[PATCH] target/riscv: Exit current TB after an sfence.vma

[Alistair Francis]

On Thu, Mar 31, 2022 at 2:36 PM Palmer Dabbelt <palmer at dabbelt.com> wrote:
>
> On Wed, 30 Mar 2022 20:23:21 PDT (-0700), alistair23 at gmail.com wrote:
> > On Thu, Mar 31, 2022 at 3:11 AM Idan Horowitz <idan.horowitz at gmail.com> wrote:
> >>
> >> On Wed, 30 Mar 2022 at 19:11, Palmer Dabbelt <palmer at dabbelt.com> wrote:
> >> >
> >> >
> >> > Presumably you mean "revert" here?  That might be the right way to go,
> >> > just to avoid breaking users (even if we fix the kernel bug, it'll take
> >> > a while to get everyone to update).  That said, this smells like the
> >> > sort of thing that's going to crop up at arbitrary times in dynamic
> >> > systems so while a revert looks like it'd work around the boot issue we
> >> > might be making more headaches for folks down the road.
> >> >
> >>
> >> The opposite in fact, I did not suggest to revert it, but rather undo
> >> the revert (as Alistair already removed it from the apply-next tree),
> >> since my original patch fixes buggy behaviour that is blocking the
> >> testing of some embedded software on QEMU.
>
> Ah, sorry -- the QEMU tree I was looking at still had the patch in
> there, must have just been an old one.
>
> > So, this is a little tricky.
> >
> > We want to apply the fix, but that will break current users.
> >
> > Once the fix is merged into Linux we can apply it here. That should
> > hopefully be right at the start of the 7.1 QEMU development window,
> > which should give time for the fix to propagate into stable kernels
> > and not break too many people by the time QEMU is released.
>
> If you think this is a Linux bug then that makes sense, but I think this
> is a QEMU bug -- I sent a patch, not sure if it went through as it didn't
> make it to lore.

Ah whoops. I saw the patch but didn't read it, then I assumed it was a
Linux bug from your diff earlier.

>
> I also think the bug will manifest without the TB exit patch, maybe in
> single-step mode and definately if we happen to exit the TB at that
> point for other reasons.  Assuming my reasoning is correct in that
> patch, we may also be hitting this as arbitrary corruption anywhere.
> I'd started to write up a "QEMU errata" Linux patch for this, but then
> convinced myself that just adding the sfence.vma was insufficient.

Yeah, looking at it now I agree, I'll send a PR for 7.0.

Alistair