Loading...
1/*
2 * Simple CPU accounting cgroup controller
3 */
4#include "sched.h"
5
6#ifdef CONFIG_IRQ_TIME_ACCOUNTING
7
8/*
9 * There are no locks covering percpu hardirq/softirq time.
10 * They are only modified in vtime_account, on corresponding CPU
11 * with interrupts disabled. So, writes are safe.
12 * They are read and saved off onto struct rq in update_rq_clock().
13 * This may result in other CPU reading this CPU's irq time and can
14 * race with irq/vtime_account on this CPU. We would either get old
15 * or new value with a side effect of accounting a slice of irq time to wrong
16 * task when irq is in progress while we read rq->clock. That is a worthy
17 * compromise in place of having locks on each irq in account_system_time.
18 */
19DEFINE_PER_CPU(struct irqtime, cpu_irqtime);
20
21static int sched_clock_irqtime;
22
23void enable_sched_clock_irqtime(void)
24{
25 sched_clock_irqtime = 1;
26}
27
28void disable_sched_clock_irqtime(void)
29{
30 sched_clock_irqtime = 0;
31}
32
33static void irqtime_account_delta(struct irqtime *irqtime, u64 delta,
34 enum cpu_usage_stat idx)
35{
36 u64 *cpustat = kcpustat_this_cpu->cpustat;
37
38 u64_stats_update_begin(&irqtime->sync);
39 cpustat[idx] += delta;
40 irqtime->total += delta;
41 irqtime->tick_delta += delta;
42 u64_stats_update_end(&irqtime->sync);
43}
44
45/*
46 * Called before incrementing preempt_count on {soft,}irq_enter
47 * and before decrementing preempt_count on {soft,}irq_exit.
48 */
49void irqtime_account_irq(struct task_struct *curr)
50{
51 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
52 s64 delta;
53 int cpu;
54
55 if (!sched_clock_irqtime)
56 return;
57
58 cpu = smp_processor_id();
59 delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
60 irqtime->irq_start_time += delta;
61
62 /*
63 * We do not account for softirq time from ksoftirqd here.
64 * We want to continue accounting softirq time to ksoftirqd thread
65 * in that case, so as not to confuse scheduler with a special task
66 * that do not consume any time, but still wants to run.
67 */
68 if (hardirq_count())
69 irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
70 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
71 irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
72}
73EXPORT_SYMBOL_GPL(irqtime_account_irq);
74
75static u64 irqtime_tick_accounted(u64 maxtime)
76{
77 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
78 u64 delta;
79
80 delta = min(irqtime->tick_delta, maxtime);
81 irqtime->tick_delta -= delta;
82
83 return delta;
84}
85
86#else /* CONFIG_IRQ_TIME_ACCOUNTING */
87
88#define sched_clock_irqtime (0)
89
90static u64 irqtime_tick_accounted(u64 dummy)
91{
92 return 0;
93}
94
95#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
96
97static inline void task_group_account_field(struct task_struct *p, int index,
98 u64 tmp)
99{
100 /*
101 * Since all updates are sure to touch the root cgroup, we
102 * get ourselves ahead and touch it first. If the root cgroup
103 * is the only cgroup, then nothing else should be necessary.
104 *
105 */
106 __this_cpu_add(kernel_cpustat.cpustat[index], tmp);
107
108 cgroup_account_cputime_field(p, index, tmp);
109}
110
111/*
112 * Account user CPU time to a process.
113 * @p: the process that the CPU time gets accounted to
114 * @cputime: the CPU time spent in user space since the last update
115 */
116void account_user_time(struct task_struct *p, u64 cputime)
117{
118 int index;
119
120 /* Add user time to process. */
121 p->utime += cputime;
122 account_group_user_time(p, cputime);
123
124 index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
125
126 /* Add user time to cpustat. */
127 task_group_account_field(p, index, cputime);
128
129 /* Account for user time used */
130 acct_account_cputime(p);
131}
132
133/*
134 * Account guest CPU time to a process.
135 * @p: the process that the CPU time gets accounted to
136 * @cputime: the CPU time spent in virtual machine since the last update
137 */
138void account_guest_time(struct task_struct *p, u64 cputime)
139{
140 u64 *cpustat = kcpustat_this_cpu->cpustat;
141
142 /* Add guest time to process. */
143 p->utime += cputime;
144 account_group_user_time(p, cputime);
145 p->gtime += cputime;
146
147 /* Add guest time to cpustat. */
148 if (task_nice(p) > 0) {
149 cpustat[CPUTIME_NICE] += cputime;
150 cpustat[CPUTIME_GUEST_NICE] += cputime;
151 } else {
152 cpustat[CPUTIME_USER] += cputime;
153 cpustat[CPUTIME_GUEST] += cputime;
154 }
155}
156
157/*
158 * Account system CPU time to a process and desired cpustat field
159 * @p: the process that the CPU time gets accounted to
160 * @cputime: the CPU time spent in kernel space since the last update
161 * @index: pointer to cpustat field that has to be updated
162 */
163void account_system_index_time(struct task_struct *p,
164 u64 cputime, enum cpu_usage_stat index)
165{
166 /* Add system time to process. */
167 p->stime += cputime;
168 account_group_system_time(p, cputime);
169
170 /* Add system time to cpustat. */
171 task_group_account_field(p, index, cputime);
172
173 /* Account for system time used */
174 acct_account_cputime(p);
175}
176
177/*
178 * Account system CPU time to a process.
179 * @p: the process that the CPU time gets accounted to
180 * @hardirq_offset: the offset to subtract from hardirq_count()
181 * @cputime: the CPU time spent in kernel space since the last update
182 */
183void account_system_time(struct task_struct *p, int hardirq_offset, u64 cputime)
184{
185 int index;
186
187 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
188 account_guest_time(p, cputime);
189 return;
190 }
191
192 if (hardirq_count() - hardirq_offset)
193 index = CPUTIME_IRQ;
194 else if (in_serving_softirq())
195 index = CPUTIME_SOFTIRQ;
196 else
197 index = CPUTIME_SYSTEM;
198
199 account_system_index_time(p, cputime, index);
200}
201
202/*
203 * Account for involuntary wait time.
204 * @cputime: the CPU time spent in involuntary wait
205 */
206void account_steal_time(u64 cputime)
207{
208 u64 *cpustat = kcpustat_this_cpu->cpustat;
209
210 cpustat[CPUTIME_STEAL] += cputime;
211}
212
213/*
214 * Account for idle time.
215 * @cputime: the CPU time spent in idle wait
216 */
217void account_idle_time(u64 cputime)
218{
219 u64 *cpustat = kcpustat_this_cpu->cpustat;
220 struct rq *rq = this_rq();
221
222 if (atomic_read(&rq->nr_iowait) > 0)
223 cpustat[CPUTIME_IOWAIT] += cputime;
224 else
225 cpustat[CPUTIME_IDLE] += cputime;
226}
227
228/*
229 * When a guest is interrupted for a longer amount of time, missed clock
230 * ticks are not redelivered later. Due to that, this function may on
231 * occasion account more time than the calling functions think elapsed.
232 */
233static __always_inline u64 steal_account_process_time(u64 maxtime)
234{
235#ifdef CONFIG_PARAVIRT
236 if (static_key_false(¶virt_steal_enabled)) {
237 u64 steal;
238
239 steal = paravirt_steal_clock(smp_processor_id());
240 steal -= this_rq()->prev_steal_time;
241 steal = min(steal, maxtime);
242 account_steal_time(steal);
243 this_rq()->prev_steal_time += steal;
244
245 return steal;
246 }
247#endif
248 return 0;
249}
250
251/*
252 * Account how much elapsed time was spent in steal, irq, or softirq time.
253 */
254static inline u64 account_other_time(u64 max)
255{
256 u64 accounted;
257
258 lockdep_assert_irqs_disabled();
259
260 accounted = steal_account_process_time(max);
261
262 if (accounted < max)
263 accounted += irqtime_tick_accounted(max - accounted);
264
265 return accounted;
266}
267
268#ifdef CONFIG_64BIT
269static inline u64 read_sum_exec_runtime(struct task_struct *t)
270{
271 return t->se.sum_exec_runtime;
272}
273#else
274static u64 read_sum_exec_runtime(struct task_struct *t)
275{
276 u64 ns;
277 struct rq_flags rf;
278 struct rq *rq;
279
280 rq = task_rq_lock(t, &rf);
281 ns = t->se.sum_exec_runtime;
282 task_rq_unlock(rq, t, &rf);
283
284 return ns;
285}
286#endif
287
288/*
289 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
290 * tasks (sum on group iteration) belonging to @tsk's group.
291 */
292void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
293{
294 struct signal_struct *sig = tsk->signal;
295 u64 utime, stime;
296 struct task_struct *t;
297 unsigned int seq, nextseq;
298 unsigned long flags;
299
300 /*
301 * Update current task runtime to account pending time since last
302 * scheduler action or thread_group_cputime() call. This thread group
303 * might have other running tasks on different CPUs, but updating
304 * their runtime can affect syscall performance, so we skip account
305 * those pending times and rely only on values updated on tick or
306 * other scheduler action.
307 */
308 if (same_thread_group(current, tsk))
309 (void) task_sched_runtime(current);
310
311 rcu_read_lock();
312 /* Attempt a lockless read on the first round. */
313 nextseq = 0;
314 do {
315 seq = nextseq;
316 flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
317 times->utime = sig->utime;
318 times->stime = sig->stime;
319 times->sum_exec_runtime = sig->sum_sched_runtime;
320
321 for_each_thread(tsk, t) {
322 task_cputime(t, &utime, &stime);
323 times->utime += utime;
324 times->stime += stime;
325 times->sum_exec_runtime += read_sum_exec_runtime(t);
326 }
327 /* If lockless access failed, take the lock. */
328 nextseq = 1;
329 } while (need_seqretry(&sig->stats_lock, seq));
330 done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
331 rcu_read_unlock();
332}
333
334#ifdef CONFIG_IRQ_TIME_ACCOUNTING
335/*
336 * Account a tick to a process and cpustat
337 * @p: the process that the CPU time gets accounted to
338 * @user_tick: is the tick from userspace
339 * @rq: the pointer to rq
340 *
341 * Tick demultiplexing follows the order
342 * - pending hardirq update
343 * - pending softirq update
344 * - user_time
345 * - idle_time
346 * - system time
347 * - check for guest_time
348 * - else account as system_time
349 *
350 * Check for hardirq is done both for system and user time as there is
351 * no timer going off while we are on hardirq and hence we may never get an
352 * opportunity to update it solely in system time.
353 * p->stime and friends are only updated on system time and not on irq
354 * softirq as those do not count in task exec_runtime any more.
355 */
356static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
357 struct rq *rq, int ticks)
358{
359 u64 other, cputime = TICK_NSEC * ticks;
360
361 /*
362 * When returning from idle, many ticks can get accounted at
363 * once, including some ticks of steal, irq, and softirq time.
364 * Subtract those ticks from the amount of time accounted to
365 * idle, or potentially user or system time. Due to rounding,
366 * other time can exceed ticks occasionally.
367 */
368 other = account_other_time(ULONG_MAX);
369 if (other >= cputime)
370 return;
371
372 cputime -= other;
373
374 if (this_cpu_ksoftirqd() == p) {
375 /*
376 * ksoftirqd time do not get accounted in cpu_softirq_time.
377 * So, we have to handle it separately here.
378 * Also, p->stime needs to be updated for ksoftirqd.
379 */
380 account_system_index_time(p, cputime, CPUTIME_SOFTIRQ);
381 } else if (user_tick) {
382 account_user_time(p, cputime);
383 } else if (p == rq->idle) {
384 account_idle_time(cputime);
385 } else if (p->flags & PF_VCPU) { /* System time or guest time */
386 account_guest_time(p, cputime);
387 } else {
388 account_system_index_time(p, cputime, CPUTIME_SYSTEM);
389 }
390}
391
392static void irqtime_account_idle_ticks(int ticks)
393{
394 struct rq *rq = this_rq();
395
396 irqtime_account_process_tick(current, 0, rq, 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 struct rq *rq, 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
408# ifndef __ARCH_HAS_VTIME_TASK_SWITCH
409void vtime_common_task_switch(struct task_struct *prev)
410{
411 if (is_idle_task(prev))
412 vtime_account_idle(prev);
413 else
414 vtime_account_system(prev);
415
416 vtime_flush(prev);
417 arch_vtime_task_switch(prev);
418}
419# endif
420#endif /* CONFIG_VIRT_CPU_ACCOUNTING */
421
422
423#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
424/*
425 * Archs that account the whole time spent in the idle task
426 * (outside irq) as idle time can rely on this and just implement
427 * vtime_account_system() and vtime_account_idle(). Archs that
428 * have other meaning of the idle time (s390 only includes the
429 * time spent by the CPU when it's in low power mode) must override
430 * vtime_account().
431 */
432#ifndef __ARCH_HAS_VTIME_ACCOUNT
433void vtime_account_irq_enter(struct task_struct *tsk)
434{
435 if (!in_interrupt() && is_idle_task(tsk))
436 vtime_account_idle(tsk);
437 else
438 vtime_account_system(tsk);
439}
440EXPORT_SYMBOL_GPL(vtime_account_irq_enter);
441#endif /* __ARCH_HAS_VTIME_ACCOUNT */
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 struct rq *rq = this_rq();
478
479 if (vtime_accounting_cpu_enabled())
480 return;
481
482 if (sched_clock_irqtime) {
483 irqtime_account_process_tick(p, user_tick, rq, 1);
484 return;
485 }
486
487 cputime = TICK_NSEC;
488 steal = steal_account_process_time(ULONG_MAX);
489
490 if (steal >= cputime)
491 return;
492
493 cputime -= steal;
494
495 if (user_tick)
496 account_user_time(p, cputime);
497 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
498 account_system_time(p, HARDIRQ_OFFSET, cputime);
499 else
500 account_idle_time(cputime);
501}
502
503/*
504 * Account multiple ticks of idle time.
505 * @ticks: number of stolen ticks
506 */
507void account_idle_ticks(unsigned long ticks)
508{
509 u64 cputime, steal;
510
511 if (sched_clock_irqtime) {
512 irqtime_account_idle_ticks(ticks);
513 return;
514 }
515
516 cputime = ticks * TICK_NSEC;
517 steal = steal_account_process_time(ULONG_MAX);
518
519 if (steal >= cputime)
520 return;
521
522 cputime -= steal;
523 account_idle_time(cputime);
524}
525
526/*
527 * Perform (stime * rtime) / total, but avoid multiplication overflow by
528 * loosing precision when the numbers are big.
529 */
530static u64 scale_stime(u64 stime, u64 rtime, u64 total)
531{
532 u64 scaled;
533
534 for (;;) {
535 /* Make sure "rtime" is the bigger of stime/rtime */
536 if (stime > rtime)
537 swap(rtime, stime);
538
539 /* Make sure 'total' fits in 32 bits */
540 if (total >> 32)
541 goto drop_precision;
542
543 /* Does rtime (and thus stime) fit in 32 bits? */
544 if (!(rtime >> 32))
545 break;
546
547 /* Can we just balance rtime/stime rather than dropping bits? */
548 if (stime >> 31)
549 goto drop_precision;
550
551 /* We can grow stime and shrink rtime and try to make them both fit */
552 stime <<= 1;
553 rtime >>= 1;
554 continue;
555
556drop_precision:
557 /* We drop from rtime, it has more bits than stime */
558 rtime >>= 1;
559 total >>= 1;
560 }
561
562 /*
563 * Make sure gcc understands that this is a 32x32->64 multiply,
564 * followed by a 64/32->64 divide.
565 */
566 scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
567 return scaled;
568}
569
570/*
571 * Adjust tick based cputime random precision against scheduler runtime
572 * accounting.
573 *
574 * Tick based cputime accounting depend on random scheduling timeslices of a
575 * task to be interrupted or not by the timer. Depending on these
576 * circumstances, the number of these interrupts may be over or
577 * under-optimistic, matching the real user and system cputime with a variable
578 * precision.
579 *
580 * Fix this by scaling these tick based values against the total runtime
581 * accounted by the CFS scheduler.
582 *
583 * This code provides the following guarantees:
584 *
585 * stime + utime == rtime
586 * stime_i+1 >= stime_i, utime_i+1 >= utime_i
587 *
588 * Assuming that rtime_i+1 >= rtime_i.
589 */
590void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
591 u64 *ut, u64 *st)
592{
593 u64 rtime, stime, utime;
594 unsigned long flags;
595
596 /* Serialize concurrent callers such that we can honour our guarantees */
597 raw_spin_lock_irqsave(&prev->lock, flags);
598 rtime = curr->sum_exec_runtime;
599
600 /*
601 * This is possible under two circumstances:
602 * - rtime isn't monotonic after all (a bug);
603 * - we got reordered by the lock.
604 *
605 * In both cases this acts as a filter such that the rest of the code
606 * can assume it is monotonic regardless of anything else.
607 */
608 if (prev->stime + prev->utime >= rtime)
609 goto out;
610
611 stime = curr->stime;
612 utime = curr->utime;
613
614 /*
615 * If either stime or utime are 0, assume all runtime is userspace.
616 * Once a task gets some ticks, the monotonicy code at 'update:'
617 * will ensure things converge to the observed ratio.
618 */
619 if (stime == 0) {
620 utime = rtime;
621 goto update;
622 }
623
624 if (utime == 0) {
625 stime = rtime;
626 goto update;
627 }
628
629 stime = scale_stime(stime, rtime, stime + utime);
630
631update:
632 /*
633 * Make sure stime doesn't go backwards; this preserves monotonicity
634 * for utime because rtime is monotonic.
635 *
636 * utime_i+1 = rtime_i+1 - stime_i
637 * = rtime_i+1 - (rtime_i - utime_i)
638 * = (rtime_i+1 - rtime_i) + utime_i
639 * >= utime_i
640 */
641 if (stime < prev->stime)
642 stime = prev->stime;
643 utime = rtime - stime;
644
645 /*
646 * Make sure utime doesn't go backwards; this still preserves
647 * monotonicity for stime, analogous argument to above.
648 */
649 if (utime < prev->utime) {
650 utime = prev->utime;
651 stime = rtime - utime;
652 }
653
654 prev->stime = stime;
655 prev->utime = utime;
656out:
657 *ut = prev->utime;
658 *st = prev->stime;
659 raw_spin_unlock_irqrestore(&prev->lock, flags);
660}
661
662void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
663{
664 struct task_cputime cputime = {
665 .sum_exec_runtime = p->se.sum_exec_runtime,
666 };
667
668 task_cputime(p, &cputime.utime, &cputime.stime);
669 cputime_adjust(&cputime, &p->prev_cputime, ut, st);
670}
671EXPORT_SYMBOL_GPL(task_cputime_adjusted);
672
673void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
674{
675 struct task_cputime cputime;
676
677 thread_group_cputime(p, &cputime);
678 cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
679}
680#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
681
682#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
683static u64 vtime_delta(struct vtime *vtime)
684{
685 unsigned long long clock;
686
687 clock = sched_clock();
688 if (clock < vtime->starttime)
689 return 0;
690
691 return clock - vtime->starttime;
692}
693
694static u64 get_vtime_delta(struct vtime *vtime)
695{
696 u64 delta = vtime_delta(vtime);
697 u64 other;
698
699 /*
700 * Unlike tick based timing, vtime based timing never has lost
701 * ticks, and no need for steal time accounting to make up for
702 * lost ticks. Vtime accounts a rounded version of actual
703 * elapsed time. Limit account_other_time to prevent rounding
704 * errors from causing elapsed vtime to go negative.
705 */
706 other = account_other_time(delta);
707 WARN_ON_ONCE(vtime->state == VTIME_INACTIVE);
708 vtime->starttime += delta;
709
710 return delta - other;
711}
712
713static void __vtime_account_system(struct task_struct *tsk,
714 struct vtime *vtime)
715{
716 vtime->stime += get_vtime_delta(vtime);
717 if (vtime->stime >= TICK_NSEC) {
718 account_system_time(tsk, irq_count(), vtime->stime);
719 vtime->stime = 0;
720 }
721}
722
723static void vtime_account_guest(struct task_struct *tsk,
724 struct vtime *vtime)
725{
726 vtime->gtime += get_vtime_delta(vtime);
727 if (vtime->gtime >= TICK_NSEC) {
728 account_guest_time(tsk, vtime->gtime);
729 vtime->gtime = 0;
730 }
731}
732
733void vtime_account_system(struct task_struct *tsk)
734{
735 struct vtime *vtime = &tsk->vtime;
736
737 if (!vtime_delta(vtime))
738 return;
739
740 write_seqcount_begin(&vtime->seqcount);
741 /* We might have scheduled out from guest path */
742 if (current->flags & PF_VCPU)
743 vtime_account_guest(tsk, vtime);
744 else
745 __vtime_account_system(tsk, vtime);
746 write_seqcount_end(&vtime->seqcount);
747}
748
749void vtime_user_enter(struct task_struct *tsk)
750{
751 struct vtime *vtime = &tsk->vtime;
752
753 write_seqcount_begin(&vtime->seqcount);
754 __vtime_account_system(tsk, vtime);
755 vtime->state = VTIME_USER;
756 write_seqcount_end(&vtime->seqcount);
757}
758
759void vtime_user_exit(struct task_struct *tsk)
760{
761 struct vtime *vtime = &tsk->vtime;
762
763 write_seqcount_begin(&vtime->seqcount);
764 vtime->utime += get_vtime_delta(vtime);
765 if (vtime->utime >= TICK_NSEC) {
766 account_user_time(tsk, vtime->utime);
767 vtime->utime = 0;
768 }
769 vtime->state = VTIME_SYS;
770 write_seqcount_end(&vtime->seqcount);
771}
772
773void vtime_guest_enter(struct task_struct *tsk)
774{
775 struct vtime *vtime = &tsk->vtime;
776 /*
777 * The flags must be updated under the lock with
778 * the vtime_starttime flush and update.
779 * That enforces a right ordering and update sequence
780 * synchronization against the reader (task_gtime())
781 * that can thus safely catch up with a tickless delta.
782 */
783 write_seqcount_begin(&vtime->seqcount);
784 __vtime_account_system(tsk, vtime);
785 current->flags |= PF_VCPU;
786 write_seqcount_end(&vtime->seqcount);
787}
788EXPORT_SYMBOL_GPL(vtime_guest_enter);
789
790void vtime_guest_exit(struct task_struct *tsk)
791{
792 struct vtime *vtime = &tsk->vtime;
793
794 write_seqcount_begin(&vtime->seqcount);
795 vtime_account_guest(tsk, vtime);
796 current->flags &= ~PF_VCPU;
797 write_seqcount_end(&vtime->seqcount);
798}
799EXPORT_SYMBOL_GPL(vtime_guest_exit);
800
801void vtime_account_idle(struct task_struct *tsk)
802{
803 account_idle_time(get_vtime_delta(&tsk->vtime));
804}
805
806void arch_vtime_task_switch(struct task_struct *prev)
807{
808 struct vtime *vtime = &prev->vtime;
809
810 write_seqcount_begin(&vtime->seqcount);
811 vtime->state = VTIME_INACTIVE;
812 write_seqcount_end(&vtime->seqcount);
813
814 vtime = ¤t->vtime;
815
816 write_seqcount_begin(&vtime->seqcount);
817 vtime->state = VTIME_SYS;
818 vtime->starttime = sched_clock();
819 write_seqcount_end(&vtime->seqcount);
820}
821
822void vtime_init_idle(struct task_struct *t, int cpu)
823{
824 struct vtime *vtime = &t->vtime;
825 unsigned long flags;
826
827 local_irq_save(flags);
828 write_seqcount_begin(&vtime->seqcount);
829 vtime->state = VTIME_SYS;
830 vtime->starttime = sched_clock();
831 write_seqcount_end(&vtime->seqcount);
832 local_irq_restore(flags);
833}
834
835u64 task_gtime(struct task_struct *t)
836{
837 struct vtime *vtime = &t->vtime;
838 unsigned int seq;
839 u64 gtime;
840
841 if (!vtime_accounting_enabled())
842 return t->gtime;
843
844 do {
845 seq = read_seqcount_begin(&vtime->seqcount);
846
847 gtime = t->gtime;
848 if (vtime->state == VTIME_SYS && t->flags & PF_VCPU)
849 gtime += vtime->gtime + vtime_delta(vtime);
850
851 } while (read_seqcount_retry(&vtime->seqcount, seq));
852
853 return gtime;
854}
855
856/*
857 * Fetch cputime raw values from fields of task_struct and
858 * add up the pending nohz execution time since the last
859 * cputime snapshot.
