Collin, I apologize for not getting to you earlier. On Wed, Sep 16, 2020 at 12:11:08 -0400, Collin Walling wrote: > On 9/16/20 3:03 AM, Michal Privoznik wrote: > > On 9/15/20 10:25 PM, Collin Walling wrote: > >> One more ping in attempt to get this in the right direction. Otherwise > >> I'll post my next idea and we can go from there :) > > > > I agree with Peter that while the idea might look correct it's too deep. > > > >> > >> Thinking about this issue, should a host-passthough CPU definition be > >> permitted for the baseline & comparison commands? The model represented > >> under this mode is not considered migration safe and it may make sense > >> to simply fail early since these commands aim to construct/determine a > >> migratable CPU model, respectively. > > > > Honestly, I don't know much about this CPU models area, but is that true > > even for two identical hosts? Say I have two desktops next to each > > other, with the same CPU and I want to migrate. I could use host model, > > couldn't I? > > > > "Host-model" is an alias for a CPU model that closely represents the > capabilities of the host machine (on s390, because this model is defined > by the hypervisor, it can also be called the "hypervisor CPU model" -- > not an important detail). > > However, a guest running with the host-passthrough mode is not > considered migration safe as that guest may covertly run with > features/capabilities that are not directly exposed to the hypervisor. Right, but migration may still be possible and working fine if both host are identical. > From what I understand regarding the hypervisor-cpu-compare and > hypervisor-cpu-baseline commands is that they aim to assist with > determining the migratability of guests based on their CPU model and > feature set (usually along with a host CPU in the equation as well). Baseline with a host-passthrough CPU is not indeed very useful, but compare could still be used and its usage is not limited to migration. For example, you can use it to check whether a domain with a guest CPU configuration can be started on a specific host before you actually try to start it. And reporting host-passthrough as incompatible would be wrong. Anyway, thanks for your patch, it was mostly correct, it just needed to be done a bit higher in the call graph. Incidentally, Tim Wiederhake [1] took this original patch and moved the change to the right place. The authorship is still yours, so if you want to append you signed-off-by tag there, I'll wait a bit before pushing Tim's patch. Jirka [1] https://www.redhat.com/archives/libvir-list/2020-September/msg01177.html
