On 11/04/2013 05:31 PM, Stefan Hajnoczi wrote:
On Wed, Oct 30, 2013 at 02:49:32PM +0100, Markus Armbruster wrote:
The first proposal is to add another parameter, say "id". Users can
then refer either to an arbitrary BDS by "id", or (for backward
compatibility) to the root BDS by "device". When the code sees
"device", it'll look up the BB, then fetch its root BDS.
CON: Existing parameter "device" becomes compatibility cruft.
PRO: Clean and obvious semantics (in my opinion).
This proposal gets my vote.
The second proposal is to press the existing parameter "device" into
service for referring to BDS node_name.
To keep backward compatibility, we obviously need to ensure that
whenever the old code accepts a value of "device", the new code accepts
it as well, and both resolve it to the same BDS.
Different legacy commands given the same device name might need to
operate on different nodes.
Could you give an example for this?
Dynamic renaming does not solve this
problem, so I'm not convinced we can always choose a device name
matching a node name.
Device name commands are higher-level than graph node commands. For
example, block_set_io_throttle makes sense on a device but less sense on
a graph node, unless we add the implicit assumption that the new
throttling node is created on top of the given node or updated in place
if the throttling node already exists (!!).
Throttling a node could be useful too, for example if we want to
throttle backing_hd which is on shared storage, but not to throttle
on the local image.
My ignorant question is: Why can't we just use one namespace, make
sure no name collision between node_name and device_name, or even
just drop device_name, so we treat the root node's node_name as
device_name? For commands that only accept a device, this can be
enforced in its implementation by checking against the whole graph
to verify this.