The first time the GBM backend's EGL context is made current after
creation, both the read and draw surfaces are set to EGL_NO_SURFACE.
This will set the GL read and draw buffers to GL_NONE in accordance with
the EGL spec.
When a real surface is later made current, however, the spec is arguably
unclear on whether the read and draw buffers should remain set to
GL_NONE or whether they should be restored to the default GL_BACK. The
Mesa driver does the latter, the NVIDIA driver does the former.
To work around this difference, Kwin has an explicit call to
glDrawBuffer in GbmSurface::makeContextCurrent. It does not have a
corresponding call to glReadBuffer, though, which can cause some desktop
effects such as background contrast to render incorrectly with the
NVIDIA driver. This change adds that missing call.
Releasing the buffers is necessary for example in the case of a GPU reset,
to make sure that the gbm surface is still properly destroyed and all buffers
with invalid content freed.
Instead of buffers being both drm framebuffers and gbm / dumb buffers, these
responsibilities are now split, which makes it possible to do zero copy
screen casting in the future.
Both framebuffers and gbm / dumb buffers also now always hold a valid underlying
resource, which simplifies code a bit.
libinput_device_get_user_data() can be used to get the associated Device
object with libinput_device. That way, we won't need to maintain a
private list of all input devices.
Format modifiers enable the graphics hardware to be much more efficient,
especially when it comes to multi-gpu transfers. With the issues regarding
bandwidth limits now solved, enable them by default to make all supported
systems benefit from them.
CCBUG: 452397
CCBUG: 452219
When explicit modifiers are used, it can happen that Mesa chooses modifiers that make
the display hardware hit bandwidth limits. In that case, atomic tests fail and the
outputs don't work, or KWin may even crash.
In order to work around that, DrmGpu now removes the used modifier whenever an atomic
test fails, and tries to find a working combination of outputs and modifiers.
While having all state be public is great for avoiding the boilerplate that
comes with setters and getters, it also exposes more state than necessary
to the rest of the backend and makes it more error-prone if more than one
part of the state needs to be changed at the same time.
Instead of passing all possible field values to the initialize()
function, pass all relevant data in a struct. With designated
initializers, it's more readable and makes code more comprehensible.
The general goal is to split Output's data in two categories - general
information about the output (e.g. edid) and mutable state (position,
mode, etc).
With this, the drm backend will be able to associate drmModeModeInfo
with Output's modes, which can be useful if there are several modes with
the same resolution and refresh rate but different flags.
This makes KWin switch to in-tree copy of KWaylandServer codebase.
KWaylandServer namespace has been left as is. It will be addressed later
by renaming classes in order to fit in the KWin namespace.
Instead of creating a gammaramp object with a fixed size, make the color
device create a color transformation object that can be used to construct
arbitrary LUTs. This is needed in order to support tiled displays well
and is useful for further color management work.
If the backend needs to apply custom logic when changing the transform,
it should override Platform::applyOutputChanges(); otherwise just update
the Output's internal transform state.
AbstractOutput is not so Abstract and it's common to avoid the word
"Abstract" in class names as it doesn't contribute any new information.
It also significantly reduces the line width in some places.
Input event flow has been refactored so all input events originate from
input devices.
The X11 backend uses InputRedirection so make it forward events to
relevant input device handlers.
Prefering alpha is needed for some interactions with video underlays; more
directly choosing formats with an alpha channel will be needed for overlay
planes, too.
The main motivation behind this change is to unify render target
representation across opengl and software renderers and avoid accessing
the render backend directory in order to get the render target.
GLRenderTarget doesn't provide a generic abstraction for framebuffer
objects, so let's call GLRenderTarget what it is - a framebuffer.
Renaming the GLRenderTarget class allows us to use the term "render
target" which abstracts fbos or shm images without creating confusion.
With these two actions being separate, RenderLoop can record the time spent
in endFrame (for example for multi-gpu transfers) without risking also recording
blocking swapbuffer calls, and endFrame can later be moved to output layer
In 5.24, the same code path is used for testing direct scanout, so that
causes false negatives. Generally though, the user setting shouldn't be
touched, it's not really proper feedback for the driver or KWin having
problems.
Some modes can have the exact frame timings but different flags.
Currently, the mode comparison function doesn't take that into account
which can result in the drm backend setting the current mode flag
incorrectly.
If the blob is fetched while there is no kernel-visible reference to it,
the driver may re-use the blob ID. When DrmProperty is created or updated,
KWin holds a reference on the blob via drmModeObjectProperties, so this
should prevent any possible issues.
CCBUG: 449285
Using the global coordinate system when specifying output layer damage
regions would be very confusing. In order to make the coordinate system
comprehensible, use the layer-local coordinate system.
The infinite region is used to tell the Compositor when it needs to
repaint the entire layer.
When modesets are necessary, they are attempted when an output on the given
GPU gets presented. With multi-gpu setups however, the situation can arise
where there is only one disabled output on a GPU; in that case KWin eternally
waits and never properly turns off the display.
In order to work around this, explicitly call DrmGpu::maybeModeset when
an output gets disabled.
BUG: 449878
FIXED-IN: 5.24.4
The .clang-format file is based on the one in ECM except the following
style options:
- AlwaysBreakBeforeMultilineStrings
- BinPackArguments
- BinPackParameters
- ColumnLimit
- BreakBeforeBraces
- KeepEmptyLinesAtTheStartOfBlocks
Virtual machines aren't properly supporting atomic mode setting yet, which
causes the cursor to be offset, and will cause more issues with overlay
planes. In order to prevent that from impacting users, fall back to legacy,
unless KWIN_DRM_NO_AMS is set.
BUG: 427060
FIXED-IN: 5.24.4
Hard to avoid as long as we don't have CI coverage yet, but that will take
a bit more time, we need KWin (and all its dependencies) to fully build
first.
The raspberry pi exposes opaque formats for the cursor plane, and interprets
them as being opaque as well... Considering that we effectively don't support
anything else with the QPainter anyways, just hardcode ARGB8888 until we paint
the cursor with OpenGl.
fbdev has been deprecated and unmaintained for a while. With Linux 5.14
including SimpleDRM driver, we can drop it. (at the time of writing this
commit message, the latest Linux version is 5.16).
[1/6] Make autotests create fake input devices
The goal of this patch set is simulating user input in unit tests via
InputDevices and no longer use the Platform to fake input. This matches
more closely with how input is processed when running a full plasma
wayland session, i.e. with the DRM and libinput backends.
Otherwise when we render it, we do so upside down and screen sharing
looks broken.
This only happens when the shadow buffer is in use, so it's not all that
common.
Instead of having the render backends manage layers, have DrmGpu and DrmPipeline
do it. This makes it possible to unify code paths for leased and normal
outputs, remove some redirection and have more freedom with assigning layers
to screens.
We already try to ensure that the surface damage is within render target
bounds. Avoid clipping surface damage in render backend, which is a bit
excessive task and perhaps it should be done an abstraction level above.
When casting from integer to pointer, promoting the integer to (u)intptr_t
will ensure that the resulting type can be converted to a pointer without
problems. These two casts changed in this commit trigger a warning when
building for CHERI-enabled architectures such as Arm Morello. This is not
just limited to CHERI, the cast from xcb_pixmap_t (uint32_t) to void*
should also be flagged by -Wint-to-void-pointer-cast when using Clang,
however, it appears that warning only handles C-style casts, and not
reinterpret_cast (https://github.com/llvm/llvm-project/issues/53964).
At this point, it's safe to assume that only X11 has weird rendering
model, which stands in the way of making rendering abstractions nice and
intuitive, so let's check operation mode. If OperationModeX11 is
dropped, this will also simplify finding X11-specific code in kwin.
The responsibilities of the Scene must be reduced to painting only so we
can move forward with the layer-based compositing.
This change moves direct scanout logic from the opengl scene to the base
scene class and the compositor. It makes the opengl scene less
overloaded and allows to share direct scanout logic.
Having a render loop in the Platform has always been awkward. Another
way to interpret the platform not supporting per screen rendering would
be that all outputs share the same render loop.
On X11, Scene::painted_screen is going to correspond to the primary
screen, we should not rely on this assumption though!
This allows us to make the GLRenderTarget a bit nicer when using it to
wrap the default fbo as we don't know what the color attachment texture
is besides its size.
This means that the responsibility of ensuring that the color attachment
outlives the fbo is now up to the caller. However, most of kwin code
has been written that way, so it's not an issue.
Because the GLRenderTarget and the GLVertexBuffer use the global
coordinate system, they are not ergonomic in render layers.
Assigning the device pixel ratio to GLRenderTarget and GLVertexBuffer is
an interesting api design choice too. Scaling is a window system
abstraction, which is absent in OpenGL or Vulkan. For example, it's not
possible to create an OpenGL texture with a scale factor of 2. It only
works with device pixels.
This change makes the GLRenderTarget and the GLVertexBuffer more
ergonomic for usages other than rendering the workspace by removing all
the global coordinate system and scaling stuff. That's the
responsibility of the users of those two classes.
In order to support layered rendering and tiled outputs KWin needs to be
able to split rendering of outputs into multiple surfaces. This commit
prepares the drm backend for that, by moving most of the code in EglGbmBackend
out to a EglGbmSurface class, which will later be used for overlay surfaces
and rendering to multiple connectors side by side.
In doing that, this commit also cleans up the code a bit, removes a lot of
now unnecessary multi-gpu stuff and potentially makes modesets a little
bit more efficient by re-using resources more often.
Currently, the Cursors::currentCursorChanged signal is wired to the
updateCursor() function which calls xcb_xfixes_hide_cursor() or
xcb_xfixes_show_cursor() depending if the cursor is hidden. However, the
currentCursorChanged signal can be emitted if the cursor changed, e.g. a
new pixmap attached, or its visibility status changes.
The zoom effect hides the pointer, but when user hovers ui elements, it
will most likely change and result in more than one xcb_xfixes_hide_cursor()
calls.
It appears like xcb_xfixes_hide_cursor() is implemented as a reference
counter, i.e. if xcb_xfixes_hide_cursor() is called two times, then you
must call xcb_xfixes_show_cursor() two times as well.
This change adds a dedicated signal to indicate whether the cursor is
hidden to avoid calling xcb_xfixes_hide_cursor() multiple times while
the screen is scaled.
BUG: 448537
They're error prone and don't really work for changing modes. Having
a current mode in DrmConnector also doesn't work well together with
the transactional style of how DrmPipeline operates
The implicit cast effectively rounds the value down, which make the
refresh rate be different from what KScreen actually wrote.
A better fix would be to use integers instead of floating point numbers
but that needs to happen in KScreen.
BUG: 448778
While in principle Mesa should already check if the buffer can be scanned
out, this may not always work. If we can't create a framebuffer object for
the buffer, fall back to compositing.
CCBUG: 448818
Encoders are not really relevant for the test result, except that one of
the encoders for the connector must be compatible with the crtc.
The kernel usually exposes only a single encoder per connector for this
reason, but if a driver exposes multiple then that means KWin will do a
lot more tests than is necessary.
In order to prevent that from happening, do fewer syscalls and simplify
code, only check supported encoders once per connector.
19c471405e7eb4b6026db24d776d205125dbc013 introduced a regression if
there are two gbm backend and the backend fail to choose drm format.
This fix does two things:
1. Current buffer format should not be reset after create new buffer,
otherwise current.format may just be empty after resetOutput.
2. force xrgb 8888 need to be set on the primary backend.
BUG: 448790
As the placeholder output gets added or removed in response to other outputs
getting enabled or disabled, the output list may change while iterating over
them and applying changes.
BUG: 448454
BUG: 448474
CCBUG: 448697
FIXED-IN: 5.24
As formats are per output and also checked on crtc changes, there
is no reason to restrict used formats to those that are supported
by all primary planes anymore.
When the crtcs get switched around between outputs, their primary
planes and thus the supported formats also get switched around. In
order to make sure that doesn't cause any problems, always check
whether or not the format+modifiers used are supported.
Userspace is expected to do a modeset and set link-status to good again,
if link-status gets bad. This is needed to prevent some black screen situations.
BUG: 448177
This adds support for reading values from a "Libinput/Defaults" group in
the input config file. This allows specifying global defaults for
devices, that are preferred over the libinput defaults. Because of the
cascading mechanisms of KConfig, this then allows distributions and
hardware vendors to supply system-wide defaults for devices.
Rather than an awkward combination of template functions, function
pointers and multiple calls to almost-the-same-but-not-entirely
functions, make ConfigData itself a template and use type erasure to
store them in the config map. This makes the data object aware of its
type and allows us to specialise the reading of config values through
template specialisation. It also removes the need for multiple
constructors and setters in the ConfigData object.
This change replaces abort() with Q_ASSERT and Q_UNREACHABLE() macros to
make kwin code base consistent. Besides that, Q_UNREACHABLE may
potentially provide the compiler more info that can be used to generate
more efficient machine code.
Instead of only allowing the current format, send the default tranches modified
to only contain formats and modifiers suitable for scanout.
In order to not fail when we can't do direct scanout with a given format
(because that may require a modeset, which we don't allow), keep a blacklist
of attempted formats and modifiers for the current client.
The fixed size of 1080p makes windows change their size and place. To
prevent that from happening, make the placeholder screen the same size
as the last disconnected screen.
This is not a bullet proof solution, only a fast one. Ideally KWin
should remember the window layout on a given monitor setup and restore
it when reconnected.
BUG: 447419
FIXED-IN: 5.24
On X11, if buffer age is unsupported, kwin can do some quirks like
copying parts of the front buffer to back buffer to avoid repainting the
whole screen.
Currently, the copying is performed in the opengl scene, which is not
perfect because it makes the scene responsible not only for painting the
scene but also some low level platform specific shenanigans.
This change moves the copying step to the glx and egl backends. It
simplifies the opengl scene, makes it less overloaded and more open to
changes, but it also duplicates code, which is not ideal. However, given
the de-facto deprecated state of the X11 platform, it's sort of acceptable
as the main focus is now on wayland session and the things that are
needed to make it fly as expected.
This commit makes 10 bits per color channel the default color depth, which
should reduce banding on outputs that support color formats with more than
8 bits per color channel. In order to support this the commit also removes
the dependency of the EglGbmBackend on a specific format and EglConfig and
instead makes those per-Output properties.
Currently, if you move the cursor really fast between two outputs, there
may be stuck cursor on the previous output.
We need to query the old cursor visibility status before updating the
cursor position, otherwise the drm backend may not schedule an update to
move the cursor offscreen.
This commit does away with the special handling of the legacy cursor and
makes it be applied directly in DrmPipeline, using the same state as the
atomic cursor and without dirty flags.
The cursor being set out-of-band with atomic commits creates problems
because it can create false positive for atomic tests, if the cursor
state gets changed in between an atomic test and its matching commit.
This commit also ports the cursor to a swapchain instead of only one
image. This is not strictly required but may prevent artifacts and
will be needed for future optimisations.
If eglSwapBuffers() fails, there won't be a buffer and so we need to
mark the frame as failed. Otherwise, the screen can be frozen.
eglSwapBuffers() can fail if some effect calls makeOpenGLContext()
between RenderBackend::beginFrame() and RenderBackend::endFrame(), which
is the case with the zoom effect. It can set wrong draw surface in
ZoomEffect::recreateTexture()
BUG: 445412
dmabuf-feedback allows the compositor to give the clients better feedback on what
formats and modifiers they should use, and for which device they should allocate.
This way they can reallocate for scanout whenever the compositor tells them to,
which makes direct scanout work for a lot more devices and applications.
While it could be useful with tiled displays, the isFormatSupported and
supportedModifier functions can be called before prepareModeset, so where
m_formats is still empty. Additionally they're neither in a hot path nor
performance critical.
Whether or not we want to use explicit modifiers for our surfaces doesn't
matter for what format+modifiers drm planes support. This way direct scanout
works by default, without having to explicitly enable modifiers
systemd takes care of setting and dropping master permissions when
sending PauseDevice and ResumeDevice signals.
When the ResumeDevice signal is received, the relevant drm device should
already have master permissions set up.
On the other hand, when the active property changes, there's still a
chance that systemd haven't granted drm master permissions to us.
In case a modeset needs to be performed, the drm backend will test all
pipelines to ensure that new mode won't cause any bandwidth issues on
other outputs, etc.
To do that, it may delay presenting frames. If the new configuration
doesn't work, it needs to notify about failed frames.
However, the relevant code that notifies the RenderLoop about failed
atomic commits doesn't check if there's actually a pending modeset
present.
When switching between VTs, systemd can revoke master permissions from
kwin. To make things even more trickier, kwin can try to present a frame
in that short time span.
EffectQuickScene is not used strictly by effects, aurorae decorations
use it too to render window decorations.
This change renames the EffectQuickView/Scene to
OffscreenQuickView/Scene to clear up the naming scheme.
Fixes a crash I have with dpms + suspend, which was caused by the udev
event for updating outputs being called before the output got enabled
again. When DrmGpu::updateOutputs got called it removed the crtc from
the inactive output and then disabled the output afterwards. Instead,
only remove crtcs if an output is really disabled.
This also allows to generalize the logic for lease outputs, and could
in the future allow for faster dpms on/off switching.
This unifies frame hooks for OpenGL and QPainter render backends. There
are a couple of reasons why it's a good idea - it provides one mental
framework to start painting a frame, the Compositor will be able to
start and submit frames. The last one is very cool because it gives the
Compositor more power over compositing.
Besides unifying frame hooks, this cleans up a bit the arg naming mess
in endFrame(). As is, "damage" and "damagedRegion" are very confusing
names. "damage" arg has been renamed to "renderedRegion," because that's
what it is. The renderedRegion arg specifies the region that has been
repainted by the Scene. It's different from the damagedRegion as that
one specifies the surface damage, i.e. the difference between the
current and the next frame, while the renderedRegion may include a
region that had to be repainted to repair the back buffer. The main
reason why we need renderedRegion is the X11 platform. On Wayland, it's
unused.
In the future, we will need to extend this api with output layers.
The ifdefs for have_gbm obfuscate the code unnecessarily - the drm backend
is not a great experience with qpainter, so in practice noone should ship
it without gbm anyways.
The Compositor contains nothing that can potentially get dirty and need
repainting.
As is, the advantages of this move aren't really noticeable, but it
makes sense with multiple scenes.
Backend parts are far from ideal, they can be improved later on as we
progress with the scene redesign.
The proprietary NVidia driver now supports gbm, which vastly improves the
user experience. For older devices that will not get gbm support dropping
EglStreams will likely not have a big impact as it has several session breaking
issues anyways.
By removing the backend a lot of logic can be simplified, most notably multi-gpu.
The main motivation behind this change is to move management of drm
blobs out of property wrappers in specialized wrappers to simplify state
management with blobs.
Connector mode blobs are created on demand.
When we switch CRTCs it can happen that a CRTC would stay enabled yet has
no connectors anymore. In this case the kernel may reject our atomic commit,
which would cause the modeset to fail. To counteract that, properly disable
unused drm objects
Currently KWin is combining modesets with presentation, which causes problems
when multiple monitors are used and crtcs need to be switched around, because
taking away a CRTC from another output causes the driver to disable the
other output. In order to avoid such problems, delay presentation until
all pipelines are ready to present and then do a modeset with a single atomic
commit. To process the resulting page flip events properly this commit also
ports KWin to page_flip_handler2 and changes how the pageFlipped and
notifyFrameFailed signals are processed.
Hardware constraints limit the number of crtcs and which connector + crtc
combinations can work together. The current code is searching for working
combinations when a hotplug happens but that's not enough, it also needs
to happen when the user enables or disables outputs and when modesets are
done, and the configuration change needs to be applied with a single atomic
commit.
This commit removes the hard dependency of DrmPipeline on crtcs by moving
the pending state of outputs from the drm objects to DrmPipeline itself,
which ensures that it's independent from the set of drm objects currently
used. It also changes requests from KScreen to be applied truly atomically.
This allows using base opengl backends in libkwin, which can be useful
later on for the purpose of moving the ownership of render backends from
the Scene class to the Compositor class.
This improves file organization in kwin by putting backends in a single
directory.
It also makes easier to discover kwin's low level components for new
contributors because the plugins directory may come as the last place to
look for. When one hears "plugin", the first thing that comes to mind is
regular plugins, not low level backends.
