Re: [Qemu-devel] Should we auto-generate IDs?

[Jeff Cody]

On Thu, Aug 27, 2015 at 08:01:12AM -0600, Eric Blake wrote:
> On 08/27/2015 07:51 AM, Daniel P. Berrange wrote:
> > On Thu, Aug 27, 2015 at 09:39:10AM -0400, Programmingkid wrote:
> >>
> >>> Better still might be fixing things to where we add a global command
> >>> line option that outright fails any attempt to create an unnamed object.
> >>> The option would be off by default for back-compat.  But management
> >>> apps like libvirt can turn it on once they are prepared to name every
> >>> object they create (which in turn may imply fixing any remaining
> >>> interfaces that cannot name an object to add in that ability for
> >>> management to pass in a name).  Then there would be no unnamed objects,
> >>> no ambiguity, and no need to generate names.
> >>
> >> I do agree with giving every device an ID, but I don't think failing if 
> >> the user
> >> forgets to give one is necessary. If libvirt doesn't give devices and ID, 
> >> it
> >> would probably benefit from having QEMU do it for libvirt.
> 
> No, you're misunderstanding our argument. The moment there is more than
> one device with an auto-assigned name is the moment that management
> doesn't know which device got which name, so it's better for management
> to pick a name in the first place.
> 
> > 
> > Libvirt always gives an explicit ID.
> 
> Except it doesn't, yet.  Libvirt still needs to be taught to name all
> node devices (and I'm slowly trying to work on patches towards that goal).
> 
> >     It is impossible to rely on QEMU
> > assigning IDs, because there is no reliable way to identify what ID
> > QEMU assigned to each device after the fact. Other management apps
> > would have the same problem, so auto-generated IDs are pretty useless
> > in that respect.
> 
> It's not to say that auto-generated names would be useless when running
> qemu manually from the command line, but I agree that management
> probably can't safely rely on auto-generated names, and therefore
> solving the issue of auto-generating names is less important.
>

I agree that the main benefit for auto-generated names in the
immediate timeframe is likely for the user directly invoking QEMU.
This can be painful and very cumbersome for a user opening a QEMU
image with a large backing chain, for instance.

However, at least in the block layer, it may be nice in the future for
QEMU to have essentially unique IDs that it knows unambiguously
identifies each BDS. Perhaps it would be for BDSs that are not created
or specified by libvirt, but done so internally, and exposed later.

Jeff