On Wed, Dec 05, 2018 at 03:02:45PM +0100, Luc Michel wrote:
> On 12/4/18 8:45 PM, Eduardo Habkost wrote:
> > On Tue, Dec 04, 2018 at 07:16:39PM +0000, Peter Maydell wrote:
> >> On Tue, 4 Dec 2018 at 19:05, Eduardo Habkost <ehabk...@redhat.com> wrote:
> >>> On Tue, Dec 04, 2018 at 06:24:19PM +0000, Peter Maydell wrote:
> >>>> A cluster is a group of CPUs which are all identical and have
> >>>> the same view of the rest of the system.
> >>
> >>> With that definition in mind, why can't QEMU cluster CPUs
> >>> automatically by looking at CPU models and address space objects?
> >>
> >> That sounds like it is in theory feasible and in practice
> >> quite tricky. You would have to look not just at the CPU
> >> model name but also introspect all its properties for
> >> ones which change features of the CPU and are set differently
> >> on different CPUs (and I don't think there's any way to
> >> automatically tell which properties are ones which make
> >> the CPU different for which-cluster purposes and which aren't).
> >> And if we automatically checked whether address space objects
> >> were the same it would rule out implementing devices with
> >> per-cpu banked memory mapped registers by mapping different
> >> things into the AS for each CPU (though that's a hypothetical
> >> at the moment -- I've thought about implementing stuff that
> >> way but we tend to implement that sort of thing by looking
> >> at current_cpu->cpu_index at the moment).
> >
> > I see.
> >
> > Can't we at least do something to make sure the cluster objects
> > make sense? e.g. by ensuring at least QOM CPU type is the same,
> > and that cpu->address_space somehow points to
> > cpu->cluster->address_space?
> Where such a check should be placed? Cluster realize function is not
> good since children can still be added after device realization. A good
> place would be when object_property_add_child() is called, but I'm not
> aware of a way of hooking into that with the QOM API...
Not sure. Maybe it can be a post-machine_init hook. Maybe this
can be done by a separate (avocado_qemu-based?) test script.
It looks like address space info is available through HMP ("info
mtree") and not QMP, so we'd need to fix that first. CCing HMP
and QMP maintainers.
I don't think this should block the series, but it would be a
welcome addition.
--
Eduardo
