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>
10 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
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.
23 DEFINE_PER_CPU(u64, cpu_hardirq_time);
24 DEFINE_PER_CPU(u64, cpu_softirq_time);
26 static DEFINE_PER_CPU(u64, irq_start_time);
27 static int sched_clock_irqtime;
29 void enable_sched_clock_irqtime(void)
31 sched_clock_irqtime = 1;
34 void disable_sched_clock_irqtime(void)
36 sched_clock_irqtime = 0;
40 DEFINE_PER_CPU(seqcount_t, irq_time_seq);
41 #endif /* CONFIG_64BIT */
44 * Called before incrementing preempt_count on {soft,}irq_enter
45 * and before decrementing preempt_count on {soft,}irq_exit.
47 void irqtime_account_irq(struct task_struct *curr)
53 if (!sched_clock_irqtime)
56 local_irq_save(flags);
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);
62 irq_time_write_begin();
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.
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);
75 local_irq_restore(flags);
77 EXPORT_SYMBOL_GPL(irqtime_account_irq);
79 static int irqtime_account_hi_update(void)
81 u64 *cpustat = kcpustat_this_cpu->cpustat;
86 local_irq_save(flags);
87 latest_ns = this_cpu_read(cpu_hardirq_time);
88 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
90 local_irq_restore(flags);
94 static int irqtime_account_si_update(void)
96 u64 *cpustat = kcpustat_this_cpu->cpustat;
101 local_irq_save(flags);
102 latest_ns = this_cpu_read(cpu_softirq_time);
103 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
105 local_irq_restore(flags);
109 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
111 #define sched_clock_irqtime (0)
113 #endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
115 static inline void task_group_account_field(struct task_struct *p, int index,
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.
124 __this_cpu_add(kernel_cpustat.cpustat[index], tmp);
126 cpuacct_account_field(p, index, tmp);
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
135 void account_user_time(struct task_struct *p, cputime_t cputime,
136 cputime_t cputime_scaled)
140 /* Add user time to process. */
142 p->utimescaled += cputime_scaled;
143 account_group_user_time(p, cputime);
145 index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
147 /* Add user time to cpustat. */
148 task_group_account_field(p, index, (__force u64) cputime);
150 /* Account for user time used */
151 acct_account_cputime(p);
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
160 static void account_guest_time(struct task_struct *p, cputime_t cputime,
161 cputime_t cputime_scaled)
163 u64 *cpustat = kcpustat_this_cpu->cpustat;
165 /* Add guest time to process. */
167 p->utimescaled += cputime_scaled;
168 account_group_user_time(p, cputime);
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;
176 cpustat[CPUTIME_USER] += (__force u64) cputime;
177 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
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
189 void __account_system_time(struct task_struct *p, cputime_t cputime,
190 cputime_t cputime_scaled, int index)
192 /* Add system time to process. */
194 p->stimescaled += cputime_scaled;
195 account_group_system_time(p, cputime);
197 /* Add system time to cpustat. */
198 task_group_account_field(p, index, (__force u64) cputime);
200 /* Account for system time used */
201 acct_account_cputime(p);
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
211 void account_system_time(struct task_struct *p, int hardirq_offset,
212 cputime_t cputime, cputime_t cputime_scaled)
216 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
217 account_guest_time(p, cputime, cputime_scaled);
221 if (hardirq_count() - hardirq_offset)
223 else if (in_serving_softirq())
224 index = CPUTIME_SOFTIRQ;
226 index = CPUTIME_SYSTEM;
228 __account_system_time(p, cputime, cputime_scaled, index);
232 * Account for involuntary wait time.
233 * @cputime: the cpu time spent in involuntary wait
235 void account_steal_time(cputime_t cputime)
237 u64 *cpustat = kcpustat_this_cpu->cpustat;
239 cpustat[CPUTIME_STEAL] += (__force u64) cputime;
243 * Account for idle time.
