Igor Mammedov <imammedo@xxxxxxxxxx> writes: > On Thu, 13 Jan 2022 22:33:38 +0000 > Sean Christopherson <seanjc@xxxxxxxxxx> wrote: > >> On Mon, Jan 03, 2022, Igor Mammedov wrote: >> > On Mon, 03 Jan 2022 09:04:29 +0100 >> > Vitaly Kuznetsov <vkuznets@xxxxxxxxxx> wrote: >> > >> > > Paolo Bonzini <pbonzini@xxxxxxxxxx> writes: >> > > >> > > > On 12/27/21 18:32, Igor Mammedov wrote: >> > > >>> Tweaked and queued nevertheless, thanks. >> > > >> it seems this patch breaks VCPU hotplug, in scenario: >> > > >> >> > > >> 1. hotunplug existing VCPU (QEMU stores VCPU file descriptor in parked cpus list) >> > > >> 2. hotplug it again (unsuspecting QEMU reuses stored file descriptor when recreating VCPU) >> > > >> >> > > >> RHBZ:https://bugzilla.redhat.com/show_bug.cgi?id=2028337#c11 >> > > >> >> > > > >> > > > The fix here would be (in QEMU) to not call KVM_SET_CPUID2 again. >> > > > However, we need to work around it in KVM, and allow KVM_SET_CPUID2 if >> > > > the data passed to the ioctl is the same that was set before. >> > > >> > > Are we sure the data is going to be *exactly* the same? In particular, >> > > when using vCPU fds from the parked list, do we keep the same >> > > APIC/x2APIC id when hotplugging? Or can we actually hotplug with a >> > > different id? >> > >> > If I recall it right, it can be a different ID easily. >> >> No, it cannot. KVM doesn't provide a way for userspace to change the APIC ID of >> a vCPU after the vCPU is created. x2APIC flat out disallows changing the APIC ID, >> and unless there's magic I'm missing, apic_mmio_write() => kvm_lapic_reg_write() >> is not reachable from userspace. >> >> The only way for userspace to set the APIC ID is to change vcpu->vcpu_id, and that >> can only be done at KVM_VCPU_CREATE. >> >> So, reusing a parked vCPU for hotplug must reuse the same APIC ID. QEMU handles >> this by stashing the vcpu_id, a.k.a. APIC ID, when parking a vCPU, and reuses a >> parked vCPU if and only if it has the same APIC ID. And because QEMU derives the >> APIC ID from topology, that means all the topology CPUID leafs must remain the >> same, otherwise the guest is hosed because it will send IPIs to the wrong vCPUs. > > Indeed, I was wrong. > I just checked all cpu unplug history in qemu. It was introduced in qemu-2.7 > and from the very beginning it did stash vcpu_id, > so there is no old QEMU that would re-plug VCPU with different apic_id. > Though tells us nothing about what other userspace implementations might do. > The genie is out of the bottle already, 5.16 is released with the change (which was promissed for some time, KVM was complaining with pr_warn_ratelimited()). I'd be brave and say that if QEMU doesn't need it then nobody else does (out of curiosity, are there KVM VMMs besides QEMU which support CPU hotplug out there?). > However, a problem of failing KVM_SET_CPUID2 during VCPU re-plug > is still there and re-plug will fail if KVM rejects repeated KVM_SET_CPUID2 > even if ioctl called with exactly the same CPUID leafs as the 1st call. > Assuming APIC id change doesn not need to be supported, I can send v2 here with an empty allowlist. -- Vitaly
