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 after incrementing preempt_count on {soft,}irq_enter
48 * and before decrementing preempt_count on {soft,}irq_exit.
49 */
50void irqtime_account_irq(struct task_struct *curr, unsigned int offset)
51{
52 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
53 unsigned int pc;
54 s64 delta;
55 int cpu;
56
57 if (!sched_clock_irqtime)
58 return;
59
60 cpu = smp_processor_id();
61 delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
62 irqtime->irq_start_time += delta;
63 pc = irq_count() - offset;
64
65 /*
66 * We do not account for softirq time from ksoftirqd here.
67 * We want to continue accounting softirq time to ksoftirqd thread
68 * in that case, so as not to confuse scheduler with a special task
69 * that do not consume any time, but still wants to run.
70 */
71 if (pc & HARDIRQ_MASK)
72 irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
73 else if ((pc & SOFTIRQ_OFFSET) && curr != this_cpu_ksoftirqd())
74 irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
75}
76
77static u64 irqtime_tick_accounted(u64 maxtime)
78{
79 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
80 u64 delta;
81
82 delta = min(irqtime->tick_delta, maxtime);
83 irqtime->tick_delta -= delta;
84
85 return delta;
86}
87
88#else /* CONFIG_IRQ_TIME_ACCOUNTING */
89
90#define sched_clock_irqtime (0)
91
92static u64 irqtime_tick_accounted(u64 dummy)
93{
94 return 0;
95}
96
97#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
98
99static inline void task_group_account_field(struct task_struct *p, int index,
100 u64 tmp)
101{
102 /*
103 * Since all updates are sure to touch the root cgroup, we
104 * get ourselves ahead and touch it first. If the root cgroup
105 * is the only cgroup, then nothing else should be necessary.
106 *
107 */
108 __this_cpu_add(kernel_cpustat.cpustat[index], tmp);
109
110 cgroup_account_cputime_field(p, index, tmp);
111}
112
113/*
114 * Account user CPU time to a process.
115 * @p: the process that the CPU time gets accounted to
116 * @cputime: the CPU time spent in user space since the last update
117 */
118void account_user_time(struct task_struct *p, u64 cputime)
119{
120 int index;
121
122 /* Add user time to process. */
123 p->utime += cputime;
124 account_group_user_time(p, cputime);
125
126 index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
127
128 /* Add user time to cpustat. */
129 task_group_account_field(p, index, cputime);
130
131 /* Account for user time used */
132 acct_account_cputime(p);
133}
134
135/*
136 * Account guest CPU time to a process.
137 * @p: the process that the CPU time gets accounted to
138 * @cputime: the CPU time spent in virtual machine since the last update
139 */
140void account_guest_time(struct task_struct *p, u64 cputime)
141{
142 u64 *cpustat = kcpustat_this_cpu->cpustat;
143
144 /* Add guest time to process. */
145 p->utime += cputime;
146 account_group_user_time(p, cputime);
147 p->gtime += cputime;
148
149 /* Add guest time to cpustat. */
150 if (task_nice(p) > 0) {
151 cpustat[CPUTIME_NICE] += cputime;
152 cpustat[CPUTIME_GUEST_NICE] += cputime;
153 } else {
154 cpustat[CPUTIME_USER] += cputime;
155 cpustat[CPUTIME_GUEST] += cputime;
156 }
157}
158
159/*
160 * Account system CPU time to a process and desired cpustat field
161 * @p: the process that the CPU time gets accounted to
162 * @cputime: the CPU time spent in kernel space since the last update
163 * @index: pointer to cpustat field that has to be updated
164 */
165void account_system_index_time(struct task_struct *p,
166 u64 cputime, enum cpu_usage_stat index)
167{
168 /* Add system time to process. */
169 p->stime += cputime;
170 account_group_system_time(p, cputime);
171
172 /* Add system time to cpustat. */
173 task_group_account_field(p, index, cputime);
174
175 /* Account for system time used */
176 acct_account_cputime(p);
177}
178
179/*
180 * Account system CPU time to a process.
181 * @p: the process that the CPU time gets accounted to
182 * @hardirq_offset: the offset to subtract from hardirq_count()
183 * @cputime: the CPU time spent in kernel space since the last update
184 */
185void account_system_time(struct task_struct *p, int hardirq_offset, u64 cputime)
186{
187 int index;
188
189 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
190 account_guest_time(p, cputime);
191 return;
192 }
193
194 if (hardirq_count() - hardirq_offset)
195 index = CPUTIME_IRQ;
196 else if (in_serving_softirq())
197 index = CPUTIME_SOFTIRQ;
198 else
199 index = CPUTIME_SYSTEM;
200
201 account_system_index_time(p, cputime, index);
202}
203
204/*
205 * Account for involuntary wait time.
206 * @cputime: the CPU time spent in involuntary wait
207 */
208void account_steal_time(u64 cputime)
209{
210 u64 *cpustat = kcpustat_this_cpu->cpustat;
211
212 cpustat[CPUTIME_STEAL] += cputime;
213}
214
215/*
216 * Account for idle time.
217 * @cputime: the CPU time spent in idle wait
218 */
219void account_idle_time(u64 cputime)
220{
221 u64 *cpustat = kcpustat_this_cpu->cpustat;
222 struct rq *rq = this_rq();
223
224 if (atomic_read(&rq->nr_iowait) > 0)
225 cpustat[CPUTIME_IOWAIT] += cputime;
226 else
227 cpustat[CPUTIME_IDLE] += cputime;
228}
229
230/*
231 * When a guest is interrupted for a longer amount of time, missed clock
232 * ticks are not redelivered later. Due to that, this function may on
233 * occasion account more time than the calling functions think elapsed.
234 */
235static __always_inline u64 steal_account_process_time(u64 maxtime)
236{
237#ifdef CONFIG_PARAVIRT
238 if (static_key_false(¶virt_steal_enabled)) {
239 u64 steal;
240
241 steal = paravirt_steal_clock(smp_processor_id());
242 steal -= this_rq()->prev_steal_time;
243 steal = min(steal, maxtime);
244 account_steal_time(steal);
245 this_rq()->prev_steal_time += steal;
246
247 return steal;
248 }
249#endif
250 return 0;
251}
252
253/*
254 * Account how much elapsed time was spent in steal, irq, or softirq time.
255 */
256static inline u64 account_other_time(u64 max)
257{
258 u64 accounted;
259
260 lockdep_assert_irqs_disabled();
261
262 accounted = steal_account_process_time(max);
263
264 if (accounted < max)
265 accounted += irqtime_tick_accounted(max - accounted);
266
267 return accounted;
268}
269
270#ifdef CONFIG_64BIT
271static inline u64 read_sum_exec_runtime(struct task_struct *t)
272{
273 return t->se.sum_exec_runtime;
274}
275#else
276static u64 read_sum_exec_runtime(struct task_struct *t)
277{
278 u64 ns;
279 struct rq_flags rf;
280 struct rq *rq;
281
282 rq = task_rq_lock(t, &rf);
283 ns = t->se.sum_exec_runtime;
284 task_rq_unlock(rq, t, &rf);
285
286 return ns;
287}
288#endif
289
290/*
291 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
292 * tasks (sum on group iteration) belonging to @tsk's group.
293 */
294void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
295{
296 struct signal_struct *sig = tsk->signal;
297 u64 utime, stime;
298 struct task_struct *t;
299 unsigned int seq, nextseq;
300 unsigned long flags;
301
302 /*
303 * Update current task runtime to account pending time since last
304 * scheduler action or thread_group_cputime() call. This thread group
305 * might have other running tasks on different CPUs, but updating
306 * their runtime can affect syscall performance, so we skip account
307 * those pending times and rely only on values updated on tick or
308 * other scheduler action.
309 */
310 if (same_thread_group(current, tsk))
311 (void) task_sched_runtime(current);
312
313 rcu_read_lock();
314 /* Attempt a lockless read on the first round. */
315 nextseq = 0;
316 do {
317 seq = nextseq;
318 flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
319 times->utime = sig->utime;
320 times->stime = sig->stime;
321 times->sum_exec_runtime = sig->sum_sched_runtime;
322
323 for_each_thread(tsk, t) {
324 task_cputime(t, &utime, &stime);
325 times->utime += utime;
326 times->stime += stime;
327 times->sum_exec_runtime += read_sum_exec_runtime(t);
328 }
329 /* If lockless access failed, take the lock. */
330 nextseq = 1;
331 } while (need_seqretry(&sig->stats_lock, seq));
332 done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
333 rcu_read_unlock();
334}
335
336#ifdef CONFIG_IRQ_TIME_ACCOUNTING
337/*
338 * Account a tick to a process and cpustat
339 * @p: the process that the CPU time gets accounted to
340 * @user_tick: is the tick from userspace
341 * @rq: the pointer to rq
342 *
343 * Tick demultiplexing follows the order
344 * - pending hardirq update
345 * - pending softirq update
346 * - user_time
347 * - idle_time
348 * - system time
349 * - check for guest_time
350 * - else account as system_time
351 *
352 * Check for hardirq is done both for system and user time as there is
353 * no timer going off while we are on hardirq and hence we may never get an
354 * opportunity to update it solely in system time.
355 * p->stime and friends are only updated on system time and not on irq
356 * softirq as those do not count in task exec_runtime any more.
357 */
358static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
359 int ticks)
360{
361 u64 other, cputime = TICK_NSEC * ticks;
362
363 /*
364 * When returning from idle, many ticks can get accounted at
365 * once, including some ticks of steal, irq, and softirq time.
366 * Subtract those ticks from the amount of time accounted to
367 * idle, or potentially user or system time. Due to rounding,
368 * other time can exceed ticks occasionally.
369 */
370 other = account_other_time(ULONG_MAX);
371 if (other >= cputime)
372 return;
373
374 cputime -= other;
375
376 if (this_cpu_ksoftirqd() == p) {
377 /*
378 * ksoftirqd time do not get accounted in cpu_softirq_time.
379 * So, we have to handle it separately here.
380 * Also, p->stime needs to be updated for ksoftirqd.
381 */
382 account_system_index_time(p, cputime, CPUTIME_SOFTIRQ);
383 } else if (user_tick) {
384 account_user_time(p, cputime);
385 } else if (p == this_rq()->idle) {
386 account_idle_time(cputime);
387 } else if (p->flags & PF_VCPU) { /* System time or guest time */
388 account_guest_time(p, cputime);
389 } else {
390 account_system_index_time(p, cputime, CPUTIME_SYSTEM);
391 }
392}
393
394static void irqtime_account_idle_ticks(int ticks)
395{
396 irqtime_account_process_tick(current, 0, ticks);
397}
398#else /* CONFIG_IRQ_TIME_ACCOUNTING */
399static inline void irqtime_account_idle_ticks(int ticks) { }
400static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
401 int nr_ticks) { }
402#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
403
404/*
405 * Use precise platform statistics if available:
406 */
407#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
408
409# ifndef __ARCH_HAS_VTIME_TASK_SWITCH
410void vtime_task_switch(struct task_struct *prev)
411{
412 if (is_idle_task(prev))
413 vtime_account_idle(prev);
414 else
415 vtime_account_kernel(prev);
416
417 vtime_flush(prev);
418 arch_vtime_task_switch(prev);
419}
420# endif
421
422void vtime_account_irq(struct task_struct *tsk, unsigned int offset)
423{
424 unsigned int pc = irq_count() - offset;
425
426 if (pc & HARDIRQ_OFFSET) {
427 vtime_account_hardirq(tsk);
428 } else if (pc & SOFTIRQ_OFFSET) {
429 vtime_account_softirq(tsk);
430 } else if (!IS_ENABLED(CONFIG_HAVE_VIRT_CPU_ACCOUNTING_IDLE) &&
431 is_idle_task(tsk)) {
432 vtime_account_idle(tsk);
433 } else {
434 vtime_account_kernel(tsk);
435 }
436}
437
438void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
439 u64 *ut, u64 *st)
440{
441 *ut = curr->utime;
442 *st = curr->stime;
443}
444
445void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
446{
447 *ut = p->utime;
448 *st = p->stime;
449}
450EXPORT_SYMBOL_GPL(task_cputime_adjusted);
451
452void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
453{
454 struct task_cputime cputime;
455
456 thread_group_cputime(p, &cputime);
457
458 *ut = cputime.utime;
459 *st = cputime.stime;
460}
461
462#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE: */
463
464/*
465 * Account a single tick of CPU time.
466 * @p: the process that the CPU time gets accounted to
467 * @user_tick: indicates if the tick is a user or a system tick
468 */
469void account_process_tick(struct task_struct *p, int user_tick)
470{
471 u64 cputime, steal;
472
473 if (vtime_accounting_enabled_this_cpu())
474 return;
475
476 if (sched_clock_irqtime) {
477 irqtime_account_process_tick(p, user_tick, 1);
478 return;
479 }
480
481 cputime = TICK_NSEC;
482 steal = steal_account_process_time(ULONG_MAX);
483
484 if (steal >= cputime)
485 return;
486
487 cputime -= steal;
488
489 if (user_tick)
490 account_user_time(p, cputime);
491 else if ((p != this_rq()->idle) || (irq_count() != HARDIRQ_OFFSET))
492 account_system_time(p, HARDIRQ_OFFSET, cputime);
493 else
494 account_idle_time(cputime);
495}
496
497/*
498 * Account multiple ticks of idle time.
499 * @ticks: number of stolen ticks
500 */
501void account_idle_ticks(unsigned long ticks)
502{
503 u64 cputime, steal;
504
505 if (sched_clock_irqtime) {
506 irqtime_account_idle_ticks(ticks);
507 return;
508 }
509
510 cputime = ticks * TICK_NSEC;
511 steal = steal_account_process_time(ULONG_MAX);
512
513 if (steal >= cputime)
514 return;
515
516 cputime -= steal;
517 account_idle_time(cputime);
518}
519
520/*
521 * Adjust tick based cputime random precision against scheduler runtime
522 * accounting.
523 *
524 * Tick based cputime accounting depend on random scheduling timeslices of a
525 * task to be interrupted or not by the timer. Depending on these
526 * circumstances, the number of these interrupts may be over or
527 * under-optimistic, matching the real user and system cputime with a variable
528 * precision.
529 *
530 * Fix this by scaling these tick based values against the total runtime
531 * accounted by the CFS scheduler.
532 *
533 * This code provides the following guarantees:
534 *
535 * stime + utime == rtime
536 * stime_i+1 >= stime_i, utime_i+1 >= utime_i
537 *
538 * Assuming that rtime_i+1 >= rtime_i.
539 */
540void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
541 u64 *ut, u64 *st)
542{
543 u64 rtime, stime, utime;
544 unsigned long flags;
545
546 /* Serialize concurrent callers such that we can honour our guarantees */
547 raw_spin_lock_irqsave(&prev->lock, flags);
548 rtime = curr->sum_exec_runtime;
549
550 /*
551 * This is possible under two circumstances:
552 * - rtime isn't monotonic after all (a bug);
553 * - we got reordered by the lock.
554 *
555 * In both cases this acts as a filter such that the rest of the code
556 * can assume it is monotonic regardless of anything else.
557 */
558 if (prev->stime + prev->utime >= rtime)
559 goto out;
560
561 stime = curr->stime;
562 utime = curr->utime;
563
564 /*
565 * If either stime or utime are 0, assume all runtime is userspace.
566 * Once a task gets some ticks, the monotonicity code at 'update:'
567 * will ensure things converge to the observed ratio.
568 */
569 if (stime == 0) {
570 utime = rtime;
571 goto update;
572 }
573
574 if (utime == 0) {
575 stime = rtime;
576 goto update;
577 }
578
579 stime = mul_u64_u64_div_u64(stime, rtime, stime + utime);
580
581update:
582 /*
583 * Make sure stime doesn't go backwards; this preserves monotonicity
584 * for utime because rtime is monotonic.
585 *
586 * utime_i+1 = rtime_i+1 - stime_i
587 * = rtime_i+1 - (rtime_i - utime_i)
588 * = (rtime_i+1 - rtime_i) + utime_i
589 * >= utime_i
590 */
591 if (stime < prev->stime)
592 stime = prev->stime;
593 utime = rtime - stime;
594
595 /*
596 * Make sure utime doesn't go backwards; this still preserves
597 * monotonicity for stime, analogous argument to above.
598 */
599 if (utime < prev->utime) {
600 utime = prev->utime;
601 stime = rtime - utime;
602 }
603
604 prev->stime = stime;
605 prev->utime = utime;
606out:
607 *ut = prev->utime;
608 *st = prev->stime;
609 raw_spin_unlock_irqrestore(&prev->lock, flags);
610}
611
612void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
613{
614 struct task_cputime cputime = {
615 .sum_exec_runtime = p->se.sum_exec_runtime,
616 };
617
618 task_cputime(p, &cputime.utime, &cputime.stime);
619 cputime_adjust(&cputime, &p->prev_cputime, ut, st);
620}
621EXPORT_SYMBOL_GPL(task_cputime_adjusted);
622
623void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
624{
625 struct task_cputime cputime;
626
627 thread_group_cputime(p, &cputime);
628 cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
629}
630#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
631
632#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
633static u64 vtime_delta(struct vtime *vtime)
634{
635 unsigned long long clock;
636
637 clock = sched_clock();
638 if (clock < vtime->starttime)
639 return 0;
640
641 return clock - vtime->starttime;
642}
643
644static u64 get_vtime_delta(struct vtime *vtime)
645{
646 u64 delta = vtime_delta(vtime);
647 u64 other;
648
649 /*
650 * Unlike tick based timing, vtime based timing never has lost
651 * ticks, and no need for steal time accounting to make up for
652 * lost ticks. Vtime accounts a rounded version of actual
653 * elapsed time. Limit account_other_time to prevent rounding
654 * errors from causing elapsed vtime to go negative.
655 */
656 other = account_other_time(delta);
657 WARN_ON_ONCE(vtime->state == VTIME_INACTIVE);
658 vtime->starttime += delta;
659
660 return delta - other;
661}
662
663static void vtime_account_system(struct task_struct *tsk,
664 struct vtime *vtime)
665{
666 vtime->stime += get_vtime_delta(vtime);
667 if (vtime->stime >= TICK_NSEC) {
668 account_system_time(tsk, irq_count(), vtime->stime);
669 vtime->stime = 0;
670 }
671}
672
673static void vtime_account_guest(struct task_struct *tsk,
674 struct vtime *vtime)
675{
676 vtime->gtime += get_vtime_delta(vtime);
677 if (vtime->gtime >= TICK_NSEC) {
678 account_guest_time(tsk, vtime->gtime);
679 vtime->gtime = 0;
680 }
681}
682
683static void __vtime_account_kernel(struct task_struct *tsk,
684 struct vtime *vtime)
685{
686 /* We might have scheduled out from guest path */
687 if (vtime->state == VTIME_GUEST)
688 vtime_account_guest(tsk, vtime);
689 else
690 vtime_account_system(tsk, vtime);
691}
692
693void vtime_account_kernel(struct task_struct *tsk)
694{
695 struct vtime *vtime = &tsk->vtime;
696
697 if (!vtime_delta(vtime))
698 return;
699
700 write_seqcount_begin(&vtime->seqcount);
701 __vtime_account_kernel(tsk, vtime);
702 write_seqcount_end(&vtime->seqcount);
703}
704
705void vtime_user_enter(struct task_struct *tsk)
706{
707 struct vtime *vtime = &tsk->vtime;
708
709 write_seqcount_begin(&vtime->seqcount);
710 vtime_account_system(tsk, vtime);
711 vtime->state = VTIME_USER;
712 write_seqcount_end(&vtime->seqcount);
713}
714
715void vtime_user_exit(struct task_struct *tsk)
716{
717 struct vtime *vtime = &tsk->vtime;
718
719 write_seqcount_begin(&vtime->seqcount);
720 vtime->utime += get_vtime_delta(vtime);
721 if (vtime->utime >= TICK_NSEC) {
722 account_user_time(tsk, vtime->utime);
723 vtime->utime = 0;
724 }
725 vtime->state = VTIME_SYS;
726 write_seqcount_end(&vtime->seqcount);
727}
728
729void vtime_guest_enter(struct task_struct *tsk)
730{
731 struct vtime *vtime = &tsk->vtime;
732 /*
733 * The flags must be updated under the lock with
734 * the vtime_starttime flush and update.
735 * That enforces a right ordering and update sequence
736 * synchronization against the reader (task_gtime())
737 * that can thus safely catch up with a tickless delta.
738 */
739 write_seqcount_begin(&vtime->seqcount);
740 vtime_account_system(tsk, vtime);
741 tsk->flags |= PF_VCPU;
742 vtime->state = VTIME_GUEST;
743 write_seqcount_end(&vtime->seqcount);
744}
745EXPORT_SYMBOL_GPL(vtime_guest_enter);
746
747void vtime_guest_exit(struct task_struct *tsk)
748{
749 struct vtime *vtime = &tsk->vtime;
750
751 write_seqcount_begin(&vtime->seqcount);
752 vtime_account_guest(tsk, vtime);
753 tsk->flags &= ~PF_VCPU;
754 vtime->state = VTIME_SYS;
755 write_seqcount_end(&vtime->seqcount);
756}
757EXPORT_SYMBOL_GPL(vtime_guest_exit);
758
759void vtime_account_idle(struct task_struct *tsk)
760{
761 account_idle_time(get_vtime_delta(&tsk->vtime));
762}
763
764void vtime_task_switch_generic(struct task_struct *prev)
765{
766 struct vtime *vtime = &prev->vtime;
767
768 write_seqcount_begin(&vtime->seqcount);
769 if (vtime->state == VTIME_IDLE)
770 vtime_account_idle(prev);
771 else
772 __vtime_account_kernel(prev, vtime);
773 vtime->state = VTIME_INACTIVE;
774 vtime->cpu = -1;
775 write_seqcount_end(&vtime->seqcount);
776
777 vtime = ¤t->vtime;
778
779 write_seqcount_begin(&vtime->seqcount);
780 if (is_idle_task(current))
781 vtime->state = VTIME_IDLE;
782 else if (current->flags & PF_VCPU)
783 vtime->state = VTIME_GUEST;
784 else
785 vtime->state = VTIME_SYS;
786 vtime->starttime = sched_clock();
787 vtime->cpu = smp_processor_id();
788 write_seqcount_end(&vtime->seqcount);
789}
790
791void vtime_init_idle(struct task_struct *t, int cpu)
792{
793 struct vtime *vtime = &t->vtime;
794 unsigned long flags;
795
796 local_irq_save(flags);
797 write_seqcount_begin(&vtime->seqcount);
798 vtime->state = VTIME_IDLE;
799 vtime->starttime = sched_clock();
800 vtime->cpu = cpu;
801 write_seqcount_end(&vtime->seqcount);
802 local_irq_restore(flags);
803}
804
805u64 task_gtime(struct task_struct *t)
806{
807 struct vtime *vtime = &t->vtime;
808 unsigned int seq;
809 u64 gtime;
810
811 if (!vtime_accounting_enabled())
812 return t->gtime;
813
814 do {
815 seq = read_seqcount_begin(&vtime->seqcount);
816
817 gtime = t->gtime;
818 if (vtime->state == VTIME_GUEST)
819 gtime += vtime->gtime + vtime_delta(vtime);
820
821 } while (read_seqcount_retry(&vtime->seqcount, seq));
822
823 return gtime;
824}
825
826/*
827 * Fetch cputime raw values from fields of task_struct and
828 * add up the pending nohz execution time since the last
829 * cputime snapshot.
830 */
831void task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
832{
833 struct vtime *vtime = &t->vtime;
834 unsigned int seq;
835 u64 delta;
836
837 if (!vtime_accounting_enabled()) {
838 *utime = t->utime;
839 *stime = t->stime;
840 return;
841 }
842
843 do {
844 seq = read_seqcount_begin(&vtime->seqcount);
845
846 *utime = t->utime;
847 *stime = t->stime;
848
849 /* Task is sleeping or idle, nothing to add */
850 if (vtime->state < VTIME_SYS)
851 continue;
852
853 delta = vtime_delta(vtime);
854
855 /*
856 * Task runs either in user (including guest) or kernel space,
857 * add pending nohz time to the right place.
858 */
859 if (vtime->state == VTIME_SYS)
860 *stime += vtime->stime + delta;
861 else
862 *utime += vtime->utime + delta;
863 } while (read_seqcount_retry(&vtime->seqcount, seq));
864}
865
866static int vtime_state_fetch(struct vtime *vtime, int cpu)
867{
868 int state = READ_ONCE(vtime->state);
869
870 /*
871 * We raced against a context switch, fetch the
872 * kcpustat task again.
873 */
874 if (vtime->cpu != cpu && vtime->cpu != -1)
875 return -EAGAIN;
876
877 /*
878 * Two possible things here:
879 * 1) We are seeing the scheduling out task (prev) or any past one.
880 * 2) We are seeing the scheduling in task (next) but it hasn't
881 * passed though vtime_task_switch() yet so the pending
882 * cputime of the prev task may not be flushed yet.
883 *
884 * Case 1) is ok but 2) is not. So wait for a safe VTIME state.
885 */
886 if (state == VTIME_INACTIVE)
887 return -EAGAIN;
888
889 return state;
890}
891
892static u64 kcpustat_user_vtime(struct vtime *vtime)
893{
894 if (vtime->state == VTIME_USER)
895 return vtime->utime + vtime_delta(vtime);
896 else if (vtime->state == VTIME_GUEST)
897 return vtime->gtime + vtime_delta(vtime);
898 return 0;
899}
900
901static int kcpustat_field_vtime(u64 *cpustat,
902 struct task_struct *tsk,
903 enum cpu_usage_stat usage,
904 int cpu, u64 *val)
905{
906 struct vtime *vtime = &tsk->vtime;
907 unsigned int seq;
908
909 do {
910 int state;
911
912 seq = read_seqcount_begin(&vtime->seqcount);
913
914 state = vtime_state_fetch(vtime, cpu);
915 if (state < 0)
916 return state;
917
918 *val = cpustat[usage];
919
920 /*
921 * Nice VS unnice cputime accounting may be inaccurate if
922 * the nice value has changed since the last vtime update.
923 * But proper fix would involve interrupting target on nice
924 * updates which is a no go on nohz_full (although the scheduler
925 * may still interrupt the target if rescheduling is needed...)
926 */
927 switch (usage) {
928 case CPUTIME_SYSTEM:
929 if (state == VTIME_SYS)
930 *val += vtime->stime + vtime_delta(vtime);
931 break;
932 case CPUTIME_USER:
933 if (task_nice(tsk) <= 0)
934 *val += kcpustat_user_vtime(vtime);
935 break;
936 case CPUTIME_NICE:
937 if (task_nice(tsk) > 0)
938 *val += kcpustat_user_vtime(vtime);
939 break;
940 case CPUTIME_GUEST:
941 if (state == VTIME_GUEST && task_nice(tsk) <= 0)
942 *val += vtime->gtime + vtime_delta(vtime);
943 break;
944 case CPUTIME_GUEST_NICE:
945 if (state == VTIME_GUEST && task_nice(tsk) > 0)
946 *val += vtime->gtime + vtime_delta(vtime);
947 break;
948 default:
949 break;
950 }
951 } while (read_seqcount_retry(&vtime->seqcount, seq));
952
953 return 0;
954}
955
956u64 kcpustat_field(struct kernel_cpustat *kcpustat,
957 enum cpu_usage_stat usage, int cpu)
958{
959 u64 *cpustat = kcpustat->cpustat;
960 u64 val = cpustat[usage];
961 struct rq *rq;
962 int err;
963
964 if (!vtime_accounting_enabled_cpu(cpu))
965 return val;
966
967 rq = cpu_rq(cpu);
968
969 for (;;) {
970 struct task_struct *curr;
971
972 rcu_read_lock();
973 curr = rcu_dereference(rq->curr);
974 if (WARN_ON_ONCE(!curr)) {
975 rcu_read_unlock();
976 return cpustat[usage];
977 }
978
979 err = kcpustat_field_vtime(cpustat, curr, usage, cpu, &val);
980 rcu_read_unlock();
981
982 if (!err)
983 return val;
984
985 cpu_relax();
986 }
987}
988EXPORT_SYMBOL_GPL(kcpustat_field);
989
990static int kcpustat_cpu_fetch_vtime(struct kernel_cpustat *dst,
991 const struct kernel_cpustat *src,
992 struct task_struct *tsk, int cpu)
993{
994 struct vtime *vtime = &tsk->vtime;
995 unsigned int seq;
996
997 do {
998 u64 *cpustat;
999 u64 delta;
1000 int state;
1001
1002 seq = read_seqcount_begin(&vtime->seqcount);
1003
1004 state = vtime_state_fetch(vtime, cpu);
1005 if (state < 0)
1006 return state;
1007
1008 *dst = *src;
1009 cpustat = dst->cpustat;
1010
1011 /* Task is sleeping, dead or idle, nothing to add */
1012 if (state < VTIME_SYS)
1013 continue;
1014
1015 delta = vtime_delta(vtime);
1016
1017 /*
1018 * Task runs either in user (including guest) or kernel space,
1019 * add pending nohz time to the right place.
1020 */
1021 if (state == VTIME_SYS) {
1022 cpustat[CPUTIME_SYSTEM] += vtime->stime + delta;
1023 } else if (state == VTIME_USER) {
1024 if (task_nice(tsk) > 0)
1025 cpustat[CPUTIME_NICE] += vtime->utime + delta;
1026 else
1027 cpustat[CPUTIME_USER] += vtime->utime + delta;
1028 } else {
1029 WARN_ON_ONCE(state != VTIME_GUEST);
1030 if (task_nice(tsk) > 0) {
1031 cpustat[CPUTIME_GUEST_NICE] += vtime->gtime + delta;
1032 cpustat[CPUTIME_NICE] += vtime->gtime + delta;
1033 } else {
1034 cpustat[CPUTIME_GUEST] += vtime->gtime + delta;
1035 cpustat[CPUTIME_USER] += vtime->gtime + delta;
1036 }
1037 }
1038 } while (read_seqcount_retry(&vtime->seqcount, seq));
1039
1040 return 0;
1041}
1042
1043void kcpustat_cpu_fetch(struct kernel_cpustat *dst, int cpu)
1044{
1045 const struct kernel_cpustat *src = &kcpustat_cpu(cpu);
1046 struct rq *rq;
1047 int err;
1048
1049 if (!vtime_accounting_enabled_cpu(cpu)) {
1050 *dst = *src;
1051 return;
1052 }
1053
1054 rq = cpu_rq(cpu);
1055
1056 for (;;) {
1057 struct task_struct *curr;
1058
1059 rcu_read_lock();
1060 curr = rcu_dereference(rq->curr);
1061 if (WARN_ON_ONCE(!curr)) {
1062 rcu_read_unlock();
1063 *dst = *src;
1064 return;
1065 }
1066
1067 err = kcpustat_cpu_fetch_vtime(dst, src, curr, cpu);
1068 rcu_read_unlock();
1069
1070 if (!err)
1071 return;
1072
1073 cpu_relax();
1074 }
1075}
1076EXPORT_SYMBOL_GPL(kcpustat_cpu_fetch);
1077
1078#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Simple CPU accounting cgroup controller
4 */
5
6#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
7 #include <asm/cputime.h>
8#endif
9
10#ifdef CONFIG_IRQ_TIME_ACCOUNTING
11
12/*
13 * There are no locks covering percpu hardirq/softirq time.
14 * They are only modified in vtime_account, on corresponding CPU
15 * with interrupts disabled. So, writes are safe.
16 * They are read and saved off onto struct rq in update_rq_clock().
17 * This may result in other CPU reading this CPU's irq time and can
18 * race with irq/vtime_account on this CPU. We would either get old
19 * or new value with a side effect of accounting a slice of irq time to wrong
20 * task when irq is in progress while we read rq->clock. That is a worthy
21 * compromise in place of having locks on each irq in account_system_time.
22 */
23DEFINE_PER_CPU(struct irqtime, cpu_irqtime);
24
25static int sched_clock_irqtime;
26
27void enable_sched_clock_irqtime(void)
28{
29 sched_clock_irqtime = 1;
30}
31
32void disable_sched_clock_irqtime(void)
33{
34 sched_clock_irqtime = 0;
35}
36
37static void irqtime_account_delta(struct irqtime *irqtime, u64 delta,
38 enum cpu_usage_stat idx)
39{
40 u64 *cpustat = kcpustat_this_cpu->cpustat;
41
42 u64_stats_update_begin(&irqtime->sync);
43 cpustat[idx] += delta;
44 irqtime->total += delta;
45 irqtime->tick_delta += delta;
46 u64_stats_update_end(&irqtime->sync);
47}
48
49/*
50 * Called after incrementing preempt_count on {soft,}irq_enter
51 * and before decrementing preempt_count on {soft,}irq_exit.
52 */
53void irqtime_account_irq(struct task_struct *curr, unsigned int offset)
54{
55 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
56 unsigned int pc;
57 s64 delta;
58 int cpu;
59
60 if (!sched_clock_irqtime)
61 return;
62
63 cpu = smp_processor_id();
64 delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
65 irqtime->irq_start_time += delta;
66 pc = irq_count() - offset;
67
68 /*
69 * We do not account for softirq time from ksoftirqd here.
70 * We want to continue accounting softirq time to ksoftirqd thread
71 * in that case, so as not to confuse scheduler with a special task
72 * that do not consume any time, but still wants to run.
73 */
74 if (pc & HARDIRQ_MASK)
75 irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
76 else if ((pc & SOFTIRQ_OFFSET) && curr != this_cpu_ksoftirqd())
77 irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
78}
79
80static u64 irqtime_tick_accounted(u64 maxtime)
81{
82 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
83 u64 delta;
84
85 delta = min(irqtime->tick_delta, maxtime);
86 irqtime->tick_delta -= delta;
87
88 return delta;
89}
90
91#else /* CONFIG_IRQ_TIME_ACCOUNTING */
92
93#define sched_clock_irqtime (0)
94
95static u64 irqtime_tick_accounted(u64 dummy)
96{
97 return 0;
98}
99
100#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
101
102static inline void task_group_account_field(struct task_struct *p, int index,
103 u64 tmp)
104{
105 /*
106 * Since all updates are sure to touch the root cgroup, we
107 * get ourselves ahead and touch it first. If the root cgroup
108 * is the only cgroup, then nothing else should be necessary.
109 *
110 */
111 __this_cpu_add(kernel_cpustat.cpustat[index], tmp);
112
113 cgroup_account_cputime_field(p, index, tmp);
114}
115
116/*
117 * Account user CPU time to a process.
118 * @p: the process that the CPU time gets accounted to
119 * @cputime: the CPU time spent in user space since the last update
120 */
121void account_user_time(struct task_struct *p, u64 cputime)
122{
123 int index;
124
125 /* Add user time to process. */
126 p->utime += cputime;
127 account_group_user_time(p, cputime);
128
129 index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
130
131 /* Add user time to cpustat. */
132 task_group_account_field(p, index, cputime);
133
134 /* Account for user time used */
135 acct_account_cputime(p);
136}
137
138/*
139 * Account guest CPU time to a process.
140 * @p: the process that the CPU time gets accounted to
141 * @cputime: the CPU time spent in virtual machine since the last update
142 */
143void account_guest_time(struct task_struct *p, u64 cputime)
144{
145 u64 *cpustat = kcpustat_this_cpu->cpustat;
146
147 /* Add guest time to process. */
148 p->utime += cputime;
149 account_group_user_time(p, cputime);
150 p->gtime += cputime;
151
152 /* Add guest time to cpustat. */
153 if (task_nice(p) > 0) {
154 task_group_account_field(p, CPUTIME_NICE, cputime);
155 cpustat[CPUTIME_GUEST_NICE] += cputime;
156 } else {
157 task_group_account_field(p, CPUTIME_USER, cputime);
158 cpustat[CPUTIME_GUEST] += cputime;
159 }
160}
161
162/*
163 * Account system CPU time to a process and desired cpustat field
164 * @p: the process that the CPU time gets accounted to
165 * @cputime: the CPU time spent in kernel space since the last update
166 * @index: pointer to cpustat field that has to be updated
167 */
168void account_system_index_time(struct task_struct *p,
169 u64 cputime, enum cpu_usage_stat index)
170{
171 /* Add system time to process. */
172 p->stime += cputime;
173 account_group_system_time(p, cputime);
174
175 /* Add system time to cpustat. */
176 task_group_account_field(p, index, cputime);
177
178 /* Account for system time used */
179 acct_account_cputime(p);
180}
181
182/*
183 * Account system CPU time to a process.
184 * @p: the process that the CPU time gets accounted to
185 * @hardirq_offset: the offset to subtract from hardirq_count()
186 * @cputime: the CPU time spent in kernel space since the last update
187 */
188void account_system_time(struct task_struct *p, int hardirq_offset, u64 cputime)
189{
190 int index;
191
192 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
193 account_guest_time(p, cputime);
194 return;
195 }
196
197 if (hardirq_count() - hardirq_offset)
198 index = CPUTIME_IRQ;
199 else if (in_serving_softirq())
200 index = CPUTIME_SOFTIRQ;
201 else
202 index = CPUTIME_SYSTEM;
203
204 account_system_index_time(p, cputime, index);
205}
206
207/*
208 * Account for involuntary wait time.
209 * @cputime: the CPU time spent in involuntary wait
210 */
211void account_steal_time(u64 cputime)
212{
213 u64 *cpustat = kcpustat_this_cpu->cpustat;
214
215 cpustat[CPUTIME_STEAL] += cputime;
216}
217
218/*
219 * Account for idle time.
220 * @cputime: the CPU time spent in idle wait
221 */
222void account_idle_time(u64 cputime)
223{
224 u64 *cpustat = kcpustat_this_cpu->cpustat;
225 struct rq *rq = this_rq();
226
227 if (atomic_read(&rq->nr_iowait) > 0)
228 cpustat[CPUTIME_IOWAIT] += cputime;
229 else
230 cpustat[CPUTIME_IDLE] += cputime;
231}
232
233
234#ifdef CONFIG_SCHED_CORE
235/*
236 * Account for forceidle time due to core scheduling.
237 *
238 * REQUIRES: schedstat is enabled.
239 */
240void __account_forceidle_time(struct task_struct *p, u64 delta)
241{
242 __schedstat_add(p->stats.core_forceidle_sum, delta);
243
244 task_group_account_field(p, CPUTIME_FORCEIDLE, delta);
245}
246#endif
247
248/*
249 * When a guest is interrupted for a longer amount of time, missed clock
250 * ticks are not redelivered later. Due to that, this function may on
251 * occasion account more time than the calling functions think elapsed.
252 */
253static __always_inline u64 steal_account_process_time(u64 maxtime)
254{
255#ifdef CONFIG_PARAVIRT
256 if (static_key_false(¶virt_steal_enabled)) {
257 u64 steal;
258
259 steal = paravirt_steal_clock(smp_processor_id());
260 steal -= this_rq()->prev_steal_time;
261 steal = min(steal, maxtime);
262 account_steal_time(steal);
263 this_rq()->prev_steal_time += steal;
264
265 return steal;
266 }
267#endif
268 return 0;
269}
270
271/*
272 * Account how much elapsed time was spent in steal, irq, or softirq time.
273 */
274static inline u64 account_other_time(u64 max)
275{
276 u64 accounted;
277
278 lockdep_assert_irqs_disabled();
279
280 accounted = steal_account_process_time(max);
281
282 if (accounted < max)
283 accounted += irqtime_tick_accounted(max - accounted);
284
285 return accounted;
286}
287
288#ifdef CONFIG_64BIT
289static inline u64 read_sum_exec_runtime(struct task_struct *t)
290{
291 return t->se.sum_exec_runtime;
292}
293#else
294static u64 read_sum_exec_runtime(struct task_struct *t)
295{
296 u64 ns;
297 struct rq_flags rf;
298 struct rq *rq;
299
300 rq = task_rq_lock(t, &rf);
301 ns = t->se.sum_exec_runtime;
302 task_rq_unlock(rq, t, &rf);
303
304 return ns;
305}
306#endif
307
308/*
309 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
310 * tasks (sum on group iteration) belonging to @tsk's group.
311 */
312void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
313{
314 struct signal_struct *sig = tsk->signal;
315 u64 utime, stime;
316 struct task_struct *t;
317 unsigned int seq, nextseq;
318 unsigned long flags;
319
320 /*
321 * Update current task runtime to account pending time since last
322 * scheduler action or thread_group_cputime() call. This thread group
323 * might have other running tasks on different CPUs, but updating
324 * their runtime can affect syscall performance, so we skip account
325 * those pending times and rely only on values updated on tick or
326 * other scheduler action.
327 */
328 if (same_thread_group(current, tsk))
329 (void) task_sched_runtime(current);
330
331 rcu_read_lock();
332 /* Attempt a lockless read on the first round. */
333 nextseq = 0;
334 do {
335 seq = nextseq;
336 flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
337 times->utime = sig->utime;
338 times->stime = sig->stime;
339 times->sum_exec_runtime = sig->sum_sched_runtime;
340
341 for_each_thread(tsk, t) {
342 task_cputime(t, &utime, &stime);
343 times->utime += utime;
344 times->stime += stime;
345 times->sum_exec_runtime += read_sum_exec_runtime(t);
346 }
347 /* If lockless access failed, take the lock. */
348 nextseq = 1;
349 } while (need_seqretry(&sig->stats_lock, seq));
350 done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
351 rcu_read_unlock();
352}
353
354#ifdef CONFIG_IRQ_TIME_ACCOUNTING
355/*
356 * Account a tick to a process and cpustat
357 * @p: the process that the CPU time gets accounted to
358 * @user_tick: is the tick from userspace
359 * @rq: the pointer to rq
360 *
361 * Tick demultiplexing follows the order
362 * - pending hardirq update
363 * - pending softirq update
364 * - user_time
365 * - idle_time
366 * - system time
367 * - check for guest_time
368 * - else account as system_time
369 *
370 * Check for hardirq is done both for system and user time as there is
371 * no timer going off while we are on hardirq and hence we may never get an
372 * opportunity to update it solely in system time.
373 * p->stime and friends are only updated on system time and not on irq
374 * softirq as those do not count in task exec_runtime any more.
375 */
376static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
377 int ticks)
378{
379 u64 other, cputime = TICK_NSEC * ticks;
380
381 /*
382 * When returning from idle, many ticks can get accounted at
383 * once, including some ticks of steal, irq, and softirq time.
384 * Subtract those ticks from the amount of time accounted to
385 * idle, or potentially user or system time. Due to rounding,
386 * other time can exceed ticks occasionally.
387 */
388 other = account_other_time(ULONG_MAX);
389 if (other >= cputime)
390 return;
391
392 cputime -= other;
393
394 if (this_cpu_ksoftirqd() == p) {
395 /*
396 * ksoftirqd time do not get accounted in cpu_softirq_time.
397 * So, we have to handle it separately here.
398 * Also, p->stime needs to be updated for ksoftirqd.
399 */
400 account_system_index_time(p, cputime, CPUTIME_SOFTIRQ);
401 } else if (user_tick) {
402 account_user_time(p, cputime);
403 } else if (p == this_rq()->idle) {
404 account_idle_time(cputime);
405 } else if (p->flags & PF_VCPU) { /* System time or guest time */
406 account_guest_time(p, cputime);
407 } else {
408 account_system_index_time(p, cputime, CPUTIME_SYSTEM);
409 }
410}
411
412static void irqtime_account_idle_ticks(int ticks)
413{
414 irqtime_account_process_tick(current, 0, ticks);
415}
416#else /* CONFIG_IRQ_TIME_ACCOUNTING */
417static inline void irqtime_account_idle_ticks(int ticks) { }
418static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
419 int nr_ticks) { }
420#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
421
422/*
423 * Use precise platform statistics if available:
424 */
425#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
426
427# ifndef __ARCH_HAS_VTIME_TASK_SWITCH
428void vtime_task_switch(struct task_struct *prev)
429{
430 if (is_idle_task(prev))
431 vtime_account_idle(prev);
432 else
433 vtime_account_kernel(prev);
434
435 vtime_flush(prev);
436 arch_vtime_task_switch(prev);
437}
438# endif
439
440void vtime_account_irq(struct task_struct *tsk, unsigned int offset)
441{
442 unsigned int pc = irq_count() - offset;
443
444 if (pc & HARDIRQ_OFFSET) {
445 vtime_account_hardirq(tsk);
446 } else if (pc & SOFTIRQ_OFFSET) {
447 vtime_account_softirq(tsk);
448 } else if (!IS_ENABLED(CONFIG_HAVE_VIRT_CPU_ACCOUNTING_IDLE) &&
449 is_idle_task(tsk)) {
450 vtime_account_idle(tsk);
451 } else {
452 vtime_account_kernel(tsk);
453 }
454}
455
456void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
457 u64 *ut, u64 *st)
458{
459 *ut = curr->utime;
460 *st = curr->stime;
461}
462
463void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
464{
465 *ut = p->utime;
466 *st = p->stime;
467}
468EXPORT_SYMBOL_GPL(task_cputime_adjusted);
469
470void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
471{
472 struct task_cputime cputime;
473
474 thread_group_cputime(p, &cputime);
475
476 *ut = cputime.utime;
477 *st = cputime.stime;
478}
479
480#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE: */
481
482/*
483 * Account a single tick of CPU time.
484 * @p: the process that the CPU time gets accounted to
485 * @user_tick: indicates if the tick is a user or a system tick
486 */
487void account_process_tick(struct task_struct *p, int user_tick)
488{
489 u64 cputime, steal;
490
491 if (vtime_accounting_enabled_this_cpu())
492 return;
493
494 if (sched_clock_irqtime) {
495 irqtime_account_process_tick(p, user_tick, 1);
496 return;
497 }
498
499 cputime = TICK_NSEC;
500 steal = steal_account_process_time(ULONG_MAX);
501
502 if (steal >= cputime)
503 return;
504
505 cputime -= steal;
506
507 if (user_tick)
508 account_user_time(p, cputime);
509 else if ((p != this_rq()->idle) || (irq_count() != HARDIRQ_OFFSET))
510 account_system_time(p, HARDIRQ_OFFSET, cputime);
511 else
512 account_idle_time(cputime);
513}
514
515/*
516 * Account multiple ticks of idle time.
517 * @ticks: number of stolen ticks
518 */
519void account_idle_ticks(unsigned long ticks)
520{
521 u64 cputime, steal;
522
523 if (sched_clock_irqtime) {
524 irqtime_account_idle_ticks(ticks);
525 return;
526 }
527
528 cputime = ticks * TICK_NSEC;
529 steal = steal_account_process_time(ULONG_MAX);
530
531 if (steal >= cputime)
532 return;
533
534 cputime -= steal;
535 account_idle_time(cputime);
536}
537
538/*
539 * Adjust tick based cputime random precision against scheduler runtime
540 * accounting.
541 *
542 * Tick based cputime accounting depend on random scheduling timeslices of a
543 * task to be interrupted or not by the timer. Depending on these
544 * circumstances, the number of these interrupts may be over or
545 * under-optimistic, matching the real user and system cputime with a variable
546 * precision.
547 *
548 * Fix this by scaling these tick based values against the total runtime
549 * accounted by the CFS scheduler.
550 *
551 * This code provides the following guarantees:
552 *
553 * stime + utime == rtime
554 * stime_i+1 >= stime_i, utime_i+1 >= utime_i
555 *
556 * Assuming that rtime_i+1 >= rtime_i.
557 */
558void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
559 u64 *ut, u64 *st)
560{
561 u64 rtime, stime, utime;
562 unsigned long flags;
563
564 /* Serialize concurrent callers such that we can honour our guarantees */
565 raw_spin_lock_irqsave(&prev->lock, flags);
566 rtime = curr->sum_exec_runtime;
567
568 /*
569 * This is possible under two circumstances:
570 * - rtime isn't monotonic after all (a bug);
571 * - we got reordered by the lock.
572 *
573 * In both cases this acts as a filter such that the rest of the code
574 * can assume it is monotonic regardless of anything else.
575 */
576 if (prev->stime + prev->utime >= rtime)
577 goto out;
578
579 stime = curr->stime;
580 utime = curr->utime;
581
582 /*
583 * If either stime or utime are 0, assume all runtime is userspace.
584 * Once a task gets some ticks, the monotonicity code at 'update:'
585 * will ensure things converge to the observed ratio.
586 */
587 if (stime == 0) {
588 utime = rtime;
589 goto update;
590 }
591
592 if (utime == 0) {
593 stime = rtime;
594 goto update;
595 }
596
597 stime = mul_u64_u64_div_u64(stime, rtime, stime + utime);
598
599update:
600 /*
601 * Make sure stime doesn't go backwards; this preserves monotonicity
602 * for utime because rtime is monotonic.
603 *
604 * utime_i+1 = rtime_i+1 - stime_i
605 * = rtime_i+1 - (rtime_i - utime_i)
606 * = (rtime_i+1 - rtime_i) + utime_i
607 * >= utime_i
608 */
609 if (stime < prev->stime)
610 stime = prev->stime;
611 utime = rtime - stime;
612
613 /*
614 * Make sure utime doesn't go backwards; this still preserves
615 * monotonicity for stime, analogous argument to above.
616 */
617 if (utime < prev->utime) {
618 utime = prev->utime;
619 stime = rtime - utime;
620 }
621
622 prev->stime = stime;
623 prev->utime = utime;
624out:
625 *ut = prev->utime;
626 *st = prev->stime;
627 raw_spin_unlock_irqrestore(&prev->lock, flags);
628}
629
630void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
631{
632 struct task_cputime cputime = {
633 .sum_exec_runtime = p->se.sum_exec_runtime,
634 };
635
636 if (task_cputime(p, &cputime.utime, &cputime.stime))
637 cputime.sum_exec_runtime = task_sched_runtime(p);
638 cputime_adjust(&cputime, &p->prev_cputime, ut, st);
639}
640EXPORT_SYMBOL_GPL(task_cputime_adjusted);
641
642void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
643{
644 struct task_cputime cputime;
645
646 thread_group_cputime(p, &cputime);
647 cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
648}
649#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
650
651#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
652static u64 vtime_delta(struct vtime *vtime)
653{
654 unsigned long long clock;
655
656 clock = sched_clock();
657 if (clock < vtime->starttime)
658 return 0;
659
660 return clock - vtime->starttime;
661}
662
663static u64 get_vtime_delta(struct vtime *vtime)
664{
665 u64 delta = vtime_delta(vtime);
666 u64 other;
667
668 /*
669 * Unlike tick based timing, vtime based timing never has lost
670 * ticks, and no need for steal time accounting to make up for
671 * lost ticks. Vtime accounts a rounded version of actual
672 * elapsed time. Limit account_other_time to prevent rounding
673 * errors from causing elapsed vtime to go negative.
674 */
675 other = account_other_time(delta);
676 WARN_ON_ONCE(vtime->state == VTIME_INACTIVE);
677 vtime->starttime += delta;
678
679 return delta - other;
680}
681
682static void vtime_account_system(struct task_struct *tsk,
683 struct vtime *vtime)
684{
685 vtime->stime += get_vtime_delta(vtime);
686 if (vtime->stime >= TICK_NSEC) {
687 account_system_time(tsk, irq_count(), vtime->stime);
688 vtime->stime = 0;
689 }
690}
691
692static void vtime_account_guest(struct task_struct *tsk,
693 struct vtime *vtime)
694{
695 vtime->gtime += get_vtime_delta(vtime);
696 if (vtime->gtime >= TICK_NSEC) {
697 account_guest_time(tsk, vtime->gtime);
698 vtime->gtime = 0;
699 }
700}
701
702static void __vtime_account_kernel(struct task_struct *tsk,
703 struct vtime *vtime)
704{
705 /* We might have scheduled out from guest path */
706 if (vtime->state == VTIME_GUEST)
707 vtime_account_guest(tsk, vtime);
708 else
709 vtime_account_system(tsk, vtime);
710}
711
712void vtime_account_kernel(struct task_struct *tsk)
713{
714 struct vtime *vtime = &tsk->vtime;
715
716 if (!vtime_delta(vtime))
717 return;
718
719 write_seqcount_begin(&vtime->seqcount);
720 __vtime_account_kernel(tsk, vtime);
721 write_seqcount_end(&vtime->seqcount);
722}
723
724void vtime_user_enter(struct task_struct *tsk)
725{
726 struct vtime *vtime = &tsk->vtime;
727
728 write_seqcount_begin(&vtime->seqcount);
729 vtime_account_system(tsk, vtime);
730 vtime->state = VTIME_USER;
731 write_seqcount_end(&vtime->seqcount);
732}
733
734void vtime_user_exit(struct task_struct *tsk)
735{
736 struct vtime *vtime = &tsk->vtime;
737
738 write_seqcount_begin(&vtime->seqcount);
739 vtime->utime += get_vtime_delta(vtime);
740 if (vtime->utime >= TICK_NSEC) {
741 account_user_time(tsk, vtime->utime);
742 vtime->utime = 0;
743 }
744 vtime->state = VTIME_SYS;
745 write_seqcount_end(&vtime->seqcount);
746}
747
748void vtime_guest_enter(struct task_struct *tsk)
749{
750 struct vtime *vtime = &tsk->vtime;
751 /*
752 * The flags must be updated under the lock with
753 * the vtime_starttime flush and update.
754 * That enforces a right ordering and update sequence
755 * synchronization against the reader (task_gtime())
756 * that can thus safely catch up with a tickless delta.
757 */
758 write_seqcount_begin(&vtime->seqcount);
759 vtime_account_system(tsk, vtime);
760 tsk->flags |= PF_VCPU;
761 vtime->state = VTIME_GUEST;
762 write_seqcount_end(&vtime->seqcount);
763}
764EXPORT_SYMBOL_GPL(vtime_guest_enter);
765
766void vtime_guest_exit(struct task_struct *tsk)
767{
768 struct vtime *vtime = &tsk->vtime;
769
770 write_seqcount_begin(&vtime->seqcount);
771 vtime_account_guest(tsk, vtime);
772 tsk->flags &= ~PF_VCPU;
773 vtime->state = VTIME_SYS;
774 write_seqcount_end(&vtime->seqcount);
775}
776EXPORT_SYMBOL_GPL(vtime_guest_exit);
777
778void vtime_account_idle(struct task_struct *tsk)
779{
780 account_idle_time(get_vtime_delta(&tsk->vtime));
781}
782
783void vtime_task_switch_generic(struct task_struct *prev)
784{
785 struct vtime *vtime = &prev->vtime;
786
787 write_seqcount_begin(&vtime->seqcount);
788 if (vtime->state == VTIME_IDLE)
789 vtime_account_idle(prev);
790 else
791 __vtime_account_kernel(prev, vtime);
792 vtime->state = VTIME_INACTIVE;
793 vtime->cpu = -1;
794 write_seqcount_end(&vtime->seqcount);
795
796 vtime = ¤t->vtime;
797
798 write_seqcount_begin(&vtime->seqcount);
799 if (is_idle_task(current))
800 vtime->state = VTIME_IDLE;
801 else if (current->flags & PF_VCPU)
802 vtime->state = VTIME_GUEST;
803 else
804 vtime->state = VTIME_SYS;
805 vtime->starttime = sched_clock();
806 vtime->cpu = smp_processor_id();
807 write_seqcount_end(&vtime->seqcount);
808}
809
810void vtime_init_idle(struct task_struct *t, int cpu)
811{
812 struct vtime *vtime = &t->vtime;
813 unsigned long flags;
814
815 local_irq_save(flags);
816 write_seqcount_begin(&vtime->seqcount);
817 vtime->state = VTIME_IDLE;
818 vtime->starttime = sched_clock();
819 vtime->cpu = cpu;
820 write_seqcount_end(&vtime->seqcount);
821 local_irq_restore(flags);
822}
823
824u64 task_gtime(struct task_struct *t)
825{
826 struct vtime *vtime = &t->vtime;
827 unsigned int seq;
828 u64 gtime;
829
830 if (!vtime_accounting_enabled())
831 return t->gtime;
832
833 do {
834 seq = read_seqcount_begin(&vtime->seqcount);
835
836 gtime = t->gtime;
837 if (vtime->state == VTIME_GUEST)
838 gtime += vtime->gtime + vtime_delta(vtime);
839
840 } while (read_seqcount_retry(&vtime->seqcount, seq));
841
842 return gtime;
843}
844
845/*
846 * Fetch cputime raw values from fields of task_struct and
847 * add up the pending nohz execution time since the last
848 * cputime snapshot.
849 */
850bool task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
851{
852 struct vtime *vtime = &t->vtime;
853 unsigned int seq;
854 u64 delta;
855 int ret;
856
857 if (!vtime_accounting_enabled()) {
858 *utime = t->utime;
859 *stime = t->stime;
860 return false;
861 }
862
863 do {
864 ret = false;
865 seq = read_seqcount_begin(&vtime->seqcount);
866
867 *utime = t->utime;
868 *stime = t->stime;
869
870 /* Task is sleeping or idle, nothing to add */
871 if (vtime->state < VTIME_SYS)
872 continue;
873
874 ret = true;
875 delta = vtime_delta(vtime);
876
877 /*
878 * Task runs either in user (including guest) or kernel space,
879 * add pending nohz time to the right place.
880 */
881 if (vtime->state == VTIME_SYS)
882 *stime += vtime->stime + delta;
883 else
884 *utime += vtime->utime + delta;
885 } while (read_seqcount_retry(&vtime->seqcount, seq));
886
887 return ret;
888}
889
890static int vtime_state_fetch(struct vtime *vtime, int cpu)
891{
892 int state = READ_ONCE(vtime->state);
893
894 /*
895 * We raced against a context switch, fetch the
896 * kcpustat task again.
897 */
898 if (vtime->cpu != cpu && vtime->cpu != -1)
899 return -EAGAIN;
900
901 /*
902 * Two possible things here:
903 * 1) We are seeing the scheduling out task (prev) or any past one.
904 * 2) We are seeing the scheduling in task (next) but it hasn't
905 * passed though vtime_task_switch() yet so the pending
906 * cputime of the prev task may not be flushed yet.
907 *
908 * Case 1) is ok but 2) is not. So wait for a safe VTIME state.
909 */
910 if (state == VTIME_INACTIVE)
911 return -EAGAIN;
912
913 return state;
914}
915
916static u64 kcpustat_user_vtime(struct vtime *vtime)
917{
918 if (vtime->state == VTIME_USER)
919 return vtime->utime + vtime_delta(vtime);
920 else if (vtime->state == VTIME_GUEST)
921 return vtime->gtime + vtime_delta(vtime);
922 return 0;
923}
924
925static int kcpustat_field_vtime(u64 *cpustat,
926 struct task_struct *tsk,
927 enum cpu_usage_stat usage,
928 int cpu, u64 *val)
929{
930 struct vtime *vtime = &tsk->vtime;
931 unsigned int seq;
932
933 do {
934 int state;
935
936 seq = read_seqcount_begin(&vtime->seqcount);
937
938 state = vtime_state_fetch(vtime, cpu);
939 if (state < 0)
940 return state;
941
942 *val = cpustat[usage];
943
944 /*
945 * Nice VS unnice cputime accounting may be inaccurate if
946 * the nice value has changed since the last vtime update.
947 * But proper fix would involve interrupting target on nice
948 * updates which is a no go on nohz_full (although the scheduler
949 * may still interrupt the target if rescheduling is needed...)
950 */
951 switch (usage) {
952 case CPUTIME_SYSTEM:
953 if (state == VTIME_SYS)
954 *val += vtime->stime + vtime_delta(vtime);
955 break;
956 case CPUTIME_USER:
957 if (task_nice(tsk) <= 0)
958 *val += kcpustat_user_vtime(vtime);
959 break;
960 case CPUTIME_NICE:
961 if (task_nice(tsk) > 0)
962 *val += kcpustat_user_vtime(vtime);
963 break;
964 case CPUTIME_GUEST:
965 if (state == VTIME_GUEST && task_nice(tsk) <= 0)
966 *val += vtime->gtime + vtime_delta(vtime);
967 break;
968 case CPUTIME_GUEST_NICE:
969 if (state == VTIME_GUEST && task_nice(tsk) > 0)
970 *val += vtime->gtime + vtime_delta(vtime);
971 break;
972 default:
973 break;
974 }
975 } while (read_seqcount_retry(&vtime->seqcount, seq));
976
977 return 0;
978}
979
980u64 kcpustat_field(struct kernel_cpustat *kcpustat,
981 enum cpu_usage_stat usage, int cpu)
982{
983 u64 *cpustat = kcpustat->cpustat;
984 u64 val = cpustat[usage];
985 struct rq *rq;
986 int err;
987
988 if (!vtime_accounting_enabled_cpu(cpu))
989 return val;
990
991 rq = cpu_rq(cpu);
992
993 for (;;) {
994 struct task_struct *curr;
995
996 rcu_read_lock();
997 curr = rcu_dereference(rq->curr);
998 if (WARN_ON_ONCE(!curr)) {
999 rcu_read_unlock();
1000 return cpustat[usage];
1001 }
1002
1003 err = kcpustat_field_vtime(cpustat, curr, usage, cpu, &val);
1004 rcu_read_unlock();
1005
1006 if (!err)
1007 return val;
1008
1009 cpu_relax();
1010 }
1011}
1012EXPORT_SYMBOL_GPL(kcpustat_field);
1013
1014static int kcpustat_cpu_fetch_vtime(struct kernel_cpustat *dst,
1015 const struct kernel_cpustat *src,
1016 struct task_struct *tsk, int cpu)
1017{
1018 struct vtime *vtime = &tsk->vtime;
1019 unsigned int seq;
1020
1021 do {
1022 u64 *cpustat;
1023 u64 delta;
1024 int state;
1025
1026 seq = read_seqcount_begin(&vtime->seqcount);
1027
1028 state = vtime_state_fetch(vtime, cpu);
1029 if (state < 0)
1030 return state;
1031
1032 *dst = *src;
1033 cpustat = dst->cpustat;
1034
1035 /* Task is sleeping, dead or idle, nothing to add */
1036 if (state < VTIME_SYS)
1037 continue;
1038
1039 delta = vtime_delta(vtime);
1040
1041 /*
1042 * Task runs either in user (including guest) or kernel space,
1043 * add pending nohz time to the right place.
1044 */
1045 if (state == VTIME_SYS) {
1046 cpustat[CPUTIME_SYSTEM] += vtime->stime + delta;
1047 } else if (state == VTIME_USER) {
1048 if (task_nice(tsk) > 0)
1049 cpustat[CPUTIME_NICE] += vtime->utime + delta;
1050 else
1051 cpustat[CPUTIME_USER] += vtime->utime + delta;
1052 } else {
1053 WARN_ON_ONCE(state != VTIME_GUEST);
1054 if (task_nice(tsk) > 0) {
1055 cpustat[CPUTIME_GUEST_NICE] += vtime->gtime + delta;
1056 cpustat[CPUTIME_NICE] += vtime->gtime + delta;
1057 } else {
1058 cpustat[CPUTIME_GUEST] += vtime->gtime + delta;
1059 cpustat[CPUTIME_USER] += vtime->gtime + delta;
1060 }
1061 }
1062 } while (read_seqcount_retry(&vtime->seqcount, seq));
1063
1064 return 0;
1065}
1066
1067void kcpustat_cpu_fetch(struct kernel_cpustat *dst, int cpu)
1068{
1069 const struct kernel_cpustat *src = &kcpustat_cpu(cpu);
1070 struct rq *rq;
1071 int err;
1072
1073 if (!vtime_accounting_enabled_cpu(cpu)) {
1074 *dst = *src;
1075 return;
1076 }
1077
1078 rq = cpu_rq(cpu);
1079
1080 for (;;) {
1081 struct task_struct *curr;
1082
1083 rcu_read_lock();
1084 curr = rcu_dereference(rq->curr);
1085 if (WARN_ON_ONCE(!curr)) {
1086 rcu_read_unlock();
1087 *dst = *src;
1088 return;
1089 }
1090
1091 err = kcpustat_cpu_fetch_vtime(dst, src, curr, cpu);
1092 rcu_read_unlock();
1093
1094 if (!err)
1095 return;
1096
1097 cpu_relax();
1098 }
1099}
1100EXPORT_SYMBOL_GPL(kcpustat_cpu_fetch);
1101
1102#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */