Re: x86: AMD Zen2 ymm registers rolling back

[Andrew Cooper]

On 22/02/2023 9:26 pm, Tavis Ormandy wrote:
> On Wed, Feb 22, 2023 at 11:09:51AM +0100, Borislav Petkov wrote:
>> On Wed, Feb 22, 2023 at 09:38:09AM +0000, Andrew Cooper wrote:
>>> This sounds suspiciously like an errata which was fixed with a ucode
>>> update last year.
>> Yes, it looks like it.
>>
>> Alternatively, you can try booting with "clearcpuid=xsaves" - that
>> should take care of your observation too but yeah, you should rather
>> update your microcode.
>>
> Thanks - confirmed, it *doesn't* repro with 0x8301055, but does repro
> with 0x830104d.

For completeness, this is erratum 1386, details of which can be found in
https://www.amd.com/system/files/TechDocs/56323-PUB_1.00.pdf amongst others.

That said, looking at the list of preconditions to tickle this erratum,
I don't see anything in your code which would cause a change in MXCSR. 
Which most likely means there's a path in the kernel modifying MXCSR and
that doesn't sound like a thing that ought to be happening by default.

> Annoyingly, I thought I was using the most recent microcode, but it seems
> like there is some bug and debian wasn't applying it at boot.
>
> That seems like a scary errata :-/

Honestly, this is tame compared to some of the things which get fixed in
ucode...

~Andrew