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