[libvirt] [PATCH] cpu_ppc64: Add support for host-model on POWER9

[Andrea Bolognani]

On Thu, 2017-06-08 at 14:12 +1000, David Gibson wrote:
> > David, are these known issues in QEMU? Are they being tracked?
> 
> It's not an issue in qemu, it's an issue in the hardware.
> 
> The various bugs in DD1 silicon mean it really isn't compliant with
> the 3.00 architecture (nor, quite, with the earlier ones).  For that
> reason we explicitly disable compatibility modes on a DD1 host.
> 
> More specifically advertising architected mode compatibility causes
> the guest kernel not to turn on some vital DD1 workarounds, which
> makes it break horribly.

Fair enough. That still doesn't account for the lack of
power9 compatibility mode, which AFAICT is simply not
defined at all in QEMU.

Are there plans to introduce it? Should I file a bug to
track it?

> > I don't see much point in updating libvirt to cope with these
> > new compat modes when QEMU itself doesn't.
> 
> Try TCG, it should work there.

Not quite:

  $ /usr/libexec/qemu-kvm \
    -nodefaults \
    -M pseries,accel=tcg \
    -cpu host,compat=power8
  qemu-kvm: Unable to find CPU definition: host

I *can* make it work with

  $ /usr/libexec/qemu-kvm \
    -nodefaults \
    -M pseries,accel=tcg \
    -cpu POWER9,compat=power8
  VNC server running on ::1:5900

but that's not helpful because <cpu mode='host-model'/>
uses the former. The same was true with POWER8, though.


One more issue I've noticed:

  $ /usr/libexec/qemu-kvm \
    -nodefaults \
    -M pseries,accel=kvm \
    -cpu POWER9
  qemu-kvm: Unable to find sPAPR CPU Core definition

Since the recommended configuration for POWER8 guests is
to use

  <cpu mode='custom'>
    <model>POWER8</model>
  </cpu>

I would expect the above command line to work so that
POWER9 guests can be configured in a similar manner. That
seems bug-worthy to me, should I go ahead and file it?

-- 
Andrea Bolognani / Red Hat / Virtualization