So this was my first participation in the WebKitGTK+ hackfest. It was great to have some time to focus on WebKitGTK+ hacking for a few days as well as meeting other colleagues face to face to discuss various related topics, specifically the one I am most interested in: Wayland support in WebKit2.
A few months back I was reviewing the status of WebKitGTK+ in Wayland and mentioned that one of the main challenges was the multi-process architecture introduced with WebKit2, the one that Web/Epiphany is currently using.
The problem is simple enough to explain: In WebKit2, scene composition is done entirely in the Web Process and then painted on the screen in the UI Process, so we need to share a graphics surface between these two processes. In X11 we do this by having the UI Process create an offscreen XWindow and sharing the window ID with the Web Process, but in Wayland there is no direct means to share a graphics surface between two Wayland clients.
The solution for this is to use the same means that a Wayland compositor uses to share graphics with its clients. The meaning of this in the context of WebKitGTK+ is that we need to implement a small Wayland compositor that we can use to share the rendering surface. The way this would work in is like this: the UI Process would play the role of Wayland compositor and the WebProcess would ask it for a surface using regular Wayland APIs. Since the UI Process implements a Wayland compositor it will have access to the graphics buffers rendered by the client (the Web Process) and we have our problem solved. This is quite a bit more of work than simply sharing a Xwindow ID though.
Some months ago I was prototyping a proof of concept for how this would work taking WebKit out of the equation to keep things easy as a first step. During the hackfest I had the opportunity to complete the work and bring it up to date with the current status of Wayland together with the help of my colleague Eduardo Lima This small prototype has two parts: a GTK+ application with a custom GtkContainer widget (which would play the role of the UIProcess in WebKitGTK+) and a separate Wayland client that renders a simple GL scene. The GTK program spawns another process to run the Wayland client when started and also implements the required bits of the Wayland compositor interface to serve Wayland requests as required by the client.
The point of the experiment was to get the GTK program to use the rendering results of the client to paint its own widget contents. This is basically what we need in WebKitGTK+, where the GtkWidget would be the WebView running in the UI Process and the client process would be the Web Process rendering the results of the scene composition to a GL texture.
The next step is to implement this solution in WebKitGTK+, which is a work in progress at the moment. It is still quite a bit of work since the WebKit code base is quite large and complex, but at this point I think it is only a matter of time to get a basic solution to work. Then of course we will have to deal with a lot of other details that this initial proof of concept did not care about, like resizing, managing surfaces for multiple windows and probably a lot more stuff that will pop up along the way.
Finally, there is another interesting consideration to make. Even if the UI Process can share a graphics surface with Web Process, it still has to render it on the GTK widget’s surface. The problem here is that GTK on Wayland uses a cairo image surface as backing or the window surface, so this process involves a copy that results in bad performance. I guess this should be fixed at some point in GTK+ so we can have the same performance we currently have on X11. In the past I tried to go around this by creating accelerated Wayland subsurfaces for the widget and render to that instead. This worked well for performance but it had to be done completely outside GTK+ , and hence it breaks a number of things (for example you have to position the surface manually within the window surface, events are not managed properly, etc), so it was a no go. I suppose that if GTK+ can provide means to manage Wayland subsurfaces for a widget natively, this would also be another option to fix the performance problem.