sched/cputime: Do not scale when utime == 0
[linux-3.10.git] / kernel / sched / cputime.c
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  */
23 DEFINE_PER_CPU(u64, cpu_hardirq_time);
24 DEFINE_PER_CPU(u64, cpu_softirq_time);
25
26 static DEFINE_PER_CPU(u64, irq_start_time);
27 static int sched_clock_irqtime;
28
29 void enable_sched_clock_irqtime(void)
30 {
31         sched_clock_irqtime = 1;
32 }
33
34 void disable_sched_clock_irqtime(void)
35 {
36         sched_clock_irqtime = 0;
37 }
38
39 #ifndef CONFIG_64BIT
40 DEFINE_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  */
47 void 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 }
77 EXPORT_SYMBOL_GPL(irqtime_account_irq);
78
79 static 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
94 static 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
115 static 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         __get_cpu_var(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  */
135 void 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  */
160 static 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  */
188 static inline
189 void __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  */
211 void 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  */
235 void 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  */
246 void 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
257 static __always_inline bool steal_account_process_tick(void)
258 {
259 #ifdef CONFIG_PARAVIRT
260         if (static_key_false(&paravirt_steal_enabled)) {
261                 u64 steal, st = 0;
262
263                 steal = paravirt_steal_clock(smp_processor_id());
264                 steal -= this_rq()->prev_steal_time;
265
266                 st = steal_ticks(steal);
267                 this_rq()->prev_steal_time += st * TICK_NSEC;
268
269                 account_steal_time(st);
270                 return st;
271         }
272 #endif
273         return false;
274 }
275
276 /*
277  * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
278  * tasks (sum on group iteration) belonging to @tsk's group.
279  */
280 void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
281 {
282         struct signal_struct *sig = tsk->signal;
283         cputime_t utime, stime;
284         struct task_struct *t;
285
286         times->utime = sig->utime;
287         times->stime = sig->stime;
288         times->sum_exec_runtime = sig->sum_sched_runtime;
289
290         rcu_read_lock();
291         /* make sure we can trust tsk->thread_group list */
292         if (!likely(pid_alive(tsk)))
293                 goto out;
294
295         t = tsk;
296         do {
297                 task_cputime(t, &utime, &stime);
298                 times->utime += utime;
299                 times->stime += stime;
300                 times->sum_exec_runtime += task_sched_runtime(t);
301         } while_each_thread(tsk, t);
302 out:
303         rcu_read_unlock();
304 }
305
306 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
307 /*
308  * Account a tick to a process and cpustat
309  * @p: the process that the cpu time gets accounted to
310  * @user_tick: is the tick from userspace
311  * @rq: the pointer to rq
312  *
313  * Tick demultiplexing follows the order
314  * - pending hardirq update
315  * - pending softirq update
316  * - user_time
317  * - idle_time
318  * - system time
319  *   - check for guest_time
320  *   - else account as system_time
321  *
322  * Check for hardirq is done both for system and user time as there is
323  * no timer going off while we are on hardirq and hence we may never get an
324  * opportunity to update it solely in system time.
325  * p->stime and friends are only updated on system time and not on irq
326  * softirq as those do not count in task exec_runtime any more.
327  */
328 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
329                                                 struct rq *rq)
330 {
331         cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
332         u64 *cpustat = kcpustat_this_cpu->cpustat;
333
334         if (steal_account_process_tick())
335                 return;
336
337         if (irqtime_account_hi_update()) {
338                 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
339         } else if (irqtime_account_si_update()) {
340                 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
341         } else if (this_cpu_ksoftirqd() == p) {
342                 /*
343                  * ksoftirqd time do not get accounted in cpu_softirq_time.
344                  * So, we have to handle it separately here.
345                  * Also, p->stime needs to be updated for ksoftirqd.
346                  */
347                 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
348                                         CPUTIME_SOFTIRQ);
349         } else if (user_tick) {
350                 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
351         } else if (p == rq->idle) {
352                 account_idle_time(cputime_one_jiffy);
353         } else if (p->flags & PF_VCPU) { /* System time or guest time */
354                 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
355         } else {
356                 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
357                                         CPUTIME_SYSTEM);
358         }
359 }
360
361 static void irqtime_account_idle_ticks(int ticks)
362 {
363         int i;
364         struct rq *rq = this_rq();
365
366         for (i = 0; i < ticks; i++)
367                 irqtime_account_process_tick(current, 0, rq);
368 }
369 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
370 static inline void irqtime_account_idle_ticks(int ticks) {}
371 static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
372                                                 struct rq *rq) {}
373 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
374
375 /*
376  * Use precise platform statistics if available:
377  */
378 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
379
380 #ifndef __ARCH_HAS_VTIME_TASK_SWITCH
381 void vtime_task_switch(struct task_struct *prev)
382 {
383         if (!vtime_accounting_enabled())
384                 return;
385
386         if (is_idle_task(prev))
387                 vtime_account_idle(prev);
388         else
389                 vtime_account_system(prev);
390
391 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
392         vtime_account_user(prev);
393 #endif
394         arch_vtime_task_switch(prev);
395 }
396 #endif
397
398 /*
399  * Archs that account the whole time spent in the idle task
400  * (outside irq) as idle time can rely on this and just implement
401  * vtime_account_system() and vtime_account_idle(). Archs that
402  * have other meaning of the idle time (s390 only includes the
403  * time spent by the CPU when it's in low power mode) must override
404  * vtime_account().
405  */
406 #ifndef __ARCH_HAS_VTIME_ACCOUNT
407 void vtime_account_irq_enter(struct task_struct *tsk)
408 {
409         if (!vtime_accounting_enabled())
410                 return;
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 }
431 EXPORT_SYMBOL_GPL(vtime_account_irq_enter);
432 #endif /* __ARCH_HAS_VTIME_ACCOUNT */
433 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
434
435
436 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
437 void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
438 {
439         *ut = p->utime;
440         *st = p->stime;
441 }
442
443 void 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  */
458 void 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);
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  */
488 void 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  */
497 void 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  */
512 static 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                         u64 tmp = rtime; rtime = stime; stime = tmp;
520                 }
521
522                 /* Make sure 'total' fits in 32 bits */
523                 if (total >> 32)
524                         goto drop_precision;
525
526                 /* Does rtime (and thus stime) fit in 32 bits? */
527                 if (!(rtime >> 32))
528                         break;
529
530                 /* Can we just balance rtime/stime rather than dropping bits? */
531                 if (stime >> 31)
532                         goto drop_precision;
533
534                 /* We can grow stime and shrink rtime and try to make them both fit */
535                 stime <<= 1;
536                 rtime >>= 1;
537                 continue;
538
539 drop_precision:
540                 /* We drop from rtime, it has more bits than stime */
541                 rtime >>= 1;
542                 total >>= 1;
543         }
544
545         /*
546          * Make sure gcc understands that this is a 32x32->64 multiply,
547          * followed by a 64/32->64 divide.
548          */
549         scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
550         return (__force cputime_t) scaled;
551 }
552
553 /*
554  * Adjust tick based cputime random precision against scheduler
555  * runtime accounting.
556  */
557 static void cputime_adjust(struct task_cputime *curr,
558                            struct cputime *prev,
559                            cputime_t *ut, cputime_t *st)
560 {
561         cputime_t rtime, stime, utime;
562
563         if (vtime_accounting_enabled()) {
564                 *ut = curr->utime;
565                 *st = curr->stime;
566                 return;
567         }
568
569         /*
570          * Tick based cputime accounting depend on random scheduling
571          * timeslices of a task to be interrupted or not by the timer.
572          * Depending on these circumstances, the number of these interrupts
573          * may be over or under-optimistic, matching the real user and system
574          * cputime with a variable precision.
575          *
576          * Fix this by scaling these tick based values against the total
577          * runtime accounted by the CFS scheduler.
578          */
579         rtime = nsecs_to_cputime(curr->sum_exec_runtime);
580
581         /*
582          * Update userspace visible utime/stime values only if actual execution
583          * time is bigger than already exported. Note that can happen, that we
584          * provided bigger values due to scaling inaccuracy on big numbers.
585          */
586         if (prev->stime + prev->utime >= rtime)
587                 goto out;
588
589         stime = curr->stime;
590         utime = curr->utime;
591
592         if (utime == 0) {
593                 stime = rtime;
594         } else if (stime == 0) {
595                 utime = rtime;
596         } else {
597                 cputime_t total = stime + utime;
598
599                 stime = scale_stime((__force u64)stime,
600                                     (__force u64)rtime, (__force u64)total);
601                 utime = rtime - stime;
602         }
603
604         /*
605          * If the tick based count grows faster than the scheduler one,
606          * the result of the scaling may go backward.
607          * Let's enforce monotonicity.
608          */
609         prev->stime = max(prev->stime, stime);
610         prev->utime = max(prev->utime, utime);
611
612 out:
613         *ut = prev->utime;
614         *st = prev->stime;
615 }
616
617 void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
618 {
619         struct task_cputime cputime = {
620                 .sum_exec_runtime = p->se.sum_exec_runtime,
621         };
622
623         task_cputime(p, &cputime.utime, &cputime.stime);
624         cputime_adjust(&cputime, &p->prev_cputime, ut, st);
625 }
626
627 /*
628  * Must be called with siglock held.
629  */
630 void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
631 {
632         struct task_cputime cputime;
633
634         thread_group_cputime(p, &cputime);
635         cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
636 }
637 #endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
638
639 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
640 static unsigned long long vtime_delta(struct task_struct *tsk)
641 {
642         unsigned long long clock;
643
644         clock = local_clock();
645         if (clock < tsk->vtime_snap)
646                 return 0;
647
648         return clock - tsk->vtime_snap;
649 }
650
651 static cputime_t get_vtime_delta(struct task_struct *tsk)
652 {
653         unsigned long long delta = vtime_delta(tsk);
654
655         WARN_ON_ONCE(tsk->vtime_snap_whence == VTIME_SLEEPING);
656         tsk->vtime_snap += delta;
657
658         /* CHECKME: always safe to convert nsecs to cputime? */
659         return nsecs_to_cputime(delta);
660 }
661
662 static void __vtime_account_system(struct task_struct *tsk)
663 {
664         cputime_t delta_cpu = get_vtime_delta(tsk);
665
666         account_system_time(tsk, irq_count(), delta_cpu, cputime_to_scaled(delta_cpu));
667 }
668
669 void vtime_account_system(struct task_struct *tsk)
670 {
671         if (!vtime_accounting_enabled())
672                 return;
673
674         write_seqlock(&tsk->vtime_seqlock);
675         __vtime_account_system(tsk);
676         write_sequnlock(&tsk->vtime_seqlock);
677 }
678
679 void vtime_account_irq_exit(struct task_struct *tsk)
680 {
681         if (!vtime_accounting_enabled())
682                 return;
683
684         write_seqlock(&tsk->vtime_seqlock);
685         if (context_tracking_in_user())
686                 tsk->vtime_snap_whence = VTIME_USER;
687         __vtime_account_system(tsk);
688         write_sequnlock(&tsk->vtime_seqlock);
689 }
690
691 void vtime_account_user(struct task_struct *tsk)
692 {
693         cputime_t delta_cpu;
694
695         if (!vtime_accounting_enabled())
696                 return;
697
698         delta_cpu = get_vtime_delta(tsk);
699
700         write_seqlock(&tsk->vtime_seqlock);
701         tsk->vtime_snap_whence = VTIME_SYS;
702         account_user_time(tsk, delta_cpu, cputime_to_scaled(delta_cpu));
703         write_sequnlock(&tsk->vtime_seqlock);
704 }
705
706 void vtime_user_enter(struct task_struct *tsk)
707 {
708         if (!vtime_accounting_enabled())
709                 return;
710
711         write_seqlock(&tsk->vtime_seqlock);
712         tsk->vtime_snap_whence = VTIME_USER;
713         __vtime_account_system(tsk);
714         write_sequnlock(&tsk->vtime_seqlock);
715 }
716
717 void vtime_guest_enter(struct task_struct *tsk)
718 {
719         write_seqlock(&tsk->vtime_seqlock);
720         __vtime_account_system(tsk);
721         current->flags |= PF_VCPU;
722         write_sequnlock(&tsk->vtime_seqlock);
723 }
724
725 void vtime_guest_exit(struct task_struct *tsk)
726 {
727         write_seqlock(&tsk->vtime_seqlock);
728         __vtime_account_system(tsk);
729         current->flags &= ~PF_VCPU;
730         write_sequnlock(&tsk->vtime_seqlock);
731 }
732
733 void vtime_account_idle(struct task_struct *tsk)
734 {
735         cputime_t delta_cpu = get_vtime_delta(tsk);
736
737         account_idle_time(delta_cpu);
738 }
739
740 bool vtime_accounting_enabled(void)
741 {
742         return context_tracking_active();
743 }
744
745 void arch_vtime_task_switch(struct task_struct *prev)
746 {
747         write_seqlock(&prev->vtime_seqlock);
748         prev->vtime_snap_whence = VTIME_SLEEPING;
749         write_sequnlock(&prev->vtime_seqlock);
750
751         write_seqlock(&current->vtime_seqlock);
752         current->vtime_snap_whence = VTIME_SYS;
753         current->vtime_snap = sched_clock_cpu(smp_processor_id());
754         write_sequnlock(&current->vtime_seqlock);
755 }
756
757 void vtime_init_idle(struct task_struct *t, int cpu)
758 {
759         unsigned long flags;
760
761         write_seqlock_irqsave(&t->vtime_seqlock, flags);
762         t->vtime_snap_whence = VTIME_SYS;
763         t->vtime_snap = sched_clock_cpu(cpu);
764         write_sequnlock_irqrestore(&t->vtime_seqlock, flags);
765 }
766
767 cputime_t task_gtime(struct task_struct *t)
768 {
769         unsigned int seq;
770         cputime_t gtime;
771
772         do {
773                 seq = read_seqbegin(&t->vtime_seqlock);
774
775                 gtime = t->gtime;
776                 if (t->flags & PF_VCPU)
777                         gtime += vtime_delta(t);
778
779         } while (read_seqretry(&t->vtime_seqlock, seq));
780
781         return gtime;
782 }
783
784 /*
785  * Fetch cputime raw values from fields of task_struct and
786  * add up the pending nohz execution time since the last
787  * cputime snapshot.
788  */
789 static void
790 fetch_task_cputime(struct task_struct *t,
791                    cputime_t *u_dst, cputime_t *s_dst,
792                    cputime_t *u_src, cputime_t *s_src,
793                    cputime_t *udelta, cputime_t *sdelta)
794 {
795         unsigned int seq;
796         unsigned long long delta;
797
798         do {
799                 *udelta = 0;
800                 *sdelta = 0;
801
802                 seq = read_seqbegin(&t->vtime_seqlock);
803
804                 if (u_dst)
805                         *u_dst = *u_src;
806                 if (s_dst)
807                         *s_dst = *s_src;
808
809                 /* Task is sleeping, nothing to add */
810                 if (t->vtime_snap_whence == VTIME_SLEEPING ||
811                     is_idle_task(t))
812                         continue;
813
814                 delta = vtime_delta(t);
815
816                 /*
817                  * Task runs either in user or kernel space, add pending nohz time to
818                  * the right place.
819                  */
820                 if (t->vtime_snap_whence == VTIME_USER || t->flags & PF_VCPU) {
821                         *udelta = delta;
822                 } else {
823                         if (t->vtime_snap_whence == VTIME_SYS)
824                                 *sdelta = delta;
825                 }
826         } while (read_seqretry(&t->vtime_seqlock, seq));
827 }
828
829
830 void task_cputime(struct task_struct *t, cputime_t *utime, cputime_t *stime)
831 {
832         cputime_t udelta, sdelta;
833
834         fetch_task_cputime(t, utime, stime, &t->utime,
835                            &t->stime, &udelta, &sdelta);
836         if (utime)
837                 *utime += udelta;
838         if (stime)
839                 *stime += sdelta;
840 }
841
842 void task_cputime_scaled(struct task_struct *t,
843                          cputime_t *utimescaled, cputime_t *stimescaled)
844 {
845         cputime_t udelta, sdelta;
846
847         fetch_task_cputime(t, utimescaled, stimescaled,
848                            &t->utimescaled, &t->stimescaled, &udelta, &sdelta);
849         if (utimescaled)
850                 *utimescaled += cputime_to_scaled(udelta);
851         if (stimescaled)
852                 *stimescaled += cputime_to_scaled(sdelta);
853 }
854 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */