[Geany-Devel] RFC: Proxy plugins

[Matthew Brush]

On 14-05-18 03:07 PM, Thomas Martitz wrote:
> Am 17.05.2014 03:43, schrieb Matthew Brush:
>> On 14-05-16 03:03 PM, Thomas Martitz wrote:
>>
>>>>> I think we want to maintain the ability to add loaders on our own,
>>
>> With Peas you can, in fact it's designed with this purpose in mind,
>> unlike the existing C-only loader.
>
> If it's designed for that, why doesn't it expose that to its users?
> peas-plugin-loader.h is neither documented nor actually shipped, which
> makes me think they want to keep their plugin architecture private
> (subject to change). Given the GNOME community reputation that really
> worries me because we totally want to add loaders to libpeas.
>

Because I guess they assume that if you're going to write another 
language loader you're going to contribute it back to them. It would be 
like adding a Scintilla lexer into Geany's source tree without 
submitting it back upstream.

>
>>
>>>>> without depending on a 3rd party project. Especially for potentially
>>
>> Meh, we depend on lots of 3rd party projects, several directly and
>> many indirectly (see `ldd geany`).
>
>
> Yes, but these generally don't affect the plugin API. When we make
> promises about the stability of the plugin API but depend on a 3rd party
> to fullfill that promise it's not an ideal situation. Especially when
> depending on the GNOME community as with libpeas.
>

This doesn't affect the plugin *API*. The only part that depends on how 
peas does it is in the .plugin ini files stuff. Nothing in the plugin's 
code even really has to know about Peas except maybe calling some type 
registration function, it's just GObject interfaces after that.

>
>>
>> Peas seems to be quite popular in the community around our
>> framework/UI toolkit and is used by several (if not dozens) of popular
>> applications, which also means it's readily available in many popular
>> distros' package repositories.
>>
>> Even if Peas becomes unmaintained/unsupported, we could roll it into
>> our own source tree and maintain it like we do ctags and tagmanager
>> and we still wouldn't have had to write it ourselves :)
>>
>
> Yes I think that would be an acceptable solution once problems arise.
> libpeas is pretty tiny isnt it?
>

I appears to be after some brief looking at the source code.

>
>>>>> creating a compat-loader for our existing plugins. It doesn't seem
>>
>> After experimenting with this a bit, I believe it doesn't make much
>> sense to make a compatibility layer because a) you still need
>> basically the same code as the existing loader code, b) you would have
>> to change/refactor/re-write lots of it c) you would still have two
>> distinct ways to implement plugin interfaces, d) increases chances of
>> breaking the existing loader/plugins and e) intertwining the two
>> loaders would probably make it harder in the distant future to just
>> drop one of them.
>
> The problem is that libpeas demands a different method of making plugins
> available (.ini vs known symbols in the .so file). We definitely want
> only one method, even if it means transitioning the symbol method to the
> .ini file.
>

Yeah, this is actually a bonus for both C and non-C plugins. For one, it 
means Geany doesn't have to actually load the .so files for C plugins to 
read the plugin info, and additionally for not-C languages, it means the 
engine can get metadata about the plugin without involving the loaders 
(ie. no need to PyInitialize() the interpreter just to get metadata of 
unloaded plugin in Plugin Manager GUI).

That being said, if we make existing plugin loader use .INI stuff it 
will break all plugins, so maybe we could make some "opt-in" way that 
doesn't.

> I don't think b) is true, and for e) it would make it actually simpler
> if the one loader is just a plugin of the other one.
>

For b) I mean mostly just shuffling around/re-factoring code. You're 
right about e), re-factoring it this way probably would improve its 
encapsulation/isolation.

>
>>
>>>>> libpeas readily supports this. Unless I'm missing something.
>>>> The loaders (except for C) are plugins, and the build-in ones are
>>>> installed in /usr/lib/libpeas-<version>/loaders/ as .so libraries.
>>>> For example, peas-plugin-loader-python.c contains:
>>>>
>>>> G_MODULE_EXPORT void
>>>> peas_register_types (PeasObjectModule *module) <-- as a regular plugin
>>>> {
>>>>    peas_object_module_register_extension_type (module,
>>>>      PEAS_TYPE_PLUGIN_LOADER,
>>>>      PEAS_TYPE_PLUGIN_LOADER_PYTHON);
>>>> }
>>>
>>> I found that too. But this is not public and not documented. It's also
>>> not installed to /usr/include/libpeas-1.0.
>>>
>>
>> Yikes! Free Software not properly documented! :)
>
>
> It's not just  a case of bad docs here. In this case it means it's not
> documented because it's private to libpeas, i.e. an implementation
> detail. For the same reason the peas-plugin-loader.h is not
> installed/distributed.
>
> While it has a plugin architecture for loaders it doesn't seem designed
> to allow for 3rd party loaders.
>

As above, I presume the idea is that new loaders should be contributed 
back upstream.

>>
>>> Since we want plugin authors to be able to create loaders we need
>>> something that's stable and documented.
>>>
>>
>> Then why do you want to write our own libpeas into the existing C
>> plugin loader? It will (at least for some time) be buggier, and less
>> widely tested/used/maintained (ie. only by us), take lots of extra
>> effort, not automate the bridge between core and the plugins'
>> language, probably be less well-documented, and after all, the end
>> result would most-likely provide an inferior experience for non-C
>> plugin authors.
>>
>
> I do not want to actually write that. However, I'm not convinced of
> libpeas yet (especially because if the .ini aspect), and then there's no
> alternative left. It only has to work well for us so I think the
> maintainance effort is manageable (and we'd have more control).
>

I think you're under-estimating the effort required, even not including 
the proxy plugin changes you want to add in as discussed originally. The 
(reverse) bindings stuff to call out to the non-C plugin languages, the 
work involved in setting up their interpreters, and only loading them 
when actually needed, a way to find meta-info about them without loading 
them (and their interpreters), and probably much more (I guess just look 
at libpeas source, most of it would have to be written, presumably).

> But if libpeas can actually help us I'm happy to adopt it. I really mean
> it.
>

I think it's a pretty good way forward without breaking existing loader 
and it also happens work nicely with adapting GObject-Introspection for 
auto-API-bindings, since that will auto-bind the peas extension points 
(reverse of API bindings), making it actually usable from non-C/GObject 
without extra work.

Cheers,
Matthew Brush