On Mon, Aug 09, 2010 at 07:38:32PM +0100, Daniel P. Berrange wrote:
On Mon, Aug 09, 2010 at 08:34:23PM +0200, Daniel Veillard wrote:
Grumpf ... that mean at the internal stucture level we need to add an extra field, that is detected at a completely different time in parsing too ... more complex in general, but I can understand the purity POV.
I don't see this as a real problem. It is no different from the way that we automatically add <controller> devices at the end of parsing, if we saw any <disk> or <channel> devices previously.
Huh ??? You can have a controller without a disk and you don't automagically add a controller even if nothing was suggested by the user, it is widely different. First you need to keep a tristate for def->balloonable, was that attribute defined, or not, then add the error case if it's not "yes" or "no", then you need to care for all the different cases of the tristate when you actually parse the devices section, how do you handle <memballon> definition when balloonable="no" was found, discard, error ? I find this different because you provide the information in 2 places at different levels and you ned to cope with conflict ... while still maintaining the current behaviour of autoadding. This then breaks nearly all the qemu XML tests due to the extra balloonable="yes" added automatically to the output since the qemu driver makes that a default. And now you have to explain in documentation all the various cases instead of a simple one liner about using "none". When data is defined at 2 different places by design, well you designed a mess. And in that case that's cerainly true. And after looking at the situation in detail, no I don't by the "purity" point of view, this actually make things *harder* both for user and from a implementation point of view. After messing with this for an hour, my patch is now 160 lines, it's very hard to tell from the code what's the behaviour is in each cases and 3 kind of regression tests need to be fixed to have all XMLs changed to add balloonable='yes' to the memory section. I firmly think that in that case the proper way to do this is to keep the information in one place and that place is the <memballoon> device, type='none' may look ugly to you, but at least the behaviour will be predictable, users will know where to look and that can be explained to the in a single line/sentence. Current patch provided, I strongly suggest to not continue down this line and apply the previous patch instead. Daniel -- Daniel Veillard | libxml Gnome XML XSLT toolkit http://xmlsoft.org/ daniel@veillard.com | Rpmfind RPM search engine http://rpmfind.net/ http://veillard.com/ | virtualization library http://libvirt.org/