[ovirt-devel] rethinking fake-kvm and faqemu

[Michal Skrivanek]

> On 26 Oct 2015, at 09:41, Dan Kenigsberg <danken at redhat.com> wrote:
> 
> On Mon, Oct 05, 2015 at 02:53:00PM +0200, Martin Polednik wrote:
>> On 05/10/15 11:31 +0200, Michal Skrivanek wrote:
>>> 
>>> On Oct 3, 2015, at 20:48 , Martin Polednik <mpolednik at redhat.com> wrote:
>>> 
>>>> Hello everyone,
>>>> 
>>>> I've been reworking the fake_kvm and faqemu VDSM hook to make them
>>>> somewhat more usable and mostly to allow testing of ppc64le on x86_64
>>>> based hosts.
>>>> 
>>>> TL;DR version: checkout [1], enable fake_kvm and happy ppc64le hacking :)
> 
> I like the initiative and the approach. Dropping faqemu bits out of
> mainline caps.py was an old task of mine. Moving code into lib/vdsm is a
> big task for our next release anyway.

one more unrelated note
current faqemu is broken on 3.6 as it removes the whole cpu tag which is needed when hotplug memory is enabled, which in turn enables numa, libvirt checks for cpu-numa mapping and fails to find any
please fix it along the way:-)

> 
>>>> 
>>>> Current fake_kvm isn't really a hook and is contained within 'caps'
>>>> module. This is wrong for multiple reasons, the most important one
>>>> being mixing optional with mainline code. Another issue appears when
>>>> one tries to move the fake_kvm code into a hook: the whole notion of
>>>> architectures within VDSM is contained within the 'caps' module.
>>>> 
>>>> The patch series, which git tip is at [1], introduces new cpuinfo
>>>> module that moves information related to architecture to 'cpuinfo'
>>>> module of the VDSM library. Intermediate benefit is that current hooks
>>>> and library code can access the information related to host's cpu.
>>>> 
>>>> This allows for moving fake_kvm code into a hook that I've called
>>>> fakearch. Fakearch is, in my opinion, more suitable name - there is
>>>> barely any KVM faking, but the host 'fakes' selected architecture.
>>>> 
>>>> Faqemu is, on the other hand, a hook. Unfortunately it wasn't really
>>>> updated and doesn't allow running VMs under fake architecture. The
>>>> series therefore try to refactor it to allow cross-arch VMs to be
>>>> started (the VM actually uses host architecture, but from engine's
>>>> point of view it's running on the faked arch).
>>> 
>>> so it will run in full emulation, or?
>> 
>> That is tricky to answer. It is running full emulation in host's
>> architecture, hiding the differences between x86_64 and ppc64le by
>> modifying the XML e.g. spice->vnc when running x86_64 fakearch on
>> ppc64le host.
>> 
>> The reason for that is the qemu-kvm-rhev doesn't seem to support
>> emulation of different arch, so the hook tries to 'fake' it.
>> 
>> The implication is the fact that the underlying (libvirt) VM is quite
>> different to from what we ask for. I wouldn't recommend doing anything
>> with the VM apart from running it and killing it.
> 
> It is called "fake" for a good reason. It should never be used for
> anything but to fake the existence of a VM to the rest of oVirt.
> 
> ... which leads me to one tiny comment: I like the name faqemu. I don't
> see a real need for establishing a new hook with a new name.
> 
> 
>> 
>>> 
>>>> 
>>>> So far tested cross-arch runs are * x86_64->x86_64 (faqemu functionality),
>>>> * x86_64->ppc64le (the most important one),
>>>> * ppc64le->x86_64,
>>>> * ppc64le->ppc64le (faqemu for Power).
>>>> 
>>>> I'm interested in your reviews and comments regarding this effort!
>>>> 
>>>> [1] https://gerrit.ovirt.org/#/c/46962/
>