On Fri, Feb 05, 2016 at 10:37:42AM +0000, Marc Zyngier wrote:
> On 05/02/16 09:23, Andrew Jones wrote:
> > On Thu, Feb 04, 2016 at 06:51:06PM +0000, Marc Zyngier wrote:
> >> Hi Drew,
> >>
> >> On 04/02/16 18:38, Andrew Jones wrote:
> >>>
> >>> Hi Marc and Andre,
> >>>
> >>> I completely understand why reset_mpidr() limits Aff0 to 16, thanks
> >>> to Andre's nice comment about ICC_SGIxR. Now, here's my question;
> >>> it seems that the Cortex-A{53,57,72} manuals want to further limit
> >>> Aff0 to 4, going so far as to say bits 7:2 are RES0. I'm looking
> >>> at userspace dictating the MPIDR for KVM. QEMU tries to model the
> >>> A57 right now, so to be true to the manual, Aff0 should only address
> >>> four PEs, but that would generate a higher trap cost for SGI broadcasts
> >>> when using KVM. Sigh... what to do?
> >>
> >> There are two things to consider:
> >>
> >> - The GICv3 architecture is perfectly happy to address 16 CPUs at Aff0.
> >> - ARM cores are designed to be grouped in clusters of at most 4, but
> >> other implementations may have very different layouts.
> >>
> >> If you want to model something matches reality, then you have to follow
> >> what Cortex-A cores do, assuming you are exposing Cortex-A cores. But
> >> absolutely nothing forces you to (after all, we're not exposing the
> >> intricacies of L2 caches, which is the actual reason why we have
> >> clusters of 4 cores).
> >
> > Thanks Marc. I'll take the question of whether or not deviation, in
> > the interest of optimal gicv3 use, is OK to QEMU.
> >
> >>
> >>> Additionally I'm looking at adding support to represent more complex
> >>> topologies in the guest MPIDR (sockets/cores/threads). I see Linux
> >>> currently expects Aff2:socket, Aff1:core, Aff0:thread when threads
> >>> are in use, and Aff1:socket, Aff0:core, when they're not. Assuming
> >>> there are never more than 4 threads to a core makes the first
> >>> expectation fine, but the second one would easily blow the 2 Aff0
> >>> bits alloted, and maybe even a 4 Aff0 bit allotment.
> >>>
> >>> So my current thinking is that always using Aff2:socket, Aff1:cluster,
> >>> Aff0:core (no threads allowed) would be nice for KVM, and allowing up
> >>> to 16 cores to be addressed in Aff0. As it seems there's no standard
> >>> for MPIDR, then that could be the KVM guest "standard".
> >>>
> >>> TCG note: I suppose threads could be allowed there, using
> >>> Aff2:socket, Aff1:core, Aff0:thread (no more than 4 threads)
> >>
> >> I'm not sure why you'd want to map a given topology to a guest (other
> >> than to give the illusion of a particular system). The affinity register
> >> does not define any of this (as you noticed). And what would Aff3 be in
> >> your design? Shelve? Rack? ;-)
> >
> > :-) Currently Aff3 would be unused, as there doesn't seem to be a need
> > for it, and as some processors don't have it, it would only complicate
> > things to use it sometimes.
>
> Careful: on a 64bit CPU, Aff3 is always present.
A57 and A72 don't appear to define it though. They have 63:32 as RES0.
>
> >>
> >> What would the benefit of defining a "socket"?
> >
> > That's a good lead in for my next question. While I don't believe
> > there needs to be any relationship between socket and numa node, I
> > suspect on real machines there is, and quite possibly socket == node.
> > Shannon is adding numa support to QEMU right now. Without special
> > configuration there's no gain other than illusion, but with pinning,
> > etc. the guest numa nodes will map to host nodes, and thus passing
> > that information on to the guest's kernel is useful. Populating a
> > socket/node affinity field seems to me like a needed step. But,
> > question time, is it? Maybe not. Also, the way Linux currently
> > handles non-thread using MPIDRs (Aff1:socket, Aff0:core) throws a
> > wrench at the Aff2:socket, Aff1:"cluster", Aff0:core(max 16) plan.
> > Either the plan or Linux would need to be changed.
>
> What I'm worried of at that stage is that we hardcode a virtual topology
> without the knowledge of the physical one. Let's take an example:
Mark's pointer to cpu-map was the piece I was missing. I didn't want to
hardcode anything, but thought we had to at least agree on the meanings
of affinity levels. I see now that the cpu-map node allows us to describe
the meanings.
>
> I (wish I) have a physical system with 2 sockets, 16 cores per socket, 8
> threads per core. I'm about to run a VM with 16 vcpus. If we're going to
> start pinning things, then we'll have to express that pinning in the
> VM's MPIDRs, and make sure we describe the mapping between the MPIDRs
> and the topology in the firmware tables (DT or ACPI).
>
> What I'm trying to say here is that there is you cannot really enforce a
> partitioning of MPIDR without considering the underlying HW, and
> communicating your expectations to the OS running in the VM.
>
> Do I make any sense?
Sure does, but, just be to be sure; so it's not crazy to want to do
this; we just need to 1) pick a topology that makes sense for the
guest/host (that's the user's/libvirt's job), and 2) make sure we
not only assign MPIDR affinities appropriately, but also describe
them with cpu-map (or the ACPI equivalent).
Is that correct?
Thanks,
drew
_______________________________________________
kvmarm mailing list
kvmarm@xxxxxxxxxxxxxxxxxxxxx
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
