Re: [patch V4 09/65] x86/fpu: Sanitize xstateregs_set()

[Andrei Vagin]

On Thu, Jul 28, 2022 at 4:32 PM Chang S. Bae <chang.seok.bae@xxxxxxxxx> wrote:
>
> On 7/25/2022 2:26 PM, Dave Hansen wrote:
> >
> > Do you happen to have a quick reproducer for this, or at least the
> > contents of the buffer that you are trying to restore?
>
> While not following this report, I think there is a regression along
> with the changes:
>
> As looking into the spec, this state load does not depend on XSTATE_BV:
>
>       RFBM := XCR0 AND EDX:EAX;
>       COMPMASK := XCOMP_BV field from XSAVE header;
>
>       IF COMPMASK[63] = 0
>           THEN
>           ...
>           IF RFBM[1] = 1 OR RFBM[2] = 1
>               THEN load MXCSR from legacy region of XSAVE area;
>           FI;
>           ...
>       ELSE
>       ...
>
> But our upstream code does reference XSTATE_BV instead of RFBM [1,2].
>
> My test case [3] fails with the upstream but works with 5.13, which is
> before the series. Then, this change looks to make it work at least for it:

gVisor test passes with this change too. Chang, are you going to send a patch?

Thanks,
Andrei