Re: [PATCH docs/manager/qemu-server v4 00/17] Add API and UI for custom CPU models

[Fiona Ebner <f.ebner@proxmox.com>]

Am 12.05.26 um 10:09 AM schrieb Arthur Bied-Charreton:
> On Mon, May 11, 2026 at 12:42:02PM +0200, Fiona Ebner wrote:
>> I found a pre-existing bug in query_supported_cpu_flags() that we should
>> fix. Namely the following replacement is wrong in general:
>>
>>>             foreach my $prop (keys %$props) {
>>>                 next if $props->{$prop} ne '1';

Oh, and we should compare against JSON::true here, since the CPU model
expansion also contains numeric, non-flag properties. While none of them
are 1 on my machine, not sure if that's the case everywhere and it's
rather for future-proofing.

>>>                 # QEMU returns some flags multiple times, with '_', '.' or '-'
>>>                 # (e.g. lahf_lm and lahf-lm; sse4.2, sse4-2 and sse4_2; ...).
>>>                 # We only keep those with underscores, to match /proc/cpuinfo
>>>                 $prop =~ s/\.|-/_/g;
>>>                 $flags->{$prop} = 1;
>>>             }
>>
>> In particular, flags like 'virt-ssbd' or 'amd-ssbd' or 'amd-no-ssb' are
>> not present multiple times, but only in their exact form. Because of the
>> replacement, we only consider e.g. 'virt_ssbd' and then mistakenly
>> detect the flags as not supported on any node.
>>
>> Let's just make an explicit map of what is actually duplicate and not do
>> such general replacement here.
> 
> After looking a little more into it I think we should also apply the
> same logic to the recognized CPU flags, or normalize the flags in 
> query_supported_cpu_flags() to match the `qemu-system-$arch -cpu help`
> output rather than /proc/cpuinfo, since that is the source we use?
> 
> For example, /proc/cpuinfo writes the sse4.2 flag as sse4_2, while
> `qemu-system-$arch -cpu help` writes it as sse4.2, so normalizing it to
> sse4_2 in query_supported_cpu_flags() means it does not match anything
> in the output of query_understood_cpu_flags(). QEMU accepts both
> versions.
> 
> Is there a reason why we need to match /proc/cpuinfo specifically that
> I am overlooking, or is it just about having a consistent output for CPU
> flags in general? 

Yes, I too suppose it was done with the expectation to somehow
incorporate information from /proc/cpuinfo, but we don't need that.
Let's align the flags with their proper, unaliased names in QEMU. See
the x86_cpu_initfn() in target/i386/cpu.c for the list of aliases. We
can just skip the aliases when iterating the result from the queried CPU
model expansion here. Of course, a code comment that QEMU uses a
different form from /proc/cpuinfo for some flags would be good.