244 * @cputime: the cpu time spent in idle wait
246 void account_idle_time(cputime_t cputime)
248 u64 *cpustat = kcpustat_this_cpu->cpustat;
249 struct rq *rq = this_rq();
251 if (atomic_read(&rq->nr_iowait) > 0)
252 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
254 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
257 static __always_inline bool steal_account_process_tick(void)
259 #ifdef CONFIG_PARAVIRT
260 if (static_key_false(¶virt_steal_enabled)) {
264 steal = paravirt_steal_clock(smp_processor_id());
265 steal -= this_rq()->prev_steal_time;
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.
272 steal_ct = nsecs_to_cputime(steal);
273 this_rq()->prev_steal_time += cputime_to_nsecs(steal_ct);
275 account_steal_time(steal_ct);
283 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
284 * tasks (sum on group iteration) belonging to @tsk's group.
286 void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
288 struct signal_struct *sig = tsk->signal;
289 cputime_t utime, stime;
290 struct task_struct *t;
291 unsigned int seq, nextseq;
295 /* Attempt a lockless read on the first round. */
299 flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
300 times->utime = sig->utime;
301 times->stime = sig->stime;
302 times->sum_exec_runtime = sig->sum_sched_runtime;
304 for_each_thread(tsk, t) {
305 task_cputime(t, &utime, &stime);
306 times->utime += utime;
307 times->stime += stime;
308 times->sum_exec_runtime += task_sched_runtime(t);
310 /* If lockless access failed, take the lock. */
312 } while (need_seqretry(&sig->stats_lock, seq));
313 done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
317 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
319 * Account a tick to a process and cpustat
320 * @p: the process that the cpu time gets accounted to
321 * @user_tick: is the tick from userspace
322 * @rq: the pointer to rq
324 * Tick demultiplexing follows the order
325 * - pending hardirq update
326 * - pending softirq update
330 * - check for guest_time
331 * - else account as system_time
333 * Check for hardirq is done both for system and user time as there is
334 * no timer going off while we are on hardirq and hence we may never get an
335 * opportunity to update it solely in system time.
336 * p->stime and friends are only updated on system time and not on irq
337 * softirq as those do not count in task exec_runtime any more.
339 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
340 struct rq *rq, int ticks)
342 cputime_t scaled = cputime_to_scaled(cputime_one_jiffy);
343 u64 cputime = (__force u64) cputime_one_jiffy;
344 u64 *cpustat = kcpustat_this_cpu->cpustat;
346 if (steal_account_process_tick())
352 if (irqtime_account_hi_update()) {
353 cpustat[CPUTIME_IRQ] += cputime;
354 } else if (irqtime_account_si_update()) {
355 cpustat[CPUTIME_SOFTIRQ] += cputime;
356 } else if (this_cpu_ksoftirqd() == p) {
358 * ksoftirqd time do not get accounted in cpu_softirq_time.
359 * So, we have to handle it separately here.
360 * Also, p->stime needs to be updated for ksoftirqd.
362 __account_system_time(p, cputime, scaled, CPUTIME_SOFTIRQ);
363 } else if (user_tick) {
364 account_user_time(p, cputime, scaled);
365 } else if (p == rq->idle) {
366 account_idle_time(cputime);
367 } else if (p->flags & PF_VCPU) { /* System time or guest time */
368 account_guest_time(p, cputime, scaled);
370 __account_system_time(p, cputime, scaled, CPUTIME_SYSTEM);
374 static void irqtime_account_idle_ticks(int ticks)
376 struct rq *rq = this_rq();
378 irqtime_account_process_tick(current, 0, rq, ticks);
380 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
381 static inline void irqtime_account_idle_ticks(int ticks) {}
382 static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
383 struct rq *rq, int nr_ticks) {}
384 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
387 * Use precise platform statistics if available:
389 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
391 #ifndef __ARCH_HAS_VTIME_TASK_SWITCH
392 void vtime_common_task_switch(struct task_struct *prev)
394 if (is_idle_task(prev))
395 vtime_account_idle(prev);
397 vtime_account_system(prev);
399 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
400 vtime_account_user(prev);
402 arch_vtime_task_switch(prev);
407 * Archs that account the whole time spent in the idle task
408 * (outside irq) as idle time can rely on this and just implement
409 * vtime_account_system() and vtime_account_idle(). Archs that
410 * have other meaning of the idle time (s390 only includes the
411 * time spent by the CPU when it's in low power mode) must override
414 #ifndef __ARCH_HAS_VTIME_ACCOUNT
415 void vtime_common_account_irq_enter(struct task_struct *tsk)
417 if (!in_interrupt()) {
419 * If we interrupted user, context_tracking_in_user()
420 * is 1 because the context tracking don't hook
421 * on irq entry/exit. This way we know if
422 * we need to flush user time on kernel entry.
424 if (context_tracking_in_user()) {
425 vtime_account_user(tsk);
429 if (is_idle_task(tsk)) {
430 vtime_account_idle(tsk);
434 vtime_account_system(tsk);
436 EXPORT_SYMBOL_GPL(vtime_common_account_irq_enter);
437 #endif /* __ARCH_HAS_VTIME_ACCOUNT */
438 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
441 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
442 void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
447 EXPORT_SYMBOL_GPL(task_cputime_adjusted);
449 void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
451 struct task_cputime cputime;
453 thread_group_cputime(p, &cputime);
458 #else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
460 * Account a single tick of cpu time.
461 * @p: the process that the cpu time gets accounted to
462 * @user_tick: indicates if the tick is a user or a system tick
464 void account_process_tick(struct task_struct *p, int user_tick)
466 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
467 struct rq *rq = this_rq();
469 if (vtime_accounting_enabled())
472 if (sched_clock_irqtime) {
473 irqtime_account_process_tick(p, user_tick, rq, 1);
477 if (steal_account_process_tick())
481 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
482 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
483 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
486 account_idle_time(cputime_one_jiffy);
490 * Account multiple ticks of steal time.
491 * @p: the process from which the cpu time has been stolen
492 * @ticks: number of stolen ticks
494 void account_steal_ticks(unsigned long ticks)
496 account_steal_time(jiffies_to_cputime(ticks));
500 * Account multiple ticks of idle time.
501 * @ticks: number of stolen ticks
503 void account_idle_ticks(unsigned long ticks)
506 if (sched_clock_irqtime) {
507 irqtime_account_idle_ticks(ticks);
511 account_idle_time(jiffies_to_cputime(ticks));
515 * Perform (stime * rtime) / total, but avoid multiplication overflow by
516 * loosing precision when the numbers are big.
518 static cputime_t scale_stime(u64 stime, u64 rtime, u64 total)
523 /* Make sure "rtime" is the bigger of stime/rtime */
527 /* Make sure 'total' fits in 32 bits */
531 /* Does rtime (and thus stime) fit in 32 bits? */
535 /* Can we just balance rtime/stime rather than dropping bits? */
539 /* We can grow stime and shrink rtime and try to make them both fit */
545 /* We drop from rtime, it has more bits than stime */
551 * Make sure gcc understands that this is a 32x32->64 multiply,
552 * followed by a 64/32->64 divide.
554 scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
555 return (__force cputime_t) scaled;
559 * Adjust tick based cputime random precision against scheduler runtime
562 * Tick based cputime accounting depend on random scheduling timeslices of a
563 * task to be interrupted or not by the timer. Depending on these
564 * circumstances, the number of these interrupts may be over or
565 * under-optimistic, matching the real user and system cputime with a variable
568 * Fix this by scaling these tick based values against the total runtime
569 * accounted by the CFS scheduler.
571 * This code provides the following guarantees:
573 * stime + utime == rtime
574 * stime_i+1 >= stime_i, utime_i+1 >= utime_i
576 * Assuming that rtime_i+1 >= rtime_i.
578 static void cputime_adjust(struct task_cputime *curr,
579 struct prev_cputime *prev,
580 cputime_t *ut, cputime_t *st)
582 cputime_t rtime, stime, utime;
585 /* Serialize concurrent callers such that we can honour our guarantees */
586 raw_spin_lock_irqsave(&prev->lock, flags);
587 rtime = nsecs_to_cputime(curr->sum_exec_runtime);
590 * This is possible under two circumstances:
591 * - rtime isn't monotonic after all (a bug);
592 * - we got reordered by the lock.
594 * In both cases this acts as a filter such that the rest of the code
595 * can assume it is monotonic regardless of anything else.
597 if (prev->stime + prev->utime >= rtime)
613 stime = scale_stime((__force u64)stime, (__force u64)rtime,
614 (__force u64)(stime + utime));
617 * Make sure stime doesn't go backwards; this preserves monotonicity
618 * for utime because rtime is monotonic.
620 * utime_i+1 = rtime_i+1 - stime_i
621 * = rtime_i+1 - (rtime_i - utime_i)
622 * = (rtime_i+1 - rtime_i) + utime_i
625 if (stime < prev->stime)
627 utime = rtime - stime;
630 * Make sure utime doesn't go backwards; this still preserves
631 * monotonicity for stime, analogous argument to above.
633 if (utime < prev->utime) {
635 stime = rtime - utime;
644 raw_spin_unlock_irqrestore(&prev->lock, flags);
647 void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
649 struct task_cputime cputime = {
650 .sum_exec_runtime = p->se.sum_exec_runtime,
653 task_cputime(p, &cputime.utime, &cputime.stime);
654 cputime_adjust(&cputime, &p->prev_cputime, ut, st);
656 EXPORT_SYMBOL_GPL(task_cputime_adjusted);
658 void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
660 struct task_cputime cputime;
662 thread_group_cputime(p, &cputime);
663 cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
665 #endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
667 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
668 static unsigned long long vtime_delta(struct task_struct *tsk)
670 unsigned long long clock;
672 clock = local_clock();
673 if (clock < tsk->vtime_snap)
676 return clock - tsk->vtime_snap;
679 static cputime_t get_vtime_delta(struct task_struct *tsk)
681 unsigned long long delta = vtime_delta(tsk);
683 WARN_ON_ONCE(tsk->vtime_snap_whence == VTIME_SLEEPING);
684 tsk->vtime_snap += delta;
686 /* CHECKME: always safe to convert nsecs to cputime? */
687 return nsecs_to_cputime(delta);
690 static void __vtime_account_system(struct task_struct *tsk)
692 cputime_t delta_cpu = get_vtime_delta(tsk);
694 account_system_time(tsk, irq_count(), delta_cpu, cputime_to_scaled(delta_cpu));
697 void vtime_account_system(struct task_struct *tsk)
699 write_seqlock(&tsk->vtime_seqlock);
700 __vtime_account_system(tsk);
701 write_sequnlock(&tsk->vtime_seqlock);
704 void vtime_gen_account_irq_exit(struct task_struct *tsk)
706 write_seqlock(&tsk->vtime_seqlock);
707 __vtime_account_system(tsk);
708 if (context_tracking_in_user())
709 tsk->vtime_snap_whence = VTIME_USER;
710 write_sequnlock(&tsk->vtime_seqlock);
713 void vtime_account_user(struct task_struct *tsk)
717 write_seqlock(&tsk->vtime_seqlock);
718 delta_cpu = get_vtime_delta(tsk);
719 tsk->vtime_snap_whence = VTIME_SYS;
720 account_user_time(tsk, delta_cpu, cputime_to_scaled(delta_cpu));
721 write_sequnlock(&tsk->vtime_seqlock);
724 void vtime_user_enter(struct task_struct *tsk)
726 write_seqlock(&tsk->vtime_seqlock);
727 __vtime_account_system(tsk);
728 tsk->vtime_snap_whence = VTIME_USER;
729 write_sequnlock(&tsk->vtime_seqlock);
732 void vtime_guest_enter(struct task_struct *tsk)
735 * The flags must be updated under the lock with
736 * the vtime_snap flush and update.
737 * That enforces a right ordering and update sequence
738 * synchronization against the reader (task_gtime())
739 * that can thus safely catch up with a tickless delta.
741 write_seqlock(&tsk->vtime_seqlock);
742 __vtime_account_system(tsk);
743 current->flags |= PF_VCPU;
744 write_sequnlock(&tsk->vtime_seqlock);
746 EXPORT_SYMBOL_GPL(vtime_guest_enter);
748 void vtime_guest_exit(struct task_struct *tsk)
750 write_seqlock(&tsk->vtime_seqlock);
751 __vtime_account_system(tsk);
752 current->flags &= ~PF_VCPU;
753 write_sequnlock(&tsk->vtime_seqlock);
755 EXPORT_SYMBOL_GPL(vtime_guest_exit);
757 void vtime_account_idle(struct task_struct *tsk)
759 cputime_t delta_cpu = get_vtime_delta(tsk);
761 account_idle_time(delta_cpu);
764 void arch_vtime_task_switch(struct task_struct *prev)
766 write_seqlock(&prev->vtime_seqlock);
767 prev->vtime_snap_whence = VTIME_SLEEPING;
768 write_sequnlock(&prev->vtime_seqlock);
770 write_seqlock(¤t->vtime_seqlock);
771 current->vtime_snap_whence = VTIME_SYS;
772 current->vtime_snap = sched_clock_cpu(smp_processor_id());
773 write_sequnlock(¤t->vtime_seqlock);
776 void vtime_init_idle(struct task_struct *t, int cpu)
780 write_seqlock_irqsave(&t->vtime_seqlock, flags);
781 t->vtime_snap_whence = VTIME_SYS;
782 t->vtime_snap = sched_clock_cpu(cpu);
783 write_sequnlock_irqrestore(&t->vtime_seqlock, flags);
786 cputime_t task_gtime(struct task_struct *t)
791 if (!context_tracking_is_enabled())
795 seq = read_seqbegin(&t->vtime_seqlock);
798 if (t->flags & PF_VCPU)
799 gtime += vtime_delta(t);
801 } while (read_seqretry(&t->vtime_seqlock, seq));
807 * Fetch cputime raw values from fields of task_struct and
808 * add up the pending nohz execution time since the last
812 fetch_task_cputime(struct task_struct *t,
813 cputime_t *u_dst, cputime_t *s_dst,
814 cputime_t *u_src, cputime_t *s_src,
815 cputime_t *udelta, cputime_t *sdelta)
818 unsigned long long delta;
824 seq = read_seqbegin(&t->vtime_seqlock);
831 /* Task is sleeping, nothing to add */
832 if (t->vtime_snap_whence == VTIME_SLEEPING ||
836 delta = vtime_delta(t);
839 * Task runs either in user or kernel space, add pending nohz time to
842 if (t->vtime_snap_whence == VTIME_USER || t->flags & PF_VCPU) {
845 if (t->vtime_snap_whence == VTIME_SYS)
848 } while (read_seqretry(&t->vtime_seqlock, seq));
852 void task_cputime(struct task_struct *t, cputime_t *utime, cputime_t *stime)
854 cputime_t udelta, sdelta;
856 fetch_task_cputime(t, utime, stime, &t->utime,
857 &t->stime, &udelta, &sdelta);
864 void task_cputime_scaled(struct task_struct *t,
865 cputime_t *utimescaled, cputime_t *stimescaled)
867 cputime_t udelta, sdelta;
869 fetch_task_cputime(t, utimescaled, stimescaled,
870 &t->utimescaled, &t->stimescaled, &udelta, &sdelta);
872 *utimescaled += cputime_to_scaled(udelta);
874 *stimescaled += cputime_to_scaled(sdelta);
876 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */