On Fri, 15 Mar 2019, Andrew Cooper wrote:
> On 14/03/2019 20:25, Thomas Gleixner wrote:
> > On Thu, 14 Mar 2019, Raj, Ashok wrote:
> >> On Thu, Mar 14, 2019 at 12:39:46PM +0000, Andrew Cooper wrote:
> >>> On late load failure, we should dump enough information to work out
> >>> exactly what went on, to determine how best to proceed, but the server
> >>> is effectively lost to us.  On late load success, the proposed new
> >>> "version" replaces the current "version".
> >>>
> >>> And again - I reiterate the point that I think it is fine to have a
> >>> simplifying assumption that we don't have mixed stepping systems to
> >>> start with, presuming this is generally in line with Intel's support
> >>> statement.  If in practice we find mixed stepping systems which are
> >>> supported by an OEM/Intel, we can see about extending the logic.
> >> Checking with Asit he says it is in fact permitted to have 1 step behind
> >> even on a multi-socket system. One could be N and other N-1 should be 
> >> supported.
> > That turns into a total disaster if N has an issue fixed ant N-1 requires
> > microcode + software workaround.
> >
> > So if N is on the boot socket, then we fail to enable the workaround
> > because CPU0 has the 'Issue fixed' bit set.
> >
> > If N-1 is on the boot socket, then we go to do the workaround nevertheless
> > on N and that might dependend on the issue just be some pointless exercise
> > or even try to access some MSR which is not available.
> >
> > *Shudder*
> 
> Intel: Are you saying that Skylake (06-55-04) is supported in
> combination with Cascade Lake B0 (06-55-05) and/or Cascade Lake B1
> (06-55-06) ?
> 
> The most insidious problem is TSX_FORCE_ABORT between the two Cascade
> Lakes.  There really will be an asymmetric existence of an MSR required
> for use in one part of the system, and unavailable in the other part of
> the system.
> 
> To a certain degree, what is technically supported by Intel is also
> tempered by what the major OS/VMM vendors are willing to boot on, as
> that is ultimately what the customer is paying for.  When the steppings
> differed only by the errata fixed, and the silicon was otherwise
> identical from software's point of view, supporting a range of adjacent
> steppings seems entirely reasonable.

Sure, if the software does not have to worry about the differences then
supporting that is a no-brainer.

> In this case you've got 3 adjacent steppings, *all* of which offer
> different architecturally defined features, and will involve software
> changes to allow mixed systems to function in a safe way.

Let's not go there. We've seen the mess which other architectures created
with big/little CPUs which expose different feature sets.

Thanks,

	tglx