Re: [libvirt] [PATCH 3/5] domaincaps: Report graphics spiceGL

On 05/09/2016 09:48 AM, Daniel P. Berrange wrote: Writing a parser for this type of XML, that maintains behavior in a future proof way, is going to be a lot of work. Hypothetical example: <video supported='yes'> <enum name='modelType'/> <value>cirrus</value> <value>qxl</value> </enum> <enum name='accel2d'/> <value>no</value> </enum> </video> 6 months later, qemu gets some accel2d support for model=qxl <video supported='yes'> <enum name='modelType'/> <value>cirrus</value> <value>qxl</value> </enum> <enum name='accel2d'/> <value>no</value> </enum> <enum name='accel2d'/> <condition> <enum name="modelType" value="qxl"> </condition> <value>yes</value> <value>no</value> </enum> </video> Another 6 months later, qemu gets accel2d support for cirrus only, but it requires <address type='pci'/> and doesn't work for <address type='isa'/> <video supported='yes'> <enum name='modelType'/> <value>cirrus</value> <value>qxl</value> </enum> <enum name='addressType'/> <value>pci</value> <value>isa</value> </enum> <enum name='accel2d'/> <value>no</value> </enum> <enum name='accel2d'/> <condition> <enum name="modelType" value="qxl"> </condition> <value>yes</value> <value>no</value> </enum> <enum name='accel2d'/> <condition> <enum name="modelType" value="cirrus"> <enum name="addressType" value="pci"> </condition> <value>yes</value> <value>no</value> </enum> </video> The app wants to answer the question 'can I specify accel2d by default for model=cirrus?' (note, cirrus, and not qxl) For the first example, the code parses the accel2d enum, doesn't see 'yes', so it doesn't enable accel2d. The code will need to contain some knowledge that enum name='accel2d' == ./video/acceleration/@accel2d but that's presently unavoidable. For the second example, code will need to be added to parse the basic condition, see that modelType=cirrus (./video/model/@type) doesn't match the condition, and so again accel2d=no For the third example, code hits the modelType=cirrus condition, but then also needs to check that addressType (./video/address/@type) == pci, and if so, sets accel2d=yes So, that's all achievable. But what kind of parser does the app write at each step that may fall over with the next XML iteration? What if the parser in step 1 doesn't expect multiple <enum name='accel2d'/> ? If it naively parses only the first instance, then nothing will regress in step #2 or step #3. step #3's conclusion will be wrong, but that's okay. Note this depends on libvirt being smart about appending conditional enums, and not putting them first. However if it parses the last instance, such as by iterating over every XML node and taking the last one that matches name='accel2d' (virtinst used to have a pattern like this when parsing capabilities XML), on step #2 the code will think the <enum> with the 'qxl' condition is the only instance. The parser doesn't know about <condition> yet, will assume accel2d is supported for everyone, and set accel2d=on for cirrus which will probably fail. Okay, starting from a working step #2. The parser knows about multiple enums, and basic <condition> statements. Upgrade to step #3. If the parser was dumb, maybe it didn't know about multiple <condition> enums, which may cause an incorrect match, setting accel2d=yes, and the config fails. Assume the parser is smart enough to know about multiple <condition> enums. It sees addressType, but it doesn't know what XML value it maps to. Does the app just ignore it and hope the config works, possibly generating an error? Does the app completely abandon the check even though it may work correctly? The latter is certainly the safest, but coding errors/naive impls could make mistakes or possible even throw other errors. Now consider that pattern expanding out with conditional or/and/not operators, maybe even numerical comparisons like <, >, etc. The only way for the parser from step #1 to not make a fatal error in step #3 is to implement <condition> parsing from the start and code very defensively. That's a lot of upfront overhead. And I think it will be tempting for apps to try and get by _without_ implementing the whole comparison engine and potentially set themselves up for future pain. Also having a very interconnected schema like this means that if we ever need to extend the XML format we need to think _very_ hard about how parsers are likely handing the existing XML, and how our additions might screw things up, because we are requiring users to perform some kind of comparison on the result. This is much different than extending the domain XML which is largely just a config file. <capabilities> XML is the closet analog but even that requires very little processing, the only tricky bit is handling that <domain type='kvm'> has its own <machine> list And even after all that, the XML is not going to be completely introspectable unless we find a way to describe addressType=./video/address/@type in the XML. So any general parser is still going to need to be extended regularly to handle new enum names. So we end up with XML that appears-introspectable-but-not-quite So let's redo the example with the unique string names and flat namespace: Step #1 is identical Step #2 addition is <enum name='qxlAccel2D'/> <value>yes</value> <value>no</value> </enum> Step #3 addition is <enum name='cirrusAccel2D'/> <value>yes</value> <value>no</value> </enum> And for step #3 we document that it's only valid for address type=pci, If address type=isa grows support later, we add a new top level string if we think it's worth it. The type=pci vs type=isa is a misleading example here, if it was some more obscure option that cirrus accel2d depended on, maybe we would want to put that in the string name. The misinterpretation problems for the app are non-existent. Yes it will need to be extended at step #3 to support cirrus accel2d=yes but that's completely fine IMO. If the next day, xen supports domcaps and wants to expose cirrusAccel2D but with sufficiently different semantics, then we would have to add a new string. Yes that sucks, but it's the safest way by far for apps. Benefits: - The app XML parser is much simpler - The app XML parser is much quicker to bring up. Important because I bet most apps will only want to check a small set of support checks - I suspect extending the libvirt code will be much simpler - We can always come up with a way to add a comparison engine later if this approach gets too out of hand Downsides: - Much less generalizable (but as I said above I don't think this there _is_ a generalizable domcaps XML schema, nor should it be the goal) - Apps/readers need to look to the docs to understand the enum semantics - The app's parser requires more maintenance in relation to how many features it wants to parse I think the guide here should largely be exposing the types of things that we already track with qemu capabilities, and at that granularity. Give that information to the user (which they presently can't find out reliably on their own), and leave the rest of the bits like interdependence of XML options up to them to figure out. So, 'qemu supports spiceGL', and not 'a working spiceGL config is spiceGL with listen=none or listen=unix and video virtio with accel3D' All that said, I'm probably missing some examples where this flat modeling especially sucks too :) So please throw them my way Thanks, Cole