Loading...
1/*
2 * kernel/sched/core.c
3 *
4 * Kernel scheduler and related syscalls
5 *
6 * Copyright (C) 1991-2002 Linus Torvalds
7 *
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
11 * by Andrea Arcangeli
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
22 * by Peter Williams
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
27 */
28
29#include <linux/mm.h>
30#include <linux/module.h>
31#include <linux/nmi.h>
32#include <linux/init.h>
33#include <linux/uaccess.h>
34#include <linux/highmem.h>
35#include <asm/mmu_context.h>
36#include <linux/interrupt.h>
37#include <linux/capability.h>
38#include <linux/completion.h>
39#include <linux/kernel_stat.h>
40#include <linux/debug_locks.h>
41#include <linux/perf_event.h>
42#include <linux/security.h>
43#include <linux/notifier.h>
44#include <linux/profile.h>
45#include <linux/freezer.h>
46#include <linux/vmalloc.h>
47#include <linux/blkdev.h>
48#include <linux/delay.h>
49#include <linux/pid_namespace.h>
50#include <linux/smp.h>
51#include <linux/threads.h>
52#include <linux/timer.h>
53#include <linux/rcupdate.h>
54#include <linux/cpu.h>
55#include <linux/cpuset.h>
56#include <linux/percpu.h>
57#include <linux/proc_fs.h>
58#include <linux/seq_file.h>
59#include <linux/sysctl.h>
60#include <linux/syscalls.h>
61#include <linux/times.h>
62#include <linux/tsacct_kern.h>
63#include <linux/kprobes.h>
64#include <linux/delayacct.h>
65#include <linux/unistd.h>
66#include <linux/pagemap.h>
67#include <linux/hrtimer.h>
68#include <linux/tick.h>
69#include <linux/debugfs.h>
70#include <linux/ctype.h>
71#include <linux/ftrace.h>
72#include <linux/slab.h>
73#include <linux/init_task.h>
74#include <linux/binfmts.h>
75
76#include <asm/switch_to.h>
77#include <asm/tlb.h>
78#include <asm/irq_regs.h>
79#include <asm/mutex.h>
80#ifdef CONFIG_PARAVIRT
81#include <asm/paravirt.h>
82#endif
83
84#include "sched.h"
85#include "../workqueue_sched.h"
86#include "../smpboot.h"
87
88#define CREATE_TRACE_POINTS
89#include <trace/events/sched.h>
90
91void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
92{
93 unsigned long delta;
94 ktime_t soft, hard, now;
95
96 for (;;) {
97 if (hrtimer_active(period_timer))
98 break;
99
100 now = hrtimer_cb_get_time(period_timer);
101 hrtimer_forward(period_timer, now, period);
102
103 soft = hrtimer_get_softexpires(period_timer);
104 hard = hrtimer_get_expires(period_timer);
105 delta = ktime_to_ns(ktime_sub(hard, soft));
106 __hrtimer_start_range_ns(period_timer, soft, delta,
107 HRTIMER_MODE_ABS_PINNED, 0);
108 }
109}
110
111DEFINE_MUTEX(sched_domains_mutex);
112DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
113
114static void update_rq_clock_task(struct rq *rq, s64 delta);
115
116void update_rq_clock(struct rq *rq)
117{
118 s64 delta;
119
120 if (rq->skip_clock_update > 0)
121 return;
122
123 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
124 rq->clock += delta;
125 update_rq_clock_task(rq, delta);
126}
127
128/*
129 * Debugging: various feature bits
130 */
131
132#define SCHED_FEAT(name, enabled) \
133 (1UL << __SCHED_FEAT_##name) * enabled |
134
135const_debug unsigned int sysctl_sched_features =
136#include "features.h"
137 0;
138
139#undef SCHED_FEAT
140
141#ifdef CONFIG_SCHED_DEBUG
142#define SCHED_FEAT(name, enabled) \
143 #name ,
144
145static const char * const sched_feat_names[] = {
146#include "features.h"
147};
148
149#undef SCHED_FEAT
150
151static int sched_feat_show(struct seq_file *m, void *v)
152{
153 int i;
154
155 for (i = 0; i < __SCHED_FEAT_NR; i++) {
156 if (!(sysctl_sched_features & (1UL << i)))
157 seq_puts(m, "NO_");
158 seq_printf(m, "%s ", sched_feat_names[i]);
159 }
160 seq_puts(m, "\n");
161
162 return 0;
163}
164
165#ifdef HAVE_JUMP_LABEL
166
167#define jump_label_key__true STATIC_KEY_INIT_TRUE
168#define jump_label_key__false STATIC_KEY_INIT_FALSE
169
170#define SCHED_FEAT(name, enabled) \
171 jump_label_key__##enabled ,
172
173struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
174#include "features.h"
175};
176
177#undef SCHED_FEAT
178
179static void sched_feat_disable(int i)
180{
181 if (static_key_enabled(&sched_feat_keys[i]))
182 static_key_slow_dec(&sched_feat_keys[i]);
183}
184
185static void sched_feat_enable(int i)
186{
187 if (!static_key_enabled(&sched_feat_keys[i]))
188 static_key_slow_inc(&sched_feat_keys[i]);
189}
190#else
191static void sched_feat_disable(int i) { };
192static void sched_feat_enable(int i) { };
193#endif /* HAVE_JUMP_LABEL */
194
195static ssize_t
196sched_feat_write(struct file *filp, const char __user *ubuf,
197 size_t cnt, loff_t *ppos)
198{
199 char buf[64];
200 char *cmp;
201 int neg = 0;
202 int i;
203
204 if (cnt > 63)
205 cnt = 63;
206
207 if (copy_from_user(&buf, ubuf, cnt))
208 return -EFAULT;
209
210 buf[cnt] = 0;
211 cmp = strstrip(buf);
212
213 if (strncmp(cmp, "NO_", 3) == 0) {
214 neg = 1;
215 cmp += 3;
216 }
217
218 for (i = 0; i < __SCHED_FEAT_NR; i++) {
219 if (strcmp(cmp, sched_feat_names[i]) == 0) {
220 if (neg) {
221 sysctl_sched_features &= ~(1UL << i);
222 sched_feat_disable(i);
223 } else {
224 sysctl_sched_features |= (1UL << i);
225 sched_feat_enable(i);
226 }
227 break;
228 }
229 }
230
231 if (i == __SCHED_FEAT_NR)
232 return -EINVAL;
233
234 *ppos += cnt;
235
236 return cnt;
237}
238
239static int sched_feat_open(struct inode *inode, struct file *filp)
240{
241 return single_open(filp, sched_feat_show, NULL);
242}
243
244static const struct file_operations sched_feat_fops = {
245 .open = sched_feat_open,
246 .write = sched_feat_write,
247 .read = seq_read,
248 .llseek = seq_lseek,
249 .release = single_release,
250};
251
252static __init int sched_init_debug(void)
253{
254 debugfs_create_file("sched_features", 0644, NULL, NULL,
255 &sched_feat_fops);
256
257 return 0;
258}
259late_initcall(sched_init_debug);
260#endif /* CONFIG_SCHED_DEBUG */
261
262/*
263 * Number of tasks to iterate in a single balance run.
264 * Limited because this is done with IRQs disabled.
265 */
266const_debug unsigned int sysctl_sched_nr_migrate = 32;
267
268/*
269 * period over which we average the RT time consumption, measured
270 * in ms.
271 *
272 * default: 1s
273 */
274const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
275
276/*
277 * period over which we measure -rt task cpu usage in us.
278 * default: 1s
279 */
280unsigned int sysctl_sched_rt_period = 1000000;
281
282__read_mostly int scheduler_running;
283
284/*
285 * part of the period that we allow rt tasks to run in us.
286 * default: 0.95s
287 */
288int sysctl_sched_rt_runtime = 950000;
289
290
291
292/*
293 * __task_rq_lock - lock the rq @p resides on.
294 */
295static inline struct rq *__task_rq_lock(struct task_struct *p)
296 __acquires(rq->lock)
297{
298 struct rq *rq;
299
300 lockdep_assert_held(&p->pi_lock);
301
302 for (;;) {
303 rq = task_rq(p);
304 raw_spin_lock(&rq->lock);
305 if (likely(rq == task_rq(p)))
306 return rq;
307 raw_spin_unlock(&rq->lock);
308 }
309}
310
311/*
312 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
313 */
314static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
315 __acquires(p->pi_lock)
316 __acquires(rq->lock)
317{
318 struct rq *rq;
319
320 for (;;) {
321 raw_spin_lock_irqsave(&p->pi_lock, *flags);
322 rq = task_rq(p);
323 raw_spin_lock(&rq->lock);
324 if (likely(rq == task_rq(p)))
325 return rq;
326 raw_spin_unlock(&rq->lock);
327 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
328 }
329}
330
331static void __task_rq_unlock(struct rq *rq)
332 __releases(rq->lock)
333{
334 raw_spin_unlock(&rq->lock);
335}
336
337static inline void
338task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
339 __releases(rq->lock)
340 __releases(p->pi_lock)
341{
342 raw_spin_unlock(&rq->lock);
343 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
344}
345
346/*
347 * this_rq_lock - lock this runqueue and disable interrupts.
348 */
349static struct rq *this_rq_lock(void)
350 __acquires(rq->lock)
351{
352 struct rq *rq;
353
354 local_irq_disable();
355 rq = this_rq();
356 raw_spin_lock(&rq->lock);
357
358 return rq;
359}
360
361#ifdef CONFIG_SCHED_HRTICK
362/*
363 * Use HR-timers to deliver accurate preemption points.
364 *
365 * Its all a bit involved since we cannot program an hrt while holding the
366 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
367 * reschedule event.
368 *
369 * When we get rescheduled we reprogram the hrtick_timer outside of the
370 * rq->lock.
371 */
372
373static void hrtick_clear(struct rq *rq)
374{
375 if (hrtimer_active(&rq->hrtick_timer))
376 hrtimer_cancel(&rq->hrtick_timer);
377}
378
379/*
380 * High-resolution timer tick.
381 * Runs from hardirq context with interrupts disabled.
382 */
383static enum hrtimer_restart hrtick(struct hrtimer *timer)
384{
385 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
386
387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
388
389 raw_spin_lock(&rq->lock);
390 update_rq_clock(rq);
391 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
392 raw_spin_unlock(&rq->lock);
393
394 return HRTIMER_NORESTART;
395}
396
397#ifdef CONFIG_SMP
398/*
399 * called from hardirq (IPI) context
400 */
401static void __hrtick_start(void *arg)
402{
403 struct rq *rq = arg;
404
405 raw_spin_lock(&rq->lock);
406 hrtimer_restart(&rq->hrtick_timer);
407 rq->hrtick_csd_pending = 0;
408 raw_spin_unlock(&rq->lock);
409}
410
411/*
412 * Called to set the hrtick timer state.
413 *
414 * called with rq->lock held and irqs disabled
415 */
416void hrtick_start(struct rq *rq, u64 delay)
417{
418 struct hrtimer *timer = &rq->hrtick_timer;
419 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
420
421 hrtimer_set_expires(timer, time);
422
423 if (rq == this_rq()) {
424 hrtimer_restart(timer);
425 } else if (!rq->hrtick_csd_pending) {
426 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
427 rq->hrtick_csd_pending = 1;
428 }
429}
430
431static int
432hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
433{
434 int cpu = (int)(long)hcpu;
435
436 switch (action) {
437 case CPU_UP_CANCELED:
438 case CPU_UP_CANCELED_FROZEN:
439 case CPU_DOWN_PREPARE:
440 case CPU_DOWN_PREPARE_FROZEN:
441 case CPU_DEAD:
442 case CPU_DEAD_FROZEN:
443 hrtick_clear(cpu_rq(cpu));
444 return NOTIFY_OK;
445 }
446
447 return NOTIFY_DONE;
448}
449
450static __init void init_hrtick(void)
451{
452 hotcpu_notifier(hotplug_hrtick, 0);
453}
454#else
455/*
456 * Called to set the hrtick timer state.
457 *
458 * called with rq->lock held and irqs disabled
459 */
460void hrtick_start(struct rq *rq, u64 delay)
461{
462 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
463 HRTIMER_MODE_REL_PINNED, 0);
464}
465
466static inline void init_hrtick(void)
467{
468}
469#endif /* CONFIG_SMP */
470
471static void init_rq_hrtick(struct rq *rq)
472{
473#ifdef CONFIG_SMP
474 rq->hrtick_csd_pending = 0;
475
476 rq->hrtick_csd.flags = 0;
477 rq->hrtick_csd.func = __hrtick_start;
478 rq->hrtick_csd.info = rq;
479#endif
480
481 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
482 rq->hrtick_timer.function = hrtick;
483}
484#else /* CONFIG_SCHED_HRTICK */
485static inline void hrtick_clear(struct rq *rq)
486{
487}
488
489static inline void init_rq_hrtick(struct rq *rq)
490{
491}
492
493static inline void init_hrtick(void)
494{
495}
496#endif /* CONFIG_SCHED_HRTICK */
497
498/*
499 * resched_task - mark a task 'to be rescheduled now'.
500 *
501 * On UP this means the setting of the need_resched flag, on SMP it
502 * might also involve a cross-CPU call to trigger the scheduler on
503 * the target CPU.
504 */
505#ifdef CONFIG_SMP
506
507#ifndef tsk_is_polling
508#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
509#endif
510
511void resched_task(struct task_struct *p)
512{
513 int cpu;
514
515 assert_raw_spin_locked(&task_rq(p)->lock);
516
517 if (test_tsk_need_resched(p))
518 return;
519
520 set_tsk_need_resched(p);
521
522 cpu = task_cpu(p);
523 if (cpu == smp_processor_id())
524 return;
525
526 /* NEED_RESCHED must be visible before we test polling */
527 smp_mb();
528 if (!tsk_is_polling(p))
529 smp_send_reschedule(cpu);
530}
531
532void resched_cpu(int cpu)
533{
534 struct rq *rq = cpu_rq(cpu);
535 unsigned long flags;
536
537 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
538 return;
539 resched_task(cpu_curr(cpu));
540 raw_spin_unlock_irqrestore(&rq->lock, flags);
541}
542
543#ifdef CONFIG_NO_HZ
544/*
545 * In the semi idle case, use the nearest busy cpu for migrating timers
546 * from an idle cpu. This is good for power-savings.
547 *
548 * We don't do similar optimization for completely idle system, as
549 * selecting an idle cpu will add more delays to the timers than intended
550 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
551 */
552int get_nohz_timer_target(void)
553{
554 int cpu = smp_processor_id();
555 int i;
556 struct sched_domain *sd;
557
558 rcu_read_lock();
559 for_each_domain(cpu, sd) {
560 for_each_cpu(i, sched_domain_span(sd)) {
561 if (!idle_cpu(i)) {
562 cpu = i;
563 goto unlock;
564 }
565 }
566 }
567unlock:
568 rcu_read_unlock();
569 return cpu;
570}
571/*
572 * When add_timer_on() enqueues a timer into the timer wheel of an
573 * idle CPU then this timer might expire before the next timer event
574 * which is scheduled to wake up that CPU. In case of a completely
575 * idle system the next event might even be infinite time into the
576 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
577 * leaves the inner idle loop so the newly added timer is taken into
578 * account when the CPU goes back to idle and evaluates the timer
579 * wheel for the next timer event.
580 */
581void wake_up_idle_cpu(int cpu)
582{
583 struct rq *rq = cpu_rq(cpu);
584
585 if (cpu == smp_processor_id())
586 return;
587
588 /*
589 * This is safe, as this function is called with the timer
590 * wheel base lock of (cpu) held. When the CPU is on the way
591 * to idle and has not yet set rq->curr to idle then it will
592 * be serialized on the timer wheel base lock and take the new
593 * timer into account automatically.
594 */
595 if (rq->curr != rq->idle)
596 return;
597
598 /*
599 * We can set TIF_RESCHED on the idle task of the other CPU
600 * lockless. The worst case is that the other CPU runs the
601 * idle task through an additional NOOP schedule()
602 */
603 set_tsk_need_resched(rq->idle);
604
605 /* NEED_RESCHED must be visible before we test polling */
606 smp_mb();
607 if (!tsk_is_polling(rq->idle))
608 smp_send_reschedule(cpu);
609}
610
611static inline bool got_nohz_idle_kick(void)
612{
613 int cpu = smp_processor_id();
614 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
615}
616
617#else /* CONFIG_NO_HZ */
618
619static inline bool got_nohz_idle_kick(void)
620{
621 return false;
622}
623
624#endif /* CONFIG_NO_HZ */
625
626void sched_avg_update(struct rq *rq)
627{
628 s64 period = sched_avg_period();
629
630 while ((s64)(rq->clock - rq->age_stamp) > period) {
631 /*
632 * Inline assembly required to prevent the compiler
633 * optimising this loop into a divmod call.
634 * See __iter_div_u64_rem() for another example of this.
635 */
636 asm("" : "+rm" (rq->age_stamp));
637 rq->age_stamp += period;
638 rq->rt_avg /= 2;
639 }
640}
641
642#else /* !CONFIG_SMP */
643void resched_task(struct task_struct *p)
644{
645 assert_raw_spin_locked(&task_rq(p)->lock);
646 set_tsk_need_resched(p);
647}
648#endif /* CONFIG_SMP */
649
650#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
651 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
652/*
653 * Iterate task_group tree rooted at *from, calling @down when first entering a
654 * node and @up when leaving it for the final time.
655 *
656 * Caller must hold rcu_lock or sufficient equivalent.
657 */
658int walk_tg_tree_from(struct task_group *from,
659 tg_visitor down, tg_visitor up, void *data)
660{
661 struct task_group *parent, *child;
662 int ret;
663
664 parent = from;
665
666down:
667 ret = (*down)(parent, data);
668 if (ret)
669 goto out;
670 list_for_each_entry_rcu(child, &parent->children, siblings) {
671 parent = child;
672 goto down;
673
674up:
675 continue;
676 }
677 ret = (*up)(parent, data);
678 if (ret || parent == from)
679 goto out;
680
681 child = parent;
682 parent = parent->parent;
683 if (parent)
684 goto up;
685out:
686 return ret;
687}
688
689int tg_nop(struct task_group *tg, void *data)
690{
691 return 0;
692}
693#endif
694
695static void set_load_weight(struct task_struct *p)
696{
697 int prio = p->static_prio - MAX_RT_PRIO;
698 struct load_weight *load = &p->se.load;
699
700 /*
701 * SCHED_IDLE tasks get minimal weight:
702 */
703 if (p->policy == SCHED_IDLE) {
704 load->weight = scale_load(WEIGHT_IDLEPRIO);
705 load->inv_weight = WMULT_IDLEPRIO;
706 return;
707 }
708
709 load->weight = scale_load(prio_to_weight[prio]);
710 load->inv_weight = prio_to_wmult[prio];
711}
712
713static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
714{
715 update_rq_clock(rq);
716 sched_info_queued(p);
717 p->sched_class->enqueue_task(rq, p, flags);
718}
719
720static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
721{
722 update_rq_clock(rq);
723 sched_info_dequeued(p);
724 p->sched_class->dequeue_task(rq, p, flags);
725}
726
727void activate_task(struct rq *rq, struct task_struct *p, int flags)
728{
729 if (task_contributes_to_load(p))
730 rq->nr_uninterruptible--;
731
732 enqueue_task(rq, p, flags);
733}
734
735void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
736{
737 if (task_contributes_to_load(p))
738 rq->nr_uninterruptible++;
739
740 dequeue_task(rq, p, flags);
741}
742
743#ifdef CONFIG_IRQ_TIME_ACCOUNTING
744
745/*
746 * There are no locks covering percpu hardirq/softirq time.
747 * They are only modified in account_system_vtime, on corresponding CPU
748 * with interrupts disabled. So, writes are safe.
749 * They are read and saved off onto struct rq in update_rq_clock().
750 * This may result in other CPU reading this CPU's irq time and can
751 * race with irq/account_system_vtime on this CPU. We would either get old
752 * or new value with a side effect of accounting a slice of irq time to wrong
753 * task when irq is in progress while we read rq->clock. That is a worthy
754 * compromise in place of having locks on each irq in account_system_time.
755 */
756static DEFINE_PER_CPU(u64, cpu_hardirq_time);
757static DEFINE_PER_CPU(u64, cpu_softirq_time);
758
759static DEFINE_PER_CPU(u64, irq_start_time);
760static int sched_clock_irqtime;
761
762void enable_sched_clock_irqtime(void)
763{
764 sched_clock_irqtime = 1;
765}
766
767void disable_sched_clock_irqtime(void)
768{
769 sched_clock_irqtime = 0;
770}
771
772#ifndef CONFIG_64BIT
773static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
774
775static inline void irq_time_write_begin(void)
776{
777 __this_cpu_inc(irq_time_seq.sequence);
778 smp_wmb();
779}
780
781static inline void irq_time_write_end(void)
782{
783 smp_wmb();
784 __this_cpu_inc(irq_time_seq.sequence);
785}
786
787static inline u64 irq_time_read(int cpu)
788{
789 u64 irq_time;
790 unsigned seq;
791
792 do {
793 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
794 irq_time = per_cpu(cpu_softirq_time, cpu) +
795 per_cpu(cpu_hardirq_time, cpu);
796 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
797
798 return irq_time;
799}
800#else /* CONFIG_64BIT */
801static inline void irq_time_write_begin(void)
802{
803}
804
805static inline void irq_time_write_end(void)
806{
807}
808
809static inline u64 irq_time_read(int cpu)
810{
811 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
812}
813#endif /* CONFIG_64BIT */
814
815/*
816 * Called before incrementing preempt_count on {soft,}irq_enter
817 * and before decrementing preempt_count on {soft,}irq_exit.
818 */
819void account_system_vtime(struct task_struct *curr)
820{
821 unsigned long flags;
822 s64 delta;
823 int cpu;
824
825 if (!sched_clock_irqtime)
826 return;
827
828 local_irq_save(flags);
829
830 cpu = smp_processor_id();
831 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
832 __this_cpu_add(irq_start_time, delta);
833
834 irq_time_write_begin();
835 /*
836 * We do not account for softirq time from ksoftirqd here.
837 * We want to continue accounting softirq time to ksoftirqd thread
838 * in that case, so as not to confuse scheduler with a special task
839 * that do not consume any time, but still wants to run.
840 */
841 if (hardirq_count())
842 __this_cpu_add(cpu_hardirq_time, delta);
843 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
844 __this_cpu_add(cpu_softirq_time, delta);
845
846 irq_time_write_end();
847 local_irq_restore(flags);
848}
849EXPORT_SYMBOL_GPL(account_system_vtime);
850
851#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
852
853#ifdef CONFIG_PARAVIRT
854static inline u64 steal_ticks(u64 steal)
855{
856 if (unlikely(steal > NSEC_PER_SEC))
857 return div_u64(steal, TICK_NSEC);
858
859 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
860}
861#endif
862
863static void update_rq_clock_task(struct rq *rq, s64 delta)
864{
865/*
866 * In theory, the compile should just see 0 here, and optimize out the call
867 * to sched_rt_avg_update. But I don't trust it...
868 */
869#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
870 s64 steal = 0, irq_delta = 0;
871#endif
872#ifdef CONFIG_IRQ_TIME_ACCOUNTING
873 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
874
875 /*
876 * Since irq_time is only updated on {soft,}irq_exit, we might run into
877 * this case when a previous update_rq_clock() happened inside a
878 * {soft,}irq region.
879 *
880 * When this happens, we stop ->clock_task and only update the
881 * prev_irq_time stamp to account for the part that fit, so that a next
882 * update will consume the rest. This ensures ->clock_task is
883 * monotonic.
884 *
885 * It does however cause some slight miss-attribution of {soft,}irq
886 * time, a more accurate solution would be to update the irq_time using
887 * the current rq->clock timestamp, except that would require using
888 * atomic ops.
889 */
890 if (irq_delta > delta)
891 irq_delta = delta;
892
893 rq->prev_irq_time += irq_delta;
894 delta -= irq_delta;
895#endif
896#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
897 if (static_key_false((¶virt_steal_rq_enabled))) {
898 u64 st;
899
900 steal = paravirt_steal_clock(cpu_of(rq));
901 steal -= rq->prev_steal_time_rq;
902
903 if (unlikely(steal > delta))
904 steal = delta;
905
906 st = steal_ticks(steal);
907 steal = st * TICK_NSEC;
908
909 rq->prev_steal_time_rq += steal;
910
911 delta -= steal;
912 }
913#endif
914
915 rq->clock_task += delta;
916
917#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
918 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
919 sched_rt_avg_update(rq, irq_delta + steal);
920#endif
921}
922
923#ifdef CONFIG_IRQ_TIME_ACCOUNTING
924static int irqtime_account_hi_update(void)
925{
926 u64 *cpustat = kcpustat_this_cpu->cpustat;
927 unsigned long flags;
928 u64 latest_ns;
929 int ret = 0;
930
931 local_irq_save(flags);
932 latest_ns = this_cpu_read(cpu_hardirq_time);
933 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
934 ret = 1;
935 local_irq_restore(flags);
936 return ret;
937}
938
939static int irqtime_account_si_update(void)
940{
941 u64 *cpustat = kcpustat_this_cpu->cpustat;
942 unsigned long flags;
943 u64 latest_ns;
944 int ret = 0;
945
946 local_irq_save(flags);
947 latest_ns = this_cpu_read(cpu_softirq_time);
948 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
949 ret = 1;
950 local_irq_restore(flags);
951 return ret;
952}
953
954#else /* CONFIG_IRQ_TIME_ACCOUNTING */
955
956#define sched_clock_irqtime (0)
957
958#endif
959
960void sched_set_stop_task(int cpu, struct task_struct *stop)
961{
962 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
963 struct task_struct *old_stop = cpu_rq(cpu)->stop;
964
965 if (stop) {
966 /*
967 * Make it appear like a SCHED_FIFO task, its something
968 * userspace knows about and won't get confused about.
969 *
970 * Also, it will make PI more or less work without too
971 * much confusion -- but then, stop work should not
972 * rely on PI working anyway.
973 */
974 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
975
976 stop->sched_class = &stop_sched_class;
977 }
978
979 cpu_rq(cpu)->stop = stop;
980
981 if (old_stop) {
982 /*
983 * Reset it back to a normal scheduling class so that
984 * it can die in pieces.
985 */
986 old_stop->sched_class = &rt_sched_class;
987 }
988}
989
990/*
991 * __normal_prio - return the priority that is based on the static prio
992 */
993static inline int __normal_prio(struct task_struct *p)
994{
995 return p->static_prio;
996}
997
998/*
999 * Calculate the expected normal priority: i.e. priority
1000 * without taking RT-inheritance into account. Might be
1001 * boosted by interactivity modifiers. Changes upon fork,
1002 * setprio syscalls, and whenever the interactivity
1003 * estimator recalculates.
1004 */
1005static inline int normal_prio(struct task_struct *p)
1006{
1007 int prio;
1008
1009 if (task_has_rt_policy(p))
1010 prio = MAX_RT_PRIO-1 - p->rt_priority;
1011 else
1012 prio = __normal_prio(p);
1013 return prio;
1014}
1015
1016/*
1017 * Calculate the current priority, i.e. the priority
1018 * taken into account by the scheduler. This value might
1019 * be boosted by RT tasks, or might be boosted by
1020 * interactivity modifiers. Will be RT if the task got
1021 * RT-boosted. If not then it returns p->normal_prio.
1022 */
1023static int effective_prio(struct task_struct *p)
1024{
1025 p->normal_prio = normal_prio(p);
1026 /*
1027 * If we are RT tasks or we were boosted to RT priority,
1028 * keep the priority unchanged. Otherwise, update priority
1029 * to the normal priority:
1030 */
1031 if (!rt_prio(p->prio))
1032 return p->normal_prio;
1033 return p->prio;
1034}
1035
1036/**
1037 * task_curr - is this task currently executing on a CPU?
1038 * @p: the task in question.
1039 */
1040inline int task_curr(const struct task_struct *p)
1041{
1042 return cpu_curr(task_cpu(p)) == p;
1043}
1044
1045static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1046 const struct sched_class *prev_class,
1047 int oldprio)
1048{
1049 if (prev_class != p->sched_class) {
1050 if (prev_class->switched_from)
1051 prev_class->switched_from(rq, p);
1052 p->sched_class->switched_to(rq, p);
1053 } else if (oldprio != p->prio)
1054 p->sched_class->prio_changed(rq, p, oldprio);
1055}
1056
1057void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1058{
1059 const struct sched_class *class;
1060
1061 if (p->sched_class == rq->curr->sched_class) {
1062 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1063 } else {
1064 for_each_class(class) {
1065 if (class == rq->curr->sched_class)
1066 break;
1067 if (class == p->sched_class) {
1068 resched_task(rq->curr);
1069 break;
1070 }
1071 }
1072 }
1073
1074 /*
1075 * A queue event has occurred, and we're going to schedule. In
1076 * this case, we can save a useless back to back clock update.
1077 */
1078 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1079 rq->skip_clock_update = 1;
1080}
1081
1082#ifdef CONFIG_SMP
1083void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1084{
1085#ifdef CONFIG_SCHED_DEBUG
1086 /*
1087 * We should never call set_task_cpu() on a blocked task,
1088 * ttwu() will sort out the placement.
1089 */
1090 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1091 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
1092
1093#ifdef CONFIG_LOCKDEP
1094 /*
1095 * The caller should hold either p->pi_lock or rq->lock, when changing
1096 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1097 *
1098 * sched_move_task() holds both and thus holding either pins the cgroup,
1099 * see task_group().
1100 *
1101 * Furthermore, all task_rq users should acquire both locks, see
1102 * task_rq_lock().
1103 */
1104 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1105 lockdep_is_held(&task_rq(p)->lock)));
1106#endif
1107#endif
1108
1109 trace_sched_migrate_task(p, new_cpu);
1110
1111 if (task_cpu(p) != new_cpu) {
1112 p->se.nr_migrations++;
1113 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1114 }
1115
1116 __set_task_cpu(p, new_cpu);
1117}
1118
1119struct migration_arg {
1120 struct task_struct *task;
1121 int dest_cpu;
1122};
1123
1124static int migration_cpu_stop(void *data);
1125
1126/*
1127 * wait_task_inactive - wait for a thread to unschedule.
1128 *
1129 * If @match_state is nonzero, it's the @p->state value just checked and
1130 * not expected to change. If it changes, i.e. @p might have woken up,
1131 * then return zero. When we succeed in waiting for @p to be off its CPU,
1132 * we return a positive number (its total switch count). If a second call
1133 * a short while later returns the same number, the caller can be sure that
1134 * @p has remained unscheduled the whole time.
1135 *
1136 * The caller must ensure that the task *will* unschedule sometime soon,
1137 * else this function might spin for a *long* time. This function can't
1138 * be called with interrupts off, or it may introduce deadlock with
1139 * smp_call_function() if an IPI is sent by the same process we are
1140 * waiting to become inactive.
1141 */
1142unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1143{
1144 unsigned long flags;
1145 int running, on_rq;
1146 unsigned long ncsw;
1147 struct rq *rq;
1148
1149 for (;;) {
1150 /*
1151 * We do the initial early heuristics without holding
1152 * any task-queue locks at all. We'll only try to get
1153 * the runqueue lock when things look like they will
1154 * work out!
1155 */
1156 rq = task_rq(p);
1157
1158 /*
1159 * If the task is actively running on another CPU
1160 * still, just relax and busy-wait without holding
1161 * any locks.
1162 *
1163 * NOTE! Since we don't hold any locks, it's not
1164 * even sure that "rq" stays as the right runqueue!
1165 * But we don't care, since "task_running()" will
1166 * return false if the runqueue has changed and p
1167 * is actually now running somewhere else!
1168 */
1169 while (task_running(rq, p)) {
1170 if (match_state && unlikely(p->state != match_state))
1171 return 0;
1172 cpu_relax();
1173 }
1174
1175 /*
1176 * Ok, time to look more closely! We need the rq
1177 * lock now, to be *sure*. If we're wrong, we'll
1178 * just go back and repeat.
1179 */
1180 rq = task_rq_lock(p, &flags);
1181 trace_sched_wait_task(p);
1182 running = task_running(rq, p);
1183 on_rq = p->on_rq;
1184 ncsw = 0;
1185 if (!match_state || p->state == match_state)
1186 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1187 task_rq_unlock(rq, p, &flags);
1188
1189 /*
1190 * If it changed from the expected state, bail out now.
1191 */
1192 if (unlikely(!ncsw))
1193 break;
1194
1195 /*
1196 * Was it really running after all now that we
1197 * checked with the proper locks actually held?
1198 *
1199 * Oops. Go back and try again..
1200 */
1201 if (unlikely(running)) {
1202 cpu_relax();
1203 continue;
1204 }
1205
1206 /*
1207 * It's not enough that it's not actively running,
1208 * it must be off the runqueue _entirely_, and not
1209 * preempted!
1210 *
1211 * So if it was still runnable (but just not actively
1212 * running right now), it's preempted, and we should
1213 * yield - it could be a while.
1214 */
1215 if (unlikely(on_rq)) {
1216 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1217
1218 set_current_state(TASK_UNINTERRUPTIBLE);
1219 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1220 continue;
1221 }
1222
1223 /*
1224 * Ahh, all good. It wasn't running, and it wasn't
1225 * runnable, which means that it will never become
1226 * running in the future either. We're all done!
1227 */
1228 break;
1229 }
1230
1231 return ncsw;
1232}
1233
1234/***
1235 * kick_process - kick a running thread to enter/exit the kernel
1236 * @p: the to-be-kicked thread
1237 *
1238 * Cause a process which is running on another CPU to enter
1239 * kernel-mode, without any delay. (to get signals handled.)
1240 *
1241 * NOTE: this function doesn't have to take the runqueue lock,
1242 * because all it wants to ensure is that the remote task enters
1243 * the kernel. If the IPI races and the task has been migrated
1244 * to another CPU then no harm is done and the purpose has been
1245 * achieved as well.
1246 */
1247void kick_process(struct task_struct *p)
1248{
1249 int cpu;
1250
1251 preempt_disable();
1252 cpu = task_cpu(p);
1253 if ((cpu != smp_processor_id()) && task_curr(p))
1254 smp_send_reschedule(cpu);
1255 preempt_enable();
1256}
1257EXPORT_SYMBOL_GPL(kick_process);
1258#endif /* CONFIG_SMP */
1259
1260#ifdef CONFIG_SMP
1261/*
1262 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1263 */
1264static int select_fallback_rq(int cpu, struct task_struct *p)
1265{
1266 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1267 enum { cpuset, possible, fail } state = cpuset;
1268 int dest_cpu;
1269
1270 /* Look for allowed, online CPU in same node. */
1271 for_each_cpu(dest_cpu, nodemask) {
1272 if (!cpu_online(dest_cpu))
1273 continue;
1274 if (!cpu_active(dest_cpu))
1275 continue;
1276 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1277 return dest_cpu;
1278 }
1279
1280 for (;;) {
1281 /* Any allowed, online CPU? */
1282 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1283 if (!cpu_online(dest_cpu))
1284 continue;
1285 if (!cpu_active(dest_cpu))
1286 continue;
1287 goto out;
1288 }
1289
1290 switch (state) {
1291 case cpuset:
1292 /* No more Mr. Nice Guy. */
1293 cpuset_cpus_allowed_fallback(p);
1294 state = possible;
1295 break;
1296
1297 case possible:
1298 do_set_cpus_allowed(p, cpu_possible_mask);
1299 state = fail;
1300 break;
1301
1302 case fail:
1303 BUG();
1304 break;
1305 }
1306 }
1307
1308out:
1309 if (state != cpuset) {
1310 /*
1311 * Don't tell them about moving exiting tasks or
1312 * kernel threads (both mm NULL), since they never
1313 * leave kernel.
1314 */
1315 if (p->mm && printk_ratelimit()) {
1316 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1317 task_pid_nr(p), p->comm, cpu);
1318 }
1319 }
1320
1321 return dest_cpu;
1322}
1323
1324/*
1325 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1326 */
1327static inline
1328int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1329{
1330 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1331
1332 /*
1333 * In order not to call set_task_cpu() on a blocking task we need
1334 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1335 * cpu.
1336 *
1337 * Since this is common to all placement strategies, this lives here.
1338 *
1339 * [ this allows ->select_task() to simply return task_cpu(p) and
1340 * not worry about this generic constraint ]
1341 */
1342 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1343 !cpu_online(cpu)))
1344 cpu = select_fallback_rq(task_cpu(p), p);
1345
1346 return cpu;
1347}
1348
1349static void update_avg(u64 *avg, u64 sample)
1350{
1351 s64 diff = sample - *avg;
1352 *avg += diff >> 3;
1353}
1354#endif
1355
1356static void
1357ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1358{
1359#ifdef CONFIG_SCHEDSTATS
1360 struct rq *rq = this_rq();
1361
1362#ifdef CONFIG_SMP
1363 int this_cpu = smp_processor_id();
1364
1365 if (cpu == this_cpu) {
1366 schedstat_inc(rq, ttwu_local);
1367 schedstat_inc(p, se.statistics.nr_wakeups_local);
1368 } else {
1369 struct sched_domain *sd;
1370
1371 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1372 rcu_read_lock();
1373 for_each_domain(this_cpu, sd) {
1374 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1375 schedstat_inc(sd, ttwu_wake_remote);
1376 break;
1377 }
1378 }
1379 rcu_read_unlock();
1380 }
1381
1382 if (wake_flags & WF_MIGRATED)
1383 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1384
1385#endif /* CONFIG_SMP */
1386
1387 schedstat_inc(rq, ttwu_count);
1388 schedstat_inc(p, se.statistics.nr_wakeups);
1389
1390 if (wake_flags & WF_SYNC)
1391 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1392
1393#endif /* CONFIG_SCHEDSTATS */
1394}
1395
1396static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1397{
1398 activate_task(rq, p, en_flags);
1399 p->on_rq = 1;
1400
1401 /* if a worker is waking up, notify workqueue */
1402 if (p->flags & PF_WQ_WORKER)
1403 wq_worker_waking_up(p, cpu_of(rq));
1404}
1405
1406/*
1407 * Mark the task runnable and perform wakeup-preemption.
1408 */
1409static void
1410ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1411{
1412 trace_sched_wakeup(p, true);
1413 check_preempt_curr(rq, p, wake_flags);
1414
1415 p->state = TASK_RUNNING;
1416#ifdef CONFIG_SMP
1417 if (p->sched_class->task_woken)
1418 p->sched_class->task_woken(rq, p);
1419
1420 if (rq->idle_stamp) {
1421 u64 delta = rq->clock - rq->idle_stamp;
1422 u64 max = 2*sysctl_sched_migration_cost;
1423
1424 if (delta > max)
1425 rq->avg_idle = max;
1426 else
1427 update_avg(&rq->avg_idle, delta);
1428 rq->idle_stamp = 0;
1429 }
1430#endif
1431}
1432
1433static void
1434ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1435{
1436#ifdef CONFIG_SMP
1437 if (p->sched_contributes_to_load)
1438 rq->nr_uninterruptible--;
1439#endif
1440
1441 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1442 ttwu_do_wakeup(rq, p, wake_flags);
1443}
1444
1445/*
1446 * Called in case the task @p isn't fully descheduled from its runqueue,
1447 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1448 * since all we need to do is flip p->state to TASK_RUNNING, since
1449 * the task is still ->on_rq.
1450 */
1451static int ttwu_remote(struct task_struct *p, int wake_flags)
1452{
1453 struct rq *rq;
1454 int ret = 0;
1455
1456 rq = __task_rq_lock(p);
1457 if (p->on_rq) {
1458 ttwu_do_wakeup(rq, p, wake_flags);
1459 ret = 1;
1460 }
1461 __task_rq_unlock(rq);
1462
1463 return ret;
1464}
1465
1466#ifdef CONFIG_SMP
1467static void sched_ttwu_pending(void)
1468{
1469 struct rq *rq = this_rq();
1470 struct llist_node *llist = llist_del_all(&rq->wake_list);
1471 struct task_struct *p;
1472
1473 raw_spin_lock(&rq->lock);
1474
1475 while (llist) {
1476 p = llist_entry(llist, struct task_struct, wake_entry);
1477 llist = llist_next(llist);
1478 ttwu_do_activate(rq, p, 0);
1479 }
1480
1481 raw_spin_unlock(&rq->lock);
1482}
1483
1484void scheduler_ipi(void)
1485{
1486 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1487 return;
1488
1489 /*
1490 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1491 * traditionally all their work was done from the interrupt return
1492 * path. Now that we actually do some work, we need to make sure
1493 * we do call them.
1494 *
1495 * Some archs already do call them, luckily irq_enter/exit nest
1496 * properly.
1497 *
1498 * Arguably we should visit all archs and update all handlers,
1499 * however a fair share of IPIs are still resched only so this would
1500 * somewhat pessimize the simple resched case.
1501 */
1502 irq_enter();
1503 sched_ttwu_pending();
1504
1505 /*
1506 * Check if someone kicked us for doing the nohz idle load balance.
1507 */
1508 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1509 this_rq()->idle_balance = 1;
1510 raise_softirq_irqoff(SCHED_SOFTIRQ);
1511 }
1512 irq_exit();
1513}
1514
1515static void ttwu_queue_remote(struct task_struct *p, int cpu)
1516{
1517 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1518 smp_send_reschedule(cpu);
1519}
1520
1521#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1522static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
1523{
1524 struct rq *rq;
1525 int ret = 0;
1526
1527 rq = __task_rq_lock(p);
1528 if (p->on_cpu) {
1529 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1530 ttwu_do_wakeup(rq, p, wake_flags);
1531 ret = 1;
1532 }
1533 __task_rq_unlock(rq);
1534
1535 return ret;
1536
1537}
1538#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1539
1540bool cpus_share_cache(int this_cpu, int that_cpu)
1541{
1542 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1543}
1544#endif /* CONFIG_SMP */
1545
1546static void ttwu_queue(struct task_struct *p, int cpu)
1547{
1548 struct rq *rq = cpu_rq(cpu);
1549
1550#if defined(CONFIG_SMP)
1551 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1552 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1553 ttwu_queue_remote(p, cpu);
1554 return;
1555 }
1556#endif
1557
1558 raw_spin_lock(&rq->lock);
1559 ttwu_do_activate(rq, p, 0);
1560 raw_spin_unlock(&rq->lock);
1561}
1562
1563/**
1564 * try_to_wake_up - wake up a thread
1565 * @p: the thread to be awakened
1566 * @state: the mask of task states that can be woken
1567 * @wake_flags: wake modifier flags (WF_*)
1568 *
1569 * Put it on the run-queue if it's not already there. The "current"
1570 * thread is always on the run-queue (except when the actual
1571 * re-schedule is in progress), and as such you're allowed to do
1572 * the simpler "current->state = TASK_RUNNING" to mark yourself
1573 * runnable without the overhead of this.
1574 *
1575 * Returns %true if @p was woken up, %false if it was already running
1576 * or @state didn't match @p's state.
1577 */
1578static int
1579try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1580{
1581 unsigned long flags;
1582 int cpu, success = 0;
1583
1584 smp_wmb();
1585 raw_spin_lock_irqsave(&p->pi_lock, flags);
1586 if (!(p->state & state))
1587 goto out;
1588
1589 success = 1; /* we're going to change ->state */
1590 cpu = task_cpu(p);
1591
1592 if (p->on_rq && ttwu_remote(p, wake_flags))
1593 goto stat;
1594
1595#ifdef CONFIG_SMP
1596 /*
1597 * If the owning (remote) cpu is still in the middle of schedule() with
1598 * this task as prev, wait until its done referencing the task.
1599 */
1600 while (p->on_cpu) {
1601#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1602 /*
1603 * In case the architecture enables interrupts in
1604 * context_switch(), we cannot busy wait, since that
1605 * would lead to deadlocks when an interrupt hits and
1606 * tries to wake up @prev. So bail and do a complete
1607 * remote wakeup.
1608 */
1609 if (ttwu_activate_remote(p, wake_flags))
1610 goto stat;
1611#else
1612 cpu_relax();
1613#endif
1614 }
1615 /*
1616 * Pairs with the smp_wmb() in finish_lock_switch().
1617 */
1618 smp_rmb();
1619
1620 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1621 p->state = TASK_WAKING;
1622
1623 if (p->sched_class->task_waking)
1624 p->sched_class->task_waking(p);
1625
1626 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1627 if (task_cpu(p) != cpu) {
1628 wake_flags |= WF_MIGRATED;
1629 set_task_cpu(p, cpu);
1630 }
1631#endif /* CONFIG_SMP */
1632
1633 ttwu_queue(p, cpu);
1634stat:
1635 ttwu_stat(p, cpu, wake_flags);
1636out:
1637 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1638
1639 return success;
1640}
1641
1642/**
1643 * try_to_wake_up_local - try to wake up a local task with rq lock held
1644 * @p: the thread to be awakened
1645 *
1646 * Put @p on the run-queue if it's not already there. The caller must
1647 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1648 * the current task.
1649 */
1650static void try_to_wake_up_local(struct task_struct *p)
1651{
1652 struct rq *rq = task_rq(p);
1653
1654 BUG_ON(rq != this_rq());
1655 BUG_ON(p == current);
1656 lockdep_assert_held(&rq->lock);
1657
1658 if (!raw_spin_trylock(&p->pi_lock)) {
1659 raw_spin_unlock(&rq->lock);
1660 raw_spin_lock(&p->pi_lock);
1661 raw_spin_lock(&rq->lock);
1662 }
1663
1664 if (!(p->state & TASK_NORMAL))
1665 goto out;
1666
1667 if (!p->on_rq)
1668 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1669
1670 ttwu_do_wakeup(rq, p, 0);
1671 ttwu_stat(p, smp_processor_id(), 0);
1672out:
1673 raw_spin_unlock(&p->pi_lock);
1674}
1675
1676/**
1677 * wake_up_process - Wake up a specific process
1678 * @p: The process to be woken up.
1679 *
1680 * Attempt to wake up the nominated process and move it to the set of runnable
1681 * processes. Returns 1 if the process was woken up, 0 if it was already
1682 * running.
1683 *
1684 * It may be assumed that this function implies a write memory barrier before
1685 * changing the task state if and only if any tasks are woken up.
1686 */
1687int wake_up_process(struct task_struct *p)
1688{
1689 return try_to_wake_up(p, TASK_ALL, 0);
1690}
1691EXPORT_SYMBOL(wake_up_process);
1692
1693int wake_up_state(struct task_struct *p, unsigned int state)
1694{
1695 return try_to_wake_up(p, state, 0);
1696}
1697
1698/*
1699 * Perform scheduler related setup for a newly forked process p.
1700 * p is forked by current.
1701 *
1702 * __sched_fork() is basic setup used by init_idle() too:
1703 */
1704static void __sched_fork(struct task_struct *p)
1705{
1706 p->on_rq = 0;
1707
1708 p->se.on_rq = 0;
1709 p->se.exec_start = 0;
1710 p->se.sum_exec_runtime = 0;
1711 p->se.prev_sum_exec_runtime = 0;
1712 p->se.nr_migrations = 0;
1713 p->se.vruntime = 0;
1714 INIT_LIST_HEAD(&p->se.group_node);
1715
1716#ifdef CONFIG_SCHEDSTATS
1717 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1718#endif
1719
1720 INIT_LIST_HEAD(&p->rt.run_list);
1721
1722#ifdef CONFIG_PREEMPT_NOTIFIERS
1723 INIT_HLIST_HEAD(&p->preempt_notifiers);
1724#endif
1725}
1726
1727/*
1728 * fork()/clone()-time setup:
1729 */
1730void sched_fork(struct task_struct *p)
1731{
1732 unsigned long flags;
1733 int cpu = get_cpu();
1734
1735 __sched_fork(p);
1736 /*
1737 * We mark the process as running here. This guarantees that
1738 * nobody will actually run it, and a signal or other external
1739 * event cannot wake it up and insert it on the runqueue either.
1740 */
1741 p->state = TASK_RUNNING;
1742
1743 /*
1744 * Make sure we do not leak PI boosting priority to the child.
1745 */
1746 p->prio = current->normal_prio;
1747
1748 /*
1749 * Revert to default priority/policy on fork if requested.
1750 */
1751 if (unlikely(p->sched_reset_on_fork)) {
1752 if (task_has_rt_policy(p)) {
1753 p->policy = SCHED_NORMAL;
1754 p->static_prio = NICE_TO_PRIO(0);
1755 p->rt_priority = 0;
1756 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1757 p->static_prio = NICE_TO_PRIO(0);
1758
1759 p->prio = p->normal_prio = __normal_prio(p);
1760 set_load_weight(p);
1761
1762 /*
1763 * We don't need the reset flag anymore after the fork. It has
1764 * fulfilled its duty:
1765 */
1766 p->sched_reset_on_fork = 0;
1767 }
1768
1769 if (!rt_prio(p->prio))
1770 p->sched_class = &fair_sched_class;
1771
1772 if (p->sched_class->task_fork)
1773 p->sched_class->task_fork(p);
1774
1775 /*
1776 * The child is not yet in the pid-hash so no cgroup attach races,
1777 * and the cgroup is pinned to this child due to cgroup_fork()
1778 * is ran before sched_fork().
1779 *
1780 * Silence PROVE_RCU.
1781 */
1782 raw_spin_lock_irqsave(&p->pi_lock, flags);
1783 set_task_cpu(p, cpu);
1784 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1785
1786#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1787 if (likely(sched_info_on()))
1788 memset(&p->sched_info, 0, sizeof(p->sched_info));
1789#endif
1790#if defined(CONFIG_SMP)
1791 p->on_cpu = 0;
1792#endif
1793#ifdef CONFIG_PREEMPT_COUNT
1794 /* Want to start with kernel preemption disabled. */
1795 task_thread_info(p)->preempt_count = 1;
1796#endif
1797#ifdef CONFIG_SMP
1798 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1799#endif
1800
1801 put_cpu();
1802}
1803
1804/*
1805 * wake_up_new_task - wake up a newly created task for the first time.
1806 *
1807 * This function will do some initial scheduler statistics housekeeping
1808 * that must be done for every newly created context, then puts the task
1809 * on the runqueue and wakes it.
1810 */
1811void wake_up_new_task(struct task_struct *p)
1812{
1813 unsigned long flags;
1814 struct rq *rq;
1815
1816 raw_spin_lock_irqsave(&p->pi_lock, flags);
1817#ifdef CONFIG_SMP
1818 /*
1819 * Fork balancing, do it here and not earlier because:
1820 * - cpus_allowed can change in the fork path
1821 * - any previously selected cpu might disappear through hotplug
1822 */
1823 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1824#endif
1825
1826 rq = __task_rq_lock(p);
1827 activate_task(rq, p, 0);
1828 p->on_rq = 1;
1829 trace_sched_wakeup_new(p, true);
1830 check_preempt_curr(rq, p, WF_FORK);
1831#ifdef CONFIG_SMP
1832 if (p->sched_class->task_woken)
1833 p->sched_class->task_woken(rq, p);
1834#endif
1835 task_rq_unlock(rq, p, &flags);
1836}
1837
1838#ifdef CONFIG_PREEMPT_NOTIFIERS
1839
1840/**
1841 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1842 * @notifier: notifier struct to register
1843 */
1844void preempt_notifier_register(struct preempt_notifier *notifier)
1845{
1846 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1847}
1848EXPORT_SYMBOL_GPL(preempt_notifier_register);
1849
1850/**
1851 * preempt_notifier_unregister - no longer interested in preemption notifications
1852 * @notifier: notifier struct to unregister
1853 *
1854 * This is safe to call from within a preemption notifier.
1855 */
1856void preempt_notifier_unregister(struct preempt_notifier *notifier)
1857{
1858 hlist_del(¬ifier->link);
1859}
1860EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1861
1862static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1863{
1864 struct preempt_notifier *notifier;
1865 struct hlist_node *node;
1866
1867 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1868 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1869}
1870
1871static void
1872fire_sched_out_preempt_notifiers(struct task_struct *curr,
1873 struct task_struct *next)
1874{
1875 struct preempt_notifier *notifier;
1876 struct hlist_node *node;
1877
1878 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1879 notifier->ops->sched_out(notifier, next);
1880}
1881
1882#else /* !CONFIG_PREEMPT_NOTIFIERS */
1883
1884static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1885{
1886}
1887
1888static void
1889fire_sched_out_preempt_notifiers(struct task_struct *curr,
1890 struct task_struct *next)
1891{
1892}
1893
1894#endif /* CONFIG_PREEMPT_NOTIFIERS */
1895
1896/**
1897 * prepare_task_switch - prepare to switch tasks
1898 * @rq: the runqueue preparing to switch
1899 * @prev: the current task that is being switched out
1900 * @next: the task we are going to switch to.
1901 *
1902 * This is called with the rq lock held and interrupts off. It must
1903 * be paired with a subsequent finish_task_switch after the context
1904 * switch.
1905 *
1906 * prepare_task_switch sets up locking and calls architecture specific
1907 * hooks.
1908 */
1909static inline void
1910prepare_task_switch(struct rq *rq, struct task_struct *prev,
1911 struct task_struct *next)
1912{
1913 sched_info_switch(prev, next);
1914 perf_event_task_sched_out(prev, next);
1915 fire_sched_out_preempt_notifiers(prev, next);
1916 prepare_lock_switch(rq, next);
1917 prepare_arch_switch(next);
1918 trace_sched_switch(prev, next);
1919}
1920
1921/**
1922 * finish_task_switch - clean up after a task-switch
1923 * @rq: runqueue associated with task-switch
1924 * @prev: the thread we just switched away from.
1925 *
1926 * finish_task_switch must be called after the context switch, paired
1927 * with a prepare_task_switch call before the context switch.
1928 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1929 * and do any other architecture-specific cleanup actions.
1930 *
1931 * Note that we may have delayed dropping an mm in context_switch(). If
1932 * so, we finish that here outside of the runqueue lock. (Doing it
1933 * with the lock held can cause deadlocks; see schedule() for
1934 * details.)
1935 */
1936static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1937 __releases(rq->lock)
1938{
1939 struct mm_struct *mm = rq->prev_mm;
1940 long prev_state;
1941
1942 rq->prev_mm = NULL;
1943
1944 /*
1945 * A task struct has one reference for the use as "current".
1946 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1947 * schedule one last time. The schedule call will never return, and
1948 * the scheduled task must drop that reference.
1949 * The test for TASK_DEAD must occur while the runqueue locks are
1950 * still held, otherwise prev could be scheduled on another cpu, die
1951 * there before we look at prev->state, and then the reference would
1952 * be dropped twice.
1953 * Manfred Spraul <manfred@colorfullife.com>
1954 */
1955 prev_state = prev->state;
1956 finish_arch_switch(prev);
1957#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1958 local_irq_disable();
1959#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1960 perf_event_task_sched_in(prev, current);
1961#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1962 local_irq_enable();
1963#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1964 finish_lock_switch(rq, prev);
1965 finish_arch_post_lock_switch();
1966
1967 fire_sched_in_preempt_notifiers(current);
1968 if (mm)
1969 mmdrop(mm);
1970 if (unlikely(prev_state == TASK_DEAD)) {
1971 /*
1972 * Remove function-return probe instances associated with this
1973 * task and put them back on the free list.
1974 */
1975 kprobe_flush_task(prev);
1976 put_task_struct(prev);
1977 }
1978}
1979
1980#ifdef CONFIG_SMP
1981
1982/* assumes rq->lock is held */
1983static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1984{
1985 if (prev->sched_class->pre_schedule)
1986 prev->sched_class->pre_schedule(rq, prev);
1987}
1988
1989/* rq->lock is NOT held, but preemption is disabled */
1990static inline void post_schedule(struct rq *rq)
1991{
1992 if (rq->post_schedule) {
1993 unsigned long flags;
1994
1995 raw_spin_lock_irqsave(&rq->lock, flags);
1996 if (rq->curr->sched_class->post_schedule)
1997 rq->curr->sched_class->post_schedule(rq);
1998 raw_spin_unlock_irqrestore(&rq->lock, flags);
1999
2000 rq->post_schedule = 0;
2001 }
2002}
2003
2004#else
2005
2006static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2007{
2008}
2009
2010static inline void post_schedule(struct rq *rq)
2011{
2012}
2013
2014#endif
2015
2016/**
2017 * schedule_tail - first thing a freshly forked thread must call.
2018 * @prev: the thread we just switched away from.
2019 */
2020asmlinkage void schedule_tail(struct task_struct *prev)
2021 __releases(rq->lock)
2022{
2023 struct rq *rq = this_rq();
2024
2025 finish_task_switch(rq, prev);
2026
2027 /*
2028 * FIXME: do we need to worry about rq being invalidated by the
2029 * task_switch?
2030 */
2031 post_schedule(rq);
2032
2033#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2034 /* In this case, finish_task_switch does not reenable preemption */
2035 preempt_enable();
2036#endif
2037 if (current->set_child_tid)
2038 put_user(task_pid_vnr(current), current->set_child_tid);
2039}
2040
2041/*
2042 * context_switch - switch to the new MM and the new
2043 * thread's register state.
2044 */
2045static inline void
2046context_switch(struct rq *rq, struct task_struct *prev,
2047 struct task_struct *next)
2048{
2049 struct mm_struct *mm, *oldmm;
2050
2051 prepare_task_switch(rq, prev, next);
2052
2053 mm = next->mm;
2054 oldmm = prev->active_mm;
2055 /*
2056 * For paravirt, this is coupled with an exit in switch_to to
2057 * combine the page table reload and the switch backend into
2058 * one hypercall.
2059 */
2060 arch_start_context_switch(prev);
2061
2062 if (!mm) {
2063 next->active_mm = oldmm;
2064 atomic_inc(&oldmm->mm_count);
2065 enter_lazy_tlb(oldmm, next);
2066 } else
2067 switch_mm(oldmm, mm, next);
2068
2069 if (!prev->mm) {
2070 prev->active_mm = NULL;
2071 rq->prev_mm = oldmm;
2072 }
2073 /*
2074 * Since the runqueue lock will be released by the next
2075 * task (which is an invalid locking op but in the case
2076 * of the scheduler it's an obvious special-case), so we
2077 * do an early lockdep release here:
2078 */
2079#ifndef __ARCH_WANT_UNLOCKED_CTXSW
2080 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2081#endif
2082
2083 /* Here we just switch the register state and the stack. */
2084 switch_to(prev, next, prev);
2085
2086 barrier();
2087 /*
2088 * this_rq must be evaluated again because prev may have moved
2089 * CPUs since it called schedule(), thus the 'rq' on its stack
2090 * frame will be invalid.
2091 */
2092 finish_task_switch(this_rq(), prev);
2093}
2094
2095/*
2096 * nr_running, nr_uninterruptible and nr_context_switches:
2097 *
2098 * externally visible scheduler statistics: current number of runnable
2099 * threads, current number of uninterruptible-sleeping threads, total
2100 * number of context switches performed since bootup.
2101 */
2102unsigned long nr_running(void)
2103{
2104 unsigned long i, sum = 0;
2105
2106 for_each_online_cpu(i)
2107 sum += cpu_rq(i)->nr_running;
2108
2109 return sum;
2110}
2111
2112unsigned long nr_uninterruptible(void)
2113{
2114 unsigned long i, sum = 0;
2115
2116 for_each_possible_cpu(i)
2117 sum += cpu_rq(i)->nr_uninterruptible;
2118
2119 /*
2120 * Since we read the counters lockless, it might be slightly
2121 * inaccurate. Do not allow it to go below zero though:
2122 */
2123 if (unlikely((long)sum < 0))
2124 sum = 0;
2125
2126 return sum;
2127}
2128
2129unsigned long long nr_context_switches(void)
2130{
2131 int i;
2132 unsigned long long sum = 0;
2133
2134 for_each_possible_cpu(i)
2135 sum += cpu_rq(i)->nr_switches;
2136
2137 return sum;
2138}
2139
2140unsigned long nr_iowait(void)
2141{
2142 unsigned long i, sum = 0;
2143
2144 for_each_possible_cpu(i)
2145 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2146
2147 return sum;
2148}
2149
2150unsigned long nr_iowait_cpu(int cpu)
2151{
2152 struct rq *this = cpu_rq(cpu);
2153 return atomic_read(&this->nr_iowait);
2154}
2155
2156unsigned long this_cpu_load(void)
2157{
2158 struct rq *this = this_rq();
2159 return this->cpu_load[0];
2160}
2161
2162
2163/*
2164 * Global load-average calculations
2165 *
2166 * We take a distributed and async approach to calculating the global load-avg
2167 * in order to minimize overhead.
2168 *
2169 * The global load average is an exponentially decaying average of nr_running +
2170 * nr_uninterruptible.
2171 *
2172 * Once every LOAD_FREQ:
2173 *
2174 * nr_active = 0;
2175 * for_each_possible_cpu(cpu)
2176 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2177 *
2178 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2179 *
2180 * Due to a number of reasons the above turns in the mess below:
2181 *
2182 * - for_each_possible_cpu() is prohibitively expensive on machines with
2183 * serious number of cpus, therefore we need to take a distributed approach
2184 * to calculating nr_active.
2185 *
2186 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2187 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2188 *
2189 * So assuming nr_active := 0 when we start out -- true per definition, we
2190 * can simply take per-cpu deltas and fold those into a global accumulate
2191 * to obtain the same result. See calc_load_fold_active().
2192 *
2193 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2194 * across the machine, we assume 10 ticks is sufficient time for every
2195 * cpu to have completed this task.
2196 *
2197 * This places an upper-bound on the IRQ-off latency of the machine. Then
2198 * again, being late doesn't loose the delta, just wrecks the sample.
2199 *
2200 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2201 * this would add another cross-cpu cacheline miss and atomic operation
2202 * to the wakeup path. Instead we increment on whatever cpu the task ran
2203 * when it went into uninterruptible state and decrement on whatever cpu
2204 * did the wakeup. This means that only the sum of nr_uninterruptible over
2205 * all cpus yields the correct result.
2206 *
2207 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2208 */
2209
2210/* Variables and functions for calc_load */
2211static atomic_long_t calc_load_tasks;
2212static unsigned long calc_load_update;
2213unsigned long avenrun[3];
2214EXPORT_SYMBOL(avenrun); /* should be removed */
2215
2216/**
2217 * get_avenrun - get the load average array
2218 * @loads: pointer to dest load array
2219 * @offset: offset to add
2220 * @shift: shift count to shift the result left
2221 *
2222 * These values are estimates at best, so no need for locking.
2223 */
2224void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2225{
2226 loads[0] = (avenrun[0] + offset) << shift;
2227 loads[1] = (avenrun[1] + offset) << shift;
2228 loads[2] = (avenrun[2] + offset) << shift;
2229}
2230
2231static long calc_load_fold_active(struct rq *this_rq)
2232{
2233 long nr_active, delta = 0;
2234
2235 nr_active = this_rq->nr_running;
2236 nr_active += (long) this_rq->nr_uninterruptible;
2237
2238 if (nr_active != this_rq->calc_load_active) {
2239 delta = nr_active - this_rq->calc_load_active;
2240 this_rq->calc_load_active = nr_active;
2241 }
2242
2243 return delta;
2244}
2245
2246/*
2247 * a1 = a0 * e + a * (1 - e)
2248 */
2249static unsigned long
2250calc_load(unsigned long load, unsigned long exp, unsigned long active)
2251{
2252 load *= exp;
2253 load += active * (FIXED_1 - exp);
2254 load += 1UL << (FSHIFT - 1);
2255 return load >> FSHIFT;
2256}
2257
2258#ifdef CONFIG_NO_HZ
2259/*
2260 * Handle NO_HZ for the global load-average.
2261 *
2262 * Since the above described distributed algorithm to compute the global
2263 * load-average relies on per-cpu sampling from the tick, it is affected by
2264 * NO_HZ.
2265 *
2266 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2267 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2268 * when we read the global state.
2269 *
2270 * Obviously reality has to ruin such a delightfully simple scheme:
2271 *
2272 * - When we go NO_HZ idle during the window, we can negate our sample
2273 * contribution, causing under-accounting.
2274 *
2275 * We avoid this by keeping two idle-delta counters and flipping them
2276 * when the window starts, thus separating old and new NO_HZ load.
2277 *
2278 * The only trick is the slight shift in index flip for read vs write.
2279 *
2280 * 0s 5s 10s 15s
2281 * +10 +10 +10 +10
2282 * |-|-----------|-|-----------|-|-----------|-|
2283 * r:0 0 1 1 0 0 1 1 0
2284 * w:0 1 1 0 0 1 1 0 0
2285 *
2286 * This ensures we'll fold the old idle contribution in this window while
2287 * accumlating the new one.
2288 *
2289 * - When we wake up from NO_HZ idle during the window, we push up our
2290 * contribution, since we effectively move our sample point to a known
2291 * busy state.
2292 *
2293 * This is solved by pushing the window forward, and thus skipping the
2294 * sample, for this cpu (effectively using the idle-delta for this cpu which
2295 * was in effect at the time the window opened). This also solves the issue
2296 * of having to deal with a cpu having been in NOHZ idle for multiple
2297 * LOAD_FREQ intervals.
2298 *
2299 * When making the ILB scale, we should try to pull this in as well.
2300 */
2301static atomic_long_t calc_load_idle[2];
2302static int calc_load_idx;
2303
2304static inline int calc_load_write_idx(void)
2305{
2306 int idx = calc_load_idx;
2307
2308 /*
2309 * See calc_global_nohz(), if we observe the new index, we also
2310 * need to observe the new update time.
2311 */
2312 smp_rmb();
2313
2314 /*
2315 * If the folding window started, make sure we start writing in the
2316 * next idle-delta.
2317 */
2318 if (!time_before(jiffies, calc_load_update))
2319 idx++;
2320
2321 return idx & 1;
2322}
2323
2324static inline int calc_load_read_idx(void)
2325{
2326 return calc_load_idx & 1;
2327}
2328
2329void calc_load_enter_idle(void)
2330{
2331 struct rq *this_rq = this_rq();
2332 long delta;
2333
2334 /*
2335 * We're going into NOHZ mode, if there's any pending delta, fold it
2336 * into the pending idle delta.
2337 */
2338 delta = calc_load_fold_active(this_rq);
2339 if (delta) {
2340 int idx = calc_load_write_idx();
2341 atomic_long_add(delta, &calc_load_idle[idx]);
2342 }
2343}
2344
2345void calc_load_exit_idle(void)
2346{
2347 struct rq *this_rq = this_rq();
2348
2349 /*
2350 * If we're still before the sample window, we're done.
2351 */
2352 if (time_before(jiffies, this_rq->calc_load_update))
2353 return;
2354
2355 /*
2356 * We woke inside or after the sample window, this means we're already
2357 * accounted through the nohz accounting, so skip the entire deal and
2358 * sync up for the next window.
2359 */
2360 this_rq->calc_load_update = calc_load_update;
2361 if (time_before(jiffies, this_rq->calc_load_update + 10))
2362 this_rq->calc_load_update += LOAD_FREQ;
2363}
2364
2365static long calc_load_fold_idle(void)
2366{
2367 int idx = calc_load_read_idx();
2368 long delta = 0;
2369
2370 if (atomic_long_read(&calc_load_idle[idx]))
2371 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2372
2373 return delta;
2374}
2375
2376/**
2377 * fixed_power_int - compute: x^n, in O(log n) time
2378 *
2379 * @x: base of the power
2380 * @frac_bits: fractional bits of @x
2381 * @n: power to raise @x to.
2382 *
2383 * By exploiting the relation between the definition of the natural power
2384 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2385 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2386 * (where: n_i \elem {0, 1}, the binary vector representing n),
2387 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2388 * of course trivially computable in O(log_2 n), the length of our binary
2389 * vector.
2390 */
2391static unsigned long
2392fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2393{
2394 unsigned long result = 1UL << frac_bits;
2395
2396 if (n) for (;;) {
2397 if (n & 1) {
2398 result *= x;
2399 result += 1UL << (frac_bits - 1);
2400 result >>= frac_bits;
2401 }
2402 n >>= 1;
2403 if (!n)
2404 break;
2405 x *= x;
2406 x += 1UL << (frac_bits - 1);
2407 x >>= frac_bits;
2408 }
2409
2410 return result;
2411}
2412
2413/*
2414 * a1 = a0 * e + a * (1 - e)
2415 *
2416 * a2 = a1 * e + a * (1 - e)
2417 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2418 * = a0 * e^2 + a * (1 - e) * (1 + e)
2419 *
2420 * a3 = a2 * e + a * (1 - e)
2421 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2422 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2423 *
2424 * ...
2425 *
2426 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2427 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2428 * = a0 * e^n + a * (1 - e^n)
2429 *
2430 * [1] application of the geometric series:
2431 *
2432 * n 1 - x^(n+1)
2433 * S_n := \Sum x^i = -------------
2434 * i=0 1 - x
2435 */
2436static unsigned long
2437calc_load_n(unsigned long load, unsigned long exp,
2438 unsigned long active, unsigned int n)
2439{
2440
2441 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2442}
2443
2444/*
2445 * NO_HZ can leave us missing all per-cpu ticks calling
2446 * calc_load_account_active(), but since an idle CPU folds its delta into
2447 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2448 * in the pending idle delta if our idle period crossed a load cycle boundary.
2449 *
2450 * Once we've updated the global active value, we need to apply the exponential
2451 * weights adjusted to the number of cycles missed.
2452 */
2453static void calc_global_nohz(void)
2454{
2455 long delta, active, n;
2456
2457 if (!time_before(jiffies, calc_load_update + 10)) {
2458 /*
2459 * Catch-up, fold however many we are behind still
2460 */
2461 delta = jiffies - calc_load_update - 10;
2462 n = 1 + (delta / LOAD_FREQ);
2463
2464 active = atomic_long_read(&calc_load_tasks);
2465 active = active > 0 ? active * FIXED_1 : 0;
2466
2467 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2468 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2469 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2470
2471 calc_load_update += n * LOAD_FREQ;
2472 }
2473
2474 /*
2475 * Flip the idle index...
2476 *
2477 * Make sure we first write the new time then flip the index, so that
2478 * calc_load_write_idx() will see the new time when it reads the new
2479 * index, this avoids a double flip messing things up.
2480 */
2481 smp_wmb();
2482 calc_load_idx++;
2483}
2484#else /* !CONFIG_NO_HZ */
2485
2486static inline long calc_load_fold_idle(void) { return 0; }
2487static inline void calc_global_nohz(void) { }
2488
2489#endif /* CONFIG_NO_HZ */
2490
2491/*
2492 * calc_load - update the avenrun load estimates 10 ticks after the
2493 * CPUs have updated calc_load_tasks.
2494 */
2495void calc_global_load(unsigned long ticks)
2496{
2497 long active, delta;
2498
2499 if (time_before(jiffies, calc_load_update + 10))
2500 return;
2501
2502 /*
2503 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2504 */
2505 delta = calc_load_fold_idle();
2506 if (delta)
2507 atomic_long_add(delta, &calc_load_tasks);
2508
2509 active = atomic_long_read(&calc_load_tasks);
2510 active = active > 0 ? active * FIXED_1 : 0;
2511
2512 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2513 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2514 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2515
2516 calc_load_update += LOAD_FREQ;
2517
2518 /*
2519 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2520 */
2521 calc_global_nohz();
2522}
2523
2524/*
2525 * Called from update_cpu_load() to periodically update this CPU's
2526 * active count.
2527 */
2528static void calc_load_account_active(struct rq *this_rq)
2529{
2530 long delta;
2531
2532 if (time_before(jiffies, this_rq->calc_load_update))
2533 return;
2534
2535 delta = calc_load_fold_active(this_rq);
2536 if (delta)
2537 atomic_long_add(delta, &calc_load_tasks);
2538
2539 this_rq->calc_load_update += LOAD_FREQ;
2540}
2541
2542/*
2543 * End of global load-average stuff
2544 */
2545
2546/*
2547 * The exact cpuload at various idx values, calculated at every tick would be
2548 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2549 *
2550 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2551 * on nth tick when cpu may be busy, then we have:
2552 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2553 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2554 *
2555 * decay_load_missed() below does efficient calculation of
2556 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2557 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2558 *
2559 * The calculation is approximated on a 128 point scale.
2560 * degrade_zero_ticks is the number of ticks after which load at any
2561 * particular idx is approximated to be zero.
2562 * degrade_factor is a precomputed table, a row for each load idx.
2563 * Each column corresponds to degradation factor for a power of two ticks,
2564 * based on 128 point scale.
2565 * Example:
2566 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2567 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2568 *
2569 * With this power of 2 load factors, we can degrade the load n times
2570 * by looking at 1 bits in n and doing as many mult/shift instead of
2571 * n mult/shifts needed by the exact degradation.
2572 */
2573#define DEGRADE_SHIFT 7
2574static const unsigned char
2575 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2576static const unsigned char
2577 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2578 {0, 0, 0, 0, 0, 0, 0, 0},
2579 {64, 32, 8, 0, 0, 0, 0, 0},
2580 {96, 72, 40, 12, 1, 0, 0},
2581 {112, 98, 75, 43, 15, 1, 0},
2582 {120, 112, 98, 76, 45, 16, 2} };
2583
2584/*
2585 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2586 * would be when CPU is idle and so we just decay the old load without
2587 * adding any new load.
2588 */
2589static unsigned long
2590decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2591{
2592 int j = 0;
2593
2594 if (!missed_updates)
2595 return load;
2596
2597 if (missed_updates >= degrade_zero_ticks[idx])
2598 return 0;
2599
2600 if (idx == 1)
2601 return load >> missed_updates;
2602
2603 while (missed_updates) {
2604 if (missed_updates % 2)
2605 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2606
2607 missed_updates >>= 1;
2608 j++;
2609 }
2610 return load;
2611}
2612
2613/*
2614 * Update rq->cpu_load[] statistics. This function is usually called every
2615 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2616 * every tick. We fix it up based on jiffies.
2617 */
2618static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2619 unsigned long pending_updates)
2620{
2621 int i, scale;
2622
2623 this_rq->nr_load_updates++;
2624
2625 /* Update our load: */
2626 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2627 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2628 unsigned long old_load, new_load;
2629
2630 /* scale is effectively 1 << i now, and >> i divides by scale */
2631
2632 old_load = this_rq->cpu_load[i];
2633 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2634 new_load = this_load;
2635 /*
2636 * Round up the averaging division if load is increasing. This
2637 * prevents us from getting stuck on 9 if the load is 10, for
2638 * example.
2639 */
2640 if (new_load > old_load)
2641 new_load += scale - 1;
2642
2643 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2644 }
2645
2646 sched_avg_update(this_rq);
2647}
2648
2649#ifdef CONFIG_NO_HZ
2650/*
2651 * There is no sane way to deal with nohz on smp when using jiffies because the
2652 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2653 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2654 *
2655 * Therefore we cannot use the delta approach from the regular tick since that
2656 * would seriously skew the load calculation. However we'll make do for those
2657 * updates happening while idle (nohz_idle_balance) or coming out of idle
2658 * (tick_nohz_idle_exit).
2659 *
2660 * This means we might still be one tick off for nohz periods.
2661 */
2662
2663/*
2664 * Called from nohz_idle_balance() to update the load ratings before doing the
2665 * idle balance.
2666 */
2667void update_idle_cpu_load(struct rq *this_rq)
2668{
2669 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2670 unsigned long load = this_rq->load.weight;
2671 unsigned long pending_updates;
2672
2673 /*
2674 * bail if there's load or we're actually up-to-date.
2675 */
2676 if (load || curr_jiffies == this_rq->last_load_update_tick)
2677 return;
2678
2679 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2680 this_rq->last_load_update_tick = curr_jiffies;
2681
2682 __update_cpu_load(this_rq, load, pending_updates);
2683}
2684
2685/*
2686 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2687 */
2688void update_cpu_load_nohz(void)
2689{
2690 struct rq *this_rq = this_rq();
2691 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2692 unsigned long pending_updates;
2693
2694 if (curr_jiffies == this_rq->last_load_update_tick)
2695 return;
2696
2697 raw_spin_lock(&this_rq->lock);
2698 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2699 if (pending_updates) {
2700 this_rq->last_load_update_tick = curr_jiffies;
2701 /*
2702 * We were idle, this means load 0, the current load might be
2703 * !0 due to remote wakeups and the sort.
2704 */
2705 __update_cpu_load(this_rq, 0, pending_updates);
2706 }
2707 raw_spin_unlock(&this_rq->lock);
2708}
2709#endif /* CONFIG_NO_HZ */
2710
2711/*
2712 * Called from scheduler_tick()
2713 */
2714static void update_cpu_load_active(struct rq *this_rq)
2715{
2716 /*
2717 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2718 */
2719 this_rq->last_load_update_tick = jiffies;
2720 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2721
2722 calc_load_account_active(this_rq);
2723}
2724
2725#ifdef CONFIG_SMP
2726
2727/*
2728 * sched_exec - execve() is a valuable balancing opportunity, because at
2729 * this point the task has the smallest effective memory and cache footprint.
2730 */
2731void sched_exec(void)
2732{
2733 struct task_struct *p = current;
2734 unsigned long flags;
2735 int dest_cpu;
2736
2737 raw_spin_lock_irqsave(&p->pi_lock, flags);
2738 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2739 if (dest_cpu == smp_processor_id())
2740 goto unlock;
2741
2742 if (likely(cpu_active(dest_cpu))) {
2743 struct migration_arg arg = { p, dest_cpu };
2744
2745 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2746 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2747 return;
2748 }
2749unlock:
2750 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2751}
2752
2753#endif
2754
2755DEFINE_PER_CPU(struct kernel_stat, kstat);
2756DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2757
2758EXPORT_PER_CPU_SYMBOL(kstat);
2759EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2760
2761/*
2762 * Return any ns on the sched_clock that have not yet been accounted in
2763 * @p in case that task is currently running.
2764 *
2765 * Called with task_rq_lock() held on @rq.
2766 */
2767static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2768{
2769 u64 ns = 0;
2770
2771 if (task_current(rq, p)) {
2772 update_rq_clock(rq);
2773 ns = rq->clock_task - p->se.exec_start;
2774 if ((s64)ns < 0)
2775 ns = 0;
2776 }
2777
2778 return ns;
2779}
2780
2781unsigned long long task_delta_exec(struct task_struct *p)
2782{
2783 unsigned long flags;
2784 struct rq *rq;
2785 u64 ns = 0;
2786
2787 rq = task_rq_lock(p, &flags);
2788 ns = do_task_delta_exec(p, rq);
2789 task_rq_unlock(rq, p, &flags);
2790
2791 return ns;
2792}
2793
2794/*
2795 * Return accounted runtime for the task.
2796 * In case the task is currently running, return the runtime plus current's
2797 * pending runtime that have not been accounted yet.
2798 */
2799unsigned long long task_sched_runtime(struct task_struct *p)
2800{
2801 unsigned long flags;
2802 struct rq *rq;
2803 u64 ns = 0;
2804
2805 rq = task_rq_lock(p, &flags);
2806 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2807 task_rq_unlock(rq, p, &flags);
2808
2809 return ns;
2810}
2811
2812#ifdef CONFIG_CGROUP_CPUACCT
2813struct cgroup_subsys cpuacct_subsys;
2814struct cpuacct root_cpuacct;
2815#endif
2816
2817static inline void task_group_account_field(struct task_struct *p, int index,
2818 u64 tmp)
2819{
2820#ifdef CONFIG_CGROUP_CPUACCT
2821 struct kernel_cpustat *kcpustat;
2822 struct cpuacct *ca;
2823#endif
2824 /*
2825 * Since all updates are sure to touch the root cgroup, we
2826 * get ourselves ahead and touch it first. If the root cgroup
2827 * is the only cgroup, then nothing else should be necessary.
2828 *
2829 */
2830 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
2831
2832#ifdef CONFIG_CGROUP_CPUACCT
2833 if (unlikely(!cpuacct_subsys.active))
2834 return;
2835
2836 rcu_read_lock();
2837 ca = task_ca(p);
2838 while (ca && (ca != &root_cpuacct)) {
2839 kcpustat = this_cpu_ptr(ca->cpustat);
2840 kcpustat->cpustat[index] += tmp;
2841 ca = parent_ca(ca);
2842 }
2843 rcu_read_unlock();
2844#endif
2845}
2846
2847
2848/*
2849 * Account user cpu time to a process.
2850 * @p: the process that the cpu time gets accounted to
2851 * @cputime: the cpu time spent in user space since the last update
2852 * @cputime_scaled: cputime scaled by cpu frequency
2853 */
2854void account_user_time(struct task_struct *p, cputime_t cputime,
2855 cputime_t cputime_scaled)
2856{
2857 int index;
2858
2859 /* Add user time to process. */
2860 p->utime += cputime;
2861 p->utimescaled += cputime_scaled;
2862 account_group_user_time(p, cputime);
2863
2864 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
2865
2866 /* Add user time to cpustat. */
2867 task_group_account_field(p, index, (__force u64) cputime);
2868
2869 /* Account for user time used */
2870 acct_update_integrals(p);
2871}
2872
2873/*
2874 * Account guest cpu time to a process.
2875 * @p: the process that the cpu time gets accounted to
2876 * @cputime: the cpu time spent in virtual machine since the last update
2877 * @cputime_scaled: cputime scaled by cpu frequency
2878 */
2879static void account_guest_time(struct task_struct *p, cputime_t cputime,
2880 cputime_t cputime_scaled)
2881{
2882 u64 *cpustat = kcpustat_this_cpu->cpustat;
2883
2884 /* Add guest time to process. */
2885 p->utime += cputime;
2886 p->utimescaled += cputime_scaled;
2887 account_group_user_time(p, cputime);
2888 p->gtime += cputime;
2889
2890 /* Add guest time to cpustat. */
2891 if (TASK_NICE(p) > 0) {
2892 cpustat[CPUTIME_NICE] += (__force u64) cputime;
2893 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
2894 } else {
2895 cpustat[CPUTIME_USER] += (__force u64) cputime;
2896 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
2897 }
2898}
2899
2900/*
2901 * Account system cpu time to a process and desired cpustat field
2902 * @p: the process that the cpu time gets accounted to
2903 * @cputime: the cpu time spent in kernel space since the last update
2904 * @cputime_scaled: cputime scaled by cpu frequency
2905 * @target_cputime64: pointer to cpustat field that has to be updated
2906 */
2907static inline
2908void __account_system_time(struct task_struct *p, cputime_t cputime,
2909 cputime_t cputime_scaled, int index)
2910{
2911 /* Add system time to process. */
2912 p->stime += cputime;
2913 p->stimescaled += cputime_scaled;
2914 account_group_system_time(p, cputime);
2915
2916 /* Add system time to cpustat. */
2917 task_group_account_field(p, index, (__force u64) cputime);
2918
2919 /* Account for system time used */
2920 acct_update_integrals(p);
2921}
2922
2923/*
2924 * Account system cpu time to a process.
2925 * @p: the process that the cpu time gets accounted to
2926 * @hardirq_offset: the offset to subtract from hardirq_count()
2927 * @cputime: the cpu time spent in kernel space since the last update
2928 * @cputime_scaled: cputime scaled by cpu frequency
2929 */
2930void account_system_time(struct task_struct *p, int hardirq_offset,
2931 cputime_t cputime, cputime_t cputime_scaled)
2932{
2933 int index;
2934
2935 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2936 account_guest_time(p, cputime, cputime_scaled);
2937 return;
2938 }
2939
2940 if (hardirq_count() - hardirq_offset)
2941 index = CPUTIME_IRQ;
2942 else if (in_serving_softirq())
2943 index = CPUTIME_SOFTIRQ;
2944 else
2945 index = CPUTIME_SYSTEM;
2946
2947 __account_system_time(p, cputime, cputime_scaled, index);
2948}
2949
2950/*
2951 * Account for involuntary wait time.
2952 * @cputime: the cpu time spent in involuntary wait
2953 */
2954void account_steal_time(cputime_t cputime)
2955{
2956 u64 *cpustat = kcpustat_this_cpu->cpustat;
2957
2958 cpustat[CPUTIME_STEAL] += (__force u64) cputime;
2959}
2960
2961/*
2962 * Account for idle time.
2963 * @cputime: the cpu time spent in idle wait
2964 */
2965void account_idle_time(cputime_t cputime)
2966{
2967 u64 *cpustat = kcpustat_this_cpu->cpustat;
2968 struct rq *rq = this_rq();
2969
2970 if (atomic_read(&rq->nr_iowait) > 0)
2971 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
2972 else
2973 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
2974}
2975
2976static __always_inline bool steal_account_process_tick(void)
2977{
2978#ifdef CONFIG_PARAVIRT
2979 if (static_key_false(¶virt_steal_enabled)) {
2980 u64 steal, st = 0;
2981
2982 steal = paravirt_steal_clock(smp_processor_id());
2983 steal -= this_rq()->prev_steal_time;
2984
2985 st = steal_ticks(steal);
2986 this_rq()->prev_steal_time += st * TICK_NSEC;
2987
2988 account_steal_time(st);
2989 return st;
2990 }
2991#endif
2992 return false;
2993}
2994
2995#ifndef CONFIG_VIRT_CPU_ACCOUNTING
2996
2997#ifdef CONFIG_IRQ_TIME_ACCOUNTING
2998/*
2999 * Account a tick to a process and cpustat
3000 * @p: the process that the cpu time gets accounted to
3001 * @user_tick: is the tick from userspace
3002 * @rq: the pointer to rq
3003 *
3004 * Tick demultiplexing follows the order
3005 * - pending hardirq update
3006 * - pending softirq update
3007 * - user_time
3008 * - idle_time
3009 * - system time
3010 * - check for guest_time
3011 * - else account as system_time
3012 *
3013 * Check for hardirq is done both for system and user time as there is
3014 * no timer going off while we are on hardirq and hence we may never get an
3015 * opportunity to update it solely in system time.
3016 * p->stime and friends are only updated on system time and not on irq
3017 * softirq as those do not count in task exec_runtime any more.
3018 */
3019static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3020 struct rq *rq)
3021{
3022 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3023 u64 *cpustat = kcpustat_this_cpu->cpustat;
3024
3025 if (steal_account_process_tick())
3026 return;
3027
3028 if (irqtime_account_hi_update()) {
3029 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
3030 } else if (irqtime_account_si_update()) {
3031 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
3032 } else if (this_cpu_ksoftirqd() == p) {
3033 /*
3034 * ksoftirqd time do not get accounted in cpu_softirq_time.
3035 * So, we have to handle it separately here.
3036 * Also, p->stime needs to be updated for ksoftirqd.
3037 */
3038 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3039 CPUTIME_SOFTIRQ);
3040 } else if (user_tick) {
3041 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3042 } else if (p == rq->idle) {
3043 account_idle_time(cputime_one_jiffy);
3044 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3045 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3046 } else {
3047 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3048 CPUTIME_SYSTEM);
3049 }
3050}
3051
3052static void irqtime_account_idle_ticks(int ticks)
3053{
3054 int i;
3055 struct rq *rq = this_rq();
3056
3057 for (i = 0; i < ticks; i++)
3058 irqtime_account_process_tick(current, 0, rq);
3059}
3060#else /* CONFIG_IRQ_TIME_ACCOUNTING */
3061static void irqtime_account_idle_ticks(int ticks) {}
3062static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3063 struct rq *rq) {}
3064#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3065
3066/*
3067 * Account a single tick of cpu time.
3068 * @p: the process that the cpu time gets accounted to
3069 * @user_tick: indicates if the tick is a user or a system tick
3070 */
3071void account_process_tick(struct task_struct *p, int user_tick)
3072{
3073 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3074 struct rq *rq = this_rq();
3075
3076 if (sched_clock_irqtime) {
3077 irqtime_account_process_tick(p, user_tick, rq);
3078 return;
3079 }
3080
3081 if (steal_account_process_tick())
3082 return;
3083
3084 if (user_tick)
3085 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3086 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3087 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3088 one_jiffy_scaled);
3089 else
3090 account_idle_time(cputime_one_jiffy);
3091}
3092
3093/*
3094 * Account multiple ticks of steal time.
3095 * @p: the process from which the cpu time has been stolen
3096 * @ticks: number of stolen ticks
3097 */
3098void account_steal_ticks(unsigned long ticks)
3099{
3100 account_steal_time(jiffies_to_cputime(ticks));
3101}
3102
3103/*
3104 * Account multiple ticks of idle time.
3105 * @ticks: number of stolen ticks
3106 */
3107void account_idle_ticks(unsigned long ticks)
3108{
3109
3110 if (sched_clock_irqtime) {
3111 irqtime_account_idle_ticks(ticks);
3112 return;
3113 }
3114
3115 account_idle_time(jiffies_to_cputime(ticks));
3116}
3117
3118#endif
3119
3120/*
3121 * Use precise platform statistics if available:
3122 */
3123#ifdef CONFIG_VIRT_CPU_ACCOUNTING
3124void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3125{
3126 *ut = p->utime;
3127 *st = p->stime;
3128}
3129
3130void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3131{
3132 struct task_cputime cputime;
3133
3134 thread_group_cputime(p, &cputime);
3135
3136 *ut = cputime.utime;
3137 *st = cputime.stime;
3138}
3139#else
3140
3141#ifndef nsecs_to_cputime
3142# define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3143#endif
3144
3145static cputime_t scale_utime(cputime_t utime, cputime_t rtime, cputime_t total)
3146{
3147 u64 temp = (__force u64) rtime;
3148
3149 temp *= (__force u64) utime;
3150
3151 if (sizeof(cputime_t) == 4)
3152 temp = div_u64(temp, (__force u32) total);
3153 else
3154 temp = div64_u64(temp, (__force u64) total);
3155
3156 return (__force cputime_t) temp;
3157}
3158
3159void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3160{
3161 cputime_t rtime, utime = p->utime, total = utime + p->stime;
3162
3163 /*
3164 * Use CFS's precise accounting:
3165 */
3166 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3167
3168 if (total)
3169 utime = scale_utime(utime, rtime, total);
3170 else
3171 utime = rtime;
3172
3173 /*
3174 * Compare with previous values, to keep monotonicity:
3175 */
3176 p->prev_utime = max(p->prev_utime, utime);
3177 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime);
3178
3179 *ut = p->prev_utime;
3180 *st = p->prev_stime;
3181}
3182
3183/*
3184 * Must be called with siglock held.
3185 */
3186void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3187{
3188 struct signal_struct *sig = p->signal;
3189 struct task_cputime cputime;
3190 cputime_t rtime, utime, total;
3191
3192 thread_group_cputime(p, &cputime);
3193
3194 total = cputime.utime + cputime.stime;
3195 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3196
3197 if (total)
3198 utime = scale_utime(cputime.utime, rtime, total);
3199 else
3200 utime = rtime;
3201
3202 sig->prev_utime = max(sig->prev_utime, utime);
3203 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime);
3204
3205 *ut = sig->prev_utime;
3206 *st = sig->prev_stime;
3207}
3208#endif
3209
3210/*
3211 * This function gets called by the timer code, with HZ frequency.
3212 * We call it with interrupts disabled.
3213 */
3214void scheduler_tick(void)
3215{
3216 int cpu = smp_processor_id();
3217 struct rq *rq = cpu_rq(cpu);
3218 struct task_struct *curr = rq->curr;
3219
3220 sched_clock_tick();
3221
3222 raw_spin_lock(&rq->lock);
3223 update_rq_clock(rq);
3224 update_cpu_load_active(rq);
3225 curr->sched_class->task_tick(rq, curr, 0);
3226 raw_spin_unlock(&rq->lock);
3227
3228 perf_event_task_tick();
3229
3230#ifdef CONFIG_SMP
3231 rq->idle_balance = idle_cpu(cpu);
3232 trigger_load_balance(rq, cpu);
3233#endif
3234}
3235
3236notrace unsigned long get_parent_ip(unsigned long addr)
3237{
3238 if (in_lock_functions(addr)) {
3239 addr = CALLER_ADDR2;
3240 if (in_lock_functions(addr))
3241 addr = CALLER_ADDR3;
3242 }
3243 return addr;
3244}
3245
3246#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3247 defined(CONFIG_PREEMPT_TRACER))
3248
3249void __kprobes add_preempt_count(int val)
3250{
3251#ifdef CONFIG_DEBUG_PREEMPT
3252 /*
3253 * Underflow?
3254 */
3255 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3256 return;
3257#endif
3258 preempt_count() += val;
3259#ifdef CONFIG_DEBUG_PREEMPT
3260 /*
3261 * Spinlock count overflowing soon?
3262 */
3263 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3264 PREEMPT_MASK - 10);
3265#endif
3266 if (preempt_count() == val)
3267 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3268}
3269EXPORT_SYMBOL(add_preempt_count);
3270
3271void __kprobes sub_preempt_count(int val)
3272{
3273#ifdef CONFIG_DEBUG_PREEMPT
3274 /*
3275 * Underflow?
3276 */
3277 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3278 return;
3279 /*
3280 * Is the spinlock portion underflowing?
3281 */
3282 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3283 !(preempt_count() & PREEMPT_MASK)))
3284 return;
3285#endif
3286
3287 if (preempt_count() == val)
3288 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3289 preempt_count() -= val;
3290}
3291EXPORT_SYMBOL(sub_preempt_count);
3292
3293#endif
3294
3295/*
3296 * Print scheduling while atomic bug:
3297 */
3298static noinline void __schedule_bug(struct task_struct *prev)
3299{
3300 if (oops_in_progress)
3301 return;
3302
3303 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3304 prev->comm, prev->pid, preempt_count());
3305
3306 debug_show_held_locks(prev);
3307 print_modules();
3308 if (irqs_disabled())
3309 print_irqtrace_events(prev);
3310 dump_stack();
3311 add_taint(TAINT_WARN);
3312}
3313
3314/*
3315 * Various schedule()-time debugging checks and statistics:
3316 */
3317static inline void schedule_debug(struct task_struct *prev)
3318{
3319 /*
3320 * Test if we are atomic. Since do_exit() needs to call into
3321 * schedule() atomically, we ignore that path for now.
3322 * Otherwise, whine if we are scheduling when we should not be.
3323 */
3324 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3325 __schedule_bug(prev);
3326 rcu_sleep_check();
3327
3328 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3329
3330 schedstat_inc(this_rq(), sched_count);
3331}
3332
3333static void put_prev_task(struct rq *rq, struct task_struct *prev)
3334{
3335 if (prev->on_rq || rq->skip_clock_update < 0)
3336 update_rq_clock(rq);
3337 prev->sched_class->put_prev_task(rq, prev);
3338}
3339
3340/*
3341 * Pick up the highest-prio task:
3342 */
3343static inline struct task_struct *
3344pick_next_task(struct rq *rq)
3345{
3346 const struct sched_class *class;
3347 struct task_struct *p;
3348
3349 /*
3350 * Optimization: we know that if all tasks are in
3351 * the fair class we can call that function directly:
3352 */
3353 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3354 p = fair_sched_class.pick_next_task(rq);
3355 if (likely(p))
3356 return p;
3357 }
3358
3359 for_each_class(class) {
3360 p = class->pick_next_task(rq);
3361 if (p)
3362 return p;
3363 }
3364
3365 BUG(); /* the idle class will always have a runnable task */
3366}
3367
3368/*
3369 * __schedule() is the main scheduler function.
3370 */
3371static void __sched __schedule(void)
3372{
3373 struct task_struct *prev, *next;
3374 unsigned long *switch_count;
3375 struct rq *rq;
3376 int cpu;
3377
3378need_resched:
3379 preempt_disable();
3380 cpu = smp_processor_id();
3381 rq = cpu_rq(cpu);
3382 rcu_note_context_switch(cpu);
3383 prev = rq->curr;
3384
3385 schedule_debug(prev);
3386
3387 if (sched_feat(HRTICK))
3388 hrtick_clear(rq);
3389
3390 raw_spin_lock_irq(&rq->lock);
3391
3392 switch_count = &prev->nivcsw;
3393 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3394 if (unlikely(signal_pending_state(prev->state, prev))) {
3395 prev->state = TASK_RUNNING;
3396 } else {
3397 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3398 prev->on_rq = 0;
3399
3400 /*
3401 * If a worker went to sleep, notify and ask workqueue
3402 * whether it wants to wake up a task to maintain
3403 * concurrency.
3404 */
3405 if (prev->flags & PF_WQ_WORKER) {
3406 struct task_struct *to_wakeup;
3407
3408 to_wakeup = wq_worker_sleeping(prev, cpu);
3409 if (to_wakeup)
3410 try_to_wake_up_local(to_wakeup);
3411 }
3412 }
3413 switch_count = &prev->nvcsw;
3414 }
3415
3416 pre_schedule(rq, prev);
3417
3418 if (unlikely(!rq->nr_running))
3419 idle_balance(cpu, rq);
3420
3421 put_prev_task(rq, prev);
3422 next = pick_next_task(rq);
3423 clear_tsk_need_resched(prev);
3424 rq->skip_clock_update = 0;
3425
3426 if (likely(prev != next)) {
3427 rq->nr_switches++;
3428 rq->curr = next;
3429 ++*switch_count;
3430
3431 context_switch(rq, prev, next); /* unlocks the rq */
3432 /*
3433 * The context switch have flipped the stack from under us
3434 * and restored the local variables which were saved when
3435 * this task called schedule() in the past. prev == current
3436 * is still correct, but it can be moved to another cpu/rq.
3437 */
3438 cpu = smp_processor_id();
3439 rq = cpu_rq(cpu);
3440 } else
3441 raw_spin_unlock_irq(&rq->lock);
3442
3443 post_schedule(rq);
3444
3445 sched_preempt_enable_no_resched();
3446 if (need_resched())
3447 goto need_resched;
3448}
3449
3450static inline void sched_submit_work(struct task_struct *tsk)
3451{
3452 if (!tsk->state || tsk_is_pi_blocked(tsk))
3453 return;
3454 /*
3455 * If we are going to sleep and we have plugged IO queued,
3456 * make sure to submit it to avoid deadlocks.
3457 */
3458 if (blk_needs_flush_plug(tsk))
3459 blk_schedule_flush_plug(tsk);
3460}
3461
3462asmlinkage void __sched schedule(void)
3463{
3464 struct task_struct *tsk = current;
3465
3466 sched_submit_work(tsk);
3467 __schedule();
3468}
3469EXPORT_SYMBOL(schedule);
3470
3471/**
3472 * schedule_preempt_disabled - called with preemption disabled
3473 *
3474 * Returns with preemption disabled. Note: preempt_count must be 1
3475 */
3476void __sched schedule_preempt_disabled(void)
3477{
3478 sched_preempt_enable_no_resched();
3479 schedule();
3480 preempt_disable();
3481}
3482
3483#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3484
3485static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3486{
3487 if (lock->owner != owner)
3488 return false;
3489
3490 /*
3491 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3492 * lock->owner still matches owner, if that fails, owner might
3493 * point to free()d memory, if it still matches, the rcu_read_lock()
3494 * ensures the memory stays valid.
3495 */
3496 barrier();
3497
3498 return owner->on_cpu;
3499}
3500
3501/*
3502 * Look out! "owner" is an entirely speculative pointer
3503 * access and not reliable.
3504 */
3505int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3506{
3507 if (!sched_feat(OWNER_SPIN))
3508 return 0;
3509
3510 rcu_read_lock();
3511 while (owner_running(lock, owner)) {
3512 if (need_resched())
3513 break;
3514
3515 arch_mutex_cpu_relax();
3516 }
3517 rcu_read_unlock();
3518
3519 /*
3520 * We break out the loop above on need_resched() and when the
3521 * owner changed, which is a sign for heavy contention. Return
3522 * success only when lock->owner is NULL.
3523 */
3524 return lock->owner == NULL;
3525}
3526#endif
3527
3528#ifdef CONFIG_PREEMPT
3529/*
3530 * this is the entry point to schedule() from in-kernel preemption
3531 * off of preempt_enable. Kernel preemptions off return from interrupt
3532 * occur there and call schedule directly.
3533 */
3534asmlinkage void __sched notrace preempt_schedule(void)
3535{
3536 struct thread_info *ti = current_thread_info();
3537
3538 /*
3539 * If there is a non-zero preempt_count or interrupts are disabled,
3540 * we do not want to preempt the current task. Just return..
3541 */
3542 if (likely(ti->preempt_count || irqs_disabled()))
3543 return;
3544
3545 do {
3546 add_preempt_count_notrace(PREEMPT_ACTIVE);
3547 __schedule();
3548 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3549
3550 /*
3551 * Check again in case we missed a preemption opportunity
3552 * between schedule and now.
3553 */
3554 barrier();
3555 } while (need_resched());
3556}
3557EXPORT_SYMBOL(preempt_schedule);
3558
3559/*
3560 * this is the entry point to schedule() from kernel preemption
3561 * off of irq context.
3562 * Note, that this is called and return with irqs disabled. This will
3563 * protect us against recursive calling from irq.
3564 */
3565asmlinkage void __sched preempt_schedule_irq(void)
3566{
3567 struct thread_info *ti = current_thread_info();
3568
3569 /* Catch callers which need to be fixed */
3570 BUG_ON(ti->preempt_count || !irqs_disabled());
3571
3572 do {
3573 add_preempt_count(PREEMPT_ACTIVE);
3574 local_irq_enable();
3575 __schedule();
3576 local_irq_disable();
3577 sub_preempt_count(PREEMPT_ACTIVE);
3578
3579 /*
3580 * Check again in case we missed a preemption opportunity
3581 * between schedule and now.
3582 */
3583 barrier();
3584 } while (need_resched());
3585}
3586
3587#endif /* CONFIG_PREEMPT */
3588
3589int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3590 void *key)
3591{
3592 return try_to_wake_up(curr->private, mode, wake_flags);
3593}
3594EXPORT_SYMBOL(default_wake_function);
3595
3596/*
3597 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3598 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3599 * number) then we wake all the non-exclusive tasks and one exclusive task.
3600 *
3601 * There are circumstances in which we can try to wake a task which has already
3602 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3603 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3604 */
3605static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3606 int nr_exclusive, int wake_flags, void *key)
3607{
3608 wait_queue_t *curr, *next;
3609
3610 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3611 unsigned flags = curr->flags;
3612
3613 if (curr->func(curr, mode, wake_flags, key) &&
3614 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3615 break;
3616 }
3617}
3618
3619/**
3620 * __wake_up - wake up threads blocked on a waitqueue.
3621 * @q: the waitqueue
3622 * @mode: which threads
3623 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3624 * @key: is directly passed to the wakeup function
3625 *
3626 * It may be assumed that this function implies a write memory barrier before
3627 * changing the task state if and only if any tasks are woken up.
3628 */
3629void __wake_up(wait_queue_head_t *q, unsigned int mode,
3630 int nr_exclusive, void *key)
3631{
3632 unsigned long flags;
3633
3634 spin_lock_irqsave(&q->lock, flags);
3635 __wake_up_common(q, mode, nr_exclusive, 0, key);
3636 spin_unlock_irqrestore(&q->lock, flags);
3637}
3638EXPORT_SYMBOL(__wake_up);
3639
3640/*
3641 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3642 */
3643void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3644{
3645 __wake_up_common(q, mode, nr, 0, NULL);
3646}
3647EXPORT_SYMBOL_GPL(__wake_up_locked);
3648
3649void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3650{
3651 __wake_up_common(q, mode, 1, 0, key);
3652}
3653EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3654
3655/**
3656 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3657 * @q: the waitqueue
3658 * @mode: which threads
3659 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3660 * @key: opaque value to be passed to wakeup targets
3661 *
3662 * The sync wakeup differs that the waker knows that it will schedule
3663 * away soon, so while the target thread will be woken up, it will not
3664 * be migrated to another CPU - ie. the two threads are 'synchronized'
3665 * with each other. This can prevent needless bouncing between CPUs.
3666 *
3667 * On UP it can prevent extra preemption.
3668 *
3669 * It may be assumed that this function implies a write memory barrier before
3670 * changing the task state if and only if any tasks are woken up.
3671 */
3672void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3673 int nr_exclusive, void *key)
3674{
3675 unsigned long flags;
3676 int wake_flags = WF_SYNC;
3677
3678 if (unlikely(!q))
3679 return;
3680
3681 if (unlikely(!nr_exclusive))
3682 wake_flags = 0;
3683
3684 spin_lock_irqsave(&q->lock, flags);
3685 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3686 spin_unlock_irqrestore(&q->lock, flags);
3687}
3688EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3689
3690/*
3691 * __wake_up_sync - see __wake_up_sync_key()
3692 */
3693void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3694{
3695 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3696}
3697EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3698
3699/**
3700 * complete: - signals a single thread waiting on this completion
3701 * @x: holds the state of this particular completion
3702 *
3703 * This will wake up a single thread waiting on this completion. Threads will be
3704 * awakened in the same order in which they were queued.
3705 *
3706 * See also complete_all(), wait_for_completion() and related routines.
3707 *
3708 * It may be assumed that this function implies a write memory barrier before
3709 * changing the task state if and only if any tasks are woken up.
3710 */
3711void complete(struct completion *x)
3712{
3713 unsigned long flags;
3714
3715 spin_lock_irqsave(&x->wait.lock, flags);
3716 x->done++;
3717 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3718 spin_unlock_irqrestore(&x->wait.lock, flags);
3719}
3720EXPORT_SYMBOL(complete);
3721
3722/**
3723 * complete_all: - signals all threads waiting on this completion
3724 * @x: holds the state of this particular completion
3725 *
3726 * This will wake up all threads waiting on this particular completion event.
3727 *
3728 * It may be assumed that this function implies a write memory barrier before
3729 * changing the task state if and only if any tasks are woken up.
3730 */
3731void complete_all(struct completion *x)
3732{
3733 unsigned long flags;
3734
3735 spin_lock_irqsave(&x->wait.lock, flags);
3736 x->done += UINT_MAX/2;
3737 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3738 spin_unlock_irqrestore(&x->wait.lock, flags);
3739}
3740EXPORT_SYMBOL(complete_all);
3741
3742static inline long __sched
3743do_wait_for_common(struct completion *x, long timeout, int state)
3744{
3745 if (!x->done) {
3746 DECLARE_WAITQUEUE(wait, current);
3747
3748 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3749 do {
3750 if (signal_pending_state(state, current)) {
3751 timeout = -ERESTARTSYS;
3752 break;
3753 }
3754 __set_current_state(state);
3755 spin_unlock_irq(&x->wait.lock);
3756 timeout = schedule_timeout(timeout);
3757 spin_lock_irq(&x->wait.lock);
3758 } while (!x->done && timeout);
3759 __remove_wait_queue(&x->wait, &wait);
3760 if (!x->done)
3761 return timeout;
3762 }
3763 x->done--;
3764 return timeout ?: 1;
3765}
3766
3767static long __sched
3768wait_for_common(struct completion *x, long timeout, int state)
3769{
3770 might_sleep();
3771
3772 spin_lock_irq(&x->wait.lock);
3773 timeout = do_wait_for_common(x, timeout, state);
3774 spin_unlock_irq(&x->wait.lock);
3775 return timeout;
3776}
3777
3778/**
3779 * wait_for_completion: - waits for completion of a task
3780 * @x: holds the state of this particular completion
3781 *
3782 * This waits to be signaled for completion of a specific task. It is NOT
3783 * interruptible and there is no timeout.
3784 *
3785 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3786 * and interrupt capability. Also see complete().
3787 */
3788void __sched wait_for_completion(struct completion *x)
3789{
3790 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3791}
3792EXPORT_SYMBOL(wait_for_completion);
3793
3794/**
3795 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3796 * @x: holds the state of this particular completion
3797 * @timeout: timeout value in jiffies
3798 *
3799 * This waits for either a completion of a specific task to be signaled or for a
3800 * specified timeout to expire. The timeout is in jiffies. It is not
3801 * interruptible.
3802 *
3803 * The return value is 0 if timed out, and positive (at least 1, or number of
3804 * jiffies left till timeout) if completed.
3805 */
3806unsigned long __sched
3807wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3808{
3809 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3810}
3811EXPORT_SYMBOL(wait_for_completion_timeout);
3812
3813/**
3814 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3815 * @x: holds the state of this particular completion
3816 *
3817 * This waits for completion of a specific task to be signaled. It is
3818 * interruptible.
3819 *
3820 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3821 */
3822int __sched wait_for_completion_interruptible(struct completion *x)
3823{
3824 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3825 if (t == -ERESTARTSYS)
3826 return t;
3827 return 0;
3828}
3829EXPORT_SYMBOL(wait_for_completion_interruptible);
3830
3831/**
3832 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3833 * @x: holds the state of this particular completion
3834 * @timeout: timeout value in jiffies
3835 *
3836 * This waits for either a completion of a specific task to be signaled or for a
3837 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3838 *
3839 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3840 * positive (at least 1, or number of jiffies left till timeout) if completed.
3841 */
3842long __sched
3843wait_for_completion_interruptible_timeout(struct completion *x,
3844 unsigned long timeout)
3845{
3846 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3847}
3848EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3849
3850/**
3851 * wait_for_completion_killable: - waits for completion of a task (killable)
3852 * @x: holds the state of this particular completion
3853 *
3854 * This waits to be signaled for completion of a specific task. It can be
3855 * interrupted by a kill signal.
3856 *
3857 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3858 */
3859int __sched wait_for_completion_killable(struct completion *x)
3860{
3861 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3862 if (t == -ERESTARTSYS)
3863 return t;
3864 return 0;
3865}
3866EXPORT_SYMBOL(wait_for_completion_killable);
3867
3868/**
3869 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3870 * @x: holds the state of this particular completion
3871 * @timeout: timeout value in jiffies
3872 *
3873 * This waits for either a completion of a specific task to be
3874 * signaled or for a specified timeout to expire. It can be
3875 * interrupted by a kill signal. The timeout is in jiffies.
3876 *
3877 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3878 * positive (at least 1, or number of jiffies left till timeout) if completed.
3879 */
3880long __sched
3881wait_for_completion_killable_timeout(struct completion *x,
3882 unsigned long timeout)
3883{
3884 return wait_for_common(x, timeout, TASK_KILLABLE);
3885}
3886EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3887
3888/**
3889 * try_wait_for_completion - try to decrement a completion without blocking
3890 * @x: completion structure
3891 *
3892 * Returns: 0 if a decrement cannot be done without blocking
3893 * 1 if a decrement succeeded.
3894 *
3895 * If a completion is being used as a counting completion,
3896 * attempt to decrement the counter without blocking. This
3897 * enables us to avoid waiting if the resource the completion
3898 * is protecting is not available.
3899 */
3900bool try_wait_for_completion(struct completion *x)
3901{
3902 unsigned long flags;
3903 int ret = 1;
3904
3905 spin_lock_irqsave(&x->wait.lock, flags);
3906 if (!x->done)
3907 ret = 0;
3908 else
3909 x->done--;
3910 spin_unlock_irqrestore(&x->wait.lock, flags);
3911 return ret;
3912}
3913EXPORT_SYMBOL(try_wait_for_completion);
3914
3915/**
3916 * completion_done - Test to see if a completion has any waiters
3917 * @x: completion structure
3918 *
3919 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3920 * 1 if there are no waiters.
3921 *
3922 */
3923bool completion_done(struct completion *x)
3924{
3925 unsigned long flags;
3926 int ret = 1;
3927
3928 spin_lock_irqsave(&x->wait.lock, flags);
3929 if (!x->done)
3930 ret = 0;
3931 spin_unlock_irqrestore(&x->wait.lock, flags);
3932 return ret;
3933}
3934EXPORT_SYMBOL(completion_done);
3935
3936static long __sched
3937sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3938{
3939 unsigned long flags;
3940 wait_queue_t wait;
3941
3942 init_waitqueue_entry(&wait, current);
3943
3944 __set_current_state(state);
3945
3946 spin_lock_irqsave(&q->lock, flags);
3947 __add_wait_queue(q, &wait);
3948 spin_unlock(&q->lock);
3949 timeout = schedule_timeout(timeout);
3950 spin_lock_irq(&q->lock);
3951 __remove_wait_queue(q, &wait);
3952 spin_unlock_irqrestore(&q->lock, flags);
3953
3954 return timeout;
3955}
3956
3957void __sched interruptible_sleep_on(wait_queue_head_t *q)
3958{
3959 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3960}
3961EXPORT_SYMBOL(interruptible_sleep_on);
3962
3963long __sched
3964interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3965{
3966 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3967}
3968EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3969
3970void __sched sleep_on(wait_queue_head_t *q)
3971{
3972 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3973}
3974EXPORT_SYMBOL(sleep_on);
3975
3976long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3977{
3978 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3979}
3980EXPORT_SYMBOL(sleep_on_timeout);
3981
3982#ifdef CONFIG_RT_MUTEXES
3983
3984/*
3985 * rt_mutex_setprio - set the current priority of a task
3986 * @p: task
3987 * @prio: prio value (kernel-internal form)
3988 *
3989 * This function changes the 'effective' priority of a task. It does
3990 * not touch ->normal_prio like __setscheduler().
3991 *
3992 * Used by the rt_mutex code to implement priority inheritance logic.
3993 */
3994void rt_mutex_setprio(struct task_struct *p, int prio)
3995{
3996 int oldprio, on_rq, running;
3997 struct rq *rq;
3998 const struct sched_class *prev_class;
3999
4000 BUG_ON(prio < 0 || prio > MAX_PRIO);
4001
4002 rq = __task_rq_lock(p);
4003
4004 /*
4005 * Idle task boosting is a nono in general. There is one
4006 * exception, when PREEMPT_RT and NOHZ is active:
4007 *
4008 * The idle task calls get_next_timer_interrupt() and holds
4009 * the timer wheel base->lock on the CPU and another CPU wants
4010 * to access the timer (probably to cancel it). We can safely
4011 * ignore the boosting request, as the idle CPU runs this code
4012 * with interrupts disabled and will complete the lock
4013 * protected section without being interrupted. So there is no
4014 * real need to boost.
4015 */
4016 if (unlikely(p == rq->idle)) {
4017 WARN_ON(p != rq->curr);
4018 WARN_ON(p->pi_blocked_on);
4019 goto out_unlock;
4020 }
4021
4022 trace_sched_pi_setprio(p, prio);
4023 oldprio = p->prio;
4024 prev_class = p->sched_class;
4025 on_rq = p->on_rq;
4026 running = task_current(rq, p);
4027 if (on_rq)
4028 dequeue_task(rq, p, 0);
4029 if (running)
4030 p->sched_class->put_prev_task(rq, p);
4031
4032 if (rt_prio(prio))
4033 p->sched_class = &rt_sched_class;
4034 else
4035 p->sched_class = &fair_sched_class;
4036
4037 p->prio = prio;
4038
4039 if (running)
4040 p->sched_class->set_curr_task(rq);
4041 if (on_rq)
4042 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4043
4044 check_class_changed(rq, p, prev_class, oldprio);
4045out_unlock:
4046 __task_rq_unlock(rq);
4047}
4048#endif
4049void set_user_nice(struct task_struct *p, long nice)
4050{
4051 int old_prio, delta, on_rq;
4052 unsigned long flags;
4053 struct rq *rq;
4054
4055 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4056 return;
4057 /*
4058 * We have to be careful, if called from sys_setpriority(),
4059 * the task might be in the middle of scheduling on another CPU.
4060 */
4061 rq = task_rq_lock(p, &flags);
4062 /*
4063 * The RT priorities are set via sched_setscheduler(), but we still
4064 * allow the 'normal' nice value to be set - but as expected
4065 * it wont have any effect on scheduling until the task is
4066 * SCHED_FIFO/SCHED_RR:
4067 */
4068 if (task_has_rt_policy(p)) {
4069 p->static_prio = NICE_TO_PRIO(nice);
4070 goto out_unlock;
4071 }
4072 on_rq = p->on_rq;
4073 if (on_rq)
4074 dequeue_task(rq, p, 0);
4075
4076 p->static_prio = NICE_TO_PRIO(nice);
4077 set_load_weight(p);
4078 old_prio = p->prio;
4079 p->prio = effective_prio(p);
4080 delta = p->prio - old_prio;
4081
4082 if (on_rq) {
4083 enqueue_task(rq, p, 0);
4084 /*
4085 * If the task increased its priority or is running and
4086 * lowered its priority, then reschedule its CPU:
4087 */
4088 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4089 resched_task(rq->curr);
4090 }
4091out_unlock:
4092 task_rq_unlock(rq, p, &flags);
4093}
4094EXPORT_SYMBOL(set_user_nice);
4095
4096/*
4097 * can_nice - check if a task can reduce its nice value
4098 * @p: task
4099 * @nice: nice value
4100 */
4101int can_nice(const struct task_struct *p, const int nice)
4102{
4103 /* convert nice value [19,-20] to rlimit style value [1,40] */
4104 int nice_rlim = 20 - nice;
4105
4106 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4107 capable(CAP_SYS_NICE));
4108}
4109
4110#ifdef __ARCH_WANT_SYS_NICE
4111
4112/*
4113 * sys_nice - change the priority of the current process.
4114 * @increment: priority increment
4115 *
4116 * sys_setpriority is a more generic, but much slower function that
4117 * does similar things.
4118 */
4119SYSCALL_DEFINE1(nice, int, increment)
4120{
4121 long nice, retval;
4122
4123 /*
4124 * Setpriority might change our priority at the same moment.
4125 * We don't have to worry. Conceptually one call occurs first
4126 * and we have a single winner.
4127 */
4128 if (increment < -40)
4129 increment = -40;
4130 if (increment > 40)
4131 increment = 40;
4132
4133 nice = TASK_NICE(current) + increment;
4134 if (nice < -20)
4135 nice = -20;
4136 if (nice > 19)
4137 nice = 19;
4138
4139 if (increment < 0 && !can_nice(current, nice))
4140 return -EPERM;
4141
4142 retval = security_task_setnice(current, nice);
4143 if (retval)
4144 return retval;
4145
4146 set_user_nice(current, nice);
4147 return 0;
4148}
4149
4150#endif
4151
4152/**
4153 * task_prio - return the priority value of a given task.
4154 * @p: the task in question.
4155 *
4156 * This is the priority value as seen by users in /proc.
4157 * RT tasks are offset by -200. Normal tasks are centered
4158 * around 0, value goes from -16 to +15.
4159 */
4160int task_prio(const struct task_struct *p)
4161{
4162 return p->prio - MAX_RT_PRIO;
4163}
4164
4165/**
4166 * task_nice - return the nice value of a given task.
4167 * @p: the task in question.
4168 */
4169int task_nice(const struct task_struct *p)
4170{
4171 return TASK_NICE(p);
4172}
4173EXPORT_SYMBOL(task_nice);
4174
4175/**
4176 * idle_cpu - is a given cpu idle currently?
4177 * @cpu: the processor in question.
4178 */
4179int idle_cpu(int cpu)
4180{
4181 struct rq *rq = cpu_rq(cpu);
4182
4183 if (rq->curr != rq->idle)
4184 return 0;
4185
4186 if (rq->nr_running)
4187 return 0;
4188
4189#ifdef CONFIG_SMP
4190 if (!llist_empty(&rq->wake_list))
4191 return 0;
4192#endif
4193
4194 return 1;
4195}
4196
4197/**
4198 * idle_task - return the idle task for a given cpu.
4199 * @cpu: the processor in question.
4200 */
4201struct task_struct *idle_task(int cpu)
4202{
4203 return cpu_rq(cpu)->idle;
4204}
4205
4206/**
4207 * find_process_by_pid - find a process with a matching PID value.
4208 * @pid: the pid in question.
4209 */
4210static struct task_struct *find_process_by_pid(pid_t pid)
4211{
4212 return pid ? find_task_by_vpid(pid) : current;
4213}
4214
4215/* Actually do priority change: must hold rq lock. */
4216static void
4217__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4218{
4219 p->policy = policy;
4220 p->rt_priority = prio;
4221 p->normal_prio = normal_prio(p);
4222 /* we are holding p->pi_lock already */
4223 p->prio = rt_mutex_getprio(p);
4224 if (rt_prio(p->prio))
4225 p->sched_class = &rt_sched_class;
4226 else
4227 p->sched_class = &fair_sched_class;
4228 set_load_weight(p);
4229}
4230
4231/*
4232 * check the target process has a UID that matches the current process's
4233 */
4234static bool check_same_owner(struct task_struct *p)
4235{
4236 const struct cred *cred = current_cred(), *pcred;
4237 bool match;
4238
4239 rcu_read_lock();
4240 pcred = __task_cred(p);
4241 match = (uid_eq(cred->euid, pcred->euid) ||
4242 uid_eq(cred->euid, pcred->uid));
4243 rcu_read_unlock();
4244 return match;
4245}
4246
4247static int __sched_setscheduler(struct task_struct *p, int policy,
4248 const struct sched_param *param, bool user)
4249{
4250 int retval, oldprio, oldpolicy = -1, on_rq, running;
4251 unsigned long flags;
4252 const struct sched_class *prev_class;
4253 struct rq *rq;
4254 int reset_on_fork;
4255
4256 /* may grab non-irq protected spin_locks */
4257 BUG_ON(in_interrupt());
4258recheck:
4259 /* double check policy once rq lock held */
4260 if (policy < 0) {
4261 reset_on_fork = p->sched_reset_on_fork;
4262 policy = oldpolicy = p->policy;
4263 } else {
4264 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4265 policy &= ~SCHED_RESET_ON_FORK;
4266
4267 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4268 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4269 policy != SCHED_IDLE)
4270 return -EINVAL;
4271 }
4272
4273 /*
4274 * Valid priorities for SCHED_FIFO and SCHED_RR are
4275 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4276 * SCHED_BATCH and SCHED_IDLE is 0.
4277 */
4278 if (param->sched_priority < 0 ||
4279 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4280 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4281 return -EINVAL;
4282 if (rt_policy(policy) != (param->sched_priority != 0))
4283 return -EINVAL;
4284
4285 /*
4286 * Allow unprivileged RT tasks to decrease priority:
4287 */
4288 if (user && !capable(CAP_SYS_NICE)) {
4289 if (rt_policy(policy)) {
4290 unsigned long rlim_rtprio =
4291 task_rlimit(p, RLIMIT_RTPRIO);
4292
4293 /* can't set/change the rt policy */
4294 if (policy != p->policy && !rlim_rtprio)
4295 return -EPERM;
4296
4297 /* can't increase priority */
4298 if (param->sched_priority > p->rt_priority &&
4299 param->sched_priority > rlim_rtprio)
4300 return -EPERM;
4301 }
4302
4303 /*
4304 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4305 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4306 */
4307 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4308 if (!can_nice(p, TASK_NICE(p)))
4309 return -EPERM;
4310 }
4311
4312 /* can't change other user's priorities */
4313 if (!check_same_owner(p))
4314 return -EPERM;
4315
4316 /* Normal users shall not reset the sched_reset_on_fork flag */
4317 if (p->sched_reset_on_fork && !reset_on_fork)
4318 return -EPERM;
4319 }
4320
4321 if (user) {
4322 retval = security_task_setscheduler(p);
4323 if (retval)
4324 return retval;
4325 }
4326
4327 /*
4328 * make sure no PI-waiters arrive (or leave) while we are
4329 * changing the priority of the task:
4330 *
4331 * To be able to change p->policy safely, the appropriate
4332 * runqueue lock must be held.
4333 */
4334 rq = task_rq_lock(p, &flags);
4335
4336 /*
4337 * Changing the policy of the stop threads its a very bad idea
4338 */
4339 if (p == rq->stop) {
4340 task_rq_unlock(rq, p, &flags);
4341 return -EINVAL;
4342 }
4343
4344 /*
4345 * If not changing anything there's no need to proceed further:
4346 */
4347 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4348 param->sched_priority == p->rt_priority))) {
4349
4350 __task_rq_unlock(rq);
4351 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4352 return 0;
4353 }
4354
4355#ifdef CONFIG_RT_GROUP_SCHED
4356 if (user) {
4357 /*
4358 * Do not allow realtime tasks into groups that have no runtime
4359 * assigned.
4360 */
4361 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4362 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4363 !task_group_is_autogroup(task_group(p))) {
4364 task_rq_unlock(rq, p, &flags);
4365 return -EPERM;
4366 }
4367 }
4368#endif
4369
4370 /* recheck policy now with rq lock held */
4371 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4372 policy = oldpolicy = -1;
4373 task_rq_unlock(rq, p, &flags);
4374 goto recheck;
4375 }
4376 on_rq = p->on_rq;
4377 running = task_current(rq, p);
4378 if (on_rq)
4379 dequeue_task(rq, p, 0);
4380 if (running)
4381 p->sched_class->put_prev_task(rq, p);
4382
4383 p->sched_reset_on_fork = reset_on_fork;
4384
4385 oldprio = p->prio;
4386 prev_class = p->sched_class;
4387 __setscheduler(rq, p, policy, param->sched_priority);
4388
4389 if (running)
4390 p->sched_class->set_curr_task(rq);
4391 if (on_rq)
4392 enqueue_task(rq, p, 0);
4393
4394 check_class_changed(rq, p, prev_class, oldprio);
4395 task_rq_unlock(rq, p, &flags);
4396
4397 rt_mutex_adjust_pi(p);
4398
4399 return 0;
4400}
4401
4402/**
4403 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4404 * @p: the task in question.
4405 * @policy: new policy.
4406 * @param: structure containing the new RT priority.
4407 *
4408 * NOTE that the task may be already dead.
4409 */
4410int sched_setscheduler(struct task_struct *p, int policy,
4411 const struct sched_param *param)
4412{
4413 return __sched_setscheduler(p, policy, param, true);
4414}
4415EXPORT_SYMBOL_GPL(sched_setscheduler);
4416
4417/**
4418 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4419 * @p: the task in question.
4420 * @policy: new policy.
4421 * @param: structure containing the new RT priority.
4422 *
4423 * Just like sched_setscheduler, only don't bother checking if the
4424 * current context has permission. For example, this is needed in
4425 * stop_machine(): we create temporary high priority worker threads,
4426 * but our caller might not have that capability.
4427 */
4428int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4429 const struct sched_param *param)
4430{
4431 return __sched_setscheduler(p, policy, param, false);
4432}
4433
4434static int
4435do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4436{
4437 struct sched_param lparam;
4438 struct task_struct *p;
4439 int retval;
4440
4441 if (!param || pid < 0)
4442 return -EINVAL;
4443 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4444 return -EFAULT;
4445
4446 rcu_read_lock();
4447 retval = -ESRCH;
4448 p = find_process_by_pid(pid);
4449 if (p != NULL)
4450 retval = sched_setscheduler(p, policy, &lparam);
4451 rcu_read_unlock();
4452
4453 return retval;
4454}
4455
4456/**
4457 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4458 * @pid: the pid in question.
4459 * @policy: new policy.
4460 * @param: structure containing the new RT priority.
4461 */
4462SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4463 struct sched_param __user *, param)
4464{
4465 /* negative values for policy are not valid */
4466 if (policy < 0)
4467 return -EINVAL;
4468
4469 return do_sched_setscheduler(pid, policy, param);
4470}
4471
4472/**
4473 * sys_sched_setparam - set/change the RT priority of a thread
4474 * @pid: the pid in question.
4475 * @param: structure containing the new RT priority.
4476 */
4477SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4478{
4479 return do_sched_setscheduler(pid, -1, param);
4480}
4481
4482/**
4483 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4484 * @pid: the pid in question.
4485 */
4486SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4487{
4488 struct task_struct *p;
4489 int retval;
4490
4491 if (pid < 0)
4492 return -EINVAL;
4493
4494 retval = -ESRCH;
4495 rcu_read_lock();
4496 p = find_process_by_pid(pid);
4497 if (p) {
4498 retval = security_task_getscheduler(p);
4499 if (!retval)
4500 retval = p->policy
4501 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4502 }
4503 rcu_read_unlock();
4504 return retval;
4505}
4506
4507/**
4508 * sys_sched_getparam - get the RT priority of a thread
4509 * @pid: the pid in question.
4510 * @param: structure containing the RT priority.
4511 */
4512SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4513{
4514 struct sched_param lp;
4515 struct task_struct *p;
4516 int retval;
4517
4518 if (!param || pid < 0)
4519 return -EINVAL;
4520
4521 rcu_read_lock();
4522 p = find_process_by_pid(pid);
4523 retval = -ESRCH;
4524 if (!p)
4525 goto out_unlock;
4526
4527 retval = security_task_getscheduler(p);
4528 if (retval)
4529 goto out_unlock;
4530
4531 lp.sched_priority = p->rt_priority;
4532 rcu_read_unlock();
4533
4534 /*
4535 * This one might sleep, we cannot do it with a spinlock held ...
4536 */
4537 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4538
4539 return retval;
4540
4541out_unlock:
4542 rcu_read_unlock();
4543 return retval;
4544}
4545
4546long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4547{
4548 cpumask_var_t cpus_allowed, new_mask;
4549 struct task_struct *p;
4550 int retval;
4551
4552 get_online_cpus();
4553 rcu_read_lock();
4554
4555 p = find_process_by_pid(pid);
4556 if (!p) {
4557 rcu_read_unlock();
4558 put_online_cpus();
4559 return -ESRCH;
4560 }
4561
4562 /* Prevent p going away */
4563 get_task_struct(p);
4564 rcu_read_unlock();
4565
4566 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4567 retval = -ENOMEM;
4568 goto out_put_task;
4569 }
4570 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4571 retval = -ENOMEM;
4572 goto out_free_cpus_allowed;
4573 }
4574 retval = -EPERM;
4575 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4576 goto out_unlock;
4577
4578 retval = security_task_setscheduler(p);
4579 if (retval)
4580 goto out_unlock;
4581
4582 cpuset_cpus_allowed(p, cpus_allowed);
4583 cpumask_and(new_mask, in_mask, cpus_allowed);
4584again:
4585 retval = set_cpus_allowed_ptr(p, new_mask);
4586
4587 if (!retval) {
4588 cpuset_cpus_allowed(p, cpus_allowed);
4589 if (!cpumask_subset(new_mask, cpus_allowed)) {
4590 /*
4591 * We must have raced with a concurrent cpuset
4592 * update. Just reset the cpus_allowed to the
4593 * cpuset's cpus_allowed
4594 */
4595 cpumask_copy(new_mask, cpus_allowed);
4596 goto again;
4597 }
4598 }
4599out_unlock:
4600 free_cpumask_var(new_mask);
4601out_free_cpus_allowed:
4602 free_cpumask_var(cpus_allowed);
4603out_put_task:
4604 put_task_struct(p);
4605 put_online_cpus();
4606 return retval;
4607}
4608
4609static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4610 struct cpumask *new_mask)
4611{
4612 if (len < cpumask_size())
4613 cpumask_clear(new_mask);
4614 else if (len > cpumask_size())
4615 len = cpumask_size();
4616
4617 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4618}
4619
4620/**
4621 * sys_sched_setaffinity - set the cpu affinity of a process
4622 * @pid: pid of the process
4623 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4624 * @user_mask_ptr: user-space pointer to the new cpu mask
4625 */
4626SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4627 unsigned long __user *, user_mask_ptr)
4628{
4629 cpumask_var_t new_mask;
4630 int retval;
4631
4632 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4633 return -ENOMEM;
4634
4635 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4636 if (retval == 0)
4637 retval = sched_setaffinity(pid, new_mask);
4638 free_cpumask_var(new_mask);
4639 return retval;
4640}
4641
4642long sched_getaffinity(pid_t pid, struct cpumask *mask)
4643{
4644 struct task_struct *p;
4645 unsigned long flags;
4646 int retval;
4647
4648 get_online_cpus();
4649 rcu_read_lock();
4650
4651 retval = -ESRCH;
4652 p = find_process_by_pid(pid);
4653 if (!p)
4654 goto out_unlock;
4655
4656 retval = security_task_getscheduler(p);
4657 if (retval)
4658 goto out_unlock;
4659
4660 raw_spin_lock_irqsave(&p->pi_lock, flags);
4661 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4662 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4663
4664out_unlock:
4665 rcu_read_unlock();
4666 put_online_cpus();
4667
4668 return retval;
4669}
4670
4671/**
4672 * sys_sched_getaffinity - get the cpu affinity of a process
4673 * @pid: pid of the process
4674 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4675 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4676 */
4677SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4678 unsigned long __user *, user_mask_ptr)
4679{
4680 int ret;
4681 cpumask_var_t mask;
4682
4683 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4684 return -EINVAL;
4685 if (len & (sizeof(unsigned long)-1))
4686 return -EINVAL;
4687
4688 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4689 return -ENOMEM;
4690
4691 ret = sched_getaffinity(pid, mask);
4692 if (ret == 0) {
4693 size_t retlen = min_t(size_t, len, cpumask_size());
4694
4695 if (copy_to_user(user_mask_ptr, mask, retlen))
4696 ret = -EFAULT;
4697 else
4698 ret = retlen;
4699 }
4700 free_cpumask_var(mask);
4701
4702 return ret;
4703}
4704
4705/**
4706 * sys_sched_yield - yield the current processor to other threads.
4707 *
4708 * This function yields the current CPU to other tasks. If there are no
4709 * other threads running on this CPU then this function will return.
4710 */
4711SYSCALL_DEFINE0(sched_yield)
4712{
4713 struct rq *rq = this_rq_lock();
4714
4715 schedstat_inc(rq, yld_count);
4716 current->sched_class->yield_task(rq);
4717
4718 /*
4719 * Since we are going to call schedule() anyway, there's
4720 * no need to preempt or enable interrupts:
4721 */
4722 __release(rq->lock);
4723 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4724 do_raw_spin_unlock(&rq->lock);
4725 sched_preempt_enable_no_resched();
4726
4727 schedule();
4728
4729 return 0;
4730}
4731
4732static inline int should_resched(void)
4733{
4734 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4735}
4736
4737static void __cond_resched(void)
4738{
4739 add_preempt_count(PREEMPT_ACTIVE);
4740 __schedule();
4741 sub_preempt_count(PREEMPT_ACTIVE);
4742}
4743
4744int __sched _cond_resched(void)
4745{
4746 if (should_resched()) {
4747 __cond_resched();
4748 return 1;
4749 }
4750 return 0;
4751}
4752EXPORT_SYMBOL(_cond_resched);
4753
4754/*
4755 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4756 * call schedule, and on return reacquire the lock.
4757 *
4758 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4759 * operations here to prevent schedule() from being called twice (once via
4760 * spin_unlock(), once by hand).
4761 */
4762int __cond_resched_lock(spinlock_t *lock)
4763{
4764 int resched = should_resched();
4765 int ret = 0;
4766
4767 lockdep_assert_held(lock);
4768
4769 if (spin_needbreak(lock) || resched) {
4770 spin_unlock(lock);
4771 if (resched)
4772 __cond_resched();
4773 else
4774 cpu_relax();
4775 ret = 1;
4776 spin_lock(lock);
4777 }
4778 return ret;
4779}
4780EXPORT_SYMBOL(__cond_resched_lock);
4781
4782int __sched __cond_resched_softirq(void)
4783{
4784 BUG_ON(!in_softirq());
4785
4786 if (should_resched()) {
4787 local_bh_enable();
4788 __cond_resched();
4789 local_bh_disable();
4790 return 1;
4791 }
4792 return 0;
4793}
4794EXPORT_SYMBOL(__cond_resched_softirq);
4795
4796/**
4797 * yield - yield the current processor to other threads.
4798 *
4799 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4800 *
4801 * The scheduler is at all times free to pick the calling task as the most
4802 * eligible task to run, if removing the yield() call from your code breaks
4803 * it, its already broken.
4804 *
4805 * Typical broken usage is:
4806 *
4807 * while (!event)
4808 * yield();
4809 *
4810 * where one assumes that yield() will let 'the other' process run that will
4811 * make event true. If the current task is a SCHED_FIFO task that will never
4812 * happen. Never use yield() as a progress guarantee!!
4813 *
4814 * If you want to use yield() to wait for something, use wait_event().
4815 * If you want to use yield() to be 'nice' for others, use cond_resched().
4816 * If you still want to use yield(), do not!
4817 */
4818void __sched yield(void)
4819{
4820 set_current_state(TASK_RUNNING);
4821 sys_sched_yield();
4822}
4823EXPORT_SYMBOL(yield);
4824
4825/**
4826 * yield_to - yield the current processor to another thread in
4827 * your thread group, or accelerate that thread toward the
4828 * processor it's on.
4829 * @p: target task
4830 * @preempt: whether task preemption is allowed or not
4831 *
4832 * It's the caller's job to ensure that the target task struct
4833 * can't go away on us before we can do any checks.
4834 *
4835 * Returns true if we indeed boosted the target task.
4836 */
4837bool __sched yield_to(struct task_struct *p, bool preempt)
4838{
4839 struct task_struct *curr = current;
4840 struct rq *rq, *p_rq;
4841 unsigned long flags;
4842 bool yielded = 0;
4843
4844 local_irq_save(flags);
4845 rq = this_rq();
4846
4847again:
4848 p_rq = task_rq(p);
4849 double_rq_lock(rq, p_rq);
4850 while (task_rq(p) != p_rq) {
4851 double_rq_unlock(rq, p_rq);
4852 goto again;
4853 }
4854
4855 if (!curr->sched_class->yield_to_task)
4856 goto out;
4857
4858 if (curr->sched_class != p->sched_class)
4859 goto out;
4860
4861 if (task_running(p_rq, p) || p->state)
4862 goto out;
4863
4864 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4865 if (yielded) {
4866 schedstat_inc(rq, yld_count);
4867 /*
4868 * Make p's CPU reschedule; pick_next_entity takes care of
4869 * fairness.
4870 */
4871 if (preempt && rq != p_rq)
4872 resched_task(p_rq->curr);
4873 } else {
4874 /*
4875 * We might have set it in task_yield_fair(), but are
4876 * not going to schedule(), so don't want to skip
4877 * the next update.
4878 */
4879 rq->skip_clock_update = 0;
4880 }
4881
4882out:
4883 double_rq_unlock(rq, p_rq);
4884 local_irq_restore(flags);
4885
4886 if (yielded)
4887 schedule();
4888
4889 return yielded;
4890}
4891EXPORT_SYMBOL_GPL(yield_to);
4892
4893/*
4894 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4895 * that process accounting knows that this is a task in IO wait state.
4896 */
4897void __sched io_schedule(void)
4898{
4899 struct rq *rq = raw_rq();
4900
4901 delayacct_blkio_start();
4902 atomic_inc(&rq->nr_iowait);
4903 blk_flush_plug(current);
4904 current->in_iowait = 1;
4905 schedule();
4906 current->in_iowait = 0;
4907 atomic_dec(&rq->nr_iowait);
4908 delayacct_blkio_end();
4909}
4910EXPORT_SYMBOL(io_schedule);
4911
4912long __sched io_schedule_timeout(long timeout)
4913{
4914 struct rq *rq = raw_rq();
4915 long ret;
4916
4917 delayacct_blkio_start();
4918 atomic_inc(&rq->nr_iowait);
4919 blk_flush_plug(current);
4920 current->in_iowait = 1;
4921 ret = schedule_timeout(timeout);
4922 current->in_iowait = 0;
4923 atomic_dec(&rq->nr_iowait);
4924 delayacct_blkio_end();
4925 return ret;
4926}
4927
4928/**
4929 * sys_sched_get_priority_max - return maximum RT priority.
4930 * @policy: scheduling class.
4931 *
4932 * this syscall returns the maximum rt_priority that can be used
4933 * by a given scheduling class.
4934 */
4935SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4936{
4937 int ret = -EINVAL;
4938
4939 switch (policy) {
4940 case SCHED_FIFO:
4941 case SCHED_RR:
4942 ret = MAX_USER_RT_PRIO-1;
4943 break;
4944 case SCHED_NORMAL:
4945 case SCHED_BATCH:
4946 case SCHED_IDLE:
4947 ret = 0;
4948 break;
4949 }
4950 return ret;
4951}
4952
4953/**
4954 * sys_sched_get_priority_min - return minimum RT priority.
4955 * @policy: scheduling class.
4956 *
4957 * this syscall returns the minimum rt_priority that can be used
4958 * by a given scheduling class.
4959 */
4960SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4961{
4962 int ret = -EINVAL;
4963
4964 switch (policy) {
4965 case SCHED_FIFO:
4966 case SCHED_RR:
4967 ret = 1;
4968 break;
4969 case SCHED_NORMAL:
4970 case SCHED_BATCH:
4971 case SCHED_IDLE:
4972 ret = 0;
4973 }
4974 return ret;
4975}
4976
4977/**
4978 * sys_sched_rr_get_interval - return the default timeslice of a process.
4979 * @pid: pid of the process.
4980 * @interval: userspace pointer to the timeslice value.
4981 *
4982 * this syscall writes the default timeslice value of a given process
4983 * into the user-space timespec buffer. A value of '0' means infinity.
4984 */
4985SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4986 struct timespec __user *, interval)
4987{
4988 struct task_struct *p;
4989 unsigned int time_slice;
4990 unsigned long flags;
4991 struct rq *rq;
4992 int retval;
4993 struct timespec t;
4994
4995 if (pid < 0)
4996 return -EINVAL;
4997
4998 retval = -ESRCH;
4999 rcu_read_lock();
5000 p = find_process_by_pid(pid);
5001 if (!p)
5002 goto out_unlock;
5003
5004 retval = security_task_getscheduler(p);
5005 if (retval)
5006 goto out_unlock;
5007
5008 rq = task_rq_lock(p, &flags);
5009 time_slice = p->sched_class->get_rr_interval(rq, p);
5010 task_rq_unlock(rq, p, &flags);
5011
5012 rcu_read_unlock();
5013 jiffies_to_timespec(time_slice, &t);
5014 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5015 return retval;
5016
5017out_unlock:
5018 rcu_read_unlock();
5019 return retval;
5020}
5021
5022static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5023
5024void sched_show_task(struct task_struct *p)
5025{
5026 unsigned long free = 0;
5027 unsigned state;
5028
5029 state = p->state ? __ffs(p->state) + 1 : 0;
5030 printk(KERN_INFO "%-15.15s %c", p->comm,
5031 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5032#if BITS_PER_LONG == 32
5033 if (state == TASK_RUNNING)
5034 printk(KERN_CONT " running ");
5035 else
5036 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5037#else
5038 if (state == TASK_RUNNING)
5039 printk(KERN_CONT " running task ");
5040 else
5041 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5042#endif
5043#ifdef CONFIG_DEBUG_STACK_USAGE
5044 free = stack_not_used(p);
5045#endif
5046 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5047 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
5048 (unsigned long)task_thread_info(p)->flags);
5049
5050 show_stack(p, NULL);
5051}
5052
5053void show_state_filter(unsigned long state_filter)
5054{
5055 struct task_struct *g, *p;
5056
5057#if BITS_PER_LONG == 32
5058 printk(KERN_INFO
5059 " task PC stack pid father\n");
5060#else
5061 printk(KERN_INFO
5062 " task PC stack pid father\n");
5063#endif
5064 rcu_read_lock();
5065 do_each_thread(g, p) {
5066 /*
5067 * reset the NMI-timeout, listing all files on a slow
5068 * console might take a lot of time:
5069 */
5070 touch_nmi_watchdog();
5071 if (!state_filter || (p->state & state_filter))
5072 sched_show_task(p);
5073 } while_each_thread(g, p);
5074
5075 touch_all_softlockup_watchdogs();
5076
5077#ifdef CONFIG_SCHED_DEBUG
5078 sysrq_sched_debug_show();
5079#endif
5080 rcu_read_unlock();
5081 /*
5082 * Only show locks if all tasks are dumped:
5083 */
5084 if (!state_filter)
5085 debug_show_all_locks();
5086}
5087
5088void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5089{
5090 idle->sched_class = &idle_sched_class;
5091}
5092
5093/**
5094 * init_idle - set up an idle thread for a given CPU
5095 * @idle: task in question
5096 * @cpu: cpu the idle task belongs to
5097 *
5098 * NOTE: this function does not set the idle thread's NEED_RESCHED
5099 * flag, to make booting more robust.
5100 */
5101void __cpuinit init_idle(struct task_struct *idle, int cpu)
5102{
5103 struct rq *rq = cpu_rq(cpu);
5104 unsigned long flags;
5105
5106 raw_spin_lock_irqsave(&rq->lock, flags);
5107
5108 __sched_fork(idle);
5109 idle->state = TASK_RUNNING;
5110 idle->se.exec_start = sched_clock();
5111
5112 do_set_cpus_allowed(idle, cpumask_of(cpu));
5113 /*
5114 * We're having a chicken and egg problem, even though we are
5115 * holding rq->lock, the cpu isn't yet set to this cpu so the
5116 * lockdep check in task_group() will fail.
5117 *
5118 * Similar case to sched_fork(). / Alternatively we could
5119 * use task_rq_lock() here and obtain the other rq->lock.
5120 *
5121 * Silence PROVE_RCU
5122 */
5123 rcu_read_lock();
5124 __set_task_cpu(idle, cpu);
5125 rcu_read_unlock();
5126
5127 rq->curr = rq->idle = idle;
5128#if defined(CONFIG_SMP)
5129 idle->on_cpu = 1;
5130#endif
5131 raw_spin_unlock_irqrestore(&rq->lock, flags);
5132
5133 /* Set the preempt count _outside_ the spinlocks! */
5134 task_thread_info(idle)->preempt_count = 0;
5135
5136 /*
5137 * The idle tasks have their own, simple scheduling class:
5138 */
5139 idle->sched_class = &idle_sched_class;
5140 ftrace_graph_init_idle_task(idle, cpu);
5141#if defined(CONFIG_SMP)
5142 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5143#endif
5144}
5145
5146#ifdef CONFIG_SMP
5147void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
5148{
5149 if (p->sched_class && p->sched_class->set_cpus_allowed)
5150 p->sched_class->set_cpus_allowed(p, new_mask);
5151
5152 cpumask_copy(&p->cpus_allowed, new_mask);
5153 p->nr_cpus_allowed = cpumask_weight(new_mask);
5154}
5155
5156/*
5157 * This is how migration works:
5158 *
5159 * 1) we invoke migration_cpu_stop() on the target CPU using
5160 * stop_one_cpu().
5161 * 2) stopper starts to run (implicitly forcing the migrated thread
5162 * off the CPU)
5163 * 3) it checks whether the migrated task is still in the wrong runqueue.
5164 * 4) if it's in the wrong runqueue then the migration thread removes
5165 * it and puts it into the right queue.
5166 * 5) stopper completes and stop_one_cpu() returns and the migration
5167 * is done.
5168 */
5169
5170/*
5171 * Change a given task's CPU affinity. Migrate the thread to a
5172 * proper CPU and schedule it away if the CPU it's executing on
5173 * is removed from the allowed bitmask.
5174 *
5175 * NOTE: the caller must have a valid reference to the task, the
5176 * task must not exit() & deallocate itself prematurely. The
5177 * call is not atomic; no spinlocks may be held.
5178 */
5179int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5180{
5181 unsigned long flags;
5182 struct rq *rq;
5183 unsigned int dest_cpu;
5184 int ret = 0;
5185
5186 rq = task_rq_lock(p, &flags);
5187
5188 if (cpumask_equal(&p->cpus_allowed, new_mask))
5189 goto out;
5190
5191 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5192 ret = -EINVAL;
5193 goto out;
5194 }
5195
5196 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
5197 ret = -EINVAL;
5198 goto out;
5199 }
5200
5201 do_set_cpus_allowed(p, new_mask);
5202
5203 /* Can the task run on the task's current CPU? If so, we're done */
5204 if (cpumask_test_cpu(task_cpu(p), new_mask))
5205 goto out;
5206
5207 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5208 if (p->on_rq) {
5209 struct migration_arg arg = { p, dest_cpu };
5210 /* Need help from migration thread: drop lock and wait. */
5211 task_rq_unlock(rq, p, &flags);
5212 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5213 tlb_migrate_finish(p->mm);
5214 return 0;
5215 }
5216out:
5217 task_rq_unlock(rq, p, &flags);
5218
5219 return ret;
5220}
5221EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5222
5223/*
5224 * Move (not current) task off this cpu, onto dest cpu. We're doing
5225 * this because either it can't run here any more (set_cpus_allowed()
5226 * away from this CPU, or CPU going down), or because we're
5227 * attempting to rebalance this task on exec (sched_exec).
5228 *
5229 * So we race with normal scheduler movements, but that's OK, as long
5230 * as the task is no longer on this CPU.
5231 *
5232 * Returns non-zero if task was successfully migrated.
5233 */
5234static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5235{
5236 struct rq *rq_dest, *rq_src;
5237 int ret = 0;
5238
5239 if (unlikely(!cpu_active(dest_cpu)))
5240 return ret;
5241
5242 rq_src = cpu_rq(src_cpu);
5243 rq_dest = cpu_rq(dest_cpu);
5244
5245 raw_spin_lock(&p->pi_lock);
5246 double_rq_lock(rq_src, rq_dest);
5247 /* Already moved. */
5248 if (task_cpu(p) != src_cpu)
5249 goto done;
5250 /* Affinity changed (again). */
5251 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
5252 goto fail;
5253
5254 /*
5255 * If we're not on a rq, the next wake-up will ensure we're
5256 * placed properly.
5257 */
5258 if (p->on_rq) {
5259 dequeue_task(rq_src, p, 0);
5260 set_task_cpu(p, dest_cpu);
5261 enqueue_task(rq_dest, p, 0);
5262 check_preempt_curr(rq_dest, p, 0);
5263 }
5264done:
5265 ret = 1;
5266fail:
5267 double_rq_unlock(rq_src, rq_dest);
5268 raw_spin_unlock(&p->pi_lock);
5269 return ret;
5270}
5271
5272/*
5273 * migration_cpu_stop - this will be executed by a highprio stopper thread
5274 * and performs thread migration by bumping thread off CPU then
5275 * 'pushing' onto another runqueue.
5276 */
5277static int migration_cpu_stop(void *data)
5278{
5279 struct migration_arg *arg = data;
5280
5281 /*
5282 * The original target cpu might have gone down and we might
5283 * be on another cpu but it doesn't matter.
5284 */
5285 local_irq_disable();
5286 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5287 local_irq_enable();
5288 return 0;
5289}
5290
5291#ifdef CONFIG_HOTPLUG_CPU
5292
5293/*
5294 * Ensures that the idle task is using init_mm right before its cpu goes
5295 * offline.
5296 */
5297void idle_task_exit(void)
5298{
5299 struct mm_struct *mm = current->active_mm;
5300
5301 BUG_ON(cpu_online(smp_processor_id()));
5302
5303 if (mm != &init_mm)
5304 switch_mm(mm, &init_mm, current);
5305 mmdrop(mm);
5306}
5307
5308/*
5309 * While a dead CPU has no uninterruptible tasks queued at this point,
5310 * it might still have a nonzero ->nr_uninterruptible counter, because
5311 * for performance reasons the counter is not stricly tracking tasks to
5312 * their home CPUs. So we just add the counter to another CPU's counter,
5313 * to keep the global sum constant after CPU-down:
5314 */
5315static void migrate_nr_uninterruptible(struct rq *rq_src)
5316{
5317 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5318
5319 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5320 rq_src->nr_uninterruptible = 0;
5321}
5322
5323/*
5324 * remove the tasks which were accounted by rq from calc_load_tasks.
5325 */
5326static void calc_global_load_remove(struct rq *rq)
5327{
5328 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5329 rq->calc_load_active = 0;
5330}
5331
5332/*
5333 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5334 * try_to_wake_up()->select_task_rq().
5335 *
5336 * Called with rq->lock held even though we'er in stop_machine() and
5337 * there's no concurrency possible, we hold the required locks anyway
5338 * because of lock validation efforts.
5339 */
5340static void migrate_tasks(unsigned int dead_cpu)
5341{
5342 struct rq *rq = cpu_rq(dead_cpu);
5343 struct task_struct *next, *stop = rq->stop;
5344 int dest_cpu;
5345
5346 /*
5347 * Fudge the rq selection such that the below task selection loop
5348 * doesn't get stuck on the currently eligible stop task.
5349 *
5350 * We're currently inside stop_machine() and the rq is either stuck
5351 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5352 * either way we should never end up calling schedule() until we're
5353 * done here.
5354 */
5355 rq->stop = NULL;
5356
5357 /* Ensure any throttled groups are reachable by pick_next_task */
5358 unthrottle_offline_cfs_rqs(rq);
5359
5360 for ( ; ; ) {
5361 /*
5362 * There's this thread running, bail when that's the only
5363 * remaining thread.
5364 */
5365 if (rq->nr_running == 1)
5366 break;
5367
5368 next = pick_next_task(rq);
5369 BUG_ON(!next);
5370 next->sched_class->put_prev_task(rq, next);
5371
5372 /* Find suitable destination for @next, with force if needed. */
5373 dest_cpu = select_fallback_rq(dead_cpu, next);
5374 raw_spin_unlock(&rq->lock);
5375
5376 __migrate_task(next, dead_cpu, dest_cpu);
5377
5378 raw_spin_lock(&rq->lock);
5379 }
5380
5381 rq->stop = stop;
5382}
5383
5384#endif /* CONFIG_HOTPLUG_CPU */
5385
5386#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5387
5388static struct ctl_table sd_ctl_dir[] = {
5389 {
5390 .procname = "sched_domain",
5391 .mode = 0555,
5392 },
5393 {}
5394};
5395
5396static struct ctl_table sd_ctl_root[] = {
5397 {
5398 .procname = "kernel",
5399 .mode = 0555,
5400 .child = sd_ctl_dir,
5401 },
5402 {}
5403};
5404
5405static struct ctl_table *sd_alloc_ctl_entry(int n)
5406{
5407 struct ctl_table *entry =
5408 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5409
5410 return entry;
5411}
5412
5413static void sd_free_ctl_entry(struct ctl_table **tablep)
5414{
5415 struct ctl_table *entry;
5416
5417 /*
5418 * In the intermediate directories, both the child directory and
5419 * procname are dynamically allocated and could fail but the mode
5420 * will always be set. In the lowest directory the names are
5421 * static strings and all have proc handlers.
5422 */
5423 for (entry = *tablep; entry->mode; entry++) {
5424 if (entry->child)
5425 sd_free_ctl_entry(&entry->child);
5426 if (entry->proc_handler == NULL)
5427 kfree(entry->procname);
5428 }
5429
5430 kfree(*tablep);
5431 *tablep = NULL;
5432}
5433
5434static void
5435set_table_entry(struct ctl_table *entry,
5436 const char *procname, void *data, int maxlen,
5437 umode_t mode, proc_handler *proc_handler)
5438{
5439 entry->procname = procname;
5440 entry->data = data;
5441 entry->maxlen = maxlen;
5442 entry->mode = mode;
5443 entry->proc_handler = proc_handler;
5444}
5445
5446static struct ctl_table *
5447sd_alloc_ctl_domain_table(struct sched_domain *sd)
5448{
5449 struct ctl_table *table = sd_alloc_ctl_entry(13);
5450
5451 if (table == NULL)
5452 return NULL;
5453
5454 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5455 sizeof(long), 0644, proc_doulongvec_minmax);
5456 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5457 sizeof(long), 0644, proc_doulongvec_minmax);
5458 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5459 sizeof(int), 0644, proc_dointvec_minmax);
5460 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5461 sizeof(int), 0644, proc_dointvec_minmax);
5462 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5463 sizeof(int), 0644, proc_dointvec_minmax);
5464 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5465 sizeof(int), 0644, proc_dointvec_minmax);
5466 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5467 sizeof(int), 0644, proc_dointvec_minmax);
5468 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5469 sizeof(int), 0644, proc_dointvec_minmax);
5470 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5471 sizeof(int), 0644, proc_dointvec_minmax);
5472 set_table_entry(&table[9], "cache_nice_tries",
5473 &sd->cache_nice_tries,
5474 sizeof(int), 0644, proc_dointvec_minmax);
5475 set_table_entry(&table[10], "flags", &sd->flags,
5476 sizeof(int), 0644, proc_dointvec_minmax);
5477 set_table_entry(&table[11], "name", sd->name,
5478 CORENAME_MAX_SIZE, 0444, proc_dostring);
5479 /* &table[12] is terminator */
5480
5481 return table;
5482}
5483
5484static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5485{
5486 struct ctl_table *entry, *table;
5487 struct sched_domain *sd;
5488 int domain_num = 0, i;
5489 char buf[32];
5490
5491 for_each_domain(cpu, sd)
5492 domain_num++;
5493 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5494 if (table == NULL)
5495 return NULL;
5496
5497 i = 0;
5498 for_each_domain(cpu, sd) {
5499 snprintf(buf, 32, "domain%d", i);
5500 entry->procname = kstrdup(buf, GFP_KERNEL);
5501 entry->mode = 0555;
5502 entry->child = sd_alloc_ctl_domain_table(sd);
5503 entry++;
5504 i++;
5505 }
5506 return table;
5507}
5508
5509static struct ctl_table_header *sd_sysctl_header;
5510static void register_sched_domain_sysctl(void)
5511{
5512 int i, cpu_num = num_possible_cpus();
5513 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5514 char buf[32];
5515
5516 WARN_ON(sd_ctl_dir[0].child);
5517 sd_ctl_dir[0].child = entry;
5518
5519 if (entry == NULL)
5520 return;
5521
5522 for_each_possible_cpu(i) {
5523 snprintf(buf, 32, "cpu%d", i);
5524 entry->procname = kstrdup(buf, GFP_KERNEL);
5525 entry->mode = 0555;
5526 entry->child = sd_alloc_ctl_cpu_table(i);
5527 entry++;
5528 }
5529
5530 WARN_ON(sd_sysctl_header);
5531 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5532}
5533
5534/* may be called multiple times per register */
5535static void unregister_sched_domain_sysctl(void)
5536{
5537 if (sd_sysctl_header)
5538 unregister_sysctl_table(sd_sysctl_header);
5539 sd_sysctl_header = NULL;
5540 if (sd_ctl_dir[0].child)
5541 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5542}
5543#else
5544static void register_sched_domain_sysctl(void)
5545{
5546}
5547static void unregister_sched_domain_sysctl(void)
5548{
5549}
5550#endif
5551
5552static void set_rq_online(struct rq *rq)
5553{
5554 if (!rq->online) {
5555 const struct sched_class *class;
5556
5557 cpumask_set_cpu(rq->cpu, rq->rd->online);
5558 rq->online = 1;
5559
5560 for_each_class(class) {
5561 if (class->rq_online)
5562 class->rq_online(rq);
5563 }
5564 }
5565}
5566
5567static void set_rq_offline(struct rq *rq)
5568{
5569 if (rq->online) {
5570 const struct sched_class *class;
5571
5572 for_each_class(class) {
5573 if (class->rq_offline)
5574 class->rq_offline(rq);
5575 }
5576
5577 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5578 rq->online = 0;
5579 }
5580}
5581
5582/*
5583 * migration_call - callback that gets triggered when a CPU is added.
5584 * Here we can start up the necessary migration thread for the new CPU.
5585 */
5586static int __cpuinit
5587migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5588{
5589 int cpu = (long)hcpu;
5590 unsigned long flags;
5591 struct rq *rq = cpu_rq(cpu);
5592
5593 switch (action & ~CPU_TASKS_FROZEN) {
5594
5595 case CPU_UP_PREPARE:
5596 rq->calc_load_update = calc_load_update;
5597 break;
5598
5599 case CPU_ONLINE:
5600 /* Update our root-domain */
5601 raw_spin_lock_irqsave(&rq->lock, flags);
5602 if (rq->rd) {
5603 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5604
5605 set_rq_online(rq);
5606 }
5607 raw_spin_unlock_irqrestore(&rq->lock, flags);
5608 break;
5609
5610#ifdef CONFIG_HOTPLUG_CPU
5611 case CPU_DYING:
5612 sched_ttwu_pending();
5613 /* Update our root-domain */
5614 raw_spin_lock_irqsave(&rq->lock, flags);
5615 if (rq->rd) {
5616 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5617 set_rq_offline(rq);
5618 }
5619 migrate_tasks(cpu);
5620 BUG_ON(rq->nr_running != 1); /* the migration thread */
5621 raw_spin_unlock_irqrestore(&rq->lock, flags);
5622
5623 migrate_nr_uninterruptible(rq);
5624 calc_global_load_remove(rq);
5625 break;
5626#endif
5627 }
5628
5629 update_max_interval();
5630
5631 return NOTIFY_OK;
5632}
5633
5634/*
5635 * Register at high priority so that task migration (migrate_all_tasks)
5636 * happens before everything else. This has to be lower priority than
5637 * the notifier in the perf_event subsystem, though.
5638 */
5639static struct notifier_block __cpuinitdata migration_notifier = {
5640 .notifier_call = migration_call,
5641 .priority = CPU_PRI_MIGRATION,
5642};
5643
5644static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5645 unsigned long action, void *hcpu)
5646{
5647 switch (action & ~CPU_TASKS_FROZEN) {
5648 case CPU_STARTING:
5649 case CPU_DOWN_FAILED:
5650 set_cpu_active((long)hcpu, true);
5651 return NOTIFY_OK;
5652 default:
5653 return NOTIFY_DONE;
5654 }
5655}
5656
5657static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5658 unsigned long action, void *hcpu)
5659{
5660 switch (action & ~CPU_TASKS_FROZEN) {
5661 case CPU_DOWN_PREPARE:
5662 set_cpu_active((long)hcpu, false);
5663 return NOTIFY_OK;
5664 default:
5665 return NOTIFY_DONE;
5666 }
5667}
5668
5669static int __init migration_init(void)
5670{
5671 void *cpu = (void *)(long)smp_processor_id();
5672 int err;
5673
5674 /* Initialize migration for the boot CPU */
5675 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5676 BUG_ON(err == NOTIFY_BAD);
5677 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5678 register_cpu_notifier(&migration_notifier);
5679
5680 /* Register cpu active notifiers */
5681 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5682 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5683
5684 return 0;
5685}
5686early_initcall(migration_init);
5687#endif
5688
5689#ifdef CONFIG_SMP
5690
5691static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5692
5693#ifdef CONFIG_SCHED_DEBUG
5694
5695static __read_mostly int sched_debug_enabled;
5696
5697static int __init sched_debug_setup(char *str)
5698{
5699 sched_debug_enabled = 1;
5700
5701 return 0;
5702}
5703early_param("sched_debug", sched_debug_setup);
5704
5705static inline bool sched_debug(void)
5706{
5707 return sched_debug_enabled;
5708}
5709
5710static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5711 struct cpumask *groupmask)
5712{
5713 struct sched_group *group = sd->groups;
5714 char str[256];
5715
5716 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5717 cpumask_clear(groupmask);
5718
5719 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5720
5721 if (!(sd->flags & SD_LOAD_BALANCE)) {
5722 printk("does not load-balance\n");
5723 if (sd->parent)
5724 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5725 " has parent");
5726 return -1;
5727 }
5728
5729 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5730
5731 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5732 printk(KERN_ERR "ERROR: domain->span does not contain "
5733 "CPU%d\n", cpu);
5734 }
5735 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5736 printk(KERN_ERR "ERROR: domain->groups does not contain"
5737 " CPU%d\n", cpu);
5738 }
5739
5740 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5741 do {
5742 if (!group) {
5743 printk("\n");
5744 printk(KERN_ERR "ERROR: group is NULL\n");
5745 break;
5746 }
5747
5748 /*
5749 * Even though we initialize ->power to something semi-sane,
5750 * we leave power_orig unset. This allows us to detect if
5751 * domain iteration is still funny without causing /0 traps.
5752 */
5753 if (!group->sgp->power_orig) {
5754 printk(KERN_CONT "\n");
5755 printk(KERN_ERR "ERROR: domain->cpu_power not "
5756 "set\n");
5757 break;
5758 }
5759
5760 if (!cpumask_weight(sched_group_cpus(group))) {
5761 printk(KERN_CONT "\n");
5762 printk(KERN_ERR "ERROR: empty group\n");
5763 break;
5764 }
5765
5766 if (!(sd->flags & SD_OVERLAP) &&
5767 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5768 printk(KERN_CONT "\n");
5769 printk(KERN_ERR "ERROR: repeated CPUs\n");
5770 break;
5771 }
5772
5773 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5774
5775 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5776
5777 printk(KERN_CONT " %s", str);
5778 if (group->sgp->power != SCHED_POWER_SCALE) {
5779 printk(KERN_CONT " (cpu_power = %d)",
5780 group->sgp->power);
5781 }
5782
5783 group = group->next;
5784 } while (group != sd->groups);
5785 printk(KERN_CONT "\n");
5786
5787 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5788 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5789
5790 if (sd->parent &&
5791 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5792 printk(KERN_ERR "ERROR: parent span is not a superset "
5793 "of domain->span\n");
5794 return 0;
5795}
5796
5797static void sched_domain_debug(struct sched_domain *sd, int cpu)
5798{
5799 int level = 0;
5800
5801 if (!sched_debug_enabled)
5802 return;
5803
5804 if (!sd) {
5805 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5806 return;
5807 }
5808
5809 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5810
5811 for (;;) {
5812 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5813 break;
5814 level++;
5815 sd = sd->parent;
5816 if (!sd)
5817 break;
5818 }
5819}
5820#else /* !CONFIG_SCHED_DEBUG */
5821# define sched_domain_debug(sd, cpu) do { } while (0)
5822static inline bool sched_debug(void)
5823{
5824 return false;
5825}
5826#endif /* CONFIG_SCHED_DEBUG */
5827
5828static int sd_degenerate(struct sched_domain *sd)
5829{
5830 if (cpumask_weight(sched_domain_span(sd)) == 1)
5831 return 1;
5832
5833 /* Following flags need at least 2 groups */
5834 if (sd->flags & (SD_LOAD_BALANCE |
5835 SD_BALANCE_NEWIDLE |
5836 SD_BALANCE_FORK |
5837 SD_BALANCE_EXEC |
5838 SD_SHARE_CPUPOWER |
5839 SD_SHARE_PKG_RESOURCES)) {
5840 if (sd->groups != sd->groups->next)
5841 return 0;
5842 }
5843
5844 /* Following flags don't use groups */
5845 if (sd->flags & (SD_WAKE_AFFINE))
5846 return 0;
5847
5848 return 1;
5849}
5850
5851static int
5852sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5853{
5854 unsigned long cflags = sd->flags, pflags = parent->flags;
5855
5856 if (sd_degenerate(parent))
5857 return 1;
5858
5859 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5860 return 0;
5861
5862 /* Flags needing groups don't count if only 1 group in parent */
5863 if (parent->groups == parent->groups->next) {
5864 pflags &= ~(SD_LOAD_BALANCE |
5865 SD_BALANCE_NEWIDLE |
5866 SD_BALANCE_FORK |
5867 SD_BALANCE_EXEC |
5868 SD_SHARE_CPUPOWER |
5869 SD_SHARE_PKG_RESOURCES);
5870 if (nr_node_ids == 1)
5871 pflags &= ~SD_SERIALIZE;
5872 }
5873 if (~cflags & pflags)
5874 return 0;
5875
5876 return 1;
5877}
5878
5879static void free_rootdomain(struct rcu_head *rcu)
5880{
5881 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5882
5883 cpupri_cleanup(&rd->cpupri);
5884 free_cpumask_var(rd->rto_mask);
5885 free_cpumask_var(rd->online);
5886 free_cpumask_var(rd->span);
5887 kfree(rd);
5888}
5889
5890static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5891{
5892 struct root_domain *old_rd = NULL;
5893 unsigned long flags;
5894
5895 raw_spin_lock_irqsave(&rq->lock, flags);
5896
5897 if (rq->rd) {
5898 old_rd = rq->rd;
5899
5900 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5901 set_rq_offline(rq);
5902
5903 cpumask_clear_cpu(rq->cpu, old_rd->span);
5904
5905 /*
5906 * If we dont want to free the old_rt yet then
5907 * set old_rd to NULL to skip the freeing later
5908 * in this function:
5909 */
5910 if (!atomic_dec_and_test(&old_rd->refcount))
5911 old_rd = NULL;
5912 }
5913
5914 atomic_inc(&rd->refcount);
5915 rq->rd = rd;
5916
5917 cpumask_set_cpu(rq->cpu, rd->span);
5918 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5919 set_rq_online(rq);
5920
5921 raw_spin_unlock_irqrestore(&rq->lock, flags);
5922
5923 if (old_rd)
5924 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5925}
5926
5927static int init_rootdomain(struct root_domain *rd)
5928{
5929 memset(rd, 0, sizeof(*rd));
5930
5931 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5932 goto out;
5933 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5934 goto free_span;
5935 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5936 goto free_online;
5937
5938 if (cpupri_init(&rd->cpupri) != 0)
5939 goto free_rto_mask;
5940 return 0;
5941
5942free_rto_mask:
5943 free_cpumask_var(rd->rto_mask);
5944free_online:
5945 free_cpumask_var(rd->online);
5946free_span:
5947 free_cpumask_var(rd->span);
5948out:
5949 return -ENOMEM;
5950}
5951
5952/*
5953 * By default the system creates a single root-domain with all cpus as
5954 * members (mimicking the global state we have today).
5955 */
5956struct root_domain def_root_domain;
5957
5958static void init_defrootdomain(void)
5959{
5960 init_rootdomain(&def_root_domain);
5961
5962 atomic_set(&def_root_domain.refcount, 1);
5963}
5964
5965static struct root_domain *alloc_rootdomain(void)
5966{
5967 struct root_domain *rd;
5968
5969 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5970 if (!rd)
5971 return NULL;
5972
5973 if (init_rootdomain(rd) != 0) {
5974 kfree(rd);
5975 return NULL;
5976 }
5977
5978 return rd;
5979}
5980
5981static void free_sched_groups(struct sched_group *sg, int free_sgp)
5982{
5983 struct sched_group *tmp, *first;
5984
5985 if (!sg)
5986 return;
5987
5988 first = sg;
5989 do {
5990 tmp = sg->next;
5991
5992 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5993 kfree(sg->sgp);
5994
5995 kfree(sg);
5996 sg = tmp;
5997 } while (sg != first);
5998}
5999
6000static void free_sched_domain(struct rcu_head *rcu)
6001{
6002 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6003
6004 /*
6005 * If its an overlapping domain it has private groups, iterate and
6006 * nuke them all.
6007 */
6008 if (sd->flags & SD_OVERLAP) {
6009 free_sched_groups(sd->groups, 1);
6010 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6011 kfree(sd->groups->sgp);
6012 kfree(sd->groups);
6013 }
6014 kfree(sd);
6015}
6016
6017static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6018{
6019 call_rcu(&sd->rcu, free_sched_domain);
6020}
6021
6022static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6023{
6024 for (; sd; sd = sd->parent)
6025 destroy_sched_domain(sd, cpu);
6026}
6027
6028/*
6029 * Keep a special pointer to the highest sched_domain that has
6030 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6031 * allows us to avoid some pointer chasing select_idle_sibling().
6032 *
6033 * Also keep a unique ID per domain (we use the first cpu number in
6034 * the cpumask of the domain), this allows us to quickly tell if
6035 * two cpus are in the same cache domain, see cpus_share_cache().
6036 */
6037DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6038DEFINE_PER_CPU(int, sd_llc_id);
6039
6040static void update_top_cache_domain(int cpu)
6041{
6042 struct sched_domain *sd;
6043 int id = cpu;
6044
6045 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6046 if (sd)
6047 id = cpumask_first(sched_domain_span(sd));
6048
6049 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6050 per_cpu(sd_llc_id, cpu) = id;
6051}
6052
6053/*
6054 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6055 * hold the hotplug lock.
6056 */
6057static void
6058cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6059{
6060 struct rq *rq = cpu_rq(cpu);
6061 struct sched_domain *tmp;
6062
6063 /* Remove the sched domains which do not contribute to scheduling. */
6064 for (tmp = sd; tmp; ) {
6065 struct sched_domain *parent = tmp->parent;
6066 if (!parent)
6067 break;
6068
6069 if (sd_parent_degenerate(tmp, parent)) {
6070 tmp->parent = parent->parent;
6071 if (parent->parent)
6072 parent->parent->child = tmp;
6073 destroy_sched_domain(parent, cpu);
6074 } else
6075 tmp = tmp->parent;
6076 }
6077
6078 if (sd && sd_degenerate(sd)) {
6079 tmp = sd;
6080 sd = sd->parent;
6081 destroy_sched_domain(tmp, cpu);
6082 if (sd)
6083 sd->child = NULL;
6084 }
6085
6086 sched_domain_debug(sd, cpu);
6087
6088 rq_attach_root(rq, rd);
6089 tmp = rq->sd;
6090 rcu_assign_pointer(rq->sd, sd);
6091 destroy_sched_domains(tmp, cpu);
6092
6093 update_top_cache_domain(cpu);
6094}
6095
6096/* cpus with isolated domains */
6097static cpumask_var_t cpu_isolated_map;
6098
6099/* Setup the mask of cpus configured for isolated domains */
6100static int __init isolated_cpu_setup(char *str)
6101{
6102 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6103 cpulist_parse(str, cpu_isolated_map);
6104 return 1;
6105}
6106
6107__setup("isolcpus=", isolated_cpu_setup);
6108
6109static const struct cpumask *cpu_cpu_mask(int cpu)
6110{
6111 return cpumask_of_node(cpu_to_node(cpu));
6112}
6113
6114struct sd_data {
6115 struct sched_domain **__percpu sd;
6116 struct sched_group **__percpu sg;
6117 struct sched_group_power **__percpu sgp;
6118};
6119
6120struct s_data {
6121 struct sched_domain ** __percpu sd;
6122 struct root_domain *rd;
6123};
6124
6125enum s_alloc {
6126 sa_rootdomain,
6127 sa_sd,
6128 sa_sd_storage,
6129 sa_none,
6130};
6131
6132struct sched_domain_topology_level;
6133
6134typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
6135typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
6136
6137#define SDTL_OVERLAP 0x01
6138
6139struct sched_domain_topology_level {
6140 sched_domain_init_f init;
6141 sched_domain_mask_f mask;
6142 int flags;
6143 int numa_level;
6144 struct sd_data data;
6145};
6146
6147/*
6148 * Build an iteration mask that can exclude certain CPUs from the upwards
6149 * domain traversal.
6150 *
6151 * Asymmetric node setups can result in situations where the domain tree is of
6152 * unequal depth, make sure to skip domains that already cover the entire
6153 * range.
6154 *
6155 * In that case build_sched_domains() will have terminated the iteration early
6156 * and our sibling sd spans will be empty. Domains should always include the
6157 * cpu they're built on, so check that.
6158 *
6159 */
6160static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6161{
6162 const struct cpumask *span = sched_domain_span(sd);
6163 struct sd_data *sdd = sd->private;
6164 struct sched_domain *sibling;
6165 int i;
6166
6167 for_each_cpu(i, span) {
6168 sibling = *per_cpu_ptr(sdd->sd, i);
6169 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6170 continue;
6171
6172 cpumask_set_cpu(i, sched_group_mask(sg));
6173 }
6174}
6175
6176/*
6177 * Return the canonical balance cpu for this group, this is the first cpu
6178 * of this group that's also in the iteration mask.
6179 */
6180int group_balance_cpu(struct sched_group *sg)
6181{
6182 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6183}
6184
6185static int
6186build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6187{
6188 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6189 const struct cpumask *span = sched_domain_span(sd);
6190 struct cpumask *covered = sched_domains_tmpmask;
6191 struct sd_data *sdd = sd->private;
6192 struct sched_domain *child;
6193 int i;
6194
6195 cpumask_clear(covered);
6196
6197 for_each_cpu(i, span) {
6198 struct cpumask *sg_span;
6199
6200 if (cpumask_test_cpu(i, covered))
6201 continue;
6202
6203 child = *per_cpu_ptr(sdd->sd, i);
6204
6205 /* See the comment near build_group_mask(). */
6206 if (!cpumask_test_cpu(i, sched_domain_span(child)))
6207 continue;
6208
6209 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6210 GFP_KERNEL, cpu_to_node(cpu));
6211
6212 if (!sg)
6213 goto fail;
6214
6215 sg_span = sched_group_cpus(sg);
6216 if (child->child) {
6217 child = child->child;
6218 cpumask_copy(sg_span, sched_domain_span(child));
6219 } else
6220 cpumask_set_cpu(i, sg_span);
6221
6222 cpumask_or(covered, covered, sg_span);
6223
6224 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
6225 if (atomic_inc_return(&sg->sgp->ref) == 1)
6226 build_group_mask(sd, sg);
6227
6228 /*
6229 * Initialize sgp->power such that even if we mess up the
6230 * domains and no possible iteration will get us here, we won't
6231 * die on a /0 trap.
6232 */
6233 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
6234
6235 /*
6236 * Make sure the first group of this domain contains the
6237 * canonical balance cpu. Otherwise the sched_domain iteration
6238 * breaks. See update_sg_lb_stats().
6239 */
6240 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6241 group_balance_cpu(sg) == cpu)
6242 groups = sg;
6243
6244 if (!first)
6245 first = sg;
6246 if (last)
6247 last->next = sg;
6248 last = sg;
6249 last->next = first;
6250 }
6251 sd->groups = groups;
6252
6253 return 0;
6254
6255fail:
6256 free_sched_groups(first, 0);
6257
6258 return -ENOMEM;
6259}
6260
6261static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6262{
6263 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6264 struct sched_domain *child = sd->child;
6265
6266 if (child)
6267 cpu = cpumask_first(sched_domain_span(child));
6268
6269 if (sg) {
6270 *sg = *per_cpu_ptr(sdd->sg, cpu);
6271 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
6272 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
6273 }
6274
6275 return cpu;
6276}
6277
6278/*
6279 * build_sched_groups will build a circular linked list of the groups
6280 * covered by the given span, and will set each group's ->cpumask correctly,
6281 * and ->cpu_power to 0.
6282 *
6283 * Assumes the sched_domain tree is fully constructed
6284 */
6285static int
6286build_sched_groups(struct sched_domain *sd, int cpu)
6287{
6288 struct sched_group *first = NULL, *last = NULL;
6289 struct sd_data *sdd = sd->private;
6290 const struct cpumask *span = sched_domain_span(sd);
6291 struct cpumask *covered;
6292 int i;
6293
6294 get_group(cpu, sdd, &sd->groups);
6295 atomic_inc(&sd->groups->ref);
6296
6297 if (cpu != cpumask_first(sched_domain_span(sd)))
6298 return 0;
6299
6300 lockdep_assert_held(&sched_domains_mutex);
6301 covered = sched_domains_tmpmask;
6302
6303 cpumask_clear(covered);
6304
6305 for_each_cpu(i, span) {
6306 struct sched_group *sg;
6307 int group = get_group(i, sdd, &sg);
6308 int j;
6309
6310 if (cpumask_test_cpu(i, covered))
6311 continue;
6312
6313 cpumask_clear(sched_group_cpus(sg));
6314 sg->sgp->power = 0;
6315 cpumask_setall(sched_group_mask(sg));
6316
6317 for_each_cpu(j, span) {
6318 if (get_group(j, sdd, NULL) != group)
6319 continue;
6320
6321 cpumask_set_cpu(j, covered);
6322 cpumask_set_cpu(j, sched_group_cpus(sg));
6323 }
6324
6325 if (!first)
6326 first = sg;
6327 if (last)
6328 last->next = sg;
6329 last = sg;
6330 }
6331 last->next = first;
6332
6333 return 0;
6334}
6335
6336/*
6337 * Initialize sched groups cpu_power.
6338 *
6339 * cpu_power indicates the capacity of sched group, which is used while
6340 * distributing the load between different sched groups in a sched domain.
6341 * Typically cpu_power for all the groups in a sched domain will be same unless
6342 * there are asymmetries in the topology. If there are asymmetries, group
6343 * having more cpu_power will pickup more load compared to the group having
6344 * less cpu_power.
6345 */
6346static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6347{
6348 struct sched_group *sg = sd->groups;
6349
6350 WARN_ON(!sd || !sg);
6351
6352 do {
6353 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6354 sg = sg->next;
6355 } while (sg != sd->groups);
6356
6357 if (cpu != group_balance_cpu(sg))
6358 return;
6359
6360 update_group_power(sd, cpu);
6361 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6362}
6363
6364int __weak arch_sd_sibling_asym_packing(void)
6365{
6366 return 0*SD_ASYM_PACKING;
6367}
6368
6369/*
6370 * Initializers for schedule domains
6371 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6372 */
6373
6374#ifdef CONFIG_SCHED_DEBUG
6375# define SD_INIT_NAME(sd, type) sd->name = #type
6376#else
6377# define SD_INIT_NAME(sd, type) do { } while (0)
6378#endif
6379
6380#define SD_INIT_FUNC(type) \
6381static noinline struct sched_domain * \
6382sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6383{ \
6384 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6385 *sd = SD_##type##_INIT; \
6386 SD_INIT_NAME(sd, type); \
6387 sd->private = &tl->data; \
6388 return sd; \
6389}
6390
6391SD_INIT_FUNC(CPU)
6392#ifdef CONFIG_SCHED_SMT
6393 SD_INIT_FUNC(SIBLING)
6394#endif
6395#ifdef CONFIG_SCHED_MC
6396 SD_INIT_FUNC(MC)
6397#endif
6398#ifdef CONFIG_SCHED_BOOK
6399 SD_INIT_FUNC(BOOK)
6400#endif
6401
6402static int default_relax_domain_level = -1;
6403int sched_domain_level_max;
6404
6405static int __init setup_relax_domain_level(char *str)
6406{
6407 if (kstrtoint(str, 0, &default_relax_domain_level))
6408 pr_warn("Unable to set relax_domain_level\n");
6409
6410 return 1;
6411}
6412__setup("relax_domain_level=", setup_relax_domain_level);
6413
6414static void set_domain_attribute(struct sched_domain *sd,
6415 struct sched_domain_attr *attr)
6416{
6417 int request;
6418
6419 if (!attr || attr->relax_domain_level < 0) {
6420 if (default_relax_domain_level < 0)
6421 return;
6422 else
6423 request = default_relax_domain_level;
6424 } else
6425 request = attr->relax_domain_level;
6426 if (request < sd->level) {
6427 /* turn off idle balance on this domain */
6428 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6429 } else {
6430 /* turn on idle balance on this domain */
6431 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6432 }
6433}
6434
6435static void __sdt_free(const struct cpumask *cpu_map);
6436static int __sdt_alloc(const struct cpumask *cpu_map);
6437
6438static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6439 const struct cpumask *cpu_map)
6440{
6441 switch (what) {
6442 case sa_rootdomain:
6443 if (!atomic_read(&d->rd->refcount))
6444 free_rootdomain(&d->rd->rcu); /* fall through */
6445 case sa_sd:
6446 free_percpu(d->sd); /* fall through */
6447 case sa_sd_storage:
6448 __sdt_free(cpu_map); /* fall through */
6449 case sa_none:
6450 break;
6451 }
6452}
6453
6454static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6455 const struct cpumask *cpu_map)
6456{
6457 memset(d, 0, sizeof(*d));
6458
6459 if (__sdt_alloc(cpu_map))
6460 return sa_sd_storage;
6461 d->sd = alloc_percpu(struct sched_domain *);
6462 if (!d->sd)
6463 return sa_sd_storage;
6464 d->rd = alloc_rootdomain();
6465 if (!d->rd)
6466 return sa_sd;
6467 return sa_rootdomain;
6468}
6469
6470/*
6471 * NULL the sd_data elements we've used to build the sched_domain and
6472 * sched_group structure so that the subsequent __free_domain_allocs()
6473 * will not free the data we're using.
6474 */
6475static void claim_allocations(int cpu, struct sched_domain *sd)
6476{
6477 struct sd_data *sdd = sd->private;
6478
6479 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6480 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6481
6482 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6483 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6484
6485 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6486 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6487}
6488
6489#ifdef CONFIG_SCHED_SMT
6490static const struct cpumask *cpu_smt_mask(int cpu)
6491{
6492 return topology_thread_cpumask(cpu);
6493}
6494#endif
6495
6496/*
6497 * Topology list, bottom-up.
6498 */
6499static struct sched_domain_topology_level default_topology[] = {
6500#ifdef CONFIG_SCHED_SMT
6501 { sd_init_SIBLING, cpu_smt_mask, },
6502#endif
6503#ifdef CONFIG_SCHED_MC
6504 { sd_init_MC, cpu_coregroup_mask, },
6505#endif
6506#ifdef CONFIG_SCHED_BOOK
6507 { sd_init_BOOK, cpu_book_mask, },
6508#endif
6509 { sd_init_CPU, cpu_cpu_mask, },
6510 { NULL, },
6511};
6512
6513static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6514
6515#ifdef CONFIG_NUMA
6516
6517static int sched_domains_numa_levels;
6518static int *sched_domains_numa_distance;
6519static struct cpumask ***sched_domains_numa_masks;
6520static int sched_domains_curr_level;
6521
6522static inline int sd_local_flags(int level)
6523{
6524 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6525 return 0;
6526
6527 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6528}
6529
6530static struct sched_domain *
6531sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6532{
6533 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6534 int level = tl->numa_level;
6535 int sd_weight = cpumask_weight(
6536 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6537
6538 *sd = (struct sched_domain){
6539 .min_interval = sd_weight,
6540 .max_interval = 2*sd_weight,
6541 .busy_factor = 32,
6542 .imbalance_pct = 125,
6543 .cache_nice_tries = 2,
6544 .busy_idx = 3,
6545 .idle_idx = 2,
6546 .newidle_idx = 0,
6547 .wake_idx = 0,
6548 .forkexec_idx = 0,
6549
6550 .flags = 1*SD_LOAD_BALANCE
6551 | 1*SD_BALANCE_NEWIDLE
6552 | 0*SD_BALANCE_EXEC
6553 | 0*SD_BALANCE_FORK
6554 | 0*SD_BALANCE_WAKE
6555 | 0*SD_WAKE_AFFINE
6556 | 0*SD_PREFER_LOCAL
6557 | 0*SD_SHARE_CPUPOWER
6558 | 0*SD_SHARE_PKG_RESOURCES
6559 | 1*SD_SERIALIZE
6560 | 0*SD_PREFER_SIBLING
6561 | sd_local_flags(level)
6562 ,
6563 .last_balance = jiffies,
6564 .balance_interval = sd_weight,
6565 };
6566 SD_INIT_NAME(sd, NUMA);
6567 sd->private = &tl->data;
6568
6569 /*
6570 * Ugly hack to pass state to sd_numa_mask()...
6571 */
6572 sched_domains_curr_level = tl->numa_level;
6573
6574 return sd;
6575}
6576
6577static const struct cpumask *sd_numa_mask(int cpu)
6578{
6579 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6580}
6581
6582static void sched_numa_warn(const char *str)
6583{
6584 static int done = false;
6585 int i,j;
6586
6587 if (done)
6588 return;
6589
6590 done = true;
6591
6592 printk(KERN_WARNING "ERROR: %s\n\n", str);
6593
6594 for (i = 0; i < nr_node_ids; i++) {
6595 printk(KERN_WARNING " ");
6596 for (j = 0; j < nr_node_ids; j++)
6597 printk(KERN_CONT "%02d ", node_distance(i,j));
6598 printk(KERN_CONT "\n");
6599 }
6600 printk(KERN_WARNING "\n");
6601}
6602
6603static bool find_numa_distance(int distance)
6604{
6605 int i;
6606
6607 if (distance == node_distance(0, 0))
6608 return true;
6609
6610 for (i = 0; i < sched_domains_numa_levels; i++) {
6611 if (sched_domains_numa_distance[i] == distance)
6612 return true;
6613 }
6614
6615 return false;
6616}
6617
6618static void sched_init_numa(void)
6619{
6620 int next_distance, curr_distance = node_distance(0, 0);
6621 struct sched_domain_topology_level *tl;
6622 int level = 0;
6623 int i, j, k;
6624
6625 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6626 if (!sched_domains_numa_distance)
6627 return;
6628
6629 /*
6630 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6631 * unique distances in the node_distance() table.
6632 *
6633 * Assumes node_distance(0,j) includes all distances in
6634 * node_distance(i,j) in order to avoid cubic time.
6635 */
6636 next_distance = curr_distance;
6637 for (i = 0; i < nr_node_ids; i++) {
6638 for (j = 0; j < nr_node_ids; j++) {
6639 for (k = 0; k < nr_node_ids; k++) {
6640 int distance = node_distance(i, k);
6641
6642 if (distance > curr_distance &&
6643 (distance < next_distance ||
6644 next_distance == curr_distance))
6645 next_distance = distance;
6646
6647 /*
6648 * While not a strong assumption it would be nice to know
6649 * about cases where if node A is connected to B, B is not
6650 * equally connected to A.
6651 */
6652 if (sched_debug() && node_distance(k, i) != distance)
6653 sched_numa_warn("Node-distance not symmetric");
6654
6655 if (sched_debug() && i && !find_numa_distance(distance))
6656 sched_numa_warn("Node-0 not representative");
6657 }
6658 if (next_distance != curr_distance) {
6659 sched_domains_numa_distance[level++] = next_distance;
6660 sched_domains_numa_levels = level;
6661 curr_distance = next_distance;
6662 } else break;
6663 }
6664
6665 /*
6666 * In case of sched_debug() we verify the above assumption.
6667 */
6668 if (!sched_debug())
6669 break;
6670 }
6671 /*
6672 * 'level' contains the number of unique distances, excluding the
6673 * identity distance node_distance(i,i).
6674 *
6675 * The sched_domains_nume_distance[] array includes the actual distance
6676 * numbers.
6677 */
6678
6679 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6680 if (!sched_domains_numa_masks)
6681 return;
6682
6683 /*
6684 * Now for each level, construct a mask per node which contains all
6685 * cpus of nodes that are that many hops away from us.
6686 */
6687 for (i = 0; i < level; i++) {
6688 sched_domains_numa_masks[i] =
6689 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6690 if (!sched_domains_numa_masks[i])
6691 return;
6692
6693 for (j = 0; j < nr_node_ids; j++) {
6694 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6695 if (!mask)
6696 return;
6697
6698 sched_domains_numa_masks[i][j] = mask;
6699
6700 for (k = 0; k < nr_node_ids; k++) {
6701 if (node_distance(j, k) > sched_domains_numa_distance[i])
6702 continue;
6703
6704 cpumask_or(mask, mask, cpumask_of_node(k));
6705 }
6706 }
6707 }
6708
6709 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6710 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6711 if (!tl)
6712 return;
6713
6714 /*
6715 * Copy the default topology bits..
6716 */
6717 for (i = 0; default_topology[i].init; i++)
6718 tl[i] = default_topology[i];
6719
6720 /*
6721 * .. and append 'j' levels of NUMA goodness.
6722 */
6723 for (j = 0; j < level; i++, j++) {
6724 tl[i] = (struct sched_domain_topology_level){
6725 .init = sd_numa_init,
6726 .mask = sd_numa_mask,
6727 .flags = SDTL_OVERLAP,
6728 .numa_level = j,
6729 };
6730 }
6731
6732 sched_domain_topology = tl;
6733}
6734#else
6735static inline void sched_init_numa(void)
6736{
6737}
6738#endif /* CONFIG_NUMA */
6739
6740static int __sdt_alloc(const struct cpumask *cpu_map)
6741{
6742 struct sched_domain_topology_level *tl;
6743 int j;
6744
6745 for (tl = sched_domain_topology; tl->init; tl++) {
6746 struct sd_data *sdd = &tl->data;
6747
6748 sdd->sd = alloc_percpu(struct sched_domain *);
6749 if (!sdd->sd)
6750 return -ENOMEM;
6751
6752 sdd->sg = alloc_percpu(struct sched_group *);
6753 if (!sdd->sg)
6754 return -ENOMEM;
6755
6756 sdd->sgp = alloc_percpu(struct sched_group_power *);
6757 if (!sdd->sgp)
6758 return -ENOMEM;
6759
6760 for_each_cpu(j, cpu_map) {
6761 struct sched_domain *sd;
6762 struct sched_group *sg;
6763 struct sched_group_power *sgp;
6764
6765 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6766 GFP_KERNEL, cpu_to_node(j));
6767 if (!sd)
6768 return -ENOMEM;
6769
6770 *per_cpu_ptr(sdd->sd, j) = sd;
6771
6772 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6773 GFP_KERNEL, cpu_to_node(j));
6774 if (!sg)
6775 return -ENOMEM;
6776
6777 sg->next = sg;
6778
6779 *per_cpu_ptr(sdd->sg, j) = sg;
6780
6781 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6782 GFP_KERNEL, cpu_to_node(j));
6783 if (!sgp)
6784 return -ENOMEM;
6785
6786 *per_cpu_ptr(sdd->sgp, j) = sgp;
6787 }
6788 }
6789
6790 return 0;
6791}
6792
6793static void __sdt_free(const struct cpumask *cpu_map)
6794{
6795 struct sched_domain_topology_level *tl;
6796 int j;
6797
6798 for (tl = sched_domain_topology; tl->init; tl++) {
6799 struct sd_data *sdd = &tl->data;
6800
6801 for_each_cpu(j, cpu_map) {
6802 struct sched_domain *sd;
6803
6804 if (sdd->sd) {
6805 sd = *per_cpu_ptr(sdd->sd, j);
6806 if (sd && (sd->flags & SD_OVERLAP))
6807 free_sched_groups(sd->groups, 0);
6808 kfree(*per_cpu_ptr(sdd->sd, j));
6809 }
6810
6811 if (sdd->sg)
6812 kfree(*per_cpu_ptr(sdd->sg, j));
6813 if (sdd->sgp)
6814 kfree(*per_cpu_ptr(sdd->sgp, j));
6815 }
6816 free_percpu(sdd->sd);
6817 sdd->sd = NULL;
6818 free_percpu(sdd->sg);
6819 sdd->sg = NULL;
6820 free_percpu(sdd->sgp);
6821 sdd->sgp = NULL;
6822 }
6823}
6824
6825struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6826 struct s_data *d, const struct cpumask *cpu_map,
6827 struct sched_domain_attr *attr, struct sched_domain *child,
6828 int cpu)
6829{
6830 struct sched_domain *sd = tl->init(tl, cpu);
6831 if (!sd)
6832 return child;
6833
6834 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6835 if (child) {
6836 sd->level = child->level + 1;
6837 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6838 child->parent = sd;
6839 }
6840 sd->child = child;
6841 set_domain_attribute(sd, attr);
6842
6843 return sd;
6844}
6845
6846/*
6847 * Build sched domains for a given set of cpus and attach the sched domains
6848 * to the individual cpus
6849 */
6850static int build_sched_domains(const struct cpumask *cpu_map,
6851 struct sched_domain_attr *attr)
6852{
6853 enum s_alloc alloc_state = sa_none;
6854 struct sched_domain *sd;
6855 struct s_data d;
6856 int i, ret = -ENOMEM;
6857
6858 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6859 if (alloc_state != sa_rootdomain)
6860 goto error;
6861
6862 /* Set up domains for cpus specified by the cpu_map. */
6863 for_each_cpu(i, cpu_map) {
6864 struct sched_domain_topology_level *tl;
6865
6866 sd = NULL;
6867 for (tl = sched_domain_topology; tl->init; tl++) {
6868 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6869 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6870 sd->flags |= SD_OVERLAP;
6871 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6872 break;
6873 }
6874
6875 while (sd->child)
6876 sd = sd->child;
6877
6878 *per_cpu_ptr(d.sd, i) = sd;
6879 }
6880
6881 /* Build the groups for the domains */
6882 for_each_cpu(i, cpu_map) {
6883 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6884 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6885 if (sd->flags & SD_OVERLAP) {
6886 if (build_overlap_sched_groups(sd, i))
6887 goto error;
6888 } else {
6889 if (build_sched_groups(sd, i))
6890 goto error;
6891 }
6892 }
6893 }
6894
6895 /* Calculate CPU power for physical packages and nodes */
6896 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6897 if (!cpumask_test_cpu(i, cpu_map))
6898 continue;
6899
6900 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6901 claim_allocations(i, sd);
6902 init_sched_groups_power(i, sd);
6903 }
6904 }
6905
6906 /* Attach the domains */
6907 rcu_read_lock();
6908 for_each_cpu(i, cpu_map) {
6909 sd = *per_cpu_ptr(d.sd, i);
6910 cpu_attach_domain(sd, d.rd, i);
6911 }
6912 rcu_read_unlock();
6913
6914 ret = 0;
6915error:
6916 __free_domain_allocs(&d, alloc_state, cpu_map);
6917 return ret;
6918}
6919
6920static cpumask_var_t *doms_cur; /* current sched domains */
6921static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6922static struct sched_domain_attr *dattr_cur;
6923 /* attribues of custom domains in 'doms_cur' */
6924
6925/*
6926 * Special case: If a kmalloc of a doms_cur partition (array of
6927 * cpumask) fails, then fallback to a single sched domain,
6928 * as determined by the single cpumask fallback_doms.
6929 */
6930static cpumask_var_t fallback_doms;
6931
6932/*
6933 * arch_update_cpu_topology lets virtualized architectures update the
6934 * cpu core maps. It is supposed to return 1 if the topology changed
6935 * or 0 if it stayed the same.
6936 */
6937int __attribute__((weak)) arch_update_cpu_topology(void)
6938{
6939 return 0;
6940}
6941
6942cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6943{
6944 int i;
6945 cpumask_var_t *doms;
6946
6947 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6948 if (!doms)
6949 return NULL;
6950 for (i = 0; i < ndoms; i++) {
6951 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6952 free_sched_domains(doms, i);
6953 return NULL;
6954 }
6955 }
6956 return doms;
6957}
6958
6959void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6960{
6961 unsigned int i;
6962 for (i = 0; i < ndoms; i++)
6963 free_cpumask_var(doms[i]);
6964 kfree(doms);
6965}
6966
6967/*
6968 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6969 * For now this just excludes isolated cpus, but could be used to
6970 * exclude other special cases in the future.
6971 */
6972static int init_sched_domains(const struct cpumask *cpu_map)
6973{
6974 int err;
6975
6976 arch_update_cpu_topology();
6977 ndoms_cur = 1;
6978 doms_cur = alloc_sched_domains(ndoms_cur);
6979 if (!doms_cur)
6980 doms_cur = &fallback_doms;
6981 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6982 err = build_sched_domains(doms_cur[0], NULL);
6983 register_sched_domain_sysctl();
6984
6985 return err;
6986}
6987
6988/*
6989 * Detach sched domains from a group of cpus specified in cpu_map
6990 * These cpus will now be attached to the NULL domain
6991 */
6992static void detach_destroy_domains(const struct cpumask *cpu_map)
6993{
6994 int i;
6995
6996 rcu_read_lock();
6997 for_each_cpu(i, cpu_map)
6998 cpu_attach_domain(NULL, &def_root_domain, i);
6999 rcu_read_unlock();
7000}
7001
7002/* handle null as "default" */
7003static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7004 struct sched_domain_attr *new, int idx_new)
7005{
7006 struct sched_domain_attr tmp;
7007
7008 /* fast path */
7009 if (!new && !cur)
7010 return 1;
7011
7012 tmp = SD_ATTR_INIT;
7013 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7014 new ? (new + idx_new) : &tmp,
7015 sizeof(struct sched_domain_attr));
7016}
7017
7018/*
7019 * Partition sched domains as specified by the 'ndoms_new'
7020 * cpumasks in the array doms_new[] of cpumasks. This compares
7021 * doms_new[] to the current sched domain partitioning, doms_cur[].
7022 * It destroys each deleted domain and builds each new domain.
7023 *
7024 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7025 * The masks don't intersect (don't overlap.) We should setup one
7026 * sched domain for each mask. CPUs not in any of the cpumasks will
7027 * not be load balanced. If the same cpumask appears both in the
7028 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7029 * it as it is.
7030 *
7031 * The passed in 'doms_new' should be allocated using
7032 * alloc_sched_domains. This routine takes ownership of it and will
7033 * free_sched_domains it when done with it. If the caller failed the
7034 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7035 * and partition_sched_domains() will fallback to the single partition
7036 * 'fallback_doms', it also forces the domains to be rebuilt.
7037 *
7038 * If doms_new == NULL it will be replaced with cpu_online_mask.
7039 * ndoms_new == 0 is a special case for destroying existing domains,
7040 * and it will not create the default domain.
7041 *
7042 * Call with hotplug lock held
7043 */
7044void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7045 struct sched_domain_attr *dattr_new)
7046{
7047 int i, j, n;
7048 int new_topology;
7049
7050 mutex_lock(&sched_domains_mutex);
7051
7052 /* always unregister in case we don't destroy any domains */
7053 unregister_sched_domain_sysctl();
7054
7055 /* Let architecture update cpu core mappings. */
7056 new_topology = arch_update_cpu_topology();
7057
7058 n = doms_new ? ndoms_new : 0;
7059
7060 /* Destroy deleted domains */
7061 for (i = 0; i < ndoms_cur; i++) {
7062 for (j = 0; j < n && !new_topology; j++) {
7063 if (cpumask_equal(doms_cur[i], doms_new[j])
7064 && dattrs_equal(dattr_cur, i, dattr_new, j))
7065 goto match1;
7066 }
7067 /* no match - a current sched domain not in new doms_new[] */
7068 detach_destroy_domains(doms_cur[i]);
7069match1:
7070 ;
7071 }
7072
7073 if (doms_new == NULL) {
7074 ndoms_cur = 0;
7075 doms_new = &fallback_doms;
7076 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7077 WARN_ON_ONCE(dattr_new);
7078 }
7079
7080 /* Build new domains */
7081 for (i = 0; i < ndoms_new; i++) {
7082 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7083 if (cpumask_equal(doms_new[i], doms_cur[j])
7084 && dattrs_equal(dattr_new, i, dattr_cur, j))
7085 goto match2;
7086 }
7087 /* no match - add a new doms_new */
7088 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7089match2:
7090 ;
7091 }
7092
7093 /* Remember the new sched domains */
7094 if (doms_cur != &fallback_doms)
7095 free_sched_domains(doms_cur, ndoms_cur);
7096 kfree(dattr_cur); /* kfree(NULL) is safe */
7097 doms_cur = doms_new;
7098 dattr_cur = dattr_new;
7099 ndoms_cur = ndoms_new;
7100
7101 register_sched_domain_sysctl();
7102
7103 mutex_unlock(&sched_domains_mutex);
7104}
7105
7106/*
7107 * Update cpusets according to cpu_active mask. If cpusets are
7108 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7109 * around partition_sched_domains().
7110 */
7111static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7112 void *hcpu)
7113{
7114 switch (action & ~CPU_TASKS_FROZEN) {
7115 case CPU_ONLINE:
7116 case CPU_DOWN_FAILED:
7117 cpuset_update_active_cpus();
7118 return NOTIFY_OK;
7119 default:
7120 return NOTIFY_DONE;
7121 }
7122}
7123
7124static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7125 void *hcpu)
7126{
7127 switch (action & ~CPU_TASKS_FROZEN) {
7128 case CPU_DOWN_PREPARE:
7129 cpuset_update_active_cpus();
7130 return NOTIFY_OK;
7131 default:
7132 return NOTIFY_DONE;
7133 }
7134}
7135
7136void __init sched_init_smp(void)
7137{
7138 cpumask_var_t non_isolated_cpus;
7139
7140 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7141 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7142
7143 sched_init_numa();
7144
7145 get_online_cpus();
7146 mutex_lock(&sched_domains_mutex);
7147 init_sched_domains(cpu_active_mask);
7148 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7149 if (cpumask_empty(non_isolated_cpus))
7150 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7151 mutex_unlock(&sched_domains_mutex);
7152 put_online_cpus();
7153
7154 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7155 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7156
7157 /* RT runtime code needs to handle some hotplug events */
7158 hotcpu_notifier(update_runtime, 0);
7159
7160 init_hrtick();
7161
7162 /* Move init over to a non-isolated CPU */
7163 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7164 BUG();
7165 sched_init_granularity();
7166 free_cpumask_var(non_isolated_cpus);
7167
7168 init_sched_rt_class();
7169}
7170#else
7171void __init sched_init_smp(void)
7172{
7173 sched_init_granularity();
7174}
7175#endif /* CONFIG_SMP */
7176
7177const_debug unsigned int sysctl_timer_migration = 1;
7178
7179int in_sched_functions(unsigned long addr)
7180{
7181 return in_lock_functions(addr) ||
7182 (addr >= (unsigned long)__sched_text_start
7183 && addr < (unsigned long)__sched_text_end);
7184}
7185
7186#ifdef CONFIG_CGROUP_SCHED
7187struct task_group root_task_group;
7188LIST_HEAD(task_groups);
7189#endif
7190
7191DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
7192
7193void __init sched_init(void)
7194{
7195 int i, j;
7196 unsigned long alloc_size = 0, ptr;
7197
7198#ifdef CONFIG_FAIR_GROUP_SCHED
7199 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7200#endif
7201#ifdef CONFIG_RT_GROUP_SCHED
7202 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7203#endif
7204#ifdef CONFIG_CPUMASK_OFFSTACK
7205 alloc_size += num_possible_cpus() * cpumask_size();
7206#endif
7207 if (alloc_size) {
7208 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7209
7210#ifdef CONFIG_FAIR_GROUP_SCHED
7211 root_task_group.se = (struct sched_entity **)ptr;
7212 ptr += nr_cpu_ids * sizeof(void **);
7213
7214 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7215 ptr += nr_cpu_ids * sizeof(void **);
7216
7217#endif /* CONFIG_FAIR_GROUP_SCHED */
7218#ifdef CONFIG_RT_GROUP_SCHED
7219 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7220 ptr += nr_cpu_ids * sizeof(void **);
7221
7222 root_task_group.rt_rq = (struct rt_rq **)ptr;
7223 ptr += nr_cpu_ids * sizeof(void **);
7224
7225#endif /* CONFIG_RT_GROUP_SCHED */
7226#ifdef CONFIG_CPUMASK_OFFSTACK
7227 for_each_possible_cpu(i) {
7228 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7229 ptr += cpumask_size();
7230 }
7231#endif /* CONFIG_CPUMASK_OFFSTACK */
7232 }
7233
7234#ifdef CONFIG_SMP
7235 init_defrootdomain();
7236#endif
7237
7238 init_rt_bandwidth(&def_rt_bandwidth,
7239 global_rt_period(), global_rt_runtime());
7240
7241#ifdef CONFIG_RT_GROUP_SCHED
7242 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7243 global_rt_period(), global_rt_runtime());
7244#endif /* CONFIG_RT_GROUP_SCHED */
7245
7246#ifdef CONFIG_CGROUP_SCHED
7247 list_add(&root_task_group.list, &task_groups);
7248 INIT_LIST_HEAD(&root_task_group.children);
7249 INIT_LIST_HEAD(&root_task_group.siblings);
7250 autogroup_init(&init_task);
7251
7252#endif /* CONFIG_CGROUP_SCHED */
7253
7254#ifdef CONFIG_CGROUP_CPUACCT
7255 root_cpuacct.cpustat = &kernel_cpustat;
7256 root_cpuacct.cpuusage = alloc_percpu(u64);
7257 /* Too early, not expected to fail */
7258 BUG_ON(!root_cpuacct.cpuusage);
7259#endif
7260 for_each_possible_cpu(i) {
7261 struct rq *rq;
7262
7263 rq = cpu_rq(i);
7264 raw_spin_lock_init(&rq->lock);
7265 rq->nr_running = 0;
7266 rq->calc_load_active = 0;
7267 rq->calc_load_update = jiffies + LOAD_FREQ;
7268 init_cfs_rq(&rq->cfs);
7269 init_rt_rq(&rq->rt, rq);
7270#ifdef CONFIG_FAIR_GROUP_SCHED
7271 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7272 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7273 /*
7274 * How much cpu bandwidth does root_task_group get?
7275 *
7276 * In case of task-groups formed thr' the cgroup filesystem, it
7277 * gets 100% of the cpu resources in the system. This overall
7278 * system cpu resource is divided among the tasks of
7279 * root_task_group and its child task-groups in a fair manner,
7280 * based on each entity's (task or task-group's) weight
7281 * (se->load.weight).
7282 *
7283 * In other words, if root_task_group has 10 tasks of weight
7284 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7285 * then A0's share of the cpu resource is:
7286 *
7287 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7288 *
7289 * We achieve this by letting root_task_group's tasks sit
7290 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7291 */
7292 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7293 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7294#endif /* CONFIG_FAIR_GROUP_SCHED */
7295
7296 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7297#ifdef CONFIG_RT_GROUP_SCHED
7298 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7299 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7300#endif
7301
7302 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7303 rq->cpu_load[j] = 0;
7304
7305 rq->last_load_update_tick = jiffies;
7306
7307#ifdef CONFIG_SMP
7308 rq->sd = NULL;
7309 rq->rd = NULL;
7310 rq->cpu_power = SCHED_POWER_SCALE;
7311 rq->post_schedule = 0;
7312 rq->active_balance = 0;
7313 rq->next_balance = jiffies;
7314 rq->push_cpu = 0;
7315 rq->cpu = i;
7316 rq->online = 0;
7317 rq->idle_stamp = 0;
7318 rq->avg_idle = 2*sysctl_sched_migration_cost;
7319
7320 INIT_LIST_HEAD(&rq->cfs_tasks);
7321
7322 rq_attach_root(rq, &def_root_domain);
7323#ifdef CONFIG_NO_HZ
7324 rq->nohz_flags = 0;
7325#endif
7326#endif
7327 init_rq_hrtick(rq);
7328 atomic_set(&rq->nr_iowait, 0);
7329 }
7330
7331 set_load_weight(&init_task);
7332
7333#ifdef CONFIG_PREEMPT_NOTIFIERS
7334 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7335#endif
7336
7337#ifdef CONFIG_RT_MUTEXES
7338 plist_head_init(&init_task.pi_waiters);
7339#endif
7340
7341 /*
7342 * The boot idle thread does lazy MMU switching as well:
7343 */
7344 atomic_inc(&init_mm.mm_count);
7345 enter_lazy_tlb(&init_mm, current);
7346
7347 /*
7348 * Make us the idle thread. Technically, schedule() should not be
7349 * called from this thread, however somewhere below it might be,
7350 * but because we are the idle thread, we just pick up running again
7351 * when this runqueue becomes "idle".
7352 */
7353 init_idle(current, smp_processor_id());
7354
7355 calc_load_update = jiffies + LOAD_FREQ;
7356
7357 /*
7358 * During early bootup we pretend to be a normal task:
7359 */
7360 current->sched_class = &fair_sched_class;
7361
7362#ifdef CONFIG_SMP
7363 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7364 /* May be allocated at isolcpus cmdline parse time */
7365 if (cpu_isolated_map == NULL)
7366 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7367 idle_thread_set_boot_cpu();
7368#endif
7369 init_sched_fair_class();
7370
7371 scheduler_running = 1;
7372}
7373
7374#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7375static inline int preempt_count_equals(int preempt_offset)
7376{
7377 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7378
7379 return (nested == preempt_offset);
7380}
7381
7382void __might_sleep(const char *file, int line, int preempt_offset)
7383{
7384 static unsigned long prev_jiffy; /* ratelimiting */
7385
7386 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7387 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7388 system_state != SYSTEM_RUNNING || oops_in_progress)
7389 return;
7390 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7391 return;
7392 prev_jiffy = jiffies;
7393
7394 printk(KERN_ERR
7395 "BUG: sleeping function called from invalid context at %s:%d\n",
7396 file, line);
7397 printk(KERN_ERR
7398 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7399 in_atomic(), irqs_disabled(),
7400 current->pid, current->comm);
7401
7402 debug_show_held_locks(current);
7403 if (irqs_disabled())
7404 print_irqtrace_events(current);
7405 dump_stack();
7406}
7407EXPORT_SYMBOL(__might_sleep);
7408#endif
7409
7410#ifdef CONFIG_MAGIC_SYSRQ
7411static void normalize_task(struct rq *rq, struct task_struct *p)
7412{
7413 const struct sched_class *prev_class = p->sched_class;
7414 int old_prio = p->prio;
7415 int on_rq;
7416
7417 on_rq = p->on_rq;
7418 if (on_rq)
7419 dequeue_task(rq, p, 0);
7420 __setscheduler(rq, p, SCHED_NORMAL, 0);
7421 if (on_rq) {
7422 enqueue_task(rq, p, 0);
7423 resched_task(rq->curr);
7424 }
7425
7426 check_class_changed(rq, p, prev_class, old_prio);
7427}
7428
7429void normalize_rt_tasks(void)
7430{
7431 struct task_struct *g, *p;
7432 unsigned long flags;
7433 struct rq *rq;
7434
7435 read_lock_irqsave(&tasklist_lock, flags);
7436 do_each_thread(g, p) {
7437 /*
7438 * Only normalize user tasks:
7439 */
7440 if (!p->mm)
7441 continue;
7442
7443 p->se.exec_start = 0;
7444#ifdef CONFIG_SCHEDSTATS
7445 p->se.statistics.wait_start = 0;
7446 p->se.statistics.sleep_start = 0;
7447 p->se.statistics.block_start = 0;
7448#endif
7449
7450 if (!rt_task(p)) {
7451 /*
7452 * Renice negative nice level userspace
7453 * tasks back to 0:
7454 */
7455 if (TASK_NICE(p) < 0 && p->mm)
7456 set_user_nice(p, 0);
7457 continue;
7458 }
7459
7460 raw_spin_lock(&p->pi_lock);
7461 rq = __task_rq_lock(p);
7462
7463 normalize_task(rq, p);
7464
7465 __task_rq_unlock(rq);
7466 raw_spin_unlock(&p->pi_lock);
7467 } while_each_thread(g, p);
7468
7469 read_unlock_irqrestore(&tasklist_lock, flags);
7470}
7471
7472#endif /* CONFIG_MAGIC_SYSRQ */
7473
7474#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7475/*
7476 * These functions are only useful for the IA64 MCA handling, or kdb.
7477 *
7478 * They can only be called when the whole system has been
7479 * stopped - every CPU needs to be quiescent, and no scheduling
7480 * activity can take place. Using them for anything else would
7481 * be a serious bug, and as a result, they aren't even visible
7482 * under any other configuration.
7483 */
7484
7485/**
7486 * curr_task - return the current task for a given cpu.
7487 * @cpu: the processor in question.
7488 *
7489 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7490 */
7491struct task_struct *curr_task(int cpu)
7492{
7493 return cpu_curr(cpu);
7494}
7495
7496#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7497
7498#ifdef CONFIG_IA64
7499/**
7500 * set_curr_task - set the current task for a given cpu.
7501 * @cpu: the processor in question.
7502 * @p: the task pointer to set.
7503 *
7504 * Description: This function must only be used when non-maskable interrupts
7505 * are serviced on a separate stack. It allows the architecture to switch the
7506 * notion of the current task on a cpu in a non-blocking manner. This function
7507 * must be called with all CPU's synchronized, and interrupts disabled, the
7508 * and caller must save the original value of the current task (see
7509 * curr_task() above) and restore that value before reenabling interrupts and
7510 * re-starting the system.
7511 *
7512 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7513 */
7514void set_curr_task(int cpu, struct task_struct *p)
7515{
7516 cpu_curr(cpu) = p;
7517}
7518
7519#endif
7520
7521#ifdef CONFIG_CGROUP_SCHED
7522/* task_group_lock serializes the addition/removal of task groups */
7523static DEFINE_SPINLOCK(task_group_lock);
7524
7525static void free_sched_group(struct task_group *tg)
7526{
7527 free_fair_sched_group(tg);
7528 free_rt_sched_group(tg);
7529 autogroup_free(tg);
7530 kfree(tg);
7531}
7532
7533/* allocate runqueue etc for a new task group */
7534struct task_group *sched_create_group(struct task_group *parent)
7535{
7536 struct task_group *tg;
7537 unsigned long flags;
7538
7539 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7540 if (!tg)
7541 return ERR_PTR(-ENOMEM);
7542
7543 if (!alloc_fair_sched_group(tg, parent))
7544 goto err;
7545
7546 if (!alloc_rt_sched_group(tg, parent))
7547 goto err;
7548
7549 spin_lock_irqsave(&task_group_lock, flags);
7550 list_add_rcu(&tg->list, &task_groups);
7551
7552 WARN_ON(!parent); /* root should already exist */
7553
7554 tg->parent = parent;
7555 INIT_LIST_HEAD(&tg->children);
7556 list_add_rcu(&tg->siblings, &parent->children);
7557 spin_unlock_irqrestore(&task_group_lock, flags);
7558
7559 return tg;
7560
7561err:
7562 free_sched_group(tg);
7563 return ERR_PTR(-ENOMEM);
7564}
7565
7566/* rcu callback to free various structures associated with a task group */
7567static void free_sched_group_rcu(struct rcu_head *rhp)
7568{
7569 /* now it should be safe to free those cfs_rqs */
7570 free_sched_group(container_of(rhp, struct task_group, rcu));
7571}
7572
7573/* Destroy runqueue etc associated with a task group */
7574void sched_destroy_group(struct task_group *tg)
7575{
7576 unsigned long flags;
7577 int i;
7578
7579 /* end participation in shares distribution */
7580 for_each_possible_cpu(i)
7581 unregister_fair_sched_group(tg, i);
7582
7583 spin_lock_irqsave(&task_group_lock, flags);
7584 list_del_rcu(&tg->list);
7585 list_del_rcu(&tg->siblings);
7586 spin_unlock_irqrestore(&task_group_lock, flags);
7587
7588 /* wait for possible concurrent references to cfs_rqs complete */
7589 call_rcu(&tg->rcu, free_sched_group_rcu);
7590}
7591
7592/* change task's runqueue when it moves between groups.
7593 * The caller of this function should have put the task in its new group
7594 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7595 * reflect its new group.
7596 */
7597void sched_move_task(struct task_struct *tsk)
7598{
7599 struct task_group *tg;
7600 int on_rq, running;
7601 unsigned long flags;
7602 struct rq *rq;
7603
7604 rq = task_rq_lock(tsk, &flags);
7605
7606 running = task_current(rq, tsk);
7607 on_rq = tsk->on_rq;
7608
7609 if (on_rq)
7610 dequeue_task(rq, tsk, 0);
7611 if (unlikely(running))
7612 tsk->sched_class->put_prev_task(rq, tsk);
7613
7614 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7615 lockdep_is_held(&tsk->sighand->siglock)),
7616 struct task_group, css);
7617 tg = autogroup_task_group(tsk, tg);
7618 tsk->sched_task_group = tg;
7619
7620#ifdef CONFIG_FAIR_GROUP_SCHED
7621 if (tsk->sched_class->task_move_group)
7622 tsk->sched_class->task_move_group(tsk, on_rq);
7623 else
7624#endif
7625 set_task_rq(tsk, task_cpu(tsk));
7626
7627 if (unlikely(running))
7628 tsk->sched_class->set_curr_task(rq);
7629 if (on_rq)
7630 enqueue_task(rq, tsk, 0);
7631
7632 task_rq_unlock(rq, tsk, &flags);
7633}
7634#endif /* CONFIG_CGROUP_SCHED */
7635
7636#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7637static unsigned long to_ratio(u64 period, u64 runtime)
7638{
7639 if (runtime == RUNTIME_INF)
7640 return 1ULL << 20;
7641
7642 return div64_u64(runtime << 20, period);
7643}
7644#endif
7645
7646#ifdef CONFIG_RT_GROUP_SCHED
7647/*
7648 * Ensure that the real time constraints are schedulable.
7649 */
7650static DEFINE_MUTEX(rt_constraints_mutex);
7651
7652/* Must be called with tasklist_lock held */
7653static inline int tg_has_rt_tasks(struct task_group *tg)
7654{
7655 struct task_struct *g, *p;
7656
7657 do_each_thread(g, p) {
7658 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7659 return 1;
7660 } while_each_thread(g, p);
7661
7662 return 0;
7663}
7664
7665struct rt_schedulable_data {
7666 struct task_group *tg;
7667 u64 rt_period;
7668 u64 rt_runtime;
7669};
7670
7671static int tg_rt_schedulable(struct task_group *tg, void *data)
7672{
7673 struct rt_schedulable_data *d = data;
7674 struct task_group *child;
7675 unsigned long total, sum = 0;
7676 u64 period, runtime;
7677
7678 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7679 runtime = tg->rt_bandwidth.rt_runtime;
7680
7681 if (tg == d->tg) {
7682 period = d->rt_period;
7683 runtime = d->rt_runtime;
7684 }
7685
7686 /*
7687 * Cannot have more runtime than the period.
7688 */
7689 if (runtime > period && runtime != RUNTIME_INF)
7690 return -EINVAL;
7691
7692 /*
7693 * Ensure we don't starve existing RT tasks.
7694 */
7695 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7696 return -EBUSY;
7697
7698 total = to_ratio(period, runtime);
7699
7700 /*
7701 * Nobody can have more than the global setting allows.
7702 */
7703 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7704 return -EINVAL;
7705
7706 /*
7707 * The sum of our children's runtime should not exceed our own.
7708 */
7709 list_for_each_entry_rcu(child, &tg->children, siblings) {
7710 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7711 runtime = child->rt_bandwidth.rt_runtime;
7712
7713 if (child == d->tg) {
7714 period = d->rt_period;
7715 runtime = d->rt_runtime;
7716 }
7717
7718 sum += to_ratio(period, runtime);
7719 }
7720
7721 if (sum > total)
7722 return -EINVAL;
7723
7724 return 0;
7725}
7726
7727static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7728{
7729 int ret;
7730
7731 struct rt_schedulable_data data = {
7732 .tg = tg,
7733 .rt_period = period,
7734 .rt_runtime = runtime,
7735 };
7736
7737 rcu_read_lock();
7738 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7739 rcu_read_unlock();
7740
7741 return ret;
7742}
7743
7744static int tg_set_rt_bandwidth(struct task_group *tg,
7745 u64 rt_period, u64 rt_runtime)
7746{
7747 int i, err = 0;
7748
7749 mutex_lock(&rt_constraints_mutex);
7750 read_lock(&tasklist_lock);
7751 err = __rt_schedulable(tg, rt_period, rt_runtime);
7752 if (err)
7753 goto unlock;
7754
7755 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7756 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7757 tg->rt_bandwidth.rt_runtime = rt_runtime;
7758
7759 for_each_possible_cpu(i) {
7760 struct rt_rq *rt_rq = tg->rt_rq[i];
7761
7762 raw_spin_lock(&rt_rq->rt_runtime_lock);
7763 rt_rq->rt_runtime = rt_runtime;
7764 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7765 }
7766 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7767unlock:
7768 read_unlock(&tasklist_lock);
7769 mutex_unlock(&rt_constraints_mutex);
7770
7771 return err;
7772}
7773
7774int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7775{
7776 u64 rt_runtime, rt_period;
7777
7778 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7779 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7780 if (rt_runtime_us < 0)
7781 rt_runtime = RUNTIME_INF;
7782
7783 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7784}
7785
7786long sched_group_rt_runtime(struct task_group *tg)
7787{
7788 u64 rt_runtime_us;
7789
7790 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7791 return -1;
7792
7793 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7794 do_div(rt_runtime_us, NSEC_PER_USEC);
7795 return rt_runtime_us;
7796}
7797
7798int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7799{
7800 u64 rt_runtime, rt_period;
7801
7802 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7803 rt_runtime = tg->rt_bandwidth.rt_runtime;
7804
7805 if (rt_period == 0)
7806 return -EINVAL;
7807
7808 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7809}
7810
7811long sched_group_rt_period(struct task_group *tg)
7812{
7813 u64 rt_period_us;
7814
7815 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7816 do_div(rt_period_us, NSEC_PER_USEC);
7817 return rt_period_us;
7818}
7819
7820static int sched_rt_global_constraints(void)
7821{
7822 u64 runtime, period;
7823 int ret = 0;
7824
7825 if (sysctl_sched_rt_period <= 0)
7826 return -EINVAL;
7827
7828 runtime = global_rt_runtime();
7829 period = global_rt_period();
7830
7831 /*
7832 * Sanity check on the sysctl variables.
7833 */
7834 if (runtime > period && runtime != RUNTIME_INF)
7835 return -EINVAL;
7836
7837 mutex_lock(&rt_constraints_mutex);
7838 read_lock(&tasklist_lock);
7839 ret = __rt_schedulable(NULL, 0, 0);
7840 read_unlock(&tasklist_lock);
7841 mutex_unlock(&rt_constraints_mutex);
7842
7843 return ret;
7844}
7845
7846int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7847{
7848 /* Don't accept realtime tasks when there is no way for them to run */
7849 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7850 return 0;
7851
7852 return 1;
7853}
7854
7855#else /* !CONFIG_RT_GROUP_SCHED */
7856static int sched_rt_global_constraints(void)
7857{
7858 unsigned long flags;
7859 int i;
7860
7861 if (sysctl_sched_rt_period <= 0)
7862 return -EINVAL;
7863
7864 /*
7865 * There's always some RT tasks in the root group
7866 * -- migration, kstopmachine etc..
7867 */
7868 if (sysctl_sched_rt_runtime == 0)
7869 return -EBUSY;
7870
7871 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7872 for_each_possible_cpu(i) {
7873 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7874
7875 raw_spin_lock(&rt_rq->rt_runtime_lock);
7876 rt_rq->rt_runtime = global_rt_runtime();
7877 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7878 }
7879 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7880
7881 return 0;
7882}
7883#endif /* CONFIG_RT_GROUP_SCHED */
7884
7885int sched_rt_handler(struct ctl_table *table, int write,
7886 void __user *buffer, size_t *lenp,
7887 loff_t *ppos)
7888{
7889 int ret;
7890 int old_period, old_runtime;
7891 static DEFINE_MUTEX(mutex);
7892
7893 mutex_lock(&mutex);
7894 old_period = sysctl_sched_rt_period;
7895 old_runtime = sysctl_sched_rt_runtime;
7896
7897 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7898
7899 if (!ret && write) {
7900 ret = sched_rt_global_constraints();
7901 if (ret) {
7902 sysctl_sched_rt_period = old_period;
7903 sysctl_sched_rt_runtime = old_runtime;
7904 } else {
7905 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7906 def_rt_bandwidth.rt_period =
7907 ns_to_ktime(global_rt_period());
7908 }
7909 }
7910 mutex_unlock(&mutex);
7911
7912 return ret;
7913}
7914
7915#ifdef CONFIG_CGROUP_SCHED
7916
7917/* return corresponding task_group object of a cgroup */
7918static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7919{
7920 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7921 struct task_group, css);
7922}
7923
7924static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7925{
7926 struct task_group *tg, *parent;
7927
7928 if (!cgrp->parent) {
7929 /* This is early initialization for the top cgroup */
7930 return &root_task_group.css;
7931 }
7932
7933 parent = cgroup_tg(cgrp->parent);
7934 tg = sched_create_group(parent);
7935 if (IS_ERR(tg))
7936 return ERR_PTR(-ENOMEM);
7937
7938 return &tg->css;
7939}
7940
7941static void cpu_cgroup_destroy(struct cgroup *cgrp)
7942{
7943 struct task_group *tg = cgroup_tg(cgrp);
7944
7945 sched_destroy_group(tg);
7946}
7947
7948static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7949 struct cgroup_taskset *tset)
7950{
7951 struct task_struct *task;
7952
7953 cgroup_taskset_for_each(task, cgrp, tset) {
7954#ifdef CONFIG_RT_GROUP_SCHED
7955 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7956 return -EINVAL;
7957#else
7958 /* We don't support RT-tasks being in separate groups */
7959 if (task->sched_class != &fair_sched_class)
7960 return -EINVAL;
7961#endif
7962 }
7963 return 0;
7964}
7965
7966static void cpu_cgroup_attach(struct cgroup *cgrp,
7967 struct cgroup_taskset *tset)
7968{
7969 struct task_struct *task;
7970
7971 cgroup_taskset_for_each(task, cgrp, tset)
7972 sched_move_task(task);
7973}
7974
7975static void
7976cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7977 struct task_struct *task)
7978{
7979 /*
7980 * cgroup_exit() is called in the copy_process() failure path.
7981 * Ignore this case since the task hasn't ran yet, this avoids
7982 * trying to poke a half freed task state from generic code.
7983 */
7984 if (!(task->flags & PF_EXITING))
7985 return;
7986
7987 sched_move_task(task);
7988}
7989
7990#ifdef CONFIG_FAIR_GROUP_SCHED
7991static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7992 u64 shareval)
7993{
7994 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7995}
7996
7997static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7998{
7999 struct task_group *tg = cgroup_tg(cgrp);
8000
8001 return (u64) scale_load_down(tg->shares);
8002}
8003
8004#ifdef CONFIG_CFS_BANDWIDTH
8005static DEFINE_MUTEX(cfs_constraints_mutex);
8006
8007const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8008const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8009
8010static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8011
8012static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8013{
8014 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8015 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8016
8017 if (tg == &root_task_group)
8018 return -EINVAL;
8019
8020 /*
8021 * Ensure we have at some amount of bandwidth every period. This is
8022 * to prevent reaching a state of large arrears when throttled via
8023 * entity_tick() resulting in prolonged exit starvation.
8024 */
8025 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8026 return -EINVAL;
8027
8028 /*
8029 * Likewise, bound things on the otherside by preventing insane quota
8030 * periods. This also allows us to normalize in computing quota
8031 * feasibility.
8032 */
8033 if (period > max_cfs_quota_period)
8034 return -EINVAL;
8035
8036 mutex_lock(&cfs_constraints_mutex);
8037 ret = __cfs_schedulable(tg, period, quota);
8038 if (ret)
8039 goto out_unlock;
8040
8041 runtime_enabled = quota != RUNTIME_INF;
8042 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8043 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
8044 raw_spin_lock_irq(&cfs_b->lock);
8045 cfs_b->period = ns_to_ktime(period);
8046 cfs_b->quota = quota;
8047
8048 __refill_cfs_bandwidth_runtime(cfs_b);
8049 /* restart the period timer (if active) to handle new period expiry */
8050 if (runtime_enabled && cfs_b->timer_active) {
8051 /* force a reprogram */
8052 cfs_b->timer_active = 0;
8053 __start_cfs_bandwidth(cfs_b);
8054 }
8055 raw_spin_unlock_irq(&cfs_b->lock);
8056
8057 for_each_possible_cpu(i) {
8058 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8059 struct rq *rq = cfs_rq->rq;
8060
8061 raw_spin_lock_irq(&rq->lock);
8062 cfs_rq->runtime_enabled = runtime_enabled;
8063 cfs_rq->runtime_remaining = 0;
8064
8065 if (cfs_rq->throttled)
8066 unthrottle_cfs_rq(cfs_rq);
8067 raw_spin_unlock_irq(&rq->lock);
8068 }
8069out_unlock:
8070 mutex_unlock(&cfs_constraints_mutex);
8071
8072 return ret;
8073}
8074
8075int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8076{
8077 u64 quota, period;
8078
8079 period = ktime_to_ns(tg->cfs_bandwidth.period);
8080 if (cfs_quota_us < 0)
8081 quota = RUNTIME_INF;
8082 else
8083 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8084
8085 return tg_set_cfs_bandwidth(tg, period, quota);
8086}
8087
8088long tg_get_cfs_quota(struct task_group *tg)
8089{
8090 u64 quota_us;
8091
8092 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8093 return -1;
8094
8095 quota_us = tg->cfs_bandwidth.quota;
8096 do_div(quota_us, NSEC_PER_USEC);
8097
8098 return quota_us;
8099}
8100
8101int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8102{
8103 u64 quota, period;
8104
8105 period = (u64)cfs_period_us * NSEC_PER_USEC;
8106 quota = tg->cfs_bandwidth.quota;
8107
8108 return tg_set_cfs_bandwidth(tg, period, quota);
8109}
8110
8111long tg_get_cfs_period(struct task_group *tg)
8112{
8113 u64 cfs_period_us;
8114
8115 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8116 do_div(cfs_period_us, NSEC_PER_USEC);
8117
8118 return cfs_period_us;
8119}
8120
8121static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
8122{
8123 return tg_get_cfs_quota(cgroup_tg(cgrp));
8124}
8125
8126static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
8127 s64 cfs_quota_us)
8128{
8129 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
8130}
8131
8132static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
8133{
8134 return tg_get_cfs_period(cgroup_tg(cgrp));
8135}
8136
8137static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8138 u64 cfs_period_us)
8139{
8140 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
8141}
8142
8143struct cfs_schedulable_data {
8144 struct task_group *tg;
8145 u64 period, quota;
8146};
8147
8148/*
8149 * normalize group quota/period to be quota/max_period
8150 * note: units are usecs
8151 */
8152static u64 normalize_cfs_quota(struct task_group *tg,
8153 struct cfs_schedulable_data *d)
8154{
8155 u64 quota, period;
8156
8157 if (tg == d->tg) {
8158 period = d->period;
8159 quota = d->quota;
8160 } else {
8161 period = tg_get_cfs_period(tg);
8162 quota = tg_get_cfs_quota(tg);
8163 }
8164
8165 /* note: these should typically be equivalent */
8166 if (quota == RUNTIME_INF || quota == -1)
8167 return RUNTIME_INF;
8168
8169 return to_ratio(period, quota);
8170}
8171
8172static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8173{
8174 struct cfs_schedulable_data *d = data;
8175 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8176 s64 quota = 0, parent_quota = -1;
8177
8178 if (!tg->parent) {
8179 quota = RUNTIME_INF;
8180 } else {
8181 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8182
8183 quota = normalize_cfs_quota(tg, d);
8184 parent_quota = parent_b->hierarchal_quota;
8185
8186 /*
8187 * ensure max(child_quota) <= parent_quota, inherit when no
8188 * limit is set
8189 */
8190 if (quota == RUNTIME_INF)
8191 quota = parent_quota;
8192 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8193 return -EINVAL;
8194 }
8195 cfs_b->hierarchal_quota = quota;
8196
8197 return 0;
8198}
8199
8200static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8201{
8202 int ret;
8203 struct cfs_schedulable_data data = {
8204 .tg = tg,
8205 .period = period,
8206 .quota = quota,
8207 };
8208
8209 if (quota != RUNTIME_INF) {
8210 do_div(data.period, NSEC_PER_USEC);
8211 do_div(data.quota, NSEC_PER_USEC);
8212 }
8213
8214 rcu_read_lock();
8215 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8216 rcu_read_unlock();
8217
8218 return ret;
8219}
8220
8221static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
8222 struct cgroup_map_cb *cb)
8223{
8224 struct task_group *tg = cgroup_tg(cgrp);
8225 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8226
8227 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
8228 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
8229 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
8230
8231 return 0;
8232}
8233#endif /* CONFIG_CFS_BANDWIDTH */
8234#endif /* CONFIG_FAIR_GROUP_SCHED */
8235
8236#ifdef CONFIG_RT_GROUP_SCHED
8237static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8238 s64 val)
8239{
8240 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8241}
8242
8243static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8244{
8245 return sched_group_rt_runtime(cgroup_tg(cgrp));
8246}
8247
8248static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8249 u64 rt_period_us)
8250{
8251 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8252}
8253
8254static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8255{
8256 return sched_group_rt_period(cgroup_tg(cgrp));
8257}
8258#endif /* CONFIG_RT_GROUP_SCHED */
8259
8260static struct cftype cpu_files[] = {
8261#ifdef CONFIG_FAIR_GROUP_SCHED
8262 {
8263 .name = "shares",
8264 .read_u64 = cpu_shares_read_u64,
8265 .write_u64 = cpu_shares_write_u64,
8266 },
8267#endif
8268#ifdef CONFIG_CFS_BANDWIDTH
8269 {
8270 .name = "cfs_quota_us",
8271 .read_s64 = cpu_cfs_quota_read_s64,
8272 .write_s64 = cpu_cfs_quota_write_s64,
8273 },
8274 {
8275 .name = "cfs_period_us",
8276 .read_u64 = cpu_cfs_period_read_u64,
8277 .write_u64 = cpu_cfs_period_write_u64,
8278 },
8279 {
8280 .name = "stat",
8281 .read_map = cpu_stats_show,
8282 },
8283#endif
8284#ifdef CONFIG_RT_GROUP_SCHED
8285 {
8286 .name = "rt_runtime_us",
8287 .read_s64 = cpu_rt_runtime_read,
8288 .write_s64 = cpu_rt_runtime_write,
8289 },
8290 {
8291 .name = "rt_period_us",
8292 .read_u64 = cpu_rt_period_read_uint,
8293 .write_u64 = cpu_rt_period_write_uint,
8294 },
8295#endif
8296 { } /* terminate */
8297};
8298
8299struct cgroup_subsys cpu_cgroup_subsys = {
8300 .name = "cpu",
8301 .create = cpu_cgroup_create,
8302 .destroy = cpu_cgroup_destroy,
8303 .can_attach = cpu_cgroup_can_attach,
8304 .attach = cpu_cgroup_attach,
8305 .exit = cpu_cgroup_exit,
8306 .subsys_id = cpu_cgroup_subsys_id,
8307 .base_cftypes = cpu_files,
8308 .early_init = 1,
8309};
8310
8311#endif /* CONFIG_CGROUP_SCHED */
8312
8313#ifdef CONFIG_CGROUP_CPUACCT
8314
8315/*
8316 * CPU accounting code for task groups.
8317 *
8318 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8319 * (balbir@in.ibm.com).
8320 */
8321
8322/* create a new cpu accounting group */
8323static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
8324{
8325 struct cpuacct *ca;
8326
8327 if (!cgrp->parent)
8328 return &root_cpuacct.css;
8329
8330 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8331 if (!ca)
8332 goto out;
8333
8334 ca->cpuusage = alloc_percpu(u64);
8335 if (!ca->cpuusage)
8336 goto out_free_ca;
8337
8338 ca->cpustat = alloc_percpu(struct kernel_cpustat);
8339 if (!ca->cpustat)
8340 goto out_free_cpuusage;
8341
8342 return &ca->css;
8343
8344out_free_cpuusage:
8345 free_percpu(ca->cpuusage);
8346out_free_ca:
8347 kfree(ca);
8348out:
8349 return ERR_PTR(-ENOMEM);
8350}
8351
8352/* destroy an existing cpu accounting group */
8353static void cpuacct_destroy(struct cgroup *cgrp)
8354{
8355 struct cpuacct *ca = cgroup_ca(cgrp);
8356
8357 free_percpu(ca->cpustat);
8358 free_percpu(ca->cpuusage);
8359 kfree(ca);
8360}
8361
8362static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8363{
8364 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8365 u64 data;
8366
8367#ifndef CONFIG_64BIT
8368 /*
8369 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8370 */
8371 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8372 data = *cpuusage;
8373 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8374#else
8375 data = *cpuusage;
8376#endif
8377
8378 return data;
8379}
8380
8381static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8382{
8383 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8384
8385#ifndef CONFIG_64BIT
8386 /*
8387 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8388 */
8389 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8390 *cpuusage = val;
8391 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8392#else
8393 *cpuusage = val;
8394#endif
8395}
8396
8397/* return total cpu usage (in nanoseconds) of a group */
8398static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8399{
8400 struct cpuacct *ca = cgroup_ca(cgrp);
8401 u64 totalcpuusage = 0;
8402 int i;
8403
8404 for_each_present_cpu(i)
8405 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8406
8407 return totalcpuusage;
8408}
8409
8410static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8411 u64 reset)
8412{
8413 struct cpuacct *ca = cgroup_ca(cgrp);
8414 int err = 0;
8415 int i;
8416
8417 if (reset) {
8418 err = -EINVAL;
8419 goto out;
8420 }
8421
8422 for_each_present_cpu(i)
8423 cpuacct_cpuusage_write(ca, i, 0);
8424
8425out:
8426 return err;
8427}
8428
8429static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8430 struct seq_file *m)
8431{
8432 struct cpuacct *ca = cgroup_ca(cgroup);
8433 u64 percpu;
8434 int i;
8435
8436 for_each_present_cpu(i) {
8437 percpu = cpuacct_cpuusage_read(ca, i);
8438 seq_printf(m, "%llu ", (unsigned long long) percpu);
8439 }
8440 seq_printf(m, "\n");
8441 return 0;
8442}
8443
8444static const char *cpuacct_stat_desc[] = {
8445 [CPUACCT_STAT_USER] = "user",
8446 [CPUACCT_STAT_SYSTEM] = "system",
8447};
8448
8449static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8450 struct cgroup_map_cb *cb)
8451{
8452 struct cpuacct *ca = cgroup_ca(cgrp);
8453 int cpu;
8454 s64 val = 0;
8455
8456 for_each_online_cpu(cpu) {
8457 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8458 val += kcpustat->cpustat[CPUTIME_USER];
8459 val += kcpustat->cpustat[CPUTIME_NICE];
8460 }
8461 val = cputime64_to_clock_t(val);
8462 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8463
8464 val = 0;
8465 for_each_online_cpu(cpu) {
8466 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8467 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8468 val += kcpustat->cpustat[CPUTIME_IRQ];
8469 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8470 }
8471
8472 val = cputime64_to_clock_t(val);
8473 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8474
8475 return 0;
8476}
8477
8478static struct cftype files[] = {
8479 {
8480 .name = "usage",
8481 .read_u64 = cpuusage_read,
8482 .write_u64 = cpuusage_write,
8483 },
8484 {
8485 .name = "usage_percpu",
8486 .read_seq_string = cpuacct_percpu_seq_read,
8487 },
8488 {
8489 .name = "stat",
8490 .read_map = cpuacct_stats_show,
8491 },
8492 { } /* terminate */
8493};
8494
8495/*
8496 * charge this task's execution time to its accounting group.
8497 *
8498 * called with rq->lock held.
8499 */
8500void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8501{
8502 struct cpuacct *ca;
8503 int cpu;
8504
8505 if (unlikely(!cpuacct_subsys.active))
8506 return;
8507
8508 cpu = task_cpu(tsk);
8509
8510 rcu_read_lock();
8511
8512 ca = task_ca(tsk);
8513
8514 for (; ca; ca = parent_ca(ca)) {
8515 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8516 *cpuusage += cputime;
8517 }
8518
8519 rcu_read_unlock();
8520}
8521
8522struct cgroup_subsys cpuacct_subsys = {
8523 .name = "cpuacct",
8524 .create = cpuacct_create,
8525 .destroy = cpuacct_destroy,
8526 .subsys_id = cpuacct_subsys_id,
8527 .base_cftypes = files,
8528};
8529#endif /* CONFIG_CGROUP_CPUACCT */
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * kernel/sched/core.c
4 *
5 * Core kernel scheduler code and related syscalls
6 *
7 * Copyright (C) 1991-2002 Linus Torvalds
8 */
9#include <linux/highmem.h>
10#include <linux/hrtimer_api.h>
11#include <linux/ktime_api.h>
12#include <linux/sched/signal.h>
13#include <linux/syscalls_api.h>
14#include <linux/debug_locks.h>
15#include <linux/prefetch.h>
16#include <linux/capability.h>
17#include <linux/pgtable_api.h>
18#include <linux/wait_bit.h>
19#include <linux/jiffies.h>
20#include <linux/spinlock_api.h>
21#include <linux/cpumask_api.h>
22#include <linux/lockdep_api.h>
23#include <linux/hardirq.h>
24#include <linux/softirq.h>
25#include <linux/refcount_api.h>
26#include <linux/topology.h>
27#include <linux/sched/clock.h>
28#include <linux/sched/cond_resched.h>
29#include <linux/sched/cputime.h>
30#include <linux/sched/debug.h>
31#include <linux/sched/hotplug.h>
32#include <linux/sched/init.h>
33#include <linux/sched/isolation.h>
34#include <linux/sched/loadavg.h>
35#include <linux/sched/mm.h>
36#include <linux/sched/nohz.h>
37#include <linux/sched/rseq_api.h>
38#include <linux/sched/rt.h>
39
40#include <linux/blkdev.h>
41#include <linux/context_tracking.h>
42#include <linux/cpuset.h>
43#include <linux/delayacct.h>
44#include <linux/init_task.h>
45#include <linux/interrupt.h>
46#include <linux/ioprio.h>
47#include <linux/kallsyms.h>
48#include <linux/kcov.h>
49#include <linux/kprobes.h>
50#include <linux/llist_api.h>
51#include <linux/mmu_context.h>
52#include <linux/mmzone.h>
53#include <linux/mutex_api.h>
54#include <linux/nmi.h>
55#include <linux/nospec.h>
56#include <linux/perf_event_api.h>
57#include <linux/profile.h>
58#include <linux/psi.h>
59#include <linux/rcuwait_api.h>
60#include <linux/rseq.h>
61#include <linux/sched/wake_q.h>
62#include <linux/scs.h>
63#include <linux/slab.h>
64#include <linux/syscalls.h>
65#include <linux/vtime.h>
66#include <linux/wait_api.h>
67#include <linux/workqueue_api.h>
68
69#ifdef CONFIG_PREEMPT_DYNAMIC
70# ifdef CONFIG_GENERIC_ENTRY
71# include <linux/entry-common.h>
72# endif
73#endif
74
75#include <uapi/linux/sched/types.h>
76
77#include <asm/irq_regs.h>
78#include <asm/switch_to.h>
79#include <asm/tlb.h>
80
81#define CREATE_TRACE_POINTS
82#include <linux/sched/rseq_api.h>
83#include <trace/events/sched.h>
84#include <trace/events/ipi.h>
85#undef CREATE_TRACE_POINTS
86
87#include "sched.h"
88#include "stats.h"
89
90#include "autogroup.h"
91#include "pelt.h"
92#include "smp.h"
93#include "stats.h"
94
95#include "../workqueue_internal.h"
96#include "../../io_uring/io-wq.h"
97#include "../smpboot.h"
98
99EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpu);
100EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpumask);
101
102/*
103 * Export tracepoints that act as a bare tracehook (ie: have no trace event
104 * associated with them) to allow external modules to probe them.
105 */
106EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
107EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
108EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
109EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
110EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
111EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_thermal_tp);
112EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
113EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
114EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
115EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
116EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
117EXPORT_TRACEPOINT_SYMBOL_GPL(sched_compute_energy_tp);
118
119DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
120
121#ifdef CONFIG_SCHED_DEBUG
122/*
123 * Debugging: various feature bits
124 *
125 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
126 * sysctl_sched_features, defined in sched.h, to allow constants propagation
127 * at compile time and compiler optimization based on features default.
128 */
129#define SCHED_FEAT(name, enabled) \
130 (1UL << __SCHED_FEAT_##name) * enabled |
131const_debug unsigned int sysctl_sched_features =
132#include "features.h"
133 0;
134#undef SCHED_FEAT
135
136/*
137 * Print a warning if need_resched is set for the given duration (if
138 * LATENCY_WARN is enabled).
139 *
140 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
141 * per boot.
142 */
143__read_mostly int sysctl_resched_latency_warn_ms = 100;
144__read_mostly int sysctl_resched_latency_warn_once = 1;
145#endif /* CONFIG_SCHED_DEBUG */
146
147/*
148 * Number of tasks to iterate in a single balance run.
149 * Limited because this is done with IRQs disabled.
150 */
151const_debug unsigned int sysctl_sched_nr_migrate = SCHED_NR_MIGRATE_BREAK;
152
153__read_mostly int scheduler_running;
154
155#ifdef CONFIG_SCHED_CORE
156
157DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
158
159/* kernel prio, less is more */
160static inline int __task_prio(const struct task_struct *p)
161{
162 if (p->sched_class == &stop_sched_class) /* trumps deadline */
163 return -2;
164
165 if (rt_prio(p->prio)) /* includes deadline */
166 return p->prio; /* [-1, 99] */
167
168 if (p->sched_class == &idle_sched_class)
169 return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
170
171 return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
172}
173
174/*
175 * l(a,b)
176 * le(a,b) := !l(b,a)
177 * g(a,b) := l(b,a)
178 * ge(a,b) := !l(a,b)
179 */
180
181/* real prio, less is less */
182static inline bool prio_less(const struct task_struct *a,
183 const struct task_struct *b, bool in_fi)
184{
185
186 int pa = __task_prio(a), pb = __task_prio(b);
187
188 if (-pa < -pb)
189 return true;
190
191 if (-pb < -pa)
192 return false;
193
194 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
195 return !dl_time_before(a->dl.deadline, b->dl.deadline);
196
197 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
198 return cfs_prio_less(a, b, in_fi);
199
200 return false;
201}
202
203static inline bool __sched_core_less(const struct task_struct *a,
204 const struct task_struct *b)
205{
206 if (a->core_cookie < b->core_cookie)
207 return true;
208
209 if (a->core_cookie > b->core_cookie)
210 return false;
211
212 /* flip prio, so high prio is leftmost */
213 if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
214 return true;
215
216 return false;
217}
218
219#define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
220
221static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
222{
223 return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
224}
225
226static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
227{
228 const struct task_struct *p = __node_2_sc(node);
229 unsigned long cookie = (unsigned long)key;
230
231 if (cookie < p->core_cookie)
232 return -1;
233
234 if (cookie > p->core_cookie)
235 return 1;
236
237 return 0;
238}
239
240void sched_core_enqueue(struct rq *rq, struct task_struct *p)
241{
242 rq->core->core_task_seq++;
243
244 if (!p->core_cookie)
245 return;
246
247 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
248}
249
250void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
251{
252 rq->core->core_task_seq++;
253
254 if (sched_core_enqueued(p)) {
255 rb_erase(&p->core_node, &rq->core_tree);
256 RB_CLEAR_NODE(&p->core_node);
257 }
258
259 /*
260 * Migrating the last task off the cpu, with the cpu in forced idle
261 * state. Reschedule to create an accounting edge for forced idle,
262 * and re-examine whether the core is still in forced idle state.
263 */
264 if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
265 rq->core->core_forceidle_count && rq->curr == rq->idle)
266 resched_curr(rq);
267}
268
269static int sched_task_is_throttled(struct task_struct *p, int cpu)
270{
271 if (p->sched_class->task_is_throttled)
272 return p->sched_class->task_is_throttled(p, cpu);
273
274 return 0;
275}
276
277static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
278{
279 struct rb_node *node = &p->core_node;
280 int cpu = task_cpu(p);
281
282 do {
283 node = rb_next(node);
284 if (!node)
285 return NULL;
286
287 p = __node_2_sc(node);
288 if (p->core_cookie != cookie)
289 return NULL;
290
291 } while (sched_task_is_throttled(p, cpu));
292
293 return p;
294}
295
296/*
297 * Find left-most (aka, highest priority) and unthrottled task matching @cookie.
298 * If no suitable task is found, NULL will be returned.
299 */
300static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
301{
302 struct task_struct *p;
303 struct rb_node *node;
304
305 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
306 if (!node)
307 return NULL;
308
309 p = __node_2_sc(node);
310 if (!sched_task_is_throttled(p, rq->cpu))
311 return p;
312
313 return sched_core_next(p, cookie);
314}
315
316/*
317 * Magic required such that:
318 *
319 * raw_spin_rq_lock(rq);
320 * ...
321 * raw_spin_rq_unlock(rq);
322 *
323 * ends up locking and unlocking the _same_ lock, and all CPUs
324 * always agree on what rq has what lock.
325 *
326 * XXX entirely possible to selectively enable cores, don't bother for now.
327 */
328
329static DEFINE_MUTEX(sched_core_mutex);
330static atomic_t sched_core_count;
331static struct cpumask sched_core_mask;
332
333static void sched_core_lock(int cpu, unsigned long *flags)
334{
335 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
336 int t, i = 0;
337
338 local_irq_save(*flags);
339 for_each_cpu(t, smt_mask)
340 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
341}
342
343static void sched_core_unlock(int cpu, unsigned long *flags)
344{
345 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
346 int t;
347
348 for_each_cpu(t, smt_mask)
349 raw_spin_unlock(&cpu_rq(t)->__lock);
350 local_irq_restore(*flags);
351}
352
353static void __sched_core_flip(bool enabled)
354{
355 unsigned long flags;
356 int cpu, t;
357
358 cpus_read_lock();
359
360 /*
361 * Toggle the online cores, one by one.
362 */
363 cpumask_copy(&sched_core_mask, cpu_online_mask);
364 for_each_cpu(cpu, &sched_core_mask) {
365 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
366
367 sched_core_lock(cpu, &flags);
368
369 for_each_cpu(t, smt_mask)
370 cpu_rq(t)->core_enabled = enabled;
371
372 cpu_rq(cpu)->core->core_forceidle_start = 0;
373
374 sched_core_unlock(cpu, &flags);
375
376 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
377 }
378
379 /*
380 * Toggle the offline CPUs.
381 */
382 for_each_cpu_andnot(cpu, cpu_possible_mask, cpu_online_mask)
383 cpu_rq(cpu)->core_enabled = enabled;
384
385 cpus_read_unlock();
386}
387
388static void sched_core_assert_empty(void)
389{
390 int cpu;
391
392 for_each_possible_cpu(cpu)
393 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
394}
395
396static void __sched_core_enable(void)
397{
398 static_branch_enable(&__sched_core_enabled);
399 /*
400 * Ensure all previous instances of raw_spin_rq_*lock() have finished
401 * and future ones will observe !sched_core_disabled().
402 */
403 synchronize_rcu();
404 __sched_core_flip(true);
405 sched_core_assert_empty();
406}
407
408static void __sched_core_disable(void)
409{
410 sched_core_assert_empty();
411 __sched_core_flip(false);
412 static_branch_disable(&__sched_core_enabled);
413}
414
415void sched_core_get(void)
416{
417 if (atomic_inc_not_zero(&sched_core_count))
418 return;
419
420 mutex_lock(&sched_core_mutex);
421 if (!atomic_read(&sched_core_count))
422 __sched_core_enable();
423
424 smp_mb__before_atomic();
425 atomic_inc(&sched_core_count);
426 mutex_unlock(&sched_core_mutex);
427}
428
429static void __sched_core_put(struct work_struct *work)
430{
431 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
432 __sched_core_disable();
433 mutex_unlock(&sched_core_mutex);
434 }
435}
436
437void sched_core_put(void)
438{
439 static DECLARE_WORK(_work, __sched_core_put);
440
441 /*
442 * "There can be only one"
443 *
444 * Either this is the last one, or we don't actually need to do any
445 * 'work'. If it is the last *again*, we rely on
446 * WORK_STRUCT_PENDING_BIT.
447 */
448 if (!atomic_add_unless(&sched_core_count, -1, 1))
449 schedule_work(&_work);
450}
451
452#else /* !CONFIG_SCHED_CORE */
453
454static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
455static inline void
456sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
457
458#endif /* CONFIG_SCHED_CORE */
459
460/*
461 * Serialization rules:
462 *
463 * Lock order:
464 *
465 * p->pi_lock
466 * rq->lock
467 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
468 *
469 * rq1->lock
470 * rq2->lock where: rq1 < rq2
471 *
472 * Regular state:
473 *
474 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
475 * local CPU's rq->lock, it optionally removes the task from the runqueue and
476 * always looks at the local rq data structures to find the most eligible task
477 * to run next.
478 *
479 * Task enqueue is also under rq->lock, possibly taken from another CPU.
480 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
481 * the local CPU to avoid bouncing the runqueue state around [ see
482 * ttwu_queue_wakelist() ]
483 *
484 * Task wakeup, specifically wakeups that involve migration, are horribly
485 * complicated to avoid having to take two rq->locks.
486 *
487 * Special state:
488 *
489 * System-calls and anything external will use task_rq_lock() which acquires
490 * both p->pi_lock and rq->lock. As a consequence the state they change is
491 * stable while holding either lock:
492 *
493 * - sched_setaffinity()/
494 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
495 * - set_user_nice(): p->se.load, p->*prio
496 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
497 * p->se.load, p->rt_priority,
498 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
499 * - sched_setnuma(): p->numa_preferred_nid
500 * - sched_move_task(): p->sched_task_group
501 * - uclamp_update_active() p->uclamp*
502 *
503 * p->state <- TASK_*:
504 *
505 * is changed locklessly using set_current_state(), __set_current_state() or
506 * set_special_state(), see their respective comments, or by
507 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
508 * concurrent self.
509 *
510 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
511 *
512 * is set by activate_task() and cleared by deactivate_task(), under
513 * rq->lock. Non-zero indicates the task is runnable, the special
514 * ON_RQ_MIGRATING state is used for migration without holding both
515 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
516 *
517 * p->on_cpu <- { 0, 1 }:
518 *
519 * is set by prepare_task() and cleared by finish_task() such that it will be
520 * set before p is scheduled-in and cleared after p is scheduled-out, both
521 * under rq->lock. Non-zero indicates the task is running on its CPU.
522 *
523 * [ The astute reader will observe that it is possible for two tasks on one
524 * CPU to have ->on_cpu = 1 at the same time. ]
525 *
526 * task_cpu(p): is changed by set_task_cpu(), the rules are:
527 *
528 * - Don't call set_task_cpu() on a blocked task:
529 *
530 * We don't care what CPU we're not running on, this simplifies hotplug,
531 * the CPU assignment of blocked tasks isn't required to be valid.
532 *
533 * - for try_to_wake_up(), called under p->pi_lock:
534 *
535 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
536 *
537 * - for migration called under rq->lock:
538 * [ see task_on_rq_migrating() in task_rq_lock() ]
539 *
540 * o move_queued_task()
541 * o detach_task()
542 *
543 * - for migration called under double_rq_lock():
544 *
545 * o __migrate_swap_task()
546 * o push_rt_task() / pull_rt_task()
547 * o push_dl_task() / pull_dl_task()
548 * o dl_task_offline_migration()
549 *
550 */
551
552void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
553{
554 raw_spinlock_t *lock;
555
556 /* Matches synchronize_rcu() in __sched_core_enable() */
557 preempt_disable();
558 if (sched_core_disabled()) {
559 raw_spin_lock_nested(&rq->__lock, subclass);
560 /* preempt_count *MUST* be > 1 */
561 preempt_enable_no_resched();
562 return;
563 }
564
565 for (;;) {
566 lock = __rq_lockp(rq);
567 raw_spin_lock_nested(lock, subclass);
568 if (likely(lock == __rq_lockp(rq))) {
569 /* preempt_count *MUST* be > 1 */
570 preempt_enable_no_resched();
571 return;
572 }
573 raw_spin_unlock(lock);
574 }
575}
576
577bool raw_spin_rq_trylock(struct rq *rq)
578{
579 raw_spinlock_t *lock;
580 bool ret;
581
582 /* Matches synchronize_rcu() in __sched_core_enable() */
583 preempt_disable();
584 if (sched_core_disabled()) {
585 ret = raw_spin_trylock(&rq->__lock);
586 preempt_enable();
587 return ret;
588 }
589
590 for (;;) {
591 lock = __rq_lockp(rq);
592 ret = raw_spin_trylock(lock);
593 if (!ret || (likely(lock == __rq_lockp(rq)))) {
594 preempt_enable();
595 return ret;
596 }
597 raw_spin_unlock(lock);
598 }
599}
600
601void raw_spin_rq_unlock(struct rq *rq)
602{
603 raw_spin_unlock(rq_lockp(rq));
604}
605
606#ifdef CONFIG_SMP
607/*
608 * double_rq_lock - safely lock two runqueues
609 */
610void double_rq_lock(struct rq *rq1, struct rq *rq2)
611{
612 lockdep_assert_irqs_disabled();
613
614 if (rq_order_less(rq2, rq1))
615 swap(rq1, rq2);
616
617 raw_spin_rq_lock(rq1);
618 if (__rq_lockp(rq1) != __rq_lockp(rq2))
619 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
620
621 double_rq_clock_clear_update(rq1, rq2);
622}
623#endif
624
625/*
626 * __task_rq_lock - lock the rq @p resides on.
627 */
628struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
629 __acquires(rq->lock)
630{
631 struct rq *rq;
632
633 lockdep_assert_held(&p->pi_lock);
634
635 for (;;) {
636 rq = task_rq(p);
637 raw_spin_rq_lock(rq);
638 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
639 rq_pin_lock(rq, rf);
640 return rq;
641 }
642 raw_spin_rq_unlock(rq);
643
644 while (unlikely(task_on_rq_migrating(p)))
645 cpu_relax();
646 }
647}
648
649/*
650 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
651 */
652struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
653 __acquires(p->pi_lock)
654 __acquires(rq->lock)
655{
656 struct rq *rq;
657
658 for (;;) {
659 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
660 rq = task_rq(p);
661 raw_spin_rq_lock(rq);
662 /*
663 * move_queued_task() task_rq_lock()
664 *
665 * ACQUIRE (rq->lock)
666 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
667 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
668 * [S] ->cpu = new_cpu [L] task_rq()
669 * [L] ->on_rq
670 * RELEASE (rq->lock)
671 *
672 * If we observe the old CPU in task_rq_lock(), the acquire of
673 * the old rq->lock will fully serialize against the stores.
674 *
675 * If we observe the new CPU in task_rq_lock(), the address
676 * dependency headed by '[L] rq = task_rq()' and the acquire
677 * will pair with the WMB to ensure we then also see migrating.
678 */
679 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
680 rq_pin_lock(rq, rf);
681 return rq;
682 }
683 raw_spin_rq_unlock(rq);
684 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
685
686 while (unlikely(task_on_rq_migrating(p)))
687 cpu_relax();
688 }
689}
690
691/*
692 * RQ-clock updating methods:
693 */
694
695static void update_rq_clock_task(struct rq *rq, s64 delta)
696{
697/*
698 * In theory, the compile should just see 0 here, and optimize out the call
699 * to sched_rt_avg_update. But I don't trust it...
700 */
701 s64 __maybe_unused steal = 0, irq_delta = 0;
702
703#ifdef CONFIG_IRQ_TIME_ACCOUNTING
704 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
705
706 /*
707 * Since irq_time is only updated on {soft,}irq_exit, we might run into
708 * this case when a previous update_rq_clock() happened inside a
709 * {soft,}irq region.
710 *
711 * When this happens, we stop ->clock_task and only update the
712 * prev_irq_time stamp to account for the part that fit, so that a next
713 * update will consume the rest. This ensures ->clock_task is
714 * monotonic.
715 *
716 * It does however cause some slight miss-attribution of {soft,}irq
717 * time, a more accurate solution would be to update the irq_time using
718 * the current rq->clock timestamp, except that would require using
719 * atomic ops.
720 */
721 if (irq_delta > delta)
722 irq_delta = delta;
723
724 rq->prev_irq_time += irq_delta;
725 delta -= irq_delta;
726 psi_account_irqtime(rq->curr, irq_delta);
727 delayacct_irq(rq->curr, irq_delta);
728#endif
729#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
730 if (static_key_false((¶virt_steal_rq_enabled))) {
731 steal = paravirt_steal_clock(cpu_of(rq));
732 steal -= rq->prev_steal_time_rq;
733
734 if (unlikely(steal > delta))
735 steal = delta;
736
737 rq->prev_steal_time_rq += steal;
738 delta -= steal;
739 }
740#endif
741
742 rq->clock_task += delta;
743
744#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
745 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
746 update_irq_load_avg(rq, irq_delta + steal);
747#endif
748 update_rq_clock_pelt(rq, delta);
749}
750
751void update_rq_clock(struct rq *rq)
752{
753 s64 delta;
754
755 lockdep_assert_rq_held(rq);
756
757 if (rq->clock_update_flags & RQCF_ACT_SKIP)
758 return;
759
760#ifdef CONFIG_SCHED_DEBUG
761 if (sched_feat(WARN_DOUBLE_CLOCK))
762 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
763 rq->clock_update_flags |= RQCF_UPDATED;
764#endif
765
766 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
767 if (delta < 0)
768 return;
769 rq->clock += delta;
770 update_rq_clock_task(rq, delta);
771}
772
773#ifdef CONFIG_SCHED_HRTICK
774/*
775 * Use HR-timers to deliver accurate preemption points.
776 */
777
778static void hrtick_clear(struct rq *rq)
779{
780 if (hrtimer_active(&rq->hrtick_timer))
781 hrtimer_cancel(&rq->hrtick_timer);
782}
783
784/*
785 * High-resolution timer tick.
786 * Runs from hardirq context with interrupts disabled.
787 */
788static enum hrtimer_restart hrtick(struct hrtimer *timer)
789{
790 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
791 struct rq_flags rf;
792
793 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
794
795 rq_lock(rq, &rf);
796 update_rq_clock(rq);
797 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
798 rq_unlock(rq, &rf);
799
800 return HRTIMER_NORESTART;
801}
802
803#ifdef CONFIG_SMP
804
805static void __hrtick_restart(struct rq *rq)
806{
807 struct hrtimer *timer = &rq->hrtick_timer;
808 ktime_t time = rq->hrtick_time;
809
810 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
811}
812
813/*
814 * called from hardirq (IPI) context
815 */
816static void __hrtick_start(void *arg)
817{
818 struct rq *rq = arg;
819 struct rq_flags rf;
820
821 rq_lock(rq, &rf);
822 __hrtick_restart(rq);
823 rq_unlock(rq, &rf);
824}
825
826/*
827 * Called to set the hrtick timer state.
828 *
829 * called with rq->lock held and irqs disabled
830 */
831void hrtick_start(struct rq *rq, u64 delay)
832{
833 struct hrtimer *timer = &rq->hrtick_timer;
834 s64 delta;
835
836 /*
837 * Don't schedule slices shorter than 10000ns, that just
838 * doesn't make sense and can cause timer DoS.
839 */
840 delta = max_t(s64, delay, 10000LL);
841 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
842
843 if (rq == this_rq())
844 __hrtick_restart(rq);
845 else
846 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
847}
848
849#else
850/*
851 * Called to set the hrtick timer state.
852 *
853 * called with rq->lock held and irqs disabled
854 */
855void hrtick_start(struct rq *rq, u64 delay)
856{
857 /*
858 * Don't schedule slices shorter than 10000ns, that just
859 * doesn't make sense. Rely on vruntime for fairness.
860 */
861 delay = max_t(u64, delay, 10000LL);
862 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
863 HRTIMER_MODE_REL_PINNED_HARD);
864}
865
866#endif /* CONFIG_SMP */
867
868static void hrtick_rq_init(struct rq *rq)
869{
870#ifdef CONFIG_SMP
871 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
872#endif
873 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
874 rq->hrtick_timer.function = hrtick;
875}
876#else /* CONFIG_SCHED_HRTICK */
877static inline void hrtick_clear(struct rq *rq)
878{
879}
880
881static inline void hrtick_rq_init(struct rq *rq)
882{
883}
884#endif /* CONFIG_SCHED_HRTICK */
885
886/*
887 * cmpxchg based fetch_or, macro so it works for different integer types
888 */
889#define fetch_or(ptr, mask) \
890 ({ \
891 typeof(ptr) _ptr = (ptr); \
892 typeof(mask) _mask = (mask); \
893 typeof(*_ptr) _val = *_ptr; \
894 \
895 do { \
896 } while (!try_cmpxchg(_ptr, &_val, _val | _mask)); \
897 _val; \
898})
899
900#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
901/*
902 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
903 * this avoids any races wrt polling state changes and thereby avoids
904 * spurious IPIs.
905 */
906static inline bool set_nr_and_not_polling(struct task_struct *p)
907{
908 struct thread_info *ti = task_thread_info(p);
909 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
910}
911
912/*
913 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
914 *
915 * If this returns true, then the idle task promises to call
916 * sched_ttwu_pending() and reschedule soon.
917 */
918static bool set_nr_if_polling(struct task_struct *p)
919{
920 struct thread_info *ti = task_thread_info(p);
921 typeof(ti->flags) val = READ_ONCE(ti->flags);
922
923 do {
924 if (!(val & _TIF_POLLING_NRFLAG))
925 return false;
926 if (val & _TIF_NEED_RESCHED)
927 return true;
928 } while (!try_cmpxchg(&ti->flags, &val, val | _TIF_NEED_RESCHED));
929
930 return true;
931}
932
933#else
934static inline bool set_nr_and_not_polling(struct task_struct *p)
935{
936 set_tsk_need_resched(p);
937 return true;
938}
939
940#ifdef CONFIG_SMP
941static inline bool set_nr_if_polling(struct task_struct *p)
942{
943 return false;
944}
945#endif
946#endif
947
948static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
949{
950 struct wake_q_node *node = &task->wake_q;
951
952 /*
953 * Atomically grab the task, if ->wake_q is !nil already it means
954 * it's already queued (either by us or someone else) and will get the
955 * wakeup due to that.
956 *
957 * In order to ensure that a pending wakeup will observe our pending
958 * state, even in the failed case, an explicit smp_mb() must be used.
959 */
960 smp_mb__before_atomic();
961 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
962 return false;
963
964 /*
965 * The head is context local, there can be no concurrency.
966 */
967 *head->lastp = node;
968 head->lastp = &node->next;
969 return true;
970}
971
972/**
973 * wake_q_add() - queue a wakeup for 'later' waking.
974 * @head: the wake_q_head to add @task to
975 * @task: the task to queue for 'later' wakeup
976 *
977 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
978 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
979 * instantly.
980 *
981 * This function must be used as-if it were wake_up_process(); IOW the task
982 * must be ready to be woken at this location.
983 */
984void wake_q_add(struct wake_q_head *head, struct task_struct *task)
985{
986 if (__wake_q_add(head, task))
987 get_task_struct(task);
988}
989
990/**
991 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
992 * @head: the wake_q_head to add @task to
993 * @task: the task to queue for 'later' wakeup
994 *
995 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
996 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
997 * instantly.
998 *
999 * This function must be used as-if it were wake_up_process(); IOW the task
1000 * must be ready to be woken at this location.
1001 *
1002 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
1003 * that already hold reference to @task can call the 'safe' version and trust
1004 * wake_q to do the right thing depending whether or not the @task is already
1005 * queued for wakeup.
1006 */
1007void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
1008{
1009 if (!__wake_q_add(head, task))
1010 put_task_struct(task);
1011}
1012
1013void wake_up_q(struct wake_q_head *head)
1014{
1015 struct wake_q_node *node = head->first;
1016
1017 while (node != WAKE_Q_TAIL) {
1018 struct task_struct *task;
1019
1020 task = container_of(node, struct task_struct, wake_q);
1021 /* Task can safely be re-inserted now: */
1022 node = node->next;
1023 task->wake_q.next = NULL;
1024
1025 /*
1026 * wake_up_process() executes a full barrier, which pairs with
1027 * the queueing in wake_q_add() so as not to miss wakeups.
1028 */
1029 wake_up_process(task);
1030 put_task_struct(task);
1031 }
1032}
1033
1034/*
1035 * resched_curr - mark rq's current task 'to be rescheduled now'.
1036 *
1037 * On UP this means the setting of the need_resched flag, on SMP it
1038 * might also involve a cross-CPU call to trigger the scheduler on
1039 * the target CPU.
1040 */
1041void resched_curr(struct rq *rq)
1042{
1043 struct task_struct *curr = rq->curr;
1044 int cpu;
1045
1046 lockdep_assert_rq_held(rq);
1047
1048 if (test_tsk_need_resched(curr))
1049 return;
1050
1051 cpu = cpu_of(rq);
1052
1053 if (cpu == smp_processor_id()) {
1054 set_tsk_need_resched(curr);
1055 set_preempt_need_resched();
1056 return;
1057 }
1058
1059 if (set_nr_and_not_polling(curr))
1060 smp_send_reschedule(cpu);
1061 else
1062 trace_sched_wake_idle_without_ipi(cpu);
1063}
1064
1065void resched_cpu(int cpu)
1066{
1067 struct rq *rq = cpu_rq(cpu);
1068 unsigned long flags;
1069
1070 raw_spin_rq_lock_irqsave(rq, flags);
1071 if (cpu_online(cpu) || cpu == smp_processor_id())
1072 resched_curr(rq);
1073 raw_spin_rq_unlock_irqrestore(rq, flags);
1074}
1075
1076#ifdef CONFIG_SMP
1077#ifdef CONFIG_NO_HZ_COMMON
1078/*
1079 * In the semi idle case, use the nearest busy CPU for migrating timers
1080 * from an idle CPU. This is good for power-savings.
1081 *
1082 * We don't do similar optimization for completely idle system, as
1083 * selecting an idle CPU will add more delays to the timers than intended
1084 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1085 */
1086int get_nohz_timer_target(void)
1087{
1088 int i, cpu = smp_processor_id(), default_cpu = -1;
1089 struct sched_domain *sd;
1090 const struct cpumask *hk_mask;
1091
1092 if (housekeeping_cpu(cpu, HK_TYPE_TIMER)) {
1093 if (!idle_cpu(cpu))
1094 return cpu;
1095 default_cpu = cpu;
1096 }
1097
1098 hk_mask = housekeeping_cpumask(HK_TYPE_TIMER);
1099
1100 guard(rcu)();
1101
1102 for_each_domain(cpu, sd) {
1103 for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1104 if (cpu == i)
1105 continue;
1106
1107 if (!idle_cpu(i))
1108 return i;
1109 }
1110 }
1111
1112 if (default_cpu == -1)
1113 default_cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
1114
1115 return default_cpu;
1116}
1117
1118/*
1119 * When add_timer_on() enqueues a timer into the timer wheel of an
1120 * idle CPU then this timer might expire before the next timer event
1121 * which is scheduled to wake up that CPU. In case of a completely
1122 * idle system the next event might even be infinite time into the
1123 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1124 * leaves the inner idle loop so the newly added timer is taken into
1125 * account when the CPU goes back to idle and evaluates the timer
1126 * wheel for the next timer event.
1127 */
1128static void wake_up_idle_cpu(int cpu)
1129{
1130 struct rq *rq = cpu_rq(cpu);
1131
1132 if (cpu == smp_processor_id())
1133 return;
1134
1135 /*
1136 * Set TIF_NEED_RESCHED and send an IPI if in the non-polling
1137 * part of the idle loop. This forces an exit from the idle loop
1138 * and a round trip to schedule(). Now this could be optimized
1139 * because a simple new idle loop iteration is enough to
1140 * re-evaluate the next tick. Provided some re-ordering of tick
1141 * nohz functions that would need to follow TIF_NR_POLLING
1142 * clearing:
1143 *
1144 * - On most archs, a simple fetch_or on ti::flags with a
1145 * "0" value would be enough to know if an IPI needs to be sent.
1146 *
1147 * - x86 needs to perform a last need_resched() check between
1148 * monitor and mwait which doesn't take timers into account.
1149 * There a dedicated TIF_TIMER flag would be required to
1150 * fetch_or here and be checked along with TIF_NEED_RESCHED
1151 * before mwait().
1152 *
1153 * However, remote timer enqueue is not such a frequent event
1154 * and testing of the above solutions didn't appear to report
1155 * much benefits.
1156 */
1157 if (set_nr_and_not_polling(rq->idle))
1158 smp_send_reschedule(cpu);
1159 else
1160 trace_sched_wake_idle_without_ipi(cpu);
1161}
1162
1163static bool wake_up_full_nohz_cpu(int cpu)
1164{
1165 /*
1166 * We just need the target to call irq_exit() and re-evaluate
1167 * the next tick. The nohz full kick at least implies that.
1168 * If needed we can still optimize that later with an
1169 * empty IRQ.
1170 */
1171 if (cpu_is_offline(cpu))
1172 return true; /* Don't try to wake offline CPUs. */
1173 if (tick_nohz_full_cpu(cpu)) {
1174 if (cpu != smp_processor_id() ||
1175 tick_nohz_tick_stopped())
1176 tick_nohz_full_kick_cpu(cpu);
1177 return true;
1178 }
1179
1180 return false;
1181}
1182
1183/*
1184 * Wake up the specified CPU. If the CPU is going offline, it is the
1185 * caller's responsibility to deal with the lost wakeup, for example,
1186 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1187 */
1188void wake_up_nohz_cpu(int cpu)
1189{
1190 if (!wake_up_full_nohz_cpu(cpu))
1191 wake_up_idle_cpu(cpu);
1192}
1193
1194static void nohz_csd_func(void *info)
1195{
1196 struct rq *rq = info;
1197 int cpu = cpu_of(rq);
1198 unsigned int flags;
1199
1200 /*
1201 * Release the rq::nohz_csd.
1202 */
1203 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1204 WARN_ON(!(flags & NOHZ_KICK_MASK));
1205
1206 rq->idle_balance = idle_cpu(cpu);
1207 if (rq->idle_balance && !need_resched()) {
1208 rq->nohz_idle_balance = flags;
1209 raise_softirq_irqoff(SCHED_SOFTIRQ);
1210 }
1211}
1212
1213#endif /* CONFIG_NO_HZ_COMMON */
1214
1215#ifdef CONFIG_NO_HZ_FULL
1216static inline bool __need_bw_check(struct rq *rq, struct task_struct *p)
1217{
1218 if (rq->nr_running != 1)
1219 return false;
1220
1221 if (p->sched_class != &fair_sched_class)
1222 return false;
1223
1224 if (!task_on_rq_queued(p))
1225 return false;
1226
1227 return true;
1228}
1229
1230bool sched_can_stop_tick(struct rq *rq)
1231{
1232 int fifo_nr_running;
1233
1234 /* Deadline tasks, even if single, need the tick */
1235 if (rq->dl.dl_nr_running)
1236 return false;
1237
1238 /*
1239 * If there are more than one RR tasks, we need the tick to affect the
1240 * actual RR behaviour.
1241 */
1242 if (rq->rt.rr_nr_running) {
1243 if (rq->rt.rr_nr_running == 1)
1244 return true;
1245 else
1246 return false;
1247 }
1248
1249 /*
1250 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1251 * forced preemption between FIFO tasks.
1252 */
1253 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1254 if (fifo_nr_running)
1255 return true;
1256
1257 /*
1258 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1259 * if there's more than one we need the tick for involuntary
1260 * preemption.
1261 */
1262 if (rq->nr_running > 1)
1263 return false;
1264
1265 /*
1266 * If there is one task and it has CFS runtime bandwidth constraints
1267 * and it's on the cpu now we don't want to stop the tick.
1268 * This check prevents clearing the bit if a newly enqueued task here is
1269 * dequeued by migrating while the constrained task continues to run.
1270 * E.g. going from 2->1 without going through pick_next_task().
1271 */
1272 if (sched_feat(HZ_BW) && __need_bw_check(rq, rq->curr)) {
1273 if (cfs_task_bw_constrained(rq->curr))
1274 return false;
1275 }
1276
1277 return true;
1278}
1279#endif /* CONFIG_NO_HZ_FULL */
1280#endif /* CONFIG_SMP */
1281
1282#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1283 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1284/*
1285 * Iterate task_group tree rooted at *from, calling @down when first entering a
1286 * node and @up when leaving it for the final time.
1287 *
1288 * Caller must hold rcu_lock or sufficient equivalent.
1289 */
1290int walk_tg_tree_from(struct task_group *from,
1291 tg_visitor down, tg_visitor up, void *data)
1292{
1293 struct task_group *parent, *child;
1294 int ret;
1295
1296 parent = from;
1297
1298down:
1299 ret = (*down)(parent, data);
1300 if (ret)
1301 goto out;
1302 list_for_each_entry_rcu(child, &parent->children, siblings) {
1303 parent = child;
1304 goto down;
1305
1306up:
1307 continue;
1308 }
1309 ret = (*up)(parent, data);
1310 if (ret || parent == from)
1311 goto out;
1312
1313 child = parent;
1314 parent = parent->parent;
1315 if (parent)
1316 goto up;
1317out:
1318 return ret;
1319}
1320
1321int tg_nop(struct task_group *tg, void *data)
1322{
1323 return 0;
1324}
1325#endif
1326
1327static void set_load_weight(struct task_struct *p, bool update_load)
1328{
1329 int prio = p->static_prio - MAX_RT_PRIO;
1330 struct load_weight *load = &p->se.load;
1331
1332 /*
1333 * SCHED_IDLE tasks get minimal weight:
1334 */
1335 if (task_has_idle_policy(p)) {
1336 load->weight = scale_load(WEIGHT_IDLEPRIO);
1337 load->inv_weight = WMULT_IDLEPRIO;
1338 return;
1339 }
1340
1341 /*
1342 * SCHED_OTHER tasks have to update their load when changing their
1343 * weight
1344 */
1345 if (update_load && p->sched_class == &fair_sched_class) {
1346 reweight_task(p, prio);
1347 } else {
1348 load->weight = scale_load(sched_prio_to_weight[prio]);
1349 load->inv_weight = sched_prio_to_wmult[prio];
1350 }
1351}
1352
1353#ifdef CONFIG_UCLAMP_TASK
1354/*
1355 * Serializes updates of utilization clamp values
1356 *
1357 * The (slow-path) user-space triggers utilization clamp value updates which
1358 * can require updates on (fast-path) scheduler's data structures used to
1359 * support enqueue/dequeue operations.
1360 * While the per-CPU rq lock protects fast-path update operations, user-space
1361 * requests are serialized using a mutex to reduce the risk of conflicting
1362 * updates or API abuses.
1363 */
1364static DEFINE_MUTEX(uclamp_mutex);
1365
1366/* Max allowed minimum utilization */
1367static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1368
1369/* Max allowed maximum utilization */
1370static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1371
1372/*
1373 * By default RT tasks run at the maximum performance point/capacity of the
1374 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1375 * SCHED_CAPACITY_SCALE.
1376 *
1377 * This knob allows admins to change the default behavior when uclamp is being
1378 * used. In battery powered devices, particularly, running at the maximum
1379 * capacity and frequency will increase energy consumption and shorten the
1380 * battery life.
1381 *
1382 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1383 *
1384 * This knob will not override the system default sched_util_clamp_min defined
1385 * above.
1386 */
1387static unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1388
1389/* All clamps are required to be less or equal than these values */
1390static struct uclamp_se uclamp_default[UCLAMP_CNT];
1391
1392/*
1393 * This static key is used to reduce the uclamp overhead in the fast path. It
1394 * primarily disables the call to uclamp_rq_{inc, dec}() in
1395 * enqueue/dequeue_task().
1396 *
1397 * This allows users to continue to enable uclamp in their kernel config with
1398 * minimum uclamp overhead in the fast path.
1399 *
1400 * As soon as userspace modifies any of the uclamp knobs, the static key is
1401 * enabled, since we have an actual users that make use of uclamp
1402 * functionality.
1403 *
1404 * The knobs that would enable this static key are:
1405 *
1406 * * A task modifying its uclamp value with sched_setattr().
1407 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1408 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1409 */
1410DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1411
1412/* Integer rounded range for each bucket */
1413#define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1414
1415#define for_each_clamp_id(clamp_id) \
1416 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1417
1418static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1419{
1420 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1421}
1422
1423static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1424{
1425 if (clamp_id == UCLAMP_MIN)
1426 return 0;
1427 return SCHED_CAPACITY_SCALE;
1428}
1429
1430static inline void uclamp_se_set(struct uclamp_se *uc_se,
1431 unsigned int value, bool user_defined)
1432{
1433 uc_se->value = value;
1434 uc_se->bucket_id = uclamp_bucket_id(value);
1435 uc_se->user_defined = user_defined;
1436}
1437
1438static inline unsigned int
1439uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1440 unsigned int clamp_value)
1441{
1442 /*
1443 * Avoid blocked utilization pushing up the frequency when we go
1444 * idle (which drops the max-clamp) by retaining the last known
1445 * max-clamp.
1446 */
1447 if (clamp_id == UCLAMP_MAX) {
1448 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1449 return clamp_value;
1450 }
1451
1452 return uclamp_none(UCLAMP_MIN);
1453}
1454
1455static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1456 unsigned int clamp_value)
1457{
1458 /* Reset max-clamp retention only on idle exit */
1459 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1460 return;
1461
1462 uclamp_rq_set(rq, clamp_id, clamp_value);
1463}
1464
1465static inline
1466unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1467 unsigned int clamp_value)
1468{
1469 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1470 int bucket_id = UCLAMP_BUCKETS - 1;
1471
1472 /*
1473 * Since both min and max clamps are max aggregated, find the
1474 * top most bucket with tasks in.
1475 */
1476 for ( ; bucket_id >= 0; bucket_id--) {
1477 if (!bucket[bucket_id].tasks)
1478 continue;
1479 return bucket[bucket_id].value;
1480 }
1481
1482 /* No tasks -- default clamp values */
1483 return uclamp_idle_value(rq, clamp_id, clamp_value);
1484}
1485
1486static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1487{
1488 unsigned int default_util_min;
1489 struct uclamp_se *uc_se;
1490
1491 lockdep_assert_held(&p->pi_lock);
1492
1493 uc_se = &p->uclamp_req[UCLAMP_MIN];
1494
1495 /* Only sync if user didn't override the default */
1496 if (uc_se->user_defined)
1497 return;
1498
1499 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1500 uclamp_se_set(uc_se, default_util_min, false);
1501}
1502
1503static void uclamp_update_util_min_rt_default(struct task_struct *p)
1504{
1505 if (!rt_task(p))
1506 return;
1507
1508 /* Protect updates to p->uclamp_* */
1509 guard(task_rq_lock)(p);
1510 __uclamp_update_util_min_rt_default(p);
1511}
1512
1513static inline struct uclamp_se
1514uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1515{
1516 /* Copy by value as we could modify it */
1517 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1518#ifdef CONFIG_UCLAMP_TASK_GROUP
1519 unsigned int tg_min, tg_max, value;
1520
1521 /*
1522 * Tasks in autogroups or root task group will be
1523 * restricted by system defaults.
1524 */
1525 if (task_group_is_autogroup(task_group(p)))
1526 return uc_req;
1527 if (task_group(p) == &root_task_group)
1528 return uc_req;
1529
1530 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1531 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1532 value = uc_req.value;
1533 value = clamp(value, tg_min, tg_max);
1534 uclamp_se_set(&uc_req, value, false);
1535#endif
1536
1537 return uc_req;
1538}
1539
1540/*
1541 * The effective clamp bucket index of a task depends on, by increasing
1542 * priority:
1543 * - the task specific clamp value, when explicitly requested from userspace
1544 * - the task group effective clamp value, for tasks not either in the root
1545 * group or in an autogroup
1546 * - the system default clamp value, defined by the sysadmin
1547 */
1548static inline struct uclamp_se
1549uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1550{
1551 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1552 struct uclamp_se uc_max = uclamp_default[clamp_id];
1553
1554 /* System default restrictions always apply */
1555 if (unlikely(uc_req.value > uc_max.value))
1556 return uc_max;
1557
1558 return uc_req;
1559}
1560
1561unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1562{
1563 struct uclamp_se uc_eff;
1564
1565 /* Task currently refcounted: use back-annotated (effective) value */
1566 if (p->uclamp[clamp_id].active)
1567 return (unsigned long)p->uclamp[clamp_id].value;
1568
1569 uc_eff = uclamp_eff_get(p, clamp_id);
1570
1571 return (unsigned long)uc_eff.value;
1572}
1573
1574/*
1575 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1576 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1577 * updates the rq's clamp value if required.
1578 *
1579 * Tasks can have a task-specific value requested from user-space, track
1580 * within each bucket the maximum value for tasks refcounted in it.
1581 * This "local max aggregation" allows to track the exact "requested" value
1582 * for each bucket when all its RUNNABLE tasks require the same clamp.
1583 */
1584static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1585 enum uclamp_id clamp_id)
1586{
1587 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1588 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1589 struct uclamp_bucket *bucket;
1590
1591 lockdep_assert_rq_held(rq);
1592
1593 /* Update task effective clamp */
1594 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1595
1596 bucket = &uc_rq->bucket[uc_se->bucket_id];
1597 bucket->tasks++;
1598 uc_se->active = true;
1599
1600 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1601
1602 /*
1603 * Local max aggregation: rq buckets always track the max
1604 * "requested" clamp value of its RUNNABLE tasks.
1605 */
1606 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1607 bucket->value = uc_se->value;
1608
1609 if (uc_se->value > uclamp_rq_get(rq, clamp_id))
1610 uclamp_rq_set(rq, clamp_id, uc_se->value);
1611}
1612
1613/*
1614 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1615 * is released. If this is the last task reference counting the rq's max
1616 * active clamp value, then the rq's clamp value is updated.
1617 *
1618 * Both refcounted tasks and rq's cached clamp values are expected to be
1619 * always valid. If it's detected they are not, as defensive programming,
1620 * enforce the expected state and warn.
1621 */
1622static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1623 enum uclamp_id clamp_id)
1624{
1625 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1626 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1627 struct uclamp_bucket *bucket;
1628 unsigned int bkt_clamp;
1629 unsigned int rq_clamp;
1630
1631 lockdep_assert_rq_held(rq);
1632
1633 /*
1634 * If sched_uclamp_used was enabled after task @p was enqueued,
1635 * we could end up with unbalanced call to uclamp_rq_dec_id().
1636 *
1637 * In this case the uc_se->active flag should be false since no uclamp
1638 * accounting was performed at enqueue time and we can just return
1639 * here.
1640 *
1641 * Need to be careful of the following enqueue/dequeue ordering
1642 * problem too
1643 *
1644 * enqueue(taskA)
1645 * // sched_uclamp_used gets enabled
1646 * enqueue(taskB)
1647 * dequeue(taskA)
1648 * // Must not decrement bucket->tasks here
1649 * dequeue(taskB)
1650 *
1651 * where we could end up with stale data in uc_se and
1652 * bucket[uc_se->bucket_id].
1653 *
1654 * The following check here eliminates the possibility of such race.
1655 */
1656 if (unlikely(!uc_se->active))
1657 return;
1658
1659 bucket = &uc_rq->bucket[uc_se->bucket_id];
1660
1661 SCHED_WARN_ON(!bucket->tasks);
1662 if (likely(bucket->tasks))
1663 bucket->tasks--;
1664
1665 uc_se->active = false;
1666
1667 /*
1668 * Keep "local max aggregation" simple and accept to (possibly)
1669 * overboost some RUNNABLE tasks in the same bucket.
1670 * The rq clamp bucket value is reset to its base value whenever
1671 * there are no more RUNNABLE tasks refcounting it.
1672 */
1673 if (likely(bucket->tasks))
1674 return;
1675
1676 rq_clamp = uclamp_rq_get(rq, clamp_id);
1677 /*
1678 * Defensive programming: this should never happen. If it happens,
1679 * e.g. due to future modification, warn and fixup the expected value.
1680 */
1681 SCHED_WARN_ON(bucket->value > rq_clamp);
1682 if (bucket->value >= rq_clamp) {
1683 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1684 uclamp_rq_set(rq, clamp_id, bkt_clamp);
1685 }
1686}
1687
1688static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1689{
1690 enum uclamp_id clamp_id;
1691
1692 /*
1693 * Avoid any overhead until uclamp is actually used by the userspace.
1694 *
1695 * The condition is constructed such that a NOP is generated when
1696 * sched_uclamp_used is disabled.
1697 */
1698 if (!static_branch_unlikely(&sched_uclamp_used))
1699 return;
1700
1701 if (unlikely(!p->sched_class->uclamp_enabled))
1702 return;
1703
1704 for_each_clamp_id(clamp_id)
1705 uclamp_rq_inc_id(rq, p, clamp_id);
1706
1707 /* Reset clamp idle holding when there is one RUNNABLE task */
1708 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1709 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1710}
1711
1712static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1713{
1714 enum uclamp_id clamp_id;
1715
1716 /*
1717 * Avoid any overhead until uclamp is actually used by the userspace.
1718 *
1719 * The condition is constructed such that a NOP is generated when
1720 * sched_uclamp_used is disabled.
1721 */
1722 if (!static_branch_unlikely(&sched_uclamp_used))
1723 return;
1724
1725 if (unlikely(!p->sched_class->uclamp_enabled))
1726 return;
1727
1728 for_each_clamp_id(clamp_id)
1729 uclamp_rq_dec_id(rq, p, clamp_id);
1730}
1731
1732static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1733 enum uclamp_id clamp_id)
1734{
1735 if (!p->uclamp[clamp_id].active)
1736 return;
1737
1738 uclamp_rq_dec_id(rq, p, clamp_id);
1739 uclamp_rq_inc_id(rq, p, clamp_id);
1740
1741 /*
1742 * Make sure to clear the idle flag if we've transiently reached 0
1743 * active tasks on rq.
1744 */
1745 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1746 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1747}
1748
1749static inline void
1750uclamp_update_active(struct task_struct *p)
1751{
1752 enum uclamp_id clamp_id;
1753 struct rq_flags rf;
1754 struct rq *rq;
1755
1756 /*
1757 * Lock the task and the rq where the task is (or was) queued.
1758 *
1759 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1760 * price to pay to safely serialize util_{min,max} updates with
1761 * enqueues, dequeues and migration operations.
1762 * This is the same locking schema used by __set_cpus_allowed_ptr().
1763 */
1764 rq = task_rq_lock(p, &rf);
1765
1766 /*
1767 * Setting the clamp bucket is serialized by task_rq_lock().
1768 * If the task is not yet RUNNABLE and its task_struct is not
1769 * affecting a valid clamp bucket, the next time it's enqueued,
1770 * it will already see the updated clamp bucket value.
1771 */
1772 for_each_clamp_id(clamp_id)
1773 uclamp_rq_reinc_id(rq, p, clamp_id);
1774
1775 task_rq_unlock(rq, p, &rf);
1776}
1777
1778#ifdef CONFIG_UCLAMP_TASK_GROUP
1779static inline void
1780uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1781{
1782 struct css_task_iter it;
1783 struct task_struct *p;
1784
1785 css_task_iter_start(css, 0, &it);
1786 while ((p = css_task_iter_next(&it)))
1787 uclamp_update_active(p);
1788 css_task_iter_end(&it);
1789}
1790
1791static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1792#endif
1793
1794#ifdef CONFIG_SYSCTL
1795#ifdef CONFIG_UCLAMP_TASK_GROUP
1796static void uclamp_update_root_tg(void)
1797{
1798 struct task_group *tg = &root_task_group;
1799
1800 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1801 sysctl_sched_uclamp_util_min, false);
1802 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1803 sysctl_sched_uclamp_util_max, false);
1804
1805 guard(rcu)();
1806 cpu_util_update_eff(&root_task_group.css);
1807}
1808#else
1809static void uclamp_update_root_tg(void) { }
1810#endif
1811
1812static void uclamp_sync_util_min_rt_default(void)
1813{
1814 struct task_struct *g, *p;
1815
1816 /*
1817 * copy_process() sysctl_uclamp
1818 * uclamp_min_rt = X;
1819 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1820 * // link thread smp_mb__after_spinlock()
1821 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1822 * sched_post_fork() for_each_process_thread()
1823 * __uclamp_sync_rt() __uclamp_sync_rt()
1824 *
1825 * Ensures that either sched_post_fork() will observe the new
1826 * uclamp_min_rt or for_each_process_thread() will observe the new
1827 * task.
1828 */
1829 read_lock(&tasklist_lock);
1830 smp_mb__after_spinlock();
1831 read_unlock(&tasklist_lock);
1832
1833 guard(rcu)();
1834 for_each_process_thread(g, p)
1835 uclamp_update_util_min_rt_default(p);
1836}
1837
1838static int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1839 void *buffer, size_t *lenp, loff_t *ppos)
1840{
1841 bool update_root_tg = false;
1842 int old_min, old_max, old_min_rt;
1843 int result;
1844
1845 guard(mutex)(&uclamp_mutex);
1846
1847 old_min = sysctl_sched_uclamp_util_min;
1848 old_max = sysctl_sched_uclamp_util_max;
1849 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1850
1851 result = proc_dointvec(table, write, buffer, lenp, ppos);
1852 if (result)
1853 goto undo;
1854 if (!write)
1855 return 0;
1856
1857 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1858 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1859 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1860
1861 result = -EINVAL;
1862 goto undo;
1863 }
1864
1865 if (old_min != sysctl_sched_uclamp_util_min) {
1866 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1867 sysctl_sched_uclamp_util_min, false);
1868 update_root_tg = true;
1869 }
1870 if (old_max != sysctl_sched_uclamp_util_max) {
1871 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1872 sysctl_sched_uclamp_util_max, false);
1873 update_root_tg = true;
1874 }
1875
1876 if (update_root_tg) {
1877 static_branch_enable(&sched_uclamp_used);
1878 uclamp_update_root_tg();
1879 }
1880
1881 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1882 static_branch_enable(&sched_uclamp_used);
1883 uclamp_sync_util_min_rt_default();
1884 }
1885
1886 /*
1887 * We update all RUNNABLE tasks only when task groups are in use.
1888 * Otherwise, keep it simple and do just a lazy update at each next
1889 * task enqueue time.
1890 */
1891 return 0;
1892
1893undo:
1894 sysctl_sched_uclamp_util_min = old_min;
1895 sysctl_sched_uclamp_util_max = old_max;
1896 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1897 return result;
1898}
1899#endif
1900
1901static int uclamp_validate(struct task_struct *p,
1902 const struct sched_attr *attr)
1903{
1904 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1905 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1906
1907 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1908 util_min = attr->sched_util_min;
1909
1910 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1911 return -EINVAL;
1912 }
1913
1914 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1915 util_max = attr->sched_util_max;
1916
1917 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1918 return -EINVAL;
1919 }
1920
1921 if (util_min != -1 && util_max != -1 && util_min > util_max)
1922 return -EINVAL;
1923
1924 /*
1925 * We have valid uclamp attributes; make sure uclamp is enabled.
1926 *
1927 * We need to do that here, because enabling static branches is a
1928 * blocking operation which obviously cannot be done while holding
1929 * scheduler locks.
1930 */
1931 static_branch_enable(&sched_uclamp_used);
1932
1933 return 0;
1934}
1935
1936static bool uclamp_reset(const struct sched_attr *attr,
1937 enum uclamp_id clamp_id,
1938 struct uclamp_se *uc_se)
1939{
1940 /* Reset on sched class change for a non user-defined clamp value. */
1941 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1942 !uc_se->user_defined)
1943 return true;
1944
1945 /* Reset on sched_util_{min,max} == -1. */
1946 if (clamp_id == UCLAMP_MIN &&
1947 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1948 attr->sched_util_min == -1) {
1949 return true;
1950 }
1951
1952 if (clamp_id == UCLAMP_MAX &&
1953 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1954 attr->sched_util_max == -1) {
1955 return true;
1956 }
1957
1958 return false;
1959}
1960
1961static void __setscheduler_uclamp(struct task_struct *p,
1962 const struct sched_attr *attr)
1963{
1964 enum uclamp_id clamp_id;
1965
1966 for_each_clamp_id(clamp_id) {
1967 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1968 unsigned int value;
1969
1970 if (!uclamp_reset(attr, clamp_id, uc_se))
1971 continue;
1972
1973 /*
1974 * RT by default have a 100% boost value that could be modified
1975 * at runtime.
1976 */
1977 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1978 value = sysctl_sched_uclamp_util_min_rt_default;
1979 else
1980 value = uclamp_none(clamp_id);
1981
1982 uclamp_se_set(uc_se, value, false);
1983
1984 }
1985
1986 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1987 return;
1988
1989 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1990 attr->sched_util_min != -1) {
1991 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1992 attr->sched_util_min, true);
1993 }
1994
1995 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1996 attr->sched_util_max != -1) {
1997 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1998 attr->sched_util_max, true);
1999 }
2000}
2001
2002static void uclamp_fork(struct task_struct *p)
2003{
2004 enum uclamp_id clamp_id;
2005
2006 /*
2007 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
2008 * as the task is still at its early fork stages.
2009 */
2010 for_each_clamp_id(clamp_id)
2011 p->uclamp[clamp_id].active = false;
2012
2013 if (likely(!p->sched_reset_on_fork))
2014 return;
2015
2016 for_each_clamp_id(clamp_id) {
2017 uclamp_se_set(&p->uclamp_req[clamp_id],
2018 uclamp_none(clamp_id), false);
2019 }
2020}
2021
2022static void uclamp_post_fork(struct task_struct *p)
2023{
2024 uclamp_update_util_min_rt_default(p);
2025}
2026
2027static void __init init_uclamp_rq(struct rq *rq)
2028{
2029 enum uclamp_id clamp_id;
2030 struct uclamp_rq *uc_rq = rq->uclamp;
2031
2032 for_each_clamp_id(clamp_id) {
2033 uc_rq[clamp_id] = (struct uclamp_rq) {
2034 .value = uclamp_none(clamp_id)
2035 };
2036 }
2037
2038 rq->uclamp_flags = UCLAMP_FLAG_IDLE;
2039}
2040
2041static void __init init_uclamp(void)
2042{
2043 struct uclamp_se uc_max = {};
2044 enum uclamp_id clamp_id;
2045 int cpu;
2046
2047 for_each_possible_cpu(cpu)
2048 init_uclamp_rq(cpu_rq(cpu));
2049
2050 for_each_clamp_id(clamp_id) {
2051 uclamp_se_set(&init_task.uclamp_req[clamp_id],
2052 uclamp_none(clamp_id), false);
2053 }
2054
2055 /* System defaults allow max clamp values for both indexes */
2056 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
2057 for_each_clamp_id(clamp_id) {
2058 uclamp_default[clamp_id] = uc_max;
2059#ifdef CONFIG_UCLAMP_TASK_GROUP
2060 root_task_group.uclamp_req[clamp_id] = uc_max;
2061 root_task_group.uclamp[clamp_id] = uc_max;
2062#endif
2063 }
2064}
2065
2066#else /* !CONFIG_UCLAMP_TASK */
2067static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
2068static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
2069static inline int uclamp_validate(struct task_struct *p,
2070 const struct sched_attr *attr)
2071{
2072 return -EOPNOTSUPP;
2073}
2074static void __setscheduler_uclamp(struct task_struct *p,
2075 const struct sched_attr *attr) { }
2076static inline void uclamp_fork(struct task_struct *p) { }
2077static inline void uclamp_post_fork(struct task_struct *p) { }
2078static inline void init_uclamp(void) { }
2079#endif /* CONFIG_UCLAMP_TASK */
2080
2081bool sched_task_on_rq(struct task_struct *p)
2082{
2083 return task_on_rq_queued(p);
2084}
2085
2086unsigned long get_wchan(struct task_struct *p)
2087{
2088 unsigned long ip = 0;
2089 unsigned int state;
2090
2091 if (!p || p == current)
2092 return 0;
2093
2094 /* Only get wchan if task is blocked and we can keep it that way. */
2095 raw_spin_lock_irq(&p->pi_lock);
2096 state = READ_ONCE(p->__state);
2097 smp_rmb(); /* see try_to_wake_up() */
2098 if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
2099 ip = __get_wchan(p);
2100 raw_spin_unlock_irq(&p->pi_lock);
2101
2102 return ip;
2103}
2104
2105static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2106{
2107 if (!(flags & ENQUEUE_NOCLOCK))
2108 update_rq_clock(rq);
2109
2110 if (!(flags & ENQUEUE_RESTORE)) {
2111 sched_info_enqueue(rq, p);
2112 psi_enqueue(p, (flags & ENQUEUE_WAKEUP) && !(flags & ENQUEUE_MIGRATED));
2113 }
2114
2115 uclamp_rq_inc(rq, p);
2116 p->sched_class->enqueue_task(rq, p, flags);
2117
2118 if (sched_core_enabled(rq))
2119 sched_core_enqueue(rq, p);
2120}
2121
2122static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2123{
2124 if (sched_core_enabled(rq))
2125 sched_core_dequeue(rq, p, flags);
2126
2127 if (!(flags & DEQUEUE_NOCLOCK))
2128 update_rq_clock(rq);
2129
2130 if (!(flags & DEQUEUE_SAVE)) {
2131 sched_info_dequeue(rq, p);
2132 psi_dequeue(p, flags & DEQUEUE_SLEEP);
2133 }
2134
2135 uclamp_rq_dec(rq, p);
2136 p->sched_class->dequeue_task(rq, p, flags);
2137}
2138
2139void activate_task(struct rq *rq, struct task_struct *p, int flags)
2140{
2141 if (task_on_rq_migrating(p))
2142 flags |= ENQUEUE_MIGRATED;
2143 if (flags & ENQUEUE_MIGRATED)
2144 sched_mm_cid_migrate_to(rq, p);
2145
2146 enqueue_task(rq, p, flags);
2147
2148 WRITE_ONCE(p->on_rq, TASK_ON_RQ_QUEUED);
2149 ASSERT_EXCLUSIVE_WRITER(p->on_rq);
2150}
2151
2152void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2153{
2154 WRITE_ONCE(p->on_rq, (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING);
2155 ASSERT_EXCLUSIVE_WRITER(p->on_rq);
2156
2157 dequeue_task(rq, p, flags);
2158}
2159
2160static inline int __normal_prio(int policy, int rt_prio, int nice)
2161{
2162 int prio;
2163
2164 if (dl_policy(policy))
2165 prio = MAX_DL_PRIO - 1;
2166 else if (rt_policy(policy))
2167 prio = MAX_RT_PRIO - 1 - rt_prio;
2168 else
2169 prio = NICE_TO_PRIO(nice);
2170
2171 return prio;
2172}
2173
2174/*
2175 * Calculate the expected normal priority: i.e. priority
2176 * without taking RT-inheritance into account. Might be
2177 * boosted by interactivity modifiers. Changes upon fork,
2178 * setprio syscalls, and whenever the interactivity
2179 * estimator recalculates.
2180 */
2181static inline int normal_prio(struct task_struct *p)
2182{
2183 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2184}
2185
2186/*
2187 * Calculate the current priority, i.e. the priority
2188 * taken into account by the scheduler. This value might
2189 * be boosted by RT tasks, or might be boosted by
2190 * interactivity modifiers. Will be RT if the task got
2191 * RT-boosted. If not then it returns p->normal_prio.
2192 */
2193static int effective_prio(struct task_struct *p)
2194{
2195 p->normal_prio = normal_prio(p);
2196 /*
2197 * If we are RT tasks or we were boosted to RT priority,
2198 * keep the priority unchanged. Otherwise, update priority
2199 * to the normal priority:
2200 */
2201 if (!rt_prio(p->prio))
2202 return p->normal_prio;
2203 return p->prio;
2204}
2205
2206/**
2207 * task_curr - is this task currently executing on a CPU?
2208 * @p: the task in question.
2209 *
2210 * Return: 1 if the task is currently executing. 0 otherwise.
2211 */
2212inline int task_curr(const struct task_struct *p)
2213{
2214 return cpu_curr(task_cpu(p)) == p;
2215}
2216
2217/*
2218 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2219 * use the balance_callback list if you want balancing.
2220 *
2221 * this means any call to check_class_changed() must be followed by a call to
2222 * balance_callback().
2223 */
2224static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2225 const struct sched_class *prev_class,
2226 int oldprio)
2227{
2228 if (prev_class != p->sched_class) {
2229 if (prev_class->switched_from)
2230 prev_class->switched_from(rq, p);
2231
2232 p->sched_class->switched_to(rq, p);
2233 } else if (oldprio != p->prio || dl_task(p))
2234 p->sched_class->prio_changed(rq, p, oldprio);
2235}
2236
2237void wakeup_preempt(struct rq *rq, struct task_struct *p, int flags)
2238{
2239 if (p->sched_class == rq->curr->sched_class)
2240 rq->curr->sched_class->wakeup_preempt(rq, p, flags);
2241 else if (sched_class_above(p->sched_class, rq->curr->sched_class))
2242 resched_curr(rq);
2243
2244 /*
2245 * A queue event has occurred, and we're going to schedule. In
2246 * this case, we can save a useless back to back clock update.
2247 */
2248 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2249 rq_clock_skip_update(rq);
2250}
2251
2252static __always_inline
2253int __task_state_match(struct task_struct *p, unsigned int state)
2254{
2255 if (READ_ONCE(p->__state) & state)
2256 return 1;
2257
2258 if (READ_ONCE(p->saved_state) & state)
2259 return -1;
2260
2261 return 0;
2262}
2263
2264static __always_inline
2265int task_state_match(struct task_struct *p, unsigned int state)
2266{
2267 /*
2268 * Serialize against current_save_and_set_rtlock_wait_state(),
2269 * current_restore_rtlock_saved_state(), and __refrigerator().
2270 */
2271 guard(raw_spinlock_irq)(&p->pi_lock);
2272 return __task_state_match(p, state);
2273}
2274
2275/*
2276 * wait_task_inactive - wait for a thread to unschedule.
2277 *
2278 * Wait for the thread to block in any of the states set in @match_state.
2279 * If it changes, i.e. @p might have woken up, then return zero. When we
2280 * succeed in waiting for @p to be off its CPU, we return a positive number
2281 * (its total switch count). If a second call a short while later returns the
2282 * same number, the caller can be sure that @p has remained unscheduled the
2283 * whole time.
2284 *
2285 * The caller must ensure that the task *will* unschedule sometime soon,
2286 * else this function might spin for a *long* time. This function can't
2287 * be called with interrupts off, or it may introduce deadlock with
2288 * smp_call_function() if an IPI is sent by the same process we are
2289 * waiting to become inactive.
2290 */
2291unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
2292{
2293 int running, queued, match;
2294 struct rq_flags rf;
2295 unsigned long ncsw;
2296 struct rq *rq;
2297
2298 for (;;) {
2299 /*
2300 * We do the initial early heuristics without holding
2301 * any task-queue locks at all. We'll only try to get
2302 * the runqueue lock when things look like they will
2303 * work out!
2304 */
2305 rq = task_rq(p);
2306
2307 /*
2308 * If the task is actively running on another CPU
2309 * still, just relax and busy-wait without holding
2310 * any locks.
2311 *
2312 * NOTE! Since we don't hold any locks, it's not
2313 * even sure that "rq" stays as the right runqueue!
2314 * But we don't care, since "task_on_cpu()" will
2315 * return false if the runqueue has changed and p
2316 * is actually now running somewhere else!
2317 */
2318 while (task_on_cpu(rq, p)) {
2319 if (!task_state_match(p, match_state))
2320 return 0;
2321 cpu_relax();
2322 }
2323
2324 /*
2325 * Ok, time to look more closely! We need the rq
2326 * lock now, to be *sure*. If we're wrong, we'll
2327 * just go back and repeat.
2328 */
2329 rq = task_rq_lock(p, &rf);
2330 trace_sched_wait_task(p);
2331 running = task_on_cpu(rq, p);
2332 queued = task_on_rq_queued(p);
2333 ncsw = 0;
2334 if ((match = __task_state_match(p, match_state))) {
2335 /*
2336 * When matching on p->saved_state, consider this task
2337 * still queued so it will wait.
2338 */
2339 if (match < 0)
2340 queued = 1;
2341 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2342 }
2343 task_rq_unlock(rq, p, &rf);
2344
2345 /*
2346 * If it changed from the expected state, bail out now.
2347 */
2348 if (unlikely(!ncsw))
2349 break;
2350
2351 /*
2352 * Was it really running after all now that we
2353 * checked with the proper locks actually held?
2354 *
2355 * Oops. Go back and try again..
2356 */
2357 if (unlikely(running)) {
2358 cpu_relax();
2359 continue;
2360 }
2361
2362 /*
2363 * It's not enough that it's not actively running,
2364 * it must be off the runqueue _entirely_, and not
2365 * preempted!
2366 *
2367 * So if it was still runnable (but just not actively
2368 * running right now), it's preempted, and we should
2369 * yield - it could be a while.
2370 */
2371 if (unlikely(queued)) {
2372 ktime_t to = NSEC_PER_SEC / HZ;
2373
2374 set_current_state(TASK_UNINTERRUPTIBLE);
2375 schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
2376 continue;
2377 }
2378
2379 /*
2380 * Ahh, all good. It wasn't running, and it wasn't
2381 * runnable, which means that it will never become
2382 * running in the future either. We're all done!
2383 */
2384 break;
2385 }
2386
2387 return ncsw;
2388}
2389
2390#ifdef CONFIG_SMP
2391
2392static void
2393__do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx);
2394
2395static int __set_cpus_allowed_ptr(struct task_struct *p,
2396 struct affinity_context *ctx);
2397
2398static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2399{
2400 struct affinity_context ac = {
2401 .new_mask = cpumask_of(rq->cpu),
2402 .flags = SCA_MIGRATE_DISABLE,
2403 };
2404
2405 if (likely(!p->migration_disabled))
2406 return;
2407
2408 if (p->cpus_ptr != &p->cpus_mask)
2409 return;
2410
2411 /*
2412 * Violates locking rules! see comment in __do_set_cpus_allowed().
2413 */
2414 __do_set_cpus_allowed(p, &ac);
2415}
2416
2417void migrate_disable(void)
2418{
2419 struct task_struct *p = current;
2420
2421 if (p->migration_disabled) {
2422 p->migration_disabled++;
2423 return;
2424 }
2425
2426 guard(preempt)();
2427 this_rq()->nr_pinned++;
2428 p->migration_disabled = 1;
2429}
2430EXPORT_SYMBOL_GPL(migrate_disable);
2431
2432void migrate_enable(void)
2433{
2434 struct task_struct *p = current;
2435 struct affinity_context ac = {
2436 .new_mask = &p->cpus_mask,
2437 .flags = SCA_MIGRATE_ENABLE,
2438 };
2439
2440 if (p->migration_disabled > 1) {
2441 p->migration_disabled--;
2442 return;
2443 }
2444
2445 if (WARN_ON_ONCE(!p->migration_disabled))
2446 return;
2447
2448 /*
2449 * Ensure stop_task runs either before or after this, and that
2450 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2451 */
2452 guard(preempt)();
2453 if (p->cpus_ptr != &p->cpus_mask)
2454 __set_cpus_allowed_ptr(p, &ac);
2455 /*
2456 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2457 * regular cpus_mask, otherwise things that race (eg.
2458 * select_fallback_rq) get confused.
2459 */
2460 barrier();
2461 p->migration_disabled = 0;
2462 this_rq()->nr_pinned--;
2463}
2464EXPORT_SYMBOL_GPL(migrate_enable);
2465
2466static inline bool rq_has_pinned_tasks(struct rq *rq)
2467{
2468 return rq->nr_pinned;
2469}
2470
2471/*
2472 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2473 * __set_cpus_allowed_ptr() and select_fallback_rq().
2474 */
2475static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2476{
2477 /* When not in the task's cpumask, no point in looking further. */
2478 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2479 return false;
2480
2481 /* migrate_disabled() must be allowed to finish. */
2482 if (is_migration_disabled(p))
2483 return cpu_online(cpu);
2484
2485 /* Non kernel threads are not allowed during either online or offline. */
2486 if (!(p->flags & PF_KTHREAD))
2487 return cpu_active(cpu) && task_cpu_possible(cpu, p);
2488
2489 /* KTHREAD_IS_PER_CPU is always allowed. */
2490 if (kthread_is_per_cpu(p))
2491 return cpu_online(cpu);
2492
2493 /* Regular kernel threads don't get to stay during offline. */
2494 if (cpu_dying(cpu))
2495 return false;
2496
2497 /* But are allowed during online. */
2498 return cpu_online(cpu);
2499}
2500
2501/*
2502 * This is how migration works:
2503 *
2504 * 1) we invoke migration_cpu_stop() on the target CPU using
2505 * stop_one_cpu().
2506 * 2) stopper starts to run (implicitly forcing the migrated thread
2507 * off the CPU)
2508 * 3) it checks whether the migrated task is still in the wrong runqueue.
2509 * 4) if it's in the wrong runqueue then the migration thread removes
2510 * it and puts it into the right queue.
2511 * 5) stopper completes and stop_one_cpu() returns and the migration
2512 * is done.
2513 */
2514
2515/*
2516 * move_queued_task - move a queued task to new rq.
2517 *
2518 * Returns (locked) new rq. Old rq's lock is released.
2519 */
2520static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2521 struct task_struct *p, int new_cpu)
2522{
2523 lockdep_assert_rq_held(rq);
2524
2525 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2526 set_task_cpu(p, new_cpu);
2527 rq_unlock(rq, rf);
2528
2529 rq = cpu_rq(new_cpu);
2530
2531 rq_lock(rq, rf);
2532 WARN_ON_ONCE(task_cpu(p) != new_cpu);
2533 activate_task(rq, p, 0);
2534 wakeup_preempt(rq, p, 0);
2535
2536 return rq;
2537}
2538
2539struct migration_arg {
2540 struct task_struct *task;
2541 int dest_cpu;
2542 struct set_affinity_pending *pending;
2543};
2544
2545/*
2546 * @refs: number of wait_for_completion()
2547 * @stop_pending: is @stop_work in use
2548 */
2549struct set_affinity_pending {
2550 refcount_t refs;
2551 unsigned int stop_pending;
2552 struct completion done;
2553 struct cpu_stop_work stop_work;
2554 struct migration_arg arg;
2555};
2556
2557/*
2558 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2559 * this because either it can't run here any more (set_cpus_allowed()
2560 * away from this CPU, or CPU going down), or because we're
2561 * attempting to rebalance this task on exec (sched_exec).
2562 *
2563 * So we race with normal scheduler movements, but that's OK, as long
2564 * as the task is no longer on this CPU.
2565 */
2566static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2567 struct task_struct *p, int dest_cpu)
2568{
2569 /* Affinity changed (again). */
2570 if (!is_cpu_allowed(p, dest_cpu))
2571 return rq;
2572
2573 rq = move_queued_task(rq, rf, p, dest_cpu);
2574
2575 return rq;
2576}
2577
2578/*
2579 * migration_cpu_stop - this will be executed by a highprio stopper thread
2580 * and performs thread migration by bumping thread off CPU then
2581 * 'pushing' onto another runqueue.
2582 */
2583static int migration_cpu_stop(void *data)
2584{
2585 struct migration_arg *arg = data;
2586 struct set_affinity_pending *pending = arg->pending;
2587 struct task_struct *p = arg->task;
2588 struct rq *rq = this_rq();
2589 bool complete = false;
2590 struct rq_flags rf;
2591
2592 /*
2593 * The original target CPU might have gone down and we might
2594 * be on another CPU but it doesn't matter.
2595 */
2596 local_irq_save(rf.flags);
2597 /*
2598 * We need to explicitly wake pending tasks before running
2599 * __migrate_task() such that we will not miss enforcing cpus_ptr
2600 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2601 */
2602 flush_smp_call_function_queue();
2603
2604 raw_spin_lock(&p->pi_lock);
2605 rq_lock(rq, &rf);
2606
2607 /*
2608 * If we were passed a pending, then ->stop_pending was set, thus
2609 * p->migration_pending must have remained stable.
2610 */
2611 WARN_ON_ONCE(pending && pending != p->migration_pending);
2612
2613 /*
2614 * If task_rq(p) != rq, it cannot be migrated here, because we're
2615 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2616 * we're holding p->pi_lock.
2617 */
2618 if (task_rq(p) == rq) {
2619 if (is_migration_disabled(p))
2620 goto out;
2621
2622 if (pending) {
2623 p->migration_pending = NULL;
2624 complete = true;
2625
2626 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2627 goto out;
2628 }
2629
2630 if (task_on_rq_queued(p)) {
2631 update_rq_clock(rq);
2632 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2633 } else {
2634 p->wake_cpu = arg->dest_cpu;
2635 }
2636
2637 /*
2638 * XXX __migrate_task() can fail, at which point we might end
2639 * up running on a dodgy CPU, AFAICT this can only happen
2640 * during CPU hotplug, at which point we'll get pushed out
2641 * anyway, so it's probably not a big deal.
2642 */
2643
2644 } else if (pending) {
2645 /*
2646 * This happens when we get migrated between migrate_enable()'s
2647 * preempt_enable() and scheduling the stopper task. At that
2648 * point we're a regular task again and not current anymore.
2649 *
2650 * A !PREEMPT kernel has a giant hole here, which makes it far
2651 * more likely.
2652 */
2653
2654 /*
2655 * The task moved before the stopper got to run. We're holding
2656 * ->pi_lock, so the allowed mask is stable - if it got
2657 * somewhere allowed, we're done.
2658 */
2659 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2660 p->migration_pending = NULL;
2661 complete = true;
2662 goto out;
2663 }
2664
2665 /*
2666 * When migrate_enable() hits a rq mis-match we can't reliably
2667 * determine is_migration_disabled() and so have to chase after
2668 * it.
2669 */
2670 WARN_ON_ONCE(!pending->stop_pending);
2671 preempt_disable();
2672 task_rq_unlock(rq, p, &rf);
2673 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2674 &pending->arg, &pending->stop_work);
2675 preempt_enable();
2676 return 0;
2677 }
2678out:
2679 if (pending)
2680 pending->stop_pending = false;
2681 task_rq_unlock(rq, p, &rf);
2682
2683 if (complete)
2684 complete_all(&pending->done);
2685
2686 return 0;
2687}
2688
2689int push_cpu_stop(void *arg)
2690{
2691 struct rq *lowest_rq = NULL, *rq = this_rq();
2692 struct task_struct *p = arg;
2693
2694 raw_spin_lock_irq(&p->pi_lock);
2695 raw_spin_rq_lock(rq);
2696
2697 if (task_rq(p) != rq)
2698 goto out_unlock;
2699
2700 if (is_migration_disabled(p)) {
2701 p->migration_flags |= MDF_PUSH;
2702 goto out_unlock;
2703 }
2704
2705 p->migration_flags &= ~MDF_PUSH;
2706
2707 if (p->sched_class->find_lock_rq)
2708 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2709
2710 if (!lowest_rq)
2711 goto out_unlock;
2712
2713 // XXX validate p is still the highest prio task
2714 if (task_rq(p) == rq) {
2715 deactivate_task(rq, p, 0);
2716 set_task_cpu(p, lowest_rq->cpu);
2717 activate_task(lowest_rq, p, 0);
2718 resched_curr(lowest_rq);
2719 }
2720
2721 double_unlock_balance(rq, lowest_rq);
2722
2723out_unlock:
2724 rq->push_busy = false;
2725 raw_spin_rq_unlock(rq);
2726 raw_spin_unlock_irq(&p->pi_lock);
2727
2728 put_task_struct(p);
2729 return 0;
2730}
2731
2732/*
2733 * sched_class::set_cpus_allowed must do the below, but is not required to
2734 * actually call this function.
2735 */
2736void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx)
2737{
2738 if (ctx->flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2739 p->cpus_ptr = ctx->new_mask;
2740 return;
2741 }
2742
2743 cpumask_copy(&p->cpus_mask, ctx->new_mask);
2744 p->nr_cpus_allowed = cpumask_weight(ctx->new_mask);
2745
2746 /*
2747 * Swap in a new user_cpus_ptr if SCA_USER flag set
2748 */
2749 if (ctx->flags & SCA_USER)
2750 swap(p->user_cpus_ptr, ctx->user_mask);
2751}
2752
2753static void
2754__do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx)
2755{
2756 struct rq *rq = task_rq(p);
2757 bool queued, running;
2758
2759 /*
2760 * This here violates the locking rules for affinity, since we're only
2761 * supposed to change these variables while holding both rq->lock and
2762 * p->pi_lock.
2763 *
2764 * HOWEVER, it magically works, because ttwu() is the only code that
2765 * accesses these variables under p->pi_lock and only does so after
2766 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2767 * before finish_task().
2768 *
2769 * XXX do further audits, this smells like something putrid.
2770 */
2771 if (ctx->flags & SCA_MIGRATE_DISABLE)
2772 SCHED_WARN_ON(!p->on_cpu);
2773 else
2774 lockdep_assert_held(&p->pi_lock);
2775
2776 queued = task_on_rq_queued(p);
2777 running = task_current(rq, p);
2778
2779 if (queued) {
2780 /*
2781 * Because __kthread_bind() calls this on blocked tasks without
2782 * holding rq->lock.
2783 */
2784 lockdep_assert_rq_held(rq);
2785 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2786 }
2787 if (running)
2788 put_prev_task(rq, p);
2789
2790 p->sched_class->set_cpus_allowed(p, ctx);
2791
2792 if (queued)
2793 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2794 if (running)
2795 set_next_task(rq, p);
2796}
2797
2798/*
2799 * Used for kthread_bind() and select_fallback_rq(), in both cases the user
2800 * affinity (if any) should be destroyed too.
2801 */
2802void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2803{
2804 struct affinity_context ac = {
2805 .new_mask = new_mask,
2806 .user_mask = NULL,
2807 .flags = SCA_USER, /* clear the user requested mask */
2808 };
2809 union cpumask_rcuhead {
2810 cpumask_t cpumask;
2811 struct rcu_head rcu;
2812 };
2813
2814 __do_set_cpus_allowed(p, &ac);
2815
2816 /*
2817 * Because this is called with p->pi_lock held, it is not possible
2818 * to use kfree() here (when PREEMPT_RT=y), therefore punt to using
2819 * kfree_rcu().
2820 */
2821 kfree_rcu((union cpumask_rcuhead *)ac.user_mask, rcu);
2822}
2823
2824static cpumask_t *alloc_user_cpus_ptr(int node)
2825{
2826 /*
2827 * See do_set_cpus_allowed() above for the rcu_head usage.
2828 */
2829 int size = max_t(int, cpumask_size(), sizeof(struct rcu_head));
2830
2831 return kmalloc_node(size, GFP_KERNEL, node);
2832}
2833
2834int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2835 int node)
2836{
2837 cpumask_t *user_mask;
2838 unsigned long flags;
2839
2840 /*
2841 * Always clear dst->user_cpus_ptr first as their user_cpus_ptr's
2842 * may differ by now due to racing.
2843 */
2844 dst->user_cpus_ptr = NULL;
2845
2846 /*
2847 * This check is racy and losing the race is a valid situation.
2848 * It is not worth the extra overhead of taking the pi_lock on
2849 * every fork/clone.
2850 */
2851 if (data_race(!src->user_cpus_ptr))
2852 return 0;
2853
2854 user_mask = alloc_user_cpus_ptr(node);
2855 if (!user_mask)
2856 return -ENOMEM;
2857
2858 /*
2859 * Use pi_lock to protect content of user_cpus_ptr
2860 *
2861 * Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent
2862 * do_set_cpus_allowed().
2863 */
2864 raw_spin_lock_irqsave(&src->pi_lock, flags);
2865 if (src->user_cpus_ptr) {
2866 swap(dst->user_cpus_ptr, user_mask);
2867 cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2868 }
2869 raw_spin_unlock_irqrestore(&src->pi_lock, flags);
2870
2871 if (unlikely(user_mask))
2872 kfree(user_mask);
2873
2874 return 0;
2875}
2876
2877static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2878{
2879 struct cpumask *user_mask = NULL;
2880
2881 swap(p->user_cpus_ptr, user_mask);
2882
2883 return user_mask;
2884}
2885
2886void release_user_cpus_ptr(struct task_struct *p)
2887{
2888 kfree(clear_user_cpus_ptr(p));
2889}
2890
2891/*
2892 * This function is wildly self concurrent; here be dragons.
2893 *
2894 *
2895 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2896 * designated task is enqueued on an allowed CPU. If that task is currently
2897 * running, we have to kick it out using the CPU stopper.
2898 *
2899 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2900 * Consider:
2901 *
2902 * Initial conditions: P0->cpus_mask = [0, 1]
2903 *
2904 * P0@CPU0 P1
2905 *
2906 * migrate_disable();
2907 * <preempted>
2908 * set_cpus_allowed_ptr(P0, [1]);
2909 *
2910 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2911 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2912 * This means we need the following scheme:
2913 *
2914 * P0@CPU0 P1
2915 *
2916 * migrate_disable();
2917 * <preempted>
2918 * set_cpus_allowed_ptr(P0, [1]);
2919 * <blocks>
2920 * <resumes>
2921 * migrate_enable();
2922 * __set_cpus_allowed_ptr();
2923 * <wakes local stopper>
2924 * `--> <woken on migration completion>
2925 *
2926 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2927 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2928 * task p are serialized by p->pi_lock, which we can leverage: the one that
2929 * should come into effect at the end of the Migrate-Disable region is the last
2930 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2931 * but we still need to properly signal those waiting tasks at the appropriate
2932 * moment.
2933 *
2934 * This is implemented using struct set_affinity_pending. The first
2935 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2936 * setup an instance of that struct and install it on the targeted task_struct.
2937 * Any and all further callers will reuse that instance. Those then wait for
2938 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2939 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2940 *
2941 *
2942 * (1) In the cases covered above. There is one more where the completion is
2943 * signaled within affine_move_task() itself: when a subsequent affinity request
2944 * occurs after the stopper bailed out due to the targeted task still being
2945 * Migrate-Disable. Consider:
2946 *
2947 * Initial conditions: P0->cpus_mask = [0, 1]
2948 *
2949 * CPU0 P1 P2
2950 * <P0>
2951 * migrate_disable();
2952 * <preempted>
2953 * set_cpus_allowed_ptr(P0, [1]);
2954 * <blocks>
2955 * <migration/0>
2956 * migration_cpu_stop()
2957 * is_migration_disabled()
2958 * <bails>
2959 * set_cpus_allowed_ptr(P0, [0, 1]);
2960 * <signal completion>
2961 * <awakes>
2962 *
2963 * Note that the above is safe vs a concurrent migrate_enable(), as any
2964 * pending affinity completion is preceded by an uninstallation of
2965 * p->migration_pending done with p->pi_lock held.
2966 */
2967static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2968 int dest_cpu, unsigned int flags)
2969 __releases(rq->lock)
2970 __releases(p->pi_lock)
2971{
2972 struct set_affinity_pending my_pending = { }, *pending = NULL;
2973 bool stop_pending, complete = false;
2974
2975 /* Can the task run on the task's current CPU? If so, we're done */
2976 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2977 struct task_struct *push_task = NULL;
2978
2979 if ((flags & SCA_MIGRATE_ENABLE) &&
2980 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2981 rq->push_busy = true;
2982 push_task = get_task_struct(p);
2983 }
2984
2985 /*
2986 * If there are pending waiters, but no pending stop_work,
2987 * then complete now.
2988 */
2989 pending = p->migration_pending;
2990 if (pending && !pending->stop_pending) {
2991 p->migration_pending = NULL;
2992 complete = true;
2993 }
2994
2995 preempt_disable();
2996 task_rq_unlock(rq, p, rf);
2997 if (push_task) {
2998 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2999 p, &rq->push_work);
3000 }
3001 preempt_enable();
3002
3003 if (complete)
3004 complete_all(&pending->done);
3005
3006 return 0;
3007 }
3008
3009 if (!(flags & SCA_MIGRATE_ENABLE)) {
3010 /* serialized by p->pi_lock */
3011 if (!p->migration_pending) {
3012 /* Install the request */
3013 refcount_set(&my_pending.refs, 1);
3014 init_completion(&my_pending.done);
3015 my_pending.arg = (struct migration_arg) {
3016 .task = p,
3017 .dest_cpu = dest_cpu,
3018 .pending = &my_pending,
3019 };
3020
3021 p->migration_pending = &my_pending;
3022 } else {
3023 pending = p->migration_pending;
3024 refcount_inc(&pending->refs);
3025 /*
3026 * Affinity has changed, but we've already installed a
3027 * pending. migration_cpu_stop() *must* see this, else
3028 * we risk a completion of the pending despite having a
3029 * task on a disallowed CPU.
3030 *
3031 * Serialized by p->pi_lock, so this is safe.
3032 */
3033 pending->arg.dest_cpu = dest_cpu;
3034 }
3035 }
3036 pending = p->migration_pending;
3037 /*
3038 * - !MIGRATE_ENABLE:
3039 * we'll have installed a pending if there wasn't one already.
3040 *
3041 * - MIGRATE_ENABLE:
3042 * we're here because the current CPU isn't matching anymore,
3043 * the only way that can happen is because of a concurrent
3044 * set_cpus_allowed_ptr() call, which should then still be
3045 * pending completion.
3046 *
3047 * Either way, we really should have a @pending here.
3048 */
3049 if (WARN_ON_ONCE(!pending)) {
3050 task_rq_unlock(rq, p, rf);
3051 return -EINVAL;
3052 }
3053
3054 if (task_on_cpu(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
3055 /*
3056 * MIGRATE_ENABLE gets here because 'p == current', but for
3057 * anything else we cannot do is_migration_disabled(), punt
3058 * and have the stopper function handle it all race-free.
3059 */
3060 stop_pending = pending->stop_pending;
3061 if (!stop_pending)
3062 pending->stop_pending = true;
3063
3064 if (flags & SCA_MIGRATE_ENABLE)
3065 p->migration_flags &= ~MDF_PUSH;
3066
3067 preempt_disable();
3068 task_rq_unlock(rq, p, rf);
3069 if (!stop_pending) {
3070 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
3071 &pending->arg, &pending->stop_work);
3072 }
3073 preempt_enable();
3074
3075 if (flags & SCA_MIGRATE_ENABLE)
3076 return 0;
3077 } else {
3078
3079 if (!is_migration_disabled(p)) {
3080 if (task_on_rq_queued(p))
3081 rq = move_queued_task(rq, rf, p, dest_cpu);
3082
3083 if (!pending->stop_pending) {
3084 p->migration_pending = NULL;
3085 complete = true;
3086 }
3087 }
3088 task_rq_unlock(rq, p, rf);
3089
3090 if (complete)
3091 complete_all(&pending->done);
3092 }
3093
3094 wait_for_completion(&pending->done);
3095
3096 if (refcount_dec_and_test(&pending->refs))
3097 wake_up_var(&pending->refs); /* No UaF, just an address */
3098
3099 /*
3100 * Block the original owner of &pending until all subsequent callers
3101 * have seen the completion and decremented the refcount
3102 */
3103 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
3104
3105 /* ARGH */
3106 WARN_ON_ONCE(my_pending.stop_pending);
3107
3108 return 0;
3109}
3110
3111/*
3112 * Called with both p->pi_lock and rq->lock held; drops both before returning.
3113 */
3114static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
3115 struct affinity_context *ctx,
3116 struct rq *rq,
3117 struct rq_flags *rf)
3118 __releases(rq->lock)
3119 __releases(p->pi_lock)
3120{
3121 const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
3122 const struct cpumask *cpu_valid_mask = cpu_active_mask;
3123 bool kthread = p->flags & PF_KTHREAD;
3124 unsigned int dest_cpu;
3125 int ret = 0;
3126
3127 update_rq_clock(rq);
3128
3129 if (kthread || is_migration_disabled(p)) {
3130 /*
3131 * Kernel threads are allowed on online && !active CPUs,
3132 * however, during cpu-hot-unplug, even these might get pushed
3133 * away if not KTHREAD_IS_PER_CPU.
3134 *
3135 * Specifically, migration_disabled() tasks must not fail the
3136 * cpumask_any_and_distribute() pick below, esp. so on
3137 * SCA_MIGRATE_ENABLE, otherwise we'll not call
3138 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
3139 */
3140 cpu_valid_mask = cpu_online_mask;
3141 }
3142
3143 if (!kthread && !cpumask_subset(ctx->new_mask, cpu_allowed_mask)) {
3144 ret = -EINVAL;
3145 goto out;
3146 }
3147
3148 /*
3149 * Must re-check here, to close a race against __kthread_bind(),
3150 * sched_setaffinity() is not guaranteed to observe the flag.
3151 */
3152 if ((ctx->flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
3153 ret = -EINVAL;
3154 goto out;
3155 }
3156
3157 if (!(ctx->flags & SCA_MIGRATE_ENABLE)) {
3158 if (cpumask_equal(&p->cpus_mask, ctx->new_mask)) {
3159 if (ctx->flags & SCA_USER)
3160 swap(p->user_cpus_ptr, ctx->user_mask);
3161 goto out;
3162 }
3163
3164 if (WARN_ON_ONCE(p == current &&
3165 is_migration_disabled(p) &&
3166 !cpumask_test_cpu(task_cpu(p), ctx->new_mask))) {
3167 ret = -EBUSY;
3168 goto out;
3169 }
3170 }
3171
3172 /*
3173 * Picking a ~random cpu helps in cases where we are changing affinity
3174 * for groups of tasks (ie. cpuset), so that load balancing is not
3175 * immediately required to distribute the tasks within their new mask.
3176 */
3177 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, ctx->new_mask);
3178 if (dest_cpu >= nr_cpu_ids) {
3179 ret = -EINVAL;
3180 goto out;
3181 }
3182
3183 __do_set_cpus_allowed(p, ctx);
3184
3185 return affine_move_task(rq, p, rf, dest_cpu, ctx->flags);
3186
3187out:
3188 task_rq_unlock(rq, p, rf);
3189
3190 return ret;
3191}
3192
3193/*
3194 * Change a given task's CPU affinity. Migrate the thread to a
3195 * proper CPU and schedule it away if the CPU it's executing on
3196 * is removed from the allowed bitmask.
3197 *
3198 * NOTE: the caller must have a valid reference to the task, the
3199 * task must not exit() & deallocate itself prematurely. The
3200 * call is not atomic; no spinlocks may be held.
3201 */
3202static int __set_cpus_allowed_ptr(struct task_struct *p,
3203 struct affinity_context *ctx)
3204{
3205 struct rq_flags rf;
3206 struct rq *rq;
3207
3208 rq = task_rq_lock(p, &rf);
3209 /*
3210 * Masking should be skipped if SCA_USER or any of the SCA_MIGRATE_*
3211 * flags are set.
3212 */
3213 if (p->user_cpus_ptr &&
3214 !(ctx->flags & (SCA_USER | SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) &&
3215 cpumask_and(rq->scratch_mask, ctx->new_mask, p->user_cpus_ptr))
3216 ctx->new_mask = rq->scratch_mask;
3217
3218 return __set_cpus_allowed_ptr_locked(p, ctx, rq, &rf);
3219}
3220
3221int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
3222{
3223 struct affinity_context ac = {
3224 .new_mask = new_mask,
3225 .flags = 0,
3226 };
3227
3228 return __set_cpus_allowed_ptr(p, &ac);
3229}
3230EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
3231
3232/*
3233 * Change a given task's CPU affinity to the intersection of its current
3234 * affinity mask and @subset_mask, writing the resulting mask to @new_mask.
3235 * If user_cpus_ptr is defined, use it as the basis for restricting CPU
3236 * affinity or use cpu_online_mask instead.
3237 *
3238 * If the resulting mask is empty, leave the affinity unchanged and return
3239 * -EINVAL.
3240 */
3241static int restrict_cpus_allowed_ptr(struct task_struct *p,
3242 struct cpumask *new_mask,
3243 const struct cpumask *subset_mask)
3244{
3245 struct affinity_context ac = {
3246 .new_mask = new_mask,
3247 .flags = 0,
3248 };
3249 struct rq_flags rf;
3250 struct rq *rq;
3251 int err;
3252
3253 rq = task_rq_lock(p, &rf);
3254
3255 /*
3256 * Forcefully restricting the affinity of a deadline task is
3257 * likely to cause problems, so fail and noisily override the
3258 * mask entirely.
3259 */
3260 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
3261 err = -EPERM;
3262 goto err_unlock;
3263 }
3264
3265 if (!cpumask_and(new_mask, task_user_cpus(p), subset_mask)) {
3266 err = -EINVAL;
3267 goto err_unlock;
3268 }
3269
3270 return __set_cpus_allowed_ptr_locked(p, &ac, rq, &rf);
3271
3272err_unlock:
3273 task_rq_unlock(rq, p, &rf);
3274 return err;
3275}
3276
3277/*
3278 * Restrict the CPU affinity of task @p so that it is a subset of
3279 * task_cpu_possible_mask() and point @p->user_cpus_ptr to a copy of the
3280 * old affinity mask. If the resulting mask is empty, we warn and walk
3281 * up the cpuset hierarchy until we find a suitable mask.
3282 */
3283void force_compatible_cpus_allowed_ptr(struct task_struct *p)
3284{
3285 cpumask_var_t new_mask;
3286 const struct cpumask *override_mask = task_cpu_possible_mask(p);
3287
3288 alloc_cpumask_var(&new_mask, GFP_KERNEL);
3289
3290 /*
3291 * __migrate_task() can fail silently in the face of concurrent
3292 * offlining of the chosen destination CPU, so take the hotplug
3293 * lock to ensure that the migration succeeds.
3294 */
3295 cpus_read_lock();
3296 if (!cpumask_available(new_mask))
3297 goto out_set_mask;
3298
3299 if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
3300 goto out_free_mask;
3301
3302 /*
3303 * We failed to find a valid subset of the affinity mask for the
3304 * task, so override it based on its cpuset hierarchy.
3305 */
3306 cpuset_cpus_allowed(p, new_mask);
3307 override_mask = new_mask;
3308
3309out_set_mask:
3310 if (printk_ratelimit()) {
3311 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3312 task_pid_nr(p), p->comm,
3313 cpumask_pr_args(override_mask));
3314 }
3315
3316 WARN_ON(set_cpus_allowed_ptr(p, override_mask));
3317out_free_mask:
3318 cpus_read_unlock();
3319 free_cpumask_var(new_mask);
3320}
3321
3322static int
3323__sched_setaffinity(struct task_struct *p, struct affinity_context *ctx);
3324
3325/*
3326 * Restore the affinity of a task @p which was previously restricted by a
3327 * call to force_compatible_cpus_allowed_ptr().
3328 *
3329 * It is the caller's responsibility to serialise this with any calls to
3330 * force_compatible_cpus_allowed_ptr(@p).
3331 */
3332void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3333{
3334 struct affinity_context ac = {
3335 .new_mask = task_user_cpus(p),
3336 .flags = 0,
3337 };
3338 int ret;
3339
3340 /*
3341 * Try to restore the old affinity mask with __sched_setaffinity().
3342 * Cpuset masking will be done there too.
3343 */
3344 ret = __sched_setaffinity(p, &ac);
3345 WARN_ON_ONCE(ret);
3346}
3347
3348void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3349{
3350#ifdef CONFIG_SCHED_DEBUG
3351 unsigned int state = READ_ONCE(p->__state);
3352
3353 /*
3354 * We should never call set_task_cpu() on a blocked task,
3355 * ttwu() will sort out the placement.
3356 */
3357 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3358
3359 /*
3360 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3361 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3362 * time relying on p->on_rq.
3363 */
3364 WARN_ON_ONCE(state == TASK_RUNNING &&
3365 p->sched_class == &fair_sched_class &&
3366 (p->on_rq && !task_on_rq_migrating(p)));
3367
3368#ifdef CONFIG_LOCKDEP
3369 /*
3370 * The caller should hold either p->pi_lock or rq->lock, when changing
3371 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3372 *
3373 * sched_move_task() holds both and thus holding either pins the cgroup,
3374 * see task_group().
3375 *
3376 * Furthermore, all task_rq users should acquire both locks, see
3377 * task_rq_lock().
3378 */
3379 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3380 lockdep_is_held(__rq_lockp(task_rq(p)))));
3381#endif
3382 /*
3383 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3384 */
3385 WARN_ON_ONCE(!cpu_online(new_cpu));
3386
3387 WARN_ON_ONCE(is_migration_disabled(p));
3388#endif
3389
3390 trace_sched_migrate_task(p, new_cpu);
3391
3392 if (task_cpu(p) != new_cpu) {
3393 if (p->sched_class->migrate_task_rq)
3394 p->sched_class->migrate_task_rq(p, new_cpu);
3395 p->se.nr_migrations++;
3396 rseq_migrate(p);
3397 sched_mm_cid_migrate_from(p);
3398 perf_event_task_migrate(p);
3399 }
3400
3401 __set_task_cpu(p, new_cpu);
3402}
3403
3404#ifdef CONFIG_NUMA_BALANCING
3405static void __migrate_swap_task(struct task_struct *p, int cpu)
3406{
3407 if (task_on_rq_queued(p)) {
3408 struct rq *src_rq, *dst_rq;
3409 struct rq_flags srf, drf;
3410
3411 src_rq = task_rq(p);
3412 dst_rq = cpu_rq(cpu);
3413
3414 rq_pin_lock(src_rq, &srf);
3415 rq_pin_lock(dst_rq, &drf);
3416
3417 deactivate_task(src_rq, p, 0);
3418 set_task_cpu(p, cpu);
3419 activate_task(dst_rq, p, 0);
3420 wakeup_preempt(dst_rq, p, 0);
3421
3422 rq_unpin_lock(dst_rq, &drf);
3423 rq_unpin_lock(src_rq, &srf);
3424
3425 } else {
3426 /*
3427 * Task isn't running anymore; make it appear like we migrated
3428 * it before it went to sleep. This means on wakeup we make the
3429 * previous CPU our target instead of where it really is.
3430 */
3431 p->wake_cpu = cpu;
3432 }
3433}
3434
3435struct migration_swap_arg {
3436 struct task_struct *src_task, *dst_task;
3437 int src_cpu, dst_cpu;
3438};
3439
3440static int migrate_swap_stop(void *data)
3441{
3442 struct migration_swap_arg *arg = data;
3443 struct rq *src_rq, *dst_rq;
3444
3445 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3446 return -EAGAIN;
3447
3448 src_rq = cpu_rq(arg->src_cpu);
3449 dst_rq = cpu_rq(arg->dst_cpu);
3450
3451 guard(double_raw_spinlock)(&arg->src_task->pi_lock, &arg->dst_task->pi_lock);
3452 guard(double_rq_lock)(src_rq, dst_rq);
3453
3454 if (task_cpu(arg->dst_task) != arg->dst_cpu)
3455 return -EAGAIN;
3456
3457 if (task_cpu(arg->src_task) != arg->src_cpu)
3458 return -EAGAIN;
3459
3460 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3461 return -EAGAIN;
3462
3463 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3464 return -EAGAIN;
3465
3466 __migrate_swap_task(arg->src_task, arg->dst_cpu);
3467 __migrate_swap_task(arg->dst_task, arg->src_cpu);
3468
3469 return 0;
3470}
3471
3472/*
3473 * Cross migrate two tasks
3474 */
3475int migrate_swap(struct task_struct *cur, struct task_struct *p,
3476 int target_cpu, int curr_cpu)
3477{
3478 struct migration_swap_arg arg;
3479 int ret = -EINVAL;
3480
3481 arg = (struct migration_swap_arg){
3482 .src_task = cur,
3483 .src_cpu = curr_cpu,
3484 .dst_task = p,
3485 .dst_cpu = target_cpu,
3486 };
3487
3488 if (arg.src_cpu == arg.dst_cpu)
3489 goto out;
3490
3491 /*
3492 * These three tests are all lockless; this is OK since all of them
3493 * will be re-checked with proper locks held further down the line.
3494 */
3495 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3496 goto out;
3497
3498 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3499 goto out;
3500
3501 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3502 goto out;
3503
3504 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3505 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3506
3507out:
3508 return ret;
3509}
3510#endif /* CONFIG_NUMA_BALANCING */
3511
3512/***
3513 * kick_process - kick a running thread to enter/exit the kernel
3514 * @p: the to-be-kicked thread
3515 *
3516 * Cause a process which is running on another CPU to enter
3517 * kernel-mode, without any delay. (to get signals handled.)
3518 *
3519 * NOTE: this function doesn't have to take the runqueue lock,
3520 * because all it wants to ensure is that the remote task enters
3521 * the kernel. If the IPI races and the task has been migrated
3522 * to another CPU then no harm is done and the purpose has been
3523 * achieved as well.
3524 */
3525void kick_process(struct task_struct *p)
3526{
3527 guard(preempt)();
3528 int cpu = task_cpu(p);
3529
3530 if ((cpu != smp_processor_id()) && task_curr(p))
3531 smp_send_reschedule(cpu);
3532}
3533EXPORT_SYMBOL_GPL(kick_process);
3534
3535/*
3536 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3537 *
3538 * A few notes on cpu_active vs cpu_online:
3539 *
3540 * - cpu_active must be a subset of cpu_online
3541 *
3542 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3543 * see __set_cpus_allowed_ptr(). At this point the newly online
3544 * CPU isn't yet part of the sched domains, and balancing will not
3545 * see it.
3546 *
3547 * - on CPU-down we clear cpu_active() to mask the sched domains and
3548 * avoid the load balancer to place new tasks on the to be removed
3549 * CPU. Existing tasks will remain running there and will be taken
3550 * off.
3551 *
3552 * This means that fallback selection must not select !active CPUs.
3553 * And can assume that any active CPU must be online. Conversely
3554 * select_task_rq() below may allow selection of !active CPUs in order
3555 * to satisfy the above rules.
3556 */
3557static int select_fallback_rq(int cpu, struct task_struct *p)
3558{
3559 int nid = cpu_to_node(cpu);
3560 const struct cpumask *nodemask = NULL;
3561 enum { cpuset, possible, fail } state = cpuset;
3562 int dest_cpu;
3563
3564 /*
3565 * If the node that the CPU is on has been offlined, cpu_to_node()
3566 * will return -1. There is no CPU on the node, and we should
3567 * select the CPU on the other node.
3568 */
3569 if (nid != -1) {
3570 nodemask = cpumask_of_node(nid);
3571
3572 /* Look for allowed, online CPU in same node. */
3573 for_each_cpu(dest_cpu, nodemask) {
3574 if (is_cpu_allowed(p, dest_cpu))
3575 return dest_cpu;
3576 }
3577 }
3578
3579 for (;;) {
3580 /* Any allowed, online CPU? */
3581 for_each_cpu(dest_cpu, p->cpus_ptr) {
3582 if (!is_cpu_allowed(p, dest_cpu))
3583 continue;
3584
3585 goto out;
3586 }
3587
3588 /* No more Mr. Nice Guy. */
3589 switch (state) {
3590 case cpuset:
3591 if (cpuset_cpus_allowed_fallback(p)) {
3592 state = possible;
3593 break;
3594 }
3595 fallthrough;
3596 case possible:
3597 /*
3598 * XXX When called from select_task_rq() we only
3599 * hold p->pi_lock and again violate locking order.
3600 *
3601 * More yuck to audit.
3602 */
3603 do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3604 state = fail;
3605 break;
3606 case fail:
3607 BUG();
3608 break;
3609 }
3610 }
3611
3612out:
3613 if (state != cpuset) {
3614 /*
3615 * Don't tell them about moving exiting tasks or
3616 * kernel threads (both mm NULL), since they never
3617 * leave kernel.
3618 */
3619 if (p->mm && printk_ratelimit()) {
3620 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3621 task_pid_nr(p), p->comm, cpu);
3622 }
3623 }
3624
3625 return dest_cpu;
3626}
3627
3628/*
3629 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3630 */
3631static inline
3632int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3633{
3634 lockdep_assert_held(&p->pi_lock);
3635
3636 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3637 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3638 else
3639 cpu = cpumask_any(p->cpus_ptr);
3640
3641 /*
3642 * In order not to call set_task_cpu() on a blocking task we need
3643 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3644 * CPU.
3645 *
3646 * Since this is common to all placement strategies, this lives here.
3647 *
3648 * [ this allows ->select_task() to simply return task_cpu(p) and
3649 * not worry about this generic constraint ]
3650 */
3651 if (unlikely(!is_cpu_allowed(p, cpu)))
3652 cpu = select_fallback_rq(task_cpu(p), p);
3653
3654 return cpu;
3655}
3656
3657void sched_set_stop_task(int cpu, struct task_struct *stop)
3658{
3659 static struct lock_class_key stop_pi_lock;
3660 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3661 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3662
3663 if (stop) {
3664 /*
3665 * Make it appear like a SCHED_FIFO task, its something
3666 * userspace knows about and won't get confused about.
3667 *
3668 * Also, it will make PI more or less work without too
3669 * much confusion -- but then, stop work should not
3670 * rely on PI working anyway.
3671 */
3672 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3673
3674 stop->sched_class = &stop_sched_class;
3675
3676 /*
3677 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3678 * adjust the effective priority of a task. As a result,
3679 * rt_mutex_setprio() can trigger (RT) balancing operations,
3680 * which can then trigger wakeups of the stop thread to push
3681 * around the current task.
3682 *
3683 * The stop task itself will never be part of the PI-chain, it
3684 * never blocks, therefore that ->pi_lock recursion is safe.
3685 * Tell lockdep about this by placing the stop->pi_lock in its
3686 * own class.
3687 */
3688 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3689 }
3690
3691 cpu_rq(cpu)->stop = stop;
3692
3693 if (old_stop) {
3694 /*
3695 * Reset it back to a normal scheduling class so that
3696 * it can die in pieces.
3697 */
3698 old_stop->sched_class = &rt_sched_class;
3699 }
3700}
3701
3702#else /* CONFIG_SMP */
3703
3704static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3705 struct affinity_context *ctx)
3706{
3707 return set_cpus_allowed_ptr(p, ctx->new_mask);
3708}
3709
3710static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3711
3712static inline bool rq_has_pinned_tasks(struct rq *rq)
3713{
3714 return false;
3715}
3716
3717static inline cpumask_t *alloc_user_cpus_ptr(int node)
3718{
3719 return NULL;
3720}
3721
3722#endif /* !CONFIG_SMP */
3723
3724static void
3725ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3726{
3727 struct rq *rq;
3728
3729 if (!schedstat_enabled())
3730 return;
3731
3732 rq = this_rq();
3733
3734#ifdef CONFIG_SMP
3735 if (cpu == rq->cpu) {
3736 __schedstat_inc(rq->ttwu_local);
3737 __schedstat_inc(p->stats.nr_wakeups_local);
3738 } else {
3739 struct sched_domain *sd;
3740
3741 __schedstat_inc(p->stats.nr_wakeups_remote);
3742
3743 guard(rcu)();
3744 for_each_domain(rq->cpu, sd) {
3745 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3746 __schedstat_inc(sd->ttwu_wake_remote);
3747 break;
3748 }
3749 }
3750 }
3751
3752 if (wake_flags & WF_MIGRATED)
3753 __schedstat_inc(p->stats.nr_wakeups_migrate);
3754#endif /* CONFIG_SMP */
3755
3756 __schedstat_inc(rq->ttwu_count);
3757 __schedstat_inc(p->stats.nr_wakeups);
3758
3759 if (wake_flags & WF_SYNC)
3760 __schedstat_inc(p->stats.nr_wakeups_sync);
3761}
3762
3763/*
3764 * Mark the task runnable.
3765 */
3766static inline void ttwu_do_wakeup(struct task_struct *p)
3767{
3768 WRITE_ONCE(p->__state, TASK_RUNNING);
3769 trace_sched_wakeup(p);
3770}
3771
3772static void
3773ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3774 struct rq_flags *rf)
3775{
3776 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3777
3778 lockdep_assert_rq_held(rq);
3779
3780 if (p->sched_contributes_to_load)
3781 rq->nr_uninterruptible--;
3782
3783#ifdef CONFIG_SMP
3784 if (wake_flags & WF_MIGRATED)
3785 en_flags |= ENQUEUE_MIGRATED;
3786 else
3787#endif
3788 if (p->in_iowait) {
3789 delayacct_blkio_end(p);
3790 atomic_dec(&task_rq(p)->nr_iowait);
3791 }
3792
3793 activate_task(rq, p, en_flags);
3794 wakeup_preempt(rq, p, wake_flags);
3795
3796 ttwu_do_wakeup(p);
3797
3798#ifdef CONFIG_SMP
3799 if (p->sched_class->task_woken) {
3800 /*
3801 * Our task @p is fully woken up and running; so it's safe to
3802 * drop the rq->lock, hereafter rq is only used for statistics.
3803 */
3804 rq_unpin_lock(rq, rf);
3805 p->sched_class->task_woken(rq, p);
3806 rq_repin_lock(rq, rf);
3807 }
3808
3809 if (rq->idle_stamp) {
3810 u64 delta = rq_clock(rq) - rq->idle_stamp;
3811 u64 max = 2*rq->max_idle_balance_cost;
3812
3813 update_avg(&rq->avg_idle, delta);
3814
3815 if (rq->avg_idle > max)
3816 rq->avg_idle = max;
3817
3818 rq->idle_stamp = 0;
3819 }
3820#endif
3821
3822 p->dl_server = NULL;
3823}
3824
3825/*
3826 * Consider @p being inside a wait loop:
3827 *
3828 * for (;;) {
3829 * set_current_state(TASK_UNINTERRUPTIBLE);
3830 *
3831 * if (CONDITION)
3832 * break;
3833 *
3834 * schedule();
3835 * }
3836 * __set_current_state(TASK_RUNNING);
3837 *
3838 * between set_current_state() and schedule(). In this case @p is still
3839 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3840 * an atomic manner.
3841 *
3842 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3843 * then schedule() must still happen and p->state can be changed to
3844 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3845 * need to do a full wakeup with enqueue.
3846 *
3847 * Returns: %true when the wakeup is done,
3848 * %false otherwise.
3849 */
3850static int ttwu_runnable(struct task_struct *p, int wake_flags)
3851{
3852 struct rq_flags rf;
3853 struct rq *rq;
3854 int ret = 0;
3855
3856 rq = __task_rq_lock(p, &rf);
3857 if (task_on_rq_queued(p)) {
3858 if (!task_on_cpu(rq, p)) {
3859 /*
3860 * When on_rq && !on_cpu the task is preempted, see if
3861 * it should preempt the task that is current now.
3862 */
3863 update_rq_clock(rq);
3864 wakeup_preempt(rq, p, wake_flags);
3865 }
3866 ttwu_do_wakeup(p);
3867 ret = 1;
3868 }
3869 __task_rq_unlock(rq, &rf);
3870
3871 return ret;
3872}
3873
3874#ifdef CONFIG_SMP
3875void sched_ttwu_pending(void *arg)
3876{
3877 struct llist_node *llist = arg;
3878 struct rq *rq = this_rq();
3879 struct task_struct *p, *t;
3880 struct rq_flags rf;
3881
3882 if (!llist)
3883 return;
3884
3885 rq_lock_irqsave(rq, &rf);
3886 update_rq_clock(rq);
3887
3888 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3889 if (WARN_ON_ONCE(p->on_cpu))
3890 smp_cond_load_acquire(&p->on_cpu, !VAL);
3891
3892 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3893 set_task_cpu(p, cpu_of(rq));
3894
3895 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3896 }
3897
3898 /*
3899 * Must be after enqueueing at least once task such that
3900 * idle_cpu() does not observe a false-negative -- if it does,
3901 * it is possible for select_idle_siblings() to stack a number
3902 * of tasks on this CPU during that window.
3903 *
3904 * It is ok to clear ttwu_pending when another task pending.
3905 * We will receive IPI after local irq enabled and then enqueue it.
3906 * Since now nr_running > 0, idle_cpu() will always get correct result.
3907 */
3908 WRITE_ONCE(rq->ttwu_pending, 0);
3909 rq_unlock_irqrestore(rq, &rf);
3910}
3911
3912/*
3913 * Prepare the scene for sending an IPI for a remote smp_call
3914 *
3915 * Returns true if the caller can proceed with sending the IPI.
3916 * Returns false otherwise.
3917 */
3918bool call_function_single_prep_ipi(int cpu)
3919{
3920 if (set_nr_if_polling(cpu_rq(cpu)->idle)) {
3921 trace_sched_wake_idle_without_ipi(cpu);
3922 return false;
3923 }
3924
3925 return true;
3926}
3927
3928/*
3929 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3930 * necessary. The wakee CPU on receipt of the IPI will queue the task
3931 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3932 * of the wakeup instead of the waker.
3933 */
3934static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3935{
3936 struct rq *rq = cpu_rq(cpu);
3937
3938 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3939
3940 WRITE_ONCE(rq->ttwu_pending, 1);
3941 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3942}
3943
3944void wake_up_if_idle(int cpu)
3945{
3946 struct rq *rq = cpu_rq(cpu);
3947
3948 guard(rcu)();
3949 if (is_idle_task(rcu_dereference(rq->curr))) {
3950 guard(rq_lock_irqsave)(rq);
3951 if (is_idle_task(rq->curr))
3952 resched_curr(rq);
3953 }
3954}
3955
3956bool cpus_equal_capacity(int this_cpu, int that_cpu)
3957{
3958 if (!sched_asym_cpucap_active())
3959 return true;
3960
3961 if (this_cpu == that_cpu)
3962 return true;
3963
3964 return arch_scale_cpu_capacity(this_cpu) == arch_scale_cpu_capacity(that_cpu);
3965}
3966
3967bool cpus_share_cache(int this_cpu, int that_cpu)
3968{
3969 if (this_cpu == that_cpu)
3970 return true;
3971
3972 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3973}
3974
3975/*
3976 * Whether CPUs are share cache resources, which means LLC on non-cluster
3977 * machines and LLC tag or L2 on machines with clusters.
3978 */
3979bool cpus_share_resources(int this_cpu, int that_cpu)
3980{
3981 if (this_cpu == that_cpu)
3982 return true;
3983
3984 return per_cpu(sd_share_id, this_cpu) == per_cpu(sd_share_id, that_cpu);
3985}
3986
3987static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
3988{
3989 /*
3990 * Do not complicate things with the async wake_list while the CPU is
3991 * in hotplug state.
3992 */
3993 if (!cpu_active(cpu))
3994 return false;
3995
3996 /* Ensure the task will still be allowed to run on the CPU. */
3997 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
3998 return false;
3999
4000 /*
4001 * If the CPU does not share cache, then queue the task on the
4002 * remote rqs wakelist to avoid accessing remote data.
4003 */
4004 if (!cpus_share_cache(smp_processor_id(), cpu))
4005 return true;
4006
4007 if (cpu == smp_processor_id())
4008 return false;
4009
4010 /*
4011 * If the wakee cpu is idle, or the task is descheduling and the
4012 * only running task on the CPU, then use the wakelist to offload
4013 * the task activation to the idle (or soon-to-be-idle) CPU as
4014 * the current CPU is likely busy. nr_running is checked to
4015 * avoid unnecessary task stacking.
4016 *
4017 * Note that we can only get here with (wakee) p->on_rq=0,
4018 * p->on_cpu can be whatever, we've done the dequeue, so
4019 * the wakee has been accounted out of ->nr_running.
4020 */
4021 if (!cpu_rq(cpu)->nr_running)
4022 return true;
4023
4024 return false;
4025}
4026
4027static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
4028{
4029 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
4030 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
4031 __ttwu_queue_wakelist(p, cpu, wake_flags);
4032 return true;
4033 }
4034
4035 return false;
4036}
4037
4038#else /* !CONFIG_SMP */
4039
4040static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
4041{
4042 return false;
4043}
4044
4045#endif /* CONFIG_SMP */
4046
4047static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
4048{
4049 struct rq *rq = cpu_rq(cpu);
4050 struct rq_flags rf;
4051
4052 if (ttwu_queue_wakelist(p, cpu, wake_flags))
4053 return;
4054
4055 rq_lock(rq, &rf);
4056 update_rq_clock(rq);
4057 ttwu_do_activate(rq, p, wake_flags, &rf);
4058 rq_unlock(rq, &rf);
4059}
4060
4061/*
4062 * Invoked from try_to_wake_up() to check whether the task can be woken up.
4063 *
4064 * The caller holds p::pi_lock if p != current or has preemption
4065 * disabled when p == current.
4066 *
4067 * The rules of saved_state:
4068 *
4069 * The related locking code always holds p::pi_lock when updating
4070 * p::saved_state, which means the code is fully serialized in both cases.
4071 *
4072 * For PREEMPT_RT, the lock wait and lock wakeups happen via TASK_RTLOCK_WAIT.
4073 * No other bits set. This allows to distinguish all wakeup scenarios.
4074 *
4075 * For FREEZER, the wakeup happens via TASK_FROZEN. No other bits set. This
4076 * allows us to prevent early wakeup of tasks before they can be run on
4077 * asymmetric ISA architectures (eg ARMv9).
4078 */
4079static __always_inline
4080bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
4081{
4082 int match;
4083
4084 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
4085 WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
4086 state != TASK_RTLOCK_WAIT);
4087 }
4088
4089 *success = !!(match = __task_state_match(p, state));
4090
4091 /*
4092 * Saved state preserves the task state across blocking on
4093 * an RT lock or TASK_FREEZABLE tasks. If the state matches,
4094 * set p::saved_state to TASK_RUNNING, but do not wake the task
4095 * because it waits for a lock wakeup or __thaw_task(). Also
4096 * indicate success because from the regular waker's point of
4097 * view this has succeeded.
4098 *
4099 * After acquiring the lock the task will restore p::__state
4100 * from p::saved_state which ensures that the regular
4101 * wakeup is not lost. The restore will also set
4102 * p::saved_state to TASK_RUNNING so any further tests will
4103 * not result in false positives vs. @success
4104 */
4105 if (match < 0)
4106 p->saved_state = TASK_RUNNING;
4107
4108 return match > 0;
4109}
4110
4111/*
4112 * Notes on Program-Order guarantees on SMP systems.
4113 *
4114 * MIGRATION
4115 *
4116 * The basic program-order guarantee on SMP systems is that when a task [t]
4117 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
4118 * execution on its new CPU [c1].
4119 *
4120 * For migration (of runnable tasks) this is provided by the following means:
4121 *
4122 * A) UNLOCK of the rq(c0)->lock scheduling out task t
4123 * B) migration for t is required to synchronize *both* rq(c0)->lock and
4124 * rq(c1)->lock (if not at the same time, then in that order).
4125 * C) LOCK of the rq(c1)->lock scheduling in task
4126 *
4127 * Release/acquire chaining guarantees that B happens after A and C after B.
4128 * Note: the CPU doing B need not be c0 or c1
4129 *
4130 * Example:
4131 *
4132 * CPU0 CPU1 CPU2
4133 *
4134 * LOCK rq(0)->lock
4135 * sched-out X
4136 * sched-in Y
4137 * UNLOCK rq(0)->lock
4138 *
4139 * LOCK rq(0)->lock // orders against CPU0
4140 * dequeue X
4141 * UNLOCK rq(0)->lock
4142 *
4143 * LOCK rq(1)->lock
4144 * enqueue X
4145 * UNLOCK rq(1)->lock
4146 *
4147 * LOCK rq(1)->lock // orders against CPU2
4148 * sched-out Z
4149 * sched-in X
4150 * UNLOCK rq(1)->lock
4151 *
4152 *
4153 * BLOCKING -- aka. SLEEP + WAKEUP
4154 *
4155 * For blocking we (obviously) need to provide the same guarantee as for
4156 * migration. However the means are completely different as there is no lock
4157 * chain to provide order. Instead we do:
4158 *
4159 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
4160 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
4161 *
4162 * Example:
4163 *
4164 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
4165 *
4166 * LOCK rq(0)->lock LOCK X->pi_lock
4167 * dequeue X
4168 * sched-out X
4169 * smp_store_release(X->on_cpu, 0);
4170 *
4171 * smp_cond_load_acquire(&X->on_cpu, !VAL);
4172 * X->state = WAKING
4173 * set_task_cpu(X,2)
4174 *
4175 * LOCK rq(2)->lock
4176 * enqueue X
4177 * X->state = RUNNING
4178 * UNLOCK rq(2)->lock
4179 *
4180 * LOCK rq(2)->lock // orders against CPU1
4181 * sched-out Z
4182 * sched-in X
4183 * UNLOCK rq(2)->lock
4184 *
4185 * UNLOCK X->pi_lock
4186 * UNLOCK rq(0)->lock
4187 *
4188 *
4189 * However, for wakeups there is a second guarantee we must provide, namely we
4190 * must ensure that CONDITION=1 done by the caller can not be reordered with
4191 * accesses to the task state; see try_to_wake_up() and set_current_state().
4192 */
4193
4194/**
4195 * try_to_wake_up - wake up a thread
4196 * @p: the thread to be awakened
4197 * @state: the mask of task states that can be woken
4198 * @wake_flags: wake modifier flags (WF_*)
4199 *
4200 * Conceptually does:
4201 *
4202 * If (@state & @p->state) @p->state = TASK_RUNNING.
4203 *
4204 * If the task was not queued/runnable, also place it back on a runqueue.
4205 *
4206 * This function is atomic against schedule() which would dequeue the task.
4207 *
4208 * It issues a full memory barrier before accessing @p->state, see the comment
4209 * with set_current_state().
4210 *
4211 * Uses p->pi_lock to serialize against concurrent wake-ups.
4212 *
4213 * Relies on p->pi_lock stabilizing:
4214 * - p->sched_class
4215 * - p->cpus_ptr
4216 * - p->sched_task_group
4217 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4218 *
4219 * Tries really hard to only take one task_rq(p)->lock for performance.
4220 * Takes rq->lock in:
4221 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
4222 * - ttwu_queue() -- new rq, for enqueue of the task;
4223 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4224 *
4225 * As a consequence we race really badly with just about everything. See the
4226 * many memory barriers and their comments for details.
4227 *
4228 * Return: %true if @p->state changes (an actual wakeup was done),
4229 * %false otherwise.
4230 */
4231int try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
4232{
4233 guard(preempt)();
4234 int cpu, success = 0;
4235
4236 if (p == current) {
4237 /*
4238 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4239 * == smp_processor_id()'. Together this means we can special
4240 * case the whole 'p->on_rq && ttwu_runnable()' case below
4241 * without taking any locks.
4242 *
4243 * In particular:
4244 * - we rely on Program-Order guarantees for all the ordering,
4245 * - we're serialized against set_special_state() by virtue of
4246 * it disabling IRQs (this allows not taking ->pi_lock).
4247 */
4248 if (!ttwu_state_match(p, state, &success))
4249 goto out;
4250
4251 trace_sched_waking(p);
4252 ttwu_do_wakeup(p);
4253 goto out;
4254 }
4255
4256 /*
4257 * If we are going to wake up a thread waiting for CONDITION we
4258 * need to ensure that CONDITION=1 done by the caller can not be
4259 * reordered with p->state check below. This pairs with smp_store_mb()
4260 * in set_current_state() that the waiting thread does.
4261 */
4262 scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
4263 smp_mb__after_spinlock();
4264 if (!ttwu_state_match(p, state, &success))
4265 break;
4266
4267 trace_sched_waking(p);
4268
4269 /*
4270 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4271 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4272 * in smp_cond_load_acquire() below.
4273 *
4274 * sched_ttwu_pending() try_to_wake_up()
4275 * STORE p->on_rq = 1 LOAD p->state
4276 * UNLOCK rq->lock
4277 *
4278 * __schedule() (switch to task 'p')
4279 * LOCK rq->lock smp_rmb();
4280 * smp_mb__after_spinlock();
4281 * UNLOCK rq->lock
4282 *
4283 * [task p]
4284 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4285 *
4286 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4287 * __schedule(). See the comment for smp_mb__after_spinlock().
4288 *
4289 * A similar smp_rmb() lives in __task_needs_rq_lock().
4290 */
4291 smp_rmb();
4292 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4293 break;
4294
4295#ifdef CONFIG_SMP
4296 /*
4297 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4298 * possible to, falsely, observe p->on_cpu == 0.
4299 *
4300 * One must be running (->on_cpu == 1) in order to remove oneself
4301 * from the runqueue.
4302 *
4303 * __schedule() (switch to task 'p') try_to_wake_up()
4304 * STORE p->on_cpu = 1 LOAD p->on_rq
4305 * UNLOCK rq->lock
4306 *
4307 * __schedule() (put 'p' to sleep)
4308 * LOCK rq->lock smp_rmb();
4309 * smp_mb__after_spinlock();
4310 * STORE p->on_rq = 0 LOAD p->on_cpu
4311 *
4312 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4313 * __schedule(). See the comment for smp_mb__after_spinlock().
4314 *
4315 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4316 * schedule()'s deactivate_task() has 'happened' and p will no longer
4317 * care about it's own p->state. See the comment in __schedule().
4318 */
4319 smp_acquire__after_ctrl_dep();
4320
4321 /*
4322 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4323 * == 0), which means we need to do an enqueue, change p->state to
4324 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4325 * enqueue, such as ttwu_queue_wakelist().
4326 */
4327 WRITE_ONCE(p->__state, TASK_WAKING);
4328
4329 /*
4330 * If the owning (remote) CPU is still in the middle of schedule() with
4331 * this task as prev, considering queueing p on the remote CPUs wake_list
4332 * which potentially sends an IPI instead of spinning on p->on_cpu to
4333 * let the waker make forward progress. This is safe because IRQs are
4334 * disabled and the IPI will deliver after on_cpu is cleared.
4335 *
4336 * Ensure we load task_cpu(p) after p->on_cpu:
4337 *
4338 * set_task_cpu(p, cpu);
4339 * STORE p->cpu = @cpu
4340 * __schedule() (switch to task 'p')
4341 * LOCK rq->lock
4342 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4343 * STORE p->on_cpu = 1 LOAD p->cpu
4344 *
4345 * to ensure we observe the correct CPU on which the task is currently
4346 * scheduling.
4347 */
4348 if (smp_load_acquire(&p->on_cpu) &&
4349 ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
4350 break;
4351
4352 /*
4353 * If the owning (remote) CPU is still in the middle of schedule() with
4354 * this task as prev, wait until it's done referencing the task.
4355 *
4356 * Pairs with the smp_store_release() in finish_task().
4357 *
4358 * This ensures that tasks getting woken will be fully ordered against
4359 * their previous state and preserve Program Order.
4360 */
4361 smp_cond_load_acquire(&p->on_cpu, !VAL);
4362
4363 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4364 if (task_cpu(p) != cpu) {
4365 if (p->in_iowait) {
4366 delayacct_blkio_end(p);
4367 atomic_dec(&task_rq(p)->nr_iowait);
4368 }
4369
4370 wake_flags |= WF_MIGRATED;
4371 psi_ttwu_dequeue(p);
4372 set_task_cpu(p, cpu);
4373 }
4374#else
4375 cpu = task_cpu(p);
4376#endif /* CONFIG_SMP */
4377
4378 ttwu_queue(p, cpu, wake_flags);
4379 }
4380out:
4381 if (success)
4382 ttwu_stat(p, task_cpu(p), wake_flags);
4383
4384 return success;
4385}
4386
4387static bool __task_needs_rq_lock(struct task_struct *p)
4388{
4389 unsigned int state = READ_ONCE(p->__state);
4390
4391 /*
4392 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4393 * the task is blocked. Make sure to check @state since ttwu() can drop
4394 * locks at the end, see ttwu_queue_wakelist().
4395 */
4396 if (state == TASK_RUNNING || state == TASK_WAKING)
4397 return true;
4398
4399 /*
4400 * Ensure we load p->on_rq after p->__state, otherwise it would be
4401 * possible to, falsely, observe p->on_rq == 0.
4402 *
4403 * See try_to_wake_up() for a longer comment.
4404 */
4405 smp_rmb();
4406 if (p->on_rq)
4407 return true;
4408
4409#ifdef CONFIG_SMP
4410 /*
4411 * Ensure the task has finished __schedule() and will not be referenced
4412 * anymore. Again, see try_to_wake_up() for a longer comment.
4413 */
4414 smp_rmb();
4415 smp_cond_load_acquire(&p->on_cpu, !VAL);
4416#endif
4417
4418 return false;
4419}
4420
4421/**
4422 * task_call_func - Invoke a function on task in fixed state
4423 * @p: Process for which the function is to be invoked, can be @current.
4424 * @func: Function to invoke.
4425 * @arg: Argument to function.
4426 *
4427 * Fix the task in it's current state by avoiding wakeups and or rq operations
4428 * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4429 * to work out what the state is, if required. Given that @func can be invoked
4430 * with a runqueue lock held, it had better be quite lightweight.
4431 *
4432 * Returns:
4433 * Whatever @func returns
4434 */
4435int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4436{
4437 struct rq *rq = NULL;
4438 struct rq_flags rf;
4439 int ret;
4440
4441 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4442
4443 if (__task_needs_rq_lock(p))
4444 rq = __task_rq_lock(p, &rf);
4445
4446 /*
4447 * At this point the task is pinned; either:
4448 * - blocked and we're holding off wakeups (pi->lock)
4449 * - woken, and we're holding off enqueue (rq->lock)
4450 * - queued, and we're holding off schedule (rq->lock)
4451 * - running, and we're holding off de-schedule (rq->lock)
4452 *
4453 * The called function (@func) can use: task_curr(), p->on_rq and
4454 * p->__state to differentiate between these states.
4455 */
4456 ret = func(p, arg);
4457
4458 if (rq)
4459 rq_unlock(rq, &rf);
4460
4461 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4462 return ret;
4463}
4464
4465/**
4466 * cpu_curr_snapshot - Return a snapshot of the currently running task
4467 * @cpu: The CPU on which to snapshot the task.
4468 *
4469 * Returns the task_struct pointer of the task "currently" running on
4470 * the specified CPU. If the same task is running on that CPU throughout,
4471 * the return value will be a pointer to that task's task_struct structure.
4472 * If the CPU did any context switches even vaguely concurrently with the
4473 * execution of this function, the return value will be a pointer to the
4474 * task_struct structure of a randomly chosen task that was running on
4475 * that CPU somewhere around the time that this function was executing.
4476 *
4477 * If the specified CPU was offline, the return value is whatever it
4478 * is, perhaps a pointer to the task_struct structure of that CPU's idle
4479 * task, but there is no guarantee. Callers wishing a useful return
4480 * value must take some action to ensure that the specified CPU remains
4481 * online throughout.
4482 *
4483 * This function executes full memory barriers before and after fetching
4484 * the pointer, which permits the caller to confine this function's fetch
4485 * with respect to the caller's accesses to other shared variables.
4486 */
4487struct task_struct *cpu_curr_snapshot(int cpu)
4488{
4489 struct task_struct *t;
4490
4491 smp_mb(); /* Pairing determined by caller's synchronization design. */
4492 t = rcu_dereference(cpu_curr(cpu));
4493 smp_mb(); /* Pairing determined by caller's synchronization design. */
4494 return t;
4495}
4496
4497/**
4498 * wake_up_process - Wake up a specific process
4499 * @p: The process to be woken up.
4500 *
4501 * Attempt to wake up the nominated process and move it to the set of runnable
4502 * processes.
4503 *
4504 * Return: 1 if the process was woken up, 0 if it was already running.
4505 *
4506 * This function executes a full memory barrier before accessing the task state.
4507 */
4508int wake_up_process(struct task_struct *p)
4509{
4510 return try_to_wake_up(p, TASK_NORMAL, 0);
4511}
4512EXPORT_SYMBOL(wake_up_process);
4513
4514int wake_up_state(struct task_struct *p, unsigned int state)
4515{
4516 return try_to_wake_up(p, state, 0);
4517}
4518
4519/*
4520 * Perform scheduler related setup for a newly forked process p.
4521 * p is forked by current.
4522 *
4523 * __sched_fork() is basic setup used by init_idle() too:
4524 */
4525static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4526{
4527 p->on_rq = 0;
4528
4529 p->se.on_rq = 0;
4530 p->se.exec_start = 0;
4531 p->se.sum_exec_runtime = 0;
4532 p->se.prev_sum_exec_runtime = 0;
4533 p->se.nr_migrations = 0;
4534 p->se.vruntime = 0;
4535 p->se.vlag = 0;
4536 p->se.slice = sysctl_sched_base_slice;
4537 INIT_LIST_HEAD(&p->se.group_node);
4538
4539#ifdef CONFIG_FAIR_GROUP_SCHED
4540 p->se.cfs_rq = NULL;
4541#endif
4542
4543#ifdef CONFIG_SCHEDSTATS
4544 /* Even if schedstat is disabled, there should not be garbage */
4545 memset(&p->stats, 0, sizeof(p->stats));
4546#endif
4547
4548 init_dl_entity(&p->dl);
4549
4550 INIT_LIST_HEAD(&p->rt.run_list);
4551 p->rt.timeout = 0;
4552 p->rt.time_slice = sched_rr_timeslice;
4553 p->rt.on_rq = 0;
4554 p->rt.on_list = 0;
4555
4556#ifdef CONFIG_PREEMPT_NOTIFIERS
4557 INIT_HLIST_HEAD(&p->preempt_notifiers);
4558#endif
4559
4560#ifdef CONFIG_COMPACTION
4561 p->capture_control = NULL;
4562#endif
4563 init_numa_balancing(clone_flags, p);
4564#ifdef CONFIG_SMP
4565 p->wake_entry.u_flags = CSD_TYPE_TTWU;
4566 p->migration_pending = NULL;
4567#endif
4568 init_sched_mm_cid(p);
4569}
4570
4571DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4572
4573#ifdef CONFIG_NUMA_BALANCING
4574
4575int sysctl_numa_balancing_mode;
4576
4577static void __set_numabalancing_state(bool enabled)
4578{
4579 if (enabled)
4580 static_branch_enable(&sched_numa_balancing);
4581 else
4582 static_branch_disable(&sched_numa_balancing);
4583}
4584
4585void set_numabalancing_state(bool enabled)
4586{
4587 if (enabled)
4588 sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
4589 else
4590 sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
4591 __set_numabalancing_state(enabled);
4592}
4593
4594#ifdef CONFIG_PROC_SYSCTL
4595static void reset_memory_tiering(void)
4596{
4597 struct pglist_data *pgdat;
4598
4599 for_each_online_pgdat(pgdat) {
4600 pgdat->nbp_threshold = 0;
4601 pgdat->nbp_th_nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
4602 pgdat->nbp_th_start = jiffies_to_msecs(jiffies);
4603 }
4604}
4605
4606static int sysctl_numa_balancing(struct ctl_table *table, int write,
4607 void *buffer, size_t *lenp, loff_t *ppos)
4608{
4609 struct ctl_table t;
4610 int err;
4611 int state = sysctl_numa_balancing_mode;
4612
4613 if (write && !capable(CAP_SYS_ADMIN))
4614 return -EPERM;
4615
4616 t = *table;
4617 t.data = &state;
4618 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4619 if (err < 0)
4620 return err;
4621 if (write) {
4622 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
4623 (state & NUMA_BALANCING_MEMORY_TIERING))
4624 reset_memory_tiering();
4625 sysctl_numa_balancing_mode = state;
4626 __set_numabalancing_state(state);
4627 }
4628 return err;
4629}
4630#endif
4631#endif
4632
4633#ifdef CONFIG_SCHEDSTATS
4634
4635DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4636
4637static void set_schedstats(bool enabled)
4638{
4639 if (enabled)
4640 static_branch_enable(&sched_schedstats);
4641 else
4642 static_branch_disable(&sched_schedstats);
4643}
4644
4645void force_schedstat_enabled(void)
4646{
4647 if (!schedstat_enabled()) {
4648 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4649 static_branch_enable(&sched_schedstats);
4650 }
4651}
4652
4653static int __init setup_schedstats(char *str)
4654{
4655 int ret = 0;
4656 if (!str)
4657 goto out;
4658
4659 if (!strcmp(str, "enable")) {
4660 set_schedstats(true);
4661 ret = 1;
4662 } else if (!strcmp(str, "disable")) {
4663 set_schedstats(false);
4664 ret = 1;
4665 }
4666out:
4667 if (!ret)
4668 pr_warn("Unable to parse schedstats=\n");
4669
4670 return ret;
4671}
4672__setup("schedstats=", setup_schedstats);
4673
4674#ifdef CONFIG_PROC_SYSCTL
4675static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4676 size_t *lenp, loff_t *ppos)
4677{
4678 struct ctl_table t;
4679 int err;
4680 int state = static_branch_likely(&sched_schedstats);
4681
4682 if (write && !capable(CAP_SYS_ADMIN))
4683 return -EPERM;
4684
4685 t = *table;
4686 t.data = &state;
4687 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4688 if (err < 0)
4689 return err;
4690 if (write)
4691 set_schedstats(state);
4692 return err;
4693}
4694#endif /* CONFIG_PROC_SYSCTL */
4695#endif /* CONFIG_SCHEDSTATS */
4696
4697#ifdef CONFIG_SYSCTL
4698static struct ctl_table sched_core_sysctls[] = {
4699#ifdef CONFIG_SCHEDSTATS
4700 {
4701 .procname = "sched_schedstats",
4702 .data = NULL,
4703 .maxlen = sizeof(unsigned int),
4704 .mode = 0644,
4705 .proc_handler = sysctl_schedstats,
4706 .extra1 = SYSCTL_ZERO,
4707 .extra2 = SYSCTL_ONE,
4708 },
4709#endif /* CONFIG_SCHEDSTATS */
4710#ifdef CONFIG_UCLAMP_TASK
4711 {
4712 .procname = "sched_util_clamp_min",
4713 .data = &sysctl_sched_uclamp_util_min,
4714 .maxlen = sizeof(unsigned int),
4715 .mode = 0644,
4716 .proc_handler = sysctl_sched_uclamp_handler,
4717 },
4718 {
4719 .procname = "sched_util_clamp_max",
4720 .data = &sysctl_sched_uclamp_util_max,
4721 .maxlen = sizeof(unsigned int),
4722 .mode = 0644,
4723 .proc_handler = sysctl_sched_uclamp_handler,
4724 },
4725 {
4726 .procname = "sched_util_clamp_min_rt_default",
4727 .data = &sysctl_sched_uclamp_util_min_rt_default,
4728 .maxlen = sizeof(unsigned int),
4729 .mode = 0644,
4730 .proc_handler = sysctl_sched_uclamp_handler,
4731 },
4732#endif /* CONFIG_UCLAMP_TASK */
4733#ifdef CONFIG_NUMA_BALANCING
4734 {
4735 .procname = "numa_balancing",
4736 .data = NULL, /* filled in by handler */
4737 .maxlen = sizeof(unsigned int),
4738 .mode = 0644,
4739 .proc_handler = sysctl_numa_balancing,
4740 .extra1 = SYSCTL_ZERO,
4741 .extra2 = SYSCTL_FOUR,
4742 },
4743#endif /* CONFIG_NUMA_BALANCING */
4744 {}
4745};
4746static int __init sched_core_sysctl_init(void)
4747{
4748 register_sysctl_init("kernel", sched_core_sysctls);
4749 return 0;
4750}
4751late_initcall(sched_core_sysctl_init);
4752#endif /* CONFIG_SYSCTL */
4753
4754/*
4755 * fork()/clone()-time setup:
4756 */
4757int sched_fork(unsigned long clone_flags, struct task_struct *p)
4758{
4759 __sched_fork(clone_flags, p);
4760 /*
4761 * We mark the process as NEW here. This guarantees that
4762 * nobody will actually run it, and a signal or other external
4763 * event cannot wake it up and insert it on the runqueue either.
4764 */
4765 p->__state = TASK_NEW;
4766
4767 /*
4768 * Make sure we do not leak PI boosting priority to the child.
4769 */
4770 p->prio = current->normal_prio;
4771
4772 uclamp_fork(p);
4773
4774 /*
4775 * Revert to default priority/policy on fork if requested.
4776 */
4777 if (unlikely(p->sched_reset_on_fork)) {
4778 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4779 p->policy = SCHED_NORMAL;
4780 p->static_prio = NICE_TO_PRIO(0);
4781 p->rt_priority = 0;
4782 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4783 p->static_prio = NICE_TO_PRIO(0);
4784
4785 p->prio = p->normal_prio = p->static_prio;
4786 set_load_weight(p, false);
4787
4788 /*
4789 * We don't need the reset flag anymore after the fork. It has
4790 * fulfilled its duty:
4791 */
4792 p->sched_reset_on_fork = 0;
4793 }
4794
4795 if (dl_prio(p->prio))
4796 return -EAGAIN;
4797 else if (rt_prio(p->prio))
4798 p->sched_class = &rt_sched_class;
4799 else
4800 p->sched_class = &fair_sched_class;
4801
4802 init_entity_runnable_average(&p->se);
4803
4804
4805#ifdef CONFIG_SCHED_INFO
4806 if (likely(sched_info_on()))
4807 memset(&p->sched_info, 0, sizeof(p->sched_info));
4808#endif
4809#if defined(CONFIG_SMP)
4810 p->on_cpu = 0;
4811#endif
4812 init_task_preempt_count(p);
4813#ifdef CONFIG_SMP
4814 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4815 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4816#endif
4817 return 0;
4818}
4819
4820void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4821{
4822 unsigned long flags;
4823
4824 /*
4825 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4826 * required yet, but lockdep gets upset if rules are violated.
4827 */
4828 raw_spin_lock_irqsave(&p->pi_lock, flags);
4829#ifdef CONFIG_CGROUP_SCHED
4830 if (1) {
4831 struct task_group *tg;
4832 tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4833 struct task_group, css);
4834 tg = autogroup_task_group(p, tg);
4835 p->sched_task_group = tg;
4836 }
4837#endif
4838 rseq_migrate(p);
4839 /*
4840 * We're setting the CPU for the first time, we don't migrate,
4841 * so use __set_task_cpu().
4842 */
4843 __set_task_cpu(p, smp_processor_id());
4844 if (p->sched_class->task_fork)
4845 p->sched_class->task_fork(p);
4846 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4847}
4848
4849void sched_post_fork(struct task_struct *p)
4850{
4851 uclamp_post_fork(p);
4852}
4853
4854unsigned long to_ratio(u64 period, u64 runtime)
4855{
4856 if (runtime == RUNTIME_INF)
4857 return BW_UNIT;
4858
4859 /*
4860 * Doing this here saves a lot of checks in all
4861 * the calling paths, and returning zero seems
4862 * safe for them anyway.
4863 */
4864 if (period == 0)
4865 return 0;
4866
4867 return div64_u64(runtime << BW_SHIFT, period);
4868}
4869
4870/*
4871 * wake_up_new_task - wake up a newly created task for the first time.
4872 *
4873 * This function will do some initial scheduler statistics housekeeping
4874 * that must be done for every newly created context, then puts the task
4875 * on the runqueue and wakes it.
4876 */
4877void wake_up_new_task(struct task_struct *p)
4878{
4879 struct rq_flags rf;
4880 struct rq *rq;
4881
4882 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4883 WRITE_ONCE(p->__state, TASK_RUNNING);
4884#ifdef CONFIG_SMP
4885 /*
4886 * Fork balancing, do it here and not earlier because:
4887 * - cpus_ptr can change in the fork path
4888 * - any previously selected CPU might disappear through hotplug
4889 *
4890 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4891 * as we're not fully set-up yet.
4892 */
4893 p->recent_used_cpu = task_cpu(p);
4894 rseq_migrate(p);
4895 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4896#endif
4897 rq = __task_rq_lock(p, &rf);
4898 update_rq_clock(rq);
4899 post_init_entity_util_avg(p);
4900
4901 activate_task(rq, p, ENQUEUE_NOCLOCK);
4902 trace_sched_wakeup_new(p);
4903 wakeup_preempt(rq, p, WF_FORK);
4904#ifdef CONFIG_SMP
4905 if (p->sched_class->task_woken) {
4906 /*
4907 * Nothing relies on rq->lock after this, so it's fine to
4908 * drop it.
4909 */
4910 rq_unpin_lock(rq, &rf);
4911 p->sched_class->task_woken(rq, p);
4912 rq_repin_lock(rq, &rf);
4913 }
4914#endif
4915 task_rq_unlock(rq, p, &rf);
4916}
4917
4918#ifdef CONFIG_PREEMPT_NOTIFIERS
4919
4920static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4921
4922void preempt_notifier_inc(void)
4923{
4924 static_branch_inc(&preempt_notifier_key);
4925}
4926EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4927
4928void preempt_notifier_dec(void)
4929{
4930 static_branch_dec(&preempt_notifier_key);
4931}
4932EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4933
4934/**
4935 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4936 * @notifier: notifier struct to register
4937 */
4938void preempt_notifier_register(struct preempt_notifier *notifier)
4939{
4940 if (!static_branch_unlikely(&preempt_notifier_key))
4941 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4942
4943 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4944}
4945EXPORT_SYMBOL_GPL(preempt_notifier_register);
4946
4947/**
4948 * preempt_notifier_unregister - no longer interested in preemption notifications
4949 * @notifier: notifier struct to unregister
4950 *
4951 * This is *not* safe to call from within a preemption notifier.
4952 */
4953void preempt_notifier_unregister(struct preempt_notifier *notifier)
4954{
4955 hlist_del(¬ifier->link);
4956}
4957EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4958
4959static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4960{
4961 struct preempt_notifier *notifier;
4962
4963 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4964 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4965}
4966
4967static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4968{
4969 if (static_branch_unlikely(&preempt_notifier_key))
4970 __fire_sched_in_preempt_notifiers(curr);
4971}
4972
4973static void
4974__fire_sched_out_preempt_notifiers(struct task_struct *curr,
4975 struct task_struct *next)
4976{
4977 struct preempt_notifier *notifier;
4978
4979 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4980 notifier->ops->sched_out(notifier, next);
4981}
4982
4983static __always_inline void
4984fire_sched_out_preempt_notifiers(struct task_struct *curr,
4985 struct task_struct *next)
4986{
4987 if (static_branch_unlikely(&preempt_notifier_key))
4988 __fire_sched_out_preempt_notifiers(curr, next);
4989}
4990
4991#else /* !CONFIG_PREEMPT_NOTIFIERS */
4992
4993static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4994{
4995}
4996
4997static inline void
4998fire_sched_out_preempt_notifiers(struct task_struct *curr,
4999 struct task_struct *next)
5000{
5001}
5002
5003#endif /* CONFIG_PREEMPT_NOTIFIERS */
5004
5005static inline void prepare_task(struct task_struct *next)
5006{
5007#ifdef CONFIG_SMP
5008 /*
5009 * Claim the task as running, we do this before switching to it
5010 * such that any running task will have this set.
5011 *
5012 * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
5013 * its ordering comment.
5014 */
5015 WRITE_ONCE(next->on_cpu, 1);
5016#endif
5017}
5018
5019static inline void finish_task(struct task_struct *prev)
5020{
5021#ifdef CONFIG_SMP
5022 /*
5023 * This must be the very last reference to @prev from this CPU. After
5024 * p->on_cpu is cleared, the task can be moved to a different CPU. We
5025 * must ensure this doesn't happen until the switch is completely
5026 * finished.
5027 *
5028 * In particular, the load of prev->state in finish_task_switch() must
5029 * happen before this.
5030 *
5031 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
5032 */
5033 smp_store_release(&prev->on_cpu, 0);
5034#endif
5035}
5036
5037#ifdef CONFIG_SMP
5038
5039static void do_balance_callbacks(struct rq *rq, struct balance_callback *head)
5040{
5041 void (*func)(struct rq *rq);
5042 struct balance_callback *next;
5043
5044 lockdep_assert_rq_held(rq);
5045
5046 while (head) {
5047 func = (void (*)(struct rq *))head->func;
5048 next = head->next;
5049 head->next = NULL;
5050 head = next;
5051
5052 func(rq);
5053 }
5054}
5055
5056static void balance_push(struct rq *rq);
5057
5058/*
5059 * balance_push_callback is a right abuse of the callback interface and plays
5060 * by significantly different rules.
5061 *
5062 * Where the normal balance_callback's purpose is to be ran in the same context
5063 * that queued it (only later, when it's safe to drop rq->lock again),
5064 * balance_push_callback is specifically targeted at __schedule().
5065 *
5066 * This abuse is tolerated because it places all the unlikely/odd cases behind
5067 * a single test, namely: rq->balance_callback == NULL.
5068 */
5069struct balance_callback balance_push_callback = {
5070 .next = NULL,
5071 .func = balance_push,
5072};
5073
5074static inline struct balance_callback *
5075__splice_balance_callbacks(struct rq *rq, bool split)
5076{
5077 struct balance_callback *head = rq->balance_callback;
5078
5079 if (likely(!head))
5080 return NULL;
5081
5082 lockdep_assert_rq_held(rq);
5083 /*
5084 * Must not take balance_push_callback off the list when
5085 * splice_balance_callbacks() and balance_callbacks() are not
5086 * in the same rq->lock section.
5087 *
5088 * In that case it would be possible for __schedule() to interleave
5089 * and observe the list empty.
5090 */
5091 if (split && head == &balance_push_callback)
5092 head = NULL;
5093 else
5094 rq->balance_callback = NULL;
5095
5096 return head;
5097}
5098
5099static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
5100{
5101 return __splice_balance_callbacks(rq, true);
5102}
5103
5104static void __balance_callbacks(struct rq *rq)
5105{
5106 do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
5107}
5108
5109static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5110{
5111 unsigned long flags;
5112
5113 if (unlikely(head)) {
5114 raw_spin_rq_lock_irqsave(rq, flags);
5115 do_balance_callbacks(rq, head);
5116 raw_spin_rq_unlock_irqrestore(rq, flags);
5117 }
5118}
5119
5120#else
5121
5122static inline void __balance_callbacks(struct rq *rq)
5123{
5124}
5125
5126static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
5127{
5128 return NULL;
5129}
5130
5131static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5132{
5133}
5134
5135#endif
5136
5137static inline void
5138prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
5139{
5140 /*
5141 * Since the runqueue lock will be released by the next
5142 * task (which is an invalid locking op but in the case
5143 * of the scheduler it's an obvious special-case), so we
5144 * do an early lockdep release here:
5145 */
5146 rq_unpin_lock(rq, rf);
5147 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
5148#ifdef CONFIG_DEBUG_SPINLOCK
5149 /* this is a valid case when another task releases the spinlock */
5150 rq_lockp(rq)->owner = next;
5151#endif
5152}
5153
5154static inline void finish_lock_switch(struct rq *rq)
5155{
5156 /*
5157 * If we are tracking spinlock dependencies then we have to
5158 * fix up the runqueue lock - which gets 'carried over' from
5159 * prev into current:
5160 */
5161 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
5162 __balance_callbacks(rq);
5163 raw_spin_rq_unlock_irq(rq);
5164}
5165
5166/*
5167 * NOP if the arch has not defined these:
5168 */
5169
5170#ifndef prepare_arch_switch
5171# define prepare_arch_switch(next) do { } while (0)
5172#endif
5173
5174#ifndef finish_arch_post_lock_switch
5175# define finish_arch_post_lock_switch() do { } while (0)
5176#endif
5177
5178static inline void kmap_local_sched_out(void)
5179{
5180#ifdef CONFIG_KMAP_LOCAL
5181 if (unlikely(current->kmap_ctrl.idx))
5182 __kmap_local_sched_out();
5183#endif
5184}
5185
5186static inline void kmap_local_sched_in(void)
5187{
5188#ifdef CONFIG_KMAP_LOCAL
5189 if (unlikely(current->kmap_ctrl.idx))
5190 __kmap_local_sched_in();
5191#endif
5192}
5193
5194/**
5195 * prepare_task_switch - prepare to switch tasks
5196 * @rq: the runqueue preparing to switch
5197 * @prev: the current task that is being switched out
5198 * @next: the task we are going to switch to.
5199 *
5200 * This is called with the rq lock held and interrupts off. It must
5201 * be paired with a subsequent finish_task_switch after the context
5202 * switch.
5203 *
5204 * prepare_task_switch sets up locking and calls architecture specific
5205 * hooks.
5206 */
5207static inline void
5208prepare_task_switch(struct rq *rq, struct task_struct *prev,
5209 struct task_struct *next)
5210{
5211 kcov_prepare_switch(prev);
5212 sched_info_switch(rq, prev, next);
5213 perf_event_task_sched_out(prev, next);
5214 rseq_preempt(prev);
5215 fire_sched_out_preempt_notifiers(prev, next);
5216 kmap_local_sched_out();
5217 prepare_task(next);
5218 prepare_arch_switch(next);
5219}
5220
5221/**
5222 * finish_task_switch - clean up after a task-switch
5223 * @prev: the thread we just switched away from.
5224 *
5225 * finish_task_switch must be called after the context switch, paired
5226 * with a prepare_task_switch call before the context switch.
5227 * finish_task_switch will reconcile locking set up by prepare_task_switch,
5228 * and do any other architecture-specific cleanup actions.
5229 *
5230 * Note that we may have delayed dropping an mm in context_switch(). If
5231 * so, we finish that here outside of the runqueue lock. (Doing it
5232 * with the lock held can cause deadlocks; see schedule() for
5233 * details.)
5234 *
5235 * The context switch have flipped the stack from under us and restored the
5236 * local variables which were saved when this task called schedule() in the
5237 * past. prev == current is still correct but we need to recalculate this_rq
5238 * because prev may have moved to another CPU.
5239 */
5240static struct rq *finish_task_switch(struct task_struct *prev)
5241 __releases(rq->lock)
5242{
5243 struct rq *rq = this_rq();
5244 struct mm_struct *mm = rq->prev_mm;
5245 unsigned int prev_state;
5246
5247 /*
5248 * The previous task will have left us with a preempt_count of 2
5249 * because it left us after:
5250 *
5251 * schedule()
5252 * preempt_disable(); // 1
5253 * __schedule()
5254 * raw_spin_lock_irq(&rq->lock) // 2
5255 *
5256 * Also, see FORK_PREEMPT_COUNT.
5257 */
5258 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
5259 "corrupted preempt_count: %s/%d/0x%x\n",
5260 current->comm, current->pid, preempt_count()))
5261 preempt_count_set(FORK_PREEMPT_COUNT);
5262
5263 rq->prev_mm = NULL;
5264
5265 /*
5266 * A task struct has one reference for the use as "current".
5267 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
5268 * schedule one last time. The schedule call will never return, and
5269 * the scheduled task must drop that reference.
5270 *
5271 * We must observe prev->state before clearing prev->on_cpu (in
5272 * finish_task), otherwise a concurrent wakeup can get prev
5273 * running on another CPU and we could rave with its RUNNING -> DEAD
5274 * transition, resulting in a double drop.
5275 */
5276 prev_state = READ_ONCE(prev->__state);
5277 vtime_task_switch(prev);
5278 perf_event_task_sched_in(prev, current);
5279 finish_task(prev);
5280 tick_nohz_task_switch();
5281 finish_lock_switch(rq);
5282 finish_arch_post_lock_switch();
5283 kcov_finish_switch(current);
5284 /*
5285 * kmap_local_sched_out() is invoked with rq::lock held and
5286 * interrupts disabled. There is no requirement for that, but the
5287 * sched out code does not have an interrupt enabled section.
5288 * Restoring the maps on sched in does not require interrupts being
5289 * disabled either.
5290 */
5291 kmap_local_sched_in();
5292
5293 fire_sched_in_preempt_notifiers(current);
5294 /*
5295 * When switching through a kernel thread, the loop in
5296 * membarrier_{private,global}_expedited() may have observed that
5297 * kernel thread and not issued an IPI. It is therefore possible to
5298 * schedule between user->kernel->user threads without passing though
5299 * switch_mm(). Membarrier requires a barrier after storing to
5300 * rq->curr, before returning to userspace, so provide them here:
5301 *
5302 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5303 * provided by mmdrop_lazy_tlb(),
5304 * - a sync_core for SYNC_CORE.
5305 */
5306 if (mm) {
5307 membarrier_mm_sync_core_before_usermode(mm);
5308 mmdrop_lazy_tlb_sched(mm);
5309 }
5310
5311 if (unlikely(prev_state == TASK_DEAD)) {
5312 if (prev->sched_class->task_dead)
5313 prev->sched_class->task_dead(prev);
5314
5315 /* Task is done with its stack. */
5316 put_task_stack(prev);
5317
5318 put_task_struct_rcu_user(prev);
5319 }
5320
5321 return rq;
5322}
5323
5324/**
5325 * schedule_tail - first thing a freshly forked thread must call.
5326 * @prev: the thread we just switched away from.
5327 */
5328asmlinkage __visible void schedule_tail(struct task_struct *prev)
5329 __releases(rq->lock)
5330{
5331 /*
5332 * New tasks start with FORK_PREEMPT_COUNT, see there and
5333 * finish_task_switch() for details.
5334 *
5335 * finish_task_switch() will drop rq->lock() and lower preempt_count
5336 * and the preempt_enable() will end up enabling preemption (on
5337 * PREEMPT_COUNT kernels).
5338 */
5339
5340 finish_task_switch(prev);
5341 preempt_enable();
5342
5343 if (current->set_child_tid)
5344 put_user(task_pid_vnr(current), current->set_child_tid);
5345
5346 calculate_sigpending();
5347}
5348
5349/*
5350 * context_switch - switch to the new MM and the new thread's register state.
5351 */
5352static __always_inline struct rq *
5353context_switch(struct rq *rq, struct task_struct *prev,
5354 struct task_struct *next, struct rq_flags *rf)
5355{
5356 prepare_task_switch(rq, prev, next);
5357
5358 /*
5359 * For paravirt, this is coupled with an exit in switch_to to
5360 * combine the page table reload and the switch backend into
5361 * one hypercall.
5362 */
5363 arch_start_context_switch(prev);
5364
5365 /*
5366 * kernel -> kernel lazy + transfer active
5367 * user -> kernel lazy + mmgrab_lazy_tlb() active
5368 *
5369 * kernel -> user switch + mmdrop_lazy_tlb() active
5370 * user -> user switch
5371 *
5372 * switch_mm_cid() needs to be updated if the barriers provided
5373 * by context_switch() are modified.
5374 */
5375 if (!next->mm) { // to kernel
5376 enter_lazy_tlb(prev->active_mm, next);
5377
5378 next->active_mm = prev->active_mm;
5379 if (prev->mm) // from user
5380 mmgrab_lazy_tlb(prev->active_mm);
5381 else
5382 prev->active_mm = NULL;
5383 } else { // to user
5384 membarrier_switch_mm(rq, prev->active_mm, next->mm);
5385 /*
5386 * sys_membarrier() requires an smp_mb() between setting
5387 * rq->curr / membarrier_switch_mm() and returning to userspace.
5388 *
5389 * The below provides this either through switch_mm(), or in
5390 * case 'prev->active_mm == next->mm' through
5391 * finish_task_switch()'s mmdrop().
5392 */
5393 switch_mm_irqs_off(prev->active_mm, next->mm, next);
5394 lru_gen_use_mm(next->mm);
5395
5396 if (!prev->mm) { // from kernel
5397 /* will mmdrop_lazy_tlb() in finish_task_switch(). */
5398 rq->prev_mm = prev->active_mm;
5399 prev->active_mm = NULL;
5400 }
5401 }
5402
5403 /* switch_mm_cid() requires the memory barriers above. */
5404 switch_mm_cid(rq, prev, next);
5405
5406 prepare_lock_switch(rq, next, rf);
5407
5408 /* Here we just switch the register state and the stack. */
5409 switch_to(prev, next, prev);
5410 barrier();
5411
5412 return finish_task_switch(prev);
5413}
5414
5415/*
5416 * nr_running and nr_context_switches:
5417 *
5418 * externally visible scheduler statistics: current number of runnable
5419 * threads, total number of context switches performed since bootup.
5420 */
5421unsigned int nr_running(void)
5422{
5423 unsigned int i, sum = 0;
5424
5425 for_each_online_cpu(i)
5426 sum += cpu_rq(i)->nr_running;
5427
5428 return sum;
5429}
5430
5431/*
5432 * Check if only the current task is running on the CPU.
5433 *
5434 * Caution: this function does not check that the caller has disabled
5435 * preemption, thus the result might have a time-of-check-to-time-of-use
5436 * race. The caller is responsible to use it correctly, for example:
5437 *
5438 * - from a non-preemptible section (of course)
5439 *
5440 * - from a thread that is bound to a single CPU
5441 *
5442 * - in a loop with very short iterations (e.g. a polling loop)
5443 */
5444bool single_task_running(void)
5445{
5446 return raw_rq()->nr_running == 1;
5447}
5448EXPORT_SYMBOL(single_task_running);
5449
5450unsigned long long nr_context_switches_cpu(int cpu)
5451{
5452 return cpu_rq(cpu)->nr_switches;
5453}
5454
5455unsigned long long nr_context_switches(void)
5456{
5457 int i;
5458 unsigned long long sum = 0;
5459
5460 for_each_possible_cpu(i)
5461 sum += cpu_rq(i)->nr_switches;
5462
5463 return sum;
5464}
5465
5466/*
5467 * Consumers of these two interfaces, like for example the cpuidle menu
5468 * governor, are using nonsensical data. Preferring shallow idle state selection
5469 * for a CPU that has IO-wait which might not even end up running the task when
5470 * it does become runnable.
5471 */
5472
5473unsigned int nr_iowait_cpu(int cpu)
5474{
5475 return atomic_read(&cpu_rq(cpu)->nr_iowait);
5476}
5477
5478/*
5479 * IO-wait accounting, and how it's mostly bollocks (on SMP).
5480 *
5481 * The idea behind IO-wait account is to account the idle time that we could
5482 * have spend running if it were not for IO. That is, if we were to improve the
5483 * storage performance, we'd have a proportional reduction in IO-wait time.
5484 *
5485 * This all works nicely on UP, where, when a task blocks on IO, we account
5486 * idle time as IO-wait, because if the storage were faster, it could've been
5487 * running and we'd not be idle.
5488 *
5489 * This has been extended to SMP, by doing the same for each CPU. This however
5490 * is broken.
5491 *
5492 * Imagine for instance the case where two tasks block on one CPU, only the one
5493 * CPU will have IO-wait accounted, while the other has regular idle. Even
5494 * though, if the storage were faster, both could've ran at the same time,
5495 * utilising both CPUs.
5496 *
5497 * This means, that when looking globally, the current IO-wait accounting on
5498 * SMP is a lower bound, by reason of under accounting.
5499 *
5500 * Worse, since the numbers are provided per CPU, they are sometimes
5501 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5502 * associated with any one particular CPU, it can wake to another CPU than it
5503 * blocked on. This means the per CPU IO-wait number is meaningless.
5504 *
5505 * Task CPU affinities can make all that even more 'interesting'.
5506 */
5507
5508unsigned int nr_iowait(void)
5509{
5510 unsigned int i, sum = 0;
5511
5512 for_each_possible_cpu(i)
5513 sum += nr_iowait_cpu(i);
5514
5515 return sum;
5516}
5517
5518#ifdef CONFIG_SMP
5519
5520/*
5521 * sched_exec - execve() is a valuable balancing opportunity, because at
5522 * this point the task has the smallest effective memory and cache footprint.
5523 */
5524void sched_exec(void)
5525{
5526 struct task_struct *p = current;
5527 struct migration_arg arg;
5528 int dest_cpu;
5529
5530 scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
5531 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5532 if (dest_cpu == smp_processor_id())
5533 return;
5534
5535 if (unlikely(!cpu_active(dest_cpu)))
5536 return;
5537
5538 arg = (struct migration_arg){ p, dest_cpu };
5539 }
5540 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5541}
5542
5543#endif
5544
5545DEFINE_PER_CPU(struct kernel_stat, kstat);
5546DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5547
5548EXPORT_PER_CPU_SYMBOL(kstat);
5549EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5550
5551/*
5552 * The function fair_sched_class.update_curr accesses the struct curr
5553 * and its field curr->exec_start; when called from task_sched_runtime(),
5554 * we observe a high rate of cache misses in practice.
5555 * Prefetching this data results in improved performance.
5556 */
5557static inline void prefetch_curr_exec_start(struct task_struct *p)
5558{
5559#ifdef CONFIG_FAIR_GROUP_SCHED
5560 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5561#else
5562 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5563#endif
5564 prefetch(curr);
5565 prefetch(&curr->exec_start);
5566}
5567
5568/*
5569 * Return accounted runtime for the task.
5570 * In case the task is currently running, return the runtime plus current's
5571 * pending runtime that have not been accounted yet.
5572 */
5573unsigned long long task_sched_runtime(struct task_struct *p)
5574{
5575 struct rq_flags rf;
5576 struct rq *rq;
5577 u64 ns;
5578
5579#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5580 /*
5581 * 64-bit doesn't need locks to atomically read a 64-bit value.
5582 * So we have a optimization chance when the task's delta_exec is 0.
5583 * Reading ->on_cpu is racy, but this is ok.
5584 *
5585 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5586 * If we race with it entering CPU, unaccounted time is 0. This is
5587 * indistinguishable from the read occurring a few cycles earlier.
5588 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5589 * been accounted, so we're correct here as well.
5590 */
5591 if (!p->on_cpu || !task_on_rq_queued(p))
5592 return p->se.sum_exec_runtime;
5593#endif
5594
5595 rq = task_rq_lock(p, &rf);
5596 /*
5597 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5598 * project cycles that may never be accounted to this
5599 * thread, breaking clock_gettime().
5600 */
5601 if (task_current(rq, p) && task_on_rq_queued(p)) {
5602 prefetch_curr_exec_start(p);
5603 update_rq_clock(rq);
5604 p->sched_class->update_curr(rq);
5605 }
5606 ns = p->se.sum_exec_runtime;
5607 task_rq_unlock(rq, p, &rf);
5608
5609 return ns;
5610}
5611
5612#ifdef CONFIG_SCHED_DEBUG
5613static u64 cpu_resched_latency(struct rq *rq)
5614{
5615 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5616 u64 resched_latency, now = rq_clock(rq);
5617 static bool warned_once;
5618
5619 if (sysctl_resched_latency_warn_once && warned_once)
5620 return 0;
5621
5622 if (!need_resched() || !latency_warn_ms)
5623 return 0;
5624
5625 if (system_state == SYSTEM_BOOTING)
5626 return 0;
5627
5628 if (!rq->last_seen_need_resched_ns) {
5629 rq->last_seen_need_resched_ns = now;
5630 rq->ticks_without_resched = 0;
5631 return 0;
5632 }
5633
5634 rq->ticks_without_resched++;
5635 resched_latency = now - rq->last_seen_need_resched_ns;
5636 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5637 return 0;
5638
5639 warned_once = true;
5640
5641 return resched_latency;
5642}
5643
5644static int __init setup_resched_latency_warn_ms(char *str)
5645{
5646 long val;
5647
5648 if ((kstrtol(str, 0, &val))) {
5649 pr_warn("Unable to set resched_latency_warn_ms\n");
5650 return 1;
5651 }
5652
5653 sysctl_resched_latency_warn_ms = val;
5654 return 1;
5655}
5656__setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5657#else
5658static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5659#endif /* CONFIG_SCHED_DEBUG */
5660
5661/*
5662 * This function gets called by the timer code, with HZ frequency.
5663 * We call it with interrupts disabled.
5664 */
5665void scheduler_tick(void)
5666{
5667 int cpu = smp_processor_id();
5668 struct rq *rq = cpu_rq(cpu);
5669 struct task_struct *curr = rq->curr;
5670 struct rq_flags rf;
5671 unsigned long thermal_pressure;
5672 u64 resched_latency;
5673
5674 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5675 arch_scale_freq_tick();
5676
5677 sched_clock_tick();
5678
5679 rq_lock(rq, &rf);
5680
5681 update_rq_clock(rq);
5682 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5683 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5684 curr->sched_class->task_tick(rq, curr, 0);
5685 if (sched_feat(LATENCY_WARN))
5686 resched_latency = cpu_resched_latency(rq);
5687 calc_global_load_tick(rq);
5688 sched_core_tick(rq);
5689 task_tick_mm_cid(rq, curr);
5690
5691 rq_unlock(rq, &rf);
5692
5693 if (sched_feat(LATENCY_WARN) && resched_latency)
5694 resched_latency_warn(cpu, resched_latency);
5695
5696 perf_event_task_tick();
5697
5698 if (curr->flags & PF_WQ_WORKER)
5699 wq_worker_tick(curr);
5700
5701#ifdef CONFIG_SMP
5702 rq->idle_balance = idle_cpu(cpu);
5703 trigger_load_balance(rq);
5704#endif
5705}
5706
5707#ifdef CONFIG_NO_HZ_FULL
5708
5709struct tick_work {
5710 int cpu;
5711 atomic_t state;
5712 struct delayed_work work;
5713};
5714/* Values for ->state, see diagram below. */
5715#define TICK_SCHED_REMOTE_OFFLINE 0
5716#define TICK_SCHED_REMOTE_OFFLINING 1
5717#define TICK_SCHED_REMOTE_RUNNING 2
5718
5719/*
5720 * State diagram for ->state:
5721 *
5722 *
5723 * TICK_SCHED_REMOTE_OFFLINE
5724 * | ^
5725 * | |
5726 * | | sched_tick_remote()
5727 * | |
5728 * | |
5729 * +--TICK_SCHED_REMOTE_OFFLINING
5730 * | ^
5731 * | |
5732 * sched_tick_start() | | sched_tick_stop()
5733 * | |
5734 * V |
5735 * TICK_SCHED_REMOTE_RUNNING
5736 *
5737 *
5738 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5739 * and sched_tick_start() are happy to leave the state in RUNNING.
5740 */
5741
5742static struct tick_work __percpu *tick_work_cpu;
5743
5744static void sched_tick_remote(struct work_struct *work)
5745{
5746 struct delayed_work *dwork = to_delayed_work(work);
5747 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5748 int cpu = twork->cpu;
5749 struct rq *rq = cpu_rq(cpu);
5750 int os;
5751
5752 /*
5753 * Handle the tick only if it appears the remote CPU is running in full
5754 * dynticks mode. The check is racy by nature, but missing a tick or
5755 * having one too much is no big deal because the scheduler tick updates
5756 * statistics and checks timeslices in a time-independent way, regardless
5757 * of when exactly it is running.
5758 */
5759 if (tick_nohz_tick_stopped_cpu(cpu)) {
5760 guard(rq_lock_irq)(rq);
5761 struct task_struct *curr = rq->curr;
5762
5763 if (cpu_online(cpu)) {
5764 update_rq_clock(rq);
5765
5766 if (!is_idle_task(curr)) {
5767 /*
5768 * Make sure the next tick runs within a
5769 * reasonable amount of time.
5770 */
5771 u64 delta = rq_clock_task(rq) - curr->se.exec_start;
5772 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5773 }
5774 curr->sched_class->task_tick(rq, curr, 0);
5775
5776 calc_load_nohz_remote(rq);
5777 }
5778 }
5779
5780 /*
5781 * Run the remote tick once per second (1Hz). This arbitrary
5782 * frequency is large enough to avoid overload but short enough
5783 * to keep scheduler internal stats reasonably up to date. But
5784 * first update state to reflect hotplug activity if required.
5785 */
5786 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5787 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5788 if (os == TICK_SCHED_REMOTE_RUNNING)
5789 queue_delayed_work(system_unbound_wq, dwork, HZ);
5790}
5791
5792static void sched_tick_start(int cpu)
5793{
5794 int os;
5795 struct tick_work *twork;
5796
5797 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5798 return;
5799
5800 WARN_ON_ONCE(!tick_work_cpu);
5801
5802 twork = per_cpu_ptr(tick_work_cpu, cpu);
5803 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5804 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5805 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5806 twork->cpu = cpu;
5807 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5808 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5809 }
5810}
5811
5812#ifdef CONFIG_HOTPLUG_CPU
5813static void sched_tick_stop(int cpu)
5814{
5815 struct tick_work *twork;
5816 int os;
5817
5818 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5819 return;
5820
5821 WARN_ON_ONCE(!tick_work_cpu);
5822
5823 twork = per_cpu_ptr(tick_work_cpu, cpu);
5824 /* There cannot be competing actions, but don't rely on stop-machine. */
5825 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5826 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5827 /* Don't cancel, as this would mess up the state machine. */
5828}
5829#endif /* CONFIG_HOTPLUG_CPU */
5830
5831int __init sched_tick_offload_init(void)
5832{
5833 tick_work_cpu = alloc_percpu(struct tick_work);
5834 BUG_ON(!tick_work_cpu);
5835 return 0;
5836}
5837
5838#else /* !CONFIG_NO_HZ_FULL */
5839static inline void sched_tick_start(int cpu) { }
5840static inline void sched_tick_stop(int cpu) { }
5841#endif
5842
5843#if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5844 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5845/*
5846 * If the value passed in is equal to the current preempt count
5847 * then we just disabled preemption. Start timing the latency.
5848 */
5849static inline void preempt_latency_start(int val)
5850{
5851 if (preempt_count() == val) {
5852 unsigned long ip = get_lock_parent_ip();
5853#ifdef CONFIG_DEBUG_PREEMPT
5854 current->preempt_disable_ip = ip;
5855#endif
5856 trace_preempt_off(CALLER_ADDR0, ip);
5857 }
5858}
5859
5860void preempt_count_add(int val)
5861{
5862#ifdef CONFIG_DEBUG_PREEMPT
5863 /*
5864 * Underflow?
5865 */
5866 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5867 return;
5868#endif
5869 __preempt_count_add(val);
5870#ifdef CONFIG_DEBUG_PREEMPT
5871 /*
5872 * Spinlock count overflowing soon?
5873 */
5874 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5875 PREEMPT_MASK - 10);
5876#endif
5877 preempt_latency_start(val);
5878}
5879EXPORT_SYMBOL(preempt_count_add);
5880NOKPROBE_SYMBOL(preempt_count_add);
5881
5882/*
5883 * If the value passed in equals to the current preempt count
5884 * then we just enabled preemption. Stop timing the latency.
5885 */
5886static inline void preempt_latency_stop(int val)
5887{
5888 if (preempt_count() == val)
5889 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5890}
5891
5892void preempt_count_sub(int val)
5893{
5894#ifdef CONFIG_DEBUG_PREEMPT
5895 /*
5896 * Underflow?
5897 */
5898 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5899 return;
5900 /*
5901 * Is the spinlock portion underflowing?
5902 */
5903 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5904 !(preempt_count() & PREEMPT_MASK)))
5905 return;
5906#endif
5907
5908 preempt_latency_stop(val);
5909 __preempt_count_sub(val);
5910}
5911EXPORT_SYMBOL(preempt_count_sub);
5912NOKPROBE_SYMBOL(preempt_count_sub);
5913
5914#else
5915static inline void preempt_latency_start(int val) { }
5916static inline void preempt_latency_stop(int val) { }
5917#endif
5918
5919static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5920{
5921#ifdef CONFIG_DEBUG_PREEMPT
5922 return p->preempt_disable_ip;
5923#else
5924 return 0;
5925#endif
5926}
5927
5928/*
5929 * Print scheduling while atomic bug:
5930 */
5931static noinline void __schedule_bug(struct task_struct *prev)
5932{
5933 /* Save this before calling printk(), since that will clobber it */
5934 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5935
5936 if (oops_in_progress)
5937 return;
5938
5939 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5940 prev->comm, prev->pid, preempt_count());
5941
5942 debug_show_held_locks(prev);
5943 print_modules();
5944 if (irqs_disabled())
5945 print_irqtrace_events(prev);
5946 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
5947 pr_err("Preemption disabled at:");
5948 print_ip_sym(KERN_ERR, preempt_disable_ip);
5949 }
5950 check_panic_on_warn("scheduling while atomic");
5951
5952 dump_stack();
5953 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5954}
5955
5956/*
5957 * Various schedule()-time debugging checks and statistics:
5958 */
5959static inline void schedule_debug(struct task_struct *prev, bool preempt)
5960{
5961#ifdef CONFIG_SCHED_STACK_END_CHECK
5962 if (task_stack_end_corrupted(prev))
5963 panic("corrupted stack end detected inside scheduler\n");
5964
5965 if (task_scs_end_corrupted(prev))
5966 panic("corrupted shadow stack detected inside scheduler\n");
5967#endif
5968
5969#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5970 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5971 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5972 prev->comm, prev->pid, prev->non_block_count);
5973 dump_stack();
5974 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5975 }
5976#endif
5977
5978 if (unlikely(in_atomic_preempt_off())) {
5979 __schedule_bug(prev);
5980 preempt_count_set(PREEMPT_DISABLED);
5981 }
5982 rcu_sleep_check();
5983 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5984
5985 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5986
5987 schedstat_inc(this_rq()->sched_count);
5988}
5989
5990static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5991 struct rq_flags *rf)
5992{
5993#ifdef CONFIG_SMP
5994 const struct sched_class *class;
5995 /*
5996 * We must do the balancing pass before put_prev_task(), such
5997 * that when we release the rq->lock the task is in the same
5998 * state as before we took rq->lock.
5999 *
6000 * We can terminate the balance pass as soon as we know there is
6001 * a runnable task of @class priority or higher.
6002 */
6003 for_class_range(class, prev->sched_class, &idle_sched_class) {
6004 if (class->balance(rq, prev, rf))
6005 break;
6006 }
6007#endif
6008
6009 put_prev_task(rq, prev);
6010}
6011
6012/*
6013 * Pick up the highest-prio task:
6014 */
6015static inline struct task_struct *
6016__pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6017{
6018 const struct sched_class *class;
6019 struct task_struct *p;
6020
6021 /*
6022 * Optimization: we know that if all tasks are in the fair class we can
6023 * call that function directly, but only if the @prev task wasn't of a
6024 * higher scheduling class, because otherwise those lose the
6025 * opportunity to pull in more work from other CPUs.
6026 */
6027 if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
6028 rq->nr_running == rq->cfs.h_nr_running)) {
6029
6030 p = pick_next_task_fair(rq, prev, rf);
6031 if (unlikely(p == RETRY_TASK))
6032 goto restart;
6033
6034 /* Assume the next prioritized class is idle_sched_class */
6035 if (!p) {
6036 put_prev_task(rq, prev);
6037 p = pick_next_task_idle(rq);
6038 }
6039
6040 /*
6041 * This is the fast path; it cannot be a DL server pick;
6042 * therefore even if @p == @prev, ->dl_server must be NULL.
6043 */
6044 if (p->dl_server)
6045 p->dl_server = NULL;
6046
6047 return p;
6048 }
6049
6050restart:
6051 put_prev_task_balance(rq, prev, rf);
6052
6053 /*
6054 * We've updated @prev and no longer need the server link, clear it.
6055 * Must be done before ->pick_next_task() because that can (re)set
6056 * ->dl_server.
6057 */
6058 if (prev->dl_server)
6059 prev->dl_server = NULL;
6060
6061 for_each_class(class) {
6062 p = class->pick_next_task(rq);
6063 if (p)
6064 return p;
6065 }
6066
6067 BUG(); /* The idle class should always have a runnable task. */
6068}
6069
6070#ifdef CONFIG_SCHED_CORE
6071static inline bool is_task_rq_idle(struct task_struct *t)
6072{
6073 return (task_rq(t)->idle == t);
6074}
6075
6076static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
6077{
6078 return is_task_rq_idle(a) || (a->core_cookie == cookie);
6079}
6080
6081static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
6082{
6083 if (is_task_rq_idle(a) || is_task_rq_idle(b))
6084 return true;
6085
6086 return a->core_cookie == b->core_cookie;
6087}
6088
6089static inline struct task_struct *pick_task(struct rq *rq)
6090{
6091 const struct sched_class *class;
6092 struct task_struct *p;
6093
6094 for_each_class(class) {
6095 p = class->pick_task(rq);
6096 if (p)
6097 return p;
6098 }
6099
6100 BUG(); /* The idle class should always have a runnable task. */
6101}
6102
6103extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
6104
6105static void queue_core_balance(struct rq *rq);
6106
6107static struct task_struct *
6108pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6109{
6110 struct task_struct *next, *p, *max = NULL;
6111 const struct cpumask *smt_mask;
6112 bool fi_before = false;
6113 bool core_clock_updated = (rq == rq->core);
6114 unsigned long cookie;
6115 int i, cpu, occ = 0;
6116 struct rq *rq_i;
6117 bool need_sync;
6118
6119 if (!sched_core_enabled(rq))
6120 return __pick_next_task(rq, prev, rf);
6121
6122 cpu = cpu_of(rq);
6123
6124 /* Stopper task is switching into idle, no need core-wide selection. */
6125 if (cpu_is_offline(cpu)) {
6126 /*
6127 * Reset core_pick so that we don't enter the fastpath when
6128 * coming online. core_pick would already be migrated to
6129 * another cpu during offline.
6130 */
6131 rq->core_pick = NULL;
6132 return __pick_next_task(rq, prev, rf);
6133 }
6134
6135 /*
6136 * If there were no {en,de}queues since we picked (IOW, the task
6137 * pointers are all still valid), and we haven't scheduled the last
6138 * pick yet, do so now.
6139 *
6140 * rq->core_pick can be NULL if no selection was made for a CPU because
6141 * it was either offline or went offline during a sibling's core-wide
6142 * selection. In this case, do a core-wide selection.
6143 */
6144 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
6145 rq->core->core_pick_seq != rq->core_sched_seq &&
6146 rq->core_pick) {
6147 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
6148
6149 next = rq->core_pick;
6150 if (next != prev) {
6151 put_prev_task(rq, prev);
6152 set_next_task(rq, next);
6153 }
6154
6155 rq->core_pick = NULL;
6156 goto out;
6157 }
6158
6159 put_prev_task_balance(rq, prev, rf);
6160
6161 smt_mask = cpu_smt_mask(cpu);
6162 need_sync = !!rq->core->core_cookie;
6163
6164 /* reset state */
6165 rq->core->core_cookie = 0UL;
6166 if (rq->core->core_forceidle_count) {
6167 if (!core_clock_updated) {
6168 update_rq_clock(rq->core);
6169 core_clock_updated = true;
6170 }
6171 sched_core_account_forceidle(rq);
6172 /* reset after accounting force idle */
6173 rq->core->core_forceidle_start = 0;
6174 rq->core->core_forceidle_count = 0;
6175 rq->core->core_forceidle_occupation = 0;
6176 need_sync = true;
6177 fi_before = true;
6178 }
6179
6180 /*
6181 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
6182 *
6183 * @task_seq guards the task state ({en,de}queues)
6184 * @pick_seq is the @task_seq we did a selection on
6185 * @sched_seq is the @pick_seq we scheduled
6186 *
6187 * However, preemptions can cause multiple picks on the same task set.
6188 * 'Fix' this by also increasing @task_seq for every pick.
6189 */
6190 rq->core->core_task_seq++;
6191
6192 /*
6193 * Optimize for common case where this CPU has no cookies
6194 * and there are no cookied tasks running on siblings.
6195 */
6196 if (!need_sync) {
6197 next = pick_task(rq);
6198 if (!next->core_cookie) {
6199 rq->core_pick = NULL;
6200 /*
6201 * For robustness, update the min_vruntime_fi for
6202 * unconstrained picks as well.
6203 */
6204 WARN_ON_ONCE(fi_before);
6205 task_vruntime_update(rq, next, false);
6206 goto out_set_next;
6207 }
6208 }
6209
6210 /*
6211 * For each thread: do the regular task pick and find the max prio task
6212 * amongst them.
6213 *
6214 * Tie-break prio towards the current CPU
6215 */
6216 for_each_cpu_wrap(i, smt_mask, cpu) {
6217 rq_i = cpu_rq(i);
6218
6219 /*
6220 * Current cpu always has its clock updated on entrance to
6221 * pick_next_task(). If the current cpu is not the core,
6222 * the core may also have been updated above.
6223 */
6224 if (i != cpu && (rq_i != rq->core || !core_clock_updated))
6225 update_rq_clock(rq_i);
6226
6227 p = rq_i->core_pick = pick_task(rq_i);
6228 if (!max || prio_less(max, p, fi_before))
6229 max = p;
6230 }
6231
6232 cookie = rq->core->core_cookie = max->core_cookie;
6233
6234 /*
6235 * For each thread: try and find a runnable task that matches @max or
6236 * force idle.
6237 */
6238 for_each_cpu(i, smt_mask) {
6239 rq_i = cpu_rq(i);
6240 p = rq_i->core_pick;
6241
6242 if (!cookie_equals(p, cookie)) {
6243 p = NULL;
6244 if (cookie)
6245 p = sched_core_find(rq_i, cookie);
6246 if (!p)
6247 p = idle_sched_class.pick_task(rq_i);
6248 }
6249
6250 rq_i->core_pick = p;
6251
6252 if (p == rq_i->idle) {
6253 if (rq_i->nr_running) {
6254 rq->core->core_forceidle_count++;
6255 if (!fi_before)
6256 rq->core->core_forceidle_seq++;
6257 }
6258 } else {
6259 occ++;
6260 }
6261 }
6262
6263 if (schedstat_enabled() && rq->core->core_forceidle_count) {
6264 rq->core->core_forceidle_start = rq_clock(rq->core);
6265 rq->core->core_forceidle_occupation = occ;
6266 }
6267
6268 rq->core->core_pick_seq = rq->core->core_task_seq;
6269 next = rq->core_pick;
6270 rq->core_sched_seq = rq->core->core_pick_seq;
6271
6272 /* Something should have been selected for current CPU */
6273 WARN_ON_ONCE(!next);
6274
6275 /*
6276 * Reschedule siblings
6277 *
6278 * NOTE: L1TF -- at this point we're no longer running the old task and
6279 * sending an IPI (below) ensures the sibling will no longer be running
6280 * their task. This ensures there is no inter-sibling overlap between
6281 * non-matching user state.
6282 */
6283 for_each_cpu(i, smt_mask) {
6284 rq_i = cpu_rq(i);
6285
6286 /*
6287 * An online sibling might have gone offline before a task
6288 * could be picked for it, or it might be offline but later
6289 * happen to come online, but its too late and nothing was
6290 * picked for it. That's Ok - it will pick tasks for itself,
6291 * so ignore it.
6292 */
6293 if (!rq_i->core_pick)
6294 continue;
6295
6296 /*
6297 * Update for new !FI->FI transitions, or if continuing to be in !FI:
6298 * fi_before fi update?
6299 * 0 0 1
6300 * 0 1 1
6301 * 1 0 1
6302 * 1 1 0
6303 */
6304 if (!(fi_before && rq->core->core_forceidle_count))
6305 task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
6306
6307 rq_i->core_pick->core_occupation = occ;
6308
6309 if (i == cpu) {
6310 rq_i->core_pick = NULL;
6311 continue;
6312 }
6313
6314 /* Did we break L1TF mitigation requirements? */
6315 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6316
6317 if (rq_i->curr == rq_i->core_pick) {
6318 rq_i->core_pick = NULL;
6319 continue;
6320 }
6321
6322 resched_curr(rq_i);
6323 }
6324
6325out_set_next:
6326 set_next_task(rq, next);
6327out:
6328 if (rq->core->core_forceidle_count && next == rq->idle)
6329 queue_core_balance(rq);
6330
6331 return next;
6332}
6333
6334static bool try_steal_cookie(int this, int that)
6335{
6336 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6337 struct task_struct *p;
6338 unsigned long cookie;
6339 bool success = false;
6340
6341 guard(irq)();
6342 guard(double_rq_lock)(dst, src);
6343
6344 cookie = dst->core->core_cookie;
6345 if (!cookie)
6346 return false;
6347
6348 if (dst->curr != dst->idle)
6349 return false;
6350
6351 p = sched_core_find(src, cookie);
6352 if (!p)
6353 return false;
6354
6355 do {
6356 if (p == src->core_pick || p == src->curr)
6357 goto next;
6358
6359 if (!is_cpu_allowed(p, this))
6360 goto next;
6361
6362 if (p->core_occupation > dst->idle->core_occupation)
6363 goto next;
6364 /*
6365 * sched_core_find() and sched_core_next() will ensure
6366 * that task @p is not throttled now, we also need to
6367 * check whether the runqueue of the destination CPU is
6368 * being throttled.
6369 */
6370 if (sched_task_is_throttled(p, this))
6371 goto next;
6372
6373 deactivate_task(src, p, 0);
6374 set_task_cpu(p, this);
6375 activate_task(dst, p, 0);
6376
6377 resched_curr(dst);
6378
6379 success = true;
6380 break;
6381
6382next:
6383 p = sched_core_next(p, cookie);
6384 } while (p);
6385
6386 return success;
6387}
6388
6389static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6390{
6391 int i;
6392
6393 for_each_cpu_wrap(i, sched_domain_span(sd), cpu + 1) {
6394 if (i == cpu)
6395 continue;
6396
6397 if (need_resched())
6398 break;
6399
6400 if (try_steal_cookie(cpu, i))
6401 return true;
6402 }
6403
6404 return false;
6405}
6406
6407static void sched_core_balance(struct rq *rq)
6408{
6409 struct sched_domain *sd;
6410 int cpu = cpu_of(rq);
6411
6412 guard(preempt)();
6413 guard(rcu)();
6414
6415 raw_spin_rq_unlock_irq(rq);
6416 for_each_domain(cpu, sd) {
6417 if (need_resched())
6418 break;
6419
6420 if (steal_cookie_task(cpu, sd))
6421 break;
6422 }
6423 raw_spin_rq_lock_irq(rq);
6424}
6425
6426static DEFINE_PER_CPU(struct balance_callback, core_balance_head);
6427
6428static void queue_core_balance(struct rq *rq)
6429{
6430 if (!sched_core_enabled(rq))
6431 return;
6432
6433 if (!rq->core->core_cookie)
6434 return;
6435
6436 if (!rq->nr_running) /* not forced idle */
6437 return;
6438
6439 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6440}
6441
6442DEFINE_LOCK_GUARD_1(core_lock, int,
6443 sched_core_lock(*_T->lock, &_T->flags),
6444 sched_core_unlock(*_T->lock, &_T->flags),
6445 unsigned long flags)
6446
6447static void sched_core_cpu_starting(unsigned int cpu)
6448{
6449 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6450 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6451 int t;
6452
6453 guard(core_lock)(&cpu);
6454
6455 WARN_ON_ONCE(rq->core != rq);
6456
6457 /* if we're the first, we'll be our own leader */
6458 if (cpumask_weight(smt_mask) == 1)
6459 return;
6460
6461 /* find the leader */
6462 for_each_cpu(t, smt_mask) {
6463 if (t == cpu)
6464 continue;
6465 rq = cpu_rq(t);
6466 if (rq->core == rq) {
6467 core_rq = rq;
6468 break;
6469 }
6470 }
6471
6472 if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6473 return;
6474
6475 /* install and validate core_rq */
6476 for_each_cpu(t, smt_mask) {
6477 rq = cpu_rq(t);
6478
6479 if (t == cpu)
6480 rq->core = core_rq;
6481
6482 WARN_ON_ONCE(rq->core != core_rq);
6483 }
6484}
6485
6486static void sched_core_cpu_deactivate(unsigned int cpu)
6487{
6488 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6489 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6490 int t;
6491
6492 guard(core_lock)(&cpu);
6493
6494 /* if we're the last man standing, nothing to do */
6495 if (cpumask_weight(smt_mask) == 1) {
6496 WARN_ON_ONCE(rq->core != rq);
6497 return;
6498 }
6499
6500 /* if we're not the leader, nothing to do */
6501 if (rq->core != rq)
6502 return;
6503
6504 /* find a new leader */
6505 for_each_cpu(t, smt_mask) {
6506 if (t == cpu)
6507 continue;
6508 core_rq = cpu_rq(t);
6509 break;
6510 }
6511
6512 if (WARN_ON_ONCE(!core_rq)) /* impossible */
6513 return;
6514
6515 /* copy the shared state to the new leader */
6516 core_rq->core_task_seq = rq->core_task_seq;
6517 core_rq->core_pick_seq = rq->core_pick_seq;
6518 core_rq->core_cookie = rq->core_cookie;
6519 core_rq->core_forceidle_count = rq->core_forceidle_count;
6520 core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6521 core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6522
6523 /*
6524 * Accounting edge for forced idle is handled in pick_next_task().
6525 * Don't need another one here, since the hotplug thread shouldn't
6526 * have a cookie.
6527 */
6528 core_rq->core_forceidle_start = 0;
6529
6530 /* install new leader */
6531 for_each_cpu(t, smt_mask) {
6532 rq = cpu_rq(t);
6533 rq->core = core_rq;
6534 }
6535}
6536
6537static inline void sched_core_cpu_dying(unsigned int cpu)
6538{
6539 struct rq *rq = cpu_rq(cpu);
6540
6541 if (rq->core != rq)
6542 rq->core = rq;
6543}
6544
6545#else /* !CONFIG_SCHED_CORE */
6546
6547static inline void sched_core_cpu_starting(unsigned int cpu) {}
6548static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6549static inline void sched_core_cpu_dying(unsigned int cpu) {}
6550
6551static struct task_struct *
6552pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6553{
6554 return __pick_next_task(rq, prev, rf);
6555}
6556
6557#endif /* CONFIG_SCHED_CORE */
6558
6559/*
6560 * Constants for the sched_mode argument of __schedule().
6561 *
6562 * The mode argument allows RT enabled kernels to differentiate a
6563 * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6564 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6565 * optimize the AND operation out and just check for zero.
6566 */
6567#define SM_NONE 0x0
6568#define SM_PREEMPT 0x1
6569#define SM_RTLOCK_WAIT 0x2
6570
6571#ifndef CONFIG_PREEMPT_RT
6572# define SM_MASK_PREEMPT (~0U)
6573#else
6574# define SM_MASK_PREEMPT SM_PREEMPT
6575#endif
6576
6577/*
6578 * __schedule() is the main scheduler function.
6579 *
6580 * The main means of driving the scheduler and thus entering this function are:
6581 *
6582 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6583 *
6584 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6585 * paths. For example, see arch/x86/entry_64.S.
6586 *
6587 * To drive preemption between tasks, the scheduler sets the flag in timer
6588 * interrupt handler scheduler_tick().
6589 *
6590 * 3. Wakeups don't really cause entry into schedule(). They add a
6591 * task to the run-queue and that's it.
6592 *
6593 * Now, if the new task added to the run-queue preempts the current
6594 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6595 * called on the nearest possible occasion:
6596 *
6597 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6598 *
6599 * - in syscall or exception context, at the next outmost
6600 * preempt_enable(). (this might be as soon as the wake_up()'s
6601 * spin_unlock()!)
6602 *
6603 * - in IRQ context, return from interrupt-handler to
6604 * preemptible context
6605 *
6606 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6607 * then at the next:
6608 *
6609 * - cond_resched() call
6610 * - explicit schedule() call
6611 * - return from syscall or exception to user-space
6612 * - return from interrupt-handler to user-space
6613 *
6614 * WARNING: must be called with preemption disabled!
6615 */
6616static void __sched notrace __schedule(unsigned int sched_mode)
6617{
6618 struct task_struct *prev, *next;
6619 unsigned long *switch_count;
6620 unsigned long prev_state;
6621 struct rq_flags rf;
6622 struct rq *rq;
6623 int cpu;
6624
6625 cpu = smp_processor_id();
6626 rq = cpu_rq(cpu);
6627 prev = rq->curr;
6628
6629 schedule_debug(prev, !!sched_mode);
6630
6631 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6632 hrtick_clear(rq);
6633
6634 local_irq_disable();
6635 rcu_note_context_switch(!!sched_mode);
6636
6637 /*
6638 * Make sure that signal_pending_state()->signal_pending() below
6639 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6640 * done by the caller to avoid the race with signal_wake_up():
6641 *
6642 * __set_current_state(@state) signal_wake_up()
6643 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6644 * wake_up_state(p, state)
6645 * LOCK rq->lock LOCK p->pi_state
6646 * smp_mb__after_spinlock() smp_mb__after_spinlock()
6647 * if (signal_pending_state()) if (p->state & @state)
6648 *
6649 * Also, the membarrier system call requires a full memory barrier
6650 * after coming from user-space, before storing to rq->curr; this
6651 * barrier matches a full barrier in the proximity of the membarrier
6652 * system call exit.
6653 */
6654 rq_lock(rq, &rf);
6655 smp_mb__after_spinlock();
6656
6657 /* Promote REQ to ACT */
6658 rq->clock_update_flags <<= 1;
6659 update_rq_clock(rq);
6660 rq->clock_update_flags = RQCF_UPDATED;
6661
6662 switch_count = &prev->nivcsw;
6663
6664 /*
6665 * We must load prev->state once (task_struct::state is volatile), such
6666 * that we form a control dependency vs deactivate_task() below.
6667 */
6668 prev_state = READ_ONCE(prev->__state);
6669 if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6670 if (signal_pending_state(prev_state, prev)) {
6671 WRITE_ONCE(prev->__state, TASK_RUNNING);
6672 } else {
6673 prev->sched_contributes_to_load =
6674 (prev_state & TASK_UNINTERRUPTIBLE) &&
6675 !(prev_state & TASK_NOLOAD) &&
6676 !(prev_state & TASK_FROZEN);
6677
6678 if (prev->sched_contributes_to_load)
6679 rq->nr_uninterruptible++;
6680
6681 /*
6682 * __schedule() ttwu()
6683 * prev_state = prev->state; if (p->on_rq && ...)
6684 * if (prev_state) goto out;
6685 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6686 * p->state = TASK_WAKING
6687 *
6688 * Where __schedule() and ttwu() have matching control dependencies.
6689 *
6690 * After this, schedule() must not care about p->state any more.
6691 */
6692 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6693
6694 if (prev->in_iowait) {
6695 atomic_inc(&rq->nr_iowait);
6696 delayacct_blkio_start();
6697 }
6698 }
6699 switch_count = &prev->nvcsw;
6700 }
6701
6702 next = pick_next_task(rq, prev, &rf);
6703 clear_tsk_need_resched(prev);
6704 clear_preempt_need_resched();
6705#ifdef CONFIG_SCHED_DEBUG
6706 rq->last_seen_need_resched_ns = 0;
6707#endif
6708
6709 if (likely(prev != next)) {
6710 rq->nr_switches++;
6711 /*
6712 * RCU users of rcu_dereference(rq->curr) may not see
6713 * changes to task_struct made by pick_next_task().
6714 */
6715 RCU_INIT_POINTER(rq->curr, next);
6716 /*
6717 * The membarrier system call requires each architecture
6718 * to have a full memory barrier after updating
6719 * rq->curr, before returning to user-space.
6720 *
6721 * Here are the schemes providing that barrier on the
6722 * various architectures:
6723 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC,
6724 * RISC-V. switch_mm() relies on membarrier_arch_switch_mm()
6725 * on PowerPC and on RISC-V.
6726 * - finish_lock_switch() for weakly-ordered
6727 * architectures where spin_unlock is a full barrier,
6728 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6729 * is a RELEASE barrier),
6730 *
6731 * The barrier matches a full barrier in the proximity of
6732 * the membarrier system call entry.
6733 *
6734 * On RISC-V, this barrier pairing is also needed for the
6735 * SYNC_CORE command when switching between processes, cf.
6736 * the inline comments in membarrier_arch_switch_mm().
6737 */
6738 ++*switch_count;
6739
6740 migrate_disable_switch(rq, prev);
6741 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6742
6743 trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
6744
6745 /* Also unlocks the rq: */
6746 rq = context_switch(rq, prev, next, &rf);
6747 } else {
6748 rq_unpin_lock(rq, &rf);
6749 __balance_callbacks(rq);
6750 raw_spin_rq_unlock_irq(rq);
6751 }
6752}
6753
6754void __noreturn do_task_dead(void)
6755{
6756 /* Causes final put_task_struct in finish_task_switch(): */
6757 set_special_state(TASK_DEAD);
6758
6759 /* Tell freezer to ignore us: */
6760 current->flags |= PF_NOFREEZE;
6761
6762 __schedule(SM_NONE);
6763 BUG();
6764
6765 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6766 for (;;)
6767 cpu_relax();
6768}
6769
6770static inline void sched_submit_work(struct task_struct *tsk)
6771{
6772 static DEFINE_WAIT_OVERRIDE_MAP(sched_map, LD_WAIT_CONFIG);
6773 unsigned int task_flags;
6774
6775 /*
6776 * Establish LD_WAIT_CONFIG context to ensure none of the code called
6777 * will use a blocking primitive -- which would lead to recursion.
6778 */
6779 lock_map_acquire_try(&sched_map);
6780
6781 task_flags = tsk->flags;
6782 /*
6783 * If a worker goes to sleep, notify and ask workqueue whether it
6784 * wants to wake up a task to maintain concurrency.
6785 */
6786 if (task_flags & PF_WQ_WORKER)
6787 wq_worker_sleeping(tsk);
6788 else if (task_flags & PF_IO_WORKER)
6789 io_wq_worker_sleeping(tsk);
6790
6791 /*
6792 * spinlock and rwlock must not flush block requests. This will
6793 * deadlock if the callback attempts to acquire a lock which is
6794 * already acquired.
6795 */
6796 SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
6797
6798 /*
6799 * If we are going to sleep and we have plugged IO queued,
6800 * make sure to submit it to avoid deadlocks.
6801 */
6802 blk_flush_plug(tsk->plug, true);
6803
6804 lock_map_release(&sched_map);
6805}
6806
6807static void sched_update_worker(struct task_struct *tsk)
6808{
6809 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER | PF_BLOCK_TS)) {
6810 if (tsk->flags & PF_BLOCK_TS)
6811 blk_plug_invalidate_ts(tsk);
6812 if (tsk->flags & PF_WQ_WORKER)
6813 wq_worker_running(tsk);
6814 else if (tsk->flags & PF_IO_WORKER)
6815 io_wq_worker_running(tsk);
6816 }
6817}
6818
6819static __always_inline void __schedule_loop(unsigned int sched_mode)
6820{
6821 do {
6822 preempt_disable();
6823 __schedule(sched_mode);
6824 sched_preempt_enable_no_resched();
6825 } while (need_resched());
6826}
6827
6828asmlinkage __visible void __sched schedule(void)
6829{
6830 struct task_struct *tsk = current;
6831
6832#ifdef CONFIG_RT_MUTEXES
6833 lockdep_assert(!tsk->sched_rt_mutex);
6834#endif
6835
6836 if (!task_is_running(tsk))
6837 sched_submit_work(tsk);
6838 __schedule_loop(SM_NONE);
6839 sched_update_worker(tsk);
6840}
6841EXPORT_SYMBOL(schedule);
6842
6843/*
6844 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6845 * state (have scheduled out non-voluntarily) by making sure that all
6846 * tasks have either left the run queue or have gone into user space.
6847 * As idle tasks do not do either, they must not ever be preempted
6848 * (schedule out non-voluntarily).
6849 *
6850 * schedule_idle() is similar to schedule_preempt_disable() except that it
6851 * never enables preemption because it does not call sched_submit_work().
6852 */
6853void __sched schedule_idle(void)
6854{
6855 /*
6856 * As this skips calling sched_submit_work(), which the idle task does
6857 * regardless because that function is a nop when the task is in a
6858 * TASK_RUNNING state, make sure this isn't used someplace that the
6859 * current task can be in any other state. Note, idle is always in the
6860 * TASK_RUNNING state.
6861 */
6862 WARN_ON_ONCE(current->__state);
6863 do {
6864 __schedule(SM_NONE);
6865 } while (need_resched());
6866}
6867
6868#if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
6869asmlinkage __visible void __sched schedule_user(void)
6870{
6871 /*
6872 * If we come here after a random call to set_need_resched(),
6873 * or we have been woken up remotely but the IPI has not yet arrived,
6874 * we haven't yet exited the RCU idle mode. Do it here manually until
6875 * we find a better solution.
6876 *
6877 * NB: There are buggy callers of this function. Ideally we
6878 * should warn if prev_state != CONTEXT_USER, but that will trigger
6879 * too frequently to make sense yet.
6880 */
6881 enum ctx_state prev_state = exception_enter();
6882 schedule();
6883 exception_exit(prev_state);
6884}
6885#endif
6886
6887/**
6888 * schedule_preempt_disabled - called with preemption disabled
6889 *
6890 * Returns with preemption disabled. Note: preempt_count must be 1
6891 */
6892void __sched schedule_preempt_disabled(void)
6893{
6894 sched_preempt_enable_no_resched();
6895 schedule();
6896 preempt_disable();
6897}
6898
6899#ifdef CONFIG_PREEMPT_RT
6900void __sched notrace schedule_rtlock(void)
6901{
6902 __schedule_loop(SM_RTLOCK_WAIT);
6903}
6904NOKPROBE_SYMBOL(schedule_rtlock);
6905#endif
6906
6907static void __sched notrace preempt_schedule_common(void)
6908{
6909 do {
6910 /*
6911 * Because the function tracer can trace preempt_count_sub()
6912 * and it also uses preempt_enable/disable_notrace(), if
6913 * NEED_RESCHED is set, the preempt_enable_notrace() called
6914 * by the function tracer will call this function again and
6915 * cause infinite recursion.
6916 *
6917 * Preemption must be disabled here before the function
6918 * tracer can trace. Break up preempt_disable() into two
6919 * calls. One to disable preemption without fear of being
6920 * traced. The other to still record the preemption latency,
6921 * which can also be traced by the function tracer.
6922 */
6923 preempt_disable_notrace();
6924 preempt_latency_start(1);
6925 __schedule(SM_PREEMPT);
6926 preempt_latency_stop(1);
6927 preempt_enable_no_resched_notrace();
6928
6929 /*
6930 * Check again in case we missed a preemption opportunity
6931 * between schedule and now.
6932 */
6933 } while (need_resched());
6934}
6935
6936#ifdef CONFIG_PREEMPTION
6937/*
6938 * This is the entry point to schedule() from in-kernel preemption
6939 * off of preempt_enable.
6940 */
6941asmlinkage __visible void __sched notrace preempt_schedule(void)
6942{
6943 /*
6944 * If there is a non-zero preempt_count or interrupts are disabled,
6945 * we do not want to preempt the current task. Just return..
6946 */
6947 if (likely(!preemptible()))
6948 return;
6949 preempt_schedule_common();
6950}
6951NOKPROBE_SYMBOL(preempt_schedule);
6952EXPORT_SYMBOL(preempt_schedule);
6953
6954#ifdef CONFIG_PREEMPT_DYNAMIC
6955#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6956#ifndef preempt_schedule_dynamic_enabled
6957#define preempt_schedule_dynamic_enabled preempt_schedule
6958#define preempt_schedule_dynamic_disabled NULL
6959#endif
6960DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
6961EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6962#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6963static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
6964void __sched notrace dynamic_preempt_schedule(void)
6965{
6966 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
6967 return;
6968 preempt_schedule();
6969}
6970NOKPROBE_SYMBOL(dynamic_preempt_schedule);
6971EXPORT_SYMBOL(dynamic_preempt_schedule);
6972#endif
6973#endif
6974
6975/**
6976 * preempt_schedule_notrace - preempt_schedule called by tracing
6977 *
6978 * The tracing infrastructure uses preempt_enable_notrace to prevent
6979 * recursion and tracing preempt enabling caused by the tracing
6980 * infrastructure itself. But as tracing can happen in areas coming
6981 * from userspace or just about to enter userspace, a preempt enable
6982 * can occur before user_exit() is called. This will cause the scheduler
6983 * to be called when the system is still in usermode.
6984 *
6985 * To prevent this, the preempt_enable_notrace will use this function
6986 * instead of preempt_schedule() to exit user context if needed before
6987 * calling the scheduler.
6988 */
6989asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6990{
6991 enum ctx_state prev_ctx;
6992
6993 if (likely(!preemptible()))
6994 return;
6995
6996 do {
6997 /*
6998 * Because the function tracer can trace preempt_count_sub()
6999 * and it also uses preempt_enable/disable_notrace(), if
7000 * NEED_RESCHED is set, the preempt_enable_notrace() called
7001 * by the function tracer will call this function again and
7002 * cause infinite recursion.
7003 *
7004 * Preemption must be disabled here before the function
7005 * tracer can trace. Break up preempt_disable() into two
7006 * calls. One to disable preemption without fear of being
7007 * traced. The other to still record the preemption latency,
7008 * which can also be traced by the function tracer.
7009 */
7010 preempt_disable_notrace();
7011 preempt_latency_start(1);
7012 /*
7013 * Needs preempt disabled in case user_exit() is traced
7014 * and the tracer calls preempt_enable_notrace() causing
7015 * an infinite recursion.
7016 */
7017 prev_ctx = exception_enter();
7018 __schedule(SM_PREEMPT);
7019 exception_exit(prev_ctx);
7020
7021 preempt_latency_stop(1);
7022 preempt_enable_no_resched_notrace();
7023 } while (need_resched());
7024}
7025EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
7026
7027#ifdef CONFIG_PREEMPT_DYNAMIC
7028#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
7029#ifndef preempt_schedule_notrace_dynamic_enabled
7030#define preempt_schedule_notrace_dynamic_enabled preempt_schedule_notrace
7031#define preempt_schedule_notrace_dynamic_disabled NULL
7032#endif
7033DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
7034EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
7035#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
7036static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
7037void __sched notrace dynamic_preempt_schedule_notrace(void)
7038{
7039 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
7040 return;
7041 preempt_schedule_notrace();
7042}
7043NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
7044EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
7045#endif
7046#endif
7047
7048#endif /* CONFIG_PREEMPTION */
7049
7050/*
7051 * This is the entry point to schedule() from kernel preemption
7052 * off of irq context.
7053 * Note, that this is called and return with irqs disabled. This will
7054 * protect us against recursive calling from irq.
7055 */
7056asmlinkage __visible void __sched preempt_schedule_irq(void)
7057{
7058 enum ctx_state prev_state;
7059
7060 /* Catch callers which need to be fixed */
7061 BUG_ON(preempt_count() || !irqs_disabled());
7062
7063 prev_state = exception_enter();
7064
7065 do {
7066 preempt_disable();
7067 local_irq_enable();
7068 __schedule(SM_PREEMPT);
7069 local_irq_disable();
7070 sched_preempt_enable_no_resched();
7071 } while (need_resched());
7072
7073 exception_exit(prev_state);
7074}
7075
7076int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
7077 void *key)
7078{
7079 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~(WF_SYNC|WF_CURRENT_CPU));
7080 return try_to_wake_up(curr->private, mode, wake_flags);
7081}
7082EXPORT_SYMBOL(default_wake_function);
7083
7084static void __setscheduler_prio(struct task_struct *p, int prio)
7085{
7086 if (dl_prio(prio))
7087 p->sched_class = &dl_sched_class;
7088 else if (rt_prio(prio))
7089 p->sched_class = &rt_sched_class;
7090 else
7091 p->sched_class = &fair_sched_class;
7092
7093 p->prio = prio;
7094}
7095
7096#ifdef CONFIG_RT_MUTEXES
7097
7098/*
7099 * Would be more useful with typeof()/auto_type but they don't mix with
7100 * bit-fields. Since it's a local thing, use int. Keep the generic sounding
7101 * name such that if someone were to implement this function we get to compare
7102 * notes.
7103 */
7104#define fetch_and_set(x, v) ({ int _x = (x); (x) = (v); _x; })
7105
7106void rt_mutex_pre_schedule(void)
7107{
7108 lockdep_assert(!fetch_and_set(current->sched_rt_mutex, 1));
7109 sched_submit_work(current);
7110}
7111
7112void rt_mutex_schedule(void)
7113{
7114 lockdep_assert(current->sched_rt_mutex);
7115 __schedule_loop(SM_NONE);
7116}
7117
7118void rt_mutex_post_schedule(void)
7119{
7120 sched_update_worker(current);
7121 lockdep_assert(fetch_and_set(current->sched_rt_mutex, 0));
7122}
7123
7124static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
7125{
7126 if (pi_task)
7127 prio = min(prio, pi_task->prio);
7128
7129 return prio;
7130}
7131
7132static inline int rt_effective_prio(struct task_struct *p, int prio)
7133{
7134 struct task_struct *pi_task = rt_mutex_get_top_task(p);
7135
7136 return __rt_effective_prio(pi_task, prio);
7137}
7138
7139/*
7140 * rt_mutex_setprio - set the current priority of a task
7141 * @p: task to boost
7142 * @pi_task: donor task
7143 *
7144 * This function changes the 'effective' priority of a task. It does
7145 * not touch ->normal_prio like __setscheduler().
7146 *
7147 * Used by the rt_mutex code to implement priority inheritance
7148 * logic. Call site only calls if the priority of the task changed.
7149 */
7150void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
7151{
7152 int prio, oldprio, queued, running, queue_flag =
7153 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7154 const struct sched_class *prev_class;
7155 struct rq_flags rf;
7156 struct rq *rq;
7157
7158 /* XXX used to be waiter->prio, not waiter->task->prio */
7159 prio = __rt_effective_prio(pi_task, p->normal_prio);
7160
7161 /*
7162 * If nothing changed; bail early.
7163 */
7164 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
7165 return;
7166
7167 rq = __task_rq_lock(p, &rf);
7168 update_rq_clock(rq);
7169 /*
7170 * Set under pi_lock && rq->lock, such that the value can be used under
7171 * either lock.
7172 *
7173 * Note that there is loads of tricky to make this pointer cache work
7174 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
7175 * ensure a task is de-boosted (pi_task is set to NULL) before the
7176 * task is allowed to run again (and can exit). This ensures the pointer
7177 * points to a blocked task -- which guarantees the task is present.
7178 */
7179 p->pi_top_task = pi_task;
7180
7181 /*
7182 * For FIFO/RR we only need to set prio, if that matches we're done.
7183 */
7184 if (prio == p->prio && !dl_prio(prio))
7185 goto out_unlock;
7186
7187 /*
7188 * Idle task boosting is a nono in general. There is one
7189 * exception, when PREEMPT_RT and NOHZ is active:
7190 *
7191 * The idle task calls get_next_timer_interrupt() and holds
7192 * the timer wheel base->lock on the CPU and another CPU wants
7193 * to access the timer (probably to cancel it). We can safely
7194 * ignore the boosting request, as the idle CPU runs this code
7195 * with interrupts disabled and will complete the lock
7196 * protected section without being interrupted. So there is no
7197 * real need to boost.
7198 */
7199 if (unlikely(p == rq->idle)) {
7200 WARN_ON(p != rq->curr);
7201 WARN_ON(p->pi_blocked_on);
7202 goto out_unlock;
7203 }
7204
7205 trace_sched_pi_setprio(p, pi_task);
7206 oldprio = p->prio;
7207
7208 if (oldprio == prio)
7209 queue_flag &= ~DEQUEUE_MOVE;
7210
7211 prev_class = p->sched_class;
7212 queued = task_on_rq_queued(p);
7213 running = task_current(rq, p);
7214 if (queued)
7215 dequeue_task(rq, p, queue_flag);
7216 if (running)
7217 put_prev_task(rq, p);
7218
7219 /*
7220 * Boosting condition are:
7221 * 1. -rt task is running and holds mutex A
7222 * --> -dl task blocks on mutex A
7223 *
7224 * 2. -dl task is running and holds mutex A
7225 * --> -dl task blocks on mutex A and could preempt the
7226 * running task
7227 */
7228 if (dl_prio(prio)) {
7229 if (!dl_prio(p->normal_prio) ||
7230 (pi_task && dl_prio(pi_task->prio) &&
7231 dl_entity_preempt(&pi_task->dl, &p->dl))) {
7232 p->dl.pi_se = pi_task->dl.pi_se;
7233 queue_flag |= ENQUEUE_REPLENISH;
7234 } else {
7235 p->dl.pi_se = &p->dl;
7236 }
7237 } else if (rt_prio(prio)) {
7238 if (dl_prio(oldprio))
7239 p->dl.pi_se = &p->dl;
7240 if (oldprio < prio)
7241 queue_flag |= ENQUEUE_HEAD;
7242 } else {
7243 if (dl_prio(oldprio))
7244 p->dl.pi_se = &p->dl;
7245 if (rt_prio(oldprio))
7246 p->rt.timeout = 0;
7247 }
7248
7249 __setscheduler_prio(p, prio);
7250
7251 if (queued)
7252 enqueue_task(rq, p, queue_flag);
7253 if (running)
7254 set_next_task(rq, p);
7255
7256 check_class_changed(rq, p, prev_class, oldprio);
7257out_unlock:
7258 /* Avoid rq from going away on us: */
7259 preempt_disable();
7260
7261 rq_unpin_lock(rq, &rf);
7262 __balance_callbacks(rq);
7263 raw_spin_rq_unlock(rq);
7264
7265 preempt_enable();
7266}
7267#else
7268static inline int rt_effective_prio(struct task_struct *p, int prio)
7269{
7270 return prio;
7271}
7272#endif
7273
7274void set_user_nice(struct task_struct *p, long nice)
7275{
7276 bool queued, running;
7277 struct rq *rq;
7278 int old_prio;
7279
7280 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
7281 return;
7282 /*
7283 * We have to be careful, if called from sys_setpriority(),
7284 * the task might be in the middle of scheduling on another CPU.
7285 */
7286 CLASS(task_rq_lock, rq_guard)(p);
7287 rq = rq_guard.rq;
7288
7289 update_rq_clock(rq);
7290
7291 /*
7292 * The RT priorities are set via sched_setscheduler(), but we still
7293 * allow the 'normal' nice value to be set - but as expected
7294 * it won't have any effect on scheduling until the task is
7295 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
7296 */
7297 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
7298 p->static_prio = NICE_TO_PRIO(nice);
7299 return;
7300 }
7301
7302 queued = task_on_rq_queued(p);
7303 running = task_current(rq, p);
7304 if (queued)
7305 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
7306 if (running)
7307 put_prev_task(rq, p);
7308
7309 p->static_prio = NICE_TO_PRIO(nice);
7310 set_load_weight(p, true);
7311 old_prio = p->prio;
7312 p->prio = effective_prio(p);
7313
7314 if (queued)
7315 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7316 if (running)
7317 set_next_task(rq, p);
7318
7319 /*
7320 * If the task increased its priority or is running and
7321 * lowered its priority, then reschedule its CPU:
7322 */
7323 p->sched_class->prio_changed(rq, p, old_prio);
7324}
7325EXPORT_SYMBOL(set_user_nice);
7326
7327/*
7328 * is_nice_reduction - check if nice value is an actual reduction
7329 *
7330 * Similar to can_nice() but does not perform a capability check.
7331 *
7332 * @p: task
7333 * @nice: nice value
7334 */
7335static bool is_nice_reduction(const struct task_struct *p, const int nice)
7336{
7337 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
7338 int nice_rlim = nice_to_rlimit(nice);
7339
7340 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE));
7341}
7342
7343/*
7344 * can_nice - check if a task can reduce its nice value
7345 * @p: task
7346 * @nice: nice value
7347 */
7348int can_nice(const struct task_struct *p, const int nice)
7349{
7350 return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE);
7351}
7352
7353#ifdef __ARCH_WANT_SYS_NICE
7354
7355/*
7356 * sys_nice - change the priority of the current process.
7357 * @increment: priority increment
7358 *
7359 * sys_setpriority is a more generic, but much slower function that
7360 * does similar things.
7361 */
7362SYSCALL_DEFINE1(nice, int, increment)
7363{
7364 long nice, retval;
7365
7366 /*
7367 * Setpriority might change our priority at the same moment.
7368 * We don't have to worry. Conceptually one call occurs first
7369 * and we have a single winner.
7370 */
7371 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
7372 nice = task_nice(current) + increment;
7373
7374 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
7375 if (increment < 0 && !can_nice(current, nice))
7376 return -EPERM;
7377
7378 retval = security_task_setnice(current, nice);
7379 if (retval)
7380 return retval;
7381
7382 set_user_nice(current, nice);
7383 return 0;
7384}
7385
7386#endif
7387
7388/**
7389 * task_prio - return the priority value of a given task.
7390 * @p: the task in question.
7391 *
7392 * Return: The priority value as seen by users in /proc.
7393 *
7394 * sched policy return value kernel prio user prio/nice
7395 *
7396 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7397 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7398 * deadline -101 -1 0
7399 */
7400int task_prio(const struct task_struct *p)
7401{
7402 return p->prio - MAX_RT_PRIO;
7403}
7404
7405/**
7406 * idle_cpu - is a given CPU idle currently?
7407 * @cpu: the processor in question.
7408 *
7409 * Return: 1 if the CPU is currently idle. 0 otherwise.
7410 */
7411int idle_cpu(int cpu)
7412{
7413 struct rq *rq = cpu_rq(cpu);
7414
7415 if (rq->curr != rq->idle)
7416 return 0;
7417
7418 if (rq->nr_running)
7419 return 0;
7420
7421#ifdef CONFIG_SMP
7422 if (rq->ttwu_pending)
7423 return 0;
7424#endif
7425
7426 return 1;
7427}
7428
7429/**
7430 * available_idle_cpu - is a given CPU idle for enqueuing work.
7431 * @cpu: the CPU in question.
7432 *
7433 * Return: 1 if the CPU is currently idle. 0 otherwise.
7434 */
7435int available_idle_cpu(int cpu)
7436{
7437 if (!idle_cpu(cpu))
7438 return 0;
7439
7440 if (vcpu_is_preempted(cpu))
7441 return 0;
7442
7443 return 1;
7444}
7445
7446/**
7447 * idle_task - return the idle task for a given CPU.
7448 * @cpu: the processor in question.
7449 *
7450 * Return: The idle task for the CPU @cpu.
7451 */
7452struct task_struct *idle_task(int cpu)
7453{
7454 return cpu_rq(cpu)->idle;
7455}
7456
7457#ifdef CONFIG_SCHED_CORE
7458int sched_core_idle_cpu(int cpu)
7459{
7460 struct rq *rq = cpu_rq(cpu);
7461
7462 if (sched_core_enabled(rq) && rq->curr == rq->idle)
7463 return 1;
7464
7465 return idle_cpu(cpu);
7466}
7467
7468#endif
7469
7470#ifdef CONFIG_SMP
7471/*
7472 * This function computes an effective utilization for the given CPU, to be
7473 * used for frequency selection given the linear relation: f = u * f_max.
7474 *
7475 * The scheduler tracks the following metrics:
7476 *
7477 * cpu_util_{cfs,rt,dl,irq}()
7478 * cpu_bw_dl()
7479 *
7480 * Where the cfs,rt and dl util numbers are tracked with the same metric and
7481 * synchronized windows and are thus directly comparable.
7482 *
7483 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7484 * which excludes things like IRQ and steal-time. These latter are then accrued
7485 * in the irq utilization.
7486 *
7487 * The DL bandwidth number otoh is not a measured metric but a value computed
7488 * based on the task model parameters and gives the minimal utilization
7489 * required to meet deadlines.
7490 */
7491unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7492 unsigned long *min,
7493 unsigned long *max)
7494{
7495 unsigned long util, irq, scale;
7496 struct rq *rq = cpu_rq(cpu);
7497
7498 scale = arch_scale_cpu_capacity(cpu);
7499
7500 /*
7501 * Early check to see if IRQ/steal time saturates the CPU, can be
7502 * because of inaccuracies in how we track these -- see
7503 * update_irq_load_avg().
7504 */
7505 irq = cpu_util_irq(rq);
7506 if (unlikely(irq >= scale)) {
7507 if (min)
7508 *min = scale;
7509 if (max)
7510 *max = scale;
7511 return scale;
7512 }
7513
7514 if (min) {
7515 /*
7516 * The minimum utilization returns the highest level between:
7517 * - the computed DL bandwidth needed with the IRQ pressure which
7518 * steals time to the deadline task.
7519 * - The minimum performance requirement for CFS and/or RT.
7520 */
7521 *min = max(irq + cpu_bw_dl(rq), uclamp_rq_get(rq, UCLAMP_MIN));
7522
7523 /*
7524 * When an RT task is runnable and uclamp is not used, we must
7525 * ensure that the task will run at maximum compute capacity.
7526 */
7527 if (!uclamp_is_used() && rt_rq_is_runnable(&rq->rt))
7528 *min = max(*min, scale);
7529 }
7530
7531 /*
7532 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7533 * CFS tasks and we use the same metric to track the effective
7534 * utilization (PELT windows are synchronized) we can directly add them
7535 * to obtain the CPU's actual utilization.
7536 */
7537 util = util_cfs + cpu_util_rt(rq);
7538 util += cpu_util_dl(rq);
7539
7540 /*
7541 * The maximum hint is a soft bandwidth requirement, which can be lower
7542 * than the actual utilization because of uclamp_max requirements.
7543 */
7544 if (max)
7545 *max = min(scale, uclamp_rq_get(rq, UCLAMP_MAX));
7546
7547 if (util >= scale)
7548 return scale;
7549
7550 /*
7551 * There is still idle time; further improve the number by using the
7552 * irq metric. Because IRQ/steal time is hidden from the task clock we
7553 * need to scale the task numbers:
7554 *
7555 * max - irq
7556 * U' = irq + --------- * U
7557 * max
7558 */
7559 util = scale_irq_capacity(util, irq, scale);
7560 util += irq;
7561
7562 return min(scale, util);
7563}
7564
7565unsigned long sched_cpu_util(int cpu)
7566{
7567 return effective_cpu_util(cpu, cpu_util_cfs(cpu), NULL, NULL);
7568}
7569#endif /* CONFIG_SMP */
7570
7571/**
7572 * find_process_by_pid - find a process with a matching PID value.
7573 * @pid: the pid in question.
7574 *
7575 * The task of @pid, if found. %NULL otherwise.
7576 */
7577static struct task_struct *find_process_by_pid(pid_t pid)
7578{
7579 return pid ? find_task_by_vpid(pid) : current;
7580}
7581
7582static struct task_struct *find_get_task(pid_t pid)
7583{
7584 struct task_struct *p;
7585 guard(rcu)();
7586
7587 p = find_process_by_pid(pid);
7588 if (likely(p))
7589 get_task_struct(p);
7590
7591 return p;
7592}
7593
7594DEFINE_CLASS(find_get_task, struct task_struct *, if (_T) put_task_struct(_T),
7595 find_get_task(pid), pid_t pid)
7596
7597/*
7598 * sched_setparam() passes in -1 for its policy, to let the functions
7599 * it calls know not to change it.
7600 */
7601#define SETPARAM_POLICY -1
7602
7603static void __setscheduler_params(struct task_struct *p,
7604 const struct sched_attr *attr)
7605{
7606 int policy = attr->sched_policy;
7607
7608 if (policy == SETPARAM_POLICY)
7609 policy = p->policy;
7610
7611 p->policy = policy;
7612
7613 if (dl_policy(policy))
7614 __setparam_dl(p, attr);
7615 else if (fair_policy(policy))
7616 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7617
7618 /*
7619 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7620 * !rt_policy. Always setting this ensures that things like
7621 * getparam()/getattr() don't report silly values for !rt tasks.
7622 */
7623 p->rt_priority = attr->sched_priority;
7624 p->normal_prio = normal_prio(p);
7625 set_load_weight(p, true);
7626}
7627
7628/*
7629 * Check the target process has a UID that matches the current process's:
7630 */
7631static bool check_same_owner(struct task_struct *p)
7632{
7633 const struct cred *cred = current_cred(), *pcred;
7634 guard(rcu)();
7635
7636 pcred = __task_cred(p);
7637 return (uid_eq(cred->euid, pcred->euid) ||
7638 uid_eq(cred->euid, pcred->uid));
7639}
7640
7641/*
7642 * Allow unprivileged RT tasks to decrease priority.
7643 * Only issue a capable test if needed and only once to avoid an audit
7644 * event on permitted non-privileged operations:
7645 */
7646static int user_check_sched_setscheduler(struct task_struct *p,
7647 const struct sched_attr *attr,
7648 int policy, int reset_on_fork)
7649{
7650 if (fair_policy(policy)) {
7651 if (attr->sched_nice < task_nice(p) &&
7652 !is_nice_reduction(p, attr->sched_nice))
7653 goto req_priv;
7654 }
7655
7656 if (rt_policy(policy)) {
7657 unsigned long rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
7658
7659 /* Can't set/change the rt policy: */
7660 if (policy != p->policy && !rlim_rtprio)
7661 goto req_priv;
7662
7663 /* Can't increase priority: */
7664 if (attr->sched_priority > p->rt_priority &&
7665 attr->sched_priority > rlim_rtprio)
7666 goto req_priv;
7667 }
7668
7669 /*
7670 * Can't set/change SCHED_DEADLINE policy at all for now
7671 * (safest behavior); in the future we would like to allow
7672 * unprivileged DL tasks to increase their relative deadline
7673 * or reduce their runtime (both ways reducing utilization)
7674 */
7675 if (dl_policy(policy))
7676 goto req_priv;
7677
7678 /*
7679 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7680 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7681 */
7682 if (task_has_idle_policy(p) && !idle_policy(policy)) {
7683 if (!is_nice_reduction(p, task_nice(p)))
7684 goto req_priv;
7685 }
7686
7687 /* Can't change other user's priorities: */
7688 if (!check_same_owner(p))
7689 goto req_priv;
7690
7691 /* Normal users shall not reset the sched_reset_on_fork flag: */
7692 if (p->sched_reset_on_fork && !reset_on_fork)
7693 goto req_priv;
7694
7695 return 0;
7696
7697req_priv:
7698 if (!capable(CAP_SYS_NICE))
7699 return -EPERM;
7700
7701 return 0;
7702}
7703
7704static int __sched_setscheduler(struct task_struct *p,
7705 const struct sched_attr *attr,
7706 bool user, bool pi)
7707{
7708 int oldpolicy = -1, policy = attr->sched_policy;
7709 int retval, oldprio, newprio, queued, running;
7710 const struct sched_class *prev_class;
7711 struct balance_callback *head;
7712 struct rq_flags rf;
7713 int reset_on_fork;
7714 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7715 struct rq *rq;
7716 bool cpuset_locked = false;
7717
7718 /* The pi code expects interrupts enabled */
7719 BUG_ON(pi && in_interrupt());
7720recheck:
7721 /* Double check policy once rq lock held: */
7722 if (policy < 0) {
7723 reset_on_fork = p->sched_reset_on_fork;
7724 policy = oldpolicy = p->policy;
7725 } else {
7726 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7727
7728 if (!valid_policy(policy))
7729 return -EINVAL;
7730 }
7731
7732 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7733 return -EINVAL;
7734
7735 /*
7736 * Valid priorities for SCHED_FIFO and SCHED_RR are
7737 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7738 * SCHED_BATCH and SCHED_IDLE is 0.
7739 */
7740 if (attr->sched_priority > MAX_RT_PRIO-1)
7741 return -EINVAL;
7742 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7743 (rt_policy(policy) != (attr->sched_priority != 0)))
7744 return -EINVAL;
7745
7746 if (user) {
7747 retval = user_check_sched_setscheduler(p, attr, policy, reset_on_fork);
7748 if (retval)
7749 return retval;
7750
7751 if (attr->sched_flags & SCHED_FLAG_SUGOV)
7752 return -EINVAL;
7753
7754 retval = security_task_setscheduler(p);
7755 if (retval)
7756 return retval;
7757 }
7758
7759 /* Update task specific "requested" clamps */
7760 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7761 retval = uclamp_validate(p, attr);
7762 if (retval)
7763 return retval;
7764 }
7765
7766 /*
7767 * SCHED_DEADLINE bandwidth accounting relies on stable cpusets
7768 * information.
7769 */
7770 if (dl_policy(policy) || dl_policy(p->policy)) {
7771 cpuset_locked = true;
7772 cpuset_lock();
7773 }
7774
7775 /*
7776 * Make sure no PI-waiters arrive (or leave) while we are
7777 * changing the priority of the task:
7778 *
7779 * To be able to change p->policy safely, the appropriate
7780 * runqueue lock must be held.
7781 */
7782 rq = task_rq_lock(p, &rf);
7783 update_rq_clock(rq);
7784
7785 /*
7786 * Changing the policy of the stop threads its a very bad idea:
7787 */
7788 if (p == rq->stop) {
7789 retval = -EINVAL;
7790 goto unlock;
7791 }
7792
7793 /*
7794 * If not changing anything there's no need to proceed further,
7795 * but store a possible modification of reset_on_fork.
7796 */
7797 if (unlikely(policy == p->policy)) {
7798 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7799 goto change;
7800 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7801 goto change;
7802 if (dl_policy(policy) && dl_param_changed(p, attr))
7803 goto change;
7804 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7805 goto change;
7806
7807 p->sched_reset_on_fork = reset_on_fork;
7808 retval = 0;
7809 goto unlock;
7810 }
7811change:
7812
7813 if (user) {
7814#ifdef CONFIG_RT_GROUP_SCHED
7815 /*
7816 * Do not allow realtime tasks into groups that have no runtime
7817 * assigned.
7818 */
7819 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7820 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7821 !task_group_is_autogroup(task_group(p))) {
7822 retval = -EPERM;
7823 goto unlock;
7824 }
7825#endif
7826#ifdef CONFIG_SMP
7827 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7828 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7829 cpumask_t *span = rq->rd->span;
7830
7831 /*
7832 * Don't allow tasks with an affinity mask smaller than
7833 * the entire root_domain to become SCHED_DEADLINE. We
7834 * will also fail if there's no bandwidth available.
7835 */
7836 if (!cpumask_subset(span, p->cpus_ptr) ||
7837 rq->rd->dl_bw.bw == 0) {
7838 retval = -EPERM;
7839 goto unlock;
7840 }
7841 }
7842#endif
7843 }
7844
7845 /* Re-check policy now with rq lock held: */
7846 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7847 policy = oldpolicy = -1;
7848 task_rq_unlock(rq, p, &rf);
7849 if (cpuset_locked)
7850 cpuset_unlock();
7851 goto recheck;
7852 }
7853
7854 /*
7855 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7856 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7857 * is available.
7858 */
7859 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7860 retval = -EBUSY;
7861 goto unlock;
7862 }
7863
7864 p->sched_reset_on_fork = reset_on_fork;
7865 oldprio = p->prio;
7866
7867 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7868 if (pi) {
7869 /*
7870 * Take priority boosted tasks into account. If the new
7871 * effective priority is unchanged, we just store the new
7872 * normal parameters and do not touch the scheduler class and
7873 * the runqueue. This will be done when the task deboost
7874 * itself.
7875 */
7876 newprio = rt_effective_prio(p, newprio);
7877 if (newprio == oldprio)
7878 queue_flags &= ~DEQUEUE_MOVE;
7879 }
7880
7881 queued = task_on_rq_queued(p);
7882 running = task_current(rq, p);
7883 if (queued)
7884 dequeue_task(rq, p, queue_flags);
7885 if (running)
7886 put_prev_task(rq, p);
7887
7888 prev_class = p->sched_class;
7889
7890 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7891 __setscheduler_params(p, attr);
7892 __setscheduler_prio(p, newprio);
7893 }
7894 __setscheduler_uclamp(p, attr);
7895
7896 if (queued) {
7897 /*
7898 * We enqueue to tail when the priority of a task is
7899 * increased (user space view).
7900 */
7901 if (oldprio < p->prio)
7902 queue_flags |= ENQUEUE_HEAD;
7903
7904 enqueue_task(rq, p, queue_flags);
7905 }
7906 if (running)
7907 set_next_task(rq, p);
7908
7909 check_class_changed(rq, p, prev_class, oldprio);
7910
7911 /* Avoid rq from going away on us: */
7912 preempt_disable();
7913 head = splice_balance_callbacks(rq);
7914 task_rq_unlock(rq, p, &rf);
7915
7916 if (pi) {
7917 if (cpuset_locked)
7918 cpuset_unlock();
7919 rt_mutex_adjust_pi(p);
7920 }
7921
7922 /* Run balance callbacks after we've adjusted the PI chain: */
7923 balance_callbacks(rq, head);
7924 preempt_enable();
7925
7926 return 0;
7927
7928unlock:
7929 task_rq_unlock(rq, p, &rf);
7930 if (cpuset_locked)
7931 cpuset_unlock();
7932 return retval;
7933}
7934
7935static int _sched_setscheduler(struct task_struct *p, int policy,
7936 const struct sched_param *param, bool check)
7937{
7938 struct sched_attr attr = {
7939 .sched_policy = policy,
7940 .sched_priority = param->sched_priority,
7941 .sched_nice = PRIO_TO_NICE(p->static_prio),
7942 };
7943
7944 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7945 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7946 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7947 policy &= ~SCHED_RESET_ON_FORK;
7948 attr.sched_policy = policy;
7949 }
7950
7951 return __sched_setscheduler(p, &attr, check, true);
7952}
7953/**
7954 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7955 * @p: the task in question.
7956 * @policy: new policy.
7957 * @param: structure containing the new RT priority.
7958 *
7959 * Use sched_set_fifo(), read its comment.
7960 *
7961 * Return: 0 on success. An error code otherwise.
7962 *
7963 * NOTE that the task may be already dead.
7964 */
7965int sched_setscheduler(struct task_struct *p, int policy,
7966 const struct sched_param *param)
7967{
7968 return _sched_setscheduler(p, policy, param, true);
7969}
7970
7971int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7972{
7973 return __sched_setscheduler(p, attr, true, true);
7974}
7975
7976int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7977{
7978 return __sched_setscheduler(p, attr, false, true);
7979}
7980EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7981
7982/**
7983 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7984 * @p: the task in question.
7985 * @policy: new policy.
7986 * @param: structure containing the new RT priority.
7987 *
7988 * Just like sched_setscheduler, only don't bother checking if the
7989 * current context has permission. For example, this is needed in
7990 * stop_machine(): we create temporary high priority worker threads,
7991 * but our caller might not have that capability.
7992 *
7993 * Return: 0 on success. An error code otherwise.
7994 */
7995int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7996 const struct sched_param *param)
7997{
7998 return _sched_setscheduler(p, policy, param, false);
7999}
8000
8001/*
8002 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
8003 * incapable of resource management, which is the one thing an OS really should
8004 * be doing.
8005 *
8006 * This is of course the reason it is limited to privileged users only.
8007 *
8008 * Worse still; it is fundamentally impossible to compose static priority
8009 * workloads. You cannot take two correctly working static prio workloads
8010 * and smash them together and still expect them to work.
8011 *
8012 * For this reason 'all' FIFO tasks the kernel creates are basically at:
8013 *
8014 * MAX_RT_PRIO / 2
8015 *
8016 * The administrator _MUST_ configure the system, the kernel simply doesn't
8017 * know enough information to make a sensible choice.
8018 */
8019void sched_set_fifo(struct task_struct *p)
8020{
8021 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
8022 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
8023}
8024EXPORT_SYMBOL_GPL(sched_set_fifo);
8025
8026/*
8027 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
8028 */
8029void sched_set_fifo_low(struct task_struct *p)
8030{
8031 struct sched_param sp = { .sched_priority = 1 };
8032 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
8033}
8034EXPORT_SYMBOL_GPL(sched_set_fifo_low);
8035
8036void sched_set_normal(struct task_struct *p, int nice)
8037{
8038 struct sched_attr attr = {
8039 .sched_policy = SCHED_NORMAL,
8040 .sched_nice = nice,
8041 };
8042 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
8043}
8044EXPORT_SYMBOL_GPL(sched_set_normal);
8045
8046static int
8047do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
8048{
8049 struct sched_param lparam;
8050
8051 if (!param || pid < 0)
8052 return -EINVAL;
8053 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
8054 return -EFAULT;
8055
8056 CLASS(find_get_task, p)(pid);
8057 if (!p)
8058 return -ESRCH;
8059
8060 return sched_setscheduler(p, policy, &lparam);
8061}
8062
8063/*
8064 * Mimics kernel/events/core.c perf_copy_attr().
8065 */
8066static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
8067{
8068 u32 size;
8069 int ret;
8070
8071 /* Zero the full structure, so that a short copy will be nice: */
8072 memset(attr, 0, sizeof(*attr));
8073
8074 ret = get_user(size, &uattr->size);
8075 if (ret)
8076 return ret;
8077
8078 /* ABI compatibility quirk: */
8079 if (!size)
8080 size = SCHED_ATTR_SIZE_VER0;
8081 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
8082 goto err_size;
8083
8084 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
8085 if (ret) {
8086 if (ret == -E2BIG)
8087 goto err_size;
8088 return ret;
8089 }
8090
8091 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
8092 size < SCHED_ATTR_SIZE_VER1)
8093 return -EINVAL;
8094
8095 /*
8096 * XXX: Do we want to be lenient like existing syscalls; or do we want
8097 * to be strict and return an error on out-of-bounds values?
8098 */
8099 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
8100
8101 return 0;
8102
8103err_size:
8104 put_user(sizeof(*attr), &uattr->size);
8105 return -E2BIG;
8106}
8107
8108static void get_params(struct task_struct *p, struct sched_attr *attr)
8109{
8110 if (task_has_dl_policy(p))
8111 __getparam_dl(p, attr);
8112 else if (task_has_rt_policy(p))
8113 attr->sched_priority = p->rt_priority;
8114 else
8115 attr->sched_nice = task_nice(p);
8116}
8117
8118/**
8119 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
8120 * @pid: the pid in question.
8121 * @policy: new policy.
8122 * @param: structure containing the new RT priority.
8123 *
8124 * Return: 0 on success. An error code otherwise.
8125 */
8126SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
8127{
8128 if (policy < 0)
8129 return -EINVAL;
8130
8131 return do_sched_setscheduler(pid, policy, param);
8132}
8133
8134/**
8135 * sys_sched_setparam - set/change the RT priority of a thread
8136 * @pid: the pid in question.
8137 * @param: structure containing the new RT priority.
8138 *
8139 * Return: 0 on success. An error code otherwise.
8140 */
8141SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
8142{
8143 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
8144}
8145
8146/**
8147 * sys_sched_setattr - same as above, but with extended sched_attr
8148 * @pid: the pid in question.
8149 * @uattr: structure containing the extended parameters.
8150 * @flags: for future extension.
8151 */
8152SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
8153 unsigned int, flags)
8154{
8155 struct sched_attr attr;
8156 int retval;
8157
8158 if (!uattr || pid < 0 || flags)
8159 return -EINVAL;
8160
8161 retval = sched_copy_attr(uattr, &attr);
8162 if (retval)
8163 return retval;
8164
8165 if ((int)attr.sched_policy < 0)
8166 return -EINVAL;
8167 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
8168 attr.sched_policy = SETPARAM_POLICY;
8169
8170 CLASS(find_get_task, p)(pid);
8171 if (!p)
8172 return -ESRCH;
8173
8174 if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
8175 get_params(p, &attr);
8176
8177 return sched_setattr(p, &attr);
8178}
8179
8180/**
8181 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
8182 * @pid: the pid in question.
8183 *
8184 * Return: On success, the policy of the thread. Otherwise, a negative error
8185 * code.
8186 */
8187SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
8188{
8189 struct task_struct *p;
8190 int retval;
8191
8192 if (pid < 0)
8193 return -EINVAL;
8194
8195 guard(rcu)();
8196 p = find_process_by_pid(pid);
8197 if (!p)
8198 return -ESRCH;
8199
8200 retval = security_task_getscheduler(p);
8201 if (!retval) {
8202 retval = p->policy;
8203 if (p->sched_reset_on_fork)
8204 retval |= SCHED_RESET_ON_FORK;
8205 }
8206 return retval;
8207}
8208
8209/**
8210 * sys_sched_getparam - get the RT priority of a thread
8211 * @pid: the pid in question.
8212 * @param: structure containing the RT priority.
8213 *
8214 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
8215 * code.
8216 */
8217SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
8218{
8219 struct sched_param lp = { .sched_priority = 0 };
8220 struct task_struct *p;
8221 int retval;
8222
8223 if (!param || pid < 0)
8224 return -EINVAL;
8225
8226 scoped_guard (rcu) {
8227 p = find_process_by_pid(pid);
8228 if (!p)
8229 return -ESRCH;
8230
8231 retval = security_task_getscheduler(p);
8232 if (retval)
8233 return retval;
8234
8235 if (task_has_rt_policy(p))
8236 lp.sched_priority = p->rt_priority;
8237 }
8238
8239 /*
8240 * This one might sleep, we cannot do it with a spinlock held ...
8241 */
8242 return copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
8243}
8244
8245/*
8246 * Copy the kernel size attribute structure (which might be larger
8247 * than what user-space knows about) to user-space.
8248 *
8249 * Note that all cases are valid: user-space buffer can be larger or
8250 * smaller than the kernel-space buffer. The usual case is that both
8251 * have the same size.
8252 */
8253static int
8254sched_attr_copy_to_user(struct sched_attr __user *uattr,
8255 struct sched_attr *kattr,
8256 unsigned int usize)
8257{
8258 unsigned int ksize = sizeof(*kattr);
8259
8260 if (!access_ok(uattr, usize))
8261 return -EFAULT;
8262
8263 /*
8264 * sched_getattr() ABI forwards and backwards compatibility:
8265 *
8266 * If usize == ksize then we just copy everything to user-space and all is good.
8267 *
8268 * If usize < ksize then we only copy as much as user-space has space for,
8269 * this keeps ABI compatibility as well. We skip the rest.
8270 *
8271 * If usize > ksize then user-space is using a newer version of the ABI,
8272 * which part the kernel doesn't know about. Just ignore it - tooling can
8273 * detect the kernel's knowledge of attributes from the attr->size value
8274 * which is set to ksize in this case.
8275 */
8276 kattr->size = min(usize, ksize);
8277
8278 if (copy_to_user(uattr, kattr, kattr->size))
8279 return -EFAULT;
8280
8281 return 0;
8282}
8283
8284/**
8285 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
8286 * @pid: the pid in question.
8287 * @uattr: structure containing the extended parameters.
8288 * @usize: sizeof(attr) for fwd/bwd comp.
8289 * @flags: for future extension.
8290 */
8291SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
8292 unsigned int, usize, unsigned int, flags)
8293{
8294 struct sched_attr kattr = { };
8295 struct task_struct *p;
8296 int retval;
8297
8298 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
8299 usize < SCHED_ATTR_SIZE_VER0 || flags)
8300 return -EINVAL;
8301
8302 scoped_guard (rcu) {
8303 p = find_process_by_pid(pid);
8304 if (!p)
8305 return -ESRCH;
8306
8307 retval = security_task_getscheduler(p);
8308 if (retval)
8309 return retval;
8310
8311 kattr.sched_policy = p->policy;
8312 if (p->sched_reset_on_fork)
8313 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
8314 get_params(p, &kattr);
8315 kattr.sched_flags &= SCHED_FLAG_ALL;
8316
8317#ifdef CONFIG_UCLAMP_TASK
8318 /*
8319 * This could race with another potential updater, but this is fine
8320 * because it'll correctly read the old or the new value. We don't need
8321 * to guarantee who wins the race as long as it doesn't return garbage.
8322 */
8323 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
8324 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
8325#endif
8326 }
8327
8328 return sched_attr_copy_to_user(uattr, &kattr, usize);
8329}
8330
8331#ifdef CONFIG_SMP
8332int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
8333{
8334 /*
8335 * If the task isn't a deadline task or admission control is
8336 * disabled then we don't care about affinity changes.
8337 */
8338 if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
8339 return 0;
8340
8341 /*
8342 * Since bandwidth control happens on root_domain basis,
8343 * if admission test is enabled, we only admit -deadline
8344 * tasks allowed to run on all the CPUs in the task's
8345 * root_domain.
8346 */
8347 guard(rcu)();
8348 if (!cpumask_subset(task_rq(p)->rd->span, mask))
8349 return -EBUSY;
8350
8351 return 0;
8352}
8353#endif
8354
8355static int
8356__sched_setaffinity(struct task_struct *p, struct affinity_context *ctx)
8357{
8358 int retval;
8359 cpumask_var_t cpus_allowed, new_mask;
8360
8361 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
8362 return -ENOMEM;
8363
8364 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
8365 retval = -ENOMEM;
8366 goto out_free_cpus_allowed;
8367 }
8368
8369 cpuset_cpus_allowed(p, cpus_allowed);
8370 cpumask_and(new_mask, ctx->new_mask, cpus_allowed);
8371
8372 ctx->new_mask = new_mask;
8373 ctx->flags |= SCA_CHECK;
8374
8375 retval = dl_task_check_affinity(p, new_mask);
8376 if (retval)
8377 goto out_free_new_mask;
8378
8379 retval = __set_cpus_allowed_ptr(p, ctx);
8380 if (retval)
8381 goto out_free_new_mask;
8382
8383 cpuset_cpus_allowed(p, cpus_allowed);
8384 if (!cpumask_subset(new_mask, cpus_allowed)) {
8385 /*
8386 * We must have raced with a concurrent cpuset update.
8387 * Just reset the cpumask to the cpuset's cpus_allowed.
8388 */
8389 cpumask_copy(new_mask, cpus_allowed);
8390
8391 /*
8392 * If SCA_USER is set, a 2nd call to __set_cpus_allowed_ptr()
8393 * will restore the previous user_cpus_ptr value.
8394 *
8395 * In the unlikely event a previous user_cpus_ptr exists,
8396 * we need to further restrict the mask to what is allowed
8397 * by that old user_cpus_ptr.
8398 */
8399 if (unlikely((ctx->flags & SCA_USER) && ctx->user_mask)) {
8400 bool empty = !cpumask_and(new_mask, new_mask,
8401 ctx->user_mask);
8402
8403 if (WARN_ON_ONCE(empty))
8404 cpumask_copy(new_mask, cpus_allowed);
8405 }
8406 __set_cpus_allowed_ptr(p, ctx);
8407 retval = -EINVAL;
8408 }
8409
8410out_free_new_mask:
8411 free_cpumask_var(new_mask);
8412out_free_cpus_allowed:
8413 free_cpumask_var(cpus_allowed);
8414 return retval;
8415}
8416
8417long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8418{
8419 struct affinity_context ac;
8420 struct cpumask *user_mask;
8421 int retval;
8422
8423 CLASS(find_get_task, p)(pid);
8424 if (!p)
8425 return -ESRCH;
8426
8427 if (p->flags & PF_NO_SETAFFINITY)
8428 return -EINVAL;
8429
8430 if (!check_same_owner(p)) {
8431 guard(rcu)();
8432 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE))
8433 return -EPERM;
8434 }
8435
8436 retval = security_task_setscheduler(p);
8437 if (retval)
8438 return retval;
8439
8440 /*
8441 * With non-SMP configs, user_cpus_ptr/user_mask isn't used and
8442 * alloc_user_cpus_ptr() returns NULL.
8443 */
8444 user_mask = alloc_user_cpus_ptr(NUMA_NO_NODE);
8445 if (user_mask) {
8446 cpumask_copy(user_mask, in_mask);
8447 } else if (IS_ENABLED(CONFIG_SMP)) {
8448 return -ENOMEM;
8449 }
8450
8451 ac = (struct affinity_context){
8452 .new_mask = in_mask,
8453 .user_mask = user_mask,
8454 .flags = SCA_USER,
8455 };
8456
8457 retval = __sched_setaffinity(p, &ac);
8458 kfree(ac.user_mask);
8459
8460 return retval;
8461}
8462
8463static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8464 struct cpumask *new_mask)
8465{
8466 if (len < cpumask_size())
8467 cpumask_clear(new_mask);
8468 else if (len > cpumask_size())
8469 len = cpumask_size();
8470
8471 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8472}
8473
8474/**
8475 * sys_sched_setaffinity - set the CPU affinity of a process
8476 * @pid: pid of the process
8477 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8478 * @user_mask_ptr: user-space pointer to the new CPU mask
8479 *
8480 * Return: 0 on success. An error code otherwise.
8481 */
8482SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8483 unsigned long __user *, user_mask_ptr)
8484{
8485 cpumask_var_t new_mask;
8486 int retval;
8487
8488 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8489 return -ENOMEM;
8490
8491 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8492 if (retval == 0)
8493 retval = sched_setaffinity(pid, new_mask);
8494 free_cpumask_var(new_mask);
8495 return retval;
8496}
8497
8498long sched_getaffinity(pid_t pid, struct cpumask *mask)
8499{
8500 struct task_struct *p;
8501 int retval;
8502
8503 guard(rcu)();
8504 p = find_process_by_pid(pid);
8505 if (!p)
8506 return -ESRCH;
8507
8508 retval = security_task_getscheduler(p);
8509 if (retval)
8510 return retval;
8511
8512 guard(raw_spinlock_irqsave)(&p->pi_lock);
8513 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8514
8515 return 0;
8516}
8517
8518/**
8519 * sys_sched_getaffinity - get the CPU affinity of a process
8520 * @pid: pid of the process
8521 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8522 * @user_mask_ptr: user-space pointer to hold the current CPU mask
8523 *
8524 * Return: size of CPU mask copied to user_mask_ptr on success. An
8525 * error code otherwise.
8526 */
8527SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8528 unsigned long __user *, user_mask_ptr)
8529{
8530 int ret;
8531 cpumask_var_t mask;
8532
8533 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8534 return -EINVAL;
8535 if (len & (sizeof(unsigned long)-1))
8536 return -EINVAL;
8537
8538 if (!zalloc_cpumask_var(&mask, GFP_KERNEL))
8539 return -ENOMEM;
8540
8541 ret = sched_getaffinity(pid, mask);
8542 if (ret == 0) {
8543 unsigned int retlen = min(len, cpumask_size());
8544
8545 if (copy_to_user(user_mask_ptr, cpumask_bits(mask), retlen))
8546 ret = -EFAULT;
8547 else
8548 ret = retlen;
8549 }
8550 free_cpumask_var(mask);
8551
8552 return ret;
8553}
8554
8555static void do_sched_yield(void)
8556{
8557 struct rq_flags rf;
8558 struct rq *rq;
8559
8560 rq = this_rq_lock_irq(&rf);
8561
8562 schedstat_inc(rq->yld_count);
8563 current->sched_class->yield_task(rq);
8564
8565 preempt_disable();
8566 rq_unlock_irq(rq, &rf);
8567 sched_preempt_enable_no_resched();
8568
8569 schedule();
8570}
8571
8572/**
8573 * sys_sched_yield - yield the current processor to other threads.
8574 *
8575 * This function yields the current CPU to other tasks. If there are no
8576 * other threads running on this CPU then this function will return.
8577 *
8578 * Return: 0.
8579 */
8580SYSCALL_DEFINE0(sched_yield)
8581{
8582 do_sched_yield();
8583 return 0;
8584}
8585
8586#if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8587int __sched __cond_resched(void)
8588{
8589 if (should_resched(0)) {
8590 preempt_schedule_common();
8591 return 1;
8592 }
8593 /*
8594 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8595 * whether the current CPU is in an RCU read-side critical section,
8596 * so the tick can report quiescent states even for CPUs looping
8597 * in kernel context. In contrast, in non-preemptible kernels,
8598 * RCU readers leave no in-memory hints, which means that CPU-bound
8599 * processes executing in kernel context might never report an
8600 * RCU quiescent state. Therefore, the following code causes
8601 * cond_resched() to report a quiescent state, but only when RCU
8602 * is in urgent need of one.
8603 */
8604#ifndef CONFIG_PREEMPT_RCU
8605 rcu_all_qs();
8606#endif
8607 return 0;
8608}
8609EXPORT_SYMBOL(__cond_resched);
8610#endif
8611
8612#ifdef CONFIG_PREEMPT_DYNAMIC
8613#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8614#define cond_resched_dynamic_enabled __cond_resched
8615#define cond_resched_dynamic_disabled ((void *)&__static_call_return0)
8616DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8617EXPORT_STATIC_CALL_TRAMP(cond_resched);
8618
8619#define might_resched_dynamic_enabled __cond_resched
8620#define might_resched_dynamic_disabled ((void *)&__static_call_return0)
8621DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8622EXPORT_STATIC_CALL_TRAMP(might_resched);
8623#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8624static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
8625int __sched dynamic_cond_resched(void)
8626{
8627 klp_sched_try_switch();
8628 if (!static_branch_unlikely(&sk_dynamic_cond_resched))
8629 return 0;
8630 return __cond_resched();
8631}
8632EXPORT_SYMBOL(dynamic_cond_resched);
8633
8634static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
8635int __sched dynamic_might_resched(void)
8636{
8637 if (!static_branch_unlikely(&sk_dynamic_might_resched))
8638 return 0;
8639 return __cond_resched();
8640}
8641EXPORT_SYMBOL(dynamic_might_resched);
8642#endif
8643#endif
8644
8645/*
8646 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8647 * call schedule, and on return reacquire the lock.
8648 *
8649 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8650 * operations here to prevent schedule() from being called twice (once via
8651 * spin_unlock(), once by hand).
8652 */
8653int __cond_resched_lock(spinlock_t *lock)
8654{
8655 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8656 int ret = 0;
8657
8658 lockdep_assert_held(lock);
8659
8660 if (spin_needbreak(lock) || resched) {
8661 spin_unlock(lock);
8662 if (!_cond_resched())
8663 cpu_relax();
8664 ret = 1;
8665 spin_lock(lock);
8666 }
8667 return ret;
8668}
8669EXPORT_SYMBOL(__cond_resched_lock);
8670
8671int __cond_resched_rwlock_read(rwlock_t *lock)
8672{
8673 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8674 int ret = 0;
8675
8676 lockdep_assert_held_read(lock);
8677
8678 if (rwlock_needbreak(lock) || resched) {
8679 read_unlock(lock);
8680 if (!_cond_resched())
8681 cpu_relax();
8682 ret = 1;
8683 read_lock(lock);
8684 }
8685 return ret;
8686}
8687EXPORT_SYMBOL(__cond_resched_rwlock_read);
8688
8689int __cond_resched_rwlock_write(rwlock_t *lock)
8690{
8691 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8692 int ret = 0;
8693
8694 lockdep_assert_held_write(lock);
8695
8696 if (rwlock_needbreak(lock) || resched) {
8697 write_unlock(lock);
8698 if (!_cond_resched())
8699 cpu_relax();
8700 ret = 1;
8701 write_lock(lock);
8702 }
8703 return ret;
8704}
8705EXPORT_SYMBOL(__cond_resched_rwlock_write);
8706
8707#ifdef CONFIG_PREEMPT_DYNAMIC
8708
8709#ifdef CONFIG_GENERIC_ENTRY
8710#include <linux/entry-common.h>
8711#endif
8712
8713/*
8714 * SC:cond_resched
8715 * SC:might_resched
8716 * SC:preempt_schedule
8717 * SC:preempt_schedule_notrace
8718 * SC:irqentry_exit_cond_resched
8719 *
8720 *
8721 * NONE:
8722 * cond_resched <- __cond_resched
8723 * might_resched <- RET0
8724 * preempt_schedule <- NOP
8725 * preempt_schedule_notrace <- NOP
8726 * irqentry_exit_cond_resched <- NOP
8727 *
8728 * VOLUNTARY:
8729 * cond_resched <- __cond_resched
8730 * might_resched <- __cond_resched
8731 * preempt_schedule <- NOP
8732 * preempt_schedule_notrace <- NOP
8733 * irqentry_exit_cond_resched <- NOP
8734 *
8735 * FULL:
8736 * cond_resched <- RET0
8737 * might_resched <- RET0
8738 * preempt_schedule <- preempt_schedule
8739 * preempt_schedule_notrace <- preempt_schedule_notrace
8740 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
8741 */
8742
8743enum {
8744 preempt_dynamic_undefined = -1,
8745 preempt_dynamic_none,
8746 preempt_dynamic_voluntary,
8747 preempt_dynamic_full,
8748};
8749
8750int preempt_dynamic_mode = preempt_dynamic_undefined;
8751
8752int sched_dynamic_mode(const char *str)
8753{
8754 if (!strcmp(str, "none"))
8755 return preempt_dynamic_none;
8756
8757 if (!strcmp(str, "voluntary"))
8758 return preempt_dynamic_voluntary;
8759
8760 if (!strcmp(str, "full"))
8761 return preempt_dynamic_full;
8762
8763 return -EINVAL;
8764}
8765
8766#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8767#define preempt_dynamic_enable(f) static_call_update(f, f##_dynamic_enabled)
8768#define preempt_dynamic_disable(f) static_call_update(f, f##_dynamic_disabled)
8769#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8770#define preempt_dynamic_enable(f) static_key_enable(&sk_dynamic_##f.key)
8771#define preempt_dynamic_disable(f) static_key_disable(&sk_dynamic_##f.key)
8772#else
8773#error "Unsupported PREEMPT_DYNAMIC mechanism"
8774#endif
8775
8776static DEFINE_MUTEX(sched_dynamic_mutex);
8777static bool klp_override;
8778
8779static void __sched_dynamic_update(int mode)
8780{
8781 /*
8782 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
8783 * the ZERO state, which is invalid.
8784 */
8785 if (!klp_override)
8786 preempt_dynamic_enable(cond_resched);
8787 preempt_dynamic_enable(might_resched);
8788 preempt_dynamic_enable(preempt_schedule);
8789 preempt_dynamic_enable(preempt_schedule_notrace);
8790 preempt_dynamic_enable(irqentry_exit_cond_resched);
8791
8792 switch (mode) {
8793 case preempt_dynamic_none:
8794 if (!klp_override)
8795 preempt_dynamic_enable(cond_resched);
8796 preempt_dynamic_disable(might_resched);
8797 preempt_dynamic_disable(preempt_schedule);
8798 preempt_dynamic_disable(preempt_schedule_notrace);
8799 preempt_dynamic_disable(irqentry_exit_cond_resched);
8800 if (mode != preempt_dynamic_mode)
8801 pr_info("Dynamic Preempt: none\n");
8802 break;
8803
8804 case preempt_dynamic_voluntary:
8805 if (!klp_override)
8806 preempt_dynamic_enable(cond_resched);
8807 preempt_dynamic_enable(might_resched);
8808 preempt_dynamic_disable(preempt_schedule);
8809 preempt_dynamic_disable(preempt_schedule_notrace);
8810 preempt_dynamic_disable(irqentry_exit_cond_resched);
8811 if (mode != preempt_dynamic_mode)
8812 pr_info("Dynamic Preempt: voluntary\n");
8813 break;
8814
8815 case preempt_dynamic_full:
8816 if (!klp_override)
8817 preempt_dynamic_disable(cond_resched);
8818 preempt_dynamic_disable(might_resched);
8819 preempt_dynamic_enable(preempt_schedule);
8820 preempt_dynamic_enable(preempt_schedule_notrace);
8821 preempt_dynamic_enable(irqentry_exit_cond_resched);
8822 if (mode != preempt_dynamic_mode)
8823 pr_info("Dynamic Preempt: full\n");
8824 break;
8825 }
8826
8827 preempt_dynamic_mode = mode;
8828}
8829
8830void sched_dynamic_update(int mode)
8831{
8832 mutex_lock(&sched_dynamic_mutex);
8833 __sched_dynamic_update(mode);
8834 mutex_unlock(&sched_dynamic_mutex);
8835}
8836
8837#ifdef CONFIG_HAVE_PREEMPT_DYNAMIC_CALL
8838
8839static int klp_cond_resched(void)
8840{
8841 __klp_sched_try_switch();
8842 return __cond_resched();
8843}
8844
8845void sched_dynamic_klp_enable(void)
8846{
8847 mutex_lock(&sched_dynamic_mutex);
8848
8849 klp_override = true;
8850 static_call_update(cond_resched, klp_cond_resched);
8851
8852 mutex_unlock(&sched_dynamic_mutex);
8853}
8854
8855void sched_dynamic_klp_disable(void)
8856{
8857 mutex_lock(&sched_dynamic_mutex);
8858
8859 klp_override = false;
8860 __sched_dynamic_update(preempt_dynamic_mode);
8861
8862 mutex_unlock(&sched_dynamic_mutex);
8863}
8864
8865#endif /* CONFIG_HAVE_PREEMPT_DYNAMIC_CALL */
8866
8867static int __init setup_preempt_mode(char *str)
8868{
8869 int mode = sched_dynamic_mode(str);
8870 if (mode < 0) {
8871 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
8872 return 0;
8873 }
8874
8875 sched_dynamic_update(mode);
8876 return 1;
8877}
8878__setup("preempt=", setup_preempt_mode);
8879
8880static void __init preempt_dynamic_init(void)
8881{
8882 if (preempt_dynamic_mode == preempt_dynamic_undefined) {
8883 if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
8884 sched_dynamic_update(preempt_dynamic_none);
8885 } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
8886 sched_dynamic_update(preempt_dynamic_voluntary);
8887 } else {
8888 /* Default static call setting, nothing to do */
8889 WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
8890 preempt_dynamic_mode = preempt_dynamic_full;
8891 pr_info("Dynamic Preempt: full\n");
8892 }
8893 }
8894}
8895
8896#define PREEMPT_MODEL_ACCESSOR(mode) \
8897 bool preempt_model_##mode(void) \
8898 { \
8899 WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
8900 return preempt_dynamic_mode == preempt_dynamic_##mode; \
8901 } \
8902 EXPORT_SYMBOL_GPL(preempt_model_##mode)
8903
8904PREEMPT_MODEL_ACCESSOR(none);
8905PREEMPT_MODEL_ACCESSOR(voluntary);
8906PREEMPT_MODEL_ACCESSOR(full);
8907
8908#else /* !CONFIG_PREEMPT_DYNAMIC */
8909
8910static inline void preempt_dynamic_init(void) { }
8911
8912#endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
8913
8914/**
8915 * yield - yield the current processor to other threads.
8916 *
8917 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8918 *
8919 * The scheduler is at all times free to pick the calling task as the most
8920 * eligible task to run, if removing the yield() call from your code breaks
8921 * it, it's already broken.
8922 *
8923 * Typical broken usage is:
8924 *
8925 * while (!event)
8926 * yield();
8927 *
8928 * where one assumes that yield() will let 'the other' process run that will
8929 * make event true. If the current task is a SCHED_FIFO task that will never
8930 * happen. Never use yield() as a progress guarantee!!
8931 *
8932 * If you want to use yield() to wait for something, use wait_event().
8933 * If you want to use yield() to be 'nice' for others, use cond_resched().
8934 * If you still want to use yield(), do not!
8935 */
8936void __sched yield(void)
8937{
8938 set_current_state(TASK_RUNNING);
8939 do_sched_yield();
8940}
8941EXPORT_SYMBOL(yield);
8942
8943/**
8944 * yield_to - yield the current processor to another thread in
8945 * your thread group, or accelerate that thread toward the
8946 * processor it's on.
8947 * @p: target task
8948 * @preempt: whether task preemption is allowed or not
8949 *
8950 * It's the caller's job to ensure that the target task struct
8951 * can't go away on us before we can do any checks.
8952 *
8953 * Return:
8954 * true (>0) if we indeed boosted the target task.
8955 * false (0) if we failed to boost the target.
8956 * -ESRCH if there's no task to yield to.
8957 */
8958int __sched yield_to(struct task_struct *p, bool preempt)
8959{
8960 struct task_struct *curr = current;
8961 struct rq *rq, *p_rq;
8962 int yielded = 0;
8963
8964 scoped_guard (irqsave) {
8965 rq = this_rq();
8966
8967again:
8968 p_rq = task_rq(p);
8969 /*
8970 * If we're the only runnable task on the rq and target rq also
8971 * has only one task, there's absolutely no point in yielding.
8972 */
8973 if (rq->nr_running == 1 && p_rq->nr_running == 1)
8974 return -ESRCH;
8975
8976 guard(double_rq_lock)(rq, p_rq);
8977 if (task_rq(p) != p_rq)
8978 goto again;
8979
8980 if (!curr->sched_class->yield_to_task)
8981 return 0;
8982
8983 if (curr->sched_class != p->sched_class)
8984 return 0;
8985
8986 if (task_on_cpu(p_rq, p) || !task_is_running(p))
8987 return 0;
8988
8989 yielded = curr->sched_class->yield_to_task(rq, p);
8990 if (yielded) {
8991 schedstat_inc(rq->yld_count);
8992 /*
8993 * Make p's CPU reschedule; pick_next_entity
8994 * takes care of fairness.
8995 */
8996 if (preempt && rq != p_rq)
8997 resched_curr(p_rq);
8998 }
8999 }
9000
9001 if (yielded)
9002 schedule();
9003
9004 return yielded;
9005}
9006EXPORT_SYMBOL_GPL(yield_to);
9007
9008int io_schedule_prepare(void)
9009{
9010 int old_iowait = current->in_iowait;
9011
9012 current->in_iowait = 1;
9013 blk_flush_plug(current->plug, true);
9014 return old_iowait;
9015}
9016
9017void io_schedule_finish(int token)
9018{
9019 current->in_iowait = token;
9020}
9021
9022/*
9023 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
9024 * that process accounting knows that this is a task in IO wait state.
9025 */
9026long __sched io_schedule_timeout(long timeout)
9027{
9028 int token;
9029 long ret;
9030
9031 token = io_schedule_prepare();
9032 ret = schedule_timeout(timeout);
9033 io_schedule_finish(token);
9034
9035 return ret;
9036}
9037EXPORT_SYMBOL(io_schedule_timeout);
9038
9039void __sched io_schedule(void)
9040{
9041 int token;
9042
9043 token = io_schedule_prepare();
9044 schedule();
9045 io_schedule_finish(token);
9046}
9047EXPORT_SYMBOL(io_schedule);
9048
9049/**
9050 * sys_sched_get_priority_max - return maximum RT priority.
9051 * @policy: scheduling class.
9052 *
9053 * Return: On success, this syscall returns the maximum
9054 * rt_priority that can be used by a given scheduling class.
9055 * On failure, a negative error code is returned.
9056 */
9057SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
9058{
9059 int ret = -EINVAL;
9060
9061 switch (policy) {
9062 case SCHED_FIFO:
9063 case SCHED_RR:
9064 ret = MAX_RT_PRIO-1;
9065 break;
9066 case SCHED_DEADLINE:
9067 case SCHED_NORMAL:
9068 case SCHED_BATCH:
9069 case SCHED_IDLE:
9070 ret = 0;
9071 break;
9072 }
9073 return ret;
9074}
9075
9076/**
9077 * sys_sched_get_priority_min - return minimum RT priority.
9078 * @policy: scheduling class.
9079 *
9080 * Return: On success, this syscall returns the minimum
9081 * rt_priority that can be used by a given scheduling class.
9082 * On failure, a negative error code is returned.
9083 */
9084SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
9085{
9086 int ret = -EINVAL;
9087
9088 switch (policy) {
9089 case SCHED_FIFO:
9090 case SCHED_RR:
9091 ret = 1;
9092 break;
9093 case SCHED_DEADLINE:
9094 case SCHED_NORMAL:
9095 case SCHED_BATCH:
9096 case SCHED_IDLE:
9097 ret = 0;
9098 }
9099 return ret;
9100}
9101
9102static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
9103{
9104 unsigned int time_slice = 0;
9105 int retval;
9106
9107 if (pid < 0)
9108 return -EINVAL;
9109
9110 scoped_guard (rcu) {
9111 struct task_struct *p = find_process_by_pid(pid);
9112 if (!p)
9113 return -ESRCH;
9114
9115 retval = security_task_getscheduler(p);
9116 if (retval)
9117 return retval;
9118
9119 scoped_guard (task_rq_lock, p) {
9120 struct rq *rq = scope.rq;
9121 if (p->sched_class->get_rr_interval)
9122 time_slice = p->sched_class->get_rr_interval(rq, p);
9123 }
9124 }
9125
9126 jiffies_to_timespec64(time_slice, t);
9127 return 0;
9128}
9129
9130/**
9131 * sys_sched_rr_get_interval - return the default timeslice of a process.
9132 * @pid: pid of the process.
9133 * @interval: userspace pointer to the timeslice value.
9134 *
9135 * this syscall writes the default timeslice value of a given process
9136 * into the user-space timespec buffer. A value of '0' means infinity.
9137 *
9138 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
9139 * an error code.
9140 */
9141SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
9142 struct __kernel_timespec __user *, interval)
9143{
9144 struct timespec64 t;
9145 int retval = sched_rr_get_interval(pid, &t);
9146
9147 if (retval == 0)
9148 retval = put_timespec64(&t, interval);
9149
9150 return retval;
9151}
9152
9153#ifdef CONFIG_COMPAT_32BIT_TIME
9154SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
9155 struct old_timespec32 __user *, interval)
9156{
9157 struct timespec64 t;
9158 int retval = sched_rr_get_interval(pid, &t);
9159
9160 if (retval == 0)
9161 retval = put_old_timespec32(&t, interval);
9162 return retval;
9163}
9164#endif
9165
9166void sched_show_task(struct task_struct *p)
9167{
9168 unsigned long free = 0;
9169 int ppid;
9170
9171 if (!try_get_task_stack(p))
9172 return;
9173
9174 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
9175
9176 if (task_is_running(p))
9177 pr_cont(" running task ");
9178#ifdef CONFIG_DEBUG_STACK_USAGE
9179 free = stack_not_used(p);
9180#endif
9181 ppid = 0;
9182 rcu_read_lock();
9183 if (pid_alive(p))
9184 ppid = task_pid_nr(rcu_dereference(p->real_parent));
9185 rcu_read_unlock();
9186 pr_cont(" stack:%-5lu pid:%-5d tgid:%-5d ppid:%-6d flags:0x%08lx\n",
9187 free, task_pid_nr(p), task_tgid_nr(p),
9188 ppid, read_task_thread_flags(p));
9189
9190 print_worker_info(KERN_INFO, p);
9191 print_stop_info(KERN_INFO, p);
9192 show_stack(p, NULL, KERN_INFO);
9193 put_task_stack(p);
9194}
9195EXPORT_SYMBOL_GPL(sched_show_task);
9196
9197static inline bool
9198state_filter_match(unsigned long state_filter, struct task_struct *p)
9199{
9200 unsigned int state = READ_ONCE(p->__state);
9201
9202 /* no filter, everything matches */
9203 if (!state_filter)
9204 return true;
9205
9206 /* filter, but doesn't match */
9207 if (!(state & state_filter))
9208 return false;
9209
9210 /*
9211 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
9212 * TASK_KILLABLE).
9213 */
9214 if (state_filter == TASK_UNINTERRUPTIBLE && (state & TASK_NOLOAD))
9215 return false;
9216
9217 return true;
9218}
9219
9220
9221void show_state_filter(unsigned int state_filter)
9222{
9223 struct task_struct *g, *p;
9224
9225 rcu_read_lock();
9226 for_each_process_thread(g, p) {
9227 /*
9228 * reset the NMI-timeout, listing all files on a slow
9229 * console might take a lot of time:
9230 * Also, reset softlockup watchdogs on all CPUs, because
9231 * another CPU might be blocked waiting for us to process
9232 * an IPI.
9233 */
9234 touch_nmi_watchdog();
9235 touch_all_softlockup_watchdogs();
9236 if (state_filter_match(state_filter, p))
9237 sched_show_task(p);
9238 }
9239
9240#ifdef CONFIG_SCHED_DEBUG
9241 if (!state_filter)
9242 sysrq_sched_debug_show();
9243#endif
9244 rcu_read_unlock();
9245 /*
9246 * Only show locks if all tasks are dumped:
9247 */
9248 if (!state_filter)
9249 debug_show_all_locks();
9250}
9251
9252/**
9253 * init_idle - set up an idle thread for a given CPU
9254 * @idle: task in question
9255 * @cpu: CPU the idle task belongs to
9256 *
9257 * NOTE: this function does not set the idle thread's NEED_RESCHED
9258 * flag, to make booting more robust.
9259 */
9260void __init init_idle(struct task_struct *idle, int cpu)
9261{
9262#ifdef CONFIG_SMP
9263 struct affinity_context ac = (struct affinity_context) {
9264 .new_mask = cpumask_of(cpu),
9265 .flags = 0,
9266 };
9267#endif
9268 struct rq *rq = cpu_rq(cpu);
9269 unsigned long flags;
9270
9271 __sched_fork(0, idle);
9272
9273 raw_spin_lock_irqsave(&idle->pi_lock, flags);
9274 raw_spin_rq_lock(rq);
9275
9276 idle->__state = TASK_RUNNING;
9277 idle->se.exec_start = sched_clock();
9278 /*
9279 * PF_KTHREAD should already be set at this point; regardless, make it
9280 * look like a proper per-CPU kthread.
9281 */
9282 idle->flags |= PF_KTHREAD | PF_NO_SETAFFINITY;
9283 kthread_set_per_cpu(idle, cpu);
9284
9285#ifdef CONFIG_SMP
9286 /*
9287 * It's possible that init_idle() gets called multiple times on a task,
9288 * in that case do_set_cpus_allowed() will not do the right thing.
9289 *
9290 * And since this is boot we can forgo the serialization.
9291 */
9292 set_cpus_allowed_common(idle, &ac);
9293#endif
9294 /*
9295 * We're having a chicken and egg problem, even though we are
9296 * holding rq->lock, the CPU isn't yet set to this CPU so the
9297 * lockdep check in task_group() will fail.
9298 *
9299 * Similar case to sched_fork(). / Alternatively we could
9300 * use task_rq_lock() here and obtain the other rq->lock.
9301 *
9302 * Silence PROVE_RCU
9303 */
9304 rcu_read_lock();
9305 __set_task_cpu(idle, cpu);
9306 rcu_read_unlock();
9307
9308 rq->idle = idle;
9309 rcu_assign_pointer(rq->curr, idle);
9310 idle->on_rq = TASK_ON_RQ_QUEUED;
9311#ifdef CONFIG_SMP
9312 idle->on_cpu = 1;
9313#endif
9314 raw_spin_rq_unlock(rq);
9315 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
9316
9317 /* Set the preempt count _outside_ the spinlocks! */
9318 init_idle_preempt_count(idle, cpu);
9319
9320 /*
9321 * The idle tasks have their own, simple scheduling class:
9322 */
9323 idle->sched_class = &idle_sched_class;
9324 ftrace_graph_init_idle_task(idle, cpu);
9325 vtime_init_idle(idle, cpu);
9326#ifdef CONFIG_SMP
9327 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
9328#endif
9329}
9330
9331#ifdef CONFIG_SMP
9332
9333int cpuset_cpumask_can_shrink(const struct cpumask *cur,
9334 const struct cpumask *trial)
9335{
9336 int ret = 1;
9337
9338 if (cpumask_empty(cur))
9339 return ret;
9340
9341 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
9342
9343 return ret;
9344}
9345
9346int task_can_attach(struct task_struct *p)
9347{
9348 int ret = 0;
9349
9350 /*
9351 * Kthreads which disallow setaffinity shouldn't be moved
9352 * to a new cpuset; we don't want to change their CPU
9353 * affinity and isolating such threads by their set of
9354 * allowed nodes is unnecessary. Thus, cpusets are not
9355 * applicable for such threads. This prevents checking for
9356 * success of set_cpus_allowed_ptr() on all attached tasks
9357 * before cpus_mask may be changed.
9358 */
9359 if (p->flags & PF_NO_SETAFFINITY)
9360 ret = -EINVAL;
9361
9362 return ret;
9363}
9364
9365bool sched_smp_initialized __read_mostly;
9366
9367#ifdef CONFIG_NUMA_BALANCING
9368/* Migrate current task p to target_cpu */
9369int migrate_task_to(struct task_struct *p, int target_cpu)
9370{
9371 struct migration_arg arg = { p, target_cpu };
9372 int curr_cpu = task_cpu(p);
9373
9374 if (curr_cpu == target_cpu)
9375 return 0;
9376
9377 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
9378 return -EINVAL;
9379
9380 /* TODO: This is not properly updating schedstats */
9381
9382 trace_sched_move_numa(p, curr_cpu, target_cpu);
9383 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
9384}
9385
9386/*
9387 * Requeue a task on a given node and accurately track the number of NUMA
9388 * tasks on the runqueues
9389 */
9390void sched_setnuma(struct task_struct *p, int nid)
9391{
9392 bool queued, running;
9393 struct rq_flags rf;
9394 struct rq *rq;
9395
9396 rq = task_rq_lock(p, &rf);
9397 queued = task_on_rq_queued(p);
9398 running = task_current(rq, p);
9399
9400 if (queued)
9401 dequeue_task(rq, p, DEQUEUE_SAVE);
9402 if (running)
9403 put_prev_task(rq, p);
9404
9405 p->numa_preferred_nid = nid;
9406
9407 if (queued)
9408 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
9409 if (running)
9410 set_next_task(rq, p);
9411 task_rq_unlock(rq, p, &rf);
9412}
9413#endif /* CONFIG_NUMA_BALANCING */
9414
9415#ifdef CONFIG_HOTPLUG_CPU
9416/*
9417 * Ensure that the idle task is using init_mm right before its CPU goes
9418 * offline.
9419 */
9420void idle_task_exit(void)
9421{
9422 struct mm_struct *mm = current->active_mm;
9423
9424 BUG_ON(cpu_online(smp_processor_id()));
9425 BUG_ON(current != this_rq()->idle);
9426
9427 if (mm != &init_mm) {
9428 switch_mm(mm, &init_mm, current);
9429 finish_arch_post_lock_switch();
9430 }
9431
9432 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
9433}
9434
9435static int __balance_push_cpu_stop(void *arg)
9436{
9437 struct task_struct *p = arg;
9438 struct rq *rq = this_rq();
9439 struct rq_flags rf;
9440 int cpu;
9441
9442 raw_spin_lock_irq(&p->pi_lock);
9443 rq_lock(rq, &rf);
9444
9445 update_rq_clock(rq);
9446
9447 if (task_rq(p) == rq && task_on_rq_queued(p)) {
9448 cpu = select_fallback_rq(rq->cpu, p);
9449 rq = __migrate_task(rq, &rf, p, cpu);
9450 }
9451
9452 rq_unlock(rq, &rf);
9453 raw_spin_unlock_irq(&p->pi_lock);
9454
9455 put_task_struct(p);
9456
9457 return 0;
9458}
9459
9460static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
9461
9462/*
9463 * Ensure we only run per-cpu kthreads once the CPU goes !active.
9464 *
9465 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
9466 * effective when the hotplug motion is down.
9467 */
9468static void balance_push(struct rq *rq)
9469{
9470 struct task_struct *push_task = rq->curr;
9471
9472 lockdep_assert_rq_held(rq);
9473
9474 /*
9475 * Ensure the thing is persistent until balance_push_set(.on = false);
9476 */
9477 rq->balance_callback = &balance_push_callback;
9478
9479 /*
9480 * Only active while going offline and when invoked on the outgoing
9481 * CPU.
9482 */
9483 if (!cpu_dying(rq->cpu) || rq != this_rq())
9484 return;
9485
9486 /*
9487 * Both the cpu-hotplug and stop task are in this case and are
9488 * required to complete the hotplug process.
9489 */
9490 if (kthread_is_per_cpu(push_task) ||
9491 is_migration_disabled(push_task)) {
9492
9493 /*
9494 * If this is the idle task on the outgoing CPU try to wake
9495 * up the hotplug control thread which might wait for the
9496 * last task to vanish. The rcuwait_active() check is
9497 * accurate here because the waiter is pinned on this CPU
9498 * and can't obviously be running in parallel.
9499 *
9500 * On RT kernels this also has to check whether there are
9501 * pinned and scheduled out tasks on the runqueue. They
9502 * need to leave the migrate disabled section first.
9503 */
9504 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
9505 rcuwait_active(&rq->hotplug_wait)) {
9506 raw_spin_rq_unlock(rq);
9507 rcuwait_wake_up(&rq->hotplug_wait);
9508 raw_spin_rq_lock(rq);
9509 }
9510 return;
9511 }
9512
9513 get_task_struct(push_task);
9514 /*
9515 * Temporarily drop rq->lock such that we can wake-up the stop task.
9516 * Both preemption and IRQs are still disabled.
9517 */
9518 preempt_disable();
9519 raw_spin_rq_unlock(rq);
9520 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
9521 this_cpu_ptr(&push_work));
9522 preempt_enable();
9523 /*
9524 * At this point need_resched() is true and we'll take the loop in
9525 * schedule(). The next pick is obviously going to be the stop task
9526 * which kthread_is_per_cpu() and will push this task away.
9527 */
9528 raw_spin_rq_lock(rq);
9529}
9530
9531static void balance_push_set(int cpu, bool on)
9532{
9533 struct rq *rq = cpu_rq(cpu);
9534 struct rq_flags rf;
9535
9536 rq_lock_irqsave(rq, &rf);
9537 if (on) {
9538 WARN_ON_ONCE(rq->balance_callback);
9539 rq->balance_callback = &balance_push_callback;
9540 } else if (rq->balance_callback == &balance_push_callback) {
9541 rq->balance_callback = NULL;
9542 }
9543 rq_unlock_irqrestore(rq, &rf);
9544}
9545
9546/*
9547 * Invoked from a CPUs hotplug control thread after the CPU has been marked
9548 * inactive. All tasks which are not per CPU kernel threads are either
9549 * pushed off this CPU now via balance_push() or placed on a different CPU
9550 * during wakeup. Wait until the CPU is quiescent.
9551 */
9552static void balance_hotplug_wait(void)
9553{
9554 struct rq *rq = this_rq();
9555
9556 rcuwait_wait_event(&rq->hotplug_wait,
9557 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
9558 TASK_UNINTERRUPTIBLE);
9559}
9560
9561#else
9562
9563static inline void balance_push(struct rq *rq)
9564{
9565}
9566
9567static inline void balance_push_set(int cpu, bool on)
9568{
9569}
9570
9571static inline void balance_hotplug_wait(void)
9572{
9573}
9574
9575#endif /* CONFIG_HOTPLUG_CPU */
9576
9577void set_rq_online(struct rq *rq)
9578{
9579 if (!rq->online) {
9580 const struct sched_class *class;
9581
9582 cpumask_set_cpu(rq->cpu, rq->rd->online);
9583 rq->online = 1;
9584
9585 for_each_class(class) {
9586 if (class->rq_online)
9587 class->rq_online(rq);
9588 }
9589 }
9590}
9591
9592void set_rq_offline(struct rq *rq)
9593{
9594 if (rq->online) {
9595 const struct sched_class *class;
9596
9597 update_rq_clock(rq);
9598 for_each_class(class) {
9599 if (class->rq_offline)
9600 class->rq_offline(rq);
9601 }
9602
9603 cpumask_clear_cpu(rq->cpu, rq->rd->online);
9604 rq->online = 0;
9605 }
9606}
9607
9608/*
9609 * used to mark begin/end of suspend/resume:
9610 */
9611static int num_cpus_frozen;
9612
9613/*
9614 * Update cpusets according to cpu_active mask. If cpusets are
9615 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9616 * around partition_sched_domains().
9617 *
9618 * If we come here as part of a suspend/resume, don't touch cpusets because we
9619 * want to restore it back to its original state upon resume anyway.
9620 */
9621static void cpuset_cpu_active(void)
9622{
9623 if (cpuhp_tasks_frozen) {
9624 /*
9625 * num_cpus_frozen tracks how many CPUs are involved in suspend
9626 * resume sequence. As long as this is not the last online
9627 * operation in the resume sequence, just build a single sched
9628 * domain, ignoring cpusets.
9629 */
9630 partition_sched_domains(1, NULL, NULL);
9631 if (--num_cpus_frozen)
9632 return;
9633 /*
9634 * This is the last CPU online operation. So fall through and
9635 * restore the original sched domains by considering the
9636 * cpuset configurations.
9637 */
9638 cpuset_force_rebuild();
9639 }
9640 cpuset_update_active_cpus();
9641}
9642
9643static int cpuset_cpu_inactive(unsigned int cpu)
9644{
9645 if (!cpuhp_tasks_frozen) {
9646 int ret = dl_bw_check_overflow(cpu);
9647
9648 if (ret)
9649 return ret;
9650 cpuset_update_active_cpus();
9651 } else {
9652 num_cpus_frozen++;
9653 partition_sched_domains(1, NULL, NULL);
9654 }
9655 return 0;
9656}
9657
9658int sched_cpu_activate(unsigned int cpu)
9659{
9660 struct rq *rq = cpu_rq(cpu);
9661 struct rq_flags rf;
9662
9663 /*
9664 * Clear the balance_push callback and prepare to schedule
9665 * regular tasks.
9666 */
9667 balance_push_set(cpu, false);
9668
9669#ifdef CONFIG_SCHED_SMT
9670 /*
9671 * When going up, increment the number of cores with SMT present.
9672 */
9673 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9674 static_branch_inc_cpuslocked(&sched_smt_present);
9675#endif
9676 set_cpu_active(cpu, true);
9677
9678 if (sched_smp_initialized) {
9679 sched_update_numa(cpu, true);
9680 sched_domains_numa_masks_set(cpu);
9681 cpuset_cpu_active();
9682 }
9683
9684 /*
9685 * Put the rq online, if not already. This happens:
9686 *
9687 * 1) In the early boot process, because we build the real domains
9688 * after all CPUs have been brought up.
9689 *
9690 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9691 * domains.
9692 */
9693 rq_lock_irqsave(rq, &rf);
9694 if (rq->rd) {
9695 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9696 set_rq_online(rq);
9697 }
9698 rq_unlock_irqrestore(rq, &rf);
9699
9700 return 0;
9701}
9702
9703int sched_cpu_deactivate(unsigned int cpu)
9704{
9705 struct rq *rq = cpu_rq(cpu);
9706 struct rq_flags rf;
9707 int ret;
9708
9709 /*
9710 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9711 * load balancing when not active
9712 */
9713 nohz_balance_exit_idle(rq);
9714
9715 set_cpu_active(cpu, false);
9716
9717 /*
9718 * From this point forward, this CPU will refuse to run any task that
9719 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9720 * push those tasks away until this gets cleared, see
9721 * sched_cpu_dying().
9722 */
9723 balance_push_set(cpu, true);
9724
9725 /*
9726 * We've cleared cpu_active_mask / set balance_push, wait for all
9727 * preempt-disabled and RCU users of this state to go away such that
9728 * all new such users will observe it.
9729 *
9730 * Specifically, we rely on ttwu to no longer target this CPU, see
9731 * ttwu_queue_cond() and is_cpu_allowed().
9732 *
9733 * Do sync before park smpboot threads to take care the rcu boost case.
9734 */
9735 synchronize_rcu();
9736
9737 rq_lock_irqsave(rq, &rf);
9738 if (rq->rd) {
9739 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9740 set_rq_offline(rq);
9741 }
9742 rq_unlock_irqrestore(rq, &rf);
9743
9744#ifdef CONFIG_SCHED_SMT
9745 /*
9746 * When going down, decrement the number of cores with SMT present.
9747 */
9748 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9749 static_branch_dec_cpuslocked(&sched_smt_present);
9750
9751 sched_core_cpu_deactivate(cpu);
9752#endif
9753
9754 if (!sched_smp_initialized)
9755 return 0;
9756
9757 sched_update_numa(cpu, false);
9758 ret = cpuset_cpu_inactive(cpu);
9759 if (ret) {
9760 balance_push_set(cpu, false);
9761 set_cpu_active(cpu, true);
9762 sched_update_numa(cpu, true);
9763 return ret;
9764 }
9765 sched_domains_numa_masks_clear(cpu);
9766 return 0;
9767}
9768
9769static void sched_rq_cpu_starting(unsigned int cpu)
9770{
9771 struct rq *rq = cpu_rq(cpu);
9772
9773 rq->calc_load_update = calc_load_update;
9774 update_max_interval();
9775}
9776
9777int sched_cpu_starting(unsigned int cpu)
9778{
9779 sched_core_cpu_starting(cpu);
9780 sched_rq_cpu_starting(cpu);
9781 sched_tick_start(cpu);
9782 return 0;
9783}
9784
9785#ifdef CONFIG_HOTPLUG_CPU
9786
9787/*
9788 * Invoked immediately before the stopper thread is invoked to bring the
9789 * CPU down completely. At this point all per CPU kthreads except the
9790 * hotplug thread (current) and the stopper thread (inactive) have been
9791 * either parked or have been unbound from the outgoing CPU. Ensure that
9792 * any of those which might be on the way out are gone.
9793 *
9794 * If after this point a bound task is being woken on this CPU then the
9795 * responsible hotplug callback has failed to do it's job.
9796 * sched_cpu_dying() will catch it with the appropriate fireworks.
9797 */
9798int sched_cpu_wait_empty(unsigned int cpu)
9799{
9800 balance_hotplug_wait();
9801 return 0;
9802}
9803
9804/*
9805 * Since this CPU is going 'away' for a while, fold any nr_active delta we
9806 * might have. Called from the CPU stopper task after ensuring that the
9807 * stopper is the last running task on the CPU, so nr_active count is
9808 * stable. We need to take the teardown thread which is calling this into
9809 * account, so we hand in adjust = 1 to the load calculation.
9810 *
9811 * Also see the comment "Global load-average calculations".
9812 */
9813static void calc_load_migrate(struct rq *rq)
9814{
9815 long delta = calc_load_fold_active(rq, 1);
9816
9817 if (delta)
9818 atomic_long_add(delta, &calc_load_tasks);
9819}
9820
9821static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9822{
9823 struct task_struct *g, *p;
9824 int cpu = cpu_of(rq);
9825
9826 lockdep_assert_rq_held(rq);
9827
9828 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9829 for_each_process_thread(g, p) {
9830 if (task_cpu(p) != cpu)
9831 continue;
9832
9833 if (!task_on_rq_queued(p))
9834 continue;
9835
9836 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9837 }
9838}
9839
9840int sched_cpu_dying(unsigned int cpu)
9841{
9842 struct rq *rq = cpu_rq(cpu);
9843 struct rq_flags rf;
9844
9845 /* Handle pending wakeups and then migrate everything off */
9846 sched_tick_stop(cpu);
9847
9848 rq_lock_irqsave(rq, &rf);
9849 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9850 WARN(true, "Dying CPU not properly vacated!");
9851 dump_rq_tasks(rq, KERN_WARNING);
9852 }
9853 rq_unlock_irqrestore(rq, &rf);
9854
9855 calc_load_migrate(rq);
9856 update_max_interval();
9857 hrtick_clear(rq);
9858 sched_core_cpu_dying(cpu);
9859 return 0;
9860}
9861#endif
9862
9863void __init sched_init_smp(void)
9864{
9865 sched_init_numa(NUMA_NO_NODE);
9866
9867 /*
9868 * There's no userspace yet to cause hotplug operations; hence all the
9869 * CPU masks are stable and all blatant races in the below code cannot
9870 * happen.
9871 */
9872 mutex_lock(&sched_domains_mutex);
9873 sched_init_domains(cpu_active_mask);
9874 mutex_unlock(&sched_domains_mutex);
9875
9876 /* Move init over to a non-isolated CPU */
9877 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0)
9878 BUG();
9879 current->flags &= ~PF_NO_SETAFFINITY;
9880 sched_init_granularity();
9881
9882 init_sched_rt_class();
9883 init_sched_dl_class();
9884
9885 sched_smp_initialized = true;
9886}
9887
9888static int __init migration_init(void)
9889{
9890 sched_cpu_starting(smp_processor_id());
9891 return 0;
9892}
9893early_initcall(migration_init);
9894
9895#else
9896void __init sched_init_smp(void)
9897{
9898 sched_init_granularity();
9899}
9900#endif /* CONFIG_SMP */
9901
9902int in_sched_functions(unsigned long addr)
9903{
9904 return in_lock_functions(addr) ||
9905 (addr >= (unsigned long)__sched_text_start
9906 && addr < (unsigned long)__sched_text_end);
9907}
9908
9909#ifdef CONFIG_CGROUP_SCHED
9910/*
9911 * Default task group.
9912 * Every task in system belongs to this group at bootup.
9913 */
9914struct task_group root_task_group;
9915LIST_HEAD(task_groups);
9916
9917/* Cacheline aligned slab cache for task_group */
9918static struct kmem_cache *task_group_cache __ro_after_init;
9919#endif
9920
9921void __init sched_init(void)
9922{
9923 unsigned long ptr = 0;
9924 int i;
9925
9926 /* Make sure the linker didn't screw up */
9927 BUG_ON(&idle_sched_class != &fair_sched_class + 1 ||
9928 &fair_sched_class != &rt_sched_class + 1 ||
9929 &rt_sched_class != &dl_sched_class + 1);
9930#ifdef CONFIG_SMP
9931 BUG_ON(&dl_sched_class != &stop_sched_class + 1);
9932#endif
9933
9934 wait_bit_init();
9935
9936#ifdef CONFIG_FAIR_GROUP_SCHED
9937 ptr += 2 * nr_cpu_ids * sizeof(void **);
9938#endif
9939#ifdef CONFIG_RT_GROUP_SCHED
9940 ptr += 2 * nr_cpu_ids * sizeof(void **);
9941#endif
9942 if (ptr) {
9943 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9944
9945#ifdef CONFIG_FAIR_GROUP_SCHED
9946 root_task_group.se = (struct sched_entity **)ptr;
9947 ptr += nr_cpu_ids * sizeof(void **);
9948
9949 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9950 ptr += nr_cpu_ids * sizeof(void **);
9951
9952 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9953 init_cfs_bandwidth(&root_task_group.cfs_bandwidth, NULL);
9954#endif /* CONFIG_FAIR_GROUP_SCHED */
9955#ifdef CONFIG_RT_GROUP_SCHED
9956 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9957 ptr += nr_cpu_ids * sizeof(void **);
9958
9959 root_task_group.rt_rq = (struct rt_rq **)ptr;
9960 ptr += nr_cpu_ids * sizeof(void **);
9961
9962#endif /* CONFIG_RT_GROUP_SCHED */
9963 }
9964
9965 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9966
9967#ifdef CONFIG_SMP
9968 init_defrootdomain();
9969#endif
9970
9971#ifdef CONFIG_RT_GROUP_SCHED
9972 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9973 global_rt_period(), global_rt_runtime());
9974#endif /* CONFIG_RT_GROUP_SCHED */
9975
9976#ifdef CONFIG_CGROUP_SCHED
9977 task_group_cache = KMEM_CACHE(task_group, 0);
9978
9979 list_add(&root_task_group.list, &task_groups);
9980 INIT_LIST_HEAD(&root_task_group.children);
9981 INIT_LIST_HEAD(&root_task_group.siblings);
9982 autogroup_init(&init_task);
9983#endif /* CONFIG_CGROUP_SCHED */
9984
9985 for_each_possible_cpu(i) {
9986 struct rq *rq;
9987
9988 rq = cpu_rq(i);
9989 raw_spin_lock_init(&rq->__lock);
9990 rq->nr_running = 0;
9991 rq->calc_load_active = 0;
9992 rq->calc_load_update = jiffies + LOAD_FREQ;
9993 init_cfs_rq(&rq->cfs);
9994 init_rt_rq(&rq->rt);
9995 init_dl_rq(&rq->dl);
9996#ifdef CONFIG_FAIR_GROUP_SCHED
9997 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9998 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9999 /*
10000 * How much CPU bandwidth does root_task_group get?
10001 *
10002 * In case of task-groups formed thr' the cgroup filesystem, it
10003 * gets 100% of the CPU resources in the system. This overall
10004 * system CPU resource is divided among the tasks of
10005 * root_task_group and its child task-groups in a fair manner,
10006 * based on each entity's (task or task-group's) weight
10007 * (se->load.weight).
10008 *
10009 * In other words, if root_task_group has 10 tasks of weight
10010 * 1024) and two child groups A0 and A1 (of weight 1024 each),
10011 * then A0's share of the CPU resource is:
10012 *
10013 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
10014 *
10015 * We achieve this by letting root_task_group's tasks sit
10016 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
10017 */
10018 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
10019#endif /* CONFIG_FAIR_GROUP_SCHED */
10020
10021 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
10022#ifdef CONFIG_RT_GROUP_SCHED
10023 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
10024#endif
10025#ifdef CONFIG_SMP
10026 rq->sd = NULL;
10027 rq->rd = NULL;
10028 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
10029 rq->balance_callback = &balance_push_callback;
10030 rq->active_balance = 0;
10031 rq->next_balance = jiffies;
10032 rq->push_cpu = 0;
10033 rq->cpu = i;
10034 rq->online = 0;
10035 rq->idle_stamp = 0;
10036 rq->avg_idle = 2*sysctl_sched_migration_cost;
10037 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
10038
10039 INIT_LIST_HEAD(&rq->cfs_tasks);
10040
10041 rq_attach_root(rq, &def_root_domain);
10042#ifdef CONFIG_NO_HZ_COMMON
10043 rq->last_blocked_load_update_tick = jiffies;
10044 atomic_set(&rq->nohz_flags, 0);
10045
10046 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
10047#endif
10048#ifdef CONFIG_HOTPLUG_CPU
10049 rcuwait_init(&rq->hotplug_wait);
10050#endif
10051#endif /* CONFIG_SMP */
10052 hrtick_rq_init(rq);
10053 atomic_set(&rq->nr_iowait, 0);
10054
10055#ifdef CONFIG_SCHED_CORE
10056 rq->core = rq;
10057 rq->core_pick = NULL;
10058 rq->core_enabled = 0;
10059 rq->core_tree = RB_ROOT;
10060 rq->core_forceidle_count = 0;
10061 rq->core_forceidle_occupation = 0;
10062 rq->core_forceidle_start = 0;
10063
10064 rq->core_cookie = 0UL;
10065#endif
10066 zalloc_cpumask_var_node(&rq->scratch_mask, GFP_KERNEL, cpu_to_node(i));
10067 }
10068
10069 set_load_weight(&init_task, false);
10070
10071 /*
10072 * The boot idle thread does lazy MMU switching as well:
10073 */
10074 mmgrab_lazy_tlb(&init_mm);
10075 enter_lazy_tlb(&init_mm, current);
10076
10077 /*
10078 * The idle task doesn't need the kthread struct to function, but it
10079 * is dressed up as a per-CPU kthread and thus needs to play the part
10080 * if we want to avoid special-casing it in code that deals with per-CPU
10081 * kthreads.
10082 */
10083 WARN_ON(!set_kthread_struct(current));
10084
10085 /*
10086 * Make us the idle thread. Technically, schedule() should not be
10087 * called from this thread, however somewhere below it might be,
10088 * but because we are the idle thread, we just pick up running again
10089 * when this runqueue becomes "idle".
10090 */
10091 init_idle(current, smp_processor_id());
10092
10093 calc_load_update = jiffies + LOAD_FREQ;
10094
10095#ifdef CONFIG_SMP
10096 idle_thread_set_boot_cpu();
10097 balance_push_set(smp_processor_id(), false);
10098#endif
10099 init_sched_fair_class();
10100
10101 psi_init();
10102
10103 init_uclamp();
10104
10105 preempt_dynamic_init();
10106
10107 scheduler_running = 1;
10108}
10109
10110#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
10111
10112void __might_sleep(const char *file, int line)
10113{
10114 unsigned int state = get_current_state();
10115 /*
10116 * Blocking primitives will set (and therefore destroy) current->state,
10117 * since we will exit with TASK_RUNNING make sure we enter with it,
10118 * otherwise we will destroy state.
10119 */
10120 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
10121 "do not call blocking ops when !TASK_RUNNING; "
10122 "state=%x set at [<%p>] %pS\n", state,
10123 (void *)current->task_state_change,
10124 (void *)current->task_state_change);
10125
10126 __might_resched(file, line, 0);
10127}
10128EXPORT_SYMBOL(__might_sleep);
10129
10130static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
10131{
10132 if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
10133 return;
10134
10135 if (preempt_count() == preempt_offset)
10136 return;
10137
10138 pr_err("Preemption disabled at:");
10139 print_ip_sym(KERN_ERR, ip);
10140}
10141
10142static inline bool resched_offsets_ok(unsigned int offsets)
10143{
10144 unsigned int nested = preempt_count();
10145
10146 nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
10147
10148 return nested == offsets;
10149}
10150
10151void __might_resched(const char *file, int line, unsigned int offsets)
10152{
10153 /* Ratelimiting timestamp: */
10154 static unsigned long prev_jiffy;
10155
10156 unsigned long preempt_disable_ip;
10157
10158 /* WARN_ON_ONCE() by default, no rate limit required: */
10159 rcu_sleep_check();
10160
10161 if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
10162 !is_idle_task(current) && !current->non_block_count) ||
10163 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
10164 oops_in_progress)
10165 return;
10166
10167 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10168 return;
10169 prev_jiffy = jiffies;
10170
10171 /* Save this before calling printk(), since that will clobber it: */
10172 preempt_disable_ip = get_preempt_disable_ip(current);
10173
10174 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
10175 file, line);
10176 pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
10177 in_atomic(), irqs_disabled(), current->non_block_count,
10178 current->pid, current->comm);
10179 pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
10180 offsets & MIGHT_RESCHED_PREEMPT_MASK);
10181
10182 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
10183 pr_err("RCU nest depth: %d, expected: %u\n",
10184 rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
10185 }
10186
10187 if (task_stack_end_corrupted(current))
10188 pr_emerg("Thread overran stack, or stack corrupted\n");
10189
10190 debug_show_held_locks(current);
10191 if (irqs_disabled())
10192 print_irqtrace_events(current);
10193
10194 print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
10195 preempt_disable_ip);
10196
10197 dump_stack();
10198 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10199}
10200EXPORT_SYMBOL(__might_resched);
10201
10202void __cant_sleep(const char *file, int line, int preempt_offset)
10203{
10204 static unsigned long prev_jiffy;
10205
10206 if (irqs_disabled())
10207 return;
10208
10209 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10210 return;
10211
10212 if (preempt_count() > preempt_offset)
10213 return;
10214
10215 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10216 return;
10217 prev_jiffy = jiffies;
10218
10219 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
10220 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
10221 in_atomic(), irqs_disabled(),
10222 current->pid, current->comm);
10223
10224 debug_show_held_locks(current);
10225 dump_stack();
10226 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10227}
10228EXPORT_SYMBOL_GPL(__cant_sleep);
10229
10230#ifdef CONFIG_SMP
10231void __cant_migrate(const char *file, int line)
10232{
10233 static unsigned long prev_jiffy;
10234
10235 if (irqs_disabled())
10236 return;
10237
10238 if (is_migration_disabled(current))
10239 return;
10240
10241 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10242 return;
10243
10244 if (preempt_count() > 0)
10245 return;
10246
10247 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10248 return;
10249 prev_jiffy = jiffies;
10250
10251 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
10252 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
10253 in_atomic(), irqs_disabled(), is_migration_disabled(current),
10254 current->pid, current->comm);
10255
10256 debug_show_held_locks(current);
10257 dump_stack();
10258 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10259}
10260EXPORT_SYMBOL_GPL(__cant_migrate);
10261#endif
10262#endif
10263
10264#ifdef CONFIG_MAGIC_SYSRQ
10265void normalize_rt_tasks(void)
10266{
10267 struct task_struct *g, *p;
10268 struct sched_attr attr = {
10269 .sched_policy = SCHED_NORMAL,
10270 };
10271
10272 read_lock(&tasklist_lock);
10273 for_each_process_thread(g, p) {
10274 /*
10275 * Only normalize user tasks:
10276 */
10277 if (p->flags & PF_KTHREAD)
10278 continue;
10279
10280 p->se.exec_start = 0;
10281 schedstat_set(p->stats.wait_start, 0);
10282 schedstat_set(p->stats.sleep_start, 0);
10283 schedstat_set(p->stats.block_start, 0);
10284
10285 if (!dl_task(p) && !rt_task(p)) {
10286 /*
10287 * Renice negative nice level userspace
10288 * tasks back to 0:
10289 */
10290 if (task_nice(p) < 0)
10291 set_user_nice(p, 0);
10292 continue;
10293 }
10294
10295 __sched_setscheduler(p, &attr, false, false);
10296 }
10297 read_unlock(&tasklist_lock);
10298}
10299
10300#endif /* CONFIG_MAGIC_SYSRQ */
10301
10302#if defined(CONFIG_KGDB_KDB)
10303/*
10304 * These functions are only useful for kdb.
10305 *
10306 * They can only be called when the whole system has been
10307 * stopped - every CPU needs to be quiescent, and no scheduling
10308 * activity can take place. Using them for anything else would
10309 * be a serious bug, and as a result, they aren't even visible
10310 * under any other configuration.
10311 */
10312
10313/**
10314 * curr_task - return the current task for a given CPU.
10315 * @cpu: the processor in question.
10316 *
10317 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10318 *
10319 * Return: The current task for @cpu.
10320 */
10321struct task_struct *curr_task(int cpu)
10322{
10323 return cpu_curr(cpu);
10324}
10325
10326#endif /* defined(CONFIG_KGDB_KDB) */
10327
10328#ifdef CONFIG_CGROUP_SCHED
10329/* task_group_lock serializes the addition/removal of task groups */
10330static DEFINE_SPINLOCK(task_group_lock);
10331
10332static inline void alloc_uclamp_sched_group(struct task_group *tg,
10333 struct task_group *parent)
10334{
10335#ifdef CONFIG_UCLAMP_TASK_GROUP
10336 enum uclamp_id clamp_id;
10337
10338 for_each_clamp_id(clamp_id) {
10339 uclamp_se_set(&tg->uclamp_req[clamp_id],
10340 uclamp_none(clamp_id), false);
10341 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
10342 }
10343#endif
10344}
10345
10346static void sched_free_group(struct task_group *tg)
10347{
10348 free_fair_sched_group(tg);
10349 free_rt_sched_group(tg);
10350 autogroup_free(tg);
10351 kmem_cache_free(task_group_cache, tg);
10352}
10353
10354static void sched_free_group_rcu(struct rcu_head *rcu)
10355{
10356 sched_free_group(container_of(rcu, struct task_group, rcu));
10357}
10358
10359static void sched_unregister_group(struct task_group *tg)
10360{
10361 unregister_fair_sched_group(tg);
10362 unregister_rt_sched_group(tg);
10363 /*
10364 * We have to wait for yet another RCU grace period to expire, as
10365 * print_cfs_stats() might run concurrently.
10366 */
10367 call_rcu(&tg->rcu, sched_free_group_rcu);
10368}
10369
10370/* allocate runqueue etc for a new task group */
10371struct task_group *sched_create_group(struct task_group *parent)
10372{
10373 struct task_group *tg;
10374
10375 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
10376 if (!tg)
10377 return ERR_PTR(-ENOMEM);
10378
10379 if (!alloc_fair_sched_group(tg, parent))
10380 goto err;
10381
10382 if (!alloc_rt_sched_group(tg, parent))
10383 goto err;
10384
10385 alloc_uclamp_sched_group(tg, parent);
10386
10387 return tg;
10388
10389err:
10390 sched_free_group(tg);
10391 return ERR_PTR(-ENOMEM);
10392}
10393
10394void sched_online_group(struct task_group *tg, struct task_group *parent)
10395{
10396 unsigned long flags;
10397
10398 spin_lock_irqsave(&task_group_lock, flags);
10399 list_add_rcu(&tg->list, &task_groups);
10400
10401 /* Root should already exist: */
10402 WARN_ON(!parent);
10403
10404 tg->parent = parent;
10405 INIT_LIST_HEAD(&tg->children);
10406 list_add_rcu(&tg->siblings, &parent->children);
10407 spin_unlock_irqrestore(&task_group_lock, flags);
10408
10409 online_fair_sched_group(tg);
10410}
10411
10412/* rcu callback to free various structures associated with a task group */
10413static void sched_unregister_group_rcu(struct rcu_head *rhp)
10414{
10415 /* Now it should be safe to free those cfs_rqs: */
10416 sched_unregister_group(container_of(rhp, struct task_group, rcu));
10417}
10418
10419void sched_destroy_group(struct task_group *tg)
10420{
10421 /* Wait for possible concurrent references to cfs_rqs complete: */
10422 call_rcu(&tg->rcu, sched_unregister_group_rcu);
10423}
10424
10425void sched_release_group(struct task_group *tg)
10426{
10427 unsigned long flags;
10428
10429 /*
10430 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
10431 * sched_cfs_period_timer()).
10432 *
10433 * For this to be effective, we have to wait for all pending users of
10434 * this task group to leave their RCU critical section to ensure no new
10435 * user will see our dying task group any more. Specifically ensure
10436 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10437 *
10438 * We therefore defer calling unregister_fair_sched_group() to
10439 * sched_unregister_group() which is guarantied to get called only after the
10440 * current RCU grace period has expired.
10441 */
10442 spin_lock_irqsave(&task_group_lock, flags);
10443 list_del_rcu(&tg->list);
10444 list_del_rcu(&tg->siblings);
10445 spin_unlock_irqrestore(&task_group_lock, flags);
10446}
10447
10448static struct task_group *sched_get_task_group(struct task_struct *tsk)
10449{
10450 struct task_group *tg;
10451
10452 /*
10453 * All callers are synchronized by task_rq_lock(); we do not use RCU
10454 * which is pointless here. Thus, we pass "true" to task_css_check()
10455 * to prevent lockdep warnings.
10456 */
10457 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10458 struct task_group, css);
10459 tg = autogroup_task_group(tsk, tg);
10460
10461 return tg;
10462}
10463
10464static void sched_change_group(struct task_struct *tsk, struct task_group *group)
10465{
10466 tsk->sched_task_group = group;
10467
10468#ifdef CONFIG_FAIR_GROUP_SCHED
10469 if (tsk->sched_class->task_change_group)
10470 tsk->sched_class->task_change_group(tsk);
10471 else
10472#endif
10473 set_task_rq(tsk, task_cpu(tsk));
10474}
10475
10476/*
10477 * Change task's runqueue when it moves between groups.
10478 *
10479 * The caller of this function should have put the task in its new group by
10480 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10481 * its new group.
10482 */
10483void sched_move_task(struct task_struct *tsk)
10484{
10485 int queued, running, queue_flags =
10486 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
10487 struct task_group *group;
10488 struct rq *rq;
10489
10490 CLASS(task_rq_lock, rq_guard)(tsk);
10491 rq = rq_guard.rq;
10492
10493 /*
10494 * Esp. with SCHED_AUTOGROUP enabled it is possible to get superfluous
10495 * group changes.
10496 */
10497 group = sched_get_task_group(tsk);
10498 if (group == tsk->sched_task_group)
10499 return;
10500
10501 update_rq_clock(rq);
10502
10503 running = task_current(rq, tsk);
10504 queued = task_on_rq_queued(tsk);
10505
10506 if (queued)
10507 dequeue_task(rq, tsk, queue_flags);
10508 if (running)
10509 put_prev_task(rq, tsk);
10510
10511 sched_change_group(tsk, group);
10512
10513 if (queued)
10514 enqueue_task(rq, tsk, queue_flags);
10515 if (running) {
10516 set_next_task(rq, tsk);
10517 /*
10518 * After changing group, the running task may have joined a
10519 * throttled one but it's still the running task. Trigger a
10520 * resched to make sure that task can still run.
10521 */
10522 resched_curr(rq);
10523 }
10524}
10525
10526static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
10527{
10528 return css ? container_of(css, struct task_group, css) : NULL;
10529}
10530
10531static struct cgroup_subsys_state *
10532cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
10533{
10534 struct task_group *parent = css_tg(parent_css);
10535 struct task_group *tg;
10536
10537 if (!parent) {
10538 /* This is early initialization for the top cgroup */
10539 return &root_task_group.css;
10540 }
10541
10542 tg = sched_create_group(parent);
10543 if (IS_ERR(tg))
10544 return ERR_PTR(-ENOMEM);
10545
10546 return &tg->css;
10547}
10548
10549/* Expose task group only after completing cgroup initialization */
10550static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10551{
10552 struct task_group *tg = css_tg(css);
10553 struct task_group *parent = css_tg(css->parent);
10554
10555 if (parent)
10556 sched_online_group(tg, parent);
10557
10558#ifdef CONFIG_UCLAMP_TASK_GROUP
10559 /* Propagate the effective uclamp value for the new group */
10560 guard(mutex)(&uclamp_mutex);
10561 guard(rcu)();
10562 cpu_util_update_eff(css);
10563#endif
10564
10565 return 0;
10566}
10567
10568static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
10569{
10570 struct task_group *tg = css_tg(css);
10571
10572 sched_release_group(tg);
10573}
10574
10575static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
10576{
10577 struct task_group *tg = css_tg(css);
10578
10579 /*
10580 * Relies on the RCU grace period between css_released() and this.
10581 */
10582 sched_unregister_group(tg);
10583}
10584
10585#ifdef CONFIG_RT_GROUP_SCHED
10586static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10587{
10588 struct task_struct *task;
10589 struct cgroup_subsys_state *css;
10590
10591 cgroup_taskset_for_each(task, css, tset) {
10592 if (!sched_rt_can_attach(css_tg(css), task))
10593 return -EINVAL;
10594 }
10595 return 0;
10596}
10597#endif
10598
10599static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10600{
10601 struct task_struct *task;
10602 struct cgroup_subsys_state *css;
10603
10604 cgroup_taskset_for_each(task, css, tset)
10605 sched_move_task(task);
10606}
10607
10608#ifdef CONFIG_UCLAMP_TASK_GROUP
10609static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10610{
10611 struct cgroup_subsys_state *top_css = css;
10612 struct uclamp_se *uc_parent = NULL;
10613 struct uclamp_se *uc_se = NULL;
10614 unsigned int eff[UCLAMP_CNT];
10615 enum uclamp_id clamp_id;
10616 unsigned int clamps;
10617
10618 lockdep_assert_held(&uclamp_mutex);
10619 SCHED_WARN_ON(!rcu_read_lock_held());
10620
10621 css_for_each_descendant_pre(css, top_css) {
10622 uc_parent = css_tg(css)->parent
10623 ? css_tg(css)->parent->uclamp : NULL;
10624
10625 for_each_clamp_id(clamp_id) {
10626 /* Assume effective clamps matches requested clamps */
10627 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10628 /* Cap effective clamps with parent's effective clamps */
10629 if (uc_parent &&
10630 eff[clamp_id] > uc_parent[clamp_id].value) {
10631 eff[clamp_id] = uc_parent[clamp_id].value;
10632 }
10633 }
10634 /* Ensure protection is always capped by limit */
10635 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10636
10637 /* Propagate most restrictive effective clamps */
10638 clamps = 0x0;
10639 uc_se = css_tg(css)->uclamp;
10640 for_each_clamp_id(clamp_id) {
10641 if (eff[clamp_id] == uc_se[clamp_id].value)
10642 continue;
10643 uc_se[clamp_id].value = eff[clamp_id];
10644 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10645 clamps |= (0x1 << clamp_id);
10646 }
10647 if (!clamps) {
10648 css = css_rightmost_descendant(css);
10649 continue;
10650 }
10651
10652 /* Immediately update descendants RUNNABLE tasks */
10653 uclamp_update_active_tasks(css);
10654 }
10655}
10656
10657/*
10658 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10659 * C expression. Since there is no way to convert a macro argument (N) into a
10660 * character constant, use two levels of macros.
10661 */
10662#define _POW10(exp) ((unsigned int)1e##exp)
10663#define POW10(exp) _POW10(exp)
10664
10665struct uclamp_request {
10666#define UCLAMP_PERCENT_SHIFT 2
10667#define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10668 s64 percent;
10669 u64 util;
10670 int ret;
10671};
10672
10673static inline struct uclamp_request
10674capacity_from_percent(char *buf)
10675{
10676 struct uclamp_request req = {
10677 .percent = UCLAMP_PERCENT_SCALE,
10678 .util = SCHED_CAPACITY_SCALE,
10679 .ret = 0,
10680 };
10681
10682 buf = strim(buf);
10683 if (strcmp(buf, "max")) {
10684 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10685 &req.percent);
10686 if (req.ret)
10687 return req;
10688 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10689 req.ret = -ERANGE;
10690 return req;
10691 }
10692
10693 req.util = req.percent << SCHED_CAPACITY_SHIFT;
10694 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10695 }
10696
10697 return req;
10698}
10699
10700static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10701 size_t nbytes, loff_t off,
10702 enum uclamp_id clamp_id)
10703{
10704 struct uclamp_request req;
10705 struct task_group *tg;
10706
10707 req = capacity_from_percent(buf);
10708 if (req.ret)
10709 return req.ret;
10710
10711 static_branch_enable(&sched_uclamp_used);
10712
10713 guard(mutex)(&uclamp_mutex);
10714 guard(rcu)();
10715
10716 tg = css_tg(of_css(of));
10717 if (tg->uclamp_req[clamp_id].value != req.util)
10718 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10719
10720 /*
10721 * Because of not recoverable conversion rounding we keep track of the
10722 * exact requested value
10723 */
10724 tg->uclamp_pct[clamp_id] = req.percent;
10725
10726 /* Update effective clamps to track the most restrictive value */
10727 cpu_util_update_eff(of_css(of));
10728
10729 return nbytes;
10730}
10731
10732static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10733 char *buf, size_t nbytes,
10734 loff_t off)
10735{
10736 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10737}
10738
10739static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10740 char *buf, size_t nbytes,
10741 loff_t off)
10742{
10743 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10744}
10745
10746static inline void cpu_uclamp_print(struct seq_file *sf,
10747 enum uclamp_id clamp_id)
10748{
10749 struct task_group *tg;
10750 u64 util_clamp;
10751 u64 percent;
10752 u32 rem;
10753
10754 scoped_guard (rcu) {
10755 tg = css_tg(seq_css(sf));
10756 util_clamp = tg->uclamp_req[clamp_id].value;
10757 }
10758
10759 if (util_clamp == SCHED_CAPACITY_SCALE) {
10760 seq_puts(sf, "max\n");
10761 return;
10762 }
10763
10764 percent = tg->uclamp_pct[clamp_id];
10765 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10766 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10767}
10768
10769static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10770{
10771 cpu_uclamp_print(sf, UCLAMP_MIN);
10772 return 0;
10773}
10774
10775static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10776{
10777 cpu_uclamp_print(sf, UCLAMP_MAX);
10778 return 0;
10779}
10780#endif /* CONFIG_UCLAMP_TASK_GROUP */
10781
10782#ifdef CONFIG_FAIR_GROUP_SCHED
10783static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10784 struct cftype *cftype, u64 shareval)
10785{
10786 if (shareval > scale_load_down(ULONG_MAX))
10787 shareval = MAX_SHARES;
10788 return sched_group_set_shares(css_tg(css), scale_load(shareval));
10789}
10790
10791static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10792 struct cftype *cft)
10793{
10794 struct task_group *tg = css_tg(css);
10795
10796 return (u64) scale_load_down(tg->shares);
10797}
10798
10799#ifdef CONFIG_CFS_BANDWIDTH
10800static DEFINE_MUTEX(cfs_constraints_mutex);
10801
10802const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10803static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10804/* More than 203 days if BW_SHIFT equals 20. */
10805static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10806
10807static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10808
10809static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10810 u64 burst)
10811{
10812 int i, ret = 0, runtime_enabled, runtime_was_enabled;
10813 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10814
10815 if (tg == &root_task_group)
10816 return -EINVAL;
10817
10818 /*
10819 * Ensure we have at some amount of bandwidth every period. This is
10820 * to prevent reaching a state of large arrears when throttled via
10821 * entity_tick() resulting in prolonged exit starvation.
10822 */
10823 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10824 return -EINVAL;
10825
10826 /*
10827 * Likewise, bound things on the other side by preventing insane quota
10828 * periods. This also allows us to normalize in computing quota
10829 * feasibility.
10830 */
10831 if (period > max_cfs_quota_period)
10832 return -EINVAL;
10833
10834 /*
10835 * Bound quota to defend quota against overflow during bandwidth shift.
10836 */
10837 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10838 return -EINVAL;
10839
10840 if (quota != RUNTIME_INF && (burst > quota ||
10841 burst + quota > max_cfs_runtime))
10842 return -EINVAL;
10843
10844 /*
10845 * Prevent race between setting of cfs_rq->runtime_enabled and
10846 * unthrottle_offline_cfs_rqs().
10847 */
10848 guard(cpus_read_lock)();
10849 guard(mutex)(&cfs_constraints_mutex);
10850
10851 ret = __cfs_schedulable(tg, period, quota);
10852 if (ret)
10853 return ret;
10854
10855 runtime_enabled = quota != RUNTIME_INF;
10856 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10857 /*
10858 * If we need to toggle cfs_bandwidth_used, off->on must occur
10859 * before making related changes, and on->off must occur afterwards
10860 */
10861 if (runtime_enabled && !runtime_was_enabled)
10862 cfs_bandwidth_usage_inc();
10863
10864 scoped_guard (raw_spinlock_irq, &cfs_b->lock) {
10865 cfs_b->period = ns_to_ktime(period);
10866 cfs_b->quota = quota;
10867 cfs_b->burst = burst;
10868
10869 __refill_cfs_bandwidth_runtime(cfs_b);
10870
10871 /*
10872 * Restart the period timer (if active) to handle new
10873 * period expiry:
10874 */
10875 if (runtime_enabled)
10876 start_cfs_bandwidth(cfs_b);
10877 }
10878
10879 for_each_online_cpu(i) {
10880 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10881 struct rq *rq = cfs_rq->rq;
10882
10883 guard(rq_lock_irq)(rq);
10884 cfs_rq->runtime_enabled = runtime_enabled;
10885 cfs_rq->runtime_remaining = 0;
10886
10887 if (cfs_rq->throttled)
10888 unthrottle_cfs_rq(cfs_rq);
10889 }
10890
10891 if (runtime_was_enabled && !runtime_enabled)
10892 cfs_bandwidth_usage_dec();
10893
10894 return 0;
10895}
10896
10897static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10898{
10899 u64 quota, period, burst;
10900
10901 period = ktime_to_ns(tg->cfs_bandwidth.period);
10902 burst = tg->cfs_bandwidth.burst;
10903 if (cfs_quota_us < 0)
10904 quota = RUNTIME_INF;
10905 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10906 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10907 else
10908 return -EINVAL;
10909
10910 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10911}
10912
10913static long tg_get_cfs_quota(struct task_group *tg)
10914{
10915 u64 quota_us;
10916
10917 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10918 return -1;
10919
10920 quota_us = tg->cfs_bandwidth.quota;
10921 do_div(quota_us, NSEC_PER_USEC);
10922
10923 return quota_us;
10924}
10925
10926static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10927{
10928 u64 quota, period, burst;
10929
10930 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10931 return -EINVAL;
10932
10933 period = (u64)cfs_period_us * NSEC_PER_USEC;
10934 quota = tg->cfs_bandwidth.quota;
10935 burst = tg->cfs_bandwidth.burst;
10936
10937 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10938}
10939
10940static long tg_get_cfs_period(struct task_group *tg)
10941{
10942 u64 cfs_period_us;
10943
10944 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10945 do_div(cfs_period_us, NSEC_PER_USEC);
10946
10947 return cfs_period_us;
10948}
10949
10950static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10951{
10952 u64 quota, period, burst;
10953
10954 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10955 return -EINVAL;
10956
10957 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10958 period = ktime_to_ns(tg->cfs_bandwidth.period);
10959 quota = tg->cfs_bandwidth.quota;
10960
10961 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10962}
10963
10964static long tg_get_cfs_burst(struct task_group *tg)
10965{
10966 u64 burst_us;
10967
10968 burst_us = tg->cfs_bandwidth.burst;
10969 do_div(burst_us, NSEC_PER_USEC);
10970
10971 return burst_us;
10972}
10973
10974static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10975 struct cftype *cft)
10976{
10977 return tg_get_cfs_quota(css_tg(css));
10978}
10979
10980static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10981 struct cftype *cftype, s64 cfs_quota_us)
10982{
10983 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10984}
10985
10986static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10987 struct cftype *cft)
10988{
10989 return tg_get_cfs_period(css_tg(css));
10990}
10991
10992static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10993 struct cftype *cftype, u64 cfs_period_us)
10994{
10995 return tg_set_cfs_period(css_tg(css), cfs_period_us);
10996}
10997
10998static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10999 struct cftype *cft)
11000{
11001 return tg_get_cfs_burst(css_tg(css));
11002}
11003
11004static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
11005 struct cftype *cftype, u64 cfs_burst_us)
11006{
11007 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
11008}
11009
11010struct cfs_schedulable_data {
11011 struct task_group *tg;
11012 u64 period, quota;
11013};
11014
11015/*
11016 * normalize group quota/period to be quota/max_period
11017 * note: units are usecs
11018 */
11019static u64 normalize_cfs_quota(struct task_group *tg,
11020 struct cfs_schedulable_data *d)
11021{
11022 u64 quota, period;
11023
11024 if (tg == d->tg) {
11025 period = d->period;
11026 quota = d->quota;
11027 } else {
11028 period = tg_get_cfs_period(tg);
11029 quota = tg_get_cfs_quota(tg);
11030 }
11031
11032 /* note: these should typically be equivalent */
11033 if (quota == RUNTIME_INF || quota == -1)
11034 return RUNTIME_INF;
11035
11036 return to_ratio(period, quota);
11037}
11038
11039static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
11040{
11041 struct cfs_schedulable_data *d = data;
11042 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11043 s64 quota = 0, parent_quota = -1;
11044
11045 if (!tg->parent) {
11046 quota = RUNTIME_INF;
11047 } else {
11048 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
11049
11050 quota = normalize_cfs_quota(tg, d);
11051 parent_quota = parent_b->hierarchical_quota;
11052
11053 /*
11054 * Ensure max(child_quota) <= parent_quota. On cgroup2,
11055 * always take the non-RUNTIME_INF min. On cgroup1, only
11056 * inherit when no limit is set. In both cases this is used
11057 * by the scheduler to determine if a given CFS task has a
11058 * bandwidth constraint at some higher level.
11059 */
11060 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
11061 if (quota == RUNTIME_INF)
11062 quota = parent_quota;
11063 else if (parent_quota != RUNTIME_INF)
11064 quota = min(quota, parent_quota);
11065 } else {
11066 if (quota == RUNTIME_INF)
11067 quota = parent_quota;
11068 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
11069 return -EINVAL;
11070 }
11071 }
11072 cfs_b->hierarchical_quota = quota;
11073
11074 return 0;
11075}
11076
11077static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
11078{
11079 struct cfs_schedulable_data data = {
11080 .tg = tg,
11081 .period = period,
11082 .quota = quota,
11083 };
11084
11085 if (quota != RUNTIME_INF) {
11086 do_div(data.period, NSEC_PER_USEC);
11087 do_div(data.quota, NSEC_PER_USEC);
11088 }
11089
11090 guard(rcu)();
11091 return walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
11092}
11093
11094static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
11095{
11096 struct task_group *tg = css_tg(seq_css(sf));
11097 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11098
11099 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
11100 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
11101 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
11102
11103 if (schedstat_enabled() && tg != &root_task_group) {
11104 struct sched_statistics *stats;
11105 u64 ws = 0;
11106 int i;
11107
11108 for_each_possible_cpu(i) {
11109 stats = __schedstats_from_se(tg->se[i]);
11110 ws += schedstat_val(stats->wait_sum);
11111 }
11112
11113 seq_printf(sf, "wait_sum %llu\n", ws);
11114 }
11115
11116 seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
11117 seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
11118
11119 return 0;
11120}
11121
11122static u64 throttled_time_self(struct task_group *tg)
11123{
11124 int i;
11125 u64 total = 0;
11126
11127 for_each_possible_cpu(i) {
11128 total += READ_ONCE(tg->cfs_rq[i]->throttled_clock_self_time);
11129 }
11130
11131 return total;
11132}
11133
11134static int cpu_cfs_local_stat_show(struct seq_file *sf, void *v)
11135{
11136 struct task_group *tg = css_tg(seq_css(sf));
11137
11138 seq_printf(sf, "throttled_time %llu\n", throttled_time_self(tg));
11139
11140 return 0;
11141}
11142#endif /* CONFIG_CFS_BANDWIDTH */
11143#endif /* CONFIG_FAIR_GROUP_SCHED */
11144
11145#ifdef CONFIG_RT_GROUP_SCHED
11146static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
11147 struct cftype *cft, s64 val)
11148{
11149 return sched_group_set_rt_runtime(css_tg(css), val);
11150}
11151
11152static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
11153 struct cftype *cft)
11154{
11155 return sched_group_rt_runtime(css_tg(css));
11156}
11157
11158static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
11159 struct cftype *cftype, u64 rt_period_us)
11160{
11161 return sched_group_set_rt_period(css_tg(css), rt_period_us);
11162}
11163
11164static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
11165 struct cftype *cft)
11166{
11167 return sched_group_rt_period(css_tg(css));
11168}
11169#endif /* CONFIG_RT_GROUP_SCHED */
11170
11171#ifdef CONFIG_FAIR_GROUP_SCHED
11172static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
11173 struct cftype *cft)
11174{
11175 return css_tg(css)->idle;
11176}
11177
11178static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
11179 struct cftype *cft, s64 idle)
11180{
11181 return sched_group_set_idle(css_tg(css), idle);
11182}
11183#endif
11184
11185static struct cftype cpu_legacy_files[] = {
11186#ifdef CONFIG_FAIR_GROUP_SCHED
11187 {
11188 .name = "shares",
11189 .read_u64 = cpu_shares_read_u64,
11190 .write_u64 = cpu_shares_write_u64,
11191 },
11192 {
11193 .name = "idle",
11194 .read_s64 = cpu_idle_read_s64,
11195 .write_s64 = cpu_idle_write_s64,
11196 },
11197#endif
11198#ifdef CONFIG_CFS_BANDWIDTH
11199 {
11200 .name = "cfs_quota_us",
11201 .read_s64 = cpu_cfs_quota_read_s64,
11202 .write_s64 = cpu_cfs_quota_write_s64,
11203 },
11204 {
11205 .name = "cfs_period_us",
11206 .read_u64 = cpu_cfs_period_read_u64,
11207 .write_u64 = cpu_cfs_period_write_u64,
11208 },
11209 {
11210 .name = "cfs_burst_us",
11211 .read_u64 = cpu_cfs_burst_read_u64,
11212 .write_u64 = cpu_cfs_burst_write_u64,
11213 },
11214 {
11215 .name = "stat",
11216 .seq_show = cpu_cfs_stat_show,
11217 },
11218 {
11219 .name = "stat.local",
11220 .seq_show = cpu_cfs_local_stat_show,
11221 },
11222#endif
11223#ifdef CONFIG_RT_GROUP_SCHED
11224 {
11225 .name = "rt_runtime_us",
11226 .read_s64 = cpu_rt_runtime_read,
11227 .write_s64 = cpu_rt_runtime_write,
11228 },
11229 {
11230 .name = "rt_period_us",
11231 .read_u64 = cpu_rt_period_read_uint,
11232 .write_u64 = cpu_rt_period_write_uint,
11233 },
11234#endif
11235#ifdef CONFIG_UCLAMP_TASK_GROUP
11236 {
11237 .name = "uclamp.min",
11238 .flags = CFTYPE_NOT_ON_ROOT,
11239 .seq_show = cpu_uclamp_min_show,
11240 .write = cpu_uclamp_min_write,
11241 },
11242 {
11243 .name = "uclamp.max",
11244 .flags = CFTYPE_NOT_ON_ROOT,
11245 .seq_show = cpu_uclamp_max_show,
11246 .write = cpu_uclamp_max_write,
11247 },
11248#endif
11249 { } /* Terminate */
11250};
11251
11252static int cpu_extra_stat_show(struct seq_file *sf,
11253 struct cgroup_subsys_state *css)
11254{
11255#ifdef CONFIG_CFS_BANDWIDTH
11256 {
11257 struct task_group *tg = css_tg(css);
11258 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11259 u64 throttled_usec, burst_usec;
11260
11261 throttled_usec = cfs_b->throttled_time;
11262 do_div(throttled_usec, NSEC_PER_USEC);
11263 burst_usec = cfs_b->burst_time;
11264 do_div(burst_usec, NSEC_PER_USEC);
11265
11266 seq_printf(sf, "nr_periods %d\n"
11267 "nr_throttled %d\n"
11268 "throttled_usec %llu\n"
11269 "nr_bursts %d\n"
11270 "burst_usec %llu\n",
11271 cfs_b->nr_periods, cfs_b->nr_throttled,
11272 throttled_usec, cfs_b->nr_burst, burst_usec);
11273 }
11274#endif
11275 return 0;
11276}
11277
11278static int cpu_local_stat_show(struct seq_file *sf,
11279 struct cgroup_subsys_state *css)
11280{
11281#ifdef CONFIG_CFS_BANDWIDTH
11282 {
11283 struct task_group *tg = css_tg(css);
11284 u64 throttled_self_usec;
11285
11286 throttled_self_usec = throttled_time_self(tg);
11287 do_div(throttled_self_usec, NSEC_PER_USEC);
11288
11289 seq_printf(sf, "throttled_usec %llu\n",
11290 throttled_self_usec);
11291 }
11292#endif
11293 return 0;
11294}
11295
11296#ifdef CONFIG_FAIR_GROUP_SCHED
11297static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
11298 struct cftype *cft)
11299{
11300 struct task_group *tg = css_tg(css);
11301 u64 weight = scale_load_down(tg->shares);
11302
11303 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
11304}
11305
11306static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
11307 struct cftype *cft, u64 weight)
11308{
11309 /*
11310 * cgroup weight knobs should use the common MIN, DFL and MAX
11311 * values which are 1, 100 and 10000 respectively. While it loses
11312 * a bit of range on both ends, it maps pretty well onto the shares
11313 * value used by scheduler and the round-trip conversions preserve
11314 * the original value over the entire range.
11315 */
11316 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
11317 return -ERANGE;
11318
11319 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
11320
11321 return sched_group_set_shares(css_tg(css), scale_load(weight));
11322}
11323
11324static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
11325 struct cftype *cft)
11326{
11327 unsigned long weight = scale_load_down(css_tg(css)->shares);
11328 int last_delta = INT_MAX;
11329 int prio, delta;
11330
11331 /* find the closest nice value to the current weight */
11332 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
11333 delta = abs(sched_prio_to_weight[prio] - weight);
11334 if (delta >= last_delta)
11335 break;
11336 last_delta = delta;
11337 }
11338
11339 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
11340}
11341
11342static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
11343 struct cftype *cft, s64 nice)
11344{
11345 unsigned long weight;
11346 int idx;
11347
11348 if (nice < MIN_NICE || nice > MAX_NICE)
11349 return -ERANGE;
11350
11351 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
11352 idx = array_index_nospec(idx, 40);
11353 weight = sched_prio_to_weight[idx];
11354
11355 return sched_group_set_shares(css_tg(css), scale_load(weight));
11356}
11357#endif
11358
11359static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
11360 long period, long quota)
11361{
11362 if (quota < 0)
11363 seq_puts(sf, "max");
11364 else
11365 seq_printf(sf, "%ld", quota);
11366
11367 seq_printf(sf, " %ld\n", period);
11368}
11369
11370/* caller should put the current value in *@periodp before calling */
11371static int __maybe_unused cpu_period_quota_parse(char *buf,
11372 u64 *periodp, u64 *quotap)
11373{
11374 char tok[21]; /* U64_MAX */
11375
11376 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
11377 return -EINVAL;
11378
11379 *periodp *= NSEC_PER_USEC;
11380
11381 if (sscanf(tok, "%llu", quotap))
11382 *quotap *= NSEC_PER_USEC;
11383 else if (!strcmp(tok, "max"))
11384 *quotap = RUNTIME_INF;
11385 else
11386 return -EINVAL;
11387
11388 return 0;
11389}
11390
11391#ifdef CONFIG_CFS_BANDWIDTH
11392static int cpu_max_show(struct seq_file *sf, void *v)
11393{
11394 struct task_group *tg = css_tg(seq_css(sf));
11395
11396 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
11397 return 0;
11398}
11399
11400static ssize_t cpu_max_write(struct kernfs_open_file *of,
11401 char *buf, size_t nbytes, loff_t off)
11402{
11403 struct task_group *tg = css_tg(of_css(of));
11404 u64 period = tg_get_cfs_period(tg);
11405 u64 burst = tg->cfs_bandwidth.burst;
11406 u64 quota;
11407 int ret;
11408
11409 ret = cpu_period_quota_parse(buf, &period, "a);
11410 if (!ret)
11411 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
11412 return ret ?: nbytes;
11413}
11414#endif
11415
11416static struct cftype cpu_files[] = {
11417#ifdef CONFIG_FAIR_GROUP_SCHED
11418 {
11419 .name = "weight",
11420 .flags = CFTYPE_NOT_ON_ROOT,
11421 .read_u64 = cpu_weight_read_u64,
11422 .write_u64 = cpu_weight_write_u64,
11423 },
11424 {
11425 .name = "weight.nice",
11426 .flags = CFTYPE_NOT_ON_ROOT,
11427 .read_s64 = cpu_weight_nice_read_s64,
11428 .write_s64 = cpu_weight_nice_write_s64,
11429 },
11430 {
11431 .name = "idle",
11432 .flags = CFTYPE_NOT_ON_ROOT,
11433 .read_s64 = cpu_idle_read_s64,
11434 .write_s64 = cpu_idle_write_s64,
11435 },
11436#endif
11437#ifdef CONFIG_CFS_BANDWIDTH
11438 {
11439 .name = "max",
11440 .flags = CFTYPE_NOT_ON_ROOT,
11441 .seq_show = cpu_max_show,
11442 .write = cpu_max_write,
11443 },
11444 {
11445 .name = "max.burst",
11446 .flags = CFTYPE_NOT_ON_ROOT,
11447 .read_u64 = cpu_cfs_burst_read_u64,
11448 .write_u64 = cpu_cfs_burst_write_u64,
11449 },
11450#endif
11451#ifdef CONFIG_UCLAMP_TASK_GROUP
11452 {
11453 .name = "uclamp.min",
11454 .flags = CFTYPE_NOT_ON_ROOT,
11455 .seq_show = cpu_uclamp_min_show,
11456 .write = cpu_uclamp_min_write,
11457 },
11458 {
11459 .name = "uclamp.max",
11460 .flags = CFTYPE_NOT_ON_ROOT,
11461 .seq_show = cpu_uclamp_max_show,
11462 .write = cpu_uclamp_max_write,
11463 },
11464#endif
11465 { } /* terminate */
11466};
11467
11468struct cgroup_subsys cpu_cgrp_subsys = {
11469 .css_alloc = cpu_cgroup_css_alloc,
11470 .css_online = cpu_cgroup_css_online,
11471 .css_released = cpu_cgroup_css_released,
11472 .css_free = cpu_cgroup_css_free,
11473 .css_extra_stat_show = cpu_extra_stat_show,
11474 .css_local_stat_show = cpu_local_stat_show,
11475#ifdef CONFIG_RT_GROUP_SCHED
11476 .can_attach = cpu_cgroup_can_attach,
11477#endif
11478 .attach = cpu_cgroup_attach,
11479 .legacy_cftypes = cpu_legacy_files,
11480 .dfl_cftypes = cpu_files,
11481 .early_init = true,
11482 .threaded = true,
11483};
11484
11485#endif /* CONFIG_CGROUP_SCHED */
11486
11487void dump_cpu_task(int cpu)
11488{
11489 if (cpu == smp_processor_id() && in_hardirq()) {
11490 struct pt_regs *regs;
11491
11492 regs = get_irq_regs();
11493 if (regs) {
11494 show_regs(regs);
11495 return;
11496 }
11497 }
11498
11499 if (trigger_single_cpu_backtrace(cpu))
11500 return;
11501
11502 pr_info("Task dump for CPU %d:\n", cpu);
11503 sched_show_task(cpu_curr(cpu));
11504}
11505
11506/*
11507 * Nice levels are multiplicative, with a gentle 10% change for every
11508 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11509 * nice 1, it will get ~10% less CPU time than another CPU-bound task
11510 * that remained on nice 0.
11511 *
11512 * The "10% effect" is relative and cumulative: from _any_ nice level,
11513 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11514 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11515 * If a task goes up by ~10% and another task goes down by ~10% then
11516 * the relative distance between them is ~25%.)
11517 */
11518const int sched_prio_to_weight[40] = {
11519 /* -20 */ 88761, 71755, 56483, 46273, 36291,
11520 /* -15 */ 29154, 23254, 18705, 14949, 11916,
11521 /* -10 */ 9548, 7620, 6100, 4904, 3906,
11522 /* -5 */ 3121, 2501, 1991, 1586, 1277,
11523 /* 0 */ 1024, 820, 655, 526, 423,
11524 /* 5 */ 335, 272, 215, 172, 137,
11525 /* 10 */ 110, 87, 70, 56, 45,
11526 /* 15 */ 36, 29, 23, 18, 15,
11527};
11528
11529/*
11530 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11531 *
11532 * In cases where the weight does not change often, we can use the
11533 * precalculated inverse to speed up arithmetics by turning divisions
11534 * into multiplications:
11535 */
11536const u32 sched_prio_to_wmult[40] = {
11537 /* -20 */ 48388, 59856, 76040, 92818, 118348,
11538 /* -15 */ 147320, 184698, 229616, 287308, 360437,
11539 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
11540 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
11541 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
11542 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
11543 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
11544 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11545};
11546
11547void call_trace_sched_update_nr_running(struct rq *rq, int count)
11548{
11549 trace_sched_update_nr_running_tp(rq, count);
11550}
11551
11552#ifdef CONFIG_SCHED_MM_CID
11553
11554/*
11555 * @cid_lock: Guarantee forward-progress of cid allocation.
11556 *
11557 * Concurrency ID allocation within a bitmap is mostly lock-free. The cid_lock
11558 * is only used when contention is detected by the lock-free allocation so
11559 * forward progress can be guaranteed.
11560 */
11561DEFINE_RAW_SPINLOCK(cid_lock);
11562
11563/*
11564 * @use_cid_lock: Select cid allocation behavior: lock-free vs spinlock.
11565 *
11566 * When @use_cid_lock is 0, the cid allocation is lock-free. When contention is
11567 * detected, it is set to 1 to ensure that all newly coming allocations are
11568 * serialized by @cid_lock until the allocation which detected contention
11569 * completes and sets @use_cid_lock back to 0. This guarantees forward progress
11570 * of a cid allocation.
11571 */
11572int use_cid_lock;
11573
11574/*
11575 * mm_cid remote-clear implements a lock-free algorithm to clear per-mm/cpu cid
11576 * concurrently with respect to the execution of the source runqueue context
11577 * switch.
11578 *
11579 * There is one basic properties we want to guarantee here:
11580 *
11581 * (1) Remote-clear should _never_ mark a per-cpu cid UNSET when it is actively
11582 * used by a task. That would lead to concurrent allocation of the cid and
11583 * userspace corruption.
11584 *
11585 * Provide this guarantee by introducing a Dekker memory ordering to guarantee
11586 * that a pair of loads observe at least one of a pair of stores, which can be
11587 * shown as:
11588 *
11589 * X = Y = 0
11590 *
11591 * w[X]=1 w[Y]=1
11592 * MB MB
11593 * r[Y]=y r[X]=x
11594 *
11595 * Which guarantees that x==0 && y==0 is impossible. But rather than using
11596 * values 0 and 1, this algorithm cares about specific state transitions of the
11597 * runqueue current task (as updated by the scheduler context switch), and the
11598 * per-mm/cpu cid value.
11599 *
11600 * Let's introduce task (Y) which has task->mm == mm and task (N) which has
11601 * task->mm != mm for the rest of the discussion. There are two scheduler state
11602 * transitions on context switch we care about:
11603 *
11604 * (TSA) Store to rq->curr with transition from (N) to (Y)
11605 *
11606 * (TSB) Store to rq->curr with transition from (Y) to (N)
11607 *
11608 * On the remote-clear side, there is one transition we care about:
11609 *
11610 * (TMA) cmpxchg to *pcpu_cid to set the LAZY flag
11611 *
11612 * There is also a transition to UNSET state which can be performed from all
11613 * sides (scheduler, remote-clear). It is always performed with a cmpxchg which
11614 * guarantees that only a single thread will succeed:
11615 *
11616 * (TMB) cmpxchg to *pcpu_cid to mark UNSET
11617 *
11618 * Just to be clear, what we do _not_ want to happen is a transition to UNSET
11619 * when a thread is actively using the cid (property (1)).
11620 *
11621 * Let's looks at the relevant combinations of TSA/TSB, and TMA transitions.
11622 *
11623 * Scenario A) (TSA)+(TMA) (from next task perspective)
11624 *
11625 * CPU0 CPU1
11626 *
11627 * Context switch CS-1 Remote-clear
11628 * - store to rq->curr: (N)->(Y) (TSA) - cmpxchg to *pcpu_id to LAZY (TMA)
11629 * (implied barrier after cmpxchg)
11630 * - switch_mm_cid()
11631 * - memory barrier (see switch_mm_cid()
11632 * comment explaining how this barrier
11633 * is combined with other scheduler
11634 * barriers)
11635 * - mm_cid_get (next)
11636 * - READ_ONCE(*pcpu_cid) - rcu_dereference(src_rq->curr)
11637 *
11638 * This Dekker ensures that either task (Y) is observed by the
11639 * rcu_dereference() or the LAZY flag is observed by READ_ONCE(), or both are
11640 * observed.
11641 *
11642 * If task (Y) store is observed by rcu_dereference(), it means that there is
11643 * still an active task on the cpu. Remote-clear will therefore not transition
11644 * to UNSET, which fulfills property (1).
11645 *
11646 * If task (Y) is not observed, but the lazy flag is observed by READ_ONCE(),
11647 * it will move its state to UNSET, which clears the percpu cid perhaps
11648 * uselessly (which is not an issue for correctness). Because task (Y) is not
11649 * observed, CPU1 can move ahead to set the state to UNSET. Because moving
11650 * state to UNSET is done with a cmpxchg expecting that the old state has the
11651 * LAZY flag set, only one thread will successfully UNSET.
11652 *
11653 * If both states (LAZY flag and task (Y)) are observed, the thread on CPU0
11654 * will observe the LAZY flag and transition to UNSET (perhaps uselessly), and
11655 * CPU1 will observe task (Y) and do nothing more, which is fine.
11656 *
11657 * What we are effectively preventing with this Dekker is a scenario where
11658 * neither LAZY flag nor store (Y) are observed, which would fail property (1)
11659 * because this would UNSET a cid which is actively used.
11660 */
11661
11662void sched_mm_cid_migrate_from(struct task_struct *t)
11663{
11664 t->migrate_from_cpu = task_cpu(t);
11665}
11666
11667static
11668int __sched_mm_cid_migrate_from_fetch_cid(struct rq *src_rq,
11669 struct task_struct *t,
11670 struct mm_cid *src_pcpu_cid)
11671{
11672 struct mm_struct *mm = t->mm;
11673 struct task_struct *src_task;
11674 int src_cid, last_mm_cid;
11675
11676 if (!mm)
11677 return -1;
11678
11679 last_mm_cid = t->last_mm_cid;
11680 /*
11681 * If the migrated task has no last cid, or if the current
11682 * task on src rq uses the cid, it means the source cid does not need
11683 * to be moved to the destination cpu.
11684 */
11685 if (last_mm_cid == -1)
11686 return -1;
11687 src_cid = READ_ONCE(src_pcpu_cid->cid);
11688 if (!mm_cid_is_valid(src_cid) || last_mm_cid != src_cid)
11689 return -1;
11690
11691 /*
11692 * If we observe an active task using the mm on this rq, it means we
11693 * are not the last task to be migrated from this cpu for this mm, so
11694 * there is no need to move src_cid to the destination cpu.
11695 */
11696 guard(rcu)();
11697 src_task = rcu_dereference(src_rq->curr);
11698 if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11699 t->last_mm_cid = -1;
11700 return -1;
11701 }
11702
11703 return src_cid;
11704}
11705
11706static
11707int __sched_mm_cid_migrate_from_try_steal_cid(struct rq *src_rq,
11708 struct task_struct *t,
11709 struct mm_cid *src_pcpu_cid,
11710 int src_cid)
11711{
11712 struct task_struct *src_task;
11713 struct mm_struct *mm = t->mm;
11714 int lazy_cid;
11715
11716 if (src_cid == -1)
11717 return -1;
11718
11719 /*
11720 * Attempt to clear the source cpu cid to move it to the destination
11721 * cpu.
11722 */
11723 lazy_cid = mm_cid_set_lazy_put(src_cid);
11724 if (!try_cmpxchg(&src_pcpu_cid->cid, &src_cid, lazy_cid))
11725 return -1;
11726
11727 /*
11728 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11729 * rq->curr->mm matches the scheduler barrier in context_switch()
11730 * between store to rq->curr and load of prev and next task's
11731 * per-mm/cpu cid.
11732 *
11733 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11734 * rq->curr->mm_cid_active matches the barrier in
11735 * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11736 * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11737 * load of per-mm/cpu cid.
11738 */
11739
11740 /*
11741 * If we observe an active task using the mm on this rq after setting
11742 * the lazy-put flag, this task will be responsible for transitioning
11743 * from lazy-put flag set to MM_CID_UNSET.
11744 */
11745 scoped_guard (rcu) {
11746 src_task = rcu_dereference(src_rq->curr);
11747 if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11748 /*
11749 * We observed an active task for this mm, there is therefore
11750 * no point in moving this cid to the destination cpu.
11751 */
11752 t->last_mm_cid = -1;
11753 return -1;
11754 }
11755 }
11756
11757 /*
11758 * The src_cid is unused, so it can be unset.
11759 */
11760 if (!try_cmpxchg(&src_pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11761 return -1;
11762 return src_cid;
11763}
11764
11765/*
11766 * Migration to dst cpu. Called with dst_rq lock held.
11767 * Interrupts are disabled, which keeps the window of cid ownership without the
11768 * source rq lock held small.
11769 */
11770void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t)
11771{
11772 struct mm_cid *src_pcpu_cid, *dst_pcpu_cid;
11773 struct mm_struct *mm = t->mm;
11774 int src_cid, dst_cid, src_cpu;
11775 struct rq *src_rq;
11776
11777 lockdep_assert_rq_held(dst_rq);
11778
11779 if (!mm)
11780 return;
11781 src_cpu = t->migrate_from_cpu;
11782 if (src_cpu == -1) {
11783 t->last_mm_cid = -1;
11784 return;
11785 }
11786 /*
11787 * Move the src cid if the dst cid is unset. This keeps id
11788 * allocation closest to 0 in cases where few threads migrate around
11789 * many cpus.
11790 *
11791 * If destination cid is already set, we may have to just clear
11792 * the src cid to ensure compactness in frequent migrations
11793 * scenarios.
11794 *
11795 * It is not useful to clear the src cid when the number of threads is
11796 * greater or equal to the number of allowed cpus, because user-space
11797 * can expect that the number of allowed cids can reach the number of
11798 * allowed cpus.
11799 */
11800 dst_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(dst_rq));
11801 dst_cid = READ_ONCE(dst_pcpu_cid->cid);
11802 if (!mm_cid_is_unset(dst_cid) &&
11803 atomic_read(&mm->mm_users) >= t->nr_cpus_allowed)
11804 return;
11805 src_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, src_cpu);
11806 src_rq = cpu_rq(src_cpu);
11807 src_cid = __sched_mm_cid_migrate_from_fetch_cid(src_rq, t, src_pcpu_cid);
11808 if (src_cid == -1)
11809 return;
11810 src_cid = __sched_mm_cid_migrate_from_try_steal_cid(src_rq, t, src_pcpu_cid,
11811 src_cid);
11812 if (src_cid == -1)
11813 return;
11814 if (!mm_cid_is_unset(dst_cid)) {
11815 __mm_cid_put(mm, src_cid);
11816 return;
11817 }
11818 /* Move src_cid to dst cpu. */
11819 mm_cid_snapshot_time(dst_rq, mm);
11820 WRITE_ONCE(dst_pcpu_cid->cid, src_cid);
11821}
11822
11823static void sched_mm_cid_remote_clear(struct mm_struct *mm, struct mm_cid *pcpu_cid,
11824 int cpu)
11825{
11826 struct rq *rq = cpu_rq(cpu);
11827 struct task_struct *t;
11828 int cid, lazy_cid;
11829
11830 cid = READ_ONCE(pcpu_cid->cid);
11831 if (!mm_cid_is_valid(cid))
11832 return;
11833
11834 /*
11835 * Clear the cpu cid if it is set to keep cid allocation compact. If
11836 * there happens to be other tasks left on the source cpu using this
11837 * mm, the next task using this mm will reallocate its cid on context
11838 * switch.
11839 */
11840 lazy_cid = mm_cid_set_lazy_put(cid);
11841 if (!try_cmpxchg(&pcpu_cid->cid, &cid, lazy_cid))
11842 return;
11843
11844 /*
11845 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11846 * rq->curr->mm matches the scheduler barrier in context_switch()
11847 * between store to rq->curr and load of prev and next task's
11848 * per-mm/cpu cid.
11849 *
11850 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11851 * rq->curr->mm_cid_active matches the barrier in
11852 * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11853 * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11854 * load of per-mm/cpu cid.
11855 */
11856
11857 /*
11858 * If we observe an active task using the mm on this rq after setting
11859 * the lazy-put flag, that task will be responsible for transitioning
11860 * from lazy-put flag set to MM_CID_UNSET.
11861 */
11862 scoped_guard (rcu) {
11863 t = rcu_dereference(rq->curr);
11864 if (READ_ONCE(t->mm_cid_active) && t->mm == mm)
11865 return;
11866 }
11867
11868 /*
11869 * The cid is unused, so it can be unset.
11870 * Disable interrupts to keep the window of cid ownership without rq
11871 * lock small.
11872 */
11873 scoped_guard (irqsave) {
11874 if (try_cmpxchg(&pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11875 __mm_cid_put(mm, cid);
11876 }
11877}
11878
11879static void sched_mm_cid_remote_clear_old(struct mm_struct *mm, int cpu)
11880{
11881 struct rq *rq = cpu_rq(cpu);
11882 struct mm_cid *pcpu_cid;
11883 struct task_struct *curr;
11884 u64 rq_clock;
11885
11886 /*
11887 * rq->clock load is racy on 32-bit but one spurious clear once in a
11888 * while is irrelevant.
11889 */
11890 rq_clock = READ_ONCE(rq->clock);
11891 pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11892
11893 /*
11894 * In order to take care of infrequently scheduled tasks, bump the time
11895 * snapshot associated with this cid if an active task using the mm is
11896 * observed on this rq.
11897 */
11898 scoped_guard (rcu) {
11899 curr = rcu_dereference(rq->curr);
11900 if (READ_ONCE(curr->mm_cid_active) && curr->mm == mm) {
11901 WRITE_ONCE(pcpu_cid->time, rq_clock);
11902 return;
11903 }
11904 }
11905
11906 if (rq_clock < pcpu_cid->time + SCHED_MM_CID_PERIOD_NS)
11907 return;
11908 sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11909}
11910
11911static void sched_mm_cid_remote_clear_weight(struct mm_struct *mm, int cpu,
11912 int weight)
11913{
11914 struct mm_cid *pcpu_cid;
11915 int cid;
11916
11917 pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11918 cid = READ_ONCE(pcpu_cid->cid);
11919 if (!mm_cid_is_valid(cid) || cid < weight)
11920 return;
11921 sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11922}
11923
11924static void task_mm_cid_work(struct callback_head *work)
11925{
11926 unsigned long now = jiffies, old_scan, next_scan;
11927 struct task_struct *t = current;
11928 struct cpumask *cidmask;
11929 struct mm_struct *mm;
11930 int weight, cpu;
11931
11932 SCHED_WARN_ON(t != container_of(work, struct task_struct, cid_work));
11933
11934 work->next = work; /* Prevent double-add */
11935 if (t->flags & PF_EXITING)
11936 return;
11937 mm = t->mm;
11938 if (!mm)
11939 return;
11940 old_scan = READ_ONCE(mm->mm_cid_next_scan);
11941 next_scan = now + msecs_to_jiffies(MM_CID_SCAN_DELAY);
11942 if (!old_scan) {
11943 unsigned long res;
11944
11945 res = cmpxchg(&mm->mm_cid_next_scan, old_scan, next_scan);
11946 if (res != old_scan)
11947 old_scan = res;
11948 else
11949 old_scan = next_scan;
11950 }
11951 if (time_before(now, old_scan))
11952 return;
11953 if (!try_cmpxchg(&mm->mm_cid_next_scan, &old_scan, next_scan))
11954 return;
11955 cidmask = mm_cidmask(mm);
11956 /* Clear cids that were not recently used. */
11957 for_each_possible_cpu(cpu)
11958 sched_mm_cid_remote_clear_old(mm, cpu);
11959 weight = cpumask_weight(cidmask);
11960 /*
11961 * Clear cids that are greater or equal to the cidmask weight to
11962 * recompact it.
11963 */
11964 for_each_possible_cpu(cpu)
11965 sched_mm_cid_remote_clear_weight(mm, cpu, weight);
11966}
11967
11968void init_sched_mm_cid(struct task_struct *t)
11969{
11970 struct mm_struct *mm = t->mm;
11971 int mm_users = 0;
11972
11973 if (mm) {
11974 mm_users = atomic_read(&mm->mm_users);
11975 if (mm_users == 1)
11976 mm->mm_cid_next_scan = jiffies + msecs_to_jiffies(MM_CID_SCAN_DELAY);
11977 }
11978 t->cid_work.next = &t->cid_work; /* Protect against double add */
11979 init_task_work(&t->cid_work, task_mm_cid_work);
11980}
11981
11982void task_tick_mm_cid(struct rq *rq, struct task_struct *curr)
11983{
11984 struct callback_head *work = &curr->cid_work;
11985 unsigned long now = jiffies;
11986
11987 if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) ||
11988 work->next != work)
11989 return;
11990 if (time_before(now, READ_ONCE(curr->mm->mm_cid_next_scan)))
11991 return;
11992 task_work_add(curr, work, TWA_RESUME);
11993}
11994
11995void sched_mm_cid_exit_signals(struct task_struct *t)
11996{
11997 struct mm_struct *mm = t->mm;
11998 struct rq *rq;
11999
12000 if (!mm)
12001 return;
12002
12003 preempt_disable();
12004 rq = this_rq();
12005 guard(rq_lock_irqsave)(rq);
12006 preempt_enable_no_resched(); /* holding spinlock */
12007 WRITE_ONCE(t->mm_cid_active, 0);
12008 /*
12009 * Store t->mm_cid_active before loading per-mm/cpu cid.
12010 * Matches barrier in sched_mm_cid_remote_clear_old().
12011 */
12012 smp_mb();
12013 mm_cid_put(mm);
12014 t->last_mm_cid = t->mm_cid = -1;
12015}
12016
12017void sched_mm_cid_before_execve(struct task_struct *t)
12018{
12019 struct mm_struct *mm = t->mm;
12020 struct rq *rq;
12021
12022 if (!mm)
12023 return;
12024
12025 preempt_disable();
12026 rq = this_rq();
12027 guard(rq_lock_irqsave)(rq);
12028 preempt_enable_no_resched(); /* holding spinlock */
12029 WRITE_ONCE(t->mm_cid_active, 0);
12030 /*
12031 * Store t->mm_cid_active before loading per-mm/cpu cid.
12032 * Matches barrier in sched_mm_cid_remote_clear_old().
12033 */
12034 smp_mb();
12035 mm_cid_put(mm);
12036 t->last_mm_cid = t->mm_cid = -1;
12037}
12038
12039void sched_mm_cid_after_execve(struct task_struct *t)
12040{
12041 struct mm_struct *mm = t->mm;
12042 struct rq *rq;
12043
12044 if (!mm)
12045 return;
12046
12047 preempt_disable();
12048 rq = this_rq();
12049 scoped_guard (rq_lock_irqsave, rq) {
12050 preempt_enable_no_resched(); /* holding spinlock */
12051 WRITE_ONCE(t->mm_cid_active, 1);
12052 /*
12053 * Store t->mm_cid_active before loading per-mm/cpu cid.
12054 * Matches barrier in sched_mm_cid_remote_clear_old().
12055 */
12056 smp_mb();
12057 t->last_mm_cid = t->mm_cid = mm_cid_get(rq, mm);
12058 }
12059 rseq_set_notify_resume(t);
12060}
12061
12062void sched_mm_cid_fork(struct task_struct *t)
12063{
12064 WARN_ON_ONCE(!t->mm || t->mm_cid != -1);
12065 t->mm_cid_active = 1;
12066}
12067#endif