Re: [patch 08/41] x86/fpu: Restrict fpstate sanitizing to legacy components

[Thomas Gleixner]

On Fri, Jun 11 2021 at 12:18, Andy Lutomirski wrote:
> On 6/11/21 12:03 PM, Andy Lutomirski wrote:
>> On 6/11/21 9:15 AM, Thomas Gleixner wrote:
>> Does this result in the mxcsr_mask and mxcsr fields being correct?
>> There is a silly number of special cases there.

Let me stare at that.

> The SDM says, in a footnote in XRSTOR:
>
> There is an exception if RFBM[1] = 0 and RFBM[2] = 1. In this case, the
> standard form of XRSTOR will load MXCSR from memory, even though MXCSR
> is part of state component 1 — SSE. The compacted form of XRSTOR does
> not make this exception.
>
> which makes me think that this code has a bug.  Also, the code is
> manifest nonsense for a different reason:
>
> if (!FP)
>   memset(the whole damn thing, 0);
>
> if (!SSE)
>   memset(just part of it, 0);
>
> which means that, if the FP bit is clear but the SSE bit is set, this
> thing will clobber the SSE state.
>
> This code is garbage.  So is the architecture that gave rise to this code.

No argument about that :)