xref: /openbmc/linux/kernel/sched/cputime.c (revision aac5987a)
1 #include <linux/export.h>
2 #include <linux/sched.h>
3 #include <linux/tsacct_kern.h>
4 #include <linux/kernel_stat.h>
5 #include <linux/static_key.h>
6 #include <linux/context_tracking.h>
7 #include <linux/sched/cputime.h>
8 #include "sched.h"
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 
27 void enable_sched_clock_irqtime(void)
28 {
29 	sched_clock_irqtime = 1;
30 }
31 
32 void disable_sched_clock_irqtime(void)
33 {
34 	sched_clock_irqtime = 0;
35 }
36 
37 /*
38  * Called before incrementing preempt_count on {soft,}irq_enter
39  * and before decrementing preempt_count on {soft,}irq_exit.
40  */
41 void irqtime_account_irq(struct task_struct *curr)
42 {
43 	struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
44 	u64 *cpustat = kcpustat_this_cpu->cpustat;
45 	s64 delta;
46 	int cpu;
47 
48 	if (!sched_clock_irqtime)
49 		return;
50 
51 	cpu = smp_processor_id();
52 	delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
53 	irqtime->irq_start_time += delta;
54 
55 	u64_stats_update_begin(&irqtime->sync);
56 	/*
57 	 * We do not account for softirq time from ksoftirqd here.
58 	 * We want to continue accounting softirq time to ksoftirqd thread
59 	 * in that case, so as not to confuse scheduler with a special task
60 	 * that do not consume any time, but still wants to run.
61 	 */
62 	if (hardirq_count()) {
63 		cpustat[CPUTIME_IRQ] += delta;
64 		irqtime->tick_delta += delta;
65 	} else if (in_serving_softirq() && curr != this_cpu_ksoftirqd()) {
66 		cpustat[CPUTIME_SOFTIRQ] += delta;
67 		irqtime->tick_delta += delta;
68 	}
69 
70 	u64_stats_update_end(&irqtime->sync);
71 }
72 EXPORT_SYMBOL_GPL(irqtime_account_irq);
73 
74 static u64 irqtime_tick_accounted(u64 maxtime)
75 {
76 	struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
77 	u64 delta;
78 
79 	delta = min(irqtime->tick_delta, maxtime);
80 	irqtime->tick_delta -= delta;
81 
82 	return delta;
83 }
84 
85 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
86 
87 #define sched_clock_irqtime	(0)
88 
89 static u64 irqtime_tick_accounted(u64 dummy)
90 {
91 	return 0;
92 }
93 
94 #endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
95 
96 static inline void task_group_account_field(struct task_struct *p, int index,
97 					    u64 tmp)
98 {
99 	/*
100 	 * Since all updates are sure to touch the root cgroup, we
101 	 * get ourselves ahead and touch it first. If the root cgroup
102 	 * is the only cgroup, then nothing else should be necessary.
103 	 *
104 	 */
105 	__this_cpu_add(kernel_cpustat.cpustat[index], tmp);
106 
107 	cpuacct_account_field(p, index, tmp);
108 }
109 
110 /*
111  * Account user cpu time to a process.
112  * @p: the process that the cpu time gets accounted to
113  * @cputime: the cpu time spent in user space since the last update
114  */
115 void account_user_time(struct task_struct *p, u64 cputime)
116 {
117 	int index;
118 
119 	/* Add user time to process. */
120 	p->utime += cputime;
121 	account_group_user_time(p, cputime);
122 
123 	index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
124 
125 	/* Add user time to cpustat. */
126 	task_group_account_field(p, index, cputime);
127 
128 	/* Account for user time used */
129 	acct_account_cputime(p);
130 }
131 
132 /*
133  * Account guest cpu time to a process.
134  * @p: the process that the cpu time gets accounted to
135  * @cputime: the cpu time spent in virtual machine since the last update
136  */
137 void account_guest_time(struct task_struct *p, u64 cputime)
138 {
139 	u64 *cpustat = kcpustat_this_cpu->cpustat;
140 
141 	/* Add guest time to process. */
142 	p->utime += cputime;
143 	account_group_user_time(p, cputime);
144 	p->gtime += cputime;
145 
146 	/* Add guest time to cpustat. */
147 	if (task_nice(p) > 0) {
148 		cpustat[CPUTIME_NICE] += cputime;
149 		cpustat[CPUTIME_GUEST_NICE] += cputime;
150 	} else {
151 		cpustat[CPUTIME_USER] += cputime;
152 		cpustat[CPUTIME_GUEST] += cputime;
153 	}
154 }
155 
156 /*
157  * Account system cpu time to a process and desired cpustat field
158  * @p: the process that the cpu time gets accounted to
159  * @cputime: the cpu time spent in kernel space since the last update
160  * @index: pointer to cpustat field that has to be updated
161  */
162 void account_system_index_time(struct task_struct *p,
163 			       u64 cputime, enum cpu_usage_stat index)
164 {
165 	/* Add system time to process. */
166 	p->stime += cputime;
167 	account_group_system_time(p, cputime);
168 
169 	/* Add system time to cpustat. */
170 	task_group_account_field(p, index, cputime);
171 
172 	/* Account for system time used */
173 	acct_account_cputime(p);
174 }
175 
176 /*
177  * Account system cpu time to a process.
178  * @p: the process that the cpu time gets accounted to
179  * @hardirq_offset: the offset to subtract from hardirq_count()
180  * @cputime: the cpu time spent in kernel space since the last update
181  */
182 void account_system_time(struct task_struct *p, int hardirq_offset, u64 cputime)
183 {
184 	int index;
185 
186 	if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
187 		account_guest_time(p, cputime);
188 		return;
189 	}
190 
191 	if (hardirq_count() - hardirq_offset)
192 		index = CPUTIME_IRQ;
193 	else if (in_serving_softirq())
194 		index = CPUTIME_SOFTIRQ;
195 	else
196 		index = CPUTIME_SYSTEM;
197 
198 	account_system_index_time(p, cputime, index);
199 }
200 
201 /*
202  * Account for involuntary wait time.
203  * @cputime: the cpu time spent in involuntary wait
204  */
205 void account_steal_time(u64 cputime)
206 {
207 	u64 *cpustat = kcpustat_this_cpu->cpustat;
208 
209 	cpustat[CPUTIME_STEAL] += cputime;
210 }
211 
212 /*
213  * Account for idle time.
214  * @cputime: the cpu time spent in idle wait
215  */
216 void account_idle_time(u64 cputime)
217 {
218 	u64 *cpustat = kcpustat_this_cpu->cpustat;
219 	struct rq *rq = this_rq();
220 
221 	if (atomic_read(&rq->nr_iowait) > 0)
222 		cpustat[CPUTIME_IOWAIT] += cputime;
223 	else
224 		cpustat[CPUTIME_IDLE] += cputime;
225 }
226 
227 /*
228  * When a guest is interrupted for a longer amount of time, missed clock
229  * ticks are not redelivered later. Due to that, this function may on
230  * occasion account more time than the calling functions think elapsed.
231  */
232 static __always_inline u64 steal_account_process_time(u64 maxtime)
233 {
234 #ifdef CONFIG_PARAVIRT
235 	if (static_key_false(&paravirt_steal_enabled)) {
236 		u64 steal;
237 
238 		steal = paravirt_steal_clock(smp_processor_id());
239 		steal -= this_rq()->prev_steal_time;
240 		steal = min(steal, maxtime);
241 		account_steal_time(steal);
242 		this_rq()->prev_steal_time += steal;
243 
244 		return steal;
245 	}
246 #endif
247 	return 0;
248 }
249 
250 /*
251  * Account how much elapsed time was spent in steal, irq, or softirq time.
252  */
253 static inline u64 account_other_time(u64 max)
254 {
255 	u64 accounted;
256 
257 	/* Shall be converted to a lockdep-enabled lightweight check */
258 	WARN_ON_ONCE(!irqs_disabled());
259 
260 	accounted = steal_account_process_time(max);
261 
262 	if (accounted < max)
263 		accounted += irqtime_tick_accounted(max - accounted);
264 
265 	return accounted;
266 }
267 
268 #ifdef CONFIG_64BIT
269 static inline u64 read_sum_exec_runtime(struct task_struct *t)
270 {
271 	return t->se.sum_exec_runtime;
272 }
273 #else
274 static u64 read_sum_exec_runtime(struct task_struct *t)
275 {
276 	u64 ns;
277 	struct rq_flags rf;
278 	struct rq *rq;
279 
280 	rq = task_rq_lock(t, &rf);
281 	ns = t->se.sum_exec_runtime;
282 	task_rq_unlock(rq, t, &rf);
283 
284 	return ns;
285 }
286 #endif
287 
288 /*
289  * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
290  * tasks (sum on group iteration) belonging to @tsk's group.
291  */
292 void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
293 {
294 	struct signal_struct *sig = tsk->signal;
295 	u64 utime, stime;
296 	struct task_struct *t;
297 	unsigned int seq, nextseq;
298 	unsigned long flags;
299 
300 	/*
301 	 * Update current task runtime to account pending time since last
302 	 * scheduler action or thread_group_cputime() call. This thread group
303 	 * might have other running tasks on different CPUs, but updating
304 	 * their runtime can affect syscall performance, so we skip account
305 	 * those pending times and rely only on values updated on tick or
306 	 * other scheduler action.
307 	 */
308 	if (same_thread_group(current, tsk))
309 		(void) task_sched_runtime(current);
310 
311 	rcu_read_lock();
312 	/* Attempt a lockless read on the first round. */
313 	nextseq = 0;
314 	do {
315 		seq = nextseq;
316 		flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
317 		times->utime = sig->utime;
318 		times->stime = sig->stime;
319 		times->sum_exec_runtime = sig->sum_sched_runtime;
320 
321 		for_each_thread(tsk, t) {
322 			task_cputime(t, &utime, &stime);
323 			times->utime += utime;
324 			times->stime += stime;
325 			times->sum_exec_runtime += read_sum_exec_runtime(t);
326 		}
327 		/* If lockless access failed, take the lock. */
328 		nextseq = 1;
329 	} while (need_seqretry(&sig->stats_lock, seq));
330 	done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
331 	rcu_read_unlock();
332 }
333 
334 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
335 /*
336  * Account a tick to a process and cpustat
337  * @p: the process that the cpu time gets accounted to
338  * @user_tick: is the tick from userspace
339  * @rq: the pointer to rq
340  *
341  * Tick demultiplexing follows the order
342  * - pending hardirq update
343  * - pending softirq update
344  * - user_time
345  * - idle_time
346  * - system time
347  *   - check for guest_time
348  *   - else account as system_time
349  *
350  * Check for hardirq is done both for system and user time as there is
351  * no timer going off while we are on hardirq and hence we may never get an
352  * opportunity to update it solely in system time.
353  * p->stime and friends are only updated on system time and not on irq
354  * softirq as those do not count in task exec_runtime any more.
355  */
356 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
357 					 struct rq *rq, int ticks)
358 {
359 	u64 other, cputime = TICK_NSEC * ticks;
360 
361 	/*
362 	 * When returning from idle, many ticks can get accounted at
363 	 * once, including some ticks of steal, irq, and softirq time.
364 	 * Subtract those ticks from the amount of time accounted to
365 	 * idle, or potentially user or system time. Due to rounding,
366 	 * other time can exceed ticks occasionally.
367 	 */
368 	other = account_other_time(ULONG_MAX);
369 	if (other >= cputime)
370 		return;
371 
372 	cputime -= other;
373 
374 	if (this_cpu_ksoftirqd() == p) {
375 		/*
376 		 * ksoftirqd time do not get accounted in cpu_softirq_time.
377 		 * So, we have to handle it separately here.
378 		 * Also, p->stime needs to be updated for ksoftirqd.
379 		 */
380 		account_system_index_time(p, cputime, CPUTIME_SOFTIRQ);
381 	} else if (user_tick) {
382 		account_user_time(p, cputime);
383 	} else if (p == rq->idle) {
384 		account_idle_time(cputime);
385 	} else if (p->flags & PF_VCPU) { /* System time or guest time */
386 		account_guest_time(p, cputime);
387 	} else {
388 		account_system_index_time(p, cputime, CPUTIME_SYSTEM);
389 	}
390 }
391 
392 static void irqtime_account_idle_ticks(int ticks)
393 {
394 	struct rq *rq = this_rq();
395 
396 	irqtime_account_process_tick(current, 0, rq, ticks);
397 }
398 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
399 static inline void irqtime_account_idle_ticks(int ticks) {}
400 static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
401 						struct rq *rq, 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
408 
409 #ifndef __ARCH_HAS_VTIME_TASK_SWITCH
410 void vtime_common_task_switch(struct task_struct *prev)
411 {
412 	if (is_idle_task(prev))
413 		vtime_account_idle(prev);
414 	else
415 		vtime_account_system(prev);
416 
417 	vtime_flush(prev);
418 	arch_vtime_task_switch(prev);
419 }
420 #endif
421 
422 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
423 
424 
425 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
426 /*
427  * Archs that account the whole time spent in the idle task
428  * (outside irq) as idle time can rely on this and just implement
429  * vtime_account_system() and vtime_account_idle(). Archs that
430  * have other meaning of the idle time (s390 only includes the
431  * time spent by the CPU when it's in low power mode) must override
432  * vtime_account().
433  */
434 #ifndef __ARCH_HAS_VTIME_ACCOUNT
435 void vtime_account_irq_enter(struct task_struct *tsk)
436 {
437 	if (!in_interrupt() && is_idle_task(tsk))
438 		vtime_account_idle(tsk);
439 	else
440 		vtime_account_system(tsk);
441 }
442 EXPORT_SYMBOL_GPL(vtime_account_irq_enter);
443 #endif /* __ARCH_HAS_VTIME_ACCOUNT */
444 
445 void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
446 {
447 	*ut = p->utime;
448 	*st = p->stime;
449 }
450 EXPORT_SYMBOL_GPL(task_cputime_adjusted);
451 
452 void 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 #else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
462 /*
463  * Account a single tick of cpu time.
464  * @p: the process that the cpu time gets accounted to
465  * @user_tick: indicates if the tick is a user or a system tick
466  */
467 void account_process_tick(struct task_struct *p, int user_tick)
468 {
469 	u64 cputime, steal;
470 	struct rq *rq = this_rq();
471 
472 	if (vtime_accounting_cpu_enabled())
473 		return;
474 
475 	if (sched_clock_irqtime) {
476 		irqtime_account_process_tick(p, user_tick, rq, 1);
477 		return;
478 	}
479 
480 	cputime = TICK_NSEC;
481 	steal = steal_account_process_time(ULONG_MAX);
482 
483 	if (steal >= cputime)
484 		return;
485 
486 	cputime -= steal;
487 
488 	if (user_tick)
489 		account_user_time(p, cputime);
490 	else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
491 		account_system_time(p, HARDIRQ_OFFSET, cputime);
492 	else
493 		account_idle_time(cputime);
494 }
495 
496 /*
497  * Account multiple ticks of idle time.
498  * @ticks: number of stolen ticks
499  */
500 void account_idle_ticks(unsigned long ticks)
501 {
502 	u64 cputime, steal;
503 
504 	if (sched_clock_irqtime) {
505 		irqtime_account_idle_ticks(ticks);
506 		return;
507 	}
508 
509 	cputime = ticks * TICK_NSEC;
510 	steal = steal_account_process_time(ULONG_MAX);
511 
512 	if (steal >= cputime)
513 		return;
514 
515 	cputime -= steal;
516 	account_idle_time(cputime);
517 }
518 
519 /*
520  * Perform (stime * rtime) / total, but avoid multiplication overflow by
521  * loosing precision when the numbers are big.
522  */
523 static u64 scale_stime(u64 stime, u64 rtime, u64 total)
524 {
525 	u64 scaled;
526 
527 	for (;;) {
528 		/* Make sure "rtime" is the bigger of stime/rtime */
529 		if (stime > rtime)
530 			swap(rtime, stime);
531 
532 		/* Make sure 'total' fits in 32 bits */
533 		if (total >> 32)
534 			goto drop_precision;
535 
536 		/* Does rtime (and thus stime) fit in 32 bits? */
537 		if (!(rtime >> 32))
538 			break;
539 
540 		/* Can we just balance rtime/stime rather than dropping bits? */
541 		if (stime >> 31)
542 			goto drop_precision;
543 
544 		/* We can grow stime and shrink rtime and try to make them both fit */
545 		stime <<= 1;
546 		rtime >>= 1;
547 		continue;
548 
549 drop_precision:
550 		/* We drop from rtime, it has more bits than stime */
551 		rtime >>= 1;
552 		total >>= 1;
553 	}
554 
555 	/*
556 	 * Make sure gcc understands that this is a 32x32->64 multiply,
557 	 * followed by a 64/32->64 divide.
558 	 */
559 	scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
560 	return scaled;
561 }
562 
563 /*
564  * Adjust tick based cputime random precision against scheduler runtime
565  * accounting.
566  *
567  * Tick based cputime accounting depend on random scheduling timeslices of a
568  * task to be interrupted or not by the timer.  Depending on these
569  * circumstances, the number of these interrupts may be over or
570  * under-optimistic, matching the real user and system cputime with a variable
571  * precision.
572  *
573  * Fix this by scaling these tick based values against the total runtime
574  * accounted by the CFS scheduler.
575  *
576  * This code provides the following guarantees:
577  *
578  *   stime + utime == rtime
579  *   stime_i+1 >= stime_i, utime_i+1 >= utime_i
580  *
581  * Assuming that rtime_i+1 >= rtime_i.
582  */
583 static void cputime_adjust(struct task_cputime *curr,
584 			   struct prev_cputime *prev,
585 			   u64 *ut, u64 *st)
586 {
587 	u64 rtime, stime, utime;
588 	unsigned long flags;
589 
590 	/* Serialize concurrent callers such that we can honour our guarantees */
591 	raw_spin_lock_irqsave(&prev->lock, flags);
592 	rtime = curr->sum_exec_runtime;
593 
594 	/*
595 	 * This is possible under two circumstances:
596 	 *  - rtime isn't monotonic after all (a bug);
597 	 *  - we got reordered by the lock.
598 	 *
599 	 * In both cases this acts as a filter such that the rest of the code
600 	 * can assume it is monotonic regardless of anything else.
601 	 */
602 	if (prev->stime + prev->utime >= rtime)
603 		goto out;
604 
605 	stime = curr->stime;
606 	utime = curr->utime;
607 
608 	/*
609 	 * If either stime or both stime and utime are 0, assume all runtime is
610 	 * userspace. Once a task gets some ticks, the monotonicy code at
611 	 * 'update' will ensure things converge to the observed ratio.
612 	 */
613 	if (stime == 0) {
614 		utime = rtime;
615 		goto update;
616 	}
617 
618 	if (utime == 0) {
619 		stime = rtime;
620 		goto update;
621 	}
622 
623 	stime = scale_stime(stime, rtime, stime + utime);
624 
625 update:
626 	/*
627 	 * Make sure stime doesn't go backwards; this preserves monotonicity
628 	 * for utime because rtime is monotonic.
629 	 *
630 	 *  utime_i+1 = rtime_i+1 - stime_i
631 	 *            = rtime_i+1 - (rtime_i - utime_i)
632 	 *            = (rtime_i+1 - rtime_i) + utime_i
633 	 *            >= utime_i
634 	 */
635 	if (stime < prev->stime)
636 		stime = prev->stime;
637 	utime = rtime - stime;
638 
639 	/*
640 	 * Make sure utime doesn't go backwards; this still preserves
641 	 * monotonicity for stime, analogous argument to above.
642 	 */
643 	if (utime < prev->utime) {
644 		utime = prev->utime;
645 		stime = rtime - utime;
646 	}
647 
648 	prev->stime = stime;
649 	prev->utime = utime;
650 out:
651 	*ut = prev->utime;
652 	*st = prev->stime;
653 	raw_spin_unlock_irqrestore(&prev->lock, flags);
654 }
655 
656 void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
657 {
658 	struct task_cputime cputime = {
659 		.sum_exec_runtime = p->se.sum_exec_runtime,
660 	};
661 
662 	task_cputime(p, &cputime.utime, &cputime.stime);
663 	cputime_adjust(&cputime, &p->prev_cputime, ut, st);
664 }
665 EXPORT_SYMBOL_GPL(task_cputime_adjusted);
666 
667 void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
668 {
669 	struct task_cputime cputime;
670 
671 	thread_group_cputime(p, &cputime);
672 	cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
673 }
674 #endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
675 
676 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
677 static u64 vtime_delta(struct task_struct *tsk)
678 {
679 	unsigned long now = READ_ONCE(jiffies);
680 
681 	if (time_before(now, (unsigned long)tsk->vtime_snap))
682 		return 0;
683 
684 	return jiffies_to_nsecs(now - tsk->vtime_snap);
685 }
686 
687 static u64 get_vtime_delta(struct task_struct *tsk)
688 {
689 	unsigned long now = READ_ONCE(jiffies);
690 	u64 delta, other;
691 
692 	/*
693 	 * Unlike tick based timing, vtime based timing never has lost
694 	 * ticks, and no need for steal time accounting to make up for
695 	 * lost ticks. Vtime accounts a rounded version of actual
696 	 * elapsed time. Limit account_other_time to prevent rounding
697 	 * errors from causing elapsed vtime to go negative.
698 	 */
699 	delta = jiffies_to_nsecs(now - tsk->vtime_snap);
700 	other = account_other_time(delta);
701 	WARN_ON_ONCE(tsk->vtime_snap_whence == VTIME_INACTIVE);
702 	tsk->vtime_snap = now;
703 
704 	return delta - other;
705 }
706 
707 static void __vtime_account_system(struct task_struct *tsk)
708 {
709 	account_system_time(tsk, irq_count(), get_vtime_delta(tsk));
710 }
711 
712 void vtime_account_system(struct task_struct *tsk)
713 {
714 	if (!vtime_delta(tsk))
715 		return;
716 
717 	write_seqcount_begin(&tsk->vtime_seqcount);
718 	__vtime_account_system(tsk);
719 	write_seqcount_end(&tsk->vtime_seqcount);
720 }
721 
722 void vtime_account_user(struct task_struct *tsk)
723 {
724 	write_seqcount_begin(&tsk->vtime_seqcount);
725 	tsk->vtime_snap_whence = VTIME_SYS;
726 	if (vtime_delta(tsk))
727 		account_user_time(tsk, get_vtime_delta(tsk));
728 	write_seqcount_end(&tsk->vtime_seqcount);
729 }
730 
731 void vtime_user_enter(struct task_struct *tsk)
732 {
733 	write_seqcount_begin(&tsk->vtime_seqcount);
734 	if (vtime_delta(tsk))
735 		__vtime_account_system(tsk);
736 	tsk->vtime_snap_whence = VTIME_USER;
737 	write_seqcount_end(&tsk->vtime_seqcount);
738 }
739 
740 void vtime_guest_enter(struct task_struct *tsk)
741 {
742 	/*
743 	 * The flags must be updated under the lock with
744 	 * the vtime_snap flush and update.
745 	 * That enforces a right ordering and update sequence
746 	 * synchronization against the reader (task_gtime())
747 	 * that can thus safely catch up with a tickless delta.
748 	 */
749 	write_seqcount_begin(&tsk->vtime_seqcount);
750 	if (vtime_delta(tsk))
751 		__vtime_account_system(tsk);
752 	current->flags |= PF_VCPU;
753 	write_seqcount_end(&tsk->vtime_seqcount);
754 }
755 EXPORT_SYMBOL_GPL(vtime_guest_enter);
756 
757 void vtime_guest_exit(struct task_struct *tsk)
758 {
759 	write_seqcount_begin(&tsk->vtime_seqcount);
760 	__vtime_account_system(tsk);
761 	current->flags &= ~PF_VCPU;
762 	write_seqcount_end(&tsk->vtime_seqcount);
763 }
764 EXPORT_SYMBOL_GPL(vtime_guest_exit);
765 
766 void vtime_account_idle(struct task_struct *tsk)
767 {
768 	account_idle_time(get_vtime_delta(tsk));
769 }
770 
771 void arch_vtime_task_switch(struct task_struct *prev)
772 {
773 	write_seqcount_begin(&prev->vtime_seqcount);
774 	prev->vtime_snap_whence = VTIME_INACTIVE;
775 	write_seqcount_end(&prev->vtime_seqcount);
776 
777 	write_seqcount_begin(&current->vtime_seqcount);
778 	current->vtime_snap_whence = VTIME_SYS;
779 	current->vtime_snap = jiffies;
780 	write_seqcount_end(&current->vtime_seqcount);
781 }
782 
783 void vtime_init_idle(struct task_struct *t, int cpu)
784 {
785 	unsigned long flags;
786 
787 	local_irq_save(flags);
788 	write_seqcount_begin(&t->vtime_seqcount);
789 	t->vtime_snap_whence = VTIME_SYS;
790 	t->vtime_snap = jiffies;
791 	write_seqcount_end(&t->vtime_seqcount);
792 	local_irq_restore(flags);
793 }
794 
795 u64 task_gtime(struct task_struct *t)
796 {
797 	unsigned int seq;
798 	u64 gtime;
799 
800 	if (!vtime_accounting_enabled())
801 		return t->gtime;
802 
803 	do {
804 		seq = read_seqcount_begin(&t->vtime_seqcount);
805 
806 		gtime = t->gtime;
807 		if (t->vtime_snap_whence == VTIME_SYS && t->flags & PF_VCPU)
808 			gtime += vtime_delta(t);
809 
810 	} while (read_seqcount_retry(&t->vtime_seqcount, seq));
811 
812 	return gtime;
813 }
814 
815 /*
816  * Fetch cputime raw values from fields of task_struct and
817  * add up the pending nohz execution time since the last
818  * cputime snapshot.
819  */
820 void task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
821 {
822 	u64 delta;
823 	unsigned int seq;
824 
825 	if (!vtime_accounting_enabled()) {
826 		*utime = t->utime;
827 		*stime = t->stime;
828 		return;
829 	}
830 
831 	do {
832 		seq = read_seqcount_begin(&t->vtime_seqcount);
833 
834 		*utime = t->utime;
835 		*stime = t->stime;
836 
837 		/* Task is sleeping, nothing to add */
838 		if (t->vtime_snap_whence == VTIME_INACTIVE || is_idle_task(t))
839 			continue;
840 
841 		delta = vtime_delta(t);
842 
843 		/*
844 		 * Task runs either in user or kernel space, add pending nohz time to
845 		 * the right place.
846 		 */
847 		if (t->vtime_snap_whence == VTIME_USER || t->flags & PF_VCPU)
848 			*utime += delta;
849 		else if (t->vtime_snap_whence == VTIME_SYS)
850 			*stime += delta;
851 	} while (read_seqcount_retry(&t->vtime_seqcount, seq));
852 }
853 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */
854