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 before 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)
51{
52 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
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
63 /*
64 * We do not account for softirq time from ksoftirqd here.
65 * We want to continue accounting softirq time to ksoftirqd thread
66 * in that case, so as not to confuse scheduler with a special task
67 * that do not consume any time, but still wants to run.
68 */
69 if (hardirq_count())
70 irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
71 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
72 irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
73}
74EXPORT_SYMBOL_GPL(irqtime_account_irq);
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 cpustat[CPUTIME_NICE] += cputime;
151 cpustat[CPUTIME_GUEST_NICE] += cputime;
152 } else {
153 cpustat[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 * When a guest is interrupted for a longer amount of time, missed clock
231 * ticks are not redelivered later. Due to that, this function may on
232 * occasion account more time than the calling functions think elapsed.
233 */
234static __always_inline u64 steal_account_process_time(u64 maxtime)
235{
236#ifdef CONFIG_PARAVIRT
237 if (static_key_false(¶virt_steal_enabled)) {
238 u64 steal;
239
240 steal = paravirt_steal_clock(smp_processor_id());
241 steal -= this_rq()->prev_steal_time;
242 steal = min(steal, maxtime);
243 account_steal_time(steal);
244 this_rq()->prev_steal_time += steal;
245
246 return steal;
247 }
248#endif
249 return 0;
250}
251
252/*
253 * Account how much elapsed time was spent in steal, irq, or softirq time.
254 */
255static inline u64 account_other_time(u64 max)
256{
257 u64 accounted;
258
259 lockdep_assert_irqs_disabled();
260
261 accounted = steal_account_process_time(max);
262
263 if (accounted < max)
264 accounted += irqtime_tick_accounted(max - accounted);
265
266 return accounted;
267}
268
269#ifdef CONFIG_64BIT
270static inline u64 read_sum_exec_runtime(struct task_struct *t)
271{
272 return t->se.sum_exec_runtime;
273}
274#else
275static u64 read_sum_exec_runtime(struct task_struct *t)
276{
277 u64 ns;
278 struct rq_flags rf;
279 struct rq *rq;
280
281 rq = task_rq_lock(t, &rf);
282 ns = t->se.sum_exec_runtime;
283 task_rq_unlock(rq, t, &rf);
284
285 return ns;
286}
287#endif
288
289/*
290 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
291 * tasks (sum on group iteration) belonging to @tsk's group.
292 */
293void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
294{
295 struct signal_struct *sig = tsk->signal;
296 u64 utime, stime;
297 struct task_struct *t;
298 unsigned int seq, nextseq;
299 unsigned long flags;
300
301 /*
302 * Update current task runtime to account pending time since last
303 * scheduler action or thread_group_cputime() call. This thread group
304 * might have other running tasks on different CPUs, but updating
305 * their runtime can affect syscall performance, so we skip account
306 * those pending times and rely only on values updated on tick or
307 * other scheduler action.
308 */
309 if (same_thread_group(current, tsk))
310 (void) task_sched_runtime(current);
311
312 rcu_read_lock();
313 /* Attempt a lockless read on the first round. */
314 nextseq = 0;
315 do {
316 seq = nextseq;
317 flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
318 times->utime = sig->utime;
319 times->stime = sig->stime;
320 times->sum_exec_runtime = sig->sum_sched_runtime;
321
322 for_each_thread(tsk, t) {
323 task_cputime(t, &utime, &stime);
324 times->utime += utime;
325 times->stime += stime;
326 times->sum_exec_runtime += read_sum_exec_runtime(t);
327 }
328 /* If lockless access failed, take the lock. */
329 nextseq = 1;
330 } while (need_seqretry(&sig->stats_lock, seq));
331 done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
332 rcu_read_unlock();
333}
334
335#ifdef CONFIG_IRQ_TIME_ACCOUNTING
336/*
337 * Account a tick to a process and cpustat
338 * @p: the process that the CPU time gets accounted to
339 * @user_tick: is the tick from userspace
340 * @rq: the pointer to rq
341 *
342 * Tick demultiplexing follows the order
343 * - pending hardirq update
344 * - pending softirq update
345 * - user_time
346 * - idle_time
347 * - system time
348 * - check for guest_time
349 * - else account as system_time
350 *
351 * Check for hardirq is done both for system and user time as there is
352 * no timer going off while we are on hardirq and hence we may never get an
353 * opportunity to update it solely in system time.
354 * p->stime and friends are only updated on system time and not on irq
355 * softirq as those do not count in task exec_runtime any more.
356 */
357static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
358 int ticks)
359{
360 u64 other, cputime = TICK_NSEC * ticks;
361
362 /*
363 * When returning from idle, many ticks can get accounted at
364 * once, including some ticks of steal, irq, and softirq time.
365 * Subtract those ticks from the amount of time accounted to
366 * idle, or potentially user or system time. Due to rounding,
367 * other time can exceed ticks occasionally.
368 */
369 other = account_other_time(ULONG_MAX);
370 if (other >= cputime)
371 return;
372
373 cputime -= other;
374
375 if (this_cpu_ksoftirqd() == p) {
376 /*
377 * ksoftirqd time do not get accounted in cpu_softirq_time.
378 * So, we have to handle it separately here.
379 * Also, p->stime needs to be updated for ksoftirqd.
380 */
381 account_system_index_time(p, cputime, CPUTIME_SOFTIRQ);
382 } else if (user_tick) {
383 account_user_time(p, cputime);
384 } else if (p == this_rq()->idle) {
385 account_idle_time(cputime);
386 } else if (p->flags & PF_VCPU) { /* System time or guest time */
387 account_guest_time(p, cputime);
388 } else {
389 account_system_index_time(p, cputime, CPUTIME_SYSTEM);
390 }
391}
392
393static void irqtime_account_idle_ticks(int ticks)
394{
395 irqtime_account_process_tick(current, 0, ticks);
396}
397#else /* CONFIG_IRQ_TIME_ACCOUNTING */
398static inline void irqtime_account_idle_ticks(int ticks) { }
399static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
400 int nr_ticks) { }
401#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
402
403/*
404 * Use precise platform statistics if available:
405 */
406#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
407
408# ifndef __ARCH_HAS_VTIME_TASK_SWITCH
409void vtime_task_switch(struct task_struct *prev)
410{
411 if (is_idle_task(prev))
412 vtime_account_idle(prev);
413 else
414 vtime_account_kernel(prev);
415
416 vtime_flush(prev);
417 arch_vtime_task_switch(prev);
418}
419# endif
420
421/*
422 * Archs that account the whole time spent in the idle task
423 * (outside irq) as idle time can rely on this and just implement
424 * vtime_account_kernel() and vtime_account_idle(). Archs that
425 * have other meaning of the idle time (s390 only includes the
426 * time spent by the CPU when it's in low power mode) must override
427 * vtime_account().
428 */
429#ifndef __ARCH_HAS_VTIME_ACCOUNT
430void vtime_account_irq_enter(struct task_struct *tsk)
431{
432 if (!in_interrupt() && is_idle_task(tsk))
433 vtime_account_idle(tsk);
434 else
435 vtime_account_kernel(tsk);
436}
437EXPORT_SYMBOL_GPL(vtime_account_irq_enter);
438#endif /* __ARCH_HAS_VTIME_ACCOUNT */
439
440void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
441 u64 *ut, u64 *st)
442{
443 *ut = curr->utime;
444 *st = curr->stime;
445}
446
447void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
448{
449 *ut = p->utime;
450 *st = p->stime;
451}
452EXPORT_SYMBOL_GPL(task_cputime_adjusted);
453
454void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
455{
456 struct task_cputime cputime;
457
458 thread_group_cputime(p, &cputime);
459
460 *ut = cputime.utime;
461 *st = cputime.stime;
462}
463
464#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE: */
465
466/*
467 * Account a single tick of CPU time.
