On Tue, May 16 2023, Marc Zyngier <maz@xxxxxxxxxx> wrote: > On Tue, 16 May 2023 15:19:00 +0100, > Cornelia Huck <cohuck@xxxxxxxxxx> wrote: >> >> On Tue, May 16 2023, Marc Zyngier <maz@xxxxxxxxxx> wrote: >> >> > On Tue, 16 May 2023 12:55:14 +0100, >> > Cornelia Huck <cohuck@xxxxxxxxxx> wrote: >> >> >> >> Do you have more concrete ideas for QEMU CPU models already? Asking >> >> because I wanted to talk about this at KVM Forum, so collecting what >> >> others would like to do seems like a good idea :) >> > >> > I'm not being asked, but I'll share my thoughts anyway! ;-) >> > >> > I don't think CPU models are necessarily the most important thing. >> > Specially when you look at the diversity of the ecosystem (and even >> > the same CPU can be configured in different ways at integration >> > time). Case in point, Neoverse N1 which can have its I/D caches made >> > coherent or not. And the guest really wants to know which one it is >> > (you can only lie in one direction). >> > >> > But being able to control the feature set exposed to the guest from >> > userspace is a huge benefit in terms of migration. >> >> Certainly; the important part is that we can keep the guest ABI >> stable... which parts match to a "CPU model" in the way other >> architectures use it is an interesting question. It almost certainly >> will look different from e.g. s390, where we only have to deal with a >> single manufacturer. >> >> I'm wondering whether we'll end up building frankenmonster CPUs. > > We already do. KVM hides a bunch of things we don't want the guest to > see, either because we don't support the feature, or that we want to > present it with a different shape (cache topology, for example), and > these combination don't really exist in any physical implementation. > > Which is why I don't really buy the "CPU model" concept as defined by > x86 and s390. We already are in a vastly different place. Yes, I agree that the "named cpu models" approach probably won't work on Arm (especially if you add other accelerators into the mix -- cpu 'foo' with tcg is unlikely to be 100% identical to cpu 'foo' with KVM.) OTOH, "these two cpus are not that different from each other, so we support migration between them with a least common denominator feature/behaviour set" seems more reasonable. > > The way I see it, you get a bunch of architectural features that can > be enabled/disabled depending on the underlying HW, hypervisor's > capabilities and userspace input. On top of that, there is a layer of > paint that tells you what is the overall implementation you could be > running on (that's what MIDR+REVIDR+AIDR tell you) so that you can > apply some unspeakable, uarch-specific hacks that keep the machine > going (got to love these CPU errata). > >> Another interesting aspect is how KVM ends up influencing what the guest >> sees on the CPU level, as in the case where we migrate across matching >> CPUs, but with a different software level. I think we want userspace to >> control that to some extent, but I'm not sure if this fully matches the >> CPU model context. > > I'm not sure I get the "different software level" part. Do you mean > VMM revisions? Yes. Basically, two (for migration purposes) identical machines with different kernel/QEMU versions, but using the same QEMU compat machine. Migrate from old to new, get more regs: works. Migrate from new to old, get less regs: boom. Expectation would be for this to work, and handling it from machine compat code seems very awkward.
