[Libguestfs] [PATCH v2v] convert: linux: Require host cpu for all RHEL-alike >= 9

[Daniel P. Berrangé]

On Mon, Feb 06, 2023 at 12:49:41PM +0000, Richard W.M. Jones wrote:
> On Mon, Feb 06, 2023 at 12:37:30PM +0000, Daniel P. Berrangé wrote:
> > On Mon, Feb 06, 2023 at 12:19:40PM +0000, Richard W.M. Jones wrote:
> > > RHEL >= 9 and compatible distros like Rocky >= 9 will not boot using
> > > the default qemu CPU.  You will see an error at boot:
> > > 
> > >   Fatal glibc error: CPU does not support x86-64-v2
> > > 
> > > Instead you need to use -cpu host.
> > > 
> > > In commit f28757c6d1 ("convert_linux: set "gcaps_default_cpu = false"
> > > for x86_64 RHEL-9.0+ guests") we fixed this specifically for RHEL >= 9.
> > > 
> > > This commit extends the same fix to all RHEL family distros.
> > 
> > If v2v is going to use  host CPU for a bunch of distros, why not use
> > it for all distros ? What's the downside in -cpu host, instead of
> > the default qemu64 CPU ?
> > 
> > Even if the distro doesn't require x86_64-v2 ABI, they'll pretty much
> > all benefit from NOT using the default QEMU CPU model, if only for
> > the fact that they gain the 'aesni' feature which increases crypto
> > performance massively. 
> 
> I don't know what we should do, but the current code does this:
> 
>  (1) If a particular CPU model is specified by the source hypervisor
>      (eg VMware), use that.

Reasonable, if there's a match

>  (2) If the guest is RHEL family >= 9, use -cpu host / host-passthrough
> 
>  (3) Otherwise don't specify any -cpu or <cpu> section
> 
> The third option seems to cause qemu / libvirt to use qemu64.  However
> it has the advantage that the guest is migratable afterwards.

I vaguely recall in previous discussion that we decided not to
care if the guest is migratable or not, on the basis that wasn't
really domain knowledge of v2v nor an user request. Thus the
use of 'host-passthrough' rather than 'host-model' to give a
"perfect" match for the host, rather than approximation.


> There may be a case for changing this so:
> 
>  (a) Rule (2) is changed so we use host-model (for the libvirt target).

All hinges on whether users expect the output of v2v to be
migratable or not.

>  (b) Rule (3) is changed so we always specify a minimum CPU, eg.
>      for x86-64 guests, choose x86-64-v2 (is that an actual CPU model??)

In the case of x86-64-v2 it is the QEMU Nehalem model is a perfect
match, and is capable of running on both Intel and AMD hosts. This
is mostly luck though. For x86-64-v3 / -v4 ABIs there is no viable
model that will run on both AMD and Intel. So once there's a guest
OS that mandates -v3, we'll be back to wanting host-model/passthrough.

> (a) seems like a no-brainer.  Not sure why we chose host-passthrough
> instead of host-model?  The explanation in the commit message is not
> very convincing:
> https://github.com/libguestfs/virt-v2v/commit/a3e45b3ea5b45e682cb4b7ad3a5646586c6c7886
> 
> I think (b) is fraught because it is my understanding that there is no
> good choice for a minimum CPU since we don't know anything about the
> target hardware (eg. if it's Intel or AMD).  And on !x86-64 we really
> have no idea at all what's a good choice.

Yeah, I'd recommend against (b).

IMHO mgmt apps / provisioning tools should only ever choose host-model
or host-passthrough for KVM. Use of an explicit named CPU model should
be something left to admins to decide based on their global knowledge of
what machines they need to migrate across. Ignoring named models also
nicely avoids the problem of needing knowledge across all arches.

If you care about TCG, you could use 'maximum' instead of 'host-passthrough'.
They are identical in the case of KVM, and for TCG 'maximum' just gives you
all possible features.

With regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|