Creating a renderer results in lots of logs. Make it clear that
the logs belong to a multi-GPU renderer (as opposed to a primary
renderer created by the compositor).
The old approach of using a signal is fundamentally broken for a common
usecase: When the waiter is ready, it's common to immediately finish and
free any resources associated with it.
Because of the semantics of wl_signal_emit_mutable() this is UB.
wl_signal_emit_mutable() always excepts that the waiter hasn't been freed
until the signal has finished being emitted.
Instead of over engineering the solution, let's just add a callback required
by wlr_drm_syncobj_timeline_waiter_init(). In this callback, the implementation
is free to finish() or free() any resource it likes.
Up until now only the DRM backend required an output commit after
updating the cursor. Unify this for all backends, because:
- Screen capture can now catch cursor updates listening for output
commits
- In the future we want to make the cursor a regular wlr_output_layer,
which would need an output commit to be updated anyways
Fixes on-screen corruption when displaying a fullscreen client
with an implicit modifier on the secondary GPU.
What happens here:
- Client allocates a buffer with an INVALID modifier on primary GPU.
- Compositor attempts to scan-out this buffer on an output connected
to secondary GPU.
- Buffer is imported to secondary GPU, and is interpreted as if it had
the secondary GPU's implicit tiling, even though it has the primary
GPU's implicit tiling.
We need to forbid cross-device imports with implicit modifiers.
The mgpu_formats list is stripped from any INVALID modifier so
checking that fixes the bug.
Using the mgpu_formats list has an additional benefit: the buffer
is rejected in the test commit if it doesn't have a format supported
by the multi-GPU renderer.
Requires this Mesa bugfix:
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31725
drmModePageFlip() will fail with EBUSY on a disabled CRTC.
We can't fix this by clearing the DRM_MODE_PAGE_FLIP_EVENT when
performing a commit which disables CRTCs, because some device-wide
commits might also page-flip other enabled CRTCs (and skipping the
page-flip event would mess up our buffer tracking).
Fix this by immediately completing page-flips which disable a CRTC
on the legacy uAPI.
Closes: https://gitlab.freedesktop.org/wlroots/wlroots/-/issues/3918
wl_buffer.release event delivery becomes undefined when using the
linux-drm-syncobj-v1 protocol, so we need to wait for buffer
release via a timeline point instead.
The protocol requires both wait and signal timelines to be set, so
we need to create one when the compositor only supplies a wait
timeline.
wlr_renderer and wlr_allocator will print out errors as they go, and end
with a final error if they fail to create anything. The caller of this
function will also log when it fails.
Skip the redundant errors emitted here.
If an mgpu device does not have a renderer, continue without one rather
than ignoring it entirely. It is not guaranteed that we will be able to
scan out to it in any particular configuration, but that is true for any
output regardless, and having the output visible is not harmful even if
it cannot light up.
To proceed safely, we strip implicit modifier support from all planes,
while avoiding duplication of the same logic for mdgpu_formats.
This helps GUD and DisplayLink scenarios.
We create a renderer for the sole purpose of blitting buffers from a
primary renderer that we might not be able to scan-out from. If we end
up with the pixman renderer, it either won't work becuase it cannot
import dmabufs from the primary renderer, or won't have any effect
because the primary renderer already uses dumb buffers.
We test for DMA-BUF capabilities specifically to make it clear what our
interest is, rather than focusing too much on the pixman renderer.
If init_drm_renderer failed, it would destroy the renderer but would not
set it to NULL, leading to use-after-free.
NULL the renderer after destroying it.
create_immed() is a bit dangerous because on failure, either the
connection is closed or the buffer is silently ignored.
Use create() with a roundtrip to figure out whether the import
succeeded.
If we are trying directly scan-out a shm buffer, we create a temporary
buffer pool to instantiate a wl_buffer from, created in accordance to
our buffers size. If the buffer has an offset, it will end up out of
bounds of the created pool.
