On Tue, Nov 26, 2024, Binbin Wu wrote: > On 11/26/2024 11:52 AM, Huang, Kai wrote: > > On Tue, 2024-11-26 at 09:44 +0800, Binbin Wu wrote: > > > > > > On 11/26/2024 6:51 AM, Sean Christopherson wrote: > > > > > > [...] > > > > When an NMI happens in non-root, the NMI is acknowledged by the CPU prior to > > > > performing VM-Exit. In regular VMX, NMIs are blocked after such VM-Exits. With > > > > TDX, that blocking happens for SEAM root, but the SEAMRET back to VMX root will > > > > load interruptibility from the SEAMCALL VMCS, and I don't see any code in the > > > > TDX-Module that propagates that blocking to SEAMCALL VMCS. > > > I see, thanks for the explanation! > > > > > > > Hmm, actually, this means that TDX has a causality inversion, which may become > > > > visible with FRED's NMI source reporting. E.g. NMI X arrives in SEAM non-root > > > > and triggers a VM-Exit. NMI X+1 becomes pending while SEAM root is active. > > > > TDX-Module SEAMRETs to VMX root, NMIs are unblocked, and so NMI X+1 is delivered > > > > and handled before NMI X. > > > This example can also cause an issue without FRED. > > > 1. NMI X arrives in SEAM non-root and triggers a VM-Exit. > > > 2. NMI X+1 becomes pending while SEAM root is active. > > > 3. TDX-Module SEAMRETs to VMX root, NMIs are unblocked. > > > 4. NMI X+1 is delivered and handled before NMI X. > > > (NMI handler could handle all NMI source events, including the source > > > triggered NMI X) > > > 5. KVM calls exc_nmi() to handle the VM Exit caused by NMI X > > > In step 5, because the source event caused NMI X has been handled, and NMI X > > > will not be detected as a second half of back-to-back NMIs, according to > > > Linux NMI handler, it will be considered as an unknown NMI. > > I don't think KVM should call exc_nmi() anymore if NMI is unblocked upon > > SEAMRET. > > IIUC, KVM has to, because the NMI triggered the VM-Exit can't trigger the > NMI handler to be invoked automatically even if NMI is unblocked upon SEAMRET. Yep. The NMI is consumed by the VM-Exit, for all intents and purposes. KVM must manually invoke the NMI handler. Which is how the ordering gets messed up: NMI X+1 arrives before KVM has a chance to manually invoke the handler for NMI X.
