1 // SPDX-License-Identifier: GPL-2.0-only
3 * Simple CPU accounting cgroup controller
7 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
10 * There are no locks covering percpu hardirq/softirq time.
11 * They are only modified in vtime_account, on corresponding CPU
12 * with interrupts disabled. So, writes are safe.
13 * They are read and saved off onto struct rq in update_rq_clock().
14 * This may result in other CPU reading this CPU's irq time and can
15 * race with irq/vtime_account on this CPU. We would either get old
16 * or new value with a side effect of accounting a slice of irq time to wrong
17 * task when irq is in progress while we read rq->clock. That is a worthy
18 * compromise in place of having locks on each irq in account_system_time.
20 DEFINE_PER_CPU(struct irqtime, cpu_irqtime);
22 static int sched_clock_irqtime;
24 void enable_sched_clock_irqtime(void)
26 sched_clock_irqtime = 1;
29 void disable_sched_clock_irqtime(void)
31 sched_clock_irqtime = 0;
34 static void irqtime_account_delta(struct irqtime *irqtime, u64 delta,
35 enum cpu_usage_stat idx)
37 u64 *cpustat = kcpustat_this_cpu->cpustat;
39 u64_stats_update_begin(&irqtime->sync);
40 cpustat[idx] += delta;
41 irqtime->total += delta;
42 irqtime->tick_delta += delta;
43 u64_stats_update_end(&irqtime->sync);
47 * Called before incrementing preempt_count on {soft,}irq_enter
48 * and before decrementing preempt_count on {soft,}irq_exit.
50 void irqtime_account_irq(struct task_struct *curr)
52 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
56 if (!sched_clock_irqtime)
59 cpu = smp_processor_id();
60 delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
61 irqtime->irq_start_time += delta;
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 irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
71 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
72 irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
74 EXPORT_SYMBOL_GPL(irqtime_account_irq);
76 static u64 irqtime_tick_accounted(u64 maxtime)
78 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
81 delta = min(irqtime->tick_delta, maxtime);
82 irqtime->tick_delta -= delta;
87 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
89 #define sched_clock_irqtime (0)
91 static u64 irqtime_tick_accounted(u64 dummy)
96 #endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
98 static inline void task_group_account_field(struct task_struct *p, int index,
102 * Since all updates are sure to touch the root cgroup, we
103 * get ourselves ahead and touch it first. If the root cgroup
104 * is the only cgroup, then nothing else should be necessary.
107 __this_cpu_add(kernel_cpustat.cpustat[index], tmp);
109 cgroup_account_cputime_field(p, index, tmp);
113 * Account user CPU time to a process.
114 * @p: the process that the CPU time gets accounted to
115 * @cputime: the CPU time spent in user space since the last update
117 void account_user_time(struct task_struct *p, u64 cputime)
121 /* Add user time to process. */
123 account_group_user_time(p, cputime);
125 index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
127 /* Add user time to cpustat. */
128 task_group_account_field(p, index, cputime);
130 /* Account for user time used */
131 acct_account_cputime(p);
135 * Account guest CPU time to a process.
136 * @p: the process that the CPU time gets accounted to
137 * @cputime: the CPU time spent in virtual machine since the last update
139 void account_guest_time(struct task_struct *p, u64 cputime)
141 u64 *cpustat = kcpustat_this_cpu->cpustat;
143 /* Add guest time to process. */
145 account_group_user_time(p, cputime);
148 /* Add guest time to cpustat. */
149 if (task_nice(p) > 0) {
150 task_group_account_field(p, CPUTIME_NICE, cputime);
151 cpustat[CPUTIME_GUEST_NICE] += cputime;
153 task_group_account_field(p, CPUTIME_USER, cputime);
154 cpustat[CPUTIME_GUEST] += cputime;
159 * Account system CPU time to a process and desired cpustat field
160 * @p: the process that the CPU time gets accounted to
161 * @cputime: the CPU time spent in kernel space since the last update
162 * @index: pointer to cpustat field that has to be updated
164 void account_system_index_time(struct task_struct *p,
165 u64 cputime, enum cpu_usage_stat index)
167 /* Add system time to process. */
169 account_group_system_time(p, cputime);
171 /* Add system time to cpustat. */
172 task_group_account_field(p, index, cputime);
174 /* Account for system time used */
175 acct_account_cputime(p);
179 * Account system CPU time to a process.
180 * @p: the process that the CPU time gets accounted to
181 * @hardirq_offset: the offset to subtract from hardirq_count()
182 * @cputime: the CPU time spent in kernel space since the last update
184 void account_system_time(struct task_struct *p, int hardirq_offset, u64 cputime)
188 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
189 account_guest_time(p, cputime);
193 if (hardirq_count() - hardirq_offset)
195 else if (in_serving_softirq())
196 index = CPUTIME_SOFTIRQ;
198 index = CPUTIME_SYSTEM;
200 account_system_index_time(p, cputime, index);
204 * Account for involuntary wait time.
205 * @cputime: the CPU time spent in involuntary wait
207 void account_steal_time(u64 cputime)
209 u64 *cpustat = kcpustat_this_cpu->cpustat;
211 cpustat[CPUTIME_STEAL] += cputime;
215 * Account for idle time.
216 * @cputime: the CPU time spent in idle wait
218 void account_idle_time(u64 cputime)
220 u64 *cpustat = kcpustat_this_cpu->cpustat;
221 struct rq *rq = this_rq();
223 if (atomic_read(&rq->nr_iowait) > 0)
224 cpustat[CPUTIME_IOWAIT] += cputime;
226 cpustat[CPUTIME_IDLE] += cputime;
230 * When a guest is interrupted for a longer amount of time, missed clock
231 * ticks are not redelivered later. Due to that, this function may on
232 * occasion account more time than the calling functions think elapsed.
234 static __always_inline u64 steal_account_process_time(u64 maxtime)
236 #ifdef CONFIG_PARAVIRT
237 if (static_key_false(¶virt_steal_enabled)) {
240 steal = paravirt_steal_clock(smp_processor_id());
241 steal -= this_rq()->prev_steal_time;
242 steal = min(steal, maxtime);
243 account_steal_time(steal);
244 this_rq()->prev_steal_time += steal;
253 * Account how much elapsed time was spent in steal, irq, or softirq time.
255 static inline u64 account_other_time(u64 max)
259 lockdep_assert_irqs_disabled();
261 accounted = steal_account_process_time(max);
264 accounted += irqtime_tick_accounted(max - accounted);
270 static inline u64 read_sum_exec_runtime(struct task_struct *t)
272 return t->se.sum_exec_runtime;
275 static u64 read_sum_exec_runtime(struct task_struct *t)
281 rq = task_rq_lock(t, &rf);
282 ns = t->se.sum_exec_runtime;
283 task_rq_unlock(rq, t, &rf);
290 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
291 * tasks (sum on group iteration) belonging to @tsk's group.
293 void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
295 struct signal_struct *sig = tsk->signal;
297 struct task_struct *t;
298 unsigned int seq, nextseq;
302 * Update current task runtime to account pending time since last
303 * scheduler action or thread_group_cputime() call. This thread group
304 * might have other running tasks on different CPUs, but updating
305 * their runtime can affect syscall performance, so we skip account
306 * those pending times and rely only on values updated on tick or
307 * other scheduler action.
309 if (same_thread_group(current, tsk))
310 (void) task_sched_runtime(current);
313 /* Attempt a lockless read on the first round. */
317 flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
318 times->utime = sig->utime;
319 times->stime = sig->stime;
320 times->sum_exec_runtime = sig->sum_sched_runtime;
322 for_each_thread(tsk, t) {
323 task_cputime(t, &utime, &stime);
324 times->utime += utime;
325 times->stime += stime;
326 times->sum_exec_runtime += read_sum_exec_runtime(t);
328 /* If lockless access failed, take the lock. */
330 } while (need_seqretry(&sig->stats_lock, seq));
331 done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
335 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
337 * Account a tick to a process and cpustat
338 * @p: the process that the CPU time gets accounted to
339 * @user_tick: is the tick from userspace
340 * @rq: the pointer to rq
342 * Tick demultiplexing follows the order
343 * - pending hardirq update
344 * - pending softirq update
348 * - check for guest_time
349 * - else account as system_time
351 * Check for hardirq is done both for system and user time as there is
352 * no timer going off while we are on hardirq and hence we may never get an
353 * opportunity to update it solely in system time.
354 * p->stime and friends are only updated on system time and not on irq
355 * softirq as those do not count in task exec_runtime any more.
357 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
358 struct rq *rq, int ticks)
360 u64 other, cputime = TICK_NSEC * ticks;
363 * When returning from idle, many ticks can get accounted at
364 * once, including some ticks of steal, irq, and softirq time.
365 * Subtract those ticks from the amount of time accounted to
366 * idle, or potentially user or system time. Due to rounding,
367 * other time can exceed ticks occasionally.
369 other = account_other_time(ULONG_MAX);
370 if (other >= cputime)
375 if (this_cpu_ksoftirqd() == p) {
377 * ksoftirqd time do not get accounted in cpu_softirq_time.
378 * So, we have to handle it separately here.
379 * Also, p->stime needs to be updated for ksoftirqd.
381 account_system_index_time(p, cputime, CPUTIME_SOFTIRQ);
382 } else if (user_tick) {
383 account_user_time(p, cputime);
384 } else if (p == rq->idle) {
385 account_idle_time(cputime);
386 } else if (p->flags & PF_VCPU) { /* System time or guest time */
387 account_guest_time(p, cputime);
389 account_system_index_time(p, cputime, CPUTIME_SYSTEM);
393 static void irqtime_account_idle_ticks(int ticks)
395 struct rq *rq = this_rq();
397 irqtime_account_process_tick(current, 0, rq, ticks);
399 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
400 static inline void irqtime_account_idle_ticks(int ticks) { }
401 static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
402 struct rq *rq, int nr_ticks) { }
403 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
406 * Use precise platform statistics if available:
408 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
409 # ifndef __ARCH_HAS_VTIME_TASK_SWITCH
410 void vtime_common_task_switch(struct task_struct *prev)
412 if (is_idle_task(prev))
413 vtime_account_idle(prev);
415 vtime_account_system(prev);
418 arch_vtime_task_switch(prev);
421 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
424 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
426 * Archs that account the whole time spent in the idle task
427 * (outside irq) as idle time can rely on this and just implement
428 * vtime_account_system() and vtime_account_idle(). Archs that
429 * have other meaning of the idle time (s390 only includes the
430 * time spent by the CPU when it's in low power mode) must override
433 #ifndef __ARCH_HAS_VTIME_ACCOUNT
434 void vtime_account_irq_enter(struct task_struct *tsk)
436 if (!in_interrupt() && is_idle_task(tsk))
437 vtime_account_idle(tsk);
439 vtime_account_system(tsk);
441 EXPORT_SYMBOL_GPL(vtime_account_irq_enter);
442 #endif /* __ARCH_HAS_VTIME_ACCOUNT */
444 void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
451 void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
456 EXPORT_SYMBOL_GPL(task_cputime_adjusted);
458 void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
460 struct task_cputime cputime;
462 thread_group_cputime(p, &cputime);
468 #else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE: */
471 * Account a single tick of CPU time.
472 * @p: the process that the CPU time gets accounted to
473 * @user_tick: indicates if the tick is a user or a system tick
475 void account_process_tick(struct task_struct *p, int user_tick)
478 struct rq *rq = this_rq();
480 if (vtime_accounting_cpu_enabled())
483 if (sched_clock_irqtime) {
484 irqtime_account_process_tick(p, user_tick, rq, 1);
489 steal = steal_account_process_time(ULONG_MAX);
491 if (steal >= cputime)
497 account_user_time(p, cputime);
498 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
499 account_system_time(p, HARDIRQ_OFFSET, cputime);
501 account_idle_time(cputime);
505 * Account multiple ticks of idle time.
506 * @ticks: number of stolen ticks
508 void account_idle_ticks(unsigned long ticks)
512 if (sched_clock_irqtime) {
513 irqtime_account_idle_ticks(ticks);
517 cputime = ticks * TICK_NSEC;
518 steal = steal_account_process_time(ULONG_MAX);
520 if (steal >= cputime)
524 account_idle_time(cputime);
528 * Perform (stime * rtime) / total, but avoid multiplication overflow by
529 * losing precision when the numbers are big.
531 static u64 scale_stime(u64 stime, u64 rtime, u64 total)
536 /* Make sure "rtime" is the bigger of stime/rtime */
540 /* Make sure 'total' fits in 32 bits */
544 /* Does rtime (and thus stime) fit in 32 bits? */
548 /* Can we just balance rtime/stime rather than dropping bits? */
552 /* We can grow stime and shrink rtime and try to make them both fit */
558 /* We drop from rtime, it has more bits than stime */
564 * Make sure gcc understands that this is a 32x32->64 multiply,
565 * followed by a 64/32->64 divide.
567 scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
572 * Adjust tick based cputime random precision against scheduler runtime
575 * Tick based cputime accounting depend on random scheduling timeslices of a
576 * task to be interrupted or not by the timer. Depending on these
577 * circumstances, the number of these interrupts may be over or
578 * under-optimistic, matching the real user and system cputime with a variable
581 * Fix this by scaling these tick based values against the total runtime
582 * accounted by the CFS scheduler.
584 * This code provides the following guarantees:
586 * stime + utime == rtime
587 * stime_i+1 >= stime_i, utime_i+1 >= utime_i
589 * Assuming that rtime_i+1 >= rtime_i.
591 void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
594 u64 rtime, stime, utime;
597 /* Serialize concurrent callers such that we can honour our guarantees */
598 raw_spin_lock_irqsave(&prev->lock, flags);
599 rtime = curr->sum_exec_runtime;
602 * This is possible under two circumstances:
603 * - rtime isn't monotonic after all (a bug);
604 * - we got reordered by the lock.
606 * In both cases this acts as a filter such that the rest of the code
607 * can assume it is monotonic regardless of anything else.
609 if (prev->stime + prev->utime >= rtime)
616 * If either stime or utime are 0, assume all runtime is userspace.
617 * Once a task gets some ticks, the monotonicy code at 'update:'
618 * will ensure things converge to the observed ratio.
630 stime = scale_stime(stime, rtime, stime + utime);
634 * Make sure stime doesn't go backwards; this preserves monotonicity
635 * for utime because rtime is monotonic.
637 * utime_i+1 = rtime_i+1 - stime_i
638 * = rtime_i+1 - (rtime_i - utime_i)
639 * = (rtime_i+1 - rtime_i) + utime_i
642 if (stime < prev->stime)
644 utime = rtime - stime;
647 * Make sure utime doesn't go backwards; this still preserves
648 * monotonicity for stime, analogous argument to above.
650 if (utime < prev->utime) {
652 stime = rtime - utime;
660 raw_spin_unlock_irqrestore(&prev->lock, flags);
663 void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
665 struct task_cputime cputime = {
666 .sum_exec_runtime = p->se.sum_exec_runtime,
669 task_cputime(p, &cputime.utime, &cputime.stime);
670 cputime_adjust(&cputime, &p->prev_cputime, ut, st);
672 EXPORT_SYMBOL_GPL(task_cputime_adjusted);
674 void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
676 struct task_cputime cputime;
678 thread_group_cputime(p, &cputime);
679 cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
681 #endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
683 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
684 static u64 vtime_delta(struct vtime *vtime)
686 unsigned long long clock;
688 clock = sched_clock();
689 if (clock < vtime->starttime)
692 return clock - vtime->starttime;
695 static u64 get_vtime_delta(struct vtime *vtime)
697 u64 delta = vtime_delta(vtime);
701 * Unlike tick based timing, vtime based timing never has lost
702 * ticks, and no need for steal time accounting to make up for
703 * lost ticks. Vtime accounts a rounded version of actual
704 * elapsed time. Limit account_other_time to prevent rounding
705 * errors from causing elapsed vtime to go negative.
707 other = account_other_time(delta);
708 WARN_ON_ONCE(vtime->state == VTIME_INACTIVE);
709 vtime->starttime += delta;
711 return delta - other;
714 static void __vtime_account_system(struct task_struct *tsk,
717 vtime->stime += get_vtime_delta(vtime);
718 if (vtime->stime >= TICK_NSEC) {
719 account_system_time(tsk, irq_count(), vtime->stime);
724 static void vtime_account_guest(struct task_struct *tsk,
727 vtime->gtime += get_vtime_delta(vtime);
728 if (vtime->gtime >= TICK_NSEC) {
729 account_guest_time(tsk, vtime->gtime);
734 void vtime_account_system(struct task_struct *tsk)
736 struct vtime *vtime = &tsk->vtime;
738 if (!vtime_delta(vtime))
741 write_seqcount_begin(&vtime->seqcount);
742 /* We might have scheduled out from guest path */
743 if (tsk->flags & PF_VCPU)
744 vtime_account_guest(tsk, vtime);
746 __vtime_account_system(tsk, vtime);
747 write_seqcount_end(&vtime->seqcount);
750 void vtime_user_enter(struct task_struct *tsk)
752 struct vtime *vtime = &tsk->vtime;
754 write_seqcount_begin(&vtime->seqcount);
755 __vtime_account_system(tsk, vtime);
756 vtime->state = VTIME_USER;
757 write_seqcount_end(&vtime->seqcount);
760 void vtime_user_exit(struct task_struct *tsk)
762 struct vtime *vtime = &tsk->vtime;
764 write_seqcount_begin(&vtime->seqcount);
765 vtime->utime += get_vtime_delta(vtime);
766 if (vtime->utime >= TICK_NSEC) {
767 account_user_time(tsk, vtime->utime);
770 vtime->state = VTIME_SYS;
771 write_seqcount_end(&vtime->seqcount);
774 void vtime_guest_enter(struct task_struct *tsk)
776 struct vtime *vtime = &tsk->vtime;
778 * The flags must be updated under the lock with
779 * the vtime_starttime flush and update.
780 * That enforces a right ordering and update sequence
781 * synchronization against the reader (task_gtime())
782 * that can thus safely catch up with a tickless delta.
784 write_seqcount_begin(&vtime->seqcount);
785 __vtime_account_system(tsk, vtime);
786 tsk->flags |= PF_VCPU;
787 write_seqcount_end(&vtime->seqcount);
789 EXPORT_SYMBOL_GPL(vtime_guest_enter);
791 void vtime_guest_exit(struct task_struct *tsk)
793 struct vtime *vtime = &tsk->vtime;
795 write_seqcount_begin(&vtime->seqcount);
796 vtime_account_guest(tsk, vtime);
797 tsk->flags &= ~PF_VCPU;
798 write_seqcount_end(&vtime->seqcount);
800 EXPORT_SYMBOL_GPL(vtime_guest_exit);
802 void vtime_account_idle(struct task_struct *tsk)
804 account_idle_time(get_vtime_delta(&tsk->vtime));
807 void arch_vtime_task_switch(struct task_struct *prev)
809 struct vtime *vtime = &prev->vtime;
811 write_seqcount_begin(&vtime->seqcount);
812 vtime->state = VTIME_INACTIVE;
813 write_seqcount_end(&vtime->seqcount);
815 vtime = ¤t->vtime;
817 write_seqcount_begin(&vtime->seqcount);
818 vtime->state = VTIME_SYS;
819 vtime->starttime = sched_clock();
820 write_seqcount_end(&vtime->seqcount);
823 void vtime_init_idle(struct task_struct *t, int cpu)
825 struct vtime *vtime = &t->vtime;
828 local_irq_save(flags);
829 write_seqcount_begin(&vtime->seqcount);
830 vtime->state = VTIME_SYS;
831 vtime->starttime = sched_clock();
832 write_seqcount_end(&vtime->seqcount);
833 local_irq_restore(flags);
836 u64 task_gtime(struct task_struct *t)
838 struct vtime *vtime = &t->vtime;
842 if (!vtime_accounting_enabled())
846 seq = read_seqcount_begin(&vtime->seqcount);
849 if (vtime->state == VTIME_SYS && t->flags & PF_VCPU)
850 gtime += vtime->gtime + vtime_delta(vtime);
852 } while (read_seqcount_retry(&vtime->seqcount, seq));
858 * Fetch cputime raw values from fields of task_struct and
859 * add up the pending nohz execution time since the last
862 void task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
864 struct vtime *vtime = &t->vtime;
868 if (!vtime_accounting_enabled()) {
875 seq = read_seqcount_begin(&vtime->seqcount);
880 /* Task is sleeping, nothing to add */
881 if (vtime->state == VTIME_INACTIVE || is_idle_task(t))
884 delta = vtime_delta(vtime);
887 * Task runs either in user or kernel space, add pending nohz time to
890 if (vtime->state == VTIME_USER || t->flags & PF_VCPU)
891 *utime += vtime->utime + delta;
892 else if (vtime->state == VTIME_SYS)
893 *stime += vtime->stime + delta;
894 } while (read_seqcount_retry(&vtime->seqcount, seq));
896 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */