On 12/1/2020 2:26 PM, Dave Hansen wrote:
On 11/30/20 3:16 PM, Yu, Yu-cheng wrote:
Do we have any other spots in the kernel where we care about:
boot_cpu_has(X86_FEATURE_SHSTK) ||
boot_cpu_has(X86_FEATURE_IBT)
? If so, we could also address this by declaring a software-defined
X86_FEATURE_CET and then setting it if SHSTK||IBT is supported, then we
just put that one feature in xsave_cpuid_features[].
These features have different CPUIDs but are complementary parts. I
don't know if someday there will be shadow-stack-only CPUs, but an
IBT-only CPU is weird. What if the kernel checks that the CPU has both
features and presents only one feature flag (X86_FEATURE_CET), no
X86_FEATURE_SHSTK or X86_FEATURE_IBT?
Logically, that's probably fine. But, X86_FEATURE_IBT/SHSTK are in a
non-scattered leaf, so we'll kinda define them whether we like it or
not. We'd have to go out of our way to *not* define them.
After more thoughts, I think it is better to just add X86_FEATURE_CET
and not more. We cannot predict what is going to happen later.
So, like what you suggested, X86_FEATURE_CET means (X86_FEATURE_SHSTK |
X86_FEATURE_IBT).
Thanks,
Yu-cheng