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 <linux/sched/cputime.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(struct irqtime, cpu_irqtime);
25 static int sched_clock_irqtime;
27 void enable_sched_clock_irqtime(void)
29 sched_clock_irqtime = 1;
32 void disable_sched_clock_irqtime(void)
34 sched_clock_irqtime = 0;
37 static void irqtime_account_delta(struct irqtime *irqtime, u64 delta,
38 enum cpu_usage_stat idx)
40 u64 *cpustat = kcpustat_this_cpu->cpustat;
42 u64_stats_update_begin(&irqtime->sync);
43 cpustat[idx] += delta;
44 irqtime->total += delta;
45 irqtime->tick_delta += delta;
46 u64_stats_update_end(&irqtime->sync);
50 * Called before incrementing preempt_count on {soft,}irq_enter
51 * and before decrementing preempt_count on {soft,}irq_exit.
53 void irqtime_account_irq(struct task_struct *curr)
55 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
59 if (!sched_clock_irqtime)
62 cpu = smp_processor_id();
63 delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
64 irqtime->irq_start_time += delta;
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.
73 irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
74 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
75 irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
77 EXPORT_SYMBOL_GPL(irqtime_account_irq);
79 static u64 irqtime_tick_accounted(u64 maxtime)
81 struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
84 delta = min(irqtime->tick_delta, maxtime);
85 irqtime->tick_delta -= delta;
90 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
92 #define sched_clock_irqtime (0)
94 static u64 irqtime_tick_accounted(u64 dummy)
99 #endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
101 static inline void task_group_account_field(struct task_struct *p, int index,
105 * Since all updates are sure to touch the root cgroup, we
106 * get ourselves ahead and touch it first. If the root cgroup
107 * is the only cgroup, then nothing else should be necessary.
110 __this_cpu_add(kernel_cpustat.cpustat[index], tmp);
112 cpuacct_account_field(p, index, tmp);
116 * Account user cpu time to a process.
117 * @p: the process that the cpu time gets accounted to
118 * @cputime: the cpu time spent in user space since the last update
120 void account_user_time(struct task_struct *p, u64 cputime)
124 /* Add user time to process. */
126 account_group_user_time(p, cputime);
128 index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
130 /* Add user time to cpustat. */
131 task_group_account_field(p, index, cputime);
133 /* Account for user time used */
134 acct_account_cputime(p);
138 * Account guest cpu time to a process.
139 * @p: the process that the cpu time gets accounted to
140 * @cputime: the cpu time spent in virtual machine since the last update
142 void account_guest_time(struct task_struct *p, u64 cputime)
144 u64 *cpustat = kcpustat_this_cpu->cpustat;
146 /* Add guest time to process. */
148 account_group_user_time(p, cputime);
151 /* Add guest time to cpustat. */
152 if (task_nice(p) > 0) {
153 cpustat[CPUTIME_NICE] += cputime;
154 cpustat[CPUTIME_GUEST_NICE] += cputime;
156 cpustat[CPUTIME_USER] += cputime;
157 cpustat[CPUTIME_GUEST] += cputime;
162 * Account system cpu time to a process and desired cpustat field
163 * @p: the process that the cpu time gets accounted to
164 * @cputime: the cpu time spent in kernel space since the last update
165 * @index: pointer to cpustat field that has to be updated
167 void account_system_index_time(struct task_struct *p,
168 u64 cputime, enum cpu_usage_stat index)
170 /* Add system time to process. */
172 account_group_system_time(p, cputime);
174 /* Add system time to cpustat. */
175 task_group_account_field(p, index, cputime);
177 /* Account for system time used */
178 acct_account_cputime(p);
182 * Account system cpu time to a process.
183 * @p: the process that the cpu time gets accounted to
184 * @hardirq_offset: the offset to subtract from hardirq_count()
185 * @cputime: the cpu time spent in kernel space since the last update
187 void account_system_time(struct task_struct *p, int hardirq_offset, u64 cputime)
191 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
192 account_guest_time(p, cputime);
196 if (hardirq_count() - hardirq_offset)
198 else if (in_serving_softirq())
199 index = CPUTIME_SOFTIRQ;
201 index = CPUTIME_SYSTEM;
203 account_system_index_time(p, cputime, index);
207 * Account for involuntary wait time.
208 * @cputime: the cpu time spent in involuntary wait
210 void account_steal_time(u64 cputime)
212 u64 *cpustat = kcpustat_this_cpu->cpustat;
214 cpustat[CPUTIME_STEAL] += cputime;
218 * Account for idle time.
219 * @cputime: the cpu time spent in idle wait
221 void account_idle_time(u64 cputime)
223 u64 *cpustat = kcpustat_this_cpu->cpustat;
224 struct rq *rq = this_rq();
226 if (atomic_read(&rq->nr_iowait) > 0)
227 cpustat[CPUTIME_IOWAIT] += cputime;
229 cpustat[CPUTIME_IDLE] += cputime;
233 * When a guest is interrupted for a longer amount of time, missed clock
234 * ticks are not redelivered later. Due to that, this function may on
235 * occasion account more time than the calling functions think elapsed.
237 static __always_inline u64 steal_account_process_time(u64 maxtime)
239 #ifdef CONFIG_PARAVIRT
240 if (static_key_false(¶virt_steal_enabled)) {
243 steal = paravirt_steal_clock(smp_processor_id());
244 steal -= this_rq()->prev_steal_time;
245 steal = min(steal, maxtime);
246 account_steal_time(steal);
247 this_rq()->prev_steal_time += steal;
256 * Account how much elapsed time was spent in steal, irq, or softirq time.
258 static inline u64 account_other_time(u64 max)
262 /* Shall be converted to a lockdep-enabled lightweight check */
263 WARN_ON_ONCE(!irqs_disabled());
265 accounted = steal_account_process_time(max);
268 accounted += irqtime_tick_accounted(max - accounted);
274 static inline u64 read_sum_exec_runtime(struct task_struct *t)
276 return t->se.sum_exec_runtime;
279 static u64 read_sum_exec_runtime(struct task_struct *t)
285 rq = task_rq_lock(t, &rf);
286 ns = t->se.sum_exec_runtime;
287 task_rq_unlock(rq, t, &rf);
294 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
295 * tasks (sum on group iteration) belonging to @tsk's group.
297 void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
299 struct signal_struct *sig = tsk->signal;
301 struct task_struct *t;
302 unsigned int seq, nextseq;
306 * Update current task runtime to account pending time since last
307 * scheduler action or thread_group_cputime() call. This thread group
308 * might have other running tasks on different CPUs, but updating
309 * their runtime can affect syscall performance, so we skip account
310 * those pending times and rely only on values updated on tick or
311 * other scheduler action.
313 if (same_thread_group(current, tsk))
314 (void) task_sched_runtime(current);
317 /* Attempt a lockless read on the first round. */
321 flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
322 times->utime = sig->utime;
323 times->stime = sig->stime;
324 times->sum_exec_runtime = sig->sum_sched_runtime;
326 for_each_thread(tsk, t) {
327 task_cputime(t, &utime, &stime);
328 times->utime += utime;
329 times->stime += stime;
330 times->sum_exec_runtime += read_sum_exec_runtime(t);
332 /* If lockless access failed, take the lock. */
334 } while (need_seqretry(&sig->stats_lock, seq));
335 done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
339 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
341 * Account a tick to a process and cpustat
342 * @p: the process that the cpu time gets accounted to
343 * @user_tick: is the tick from userspace
344 * @rq: the pointer to rq
346 * Tick demultiplexing follows the order
347 * - pending hardirq update
348 * - pending softirq update
352 * - check for guest_time
353 * - else account as system_time
355 * Check for hardirq is done both for system and user time as there is
356 * no timer going off while we are on hardirq and hence we may never get an
357 * opportunity to update it solely in system time.
358 * p->stime and friends are only updated on system time and not on irq
359 * softirq as those do not count in task exec_runtime any more.
361 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
362 struct rq *rq, int ticks)
364 u64 other, cputime = TICK_NSEC * ticks;
367 * When returning from idle, many ticks can get accounted at
368 * once, including some ticks of steal, irq, and softirq time.
369 * Subtract those ticks from the amount of time accounted to
370 * idle, or potentially user or system time. Due to rounding,
371 * other time can exceed ticks occasionally.
373 other = account_other_time(ULONG_MAX);
374 if (other >= cputime)
379 if (this_cpu_ksoftirqd() == p) {
381 * ksoftirqd time do not get accounted in cpu_softirq_time.
382 * So, we have to handle it separately here.
383 * Also, p->stime needs to be updated for ksoftirqd.
385 account_system_index_time(p, cputime, CPUTIME_SOFTIRQ);
386 } else if (user_tick) {
387 account_user_time(p, cputime);
388 } else if (p == rq->idle) {
389 account_idle_time(cputime);
390 } else if (p->flags & PF_VCPU) { /* System time or guest time */
391 account_guest_time(p, cputime);
393 account_system_index_time(p, cputime, CPUTIME_SYSTEM);
397 static void irqtime_account_idle_ticks(int ticks)
399 struct rq *rq = this_rq();
401 irqtime_account_process_tick(current, 0, rq, ticks);
403 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
404 static inline void irqtime_account_idle_ticks(int ticks) {}
405 static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
406 struct rq *rq, int nr_ticks) {}
407 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
410 * Use precise platform statistics if available:
412 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
414 #ifndef __ARCH_HAS_VTIME_TASK_SWITCH
415 void vtime_common_task_switch(struct task_struct *prev)
417 if (is_idle_task(prev))
418 vtime_account_idle(prev);
420 vtime_account_system(prev);
423 arch_vtime_task_switch(prev);
427 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
430 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
432 * Archs that account the whole time spent in the idle task
433 * (outside irq) as idle time can rely on this and just implement
434 * vtime_account_system() and vtime_account_idle(). Archs that
435 * have other meaning of the idle time (s390 only includes the
436 * time spent by the CPU when it's in low power mode) must override
439 #ifndef __ARCH_HAS_VTIME_ACCOUNT
440 void vtime_account_irq_enter(struct task_struct *tsk)
442 if (!in_interrupt() && is_idle_task(tsk))
443 vtime_account_idle(tsk);
445 vtime_account_system(tsk);
447 EXPORT_SYMBOL_GPL(vtime_account_irq_enter);
448 #endif /* __ARCH_HAS_VTIME_ACCOUNT */
450 void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
455 EXPORT_SYMBOL_GPL(task_cputime_adjusted);
457 void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
459 struct task_cputime cputime;
461 thread_group_cputime(p, &cputime);
466 #else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
468 * Account a single tick of cpu time.
469 * @p: the process that the cpu time gets accounted to
470 * @user_tick: indicates if the tick is a user or a system tick
472 void account_process_tick(struct task_struct *p, int user_tick)
475 struct rq *rq = this_rq();
477 if (vtime_accounting_cpu_enabled())
480 if (sched_clock_irqtime) {
481 irqtime_account_process_tick(p, user_tick, rq, 1);
486 steal = steal_account_process_time(ULONG_MAX);
488 if (steal >= cputime)
494 account_user_time(p, cputime);
495 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
496 account_system_time(p, HARDIRQ_OFFSET, cputime);
498 account_idle_time(cputime);
502 * Account multiple ticks of idle time.
503 * @ticks: number of stolen ticks
505 void account_idle_ticks(unsigned long ticks)
509 if (sched_clock_irqtime) {
510 irqtime_account_idle_ticks(ticks);
514 cputime = ticks * TICK_NSEC;
515 steal = steal_account_process_time(ULONG_MAX);
517 if (steal >= cputime)
521 account_idle_time(cputime);
525 * Perform (stime * rtime) / total, but avoid multiplication overflow by
526 * loosing precision when the numbers are big.
528 static u64 scale_stime(u64 stime, u64 rtime, u64 total)
533 /* Make sure "rtime" is the bigger of stime/rtime */
537 /* Make sure 'total' fits in 32 bits */
541 /* Does rtime (and thus stime) fit in 32 bits? */
545 /* Can we just balance rtime/stime rather than dropping bits? */
549 /* We can grow stime and shrink rtime and try to make them both fit */
555 /* We drop from rtime, it has more bits than stime */
561 * Make sure gcc understands that this is a 32x32->64 multiply,
562 * followed by a 64/32->64 divide.
564 scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
569 * Adjust tick based cputime random precision against scheduler runtime
572 * Tick based cputime accounting depend on random scheduling timeslices of a
573 * task to be interrupted or not by the timer. Depending on these
574 * circumstances, the number of these interrupts may be over or
575 * under-optimistic, matching the real user and system cputime with a variable
578 * Fix this by scaling these tick based values against the total runtime
579 * accounted by the CFS scheduler.
581 * This code provides the following guarantees:
583 * stime + utime == rtime
584 * stime_i+1 >= stime_i, utime_i+1 >= utime_i
586 * Assuming that rtime_i+1 >= rtime_i.
588 static void cputime_adjust(struct task_cputime *curr,
589 struct prev_cputime *prev,
592 u64 rtime, stime, utime;
595 /* Serialize concurrent callers such that we can honour our guarantees */
596 raw_spin_lock_irqsave(&prev->lock, flags);
597 rtime = curr->sum_exec_runtime;
600 * This is possible under two circumstances:
601 * - rtime isn't monotonic after all (a bug);
602 * - we got reordered by the lock.
604 * In both cases this acts as a filter such that the rest of the code
605 * can assume it is monotonic regardless of anything else.
607 if (prev->stime + prev->utime >= rtime)
614 * If either stime or both stime and utime are 0, assume all runtime is
615 * userspace. Once a task gets some ticks, the monotonicy code at
616 * 'update' will ensure things converge to the observed ratio.
628 stime = scale_stime(stime, rtime, stime + utime);
632 * Make sure stime doesn't go backwards; this preserves monotonicity
633 * for utime because rtime is monotonic.
635 * utime_i+1 = rtime_i+1 - stime_i
636 * = rtime_i+1 - (rtime_i - utime_i)
637 * = (rtime_i+1 - rtime_i) + utime_i
640 if (stime < prev->stime)
642 utime = rtime - stime;
645 * Make sure utime doesn't go backwards; this still preserves
646 * monotonicity for stime, analogous argument to above.
648 if (utime < prev->utime) {
650 stime = rtime - utime;
658 raw_spin_unlock_irqrestore(&prev->lock, flags);
661 void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
663 struct task_cputime cputime = {
664 .sum_exec_runtime = p->se.sum_exec_runtime,
667 task_cputime(p, &cputime.utime, &cputime.stime);
668 cputime_adjust(&cputime, &p->prev_cputime, ut, st);
670 EXPORT_SYMBOL_GPL(task_cputime_adjusted);
672 void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
674 struct task_cputime cputime;
676 thread_group_cputime(p, &cputime);
677 cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
679 #endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
681 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
682 static u64 vtime_delta(struct task_struct *tsk)
684 unsigned long now = READ_ONCE(jiffies);
686 if (time_before(now, (unsigned long)tsk->vtime_snap))
689 return jiffies_to_nsecs(now - tsk->vtime_snap);
692 static u64 get_vtime_delta(struct task_struct *tsk)
694 unsigned long now = READ_ONCE(jiffies);
698 * Unlike tick based timing, vtime based timing never has lost
699 * ticks, and no need for steal time accounting to make up for
700 * lost ticks. Vtime accounts a rounded version of actual
701 * elapsed time. Limit account_other_time to prevent rounding
702 * errors from causing elapsed vtime to go negative.
704 delta = jiffies_to_nsecs(now - tsk->vtime_snap);
705 other = account_other_time(delta);
706 WARN_ON_ONCE(tsk->vtime_snap_whence == VTIME_INACTIVE);
707 tsk->vtime_snap = now;
709 return delta - other;
712 static void __vtime_account_system(struct task_struct *tsk)
714 account_system_time(tsk, irq_count(), get_vtime_delta(tsk));
717 void vtime_account_system(struct task_struct *tsk)
719 if (!vtime_delta(tsk))
722 write_seqcount_begin(&tsk->vtime_seqcount);
723 __vtime_account_system(tsk);
724 write_seqcount_end(&tsk->vtime_seqcount);
727 void vtime_account_user(struct task_struct *tsk)
729 write_seqcount_begin(&tsk->vtime_seqcount);
730 tsk->vtime_snap_whence = VTIME_SYS;
731 if (vtime_delta(tsk))
732 account_user_time(tsk, get_vtime_delta(tsk));
733 write_seqcount_end(&tsk->vtime_seqcount);
736 void vtime_user_enter(struct task_struct *tsk)
738 write_seqcount_begin(&tsk->vtime_seqcount);
739 if (vtime_delta(tsk))
740 __vtime_account_system(tsk);
741 tsk->vtime_snap_whence = VTIME_USER;
742 write_seqcount_end(&tsk->vtime_seqcount);
745 void vtime_guest_enter(struct task_struct *tsk)
748 * The flags must be updated under the lock with
749 * the vtime_snap flush and update.
750 * That enforces a right ordering and update sequence
751 * synchronization against the reader (task_gtime())
752 * that can thus safely catch up with a tickless delta.
754 write_seqcount_begin(&tsk->vtime_seqcount);
755 if (vtime_delta(tsk))
756 __vtime_account_system(tsk);
757 current->flags |= PF_VCPU;
758 write_seqcount_end(&tsk->vtime_seqcount);
760 EXPORT_SYMBOL_GPL(vtime_guest_enter);
762 void vtime_guest_exit(struct task_struct *tsk)
764 write_seqcount_begin(&tsk->vtime_seqcount);
765 __vtime_account_system(tsk);
766 current->flags &= ~PF_VCPU;
767 write_seqcount_end(&tsk->vtime_seqcount);
769 EXPORT_SYMBOL_GPL(vtime_guest_exit);
771 void vtime_account_idle(struct task_struct *tsk)
773 account_idle_time(get_vtime_delta(tsk));
776 void arch_vtime_task_switch(struct task_struct *prev)
778 write_seqcount_begin(&prev->vtime_seqcount);
779 prev->vtime_snap_whence = VTIME_INACTIVE;
780 write_seqcount_end(&prev->vtime_seqcount);
782 write_seqcount_begin(¤t->vtime_seqcount);
783 current->vtime_snap_whence = VTIME_SYS;
784 current->vtime_snap = jiffies;
785 write_seqcount_end(¤t->vtime_seqcount);
788 void vtime_init_idle(struct task_struct *t, int cpu)
792 local_irq_save(flags);
793 write_seqcount_begin(&t->vtime_seqcount);
794 t->vtime_snap_whence = VTIME_SYS;
795 t->vtime_snap = jiffies;
796 write_seqcount_end(&t->vtime_seqcount);
797 local_irq_restore(flags);
800 u64 task_gtime(struct task_struct *t)
805 if (!vtime_accounting_enabled())
809 seq = read_seqcount_begin(&t->vtime_seqcount);
812 if (t->vtime_snap_whence == VTIME_SYS && t->flags & PF_VCPU)
813 gtime += vtime_delta(t);
815 } while (read_seqcount_retry(&t->vtime_seqcount, seq));
821 * Fetch cputime raw values from fields of task_struct and
822 * add up the pending nohz execution time since the last
825 void task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
830 if (!vtime_accounting_enabled()) {
837 seq = read_seqcount_begin(&t->vtime_seqcount);
842 /* Task is sleeping, nothing to add */
843 if (t->vtime_snap_whence == VTIME_INACTIVE || is_idle_task(t))
846 delta = vtime_delta(t);
849 * Task runs either in user or kernel space, add pending nohz time to
852 if (t->vtime_snap_whence == VTIME_USER || t->flags & PF_VCPU)
854 else if (t->vtime_snap_whence == VTIME_SYS)
856 } while (read_seqcount_retry(&t->vtime_seqcount, seq));
858 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */