On 5/5/25 5:58 AM, Andrea Bolognani wrote:
On Tue, Apr 29, 2025 at 05:41:43PM +0200, Jiří Denemark wrote:
On Tue, Apr 29, 2025 at 13:03:12 +0100, Daniel P. Berrangé wrote:
On Tue, Apr 29, 2025 at 01:42:22PM +0200, Peter Krempa wrote:
On Tue, Apr 29, 2025 at 12:36:40 +0100, Daniel P. Berrangé wrote:
On Tue, Apr 15, 2025 at 04:13:08AM -0400, Narayana Murty N wrote:
<model usable='unknown' vendor='IBM'>POWER7</model>
<model usable='unknown' vendor='IBM'>POWER8</model>
<model usable='unknown' vendor='IBM'>POWER9</model>
+ <model usable='unknown' vendor='IBM'>power11</model>
QEMU allows both upper & lowercase for CPU names, and libvirt
has stuck with uppercase historically.
What's the justification for changing this approach for
power11 ?
qemu itself reports the model lowercase now. In v2 of this series I've
pointed that out because defining this caused test data to change which
was suspiciuous.
By 'reports the model lowercase' i presume you're referring to '-cpu help'
and/or 'query-cpu-definitions' ? That doesn't seem to be a new change
either though. I've not gone far back, but qemu 8.2.0 seems to already
be using lowercase.
So I'm still wondering why we're changing our decision about CPU naming
/now/ with power11 ? The commit message needs to explain and justify
this change.
Well, when I replied to the original patch, my memory was telling me
QEMU switched to lower case names and had to add compatibility code to
translate our existing upper case names. And adding a new CPU model
looked like a good excuse to stop doing that (for new models of course)
and just define the model in lower case. Now that you're questioning
this I tried to find evidence and I could not find it :-( So there are
two options... either I'm bad at searching or I my memory confused this
with dropping version suffixes (_vX.Y) in virCPUppc64ConvertLegacy. Or
both options are true at the same time :-)
Maybe Andrea can remember more, I think he was involved in all this ages
ago.
It seems that CPU models have been reported in lower case at least
going all the way back to QEMU 3.0.0 from 2018:
https://gitlab.com/libvirt/libvirt/-/blob/c7e09b7b5fe7f64f12318d150b281343c6d64fd5/tests/qemucapabilitiesdata/caps_3.0.0.ppc64.replies
It would probably make sense for libvirt to align with this decision,
and accept/generate lower case CPU model names. Of course upper case
CPU models would still need to be accepted for backwards
compatibility reasons.
Note that I'm suggesting we adopt this approach consistently for all
CPU models, including existing ones. So power8 should be accepted, as
should be POWER11.
CC'ing Daniel and David, both of which have worked on ppc64 in the
past and might be able to offer additional insights / historical
context.
It has to do with IBM marketing.
POWER is an acronym (Performance Optimization With Enhanced RISC) and all
marketing material demanded that it was mentioned in full caps, e.g. POWER7,
POWER8, POWER9. The reason why libvirt kept using full caps in libvirt, even
after QEMU changed it to lower, was due to IBM marketing demands. IBM wanted
users to use the full caps name in the domain XML, which is/was the official way
customers would interact with the virtualization stack.
In power10 it seems like IBM marketing changed its mind [1] and now the model
name is camel case, so now we're supposed to call them Power10, Power11, etc.
Narayana can surely provide more details than I can, but I believe this is the
reason why we would want a change for case insensitive CPU models in the IBM
PowerPC family in libvirt. There are domains in the wild that are using full
caps names, and we can't break those, but now IBM would tell customers to use
<model fallback='forbid'>power11</model>"
or similar. Ignoring case seems a sensible change to make since we do not want
to be hostage of IBM marketing whims.
But, as Daniel said, this should be a properly justified and documented change.
Thanks,
The other Daniel
[1] https://en.wikipedia.org/wiki/Power10, section "Branding"