Extend the temporary pool by the buffer offset to compensate. Matching
the original pool size does matter unless we want to optimize away the
temporary pool to reduce mappings in the parent compositor.
timeline feature will never be applicable to backends without it, so
don't check it. Before this commit, it would cause timeline to be set
to false if libinput ended up being the last item in the list (which is
the case currently).
We store both queued and current buffers to be able to retain both the
framebuffer currently on screen and the one queued to replace it. From a
re-use perspective, we only care about the last committed framebuffer.
The viewport is only stored in order to be re-used together with the
last committed framebuffer, so do away with the queued/current
distinction and store a single viewport updated every time a commit
completes.
Instead of trying to restore the drm state when the session is activated
again, just disconnect all outputs when the session is deactivated. The
scan that triggers on session activation will rediscover the connectors.
Accessing the output state viewport require a buffer, and that might not
have a state with a buffer when preparing the plane properties for an
atomic commit.
Instead, store the properties at the same time as the fb, and use a
similar mechanism to carry the state around.
If our session is re-activated during scanout, restore_drm_device will
reset planes and then attempt an enabling modeset commit without a
buffer. The new plane transform logic requires a committed buffer to be
present to calculate the boxes if they were not explicitly provided, and
at least amdgpu rejects commits that try to use 0 as default.
Skip updating plane props instead of segfaulting if no buffer is set.
A better fix would be to not rely on restore_drm_device at all and
instead require compositors to modeset in response to session
activation.
Fixes: https://gitlab.freedesktop.org/wlroots/wlroots/-/issues/3912
We were checking whether the damage region was empty before
clipping. However a non-empty damage region can become empty after
clipping. Instead, check whether the clipped region is empty.
Fixes: 4339c37f99 ("backend/drm: clip FB damage")
Currently the headless backend does not actually implement the
wlr_output_test function, causing tests containing output state
unsupported by the headless backend to succeed while committing the same
state will always fail.
This commit fixes that by actually hooking up the already exisiting test
function.
References: https://codeberg.org/river/river/issues/1154
Enable scene-tree direct scanout of a single buffer with various options
for scaling and source crop. This is intended to support direct scanout
for fullscreen video with/without scaling, letterboxing/pillarboxing
(e.g. 4:3 content on a 16:9 display), and source crop (e.g. when
1920x1088 planes are used for 1920x1080 video).
This works by explicitly specifying the source crop and destination box
for the primary buffer in the output state. DRM atomic and libliftoff
backends will turn this into a crop and scale of the plane (assuming the
hardware supports that). For the Wayland/X11/DRM-legacy backends I just
reject this so scanout will be disabled.
The previous behaviour is preserved if buffer_src_box and buffer_dst_box
are unset: the buffer is displayed at native size at the top-left of the
output with no crop.
The change to `struct wlr_output_state` makes this a binary breaking
change (but this works transparently for scene-tree compositors like
labwc after a recompile).
wlr_multi_backend sorts the states it is given and tries to perform
sequential backend-wide commits for each sub-backend with the states
that belong to it.
It did not manage the index correctly for the next iteration, so given N
states for a backend it would perform N backend-wide commits.
Clarify the logic by calculating a length rather than an end pointer and
update the index after each iteration.
This piece of code checks for multi-GPU renderer support, so it
needs to run after the renderer is initialized.
Fixes: 514c4b4cce ("backend: add timeline feature flag")
Closes: https://github.com/swaywm/sway/issues/8382
The output feature flag has a flaw: it's not possible to check
whether the backend supports timelines during compositor
initialization when we need to figure out whether we want to enable
the linux-drm-syncobj-v1 protocol.
Introduce a backend-wide feature flag to indicate support for
timelines to address this defect.
Closes: https://gitlab.freedesktop.org/wlroots/wlroots/-/issues/3904
After a connector scan, new connectors might have appeared and old ones
gone away. At this point, old CRTC allocations are already gone, while
new allocations are not yet needed. Skip the call.
