Linux Audio

Check our new training course

Loading...
v4.17
  1/*
  2 * Simple CPU accounting cgroup controller
  3 */
 
 
 
  4#include "sched.h"
 
 
 
 
  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 */
 19DEFINE_PER_CPU(struct irqtime, cpu_irqtime);
 
 20
 
 21static int sched_clock_irqtime;
 22
 23void enable_sched_clock_irqtime(void)
 24{
 25	sched_clock_irqtime = 1;
 26}
 27
 28void disable_sched_clock_irqtime(void)
 29{
 30	sched_clock_irqtime = 0;
 31}
 32
 33static 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 before incrementing preempt_count on {soft,}irq_enter
 47 * and before decrementing preempt_count on {soft,}irq_exit.
 48 */
 49void irqtime_account_irq(struct task_struct *curr)
 50{
 51	struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
 52	s64 delta;
 53	int cpu;
 54
 55	if (!sched_clock_irqtime)
 56		return;
 57
 
 
 58	cpu = smp_processor_id();
 59	delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
 60	irqtime->irq_start_time += delta;
 61
 
 62	/*
 63	 * We do not account for softirq time from ksoftirqd here.
 64	 * We want to continue accounting softirq time to ksoftirqd thread
 65	 * in that case, so as not to confuse scheduler with a special task
 66	 * that do not consume any time, but still wants to run.
 67	 */
 68	if (hardirq_count())
 69		irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
 70	else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
 71		irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
 
 
 
 72}
 73EXPORT_SYMBOL_GPL(irqtime_account_irq);
 74
 75static u64 irqtime_tick_accounted(u64 maxtime)
 76{
 77	struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
 78	u64 delta;
 
 
 79
 80	delta = min(irqtime->tick_delta, maxtime);
 81	irqtime->tick_delta -= delta;
 
 
 
 
 
 
 
 
 
 
 
 
 82
 83	return delta;
 
 
 
 
 
 84}
 85
 86#else /* CONFIG_IRQ_TIME_ACCOUNTING */
 87
 88#define sched_clock_irqtime	(0)
 89
 90static u64 irqtime_tick_accounted(u64 dummy)
 91{
 92	return 0;
 93}
 94
 95#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
 96
 97static inline void task_group_account_field(struct task_struct *p, int index,
 98					    u64 tmp)
 99{
100	/*
101	 * Since all updates are sure to touch the root cgroup, we
102	 * get ourselves ahead and touch it first. If the root cgroup
103	 * is the only cgroup, then nothing else should be necessary.
104	 *
105	 */
106	__this_cpu_add(kernel_cpustat.cpustat[index], tmp);
107
108	cgroup_account_cputime_field(p, index, tmp);
109}
110
111/*
112 * Account user CPU time to a process.
113 * @p: the process that the CPU time gets accounted to
114 * @cputime: the CPU time spent in user space since the last update
 
115 */
116void account_user_time(struct task_struct *p, u64 cputime)
 
117{
118	int index;
119
120	/* Add user time to process. */
121	p->utime += cputime;
 
122	account_group_user_time(p, cputime);
123
124	index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
125
126	/* Add user time to cpustat. */
127	task_group_account_field(p, index, cputime);
128
129	/* Account for user time used */
130	acct_account_cputime(p);
131}
132
133/*
134 * Account guest CPU time to a process.
135 * @p: the process that the CPU time gets accounted to
136 * @cputime: the CPU time spent in virtual machine since the last update
 
137 */
138void account_guest_time(struct task_struct *p, u64 cputime)
 
139{
140	u64 *cpustat = kcpustat_this_cpu->cpustat;
141
142	/* Add guest time to process. */
143	p->utime += cputime;
 
144	account_group_user_time(p, cputime);
145	p->gtime += cputime;
146
147	/* Add guest time to cpustat. */
148	if (task_nice(p) > 0) {
149		cpustat[CPUTIME_NICE] += cputime;
150		cpustat[CPUTIME_GUEST_NICE] += cputime;
151	} else {
152		cpustat[CPUTIME_USER] += cputime;
153		cpustat[CPUTIME_GUEST] += cputime;
154	}
155}
156
157/*
158 * Account system CPU time to a process and desired cpustat field
159 * @p: the process that the CPU time gets accounted to
160 * @cputime: the CPU time spent in kernel space since the last update
161 * @index: pointer to cpustat field that has to be updated
 
162 */
163void account_system_index_time(struct task_struct *p,
164			       u64 cputime, enum cpu_usage_stat index)
 
165{
166	/* Add system time to process. */
167	p->stime += cputime;
 
168	account_group_system_time(p, cputime);
169
170	/* Add system time to cpustat. */
171	task_group_account_field(p, index, cputime);
172
173	/* Account for system time used */
174	acct_account_cputime(p);
175}
176
177/*
178 * Account system CPU time to a process.
179 * @p: the process that the CPU time gets accounted to
180 * @hardirq_offset: the offset to subtract from hardirq_count()
181 * @cputime: the CPU time spent in kernel space since the last update
 
182 */
183void account_system_time(struct task_struct *p, int hardirq_offset, u64 cputime)
 
184{
185	int index;
186
187	if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
188		account_guest_time(p, cputime);
189		return;
190	}
191
192	if (hardirq_count() - hardirq_offset)
193		index = CPUTIME_IRQ;
194	else if (in_serving_softirq())
195		index = CPUTIME_SOFTIRQ;
196	else
197		index = CPUTIME_SYSTEM;
198
199	account_system_index_time(p, cputime, index);
200}
201
202/*
203 * Account for involuntary wait time.
204 * @cputime: the CPU time spent in involuntary wait
205 */
206void account_steal_time(u64 cputime)
207{
208	u64 *cpustat = kcpustat_this_cpu->cpustat;
209
210	cpustat[CPUTIME_STEAL] += cputime;
211}
212
213/*
214 * Account for idle time.
215 * @cputime: the CPU time spent in idle wait
216 */
217void account_idle_time(u64 cputime)
218{
219	u64 *cpustat = kcpustat_this_cpu->cpustat;
220	struct rq *rq = this_rq();
221
222	if (atomic_read(&rq->nr_iowait) > 0)
223		cpustat[CPUTIME_IOWAIT] += cputime;
224	else
225		cpustat[CPUTIME_IDLE] += cputime;
226}
227
228/*
229 * When a guest is interrupted for a longer amount of time, missed clock
230 * ticks are not redelivered later. Due to that, this function may on
231 * occasion account more time than the calling functions think elapsed.
232 */
233static __always_inline u64 steal_account_process_time(u64 maxtime)
234{
235#ifdef CONFIG_PARAVIRT
236	if (static_key_false(&paravirt_steal_enabled)) {
237		u64 steal;
 
238
239		steal = paravirt_steal_clock(smp_processor_id());
240		steal -= this_rq()->prev_steal_time;
241		steal = min(steal, maxtime);
242		account_steal_time(steal);
243		this_rq()->prev_steal_time += steal;
244
245		return steal;
 
 
 
 
 
 
 
 
 
246	}
247#endif
248	return 0;
249}
250
251/*
252 * Account how much elapsed time was spent in steal, irq, or softirq time.
253 */
254static inline u64 account_other_time(u64 max)
255{
256	u64 accounted;
257
258	lockdep_assert_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
269static inline u64 read_sum_exec_runtime(struct task_struct *t)
270{
271	return t->se.sum_exec_runtime;
272}
273#else
274static 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 */
292void 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 */
356static 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
392static 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 */
399static inline void irqtime_account_idle_ticks(int ticks) { }
400static 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# ifndef __ARCH_HAS_VTIME_TASK_SWITCH
 
409void vtime_common_task_switch(struct task_struct *prev)
410{
411	if (is_idle_task(prev))
412		vtime_account_idle(prev);
413	else
414		vtime_account_system(prev);
415
416	vtime_flush(prev);
 
 
417	arch_vtime_task_switch(prev);
418}
419# endif
420#endif /* CONFIG_VIRT_CPU_ACCOUNTING */
421
422
423#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
424/*
425 * Archs that account the whole time spent in the idle task
426 * (outside irq) as idle time can rely on this and just implement
427 * vtime_account_system() and vtime_account_idle(). Archs that
428 * have other meaning of the idle time (s390 only includes the
429 * time spent by the CPU when it's in low power mode) must override
430 * vtime_account().
431 */
432#ifndef __ARCH_HAS_VTIME_ACCOUNT
433void vtime_account_irq_enter(struct task_struct *tsk)
434{
435	if (!in_interrupt() && is_idle_task(tsk))
436		vtime_account_idle(tsk);
437	else
438		vtime_account_system(tsk);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
439}
440EXPORT_SYMBOL_GPL(vtime_account_irq_enter);
441#endif /* __ARCH_HAS_VTIME_ACCOUNT */
 
442
443void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
444		    u64 *ut, u64 *st)
445{
446	*ut = curr->utime;
447	*st = curr->stime;
448}
449
450void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
 
451{
452	*ut = p->utime;
453	*st = p->stime;
454}
455EXPORT_SYMBOL_GPL(task_cputime_adjusted);
456
457void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
458{
459	struct task_cputime cputime;
460
461	thread_group_cputime(p, &cputime);
462
463	*ut = cputime.utime;
464	*st = cputime.stime;
465}
466
467#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE: */
468
469/*
470 * Account a single tick of CPU time.
471 * @p: the process that the CPU time gets accounted to
472 * @user_tick: indicates if the tick is a user or a system tick
473 */
474void account_process_tick(struct task_struct *p, int user_tick)
475{
476	u64 cputime, steal;
477	struct rq *rq = this_rq();
478
479	if (vtime_accounting_cpu_enabled())
480		return;
481
482	if (sched_clock_irqtime) {
483		irqtime_account_process_tick(p, user_tick, rq, 1);
484		return;
485	}
486
487	cputime = TICK_NSEC;
488	steal = steal_account_process_time(ULONG_MAX);
489
490	if (steal >= cputime)
491		return;
492
493	cputime -= steal;
494
495	if (user_tick)
496		account_user_time(p, cputime);
497	else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
498		account_system_time(p, HARDIRQ_OFFSET, cputime);
 
499	else
500		account_idle_time(cputime);
 
 
 
 
 
 
 
 
 
 
501}
502
503/*
504 * Account multiple ticks of idle time.
505 * @ticks: number of stolen ticks
506 */
507void account_idle_ticks(unsigned long ticks)
508{
509	u64 cputime, steal;
510
511	if (sched_clock_irqtime) {
512		irqtime_account_idle_ticks(ticks);
513		return;
514	}
515
516	cputime = ticks * TICK_NSEC;
517	steal = steal_account_process_time(ULONG_MAX);
518
519	if (steal >= cputime)
520		return;
521
522	cputime -= steal;
523	account_idle_time(cputime);
524}
525
526/*
527 * Perform (stime * rtime) / total, but avoid multiplication overflow by
528 * loosing precision when the numbers are big.
529 */
530static u64 scale_stime(u64 stime, u64 rtime, u64 total)
531{
532	u64 scaled;
533
534	for (;;) {
535		/* Make sure "rtime" is the bigger of stime/rtime */
536		if (stime > rtime)
537			swap(rtime, stime);
538
539		/* Make sure 'total' fits in 32 bits */
540		if (total >> 32)
541			goto drop_precision;
542
543		/* Does rtime (and thus stime) fit in 32 bits? */
544		if (!(rtime >> 32))
545			break;
546
547		/* Can we just balance rtime/stime rather than dropping bits? */
548		if (stime >> 31)
549			goto drop_precision;
550
551		/* We can grow stime and shrink rtime and try to make them both fit */
552		stime <<= 1;
553		rtime >>= 1;
554		continue;
555
556drop_precision:
557		/* We drop from rtime, it has more bits than stime */
558		rtime >>= 1;
559		total >>= 1;
560	}
561
562	/*
563	 * Make sure gcc understands that this is a 32x32->64 multiply,
564	 * followed by a 64/32->64 divide.
565	 */
566	scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
567	return scaled;
568}
569
570/*
571 * Adjust tick based cputime random precision against scheduler runtime
572 * accounting.
573 *
574 * Tick based cputime accounting depend on random scheduling timeslices of a
575 * task to be interrupted or not by the timer.  Depending on these
576 * circumstances, the number of these interrupts may be over or
577 * under-optimistic, matching the real user and system cputime with a variable
578 * precision.
579 *
580 * Fix this by scaling these tick based values against the total runtime
581 * accounted by the CFS scheduler.
582 *
583 * This code provides the following guarantees:
584 *
585 *   stime + utime == rtime
586 *   stime_i+1 >= stime_i, utime_i+1 >= utime_i
587 *
588 * Assuming that rtime_i+1 >= rtime_i.
589 */
590void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
591		    u64 *ut, u64 *st)
 
