Linux Audio

Check our new training course

Loading...
v4.10.11
  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(struct irqtime, cpu_irqtime);
 
 27
 
 28static int sched_clock_irqtime;
 29
 30void enable_sched_clock_irqtime(void)
 31{
 32	sched_clock_irqtime = 1;
 33}
 34
 35void disable_sched_clock_irqtime(void)
 36{
 37	sched_clock_irqtime = 0;
 38}
 39
 
 
 
 
 40/*
 41 * Called before incrementing preempt_count on {soft,}irq_enter
 42 * and before decrementing preempt_count on {soft,}irq_exit.
 43 */
 44void irqtime_account_irq(struct task_struct *curr)
 45{
 46	struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
 47	s64 delta;
 48	int cpu;
 49
 50	if (!sched_clock_irqtime)
 51		return;
 52
 
 
 53	cpu = smp_processor_id();
 54	delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
 55	irqtime->irq_start_time += delta;
 56
 57	u64_stats_update_begin(&irqtime->sync);
 58	/*
 59	 * We do not account for softirq time from ksoftirqd here.
 60	 * We want to continue accounting softirq time to ksoftirqd thread
 61	 * in that case, so as not to confuse scheduler with a special task
 62	 * that do not consume any time, but still wants to run.
 63	 */
 64	if (hardirq_count())
 65		irqtime->hardirq_time += delta;
 66	else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
 67		irqtime->softirq_time += delta;
 68
 69	u64_stats_update_end(&irqtime->sync);
 
 70}
 71EXPORT_SYMBOL_GPL(irqtime_account_irq);
 72
 73static cputime_t irqtime_account_update(u64 irqtime, int idx, cputime_t maxtime)
 74{
 75	u64 *cpustat = kcpustat_this_cpu->cpustat;
 76	cputime_t irq_cputime;
 77
 78	irq_cputime = nsecs_to_cputime64(irqtime) - cpustat[idx];
 79	irq_cputime = min(irq_cputime, maxtime);
 80	cpustat[idx] += irq_cputime;
 81
 82	return irq_cputime;
 
 
 
 
 
 83}
 84
 85static cputime_t irqtime_account_hi_update(cputime_t maxtime)
 86{
 87	return irqtime_account_update(__this_cpu_read(cpu_irqtime.hardirq_time),
 88				      CPUTIME_IRQ, maxtime);
 89}
 
 90
 91static cputime_t irqtime_account_si_update(cputime_t maxtime)
 92{
 93	return irqtime_account_update(__this_cpu_read(cpu_irqtime.softirq_time),
 94				      CPUTIME_SOFTIRQ, maxtime);
 
 
 95}
 96
 97#else /* CONFIG_IRQ_TIME_ACCOUNTING */
 98
 99#define sched_clock_irqtime	(0)
100
101static cputime_t irqtime_account_hi_update(cputime_t dummy)
102{
103	return 0;
104}
105
106static cputime_t irqtime_account_si_update(cputime_t dummy)
107{
108	return 0;
109}
110
111#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
112
113static inline void task_group_account_field(struct task_struct *p, int index,
114					    u64 tmp)
115{
116	/*
117	 * Since all updates are sure to touch the root cgroup, we
118	 * get ourselves ahead and touch it first. If the root cgroup
119	 * is the only cgroup, then nothing else should be necessary.
120	 *
121	 */
122	__this_cpu_add(kernel_cpustat.cpustat[index], tmp);
123
124	cpuacct_account_field(p, index, tmp);
125}
126
127/*
128 * Account user cpu time to a process.
129 * @p: the process that the cpu time gets accounted to
130 * @cputime: the cpu time spent in user space since the last update
 
131 */
132void account_user_time(struct task_struct *p, cputime_t cputime)
 
133{
134	int index;
135
136	/* Add user time to process. */
137	p->utime += cputime;
 
138	account_group_user_time(p, cputime);
139
140	index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
141
142	/* Add user time to cpustat. */
143	task_group_account_field(p, index, (__force u64) cputime);
144
145	/* Account for user time used */
146	acct_account_cputime(p);
147}
148
149/*
150 * Account guest cpu time to a process.
151 * @p: the process that the cpu time gets accounted to
152 * @cputime: the cpu time spent in virtual machine since the last update
 
153 */
154static void account_guest_time(struct task_struct *p, cputime_t cputime)
 
155{
156	u64 *cpustat = kcpustat_this_cpu->cpustat;
157
158	/* Add guest time to process. */
159	p->utime += cputime;
 
160	account_group_user_time(p, cputime);
161	p->gtime += cputime;
162
163	/* Add guest time to cpustat. */
164	if (task_nice(p) > 0) {
165		cpustat[CPUTIME_NICE] += (__force u64) cputime;
166		cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
167	} else {
168		cpustat[CPUTIME_USER] += (__force u64) cputime;
169		cpustat[CPUTIME_GUEST] += (__force u64) cputime;
170	}
171}
172
173/*
174 * Account system cpu time to a process and desired cpustat field
175 * @p: the process that the cpu time gets accounted to
176 * @cputime: the cpu time spent in kernel space since the last update
177 * @index: pointer to cpustat field that has to be updated
 
178 */
179static inline
180void __account_system_time(struct task_struct *p, cputime_t cputime, int index)
 
181{
182	/* Add system time to process. */
183	p->stime += cputime;
 
184	account_group_system_time(p, cputime);
185
186	/* Add system time to cpustat. */
187	task_group_account_field(p, index, (__force u64) cputime);
188
189	/* Account for system time used */
190	acct_account_cputime(p);
191}
192
193/*
194 * Account system cpu time to a process.
195 * @p: the process that the cpu time gets accounted to
196 * @hardirq_offset: the offset to subtract from hardirq_count()
197 * @cputime: the cpu time spent in kernel space since the last update
 
198 */
199void account_system_time(struct task_struct *p, int hardirq_offset,
200			 cputime_t cputime)
201{
202	int index;
203
204	if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
205		account_guest_time(p, cputime);
206		return;
207	}
208
209	if (hardirq_count() - hardirq_offset)
210		index = CPUTIME_IRQ;
211	else if (in_serving_softirq())
212		index = CPUTIME_SOFTIRQ;
213	else
214		index = CPUTIME_SYSTEM;
215
216	__account_system_time(p, cputime, index);
217}
218
219/*
220 * Account for involuntary wait time.
221 * @cputime: the cpu time spent in involuntary wait
222 */
223void account_steal_time(cputime_t cputime)
224{
225	u64 *cpustat = kcpustat_this_cpu->cpustat;
226
227	cpustat[CPUTIME_STEAL] += (__force u64) cputime;
228}
229
230/*
231 * Account for idle time.
232 * @cputime: the cpu time spent in idle wait
233 */
234void account_idle_time(cputime_t cputime)
235{
236	u64 *cpustat = kcpustat_this_cpu->cpustat;
237	struct rq *rq = this_rq();
238
239	if (atomic_read(&rq->nr_iowait) > 0)
240		cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
241	else
242		cpustat[CPUTIME_IDLE] += (__force u64) cputime;
243}
244
245/*
246 * When a guest is interrupted for a longer amount of time, missed clock
247 * ticks are not redelivered later. Due to that, this function may on
248 * occasion account more time than the calling functions think elapsed.
249 */
250static __always_inline cputime_t steal_account_process_time(cputime_t maxtime)
251{
252#ifdef CONFIG_PARAVIRT
253	if (static_key_false(&paravirt_steal_enabled)) {
254		cputime_t steal_cputime;
255		u64 steal;
 
256
257		steal = paravirt_steal_clock(smp_processor_id());
258		steal -= this_rq()->prev_steal_time;
259
260		steal_cputime = min(nsecs_to_cputime(steal), maxtime);
261		account_steal_time(steal_cputime);
262		this_rq()->prev_steal_time += cputime_to_nsecs(steal_cputime);
 
 
 
 
263
264		return steal_cputime;
 
265	}
266#endif
267	return 0;
268}
269
270/*
271 * Account how much elapsed time was spent in steal, irq, or softirq time.
272 */
273static inline cputime_t account_other_time(cputime_t max)
274{
275	cputime_t accounted;
276
277	/* Shall be converted to a lockdep-enabled lightweight check */
278	WARN_ON_ONCE(!irqs_disabled());
279
280	accounted = steal_account_process_time(max);
281
282	if (accounted < max)
283		accounted += irqtime_account_hi_update(max - accounted);
284
285	if (accounted < max)
286		accounted += irqtime_account_si_update(max - accounted);
287
288	return accounted;
289}
290
291#ifdef CONFIG_64BIT
292static inline u64 read_sum_exec_runtime(struct task_struct *t)
293{
294	return t->se.sum_exec_runtime;
295}
296#else
297static u64 read_sum_exec_runtime(struct task_struct *t)
298{
299	u64 ns;
300	struct rq_flags rf;
301	struct rq *rq;
302
303	rq = task_rq_lock(t, &rf);
304	ns = t->se.sum_exec_runtime;
305	task_rq_unlock(rq, t, &rf);
306
307	return ns;
308}
309#endif
310
311/*
312 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
313 * tasks (sum on group iteration) belonging to @tsk's group.
314 */
315void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
316{
317	struct signal_struct *sig = tsk->signal;
318	cputime_t utime, stime;
319	struct task_struct *t;
320	unsigned int seq, nextseq;
321	unsigned long flags;
322
323	/*
324	 * Update current task runtime to account pending time since last
325	 * scheduler action or thread_group_cputime() call. This thread group
326	 * might have other running tasks on different CPUs, but updating
327	 * their runtime can affect syscall performance, so we skip account
328	 * those pending times and rely only on values updated on tick or
329	 * other scheduler action.
330	 */
331	if (same_thread_group(current, tsk))
332		(void) task_sched_runtime(current);
333
334	rcu_read_lock();
335	/* Attempt a lockless read on the first round. */
336	nextseq = 0;
 
 
 
337	do {
338		seq = nextseq;
339		flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
340		times->utime = sig->utime;
341		times->stime = sig->stime;
342		times->sum_exec_runtime = sig->sum_sched_runtime;
343
344		for_each_thread(tsk, t) {
345			task_cputime(t, &utime, &stime);
346			times->utime += utime;
347			times->stime += stime;
348			times->sum_exec_runtime += read_sum_exec_runtime(t);
349		}
350		/* If lockless access failed, take the lock. */
351		nextseq = 1;
352	} while (need_seqretry(&sig->stats_lock, seq));
353	done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
354	rcu_read_unlock();
355}
356
357#ifdef CONFIG_IRQ_TIME_ACCOUNTING
358/*
359 * Account a tick to a process and cpustat
360 * @p: the process that the cpu time gets accounted to
361 * @user_tick: is the tick from userspace
362 * @rq: the pointer to rq
363 *
364 * Tick demultiplexing follows the order
365 * - pending hardirq update
366 * - pending softirq update
367 * - user_time
368 * - idle_time
369 * - system time
370 *   - check for guest_time
371 *   - else account as system_time
372 *
373 * Check for hardirq is done both for system and user time as there is
374 * no timer going off while we are on hardirq and hence we may never get an
375 * opportunity to update it solely in system time.
376 * p->stime and friends are only updated on system time and not on irq
377 * softirq as those do not count in task exec_runtime any more.
378 */
379static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
380					 struct rq *rq, int ticks)
381{
382	u64 cputime = (__force u64) cputime_one_jiffy * ticks;
383	cputime_t other;
 
384
385	/*
386	 * When returning from idle, many ticks can get accounted at
387	 * once, including some ticks of steal, irq, and softirq time.
388	 * Subtract those ticks from the amount of time accounted to
389	 * idle, or potentially user or system time. Due to rounding,
390	 * other time can exceed ticks occasionally.
391	 */
392	other = account_other_time(ULONG_MAX);
393	if (other >= cputime)
394		return;
395	cputime -= other;
396
397	if (this_cpu_ksoftirqd() == p) {
 
 
 
 
 
 
 
398		/*
399		 * ksoftirqd time do not get accounted in cpu_softirq_time.
400		 * So, we have to handle it separately here.
401		 * Also, p->stime needs to be updated for ksoftirqd.
402		 */
403		__account_system_time(p, cputime, CPUTIME_SOFTIRQ);
404	} else if (user_tick) {
405		account_user_time(p, cputime);
406	} else if (p == rq->idle) {
407		account_idle_time(cputime);
408	} else if (p->flags & PF_VCPU) { /* System time or guest time */
409		account_guest_time(p, cputime);
410	} else {
411		__account_system_time(p, cputime, CPUTIME_SYSTEM);
412	}
413}
414
415static void irqtime_account_idle_ticks(int ticks)
416{
417	struct rq *rq = this_rq();
418
419	irqtime_account_process_tick(current, 0, rq, ticks);
420}
421#else /* CONFIG_IRQ_TIME_ACCOUNTING */
422static inline void irqtime_account_idle_ticks(int ticks) {}
423static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
424						struct rq *rq, int nr_ticks) {}
425#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
426
427/*
428 * Use precise platform statistics if available:
429 */
430#ifdef CONFIG_VIRT_CPU_ACCOUNTING
431
432#ifndef __ARCH_HAS_VTIME_TASK_SWITCH
433void vtime_common_task_switch(struct task_struct *prev)
434{
435	if (is_idle_task(prev))
436		vtime_account_idle(prev);
437	else
438		vtime_account_system(prev);
439
440#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
441	vtime_account_user(prev);
442#endif
443	arch_vtime_task_switch(prev);
444}
445#endif
446
447#endif /* CONFIG_VIRT_CPU_ACCOUNTING */
448
449
450#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
451/*
452 * Archs that account the whole time spent in the idle task
453 * (outside irq) as idle time can rely on this and just implement
454 * vtime_account_system() and vtime_account_idle(). Archs that
455 * have other meaning of the idle time (s390 only includes the
456 * time spent by the CPU when it's in low power mode) must override
457 * vtime_account().
458 */
459#ifndef __ARCH_HAS_VTIME_ACCOUNT
460void vtime_account_irq_enter(struct task_struct *tsk)
461{
462	if (!in_interrupt() && is_idle_task(tsk))
463		vtime_account_idle(tsk);
464	else
465		vtime_account_system(tsk);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
466}
467EXPORT_SYMBOL_GPL(vtime_account_irq_enter);
468#endif /* __ARCH_HAS_VTIME_ACCOUNT */
 
 
469
 
470void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
471{
472	*ut = p->utime;
473	*st = p->stime;
474}
475EXPORT_SYMBOL_GPL(task_cputime_adjusted);
476
477void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
478{
479	struct task_cputime cputime;
480
481	thread_group_cputime(p, &cputime);
482
483	*ut = cputime.utime;
484	*st = cputime.stime;
485}
486#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
487/*
488 * Account a single tick of cpu time.
489 * @p: the process that the cpu time gets accounted to
490 * @user_tick: indicates if the tick is a user or a system tick
491 */
492void account_process_tick(struct task_struct *p, int user_tick)
493{
494	cputime_t cputime, steal;
495	struct rq *rq = this_rq();
496
497	if (vtime_accounting_cpu_enabled())
498		return;
499
500	if (sched_clock_irqtime) {
501		irqtime_account_process_tick(p, user_tick, rq, 1);
502		return;
503	}
504
505	cputime = cputime_one_jiffy;
506	steal = steal_account_process_time(ULONG_MAX);
507
508	if (steal >= cputime)
509		return;
510
511	cputime -= steal;
512
513	if (user_tick)
514		account_user_time(p, cputime);
515	else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
516		account_system_time(p, HARDIRQ_OFFSET, cputime);
 
517	else
518		account_idle_time(cputime);
 
 
 
 
 
 
 
 
 
 
519}
520
521/*
522 * Account multiple ticks of idle time.
523 * @ticks: number of stolen ticks
524 */
525void account_idle_ticks(unsigned long ticks)
526{
527	cputime_t cputime, steal;
528
529	if (sched_clock_irqtime) {
530		irqtime_account_idle_ticks(ticks);
531		return;
532	}
533
534	cputime = jiffies_to_cputime(ticks);
535	steal = steal_account_process_time(ULONG_MAX);
536
537	if (steal >= cputime)
538		return;
539
540	cputime -= steal;
541	account_idle_time(cputime);
542}
543
544/*
545 * Perform (stime * rtime) / total, but avoid multiplication overflow by
546 * loosing precision when the numbers are big.
547 */
548static cputime_t scale_stime(u64 stime, u64 rtime, u64 total)
549{
550	u64 scaled;
551
552	for (;;) {
553		/* Make sure "rtime" is the bigger of stime/rtime */
554		if (stime > rtime)
555			swap(rtime, stime);
556
557		/* Make sure 'total' fits in 32 bits */
558		if (total >> 32)
559			goto drop_precision;
560
561		/* Does rtime (and thus stime) fit in 32 bits? */
562		if (!(rtime >> 32))
563			break;
564
565		/* Can we just balance rtime/stime rather than dropping bits? */
566		if (stime >> 31)
567			goto drop_precision;
568
569		/* We can grow stime and shrink rtime and try to make them both fit */
570		stime <<= 1;
571		rtime >>= 1;
572		continue;
573
574drop_precision:
575		/* We drop from rtime, it has more bits than stime */
576		rtime >>= 1;
577		total >>= 1;
578	}
579
580	/*
581	 * Make sure gcc understands that this is a 32x32->64 multiply,
582	 * followed by a 64/32->64 divide.
583	 */
584	scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
585	return (__force cputime_t) scaled;
586}
587
588/*
589 * Adjust tick based cputime random precision against scheduler runtime
590 * accounting.
591 *
592 * Tick based cputime accounting depend on random scheduling timeslices of a
593 * task to be interrupted or not by the timer.  Depending on these
594 * circumstances, the number of these interrupts may be over or
595 * under-optimistic, matching the real user and system cputime with a variable
596 * precision.
597 *
598 * Fix this by scaling these tick based values against the total runtime
599 * accounted by the CFS scheduler.
600 *
601 * This code provides the following guarantees:
602 *
603 *   stime + utime == rtime
604 *   stime_i+1 >= stime_i, utime_i+1 >= utime_i
605 *
606 * Assuming that rtime_i+1 >= rtime_i.
607 */
608static void cputime_adjust(struct task_cputime *curr,
609			   struct prev_cputime *prev,
610			   cputime_t *ut, cputime_t *st)
611{
612	cputime_t rtime, stime, utime;
613	unsigned long flags;
614
615	/* Serialize concurrent callers such that we can honour our guarantees */
616	raw_spin_lock_irqsave(&prev->lock, flags);
 
 
 
 
 
 
 
 
617	rtime = nsecs_to_cputime(curr->sum_exec_runtime);
618
619	/*
620	 * This is possible under two circumstances:
621	 *  - rtime isn't monotonic after all (a bug);
622	 *  - we got reordered by the lock.
623	 *
624	 * In both cases this acts as a filter such that the rest of the code
625	 * can assume it is monotonic regardless of anything else.
626	 */
627	if (prev->stime + prev->utime >= rtime)
628		goto out;
629
630	stime = curr->stime;
631	utime = curr->utime;
632
633	/*
634	 * If either stime or both stime and utime are 0, assume all runtime is
635	 * userspace. Once a task gets some ticks, the monotonicy code at
636	 * 'update' will ensure things converge to the observed ratio.
637	 */
638	if (stime == 0) {
639		utime = rtime;
640		goto update;
641	}
642
643	if (utime == 0) {
644		stime = rtime;
645		goto update;
646	}
647
648	stime = scale_stime((__force u64)stime, (__force u64)rtime,
649			    (__force u64)(stime + utime));
650
651update:
652	/*
653	 * Make sure stime doesn't go backwards; this preserves monotonicity
654	 * for utime because rtime is monotonic.
655	 *
656	 *  utime_i+1 = rtime_i+1 - stime_i
657	 *            = rtime_i+1 - (rtime_i - utime_i)
658	 *            = (rtime_i+1 - rtime_i) + utime_i
659	 *            >= utime_i
660	 */
661	if (stime < prev->stime)
662		stime = prev->stime;
663	utime = rtime - stime;
664
665	/*
666	 * Make sure utime doesn't go backwards; this still preserves
667	 * monotonicity for stime, analogous argument to above.
 
668	 */
669	if (utime < prev->utime) {
670		utime = prev->utime;
671		stime = rtime - utime;
672	}
673
674	prev->stime = stime;
675	prev->utime = utime;
676out:
677	*ut = prev->utime;
678	*st = prev->stime;
679	raw_spin_unlock_irqrestore(&prev->lock, flags);
680}
681
682void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
683{
684	struct task_cputime cputime = {
685		.sum_exec_runtime = p->se.sum_exec_runtime,
686	};
687
688	task_cputime(p, &cputime.utime, &cputime.stime);
689	cputime_adjust(&cputime, &p->prev_cputime, ut, st);
690}
691EXPORT_SYMBOL_GPL(task_cputime_adjusted);
692
 
 
 
693void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
694{
695	struct task_cputime cputime;
696
697	thread_group_cputime(p, &cputime);
698	cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
699}
700#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
701
702#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
703static cputime_t vtime_delta(struct task_struct *tsk)
704{
705	unsigned long now = READ_ONCE(jiffies);
706
707	if (time_before(now, (unsigned long)tsk->vtime_snap))
 
708		return 0;
709
710	return jiffies_to_cputime(now - tsk->vtime_snap);
711}
712
713static cputime_t get_vtime_delta(struct task_struct *tsk)
714{
715	unsigned long now = READ_ONCE(jiffies);
716	cputime_t delta, other;
717
718	/*
719	 * Unlike tick based timing, vtime based timing never has lost
720	 * ticks, and no need for steal time accounting to make up for
721	 * lost ticks. Vtime accounts a rounded version of actual
722	 * elapsed time. Limit account_other_time to prevent rounding
723	 * errors from causing elapsed vtime to go negative.
724	 */
725	delta = jiffies_to_cputime(now - tsk->vtime_snap);
726	other = account_other_time(delta);
727	WARN_ON_ONCE(tsk->vtime_snap_whence == VTIME_INACTIVE);
728	tsk->vtime_snap = now;
729
730	return delta - other;
 
731}
732
733static void __vtime_account_system(struct task_struct *tsk)
734{
735	cputime_t delta_cpu = get_vtime_delta(tsk);
736
737	account_system_time(tsk, irq_count(), delta_cpu);
738}
739
740void vtime_account_system(struct task_struct *tsk)
741{
742	if (!vtime_delta(tsk))
743		return;
 
 
744
745	write_seqcount_begin(&tsk->vtime_seqcount);
 
 
746	__vtime_account_system(tsk);
747	write_seqcount_end(&tsk->vtime_seqcount);
 
 
748}
749
750void vtime_account_user(struct task_struct *tsk)
751{
752	cputime_t delta_cpu;
753
754	write_seqcount_begin(&tsk->vtime_seqcount);
 
755	tsk->vtime_snap_whence = VTIME_SYS;
756	if (vtime_delta(tsk)) {
757		delta_cpu = get_vtime_delta(tsk);
758		account_user_time(tsk, delta_cpu);
759	}
760	write_seqcount_end(&tsk->vtime_seqcount);
761}
762
763void vtime_user_enter(struct task_struct *tsk)
764{
765	write_seqcount_begin(&tsk->vtime_seqcount);
766	if (vtime_delta(tsk))
767		__vtime_account_system(tsk);
768	tsk->vtime_snap_whence = VTIME_USER;
769	write_seqcount_end(&tsk->vtime_seqcount);
770}
771
772void vtime_guest_enter(struct task_struct *tsk)
773{
774	/*
775	 * The flags must be updated under the lock with
776	 * the vtime_snap flush and update.
777	 * That enforces a right ordering and update sequence
778	 * synchronization against the reader (task_gtime())
779	 * that can thus safely catch up with a tickless delta.
780	 */
781	write_seqcount_begin(&tsk->vtime_seqcount);
782	if (vtime_delta(tsk))
783		__vtime_account_system(tsk);
784	current->flags |= PF_VCPU;
785	write_seqcount_end(&tsk->vtime_seqcount);
786}
787EXPORT_SYMBOL_GPL(vtime_guest_enter);
788
789void vtime_guest_exit(struct task_struct *tsk)
790{
791	write_seqcount_begin(&tsk->vtime_seqcount);
792	__vtime_account_system(tsk);
793	current->flags &= ~PF_VCPU;
794	write_seqcount_end(&tsk->vtime_seqcount);
795}
796EXPORT_SYMBOL_GPL(vtime_guest_exit);
797
798void vtime_account_idle(struct task_struct *tsk)
799{
800	cputime_t delta_cpu = get_vtime_delta(tsk);
801
802	account_idle_time(delta_cpu);
803}
804
805void arch_vtime_task_switch(struct task_struct *prev)
806{
807	write_seqcount_begin(&prev->vtime_seqcount);
808	prev->vtime_snap_whence = VTIME_INACTIVE;
809	write_seqcount_end(&prev->vtime_seqcount);
810
811	write_seqcount_begin(&current->vtime_seqcount);
812	current->vtime_snap_whence = VTIME_SYS;
813	current->vtime_snap = jiffies;
814	write_seqcount_end(&current->vtime_seqcount);
815}
816
817void vtime_init_idle(struct task_struct *t, int cpu)
818{
819	unsigned long flags;
820
821	local_irq_save(flags);
822	write_seqcount_begin(&t->vtime_seqcount);
823	t->vtime_snap_whence = VTIME_SYS;
824	t->vtime_snap = jiffies;
825	write_seqcount_end(&t->vtime_seqcount);
826	local_irq_restore(flags);
827}
828
829cputime_t task_gtime(struct task_struct *t)
830{
831	unsigned int seq;
832	cputime_t gtime;
833
834	if (!vtime_accounting_enabled())
835		return t->gtime;
836
837	do {
838		seq = read_seqcount_begin(&t->vtime_seqcount);
839
840		gtime = t->gtime;
841		if (t->vtime_snap_whence == VTIME_SYS && t->flags & PF_VCPU)
842			gtime += vtime_delta(t);
843
844	} while (read_seqcount_retry(&t->vtime_seqcount, seq));
845
846	return gtime;
847}
848
849/*
850 * Fetch cputime raw values from fields of task_struct and
851 * add up the pending nohz execution time since the last
852 * cputime snapshot.
853 */
854void task_cputime(struct task_struct *t, cputime_t *utime, cputime_t *stime)
 
 
 
 
855{
856	cputime_t delta;
857	unsigned int seq;
858
859	if (!vtime_accounting_enabled()) {
860		*utime = t->utime;
861		*stime = t->stime;
862		return;
863	}
864
865	do {
866		seq = read_seqcount_begin(&t->vtime_seqcount);
 
867
868		*utime = t->utime;
869		*stime = t->stime;
 
 
 
 
870
871		/* Task is sleeping, nothing to add */
872		if (t->vtime_snap_whence == VTIME_INACTIVE || is_idle_task(t))
 
873			continue;
874
875		delta = vtime_delta(t);
876
877		/*
878		 * Task runs either in user or kernel space, add pending nohz time to
879		 * the right place.
880		 */
881		if (t->vtime_snap_whence == VTIME_USER || t->flags & PF_VCPU)
882			*utime += delta;
883		else if (t->vtime_snap_whence == VTIME_SYS)
884			*stime += delta;
885	} while (read_seqcount_retry(&t->vtime_seqcount, seq));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
886}
887#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 */