On Tue, Jul 17, 2018 at 05:56:53AM -0700, Paul E. McKenney wrote:
> On Tue, Jul 17, 2018 at 10:19:08AM +0200, David Woodhouse wrote:
> > On Mon, 2018-07-16 at 08:40 -0700, Paul E. McKenney wrote:
> > > Most of the weekend was devoted to testing today's upcoming pull request,
> > > but I did get a bit more testing done on this.
> > >
> > > I was able to make this happen more often by tweaking rcutorture a
> > > bit, but I still do not yet have statistically significant results.
> > > Nevertheless, I have thus far only seen failures with David's patch or
> > > with both David's and my patch. And I actually got a full-up rcutorture
> > > failure (a too-short grace period) in addition to the aforementioned
> > > close calls.
> > >
> > > Over this coming week I expect to devote significant testing time to
> > > the commit just prior to David's in my stack. If I don't see failures
> > > on that commit, we will need to spent some quality time with the KVM
> > > folks on whether or not kvm_x86_ops->run() and friends have the option of
> > > failing to return, but instead causing control to pop up somewhere else.
> > > Or someone could tell me how I am being blind to some obvious bug in
> > > the two commits that allow RCU to treat KVM guest-OS execution as an
> > > extended quiescent state. ;-)
> >
> > One thing we can try, if my patch is implicated, is moving the calls to
> > rcu_kvm_en{ter,xit} closer to the actual VM entry. Let's try putting
> > them around the large asm block in arch/x86/kvm/vmx.c::vmx_vcpu_run()
> > for example. If that fixes it, then we know we've missed something else
> > interesting that's happening in the middle.
>
> I don't have enough data to say anything with too much certainty, but
> my patch has rcu_kvm_en{ter,xit}() quite a bit farther apart than yours
> does, and I am not seeing massive increases in error rate in my patch
> compared to yours. Which again might or might not mean anything.
>
> Plus I haven't proven that your patch isn't an innocent bystander yet.
> If it isn't just an innocent bystander, that will take most of this
> week do demonstrate given current failure rates.
And I finally did get some near misses from an earlier commit, so we
should consider your patch to be officially off the hook.
Thanx, Paul
> I am also working on improving rcutorture diagnostics which should help
> me work out how to change rcutorture so as to find this more quickly.
>
> > Testing on Skylake shows a guest CPUID goes from ~3000 cycles to ~3500
> > with this patch, so in the next iteration it definitely needs to be
> > ifdef CONFIG_NO_HZ_FULL anyway, because it's actually required there
> > (AFAICT) and it's too expensive otherwise as Christian pointed out.
>
> Makes sense!
>
> Thanx, Paul
