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 before incrementing preempt_count on {soft,}irq_enter
48 * and before decrementing preempt_count on {soft,}irq_exit.
49 */
50void irqtime_account_irq(struct task_struct *curr)
51{
52 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
53 s64 delta;
54 int cpu;
55
56 if (!sched_clock_irqtime)
57 return;
58
59 cpu = smp_processor_id();
60 delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
61 irqtime->irq_start_time += delta;
62
63 /*
64 * We do not account for softirq time from ksoftirqd here.
65 * We want to continue accounting softirq time to ksoftirqd thread
66 * in that case, so as not to confuse scheduler with a special task
67 * that do not consume any time, but still wants to run.
68 */
69 if (hardirq_count())
70 irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
71 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
72 irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
73}
74EXPORT_SYMBOL_GPL(irqtime_account_irq);
75
76static u64 irqtime_tick_accounted(u64 maxtime)
77{
78 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
79 u64 delta;
80
81 delta = min(irqtime->tick_delta, maxtime);
82 irqtime->tick_delta -= delta;
83
84 return delta;
85}
86
87#else /* CONFIG_IRQ_TIME_ACCOUNTING */
88
89#define sched_clock_irqtime (0)
90
91static u64 irqtime_tick_accounted(u64 dummy)
92{
93 return 0;
94}
95
96#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
97
98static inline void task_group_account_field(struct task_struct *p, int index,
99 u64 tmp)
100{
101 /*
102 * Since all updates are sure to touch the root cgroup, we
103 * get ourselves ahead and touch it first. If the root cgroup
104 * is the only cgroup, then nothing else should be necessary.
105 *
106 */
107 __this_cpu_add(kernel_cpustat.cpustat[index], tmp);
108
109 cgroup_account_cputime_field(p, index, tmp);
110}
111
112/*
113 * Account user CPU time to a process.
114 * @p: the process that the CPU time gets accounted to
115 * @cputime: the CPU time spent in user space since the last update
116 */
117void account_user_time(struct task_struct *p, u64 cputime)
118{
119 int index;
120
121 /* Add user time to process. */
122 p->utime += cputime;
123 account_group_user_time(p, cputime);
124
125 index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
126
127 /* Add user time to cpustat. */
128 task_group_account_field(p, index, cputime);
129
130 /* Account for user time used */
131 acct_account_cputime(p);
132}
133
134/*
135 * Account guest CPU time to a process.
136 * @p: the process that the CPU time gets accounted to
137 * @cputime: the CPU time spent in virtual machine since the last update
138 */
139void account_guest_time(struct task_struct *p, u64 cputime)
140{
141 u64 *cpustat = kcpustat_this_cpu->cpustat;
142
143 /* Add guest time to process. */
144 p->utime += cputime;
145 account_group_user_time(p, cputime);
146 p->gtime += cputime;
147
148 /* Add guest time to cpustat. */
149 if (task_nice(p) > 0) {
150 cpustat[CPUTIME_NICE] += cputime;
151 cpustat[CPUTIME_GUEST_NICE] += cputime;
152 } else {
153 cpustat[CPUTIME_USER] += cputime;
154 cpustat[CPUTIME_GUEST] += cputime;
155 }
156}
157
158/*
159 * Account system CPU time to a process and desired cpustat field
160 * @p: the process that the CPU time gets accounted to
161 * @cputime: the CPU time spent in kernel space since the last update
162 * @index: pointer to cpustat field that has to be updated
163 */
164void account_system_index_time(struct task_struct *p,
165 u64 cputime, enum cpu_usage_stat index)
166{
167 /* Add system time to process. */
168 p->stime += cputime;
169 account_group_system_time(p, cputime);
170
171 /* Add system time to cpustat. */
172 task_group_account_field(p, index, cputime);
173
174 /* Account for system time used */
175 acct_account_cputime(p);
176}
177
178/*
179 * Account system CPU time to a process.
