On Wed, Dec 13, 2023 at 3:10 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote: > > On Thu, Dec 14, 2023, Maxim Levitsky wrote: > > On Mon, 2023-11-13 at 20:35 -0800, isaku.yamahata@xxxxxxxxx wrote: > > > From: Isaku Yamahata <isaku.yamahata@xxxxxxxxx> > > > > > > TDX virtualizes the advertised APIC bus frequency to be 25MHz. > > > > Can you explain a bit better why TDX needs this? I am not familiar > > with TDX well enough yet to fully understand. > > TDX (the module/architecture) hardcodes the core crystal frequency to 25Mhz, > whereas KVM hardcodes the APIC bus frequency to 1Ghz. And TDX (again, the module) > *unconditionally* enumerates CPUID 0x15 to TDX guests, i.e. _tells_ the guest that > the frequency is 25MHz regardless of what the VMM/hypervisor actually emulates. > And so the guest skips calibrating the APIC timer, which results in the guest > scheduling timer interrupts waaaaaaay too frequently, i.e. the guest ends up > gettings interrupts at 40x the rate it wants. > > Upstream KVM's non-TDX behavior is fine, because KVM doesn't advertise support > for CPUID 0x15, i.e. doesn't announce to host userspace that it's safe to expose > CPUID 0x15 to the guest. Because TDX makes exposing CPUID 0x15 mandatory, KVM > needs to be taught to correctly emulate the guest's APIC bus frequency, a.k.a. > the TDX guest core crystal frequency of 25Mhz. Aside from placating a broken guest infrastructure that ignores a 17-year old contract between KVM and its guests, what are the advantages to supporting a range of APIC bus frequencies? > I halfheartedly floated the idea of "fixing" the TDX module/architecture to either > use 1Ghz as the base frequency (off list), but it definitely isn't a hill worth > dying on since the KVM changes are relatively simple. Not making the KVM changes is even simpler. :)
