compositing: Move compositing off the main thread.

[pcwalton]

Because the compositor no longer runs on the main thread, it can now use
plain old Rust channels for communication, and the main thread instead
receives the special "proxy" treatment. This change also makes the
compositor entirely optional; in headless mode, no compositor is started
at all.

This change will allow us to support multiple windows, since those must
have separate compositors. It should also improve our event dispatch
latency, since even less runs on the main thread, and we can handle
events while waiting for buffer swaps.

Additionally, this patch refactors many of the task startup functions,
which took too many arguments, to use initialization structures. This
makes the code easier to read.

Closes #4228.

r? @glennw

[hoppipolla-critic-bot]

Critic review: https://critic.hoppipolla.co.uk/r/3409

This is an external review system which you may optionally use for the code review of your pull request.

In order to help critic track your changes, please do not make in-place history rewrites (e.g. via git rebase -i or git commit --amend) when updating this pull request.

[pcwalton]

No OS specific stuff was needed because GLFW supplies the necessary functions. This was nice.

[pcwalton]

@glennw Updated with @Ms2ger's comments addressed.

[glennw]

@pcwalton I checked out your branch, but the running the ref-tests hang on linux, and pressing escape or clicking close on about-mozilla.html gives the following:

task '' panicked at 'You should have disposed of the pixmap properly with destroy()! This pixmap will leak!', /home/gw/.cargo/git/checkouts/rust-layers-3d9d5f0b4e4d9002/master/src/platform/linux/surface.rs:206
task 'Constellation' panicked at 'sending on a closed channel', /home/jack/src/rust/src/libsync/comm/mod.rs:571

[glennw]

@pcwalton I had a brief look at this. The root of the problem appears to be that compositor.shutdown() is never called with this patch.

This means that forget_all_tiles() isn't called, resulting in the panic above. If I comment out the panic!() when dropping a leaked surface, then the ref-tests run and pass.

With that change, there is another panic that occurs when closing the window, but I suspect this will be solved when a fix is made to shutdown the compositor as mentioned above.

[pcwalton]

@glennw Glutin has landed while this patch was outstanding. Is there a way to push events to the event queue from another thread with it?

[glennw]

@pcwalton Do you just need to wake up the event queue from another thread? You can use the glutin WindowProxy interface for that. If you need something more than just a wakeup event, let me know what's required and I'll add it.

[metajack]

Assigning @glennw to this, but looks like @pcwalton needs to make the necessary change and rebase.

[glennw]

@pcwalton I think this just needs to be rebased and updated for glutin based on my last comment / question above.

[metajack]

@pcwalton Sounds like this should be easy to land.

[nox]

Is this PR even relevant with compositor being IpcSender<_> now?

[pcwalton]

Yes. We will need this in order to support multiple windows, for example.

[nox]

@pcwalton Would it help if we split this branch in more commits? I can try to separate this:

Additionally, this patch refactors many of the task startup functions, which took too many arguments, to use initialization structures. This makes the code easier to read.

[Ms2ger]

I think that's worth doing independently of this PR.

[nox]

@Ms2ger Yes that's what I was saying meaning.

[nox]

Things I remarked so far:

• many messages sent to the compositor should be sent to the main thread directly (everything where the compositor used to just call something on self.window);
• the compositor doesn't seem to need to send_opt() to the constellation on shutdown, because the compositor doesn't instantiate shutdown ever anymore;
• I have absolutely no idea what to do about WindowMethods::prepare_for_compositing().

[glennw]

@pcwalton Should we close this and re-open it when you have time to look at this again?