180 * @p: the process that the CPU time gets accounted to
181 * @hardirq_offset: the offset to subtract from hardirq_count()
182 * @cputime: the CPU time spent in kernel space since the last update
183 */
184void account_system_time(struct task_struct *p, int hardirq_offset, u64 cputime)
185{
186 int index;
187
188 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
189 account_guest_time(p, cputime);
190 return;
191 }
192
193 if (hardirq_count() - hardirq_offset)
194 index = CPUTIME_IRQ;
195 else if (in_serving_softirq())
196 index = CPUTIME_SOFTIRQ;
197 else
198 index = CPUTIME_SYSTEM;
199
200 account_system_index_time(p, cputime, index);
201}
202
203/*
204 * Account for involuntary wait time.
205 * @cputime: the CPU time spent in involuntary wait
206 */
207void account_steal_time(u64 cputime)
208{
209 u64 *cpustat = kcpustat_this_cpu->cpustat;
210
211 cpustat[CPUTIME_STEAL] += cputime;
212}
213
214/*
215 * Account for idle time.
216 * @cputime: the CPU time spent in idle wait
217 */
218void account_idle_time(u64 cputime)
219{
220 u64 *cpustat = kcpustat_this_cpu->cpustat;
221 struct rq *rq = this_rq();
222
223 if (atomic_read(&rq->nr_iowait) > 0)
224 cpustat[CPUTIME_IOWAIT] += cputime;
225 else
226 cpustat[CPUTIME_IDLE] += cputime;
227}
228
229/*
230 * When a guest is interrupted for a longer amount of time, missed clock
231 * ticks are not redelivered later. Due to that, this function may on
232 * occasion account more time than the calling functions think elapsed.
233 */
234static __always_inline u64 steal_account_process_time(u64 maxtime)
235{
236#ifdef CONFIG_PARAVIRT
237 if (static_key_false(¶virt_steal_enabled)) {
238 u64 steal;
239
240 steal = paravirt_steal_clock(smp_processor_id());
241 steal -= this_rq()->prev_steal_time;
242 steal = min(steal, maxtime);
243 account_steal_time(steal);
244 this_rq()->prev_steal_time += steal;
245
246 return steal;
247 }
248#endif
249 return 0;
250}
251
252/*
253 * Account how much elapsed time was spent in steal, irq, or softirq time.
254 */
255static inline u64 account_other_time(u64 max)
256{
257 u64 accounted;
258
259 lockdep_assert_irqs_disabled();
260
261 accounted = steal_account_process_time(max);
262
263 if (accounted < max)
264 accounted += irqtime_tick_accounted(max - accounted);
265
266 return accounted;
267}
268
269#ifdef CONFIG_64BIT
270static inline u64 read_sum_exec_runtime(struct task_struct *t)
271{
272 return t->se.sum_exec_runtime;
273}
274#else
275static u64 read_sum_exec_runtime(struct task_struct *t)
276{
277 u64 ns;
278 struct rq_flags rf;
279 struct rq *rq;
280
281 rq = task_rq_lock(t, &rf);
282 ns = t->se.sum_exec_runtime;
283 task_rq_unlock(rq, t, &rf);
284
285 return ns;
286}
287#endif
288
289/*
290 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
291 * tasks (sum on group iteration) belonging to @tsk's group.
292 */
293void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
294{
295 struct signal_struct *sig = tsk->signal;
296 u64 utime, stime;
297 struct task_struct *t;
298 unsigned int seq, nextseq;
299 unsigned long flags;
300
301 /*
302 * Update current task runtime to account pending time since last
303 * scheduler action or thread_group_cputime() call. This thread group
304 * might have other running tasks on different CPUs, but updating
305 * their runtime can affect syscall performance, so we skip account
306 * those pending times and rely only on values updated on tick or
307 * other scheduler action.
308 */
309 if (same_thread_group(current, tsk))
310 (void) task_sched_runtime(current);
311
312 rcu_read_lock();
313 /* Attempt a lockless read on the first round. */
314 nextseq = 0;
315 do {
316 seq = nextseq;
317 flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
318 times->utime = sig->utime;
319 times->stime = sig->stime;
320 times->sum_exec_runtime = sig->sum_sched_runtime;
321
322 for_each_thread(tsk, t) {
323 task_cputime(t, &utime, &stime);
324 times->utime += utime;
325 times->stime += stime;
326 times->sum_exec_runtime += read_sum_exec_runtime(t);
327 }
328 /* If lockless access failed, take the lock. */
329 nextseq = 1;
330 } while (need_seqretry(&sig->stats_lock, seq));
331 done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
332 rcu_read_unlock();
333}
334
335#ifdef CONFIG_IRQ_TIME_ACCOUNTING
336/*
337 * Account a tick to a process and cpustat
338 * @p: the process that the CPU time gets accounted to
339 * @user_tick: is the tick from userspace
340 * @rq: the pointer to rq
341 *
342 * Tick demultiplexing follows the order
343 * - pending hardirq update
344 * - pending softirq update
345 * - user_time
346 * - idle_time
347 * - system time
348 * - check for guest_time
349 * - else account as system_time
350 *
351 * Check for hardirq is done both for system and user time as there is
352 * no timer going off while we are on hardirq and hence we may never get an
353 * opportunity to update it solely in system time.
354 * p->stime and friends are only updated on system time and not on irq
355 * softirq as those do not count in task exec_runtime any more.
356 */
357static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
358 int ticks)
359{
360 u64 other, cputime = TICK_NSEC * ticks;
361
362 /*
363 * When returning from idle, many ticks can get accounted at
364 * once, including some ticks of steal, irq, and softirq time.
365 * Subtract those ticks from the amount of time accounted to
366 * idle, or potentially user or system time. Due to rounding,
367 * other time can exceed ticks occasionally.
368 */
369 other = account_other_time(ULONG_MAX);
370 if (other >= cputime)
371 return;
372
373 cputime -= other;
374
375 if (this_cpu_ksoftirqd() == p) {
376 /*
377 * ksoftirqd time do not get accounted in cpu_softirq_time.
378 * So, we have to handle it separately here.
379 * Also, p->stime needs to be updated for ksoftirqd.
380 */
381 account_system_index_time(p, cputime, CPUTIME_SOFTIRQ);
382 } else if (user_tick) {
383 account_user_time(p, cputime);
384 } else if (p == this_rq()->idle) {
385 account_idle_time(cputime);
386 } else if (p->flags & PF_VCPU) { /* System time or guest time */
387 account_guest_time(p, cputime);
388 } else {
389 account_system_index_time(p, cputime, CPUTIME_SYSTEM);
390 }
391}
392
393static void irqtime_account_idle_ticks(int ticks)
394{
395 irqtime_account_process_tick(current, 0, ticks);
396}
397#else /* CONFIG_IRQ_TIME_ACCOUNTING */
398static inline void irqtime_account_idle_ticks(int ticks) { }
399static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
400 int nr_ticks) { }
401#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
402
403/*
404 * Use precise platform statistics if available:
405 */
406#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
407
408# ifndef __ARCH_HAS_VTIME_TASK_SWITCH
409void vtime_task_switch(struct task_struct *prev)
410{
411 if (is_idle_task(prev))
412 vtime_account_idle(prev);
413 else
414 vtime_account_kernel(prev);
415
416 vtime_flush(prev);
417 arch_vtime_task_switch(prev);
418}
419# endif
420
421/*
422 * Archs that account the whole time spent in the idle task
423 * (outside irq) as idle time can rely on this and just implement
424 * vtime_account_kernel() and vtime_account_idle(). Archs that
425 * have other meaning of the idle time (s390 only includes the
426 * time spent by the CPU when it's in low power mode) must override
427 * vtime_account().
428 */
429#ifndef __ARCH_HAS_VTIME_ACCOUNT
430void vtime_account_irq_enter(struct task_struct *tsk)
431{
432 if (!in_interrupt() && is_idle_task(tsk))
433 vtime_account_idle(tsk);
434 else
435 vtime_account_kernel(tsk);
436}
437EXPORT_SYMBOL_GPL(vtime_account_irq_enter);
438#endif /* __ARCH_HAS_VTIME_ACCOUNT */
439
440void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
441 u64 *ut, u64 *st)
442{
443 *ut = curr->utime;
444 *st = curr->stime;
445}
446
447void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
448{
449 *ut = p->utime;
450 *st = p->stime;
451}
452EXPORT_SYMBOL_GPL(task_cputime_adjusted);
453
454void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
455{
456 struct task_cputime cputime;
457
458 thread_group_cputime(p, &cputime);
459
460 *ut = cputime.utime;
461 *st = cputime.stime;
462}
463
464#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE: */
465
466/*
467 * Account a single tick of CPU time.
468 * @p: the process that the CPU time gets accounted to
469 * @user_tick: indicates if the tick is a user or a system tick
470 */
471void account_process_tick(struct task_struct *p, int user_tick)
472{
473 u64 cputime, steal;
474
475 if (vtime_accounting_enabled_this_cpu())
476 return;
477
478 if (sched_clock_irqtime) {
479 irqtime_account_process_tick(p, user_tick, 1);
480 return;
481 }
482
483 cputime = TICK_NSEC;
484 steal = steal_account_process_time(ULONG_MAX);
485
486 if (steal >= cputime)
487 return;
488
489 cputime -= steal;
490
491 if (user_tick)
492 account_user_time(p, cputime);
493 else if ((p != this_rq()->idle) || (irq_count() != HARDIRQ_OFFSET))
494 account_system_time(p, HARDIRQ_OFFSET, cputime);
495 else
496 account_idle_time(cputime);
497}
498
499/*
500 * Account multiple ticks of idle time.
501 * @ticks: number of stolen ticks
502 */
503void account_idle_ticks(unsigned long ticks)
504{
505 u64 cputime, steal;
506
507 if (sched_clock_irqtime) {
508 irqtime_account_idle_ticks(ticks);
509 return;
510 }
511
512 cputime = ticks * TICK_NSEC;
513 steal = steal_account_process_time(ULONG_MAX);
514
515 if (steal >= cputime)
516 return;
517
518 cputime -= steal;
519 account_idle_time(cputime);
520}
521
522/*
523 * Adjust tick based cputime random precision against scheduler runtime
524 * accounting.
525 *
526 * Tick based cputime accounting depend on random scheduling timeslices of a
527 * task to be interrupted or not by the timer. Depending on these
528 * circumstances, the number of these interrupts may be over or
529 * under-optimistic, matching the real user and system cputime with a variable
530 * precision.
531 *
532 * Fix this by scaling these tick based values against the total runtime
533 * accounted by the CFS scheduler.
534 *
535 * This code provides the following guarantees:
536 *
537 * stime + utime == rtime
538 * stime_i+1 >= stime_i, utime_i+1 >= utime_i
539 *
540 * Assuming that rtime_i+1 >= rtime_i.
541 */
542void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
543 u64 *ut, u64 *st)
544{
545 u64 rtime, stime, utime;
546 unsigned long flags;
547
548 /* Serialize concurrent callers such that we can honour our guarantees */
549 raw_spin_lock_irqsave(&prev->lock, flags);
550 rtime = curr->sum_exec_runtime;
551
552 /*
553 * This is possible under two circumstances:
554 * - rtime isn't monotonic after all (a bug);
555 * - we got reordered by the lock.
556 *
557 * In both cases this acts as a filter such that the rest of the code
558 * can assume it is monotonic regardless of anything else.
559 */
560 if (prev->stime + prev->utime >= rtime)
561 goto out;
562
563 stime = curr->stime;
564 utime = curr->utime;
565
566 /*
567 * If either stime or utime are 0, assume all runtime is userspace.
568 * Once a task gets some ticks, the monotonicy code at 'update:'
569 * will ensure things converge to the observed ratio.
570 */
571 if (stime == 0) {
572 utime = rtime;
573 goto update;
574 }
575
576 if (utime == 0) {
577 stime = rtime;
578 goto update;
579 }
580
581 stime = mul_u64_u64_div_u64(stime, rtime, stime + utime);
582
583update:
584 /*
585 * Make sure stime doesn't go backwards; this preserves monotonicity
586 * for utime because rtime is monotonic.
587 *
588 * utime_i+1 = rtime_i+1 - stime_i
589 * = rtime_i+1 - (rtime_i - utime_i)
590 * = (rtime_i+1 - rtime_i) + utime_i
591 * >= utime_i
592 */
593 if (stime < prev->stime)
594 stime = prev->stime;
595 utime = rtime - stime;
596
597 /*
598 * Make sure utime doesn't go backwards; this still preserves
599 * monotonicity for stime, analogous argument to above.
600 */
601 if (utime < prev->utime) {
602 utime = prev->utime;
603 stime = rtime - utime;
604 }
605
606 prev->stime = stime;
607 prev->utime = utime;
608out:
609 *ut = prev->utime;
610 *st = prev->stime;
611 raw_spin_unlock_irqrestore(&prev->lock, flags);
612}
613
614void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
615{
616 struct task_cputime cputime = {
617 .sum_exec_runtime = p->se.sum_exec_runtime,
618 };
619
620 task_cputime(p, &cputime.utime, &cputime.stime);
621 cputime_adjust(&cputime, &p->prev_cputime, ut, st);
622}
623EXPORT_SYMBOL_GPL(task_cputime_adjusted);
624
625void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
626{
627 struct task_cputime cputime;
628
629 thread_group_cputime(p, &cputime);
630 cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
631}
632#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
633
634#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
635static u64 vtime_delta(struct vtime *vtime)
636{
637 unsigned long long clock;
638
639 clock = sched_clock();
640 if (clock < vtime->starttime)
641 return 0;
642
643 return clock - vtime->starttime;
644}
645
646static u64 get_vtime_delta(struct vtime *vtime)
647{
648 u64 delta = vtime_delta(vtime);
649 u64 other;
650
651 /*
652 * Unlike tick based timing, vtime based timing never has lost
653 * ticks, and no need for steal time accounting to make up for
654 * lost ticks. Vtime accounts a rounded version of actual
655 * elapsed time. Limit account_other_time to prevent rounding
656 * errors from causing elapsed vtime to go negative.
657 */
658 other = account_other_time(delta);
659 WARN_ON_ONCE(vtime->state == VTIME_INACTIVE);
660 vtime->starttime += delta;
661
662 return delta - other;
663}
664
665static void vtime_account_system(struct task_struct *tsk,
666 struct vtime *vtime)
667{
668 vtime->stime += get_vtime_delta(vtime);
669 if (vtime->stime >= TICK_NSEC) {
670 account_system_time(tsk, irq_count(), vtime->stime);
671 vtime->stime = 0;
672 }
673}
674
675static void vtime_account_guest(struct task_struct *tsk,
676 struct vtime *vtime)
677{
678 vtime->gtime += get_vtime_delta(vtime);
679 if (vtime->gtime >= TICK_NSEC) {
680 account_guest_time(tsk, vtime->gtime);
681 vtime->gtime = 0;
682 }
683}
684
685static void __vtime_account_kernel(struct task_struct *tsk,
686 struct vtime *vtime)
687{
688 /* We might have scheduled out from guest path */
689 if (vtime->state == VTIME_GUEST)
690 vtime_account_guest(tsk, vtime);
691 else
692 vtime_account_system(tsk, vtime);
693}
694
695void vtime_account_kernel(struct task_struct *tsk)
696{
697 struct vtime *vtime = &tsk->vtime;
698
699 if (!vtime_delta(vtime))
700 return;
701
702 write_seqcount_begin(&vtime->seqcount);
703 __vtime_account_kernel(tsk, vtime);
704 write_seqcount_end(&vtime->seqcount);
705}
706
707void vtime_user_enter(struct task_struct *tsk)
708{
709 struct vtime *vtime = &tsk->vtime;
710
711 write_seqcount_begin(&vtime->seqcount);
712 vtime_account_system(tsk, vtime);
713 vtime->state = VTIME_USER;
714 write_seqcount_end(&vtime->seqcount);
715}
716
717void vtime_user_exit(struct task_struct *tsk)
718{
719 struct vtime *vtime = &tsk->vtime;
720
721 write_seqcount_begin(&vtime->seqcount);
722 vtime->utime += get_vtime_delta(vtime);
723 if (vtime->utime >= TICK_NSEC) {
724 account_user_time(tsk, vtime->utime);
725 vtime->utime = 0;
726 }
727 vtime->state = VTIME_SYS;
728 write_seqcount_end(&vtime->seqcount);
729}
730
731void vtime_guest_enter(struct task_struct *tsk)
732{
733 struct vtime *vtime = &tsk->vtime;
734 /*
735 * The flags must be updated under the lock with
736 * the vtime_starttime flush and update.
737 * That enforces a right ordering and update sequence
738 * synchronization against the reader (task_gtime())
739 * that can thus safely catch up with a tickless delta.
740 */
741 write_seqcount_begin(&vtime->seqcount);
742 vtime_account_system(tsk, vtime);
743 tsk->flags |= PF_VCPU;
744 vtime->state = VTIME_GUEST;
745 write_seqcount_end(&vtime->seqcount);
746}
747EXPORT_SYMBOL_GPL(vtime_guest_enter);
748
749void vtime_guest_exit(struct task_struct *tsk)
750{
751 struct vtime *vtime = &tsk->vtime;
752
753 write_seqcount_begin(&vtime->seqcount);
754 vtime_account_guest(tsk, vtime);
755 tsk->flags &= ~PF_VCPU;
756 vtime->state = VTIME_SYS;
757 write_seqcount_end(&vtime->seqcount);
758}
759EXPORT_SYMBOL_GPL(vtime_guest_exit);
760
761void vtime_account_idle(struct task_struct *tsk)
762{
763 account_idle_time(get_vtime_delta(&tsk->vtime));
764}
765
766void vtime_task_switch_generic(struct task_struct *prev)
767{
768 struct vtime *vtime = &prev->vtime;
769
770 write_seqcount_begin(&vtime->seqcount);
771 if (vtime->state == VTIME_IDLE)
772 vtime_account_idle(prev);
773 else
774 __vtime_account_kernel(prev, vtime);
775 vtime->state = VTIME_INACTIVE;
776 vtime->cpu = -1;
777 write_seqcount_end(&vtime->seqcount);
778
779 vtime = ¤t->vtime;
780
781 write_seqcount_begin(&vtime->seqcount);
782 if (is_idle_task(current))
783 vtime->state = VTIME_IDLE;
784 else if (current->flags & PF_VCPU)
785 vtime->state = VTIME_GUEST;
786 else
787 vtime->state = VTIME_SYS;
788 vtime->starttime = sched_clock();
789 vtime->cpu = smp_processor_id();
790 write_seqcount_end(&vtime->seqcount);
791}
792
793void vtime_init_idle(struct task_struct *t, int cpu)
794{
795 struct vtime *vtime = &t->vtime;
796 unsigned long flags;
797
798 local_irq_save(flags);
799 write_seqcount_begin(&vtime->seqcount);
800 vtime->state = VTIME_IDLE;
801 vtime->starttime = sched_clock();
802 vtime->cpu = cpu;
803 write_seqcount_end(&vtime->seqcount);
804 local_irq_restore(flags);
805}
806
807u64 task_gtime(struct task_struct *t)
808{
809 struct vtime *vtime = &t->vtime;
810 unsigned int seq;
811 u64 gtime;
812
813 if (!vtime_accounting_enabled())
814 return t->gtime;
815
816 do {
817 seq = read_seqcount_begin(&vtime->seqcount);
818
819 gtime = t->gtime;
820 if (vtime->state == VTIME_GUEST)
821 gtime += vtime->gtime + vtime_delta(vtime);
822
823 } while (read_seqcount_retry(&vtime->seqcount, seq));
824
825 return gtime;
826}
827
828/*
829 * Fetch cputime raw values from fields of task_struct and
830 * add up the pending nohz execution time since the last
831 * cputime snapshot.
832 */
833void task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
834{
835 struct vtime *vtime = &t->vtime;
836 unsigned int seq;
837 u64 delta;
838
839 if (!vtime_accounting_enabled()) {
840 *utime = t->utime;
841 *stime = t->stime;
842 return;
843 }
844
845 do {
846 seq = read_seqcount_begin(&vtime->seqcount);
847
848 *utime = t->utime;
849 *stime = t->stime;
850
851 /* Task is sleeping or idle, nothing to add */
852 if (vtime->state < VTIME_SYS)
853 continue;
854
855 delta = vtime_delta(vtime);
856
857 /*
858 * Task runs either in user (including guest) or kernel space,
859 * add pending nohz time to the right place.
860 */
861 if (vtime->state == VTIME_SYS)
862 *stime += vtime->stime + delta;
863 else
864 *utime += vtime->utime + delta;
865 } while (read_seqcount_retry(&vtime->seqcount, seq));
866}
867
868static int vtime_state_fetch(struct vtime *vtime, int cpu)
869{
870 int state = READ_ONCE(vtime->state);
871
872 /*
873 * We raced against a context switch, fetch the
874 * kcpustat task again.
875 */
876 if (vtime->cpu != cpu && vtime->cpu != -1)
877 return -EAGAIN;
878
879 /*
880 * Two possible things here:
881 * 1) We are seeing the scheduling out task (prev) or any past one.
882 * 2) We are seeing the scheduling in task (next) but it hasn't
883 * passed though vtime_task_switch() yet so the pending
884 * cputime of the prev task may not be flushed yet.
885 *
886 * Case 1) is ok but 2) is not. So wait for a safe VTIME state.
887 */
888 if (state == VTIME_INACTIVE)
889 return -EAGAIN;
890
891 return state;
892}
893
894static u64 kcpustat_user_vtime(struct vtime *vtime)
895{
896 if (vtime->state == VTIME_USER)
897 return vtime->utime + vtime_delta(vtime);
898 else if (vtime->state == VTIME_GUEST)
899 return vtime->gtime + vtime_delta(vtime);
900 return 0;
901}
902
903static int kcpustat_field_vtime(u64 *cpustat,
904 struct task_struct *tsk,
905 enum cpu_usage_stat usage,
906 int cpu, u64 *val)
907{
908 struct vtime *vtime = &tsk->vtime;
909 unsigned int seq;
910
911 do {
912 int state;
913
914 seq = read_seqcount_begin(&vtime->seqcount);
915
916 state = vtime_state_fetch(vtime, cpu);
917 if (state < 0)
918 return state;
919
920 *val = cpustat[usage];
921
922 /*
923 * Nice VS unnice cputime accounting may be inaccurate if
924 * the nice value has changed since the last vtime update.
925 * But proper fix would involve interrupting target on nice
926 * updates which is a no go on nohz_full (although the scheduler
927 * may still interrupt the target if rescheduling is needed...)
928 */
929 switch (usage) {
930 case CPUTIME_SYSTEM:
931 if (state == VTIME_SYS)
932 *val += vtime->stime + vtime_delta(vtime);
933 break;
934 case CPUTIME_USER:
935 if (task_nice(tsk) <= 0)
936 *val += kcpustat_user_vtime(vtime);
937 break;
938 case CPUTIME_NICE:
939 if (task_nice(tsk) > 0)
940 *val += kcpustat_user_vtime(vtime);
941 break;
942 case CPUTIME_GUEST:
943 if (state == VTIME_GUEST && task_nice(tsk) <= 0)
944 *val += vtime->gtime + vtime_delta(vtime);
945 break;
946 case CPUTIME_GUEST_NICE:
947 if (state == VTIME_GUEST && task_nice(tsk) > 0)
948 *val += vtime->gtime + vtime_delta(vtime);
949 break;
950 default:
951 break;
952 }
953 } while (read_seqcount_retry(&vtime->seqcount, seq));
954
955 return 0;
956}
957
958u64 kcpustat_field(struct kernel_cpustat *kcpustat,
959 enum cpu_usage_stat usage, int cpu)
960{
961 u64 *cpustat = kcpustat->cpustat;
962 u64 val = cpustat[usage];
963 struct rq *rq;
964 int err;
965
966 if (!vtime_accounting_enabled_cpu(cpu))
967 return val;
968
969 rq = cpu_rq(cpu);
970
971 for (;;) {
972 struct task_struct *curr;
973
974 rcu_read_lock();
975 curr = rcu_dereference(rq->curr);
976 if (WARN_ON_ONCE(!curr)) {
977 rcu_read_unlock();
978 return cpustat[usage];
979 }
980
981 err = kcpustat_field_vtime(cpustat, curr, usage, cpu, &val);
982 rcu_read_unlock();
983
984 if (!err)
985 return val;
986
987 cpu_relax();
988 }
989}
990EXPORT_SYMBOL_GPL(kcpustat_field);
991
992static int kcpustat_cpu_fetch_vtime(struct kernel_cpustat *dst,
993 const struct kernel_cpustat *src,
994 struct task_struct *tsk, int cpu)
995{
996 struct vtime *vtime = &tsk->vtime;
997 unsigned int seq;
998
999 do {
1000 u64 *cpustat;
1001 u64 delta;
1002 int state;
1003
1004 seq = read_seqcount_begin(&vtime->seqcount);
1005
1006 state = vtime_state_fetch(vtime, cpu);
1007 if (state < 0)
1008 return state;
1009
1010 *dst = *src;
1011 cpustat = dst->cpustat;
1012
1013 /* Task is sleeping, dead or idle, nothing to add */
1014 if (state < VTIME_SYS)
1015 continue;
1016
1017 delta = vtime_delta(vtime);
1018
1019 /*
1020 * Task runs either in user (including guest) or kernel space,
1021 * add pending nohz time to the right place.
1022 */
1023 if (state == VTIME_SYS) {
1024 cpustat[CPUTIME_SYSTEM] += vtime->stime + delta;
1025 } else if (state == VTIME_USER) {
1026 if (task_nice(tsk) > 0)
1027 cpustat[CPUTIME_NICE] += vtime->utime + delta;
1028 else
1029 cpustat[CPUTIME_USER] += vtime->utime + delta;
1030 } else {
1031 WARN_ON_ONCE(state != VTIME_GUEST);
1032 if (task_nice(tsk) > 0) {
1033 cpustat[CPUTIME_GUEST_NICE] += vtime->gtime + delta;
1034 cpustat[CPUTIME_NICE] += vtime->gtime + delta;
1035 } else {
1036 cpustat[CPUTIME_GUEST] += vtime->gtime + delta;
1037 cpustat[CPUTIME_USER] += vtime->gtime + delta;
1038 }
1039 }
1040 } while (read_seqcount_retry(&vtime->seqcount, seq));
1041
1042 return 0;
1043}
1044
1045void kcpustat_cpu_fetch(struct kernel_cpustat *dst, int cpu)
1046{
1047 const struct kernel_cpustat *src = &kcpustat_cpu(cpu);
1048 struct rq *rq;
1049 int err;
1050
1051 if (!vtime_accounting_enabled_cpu(cpu)) {
1052 *dst = *src;
1053 return;
1054 }
1055
1056 rq = cpu_rq(cpu);
1057
1058 for (;;) {
1059 struct task_struct *curr;
1060
1061 rcu_read_lock();
1062 curr = rcu_dereference(rq->curr);
1063 if (WARN_ON_ONCE(!curr)) {
1064 rcu_read_unlock();
1065 *dst = *src;
1066 return;
1067 }
1068
1069 err = kcpustat_cpu_fetch_vtime(dst, src, curr, cpu);
1070 rcu_read_unlock();
1071
1072 if (!err)
1073 return;
1074
1075 cpu_relax();
1076 }
1077}
1078EXPORT_SYMBOL_GPL(kcpustat_cpu_fetch);
1079
1080#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */