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 */
v5.4
  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * Simple CPU accounting cgroup controller
  4 */
 
 
  5#include "sched.h"
 
 
 
 
  6
  7#ifdef CONFIG_IRQ_TIME_ACCOUNTING
  8
  9/*
 10 * There are no locks covering percpu hardirq/softirq time.
 11 * They are only modified in vtime_account, on corresponding CPU
 12 * with interrupts disabled. So, writes are safe.
 13 * They are read and saved off onto struct rq in update_rq_clock().
 14 * This may result in other CPU reading this CPU's irq time and can
 15 * race with irq/vtime_account on this CPU. We would either get old
 16 * or new value with a side effect of accounting a slice of irq time to wrong
 17 * task when irq is in progress while we read rq->clock. That is a worthy
 18 * compromise in place of having locks on each irq in account_system_time.
 19 */
 20DEFINE_PER_CPU(struct irqtime, cpu_irqtime);
 21
 22static int sched_clock_irqtime;
 23
 24void enable_sched_clock_irqtime(void)
 25{
 26	sched_clock_irqtime = 1;
 27}
 28
 29void disable_sched_clock_irqtime(void)
 30{
 31	sched_clock_irqtime = 0;
 32}
 33
 34static void irqtime_account_delta(struct irqtime *irqtime, u64 delta,
 35				  enum cpu_usage_stat idx)
 36{
 37	u64 *cpustat = kcpustat_this_cpu->cpustat;
 38
 39	u64_stats_update_begin(&irqtime->sync);
 40	cpustat[idx] += delta;
 41	irqtime->total += delta;
 42	irqtime->tick_delta += delta;
 43	u64_stats_update_end(&irqtime->sync);
 44}
 45
 46/*
 47 * Called before incrementing preempt_count on {soft,}irq_enter
 48 * and before decrementing preempt_count on {soft,}irq_exit.
 49 */
 50void irqtime_account_irq(struct task_struct *curr)
 51{
 52	struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
 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
 
 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		irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
 71	else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
 72		irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
 
 
 73}
 74EXPORT_SYMBOL_GPL(irqtime_account_irq);
 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		cpustat[CPUTIME_NICE] += cputime;
151		cpustat[CPUTIME_GUEST_NICE] += cputime;
152	} else {
153		cpustat[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 * When a guest is interrupted for a longer amount of time, missed clock
231 * ticks are not redelivered later. Due to that, this function may on
232 * occasion account more time than the calling functions think elapsed.
233 */
234static __always_inline u64 steal_account_process_time(u64 maxtime)
235{
236#ifdef CONFIG_PARAVIRT
237	if (static_key_false(&paravirt_steal_enabled)) {
 
238		u64 steal;
239
240		steal = paravirt_steal_clock(smp_processor_id());
241		steal -= this_rq()->prev_steal_time;
242		steal = min(steal, maxtime);
243		account_steal_time(steal);
244		this_rq()->prev_steal_time += steal;
245
246		return steal;
 
 
 
 
247	}
248#endif
249	return 0;
250}
251
252/*
253 * Account how much elapsed time was spent in steal, irq, or softirq time.
254 */
255static inline u64 account_other_time(u64 max)
256{
257	u64 accounted;
258
259	lockdep_assert_irqs_disabled();
 
260
261	accounted = steal_account_process_time(max);
262
263	if (accounted < max)
264		accounted += irqtime_tick_accounted(max - accounted);
 
 
 
265
266	return accounted;
267}
268
269#ifdef CONFIG_64BIT
270static inline u64 read_sum_exec_runtime(struct task_struct *t)
271{
272	return t->se.sum_exec_runtime;
273}
274#else
275static u64 read_sum_exec_runtime(struct task_struct *t)
276{
277	u64 ns;
278	struct rq_flags rf;
279	struct rq *rq;
280
281	rq = task_rq_lock(t, &rf);
282	ns = t->se.sum_exec_runtime;
283	task_rq_unlock(rq, t, &rf);
284
285	return ns;
286}
287#endif
288
289/*
290 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
291 * tasks (sum on group iteration) belonging to @tsk's group.
292 */
293void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
294{
295	struct signal_struct *sig = tsk->signal;
296	u64 utime, stime;
297	struct task_struct *t;
298	unsigned int seq, nextseq;
299	unsigned long flags;
300
301	/*
302	 * Update current task runtime to account pending time since last
303	 * scheduler action or thread_group_cputime() call. This thread group
304	 * might have other running tasks on different CPUs, but updating
305	 * their runtime can affect syscall performance, so we skip account
306	 * those pending times and rely only on values updated on tick or
307	 * other scheduler action.
308	 */
309	if (same_thread_group(current, tsk))
310		(void) task_sched_runtime(current);
311
312	rcu_read_lock();
313	/* Attempt a lockless read on the first round. */
314	nextseq = 0;
315	do {
316		seq = nextseq;
317		flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
318		times->utime = sig->utime;
319		times->stime = sig->stime;
320		times->sum_exec_runtime = sig->sum_sched_runtime;
321
322		for_each_thread(tsk, t) {
323			task_cputime(t, &utime, &stime);
324			times->utime += utime;
325			times->stime += stime;
326			times->sum_exec_runtime += read_sum_exec_runtime(t);
327		}
328		/* If lockless access failed, take the lock. */
329		nextseq = 1;
330	} while (need_seqretry(&sig->stats_lock, seq));
331	done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
332	rcu_read_unlock();
333}
334
335#ifdef CONFIG_IRQ_TIME_ACCOUNTING
336/*
337 * Account a tick to a process and cpustat
338 * @p: the process that the CPU time gets accounted to
339 * @user_tick: is the tick from userspace
340 * @rq: the pointer to rq
341 *
342 * Tick demultiplexing follows the order
343 * - pending hardirq update
344 * - pending softirq update
345 * - user_time
346 * - idle_time
347 * - system time
348 *   - check for guest_time
349 *   - else account as system_time
350 *
351 * Check for hardirq is done both for system and user time as there is
352 * no timer going off while we are on hardirq and hence we may never get an
353 * opportunity to update it solely in system time.
354 * p->stime and friends are only updated on system time and not on irq
355 * softirq as those do not count in task exec_runtime any more.
356 */
357static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
358					 struct rq *rq, int ticks)
359{
360	u64 other, cputime = TICK_NSEC * ticks;
 
361
362	/*
363	 * When returning from idle, many ticks can get accounted at
364	 * once, including some ticks of steal, irq, and softirq time.
365	 * Subtract those ticks from the amount of time accounted to
366	 * idle, or potentially user or system time. Due to rounding,
367	 * other time can exceed ticks occasionally.
368	 */
369	other = account_other_time(ULONG_MAX);
370	if (other >= cputime)
371		return;
372
373	cputime -= other;
374
375	if (this_cpu_ksoftirqd() == p) {
376		/*
377		 * ksoftirqd time do not get accounted in cpu_softirq_time.
378		 * So, we have to handle it separately here.
379		 * Also, p->stime needs to be updated for ksoftirqd.
380		 */
381		account_system_index_time(p, cputime, CPUTIME_SOFTIRQ);
382	} else if (user_tick) {
383		account_user_time(p, cputime);
384	} else if (p == rq->idle) {
385		account_idle_time(cputime);
386	} else if (p->flags & PF_VCPU) { /* System time or guest time */
387		account_guest_time(p, cputime);
388	} else {
389		account_system_index_time(p, cputime, CPUTIME_SYSTEM);
390	}
391}
392
393static void irqtime_account_idle_ticks(int ticks)
394{
395	struct rq *rq = this_rq();
396
397	irqtime_account_process_tick(current, 0, rq, ticks);
398}
399#else /* CONFIG_IRQ_TIME_ACCOUNTING */
400static inline void irqtime_account_idle_ticks(int ticks) { }
401static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
402						struct rq *rq, int nr_ticks) { }
403#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
404
405/*
406 * Use precise platform statistics if available:
407 */
408#ifdef CONFIG_VIRT_CPU_ACCOUNTING
409# ifndef __ARCH_HAS_VTIME_TASK_SWITCH
 
410void vtime_common_task_switch(struct task_struct *prev)
411{
412	if (is_idle_task(prev))
413		vtime_account_idle(prev);
414	else
415		vtime_account_system(prev);
416
417	vtime_flush(prev);
 
 
418	arch_vtime_task_switch(prev);
419}
420# endif
 
421#endif /* CONFIG_VIRT_CPU_ACCOUNTING */
422
423
424#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
425/*
426 * Archs that account the whole time spent in the idle task
427 * (outside irq) as idle time can rely on this and just implement
428 * vtime_account_system() and vtime_account_idle(). Archs that
429 * have other meaning of the idle time (s390 only includes the
430 * time spent by the CPU when it's in low power mode) must override
431 * vtime_account().
432 */
433#ifndef __ARCH_HAS_VTIME_ACCOUNT
434void vtime_account_irq_enter(struct task_struct *tsk)
435{
436	if (!in_interrupt() && is_idle_task(tsk))
437		vtime_account_idle(tsk);
438	else
439		vtime_account_system(tsk);
440}
441EXPORT_SYMBOL_GPL(vtime_account_irq_enter);
442#endif /* __ARCH_HAS_VTIME_ACCOUNT */
443
444void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
445		    u64 *ut, u64 *st)
446{
447	*ut = curr->utime;
448	*st = curr->stime;
449}
450
451void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
452{
453	*ut = p->utime;
454	*st = p->stime;
455}
456EXPORT_SYMBOL_GPL(task_cputime_adjusted);
457
458void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
459{
460	struct task_cputime cputime;
461
462	thread_group_cputime(p, &cputime);
463
464	*ut = cputime.utime;
465	*st = cputime.stime;
466}
467
468#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE: */
469
470/*
471 * Account a single tick of CPU time.
472 * @p: the process that the CPU time gets accounted to
473 * @user_tick: indicates if the tick is a user or a system tick
474 */
475void account_process_tick(struct task_struct *p, int user_tick)
476{
477	u64 cputime, steal;
478	struct rq *rq = this_rq();
479
480	if (vtime_accounting_cpu_enabled())
481		return;
482
483	if (sched_clock_irqtime) {
484		irqtime_account_process_tick(p, user_tick, rq, 1);
485		return;
486	}
487
488	cputime = TICK_NSEC;
489	steal = steal_account_process_time(ULONG_MAX);
490
491	if (steal >= cputime)
492		return;
493
494	cputime -= steal;
495
496	if (user_tick)
497		account_user_time(p, cputime);
498	else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
499		account_system_time(p, HARDIRQ_OFFSET, cputime);
500	else
501		account_idle_time(cputime);
502}
503
504/*
505 * Account multiple ticks of idle time.
506 * @ticks: number of stolen ticks
507 */
508void account_idle_ticks(unsigned long ticks)
509{
510	u64 cputime, steal;
511
512	if (sched_clock_irqtime) {
513		irqtime_account_idle_ticks(ticks);
514		return;
515	}
516
517	cputime = ticks * TICK_NSEC;
518	steal = steal_account_process_time(ULONG_MAX);
519
520	if (steal >= cputime)
521		return;
522
523	cputime -= steal;
524	account_idle_time(cputime);
525}
526
527/*
528 * Perform (stime * rtime) / total, but avoid multiplication overflow by
529 * losing precision when the numbers are big.
530 */
531static u64 scale_stime(u64 stime, u64 rtime, u64 total)
532{
533	u64 scaled;
534
535	for (;;) {
536		/* Make sure "rtime" is the bigger of stime/rtime */
537		if (stime > rtime)
538			swap(rtime, stime);
539
540		/* Make sure 'total' fits in 32 bits */
541		if (total >> 32)
542			goto drop_precision;
543
544		/* Does rtime (and thus stime) fit in 32 bits? */
545		if (!(rtime >> 32))
546			break;
547
548		/* Can we just balance rtime/stime rather than dropping bits? */
549		if (stime >> 31)
550			goto drop_precision;
551
552		/* We can grow stime and shrink rtime and try to make them both fit */
553		stime <<= 1;
554		rtime >>= 1;
555		continue;
556
557drop_precision:
558		/* We drop from rtime, it has more bits than stime */
559		rtime >>= 1;
560		total >>= 1;
561	}
562
563	/*
564	 * Make sure gcc understands that this is a 32x32->64 multiply,
565	 * followed by a 64/32->64 divide.
566	 */
567	scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
568	return scaled;
569}
570
571/*
572 * Adjust tick based cputime random precision against scheduler runtime
573 * accounting.
574 *
575 * Tick based cputime accounting depend on random scheduling timeslices of a
576 * task to be interrupted or not by the timer.  Depending on these
577 * circumstances, the number of these interrupts may be over or
578 * under-optimistic, matching the real user and system cputime with a variable
579 * precision.
580 *
581 * Fix this by scaling these tick based values against the total runtime
582 * accounted by the CFS scheduler.
583 *
584 * This code provides the following guarantees:
585 *
586 *   stime + utime == rtime
587 *   stime_i+1 >= stime_i, utime_i+1 >= utime_i
588 *
589 * Assuming that rtime_i+1 >= rtime_i.
590 */
591void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
592		    u64 *ut, u64 *st)
 
593{
594	u64 rtime, stime, utime;
595	unsigned long flags;
596
597	/* Serialize concurrent callers such that we can honour our guarantees */
598	raw_spin_lock_irqsave(&prev->lock, flags);
599	rtime = curr->sum_exec_runtime;
600
601	/*
602	 * This is possible under two circumstances:
603	 *  - rtime isn't monotonic after all (a bug);
604	 *  - we got reordered by the lock.
605	 *
606	 * In both cases this acts as a filter such that the rest of the code
607	 * can assume it is monotonic regardless of anything else.
608	 */
609	if (prev->stime + prev->utime >= rtime)
610		goto out;
611
612	stime = curr->stime;
613	utime = curr->utime;
614
615	/*
616	 * If either stime or utime are 0, assume all runtime is userspace.
617	 * Once a task gets some ticks, the monotonicy code at 'update:'
618	 * will ensure things converge to the observed ratio.
619	 */
620	if (stime == 0) {
621		utime = rtime;
622		goto update;
623	}
624
625	if (utime == 0) {
626		stime = rtime;
627		goto update;
628	}
629
630	stime = scale_stime(stime, rtime, stime + utime);
 
631
632update:
633	/*
634	 * Make sure stime doesn't go backwards; this preserves monotonicity
635	 * for utime because rtime is monotonic.
636	 *
637	 *  utime_i+1 = rtime_i+1 - stime_i
638	 *            = rtime_i+1 - (rtime_i - utime_i)
639	 *            = (rtime_i+1 - rtime_i) + utime_i
640	 *            >= utime_i
641	 */
642	if (stime < prev->stime)
643		stime = prev->stime;
644	utime = rtime - stime;
645
646	/*
647	 * Make sure utime doesn't go backwards; this still preserves
648	 * monotonicity for stime, analogous argument to above.
649	 */
650	if (utime < prev->utime) {
651		utime = prev->utime;
652		stime = rtime - utime;
653	}
654
655	prev->stime = stime;
656	prev->utime = utime;
657out:
658	*ut = prev->utime;
659	*st = prev->stime;
660	raw_spin_unlock_irqrestore(&prev->lock, flags);
661}
662
663void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
664{
665	struct task_cputime cputime = {
666		.sum_exec_runtime = p->se.sum_exec_runtime,
667	};
668
669	task_cputime(p, &cputime.utime, &cputime.stime);
670	cputime_adjust(&cputime, &p->prev_cputime, ut, st);
671}
672EXPORT_SYMBOL_GPL(task_cputime_adjusted);
673
674void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
675{
676	struct task_cputime cputime;
677
678	thread_group_cputime(p, &cputime);
679	cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
680}
681#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
682
683#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
684static u64 vtime_delta(struct vtime *vtime)
685{
686	unsigned long long clock;
687
688	clock = sched_clock();
689	if (clock < vtime->starttime)
690		return 0;
691
692	return clock - vtime->starttime;
693}
694
695static u64 get_vtime_delta(struct vtime *vtime)
696{
697	u64 delta = vtime_delta(vtime);
698	u64 other;
699
700	/*
701	 * Unlike tick based timing, vtime based timing never has lost
702	 * ticks, and no need for steal time accounting to make up for
703	 * lost ticks. Vtime accounts a rounded version of actual
704	 * elapsed time. Limit account_other_time to prevent rounding
705	 * errors from causing elapsed vtime to go negative.
706	 */
 
707	other = account_other_time(delta);
708	WARN_ON_ONCE(vtime->state == VTIME_INACTIVE);
709	vtime->starttime += delta;
710
711	return delta - other;
712}
713
714static void __vtime_account_system(struct task_struct *tsk,
715				   struct vtime *vtime)
716{
717	vtime->stime += get_vtime_delta(vtime);
718	if (vtime->stime >= TICK_NSEC) {
719		account_system_time(tsk, irq_count(), vtime->stime);
720		vtime->stime = 0;
721	}
722}
723
724static void vtime_account_guest(struct task_struct *tsk,
725				struct vtime *vtime)
726{
727	vtime->gtime += get_vtime_delta(vtime);
728	if (vtime->gtime >= TICK_NSEC) {
729		account_guest_time(tsk, vtime->gtime);
730		vtime->gtime = 0;
731	}
732}
733
734void vtime_account_system(struct task_struct *tsk)
735{
736	struct vtime *vtime = &tsk->vtime;
737
738	if (!vtime_delta(vtime))
739		return;
740
741	write_seqcount_begin(&vtime->seqcount);
742	/* We might have scheduled out from guest path */
743	if (tsk->flags & PF_VCPU)
744		vtime_account_guest(tsk, vtime);
745	else
746		__vtime_account_system(tsk, vtime);
747	write_seqcount_end(&vtime->seqcount);
748}
749
750void vtime_user_enter(struct task_struct *tsk)
751{
752	struct vtime *vtime = &tsk->vtime;
753
754	write_seqcount_begin(&vtime->seqcount);
755	__vtime_account_system(tsk, vtime);
756	vtime->state = VTIME_USER;
757	write_seqcount_end(&vtime->seqcount);
 
 
 
758}
759
760void vtime_user_exit(struct task_struct *tsk)
761{
762	struct vtime *vtime = &tsk->vtime;
763
764	write_seqcount_begin(&vtime->seqcount);
765	vtime->utime += get_vtime_delta(vtime);
766	if (vtime->utime >= TICK_NSEC) {
767		account_user_time(tsk, vtime->utime);
768		vtime->utime = 0;
769	}
770	vtime->state = VTIME_SYS;
771	write_seqcount_end(&vtime->seqcount);
772}
773
774void vtime_guest_enter(struct task_struct *tsk)
775{
776	struct vtime *vtime = &tsk->vtime;
777	/*
778	 * The flags must be updated under the lock with
779	 * the vtime_starttime flush and update.
780	 * That enforces a right ordering and update sequence
781	 * synchronization against the reader (task_gtime())
782	 * that can thus safely catch up with a tickless delta.
783	 */
784	write_seqcount_begin(&vtime->seqcount);
785	__vtime_account_system(tsk, vtime);
786	tsk->flags |= PF_VCPU;
787	write_seqcount_end(&vtime->seqcount);
 
788}
789EXPORT_SYMBOL_GPL(vtime_guest_enter);
790
791void vtime_guest_exit(struct task_struct *tsk)
792{
793	struct vtime *vtime = &tsk->vtime;
794
795	write_seqcount_begin(&vtime->seqcount);
796	vtime_account_guest(tsk, vtime);
797	tsk->flags &= ~PF_VCPU;
798	write_seqcount_end(&vtime->seqcount);
799}
800EXPORT_SYMBOL_GPL(vtime_guest_exit);
801
802void vtime_account_idle(struct task_struct *tsk)
803{
804	account_idle_time(get_vtime_delta(&tsk->vtime));
 
 
805}
806
807void arch_vtime_task_switch(struct task_struct *prev)
808{
809	struct vtime *vtime = &prev->vtime;
810
811	write_seqcount_begin(&vtime->seqcount);
812	vtime->state = VTIME_INACTIVE;
813	write_seqcount_end(&vtime->seqcount);
814
815	vtime = &current->vtime;
816
817	write_seqcount_begin(&vtime->seqcount);
818	vtime->state = VTIME_SYS;
819	vtime->starttime = sched_clock();
820	write_seqcount_end(&vtime->seqcount);
821}
822
823void vtime_init_idle(struct task_struct *t, int cpu)
824{
825	struct vtime *vtime = &t->vtime;
826	unsigned long flags;
827
828	local_irq_save(flags);
829	write_seqcount_begin(&vtime->seqcount);
830	vtime->state = VTIME_SYS;
831	vtime->starttime = sched_clock();
832	write_seqcount_end(&vtime->seqcount);
833	local_irq_restore(flags);
834}
835
836u64 task_gtime(struct task_struct *t)
837{
838	struct vtime *vtime = &t->vtime;
839	unsigned int seq;
840	u64 gtime;
841
842	if (!vtime_accounting_enabled())
843		return t->gtime;
844
845	do {
846		seq = read_seqcount_begin(&vtime->seqcount);
847
848		gtime = t->gtime;
849		if (vtime->state == VTIME_SYS && t->flags & PF_VCPU)
850			gtime += vtime->gtime + vtime_delta(vtime);
851
852	} while (read_seqcount_retry(&vtime->seqcount, seq));
853
854	return gtime;
855}
856
857/*
858 * Fetch cputime raw values from fields of task_struct and
859 * add up the pending nohz execution time since the last
860 * cputime snapshot.
861 */
862void task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
863{
864	struct vtime *vtime = &t->vtime;
865	unsigned int seq;
866	u64 delta;
867
868	if (!vtime_accounting_enabled()) {
869		*utime = t->utime;
870		*stime = t->stime;
871		return;
872	}
873
874	do {
875		seq = read_seqcount_begin(&vtime->seqcount);
876
877		*utime = t->utime;
878		*stime = t->stime;
879
880		/* Task is sleeping, nothing to add */
881		if (vtime->state == VTIME_INACTIVE || is_idle_task(t))
882			continue;
883
884		delta = vtime_delta(vtime);
885
886		/*
887		 * Task runs either in user or kernel space, add pending nohz time to
888		 * the right place.
889		 */
890		if (vtime->state == VTIME_USER || t->flags & PF_VCPU)
891			*utime += vtime->utime + delta;
892		else if (vtime->state == VTIME_SYS)
893			*stime += vtime->stime + delta;
894	} while (read_seqcount_retry(&vtime->seqcount, seq));
895}
896#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */