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