On 11/25/2015 05:23 PM, Eric Blake wrote: > Cache the visitor in a local variable instead of repeatedly > calling the accessor. Pass NULL for the visit_start_struct() > object (which matches the fact that we were already passing 0 > for the size argument, because we aren't using the visit to > allocate a qapi struct). Pass "object" for the struct name, > for better error messages. Reflow the logic so that we don't > have to undo an object_add(). > > A later patch will then split the error detection currently > in visit_struct_end(), at which point we can again hoist the > object_add() to occur before the label as one of the cleanups > enabled by that split. > > Signed-off-by: Eric Blake <ebl...@redhat.com> >
>
> - object_add(type, id, pdict, opts_get_visitor(ov), &err);
> -
> out_end:
> - visit_end_struct(opts_get_visitor(ov), &err_end);
> - if (!err && err_end) {
> - qmp_object_del(id, NULL);
> + visit_end_struct(v, &err_end);
> + if (!err && !err_end) {
> + object_add(type, id, pdict, v, &err);
> }
The attempt to avoid a qmp_object_del() cleanup on error was honorable,
but wrong. Calling visit_end_struct() prior to passing 'v' to
object_add() means that object_add() is now visiting a different level
of {} than it was pre-patch, which showed up as a testsuite breakage
under 'make check-qtest'. I'm reworking this patch (and the similar
change to vl.c in 10/23) for v7.
--
Eric Blake eblake redhat com +1-919-301-3266
Libvirt virtualization library http://libvirt.org
signature.asc
Description: OpenPGP digital signature
