On Thu, Dec 12, 2024, Doug Covelli wrote: > On Thu, Nov 14, 2024 at 10:45 AM Doug Covelli <doug.covelli@xxxxxxxxxxxx> wrote: > > > For LINT1, it should be less performance critical; if it's possible > > > to just go through all vCPUs, and do KVM_GET_LAPIC to check who you > > > should send a KVM_NMI to, then I'd do that. I'd also accept a patch > > > that adds a VM-wide KVM_NMI ioctl that does the same in the hypervisor > > > if it's useful for you. > > > > Thanks for the patch - I'll get it a try but it might not be right away. > > > > > And since I've been proven wrong already, what do you need INIT/SIPI for? > > > > I don't think this one is as critical. I believe the reason it was > > added was so that we can synchronize startup of the APs with execution > > of the BSP for guests that do not do a good job of that (Windows). > > > > Doug > > We were able to get the in-kernel APIC working with our code using the split > IRQ chip option with our virtual EFI FW even w/o the traps for SVR and LVT0 > writes. Performance of Windows VMs is greatly improved as expected. > Unfortunately our ancient legacy BIOS will not work with > 1 VCPU due to lack > of support for IPIs with an archaic delivery mode of remote read which it uses > to discover APs by attempting to read their APIC ID register. MSFT WHP supports > this functionality via an option, WHvPartitionPropertyCodeApicRemoteReadSupport. > > Changing our legacy BIOS is not an option so in order to support Windows VMs > with the legacy BIOS with decent performance we would either need to add support > for remote reads of the APIC ID register to KVM or support CR8 accesses w/o > exiting w/o the in-kernel APIC in order. Do you have a preference? I didn't quite follow the CR8 access thing. If the choice is between emulating Remote Read IPIs and using a userspace local APIC, then I vote with both hands for emulating Remote Reads, especially if we can do a half-assed version that provides only what your crazy BIOS needs :-) The biggest wrinkle I can think of is that KVM uses the Remote Read IPI encoding for a paravirt vCPU kick feature, but I doubt that's used by Windows guests and so can be sacrificed on the Altar of Ancient BIOS.
