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 */
v3.15
  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
  9
 10#ifdef CONFIG_IRQ_TIME_ACCOUNTING
 11
 12/*
 13 * There are no locks covering percpu hardirq/softirq time.
 14 * They are only modified in vtime_account, on corresponding CPU
 15 * with interrupts disabled. So, writes are safe.
 16 * They are read and saved off onto struct rq in update_rq_clock().
 17 * This may result in other CPU reading this CPU's irq time and can
 18 * race with irq/vtime_account on this CPU. We would either get old
 19 * or new value with a side effect of accounting a slice of irq time to wrong
 20 * task when irq is in progress while we read rq->clock. That is a worthy
 21 * compromise in place of having locks on each irq in account_system_time.
 22 */
 23DEFINE_PER_CPU(u64, cpu_hardirq_time);
 24DEFINE_PER_CPU(u64, cpu_softirq_time);
 25
 26static DEFINE_PER_CPU(u64, irq_start_time);
 27static int sched_clock_irqtime;
 28
 29void enable_sched_clock_irqtime(void)
 30{
 31	sched_clock_irqtime = 1;
 32}
 33
 34void disable_sched_clock_irqtime(void)
 35{
 36	sched_clock_irqtime = 0;
 37}
 38
 39#ifndef CONFIG_64BIT
 40DEFINE_PER_CPU(seqcount_t, irq_time_seq);
 41#endif /* CONFIG_64BIT */
 
 
 
 
 
 
 
 
 42
 43/*
 44 * Called before incrementing preempt_count on {soft,}irq_enter
 45 * and before decrementing preempt_count on {soft,}irq_exit.
 46 */
 47void irqtime_account_irq(struct task_struct *curr)
 48{
 49	unsigned long flags;
 50	s64 delta;
 51	int cpu;
 52
 53	if (!sched_clock_irqtime)
 54		return;
 55
 56	local_irq_save(flags);
 57
 58	cpu = smp_processor_id();
 59	delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
 60	__this_cpu_add(irq_start_time, delta);
 61
 62	irq_time_write_begin();
 63	/*
 64	 * We do not account for softirq time from ksoftirqd here.
 65	 * We want to continue accounting softirq time to ksoftirqd thread
 66	 * in that case, so as not to confuse scheduler with a special task
 67	 * that do not consume any time, but still wants to run.
 68	 */
 69	if (hardirq_count())
 70		__this_cpu_add(cpu_hardirq_time, delta);
 71	else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
 72		__this_cpu_add(cpu_softirq_time, delta);
 73
 74	irq_time_write_end();
 75	local_irq_restore(flags);
 76}
 77EXPORT_SYMBOL_GPL(irqtime_account_irq);
 78
 79static int irqtime_account_hi_update(void)
 80{
 81	u64 *cpustat = kcpustat_this_cpu->cpustat;
 82	unsigned long flags;
 83	u64 latest_ns;
 84	int ret = 0;
 85
 86	local_irq_save(flags);
 87	latest_ns = this_cpu_read(cpu_hardirq_time);
 88	if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
 89		ret = 1;
 90	local_irq_restore(flags);
 91	return ret;
 92}
 93
 94static int irqtime_account_si_update(void)
 95{
 96	u64 *cpustat = kcpustat_this_cpu->cpustat;
 97	unsigned long flags;
 98	u64 latest_ns;
 99	int ret = 0;
100
101	local_irq_save(flags);
102	latest_ns = this_cpu_read(cpu_softirq_time);
103	if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
104		ret = 1;
105	local_irq_restore(flags);
106	return ret;
107}
108
109#else /* CONFIG_IRQ_TIME_ACCOUNTING */
110
111#define sched_clock_irqtime	(0)
112
 
 
 
 
 
113#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
114
115static inline void task_group_account_field(struct task_struct *p, int index,
116					    u64 tmp)
117{
118	/*
119	 * Since all updates are sure to touch the root cgroup, we
120	 * get ourselves ahead and touch it first. If the root cgroup
121	 * is the only cgroup, then nothing else should be necessary.
122	 *
123	 */
124	__this_cpu_add(kernel_cpustat.cpustat[index], tmp);
125
126	cpuacct_account_field(p, index, tmp);
127}
128
129/*
130 * Account user cpu time to a process.
131 * @p: the process that the cpu time gets accounted to
132 * @cputime: the cpu time spent in user space since the last update
133 * @cputime_scaled: cputime scaled by cpu frequency
134 */
135void account_user_time(struct task_struct *p, cputime_t cputime,
136		       cputime_t cputime_scaled)
137{
138	int index;
139
140	/* Add user time to process. */
141	p->utime += cputime;
142	p->utimescaled += cputime_scaled;
143	account_group_user_time(p, cputime);
144
145	index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
146
147	/* Add user time to cpustat. */
148	task_group_account_field(p, index, (__force u64) cputime);
149
150	/* Account for user time used */
151	acct_account_cputime(p);
152}
153
154/*
155 * Account guest cpu time to a process.
156 * @p: the process that the cpu time gets accounted to
157 * @cputime: the cpu time spent in virtual machine since the last update
158 * @cputime_scaled: cputime scaled by cpu frequency
159 */
160static void account_guest_time(struct task_struct *p, cputime_t cputime,
161			       cputime_t cputime_scaled)
162{
163	u64 *cpustat = kcpustat_this_cpu->cpustat;
164
165	/* Add guest time to process. */
166	p->utime += cputime;
167	p->utimescaled += cputime_scaled;
168	account_group_user_time(p, cputime);
169	p->gtime += cputime;
170
171	/* Add guest time to cpustat. */
172	if (task_nice(p) > 0) {
173		cpustat[CPUTIME_NICE] += (__force u64) cputime;
174		cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
175	} else {
176		cpustat[CPUTIME_USER] += (__force u64) cputime;
177		cpustat[CPUTIME_GUEST] += (__force u64) cputime;
178	}
179}
180
181/*
182 * Account system cpu time to a process and desired cpustat field
183 * @p: the process that the cpu time gets accounted to
184 * @cputime: the cpu time spent in kernel space since the last update
185 * @cputime_scaled: cputime scaled by cpu frequency
186 * @target_cputime64: pointer to cpustat field that has to be updated
187 */
188static inline
189void __account_system_time(struct task_struct *p, cputime_t cputime,
190			cputime_t cputime_scaled, int index)
191{
192	/* Add system time to process. */
193	p->stime += cputime;
194	p->stimescaled += cputime_scaled;
195	account_group_system_time(p, cputime);
196
197	/* Add system time to cpustat. */
198	task_group_account_field(p, index, (__force u64) cputime);
199
200	/* Account for system time used */
201	acct_account_cputime(p);
202}
203
204/*
205 * Account system cpu time to a process.
206 * @p: the process that the cpu time gets accounted to
207 * @hardirq_offset: the offset to subtract from hardirq_count()
208 * @cputime: the cpu time spent in kernel space since the last update
209 * @cputime_scaled: cputime scaled by cpu frequency
210 */
211void account_system_time(struct task_struct *p, int hardirq_offset,
212			 cputime_t cputime, cputime_t cputime_scaled)
213{
214	int index;
215
216	if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
217		account_guest_time(p, cputime, cputime_scaled);
218		return;
219	}
220
221	if (hardirq_count() - hardirq_offset)
222		index = CPUTIME_IRQ;
223	else if (in_serving_softirq())
224		index = CPUTIME_SOFTIRQ;
225	else
226		index = CPUTIME_SYSTEM;
227
228	__account_system_time(p, cputime, cputime_scaled, index);
229}
230
231/*
232 * Account for involuntary wait time.
233 * @cputime: the cpu time spent in involuntary wait
234 */
235void account_steal_time(cputime_t cputime)
236{
237	u64 *cpustat = kcpustat_this_cpu->cpustat;
238
239	cpustat[CPUTIME_STEAL] += (__force u64) cputime;
240}
241
242/*
243 * Account for idle time.
244 * @cputime: the cpu time spent in idle wait
245 */
246void account_idle_time(cputime_t cputime)
247{
248	u64 *cpustat = kcpustat_this_cpu->cpustat;
249	struct rq *rq = this_rq();
250
251	if (atomic_read(&rq->nr_iowait) > 0)
252		cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
253	else
254		cpustat[CPUTIME_IDLE] += (__force u64) cputime;
255}
256
257static __always_inline bool steal_account_process_tick(void)
 
 
 
 
 
258{
259#ifdef CONFIG_PARAVIRT
260	if (static_key_false(&paravirt_steal_enabled)) {
261		u64 steal;
262		cputime_t steal_ct;
263
264		steal = paravirt_steal_clock(smp_processor_id());
265		steal -= this_rq()->prev_steal_time;
 
 
 
266
267		/*
268		 * cputime_t may be less precise than nsecs (eg: if it's
269		 * based on jiffies). Lets cast the result to cputime
270		 * granularity and account the rest on the next rounds.
271		 */
272		steal_ct = nsecs_to_cputime(steal);
273		this_rq()->prev_steal_time += cputime_to_nsecs(steal_ct);
274
275		account_steal_time(steal_ct);
276		return steal_ct;
277	}
278#endif
279	return false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
280}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
281
282/*
283 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
284 * tasks (sum on group iteration) belonging to @tsk's group.
285 */
286void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
287{
288	struct signal_struct *sig = tsk->signal;
289	cputime_t utime, stime;
290	struct task_struct *t;
 
 
291
292	times->utime = sig->utime;
293	times->stime = sig->stime;
294	times->sum_exec_runtime = sig->sum_sched_runtime;
 
 
 
 
 
 
 
295
296	rcu_read_lock();
297	/* make sure we can trust tsk->thread_group list */
298	if (!likely(pid_alive(tsk)))
299		goto out;
300
301	t = tsk;
302	do {
303		task_cputime(t, &utime, &stime);
304		times->utime += utime;
305		times->stime += stime;
306		times->sum_exec_runtime += task_sched_runtime(t);
307	} while_each_thread(tsk, t);
308out:
 
 
 
 
 
 
 
 
 
 
309	rcu_read_unlock();
310}
311
312#ifdef CONFIG_IRQ_TIME_ACCOUNTING
313/*
314 * Account a tick to a process and cpustat
315 * @p: the process that the cpu time gets accounted to
316 * @user_tick: is the tick from userspace
317 * @rq: the pointer to rq
318 *
319 * Tick demultiplexing follows the order
320 * - pending hardirq update
321 * - pending softirq update
322 * - user_time
323 * - idle_time
324 * - system time
325 *   - check for guest_time
326 *   - else account as system_time
327 *
328 * Check for hardirq is done both for system and user time as there is
329 * no timer going off while we are on hardirq and hence we may never get an
330 * opportunity to update it solely in system time.
331 * p->stime and friends are only updated on system time and not on irq
332 * softirq as those do not count in task exec_runtime any more.
333 */
334static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
335					 struct rq *rq, int ticks)
336{
337	cputime_t scaled = cputime_to_scaled(cputime_one_jiffy);
338	u64 cputime = (__force u64) cputime_one_jiffy;
339	u64 *cpustat = kcpustat_this_cpu->cpustat;
340
341	if (steal_account_process_tick())
 
 
 
 
 
 
 
 
342		return;
343
344	cputime *= ticks;
345	scaled *= ticks;
346
347	if (irqtime_account_hi_update()) {
348		cpustat[CPUTIME_IRQ] += cputime;
349	} else if (irqtime_account_si_update()) {
350		cpustat[CPUTIME_SOFTIRQ] += cputime;
351	} else if (this_cpu_ksoftirqd() == p) {
352		/*
353		 * ksoftirqd time do not get accounted in cpu_softirq_time.
354		 * So, we have to handle it separately here.
355		 * Also, p->stime needs to be updated for ksoftirqd.
356		 */
357		__account_system_time(p, cputime, scaled, CPUTIME_SOFTIRQ);
358	} else if (user_tick) {
359		account_user_time(p, cputime, scaled);
360	} else if (p == rq->idle) {
361		account_idle_time(cputime);
362	} else if (p->flags & PF_VCPU) { /* System time or guest time */
363		account_guest_time(p, cputime, scaled);
364	} else {
365		__account_system_time(p, cputime, scaled,	CPUTIME_SYSTEM);
366	}
367}
368
369static void irqtime_account_idle_ticks(int ticks)
370{
371	struct rq *rq = this_rq();
372
373	irqtime_account_process_tick(current, 0, rq, ticks);
374}
375#else /* CONFIG_IRQ_TIME_ACCOUNTING */
376static inline void irqtime_account_idle_ticks(int ticks) {}
377static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
378						struct rq *rq, int nr_ticks) {}
379#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
380
381/*
382 * Use precise platform statistics if available:
383 */
384#ifdef CONFIG_VIRT_CPU_ACCOUNTING
385
386#ifndef __ARCH_HAS_VTIME_TASK_SWITCH
387void vtime_common_task_switch(struct task_struct *prev)
388{
389	if (is_idle_task(prev))
390		vtime_account_idle(prev);
391	else
392		vtime_account_system(prev);
393
394#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
395	vtime_account_user(prev);
396#endif
397	arch_vtime_task_switch(prev);
398}
399#endif
 
 
400
 
401/*
402 * Archs that account the whole time spent in the idle task
403 * (outside irq) as idle time can rely on this and just implement
404 * vtime_account_system() and vtime_account_idle(). Archs that
405 * have other meaning of the idle time (s390 only includes the
406 * time spent by the CPU when it's in low power mode) must override
407 * vtime_account().
408 */
409#ifndef __ARCH_HAS_VTIME_ACCOUNT
410void vtime_common_account_irq_enter(struct task_struct *tsk)
411{
412	if (!in_interrupt()) {
413		/*
414		 * If we interrupted user, context_tracking_in_user()
415		 * is 1 because the context tracking don't hook
416		 * on irq entry/exit. This way we know if
417		 * we need to flush user time on kernel entry.
418		 */
419		if (context_tracking_in_user()) {
420			vtime_account_user(tsk);
421			return;
422		}
423
424		if (is_idle_task(tsk)) {
425			vtime_account_idle(tsk);
426			return;
427		}
428	}
429	vtime_account_system(tsk);
430}
431EXPORT_SYMBOL_GPL(vtime_common_account_irq_enter);
432#endif /* __ARCH_HAS_VTIME_ACCOUNT */
433#endif /* CONFIG_VIRT_CPU_ACCOUNTING */
434
 
 
 
 
 
 
435
436#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
437void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
438{
439	*ut = p->utime;
440	*st = p->stime;
441}
 
442
443void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
444{
445	struct task_cputime cputime;
446
447	thread_group_cputime(p, &cputime);
448
449	*ut = cputime.utime;
450	*st = cputime.stime;
451}
452#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
 
 
453/*
454 * Account a single tick of cpu time.
455 * @p: the process that the cpu time gets accounted to
456 * @user_tick: indicates if the tick is a user or a system tick
457 */
458void account_process_tick(struct task_struct *p, int user_tick)
459{
460	cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
461	struct rq *rq = this_rq();
462
463	if (vtime_accounting_enabled())
464		return;
465
466	if (sched_clock_irqtime) {
467		irqtime_account_process_tick(p, user_tick, rq, 1);
468		return;
469	}
470
471	if (steal_account_process_tick())
 
 
 
472		return;
473
 
 
474	if (user_tick)
475		account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
476	else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
477		account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
478				    one_jiffy_scaled);
479	else
480		account_idle_time(cputime_one_jiffy);
481}
482
483/*
484 * Account multiple ticks of steal time.
485 * @p: the process from which the cpu time has been stolen
486 * @ticks: number of stolen ticks
487 */
488void account_steal_ticks(unsigned long ticks)
489{
490	account_steal_time(jiffies_to_cputime(ticks));
491}
492
493/*
494 * Account multiple ticks of idle time.
495 * @ticks: number of stolen ticks
496 */
497void account_idle_ticks(unsigned long ticks)
498{
 
499
500	if (sched_clock_irqtime) {
501		irqtime_account_idle_ticks(ticks);
502		return;
503	}
504
505	account_idle_time(jiffies_to_cputime(ticks));
 
 
 
 
 
 
 
506}
507
508/*
509 * Perform (stime * rtime) / total, but avoid multiplication overflow by
510 * loosing precision when the numbers are big.
511 */
512static cputime_t scale_stime(u64 stime, u64 rtime, u64 total)
513{
514	u64 scaled;
515
516	for (;;) {
517		/* Make sure "rtime" is the bigger of stime/rtime */
518		if (stime > rtime)
519			swap(rtime, stime);
520
521		/* Make sure 'total' fits in 32 bits */
522		if (total >> 32)
523			goto drop_precision;
524
525		/* Does rtime (and thus stime) fit in 32 bits? */
526		if (!(rtime >> 32))
527			break;
528
529		/* Can we just balance rtime/stime rather than dropping bits? */
530		if (stime >> 31)
531			goto drop_precision;
532
533		/* We can grow stime and shrink rtime and try to make them both fit */
534		stime <<= 1;
535		rtime >>= 1;
536		continue;
537
538drop_precision:
539		/* We drop from rtime, it has more bits than stime */
540		rtime >>= 1;
541		total >>= 1;
542	}
543
544	/*
545	 * Make sure gcc understands that this is a 32x32->64 multiply,
546	 * followed by a 64/32->64 divide.
547	 */
548	scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
549	return (__force cputime_t) scaled;
550}
551
552/*
553 * Adjust tick based cputime random precision against scheduler
554 * runtime accounting.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
555 */
556static void cputime_adjust(struct task_cputime *curr,
557			   struct cputime *prev,
558			   cputime_t *ut, cputime_t *st)
559{
560	cputime_t rtime, stime, utime;
 
 
 
 
 
561
562	/*
563	 * Tick based cputime accounting depend on random scheduling
564	 * timeslices of a task to be interrupted or not by the timer.
565	 * Depending on these circumstances, the number of these interrupts
566	 * may be over or under-optimistic, matching the real user and system
567	 * cputime with a variable precision.
568	 *
569	 * Fix this by scaling these tick based values against the total
570	 * runtime accounted by the CFS scheduler.
571	 */
572	rtime = nsecs_to_cputime(curr->sum_exec_runtime);
573
574	/*
575	 * Update userspace visible utime/stime values only if actual execution
576	 * time is bigger than already exported. Note that can happen, that we
577	 * provided bigger values due to scaling inaccuracy on big numbers.
578	 */
579	if (prev->stime + prev->utime >= rtime)
580		goto out;
581
582	stime = curr->stime;
583	utime = curr->utime;
584
 
 
 
 
 
 
 
 
 
 
585	if (utime == 0) {
586		stime = rtime;
587	} else if (stime == 0) {
588		utime = rtime;
589	} else {
590		cputime_t total = stime + utime;
591
592		stime = scale_stime((__force u64)stime,
593				    (__force u64)rtime, (__force u64)total);
594		utime = rtime - stime;
595	}
 
 
 
 
 
 
 
 
 
596
597	/*
598	 * If the tick based count grows faster than the scheduler one,
599	 * the result of the scaling may go backward.
600	 * Let's enforce monotonicity.
601	 */
602	prev->stime = max(prev->stime, stime);
603	prev->utime = max(prev->utime, utime);
 
 
604
 
 
605out:
606	*ut = prev->utime;
607	*st = prev->stime;
 
608}
609
610void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
611{
612	struct task_cputime cputime = {
613		.sum_exec_runtime = p->se.sum_exec_runtime,
614	};
615
616	task_cputime(p, &cputime.utime, &cputime.stime);
617	cputime_adjust(&cputime, &p->prev_cputime, ut, st);
618}
 
619
620/*
621 * Must be called with siglock held.
622 */
623void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *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
633static unsigned long long vtime_delta(struct task_struct *tsk)
634{
635	unsigned long long clock;
636
637	clock = local_clock();
638	if (clock < tsk->vtime_snap)
639		return 0;
640
641	return clock - tsk->vtime_snap;
642}
643
644static cputime_t get_vtime_delta(struct task_struct *tsk)
645{
646	unsigned long long delta = vtime_delta(tsk);
 
647
648	WARN_ON_ONCE(tsk->vtime_snap_whence == VTIME_SLEEPING);
649	tsk->vtime_snap += delta;
 
 
 
 
 
 
 
 
650
651	/* CHECKME: always safe to convert nsecs to cputime? */
652	return nsecs_to_cputime(delta);
653}
654
655static void __vtime_account_system(struct task_struct *tsk)
 
656{
657	cputime_t delta_cpu = get_vtime_delta(tsk);
 
 
 
 
 
658
659	account_system_time(tsk, irq_count(), delta_cpu, cputime_to_scaled(delta_cpu));
 
 
 
 
 
 
 
660}
661
662void vtime_account_system(struct task_struct *tsk)
663{
664	write_seqlock(&tsk->vtime_seqlock);
665	__vtime_account_system(tsk);
666	write_sequnlock(&tsk->vtime_seqlock);
667}
668
669void vtime_gen_account_irq_exit(struct task_struct *tsk)
670{
671	write_seqlock(&tsk->vtime_seqlock);
672	__vtime_account_system(tsk);
673	if (context_tracking_in_user())
674		tsk->vtime_snap_whence = VTIME_USER;
675	write_sequnlock(&tsk->vtime_seqlock);
676}
677
678void vtime_account_user(struct task_struct *tsk)
679{
680	cputime_t delta_cpu;
681
682	write_seqlock(&tsk->vtime_seqlock);
683	delta_cpu = get_vtime_delta(tsk);
684	tsk->vtime_snap_whence = VTIME_SYS;
685	account_user_time(tsk, delta_cpu, cputime_to_scaled(delta_cpu));
686	write_sequnlock(&tsk->vtime_seqlock);
687}
688
689void vtime_user_enter(struct task_struct *tsk)
690{
691	write_seqlock(&tsk->vtime_seqlock);
692	__vtime_account_system(tsk);
693	tsk->vtime_snap_whence = VTIME_USER;
694	write_sequnlock(&tsk->vtime_seqlock);
 
 
 
 
 
 
695}
696
697void vtime_guest_enter(struct task_struct *tsk)
698{
 
699	/*
700	 * The flags must be updated under the lock with
701	 * the vtime_snap flush and update.
702	 * That enforces a right ordering and update sequence
703	 * synchronization against the reader (task_gtime())
704	 * that can thus safely catch up with a tickless delta.
705	 */
706	write_seqlock(&tsk->vtime_seqlock);
707	__vtime_account_system(tsk);
708	current->flags |= PF_VCPU;
709	write_sequnlock(&tsk->vtime_seqlock);
710}
711EXPORT_SYMBOL_GPL(vtime_guest_enter);
712
713void vtime_guest_exit(struct task_struct *tsk)
714{
715	write_seqlock(&tsk->vtime_seqlock);
716	__vtime_account_system(tsk);
 
 
717	current->flags &= ~PF_VCPU;
718	write_sequnlock(&tsk->vtime_seqlock);
719}
720EXPORT_SYMBOL_GPL(vtime_guest_exit);
721
722void vtime_account_idle(struct task_struct *tsk)
723{
724	cputime_t delta_cpu = get_vtime_delta(tsk);
725
726	account_idle_time(delta_cpu);
727}
728
729void arch_vtime_task_switch(struct task_struct *prev)
730{
731	write_seqlock(&prev->vtime_seqlock);
732	prev->vtime_snap_whence = VTIME_SLEEPING;
733	write_sequnlock(&prev->vtime_seqlock);
734
735	write_seqlock(&current->vtime_seqlock);
736	current->vtime_snap_whence = VTIME_SYS;
737	current->vtime_snap = sched_clock_cpu(smp_processor_id());
738	write_sequnlock(&current->vtime_seqlock);
 
 
 
 
739}
740
741void vtime_init_idle(struct task_struct *t, int cpu)
742{
 
743	unsigned long flags;
744
745	write_seqlock_irqsave(&t->vtime_seqlock, flags);
746	t->vtime_snap_whence = VTIME_SYS;
747	t->vtime_snap = sched_clock_cpu(cpu);
748	write_sequnlock_irqrestore(&t->vtime_seqlock, flags);
 
 
749}
750
751cputime_t task_gtime(struct task_struct *t)
752{
 
753	unsigned int seq;
754	cputime_t gtime;
 
 
 
755
756	do {
757		seq = read_seqbegin(&t->vtime_seqlock);
758
759		gtime = t->gtime;
760		if (t->flags & PF_VCPU)
761			gtime += vtime_delta(t);
762
763	} while (read_seqretry(&t->vtime_seqlock, seq));
764
765	return gtime;
766}
767
768/*
769 * Fetch cputime raw values from fields of task_struct and
770 * add up the pending nohz execution time since the last
771 * cputime snapshot.
772 */
773static void
774fetch_task_cputime(struct task_struct *t,
775		   cputime_t *u_dst, cputime_t *s_dst,
776		   cputime_t *u_src, cputime_t *s_src,
777		   cputime_t *udelta, cputime_t *sdelta)
778{
 
779	unsigned int seq;
780	unsigned long long delta;
 
 
 
 
 
 
781
782	do {
783		*udelta = 0;
784		*sdelta = 0;
785
786		seq = read_seqbegin(&t->vtime_seqlock);
787
788		if (u_dst)
789			*u_dst = *u_src;
790		if (s_dst)
791			*s_dst = *s_src;
792
793		/* Task is sleeping, nothing to add */
794		if (t->vtime_snap_whence == VTIME_SLEEPING ||
795		    is_idle_task(t))
796			continue;
797
798		delta = vtime_delta(t);
799
800		/*
801		 * Task runs either in user or kernel space, add pending nohz time to
802		 * the right place.
803		 */
804		if (t->vtime_snap_whence == VTIME_USER || t->flags & PF_VCPU) {
805			*udelta = delta;
806		} else {
807			if (t->vtime_snap_whence == VTIME_SYS)
808				*sdelta = delta;
809		}
810	} while (read_seqretry(&t->vtime_seqlock, seq));
811}
812
813
814void task_cputime(struct task_struct *t, cputime_t *utime, cputime_t *stime)
815{
816	cputime_t udelta, sdelta;
817
818	fetch_task_cputime(t, utime, stime, &t->utime,
819			   &t->stime, &udelta, &sdelta);
820	if (utime)
821		*utime += udelta;
822	if (stime)
823		*stime += sdelta;
824}
825
826void task_cputime_scaled(struct task_struct *t,
827			 cputime_t *utimescaled, cputime_t *stimescaled)
828{
829	cputime_t udelta, sdelta;
830
831	fetch_task_cputime(t, utimescaled, stimescaled,
832			   &t->utimescaled, &t->stimescaled, &udelta, &sdelta);
833	if (utimescaled)
834		*utimescaled += cputime_to_scaled(udelta);
835	if (stimescaled)
836		*stimescaled += cputime_to_scaled(sdelta);
837}
838#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */