Re: [Qemu-block] [PATCH v2 00/16] qapi: Allow blockdev-add for NBD

[Kevin Wolf]

Am 01.03.2016 um 11:00 hat Daniel P. Berrange geschrieben:
> On Tue, Mar 01, 2016 at 12:37:14AM +0100, Max Reitz wrote:
> > On 01.03.2016 00:24, Eric Blake wrote:
> > > On 02/29/2016 04:19 PM, Max Reitz wrote:
> > >> Turns out NBD is not so simple to do if you do it right. Anyway, this
> > >> series adds blockdev-add support for NBD clients.
> > >>
> > >> Patches 1 and 2 add one less and one more complicated QDict function,
> > >> respectively, which I needed in later NBD patches: Patch 1 for handling
> > >> legacy options (move "host" to "address.data.host" etc.) and patch 2
> > >> because I'd like to use the input visitor for transforming the NBD
> > >> options into a SocketAddress. Unfortunately, the block layer uses
> > >> flattened QDicts everywhere, so we'll have to unflatten (erect?) them
> > >> before we can use that visitor.
> > > 
> > > Dan had a patch proposal that called the operation "crumple"; I need to
> > > review both proposals and see which one I like.
> > > https://lists.gnu.org/archive/html/qemu-devel/2016-02/msg04618.html
> > 
> > Well, here I go again, not looking at patches on the list...
> > 
> > Looking at the design, I like his idea of having an escape sequence;
> > also, his qdict_crumple() can return boths lists and dicts where my
> > qdict_unflatten() only returns dicts (then again, this is what
> > qdict_flatten() always works on). And his patch doesn't suffer from as
> > much indentation as mine does.
> 
> The escape sequence is critical to support for my use case, because
> sadly some object properties have '.' in their name :-(
> 
> > What I like more about my patch, however, is that I'm reusing
> > qdict_array_split() and qdict_array_entries(). That is mostly because my
> > function modifies the given QDict, where Dan's does not.
> 
> The reason for that is that the use context in which I need to call
> qdict_crumple() has a const QDict, so modifying the original QDict
> was not an option.

You can always clone and modify if modifying an existing QDict turns out
to be nicer to implement.

> Second, for error handling, if there is a problem we can't resolve
> half way through the unflattening process, then if you're modifying
> the original QDict you end up with a QDict that is a hybrid between
> the flat & unflat forms. I think it is pretty bad practice for API
> design / behaviour to leave inputs in such a state on error. ie if
> the code isn't capable of rolling back to the original state it
> should not be modifying the input arg.

I think we generally abort the whole action in such error cases. Then
it doesn't really matter in what state the to be freed QDict is.

Kevin