592{
593	u64 rtime, stime, utime;
594	unsigned long flags;
595
596	/* Serialize concurrent callers such that we can honour our guarantees */
597	raw_spin_lock_irqsave(&prev->lock, flags);
598	rtime = curr->sum_exec_runtime;
599
600	/*
601	 * This is possible under two circumstances:
602	 *  - rtime isn't monotonic after all (a bug);
603	 *  - we got reordered by the lock.
604	 *
605	 * In both cases this acts as a filter such that the rest of the code
606	 * can assume it is monotonic regardless of anything else.
607	 */
608	if (prev->stime + prev->utime >= rtime)
609		goto out;
610
611	stime = curr->stime;
612	utime = curr->utime;
613
614	/*
615	 * If either stime or utime are 0, assume all runtime is userspace.
616	 * Once a task gets some ticks, the monotonicy code at 'update:'
617	 * will ensure things converge to the observed ratio.
618	 */
619	if (stime == 0) {
620		utime = rtime;
621		goto update;
622	}
623
624	if (utime == 0) {
625		stime = rtime;
626		goto update;
627	}
628
629	stime = scale_stime(stime, rtime, stime + utime);
 
630
631update:
632	/*
633	 * Make sure stime doesn't go backwards; this preserves monotonicity
634	 * for utime because rtime is monotonic.
635	 *
636	 *  utime_i+1 = rtime_i+1 - stime_i
637	 *            = rtime_i+1 - (rtime_i - utime_i)
638	 *            = (rtime_i+1 - rtime_i) + utime_i
639	 *            >= utime_i
640	 */
641	if (stime < prev->stime)
642		stime = prev->stime;
643	utime = rtime - stime;
644
645	/*
646	 * Make sure utime doesn't go backwards; this still preserves
647	 * monotonicity for stime, analogous argument to above.
648	 */
649	if (utime < prev->utime) {
650		utime = prev->utime;
651		stime = rtime - utime;
652	}
653
 
654	prev->stime = stime;
655	prev->utime = utime;
656out:
657	*ut = prev->utime;
658	*st = prev->stime;
659	raw_spin_unlock_irqrestore(&prev->lock, flags);
660}
661
662void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
663{
664	struct task_cputime cputime = {
665		.sum_exec_runtime = p->se.sum_exec_runtime,
666	};
667
668	task_cputime(p, &cputime.utime, &cputime.stime);
669	cputime_adjust(&cputime, &p->prev_cputime, ut, st);
670}
671EXPORT_SYMBOL_GPL(task_cputime_adjusted);
672
673void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
674{
675	struct task_cputime cputime;
676
677	thread_group_cputime(p, &cputime);
678	cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
679}
680#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
681
682#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
683static u64 vtime_delta(struct vtime *vtime)
684{
685	unsigned long long clock;
686
687	clock = sched_clock();
688	if (clock < vtime->starttime)
689		return 0;
690
691	return clock - vtime->starttime;
692}
693
694static u64 get_vtime_delta(struct vtime *vtime)
695{
696	u64 delta = vtime_delta(vtime);
697	u64 other;
698
699	/*
700	 * Unlike tick based timing, vtime based timing never has lost
701	 * ticks, and no need for steal time accounting to make up for
702	 * lost ticks. Vtime accounts a rounded version of actual
703	 * elapsed time. Limit account_other_time to prevent rounding
704	 * errors from causing elapsed vtime to go negative.
705	 */
706	other = account_other_time(delta);
707	WARN_ON_ONCE(vtime->state == VTIME_INACTIVE);
708	vtime->starttime += delta;
709
710	return delta - other;
711}
712
713static void __vtime_account_system(struct task_struct *tsk,
714				   struct vtime *vtime)
715{
716	vtime->stime += get_vtime_delta(vtime);
717	if (vtime->stime >= TICK_NSEC) {
718		account_system_time(tsk, irq_count(), vtime->stime);
719		vtime->stime = 0;
720	}
721}
722
723static void vtime_account_guest(struct task_struct *tsk,
724				struct vtime *vtime)
725{
726	vtime->gtime += get_vtime_delta(vtime);
727	if (vtime->gtime >= TICK_NSEC) {
728		account_guest_time(tsk, vtime->gtime);
729		vtime->gtime = 0;
730	}
731}
732
733void vtime_account_system(struct task_struct *tsk)
734{
735	struct vtime *vtime = &tsk->vtime;
736
737	if (!vtime_delta(vtime))
738		return;
739
740	write_seqcount_begin(&vtime->seqcount);
741	/* We might have scheduled out from guest path */
742	if (current->flags & PF_VCPU)
743		vtime_account_guest(tsk, vtime);
744	else
745		__vtime_account_system(tsk, vtime);
746	write_seqcount_end(&vtime->seqcount);
747}
748
749void vtime_user_enter(struct task_struct *tsk)
750{
751	struct vtime *vtime = &tsk->vtime;
752
753	write_seqcount_begin(&vtime->seqcount);
754	__vtime_account_system(tsk, vtime);
755	vtime->state = VTIME_USER;
756	write_seqcount_end(&vtime->seqcount);
757}
758
759void vtime_user_exit(struct task_struct *tsk)
760{
761	struct vtime *vtime = &tsk->vtime;
762
763	write_seqcount_begin(&vtime->seqcount);
764	vtime->utime += get_vtime_delta(vtime);
765	if (vtime->utime >= TICK_NSEC) {
766		account_user_time(tsk, vtime->utime);
767		vtime->utime = 0;
768	}
769	vtime->state = VTIME_SYS;
770	write_seqcount_end(&vtime->seqcount);
 
 
 
 
 
 
 
 
771}
772
773void vtime_guest_enter(struct task_struct *tsk)
774{
775	struct vtime *vtime = &tsk->vtime;
776	/*
777	 * The flags must be updated under the lock with
778	 * the vtime_starttime flush and update.
779	 * That enforces a right ordering and update sequence
780	 * synchronization against the reader (task_gtime())
781	 * that can thus safely catch up with a tickless delta.
782	 */
783	write_seqcount_begin(&vtime->seqcount);
784	__vtime_account_system(tsk, vtime);
 
785	current->flags |= PF_VCPU;
786	write_seqcount_end(&vtime->seqcount);
787}
788EXPORT_SYMBOL_GPL(vtime_guest_enter);
789
790void vtime_guest_exit(struct task_struct *tsk)
791{
792	struct vtime *vtime = &tsk->vtime;
793
794	write_seqcount_begin(&vtime->seqcount);
795	vtime_account_guest(tsk, vtime);
796	current->flags &= ~PF_VCPU;
797	write_seqcount_end(&vtime->seqcount);
798}
799EXPORT_SYMBOL_GPL(vtime_guest_exit);
800
801void vtime_account_idle(struct task_struct *tsk)
802{
803	account_idle_time(get_vtime_delta(&tsk->vtime));
 
 
804}
805
806void arch_vtime_task_switch(struct task_struct *prev)
807{
808	struct vtime *vtime = &prev->vtime;
809
810	write_seqcount_begin(&vtime->seqcount);
811	vtime->state = VTIME_INACTIVE;
812	write_seqcount_end(&vtime->seqcount);
813
814	vtime = &current->vtime;
815
816	write_seqcount_begin(&vtime->seqcount);
817	vtime->state = VTIME_SYS;
818	vtime->starttime = sched_clock();
819	write_seqcount_end(&vtime->seqcount);
820}
821
822void vtime_init_idle(struct task_struct *t, int cpu)
823{
824	struct vtime *vtime = &t->vtime;
825	unsigned long flags;
826
827	local_irq_save(flags);
828	write_seqcount_begin(&vtime->seqcount);
829	vtime->state = VTIME_SYS;
830	vtime->starttime = sched_clock();
831	write_seqcount_end(&vtime->seqcount);
832	local_irq_restore(flags);
833}
834
835u64 task_gtime(struct task_struct *t)
836{
837	struct vtime *vtime = &t->vtime;
838	unsigned int seq;
839	u64 gtime;
840
841	if (!vtime_accounting_enabled())
842		return t->gtime;
843
844	do {
845		seq = read_seqcount_begin(&vtime->seqcount);
846
847		gtime = t->gtime;
848		if (vtime->state == VTIME_SYS && t->flags & PF_VCPU)
849			gtime += vtime->gtime + vtime_delta(vtime);
850
851	} while (read_seqcount_retry(&vtime->seqcount, seq));
852
853	return gtime;
854}
855
856/*
857 * Fetch cputime raw values from fields of task_struct and
858 * add up the pending nohz execution time since the last
859 * cputime snapshot.
860 */
861void task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
 
 
 
 
862{
863	struct vtime *vtime = &t->vtime;
864	unsigned int seq;
865	u64 delta;
866
867	if (!vtime_accounting_enabled()) {
868		*utime = t->utime;
869		*stime = t->stime;
870		return;
871	}
872
873	do {
874		seq = read_seqcount_begin(&vtime->seqcount);
 
875
876		*utime = t->utime;
877		*stime = t->stime;
 
 
 
 
878
879		/* Task is sleeping, nothing to add */
880		if (vtime->state == VTIME_INACTIVE || is_idle_task(t))
 
881			continue;
882
883		delta = vtime_delta(vtime);
884
885		/*
886		 * Task runs either in user or kernel space, add pending nohz time to
887		 * the right place.
888		 */
889		if (vtime->state == VTIME_USER || t->flags & PF_VCPU)
890			*utime += vtime->utime + delta;
891		else if (vtime->state == VTIME_SYS)
892			*stime += vtime->stime + delta;
893	} while (read_seqcount_retry(&vtime->seqcount, seq));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
894}
895#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */
v4.6
  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 "sched.h"
  8#ifdef CONFIG_PARAVIRT
  9#include <asm/paravirt.h>
 10#endif
 11
 12
 13#ifdef CONFIG_IRQ_TIME_ACCOUNTING
 14
 15/*
 16 * There are no locks covering percpu hardirq/softirq time.
 17 * They are only modified in vtime_account, on corresponding CPU
 18 * with interrupts disabled. So, writes are safe.
 19 * They are read and saved off onto struct rq in update_rq_clock().
 20 * This may result in other CPU reading this CPU's irq time and can
 21 * race with irq/vtime_account on this CPU. We would either get old
 22 * or new value with a side effect of accounting a slice of irq time to wrong
 23 * task when irq is in progress while we read rq->clock. That is a worthy
 24 * compromise in place of having locks on each irq in account_system_time.
 25 */
 26DEFINE_PER_CPU(u64, cpu_hardirq_time);
 27DEFINE_PER_CPU(u64, cpu_softirq_time);
 28
 29static DEFINE_PER_CPU(u64, irq_start_time);
 30static int sched_clock_irqtime;
 31
 32void enable_sched_clock_irqtime(void)
 33{
 34	sched_clock_irqtime = 1;
 35}
 36
 37void disable_sched_clock_irqtime(void)
 38{
 39	sched_clock_irqtime = 0;
 40}
 41
 42#ifndef CONFIG_64BIT
 43DEFINE_PER_CPU(seqcount_t, irq_time_seq);
 44#endif /* CONFIG_64BIT */
 
 
 
 
 
 
 
 
 45
 46/*
 47 * Called before incrementing preempt_count on {soft,}irq_enter
 48 * and before decrementing preempt_count on {soft,}irq_exit.
 49 */
 50void irqtime_account_irq(struct task_struct *curr)
 51{
 52	unsigned long flags;
 53	s64 delta;
 54	int cpu;
 55
 56	if (!sched_clock_irqtime)
 57		return;
 58
 59	local_irq_save(flags);
 60
 61	cpu = smp_processor_id();
 62	delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
 63	__this_cpu_add(irq_start_time, delta);
 64
 65	irq_time_write_begin();
 66	/*
 67	 * We do not account for softirq time from ksoftirqd here.
 68	 * We want to continue accounting softirq time to ksoftirqd thread
 69	 * in that case, so as not to confuse scheduler with a special task
 70	 * that do not consume any time, but still wants to run.
 71	 */
 72	if (hardirq_count())
 73		__this_cpu_add(cpu_hardirq_time, delta);
 74	else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
 75		__this_cpu_add(cpu_softirq_time, delta);
 76
 77	irq_time_write_end();
 78	local_irq_restore(flags);
 79}
 80EXPORT_SYMBOL_GPL(irqtime_account_irq);
 81
 82static int irqtime_account_hi_update(void)
 83{
 84	u64 *cpustat = kcpustat_this_cpu->cpustat;
 85	unsigned long flags;
 86	u64 latest_ns;
 87	int ret = 0;
 88
 89	local_irq_save(flags);
 90	latest_ns = this_cpu_read(cpu_hardirq_time);
 91	if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
 92		ret = 1;
 93	local_irq_restore(flags);
 94	return ret;
 95}
 96
 97static int irqtime_account_si_update(void)
 98{
 99	u64 *cpustat = kcpustat_this_cpu->cpustat;
100	unsigned long flags;
101	u64 latest_ns;
102	int ret = 0;
103
104	local_irq_save(flags);
105	latest_ns = this_cpu_read(cpu_softirq_time);
106	if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
107		ret = 1;
108	local_irq_restore(flags);
109	return ret;
110}
111
112#else /* CONFIG_IRQ_TIME_ACCOUNTING */
113
114#define sched_clock_irqtime	(0)
115
 
 
 
 
 
116#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
117
118static inline void task_group_account_field(struct task_struct *p, int index,
119					    u64 tmp)
120{
121	/*
122	 * Since all updates are sure to touch the root cgroup, we
123	 * get ourselves ahead and touch it first. If the root cgroup
124	 * is the only cgroup, then nothing else should be necessary.
125	 *
126	 */
127	__this_cpu_add(kernel_cpustat.cpustat[index], tmp);
128
129	cpuacct_account_field(p, index, tmp);
130}
131
132/*
133 * Account user cpu time to a process.
134 * @p: the process that the cpu time gets accounted to
135 * @cputime: the cpu time spent in user space since the last update
136 * @cputime_scaled: cputime scaled by cpu frequency
137 */
138void account_user_time(struct task_struct *p, cputime_t cputime,
139		       cputime_t cputime_scaled)
140{
141	int index;
142
143	/* Add user time to process. */
144	p->utime += cputime;
145	p->utimescaled += cputime_scaled;
146	account_group_user_time(p, cputime);
147
148	index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
149
150	/* Add user time to cpustat. */
151	task_group_account_field(p, index, (__force u64) cputime);
152
153	/* Account for user time used */
154	acct_account_cputime(p);
155}
156
157/*
158 * Account guest cpu time to a process.
159 * @p: the process that the cpu time gets accounted to
160 * @cputime: the cpu time spent in virtual machine since the last update
161 * @cputime_scaled: cputime scaled by cpu frequency
162 */
163static void account_guest_time(struct task_struct *p, cputime_t cputime,
164			       cputime_t cputime_scaled)
165{
166	u64 *cpustat = kcpustat_this_cpu->cpustat;
167
168	/* Add guest time to process. */
169	p->utime += cputime;
170	p->utimescaled += cputime_scaled;
171	account_group_user_time(p, cputime);
172	p->gtime += cputime;
173
174	/* Add guest time to cpustat. */
175	if (task_nice(p) > 0) {
176		cpustat[CPUTIME_NICE] += (__force u64) cputime;
177		cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
178	} else {
179		cpustat[CPUTIME_USER] += (__force u64) cputime;
180		cpustat[CPUTIME_GUEST] += (__force u64) cputime;
181	}
182}
183
184/*
185 * Account system cpu time to a process and desired cpustat field
186 * @p: the process that the cpu time gets accounted to
187 * @cputime: the cpu time spent in kernel space since the last update
188 * @cputime_scaled: cputime scaled by cpu frequency
189 * @target_cputime64: pointer to cpustat field that has to be updated
190 */
191static inline
192void __account_system_time(struct task_struct *p, cputime_t cputime,
193			cputime_t cputime_scaled, int index)
194{
195	/* Add system time to process. */
196	p->stime += cputime;
197	p->stimescaled += cputime_scaled;
198	account_group_system_time(p, cputime);
199
200	/* Add system time to cpustat. */
201	task_group_account_field(p, index, (__force u64) cputime);
202
203	/* Account for system time used */
204	acct_account_cputime(p);
205}
206
207/*
208 * Account system cpu time to a process.
209 * @p: the process that the cpu time gets accounted to
210 * @hardirq_offset: the offset to subtract from hardirq_count()
211 * @cputime: the cpu time spent in kernel space since the last update
212 * @cputime_scaled: cputime scaled by cpu frequency
213 */
214void account_system_time(struct task_struct *p, int hardirq_offset,
215			 cputime_t cputime, cputime_t cputime_scaled)
216{
217	int index;
218
219	if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
220		account_guest_time(p, cputime, cputime_scaled);
221		return;
222	}
223
224	if (hardirq_count() - hardirq_offset)
225		index = CPUTIME_IRQ;
226	else if (in_serving_softirq())
227		index = CPUTIME_SOFTIRQ;
228	else
229		index = CPUTIME_SYSTEM;
230
231	__account_system_time(p, cputime, cputime_scaled, index);
232}
233
234/*
235 * Account for involuntary wait time.
236 * @cputime: the cpu time spent in involuntary wait
237 */
238void account_steal_time(cputime_t cputime)
239{
240	u64 *cpustat = kcpustat_this_cpu->cpustat;
241
242	cpustat[CPUTIME_STEAL] += (__force u64) cputime;
243}
244
245/*
246 * Account for idle time.
247 * @cputime: the cpu time spent in idle wait
248 */
249void account_idle_time(cputime_t cputime)
250{
251	u64 *cpustat = kcpustat_this_cpu->cpustat;
252	struct rq *rq = this_rq();
253
254	if (atomic_read(&rq->nr_iowait) > 0)
255		cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
256	else
257		cpustat[CPUTIME_IDLE] += (__force u64) cputime;
258}
259
260static __always_inline bool steal_account_process_tick(void)
 
 
 
 
 
261{
262#ifdef CONFIG_PARAVIRT
263	if (static_key_false(&paravirt_steal_enabled)) {
264		u64 steal;
265		unsigned long steal_jiffies;
266
267		steal = paravirt_steal_clock(smp_processor_id());
268		steal -= this_rq()->prev_steal_time;
 
 
 
269
270		/*
271		 * steal is in nsecs but our caller is expecting steal
272		 * time in jiffies. Lets cast the result to jiffies
273		 * granularity and account the rest on the next rounds.
274		 */
275		steal_jiffies = nsecs_to_jiffies(steal);
276		this_rq()->prev_steal_time += jiffies_to_nsecs(steal_jiffies);
277
278		account_steal_time(jiffies_to_cputime(steal_jiffies));
279		return steal_jiffies;
280	}
281#endif
282	return false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
283}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
284
285/*
286 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
287 * tasks (sum on group iteration) belonging to @tsk's group.
288 */
289void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
290{
291	struct signal_struct *sig = tsk->signal;
292	cputime_t utime, stime;
293	struct task_struct *t;
294	unsigned int seq, nextseq;
295	unsigned long flags;
296
 
 
 
 
 
 
 
 
 
 
 
297	rcu_read_lock();
298	/* Attempt a lockless read on the first round. */
299	nextseq = 0;
300	do {
301		seq = nextseq;
302		flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
303		times->utime = sig->utime;
304		times->stime = sig->stime;
305		times->sum_exec_runtime = sig->sum_sched_runtime;
306
307		for_each_thread(tsk, t) {
308			task_cputime(t, &utime, &stime);
309			times->utime += utime;
310			times->stime += stime;
311			times->sum_exec_runtime += task_sched_runtime(t);
312		}
313		/* If lockless access failed, take the lock. */
314		nextseq = 1;
315	} while (need_seqretry(&sig->stats_lock, seq));
316	done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
317	rcu_read_unlock();
318}
319
320#ifdef CONFIG_IRQ_TIME_ACCOUNTING
321/*
322 * Account a tick to a process and cpustat
323 * @p: the process that the cpu time gets accounted to
324 * @user_tick: is the tick from userspace
325 * @rq: the pointer to rq
326 *
327 * Tick demultiplexing follows the order
328 * - pending hardirq update
329 * - pending softirq update
330 * - user_time
331 * - idle_time
332 * - system time
333 *   - check for guest_time
334 *   - else account as system_time
335 *
336 * Check for hardirq is done both for system and user time as there is
337 * no timer going off while we are on hardirq and hence we may never get an
338 * opportunity to update it solely in system time.
339 * p->stime and friends are only updated on system time and not on irq
340 * softirq as those do not count in task exec_runtime any more.
341 */
342static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
343					 struct rq *rq, int ticks)
344{
345	cputime_t scaled = cputime_to_scaled(cputime_one_jiffy);
346	u64 cputime = (__force u64) cputime_one_jiffy;
347	u64 *cpustat = kcpustat_this_cpu->cpustat;
348
349	if (steal_account_process_tick())
 
 
 
 
 
 
 
 
350		return;
351
352	cputime *= ticks;
353	scaled *= ticks;
354
355	if (irqtime_account_hi_update()) {
356		cpustat[CPUTIME_IRQ] += cputime;
357	} else if (irqtime_account_si_update()) {
358		cpustat[CPUTIME_SOFTIRQ] += cputime;
359	} else if (this_cpu_ksoftirqd() == p) {
360		/*
361		 * ksoftirqd time do not get accounted in cpu_softirq_time.
362		 * So, we have to handle it separately here.
363		 * Also, p->stime needs to be updated for ksoftirqd.
364		 */
365		__account_system_time(p, cputime, scaled, CPUTIME_SOFTIRQ);
366	} else if (user_tick) {
367		account_user_time(p, cputime, scaled);
368	} else if (p == rq->idle) {
369		account_idle_time(cputime);
370	} else if (p->flags & PF_VCPU) { /* System time or guest time */
371		account_guest_time(p, cputime, scaled);
372	} else {
373		__account_system_time(p, cputime, scaled,	CPUTIME_SYSTEM);
374	}
375}
376
377static void irqtime_account_idle_ticks(int ticks)
378{
379	struct rq *rq = this_rq();
380
381	irqtime_account_process_tick(current, 0, rq, ticks);
382}
383#else /* CONFIG_IRQ_TIME_ACCOUNTING */
384static inline void irqtime_account_idle_ticks(int ticks) {}
385static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
386						struct rq *rq, int nr_ticks) {}
387#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
388
389/*
390 * Use precise platform statistics if available:
391 */
392#ifdef CONFIG_VIRT_CPU_ACCOUNTING
393
394#ifndef __ARCH_HAS_VTIME_TASK_SWITCH
395void vtime_common_task_switch(struct task_struct *prev)
396{
397	if (is_idle_task(prev))
398		vtime_account_idle(prev);
399	else
400		vtime_account_system(prev);
401
402#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
403	vtime_account_user(prev);
404#endif
405	arch_vtime_task_switch(prev);
406}
407#endif
 
 
408
 
409/*
410 * Archs that account the whole time spent in the idle task
411 * (outside irq) as idle time can rely on this and just implement
412 * vtime_account_system() and vtime_account_idle(). Archs that
413 * have other meaning of the idle time (s390 only includes the
414 * time spent by the CPU when it's in low power mode) must override
415 * vtime_account().
416 */
417#ifndef __ARCH_HAS_VTIME_ACCOUNT
418void vtime_common_account_irq_enter(struct task_struct *tsk)
419{
420	if (!in_interrupt()) {
421		/*
422		 * If we interrupted user, context_tracking_in_user()
423		 * is 1 because the context tracking don't hook
424		 * on irq entry/exit. This way we know if
425		 * we need to flush user time on kernel entry.
426		 */
427		if (context_tracking_in_user()) {
428			vtime_account_user(tsk);
429			return;
430		}
431
432		if (is_idle_task(tsk)) {
433			vtime_account_idle(tsk);
434			return;
435		}
436	}
437	vtime_account_system(tsk);
438}
439EXPORT_SYMBOL_GPL(vtime_common_account_irq_enter);
440#endif /* __ARCH_HAS_VTIME_ACCOUNT */
441#endif /* CONFIG_VIRT_CPU_ACCOUNTING */
442
 
 
 
 
 
 
443
444#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
445void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
446{
447	*ut = p->utime;
448	*st = p->stime;
449}
450EXPORT_SYMBOL_GPL(task_cputime_adjusted);
451
452void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *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 */
467void account_process_tick(struct task_struct *p, int user_tick)
468{
469	cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
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	if (steal_account_process_tick())
 
 
 
481		return;
482
 
 
483	if (user_tick)
484		account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
485	else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
486		account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
487				    one_jiffy_scaled);
488	else
489		account_idle_time(cputime_one_jiffy);
490}
491
492/*
493 * Account multiple ticks of steal time.
494 * @p: the process from which the cpu time has been stolen
495 * @ticks: number of stolen ticks
496 */
497void account_steal_ticks(unsigned long ticks)
498{
499	account_steal_time(jiffies_to_cputime(ticks));
500}
501
502/*
503 * Account multiple ticks of idle time.
504 * @ticks: number of stolen ticks
505 */
506void account_idle_ticks(unsigned long ticks)
507{
 
508
509	if (sched_clock_irqtime) {
510		irqtime_account_idle_ticks(ticks);
511		return;
512	}
513
514	account_idle_time(jiffies_to_cputime(ticks));
 
 
 
 
 
 
 
515}
516
517/*
518 * Perform (stime * rtime) / total, but avoid multiplication overflow by
519 * loosing precision when the numbers are big.
520 */
521static cputime_t scale_stime(u64 stime, u64 rtime, u64 total)
522{
523	u64 scaled;
524
525	for (;;) {
526		/* Make sure "rtime" is the bigger of stime/rtime */
527		if (stime > rtime)
528			swap(rtime, stime);
529
530		/* Make sure 'total' fits in 32 bits */
531		if (total >> 32)
532			goto drop_precision;
533
534		/* Does rtime (and thus stime) fit in 32 bits? */
535		if (!(rtime >> 32))
536			break;
537
538		/* Can we just balance rtime/stime rather than dropping bits? */
539		if (stime >> 31)
540			goto drop_precision;
541
542		/* We can grow stime and shrink rtime and try to make them both fit */
543		stime <<= 1;
544		rtime >>= 1;
545		continue;
546
547drop_precision:
548		/* We drop from rtime, it has more bits than stime */
549		rtime >>= 1;
550		total >>= 1;
551	}
552
553	/*
554	 * Make sure gcc understands that this is a 32x32->64 multiply,
555	 * followed by a 64/32->64 divide.
556	 */
557	scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
558	return (__force cputime_t) scaled;
559}
560
561/*
562 * Adjust tick based cputime random precision against scheduler runtime
563 * accounting.
564 *
565 * Tick based cputime accounting depend on random scheduling timeslices of a
566 * task to be interrupted or not by the timer.  Depending on these
567 * circumstances, the number of these interrupts may be over or
568 * under-optimistic, matching the real user and system cputime with a variable
569 * precision.
570 *
571 * Fix this by scaling these tick based values against the total runtime
572 * accounted by the CFS scheduler.
573 *
574 * This code provides the following guarantees:
575 *
576 *   stime + utime == rtime
577 *   stime_i+1 >= stime_i, utime_i+1 >= utime_i
578 *
579 * Assuming that rtime_i+1 >= rtime_i.
580 */
581static void cputime_adjust(struct task_cputime *curr,
582			   struct prev_cputime *prev,
583			   cputime_t *ut, cputime_t *st)
584{
585	cputime_t rtime, stime, utime;
586	unsigned long flags;
587
588	/* Serialize concurrent callers such that we can honour our guarantees */
589	raw_spin_lock_irqsave(&prev->lock, flags);
590	rtime = nsecs_to_cputime(curr->sum_exec_runtime);
591
592	/*
593	 * This is possible under two circumstances:
594	 *  - rtime isn't monotonic after all (a bug);
595	 *  - we got reordered by the lock.
596	 *
597	 * In both cases this acts as a filter such that the rest of the code
598	 * can assume it is monotonic regardless of anything else.
599	 */
600	if (prev->stime + prev->utime >= rtime)
601		goto out;
602
603	stime = curr->stime;
604	utime = curr->utime;
605
606	if (utime == 0) {
607		stime = rtime;
 
 
 
 
 
608		goto update;
609	}
610
611	if (stime == 0) {
612		utime = rtime;
613		goto update;
614	}
615
616	stime = scale_stime((__force u64)stime, (__force u64)rtime,
617			    (__force u64)(stime + utime));
618
 
619	/*
620	 * Make sure stime doesn't go backwards; this preserves monotonicity
621	 * for utime because rtime is monotonic.
622	 *
623	 *  utime_i+1 = rtime_i+1 - stime_i
624	 *            = rtime_i+1 - (rtime_i - utime_i)
625	 *            = (rtime_i+1 - rtime_i) + utime_i
626	 *            >= utime_i
627	 */
628	if (stime < prev->stime)
629		stime = prev->stime;
630	utime = rtime - stime;
631
632	/*
633	 * Make sure utime doesn't go backwards; this still preserves
634	 * monotonicity for stime, analogous argument to above.
635	 */
636	if (utime < prev->utime) {
637		utime = prev->utime;
638		stime = rtime - utime;
639	}
640
641update:
642	prev->stime = stime;
643	prev->utime = utime;
644out:
645	*ut = prev->utime;
646	*st = prev->stime;
647	raw_spin_unlock_irqrestore(&prev->lock, flags);
648}
649
650void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
651{
652	struct task_cputime cputime = {
653		.sum_exec_runtime = p->se.sum_exec_runtime,
654	};
655
656	task_cputime(p, &cputime.utime, &cputime.stime);
657	cputime_adjust(&cputime, &p->prev_cputime, ut, st);
658}
659EXPORT_SYMBOL_GPL(task_cputime_adjusted);
660
661void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
662{
663	struct task_cputime cputime;
664
665	thread_group_cputime(p, &cputime);
666	cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
667}
668#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
669
670#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
671static cputime_t vtime_delta(struct task_struct *tsk)
672{
673	unsigned long now = READ_ONCE(jiffies);
674
675	if (time_before(now, (unsigned long)tsk->vtime_snap))
 
676		return 0;
677
678	return jiffies_to_cputime(now - tsk->vtime_snap);
679}
680
681static cputime_t get_vtime_delta(struct task_struct *tsk)
682{
683	unsigned long now = READ_ONCE(jiffies);
684	unsigned long delta = now - tsk->vtime_snap;
685
686	WARN_ON_ONCE(tsk->vtime_snap_whence == VTIME_INACTIVE);
687	tsk->vtime_snap = now;
 
 
 
 
 
 
 
 
688
689	return jiffies_to_cputime(delta);
690}
691
692static void __vtime_account_system(struct task_struct *tsk)
 
693{
694	cputime_t delta_cpu = get_vtime_delta(tsk);
 
 
 
 
 
695
696	account_system_time(tsk, irq_count(), delta_cpu, cputime_to_scaled(delta_cpu));
 
 
 
 
 
 
 
697}
698
699void vtime_account_system(struct task_struct *tsk)
700{
701	if (!vtime_delta(tsk))
 
 
702		return;
703
704	write_seqcount_begin(&tsk->vtime_seqcount);
705	__vtime_account_system(tsk);
706	write_seqcount_end(&tsk->vtime_seqcount);
 
 
 
 
707}
708
709void vtime_gen_account_irq_exit(struct task_struct *tsk)
710{
711	write_seqcount_begin(&tsk->vtime_seqcount);
712	if (vtime_delta(tsk))
713		__vtime_account_system(tsk);
714	if (context_tracking_in_user())
715		tsk->vtime_snap_whence = VTIME_USER;
716	write_seqcount_end(&tsk->vtime_seqcount);
717}
718
719void vtime_account_user(struct task_struct *tsk)
720{
721	cputime_t delta_cpu;
722
723	write_seqcount_begin(&tsk->vtime_seqcount);
724	tsk->vtime_snap_whence = VTIME_SYS;
725	if (vtime_delta(tsk)) {
726		delta_cpu = get_vtime_delta(tsk);
727		account_user_time(tsk, delta_cpu, cputime_to_scaled(delta_cpu));
728	}
729	write_seqcount_end(&tsk->vtime_seqcount);
730}
731
732void vtime_user_enter(struct task_struct *tsk)
733{
734	write_seqcount_begin(&tsk->vtime_seqcount);
735	if (vtime_delta(tsk))
736		__vtime_account_system(tsk);
737	tsk->vtime_snap_whence = VTIME_USER;
738	write_seqcount_end(&tsk->vtime_seqcount);
739}
740
741void vtime_guest_enter(struct task_struct *tsk)
742{
 
743	/*
744	 * The flags must be updated under the lock with
745	 * the vtime_snap flush and update.
746	 * That enforces a right ordering and update sequence
747	 * synchronization against the reader (task_gtime())
748	 * that can thus safely catch up with a tickless delta.
749	 */
750	write_seqcount_begin(&tsk->vtime_seqcount);
751	if (vtime_delta(tsk))
752		__vtime_account_system(tsk);
753	current->flags |= PF_VCPU;
754	write_seqcount_end(&tsk->vtime_seqcount);
755}
756EXPORT_SYMBOL_GPL(vtime_guest_enter);
757
758void vtime_guest_exit(struct task_struct *tsk)
759{
760	write_seqcount_begin(&tsk->vtime_seqcount);
761	__vtime_account_system(tsk);
 
 
762	current->flags &= ~PF_VCPU;
763	write_seqcount_end(&tsk->vtime_seqcount);
764}
765EXPORT_SYMBOL_GPL(vtime_guest_exit);
766
767void vtime_account_idle(struct task_struct *tsk)
768{
769	cputime_t delta_cpu = get_vtime_delta(tsk);
770
771	account_idle_time(delta_cpu);
772}
773
774void arch_vtime_task_switch(struct task_struct *prev)
775{
776	write_seqcount_begin(&prev->vtime_seqcount);
777	prev->vtime_snap_whence = VTIME_INACTIVE;
778	write_seqcount_end(&prev->vtime_seqcount);
779
780	write_seqcount_begin(&current->vtime_seqcount);
781	current->vtime_snap_whence = VTIME_SYS;
782	current->vtime_snap = jiffies;
783	write_seqcount_end(&current->vtime_seqcount);
 
 
 
 
784}
785
786void vtime_init_idle(struct task_struct *t, int cpu)
787{
 
788	unsigned long flags;
789
790	local_irq_save(flags);
791	write_seqcount_begin(&t->vtime_seqcount);
792	t->vtime_snap_whence = VTIME_SYS;
793	t->vtime_snap = jiffies;
794	write_seqcount_end(&t->vtime_seqcount);
795	local_irq_restore(flags);
796}
797
798cputime_t task_gtime(struct task_struct *t)
799{
 
800	unsigned int seq;
801	cputime_t gtime;
802
803	if (!vtime_accounting_enabled())
804		return t->gtime;
805
806	do {
807		seq = read_seqcount_begin(&t->vtime_seqcount);
808
809		gtime = t->gtime;
810		if (t->vtime_snap_whence == VTIME_SYS && t->flags & PF_VCPU)
811			gtime += vtime_delta(t);
812
813	} while (read_seqcount_retry(&t->vtime_seqcount, seq));
814
815	return gtime;
816}
817
818/*
819 * Fetch cputime raw values from fields of task_struct and
820 * add up the pending nohz execution time since the last
821 * cputime snapshot.
822 */
823static void
824fetch_task_cputime(struct task_struct *t,
825		   cputime_t *u_dst, cputime_t *s_dst,
826		   cputime_t *u_src, cputime_t *s_src,
827		   cputime_t *udelta, cputime_t *sdelta)
828{
 
829	unsigned int seq;
830	unsigned long long delta;
 
 
 
 
 
 
831
832	do {
833		*udelta = 0;
834		*sdelta = 0;
835
836		seq = read_seqcount_begin(&t->vtime_seqcount);
837
838		if (u_dst)
839			*u_dst = *u_src;
840		if (s_dst)
841			*s_dst = *s_src;
842
843		/* Task is sleeping, nothing to add */
844		if (t->vtime_snap_whence == VTIME_INACTIVE ||
845		    is_idle_task(t))
846			continue;
847
848		delta = vtime_delta(t);
849
850		/*
851		 * Task runs either in user or kernel space, add pending nohz time to
852		 * the right place.
853		 */
854		if (t->vtime_snap_whence == VTIME_USER || t->flags & PF_VCPU) {
855			*udelta = delta;
856		} else {
857			if (t->vtime_snap_whence == VTIME_SYS)
858				*sdelta = delta;
859		}
860	} while (read_seqcount_retry(&t->vtime_seqcount, seq));
861}
862
863
864void task_cputime(struct task_struct *t, cputime_t *utime, cputime_t *stime)
865{
866	cputime_t udelta, sdelta;
867
868	if (!vtime_accounting_enabled()) {
869		if (utime)
870			*utime = t->utime;
871		if (stime)
872			*stime = t->stime;
873		return;
874	}
875
876	fetch_task_cputime(t, utime, stime, &t->utime,
877			   &t->stime, &udelta, &sdelta);
878	if (utime)
879		*utime += udelta;
880	if (stime)
881		*stime += sdelta;
882}
883
884void task_cputime_scaled(struct task_struct *t,
885			 cputime_t *utimescaled, cputime_t *stimescaled)
886{
887	cputime_t udelta, sdelta;
888
889	if (!vtime_accounting_enabled()) {
890		if (utimescaled)
891			*utimescaled = t->utimescaled;
892		if (stimescaled)
893			*stimescaled = t->stimescaled;
894		return;
895	}
896
897	fetch_task_cputime(t, utimescaled, stimescaled,
898			   &t->utimescaled, &t->stimescaled, &udelta, &sdelta);
899	if (utimescaled)
900		*utimescaled += cputime_to_scaled(udelta);
901	if (stimescaled)
902		*stimescaled += cputime_to_scaled(sdelta);
903}
904#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */