gbm.rs' lock_front_buffer calls has_free_buffers before calling the underling C lock_front_buffer function on the gbm surface. This has_free_buffers() check succeeds the first time but always fails after that.
However apart from this check, everything else appears to be working perfectly fine. Projects like weston, ksmcube, Qt, or mutter don't do this check beforehand either, so this patch works around gbm.rs by skipping this.
Unfortunately this requires going to gbm-sys directly and implementing the drm traits that require calling the various gbm_bo_get_* functions to retrieve the buffer details for posting to DRM.
As discussed a few days ago, let's define rotation to be always
clockwise, as it is done in other graphics frameworks as well as our
own image rotation. This means changing the definition for the
software renderer. Also fixed docs for the enum in the linuxkms backend
that was wrong.
The event loop roughly loops like this:
loop {
update_timers_and_animations();
render();
sleep_or_dispatch_events();
}
sleep_or_dispatch_events() will wake up when we receive a queued callback from invoke_from_event_loop, and we would
run the callback right away. If that callback sets animated properties, the start time for animations would be incorrect, as we haven't called update_timers_and_animations() yet.
So instead, let's keep track of any received callbacks per sleep cycle, update the animation tick, and then invoke the callbacks.
We trigger update_timers_and_animations() every time the event loop wakes up.
If there's a real animation in the UI going on, that'll also trigger a
request_redraw() call and all will be well.
But when the only source of animation is set_has_active_animations() from
RenderingMetricsCollector, then we might end up in a situation where the event
loop is woken up before the rendered frame is on the screen, which means we
invoke update_timers_and_animations(), but we won't render because the
previous frame isn't shown yet. When it is shown and our presentation callback
is invoked, has_active_animations() is false again.
So instead, remember if we need to re-render once after rendering()
and use that to decide if we need to re-render in the presentation callback.
This is the initial implementation based on DRM dumb buffers,
which cleans up various things at the same time:
- DRM output is separate from GBM and DRM dumb buffer displays.
The latter two re-use the former.
- The timer refresh for the Vulkan KHR Display rendering lives
now with the KHR display code, in hope to be replaced with a
better mechanism in the future.
What's missing still is support for partial updates as well as
rendering with our own software renderer.
Remove the hack of calling present() from our swap_buffers() forwarder (and let it really only forward),
and instead call it explicitly through the shared egl_display member.
The way we can pass along the callback instead of storing it separately.
We don't need to wait 16ms, we could render right away. In theory this could also use invoke_from_event_loop, but then
the callback needs to be sync, which is an unnecessary
complication.
That's less boiler plate for us and better error handling (note how the receive_events()
call on the DRM device now propagates the error). And this also unregisters automatically
on drop.
... instead pass the fd into the calloop event loop and change state when we receive activity on it.
This improves performance slightly, and is a necessary to be able to implement refresh rate throttled rendering (in the next commits).
This patch adds support for the `SLINT_KMS_ROTATION` environment
variable, that instructs the Skia/FemtoVG renderers to rotate
the scene before rendering.
This might fix the issue with resizing causing swapchain re-creation
errors, because "Swapchain creation no longer returns an error when the
swapchain extent doesn't match the current surface extent."
cc #3850
Make the size argument to svg::render optional to mean that it is the
default size of the image.
Otherwise, passing None as the size to ImageInner::render_to_buffer
would not render the image which is possible in some backend (eg: the
button image icon with the qt backend)
And if the image is really rendered on an empty because of layouting or
so, we don't need to show a warning anyway.
Fix#3790
This has two benefits:
- It improves start-up time for device that are touch-only because
the xkb parsing is delayed until a keyboard is attached and key is
pressed
- It fixes#3678 by working around the situation where xkb would
crash if the xkbcommon data files (in /usr/share/X11/xkb) aren't
there. Sadly, no error is reported in that case (that we could
handle), instead just crashes.
This reverts commit 17afe1aa8a.
This causes
thread '<unnamed>' panicked at 'assertion failed: (has_modifier && modifier.is_some()) || (!has_modifier && modifier.is_none())', drm-0.10.0/src/control/mod.rs:324:9
when running on qemu-virgl.
We enable libseat's custom_logger feature by default, which redirects
log output from libseat into the Rust log facade. That's nice, but as
part of the implementation of this feature unfortunately libseat's
build.rs unconditionally adds /usr/local/include to the list of the
search include paths when compiling the log_handler.c file. That breaks
the Yocto build, rightly so:
warning: cc1: error: include location "/usr/local/include" is unsafe for cross-compilation [-Werror=poison-system-directories]
We can do without this feature, so let's not enable it.
This moves most of the version information we need to update into one
place.
Note that the workplace dependency features are in *addition* to any
feature set when using the workspace dependency. So we have all
workspace dependencies defined with `no-default-features = true`.
