On Wed, Sep 25, 2019 at 08:31:36PM +0200, Borislav Petkov wrote: > On Wed, Sep 25, 2019 at 10:18:24AM -0700, Sean Christopherson wrote: > > Realistically, there will likely be a non-trivial number of systems with > > SGX_LE_WR=0 but SGX enabled. > > Well no. We won't support those. I remember very vividly at Tech Days a > couple of years ago where we said we won't support locked down systems. Yep, that's our intent as well. > > It's inevitable that some systems will lock down the LE hash MSRs, either > > intentionally or due to lack of support for SGX_LE_WR. The latter is > > probably going to be more common than OEMs intentionally locking the MSRs, > > because some Intel reference BIOSes simply don't support SGX_LE_WR, e.g. I > > have a Coffee Lake SDP that has hardware support for SGX_LC, but the BIOS > > doesn't provide any way to set SGX_LE_WR or leave FEATURE_CONTROL unlocked. > > We won't support those too. Nothing changes since a couple of years ago. > We won't support locked down systems and unfinished BIOS systems. Yep. > ... reading your other mail about KVM... > > I guess KVM could be an exception here if people wanna run different > OSes in the guest. IMHO. > > For that, though, we should still clear all SGX feature bits in the > host, I'd say, and let the kvm module rediscover everything itself > through CPUID directly and not using *cpu_has* > > Why, you ask? Because otherwise users will start asking why do they have > "sgx" in /proc/cpuinfo but they can't run their own enclaves. That makes sense. I was thinking it'd be helpful to leave the bits set, e.g. for users to differentiate between "I don't have SGX" and "I can't use SGX because SGX_LC is disabled". But I'm probably being slightly optomistic... > But maybe someone has a better idea. > > In any case, I think it would be bad idea to show only a subset of > features in /proc/cpuinfo of a locked-down system and have to explain it > to users why they can't do own enclaves. > > But again, someone might have a better idea. I'm 99% certain this won't even require a change to the proposed KVM patches, as KVM mostly pulls SGX support directly from CPUID. The only thing it checks via cpu_has() is SGX_LC to query whether or not the MSRs are fully writable. Keeping the SGX feature bits set was more about reflecting hardware capabilities than it was a functional requirement.
