Re: [edk2-devel] [PATCH v2 2/2] ArmVirtPkg/ArmVirtQemu: Avoid early ID map on ThunderX

["Ard Biesheuvel" <ardb@kernel.org>]

On Sat, 20 May 2023 at 10:37, Oliver Steffen <osteffen@redhat.com> wrote:
>
> Quoting Ard Biesheuvel (2023-05-19 23:36:53)
> > On Fri, 19 May 2023 at 18:32, Oliver Steffen <osteffen@redhat.com> wrote:
> > >
> > >
> > > Hi all,
> > >
> > > I had another look at this and I can now reproduce the issue consistently,
> > > with a quite minimal setup, on recent Linux kernel, Qemu, and EDK2.
> > > It requires rebooting the guest in a tight loop. It happens in silent
> > > and verbose
> > > builds alike, but since the verbose ones are slowed down by the serial
> > > output, it
> > > takes longer to hit the issue.
> > > It is possible to reproduce it with the silent builds within a few minutes.
> > > For the verbose case I recommend running multiple Qemu instances in parallel (as
> > > many as the machine allows, in my case ~100).
> > >
> >
> > Thanks a lot for all these details, this is extremely helpful.
> >
> > So what appears to be happening is that we split the 2M block mapping
> > that covers the code that we were called from, and hit a level 2
> > translation fault because the updated page table entry is still
> > observed to be in its transient 'invalid' state as we return to it.
> >
> > Could you please check whether this makes a difference?
> >
> > --- a/ArmPkg/Library/ArmMmuLib/AArch64/ArmMmuLibReplaceEntry.S
> > +++ b/ArmPkg/Library/ArmMmuLib/AArch64/ArmMmuLibReplaceEntry.S
> > @@ -65,6 +65,7 @@
> >    // write updated entry
> >    str   x1, [x0]
> >    dsb   nshst
> > +  isb
> >
> >  .L2_\@:
> >    .endm
>
> That fixes it - no crash observed within 150k iterations.
> Thanks, Ard!
>

Fantastic! Thanks a lot for all the effort in tracking this down.