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#include <linux/export.h>
2#include <linux/sched.h>
3#include <linux/tsacct_kern.h>
4#include <linux/kernel_stat.h>
5#include <linux/static_key.h>
6#include <linux/context_tracking.h>
7#include "sched.h"
8#ifdef CONFIG_PARAVIRT
9#include <asm/paravirt.h>
10#endif
11
12
13#ifdef CONFIG_IRQ_TIME_ACCOUNTING
14
15/*
16 * There are no locks covering percpu hardirq/softirq time.
17 * They are only modified in vtime_account, on corresponding CPU
18 * with interrupts disabled. So, writes are safe.
19 * They are read and saved off onto struct rq in update_rq_clock().
20 * This may result in other CPU reading this CPU's irq time and can
21 * race with irq/vtime_account on this CPU. We would either get old
22 * or new value with a side effect of accounting a slice of irq time to wrong
23 * task when irq is in progress while we read rq->clock. That is a worthy
24 * compromise in place of having locks on each irq in account_system_time.
25 */
26DEFINE_PER_CPU(struct irqtime, cpu_irqtime);
27
28static int sched_clock_irqtime;
29
30void enable_sched_clock_irqtime(void)
31{
32 sched_clock_irqtime = 1;
33}
34
35void disable_sched_clock_irqtime(void)
36{
37 sched_clock_irqtime = 0;
38}
39
40/*
41 * Called before incrementing preempt_count on {soft,}irq_enter
42 * and before decrementing preempt_count on {soft,}irq_exit.
43 */
44void irqtime_account_irq(struct task_struct *curr)
45{
46 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
47 s64 delta;
48 int cpu;
49
50 if (!sched_clock_irqtime)
51 return;
52
53 cpu = smp_processor_id();
54 delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
55 irqtime->irq_start_time += delta;
56
57 u64_stats_update_begin(&irqtime->sync);
58 /*
59 * We do not account for softirq time from ksoftirqd here.
60 * We want to continue accounting softirq time to ksoftirqd thread
61 * in that case, so as not to confuse scheduler with a special task
62 * that do not consume any time, but still wants to run.
63 */
64 if (hardirq_count())
65 irqtime->hardirq_time += delta;
66 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
67 irqtime->softirq_time += delta;
68
69 u64_stats_update_end(&irqtime->sync);
70}
71EXPORT_SYMBOL_GPL(irqtime_account_irq);
72
73static cputime_t irqtime_account_update(u64 irqtime, int idx, cputime_t maxtime)
74{
75 u64 *cpustat = kcpustat_this_cpu->cpustat;
76 cputime_t irq_cputime;
77
78 irq_cputime = nsecs_to_cputime64(irqtime) - cpustat[idx];
79 irq_cputime = min(irq_cputime, maxtime);
80 cpustat[idx] += irq_cputime;
81
82 return irq_cputime;
83}
84
85static cputime_t irqtime_account_hi_update(cputime_t maxtime)
86{
87 return irqtime_account_update(__this_cpu_read(cpu_irqtime.hardirq_time),
88 CPUTIME_IRQ, maxtime);
89}
90
91static cputime_t irqtime_account_si_update(cputime_t maxtime)
92{
93 return irqtime_account_update(__this_cpu_read(cpu_irqtime.softirq_time),
94 CPUTIME_SOFTIRQ, maxtime);
95}
96
97#else /* CONFIG_IRQ_TIME_ACCOUNTING */
98
99#define sched_clock_irqtime (0)
100
101static cputime_t irqtime_account_hi_update(cputime_t dummy)
102{
103 return 0;
104}
105
106static cputime_t irqtime_account_si_update(cputime_t dummy)
107{
108 return 0;
109}
110
111#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
112
113static inline void task_group_account_field(struct task_struct *p, int index,
114 u64 tmp)
115{
116 /*
117 * Since all updates are sure to touch the root cgroup, we
118 * get ourselves ahead and touch it first. If the root cgroup
119 * is the only cgroup, then nothing else should be necessary.
120 *
121 */
122 __this_cpu_add(kernel_cpustat.cpustat[index], tmp);
123
124 cpuacct_account_field(p, index, tmp);
125}
126
127/*
128 * Account user cpu time to a process.
129 * @p: the process that the cpu time gets accounted to
130 * @cputime: the cpu time spent in user space since the last update
131 */
132void account_user_time(struct task_struct *p, cputime_t cputime)
133{
134 int index;
135
136 /* Add user time to process. */
137 p->utime += cputime;
138 account_group_user_time(p, cputime);
139
140 index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
141
142 /* Add user time to cpustat. */
143 task_group_account_field(p, index, (__force u64) cputime);
144
145 /* Account for user time used */
146 acct_account_cputime(p);
147}
148
149/*
150 * Account guest cpu time to a process.
151 * @p: the process that the cpu time gets accounted to
152 * @cputime: the cpu time spent in virtual machine since the last update
153 */
154static void account_guest_time(struct task_struct *p, cputime_t cputime)
155{
156 u64 *cpustat = kcpustat_this_cpu->cpustat;
157
158 /* Add guest time to process. */
159 p->utime += cputime;
160 account_group_user_time(p, cputime);
161 p->gtime += cputime;
162
163 /* Add guest time to cpustat. */
164 if (task_nice(p) > 0) {
165 cpustat[CPUTIME_NICE] += (__force u64) cputime;
166 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
167 } else {
168 cpustat[CPUTIME_USER] += (__force u64) cputime;
169 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
170 }
171}
172
173/*
174 * Account system cpu time to a process and desired cpustat field
175 * @p: the process that the cpu time gets accounted to
176 * @cputime: the cpu time spent in kernel space since the last update
177 * @index: pointer to cpustat field that has to be updated
178 */
179static inline
180void __account_system_time(struct task_struct *p, cputime_t cputime, int index)
181{
182 /* Add system time to process. */
183 p->stime += cputime;
184 account_group_system_time(p, cputime);
185
186 /* Add system time to cpustat. */
187 task_group_account_field(p, index, (__force u64) cputime);
188
189 /* Account for system time used */
190 acct_account_cputime(p);
191}
192
193/*
194 * Account system cpu time to a process.
195 * @p: the process that the cpu time gets accounted to
196 * @hardirq_offset: the offset to subtract from hardirq_count()
197 * @cputime: the cpu time spent in kernel space since the last update
198 */
199void account_system_time(struct task_struct *p, int hardirq_offset,
200 cputime_t cputime)
201{
202 int index;
203
204 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
205 account_guest_time(p, cputime);
206 return;
207 }
208
209 if (hardirq_count() - hardirq_offset)
210 index = CPUTIME_IRQ;
211 else if (in_serving_softirq())
212 index = CPUTIME_SOFTIRQ;
213 else
214 index = CPUTIME_SYSTEM;
215
216 __account_system_time(p, cputime, index);
217}
218
219/*
220 * Account for involuntary wait time.
221 * @cputime: the cpu time spent in involuntary wait
222 */
223void account_steal_time(cputime_t cputime)
224{
225 u64 *cpustat = kcpustat_this_cpu->cpustat;
226
227 cpustat[CPUTIME_STEAL] += (__force u64) cputime;
228}
229
230/*
231 * Account for idle time.
232 * @cputime: the cpu time spent in idle wait
233 */
234void account_idle_time(cputime_t cputime)
235{
236 u64 *cpustat = kcpustat_this_cpu->cpustat;
237 struct rq *rq = this_rq();
238
239 if (atomic_read(&rq->nr_iowait) > 0)
240 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
241 else
242 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
243}
244
245/*
246 * When a guest is interrupted for a longer amount of time, missed clock
247 * ticks are not redelivered later. Due to that, this function may on
248 * occasion account more time than the calling functions think elapsed.
249 */
250static __always_inline cputime_t steal_account_process_time(cputime_t maxtime)
251{
252#ifdef CONFIG_PARAVIRT
253 if (static_key_false(¶virt_steal_enabled)) {
254 cputime_t steal_cputime;
255 u64 steal;
256
257 steal = paravirt_steal_clock(smp_processor_id());
258 steal -= this_rq()->prev_steal_time;
259
260 steal_cputime = min(nsecs_to_cputime(steal), maxtime);
261 account_steal_time(steal_cputime);
262 this_rq()->prev_steal_time += cputime_to_nsecs(steal_cputime);
263
264 return steal_cputime;
265 }
266#endif
267 return 0;
268}
269
270/*
271 * Account how much elapsed time was spent in steal, irq, or softirq time.
272 */
273static inline cputime_t account_other_time(cputime_t max)
274{
275 cputime_t accounted;
276
277 /* Shall be converted to a lockdep-enabled lightweight check */
278 WARN_ON_ONCE(!irqs_disabled());
279
280 accounted = steal_account_process_time(max);
281
282 if (accounted < max)
283 accounted += irqtime_account_hi_update(max - accounted);
284
285 if (accounted < max)
286 accounted += irqtime_account_si_update(max - accounted);
287
288 return accounted;
289}
290
291#ifdef CONFIG_64BIT
292static inline u64 read_sum_exec_runtime(struct task_struct *t)
293{
294 return t->se.sum_exec_runtime;
295}
296#else
297static u64 read_sum_exec_runtime(struct task_struct *t)
298{
299 u64 ns;
300 struct rq_flags rf;
301 struct rq *rq;
302
303 rq = task_rq_lock(t, &rf);
304 ns = t->se.sum_exec_runtime;
305 task_rq_unlock(rq, t, &rf);
306
307 return ns;
308}
309#endif
310
311/*
312 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
313 * tasks (sum on group iteration) belonging to @tsk's group.
314 */
315void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
316{
317 struct signal_struct *sig = tsk->signal;
318 cputime_t utime, stime;
319 struct task_struct *t;
320 unsigned int seq, nextseq;
321 unsigned long flags;
322
323 /*
324 * Update current task runtime to account pending time since last
325 * scheduler action or thread_group_cputime() call. This thread group
326 * might have other running tasks on different CPUs, but updating
327 * their runtime can affect syscall performance, so we skip account
328 * those pending times and rely only on values updated on tick or
329 * other scheduler action.
330 */
331 if (same_thread_group(current, tsk))
332 (void) task_sched_runtime(current);
333
334 rcu_read_lock();
335 /* Attempt a lockless read on the first round. */
336 nextseq = 0;
337 do {
338 seq = nextseq;
339 flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
340 times->utime = sig->utime;
341 times->stime = sig->stime;
342 times->sum_exec_runtime = sig->sum_sched_runtime;
343
344 for_each_thread(tsk, t) {
345 task_cputime(t, &utime, &stime);
346 times->utime += utime;
347 times->stime += stime;
348 times->sum_exec_runtime += read_sum_exec_runtime(t);
349 }
350 /* If lockless access failed, take the lock. */
351 nextseq = 1;
352 } while (need_seqretry(&sig->stats_lock, seq));
353 done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
354 rcu_read_unlock();
355}
356
357#ifdef CONFIG_IRQ_TIME_ACCOUNTING
358/*
359 * Account a tick to a process and cpustat
360 * @p: the process that the cpu time gets accounted to
361 * @user_tick: is the tick from userspace
362 * @rq: the pointer to rq
363 *
364 * Tick demultiplexing follows the order
365 * - pending hardirq update
366 * - pending softirq update
367 * - user_time
368 * - idle_time
369 * - system time
370 * - check for guest_time
371 * - else account as system_time
372 *
373 * Check for hardirq is done both for system and user time as there is
374 * no timer going off while we are on hardirq and hence we may never get an
375 * opportunity to update it solely in system time.
376 * p->stime and friends are only updated on system time and not on irq
377 * softirq as those do not count in task exec_runtime any more.
378 */
379static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
380 struct rq *rq, int ticks)
381{
382 u64 cputime = (__force u64) cputime_one_jiffy * ticks;
383 cputime_t other;
384
385 /*
386 * When returning from idle, many ticks can get accounted at
387 * once, including some ticks of steal, irq, and softirq time.
