Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Simple CPU accounting cgroup controller
4 */
5
6#ifdef CONFIG_IRQ_TIME_ACCOUNTING
7
8/*
9 * There are no locks covering percpu hardirq/softirq time.
10 * They are only modified in vtime_account, on corresponding CPU
11 * with interrupts disabled. So, writes are safe.
12 * They are read and saved off onto struct rq in update_rq_clock().
13 * This may result in other CPU reading this CPU's irq time and can
14 * race with irq/vtime_account on this CPU. We would either get old
15 * or new value with a side effect of accounting a slice of irq time to wrong
16 * task when irq is in progress while we read rq->clock. That is a worthy
17 * compromise in place of having locks on each irq in account_system_time.
18 */
19DEFINE_PER_CPU(struct irqtime, cpu_irqtime);
20
21static int sched_clock_irqtime;
22
23void enable_sched_clock_irqtime(void)
24{
25 sched_clock_irqtime = 1;
26}
27
28void disable_sched_clock_irqtime(void)
29{
30 sched_clock_irqtime = 0;
31}
32
33static void irqtime_account_delta(struct irqtime *irqtime, u64 delta,
34 enum cpu_usage_stat idx)
35{
36 u64 *cpustat = kcpustat_this_cpu->cpustat;
37
38 u64_stats_update_begin(&irqtime->sync);
39 cpustat[idx] += delta;
40 irqtime->total += delta;
41 irqtime->tick_delta += delta;
42 u64_stats_update_end(&irqtime->sync);
43}
44
45/*
46 * Called after incrementing preempt_count on {soft,}irq_enter
47 * and before decrementing preempt_count on {soft,}irq_exit.
48 */
49void irqtime_account_irq(struct task_struct *curr, unsigned int offset)
50{
51 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
52 unsigned int pc;
53 s64 delta;
54 int cpu;
55
56 if (!sched_clock_irqtime)
57 return;
58
59 cpu = smp_processor_id();
60 delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
61 irqtime->irq_start_time += delta;
62 pc = irq_count() - offset;
63
64 /*
65 * We do not account for softirq time from ksoftirqd here.
66 * We want to continue accounting softirq time to ksoftirqd thread
67 * in that case, so as not to confuse scheduler with a special task
68 * that do not consume any time, but still wants to run.
69 */
70 if (pc & HARDIRQ_MASK)
71 irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
72 else if ((pc & SOFTIRQ_OFFSET) && curr != this_cpu_ksoftirqd())
73 irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
74}
75
76static u64 irqtime_tick_accounted(u64 maxtime)
77{
78 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
79 u64 delta;
80
81 delta = min(irqtime->tick_delta, maxtime);
82 irqtime->tick_delta -= delta;
83
84 return delta;
85}
86
87#else /* CONFIG_IRQ_TIME_ACCOUNTING */
88
89#define sched_clock_irqtime (0)
90
91static u64 irqtime_tick_accounted(u64 dummy)
92{
93 return 0;
94}
95
96#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
97
98static inline void task_group_account_field(struct task_struct *p, int index,
99 u64 tmp)
100{
101 /*
102 * Since all updates are sure to touch the root cgroup, we
103 * get ourselves ahead and touch it first. If the root cgroup
104 * is the only cgroup, then nothing else should be necessary.
105 *
106 */
107 __this_cpu_add(kernel_cpustat.cpustat[index], tmp);
108
109 cgroup_account_cputime_field(p, index, tmp);
110}
111
112/*
113 * Account user CPU time to a process.
114 * @p: the process that the CPU time gets accounted to
115 * @cputime: the CPU time spent in user space since the last update
116 */
117void account_user_time(struct task_struct *p, u64 cputime)
118{
119 int index;
120
121 /* Add user time to process. */
122 p->utime += cputime;
123 account_group_user_time(p, cputime);
124
125 index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
126
127 /* Add user time to cpustat. */
128 task_group_account_field(p, index, cputime);
129
130 /* Account for user time used */
131 acct_account_cputime(p);
132}
133
134/*
135 * Account guest CPU time to a process.
136 * @p: the process that the CPU time gets accounted to
137 * @cputime: the CPU time spent in virtual machine since the last update
138 */
139void account_guest_time(struct task_struct *p, u64 cputime)
140{
141 u64 *cpustat = kcpustat_this_cpu->cpustat;
142
143 /* Add guest time to process. */
144 p->utime += cputime;
145 account_group_user_time(p, cputime);
146 p->gtime += cputime;
147
148 /* Add guest time to cpustat. */
149 if (task_nice(p) > 0) {
150 task_group_account_field(p, CPUTIME_NICE, cputime);
151 cpustat[CPUTIME_GUEST_NICE] += cputime;
152 } else {
153 task_group_account_field(p, CPUTIME_USER, cputime);
154 cpustat[CPUTIME_GUEST] += cputime;
155 }
156}
157
158/*
159 * Account system CPU time to a process and desired cpustat field
160 * @p: the process that the CPU time gets accounted to
161 * @cputime: the CPU time spent in kernel space since the last update
162 * @index: pointer to cpustat field that has to be updated
163 */
164void account_system_index_time(struct task_struct *p,
165 u64 cputime, enum cpu_usage_stat index)
166{
167 /* Add system time to process. */
168 p->stime += cputime;
169 account_group_system_time(p, cputime);
170
171 /* Add system time to cpustat. */
172 task_group_account_field(p, index, cputime);
173
174 /* Account for system time used */
175 acct_account_cputime(p);
176}
177
178/*
179 * Account system CPU time to a process.
180 * @p: the process that the CPU time gets accounted to
181 * @hardirq_offset: the offset to subtract from hardirq_count()
182 * @cputime: the CPU time spent in kernel space since the last update
183 */
184void account_system_time(struct task_struct *p, int hardirq_offset, u64 cputime)
185{
186 int index;
187
188 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
189 account_guest_time(p, cputime);
190 return;
191 }
192
193 if (hardirq_count() - hardirq_offset)
194 index = CPUTIME_IRQ;
195 else if (in_serving_softirq())
196 index = CPUTIME_SOFTIRQ;
197 else
198 index = CPUTIME_SYSTEM;
199
200 account_system_index_time(p, cputime, index);
201}
202
203/*
204 * Account for involuntary wait time.
205 * @cputime: the CPU time spent in involuntary wait
206 */
207void account_steal_time(u64 cputime)
208{
209 u64 *cpustat = kcpustat_this_cpu->cpustat;
210
211 cpustat[CPUTIME_STEAL] += cputime;
212}
213
214/*
215 * Account for idle time.
216 * @cputime: the CPU time spent in idle wait
217 */
218void account_idle_time(u64 cputime)
219{
220 u64 *cpustat = kcpustat_this_cpu->cpustat;
221 struct rq *rq = this_rq();
222
223 if (atomic_read(&rq->nr_iowait) > 0)
224 cpustat[CPUTIME_IOWAIT] += cputime;
225 else
226 cpustat[CPUTIME_IDLE] += cputime;
227}
228
229
230#ifdef CONFIG_SCHED_CORE
231/*
232 * Account for forceidle time due to core scheduling.
233 *
234 * REQUIRES: schedstat is enabled.
235 */
236void __account_forceidle_time(struct task_struct *p, u64 delta)
237{
238 __schedstat_add(p->stats.core_forceidle_sum, delta);
239
240 task_group_account_field(p, CPUTIME_FORCEIDLE, delta);
241}
242#endif
243
244/*
245 * When a guest is interrupted for a longer amount of time, missed clock
246 * ticks are not redelivered later. Due to that, this function may on
247 * occasion account more time than the calling functions think elapsed.
248 */
249static __always_inline u64 steal_account_process_time(u64 maxtime)
250{
251#ifdef CONFIG_PARAVIRT
252 if (static_key_false(¶virt_steal_enabled)) {
253 u64 steal;
254
255 steal = paravirt_steal_clock(smp_processor_id());
256 steal -= this_rq()->prev_steal_time;
257 steal = min(steal, maxtime);
258 account_steal_time(steal);
259 this_rq()->prev_steal_time += steal;
260
261 return steal;
262 }
263#endif
264 return 0;
265}
266
267/*
268 * Account how much elapsed time was spent in steal, irq, or softirq time.
