Re: spapr watchdog vs watchdog_perform_action() / QMP watchdog-set-action

[Peter Maydell]

On Mon, 29 Jan 2024 at 04:02, David Gibson <david@gibson.dropbear.id.au> wrote:
>
> On Sat, Jan 27, 2024 at 01:08:02PM +0000, Peter Maydell wrote:
> > On Fri, 26 Jan 2024 at 20:49, Markus Armbruster <armbru@redhat.com> wrote:
> > >
> > > Peter Maydell <peter.maydell@linaro.org> writes:
> > >
> > > > Hi; one of the "bitesized tasks" we have listed is to convert
> > > > watchdog timers which directly call qemu_system_reset_request() on
> > > > watchdog timeout to call watchdog_perform_action() instead. This
> > > > means they honour the QMP commands that let the user specifiy
> > > > the behaviour on watchdog expiry:
> > > > https://www.qemu.org/docs/master/interop/qemu-qmp-ref.html#qapidoc-141
> > > > https://www.qemu.org/docs/master/interop/qemu-qmp-ref.html#qapidoc-129
> > > > (choices include reset, power off the system, do nothing, etc).
> > > >
> > > > There are only a few remaining watchdogs that don't use the
> > > > watchdog_perform_action() function. In most cases the change
> > > > is obvious and easy: just make them do that instead of calling
> > > > qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET).
> > > >
> > > > However, the hw/watchdog/spapr_watchdog.c case is trickier. As
> > > > far as I can tell from the sources, this is a watchdog set up via
> > > > a hypercall, and the guest makes a choice of "power off, restart,
> > > > or dump and restart" for its on-expiry action.
> > > >
> > > > What should this watchdog's interaction with the watchdog-set-action
> > > > QMP command be? If the user says "do X" and the guest says "do Y",
> > > > which do we do? (With the current code, we always honour what
> > > > the guest asks for and ignore what the user asks for.)
> > >
> > > Gut reaction: when the user says "do X", the guest should not get a say.
> > > But one of the values of X could be "whatever the guest says".
>
> That would also be my inclination.
>
> > Mmm. Slightly awkwardly, we don't currently distinguish between
> > "action is reset because the user never expressed a preference"
> > and "action is reset because the user specifically asked for that",
> > but I guess in theory we could make that distinction. (Conveniently
> > there is no QMP action for "query current watchdog-action state",
> > so we don't need to worry about reflecting that distinction in the
> > QMP interface if we make it.)
>
> I think that change is necessary in order to accomodate this sort of
> watchdog with guest-progammable behaviour (which is part of the PAPR
> spec, so we shouldn't just ignore it).

I've now filed https://gitlab.com/qemu-project/qemu/-/issues/2185
to track the spapr-watchdog vs watchdog-set-action missing feature
that we've discussed in this thread. (I haven't marked it as a
"bite-sized feature", though I think it probably could be one if
somebody familiar with at least one of spapr or QMP wanted to
flesh out the required changes a bit.)

-- PMM