Re: [libvirt] [PATCH 00/13] PCIe fixes + new PCI controllers w/RFC

On Thu, Jun 25, 2015 at 08:44:17AM +0200, Martin Kletzander wrote: > On Wed, Jun 24, 2015 at 12:04:33PM -0400, Laine Stump wrote: > >At this point you can see that it all comes down to what name we want to > >give the subelement; within that, we give the exact name of the qemu > >device, and the exact name/value of any qemu options that we set to > >non-default values. > > > >Somebody *please* have an opinion on the name for this, because none of > >these strikes me as better or worse (well, I think I dislike <driver> > > > > To be honest, I kinds dislike all of them. Not that they would be > chosen poorly, no, it's simply because the good sensible choice is > unavailable due to another poor decision in the past (this may be > another point for Michal's talk on KVM Forum). Thinking about it I > must say I don't like how target (which is supposed to match a place > where the device appears for the guest) is used for the model > specification, on the other hand (ab)using 'model' element for the > specification of an "address" in guest (that's what I understand > chassis and port are) doesn't feel any better. What if we go with two > of those elements? Would that be too much pain? E.g.: > > <controller type='pci model='pci-root-port' index='3'> > <address type='pci' bus='0' slot='4' function='1'> > <model type='ioh3420'/> > <target chassis='3' port='0x21'/> > </controller> I like this, essentially a subModel without the camelCase. One small nit: <model name='ioh4320'/> would look nicer to me, but we're using <model type=/> in at least one other place, so I'm fine with both.

Jan > > I understand this might look like an overkill, but I think it's better > safe then sorry, I guess I just see us not so far in the future > regretting any decision made now. >