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