Re: Utime and stime are less when getrusage (RUSAGE_THREAD) is executed on a tickless CPU.

[Frederic Weisbecker <frederic@kernel.org>]

On Wed, May 19, 2021 at 11:24:58AM +0200, Peter Zijlstra wrote:
> On Wed, May 19, 2021 at 06:30:36AM +0000, hasegawa-hitomi@fujitsu.com wrote:
> > Hi Ingo, Peter, Juri, and Vincent.
> > 
> > 
> > > Your email is malformed.
> > 
> > I'm sorry. I was sent in the wrong format. I correct it and resend.
> > Thank you, Peter, for pointing this out.
> > 
> > 
> > I found that when I run getrusage(RUSAGE_THREAD) on a tickless CPU,
> > the utime and stime I get are less than the actual time, unlike when I run
> > getrusage(RUSAGE_SELF) on a single thread.
> > This problem seems to be caused by the fact that se.sum_exec_runtime is not
> > updated just before getting the information from 'current'.
> > In the current implementation, task_cputime_adjusted() calls task_cputime() to
> > get the 'current' utime and stime, then calls cputime_adjust() to adjust the
> > sum of utime and stime to be equal to cputime.sum_exec_runtime. On a tickless
> > CPU, sum_exec_runtime is not updated periodically, so there seems to be a
> > discrepancy with the actual time.
> > Therefore, I think I should include a process to update se.sum_exec_runtime
> > just before getting the information from 'current' (as in other processes
> > except RUSAGE_THREAD). I'm thinking of the following improvement.
> > 
> > @@ void getrusage(struct task_struct *p, int who, struct rusage *r)
> >         if (who == RUSAGE_THREAD) {
> > +               task_sched_runtime(current);
> >                 task_cputime_adjusted(current, &utime, &stime);
> > 
> > Is there any possible problem with this?
> 
> Would be superfluous for CONFIG_VIRT_CPU_ACCOUNTING_NATIVE=y
> architectures at the very least.
> 
> It also doesn't help any of the other callers, like for example procfs.
> 
> Something like the below ought to work and fix all variants I think. But
> it does make the call significantly more expensive.
> 
> Looking at thread_group_cputime() that already does something like this,
> but that's also susceptible to a variant of this very same issue; since
> it doesn't call it unconditionally, nor on all tasks, so if current
> isn't part of the threadgroup and/or another task is on a nohz_full cpu,
> things will go wobbly again.
> 
> There's a note about syscall performance there, so clearly someone seems
> to care about that aspect of things, but it does suck for nohz_full.
> 
> Frederic, didn't we have remote ticks that should help with this stuff?
> 
> And mostly I think the trade-off here is that if you run on nohz_full,
> you're not expected to go do syscalls anyway (because they're sodding
> expensive) and hence the accuracy of these sort of things is mostly
> irrelevant.
> 
> So it might be the use-case is just fundamentally bonkers and we
> shouldn't really bother fixing this.
> 
> Anyway?
> 
> ---
> diff --git a/kernel/sched/cputime.c b/kernel/sched/cputime.c
> index 872e481d5098..620871c8e4f8 100644
> --- a/kernel/sched/cputime.c
> +++ b/kernel/sched/cputime.c
> @@ -612,7 +612,7 @@ void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
>  void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
>  {
>  	struct task_cputime cputime = {
> -		.sum_exec_runtime = p->se.sum_exec_runtime,
> +		.sum_exec_runtime = task_sched_runtime(p),
>  	};
>  
>  	task_cputime(p, &cputime.utime, &cputime.stime);

If necessary I guess we can do something like the below, which
would only add the overhead where it's required:

diff --git a/include/linux/sched/cputime.h b/include/linux/sched/cputime.h
index 6c9f19a33865..ce3c58286062 100644
--- a/include/linux/sched/cputime.h
+++ b/include/linux/sched/cputime.h
@@ -18,15 +18,16 @@
 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
 
 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
-extern void task_cputime(struct task_struct *t,
+extern bool task_cputime(struct task_struct *t,
 			 u64 *utime, u64 *stime);
 extern u64 task_gtime(struct task_struct *t);
 #else
-static inline void task_cputime(struct task_struct *t,
+static inline bool task_cputime(struct task_struct *t,
 				u64 *utime, u64 *stime)
 {
 	*utime = t->utime;
 	*stime = t->stime;
+	return false;
 }
 
 static inline u64 task_gtime(struct task_struct *t)
diff --git a/kernel/sched/cputime.c b/kernel/sched/cputime.c
index 872e481d5098..9392aea1804e 100644
--- a/kernel/sched/cputime.c
+++ b/kernel/sched/cputime.c
@@ -615,7 +615,8 @@ void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
 		.sum_exec_runtime = p->se.sum_exec_runtime,
 	};
 
-	task_cputime(p, &cputime.utime, &cputime.stime);
+	if (task_cputime(p, &cputime.utime, &cputime.stime))
+		cputime.sum_exec_runtime = task_sched_runtime(p);
 	cputime_adjust(&cputime, &p->prev_cputime, ut, st);
 }
 EXPORT_SYMBOL_GPL(task_cputime_adjusted);
@@ -828,19 +829,21 @@ u64 task_gtime(struct task_struct *t)
  * add up the pending nohz execution time since the last
  * cputime snapshot.
  */
-void task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
+bool task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
 {
 	struct vtime *vtime = &t->vtime;
 	unsigned int seq;
 	u64 delta;
+	int ret;
 
 	if (!vtime_accounting_enabled()) {
 		*utime = t->utime;
 		*stime = t->stime;
-		return;
+		return false;
 	}
 
 	do {
+		ret = false;
 		seq = read_seqcount_begin(&vtime->seqcount);
 
 		*utime = t->utime;
@@ -850,6 +853,7 @@ void task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
 		if (vtime->state < VTIME_SYS)
 			continue;
 
+		ret = true;
 		delta = vtime_delta(vtime);
 
 		/*
@@ -861,6 +865,8 @@ void task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
 		else
 			*utime += vtime->utime + delta;
 	} while (read_seqcount_retry(&vtime->seqcount, seq));
+
+	return ret;
 }
 
 static int vtime_state_fetch(struct vtime *vtime, int cpu)