Re: [Qemu-devel] Review of monitor commands identifying BDS / BB by name

[Kevin Wolf <kwolf@redhat.com>]

Am 05.12.2014 um 10:34 hat Markus Armbruster geschrieben:
> Eric Blake <eblake@redhat.com> writes:
> 
> > On 12/04/2014 08:56 AM, Markus Armbruster wrote:
> >
> >> 
> >> @device is a sub-optimal name for this single parameter.  Either we
> >> accept that and move on, or we deprecate it in favor of a new parameter
> >> with a better name.  I guess the better name isn't worth that much
> >> trouble, in particular since the command name is already ugly.
> >> 
> >> When @node-name is given, @device must not be given.  So @device is
> >> mandatory *except* in the @node-name usage we retain for compatibility.
> >> Deprecate the usage.
> >> 
> >> Command documentation could then look like this:
> >> 
> >> ##
> >> # @block-resize
> >> #
> >> # Resize a block image while a guest is running.
> >> #
> >> # @device: the name of the block backend or node to resize (node names
> >> # supported since 2.3)
> >> #
> >> # @size: new image size in bytes
> >> #
> >> # Deprecated usage (since 2.3):
> >> #
> >> # @device: #optional the name of the block backend to resize
> >> #
> >> # @node-name: #optional name of the node to resize (since 2.0)
> >> #
> >> # Either @device or @node-name must be set but not both.
> >
> > But this isn't discoverable.  You aren't changing whether any parameters
> > are optional, only enhancing the semantics of existing parameters.  How
> > is libvirt supposed to know if qemu is old (node names have to go
> > through node-name) or new (node names are preferred through device)?
> 
> Good point.
> 
> > Just blindly try the 'device' argument, and if it fails, fall back to an
> > attempt with node-name?
> 
> Works, but "try interfaces one after the other until you find one that
> works" is a rather lame discovery technique.

As long as we don't have introspection, it's the only one we have.

> > On the other hand, if 'node-name' becomes the preferred interface, then
> > libvirt has a solution: if node-name is present (it is easy during
> > up-front QMP probing to determine whether 'node-name' is a recognized
> > optional argument or an unknown argument), then always use node-name.
> > As long as libvirt always names the nodes of each device (which it
> > should be doing, but that's a patch series for another day and another
> > list), then a device lookup is never needed.  If 'node-name' is not
> > present, then only the 'device' form is supported, and libvirt can only
> > manage the top image of a backend (but can make that point obvious to
> > the end user that they should upgrade qemu for more functionality).
> 
> If we deprecate @device instead of @node-name, we make the appropriate
> (non-deprecated) use of backend names rather than node names hard to
> probe.
> 
> Command argument introspection could help only if it carried
> "deprecated" flags.  Might be a good idea anyway, but command
> introspection is one of those nice-to-haves we can't seem to deliver.
> 
> A possible alternative is our common way to cheat at probing: when
> probing for A is hard, probe for B, and assume support for B implies
> support for A.
> 
> My guess that a "better name [than @device for the single parameter]
> isn't worth that much trouble" needs to be reevaluated with
> discoverability in mind.  Yes, it's troublesome, but it's also easily
> discoverable.

Still requires trying the new argument and then falling back to @device/
@node-name.

But as long as libvirt doesn't support the node-name interface yet
anyway, I think this discussion is mostly moot.

Kevin