On Tue, Sep 24, 2019 at 05:52:32PM +0200, Borislav Petkov wrote:
> On Tue, Sep 03, 2019 at 05:26:33PM +0300, Jarkko Sakkinen wrote:
> > From: Kai Huang <kai.huang@xxxxxxxxxxxxxxx>
> >
> > Add X86_FEATURE_SGX_LC, which informs whether or not the CPU supports SGX
> > Launch Control.
> >
> > Add MSR_IA32_SGXLEPUBKEYHASH{0, 1, 2, 3}, which when combined contain a
> > SHA256 hash of a 3072-bit RSA public key. SGX backed software packages, so
> > called enclaves, are always signed. All enclaves signed with the public key
> > are unconditionally allowed to initialize. [1]
> >
> > Add FEATURE_CONTROL_SGX_LE_WR bit of the feature control MSR, which informs
> > whether the formentioned MSRs are writable or not. If the bit is off, the
> > public key MSRs are read-only for the OS.
> >
> > If the MSRs are read-only, the platform must provide a launch enclave (LE).
> > LE can create cryptographic tokens for other enclaves that they can pass
> > together with their signature to the ENCLS(EINIT) opcode, which is used
> > to initialize enclaves.
> >
> > Linux is unlikely to support the locked configuration because it takes away
> > the control of the launch decisions from the kernel.
>
> Right, who has control over FEATURE_CONTROL_SGX_LE_WR? Can the
> kernel set it and put another hash in there or there will be locked
> configurations where setting that bit will trap?
Short answer, BIOS controls SGX_LE_WR.
The approach we chose (patch 04, which we were discussing) is to disable
SGX if SGX_LE_WR is not set, i.e. disallow SGX unless the hash MSRs exist
and are fully writable.
WRMSR will #GP if FEATURE_CONTROL is locked (bit 0), e.g. attempting to
set SGX_LE_WR will trap if FEATURE_CONTROL was locked by BIOS. And
conversely, the various enable bits in FEATURE_CONTROL don't take effect
until FEATURE_CONTROL is locked, e.g. the LE hash MSRs aren't writable if
FEATURE_CONTROL is unlocked, regardless of whether SGX_LE_WR is set.
> I don't want to leave anything in the hands of the BIOS controlling
> whether the platform can set its own key because BIOS is known to f*ck
> it up almost every time. And so I'd like for us to be able to fix up
> things without depending on the mood of some OEM vendor's BIOS fixing
> desire.
Sadly, because FEATURE_CONTROL must be locked to fully enable SGX, the
reality is that any BIOS that supports SGX will lock FEATURE_CONTROL.
That's the status quo today as well since VMX (and SMX/TXT) is also
enabled via FEATURE_CONTROL. KVM does have logic to enable VMX and lock
FEATURE_CONTROL if the MSR isn't locked, but AIUI that exists only to work
with old BIOSes.
If we want to support setting and locking FEATURE_CONTROL in the extremely
unlikely scenario that BIOS left it unlocked, the proper change would be
to move the existing KVM FEATURE_CONTROL logic into the early-ish boot
flow and try to set all known bits before locking FEATURE_CONTROL. I
don't have a strong preference either way. We opted not to try and set
FEATURE_CONTROL as we felt that doing so was more likely to cause breakage
than it was to actually "fix" a broken BIOS.
> > [1] Intel SDM: 38.1.4 Intel SGX Launch Control Configuration
One note on Launch Control that isn't covered in the SDM: the LE hash
MSRs can also be written before SGX is activated. SGX activation must
occur before FEATURE_CONTROL is locked, meaning BIOS can set the LE
hash MSRs to a non-intel and then lock FEATURE_CONTROL with SGX_LE_WR=0.
There's a blurb on SGX activation in the kernel docs (patch 23).
> > diff --git a/arch/x86/include/asm/cpufeatures.h b/arch/x86/include/asm/cpufeatures.h
> > index c5582e766121..ca82226e25ec 100644
> > --- a/arch/x86/include/asm/cpufeatures.h
> > +++ b/arch/x86/include/asm/cpufeatures.h
> > @@ -355,6 +355,7 @@
> > #define X86_FEATURE_CLDEMOTE (16*32+25) /* CLDEMOTE instruction */
> > #define X86_FEATURE_MOVDIRI (16*32+27) /* MOVDIRI instruction */
> > #define X86_FEATURE_MOVDIR64B (16*32+28) /* MOVDIR64B instruction */
> > +#define X86_FEATURE_SGX_LC (16*32+30) /* Software Guard Extensions Launch Control */
>
> Amazing. SGX feature bits are spread around at least three CPUID leafs:
>
> 7_EBX, 7_ECX, 12_EAX. Maybe there's a 4th somewhere because hey... :-\
Heh, why stop at 4? 12_EBX, 12_1_ECX and 12_1_EDX are effectively feature
leafs as well, although the kernel can ignore them for the most part.
