On Tue, May 19, 2020 at 09:43:23AM +0200, Paolo Bonzini wrote: > On 19/05/20 08:02, Sean Christopherson wrote: > > On Mon, May 18, 2020 at 07:37:08PM +0200, Paolo Bonzini wrote: > >> On 18/05/20 18:07, Sean Christopherson wrote: > >>> On Fri, May 15, 2020 at 12:19:19PM -0400, Paolo Bonzini wrote: > >>>> Instructions starting with 0f18 up to 0f1f are reserved nops, except those > >>>> that were assigned to MPX. > >>> Well, they're probably reserved NOPs again :-D. > >> > >> So are you suggesting adding them back to the list as well? > > > > Doesn't KVM still support MPX? > > > >>>> These include the endbr markers used by CET. > >>> And RDSPP. Wouldn't it make sense to treat RDSPP as a #UD even though it's > >>> a NOP if CET is disabled? The logic being that a sane guest will execute > >>> RDSSP iff CET is enabled, and in that case it'd be better to inject a #UD > >>> than to silently break the guest. > >> > >> We cannot assume that guests will bother checking CPUID before invoking > >> RDSPP. This is especially true userspace, which needs to check if CET > >> is enable for itself and can only use RDSPP to do so. > > > > Ugh, yeah, just read through the CET enabling thread that showed code snippets > > that do exactly this. > > > > I assume it would be best to make SHSTK dependent on unrestricted guest? > > Emulating RDSPP by reading vmcs.GUEST_SSP seems pointless as it will become > > statle apart on the first emulated CALL/RET. > > Running arbitrary code under the emulator is problematic anyway with > CET, since you won't be checking ENDBR markers or updating the state > machine. So perhaps in addition to what you say we should have a mode > where, unless unrestricted guest is disabled, the emulator only accepts > I/O, MOV and ALU instructions. Doh, I forgot all about those pesky ENDBR markers. I think a slimmed down emulator makes sense? Tangentially related, isn't the whole fastop thing doomed once CET kernel support lands?
