On Tue, Mar 24, 2026, Tycho Andersen wrote: > From: "Tycho Andersen (AMD)" <[email protected]> > > In some configurations not all VM types are supported by the firmware. > Reflect this information in the supported_vm_types that KVM exports. > > Link: https://lore.kernel.org/all/[email protected]/ > Suggested-by: Sean Christopherson <[email protected]> > Signed-off-by: Tycho Andersen (AMD) <[email protected]> > --- > arch/x86/kvm/svm/sev.c | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/arch/x86/kvm/svm/sev.c b/arch/x86/kvm/svm/sev.c > index 37490803f2e8..0fe9515db1e7 100644 > --- a/arch/x86/kvm/svm/sev.c > +++ b/arch/x86/kvm/svm/sev.c > @@ -2976,7 +2976,8 @@ void __init sev_set_cpu_caps(void) > supported_vm_types |= BIT(KVM_X86_SNP_VM); > } > > - kvm_caps.supported_vm_types |= supported_vm_types; > + kvm_caps.supported_vm_types |= (supported_vm_types & > + sev_firmware_supported_vm_types());
This is slightly flawed, in that sev_hardware_setup() still reports SEV-ES as fully enabled, whereas the other cases (ASID exhaustation) clear the VM type *and* report the feature as "unusable". Addressing that is actually a great opportunity to dedup some code in the previous path. I.e. if we first relocate the supported_vm_types updates to sev_hardware_setup(), then there's no need to copy+paste the ASID checks. And then restricting VM types based on firmware support Just Works. E.g. to yield: [ 1813.863571] ccp 0000:24:00.1: SEV-SNP API:1.58 build:1 [ 1813.876790] kvm_amd: SEV enabled (ASIDs 254 - 509) [ 1813.881595] kvm_amd: SEV-ES unusable (ASIDs 1 - 253) [ 1813.886574] kvm_amd: SEV-SNP enabled (ASIDs 1 - 253) [ 1813.891549] kvm_amd: Virtual VMLOAD VMSAVE supported [ 1813.896522] kvm_amd: Virtual GIF supported I'll send a v3, I've got everything coded up (I wanted to make sure my idea actually worked before suggesting it :-) ).