860 */
861void task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
862{
863 struct vtime *vtime = &t->vtime;
864 unsigned int seq;
865 u64 delta;
866
867 if (!vtime_accounting_enabled()) {
868 *utime = t->utime;
869 *stime = t->stime;
870 return;
871 }
872
873 do {
874 seq = read_seqcount_begin(&vtime->seqcount);
875
876 *utime = t->utime;
877 *stime = t->stime;
878
879 /* Task is sleeping, nothing to add */
880 if (vtime->state == VTIME_INACTIVE || is_idle_task(t))
881 continue;
882
883 delta = vtime_delta(vtime);
884
885 /*
886 * Task runs either in user or kernel space, add pending nohz time to
887 * the right place.
888 */
889 if (vtime->state == VTIME_USER || t->flags & PF_VCPU)
890 *utime += vtime->utime + delta;
891 else if (vtime->state == VTIME_SYS)
892 *stime += vtime->stime + delta;
893 } while (read_seqcount_retry(&vtime->seqcount, seq));
894}
895#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Simple CPU accounting cgroup controller
4 */
5
6#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
7 #include <asm/cputime.h>
8#endif
9
10#ifdef CONFIG_IRQ_TIME_ACCOUNTING
11
12/*
13 * There are no locks covering percpu hardirq/softirq time.
14 * They are only modified in vtime_account, on corresponding CPU
15 * with interrupts disabled. So, writes are safe.
16 * They are read and saved off onto struct rq in update_rq_clock().
17 * This may result in other CPU reading this CPU's IRQ time and can
18 * race with irq/vtime_account on this CPU. We would either get old
19 * or new value with a side effect of accounting a slice of IRQ time to wrong
20 * task when IRQ is in progress while we read rq->clock. That is a worthy
21 * compromise in place of having locks on each IRQ in account_system_time.
22 */
23DEFINE_PER_CPU(struct irqtime, cpu_irqtime);
24
25static int sched_clock_irqtime;
26
27void enable_sched_clock_irqtime(void)
28{
29 sched_clock_irqtime = 1;
30}
31
32void disable_sched_clock_irqtime(void)
33{
34 sched_clock_irqtime = 0;
35}
36
37static void irqtime_account_delta(struct irqtime *irqtime, u64 delta,
38 enum cpu_usage_stat idx)
39{
40 u64 *cpustat = kcpustat_this_cpu->cpustat;
41
42 u64_stats_update_begin(&irqtime->sync);
43 cpustat[idx] += delta;
44 irqtime->total += delta;
45 irqtime->tick_delta += delta;
46 u64_stats_update_end(&irqtime->sync);
47}
48
49/*
50 * Called after incrementing preempt_count on {soft,}irq_enter
51 * and before decrementing preempt_count on {soft,}irq_exit.
52 */
53void irqtime_account_irq(struct task_struct *curr, unsigned int offset)
54{
55 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
56 unsigned int pc;
57 s64 delta;
58 int cpu;
59
60 if (!sched_clock_irqtime)
61 return;
62
63 cpu = smp_processor_id();
64 delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
65 irqtime->irq_start_time += delta;
66 pc = irq_count() - offset;
67
68 /*
69 * We do not account for softirq time from ksoftirqd here.
70 * We want to continue accounting softirq time to ksoftirqd thread
71 * in that case, so as not to confuse scheduler with a special task
72 * that do not consume any time, but still wants to run.
73 */
74 if (pc & HARDIRQ_MASK)
75 irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
76 else if ((pc & SOFTIRQ_OFFSET) && curr != this_cpu_ksoftirqd())
77 irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
78}
79
80static u64 irqtime_tick_accounted(u64 maxtime)
81{
82 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
83 u64 delta;
84
85 delta = min(irqtime->tick_delta, maxtime);
86 irqtime->tick_delta -= delta;
87
88 return delta;
89}
90
91#else /* CONFIG_IRQ_TIME_ACCOUNTING */
92
93#define sched_clock_irqtime (0)
94
95static u64 irqtime_tick_accounted(u64 dummy)
96{
97 return 0;
98}
99
100#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
101
102static inline void task_group_account_field(struct task_struct *p, int index,
103 u64 tmp)
104{
105 /*
106 * Since all updates are sure to touch the root cgroup, we
107 * get ourselves ahead and touch it first. If the root cgroup
108 * is the only cgroup, then nothing else should be necessary.
109 *
110 */
111 __this_cpu_add(kernel_cpustat.cpustat[index], tmp);
112
113 cgroup_account_cputime_field(p, index, tmp);
114}
115
116/*
117 * Account user CPU time to a process.
118 * @p: the process that the CPU time gets accounted to
119 * @cputime: the CPU time spent in user space since the last update
120 */
121void account_user_time(struct task_struct *p, u64 cputime)
122{
123 int index;
124
125 /* Add user time to process. */
126 p->utime += cputime;
127 account_group_user_time(p, cputime);
128
129 index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
130
131 /* Add user time to cpustat. */
132 task_group_account_field(p, index, cputime);
133
134 /* Account for user time used */
135 acct_account_cputime(p);
136}
137
138/*
139 * Account guest CPU time to a process.
140 * @p: the process that the CPU time gets accounted to
141 * @cputime: the CPU time spent in virtual machine since the last update
142 */
143void account_guest_time(struct task_struct *p, u64 cputime)
144{
145 u64 *cpustat = kcpustat_this_cpu->cpustat;
146
147 /* Add guest time to process. */
148 p->utime += cputime;
149 account_group_user_time(p, cputime);
150 p->gtime += cputime;
151
152 /* Add guest time to cpustat. */
153 if (task_nice(p) > 0) {
154 task_group_account_field(p, CPUTIME_NICE, cputime);
155 cpustat[CPUTIME_GUEST_NICE] += cputime;
156 } else {
157 task_group_account_field(p, CPUTIME_USER, cputime);
158 cpustat[CPUTIME_GUEST] += cputime;
159 }
160}
161
162/*
163 * Account system CPU time to a process and desired cpustat field
164 * @p: the process that the CPU time gets accounted to
165 * @cputime: the CPU time spent in kernel space since the last update
166 * @index: pointer to cpustat field that has to be updated
167 */
168void account_system_index_time(struct task_struct *p,
169 u64 cputime, enum cpu_usage_stat index)
170{
171 /* Add system time to process. */
172 p->stime += cputime;
173 account_group_system_time(p, cputime);
174
175 /* Add system time to cpustat. */
176 task_group_account_field(p, index, cputime);
177
178 /* Account for system time used */
179 acct_account_cputime(p);
180}
181
182/*
183 * Account system CPU time to a process.
184 * @p: the process that the CPU time gets accounted to
185 * @hardirq_offset: the offset to subtract from hardirq_count()
186 * @cputime: the CPU time spent in kernel space since the last update
187 */
188void account_system_time(struct task_struct *p, int hardirq_offset, u64 cputime)
189{
190 int index;
191
192 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
193 account_guest_time(p, cputime);
194 return;
195 }
196
197 if (hardirq_count() - hardirq_offset)
198 index = CPUTIME_IRQ;
199 else if (in_serving_softirq())
200 index = CPUTIME_SOFTIRQ;
201 else
202 index = CPUTIME_SYSTEM;
203
204 account_system_index_time(p, cputime, index);
205}
206
207/*
208 * Account for involuntary wait time.
209 * @cputime: the CPU time spent in involuntary wait
210 */
211void account_steal_time(u64 cputime)
212{
213 u64 *cpustat = kcpustat_this_cpu->cpustat;
214
215 cpustat[CPUTIME_STEAL] += cputime;
216}
217
218/*
219 * Account for idle time.
220 * @cputime: the CPU time spent in idle wait
221 */
222void account_idle_time(u64 cputime)
223{
224 u64 *cpustat = kcpustat_this_cpu->cpustat;
225 struct rq *rq = this_rq();
226
227 if (atomic_read(&rq->nr_iowait) > 0)
228 cpustat[CPUTIME_IOWAIT] += cputime;
229 else
230 cpustat[CPUTIME_IDLE] += cputime;
231}
232
233
234#ifdef CONFIG_SCHED_CORE
235/*
236 * Account for forceidle time due to core scheduling.
237 *
238 * REQUIRES: schedstat is enabled.
239 */
240void __account_forceidle_time(struct task_struct *p, u64 delta)
241{
242 __schedstat_add(p->stats.core_forceidle_sum, delta);
243
244 task_group_account_field(p, CPUTIME_FORCEIDLE, delta);
245}
246#endif
247
248/*
249 * When a guest is interrupted for a longer amount of time, missed clock
250 * ticks are not redelivered later. Due to that, this function may on
251 * occasion account more time than the calling functions think elapsed.
252 */
253static __always_inline u64 steal_account_process_time(u64 maxtime)
254{
255#ifdef CONFIG_PARAVIRT
256 if (static_key_false(¶virt_steal_enabled)) {
257 u64 steal;
258
259 steal = paravirt_steal_clock(smp_processor_id());
260 steal -= this_rq()->prev_steal_time;
261 steal = min(steal, maxtime);
262 account_steal_time(steal);
263 this_rq()->prev_steal_time += steal;
264
265 return steal;
266 }
267#endif
268 return 0;
269}
270
271/*
272 * Account how much elapsed time was spent in steal, IRQ, or softirq time.
273 */
274static inline u64 account_other_time(u64 max)
275{
276 u64 accounted;
277
278 lockdep_assert_irqs_disabled();
279
280 accounted = steal_account_process_time(max);
281
282 if (accounted < max)
283 accounted += irqtime_tick_accounted(max - accounted);
284
285 return accounted;
286}
287
288#ifdef CONFIG_64BIT
289static inline u64 read_sum_exec_runtime(struct task_struct *t)
290{
291 return t->se.sum_exec_runtime;
292}
293#else
294static u64 read_sum_exec_runtime(struct task_struct *t)
295{
296 u64 ns;
297 struct rq_flags rf;
298 struct rq *rq;
299
300 rq = task_rq_lock(t, &rf);
301 ns = t->se.sum_exec_runtime;
302 task_rq_unlock(rq, t, &rf);
303
304 return ns;
305}
306#endif
307
308/*
309 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
310 * tasks (sum on group iteration) belonging to @tsk's group.
311 */
312void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
313{
314 struct signal_struct *sig = tsk->signal;
315 u64 utime, stime;
316 struct task_struct *t;
317 unsigned int seq, nextseq;
318 unsigned long flags;
319
320 /*
321 * Update current task runtime to account pending time since last
322 * scheduler action or thread_group_cputime() call. This thread group
323 * might have other running tasks on different CPUs, but updating
324 * their runtime can affect syscall performance, so we skip account
325 * those pending times and rely only on values updated on tick or
326 * other scheduler action.
327 */
328 if (same_thread_group(current, tsk))
329 (void) task_sched_runtime(current);
330
331 rcu_read_lock();
332 /* Attempt a lockless read on the first round. */
333 nextseq = 0;
334 do {
335 seq = nextseq;
336 flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
337 times->utime = sig->utime;
338 times->stime = sig->stime;
339 times->sum_exec_runtime = sig->sum_sched_runtime;
340
341 for_each_thread(tsk, t) {
342 task_cputime(t, &utime, &stime);
343 times->utime += utime;
344 times->stime += stime;
345 times->sum_exec_runtime += read_sum_exec_runtime(t);
346 }
347 /* If lockless access failed, take the lock. */
348 nextseq = 1;
349 } while (need_seqretry(&sig->stats_lock, seq));
350 done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
351 rcu_read_unlock();
352}
353
354#ifdef CONFIG_IRQ_TIME_ACCOUNTING
355/*
356 * Account a tick to a process and cpustat
357 * @p: the process that the CPU time gets accounted to
358 * @user_tick: is the tick from userspace
359 * @rq: the pointer to rq
360 *
361 * Tick demultiplexing follows the order
362 * - pending hardirq update
363 * - pending softirq update
364 * - user_time
365 * - idle_time
366 * - system time
367 * - check for guest_time
368 * - else account as system_time
369 *
370 * Check for hardirq is done both for system and user time as there is
371 * no timer going off while we are on hardirq and hence we may never get an
372 * opportunity to update it solely in system time.
373 * p->stime and friends are only updated on system time and not on IRQ
374 * softirq as those do not count in task exec_runtime any more.
375 */
376static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
377 int ticks)
378{
379 u64 other, cputime = TICK_NSEC * ticks;
380
381 /*
382 * When returning from idle, many ticks can get accounted at
383 * once, including some ticks of steal, IRQ, and softirq time.
384 * Subtract those ticks from the amount of time accounted to
385 * idle, or potentially user or system time. Due to rounding,
386 * other time can exceed ticks occasionally.
387 */
388 other = account_other_time(ULONG_MAX);
389 if (other >= cputime)
390 return;
391
392 cputime -= other;
393
394 if (this_cpu_ksoftirqd() == p) {
395 /*
396 * ksoftirqd time do not get accounted in cpu_softirq_time.
397 * So, we have to handle it separately here.
398 * Also, p->stime needs to be updated for ksoftirqd.
399 */
400 account_system_index_time(p, cputime, CPUTIME_SOFTIRQ);
401 } else if (user_tick) {
402 account_user_time(p, cputime);
403 } else if (p == this_rq()->idle) {
404 account_idle_time(cputime);
405 } else if (p->flags & PF_VCPU) { /* System time or guest time */
406 account_guest_time(p, cputime);
407 } else {
408 account_system_index_time(p, cputime, CPUTIME_SYSTEM);
409 }
410}
411
412static void irqtime_account_idle_ticks(int ticks)
413{
414 irqtime_account_process_tick(current, 0, ticks);
415}
416#else /* CONFIG_IRQ_TIME_ACCOUNTING */
417static inline void irqtime_account_idle_ticks(int ticks) { }
418static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
419 int nr_ticks) { }
420#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
421
422/*
423 * Use precise platform statistics if available:
424 */
425#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
426
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 */