468 * @p: the process that the CPU time gets accounted to
469 * @user_tick: indicates if the tick is a user or a system tick
470 */
471void account_process_tick(struct task_struct *p, int user_tick)
472{
473 u64 cputime, steal;
474
475 if (vtime_accounting_enabled_this_cpu())
476 return;
477
478 if (sched_clock_irqtime) {
479 irqtime_account_process_tick(p, user_tick, 1);
480 return;
481 }
482
483 cputime = TICK_NSEC;
484 steal = steal_account_process_time(ULONG_MAX);
485
486 if (steal >= cputime)
487 return;
488
489 cputime -= steal;
490
491 if (user_tick)
492 account_user_time(p, cputime);
493 else if ((p != this_rq()->idle) || (irq_count() != HARDIRQ_OFFSET))
494 account_system_time(p, HARDIRQ_OFFSET, cputime);
495 else
496 account_idle_time(cputime);
497}
498
499/*
500 * Account multiple ticks of idle time.
501 * @ticks: number of stolen ticks
502 */
503void account_idle_ticks(unsigned long ticks)
504{
505 u64 cputime, steal;
506
507 if (sched_clock_irqtime) {
508 irqtime_account_idle_ticks(ticks);
509 return;
510 }
511
512 cputime = ticks * TICK_NSEC;
513 steal = steal_account_process_time(ULONG_MAX);
514
515 if (steal >= cputime)
516 return;
517
518 cputime -= steal;
519 account_idle_time(cputime);
520}
521
522/*
523 * Adjust tick based cputime random precision against scheduler runtime
524 * accounting.
525 *
526 * Tick based cputime accounting depend on random scheduling timeslices of a
527 * task to be interrupted or not by the timer. Depending on these
528 * circumstances, the number of these interrupts may be over or
529 * under-optimistic, matching the real user and system cputime with a variable
530 * precision.
531 *
532 * Fix this by scaling these tick based values against the total runtime
533 * accounted by the CFS scheduler.
534 *
535 * This code provides the following guarantees:
536 *
537 * stime + utime == rtime
538 * stime_i+1 >= stime_i, utime_i+1 >= utime_i
539 *
540 * Assuming that rtime_i+1 >= rtime_i.
541 */
542void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
543 u64 *ut, u64 *st)
544{
545 u64 rtime, stime, utime;
546 unsigned long flags;
547
548 /* Serialize concurrent callers such that we can honour our guarantees */
549 raw_spin_lock_irqsave(&prev->lock, flags);
550 rtime = curr->sum_exec_runtime;
551
552 /*
553 * This is possible under two circumstances:
554 * - rtime isn't monotonic after all (a bug);
555 * - we got reordered by the lock.
556 *
557 * In both cases this acts as a filter such that the rest of the code
558 * can assume it is monotonic regardless of anything else.
559 */
560 if (prev->stime + prev->utime >= rtime)
561 goto out;
562
563 stime = curr->stime;
564 utime = curr->utime;
565
566 /*
567 * If either stime or utime are 0, assume all runtime is userspace.
568 * Once a task gets some ticks, the monotonicy code at 'update:'
569 * will ensure things converge to the observed ratio.
570 */
571 if (stime == 0) {
572 utime = rtime;
573 goto update;
574 }
575
576 if (utime == 0) {
577 stime = rtime;
578 goto update;
579 }
580
581 stime = mul_u64_u64_div_u64(stime, rtime, stime + utime);
582
583update:
584 /*
585 * Make sure stime doesn't go backwards; this preserves monotonicity
586 * for utime because rtime is monotonic.
587 *
588 * utime_i+1 = rtime_i+1 - stime_i
589 * = rtime_i+1 - (rtime_i - utime_i)
590 * = (rtime_i+1 - rtime_i) + utime_i
591 * >= utime_i
592 */
593 if (stime < prev->stime)
594 stime = prev->stime;
595 utime = rtime - stime;
596
597 /*
598 * Make sure utime doesn't go backwards; this still preserves
599 * monotonicity for stime, analogous argument to above.
600 */
601 if (utime < prev->utime) {
602 utime = prev->utime;
603 stime = rtime - utime;
604 }
605
606 prev->stime = stime;
607 prev->utime = utime;
608out:
609 *ut = prev->utime;
610 *st = prev->stime;
611 raw_spin_unlock_irqrestore(&prev->lock, flags);
612}
613
614void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
615{
616 struct task_cputime cputime = {
617 .sum_exec_runtime = p->se.sum_exec_runtime,
618 };
619
620 task_cputime(p, &cputime.utime, &cputime.stime);
621 cputime_adjust(&cputime, &p->prev_cputime, ut, st);
622}
623EXPORT_SYMBOL_GPL(task_cputime_adjusted);
624
625void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
626{
627 struct task_cputime cputime;
628
629 thread_group_cputime(p, &cputime);
630 cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
631}
632#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
633
634#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
635static u64 vtime_delta(struct vtime *vtime)
636{
637 unsigned long long clock;
638
639 clock = sched_clock();
640 if (clock < vtime->starttime)
641 return 0;
642
643 return clock - vtime->starttime;
644}
645
646static u64 get_vtime_delta(struct vtime *vtime)
647{
648 u64 delta = vtime_delta(vtime);
649 u64 other;
650
651 /*
652 * Unlike tick based timing, vtime based timing never has lost
653 * ticks, and no need for steal time accounting to make up for
654 * lost ticks. Vtime accounts a rounded version of actual
655 * elapsed time. Limit account_other_time to prevent rounding
656 * errors from causing elapsed vtime to go negative.
657 */
658 other = account_other_time(delta);
659 WARN_ON_ONCE(vtime->state == VTIME_INACTIVE);
660 vtime->starttime += delta;
661
662 return delta - other;
663}
664
665static void vtime_account_system(struct task_struct *tsk,
666 struct vtime *vtime)
667{
668 vtime->stime += get_vtime_delta(vtime);
669 if (vtime->stime >= TICK_NSEC) {
670 account_system_time(tsk, irq_count(), vtime->stime);
671 vtime->stime = 0;
672 }
673}
674
675static void vtime_account_guest(struct task_struct *tsk,
676 struct vtime *vtime)
677{
678 vtime->gtime += get_vtime_delta(vtime);
679 if (vtime->gtime >= TICK_NSEC) {
680 account_guest_time(tsk, vtime->gtime);
681 vtime->gtime = 0;
682 }
683}
684
685static void __vtime_account_kernel(struct task_struct *tsk,
686 struct vtime *vtime)
687{
688 /* We might have scheduled out from guest path */
689 if (vtime->state == VTIME_GUEST)
690 vtime_account_guest(tsk, vtime);
691 else
692 vtime_account_system(tsk, vtime);
693}
694
695void vtime_account_kernel(struct task_struct *tsk)
696{
697 struct vtime *vtime = &tsk->vtime;
698
699 if (!vtime_delta(vtime))
700 return;
701
702 write_seqcount_begin(&vtime->seqcount);
703 __vtime_account_kernel(tsk, vtime);
704 write_seqcount_end(&vtime->seqcount);
705}
706
707void vtime_user_enter(struct task_struct *tsk)
708{
709 struct vtime *vtime = &tsk->vtime;
710
711 write_seqcount_begin(&vtime->seqcount);
712 vtime_account_system(tsk, vtime);
713 vtime->state = VTIME_USER;
714 write_seqcount_end(&vtime->seqcount);
715}
716
717void vtime_user_exit(struct task_struct *tsk)
718{
719 struct vtime *vtime = &tsk->vtime;
720
721 write_seqcount_begin(&vtime->seqcount);
722 vtime->utime += get_vtime_delta(vtime);
723 if (vtime->utime >= TICK_NSEC) {
724 account_user_time(tsk, vtime->utime);
725 vtime->utime = 0;
726 }
727 vtime->state = VTIME_SYS;
728 write_seqcount_end(&vtime->seqcount);
729}
730
731void vtime_guest_enter(struct task_struct *tsk)
732{
733 struct vtime *vtime = &tsk->vtime;
734 /*
735 * The flags must be updated under the lock with
736 * the vtime_starttime flush and update.
737 * That enforces a right ordering and update sequence
738 * synchronization against the reader (task_gtime())
739 * that can thus safely catch up with a tickless delta.
740 */
741 write_seqcount_begin(&vtime->seqcount);
742 vtime_account_system(tsk, vtime);
743 tsk->flags |= PF_VCPU;
744 vtime->state = VTIME_GUEST;
745 write_seqcount_end(&vtime->seqcount);
746}
747EXPORT_SYMBOL_GPL(vtime_guest_enter);
748
749void vtime_guest_exit(struct task_struct *tsk)
750{
751 struct vtime *vtime = &tsk->vtime;
752
753 write_seqcount_begin(&vtime->seqcount);
754 vtime_account_guest(tsk, vtime);
755 tsk->flags &= ~PF_VCPU;
756 vtime->state = VTIME_SYS;
757 write_seqcount_end(&vtime->seqcount);
758}
759EXPORT_SYMBOL_GPL(vtime_guest_exit);
760
761void vtime_account_idle(struct task_struct *tsk)
762{
763 account_idle_time(get_vtime_delta(&tsk->vtime));
764}
765
766void vtime_task_switch_generic(struct task_struct *prev)
767{
768 struct vtime *vtime = &prev->vtime;
769
770 write_seqcount_begin(&vtime->seqcount);
771 if (vtime->state == VTIME_IDLE)
772 vtime_account_idle(prev);
773 else
774 __vtime_account_kernel(prev, vtime);
775 vtime->state = VTIME_INACTIVE;
776 vtime->cpu = -1;
777 write_seqcount_end(&vtime->seqcount);
778
779 vtime = ¤t->vtime;
780
781 write_seqcount_begin(&vtime->seqcount);
782 if (is_idle_task(current))
783 vtime->state = VTIME_IDLE;
784 else if (current->flags & PF_VCPU)
785 vtime->state = VTIME_GUEST;
786 else
787 vtime->state = VTIME_SYS;
788 vtime->starttime = sched_clock();
789 vtime->cpu = smp_processor_id();
790 write_seqcount_end(&vtime->seqcount);
791}
792
793void vtime_init_idle(struct task_struct *t, int cpu)
794{
795 struct vtime *vtime = &t->vtime;
796 unsigned long flags;
797
798 local_irq_save(flags);
799 write_seqcount_begin(&vtime->seqcount);
800 vtime->state = VTIME_IDLE;
801 vtime->starttime = sched_clock();
802 vtime->cpu = cpu;
803 write_seqcount_end(&vtime->seqcount);
804 local_irq_restore(flags);
805}
806
807u64 task_gtime(struct task_struct *t)
808{
809 struct vtime *vtime = &t->vtime;
810 unsigned int seq;
811 u64 gtime;
812
813 if (!vtime_accounting_enabled())
814 return t->gtime;
815
816 do {
817 seq = read_seqcount_begin(&vtime->seqcount);
818
819 gtime = t->gtime;
820 if (vtime->state == VTIME_GUEST)
821 gtime += vtime->gtime + vtime_delta(vtime);
822
823 } while (read_seqcount_retry(&vtime->seqcount, seq));
824
825 return gtime;
826}
827
828/*
829 * Fetch cputime raw values from fields of task_struct and
830 * add up the pending nohz execution time since the last
831 * cputime snapshot.
832 */
833void task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
834{
835 struct vtime *vtime = &t->vtime;
836 unsigned int seq;
837 u64 delta;
838
839 if (!vtime_accounting_enabled()) {
840 *utime = t->utime;
841 *stime = t->stime;
842 return;
843 }
844
845 do {
846 seq = read_seqcount_begin(&vtime->seqcount);
847
848 *utime = t->utime;
849 *stime = t->stime;
850
851 /* Task is sleeping or idle, nothing to add */
852 if (vtime->state < VTIME_SYS)
853 continue;
854
855 delta = vtime_delta(vtime);
856
857 /*
858 * Task runs either in user (including guest) or kernel space,
859 * add pending nohz time to the right place.
860 */
861 if (vtime->state == VTIME_SYS)
862 *stime += vtime->stime + delta;
863 else
864 *utime += vtime->utime + delta;
865 } while (read_seqcount_retry(&vtime->seqcount, seq));
866}
867
868static int vtime_state_fetch(struct vtime *vtime, int cpu)
869{
870 int state = READ_ONCE(vtime->state);
871
872 /*
873 * We raced against a context switch, fetch the
874 * kcpustat task again.
875 */
876 if (vtime->cpu != cpu && vtime->cpu != -1)
877 return -EAGAIN;
878
879 /*
880 * Two possible things here:
881 * 1) We are seeing the scheduling out task (prev) or any past one.
882 * 2) We are seeing the scheduling in task (next) but it hasn't
883 * passed though vtime_task_switch() yet so the pending
884 * cputime of the prev task may not be flushed yet.
885 *
886 * Case 1) is ok but 2) is not. So wait for a safe VTIME state.
887 */
888 if (state == VTIME_INACTIVE)
889 return -EAGAIN;
890
891 return state;
892}
893
894static u64 kcpustat_user_vtime(struct vtime *vtime)
895{
896 if (vtime->state == VTIME_USER)
897 return vtime->utime + vtime_delta(vtime);
898 else if (vtime->state == VTIME_GUEST)
899 return vtime->gtime + vtime_delta(vtime);
900 return 0;
901}
902
903static int kcpustat_field_vtime(u64 *cpustat,
904 struct task_struct *tsk,
905 enum cpu_usage_stat usage,
906 int cpu, u64 *val)
907{
908 struct vtime *vtime = &tsk->vtime;
909 unsigned int seq;
910
911 do {
912 int state;
913
914 seq = read_seqcount_begin(&vtime->seqcount);
915
916 state = vtime_state_fetch(vtime, cpu);
917 if (state < 0)
918 return state;
919
920 *val = cpustat[usage];
921
922 /*
923 * Nice VS unnice cputime accounting may be inaccurate if
924 * the nice value has changed since the last vtime update.
925 * But proper fix would involve interrupting target on nice
926 * updates which is a no go on nohz_full (although the scheduler
927 * may still interrupt the target if rescheduling is needed...)
928 */
929 switch (usage) {
930 case CPUTIME_SYSTEM:
931 if (state == VTIME_SYS)
932 *val += vtime->stime + vtime_delta(vtime);
933 break;
934 case CPUTIME_USER:
935 if (task_nice(tsk) <= 0)
936 *val += kcpustat_user_vtime(vtime);
937 break;
938 case CPUTIME_NICE:
939 if (task_nice(tsk) > 0)
940 *val += kcpustat_user_vtime(vtime);
941 break;
942 case CPUTIME_GUEST:
943 if (state == VTIME_GUEST && task_nice(tsk) <= 0)
944 *val += vtime->gtime + vtime_delta(vtime);
945 break;
946 case CPUTIME_GUEST_NICE:
947 if (state == VTIME_GUEST && task_nice(tsk) > 0)
948 *val += vtime->gtime + vtime_delta(vtime);
949 break;
950 default:
951 break;
952 }
953 } while (read_seqcount_retry(&vtime->seqcount, seq));
954
955 return 0;
956}
957
958u64 kcpustat_field(struct kernel_cpustat *kcpustat,
959 enum cpu_usage_stat usage, int cpu)
960{
961 u64 *cpustat = kcpustat->cpustat;
962 u64 val = cpustat[usage];
963 struct rq *rq;
964 int err;
965
966 if (!vtime_accounting_enabled_cpu(cpu))
967 return val;
968
969 rq = cpu_rq(cpu);
970
971 for (;;) {
972 struct task_struct *curr;
973
974 rcu_read_lock();
975 curr = rcu_dereference(rq->curr);
976 if (WARN_ON_ONCE(!curr)) {
977 rcu_read_unlock();
978 return cpustat[usage];
979 }
980
981 err = kcpustat_field_vtime(cpustat, curr, usage, cpu, &val);
982 rcu_read_unlock();
983
984 if (!err)
985 return val;
986
987 cpu_relax();
988 }
989}
990EXPORT_SYMBOL_GPL(kcpustat_field);
991
992static int kcpustat_cpu_fetch_vtime(struct kernel_cpustat *dst,
993 const struct kernel_cpustat *src,
994 struct task_struct *tsk, int cpu)
995{
996 struct vtime *vtime = &tsk->vtime;
997 unsigned int seq;
998
999 do {
1000 u64 *cpustat;
1001 u64 delta;
1002 int state;
1003
1004 seq = read_seqcount_begin(&vtime->seqcount);
1005
1006 state = vtime_state_fetch(vtime, cpu);
1007 if (state < 0)
1008 return state;
1009
1010 *dst = *src;
1011 cpustat = dst->cpustat;
1012
1013 /* Task is sleeping, dead or idle, nothing to add */
1014 if (state < VTIME_SYS)
1015 continue;
1016
1017 delta = vtime_delta(vtime);
1018
1019 /*
1020 * Task runs either in user (including guest) or kernel space,
1021 * add pending nohz time to the right place.
1022 */
1023 if (state == VTIME_SYS) {
1024 cpustat[CPUTIME_SYSTEM] += vtime->stime + delta;
1025 } else if (state == VTIME_USER) {
1026 if (task_nice(tsk) > 0)
1027 cpustat[CPUTIME_NICE] += vtime->utime + delta;
1028 else
1029 cpustat[CPUTIME_USER] += vtime->utime + delta;
1030 } else {
1031 WARN_ON_ONCE(state != VTIME_GUEST);
1032 if (task_nice(tsk) > 0) {
1033 cpustat[CPUTIME_GUEST_NICE] += vtime->gtime + delta;
1034 cpustat[CPUTIME_NICE] += vtime->gtime + delta;
1035 } else {
1036 cpustat[CPUTIME_GUEST] += vtime->gtime + delta;
1037 cpustat[CPUTIME_USER] += vtime->gtime + delta;
1038 }
1039 }
1040 } while (read_seqcount_retry(&vtime->seqcount, seq));
1041
1042 return 0;
1043}
1044
1045void kcpustat_cpu_fetch(struct kernel_cpustat *dst, int cpu)
1046{
1047 const struct kernel_cpustat *src = &kcpustat_cpu(cpu);
1048 struct rq *rq;
1049 int err;
1050
1051 if (!vtime_accounting_enabled_cpu(cpu)) {
1052 *dst = *src;
1053 return;
1054 }
1055
1056 rq = cpu_rq(cpu);
1057
1058 for (;;) {
1059 struct task_struct *curr;
1060
1061 rcu_read_lock();
1062 curr = rcu_dereference(rq->curr);
1063 if (WARN_ON_ONCE(!curr)) {
1064 rcu_read_unlock();
1065 *dst = *src;
1066 return;
1067 }
1068
1069 err = kcpustat_cpu_fetch_vtime(dst, src, curr, cpu);
1070 rcu_read_unlock();
1071
1072 if (!err)
1073 return;
1074
1075 cpu_relax();
1076 }
1077}
1078EXPORT_SYMBOL_GPL(kcpustat_cpu_fetch);
1079
1080#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */