Loading...
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 struct rq *rq, 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 == 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 struct rq *rq = this_rq();
396
397 irqtime_account_process_tick(current, 0, rq, ticks);
398}
399#else /* CONFIG_IRQ_TIME_ACCOUNTING */
400static inline void irqtime_account_idle_ticks(int ticks) { }
401static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
402 struct rq *rq, int nr_ticks) { }
403#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
404
405/*
406 * Use precise platform statistics if available:
407 */
408#ifdef CONFIG_VIRT_CPU_ACCOUNTING
409# ifndef __ARCH_HAS_VTIME_TASK_SWITCH
410void vtime_common_task_switch(struct task_struct *prev)
411{
412 if (is_idle_task(prev))
413 vtime_account_idle(prev);
414 else
415 vtime_account_system(prev);
416
417 vtime_flush(prev);
418 arch_vtime_task_switch(prev);
419}
420# endif
421#endif /* CONFIG_VIRT_CPU_ACCOUNTING */
422
423
424#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
425/*
426 * Archs that account the whole time spent in the idle task
427 * (outside irq) as idle time can rely on this and just implement
428 * vtime_account_system() and vtime_account_idle(). Archs that
429 * have other meaning of the idle time (s390 only includes the
430 * time spent by the CPU when it's in low power mode) must override
431 * vtime_account().
432 */
433#ifndef __ARCH_HAS_VTIME_ACCOUNT
434void vtime_account_irq_enter(struct task_struct *tsk)
435{
436 if (!in_interrupt() && is_idle_task(tsk))
437 vtime_account_idle(tsk);
438 else
439 vtime_account_system(tsk);
440}
441EXPORT_SYMBOL_GPL(vtime_account_irq_enter);
442#endif /* __ARCH_HAS_VTIME_ACCOUNT */
443
444void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
445 u64 *ut, u64 *st)
446{
447 *ut = curr->utime;
448 *st = curr->stime;
449}
450
451void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
452{
453 *ut = p->utime;
454 *st = p->stime;
455}
456EXPORT_SYMBOL_GPL(task_cputime_adjusted);
457
458void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
459{
460 struct task_cputime cputime;
461
462 thread_group_cputime(p, &cputime);
463
464 *ut = cputime.utime;
465 *st = cputime.stime;
466}
467
468#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE: */
469
470/*
471 * Account a single tick of CPU time.
472 * @p: the process that the CPU time gets accounted to
473 * @user_tick: indicates if the tick is a user or a system tick
474 */
475void account_process_tick(struct task_struct *p, int user_tick)
476{
477 u64 cputime, steal;
478 struct rq *rq = this_rq();
479
480 if (vtime_accounting_cpu_enabled())
481 return;
482
483 if (sched_clock_irqtime) {
484 irqtime_account_process_tick(p, user_tick, rq, 1);
485 return;
486 }
487
488 cputime = TICK_NSEC;
489 steal = steal_account_process_time(ULONG_MAX);
490
491 if (steal >= cputime)
492 return;
493
494 cputime -= steal;
495
496 if (user_tick)
497 account_user_time(p, cputime);
498 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
499 account_system_time(p, HARDIRQ_OFFSET, cputime);
500 else
501 account_idle_time(cputime);
502}
503
504/*
505 * Account multiple ticks of idle time.
506 * @ticks: number of stolen ticks
507 */
508void account_idle_ticks(unsigned long ticks)
509{
510 u64 cputime, steal;
511
512 if (sched_clock_irqtime) {
513 irqtime_account_idle_ticks(ticks);
514 return;
515 }
516
517 cputime = ticks * TICK_NSEC;
518 steal = steal_account_process_time(ULONG_MAX);
519
520 if (steal >= cputime)
521 return;
522
523 cputime -= steal;
524 account_idle_time(cputime);
525}
526
527/*
528 * Perform (stime * rtime) / total, but avoid multiplication overflow by
529 * losing precision when the numbers are big.
530 */
531static u64 scale_stime(u64 stime, u64 rtime, u64 total)
532{
533 u64 scaled;
534
535 for (;;) {
536 /* Make sure "rtime" is the bigger of stime/rtime */
537 if (stime > rtime)
538 swap(rtime, stime);
539
540 /* Make sure 'total' fits in 32 bits */
541 if (total >> 32)
542 goto drop_precision;
543
544 /* Does rtime (and thus stime) fit in 32 bits? */
545 if (!(rtime >> 32))
546 break;
547
548 /* Can we just balance rtime/stime rather than dropping bits? */
549 if (stime >> 31)
550 goto drop_precision;
551
552 /* We can grow stime and shrink rtime and try to make them both fit */
553 stime <<= 1;
554 rtime >>= 1;
555 continue;
556
557drop_precision:
558 /* We drop from rtime, it has more bits than stime */
559 rtime >>= 1;
560 total >>= 1;
561 }
562
563 /*
564 * Make sure gcc understands that this is a 32x32->64 multiply,
565 * followed by a 64/32->64 divide.
566 */
567 scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
568 return scaled;
569}
570
571/*
572 * Adjust tick based cputime random precision against scheduler runtime
573 * accounting.
574 *
575 * Tick based cputime accounting depend on random scheduling timeslices of a
576 * task to be interrupted or not by the timer. Depending on these
577 * circumstances, the number of these interrupts may be over or
578 * under-optimistic, matching the real user and system cputime with a variable
579 * precision.
580 *
581 * Fix this by scaling these tick based values against the total runtime
582 * accounted by the CFS scheduler.
583 *
584 * This code provides the following guarantees:
585 *
586 * stime + utime == rtime
587 * stime_i+1 >= stime_i, utime_i+1 >= utime_i
588 *
589 * Assuming that rtime_i+1 >= rtime_i.
590 */
591void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
592 u64 *ut, u64 *st)
593{
594 u64 rtime, stime, utime;
595 unsigned long flags;
596
597 /* Serialize concurrent callers such that we can honour our guarantees */
598 raw_spin_lock_irqsave(&prev->lock, flags);
599 rtime = curr->sum_exec_runtime;
600
601 /*
602 * This is possible under two circumstances:
603 * - rtime isn't monotonic after all (a bug);
604 * - we got reordered by the lock.
605 *
606 * In both cases this acts as a filter such that the rest of the code
607 * can assume it is monotonic regardless of anything else.
608 */
609 if (prev->stime + prev->utime >= rtime)
610 goto out;
611
612 stime = curr->stime;
613 utime = curr->utime;
614
615 /*
616 * If either stime or utime are 0, assume all runtime is userspace.
617 * Once a task gets some ticks, the monotonicy code at 'update:'
618 * will ensure things converge to the observed ratio.
619 */
620 if (stime == 0) {
621 utime = rtime;
622 goto update;
623 }
624
625 if (utime == 0) {
626 stime = rtime;
627 goto update;
628 }
629
630 stime = scale_stime(stime, rtime, stime + utime);
631
632update:
633 /*
634 * Make sure stime doesn't go backwards; this preserves monotonicity
635 * for utime because rtime is monotonic.
636 *
637 * utime_i+1 = rtime_i+1 - stime_i
638 * = rtime_i+1 - (rtime_i - utime_i)
639 * = (rtime_i+1 - rtime_i) + utime_i
640 * >= utime_i
641 */
642 if (stime < prev->stime)
643 stime = prev->stime;
644 utime = rtime - stime;
645
646 /*
647 * Make sure utime doesn't go backwards; this still preserves
648 * monotonicity for stime, analogous argument to above.
649 */
650 if (utime < prev->utime) {
651 utime = prev->utime;
652 stime = rtime - utime;
653 }
654
655 prev->stime = stime;
656 prev->utime = utime;
657out:
658 *ut = prev->utime;
659 *st = prev->stime;
660 raw_spin_unlock_irqrestore(&prev->lock, flags);
661}
662
663void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
664{
665 struct task_cputime cputime = {
666 .sum_exec_runtime = p->se.sum_exec_runtime,
667 };
668
669 task_cputime(p, &cputime.utime, &cputime.stime);
670 cputime_adjust(&cputime, &p->prev_cputime, ut, st);
671}
672EXPORT_SYMBOL_GPL(task_cputime_adjusted);
673
674void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
675{
676 struct task_cputime cputime;
677
678 thread_group_cputime(p, &cputime);
679 cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
680}
681#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
682
683#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
684static u64 vtime_delta(struct vtime *vtime)
685{
686 unsigned long long clock;
687
688 clock = sched_clock();
689 if (clock < vtime->starttime)
690 return 0;
691
692 return clock - vtime->starttime;
693}
694
695static u64 get_vtime_delta(struct vtime *vtime)
696{
697 u64 delta = vtime_delta(vtime);
698 u64 other;
699
700 /*
701 * Unlike tick based timing, vtime based timing never has lost
702 * ticks, and no need for steal time accounting to make up for
703 * lost ticks. Vtime accounts a rounded version of actual
704 * elapsed time. Limit account_other_time to prevent rounding
705 * errors from causing elapsed vtime to go negative.
706 */
707 other = account_other_time(delta);
708 WARN_ON_ONCE(vtime->state == VTIME_INACTIVE);
709 vtime->starttime += delta;
710
711 return delta - other;
712}
713
714static void __vtime_account_system(struct task_struct *tsk,
715 struct vtime *vtime)
716{
717 vtime->stime += get_vtime_delta(vtime);
718 if (vtime->stime >= TICK_NSEC) {
719 account_system_time(tsk, irq_count(), vtime->stime);
720 vtime->stime = 0;
721 }
722}
723
724static void vtime_account_guest(struct task_struct *tsk,
725 struct vtime *vtime)
726{
727 vtime->gtime += get_vtime_delta(vtime);
728 if (vtime->gtime >= TICK_NSEC) {
729 account_guest_time(tsk, vtime->gtime);
730 vtime->gtime = 0;
731 }
732}
733
734void vtime_account_system(struct task_struct *tsk)
735{
736 struct vtime *vtime = &tsk->vtime;
737
738 if (!vtime_delta(vtime))
739 return;
740
741 write_seqcount_begin(&vtime->seqcount);
742 /* We might have scheduled out from guest path */
743 if (tsk->flags & PF_VCPU)
744 vtime_account_guest(tsk, vtime);
745 else
746 __vtime_account_system(tsk, vtime);
747 write_seqcount_end(&vtime->seqcount);
748}
749
750void vtime_user_enter(struct task_struct *tsk)
751{
752 struct vtime *vtime = &tsk->vtime;
753
754 write_seqcount_begin(&vtime->seqcount);
755 __vtime_account_system(tsk, vtime);
756 vtime->state = VTIME_USER;
757 write_seqcount_end(&vtime->seqcount);
758}
759
760void vtime_user_exit(struct task_struct *tsk)
761{
762 struct vtime *vtime = &tsk->vtime;
763
764 write_seqcount_begin(&vtime->seqcount);
765 vtime->utime += get_vtime_delta(vtime);
766 if (vtime->utime >= TICK_NSEC) {
767 account_user_time(tsk, vtime->utime);
768 vtime->utime = 0;
769 }
770 vtime->state = VTIME_SYS;
771 write_seqcount_end(&vtime->seqcount);
772}
773
774void vtime_guest_enter(struct task_struct *tsk)
775{
776 struct vtime *vtime = &tsk->vtime;
777 /*
778 * The flags must be updated under the lock with
779 * the vtime_starttime flush and update.