388 * Subtract those ticks from the amount of time accounted to
389 * idle, or potentially user or system time. Due to rounding,
390 * other time can exceed ticks occasionally.
391 */
392 other = account_other_time(ULONG_MAX);
393 if (other >= cputime)
394 return;
395 cputime -= other;
396
397 if (this_cpu_ksoftirqd() == p) {
398 /*
399 * ksoftirqd time do not get accounted in cpu_softirq_time.
400 * So, we have to handle it separately here.
401 * Also, p->stime needs to be updated for ksoftirqd.
402 */
403 __account_system_time(p, cputime, CPUTIME_SOFTIRQ);
404 } else if (user_tick) {
405 account_user_time(p, cputime);
406 } else if (p == rq->idle) {
407 account_idle_time(cputime);
408 } else if (p->flags & PF_VCPU) { /* System time or guest time */
409 account_guest_time(p, cputime);
410 } else {
411 __account_system_time(p, cputime, CPUTIME_SYSTEM);
412 }
413}
414
415static void irqtime_account_idle_ticks(int ticks)
416{
417 struct rq *rq = this_rq();
418
419 irqtime_account_process_tick(current, 0, rq, ticks);
420}
421#else /* CONFIG_IRQ_TIME_ACCOUNTING */
422static inline void irqtime_account_idle_ticks(int ticks) {}
423static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
424 struct rq *rq, int nr_ticks) {}
425#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
426
427/*
428 * Use precise platform statistics if available:
429 */
430#ifdef CONFIG_VIRT_CPU_ACCOUNTING
431
432#ifndef __ARCH_HAS_VTIME_TASK_SWITCH
433void vtime_common_task_switch(struct task_struct *prev)
434{
435 if (is_idle_task(prev))
436 vtime_account_idle(prev);
437 else
438 vtime_account_system(prev);
439
440#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
441 vtime_account_user(prev);
442#endif
443 arch_vtime_task_switch(prev);
444}
445#endif
446
447#endif /* CONFIG_VIRT_CPU_ACCOUNTING */
448
449
450#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
451/*
452 * Archs that account the whole time spent in the idle task
453 * (outside irq) as idle time can rely on this and just implement
454 * vtime_account_system() and vtime_account_idle(). Archs that
455 * have other meaning of the idle time (s390 only includes the
456 * time spent by the CPU when it's in low power mode) must override
457 * vtime_account().
458 */
459#ifndef __ARCH_HAS_VTIME_ACCOUNT
460void vtime_account_irq_enter(struct task_struct *tsk)
461{
462 if (!in_interrupt() && is_idle_task(tsk))
463 vtime_account_idle(tsk);
464 else
465 vtime_account_system(tsk);
466}
467EXPORT_SYMBOL_GPL(vtime_account_irq_enter);
468#endif /* __ARCH_HAS_VTIME_ACCOUNT */
469
470void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
471{
472 *ut = p->utime;
473 *st = p->stime;
474}
475EXPORT_SYMBOL_GPL(task_cputime_adjusted);
476
477void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
478{
479 struct task_cputime cputime;
480
481 thread_group_cputime(p, &cputime);
482
483 *ut = cputime.utime;
484 *st = cputime.stime;
485}
486#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
487/*
488 * Account a single tick of cpu time.
489 * @p: the process that the cpu time gets accounted to
490 * @user_tick: indicates if the tick is a user or a system tick
491 */
492void account_process_tick(struct task_struct *p, int user_tick)
493{
494 cputime_t cputime, steal;
495 struct rq *rq = this_rq();
496
497 if (vtime_accounting_cpu_enabled())
498 return;
499
500 if (sched_clock_irqtime) {
501 irqtime_account_process_tick(p, user_tick, rq, 1);
502 return;
503 }
504
505 cputime = cputime_one_jiffy;
506 steal = steal_account_process_time(ULONG_MAX);
507
508 if (steal >= cputime)
509 return;
510
511 cputime -= steal;
512
513 if (user_tick)
514 account_user_time(p, cputime);
515 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
516 account_system_time(p, HARDIRQ_OFFSET, cputime);
517 else
518 account_idle_time(cputime);
519}
520
521/*
522 * Account multiple ticks of idle time.
523 * @ticks: number of stolen ticks
524 */
525void account_idle_ticks(unsigned long ticks)
526{
527 cputime_t cputime, steal;
528
529 if (sched_clock_irqtime) {
530 irqtime_account_idle_ticks(ticks);
531 return;
532 }
533
534 cputime = jiffies_to_cputime(ticks);
535 steal = steal_account_process_time(ULONG_MAX);
536
537 if (steal >= cputime)
538 return;
539
540 cputime -= steal;
541 account_idle_time(cputime);
542}
543
544/*
545 * Perform (stime * rtime) / total, but avoid multiplication overflow by
546 * loosing precision when the numbers are big.
547 */
548static cputime_t scale_stime(u64 stime, u64 rtime, u64 total)
549{
550 u64 scaled;
551
552 for (;;) {
553 /* Make sure "rtime" is the bigger of stime/rtime */
554 if (stime > rtime)
555 swap(rtime, stime);
556
557 /* Make sure 'total' fits in 32 bits */
558 if (total >> 32)
559 goto drop_precision;
560
561 /* Does rtime (and thus stime) fit in 32 bits? */
562 if (!(rtime >> 32))
563 break;
564
565 /* Can we just balance rtime/stime rather than dropping bits? */
566 if (stime >> 31)
567 goto drop_precision;
568
569 /* We can grow stime and shrink rtime and try to make them both fit */
570 stime <<= 1;
571 rtime >>= 1;
572 continue;
573
574drop_precision:
575 /* We drop from rtime, it has more bits than stime */
576 rtime >>= 1;
577 total >>= 1;
578 }
579
580 /*
581 * Make sure gcc understands that this is a 32x32->64 multiply,
582 * followed by a 64/32->64 divide.
583 */
584 scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
585 return (__force cputime_t) scaled;
586}
587
588/*
589 * Adjust tick based cputime random precision against scheduler runtime
590 * accounting.
591 *
592 * Tick based cputime accounting depend on random scheduling timeslices of a
593 * task to be interrupted or not by the timer. Depending on these
594 * circumstances, the number of these interrupts may be over or
595 * under-optimistic, matching the real user and system cputime with a variable
596 * precision.
597 *
598 * Fix this by scaling these tick based values against the total runtime
599 * accounted by the CFS scheduler.
600 *
601 * This code provides the following guarantees:
602 *
603 * stime + utime == rtime
604 * stime_i+1 >= stime_i, utime_i+1 >= utime_i
605 *
606 * Assuming that rtime_i+1 >= rtime_i.
607 */
608static void cputime_adjust(struct task_cputime *curr,
609 struct prev_cputime *prev,
610 cputime_t *ut, cputime_t *st)
611{
612 cputime_t rtime, stime, utime;
613 unsigned long flags;
614
615 /* Serialize concurrent callers such that we can honour our guarantees */
616 raw_spin_lock_irqsave(&prev->lock, flags);
617 rtime = nsecs_to_cputime(curr->sum_exec_runtime);
618
619 /*
620 * This is possible under two circumstances:
621 * - rtime isn't monotonic after all (a bug);
622 * - we got reordered by the lock.
623 *
624 * In both cases this acts as a filter such that the rest of the code
625 * can assume it is monotonic regardless of anything else.
626 */
627 if (prev->stime + prev->utime >= rtime)
628 goto out;
629
630 stime = curr->stime;
631 utime = curr->utime;
632
633 /*
634 * If either stime or both stime and utime are 0, assume all runtime is
635 * userspace. Once a task gets some ticks, the monotonicy code at
636 * 'update' will ensure things converge to the observed ratio.
637 */
638 if (stime == 0) {
639 utime = rtime;
640 goto update;
641 }
642
643 if (utime == 0) {
644 stime = rtime;
645 goto update;
646 }
647
648 stime = scale_stime((__force u64)stime, (__force u64)rtime,
649 (__force u64)(stime + utime));
650
651update:
652 /*
653 * Make sure stime doesn't go backwards; this preserves monotonicity
654 * for utime because rtime is monotonic.
655 *
656 * utime_i+1 = rtime_i+1 - stime_i
657 * = rtime_i+1 - (rtime_i - utime_i)
658 * = (rtime_i+1 - rtime_i) + utime_i
659 * >= utime_i
660 */
661 if (stime < prev->stime)
662 stime = prev->stime;
663 utime = rtime - stime;
664
665 /*
666 * Make sure utime doesn't go backwards; this still preserves
667 * monotonicity for stime, analogous argument to above.
668 */
669 if (utime < prev->utime) {
670 utime = prev->utime;
671 stime = rtime - utime;
672 }
673
674 prev->stime = stime;
675 prev->utime = utime;
676out:
677 *ut = prev->utime;
678 *st = prev->stime;
679 raw_spin_unlock_irqrestore(&prev->lock, flags);
680}
681
682void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
683{
684 struct task_cputime cputime = {
685 .sum_exec_runtime = p->se.sum_exec_runtime,
686 };
687
688 task_cputime(p, &cputime.utime, &cputime.stime);
689 cputime_adjust(&cputime, &p->prev_cputime, ut, st);
690}
691EXPORT_SYMBOL_GPL(task_cputime_adjusted);
692
693void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
694{
695 struct task_cputime cputime;
696
697 thread_group_cputime(p, &cputime);
698 cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
699}
700#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
701
702#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
703static cputime_t vtime_delta(struct task_struct *tsk)
704{
705 unsigned long now = READ_ONCE(jiffies);
706
707 if (time_before(now, (unsigned long)tsk->vtime_snap))
708 return 0;
709
710 return jiffies_to_cputime(now - tsk->vtime_snap);
711}
712
713static cputime_t get_vtime_delta(struct task_struct *tsk)
714{
715 unsigned long now = READ_ONCE(jiffies);
716 cputime_t delta, other;
717
718 /*
719 * Unlike tick based timing, vtime based timing never has lost
720 * ticks, and no need for steal time accounting to make up for
721 * lost ticks. Vtime accounts a rounded version of actual
722 * elapsed time. Limit account_other_time to prevent rounding
723 * errors from causing elapsed vtime to go negative.
724 */
725 delta = jiffies_to_cputime(now - tsk->vtime_snap);
726 other = account_other_time(delta);
727 WARN_ON_ONCE(tsk->vtime_snap_whence == VTIME_INACTIVE);
728 tsk->vtime_snap = now;
729
730 return delta - other;
731}
732
733static void __vtime_account_system(struct task_struct *tsk)
734{
735 cputime_t delta_cpu = get_vtime_delta(tsk);
736
737 account_system_time(tsk, irq_count(), delta_cpu);
738}
739
740void vtime_account_system(struct task_struct *tsk)
741{
742 if (!vtime_delta(tsk))
743 return;
744
745 write_seqcount_begin(&tsk->vtime_seqcount);
746 __vtime_account_system(tsk);
747 write_seqcount_end(&tsk->vtime_seqcount);
748}
749
750void vtime_account_user(struct task_struct *tsk)
751{
752 cputime_t delta_cpu;
753
754 write_seqcount_begin(&tsk->vtime_seqcount);
755 tsk->vtime_snap_whence = VTIME_SYS;
756 if (vtime_delta(tsk)) {
757 delta_cpu = get_vtime_delta(tsk);
758 account_user_time(tsk, delta_cpu);
759 }
760 write_seqcount_end(&tsk->vtime_seqcount);
761}
762
763void vtime_user_enter(struct task_struct *tsk)
764{
765 write_seqcount_begin(&tsk->vtime_seqcount);
766 if (vtime_delta(tsk))
767 __vtime_account_system(tsk);
768 tsk->vtime_snap_whence = VTIME_USER;
769 write_seqcount_end(&tsk->vtime_seqcount);
770}
771
772void vtime_guest_enter(struct task_struct *tsk)
773{
774 /*
775 * The flags must be updated under the lock with
776 * the vtime_snap flush and update.
777 * That enforces a right ordering and update sequence
778 * synchronization against the reader (task_gtime())
779 * that can thus safely catch up with a tickless delta.
780 */
781 write_seqcount_begin(&tsk->vtime_seqcount);
782 if (vtime_delta(tsk))
783 __vtime_account_system(tsk);
784 current->flags |= PF_VCPU;
785 write_seqcount_end(&tsk->vtime_seqcount);
786}
787EXPORT_SYMBOL_GPL(vtime_guest_enter);
788
789void vtime_guest_exit(struct task_struct *tsk)
790{
791 write_seqcount_begin(&tsk->vtime_seqcount);
792 __vtime_account_system(tsk);
793 current->flags &= ~PF_VCPU;
794 write_seqcount_end(&tsk->vtime_seqcount);
795}
796EXPORT_SYMBOL_GPL(vtime_guest_exit);
797
798void vtime_account_idle(struct task_struct *tsk)
799{
800 cputime_t delta_cpu = get_vtime_delta(tsk);
801
802 account_idle_time(delta_cpu);
803}
804
805void arch_vtime_task_switch(struct task_struct *prev)
806{
807 write_seqcount_begin(&prev->vtime_seqcount);
808 prev->vtime_snap_whence = VTIME_INACTIVE;
809 write_seqcount_end(&prev->vtime_seqcount);
810
811 write_seqcount_begin(¤t->vtime_seqcount);
812 current->vtime_snap_whence = VTIME_SYS;
813 current->vtime_snap = jiffies;
814 write_seqcount_end(¤t->vtime_seqcount);
815}
816
817void vtime_init_idle(struct task_struct *t, int cpu)
818{
819 unsigned long flags;
820
821 local_irq_save(flags);
822 write_seqcount_begin(&t->vtime_seqcount);
823 t->vtime_snap_whence = VTIME_SYS;
824 t->vtime_snap = jiffies;
825 write_seqcount_end(&t->vtime_seqcount);
826 local_irq_restore(flags);
827}
828
829cputime_t task_gtime(struct task_struct *t)
830{
831 unsigned int seq;
832 cputime_t gtime;
833
834 if (!vtime_accounting_enabled())
835 return t->gtime;
836
837 do {
838 seq = read_seqcount_begin(&t->vtime_seqcount);
839
840 gtime = t->gtime;
841 if (t->vtime_snap_whence == VTIME_SYS && t->flags & PF_VCPU)
842 gtime += vtime_delta(t);
843
844 } while (read_seqcount_retry(&t->vtime_seqcount, seq));
845
846 return gtime;
847}
848
849/*
850 * Fetch cputime raw values from fields of task_struct and
851 * add up the pending nohz execution time since the last
852 * cputime snapshot.
853 */
854void task_cputime(struct task_struct *t, cputime_t *utime, cputime_t *stime)
855{
856 cputime_t delta;
857 unsigned int seq;
858
859 if (!vtime_accounting_enabled()) {
860 *utime = t->utime;
861 *stime = t->stime;
862 return;
863 }
864
865 do {
866 seq = read_seqcount_begin(&t->vtime_seqcount);
867
868 *utime = t->utime;
869 *stime = t->stime;
870
871 /* Task is sleeping, nothing to add */
872 if (t->vtime_snap_whence == VTIME_INACTIVE || is_idle_task(t))
873 continue;
874
875 delta = vtime_delta(t);
876
877 /*
878 * Task runs either in user or kernel space, add pending nohz time to
879 * the right place.
880 */
881 if (t->vtime_snap_whence == VTIME_USER || t->flags & PF_VCPU)
882 *utime += delta;
883 else if (t->vtime_snap_whence == VTIME_SYS)
884 *stime += delta;
885 } while (read_seqcount_retry(&t->vtime_seqcount, seq));
886}
887#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */