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#include "sched.h"
6
7#ifdef CONFIG_IRQ_TIME_ACCOUNTING
8
9/*
10 * There are no locks covering percpu hardirq/softirq time.
11 * They are only modified in vtime_account, on corresponding CPU
12 * with interrupts disabled. So, writes are safe.
13 * They are read and saved off onto struct rq in update_rq_clock().
14 * This may result in other CPU reading this CPU's irq time and can
15 * race with irq/vtime_account on this CPU. We would either get old
16 * or new value with a side effect of accounting a slice of irq time to wrong
17 * task when irq is in progress while we read rq->clock. That is a worthy
18 * compromise in place of having locks on each irq in account_system_time.
19 */
20DEFINE_PER_CPU(struct irqtime, cpu_irqtime);
21
22static int sched_clock_irqtime;
23
24void enable_sched_clock_irqtime(void)
25{
26 sched_clock_irqtime = 1;
27}
28
29void disable_sched_clock_irqtime(void)
30{
31 sched_clock_irqtime = 0;
32}
33
34static void irqtime_account_delta(struct irqtime *irqtime, u64 delta,
35 enum cpu_usage_stat idx)
36{
37 u64 *cpustat = kcpustat_this_cpu->cpustat;
38
39 u64_stats_update_begin(&irqtime->sync);
40 cpustat[idx] += delta;
41 irqtime->total += delta;
42 irqtime->tick_delta += delta;
43 u64_stats_update_end(&irqtime->sync);
44}
45
46/*
47 * Called before incrementing preempt_count on {soft,}irq_enter
48 * and before decrementing preempt_count on {soft,}irq_exit.
49 */
50void irqtime_account_irq(struct task_struct *curr)
51{
52 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
53 s64 delta;
54 int cpu;
55
56 if (!sched_clock_irqtime)
57 return;
58
59 cpu = smp_processor_id();
60 delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
61 irqtime->irq_start_time += delta;
62
63 /*
64 * We do not account for softirq time from ksoftirqd here.
65 * We want to continue accounting softirq time to ksoftirqd thread
66 * in that case, so as not to confuse scheduler with a special task
67 * that do not consume any time, but still wants to run.
68 */
69 if (hardirq_count())
70 irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
71 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
72 irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
73}
74EXPORT_SYMBOL_GPL(irqtime_account_irq);
75
76static u64 irqtime_tick_accounted(u64 maxtime)
77{
78 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
79 u64 delta;
80
81 delta = min(irqtime->tick_delta, maxtime);
82 irqtime->tick_delta -= delta;
83
84 return delta;
85}
86
87#else /* CONFIG_IRQ_TIME_ACCOUNTING */
88
89#define sched_clock_irqtime (0)
90
91static u64 irqtime_tick_accounted(u64 dummy)
92{
93 return 0;
94}
95
96#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
97
98static inline void task_group_account_field(struct task_struct *p, int index,
99 u64 tmp)
100{
101 /*
102 * Since all updates are sure to touch the root cgroup, we
103 * get ourselves ahead and touch it first. If the root cgroup
104 * is the only cgroup, then nothing else should be necessary.
105 *
106 */
107 __this_cpu_add(kernel_cpustat.cpustat[index], tmp);
108
109 cgroup_account_cputime_field(p, index, tmp);
110}
111
112/*
113 * Account user CPU time to a process.
114 * @p: the process that the CPU time gets accounted to
115 * @cputime: the CPU time spent in user space since the last update
116 */
117void account_user_time(struct task_struct *p, u64 cputime)
118{
119 int index;
120
121 /* Add user time to process. */
122 p->utime += cputime;
123 account_group_user_time(p, cputime);
124
125 index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
126
127 /* Add user time to cpustat. */
128 task_group_account_field(p, index, cputime);
129
130 /* Account for user time used */
131 acct_account_cputime(p);
132}
133
134/*
135 * Account guest CPU time to a process.
