Fix "modelstepping" handling in verification #85
Conversation
kds/kds.go
Outdated
| // The string is not fully programmable, given the processor programming reference revisions. | ||
| switch product.Name { | ||
| case pb.SevProduct_SEV_PRODUCT_MILAN: | ||
| return fmt.Sprintf("Milan-B%X", stepping) |
There was a problem hiding this comment.
Just out of curiosity: I saw Milan-B0/Genoa-A0/etc. Is it documented in the spec that Milan revision will always start with B and Genoa revision will always start with A?
There was a problem hiding this comment.
I specifically don't know for sure. I'm using the A%X naming provisionally given the Zen4 silicon revision was A0 last year, but I see B0 could be coming or is already here. I'm unsure how to differentiate between the two and I've asked AMD what the deal is. It might be model == 0 => "A", and model == 1 => "B", but I don't know.
There was a problem hiding this comment.
I checked in with one of our KVM engineers and learned the stepping <-> stepping version translation is not algorithmic, but just a matter of manufacturer decision whenever they change the CPUID stepping value. Updated the logic accordingly with lookup tables given information known about taped out stepping versions.
The "Milan-B0" naming in a product name is not directly derivable from the cpuid(1) values. Instead this B0 naming is from a different naming convention that comes from the development lifecycle that the manufacturer decides. Every stepping number that is visible to consumers is subject to a table lookup. That table can only be built from a piecemeal toilsome effort of following the manufacturers' publications about its product lines. This fix has been verified in hardware tests. Signed-off-by: Dionna Glaze <dionnaglaze@google.com>
The "Milan-B0" naming in a product name is not directly derivable from the cpuid(1) values. Instead this B0 naming is from a different naming convention I cannot yet determine. Despite this, I can at least overfit the verification function for now so that if this model is off, tests will fail loudly and we can fix it.
This fix has been verified in hardware tests.