Yet another GTK+ update

GTK+ 3.20 was released a while ago; we’re up to 3.20.3 now.  As I tried to explain in earlier posts here and here, this was a pretty active development cycle for GTK+. We landed a lot of of new stuff, and many things have changed.

I’m using the neutral term changed here for a reason. How you view changes depends a lot on your perspective. Us, who implemented the changes, are of course convinced that they are great improvements. Others who maintain GTK+ themes or applications may have a different take, since changes often imply that they have to do work to adapt.

What changed in GTK+

A big set of changes is related to the inner workings of GTK+ CSS.

The CSS box model is much better supported in widgets. This includes padding, margins, borders, shadows, and the min-width and min-height properties. Since many widgets are complex, they typically have many CSS boxes. Here is how the box tree GtkNotebook looks:


In the past (up to and including GTK+ 3.18), we used a mixture of widget class names (like GtkNotebook), style classes (like .button) and widget names (like #cancel_button) for matching styles to widgets. Now, we are using element names for each box (e.g. header, tabs and tab in the screenshot above). Style classes are still used for optional things and variants.

The themes that are included in GTK+ (Adwaita, Adwaita dark, HighContrast, HighContrastInverse and the win32 theme) have of course been updated to follow this new naming scheme. Third-party themes need and application-specific CSS need to be updated for this too.

To help with this, we have expanded both the general documentation about CSS support in GTK+ here and here, and we have documented the element names, style classes and the node hierarchy for each widget. Here, for example, is the notebook documentation.

The documentation is also a good place to learn about style properties that have been deprecated in favor of equivalent CSS properties, like the secondary cursor color property. We warn about deprecated style properties that are used in themes or custom CSS, so it is easy to find and replace them:

(gtk3-demo:14116): Gtk-WARNING **: Theme parsing error: gtk-contained.css:18:37: The style property GtkWidget:secondary-cursor-color is deprecated and shouldn't be used anymore. It will be removed in a future version

There’s also a number of new features in CSS. We do support the CSS syntax for radial gradients, we let you load and recolor symbolic icons, image() and calc() are supported, as well as the rem (‘root em’) unit.

Beyond CSS, the drag-and-drop code as been rearchitected to move the drag cancel animation and most input handling into GDK, thereby dropping most of the platform-dependent code out of GTK+.  The main reason for doing this was to enable a complete DND implementation for Wayland. As a side-effect, we gained the ability to use non-toplevel widgets as drag icons, and we dropped the X11 specific feature to use RGBA cursors as drag icons.

The Wayland backend has grown most features that it was missing compared to X11:  the already mentioned full DND support, kinetic scrolling, startup notification, primary selection, presenting windows, a bell.

Changes in applications

Here is an unsorted list of issues that may show up in applications with GTK+ 3.20, with some advice on how to handle them.

One of the motivations for the changes is to enable animations and transitions. If you use gtk_style_context_save/restore in your draw() function, that prevents GTK+ from keeping the state that is needed to support animations; so you should avoid it when you can.

There is one place where you need to use gtk_style_context_save(), though: when using “theme colors”.  The function gtk_style_context_get_color() will warn when you pass a state other than the current state of the context. To avoid the warning, save the context and set the state:

gtk_style_context_save (context);
gtk_style_context_set_state (context, state);
gtk_style_context_get_color (context, state, &color);
gtk_style_context_restore (context);

And yes, it has been pointed out repeatedly that this change makes the state parameter of gtk_style_context_get_color() and similar functions largely useless – this API has been around sinc e 3.0, when the CSS machinery was much less developed than it is now. Back then, passing in a different state was not a problem (because animations were not really supported).

Another word of caution about  “theme colors”: CSS has no concept of foreground/background color pairs. The CSS background is just an image, which is why gtk_style_context_get_background_color() is deprecated and we cannot generally make it return a useful color. The proper way to have a theme-provided background in a widget is to call gtk_widget_render_background() in your draw() function.

If you are using type names of GTK+ widgets in your CSS, look up the element names in the documentation and use them instead. For your own widgets, use gtk_widget_class_set_css_name() to give them an element name, and use it in the CSS.

A problem that we’ve seen in some applications is the interaction between size_allocate() and draw(). GTK+’s CSS boxes need to know their size before they can draw. If you derive from a GTK+ widget and override size_allocate without chaining up, then GTK+ does not get a chance to assign sizes to the boxes. This will lead to critical warnings from GTK+’s draw() function if you don’t override it. The possible solutions to this problem are either to chain up in size_allocate or to provide your own draw implementation.

If you are using GTK+ just for themed drawing without using GTK+ widgets, you probably need to make some changes in the way you are getting theme information. We have added a foreing drawing example to gtk3-demo that shows how this can be done. The example was written with the help of libreoffice and firefox developers, and we intend to keep it up-to-date to ensure that this use case is not neglected.

A plea

If you are maintaining a GTK+ application (in particular, a big one like, say, inkscape), and you are looking at porting from GTK+ 2 to GTK+ 3, or updating it to keep up with the changes in 3.20, please let us know about the issues you find. Such feedback will be useful input for us when we get together for a GTK+ hackfest in a few weeks.

Whats coming

One of the big incoming changes for 3.22 is a GL-based renderer and scene graph. Emmanuele has been working on this on-and-off for quite a while – you may have seen some of his earlier presentations. Together with the recent merge of (copies of) clutter and cogl into mutter, this will put clutter on the path towards retirement.


My first xdg-app

A few days ago, I set out to get some experience with building an application as an xdg-app.  In this post, I’m collecting some of the lessons I learned.

Since I didn’t want pick a too easy test case, I chose terminix, a promising terminal emulator for GNOME. Terminix uses GTK+ and vte, which means that most dependencies are already present in the GNOME runtime.


However, terminix is written in D, and the GNOME sdk does not include D support.  So, the challenge here is to build a compiler and runtime for this language, and any other required language-specific utilities. The DMD D compiler is written in D (naturally), so some bootstrapping was required.

Build tools

xdg-app comes with low-level build support in the form of various xdg-app build commands. But you really want to use the newer xdg-app-builder tool. It is really nice.

xdg-app-builder downloads and builds the application and all its dependencies, according to a JSON manifest.  Thats par for the course for modern build tools, of course. But xdg-app-builder also has smart caching: It keeps git checkouts  of all sources (if they are in git), and only rebuilds them when they change. It also keeps the results of each modules’ build in an ostree repository, so reusing a previous build is really fast.

All the caches are kept in .xdg-app-builder/ in the directory where you run the build. If you have many dependencies, this hidden directory can grow large, so you might want to keep an eye on it and clean it out every now and then (remember, it is just a cache).

You can take a look at the JSON file I came up with.

Build API

Similar to the GNOME Continuous build system, xdg-app-builder assumes that each module in your JSON supports the ‘build api’ which consists of configure & make & make install. The world is of course more diverse than that, and rolling your own build system is irresistable for some.

Here is a way to quickly add the required build api support to almost any module (I’ve stolen this setup from the pitivi xdg-app build scripts):

Place a foo-configure script next to your JSON recipe that looks like this (note that Makefile syntax requires tabs that were eaten when I pasted this content in here):


cat <<EOF >Makefile
all: whatever is needed to build foo

        ...commands to install foo go here


In the JSON  fragment for the foo module, you add this file as an extra source (we are taking advantage of the fact that xdg-app-builder allows multiple sources for a module):

"modules": [
        "name": "foo",
        "sources": [
                "type": "git",
                 "url": "",
                 "branch": "master"
                "type": "file",
                "path": "foo-configure",
                "dest-filename": "configure"

I guess you could just as well have the Makefile as yet another source; this approach is following the traditional role of configure scripts to produce Makefiles.

Network access

As I mentioned already, the first step in my build was to build a D compiler written in D. Thankfully, the build script of the dmd compiler is prepared for handling this kind of bootstrapping. It does so by downloading a pre-built D compiler that is used to build the sources.

Downloading things during the build is not great for trusted and repeatable builds. And xdg-app’s build support is set up to produce such builds by running the build in a controlled, sandboxed environment, which doesn’t have network access.

So, in order to get the D compiler built, had to weaken the sandboxing for this module, and grant it network access.  It took me a little while to find out that the build-args field in the build-options does this job:

"modules": [
        "name": "dmd",
                "build-args": ["--share=network"]

Shedding weight

After navigating around other hurdles, I eventually succeeded in having a build of my entire JSON recipe run through the end. Yay! But I quickly discovered that the build directory was quite heavy. It came to over 200M, a bit much for a terminal.

xdg-app-builder creates the final build by combining the build results from all the modules in the JSON recipe. That means my terminix build included a D compiler, static and shared libraries for the D runtime, build utilties, etc.

To fix this, I added a couple of cleanup commands to the JSON. These are run after all the modules have been built, and can remove things that are no longer needed.

"cleanup-commands": ["rm -rf /app/lib",
                     "rm -rf /app/src",
                      rm -rf /app/share/nautilus-python",
                      "rm /app/bin/dmd",

Note that the paths start with /app, which is the prefix that xdg-app apps are put in (to avoid interference with /usr).

After these cleanups, my build weighed less than a megabyte, which is more acceptable.

Trying it out

The best way to distribute an xdg-app is via an OSTree repository. Since I don’t have a good place to put one, and Terminix is developed on github, I decided to turn my xdg-app into a bundle, using this command:

xdg-app build-bundle ~/xdg-app-repos/terminix \
                     terminix.x86_64.xdgapp \
                     com.gexperts.Terminix \ 

Since github has a concept of releases, I’ve just put the bundle there: