On Tue, Jan 16, 2018 at 09:01:49PM +0100, Borislav Petkov wrote:
> On Tue, Jan 16, 2018 at 05:24:27PM +0000, Luck, Tony wrote:
> > > I'll look for someone who can confirm the 2.5MB/core detail.
> >
> > Ok ... re-read the erratum. The 2.5MB/core is clear. The E5+E7 is clear.
> >
> > No mention of the platform ID, but Jia is dropping that part.
> >
> > Boris ... what specific questions remain?
>
> This magic:
>
> llc_size_per_core(c) > 2621440
>
> as a reliable detection characteristic whether the patch is good to
> apply late. There must be a more reliable way to detect that.
>
> Also, the testing order is:
>
> llc_size_per_core(c) > 2621440 &&
> c->microcode < 0x0b000021) {
>
> so if the LLC size per core check fails, the microcode revision being <
> 0x0b000021 doesn't matter. I.e., on machines with LLC-per-core < 2.5M,
> we can update even with revisions < 0x0b000021.
>
> Is that ordering correct?
I think so. The erratum (see below) says the problem only occurs
on the large-cache SKUs. So we only need to avoid the update if
we are on a big cache SKU that is also running old microcode.
> Also, this heuristic is not documented in the public doc AFAICT - I'm
> guessing that'll change soon...?
Here's what I see in the public doc. for BDF90:
Problem: An uncorrectable error (IA32_MC3_STATUS.MCACOD=0400 and
IA32_MC3_STATUS.MSCOD=0080) may be logged for processors that have more
than 2.5MB last-level-cache per core on attempting to load a microcode
update or execute an authenticated code module. This issue does not
occur with microcode updates with a signature of 0x0b000021 and greater.
-Tony
