We have been caching the paths to the plugins but before we would only
cache the specific plugin that we were looking for. This meant that we
might search through all of the plugin directories before finding the
specific module we were interested in. The next plugin we needed we
might again search through all the plugin directories before finding the
plugin we wanted.
This new code will cache all the potential plugins in each directory
searched. For a play that only uses one plugin we'll only search
through directories until we encounter the directory that has the proper
plugin. For a large playbook with many plugins used we'll search
through each directory at most once each.
This should alleviate #10165
'_', they will be loaded after non prefixed modules are checked they can
be full modules or symlinks to existing ones (alias)
also updated ansible doc to ignore these, will eventually add selective
display
by ensuring all basedirs, plugin paths and extra
paths are handled as absolute paths and are checked
to not add any doubles.
This fixes the corner case where e.g. the user has
an additional plugin path configured to a dir
relative to his playbooks or inventory location,
which also matches the _plugin subdir relative to
one of the basedirs in the play.
For most plugins this doesn't show as an obvious issue
except for callback_plugins, which might fire more
than once. Other plugins (inventory and template
plugins) might unnecessarily be ran twice.
e.g. ansible.cfg has
callback_plugins = ./plays/callback_plugins
and plays/ contains a playbook file:
.
├── ansible.cfg
├── inventory
└── plays
├── callback_plugins
│ └── timestamp.py
└── site.yml
modified: lib/ansible/utils/plugins.py
ansible.constants was calling expanduser (by way of shell_expand_path)
on the entire configured value for the library and *_plugins
configuration values, but these values have always been interpreted as
multiple directories separated by os.pathsep. Thus, if you supplied
multiple directories for one of these values, typically only the first
(at least on *nix) would have e.g. "~" expanded to HOME.
Now PluginLoader does expansion on each individual path in each of
these variables.
PluginLoader._get_paths, as of 391fb98e, was only finding plug-ins that
were in a subdirectory of one of the basedirs (i.e. in a category
directory). For example, action_plugins/foo.py would never be loaded,
but action_plugins/bar/foo.py would work.
This makes it so that "uncategorized" plug-ins in the top level of a
directory such as action_plugins will be loaded, though plug-ins in a
"category" subdirectory will still be preferred. For example,
action_plugins/bar/foo.py would be preferred over action_plugins/foo.py.
PluginLoader.add_directory() can receive None from, for example,
Inventory.add_directory(self.basedir()) if host_list is a custom list.
None has no reasonable interpretation other than ignore it.