136 * @p: the process that the CPU time gets accounted to
137 * @cputime: the CPU time spent in virtual machine since the last update
138 */
139void account_guest_time(struct task_struct *p, u64 cputime)
140{
141 u64 *cpustat = kcpustat_this_cpu->cpustat;
142
143 /* Add guest time to process. */
144 p->utime += cputime;
145 account_group_user_time(p, cputime);
146 p->gtime += cputime;
147
148 /* Add guest time to cpustat. */
149 if (task_nice(p) > 0) {
150 cpustat[CPUTIME_NICE] += cputime;
151 cpustat[CPUTIME_GUEST_NICE] += cputime;
152 } else {
153 cpustat[CPUTIME_USER] += cputime;
154 cpustat[CPUTIME_GUEST] += cputime;
155 }
156}
157
158/*
159 * Account system CPU time to a process and desired cpustat field
160 * @p: the process that the CPU time gets accounted to
161 * @cputime: the CPU time spent in kernel space since the last update
162 * @index: pointer to cpustat field that has to be updated
163 */
164void account_system_index_time(struct task_struct *p,
165 u64 cputime, enum cpu_usage_stat index)
166{
167 /* Add system time to process. */
168 p->stime += cputime;
169 account_group_system_time(p, cputime);
170
171 /* Add system time to cpustat. */
172 task_group_account_field(p, index, cputime);
173
174 /* Account for system time used */
175 acct_account_cputime(p);
176}
177
178/*
179 * Account system CPU time to a process.
180 * @p: the process that the CPU time gets accounted to
181 * @hardirq_offset: the offset to subtract from hardirq_count()
182 * @cputime: the CPU time spent in kernel space since the last update
183 */
184void account_system_time(struct task_struct *p, int hardirq_offset, u64 cputime)
185{
186 int index;
187
188 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
189 account_guest_time(p, cputime);
190 return;
191 }
192
193 if (hardirq_count() - hardirq_offset)
194 index = CPUTIME_IRQ;
195 else if (in_serving_softirq())
196 index = CPUTIME_SOFTIRQ;
197 else
198 index = CPUTIME_SYSTEM;
199
200 account_system_index_time(p, cputime, index);
201}
202
203/*
204 * Account for involuntary wait time.
205 * @cputime: the CPU time spent in involuntary wait
206 */
207void account_steal_time(u64 cputime)
208{
209 u64 *cpustat = kcpustat_this_cpu->cpustat;
210
211 cpustat[CPUTIME_STEAL] += cputime;
212}
213
214/*
215 * Account for idle time.
216 * @cputime: the CPU time spent in idle wait
217 */
218void account_idle_time(u64 cputime)
219{
220 u64 *cpustat = kcpustat_this_cpu->cpustat;
221 struct rq *rq = this_rq();
222
223 if (atomic_read(&rq->nr_iowait) > 0)
224 cpustat[CPUTIME_IOWAIT] += cputime;
225 else
226 cpustat[CPUTIME_IDLE] += cputime;
227}
228
229/*
230 * When a guest is interrupted for a longer amount of time, missed clock
231 * ticks are not redelivered later. Due to that, this function may on
232 * occasion account more time than the calling functions think elapsed.
233 */
234static __always_inline u64 steal_account_process_time(u64 maxtime)
235{
236#ifdef CONFIG_PARAVIRT
237 if (static_key_false(¶virt_steal_enabled)) {
238 u64 steal;
239
240 steal = paravirt_steal_clock(smp_processor_id());
241 steal -= this_rq()->prev_steal_time;
242 steal = min(steal, maxtime);
243 account_steal_time(steal);
244 this_rq()->prev_steal_time += steal;
245
246 return steal;
247 }
248#endif
249 return 0;
250}
251
252/*
253 * Account how much elapsed time was spent in steal, irq, or softirq time.
254 */
255static inline u64 account_other_time(u64 max)
256{
257 u64 accounted;
258
259 lockdep_assert_irqs_disabled();
260
261 accounted = steal_account_process_time(max);
262
263 if (accounted < max)
264 accounted += irqtime_tick_accounted(max - accounted);
265
266 return accounted;
267}
268
269#ifdef CONFIG_64BIT
270static inline u64 read_sum_exec_runtime(struct task_struct *t)
271{
272 return t->se.sum_exec_runtime;
273}
274#else
275static u64 read_sum_exec_runtime(struct task_struct *t)
276{
277 u64 ns;
278 struct rq_flags rf;
279 struct rq *rq;
280
281 rq = task_rq_lock(t, &rf);
282 ns = t->se.sum_exec_runtime;
283 task_rq_unlock(rq, t, &rf);
284
285 return ns;
286}
287#endif
288
289/*
290 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
291 * tasks (sum on group iteration) belonging to @tsk's group.
292 */
293void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
294{
295 struct signal_struct *sig = tsk->signal;
296 u64 utime, stime;
297 struct task_struct *t;
298 unsigned int seq, nextseq;
299 unsigned long flags;
300
301 /*
302 * Update current task runtime to account pending time since last
303 * scheduler action or thread_group_cputime() call. This thread group
304 * might have other running tasks on different CPUs, but updating
305 * their runtime can affect syscall performance, so we skip account
306 * those pending times and rely only on values updated on tick or
307 * other scheduler action.
308 */
309 if (same_thread_group(current, tsk))
310 (void) task_sched_runtime(current);
311
312 rcu_read_lock();
313 /* Attempt a lockless read on the first round. */
314 nextseq = 0;
315 do {
316 seq = nextseq;
317 flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
318 times->utime = sig->utime;
319 times->stime = sig->stime;
320 times->sum_exec_runtime = sig->sum_sched_runtime;
321
322 for_each_thread(tsk, t) {
323 task_cputime(t, &utime, &stime);
324 times->utime += utime;
325 times->stime += stime;
326 times->sum_exec_runtime += read_sum_exec_runtime(t);
327 }
328 /* If lockless access failed, take the lock. */
329 nextseq = 1;
330 } while (need_seqretry(&sig->stats_lock, seq));
331 done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
332 rcu_read_unlock();
333}
334
335#ifdef CONFIG_IRQ_TIME_ACCOUNTING
336/*
337 * Account a tick to a process and cpustat
338 * @p: the process that the CPU time gets accounted to
339 * @user_tick: is the tick from userspace
340 * @rq: the pointer to rq
341 *
342 * Tick demultiplexing follows the order
343 * - pending hardirq update
344 * - pending softirq update
345 * - user_time
346 * - idle_time
347 * - system time
348 * - check for guest_time
349 * - else account as system_time
350 *
351 * Check for hardirq is done both for system and user time as there is
352 * no timer going off while we are on hardirq and hence we may never get an
353 * opportunity to update it solely in system time.
354 * p->stime and friends are only updated on system time and not on irq
355 * softirq as those do not count in task exec_runtime any more.
356 */
357static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
358 int ticks)
359{
360 u64 other, cputime = TICK_NSEC * ticks;
361
362 /*
363 * When returning from idle, many ticks can get accounted at
364 * once, including some ticks of steal, irq, and softirq time.
365 * Subtract those ticks from the amount of time accounted to
366 * idle, or potentially user or system time. Due to rounding,
367 * other time can exceed ticks occasionally.
368 */
369 other = account_other_time(ULONG_MAX);
370 if (other >= cputime)
371 return;
372
373 cputime -= other;
374
375 if (this_cpu_ksoftirqd() == p) {
376 /*
377 * ksoftirqd time do not get accounted in cpu_softirq_time.
378 * So, we have to handle it separately here.
379 * Also, p->stime needs to be updated for ksoftirqd.
380 */
381 account_system_index_time(p, cputime, CPUTIME_SOFTIRQ);
382 } else if (user_tick) {
383 account_user_time(p, cputime);
384 } else if (p == this_rq()->idle) {
385 account_idle_time(cputime);
386 } else if (p->flags & PF_VCPU) { /* System time or guest time */
387 account_guest_time(p, cputime);
388 } else {
389 account_system_index_time(p, cputime, CPUTIME_SYSTEM);
390 }
391}
392
393static void irqtime_account_idle_ticks(int ticks)
394{
395 irqtime_account_process_tick(current, 0, ticks);
396}
397#else /* CONFIG_IRQ_TIME_ACCOUNTING */
398static inline void irqtime_account_idle_ticks(int ticks) { }
399static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
400 int nr_ticks) { }
401#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
402
403/*
404 * Use precise platform statistics if available:
405 */
406#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
407
408# ifndef __ARCH_HAS_VTIME_TASK_SWITCH
409void vtime_task_switch(struct task_struct *prev)
410{
411 if (is_idle_task(prev))
412 vtime_account_idle(prev);
413 else
414 vtime_account_kernel(prev);
415
416 vtime_flush(prev);
417 arch_vtime_task_switch(prev);
418}
419# endif
420
421/*
422 * Archs that account the whole time spent in the idle task
423 * (outside irq) as idle time can rely on this and just implement
424 * vtime_account_kernel() and vtime_account_idle(). Archs that
425 * have other meaning of the idle time (s390 only includes the
426 * time spent by the CPU when it's in low power mode) must override
427 * vtime_account().
428 */
429#ifndef __ARCH_HAS_VTIME_ACCOUNT
430void vtime_account_irq_enter(struct task_struct *tsk)
431{
432 if (!in_interrupt() && is_idle_task(tsk))
433 vtime_account_idle(tsk);
434 else
435 vtime_account_kernel(tsk);
436}
437EXPORT_SYMBOL_GPL(vtime_account_irq_enter);
438#endif /* __ARCH_HAS_VTIME_ACCOUNT */
439
440void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
441 u64 *ut, u64 *st)
442{
443 *ut = curr->utime;
444 *st = curr->stime;
445}
446
447void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
448{
449 *ut = p->utime;
450 *st = p->stime;
451}
452EXPORT_SYMBOL_GPL(task_cputime_adjusted);
453
454void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
455{
456 struct task_cputime cputime;
457
458 thread_group_cputime(p, &cputime);
459
460 *ut = cputime.utime;
461 *st = cputime.stime;
462}
463
464#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE: */
465
466/*
467 * Account a single tick of CPU time.
468 * @p: the process that the CPU time gets accounted to
469 * @user_tick: indicates if the tick is a user or a system tick
470 */
471void account_process_tick(struct task_struct *p, int user_tick)
472{
473 u64 cputime, steal;
474
475 if (vtime_accounting_enabled_this_cpu())
476 return;
477
478 if (sched_clock_irqtime) {
479 irqtime_account_process_tick(p, user_tick, 1);
480 return;
481 }
482
483 cputime = TICK_NSEC;
484 steal = steal_account_process_time(ULONG_MAX);
485
486 if (steal >= cputime)
487 return;
488
489 cputime -= steal;
490
491 if (user_tick)
492 account_user_time(p, cputime);
493 else if ((p != this_rq()->idle) || (irq_count() != HARDIRQ_OFFSET))
494 account_system_time(p, HARDIRQ_OFFSET, cputime);
495 else
496 account_idle_time(cputime);
497}
498
499/*
500 * Account multiple ticks of idle time.
501 * @ticks: number of stolen ticks
502 */
503void account_idle_ticks(unsigned long ticks)
504{
505 u64 cputime, steal;
506
507 if (sched_clock_irqtime) {
508 irqtime_account_idle_ticks(ticks);
509 return;
510 }
511
512 cputime = ticks * TICK_NSEC;
513 steal = steal_account_process_time(ULONG_MAX);
514
515 if (steal >= cputime)
516 return;
517
518 cputime -= steal;
519 account_idle_time(cputime);
520}
521
522/*
523 * Adjust tick based cputime random precision against scheduler runtime
524 * accounting.
525 *
526 * Tick based cputime accounting depend on random scheduling timeslices of a
527 * task to be interrupted or not by the timer. Depending on these
528 * circumstances, the number of these interrupts may be over or
529 * under-optimistic, matching the real user and system cputime with a variable
530 * precision.
531 *
532 * Fix this by scaling these tick based values against the total runtime
533 * accounted by the CFS scheduler.
534 *
535 * This code provides the following guarantees:
536 *
537 * stime + utime == rtime
538 * stime_i+1 >= stime_i, utime_i+1 >= utime_i
539 *
540 * Assuming that rtime_i+1 >= rtime_i.
541 */
542void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
543 u64 *ut, u64 *st)
544{
545 u64 rtime, stime, utime;
546 unsigned long flags;
547
548 /* Serialize concurrent callers such that we can honour our guarantees */
549 raw_spin_lock_irqsave(&prev->lock, flags);
550 rtime = curr->sum_exec_runtime;
551
552 /*
553 * This is possible under two circumstances:
554 * - rtime isn't monotonic after all (a bug);
555 * - we got reordered by the lock.
556 *
557 * In both cases this acts as a filter such that the rest of the code
558 * can assume it is monotonic regardless of anything else.
559 */
560 if (prev->stime + prev->utime >= rtime)
561 goto out;
562
563 stime = curr->stime;
564 utime = curr->utime;
565
566 /*
567 * If either stime or utime are 0, assume all runtime is userspace.
568 * Once a task gets some ticks, the monotonicy code at 'update:'
569 * will ensure things converge to the observed ratio.
570 */
571 if (stime == 0) {
572 utime = rtime;
573 goto update;
574 }
575
576 if (utime == 0) {
577 stime = rtime;
578 goto update;
579 }
580
581 stime = mul_u64_u64_div_u64(stime, rtime, stime + utime);
582
583update:
584 /*
585 * Make sure stime doesn't go backwards; this preserves monotonicity
586 * for utime because rtime is monotonic.
587 *
588 * utime_i+1 = rtime_i+1 - stime_i
589 * = rtime_i+1 - (rtime_i - utime_i)
590 * = (rtime_i+1 - rtime_i) + utime_i
591 * >= utime_i
592 */
593 if (stime < prev->stime)
594 stime = prev->stime;
595 utime = rtime - stime;
596
597 /*
598 * Make sure utime doesn't go backwards; this still preserves
599 * monotonicity for stime, analogous argument to above.
600 */
601 if (utime < prev->utime) {
602 utime = prev->utime;
603 stime = rtime - utime;
604 }
605
606 prev->stime = stime;
607 prev->utime = utime;
608out:
609 *ut = prev->utime;
610 *st = prev->stime;
611 raw_spin_unlock_irqrestore(&prev->lock, flags);
612}
613
614void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
615{
616 struct task_cputime cputime = {
617 .sum_exec_runtime = p->se.sum_exec_runtime,
618 };
619
620 task_cputime(p, &cputime.utime, &cputime.stime);
621 cputime_adjust(&cputime, &p->prev_cputime, ut, st);
622}
623EXPORT_SYMBOL_GPL(task_cputime_adjusted);
624
625void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
626{
627 struct task_cputime cputime;
628
629 thread_group_cputime(p, &cputime);
630 cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
631}
632#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
633
634#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
635static u64 vtime_delta(struct vtime *vtime)
636{
637 unsigned long long clock;
638
639 clock = sched_clock();
640 if (clock < vtime->starttime)
641 return 0;
642
643 return clock - vtime->starttime;
644}
645
646static u64 get_vtime_delta(struct vtime *vtime)
647{
648 u64 delta = vtime_delta(vtime);
649 u64 other;
650
651 /*
652 * Unlike tick based timing, vtime based timing never has lost
653 * ticks, and no need for steal time accounting to make up for
654 * lost ticks. Vtime accounts a rounded version of actual
655 * elapsed time. Limit account_other_time to prevent rounding
656 * errors from causing elapsed vtime to go negative.
657 */
658 other = account_other_time(delta);
659 WARN_ON_ONCE(vtime->state == VTIME_INACTIVE);
660 vtime->starttime += delta;
661
662 return delta - other;
663}
664
665static void vtime_account_system(struct task_struct *tsk,
666 struct vtime *vtime)
667{
668 vtime->stime += get_vtime_delta(vtime);
669 if (vtime->stime >= TICK_NSEC) {
670 account_system_time(tsk, irq_count(), vtime->stime);
671 vtime->stime = 0;
672 }
673}
674
675static void vtime_account_guest(struct task_struct *tsk,
676 struct vtime *vtime)
677{
678 vtime->gtime += get_vtime_delta(vtime);
679 if (vtime->gtime >= TICK_NSEC) {
680 account_guest_time(tsk, vtime->gtime);
681 vtime->gtime = 0;
682 }
683}
684
685static void __vtime_account_kernel(struct task_struct *tsk,
686 struct vtime *vtime)
687{
688 /* We might have scheduled out from guest path */
689 if (vtime->state == VTIME_GUEST)
690 vtime_account_guest(tsk, vtime);
691 else
692 vtime_account_system(tsk, vtime);
693}
694
695void vtime_account_kernel(struct task_struct *tsk)
696{
697 struct vtime *vtime = &tsk->vtime;
698
699 if (!vtime_delta(vtime))
700 return;
701
702 write_seqcount_begin(&vtime->seqcount);
703 __vtime_account_kernel(tsk, vtime);
704 write_seqcount_end(&vtime->seqcount);
705}
706
707void vtime_user_enter(struct task_struct *tsk)
708{
709 struct vtime *vtime = &tsk->vtime;
710
711 write_seqcount_begin(&vtime->seqcount);
712 vtime_account_system(tsk, vtime);
713 vtime->state = VTIME_USER;
714 write_seqcount_end(&vtime->seqcount);
715}
716
717void vtime_user_exit(struct task_struct *tsk)
718{
719 struct vtime *vtime = &tsk->vtime;
720
721 write_seqcount_begin(&vtime->seqcount);
722 vtime->utime += get_vtime_delta(vtime);
723 if (vtime->utime >= TICK_NSEC) {
724 account_user_time(tsk, vtime->utime);
725 vtime->utime = 0;
726 }
727 vtime->state = VTIME_SYS;
728 write_seqcount_end(&vtime->seqcount);
729}
730
731void vtime_guest_enter(struct task_struct *tsk)
732{
733 struct vtime *vtime = &tsk->vtime;
734 /*
735 * The flags must be updated under the lock with
736 * the vtime_starttime flush and update.
737 * That enforces a right ordering and update sequence
738 * synchronization against the reader (task_gtime())
739 * that can thus safely catch up with a tickless delta.
740 */
741 write_seqcount_begin(&vtime->seqcount);
742 vtime_account_system(tsk, vtime);
743 tsk->flags |= PF_VCPU;
744 vtime->state = VTIME_GUEST;
745 write_seqcount_end(&vtime->seqcount);
746}
747EXPORT_SYMBOL_GPL(vtime_guest_enter);
748
749void vtime_guest_exit(struct task_struct *tsk)
750{
751 struct vtime *vtime = &tsk->vtime;
752
753 write_seqcount_begin(&vtime->seqcount);
754 vtime_account_guest(tsk, vtime);
755 tsk->flags &= ~PF_VCPU;
756 vtime->state = VTIME_SYS;
757 write_seqcount_end(&vtime->seqcount);
758}
759EXPORT_SYMBOL_GPL(vtime_guest_exit);
760
761void vtime_account_idle(struct task_struct *tsk)
762{
763 account_idle_time(get_vtime_delta(&tsk->vtime));
764}
765
766void vtime_task_switch_generic(struct task_struct *prev)
767{
768 struct vtime *vtime = &prev->vtime;
769
770 write_seqcount_begin(&vtime->seqcount);
771 if (vtime->state == VTIME_IDLE)
772 vtime_account_idle(prev);
773 else
774 __vtime_account_kernel(prev, vtime);
775 vtime->state = VTIME_INACTIVE;
776 vtime->cpu = -1;
777 write_seqcount_end(&vtime->seqcount);
778
779 vtime = ¤t->vtime;
780
781 write_seqcount_begin(&vtime->seqcount);
782 if (is_idle_task(current))
783 vtime->state = VTIME_IDLE;
784 else if (current->flags & PF_VCPU)
785 vtime->state = VTIME_GUEST;
786 else
787 vtime->state = VTIME_SYS;
788 vtime->starttime = sched_clock();
789 vtime->cpu = smp_processor_id();
790 write_seqcount_end(&vtime->seqcount);
791}
792
793void vtime_init_idle(struct task_struct *t, int cpu)
794{
795 struct vtime *vtime = &t->vtime;
796 unsigned long flags;
797
798 local_irq_save(flags);
799 write_seqcount_begin(&vtime->seqcount);
800 vtime->state = VTIME_IDLE;
801 vtime->starttime = sched_clock();
802 vtime->cpu = cpu;
803 write_seqcount_end(&vtime->seqcount);
804 local_irq_restore(flags);
805}
806
807u64 task_gtime(struct task_struct *t)
808{
809 struct vtime *vtime = &t->vtime;
810 unsigned int seq;
811 u64 gtime;
812
813 if (!vtime_accounting_enabled())
814 return t->gtime;
815
816 do {
817 seq = read_seqcount_begin(&vtime->seqcount);
818
819 gtime = t->gtime;
820 if (vtime->state == VTIME_GUEST)
821 gtime += vtime->gtime + vtime_delta(vtime);
822
823 } while (read_seqcount_retry(&vtime->seqcount, seq));
824
825 return gtime;
826}
827
828/*
829 * Fetch cputime raw values from fields of task_struct and
830 * add up the pending nohz execution time since the last
831 * cputime snapshot.
832 */
833void task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
834{
835 struct vtime *vtime = &t->vtime;
836 unsigned int seq;
837 u64 delta;
838
839 if (!vtime_accounting_enabled()) {
840 *utime = t->utime;
841 *stime = t->stime;
842 return;
843 }
844
845 do {
846 seq = read_seqcount_begin(&vtime->seqcount);
847
848 *utime = t->utime;
849 *stime = t->stime;
850
851 /* Task is sleeping or idle, nothing to add */
852 if (vtime->state < VTIME_SYS)
853 continue;
854
855 delta = vtime_delta(vtime);
856
857 /*
858 * Task runs either in user (including guest) or kernel space,
859 * add pending nohz time to the right place.
860 */
861 if (vtime->state == VTIME_SYS)
862 *stime += vtime->stime + delta;
863 else
864 *utime += vtime->utime + delta;
865 } while (read_seqcount_retry(&vtime->seqcount, seq));
866}
867
868static int vtime_state_fetch(struct vtime *vtime, int cpu)
869{
870 int state = READ_ONCE(vtime->state);
871
872 /*
873 * We raced against a context switch, fetch the
874 * kcpustat task again.
875 */
876 if (vtime->cpu != cpu && vtime->cpu != -1)
877 return -EAGAIN;
878
879 /*
880 * Two possible things here:
881 * 1) We are seeing the scheduling out task (prev) or any past one.
882 * 2) We are seeing the scheduling in task (next) but it hasn't
883 * passed though vtime_task_switch() yet so the pending
884 * cputime of the prev task may not be flushed yet.
885 *
886 * Case 1) is ok but 2) is not. So wait for a safe VTIME state.
887 */
888 if (state == VTIME_INACTIVE)
889 return -EAGAIN;
890
891 return state;
892}
893
894static u64 kcpustat_user_vtime(struct vtime *vtime)
895{
896 if (vtime->state == VTIME_USER)
897 return vtime->utime + vtime_delta(vtime);
898 else if (vtime->state == VTIME_GUEST)
899 return vtime->gtime + vtime_delta(vtime);
900 return 0;
901}
902
903static int kcpustat_field_vtime(u64 *cpustat,
904 struct task_struct *tsk,
905 enum cpu_usage_stat usage,
906 int cpu, u64 *val)
907{
908 struct vtime *vtime = &tsk->vtime;
909 unsigned int seq;
910
911 do {
912 int state;
913
914 seq = read_seqcount_begin(&vtime->seqcount);
915
916 state = vtime_state_fetch(vtime, cpu);
917 if (state < 0)
918 return state;
919
920 *val = cpustat[usage];
921
922 /*
923 * Nice VS unnice cputime accounting may be inaccurate if
924 * the nice value has changed since the last vtime update.
925 * But proper fix would involve interrupting target on nice
926 * updates which is a no go on nohz_full (although the scheduler
927 * may still interrupt the target if rescheduling is needed...)
928 */
929 switch (usage) {
930 case CPUTIME_SYSTEM:
931 if (state == VTIME_SYS)
932 *val += vtime->stime + vtime_delta(vtime);
933 break;
934 case CPUTIME_USER:
935 if (task_nice(tsk) <= 0)
936 *val += kcpustat_user_vtime(vtime);
937 break;
938 case CPUTIME_NICE:
939 if (task_nice(tsk) > 0)
940 *val += kcpustat_user_vtime(vtime);
941 break;
942 case CPUTIME_GUEST:
943 if (state == VTIME_GUEST && task_nice(tsk) <= 0)
944 *val += vtime->gtime + vtime_delta(vtime);
945 break;
946 case CPUTIME_GUEST_NICE:
947 if (state == VTIME_GUEST && task_nice(tsk) > 0)
948 *val += vtime->gtime + vtime_delta(vtime);
949 break;
950 default:
951 break;
952 }
953 } while (read_seqcount_retry(&vtime->seqcount, seq));
954
955 return 0;
956}
957
958u64 kcpustat_field(struct kernel_cpustat *kcpustat,
959 enum cpu_usage_stat usage, int cpu)
960{
961 u64 *cpustat = kcpustat->cpustat;
962 u64 val = cpustat[usage];
963 struct rq *rq;
964 int err;
965
966 if (!vtime_accounting_enabled_cpu(cpu))
967 return val;
968
969 rq = cpu_rq(cpu);
970
971 for (;;) {
972 struct task_struct *curr;
973
974 rcu_read_lock();
975 curr = rcu_dereference(rq->curr);
976 if (WARN_ON_ONCE(!curr)) {
977 rcu_read_unlock();
978 return cpustat[usage];
979 }
980
981 err = kcpustat_field_vtime(cpustat, curr, usage, cpu, &val);
982 rcu_read_unlock();
983
984 if (!err)
985 return val;
986
987 cpu_relax();
988 }
989}
990EXPORT_SYMBOL_GPL(kcpustat_field);
991
992static int kcpustat_cpu_fetch_vtime(struct kernel_cpustat *dst,
993 const struct kernel_cpustat *src,
994 struct task_struct *tsk, int cpu)
995{
996 struct vtime *vtime = &tsk->vtime;
997 unsigned int seq;
998
999 do {
1000 u64 *cpustat;
1001 u64 delta;
1002 int state;
1003
1004 seq = read_seqcount_begin(&vtime->seqcount);
1005
1006 state = vtime_state_fetch(vtime, cpu);
1007 if (state < 0)
1008 return state;
1009
1010 *dst = *src;
1011 cpustat = dst->cpustat;
1012
1013 /* Task is sleeping, dead or idle, nothing to add */
1014 if (state < VTIME_SYS)
1015 continue;
1016
1017 delta = vtime_delta(vtime);
1018
1019 /*
1020 * Task runs either in user (including guest) or kernel space,
1021 * add pending nohz time to the right place.
1022 */
1023 if (state == VTIME_SYS) {
1024 cpustat[CPUTIME_SYSTEM] += vtime->stime + delta;
1025 } else if (state == VTIME_USER) {
1026 if (task_nice(tsk) > 0)
1027 cpustat[CPUTIME_NICE] += vtime->utime + delta;
1028 else
1029 cpustat[CPUTIME_USER] += vtime->utime + delta;
1030 } else {
1031 WARN_ON_ONCE(state != VTIME_GUEST);
1032 if (task_nice(tsk) > 0) {
1033 cpustat[CPUTIME_GUEST_NICE] += vtime->gtime + delta;
1034 cpustat[CPUTIME_NICE] += vtime->gtime + delta;
1035 } else {
1036 cpustat[CPUTIME_GUEST] += vtime->gtime + delta;
1037 cpustat[CPUTIME_USER] += vtime->gtime + delta;
1038 }
1039 }
1040 } while (read_seqcount_retry(&vtime->seqcount, seq));
1041
1042 return 0;
1043}
1044
1045void kcpustat_cpu_fetch(struct kernel_cpustat *dst, int cpu)
1046{
1047 const struct kernel_cpustat *src = &kcpustat_cpu(cpu);
1048 struct rq *rq;
1049 int err;
1050
1051 if (!vtime_accounting_enabled_cpu(cpu)) {
1052 *dst = *src;
1053 return;
1054 }
1055
1056 rq = cpu_rq(cpu);
1057
1058 for (;;) {
1059 struct task_struct *curr;
1060
1061 rcu_read_lock();
1062 curr = rcu_dereference(rq->curr);
1063 if (WARN_ON_ONCE(!curr)) {
1064 rcu_read_unlock();
1065 *dst = *src;
1066 return;
1067 }
1068
1069 err = kcpustat_cpu_fetch_vtime(dst, src, curr, cpu);
1070 rcu_read_unlock();
1071
1072 if (!err)
1073 return;
1074
1075 cpu_relax();
1076 }
1077}
1078EXPORT_SYMBOL_GPL(kcpustat_cpu_fetch);
1079
1080#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */