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 */
23 DEFINE_PER_CPU(struct irqtime, cpu_irqtime);
24
25 static int sched_clock_irqtime;
26
enable_sched_clock_irqtime(void)27 void enable_sched_clock_irqtime(void)
28 {
29 sched_clock_irqtime = 1;
30 }
31
disable_sched_clock_irqtime(void)32 void disable_sched_clock_irqtime(void)
33 {
34 sched_clock_irqtime = 0;
35 }
36
irqtime_account_delta(struct irqtime * irqtime,u64 delta,enum cpu_usage_stat idx)37 static 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 */
irqtime_account_irq(struct task_struct * curr,unsigned int offset)53 void 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
irqtime_tick_accounted(u64 maxtime)80 static 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
irqtime_tick_accounted(u64 dummy)95 static u64 irqtime_tick_accounted(u64 dummy)
96 {
97 return 0;
98 }
99
100 #endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
101
task_group_account_field(struct task_struct * p,int index,u64 tmp)102 static 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 */
account_user_time(struct task_struct * p,u64 cputime)121 void 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 */
account_guest_time(struct task_struct * p,u64 cputime)143 void 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 */
account_system_index_time(struct task_struct * p,u64 cputime,enum cpu_usage_stat index)168 void 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 */
account_system_time(struct task_struct * p,int hardirq_offset,u64 cputime)188 void 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 */
account_steal_time(u64 cputime)211 void 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 */
account_idle_time(u64 cputime)222 void 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 */
__account_forceidle_time(struct task_struct * p,u64 delta)240 void __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 */
steal_account_process_time(u64 maxtime)253 static __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 */
account_other_time(u64 max)274 static 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
read_sum_exec_runtime(struct task_struct * t)289 static inline u64 read_sum_exec_runtime(struct task_struct *t)
290 {
291 return t->se.sum_exec_runtime;
292 }
293 #else
read_sum_exec_runtime(struct task_struct * t)294 static 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 */
thread_group_cputime(struct task_struct * tsk,struct task_cputime * times)312 void 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 */
irqtime_account_process_tick(struct task_struct * p,int user_tick,int ticks)376 static 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
irqtime_account_idle_ticks(int ticks)412 static void irqtime_account_idle_ticks(int ticks)
413 {
414 irqtime_account_process_tick(current, 0, ticks);
415 }
416 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
irqtime_account_idle_ticks(int ticks)417 static inline void irqtime_account_idle_ticks(int ticks) { }
irqtime_account_process_tick(struct task_struct * p,int user_tick,int nr_ticks)418 static 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
vtime_account_irq(struct task_struct * tsk,unsigned int offset)427 void vtime_account_irq(struct task_struct *tsk, unsigned int offset)
428 {
429 unsigned int pc = irq_count() - offset;
430
431 if (pc & HARDIRQ_OFFSET) {
432 vtime_account_hardirq(tsk);
433 } else if (pc & SOFTIRQ_OFFSET) {
434 vtime_account_softirq(tsk);
435 } else if (!IS_ENABLED(CONFIG_HAVE_VIRT_CPU_ACCOUNTING_IDLE) &&
436 is_idle_task(tsk)) {
437 vtime_account_idle(tsk);
438 } else {
439 vtime_account_kernel(tsk);
440 }
441 }
442
cputime_adjust(struct task_cputime * curr,struct prev_cputime * prev,u64 * ut,u64 * st)443 void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
444 u64 *ut, u64 *st)
445 {
446 *ut = curr->utime;
447 *st = curr->stime;
448 }
449
task_cputime_adjusted(struct task_struct * p,u64 * ut,u64 * st)450 void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
451 {
452 *ut = p->utime;
453 *st = p->stime;
454 }
455 EXPORT_SYMBOL_GPL(task_cputime_adjusted);
456
thread_group_cputime_adjusted(struct task_struct * p,u64 * ut,u64 * st)457 void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
458 {
459 struct task_cputime cputime;
460
461 thread_group_cputime(p, &cputime);
462
463 *ut = cputime.utime;
464 *st = cputime.stime;
465 }
466
467 #else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE: */
468
469 /*
470 * Account a single tick of CPU time.
471 * @p: the process that the CPU time gets accounted to
472 * @user_tick: indicates if the tick is a user or a system tick
473 */
account_process_tick(struct task_struct * p,int user_tick)474 void account_process_tick(struct task_struct *p, int user_tick)
475 {
476 u64 cputime, steal;
477
478 if (vtime_accounting_enabled_this_cpu())
479 return;
480
481 if (sched_clock_irqtime) {
482 irqtime_account_process_tick(p, user_tick, 1);
483 return;
484 }
485
486 cputime = TICK_NSEC;
487 steal = steal_account_process_time(ULONG_MAX);
488
489 if (steal >= cputime)
490 return;
491
492 cputime -= steal;
493
494 if (user_tick)
495 account_user_time(p, cputime);
496 else if ((p != this_rq()->idle) || (irq_count() != HARDIRQ_OFFSET))
497 account_system_time(p, HARDIRQ_OFFSET, cputime);
498 else
499 account_idle_time(cputime);
500 }
501
502 /*
503 * Account multiple ticks of idle time.
504 * @ticks: number of stolen ticks
505 */
account_idle_ticks(unsigned long ticks)506 void account_idle_ticks(unsigned long ticks)
507 {
508 u64 cputime, steal;
509
510 if (sched_clock_irqtime) {
511 irqtime_account_idle_ticks(ticks);
512 return;
513 }
514
515 cputime = ticks * TICK_NSEC;
516 steal = steal_account_process_time(ULONG_MAX);
517
518 if (steal >= cputime)
519 return;
520
521 cputime -= steal;
522 account_idle_time(cputime);
523 }
524
525 /*
526 * Adjust tick based cputime random precision against scheduler runtime
527 * accounting.
528 *
529 * Tick based cputime accounting depend on random scheduling timeslices of a
530 * task to be interrupted or not by the timer. Depending on these
531 * circumstances, the number of these interrupts may be over or
532 * under-optimistic, matching the real user and system cputime with a variable
533 * precision.
534 *
535 * Fix this by scaling these tick based values against the total runtime
536 * accounted by the CFS scheduler.
537 *
538 * This code provides the following guarantees:
539 *
540 * stime + utime == rtime
541 * stime_i+1 >= stime_i, utime_i+1 >= utime_i
542 *
543 * Assuming that rtime_i+1 >= rtime_i.
544 */
cputime_adjust(struct task_cputime * curr,struct prev_cputime * prev,u64 * ut,u64 * st)545 void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
546 u64 *ut, u64 *st)
547 {
548 u64 rtime, stime, utime;
549 unsigned long flags;
550
551 /* Serialize concurrent callers such that we can honour our guarantees */
552 raw_spin_lock_irqsave(&prev->lock, flags);
553 rtime = curr->sum_exec_runtime;
554
555 /*
556 * This is possible under two circumstances:
557 * - rtime isn't monotonic after all (a bug);
558 * - we got reordered by the lock.
559 *
560 * In both cases this acts as a filter such that the rest of the code
561 * can assume it is monotonic regardless of anything else.
562 */
563 if (prev->stime + prev->utime >= rtime)
564 goto out;
565
566 stime = curr->stime;
567 utime = curr->utime;
568
569 /*
570 * If either stime or utime are 0, assume all runtime is userspace.
571 * Once a task gets some ticks, the monotonicity code at 'update:'
572 * will ensure things converge to the observed ratio.
573 */
574 if (stime == 0) {
575 utime = rtime;
576 goto update;
577 }
578
579 if (utime == 0) {
580 stime = rtime;
581 goto update;
582 }
583
584 stime = mul_u64_u64_div_u64(stime, rtime, stime + utime);
585
586 update:
587 /*
588 * Make sure stime doesn't go backwards; this preserves monotonicity
589 * for utime because rtime is monotonic.
590 *
591 * utime_i+1 = rtime_i+1 - stime_i
592 * = rtime_i+1 - (rtime_i - utime_i)
593 * = (rtime_i+1 - rtime_i) + utime_i
594 * >= utime_i
595 */
596 if (stime < prev->stime)
597 stime = prev->stime;
598 utime = rtime - stime;
599
600 /*
601 * Make sure utime doesn't go backwards; this still preserves
602 * monotonicity for stime, analogous argument to above.
603 */
604 if (utime < prev->utime) {
605 utime = prev->utime;
606 stime = rtime - utime;
607 }
608
609 prev->stime = stime;
610 prev->utime = utime;
611 out:
612 *ut = prev->utime;
613 *st = prev->stime;
614 raw_spin_unlock_irqrestore(&prev->lock, flags);
615 }
616
task_cputime_adjusted(struct task_struct * p,u64 * ut,u64 * st)617 void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
618 {
619 struct task_cputime cputime = {
620 .sum_exec_runtime = p->se.sum_exec_runtime,
621 };
622
623 if (task_cputime(p, &cputime.utime, &cputime.stime))
624 cputime.sum_exec_runtime = task_sched_runtime(p);
625 cputime_adjust(&cputime, &p->prev_cputime, ut, st);
626 }
627 EXPORT_SYMBOL_GPL(task_cputime_adjusted);
628
thread_group_cputime_adjusted(struct task_struct * p,u64 * ut,u64 * st)629 void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
630 {
631 struct task_cputime cputime;
632
633 thread_group_cputime(p, &cputime);
634 cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
635 }
636 #endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
637
638 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
vtime_delta(struct vtime * vtime)639 static u64 vtime_delta(struct vtime *vtime)
640 {
641 unsigned long long clock;
642
643 clock = sched_clock();
644 if (clock < vtime->starttime)
645 return 0;
646
647 return clock - vtime->starttime;
648 }
649
get_vtime_delta(struct vtime * vtime)650 static u64 get_vtime_delta(struct vtime *vtime)
651 {
652 u64 delta = vtime_delta(vtime);
653 u64 other;
654
655 /*
656 * Unlike tick based timing, vtime based timing never has lost
657 * ticks, and no need for steal time accounting to make up for
658 * lost ticks. Vtime accounts a rounded version of actual
659 * elapsed time. Limit account_other_time to prevent rounding
660 * errors from causing elapsed vtime to go negative.
661 */
662 other = account_other_time(delta);
663 WARN_ON_ONCE(vtime->state == VTIME_INACTIVE);
664 vtime->starttime += delta;
665
666 return delta - other;
667 }
668
vtime_account_system(struct task_struct * tsk,struct vtime * vtime)669 static void vtime_account_system(struct task_struct *tsk,
670 struct vtime *vtime)
671 {
672 vtime->stime += get_vtime_delta(vtime);
673 if (vtime->stime >= TICK_NSEC) {
674 account_system_time(tsk, irq_count(), vtime->stime);
675 vtime->stime = 0;
676 }
677 }
678
vtime_account_guest(struct task_struct * tsk,struct vtime * vtime)679 static void vtime_account_guest(struct task_struct *tsk,
680 struct vtime *vtime)
681 {
682 vtime->gtime += get_vtime_delta(vtime);
683 if (vtime->gtime >= TICK_NSEC) {
684 account_guest_time(tsk, vtime->gtime);
685 vtime->gtime = 0;
686 }
687 }
688
__vtime_account_kernel(struct task_struct * tsk,struct vtime * vtime)689 static void __vtime_account_kernel(struct task_struct *tsk,
690 struct vtime *vtime)
691 {
692 /* We might have scheduled out from guest path */
693 if (vtime->state == VTIME_GUEST)
694 vtime_account_guest(tsk, vtime);
695 else
696 vtime_account_system(tsk, vtime);
697 }
698
vtime_account_kernel(struct task_struct * tsk)699 void vtime_account_kernel(struct task_struct *tsk)
700 {
701 struct vtime *vtime = &tsk->vtime;
702
703 if (!vtime_delta(vtime))
704 return;
705
706 write_seqcount_begin(&vtime->seqcount);
707 __vtime_account_kernel(tsk, vtime);
708 write_seqcount_end(&vtime->seqcount);
709 }
710
vtime_user_enter(struct task_struct * tsk)711 void vtime_user_enter(struct task_struct *tsk)
712 {
713 struct vtime *vtime = &tsk->vtime;
714
715 write_seqcount_begin(&vtime->seqcount);
716 vtime_account_system(tsk, vtime);
717 vtime->state = VTIME_USER;
718 write_seqcount_end(&vtime->seqcount);
719 }
720
vtime_user_exit(struct task_struct * tsk)721 void vtime_user_exit(struct task_struct *tsk)
722 {
723 struct vtime *vtime = &tsk->vtime;
724
725 write_seqcount_begin(&vtime->seqcount);
726 vtime->utime += get_vtime_delta(vtime);
727 if (vtime->utime >= TICK_NSEC) {
728 account_user_time(tsk, vtime->utime);
729 vtime->utime = 0;
730 }
731 vtime->state = VTIME_SYS;
732 write_seqcount_end(&vtime->seqcount);
733 }
734
vtime_guest_enter(struct task_struct * tsk)735 void vtime_guest_enter(struct task_struct *tsk)
736 {
737 struct vtime *vtime = &tsk->vtime;
738 /*
739 * The flags must be updated under the lock with
740 * the vtime_starttime flush and update.
741 * That enforces a right ordering and update sequence
742 * synchronization against the reader (task_gtime())
743 * that can thus safely catch up with a tickless delta.
744 */
745 write_seqcount_begin(&vtime->seqcount);
746 vtime_account_system(tsk, vtime);
747 tsk->flags |= PF_VCPU;
748 vtime->state = VTIME_GUEST;
749 write_seqcount_end(&vtime->seqcount);
750 }
751 EXPORT_SYMBOL_GPL(vtime_guest_enter);
752
vtime_guest_exit(struct task_struct * tsk)753 void vtime_guest_exit(struct task_struct *tsk)
754 {
755 struct vtime *vtime = &tsk->vtime;
756
757 write_seqcount_begin(&vtime->seqcount);
758 vtime_account_guest(tsk, vtime);
759 tsk->flags &= ~PF_VCPU;
760 vtime->state = VTIME_SYS;
761 write_seqcount_end(&vtime->seqcount);
762 }
763 EXPORT_SYMBOL_GPL(vtime_guest_exit);
764
vtime_account_idle(struct task_struct * tsk)765 void vtime_account_idle(struct task_struct *tsk)
766 {
767 account_idle_time(get_vtime_delta(&tsk->vtime));
768 }
769
vtime_task_switch_generic(struct task_struct * prev)770 void vtime_task_switch_generic(struct task_struct *prev)
771 {
772 struct vtime *vtime = &prev->vtime;
773
774 write_seqcount_begin(&vtime->seqcount);
775 if (vtime->state == VTIME_IDLE)
776 vtime_account_idle(prev);
777 else
778 __vtime_account_kernel(prev, vtime);
779 vtime->state = VTIME_INACTIVE;
780 vtime->cpu = -1;
781 write_seqcount_end(&vtime->seqcount);
782
783 vtime = ¤t->vtime;
784
785 write_seqcount_begin(&vtime->seqcount);
786 if (is_idle_task(current))
787 vtime->state = VTIME_IDLE;
788 else if (current->flags & PF_VCPU)
789 vtime->state = VTIME_GUEST;
790 else
791 vtime->state = VTIME_SYS;
792 vtime->starttime = sched_clock();
793 vtime->cpu = smp_processor_id();
794 write_seqcount_end(&vtime->seqcount);
795 }
796
vtime_init_idle(struct task_struct * t,int cpu)797 void vtime_init_idle(struct task_struct *t, int cpu)
798 {
799 struct vtime *vtime = &t->vtime;
800 unsigned long flags;
801
802 local_irq_save(flags);
803 write_seqcount_begin(&vtime->seqcount);
804 vtime->state = VTIME_IDLE;
805 vtime->starttime = sched_clock();
806 vtime->cpu = cpu;
807 write_seqcount_end(&vtime->seqcount);
808 local_irq_restore(flags);
809 }
810
task_gtime(struct task_struct * t)811 u64 task_gtime(struct task_struct *t)
812 {
813 struct vtime *vtime = &t->vtime;
814 unsigned int seq;
815 u64 gtime;
816
817 if (!vtime_accounting_enabled())
818 return t->gtime;
819
820 do {
821 seq = read_seqcount_begin(&vtime->seqcount);
822
823 gtime = t->gtime;
824 if (vtime->state == VTIME_GUEST)
825 gtime += vtime->gtime + vtime_delta(vtime);
826
827 } while (read_seqcount_retry(&vtime->seqcount, seq));
828
829 return gtime;
830 }
831
832 /*
833 * Fetch cputime raw values from fields of task_struct and
834 * add up the pending nohz execution time since the last
835 * cputime snapshot.
836 */
task_cputime(struct task_struct * t,u64 * utime,u64 * stime)837 bool task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
838 {
839 struct vtime *vtime = &t->vtime;
840 unsigned int seq;
841 u64 delta;
842 int ret;
843
844 if (!vtime_accounting_enabled()) {
845 *utime = t->utime;
846 *stime = t->stime;
847 return false;
848 }
849
850 do {
851 ret = false;
852 seq = read_seqcount_begin(&vtime->seqcount);
853
854 *utime = t->utime;
855 *stime = t->stime;
856
857 /* Task is sleeping or idle, nothing to add */
858 if (vtime->state < VTIME_SYS)
859 continue;
860
861 ret = true;
862 delta = vtime_delta(vtime);
863
864 /*
865 * Task runs either in user (including guest) or kernel space,
866 * add pending nohz time to the right place.
867 */
868 if (vtime->state == VTIME_SYS)
869 *stime += vtime->stime + delta;
870 else
871 *utime += vtime->utime + delta;
872 } while (read_seqcount_retry(&vtime->seqcount, seq));
873
874 return ret;
875 }
876
vtime_state_fetch(struct vtime * vtime,int cpu)877 static int vtime_state_fetch(struct vtime *vtime, int cpu)
878 {
879 int state = READ_ONCE(vtime->state);
880
881 /*
882 * We raced against a context switch, fetch the
883 * kcpustat task again.
884 */
885 if (vtime->cpu != cpu && vtime->cpu != -1)
886 return -EAGAIN;
887
888 /*
889 * Two possible things here:
890 * 1) We are seeing the scheduling out task (prev) or any past one.
891 * 2) We are seeing the scheduling in task (next) but it hasn't
892 * passed though vtime_task_switch() yet so the pending
893 * cputime of the prev task may not be flushed yet.
894 *
895 * Case 1) is ok but 2) is not. So wait for a safe VTIME state.
896 */
897 if (state == VTIME_INACTIVE)
898 return -EAGAIN;
899
900 return state;
901 }
902
kcpustat_user_vtime(struct vtime * vtime)903 static u64 kcpustat_user_vtime(struct vtime *vtime)
904 {
905 if (vtime->state == VTIME_USER)
906 return vtime->utime + vtime_delta(vtime);
907 else if (vtime->state == VTIME_GUEST)
908 return vtime->gtime + vtime_delta(vtime);
909 return 0;
910 }
911
kcpustat_field_vtime(u64 * cpustat,struct task_struct * tsk,enum cpu_usage_stat usage,int cpu,u64 * val)912 static int kcpustat_field_vtime(u64 *cpustat,
913 struct task_struct *tsk,
914 enum cpu_usage_stat usage,
915 int cpu, u64 *val)
916 {
917 struct vtime *vtime = &tsk->vtime;
918 unsigned int seq;
919
920 do {
921 int state;
922
923 seq = read_seqcount_begin(&vtime->seqcount);
924
925 state = vtime_state_fetch(vtime, cpu);
926 if (state < 0)
927 return state;
928
929 *val = cpustat[usage];
930
931 /*
932 * Nice VS unnice cputime accounting may be inaccurate if
933 * the nice value has changed since the last vtime update.
934 * But proper fix would involve interrupting target on nice
935 * updates which is a no go on nohz_full (although the scheduler
936 * may still interrupt the target if rescheduling is needed...)
