Re: [PATCH] x86: vmx: Allow direct access to MSR_IA32_SPEC_CTRL

[Paolo Bonzini]

On 29/01/2018 11:37, David Woodhouse wrote:
> On Mon, 2018-01-29 at 10:43 +0100, KarimAllah Ahmed wrote:
>> On 01/29/2018 09:46 AM, David Woodhouse wrote:
>>> Reading the code and comparing with the SDM, I can't see where we're
>>> ever setting VM_EXIT_MSR_STORE_{ADDR,COUNT} except in the nested
>>> case...
>> Hmmm ... you are probably right! I think all users of this interface
>> always trap + update save area and never passthrough the MSR. That is
>> why only LOAD is needed *so far*.
>>
>> Okay, let me sort this out in v3 then.
> 
> I'm starting to think a variant of Ashok's patch might actually be the
> simpler approach, and not "premature optimisation". Especially if we
> need to support the !cpu_has_vmx_msr_bitmaps() case?

That case is awfully slow anyway, it doesn't matter, but the
direct-access flag would simply be always 0 if you have no MSR bitmaps.

> Start with vmx->spec_ctrl set to zero. When first touched, make it
> passthrough (but not atomically switched) and set a flag (e.g.
> "spec_ctrl_live") which triggers the 'restore_branch_speculation' and
> 'save_and_restrict_branch_speculation' behaviours. Except don't use
> those macros. Those can look something like
> 
> Â/* If this vCPU has touched SPEC_CTRL then restore its value if needed */
> Âif (vmx->spec_ctrl_live && vmx->spec_ctrl)
> Â Â Âwrmsrl(MSR_IA32_SPEC_CTRL, vmx->spec_ctrl);
> Â/* vmentry is serialising on affected CPUs, so the conditional branch is safe */
> 
> 
> ... and, respectively, ...
> 
>  /* If this vCPU has touched SPEC_CTRL then save its value and ensure we have zero */
>  if (vmx->spec_ctrl_live) {
>      rdmsrl(MSR_IA32_SPEC_CTRL, vmx->spec_ctrl);
>      if (vmx->spec_ctrl)
>          wrmsrl(MSR_IA32_SPEC_CTRL, 0);
>  }
> 
> Perhaps we can ditch the separate 'spec_ctrl_live' flag and check the
> pass-through MSR bitmap directly, in the case that it exists?Â

Probably a cache miss, or even a TLB miss if you're unlucky, so the
separate flag is okay.

Paolo