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