Linux Audio

Check our new training course

Loading...
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 */
v6.2
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Simple CPU accounting cgroup controller
   4 */
 
 
 
 
   5
   6#ifdef CONFIG_IRQ_TIME_ACCOUNTING
   7
   8/*
   9 * There are no locks covering percpu hardirq/softirq time.
  10 * They are only modified in vtime_account, on corresponding CPU
  11 * with interrupts disabled. So, writes are safe.
  12 * They are read and saved off onto struct rq in update_rq_clock().
  13 * This may result in other CPU reading this CPU's irq time and can
  14 * race with irq/vtime_account on this CPU. We would either get old
  15 * or new value with a side effect of accounting a slice of irq time to wrong
  16 * task when irq is in progress while we read rq->clock. That is a worthy
  17 * compromise in place of having locks on each irq in account_system_time.
  18 */
  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 after 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, unsigned int offset)
  50{
  51	struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
  52	unsigned int pc;
  53	s64 delta;
  54	int cpu;
  55
  56	if (!sched_clock_irqtime)
  57		return;
  58
 
 
  59	cpu = smp_processor_id();
  60	delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
  61	irqtime->irq_start_time += delta;
  62	pc = irq_count() - offset;
  63
 
  64	/*
  65	 * We do not account for softirq time from ksoftirqd here.
  66	 * We want to continue accounting softirq time to ksoftirqd thread
  67	 * in that case, so as not to confuse scheduler with a special task
  68	 * that do not consume any time, but still wants to run.
  69	 */
  70	if (pc & HARDIRQ_MASK)
  71		irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
  72	else if ((pc & SOFTIRQ_OFFSET) && curr != this_cpu_ksoftirqd())
  73		irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
 
 
 
  74}
 
  75
  76static u64 irqtime_tick_accounted(u64 maxtime)
  77{
  78	struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
  79	u64 delta;
 
 
 
 
 
 
 
 
 
 
  80
  81	delta = min(irqtime->tick_delta, maxtime);
  82	irqtime->tick_delta -= delta;
 
 
 
 
  83
  84	return delta;
 
 
 
 
 
  85}
  86
  87#else /* CONFIG_IRQ_TIME_ACCOUNTING */
  88
  89#define sched_clock_irqtime	(0)
  90
  91static u64 irqtime_tick_accounted(u64 dummy)
  92{
  93	return 0;
  94}
  95
  96#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
  97
  98static inline void task_group_account_field(struct task_struct *p, int index,
  99					    u64 tmp)
 100{
 101	/*
 102	 * Since all updates are sure to touch the root cgroup, we
 103	 * get ourselves ahead and touch it first. If the root cgroup
 104	 * is the only cgroup, then nothing else should be necessary.
 105	 *
 106	 */
 107	__this_cpu_add(kernel_cpustat.cpustat[index], tmp);
 108
 109	cgroup_account_cputime_field(p, index, tmp);
 110}
 111
 112/*
 113 * Account user CPU time to a process.
 114 * @p: the process that the CPU time gets accounted to
 115 * @cputime: the CPU time spent in user space since the last update
 
 116 */
 117void account_user_time(struct task_struct *p, u64 cputime)
 
 118{
 119	int index;
 120
 121	/* Add user time to process. */
 122	p->utime += cputime;
 
 123	account_group_user_time(p, cputime);
 124
 125	index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
 126
 127	/* Add user time to cpustat. */
 128	task_group_account_field(p, index, cputime);
 129
 130	/* Account for user time used */
 131	acct_account_cputime(p);
 132}
 133
 134/*
 135 * Account guest CPU time to a process.
 136 * @p: the process that the CPU time gets accounted to
 137 * @cputime: the CPU time spent in virtual machine since the last update
 
 138 */
 139void account_guest_time(struct task_struct *p, u64 cputime)
 
 140{
 141	u64 *cpustat = kcpustat_this_cpu->cpustat;
 142
 143	/* Add guest time to process. */
 144	p->utime += cputime;
 
 145	account_group_user_time(p, cputime);
 146	p->gtime += cputime;
 147
 148	/* Add guest time to cpustat. */
 149	if (task_nice(p) > 0) {
 150		task_group_account_field(p, CPUTIME_NICE, cputime);
 151		cpustat[CPUTIME_GUEST_NICE] += cputime;
 152	} else {
 153		task_group_account_field(p, CPUTIME_USER, cputime);
 154		cpustat[CPUTIME_GUEST] += cputime;
 155	}
 156}
 157
 158/*
 159 * Account system CPU time to a process and desired cpustat field
 160 * @p: the process that the CPU time gets accounted to
 161 * @cputime: the CPU time spent in kernel space since the last update
 162 * @index: pointer to cpustat field that has to be updated
 
 163 */
 164void account_system_index_time(struct task_struct *p,
 165			       u64 cputime, enum cpu_usage_stat index)
 
 166{
 167	/* Add system time to process. */
 168	p->stime += cputime;
 
 169	account_group_system_time(p, cputime);
 170
 171	/* Add system time to cpustat. */
 172	task_group_account_field(p, index, cputime);
 173
 174	/* Account for system time used */
 175	acct_account_cputime(p);
 176}
 177
 178/*
 179 * Account system CPU time to a process.
 180 * @p: the process that the CPU time gets accounted to
 181 * @hardirq_offset: the offset to subtract from hardirq_count()
 182 * @cputime: the CPU time spent in kernel space since the last update
 
 183 */
 184void account_system_time(struct task_struct *p, int hardirq_offset, u64 cputime)
 
 185{
 186	int index;
 187
 188	if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
 189		account_guest_time(p, cputime);
 190		return;
 191	}
 192
 193	if (hardirq_count() - hardirq_offset)
 194		index = CPUTIME_IRQ;
 195	else if (in_serving_softirq())
 196		index = CPUTIME_SOFTIRQ;
 197	else
 198		index = CPUTIME_SYSTEM;
 199
 200	account_system_index_time(p, cputime, index);
 201}
 202
 203/*
 204 * Account for involuntary wait time.
 205 * @cputime: the CPU time spent in involuntary wait
 206 */
 207void account_steal_time(u64 cputime)
 208{
 209	u64 *cpustat = kcpustat_this_cpu->cpustat;
 210
 211	cpustat[CPUTIME_STEAL] += cputime;
 212}
 213
 214/*
 215 * Account for idle time.
 216 * @cputime: the CPU time spent in idle wait
 217 */
 218void account_idle_time(u64 cputime)
 219{
 220	u64 *cpustat = kcpustat_this_cpu->cpustat;
 221	struct rq *rq = this_rq();
 222
 223	if (atomic_read(&rq->nr_iowait) > 0)
 224		cpustat[CPUTIME_IOWAIT] += cputime;
 225	else
 226		cpustat[CPUTIME_IDLE] += cputime;
 227}
 228
 229
 230#ifdef CONFIG_SCHED_CORE
 231/*
 232 * Account for forceidle time due to core scheduling.
 233 *
 234 * REQUIRES: schedstat is enabled.
 235 */
 236void __account_forceidle_time(struct task_struct *p, u64 delta)
 237{
 238	__schedstat_add(p->stats.core_forceidle_sum, delta);
 239
 240	task_group_account_field(p, CPUTIME_FORCEIDLE, delta);
 241}
 242#endif
 243
 244/*
 245 * When a guest is interrupted for a longer amount of time, missed clock
 246 * ticks are not redelivered later. Due to that, this function may on
 247 * occasion account more time than the calling functions think elapsed.
 248 */
 249static __always_inline u64 steal_account_process_time(u64 maxtime)
 250{
 251#ifdef CONFIG_PARAVIRT
 252	if (static_key_false(&paravirt_steal_enabled)) {
 253		u64 steal;
 
 254
 255		steal = paravirt_steal_clock(smp_processor_id());
 256		steal -= this_rq()->prev_steal_time;
 257		steal = min(steal, maxtime);
 258		account_steal_time(steal);
 259		this_rq()->prev_steal_time += steal;
 260
 261		return steal;
 
 
 
 
 
 
 
 
 
 262	}
 263#endif
 264	return 0;
 265}
 266
 267/*
 268 * Account how much elapsed time was spent in steal, irq, or softirq time.
 269 */
 270static inline u64 account_other_time(u64 max)
 271{
 272	u64 accounted;
 273
 274	lockdep_assert_irqs_disabled();
 275
 276	accounted = steal_account_process_time(max);
 277
 278	if (accounted < max)
 279		accounted += irqtime_tick_accounted(max - accounted);
 280
 281	return accounted;
 282}
 283
 284#ifdef CONFIG_64BIT
 285static inline u64 read_sum_exec_runtime(struct task_struct *t)
 286{
 287	return t->se.sum_exec_runtime;
 288}
 289#else
 290static u64 read_sum_exec_runtime(struct task_struct *t)
 291{
 292	u64 ns;
 293	struct rq_flags rf;
 294	struct rq *rq;
 295
 296	rq = task_rq_lock(t, &rf);
 297	ns = t->se.sum_exec_runtime;
 298	task_rq_unlock(rq, t, &rf);
 299
 300	return ns;
 301}
 302#endif
 303
 304/*
 305 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
 306 * tasks (sum on group iteration) belonging to @tsk's group.
 307 */
 308void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
 309{
 310	struct signal_struct *sig = tsk->signal;
 311	u64 utime, stime;
 312	struct task_struct *t;
 313	unsigned int seq, nextseq;
 314	unsigned long flags;
 315
 316	/*
 317	 * Update current task runtime to account pending time since last
 318	 * scheduler action or thread_group_cputime() call. This thread group
 319	 * might have other running tasks on different CPUs, but updating
 320	 * their runtime can affect syscall performance, so we skip account
 321	 * those pending times and rely only on values updated on tick or
 322	 * other scheduler action.
 323	 */
 324	if (same_thread_group(current, tsk))
 325		(void) task_sched_runtime(current);
 326
 327	rcu_read_lock();
 328	/* Attempt a lockless read on the first round. */
 329	nextseq = 0;
 
 
 
 330	do {
 331		seq = nextseq;
 332		flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
 333		times->utime = sig->utime;
 334		times->stime = sig->stime;
 335		times->sum_exec_runtime = sig->sum_sched_runtime;
 336
 337		for_each_thread(tsk, t) {
 338			task_cputime(t, &utime, &stime);
 339			times->utime += utime;
 340			times->stime += stime;
 341			times->sum_exec_runtime += read_sum_exec_runtime(t);
 342		}
 343		/* If lockless access failed, take the lock. */
 344		nextseq = 1;
 345	} while (need_seqretry(&sig->stats_lock, seq));
 346	done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
 347	rcu_read_unlock();
 348}
 349
 350#ifdef CONFIG_IRQ_TIME_ACCOUNTING
 351/*
 352 * Account a tick to a process and cpustat
 353 * @p: the process that the CPU time gets accounted to
 354 * @user_tick: is the tick from userspace
 355 * @rq: the pointer to rq
 356 *
 357 * Tick demultiplexing follows the order
 358 * - pending hardirq update
 359 * - pending softirq update
 360 * - user_time
 361 * - idle_time
 362 * - system time
 363 *   - check for guest_time
 364 *   - else account as system_time
 365 *
 366 * Check for hardirq is done both for system and user time as there is
 367 * no timer going off while we are on hardirq and hence we may never get an
 368 * opportunity to update it solely in system time.
 369 * p->stime and friends are only updated on system time and not on irq
 370 * softirq as those do not count in task exec_runtime any more.
 371 */
 372static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
 373					 int ticks)
 374{
 375	u64 other, cputime = TICK_NSEC * ticks;
 
 
 376
 377	/*
 378	 * When returning from idle, many ticks can get accounted at
 379	 * once, including some ticks of steal, irq, and softirq time.
 380	 * Subtract those ticks from the amount of time accounted to
 381	 * idle, or potentially user or system time. Due to rounding,
 382	 * other time can exceed ticks occasionally.
 383	 */
 384	other = account_other_time(ULONG_MAX);
 385	if (other >= cputime)
 386		return;
 387
 388	cputime -= other;
 
 389
 390	if (this_cpu_ksoftirqd() == p) {
 
 
 
 
 391		/*
 392		 * ksoftirqd time do not get accounted in cpu_softirq_time.
 393		 * So, we have to handle it separately here.
 394		 * Also, p->stime needs to be updated for ksoftirqd.
 395		 */
 396		account_system_index_time(p, cputime, CPUTIME_SOFTIRQ);
 397	} else if (user_tick) {
 398		account_user_time(p, cputime);
 399	} else if (p == this_rq()->idle) {
 400		account_idle_time(cputime);
 401	} else if (p->flags & PF_VCPU) { /* System time or guest time */
 402		account_guest_time(p, cputime);
 403	} else {
 404		account_system_index_time(p, cputime, CPUTIME_SYSTEM);
 405	}
 406}
 407
 408static void irqtime_account_idle_ticks(int ticks)
 409{
 410	irqtime_account_process_tick(current, 0, ticks);
 
 
 411}
 412#else /* CONFIG_IRQ_TIME_ACCOUNTING */
 413static inline void irqtime_account_idle_ticks(int ticks) { }
 414static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
 415						int nr_ticks) { }
 416#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
 417
 418/*
 419 * Use precise platform statistics if available:
 420 */
 421#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
 422
 423# ifndef __ARCH_HAS_VTIME_TASK_SWITCH
 424void vtime_task_switch(struct task_struct *prev)
 425{
 426	if (is_idle_task(prev))
 427		vtime_account_idle(prev);
 428	else
 429		vtime_account_kernel(prev);
 430
 431	vtime_flush(prev);
 
 
 432	arch_vtime_task_switch(prev);
 433}
 434# endif
 435
 436void vtime_account_irq(struct task_struct *tsk, unsigned int offset)
 
 
 
 
 
 
 
 
 
 437{
 438	unsigned int pc = irq_count() - offset;
 
 
 
 
 
 
 
 
 
 
 439
 440	if (pc & HARDIRQ_OFFSET) {
 441		vtime_account_hardirq(tsk);
 442	} else if (pc & SOFTIRQ_OFFSET) {
 443		vtime_account_softirq(tsk);
 444	} else if (!IS_ENABLED(CONFIG_HAVE_VIRT_CPU_ACCOUNTING_IDLE) &&
 445		   is_idle_task(tsk)) {
 446		vtime_account_idle(tsk);
 447	} else {
 448		vtime_account_kernel(tsk);
 449	}
 
 450}
 
 
 
 451
 452void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
 453		    u64 *ut, u64 *st)
 454{
 455	*ut = curr->utime;
 456	*st = curr->stime;
 457}
 458
 459void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
 
 460{
 461	*ut = p->utime;
 462	*st = p->stime;
 463}
 464EXPORT_SYMBOL_GPL(task_cputime_adjusted);
 465
 466void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
 467{
 468	struct task_cputime cputime;
 469
 470	thread_group_cputime(p, &cputime);
 471
 472	*ut = cputime.utime;
 473	*st = cputime.stime;
 474}
 475
 476#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE: */
 477
 478/*
 479 * Account a single tick of CPU time.
 480 * @p: the process that the CPU time gets accounted to
 481 * @user_tick: indicates if the tick is a user or a system tick
 482 */
 483void account_process_tick(struct task_struct *p, int user_tick)
 484{
 485	u64 cputime, steal;
 
 486
 487	if (vtime_accounting_enabled_this_cpu())
 488		return;
 489
 490	if (sched_clock_irqtime) {
 491		irqtime_account_process_tick(p, user_tick, 1);
 492		return;
 493	}
 494
 495	cputime = TICK_NSEC;
 496	steal = steal_account_process_time(ULONG_MAX);
 497
 498	if (steal >= cputime)
 499		return;
 500
 501	cputime -= steal;
 502
 503	if (user_tick)
 504		account_user_time(p, cputime);
 505	else if ((p != this_rq()->idle) || (irq_count() != HARDIRQ_OFFSET))
 506		account_system_time(p, HARDIRQ_OFFSET, cputime);
 
 507	else
 508		account_idle_time(cputime);
 
 
 
 
 
 
 
 
 
 
 509}
 510
 511/*
 512 * Account multiple ticks of idle time.
 513 * @ticks: number of stolen ticks
 514 */
 515void account_idle_ticks(unsigned long ticks)
 516{
 517	u64 cputime, steal;
 518
 519	if (sched_clock_irqtime) {
 520		irqtime_account_idle_ticks(ticks);
 521		return;
 522	}
 523
 524	cputime = ticks * TICK_NSEC;
 525	steal = steal_account_process_time(ULONG_MAX);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 526
 527	if (steal >= cputime)
 528		return;
 
 
 
 
 
 
 
 
 529
 530	cputime -= steal;
 531	account_idle_time(cputime);
 
 
 
 
 532}
 533
 534/*
 535 * Adjust tick based cputime random precision against scheduler runtime
 536 * accounting.
 537 *
 538 * Tick based cputime accounting depend on random scheduling timeslices of a
 539 * task to be interrupted or not by the timer.  Depending on these
 540 * circumstances, the number of these interrupts may be over or
 541 * under-optimistic, matching the real user and system cputime with a variable
 542 * precision.
 543 *
 544 * Fix this by scaling these tick based values against the total runtime
 545 * accounted by the CFS scheduler.
 546 *
 547 * This code provides the following guarantees:
 548 *
 549 *   stime + utime == rtime
 550 *   stime_i+1 >= stime_i, utime_i+1 >= utime_i
 551 *
 552 * Assuming that rtime_i+1 >= rtime_i.
 553 */
 554void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
 555		    u64 *ut, u64 *st)
 
 556{
 557	u64 rtime, stime, utime;
 558	unsigned long flags;
 559
 560	/* Serialize concurrent callers such that we can honour our guarantees */
 561	raw_spin_lock_irqsave(&prev->lock, flags);
 562	rtime = curr->sum_exec_runtime;
 
 
 
 
 
 
 
 
 563
 564	/*
 565	 * This is possible under two circumstances:
 566	 *  - rtime isn't monotonic after all (a bug);
 567	 *  - we got reordered by the lock.
 568	 *
 569	 * In both cases this acts as a filter such that the rest of the code
 570	 * can assume it is monotonic regardless of anything else.
 571	 */
 572	if (prev->stime + prev->utime >= rtime)
 573		goto out;
 574
 575	stime = curr->stime;
 576	utime = curr->utime;
 577
 578	/*
 579	 * If either stime or utime are 0, assume all runtime is userspace.
 580	 * Once a task gets some ticks, the monotonicity code at 'update:'
 581	 * will ensure things converge to the observed ratio.
 582	 */
 583	if (stime == 0) {
 584		utime = rtime;
 585		goto update;
 586	}
 587
 588	if (utime == 0) {
 589		stime = rtime;
 590		goto update;
 591	}
 592
 593	stime = mul_u64_u64_div_u64(stime, rtime, stime + utime);
 594
 595update:
 596	/*
 597	 * Make sure stime doesn't go backwards; this preserves monotonicity
 598	 * for utime because rtime is monotonic.
 599	 *
 600	 *  utime_i+1 = rtime_i+1 - stime_i
 601	 *            = rtime_i+1 - (rtime_i - utime_i)
 602	 *            = (rtime_i+1 - rtime_i) + utime_i
 603	 *            >= utime_i
 604	 */
 605	if (stime < prev->stime)
 606		stime = prev->stime;
 607	utime = rtime - stime;
 608
 609	/*
 610	 * Make sure utime doesn't go backwards; this still preserves
 611	 * monotonicity for stime, analogous argument to above.
 
 612	 */
 613	if (utime < prev->utime) {
 614		utime = prev->utime;
 615		stime = rtime - utime;
 616	}
 617
 618	prev->stime = stime;
 619	prev->utime = utime;
 620out:
 621	*ut = prev->utime;
 622	*st = prev->stime;
 623	raw_spin_unlock_irqrestore(&prev->lock, flags);
 624}
 625
 626void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
 627{
 628	struct task_cputime cputime = {
 629		.sum_exec_runtime = p->se.sum_exec_runtime,
 630	};
 631
 632	if (task_cputime(p, &cputime.utime, &cputime.stime))
 633		cputime.sum_exec_runtime = task_sched_runtime(p);
 634	cputime_adjust(&cputime, &p->prev_cputime, ut, st);
 635}
 636EXPORT_SYMBOL_GPL(task_cputime_adjusted);
 637
 638void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
 
 
 
 639{
 640	struct task_cputime cputime;
 641
 642	thread_group_cputime(p, &cputime);
 643	cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
 644}
 645#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
 646
 647#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
 648static u64 vtime_delta(struct vtime *vtime)
 649{
 650	unsigned long long clock;
 651
 652	clock = sched_clock();
 653	if (clock < vtime->starttime)
 654		return 0;
 655
 656	return clock - vtime->starttime;
 657}
 658
 659static u64 get_vtime_delta(struct vtime *vtime)
 660{
 661	u64 delta = vtime_delta(vtime);
 662	u64 other;
 663
 664	/*
 665	 * Unlike tick based timing, vtime based timing never has lost
 666	 * ticks, and no need for steal time accounting to make up for
 667	 * lost ticks. Vtime accounts a rounded version of actual
 668	 * elapsed time. Limit account_other_time to prevent rounding
 669	 * errors from causing elapsed vtime to go negative.
 670	 */
 671	other = account_other_time(delta);
 672	WARN_ON_ONCE(vtime->state == VTIME_INACTIVE);
 673	vtime->starttime += delta;
 674
 675	return delta - other;
 
 676}
 677
 678static void vtime_account_system(struct task_struct *tsk,
 679				 struct vtime *vtime)
 680{
 681	vtime->stime += get_vtime_delta(vtime);
 682	if (vtime->stime >= TICK_NSEC) {
 683		account_system_time(tsk, irq_count(), vtime->stime);
 684		vtime->stime = 0;
 685	}
 686}
 687
 688static void vtime_account_guest(struct task_struct *tsk,
 689				struct vtime *vtime)
 690{
 691	vtime->gtime += get_vtime_delta(vtime);
 692	if (vtime->gtime >= TICK_NSEC) {
 693		account_guest_time(tsk, vtime->gtime);
 694		vtime->gtime = 0;
 695	}
 696}
 697
 698static void __vtime_account_kernel(struct task_struct *tsk,
 699				   struct vtime *vtime)
 700{
 701	/* We might have scheduled out from guest path */
 702	if (vtime->state == VTIME_GUEST)
 703		vtime_account_guest(tsk, vtime);
 704	else
 705		vtime_account_system(tsk, vtime);
 706}
 707
 708void vtime_account_kernel(struct task_struct *tsk)
 709{
 710	struct vtime *vtime = &tsk->vtime;
 711
 712	if (!vtime_delta(vtime))
 713		return;
 714
 715	write_seqcount_begin(&vtime->seqcount);
 716	__vtime_account_kernel(tsk, vtime);
 717	write_seqcount_end(&vtime->seqcount);
 718}
 719
 720void vtime_user_enter(struct task_struct *tsk)
 721{
 722	struct vtime *vtime = &tsk->vtime;
 723
 724	write_seqcount_begin(&vtime->seqcount);
 725	vtime_account_system(tsk, vtime);
 726	vtime->state = VTIME_USER;
 727	write_seqcount_end(&vtime->seqcount);
 728}
 729
 730void vtime_user_exit(struct task_struct *tsk)
 731{
 732	struct vtime *vtime = &tsk->vtime;
 733
 734	write_seqcount_begin(&vtime->seqcount);
 735	vtime->utime += get_vtime_delta(vtime);
 736	if (vtime->utime >= TICK_NSEC) {
 737		account_user_time(tsk, vtime->utime);
 738		vtime->utime = 0;
 739	}
 740	vtime->state = VTIME_SYS;
 741	write_seqcount_end(&vtime->seqcount);
 742}
 743
 744void vtime_guest_enter(struct task_struct *tsk)
 745{
 746	struct vtime *vtime = &tsk->vtime;
 747	/*
 748	 * The flags must be updated under the lock with
 749	 * the vtime_starttime flush and update.
 750	 * That enforces a right ordering and update sequence
 751	 * synchronization against the reader (task_gtime())
 752	 * that can thus safely catch up with a tickless delta.
 753	 */
 754	write_seqcount_begin(&vtime->seqcount);
 755	vtime_account_system(tsk, vtime);
 756	tsk->flags |= PF_VCPU;
 757	vtime->state = VTIME_GUEST;
 758	write_seqcount_end(&vtime->seqcount);
 759}
 760EXPORT_SYMBOL_GPL(vtime_guest_enter);
 761
 762void vtime_guest_exit(struct task_struct *tsk)
 763{
 764	struct vtime *vtime = &tsk->vtime;
 765
 766	write_seqcount_begin(&vtime->seqcount);
 767	vtime_account_guest(tsk, vtime);
 768	tsk->flags &= ~PF_VCPU;
 769	vtime->state = VTIME_SYS;
 770	write_seqcount_end(&vtime->seqcount);
 771}
 772EXPORT_SYMBOL_GPL(vtime_guest_exit);
 773
 774void vtime_account_idle(struct task_struct *tsk)
 775{
 776	account_idle_time(get_vtime_delta(&tsk->vtime));
 
 
 777}
 778
 779void vtime_task_switch_generic(struct task_struct *prev)
 780{
 781	struct vtime *vtime = &prev->vtime;
 
 
 782
 783	write_seqcount_begin(&vtime->seqcount);
 784	if (vtime->state == VTIME_IDLE)
 785		vtime_account_idle(prev);
 786	else
 787		__vtime_account_kernel(prev, vtime);
 788	vtime->state = VTIME_INACTIVE;
 789	vtime->cpu = -1;
 790	write_seqcount_end(&vtime->seqcount);
 791
 792	vtime = &current->vtime;
 793
 794	write_seqcount_begin(&vtime->seqcount);
 795	if (is_idle_task(current))
 796		vtime->state = VTIME_IDLE;
 797	else if (current->flags & PF_VCPU)
 798		vtime->state = VTIME_GUEST;
 799	else
 800		vtime->state = VTIME_SYS;
 801	vtime->starttime = sched_clock();
 802	vtime->cpu = smp_processor_id();
 803	write_seqcount_end(&vtime->seqcount);
 804}
 805
 806void vtime_init_idle(struct task_struct *t, int cpu)
 807{
 808	struct vtime *vtime = &t->vtime;
 809	unsigned long flags;
 810
 811	local_irq_save(flags);
 812	write_seqcount_begin(&vtime->seqcount);
 813	vtime->state = VTIME_IDLE;
 814	vtime->starttime = sched_clock();
 815	vtime->cpu = cpu;
 816	write_seqcount_end(&vtime->seqcount);
 817	local_irq_restore(flags);
 818}
 819
 820u64 task_gtime(struct task_struct *t)
 821{
 822	struct vtime *vtime = &t->vtime;
 823	unsigned int seq;
 824	u64 gtime;
 825
 826	if (!vtime_accounting_enabled())
 827		return t->gtime;
 828
 829	do {
 830		seq = read_seqcount_begin(&vtime->seqcount);
 831
 832		gtime = t->gtime;
 833		if (vtime->state == VTIME_GUEST)
 834			gtime += vtime->gtime + vtime_delta(vtime);
 835
 836	} while (read_seqcount_retry(&vtime->seqcount, seq));
 837
 838	return gtime;
 839}
 840
 841/*
 842 * Fetch cputime raw values from fields of task_struct and
 843 * add up the pending nohz execution time since the last
 844 * cputime snapshot.
 845 */
 846bool task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
 
 
 
 
 847{
 848	struct vtime *vtime = &t->vtime;
 849	unsigned int seq;
 850	u64 delta;
 851	int ret;
 852
 853	if (!vtime_accounting_enabled()) {
 854		*utime = t->utime;
 855		*stime = t->stime;
 856		return false;
 857	}
 858
 859	do {
 860		ret = false;
 861		seq = read_seqcount_begin(&vtime->seqcount);
 862
 863		*utime = t->utime;
 864		*stime = t->stime;
 865
 866		/* Task is sleeping or idle, nothing to add */
 867		if (vtime->state < VTIME_SYS)
 
 
 
 
 
 
 868			continue;
 869
 870		ret = true;
 871		delta = vtime_delta(vtime);
 872
 873		/*
 874		 * Task runs either in user (including guest) or kernel space,
 875		 * add pending nohz time to the right place.
 876		 */
 877		if (vtime->state == VTIME_SYS)
 878			*stime += vtime->stime + delta;
 879		else
 880			*utime += vtime->utime + delta;
 881	} while (read_seqcount_retry(&vtime->seqcount, seq));
 882
 883	return ret;
 884}
 885
 886static int vtime_state_fetch(struct vtime *vtime, int cpu)
 887{
 888	int state = READ_ONCE(vtime->state);
 889
 890	/*
 891	 * We raced against a context switch, fetch the
 892	 * kcpustat task again.
 893	 */
 894	if (vtime->cpu != cpu && vtime->cpu != -1)
 895		return -EAGAIN;
 896
 897	/*
 898	 * Two possible things here:
 899	 * 1) We are seeing the scheduling out task (prev) or any past one.
 900	 * 2) We are seeing the scheduling in task (next) but it hasn't
 901	 *    passed though vtime_task_switch() yet so the pending
 902	 *    cputime of the prev task may not be flushed yet.
 903	 *
 904	 * Case 1) is ok but 2) is not. So wait for a safe VTIME state.
 905	 */
 906	if (state == VTIME_INACTIVE)
 907		return -EAGAIN;
 908
 909	return state;
 910}
 911
 912static u64 kcpustat_user_vtime(struct vtime *vtime)
 913{
 914	if (vtime->state == VTIME_USER)
 915		return vtime->utime + vtime_delta(vtime);
 916	else if (vtime->state == VTIME_GUEST)
 917		return vtime->gtime + vtime_delta(vtime);
 918	return 0;
 919}
 920
 921static int kcpustat_field_vtime(u64 *cpustat,
 922				struct task_struct *tsk,
 923				enum cpu_usage_stat usage,
 924				int cpu, u64 *val)
 925{
 926	struct vtime *vtime = &tsk->vtime;
 927	unsigned int seq;
 928
 929	do {
 930		int state;
 931
 932		seq = read_seqcount_begin(&vtime->seqcount);
 933
 934		state = vtime_state_fetch(vtime, cpu);
 935		if (state < 0)
 936			return state;
 937
 938		*val = cpustat[usage];
 939
 940		/*
 941		 * Nice VS unnice cputime accounting may be inaccurate if
 942		 * the nice value has changed since the last vtime update.
 943		 * But proper fix would involve interrupting target on nice
 944		 * updates which is a no go on nohz_full (although the scheduler
 945		 * may still interrupt the target if rescheduling is needed...)
 946		 */
 947		switch (usage) {
 948		case CPUTIME_SYSTEM:
 949			if (state == VTIME_SYS)
 950				*val += vtime->stime + vtime_delta(vtime);
 951			break;
 952		case CPUTIME_USER:
 953			if (task_nice(tsk) <= 0)
 954				*val += kcpustat_user_vtime(vtime);
 955			break;
 956		case CPUTIME_NICE:
 957			if (task_nice(tsk) > 0)
 958				*val += kcpustat_user_vtime(vtime);
 959			break;
 960		case CPUTIME_GUEST:
 961			if (state == VTIME_GUEST && task_nice(tsk) <= 0)
 962				*val += vtime->gtime + vtime_delta(vtime);
 963			break;
 964		case CPUTIME_GUEST_NICE:
 965			if (state == VTIME_GUEST && task_nice(tsk) > 0)
 966				*val += vtime->gtime + vtime_delta(vtime);
 967			break;
 968		default:
 969			break;
 970		}
 971	} while (read_seqcount_retry(&vtime->seqcount, seq));
 972
 973	return 0;
 974}
 975
 976u64 kcpustat_field(struct kernel_cpustat *kcpustat,
 977		   enum cpu_usage_stat usage, int cpu)
 978{
 979	u64 *cpustat = kcpustat->cpustat;
 980	u64 val = cpustat[usage];
 981	struct rq *rq;
 982	int err;
 983
 984	if (!vtime_accounting_enabled_cpu(cpu))
 985		return val;
 986
 987	rq = cpu_rq(cpu);
 988
 989	for (;;) {
 990		struct task_struct *curr;
 991
 992		rcu_read_lock();
 993		curr = rcu_dereference(rq->curr);
 994		if (WARN_ON_ONCE(!curr)) {
 995			rcu_read_unlock();
 996			return cpustat[usage];
 997		}
 998
 999		err = kcpustat_field_vtime(cpustat, curr, usage, cpu, &val);
1000		rcu_read_unlock();
1001
1002		if (!err)
1003			return val;
1004
1005		cpu_relax();
1006	}
1007}
1008EXPORT_SYMBOL_GPL(kcpustat_field);
1009
1010static int kcpustat_cpu_fetch_vtime(struct kernel_cpustat *dst,
1011				    const struct kernel_cpustat *src,
1012				    struct task_struct *tsk, int cpu)
1013{
1014	struct vtime *vtime = &tsk->vtime;
1015	unsigned int seq;
1016
1017	do {
1018		u64 *cpustat;
1019		u64 delta;
1020		int state;
1021
1022		seq = read_seqcount_begin(&vtime->seqcount);
1023
1024		state = vtime_state_fetch(vtime, cpu);
1025		if (state < 0)
1026			return state;
1027
1028		*dst = *src;
1029		cpustat = dst->cpustat;
1030
1031		/* Task is sleeping, dead or idle, nothing to add */
1032		if (state < VTIME_SYS)
1033			continue;
1034
1035		delta = vtime_delta(vtime);
1036
1037		/*
1038		 * Task runs either in user (including guest) or kernel space,
1039		 * add pending nohz time to the right place.
1040		 */
1041		if (state == VTIME_SYS) {
1042			cpustat[CPUTIME_SYSTEM] += vtime->stime + delta;
1043		} else if (state == VTIME_USER) {
1044			if (task_nice(tsk) > 0)
1045				cpustat[CPUTIME_NICE] += vtime->utime + delta;
1046			else
1047				cpustat[CPUTIME_USER] += vtime->utime + delta;
1048		} else {
1049			WARN_ON_ONCE(state != VTIME_GUEST);
1050			if (task_nice(tsk) > 0) {
1051				cpustat[CPUTIME_GUEST_NICE] += vtime->gtime + delta;
1052				cpustat[CPUTIME_NICE] += vtime->gtime + delta;
1053			} else {
1054				cpustat[CPUTIME_GUEST] += vtime->gtime + delta;
1055				cpustat[CPUTIME_USER] += vtime->gtime + delta;
1056			}
1057		}
1058	} while (read_seqcount_retry(&vtime->seqcount, seq));
1059
1060	return 0;
 
 
 
 
 
1061}
1062
1063void kcpustat_cpu_fetch(struct kernel_cpustat *dst, int cpu)
 
1064{
1065	const struct kernel_cpustat *src = &kcpustat_cpu(cpu);
1066	struct rq *rq;
1067	int err;
1068
1069	if (!vtime_accounting_enabled_cpu(cpu)) {
1070		*dst = *src;
1071		return;
1072	}
1073
1074	rq = cpu_rq(cpu);
1075
1076	for (;;) {
1077		struct task_struct *curr;
1078
1079		rcu_read_lock();
1080		curr = rcu_dereference(rq->curr);
1081		if (WARN_ON_ONCE(!curr)) {
1082			rcu_read_unlock();
1083			*dst = *src;
1084			return;
1085		}
1086
1087		err = kcpustat_cpu_fetch_vtime(dst, src, curr, cpu);
1088		rcu_read_unlock();
1089
1090		if (!err)
1091			return;
1092
1093		cpu_relax();
1094	}
1095}
1096EXPORT_SYMBOL_GPL(kcpustat_cpu_fetch);
1097
1098#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */