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