On 26/03/2024 15:59, mlevitsk@xxxxxxxxxx wrote: > On Mon, 2024-03-25 at 20:15 -0700, Jim Mattson wrote: >>> On Thu, Sep 28, 2023 at 8:05 AM Maxim Levitsky <mlevitsk@xxxxxxxxxx> wrote: >>>>> >>>>> Hi! >>>>> >>>>> This patch series includes several fixes to AVIC I found while working >>>>> on a new version of nested AVIC code. >>>>> >>>>> Also while developing it I realized that a very simple workaround for >>>>> AVIC's errata #1235 exists and included it in this patch series as well. >>>>> >>>>> Best regards, >>>>> Maxim Levitsky >>> >>> Can someone explain why we're still unwilling to enable AVIC by >>> default? Have the performance issues that plagued the Rome >>> implementation been fixed? What is AMD's guidance? >>> > Hi > > This is what I know: > > Zen1: > I never tested it, so I don't know how well AVIC works there and if it has any erratas. > > Zen2: > Has CPU errata in regard to IPI virtualization that makes it unusable in production, > but if AVIC's IPI virtualization (borrowing the Intel term here) is disabled, > then it works just fine and 1:1 equivalent to APICv without IPI. > > I posted patches for this several times, latest version is here, it still applies I think: > https://lkml.iu.edu/hypermail/linux/kernel/2310.0/00790.html > > Zen3: > For some reason AVIC got disabled by AMD in CPUID. It is still there though and force_avic=1 kvm_amd option > can make KVM use it and AFAIK it works just fine. > > It is possible that it got disabled due to Zen2 errata that is fixed on Zen3, > but maybe AMD wasn't sure back then that it will be fixed or it might be due to performance issues with broadcast > IPIs which I think ended up being a software issue and was fixed a long time ago. > > Zen4+ > I haven't tested it much, but AFAIK it should work out of the box. It also got x2avic mode which allows > to use AVIC with VMs that have more that 254 vCPUs. > > IMHO if we merge the workaround I have for IPI virtualization and make IPI virtualization off for Zen2 > (and maybe Zen1 as well), then I don't see why we can't make AVIC be the default on. Additionally, I think right now with avic=1 it fails vcpu creation when creating more a vcpu with an id bigger than what's supported i.e. the MAX_VCPUS we currently advertised in UAPI isn't quite honored. So that's the only wrinkle at least I am aware. Sean had send this one series to inhibit AVIC when such config: https://lore.kernel.org/linux-iommu/20230815213533.548732-1-seanjc@xxxxxxxxxx/#r But I am not sure if it was respinned. Joao