269 */
270static inline u64 account_other_time(u64 max)
271{
272 u64 accounted;
273
274 lockdep_assert_irqs_disabled();
275
276 accounted = steal_account_process_time(max);
277
278 if (accounted < max)
279 accounted += irqtime_tick_accounted(max - accounted);
280
281 return accounted;
282}
283
284#ifdef CONFIG_64BIT
285static inline u64 read_sum_exec_runtime(struct task_struct *t)
286{
287 return t->se.sum_exec_runtime;
288}
289#else
290static u64 read_sum_exec_runtime(struct task_struct *t)
291{
292 u64 ns;
293 struct rq_flags rf;
294 struct rq *rq;
295
296 rq = task_rq_lock(t, &rf);
297 ns = t->se.sum_exec_runtime;
298 task_rq_unlock(rq, t, &rf);
299
300 return ns;
301}
302#endif
303
304/*
305 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
306 * tasks (sum on group iteration) belonging to @tsk's group.
307 */
308void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
309{
310 struct signal_struct *sig = tsk->signal;
311 u64 utime, stime;
312 struct task_struct *t;
313 unsigned int seq, nextseq;
314 unsigned long flags;
315
316 /*
317 * Update current task runtime to account pending time since last
318 * scheduler action or thread_group_cputime() call. This thread group
319 * might have other running tasks on different CPUs, but updating
320 * their runtime can affect syscall performance, so we skip account
321 * those pending times and rely only on values updated on tick or
322 * other scheduler action.
323 */
324 if (same_thread_group(current, tsk))
325 (void) task_sched_runtime(current);
326
327 rcu_read_lock();
328 /* Attempt a lockless read on the first round. */
329 nextseq = 0;
330 do {
331 seq = nextseq;
332 flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
333 times->utime = sig->utime;
334 times->stime = sig->stime;
335 times->sum_exec_runtime = sig->sum_sched_runtime;
336
337 for_each_thread(tsk, t) {
338 task_cputime(t, &utime, &stime);
339 times->utime += utime;
340 times->stime += stime;
341 times->sum_exec_runtime += read_sum_exec_runtime(t);
342 }
343 /* If lockless access failed, take the lock. */
344 nextseq = 1;
345 } while (need_seqretry(&sig->stats_lock, seq));
346 done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
347 rcu_read_unlock();
348}
349
350#ifdef CONFIG_IRQ_TIME_ACCOUNTING
351/*
352 * Account a tick to a process and cpustat
353 * @p: the process that the CPU time gets accounted to
354 * @user_tick: is the tick from userspace
355 * @rq: the pointer to rq
356 *
357 * Tick demultiplexing follows the order
358 * - pending hardirq update
359 * - pending softirq update
360 * - user_time
361 * - idle_time
362 * - system time
363 * - check for guest_time
364 * - else account as system_time
365 *
366 * Check for hardirq is done both for system and user time as there is
367 * no timer going off while we are on hardirq and hence we may never get an
368 * opportunity to update it solely in system time.
369 * p->stime and friends are only updated on system time and not on irq
370 * softirq as those do not count in task exec_runtime any more.
371 */
372static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
373 int ticks)
374{
375 u64 other, cputime = TICK_NSEC * ticks;
376
377 /*
378 * When returning from idle, many ticks can get accounted at
379 * once, including some ticks of steal, irq, and softirq time.
380 * Subtract those ticks from the amount of time accounted to
381 * idle, or potentially user or system time. Due to rounding,
382 * other time can exceed ticks occasionally.
383 */
384 other = account_other_time(ULONG_MAX);
385 if (other >= cputime)
386 return;
387
388 cputime -= other;
389
390 if (this_cpu_ksoftirqd() == p) {
391 /*
392 * ksoftirqd time do not get accounted in cpu_softirq_time.
393 * So, we have to handle it separately here.
394 * Also, p->stime needs to be updated for ksoftirqd.
395 */
396 account_system_index_time(p, cputime, CPUTIME_SOFTIRQ);
397 } else if (user_tick) {
398 account_user_time(p, cputime);
399 } else if (p == this_rq()->idle) {
400 account_idle_time(cputime);
401 } else if (p->flags & PF_VCPU) { /* System time or guest time */
402 account_guest_time(p, cputime);
403 } else {
404 account_system_index_time(p, cputime, CPUTIME_SYSTEM);
405 }
406}
407
408static void irqtime_account_idle_ticks(int ticks)
409{
410 irqtime_account_process_tick(current, 0, ticks);
411}
412#else /* CONFIG_IRQ_TIME_ACCOUNTING */
413static inline void irqtime_account_idle_ticks(int ticks) { }
414static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
415 int nr_ticks) { }
416#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
417
418/*
419 * Use precise platform statistics if available:
420 */
421#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
422
423# ifndef __ARCH_HAS_VTIME_TASK_SWITCH
424void vtime_task_switch(struct task_struct *prev)
425{
426 if (is_idle_task(prev))
427 vtime_account_idle(prev);
428 else
429 vtime_account_kernel(prev);
430
431 vtime_flush(prev);
432 arch_vtime_task_switch(prev);
433}
434# endif
435
436void vtime_account_irq(struct task_struct *tsk, unsigned int offset)
437{
438 unsigned int pc = irq_count() - offset;
439
440 if (pc & HARDIRQ_OFFSET) {
441 vtime_account_hardirq(tsk);
442 } else if (pc & SOFTIRQ_OFFSET) {
443 vtime_account_softirq(tsk);
444 } else if (!IS_ENABLED(CONFIG_HAVE_VIRT_CPU_ACCOUNTING_IDLE) &&
445 is_idle_task(tsk)) {
446 vtime_account_idle(tsk);
447 } else {
448 vtime_account_kernel(tsk);
449 }
450}
451
452void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
453 u64 *ut, u64 *st)
454{
455 *ut = curr->utime;
456 *st = curr->stime;
457}
458
459void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
460{
461 *ut = p->utime;
462 *st = p->stime;
463}
464EXPORT_SYMBOL_GPL(task_cputime_adjusted);
465
466void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
467{
468 struct task_cputime cputime;
469
470 thread_group_cputime(p, &cputime);
471
472 *ut = cputime.utime;
473 *st = cputime.stime;
474}
475
476#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE: */
477
478/*
479 * Account a single tick of CPU time.
480 * @p: the process that the CPU time gets accounted to
481 * @user_tick: indicates if the tick is a user or a system tick
482 */
483void account_process_tick(struct task_struct *p, int user_tick)
484{
485 u64 cputime, steal;
486
487 if (vtime_accounting_enabled_this_cpu())
488 return;
489
490 if (sched_clock_irqtime) {
491 irqtime_account_process_tick(p, user_tick, 1);
492 return;
493 }
494
495 cputime = TICK_NSEC;
496 steal = steal_account_process_time(ULONG_MAX);
497
498 if (steal >= cputime)
499 return;
500
501 cputime -= steal;
502
503 if (user_tick)
504 account_user_time(p, cputime);
505 else if ((p != this_rq()->idle) || (irq_count() != HARDIRQ_OFFSET))
506 account_system_time(p, HARDIRQ_OFFSET, cputime);
507 else
508 account_idle_time(cputime);
509}
510
511/*
512 * Account multiple ticks of idle time.
513 * @ticks: number of stolen ticks
514 */
515void account_idle_ticks(unsigned long ticks)
516{
517 u64 cputime, steal;
518
519 if (sched_clock_irqtime) {
520 irqtime_account_idle_ticks(ticks);
521 return;
522 }
523
524 cputime = ticks * TICK_NSEC;
525 steal = steal_account_process_time(ULONG_MAX);
526
527 if (steal >= cputime)
528 return;
529
530 cputime -= steal;
531 account_idle_time(cputime);
532}
533
534/*
535 * Adjust tick based cputime random precision against scheduler runtime
536 * accounting.
537 *
538 * Tick based cputime accounting depend on random scheduling timeslices of a
539 * task to be interrupted or not by the timer. Depending on these
540 * circumstances, the number of these interrupts may be over or
541 * under-optimistic, matching the real user and system cputime with a variable
542 * precision.
543 *
544 * Fix this by scaling these tick based values against the total runtime
545 * accounted by the CFS scheduler.
546 *
547 * This code provides the following guarantees:
548 *
549 * stime + utime == rtime
550 * stime_i+1 >= stime_i, utime_i+1 >= utime_i
551 *
552 * Assuming that rtime_i+1 >= rtime_i.
553 */
554void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
555 u64 *ut, u64 *st)
556{
557 u64 rtime, stime, utime;
558 unsigned long flags;
559
560 /* Serialize concurrent callers such that we can honour our guarantees */
561 raw_spin_lock_irqsave(&prev->lock, flags);
562 rtime = curr->sum_exec_runtime;
563
564 /*
565 * This is possible under two circumstances:
566 * - rtime isn't monotonic after all (a bug);
567 * - we got reordered by the lock.
568 *
569 * In both cases this acts as a filter such that the rest of the code
570 * can assume it is monotonic regardless of anything else.
571 */
572 if (prev->stime + prev->utime >= rtime)
573 goto out;
574
575 stime = curr->stime;
576 utime = curr->utime;
577
578 /*
579 * If either stime or utime are 0, assume all runtime is userspace.
580 * Once a task gets some ticks, the monotonicity code at 'update:'
581 * will ensure things converge to the observed ratio.
582 */
583 if (stime == 0) {
584 utime = rtime;
585 goto update;
586 }
587
588 if (utime == 0) {
589 stime = rtime;
590 goto update;
591 }
592
593 stime = mul_u64_u64_div_u64(stime, rtime, stime + utime);
594
595update:
596 /*
597 * Make sure stime doesn't go backwards; this preserves monotonicity
598 * for utime because rtime is monotonic.
599 *
600 * utime_i+1 = rtime_i+1 - stime_i
601 * = rtime_i+1 - (rtime_i - utime_i)
602 * = (rtime_i+1 - rtime_i) + utime_i
603 * >= utime_i
604 */
605 if (stime < prev->stime)
606 stime = prev->stime;
607 utime = rtime - stime;
608
609 /*
610 * Make sure utime doesn't go backwards; this still preserves
611 * monotonicity for stime, analogous argument to above.
612 */
613 if (utime < prev->utime) {
614 utime = prev->utime;
615 stime = rtime - utime;
616 }
617
618 prev->stime = stime;
619 prev->utime = utime;
620out:
621 *ut = prev->utime;
622 *st = prev->stime;
623 raw_spin_unlock_irqrestore(&prev->lock, flags);
624}
625
626void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
627{
628 struct task_cputime cputime = {
629 .sum_exec_runtime = p->se.sum_exec_runtime,
630 };
631
632 if (task_cputime(p, &cputime.utime, &cputime.stime))
633 cputime.sum_exec_runtime = task_sched_runtime(p);
634 cputime_adjust(&cputime, &p->prev_cputime, ut, st);
635}
636EXPORT_SYMBOL_GPL(task_cputime_adjusted);
637
638void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
639{
640 struct task_cputime cputime;
641
642 thread_group_cputime(p, &cputime);
643 cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
644}
645#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
646
647#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
648static u64 vtime_delta(struct vtime *vtime)
649{
650 unsigned long long clock;
651
652 clock = sched_clock();
653 if (clock < vtime->starttime)
654 return 0;
655
656 return clock - vtime->starttime;
657}
658
659static u64 get_vtime_delta(struct vtime *vtime)
660{
661 u64 delta = vtime_delta(vtime);
662 u64 other;
663
664 /*
665 * Unlike tick based timing, vtime based timing never has lost
666 * ticks, and no need for steal time accounting to make up for
667 * lost ticks. Vtime accounts a rounded version of actual
668 * elapsed time. Limit account_other_time to prevent rounding
669 * errors from causing elapsed vtime to go negative.
670 */
671 other = account_other_time(delta);
672 WARN_ON_ONCE(vtime->state == VTIME_INACTIVE);
673 vtime->starttime += delta;
674
675 return delta - other;
676}
677
678static void vtime_account_system(struct task_struct *tsk,
679 struct vtime *vtime)
680{
681 vtime->stime += get_vtime_delta(vtime);
682 if (vtime->stime >= TICK_NSEC) {
683 account_system_time(tsk, irq_count(), vtime->stime);
684 vtime->stime = 0;
685 }
686}
687
688static void vtime_account_guest(struct task_struct *tsk,
689 struct vtime *vtime)
690{
691 vtime->gtime += get_vtime_delta(vtime);
692 if (vtime->gtime >= TICK_NSEC) {
693 account_guest_time(tsk, vtime->gtime);
694 vtime->gtime = 0;
695 }
696}
697
698static void __vtime_account_kernel(struct task_struct *tsk,
699 struct vtime *vtime)
700{
701 /* We might have scheduled out from guest path */
702 if (vtime->state == VTIME_GUEST)
703 vtime_account_guest(tsk, vtime);
704 else
705 vtime_account_system(tsk, vtime);
706}
707
708void vtime_account_kernel(struct task_struct *tsk)
709{
710 struct vtime *vtime = &tsk->vtime;
711
712 if (!vtime_delta(vtime))
713 return;
714
715 write_seqcount_begin(&vtime->seqcount);
716 __vtime_account_kernel(tsk, vtime);
717 write_seqcount_end(&vtime->seqcount);
718}
719
720void vtime_user_enter(struct task_struct *tsk)
721{
722 struct vtime *vtime = &tsk->vtime;
723
724 write_seqcount_begin(&vtime->seqcount);
725 vtime_account_system(tsk, vtime);
726 vtime->state = VTIME_USER;
727 write_seqcount_end(&vtime->seqcount);
728}
729
730void vtime_user_exit(struct task_struct *tsk)
731{
732 struct vtime *vtime = &tsk->vtime;
733
734 write_seqcount_begin(&vtime->seqcount);
735 vtime->utime += get_vtime_delta(vtime);
736 if (vtime->utime >= TICK_NSEC) {
737 account_user_time(tsk, vtime->utime);
738 vtime->utime = 0;
739 }
740 vtime->state = VTIME_SYS;
741 write_seqcount_end(&vtime->seqcount);
742}
743
744void vtime_guest_enter(struct task_struct *tsk)
745{
746 struct vtime *vtime = &tsk->vtime;
747 /*
748 * The flags must be updated under the lock with
749 * the vtime_starttime flush and update.
750 * That enforces a right ordering and update sequence
751 * synchronization against the reader (task_gtime())
752 * that can thus safely catch up with a tickless delta.
753 */
754 write_seqcount_begin(&vtime->seqcount);
755 vtime_account_system(tsk, vtime);
756 tsk->flags |= PF_VCPU;
757 vtime->state = VTIME_GUEST;
758 write_seqcount_end(&vtime->seqcount);
759}
760EXPORT_SYMBOL_GPL(vtime_guest_enter);
761
762void vtime_guest_exit(struct task_struct *tsk)
763{
764 struct vtime *vtime = &tsk->vtime;
765
766 write_seqcount_begin(&vtime->seqcount);
767 vtime_account_guest(tsk, vtime);
768 tsk->flags &= ~PF_VCPU;
769 vtime->state = VTIME_SYS;
770 write_seqcount_end(&vtime->seqcount);
771}
772EXPORT_SYMBOL_GPL(vtime_guest_exit);
773
774void vtime_account_idle(struct task_struct *tsk)
775{
776 account_idle_time(get_vtime_delta(&tsk->vtime));
777}
778
779void vtime_task_switch_generic(struct task_struct *prev)
780{
781 struct vtime *vtime = &prev->vtime;
782
783 write_seqcount_begin(&vtime->seqcount);
784 if (vtime->state == VTIME_IDLE)
785 vtime_account_idle(prev);
786 else
787 __vtime_account_kernel(prev, vtime);
788 vtime->state = VTIME_INACTIVE;
789 vtime->cpu = -1;
790 write_seqcount_end(&vtime->seqcount);
791
792 vtime = ¤t->vtime;
793
794 write_seqcount_begin(&vtime->seqcount);
795 if (is_idle_task(current))
796 vtime->state = VTIME_IDLE;
797 else if (current->flags & PF_VCPU)
798 vtime->state = VTIME_GUEST;
799 else
800 vtime->state = VTIME_SYS;
801 vtime->starttime = sched_clock();
802 vtime->cpu = smp_processor_id();
803 write_seqcount_end(&vtime->seqcount);
804}
805
806void vtime_init_idle(struct task_struct *t, int cpu)
807{
808 struct vtime *vtime = &t->vtime;
809 unsigned long flags;
810
811 local_irq_save(flags);
812 write_seqcount_begin(&vtime->seqcount);
813 vtime->state = VTIME_IDLE;
814 vtime->starttime = sched_clock();
815 vtime->cpu = cpu;
816 write_seqcount_end(&vtime->seqcount);
817 local_irq_restore(flags);
818}
819
820u64 task_gtime(struct task_struct *t)
821{
822 struct vtime *vtime = &t->vtime;
823 unsigned int seq;
824 u64 gtime;
825
826 if (!vtime_accounting_enabled())
827 return t->gtime;
828
829 do {
830 seq = read_seqcount_begin(&vtime->seqcount);
831
832 gtime = t->gtime;
833 if (vtime->state == VTIME_GUEST)
834 gtime += vtime->gtime + vtime_delta(vtime);
835
836 } while (read_seqcount_retry(&vtime->seqcount, seq));
837
838 return gtime;
839}
840
841/*
842 * Fetch cputime raw values from fields of task_struct and
843 * add up the pending nohz execution time since the last
844 * cputime snapshot.
845 */
846bool task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
847{
848 struct vtime *vtime = &t->vtime;
849 unsigned int seq;
850 u64 delta;
851 int ret;
852
853 if (!vtime_accounting_enabled()) {
854 *utime = t->utime;
855 *stime = t->stime;
856 return false;
857 }
858
859 do {
860 ret = false;
861 seq = read_seqcount_begin(&vtime->seqcount);
862
863 *utime = t->utime;
864 *stime = t->stime;
865
866 /* Task is sleeping or idle, nothing to add */
867 if (vtime->state < VTIME_SYS)
868 continue;
869
870 ret = true;
871 delta = vtime_delta(vtime);
872
873 /*
874 * Task runs either in user (including guest) or kernel space,
875 * add pending nohz time to the right place.
876 */
877 if (vtime->state == VTIME_SYS)
878 *stime += vtime->stime + delta;
879 else
880 *utime += vtime->utime + delta;
881 } while (read_seqcount_retry(&vtime->seqcount, seq));
882
883 return ret;
884}
885
886static int vtime_state_fetch(struct vtime *vtime, int cpu)
887{
888 int state = READ_ONCE(vtime->state);
889
890 /*
891 * We raced against a context switch, fetch the
892 * kcpustat task again.
893 */
894 if (vtime->cpu != cpu && vtime->cpu != -1)
895 return -EAGAIN;
896
897 /*
898 * Two possible things here:
899 * 1) We are seeing the scheduling out task (prev) or any past one.
900 * 2) We are seeing the scheduling in task (next) but it hasn't
901 * passed though vtime_task_switch() yet so the pending
902 * cputime of the prev task may not be flushed yet.
903 *
904 * Case 1) is ok but 2) is not. So wait for a safe VTIME state.
905 */
906 if (state == VTIME_INACTIVE)
907 return -EAGAIN;
908
909 return state;
910}
911
912static u64 kcpustat_user_vtime(struct vtime *vtime)
913{
914 if (vtime->state == VTIME_USER)
915 return vtime->utime + vtime_delta(vtime);
916 else if (vtime->state == VTIME_GUEST)
917 return vtime->gtime + vtime_delta(vtime);
918 return 0;
919}
920
921static int kcpustat_field_vtime(u64 *cpustat,
922 struct task_struct *tsk,
923 enum cpu_usage_stat usage,
924 int cpu, u64 *val)
925{
926 struct vtime *vtime = &tsk->vtime;
927 unsigned int seq;
928
929 do {
930 int state;
931
932 seq = read_seqcount_begin(&vtime->seqcount);
933
934 state = vtime_state_fetch(vtime, cpu);
935 if (state < 0)
936 return state;
937
938 *val = cpustat[usage];
939
940 /*
941 * Nice VS unnice cputime accounting may be inaccurate if
942 * the nice value has changed since the last vtime update.
943 * But proper fix would involve interrupting target on nice
944 * updates which is a no go on nohz_full (although the scheduler
945 * may still interrupt the target if rescheduling is needed...)
946 */
947 switch (usage) {
948 case CPUTIME_SYSTEM:
949 if (state == VTIME_SYS)
950 *val += vtime->stime + vtime_delta(vtime);
951 break;
952 case CPUTIME_USER:
953 if (task_nice(tsk) <= 0)
954 *val += kcpustat_user_vtime(vtime);
955 break;
956 case CPUTIME_NICE:
957 if (task_nice(tsk) > 0)
958 *val += kcpustat_user_vtime(vtime);
959 break;
960 case CPUTIME_GUEST:
961 if (state == VTIME_GUEST && task_nice(tsk) <= 0)
962 *val += vtime->gtime + vtime_delta(vtime);
963 break;
964 case CPUTIME_GUEST_NICE:
965 if (state == VTIME_GUEST && task_nice(tsk) > 0)
966 *val += vtime->gtime + vtime_delta(vtime);
967 break;
968 default:
969 break;
970 }
971 } while (read_seqcount_retry(&vtime->seqcount, seq));
972
973 return 0;
974}
975
976u64 kcpustat_field(struct kernel_cpustat *kcpustat,
977 enum cpu_usage_stat usage, int cpu)
978{
979 u64 *cpustat = kcpustat->cpustat;
980 u64 val = cpustat[usage];
981 struct rq *rq;
982 int err;
983
984 if (!vtime_accounting_enabled_cpu(cpu))
985 return val;
986
987 rq = cpu_rq(cpu);
988
989 for (;;) {
990 struct task_struct *curr;
991
992 rcu_read_lock();
993 curr = rcu_dereference(rq->curr);
994 if (WARN_ON_ONCE(!curr)) {
995 rcu_read_unlock();
996 return cpustat[usage];
997 }
998
999 err = kcpustat_field_vtime(cpustat, curr, usage, cpu, &val);
1000 rcu_read_unlock();
1001
1002 if (!err)
1003 return val;
1004
1005 cpu_relax();
1006 }
1007}
1008EXPORT_SYMBOL_GPL(kcpustat_field);
1009
1010static int kcpustat_cpu_fetch_vtime(struct kernel_cpustat *dst,
1011 const struct kernel_cpustat *src,
1012 struct task_struct *tsk, int cpu)
1013{
1014 struct vtime *vtime = &tsk->vtime;
1015 unsigned int seq;
1016
1017 do {
1018 u64 *cpustat;
1019 u64 delta;
1020 int state;
1021
1022 seq = read_seqcount_begin(&vtime->seqcount);
1023
1024 state = vtime_state_fetch(vtime, cpu);
1025 if (state < 0)
1026 return state;
1027
1028 *dst = *src;
1029 cpustat = dst->cpustat;
1030
1031 /* Task is sleeping, dead or idle, nothing to add */
1032 if (state < VTIME_SYS)
1033 continue;
1034
1035 delta = vtime_delta(vtime);
1036
1037 /*
1038 * Task runs either in user (including guest) or kernel space,
1039 * add pending nohz time to the right place.
1040 */
1041 if (state == VTIME_SYS) {
1042 cpustat[CPUTIME_SYSTEM] += vtime->stime + delta;
1043 } else if (state == VTIME_USER) {
1044 if (task_nice(tsk) > 0)
1045 cpustat[CPUTIME_NICE] += vtime->utime + delta;
1046 else
1047 cpustat[CPUTIME_USER] += vtime->utime + delta;
1048 } else {
1049 WARN_ON_ONCE(state != VTIME_GUEST);
1050 if (task_nice(tsk) > 0) {
1051 cpustat[CPUTIME_GUEST_NICE] += vtime->gtime + delta;
1052 cpustat[CPUTIME_NICE] += vtime->gtime + delta;
1053 } else {
1054 cpustat[CPUTIME_GUEST] += vtime->gtime + delta;
1055 cpustat[CPUTIME_USER] += vtime->gtime + delta;
1056 }
1057 }
1058 } while (read_seqcount_retry(&vtime->seqcount, seq));
1059
1060 return 0;
1061}
1062
1063void kcpustat_cpu_fetch(struct kernel_cpustat *dst, int cpu)
1064{
1065 const struct kernel_cpustat *src = &kcpustat_cpu(cpu);
1066 struct rq *rq;
1067 int err;
1068
1069 if (!vtime_accounting_enabled_cpu(cpu)) {
1070 *dst = *src;
1071 return;
1072 }
1073
1074 rq = cpu_rq(cpu);
1075
1076 for (;;) {
1077 struct task_struct *curr;
1078
1079 rcu_read_lock();
1080 curr = rcu_dereference(rq->curr);
1081 if (WARN_ON_ONCE(!curr)) {
1082 rcu_read_unlock();
1083 *dst = *src;
1084 return;
1085 }
1086
1087 err = kcpustat_cpu_fetch_vtime(dst, src, curr, cpu);
1088 rcu_read_unlock();
1089
1090 if (!err)
1091 return;
1092
1093 cpu_relax();
1094 }
1095}
1096EXPORT_SYMBOL_GPL(kcpustat_cpu_fetch);
1097
1098#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Simple CPU accounting cgroup controller
4 */
5
6#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
7 #include <asm/cputime.h>
8#endif
9
10#ifdef CONFIG_IRQ_TIME_ACCOUNTING
11
12/*
13 * There are no locks covering percpu hardirq/softirq time.
14 * They are only modified in vtime_account, on corresponding CPU
15 * with interrupts disabled. So, writes are safe.
16 * They are read and saved off onto struct rq in update_rq_clock().
17 * This may result in other CPU reading this CPU's irq time and can
18 * race with irq/vtime_account on this CPU. We would either get old
19 * or new value with a side effect of accounting a slice of irq time to wrong
20 * task when irq is in progress while we read rq->clock. That is a worthy
21 * compromise in place of having locks on each irq in account_system_time.
22 */
23DEFINE_PER_CPU(struct irqtime, cpu_irqtime);
24
25static int sched_clock_irqtime;
26
27void enable_sched_clock_irqtime(void)
28{
29 sched_clock_irqtime = 1;
30}
31
32void disable_sched_clock_irqtime(void)
33{
34 sched_clock_irqtime = 0;
35}
36
37static void irqtime_account_delta(struct irqtime *irqtime, u64 delta,
38 enum cpu_usage_stat idx)
39{
40 u64 *cpustat = kcpustat_this_cpu->cpustat;
41
42 u64_stats_update_begin(&irqtime->sync);
43 cpustat[idx] += delta;
44 irqtime->total += delta;
45 irqtime->tick_delta += delta;
46 u64_stats_update_end(&irqtime->sync);
47}
48
49/*
50 * Called after incrementing preempt_count on {soft,}irq_enter
51 * and before decrementing preempt_count on {soft,}irq_exit.
52 */
53void irqtime_account_irq(struct task_struct *curr, unsigned int offset)
54{
55 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
56 unsigned int pc;
57 s64 delta;
58 int cpu;
59
60 if (!sched_clock_irqtime)
61 return;
62
63 cpu = smp_processor_id();
64 delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
65 irqtime->irq_start_time += delta;
66 pc = irq_count() - offset;
67
68 /*
69 * We do not account for softirq time from ksoftirqd here.
70 * We want to continue accounting softirq time to ksoftirqd thread
71 * in that case, so as not to confuse scheduler with a special task
72 * that do not consume any time, but still wants to run.
73 */
74 if (pc & HARDIRQ_MASK)
75 irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
76 else if ((pc & SOFTIRQ_OFFSET) && curr != this_cpu_ksoftirqd())
77 irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
78}
79
80static u64 irqtime_tick_accounted(u64 maxtime)
81{
82 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
83 u64 delta;
84
85 delta = min(irqtime->tick_delta, maxtime);
86 irqtime->tick_delta -= delta;
87
88 return delta;
89}
90
91#else /* CONFIG_IRQ_TIME_ACCOUNTING */
92
93#define sched_clock_irqtime (0)
94
95static u64 irqtime_tick_accounted(u64 dummy)
96{
97 return 0;
98}
99
100#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
101
102static inline void task_group_account_field(struct task_struct *p, int index,
103 u64 tmp)
104{
105 /*
106 * Since all updates are sure to touch the root cgroup, we
107 * get ourselves ahead and touch it first. If the root cgroup
108 * is the only cgroup, then nothing else should be necessary.
109 *
110 */
111 __this_cpu_add(kernel_cpustat.cpustat[index], tmp);
112
113 cgroup_account_cputime_field(p, index, tmp);
114}
115
116/*
117 * Account user CPU time to a process.
118 * @p: the process that the CPU time gets accounted to
119 * @cputime: the CPU time spent in user space since the last update
120 */
121void account_user_time(struct task_struct *p, u64 cputime)
122{
123 int index;
124
125 /* Add user time to process. */
126 p->utime += cputime;
127 account_group_user_time(p, cputime);
128
129 index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
130
131 /* Add user time to cpustat. */
132 task_group_account_field(p, index, cputime);
133
134 /* Account for user time used */
135 acct_account_cputime(p);
136}
137
138/*
139 * Account guest CPU time to a process.
140 * @p: the process that the CPU time gets accounted to
141 * @cputime: the CPU time spent in virtual machine since the last update
142 */
143void account_guest_time(struct task_struct *p, u64 cputime)
144{
145 u64 *cpustat = kcpustat_this_cpu->cpustat;
146
147 /* Add guest time to process. */
148 p->utime += cputime;
149 account_group_user_time(p, cputime);
150 p->gtime += cputime;
151
152 /* Add guest time to cpustat. */
153 if (task_nice(p) > 0) {
154 task_group_account_field(p, CPUTIME_NICE, cputime);
155 cpustat[CPUTIME_GUEST_NICE] += cputime;
156 } else {
157 task_group_account_field(p, CPUTIME_USER, cputime);
158 cpustat[CPUTIME_GUEST] += cputime;
159 }
160}
161
162/*
163 * Account system CPU time to a process and desired cpustat field
164 * @p: the process that the CPU time gets accounted to
165 * @cputime: the CPU time spent in kernel space since the last update
166 * @index: pointer to cpustat field that has to be updated
167 */
168void account_system_index_time(struct task_struct *p,
169 u64 cputime, enum cpu_usage_stat index)
170{
171 /* Add system time to process. */
172 p->stime += cputime;
173 account_group_system_time(p, cputime);
174
175 /* Add system time to cpustat. */
176 task_group_account_field(p, index, cputime);
177
178 /* Account for system time used */
179 acct_account_cputime(p);
180}
181
182/*
183 * Account system CPU time to a process.
184 * @p: the process that the CPU time gets accounted to
185 * @hardirq_offset: the offset to subtract from hardirq_count()
186 * @cputime: the CPU time spent in kernel space since the last update
187 */
188void account_system_time(struct task_struct *p, int hardirq_offset, u64 cputime)
189{
190 int index;
191
192 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
193 account_guest_time(p, cputime);
194 return;
195 }
196
197 if (hardirq_count() - hardirq_offset)
198 index = CPUTIME_IRQ;
199 else if (in_serving_softirq())
200 index = CPUTIME_SOFTIRQ;
201 else
202 index = CPUTIME_SYSTEM;
203
204 account_system_index_time(p, cputime, index);
205}
206
207/*
208 * Account for involuntary wait time.
209 * @cputime: the CPU time spent in involuntary wait
210 */
211void account_steal_time(u64 cputime)
212{
213 u64 *cpustat = kcpustat_this_cpu->cpustat;
214
215 cpustat[CPUTIME_STEAL] += cputime;
216}
217
218/*
219 * Account for idle time.
220 * @cputime: the CPU time spent in idle wait
221 */
222void account_idle_time(u64 cputime)
223{
224 u64 *cpustat = kcpustat_this_cpu->cpustat;
225 struct rq *rq = this_rq();
226
227 if (atomic_read(&rq->nr_iowait) > 0)
228 cpustat[CPUTIME_IOWAIT] += cputime;
229 else
230 cpustat[CPUTIME_IDLE] += cputime;
231}
232
233
234#ifdef CONFIG_SCHED_CORE
235/*
236 * Account for forceidle time due to core scheduling.
237 *
238 * REQUIRES: schedstat is enabled.
239 */
240void __account_forceidle_time(struct task_struct *p, u64 delta)
241{
242 __schedstat_add(p->stats.core_forceidle_sum, delta);
243
244 task_group_account_field(p, CPUTIME_FORCEIDLE, delta);
245}
246#endif
247
248/*
249 * When a guest is interrupted for a longer amount of time, missed clock
250 * ticks are not redelivered later. Due to that, this function may on
251 * occasion account more time than the calling functions think elapsed.
252 */
253static __always_inline u64 steal_account_process_time(u64 maxtime)
254{
255#ifdef CONFIG_PARAVIRT
256 if (static_key_false(¶virt_steal_enabled)) {
257 u64 steal;
258
259 steal = paravirt_steal_clock(smp_processor_id());
260 steal -= this_rq()->prev_steal_time;
261 steal = min(steal, maxtime);
262 account_steal_time(steal);
263 this_rq()->prev_steal_time += steal;
264
265 return steal;
266 }
267#endif
268 return 0;
269}
270
271/*
272 * Account how much elapsed time was spent in steal, irq, or softirq time.
273 */
274static inline u64 account_other_time(u64 max)
275{
276 u64 accounted;
277
278 lockdep_assert_irqs_disabled();
279
280 accounted = steal_account_process_time(max);
281
282 if (accounted < max)
283 accounted += irqtime_tick_accounted(max - accounted);
284
285 return accounted;
286}
287
288#ifdef CONFIG_64BIT
289static inline u64 read_sum_exec_runtime(struct task_struct *t)
290{
291 return t->se.sum_exec_runtime;
292}
293#else
294static u64 read_sum_exec_runtime(struct task_struct *t)
295{
296 u64 ns;
297 struct rq_flags rf;
298 struct rq *rq;
299
300 rq = task_rq_lock(t, &rf);
301 ns = t->se.sum_exec_runtime;
302 task_rq_unlock(rq, t, &rf);
303
304 return ns;
305}
306#endif
307
308/*
309 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
310 * tasks (sum on group iteration) belonging to @tsk's group.
311 */
312void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
313{
314 struct signal_struct *sig = tsk->signal;
315 u64 utime, stime;
316 struct task_struct *t;
317 unsigned int seq, nextseq;
318 unsigned long flags;
319
320 /*
321 * Update current task runtime to account pending time since last
322 * scheduler action or thread_group_cputime() call. This thread group
323 * might have other running tasks on different CPUs, but updating
324 * their runtime can affect syscall performance, so we skip account
325 * those pending times and rely only on values updated on tick or
326 * other scheduler action.
327 */
328 if (same_thread_group(current, tsk))
329 (void) task_sched_runtime(current);
330
331 rcu_read_lock();
332 /* Attempt a lockless read on the first round. */
333 nextseq = 0;
334 do {
335 seq = nextseq;
336 flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
337 times->utime = sig->utime;
338 times->stime = sig->stime;
339 times->sum_exec_runtime = sig->sum_sched_runtime;
340
341 for_each_thread(tsk, t) {
342 task_cputime(t, &utime, &stime);
343 times->utime += utime;
344 times->stime += stime;
345 times->sum_exec_runtime += read_sum_exec_runtime(t);
346 }
347 /* If lockless access failed, take the lock. */
348 nextseq = 1;
349 } while (need_seqretry(&sig->stats_lock, seq));
350 done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
351 rcu_read_unlock();
352}
353
354#ifdef CONFIG_IRQ_TIME_ACCOUNTING
355/*
356 * Account a tick to a process and cpustat
357 * @p: the process that the CPU time gets accounted to
358 * @user_tick: is the tick from userspace
359 * @rq: the pointer to rq
360 *
361 * Tick demultiplexing follows the order
362 * - pending hardirq update
363 * - pending softirq update
364 * - user_time
365 * - idle_time
366 * - system time
367 * - check for guest_time
368 * - else account as system_time
369 *
370 * Check for hardirq is done both for system and user time as there is
371 * no timer going off while we are on hardirq and hence we may never get an
372 * opportunity to update it solely in system time.
373 * p->stime and friends are only updated on system time and not on irq
374 * softirq as those do not count in task exec_runtime any more.
375 */
376static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
377 int ticks)
378{
379 u64 other, cputime = TICK_NSEC * ticks;
380
381 /*
382 * When returning from idle, many ticks can get accounted at
383 * once, including some ticks of steal, irq, and softirq time.
384 * Subtract those ticks from the amount of time accounted to
385 * idle, or potentially user or system time. Due to rounding,
386 * other time can exceed ticks occasionally.
387 */
388 other = account_other_time(ULONG_MAX);
389 if (other >= cputime)
390 return;
391
392 cputime -= other;
393
394 if (this_cpu_ksoftirqd() == p) {
395 /*
396 * ksoftirqd time do not get accounted in cpu_softirq_time.
397 * So, we have to handle it separately here.
398 * Also, p->stime needs to be updated for ksoftirqd.
399 */
400 account_system_index_time(p, cputime, CPUTIME_SOFTIRQ);
401 } else if (user_tick) {
402 account_user_time(p, cputime);
403 } else if (p == this_rq()->idle) {
404 account_idle_time(cputime);
405 } else if (p->flags & PF_VCPU) { /* System time or guest time */
406 account_guest_time(p, cputime);
407 } else {
408 account_system_index_time(p, cputime, CPUTIME_SYSTEM);
409 }
410}
411
412static void irqtime_account_idle_ticks(int ticks)
413{
414 irqtime_account_process_tick(current, 0, ticks);
415}
416#else /* CONFIG_IRQ_TIME_ACCOUNTING */
417static inline void irqtime_account_idle_ticks(int ticks) { }
418static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
419 int nr_ticks) { }
420#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
421
422/*
423 * Use precise platform statistics if available:
424 */
425#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
426
427# ifndef __ARCH_HAS_VTIME_TASK_SWITCH
428void vtime_task_switch(struct task_struct *prev)
429{
430 if (is_idle_task(prev))
431 vtime_account_idle(prev);
432 else
433 vtime_account_kernel(prev);
434
435 vtime_flush(prev);
436 arch_vtime_task_switch(prev);
437}
438# endif
439
440void vtime_account_irq(struct task_struct *tsk, unsigned int offset)
441{
442 unsigned int pc = irq_count() - offset;
443
444 if (pc & HARDIRQ_OFFSET) {
445 vtime_account_hardirq(tsk);
446 } else if (pc & SOFTIRQ_OFFSET) {
447 vtime_account_softirq(tsk);
448 } else if (!IS_ENABLED(CONFIG_HAVE_VIRT_CPU_ACCOUNTING_IDLE) &&
449 is_idle_task(tsk)) {
450 vtime_account_idle(tsk);
451 } else {
452 vtime_account_kernel(tsk);
453 }
454}
455
456void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
457 u64 *ut, u64 *st)
458{
459 *ut = curr->utime;
460 *st = curr->stime;
461}
462
463void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
464{
465 *ut = p->utime;
466 *st = p->stime;
467}
468EXPORT_SYMBOL_GPL(task_cputime_adjusted);
469
470void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
471{
472 struct task_cputime cputime;
473
474 thread_group_cputime(p, &cputime);
475
476 *ut = cputime.utime;
477 *st = cputime.stime;
478}
479
480#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE: */
481
482/*
483 * Account a single tick of CPU time.
484 * @p: the process that the CPU time gets accounted to
485 * @user_tick: indicates if the tick is a user or a system tick
486 */
487void account_process_tick(struct task_struct *p, int user_tick)
488{
489 u64 cputime, steal;
490
491 if (vtime_accounting_enabled_this_cpu())
492 return;
493
494 if (sched_clock_irqtime) {
495 irqtime_account_process_tick(p, user_tick, 1);
496 return;
497 }
498
499 cputime = TICK_NSEC;
500 steal = steal_account_process_time(ULONG_MAX);
501
502 if (steal >= cputime)
503 return;
504
505 cputime -= steal;
506
507 if (user_tick)
508 account_user_time(p, cputime);
509 else if ((p != this_rq()->idle) || (irq_count() != HARDIRQ_OFFSET))
510 account_system_time(p, HARDIRQ_OFFSET, cputime);
511 else
512 account_idle_time(cputime);
513}
514
515/*
516 * Account multiple ticks of idle time.
517 * @ticks: number of stolen ticks
518 */
519void account_idle_ticks(unsigned long ticks)
520{
521 u64 cputime, steal;
522
523 if (sched_clock_irqtime) {
524 irqtime_account_idle_ticks(ticks);
525 return;
526 }
527
528 cputime = ticks * TICK_NSEC;
529 steal = steal_account_process_time(ULONG_MAX);
530
531 if (steal >= cputime)
532 return;
533
534 cputime -= steal;
535 account_idle_time(cputime);
536}
537
538/*
539 * Adjust tick based cputime random precision against scheduler runtime
540 * accounting.
541 *
542 * Tick based cputime accounting depend on random scheduling timeslices of a
543 * task to be interrupted or not by the timer. Depending on these
544 * circumstances, the number of these interrupts may be over or
545 * under-optimistic, matching the real user and system cputime with a variable
546 * precision.
547 *
548 * Fix this by scaling these tick based values against the total runtime
549 * accounted by the CFS scheduler.
550 *
551 * This code provides the following guarantees:
552 *
553 * stime + utime == rtime
554 * stime_i+1 >= stime_i, utime_i+1 >= utime_i
555 *
556 * Assuming that rtime_i+1 >= rtime_i.
557 */
558void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
559 u64 *ut, u64 *st)
560{
561 u64 rtime, stime, utime;
562 unsigned long flags;
563
564 /* Serialize concurrent callers such that we can honour our guarantees */
565 raw_spin_lock_irqsave(&prev->lock, flags);
566 rtime = curr->sum_exec_runtime;
567
568 /*
569 * This is possible under two circumstances:
570 * - rtime isn't monotonic after all (a bug);
571 * - we got reordered by the lock.
572 *
573 * In both cases this acts as a filter such that the rest of the code
574 * can assume it is monotonic regardless of anything else.
575 */
576 if (prev->stime + prev->utime >= rtime)
577 goto out;
578
579 stime = curr->stime;
580 utime = curr->utime;
581
582 /*
583 * If either stime or utime are 0, assume all runtime is userspace.
584 * Once a task gets some ticks, the monotonicity code at 'update:'
585 * will ensure things converge to the observed ratio.
586 */
587 if (stime == 0) {
588 utime = rtime;
589 goto update;
590 }
591
592 if (utime == 0) {
593 stime = rtime;
594 goto update;
595 }
596
597 stime = mul_u64_u64_div_u64(stime, rtime, stime + utime);
598
599update:
600 /*
601 * Make sure stime doesn't go backwards; this preserves monotonicity
602 * for utime because rtime is monotonic.
603 *
604 * utime_i+1 = rtime_i+1 - stime_i
605 * = rtime_i+1 - (rtime_i - utime_i)
606 * = (rtime_i+1 - rtime_i) + utime_i
607 * >= utime_i
608 */
609 if (stime < prev->stime)
610 stime = prev->stime;
611 utime = rtime - stime;
612
613 /*
614 * Make sure utime doesn't go backwards; this still preserves
615 * monotonicity for stime, analogous argument to above.
616 */
617 if (utime < prev->utime) {
618 utime = prev->utime;
619 stime = rtime - utime;
620 }
621
622 prev->stime = stime;
623 prev->utime = utime;
624out:
625 *ut = prev->utime;
626 *st = prev->stime;
627 raw_spin_unlock_irqrestore(&prev->lock, flags);
628}
629
630void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
631{
632 struct task_cputime cputime = {
633 .sum_exec_runtime = p->se.sum_exec_runtime,
634 };
635
636 if (task_cputime(p, &cputime.utime, &cputime.stime))
637 cputime.sum_exec_runtime = task_sched_runtime(p);
638 cputime_adjust(&cputime, &p->prev_cputime, ut, st);
639}
640EXPORT_SYMBOL_GPL(task_cputime_adjusted);
641
642void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
643{
644 struct task_cputime cputime;
645
646 thread_group_cputime(p, &cputime);
647 cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
648}
649#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
650
651#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
652static u64 vtime_delta(struct vtime *vtime)
653{
654 unsigned long long clock;
655
656 clock = sched_clock();
657 if (clock < vtime->starttime)
658 return 0;
659
660 return clock - vtime->starttime;
661}
662
663static u64 get_vtime_delta(struct vtime *vtime)
664{
665 u64 delta = vtime_delta(vtime);
666 u64 other;
667
668 /*
669 * Unlike tick based timing, vtime based timing never has lost
670 * ticks, and no need for steal time accounting to make up for
671 * lost ticks. Vtime accounts a rounded version of actual
672 * elapsed time. Limit account_other_time to prevent rounding
673 * errors from causing elapsed vtime to go negative.
674 */
675 other = account_other_time(delta);
676 WARN_ON_ONCE(vtime->state == VTIME_INACTIVE);
677 vtime->starttime += delta;
678
679 return delta - other;
680}
681
682static void vtime_account_system(struct task_struct *tsk,
683 struct vtime *vtime)
684{
685 vtime->stime += get_vtime_delta(vtime);
686 if (vtime->stime >= TICK_NSEC) {
687 account_system_time(tsk, irq_count(), vtime->stime);
688 vtime->stime = 0;
689 }
690}
691
692static void vtime_account_guest(struct task_struct *tsk,
693 struct vtime *vtime)
694{
695 vtime->gtime += get_vtime_delta(vtime);
696 if (vtime->gtime >= TICK_NSEC) {
697 account_guest_time(tsk, vtime->gtime);
698 vtime->gtime = 0;
699 }
700}
701
702static void __vtime_account_kernel(struct task_struct *tsk,
703 struct vtime *vtime)
704{
705 /* We might have scheduled out from guest path */
706 if (vtime->state == VTIME_GUEST)
707 vtime_account_guest(tsk, vtime);
708 else
709 vtime_account_system(tsk, vtime);
710}
711
712void vtime_account_kernel(struct task_struct *tsk)
713{
714 struct vtime *vtime = &tsk->vtime;
715
716 if (!vtime_delta(vtime))
717 return;
718
719 write_seqcount_begin(&vtime->seqcount);
720 __vtime_account_kernel(tsk, vtime);
721 write_seqcount_end(&vtime->seqcount);
722}
723
724void vtime_user_enter(struct task_struct *tsk)
725{
726 struct vtime *vtime = &tsk->vtime;
727
728 write_seqcount_begin(&vtime->seqcount);
729 vtime_account_system(tsk, vtime);
730 vtime->state = VTIME_USER;
731 write_seqcount_end(&vtime->seqcount);
732}
733
734void vtime_user_exit(struct task_struct *tsk)
735{
736 struct vtime *vtime = &tsk->vtime;
737
738 write_seqcount_begin(&vtime->seqcount);
739 vtime->utime += get_vtime_delta(vtime);
740 if (vtime->utime >= TICK_NSEC) {
741 account_user_time(tsk, vtime->utime);
742 vtime->utime = 0;
743 }
744 vtime->state = VTIME_SYS;
745 write_seqcount_end(&vtime->seqcount);
746}
747
748void vtime_guest_enter(struct task_struct *tsk)
749{
750 struct vtime *vtime = &tsk->vtime;
751 /*
752 * The flags must be updated under the lock with
753 * the vtime_starttime flush and update.
754 * That enforces a right ordering and update sequence
755 * synchronization against the reader (task_gtime())
756 * that can thus safely catch up with a tickless delta.
757 */
758 write_seqcount_begin(&vtime->seqcount);
759 vtime_account_system(tsk, vtime);
760 tsk->flags |= PF_VCPU;
761 vtime->state = VTIME_GUEST;
762 write_seqcount_end(&vtime->seqcount);
763}
764EXPORT_SYMBOL_GPL(vtime_guest_enter);
765
766void vtime_guest_exit(struct task_struct *tsk)
767{
768 struct vtime *vtime = &tsk->vtime;
769
770 write_seqcount_begin(&vtime->seqcount);
771 vtime_account_guest(tsk, vtime);
772 tsk->flags &= ~PF_VCPU;
773 vtime->state = VTIME_SYS;
774 write_seqcount_end(&vtime->seqcount);
775}
776EXPORT_SYMBOL_GPL(vtime_guest_exit);
777
778void vtime_account_idle(struct task_struct *tsk)
779{
780 account_idle_time(get_vtime_delta(&tsk->vtime));
781}
782
783void vtime_task_switch_generic(struct task_struct *prev)
784{
785 struct vtime *vtime = &prev->vtime;
786
787 write_seqcount_begin(&vtime->seqcount);
788 if (vtime->state == VTIME_IDLE)
789 vtime_account_idle(prev);
790 else
791 __vtime_account_kernel(prev, vtime);
792 vtime->state = VTIME_INACTIVE;
793 vtime->cpu = -1;
794 write_seqcount_end(&vtime->seqcount);
795
796 vtime = ¤t->vtime;
797
798 write_seqcount_begin(&vtime->seqcount);
799 if (is_idle_task(current))
800 vtime->state = VTIME_IDLE;
801 else if (current->flags & PF_VCPU)
802 vtime->state = VTIME_GUEST;
803 else
804 vtime->state = VTIME_SYS;
805 vtime->starttime = sched_clock();
806 vtime->cpu = smp_processor_id();
807 write_seqcount_end(&vtime->seqcount);
808}
809
810void vtime_init_idle(struct task_struct *t, int cpu)
811{
812 struct vtime *vtime = &t->vtime;
813 unsigned long flags;
814
815 local_irq_save(flags);
816 write_seqcount_begin(&vtime->seqcount);
817 vtime->state = VTIME_IDLE;
818 vtime->starttime = sched_clock();
819 vtime->cpu = cpu;
820 write_seqcount_end(&vtime->seqcount);
821 local_irq_restore(flags);
822}
823
824u64 task_gtime(struct task_struct *t)
825{
826 struct vtime *vtime = &t->vtime;
827 unsigned int seq;
828 u64 gtime;
829
830 if (!vtime_accounting_enabled())
831 return t->gtime;
832
833 do {
834 seq = read_seqcount_begin(&vtime->seqcount);
835
836 gtime = t->gtime;
837 if (vtime->state == VTIME_GUEST)
838 gtime += vtime->gtime + vtime_delta(vtime);
839
840 } while (read_seqcount_retry(&vtime->seqcount, seq));
841
842 return gtime;
843}
844
845/*
846 * Fetch cputime raw values from fields of task_struct and
847 * add up the pending nohz execution time since the last
848 * cputime snapshot.
849 */
850bool task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
851{
852 struct vtime *vtime = &t->vtime;
853 unsigned int seq;
854 u64 delta;
855 int ret;
856
857 if (!vtime_accounting_enabled()) {
858 *utime = t->utime;
859 *stime = t->stime;
860 return false;
861 }
862
863 do {
864 ret = false;
865 seq = read_seqcount_begin(&vtime->seqcount);
866
867 *utime = t->utime;
868 *stime = t->stime;
869
870 /* Task is sleeping or idle, nothing to add */
871 if (vtime->state < VTIME_SYS)
872 continue;
873
874 ret = true;
875 delta = vtime_delta(vtime);
876
877 /*
878 * Task runs either in user (including guest) or kernel space,
879 * add pending nohz time to the right place.
880 */
881 if (vtime->state == VTIME_SYS)
882 *stime += vtime->stime + delta;
883 else
884 *utime += vtime->utime + delta;
885 } while (read_seqcount_retry(&vtime->seqcount, seq));
886
887 return ret;
888}
889
890static int vtime_state_fetch(struct vtime *vtime, int cpu)
891{
892 int state = READ_ONCE(vtime->state);
893
894 /*
895 * We raced against a context switch, fetch the
896 * kcpustat task again.
897 */
898 if (vtime->cpu != cpu && vtime->cpu != -1)
899 return -EAGAIN;
900
901 /*
902 * Two possible things here:
903 * 1) We are seeing the scheduling out task (prev) or any past one.
904 * 2) We are seeing the scheduling in task (next) but it hasn't
905 * passed though vtime_task_switch() yet so the pending
906 * cputime of the prev task may not be flushed yet.
907 *
908 * Case 1) is ok but 2) is not. So wait for a safe VTIME state.
909 */
910 if (state == VTIME_INACTIVE)
911 return -EAGAIN;
912
913 return state;
914}
915
916static u64 kcpustat_user_vtime(struct vtime *vtime)
917{
918 if (vtime->state == VTIME_USER)
919 return vtime->utime + vtime_delta(vtime);
920 else if (vtime->state == VTIME_GUEST)
921 return vtime->gtime + vtime_delta(vtime);
922 return 0;
923}
924
925static int kcpustat_field_vtime(u64 *cpustat,
926 struct task_struct *tsk,
927 enum cpu_usage_stat usage,
928 int cpu, u64 *val)
929{
930 struct vtime *vtime = &tsk->vtime;
931 unsigned int seq;
932
933 do {
934 int state;
935
936 seq = read_seqcount_begin(&vtime->seqcount);
937
938 state = vtime_state_fetch(vtime, cpu);
939 if (state < 0)
940 return state;
941
942 *val = cpustat[usage];
943
944 /*
945 * Nice VS unnice cputime accounting may be inaccurate if
946 * the nice value has changed since the last vtime update.
947 * But proper fix would involve interrupting target on nice
948 * updates which is a no go on nohz_full (although the scheduler
949 * may still interrupt the target if rescheduling is needed...)
950 */
951 switch (usage) {
952 case CPUTIME_SYSTEM:
953 if (state == VTIME_SYS)
954 *val += vtime->stime + vtime_delta(vtime);
955 break;
956 case CPUTIME_USER:
957 if (task_nice(tsk) <= 0)
958 *val += kcpustat_user_vtime(vtime);
959 break;
960 case CPUTIME_NICE:
961 if (task_nice(tsk) > 0)
962 *val += kcpustat_user_vtime(vtime);
963 break;
964 case CPUTIME_GUEST:
965 if (state == VTIME_GUEST && task_nice(tsk) <= 0)
966 *val += vtime->gtime + vtime_delta(vtime);
967 break;
968 case CPUTIME_GUEST_NICE:
969 if (state == VTIME_GUEST && task_nice(tsk) > 0)
970 *val += vtime->gtime + vtime_delta(vtime);
971 break;
972 default:
973 break;
974 }
975 } while (read_seqcount_retry(&vtime->seqcount, seq));
976
977 return 0;
978}
979
980u64 kcpustat_field(struct kernel_cpustat *kcpustat,
981 enum cpu_usage_stat usage, int cpu)
982{
983 u64 *cpustat = kcpustat->cpustat;
984 u64 val = cpustat[usage];
985 struct rq *rq;
986 int err;
987
988 if (!vtime_accounting_enabled_cpu(cpu))
989 return val;
990
991 rq = cpu_rq(cpu);
992
993 for (;;) {
994 struct task_struct *curr;
995
996 rcu_read_lock();
997 curr = rcu_dereference(rq->curr);
998 if (WARN_ON_ONCE(!curr)) {
999 rcu_read_unlock();
1000 return cpustat[usage];
1001 }
1002
1003 err = kcpustat_field_vtime(cpustat, curr, usage, cpu, &val);
1004 rcu_read_unlock();
1005
1006 if (!err)
1007 return val;
1008
1009 cpu_relax();
1010 }
1011}
1012EXPORT_SYMBOL_GPL(kcpustat_field);
1013
1014static int kcpustat_cpu_fetch_vtime(struct kernel_cpustat *dst,
1015 const struct kernel_cpustat *src,
1016 struct task_struct *tsk, int cpu)
1017{
1018 struct vtime *vtime = &tsk->vtime;
1019 unsigned int seq;
1020
1021 do {
1022 u64 *cpustat;
1023 u64 delta;
1024 int state;
1025
1026 seq = read_seqcount_begin(&vtime->seqcount);
1027
1028 state = vtime_state_fetch(vtime, cpu);
1029 if (state < 0)
1030 return state;
1031
1032 *dst = *src;
1033 cpustat = dst->cpustat;
1034
1035 /* Task is sleeping, dead or idle, nothing to add */
1036 if (state < VTIME_SYS)
1037 continue;
1038
1039 delta = vtime_delta(vtime);
1040
1041 /*
1042 * Task runs either in user (including guest) or kernel space,
1043 * add pending nohz time to the right place.
1044 */
1045 if (state == VTIME_SYS) {
1046 cpustat[CPUTIME_SYSTEM] += vtime->stime + delta;
1047 } else if (state == VTIME_USER) {
1048 if (task_nice(tsk) > 0)
1049 cpustat[CPUTIME_NICE] += vtime->utime + delta;
1050 else
1051 cpustat[CPUTIME_USER] += vtime->utime + delta;
1052 } else {
1053 WARN_ON_ONCE(state != VTIME_GUEST);
1054 if (task_nice(tsk) > 0) {
1055 cpustat[CPUTIME_GUEST_NICE] += vtime->gtime + delta;
1056 cpustat[CPUTIME_NICE] += vtime->gtime + delta;
1057 } else {
1058 cpustat[CPUTIME_GUEST] += vtime->gtime + delta;
1059 cpustat[CPUTIME_USER] += vtime->gtime + delta;
1060 }
1061 }
1062 } while (read_seqcount_retry(&vtime->seqcount, seq));
1063
1064 return 0;
1065}
1066
1067void kcpustat_cpu_fetch(struct kernel_cpustat *dst, int cpu)
1068{
1069 const struct kernel_cpustat *src = &kcpustat_cpu(cpu);
1070 struct rq *rq;
1071 int err;
1072
1073 if (!vtime_accounting_enabled_cpu(cpu)) {
1074 *dst = *src;
1075 return;
1076 }
1077
1078 rq = cpu_rq(cpu);
1079
1080 for (;;) {
1081 struct task_struct *curr;
1082
1083 rcu_read_lock();
1084 curr = rcu_dereference(rq->curr);
1085 if (WARN_ON_ONCE(!curr)) {
1086 rcu_read_unlock();
1087 *dst = *src;
1088 return;
1089 }
1090
1091 err = kcpustat_cpu_fetch_vtime(dst, src, curr, cpu);
1092 rcu_read_unlock();
1093
1094 if (!err)
1095 return;
1096
1097 cpu_relax();
1098 }
1099}
1100EXPORT_SYMBOL_GPL(kcpustat_cpu_fetch);
1101
1102#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */