1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Simple CPU accounting cgroup controller
4 */
5#include <linux/sched/cputime.h>
6#include <linux/tsacct_kern.h>
7#include "sched.h"
8
9#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
10 #include <asm/cputime.h>
11#endif
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
28int 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
40static void irqtime_account_delta(struct irqtime *irqtime, u64 delta,
41 enum cpu_usage_stat idx)
42{
43 u64 *cpustat = kcpustat_this_cpu->cpustat;
44
45 u64_stats_update_begin(&irqtime->sync);
46 cpustat[idx] += delta;
47 irqtime->total += delta;
48 irqtime->tick_delta += delta;
49 u64_stats_update_end(&irqtime->sync);
50}
51
52/*
53 * Called after incrementing preempt_count on {soft,}irq_enter
54 * and before decrementing preempt_count on {soft,}irq_exit.
55 */
56void irqtime_account_irq(struct task_struct *curr, unsigned int offset)
57{
58 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
59 unsigned int pc;
60 s64 delta;
61 int cpu;
62
63 if (!irqtime_enabled())
64 return;
65
66 cpu = smp_processor_id();
67 delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
68 irqtime->irq_start_time += delta;
69 pc = irq_count() - offset;
70
71 /*
72 * We do not account for softirq time from ksoftirqd here.
73 * We want to continue accounting softirq time to ksoftirqd thread
74 * in that case, so as not to confuse scheduler with a special task
75 * that do not consume any time, but still wants to run.
76 */
77 if (pc & HARDIRQ_MASK)
78 irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
79 else if ((pc & SOFTIRQ_OFFSET) && curr != this_cpu_ksoftirqd())
80 irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
81}
82
83static u64 irqtime_tick_accounted(u64 maxtime)
84{
85 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
86 u64 delta;
87
88 delta = min(irqtime->tick_delta, maxtime);
89 irqtime->tick_delta -= delta;
90
91 return delta;
92}
93
94#else /* !CONFIG_IRQ_TIME_ACCOUNTING: */
95
96static u64 irqtime_tick_accounted(u64 dummy)
97{
98 return 0;
99}
100
101#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
102
103static inline void task_group_account_field(struct task_struct *p, int index,
104 u64 tmp)
105{
106 /*
107 * Since all updates are sure to touch the root cgroup, we
108 * get ourselves ahead and touch it first. If the root cgroup
109 * is the only cgroup, then nothing else should be necessary.
110 *
111 */
112 __this_cpu_add(kernel_cpustat.cpustat[index], tmp);
113
114 cgroup_account_cputime_field(task: p, index, delta_exec: tmp);
115}
116
117/*
118 * Account user CPU time to a process.
119 * @p: the process that the CPU time gets accounted to
120 * @cputime: the CPU time spent in user space since the last update
121 */
122void account_user_time(struct task_struct *p, u64 cputime)
123{
124 int index;
125
126 /* Add user time to process. */
127 p->utime += cputime;
128 account_group_user_time(tsk: p, cputime);
129
130 index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
131
132 /* Add user time to cpustat. */
133 task_group_account_field(p, index, tmp: cputime);
134
135 /* Account for user time used */
136 acct_account_cputime(tsk: p);
137}
138
139/*
140 * Account guest CPU time to a process.
141 * @p: the process that the CPU time gets accounted to
142 * @cputime: the CPU time spent in virtual machine since the last update
143 */
144void account_guest_time(struct task_struct *p, u64 cputime)
145{
146 u64 *cpustat = kcpustat_this_cpu->cpustat;
147
148 /* Add guest time to process. */
149 p->utime += cputime;
150 account_group_user_time(tsk: p, cputime);
151 p->gtime += cputime;
152
153 /* Add guest time to cpustat. */
154 if (task_nice(p) > 0) {
155 task_group_account_field(p, index: CPUTIME_NICE, tmp: cputime);
156 cpustat[CPUTIME_GUEST_NICE] += cputime;
157 } else {
158 task_group_account_field(p, index: CPUTIME_USER, tmp: cputime);
159 cpustat[CPUTIME_GUEST] += cputime;
160 }
161}
162
163/*
164 * Account system CPU time to a process and desired cpustat field
165 * @p: the process that the CPU time gets accounted to
166 * @cputime: the CPU time spent in kernel space since the last update
167 * @index: pointer to cpustat field that has to be updated
168 */
169void account_system_index_time(struct task_struct *p,
170 u64 cputime, enum cpu_usage_stat index)
171{
172 /* Add system time to process. */
173 p->stime += cputime;
174 account_group_system_time(tsk: p, cputime);
175
176 /* Add system time to cpustat. */
177 task_group_account_field(p, index, tmp: cputime);
178
179 /* Account for system time used */
180 acct_account_cputime(tsk: p);
181}
182
183/*
184 * Account system CPU time to a process.
185 * @p: the process that the CPU time gets accounted to
186 * @hardirq_offset: the offset to subtract from hardirq_count()
187 * @cputime: the CPU time spent in kernel space since the last update
188 */
189void account_system_time(struct task_struct *p, int hardirq_offset, u64 cputime)
190{
191 int index;
192
193 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
194 account_guest_time(p, cputime);
195 return;
196 }
197
198 if (hardirq_count() - hardirq_offset)
199 index = CPUTIME_IRQ;
200 else if (in_serving_softirq())
201 index = CPUTIME_SOFTIRQ;
202 else
203 index = CPUTIME_SYSTEM;
204
205 account_system_index_time(p, cputime, index);
206}
207
208/*
209 * Account for involuntary wait time.
210 * @cputime: the CPU time spent in involuntary wait
211 */
212void account_steal_time(u64 cputime)
213{
214 u64 *cpustat = kcpustat_this_cpu->cpustat;
215
216 cpustat[CPUTIME_STEAL] += cputime;
217}
218
219/*
220 * Account for idle time.
221 * @cputime: the CPU time spent in idle wait
222 */
223void account_idle_time(u64 cputime)
224{
225 u64 *cpustat = kcpustat_this_cpu->cpustat;
226 struct rq *rq = this_rq();
227
228 if (atomic_read(v: &rq->nr_iowait) > 0)
229 cpustat[CPUTIME_IOWAIT] += cputime;
230 else
231 cpustat[CPUTIME_IDLE] += cputime;
232}
233
234
235#ifdef CONFIG_SCHED_CORE
236/*
237 * Account for forceidle time due to core scheduling.
238 *
239 * REQUIRES: schedstat is enabled.
240 */
241void __account_forceidle_time(struct task_struct *p, u64 delta)
242{
243 __schedstat_add(p->stats.core_forceidle_sum, delta);
244
245 task_group_account_field(p, CPUTIME_FORCEIDLE, delta);
246}
247#endif /* CONFIG_SCHED_CORE */
248
249/*
250 * When a guest is interrupted for a longer amount of time, missed clock
251 * ticks are not redelivered later. Due to that, this function may on
252 * occasion account more time than the calling functions think elapsed.
253 */
254static __always_inline u64 steal_account_process_time(u64 maxtime)
255{
256#ifdef CONFIG_PARAVIRT
257 if (static_key_false(key: &paravirt_steal_enabled)) {
258 u64 steal;
259
260 steal = paravirt_steal_clock(smp_processor_id());
261 steal -= this_rq()->prev_steal_time;
262 steal = min(steal, maxtime);
263 account_steal_time(cputime: steal);
264 this_rq()->prev_steal_time += steal;
265
266 return steal;
267 }
268#endif /* CONFIG_PARAVIRT */
269 return 0;
270}
271
272/*
273 * Account how much elapsed time was spent in steal, IRQ, or softirq time.
274 */
275static inline u64 account_other_time(u64 max)
276{
277 u64 accounted;
278
279 lockdep_assert_irqs_disabled();
280
281 accounted = steal_account_process_time(maxtime: max);
282
283 if (accounted < max)
284 accounted += irqtime_tick_accounted(dummy: max - accounted);
285
286 return accounted;
287}
288
289#ifdef CONFIG_64BIT
290static inline u64 read_sum_exec_runtime(struct task_struct *t)
291{
292 return t->se.sum_exec_runtime;
293}
294#else /* !CONFIG_64BIT: */
295static u64 read_sum_exec_runtime(struct task_struct *t)
296{
297 u64 ns;
298 struct rq_flags rf;
299 struct rq *rq;
300
301 rq = task_rq_lock(t, &rf);
302 ns = t->se.sum_exec_runtime;
303 task_rq_unlock(rq, t, &rf);
304
305 return ns;
306}
307#endif /* !CONFIG_64BIT */
308
309/*
310 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
311 * tasks (sum on group iteration) belonging to @tsk's group.
312 */
313void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
314{
315 struct signal_struct *sig = tsk->signal;
316 u64 utime, stime;
317 struct task_struct *t;
318 unsigned int seq, nextseq;
319 unsigned long flags;
320
321 /*
322 * Update current task runtime to account pending time since last
323 * scheduler action or thread_group_cputime() call. This thread group
324 * might have other running tasks on different CPUs, but updating
325 * their runtime can affect syscall performance, so we skip account
326 * those pending times and rely only on values updated on tick or
327 * other scheduler action.
328 */
329 if (same_thread_group(current, p2: tsk))
330 (void) task_sched_runtime(current);
331
332 rcu_read_lock();
333 /* Attempt a lockless read on the first round. */
334 nextseq = 0;
335 do {
336 seq = nextseq;
337 flags = read_seqbegin_or_lock_irqsave(lock: &sig->stats_lock, seq: &seq);
338 times->utime = sig->utime;
339 times->stime = sig->stime;
340 times->sum_exec_runtime = sig->sum_sched_runtime;
341
342 for_each_thread(tsk, t) {
343 task_cputime(t, utime: &utime, stime: &stime);
344 times->utime += utime;
345 times->stime += stime;
346 times->sum_exec_runtime += read_sum_exec_runtime(t);
347 }
348 /* If lockless access failed, take the lock. */
349 nextseq = 1;
350 } while (need_seqretry(lock: &sig->stats_lock, seq));
351 done_seqretry_irqrestore(lock: &sig->stats_lock, seq, flags);
352 rcu_read_unlock();
353}
354
355#ifdef CONFIG_IRQ_TIME_ACCOUNTING
356/*
357 * Account a tick to a process and cpustat
358 * @p: the process that the CPU time gets accounted to
359 * @user_tick: is the tick from userspace
360 * @rq: the pointer to rq
361 *
362 * Tick demultiplexing follows the order
363 * - pending hardirq update
364 * - pending softirq update
365 * - user_time
366 * - idle_time
367 * - system time
368 * - check for guest_time
369 * - else account as system_time
370 *
371 * Check for hardirq is done both for system and user time as there is
372 * no timer going off while we are on hardirq and hence we may never get an
373 * opportunity to update it solely in system time.
374 * p->stime and friends are only updated on system time and not on IRQ
375 * softirq as those do not count in task exec_runtime any more.
376 */
377static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
378 int ticks)
379{
380 u64 other, cputime = TICK_NSEC * ticks;
381
382 /*
383 * When returning from idle, many ticks can get accounted at
384 * once, including some ticks of steal, IRQ, and softirq time.
385 * Subtract those ticks from the amount of time accounted to
386 * idle, or potentially user or system time. Due to rounding,
387 * other time can exceed ticks occasionally.
388 */
389 other = account_other_time(ULONG_MAX);
390 if (other >= cputime)
391 return;
392
393 cputime -= other;
394
395 if (this_cpu_ksoftirqd() == p) {
396 /*
397 * ksoftirqd time do not get accounted in cpu_softirq_time.
398 * So, we have to handle it separately here.
399 * Also, p->stime needs to be updated for ksoftirqd.
400 */
401 account_system_index_time(p, cputime, CPUTIME_SOFTIRQ);
402 } else if (user_tick) {
403 account_user_time(p, cputime);
404 } else if (p == this_rq()->idle) {
405 account_idle_time(cputime);
406 } else if (p->flags & PF_VCPU) { /* System time or guest time */
407 account_guest_time(p, cputime);
408 } else {
409 account_system_index_time(p, cputime, CPUTIME_SYSTEM);
410 }
411}
412
413static void irqtime_account_idle_ticks(int ticks)
414{
415 irqtime_account_process_tick(current, 0, ticks);
416}
417#else /* !CONFIG_IRQ_TIME_ACCOUNTING: */
418static inline void irqtime_account_idle_ticks(int ticks) { }
419static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
420 int nr_ticks) { }
421#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
422
423/*
424 * Use precise platform statistics if available:
425 */
426#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
427
428void vtime_account_irq(struct task_struct *tsk, unsigned int offset)
429{
430 unsigned int pc = irq_count() - offset;
431
432 if (pc & HARDIRQ_OFFSET) {
433 vtime_account_hardirq(tsk);
434 } else if (pc & SOFTIRQ_OFFSET) {
435 vtime_account_softirq(tsk);
436 } else if (!IS_ENABLED(CONFIG_HAVE_VIRT_CPU_ACCOUNTING_IDLE) &&
437 is_idle_task(tsk)) {
438 vtime_account_idle(tsk);
439 } else {
440 vtime_account_kernel(tsk);
441 }
442}
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
479 if (vtime_accounting_enabled_this_cpu())
480 return;
481
482 if (irqtime_enabled()) {
483 irqtime_account_process_tick(p, user_tick, nr_ticks: 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 != this_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 (irqtime_enabled()) {
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 * Adjust tick based cputime random precision against scheduler runtime
528 * accounting.
529 *
530 * Tick based cputime accounting depend on random scheduling timeslices of a
531 * task to be interrupted or not by the timer. Depending on these
532 * circumstances, the number of these interrupts may be over or
533 * under-optimistic, matching the real user and system cputime with a variable
534 * precision.
535 *
536 * Fix this by scaling these tick based values against the total runtime
537 * accounted by the CFS scheduler.
538 *
539 * This code provides the following guarantees:
540 *
541 * stime + utime == rtime
542 * stime_i+1 >= stime_i, utime_i+1 >= utime_i
543 *
544 * Assuming that rtime_i+1 >= rtime_i.
545 */
546void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
547 u64 *ut, u64 *st)
548{
549 u64 rtime, stime, utime;
550 unsigned long flags;
551
552 /* Serialize concurrent callers such that we can honour our guarantees */
553 raw_spin_lock_irqsave(&prev->lock, flags);
554 rtime = curr->sum_exec_runtime;
555
556 /*
557 * This is possible under two circumstances:
558 * - rtime isn't monotonic after all (a bug);
559 * - we got reordered by the lock.
560 *
561 * In both cases this acts as a filter such that the rest of the code
562 * can assume it is monotonic regardless of anything else.
563 */
564 if (prev->stime + prev->utime >= rtime)
565 goto out;
566
567 stime = curr->stime;
568 utime = curr->utime;
569
570 /*
571 * If either stime or utime are 0, assume all runtime is userspace.
572 * Once a task gets some ticks, the monotonicity code at 'update:'
573 * will ensure things converge to the observed ratio.
574 */
575 if (stime == 0) {
576 utime = rtime;
577 goto update;
578 }
579
580 if (utime == 0) {
581 stime = rtime;
582 goto update;
583 }
584
585 stime = mul_u64_u64_div_u64(a: stime, mul: rtime, div: stime + utime);
586 /*
587 * Because mul_u64_u64_div_u64() can approximate on some
588 * achitectures; enforce the constraint that: a*b/(b+c) <= a.
589 */
590 if (unlikely(stime > rtime))
591 stime = rtime;
592
593update:
594 /*
595 * Make sure stime doesn't go backwards; this preserves monotonicity
596 * for utime because rtime is monotonic.
597 *
598 * utime_i+1 = rtime_i+1 - stime_i
599 * = rtime_i+1 - (rtime_i - utime_i)
600 * = (rtime_i+1 - rtime_i) + utime_i
601 * >= utime_i
602 */
603 if (stime < prev->stime)
604 stime = prev->stime;
605 utime = rtime - stime;
606
607 /*
608 * Make sure utime doesn't go backwards; this still preserves
609 * monotonicity for stime, analogous argument to above.
610 */
611 if (utime < prev->utime) {
612 utime = prev->utime;
613 stime = rtime - utime;
614 }
615
616 prev->stime = stime;
617 prev->utime = utime;
618out:
619 *ut = prev->utime;
620 *st = prev->stime;
621 raw_spin_unlock_irqrestore(&prev->lock, flags);
622}
623
624void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
625{
626 struct task_cputime cputime = {
627 .sum_exec_runtime = p->se.sum_exec_runtime,
628 };
629
630 if (task_cputime(t: p, utime: &cputime.utime, stime: &cputime.stime))
631 cputime.sum_exec_runtime = task_sched_runtime(task: p);
632 cputime_adjust(curr: &cputime, prev: &p->prev_cputime, ut, st);
633}
634EXPORT_SYMBOL_GPL(task_cputime_adjusted);
635
636void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
637{
638 struct task_cputime cputime;
639
640 thread_group_cputime(tsk: p, times: &cputime);
641 cputime_adjust(curr: &cputime, prev: &p->signal->prev_cputime, ut, st);
642}
643#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
644
645#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
646static u64 vtime_delta(struct vtime *vtime)
647{
648 unsigned long long clock;
649
650 clock = sched_clock();
651 if (clock < vtime->starttime)
652 return 0;
653
654 return clock - vtime->starttime;
655}
656
657static u64 get_vtime_delta(struct vtime *vtime)
658{
659 u64 delta = vtime_delta(vtime);
660 u64 other;
661
662 /*
663 * Unlike tick based timing, vtime based timing never has lost
664 * ticks, and no need for steal time accounting to make up for
665 * lost ticks. Vtime accounts a rounded version of actual
666 * elapsed time. Limit account_other_time to prevent rounding
667 * errors from causing elapsed vtime to go negative.
668 */
669 other = account_other_time(delta);
670 WARN_ON_ONCE(vtime->state == VTIME_INACTIVE);
671 vtime->starttime += delta;
672
673 return delta - other;
674}
675
676static void vtime_account_system(struct task_struct *tsk,
677 struct vtime *vtime)
678{
679 vtime->stime += get_vtime_delta(vtime);
680 if (vtime->stime >= TICK_NSEC) {
681 account_system_time(tsk, irq_count(), vtime->stime);
682 vtime->stime = 0;
683 }
684}
685
686static void vtime_account_guest(struct task_struct *tsk,
687 struct vtime *vtime)
688{
689 vtime->gtime += get_vtime_delta(vtime);
690 if (vtime->gtime >= TICK_NSEC) {
691 account_guest_time(tsk, vtime->gtime);
692 vtime->gtime = 0;
693 }
694}
695
696static void __vtime_account_kernel(struct task_struct *tsk,
697 struct vtime *vtime)
698{
699 /* We might have scheduled out from guest path */
700 if (vtime->state == VTIME_GUEST)
701 vtime_account_guest(tsk, vtime);
702 else
703 vtime_account_system(tsk, vtime);
704}
705
706void vtime_account_kernel(struct task_struct *tsk)
707{
708 struct vtime *vtime = &tsk->vtime;
709
710 if (!vtime_delta(vtime))
711 return;
712
713 write_seqcount_begin(&vtime->seqcount);
714 __vtime_account_kernel(tsk, vtime);
715 write_seqcount_end(&vtime->seqcount);
716}
717
718void vtime_user_enter(struct task_struct *tsk)
719{
720 struct vtime *vtime = &tsk->vtime;
721
722 write_seqcount_begin(&vtime->seqcount);
723 vtime_account_system(tsk, vtime);
724 vtime->state = VTIME_USER;
725 write_seqcount_end(&vtime->seqcount);
726}
727
728void vtime_user_exit(struct task_struct *tsk)
729{
730 struct vtime *vtime = &tsk->vtime;
731
732 write_seqcount_begin(&vtime->seqcount);
733 vtime->utime += get_vtime_delta(vtime);
734 if (vtime->utime >= TICK_NSEC) {
735 account_user_time(tsk, vtime->utime);
736 vtime->utime = 0;
737 }
738 vtime->state = VTIME_SYS;
739 write_seqcount_end(&vtime->seqcount);
740}
741
742void vtime_guest_enter(struct task_struct *tsk)
743{
744 struct vtime *vtime = &tsk->vtime;
745 /*
746 * The flags must be updated under the lock with
747 * the vtime_starttime flush and update.
748 * That enforces a right ordering and update sequence
749 * synchronization against the reader (task_gtime())
750 * that can thus safely catch up with a tickless delta.
751 */
752 write_seqcount_begin(&vtime->seqcount);
753 vtime_account_system(tsk, vtime);
754 tsk->flags |= PF_VCPU;
755 vtime->state = VTIME_GUEST;
756 write_seqcount_end(&vtime->seqcount);
757}
758EXPORT_SYMBOL_GPL(vtime_guest_enter);
759
760void vtime_guest_exit(struct task_struct *tsk)
761{
762 struct vtime *vtime = &tsk->vtime;
763
764 write_seqcount_begin(&vtime->seqcount);
765 vtime_account_guest(tsk, vtime);
766 tsk->flags &= ~PF_VCPU;
767 vtime->state = VTIME_SYS;
768 write_seqcount_end(&vtime->seqcount);
769}
770EXPORT_SYMBOL_GPL(vtime_guest_exit);
771
772void vtime_account_idle(struct task_struct *tsk)
773{
774 account_idle_time(get_vtime_delta(&tsk->vtime));
775}
776
777void vtime_task_switch_generic(struct task_struct *prev)
778{
779 struct vtime *vtime = &prev->vtime;
780
781 write_seqcount_begin(&vtime->seqcount);
782 if (vtime->state == VTIME_IDLE)
783 vtime_account_idle(prev);
784 else
785 __vtime_account_kernel(prev, vtime);
786 vtime->state = VTIME_INACTIVE;
787 vtime->cpu = -1;
788 write_seqcount_end(&vtime->seqcount);
789
790 vtime = &current->vtime;
791
792 write_seqcount_begin(&vtime->seqcount);
793 if (is_idle_task(current))
794 vtime->state = VTIME_IDLE;
795 else if (current->flags & PF_VCPU)
796 vtime->state = VTIME_GUEST;
797 else
798 vtime->state = VTIME_SYS;
799 vtime->starttime = sched_clock();
800 vtime->cpu = smp_processor_id();
801 write_seqcount_end(&vtime->seqcount);
802}
803
804void vtime_init_idle(struct task_struct *t, int cpu)
805{
806 struct vtime *vtime = &t->vtime;
807 unsigned long flags;
808
809 local_irq_save(flags);
810 write_seqcount_begin(&vtime->seqcount);
811 vtime->state = VTIME_IDLE;
812 vtime->starttime = sched_clock();
813 vtime->cpu = cpu;
814 write_seqcount_end(&vtime->seqcount);
815 local_irq_restore(flags);
816}
817
818u64 task_gtime(struct task_struct *t)
819{
820 struct vtime *vtime = &t->vtime;
821 unsigned int seq;
822 u64 gtime;
823
824 if (!vtime_accounting_enabled())
825 return t->gtime;
826
827 do {
828 seq = read_seqcount_begin(&vtime->seqcount);
829
830 gtime = t->gtime;
831 if (vtime->state == VTIME_GUEST)
832 gtime += vtime->gtime + vtime_delta(vtime);
833
834 } while (read_seqcount_retry(&vtime->seqcount, seq));
835
836 return gtime;
837}
838
839/*
840 * Fetch cputime raw values from fields of task_struct and
841 * add up the pending nohz execution time since the last
842 * cputime snapshot.
843 */
844bool task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
845{
846 struct vtime *vtime = &t->vtime;
847 unsigned int seq;
848 u64 delta;
849 int ret;
850
851 if (!vtime_accounting_enabled()) {
852 *utime = t->utime;
853 *stime = t->stime;
854 return false;
855 }
856
857 do {
858 ret = false;
859 seq = read_seqcount_begin(&vtime->seqcount);
860
861 *utime = t->utime;
862 *stime = t->stime;
863
864 /* Task is sleeping or idle, nothing to add */
865 if (vtime->state < VTIME_SYS)
866 continue;
867
868 ret = true;
869 delta = vtime_delta(vtime);
870
871 /*
872 * Task runs either in user (including guest) or kernel space,
873 * add pending nohz time to the right place.
874 */
875 if (vtime->state == VTIME_SYS)
876 *stime += vtime->stime + delta;
877 else
878 *utime += vtime->utime + delta;
879 } while (read_seqcount_retry(&vtime->seqcount, seq));
880
881 return ret;
882}
883
884static int vtime_state_fetch(struct vtime *vtime, int cpu)
885{
886 int state = READ_ONCE(vtime->state);
887
888 /*
889 * We raced against a context switch, fetch the
890 * kcpustat task again.
891 */
892 if (vtime->cpu != cpu && vtime->cpu != -1)
893 return -EAGAIN;
894
895 /*
896 * Two possible things here:
897 * 1) We are seeing the scheduling out task (prev) or any past one.
898 * 2) We are seeing the scheduling in task (next) but it hasn't
899 * passed though vtime_task_switch() yet so the pending
900 * cputime of the prev task may not be flushed yet.
901 *
902 * Case 1) is ok but 2) is not. So wait for a safe VTIME state.
903 */
904 if (state == VTIME_INACTIVE)
905 return -EAGAIN;
906
907 return state;
908}
909
910static u64 kcpustat_user_vtime(struct vtime *vtime)
911{
912 if (vtime->state == VTIME_USER)
913 return vtime->utime + vtime_delta(vtime);
914 else if (vtime->state == VTIME_GUEST)
915 return vtime->gtime + vtime_delta(vtime);
916 return 0;
917}
918
919static int kcpustat_field_vtime(u64 *cpustat,
920 struct task_struct *tsk,
921 enum cpu_usage_stat usage,
922 int cpu, u64 *val)
923{
924 struct vtime *vtime = &tsk->vtime;
925 unsigned int seq;
926
927 do {
928 int state;
929
930 seq = read_seqcount_begin(&vtime->seqcount);
931
932 state = vtime_state_fetch(vtime, cpu);
933 if (state < 0)
934 return state;
935
936 *val = cpustat[usage];
937
938 /*
939 * Nice VS unnice cputime accounting may be inaccurate if
940 * the nice value has changed since the last vtime update.
941 * But proper fix would involve interrupting target on nice
942 * updates which is a no go on nohz_full (although the scheduler
943 * may still interrupt the target if rescheduling is needed...)
944 */
945 switch (usage) {
946 case CPUTIME_SYSTEM:
947 if (state == VTIME_SYS)
948 *val += vtime->stime + vtime_delta(vtime);
949 break;
950 case CPUTIME_USER:
951 if (task_nice(tsk) <= 0)
952 *val += kcpustat_user_vtime(vtime);
953 break;
954 case CPUTIME_NICE:
955 if (task_nice(tsk) > 0)
956 *val += kcpustat_user_vtime(vtime);
957 break;
958 case CPUTIME_GUEST:
959 if (state == VTIME_GUEST && task_nice(tsk) <= 0)
960 *val += vtime->gtime + vtime_delta(vtime);
961 break;
962 case CPUTIME_GUEST_NICE:
963 if (state == VTIME_GUEST && task_nice(tsk) > 0)
964 *val += vtime->gtime + vtime_delta(vtime);
965 break;
966 default:
967 break;
968 }
969 } while (read_seqcount_retry(&vtime->seqcount, seq));
970
971 return 0;
972}
973
974u64 kcpustat_field(struct kernel_cpustat *kcpustat,
975 enum cpu_usage_stat usage, int cpu)
976{
977 u64 *cpustat = kcpustat->cpustat;
978 u64 val = cpustat[usage];
979 struct rq *rq;
980 int err;
981
982 if (!vtime_accounting_enabled_cpu(cpu))
983 return val;
984
985 rq = cpu_rq(cpu);
986
987 for (;;) {
988 struct task_struct *curr;
989
990 rcu_read_lock();
991 curr = rcu_dereference(rq->curr);
992 if (WARN_ON_ONCE(!curr)) {
993 rcu_read_unlock();
994 return cpustat[usage];
995 }
996
997 err = kcpustat_field_vtime(cpustat, curr, usage, cpu, &val);
998 rcu_read_unlock();
999
1000 if (!err)
1001 return val;
1002
1003 cpu_relax();
1004 }
1005}
1006EXPORT_SYMBOL_GPL(kcpustat_field);
1007
1008static int kcpustat_cpu_fetch_vtime(struct kernel_cpustat *dst,
1009 const struct kernel_cpustat *src,
1010 struct task_struct *tsk, int cpu)
1011{
1012 struct vtime *vtime = &tsk->vtime;
1013 unsigned int seq;
1014
1015 do {
1016 u64 *cpustat;
1017 u64 delta;
1018 int state;
1019
1020 seq = read_seqcount_begin(&vtime->seqcount);
1021
1022 state = vtime_state_fetch(vtime, cpu);
1023 if (state < 0)
1024 return state;
1025
1026 *dst = *src;
1027 cpustat = dst->cpustat;
1028
1029 /* Task is sleeping, dead or idle, nothing to add */
1030 if (state < VTIME_SYS)
1031 continue;
1032
1033 delta = vtime_delta(vtime);
1034
1035 /*
1036 * Task runs either in user (including guest) or kernel space,
1037 * add pending nohz time to the right place.
1038 */
1039 if (state == VTIME_SYS) {
1040 cpustat[CPUTIME_SYSTEM] += vtime->stime + delta;
1041 } else if (state == VTIME_USER) {
1042 if (task_nice(tsk) > 0)
1043 cpustat[CPUTIME_NICE] += vtime->utime + delta;
1044 else
1045 cpustat[CPUTIME_USER] += vtime->utime + delta;
1046 } else {
1047 WARN_ON_ONCE(state != VTIME_GUEST);
1048 if (task_nice(tsk) > 0) {
1049 cpustat[CPUTIME_GUEST_NICE] += vtime->gtime + delta;
1050 cpustat[CPUTIME_NICE] += vtime->gtime + delta;
1051 } else {
1052 cpustat[CPUTIME_GUEST] += vtime->gtime + delta;
1053 cpustat[CPUTIME_USER] += vtime->gtime + delta;
1054 }
1055 }
1056 } while (read_seqcount_retry(&vtime->seqcount, seq));
1057
1058 return 0;
1059}
1060
1061void kcpustat_cpu_fetch(struct kernel_cpustat *dst, int cpu)
1062{
1063 const struct kernel_cpustat *src = &kcpustat_cpu(cpu);
1064 struct rq *rq;
1065 int err;
1066
1067 if (!vtime_accounting_enabled_cpu(cpu)) {
1068 *dst = *src;
1069 return;
1070 }
1071
1072 rq = cpu_rq(cpu);
1073
1074 for (;;) {
1075 struct task_struct *curr;
1076
1077 rcu_read_lock();
1078 curr = rcu_dereference(rq->curr);
1079 if (WARN_ON_ONCE(!curr)) {
1080 rcu_read_unlock();
1081 *dst = *src;
1082 return;
1083 }
1084
1085 err = kcpustat_cpu_fetch_vtime(dst, src, curr, cpu);
1086 rcu_read_unlock();
1087
1088 if (!err)
1089 return;
1090
1091 cpu_relax();
1092 }
1093}
1094EXPORT_SYMBOL_GPL(kcpustat_cpu_fetch);
1095
1096#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */
1097