The new wlr_renderer_autocreate API is great for compositors, however
it causes some issues with DRM multi-GPU support.
A DRM child backend wants the compositor to use the parent GPU, so it
exposes the parent's DRM FD in get_drm_fd. However, in order to be able
to perform multi-GPU buffer copies, the child DRM backend still needs to
create a local renderer.
Use the new private wlr_renderer_autocreate_with_drm_fd function to
avoid creating a renderer for the parent GPU.
Fixes: e128e6c08d ("render: drop egl parameters from wlr_renderer_autocreate")
Instead of requiring callers to manually make the EGL context current
before binding a buffer and unsetting it after unbinding a buffer, do
it inside wlr_renderer_bind_buffer.
This hides renderer-specific implementation details inside the
wlr_renderer interface. Non-GLES2 renderers may not use EGL.
This removes all EGL dependencies from the backends.
References: https://github.com/swaywm/wlroots/issues/2618
References: https://github.com/swaywm/wlroots/pull/2615#issuecomment-756687006
Since 5b1b43c68c ("backend/drm: make wlr_drm_plane.{pending,queued,current}_fb
pointers"), current_fb can be NULL if there's no buffer. If current_fb
is not NULL, current_fb->wlr_buf is guaranteed to not be NULL.
Closes: https://github.com/swaywm/wlroots/issues/2634
Stop using wlr_drm_surface.{width,height} to figure out the size of a
gbm_bo. In the future we'll stop using wlr_drm_plane.surf, so these will
be zero. Instead, rely on gbm_bo_get_{width,height}.
Stop keeping track of buffers on the parent GPU when multi-GPU is used.
This removes support for export_dmabuf on secondary GPUs, but renderer
v6 will bring this back by managing the swapchains in wlr_output instead
of the backends.
keyboard_handle_leave would always process 1 keycode more than was
pending, which meant reading uninitialized memory from the "pressed"
array.
Found by valgrind.
Sometimes wlr_session_find_gpus will encounter an error. This is
different from finding zero GPUs.
On error, wlr_session_find_gpus already returns -1. However, this is
casted to size_t, so callers uncorrectly assume this is a success.
Instead, make wlr_session_find_gpus return a ssize_t and allow callers
to handle the error accordingly.
wlr_drm_connector.crtc may be updated by the DRM backend while a
page-flip is pending. In this case, the page-flip handler won't be able
to find the right wlr_drm_connector from the CRTC ID.
Save the CRTC when performing a page-flip to ensure we always find the
right connector when we get the event.
When the session is inactive, we can't change the KMS state. Ignore
hotplug events so that compositors don't try to perform a modeset when
a connector is plugged in. We already re-scan connectors when the
session becomes active.
To test, run a wlroots compositor on VT 1, switch to VT 2, unplug a
connector, re-plug it, switch back to VT 1. Without this patch the
screen is black on VT 1.
References: https://github.com/swaywm/wlroots/issues/2370
If we get an authenticated primary node from the X11 server, don't use
it because we can't authenticate our Wayland clients with it. Instead,
open a render node.
Closes: https://github.com/swaywm/wlroots/issues/2576
This callback allowed compositors to customize the EGL config used by
the renderer. However with renderer v6 EGL configs aren't used anymore.
Instead, buffers are allocated via GBM and GL FBOs are rendered to. So
customizing the EGL config is a no-op.
We now properly mark the cursor plane's formats as linear-only, and we
now have a version of wlr_drm_format_intersect that handles the case of
linear-only formats and implicit modifiers.
We can remove the special drm_plane_init_surface flag we had for cursor
planes. This also allows us to use a non-linear layout for cursor planes
on drivers that support it.
Tested on amdgpu GFX9.
If the kernel driver doesn't support modifiers, it still expects cursor
FBs to have a LINEAR layout. See [1] for expectations for framebuffers
attached to the cursor plane.
[1]: https://patchwork.freedesktop.org/patch/408512/
The Present protocol states:
> An event context is associated with a specific window; using an existing
> event context with a different window generates a Match error.
Instead of a global event context, use a per-window event context to fix
this error:
[backend/x11/backend.c:608] X11 error: op Present (SelectInput), code Match (no extension), sequence 63, value 4194307
Closes: https://github.com/swaywm/wlroots/issues/2577
Instead of using a timer, rely on X11 Present events and send a new
frame event when the parent compositor displays a new frame on screen.
The previous attempt at doing this [1] hit issues with EGLSurface, but
we don't use that anymore.
[1]: https://github.com/swaywm/wlroots/pull/1894
Parse WLR_DRM_NO_MODIFIERS at startup. Don't parse IN_FORMATS when
WLR_DRM_NO_MODIFIERS is set, so that the legacy behaviour is better
reproduced.
When modifiers aren't supported, try the initial page-flip once only.
This was lost during the refactoring. We were previously calling
wlr_output_destroy, which destroyed the connector as well.
Fixes: 248c7787c7 ("backend/drm: refactor wlr_output destruction")
The DRM backend is a little special when it comes to wlr_outputs: the
wlr_drm_connectors are long-lived and are created even when no screen is
connected.
A wlr_drm_connector only advertises a wlr_output to the compositor when
a screen is connected. As such, most of wlr_output's state is invalid
when the connector is disconnected.
We want to stop using wlr_output state on disconnected connectors.
Introduce wlr_drm_connector.name which is always valid regardless of the
connector status to avoid reading wlr_output.name when disconnected.
Simplify and unify connector-specific logging with a new
wlr_drm_conn_log macro. This makes it easier to understand which
connector a failure is about, without having to explicitly integrate the
connector name in each log message.
The workaround is broken because drm_fb_acquire doesn't leave the EGL
context current anymore. We'll need to re-introduce it.
References: https://github.com/swaywm/wlroots/issues/2525
We queried DRI3 formats, but we weren't using them. Because of a typo,
only render formats were used.
Fixes: c59aacf944 ("backend/x11: query modifiers supported by X11 server")
Closes: https://github.com/swaywm/wlroots/issues/2552
Backends will eventually stop creating their renderer. To prepare for
this, stop using EGL_PLATFORM_SURFACELESS_MESA in the headless renderer.
Pick a render node using libdrm.
The new allocator/renderer creation logic looks very much like what will
end up in common code.
gbm_bo_create_with_modifiers doesn't take GBM flags, so our
wlr_gbm_allocator interface doesn't either. We were still internally
using GBM flags in the DRM backend, leading to awkward back-and-forth
conversions.
The only flag passed to drm_plane_init_surface was GBM_BO_USE_LINEAR, so
turn that into a bool to make sure other flags can't be passed in.
Move the "force linear" logic out of init_drm_surface, because the
supplied wlr_drm_format should already contain that information.
- The DRM backend initially doesn't have a frame scheduled initially.
However the compositor is expected to set a mode to start the
rendering loop (frame_pending is set to true in drm_crtc_pageflip).
- The headless and X11 backends have a timer to schedule frames, so they
ignore this hint completely.
- The Wayland backend renders a fake frame to start the rendering loop.
It's the only case where a frame is pending on init, move the
assumption there.
I was about to add a check to fail instead of crash when the compositor
uses direct scan-out, but with renderer v6 it's so simple to just add
support for direct scan-out, why bother?
Closes: https://github.com/swaywm/wlroots/issues/2523
drm_fb_import_wlr may need to change the current EGL context. For
instance by calling drm_fb_clear, which calls wlr_buffer_unlock, which
may destroy a buffer if the cursor swapchain size has changed, which
calls gles2's destroy_buffer, which calls glDeleteFramebuffers.
Closes: https://github.com/swaywm/wlroots/issues/2479
This is just the vendor-agnostic define for the GBM platform. It has the
same value as EGL_PLATFORM_GBM_MESA, so should work with old drivers
that only offer the Mesa-vendored extension too.
On some platforms it's possible that the display engine supports
modifiers not supported by the render engine.
Query render formats and intersect them with plane formats to accommodate
for this.
The Wayland platform doesn't have visuals. By chance,
WL_SHM_FORMAT_ARGB8888 is zero, which means egl_get_config was ignoring
it and everything worked fine.
After discussing with Pekka and Daniel on #dri-devel, we concluded [1]
that user-space shouldn't need to force-probe connectors. Force-probing
can take some time, so using drmModeGetConnectorCurrent can result in
faster start-up.
Users can manually trigger a force-probe if necessary:
echo detect | sudo tee /sys/class/drm/card0-DP-1/status
Or just by running a tool like drm_info.
A similar change has been submitted to Weston [2].
[1]: https://lists.freedesktop.org/archives/dri-devel/2020-November/287728.html
[2]: https://gitlab.freedesktop.org/wayland/weston/-/issues/437
Wait for a DRM device if none is found in wlr_session_find_gpus. This
can happen if the compositor is loaded before the display kernel driver.
This supersedes the logind CanGraphical property.
To test, e.g. with i915 and sway:
rmmod -f i915
sway &
modprobe i915
Closes: https://github.com/swaywm/wlroots/issues/2093
We're only interested in card devices. The loop over wlr_session.devices
would take care of ignoring non-card events, but a future patch will
listen to udev "add" events as well.
Instead of operating on FDs in {open,close}_device, operate on
wlr_devices. This avoids the device lookup in wlr_session and allows
callers to have access to wlr_device fields.
For now, we use it to remove wlr_session_signal_add and replace it with
a more idiomatic wlr_session.events.change field. In the future, other
events will be added.
We would always return the GAMMA_LUT_SIZE property if available, and
only fall back to legacy gamma size otherwise. This leads to issues if
both are available in differs in size while we use the legacy backend.
Ensure that we only return the legacy size if we're using the legacy
backend.
Closes: https://github.com/swaywm/wlroots/issues/2429
We have the policy of requiring up-to-date dependencies instead of
adding conditionals for older versions. libinput 1.14 was published more
than 1 year ago.
Every host seat with pointer capability propagates events to one of
sub-pointer depending which output window we entered.
active_pointer tracks reference to sub-pointer on enter/leave events to
avoid lookup for it on every move events.
Fixesswaywm/wlroots#1499
This effectively gets swaywm/wlroots#1499 to the point where
functionality somewhat preserved and no crash happens.
We still can have only one cursor, but we can control it from multiple
seats in time-sharing manner by entering/leaving output.
When performing a modeset, the DRM backend will request a page-flip
event. However frame_pending wasn't set to true, so any subsequent
wlr_output_schedule_frame calls would imemdiately trigger a synthetic
frame event, asking the compositor to submit a new frame. Committing the
new frame fails with "a page-flip is already pending" error in the DRM
backend.
When an output is disabled one last pageflip will happen to disable it.
Currently this pageflip causes a frame event.
Since the output is disabled we don't want to send this frame event.
I'm not sure what this was used for, but it's not used by libwayland.
Setting _WAYLAND_DISPLAY would result in the Wayland backend being
picked but would ignore the actual value of the env variable.
When starting a compositor that's using the "direct" session backend,
wlroots needs to handle calls to `drmSetMaster()` and `drmDropMaster()`.
As both calls used to require `CAP_SYS_ADMIN`, wlroots thus simply
refused starting in case the process doesn't enjoy evelated privileges.
Permission rules have changed since linux.git commit 45bc3d26c95a (drm:
rework SET_MASTER and DROP_MASTER perm handling, 2020-03-19). As a
result, starting with Linux v5.8, both ioctls will now also succeed if
the process is currently or has been the DRM master. And as the first
process to open render nodes will become the DRM master automatically,
this effectively means that process elevation is not strictly required
in all setups anymore.
So let's drop the `geteuid() != 0` permission check to allow those new
rules to do their magic.
bad1e9afa8 ("session: Add libseat backend") introduced a change to to
how session backends initialize, but failed to update the FreeBSD
specific version of the direct backend accordingly.
Closes: https://github.com/swaywm/wlroots/issues/2376
get_drm_prop_blob does not set path_len if it returns NULL. Check the
return value before path_len to avoid reading uninitialized memory.
(Granted, this doesn't change the logic at all, but it does make
Valgrind a bit happier.)
drm_connector_set_cursor wasn't checking the return value of the
drm_surface_make_current call. On failure, this results in a failed
assertion in wlr_renderer_begin (because no rendering context is
current).
The output backend API is now mostly state-less thanks to the atomic
hooks (commit and test). There is one exception though: attach_render.
This function makes the rendering context current. However sometimes the
compositor might decide not to render after attach_render (e.g. when
there's nothing new to render to the back buffer). Thus
wlr_output_rollback has been introduced to revert the pending state.
Because the output backend API is mostly state-less, the only thing
wlr_output_impl.rollback needs to do is revert the current rendering
context. Rename the function to rollback_render to make this clear. Add
a check in the common wlr_output code to only call rollback_render when
attach_buffer has been previously called.
On the long term, we'll be able to remove attach_render and
rollback_render together.
This commit makes `get_current_time_msec` correctly return milliseconds
as opposed to microseconds. It also considers the value of `tv_sec`, so
we don't lose occasionally go back in time by one second. Finally, the
function is moved into `util/time.cc` so that it can be reused elsewhere
without having to consider these pitfalls.
This patch fixes this failure:
01:57:16.642 [ERROR] [backend/drm/drm.c:360] Failed to page-flip output 'eDP-1': a page-flip is already pending
01:57:16.684 [ERROR] [backend/drm/drm.c:360] Failed to page-flip output 'eDP-1': a page-flip is already pending
01:57:16.684 [ERROR] [backend/drm/drm.c:732] Failed to initialize renderer on connector 'eDP-1': initial page-flip failed
01:57:16.684 [ERROR] [backend/drm/drm.c:805] Failed to initialize renderer for plane
01:57:16.684 [sway/config/output.c:423] Failed to commit output eDP-1
References: https://github.com/swaywm/sway/issues/5101
When the mode, status or buffer hasn't changed, drm_connector_commit was
a no-op. Because of this individual changes to the gamma LUT or adaptive
sync status were ignored (if committed without a buffer).
Instead, perform a commit to apply the changes.
A test commit doesn't apply the pending state.
The CRTC state will be populated again if the compositor decides to
perform a regular commit afterwards.
Previously, we only had the pending state (crtc->pending, crtc->mode and
crtc->active). This causes issues when a commit fails: the pending state
is left as-is, and the next commit may read stale data from it.
This will also cause issues when implementing test-only commits: we need
to rollback the pending CRTC state after a test-only commit.
Introduce separate pending and current CRTC states. Properly update the
current state after a commit.
retry_pageflip is now dead code, since drm_connector_start_renderer
isn't called anymore. It was previously called when enabling an output.
The name "retry_pageflip" was a little confusing because the function
retried a modeset and the timer wasn't set up while performing a simple
page-flip.
Let's just remove this altogether for now. We can discuss whether it's
worth it to bring it back. Should we only do it on failed page-flips?
Should we only do it on EBUSY?
If the atomic commit fails or is test-only, rollback
wlr_drm_crtc.{mode_id,gamma_lut} blob IDs to their previous value. This
prevents the next commits from failing or applying test-only state.
After swapping buffers, it doesn't make sense to perform more rendering
operations. Unset the context to reflect this.
This commit makes it so the context is always only current between
wlr_egl_make_current and wlr_egl_swap_buffers.
This is an alternative to [1].
[1]: https://github.com/swaywm/wlroots/pull/2212
This function can be called after wlr_egl_make_current to cleanup the
EGL context. This avoids having lingering EGL contexts that make things
work by chance.
Closes: https://github.com/swaywm/wlroots/issues/2197
When calling assert(0) instead of returning a value, -Wreturn-type
warnings are triggered because assertions can be disabled. Replace these
assertions with abort().
We already have the logic to fallback to the legacy interface above. We
just need to avoid calling atomic_add with a zero prop ID.
Closes: https://github.com/swaywm/wlroots/issues/2187
dealloc_crtc was destroying GBM surfaces, but the cursor_enabled flag
was left as-is. When re-enabling the output, atomic_crtc_pageflip would
try enabling the cursor plane, but would fail because no framebuffer is
available.
Closes: https://github.com/swaywm/wlroots/issues/2150
When an output is enabled and modeset at the same time,
drm_connector_commit would first try to modeset then try to commit. This
won't work because both will trigger a page-flip. KMS will reject that.
Change the logic to only enable an output if no modeset has been
requested. The logic in wlr_output already checks that the user isn't
doing a modeset and disabling the output at the same time.
We have a workaround for legacy drivers that support drmModeSetCursor
without exposing a cursor plane. It doesn't work anymore now that we've
moved to a more atomic interface.
Let's just remove this workaround and fallback to software cursors.
Closes: https://github.com/swaywm/wlroots/issues/2166
conn->crtc is NULL in case the output is disabled.
However, the DRM backend will set the GAMMA_LUT property anyway. On each
commit the whole state is sent to KMS. Adding WLR_DRM_CRTC_GAMMA_LUT to
the pending state would just make the backend re-create the blob
containing the gamma LUT.
We don't need a per-CRTC atomic request anymore. Let's make the request
per-commit so that it's easier to debug.
This is also groundwork for supporting wlr_output_test properly.
GAMMA_LUT_SIZE isn't an atomic property. It can be used with the legacy
interface too. So we can unify both codepaths and remove
wlr_drm_interface.crtc_get_gamma_size.
It's no guaranteed to exist though, so we still need to keep the
fallback.
When the headless backend uses an already-existing renderer, it doesn't
have ownership over the renderer. When the renderer is destroyed, the
headless backend needs to destroy itself.
The logs don't currently display the importance of a line. It's easy to
read "skipping page-flip" as a debug message instead of an error
message.
Change the error message to make it more explicit.
References: https://github.com/swaywm/wlroots/pull/2147
Reconnecting an external display fails on initial
page-flip. This occurs when the pageflip_pending flag
has been set on the connector at the time the monitor
is removed. Once the connector's CRTC has been deallocated,
page_flip_handler cannot find the connector by its crtc_id,
so pageflip_pending will remain set. This causes
initialization to fail when the monitor is reconnected
with the error messages:
Skipping pageflip on output 'X'
Failed to initialize renderer on connector 'X': initial page-flip failed
To fix this problem, clear the pageflip_pending flag
when cleaning up the connector.
This is simply a false sense of security, and is worse than just using
setuid. CAP_SYS_ADMIN is an extremely serious capability that is
effectively as powerful as root.
It also required users to be in the input group, which allows any
process to keylog the entire system.
Some primary planes don't support ARGB8888, they only support XRGB8888
(for instance on older Intel hardware). The DRM backend would fail to
initialize.
When that's the case, try to allocate buffers without an alpha channel.
This is a type which manages gbm_surfaces and imported dmabufs in the
same place, and makes the lifetime management between the two shared. It
should lead to easier to understand code, and fewer special cases.
This also contains a fair bit of refactoring to start using this new
type.
Co-authored-by: Simon Ser <contact@emersion.fr>
This commit doesn't fix any issue that I'm aware of, since sway
incidentally does not use these fields. Still, they exist, so they
should probably be filled in to prevent fun surprises in the future.
Most of the pending output state is not forwarded to the backend prior
to an output commit. For instance, wlr_output_set_mode just stashes the
mode without calling any wlr_output_impl function.
wlr_output_impl.commit is responsible for applying the pending mode.
However, there are exceptions to this rule. The first one is
wlr_output_attach_render. It won't go away before renderer v6 is
complete, because it needs to set the current EGL surface.
The second one is wlr_output_attach_buffer.
wlr_output_impl.attach_buffer is removed in [1].
When wlr_output_rollback is called, all pending state is supposed to be
cleared. This works for all the state except the two exceptions
mentionned above. To fix this, introduce wlr_output_impl.rollback.
Right now, the backend resets the current EGL surface. This prevents GL
commands from affecting the output after wlr_output_rollback.
This patch is required for FBO-based outputs to work properly. The
compositor might be using FBOs for its own purposes [2], having leftover
FBO state can have bad consequences.
[1]: https://github.com/swaywm/wlroots/pull/2097
[2]: https://github.com/swaywm/wlroots/pull/2063#issuecomment-597614312
Check that buffer can be scanned out in wlr_output_test instead of
wlr_output_attach_buffer. This allows the backend to have access to the
whole pending state when performing the check.
This brings the wlr_output API more in line with the KMS API.
This removes the need for wlr_output_attach_buffer to return a value,
and for wlr_output_impl.attach_buffer.
Consumers call wlr_buffer_lock. Once all consumers are done with the
buffer, only the producer should have a reference to the buffer. In this
case, we can release the buffer (and let the producer re-use it).
The logind provider defaulted to systemd and in order to use elogind,
-Dlogin-provider=elogind was required. This adds 'auto' as a choice
for the login-provider option and sets it as default. Using 'auto',
the build will check for systemd first and if it's not found, try
to find and use elogind automatically.
Previously, each time a wl_seat.capabilities event was received the
Wayland backend created new input devices. It now only does so the first
time.
Input devices are now destroyed when the cap is removed.
Closes: https://github.com/swaywm/sway/issues/5055
This function allowed backends to provide a custom function for frame
scheduling. Before resuming the rendering loop, the DRM and Wayland
backends would wait for vsync.
There isn't a clear benefit of doing this. The only upside is that we
get more stable timings: the delay between two repaints doesn't change too
much and is close to a mutliple of the refresh rate.
However this introduces latency, especially when a client misses a
frame. For instance a fullscreen game missing vblank will need to wait
more than a whole frame before being able to display new content. This
worst case scenario happens as follows:
- Client is still rendering its frame and cannot submit it in time
- Deadline is reached
- Compositor decides to stop the rendering loop since nothing changed on
screen
- Client finally manages to render its frame, submits it
- Compositor calls wlr_output_schedule_frame
- DRM backend waits for next vblank
- The wlr_output frame event is fired, compositor draws new content on screen
- On the second next vblank, the new content reaches the screen
With this patch, the wlr_output frame event is fired immediately when
the client submits its late frame.
This change also makes it easier to support variable refresh rate, since
VRR is all about being able to present too-late frames earlier.
References: https://github.com/swaywm/wlroots/issues/1925
Simon Ser
Simon Zeni
Kenny Levinsen
Ilia Bozhinov
Ariadne Conill
Ronan Pigott
Stephane Chauveau
Isaac Freund
Mykola Orliuk
Marten Ringwelski
nerdopolis
Patrick Steinhardt
Ryan Walklin
Devin J. Pohly
Antonin Décimo
Andri Yngvason
Tudor Brindus
JonnyMako
Will Daly
Scott Anderson
Scott Moreau
Filip Sandborg