Re: [Qemu-devel] [PATCH V6 8/8] block: Use graph node name as reference in bdrv_file_open().

[Kevin Wolf <kwolf@redhat.com>]

Am 31.01.2014 um 22:37 hat Benoît Canet geschrieben:
> Le Friday 31 Jan 2014 à 21:32:34 (+0100), Max Reitz a écrit :
> > On 28.01.2014 01:04, Benoît Canet wrote:
> > >Le Monday 27 Jan 2014 à 20:11:59 (+0100), Max Reitz a écrit :
> > >>On 27.01.2014 15:36, Benoît Canet wrote:
> > >>>Le Friday 24 Jan 2014 à 15:54:39 (+0100), Max Reitz a écrit :
> > >>>>On 24.01.2014 15:48, Kevin Wolf wrote:
> > >>>>>Am 24.01.2014 um 14:37 hat Max Reitz geschrieben:
> > >>>>>>On 24.01.2014 14:26, Kevin Wolf wrote:
> > >>>>>>>Am 23.01.2014 um 21:31 hat Benoît Canet geschrieben:
> > >>>>>>>>Signed-off-by: Benoit Canet <benoit@irqsave.net>
> > >>>>>>>>---
> > >>>>>>>>  block.c | 6 +++---
> > >>>>>>>>  1 file changed, 3 insertions(+), 3 deletions(-)
> > >>>>>>>I'm not going to merge this one yet. It breaks qemu-iotests case 071,
> > >>>>>>>which would have to be adapted.
> > >>>>>>>
> > >>>>>>>However, first of all I'd like to hear the opinions of at least Eric and
> > >>>>>>>Max on what BlockRef should really refer to. I think node names make
> > >>>>>>>most sense, but perhaps it's a bit inconvenient and the command line
> > >>>>>>>should default to node-name = id when id is set, but node-name isn't?
> > >>>>>>The QAPI schema is pretty clear about this: “references the ID of an
> > >>>>>>existing block device.”
> > >>>>>Sure, that's because I wrote that text before we had a node name.
> > >>>>>
> > >>>>>However, in 1.7 references didn't work yet, so we still have all freedom
> > >>>>>to change the interface as we like.
> > >>>>Yes, that's right.
> > >>>>
> > >>>>>>However, if the ID cannot be found, I think
> > >>>>>>we should interpret it as a reference to the node name.
> > >>>>>>
> > >>>>>>Therefore, I'd first try bdrv_find() and if that returns NULL, try
> > >>>>>>again with bdrv_find_node().
> > >>>>>I think I would prefer to avoid such ambiguities. Otherwise a management
> > >>>>>tool that wants to use the node name needs to check first if it's not
> > >>>>>already used as a device name somewhere else and would therefore operate
> > >>>>>on the wrong device.
> > >>>>>
> > >>>>>On the other hand, a management tool using the same names for devices
> > >>>>>and nodes just gets what it deserves.
> > >>>>>
> > >>>>>Perhaps we should use a common namespace for both, i.e. you get an error
> > >>>>>if you try to assign a node name that is already a device name and vice
> > >>>>>versa?
> > >>>>This is what I would go for. However, then I don't really know why
> > >>>>we should separate the ID and the node name in the first place
> > >>>>(although that's probably because I haven't followed the discussion
> > >>>>around node names).
> > >>>>
> > >>>>Max
> > >>>Ping,
> > >>>
> > >>>I still want to make quorum merge.
> > >>>What should be done for the references ?
> > >>>
> > >>>Best regards
> > >>>
> > >>>Benoît
> > >>My only problem is that I don't really know what IDs are for, then. ;-)
> > >>
> > > From the understanding I have ID are for block backend top level bds and
> > >node-name naming all the bds burried in the graph.
> > >
> > >So my personal opinion would be to relax the constraint on bdrv_lookup_bs
> > >and use it for references.
> > >
> > >Kevin && Max: what do you think of this scheme ?
> > 
> > I agree. For example, we could change the constraint to report an
> > error only if both ID and node name are actually valid (and point to
> > different devices), that is, bdrv_find() and bdrv_find_node() return
> > different non-NULL values.
> 
> Ok I will write patch doing this on top of quorum patches.

Yes, I think allowing bdrv_lookup_bs() to find both node names and
device names makes sense.

I would still use a common namespace and forbid using the same name for
a device and a node.

Kevin