780 * That enforces a right ordering and update sequence
781 * synchronization against the reader (task_gtime())
782 * that can thus safely catch up with a tickless delta.
783 */
784 write_seqcount_begin(&vtime->seqcount);
785 __vtime_account_system(tsk, vtime);
786 tsk->flags |= PF_VCPU;
787 write_seqcount_end(&vtime->seqcount);
788}
789EXPORT_SYMBOL_GPL(vtime_guest_enter);
790
791void vtime_guest_exit(struct task_struct *tsk)
792{
793 struct vtime *vtime = &tsk->vtime;
794
795 write_seqcount_begin(&vtime->seqcount);
796 vtime_account_guest(tsk, vtime);
797 tsk->flags &= ~PF_VCPU;
798 write_seqcount_end(&vtime->seqcount);
799}
800EXPORT_SYMBOL_GPL(vtime_guest_exit);
801
802void vtime_account_idle(struct task_struct *tsk)
803{
804 account_idle_time(get_vtime_delta(&tsk->vtime));
805}
806
807void arch_vtime_task_switch(struct task_struct *prev)
808{
809 struct vtime *vtime = &prev->vtime;
810
811 write_seqcount_begin(&vtime->seqcount);
812 vtime->state = VTIME_INACTIVE;
813 write_seqcount_end(&vtime->seqcount);
814
815 vtime = ¤t->vtime;
816
817 write_seqcount_begin(&vtime->seqcount);
818 vtime->state = VTIME_SYS;
819 vtime->starttime = sched_clock();
820 write_seqcount_end(&vtime->seqcount);
821}
822
823void vtime_init_idle(struct task_struct *t, int cpu)
824{
825 struct vtime *vtime = &t->vtime;
826 unsigned long flags;
827
828 local_irq_save(flags);
829 write_seqcount_begin(&vtime->seqcount);
830 vtime->state = VTIME_SYS;
831 vtime->starttime = sched_clock();
832 write_seqcount_end(&vtime->seqcount);
833 local_irq_restore(flags);
834}
835
836u64 task_gtime(struct task_struct *t)
837{
838 struct vtime *vtime = &t->vtime;
839 unsigned int seq;
840 u64 gtime;
841
842 if (!vtime_accounting_enabled())
843 return t->gtime;
844
845 do {
846 seq = read_seqcount_begin(&vtime->seqcount);
847
848 gtime = t->gtime;
849 if (vtime->state == VTIME_SYS && t->flags & PF_VCPU)
850 gtime += vtime->gtime + vtime_delta(vtime);
851
852 } while (read_seqcount_retry(&vtime->seqcount, seq));
853
854 return gtime;
855}
856
857/*
858 * Fetch cputime raw values from fields of task_struct and
859 * add up the pending nohz execution time since the last
860 * cputime snapshot.
861 */
862void task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
863{
864 struct vtime *vtime = &t->vtime;
865 unsigned int seq;
866 u64 delta;
867
868 if (!vtime_accounting_enabled()) {
869 *utime = t->utime;
870 *stime = t->stime;
871 return;
872 }
873
874 do {
875 seq = read_seqcount_begin(&vtime->seqcount);
876
877 *utime = t->utime;
878 *stime = t->stime;
879
880 /* Task is sleeping, nothing to add */
881 if (vtime->state == VTIME_INACTIVE || is_idle_task(t))
882 continue;
883
884 delta = vtime_delta(vtime);
885
886 /*
887 * Task runs either in user or kernel space, add pending nohz time to
888 * the right place.
889 */
890 if (vtime->state == VTIME_USER || t->flags & PF_VCPU)
891 *utime += vtime->utime + delta;
892 else if (vtime->state == VTIME_SYS)
893 *stime += vtime->stime + delta;
894 } while (read_seqcount_retry(&vtime->seqcount, seq));
895}
896#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(u64, cpu_hardirq_time);
27DEFINE_PER_CPU(u64, cpu_softirq_time);
28
29static DEFINE_PER_CPU(u64, irq_start_time);
30static int sched_clock_irqtime;
31
32void enable_sched_clock_irqtime(void)
33{
34 sched_clock_irqtime = 1;
35}
36
37void disable_sched_clock_irqtime(void)
38{
39 sched_clock_irqtime = 0;
40}
41
42#ifndef CONFIG_64BIT
43DEFINE_PER_CPU(seqcount_t, irq_time_seq);
44#endif /* CONFIG_64BIT */
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 unsigned long flags;
53 s64 delta;
54 int cpu;
55
56 if (!sched_clock_irqtime)
57 return;
58
59 local_irq_save(flags);
60
61 cpu = smp_processor_id();
62 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
63 __this_cpu_add(irq_start_time, delta);
64
65 irq_time_write_begin();
66 /*
67 * We do not account for softirq time from ksoftirqd here.
68 * We want to continue accounting softirq time to ksoftirqd thread
69 * in that case, so as not to confuse scheduler with a special task
70 * that do not consume any time, but still wants to run.
71 */
72 if (hardirq_count())
73 __this_cpu_add(cpu_hardirq_time, delta);
74 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
75 __this_cpu_add(cpu_softirq_time, delta);
76
77 irq_time_write_end();
78 local_irq_restore(flags);
79}
80EXPORT_SYMBOL_GPL(irqtime_account_irq);
81
82static int irqtime_account_hi_update(void)
83{
84 u64 *cpustat = kcpustat_this_cpu->cpustat;
85 unsigned long flags;
86 u64 latest_ns;
87 int ret = 0;
88
89 local_irq_save(flags);
90 latest_ns = this_cpu_read(cpu_hardirq_time);
91 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
92 ret = 1;
93 local_irq_restore(flags);
94 return ret;
95}
96
97static int irqtime_account_si_update(void)
98{
99 u64 *cpustat = kcpustat_this_cpu->cpustat;
100 unsigned long flags;
101 u64 latest_ns;
102 int ret = 0;
103
104 local_irq_save(flags);
105 latest_ns = this_cpu_read(cpu_softirq_time);
106 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
107 ret = 1;
108 local_irq_restore(flags);
109 return ret;
110}
111
112#else /* CONFIG_IRQ_TIME_ACCOUNTING */
113
114#define sched_clock_irqtime (0)
115
116#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
117
118static inline void task_group_account_field(struct task_struct *p, int index,
119 u64 tmp)
120{
121 /*
122 * Since all updates are sure to touch the root cgroup, we
123 * get ourselves ahead and touch it first. If the root cgroup
124 * is the only cgroup, then nothing else should be necessary.
125 *
126 */
127 __this_cpu_add(kernel_cpustat.cpustat[index], tmp);
128
129 cpuacct_account_field(p, index, tmp);
130}
131
132/*
133 * Account user cpu time to a process.
134 * @p: the process that the cpu time gets accounted to
135 * @cputime: the cpu time spent in user space since the last update
136 * @cputime_scaled: cputime scaled by cpu frequency
137 */
138void account_user_time(struct task_struct *p, cputime_t cputime,
139 cputime_t cputime_scaled)
140{
141 int index;
142
143 /* Add user time to process. */
144 p->utime += cputime;
145 p->utimescaled += cputime_scaled;
146 account_group_user_time(p, cputime);
147
148 index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
149
150 /* Add user time to cpustat. */
151 task_group_account_field(p, index, (__force u64) cputime);
152
153 /* Account for user time used */
154 acct_account_cputime(p);
155}
156
157/*
158 * Account guest cpu time to a process.
159 * @p: the process that the cpu time gets accounted to
160 * @cputime: the cpu time spent in virtual machine since the last update
161 * @cputime_scaled: cputime scaled by cpu frequency
162 */
163static void account_guest_time(struct task_struct *p, cputime_t cputime,
164 cputime_t cputime_scaled)
165{
166 u64 *cpustat = kcpustat_this_cpu->cpustat;
167
168 /* Add guest time to process. */
169 p->utime += cputime;
170 p->utimescaled += cputime_scaled;
171 account_group_user_time(p, cputime);
172 p->gtime += cputime;
173
174 /* Add guest time to cpustat. */
175 if (task_nice(p) > 0) {
176 cpustat[CPUTIME_NICE] += (__force u64) cputime;
177 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
178 } else {
179 cpustat[CPUTIME_USER] += (__force u64) cputime;
180 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
181 }
182}
183
184/*
185 * Account system cpu time to a process and desired cpustat field
186 * @p: the process that the cpu time gets accounted to
187 * @cputime: the cpu time spent in kernel space since the last update
188 * @cputime_scaled: cputime scaled by cpu frequency
189 * @target_cputime64: pointer to cpustat field that has to be updated
190 */
191static inline
192void __account_system_time(struct task_struct *p, cputime_t cputime,
193 cputime_t cputime_scaled, int index)
194{
195 /* Add system time to process. */
196 p->stime += cputime;
197 p->stimescaled += cputime_scaled;
198 account_group_system_time(p, cputime);
199
200 /* Add system time to cpustat. */
201 task_group_account_field(p, index, (__force u64) cputime);
202
203 /* Account for system time used */
204 acct_account_cputime(p);
205}
206
207/*
208 * Account system cpu time to a process.
209 * @p: the process that the cpu time gets accounted to
210 * @hardirq_offset: the offset to subtract from hardirq_count()
211 * @cputime: the cpu time spent in kernel space since the last update
212 * @cputime_scaled: cputime scaled by cpu frequency
213 */
214void account_system_time(struct task_struct *p, int hardirq_offset,
215 cputime_t cputime, cputime_t cputime_scaled)
216{
217 int index;
218
219 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
220 account_guest_time(p, cputime, cputime_scaled);
221 return;
222 }
223
224 if (hardirq_count() - hardirq_offset)
225 index = CPUTIME_IRQ;
226 else if (in_serving_softirq())
227 index = CPUTIME_SOFTIRQ;
228 else
229 index = CPUTIME_SYSTEM;
230
231 __account_system_time(p, cputime, cputime_scaled, index);
232}
233
234/*
235 * Account for involuntary wait time.
236 * @cputime: the cpu time spent in involuntary wait
237 */
238void account_steal_time(cputime_t cputime)
239{
240 u64 *cpustat = kcpustat_this_cpu->cpustat;
241
242 cpustat[CPUTIME_STEAL] += (__force u64) cputime;
243}
244
245/*
246 * Account for idle time.
247 * @cputime: the cpu time spent in idle wait
248 */
249void account_idle_time(cputime_t cputime)
250{
251 u64 *cpustat = kcpustat_this_cpu->cpustat;
252 struct rq *rq = this_rq();
253
254 if (atomic_read(&rq->nr_iowait) > 0)
255 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
256 else
257 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
258}
259
260static __always_inline bool steal_account_process_tick(void)
261{
262#ifdef CONFIG_PARAVIRT
263 if (static_key_false(¶virt_steal_enabled)) {
264 u64 steal;
265 unsigned long steal_jiffies;
266
267 steal = paravirt_steal_clock(smp_processor_id());
268 steal -= this_rq()->prev_steal_time;
269
270 /*
271 * steal is in nsecs but our caller is expecting steal
272 * time in jiffies. Lets cast the result to jiffies
273 * granularity and account the rest on the next rounds.
274 */
275 steal_jiffies = nsecs_to_jiffies(steal);
276 this_rq()->prev_steal_time += jiffies_to_nsecs(steal_jiffies);
277
278 account_steal_time(jiffies_to_cputime(steal_jiffies));
279 return steal_jiffies;
280 }
281#endif
282 return false;
283}
284
285/*
286 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
287 * tasks (sum on group iteration) belonging to @tsk's group.
288 */
289void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
290{
291 struct signal_struct *sig = tsk->signal;
292 cputime_t utime, stime;
293 struct task_struct *t;
294 unsigned int seq, nextseq;
295 unsigned long flags;
296
297 rcu_read_lock();
298 /* Attempt a lockless read on the first round. */
299 nextseq = 0;
300 do {
301 seq = nextseq;
302 flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
303 times->utime = sig->utime;
304 times->stime = sig->stime;
305 times->sum_exec_runtime = sig->sum_sched_runtime;
306
307 for_each_thread(tsk, t) {
308 task_cputime(t, &utime, &stime);
309 times->utime += utime;
310 times->stime += stime;
311 times->sum_exec_runtime += task_sched_runtime(t);
312 }
313 /* If lockless access failed, take the lock. */
314 nextseq = 1;
315 } while (need_seqretry(&sig->stats_lock, seq));
316 done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
317 rcu_read_unlock();
318}
319
320#ifdef CONFIG_IRQ_TIME_ACCOUNTING
321/*
322 * Account a tick to a process and cpustat
323 * @p: the process that the cpu time gets accounted to
324 * @user_tick: is the tick from userspace
325 * @rq: the pointer to rq
326 *
327 * Tick demultiplexing follows the order
328 * - pending hardirq update
329 * - pending softirq update
330 * - user_time
331 * - idle_time
332 * - system time
333 * - check for guest_time
334 * - else account as system_time
335 *
336 * Check for hardirq is done both for system and user time as there is
337 * no timer going off while we are on hardirq and hence we may never get an
338 * opportunity to update it solely in system time.
339 * p->stime and friends are only updated on system time and not on irq
340 * softirq as those do not count in task exec_runtime any more.
341 */
342static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
343 struct rq *rq, int ticks)
344{
345 cputime_t scaled = cputime_to_scaled(cputime_one_jiffy);
346 u64 cputime = (__force u64) cputime_one_jiffy;
347 u64 *cpustat = kcpustat_this_cpu->cpustat;
348
349 if (steal_account_process_tick())
350 return;
351
352 cputime *= ticks;
353 scaled *= ticks;
354
355 if (irqtime_account_hi_update()) {
356 cpustat[CPUTIME_IRQ] += cputime;
357 } else if (irqtime_account_si_update()) {
358 cpustat[CPUTIME_SOFTIRQ] += cputime;
359 } else if (this_cpu_ksoftirqd() == p) {
360 /*
361 * ksoftirqd time do not get accounted in cpu_softirq_time.
362 * So, we have to handle it separately here.
363 * Also, p->stime needs to be updated for ksoftirqd.
364 */
365 __account_system_time(p, cputime, scaled, CPUTIME_SOFTIRQ);
366 } else if (user_tick) {
367 account_user_time(p, cputime, scaled);
368 } else if (p == rq->idle) {
369 account_idle_time(cputime);
370 } else if (p->flags & PF_VCPU) { /* System time or guest time */
371 account_guest_time(p, cputime, scaled);
372 } else {
373 __account_system_time(p, cputime, scaled, CPUTIME_SYSTEM);
374 }
375}
376
377static void irqtime_account_idle_ticks(int ticks)
378{
379 struct rq *rq = this_rq();
380
381 irqtime_account_process_tick(current, 0, rq, ticks);
382}
383#else /* CONFIG_IRQ_TIME_ACCOUNTING */
384static inline void irqtime_account_idle_ticks(int ticks) {}
385static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
386 struct rq *rq, int nr_ticks) {}
387#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
388
389/*
390 * Use precise platform statistics if available:
391 */
392#ifdef CONFIG_VIRT_CPU_ACCOUNTING
393
394#ifndef __ARCH_HAS_VTIME_TASK_SWITCH
395void vtime_common_task_switch(struct task_struct *prev)
396{
397 if (is_idle_task(prev))
398 vtime_account_idle(prev);
399 else
400 vtime_account_system(prev);
401
402#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
403 vtime_account_user(prev);
404#endif
405 arch_vtime_task_switch(prev);
406}
407#endif
408
409/*
410 * Archs that account the whole time spent in the idle task
411 * (outside irq) as idle time can rely on this and just implement
412 * vtime_account_system() and vtime_account_idle(). Archs that
413 * have other meaning of the idle time (s390 only includes the
414 * time spent by the CPU when it's in low power mode) must override
415 * vtime_account().
416 */
417#ifndef __ARCH_HAS_VTIME_ACCOUNT
418void vtime_common_account_irq_enter(struct task_struct *tsk)
419{
420 if (!in_interrupt()) {
421 /*
422 * If we interrupted user, context_tracking_in_user()
423 * is 1 because the context tracking don't hook
424 * on irq entry/exit. This way we know if
425 * we need to flush user time on kernel entry.
426 */
427 if (context_tracking_in_user()) {
428 vtime_account_user(tsk);
429 return;
430 }
431
432 if (is_idle_task(tsk)) {
433 vtime_account_idle(tsk);
434 return;
435 }
436 }
437 vtime_account_system(tsk);
438}
439EXPORT_SYMBOL_GPL(vtime_common_account_irq_enter);
440#endif /* __ARCH_HAS_VTIME_ACCOUNT */
441#endif /* CONFIG_VIRT_CPU_ACCOUNTING */
442
443
444#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
445void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
446{
447 *ut = p->utime;
448 *st = p->stime;
449}
450EXPORT_SYMBOL_GPL(task_cputime_adjusted);
451
452void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
453{
454 struct task_cputime cputime;
455
456 thread_group_cputime(p, &cputime);
457
458 *ut = cputime.utime;
459 *st = cputime.stime;
460}
461#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
462/*
463 * Account a single tick of cpu time.
464 * @p: the process that the cpu time gets accounted to
465 * @user_tick: indicates if the tick is a user or a system tick
466 */
467void account_process_tick(struct task_struct *p, int user_tick)
468{
469 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
470 struct rq *rq = this_rq();
471
472 if (vtime_accounting_cpu_enabled())
473 return;
474
475 if (sched_clock_irqtime) {
476 irqtime_account_process_tick(p, user_tick, rq, 1);
477 return;
478 }
479
480 if (steal_account_process_tick())
481 return;
482
483 if (user_tick)
484 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
485 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
486 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
487 one_jiffy_scaled);
488 else
489 account_idle_time(cputime_one_jiffy);
490}
491
492/*
493 * Account multiple ticks of steal time.
494 * @p: the process from which the cpu time has been stolen
495 * @ticks: number of stolen ticks
496 */
497void account_steal_ticks(unsigned long ticks)
498{
499 account_steal_time(jiffies_to_cputime(ticks));
500}
501
502/*
503 * Account multiple ticks of idle time.
504 * @ticks: number of stolen ticks
505 */
506void account_idle_ticks(unsigned long ticks)
507{
508
509 if (sched_clock_irqtime) {
510 irqtime_account_idle_ticks(ticks);
511 return;
512 }
513
514 account_idle_time(jiffies_to_cputime(ticks));
515}
516
517/*
518 * Perform (stime * rtime) / total, but avoid multiplication overflow by
519 * loosing precision when the numbers are big.
520 */
521static cputime_t scale_stime(u64 stime, u64 rtime, u64 total)
522{
523 u64 scaled;
524
525 for (;;) {
526 /* Make sure "rtime" is the bigger of stime/rtime */
527 if (stime > rtime)
528 swap(rtime, stime);
529
530 /* Make sure 'total' fits in 32 bits */
531 if (total >> 32)
532 goto drop_precision;
533
534 /* Does rtime (and thus stime) fit in 32 bits? */
535 if (!(rtime >> 32))
536 break;
537
538 /* Can we just balance rtime/stime rather than dropping bits? */
539 if (stime >> 31)
540 goto drop_precision;
541
542 /* We can grow stime and shrink rtime and try to make them both fit */
543 stime <<= 1;
544 rtime >>= 1;
545 continue;
546
547drop_precision:
548 /* We drop from rtime, it has more bits than stime */
549 rtime >>= 1;
550 total >>= 1;
551 }
552
553 /*
554 * Make sure gcc understands that this is a 32x32->64 multiply,
555 * followed by a 64/32->64 divide.
556 */
557 scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
558 return (__force cputime_t) scaled;
559}
560
561/*
562 * Adjust tick based cputime random precision against scheduler runtime
563 * accounting.
564 *
565 * Tick based cputime accounting depend on random scheduling timeslices of a
566 * task to be interrupted or not by the timer. Depending on these
567 * circumstances, the number of these interrupts may be over or
568 * under-optimistic, matching the real user and system cputime with a variable
569 * precision.
570 *
571 * Fix this by scaling these tick based values against the total runtime
572 * accounted by the CFS scheduler.
573 *
574 * This code provides the following guarantees:
575 *
576 * stime + utime == rtime
577 * stime_i+1 >= stime_i, utime_i+1 >= utime_i
578 *
579 * Assuming that rtime_i+1 >= rtime_i.
580 */
581static void cputime_adjust(struct task_cputime *curr,
582 struct prev_cputime *prev,
583 cputime_t *ut, cputime_t *st)
584{
585 cputime_t rtime, stime, utime;
586 unsigned long flags;
587
588 /* Serialize concurrent callers such that we can honour our guarantees */
589 raw_spin_lock_irqsave(&prev->lock, flags);
590 rtime = nsecs_to_cputime(curr->sum_exec_runtime);
591
592 /*
593 * This is possible under two circumstances:
594 * - rtime isn't monotonic after all (a bug);
595 * - we got reordered by the lock.
596 *
597 * In both cases this acts as a filter such that the rest of the code
598 * can assume it is monotonic regardless of anything else.
599 */
600 if (prev->stime + prev->utime >= rtime)
601 goto out;
602
603 stime = curr->stime;
604 utime = curr->utime;
605
606 if (utime == 0) {
607 stime = rtime;
608 goto update;
609 }
610
611 if (stime == 0) {
612 utime = rtime;
613 goto update;
614 }
615
616 stime = scale_stime((__force u64)stime, (__force u64)rtime,
617 (__force u64)(stime + utime));
618
619 /*
620 * Make sure stime doesn't go backwards; this preserves monotonicity
621 * for utime because rtime is monotonic.
622 *
623 * utime_i+1 = rtime_i+1 - stime_i
624 * = rtime_i+1 - (rtime_i - utime_i)
625 * = (rtime_i+1 - rtime_i) + utime_i
626 * >= utime_i
627 */
628 if (stime < prev->stime)
629 stime = prev->stime;
630 utime = rtime - stime;
631
632 /*
633 * Make sure utime doesn't go backwards; this still preserves
634 * monotonicity for stime, analogous argument to above.
635 */
636 if (utime < prev->utime) {
637 utime = prev->utime;
638 stime = rtime - utime;
639 }
640
641update:
642 prev->stime = stime;
643 prev->utime = utime;
644out:
645 *ut = prev->utime;
646 *st = prev->stime;
647 raw_spin_unlock_irqrestore(&prev->lock, flags);
648}
649
650void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
651{
652 struct task_cputime cputime = {
653 .sum_exec_runtime = p->se.sum_exec_runtime,
654 };
655
656 task_cputime(p, &cputime.utime, &cputime.stime);
657 cputime_adjust(&cputime, &p->prev_cputime, ut, st);
658}
659EXPORT_SYMBOL_GPL(task_cputime_adjusted);
660
661void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
662{
663 struct task_cputime cputime;
664
665 thread_group_cputime(p, &cputime);
666 cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
667}
668#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
669
670#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
671static cputime_t vtime_delta(struct task_struct *tsk)
672{
673 unsigned long now = READ_ONCE(jiffies);
674
675 if (time_before(now, (unsigned long)tsk->vtime_snap))
676 return 0;
677
678 return jiffies_to_cputime(now - tsk->vtime_snap);
679}
680
681static cputime_t get_vtime_delta(struct task_struct *tsk)
682{
683 unsigned long now = READ_ONCE(jiffies);
684 unsigned long delta = now - tsk->vtime_snap;
685
686 WARN_ON_ONCE(tsk->vtime_snap_whence == VTIME_INACTIVE);
687 tsk->vtime_snap = now;
688
689 return jiffies_to_cputime(delta);
690}
691
692static void __vtime_account_system(struct task_struct *tsk)
693{
694 cputime_t delta_cpu = get_vtime_delta(tsk);
695
696 account_system_time(tsk, irq_count(), delta_cpu, cputime_to_scaled(delta_cpu));
697}
698
699void vtime_account_system(struct task_struct *tsk)
700{
701 if (!vtime_delta(tsk))
702 return;
703
704 write_seqcount_begin(&tsk->vtime_seqcount);
705 __vtime_account_system(tsk);
706 write_seqcount_end(&tsk->vtime_seqcount);
707}
708
709void vtime_gen_account_irq_exit(struct task_struct *tsk)
710{
711 write_seqcount_begin(&tsk->vtime_seqcount);
712 if (vtime_delta(tsk))
713 __vtime_account_system(tsk);
714 if (context_tracking_in_user())
715 tsk->vtime_snap_whence = VTIME_USER;
716 write_seqcount_end(&tsk->vtime_seqcount);
717}
718
719void vtime_account_user(struct task_struct *tsk)
720{
721 cputime_t delta_cpu;
722
723 write_seqcount_begin(&tsk->vtime_seqcount);
724 tsk->vtime_snap_whence = VTIME_SYS;
725 if (vtime_delta(tsk)) {
726 delta_cpu = get_vtime_delta(tsk);
727 account_user_time(tsk, delta_cpu, cputime_to_scaled(delta_cpu));
728 }
729 write_seqcount_end(&tsk->vtime_seqcount);
730}
731
732void vtime_user_enter(struct task_struct *tsk)
733{
734 write_seqcount_begin(&tsk->vtime_seqcount);
735 if (vtime_delta(tsk))
736 __vtime_account_system(tsk);
737 tsk->vtime_snap_whence = VTIME_USER;
738 write_seqcount_end(&tsk->vtime_seqcount);
739}
740
741void vtime_guest_enter(struct task_struct *tsk)
742{
743 /*
744 * The flags must be updated under the lock with
745 * the vtime_snap flush and update.
746 * That enforces a right ordering and update sequence
747 * synchronization against the reader (task_gtime())
748 * that can thus safely catch up with a tickless delta.
749 */
750 write_seqcount_begin(&tsk->vtime_seqcount);
751 if (vtime_delta(tsk))
752 __vtime_account_system(tsk);
753 current->flags |= PF_VCPU;
754 write_seqcount_end(&tsk->vtime_seqcount);
755}
756EXPORT_SYMBOL_GPL(vtime_guest_enter);
757
758void vtime_guest_exit(struct task_struct *tsk)
759{
760 write_seqcount_begin(&tsk->vtime_seqcount);
761 __vtime_account_system(tsk);
762 current->flags &= ~PF_VCPU;
763 write_seqcount_end(&tsk->vtime_seqcount);
764}
765EXPORT_SYMBOL_GPL(vtime_guest_exit);
766
767void vtime_account_idle(struct task_struct *tsk)
768{
769 cputime_t delta_cpu = get_vtime_delta(tsk);
770
771 account_idle_time(delta_cpu);
772}
773
774void arch_vtime_task_switch(struct task_struct *prev)
775{
776 write_seqcount_begin(&prev->vtime_seqcount);
777 prev->vtime_snap_whence = VTIME_INACTIVE;
778 write_seqcount_end(&prev->vtime_seqcount);
779
780 write_seqcount_begin(¤t->vtime_seqcount);
781 current->vtime_snap_whence = VTIME_SYS;
782 current->vtime_snap = jiffies;
783 write_seqcount_end(¤t->vtime_seqcount);
784}
785
786void vtime_init_idle(struct task_struct *t, int cpu)
787{
788 unsigned long flags;
789
790 local_irq_save(flags);
791 write_seqcount_begin(&t->vtime_seqcount);
792 t->vtime_snap_whence = VTIME_SYS;
793 t->vtime_snap = jiffies;
794 write_seqcount_end(&t->vtime_seqcount);
795 local_irq_restore(flags);
796}
797
798cputime_t task_gtime(struct task_struct *t)
799{
800 unsigned int seq;
801 cputime_t gtime;
802
803 if (!vtime_accounting_enabled())
804 return t->gtime;
805
806 do {
807 seq = read_seqcount_begin(&t->vtime_seqcount);
808
809 gtime = t->gtime;
810 if (t->vtime_snap_whence == VTIME_SYS && t->flags & PF_VCPU)
811 gtime += vtime_delta(t);
812
813 } while (read_seqcount_retry(&t->vtime_seqcount, seq));
814
815 return gtime;
816}
817
818/*
819 * Fetch cputime raw values from fields of task_struct and
820 * add up the pending nohz execution time since the last
821 * cputime snapshot.
822 */
823static void
824fetch_task_cputime(struct task_struct *t,
825 cputime_t *u_dst, cputime_t *s_dst,
826 cputime_t *u_src, cputime_t *s_src,
827 cputime_t *udelta, cputime_t *sdelta)
828{
829 unsigned int seq;
830 unsigned long long delta;
831
832 do {
833 *udelta = 0;
834 *sdelta = 0;
835
836 seq = read_seqcount_begin(&t->vtime_seqcount);
837
838 if (u_dst)
839 *u_dst = *u_src;
840 if (s_dst)
841 *s_dst = *s_src;
842
843 /* Task is sleeping, nothing to add */
844 if (t->vtime_snap_whence == VTIME_INACTIVE ||
845 is_idle_task(t))
846 continue;
847
848 delta = vtime_delta(t);
849
850 /*
851 * Task runs either in user or kernel space, add pending nohz time to
852 * the right place.
853 */
854 if (t->vtime_snap_whence == VTIME_USER || t->flags & PF_VCPU) {
855 *udelta = delta;
856 } else {
857 if (t->vtime_snap_whence == VTIME_SYS)
858 *sdelta = delta;
859 }
860 } while (read_seqcount_retry(&t->vtime_seqcount, seq));
861}
862
863
864void task_cputime(struct task_struct *t, cputime_t *utime, cputime_t *stime)
865{
866 cputime_t udelta, sdelta;
867
868 if (!vtime_accounting_enabled()) {
869 if (utime)
870 *utime = t->utime;
871 if (stime)
872 *stime = t->stime;
873 return;
874 }
875
876 fetch_task_cputime(t, utime, stime, &t->utime,
877 &t->stime, &udelta, &sdelta);
878 if (utime)
879 *utime += udelta;
880 if (stime)
881 *stime += sdelta;
882}
883
884void task_cputime_scaled(struct task_struct *t,
885 cputime_t *utimescaled, cputime_t *stimescaled)
886{
887 cputime_t udelta, sdelta;
888
889 if (!vtime_accounting_enabled()) {
890 if (utimescaled)
891 *utimescaled = t->utimescaled;
892 if (stimescaled)
893 *stimescaled = t->stimescaled;
894 return;
895 }
896
897 fetch_task_cputime(t, utimescaled, stimescaled,
898 &t->utimescaled, &t->stimescaled, &udelta, &sdelta);
899 if (utimescaled)
900 *utimescaled += cputime_to_scaled(udelta);
901 if (stimescaled)
902 *stimescaled += cputime_to_scaled(sdelta);
903}
904#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */