In the future we might want to enable calling it again to set new supported
plugin types/extensions. This is not implemented yet, but in order to
allow this in the future we have to prevent it now, otherwise we'd
need to break the API.
g_ptr_array_insert() is too recent (2.40), but prepending is required. GQueue
is a fine replacement with better old-glib support, at the expense of working
with a doubly-linked list instead of plain array.
Geany now remembers how many plugins depend on a pluxy. It uses this
information to disable the "Active" checkbox in the PM dialog.
Additionally, the PM dialog displays plugins in a hierarchical manner, so that
sub-plugins are shown next to pluxy. This is espcially handy since it makes
the sub-plugin <-> pluxy relationship really obvious, and it's easier to spot
which plugins need to be disabled before the pluxy can be disabled. This allows
to remove code to re-select the plugin because the row (respective to the
hierarchy level) does not change anymore.
This demo proxy does not actually do anything useful. It simply loads
pseudo-plugins from an ini-style file. The point is that there will be a plugin
in the PM dialog for each ini. Each ini-plugin also causes a menu item to be
generated.
When enabling/disabling pluxys in the PM dialog the list of available
plugins might change. If plugins before the pluxy go/come then the wrong
plugin becomes selected (the selected row number stays the same). Re-apply
the selection to the current one in the toggle callback to overcome this issue.
This function finally allows plugins to register themselves as a proxy
for one or more file extensions.
Lots of documentation is added to doc/plugins.dox, please refer to that for more
details.
When a file extension alone is ambigious as to whether a potential plugin is
really handled then the proxy should use the probe hook to find out. This can
be especially helpful when two pluxies work on the same file extension.
The proxy's probe() should return PROXY_IGNORED or PROXY_MATCHED accordingly.
A special flag value, PROXY_NOLOAD, can be or'ed into PROXY_MATCHED to say
that the file belongs to the proxy, but isn't directly loaded and should not
be handled by any other proxy or geany itself.
Example for PROXY_IGNORED:
geanypy only supports python2 at the moment. So, scripts written
for python3 aren't handled by it and should be skipped for the PM dialog.
Or perhaps they are handled by another proxy that supports python3.
Example for PROXY_NOLOAD:
A pluxy registers for the metadata file extension (.plugin) where author etc
is in. The actual implmentation is in a python script (.py). The .py file
is tied to the .plugin and should not be processed by other pluxies. Thus,
the pluxy also registers for the .py extension but returns
PROXY_MATCHED|PROXY_NOLOAD for it (if it would return only PROXY_MATCHED
the sub-plugin would show up twice in the PM dialog).
During the loading of the active plugins they are also initialized (done at
startup). As a result, these plugins could be pluxys and make more plugins
available, some of which may be active as well.
Because of this the loop has to be restarted if pluxies become
available to also load active plugins that depend on the pluxy.
The loop is only restarted at the end so only nested pluxys could possibly
cause the loop to be run more than twice.
Being a GModule is actually a detail of standard plugins. Future proxy plugins
might need different handles. Therefore replace the module field with a more
generic pointer and encapsulate the GModule detail further.
This pointer shall be returned from GeanyProxyFuncs::load and is passed back
to GeanyProxyFuncs::unload, and isn't interpreted by Geany.
Currently they encapsulate loading and unloading of standard plugins. In
the future plugins can provide such functions to load their types of plugins.
Such a dummy proxy plugin is implemented now to load standard plugins so
that these aren't going to be specially handled.
GVFS uses different backends for "native" GIO operations and POSIX
operations which use the FUSE backend. If the two kinds of operations are
mixed, we may get races.
The patch checks the value of file_prefs.use_gio_unsafe_file_saving and
based on it either uses GIO operations or POSIX operations for file loading,
saving and checking modification time.
This also includes obvious changes to msgstr of some languages.
Languages that didn't explicitly use 'l' as previous shortcut or
didn't have 'g' on their translation were left untouched.
The current shorcut key for "Use multi-line matching" conflicts with
"In Selection" when "Use regular expressions" is enabled. It should be
convenient if we change it.
We choose 'g' since other letters are already in use:
u: "Use regular expressions"
s: "Search for"
e: "Use escape sequences"
m: "Mark"
l: "In Selection"
t: "Match from start of word"
i: "In Document"
n: "Replace & Find"
a: "Case sensitive"
c: "Close"
h: "Replace with"
Since geany_load_module() is called for non-enabled plugins you may not
use the plugin API here yet. The only exceptions to this rule are API functions
required for plugin registration.
This rule is hard to enforce (would need to g_return_if_val(PLUGIN_LOADED_OK(p))
for all API functions (well, those taking a plugin pointer anyway), so this
rule is only documented for now.
It was found that because geany_plugin_set_data() could be used by both
plugin's init() and geany_load_module(), that it introduced some uncertainty
as to when to call the free_func. init() callers might expect the call
around the same time as cleanup() is called, while geany_load_module()
callers expected the call at module unload time.
It was indeed called at module unload time. But that means that init() callers
cannot call it again reliably after in a init()->cleanup()->init() flow (when
toggling the plugin) without fully unloading the plugin (which is what we do
currently but that's we would want to change).
With the separation we can actually destroy the data depending on where
it was set and do everything unambigiously.
The documentation provides a quite detailed description of the new loader
In addition it adds a "how to transition" that briefly describes the old
loader (for curious newcomers) and lots of hints for porting legacy
plugins to the new loader.
If the plugin did not set its own user_data we set it to whatever it set
with geany_plugin_register_full() or geany_plugin_set_data().
This is particularly convinient because PluginCallback is usually statically
allocated, at which point dynamically allocated plugin data doesn't exists yet.
- The return value from geany_load_module is removed (void). It was ignored
anyway and we have to check separately whether the plugin loaded OK or not
anyway. If the plugin specific code fails it should simply not call
geany_plugin_register() (which it should only call iff all other conditions
are good).
- GeanyPluginFuncs::init() now returns a bool to allow failing initialization.
Some plugins might want to defer work to their init() (i.e. only do
it when the plugin was activated by the user), and some of that work can
possibly fail (e.g. GtkBuilder fails to load .xml).
Note that the GUI integration of the latter is less than ideal but this kind
of GUI/policy work is out of scope for this patch set. Therefore a plugin
failing to init is simply removed from the PM dialog as if it became
incompatible. However, as the code that generates the list does not call init
they will show up again if the PM dialog is re-opened.
With geany_plugin_set_data() the legacy plugin support can be made
more transparent by using wrapper functions that call the actual plugin_*
functions. This allows to remove the differentiation in code that's not
directly concerned with actually loading plugins.
This commit doesn't change anything except for one thing: legacy plugins now
cannot call geany_plugin_set_data(). But it is meant for new-style plugins
anyway.
The API function adds a free_func parameter, and can also be called
after geany_plugin_register(), i.e. in the plugin's init() callback. This
fixes a by-design memory leak and gives greater flexibility.
This is easier to handle if we decide to add callbacks. Since we can
zero-initialize callbacks before passing it to the plugin we can be certain as
to which callbacks the plugin knew about when it was compiled. This is exactly
the same method used for GeanyPlugin::info already and easier than inspecting
the API version.
The old plugin loader has a number of deficiencies:
- plugins need to export a couple of callback functions into the global namespace
- plugins need to export data pointers, that are written by Geany
- the exported functions have no user_data param, so there is no way to
pass context/state back to the plugin (it needs global storage for that)
- plugin registration is implicit, plugins have no way to not register themselves
(it may want that due to missing runtime dependencies)
- plugins perform the ABI/API verification, and even though we provide a
convinience wrapper, it may get that wrong
As a result, I designed a new loader with the following design principles
- semantics of callbacks should not change, but they they shouldn't be mess
with the global namespace
- each callback receives a self-identifying param (the GeanyPlugin instance) and
a plugin-defined data pointer for their own use
- explicit registration through a new API function
- in-core API/ABI checks
The following principles shall be left unchanged:
- The scan is done on startup and when the PM dialog is opened
- Geany allocates GeanyPluginPrivate for each plugin, and GeanyPlugin is
a member of it
- Geany initially probes for the validity of the plugin, including file type
and API/ABI check, thus Geany has the last word in determining what a
plugin is
- the PM dialog is updated with the proper, translated plugin information
- the PM dialog GUI and user interaction in general is unchanged
With the redesign, plugins export a single function: geany_load_module().
This is called when the GModule is loaded. The main purpose of this function
is to call geany_plugin_register() (new API function) to register the plugin.
This is the only function that is learned about through g_module_symbol().
Within this call the plugin should
a) set the localized info fields of GeanyPlugin::info
b) pass compiled-against and minimum API version as well as compiled-against
ABI version, to allow Geany to verify compatibility
c) pass a pointer to an instance of GeanyPluginFuncs
which holds pointers to enhanced versions of the known callbacks (except
configure_single which is dropped).
d) optionally pass a plugin-private data pointer for later callbacks
Enhanced means that all callbacks receive the GeanyPlugin pointer as the first
and a pdata pointer as the last. pdata is private to the plugin and is set
by geany_plugin_register().
The callbacks need (should) not be globally defined anymore, and the global
GeanyData, GeanyPlugin and GeanyFunctions pointers are ignored and not set
anymore. GeanyData is available through GeanyPlugin::geany_data.
As the edited file can be a remote file on a server with a different time
zone, the mtime can actually be in the future. In this case the check not
only shows the misleading warning but more importantly the
doc->priv->mtime < st.st_mtime
check never happens and the user doesn't get the modified file prompt.
Setting
doc->priv->mtime = time(NULL);
to the current time on file creation isn't harmful in any way because the
saved file's mtime is taken but it's a bit misleading so better to set it
to 0.
