This change ensures that the event loop prioritizes enqueueing another
render or handling user input over executing runnables.
It's a subtle change as a result of a week of digging into performance
on X11. It's also not perfect: ideally we'd get rid of the intermediate
channel here and had more control over when and how we run runnables vs.
X11 events, but I think short of rewriting how we use an event loop,
this is good cost/benefit change.
To illustrate:
Before this change, it was possible to block the app from rendering for
a long time by just creating a ton of futures that were executed on the
"main" thread (we don't have a "main" thread on Linux, but we have a
single thread in which we run the event loop).
That was relatively easy to reproduce by opening the `zed` repository
and starting `rust-analyzer`: at some point `rust-analyzer` sends us so
many notifications, that are all handled in futures, that the event loop
is busy just working off the runnables, never getting to the events that
X11 sends us or our own timer to re-enqueue another render.
When you put print statements into the code to show when which event was
handled, you'd see something like this **before this change**:
```
[ ... hundreds of runnable.run() ... ]
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
new render tick timer. lag: 56.942049ms
X11 event
new render tick timer. lag: 9.668µs
X11 event
new render tick timer. lag: 9.955µs
X11 event
runnable.run()
runnable.run()
runnable.run()
runnable.run()
new render tick timer. lag: 12.462µs
X11 event
new render tick timer. lag: 14.868µs
X11 event
new render tick timer. lag: 11.234µs
X11 event
new render tick timer. lag: 11.681µs
X11 event
new render tick timer. lag: 13.926µs
X11 event
```
Note the `lag: 56ms`: that's the difference between when we wanted to
execute the callback that enqueues another render and when it ran.
Longer lags are possible, this is just the first one I grabbed from the
logs.
Now, compare this with the logs **after this change**:
```
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
new render tick timer. lag: 36.051µs
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
X11 event
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
runnable.run()
```
In-between many `runnable.run()` we'll always handle events.
So, in essence, what this change does is to introduce 2 priorities into
the X11 event queue:
- high: X11 events (user events, render events, ...), render tick, XIM
events, ...
- low: all async rust code
I've tested this with a debug build and release build and I think the
app now feels more responsive. It doesn't feel perfect still, especially
in the slow debug builds, but I couldn't observe 10s lockups anymore.
Since it's a pretty small change, I think we should go for it and see
how it behaves.
Thanks to @maan2003 this now also includes the same change to Wayland.
Release Notes:
- N/A
---------
Co-authored-by: maan2003 <manmeetmann2003@gmail.com>
43d1a8040d