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