On Mon, Sep 09, 2013 at 11:59:25AM +0100, Richard W.M. Jones wrote:
On Mon, Sep 09, 2013 at 10:28:39AM +0100, Daniel P. Berrange wrote:
> On Mon, Sep 09, 2013 at 09:09:56AM +0100, Richard W.M. Jones wrote:
> >
> > I tried this patchset and it works at least as far as being able to
> > get the libguestfs appliance up and running on ARM (with KVM) using
> > the libvirt backend. Virtio-scsi & virtio-serial work.
> >
> > Some issues however:
> >
> > - Why is arch="armv7l"? Why not just "arm", or
"armv7hl"?
>
> The kernel reported uname is 'armv7l', so libvirt matches that.
>
> IIUC 'armv7hl' is an RPM invention to show that the armv7l binaries
> were compiled with hard floating point impl.
>
> 'arm' alone is too generic - there are ABI incompatible versions
> armv5, armv6 which we need to distinguish, as well as big-endian
> v7 (eg armv7b).
>
> > - I had to force the <emulator> to be /usr/bin/qemu-system-arm.
> > Otherwise libvirt tries to run /usr/bin/qemu-kvm. May be related
> > to RHBZ#1005581.
> >
https://github.com/libguestfs/libguestfs/commit/37e1e0da5acb34c902d55cde1...
> >
https://bugzilla.redhat.com/show_bug.cgi?id=1005581
>
> If it exists, /usr/bin/qemu-kvm should always point to the
> qemu-system-NNN which matches the host arch. So if it is
> pointing to the x86 emulator that is broken Fedora packaging
>
> > - I had to add -machine kernel_irqchip=off (for KVM):
> >
https://github.com/libguestfs/libguestfs/commit/f8d89bc42fbd930c61e4f6a81...
> >
http://www.mail-archive.com/arm@lists.fedoraproject.org/msg05546.html
>
> I'd say that's something for QEMU / KVM guys to fix. If the kernel
> IRQ chip is known broken for ARM, QEMU shouldn't be trying to use it.
>
> > - Setting <cpu mode="host-model"/> confuses libvirt
completely.
>
> Yeah, CPU model stuff for non-x86 seems somewhat problematic in my
> experiance.
One more:
- Hotplugging (adding a virtio-scsi disk) doesn't work on ARM:
could not attach disk to libvirt domain: internal error: SCSI controller 0 was missing
its PCI address [code=1 domain=10] at
/home/rjones/d/libguestfs/tests/hotplug/test-hot-add.pl line 50.
Sounds like something is mistakenly assuming PCI here. Please file a
bug for this one against libvirt.
Daniel
--
|: