tg_get_cfs_period(css_tg(css));
7819 }
7820
7821 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7822 struct cftype *cftype, u64 cfs_period_us)
7823 {
7824 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7825 }
7826
7827 struct cfs_schedulable_data {
7828 struct task_group *tg;
7829 u64 period, quota;
7830 };
7831
7832 /*
7833 * normalize group quota/period to be quota/max_period
7834 * note: units are usecs
7835 */
7836 static u64 normalize_cfs_quota(struct task_group *tg,
7837 struct cfs_schedulable_data *d)
7838 {
7839 u64 quota, period;
7840
7841 if (tg == d->tg) {
7842 period = d->period;
7843 quota = d->quota;
7844 } else {
7845 period = tg_get_cfs_period(tg);
7846 quota = tg_get_cfs_quota(tg);
7847 }
7848
7849 /* note: these should typically be equivalent */
7850 if (quota == RUNTIME_INF || quota == -1)
7851 return RUNTIME_INF;
7852
7853 return to_ratio(period, quota);
7854 }
7855
7856 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7857 {
7858 struct cfs_schedulable_data *d = data;
7859 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7860 s64 quota = 0, parent_quota = -1;
7861
7862 if (!tg->parent) {
7863 quota = RUNTIME_INF;
7864 } else {
7865 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7866
7867 quota = normalize_cfs_quota(tg, d);
7868 parent_quota = parent_b->hierarchal_quota;
7869
7870 /*
7871 * ensure max(child_quota) <= parent_quota, inherit when no
7872 * limit is set
7873 */
7874 if (quota == RUNTIME_INF)
7875 quota = parent_quota;
7876 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7877 return -EINVAL;
7878 }
7879 cfs_b->hierarchal_quota = quota;
7880
7881 return 0;
7882 }
7883
7884 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7885 {
7886 int ret;
7887 struct cfs_schedulable_data data = {
7888 .tg = tg,
7889 .period = period,
7890 .quota = quota,
7891 };
7892
7893 if (quota != RUNTIME_INF) {
7894 do_div(data.period, NSEC_PER_USEC);
7895 do_div(data.quota, NSEC_PER_USEC);
7896 }
7897
7898 rcu_read_lock();
7899 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7900 rcu_read_unlock();
7901
7902 return ret;
7903 }
7904
7905 static int cpu_stats_show(struct seq_file *sf, void *v)
7906 {
7907 struct task_group *tg = css_tg(seq_css(sf));
7908 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7909
7910 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7911 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7912 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7913
7914 return 0;
7915 }
7916 #endif /* CONFIG_CFS_BANDWIDTH */
7917 #endif /* CONFIG_FAIR_GROUP_SCHED */
7918
7919 #ifdef CONFIG_RT_GROUP_SCHED
7920 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7921 struct cftype *cft, s64 val)
7922 {
7923 return sched_group_set_rt_runtime(css_tg(css), val);
7924 }
7925
7926 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7927 struct cftype *cft)
7928 {
7929 return sched_group_rt_runtime(css_tg(css));
7930 }
7931
7932 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7933 struct cftype *cftype, u64 rt_period_us)
7934 {
7935 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7936 }
7937
7938 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7939 struct cftype *cft)
7940 {
7941 return sched_group_rt_period(css_tg(css));
7942 }
7943 #endif /* CONFIG_RT_GROUP_SCHED */
7944
7945 static struct cftype cpu_files[] = {
7946 #ifdef CONFIG_FAIR_GROUP_SCHED
7947 {
7948 .name = "shares",
7949 .read_u64 = cpu_shares_read_u64,
7950 .write_u64 = cpu_shares_write_u64,
7951 },
7952 #endif
7953 #ifdef CONFIG_CFS_BANDWIDTH
7954 {
7955 .name = "cfs_quota_us",
7956 .read_s64 = cpu_cfs_quota_read_s64,
7957 .write_s64 = cpu_cfs_quota_write_s64,
7958 },
7959 {
7960 .name = "cfs_period_us",
7961 .read_u64 = cpu_cfs_period_read_u64,
7962 .write_u64 = cpu_cfs_period_write_u64,
7963 },
7964 {
7965 .name = "stat",
7966 .seq_show = cpu_stats_show,
7967 },
7968 #endif
7969 #ifdef CONFIG_RT_GROUP_SCHED
7970 {
7971 .name = "rt_runtime_us",
7972 .read_s64 = cpu_rt_runtime_read,
7973 .write_s64 = cpu_rt_runtime_write,
7974 },
7975 {
7976 .name = "rt_period_us",
7977 .read_u64 = cpu_rt_period_read_uint,
7978 .write_u64 = cpu_rt_period_write_uint,
7979 },
7980 #endif
7981 { } /* terminate */
7982 };
7983
7984 struct cgroup_subsys cpu_cgrp_subsys = {
7985 .css_alloc = cpu_cgroup_css_alloc,
7986 .css_free = cpu_cgroup_css_free,
7987 .css_online = cpu_cgroup_css_online,
7988 .css_offline = cpu_cgroup_css_offline,
7989 .can_attach = cpu_cgroup_can_attach,
7990 .attach = cpu_cgroup_attach,
7991 .exit = cpu_cgroup_exit,
7992 .base_cftypes = cpu_files,
7993 .early_init = 1,
7994 };
7995
7996 #endif /* CONFIG_CGROUP_SCHED */
7997
7998 void dump_cpu_task(int cpu)
7999 {
8000 pr_info("Task dump for CPU %d:\n", cpu);
8001 sched_show_task(cpu_curr(cpu));
8002 }