Re: [PATCH] x86/pkeys: Manually set X86_FEATURE_OSPKE to preserve existing changes

[Jacob Keller]

On 2/26/2020 3:16 PM, Sean Christopherson wrote:
> Explicitly set X86_FEATURE_OSPKE via set_cpu_cap() instead of calling
> get_cpu_cap() to pull the feature bit from CPUID after enabling CR4.PKE.
> Invoking get_cpu_cap() effectively wipes out any {set,clear}_cpu_cap()
> changes that were made between this_cpu->c_init() and setup_pku(), as
> all non-synthetic feature words are reinitialized from the CPU's CPUID
> values.
> 
> Blasting away capability updates manifests most visibility when running
> on a VMX capable CPU, but with VMX disabled by BIOS.  To indicate that
> VMX is disabled, init_ia32_feat_ctl() clears X86_FEATURE_VMX, using
> clear_cpu_cap() instead of setup_clear_cpu_cap() so that KVM can report
> which CPU is misconfigured (KVM needs to probe every CPU anyways).
> Restoring X86_FEATURE_VMX from CPUID causes KVM to think VMX is enabled,
> ultimately leading to an unexpected #GP when KVM attempts to do VMXON.
> 
> Arguably, init_ia32_feat_ctl() should use setup_clear_cpu_cap() and let
> KVM figure out a different way to report the misconfigured CPU, but VMX
> is not the only feature bit that is affected, i.e. there is precedent
> that tweaking feature bits via {set,clear}_cpu_cap() after ->c_init()
> is expected to work.  Most notably, x86_init_rdrand()'s clearing of
> X86_FEATURE_RDRAND when RDRAND malfunctions is also overwritten.
> 
> Fixes: 0697694564c8 ("x86/mm/pkeys: Actually enable Memory Protection Keys in the CPU")
> Cc: stable@xxxxxxxxxxxxxxx
> Cc: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>
> Reported-by: Jacob Keller <jacob.e.keller@xxxxxxxxx>
> Signed-off-by: Sean Christopherson <sean.j.christopherson@xxxxxxxxx>
> ---
> 

I tested this and it resolves my report! Thanks for a timely fix.

I agree with the analysis. Perhaps it would make sense in the long term
to find a solution where get_cpu_cap can remember what was cleared for
each CPU and restore those? It already does this using the global
variables...

Thanks,
Jake