This closes a race between setTerms and a slow GetInitialResultSet.
The bug manifests as follows:
- initial search for a short string
- previous results === undefined, call GetInitialResultSet
- user types more, cancel previous search in setTerms()
- mainloop, then _gotResults([])
- previous results === [], !!previous results === true
- therefore call GetSubsearchResultSet with an empty list of results
- _gotResults() from GetSubsearchResultSet is empty
- much later, return from GetInitialResultSet is discarded by
cancellable
- user unhappy because what he searched for is not there
After this fix, the flow is:
- initial search for a short string
- previous results === undefined, call GetInitialResultSet
- user types more, cancel previous search in setTerms()
- mainloop, but no _gotResults
- previous results === undefined, call GetInitialResultSet again with
longer string
- some time later, return from first GetInitialResultSet is discarded
by cancellable
- soon after, return from second GetInitialResultSet comes with good
results
- user happy
https://bugzilla.gnome.org/show_bug.cgi?id=745861
Search providers that should be disabled by default come with
a DefaultDisabled=true key in their keyfile, and are enabled
with the "enabled" whitelist, not with the "disabled" blacklist.
https://bugzilla.gnome.org/show_bug.cgi?id=734110
The asynchronous nature of extension loading, session loading, and more,
makes the code racy as to what is initialized first, and hard to debug.
Additionally, since gjs is single-threaded, the only code we're running
in a thread anyway is readdir, which is going to be I/O bound, so the
code here is actually likely to be faster.
Drop this in favor of some good old fashioned synchronous loading.
Long ago, the search system worked in a synchronous manner: providers
were given a query, and results were collected in a single array of
[provider, results] pairs, and then the search display was updated
from that.
We introduced an asynchronous search system when we wanted to potentially
add a Zeitgeist search provider to the Shell in 3.2. For a while, search
providers were either async or sync, which worked by storing a dummy array
in the results, and adding a method for search providers to add results
later.
Later, we removed the search system entirely and ported the remaining
search providers to simply use the API to modify the empty array, but the
remains of the synchronous search system with its silly array still
lingered.
Finally, it's time to modernize. Promises^WCallbacks are the future.
Port the one remaining in-shell search engine (app search) to the new
callback based system, and simplify the remote search system in the
process.
Allow the prefix 'special:' applied to result IDs to mark results
that should be always shown, even when they would overflow the
maximum results cap. This will be used by epiphany for the special
"Search the Web" result.
https://bugzilla.gnome.org/show_bug.cgi?id=707055
As we only reload search providers on startup or when the sort order changes,
and given the small number of search providers we'll actually load, I doubt
we'll see any speed decrease.
The simplicity of synchronous code is also much clearer, and fully avoids
all the possible bugs about in-flight requests or similar.
This also prevents issues with multiple search providers showing up at once,
which happen when multiple requests to reload search providers get called
immediately, with the existing in-flight async requests never cancelled.
https://bugzilla.gnome.org/show_bug.cgi?id=700283
pushResults, and the original async search API, were originally intended
so search results that weren't immediate could be added as they come in.
Since then, we've decided that the design of search results is that they
should finish at once with all results. Thus, the code was modified so
that pushResults always overwrote the current result set. As such, it makes
sense to rename the method so that the name matches the behavior.
https://bugzilla.gnome.org/show_bug.cgi?id=693836
Since [1], GIO supports generic serialization and deserialization of a
GIcon into a GVariant. This is also implemented by GdkPixbuf and could be
used instead of our homegrown code for it.
This commit adds support to another 'icon' key in the metas dictionary
returned by applications for it. The previous 'gicon' and 'icon-data'
keys are still parsed and supported as before, but are now deprecated.
[1]
https://git.gnome.org/browse/glib/commit/?id=c16f914b40c749b938490a4e10a3c54ec1855c42https://bugzilla.gnome.org/show_bug.cgi?id=698761
"description" is documented as a valid field for search result metas,
and ListSearchResults implements it, so pass it down to be used.
Also, don't wrap the description in quotes, so that the search provider
can decide if it is an excerpt from the searched text or something else.
And to that extent, set use_markup to true, so that terms can be
highlighted.
https://bugzilla.gnome.org/show_bug.cgi?id=694906
This is causing more confusion than anything else these days; the DBus
API is properly documented now and that's what people are expected to
use, the rest are implementation details we're not interested in
exposing.
https://bugzilla.gnome.org/show_bug.cgi?id=681797
Display a '+' icon on the provider icon if there are more results that are
hidden. If the provider icon is clicked, ask the provider to launch itself and
perform a search with the current terms.
https://bugzilla.gnome.org/show_bug.cgi?id=681797
This allows us to fix the shortcomings of the original ActivateResult()
method. In particular:
- allow to pass the search terms to the provider
- allow to pass a user interaction timestamp
https://bugzilla.gnome.org/show_bug.cgi?id=689735
We read the implemented version from the search provider's keyfile, and
then create a RemoteSearchProvider object from the right class
accordingly.
Wire ActivateResult() to the new method (without actually passing the
new parameters along) - an actual implementation will be added in a
future commit.
https://bugzilla.gnome.org/show_bug.cgi?id=689735
Instead of adding search providers to the system as we find them, wait
until we loaded information from all the directories, and then add all
providers at once.
This will be useful when we will sort the providers information
according to the sort order saved in GSettings.
https://bugzilla.gnome.org/show_bug.cgi?id=687491
Initializing this synchronously means that we will possibly wait for the
process to be auto-activated and answering to our call.
If the process is already running it also might not answer immediately
our request, as it might be doing sync I/O.
The right thing to do is to initialize the proxy asynchronously; there
are try/catch blocks in place for when the object is not available, or
not properly initialized.
https://bugzilla.gnome.org/show_bug.cgi?id=687491
This is called in the main thread, which we should never block for
synchronous I/O.
Since the operation we're wrapping is async already, just use
g_file_query_info_async() instead.
https://bugzilla.gnome.org/show_bug.cgi?id=687491
Remote search providers install an auxiliary keyfile to specify
static information, such as the object path/bus name needed to activate
the binary. Such keyfiles also specify the application the providers
pushes results for; currently, we support two formats for application
information
- two fields, "Title" and "Icon" that specify a (translatable) title and
an icon name for display
- one field "DesktopId" that specifies the desktop file name of the
application backing the provider, which obsoletes the previous
Title/Icon syntax
Since all providers in GNOME use DesktopId now, and we need to ensure a
remote search providers is always backed by an application for future
development, this commit drops the support for the older syntax.
https://bugzilla.gnome.org/show_bug.cgi?id=687491
Currently we load all search providers from XDG_DATA_DIRS, so if
the same provider is installed several times in different directories,
we happily show duplicate results. To fix, keep track of all remote
providers we add and skip those that already have been loaded from a
different directory.
https://bugzilla.gnome.org/show_bug.cgi?id=682470
We strongly expect applications to use the same values for the
'Title'/'Icon' fields in their search provider .ini file as the
'Name'/'Icon' fields in their .desktop file. Rather than requiring
applications to duplicate those fields, allow them to specify a
'DesktopId' field instead to point to the corresponding .desktop
file, which makes it possible to ship search provider files without
translatable strings (which is nice given that merging translations
into search provider files lacks a standard rule).
https://bugzilla.gnome.org/show_bug.cgi?id=678816
Currently we pass an icon name for the 'icon' parameter of the
RemoteSearchProvider constructor. In hindsight, using a GIcon
instead will give us a bit more flexibility, so change it.
https://bugzilla.gnome.org/show_bug.cgi?id=678816
As shown in the previous commits, synchronous search is easily implemented
by the asynchronous search API. The only reason we still have a
synchronous search API is of historical reasons. Well, we're not a museum,
and git log can keep our fossils safe if need be....
https://bugzilla.gnome.org/show_bug.cgi?id=675328
Allow applications to register search providers by dropping a keyfile
into a well-known directory. For now, initialize all found providers;
long term, we probably want to give users the ability to restrict the
set of active search providers.
https://bugzilla.gnome.org/show_bug.cgi?id=663125
Add an asynchronous search provider for results from a DBus service
implementing the org.gnome.Shell.SearchProvider interface; this
will allow applications to hook into the Shell's search without
implementing it in Shell itself or requiring an extension.
https://bugzilla.gnome.org/show_bug.cgi?id=663125