937 */
938 switch (usage) {
939 case CPUTIME_SYSTEM:
940 if (state == VTIME_SYS)
941 *val += vtime->stime + vtime_delta(vtime);
942 break;
943 case CPUTIME_USER:
944 if (task_nice(tsk) <= 0)
945 *val += kcpustat_user_vtime(vtime);
946 break;
947 case CPUTIME_NICE:
948 if (task_nice(tsk) > 0)
949 *val += kcpustat_user_vtime(vtime);
950 break;
951 case CPUTIME_GUEST:
952 if (state == VTIME_GUEST && task_nice(tsk) <= 0)
953 *val += vtime->gtime + vtime_delta(vtime);
954 break;
955 case CPUTIME_GUEST_NICE:
956 if (state == VTIME_GUEST && task_nice(tsk) > 0)
957 *val += vtime->gtime + vtime_delta(vtime);
958 break;
959 default:
960 break;
961 }
962 } while (read_seqcount_retry(&vtime->seqcount, seq));
963
964 return 0;
965 }
966
kcpustat_field(struct kernel_cpustat * kcpustat,enum cpu_usage_stat usage,int cpu)967 u64 kcpustat_field(struct kernel_cpustat *kcpustat,
968 enum cpu_usage_stat usage, int cpu)
969 {
970 u64 *cpustat = kcpustat->cpustat;
971 u64 val = cpustat[usage];
972 struct rq *rq;
973 int err;
974
975 if (!vtime_accounting_enabled_cpu(cpu))
976 return val;
977
978 rq = cpu_rq(cpu);
979
980 for (;;) {
981 struct task_struct *curr;
982
983 rcu_read_lock();
984 curr = rcu_dereference(rq->curr);
985 if (WARN_ON_ONCE(!curr)) {
986 rcu_read_unlock();
987 return cpustat[usage];
988 }
989
990 err = kcpustat_field_vtime(cpustat, curr, usage, cpu, &val);
991 rcu_read_unlock();
992
993 if (!err)
994 return val;
995
996 cpu_relax();
997 }
998 }
999 EXPORT_SYMBOL_GPL(kcpustat_field);
1000
kcpustat_cpu_fetch_vtime(struct kernel_cpustat * dst,const struct kernel_cpustat * src,struct task_struct * tsk,int cpu)1001 static int kcpustat_cpu_fetch_vtime(struct kernel_cpustat *dst,
1002 const struct kernel_cpustat *src,
1003 struct task_struct *tsk, int cpu)
1004 {
1005 struct vtime *vtime = &tsk->vtime;
1006 unsigned int seq;
1007
1008 do {
1009 u64 *cpustat;
1010 u64 delta;
1011 int state;
1012
1013 seq = read_seqcount_begin(&vtime->seqcount);
1014
1015 state = vtime_state_fetch(vtime, cpu);
1016 if (state < 0)
1017 return state;
1018
1019 *dst = *src;
1020 cpustat = dst->cpustat;
1021
1022 /* Task is sleeping, dead or idle, nothing to add */
1023 if (state < VTIME_SYS)
1024 continue;
1025
1026 delta = vtime_delta(vtime);
1027
1028 /*
1029 * Task runs either in user (including guest) or kernel space,
1030 * add pending nohz time to the right place.
1031 */
1032 if (state == VTIME_SYS) {
1033 cpustat[CPUTIME_SYSTEM] += vtime->stime + delta;
1034 } else if (state == VTIME_USER) {
1035 if (task_nice(tsk) > 0)
1036 cpustat[CPUTIME_NICE] += vtime->utime + delta;
1037 else
1038 cpustat[CPUTIME_USER] += vtime->utime + delta;
1039 } else {
1040 WARN_ON_ONCE(state != VTIME_GUEST);
1041 if (task_nice(tsk) > 0) {
1042 cpustat[CPUTIME_GUEST_NICE] += vtime->gtime + delta;
1043 cpustat[CPUTIME_NICE] += vtime->gtime + delta;
1044 } else {
1045 cpustat[CPUTIME_GUEST] += vtime->gtime + delta;
1046 cpustat[CPUTIME_USER] += vtime->gtime + delta;
1047 }
1048 }
1049 } while (read_seqcount_retry(&vtime->seqcount, seq));
1050
1051 return 0;
1052 }
1053
kcpustat_cpu_fetch(struct kernel_cpustat * dst,int cpu)1054 void kcpustat_cpu_fetch(struct kernel_cpustat *dst, int cpu)
1055 {
1056 const struct kernel_cpustat *src = &kcpustat_cpu(cpu);
1057 struct rq *rq;
1058 int err;
1059
1060 if (!vtime_accounting_enabled_cpu(cpu)) {
1061 *dst = *src;
1062 return;
1063 }
1064
1065 rq = cpu_rq(cpu);
1066
1067 for (;;) {
1068 struct task_struct *curr;
1069
1070 rcu_read_lock();
1071 curr = rcu_dereference(rq->curr);
1072 if (WARN_ON_ONCE(!curr)) {
1073 rcu_read_unlock();
1074 *dst = *src;
1075 return;
1076 }
1077
1078 err = kcpustat_cpu_fetch_vtime(dst, src, curr, cpu);
1079 rcu_read_unlock();
1080
1081 if (!err)
1082 return;
1083
1084 cpu_relax();
1085 }
1086 }
1087 EXPORT_SYMBOL_GPL(kcpustat_cpu_fetch);
1088
1089 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */
1090