This changes the implementation of the X11 client to use `mio`, as a
polling mechanism, and a custom run loop instead of `calloop` and its
callback-based approach.
We're doing this for one big reason: more control over how we handle
events.
With `calloop` we don't have any control over which events are processed
when and how long they're processes for. For example: we could be
blasted with 150 input events from X11 and miss a frame while processing
them, but instead of then drawing a new frame, calloop could decide to
work off the runnables that were generated from application-level code,
which would then again cause us to be behind.
We kinda worked around some of that in
https://github.com/zed-industries/zed/pull/12839 but the problem still
persists.
So what we're doing here is to use `mio` as a polling-mechanism. `mio`
notifies us if there are X11 on the XCB connection socket to be
processed. We also use its timeout mechanism to make sure that we don't
wait for events when we should render frames.
On top of `mio` we now have a custom run loop that allows us to decide
how much time to spend on what — input events, rendering windows, XDG
events, runnables — and in what order we work things off.
This custom run loop is consciously "dumb": we render all windows at the
highest frame rate right now, because we want to keep things predictable
for now while we test this approach more. We can then always switch to
more granular timings. But considering that our loop runs and checks for
windows to be redrawn whenever there's an event, this is more an
optimization than a requirement.
One reason for why we're doing this for X11 but not for Wayland is due
to how peculiar X11's event handling is: it's asynchronous and by
default X11 generates synthetic events when a key is held down. That can
lead to us being flooded with input events if someone keeps a key
pressed.
So another optimization that's in here is inspired by [GLFW's X11 input
handling](https://github.com/glfw/glfw/blob/b35641f4a3c62aa86a0b3c983d163bc0fe36026d/src/x11_window.c#L1321-L1349):
based on a heuristic we detect whether a `KeyRelease` event was
auto-generated and if so, we drop it. That essentially halves the amount
of events we have to process when someone keeps a key pressed.
Release Notes:
- N/A
---------
Co-authored-by: Conrad Irwin <conrad.irwin@gmail.com>
Co-authored-by: Conrad <conrad@zed.dev>
Release Notes:
- N/A
Fixed#13463 Fixed crash when the locale was non UTF-8 and fixed the
fallback locale.
Fixed#13010 Fixed crash when `compose.keysym()` was `XKB_KEY_NoSymbol`
I also extracted the `xkb_compose_state` to a single place
The clipboard library we use for X11 doesn't yet support multiple
formats on the clipboard, so for now we just store this in memory for
the current zed process, as we do for Wayland.
Fixes: #11971
Release Notes:
- N/A
---------
Co-authored-by: Mikayla Maki <mikayla@zed.dev>
This is a small change that aims to address frames being dropped when we
get a ton of X11 input events.
What it does, in short, is to first read all X11 input events and then
prioritize the rendering.
In my testing, it causes less frames to be dropped when the system is
under heavy load and lots of input events are being created.
Release Notes:
- N/A
This reverts commit f69c8ca74e after it
has already been partially reverted in
https://github.com/zed-industries/zed/pull/13458.
Why the revert?
The changes in that commit/PR fix one type of problem — dropping of
frames when being blasted with input events — but trades it for another
one that I can't explain yet: when the system is under load, then input
becomes _laggy_ and input events seem to be delayed.
Two examples of how that shows up:
1. When the system is under load* and you hold down the `down` key to
scroll, then lift the finger, the cursor stops sometimes. If you then
produce another input event by jiggling the mouse cursor you'll see more
`down`-key events coming up and the cursor moving down. It feels as if
the event loop is not being woken up even though there are still events.
I suspect it might have something to do with XIM, because if it's
disabled, it seems as if problems become less severe.
2. When the system is under load* and you click-and-drag a selection in
the editor, you can see how the selection is delayed and takes 500ms-1s
to catch up to where the cursor is.
* system under load: start Zed, then in another terminal window create a
release build of Zed, for example.
With the changes reverted, the failure mode looks different: we skip
frames. But that, I think, is the better of two bad options, because
skipping frames means that you see what's happening vs. input events
seemingly still coming in seconds after you stopped using the keyboard.
Release Notes:
- N/A
This change ensures that we always render a window according to its
refresh rate, even if there are a lot of X11 events.
We're working around some limitations of `calloop`. In the future, we
think we should revisit how the event loop is implemented on X11, so
that we can ensure proper prioritization of input events vs. rendering.
Release Notes:
- N/A
Co-authored-by: Antonio <me@as-cii.com>
We saw this panic come up:
```
called `Result::unwrap()` on an `Err` value: IoError(Custom { kind: Other, error: UnknownError })
core::panicking::panic_fmt
core::result::unwrap_failed
<gpui::platform::linux::x11::window::X11Window as core::ops::drop::Drop>::drop
core::ptr::drop_in_place<gpui::platform::linux::x11::window::X11Window>
core::ptr::drop_in_place<gpui::window::Window>
gpui::app::AppContext::shutdown
gpui::app::AppContext::new::{{closure}}
gpui::platform::linux::platform::<impl gpui::platform::Platform for P>::run
gpui::app::App::run
zed::main
std::sys_common::backtrace::__rust_begin_short_backtrace
std::rt::lang_start::{{closure}}
std::rt::lang_start_internal
main
__libc_start_call_main
__libc_start_main_impl
_start
```
I'm not sure where exactly that error comes from, except from the X11
stuff. So let's be defensive and log error and only then tear down
everything.
I _think_ that if the error is repeatable that means we won't close the
window but instead just log errors, but I do think that's better than
panicking right now.
Release Notes:
- N/A
Fixes https://github.com/zed-industries/zed/issues/12054
Replaces the `copypasta`/`smithay-clipboard` implementation with a new,
custom one
TODO list:
- [x] Cleanup code
- [x] Remove `smithay-clipboard`
- [x] Add more mime types to the supported list
Release Notes:
- Fixed drag and drop on Gnome
- Fixed clipboard paste on Hyprland
This fixes everything but the main Zed window (GPUI examples, prompt
library, etc.) not being closable by clicking on the X in X11.
We had a dangling reference before: we would remove the window from the
X11 state, but GPUI itself would still have the window in its
references.
In order to fix this we have to call `window.close()`, which ends up
calling `cx.remove_window()`, which removes the reference.
That in turn then causes the reference to be dropped, which cleans up
the X11 state for the window.
Release Notes:
- N/A
On most platforms, things were working correctly, but had the wrong
type. On X11, there were some problems with window and display size
calculations.
Release Notes:
- Fixed issues with window positioning on X11
---------
Co-authored-by: Conrad Irwin <conrad.irwin@gmail.com>
Co-authored-by: Mikayla <mikayla@zed.dev>
This PR adds support for `org.gnome.desktop.interface`'s `cursor-theme`
setting on Wayland. This should fix cursors not showing up on some GNOME
installs. This PR also adds the wiring to watch the current cursor theme
value.
Thanks to @apricotbucket28 for helping debug the issue.
Release Notes:
- N/A
TODO:
- [x] Finish GPUI changes on other operating systems
This is a largely internal change to how we report data to our
diagnostics and telemetry. This PR also includes an update to our blade
backend which allows us to report errors in a more useful way when
failing to initialize blade.
Release Notes:
- N/A
---------
Co-authored-by: Conrad Irwin <conrad.irwin@gmail.com>
Turns out we still get FocusOut and UnmapNotify events after the window
has been destroyed, which resulted in error messages popping up because
we can't find the window anymore that we want to mark as unfocused.
Release Notes:
- N/A
I noticed that when I use my mouse wheel, we get a ton of the
`XkbStateNotify` events, but the modifiers don't change, so we add a ton
of useless input events for the window.
Release Notes:
- N/A
This helps with the problem of keyboard input feeling laggy when the
event loop is under load.
What would previously happen is:
- N events from X11 arrive
- N events get forwarded to XIM
- N events are handled in N iterations of the event loop (sadly, yes: we
only seem to be getting back one `ClientMessage` per poll from XCB
connection)
- Each event is pushed into the channel
- N event loop iterations are needed to get the events off the channel
and handle them
With this change, we get rid of the last 2 steps: instead of pushing the
event onto a channel, we store it on the XIM handler itself, and then
work it off synchronously.
Usually one shouldn't block the event loop, but I think in this case -
user input! - it's better to handle the events directly instead of
re-enqueuing them again in a channel, where they can accumulate and need
multiple iterations of the loop to be worked off.
This does *not* fix the problem of input feeling choppy/slower when the
system is under load, but it makes the behavior now feel exactly the
same as when XIM is disabled.
I also think the code is easier to understand since it's more
straightforward.
Release Notes:
- N/A
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>
This fixes#11236 by ignoring the `bounds.origin` values when the window
is only being resized.
The cause for the issue was that the `ConfigureNotify` event would
contain "wrong" values when the window was being resized (by dragging a
corner).
In my case it would *always* contain x:14/y:49, which is I think might
map to the origin of the top bar in GNOME.
We would then persist these wrong values when serializing the workspace.
On restart, we'd use these values and end up with the window decorations
in the wrong place.
What I still don't know:
1. What exactly the 14/49 map to, because it's not the origin of the top
bar in GNOME. I also tried the X11 TranslateCoordinates call but
couldn't get meaningful results back (even taking scale factor into
account).
2. Why the window decorations end up looking wrong vs. the window being
in the first place. But if you look at my screenshot in #11236, it looks
like the decorations are off exactly by 14/49px.
That being said, I think the solution here is a good one for now: we
don't do an additional X11 call and when we're resizing, we're not
interested in the origin changing.
Release Notes:
- N/A
Proof:
[Screencast from 2024-06-03
15-08-36.webm](https://github.com/zed-industries/zed/assets/1185253/90efccfc-8ec6-42d2-8380-1625eff57805)
Release Notes:
- N/A
Fixes#12198 and some minor fixes:
* IBus was intercepting normal keys like `a`, `k` which caused some
problems in vim mode.
* Wayland: Trying to commit the pre_edit on click wasn't working
properly, should be fixed now.
* X11: The pre_edit was supposed to be cleared when losing keyboard
focus.
* X11: We should commit the pre_edit on click.
---------
Co-authored-by: Mikayla Maki <mikayla@zed.dev>
The method has been tested on:
- Gnome 46 (Working)
- Gnome 40 (Not supported)
Tasks
- [x] Implements a draft which get and provides the user theme to
components which needs it
- [x] Implements a way to call the callback function when the theme is
updated
- [X] Cleans the code
Release notes:
- N/A
This (mostly) allows the CSD added in
https://github.com/zed-industries/zed/pull/11525 to work in X11. It's
still a bit buggy as it detects a second window drag right after the
first one finishes, but it's probably better to change the way window
drags are detected in the title bar itself (as that causes other
issues).
The CSD can be tested by changing the return value of
`should_render_window_controls` to true.
Also fixes F11 crashing.
Release Notes:
- N/A
This pull request adds XIM (X Input Method) support to x11 platform.
The implementation utilizes [xim-rs](https://crates.io/crates/xim), a
XIM library written entirely in Rust, to provide asynchronous XIM
communication.
Preedit and candidate positioning are fully supported in the editor
interface, yet notably absent in the terminal environment.
This work is sponsored by [Rainlab Inc.](https://rainlab.co.jp/en/)
Release Notes:
- N/A
---------
Signed-off-by: npmania <np@mkv.li>
Now, regardless of how the Zed window is closed, Zed can remember the
window's restore size.
- [x] Windows implementation
- [x] macOS implementation
- [x] Linux implementation (partial)
- [x] update SQL data base (mark column `fullscreen` as deprecated)
The current implementation on Linux is basic, and I'm not sure if it's
correct.
The variable `fullscreen` in SQL can be removed, but I'm unsure how to
do it.
edit: mark `fullscreen` as deprecated
### Case 1
When the window is closed as maximized, reopening it will open in the
maximized state, and returning from maximized state will restore the
position and size it had when it was maximized.
https://github.com/zed-industries/zed/assets/14981363/7207752e-878a-4d43-93a7-41ad1fdb3a06
### Case 2
When the window is closed as fullscreen, reopening it will open in
fullscreen mode, and toggling fullscreen will restore the position and
size it had when it entered fullscreen (note that the fullscreen
application was not recorded in the video, showing a black screen, but
it had actually entered fullscreen mode).
https://github.com/zed-industries/zed/assets/14981363/ea5aa70d-b296-462a-afb3-4c3372883ea3
### What's more
- As English is not my native language, some variable and struct names
may need to be modified to match their actual meaning.
- I am not familiar with the APIs related to macOS and Linux, so
implementation for these two platforms has not been done for now.
- Any suggestions and ideas are welcome.
Release Notes:
- N/A
This PR replaces all pointer events on X11 with their XI2 equivalents,
which fixes problems with scroll events not being reported when a mouse
button is down. Additionally it closes#11206 by resetting the tracked
global scroll valulator position with `None` on a leave event to prevent
a large scroll delta if scrolling is done outside the window. Lastly, it
resolves the bad window issue kvark was having.
Release Notes:
- Fixed X11 Scroll snapping (#11206 ).
---------
Co-authored-by: Mikayla Maki <mikayla@zed.dev>
Notable things I've had to fix due to 1.78:
- Better detection of unused items
- New clippy lint (`assigning_clones`) that points out places where assignment operations with clone rhs could be replaced with more performant `clone_into`
Release Notes:
- N/A
Since Wayland doesn't have a way for windows to activate themselves,
currently, when you click on a link in Zed, the browser window opens in
the background.
This PR implements the `xdg-activation` protocol to get an activation
token, which the browser can use to raise its window.
https://github.com/zed-industries/zed/assets/71973804/8b3456c0-89f8-4201-b1cb-633a149796b7
Release Notes:
- N/A
This should have fixed the problems that some users were reporting with
https://github.com/zed-industries/zed/pull/10695 .
The problem was with devices which send more than one valuator axis in a
single event whereas the original PR assumed there would only ever be
one axis per event. This version also does away with the complicated
device selection and instead just uses the master pointer device, which
automatically uses all sub-pointers.
Edit: Confirmed working for one of the user's which the first attempt
was broken for.
Release Notes:
- Added smooth scrolling for X11 on Linux
- Added horizontal scrolling for X11 on Linux
fixes https://github.com/zed-industries/zed/issues/9132
By setting the app id, window managers like `sway` can apply custom
configuration like `for_window [app_id="zed"] floating enable`.
Tested using `wlprop`/`hyprctl activewindow` for wayland, `xprop` for
x11.
Release Notes:
- Zed now sets the window app id / class, which can be used e.g. in
window managers like `sway`/`i3` to define custom rules
Release Notes:
- N/A
Picks up https://github.com/kvark/blade/pull/113 and a bunch of other
fixes.
Should prevent the exclusive full-screen on Vulkan - related to #9728
cc @kazatsuyu
Note: this PR doesn't enable transparency, this is left to follow-up
Reverts zed-industries/zed#10695
Some users are experiencing broken scrolling due to the changes from
this PR, so it should be reverted while I investigate what causes the
problems.
Release Notes:
- N/A
Fernando Tagawa
citorva
Jakob Hellermann
npmania
张小白
Dzmitry MalyshauandJan Solanti
Piotr Osiewicz