[libvirt] [PATCH 0/2] smbios fixes

[Gleb Natapov]

On Thu, Dec 02, 2010 at 08:12:23PM +0000, Daniel P. Berrange wrote:
> On Thu, Dec 02, 2010 at 10:04:59PM +0200, Gleb Natapov wrote:
> > On Thu, Dec 02, 2010 at 07:56:03PM +0000, Daniel P. Berrange wrote:
> > > On Thu, Dec 02, 2010 at 09:44:47PM +0200, Gleb Natapov wrote:
> > > > On Thu, Dec 02, 2010 at 10:36:59AM -0700, Eric Blake wrote:
> > > > > On 12/02/2010 05:47 AM, Daniel P. Berrange wrote:
> > > > > > On Wed, Dec 01, 2010 at 05:12:08PM -0700, Eric Blake wrote:
> > > > > >> In testing <smbios mode='host'/>, I noticed a couple of problems.
> > > > > >> First, qemu rejected the command line we gave it (invalid UUID);
> > > > > >> ifixingthat showed that all other arguments were being given literal
> > > > > >> "" which then made the guest smbios differ from the host.  Second,
> > > > > >> the qemu option -smbios type=1,family=string wasn't supported, which
> > > > > >> lead to a spurious difference between host and guest.
> > > > > >>
> > > > > >> Meanwhile, qemu has a bug where '-smbios type=1,uuid=uuid' must parse
> > > > > >> as a valid uuid, but is otherwise ignored.  The current qemu.git code
> > > > > >> base _only_ sets smbios uuid from the '-uuid uuid' argument.  I've
> > > > > >> filed a bug against the qemu folks asking whether this is intended (in
> > > > > >> which case we have to modify libvirt to change the -uuid argument it
> > > > > >> outputs when smbios is specified), or an oversight (hopefully the
> > > > > >> latter, since it's still nice to correlate
> > > > > >> /var/log/libvirt/qemu/....log with the XML uuid even when that differs
> > > > > >> from the smbios uuid presented to the guest.)
> > > > > > 
> > > > > > Hmm, I thought the UUID we were passing in was a host UUID, but
> > > > > > on closer inspection the type=1 table is definitely intended to
> > > > > > carry the guest UUID. As such it doesn't make sense to populate
> > > > > > that with anything other than the '<uuid>' from the guest XML.
> > > > > > A host UUID should live elsewhere in the SMBIOS tables, likely
> > > > > > in the type2 or type3 blocks
> > > > > 
> > > > > What does that mean to our XML?  Should we reject XML files where both
> > > > > domain/uuid and domain/sysinfo/system/entry/uuid exist, but differ?
> > > > > Both elements are optional, so it's feasible to see a guest uuid in
> > > > > either location.  Meanwhile, I'm waiting on resolution to
> > > > > https://bugzilla.redhat.com/show_bug.cgi?id=659122 to see if qemu will
> > > > > be patched to let us stick the host's uuid in block 2 (base board) or
> > > > > block 3 (chassis), in which case, we'll need to expand our XML to
> > > > > support that notion.
> > > > > 
> > > > As I commented on the BZ use OEM strings type 11 smbios table to pass
> > > > anything you want into a guest.
> > > > 
> > > > > I've also discovered that with current qemu, if both '-uuid uuid' and
> > > > > '-smbios type=1,uuid=uuid' are specified, then -uuid takes precedence;
> > > > > but either one of the two in isolation serves to set smbios block 1 with
> > > > > the guest's uuid.
> > > > > 
> > > > I am surprised that libvirt still uses -uuid.
> > > 
> > > SMBIOS tables are only setup on x86 QEMU binaries, doing fullvirt.
> > > On non-x86, or QEMU used for Xen paravirt, the -uuid arg value
> > > would be used in other ways unrelated to SMBIOS. Thus replacing
> > > -uuid $UUID with -smbiios type=1,uuid=$UUID would be wrong.
> > > 
> > What non-x86 platforms libvirt supports? With Xen use -uuid or whatever
> > they support. With KVM use only smbios type=1,uuid=$UUID. There is not
> > valid reason that I see to use both. But if you use both of them for some
> > strange reason please make sure you provide the same value for both.
> 
> libvirt aims to support any and all QEMU target architectures
> not merely x86. There's no benefit to using -smbios type=1,uuid=$UUID
> with KVM, over -uuid $UUID.
So why do you use them both? Even more interesting question why do you
use them both with different $UUID? If you do, do not complain that there
should be some kind of order between them.

>                              Changing it only for KVM would needlessly
> complicate the code.
> 
Doing things correctly is more important that simple code. And this is
not related to KVM at all. You cannot expect qemu-sparc to work
with qemu-x86 command line, so if you aim to support any and all qemu
targets you should know how to build correct parameter list for any and
all of them.

--
			Gleb.